diff --git a/sys/kern/subr_trap.c b/sys/kern/subr_trap.c index 8981091b50ed..d0f616d037c5 100644 --- a/sys/kern/subr_trap.c +++ b/sys/kern/subr_trap.c @@ -1,376 +1,386 @@ /*- * SPDX-License-Identifier: BSD-4-Clause * * Copyright (C) 1994, David Greenman * Copyright (c) 1990, 1993 * The Regents of the University of California. All rights reserved. * Copyright (c) 2007 The FreeBSD Foundation * * This code is derived from software contributed to Berkeley by * the University of Utah, and William Jolitz. * * Portions of this software were developed by A. Joseph Koshy under * sponsorship from the FreeBSD Foundation and Google, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)trap.c 7.4 (Berkeley) 5/13/91 */ #include __FBSDID("$FreeBSD$"); #include "opt_hwpmc_hooks.h" #include "opt_ktrace.h" #include "opt_sched.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KTRACE #include #include #endif #include #include #ifdef VIMAGE #include #endif #ifdef HWPMC_HOOKS #include #endif #include void (*softdep_ast_cleanup)(struct thread *); /* * Define the code needed before returning to user mode, for trap and * syscall. */ void userret(struct thread *td, struct trapframe *frame) { struct proc *p = td->td_proc; CTR3(KTR_SYSC, "userret: thread %p (pid %d, %s)", td, p->p_pid, td->td_name); KASSERT((p->p_flag & P_WEXIT) == 0, ("Exiting process returns to usermode")); #ifdef DIAGNOSTIC /* * Check that we called signotify() enough. For * multi-threaded processes, where signal distribution might * change due to other threads changing sigmask, the check is * racy and cannot be performed reliably. * If current process is vfork child, indicated by P_PPWAIT, then * issignal() ignores stops, so we block the check to avoid * classifying pending signals. */ if (p->p_numthreads == 1) { PROC_LOCK(p); thread_lock(td); if ((p->p_flag & P_PPWAIT) == 0 && (td->td_pflags & TDP_SIGFASTBLOCK) == 0) { if (SIGPENDING(td) && (td->td_flags & (TDF_NEEDSIGCHK | TDF_ASTPENDING)) != (TDF_NEEDSIGCHK | TDF_ASTPENDING)) { thread_unlock(td); panic( "failed to set signal flags for ast p %p td %p fl %x", p, td, td->td_flags); } } thread_unlock(td); PROC_UNLOCK(p); } #endif /* * Charge system time if profiling. */ if (__predict_false(p->p_flag & P_PROFIL)) addupc_task(td, TRAPF_PC(frame), td->td_pticks * psratio); #ifdef HWPMC_HOOKS if (PMC_THREAD_HAS_SAMPLES(td)) PMC_CALL_HOOK(td, PMC_FN_THR_USERRET, NULL); +#endif +#ifdef TCPHPTS + /* + * @gallatin is adament that this needs to go here, I + * am not so sure. Running hpts is a lot like + * a lro_flush() that happens while a user process + * is running. But he may know best so I will go + * with his view of accounting. :-) + */ + tcp_run_hpts(); #endif /* * Let the scheduler adjust our priority etc. */ sched_userret(td); /* * Check for misbehavior. * * In case there is a callchain tracing ongoing because of * hwpmc(4), skip the scheduler pinning check. * hwpmc(4) subsystem, infact, will collect callchain informations * at ast() checkpoint, which is past userret(). */ WITNESS_WARN(WARN_PANIC, NULL, "userret: returning"); KASSERT(td->td_critnest == 0, ("userret: Returning in a critical section")); KASSERT(td->td_locks == 0, ("userret: Returning with %d locks held", td->td_locks)); KASSERT(td->td_rw_rlocks == 0, ("userret: Returning with %d rwlocks held in read mode", td->td_rw_rlocks)); KASSERT(td->td_sx_slocks == 0, ("userret: Returning with %d sx locks held in shared mode", td->td_sx_slocks)); KASSERT(td->td_lk_slocks == 0, ("userret: Returning with %d lockmanager locks held in shared mode", td->td_lk_slocks)); KASSERT((td->td_pflags & TDP_NOFAULTING) == 0, ("userret: Returning with pagefaults disabled")); if (__predict_false(!THREAD_CAN_SLEEP())) { #ifdef EPOCH_TRACE epoch_trace_list(curthread); #endif KASSERT(0, ("userret: Returning with sleep disabled")); } KASSERT(td->td_pinned == 0 || (td->td_pflags & TDP_CALLCHAIN) != 0, ("userret: Returning with with pinned thread")); KASSERT(td->td_vp_reserved == NULL, ("userret: Returning with preallocated vnode")); KASSERT((td->td_flags & (TDF_SBDRY | TDF_SEINTR | TDF_SERESTART)) == 0, ("userret: Returning with stop signals deferred")); KASSERT(td->td_vslock_sz == 0, ("userret: Returning with vslock-wired space")); #ifdef VIMAGE /* Unfortunately td_vnet_lpush needs VNET_DEBUG. */ VNET_ASSERT(curvnet == NULL, ("%s: Returning on td %p (pid %d, %s) with vnet %p set in %s", __func__, td, p->p_pid, td->td_name, curvnet, (td->td_vnet_lpush != NULL) ? td->td_vnet_lpush : "N/A")); #endif } /* * Process an asynchronous software trap. * This is relatively easy. * This function will return with preemption disabled. */ void ast(struct trapframe *framep) { struct thread *td; struct proc *p; int flags, sig; bool resched_sigs; td = curthread; p = td->td_proc; CTR3(KTR_SYSC, "ast: thread %p (pid %d, %s)", td, p->p_pid, p->p_comm); KASSERT(TRAPF_USERMODE(framep), ("ast in kernel mode")); WITNESS_WARN(WARN_PANIC, NULL, "Returning to user mode"); mtx_assert(&Giant, MA_NOTOWNED); THREAD_LOCK_ASSERT(td, MA_NOTOWNED); td->td_frame = framep; td->td_pticks = 0; /* * This updates the td_flag's for the checks below in one * "atomic" operation with turning off the astpending flag. * If another AST is triggered while we are handling the * AST's saved in flags, the astpending flag will be set and * ast() will be called again. */ thread_lock(td); flags = td->td_flags; td->td_flags &= ~(TDF_ASTPENDING | TDF_NEEDSIGCHK | TDF_NEEDSUSPCHK | TDF_NEEDRESCHED | TDF_ALRMPEND | TDF_PROFPEND | TDF_MACPEND); thread_unlock(td); VM_CNT_INC(v_trap); if (td->td_cowgen != p->p_cowgen) thread_cow_update(td); if (td->td_pflags & TDP_OWEUPC && p->p_flag & P_PROFIL) { addupc_task(td, td->td_profil_addr, td->td_profil_ticks); td->td_profil_ticks = 0; td->td_pflags &= ~TDP_OWEUPC; } #ifdef HWPMC_HOOKS /* Handle Software PMC callchain capture. */ if (PMC_IS_PENDING_CALLCHAIN(td)) PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_USER_CALLCHAIN_SOFT, (void *) framep); #endif if (flags & TDF_ALRMPEND) { PROC_LOCK(p); kern_psignal(p, SIGVTALRM); PROC_UNLOCK(p); } if (flags & TDF_PROFPEND) { PROC_LOCK(p); kern_psignal(p, SIGPROF); PROC_UNLOCK(p); } #ifdef MAC if (flags & TDF_MACPEND) mac_thread_userret(td); #endif if (flags & TDF_NEEDRESCHED) { #ifdef KTRACE if (KTRPOINT(td, KTR_CSW)) ktrcsw(1, 1, __func__); #endif thread_lock(td); sched_prio(td, td->td_user_pri); mi_switch(SW_INVOL | SWT_NEEDRESCHED); #ifdef KTRACE if (KTRPOINT(td, KTR_CSW)) ktrcsw(0, 1, __func__); #endif } td_softdep_cleanup(td); MPASS(td->td_su == NULL); /* * If this thread tickled GEOM, we need to wait for the giggling to * stop before we return to userland */ if (__predict_false(td->td_pflags & TDP_GEOM)) g_waitidle(); #ifdef DIAGNOSTIC if (p->p_numthreads == 1 && (flags & TDF_NEEDSIGCHK) == 0) { PROC_LOCK(p); thread_lock(td); /* * Note that TDF_NEEDSIGCHK should be re-read from * td_flags, since signal might have been delivered * after we cleared td_flags above. This is one of * the reason for looping check for AST condition. * See comment in userret() about P_PPWAIT. */ if ((p->p_flag & P_PPWAIT) == 0 && (td->td_pflags & TDP_SIGFASTBLOCK) == 0) { if (SIGPENDING(td) && (td->td_flags & (TDF_NEEDSIGCHK | TDF_ASTPENDING)) != (TDF_NEEDSIGCHK | TDF_ASTPENDING)) { thread_unlock(td); /* fix dumps */ panic( "failed2 to set signal flags for ast p %p td %p fl %x %x", p, td, flags, td->td_flags); } } thread_unlock(td); PROC_UNLOCK(p); } #endif /* * Check for signals. Unlocked reads of p_pendingcnt or * p_siglist might cause process-directed signal to be handled * later. */ if (flags & TDF_NEEDSIGCHK || p->p_pendingcnt > 0 || !SIGISEMPTY(p->p_siglist)) { sigfastblock_fetch(td); PROC_LOCK(p); mtx_lock(&p->p_sigacts->ps_mtx); while ((sig = cursig(td)) != 0) { KASSERT(sig >= 0, ("sig %d", sig)); postsig(sig); } mtx_unlock(&p->p_sigacts->ps_mtx); PROC_UNLOCK(p); resched_sigs = true; } else { resched_sigs = false; } /* * Handle deferred update of the fast sigblock value, after * the postsig() loop was performed. */ sigfastblock_setpend(td, resched_sigs); #ifdef KTRACE KTRUSERRET(td); #endif /* * We need to check to see if we have to exit or wait due to a * single threading requirement or some other STOP condition. */ if (flags & TDF_NEEDSUSPCHK) { PROC_LOCK(p); thread_suspend_check(0); PROC_UNLOCK(p); } if (td->td_pflags & TDP_OLDMASK) { td->td_pflags &= ~TDP_OLDMASK; kern_sigprocmask(td, SIG_SETMASK, &td->td_oldsigmask, NULL, 0); } #ifdef RACCT if (__predict_false(racct_enable && p->p_throttled != 0)) racct_proc_throttled(p); #endif userret(td, framep); } const char * syscallname(struct proc *p, u_int code) { static const char unknown[] = "unknown"; struct sysentvec *sv; sv = p->p_sysent; if (sv->sv_syscallnames == NULL || code >= sv->sv_size) return (unknown); return (sv->sv_syscallnames[code]); } diff --git a/sys/netinet/in_pcb.h b/sys/netinet/in_pcb.h index 923c5ba4edff..c20c516628b8 100644 --- a/sys/netinet/in_pcb.h +++ b/sys/netinet/in_pcb.h @@ -1,897 +1,899 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1990, 1993 * The Regents of the University of California. * Copyright (c) 2010-2011 Juniper Networks, Inc. * All rights reserved. * * Portions of this software were developed by Robert N. M. Watson under * contract to Juniper Networks, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)in_pcb.h 8.1 (Berkeley) 6/10/93 * $FreeBSD$ */ #ifndef _NETINET_IN_PCB_H_ #define _NETINET_IN_PCB_H_ #include #include #include #include #include #include #ifdef _KERNEL #include #include #include #include #endif #include /* * struct inpcb is the common protocol control block structure used in most * IP transport protocols. * * Pointers to local and foreign host table entries, local and foreign socket * numbers, and pointers up (to a socket structure) and down (to a * protocol-specific control block) are stored here. */ CK_LIST_HEAD(inpcbhead, inpcb); CK_LIST_HEAD(inpcbporthead, inpcbport); CK_LIST_HEAD(inpcblbgrouphead, inpcblbgroup); typedef uint64_t inp_gen_t; /* * PCB with AF_INET6 null bind'ed laddr can receive AF_INET input packet. * So, AF_INET6 null laddr is also used as AF_INET null laddr, by utilizing * the following structure. */ struct in_addr_4in6 { u_int32_t ia46_pad32[3]; struct in_addr ia46_addr4; }; union in_dependaddr { struct in_addr_4in6 id46_addr; struct in6_addr id6_addr; }; /* * NOTE: ipv6 addrs should be 64-bit aligned, per RFC 2553. in_conninfo has * some extra padding to accomplish this. * NOTE 2: tcp_syncache.c uses first 5 32-bit words, which identify fport, * lport, faddr to generate hash, so these fields shouldn't be moved. */ struct in_endpoints { u_int16_t ie_fport; /* foreign port */ u_int16_t ie_lport; /* local port */ /* protocol dependent part, local and foreign addr */ union in_dependaddr ie_dependfaddr; /* foreign host table entry */ union in_dependaddr ie_dependladdr; /* local host table entry */ #define ie_faddr ie_dependfaddr.id46_addr.ia46_addr4 #define ie_laddr ie_dependladdr.id46_addr.ia46_addr4 #define ie6_faddr ie_dependfaddr.id6_addr #define ie6_laddr ie_dependladdr.id6_addr u_int32_t ie6_zoneid; /* scope zone id */ }; /* * XXX The defines for inc_* are hacks and should be changed to direct * references. */ struct in_conninfo { u_int8_t inc_flags; u_int8_t inc_len; u_int16_t inc_fibnum; /* XXX was pad, 16 bits is plenty */ /* protocol dependent part */ struct in_endpoints inc_ie; }; /* * Flags for inc_flags. */ #define INC_ISIPV6 0x01 #define INC_IPV6MINMTU 0x02 #define inc_fport inc_ie.ie_fport #define inc_lport inc_ie.ie_lport #define inc_faddr inc_ie.ie_faddr #define inc_laddr inc_ie.ie_laddr #define inc6_faddr inc_ie.ie6_faddr #define inc6_laddr inc_ie.ie6_laddr #define inc6_zoneid inc_ie.ie6_zoneid #if defined(_KERNEL) || defined(_WANT_INPCB) /* * struct inpcb captures the network layer state for TCP, UDP, and raw IPv4 and * IPv6 sockets. In the case of TCP and UDP, further per-connection state is * hung off of inp_ppcb most of the time. Almost all fields of struct inpcb * are static after creation or protected by a per-inpcb rwlock, inp_lock. A * few fields are protected by multiple locks as indicated in the locking notes * below. For these fields, all of the listed locks must be write-locked for * any modifications. However, these fields can be safely read while any one of * the listed locks are read-locked. This model can permit greater concurrency * for read operations. For example, connections can be looked up while only * holding a read lock on the global pcblist lock. This is important for * performance when attempting to find the connection for a packet given its IP * and port tuple. * * One noteworthy exception is that the global pcbinfo lock follows a different * set of rules in relation to the inp_list field. Rather than being * write-locked for modifications and read-locked for list iterations, it must * be read-locked during modifications and write-locked during list iterations. * This ensures that the relatively rare global list iterations safely walk a * stable snapshot of connections while allowing more common list modifications * to safely grab the pcblist lock just while adding or removing a connection * from the global list. * * Key: * (b) - Protected by the hpts lock. * (c) - Constant after initialization * (e) - Protected by the net_epoch_prempt epoch * (g) - Protected by the pcbgroup lock * (i) - Protected by the inpcb lock * (p) - Protected by the pcbinfo lock for the inpcb * (l) - Protected by the pcblist lock for the inpcb * (h) - Protected by the pcbhash lock for the inpcb * (s) - Protected by another subsystem's locks * (x) - Undefined locking * * Notes on the tcp_hpts: * * First Hpts lock order is * 1) INP_WLOCK() * 2) HPTS_LOCK() i.e. hpts->pmtx * * To insert a TCB on the hpts you *must* be holding the INP_WLOCK(). * You may check the inp->inp_in_hpts flag without the hpts lock. * The hpts is the only one that will clear this flag holding * only the hpts lock. This means that in your tcp_output() * routine when you test for the inp_in_hpts flag to be 1 * it may be transitioning to 0 (by the hpts). * That's ok since that will just mean an extra call to tcp_output * that most likely will find the call you executed * (when the mis-match occured) will have put the TCB back * on the hpts and it will return. If your * call did not add the inp back to the hpts then you will either * over-send or the cwnd will block you from sending more. * * Note you should also be holding the INP_WLOCK() when you * call the remove from the hpts as well. Though usually * you are either doing this from a timer, where you need and have * the INP_WLOCK() or from destroying your TCB where again * you should already have the INP_WLOCK(). * * The inp_hpts_cpu, inp_hpts_cpu_set, inp_input_cpu and * inp_input_cpu_set fields are controlled completely by * the hpts. Do not ever set these. The inp_hpts_cpu_set * and inp_input_cpu_set fields indicate if the hpts has * setup the respective cpu field. It is advised if this * field is 0, to enqueue the packet with the appropriate * hpts_immediate() call. If the _set field is 1, then * you may compare the inp_*_cpu field to the curcpu and * may want to again insert onto the hpts if these fields * are not equal (i.e. you are not on the expected CPU). * * A note on inp_hpts_calls and inp_input_calls, these * flags are set when the hpts calls either the output * or do_segment routines respectively. If the routine * being called wants to use this, then it needs to * clear the flag before returning. The hpts will not * clear the flag. The flags can be used to tell if * the hpts is the function calling the respective * routine. * * A few other notes: * * When a read lock is held, stability of the field is guaranteed; to write * to a field, a write lock must generally be held. * * netinet/netinet6-layer code should not assume that the inp_socket pointer * is safe to dereference without inp_lock being held, even for protocols * other than TCP (where the inpcb persists during TIMEWAIT even after the * socket has been freed), or there may be close(2)-related races. * * The inp_vflag field is overloaded, and would otherwise ideally be (c). * * TODO: Currently only the TCP stack is leveraging the global pcbinfo lock * read-lock usage during modification, this model can be applied to other * protocols (especially SCTP). */ struct icmp6_filter; struct inpcbpolicy; struct m_snd_tag; struct inpcb { /* Cache line #1 (amd64) */ CK_LIST_ENTRY(inpcb) inp_hash; /* [w](h/i) [r](e/i) hash list */ CK_LIST_ENTRY(inpcb) inp_pcbgrouphash; /* (g/i) hash list */ struct rwlock inp_lock; /* Cache line #2 (amd64) */ #define inp_start_zero inp_hpts #define inp_zero_size (sizeof(struct inpcb) - \ offsetof(struct inpcb, inp_start_zero)) TAILQ_ENTRY(inpcb) inp_hpts; /* pacing out queue next lock(b) */ uint32_t inp_hpts_request; /* Current hpts request, zero if * fits in the pacing window (i&b). */ /* * Note the next fields are protected by a * different lock (hpts-lock). This means that * they must correspond in size to the smallest * protectable bit field (uint8_t on x86, and * other platfomrs potentially uint32_t?). Also * since CPU switches can occur at different times the two * fields can *not* be collapsed into a signal bit field. */ #if defined(__amd64__) || defined(__i386__) volatile uint8_t inp_in_hpts; /* on output hpts (lock b) */ volatile uint8_t inp_in_input; /* on input hpts (lock b) */ #else volatile uint32_t inp_in_hpts; /* on output hpts (lock b) */ volatile uint32_t inp_in_input; /* on input hpts (lock b) */ #endif volatile uint16_t inp_hpts_cpu; /* Lock (i) */ + volatile uint16_t inp_irq_cpu; /* Set by LRO in behalf of or the driver */ u_int inp_refcount; /* (i) refcount */ int inp_flags; /* (i) generic IP/datagram flags */ int inp_flags2; /* (i) generic IP/datagram flags #2*/ volatile uint16_t inp_input_cpu; /* Lock (i) */ volatile uint8_t inp_hpts_cpu_set :1, /* on output hpts (i) */ inp_input_cpu_set : 1, /* on input hpts (i) */ inp_hpts_calls :1, /* (i) from output hpts */ inp_input_calls :1, /* (i) from input hpts */ - inp_spare_bits2 : 4; + inp_irq_cpu_set :1, /* (i) from LRO/Driver */ + inp_spare_bits2 : 3; uint8_t inp_numa_domain; /* numa domain */ void *inp_ppcb; /* (i) pointer to per-protocol pcb */ struct socket *inp_socket; /* (i) back pointer to socket */ uint32_t inp_hptsslot; /* Hpts wheel slot this tcb is Lock(i&b) */ uint32_t inp_hpts_drop_reas; /* reason we are dropping the PCB (lock i&b) */ TAILQ_ENTRY(inpcb) inp_input; /* pacing in queue next lock(b) */ struct inpcbinfo *inp_pcbinfo; /* (c) PCB list info */ struct inpcbgroup *inp_pcbgroup; /* (g/i) PCB group list */ CK_LIST_ENTRY(inpcb) inp_pcbgroup_wild; /* (g/i/h) group wildcard entry */ struct ucred *inp_cred; /* (c) cache of socket cred */ u_int32_t inp_flow; /* (i) IPv6 flow information */ u_char inp_vflag; /* (i) IP version flag (v4/v6) */ u_char inp_ip_ttl; /* (i) time to live proto */ u_char inp_ip_p; /* (c) protocol proto */ u_char inp_ip_minttl; /* (i) minimum TTL or drop */ uint32_t inp_flowid; /* (x) flow id / queue id */ struct m_snd_tag *inp_snd_tag; /* (i) send tag for outgoing mbufs */ uint32_t inp_flowtype; /* (x) M_HASHTYPE value */ uint32_t inp_rss_listen_bucket; /* (x) overridden RSS listen bucket */ /* Local and foreign ports, local and foreign addr. */ struct in_conninfo inp_inc; /* (i) list for PCB's local port */ /* MAC and IPSEC policy information. */ struct label *inp_label; /* (i) MAC label */ struct inpcbpolicy *inp_sp; /* (s) for IPSEC */ /* Protocol-dependent part; options. */ struct { u_char inp_ip_tos; /* (i) type of service proto */ struct mbuf *inp_options; /* (i) IP options */ struct ip_moptions *inp_moptions; /* (i) mcast options */ }; struct { /* (i) IP options */ struct mbuf *in6p_options; /* (i) IP6 options for outgoing packets */ struct ip6_pktopts *in6p_outputopts; /* (i) IP multicast options */ struct ip6_moptions *in6p_moptions; /* (i) ICMPv6 code type filter */ struct icmp6_filter *in6p_icmp6filt; /* (i) IPV6_CHECKSUM setsockopt */ int in6p_cksum; short in6p_hops; }; CK_LIST_ENTRY(inpcb) inp_portlist; /* (i/h) */ struct inpcbport *inp_phd; /* (i/h) head of this list */ inp_gen_t inp_gencnt; /* (c) generation count */ void *spare_ptr; /* Spare pointer. */ rt_gen_t inp_rt_cookie; /* generation for route entry */ union { /* cached L3 information */ struct route inp_route; struct route_in6 inp_route6; }; CK_LIST_ENTRY(inpcb) inp_list; /* (p/l) list for all PCBs for proto */ /* (e[r]) for list iteration */ /* (p[w]/l) for addition/removal */ struct epoch_context inp_epoch_ctx; }; #endif /* _KERNEL */ #define inp_fport inp_inc.inc_fport #define inp_lport inp_inc.inc_lport #define inp_faddr inp_inc.inc_faddr #define inp_laddr inp_inc.inc_laddr #define in6p_faddr inp_inc.inc6_faddr #define in6p_laddr inp_inc.inc6_laddr #define in6p_zoneid inp_inc.inc6_zoneid #define inp_vnet inp_pcbinfo->ipi_vnet /* * The range of the generation count, as used in this implementation, is 9e19. * We would have to create 300 billion connections per second for this number * to roll over in a year. This seems sufficiently unlikely that we simply * don't concern ourselves with that possibility. */ /* * Interface exported to userland by various protocols which use inpcbs. Hack * alert -- only define if struct xsocket is in scope. * Fields prefixed with "xi_" are unique to this structure, and the rest * match fields in the struct inpcb, to ease coding and porting. * * Legend: * (s) - used by userland utilities in src * (p) - used by utilities in ports * (3) - is known to be used by third party software not in ports * (n) - no known usage */ #ifdef _SYS_SOCKETVAR_H_ struct xinpcb { ksize_t xi_len; /* length of this structure */ struct xsocket xi_socket; /* (s,p) */ struct in_conninfo inp_inc; /* (s,p) */ uint64_t inp_gencnt; /* (s,p) */ kvaddr_t inp_ppcb; /* (s) netstat(1) */ int64_t inp_spare64[4]; uint32_t inp_flow; /* (s) */ uint32_t inp_flowid; /* (s) */ uint32_t inp_flowtype; /* (s) */ int32_t inp_flags; /* (s,p) */ int32_t inp_flags2; /* (s) */ int32_t inp_rss_listen_bucket; /* (n) */ int32_t in6p_cksum; /* (n) */ int32_t inp_spare32[4]; uint16_t in6p_hops; /* (n) */ uint8_t inp_ip_tos; /* (n) */ int8_t pad8; uint8_t inp_vflag; /* (s,p) */ uint8_t inp_ip_ttl; /* (n) */ uint8_t inp_ip_p; /* (n) */ uint8_t inp_ip_minttl; /* (n) */ int8_t inp_spare8[4]; } __aligned(8); struct xinpgen { ksize_t xig_len; /* length of this structure */ u_int xig_count; /* number of PCBs at this time */ uint32_t _xig_spare32; inp_gen_t xig_gen; /* generation count at this time */ so_gen_t xig_sogen; /* socket generation count this time */ uint64_t _xig_spare64[4]; } __aligned(8); #ifdef _KERNEL void in_pcbtoxinpcb(const struct inpcb *, struct xinpcb *); #endif #endif /* _SYS_SOCKETVAR_H_ */ struct inpcbport { struct epoch_context phd_epoch_ctx; CK_LIST_ENTRY(inpcbport) phd_hash; struct inpcbhead phd_pcblist; u_short phd_port; }; struct in_pcblist { int il_count; struct epoch_context il_epoch_ctx; struct inpcbinfo *il_pcbinfo; struct inpcb *il_inp_list[0]; }; /*- * Global data structure for each high-level protocol (UDP, TCP, ...) in both * IPv4 and IPv6. Holds inpcb lists and information for managing them. * * Each pcbinfo is protected by three locks: ipi_lock, ipi_hash_lock and * ipi_list_lock: * - ipi_lock covering the global pcb list stability during loop iteration, * - ipi_hash_lock covering the hashed lookup tables, * - ipi_list_lock covering mutable global fields (such as the global * pcb list) * * The lock order is: * * ipi_lock (before) * inpcb locks (before) * ipi_list locks (before) * {ipi_hash_lock, pcbgroup locks} * * Locking key: * * (c) Constant or nearly constant after initialisation * (e) - Protected by the net_epoch_prempt epoch * (g) Locked by ipi_lock * (l) Locked by ipi_list_lock * (h) Read using either net_epoch_preempt or inpcb lock; write requires both ipi_hash_lock and inpcb lock * (p) Protected by one or more pcbgroup locks * (x) Synchronisation properties poorly defined */ struct inpcbinfo { /* * Global lock protecting inpcb list modification */ struct mtx ipi_lock; /* * Global list of inpcbs on the protocol. */ struct inpcbhead *ipi_listhead; /* [r](e) [w](g/l) */ u_int ipi_count; /* (l) */ /* * Generation count -- incremented each time a connection is allocated * or freed. */ u_quad_t ipi_gencnt; /* (l) */ /* * Fields associated with port lookup and allocation. */ u_short ipi_lastport; /* (x) */ u_short ipi_lastlow; /* (x) */ u_short ipi_lasthi; /* (x) */ /* * UMA zone from which inpcbs are allocated for this protocol. */ struct uma_zone *ipi_zone; /* (c) */ /* * Connection groups associated with this protocol. These fields are * constant, but pcbgroup structures themselves are protected by * per-pcbgroup locks. */ struct inpcbgroup *ipi_pcbgroups; /* (c) */ u_int ipi_npcbgroups; /* (c) */ u_int ipi_hashfields; /* (c) */ /* * Global lock protecting modification non-pcbgroup hash lookup tables. */ struct mtx ipi_hash_lock; /* * Global hash of inpcbs, hashed by local and foreign addresses and * port numbers. */ struct inpcbhead *ipi_hashbase; /* (h) */ u_long ipi_hashmask; /* (h) */ /* * Global hash of inpcbs, hashed by only local port number. */ struct inpcbporthead *ipi_porthashbase; /* (h) */ u_long ipi_porthashmask; /* (h) */ /* * List of wildcard inpcbs for use with pcbgroups. In the past, was * per-pcbgroup but is now global. All pcbgroup locks must be held * to modify the list, so any is sufficient to read it. */ struct inpcbhead *ipi_wildbase; /* (p) */ u_long ipi_wildmask; /* (p) */ /* * Load balance groups used for the SO_REUSEPORT_LB option, * hashed by local port. */ struct inpcblbgrouphead *ipi_lbgrouphashbase; /* (h) */ u_long ipi_lbgrouphashmask; /* (h) */ /* * Pointer to network stack instance */ struct vnet *ipi_vnet; /* (c) */ /* * general use 2 */ void *ipi_pspare[2]; /* * Global lock protecting global inpcb list, inpcb count, etc. */ struct rwlock ipi_list_lock; }; #ifdef _KERNEL /* * Connection groups hold sets of connections that have similar CPU/thread * affinity. Each connection belongs to exactly one connection group. */ struct inpcbgroup { /* * Per-connection group hash of inpcbs, hashed by local and foreign * addresses and port numbers. */ struct inpcbhead *ipg_hashbase; /* (c) */ u_long ipg_hashmask; /* (c) */ /* * Notional affinity of this pcbgroup. */ u_int ipg_cpu; /* (p) */ /* * Per-connection group lock, not to be confused with ipi_lock. * Protects the hash table hung off the group, but also the global * wildcard list in inpcbinfo. */ struct mtx ipg_lock; } __aligned(CACHE_LINE_SIZE); /* * Load balance groups used for the SO_REUSEPORT_LB socket option. Each group * (or unique address:port combination) can be re-used at most * INPCBLBGROUP_SIZMAX (256) times. The inpcbs are stored in il_inp which * is dynamically resized as processes bind/unbind to that specific group. */ struct inpcblbgroup { CK_LIST_ENTRY(inpcblbgroup) il_list; struct epoch_context il_epoch_ctx; uint16_t il_lport; /* (c) */ u_char il_vflag; /* (c) */ u_int8_t il_numa_domain; uint32_t il_pad2; union in_dependaddr il_dependladdr; /* (c) */ #define il_laddr il_dependladdr.id46_addr.ia46_addr4 #define il6_laddr il_dependladdr.id6_addr uint32_t il_inpsiz; /* max count in il_inp[] (h) */ uint32_t il_inpcnt; /* cur count in il_inp[] (h) */ struct inpcb *il_inp[]; /* (h) */ }; #define INP_LOCK_INIT(inp, d, t) \ rw_init_flags(&(inp)->inp_lock, (t), RW_RECURSE | RW_DUPOK) #define INP_LOCK_DESTROY(inp) rw_destroy(&(inp)->inp_lock) #define INP_RLOCK(inp) rw_rlock(&(inp)->inp_lock) #define INP_WLOCK(inp) rw_wlock(&(inp)->inp_lock) #define INP_TRY_RLOCK(inp) rw_try_rlock(&(inp)->inp_lock) #define INP_TRY_WLOCK(inp) rw_try_wlock(&(inp)->inp_lock) #define INP_RUNLOCK(inp) rw_runlock(&(inp)->inp_lock) #define INP_WUNLOCK(inp) rw_wunlock(&(inp)->inp_lock) #define INP_UNLOCK(inp) rw_unlock(&(inp)->inp_lock) #define INP_TRY_UPGRADE(inp) rw_try_upgrade(&(inp)->inp_lock) #define INP_DOWNGRADE(inp) rw_downgrade(&(inp)->inp_lock) #define INP_WLOCKED(inp) rw_wowned(&(inp)->inp_lock) #define INP_LOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_LOCKED) #define INP_RLOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_RLOCKED) #define INP_WLOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_WLOCKED) #define INP_UNLOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_UNLOCKED) /* * These locking functions are for inpcb consumers outside of sys/netinet, * more specifically, they were added for the benefit of TOE drivers. The * macros are reserved for use by the stack. */ void inp_wlock(struct inpcb *); void inp_wunlock(struct inpcb *); void inp_rlock(struct inpcb *); void inp_runlock(struct inpcb *); #ifdef INVARIANT_SUPPORT void inp_lock_assert(struct inpcb *); void inp_unlock_assert(struct inpcb *); #else #define inp_lock_assert(inp) do {} while (0) #define inp_unlock_assert(inp) do {} while (0) #endif void inp_apply_all(void (*func)(struct inpcb *, void *), void *arg); int inp_ip_tos_get(const struct inpcb *inp); void inp_ip_tos_set(struct inpcb *inp, int val); struct socket * inp_inpcbtosocket(struct inpcb *inp); struct tcpcb * inp_inpcbtotcpcb(struct inpcb *inp); void inp_4tuple_get(struct inpcb *inp, uint32_t *laddr, uint16_t *lp, uint32_t *faddr, uint16_t *fp); int inp_so_options(const struct inpcb *inp); #endif /* _KERNEL */ #define INP_INFO_LOCK_INIT(ipi, d) \ mtx_init(&(ipi)->ipi_lock, (d), NULL, MTX_DEF| MTX_RECURSE) #define INP_INFO_LOCK_DESTROY(ipi) mtx_destroy(&(ipi)->ipi_lock) #define INP_INFO_WLOCK(ipi) mtx_lock(&(ipi)->ipi_lock) #define INP_INFO_TRY_WLOCK(ipi) mtx_trylock(&(ipi)->ipi_lock) #define INP_INFO_WLOCKED(ipi) mtx_owned(&(ipi)->ipi_lock) #define INP_INFO_WUNLOCK(ipi) mtx_unlock(&(ipi)->ipi_lock) #define INP_INFO_LOCK_ASSERT(ipi) MPASS(in_epoch(net_epoch_preempt) || mtx_owned(&(ipi)->ipi_lock)) #define INP_INFO_WLOCK_ASSERT(ipi) mtx_assert(&(ipi)->ipi_lock, MA_OWNED) #define INP_INFO_WUNLOCK_ASSERT(ipi) \ mtx_assert(&(ipi)->ipi_lock, MA_NOTOWNED) #define INP_LIST_LOCK_INIT(ipi, d) \ rw_init_flags(&(ipi)->ipi_list_lock, (d), 0) #define INP_LIST_LOCK_DESTROY(ipi) rw_destroy(&(ipi)->ipi_list_lock) #define INP_LIST_RLOCK(ipi) rw_rlock(&(ipi)->ipi_list_lock) #define INP_LIST_WLOCK(ipi) rw_wlock(&(ipi)->ipi_list_lock) #define INP_LIST_TRY_RLOCK(ipi) rw_try_rlock(&(ipi)->ipi_list_lock) #define INP_LIST_TRY_WLOCK(ipi) rw_try_wlock(&(ipi)->ipi_list_lock) #define INP_LIST_TRY_UPGRADE(ipi) rw_try_upgrade(&(ipi)->ipi_list_lock) #define INP_LIST_RUNLOCK(ipi) rw_runlock(&(ipi)->ipi_list_lock) #define INP_LIST_WUNLOCK(ipi) rw_wunlock(&(ipi)->ipi_list_lock) #define INP_LIST_LOCK_ASSERT(ipi) \ rw_assert(&(ipi)->ipi_list_lock, RA_LOCKED) #define INP_LIST_RLOCK_ASSERT(ipi) \ rw_assert(&(ipi)->ipi_list_lock, RA_RLOCKED) #define INP_LIST_WLOCK_ASSERT(ipi) \ rw_assert(&(ipi)->ipi_list_lock, RA_WLOCKED) #define INP_LIST_UNLOCK_ASSERT(ipi) \ rw_assert(&(ipi)->ipi_list_lock, RA_UNLOCKED) #define INP_HASH_LOCK_INIT(ipi, d) mtx_init(&(ipi)->ipi_hash_lock, (d), NULL, MTX_DEF) #define INP_HASH_LOCK_DESTROY(ipi) mtx_destroy(&(ipi)->ipi_hash_lock) #define INP_HASH_WLOCK(ipi) mtx_lock(&(ipi)->ipi_hash_lock) #define INP_HASH_WUNLOCK(ipi) mtx_unlock(&(ipi)->ipi_hash_lock) #define INP_HASH_LOCK_ASSERT(ipi) MPASS(in_epoch(net_epoch_preempt) || mtx_owned(&(ipi)->ipi_hash_lock)) #define INP_HASH_WLOCK_ASSERT(ipi) mtx_assert(&(ipi)->ipi_hash_lock, MA_OWNED); #define INP_GROUP_LOCK_INIT(ipg, d) mtx_init(&(ipg)->ipg_lock, (d), NULL, \ MTX_DEF | MTX_DUPOK) #define INP_GROUP_LOCK_DESTROY(ipg) mtx_destroy(&(ipg)->ipg_lock) #define INP_GROUP_LOCK(ipg) mtx_lock(&(ipg)->ipg_lock) #define INP_GROUP_LOCK_ASSERT(ipg) mtx_assert(&(ipg)->ipg_lock, MA_OWNED) #define INP_GROUP_UNLOCK(ipg) mtx_unlock(&(ipg)->ipg_lock) #define INP_PCBHASH(faddr, lport, fport, mask) \ (((faddr) ^ ((faddr) >> 16) ^ ntohs((lport) ^ (fport))) & (mask)) #define INP_PCBPORTHASH(lport, mask) \ (ntohs((lport)) & (mask)) #define INP_PCBLBGROUP_PKTHASH(faddr, lport, fport) \ ((faddr) ^ ((faddr) >> 16) ^ ntohs((lport) ^ (fport))) #define INP6_PCBHASHKEY(faddr) ((faddr)->s6_addr32[3]) /* * Flags for inp_vflags -- historically version flags only */ #define INP_IPV4 0x1 #define INP_IPV6 0x2 #define INP_IPV6PROTO 0x4 /* opened under IPv6 protocol */ /* * Flags for inp_flags. */ #define INP_RECVOPTS 0x00000001 /* receive incoming IP options */ #define INP_RECVRETOPTS 0x00000002 /* receive IP options for reply */ #define INP_RECVDSTADDR 0x00000004 /* receive IP dst address */ #define INP_HDRINCL 0x00000008 /* user supplies entire IP header */ #define INP_HIGHPORT 0x00000010 /* user wants "high" port binding */ #define INP_LOWPORT 0x00000020 /* user wants "low" port binding */ #define INP_ANONPORT 0x00000040 /* port chosen for user */ #define INP_RECVIF 0x00000080 /* receive incoming interface */ #define INP_MTUDISC 0x00000100 /* user can do MTU discovery */ /* 0x000200 unused: was INP_FAITH */ #define INP_RECVTTL 0x00000400 /* receive incoming IP TTL */ #define INP_DONTFRAG 0x00000800 /* don't fragment packet */ #define INP_BINDANY 0x00001000 /* allow bind to any address */ #define INP_INHASHLIST 0x00002000 /* in_pcbinshash() has been called */ #define INP_RECVTOS 0x00004000 /* receive incoming IP TOS */ #define IN6P_IPV6_V6ONLY 0x00008000 /* restrict AF_INET6 socket for v6 */ #define IN6P_PKTINFO 0x00010000 /* receive IP6 dst and I/F */ #define IN6P_HOPLIMIT 0x00020000 /* receive hoplimit */ #define IN6P_HOPOPTS 0x00040000 /* receive hop-by-hop options */ #define IN6P_DSTOPTS 0x00080000 /* receive dst options after rthdr */ #define IN6P_RTHDR 0x00100000 /* receive routing header */ #define IN6P_RTHDRDSTOPTS 0x00200000 /* receive dstoptions before rthdr */ #define IN6P_TCLASS 0x00400000 /* receive traffic class value */ #define IN6P_AUTOFLOWLABEL 0x00800000 /* attach flowlabel automatically */ #define INP_TIMEWAIT 0x01000000 /* in TIMEWAIT, ppcb is tcptw */ #define INP_ONESBCAST 0x02000000 /* send all-ones broadcast */ #define INP_DROPPED 0x04000000 /* protocol drop flag */ #define INP_SOCKREF 0x08000000 /* strong socket reference */ #define INP_RESERVED_0 0x10000000 /* reserved field */ #define INP_RESERVED_1 0x20000000 /* reserved field */ #define IN6P_RFC2292 0x40000000 /* used RFC2292 API on the socket */ #define IN6P_MTU 0x80000000 /* receive path MTU */ #define INP_CONTROLOPTS (INP_RECVOPTS|INP_RECVRETOPTS|INP_RECVDSTADDR|\ INP_RECVIF|INP_RECVTTL|INP_RECVTOS|\ IN6P_PKTINFO|IN6P_HOPLIMIT|IN6P_HOPOPTS|\ IN6P_DSTOPTS|IN6P_RTHDR|IN6P_RTHDRDSTOPTS|\ IN6P_TCLASS|IN6P_AUTOFLOWLABEL|IN6P_RFC2292|\ IN6P_MTU) /* * Flags for inp_flags2. */ #define INP_MBUF_L_ACKS 0x00000001 /* We need large mbufs for ack compression */ #define INP_MBUF_ACKCMP 0x00000002 /* TCP mbuf ack compression ok */ #define INP_PCBGROUPWILD 0x00000004 /* in pcbgroup wildcard list */ #define INP_REUSEPORT 0x00000008 /* SO_REUSEPORT option is set */ #define INP_FREED 0x00000010 /* inp itself is not valid */ #define INP_REUSEADDR 0x00000020 /* SO_REUSEADDR option is set */ #define INP_BINDMULTI 0x00000040 /* IP_BINDMULTI option is set */ #define INP_RSS_BUCKET_SET 0x00000080 /* IP_RSS_LISTEN_BUCKET is set */ #define INP_RECVFLOWID 0x00000100 /* populate recv datagram with flow info */ #define INP_RECVRSSBUCKETID 0x00000200 /* populate recv datagram with bucket id */ #define INP_RATE_LIMIT_CHANGED 0x00000400 /* rate limit needs attention */ #define INP_ORIGDSTADDR 0x00000800 /* receive IP dst address/port */ #define INP_CANNOT_DO_ECN 0x00001000 /* The stack does not do ECN */ #define INP_REUSEPORT_LB 0x00002000 /* SO_REUSEPORT_LB option is set */ #define INP_SUPPORTS_MBUFQ 0x00004000 /* Supports the mbuf queue method of LRO */ #define INP_MBUF_QUEUE_READY 0x00008000 /* The transport is pacing, inputs can be queued */ #define INP_DONT_SACK_QUEUE 0x00010000 /* If a sack arrives do not wake me */ #define INP_2PCP_SET 0x00020000 /* If the Eth PCP should be set explicitly */ #define INP_2PCP_BIT0 0x00040000 /* Eth PCP Bit 0 */ #define INP_2PCP_BIT1 0x00080000 /* Eth PCP Bit 1 */ #define INP_2PCP_BIT2 0x00100000 /* Eth PCP Bit 2 */ #define INP_2PCP_BASE INP_2PCP_BIT0 #define INP_2PCP_MASK (INP_2PCP_BIT0 | INP_2PCP_BIT1 | INP_2PCP_BIT2) #define INP_2PCP_SHIFT 18 /* shift PCP field in/out of inp_flags2 */ /* * Flags passed to in_pcblookup*() functions. */ #define INPLOOKUP_WILDCARD 0x00000001 /* Allow wildcard sockets. */ #define INPLOOKUP_RLOCKPCB 0x00000002 /* Return inpcb read-locked. */ #define INPLOOKUP_WLOCKPCB 0x00000004 /* Return inpcb write-locked. */ #define INPLOOKUP_MASK (INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB | \ INPLOOKUP_WLOCKPCB) #define sotoinpcb(so) ((struct inpcb *)(so)->so_pcb) #define INP_SOCKAF(so) so->so_proto->pr_domain->dom_family #define INP_CHECK_SOCKAF(so, af) (INP_SOCKAF(so) == af) /* * Constants for pcbinfo.ipi_hashfields. */ #define IPI_HASHFIELDS_NONE 0 #define IPI_HASHFIELDS_2TUPLE 1 #define IPI_HASHFIELDS_4TUPLE 2 #ifdef _KERNEL VNET_DECLARE(int, ipport_reservedhigh); VNET_DECLARE(int, ipport_reservedlow); VNET_DECLARE(int, ipport_lowfirstauto); VNET_DECLARE(int, ipport_lowlastauto); VNET_DECLARE(int, ipport_firstauto); VNET_DECLARE(int, ipport_lastauto); VNET_DECLARE(int, ipport_hifirstauto); VNET_DECLARE(int, ipport_hilastauto); VNET_DECLARE(int, ipport_randomized); VNET_DECLARE(int, ipport_randomcps); VNET_DECLARE(int, ipport_randomtime); VNET_DECLARE(int, ipport_stoprandom); VNET_DECLARE(int, ipport_tcpallocs); #define V_ipport_reservedhigh VNET(ipport_reservedhigh) #define V_ipport_reservedlow VNET(ipport_reservedlow) #define V_ipport_lowfirstauto VNET(ipport_lowfirstauto) #define V_ipport_lowlastauto VNET(ipport_lowlastauto) #define V_ipport_firstauto VNET(ipport_firstauto) #define V_ipport_lastauto VNET(ipport_lastauto) #define V_ipport_hifirstauto VNET(ipport_hifirstauto) #define V_ipport_hilastauto VNET(ipport_hilastauto) #define V_ipport_randomized VNET(ipport_randomized) #define V_ipport_randomcps VNET(ipport_randomcps) #define V_ipport_randomtime VNET(ipport_randomtime) #define V_ipport_stoprandom VNET(ipport_stoprandom) #define V_ipport_tcpallocs VNET(ipport_tcpallocs) void in_pcbinfo_destroy(struct inpcbinfo *); void in_pcbinfo_init(struct inpcbinfo *, const char *, struct inpcbhead *, int, int, char *, uma_init, u_int); int in_pcbbind_check_bindmulti(const struct inpcb *ni, const struct inpcb *oi); struct inpcbgroup * in_pcbgroup_byhash(struct inpcbinfo *, u_int, uint32_t); struct inpcbgroup * in_pcbgroup_byinpcb(struct inpcb *); struct inpcbgroup * in_pcbgroup_bytuple(struct inpcbinfo *, struct in_addr, u_short, struct in_addr, u_short); void in_pcbgroup_destroy(struct inpcbinfo *); int in_pcbgroup_enabled(struct inpcbinfo *); void in_pcbgroup_init(struct inpcbinfo *, u_int, int); void in_pcbgroup_remove(struct inpcb *); void in_pcbgroup_update(struct inpcb *); void in_pcbgroup_update_mbuf(struct inpcb *, struct mbuf *); void in_pcbpurgeif0(struct inpcbinfo *, struct ifnet *); int in_pcballoc(struct socket *, struct inpcbinfo *); int in_pcbbind(struct inpcb *, struct sockaddr *, struct ucred *); int in_pcb_lport_dest(struct inpcb *inp, struct sockaddr *lsa, u_short *lportp, struct sockaddr *fsa, u_short fport, struct ucred *cred, int lookupflags); int in_pcb_lport(struct inpcb *, struct in_addr *, u_short *, struct ucred *, int); int in_pcbbind_setup(struct inpcb *, struct sockaddr *, in_addr_t *, u_short *, struct ucred *); int in_pcbconnect(struct inpcb *, struct sockaddr *, struct ucred *); int in_pcbconnect_mbuf(struct inpcb *, struct sockaddr *, struct ucred *, struct mbuf *, bool); int in_pcbconnect_setup(struct inpcb *, struct sockaddr *, in_addr_t *, u_short *, in_addr_t *, u_short *, struct inpcb **, struct ucred *); void in_pcbdetach(struct inpcb *); void in_pcbdisconnect(struct inpcb *); void in_pcbdrop(struct inpcb *); void in_pcbfree(struct inpcb *); int in_pcbinshash(struct inpcb *); int in_pcbinshash_mbuf(struct inpcb *, struct mbuf *); int in_pcbladdr(struct inpcb *, struct in_addr *, struct in_addr *, struct ucred *); int in_pcblbgroup_numa(struct inpcb *, int arg); struct inpcb * in_pcblookup_local(struct inpcbinfo *, struct in_addr, u_short, int, struct ucred *); struct inpcb * in_pcblookup(struct inpcbinfo *, struct in_addr, u_int, struct in_addr, u_int, int, struct ifnet *); struct inpcb * in_pcblookup_mbuf(struct inpcbinfo *, struct in_addr, u_int, struct in_addr, u_int, int, struct ifnet *, struct mbuf *); void in_pcbnotifyall(struct inpcbinfo *pcbinfo, struct in_addr, int, struct inpcb *(*)(struct inpcb *, int)); void in_pcbref(struct inpcb *); void in_pcbrehash(struct inpcb *); void in_pcbrehash_mbuf(struct inpcb *, struct mbuf *); int in_pcbrele(struct inpcb *); int in_pcbrele_rlocked(struct inpcb *); int in_pcbrele_wlocked(struct inpcb *); void in_pcblist_rele_rlocked(epoch_context_t ctx); void in_losing(struct inpcb *); void in_pcbsetsolabel(struct socket *so); int in_getpeeraddr(struct socket *so, struct sockaddr **nam); int in_getsockaddr(struct socket *so, struct sockaddr **nam); struct sockaddr * in_sockaddr(in_port_t port, struct in_addr *addr); void in_pcbsosetlabel(struct socket *so); #ifdef RATELIMIT int in_pcboutput_txrtlmt_locked(struct inpcb *, struct ifnet *, struct mbuf *, uint32_t); int in_pcbattach_txrtlmt(struct inpcb *, struct ifnet *, uint32_t, uint32_t, uint32_t, struct m_snd_tag **); void in_pcbdetach_txrtlmt(struct inpcb *); void in_pcbdetach_tag(struct m_snd_tag *); int in_pcbmodify_txrtlmt(struct inpcb *, uint32_t); int in_pcbquery_txrtlmt(struct inpcb *, uint32_t *); int in_pcbquery_txrlevel(struct inpcb *, uint32_t *); void in_pcboutput_txrtlmt(struct inpcb *, struct ifnet *, struct mbuf *); void in_pcboutput_eagain(struct inpcb *); #endif #endif /* _KERNEL */ #endif /* !_NETINET_IN_PCB_H_ */ diff --git a/sys/netinet/tcp_hpts.c b/sys/netinet/tcp_hpts.c index fd8b66b9ccdb..9a390e3a85c8 100644 --- a/sys/netinet/tcp_hpts.c +++ b/sys/netinet/tcp_hpts.c @@ -1,2016 +1,2299 @@ /*- * Copyright (c) 2016-2018 Netflix, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_rss.h" #include "opt_tcpdebug.h" /** * Some notes about usage. * * The tcp_hpts system is designed to provide a high precision timer * system for tcp. Its main purpose is to provide a mechanism for * pacing packets out onto the wire. It can be used in two ways * by a given TCP stack (and those two methods can be used simultaneously). * * First, and probably the main thing its used by Rack and BBR, it can * be used to call tcp_output() of a transport stack at some time in the future. * The normal way this is done is that tcp_output() of the stack schedules * itself to be called again by calling tcp_hpts_insert(tcpcb, slot). The * slot is the time from now that the stack wants to be called but it * must be converted to tcp_hpts's notion of slot. This is done with * one of the macros HPTS_MS_TO_SLOTS or HPTS_USEC_TO_SLOTS. So a typical * call from the tcp_output() routine might look like: * * tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(550)); * * The above would schedule tcp_ouput() to be called in 550 useconds. * Note that if using this mechanism the stack will want to add near * its top a check to prevent unwanted calls (from user land or the * arrival of incoming ack's). So it would add something like: * * if (inp->inp_in_hpts) * return; * * to prevent output processing until the time alotted has gone by. * Of course this is a bare bones example and the stack will probably * have more consideration then just the above. * * Now the second function (actually two functions I guess :D) * the tcp_hpts system provides is the ability to either abort * a connection (later) or process input on a connection. * Why would you want to do this? To keep processor locality * and or not have to worry about untangling any recursive * locks. The input function now is hooked to the new LRO * system as well. * * In order to use the input redirection function the * tcp stack must define an input function for * tfb_do_queued_segments(). This function understands * how to dequeue a array of packets that were input and * knows how to call the correct processing routine. * * Locking in this is important as well so most likely the * stack will need to define the tfb_do_segment_nounlock() * splitting tfb_do_segment() into two parts. The main processing * part that does not unlock the INP and returns a value of 1 or 0. * It returns 0 if all is well and the lock was not released. It * returns 1 if we had to destroy the TCB (a reset received etc). * The remains of tfb_do_segment() then become just a simple call * to the tfb_do_segment_nounlock() function and check the return * code and possibly unlock. * * The stack must also set the flag on the INP that it supports this * feature i.e. INP_SUPPORTS_MBUFQ. The LRO code recoginizes * this flag as well and will queue packets when it is set. * There are other flags as well INP_MBUF_QUEUE_READY and * INP_DONT_SACK_QUEUE. The first flag tells the LRO code * that we are in the pacer for output so there is no * need to wake up the hpts system to get immediate * input. The second tells the LRO code that its okay * if a SACK arrives you can still defer input and let * the current hpts timer run (this is usually set when * a rack timer is up so we know SACK's are happening * on the connection already and don't want to wakeup yet). * * There is a common functions within the rack_bbr_common code * version i.e. ctf_do_queued_segments(). This function * knows how to take the input queue of packets from * tp->t_in_pkts and process them digging out * all the arguments, calling any bpf tap and * calling into tfb_do_segment_nounlock(). The common * function (ctf_do_queued_segments()) requires that * you have defined the tfb_do_segment_nounlock() as * described above. * * The second feature of the input side of hpts is the * dropping of a connection. This is due to the way that * locking may have occured on the INP_WLOCK. So if * a stack wants to drop a connection it calls: * * tcp_set_inp_to_drop(tp, ETIMEDOUT) * * To schedule the tcp_hpts system to call * * tcp_drop(tp, drop_reason) * * at a future point. This is quite handy to prevent locking * issues when dropping connections. * */ #include #include #include #include #include #include #include #include #include /* for proc0 declaration */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef RSS #include #include #endif #define TCPSTATES /* for logging */ #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef tcpdebug #include #endif /* tcpdebug */ #ifdef tcp_offload #include #endif MALLOC_DEFINE(M_TCPHPTS, "tcp_hpts", "TCP hpts"); #ifdef RSS static int tcp_bind_threads = 1; #else static int tcp_bind_threads = 2; #endif -TUNABLE_INT("net.inet.tcp.bind_hptss", &tcp_bind_threads); - +static int tcp_use_irq_cpu = 0; static struct tcp_hptsi tcp_pace; +static uint32_t *cts_last_ran; static int hpts_does_tp_logging = 0; +static int hpts_use_assigned_cpu = 1; +static int32_t hpts_uses_oldest = OLDEST_THRESHOLD; -static void tcp_wakehpts(struct tcp_hpts_entry *p); -static void tcp_wakeinput(struct tcp_hpts_entry *p); static void tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv); -static void tcp_hptsi(struct tcp_hpts_entry *hpts); +static int32_t tcp_hptsi(struct tcp_hpts_entry *hpts, int from_callout); static void tcp_hpts_thread(void *ctx); static void tcp_init_hptsi(void *st); int32_t tcp_min_hptsi_time = DEFAULT_MIN_SLEEP; -static int32_t tcp_hpts_callout_skip_swi = 0; +static int conn_cnt_thresh = DEFAULT_CONNECTION_THESHOLD; +static int32_t dynamic_min_sleep = DYNAMIC_MIN_SLEEP; +static int32_t dynamic_max_sleep = DYNAMIC_MAX_SLEEP; + + SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hpts, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TCP Hpts controls"); +SYSCTL_NODE(_net_inet_tcp_hpts, OID_AUTO, stats, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, + "TCP Hpts statistics"); #define timersub(tvp, uvp, vvp) \ do { \ (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \ (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \ if ((vvp)->tv_usec < 0) { \ (vvp)->tv_sec--; \ (vvp)->tv_usec += 1000000; \ } \ } while (0) static int32_t tcp_hpts_precision = 120; struct hpts_domain_info { int count; int cpu[MAXCPU]; }; struct hpts_domain_info hpts_domains[MAXMEMDOM]; -SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, precision, CTLFLAG_RW, - &tcp_hpts_precision, 120, - "Value for PRE() precision of callout"); - counter_u64_t hpts_hopelessly_behind; -SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, hopeless, CTLFLAG_RD, +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, hopeless, CTLFLAG_RD, &hpts_hopelessly_behind, "Number of times hpts could not catch up and was behind hopelessly"); counter_u64_t hpts_loops; -SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, loops, CTLFLAG_RD, +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, loops, CTLFLAG_RD, &hpts_loops, "Number of times hpts had to loop to catch up"); counter_u64_t back_tosleep; -SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, no_tcbsfound, CTLFLAG_RD, +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, no_tcbsfound, CTLFLAG_RD, &back_tosleep, "Number of times hpts found no tcbs"); counter_u64_t combined_wheel_wrap; -SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, comb_wheel_wrap, CTLFLAG_RD, +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, comb_wheel_wrap, CTLFLAG_RD, &combined_wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap"); counter_u64_t wheel_wrap; -SYSCTL_COUNTER_U64(_net_inet_tcp_hpts, OID_AUTO, wheel_wrap, CTLFLAG_RD, +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, wheel_wrap, CTLFLAG_RD, &wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap"); -static int32_t out_ts_percision = 0; +counter_u64_t hpts_direct_call; +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, direct_call, CTLFLAG_RD, + &hpts_direct_call, "Number of times hpts was called by syscall/trap or other entry"); + +counter_u64_t hpts_wake_timeout; + +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, timeout_wakeup, CTLFLAG_RD, + &hpts_wake_timeout, "Number of times hpts threads woke up via the callout expiring"); + +counter_u64_t hpts_direct_awakening; + +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, direct_awakening, CTLFLAG_RD, + &hpts_direct_awakening, "Number of times hpts threads woke up via the callout expiring"); -SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, out_tspercision, CTLFLAG_RW, - &out_ts_percision, 0, - "Do we use a percise timestamp for every output cts"); +counter_u64_t hpts_back_tosleep; + +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, back_tosleep, CTLFLAG_RD, + &hpts_back_tosleep, "Number of times hpts threads woke up via the callout expiring and went back to sleep no work"); + +counter_u64_t cpu_uses_flowid; +counter_u64_t cpu_uses_random; + +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, cpusel_flowid, CTLFLAG_RD, + &cpu_uses_flowid, "Number of times when setting cpuid we used the flowid field"); +SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, cpusel_random, CTLFLAG_RD, + &cpu_uses_random, "Number of times when setting cpuid we used the a random value"); + +TUNABLE_INT("net.inet.tcp.bind_hptss", &tcp_bind_threads); +TUNABLE_INT("net.inet.tcp.use_irq", &tcp_use_irq_cpu); +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, bind_hptss, CTLFLAG_RD, + &tcp_bind_threads, 2, + "Thread Binding tunable"); +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, use_irq, CTLFLAG_RD, + &tcp_use_irq_cpu, 0, + "Use of irq CPU tunable"); +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, precision, CTLFLAG_RW, + &tcp_hpts_precision, 120, + "Value for PRE() precision of callout"); +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, cnt_thresh, CTLFLAG_RW, + &conn_cnt_thresh, 0, + "How many connections (below) make us use the callout based mechanism"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, logging, CTLFLAG_RW, &hpts_does_tp_logging, 0, "Do we add to any tp that has logging on pacer logs"); +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, use_assigned_cpu, CTLFLAG_RW, + &hpts_use_assigned_cpu, 0, + "Do we start any hpts timer on the assigned cpu?"); +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, use_oldest, CTLFLAG_RW, + &hpts_uses_oldest, OLDEST_THRESHOLD, + "Do syscalls look for the hpts that has been the longest since running (or just use cpu no if 0)?"); +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, dyn_minsleep, CTLFLAG_RW, + &dynamic_min_sleep, 250, + "What is the dynamic minsleep value?"); +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, dyn_maxsleep, CTLFLAG_RW, + &dynamic_max_sleep, 5000, + "What is the dynamic maxsleep value?"); + + + + static int32_t max_pacer_loops = 10; SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, loopmax, CTLFLAG_RW, &max_pacer_loops, 10, "What is the maximum number of times the pacer will loop trying to catch up"); #define HPTS_MAX_SLEEP_ALLOWED (NUM_OF_HPTSI_SLOTS/2) static uint32_t hpts_sleep_max = HPTS_MAX_SLEEP_ALLOWED; static int sysctl_net_inet_tcp_hpts_max_sleep(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = hpts_sleep_max; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { - if ((new < (NUM_OF_HPTSI_SLOTS / 4)) || + if ((new < dynamic_min_sleep) || (new > HPTS_MAX_SLEEP_ALLOWED)) error = EINVAL; else hpts_sleep_max = new; } return (error); } +static int +sysctl_net_inet_tcp_hpts_min_sleep(SYSCTL_HANDLER_ARGS) +{ + int error; + uint32_t new; + + new = tcp_min_hptsi_time; + error = sysctl_handle_int(oidp, &new, 0, req); + if (error == 0 && req->newptr) { + if (new < LOWEST_SLEEP_ALLOWED) + error = EINVAL; + else + tcp_min_hptsi_time = new; + } + return (error); +} + SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, maxsleep, CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &hpts_sleep_max, 0, &sysctl_net_inet_tcp_hpts_max_sleep, "IU", "Maximum time hpts will sleep"); -SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, minsleep, CTLFLAG_RW, +SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, minsleep, + CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_min_hptsi_time, 0, + &sysctl_net_inet_tcp_hpts_min_sleep, "IU", "The minimum time the hpts must sleep before processing more slots"); -SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, skip_swi, CTLFLAG_RW, - &tcp_hpts_callout_skip_swi, 0, - "Do we have the callout call directly to the hpts?"); +static int ticks_indicate_more_sleep = TICKS_INDICATE_MORE_SLEEP; +static int ticks_indicate_less_sleep = TICKS_INDICATE_LESS_SLEEP; +static int tcp_hpts_no_wake_over_thresh = 1; + +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, more_sleep, CTLFLAG_RW, + &ticks_indicate_more_sleep, 0, + "If we only process this many or less on a timeout, we need longer sleep on the next callout"); +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, less_sleep, CTLFLAG_RW, + &ticks_indicate_less_sleep, 0, + "If we process this many or more on a timeout, we need less sleep on the next callout"); +SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, nowake_over_thresh, CTLFLAG_RW, + &tcp_hpts_no_wake_over_thresh, 0, + "When we are over the threshold on the pacer do we prohibit wakeups?"); static void tcp_hpts_log(struct tcp_hpts_entry *hpts, struct tcpcb *tp, struct timeval *tv, - int ticks_to_run, int idx) + int slots_to_run, int idx, int from_callout) { union tcp_log_stackspecific log; - + /* + * Unused logs are + * 64 bit - delRate, rttProp, bw_inuse + * 16 bit - cwnd_gain + * 8 bit - bbr_state, bbr_substate, inhpts, ininput; + */ memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = hpts->p_nxt_slot; log.u_bbr.flex2 = hpts->p_cur_slot; log.u_bbr.flex3 = hpts->p_prev_slot; log.u_bbr.flex4 = idx; log.u_bbr.flex5 = hpts->p_curtick; log.u_bbr.flex6 = hpts->p_on_queue_cnt; - log.u_bbr.use_lt_bw = 1; - log.u_bbr.inflight = ticks_to_run; + log.u_bbr.flex7 = hpts->p_cpu; + log.u_bbr.flex8 = (uint8_t)from_callout; + log.u_bbr.inflight = slots_to_run; log.u_bbr.applimited = hpts->overidden_sleep; log.u_bbr.delivered = hpts->saved_curtick; log.u_bbr.timeStamp = tcp_tv_to_usectick(tv); log.u_bbr.epoch = hpts->saved_curslot; log.u_bbr.lt_epoch = hpts->saved_prev_slot; log.u_bbr.pkts_out = hpts->p_delayed_by; log.u_bbr.lost = hpts->p_hpts_sleep_time; - log.u_bbr.cur_del_rate = hpts->p_runningtick; + log.u_bbr.pacing_gain = hpts->p_cpu; + log.u_bbr.pkt_epoch = hpts->p_runningslot; + log.u_bbr.use_lt_bw = 1; TCP_LOG_EVENTP(tp, NULL, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, BBR_LOG_HPTSDIAG, 0, 0, &log, false, tv); } static void -hpts_timeout_swi(void *arg) +tcp_wakehpts(struct tcp_hpts_entry *hpts) { - struct tcp_hpts_entry *hpts; + HPTS_MTX_ASSERT(hpts); - hpts = (struct tcp_hpts_entry *)arg; - swi_sched(hpts->ie_cookie, 0); + if (tcp_hpts_no_wake_over_thresh && (hpts->p_on_queue_cnt >= conn_cnt_thresh)) { + hpts->p_direct_wake = 0; + return; + } + if (hpts->p_hpts_wake_scheduled == 0) { + hpts->p_hpts_wake_scheduled = 1; + swi_sched(hpts->ie_cookie, 0); + } } static void -hpts_timeout_dir(void *arg) +hpts_timeout_swi(void *arg) { - tcp_hpts_thread(arg); + struct tcp_hpts_entry *hpts; + + hpts = (struct tcp_hpts_entry *)arg; + swi_sched(hpts->ie_cookie, 0); } static inline void hpts_sane_pace_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int clear) { -#ifdef INVARIANTS - if (mtx_owned(&hpts->p_mtx) == 0) { - /* We don't own the mutex? */ - panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); - } - if (hpts->p_cpu != inp->inp_hpts_cpu) { - /* It is not the right cpu/mutex? */ - panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); - } - if (inp->inp_in_hpts == 0) { - /* We are not on the hpts? */ - panic("%s: hpts:%p inp:%p not on the hpts?", __FUNCTION__, hpts, inp); - } -#endif + HPTS_MTX_ASSERT(hpts); + KASSERT(hpts->p_cpu == inp->inp_hpts_cpu, ("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp)); + KASSERT(inp->inp_in_hpts != 0, ("%s: hpts:%p inp:%p not on the hpts?", __FUNCTION__, hpts, inp)); TAILQ_REMOVE(head, inp, inp_hpts); hpts->p_on_queue_cnt--; - if (hpts->p_on_queue_cnt < 0) { - /* Count should not go negative .. */ -#ifdef INVARIANTS - panic("Hpts goes negative inp:%p hpts:%p", - inp, hpts); -#endif - hpts->p_on_queue_cnt = 0; - } + KASSERT(hpts->p_on_queue_cnt >= 0, + ("Hpts goes negative inp:%p hpts:%p", + inp, hpts)); if (clear) { inp->inp_hpts_request = 0; inp->inp_in_hpts = 0; } } static inline void hpts_sane_pace_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, struct hptsh *head, int line, int noref) { -#ifdef INVARIANTS - if (mtx_owned(&hpts->p_mtx) == 0) { - /* We don't own the mutex? */ - panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); - } - if (hpts->p_cpu != inp->inp_hpts_cpu) { - /* It is not the right cpu/mutex? */ - panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); - } - if ((noref == 0) && (inp->inp_in_hpts == 1)) { - /* We are already on the hpts? */ - panic("%s: hpts:%p inp:%p already on the hpts?", __FUNCTION__, hpts, inp); - } -#endif + HPTS_MTX_ASSERT(hpts); + KASSERT(hpts->p_cpu == inp->inp_hpts_cpu, + ("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp)); + KASSERT(((noref == 1) && (inp->inp_in_hpts == 1)) || + ((noref == 0) && (inp->inp_in_hpts == 0)), + ("%s: hpts:%p inp:%p already on the hpts?", + __FUNCTION__, hpts, inp)); TAILQ_INSERT_TAIL(head, inp, inp_hpts); inp->inp_in_hpts = 1; hpts->p_on_queue_cnt++; if (noref == 0) { in_pcbref(inp); } } static inline void hpts_sane_input_remove(struct tcp_hpts_entry *hpts, struct inpcb *inp, int clear) { -#ifdef INVARIANTS - if (mtx_owned(&hpts->p_mtx) == 0) { - /* We don't own the mutex? */ - panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); - } - if (hpts->p_cpu != inp->inp_input_cpu) { - /* It is not the right cpu/mutex? */ - panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); - } - if (inp->inp_in_input == 0) { - /* We are not on the input hpts? */ - panic("%s: hpts:%p inp:%p not on the input hpts?", __FUNCTION__, hpts, inp); - } -#endif + HPTS_MTX_ASSERT(hpts); + KASSERT(hpts->p_cpu == inp->inp_hpts_cpu, + ("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp)); + KASSERT(inp->inp_in_input != 0, + ("%s: hpts:%p inp:%p not on the input hpts?", __FUNCTION__, hpts, inp)); TAILQ_REMOVE(&hpts->p_input, inp, inp_input); hpts->p_on_inqueue_cnt--; - if (hpts->p_on_inqueue_cnt < 0) { -#ifdef INVARIANTS - panic("Hpts in goes negative inp:%p hpts:%p", - inp, hpts); -#endif - hpts->p_on_inqueue_cnt = 0; - } -#ifdef INVARIANTS - if (TAILQ_EMPTY(&hpts->p_input) && - (hpts->p_on_inqueue_cnt != 0)) { - /* We should not be empty with a queue count */ - panic("%s hpts:%p in_hpts input empty but cnt:%d", - __FUNCTION__, hpts, hpts->p_on_inqueue_cnt); - } -#endif + KASSERT(hpts->p_on_inqueue_cnt >= 0, + ("Hpts in goes negative inp:%p hpts:%p", + inp, hpts)); + KASSERT((((TAILQ_EMPTY(&hpts->p_input) != 0) && (hpts->p_on_inqueue_cnt == 0)) || + ((TAILQ_EMPTY(&hpts->p_input) == 0) && (hpts->p_on_inqueue_cnt > 0))), + ("%s hpts:%p input cnt (p_on_inqueue):%d and queue state mismatch", + __FUNCTION__, hpts, hpts->p_on_inqueue_cnt)); if (clear) inp->inp_in_input = 0; } static inline void hpts_sane_input_insert(struct tcp_hpts_entry *hpts, struct inpcb *inp, int line) { -#ifdef INVARIANTS - if (mtx_owned(&hpts->p_mtx) == 0) { - /* We don't own the mutex? */ - panic("%s: hpts:%p inp:%p no hpts mutex", __FUNCTION__, hpts, inp); - } - if (hpts->p_cpu != inp->inp_input_cpu) { - /* It is not the right cpu/mutex? */ - panic("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp); - } - if (inp->inp_in_input == 1) { - /* We are already on the input hpts? */ - panic("%s: hpts:%p inp:%p already on the input hpts?", __FUNCTION__, hpts, inp); - } -#endif + HPTS_MTX_ASSERT(hpts); + KASSERT(hpts->p_cpu == inp->inp_hpts_cpu, + ("%s: hpts:%p inp:%p incorrect CPU", __FUNCTION__, hpts, inp)); + KASSERT(inp->inp_in_input == 0, + ("%s: hpts:%p inp:%p already on the input hpts?", __FUNCTION__, hpts, inp)); TAILQ_INSERT_TAIL(&hpts->p_input, inp, inp_input); inp->inp_in_input = 1; hpts->p_on_inqueue_cnt++; in_pcbref(inp); } -static void -tcp_wakehpts(struct tcp_hpts_entry *hpts) -{ - HPTS_MTX_ASSERT(hpts); - if (hpts->p_hpts_wake_scheduled == 0) { - hpts->p_hpts_wake_scheduled = 1; - swi_sched(hpts->ie_cookie, 0); - } -} - -static void -tcp_wakeinput(struct tcp_hpts_entry *hpts) -{ - HPTS_MTX_ASSERT(hpts); - if (hpts->p_hpts_wake_scheduled == 0) { - hpts->p_hpts_wake_scheduled = 1; - swi_sched(hpts->ie_cookie, 0); - } -} - struct tcp_hpts_entry * tcp_cur_hpts(struct inpcb *inp) { int32_t hpts_num; struct tcp_hpts_entry *hpts; hpts_num = inp->inp_hpts_cpu; hpts = tcp_pace.rp_ent[hpts_num]; return (hpts); } struct tcp_hpts_entry * tcp_hpts_lock(struct inpcb *inp) { struct tcp_hpts_entry *hpts; int32_t hpts_num; again: hpts_num = inp->inp_hpts_cpu; hpts = tcp_pace.rp_ent[hpts_num]; -#ifdef INVARIANTS - if (mtx_owned(&hpts->p_mtx)) { - panic("Hpts:%p owns mtx prior-to lock line:%d", - hpts, __LINE__); - } -#endif + KASSERT(mtx_owned(&hpts->p_mtx) == 0, + ("Hpts:%p owns mtx prior-to lock line:%d", + hpts, __LINE__)); mtx_lock(&hpts->p_mtx); if (hpts_num != inp->inp_hpts_cpu) { mtx_unlock(&hpts->p_mtx); goto again; } return (hpts); } struct tcp_hpts_entry * tcp_input_lock(struct inpcb *inp) { struct tcp_hpts_entry *hpts; int32_t hpts_num; again: hpts_num = inp->inp_input_cpu; hpts = tcp_pace.rp_ent[hpts_num]; -#ifdef INVARIANTS - if (mtx_owned(&hpts->p_mtx)) { - panic("Hpts:%p owns mtx prior-to lock line:%d", - hpts, __LINE__); - } -#endif + KASSERT(mtx_owned(&hpts->p_mtx) == 0, + ("Hpts:%p owns mtx prior-to lock line:%d", + hpts, __LINE__)); mtx_lock(&hpts->p_mtx); if (hpts_num != inp->inp_input_cpu) { mtx_unlock(&hpts->p_mtx); goto again; } return (hpts); } static void tcp_remove_hpts_ref(struct inpcb *inp, struct tcp_hpts_entry *hpts, int line) { int32_t add_freed; + int32_t ret; if (inp->inp_flags2 & INP_FREED) { /* * Need to play a special trick so that in_pcbrele_wlocked * does not return 1 when it really should have returned 0. */ add_freed = 1; inp->inp_flags2 &= ~INP_FREED; } else { add_freed = 0; } #ifndef INP_REF_DEBUG - if (in_pcbrele_wlocked(inp)) { - /* - * This should not happen. We have the inpcb referred to by - * the main socket (why we are called) and the hpts. It - * should always return 0. - */ - panic("inpcb:%p release ret 1", - inp); - } + ret = in_pcbrele_wlocked(inp); #else - if (__in_pcbrele_wlocked(inp, line)) { - /* - * This should not happen. We have the inpcb referred to by - * the main socket (why we are called) and the hpts. It - * should always return 0. - */ - panic("inpcb:%p release ret 1", - inp); - } + ret = __in_pcbrele_wlocked(inp, line); #endif + KASSERT(ret != 1, ("inpcb:%p release ret 1", inp)); if (add_freed) { inp->inp_flags2 |= INP_FREED; } } static void tcp_hpts_remove_locked_output(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line) { if (inp->inp_in_hpts) { hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], 1); tcp_remove_hpts_ref(inp, hpts, line); } } static void tcp_hpts_remove_locked_input(struct tcp_hpts_entry *hpts, struct inpcb *inp, int32_t flags, int32_t line) { HPTS_MTX_ASSERT(hpts); if (inp->inp_in_input) { hpts_sane_input_remove(hpts, inp, 1); tcp_remove_hpts_ref(inp, hpts, line); } } /* * Called normally with the INP_LOCKED but it * does not matter, the hpts lock is the key * but the lock order allows us to hold the * INP lock and then get the hpts lock. * * Valid values in the flags are * HPTS_REMOVE_OUTPUT - remove from the output of the hpts. * HPTS_REMOVE_INPUT - remove from the input of the hpts. * Note that you can use one or both values together * and get two actions. */ void __tcp_hpts_remove(struct inpcb *inp, int32_t flags, int32_t line) { struct tcp_hpts_entry *hpts; INP_WLOCK_ASSERT(inp); if (flags & HPTS_REMOVE_OUTPUT) { hpts = tcp_hpts_lock(inp); tcp_hpts_remove_locked_output(hpts, inp, flags, line); mtx_unlock(&hpts->p_mtx); } if (flags & HPTS_REMOVE_INPUT) { hpts = tcp_input_lock(inp); tcp_hpts_remove_locked_input(hpts, inp, flags, line); mtx_unlock(&hpts->p_mtx); } } static inline int -hpts_tick(uint32_t wheel_tick, uint32_t plus) +hpts_slot(uint32_t wheel_slot, uint32_t plus) { /* * Given a slot on the wheel, what slot * is that plus ticks out? */ - KASSERT(wheel_tick < NUM_OF_HPTSI_SLOTS, ("Invalid tick %u not on wheel", wheel_tick)); - return ((wheel_tick + plus) % NUM_OF_HPTSI_SLOTS); + KASSERT(wheel_slot < NUM_OF_HPTSI_SLOTS, ("Invalid tick %u not on wheel", wheel_slot)); + return ((wheel_slot + plus) % NUM_OF_HPTSI_SLOTS); } static inline int tick_to_wheel(uint32_t cts_in_wticks) { /* - * Given a timestamp in wheel ticks (10usec inc's) - * map it to our limited space wheel. + * Given a timestamp in ticks (so by + * default to get it to a real time one + * would multiply by 10.. i.e the number + * of ticks in a slot) map it to our limited + * space wheel. */ return (cts_in_wticks % NUM_OF_HPTSI_SLOTS); } static inline int -hpts_ticks_diff(int prev_tick, int tick_now) +hpts_slots_diff(int prev_slot, int slot_now) { /* - * Given two ticks that are someplace + * Given two slots that are someplace * on our wheel. How far are they apart? */ - if (tick_now > prev_tick) - return (tick_now - prev_tick); - else if (tick_now == prev_tick) + if (slot_now > prev_slot) + return (slot_now - prev_slot); + else if (slot_now == prev_slot) /* * Special case, same means we can go all of our * wheel less one slot. */ return (NUM_OF_HPTSI_SLOTS - 1); else - return ((NUM_OF_HPTSI_SLOTS - prev_tick) + tick_now); + return ((NUM_OF_HPTSI_SLOTS - prev_slot) + slot_now); } /* - * Given a tick on the wheel that is the current time - * mapped to the wheel (wheel_tick), what is the maximum + * Given a slot on the wheel that is the current time + * mapped to the wheel (wheel_slot), what is the maximum * distance forward that can be obtained without - * wrapping past either prev_tick or running_tick + * wrapping past either prev_slot or running_slot * depending on the htps state? Also if passed - * a uint32_t *, fill it with the tick location. + * a uint32_t *, fill it with the slot location. * * Note if you do not give this function the current - * time (that you think it is) mapped to the wheel + * time (that you think it is) mapped to the wheel slot * then the results will not be what you expect and * could lead to invalid inserts. */ static inline int32_t -max_ticks_available(struct tcp_hpts_entry *hpts, uint32_t wheel_tick, uint32_t *target_tick) +max_slots_available(struct tcp_hpts_entry *hpts, uint32_t wheel_slot, uint32_t *target_slot) { - uint32_t dis_to_travel, end_tick, pacer_to_now, avail_on_wheel; + uint32_t dis_to_travel, end_slot, pacer_to_now, avail_on_wheel; if ((hpts->p_hpts_active == 1) && (hpts->p_wheel_complete == 0)) { - end_tick = hpts->p_runningtick; + end_slot = hpts->p_runningslot; /* Back up one tick */ - if (end_tick == 0) - end_tick = NUM_OF_HPTSI_SLOTS - 1; + if (end_slot == 0) + end_slot = NUM_OF_HPTSI_SLOTS - 1; else - end_tick--; - if (target_tick) - *target_tick = end_tick; + end_slot--; + if (target_slot) + *target_slot = end_slot; } else { /* * For the case where we are * not active, or we have * completed the pass over * the wheel, we can use the * prev tick and subtract one from it. This puts us * as far out as possible on the wheel. */ - end_tick = hpts->p_prev_slot; - if (end_tick == 0) - end_tick = NUM_OF_HPTSI_SLOTS - 1; + end_slot = hpts->p_prev_slot; + if (end_slot == 0) + end_slot = NUM_OF_HPTSI_SLOTS - 1; else - end_tick--; - if (target_tick) - *target_tick = end_tick; + end_slot--; + if (target_slot) + *target_slot = end_slot; /* * Now we have close to the full wheel left minus the * time it has been since the pacer went to sleep. Note * that wheel_tick, passed in, should be the current time * from the perspective of the caller, mapped to the wheel. */ - if (hpts->p_prev_slot != wheel_tick) - dis_to_travel = hpts_ticks_diff(hpts->p_prev_slot, wheel_tick); + if (hpts->p_prev_slot != wheel_slot) + dis_to_travel = hpts_slots_diff(hpts->p_prev_slot, wheel_slot); else dis_to_travel = 1; /* * dis_to_travel in this case is the space from when the - * pacer stopped (p_prev_slot) and where our wheel_tick + * pacer stopped (p_prev_slot) and where our wheel_slot * is now. To know how many slots we can put it in we * subtract from the wheel size. We would not want * to place something after p_prev_slot or it will * get ran too soon. */ return (NUM_OF_HPTSI_SLOTS - dis_to_travel); } /* - * So how many slots are open between p_runningtick -> p_cur_slot + * So how many slots are open between p_runningslot -> p_cur_slot * that is what is currently un-available for insertion. Special * case when we are at the last slot, this gets 1, so that * the answer to how many slots are available is all but 1. */ - if (hpts->p_runningtick == hpts->p_cur_slot) + if (hpts->p_runningslot == hpts->p_cur_slot) dis_to_travel = 1; else - dis_to_travel = hpts_ticks_diff(hpts->p_runningtick, hpts->p_cur_slot); + dis_to_travel = hpts_slots_diff(hpts->p_runningslot, hpts->p_cur_slot); /* * How long has the pacer been running? */ - if (hpts->p_cur_slot != wheel_tick) { + if (hpts->p_cur_slot != wheel_slot) { /* The pacer is a bit late */ - pacer_to_now = hpts_ticks_diff(hpts->p_cur_slot, wheel_tick); + pacer_to_now = hpts_slots_diff(hpts->p_cur_slot, wheel_slot); } else { /* The pacer is right on time, now == pacers start time */ pacer_to_now = 0; } /* * To get the number left we can insert into we simply * subract the distance the pacer has to run from how * many slots there are. */ avail_on_wheel = NUM_OF_HPTSI_SLOTS - dis_to_travel; /* * Now how many of those we will eat due to the pacer's * time (p_cur_slot) of start being behind the - * real time (wheel_tick)? + * real time (wheel_slot)? */ if (avail_on_wheel <= pacer_to_now) { /* * Wheel wrap, we can't fit on the wheel, that * is unusual the system must be way overloaded! - * Insert into the assured tick, and return special + * Insert into the assured slot, and return special * "0". */ counter_u64_add(combined_wheel_wrap, 1); - *target_tick = hpts->p_nxt_slot; + *target_slot = hpts->p_nxt_slot; return (0); } else { /* * We know how many slots are open * on the wheel (the reverse of what * is left to run. Take away the time - * the pacer started to now (wheel_tick) + * the pacer started to now (wheel_slot) * and that tells you how many slots are * open that can be inserted into that won't * be touched by the pacer until later. */ return (avail_on_wheel - pacer_to_now); } } static int tcp_queue_to_hpts_immediate_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line, int32_t noref) { uint32_t need_wake = 0; HPTS_MTX_ASSERT(hpts); if (inp->inp_in_hpts == 0) { /* Ok we need to set it on the hpts in the current slot */ inp->inp_hpts_request = 0; if ((hpts->p_hpts_active == 0) || (hpts->p_wheel_complete)) { /* * A sleeping hpts we want in next slot to run * note that in this state p_prev_slot == p_cur_slot */ - inp->inp_hptsslot = hpts_tick(hpts->p_prev_slot, 1); + inp->inp_hptsslot = hpts_slot(hpts->p_prev_slot, 1); if ((hpts->p_on_min_sleep == 0) && (hpts->p_hpts_active == 0)) need_wake = 1; } else if ((void *)inp == hpts->p_inp) { /* * The hpts system is running and the caller * was awoken by the hpts system. * We can't allow you to go into the same slot we * are in (we don't want a loop :-D). */ inp->inp_hptsslot = hpts->p_nxt_slot; } else - inp->inp_hptsslot = hpts->p_runningtick; + inp->inp_hptsslot = hpts->p_runningslot; hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, noref); if (need_wake) { /* * Activate the hpts if it is sleeping and its * timeout is not 1. */ hpts->p_direct_wake = 1; tcp_wakehpts(hpts); } } return (need_wake); } int __tcp_queue_to_hpts_immediate(struct inpcb *inp, int32_t line) { int32_t ret; struct tcp_hpts_entry *hpts; INP_WLOCK_ASSERT(inp); hpts = tcp_hpts_lock(inp); ret = tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0); mtx_unlock(&hpts->p_mtx); return (ret); } #ifdef INVARIANTS static void check_if_slot_would_be_wrong(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t inp_hptsslot, int line) { /* * Sanity checks for the pacer with invariants * on insert. */ - if (inp_hptsslot >= NUM_OF_HPTSI_SLOTS) - panic("hpts:%p inp:%p slot:%d > max", - hpts, inp, inp_hptsslot); + KASSERT(inp_hptsslot < NUM_OF_HPTSI_SLOTS, + ("hpts:%p inp:%p slot:%d > max", + hpts, inp, inp_hptsslot)); if ((hpts->p_hpts_active) && (hpts->p_wheel_complete == 0)) { /* * If the pacer is processing a arc * of the wheel, we need to make * sure we are not inserting within * that arc. */ int distance, yet_to_run; - distance = hpts_ticks_diff(hpts->p_runningtick, inp_hptsslot); - if (hpts->p_runningtick != hpts->p_cur_slot) - yet_to_run = hpts_ticks_diff(hpts->p_runningtick, hpts->p_cur_slot); + distance = hpts_slots_diff(hpts->p_runningslot, inp_hptsslot); + if (hpts->p_runningslot != hpts->p_cur_slot) + yet_to_run = hpts_slots_diff(hpts->p_runningslot, hpts->p_cur_slot); else yet_to_run = 0; /* processing last slot */ - if (yet_to_run > distance) { - panic("hpts:%p inp:%p slot:%d distance:%d yet_to_run:%d rs:%d cs:%d", - hpts, inp, inp_hptsslot, - distance, yet_to_run, - hpts->p_runningtick, hpts->p_cur_slot); - } + KASSERT(yet_to_run <= distance, + ("hpts:%p inp:%p slot:%d distance:%d yet_to_run:%d rs:%d cs:%d", + hpts, inp, inp_hptsslot, + distance, yet_to_run, + hpts->p_runningslot, hpts->p_cur_slot)); } } #endif static void tcp_hpts_insert_locked(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag, struct timeval *tv) { uint32_t need_new_to = 0; - uint32_t wheel_cts, last_tick; - int32_t wheel_tick, maxticks; + uint32_t wheel_cts; + int32_t wheel_slot, maxslots, last_slot; + int cpu; int8_t need_wakeup = 0; HPTS_MTX_ASSERT(hpts); if (diag) { memset(diag, 0, sizeof(struct hpts_diag)); diag->p_hpts_active = hpts->p_hpts_active; diag->p_prev_slot = hpts->p_prev_slot; - diag->p_runningtick = hpts->p_runningtick; + diag->p_runningslot = hpts->p_runningslot; diag->p_nxt_slot = hpts->p_nxt_slot; diag->p_cur_slot = hpts->p_cur_slot; diag->p_curtick = hpts->p_curtick; diag->p_lasttick = hpts->p_lasttick; diag->slot_req = slot; diag->p_on_min_sleep = hpts->p_on_min_sleep; diag->hpts_sleep_time = hpts->p_hpts_sleep_time; } - if (inp->inp_in_hpts == 0) { - if (slot == 0) { - /* Immediate */ - tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0); - return; - } - /* Get the current time relative to the wheel */ - wheel_cts = tcp_tv_to_hptstick(tv); - /* Map it onto the wheel */ - wheel_tick = tick_to_wheel(wheel_cts); - /* Now what's the max we can place it at? */ - maxticks = max_ticks_available(hpts, wheel_tick, &last_tick); - if (diag) { - diag->wheel_tick = wheel_tick; - diag->maxticks = maxticks; - diag->wheel_cts = wheel_cts; + KASSERT(inp->inp_in_hpts == 0, ("Hpts:%p tp:%p already on hpts and add?", hpts, inp)); + if (slot == 0) { + /* Immediate */ + tcp_queue_to_hpts_immediate_locked(inp, hpts, line, 0); + return; + } + /* Get the current time relative to the wheel */ + wheel_cts = tcp_tv_to_hptstick(tv); + /* Map it onto the wheel */ + wheel_slot = tick_to_wheel(wheel_cts); + /* Now what's the max we can place it at? */ + maxslots = max_slots_available(hpts, wheel_slot, &last_slot); + if (diag) { + diag->wheel_slot = wheel_slot; + diag->maxslots = maxslots; + diag->wheel_cts = wheel_cts; + } + if (maxslots == 0) { + /* The pacer is in a wheel wrap behind, yikes! */ + if (slot > 1) { + /* + * Reduce by 1 to prevent a forever loop in + * case something else is wrong. Note this + * probably does not hurt because the pacer + * if its true is so far behind we will be + * > 1second late calling anyway. + */ + slot--; } - if (maxticks == 0) { - /* The pacer is in a wheel wrap behind, yikes! */ - if (slot > 1) { - /* - * Reduce by 1 to prevent a forever loop in - * case something else is wrong. Note this - * probably does not hurt because the pacer - * if its true is so far behind we will be - * > 1second late calling anyway. - */ - slot--; - } - inp->inp_hptsslot = last_tick; - inp->inp_hpts_request = slot; - } else if (maxticks >= slot) { - /* It all fits on the wheel */ - inp->inp_hpts_request = 0; - inp->inp_hptsslot = hpts_tick(wheel_tick, slot); - } else { - /* It does not fit */ - inp->inp_hpts_request = slot - maxticks; - inp->inp_hptsslot = last_tick; + inp->inp_hptsslot = last_slot; + inp->inp_hpts_request = slot; + } else if (maxslots >= slot) { + /* It all fits on the wheel */ + inp->inp_hpts_request = 0; + inp->inp_hptsslot = hpts_slot(wheel_slot, slot); + } else { + /* It does not fit */ + inp->inp_hpts_request = slot - maxslots; + inp->inp_hptsslot = last_slot; + } + if (diag) { + diag->slot_remaining = inp->inp_hpts_request; + diag->inp_hptsslot = inp->inp_hptsslot; + } +#ifdef INVARIANTS + check_if_slot_would_be_wrong(hpts, inp, inp->inp_hptsslot, line); +#endif + hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, 0); + if ((hpts->p_hpts_active == 0) && + (inp->inp_hpts_request == 0) && + (hpts->p_on_min_sleep == 0)) { + /* + * The hpts is sleeping and NOT on a minimum + * sleep time, we need to figure out where + * it will wake up at and if we need to reschedule + * its time-out. + */ + uint32_t have_slept, yet_to_sleep; + + /* Now do we need to restart the hpts's timer? */ + have_slept = hpts_slots_diff(hpts->p_prev_slot, wheel_slot); + if (have_slept < hpts->p_hpts_sleep_time) + yet_to_sleep = hpts->p_hpts_sleep_time - have_slept; + else { + /* We are over-due */ + yet_to_sleep = 0; + need_wakeup = 1; } if (diag) { - diag->slot_remaining = inp->inp_hpts_request; - diag->inp_hptsslot = inp->inp_hptsslot; + diag->have_slept = have_slept; + diag->yet_to_sleep = yet_to_sleep; } -#ifdef INVARIANTS - check_if_slot_would_be_wrong(hpts, inp, inp->inp_hptsslot, line); -#endif - hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], line, 0); - if ((hpts->p_hpts_active == 0) && - (inp->inp_hpts_request == 0) && - (hpts->p_on_min_sleep == 0)) { + if (yet_to_sleep && + (yet_to_sleep > slot)) { /* - * The hpts is sleeping and not on a minimum - * sleep time, we need to figure out where - * it will wake up at and if we need to reschedule - * its time-out. + * We need to reschedule the hpts's time-out. */ - uint32_t have_slept, yet_to_sleep; - - /* Now do we need to restart the hpts's timer? */ - have_slept = hpts_ticks_diff(hpts->p_prev_slot, wheel_tick); - if (have_slept < hpts->p_hpts_sleep_time) - yet_to_sleep = hpts->p_hpts_sleep_time - have_slept; - else { - /* We are over-due */ - yet_to_sleep = 0; - need_wakeup = 1; - } - if (diag) { - diag->have_slept = have_slept; - diag->yet_to_sleep = yet_to_sleep; - } - if (yet_to_sleep && - (yet_to_sleep > slot)) { - /* - * We need to reschedule the hpts's time-out. - */ - hpts->p_hpts_sleep_time = slot; - need_new_to = slot * HPTS_TICKS_PER_USEC; - } + hpts->p_hpts_sleep_time = slot; + need_new_to = slot * HPTS_TICKS_PER_SLOT; } - /* - * Now how far is the hpts sleeping to? if active is 1, its - * up and ticking we do nothing, otherwise we may need to - * reschedule its callout if need_new_to is set from above. - */ - if (need_wakeup) { - hpts->p_direct_wake = 1; - tcp_wakehpts(hpts); - if (diag) { - diag->need_new_to = 0; - diag->co_ret = 0xffff0000; - } - } else if (need_new_to) { - int32_t co_ret; - struct timeval tv; - sbintime_t sb; + } + /* + * Now how far is the hpts sleeping to? if active is 1, its + * up and ticking we do nothing, otherwise we may need to + * reschedule its callout if need_new_to is set from above. + */ + if (need_wakeup) { + hpts->p_direct_wake = 1; + tcp_wakehpts(hpts); + if (diag) { + diag->need_new_to = 0; + diag->co_ret = 0xffff0000; + } + } else if (need_new_to) { + int32_t co_ret; + struct timeval tv; + sbintime_t sb; - tv.tv_sec = 0; - tv.tv_usec = 0; - while (need_new_to > HPTS_USEC_IN_SEC) { - tv.tv_sec++; - need_new_to -= HPTS_USEC_IN_SEC; - } - tv.tv_usec = need_new_to; - sb = tvtosbt(tv); - if (tcp_hpts_callout_skip_swi == 0) { - co_ret = callout_reset_sbt_on(&hpts->co, sb, 0, - hpts_timeout_swi, hpts, hpts->p_cpu, - (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); - } else { - co_ret = callout_reset_sbt_on(&hpts->co, sb, 0, - hpts_timeout_dir, hpts, - hpts->p_cpu, - C_PREL(tcp_hpts_precision)); - } - if (diag) { - diag->need_new_to = need_new_to; - diag->co_ret = co_ret; - } + tv.tv_sec = 0; + tv.tv_usec = 0; + while (need_new_to > HPTS_USEC_IN_SEC) { + tv.tv_sec++; + need_new_to -= HPTS_USEC_IN_SEC; + } + tv.tv_usec = need_new_to; + sb = tvtosbt(tv); + cpu = (tcp_bind_threads || hpts_use_assigned_cpu) ? hpts->p_cpu : curcpu; + co_ret = callout_reset_sbt_on(&hpts->co, sb, 0, + hpts_timeout_swi, hpts, cpu, + (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); + if (diag) { + diag->need_new_to = need_new_to; + diag->co_ret = co_ret; } - } else { -#ifdef INVARIANTS - panic("Hpts:%p tp:%p already on hpts and add?", hpts, inp); -#endif } } uint32_t tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag) { struct tcp_hpts_entry *hpts; uint32_t slot_on; struct timeval tv; /* * We now return the next-slot the hpts will be on, beyond its * current run (if up) or where it was when it stopped if it is * sleeping. */ INP_WLOCK_ASSERT(inp); hpts = tcp_hpts_lock(inp); microuptime(&tv); tcp_hpts_insert_locked(hpts, inp, slot, line, diag, &tv); slot_on = hpts->p_nxt_slot; mtx_unlock(&hpts->p_mtx); return (slot_on); } uint32_t __tcp_hpts_insert(struct inpcb *inp, uint32_t slot, int32_t line){ return (tcp_hpts_insert_diag(inp, slot, line, NULL)); } + int __tcp_queue_to_input_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line) { int32_t retval = 0; HPTS_MTX_ASSERT(hpts); if (inp->inp_in_input == 0) { /* Ok we need to set it on the hpts in the current slot */ hpts_sane_input_insert(hpts, inp, line); retval = 1; - if (hpts->p_hpts_active == 0) { + if ((hpts->p_hpts_active == 0) && + (hpts->p_on_min_sleep == 0)){ /* * Activate the hpts if it is sleeping. */ retval = 2; hpts->p_direct_wake = 1; - tcp_wakeinput(hpts); + tcp_wakehpts(hpts); } - } else if (hpts->p_hpts_active == 0) { + } else if ((hpts->p_hpts_active == 0) && + (hpts->p_on_min_sleep == 0)){ retval = 4; hpts->p_direct_wake = 1; - tcp_wakeinput(hpts); + tcp_wakehpts(hpts); } return (retval); } int32_t __tcp_queue_to_input(struct inpcb *inp, int line) { struct tcp_hpts_entry *hpts; int32_t ret; hpts = tcp_input_lock(inp); ret = __tcp_queue_to_input_locked(inp, hpts, line); mtx_unlock(&hpts->p_mtx); return (ret); } void __tcp_set_inp_to_drop(struct inpcb *inp, uint16_t reason, int32_t line) { struct tcp_hpts_entry *hpts; struct tcpcb *tp; tp = intotcpcb(inp); hpts = tcp_input_lock(tp->t_inpcb); if (inp->inp_in_input == 0) { /* Ok we need to set it on the hpts in the current slot */ hpts_sane_input_insert(hpts, inp, line); - if (hpts->p_hpts_active == 0) { + if ((hpts->p_hpts_active == 0) && + (hpts->p_on_min_sleep == 0)){ /* * Activate the hpts if it is sleeping. */ hpts->p_direct_wake = 1; - tcp_wakeinput(hpts); + tcp_wakehpts(hpts); } - } else if (hpts->p_hpts_active == 0) { + } else if ((hpts->p_hpts_active == 0) && + (hpts->p_on_min_sleep == 0)){ hpts->p_direct_wake = 1; - tcp_wakeinput(hpts); + tcp_wakehpts(hpts); } inp->inp_hpts_drop_reas = reason; mtx_unlock(&hpts->p_mtx); } -static uint16_t +uint16_t hpts_random_cpu(struct inpcb *inp){ /* * No flow type set distribute the load randomly. */ uint16_t cpuid; uint32_t ran; /* * If one has been set use it i.e. we want both in and out on the * same hpts. */ if (inp->inp_input_cpu_set) { return (inp->inp_input_cpu); } else if (inp->inp_hpts_cpu_set) { return (inp->inp_hpts_cpu); } /* Nothing set use a random number */ ran = arc4random(); - cpuid = (ran & 0xffff) % mp_ncpus; + cpuid = (((ran & 0xffff) % mp_ncpus) % tcp_pace.rp_num_hptss); return (cpuid); } static uint16_t -hpts_cpuid(struct inpcb *inp) +hpts_cpuid(struct inpcb *inp, int *failed) { u_int cpuid; #if !defined(RSS) && defined(NUMA) struct hpts_domain_info *di; #endif + *failed = 0; /* * If one has been set use it i.e. we want both in and out on the * same hpts. */ if (inp->inp_input_cpu_set) { return (inp->inp_input_cpu); } else if (inp->inp_hpts_cpu_set) { return (inp->inp_hpts_cpu); } + /* + * If we are using the irq cpu set by LRO or + * the driver then it overrides all other domains. + */ + if (tcp_use_irq_cpu) { + if (inp->inp_irq_cpu_set == 0) { + *failed = 1; + return(0); + } + return(inp->inp_irq_cpu); + } /* If one is set the other must be the same */ #ifdef RSS cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); if (cpuid == NETISR_CPUID_NONE) return (hpts_random_cpu(inp)); else return (cpuid); #else /* * We don't have a flowid -> cpuid mapping, so cheat and just map * unknown cpuids to curcpu. Not the best, but apparently better * than defaulting to swi 0. */ - - if (inp->inp_flowtype == M_HASHTYPE_NONE) + if (inp->inp_flowtype == M_HASHTYPE_NONE) { + counter_u64_add(cpu_uses_random, 1); return (hpts_random_cpu(inp)); + } /* * Hash to a thread based on the flowid. If we are using numa, * then restrict the hash to the numa domain where the inp lives. */ #ifdef NUMA if (tcp_bind_threads == 2 && inp->inp_numa_domain != M_NODOM) { di = &hpts_domains[inp->inp_numa_domain]; cpuid = di->cpu[inp->inp_flowid % di->count]; } else #endif cpuid = inp->inp_flowid % mp_ncpus; - + counter_u64_add(cpu_uses_flowid, 1); return (cpuid); #endif } static void tcp_drop_in_pkts(struct tcpcb *tp) { struct mbuf *m, *n; m = tp->t_in_pkt; if (m) n = m->m_nextpkt; else n = NULL; tp->t_in_pkt = NULL; while (m) { m_freem(m); m = n; if (m) n = m->m_nextpkt; } } /* * Do NOT try to optimize the processing of inp's * by first pulling off all the inp's into a temporary * list (e.g. TAILQ_CONCAT). If you do that the subtle * interactions of switching CPU's will kill because of * problems in the linked list manipulation. Basically * you would switch cpu's with the hpts mutex locked * but then while you were processing one of the inp's * some other one that you switch will get a new * packet on the different CPU. It will insert it * on the new hpts's input list. Creating a temporary * link in the inp will not fix it either, since * the other hpts will be doing the same thing and * you will both end up using the temporary link. * * You will die in an ASSERT for tailq corruption if you * run INVARIANTS or you will die horribly without * INVARIANTS in some unknown way with a corrupt linked * list. */ static void tcp_input_data(struct tcp_hpts_entry *hpts, struct timeval *tv) { struct tcpcb *tp; struct inpcb *inp; uint16_t drop_reason; int16_t set_cpu; uint32_t did_prefetch = 0; int dropped; HPTS_MTX_ASSERT(hpts); NET_EPOCH_ASSERT(); while ((inp = TAILQ_FIRST(&hpts->p_input)) != NULL) { HPTS_MTX_ASSERT(hpts); hpts_sane_input_remove(hpts, inp, 0); if (inp->inp_input_cpu_set == 0) { set_cpu = 1; } else { set_cpu = 0; } hpts->p_inp = inp; drop_reason = inp->inp_hpts_drop_reas; inp->inp_in_input = 0; mtx_unlock(&hpts->p_mtx); INP_WLOCK(inp); #ifdef VIMAGE CURVNET_SET(inp->inp_vnet); #endif if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) || (inp->inp_flags2 & INP_FREED)) { out: hpts->p_inp = NULL; if (in_pcbrele_wlocked(inp) == 0) { INP_WUNLOCK(inp); } #ifdef VIMAGE CURVNET_RESTORE(); #endif mtx_lock(&hpts->p_mtx); continue; } tp = intotcpcb(inp); if ((tp == NULL) || (tp->t_inpcb == NULL)) { goto out; } if (drop_reason) { /* This tcb is being destroyed for drop_reason */ tcp_drop_in_pkts(tp); tp = tcp_drop(tp, drop_reason); if (tp == NULL) { INP_WLOCK(inp); } if (in_pcbrele_wlocked(inp) == 0) INP_WUNLOCK(inp); #ifdef VIMAGE CURVNET_RESTORE(); #endif mtx_lock(&hpts->p_mtx); continue; } if (set_cpu) { /* * Setup so the next time we will move to the right * CPU. This should be a rare event. It will * sometimes happens when we are the client side * (usually not the server). Somehow tcp_output() * gets called before the tcp_do_segment() sets the * intial state. This means the r_cpu and r_hpts_cpu * is 0. We get on the hpts, and then tcp_input() * gets called setting up the r_cpu to the correct * value. The hpts goes off and sees the mis-match. * We simply correct it here and the CPU will switch * to the new hpts nextime the tcb gets added to the * the hpts (not this time) :-) */ tcp_set_hpts(inp); } if (tp->t_fb_ptr != NULL) { kern_prefetch(tp->t_fb_ptr, &did_prefetch); did_prefetch = 1; } - if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) { + if ((tp->t_fb->tfb_do_queued_segments != NULL) && tp->t_in_pkt) { if (inp->inp_in_input) tcp_hpts_remove(inp, HPTS_REMOVE_INPUT); dropped = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0); if (dropped) { /* Re-acquire the wlock so we can release the reference */ INP_WLOCK(inp); } } else if (tp->t_in_pkt) { /* * We reach here only if we had a * stack that supported INP_SUPPORTS_MBUFQ * and then somehow switched to a stack that * does not. The packets are basically stranded * and would hang with the connection until * cleanup without this code. Its not the * best way but I know of no other way to * handle it since the stack needs functions * it does not have to handle queued packets. */ tcp_drop_in_pkts(tp); } if (in_pcbrele_wlocked(inp) == 0) INP_WUNLOCK(inp); INP_UNLOCK_ASSERT(inp); #ifdef VIMAGE CURVNET_RESTORE(); #endif mtx_lock(&hpts->p_mtx); hpts->p_inp = NULL; } } static void -tcp_hptsi(struct tcp_hpts_entry *hpts) +tcp_hpts_set_max_sleep(struct tcp_hpts_entry *hpts, int wrap_loop_cnt) +{ + uint32_t t = 0, i, fnd = 0; + + if ((hpts->p_on_queue_cnt) && (wrap_loop_cnt < 2)) { + /* + * Find next slot that is occupied and use that to + * be the sleep time. + */ + for (i = 0, t = hpts_slot(hpts->p_cur_slot, 1); i < NUM_OF_HPTSI_SLOTS; i++) { + if (TAILQ_EMPTY(&hpts->p_hptss[t]) == 0) { + fnd = 1; + break; + } + t = (t + 1) % NUM_OF_HPTSI_SLOTS; + } + KASSERT(fnd != 0, ("Hpts:%p cnt:%d but none found", hpts, hpts->p_on_queue_cnt)); + hpts->p_hpts_sleep_time = min((i + 1), hpts_sleep_max); + } else { + /* No one on the wheel sleep for all but 400 slots or sleep max */ + hpts->p_hpts_sleep_time = hpts_sleep_max; + } +} + +static int32_t +tcp_hptsi(struct tcp_hpts_entry *hpts, int from_callout) { struct tcpcb *tp; struct inpcb *inp = NULL, *ninp; struct timeval tv; - int32_t ticks_to_run, i, error; + uint64_t total_slots_processed = 0; + int32_t slots_to_run, i, error; int32_t paced_cnt = 0; int32_t loop_cnt = 0; int32_t did_prefetch = 0; int32_t prefetch_ninp = 0; int32_t prefetch_tp = 0; int32_t wrap_loop_cnt = 0; + int32_t slot_pos_of_endpoint = 0; + int32_t orig_exit_slot; int16_t set_cpu; + int8_t completed_measure = 0, seen_endpoint = 0; HPTS_MTX_ASSERT(hpts); NET_EPOCH_ASSERT(); - /* record previous info for any logging */ hpts->saved_lasttick = hpts->p_lasttick; hpts->saved_curtick = hpts->p_curtick; hpts->saved_curslot = hpts->p_cur_slot; hpts->saved_prev_slot = hpts->p_prev_slot; hpts->p_lasttick = hpts->p_curtick; hpts->p_curtick = tcp_gethptstick(&tv); - hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); + cts_last_ran[hpts->p_num] = tcp_tv_to_usectick(&tv); + orig_exit_slot = hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); if ((hpts->p_on_queue_cnt == 0) || (hpts->p_lasttick == hpts->p_curtick)) { /* * No time has yet passed, * or nothing to do. */ hpts->p_prev_slot = hpts->p_cur_slot; hpts->p_lasttick = hpts->p_curtick; goto no_run; } again: hpts->p_wheel_complete = 0; HPTS_MTX_ASSERT(hpts); - ticks_to_run = hpts_ticks_diff(hpts->p_prev_slot, hpts->p_cur_slot); - if (((hpts->p_curtick - hpts->p_lasttick) > ticks_to_run) && + slots_to_run = hpts_slots_diff(hpts->p_prev_slot, hpts->p_cur_slot); + if (((hpts->p_curtick - hpts->p_lasttick) > + ((NUM_OF_HPTSI_SLOTS-1) * HPTS_TICKS_PER_SLOT)) && (hpts->p_on_queue_cnt != 0)) { /* * Wheel wrap is occuring, basically we * are behind and the distance between * run's has spread so much it has exceeded * the time on the wheel (1.024 seconds). This * is ugly and should NOT be happening. We * need to run the entire wheel. We last processed * p_prev_slot, so that needs to be the last slot * we run. The next slot after that should be our * reserved first slot for new, and then starts * the running postion. Now the problem is the * reserved "not to yet" place does not exist * and there may be inp's in there that need * running. We can merge those into the * first slot at the head. */ wrap_loop_cnt++; - hpts->p_nxt_slot = hpts_tick(hpts->p_prev_slot, 1); - hpts->p_runningtick = hpts_tick(hpts->p_prev_slot, 2); + hpts->p_nxt_slot = hpts_slot(hpts->p_prev_slot, 1); + hpts->p_runningslot = hpts_slot(hpts->p_prev_slot, 2); /* * Adjust p_cur_slot to be where we are starting from * hopefully we will catch up (fat chance if something * is broken this bad :( ) */ hpts->p_cur_slot = hpts->p_prev_slot; /* * The next slot has guys to run too, and that would * be where we would normally start, lets move them into * the next slot (p_prev_slot + 2) so that we will * run them, the extra 10usecs of late (by being * put behind) does not really matter in this situation. */ #ifdef INVARIANTS /* * To prevent a panic we need to update the inpslot to the * new location. This is safe since it takes both the * INP lock and the pacer mutex to change the inp_hptsslot. */ TAILQ_FOREACH(inp, &hpts->p_hptss[hpts->p_nxt_slot], inp_hpts) { - inp->inp_hptsslot = hpts->p_runningtick; + inp->inp_hptsslot = hpts->p_runningslot; } #endif - TAILQ_CONCAT(&hpts->p_hptss[hpts->p_runningtick], + TAILQ_CONCAT(&hpts->p_hptss[hpts->p_runningslot], &hpts->p_hptss[hpts->p_nxt_slot], inp_hpts); - ticks_to_run = NUM_OF_HPTSI_SLOTS - 1; + slots_to_run = NUM_OF_HPTSI_SLOTS - 1; counter_u64_add(wheel_wrap, 1); } else { /* - * Nxt slot is always one after p_runningtick though + * Nxt slot is always one after p_runningslot though * its not used usually unless we are doing wheel wrap. */ hpts->p_nxt_slot = hpts->p_prev_slot; - hpts->p_runningtick = hpts_tick(hpts->p_prev_slot, 1); - } -#ifdef INVARIANTS - if (TAILQ_EMPTY(&hpts->p_input) && - (hpts->p_on_inqueue_cnt != 0)) { - panic("tp:%p in_hpts input empty but cnt:%d", - hpts, hpts->p_on_inqueue_cnt); + hpts->p_runningslot = hpts_slot(hpts->p_prev_slot, 1); } -#endif + KASSERT((((TAILQ_EMPTY(&hpts->p_input) != 0) && (hpts->p_on_inqueue_cnt == 0)) || + ((TAILQ_EMPTY(&hpts->p_input) == 0) && (hpts->p_on_inqueue_cnt > 0))), + ("%s hpts:%p in_hpts cnt:%d and queue state mismatch", + __FUNCTION__, hpts, hpts->p_on_inqueue_cnt)); HPTS_MTX_ASSERT(hpts); if (hpts->p_on_queue_cnt == 0) { goto no_one; } HPTS_MTX_ASSERT(hpts); - for (i = 0; i < ticks_to_run; i++) { + for (i = 0; i < slots_to_run; i++) { /* * Calculate our delay, if there are no extra ticks there - * was not any (i.e. if ticks_to_run == 1, no delay). + * was not any (i.e. if slots_to_run == 1, no delay). */ - hpts->p_delayed_by = (ticks_to_run - (i + 1)) * HPTS_TICKS_PER_USEC; + hpts->p_delayed_by = (slots_to_run - (i + 1)) * HPTS_TICKS_PER_SLOT; HPTS_MTX_ASSERT(hpts); - while ((inp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningtick])) != NULL) { + while ((inp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningslot])) != NULL) { + HPTS_MTX_ASSERT(hpts); /* For debugging */ + if (seen_endpoint == 0) { + seen_endpoint = 1; + orig_exit_slot = slot_pos_of_endpoint = hpts->p_runningslot; + } else if (completed_measure == 0) { + /* Record the new position */ + orig_exit_slot = hpts->p_runningslot; + } + total_slots_processed++; hpts->p_inp = inp; paced_cnt++; -#ifdef INVARIANTS - if (hpts->p_runningtick != inp->inp_hptsslot) { - panic("Hpts:%p inp:%p slot mis-aligned %u vs %u", - hpts, inp, hpts->p_runningtick, inp->inp_hptsslot); - } -#endif + KASSERT(hpts->p_runningslot == inp->inp_hptsslot, + ("Hpts:%p inp:%p slot mis-aligned %u vs %u", + hpts, inp, hpts->p_runningslot, inp->inp_hptsslot)); /* Now pull it */ if (inp->inp_hpts_cpu_set == 0) { set_cpu = 1; } else { set_cpu = 0; } - hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[hpts->p_runningtick], 0); - if ((ninp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningtick])) != NULL) { + hpts_sane_pace_remove(hpts, inp, &hpts->p_hptss[hpts->p_runningslot], 0); + if ((ninp = TAILQ_FIRST(&hpts->p_hptss[hpts->p_runningslot])) != NULL) { /* We prefetch the next inp if possible */ kern_prefetch(ninp, &prefetch_ninp); prefetch_ninp = 1; } if (inp->inp_hpts_request) { /* * This guy is deferred out further in time * then our wheel had available on it. * Push him back on the wheel or run it * depending. */ - uint32_t maxticks, last_tick, remaining_slots; + uint32_t maxslots, last_slot, remaining_slots; - remaining_slots = ticks_to_run - (i + 1); + remaining_slots = slots_to_run - (i + 1); if (inp->inp_hpts_request > remaining_slots) { /* * How far out can we go? */ - maxticks = max_ticks_available(hpts, hpts->p_cur_slot, &last_tick); - if (maxticks >= inp->inp_hpts_request) { + maxslots = max_slots_available(hpts, hpts->p_cur_slot, &last_slot); + if (maxslots >= inp->inp_hpts_request) { /* we can place it finally to be processed */ - inp->inp_hptsslot = hpts_tick(hpts->p_runningtick, inp->inp_hpts_request); + inp->inp_hptsslot = hpts_slot(hpts->p_runningslot, inp->inp_hpts_request); inp->inp_hpts_request = 0; } else { /* Work off some more time */ - inp->inp_hptsslot = last_tick; - inp->inp_hpts_request-= maxticks; + inp->inp_hptsslot = last_slot; + inp->inp_hpts_request-= maxslots; } hpts_sane_pace_insert(hpts, inp, &hpts->p_hptss[inp->inp_hptsslot], __LINE__, 1); hpts->p_inp = NULL; continue; } inp->inp_hpts_request = 0; /* Fall through we will so do it now */ } /* * We clear the hpts flag here after dealing with * remaining slots. This way anyone looking with the * TCB lock will see its on the hpts until just * before we unlock. */ inp->inp_in_hpts = 0; mtx_unlock(&hpts->p_mtx); INP_WLOCK(inp); if (in_pcbrele_wlocked(inp)) { mtx_lock(&hpts->p_mtx); hpts->p_inp = NULL; continue; } if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) || (inp->inp_flags2 & INP_FREED)) { out_now: -#ifdef INVARIANTS - if (mtx_owned(&hpts->p_mtx)) { - panic("Hpts:%p owns mtx prior-to lock line:%d", - hpts, __LINE__); - } -#endif + KASSERT(mtx_owned(&hpts->p_mtx) == 0, + ("Hpts:%p owns mtx prior-to lock line:%d", + hpts, __LINE__)); INP_WUNLOCK(inp); mtx_lock(&hpts->p_mtx); hpts->p_inp = NULL; continue; } tp = intotcpcb(inp); if ((tp == NULL) || (tp->t_inpcb == NULL)) { goto out_now; } if (set_cpu) { /* * Setup so the next time we will move to * the right CPU. This should be a rare * event. It will sometimes happens when we * are the client side (usually not the * server). Somehow tcp_output() gets called * before the tcp_do_segment() sets the * intial state. This means the r_cpu and * r_hpts_cpu is 0. We get on the hpts, and * then tcp_input() gets called setting up * the r_cpu to the correct value. The hpts * goes off and sees the mis-match. We * simply correct it here and the CPU will * switch to the new hpts nextime the tcb * gets added to the the hpts (not this one) * :-) */ tcp_set_hpts(inp); } #ifdef VIMAGE CURVNET_SET(inp->inp_vnet); #endif /* Lets do any logging that we might want to */ if (hpts_does_tp_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { - tcp_hpts_log(hpts, tp, &tv, ticks_to_run, i); + tcp_hpts_log(hpts, tp, &tv, slots_to_run, i, from_callout); } /* * There is a hole here, we get the refcnt on the * inp so it will still be preserved but to make * sure we can get the INP we need to hold the p_mtx * above while we pull out the tp/inp, as long as * fini gets the lock first we are assured of having * a sane INP we can lock and test. */ -#ifdef INVARIANTS - if (mtx_owned(&hpts->p_mtx)) { - panic("Hpts:%p owns mtx before tcp-output:%d", - hpts, __LINE__); - } -#endif + KASSERT(mtx_owned(&hpts->p_mtx) == 0, + ("Hpts:%p owns mtx prior-to tcp_output call line:%d", + hpts, __LINE__)); + if (tp->t_fb_ptr != NULL) { kern_prefetch(tp->t_fb_ptr, &did_prefetch); did_prefetch = 1; } if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) { error = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0); if (error) { /* The input killed the connection */ goto skip_pacing; } } inp->inp_hpts_calls = 1; error = tp->t_fb->tfb_tcp_output(tp); inp->inp_hpts_calls = 0; if (ninp && ninp->inp_ppcb) { /* * If we have a nxt inp, see if we can * prefetch its ppcb. Note this may seem * "risky" since we have no locks (other * than the previous inp) and there no * assurance that ninp was not pulled while * we were processing inp and freed. If this * occured it could mean that either: * * a) Its NULL (which is fine we won't go * here) b) Its valid (which is cool we * will prefetch it) c) The inp got * freed back to the slab which was * reallocated. Then the piece of memory was * re-used and something else (not an * address) is in inp_ppcb. If that occurs * we don't crash, but take a TLB shootdown * performance hit (same as if it was NULL * and we tried to pre-fetch it). * * Considering that the likelyhood of is * quite rare we will take a risk on doing * this. If performance drops after testing * we can always take this out. NB: the * kern_prefetch on amd64 actually has * protection against a bad address now via * the DMAP_() tests. This will prevent the * TLB hit, and instead if occurs just * cause us to load cache with a useless * address (to us). */ kern_prefetch(ninp->inp_ppcb, &prefetch_tp); prefetch_tp = 1; } INP_WUNLOCK(inp); skip_pacing: #ifdef VIMAGE CURVNET_RESTORE(); #endif INP_UNLOCK_ASSERT(inp); -#ifdef INVARIANTS - if (mtx_owned(&hpts->p_mtx)) { - panic("Hpts:%p owns mtx prior-to lock line:%d", - hpts, __LINE__); - } -#endif + KASSERT(mtx_owned(&hpts->p_mtx) == 0, + ("Hpts:%p owns mtx prior-to lock line:%d", + hpts, __LINE__)); mtx_lock(&hpts->p_mtx); hpts->p_inp = NULL; } + if (seen_endpoint) { + /* + * We now have a accurate distance between + * slot_pos_of_endpoint <-> orig_exit_slot + * to tell us how late we were, orig_exit_slot + * is where we calculated the end of our cycle to + * be when we first entered. + */ + completed_measure = 1; + } HPTS_MTX_ASSERT(hpts); hpts->p_inp = NULL; - hpts->p_runningtick++; - if (hpts->p_runningtick >= NUM_OF_HPTSI_SLOTS) { - hpts->p_runningtick = 0; + hpts->p_runningslot++; + if (hpts->p_runningslot >= NUM_OF_HPTSI_SLOTS) { + hpts->p_runningslot = 0; } } no_one: HPTS_MTX_ASSERT(hpts); hpts->p_delayed_by = 0; /* * Check to see if we took an excess amount of time and need to run * more ticks (if we did not hit eno-bufs). */ -#ifdef INVARIANTS - if (TAILQ_EMPTY(&hpts->p_input) && - (hpts->p_on_inqueue_cnt != 0)) { - panic("tp:%p in_hpts input empty but cnt:%d", - hpts, hpts->p_on_inqueue_cnt); - } -#endif + KASSERT((((TAILQ_EMPTY(&hpts->p_input) != 0) && (hpts->p_on_inqueue_cnt == 0)) || + ((TAILQ_EMPTY(&hpts->p_input) == 0) && (hpts->p_on_inqueue_cnt > 0))), + ("%s hpts:%p in_hpts cnt:%d queue state mismatch", + __FUNCTION__, hpts, hpts->p_on_inqueue_cnt)); hpts->p_prev_slot = hpts->p_cur_slot; hpts->p_lasttick = hpts->p_curtick; - if (loop_cnt > max_pacer_loops) { + if ((from_callout == 0) || (loop_cnt > max_pacer_loops)) { /* * Something is serious slow we have * looped through processing the wheel * and by the time we cleared the * needs to run max_pacer_loops time * we still needed to run. That means * the system is hopelessly behind and * can never catch up :( * * We will just lie to this thread * and let it thing p_curtick is * correct. When it next awakens * it will find itself further behind. */ - counter_u64_add(hpts_hopelessly_behind, 1); + if (from_callout) + counter_u64_add(hpts_hopelessly_behind, 1); goto no_run; } hpts->p_curtick = tcp_gethptstick(&tv); hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); + if (seen_endpoint == 0) { + /* We saw no endpoint but we may be looping */ + orig_exit_slot = hpts->p_cur_slot; + } if ((wrap_loop_cnt < 2) && (hpts->p_lasttick != hpts->p_curtick)) { counter_u64_add(hpts_loops, 1); loop_cnt++; goto again; } no_run: + cts_last_ran[hpts->p_num] = tcp_tv_to_usectick(&tv); /* * Set flag to tell that we are done for * any slot input that happens during * input. */ hpts->p_wheel_complete = 1; /* * Run any input that may be there not covered * in running data. */ if (!TAILQ_EMPTY(&hpts->p_input)) { tcp_input_data(hpts, &tv); /* - * Now did we spend too long running - * input and need to run more ticks? + * Now did we spend too long running input and need to run more ticks? + * Note that if wrap_loop_cnt < 2 then we should have the conditions + * in the KASSERT's true. But if the wheel is behind i.e. wrap_loop_cnt + * is greater than 2, then the condtion most likely are *not* true. Also + * if we are called not from the callout, we don't run the wheel multiple + * times so the slots may not align either. */ - KASSERT(hpts->p_prev_slot == hpts->p_cur_slot, + KASSERT(((hpts->p_prev_slot == hpts->p_cur_slot) || + (wrap_loop_cnt >= 2) || (from_callout == 0)), ("H:%p p_prev_slot:%u not equal to p_cur_slot:%u", hpts, hpts->p_prev_slot, hpts->p_cur_slot)); - KASSERT(hpts->p_lasttick == hpts->p_curtick, + KASSERT(((hpts->p_lasttick == hpts->p_curtick) + || (wrap_loop_cnt >= 2) || (from_callout == 0)), ("H:%p p_lasttick:%u not equal to p_curtick:%u", hpts, hpts->p_lasttick, hpts->p_curtick)); - hpts->p_curtick = tcp_gethptstick(&tv); - if (hpts->p_lasttick != hpts->p_curtick) { + if (from_callout && (hpts->p_lasttick != hpts->p_curtick)) { + hpts->p_curtick = tcp_gethptstick(&tv); counter_u64_add(hpts_loops, 1); hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); goto again; } } - { - uint32_t t = 0, i, fnd = 0; - - if ((hpts->p_on_queue_cnt) && (wrap_loop_cnt < 2)) { - /* - * Find next slot that is occupied and use that to - * be the sleep time. - */ - for (i = 0, t = hpts_tick(hpts->p_cur_slot, 1); i < NUM_OF_HPTSI_SLOTS; i++) { - if (TAILQ_EMPTY(&hpts->p_hptss[t]) == 0) { - fnd = 1; - break; - } - t = (t + 1) % NUM_OF_HPTSI_SLOTS; - } - if (fnd) { - hpts->p_hpts_sleep_time = min((i + 1), hpts_sleep_max); - } else { -#ifdef INVARIANTS - panic("Hpts:%p cnt:%d but none found", hpts, hpts->p_on_queue_cnt); -#endif - counter_u64_add(back_tosleep, 1); - hpts->p_on_queue_cnt = 0; - goto non_found; - } - } else if (wrap_loop_cnt >= 2) { - /* Special case handling */ - hpts->p_hpts_sleep_time = tcp_min_hptsi_time; - } else { - /* No one on the wheel sleep for all but 400 slots or sleep max */ - non_found: - hpts->p_hpts_sleep_time = hpts_sleep_max; - } + if (from_callout){ + tcp_hpts_set_max_sleep(hpts, wrap_loop_cnt); } + if (seen_endpoint) + return(hpts_slots_diff(slot_pos_of_endpoint, orig_exit_slot)); + else + return (0); } void __tcp_set_hpts(struct inpcb *inp, int32_t line) { struct tcp_hpts_entry *hpts; + int failed; INP_WLOCK_ASSERT(inp); hpts = tcp_hpts_lock(inp); if ((inp->inp_in_hpts == 0) && (inp->inp_hpts_cpu_set == 0)) { - inp->inp_hpts_cpu = hpts_cpuid(inp); - inp->inp_hpts_cpu_set = 1; + inp->inp_hpts_cpu = hpts_cpuid(inp, &failed); + if (failed == 0) + inp->inp_hpts_cpu_set = 1; } mtx_unlock(&hpts->p_mtx); hpts = tcp_input_lock(inp); if ((inp->inp_input_cpu_set == 0) && (inp->inp_in_input == 0)) { - inp->inp_input_cpu = hpts_cpuid(inp); - inp->inp_input_cpu_set = 1; + inp->inp_input_cpu = hpts_cpuid(inp, &failed); + if (failed == 0) + inp->inp_input_cpu_set = 1; } mtx_unlock(&hpts->p_mtx); } uint16_t tcp_hpts_delayedby(struct inpcb *inp){ return (tcp_pace.rp_ent[inp->inp_hpts_cpu]->p_delayed_by); } +static void +__tcp_run_hpts(struct tcp_hpts_entry *hpts) +{ + int ticks_ran; + + if (hpts->p_hpts_active) { + /* Already active */ + return; + } + if (mtx_trylock(&hpts->p_mtx) == 0) { + /* Someone else got the lock */ + return; + } + if (hpts->p_hpts_active) + goto out_with_mtx; + hpts->syscall_cnt++; + counter_u64_add(hpts_direct_call, 1); + hpts->p_hpts_active = 1; + ticks_ran = tcp_hptsi(hpts, 0); + /* We may want to adjust the sleep values here */ + if (hpts->p_on_queue_cnt >= conn_cnt_thresh) { + if (ticks_ran > ticks_indicate_less_sleep) { + struct timeval tv; + sbintime_t sb; + int cpu; + + hpts->p_mysleep.tv_usec /= 2; + if (hpts->p_mysleep.tv_usec < dynamic_min_sleep) + hpts->p_mysleep.tv_usec = dynamic_min_sleep; + /* Reschedule with new to value */ + tcp_hpts_set_max_sleep(hpts, 0); + tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_SLOT; + /* Validate its in the right ranges */ + if (tv.tv_usec < hpts->p_mysleep.tv_usec) { + hpts->overidden_sleep = tv.tv_usec; + tv.tv_usec = hpts->p_mysleep.tv_usec; + } else if (tv.tv_usec > dynamic_max_sleep) { + /* Lets not let sleep get above this value */ + hpts->overidden_sleep = tv.tv_usec; + tv.tv_usec = dynamic_max_sleep; + } + /* + * In this mode the timer is a backstop to + * all the userret/lro_flushes so we use + * the dynamic value and set the on_min_sleep + * flag so we will not be awoken. + */ + sb = tvtosbt(tv); + cpu = (tcp_bind_threads || hpts_use_assigned_cpu) ? hpts->p_cpu : curcpu; + /* Store off to make visible the actual sleep time */ + hpts->sleeping = tv.tv_usec; + callout_reset_sbt_on(&hpts->co, sb, 0, + hpts_timeout_swi, hpts, cpu, + (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); + } else if (ticks_ran < ticks_indicate_more_sleep) { + /* For the further sleep, don't reschedule hpts */ + hpts->p_mysleep.tv_usec *= 2; + if (hpts->p_mysleep.tv_usec > dynamic_max_sleep) + hpts->p_mysleep.tv_usec = dynamic_max_sleep; + } + hpts->p_on_min_sleep = 1; + } + hpts->p_hpts_active = 0; +out_with_mtx: + HPTS_MTX_ASSERT(hpts); + mtx_unlock(&hpts->p_mtx); +} + +static struct tcp_hpts_entry * +tcp_choose_hpts_to_run() +{ + int i, oldest_idx; + uint32_t cts, time_since_ran, calc; + + if ((hpts_uses_oldest == 0) || + ((hpts_uses_oldest > 1) && + (tcp_pace.rp_ent[(tcp_pace.rp_num_hptss-1)]->p_on_queue_cnt >= hpts_uses_oldest))) { + /* + * We have either disabled the feature (0), or + * we have crossed over the oldest threshold on the + * last hpts. We use the last one for simplification + * since we don't want to use the first one (it may + * have starting connections that have not settled + * on the cpu yet). + */ + return(tcp_pace.rp_ent[(curcpu % tcp_pace.rp_num_hptss)]); + } + /* Lets find the oldest hpts to attempt to run */ + cts = tcp_get_usecs(NULL); + time_since_ran = 0; + oldest_idx = -1; + for (i = 0; i < tcp_pace.rp_num_hptss; i++) { + if (TSTMP_GT(cts, cts_last_ran[i])) + calc = cts - cts_last_ran[i]; + else + calc = 0; + if (calc > time_since_ran) { + oldest_idx = i; + time_since_ran = calc; + } + } + if (oldest_idx >= 0) + return(tcp_pace.rp_ent[oldest_idx]); + else + return(tcp_pace.rp_ent[(curcpu % tcp_pace.rp_num_hptss)]); +} + + +void +tcp_run_hpts(void) +{ + static struct tcp_hpts_entry *hpts; + struct epoch_tracker et; + + NET_EPOCH_ENTER(et); + hpts = tcp_choose_hpts_to_run(); + __tcp_run_hpts(hpts); + NET_EPOCH_EXIT(et); +} + + static void tcp_hpts_thread(void *ctx) { struct tcp_hpts_entry *hpts; struct epoch_tracker et; struct timeval tv; sbintime_t sb; + int cpu, ticks_ran; hpts = (struct tcp_hpts_entry *)ctx; mtx_lock(&hpts->p_mtx); if (hpts->p_direct_wake) { - /* Signaled by input */ + /* Signaled by input or output with low occupancy count. */ callout_stop(&hpts->co); + counter_u64_add(hpts_direct_awakening, 1); } else { - /* Timed out */ + /* Timed out, the normal case. */ + counter_u64_add(hpts_wake_timeout, 1); if (callout_pending(&hpts->co) || !callout_active(&hpts->co)) { mtx_unlock(&hpts->p_mtx); return; } - callout_deactivate(&hpts->co); } + callout_deactivate(&hpts->co); hpts->p_hpts_wake_scheduled = 0; - hpts->p_hpts_active = 1; NET_EPOCH_ENTER(et); - tcp_hptsi(hpts); - NET_EPOCH_EXIT(et); - HPTS_MTX_ASSERT(hpts); + if (hpts->p_hpts_active) { + /* + * We are active already. This means that a syscall + * trap or LRO is running in behalf of hpts. In that case + * we need to double our timeout since there seems to be + * enough activity in the system that we don't need to + * run as often (if we were not directly woken). + */ + if (hpts->p_direct_wake == 0) { + counter_u64_add(hpts_back_tosleep, 1); + if (hpts->p_on_queue_cnt >= conn_cnt_thresh) { + hpts->p_mysleep.tv_usec *= 2; + if (hpts->p_mysleep.tv_usec > dynamic_max_sleep) + hpts->p_mysleep.tv_usec = dynamic_max_sleep; + tv.tv_usec = hpts->p_mysleep.tv_usec; + hpts->p_on_min_sleep = 1; + } else { + /* + * Here we have low count on the wheel, but + * somehow we still collided with one of the + * connections. Lets go back to sleep for a + * min sleep time, but clear the flag so we + * can be awoken by insert. + */ + hpts->p_on_min_sleep = 0; + tv.tv_usec = tcp_min_hptsi_time; + } + } else { + /* + * Directly woken most likely to reset the + * callout time. + */ + tv.tv_sec = 0; + tv.tv_usec = hpts->p_mysleep.tv_usec; + } + goto back_to_sleep; + } + hpts->sleeping = 0; + hpts->p_hpts_active = 1; + ticks_ran = tcp_hptsi(hpts, 1); tv.tv_sec = 0; - tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC; - if (tcp_min_hptsi_time && (tv.tv_usec < tcp_min_hptsi_time)) { - hpts->overidden_sleep = tv.tv_usec; - tv.tv_usec = tcp_min_hptsi_time; + tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_SLOT; + if (hpts->p_on_queue_cnt >= conn_cnt_thresh) { + if(hpts->p_direct_wake == 0) { + /* + * Only adjust sleep time if we were + * called from the callout i.e. direct_wake == 0. + */ + if (ticks_ran < ticks_indicate_more_sleep) { + hpts->p_mysleep.tv_usec *= 2; + if (hpts->p_mysleep.tv_usec > dynamic_max_sleep) + hpts->p_mysleep.tv_usec = dynamic_max_sleep; + } else if (ticks_ran > ticks_indicate_less_sleep) { + hpts->p_mysleep.tv_usec /= 2; + if (hpts->p_mysleep.tv_usec < dynamic_min_sleep) + hpts->p_mysleep.tv_usec = dynamic_min_sleep; + } + } + if (tv.tv_usec < hpts->p_mysleep.tv_usec) { + hpts->overidden_sleep = tv.tv_usec; + tv.tv_usec = hpts->p_mysleep.tv_usec; + } else if (tv.tv_usec > dynamic_max_sleep) { + /* Lets not let sleep get above this value */ + hpts->overidden_sleep = tv.tv_usec; + tv.tv_usec = dynamic_max_sleep; + } + /* + * In this mode the timer is a backstop to + * all the userret/lro_flushes so we use + * the dynamic value and set the on_min_sleep + * flag so we will not be awoken. + */ hpts->p_on_min_sleep = 1; - } else { - /* Clear the min sleep flag */ - hpts->overidden_sleep = 0; + } else if (hpts->p_on_queue_cnt == 0) { + /* + * No one on the wheel, please wake us up + * if you insert on the wheel. + */ hpts->p_on_min_sleep = 0; - } - hpts->p_hpts_active = 0; - sb = tvtosbt(tv); - if (tcp_hpts_callout_skip_swi == 0) { - callout_reset_sbt_on(&hpts->co, sb, 0, - hpts_timeout_swi, hpts, hpts->p_cpu, - (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); + hpts->overidden_sleep = 0; } else { - callout_reset_sbt_on(&hpts->co, sb, 0, - hpts_timeout_dir, hpts, - hpts->p_cpu, - C_PREL(tcp_hpts_precision)); + /* + * We hit here when we have a low number of + * clients on the wheel (our else clause). + * We may need to go on min sleep, if we set + * the flag we will not be awoken if someone + * is inserted ahead of us. Clearing the flag + * means we can be awoken. This is "old mode" + * where the timer is what runs hpts mainly. + */ + if (tv.tv_usec < tcp_min_hptsi_time) { + /* + * Yes on min sleep, which means + * we cannot be awoken. + */ + hpts->overidden_sleep = tv.tv_usec; + tv.tv_usec = tcp_min_hptsi_time; + hpts->p_on_min_sleep = 1; + } else { + /* Clear the min sleep flag */ + hpts->overidden_sleep = 0; + hpts->p_on_min_sleep = 0; + } } + HPTS_MTX_ASSERT(hpts); + hpts->p_hpts_active = 0; +back_to_sleep: hpts->p_direct_wake = 0; + sb = tvtosbt(tv); + cpu = (tcp_bind_threads || hpts_use_assigned_cpu) ? hpts->p_cpu : curcpu; + /* Store off to make visible the actual sleep time */ + hpts->sleeping = tv.tv_usec; + callout_reset_sbt_on(&hpts->co, sb, 0, + hpts_timeout_swi, hpts, cpu, + (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); + NET_EPOCH_EXIT(et); mtx_unlock(&hpts->p_mtx); } #undef timersub static void tcp_init_hptsi(void *st) { int32_t i, j, error, bound = 0, created = 0; size_t sz, asz; struct timeval tv; sbintime_t sb; struct tcp_hpts_entry *hpts; struct pcpu *pc; cpuset_t cs; char unit[16]; uint32_t ncpus = mp_ncpus ? mp_ncpus : MAXCPU; - int count, domain; + int count, domain, cpu; tcp_pace.rp_proc = NULL; tcp_pace.rp_num_hptss = ncpus; hpts_hopelessly_behind = counter_u64_alloc(M_WAITOK); hpts_loops = counter_u64_alloc(M_WAITOK); back_tosleep = counter_u64_alloc(M_WAITOK); combined_wheel_wrap = counter_u64_alloc(M_WAITOK); wheel_wrap = counter_u64_alloc(M_WAITOK); + hpts_wake_timeout = counter_u64_alloc(M_WAITOK); + hpts_direct_awakening = counter_u64_alloc(M_WAITOK); + hpts_back_tosleep = counter_u64_alloc(M_WAITOK); + hpts_direct_call = counter_u64_alloc(M_WAITOK); + cpu_uses_flowid = counter_u64_alloc(M_WAITOK); + cpu_uses_random = counter_u64_alloc(M_WAITOK); + + sz = (tcp_pace.rp_num_hptss * sizeof(struct tcp_hpts_entry *)); tcp_pace.rp_ent = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO); + sz = (sizeof(uint32_t) * tcp_pace.rp_num_hptss); + cts_last_ran = malloc(sz, M_TCPHPTS, M_WAITOK); asz = sizeof(struct hptsh) * NUM_OF_HPTSI_SLOTS; for (i = 0; i < tcp_pace.rp_num_hptss; i++) { tcp_pace.rp_ent[i] = malloc(sizeof(struct tcp_hpts_entry), M_TCPHPTS, M_WAITOK | M_ZERO); tcp_pace.rp_ent[i]->p_hptss = malloc(asz, M_TCPHPTS, M_WAITOK); hpts = tcp_pace.rp_ent[i]; /* * Init all the hpts structures that are not specifically * zero'd by the allocations. Also lets attach them to the * appropriate sysctl block as well. */ mtx_init(&hpts->p_mtx, "tcp_hpts_lck", "hpts", MTX_DEF | MTX_DUPOK); TAILQ_INIT(&hpts->p_input); for (j = 0; j < NUM_OF_HPTSI_SLOTS; j++) { TAILQ_INIT(&hpts->p_hptss[j]); } sysctl_ctx_init(&hpts->hpts_ctx); sprintf(unit, "%d", i); hpts->hpts_root = SYSCTL_ADD_NODE(&hpts->hpts_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp_hpts), OID_AUTO, unit, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); SYSCTL_ADD_INT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "in_qcnt", CTLFLAG_RD, &hpts->p_on_inqueue_cnt, 0, "Count TCB's awaiting input processing"); SYSCTL_ADD_INT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "out_qcnt", CTLFLAG_RD, &hpts->p_on_queue_cnt, 0, "Count TCB's awaiting output processing"); SYSCTL_ADD_U16(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "active", CTLFLAG_RD, &hpts->p_hpts_active, 0, "Is the hpts active"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "curslot", CTLFLAG_RD, &hpts->p_cur_slot, 0, "What the current running pacers goal"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "runtick", CTLFLAG_RD, - &hpts->p_runningtick, 0, + &hpts->p_runningslot, 0, "What the running pacers current slot is"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "curtick", CTLFLAG_RD, &hpts->p_curtick, 0, "What the running pacers last tick mapped to the wheel was"); + SYSCTL_ADD_UINT(&hpts->hpts_ctx, + SYSCTL_CHILDREN(hpts->hpts_root), + OID_AUTO, "lastran", CTLFLAG_RD, + &cts_last_ran[i], 0, + "The last usec tick that this hpts ran"); + SYSCTL_ADD_U64(&hpts->hpts_ctx, + SYSCTL_CHILDREN(hpts->hpts_root), + OID_AUTO, "cur_min_sleep", CTLFLAG_RD, + &hpts->p_mysleep.tv_usec, 0, + "What the running pacers is using for p_mysleep.tv_usec"); + SYSCTL_ADD_U64(&hpts->hpts_ctx, + SYSCTL_CHILDREN(hpts->hpts_root), + OID_AUTO, "now_sleeping", CTLFLAG_RD, + &hpts->sleeping, 0, + "What the running pacers is actually sleeping for"); + SYSCTL_ADD_U64(&hpts->hpts_ctx, + SYSCTL_CHILDREN(hpts->hpts_root), + OID_AUTO, "syscall_cnt", CTLFLAG_RD, + &hpts->syscall_cnt, 0, + "How many times we had syscalls on this hpts"); + hpts->p_hpts_sleep_time = hpts_sleep_max; hpts->p_num = i; hpts->p_curtick = tcp_gethptstick(&tv); + cts_last_ran[i] = tcp_tv_to_usectick(&tv); hpts->p_prev_slot = hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); hpts->p_cpu = 0xffff; - hpts->p_nxt_slot = hpts_tick(hpts->p_cur_slot, 1); + hpts->p_nxt_slot = hpts_slot(hpts->p_cur_slot, 1); callout_init(&hpts->co, 1); } /* Don't try to bind to NUMA domains if we don't have any */ if (vm_ndomains == 1 && tcp_bind_threads == 2) tcp_bind_threads = 0; /* * Now lets start ithreads to handle the hptss. */ - CPU_FOREACH(i) { + for (i = 0; i < tcp_pace.rp_num_hptss; i++) { hpts = tcp_pace.rp_ent[i]; hpts->p_cpu = i; error = swi_add(&hpts->ie, "hpts", tcp_hpts_thread, (void *)hpts, SWI_NET, INTR_MPSAFE, &hpts->ie_cookie); - if (error) { - panic("Can't add hpts:%p i:%d err:%d", - hpts, i, error); - } + KASSERT(error == 0, + ("Can't add hpts:%p i:%d err:%d", + hpts, i, error)); created++; + hpts->p_mysleep.tv_sec = 0; + hpts->p_mysleep.tv_usec = tcp_min_hptsi_time; if (tcp_bind_threads == 1) { if (intr_event_bind(hpts->ie, i) == 0) bound++; } else if (tcp_bind_threads == 2) { pc = pcpu_find(i); domain = pc->pc_domain; CPU_COPY(&cpuset_domain[domain], &cs); if (intr_event_bind_ithread_cpuset(hpts->ie, &cs) == 0) { bound++; count = hpts_domains[domain].count; hpts_domains[domain].cpu[count] = i; hpts_domains[domain].count++; } } tv.tv_sec = 0; - tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_USEC; + tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_SLOT; + hpts->sleeping = tv.tv_usec; sb = tvtosbt(tv); - if (tcp_hpts_callout_skip_swi == 0) { - callout_reset_sbt_on(&hpts->co, sb, 0, - hpts_timeout_swi, hpts, hpts->p_cpu, - (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); - } else { - callout_reset_sbt_on(&hpts->co, sb, 0, - hpts_timeout_dir, hpts, - hpts->p_cpu, - C_PREL(tcp_hpts_precision)); - } + cpu = (tcp_bind_threads || hpts_use_assigned_cpu) ? hpts->p_cpu : curcpu; + callout_reset_sbt_on(&hpts->co, sb, 0, + hpts_timeout_swi, hpts, cpu, + (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); } /* * If we somehow have an empty domain, fall back to choosing * among all htps threads. */ for (i = 0; i < vm_ndomains; i++) { if (hpts_domains[i].count == 0) { tcp_bind_threads = 0; break; } } - printf("TCP Hpts created %d swi interrupt threads and bound %d to %s\n", created, bound, tcp_bind_threads == 2 ? "NUMA domains" : "cpus"); +#ifdef INVARIANTS + printf("HPTS is in INVARIANT mode!!\n"); +#endif } -SYSINIT(tcphptsi, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, tcp_init_hptsi, NULL); +SYSINIT(tcphptsi, SI_SUB_SOFTINTR, SI_ORDER_ANY, tcp_init_hptsi, NULL); MODULE_VERSION(tcphpts, 1); diff --git a/sys/netinet/tcp_hpts.h b/sys/netinet/tcp_hpts.h index 772a7ec67202..bfbd34e65ff5 100644 --- a/sys/netinet/tcp_hpts.h +++ b/sys/netinet/tcp_hpts.h @@ -1,267 +1,316 @@ /*- * Copyright (c) 2016-2018 Netflix, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #ifndef __tcp_hpts_h__ #define __tcp_hpts_h__ /* * The hpts uses a 102400 wheel. The wheel * defines the time in 10 usec increments (102400 x 10). * This gives a range of 10usec - 1024ms to place * an entry within. If the user requests more than * 1.024 second, a remaineder is attached and the hpts * when seeing the remainder will re-insert the * inpcb forward in time from where it is until * the remainder is zero. */ #define NUM_OF_HPTSI_SLOTS 102400 TAILQ_HEAD(hptsh, inpcb); /* Number of useconds in a hpts tick */ -#define HPTS_TICKS_PER_USEC 10 +#define HPTS_TICKS_PER_SLOT 10 #define HPTS_MS_TO_SLOTS(x) ((x * 100) + 1) #define HPTS_USEC_TO_SLOTS(x) ((x+9) /10) #define HPTS_USEC_IN_SEC 1000000 #define HPTS_MSEC_IN_SEC 1000 #define HPTS_USEC_IN_MSEC 1000 struct hpts_diag { uint32_t p_hpts_active; /* bbr->flex7 x */ uint32_t p_nxt_slot; /* bbr->flex1 x */ uint32_t p_cur_slot; /* bbr->flex2 x */ uint32_t p_prev_slot; /* bbr->delivered */ - uint32_t p_runningtick; /* bbr->inflight */ + uint32_t p_runningslot; /* bbr->inflight */ uint32_t slot_req; /* bbr->flex3 x */ uint32_t inp_hptsslot; /* bbr->flex4 x */ uint32_t slot_remaining; /* bbr->flex5 x */ uint32_t have_slept; /* bbr->epoch x */ uint32_t hpts_sleep_time; /* bbr->applimited x */ uint32_t yet_to_sleep; /* bbr->lt_epoch x */ uint32_t need_new_to; /* bbr->flex6 x */ - uint32_t wheel_tick; /* bbr->bw_inuse x */ - uint32_t maxticks; /* bbr->delRate x */ + uint32_t wheel_slot; /* bbr->bw_inuse x */ + uint32_t maxslots; /* bbr->delRate x */ uint32_t wheel_cts; /* bbr->rttProp x */ int32_t co_ret; /* bbr->pkts_out x */ uint32_t p_curtick; /* upper bbr->cur_del_rate */ uint32_t p_lasttick; /* lower bbr->cur_del_rate */ uint8_t p_on_min_sleep; /* bbr->flex8 x */ }; /* Magic flags to tell whats cooking on the pacing wheel */ #define PACE_TMR_DELACK 0x01 /* Delayed ack timer running */ #define PACE_TMR_RACK 0x02 /* RACK timer running */ #define PACE_TMR_TLP 0x04 /* TLP timer running */ #define PACE_TMR_RXT 0x08 /* Retransmit timer running */ #define PACE_TMR_PERSIT 0x10 /* Persists timer running */ #define PACE_TMR_KEEP 0x20 /* Keep alive timer running */ #define PACE_PKT_OUTPUT 0x40 /* Output Packets being paced */ #define PACE_TMR_MASK (PACE_TMR_KEEP|PACE_TMR_PERSIT|PACE_TMR_RXT|PACE_TMR_TLP|PACE_TMR_RACK|PACE_TMR_DELACK) +#define DEFAULT_CONNECTION_THESHOLD 100 + #ifdef _KERNEL /* Each hpts has its own p_mtx which is used for locking */ struct tcp_hpts_entry { /* Cache line 0x00 */ struct mtx p_mtx; /* Mutex for hpts */ + struct timeval p_mysleep; /* Our min sleep time */ + uint64_t syscall_cnt; + uint64_t sleeping; /* What the actual sleep was (if sleeping) */ uint16_t p_hpts_active; /* Flag that says hpts is awake */ - uint8_t p_hpts_wake_scheduled; /* Have we scheduled a wakeup? */ uint8_t p_wheel_complete; /* have we completed the wheel arc walk? */ uint32_t p_curtick; /* Tick in 10 us the hpts is going to */ - uint32_t p_runningtick; /* Current tick we are at if we are running */ + uint32_t p_runningslot; /* Current tick we are at if we are running */ uint32_t p_prev_slot; /* Previous slot we were on */ uint32_t p_cur_slot; /* Current slot in wheel hpts is draining */ uint32_t p_nxt_slot; /* The next slot outside the current range of * slots that the hpts is running on. */ int32_t p_on_queue_cnt; /* Count on queue in this hpts */ uint32_t p_lasttick; /* Last tick before the current one */ uint8_t p_direct_wake :1, /* boolean */ p_on_min_sleep:1, /* boolean */ - p_avail:6; + p_hpts_wake_scheduled:1, /* boolean */ + p_avail:5; uint8_t p_fill[3]; /* Fill to 32 bits */ /* Cache line 0x40 */ void *p_inp; struct hptsh p_input; /* For the tcp-input runner */ /* Hptsi wheel */ struct hptsh *p_hptss; int32_t p_on_inqueue_cnt; /* Count on input queue in this hpts */ - uint32_t hit_no_enobuf; - uint32_t p_dyn_adjust; uint32_t p_hpts_sleep_time; /* Current sleep interval having a max * of 255ms */ uint32_t overidden_sleep; /* what was overrided by min-sleep for logging */ uint32_t saved_lasttick; /* for logging */ uint32_t saved_curtick; /* for logging */ uint32_t saved_curslot; /* for logging */ uint32_t saved_prev_slot; /* for logging */ uint32_t p_delayed_by; /* How much were we delayed by */ /* Cache line 0x80 */ struct sysctl_ctx_list hpts_ctx; struct sysctl_oid *hpts_root; struct intr_event *ie; void *ie_cookie; uint16_t p_num; /* The hpts number one per cpu */ uint16_t p_cpu; /* The hpts CPU */ /* There is extra space in here */ /* Cache line 0x100 */ struct callout co __aligned(CACHE_LINE_SIZE); } __aligned(CACHE_LINE_SIZE); struct tcp_hptsi { struct proc *rp_proc; /* Process structure for hpts */ struct tcp_hpts_entry **rp_ent; /* Array of hptss */ + uint32_t *cts_last_ran; uint32_t rp_num_hptss; /* Number of hpts threads */ }; #endif #define HPTS_REMOVE_INPUT 0x01 #define HPTS_REMOVE_OUTPUT 0x02 #define HPTS_REMOVE_ALL (HPTS_REMOVE_INPUT | HPTS_REMOVE_OUTPUT) /* * When using the hpts, a TCP stack must make sure * that once a INP_DROPPED flag is applied to a INP * that it does not expect tcp_output() to ever be * called by the hpts. The hpts will *not* call * any output (or input) functions on a TCB that * is in the DROPPED state. * * This implies final ACK's and RST's that might * be sent when a TCB is still around must be * sent from a routine like tcp_respond(). */ +#define LOWEST_SLEEP_ALLOWED 50 #define DEFAULT_MIN_SLEEP 250 /* How many usec's is default for hpts sleep * this determines min granularity of the - * hpts. If 0, granularity is 10useconds at - * the cost of more CPU (context switching). */ + * hpts. If 1, granularity is 10useconds at + * the cost of more CPU (context switching). + * Note do not set this to 0. + */ +#define DYNAMIC_MIN_SLEEP DEFAULT_MIN_SLEEP +#define DYNAMIC_MAX_SLEEP 100000 /* 100ms */ +/* No of connections when wee start aligning to the cpu from syscalls */ +#define OLDEST_THRESHOLD 1200 +/* Thresholds for raising/lowering sleep */ +#define TICKS_INDICATE_MORE_SLEEP 100 /* This would be 1ms */ +#define TICKS_INDICATE_LESS_SLEEP 1000 /* This would indicate 10ms */ +/** + * + * Dynamic adjustment of sleeping times is done in "new" mode + * where we are depending on syscall returns and lro returns + * to push hpts forward mainly and the timer is only a backstop. + * + * When we are in the "new" mode i.e. conn_cnt > conn_cnt_thresh + * then we do a dynamic adjustment on the time we sleep. + * Our threshold is if the lateness of the first client served (in ticks) is + * greater than or equal too ticks_indicate_more_sleep (10ms + * or 10000 ticks). If we were that late, the actual sleep time + * is adjusted down by 50%. If the ticks_ran is less than + * ticks_indicate_more_sleep (100 ticks or 1000usecs). + * + */ + + #ifdef _KERNEL #define HPTS_MTX_ASSERT(hpts) mtx_assert(&(hpts)->p_mtx, MA_OWNED) struct tcp_hpts_entry *tcp_hpts_lock(struct inpcb *inp); struct tcp_hpts_entry *tcp_input_lock(struct inpcb *inp); int __tcp_queue_to_hpts_immediate(struct inpcb *inp, int32_t line); #define tcp_queue_to_hpts_immediate(a)__tcp_queue_to_hpts_immediate(a, __LINE__) struct tcp_hpts_entry *tcp_cur_hpts(struct inpcb *inp); #define tcp_hpts_remove(a, b) __tcp_hpts_remove(a, b, __LINE__) void __tcp_hpts_remove(struct inpcb *inp, int32_t flags, int32_t line); /* * To insert a TCB on the hpts you *must* be holding the * INP_WLOCK(). The hpts insert code will then acqurire * the hpts's lock and insert the TCB on the requested * slot possibly waking up the hpts if you are requesting * a time earlier than what the hpts is sleeping to (if * the hpts is sleeping). You may check the inp->inp_in_hpts * flag without the hpts lock. The hpts is the only one * that will clear this flag holding only the hpts lock. This * means that in your tcp_output() routine when you test for * it to be 1 (so you wont call output) it may be transitioning * to 0 (by the hpts). That will be fine since that will just * mean an extra call to tcp_output that most likely will find * the call you executed (when the mis-match occured) will have * put the TCB back on the hpts and it will return. If your * call did not add it back to the hpts then you will either * over-send or the cwnd will block you from sending more. * * Note you should also be holding the INP_WLOCK() when you * call the remove from the hpts as well. Thoug usually * you are either doing this from a timer, where you need * that INP_WLOCK() or from destroying your TCB where again * you should already have the INP_WLOCK(). */ uint32_t __tcp_hpts_insert(struct inpcb *inp, uint32_t slot, int32_t line); #define tcp_hpts_insert(a, b) __tcp_hpts_insert(a, b, __LINE__) uint32_t tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag); int __tcp_queue_to_input_locked(struct inpcb *inp, struct tcp_hpts_entry *hpts, int32_t line); #define tcp_queue_to_input_locked(a, b) __tcp_queue_to_input_locked(a, b, __LINE__); int __tcp_queue_to_input(struct inpcb *inp, int32_t line); #define tcp_queue_to_input(a) __tcp_queue_to_input(a, __LINE__) uint16_t tcp_hpts_delayedby(struct inpcb *inp); void __tcp_set_hpts(struct inpcb *inp, int32_t line); #define tcp_set_hpts(a) __tcp_set_hpts(a, __LINE__) void __tcp_set_inp_to_drop(struct inpcb *inp, uint16_t reason, int32_t line); #define tcp_set_inp_to_drop(a, b) __tcp_set_inp_to_drop(a, b, __LINE__) +void tcp_run_hpts(void); + +uint16_t hpts_random_cpu(struct inpcb *inp); + extern int32_t tcp_min_hptsi_time; -static __inline uint32_t -tcp_tv_to_hptstick(struct timeval *sv) -{ - return ((sv->tv_sec * 100000) + (sv->tv_usec / 10)); -} +#endif /* _KERNEL */ +/* + * The following functions should also be available + * to userspace as well. + */ static __inline uint32_t -tcp_gethptstick(struct timeval *sv) +tcp_tv_to_hptstick(const struct timeval *sv) { - struct timeval tv; - - if (sv == NULL) - sv = &tv; - microuptime(sv); - return (tcp_tv_to_hptstick(sv)); + return ((sv->tv_sec * 100000) + (sv->tv_usec / HPTS_TICKS_PER_SLOT)); } static __inline uint32_t -tcp_tv_to_usectick(struct timeval *sv) +tcp_tv_to_usectick(const struct timeval *sv) { return ((uint32_t) ((sv->tv_sec * HPTS_USEC_IN_SEC) + sv->tv_usec)); } static __inline uint32_t -tcp_tv_to_mssectick(struct timeval *sv) +tcp_tv_to_mssectick(const struct timeval *sv) { return ((uint32_t) ((sv->tv_sec * HPTS_MSEC_IN_SEC) + (sv->tv_usec/HPTS_USEC_IN_MSEC))); } +static __inline uint64_t +tcp_tv_to_lusectick(const struct timeval *sv) +{ + return ((uint64_t)((sv->tv_sec * HPTS_USEC_IN_SEC) + sv->tv_usec)); +} + +#ifdef _KERNEL + static __inline void tcp_hpts_unlock(struct tcp_hpts_entry *hpts) { mtx_unlock(&hpts->p_mtx); } +static __inline uint32_t +tcp_gethptstick(struct timeval *sv) +{ + struct timeval tv; + + if (sv == NULL) + sv = &tv; + microuptime(sv); + return (tcp_tv_to_hptstick(sv)); +} + static __inline uint32_t tcp_get_usecs(struct timeval *tv) { struct timeval tvd; if (tv == NULL) tv = &tvd; microuptime(tv); return (tcp_tv_to_usectick(tv)); } #endif /* _KERNEL */ #endif /* __tcp_hpts_h__ */ diff --git a/sys/netinet/tcp_lro.c b/sys/netinet/tcp_lro.c index ea84ca191eca..4e1480f2c002 100644 --- a/sys/netinet/tcp_lro.c +++ b/sys/netinet/tcp_lro.c @@ -1,1882 +1,1904 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2007, Myricom Inc. * Copyright (c) 2008, Intel Corporation. * Copyright (c) 2012 The FreeBSD Foundation * Copyright (c) 2016-2021 Mellanox Technologies. * All rights reserved. * * Portions of this software were developed by Bjoern Zeeb * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures"); #define TCP_LRO_TS_OPTION \ ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \ (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP) static void tcp_lro_rx_done(struct lro_ctrl *lc); static int tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash); #ifdef TCPHPTS static bool do_bpf_strip_and_compress(struct inpcb *, struct lro_ctrl *, struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **, bool *, bool); #endif SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TCP LRO"); static long tcplro_stacks_wanting_mbufq; counter_u64_t tcp_inp_lro_direct_queue; counter_u64_t tcp_inp_lro_wokeup_queue; counter_u64_t tcp_inp_lro_compressed; counter_u64_t tcp_inp_lro_locks_taken; counter_u64_t tcp_extra_mbuf; counter_u64_t tcp_would_have_but; counter_u64_t tcp_comp_total; counter_u64_t tcp_uncomp_total; static unsigned tcp_lro_entries = TCP_LRO_ENTRIES; SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries, CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0, "default number of LRO entries"); +static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH; +SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold, + CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0, + "Number of interrups in a row on the same CPU that will make us declare an 'affinity' cpu?"); + SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD, &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD, &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD, &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD, &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD, &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD, &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD, &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD, &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP"); void tcp_lro_reg_mbufq(void) { atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1); } void tcp_lro_dereg_mbufq(void) { atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1); } static __inline void tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket, struct lro_entry *le) { LIST_INSERT_HEAD(&lc->lro_active, le, next); LIST_INSERT_HEAD(bucket, le, hash_next); } static __inline void tcp_lro_active_remove(struct lro_entry *le) { LIST_REMOVE(le, next); /* active list */ LIST_REMOVE(le, hash_next); /* hash bucket */ } int tcp_lro_init(struct lro_ctrl *lc) { return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0)); } int tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp, unsigned lro_entries, unsigned lro_mbufs) { struct lro_entry *le; size_t size; unsigned i, elements; lc->lro_bad_csum = 0; lc->lro_queued = 0; lc->lro_flushed = 0; lc->lro_mbuf_count = 0; lc->lro_mbuf_max = lro_mbufs; lc->lro_cnt = lro_entries; lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX; lc->lro_length_lim = TCP_LRO_LENGTH_MAX; lc->ifp = ifp; LIST_INIT(&lc->lro_free); LIST_INIT(&lc->lro_active); /* create hash table to accelerate entry lookup */ if (lro_entries > lro_mbufs) elements = lro_entries; else elements = lro_mbufs; lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz, HASH_NOWAIT); if (lc->lro_hash == NULL) { memset(lc, 0, sizeof(*lc)); return (ENOMEM); } /* compute size to allocate */ size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) + (lro_entries * sizeof(*le)); lc->lro_mbuf_data = (struct lro_mbuf_sort *) malloc(size, M_LRO, M_NOWAIT | M_ZERO); /* check for out of memory */ if (lc->lro_mbuf_data == NULL) { free(lc->lro_hash, M_LRO); memset(lc, 0, sizeof(*lc)); return (ENOMEM); } /* compute offset for LRO entries */ le = (struct lro_entry *) (lc->lro_mbuf_data + lro_mbufs); /* setup linked list */ for (i = 0; i != lro_entries; i++) LIST_INSERT_HEAD(&lc->lro_free, le + i, next); return (0); } struct vxlan_header { uint32_t vxlh_flags; uint32_t vxlh_vni; }; static inline void * tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan) { const struct ether_vlan_header *eh; void *old; uint16_t eth_type; if (update_data) memset(parser, 0, sizeof(*parser)); old = ptr; if (is_vxlan) { const struct vxlan_header *vxh; vxh = ptr; ptr = (uint8_t *)ptr + sizeof(*vxh); if (update_data) { parser->data.vxlan_vni = vxh->vxlh_vni & htonl(0xffffff00); } } eh = ptr; if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) { eth_type = eh->evl_proto; if (update_data) { /* strip priority and keep VLAN ID only */ parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK); } /* advance to next header */ ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; } else { eth_type = eh->evl_encap_proto; /* advance to next header */ ptr = (uint8_t *)ptr + ETHER_HDR_LEN; } switch (eth_type) { #ifdef INET case htons(ETHERTYPE_IP): parser->ip4 = ptr; /* Ensure there are no IPv4 options. */ if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4)) break; /* .. and the packet is not fragmented. */ if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK)) break; ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2); if (update_data) { parser->data.s_addr.v4 = parser->ip4->ip_src; parser->data.d_addr.v4 = parser->ip4->ip_dst; } switch (parser->ip4->ip_p) { case IPPROTO_UDP: parser->udp = ptr; if (update_data) { parser->data.lro_type = LRO_TYPE_IPV4_UDP; parser->data.s_port = parser->udp->uh_sport; parser->data.d_port = parser->udp->uh_dport; } else { MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP); } ptr = ((uint8_t *)ptr + sizeof(*parser->udp)); parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; return (ptr); case IPPROTO_TCP: parser->tcp = ptr; if (update_data) { parser->data.lro_type = LRO_TYPE_IPV4_TCP; parser->data.s_port = parser->tcp->th_sport; parser->data.d_port = parser->tcp->th_dport; } else { MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP); } ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2); parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; return (ptr); default: break; } break; #endif #ifdef INET6 case htons(ETHERTYPE_IPV6): parser->ip6 = ptr; ptr = (uint8_t *)ptr + sizeof(*parser->ip6); if (update_data) { parser->data.s_addr.v6 = parser->ip6->ip6_src; parser->data.d_addr.v6 = parser->ip6->ip6_dst; } switch (parser->ip6->ip6_nxt) { case IPPROTO_UDP: parser->udp = ptr; if (update_data) { parser->data.lro_type = LRO_TYPE_IPV6_UDP; parser->data.s_port = parser->udp->uh_sport; parser->data.d_port = parser->udp->uh_dport; } else { MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP); } ptr = (uint8_t *)ptr + sizeof(*parser->udp); parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; return (ptr); case IPPROTO_TCP: parser->tcp = ptr; if (update_data) { parser->data.lro_type = LRO_TYPE_IPV6_TCP; parser->data.s_port = parser->tcp->th_sport; parser->data.d_port = parser->tcp->th_dport; } else { MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP); } ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2); parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; return (ptr); default: break; } break; #endif default: break; } /* Invalid packet - cannot parse */ return (NULL); } static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID | CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID; static inline struct lro_parser * tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data) { void *data_ptr; /* Try to parse outer headers first. */ data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false); if (data_ptr == NULL || po->total_hdr_len > m->m_len) return (NULL); if (update_data) { /* Store VLAN ID, if any. */ if (__predict_false(m->m_flags & M_VLANTAG)) { po->data.vlan_id = htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK); } } switch (po->data.lro_type) { case LRO_TYPE_IPV4_UDP: case LRO_TYPE_IPV6_UDP: /* Check for VXLAN headers. */ if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum) break; /* Try to parse inner headers. */ data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true); if (data_ptr == NULL || pi->total_hdr_len > m->m_len) break; /* Verify supported header types. */ switch (pi->data.lro_type) { case LRO_TYPE_IPV4_TCP: case LRO_TYPE_IPV6_TCP: return (pi); default: break; } break; case LRO_TYPE_IPV4_TCP: case LRO_TYPE_IPV6_TCP: if (update_data) memset(pi, 0, sizeof(*pi)); return (po); default: break; } return (NULL); } static inline int tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po) { int len; switch (po->data.lro_type) { #ifdef INET case LRO_TYPE_IPV4_TCP: len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) + ntohs(po->ip4->ip_len); break; #endif #ifdef INET6 case LRO_TYPE_IPV6_TCP: len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) + ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6); break; #endif default: return (TCP_LRO_CANNOT); } /* * If the frame is padded beyond the end of the IP packet, * then trim the extra bytes off: */ if (__predict_true(m->m_pkthdr.len == len)) { return (0); } else if (m->m_pkthdr.len > len) { m_adj(m, len - m->m_pkthdr.len); return (0); } return (TCP_LRO_CANNOT); } static struct tcphdr * tcp_lro_get_th(struct mbuf *m) { return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off)); } static void lro_free_mbuf_chain(struct mbuf *m) { struct mbuf *save; while (m) { save = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); m = save; } } void tcp_lro_free(struct lro_ctrl *lc) { struct lro_entry *le; unsigned x; /* reset LRO free list */ LIST_INIT(&lc->lro_free); /* free active mbufs, if any */ while ((le = LIST_FIRST(&lc->lro_active)) != NULL) { tcp_lro_active_remove(le); lro_free_mbuf_chain(le->m_head); } /* free hash table */ free(lc->lro_hash, M_LRO); lc->lro_hash = NULL; lc->lro_hashsz = 0; /* free mbuf array, if any */ for (x = 0; x != lc->lro_mbuf_count; x++) m_freem(lc->lro_mbuf_data[x].mb); lc->lro_mbuf_count = 0; /* free allocated memory, if any */ free(lc->lro_mbuf_data, M_LRO); lc->lro_mbuf_data = NULL; } static uint16_t tcp_lro_rx_csum_tcphdr(const struct tcphdr *th) { const uint16_t *ptr; uint32_t csum; uint16_t len; csum = -th->th_sum; /* exclude checksum field */ len = th->th_off; ptr = (const uint16_t *)th; while (len--) { csum += *ptr; ptr++; csum += *ptr; ptr++; } while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); return (csum); } static uint16_t tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum) { uint32_t c; uint16_t cs; c = tcp_csum; switch (pa->data.lro_type) { #ifdef INET6 case LRO_TYPE_IPV6_TCP: /* Compute full pseudo IPv6 header checksum. */ cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0); break; #endif #ifdef INET case LRO_TYPE_IPV4_TCP: /* Compute full pseudo IPv4 header checsum. */ cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP); cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs)); break; #endif default: cs = 0; /* Keep compiler happy. */ break; } /* Complement checksum. */ cs = ~cs; c += cs; /* Remove TCP header checksum. */ cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp); c += cs; /* Compute checksum remainder. */ while (c > 0xffff) c = (c >> 16) + (c & 0xffff); return (c); } static void tcp_lro_rx_done(struct lro_ctrl *lc) { struct lro_entry *le; while ((le = LIST_FIRST(&lc->lro_active)) != NULL) { tcp_lro_active_remove(le); tcp_lro_flush(lc, le); } } void tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout) { struct lro_entry *le, *le_tmp; uint64_t now, tov; struct bintime bt; if (LIST_EMPTY(&lc->lro_active)) return; /* get timeout time and current time in ns */ binuptime(&bt); now = bintime2ns(&bt); tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000)); LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) { if (now >= (bintime2ns(&le->alloc_time) + tov)) { tcp_lro_active_remove(le); tcp_lro_flush(lc, le); } } } #ifdef INET static int tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4) { uint16_t csum; /* Legacy IP has a header checksum that needs to be correct. */ if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) { if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) { lc->lro_bad_csum++; return (TCP_LRO_CANNOT); } } else { csum = in_cksum_hdr(ip4); if (__predict_false(csum != 0)) { lc->lro_bad_csum++; return (TCP_LRO_CANNOT); } } return (0); } #endif #ifdef TCPHPTS static void tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc, const struct lro_entry *le, const struct mbuf *m, int frm, int32_t tcp_data_len, uint32_t th_seq, uint32_t th_ack, uint16_t th_win) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv, btv; uint32_t cts; cts = tcp_get_usecs(&tv); memset(&log, 0, sizeof(union tcp_log_stackspecific)); log.u_bbr.flex8 = frm; log.u_bbr.flex1 = tcp_data_len; if (m) log.u_bbr.flex2 = m->m_pkthdr.len; else log.u_bbr.flex2 = 0; log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs; log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len; if (le->m_head) { log.u_bbr.flex5 = le->m_head->m_pkthdr.len; log.u_bbr.delRate = le->m_head->m_flags; log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp; } log.u_bbr.inflight = th_seq; + log.u_bbr.delivered = th_ack; log.u_bbr.timeStamp = cts; log.u_bbr.epoch = le->next_seq; - log.u_bbr.delivered = th_ack; log.u_bbr.lt_epoch = le->ack_seq; log.u_bbr.pacing_gain = th_win; log.u_bbr.cwnd_gain = le->window; + log.u_bbr.lost = curcpu; log.u_bbr.cur_del_rate = (uintptr_t)m; log.u_bbr.bw_inuse = (uintptr_t)le->m_head; bintime2timeval(&lc->lro_last_queue_time, &btv); log.u_bbr.flex6 = tcp_tv_to_usectick(&btv); log.u_bbr.flex7 = le->compressed; log.u_bbr.pacing_gain = le->uncompressed; if (in_epoch(net_epoch_preempt)) log.u_bbr.inhpts = 1; else log.u_bbr.inhpts = 0; TCP_LOG_EVENTP(tp, NULL, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_LRO, 0, 0, &log, false, &tv); } } #endif static inline void tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum) { uint32_t csum; csum = 0xffff - *ptr + value; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); *ptr = value; *psum = csum; } static uint16_t tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le, uint16_t payload_len, uint16_t delta_sum) { uint32_t csum; uint16_t tlen; uint16_t temp[5] = {}; switch (pa->data.lro_type) { case LRO_TYPE_IPV4_TCP: /* Compute new IPv4 length. */ tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len; tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]); /* Subtract delta from current IPv4 checksum. */ csum = pa->ip4->ip_sum + 0xffff - temp[0]; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]); goto update_tcp_header; case LRO_TYPE_IPV6_TCP: /* Compute new IPv6 length. */ tlen = (pa->tcp->th_off << 2) + payload_len; tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]); goto update_tcp_header; case LRO_TYPE_IPV4_UDP: /* Compute new IPv4 length. */ tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len; tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]); /* Subtract delta from current IPv4 checksum. */ csum = pa->ip4->ip_sum + 0xffff - temp[0]; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]); goto update_udp_header; case LRO_TYPE_IPV6_UDP: /* Compute new IPv6 length. */ tlen = sizeof(*pa->udp) + payload_len; tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]); goto update_udp_header; default: return (0); } update_tcp_header: /* Compute current TCP header checksum. */ temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp); /* Incorporate the latest ACK into the TCP header. */ pa->tcp->th_ack = le->ack_seq; pa->tcp->th_win = le->window; /* Incorporate latest timestamp into the TCP header. */ if (le->timestamp != 0) { uint32_t *ts_ptr; ts_ptr = (uint32_t *)(pa->tcp + 1); ts_ptr[1] = htonl(le->tsval); ts_ptr[2] = le->tsecr; } /* Compute new TCP header checksum. */ temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp); /* Compute new TCP checksum. */ csum = pa->tcp->th_sum + 0xffff - delta_sum + 0xffff - temp[0] + 0xffff - temp[3] + temp[2]; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); /* Assign new TCP checksum. */ tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]); /* Compute all modififications affecting next checksum. */ csum = temp[0] + temp[1] + 0xffff - temp[2] + temp[3] + temp[4] + delta_sum; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); /* Return delta checksum to next stage, if any. */ return (csum); update_udp_header: tlen = sizeof(*pa->udp) + payload_len; /* Assign new UDP length and compute checksum delta. */ tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]); /* Check if there is a UDP checksum. */ if (__predict_false(pa->udp->uh_sum != 0)) { /* Compute new UDP checksum. */ csum = pa->udp->uh_sum + 0xffff - delta_sum + 0xffff - temp[0] + 0xffff - temp[2]; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); /* Assign new UDP checksum. */ tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]); } /* Compute all modififications affecting next checksum. */ csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); /* Return delta checksum to next stage, if any. */ return (csum); } static void tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le) { /* Check if we need to recompute any checksums. */ if (le->m_head->m_pkthdr.lro_nsegs > 1) { uint16_t csum; switch (le->inner.data.lro_type) { case LRO_TYPE_IPV4_TCP: csum = tcp_lro_update_checksum(&le->inner, le, le->m_head->m_pkthdr.lro_tcp_d_len, le->m_head->m_pkthdr.lro_tcp_d_csum); csum = tcp_lro_update_checksum(&le->outer, NULL, le->m_head->m_pkthdr.lro_tcp_d_len + le->inner.total_hdr_len, csum); le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID; le->m_head->m_pkthdr.csum_data = 0xffff; break; case LRO_TYPE_IPV6_TCP: csum = tcp_lro_update_checksum(&le->inner, le, le->m_head->m_pkthdr.lro_tcp_d_len, le->m_head->m_pkthdr.lro_tcp_d_csum); csum = tcp_lro_update_checksum(&le->outer, NULL, le->m_head->m_pkthdr.lro_tcp_d_len + le->inner.total_hdr_len, csum); le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | CSUM_PSEUDO_HDR; le->m_head->m_pkthdr.csum_data = 0xffff; break; case LRO_TYPE_NONE: switch (le->outer.data.lro_type) { case LRO_TYPE_IPV4_TCP: csum = tcp_lro_update_checksum(&le->outer, le, le->m_head->m_pkthdr.lro_tcp_d_len, le->m_head->m_pkthdr.lro_tcp_d_csum); le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID; le->m_head->m_pkthdr.csum_data = 0xffff; break; case LRO_TYPE_IPV6_TCP: csum = tcp_lro_update_checksum(&le->outer, le, le->m_head->m_pkthdr.lro_tcp_d_len, le->m_head->m_pkthdr.lro_tcp_d_csum); le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | CSUM_PSEUDO_HDR; le->m_head->m_pkthdr.csum_data = 0xffff; break; default: break; } break; default: break; } } /* * Break any chain, this is not set to NULL on the singleton * case m_nextpkt points to m_head. Other case set them * m_nextpkt to NULL in push_and_replace. */ le->m_head->m_nextpkt = NULL; lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs; (*lc->ifp->if_input)(lc->ifp, le->m_head); } static void tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m, struct tcphdr *th) { uint32_t *ts_ptr; uint16_t tcp_data_len; uint16_t tcp_opt_len; ts_ptr = (uint32_t *)(th + 1); tcp_opt_len = (th->th_off << 2); tcp_opt_len -= sizeof(*th); /* Check if there is a timestamp option. */ if (tcp_opt_len == 0 || __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA || *ts_ptr != TCP_LRO_TS_OPTION)) { /* We failed to find the timestamp option. */ le->timestamp = 0; } else { le->timestamp = 1; le->tsval = ntohl(*(ts_ptr + 1)); le->tsecr = *(ts_ptr + 2); } tcp_data_len = m->m_pkthdr.lro_tcp_d_len; /* Pull out TCP sequence numbers and window size. */ le->next_seq = ntohl(th->th_seq) + tcp_data_len; le->ack_seq = th->th_ack; le->window = th->th_win; /* Setup new data pointers. */ le->m_head = m; le->m_tail = m_last(m); } static void tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m) { struct lro_parser *pa; /* * Push up the stack of the current entry * and replace it with "m". */ struct mbuf *msave; /* Grab off the next and save it */ msave = le->m_head->m_nextpkt; le->m_head->m_nextpkt = NULL; /* Now push out the old entry */ tcp_flush_out_entry(lc, le); /* Re-parse new header, should not fail. */ pa = tcp_lro_parser(m, &le->outer, &le->inner, false); KASSERT(pa != NULL, ("tcp_push_and_replace: LRO parser failed on m=%p\n", m)); /* * Now to replace the data properly in the entry * we have to reset the TCP header and * other fields. */ tcp_set_entry_to_mbuf(lc, le, m, pa->tcp); /* Restore the next list */ m->m_nextpkt = msave; } static void tcp_lro_mbuf_append_pkthdr(struct mbuf *m, const struct mbuf *p) { uint32_t csum; if (m->m_pkthdr.lro_nsegs == 1) { /* Compute relative checksum. */ csum = p->m_pkthdr.lro_tcp_d_csum; } else { /* Merge TCP data checksums. */ csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum + (uint32_t)p->m_pkthdr.lro_tcp_d_csum; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); } /* Update various counters. */ m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len; m->m_pkthdr.lro_tcp_d_csum = csum; m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len; m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs; } static void tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le) { /* * Walk through the mbuf chain we * have on tap and compress/condense * as required. */ uint32_t *ts_ptr; struct mbuf *m; struct tcphdr *th; uint32_t tcp_data_len_total; uint32_t tcp_data_seg_total; uint16_t tcp_data_len; uint16_t tcp_opt_len; /* * First we must check the lead (m_head) * we must make sure that it is *not* * something that should be sent up * right away (sack etc). */ again: m = le->m_head->m_nextpkt; if (m == NULL) { /* Just one left. */ return; } th = tcp_lro_get_th(m); tcp_opt_len = (th->th_off << 2); tcp_opt_len -= sizeof(*th); ts_ptr = (uint32_t *)(th + 1); if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA || *ts_ptr != TCP_LRO_TS_OPTION)) { /* * Its not the timestamp. We can't * use this guy as the head. */ le->m_head->m_nextpkt = m->m_nextpkt; tcp_push_and_replace(lc, le, m); goto again; } if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) { /* * Make sure that previously seen segements/ACKs are delivered * before this segment, e.g. FIN. */ le->m_head->m_nextpkt = m->m_nextpkt; tcp_push_and_replace(lc, le, m); goto again; } while((m = le->m_head->m_nextpkt) != NULL) { /* * condense m into le, first * pull m out of the list. */ le->m_head->m_nextpkt = m->m_nextpkt; m->m_nextpkt = NULL; /* Setup my data */ tcp_data_len = m->m_pkthdr.lro_tcp_d_len; th = tcp_lro_get_th(m); ts_ptr = (uint32_t *)(th + 1); tcp_opt_len = (th->th_off << 2); tcp_opt_len -= sizeof(*th); tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len; tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs; if (tcp_data_seg_total >= lc->lro_ackcnt_lim || tcp_data_len_total >= lc->lro_length_lim) { /* Flush now if appending will result in overflow. */ tcp_push_and_replace(lc, le, m); goto again; } if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA || *ts_ptr != TCP_LRO_TS_OPTION)) { /* * Maybe a sack in the new one? We need to * start all over after flushing the * current le. We will go up to the beginning * and flush it (calling the replace again possibly * or just returning). */ tcp_push_and_replace(lc, le, m); goto again; } if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) { tcp_push_and_replace(lc, le, m); goto again; } if (tcp_opt_len != 0) { uint32_t tsval = ntohl(*(ts_ptr + 1)); /* Make sure timestamp values are increasing. */ if (TSTMP_GT(le->tsval, tsval)) { tcp_push_and_replace(lc, le, m); goto again; } le->tsval = tsval; le->tsecr = *(ts_ptr + 2); } /* Try to append the new segment. */ if (__predict_false(ntohl(th->th_seq) != le->next_seq || (tcp_data_len == 0 && le->ack_seq == th->th_ack && le->window == th->th_win))) { /* Out of order packet or duplicate ACK. */ tcp_push_and_replace(lc, le, m); goto again; } if (tcp_data_len != 0 || SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) { le->next_seq += tcp_data_len; le->ack_seq = th->th_ack; le->window = th->th_win; } else if (th->th_ack == le->ack_seq) { le->window = WIN_MAX(le->window, th->th_win); } if (tcp_data_len == 0) { m_freem(m); continue; } /* Merge TCP data checksum and length to head mbuf. */ tcp_lro_mbuf_append_pkthdr(le->m_head, m); /* * Adjust the mbuf so that m_data points to the first byte of * the ULP payload. Adjust the mbuf to avoid complications and * append new segment to existing mbuf chain. */ m_adj(m, m->m_pkthdr.len - tcp_data_len); m_demote_pkthdr(m); le->m_tail->m_next = m; le->m_tail = m_last(m); } } #ifdef TCPHPTS static void tcp_queue_pkts(struct inpcb *inp, struct tcpcb *tp, struct lro_entry *le) { INP_WLOCK_ASSERT(inp); if (tp->t_in_pkt == NULL) { /* Nothing yet there */ tp->t_in_pkt = le->m_head; tp->t_tail_pkt = le->m_last_mbuf; } else { /* Already some there */ tp->t_tail_pkt->m_nextpkt = le->m_head; tp->t_tail_pkt = le->m_last_mbuf; } le->m_head = NULL; le->m_last_mbuf = NULL; } static struct mbuf * tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le, struct inpcb *inp, int32_t *new_m) { struct tcpcb *tp; struct mbuf *m; tp = intotcpcb(inp); if (__predict_false(tp == NULL)) return (NULL); /* Look at the last mbuf if any in queue */ m = tp->t_tail_pkt; if (m != NULL && (m->m_flags & M_ACKCMP) != 0) { if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) { tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0); *new_m = 0; counter_u64_add(tcp_extra_mbuf, 1); return (m); } else { /* Mark we ran out of space */ inp->inp_flags2 |= INP_MBUF_L_ACKS; } } /* Decide mbuf size. */ if (inp->inp_flags2 & INP_MBUF_L_ACKS) m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR); else m = m_gethdr(M_NOWAIT, MT_DATA); if (__predict_false(m == NULL)) { counter_u64_add(tcp_would_have_but, 1); return (NULL); } counter_u64_add(tcp_comp_total, 1); m->m_flags |= M_ACKCMP; *new_m = 1; return (m); } static struct inpcb * tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa) { struct inpcb *inp; NET_EPOCH_ASSERT(); switch (pa->data.lro_type) { #ifdef INET6 case LRO_TYPE_IPV6_TCP: inp = in6_pcblookup(&V_tcbinfo, &pa->data.s_addr.v6, pa->data.s_port, &pa->data.d_addr.v6, pa->data.d_port, INPLOOKUP_WLOCKPCB, ifp); break; #endif #ifdef INET case LRO_TYPE_IPV4_TCP: inp = in_pcblookup(&V_tcbinfo, pa->data.s_addr.v4, pa->data.s_port, pa->data.d_addr.v4, pa->data.d_port, INPLOOKUP_WLOCKPCB, ifp); break; #endif default: inp = NULL; break; } return (inp); } static inline bool tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts) { /* * This function returns two bits of valuable information. * a) Is what is present capable of being ack-compressed, * we can ack-compress if there is no options or just * a timestamp option, and of course the th_flags must * be correct as well. * b) Our other options present such as SACK. This is * used to determine if we want to wakeup or not. */ bool ret = true; switch (th->th_off << 2) { case (sizeof(*th) + TCPOLEN_TSTAMP_APPA): *ppts = (uint32_t *)(th + 1); /* Check if we have only one timestamp option. */ if (**ppts == TCP_LRO_TS_OPTION) *other_opts = false; else { *other_opts = true; ret = false; } break; case (sizeof(*th)): /* No options. */ *ppts = NULL; *other_opts = false; break; default: *ppts = NULL; *other_opts = true; ret = false; break; } /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */ if ((th->th_flags & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0) ret = false; /* If it has data on it we cannot compress it */ if (m->m_pkthdr.lro_tcp_d_len) ret = false; /* ACK flag must be set. */ if (!(th->th_flags & TH_ACK)) ret = false; return (ret); } static int tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le) { struct inpcb *inp; struct tcpcb *tp; struct mbuf **pp, *cmp, *mv_to; bool bpf_req, should_wake; /* Check if packet doesn't belongs to our network interface. */ if ((tcplro_stacks_wanting_mbufq == 0) || (le->outer.data.vlan_id != 0) || (le->inner.data.lro_type != LRO_TYPE_NONE)) return (TCP_LRO_CANNOT); #ifdef INET6 /* * Be proactive about unspecified IPv6 address in source. As * we use all-zero to indicate unbounded/unconnected pcb, * unspecified IPv6 address can be used to confuse us. * * Note that packets with unspecified IPv6 destination is * already dropped in ip6_input. */ if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP && IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6))) return (TCP_LRO_CANNOT); if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP && IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6))) return (TCP_LRO_CANNOT); #endif /* Lookup inp, if any. */ inp = tcp_lro_lookup(lc->ifp, (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner); if (inp == NULL) return (TCP_LRO_CANNOT); counter_u64_add(tcp_inp_lro_locks_taken, 1); /* Get TCP control structure. */ tp = intotcpcb(inp); /* Check if the inp is dead, Jim. */ if (tp == NULL || (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) || (inp->inp_flags2 & INP_FREED)) { INP_WUNLOCK(inp); return (TCP_LRO_CANNOT); } - + if ((inp->inp_irq_cpu_set == 0) && (lc->lro_cpu_is_set == 1)) { + inp->inp_irq_cpu = lc->lro_last_cpu; + inp->inp_irq_cpu_set = 1; + } /* Check if the transport doesn't support the needed optimizations. */ if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ | INP_MBUF_ACKCMP)) == 0) { INP_WUNLOCK(inp); return (TCP_LRO_CANNOT); } if (inp->inp_flags2 & INP_MBUF_QUEUE_READY) should_wake = false; else should_wake = true; /* Check if packets should be tapped to BPF. */ bpf_req = bpf_peers_present(lc->ifp->if_bpf); /* Strip and compress all the incoming packets. */ cmp = NULL; for (pp = &le->m_head; *pp != NULL; ) { mv_to = NULL; if (do_bpf_strip_and_compress(inp, lc, le, pp, &cmp, &mv_to, &should_wake, bpf_req ) == false) { /* Advance to next mbuf. */ pp = &(*pp)->m_nextpkt; } else if (mv_to != NULL) { /* We are asked to move pp up */ pp = &mv_to->m_nextpkt; } } /* Update "m_last_mbuf", if any. */ if (pp == &le->m_head) le->m_last_mbuf = *pp; else le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt); /* Check if any data mbufs left. */ if (le->m_head != NULL) { counter_u64_add(tcp_inp_lro_direct_queue, 1); tcp_lro_log(tp, lc, le, NULL, 22, 1, inp->inp_flags2, inp->inp_in_input, 1); tcp_queue_pkts(inp, tp, le); } if (should_wake) { /* Wakeup */ counter_u64_add(tcp_inp_lro_wokeup_queue, 1); if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0)) inp = NULL; } if (inp != NULL) INP_WUNLOCK(inp); return (0); /* Success. */ } #endif void tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le) { /* Only optimise if there are multiple packets waiting. */ #ifdef TCPHPTS int error; CURVNET_SET(lc->ifp->if_vnet); error = tcp_lro_flush_tcphpts(lc, le); CURVNET_RESTORE(); if (error != 0) { #endif tcp_lro_condense(lc, le); tcp_flush_out_entry(lc, le); #ifdef TCPHPTS } #endif lc->lro_flushed++; bzero(le, sizeof(*le)); LIST_INSERT_HEAD(&lc->lro_free, le, next); } #ifdef HAVE_INLINE_FLSLL #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1)) #else static inline uint64_t tcp_lro_msb_64(uint64_t x) { x |= (x >> 1); x |= (x >> 2); x |= (x >> 4); x |= (x >> 8); x |= (x >> 16); x |= (x >> 32); return (x & ~(x >> 1)); } #endif /* * The tcp_lro_sort() routine is comparable to qsort(), except it has * a worst case complexity limit of O(MIN(N,64)*N), where N is the * number of elements to sort and 64 is the number of sequence bits * available. The algorithm is bit-slicing the 64-bit sequence number, * sorting one bit at a time from the most significant bit until the * least significant one, skipping the constant bits. This is * typically called a radix sort. */ static void tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size) { struct lro_mbuf_sort temp; uint64_t ones; uint64_t zeros; uint32_t x; uint32_t y; repeat: /* for small arrays insertion sort is faster */ if (size <= 12) { for (x = 1; x < size; x++) { temp = parray[x]; for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--) parray[y] = parray[y - 1]; parray[y] = temp; } return; } /* compute sequence bits which are constant */ ones = 0; zeros = 0; for (x = 0; x != size; x++) { ones |= parray[x].seq; zeros |= ~parray[x].seq; } /* compute bits which are not constant into "ones" */ ones &= zeros; if (ones == 0) return; /* pick the most significant bit which is not constant */ ones = tcp_lro_msb_64(ones); /* * Move entries having cleared sequence bits to the beginning * of the array: */ for (x = y = 0; y != size; y++) { /* skip set bits */ if (parray[y].seq & ones) continue; /* swap entries */ temp = parray[x]; parray[x] = parray[y]; parray[y] = temp; x++; } KASSERT(x != 0 && x != size, ("Memory is corrupted\n")); /* sort zeros */ tcp_lro_sort(parray, x); /* sort ones */ parray += x; size -= x; goto repeat; } void tcp_lro_flush_all(struct lro_ctrl *lc) { uint64_t seq; uint64_t nseq; unsigned x; /* check if no mbufs to flush */ if (lc->lro_mbuf_count == 0) goto done; - + if (lc->lro_cpu_is_set == 0) { + if (lc->lro_last_cpu == curcpu) { + lc->lro_cnt_of_same_cpu++; + /* Have we reached the threshold to declare a cpu? */ + if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh) + lc->lro_cpu_is_set = 1; + } else { + lc->lro_last_cpu = curcpu; + lc->lro_cnt_of_same_cpu = 0; + } + } CURVNET_SET(lc->ifp->if_vnet); /* get current time */ binuptime(&lc->lro_last_queue_time); /* sort all mbufs according to stream */ tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count); /* input data into LRO engine, stream by stream */ seq = 0; for (x = 0; x != lc->lro_mbuf_count; x++) { struct mbuf *mb; /* get mbuf */ mb = lc->lro_mbuf_data[x].mb; /* get sequence number, masking away the packet index */ nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24); /* check for new stream */ if (seq != nseq) { seq = nseq; /* flush active streams */ tcp_lro_rx_done(lc); } /* add packet to LRO engine */ if (tcp_lro_rx_common(lc, mb, 0, false) != 0) { /* input packet to network layer */ (*lc->ifp->if_input)(lc->ifp, mb); lc->lro_queued++; lc->lro_flushed++; } } CURVNET_RESTORE(); done: /* flush active streams */ tcp_lro_rx_done(lc); +#ifdef TCPHPTS + tcp_run_hpts(); +#endif lc->lro_mbuf_count = 0; } #ifdef TCPHPTS static void build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m, uint32_t *ts_ptr, uint16_t iptos) { /* * Given a TCP ACK, summarize it down into the small TCP ACK * entry. */ ae->timestamp = m->m_pkthdr.rcv_tstmp; if (m->m_flags & M_TSTMP_LRO) ae->flags = TSTMP_LRO; else if (m->m_flags & M_TSTMP) ae->flags = TSTMP_HDWR; ae->seq = ntohl(th->th_seq); ae->ack = ntohl(th->th_ack); ae->flags |= th->th_flags; if (ts_ptr != NULL) { ae->ts_value = ntohl(ts_ptr[1]); ae->ts_echo = ntohl(ts_ptr[2]); ae->flags |= HAS_TSTMP; } ae->win = ntohs(th->th_win); ae->codepoint = iptos; } /* * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets * and strip all, but the IPv4/IPv6 header. */ static bool do_bpf_strip_and_compress(struct inpcb *inp, struct lro_ctrl *lc, struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to, bool *should_wake, bool bpf_req) { union { void *ptr; struct ip *ip4; struct ip6_hdr *ip6; } l3; struct mbuf *m; struct mbuf *nm; struct tcphdr *th; struct tcp_ackent *ack_ent; uint32_t *ts_ptr; int32_t n_mbuf; bool other_opts, can_compress; uint16_t lro_type; uint16_t iptos; int tcp_hdr_offset; int idx; /* Get current mbuf. */ m = *pp; /* Let the BPF see the packet */ if (__predict_false(bpf_req)) ETHER_BPF_MTAP(lc->ifp, m); tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off; lro_type = le->inner.data.lro_type; switch (lro_type) { case LRO_TYPE_NONE: lro_type = le->outer.data.lro_type; switch (lro_type) { case LRO_TYPE_IPV4_TCP: tcp_hdr_offset -= sizeof(*le->outer.ip4); m->m_pkthdr.lro_etype = ETHERTYPE_IP; break; case LRO_TYPE_IPV6_TCP: tcp_hdr_offset -= sizeof(*le->outer.ip6); m->m_pkthdr.lro_etype = ETHERTYPE_IPV6; break; default: goto compressed; } break; case LRO_TYPE_IPV4_TCP: tcp_hdr_offset -= sizeof(*le->outer.ip4); m->m_pkthdr.lro_etype = ETHERTYPE_IP; break; case LRO_TYPE_IPV6_TCP: tcp_hdr_offset -= sizeof(*le->outer.ip6); m->m_pkthdr.lro_etype = ETHERTYPE_IPV6; break; default: goto compressed; } MPASS(tcp_hdr_offset >= 0); m_adj(m, tcp_hdr_offset); m->m_flags |= M_LRO_EHDRSTRP; m->m_flags &= ~M_ACKCMP; m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset; th = tcp_lro_get_th(m); th->th_sum = 0; /* TCP checksum is valid. */ /* Check if ACK can be compressed */ can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts); /* Now lets look at the should wake states */ if ((other_opts == true) && ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) == 0)) { /* * If there are other options (SACK?) and the * tcp endpoint has not expressly told us it does * not care about SACKS, then we should wake up. */ *should_wake = true; } /* Is the ack compressable? */ if (can_compress == false) goto done; /* Does the TCP endpoint support ACK compression? */ if ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0) goto done; /* Lets get the TOS/traffic class field */ l3.ptr = mtod(m, void *); switch (lro_type) { case LRO_TYPE_IPV4_TCP: iptos = l3.ip4->ip_tos; break; case LRO_TYPE_IPV6_TCP: iptos = IPV6_TRAFFIC_CLASS(l3.ip6); break; default: iptos = 0; /* Keep compiler happy. */ break; } /* Now lets get space if we don't have some already */ if (*cmp == NULL) { new_one: nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, &n_mbuf); if (__predict_false(nm == NULL)) goto done; *cmp = nm; if (n_mbuf) { /* * Link in the new cmp ack to our in-order place, * first set our cmp ack's next to where we are. */ nm->m_nextpkt = m; (*pp) = nm; /* * Set it up so mv_to is advanced to our * compressed ack. This way the caller can * advance pp to the right place. */ *mv_to = nm; /* * Advance it here locally as well. */ pp = &nm->m_nextpkt; } } else { /* We have one already we are working on */ nm = *cmp; if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) { /* We ran out of space */ inp->inp_flags2 |= INP_MBUF_L_ACKS; goto new_one; } } MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent)); counter_u64_add(tcp_inp_lro_compressed, 1); le->compressed++; /* We can add in to the one on the tail */ ack_ent = mtod(nm, struct tcp_ackent *); idx = (nm->m_len / sizeof(struct tcp_ackent)); build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos); /* Bump the size of both pkt-hdr and len */ nm->m_len += sizeof(struct tcp_ackent); nm->m_pkthdr.len += sizeof(struct tcp_ackent); compressed: /* Advance to next mbuf before freeing. */ *pp = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); return (true); done: counter_u64_add(tcp_uncomp_total, 1); le->uncompressed++; return (false); } #endif static struct lro_head * tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser) { u_long hash; if (M_HASHTYPE_ISHASH(m)) { hash = m->m_pkthdr.flowid; } else { for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++) hash += parser->data.raw[i]; } return (&lc->lro_hash[hash % lc->lro_hashsz]); } static int tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash) { struct lro_parser pi; /* inner address data */ struct lro_parser po; /* outer address data */ struct lro_parser *pa; /* current parser for TCP stream */ struct lro_entry *le; struct lro_head *bucket; struct tcphdr *th; int tcp_data_len; int tcp_opt_len; int error; uint16_t tcp_data_sum; #ifdef INET /* Quickly decide if packet cannot be LRO'ed */ if (__predict_false(V_ipforwarding != 0)) return (TCP_LRO_CANNOT); #endif #ifdef INET6 /* Quickly decide if packet cannot be LRO'ed */ if (__predict_false(V_ip6_forwarding != 0)) return (TCP_LRO_CANNOT); #endif /* We expect a contiguous header [eh, ip, tcp]. */ pa = tcp_lro_parser(m, &po, &pi, true); if (__predict_false(pa == NULL)) return (TCP_LRO_NOT_SUPPORTED); /* We don't expect any padding. */ error = tcp_lro_trim_mbuf_chain(m, pa); if (__predict_false(error != 0)) return (error); #ifdef INET switch (pa->data.lro_type) { case LRO_TYPE_IPV4_TCP: error = tcp_lro_rx_ipv4(lc, m, pa->ip4); if (__predict_false(error != 0)) return (error); break; default: break; } #endif /* If no hardware or arrival stamp on the packet add timestamp */ if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) { m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time); m->m_flags |= M_TSTMP_LRO; } /* Get pointer to TCP header. */ th = pa->tcp; /* Don't process SYN packets. */ if (__predict_false(th->th_flags & TH_SYN)) return (TCP_LRO_CANNOT); /* Get total TCP header length and compute payload length. */ tcp_opt_len = (th->th_off << 2); tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th - (uint8_t *)m->m_data) - tcp_opt_len; tcp_opt_len -= sizeof(*th); /* Don't process invalid TCP headers. */ if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0)) return (TCP_LRO_CANNOT); /* Compute TCP data only checksum. */ if (tcp_data_len == 0) tcp_data_sum = 0; /* no data, no checksum */ else if (__predict_false(csum != 0)) tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum); else tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum); /* Save TCP info in mbuf. */ m->m_nextpkt = NULL; m->m_pkthdr.rcvif = lc->ifp; m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum; m->m_pkthdr.lro_tcp_d_len = tcp_data_len; m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data); m->m_pkthdr.lro_nsegs = 1; /* Get hash bucket. */ if (!use_hash) { bucket = &lc->lro_hash[0]; } else { bucket = tcp_lro_rx_get_bucket(lc, m, pa); } /* Try to find a matching previous segment. */ LIST_FOREACH(le, bucket, hash_next) { /* Compare addresses and ports. */ if (lro_address_compare(&po.data, &le->outer.data) == false || lro_address_compare(&pi.data, &le->inner.data) == false) continue; /* Check if no data and old ACK. */ if (tcp_data_len == 0 && SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) { m_freem(m); return (0); } /* Mark "m" in the last spot. */ le->m_last_mbuf->m_nextpkt = m; /* Now set the tail to "m". */ le->m_last_mbuf = m; return (0); } /* Try to find an empty slot. */ if (LIST_EMPTY(&lc->lro_free)) return (TCP_LRO_NO_ENTRIES); /* Start a new segment chain. */ le = LIST_FIRST(&lc->lro_free); LIST_REMOVE(le, next); tcp_lro_active_insert(lc, bucket, le); /* Make sure the headers are set. */ le->inner = pi; le->outer = po; /* Store time this entry was allocated. */ le->alloc_time = lc->lro_last_queue_time; tcp_set_entry_to_mbuf(lc, le, m, th); /* Now set the tail to "m". */ le->m_last_mbuf = m; return (0); } int tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum) { int error; /* get current time */ binuptime(&lc->lro_last_queue_time); CURVNET_SET(lc->ifp->if_vnet); error = tcp_lro_rx_common(lc, m, csum, true); CURVNET_RESTORE(); return (error); } void tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb) { /* sanity checks */ if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL || lc->lro_mbuf_max == 0)) { /* packet drop */ m_freem(mb); return; } /* check if packet is not LRO capable */ if (__predict_false(mb->m_pkthdr.csum_flags == 0 || (lc->ifp->if_capenable & IFCAP_LRO) == 0)) { /* input packet to network layer */ (*lc->ifp->if_input) (lc->ifp, mb); return; } /* create sequence number */ lc->lro_mbuf_data[lc->lro_mbuf_count].seq = (((uint64_t)M_HASHTYPE_GET(mb)) << 56) | (((uint64_t)mb->m_pkthdr.flowid) << 24) | ((uint64_t)lc->lro_mbuf_count); /* enter mbuf */ lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb; /* flush if array is full */ if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max)) tcp_lro_flush_all(lc); } /* end */ diff --git a/sys/netinet/tcp_lro.h b/sys/netinet/tcp_lro.h index 5ff15e2dc97e..75ab78c7fc6f 100644 --- a/sys/netinet/tcp_lro.h +++ b/sys/netinet/tcp_lro.h @@ -1,208 +1,216 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2006, Myricom Inc. * Copyright (c) 2008, Intel Corporation. * Copyright (c) 2016-2021 Mellanox Technologies. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #ifndef _TCP_LRO_H_ #define _TCP_LRO_H_ #include #ifndef TCP_LRO_ENTRIES /* Define default number of LRO entries per RX queue */ #define TCP_LRO_ENTRIES 8 #endif /* * Flags for ACK entry for compression * the bottom 8 bits has the th_flags. * LRO itself adds only the TSTMP flags * to indicate if either of the types * of timestamps are filled and the * HAS_TSTMP option to indicate if the * TCP timestamp option is valid. * * The other 5 flag bits are for processing * by a stack. * */ #define TSTMP_LRO 0x0100 #define TSTMP_HDWR 0x0200 #define HAS_TSTMP 0x0400 +/* + * Default number of interrupts on the same cpu in a row + * that will cause us to declare a "affinity cpu". + */ +#define TCP_LRO_CPU_DECLARATION_THRESH 50 struct inpcb; union lro_address { u_long raw[1]; struct { uint16_t lro_type; /* internal */ #define LRO_TYPE_NONE 0 #define LRO_TYPE_IPV4_TCP 1 #define LRO_TYPE_IPV6_TCP 2 #define LRO_TYPE_IPV4_UDP 3 #define LRO_TYPE_IPV6_UDP 4 uint16_t vlan_id; /* VLAN identifier */ uint16_t s_port; /* source TCP/UDP port */ uint16_t d_port; /* destination TCP/UDP port */ uint32_t vxlan_vni; /* VXLAN virtual network identifier */ union { #ifdef INET struct in_addr v4; #endif #ifdef INET6 struct in6_addr v6; #endif } s_addr; /* source IPv4/IPv6 address */ union { #ifdef INET struct in_addr v4; #endif #ifdef INET6 struct in6_addr v6; #endif } d_addr; /* destination IPv4/IPv6 address */ }; } __aligned(sizeof(u_long)); #define LRO_RAW_ADDRESS_MAX \ (sizeof(union lro_address) / sizeof(u_long)) /* Optimize address comparison by comparing one unsigned long at a time: */ static inline bool lro_address_compare(const union lro_address *pa, const union lro_address *pb) { if (pa->lro_type == LRO_TYPE_NONE && pb->lro_type == LRO_TYPE_NONE) { return (true); } else for (unsigned i = 0; i < LRO_RAW_ADDRESS_MAX; i++) { if (pa->raw[i] != pb->raw[i]) return (false); } return (true); } struct lro_parser { union lro_address data; union { uint8_t *l3; struct ip *ip4; struct ip6_hdr *ip6; }; union { uint8_t *l4; struct tcphdr *tcp; struct udphdr *udp; }; uint16_t total_hdr_len; }; /* This structure is zeroed frequently, try to keep it small. */ struct lro_entry { LIST_ENTRY(lro_entry) next; LIST_ENTRY(lro_entry) hash_next; struct mbuf *m_head; struct mbuf *m_tail; struct mbuf *m_last_mbuf; struct lro_parser outer; struct lro_parser inner; uint32_t next_seq; /* tcp_seq */ uint32_t ack_seq; /* tcp_seq */ uint32_t tsval; uint32_t tsecr; uint16_t compressed; uint16_t uncompressed; uint16_t window; uint16_t timestamp; /* flag, not a TCP hdr field. */ struct bintime alloc_time; /* time when entry was allocated */ }; LIST_HEAD(lro_head, lro_entry); struct lro_mbuf_sort { uint64_t seq; struct mbuf *mb; }; /* NB: This is part of driver structs. */ struct lro_ctrl { struct ifnet *ifp; struct lro_mbuf_sort *lro_mbuf_data; struct bintime lro_last_queue_time; /* last time data was queued */ uint64_t lro_queued; uint64_t lro_flushed; uint64_t lro_bad_csum; unsigned lro_cnt; unsigned lro_mbuf_count; unsigned lro_mbuf_max; unsigned short lro_ackcnt_lim; /* max # of aggregated ACKs */ + unsigned short lro_cpu; /* Guess at the cpu we have affinity too */ unsigned lro_length_lim; /* max len of aggregated data */ - u_long lro_hashsz; + uint32_t lro_last_cpu; + uint32_t lro_cnt_of_same_cpu; struct lro_head *lro_hash; struct lro_head lro_active; struct lro_head lro_free; + uint8_t lro_cpu_is_set; /* Flag to say its ok to set the CPU on the inp */ }; struct tcp_ackent { uint64_t timestamp; /* hardware or sofware timestamp, valid if TSTMP_LRO or TSTMP_HDRW set */ uint32_t seq; /* th_seq value */ uint32_t ack; /* th_ack value */ uint32_t ts_value; /* If ts option value, valid if HAS_TSTMP is set */ uint32_t ts_echo; /* If ts option echo, valid if HAS_TSTMP is set */ uint16_t win; /* TCP window */ uint16_t flags; /* Flags to say if TS is present and type of timestamp and th_flags */ uint8_t codepoint; /* IP level codepoint including ECN bits */ uint8_t ack_val_set; /* Classification of ack used by the stack */ uint8_t pad[2]; /* To 32 byte boundary */ }; /* We use two M_PROTO on the mbuf */ #define M_ACKCMP M_PROTO4 /* Indicates LRO is sending in a Ack-compression mbuf */ #define M_LRO_EHDRSTRP M_PROTO6 /* Indicates that LRO has stripped the etherenet header */ #define TCP_LRO_LENGTH_MAX (65535 - 255) /* safe value with room for outer headers */ #define TCP_LRO_ACKCNT_MAX 65535 /* unlimited */ int tcp_lro_init(struct lro_ctrl *); int tcp_lro_init_args(struct lro_ctrl *, struct ifnet *, unsigned, unsigned); void tcp_lro_free(struct lro_ctrl *); void tcp_lro_flush_inactive(struct lro_ctrl *, const struct timeval *); void tcp_lro_flush(struct lro_ctrl *, struct lro_entry *); void tcp_lro_flush_all(struct lro_ctrl *); int tcp_lro_rx(struct lro_ctrl *, struct mbuf *, uint32_t); void tcp_lro_queue_mbuf(struct lro_ctrl *, struct mbuf *); void tcp_lro_reg_mbufq(void); void tcp_lro_dereg_mbufq(void); #define TCP_LRO_NO_ENTRIES -2 #define TCP_LRO_CANNOT -1 #define TCP_LRO_NOT_SUPPORTED 1 #endif /* _TCP_LRO_H_ */ diff --git a/sys/netinet/tcp_stacks/bbr.c b/sys/netinet/tcp_stacks/bbr.c index f8cd0bec1c95..8969e4e47ba1 100644 --- a/sys/netinet/tcp_stacks/bbr.c +++ b/sys/netinet/tcp_stacks/bbr.c @@ -1,14915 +1,14915 @@ /*- * Copyright (c) 2016-2020 Netflix, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ /** * Author: Randall Stewart * This work is based on the ACM Queue paper * BBR - Congestion Based Congestion Control * and also numerous discussions with Neal, Yuchung and Van. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include "opt_ratelimit.h" #include #include #include #include #include #ifdef TCP_HHOOK #include #endif #include #include #include #include #include #include #include #ifdef STATS #include #include #include /* Must come after qmath.h and tree.h */ #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES /* for logging */ #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #define TCPOUTFLAGS #include #include #include #include #include #include #include #include #include #include #include #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #ifdef INET6 #include #endif #include #include #include #include #include #if defined(IPSEC) || defined(IPSEC_SUPPORT) #include #include #endif /* IPSEC */ #include #include #include #ifdef MAC #include #endif #include "sack_filter.h" #include "tcp_bbr.h" #include "rack_bbr_common.h" uma_zone_t bbr_zone; uma_zone_t bbr_pcb_zone; struct sysctl_ctx_list bbr_sysctl_ctx; struct sysctl_oid *bbr_sysctl_root; #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \ (tv) = (value); \ if ((u_long)(tv) < (u_long)(tvmin)) \ (tv) = (tvmin); \ if ((u_long)(tv) > (u_long)(tvmax)) \ (tv) = (tvmax); \ } while(0) /*#define BBR_INVARIANT 1*/ /* * initial window */ static uint32_t bbr_def_init_win = 10; static int32_t bbr_persist_min = 250000; /* 250ms */ static int32_t bbr_persist_max = 1000000; /* 1 Second */ static int32_t bbr_cwnd_may_shrink = 0; static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP; static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT; static int32_t bbr_hardware_pacing_limit = 8000; static int32_t bbr_quanta = 3; /* How much extra quanta do we get? */ static int32_t bbr_no_retran = 0; static int32_t bbr_error_base_paceout = 10000; /* usec to pace */ static int32_t bbr_max_net_error_cnt = 10; /* Should the following be dynamic too -- loss wise */ static int32_t bbr_rtt_gain_thresh = 0; /* Measurement controls */ static int32_t bbr_use_google_algo = 1; static int32_t bbr_ts_limiting = 1; static int32_t bbr_ts_can_raise = 0; static int32_t bbr_do_red = 600; static int32_t bbr_red_scale = 20000; static int32_t bbr_red_mul = 1; static int32_t bbr_red_div = 2; static int32_t bbr_red_growth_restrict = 1; static int32_t bbr_target_is_bbunit = 0; static int32_t bbr_drop_limit = 0; /* * How much gain do we need to see to * stay in startup? */ static int32_t bbr_marks_rxt_sack_passed = 0; static int32_t bbr_start_exit = 25; static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */ static int32_t bbr_startup_loss_thresh = 2000; /* 20.00% loss */ static int32_t bbr_hptsi_max_mul = 1; /* These two mul/div assure a min pacing */ static int32_t bbr_hptsi_max_div = 2; /* time, 0 means turned off. We need this * if we go back ever to where the pacer * has priority over timers. */ static int32_t bbr_policer_call_from_rack_to = 0; static int32_t bbr_policer_detection_enabled = 1; static int32_t bbr_min_measurements_req = 1; /* We need at least 2 * measurments before we are * "good" note that 2 == 1. * This is because we use a > * comparison. This means if * min_measure was 0, it takes * num-measures > min(0) and * you get 1 measurement and * you are good. Set to 1, you * have to have two * measurements (this is done * to prevent it from being ok * to have no measurements). */ static int32_t bbr_no_pacing_until = 4; static int32_t bbr_min_usec_delta = 20000; /* 20,000 usecs */ static int32_t bbr_min_peer_delta = 20; /* 20 units */ static int32_t bbr_delta_percent = 150; /* 15.0 % */ static int32_t bbr_target_cwnd_mult_limit = 8; /* * bbr_cwnd_min_val is the number of * segments we hold to in the RTT probe * state typically 4. */ static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS; static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS; static int32_t bbr_gain_to_target = 1; static int32_t bbr_gain_gets_extra_too = 1; /* * bbr_high_gain is the 2/ln(2) value we need * to double the sending rate in startup. This * is used for both cwnd and hptsi gain's. */ static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1; static int32_t bbr_use_lower_gain_in_startup = 1; /* thresholds for reduction on drain in sub-states/drain */ static int32_t bbr_drain_rtt = BBR_SRTT; static int32_t bbr_drain_floor = 88; static int32_t google_allow_early_out = 1; static int32_t google_consider_lost = 1; static int32_t bbr_drain_drop_mul = 4; static int32_t bbr_drain_drop_div = 5; static int32_t bbr_rand_ot = 50; static int32_t bbr_can_force_probertt = 0; static int32_t bbr_can_adjust_probertt = 1; static int32_t bbr_probertt_sets_rtt = 0; static int32_t bbr_can_use_ts_for_rtt = 1; static int32_t bbr_is_ratio = 0; static int32_t bbr_sub_drain_app_limit = 1; static int32_t bbr_prtt_slam_cwnd = 1; static int32_t bbr_sub_drain_slam_cwnd = 1; static int32_t bbr_slam_cwnd_in_main_drain = 1; static int32_t bbr_filter_len_sec = 6; /* How long does the rttProp filter * hold */ static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4); /* * bbr_drain_gain is the reverse of the high_gain * designed to drain back out the standing queue * that is formed in startup by causing a larger * hptsi gain and thus drainging the packets * in flight. */ static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885; static int32_t bbr_rttprobe_gain = 192; /* * The cwnd_gain is the default cwnd gain applied when * calculating a target cwnd. Note that the cwnd is * a secondary factor in the way BBR works (see the * paper and think about it, it will take some time). * Basically the hptsi_gain spreads the packets out * so you never get more than BDP to the peer even * if the cwnd is high. In our implemenation that * means in non-recovery/retransmission scenarios * cwnd will never be reached by the flight-size. */ static int32_t bbr_cwnd_gain = BBR_UNIT * 2; static int32_t bbr_tlp_type_to_use = BBR_SRTT; static int32_t bbr_delack_time = 100000; /* 100ms in useconds */ static int32_t bbr_sack_not_required = 0; /* set to one to allow non-sack to use bbr */ static int32_t bbr_initial_bw_bps = 62500; /* 500kbps in bytes ps */ static int32_t bbr_ignore_data_after_close = 1; static int16_t bbr_hptsi_gain[] = { (BBR_UNIT *5 / 4), (BBR_UNIT * 3 / 4), BBR_UNIT, BBR_UNIT, BBR_UNIT, BBR_UNIT, BBR_UNIT, BBR_UNIT }; int32_t bbr_use_rack_resend_cheat = 1; int32_t bbr_sends_full_iwnd = 1; #define BBR_HPTSI_GAIN_MAX 8 /* * The BBR module incorporates a number of * TCP ideas that have been put out into the IETF * over the last few years: * - Yuchung Cheng's RACK TCP (for which its named) that * will stop us using the number of dup acks and instead * use time as the gage of when we retransmit. * - Reorder Detection of RFC4737 and the Tail-Loss probe draft * of Dukkipati et.al. * - Van Jacobson's et.al BBR. * * RACK depends on SACK, so if an endpoint arrives that * cannot do SACK the state machine below will shuttle the * connection back to using the "default" TCP stack that is * in FreeBSD. * * To implement BBR and RACK the original TCP stack was first decomposed * into a functional state machine with individual states * for each of the possible TCP connection states. The do_segement * functions role in life is to mandate the connection supports SACK * initially and then assure that the RACK state matches the conenction * state before calling the states do_segment function. Data processing * of inbound segments also now happens in the hpts_do_segment in general * with only one exception. This is so we can keep the connection on * a single CPU. * * Each state is simplified due to the fact that the original do_segment * has been decomposed and we *know* what state we are in (no * switches on the state) and all tests for SACK are gone. This * greatly simplifies what each state does. * * TCP output is also over-written with a new version since it * must maintain the new rack scoreboard and has had hptsi * integrated as a requirment. Still todo is to eliminate the * use of the callout_() system and use the hpts for all * timers as well. */ static uint32_t bbr_rtt_probe_time = 200000; /* 200ms in micro seconds */ static uint32_t bbr_rtt_probe_cwndtarg = 4; /* How many mss's outstanding */ static const int32_t bbr_min_req_free = 2; /* The min we must have on the * free list */ static int32_t bbr_tlp_thresh = 1; static int32_t bbr_reorder_thresh = 2; static int32_t bbr_reorder_fade = 60000000; /* 0 - never fade, def * 60,000,000 - 60 seconds */ static int32_t bbr_pkt_delay = 1000; static int32_t bbr_min_to = 1000; /* Number of usec's minimum timeout */ static int32_t bbr_incr_timers = 1; static int32_t bbr_tlp_min = 10000; /* 10ms in usecs */ static int32_t bbr_delayed_ack_time = 200000; /* 200ms in usecs */ static int32_t bbr_exit_startup_at_loss = 1; /* * bbr_lt_bw_ratio is 1/8th * bbr_lt_bw_diff is < 4 Kbit/sec */ static uint64_t bbr_lt_bw_diff = 4000 / 8; /* In bytes per second */ static uint64_t bbr_lt_bw_ratio = 8; /* For 1/8th */ static uint32_t bbr_lt_bw_max_rtts = 48; /* How many rtt's do we use * the lt_bw for */ static uint32_t bbr_lt_intvl_min_rtts = 4; /* Min num of RTT's to measure * lt_bw */ static int32_t bbr_lt_intvl_fp = 0; /* False positive epoch diff */ static int32_t bbr_lt_loss_thresh = 196; /* Lost vs delivered % */ static int32_t bbr_lt_fd_thresh = 100; /* false detection % */ static int32_t bbr_verbose_logging = 0; /* * Currently regular tcp has a rto_min of 30ms * the backoff goes 12 times so that ends up * being a total of 122.850 seconds before a * connection is killed. */ static int32_t bbr_rto_min_ms = 30; /* 30ms same as main freebsd */ static int32_t bbr_rto_max_sec = 4; /* 4 seconds */ /****************************************************/ /* DEFAULT TSO SIZING (cpu performance impacting) */ /****************************************************/ /* What amount is our formula using to get TSO size */ static int32_t bbr_hptsi_per_second = 1000; /* * For hptsi under bbr_cross_over connections what is delay * target 7ms (in usec) combined with a seg_max of 2 * gets us close to identical google behavior in * TSO size selection (possibly more 1MSS sends). */ static int32_t bbr_hptsi_segments_delay_tar = 7000; /* Does pacing delay include overhead's in its time calculations? */ static int32_t bbr_include_enet_oh = 0; static int32_t bbr_include_ip_oh = 1; static int32_t bbr_include_tcp_oh = 1; static int32_t bbr_google_discount = 10; /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */ static int32_t bbr_state_is_pkt_epoch = 0; static int32_t bbr_state_drain_2_tar = 1; /* What is the max the 0 - bbr_cross_over MBPS TSO target * can reach using our delay target. Note that this * value becomes the floor for the cross over * algorithm. */ static int32_t bbr_hptsi_segments_max = 2; static int32_t bbr_hptsi_segments_floor = 1; static int32_t bbr_hptsi_utter_max = 0; /* What is the min the 0 - bbr_cross-over MBPS TSO target can be */ static int32_t bbr_hptsi_bytes_min = 1460; static int32_t bbr_all_get_min = 0; /* Cross over point from algo-a to algo-b */ static uint32_t bbr_cross_over = TWENTY_THREE_MBPS; /* Do we deal with our restart state? */ static int32_t bbr_uses_idle_restart = 0; static int32_t bbr_idle_restart_threshold = 100000; /* 100ms in useconds */ /* Do we allow hardware pacing? */ static int32_t bbr_allow_hdwr_pacing = 0; static int32_t bbr_hdwr_pace_adjust = 2; /* multipler when we calc the tso size */ static int32_t bbr_hdwr_pace_floor = 1; static int32_t bbr_hdwr_pacing_delay_cnt = 10; /****************************************************/ static int32_t bbr_resends_use_tso = 0; static int32_t bbr_tlp_max_resend = 2; static int32_t bbr_sack_block_limit = 128; #define BBR_MAX_STAT 19 counter_u64_t bbr_state_time[BBR_MAX_STAT]; counter_u64_t bbr_state_lost[BBR_MAX_STAT]; counter_u64_t bbr_state_resend[BBR_MAX_STAT]; counter_u64_t bbr_stat_arry[BBR_STAT_SIZE]; counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE]; counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE]; counter_u64_t bbr_flows_whdwr_pacing; counter_u64_t bbr_flows_nohdwr_pacing; counter_u64_t bbr_nohdwr_pacing_enobuf; counter_u64_t bbr_hdwr_pacing_enobuf; static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr); /* * Static defintions we need for forward declarations. */ static uint32_t bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw); static uint32_t bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain); static void bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win); static void bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses); static void bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line, int dolog); static uint32_t bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain); static void bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses); static uint32_t bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm); static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp); static uint32_t bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts); static void bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line); static void bbr_set_state_target(struct tcp_bbr *bbr, int line); static void bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line); static void bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line); static void tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts); static void bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts); static void bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt, uint32_t line, uint8_t is_start, uint16_t set); static struct bbr_sendmap * bbr_find_lowest_rsm(struct tcp_bbr *bbr); static __inline uint32_t bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type); static void bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which); static void bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, uint32_t thresh, uint32_t to); static void bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag); static void bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, uint32_t del_by, uint32_t cts, uint32_t sloton, uint32_t prev_delay); static void bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line); static void bbr_stop_all_timers(struct tcpcb *tp); static void bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts); static void bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts); static void bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts); static void bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod); static int bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp); static inline uint8_t bbr_state_val(struct tcp_bbr *bbr) { return(bbr->rc_bbr_substate); } static inline uint32_t get_min_cwnd(struct tcp_bbr *bbr) { int mss; mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED) return (bbr_cwnd_min_val_hs * mss); else return (bbr_cwnd_min_val * mss); } static uint32_t bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr) { uint64_t srtt, var; uint64_t ret_val; bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; if (tp->t_srtt == 0) { srtt = (uint64_t)BBR_INITIAL_RTO; var = 0; } else { srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT); } TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]), bbr_persist_min, bbr_persist_max); return ((uint32_t)ret_val); } static uint32_t bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * Start the FR timer, we do this based on getting the first one in * the rc_tmap. Note that if its NULL we must stop the timer. in all * events we need to stop the running timer (if its running) before * starting the new one. */ uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; int32_t idx; int32_t is_tlp_timer = 0; struct bbr_sendmap *rsm; if (bbr->rc_all_timers_stopped) { /* All timers have been stopped none are to run */ return (0); } if (bbr->rc_in_persist) { /* We can't start any timer in persists */ return (bbr_get_persists_timer_val(tp, bbr)); } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if ((rsm == NULL) || ((tp->t_flags & TF_SACK_PERMIT) == 0) || (tp->t_state < TCPS_ESTABLISHED)) { /* Nothing on the send map */ activate_rxt: if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { uint64_t tov; time_since_sent = 0; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (rsm) { idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = rsm->r_tim_lastsent[idx]; else tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(tstmp_touse, cts)) time_since_sent = cts - tstmp_touse; } bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; if (tp->t_srtt == 0) tov = BBR_INITIAL_RTO; else tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) + ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT); if (tp->t_rxtshift) tov *= tcp_backoff[tp->t_rxtshift]; if (tov > time_since_sent) tov -= time_since_sent; else tov = bbr->r_ctl.rc_min_to; TCPT_RANGESET_NOSLOP(to, tov, (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC), (bbr->rc_max_rto_sec * USECS_IN_SECOND)); bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to); return (to); } return (0); } if (rsm->r_flags & BBR_ACKED) { rsm = bbr_find_lowest_rsm(bbr); if (rsm == NULL) { /* No lowest? */ goto activate_rxt; } } /* Convert from ms to usecs */ if (rsm->r_flags & BBR_SACK_PASSED) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & BBR_HAS_FIN)) { /* * We don't start a bbr rack timer if all we have is * a FIN outstanding. */ goto activate_rxt; } srtt = bbr_get_rtt(bbr, BBR_RTT_RACK); thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm); idx = rsm->r_rtr_cnt - 1; exp = rsm->r_tim_lastsent[idx] + thresh; if (SEQ_GEQ(exp, cts)) { to = exp - cts; if (to < bbr->r_ctl.rc_min_to) { to = bbr->r_ctl.rc_min_to; } } else { to = bbr->r_ctl.rc_min_to; } } else { /* Ok we need to do a TLP not RACK */ if (bbr->rc_tlp_in_progress != 0) { /* * The previous send was a TLP. */ goto activate_rxt; } rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); if (rsm == NULL) { /* We found no rsm to TLP with. */ goto activate_rxt; } if (rsm->r_flags & BBR_HAS_FIN) { /* If its a FIN we don't do TLP */ rsm = NULL; goto activate_rxt; } time_since_sent = 0; idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = rsm->r_tim_lastsent[idx]; else tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(tstmp_touse, cts)) time_since_sent = cts - tstmp_touse; is_tlp_timer = 1; srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use); thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts); if (thresh > time_since_sent) to = thresh - time_since_sent; else to = bbr->r_ctl.rc_min_to; if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { /* * If the TLP time works out to larger than the max * RTO lets not do TLP.. just RTO. */ goto activate_rxt; } if ((bbr->rc_tlp_rtx_out == 1) && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) { /* * Second retransmit of the same TLP * lets not. */ bbr->rc_tlp_rtx_out = 0; goto activate_rxt; } if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) { /* * The tail is no longer the last one I did a probe * on */ bbr->r_ctl.rc_tlp_seg_send_cnt = 0; bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; } } if (is_tlp_timer == 0) { BBR_STAT_INC(bbr_to_arm_rack); bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; } else { bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to); if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { /* * We have exceeded how many times we can retran the * current TLP timer, switch to the RTO timer. */ goto activate_rxt; } else { BBR_STAT_INC(bbr_to_arm_tlp); bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; } } return (to); } static inline int32_t bbr_minseg(struct tcp_bbr *bbr) { return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options); } static void bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len) { struct inpcb *inp; struct hpts_diag diag; uint32_t delayed_ack = 0; uint32_t left = 0; uint32_t hpts_timeout; uint8_t stopped; int32_t delay_calc = 0; uint32_t prev_delay = 0; inp = tp->t_inpcb; if (inp->inp_in_hpts) { /* A previous call is already set up */ return; } if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { return; } stopped = bbr->rc_tmr_stopped; if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { left = bbr->r_ctl.rc_timer_exp - cts; } bbr->r_ctl.rc_hpts_flags = 0; bbr->r_ctl.rc_timer_exp = 0; prev_delay = bbr->r_ctl.rc_last_delay_val; if (bbr->r_ctl.rc_last_delay_val && (slot == 0)) { /* * If a previous pacer delay was in place we * are not coming from the output side (where * we calculate a delay, more likely a timer). */ slot = bbr->r_ctl.rc_last_delay_val; if (TSTMP_GT(cts, bbr->rc_pacer_started)) { /* Compensate for time passed */ delay_calc = cts - bbr->rc_pacer_started; if (delay_calc <= slot) slot -= delay_calc; } } /* Do we have early to make up for by pushing out the pacing time? */ if (bbr->r_agg_early_set) { bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2); slot += bbr->r_ctl.rc_agg_early; bbr->r_ctl.rc_agg_early = 0; bbr->r_agg_early_set = 0; } /* Are we running a total debt that needs to be compensated for? */ if (bbr->r_ctl.rc_hptsi_agg_delay) { if (slot > bbr->r_ctl.rc_hptsi_agg_delay) { /* We nuke the delay */ slot -= bbr->r_ctl.rc_hptsi_agg_delay; bbr->r_ctl.rc_hptsi_agg_delay = 0; } else { /* We nuke some of the delay, put in a minimal 100usecs */ bbr->r_ctl.rc_hptsi_agg_delay -= slot; bbr->r_ctl.rc_last_delay_val = slot = 100; } } bbr->r_ctl.rc_last_delay_val = slot; hpts_timeout = bbr_timer_start(tp, bbr, cts); if (tp->t_flags & TF_DELACK) { if (bbr->rc_in_persist == 0) { delayed_ack = bbr_delack_time; } else { /* * We are in persists and have * gotten a new data element. */ if (hpts_timeout > bbr_delack_time) { /* * Lets make the persists timer (which acks) * be the smaller of hpts_timeout and bbr_delack_time. */ hpts_timeout = bbr_delack_time; } } } if (delayed_ack && ((hpts_timeout == 0) || (delayed_ack < hpts_timeout))) { /* We need a Delayed ack timer */ bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; hpts_timeout = delayed_ack; } if (slot) { /* Mark that we have a pacing timer up */ BBR_STAT_INC(bbr_paced_segments); bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; } /* * If no timers are going to run and we will fall off thfe hptsi * wheel, we resort to a keep-alive timer if its configured. */ if ((hpts_timeout == 0) && (slot == 0)) { if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) { /* * Ok we have no timer (persists, rack, tlp, rxt or * del-ack), we don't have segments being paced. So * all that is left is the keepalive timer. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); } else { hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); } bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; } } if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { /* * RACK, TLP, persists and RXT timers all are restartable * based on actions input .. i.e we received a packet (ack * or sack) and that changes things (rw, or snd_una etc). * Thus we can restart them with a new value. For * keep-alive, delayed_ack we keep track of what was left * and restart the timer with a smaller value. */ if (left < hpts_timeout) hpts_timeout = left; } if (bbr->r_ctl.rc_incr_tmrs && slot && (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { /* * If configured to do so, and the timer is either * the TLP or RXT timer, we need to increase the timeout * by the pacing time. Consider the bottleneck at my * machine as an example, we are sending something * to start a TLP on. The last packet won't be emitted * fully until the pacing time (the bottleneck will hold * the data in place). Once the packet is emitted that * is when we want to start waiting for the TLP. This * is most evident with hardware pacing (where the nic * is holding the packet(s) before emitting). But it * can also show up in the network so we do it for all * cases. Technically we would take off one packet from * this extra delay but this is easier and being more * conservative is probably better. */ hpts_timeout += slot; } if (hpts_timeout) { /* * Hack alert for now we can't time-out over 2147 seconds (a * bit more than 35min) */ if (hpts_timeout > 0x7ffffffe) hpts_timeout = 0x7ffffffe; bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; } else bbr->r_ctl.rc_timer_exp = 0; if ((slot) && (bbr->rc_use_google || bbr->output_error_seen || (slot <= hpts_timeout)) ) { /* * Tell LRO that it can queue packets while * we pace. */ bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && (bbr->rc_cwnd_limited == 0)) { /* * If we are not cwnd limited and we * are running a rack timer we put on * the do not disturbe even for sack. */ inp->inp_flags2 |= INP_DONT_SACK_QUEUE; } else inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; bbr->rc_pacer_started = cts; (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot), __LINE__, &diag); bbr->rc_timer_first = 0; bbr->bbr_timer_src = frm; bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1); bbr_log_hpts_diag(bbr, cts, &diag); } else if (hpts_timeout) { (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout), __LINE__, &diag); /* * We add the flag here as well if the slot is set, * since hpts will call in to clear the queue first before * calling the output routine (which does our timers). * We don't want to set the flag if its just a timer * else the arrival of data might (that causes us * to send more) might get delayed. Imagine being * on a keep-alive timer and a request comes in for * more data. */ if (slot) bbr->rc_pacer_started = cts; if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && (bbr->rc_cwnd_limited == 0)) { /* * For a rack timer, don't wake us even * if a sack arrives as long as we are * not cwnd limited. */ bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; inp->inp_flags2 |= INP_DONT_SACK_QUEUE; } else { /* All other timers wake us up */ bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; } bbr->bbr_timer_src = frm; bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0); bbr_log_hpts_diag(bbr, cts, &diag); bbr->rc_timer_first = 1; } bbr->rc_tmr_stopped = 0; bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay); } static void bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb) { /* * We received an ack, and then did not call send or were bounced * out due to the hpts was running. Now a timer is up as well, is it * the right timer? */ struct inpcb *inp; struct bbr_sendmap *rsm; uint32_t hpts_timeout; int tmr_up; tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) return; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && (tmr_up == PACE_TMR_RXT)) { /* Should be an RXT */ return; } inp = bbr->rc_inp; if (rsm == NULL) { /* Nothing outstanding? */ if (tp->t_flags & TF_DELACK) { if (tmr_up == PACE_TMR_DELACK) /* * We are supposed to have delayed ack up * and we do */ return; } else if (sbavail(&inp->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { /* * if we hit enobufs then we would expect the * possiblity of nothing outstanding and the RXT up * (and the hptsi timer). */ return; } else if (((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) && (tmr_up == PACE_TMR_KEEP) && (tp->snd_max == tp->snd_una)) { /* We should have keep alive up and we do */ return; } } if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & BBR_HAS_FIN)) { /* needs to be a RXT */ if (tmr_up == PACE_TMR_RXT) return; else goto wrong_timer; } else if (tmr_up == PACE_TMR_RACK) return; else goto wrong_timer; } else if (rsm && (tmr_up == PACE_TMR_RACK)) { /* Rack timer has priority if we have data out */ return; } else if (SEQ_GT(tp->snd_max, tp->snd_una) && ((tmr_up == PACE_TMR_TLP) || (tmr_up == PACE_TMR_RXT))) { /* * Either a TLP or RXT is fine if no sack-passed is in place * and data is outstanding. */ return; } else if (tmr_up == PACE_TMR_DELACK) { /* * If the delayed ack was going to go off before the * rtx/tlp/rack timer were going to expire, then that would * be the timer in control. Note we don't check the time * here trusting the code is correct. */ return; } if (SEQ_GT(tp->snd_max, tp->snd_una) && ((tmr_up == PACE_TMR_RXT) || (tmr_up == PACE_TMR_TLP) || (tmr_up == PACE_TMR_RACK))) { /* * We have outstanding data and * we *do* have a RACK, TLP or RXT * timer running. We won't restart * anything here since thats probably ok we * will get called with some timer here shortly. */ return; } /* * Ok the timer originally started is not what we want now. We will * force the hpts to be stopped if any, and restart with the slot * set to what was in the saved slot. */ wrong_timer: if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) { if (inp->inp_in_hpts) tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); bbr_timer_cancel(bbr, __LINE__, cts); bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val, 0); } else { /* * Output is hptsi so we just need to switch the type of * timer. We don't bother with keep-alive, since when we * jump through the output, it will start the keep-alive if * nothing is sent. * * We only need a delayed-ack added and or the hpts_timeout. */ hpts_timeout = bbr_timer_start(tp, bbr, cts); if (tp->t_flags & TF_DELACK) { if (hpts_timeout == 0) { hpts_timeout = bbr_delack_time; bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; } else if (hpts_timeout > bbr_delack_time) { hpts_timeout = bbr_delack_time; bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; } } if (hpts_timeout) { if (hpts_timeout > 0x7ffffffe) hpts_timeout = 0x7ffffffe; bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; } } } int32_t bbr_clear_lost = 0; /* * Considers the two time values now (cts) and earlier. * If cts is smaller than earlier, we could have * had a sequence wrap (our counter wraps every * 70 min or so) or it could be just clock skew * getting us two differnt time values. Clock skew * will show up within 10ms or so. So in such * a case (where cts is behind earlier time by * less than 10ms) we return 0. Otherwise we * return the true difference between them. */ static inline uint32_t bbr_calc_time(uint32_t cts, uint32_t earlier_time) { /* * Given two timestamps, the current time stamp cts, and some other * time-stamp taken in theory earlier return the difference. The * trick is here sometimes locking will get the other timestamp * after the cts. If this occurs we need to return 0. */ if (TSTMP_GEQ(cts, earlier_time)) return (cts - earlier_time); /* * cts is behind earlier_time if its less than 10ms consider it 0. * If its more than 10ms difference then we had a time wrap. Else * its just the normal locking foo. I wonder if we should not go to * 64bit TS and get rid of this issue. */ if (TSTMP_GEQ((cts + 10000), earlier_time)) return (0); /* * Ok the time must have wrapped. So we need to answer a large * amount of time, which the normal subtraction should do. */ return (cts - earlier_time); } static int sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS) { uint32_t stat; int32_t error; error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t)); if (error || req->newptr == NULL) return error; error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); if (error) return (error); if (stat == 1) { #ifdef BBR_INVARIANTS printf("Clearing BBR lost counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT); COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT); COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT); } else if (stat == 2) { #ifdef BBR_INVARIANTS printf("Clearing BBR option counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE); } else if (stat == 3) { #ifdef BBR_INVARIANTS printf("Clearing BBR stats counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE); } else if (stat == 4) { #ifdef BBR_INVARIANTS printf("Clearing BBR out-size counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE); } bbr_clear_lost = 0; return (0); } static void bbr_init_sysctls(void) { struct sysctl_oid *bbr_probertt; struct sysctl_oid *bbr_hptsi; struct sysctl_oid *bbr_measure; struct sysctl_oid *bbr_cwnd; struct sysctl_oid *bbr_timeout; struct sysctl_oid *bbr_states; struct sysctl_oid *bbr_startup; struct sysctl_oid *bbr_policer; /* Probe rtt controls */ bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "probertt", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "gain", CTLFLAG_RW, &bbr_rttprobe_gain, 192, "What is the filter gain drop in probe_rtt (0=disable)?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "cwnd", CTLFLAG_RW, &bbr_rtt_probe_cwndtarg, 4, "How many mss's are outstanding during probe-rtt"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "int", CTLFLAG_RW, &bbr_rtt_probe_limit, 4000000, "If RTT has not shrank in this many micro-seconds enter probe-rtt"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "mintime", CTLFLAG_RW, &bbr_rtt_probe_time, 200000, "How many microseconds in probe-rtt"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "filter_len_sec", CTLFLAG_RW, &bbr_filter_len_sec, 6, "How long in seconds does the rttProp filter run?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "drain_rtt", CTLFLAG_RW, &bbr_drain_rtt, BBR_SRTT, "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "can_force", CTLFLAG_RW, &bbr_can_force_probertt, 0, "If we keep setting new low rtt's but delay going in probe-rtt can we force in??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "enter_sets_force", CTLFLAG_RW, &bbr_probertt_sets_rtt, 0, "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "can_adjust", CTLFLAG_RW, &bbr_can_adjust_probertt, 1, "Can we dynamically adjust the probe-rtt limits and times?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "is_ratio", CTLFLAG_RW, &bbr_is_ratio, 0, "is the limit to filter a ratio?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "use_cwnd", CTLFLAG_RW, &bbr_prtt_slam_cwnd, 0, "Should we set/recover cwnd?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "can_use_ts", CTLFLAG_RW, &bbr_can_use_ts_for_rtt, 1, "Can we use the ms timestamp if available for retransmistted rtt calculations?"); /* Pacing controls */ bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "pacing", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing", CTLFLAG_RW, &bbr_allow_hdwr_pacing, 1, "Do we allow hardware pacing?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_limit", CTLFLAG_RW, &bbr_hardware_pacing_limit, 4000, "Do we have a limited number of connections for pacing chelsio (0=no limit)?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_adj", CTLFLAG_RW, &bbr_hdwr_pace_adjust, 2, "Multiplier to calculated tso size?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_floor", CTLFLAG_RW, &bbr_hdwr_pace_floor, 1, "Do we invoke the hardware pacing floor?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW, &bbr_hdwr_pacing_delay_cnt, 10, "How many packets must be sent after hdwr pacing is enabled"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "bw_cross", CTLFLAG_RW, &bbr_cross_over, 3000000, "What is the point where we cross over to linux like TSO size set"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_deltarg", CTLFLAG_RW, &bbr_hptsi_segments_delay_tar, 7000, "What is the worse case delay target for hptsi < 48Mbp connections"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "enet_oh", CTLFLAG_RW, &bbr_include_enet_oh, 0, "Do we include the ethernet overhead in calculating pacing delay?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "ip_oh", CTLFLAG_RW, &bbr_include_ip_oh, 1, "Do we include the IP overhead in calculating pacing delay?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "tcp_oh", CTLFLAG_RW, &bbr_include_tcp_oh, 0, "Do we include the TCP overhead in calculating pacing delay?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "google_discount", CTLFLAG_RW, &bbr_google_discount, 10, "What is the default google discount percentage wise for pacing (11 = 1.1%%)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "all_get_min", CTLFLAG_RW, &bbr_all_get_min, 0, "If you are less than a MSS do you just get the min?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "tso_min", CTLFLAG_RW, &bbr_hptsi_bytes_min, 1460, "For 0 -> 24Mbps what is floor number of segments for TSO"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_tso_max", CTLFLAG_RW, &bbr_hptsi_segments_max, 6, "For 0 -> 24Mbps what is top number of segments for TSO"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_floor", CTLFLAG_RW, &bbr_hptsi_segments_floor, 1, "Minimum TSO size we will fall too in segments"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "utter_max", CTLFLAG_RW, &bbr_hptsi_utter_max, 0, "The absolute maximum that any pacing (outside of hardware) can be"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_divisor", CTLFLAG_RW, &bbr_hptsi_per_second, 100, "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps "); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "srtt_mul", CTLFLAG_RW, &bbr_hptsi_max_mul, 1, "The multiplier for pace len max"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "srtt_div", CTLFLAG_RW, &bbr_hptsi_max_div, 2, "The divisor for pace len max"); /* Measurement controls */ bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "measure", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Measurement controls"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "min_i_bw", CTLFLAG_RW, &bbr_initial_bw_bps, 62500, "Minimum initial b/w in bytes per second"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "no_sack_needed", CTLFLAG_RW, &bbr_sack_not_required, 0, "Do we allow bbr to run on connections not supporting SACK?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "use_google", CTLFLAG_RW, &bbr_use_google_algo, 0, "Use has close to google V1.0 has possible?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_limiting", CTLFLAG_RW, &bbr_ts_limiting, 1, "Do we attempt to use the peers timestamp to limit b/w caculations?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_can_raise", CTLFLAG_RW, &bbr_ts_can_raise, 0, "Can we raise the b/w via timestamp b/w calculation?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_delta", CTLFLAG_RW, &bbr_min_usec_delta, 20000, "How long in usec between ts of our sends in ts validation code?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_peer_delta", CTLFLAG_RW, &bbr_min_peer_delta, 20, "What min numerical value should be between the peer deltas?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_delta_percent", CTLFLAG_RW, &bbr_delta_percent, 150, "What percentage (150 = 15.0) do we allow variance for?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "min_measure_good_bw", CTLFLAG_RW, &bbr_min_measurements_req, 1, "What is the minimum measurment count we need before we switch to our b/w estimate"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "min_measure_before_pace", CTLFLAG_RW, &bbr_no_pacing_until, 4, "How many pkt-epoch's (0 is off) do we need before pacing is on?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "quanta", CTLFLAG_RW, &bbr_quanta, 2, "Extra quanta to add when calculating the target (ID section 4.2.3.2)."); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "noretran", CTLFLAG_RW, &bbr_no_retran, 0, "Should google mode not use retransmission measurements for the b/w estimation?"); /* State controls */ bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "states", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "State controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "idle_restart", CTLFLAG_RW, &bbr_uses_idle_restart, 0, "Do we use a new special idle_restart state to ramp back up quickly?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "idle_restart_threshold", CTLFLAG_RW, &bbr_idle_restart_threshold, 100000, "How long must we be idle before we restart??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "use_pkt_epoch", CTLFLAG_RW, &bbr_state_is_pkt_epoch, 0, "Do we use a pkt-epoch for substate if 0 rttProp?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "startup_rtt_gain", CTLFLAG_RW, &bbr_rtt_gain_thresh, 0, "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "drain_floor", CTLFLAG_RW, &bbr_drain_floor, 88, "What is the lowest we can drain (pg) too?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "drain_2_target", CTLFLAG_RW, &bbr_state_drain_2_tar, 1, "Do we drain to target in drain substate?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "gain_2_target", CTLFLAG_RW, &bbr_gain_to_target, 1, "Does probe bw gain to target??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "gain_extra_time", CTLFLAG_RW, &bbr_gain_gets_extra_too, 1, "Does probe bw gain get the extra time too?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "ld_div", CTLFLAG_RW, &bbr_drain_drop_div, 5, "Long drain drop divider?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "ld_mul", CTLFLAG_RW, &bbr_drain_drop_mul, 4, "Long drain drop multiplier?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "rand_ot_disc", CTLFLAG_RW, &bbr_rand_ot, 50, "Random discount of the ot?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "dr_filter_life", CTLFLAG_RW, &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT, "How many packet-epochs does the b/w delivery rate last?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "subdrain_applimited", CTLFLAG_RW, &bbr_sub_drain_app_limit, 0, "Does our sub-state drain invoke app limited if its long?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW, &bbr_sub_drain_slam_cwnd, 0, "Should we set/recover cwnd for sub-state drain?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW, &bbr_slam_cwnd_in_main_drain, 0, "Should we set/recover cwnd for main-state drain?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "google_gets_earlyout", CTLFLAG_RW, &google_allow_early_out, 1, "Should we allow google probe-bw/drain to exit early at flight target?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "google_exit_loss", CTLFLAG_RW, &google_consider_lost, 1, "Should we have losses exit gain of probebw in google mode??"); /* Startup controls */ bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "startup", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Startup controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "cheat_iwnd", CTLFLAG_RW, &bbr_sends_full_iwnd, 1, "Do we not pace but burst out initial windows has our TSO size?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "loss_threshold", CTLFLAG_RW, &bbr_startup_loss_thresh, 2000, "In startup what is the loss threshold in a pe that will exit us from startup?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "use_lowerpg", CTLFLAG_RW, &bbr_use_lower_gain_in_startup, 1, "Should we use a lower hptsi gain if we see loss in startup?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "gain", CTLFLAG_RW, &bbr_start_exit, 25, "What gain percent do we need to see to stay in startup??"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "low_gain", CTLFLAG_RW, &bbr_low_start_exit, 15, "What gain percent do we need to see to stay in the lower gain startup??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "loss_exit", CTLFLAG_RW, &bbr_exit_startup_at_loss, 1, "Should we exit startup at loss in an epoch if we are not gaining?"); /* CWND controls */ bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "cwnd", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Cwnd controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "tar_rtt", CTLFLAG_RW, &bbr_cwndtarget_rtt_touse, 0, "Target cwnd rtt measurment to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "may_shrink", CTLFLAG_RW, &bbr_cwnd_may_shrink, 0, "Can the cwnd shrink if it would grow to more than the target?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "max_target_limit", CTLFLAG_RW, &bbr_target_cwnd_mult_limit, 8, "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "highspeed_min", CTLFLAG_RW, &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS, "What is the high-speed min cwnd (rttProp under 1ms)"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "lowspeed_min", CTLFLAG_RW, &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS, "What is the min cwnd (rttProp > 1ms)"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "initwin", CTLFLAG_RW, &bbr_def_init_win, 10, "What is the BBR initial window, if 0 use tcp version"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "do_loss_red", CTLFLAG_RW, &bbr_do_red, 600, "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_scale", CTLFLAG_RW, &bbr_red_scale, 20000, "What RTT do we scale with?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_growslow", CTLFLAG_RW, &bbr_red_growth_restrict, 1, "Do we restrict cwnd growth for whats in flight?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_div", CTLFLAG_RW, &bbr_red_div, 2, "If we reduce whats the divisor?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_mul", CTLFLAG_RW, &bbr_red_mul, 1, "If we reduce whats the mulitiplier?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "target_is_unit", CTLFLAG_RW, &bbr_target_is_bbunit, 0, "Is the state target the pacing_gain or BBR_UNIT?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "drop_limit", CTLFLAG_RW, &bbr_drop_limit, 0, "Number of segments limit for drop (0=use min_cwnd w/flight)?"); /* Timeout controls */ bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "timeout", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Time out controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "delack", CTLFLAG_RW, &bbr_delack_time, 100000, "BBR's delayed ack time"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_uses", CTLFLAG_RW, &bbr_tlp_type_to_use, 3, "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "persmin", CTLFLAG_RW, &bbr_persist_min, 250000, "What is the minimum time in microseconds between persists"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "persmax", CTLFLAG_RW, &bbr_persist_max, 1000000, "What is the largest delay in microseconds between persists"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_minto", CTLFLAG_RW, &bbr_tlp_min, 10000, "TLP Min timeout in usecs"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_dack_time", CTLFLAG_RW, &bbr_delayed_ack_time, 200000, "TLP delayed ack compensation value"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "minrto", CTLFLAG_RW, &bbr_rto_min_ms, 30, "Minimum RTO in ms"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "maxrto", CTLFLAG_RW, &bbr_rto_max_sec, 4, "Maxiumum RTO in seconds -- should be at least as large as min_rto"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_retry", CTLFLAG_RW, &bbr_tlp_max_resend, 2, "How many times does TLP retry a single segment or multiple with no ACK"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "minto", CTLFLAG_RW, &bbr_min_to, 1000, "Minimum rack timeout in useconds"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "pktdelay", CTLFLAG_RW, &bbr_pkt_delay, 1000, "Extra RACK time (in useconds) besides reordering thresh"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "incr_tmrs", CTLFLAG_RW, &bbr_incr_timers, 1, "Increase the RXT/TLP timer by the pacing time used?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW, &bbr_marks_rxt_sack_passed, 0, "Mark sack passed on all those not ack'd when a RXT hits?"); /* Policer controls */ bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "policer", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Policer controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "detect_enable", CTLFLAG_RW, &bbr_policer_detection_enabled, 1, "Is policer detection enabled??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "min_pes", CTLFLAG_RW, &bbr_lt_intvl_min_rtts, 4, "Minimum number of PE's?"); SYSCTL_ADD_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "bwdiff", CTLFLAG_RW, &bbr_lt_bw_diff, (4000/8), "Minimal bw diff?"); SYSCTL_ADD_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "bwratio", CTLFLAG_RW, &bbr_lt_bw_ratio, 8, "Minimal bw diff?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "from_rack_rxt", CTLFLAG_RW, &bbr_policer_call_from_rack_to, 0, "Do we call the policer detection code from a rack-timeout?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "false_postive", CTLFLAG_RW, &bbr_lt_intvl_fp, 0, "What packet epoch do we do false-postive detection at (0=no)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "loss_thresh", CTLFLAG_RW, &bbr_lt_loss_thresh, 196, "Loss threshold 196 = 19.6%?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "false_postive_thresh", CTLFLAG_RW, &bbr_lt_fd_thresh, 100, "What percentage is the false detection threshold (150=15.0)?"); /* All the rest */ SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "cheat_rxt", CTLFLAG_RW, &bbr_use_rack_resend_cheat, 0, "Do we burst 1ms between sends on retransmissions (like rack)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "error_paceout", CTLFLAG_RW, &bbr_error_base_paceout, 10000, "When we hit an error what is the min to pace out in usec's?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "kill_paceout", CTLFLAG_RW, &bbr_max_net_error_cnt, 10, "When we hit this many errors in a row, kill the session?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "data_after_close", CTLFLAG_RW, &bbr_ignore_data_after_close, 1, "Do we hold off sending a RST until all pending data is ack'd"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "resend_use_tso", CTLFLAG_RW, &bbr_resends_use_tso, 0, "Can resends use TSO?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "sblklimit", CTLFLAG_RW, &bbr_sack_block_limit, 128, "When do we start ignoring small sack blocks"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "bb_verbose", CTLFLAG_RW, &bbr_verbose_logging, 0, "Should BBR black box logging be verbose"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "reorder_thresh", CTLFLAG_RW, &bbr_reorder_thresh, 2, "What factor for rack will be added when seeing reordering (shift right)"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "reorder_fade", CTLFLAG_RW, &bbr_reorder_fade, 0, "Does reorder detection fade, if so how many ms (0 means never)"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, &bbr_tlp_thresh, 1, "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); /* Stats and counters */ /* The pacing counters for hdwr/software can't be in the array */ bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD, &bbr_hdwr_pacing_enobuf, "Total number of enobufs for hardware paced flows"); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD, &bbr_nohdwr_pacing_enobuf, "Total number of enobufs for non-hardware paced flows"); bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "hdwr_pacing", CTLFLAG_RD, &bbr_flows_whdwr_pacing, "Total number of hardware paced flows"); bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "software_pacing", CTLFLAG_RD, &bbr_flows_nohdwr_pacing, "Total number of software paced flows"); COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "stats", CTLFLAG_RD, bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats"); COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "opts", CTLFLAG_RD, bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats"); COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "lost", CTLFLAG_RD, bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur"); COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "stateresend", CTLFLAG_RD, bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend"); COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "statetime", CTLFLAG_RD, bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states"); COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "outsize", CTLFLAG_RD, bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls"); SYSCTL_ADD_PROC(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters"); } static void bbr_counter_destroy(void) { COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE); COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE); COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE); COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT); COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT); COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT); counter_u64_free(bbr_nohdwr_pacing_enobuf); counter_u64_free(bbr_hdwr_pacing_enobuf); counter_u64_free(bbr_flows_whdwr_pacing); counter_u64_free(bbr_flows_nohdwr_pacing); } static __inline void bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts) { memset(l, 0, sizeof(union tcp_log_stackspecific)); l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate; l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate); l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop); l->bw_inuse = bbr_get_bw(bbr); l->inflight = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); l->applimited = bbr->r_ctl.r_app_limited_until; l->delivered = bbr->r_ctl.rc_delivered; l->timeStamp = cts; l->lost = bbr->r_ctl.rc_lost; l->bbr_state = bbr->rc_bbr_state; l->bbr_substate = bbr_state_val(bbr); l->epoch = bbr->r_ctl.rc_rtt_epoch; l->lt_epoch = bbr->r_ctl.rc_lt_epoch; l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain; l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain; l->inhpts = bbr->rc_inp->inp_in_hpts; l->ininput = bbr->rc_inp->inp_in_input; l->use_lt_bw = bbr->rc_lt_use_bw; l->pkts_out = bbr->r_ctl.rc_flight_at_input; l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch; } static void bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = 0; log.u_bbr.flex2 = 0; log.u_bbr.flex5 = 0; log.u_bbr.flex3 = 0; log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate; log.u_bbr.flex7 = reason; log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt; log.u_bbr.flex8 = 0; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BW_RED_EV, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = seq; log.u_bbr.flex2 = count; log.u_bbr.flex8 = mode; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_LOWGAIN, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling, uint8_t reason, uint32_t p_maxseg, int len) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = p_maxseg; log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; log.u_bbr.flex4 = reason; log.u_bbr.flex5 = bbr->rc_in_persist; log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val; log.u_bbr.flex7 = p_maxseg; log.u_bbr.flex8 = bbr->rc_in_persist; log.u_bbr.pkts_out = 0; log.u_bbr.applimited = len; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_JUSTRET, 0, tlen, &log, false, &bbr->rc_tv); } } static void bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = seq; log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_ENTREC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_msgsize_fail(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t len, uint32_t maxseg, uint32_t mtu, int32_t csum_flags, int32_t tso, uint32_t cts) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = tso; log.u_bbr.flex2 = maxseg; log.u_bbr.flex3 = mtu; log.u_bbr.flex4 = csum_flags; TCP_LOG_EVENTP(tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_MSGSIZE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_flowend(struct tcp_bbr *bbr) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct sockbuf *r, *s; struct timeval tv; if (bbr->rc_inp->inp_socket) { r = &bbr->rc_inp->inp_socket->so_rcv; s = &bbr->rc_inp->inp_socket->so_snd; } else { r = s = NULL; } bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv)); TCP_LOG_EVENTP(bbr->rc_tp, NULL, r, s, TCP_LOG_FLOWEND, 0, 0, &log, false, &tv); } } static void bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t lost, uint32_t del) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = lost; log.u_bbr.flex2 = del; log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw; log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt; log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; log.u_bbr.flex7 = line; log.u_bbr.flex8 = 0; log.u_bbr.inflight = bbr->r_ctl.r_measurement_count; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_PKT_EPOCH, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->r_ctl.rc_lost; log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat; log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat; log.u_bbr.flex7 = line; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIME_EPOCH, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex2 = new_tar; log.u_bbr.flex3 = line; log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; log.u_bbr.flex5 = bbr_quanta; log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs; log.u_bbr.flex7 = bbr->rc_last_options; log.u_bbr.flex8 = meth; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_STATE_TARGET, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; if (bbr_state_is_pkt_epoch) log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); else log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP); log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000); log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra; log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_STATE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt; log.u_bbr.flex4 = applied; log.u_bbr.flex5 = rtt; log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex7 = cond; log.u_bbr.flex8 = reas; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_RTT_SHRINKS, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_exit_rec(struct tcp_bbr *bbr) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start; log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_EXITREC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg, uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = line; log.u_bbr.flex2 = prev_acked; log.u_bbr.flex3 = bytes_this_ack; log.u_bbr.flex4 = chg; log.u_bbr.flex5 = th_ack; log.u_bbr.flex6 = target; log.u_bbr.flex8 = meth; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_CWND, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin) { /* * Log the rtt sample we are applying to the srtt algorithm in * useconds. */ if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = rtt; log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time; log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay; log.u_bbr.flex4 = bbr->rc_tp->ts_offset; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv); log.u_bbr.flex6 = tsin; log.u_bbr.flex7 = 0; log.u_bbr.flex8 = bbr->rc_ack_was_delayed; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, TCP_LOG_RTT, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = time_in; log.u_bbr.flex2 = line; log.u_bbr.flex8 = enter_exit; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_PERSIST, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age; log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_ACKCLEAR, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen, uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = nsegs; log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes; if (m) { struct timespec ts; log.u_bbr.flex3 = m->m_flags; if (m->m_flags & M_TSTMP) { mbuf_tstmp2timespec(m, &ts); tv.tv_sec = ts.tv_sec; tv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv); } else { log.u_bbr.lt_epoch = 0; } if (m->m_flags & M_TSTMP_LRO) { tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000; tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000; log.u_bbr.flex5 = tcp_tv_to_usectick(&tv); } else { /* No arrival timestamp */ log.u_bbr.flex5 = 0; } log.u_bbr.pkts_out = tcp_get_usecs(&tv); } else { log.u_bbr.flex3 = 0; log.u_bbr.flex5 = 0; log.u_bbr.flex6 = 0; log.u_bbr.pkts_out = 0; } log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex7 = bbr->r_wanted_output; log.u_bbr.flex8 = bbr->rc_in_persist; TCP_LOG_EVENTP(bbr->rc_tp, th, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, TCP_LOG_IN, 0, tlen, &log, true, &bbr->rc_tv); } } static void bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = did_out; log.u_bbr.flex2 = nxt_pkt; log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val; log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes; log.u_bbr.flex7 = bbr->r_wanted_output; log.u_bbr.flex8 = bbr->rc_in_persist; log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_DOSEG_DONE, 0, 0, &log, true, &bbr->rc_tv); } } static void bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts, int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = o_len; log.u_bbr.flex3 = segcnt; log.u_bbr.flex4 = segsiz; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_ENOBUF_JMP, ENOBUFS, len, &log, true, &bbr->rc_tv); } } static void bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = timers; log.u_bbr.flex2 = ret; log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex5 = cts; log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex8 = hpts_calling; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TO_PROCESS, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; uint64_t ar; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->bbr_timer_src; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; ar = (uint64_t)(bbr->r_ctl.rc_resend); ar >>= 32; ar &= 0x00000000ffffffff; log.u_bbr.flex4 = (uint32_t)ar; ar = (uint64_t)bbr->r_ctl.rc_resend; ar &= 0x00000000ffffffff; log.u_bbr.flex5 = (uint32_t)ar; log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.flex8 = to_num; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_RTO, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = flex1; log.u_bbr.flex2 = flex2; log.u_bbr.flex3 = flex3; log.u_bbr.flex4 = 0; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; log.u_bbr.flex8 = reason; log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_REDUCE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = diag->p_nxt_slot; log.u_bbr.flex2 = diag->p_cur_slot; log.u_bbr.flex3 = diag->slot_req; log.u_bbr.flex4 = diag->inp_hptsslot; log.u_bbr.flex5 = diag->slot_remaining; log.u_bbr.flex6 = diag->need_new_to; log.u_bbr.flex7 = diag->p_hpts_active; log.u_bbr.flex8 = diag->p_on_min_sleep; /* Hijack other fields as needed */ log.u_bbr.epoch = diag->have_slept; log.u_bbr.lt_epoch = diag->yet_to_sleep; log.u_bbr.pkts_out = diag->co_ret; log.u_bbr.applimited = diag->hpts_sleep_time; log.u_bbr.delivered = diag->p_prev_slot; - log.u_bbr.inflight = diag->p_runningtick; - log.u_bbr.bw_inuse = diag->wheel_tick; + log.u_bbr.inflight = diag->p_runningslot; + log.u_bbr.bw_inuse = diag->wheel_slot; log.u_bbr.rttProp = diag->wheel_cts; - log.u_bbr.delRate = diag->maxticks; + log.u_bbr.delRate = diag->maxslots; log.u_bbr.cur_del_rate = diag->p_curtick; log.u_bbr.cur_del_rate <<= 32; log.u_bbr.cur_del_rate |= diag->p_lasttick; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSDIAG, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, uint32_t thresh, uint32_t to) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->rc_tp->t_rttvar; log.u_bbr.flex2 = time_since_sent; log.u_bbr.flex3 = srtt; log.u_bbr.flex4 = thresh; log.u_bbr.flex5 = to; log.u_bbr.flex6 = bbr->rc_tp->t_srtt; log.u_bbr.flex8 = mode; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERPREP, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = usecs; log.u_bbr.flex2 = len; log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff); log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff); if (override) log.u_bbr.flex5 = (1 << 2); else log.u_bbr.flex5 = 0; log.u_bbr.flex6 = override; log.u_bbr.flex7 = gain; log.u_bbr.flex8 = mod; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSI_CALC, 0, len, &log, false, &bbr->rc_tv); } } static void bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->bbr_timer_src; log.u_bbr.flex2 = to; log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = slot; log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot; log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2; log.u_bbr.flex8 = which; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERSTAR, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_thresh_choice(struct tcp_bbr *bbr, uint32_t cts, uint32_t thresh, uint32_t lro, uint32_t srtt, struct bbr_sendmap *rsm, uint8_t frm) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = thresh; log.u_bbr.flex2 = lro; log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts; log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.flex6 = srtt; log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift; log.u_bbr.flex8 = frm; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_THRESH_CALC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = bbr->bbr_timer_src; log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = bbr->rc_in_persist; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.flex8 = hpts_removed; log.u_bbr.pkts_out = bbr->rc_pacer_started; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERCANC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio; log.u_bbr.flex2 = (peer_delta >> 32); log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff); log.u_bbr.flex4 = (delta >> 32); log.u_bbr.flex5 = (delta & 0x00000000ffffffff); log.u_bbr.flex7 = bbr->rc_ts_clock_set; log.u_bbr.flex8 = bbr->rc_ts_cant_be_used; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TSTMP_VAL, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_tsosize(struct tcp_bbr *bbr, uint32_t cts, uint32_t tsosz, uint32_t tls, uint32_t old_val, uint32_t maxseg, int hdwr) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = tsosz; log.u_bbr.flex2 = tls; log.u_bbr.flex3 = tcp_min_hptsi_time; log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min; log.u_bbr.flex5 = old_val; log.u_bbr.flex6 = maxseg; log.u_bbr.flex7 = bbr->rc_no_pacing; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= bbr->rc_past_init_win; if (hdwr) log.u_bbr.flex8 = 0x80 | bbr->rc_use_google; else log.u_bbr.flex8 = bbr->rc_use_google; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRTSO, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, uint32_t flags, uint32_t line) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = rsm->r_start; log.u_bbr.flex3 = rsm->r_end; log.u_bbr.flex4 = rsm->r_delivered; log.u_bbr.flex5 = rsm->r_rtr_cnt; log.u_bbr.flex6 = rsm->r_dupack; log.u_bbr.flex7 = rsm->r_tim_lastsent[0]; log.u_bbr.flex8 = rsm->r_flags; /* Hijack the pkts_out fids */ log.u_bbr.applimited = flags; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_RSM_CLEARED, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts, uint32_t flex3, uint32_t flex2, uint32_t flex5, uint32_t flex6, uint32_t pkts_out, int flex7, uint32_t flex4, uint32_t flex1) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = flex1; log.u_bbr.flex2 = flex2; log.u_bbr.flex3 = flex3; log.u_bbr.flex4 = flex4; log.u_bbr.flex5 = flex5; log.u_bbr.flex6 = flex6; log.u_bbr.flex7 = flex7; /* Hijack the pkts_out fids */ log.u_bbr.pkts_out = pkts_out; log.u_bbr.flex8 = flex8; if (bbr->rc_ack_was_delayed) log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay; else log.u_bbr.epoch = 0; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRUPD, 0, flex2, &log, false, &bbr->rc_tv); } } static void bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason, uint32_t newbw, uint32_t obw, uint32_t diff, uint32_t tim) { if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = reason; log.u_bbr.flex2 = newbw; log.u_bbr.flex3 = obw; log.u_bbr.flex4 = diff; log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost; log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del; log.u_bbr.flex7 = bbr->rc_lt_is_sampling; log.u_bbr.pkts_out = tim; log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw; if (bbr->rc_lt_use_bw == 0) log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; else log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BWSAMP, 0, 0, &log, false, &bbr->rc_tv); } } static inline void bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = line; log.u_bbr.flex2 = tick; log.u_bbr.flex3 = tp->t_maxunacktime; log.u_bbr.flex4 = tp->t_acktime; log.u_bbr.flex8 = event; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_PROGRESS, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp, uint64_t rate, uint64_t hw_rate, int line, uint32_t cts, int error) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); log.u_bbr.bw_inuse = rate; log.u_bbr.flex5 = line; log.u_bbr.flex6 = error; log.u_bbr.flex8 = bbr->skip_gain; log.u_bbr.flex8 <<= 1; log.u_bbr.flex8 |= bbr->gain_is_limited; log.u_bbr.flex8 <<= 1; log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing; log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_HDWR_PACE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, uint32_t del_by, uint32_t cts, uint32_t line, uint32_t prev_delay) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = slot; log.u_bbr.flex2 = del_by; log.u_bbr.flex3 = prev_delay; log.u_bbr.flex4 = line; log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val; log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay; log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags); log.u_bbr.flex8 = bbr->rc_in_persist; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRSND, 0, len, &log, false, &bbr->rc_tv); } } static void bbr_log_type_bbrrttprop(struct tcp_bbr *bbr, uint32_t t, uint32_t end, uint32_t tsconv, uint32_t cts, int32_t match, uint32_t seq, uint8_t flags) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->r_ctl.rc_delivered; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt; log.u_bbr.flex4 = end; log.u_bbr.flex5 = seq; log.u_bbr.flex6 = t; log.u_bbr.flex7 = match; log.u_bbr.flex8 = flags; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRRTT, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options); log.u_bbr.flex3 = bbr->r_ctl.gain_epoch; log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs; log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight; log.u_bbr.flex7 = 0; log.u_bbr.flex8 = entry_method; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_EXIT_GAIN, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); /* R-HU */ log.u_bbr.flex1 = 0; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = 0; log.u_bbr.flex4 = 0; log.u_bbr.flex7 = 0; log.u_bbr.flex8 = settings_desired; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_SETTINGS_CHG, 0, 0, &log, false, &bbr->rc_tv); } } /* * Returns the bw from the our filter. */ static inline uint64_t bbr_get_full_bw(struct tcp_bbr *bbr) { uint64_t bw; bw = get_filter_value(&bbr->r_ctl.rc_delrate); return (bw); } static inline void bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { uint64_t calclr; uint32_t lost, del; if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch) lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch; else lost = 0; del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del; if (lost == 0) { calclr = 0; } else if (del) { calclr = lost; calclr *= (uint64_t)1000; calclr /= (uint64_t)del; } else { /* Nothing delivered? 100.0% loss */ calclr = 1000; } bbr->r_ctl.rc_pkt_epoch_loss_rate = (uint32_t)calclr; if (IN_RECOVERY(bbr->rc_tp->t_flags)) bbr->r_ctl.recovery_lr += (uint32_t)calclr; bbr->r_ctl.rc_pkt_epoch++; if (bbr->rc_no_pacing && (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) { bbr->rc_no_pacing = 0; tcp_bbr_tso_size_check(bbr, cts); } bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time); bbr->r_ctl.rc_pkt_epoch_time = cts; /* What was our loss rate */ bbr_log_pkt_epoch(bbr, cts, line, lost, del); bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered; bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost; } static inline void bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { uint32_t epoch_time; /* Tick the RTT clock */ bbr->r_ctl.rc_rtt_epoch++; epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start; bbr_log_time_epoch(bbr, cts, line, epoch_time); bbr->r_ctl.rc_rcv_epoch_start = cts; } static inline void bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked) { if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) { bbr->rc_is_pkt_epoch_now = 1; } } /* * Returns the bw from either the b/w filter * or from the lt_bw (if the connection is being * policed). */ static inline uint64_t __bbr_get_bw(struct tcp_bbr *bbr) { uint64_t bw, min_bw; uint64_t rtt; int gm_measure_cnt = 1; /* * For startup we make, like google, a * minimum b/w. This is generated from the * IW and the rttProp. We do fall back to srtt * if for some reason (initial handshake) we don't * have a rttProp. We, in the worst case, fall back * to the configured min_bw (rc_initial_hptsi_bw). */ if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { /* Attempt first to use rttProp */ rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (rtt && (rtt < 0xffffffff)) { measure: min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * ((uint64_t)1000000); min_bw /= rtt; if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) { min_bw = bbr->r_ctl.rc_initial_hptsi_bw; } } else if (bbr->rc_tp->t_srtt != 0) { /* No rttProp, use srtt? */ rtt = bbr_get_rtt(bbr, BBR_SRTT); goto measure; } else { min_bw = bbr->r_ctl.rc_initial_hptsi_bw; } } else min_bw = 0; if ((bbr->rc_past_init_win == 0) && (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp))) bbr->rc_past_init_win = 1; if ((bbr->rc_use_google) && (bbr->r_ctl.r_measurement_count >= 1)) gm_measure_cnt = 0; if (gm_measure_cnt && ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) || (bbr->rc_past_init_win == 0))) { /* For google we use our guess rate until we get 1 measurement */ use_initial_window: rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (rtt && (rtt < 0xffffffff)) { /* * We have an RTT measurment. Use that in * combination with our initial window to calculate * a b/w. */ bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * ((uint64_t)1000000); bw /= rtt; if (bw < bbr->r_ctl.rc_initial_hptsi_bw) { bw = bbr->r_ctl.rc_initial_hptsi_bw; } } else { /* Drop back to the 40 and punt to a default */ bw = bbr->r_ctl.rc_initial_hptsi_bw; } if (bw < 1) /* Probably should panic */ bw = 1; if (bw > min_bw) return (bw); else return (min_bw); } if (bbr->rc_lt_use_bw) bw = bbr->r_ctl.rc_lt_bw; else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0)) bw = bbr->r_ctl.red_bw; else bw = get_filter_value(&bbr->r_ctl.rc_delrate); if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) { /* * Enforce user set rate limit, keep in mind that * t_peakrate_thr is in B/s already */ bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw); } if (bw == 0) { /* We should not be at 0, go to the initial window then */ goto use_initial_window; } if (bw < 1) /* Probably should panic */ bw = 1; if (bw < min_bw) bw = min_bw; return (bw); } static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr) { uint64_t bw; bw = __bbr_get_bw(bbr); return (bw); } static inline void bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts) { bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch; bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time; bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } static inline void bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts) { bbr->rc_lt_is_sampling = 0; bbr->rc_lt_use_bw = 0; bbr->r_ctl.rc_lt_bw = 0; bbr_reset_lt_bw_interval(bbr, cts); } static inline void bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin) { uint64_t diff; /* Do we have a previous sample? */ if (bbr->r_ctl.rc_lt_bw) { /* Get the diff in bytes per second */ if (bbr->r_ctl.rc_lt_bw > bw) diff = bbr->r_ctl.rc_lt_bw - bw; else diff = bw - bbr->r_ctl.rc_lt_bw; if ((diff <= bbr_lt_bw_diff) || (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) { /* Consider us policed */ uint32_t saved_bw; saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw; bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2; /* average of two */ bbr->rc_lt_use_bw = 1; bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; /* * Use pkt based epoch for measuring length of * policer up */ bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch; /* * reason 4 is we need to start consider being * policed */ bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin); return; } } bbr->r_ctl.rc_lt_bw = bw; bbr_reset_lt_bw_interval(bbr, cts); bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin); } static void bbr_randomize_extra_state_time(struct tcp_bbr *bbr) { uint32_t ran, deduct; ran = arc4random_uniform(bbr_rand_ot); if (ran) { deduct = bbr->r_ctl.rc_level_state_extra / ran; bbr->r_ctl.rc_level_state_extra -= deduct; } } /* * Return randomly the starting state * to use in probebw. */ static uint8_t bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts) { uint32_t ran; uint8_t ret_val; /* Initialize the offset to 0 */ bbr->r_ctl.rc_exta_time_gd = 0; bbr->rc_hit_state_1 = 0; bbr->r_ctl.rc_level_state_extra = 0; ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1)); /* * The math works funny here :) the return value is used to set the * substate and then the state change is called which increments by * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when * we fully enter the state. Note that the (8 - 1 - ran) assures that * we return 1 - 7, so we dont return 0 and end up starting in * state 1 (DRAIN). */ ret_val = BBR_SUBSTATE_COUNT - 1 - ran; /* Set an epoch */ if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) bbr_set_epoch(bbr, cts, __LINE__); bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; return (ret_val); } static void bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected) { uint32_t diff, d_time; uint64_t del_time, bw, lost, delivered; if (bbr->r_use_policer == 0) return; if (bbr->rc_lt_use_bw) { /* We are using lt bw do we stop yet? */ diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; if (diff > bbr_lt_bw_max_rtts) { /* Reset it all */ reset_all: bbr_reset_lt_bw_sampling(bbr, cts); if (bbr->rc_filled_pipe) { bbr_set_epoch(bbr, cts, __LINE__); bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr_substate_change(bbr, cts, __LINE__, 0); bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr_log_type_statechange(bbr, cts, __LINE__); } else { /* * This should not happen really * unless we remove the startup/drain * restrictions above. */ bbr->rc_bbr_state = BBR_STATE_STARTUP; bbr_set_epoch(bbr, cts, __LINE__); bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; bbr_set_state_target(bbr, __LINE__); bbr_log_type_statechange(bbr, cts, __LINE__); } /* reason 0 is to stop using lt-bw */ bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0); return; } if (bbr_lt_intvl_fp == 0) { /* Not doing false-postive detection */ return; } /* False positive detection */ if (diff == bbr_lt_intvl_fp) { /* At bbr_lt_intvl_fp we record the lost */ bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) { /* Now is our loss rate still high? */ lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; if ((delivered == 0) || (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) { /* No still below our threshold */ bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0); } else { /* Yikes its still high, it must be a false positive */ bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0); goto reset_all; } } return; } /* * Wait for the first loss before sampling, to let the policer * exhaust its tokens and estimate the steady-state rate allowed by * the policer. Starting samples earlier includes bursts that * over-estimate the bw. */ if (bbr->rc_lt_is_sampling == 0) { /* reason 1 is to begin doing the sampling */ if (loss_detected == 0) return; bbr_reset_lt_bw_interval(bbr, cts); bbr->rc_lt_is_sampling = 1; bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0); return; } /* Now how long were we delivering long term last> */ if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time)) d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time; else d_time = 0; /* To avoid underestimates, reset sampling if we run out of data. */ if (bbr->r_ctl.r_app_limited_until) { /* Can not measure in app-limited state */ bbr_reset_lt_bw_sampling(bbr, cts); /* reason 2 is to reset sampling due to app limits */ bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time); return; } diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; if (diff < bbr_lt_intvl_min_rtts) { /* * need more samples (we don't * start on a round like linux so * we need 1 more). */ /* 6 is not_enough time or no-loss */ bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); return; } if (diff > (4 * bbr_lt_intvl_min_rtts)) { /* * For now if we wait too long, reset all sampling. We need * to do some research here, its possible that we should * base this on how much loss as occurred.. something like * if its under 10% (or some thresh) reset all otherwise * don't. Thats for phase II I guess. */ bbr_reset_lt_bw_sampling(bbr, cts); /* reason 3 is to reset sampling due too long of sampling */ bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); return; } /* * End sampling interval when a packet is lost, so we estimate the * policer tokens were exhausted. Stopping the sampling before the * tokens are exhausted under-estimates the policed rate. */ if (loss_detected == 0) { /* 6 is not_enough time or no-loss */ bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); return; } /* Calculate packets lost and delivered in sampling interval. */ lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; if ((delivered == 0) || (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) { bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time); return; } if (d_time < 1000) { /* Not enough time. wait */ /* 6 is not_enough time or no-loss */ bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); return; } if (d_time >= (0xffffffff / USECS_IN_MSEC)) { /* Too long */ bbr_reset_lt_bw_sampling(bbr, cts); /* reason 3 is to reset sampling due too long of sampling */ bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); return; } del_time = d_time; bw = delivered; bw *= (uint64_t)USECS_IN_SECOND; bw /= del_time; bbr_lt_bw_samp_done(bbr, bw, cts, d_time); } /* * Allocate a sendmap from our zone. */ static struct bbr_sendmap * bbr_alloc(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; BBR_STAT_INC(bbr_to_alloc); rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO)); if (rsm) { bbr->r_ctl.rc_num_maps_alloced++; return (rsm); } if (bbr->r_ctl.rc_free_cnt) { BBR_STAT_INC(bbr_to_alloc_emerg); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); bbr->r_ctl.rc_free_cnt--; return (rsm); } BBR_STAT_INC(bbr_to_alloc_failed); return (NULL); } static struct bbr_sendmap * bbr_alloc_full_limit(struct tcp_bbr *bbr) { if ((V_tcp_map_entries_limit > 0) && (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { BBR_STAT_INC(bbr_alloc_limited); if (!bbr->alloc_limit_reported) { bbr->alloc_limit_reported = 1; BBR_STAT_INC(bbr_alloc_limited_conns); } return (NULL); } return (bbr_alloc(bbr)); } /* wrapper to allocate a sendmap entry, subject to a specific limit */ static struct bbr_sendmap * bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type) { struct bbr_sendmap *rsm; if (limit_type) { /* currently there is only one limit type */ if (V_tcp_map_split_limit > 0 && bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { BBR_STAT_INC(bbr_split_limited); if (!bbr->alloc_limit_reported) { bbr->alloc_limit_reported = 1; BBR_STAT_INC(bbr_alloc_limited_conns); } return (NULL); } } /* allocate and mark in the limit type, if set */ rsm = bbr_alloc(bbr); if (rsm != NULL && limit_type) { rsm->r_limit_type = limit_type; bbr->r_ctl.rc_num_split_allocs++; } return (rsm); } static void bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) { if (rsm->r_limit_type) { /* currently there is only one limit type */ bbr->r_ctl.rc_num_split_allocs--; } if (rsm->r_is_smallmap) bbr->r_ctl.rc_num_small_maps_alloced--; if (bbr->r_ctl.rc_tlp_send == rsm) bbr->r_ctl.rc_tlp_send = NULL; if (bbr->r_ctl.rc_resend == rsm) { bbr->r_ctl.rc_resend = NULL; } if (bbr->r_ctl.rc_next == rsm) bbr->r_ctl.rc_next = NULL; if (bbr->r_ctl.rc_sacklast == rsm) bbr->r_ctl.rc_sacklast = NULL; if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { memset(rsm, 0, sizeof(struct bbr_sendmap)); TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); rsm->r_limit_type = 0; bbr->r_ctl.rc_free_cnt++; return; } bbr->r_ctl.rc_num_maps_alloced--; uma_zfree(bbr_zone, rsm); } /* * Returns the BDP. */ static uint64_t bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) { /* * Calculate the bytes in flight needed given the bw (in bytes per * second) and the specifyed rtt in useconds. We need to put out the * returned value per RTT to match that rate. Gain will normally * raise it up from there. * * This should not overflow as long as the bandwidth is below 1 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30). */ uint64_t usec_per_sec; usec_per_sec = USECS_IN_SECOND; return ((rtt * bw) / usec_per_sec); } /* * Return the initial cwnd. */ static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp) { uint32_t i_cwnd; if (bbr->rc_init_win) { i_cwnd = bbr->rc_init_win * tp->t_maxseg; } else if (V_tcp_initcwnd_segments) i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), max(2 * tp->t_maxseg, 14600)); else if (V_tcp_do_rfc3390) i_cwnd = min(4 * tp->t_maxseg, max(2 * tp->t_maxseg, 4380)); else { /* Per RFC5681 Section 3.1 */ if (tp->t_maxseg > 2190) i_cwnd = 2 * tp->t_maxseg; else if (tp->t_maxseg > 1095) i_cwnd = 3 * tp->t_maxseg; else i_cwnd = 4 * tp->t_maxseg; } return (i_cwnd); } /* * Given a specified gain, return the target * cwnd based on that gain. */ static uint32_t bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw) { uint64_t bdp, rtt; uint32_t cwnd; if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) || (bbr_get_full_bw(bbr) == 0)) { /* No measurements yet */ return (bbr_initial_cwnd(bbr, bbr->rc_tp)); } /* * Get bytes per RTT needed (rttProp is normally in * bbr_cwndtarget_rtt_touse) */ rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse); /* Get the bdp from the two values */ bdp = bbr_get_bw_delay_prod(rtt, bw); /* Now apply the gain */ cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT)); return (cwnd); } static uint32_t bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain) { uint32_t cwnd, mss; mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); /* Get the base cwnd with gain rounded to a mss */ cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss); /* * Add in N (2 default since we do not have a * fq layer to trap packets in) quanta's per the I-D * section 4.2.3.2 quanta adjust. */ cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs); if (bbr->rc_use_google) { if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && (bbr_state_val(bbr) == BBR_SUB_GAIN)) { /* * The linux implementation adds * an extra 2 x mss in gain cycle which * is documented no-where except in the code. * so we add more for Neal undocumented feature */ cwnd += 2 * mss; } if ((cwnd / mss) & 0x1) { /* Round up for odd num mss */ cwnd += mss; } } /* Are we below the min cwnd? */ if (cwnd < get_min_cwnd(bbr)) return (get_min_cwnd(bbr)); return (cwnd); } static uint16_t bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain) { if (gain < 1) gain = 1; return (gain); } static uint32_t bbr_get_header_oh(struct tcp_bbr *bbr) { int seg_oh; seg_oh = 0; if (bbr->r_ctl.rc_inc_tcp_oh) { /* Do we include TCP overhead? */ seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr)); } if (bbr->r_ctl.rc_inc_ip_oh) { /* Do we include IP overhead? */ #ifdef INET6 if (bbr->r_is_v6) seg_oh += sizeof(struct ip6_hdr); else #endif #ifdef INET seg_oh += sizeof(struct ip); #endif } if (bbr->r_ctl.rc_inc_enet_oh) { /* Do we include the ethernet overhead? */ seg_oh += sizeof(struct ether_header); } return(seg_oh); } static uint32_t bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw) { uint64_t divor, res, tim; if (useconds_time == 0) return (0); gain = bbr_gain_adjust(bbr, gain); divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT; tim = useconds_time; res = (tim * bw * gain) / divor; if (res == 0) res = 1; return ((uint32_t)res); } /* * Given a gain and a length return the delay in useconds that * should be used to evenly space out packets * on the connection (based on the gain factor). */ static uint32_t bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog) { uint64_t bw, lentim, res; uint32_t usecs, srtt, over = 0; uint32_t seg_oh, num_segs, maxseg; if (len == 0) return (0); maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; num_segs = (len + maxseg - 1) / maxseg; if (bbr->rc_use_google == 0) { seg_oh = bbr_get_header_oh(bbr); len += (num_segs * seg_oh); } gain = bbr_gain_adjust(bbr, gain); bw = bbr_get_bw(bbr); if (bbr->rc_use_google) { uint64_t cbw; /* * Reduce the b/w by the google discount * factor 10 = 1%. */ cbw = bw * (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount); cbw /= (uint64_t)1000; /* We don't apply a discount if it results in 0 */ if (cbw > 0) bw = cbw; } lentim = ((uint64_t)len * (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT); res = lentim / ((uint64_t)gain * bw); if (res == 0) res = 1; usecs = (uint32_t)res; srtt = bbr_get_rtt(bbr, BBR_SRTT); if (bbr_hptsi_max_mul && bbr_hptsi_max_div && (bbr->rc_use_google == 0) && (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) { /* * We cannot let the delay be more than 1/2 the srtt time. * Otherwise we cannot pace out or send properly. */ over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div; BBR_STAT_INC(bbr_hpts_min_time); } if (!nolog) bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1); return (usecs); } static void bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack, uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses) { INP_WLOCK_ASSERT(tp->t_inpcb); uint64_t bw; uint32_t cwnd, target_cwnd, saved_bytes, maxseg; int32_t meth; #ifdef STATS if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GEQ(th->th_ack, tp->gput_ack)) { /* * Strech acks and compressed acks will cause this to * oscillate but we are doing it the same way as the main * stack so it will be compariable (though possibly not * ideal). */ int32_t cgput; int64_t gput, time_stamp; gput = (int64_t) (th->th_ack - tp->gput_seq) * 8; time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000)); cgput = gput / time_stamp; stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, cgput); if (tp->t_stats_gput_prev > 0) stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_GPUT_ND, ((gput - tp->t_stats_gput_prev) * 100) / tp->t_stats_gput_prev); tp->t_flags &= ~TF_GPUTINPROG; tp->t_stats_gput_prev = cgput; } #endif if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { /* We don't change anything in probe-rtt */ return; } maxseg = tp->t_maxseg - bbr->rc_last_options; saved_bytes = bytes_this_ack; bytes_this_ack += sack_changed; if (bytes_this_ack > prev_acked) { bytes_this_ack -= prev_acked; /* * A byte ack'd gives us a full mss * to be like linux i.e. they count packets. */ if ((bytes_this_ack < maxseg) && bbr->rc_use_google) bytes_this_ack = maxseg; } else { /* Unlikely */ bytes_this_ack = 0; } cwnd = tp->snd_cwnd; bw = get_filter_value(&bbr->r_ctl.rc_delrate); if (bw) target_cwnd = bbr_get_target_cwnd(bbr, bw, (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain); else target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp); if (IN_RECOVERY(tp->t_flags) && (bbr->bbr_prev_in_rec == 0)) { /* * We are entering recovery and * thus packet conservation. */ bbr->pkt_conservation = 1; bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime; cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + bytes_this_ack; } if (IN_RECOVERY(tp->t_flags)) { uint32_t flight; bbr->bbr_prev_in_rec = 1; if (cwnd > losses) { cwnd -= losses; if (cwnd < maxseg) cwnd = maxseg; } else cwnd = maxseg; flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr_log_type_cwndupd(bbr, flight, 0, losses, 10, 0, 0, line); if (bbr->pkt_conservation) { uint32_t time_in; if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start)) time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start; else time_in = 0; if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { /* Clear packet conservation after an rttProp */ bbr->pkt_conservation = 0; } else { if ((flight + bytes_this_ack) > cwnd) cwnd = flight + bytes_this_ack; if (cwnd < get_min_cwnd(bbr)) cwnd = get_min_cwnd(bbr); tp->snd_cwnd = cwnd; bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, 1, target_cwnd, th->th_ack, line); return; } } } else bbr->bbr_prev_in_rec = 0; if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) { bbr->r_ctl.restrict_growth--; if (bytes_this_ack > maxseg) bytes_this_ack = maxseg; } if (bbr->rc_filled_pipe) { /* * Here we have exited startup and filled the pipe. We will * thus allow the cwnd to shrink to the target. We hit here * mostly. */ uint32_t s_cwnd; meth = 2; s_cwnd = min((cwnd + bytes_this_ack), target_cwnd); if (s_cwnd > cwnd) cwnd = s_cwnd; else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing) cwnd = s_cwnd; } else { /* * Here we are still in startup, we increase cwnd by what * has been acked. */ if ((cwnd < target_cwnd) || (bbr->rc_past_init_win == 0)) { meth = 3; cwnd += bytes_this_ack; } else { /* * Method 4 means we are at target so no gain in * startup and past the initial window. */ meth = 4; } } tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr)); bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line); } static void tcp_bbr_partialack(struct tcpcb *tp) { struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= tp->snd_cwnd) { bbr->r_wanted_output = 1; } } static void bbr_post_recovery(struct tcpcb *tp) { struct tcp_bbr *bbr; uint32_t flight; INP_WLOCK_ASSERT(tp->t_inpcb); bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* * Here we just exit recovery. */ EXIT_RECOVERY(tp->t_flags); /* Lock in our b/w reduction for the specified number of pkt-epochs */ bbr->r_recovery_bw = 0; tp->snd_recover = tp->snd_una; tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); bbr->pkt_conservation = 0; if (bbr->rc_use_google == 0) { /* * For non-google mode lets * go ahead and make sure we clear * the recovery state so if we * bounce back in to recovery we * will do PC. */ bbr->bbr_prev_in_rec = 0; } bbr_log_type_exit_rec(bbr); if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__); } else { /* For probe-rtt case lets fix up its saved_cwnd */ if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) { bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent; bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__); } } flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if ((bbr->rc_use_google == 0) && bbr_do_red) { uint64_t val, lr2use; uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd; uint32_t *cwnd_p; if (bbr_get_rtt(bbr, BBR_SRTT)) { val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000); val /= bbr_get_rtt(bbr, BBR_SRTT); ratio = (uint32_t)val; } else ratio = 1000; bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, bbr->r_ctl.recovery_lr, 21, ratio, bbr->r_ctl.rc_red_cwnd_pe, __LINE__); if ((ratio < bbr_do_red) || (bbr_do_red == 0)) goto done; if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && bbr_prtt_slam_cwnd) || (bbr_sub_drain_slam_cwnd && (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && bbr->rc_hit_state_1 && (bbr_state_val(bbr) == BBR_SUB_DRAIN)) || ((bbr->rc_bbr_state == BBR_STATE_DRAIN) && bbr_slam_cwnd_in_main_drain)) { /* * Here we must poke at the saved cwnd * as well as the cwnd. */ cwnd = bbr->r_ctl.rc_saved_cwnd; cwnd_p = &bbr->r_ctl.rc_saved_cwnd; } else { cwnd = tp->snd_cwnd; cwnd_p = &tp->snd_cwnd; } maxseg = tp->t_maxseg - bbr->rc_last_options; /* Add the overall lr with the recovery lr */ if (bbr->r_ctl.rc_lost == 0) lr2use = 0; else if (bbr->r_ctl.rc_delivered == 0) lr2use = 1000; else { lr2use = bbr->r_ctl.rc_lost * 1000; lr2use /= bbr->r_ctl.rc_delivered; } lr2use += bbr->r_ctl.recovery_lr; acks_inflight = (flight / (maxseg * 2)); if (bbr_red_scale) { lr2use *= bbr_get_rtt(bbr, BBR_SRTT); lr2use /= bbr_red_scale; if ((bbr_red_growth_restrict) && ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1)) bbr->r_ctl.restrict_growth += acks_inflight; } if (lr2use) { val = (uint64_t)cwnd * lr2use; val /= 1000; if (cwnd > val) newcwnd = roundup((cwnd - val), maxseg); else newcwnd = maxseg; } else { val = (uint64_t)cwnd * (uint64_t)bbr_red_mul; val /= (uint64_t)bbr_red_div; newcwnd = roundup((uint32_t)val, maxseg); } /* with standard delayed acks how many acks can I expect? */ if (bbr_drop_limit == 0) { /* * Anticpate how much we will * raise the cwnd based on the acks. */ if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) { /* We do enforce the min (with the acks) */ newcwnd = (get_min_cwnd(bbr) - acks_inflight); } } else { /* * A strict drop limit of N is is inplace */ if (newcwnd < (bbr_drop_limit * maxseg)) { newcwnd = bbr_drop_limit * maxseg; } } /* For the next N acks do we restrict the growth */ *cwnd_p = newcwnd; if (tp->snd_cwnd > newcwnd) tp->snd_cwnd = newcwnd; bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22, (uint32_t)lr2use, bbr_get_rtt(bbr, BBR_SRTT), __LINE__); bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch; } done: bbr->r_ctl.recovery_lr = 0; if (flight <= tp->snd_cwnd) { bbr->r_wanted_output = 1; } tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); } static void bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts) { bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate); /* Limit the drop in b/w to 1/2 our current filter. */ if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate) bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate; if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2)) bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2; tcp_bbr_tso_size_check(bbr, cts); } static void bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm) { struct tcp_bbr *bbr; INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); #endif bbr = (struct tcp_bbr *)tp->t_fb_ptr; switch (type) { case CC_NDUPACK: if (!IN_RECOVERY(tp->t_flags)) { tp->snd_recover = tp->snd_max; /* Start a new epoch */ bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__); if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) { /* * Move forward the lt epoch * so it won't count the truncated * epoch. */ bbr->r_ctl.rc_lt_epoch++; } if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { /* * Just like the policer detection code * if we are in startup we must push * forward the last startup epoch * to hide the truncated PE. */ bbr->r_ctl.rc_bbr_last_startup_epoch++; } bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd; ENTER_RECOVERY(tp->t_flags); bbr->rc_tlp_rtx_out = 0; bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate; tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); if (bbr->rc_inp->inp_in_hpts && ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) { /* * When we enter recovery, we need to restart * any timers. This may mean we gain an agg * early, which will be made up for at the last * rxt out. */ bbr->rc_timer_first = 1; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } /* * Calculate a new cwnd based on to the current * delivery rate with no gain. We get the bdp * without gaining it up like we normally would and * we use the last cur_del_rate. */ if ((bbr->rc_use_google == 0) && (bbr->r_ctl.bbr_rttprobe_gain_val || (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) { tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + (tp->t_maxseg - bbr->rc_last_options); if (tp->snd_cwnd < get_min_cwnd(bbr)) { /* We always gate to min cwnd */ tp->snd_cwnd = get_min_cwnd(bbr); } bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__); } bbr_log_type_enter_rec(bbr, rsm->r_start); } break; case CC_RTO_ERR: KMOD_TCPSTAT_INC(tcps_sndrexmitbad); /* RTO was unnecessary, so reset everything. */ bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime); if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { tp->snd_cwnd = tp->snd_cwnd_prev; tp->snd_ssthresh = tp->snd_ssthresh_prev; tp->snd_recover = tp->snd_recover_prev; tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__); } tp->t_badrxtwin = 0; break; } } /* * Indicate whether this ack should be delayed. We can delay the ack if * following conditions are met: * - There is no delayed ack timer in progress. * - Our last ack wasn't a 0-sized window. We never want to delay * the ack that opens up a 0-sized window. * - LRO wasn't used for this segment. We make sure by checking that the * segment size is not larger than the MSS. * - Delayed acks are enabled or this is a half-synchronized T/TCP * connection. * - The data being acked is less than a full segment (a stretch ack * of more than a segment we should ack. * - nsegs is 1 (if its more than that we received more than 1 ack). */ #define DELAY_ACK(tp, bbr, nsegs) \ (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ ((tp->t_flags & TF_DELACK) == 0) && \ ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) && \ (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) /* * Return the lowest RSM in the map of * packets still in flight that is not acked. * This should normally find on the first one * since we remove packets from the send * map after they are marked ACKED. */ static struct bbr_sendmap * bbr_find_lowest_rsm(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; /* * Walk the time-order transmitted list looking for an rsm that is * not acked. This will be the one that was sent the longest time * ago that is still outstanding. */ TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) { if (rsm->r_flags & BBR_ACKED) { continue; } goto finish; } finish: return (rsm); } static struct bbr_sendmap * bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) { struct bbr_sendmap *prsm; /* * Walk the sequence order list backward until we hit and arrive at * the highest seq not acked. In theory when this is called it * should be the last segment (which it was not). */ prsm = rsm; TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) { continue; } return (prsm); } return (NULL); } /* * Returns to the caller the number of microseconds that * the packet can be outstanding before we think we * should have had an ack returned. */ static uint32_t bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm) { /* * lro is the flag we use to determine if we have seen reordering. * If it gets set we have seen reordering. The reorder logic either * works in one of two ways: * * If reorder-fade is configured, then we track the last time we saw * re-ordering occur. If we reach the point where enough time as * passed we no longer consider reordering has occuring. * * Or if reorder-face is 0, then once we see reordering we consider * the connection to alway be subject to reordering and just set lro * to 1. * * In the end if lro is non-zero we add the extra time for * reordering in. */ int32_t lro; uint32_t thresh, t_rxtcur; if (srtt == 0) srtt = 1; if (bbr->r_ctl.rc_reorder_ts) { if (bbr->r_ctl.rc_reorder_fade) { if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) { lro = cts - bbr->r_ctl.rc_reorder_ts; if (lro == 0) { /* * No time as passed since the last * reorder, mark it as reordering. */ lro = 1; } } else { /* Negative time? */ lro = 0; } if (lro > bbr->r_ctl.rc_reorder_fade) { /* Turn off reordering seen too */ bbr->r_ctl.rc_reorder_ts = 0; lro = 0; } } else { /* Reodering does not fade */ lro = 1; } } else { lro = 0; } thresh = srtt + bbr->r_ctl.rc_pkt_delay; if (lro) { /* It must be set, if not you get 1/4 rtt */ if (bbr->r_ctl.rc_reorder_shift) thresh += (srtt >> bbr->r_ctl.rc_reorder_shift); else thresh += (srtt >> 2); } else { thresh += 1000; } /* We don't let the rack timeout be above a RTO */ if ((bbr->rc_tp)->t_srtt == 0) t_rxtcur = BBR_INITIAL_RTO; else t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); if (thresh > t_rxtcur) { thresh = t_rxtcur; } /* And we don't want it above the RTO max either */ if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); } bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK); return (thresh); } /* * Return to the caller the amount of time in mico-seconds * that should be used for the TLP timer from the last * send time of this packet. */ static uint32_t bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts) { uint32_t thresh, len, maxseg, t_rxtcur; struct bbr_sendmap *prsm; if (srtt == 0) srtt = 1; if (bbr->rc_tlp_threshold) thresh = srtt + (srtt / bbr->rc_tlp_threshold); else thresh = (srtt * 2); maxseg = tp->t_maxseg - bbr->rc_last_options; /* Get the previous sent packet, if any */ len = rsm->r_end - rsm->r_start; /* 2.1 behavior */ prsm = TAILQ_PREV(rsm, bbr_head, r_tnext); if (prsm && (len <= maxseg)) { /* * Two packets outstanding, thresh should be (2*srtt) + * possible inter-packet delay (if any). */ uint32_t inter_gap = 0; int idx, nidx; idx = rsm->r_rtr_cnt - 1; nidx = prsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { /* Yes it was sent later (or at the same time) */ inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; } thresh += inter_gap; } else if (len <= maxseg) { /* * Possibly compensate for delayed-ack. */ uint32_t alt_thresh; alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } /* Not above the current RTO */ if (tp->t_srtt == 0) t_rxtcur = BBR_INITIAL_RTO; else t_rxtcur = TICKS_2_USEC(tp->t_rxtcur); bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP); /* Not above an RTO */ if (thresh > t_rxtcur) { thresh = t_rxtcur; } /* Not above a RTO max */ if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); } /* And now apply the user TLP min */ if (thresh < bbr_tlp_min) { thresh = bbr_tlp_min; } return (thresh); } /* * Return one of three RTTs to use (in microseconds). */ static __inline uint32_t bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type) { uint32_t f_rtt; uint32_t srtt; f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) { /* We have no rtt at all */ if (bbr->rc_tp->t_srtt == 0) f_rtt = BBR_INITIAL_RTO; else f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); /* * Since we don't know how good the rtt is apply a * delayed-ack min */ if (f_rtt < bbr_delayed_ack_time) { f_rtt = bbr_delayed_ack_time; } } /* Take the filter version or last measured pkt-rtt */ if (rtt_type == BBR_RTT_PROP) { srtt = f_rtt; } else if (rtt_type == BBR_RTT_PKTRTT) { if (bbr->r_ctl.rc_pkt_epoch_rtt) { srtt = bbr->r_ctl.rc_pkt_epoch_rtt; } else { /* No pkt rtt yet */ srtt = f_rtt; } } else if (rtt_type == BBR_RTT_RACK) { srtt = bbr->r_ctl.rc_last_rtt; /* We need to add in any internal delay for our timer */ if (bbr->rc_ack_was_delayed) srtt += bbr->r_ctl.rc_ack_hdwr_delay; } else if (rtt_type == BBR_SRTT) { srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); } else { /* TSNH */ srtt = f_rtt; #ifdef BBR_INVARIANTS panic("Unknown rtt request type %d", rtt_type); #endif } return (srtt); } static int bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts) { uint32_t thresh; thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK), cts, rsm); if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) { /* It is lost (past time) */ return (1); } return (0); } /* * Return a sendmap if we need to retransmit something. */ static struct bbr_sendmap * bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * Check to see that we don't need to fall into recovery. We will * need to do so if our oldest transmit is past the time we should * have had an ack. */ struct bbr_sendmap *rsm; int32_t idx; if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) { /* Nothing outstanding that we know of */ return (NULL); } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (rsm == NULL) { /* Nothing in the transmit map */ return (NULL); } if (tp->t_flags & TF_SENTFIN) { /* Fin restricted, don't find anything once a fin is sent */ return (NULL); } if (rsm->r_flags & BBR_ACKED) { /* * Ok the first one is acked (this really should not happen * since we remove the from the tmap once they are acked) */ rsm = bbr_find_lowest_rsm(bbr); if (rsm == NULL) return (NULL); } idx = rsm->r_rtr_cnt - 1; if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) { /* Send timestamp is the same or less? can't be ready */ return (NULL); } /* Get our RTT time */ if (bbr_is_lost(bbr, rsm, cts) && ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || (rsm->r_flags & BBR_SACK_PASSED))) { if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { rsm->r_flags |= BBR_MARKED_LOST; bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; } bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm); #ifdef BBR_INVARIANTS if ((rsm->r_end - rsm->r_start) == 0) panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm); #endif return (rsm); } return (NULL); } /* * RACK Timer, here we simply do logging and house keeping. * the normal bbr_output_wtime() function will call the * appropriate thing to check if we need to do a RACK retransmit. * We return 1, saying don't proceed with bbr_output_wtime only * when all timers have been stopped (destroyed PCB?). */ static int bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * This timer simply provides an internal trigger to send out data. * The check_recovery_mode call will see if there are needed * retransmissions, if so we will enter fast-recovery. The output * call may or may not do the same thing depending on sysctl * settings. */ uint32_t lost; if (bbr->rc_all_timers_stopped) { return (1); } if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } BBR_STAT_INC(bbr_to_tot); lost = bbr->r_ctl.rc_lost; if (bbr->r_state && (bbr->r_state != tp->t_state)) bbr_set_state(tp, bbr, 0); bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK); if (bbr->r_ctl.rc_resend == NULL) { /* Lets do the check here */ bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); } if (bbr_policer_call_from_rack_to) bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; return (0); } static __inline void bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start) { int idx; nrsm->r_start = start; nrsm->r_end = rsm->r_end; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed; nrsm->r_flags = rsm->r_flags; /* We don't transfer forward the SYN flag */ nrsm->r_flags &= ~BBR_HAS_SYN; /* We move forward the FIN flag, not that this should happen */ rsm->r_flags &= ~BBR_HAS_FIN; nrsm->r_dupack = rsm->r_dupack; nrsm->r_rtr_bytes = 0; nrsm->r_is_gain = rsm->r_is_gain; nrsm->r_is_drain = rsm->r_is_drain; nrsm->r_delivered = rsm->r_delivered; nrsm->r_ts_valid = rsm->r_ts_valid; nrsm->r_del_ack_ts = rsm->r_del_ack_ts; nrsm->r_del_time = rsm->r_del_time; nrsm->r_app_limited = rsm->r_app_limited; nrsm->r_first_sent_time = rsm->r_first_sent_time; nrsm->r_flight_at_send = rsm->r_flight_at_send; /* We split a piece the lower section looses any just_ret flag. */ nrsm->r_bbr_state = rsm->r_bbr_state; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } rsm->r_end = nrsm->r_start; idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); idx /= 8; /* Check if we got too small */ if ((rsm->r_is_smallmap == 0) && ((rsm->r_end - rsm->r_start) <= idx)) { bbr->r_ctl.rc_num_small_maps_alloced++; rsm->r_is_smallmap = 1; } /* Check the new one as well */ if ((nrsm->r_end - nrsm->r_start) <= idx) { bbr->r_ctl.rc_num_small_maps_alloced++; nrsm->r_is_smallmap = 1; } } static int bbr_sack_mergable(struct bbr_sendmap *at, uint32_t start, uint32_t end) { /* * Given a sack block defined by * start and end, and a current postion * at. Return 1 if either side of at * would show that the block is mergable * to that side. A block to be mergable * must have overlap with the start/end * and be in the SACK'd state. */ struct bbr_sendmap *l_rsm; struct bbr_sendmap *r_rsm; /* first get the either side blocks */ l_rsm = TAILQ_PREV(at, bbr_head, r_next); r_rsm = TAILQ_NEXT(at, r_next); if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) { /* Potentially mergeable */ if ((l_rsm->r_end == start) || (SEQ_LT(start, l_rsm->r_end) && SEQ_GT(end, l_rsm->r_end))) { /* * map blk |------| * sack blk |------| * * map blk |------| * sack blk |------| */ return (1); } } if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) { /* Potentially mergeable */ if ((r_rsm->r_start == end) || (SEQ_LT(start, r_rsm->r_start) && SEQ_GT(end, r_rsm->r_start))) { /* * map blk |---------| * sack blk |----| * * map blk |---------| * sack blk |-------| */ return (1); } } return (0); } static struct bbr_sendmap * bbr_merge_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *l_rsm, struct bbr_sendmap *r_rsm) { /* * We are merging two ack'd RSM's, * the l_rsm is on the left (lower seq * values) and the r_rsm is on the right * (higher seq value). The simplest way * to merge these is to move the right * one into the left. I don't think there * is any reason we need to try to find * the oldest (or last oldest retransmitted). */ l_rsm->r_end = r_rsm->r_end; if (l_rsm->r_dupack < r_rsm->r_dupack) l_rsm->r_dupack = r_rsm->r_dupack; if (r_rsm->r_rtr_bytes) l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; if (r_rsm->r_in_tmap) { /* This really should not happen */ TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext); } if (r_rsm->r_app_limited) l_rsm->r_app_limited = r_rsm->r_app_limited; /* Now the flags */ if (r_rsm->r_flags & BBR_HAS_FIN) l_rsm->r_flags |= BBR_HAS_FIN; if (r_rsm->r_flags & BBR_TLP) l_rsm->r_flags |= BBR_TLP; if (r_rsm->r_flags & BBR_RWND_COLLAPSED) l_rsm->r_flags |= BBR_RWND_COLLAPSED; if (r_rsm->r_flags & BBR_MARKED_LOST) { /* This really should not happen */ bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start; } TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next); if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { /* Transfer the split limit to the map we free */ r_rsm->r_limit_type = l_rsm->r_limit_type; l_rsm->r_limit_type = 0; } bbr_free(bbr, r_rsm); return(l_rsm); } /* * TLP Timer, here we simply setup what segment we want to * have the TLP expire on, the normal bbr_output_wtime() will then * send it out. * * We return 1, saying don't proceed with bbr_output_wtime only * when all timers have been stopped (destroyed PCB?). */ static int bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * Tail Loss Probe. */ struct bbr_sendmap *rsm = NULL; struct socket *so; uint32_t amm; uint32_t out, avail; uint32_t maxseg; int collapsed_win = 0; if (bbr->rc_all_timers_stopped) { return (1); } if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); return (1); } /* Did we somehow get into persists? */ if (bbr->rc_in_persist) { return (0); } if (bbr->r_state && (bbr->r_state != tp->t_state)) bbr_set_state(tp, bbr, 0); BBR_STAT_INC(bbr_tlp_tot); maxseg = tp->t_maxseg - bbr->rc_last_options; /* * A TLP timer has expired. We have been idle for 2 rtts. So we now * need to figure out how to force a full MSS segment out. */ so = tp->t_inpcb->inp_socket; avail = sbavail(&so->so_snd); out = ctf_outstanding(tp); if (out > tp->snd_wnd) { /* special case, we need a retransmission */ collapsed_win = 1; goto need_retran; } if (avail > out) { /* New data is available */ amm = avail - out; if (amm > maxseg) { amm = maxseg; } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { /* not enough to fill a MTU and no-delay is off */ goto need_retran; } /* Set the send-new override */ if ((out + amm) <= tp->snd_wnd) { bbr->rc_tlp_new_data = 1; } else { goto need_retran; } bbr->r_ctl.rc_tlp_seg_send_cnt = 0; bbr->r_ctl.rc_last_tlp_seq = tp->snd_max; bbr->r_ctl.rc_tlp_send = NULL; /* cap any slots */ BBR_STAT_INC(bbr_tlp_newdata); goto send; } need_retran: /* * Ok we need to arrange the last un-acked segment to be re-sent, or * optionally the first un-acked segment. */ if (collapsed_win == 0) { rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); if (rsm && (BBR_ACKED | BBR_HAS_FIN)) { rsm = bbr_find_high_nonack(bbr, rsm); } if (rsm == NULL) { goto restore; } } else { /* * We must find the last segment * that was acceptable by the client. */ TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) { /* Found one */ break; } } if (rsm == NULL) { /* None? if so send the first */ rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm == NULL) goto restore; } } if ((rsm->r_end - rsm->r_start) > maxseg) { /* * We need to split this the last segment in two. */ struct bbr_sendmap *nrsm; nrsm = bbr_alloc_full_limit(bbr); if (nrsm == NULL) { /* * We can't get memory to split, we can either just * not split it. Or retransmit the whole piece, lets * do the large send (BTLP :-) ). */ goto go_for_it; } bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg)); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); rsm = nrsm; } go_for_it: bbr->r_ctl.rc_tlp_send = rsm; bbr->rc_tlp_rtx_out = 1; if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) { bbr->r_ctl.rc_tlp_seg_send_cnt++; tp->t_rxtshift++; } else { bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; bbr->r_ctl.rc_tlp_seg_send_cnt = 1; } send: if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { /* * Can't [re]/transmit a segment we have retranmitted the * max times. We need the retransmit timer to take over. */ restore: bbr->rc_tlp_new_data = 0; bbr->r_ctl.rc_tlp_send = NULL; if (rsm) rsm->r_flags &= ~BBR_TLP; BBR_STAT_INC(bbr_tlp_retran_fail); return (0); } else if (rsm) { rsm->r_flags |= BBR_TLP; } if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) && (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) { /* * We have retransmitted to many times for TLP. Switch to * the regular RTO timer */ goto restore; } bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); } /* * Delayed ack Timer, here we simply need to setup the * ACK_NOW flag and remove the DELACK flag. From there * the output routine will send the ack out. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { if (bbr->rc_all_timers_stopped) { return (1); } bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK); tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; KMOD_TCPSTAT_INC(tcps_delack); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; return (0); } /* * Here we send a KEEP-ALIVE like probe to the * peer, we do not send data. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { struct tcptemp *t_template; int32_t retval = 1; if (bbr->rc_all_timers_stopped) { return (1); } if (bbr->rc_in_persist == 0) return (0); KASSERT(tp->t_inpcb != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); /* * Persistence timer into zero window. Force a byte to be output, if * possible. */ bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; KMOD_TCPSTAT_INC(tcps_persisttimeo); /* * Have we exceeded the user specified progress time? */ if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } /* * Hack: if the peer is dead/unreachable, we do not time out if the * window is closed. After a full backoff, drop the connection if * the idle time (no responses to probes) reaches the maximum * backoff that we would use if retransmitting. */ if (tp->t_rxtshift == TCP_MAXRXTSHIFT && (ticks - tp->t_rcvtime >= tcp_maxpersistidle || ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) && tp->snd_una == tp->snd_max) { bbr_exit_persist(tp, bbr, cts, __LINE__); retval = 0; goto out; } /* * If the user has closed the socket then drop a persisting * connection after a much reduced timeout. */ if (tp->t_state > TCPS_CLOSE_WAIT && (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } t_template = tcpip_maketemplate(bbr->rc_inp); if (t_template) { tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); /* This sends an ack */ if (tp->t_flags & TF_DELACK) tp->t_flags &= ~TF_DELACK; free(t_template, M_TEMP); } if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0); out: return (retval); } /* * If a keepalive goes off, we had no other timers * happening. We always return 1 here since this * routine either drops the connection or sends * out a segment with respond. */ static int bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { struct tcptemp *t_template; struct inpcb *inp; if (bbr->rc_all_timers_stopped) { return (1); } bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; inp = tp->t_inpcb; bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP); /* * Keep-alive timer went off; send something or drop connection if * idle for too long. */ KMOD_TCPSTAT_INC(tcps_keeptimeo); if (tp->t_state < TCPS_ESTABLISHED) goto dropit; if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && tp->t_state <= TCPS_CLOSING) { if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) goto dropit; /* * Send a packet designed to force a response if the peer is * up and reachable: either an ACK if the connection is * still alive, or an RST if the peer has closed the * connection due to timeout or reboot. Using sequence * number tp->snd_una-1 causes the transmitted zero-length * segment to lie outside the receive window; by the * protocol spec, this requires the correspondent TCP to * respond. */ KMOD_TCPSTAT_INC(tcps_keepprobe); t_template = tcpip_maketemplate(inp); if (t_template) { tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); free(t_template, M_TEMP); } } bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0); return (1); dropit: KMOD_TCPSTAT_INC(tcps_keepdrops); tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); return (1); } /* * Retransmit helper function, clear up all the ack * flags and take care of important book keeping. */ static void bbr_remxt_tmr(struct tcpcb *tp) { /* * The retransmit timer went off, all sack'd blocks must be * un-acked. */ struct bbr_sendmap *rsm, *trsm = NULL; struct tcp_bbr *bbr; uint32_t cts, lost; bbr = (struct tcp_bbr *)tp->t_fb_ptr; cts = tcp_get_usecs(&bbr->rc_tv); lost = bbr->r_ctl.rc_lost; if (bbr->r_state && (bbr->r_state != tp->t_state)) bbr_set_state(tp, bbr, 0); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { if (rsm->r_flags & BBR_ACKED) { uint32_t old_flags; rsm->r_dupack = 0; if (rsm->r_in_tmap == 0) { /* We must re-add it back to the tlist */ if (trsm == NULL) { TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); } else { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext); } rsm->r_in_tmap = 1; } old_flags = rsm->r_flags; rsm->r_flags |= BBR_RXT_CLEARED; rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS); bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); } else { if ((tp->t_state < TCPS_ESTABLISHED) && (rsm->r_start == tp->snd_una)) { /* * Special case for TCP FO. Where * we sent more data beyond the snd_max. * We don't mark that as lost and stop here. */ break; } if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; } if (bbr_marks_rxt_sack_passed) { /* * With this option, we will rack out * in 1ms increments the rest of the packets. */ rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST; rsm->r_flags &= ~BBR_WAS_SACKPASS; } else { /* * With this option we only mark them lost * and remove all sack'd markings. We will run * another RXT or a TLP. This will cause * us to eventually send more based on what * ack's come in. */ rsm->r_flags |= BBR_MARKED_LOST; rsm->r_flags &= ~BBR_WAS_SACKPASS; rsm->r_flags &= ~BBR_SACK_PASSED; } } trsm = rsm; } bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map); /* Clear the count (we just un-acked them) */ bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR); bbr->rc_tlp_new_data = 0; bbr->r_ctl.rc_tlp_seg_send_cnt = 0; /* zap the behindness on a rxt */ bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; bbr->rc_tlp_rtx_out = 0; bbr->r_ctl.rc_sacked = 0; bbr->r_ctl.rc_sacklast = NULL; bbr->r_timer_override = 1; bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); } /* * Re-transmit timeout! If we drop the PCB we will return 1, otherwise * we will setup to retransmit the lowest seq number outstanding. */ static int bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { int32_t rexmt; int32_t retval = 0; bool isipv6; bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; if (bbr->rc_all_timers_stopped) { return (1); } if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_una == tp->snd_max)) { /* Nothing outstanding .. nothing to do */ return (0); } /* * Retransmission timer went off. Message has not been acked within * retransmit interval. Back off to a longer retransmit interval * and retransmit one segment. */ if (ctf_progress_timeout_check(tp, true)) { retval = 1; bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } bbr_remxt_tmr(tp); if ((bbr->r_ctl.rc_resend == NULL) || ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) { /* * If the rwnd collapsed on * the one we are retransmitting * it does not count against the * rxt count. */ tp->t_rxtshift++; } if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { tp->t_rxtshift = TCP_MAXRXTSHIFT; KMOD_TCPSTAT_INC(tcps_timeoutdrop); retval = 1; tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); tcp_set_inp_to_drop(bbr->rc_inp, (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); goto out; } if (tp->t_state == TCPS_SYN_SENT) { /* * If the SYN was retransmitted, indicate CWND to be limited * to 1 segment in cc_conn_init(). */ tp->snd_cwnd = 1; } else if (tp->t_rxtshift == 1) { /* * first retransmit; record ssthresh and cwnd so they can be * recovered if this turns out to be a "bad" retransmit. A * retransmit is considered "bad" if an ACK for this segment * is received within RTT/2 interval; the assumption here is * that the ACK was already in flight. See "On Estimating * End-to-End Network Path Properties" by Allman and Paxson * for more details. */ tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; if (!IN_RECOVERY(tp->t_flags)) { tp->snd_cwnd_prev = tp->snd_cwnd; tp->snd_ssthresh_prev = tp->snd_ssthresh; tp->snd_recover_prev = tp->snd_recover; tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); tp->t_flags |= TF_PREVVALID; } else { tp->t_flags &= ~TF_PREVVALID; } tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; } else { tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; tp->t_flags &= ~TF_PREVVALID; } KMOD_TCPSTAT_INC(tcps_rexmttimeo); if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift]; else rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; TCPT_RANGESET(tp->t_rxtcur, rexmt, MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); /* * We enter the path for PLMTUD if connection is established or, if * connection is FIN_WAIT_1 status, reason for the last is that if * amount of data we send is very small, we could send it in couple * of packets and process straight to FIN. In that case we won't * catch ESTABLISHED state. */ #ifdef INET6 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false; #else isipv6 = false; #endif if (((V_tcp_pmtud_blackhole_detect == 1) || (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && ((tp->t_state == TCPS_ESTABLISHED) || (tp->t_state == TCPS_FIN_WAIT_1))) { /* * Idea here is that at each stage of mtu probe (usually, * 1448 -> 1188 -> 524) should be given 2 chances to recover * before further clamping down. 'tp->t_rxtshift % 2 == 0' * should take care of that. */ if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && tp->t_rxtshift % 2 == 0)) { /* * Enter Path MTU Black-hole Detection mechanism: - * Disable Path MTU Discovery (IP "DF" bit). - * Reduce MTU to lower value than what we negotiated * with peer. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { /* * Record that we may have found a black * hole. */ tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; /* Keep track of previous MSS. */ tp->t_pmtud_saved_maxseg = tp->t_maxseg; } /* * Reduce the MSS to blackhole value or to the * default in an attempt to retransmit. */ #ifdef INET6 isipv6 = bbr->r_is_v6; if (isipv6 && tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else if (isipv6) { /* Use the default MSS. */ tp->t_maxseg = V_tcp_v6mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else { /* Use the default MSS. */ tp->t_maxseg = V_tcp_mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif } else { /* * If further retransmissions are still unsuccessful * with a lowered MTU, maybe this isn't a blackhole * and we restore the previous MSS and blackhole * detection flags. The limit '6' is determined by * giving each probe stage (1448, 1188, 524) 2 * chances to recover. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && (tp->t_rxtshift >= 6)) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; tp->t_maxseg = tp->t_pmtud_saved_maxseg; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); } } } /* * Disable RFC1323 and SACK if we haven't got any response to our * third SYN to work-around some broken terminal servers (most of * which have hopefully been retired) that have bad VJ header * compression code which trashes TCP segments containing * unknown-to-them TCP options. */ if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && (tp->t_rxtshift == 3)) tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); /* * If we backed off this far, our srtt estimate is probably bogus. * Clobber it so we'll take the next rtt measurement as our srtt; * move the current srtt into rttvar to keep the current retransmit * times until then. */ if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { #ifdef INET6 if (bbr->r_is_v6) in6_losing(tp->t_inpcb); else #endif in_losing(tp->t_inpcb); tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); tp->t_srtt = 0; } sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); tp->snd_recover = tp->snd_max; tp->t_flags |= TF_ACKNOW; tp->t_rtttime = 0; out: return (retval); } static int bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling) { int32_t ret = 0; int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK); if (timers == 0) { return (0); } if (tp->t_state == TCPS_LISTEN) { /* no timers on listen sockets */ if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) return (0); return (1); } if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { uint32_t left; if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { ret = -1; bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); return (0); } if (hpts_calling == 0) { ret = -2; bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); return (0); } /* * Ok our timer went off early and we are not paced false * alarm, go back to sleep. */ left = bbr->r_ctl.rc_timer_exp - cts; ret = -3; bbr_log_to_processing(bbr, cts, ret, left, hpts_calling); tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left)); return (1); } bbr->rc_tmr_stopped = 0; bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; if (timers & PACE_TMR_DELACK) { ret = bbr_timeout_delack(tp, bbr, cts); } else if (timers & PACE_TMR_PERSIT) { ret = bbr_timeout_persist(tp, bbr, cts); } else if (timers & PACE_TMR_RACK) { bbr->r_ctl.rc_tlp_rxt_last_time = cts; ret = bbr_timeout_rack(tp, bbr, cts); } else if (timers & PACE_TMR_TLP) { bbr->r_ctl.rc_tlp_rxt_last_time = cts; ret = bbr_timeout_tlp(tp, bbr, cts); } else if (timers & PACE_TMR_RXT) { bbr->r_ctl.rc_tlp_rxt_last_time = cts; ret = bbr_timeout_rxt(tp, bbr, cts); } else if (timers & PACE_TMR_KEEP) { ret = bbr_timeout_keepalive(tp, bbr, cts); } bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling); return (ret); } static void bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts) { if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { uint8_t hpts_removed = 0; if (bbr->rc_inp->inp_in_hpts && (bbr->rc_timer_first == 1)) { /* * If we are canceling timer's when we have the * timer ahead of the output being paced. We also * must remove ourselves from the hpts. */ hpts_removed = 1; tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT); if (bbr->r_ctl.rc_last_delay_val) { /* Update the last hptsi delay too */ uint32_t time_since_send; if (TSTMP_GT(cts, bbr->rc_pacer_started)) time_since_send = cts - bbr->rc_pacer_started; else time_since_send = 0; if (bbr->r_ctl.rc_last_delay_val > time_since_send) { /* Cut down our slot time */ bbr->r_ctl.rc_last_delay_val -= time_since_send; } else { bbr->r_ctl.rc_last_delay_val = 0; } bbr->rc_pacer_started = cts; } } bbr->rc_timer_first = 0; bbr_log_to_cancel(bbr, line, cts, hpts_removed); bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); } } static void bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type) { struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; bbr->rc_all_timers_stopped = 1; return; } /* * stop all timers always returning 0. */ static int bbr_stopall(struct tcpcb *tp) { return (0); } static void bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) { return; } /* * return true if a bbr timer (rack or tlp) is active. */ static int bbr_timer_active(struct tcpcb *tp, uint32_t timer_type) { return (0); } static uint32_t bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts) { struct bbr_sendmap *rsm; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if ((rsm == NULL) || (u_rsm == rsm)) return (cts); return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); } static void bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time) { int32_t idx; rsm->r_rtr_cnt++; rsm->r_dupack = 0; if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) { rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS; rsm->r_flags |= BBR_OVERMAX; } if (rsm->r_flags & BBR_RWND_COLLAPSED) { /* Take off the collapsed flag at rxt */ rsm->r_flags &= ~BBR_RWND_COLLAPSED; } if (rsm->r_flags & BBR_MARKED_LOST) { /* We have retransmitted, its no longer lost */ rsm->r_flags &= ~BBR_MARKED_LOST; bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } if (rsm->r_flags & BBR_RXT_CLEARED) { /* * We hit a RXT timer on it and * we cleared the "acked" flag. * We now have it going back into * flight, we can remove the cleared * flag and possibly do accounting on * this piece. */ rsm->r_flags &= ~BBR_RXT_CLEARED; } if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) { bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); } idx = rsm->r_rtr_cnt - 1; rsm->r_tim_lastsent[idx] = cts; rsm->r_pacing_delay = pacing_time; rsm->r_delivered = bbr->r_ctl.rc_delivered; rsm->r_ts_valid = bbr->rc_ts_valid; if (bbr->rc_ts_valid) rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; if (bbr->r_ctl.r_app_limited_until) rsm->r_app_limited = 1; else rsm->r_app_limited = 0; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) rsm->r_bbr_state = bbr_state_val(bbr); else rsm->r_bbr_state = 8; if (rsm->r_flags & BBR_ACKED) { /* Problably MTU discovery messing with us */ uint32_t old_flags; old_flags = rsm->r_flags; rsm->r_flags &= ~BBR_ACKED; bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); if (bbr->r_ctl.rc_sacked == 0) bbr->r_ctl.rc_sacklast = NULL; } if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); } TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (rsm->r_flags & BBR_SACK_PASSED) { /* We have retransmitted due to the SACK pass */ rsm->r_flags &= ~BBR_SACK_PASSED; rsm->r_flags |= BBR_WAS_SACKPASS; } rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { rsm->r_is_gain = 1; rsm->r_is_drain = 0; } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { rsm->r_is_drain = 1; rsm->r_is_gain = 0; } else { rsm->r_is_drain = 0; rsm->r_is_gain = 0; } rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */ } /* * Returns 0, or the sequence where we stopped * updating. We also update the lenp to be the amount * of data left. */ static uint32_t bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time) { /* * We (re-)transmitted starting at rsm->r_start for some length * (possibly less than r_end. */ struct bbr_sendmap *nrsm; uint32_t c_end; int32_t len; len = *lenp; c_end = rsm->r_start + len; if (SEQ_GEQ(c_end, rsm->r_end)) { /* * We retransmitted the whole piece or more than the whole * slopping into the next rsm. */ bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); if (c_end == rsm->r_end) { *lenp = 0; return (0); } else { int32_t act_len; /* Hangs over the end return whats left */ act_len = rsm->r_end - rsm->r_start; *lenp = (len - act_len); return (rsm->r_end); } /* We don't get out of this block. */ } /* * Here we retransmitted less than the whole thing which means we * have to split this into what was transmitted and what was not. */ nrsm = bbr_alloc_full_limit(bbr); if (nrsm == NULL) { *lenp = 0; return (0); } /* * So here we are going to take the original rsm and make it what we * retransmitted. nrsm will be the tail portion we did not * retransmit. For example say the chunk was 1, 11 (10 bytes). And * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to * 1, 6 and the new piece will be 6, 11. */ bbr_clone_rsm(bbr, nrsm, rsm, c_end); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); nrsm->r_dupack = 0; if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); *lenp = 0; return (0); } static uint64_t bbr_get_hardware_rate(struct tcp_bbr *bbr) { uint64_t bw; bw = bbr_get_bw(bbr); bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]; bw /= (uint64_t)BBR_UNIT; return(bw); } static void bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts, uint64_t act_rate, uint64_t rate_wanted) { /* * We could not get a full gains worth * of rate. */ if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) { /* we can't even get the real rate */ uint64_t red; bbr->skip_gain = 1; bbr->gain_is_limited = 0; red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate; if (red) filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts); } else { /* We can use a lower gain */ bbr->skip_gain = 0; bbr->gain_is_limited = 1; } } static void bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts) { const struct tcp_hwrate_limit_table *nrte; int error, rate = -1; if (bbr->r_ctl.crte == NULL) return; if ((bbr->rc_inp->inp_route.ro_nh == NULL) || (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) { /* Lost our routes? */ /* Clear the way for a re-attempt */ bbr->bbr_attempt_hdwr_pace = 0; lost_rate: bbr->gain_is_limited = 0; bbr->skip_gain = 0; bbr->bbr_hdrw_pacing = 0; counter_u64_add(bbr_flows_whdwr_pacing, -1); counter_u64_add(bbr_flows_nohdwr_pacing, 1); tcp_bbr_tso_size_check(bbr, cts); return; } rate = bbr_get_hardware_rate(bbr); nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp, bbr->rc_inp->inp_route.ro_nh->nh_ifp, rate, (RS_PACING_GEQ|RS_PACING_SUB_OK), &error, NULL); if (nrte == NULL) { goto lost_rate; } if (nrte != bbr->r_ctl.crte) { bbr->r_ctl.crte = nrte; if (error == 0) { BBR_STAT_INC(bbr_hdwr_rl_mod_ok); if (bbr->r_ctl.crte->rate < rate) { /* We have a problem */ bbr_setup_less_of_rate(bbr, cts, bbr->r_ctl.crte->rate, rate); } else { /* We are good */ bbr->gain_is_limited = 0; bbr->skip_gain = 0; } } else { /* A failure should release the tag */ BBR_STAT_INC(bbr_hdwr_rl_mod_fail); bbr->gain_is_limited = 0; bbr->skip_gain = 0; bbr->bbr_hdrw_pacing = 0; } bbr_type_log_hdwr_pacing(bbr, bbr->r_ctl.crte->ptbl->rs_ifp, rate, ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate), __LINE__, cts, error); } } static void bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts) { /* * If we have hardware pacing support * we need to factor that in for our * TSO size. */ const struct tcp_hwrate_limit_table *rlp; uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay; if ((bbr->bbr_hdrw_pacing == 0) || (IN_RECOVERY(bbr->rc_tp->t_flags)) || (bbr->r_ctl.crte == NULL)) return; if (bbr->hw_pacing_set == 0) { /* Not yet by the hdwr pacing count delay */ return; } if (bbr_hdwr_pace_adjust == 0) { /* No adjustment */ return; } rlp = bbr->r_ctl.crte; if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; else maxseg = BBR_MIN_SEG - bbr->rc_last_options; /* * So lets first get the * time we will take between * TSO sized sends currently without * hardware help. */ cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT, bbr->r_ctl.rc_pace_max_segs, cts, 1); hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg; hdwr_delay *= rlp->time_between; if (cur_delay > hdwr_delay) delta = cur_delay - hdwr_delay; else delta = 0; bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay, (bbr->r_ctl.rc_pace_max_segs / maxseg), 1); if (delta && (delta < (max(rlp->time_between, bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) { /* * Now lets divide by the pacing * time between each segment the * hardware sends rounding up and * derive a bytes from that. We multiply * that by bbr_hdwr_pace_adjust to get * more bang for our buck. * * The goal is to have the software pacer * waiting no more than an additional * pacing delay if we can (without the * compensation i.e. x bbr_hdwr_pace_adjust). */ seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between), (bbr->r_ctl.rc_pace_max_segs/maxseg)); seg_sz *= bbr_hdwr_pace_adjust; if (bbr_hdwr_pace_floor && (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { /* Currently hardware paces * out rs_min_seg segments at a time. * We need to make sure we always send at least * a full burst of bbr_hdwr_pace_floor down. */ seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; } seg_sz *= maxseg; } else if (delta == 0) { /* * The highest pacing rate is * above our b/w gained. This means * we probably are going quite fast at * the hardware highest rate. Lets just multiply * the calculated TSO size by the * multiplier factor (its probably * 4 segments in the default config for * mlx). */ seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust; if (bbr_hdwr_pace_floor && (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { /* Currently hardware paces * out rs_min_seg segments at a time. * We need to make sure we always send at least * a full burst of bbr_hdwr_pace_floor down. */ seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; } } else { /* * The pacing time difference is so * big that the hardware will * pace out more rapidly then we * really want and then we * will have a long delay. Lets just keep * the same TSO size so its as if * we were not using hdwr pacing (we * just gain a bit of spacing from the * hardware if seg_sz > 1). */ seg_sz = bbr->r_ctl.rc_pace_max_segs; } if (seg_sz > bbr->r_ctl.rc_pace_max_segs) new_tso = seg_sz; else new_tso = bbr->r_ctl.rc_pace_max_segs; if (new_tso >= (PACE_MAX_IP_BYTES-maxseg)) new_tso = PACE_MAX_IP_BYTES - maxseg; if (new_tso != bbr->r_ctl.rc_pace_max_segs) { bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0); bbr->r_ctl.rc_pace_max_segs = new_tso; } } static void tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts) { uint64_t bw; uint32_t old_tso = 0, new_tso; uint32_t maxseg, bytes; uint32_t tls_seg=0; /* * Google/linux uses the following algorithm to determine * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18): * * bytes = bw_in_bytes_per_second / 1000 * bytes = min(bytes, 64k) * tso_segs = bytes / MSS * if (bw < 1.2Mbs) * min_tso_segs = 1 * else * min_tso_segs = 2 * tso_segs = max(tso_segs, min_tso_segs) * * * Note apply a device specific limit (we apply this in the * tcp_m_copym). * Note that before the initial measurement is made google bursts out * a full iwnd just like new-reno/cubic. * * We do not use this algorithm. Instead we * use a two phased approach: * * if ( bw <= per-tcb-cross-over) * goal_tso = calculate how much with this bw we * can send in goal-time seconds. * if (goal_tso > mss) * seg = goal_tso / mss * tso = seg * mss * else * tso = mss * if (tso > per-tcb-max) * tso = per-tcb-max * else if ( bw > 512Mbps) * tso = max-tso (64k/mss) * else * goal_tso = bw / per-tcb-divsor * seg = (goal_tso + mss-1)/mss * tso = seg * mss * * if (tso < per-tcb-floor) * tso = per-tcb-floor * if (tso > per-tcb-utter_max) * tso = per-tcb-utter_max * * Note the default per-tcb-divisor is 1000 (same as google). * the goal cross over is 30Mbps however. To recreate googles * algorithm you need to set: * * cross-over = 23,168,000 bps * goal-time = 18000 * per-tcb-max = 2 * per-tcb-divisor = 1000 * per-tcb-floor = 1 * * This will get you "google bbr" behavior with respect to tso size. * * Note we do set anything TSO size until we are past the initial * window. Before that we gnerally use either a single MSS * or we use the full IW size (so we burst a IW at a time) */ if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) { maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; } else { maxseg = BBR_MIN_SEG - bbr->rc_last_options; } old_tso = bbr->r_ctl.rc_pace_max_segs; if (bbr->rc_past_init_win == 0) { /* * Not enough data has been acknowledged to make a * judgement. Set up the initial TSO based on if we * are sending a full IW at once or not. */ if (bbr->rc_use_google) bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2); else if (bbr->bbr_init_win_cheat) bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp); else bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options; if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg) bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg; if (bbr->r_ctl.rc_pace_max_segs == 0) { bbr->r_ctl.rc_pace_max_segs = maxseg; } bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0); bbr_adjust_for_hw_pacing(bbr, cts); return; } /** * Now lets set the TSO goal based on our delivery rate in * bytes per second. Note we only do this if * we have acked at least the initial cwnd worth of data. */ bw = bbr_get_bw(bbr); if (IN_RECOVERY(bbr->rc_tp->t_flags) && (bbr->rc_use_google == 0)) { /* We clamp to one MSS in recovery */ new_tso = maxseg; } else if (bbr->rc_use_google) { int min_tso_segs; /* Google considers the gain too */ if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) { bw *= bbr->r_ctl.rc_bbr_hptsi_gain; bw /= BBR_UNIT; } bytes = bw / 1024; if (bytes > (64 * 1024)) bytes = 64 * 1024; new_tso = bytes / maxseg; if (bw < ONE_POINT_TWO_MEG) min_tso_segs = 1; else min_tso_segs = 2; if (new_tso < min_tso_segs) new_tso = min_tso_segs; new_tso *= maxseg; } else if (bbr->rc_no_pacing) { new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg; } else if (bw <= bbr->r_ctl.bbr_cross_over) { /* * Calculate the worse case b/w TSO if we are inserting no * more than a delay_target number of TSO's. */ uint32_t tso_len, min_tso; tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw); if (tso_len > maxseg) { new_tso = tso_len / maxseg; if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max) new_tso = bbr->r_ctl.bbr_hptsi_segments_max; new_tso *= maxseg; } else { /* * less than a full sized frame yikes.. long rtt or * low bw? */ min_tso = bbr_minseg(bbr); if ((tso_len > min_tso) && (bbr_all_get_min == 0)) new_tso = rounddown(tso_len, min_tso); else new_tso = min_tso; } } else if (bw > FIVETWELVE_MBPS) { /* * This guy is so fast b/w wise that we can TSO as large as * possible of segments that the NIC will allow. */ new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); } else { /* * This formula is based on attempting to send a segment or * more every bbr_hptsi_per_second. The default is 1000 * which means you are targeting what you can send every 1ms * based on the peers bw. * * If the number drops to say 500, then you are looking more * at 2ms and you will raise how much we send in a single * TSO thus saving CPU (less bbr_output_wtime() calls). The * trade off of course is you will send more at once and * thus tend to clump up the sends into larger "bursts" * building a queue. */ bw /= bbr->r_ctl.bbr_hptsi_per_second; new_tso = roundup(bw, (uint64_t)maxseg); /* * Gate the floor to match what our lower than 48Mbps * algorithm does. The ceiling (bbr_hptsi_segments_max) thus * becomes the floor for this calculation. */ if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg)) new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg); } if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor))) new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor; if (new_tso > PACE_MAX_IP_BYTES) new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); /* Enforce an utter maximum. */ if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) { new_tso = bbr->r_ctl.bbr_utter_max * maxseg; } if (old_tso != new_tso) { /* Only log changes */ bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0); bbr->r_ctl.rc_pace_max_segs = new_tso; } /* We have hardware pacing! */ bbr_adjust_for_hw_pacing(bbr, cts); } static void bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t cts, struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc, struct sockbuf *sb) { struct bbr_sendmap *rsm, *nrsm; register uint32_t snd_max, snd_una; uint32_t pacing_time; /* * Add to the RACK log of packets in flight or retransmitted. If * there is a TS option we will use the TS echoed, if not we will * grab a TS. * * Retransmissions will increment the count and move the ts to its * proper place. Note that if options do not include TS's then we * won't be able to effectively use the ACK for an RTT on a retran. * * Notes about r_start and r_end. Lets consider a send starting at * sequence 1 for 10 bytes. In such an example the r_start would be * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. * This means that r_end is actually the first sequence for the next * slot (11). * */ INP_WLOCK_ASSERT(tp->t_inpcb); if (err) { /* * We don't log errors -- we could but snd_max does not * advance in this case either. */ return; } if (th_flags & TH_RST) { /* * We don't log resets and we return immediately from * sending */ *abandon = 1; return; } snd_una = tp->snd_una; if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) { /* * The call to bbr_log_output is made before bumping * snd_max. This means we can record one extra byte on a SYN * or FIN if seq_out is adding more on and a FIN is present * (and we are not resending). */ if ((th_flags & TH_SYN) && (tp->iss == seq_out)) len++; if (th_flags & TH_FIN) len++; } if (SEQ_LEQ((seq_out + len), snd_una)) { /* Are sending an old segment to induce an ack (keep-alive)? */ return; } if (SEQ_LT(seq_out, snd_una)) { /* huh? should we panic? */ uint32_t end; end = seq_out + len; seq_out = snd_una; len = end - seq_out; } snd_max = tp->snd_max; if (len == 0) { /* We don't log zero window probes */ return; } pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1); /* First question is it a retransmission? */ if (seq_out == snd_max) { again: rsm = bbr_alloc(bbr); if (rsm == NULL) { return; } rsm->r_flags = 0; if (th_flags & TH_SYN) rsm->r_flags |= BBR_HAS_SYN; if (th_flags & TH_FIN) rsm->r_flags |= BBR_HAS_FIN; rsm->r_tim_lastsent[0] = cts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; rsm->r_start = seq_out; rsm->r_end = rsm->r_start + len; rsm->r_dupack = 0; rsm->r_delivered = bbr->r_ctl.rc_delivered; rsm->r_pacing_delay = pacing_time; rsm->r_ts_valid = bbr->rc_ts_valid; if (bbr->rc_ts_valid) rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; rsm->r_del_time = bbr->r_ctl.rc_del_time; if (bbr->r_ctl.r_app_limited_until) rsm->r_app_limited = 1; else rsm->r_app_limited = 0; rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); /* * Here we must also add in this rsm since snd_max * is updated after we return from a new send. */ rsm->r_flight_at_send += len; TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) rsm->r_bbr_state = bbr_state_val(bbr); else rsm->r_bbr_state = 8; if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { rsm->r_is_gain = 1; rsm->r_is_drain = 0; } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { rsm->r_is_drain = 1; rsm->r_is_gain = 0; } else { rsm->r_is_drain = 0; rsm->r_is_gain = 0; } return; } /* * If we reach here its a retransmission and we need to find it. */ more: if (hintrsm && (hintrsm->r_start == seq_out)) { rsm = hintrsm; hintrsm = NULL; } else if (bbr->r_ctl.rc_next) { /* We have a hint from a previous run */ rsm = bbr->r_ctl.rc_next; } else { /* No hints sorry */ rsm = NULL; } if ((rsm) && (rsm->r_start == seq_out)) { /* * We used rc_next or hintrsm to retransmit, hopefully the * likely case. */ seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); if (len == 0) { return; } else { goto more; } } /* Ok it was not the last pointer go through it the hard way. */ TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { if (rsm->r_start == seq_out) { seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); if (len == 0) { return; } else { continue; } } if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { /* Transmitted within this piece */ /* * Ok we must split off the front and then let the * update do the rest */ nrsm = bbr_alloc_full_limit(bbr); if (nrsm == NULL) { bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); return; } /* * copy rsm to nrsm and then trim the front of rsm * to not include this part. */ bbr_clone_rsm(bbr, nrsm, rsm, seq_out); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time); if (len == 0) { return; } } } /* * Hmm not found in map did they retransmit both old and on into the * new? */ if (seq_out == tp->snd_max) { goto again; } else if (SEQ_LT(seq_out, tp->snd_max)) { #ifdef BBR_INVARIANTS printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", seq_out, len, tp->snd_una, tp->snd_max); printf("Starting Dump of all rack entries\n"); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { printf("rsm:%p start:%u end:%u\n", rsm, rsm->r_start, rsm->r_end); } printf("Dump complete\n"); panic("seq_out not found rack:%p tp:%p", bbr, tp); #endif } else { #ifdef BBR_INVARIANTS /* * Hmm beyond sndmax? (only if we are using the new rtt-pack * flag) */ panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", seq_out, len, tp->snd_max, tp); #endif } } static void bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt) { /* * Collapse timeout back the cum-ack moved. */ tp->t_rxtshift = 0; tp->t_softerror = 0; } static void tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin) { bbr->rtt_valid = 1; bbr->r_ctl.cur_rtt = rtt_usecs; bbr->r_ctl.ts_in = tsin; if (rsm_send_time) bbr->r_ctl.cur_rtt_send_time = rsm_send_time; } static void bbr_make_timestamp_determination(struct tcp_bbr *bbr) { /** * We have in our bbr control: * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp). * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts). * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts) * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time) * * Now we can calculate the time between the sends by doing: * * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts * * And the peer's time between receiving them by doing: * * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp * * We want to figure out if the timestamp values are in msec, 10msec or usec. * We also may find that we can't use the timestamps if say we see * that the peer_delta indicates that though we may have taken 10ms to * pace out the data, it only saw 1ms between the two packets. This would * indicate that somewhere on the path is a batching entity that is giving * out time-slices of the actual b/w. This would mean we could not use * reliably the peers timestamps. * * We expect delta > peer_delta initially. Until we figure out the * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio. * If we place 1000 there then its a ms vs our usec. If we place 10000 there * then its 10ms vs our usec. If the peer is running a usec clock we would * put a 1 there. If the value is faster then ours, we will disable the * use of timestamps (though we could revist this later if we find it to be not * just an isolated one or two flows)). * * To detect the batching middle boxes we will come up with our compensation and * if with it in place, we find the peer is drastically off (by some margin) in * the smaller direction, then we will assume the worst case and disable use of timestamps. * */ uint64_t delta, peer_delta, delta_up; delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts; if (delta < bbr_min_usec_delta) { /* * Have not seen a min amount of time * between our send times so we can * make a determination of the timestamp * yet. */ return; } peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp; if (peer_delta < bbr_min_peer_delta) { /* * We may have enough in the form of * our delta but the peers number * has not changed that much. It could * be its clock ratio is such that * we need more data (10ms tick) or * there may be other compression scenarios * going on. In any event we need the * spread to be larger. */ return; } /* Ok lets first see which way our delta is going */ if (peer_delta > delta) { /* Very unlikely, the peer without * compensation shows that it saw * the two sends arrive further apart * then we saw then in micro-seconds. */ if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) { /* well it looks like the peer is a micro-second clock. */ bbr->rc_ts_clock_set = 1; bbr->r_ctl.bbr_peer_tsratio = 1; } else { bbr->rc_ts_cant_be_used = 1; bbr->rc_ts_clock_set = 1; } return; } /* Ok we know that the peer_delta is smaller than our send distance */ bbr->rc_ts_clock_set = 1; /* First question is it within the percentage that they are using usec time? */ delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent; if ((peer_delta + delta_up) >= delta) { /* Its a usec clock */ bbr->r_ctl.bbr_peer_tsratio = 1; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* Ok if not usec, what about 10usec (though unlikely)? */ delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent; if (((peer_delta * 10) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 10; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* And what about 100usec (though again unlikely)? */ delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent; if (((peer_delta * 100) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 100; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* And how about 1 msec (the most likely one)? */ delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent; if (((peer_delta * 1000) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 1000; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* Ok if not msec could it be 10 msec? */ delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent; if (((peer_delta * 10000) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 10000; return; } /* If we fall down here the clock tick so slowly we can't use it */ bbr->rc_ts_cant_be_used = 1; bbr->r_ctl.bbr_peer_tsratio = 0; bbr_log_tstmp_validation(bbr, peer_delta, delta); } /* * Collect new round-trip time estimate * and update averages and current timeout. */ static void tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts) { int32_t delta; uint32_t rtt, tsin; int32_t rtt_ticks; if (bbr->rtt_valid == 0) /* No valid sample */ return; rtt = bbr->r_ctl.cur_rtt; tsin = bbr->r_ctl.ts_in; if (bbr->rc_prtt_set_ts) { /* * We are to force feed the rttProp filter due * to an entry into PROBE_RTT. This assures * that the times are sync'd between when we * go into PROBE_RTT and the filter expiration. * * Google does not use a true filter, so they do * this implicitly since they only keep one value * and when they enter probe-rtt they update the * value to the newest rtt. */ uint32_t rtt_prop; bbr->rc_prtt_set_ts = 0; rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (rtt > rtt_prop) filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts); else apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); } if (bbr->rc_ack_was_delayed) rtt += bbr->r_ctl.rc_ack_hdwr_delay; if (rtt < bbr->r_ctl.rc_lowest_rtt) bbr->r_ctl.rc_lowest_rtt = rtt; bbr_log_rtt_sample(bbr, rtt, tsin); if (bbr->r_init_rtt) { /* * The initial rtt is not-trusted, nuke it and lets get * our first valid measurement in. */ bbr->r_init_rtt = 0; tp->t_srtt = 0; } if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) { /* * So we have not yet figured out * what the peers TSTMP value is * in (most likely ms). We need a * series of cum-ack's to determine * this reliably. */ if (bbr->rc_ack_is_cumack) { if (bbr->rc_ts_data_set) { /* Lets attempt to determine the timestamp granularity. */ bbr_make_timestamp_determination(bbr); } else { bbr->rc_ts_data_set = 1; bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts; bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time; } } else { /* * We have to have consecutive acks * reset any "filled" state to none. */ bbr->rc_ts_data_set = 0; } } /* Round it up */ rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1))); if (rtt_ticks == 0) rtt_ticks = 1; if (tp->t_srtt != 0) { /* * srtt is stored as fixed point with 5 bits after the * binary point (i.e., scaled by 8). The following magic is * equivalent to the smoothing algorithm in rfc793 with an * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). * Adjust rtt to origin 0. */ delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT) - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); tp->t_srtt += delta; if (tp->t_srtt <= 0) tp->t_srtt = 1; /* * We accumulate a smoothed rtt variance (actually, a * smoothed mean difference), then set the retransmit timer * to smoothed rtt + 4 times the smoothed variance. rttvar * is stored as fixed point with 4 bits after the binary * point (scaled by 16). The following is equivalent to * rfc793 smoothing with an alpha of .75 (rttvar = * rttvar*3/4 + |delta| / 4). This replaces rfc793's * wired-in beta. */ if (delta < 0) delta = -delta; delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); tp->t_rttvar += delta; if (tp->t_rttvar <= 0) tp->t_rttvar = 1; if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } else { /* * No rtt measurement yet - use the unsmoothed rtt. Set the * variance to half the rtt (so our first retransmit happens * at 3*rtt). */ tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT; tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1); tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } KMOD_TCPSTAT_INC(tcps_rttupdated); tp->t_rttupdated++; #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks)); #endif /* * the retransmit should happen at rtt + 4 * rttvar. Because of the * way we do the smoothing, srtt and rttvar will each average +1/2 * tick of bias. When we compute the retransmit timer, we want 1/2 * tick of rounding and 1 extra tick because of +-1/2 tick * uncertainty in the firing of the timer. The bias will give us * exactly the 1.5 tick we need. But, because the bias is * statistical, we have to test that we don't drop below the minimum * feasible timer (which is 2 ticks). */ TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2), MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); /* * We received an ack for a packet that wasn't retransmitted; it is * probably safe to discard any error indications we've received * recently. This isn't quite right, but close enough for now (a * route might have failed after we sent a segment, and the return * path might not be symmetrical). */ tp->t_softerror = 0; rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt) bbr->r_ctl.bbr_smallest_srtt_this_state = rtt; } static void bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line) { bbr->r_ctl.rc_rtt_shrinks = cts; if (bbr_can_force_probertt && (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { /* * We should enter probe-rtt its been too long * since we have been there. */ bbr_enter_probe_rtt(bbr, cts, __LINE__); } else bbr_check_probe_rtt_limits(bbr, cts); } static void tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts) { uint64_t orig_bw; if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) { /* We never apply a zero measurment */ bbr_log_type_bbrupd(bbr, 20, cts, 0, 0, 0, 0, 0, 0, 0, 0); return; } if (bbr->r_ctl.r_measurement_count < 0xffffffff) bbr->r_ctl.r_measurement_count++; orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate); apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch); bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw, (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate), 0, 0, 0, 0, 0, 0); if (orig_bw && (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) { if (bbr->bbr_hdrw_pacing) { /* * Apply a new rate to the hardware * possibly. */ bbr_update_hardware_pacing_rate(bbr, cts); } bbr_set_state_target(bbr, __LINE__); tcp_bbr_tso_size_check(bbr, cts); if (bbr->r_recovery_bw) { bbr_setup_red_bw(bbr, cts); bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW); } } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate)) tcp_bbr_tso_size_check(bbr, cts); } static void bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) { if (bbr->rc_in_persist == 0) { /* We log only when not in persist */ /* Translate to a Bytes Per Second */ uint64_t tim, bw, ts_diff, ts_bw; uint32_t upper, lower, delivered; if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); else tim = 1; /* * Now that we have processed the tim (skipping the sample * or possibly updating the time, go ahead and * calculate the cdr. */ delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); bw = (uint64_t)delivered; bw *= (uint64_t)USECS_IN_SECOND; bw /= tim; if (bw == 0) { /* We must have a calculatable amount */ return; } upper = (bw >> 32) & 0x00000000ffffffff; lower = bw & 0x00000000ffffffff; /* * If we are using this b/w shove it in now so we * can see in the trace viewer if it gets over-ridden. */ if (rsm->r_ts_valid && bbr->rc_ts_valid && bbr->rc_ts_clock_set && (bbr->rc_ts_cant_be_used == 0) && bbr->rc_use_ts_limit) { ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1); ts_diff *= bbr->r_ctl.bbr_peer_tsratio; if ((delivered == 0) || (rtt < 1000)) { /* Can't use the ts */ bbr_log_type_bbrupd(bbr, 61, cts, ts_diff, bbr->r_ctl.last_inbound_ts, rsm->r_del_ack_ts, 0, 0, 0, 0, delivered); } else { ts_bw = (uint64_t)delivered; ts_bw *= (uint64_t)USECS_IN_SECOND; ts_bw /= ts_diff; bbr_log_type_bbrupd(bbr, 62, cts, (ts_bw >> 32), (ts_bw & 0xffffffff), 0, 0, 0, 0, ts_diff, delivered); if ((bbr->ts_can_raise) && (ts_bw > bw)) { bbr_log_type_bbrupd(bbr, 8, cts, delivered, ts_diff, (bw >> 32), (bw & 0x00000000ffffffff), 0, 0, 0, 0); bw = ts_bw; } else if (ts_bw && (ts_bw < bw)) { bbr_log_type_bbrupd(bbr, 7, cts, delivered, ts_diff, (bw >> 32), (bw & 0x00000000ffffffff), 0, 0, 0, 0); bw = ts_bw; } } } if (rsm->r_first_sent_time && TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { uint64_t sbw, sti; /* * We use what was in flight at the time of our * send and the size of this send to figure * out what we have been sending at (amount). * For the time we take from the time of * the send of the first send outstanding * until this send plus this sends pacing * time. This gives us a good calculation * as to the rate we have been sending at. */ sbw = (uint64_t)(rsm->r_flight_at_send); sbw *= (uint64_t)USECS_IN_SECOND; sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; sti += rsm->r_pacing_delay; sbw /= sti; if (sbw < bw) { bbr_log_type_bbrupd(bbr, 6, cts, delivered, (uint32_t)sti, (bw >> 32), (uint32_t)bw, rsm->r_first_sent_time, 0, (sbw >> 32), (uint32_t)sbw); bw = sbw; } } /* Use the google algorithm for b/w measurements */ bbr->r_ctl.rc_bbr_cur_del_rate = bw; if ((rsm->r_app_limited == 0) || (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) { tcp_bbr_commit_bw(bbr, cts); bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); } } } static void bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) { if (bbr->rc_in_persist == 0) { /* We log only when not in persist */ /* Translate to a Bytes Per Second */ uint64_t tim, bw; uint32_t upper, lower, delivered; int no_apply = 0; if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); else tim = 1; /* * Now that we have processed the tim (skipping the sample * or possibly updating the time, go ahead and * calculate the cdr. */ delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); bw = (uint64_t)delivered; bw *= (uint64_t)USECS_IN_SECOND; bw /= tim; if (tim < bbr->r_ctl.rc_lowest_rtt) { bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); no_apply = 1; } upper = (bw >> 32) & 0x00000000ffffffff; lower = bw & 0x00000000ffffffff; /* * If we are using this b/w shove it in now so we * can see in the trace viewer if it gets over-ridden. */ bbr->r_ctl.rc_bbr_cur_del_rate = bw; /* Gate by the sending rate */ if (rsm->r_first_sent_time && TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { uint64_t sbw, sti; /* * We use what was in flight at the time of our * send and the size of this send to figure * out what we have been sending at (amount). * For the time we take from the time of * the send of the first send outstanding * until this send plus this sends pacing * time. This gives us a good calculation * as to the rate we have been sending at. */ sbw = (uint64_t)(rsm->r_flight_at_send); sbw *= (uint64_t)USECS_IN_SECOND; sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; sti += rsm->r_pacing_delay; sbw /= sti; if (sbw < bw) { bbr_log_type_bbrupd(bbr, 6, cts, delivered, (uint32_t)sti, (bw >> 32), (uint32_t)bw, rsm->r_first_sent_time, 0, (sbw >> 32), (uint32_t)sbw); bw = sbw; } if ((sti > tim) && (sti < bbr->r_ctl.rc_lowest_rtt)) { bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); no_apply = 1; } else no_apply = 0; } bbr->r_ctl.rc_bbr_cur_del_rate = bw; if ((no_apply == 0) && ((rsm->r_app_limited == 0) || (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) { tcp_bbr_commit_bw(bbr, cts); bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); } } } static void bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin, uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to) { uint64_t old_rttprop; /* Update our delivery time and amount */ bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_del_time = cts; if (rtt == 0) { /* * 0 means its a retransmit, for now we don't use these for * the rest of BBR. */ return; } if ((bbr->rc_use_google == 0) && (match != BBR_RTT_BY_EXACTMATCH) && (match != BBR_RTT_BY_TIMESTAMP)){ /* * We get a lot of rtt updates, lets not pay attention to * any that are not an exact match. That way we don't have * to worry about timestamps and the whole nonsense of * unsure if its a retransmission etc (if we ever had the * timestamp fixed to always have the last thing sent this * would not be a issue). */ return; } if ((bbr_no_retran && bbr->rc_use_google) && (match != BBR_RTT_BY_EXACTMATCH) && (match != BBR_RTT_BY_TIMESTAMP)){ /* * We only do measurements in google mode * with bbr_no_retran on for sure things. */ return; } /* Only update srtt if we know by exact match */ tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin); if (ack_type == BBR_CUM_ACKED) bbr->rc_ack_is_cumack = 1; else bbr->rc_ack_is_cumack = 0; old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP); /* * Note the following code differs to the original * BBR spec. It calls for <= not <. However after a * long discussion in email with Neal, he acknowledged * that it should be < than so that we will have flows * going into probe-rtt (we were seeing cases where that * did not happen and caused ugly things to occur). We * have added this agreed upon fix to our code base. */ if (rtt < old_rttprop) { /* Update when we last saw a rtt drop */ bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts, match, rsm->r_start, rsm->r_flags); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) { /* * The RTT-prop moved, reset the target (may be a * nop for some states). */ bbr_set_state_target(bbr, __LINE__); if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_NEW_TARGET, 0); else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP)) /* It went up */ bbr_check_probe_rtt_limits(bbr, cts); } if ((bbr->rc_use_google == 0) && (match == BBR_RTT_BY_TIMESTAMP)) { /* * We don't do b/w update with * these since they are not really * reliable. */ return; } if (bbr->r_ctl.r_app_limited_until && (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) { /* We are no longer app-limited */ bbr->r_ctl.r_app_limited_until = 0; } if (bbr->rc_use_google) { bbr_google_measurement(bbr, rsm, rtt, cts); } else { bbr_nf_measurement(bbr, rsm, rtt, cts); } } /* * Convert a timestamp that the main stack * uses (milliseconds) into one that bbr uses * (microseconds). Return that converted timestamp. */ static uint32_t bbr_ts_convert(uint32_t cts) { uint32_t sec, msec; sec = cts / MS_IN_USEC; msec = cts - (MS_IN_USEC * sec); return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC)); } /* * Return 0 if we did not update the RTT time, return * 1 if we did. */ static int bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack) { int32_t i; uint32_t t, uts = 0; if ((rsm->r_flags & BBR_ACKED) || (rsm->r_flags & BBR_WAS_RENEGED) || (rsm->r_flags & BBR_RXT_CLEARED)) { /* Already done */ return (0); } if (rsm->r_rtt_not_allowed) { /* Not allowed */ return (0); } if (rsm->r_rtr_cnt == 1) { /* * Only one transmit. Hopefully the normal case. */ if (TSTMP_GT(cts, rsm->r_tim_lastsent[0])) t = cts - rsm->r_tim_lastsent[0]; else t = 1; if ((int)t <= 0) t = 1; bbr->r_ctl.rc_last_rtt = t; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to); return (1); } /* Convert to usecs */ if ((bbr_can_use_ts_for_rtt == 1) && (bbr->rc_use_google == 1) && (ack_type == BBR_CUM_ACKED) && (to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr; if (t < 1) t = 1; t *= MS_IN_USEC; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, BBR_RTT_BY_TIMESTAMP, rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)], ack_type, to); return (1); } uts = bbr_ts_convert(to->to_tsecr); if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0) && (ack_type == BBR_CUM_ACKED) && ((rsm->r_flags & BBR_OVERMAX) == 0)) { /* * Now which timestamp does it match? In this block the ACK * may be coming from a previous transmission. */ uint32_t fudge; fudge = BBR_TIMER_FUDGE; for (i = 0; i < rsm->r_rtr_cnt; i++) { if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) && (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) { if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) t = cts - rsm->r_tim_lastsent[i]; else t = 1; if ((int)t <= 0) t = 1; bbr->r_ctl.rc_last_rtt = t; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING, rsm->r_tim_lastsent[i], ack_type, to); if ((i + 1) < rsm->r_rtr_cnt) { /* Likely */ return (0); } else if (rsm->r_flags & BBR_TLP) { bbr->rc_tlp_rtx_out = 0; } return (1); } } /* Fall through if we can't find a matching timestamp */ } /* * Ok its a SACK block that we retransmitted. or a windows * machine without timestamps. We can tell nothing from the * time-stamp since its not there or the time the peer last * recieved a segment that moved forward its cum-ack point. * * Lets look at the last retransmit and see what we can tell * (with BBR for space we only keep 2 note we have to keep * at least 2 so the map can not be condensed more). */ i = rsm->r_rtr_cnt - 1; if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) t = cts - rsm->r_tim_lastsent[i]; else goto not_sure; if (t < bbr->r_ctl.rc_lowest_rtt) { /* * We retransmitted and the ack came back in less * than the smallest rtt we have observed in the * windowed rtt. We most likey did an improper * retransmit as outlined in 4.2 Step 3 point 2 in * the rack-draft. * * Use the prior transmission to update all the * information as long as there is only one prior * transmission. */ if ((rsm->r_flags & BBR_OVERMAX) == 0) { #ifdef BBR_INVARIANTS if (rsm->r_rtr_cnt == 1) panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags); #endif i = rsm->r_rtr_cnt - 2; if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) t = cts - rsm->r_tim_lastsent[i]; else t = 1; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET, rsm->r_tim_lastsent[i], ack_type, to); return (0); } else { /* * Too many prior transmissions, just * updated BBR delivered */ not_sure: bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); } } else { /* * We retransmitted it and the retransmit did the * job. */ if (rsm->r_flags & BBR_TLP) bbr->rc_tlp_rtx_out = 0; if ((rsm->r_flags & BBR_OVERMAX) == 0) bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to); else bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); return (1); } return (0); } /* * Mark the SACK_PASSED flag on all entries prior to rsm send wise. */ static void bbr_log_sack_passed(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm) { struct bbr_sendmap *nrsm; nrsm = rsm; TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap, bbr_head, r_tnext) { if (nrsm == rsm) { /* Skip orginal segment he is acked */ continue; } if (nrsm->r_flags & BBR_ACKED) { /* Skip ack'd segments */ continue; } if (nrsm->r_flags & BBR_SACK_PASSED) { /* * We found one that is already marked * passed, we have been here before and * so all others below this are marked. */ break; } BBR_STAT_INC(bbr_sack_passed); nrsm->r_flags |= BBR_SACK_PASSED; if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) && bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) { bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start; bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start; nrsm->r_flags |= BBR_MARKED_LOST; } nrsm->r_flags &= ~BBR_WAS_SACKPASS; } } /* * Returns the number of bytes that were * newly ack'd by sack blocks. */ static uint32_t bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack, struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts) { int32_t times = 0; uint32_t start, end, maxseg, changed = 0; struct bbr_sendmap *rsm, *nrsm; int32_t used_ref = 1; uint8_t went_back = 0, went_fwd = 0; maxseg = tp->t_maxseg - bbr->rc_last_options; start = sack->start; end = sack->end; rsm = *prsm; if (rsm == NULL) used_ref = 0; /* Do we locate the block behind where we last were? */ if (rsm && SEQ_LT(start, rsm->r_start)) { went_back = 1; TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if (SEQ_GEQ(start, rsm->r_start) && SEQ_LT(start, rsm->r_end)) { goto do_rest_ofb; } } } start_at_beginning: went_fwd = 1; /* * Ok lets locate the block where this guy is fwd from rsm (if its * set) */ TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) { if (SEQ_GEQ(start, rsm->r_start) && SEQ_LT(start, rsm->r_end)) { break; } } do_rest_ofb: if (rsm == NULL) { /* * This happens when we get duplicate sack blocks with the * same end. For example SACK 4: 100 SACK 3: 100 The sort * will not change there location so we would just start at * the end of the first one and get lost. */ if (tp->t_flags & TF_SENTFIN) { /* * Check to see if we have not logged the FIN that * went out. */ nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { /* * Ok we did not get the FIN logged. */ nrsm->r_end++; rsm = nrsm; goto do_rest_ofb; } } if (times == 1) { #ifdef BBR_INVARIANTS panic("tp:%p bbr:%p sack:%p to:%p prsm:%p", tp, bbr, sack, to, prsm); #else goto out; #endif } times++; BBR_STAT_INC(bbr_sack_proc_restart); rsm = NULL; goto start_at_beginning; } /* Ok we have an ACK for some piece of rsm */ if (rsm->r_start != start) { /* * Need to split this in two pieces the before and after. */ if (bbr_sack_mergable(rsm, start, end)) nrsm = bbr_alloc_full_limit(bbr); else nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* We could not allocate ignore the sack */ struct sackblk blk; blk.start = start; blk.end = end; sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); goto out; } bbr_clone_rsm(bbr, nrsm, rsm, start); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); rsm = nrsm; } if (SEQ_GEQ(end, rsm->r_end)) { /* * The end of this block is either beyond this guy or right * at this guy. */ if ((rsm->r_flags & BBR_ACKED) == 0) { bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); changed += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); bbr_log_sack_passed(tp, bbr, rsm); if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } /* Is Reordering occuring? */ if (rsm->r_flags & BBR_SACK_PASSED) { BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } } rsm->r_flags |= BBR_ACKED; rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } } bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); if (end == rsm->r_end) { /* This block only - done */ goto out; } /* There is more not coverend by this rsm move on */ start = rsm->r_end; nrsm = TAILQ_NEXT(rsm, r_next); rsm = nrsm; times = 0; goto do_rest_ofb; } if (rsm->r_flags & BBR_ACKED) { /* Been here done that */ goto out; } /* Ok we need to split off this one at the tail */ if (bbr_sack_mergable(rsm, start, end)) nrsm = bbr_alloc_full_limit(bbr); else nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* failed XXXrrs what can we do but loose the sack info? */ struct sackblk blk; blk.start = start; blk.end = end; sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); goto out; } /* Clone it */ bbr_clone_rsm(bbr, nrsm, rsm, end); /* The sack block does not cover this guy fully */ rsm->r_flags &= (~BBR_HAS_FIN); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } nrsm->r_dupack = 0; bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); changed += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); bbr_log_sack_passed(tp, bbr, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } if (rsm->r_flags & BBR_SACK_PASSED) { BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } } rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); rsm->r_flags |= BBR_ACKED; if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } out: if (rsm && (rsm->r_flags & BBR_ACKED)) { /* * Now can we merge this newly acked * block with either the previous or * next block? */ nrsm = TAILQ_NEXT(rsm, r_next); if (nrsm && (nrsm->r_flags & BBR_ACKED)) { /* yep this and next can be merged */ rsm = bbr_merge_rsm(bbr, rsm, nrsm); } /* Now what about the previous? */ nrsm = TAILQ_PREV(rsm, bbr_head, r_next); if (nrsm && (nrsm->r_flags & BBR_ACKED)) { /* yep the previous and this can be merged */ rsm = bbr_merge_rsm(bbr, nrsm, rsm); } } if (used_ref == 0) { BBR_STAT_INC(bbr_sack_proc_all); } else { BBR_STAT_INC(bbr_sack_proc_short); } if (went_fwd && went_back) { BBR_STAT_INC(bbr_sack_search_both); } else if (went_fwd) { BBR_STAT_INC(bbr_sack_search_fwd); } else if (went_back) { BBR_STAT_INC(bbr_sack_search_back); } /* Save off where the next seq is */ if (rsm) bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); else bbr->r_ctl.rc_sacklast = NULL; *prsm = rsm; return (changed); } static void inline bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack) { struct bbr_sendmap *tmap; BBR_STAT_INC(bbr_reneges_seen); tmap = NULL; while (rsm && (rsm->r_flags & BBR_ACKED)) { /* Its no longer sacked, mark it so */ uint32_t oflags; bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); #ifdef BBR_INVARIANTS if (rsm->r_in_tmap) { panic("bbr:%p rsm:%p flags:0x%x in tmap?", bbr, rsm, rsm->r_flags); } #endif oflags = rsm->r_flags; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST); rsm->r_flags |= BBR_WAS_RENEGED; rsm->r_flags |= BBR_RXT_CLEARED; bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__); /* Rebuild it into our tmap */ if (tmap == NULL) { TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); tmap = rsm; } else { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext); tmap = rsm; } tmap->r_in_tmap = 1; /* * XXXrrs Delivered? Should we do anything here? * * Of course we don't on a rxt timeout so maybe its ok that * we don't? * * For now lets not. */ rsm = TAILQ_NEXT(rsm, r_next); } /* * Now lets possibly clear the sack filter so we start recognizing * sacks that cover this area. */ sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack); } static void bbr_log_syn(struct tcpcb *tp, struct tcpopt *to) { struct tcp_bbr *bbr; struct bbr_sendmap *rsm; uint32_t cts; bbr = (struct tcp_bbr *)tp->t_fb_ptr; cts = bbr->r_ctl.rc_rcvtime; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm && (rsm->r_flags & BBR_HAS_SYN)) { if ((rsm->r_end - rsm->r_start) <= 1) { /* Log out the SYN completely */ bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } if (bbr->r_ctl.rc_next == rsm) { /* scoot along the marker */ bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); } if (to != NULL) bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0); bbr_free(bbr, rsm); } else { /* There is more (Fast open)? strip out SYN. */ rsm->r_flags &= ~BBR_HAS_SYN; rsm->r_start++; } } } /* * Returns the number of bytes that were * acknowledged by SACK blocks. */ static uint32_t bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, uint32_t *prev_acked) { uint32_t changed, last_seq, entered_recovery = 0; struct tcp_bbr *bbr; struct bbr_sendmap *rsm; struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; register uint32_t th_ack; int32_t i, j, k, new_sb, num_sack_blks = 0; uint32_t cts, acked, ack_point, sack_changed = 0; uint32_t p_maxseg, maxseg, p_acked = 0; INP_WLOCK_ASSERT(tp->t_inpcb); if (th->th_flags & TH_RST) { /* We don't log resets */ return (0); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; cts = bbr->r_ctl.rc_rcvtime; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); changed = 0; maxseg = tp->t_maxseg - bbr->rc_last_options; p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg); th_ack = th->th_ack; if (SEQ_GT(th_ack, tp->snd_una)) { acked = th_ack - tp->snd_una; bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__); bbr->rc_tp->t_acktime = ticks; } else acked = 0; if (SEQ_LEQ(th_ack, tp->snd_una)) { /* Only sent here for sack processing */ goto proc_sack; } if (rsm && SEQ_GT(th_ack, rsm->r_start)) { changed = th_ack - rsm->r_start; } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) { /* * For the SYN incoming case we will not have called * tcp_output for the sending of the SYN, so there will be * no map. All other cases should probably be a panic. */ if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { /* * We have a timestamp that can be used to generate * an initial RTT. */ uint32_t ts, now, rtt; ts = bbr_ts_convert(to->to_tsecr); now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv)); rtt = now - ts; if (rtt < 1) rtt = 1; bbr_log_type_bbrrttprop(bbr, rtt, tp->iss, 0, cts, BBR_RTT_BY_TIMESTAMP, tp->iss, 0); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); changed = 1; bbr->r_wanted_output = 1; goto out; } goto proc_sack; } else if (rsm == NULL) { goto out; } if (changed) { /* * The ACK point is advancing to th_ack, we must drop off * the packets in the rack log and calculate any eligble * RTT's. */ bbr->r_wanted_output = 1; more: if (rsm == NULL) { if (tp->t_flags & TF_SENTFIN) { /* if we send a FIN we will not hav a map */ goto proc_sack; } #ifdef BBR_INVARIANTS panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n", tp, th, tp->t_state, bbr, tp->snd_una, tp->snd_max, changed); #endif goto proc_sack; } } if (SEQ_LT(th_ack, rsm->r_start)) { /* Huh map is missing this */ #ifdef BBR_INVARIANTS printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n", rsm->r_start, th_ack, tp->t_state, bbr->r_state, bbr); panic("th-ack is bad bbr:%p tp:%p", bbr, tp); #endif goto proc_sack; } else if (th_ack == rsm->r_start) { /* None here to ack */ goto proc_sack; } /* * Clear the dup ack counter, it will * either be freed or if there is some * remaining we need to start it at zero. */ rsm->r_dupack = 0; /* Now do we consume the whole thing? */ if (SEQ_GEQ(th_ack, rsm->r_end)) { /* Its all consumed. */ uint32_t left; if (rsm->r_flags & BBR_ACKED) { /* * It was acked on the scoreboard -- remove it from * total */ p_acked += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); if (bbr->r_ctl.rc_sacked == 0) bbr->r_ctl.rc_sacklast = NULL; } else { bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack); if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } if (rsm->r_flags & BBR_SACK_PASSED) { /* * There are acked segments ACKED on the * scoreboard further up. We are seeing * reordering. */ BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } } rsm->r_flags &= ~BBR_MARKED_LOST; } bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } if (bbr->r_ctl.rc_next == rsm) { /* scoot along the marker */ bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); } bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); /* Adjust the packet counts */ left = th_ack - rsm->r_end; /* Free back to zone */ bbr_free(bbr, rsm); if (left) { rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); goto more; } goto proc_sack; } if (rsm->r_flags & BBR_ACKED) { /* * It was acked on the scoreboard -- remove it from total * for the part being cum-acked. */ p_acked += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start); if (bbr->r_ctl.rc_sacked == 0) bbr->r_ctl.rc_sacklast = NULL; } else { /* * It was acked up to th_ack point for the first time */ struct bbr_sendmap lrsm; memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap)); lrsm.r_end = th_ack; bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack); } if ((rsm->r_flags & BBR_MARKED_LOST) && ((rsm->r_flags & BBR_ACKED) == 0)) { /* * It was marked lost and partly ack'd now * for the first time. We lower the rc_lost_bytes * and still leave it MARKED. */ bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start; } bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; /* adjust packet count */ rsm->r_start = th_ack; proc_sack: /* Check for reneging */ rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) { /* * The peer has moved snd_una up to the edge of this send, * i.e. one that it had previously acked. The only way that * can be true if the peer threw away data (space issues) * that it had previously sacked (else it would have given * us snd_una up to (rsm->r_end). We need to undo the acked * markings here. * * Note we have to look to make sure th_ack is our * rsm->r_start in case we get an old ack where th_ack is * behind snd_una. */ bbr_peer_reneges(bbr, rsm, th->th_ack); } if ((to->to_flags & TOF_SACK) == 0) { /* We are done nothing left to log */ goto out; } rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); if (rsm) { last_seq = rsm->r_end; } else { last_seq = tp->snd_max; } /* Sack block processing */ if (SEQ_GT(th_ack, tp->snd_una)) ack_point = th_ack; else ack_point = tp->snd_una; for (i = 0; i < to->to_nsacks; i++) { bcopy((to->to_sacks + i * TCPOLEN_SACK), &sack, sizeof(sack)); sack.start = ntohl(sack.start); sack.end = ntohl(sack.end); if (SEQ_GT(sack.end, sack.start) && SEQ_GT(sack.start, ack_point) && SEQ_LT(sack.start, tp->snd_max) && SEQ_GT(sack.end, ack_point) && SEQ_LEQ(sack.end, tp->snd_max)) { if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) && (SEQ_LT(sack.end, last_seq)) && ((sack.end - sack.start) < (p_maxseg / 8))) { /* * Not the last piece and its smaller than * 1/8th of a p_maxseg. We ignore this. */ BBR_STAT_INC(bbr_runt_sacks); continue; } sack_blocks[num_sack_blks] = sack; num_sack_blks++; #ifdef NETFLIX_STATS } else if (SEQ_LEQ(sack.start, th_ack) && SEQ_LEQ(sack.end, th_ack)) { /* * Its a D-SACK block. */ tcp_record_dsack(sack.start, sack.end); #endif } } if (num_sack_blks == 0) goto out; /* * Sort the SACK blocks so we can update the rack scoreboard with * just one pass. */ new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks, num_sack_blks, th->th_ack); ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks); BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks); BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb)); num_sack_blks = new_sb; if (num_sack_blks < 2) { goto do_sack_work; } /* Sort the sacks */ for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { sack = sack_blocks[i]; sack_blocks[i] = sack_blocks[j]; sack_blocks[j] = sack; } } } /* * Now are any of the sack block ends the same (yes some * implememtations send these)? */ again: if (num_sack_blks > 1) { for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (sack_blocks[i].end == sack_blocks[j].end) { /* * Ok these two have the same end we * want the smallest end and then * throw away the larger and start * again. */ if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { /* * The second block covers * more area use that */ sack_blocks[i].start = sack_blocks[j].start; } /* * Now collapse out the dup-sack and * lower the count */ for (k = (j + 1); k < num_sack_blks; k++) { sack_blocks[j].start = sack_blocks[k].start; sack_blocks[j].end = sack_blocks[k].end; j++; } num_sack_blks--; goto again; } } } } do_sack_work: rsm = bbr->r_ctl.rc_sacklast; for (i = 0; i < num_sack_blks; i++) { acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts); if (acked) { bbr->r_wanted_output = 1; changed += acked; sack_changed += acked; } } out: *prev_acked = p_acked; if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { /* * Ok we have a high probability that we need to go in to * recovery since we have data sack'd */ struct bbr_sendmap *rsm; rsm = bbr_check_recovery_mode(tp, bbr, cts); if (rsm) { /* Enter recovery */ entered_recovery = 1; bbr->r_wanted_output = 1; /* * When we enter recovery we need to assure we send * one packet. */ if (bbr->r_ctl.rc_resend == NULL) { bbr->r_ctl.rc_resend = rsm; } } } if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { /* * See if we need to rack-retransmit anything if so set it * up as the thing to resend assuming something else is not * already in that position. */ if (bbr->r_ctl.rc_resend == NULL) { bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); } } /* * We return the amount that changed via sack, this is used by the * ack-received code to augment what was changed between th_ack <-> * snd_una. */ return (sack_changed); } static void bbr_strike_dupack(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (rsm && (rsm->r_dupack < 0xff)) { rsm->r_dupack++; if (rsm->r_dupack >= DUP_ACK_THRESHOLD) bbr->r_wanted_output = 1; } } /* * Return value of 1, we do not need to call bbr_process_data(). * return value of 0, bbr_process_data can be called. * For ret_val if its 0 the TCB is locked and valid, if its non-zero * its unlocked and probably unsafe to touch the TCB. */ static int bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val) { int32_t ourfinisacked = 0; int32_t acked_amount; uint16_t nsegs; int32_t acked; uint32_t lost, sack_changed = 0; struct mbuf *mfree; struct tcp_bbr *bbr; uint32_t prev_acked = 0; bbr = (struct tcp_bbr *)tp->t_fb_ptr; lost = bbr->r_ctl.rc_lost; nsegs = max(1, m->m_pkthdr.lro_nsegs); if (SEQ_GT(th->th_ack, tp->snd_max)) { ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); bbr->r_wanted_output = 1; return (1); } if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { /* Process the ack */ if (bbr->rc_in_persist) tp->t_rxtshift = 0; if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) bbr_strike_dupack(bbr); sack_changed = bbr_log_ack(tp, to, th, &prev_acked); } bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost)); if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* * Old ack, behind the last one rcv'd or a duplicate ack * with SACK info. */ if (th->th_ack == tp->snd_una) { bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0); if (bbr->r_state == TCPS_SYN_SENT) { /* * Special case on where we sent SYN. When * the SYN-ACK is processed in syn_sent * state it bumps the snd_una. This causes * us to hit here even though we did ack 1 * byte. * * Go through the nothing left case so we * send data. */ goto nothing_left; } } return (0); } /* * If we reach this point, ACK is not a duplicate, i.e., it ACKs * something we sent. */ if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our SYN has * been ACK'd (so connection is now fully synchronized). Go * to non-starred state, increment snd_una for ACK of SYN, * and check if we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; /* Send window already scaled. */ } } INP_WLOCK_ASSERT(tp->t_inpcb); acked = BYTES_THIS_ACK(tp, th); KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); /* * If we just performed our first retransmit, and the ACK arrives * within our recovery window, then it was a mistake to do the * retransmit in the first place. Recover our original cwnd and * ssthresh, and proceed to transmit where we left off. */ if (tp->t_flags & TF_PREVVALID) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); } SOCKBUF_LOCK(&so->so_snd); acked_amount = min(acked, (int)sbavail(&so->so_snd)); tp->snd_wnd -= acked_amount; mfree = sbcut_locked(&so->so_snd, acked_amount); /* NB: sowwakeup_locked() does an implicit unlock. */ sowwakeup_locked(so); m_freem(mfree); if (SEQ_GT(th->th_ack, tp->snd_una)) { bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); } tp->snd_una = th->th_ack; bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost)); if (IN_RECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover) && (SEQ_LT(th->th_ack, tp->snd_max))) { tcp_bbr_partialack(tp); } else { bbr_post_recovery(tp); } } if (SEQ_GT(tp->snd_una, tp->snd_recover)) { tp->snd_recover = tp->snd_una; } if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { tp->snd_nxt = tp->snd_max; } if (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ nothing_left: bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) bbr->rc_tp->t_acktime = 0; if ((sbused(&so->so_snd) == 0) && (tp->t_flags & TF_SENTFIN)) { ourfinisacked = 1; } bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); if (bbr->rc_in_persist == 0) { bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; } sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); /* * We invalidate the last ack here since we * don't want to transfer forward the time * for our sum's calculations. */ if ((tp->t_state >= TCPS_FIN_WAIT_1) && (sbavail(&so->so_snd) == 0) && (tp->t_flags2 & TF2_DROP_AF_DATA)) { /* * The socket was gone and the peer sent data, time * to reset him. */ *ret_val = 1; tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); BBR_STAT_INC(bbr_dropped_af_data); return (1); } /* Set need output so persist might get set */ bbr->r_wanted_output = 1; } if (ofia) *ofia = ourfinisacked; return (0); } static void bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) { if (bbr->rc_in_persist == 0) { bbr_timer_cancel(bbr, __LINE__, cts); bbr->r_ctl.rc_last_delay_val = 0; tp->t_rxtshift = 0; bbr->rc_in_persist = 1; bbr->r_ctl.rc_went_idle_time = cts; /* We should be capped when rw went to 0 but just in case */ bbr_log_type_pesist(bbr, cts, 0, line, 1); /* Time freezes for the state, so do the accounting now */ if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { int32_t idx; idx = bbr_state_val(bbr); counter_u64_add(bbr_state_time[(idx + 5)], time_in); } else { counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } } bbr->r_ctl.rc_bbr_state_time = cts; } } static void bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time) { /* * Note that if idle time does not exceed our * threshold, we do nothing continuing the state * transitions we were last walking through. */ if (idle_time >= bbr_idle_restart_threshold) { if (bbr->rc_use_idle_restart) { bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT; /* * Set our target using BBR_UNIT, so * we increase at a dramatic rate but * we stop when we get the pipe * full again for our current b/w estimate. */ bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_set_state_target(bbr, __LINE__); /* Now setup our gains to ramp up */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; bbr_log_type_statechange(bbr, cts, __LINE__); } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { bbr_substate_change(bbr, cts, __LINE__, 1); } } } static void bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) { uint32_t idle_time; if (bbr->rc_in_persist == 0) return; idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time); bbr->rc_in_persist = 0; bbr->rc_hit_state_1 = 0; bbr->r_ctl.rc_del_time = cts; /* * We invalidate the last ack here since we * don't want to transfer forward the time * for our sum's calculations. */ if (bbr->rc_inp->inp_in_hpts) { tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT); bbr->rc_timer_first = 0; bbr->r_ctl.rc_hpts_flags = 0; bbr->r_ctl.rc_last_delay_val = 0; bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; } bbr_log_type_pesist(bbr, cts, idle_time, line, 0); if (idle_time >= bbr_rtt_probe_time) { /* * This qualifies as a RTT_PROBE session since we drop the * data outstanding to nothing and waited more than * bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0); bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts; } tp->t_rxtshift = 0; /* * If in probeBW and we have persisted more than an RTT lets do * special handling. */ /* Force a time based epoch */ bbr_set_epoch(bbr, cts, __LINE__); /* * Setup the lost so we don't count anything against the guy * we have been stuck with during persists. */ bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; /* Time un-freezes for the state */ bbr->r_ctl.rc_bbr_state_time = cts; if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) || (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) { /* * If we are going back to probe-bw * or probe_rtt, we may need to possibly * do a fast restart. */ bbr_restart_after_idle(bbr, cts, idle_time); } } static void bbr_collapsed_window(struct tcp_bbr *bbr) { /* * Now we must walk the * send map and divide the * ones left stranded. These * guys can't cause us to abort * the connection and are really * "unsent". However if a buggy * client actually did keep some * of the data i.e. collapsed the win * and refused to ack and then opened * the win and acked that data. We would * get into an ack war, the simplier * method then of just pretending we * did not send those segments something * won't work. */ struct bbr_sendmap *rsm, *nrsm; tcp_seq max_seq; uint32_t maxseg; int can_split = 0; int fnd = 0; maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd; bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { /* Find the first seq past or at maxseq */ if (rsm->r_flags & BBR_RWND_COLLAPSED) rsm->r_flags &= ~BBR_RWND_COLLAPSED; if (SEQ_GEQ(max_seq, rsm->r_start) && SEQ_GEQ(rsm->r_end, max_seq)) { fnd = 1; break; } } bbr->rc_has_collapsed = 0; if (!fnd) { /* Nothing to do strange */ return; } /* * Now can we split? * * We don't want to split if splitting * would generate too many small segments * less we let an attacker fragment our * send_map and leave us out of memory. */ if ((max_seq != rsm->r_start) && (max_seq != rsm->r_end)){ /* can we split? */ int res1, res2; res1 = max_seq - rsm->r_start; res2 = rsm->r_end - max_seq; if ((res1 >= (maxseg/8)) && (res2 >= (maxseg/8))) { /* No small pieces here */ can_split = 1; } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) { /* We are under the limit */ can_split = 1; } } /* Ok do we need to split this rsm? */ if (max_seq == rsm->r_start) { /* It's this guy no split required */ nrsm = rsm; } else if (max_seq == rsm->r_end) { /* It's the next one no split required. */ nrsm = TAILQ_NEXT(rsm, r_next); if (nrsm == NULL) { /* Huh? */ return; } } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) { /* yep we need to split it */ nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* failed XXXrrs what can we do mark the whole? */ nrsm = rsm; goto no_split; } /* Clone it */ bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0); bbr_clone_rsm(bbr, nrsm, rsm, max_seq); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } } else { /* * Split not allowed just start here just * use this guy. */ nrsm = rsm; } no_split: BBR_STAT_INC(bbr_collapsed_win); /* reuse fnd as a count */ fnd = 0; TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) { nrsm->r_flags |= BBR_RWND_COLLAPSED; fnd++; bbr->rc_has_collapsed = 1; } bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd); } static void bbr_un_collapse_window(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; int cleared = 0; TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if (rsm->r_flags & BBR_RWND_COLLAPSED) { /* Clear the flag */ rsm->r_flags &= ~BBR_RWND_COLLAPSED; cleared++; } else break; } bbr_log_type_rwnd_collapse(bbr, (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared); bbr->rc_has_collapsed = 0; } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { /* * Update window information. Don't look at window if no ACK: TAC's * send garbage on first SYN. */ uint16_t nsegs; int32_t tfo_syn; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); nsegs = max(1, m->m_pkthdr.lro_nsegs); if ((thflags & TH_ACK) && (SEQ_LT(tp->snd_wl1, th->th_seq) || (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if (tlen == 0 && tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) KMOD_TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; bbr->r_wanted_output = 1; } else if (thflags & TH_ACK) { if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; } } if (tp->snd_wnd < ctf_outstanding(tp)) /* The peer collapsed its window on us */ bbr_collapsed_window(bbr); else if (bbr->rc_has_collapsed) bbr_un_collapse_window(bbr); /* Was persist timer active and now we have window space? */ if ((bbr->rc_in_persist != 0) && (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr)))) { /* * Make the rate persist at end of persist mode if idle long * enough */ bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); /* Make sure we output to start the timer */ bbr->r_wanted_output = 1; } /* Do we need to enter persist? */ if ((bbr->rc_in_persist == 0) && (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* No send window.. we must enter persist */ bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); } if (tp->t_flags2 & TF2_DROP_AF_DATA) { m_freem(m); return (0); } /* * We don't support urgent data but * drag along the up just to make sure * if there is a stack switch no one * is surprised. */ tp->rcv_up = tp->rcv_nxt; INP_WLOCK_ASSERT(tp->t_inpcb); /* * Process the segment text, merging it into the TCP sequencing * queue, and arranging for acknowledgment of receipt if necessary. * This process logically involves adjusting tp->rcv_wnd as data is * presented to the user (this happens in tcp_usrreq.c, case * PRU_RCVD). If a FIN has already been received on this connection * then we just ignore the text. */ tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && IS_FASTOPEN(tp->t_flags)); if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && TCPS_HAVERCVDFIN(tp->t_state) == 0) { tcp_seq save_start = th->th_seq; tcp_seq save_rnxt = tp->rcv_nxt; int save_tlen = tlen; m_adj(m, drop_hdrlen); /* delayed header drop */ /* * Insert segment which includes th into TCP reassembly * queue with control block tp. Set thflags to whether * reassembly now includes a segment with FIN. This handles * the common case inline (segment is the next to be * received on an established connection, and the queue is * empty), avoiding linkage into and removal from the queue * and repetition of various conversions. Set DELACK for * segments received in order, but ack immediately when * segments are out of order (so fast retransmit can work). */ if (th->th_seq == tp->rcv_nxt && SEGQ_EMPTY(tp) && (TCPS_HAVEESTABLISHED(tp->t_state) || tfo_syn)) { #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (0); } } #endif if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) { bbr->bbr_segs_rcvd += max(1, nsegs); tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { bbr->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } thflags = th->th_flags & TH_FIN; KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) m_freem(m); else #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && appended != mcnt) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif } else { /* * XXX: Due to the header drop above "th" is * theoretically invalid by now. Fortunately * m_adj() doesn't actually frees any mbufs when * trimming from the head. */ tcp_seq temp = save_start; thflags = tcp_reass(tp, th, &temp, &tlen, m); tp->t_flags |= TF_ACKNOW; if (tp->t_flags & TF_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } } if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { /* * DSACK actually handled in the fastpath * above. */ tcp_update_sack_list(tp, save_start, save_start + save_tlen); } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { if ((tp->rcv_numsacks >= 1) && (tp->sackblks[0].end == save_start)) { /* * Partial overlap, recorded at todrop * above. */ tcp_update_sack_list(tp, tp->sackblks[0].start, tp->sackblks[0].end); } else { tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } } else if (tlen >= save_tlen) { /* Update of sackblks. */ tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } else if (tlen > 0) { tcp_update_dsack_list(tp, save_start, save_start + tlen); } } } else { m_freem(m); thflags &= ~TH_FIN; } /* * If FIN is received ACK the FIN and let the user know that the * connection is closing. */ if (thflags & TH_FIN) { if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { /* The socket upcall is handled by socantrcvmore. */ socantrcvmore(so); /* * If connection is half-synchronized (ie NEEDSYN * flag on) then delay ACK, so it may be piggybacked * when SYN is sent. Otherwise, since we received a * FIN then no more input can be expected, send ACK * now. */ if (tp->t_flags & TF_NEEDSYN) { tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt++; } switch (tp->t_state) { /* * In SYN_RECEIVED and ESTABLISHED STATES enter the * CLOSE_WAIT state. */ case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: tcp_state_change(tp, TCPS_CLOSE_WAIT); break; /* * If still in FIN_WAIT_1 STATE FIN has not been * acked so enter the CLOSING state. */ case TCPS_FIN_WAIT_1: tcp_state_change(tp, TCPS_CLOSING); break; /* * In FIN_WAIT_2 state enter the TIME_WAIT state, * starting the time-wait timer, turning off the * other standard timers. */ case TCPS_FIN_WAIT_2: bbr->rc_timer_first = 1; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); INP_WLOCK_ASSERT(tp->t_inpcb); tcp_twstart(tp); return (1); } } /* * Return any desired output. */ if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > ctf_outstanding(tp))) { bbr->r_wanted_output = 1; } INP_WLOCK_ASSERT(tp->t_inpcb); return (0); } /* * Here nothing is really faster, its just that we * have broken out the fast-data path also just like * the fast-ack. Return 1 if we processed the packet * return 0 if you need to take the "slow-path". */ static int bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt) { uint16_t nsegs; int32_t newsize = 0; /* automatic sockbuf scaling */ struct tcp_bbr *bbr; #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; #endif #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif /* On the hpts and we would have called output */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if (bbr->r_ctl.rc_resend != NULL) { return (0); } if (tiwin && tiwin != tp->snd_wnd) { return (0); } if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { return (0); } if (__predict_false((to->to_flags & TOF_TS) && (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { return (0); } if (__predict_false((th->th_ack != tp->snd_una))) { return (0); } if (__predict_false(tlen > sbspace(&so->so_rcv))) { return (0); } if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * This is a pure, in-sequence data packet with nothing on the * reassembly queue and we have enough buffer space to take it. */ nsegs = max(1, m->m_pkthdr.lro_nsegs); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (1); } } #endif /* Clean receiver SACK report if present */ if (tp->rcv_numsacks) tcp_clean_sackreport(tp); KMOD_TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } /* * Pull snd_wl1 up to prevent seq wrap relative to th_seq. */ tp->snd_wl1 = th->th_seq; /* * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. */ tp->rcv_up = tp->rcv_nxt; KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else { /* * Set new socket buffer size. Give up when limit is * reached. */ if (newsize) if (!sbreserve_locked(&so->so_rcv, newsize, so, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); ctf_calc_rwin(so, tp); } /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && mcnt != appended) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif if (DELAY_ACK(tp, bbr, nsegs)) { bbr->bbr_segs_rcvd += max(1, nsegs); tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { bbr->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; } return (1); } /* * This subfunction is used to try to highly optimize the * fast path. We again allow window updates that are * in sequence to remain in the fast-path. We also add * in the __predict's to attempt to help the compiler. * Note that if we return a 0, then we can *not* process * it and the caller should push the packet into the * slow-path. If we return 1, then all is well and * the packet is fully processed. */ static int bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) { int32_t acked; uint16_t nsegs; uint32_t sack_changed; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif uint32_t prev_acked = 0; struct tcp_bbr *bbr; if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* Old ack, behind (or duplicate to) the last one rcv'd */ return (0); } if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { /* Above what we have sent? */ return (0); } if (__predict_false(tiwin == 0)) { /* zero window */ return (0); } if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { /* We need a SYN or a FIN, unlikely.. */ return (0); } if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { /* Timestamp is behind .. old ack with seq wrap? */ return (0); } if (__predict_false(IN_RECOVERY(tp->t_flags))) { /* Still recovering */ return (0); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (__predict_false(bbr->r_ctl.rc_resend != NULL)) { /* We are retransmitting */ return (0); } if (__predict_false(bbr->rc_in_persist != 0)) { /* In persist mode */ return (0); } if (bbr->r_ctl.rc_sacked) { /* We have sack holes on our scoreboard */ return (0); } /* Ok if we reach here, we can process a fast-ack */ nsegs = max(1, m->m_pkthdr.lro_nsegs); sack_changed = bbr_log_ack(tp, to, th, &prev_acked); /* * We never detect loss in fast ack [we can't * have a sack and can't be in recovery so * we always pass 0 (nothing detected)]. */ bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0); /* Did the window get updated? */ if (tiwin != tp->snd_wnd) { tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; } /* Do we need to exit persists? */ if ((bbr->rc_in_persist != 0) && (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr)))) { bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); bbr->r_wanted_output = 1; } /* Do we need to enter persists? */ if ((bbr->rc_in_persist == 0) && (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* No send window.. we must enter persist */ bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); } /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = bbr->r_ctl.rc_rcvtime; tp->ts_recent = to->to_tsval; } /* * This is a pure ack for outstanding data. */ KMOD_TCPSTAT_INC(tcps_predack); /* * "bad retransmit" recovery. */ if (tp->t_flags & TF_PREVVALID) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); } /* * Recalculate the transmit timer / rtt. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ acked = BYTES_THIS_ACK(tp, th); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, to); #endif KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); sbdrop(&so->so_snd, acked); if (SEQ_GT(th->th_ack, tp->snd_una)) bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); tp->snd_una = th->th_ack; if (tp->snd_wnd < ctf_outstanding(tp)) /* The peer collapsed its window on us */ bbr_collapsed_window(bbr); else if (bbr->rc_has_collapsed) bbr_un_collapse_window(bbr); if (SEQ_GT(tp->snd_una, tp->snd_recover)) { tp->snd_recover = tp->snd_una; } bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0); /* * Pull snd_wl2 up to prevent seq wrap relative to th_ack. */ tp->snd_wl2 = th->th_ack; m_freem(m); /* * If all outstanding data are acked, stop retransmit timer, * otherwise restart timer using current (possibly backed-off) * value. If process is waiting for space, wakeup/selwakeup/signal. * If data are ready to send, let tcp_output decide between more * output or persist. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif /* Wake up the socket if we have room to write more */ sowwakeup(so); if (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) bbr->rc_tp->t_acktime = 0; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); if (bbr->rc_in_persist == 0) { bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; } sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); /* * We invalidate the last ack here since we * don't want to transfer forward the time * for our sum's calculations. */ bbr->r_wanted_output = 1; } if (sbavail(&so->so_snd)) { bbr->r_wanted_output = 1; } return (1); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t todrop; int32_t ourfinisacked = 0; struct tcp_bbr *bbr; int32_t ret_val = 0; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); /* * If the state is SYN_SENT: if seg contains an ACK, but not for our * SYN, drop the input. if seg contains a RST, then drop the * connection. if seg does not contain SYN, then drop it. Otherwise * this is an acceptable SYN segment initialize tp->rcv_nxt and * tp->irs if seg contains ack then advance tp->snd_una. BRR does * not support ECN so we will not say we are capable. if SYN has * been acked change to ESTABLISHED else SYN_RCVD state arrange for * segment to be acked (eventually) continue processing rest of * data/controls, beginning with URG */ if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { TCP_PROBE5(connect__refused, NULL, tp, mtod(m, const char *), tp, th); tp = tcp_drop(tp, ECONNREFUSED); ctf_do_drop(m, tp); return (1); } if (thflags & TH_RST) { ctf_do_drop(m, tp); return (1); } if (!(thflags & TH_SYN)) { ctf_do_drop(m, tp); return (1); } tp->irs = th->th_seq; tcp_rcvseqinit(tp); if (thflags & TH_ACK) { int tfo_partial = 0; KMOD_TCPSTAT_INC(tcps_connects); soisconnected(so); #ifdef MAC mac_socketpeer_set_from_mbuf(m, so); #endif /* Do window scaling on this connection? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } tp->rcv_adv += min(tp->rcv_wnd, TCP_MAXWIN << tp->rcv_scale); /* * If not all the data that was sent in the TFO SYN * has been acked, resend the remainder right away. */ if (IS_FASTOPEN(tp->t_flags) && (tp->snd_una != tp->snd_max)) { tp->snd_nxt = th->th_ack; tfo_partial = 1; } /* * If there's data, delay ACK; if there's also a FIN ACKNOW * will be turned on later. */ if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) { bbr->bbr_segs_rcvd += 1; tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { bbr->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; } if (SEQ_GT(th->th_ack, tp->iss)) { /* * The SYN is acked * handle it specially. */ bbr_log_syn(tp, to); } if (SEQ_GT(th->th_ack, tp->snd_una)) { /* * We advance snd_una for the * fast open case. If th_ack is * acknowledging data beyond * snd_una we can't just call * ack-processing since the * data stream in our send-map * will start at snd_una + 1 (one * beyond the SYN). If its just * equal we don't need to do that * and there is no send_map. */ tp->snd_una++; } /* * Received in SYN_SENT[*] state. Transitions: * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; thflags &= ~TH_SYN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(connect__established, NULL, tp, mtod(m, const char *), tp, th); cc_conn_init(tp); } } else { /* * Received initial SYN in SYN-SENT[*] state => simultaneous * open. If segment contains CC option and there is a * cached CC, apply TAO test. If it succeeds, connection is * * half-synchronized. Otherwise, do 3-way handshake: * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If * there was no CC option, clear cached CC value. */ tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); tcp_state_change(tp, TCPS_SYN_RECEIVED); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * Advance th->th_seq to correspond to first data byte. If data, * trim to stay within window, dropping FIN if necessary. */ th->th_seq++; if (tlen > tp->rcv_wnd) { todrop = tlen - tp->rcv_wnd; m_adj(m, -todrop); tlen = tp->rcv_wnd; thflags &= ~TH_FIN; KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); } tp->snd_wl1 = th->th_seq - 1; tp->rcv_up = th->th_seq; /* * Client side of transaction: already sent SYN and data. If the * remote host used T/TCP to validate the SYN, our data will be * ACK'd; if so, enter normal data segment processing in the middle * of step 5, ack processing. Otherwise, goto step 6. */ if (thflags & TH_ACK) { if ((to->to_flags & TOF_TS) != 0) { uint32_t t, rtt; t = tcp_tv_to_mssectick(&bbr->rc_tv); if (TSTMP_GEQ(t, to->to_tsecr)) { rtt = t - to->to_tsecr; if (rtt == 0) { rtt = 1; } rtt *= MS_IN_USEC; tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); } } if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) return (ret_val); /* We may have changed to FIN_WAIT_1 above */ if (tp->t_state == TCPS_FIN_WAIT_1) { /* * In FIN_WAIT_1 STATE in addition to the processing * for the ESTABLISHED state if our FIN is now * acknowledged then enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then * closing user can proceed. Starting the * timer is contrary to the specification, * but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and * use a compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if (IS_FASTOPEN(tp->t_flags)) { /* * When a TFO connection is in SYN_RECEIVED, the only valid * packets are the initial SYN, a retransmit/copy of the * initial SYN (possibly with a subset of the original * data), a valid ACK, a FIN, or a RST. */ if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } else if (thflags & TH_SYN) { /* non-initial SYN is ignored */ if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { ctf_do_drop(m, NULL); return (0); } } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { ctf_do_drop(m, NULL); return (0); } } if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } /* * In the SYN-RECEIVED state, validate that the packet belongs to * this connection before trimming the data to fit the receive * window. Check the sequence number versus IRS since we know the * sequence numbers haven't wrapped. This is a partial fix for the * "LAND" DoS attack. */ if (SEQ_LT(th->th_seq, tp->irs)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } tp->snd_wnd = tiwin; /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (IS_FASTOPEN(tp->t_flags)) { cc_conn_init(tp); } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } KMOD_TCPSTAT_INC(tcps_connects); soisconnected(so); /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } /* * ok for the first time in lets see if we can use the ts to figure * out what the initial RTT was. */ if ((to->to_flags & TOF_TS) != 0) { uint32_t t, rtt; t = tcp_tv_to_mssectick(&bbr->rc_tv); if (TSTMP_GEQ(t, to->to_tsecr)) { rtt = t - to->to_tsecr; if (rtt == 0) { rtt = 1; } rtt *= MS_IN_USEC; tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); } } /* Drop off any SYN in the send map (probably not there) */ if (thflags & TH_ACK) bbr_log_syn(tp, to); if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; } /* * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> * FIN-WAIT-1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(accept__established, NULL, tp, mtod(m, const char *), tp, th); /* * TFO connections call cc_conn_init() during SYN * processing. Calling it again here for such connections * is not harmless as it would undo the snd_cwnd reduction * that occurs when a TFO SYN|ACK is retransmitted. */ if (!IS_FASTOPEN(tp->t_flags)) cc_conn_init(tp); } /* * Account for the ACK of our SYN prior to * regular ACK processing below, except for * simultaneous SYN, which is handled later. */ if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) tp->snd_una++; /* * If segment contains data or ACK, will call tcp_reass() later; if * not, do so now to pass queued data to user. */ if (tlen == 0 && (thflags & TH_FIN) == 0) { (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, (struct mbuf *)0); if (tp->t_flags & TF_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } } tp->snd_wl1 = th->th_seq - 1; if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (tp->t_state == TCPS_FIN_WAIT_1) { /* We could have went to FIN_WAIT_1 (or EST) above */ /* * In FIN_WAIT_1 STATE in addition to the processing for the * ESTABLISHED state if our FIN is now acknowledged then * enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then closing * user can proceed. Starting the timer is contrary * to the specification, but if we don't get a FIN * we'll hang forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { struct tcp_bbr *bbr; int32_t ret_val; /* * Header prediction: check for the two common cases of a * uni-directional data xfer. If the packet has no control flags, * is in-sequence, the window didn't change and we're not * retransmitting, it's a candidate. If the length is zero and the * ack moved forward, we're the sender side of the xfer. Just free * the data acked & wake any higher level process that was blocked * waiting for space. If the length is non-zero and the ack didn't * move, we're the receiver side. If we're getting packets in-order * (the reassembly queue is empty), add the data toc The socket * buffer and note that we need a delayed ack. Make sure that the * hidden state-flags are also off. Since we check for * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) { /* * If we have delived under 4 segments increase the initial * window if raised by the peer. We use this to determine * dynamic and static rwnd's at the end of a connection. */ bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd); } if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && __predict_true(SEGQ_EMPTY(tp)) && __predict_true(th->th_seq == tp->rcv_nxt)) { if (tlen == 0) { if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt, iptos)) { return (0); } } else { if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt)) { return (0); } } } ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } /* State changes only happen in bbr_process_data() */ return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { struct tcp_bbr *bbr; int32_t ret_val; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static int bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr, struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so) { if (bbr->rc_allow_data_af_clo == 0) { close_now: tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); KMOD_TCPSTAT_INC(tcps_rcvafterclose); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); return (1); } if (sbavail(&so->so_snd) == 0) goto close_now; /* Ok we allow data that is ignored and a followup reset */ tp->rcv_nxt = th->th_seq + *tlen; tp->t_flags2 |= TF2_DROP_AF_DATA; bbr->r_wanted_output = 1; *tlen = 0; return (0); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { /* * If we can't receive any more data, then closing user can * proceed. Starting the timer is contrary to the * specification, but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tcp_twstart(tp); m_freem(m); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * case TCPS_LAST_ACK: Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tp = tcp_close(tp); ctf_do_drop(m, tp); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); /* Reset receive buffer auto scaling when not in bulk receive mode. */ if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then we may RST the other end depending on the outcome * of bbr_check_data_after_close. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ INP_WLOCK_ASSERT(tp->t_inpcb); if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ INP_WLOCK_ASSERT(tp->t_inpcb); if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } INP_WLOCK_ASSERT(tp->t_inpcb); return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static void bbr_stop_all_timers(struct tcpcb *tp) { struct tcp_bbr *bbr; /* * Assure no timers are running. */ if (tcp_timer_active(tp, TT_PERSIST)) { /* We enter in persists, set the flag appropriately */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; bbr->rc_in_persist = 1; } tcp_timer_suspend(tp, TT_PERSIST); tcp_timer_suspend(tp, TT_REXMT); tcp_timer_suspend(tp, TT_KEEP); tcp_timer_suspend(tp, TT_DELACK); } static void bbr_google_mode_on(struct tcp_bbr *bbr) { bbr->rc_use_google = 1; bbr->rc_no_pacing = 0; bbr->r_ctl.bbr_google_discount = bbr_google_discount; bbr->r_use_policer = bbr_policer_detection_enabled; bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); bbr->bbr_use_rack_cheat = 0; bbr->r_ctl.rc_incr_tmrs = 0; bbr->r_ctl.rc_inc_tcp_oh = 0; bbr->r_ctl.rc_inc_ip_oh = 0; bbr->r_ctl.rc_inc_enet_oh = 0; reset_time(&bbr->r_ctl.rc_delrate, BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); reset_time_small(&bbr->r_ctl.rc_rttprop, (11 * USECS_IN_SECOND)); tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); } static void bbr_google_mode_off(struct tcp_bbr *bbr) { bbr->rc_use_google = 0; bbr->r_ctl.bbr_google_discount = 0; bbr->no_pacing_until = bbr_no_pacing_until; bbr->r_use_policer = 0; if (bbr->no_pacing_until) bbr->rc_no_pacing = 1; else bbr->rc_no_pacing = 0; if (bbr_use_rack_resend_cheat) bbr->bbr_use_rack_cheat = 1; else bbr->bbr_use_rack_cheat = 0; if (bbr_incr_timers) bbr->r_ctl.rc_incr_tmrs = 1; else bbr->r_ctl.rc_incr_tmrs = 0; if (bbr_include_tcp_oh) bbr->r_ctl.rc_inc_tcp_oh = 1; else bbr->r_ctl.rc_inc_tcp_oh = 0; if (bbr_include_ip_oh) bbr->r_ctl.rc_inc_ip_oh = 1; else bbr->r_ctl.rc_inc_ip_oh = 0; if (bbr_include_enet_oh) bbr->r_ctl.rc_inc_enet_oh = 1; else bbr->r_ctl.rc_inc_enet_oh = 0; bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; reset_time(&bbr->r_ctl.rc_delrate, bbr_num_pktepo_for_del_limit); reset_time_small(&bbr->r_ctl.rc_rttprop, (bbr_filter_len_sec * USECS_IN_SECOND)); tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); } /* * Return 0 on success, non-zero on failure * which indicates the error (usually no memory). */ static int bbr_init(struct tcpcb *tp) { struct tcp_bbr *bbr = NULL; struct inpcb *inp; uint32_t cts; tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO)); if (tp->t_fb_ptr == NULL) { /* * We need to allocate memory but cant. The INP and INP_INFO * locks and they are recusive (happens during setup. So a * scheme to drop the locks fails :( * */ return (ENOMEM); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; bbr->rtt_valid = 0; inp = tp->t_inpcb; inp->inp_flags2 |= INP_CANNOT_DO_ECN; inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; TAILQ_INIT(&bbr->r_ctl.rc_map); TAILQ_INIT(&bbr->r_ctl.rc_free); TAILQ_INIT(&bbr->r_ctl.rc_tmap); bbr->rc_tp = tp; if (tp->t_inpcb) { bbr->rc_inp = tp->t_inpcb; } cts = tcp_get_usecs(&bbr->rc_tv); tp->t_acktime = 0; bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close; bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade; bbr->rc_tlp_threshold = bbr_tlp_thresh; bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh; bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay; bbr->r_ctl.rc_min_to = bbr_min_to; bbr->rc_bbr_state = BBR_STATE_STARTUP; bbr->r_ctl.bbr_lost_at_state = 0; bbr->r_ctl.rc_lost_at_startup = 0; bbr->rc_all_timers_stopped = 0; bbr->r_ctl.rc_bbr_lastbtlbw = 0; bbr->r_ctl.rc_pkt_epoch_del = 0; bbr->r_ctl.rc_pkt_epoch = 0; bbr->r_ctl.rc_lowest_rtt = 0xffffffff; bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain; bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; bbr->r_ctl.rc_went_idle_time = cts; bbr->rc_pacer_started = cts; bbr->r_ctl.rc_pkt_epoch_time = cts; bbr->r_ctl.rc_rcvtime = cts; bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.rc_del_time = cts; bbr->r_ctl.rc_tlp_rxt_last_time = cts; bbr->r_ctl.last_in_probertt = cts; bbr->skip_gain = 0; bbr->gain_is_limited = 0; bbr->no_pacing_until = bbr_no_pacing_until; if (bbr->no_pacing_until) bbr->rc_no_pacing = 1; if (bbr_use_google_algo) { bbr->rc_no_pacing = 0; bbr->rc_use_google = 1; bbr->r_ctl.bbr_google_discount = bbr_google_discount; bbr->r_use_policer = bbr_policer_detection_enabled; } else { bbr->rc_use_google = 0; bbr->r_ctl.bbr_google_discount = 0; bbr->r_use_policer = 0; } if (bbr_ts_limiting) bbr->rc_use_ts_limit = 1; else bbr->rc_use_ts_limit = 0; if (bbr_ts_can_raise) bbr->ts_can_raise = 1; else bbr->ts_can_raise = 0; if (V_tcp_delack_enabled == 1) tp->t_delayed_ack = 2; else if (V_tcp_delack_enabled == 0) tp->t_delayed_ack = 0; else if (V_tcp_delack_enabled < 100) tp->t_delayed_ack = V_tcp_delack_enabled; else tp->t_delayed_ack = 2; if (bbr->rc_use_google == 0) bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; else bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms; bbr->rc_max_rto_sec = bbr_rto_max_sec; bbr->rc_init_win = bbr_def_init_win; if (tp->t_flags & TF_REQ_TSTMP) bbr->rc_last_options = TCP_TS_OVERHEAD; bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options; bbr->r_ctl.rc_high_rwnd = tp->snd_wnd; bbr->r_init_rtt = 1; counter_u64_add(bbr_flows_nohdwr_pacing, 1); if (bbr_allow_hdwr_pacing) bbr->bbr_hdw_pace_ena = 1; else bbr->bbr_hdw_pace_ena = 0; if (bbr_sends_full_iwnd) bbr->bbr_init_win_cheat = 1; else bbr->bbr_init_win_cheat = 0; bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max; bbr->r_ctl.rc_drain_pg = bbr_drain_gain; bbr->r_ctl.rc_startup_pg = bbr_high_gain; bbr->rc_loss_exit = bbr_exit_startup_at_loss; bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain; bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second; bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar; bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max; bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor; bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min; bbr->r_ctl.bbr_cross_over = bbr_cross_over; bbr->r_ctl.rc_rtt_shrinks = cts; if (bbr->rc_use_google) { setup_time_filter(&bbr->r_ctl.rc_delrate, FILTER_TYPE_MAX, BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); setup_time_filter_small(&bbr->r_ctl.rc_rttprop, FILTER_TYPE_MIN, (11 * USECS_IN_SECOND)); } else { setup_time_filter(&bbr->r_ctl.rc_delrate, FILTER_TYPE_MAX, bbr_num_pktepo_for_del_limit); setup_time_filter_small(&bbr->r_ctl.rc_rttprop, FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND)); } bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0); if (bbr_uses_idle_restart) bbr->rc_use_idle_restart = 1; else bbr->rc_use_idle_restart = 0; bbr->r_ctl.rc_bbr_cur_del_rate = 0; bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps; if (bbr_resends_use_tso) bbr->rc_resends_use_tso = 1; #ifdef NETFLIX_PEAKRATE tp->t_peakrate_thr = tp->t_maxpeakrate; #endif if (tp->snd_una != tp->snd_max) { /* Create a send map for the current outstanding data */ struct bbr_sendmap *rsm; rsm = bbr_alloc(bbr); if (rsm == NULL) { uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; return (ENOMEM); } rsm->r_rtt_not_allowed = 1; rsm->r_tim_lastsent[0] = cts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; rsm->r_start = tp->snd_una; rsm->r_end = tp->snd_max; rsm->r_dupack = 0; rsm->r_delivered = bbr->r_ctl.rc_delivered; rsm->r_ts_valid = 0; rsm->r_del_ack_ts = tp->ts_recent; rsm->r_del_time = cts; if (bbr->r_ctl.r_app_limited_until) rsm->r_app_limited = 1; else rsm->r_app_limited = 0; TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) rsm->r_bbr_state = bbr_state_val(bbr); else rsm->r_bbr_state = 8; } if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0)) bbr->bbr_use_rack_cheat = 1; if (bbr_incr_timers && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_incr_tmrs = 1; if (bbr_include_tcp_oh && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_inc_tcp_oh = 1; if (bbr_include_ip_oh && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_inc_ip_oh = 1; if (bbr_include_enet_oh && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_inc_enet_oh = 1; bbr_log_type_statechange(bbr, cts, __LINE__); if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_srtt)) { uint32_t rtt; rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); } /* announce the settings and state */ bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT); tcp_bbr_tso_size_check(bbr, cts); /* * Now call the generic function to start a timer. This will place * the TCB on the hptsi wheel if a timer is needed with appropriate * flags. */ bbr_stop_all_timers(tp); bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0); return (0); } /* * Return 0 if we can accept the connection. Return * non-zero if we can't handle the connection. A EAGAIN * means you need to wait until the connection is up. * a EADDRNOTAVAIL means we can never handle the connection * (no SACK). */ static int bbr_handoff_ok(struct tcpcb *tp) { if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { /* Sure no problem though it may not stick */ return (0); } if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) { /* * We really don't know you have to get to ESTAB or beyond * to tell. */ return (EAGAIN); } if (tp->t_flags & TF_SENTFIN) return (EINVAL); if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) { return (0); } /* * If we reach here we don't do SACK on this connection so we can * never do rack. */ return (EINVAL); } static void bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged) { if (tp->t_fb_ptr) { uint32_t calc; struct tcp_bbr *bbr; struct bbr_sendmap *rsm; bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (bbr->r_ctl.crte) tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); bbr_log_flowend(bbr); bbr->rc_tp = NULL; if (tp->t_inpcb) { /* Backout any flags2 we applied */ tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN; tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY; } if (bbr->bbr_hdrw_pacing) counter_u64_add(bbr_flows_whdwr_pacing, -1); else counter_u64_add(bbr_flows_nohdwr_pacing, -1); if (bbr->r_ctl.crte != NULL) { tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); bbr->r_ctl.crte = NULL; } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); while (rsm) { TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); uma_zfree(bbr_zone, rsm); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); while (rsm) { TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); uma_zfree(bbr_zone, rsm); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); } calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd; if (calc > (bbr->r_ctl.rc_init_rwnd / 10)) BBR_STAT_INC(bbr_dynamic_rwnd); else BBR_STAT_INC(bbr_static_rwnd); bbr->r_ctl.rc_free_cnt = 0; uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; } /* Make sure snd_nxt is correctly set */ tp->snd_nxt = tp->snd_max; } static void bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win) { switch (tp->t_state) { case TCPS_SYN_SENT: bbr->r_state = TCPS_SYN_SENT; bbr->r_substate = bbr_do_syn_sent; break; case TCPS_SYN_RECEIVED: bbr->r_state = TCPS_SYN_RECEIVED; bbr->r_substate = bbr_do_syn_recv; break; case TCPS_ESTABLISHED: bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd); bbr->r_state = TCPS_ESTABLISHED; bbr->r_substate = bbr_do_established; break; case TCPS_CLOSE_WAIT: bbr->r_state = TCPS_CLOSE_WAIT; bbr->r_substate = bbr_do_close_wait; break; case TCPS_FIN_WAIT_1: bbr->r_state = TCPS_FIN_WAIT_1; bbr->r_substate = bbr_do_fin_wait_1; break; case TCPS_CLOSING: bbr->r_state = TCPS_CLOSING; bbr->r_substate = bbr_do_closing; break; case TCPS_LAST_ACK: bbr->r_state = TCPS_LAST_ACK; bbr->r_substate = bbr_do_lastack; break; case TCPS_FIN_WAIT_2: bbr->r_state = TCPS_FIN_WAIT_2; bbr->r_substate = bbr_do_fin_wait_2; break; case TCPS_LISTEN: case TCPS_CLOSED: case TCPS_TIME_WAIT: default: break; }; } static void bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog) { /* * Now what state are we going into now? Is there adjustments * needed? */ int32_t old_state, old_gain; old_state = bbr_state_val(bbr); old_gain = bbr->r_ctl.rc_bbr_hptsi_gain; if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) { /* Save the lowest srtt we saw in our end of the sub-state */ bbr->rc_hit_state_1 = 0; if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff) bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state; } bbr->rc_bbr_substate++; if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) { /* Cycle back to first state-> gain */ bbr->rc_bbr_substate = 0; } if (bbr_state_val(bbr) == BBR_SUB_GAIN) { /* * We enter the gain(5/4) cycle (possibly less if * shallow buffer detection is enabled) */ if (bbr->skip_gain) { /* * Hardware pacing has set our rate to * the max and limited our b/w just * do level i.e. no gain. */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1]; } else if (bbr->gain_is_limited && bbr->bbr_hdrw_pacing && bbr->r_ctl.crte) { /* * We can't gain above the hardware pacing * rate which is less than our rate + the gain * calculate the gain needed to reach the hardware * pacing rate.. */ uint64_t bw, rate, gain_calc; bw = bbr_get_bw(bbr); rate = bbr->r_ctl.crte->rate; if ((rate > bw) && (((bw * (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) { gain_calc = (rate * BBR_UNIT) / bw; if (gain_calc < BBR_UNIT) gain_calc = BBR_UNIT; bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc; } else { bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; } } else bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) { bbr->r_ctl.rc_bbr_state_atflight = cts; } else bbr->r_ctl.rc_bbr_state_atflight = 0; } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { bbr->rc_hit_state_1 = 1; bbr->r_ctl.rc_exta_time_gd = 0; bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (bbr_state_drain_2_tar) { bbr->r_ctl.rc_bbr_state_atflight = 0; } else bbr->r_ctl.rc_bbr_state_atflight = cts; bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN]; } else { /* All other cycles hit here 2-7 */ if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) { if (bbr_sub_drain_slam_cwnd && (bbr->rc_use_google == 0) && (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP)) bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) - bbr_get_rtt(bbr, BBR_RTT_PROP)); else bbr->r_ctl.rc_exta_time_gd = 0; if (bbr->r_ctl.rc_exta_time_gd) { bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd; /* Now chop up the time for each state (div by 7) */ bbr->r_ctl.rc_level_state_extra /= 7; if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) { /* Add a randomization */ bbr_randomize_extra_state_time(bbr); } } } bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)]; } if (bbr->rc_use_google) { bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); } bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; if (dolog) bbr_log_type_statechange(bbr, cts, line); if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { counter_u64_add(bbr_state_time[(old_state + 5)], time_in); } else { counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } } bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; bbr_set_state_target(bbr, __LINE__); if (bbr_sub_drain_slam_cwnd && (bbr->rc_use_google == 0) && (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { /* Slam down the cwnd */ bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; if (bbr_sub_drain_app_limit) { /* Go app limited if we are on a long drain */ bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes))); } bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (bbr->rc_lt_use_bw) { /* In policed mode we clamp pacing_gain to BBR_UNIT */ bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; } /* Google changes TSO size every cycle */ if (bbr->rc_use_google) tcp_bbr_tso_size_check(bbr, cts); bbr->r_ctl.gain_epoch = cts; bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch; } static void bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) { if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) && (google_allow_early_out == 1) && (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) { /* We have reached out target flight size possibly early */ goto change_state; } if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) { return; } if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) { /* * Must be a rttProp movement forward before * we can change states. */ return; } if (bbr_state_val(bbr) == BBR_SUB_GAIN) { /* * The needed time has passed but for * the gain cycle extra rules apply: * 1) If we have seen loss, we exit * 2) If we have not reached the target * we stay in GAIN (gain-to-target). */ if (google_consider_lost && losses) goto change_state; if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) { return; } } change_state: /* For gain we must reach our target, all others last 1 rttProp */ bbr_substate_change(bbr, cts, __LINE__, 1); } static void bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) { uint32_t flight, bbr_cur_cycle_time; if (bbr->rc_use_google) { bbr_set_probebw_google_gains(bbr, cts, losses); return; } if (cts == 0) { /* * Never alow cts to be 0 we * do this so we can judge if * we have set a timestamp. */ cts = 1; } if (bbr_state_is_pkt_epoch) bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); else bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP); if (bbr->r_ctl.rc_bbr_state_atflight == 0) { if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { flight = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) { /* Keep it slam down */ if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (bbr_sub_drain_app_limit) { /* Go app limited if we are on a long drain */ bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight); } } if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) && (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) || (flight >= bbr->r_ctl.flightsize_at_drain))) { /* * Still here after the same time as * the gain. We need to drain harder * for the next srtt. Reduce by a set amount * the gain drop is capped at DRAIN states * value (88). */ bbr->r_ctl.flightsize_at_drain = flight; if (bbr_drain_drop_mul && bbr_drain_drop_div && (bbr_drain_drop_mul < bbr_drain_drop_div)) { /* Use your specific drop value (def 4/5 = 20%) */ bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul; bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div; } else { /* You get drop of 20% */ bbr->r_ctl.rc_bbr_hptsi_gain *= 4; bbr->r_ctl.rc_bbr_hptsi_gain /= 5; } if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) { /* Reduce our gain again to the bottom */ bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); } bbr_log_exit_gain(bbr, cts, 4); /* * Extend out so we wait another * epoch before dropping again. */ bbr->r_ctl.gain_epoch = cts; } if (flight <= bbr->r_ctl.rc_target_at_state) { if (bbr_sub_drain_slam_cwnd && (bbr->rc_use_google == 0) && (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); bbr_log_exit_gain(bbr, cts, 3); } } else { /* Its a gain */ if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) { bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); goto change_state; } if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) || ((ctf_outstanding(bbr->rc_tp) + bbr->rc_tp->t_maxseg - 1) >= bbr->rc_tp->snd_wnd)) { bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); bbr_log_exit_gain(bbr, cts, 2); } } /** * We fall through and return always one of two things has * occurred. * 1) We are still not at target * * 2) We reached the target and set rc_bbr_state_atflight * which means we no longer hit this block * next time we are called. */ return; } change_state: if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) return; if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) { /* Less than a full time-period has passed */ return; } if (bbr->r_ctl.rc_level_state_extra && (bbr_state_val(bbr) > BBR_SUB_DRAIN) && ((cts - bbr->r_ctl.rc_bbr_state_time) < (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { /* Less than a full time-period + extra has passed */ return; } if (bbr_gain_gets_extra_too && bbr->r_ctl.rc_level_state_extra && (bbr_state_val(bbr) == BBR_SUB_GAIN) && ((cts - bbr->r_ctl.rc_bbr_state_time) < (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { /* Less than a full time-period + extra has passed */ return; } bbr_substate_change(bbr, cts, __LINE__, 1); } static uint32_t bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain) { uint32_t mss, tar; if (bbr->rc_use_google) { /* Google just uses the cwnd target */ tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain); } else { mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); /* Get the base cwnd with gain rounded to a mss */ tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr), gain), mss); /* Make sure it is within our min */ if (tar < get_min_cwnd(bbr)) return (get_min_cwnd(bbr)); } return (tar); } static void bbr_set_state_target(struct tcp_bbr *bbr, int line) { uint32_t tar, meth; if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { /* Special case using old probe-rtt method */ tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); meth = 1; } else { /* Non-probe-rtt case and reduced probe-rtt */ if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) { /* For gain cycle we use the hptsi gain */ tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); meth = 2; } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) { /* * If configured, or for google all other states * get BBR_UNIT. */ tar = bbr_get_a_state_target(bbr, BBR_UNIT); meth = 3; } else { /* * Or we set a target based on the pacing gain * for non-google mode and default (non-configured). * Note we don't set a target goal below drain (192). */ if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN]) { tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]); meth = 4; } else { tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); meth = 5; } } } bbr_log_set_of_state_target(bbr, tar, line, meth); bbr->r_ctl.rc_target_at_state = tar; } static void bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { /* Change to probe_rtt */ uint32_t time_in; bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain + bbr->r_ctl.rc_delivered); /* Setup so we force feed the filter */ if (bbr->rc_use_google || bbr_probertt_sets_rtt) bbr->rc_prtt_set_ts = 1; if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0); bbr->r_ctl.rc_rtt_shrinks = cts; bbr->r_ctl.last_in_probertt = cts; bbr->r_ctl.rc_probertt_srttchktim = cts; bbr->r_ctl.rc_bbr_state_time = cts; bbr->rc_bbr_state = BBR_STATE_PROBE_RTT; /* We need to force the filter to update */ if ((bbr_sub_drain_slam_cwnd) && bbr->rc_hit_state_1 && (bbr->rc_use_google == 0) && (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd) bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; } else bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; /* Update the lost */ bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){ /* Set to the non-configurable default of 4 (PROBE_RTT_MIN) */ bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6); bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd; } else { /* * We bring it down slowly by using a hptsi gain that is * probably 75%. This will slowly float down our outstanding * without tampering with the cwnd. */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_set_state_target(bbr, __LINE__); if (bbr_prtt_slam_cwnd && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } } if (ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= bbr->r_ctl.rc_target_at_state) { /* We are at target */ bbr->r_ctl.rc_bbr_enters_probertt = cts; } else { /* We need to come down to reach target before our time begins */ bbr->r_ctl.rc_bbr_enters_probertt = 0; } bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch; BBR_STAT_INC(bbr_enter_probertt); bbr_log_exit_gain(bbr, cts, 0); bbr_log_type_statechange(bbr, cts, line); } static void bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts) { /* * Sanity check on probe-rtt intervals. * In crazy situations where we are competing * against new-reno flows with huge buffers * our rtt-prop interval could come to dominate * things if we can't get through a full set * of cycles, we need to adjust it. */ if (bbr_can_adjust_probertt && (bbr->rc_use_google == 0)) { uint16_t val = 0; uint32_t cur_rttp, fval, newval, baseval; /* Are we to small and go into probe-rtt to often? */ baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1)); cur_rttp = roundup(baseval, USECS_IN_SECOND); fval = bbr_filter_len_sec * USECS_IN_SECOND; if (bbr_is_ratio == 0) { if (fval > bbr_rtt_probe_limit) newval = cur_rttp + (fval - bbr_rtt_probe_limit); else newval = cur_rttp; } else { int mul; mul = fval / bbr_rtt_probe_limit; newval = cur_rttp * mul; } if (cur_rttp > bbr->r_ctl.rc_probertt_int) { bbr->r_ctl.rc_probertt_int = cur_rttp; reset_time_small(&bbr->r_ctl.rc_rttprop, newval); val = 1; } else { /* * No adjustments were made * do we need to shrink it? */ if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) { if (cur_rttp <= bbr_rtt_probe_limit) { /* * Things have calmed down lets * shrink all the way to default */ bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; reset_time_small(&bbr->r_ctl.rc_rttprop, (bbr_filter_len_sec * USECS_IN_SECOND)); cur_rttp = bbr_rtt_probe_limit; newval = (bbr_filter_len_sec * USECS_IN_SECOND); val = 2; } else { /* * Well does some adjustment make sense? */ if (cur_rttp < bbr->r_ctl.rc_probertt_int) { /* We can reduce interval time some */ bbr->r_ctl.rc_probertt_int = cur_rttp; reset_time_small(&bbr->r_ctl.rc_rttprop, newval); val = 3; } } } } if (val) bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val); } } static void bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* Exit probe-rtt */ if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) { tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } bbr_log_exit_gain(bbr, cts, 1); bbr->rc_hit_state_1 = 0; bbr->r_ctl.rc_rtt_shrinks = cts; bbr->r_ctl.last_in_probertt = cts; bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0); bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } if (bbr->rc_filled_pipe) { /* Switch to probe_bw */ bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; bbr_substate_change(bbr, cts, __LINE__, 0); bbr_log_type_statechange(bbr, cts, __LINE__); } else { /* Back to startup */ bbr->rc_bbr_state = BBR_STATE_STARTUP; bbr->r_ctl.rc_bbr_state_time = cts; /* * We don't want to give a complete free 3 * measurements until we exit, so we use * the number of pe's we were in probe-rtt * to add to the startup_epoch. That way * we will still retain the old state. */ bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt); bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; /* Make sure to use the lower pg when shifting back in */ if (bbr->r_ctl.rc_lost && bbr_use_lower_gain_in_startup && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; else bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; /* Probably not needed but set it anyway */ bbr_set_state_target(bbr, __LINE__); bbr_log_type_statechange(bbr, cts, __LINE__); bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0); } bbr_check_probe_rtt_limits(bbr, cts); } static int32_t inline bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts) { if ((bbr->rc_past_init_win == 1) && (bbr->rc_in_persist == 0) && (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) { return (1); } if (bbr_can_force_probertt && (bbr->rc_in_persist == 0) && (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { return (1); } return (0); } static int32_t bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t pkt_epoch) { uint64_t btlbw, gain; if (pkt_epoch == 0) { /* * Need to be on a pkt-epoch to continue. */ return (0); } btlbw = bbr_get_full_bw(bbr); gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; if (btlbw >= gain) { bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; } if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) return (1); bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); return(0); } static int32_t inline bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch) { /* Have we gained 25% in the last 3 packet based epoch's? */ uint64_t btlbw, gain; int do_exit; int delta, rtt_gain; if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { /* * This qualifies as a RTT_PROBE session since we drop the * data outstanding to nothing and waited more than * bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } if (bbr_should_enter_probe_rtt(bbr, cts)) { bbr_enter_probe_rtt(bbr, cts, __LINE__); return (0); } if (bbr->rc_use_google) return (bbr_google_startup(bbr, cts, pkt_epoch)); if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && (bbr_use_lower_gain_in_startup)) { /* Drop to a lower gain 1.5 x since we saw loss */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; } if (pkt_epoch == 0) { /* * Need to be on a pkt-epoch to continue. */ return (0); } if (bbr_rtt_gain_thresh) { /* * Do we allow a flow to stay * in startup with no loss and no * gain in rtt over a set threshold? */ if (bbr->r_ctl.rc_pkt_epoch_rtt && bbr->r_ctl.startup_last_srtt && (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) { delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt; rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt; } else rtt_gain = 0; if ((bbr->r_ctl.startup_last_srtt == 0) || (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt)) /* First time or new lower value */ bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt; if ((bbr->r_ctl.rc_lost == 0) && (rtt_gain < bbr_rtt_gain_thresh)) { /* * No loss, and we are under * our gain threhold for * increasing RTT. */ if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) bbr->r_ctl.rc_bbr_last_startup_epoch++; bbr_log_startup_event(bbr, cts, rtt_gain, delta, bbr->r_ctl.startup_last_srtt, 10); return (0); } } if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) && (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) && (!IN_RECOVERY(bbr->rc_tp->t_flags))) { /* * We only assess if we have a new measurment when * we have no loss and are not in recovery. * Drag up by one our last_startup epoch so we will hold * the number of non-gain we have already accumulated. */ if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) bbr->r_ctl.rc_bbr_last_startup_epoch++; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9); return (0); } /* Case where we reduced the lost (bad retransmit) */ if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost) bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count; btlbw = bbr_get_full_bw(bbr); if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower) gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; else gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; do_exit = 0; if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw) bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; if (btlbw >= gain) { bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; /* Update the lost so we won't exit in next set of tests */ bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); } if ((bbr->rc_loss_exit && (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) && ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) { /* * If we had no gain, we had loss and that loss was above * our threshould, the rwnd is not constrained, and we have * had at least 3 packet epochs exit. Note that this is * switched off by sysctl. Google does not do this by the * way. */ if ((ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) { do_exit = 1; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4); } else { /* Just record an updated loss value */ bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5); } } else bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) || do_exit) { /* Return 1 to exit the startup state. */ return (1); } /* Stay in startup */ bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); return (0); } static void bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses) { /* * A tick occurred in the rtt epoch do we need to do anything? */ #ifdef BBR_INVARIANTS if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) && (bbr->rc_bbr_state != BBR_STATE_DRAIN) && (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) && (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) { /* Debug code? */ panic("Unknown BBR state %d?\n", bbr->rc_bbr_state); } #endif if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { /* Do we exit the startup state? */ if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) { uint32_t time_in; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6); bbr->rc_filled_pipe = 1; bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } else time_in = 0; if (bbr->rc_no_pacing) bbr->rc_no_pacing = 0; bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg; bbr->rc_bbr_state = BBR_STATE_DRAIN; bbr_set_state_target(bbr, __LINE__); if ((bbr->rc_use_google == 0) && bbr_slam_cwnd_in_main_drain) { /* Here we don't have to worry about probe-rtt */ bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; bbr_log_type_statechange(bbr, cts, __LINE__); if (ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= bbr->r_ctl.rc_target_at_state) { /* * Switch to probe_bw if we are already * there */ bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr_substate_change(bbr, cts, __LINE__, 0); bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr_log_type_statechange(bbr, cts, __LINE__); } } } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) { uint32_t inflight; struct tcpcb *tp; tp = bbr->rc_tp; inflight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (inflight >= bbr->r_ctl.rc_target_at_state) { /* We have reached a flight of the cwnd target */ bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_set_state_target(bbr, __LINE__); /* * Rig it so we don't do anything crazy and * start fresh with a new randomization. */ bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; bbr->rc_bbr_substate = BBR_SUB_LEVEL6; bbr_substate_change(bbr, cts, __LINE__, 1); } } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) { /* Has in-flight reached the bdp (or less)? */ uint32_t inflight; struct tcpcb *tp; tp = bbr->rc_tp; inflight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if ((bbr->rc_use_google == 0) && bbr_slam_cwnd_in_main_drain && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { /* * Here we don't have to worry about probe-rtt * re-slam it, but keep it slammed down. */ bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (inflight <= bbr->r_ctl.rc_target_at_state) { /* We have drained */ bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } if ((bbr->rc_use_google == 0) && bbr_slam_cwnd_in_main_drain && (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { /* Restore the cwnd */ tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } /* Setup probe-rtt has being done now RRS-HERE */ bbr->r_ctl.rc_rtt_shrinks = cts; bbr->r_ctl.last_in_probertt = cts; bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0); /* Randomly pick a sub-state */ bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr_substate_change(bbr, cts, __LINE__, 0); bbr_log_type_statechange(bbr, cts, __LINE__); } } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) { uint32_t flight; flight = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered); if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { /* * We must keep cwnd at the desired MSS. */ bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } else if ((bbr_prtt_slam_cwnd) && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { /* Re-slam it */ bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (bbr->r_ctl.rc_bbr_enters_probertt == 0) { /* Has outstanding reached our target? */ if (flight <= bbr->r_ctl.rc_target_at_state) { bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0); bbr->r_ctl.rc_bbr_enters_probertt = cts; /* If time is exactly 0, be 1usec off */ if (bbr->r_ctl.rc_bbr_enters_probertt == 0) bbr->r_ctl.rc_bbr_enters_probertt = 1; if (bbr->rc_use_google == 0) { /* * Restore any lowering that as occurred to * reach here */ if (bbr->r_ctl.bbr_rttprobe_gain_val) bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; else bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; } } if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) && (bbr->rc_use_google == 0) && bbr->r_ctl.bbr_rttprobe_gain_val && (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) || (flight >= bbr->r_ctl.flightsize_at_drain))) { /* * We have doddled with our current hptsi * gain an srtt and have still not made it * to target, or we have increased our flight. * Lets reduce the gain by xx% * flooring the reduce at DRAIN (based on * mul/div) */ int red; bbr->r_ctl.flightsize_at_drain = flight; bbr->r_ctl.rc_probertt_srttchktim = cts; red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1); if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) { /* Reduce our gain again */ bbr->r_ctl.rc_bbr_hptsi_gain -= red; bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0); } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) { /* one more chance before we give up */ bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0); } else { /* At the very bottom */ bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1); } } } if (bbr->r_ctl.rc_bbr_enters_probertt && (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) && ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) { /* Time to exit probe RTT normally */ bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts); } } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { /* * This qualifies as a RTT_PROBE session since we * drop the data outstanding to nothing and waited * more than bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } if (bbr_should_enter_probe_rtt(bbr, cts)) { bbr_enter_probe_rtt(bbr, cts, __LINE__); } else { bbr_set_probebw_gains(bbr, cts, losses); } } } static void bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses) { int32_t epoch = 0; if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { bbr_set_epoch(bbr, cts, line); /* At each epoch doe lt bw sampling */ epoch = 1; } bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses); } static int bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt, struct timeval *tv) { int32_t thflags, retval; uint32_t cts, lcts; uint32_t tiwin; struct tcpopt to; struct tcp_bbr *bbr; struct bbr_sendmap *rsm; struct timeval ltv; int32_t did_out = 0; int32_t in_recovery; uint16_t nsegs; int32_t prev_state; uint32_t lost; nsegs = max(1, m->m_pkthdr.lro_nsegs); bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* add in our stats */ kern_prefetch(bbr, &prev_state); prev_state = 0; thflags = th->th_flags; /* * If this is either a state-changing packet or current state isn't * established, we require a write lock on tcbinfo. Otherwise, we * allow the tcbinfo to be in either alocked or unlocked, as the * caller may have unnecessarily acquired a write lock due to a * race. */ INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", __func__)); tp->t_rcvtime = ticks; /* * Unscale the window into a 32-bit value. For the SYN_SENT state * the scale is zero. */ tiwin = th->th_win << tp->snd_scale; #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); #endif if (m->m_flags & M_TSTMP) { /* Prefer the hardware timestamp if present */ struct timespec ts; mbuf_tstmp2timespec(m, &ts); bbr->rc_tv.tv_sec = ts.tv_sec; bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); } else if (m->m_flags & M_TSTMP_LRO) { /* Next the arrival timestamp */ struct timespec ts; mbuf_tstmp2timespec(m, &ts); bbr->rc_tv.tv_sec = ts.tv_sec; bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); } else { /* * Ok just get the current time. */ bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv); } /* * Parse options on any incoming segment. */ tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); /* * If timestamps were negotiated during SYN/ACK and a * segment without a timestamp is received, silently drop * the segment, unless it is a RST segment or missing timestamps are * tolerated. * See section 3.2 of RFC 7323. */ if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { retval = 0; m_freem(m); goto done_with_input; } /* * If echoed timestamp is later than the current time, fall back to * non RFC1323 RTT calculation. Normalize timestamp if syncookies * were used when this connection was established. */ if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { to.to_tsecr -= tp->ts_offset; if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv))) to.to_tsecr = 0; } /* * If its the first time in we need to take care of options and * verify we can do SACK for rack! */ if (bbr->r_state == 0) { /* * Process options only when we get SYN/ACK back. The SYN * case for incoming connections is handled in tcp_syncache. * According to RFC1323 the window field in a SYN (i.e., a * or ) segment itself is never scaled. XXX * this is traditional behavior, may need to be cleaned up. */ if (bbr->rc_inp == NULL) { bbr->rc_inp = tp->t_inpcb; } /* * We need to init rc_inp here since its not init'd when * bbr_init is called */ if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) { tp->t_flags |= TF_RCVD_SCALE; tp->snd_scale = to.to_wscale; } else tp->t_flags &= ~TF_REQ_SCALE; /* * Initial send window. It will be updated with the * next incoming segment to the scaled value. */ tp->snd_wnd = th->th_win; if ((to.to_flags & TOF_TS) && (tp->t_flags & TF_REQ_TSTMP)) { tp->t_flags |= TF_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); } else tp->t_flags &= ~TF_REQ_TSTMP; if (to.to_flags & TOF_MSS) tcp_mss(tp, to.to_mss); if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACKPERM) == 0) tp->t_flags &= ~TF_SACK_PERMIT; if (IS_FASTOPEN(tp->t_flags)) { if (to.to_flags & TOF_FASTOPEN) { uint16_t mss; if (to.to_flags & TOF_MSS) mss = to.to_mss; else if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) mss = TCP6_MSS; else mss = TCP_MSS; tcp_fastopen_update_cache(tp, mss, to.to_tfo_len, to.to_tfo_cookie); } else tcp_fastopen_disable_path(tp); } } /* * At this point we are at the initial call. Here we decide * if we are doing RACK or not. We do this by seeing if * TF_SACK_PERMIT is set, if not rack is *not* possible and * we switch to the default code. */ if ((tp->t_flags & TF_SACK_PERMIT) == 0) { /* Bail */ tcp_switch_back_to_default(tp); (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, tlen, iptos); return (1); } /* Set the flag */ bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; tcp_set_hpts(tp->t_inpcb); sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack); } if (thflags & TH_ACK) { /* Track ack types */ if (to.to_flags & TOF_SACK) BBR_STAT_INC(bbr_acks_with_sacks); else BBR_STAT_INC(bbr_plain_acks); } /* * This is the one exception case where we set the rack state * always. All other times (timers etc) we must have a rack-state * set (so we assure we have done the checks above for SACK). */ if (thflags & TH_FIN) tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); if (bbr->r_state != tp->t_state) bbr_set_state(tp, bbr, tiwin); if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL) kern_prefetch(rsm, &prev_state); prev_state = bbr->r_state; bbr->rc_ack_was_delayed = 0; lost = bbr->r_ctl.rc_lost; bbr->rc_is_pkt_epoch_now = 0; if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) { /* Get the real time into lcts and figure the real delay */ lcts = tcp_get_usecs(<v); if (TSTMP_GT(lcts, cts)) { bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts; bbr->rc_ack_was_delayed = 1; if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay, bbr->r_ctl.highest_hdwr_delay)) bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay; } else { bbr->r_ctl.rc_ack_hdwr_delay = 0; bbr->rc_ack_was_delayed = 0; } } else { bbr->r_ctl.rc_ack_hdwr_delay = 0; bbr->rc_ack_was_delayed = 0; } bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m); if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { retval = 0; m_freem(m); goto done_with_input; } /* * If a segment with the ACK-bit set arrives in the SYN-SENT state * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. */ if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } in_recovery = IN_RECOVERY(tp->t_flags); if (tiwin > bbr->r_ctl.rc_high_rwnd) bbr->r_ctl.rc_high_rwnd = tiwin; #ifdef BBR_INVARIANTS if ((tp->t_inpcb->inp_flags & INP_DROPPED) || (tp->t_inpcb->inp_flags2 & INP_FREED)) { panic("tp:%p bbr:%p given a dropped inp:%p", tp, bbr, tp->t_inpcb); } #endif bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->rtt_valid = 0; if (to.to_flags & TOF_TS) { bbr->rc_ts_valid = 1; bbr->r_ctl.last_inbound_ts = to.to_tsval; } else { bbr->rc_ts_valid = 0; bbr->r_ctl.last_inbound_ts = 0; } retval = (*bbr->r_substate) (m, th, so, tp, &to, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt, iptos); #ifdef BBR_INVARIANTS if ((retval == 0) && (tp->t_inpcb == NULL)) { panic("retval:%d tp:%p t_inpcb:NULL state:%d", retval, tp, prev_state); } #endif if (nxt_pkt == 0) BBR_STAT_INC(bbr_rlock_left_ret0); else BBR_STAT_INC(bbr_rlock_left_ret1); if (retval == 0) { /* * If retval is 1 the tcb is unlocked and most likely the tp * is gone. */ INP_WLOCK_ASSERT(tp->t_inpcb); tcp_bbr_xmit_timer_commit(bbr, tp, cts); if (bbr->rc_is_pkt_epoch_now) bbr_set_pktepoch(bbr, cts, __LINE__); bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost)); if (nxt_pkt == 0) { if (bbr->r_wanted_output != 0) { bbr->rc_output_starts_timer = 0; did_out = 1; (void)tp->t_fb->tfb_tcp_output(tp); } else bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0); } if ((nxt_pkt == 0) && ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && (SEQ_GT(tp->snd_max, tp->snd_una) || (tp->t_flags & TF_DELACK) || ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)))) { /* * We could not send (probably in the hpts but * stopped the timer)? */ if ((tp->snd_max == tp->snd_una) && ((tp->t_flags & TF_DELACK) == 0) && (bbr->rc_inp->inp_in_hpts) && (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* * keep alive not needed if we are hptsi * output yet */ ; } else { if (bbr->rc_inp->inp_in_hpts) { tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT); if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && (TSTMP_GT(lcts, bbr->rc_pacer_started))) { uint32_t del; del = lcts - bbr->rc_pacer_started; if (bbr->r_ctl.rc_last_delay_val > del) { BBR_STAT_INC(bbr_force_timer_start); bbr->r_ctl.rc_last_delay_val -= del; bbr->rc_pacer_started = lcts; } else { /* We are late */ bbr->r_ctl.rc_last_delay_val = 0; BBR_STAT_INC(bbr_force_output); (void)tp->t_fb->tfb_tcp_output(tp); } } } bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val, 0); } } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) { /* Do we have the correct timer running? */ bbr_timer_audit(tp, bbr, lcts, &so->so_snd); } /* Do we have a new state */ if (bbr->r_state != tp->t_state) bbr_set_state(tp, bbr, tiwin); done_with_input: bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out); if (did_out) bbr->r_wanted_output = 0; #ifdef BBR_INVARIANTS if (tp->t_inpcb == NULL) { panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", did_out, retval, tp, prev_state); } #endif } return (retval); } static void bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) { struct timeval tv; int retval; /* First lets see if we have old packets */ if (tp->t_in_pkt) { if (ctf_do_queued_segments(so, tp, 1)) { m_freem(m); return; } } if (m->m_flags & M_TSTMP_LRO) { tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000; tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000; } else { /* Should not be should we kassert instead? */ tcp_get_usecs(&tv); } retval = bbr_do_segment_nounlock(m, th, so, tp, drop_hdrlen, tlen, iptos, 0, &tv); if (retval == 0) { INP_WUNLOCK(tp->t_inpcb); } } /* * Return how much data can be sent without violating the * cwnd or rwnd. */ static inline uint32_t bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin, uint32_t avail, int32_t sb_offset, uint32_t cts) { uint32_t len; if (ctf_outstanding(tp) >= tp->snd_wnd) { /* We never want to go over our peers rcv-window */ len = 0; } else { uint32_t flight; flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (flight >= sendwin) { /* * We have in flight what we are allowed by cwnd (if * it was rwnd blocking it would have hit above out * >= tp->snd_wnd). */ return (0); } len = sendwin - flight; if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { /* We would send too much (beyond the rwnd) */ len = tp->snd_wnd - ctf_outstanding(tp); } if ((len + sb_offset) > avail) { /* * We don't have that much in the SB, how much is * there? */ len = avail - sb_offset; } } return (len); } static inline void bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) { #ifdef NETFLIX_STATS KMOD_TCPSTAT_INC(tcps_sndpack_error); KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len); #endif } static inline void bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) { if (error) { bbr_do_error_accounting(tp, bbr, rsm, len, error); return; } if (rsm) { if (rsm->r_flags & BBR_TLP) { /* * TLP should not count in retran count, but in its * own bin */ #ifdef NETFLIX_STATS KMOD_TCPSTAT_INC(tcps_tlpresends); KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len); #endif } else { /* Retransmit */ tp->t_sndrexmitpack++; KMOD_TCPSTAT_INC(tcps_sndrexmitpack); KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif } /* * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is * sub-state */ counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len); if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) { /* Non probe_bw log in 1, 2, or 4. */ counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len); } else { /* * Log our probe state 3, and log also 5-13 to show * us the recovery sub-state for the send. This * means that 3 == (5+6+7+8+9+10+11+12+13) */ counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len); counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len); } /* Place in both 16's the totals of retransmitted */ counter_u64_add(bbr_state_lost[16], len); counter_u64_add(bbr_state_resend[16], len); /* Place in 17's the total sent */ counter_u64_add(bbr_state_resend[17], len); counter_u64_add(bbr_state_lost[17], len); } else { /* New sends */ KMOD_TCPSTAT_INC(tcps_sndpack); KMOD_TCPSTAT_ADD(tcps_sndbyte, len); /* Place in 17's the total sent */ counter_u64_add(bbr_state_resend[17], len); counter_u64_add(bbr_state_lost[17], len); #ifdef STATS stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif } } static void bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level) { if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) { /* * Limit the cwnd to not be above N x the target plus whats * is outstanding. The target is based on the current b/w * estimate. */ uint32_t target; target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT); target += ctf_outstanding(tp); target *= bbr_target_cwnd_mult_limit; if (tp->snd_cwnd > target) tp->snd_cwnd = target; bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__); } } static int bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg) { /* * "adv" is the amount we could increase the window, taking into * account that we are limited by TCP_MAXWIN << tp->rcv_scale. */ int32_t adv; int32_t oldwin; adv = recwin; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { oldwin = (tp->rcv_adv - tp->rcv_nxt); if (adv > oldwin) adv -= oldwin; else { /* We can't increase the window */ adv = 0; } } else oldwin = 0; /* * If the new window size ends up being the same as or less * than the old size when it is scaled, then don't force * a window update. */ if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) return (0); if (adv >= (2 * maxseg) && (adv >= (so->so_rcv.sb_hiwat / 4) || recwin <= (so->so_rcv.sb_hiwat / 8) || so->so_rcv.sb_hiwat <= 8 * maxseg)) { return (1); } if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) return (1); return (0); } /* * Return 0 on success and a errno on failure to send. * Note that a 0 return may not mean we sent anything * if the TCB was on the hpts. A non-zero return * does indicate the error we got from ip[6]_output. */ static int bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv) { struct socket *so; int32_t len; uint32_t cts; uint32_t recwin, sendwin; int32_t sb_offset; int32_t flags, abandon, error = 0; struct tcp_log_buffer *lgb = NULL; struct mbuf *m; struct mbuf *mb; uint32_t if_hw_tsomaxsegcount = 0; uint32_t if_hw_tsomaxsegsize = 0; uint32_t if_hw_tsomax = 0; struct ip *ip = NULL; #ifdef TCPDEBUG struct ipovly *ipov = NULL; #endif struct tcp_bbr *bbr; struct tcphdr *th; struct udphdr *udp = NULL; u_char opt[TCP_MAXOLEN]; unsigned ipoptlen, optlen, hdrlen; unsigned ulen; uint32_t bbr_seq; uint32_t delay_calc=0; uint8_t doing_tlp = 0; uint8_t local_options; #ifdef BBR_INVARIANTS uint8_t doing_retran_from = 0; uint8_t picked_up_retran = 0; #endif uint8_t wanted_cookie = 0; uint8_t more_to_rxt=0; int32_t prefetch_so_done = 0; int32_t prefetch_rsm = 0; uint32_t what_we_can = 0; uint32_t tot_len = 0; uint32_t rtr_cnt = 0; uint32_t maxseg, pace_max_segs, p_maxseg; int32_t csum_flags; int32_t hw_tls; #if defined(IPSEC) || defined(IPSEC_SUPPORT) unsigned ipsec_optlen = 0; #endif volatile int32_t sack_rxmit; struct bbr_sendmap *rsm = NULL; int32_t tso, mtu; struct tcpopt to; int32_t slot = 0; struct inpcb *inp; struct sockbuf *sb; uint32_t hpts_calling; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int32_t isipv6; #endif uint8_t app_limited = BBR_JR_SENT_DATA; uint8_t filled_all = 0; bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* We take a cache hit here */ memcpy(&bbr->rc_tv, tv, sizeof(struct timeval)); cts = tcp_tv_to_usectick(&bbr->rc_tv); inp = bbr->rc_inp; so = inp->inp_socket; sb = &so->so_snd; if (sb->sb_flags & SB_TLS_IFNET) hw_tls = 1; else hw_tls = 0; kern_prefetch(sb, &maxseg); maxseg = tp->t_maxseg - bbr->rc_last_options; if (bbr_minseg(bbr) < maxseg) { tcp_bbr_tso_size_check(bbr, cts); } /* Remove any flags that indicate we are pacing on the inp */ pace_max_segs = bbr->r_ctl.rc_pace_max_segs; p_maxseg = min(maxseg, pace_max_segs); INP_WLOCK_ASSERT(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) return (tcp_offload_output(tp)); #endif #ifdef INET6 if (bbr->r_state) { /* Use the cache line loaded if possible */ isipv6 = bbr->r_is_v6; } else { isipv6 = (inp->inp_vflag & INP_IPV6) != 0; } #endif if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && inp->inp_in_hpts) { /* * We are on the hpts for some timer but not hptsi output. * Possibly remove from the hpts so we can send/recv etc. */ if ((tp->t_flags & TF_ACKNOW) == 0) { /* * No immediate demand right now to send an ack, but * the user may have read, making room for new data * (a window update). If so we may want to cancel * whatever timer is running (KEEP/DEL-ACK?) and * continue to send out a window update. Or we may * have gotten more data into the socket buffer to * send. */ recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) && ((tcp_outflags[tp->t_state] & TH_RST) == 0) && ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <= (tp->snd_max - tp->snd_una))) { /* * Nothing new to send and no window update * is needed to send. Lets just return and * let the timer-run off. */ return (0); } } tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); bbr_timer_cancel(bbr, __LINE__, cts); } if (bbr->r_ctl.rc_last_delay_val) { /* Calculate a rough delay for early escape to sending */ if (SEQ_GT(cts, bbr->rc_pacer_started)) delay_calc = cts - bbr->rc_pacer_started; if (delay_calc >= bbr->r_ctl.rc_last_delay_val) delay_calc -= bbr->r_ctl.rc_last_delay_val; else delay_calc = 0; } /* Mark that we have called bbr_output(). */ if ((bbr->r_timer_override) || (tp->t_state < TCPS_ESTABLISHED)) { /* Timeouts or early states are exempt */ if (inp->inp_in_hpts) tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); } else if (inp->inp_in_hpts) { if ((bbr->r_ctl.rc_last_delay_val) && (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && delay_calc) { /* * We were being paced for output and the delay has * already exceeded when we were supposed to be * called, lets go ahead and pull out of the hpts * and call output. */ counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1); bbr->r_ctl.rc_last_delay_val = 0; tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); } else if (tp->t_state == TCPS_CLOSED) { bbr->r_ctl.rc_last_delay_val = 0; tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); } else { /* * On the hpts, you shall not pass! even if ACKNOW * is on, we will when the hpts fires, unless of * course we are overdue. */ counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1); return (0); } } bbr->rc_cwnd_limited = 0; if (bbr->r_ctl.rc_last_delay_val) { /* recalculate the real delay and deal with over/under */ if (SEQ_GT(cts, bbr->rc_pacer_started)) delay_calc = cts - bbr->rc_pacer_started; else delay_calc = 0; if (delay_calc >= bbr->r_ctl.rc_last_delay_val) /* Setup the delay which will be added in */ delay_calc -= bbr->r_ctl.rc_last_delay_val; else { /* * We are early setup to adjust * our slot time. */ uint64_t merged_val; bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc); bbr->r_agg_early_set = 1; if (bbr->r_ctl.rc_hptsi_agg_delay) { if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) { /* Nope our previous late cancels out the early */ bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; } else { bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay; bbr->r_ctl.rc_hptsi_agg_delay = 0; } } merged_val = bbr->rc_pacer_started; merged_val <<= 32; merged_val |= bbr->r_ctl.rc_last_delay_val; bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls, bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val, bbr->r_agg_early_set, 3); bbr->r_ctl.rc_last_delay_val = 0; BBR_STAT_INC(bbr_early); delay_calc = 0; } } else { /* We were not delayed due to hptsi */ if (bbr->r_agg_early_set) bbr->r_ctl.rc_agg_early = 0; bbr->r_agg_early_set = 0; delay_calc = 0; } if (delay_calc) { /* * We had a hptsi delay which means we are falling behind on * sending at the expected rate. Calculate an extra amount * of data we can send, if any, to put us back on track. */ if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay) bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff; else bbr->r_ctl.rc_hptsi_agg_delay += delay_calc; } sendwin = min(tp->snd_wnd, tp->snd_cwnd); if ((tp->snd_una == tp->snd_max) && (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && (sbavail(sb))) { /* * Ok we have been idle with nothing outstanding * we possibly need to start fresh with either a new * suite of states or a fast-ramp up. */ bbr_restart_after_idle(bbr, cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time)); } /* * Now was there a hptsi delay where we are behind? We only count * being behind if: a) We are not in recovery. b) There was a delay. * c) We had room to send something. * */ hpts_calling = inp->inp_hpts_calls; inp->inp_hpts_calls = 0; if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { if (bbr_process_timers(tp, bbr, cts, hpts_calling)) { counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1); return (0); } } bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; if (hpts_calling && (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { bbr->r_ctl.rc_last_delay_val = 0; } bbr->r_timer_override = 0; bbr->r_wanted_output = 0; /* * For TFO connections in SYN_RECEIVED, only allow the initial * SYN|ACK and those sent by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && ((tp->t_state == TCPS_SYN_RECEIVED) || (tp->t_state == TCPS_SYN_SENT)) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ (tp->t_rxtshift == 0)) { /* not a retransmit */ len = 0; goto just_return_nolock; } /* * Before sending anything check for a state update. For hpts * calling without input this is important. If its input calling * then this was already done. */ if (bbr->rc_use_google == 0) bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); again: /* * If we've recently taken a timeout, snd_max will be greater than * snd_max. BBR in general does not pay much attention to snd_nxt * for historic reasons the persist timer still uses it. This means * we have to look at it. All retransmissions that are not persits * use the rsm that needs to be sent so snd_nxt is ignored. At the * end of this routine we pull snd_nxt always up to snd_max. */ doing_tlp = 0; #ifdef BBR_INVARIANTS doing_retran_from = picked_up_retran = 0; #endif error = 0; tso = 0; slot = 0; mtu = 0; sendwin = min(tp->snd_wnd, tp->snd_cwnd); sb_offset = tp->snd_max - tp->snd_una; flags = tcp_outflags[tp->t_state]; sack_rxmit = 0; len = 0; rsm = NULL; if (flags & TH_RST) { SOCKBUF_LOCK(sb); goto send; } recheck_resend: while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { /* We need to always have one in reserve */ rsm = bbr_alloc(bbr); if (rsm == NULL) { error = ENOMEM; /* Lie to get on the hpts */ tot_len = tp->t_maxseg; if (hpts_calling) /* Retry in a ms */ slot = 1001; goto just_return_nolock; } TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); bbr->r_ctl.rc_free_cnt++; rsm = NULL; } /* What do we send, a resend? */ if (bbr->r_ctl.rc_resend == NULL) { /* Check for rack timeout */ bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); if (bbr->r_ctl.rc_resend) { #ifdef BBR_INVARIANTS picked_up_retran = 1; #endif bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend); } } if (bbr->r_ctl.rc_resend) { rsm = bbr->r_ctl.rc_resend; #ifdef BBR_INVARIANTS doing_retran_from = 1; #endif /* Remove any TLP flags its a RACK or T-O */ rsm->r_flags &= ~BBR_TLP; bbr->r_ctl.rc_resend = NULL; if (SEQ_LT(rsm->r_start, tp->snd_una)) { #ifdef BBR_INVARIANTS panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n", tp, bbr, rsm, rsm->r_start, tp->snd_una); goto recheck_resend; #else /* TSNH */ rsm = NULL; goto recheck_resend; #endif } rtr_cnt++; if (rsm->r_flags & BBR_HAS_SYN) { /* Only retransmit a SYN by itself */ len = 0; if ((flags & TH_SYN) == 0) { /* Huh something is wrong */ rsm->r_start++; if (rsm->r_start == rsm->r_end) { /* Clean it up, somehow we missed the ack? */ bbr_log_syn(tp, NULL); } else { /* TFO with data? */ rsm->r_flags &= ~BBR_HAS_SYN; len = rsm->r_end - rsm->r_start; } } else { /* Retransmitting SYN */ rsm = NULL; SOCKBUF_LOCK(sb); goto send; } } else len = rsm->r_end - rsm->r_start; if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) { len = maxseg; more_to_rxt = 1; } sb_offset = rsm->r_start - tp->snd_una; if (len > 0) { sack_rxmit = 1; KMOD_TCPSTAT_INC(tcps_sack_rexmits); KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, min(len, maxseg)); } else { /* I dont think this can happen */ rsm = NULL; goto recheck_resend; } BBR_STAT_INC(bbr_resends_set); } else if (bbr->r_ctl.rc_tlp_send) { /* * Tail loss probe */ doing_tlp = 1; rsm = bbr->r_ctl.rc_tlp_send; bbr->r_ctl.rc_tlp_send = NULL; sack_rxmit = 1; len = rsm->r_end - rsm->r_start; rtr_cnt++; if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) len = maxseg; if (SEQ_GT(tp->snd_una, rsm->r_start)) { #ifdef BBR_INVARIANTS panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u", tp, bbr, tp->snd_una, rsm, rsm->r_start); #else /* TSNH */ rsm = NULL; goto recheck_resend; #endif } sb_offset = rsm->r_start - tp->snd_una; BBR_STAT_INC(bbr_tlp_set); } /* * Enforce a connection sendmap count limit if set * as long as we are not retransmiting. */ if ((rsm == NULL) && (V_tcp_map_entries_limit > 0) && (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { BBR_STAT_INC(bbr_alloc_limited); if (!bbr->alloc_limit_reported) { bbr->alloc_limit_reported = 1; BBR_STAT_INC(bbr_alloc_limited_conns); } goto just_return_nolock; } #ifdef BBR_INVARIANTS if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) { panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u", tp, bbr, rsm, sb_offset, len); } #endif /* * Get standard flags, and add SYN or FIN if requested by 'hidden' * state flags. */ if (tp->t_flags & TF_NEEDFIN && (rsm == NULL)) flags |= TH_FIN; if (tp->t_flags & TF_NEEDSYN) flags |= TH_SYN; if (rsm && (rsm->r_flags & BBR_HAS_FIN)) { /* we are retransmitting the fin */ len--; if (len) { /* * When retransmitting data do *not* include the * FIN. This could happen from a TLP probe if we * allowed data with a FIN. */ flags &= ~TH_FIN; } } else if (rsm) { if (flags & TH_FIN) flags &= ~TH_FIN; } if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { void *end_rsm; end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); if (end_rsm) kern_prefetch(end_rsm, &prefetch_rsm); prefetch_rsm = 1; } SOCKBUF_LOCK(sb); /* * If snd_nxt == snd_max and we have transmitted a FIN, the * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a * negative length. This can also occur when TCP opens up its * congestion window while receiving additional duplicate acks after * fast-retransmit because TCP will reset snd_nxt to snd_max after * the fast-retransmit. * * In the normal retransmit-FIN-only case, however, snd_nxt will be * set to snd_una, the sb_offset will be 0, and the length may wind * up 0. * * If sack_rxmit is true we are retransmitting from the scoreboard * in which case len is already set. */ if (sack_rxmit == 0) { uint32_t avail; avail = sbavail(sb); if (SEQ_GT(tp->snd_max, tp->snd_una)) sb_offset = tp->snd_max - tp->snd_una; else sb_offset = 0; if (bbr->rc_tlp_new_data) { /* TLP is forcing out new data */ uint32_t tlplen; doing_tlp = 1; tlplen = maxseg; if (tlplen > (uint32_t)(avail - sb_offset)) { tlplen = (uint32_t)(avail - sb_offset); } if (tlplen > tp->snd_wnd) { len = tp->snd_wnd; } else { len = tlplen; } bbr->rc_tlp_new_data = 0; } else { what_we_can = len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts); if ((len < p_maxseg) && (bbr->rc_in_persist == 0) && (ctf_outstanding(tp) >= (2 * p_maxseg)) && ((avail - sb_offset) >= p_maxseg)) { /* * We are not completing whats in the socket * buffer (i.e. there is at least a segment * waiting to send) and we have 2 or more * segments outstanding. There is no sense * of sending a little piece. Lets defer and * and wait until we can send a whole * segment. */ len = 0; } if (bbr->rc_in_persist) { /* * We are in persists, figure out if * a retransmit is available (maybe the previous * persists we sent) or if we have to send new * data. */ rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm) { len = rsm->r_end - rsm->r_start; if (rsm->r_flags & BBR_HAS_FIN) len--; if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) len = maxseg; if (len > 1) BBR_STAT_INC(bbr_persist_reneg); /* * XXXrrs we could force the len to * 1 byte here to cause the chunk to * split apart.. but that would then * mean we always retransmit it as * one byte even after the window * opens. */ sack_rxmit = 1; sb_offset = rsm->r_start - tp->snd_una; } else { /* * First time through in persists or peer * acked our one byte. Though we do have * to have something in the sb. */ len = 1; sb_offset = 0; if (avail == 0) len = 0; } } } } if (prefetch_so_done == 0) { kern_prefetch(so, &prefetch_so_done); prefetch_so_done = 1; } /* * Lop off SYN bit if it has already been sent. However, if this is * SYN-SENT state and if segment contains data and if we don't know * that foreign host supports TAO, suppress sending segment. */ if ((flags & TH_SYN) && (rsm == NULL) && SEQ_GT(tp->snd_max, tp->snd_una)) { if (tp->t_state != TCPS_SYN_RECEIVED) flags &= ~TH_SYN; /* * When sending additional segments following a TFO SYN|ACK, * do not include the SYN bit. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED)) flags &= ~TH_SYN; sb_offset--, len++; if (sbavail(sb) == 0) len = 0; } else if ((flags & TH_SYN) && rsm) { /* * Subtract one from the len for the SYN being * retransmitted. */ len--; } /* * Be careful not to send data and/or FIN on SYN segments. This * measure is needed to prevent interoperability problems with not * fully conformant TCP implementations. */ if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { len = 0; flags &= ~TH_FIN; } /* * On TFO sockets, ensure no data is sent in the following cases: * * - When retransmitting SYN|ACK on a passively-created socket * - When retransmitting SYN on an actively created socket * - When sending a zero-length cookie (cookie request) on an * actively created socket * - When the socket is in the CLOSED state (RST is being sent) */ if (IS_FASTOPEN(tp->t_flags) && (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || ((tp->t_state == TCPS_SYN_SENT) && (tp->t_tfo_client_cookie_len == 0)) || (flags & TH_RST))) { len = 0; sack_rxmit = 0; rsm = NULL; } /* Without fast-open there should never be data sent on a SYN */ if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) len = 0; if (len <= 0) { /* * If FIN has been sent but not acked, but we haven't been * called to retransmit, len will be < 0. Otherwise, window * shrank after we sent into it. If window shrank to 0, * cancel pending retransmit, pull snd_nxt back to (closed) * window, and set the persist timer if it isn't already * going. If the window didn't close completely, just wait * for an ACK. * * We also do a general check here to ensure that we will * set the persist timer when we have data to send, but a * 0-byte window. This makes sure the persist timer is set * even if the packet hits one of the "goto send" lines * below. */ len = 0; if ((tp->snd_wnd == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (tp->snd_una == tp->snd_max) && (sb_offset < (int)sbavail(sb))) { /* * Not enough room in the rwnd to send * a paced segment out. */ bbr_enter_persist(tp, bbr, cts, __LINE__); } } else if ((rsm == NULL) && (doing_tlp == 0) && (len < bbr->r_ctl.rc_pace_max_segs)) { /* * We are not sending a full segment for * some reason. Should we not send anything (think * sws or persists)? */ if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (len < (int)(sbavail(sb) - sb_offset))) { /* * Here the rwnd is less than * the pacing size, this is not a retransmit, * we are established and * the send is not the last in the socket buffer * lets not send, and possibly enter persists. */ len = 0; if (tp->snd_max == tp->snd_una) bbr_enter_persist(tp, bbr, cts, __LINE__); } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) && (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && (len < (int)(sbavail(sb) - sb_offset)) && (len < bbr_minseg(bbr))) { /* * Here we are not retransmitting, and * the cwnd is not so small that we could * not send at least a min size (rxt timer * not having gone off), We have 2 segments or * more already in flight, its not the tail end * of the socket buffer and the cwnd is blocking * us from sending out minimum pacing segment size. * Lets not send anything. */ bbr->rc_cwnd_limited = 1; len = 0; } else if (((tp->snd_wnd - ctf_outstanding(tp)) < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && (len < (int)(sbavail(sb) - sb_offset)) && (TCPS_HAVEESTABLISHED(tp->t_state))) { /* * Here we have a send window but we have * filled it up and we can't send another pacing segment. * We also have in flight more than 2 segments * and we are not completing the sb i.e. we allow * the last bytes of the sb to go out even if * its not a full pacing segment. */ len = 0; } } /* len will be >= 0 after this point. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); tcp_sndbuf_autoscale(tp, so, sendwin); /* * */ if (bbr->rc_in_persist && len && (rsm == NULL) && (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) { /* * We are in persist, not doing a retransmit and don't have enough space * yet to send a full TSO. So is it at the end of the sb * if so we need to send else nuke to 0 and don't send. */ int sbleft; if (sbavail(sb) > sb_offset) sbleft = sbavail(sb) - sb_offset; else sbleft = 0; if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) { /* not at end of sb lets not send */ len = 0; } } /* * Decide if we can use TCP Segmentation Offloading (if supported by * hardware). * * TSO may only be used if we are in a pure bulk sending state. The * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP * options prevent using TSO. With TSO the TCP header is the same * (except for the sequence number) for all generated packets. This * makes it impossible to transmit any options which vary per * generated segment or packet. * * IPv4 handling has a clear separation of ip options and ip header * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() * does the right thing below to provide length of just ip options * and thus checking for ipoptlen is enough to decide if ip options * are present. */ #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(inp); else #endif if (inp->inp_options) ipoptlen = inp->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* * Pre-calculate here as we save another lookup into the darknesses * of IPsec that way and can actually decide if TSO is ok. */ #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6)) ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4)) ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); #endif /* INET */ #endif /* IPSEC */ #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && (len > maxseg) && (tp->t_port == 0) && ((tp->t_flags & TF_SIGNATURE) == 0) && tp->rcv_numsacks == 0 && ipoptlen == 0) tso = 1; recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); /* * Sender silly window avoidance. We transmit under the following * conditions when len is non-zero: * * - We have a full segment (or more with TSO) - This is the last * buffer in a write()/send() and we are either idle or running * NODELAY - we've timed out (e.g. persist timer) - we have more * then 1/2 the maximum send window's worth of data (receiver may be * limited the window size) - we need to retransmit */ if (rsm) goto send; if (len) { if (sack_rxmit) goto send; if (len >= p_maxseg) goto send; /* * NOTE! on localhost connections an 'ack' from the remote * end may occur synchronously with the output and cause us * to flush a buffer queued with moretocome. XXX * */ if (((tp->t_flags & TF_MORETOCOME) == 0) && /* normal case */ ((tp->t_flags & TF_NODELAY) || ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) && (tp->t_flags & TF_NOPUSH) == 0) { goto send; } if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ goto send; } if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { goto send; } } /* * Sending of standalone window updates. * * Window updates are important when we close our window due to a * full socket buffer and are opening it again after the application * reads data from it. Once the window has opened again and the * remote end starts to send again the ACK clock takes over and * provides the most current window information. * * We must avoid the silly window syndrome whereas every read from * the receive buffer, no matter how small, causes a window update * to be sent. We also should avoid sending a flurry of window * updates when the socket buffer had queued a lot of data and the * application is doing small reads. * * Prevent a flurry of pointless window updates by only sending an * update when we can increase the advertized window by more than * 1/4th of the socket buffer capacity. When the buffer is getting * full or is very small be more aggressive and send an update * whenever we can increase by two mss sized segments. In all other * situations the ACK's to new incoming data will carry further * window increases. * * Don't send an independent window update if a delayed ACK is * pending (it will get piggy-backed on it) or the remote side * already has done a half-close and won't send more data. Skip * this if the connection is in T/TCP half-open state. */ if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && !(tp->t_flags & TF_DELACK) && !TCPS_HAVERCVDFIN(tp->t_state)) { /* Check to see if we should do a window update */ if (bbr_window_update_needed(tp, so, recwin, maxseg)) goto send; } /* * Send if we owe the peer an ACK, RST, SYN. ACKNOW * is also a catch-all for the retransmit timer timeout case. */ if (tp->t_flags & TF_ACKNOW) { goto send; } if (flags & TH_RST) { /* Always send a RST if one is due */ goto send; } if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) { goto send; } /* * If our state indicates that FIN should be sent and we have not * yet done so, then we need to send. */ if (flags & TH_FIN && ((tp->t_flags & TF_SENTFIN) == 0)) { goto send; } /* * No reason to send a segment, just return. */ just_return: SOCKBUF_UNLOCK(sb); just_return_nolock: if (tot_len) slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); if (bbr->rc_no_pacing) slot = 0; if (tot_len == 0) { if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >= tp->snd_wnd) { BBR_STAT_INC(bbr_rwnd_limited); app_limited = BBR_JR_RWND_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); if ((bbr->rc_in_persist == 0) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* No send window.. we must enter persist */ bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); } } else if (ctf_outstanding(tp) >= sbavail(sb)) { BBR_STAT_INC(bbr_app_limited); app_limited = BBR_JR_APP_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) { BBR_STAT_INC(bbr_cwnd_limited); app_limited = BBR_JR_CWND_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes))); bbr->rc_cwnd_limited = 1; } else { BBR_STAT_INC(bbr_app_limited); app_limited = BBR_JR_APP_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); } bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; bbr->r_ctl.rc_last_delay_val = 0; } else if (bbr->rc_use_google == 0) bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); /* Are we app limited? */ if ((app_limited == BBR_JR_APP_LIMITED) || (app_limited == BBR_JR_RWND_LIMITED)) { /** * We are application limited. */ bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); } if (tot_len == 0) counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1); /* Dont update the time if we did not send */ bbr->r_ctl.rc_last_delay_val = 0; bbr->rc_output_starts_timer = 1; bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len); bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* Make sure snd_nxt is drug up */ tp->snd_nxt = tp->snd_max; } return (error); send: if (doing_tlp == 0) { /* * Data not a TLP, and its not the rxt firing. If it is the * rxt firing, we want to leave the tlp_in_progress flag on * so we don't send another TLP. It has to be a rack timer * or normal send (response to acked data) to clear the tlp * in progress flag. */ bbr->rc_tlp_in_progress = 0; bbr->rc_tlp_rtx_out = 0; } else { /* * Its a TLP. */ bbr->rc_tlp_in_progress = 1; } bbr_timer_cancel(bbr, __LINE__, cts); if (rsm == NULL) { if (sbused(sb) > 0) { /* * This is sub-optimal. We only send a stand alone * FIN on its own segment. */ if (flags & TH_FIN) { flags &= ~TH_FIN; if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) { /* Lets not send this */ slot = 0; goto just_return; } } } } else { /* * We do *not* send a FIN on a retransmit if it has data. * The if clause here where len > 1 should never come true. */ if ((len > 0) && (((rsm->r_flags & BBR_HAS_FIN) == 0) && (flags & TH_FIN))) { flags &= ~TH_FIN; len--; } } SOCKBUF_LOCK_ASSERT(sb); if (len > 0) { if ((tp->snd_una == tp->snd_max) && (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { /* * This qualifies as a RTT_PROBE session since we * drop the data outstanding to nothing and waited * more than bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } if (len >= maxseg) tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; else tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; } /* * Before ESTABLISHED, force sending of initial options unless TCP * set not to do any options. NOTE: we assume that the IP/TCP header * plus TCP options always fit in a single mbuf, leaving room for a * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) * + optlen <= MCLBYTES */ optlen = 0; #ifdef INET6 if (isipv6) hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else #endif hdrlen = sizeof(struct tcpiphdr); /* * Compute options for segment. We only have to care about SYN and * established connection segments. Options for SYN-ACK segments * are handled in TCP syncache. */ to.to_flags = 0; local_options = 0; if ((tp->t_flags & TF_NOOPT) == 0) { /* Maximum segment size. */ if (flags & TH_SYN) { to.to_mss = tcp_mssopt(&inp->inp_inc); if (tp->t_port) to.to_mss -= V_tcp_udp_tunneling_overhead; to.to_flags |= TOF_MSS; /* * On SYN or SYN|ACK transmits on TFO connections, * only include the TFO option if it is not a * retransmit, as the presence of the TFO option may * have caused the original SYN or SYN|ACK to have * been dropped by a middlebox. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_rxtshift == 0)) { if (tp->t_state == TCPS_SYN_RECEIVED) { to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; to.to_tfo_cookie = (u_int8_t *)&tp->t_tfo_cookie.server; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } else if (tp->t_state == TCPS_SYN_SENT) { to.to_tfo_len = tp->t_tfo_client_cookie_len; to.to_tfo_cookie = tp->t_tfo_cookie.client; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } } } /* Window scaling. */ if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { to.to_wscale = tp->request_r_scale; to.to_flags |= TOF_SCALE; } /* Timestamps. */ if ((tp->t_flags & TF_RCVD_TSTMP) || ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { to.to_tsval = tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; local_options += TCPOLEN_TIMESTAMP + 2; } /* Set receive buffer autosizing timestamp. */ if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) tp->rfbuf_ts = tcp_tv_to_mssectick(&bbr->rc_tv); /* Selective ACK's. */ if (flags & TH_SYN) to.to_flags |= TOF_SACKPERM; else if (TCPS_HAVEESTABLISHED(tp->t_state) && tp->rcv_numsacks > 0) { to.to_flags |= TOF_SACK; to.to_nsacks = tp->rcv_numsacks; to.to_sacks = (u_char *)tp->sackblks; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* TCP-MD5 (RFC2385). */ if (tp->t_flags & TF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif /* TCP_SIGNATURE */ /* Processing the options. */ hdrlen += (optlen = tcp_addoptions(&to, opt)); /* * If we wanted a TFO option to be added, but it was unable * to fit, ensure no data is sent. */ if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && !(to.to_flags & TOF_FASTOPEN)) len = 0; } if (tp->t_port) { if (V_tcp_udp_tunneling_port == 0) { /* The port was removed?? */ SOCKBUF_UNLOCK(&so->so_snd); return (EHOSTUNREACH); } hdrlen += sizeof(struct udphdr); } #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(tp->t_inpcb); else #endif if (tp->t_inpcb->inp_options) ipoptlen = tp->t_inpcb->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if (bbr->rc_last_options != local_options) { /* * Cache the options length this generally does not change * on a connection. We use this to calculate TSO. */ bbr->rc_last_options = local_options; } maxseg = tp->t_maxseg - (ipoptlen + optlen); p_maxseg = min(maxseg, pace_max_segs); /* * Adjust data length if insertion of options will bump the packet * length beyond the t_maxseg length. Clear the FIN bit because we * cut off the tail of the segment. */ if (len > maxseg) { if (len != 0 && (flags & TH_FIN)) { flags &= ~TH_FIN; } if (tso) { uint32_t moff; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; KASSERT(ipoptlen == 0, ("%s: TSO can't do IP options", __func__)); /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { len = 0; } else if (len > max_len) { len = max_len; } } /* * Prevent the last segment from being fractional * unless the send sockbuf can be emptied: */ if ((sb_offset + len) < sbavail(sb)) { moff = len % (uint32_t)maxseg; if (moff != 0) { len -= moff; } } /* * In case there are too many small fragments don't * use TSO: */ if (len <= maxseg) { len = maxseg; tso = 0; } } else { /* Not doing TSO */ if (optlen + ipoptlen >= tp->t_maxseg) { /* * Since we don't have enough space to put * the IP header chain and the TCP header in * one packet as required by RFC 7112, don't * send it. Also ensure that at least one * byte of the payload can be put into the * TCP segment. */ SOCKBUF_UNLOCK(&so->so_snd); error = EMSGSIZE; sack_rxmit = 0; goto out; } len = maxseg; } } else { /* Not doing TSO */ if_hw_tsomaxsegcount = 0; tso = 0; } KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, ("%s: len > IP_MAXPACKET", __func__)); #ifdef DIAGNOSTIC #ifdef INET6 if (max_linkhdr + hdrlen > MCLBYTES) #else if (max_linkhdr + hdrlen > MHLEN) #endif panic("tcphdr too big"); #endif /* * This KASSERT is here to catch edge cases at a well defined place. * Before, those had triggered (random) panic conditions further * down. */ #ifdef BBR_INVARIANTS if (sack_rxmit) { if (SEQ_LT(rsm->r_start, tp->snd_una)) { panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u", rsm, tp, bbr, rsm->r_start, tp->snd_una); } } #endif KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); if ((len == 0) && (flags & TH_FIN) && (sbused(sb))) { /* * We have outstanding data, don't send a fin by itself!. */ slot = 0; goto just_return; } /* * Grab a header mbuf, attaching a copy of data to be transmitted, * and initialize the header from the template for sends on this * connection. */ if (len) { uint32_t moff; uint32_t orig_len; /* * We place a limit on sending with hptsi. */ if ((rsm == NULL) && len > pace_max_segs) len = pace_max_segs; if (len <= maxseg) tso = 0; #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { BBR_STAT_INC(bbr_failed_mbuf_aloc); bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); SOCKBUF_UNLOCK(sb); error = ENOBUFS; sack_rxmit = 0; goto out; } m->m_data += max_linkhdr; m->m_len = hdrlen; /* * Start the m_copy functions from the closest mbuf to the * sb_offset in the socket buffer chain. */ if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) { #ifdef BBR_INVARIANTS if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u", tp, bbr, len, sb_offset, sbavail(sb), rsm, doing_retran_from, picked_up_retran, doing_tlp); #endif /* * In this messed up situation we have two choices, * a) pretend the send worked, and just start timers * and what not (not good since that may lead us * back here a lot). b) Send the lowest segment * in the map. c) Drop the connection. Lets do * which if it continues to happen will lead to * via timeouts. */ BBR_STAT_INC(bbr_offset_recovery); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); sb_offset = 0; if (rsm == NULL) { sack_rxmit = 0; len = sbavail(sb); } else { sack_rxmit = 1; if (rsm->r_start != tp->snd_una) { /* * Things are really messed up, * is the only thing to do. */ BBR_STAT_INC(bbr_offset_drop); tcp_set_inp_to_drop(inp, EFAULT); SOCKBUF_UNLOCK(sb); (void)m_free(m); return (0); } len = rsm->r_end - rsm->r_start; } if (len > sbavail(sb)) len = sbavail(sb); if (len > maxseg) len = maxseg; } mb = sbsndptr_noadv(sb, sb_offset, &moff); if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { m_copydata(mb, moff, (int)len, mtod(m, caddr_t)+hdrlen); if (rsm == NULL) sbsndptr_adv(sb, mb, len); m->m_len += len; } else { struct sockbuf *msb; if (rsm) msb = NULL; else msb = sb; #ifdef BBR_INVARIANTS if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) { if (rsm) { panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u rsm:%p snd_una:%u rsm_start:%u flg:%x %u:%u:%u sr:%d ", tp, bbr, len, moff, sbavail(sb), rsm, tp->snd_una, rsm->r_flags, rsm->r_start, doing_retran_from, picked_up_retran, doing_tlp, sack_rxmit); } else { panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u", tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una); } } #endif orig_len = len; m->m_next = tcp_m_copym( mb, moff, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb, ((rsm == NULL) ? hw_tls : 0) #ifdef NETFLIX_COPY_ARGS , &filled_all #endif ); if (len <= maxseg) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if (m->m_next == NULL) { SOCKBUF_UNLOCK(sb); (void)m_free(m); error = ENOBUFS; sack_rxmit = 0; goto out; } } #ifdef BBR_INVARIANTS if (tso && len < maxseg) { panic("tp:%p tso on, but len:%d < maxseg:%d", tp, len, maxseg); } if (tso && if_hw_tsomaxsegcount) { int32_t seg_cnt = 0; struct mbuf *foo; foo = m; while (foo) { seg_cnt++; foo = foo->m_next; } if (seg_cnt > if_hw_tsomaxsegcount) { panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount); } } #endif /* * If we're sending everything we've got, set PUSH. (This * will keep happy those implementations which only give * data to the user when a buffer fills or a PUSH comes in.) */ if (sb_offset + len == sbused(sb) && sbused(sb) && !(flags & TH_SYN)) { flags |= TH_PUSH; } SOCKBUF_UNLOCK(sb); } else { SOCKBUF_UNLOCK(sb); if (tp->t_flags & TF_ACKNOW) KMOD_TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN | TH_FIN | TH_RST)) KMOD_TCPSTAT_INC(tcps_sndctrl); else KMOD_TCPSTAT_INC(tcps_sndwinup); m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { BBR_STAT_INC(bbr_failed_mbuf_aloc); bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); error = ENOBUFS; /* Fudge the send time since we could not send */ sack_rxmit = 0; goto out; } #ifdef INET6 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && MHLEN >= hdrlen) { M_ALIGN(m, hdrlen); } else #endif m->m_data += max_linkhdr; m->m_len = hdrlen; } SOCKBUF_UNLOCK_ASSERT(sb); m->m_pkthdr.rcvif = (struct ifnet *)0; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip6_hdr); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else { th = (struct tcphdr *)(ip6 + 1); } tcpip_fillheaders(inp, tp->t_port, ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); #ifdef TCPDEBUG ipov = (struct ipovly *)ip; #endif if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else { th = (struct tcphdr *)(ip + 1); } tcpip_fillheaders(inp, tp->t_port, ip, th); } /* * If we are doing retransmissions, then snd_nxt will not reflect * the first unsent octet. For ACK only packets, we do not want the * sequence number of the retransmitted packet, we want the sequence * number of the next unsent octet. So, if there is no data (and no * SYN or FIN), use snd_max instead of snd_nxt when filling in * ti_seq. But if we are in persist state, snd_max might reflect * one byte beyond the right edge of the window, so use snd_nxt in * that case, since we know we aren't doing a retransmission. * (retransmit and persist are mutually exclusive...) */ if (sack_rxmit == 0) { if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) { /* New data (including new persists) */ th->th_seq = htonl(tp->snd_max); bbr_seq = tp->snd_max; } else if (flags & TH_SYN) { /* Syn's always send from iss */ th->th_seq = htonl(tp->iss); bbr_seq = tp->iss; } else if (flags & TH_FIN) { if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) { /* * If we sent the fin already its 1 minus * snd_max */ th->th_seq = (htonl(tp->snd_max - 1)); bbr_seq = (tp->snd_max - 1); } else { /* First time FIN use snd_max */ th->th_seq = htonl(tp->snd_max); bbr_seq = tp->snd_max; } } else { /* * len == 0 and not persist we use snd_max, sending * an ack unless we have sent the fin then its 1 * minus. */ /* * XXXRRS Question if we are in persists and we have * nothing outstanding to send and we have not sent * a FIN, we will send an ACK. In such a case it * might be better to send (tp->snd_una - 1) which * would force the peer to ack. */ if (tp->t_flags & TF_SENTFIN) { th->th_seq = htonl(tp->snd_max - 1); bbr_seq = (tp->snd_max - 1); } else { th->th_seq = htonl(tp->snd_max); bbr_seq = tp->snd_max; } } } else { /* All retransmits use the rsm to guide the send */ th->th_seq = htonl(rsm->r_start); bbr_seq = rsm->r_start; } th->th_ack = htonl(tp->rcv_nxt); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } th->th_flags = flags; /* * Calculate receive window. Don't shrink window, but avoid silly * window syndrome. */ if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) && recwin < maxseg))) recwin = 0; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && recwin < (tp->rcv_adv - tp->rcv_nxt)) recwin = (tp->rcv_adv - tp->rcv_nxt); if (recwin > TCP_MAXWIN << tp->rcv_scale) recwin = TCP_MAXWIN << tp->rcv_scale; /* * According to RFC1323 the window field in a SYN (i.e., a or * ) segment itself is never scaled. The case is * handled in syncache. */ if (flags & TH_SYN) th->th_win = htons((u_short) (min(sbspace(&so->so_rcv), TCP_MAXWIN))); else { /* Avoid shrinking window with window scaling. */ recwin = roundup2(recwin, 1 << tp->rcv_scale); th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); } /* * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 * window. This may cause the remote transmitter to stall. This * flag tells soreceive() to disable delayed acknowledgements when * draining the buffer. This can occur if the receiver is * attempting to read more data than can be buffered prior to * transmitting on the connection. */ if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; /* * We don't support urgent data, but drag along * the pointer in case of a stack switch. */ tp->snd_up = tp->snd_una; #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (to.to_flags & TOF_SIGNATURE) { /* * Calculate MD5 signature and put it into the place * determined before. NOTE: since TCP options buffer doesn't * point into mbuf's data, calculate offset and use it. */ if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { /* * Do not send segment if the calculation of MD5 * digest has failed. */ goto out; } } #endif /* * Put TCP length in extended header, and then checksum extended * header and data. */ m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (isipv6) { /* * ip6_plen is not need to be filled now, and will be filled * in ip6_output. */ if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif /* * Enable TSO and specify the size of the segments. The TCP pseudo * header checksum is always provided. XXX: Fixme: This is currently * not the case for IPv6. */ if (tso) { KASSERT(len > maxseg, ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg)); m->m_pkthdr.csum_flags |= CSUM_TSO; csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = maxseg; } KASSERT(len + hdrlen == m_length(m, NULL), ("%s: mbuf chain different than expected: %d + %u != %u", __func__, len, hdrlen, m_length(m, NULL))); #ifdef TCP_HHOOK /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */ hhook_run_tcp_est_out(tp, th, &to, len, tso); #endif #ifdef TCPDEBUG /* * Trace. */ if (so->so_options & SO_DEBUG) { u_short save = 0; #ifdef INET6 if (!isipv6) #endif { save = ipov->ih_len; ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + * (th->th_off << 2) */ ); } tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); #ifdef INET6 if (!isipv6) #endif ipov->ih_len = save; } #endif /* TCPDEBUG */ /* Log to the black box */ if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); /* Record info on type of transmission */ log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay; log.u_bbr.flex2 = (bbr->r_recovery_bw << 3); log.u_bbr.flex3 = maxseg; log.u_bbr.flex4 = delay_calc; /* Encode filled_all into the upper flex5 bit */ log.u_bbr.flex5 = bbr->rc_past_init_win; log.u_bbr.flex5 <<= 1; log.u_bbr.flex5 |= bbr->rc_no_pacing; log.u_bbr.flex5 <<= 29; if (filled_all) log.u_bbr.flex5 |= 0x80000000; log.u_bbr.flex5 |= tp->t_maxseg; log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs; log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr); /* lets poke in the low and the high here for debugging */ log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; if (rsm || sack_rxmit) { if (doing_tlp) log.u_bbr.flex8 = 2; else log.u_bbr.flex8 = 1; } else { log.u_bbr.flex8 = 0; } lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, tv); } else { lgb = NULL; } /* * Fill in IP length and desired time to live and send to IP level. * There should be a better way to handle ttl and tos; we could keep * them in the template, but need a way to checksum without them. */ /* * m->m_pkthdr.len should have been set before cksum calcuration, * because in6_cksum() need it. */ #ifdef INET6 if (isipv6) { /* * we separately set hoplimit for every segment, since the * user might want to change the value via setsockopt. Also, * desired default hop limit might be changed via Neighbor * Discovery. */ ip6->ip6_hlim = in6_selecthlim(inp, NULL); /* * Set the packet size here for the benefit of DTrace * probes. ip6_output() will set it properly; it's supposed * to include the option header lengths as well. */ ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); TCP_PROBE5(send, NULL, tp, ip6, tp, th); /* TODO: IPv6 IP6TOS_ECT bit on */ error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), NULL, NULL, inp); if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) mtu = inp->inp_route6.ro_nh->nh_mtu; } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); #ifdef INET6 if (isipv6) ip->ip_ttl = in6_selecthlim(inp, NULL); #endif /* INET6 */ /* * If we do path MTU discovery, then we set DF on every * packet. This might not be the best thing to do according * to RFC3390 Section 2. However the tcp hostcache migitates * the problem so it affects only the first tcp connection * with a host. * * NB: Don't set DF on small MTU/MSS to have a safe * fallback. */ if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); TCP_PROBE5(send, NULL, tp, ip, tp, th); error = ip_output(m, inp->inp_options, &inp->inp_route, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, inp); if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) mtu = inp->inp_route.ro_nh->nh_mtu; } #endif /* INET */ out: if (lgb) { lgb->tlb_errno = error; lgb = NULL; } /* * In transmit state, time the transmission and arrange for the * retransmit. In persist state, just set snd_max. */ if (error == 0) { tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0) tcp_clean_dsack_blocks(tp); /* We sent an ack clear the bbr_segs_rcvd count */ bbr->output_error_seen = 0; bbr->oerror_cnt = 0; bbr->bbr_segs_rcvd = 0; if (len == 0) counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1); /* Do accounting for new sends */ if ((len > 0) && (rsm == NULL)) { int idx; if (tp->snd_una == tp->snd_max) { /* * Special case to match google, when * nothing is in flight the delivered * time does get updated to the current * time (see tcp_rate_bsd.c). */ bbr->r_ctl.rc_del_time = cts; } if (len >= maxseg) { idx = (len / maxseg) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1); else counter_u64_add(bbr_out_size[idx], 1); } else { /* smaller than a MSS */ idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options); if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV) idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1); counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1); } } } abandon = 0; /* * We must do the send accounting before we log the output, * otherwise the state of the rsm could change and we account to the * wrong bucket. */ if (len > 0) { bbr_do_send_accounting(tp, bbr, rsm, len, error); if (error == 0) { if (tp->snd_una == tp->snd_max) bbr->r_ctl.rc_tlp_rxt_last_time = cts; } } bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error, cts, mb, &abandon, rsm, 0, sb); if (abandon) { /* * If bbr_log_output destroys the TCB or sees a TH_RST being * sent we should hit this condition. */ return (0); } if (bbr->rc_in_persist == 0) { /* * Advance snd_nxt over sequence space of this segment. */ if (error) /* We don't log or do anything with errors */ goto skip_upd; if (tp->snd_una == tp->snd_max && (len || (flags & (TH_SYN | TH_FIN)))) { /* * Update the time we just added data since none was * outstanding. */ bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); bbr->rc_tp->t_acktime = ticks; } if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) { if (flags & TH_SYN) { /* * Smack the snd_max to iss + 1 * if its a FO we will add len below. */ tp->snd_max = tp->iss + 1; } if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { tp->snd_max++; tp->t_flags |= TF_SENTFIN; } } if (sack_rxmit == 0) tp->snd_max += len; skip_upd: if ((error == 0) && len) tot_len += len; } else { /* Persists case */ int32_t xlen = len; if (error) goto nomore; if (flags & TH_SYN) ++xlen; if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { ++xlen; tp->t_flags |= TF_SENTFIN; } if (xlen && (tp->snd_una == tp->snd_max)) { /* * Update the time we just added data since none was * outstanding. */ bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); bbr->rc_tp->t_acktime = ticks; } if (sack_rxmit == 0) tp->snd_max += xlen; tot_len += (len + optlen + ipoptlen); } nomore: if (error) { /* * Failures do not advance the seq counter above. For the * case of ENOBUFS we will fall out and become ack-clocked. * capping the cwnd at the current flight. * Everything else will just have to retransmit with the timer * (no pacer). */ SOCKBUF_UNLOCK_ASSERT(sb); BBR_STAT_INC(bbr_saw_oerr); /* Clear all delay/early tracks */ bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_ctl.rc_agg_early = 0; bbr->r_agg_early_set = 0; bbr->output_error_seen = 1; if (bbr->oerror_cnt < 0xf) bbr->oerror_cnt++; if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) { /* drop the session */ tcp_set_inp_to_drop(inp, ENETDOWN); } switch (error) { case ENOBUFS: /* * Make this guy have to get ack's to send * more but lets make sure we don't * slam him below a T-O (1MSS). */ if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) - maxseg; if (tp->snd_cwnd < maxseg) tp->snd_cwnd = maxseg; } slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt; BBR_STAT_INC(bbr_saw_enobuf); if (bbr->bbr_hdrw_pacing) counter_u64_add(bbr_hdwr_pacing_enobuf, 1); else counter_u64_add(bbr_nohdwr_pacing_enobuf, 1); /* * Here even in the enobuf's case we want to do our * state update. The reason being we may have been * called by the input function. If so we have had * things change. */ error = 0; goto enobufs; case EMSGSIZE: /* * For some reason the interface we used initially * to send segments changed to another or lowered * its MTU. If TSO was active we either got an * interface without TSO capabilits or TSO was * turned off. If we obtained mtu from ip_output() * then update it and try again. */ /* Turn on tracing (or try to) */ { int old_maxseg; old_maxseg = tp->t_maxseg; BBR_STAT_INC(bbr_saw_emsgsiz); bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts); if (mtu != 0) tcp_mss_update(tp, -1, mtu, NULL, NULL); if (old_maxseg <= tp->t_maxseg) { /* Huh it did not shrink? */ tp->t_maxseg = old_maxseg - 40; bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts); } /* * Nuke all other things that can interfere * with slot */ if ((tot_len + len) && (len >= tp->t_maxseg)) { slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, (tot_len + len), cts, 0); if (slot < bbr_error_base_paceout) slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; } else slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; bbr->rc_output_starts_timer = 1; bbr_start_hpts_timer(bbr, tp, cts, 10, slot, tot_len); return (error); } case EPERM: tp->t_softerror = error; /* Fall through */ case EHOSTDOWN: case EHOSTUNREACH: case ENETDOWN: case ENETUNREACH: if (TCPS_HAVERCVDSYN(tp->t_state)) { tp->t_softerror = error; } /* FALLTHROUGH */ default: slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt; bbr->rc_output_starts_timer = 1; bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0); return (error); } #ifdef STATS } else if (((tp->t_flags & TF_GPUTINPROG) == 0) && len && (rsm == NULL) && (bbr->rc_in_persist == 0)) { tp->gput_seq = bbr_seq; tp->gput_ack = bbr_seq + min(sbavail(&so->so_snd) - sb_offset, sendwin); tp->gput_ts = cts; tp->t_flags |= TF_GPUTINPROG; #endif } KMOD_TCPSTAT_INC(tcps_sndtotal); if ((bbr->bbr_hdw_pace_ena) && (bbr->bbr_attempt_hdwr_pace == 0) && (bbr->rc_past_init_win) && (bbr->rc_bbr_state != BBR_STATE_STARTUP) && (get_filter_value(&bbr->r_ctl.rc_delrate)) && (inp->inp_route.ro_nh && inp->inp_route.ro_nh->nh_ifp)) { /* * We are past the initial window and * have at least one measurement so we * could use hardware pacing if its available. * We have an interface and we have not attempted * to setup hardware pacing, lets try to now. */ uint64_t rate_wanted; int err = 0; rate_wanted = bbr_get_hardware_rate(bbr); bbr->bbr_attempt_hdwr_pace = 1; bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp, inp->inp_route.ro_nh->nh_ifp, rate_wanted, (RS_PACING_GEQ|RS_PACING_SUB_OK), &err, NULL); if (bbr->r_ctl.crte) { bbr_type_log_hdwr_pacing(bbr, bbr->r_ctl.crte->ptbl->rs_ifp, rate_wanted, bbr->r_ctl.crte->rate, __LINE__, cts, err); BBR_STAT_INC(bbr_hdwr_rl_add_ok); counter_u64_add(bbr_flows_nohdwr_pacing, -1); counter_u64_add(bbr_flows_whdwr_pacing, 1); bbr->bbr_hdrw_pacing = 1; /* Now what is our gain status? */ if (bbr->r_ctl.crte->rate < rate_wanted) { /* We have a problem */ bbr_setup_less_of_rate(bbr, cts, bbr->r_ctl.crte->rate, rate_wanted); } else { /* We are good */ bbr->gain_is_limited = 0; bbr->skip_gain = 0; } tcp_bbr_tso_size_check(bbr, cts); } else { bbr_type_log_hdwr_pacing(bbr, inp->inp_route.ro_nh->nh_ifp, rate_wanted, 0, __LINE__, cts, err); BBR_STAT_INC(bbr_hdwr_rl_add_fail); } } if (bbr->bbr_hdrw_pacing) { /* * Worry about cases where the route * changes or something happened that we * lost our hardware pacing possibly during * the last ip_output call. */ if (inp->inp_snd_tag == NULL) { /* A change during ip output disabled hw pacing? */ bbr->bbr_hdrw_pacing = 0; } else if ((inp->inp_route.ro_nh == NULL) || (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) { /* * We had an interface or route change, * detach from the current hdwr pacing * and setup to re-attempt next go * round. */ bbr->bbr_hdrw_pacing = 0; bbr->bbr_attempt_hdwr_pace = 0; tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); tcp_bbr_tso_size_check(bbr, cts); } } /* * Data sent (as far as we can tell). If this advertises a larger * window than any other segment, then remember the size of the * advertised window. Any pending ACK has now been sent. */ if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) tp->rcv_adv = tp->rcv_nxt + recwin; tp->last_ack_sent = tp->rcv_nxt; if ((error == 0) && (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) && (doing_tlp == 0) && (tso == 0) && (len > 0) && ((flags & TH_RST) == 0) && ((flags & TH_SYN) == 0) && (IN_RECOVERY(tp->t_flags) == 0) && (bbr->rc_in_persist == 0) && (tot_len < bbr->r_ctl.rc_pace_max_segs)) { /* * For non-tso we need to goto again until we have sent out * enough data to match what we are hptsi out every hptsi * interval. */ if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* Make sure snd_nxt is drug up */ tp->snd_nxt = tp->snd_max; } if (rsm != NULL) { rsm = NULL; goto skip_again; } rsm = NULL; sack_rxmit = 0; tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); goto again; } skip_again: if ((error == 0) && (flags & TH_FIN)) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); if ((error == 0) && (flags & TH_RST)) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) { /* * Calculate/Re-Calculate the hptsi slot in usecs based on * what we have sent so far */ slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); if (bbr->rc_no_pacing) slot = 0; } tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); enobufs: if (bbr->rc_use_google == 0) bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes))); bbr->rc_output_starts_timer = 1; if (bbr->bbr_use_rack_cheat && (more_to_rxt || ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) { /* Rack cheats and shotguns out all rxt's 1ms apart */ if (slot > 1000) slot = 1000; } if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) { /* * We don't change the tso size until some number of sends * to give the hardware commands time to get down * to the interface. */ bbr->r_ctl.bbr_hdwr_cnt_noset_snt++; if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) { bbr->hw_pacing_set = 1; tcp_bbr_tso_size_check(bbr, cts); } } bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* Make sure snd_nxt is drug up */ tp->snd_nxt = tp->snd_max; } return (error); } /* * See bbr_output_wtime() for return values. */ static int bbr_output(struct tcpcb *tp) { int32_t ret; struct timeval tv; struct tcp_bbr *bbr; NET_EPOCH_ASSERT(); bbr = (struct tcp_bbr *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); (void)tcp_get_usecs(&tv); ret = bbr_output_wtime(tp, &tv); return (ret); } static void bbr_mtu_chg(struct tcpcb *tp) { struct tcp_bbr *bbr; struct bbr_sendmap *rsm, *frsm = NULL; uint32_t maxseg; /* * The MTU has changed. a) Clear the sack filter. b) Mark everything * over the current size as SACK_PASS so a retransmit will occur. */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; maxseg = tp->t_maxseg - bbr->rc_last_options; sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { /* Don't mess with ones acked (by sack?) */ if (rsm->r_flags & BBR_ACKED) continue; if ((rsm->r_end - rsm->r_start) > maxseg) { /* * We mark sack-passed on all the previous large * sends we did. This will force them to retransmit. */ rsm->r_flags |= BBR_SACK_PASSED; if (((rsm->r_flags & BBR_MARKED_LOST) == 0) && bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) { bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; rsm->r_flags |= BBR_MARKED_LOST; } if (frsm == NULL) frsm = rsm; } } if (frsm) { bbr->r_ctl.rc_resend = frsm; } } static int bbr_pru_options(struct tcpcb *tp, int flags) { if (flags & PRUS_OOB) return (EOPNOTSUPP); return (0); } struct tcp_function_block __tcp_bbr = { .tfb_tcp_block_name = __XSTRING(STACKNAME), .tfb_tcp_output = bbr_output, .tfb_do_queued_segments = ctf_do_queued_segments, .tfb_do_segment_nounlock = bbr_do_segment_nounlock, .tfb_tcp_do_segment = bbr_do_segment, .tfb_tcp_ctloutput = bbr_ctloutput, .tfb_tcp_fb_init = bbr_init, .tfb_tcp_fb_fini = bbr_fini, .tfb_tcp_timer_stop_all = bbr_stopall, .tfb_tcp_timer_activate = bbr_timer_activate, .tfb_tcp_timer_active = bbr_timer_active, .tfb_tcp_timer_stop = bbr_timer_stop, .tfb_tcp_rexmit_tmr = bbr_remxt_tmr, .tfb_tcp_handoff_ok = bbr_handoff_ok, .tfb_tcp_mtu_chg = bbr_mtu_chg, .tfb_pru_options = bbr_pru_options, }; /* * bbr_ctloutput() must drop the inpcb lock before performing copyin on * socket option arguments. When it re-acquires the lock after the copy, it * has to revalidate that the connection is still valid for the socket * option. */ static int bbr_set_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr) { struct epoch_tracker et; int32_t error = 0, optval; switch (sopt->sopt_name) { case TCP_RACK_PACE_MAX_SEG: case TCP_RACK_MIN_TO: case TCP_RACK_REORD_THRESH: case TCP_RACK_REORD_FADE: case TCP_RACK_TLP_THRESH: case TCP_RACK_PKT_DELAY: case TCP_BBR_ALGORITHM: case TCP_BBR_TSLIMITS: case TCP_BBR_IWINTSO: case TCP_BBR_RECFORCE: case TCP_BBR_STARTUP_PG: case TCP_BBR_DRAIN_PG: case TCP_BBR_RWND_IS_APP: case TCP_BBR_PROBE_RTT_INT: case TCP_BBR_PROBE_RTT_GAIN: case TCP_BBR_PROBE_RTT_LEN: case TCP_BBR_STARTUP_LOSS_EXIT: case TCP_BBR_USEDEL_RATE: case TCP_BBR_MIN_RTO: case TCP_BBR_MAX_RTO: case TCP_BBR_PACE_PER_SEC: case TCP_DELACK: case TCP_BBR_PACE_DEL_TAR: case TCP_BBR_SEND_IWND_IN_TSO: case TCP_BBR_EXTRA_STATE: case TCP_BBR_UTTER_MAX_TSO: case TCP_BBR_MIN_TOPACEOUT: case TCP_BBR_FLOOR_MIN_TSO: case TCP_BBR_TSTMP_RAISES: case TCP_BBR_POLICER_DETECT: case TCP_BBR_USE_RACK_CHEAT: case TCP_DATA_AFTER_CLOSE: case TCP_BBR_HDWR_PACE: case TCP_BBR_PACE_SEG_MAX: case TCP_BBR_PACE_SEG_MIN: case TCP_BBR_PACE_CROSS: case TCP_BBR_PACE_OH: #ifdef NETFLIX_PEAKRATE case TCP_MAXPEAKRATE: #endif case TCP_BBR_TMR_PACE_OH: case TCP_BBR_RACK_RTT_USE: case TCP_BBR_RETRAN_WTSO: break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); if (error) return (error); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); if (tp->t_fb != &__tcp_bbr) { INP_WUNLOCK(inp); return (ENOPROTOOPT); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; switch (sopt->sopt_name) { case TCP_BBR_PACE_PER_SEC: BBR_OPTS_INC(tcp_bbr_pace_per_sec); bbr->r_ctl.bbr_hptsi_per_second = optval; break; case TCP_BBR_PACE_DEL_TAR: BBR_OPTS_INC(tcp_bbr_pace_del_tar); bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval; break; case TCP_BBR_PACE_SEG_MAX: BBR_OPTS_INC(tcp_bbr_pace_seg_max); bbr->r_ctl.bbr_hptsi_segments_max = optval; break; case TCP_BBR_PACE_SEG_MIN: BBR_OPTS_INC(tcp_bbr_pace_seg_min); bbr->r_ctl.bbr_hptsi_bytes_min = optval; break; case TCP_BBR_PACE_CROSS: BBR_OPTS_INC(tcp_bbr_pace_cross); bbr->r_ctl.bbr_cross_over = optval; break; case TCP_BBR_ALGORITHM: BBR_OPTS_INC(tcp_bbr_algorithm); if (optval && (bbr->rc_use_google == 0)) { /* Turn on the google mode */ bbr_google_mode_on(bbr); if ((optval > 3) && (optval < 500)) { /* * Must be at least greater than .3% * and must be less than 50.0%. */ bbr->r_ctl.bbr_google_discount = optval; } } else if ((optval == 0) && (bbr->rc_use_google == 1)) { /* Turn off the google mode */ bbr_google_mode_off(bbr); } break; case TCP_BBR_TSLIMITS: BBR_OPTS_INC(tcp_bbr_tslimits); if (optval == 1) bbr->rc_use_ts_limit = 1; else if (optval == 0) bbr->rc_use_ts_limit = 0; else error = EINVAL; break; case TCP_BBR_IWINTSO: BBR_OPTS_INC(tcp_bbr_iwintso); if ((optval >= 0) && (optval < 128)) { uint32_t twin; bbr->rc_init_win = optval; twin = bbr_initial_cwnd(bbr, tp); if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd)) tp->snd_cwnd = twin; else error = EBUSY; } else error = EINVAL; break; case TCP_BBR_STARTUP_PG: BBR_OPTS_INC(tcp_bbr_startup_pg); if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) { bbr->r_ctl.rc_startup_pg = optval; if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { bbr->r_ctl.rc_bbr_hptsi_gain = optval; } } else error = EINVAL; break; case TCP_BBR_DRAIN_PG: BBR_OPTS_INC(tcp_bbr_drain_pg); if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) bbr->r_ctl.rc_drain_pg = optval; else error = EINVAL; break; case TCP_BBR_PROBE_RTT_LEN: BBR_OPTS_INC(tcp_bbr_probertt_len); if (optval <= 1) reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND)); else error = EINVAL; break; case TCP_BBR_PROBE_RTT_GAIN: BBR_OPTS_INC(tcp_bbr_probertt_gain); if (optval <= BBR_UNIT) bbr->r_ctl.bbr_rttprobe_gain_val = optval; else error = EINVAL; break; case TCP_BBR_PROBE_RTT_INT: BBR_OPTS_INC(tcp_bbr_probe_rtt_int); if (optval > 1000) bbr->r_ctl.rc_probertt_int = optval; else error = EINVAL; break; case TCP_BBR_MIN_TOPACEOUT: BBR_OPTS_INC(tcp_bbr_topaceout); if (optval == 0) { bbr->no_pacing_until = 0; bbr->rc_no_pacing = 0; } else if (optval <= 0x00ff) { bbr->no_pacing_until = optval; if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) && (bbr->rc_bbr_state == BBR_STATE_STARTUP)){ /* Turn on no pacing */ bbr->rc_no_pacing = 1; } } else error = EINVAL; break; case TCP_BBR_STARTUP_LOSS_EXIT: BBR_OPTS_INC(tcp_bbr_startup_loss_exit); bbr->rc_loss_exit = optval; break; case TCP_BBR_USEDEL_RATE: error = EINVAL; break; case TCP_BBR_MIN_RTO: BBR_OPTS_INC(tcp_bbr_min_rto); bbr->r_ctl.rc_min_rto_ms = optval; break; case TCP_BBR_MAX_RTO: BBR_OPTS_INC(tcp_bbr_max_rto); bbr->rc_max_rto_sec = optval; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ BBR_OPTS_INC(tcp_rack_min_to); bbr->r_ctl.rc_min_to = optval; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ BBR_OPTS_INC(tcp_rack_reord_thresh); if ((optval > 0) && (optval < 31)) bbr->r_ctl.rc_reorder_shift = optval; else error = EINVAL; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ BBR_OPTS_INC(tcp_rack_reord_fade); bbr->r_ctl.rc_reorder_fade = optval; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ BBR_OPTS_INC(tcp_rack_tlp_thresh); if (optval) bbr->rc_tlp_threshold = optval; else error = EINVAL; break; case TCP_BBR_USE_RACK_CHEAT: BBR_OPTS_INC(tcp_use_rackcheat); if (bbr->rc_use_google) { error = EINVAL; break; } BBR_OPTS_INC(tcp_rack_cheat); if (optval) bbr->bbr_use_rack_cheat = 1; else bbr->bbr_use_rack_cheat = 0; break; case TCP_BBR_FLOOR_MIN_TSO: BBR_OPTS_INC(tcp_utter_max_tso); if ((optval >= 0) && (optval < 40)) bbr->r_ctl.bbr_hptsi_segments_floor = optval; else error = EINVAL; break; case TCP_BBR_UTTER_MAX_TSO: BBR_OPTS_INC(tcp_utter_max_tso); if ((optval >= 0) && (optval < 0xffff)) bbr->r_ctl.bbr_utter_max = optval; else error = EINVAL; break; case TCP_BBR_EXTRA_STATE: BBR_OPTS_INC(tcp_extra_state); if (optval) bbr->rc_use_idle_restart = 1; else bbr->rc_use_idle_restart = 0; break; case TCP_BBR_SEND_IWND_IN_TSO: BBR_OPTS_INC(tcp_iwnd_tso); if (optval) { bbr->bbr_init_win_cheat = 1; if (bbr->rc_past_init_win == 0) { uint32_t cts; cts = tcp_get_usecs(&bbr->rc_tv); tcp_bbr_tso_size_check(bbr, cts); } } else bbr->bbr_init_win_cheat = 0; break; case TCP_BBR_HDWR_PACE: BBR_OPTS_INC(tcp_hdwr_pacing); if (optval){ bbr->bbr_hdw_pace_ena = 1; bbr->bbr_attempt_hdwr_pace = 0; } else { bbr->bbr_hdw_pace_ena = 0; #ifdef RATELIMIT if (bbr->r_ctl.crte != NULL) { tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); bbr->r_ctl.crte = NULL; } #endif } break; case TCP_DELACK: BBR_OPTS_INC(tcp_delack); if (optval < 100) { if (optval == 0) /* off */ tp->t_delayed_ack = 0; else if (optval == 1) /* on which is 2 */ tp->t_delayed_ack = 2; else /* higher than 2 and less than 100 */ tp->t_delayed_ack = optval; if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; NET_EPOCH_ENTER(et); bbr_output(tp); NET_EPOCH_EXIT(et); } } else error = EINVAL; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ BBR_OPTS_INC(tcp_rack_pkt_delay); bbr->r_ctl.rc_pkt_delay = optval; break; #ifdef NETFLIX_PEAKRATE case TCP_MAXPEAKRATE: BBR_OPTS_INC(tcp_maxpeak); error = tcp_set_maxpeakrate(tp, optval); if (!error) tp->t_peakrate_thr = tp->t_maxpeakrate; break; #endif case TCP_BBR_RETRAN_WTSO: BBR_OPTS_INC(tcp_retran_wtso); if (optval) bbr->rc_resends_use_tso = 1; else bbr->rc_resends_use_tso = 0; break; case TCP_DATA_AFTER_CLOSE: BBR_OPTS_INC(tcp_data_ac); if (optval) bbr->rc_allow_data_af_clo = 1; else bbr->rc_allow_data_af_clo = 0; break; case TCP_BBR_POLICER_DETECT: BBR_OPTS_INC(tcp_policer_det); if (bbr->rc_use_google == 0) error = EINVAL; else if (optval) bbr->r_use_policer = 1; else bbr->r_use_policer = 0; break; case TCP_BBR_TSTMP_RAISES: BBR_OPTS_INC(tcp_ts_raises); if (optval) bbr->ts_can_raise = 1; else bbr->ts_can_raise = 0; break; case TCP_BBR_TMR_PACE_OH: BBR_OPTS_INC(tcp_pacing_oh_tmr); if (bbr->rc_use_google) { error = EINVAL; } else { if (optval) bbr->r_ctl.rc_incr_tmrs = 1; else bbr->r_ctl.rc_incr_tmrs = 0; } break; case TCP_BBR_PACE_OH: BBR_OPTS_INC(tcp_pacing_oh); if (bbr->rc_use_google) { error = EINVAL; } else { if (optval > (BBR_INCL_TCP_OH| BBR_INCL_IP_OH| BBR_INCL_ENET_OH)) { error = EINVAL; break; } if (optval & BBR_INCL_TCP_OH) bbr->r_ctl.rc_inc_tcp_oh = 1; else bbr->r_ctl.rc_inc_tcp_oh = 0; if (optval & BBR_INCL_IP_OH) bbr->r_ctl.rc_inc_ip_oh = 1; else bbr->r_ctl.rc_inc_ip_oh = 0; if (optval & BBR_INCL_ENET_OH) bbr->r_ctl.rc_inc_enet_oh = 1; else bbr->r_ctl.rc_inc_enet_oh = 0; } break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } #ifdef NETFLIX_STATS tcp_log_socket_option(tp, sopt->sopt_name, optval, error); #endif INP_WUNLOCK(inp); return (error); } /* * return 0 on success, error-num on failure */ static int bbr_get_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr) { int32_t error, optval; /* * Because all our options are either boolean or an int, we can just * pull everything into optval and then unlock and copy. If we ever * add a option that is not a int, then this will have quite an * impact to this routine. */ switch (sopt->sopt_name) { case TCP_BBR_PACE_PER_SEC: optval = bbr->r_ctl.bbr_hptsi_per_second; break; case TCP_BBR_PACE_DEL_TAR: optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar; break; case TCP_BBR_PACE_SEG_MAX: optval = bbr->r_ctl.bbr_hptsi_segments_max; break; case TCP_BBR_MIN_TOPACEOUT: optval = bbr->no_pacing_until; break; case TCP_BBR_PACE_SEG_MIN: optval = bbr->r_ctl.bbr_hptsi_bytes_min; break; case TCP_BBR_PACE_CROSS: optval = bbr->r_ctl.bbr_cross_over; break; case TCP_BBR_ALGORITHM: optval = bbr->rc_use_google; break; case TCP_BBR_TSLIMITS: optval = bbr->rc_use_ts_limit; break; case TCP_BBR_IWINTSO: optval = bbr->rc_init_win; break; case TCP_BBR_STARTUP_PG: optval = bbr->r_ctl.rc_startup_pg; break; case TCP_BBR_DRAIN_PG: optval = bbr->r_ctl.rc_drain_pg; break; case TCP_BBR_PROBE_RTT_INT: optval = bbr->r_ctl.rc_probertt_int; break; case TCP_BBR_PROBE_RTT_LEN: optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND); break; case TCP_BBR_PROBE_RTT_GAIN: optval = bbr->r_ctl.bbr_rttprobe_gain_val; break; case TCP_BBR_STARTUP_LOSS_EXIT: optval = bbr->rc_loss_exit; break; case TCP_BBR_USEDEL_RATE: error = EINVAL; break; case TCP_BBR_MIN_RTO: optval = bbr->r_ctl.rc_min_rto_ms; break; case TCP_BBR_MAX_RTO: optval = bbr->rc_max_rto_sec; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments in a pace */ optval = bbr->r_ctl.rc_pace_max_segs; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ optval = bbr->r_ctl.rc_min_to; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ optval = bbr->r_ctl.rc_reorder_shift; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ optval = bbr->r_ctl.rc_reorder_fade; break; case TCP_BBR_USE_RACK_CHEAT: /* Do we use the rack cheat for rxt */ optval = bbr->bbr_use_rack_cheat; break; case TCP_BBR_FLOOR_MIN_TSO: optval = bbr->r_ctl.bbr_hptsi_segments_floor; break; case TCP_BBR_UTTER_MAX_TSO: optval = bbr->r_ctl.bbr_utter_max; break; case TCP_BBR_SEND_IWND_IN_TSO: /* Do we send TSO size segments initially */ optval = bbr->bbr_init_win_cheat; break; case TCP_BBR_EXTRA_STATE: optval = bbr->rc_use_idle_restart; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ optval = bbr->rc_tlp_threshold; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ optval = bbr->r_ctl.rc_pkt_delay; break; case TCP_BBR_RETRAN_WTSO: optval = bbr->rc_resends_use_tso; break; case TCP_DATA_AFTER_CLOSE: optval = bbr->rc_allow_data_af_clo; break; case TCP_DELACK: optval = tp->t_delayed_ack; break; case TCP_BBR_HDWR_PACE: optval = bbr->bbr_hdw_pace_ena; break; case TCP_BBR_POLICER_DETECT: optval = bbr->r_use_policer; break; case TCP_BBR_TSTMP_RAISES: optval = bbr->ts_can_raise; break; case TCP_BBR_TMR_PACE_OH: optval = bbr->r_ctl.rc_incr_tmrs; break; case TCP_BBR_PACE_OH: optval = 0; if (bbr->r_ctl.rc_inc_tcp_oh) optval |= BBR_INCL_TCP_OH; if (bbr->r_ctl.rc_inc_ip_oh) optval |= BBR_INCL_IP_OH; if (bbr->r_ctl.rc_inc_enet_oh) optval |= BBR_INCL_ENET_OH; break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); return (error); } /* * return 0 on success, error-num on failure */ static int bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) { int32_t error = EINVAL; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (bbr == NULL) { /* Huh? */ goto out; } if (sopt->sopt_dir == SOPT_SET) { return (bbr_set_sockopt(so, sopt, inp, tp, bbr)); } else if (sopt->sopt_dir == SOPT_GET) { return (bbr_get_sockopt(so, sopt, inp, tp, bbr)); } out: INP_WUNLOCK(inp); return (error); } static const char *bbr_stack_names[] = { __XSTRING(STACKNAME), #ifdef STACKALIAS __XSTRING(STACKALIAS), #endif }; static bool bbr_mod_inited = false; static int tcp_addbbr(module_t mod, int32_t type, void *data) { int32_t err = 0; int num_stacks; switch (type) { case MOD_LOAD: printf("Attempting to load " __XSTRING(MODNAME) "\n"); bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map", sizeof(struct bbr_sendmap), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", sizeof(struct tcp_bbr), NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); sysctl_ctx_init(&bbr_sysctl_ctx); bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp), OID_AUTO, #ifdef STACKALIAS __XSTRING(STACKALIAS), #else __XSTRING(STACKNAME), #endif CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); if (bbr_sysctl_root == NULL) { printf("Failed to add sysctl node\n"); err = EFAULT; goto free_uma; } bbr_init_sysctls(); num_stacks = nitems(bbr_stack_names); err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK, bbr_stack_names, &num_stacks); if (err) { printf("Failed to register %s stack name for " "%s module\n", bbr_stack_names[num_stacks], __XSTRING(MODNAME)); sysctl_ctx_free(&bbr_sysctl_ctx); free_uma: uma_zdestroy(bbr_zone); uma_zdestroy(bbr_pcb_zone); bbr_counter_destroy(); printf("Failed to register " __XSTRING(MODNAME) " module err:%d\n", err); return (err); } tcp_lro_reg_mbufq(); bbr_mod_inited = true; printf(__XSTRING(MODNAME) " is now available\n"); break; case MOD_QUIESCE: err = deregister_tcp_functions(&__tcp_bbr, true, false); break; case MOD_UNLOAD: err = deregister_tcp_functions(&__tcp_bbr, false, true); if (err == EBUSY) break; if (bbr_mod_inited) { uma_zdestroy(bbr_zone); uma_zdestroy(bbr_pcb_zone); sysctl_ctx_free(&bbr_sysctl_ctx); bbr_counter_destroy(); printf(__XSTRING(MODNAME) " is now no longer available\n"); bbr_mod_inited = false; } tcp_lro_dereg_mbufq(); err = 0; break; default: return (EOPNOTSUPP); } return (err); } static moduledata_t tcp_bbr = { .name = __XSTRING(MODNAME), .evhand = tcp_addbbr, .priv = 0 }; MODULE_VERSION(MODNAME, 1); DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); diff --git a/sys/netinet/tcp_stacks/rack.c b/sys/netinet/tcp_stacks/rack.c index b5727f7873fd..75287282cf3e 100644 --- a/sys/netinet/tcp_stacks/rack.c +++ b/sys/netinet/tcp_stacks/rack.c @@ -1,20217 +1,20217 @@ /*- * Copyright (c) 2016-2020 Netflix, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include "opt_ratelimit.h" #include "opt_kern_tls.h" #include #include #include #include #ifdef TCP_HHOOK #include #endif #include #include #include #include #include #include /* for proc0 declaration */ #include #include #include #include #ifdef STATS #include #include #include /* Must come after qmath.h and tree.h */ #else #include #endif #include #include #include #include #include #include #include #ifdef TCP_ACCOUNTING #include #include #endif #include #include #include #include #define TCPSTATES /* for logging */ #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #include #define TCPOUTFLAGS #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef NETFLIX_SHARED_CWND #include #endif #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #ifdef INET6 #include #endif #include #if defined(IPSEC) || defined(IPSEC_SUPPORT) #include #include #endif /* IPSEC */ #include #include #include #ifdef MAC #include #endif #include "sack_filter.h" #include "tcp_rack.h" #include "rack_bbr_common.h" uma_zone_t rack_zone; uma_zone_t rack_pcb_zone; #ifndef TICKS2SBT #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t))) #endif VNET_DECLARE(uint32_t, newreno_beta); VNET_DECLARE(uint32_t, newreno_beta_ecn); #define V_newreno_beta VNET(newreno_beta) #define V_newreno_beta_ecn VNET(newreno_beta_ecn) MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block"); MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options"); struct sysctl_ctx_list rack_sysctl_ctx; struct sysctl_oid *rack_sysctl_root; #define CUM_ACKED 1 #define SACKED 2 /* * The RACK module incorporates a number of * TCP ideas that have been put out into the IETF * over the last few years: * - Matt Mathis's Rate Halving which slowly drops * the congestion window so that the ack clock can * be maintained during a recovery. * - Yuchung Cheng's RACK TCP (for which its named) that * will stop us using the number of dup acks and instead * use time as the gage of when we retransmit. * - Reorder Detection of RFC4737 and the Tail-Loss probe draft * of Dukkipati et.al. * RACK depends on SACK, so if an endpoint arrives that * cannot do SACK the state machine below will shuttle the * connection back to using the "default" TCP stack that is * in FreeBSD. * * To implement RACK the original TCP stack was first decomposed * into a functional state machine with individual states * for each of the possible TCP connection states. The do_segement * functions role in life is to mandate the connection supports SACK * initially and then assure that the RACK state matches the conenction * state before calling the states do_segment function. Each * state is simplified due to the fact that the original do_segment * has been decomposed and we *know* what state we are in (no * switches on the state) and all tests for SACK are gone. This * greatly simplifies what each state does. * * TCP output is also over-written with a new version since it * must maintain the new rack scoreboard. * */ static int32_t rack_tlp_thresh = 1; static int32_t rack_tlp_limit = 2; /* No more than 2 TLPs w-out new data */ static int32_t rack_tlp_use_greater = 1; static int32_t rack_reorder_thresh = 2; static int32_t rack_reorder_fade = 60000000; /* 0 - never fade, def 60,000,000 * - 60 seconds */ static uint8_t rack_req_measurements = 1; /* Attack threshold detections */ static uint32_t rack_highest_sack_thresh_seen = 0; static uint32_t rack_highest_move_thresh_seen = 0; static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */ static int32_t rack_hw_pace_extra_slots = 2; /* 2 extra MSS time betweens */ static int32_t rack_hw_rate_caps = 1; /* 1; */ static int32_t rack_hw_rate_min = 0; /* 1500000;*/ static int32_t rack_hw_rate_to_low = 0; /* 1200000; */ static int32_t rack_hw_up_only = 1; static int32_t rack_stats_gets_ms_rtt = 1; static int32_t rack_prr_addbackmax = 2; static int32_t rack_pkt_delay = 1000; static int32_t rack_send_a_lot_in_prr = 1; static int32_t rack_min_to = 1000; /* Number of microsecond min timeout */ static int32_t rack_verbose_logging = 0; static int32_t rack_ignore_data_after_close = 1; static int32_t rack_enable_shared_cwnd = 1; static int32_t rack_use_cmp_acks = 1; static int32_t rack_use_fsb = 1; static int32_t rack_use_rfo = 1; static int32_t rack_use_rsm_rfo = 1; static int32_t rack_max_abc_post_recovery = 2; static int32_t rack_client_low_buf = 0; #ifdef TCP_ACCOUNTING static int32_t rack_tcp_accounting = 0; #endif static int32_t rack_limits_scwnd = 1; static int32_t rack_enable_mqueue_for_nonpaced = 0; static int32_t rack_disable_prr = 0; static int32_t use_rack_rr = 1; static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */ static int32_t rack_persist_min = 250000; /* 250usec */ static int32_t rack_persist_max = 2000000; /* 2 Second in usec's */ static int32_t rack_sack_not_required = 1; /* set to one to allow non-sack to use rack */ static int32_t rack_default_init_window = 0; /* Use system default */ static int32_t rack_limit_time_with_srtt = 0; static int32_t rack_autosndbuf_inc = 20; /* In percentage form */ static int32_t rack_enobuf_hw_boost_mult = 2; /* How many times the hw rate we boost slot using time_between */ static int32_t rack_enobuf_hw_max = 12000; /* 12 ms in usecs */ static int32_t rack_enobuf_hw_min = 10000; /* 10 ms in usecs */ static int32_t rack_hw_rwnd_factor = 2; /* How many max_segs the rwnd must be before we hold off sending */ /* * Currently regular tcp has a rto_min of 30ms * the backoff goes 12 times so that ends up * being a total of 122.850 seconds before a * connection is killed. */ static uint32_t rack_def_data_window = 20; static uint32_t rack_goal_bdp = 2; static uint32_t rack_min_srtts = 1; static uint32_t rack_min_measure_usec = 0; static int32_t rack_tlp_min = 10000; /* 10ms */ static int32_t rack_rto_min = 30000; /* 30,000 usec same as main freebsd */ static int32_t rack_rto_max = 4000000; /* 4 seconds in usec's */ static const int32_t rack_free_cache = 2; static int32_t rack_hptsi_segments = 40; static int32_t rack_rate_sample_method = USE_RTT_LOW; static int32_t rack_pace_every_seg = 0; static int32_t rack_delayed_ack_time = 40000; /* 40ms in usecs */ static int32_t rack_slot_reduction = 4; static int32_t rack_wma_divisor = 8; /* For WMA calculation */ static int32_t rack_cwnd_block_ends_measure = 0; static int32_t rack_rwnd_block_ends_measure = 0; static int32_t rack_def_profile = 0; static int32_t rack_lower_cwnd_at_tlp = 0; static int32_t rack_limited_retran = 0; static int32_t rack_always_send_oldest = 0; static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE; static uint16_t rack_per_of_gp_ss = 250; /* 250 % slow-start */ static uint16_t rack_per_of_gp_ca = 200; /* 200 % congestion-avoidance */ static uint16_t rack_per_of_gp_rec = 200; /* 200 % of bw */ /* Probertt */ static uint16_t rack_per_of_gp_probertt = 60; /* 60% of bw */ static uint16_t rack_per_of_gp_lowthresh = 40; /* 40% is bottom */ static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */ static uint16_t rack_atexit_prtt_hbp = 130; /* Clamp to 130% on exit prtt if highly buffered path */ static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */ static uint32_t rack_max_drain_wait = 2; /* How man gp srtt's before we give up draining */ static uint32_t rack_must_drain = 1; /* How many GP srtt's we *must* wait */ static uint32_t rack_probertt_use_min_rtt_entry = 1; /* Use the min to calculate the goal else gp_srtt */ static uint32_t rack_probertt_use_min_rtt_exit = 0; static uint32_t rack_probe_rtt_sets_cwnd = 0; static uint32_t rack_probe_rtt_safety_val = 2000000; /* No more than 2 sec in probe-rtt */ static uint32_t rack_time_between_probertt = 9600000; /* 9.6 sec in usecs */ static uint32_t rack_probertt_gpsrtt_cnt_mul = 0; /* How many srtt periods does probe-rtt last top fraction */ static uint32_t rack_probertt_gpsrtt_cnt_div = 0; /* How many srtt periods does probe-rtt last bottom fraction */ static uint32_t rack_min_probertt_hold = 40000; /* Equal to delayed ack time */ static uint32_t rack_probertt_filter_life = 10000000; static uint32_t rack_probertt_lower_within = 10; static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds) to count as a lowering */ static int32_t rack_pace_one_seg = 0; /* Shall we pace for less than 1.4Meg 1MSS at a time */ static int32_t rack_probertt_clear_is = 1; static int32_t rack_max_drain_hbp = 1; /* Extra drain times gpsrtt for highly buffered paths */ static int32_t rack_hbp_thresh = 3; /* what is the divisor max_rtt/min_rtt to decided a hbp */ /* Part of pacing */ static int32_t rack_max_per_above = 30; /* When we go to increment stop if above 100+this% */ /* Timely information */ /* Combine these two gives the range of 'no change' to bw */ /* ie the up/down provide the upper and lower bound */ static int32_t rack_gp_per_bw_mul_up = 2; /* 2% */ static int32_t rack_gp_per_bw_mul_down = 4; /* 4% */ static int32_t rack_gp_rtt_maxmul = 3; /* 3 x maxmin */ static int32_t rack_gp_rtt_minmul = 1; /* minrtt + (minrtt/mindiv) is lower rtt */ static int32_t rack_gp_rtt_mindiv = 4; /* minrtt + (minrtt * minmul/mindiv) is lower rtt */ static int32_t rack_gp_decrease_per = 20; /* 20% decrease in multipler */ static int32_t rack_gp_increase_per = 2; /* 2% increase in multipler */ static int32_t rack_per_lower_bound = 50; /* Don't allow to drop below this multiplier */ static int32_t rack_per_upper_bound_ss = 0; /* Don't allow SS to grow above this */ static int32_t rack_per_upper_bound_ca = 0; /* Don't allow CA to grow above this */ static int32_t rack_do_dyn_mul = 0; /* Are the rack gp multipliers dynamic */ static int32_t rack_gp_no_rec_chg = 1; /* Prohibit recovery from reducing it's multiplier */ static int32_t rack_timely_dec_clear = 6; /* Do we clear decrement count at a value (6)? */ static int32_t rack_timely_max_push_rise = 3; /* One round of pushing */ static int32_t rack_timely_max_push_drop = 3; /* Three round of pushing */ static int32_t rack_timely_min_segs = 4; /* 4 segment minimum */ static int32_t rack_use_max_for_nobackoff = 0; static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */ static int32_t rack_timely_no_stopping = 0; static int32_t rack_down_raise_thresh = 100; static int32_t rack_req_segs = 1; static uint64_t rack_bw_rate_cap = 0; /* Weird delayed ack mode */ static int32_t rack_use_imac_dack = 0; /* Rack specific counters */ counter_u64_t rack_badfr; counter_u64_t rack_badfr_bytes; counter_u64_t rack_rtm_prr_retran; counter_u64_t rack_rtm_prr_newdata; counter_u64_t rack_timestamp_mismatch; counter_u64_t rack_reorder_seen; counter_u64_t rack_paced_segments; counter_u64_t rack_unpaced_segments; counter_u64_t rack_calc_zero; counter_u64_t rack_calc_nonzero; counter_u64_t rack_saw_enobuf; counter_u64_t rack_saw_enobuf_hw; counter_u64_t rack_saw_enetunreach; counter_u64_t rack_per_timer_hole; counter_u64_t rack_large_ackcmp; counter_u64_t rack_small_ackcmp; #ifdef INVARIANTS counter_u64_t rack_adjust_map_bw; #endif /* Tail loss probe counters */ counter_u64_t rack_tlp_tot; counter_u64_t rack_tlp_newdata; counter_u64_t rack_tlp_retran; counter_u64_t rack_tlp_retran_bytes; counter_u64_t rack_tlp_retran_fail; counter_u64_t rack_to_tot; counter_u64_t rack_to_arm_rack; counter_u64_t rack_to_arm_tlp; counter_u64_t rack_hot_alloc; counter_u64_t rack_to_alloc; counter_u64_t rack_to_alloc_hard; counter_u64_t rack_to_alloc_emerg; counter_u64_t rack_to_alloc_limited; counter_u64_t rack_alloc_limited_conns; counter_u64_t rack_split_limited; #define MAX_NUM_OF_CNTS 13 counter_u64_t rack_proc_comp_ack[MAX_NUM_OF_CNTS]; counter_u64_t rack_multi_single_eq; counter_u64_t rack_proc_non_comp_ack; counter_u64_t rack_fto_send; counter_u64_t rack_fto_rsm_send; counter_u64_t rack_nfto_resend; counter_u64_t rack_non_fto_send; counter_u64_t rack_extended_rfo; counter_u64_t rack_sack_proc_all; counter_u64_t rack_sack_proc_short; counter_u64_t rack_sack_proc_restart; counter_u64_t rack_sack_attacks_detected; counter_u64_t rack_sack_attacks_reversed; counter_u64_t rack_sack_used_next_merge; counter_u64_t rack_sack_splits; counter_u64_t rack_sack_used_prev_merge; counter_u64_t rack_sack_skipped_acked; counter_u64_t rack_ack_total; counter_u64_t rack_express_sack; counter_u64_t rack_sack_total; counter_u64_t rack_move_none; counter_u64_t rack_move_some; counter_u64_t rack_used_tlpmethod; counter_u64_t rack_used_tlpmethod2; counter_u64_t rack_enter_tlp_calc; counter_u64_t rack_input_idle_reduces; counter_u64_t rack_collapsed_win; counter_u64_t rack_tlp_does_nada; counter_u64_t rack_try_scwnd; counter_u64_t rack_hw_pace_init_fail; counter_u64_t rack_hw_pace_lost; counter_u64_t rack_sbsndptr_right; counter_u64_t rack_sbsndptr_wrong; /* Temp CPU counters */ counter_u64_t rack_find_high; counter_u64_t rack_progress_drops; counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE]; counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2))) #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do { \ (tv) = (value) + slop; \ if ((u_long)(tv) < (u_long)(tvmin)) \ (tv) = (tvmin); \ if ((u_long)(tv) > (u_long)(tvmax)) \ (tv) = (tvmax); \ } while (0) static void rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line); static int rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val); static int rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static void rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery); static struct rack_sendmap *rack_alloc(struct tcp_rack *rack); static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type); static struct rack_sendmap * rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused); static void rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack); static void rack_counter_destroy(void); static int rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp); static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how); static void rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override); static void rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos); static void rack_dtor(void *mem, int32_t size, void *arg); static void rack_log_alt_to_to_cancel(struct tcp_rack *rack, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint32_t flex4, uint32_t flex5, uint32_t flex6, uint16_t flex7, uint8_t mod); static void rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm, uint8_t quality); static struct rack_sendmap * rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm); static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack); static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm); static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged); static int rack_get_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); static void rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, int line, uint8_t quality); static uint32_t rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss); static int32_t rack_handoff_ok(struct tcpcb *tp); static int32_t rack_init(struct tcpcb *tp); static void rack_init_sysctls(void); static void rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_rec, int dup_ack_struck); static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t ts, struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls); static void rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm); static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm); static int32_t rack_output(struct tcpcb *tp); static uint32_t rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two); static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq); static void rack_remxt_tmr(struct tcpcb *tp); static int rack_set_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack); static int32_t rack_stopall(struct tcpcb *tp); static void rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta); static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type); static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line); static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type); static uint32_t rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag); static void rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag); static int rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack); static int32_t tcp_addrack(module_t mod, int32_t type, void *data); static int rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos); static int rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); struct rack_sendmap * tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused); static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt); static void tcp_rack_partialack(struct tcpcb *tp); static int rack_set_profile(struct tcp_rack *rack, int prof); static void rack_apply_deferred_options(struct tcp_rack *rack); int32_t rack_clear_counter=0; static void rack_set_cc_pacing(struct tcp_rack *rack) { struct sockopt sopt; struct cc_newreno_opts opt; struct newreno old, *ptr; struct tcpcb *tp; int error; if (rack->rc_pacing_cc_set) return; tp = rack->rc_tp; if (tp->cc_algo == NULL) { /* Tcb is leaving */ printf("No cc algorithm?\n"); return; } rack->rc_pacing_cc_set = 1; if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { /* Not new-reno we can't play games with beta! */ goto out; } ptr = ((struct newreno *)tp->ccv->cc_data); if (CC_ALGO(tp)->ctl_output == NULL) { /* Huh, why does new_reno no longer have a set function? */ printf("no ctl_output for algo:%s\n", tp->cc_algo->name); goto out; } if (ptr == NULL) { /* Just the default values */ old.beta = V_newreno_beta_ecn; old.beta_ecn = V_newreno_beta_ecn; old.newreno_flags = 0; } else { old.beta = ptr->beta; old.beta_ecn = ptr->beta_ecn; old.newreno_flags = ptr->newreno_flags; } sopt.sopt_valsize = sizeof(struct cc_newreno_opts); sopt.sopt_dir = SOPT_SET; opt.name = CC_NEWRENO_BETA; opt.val = rack->r_ctl.rc_saved_beta.beta; error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); if (error) { printf("Error returned by ctl_output %d\n", error); goto out; } /* * Hack alert we need to set in our newreno_flags * so that Abe behavior is also applied. */ ((struct newreno *)tp->ccv->cc_data)->newreno_flags = CC_NEWRENO_BETA_ECN; opt.name = CC_NEWRENO_BETA_ECN; opt.val = rack->r_ctl.rc_saved_beta.beta_ecn; error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); if (error) { printf("Error returned by ctl_output %d\n", error); goto out; } /* Save off the original values for restoral */ memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno)); out: if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; ptr = ((struct newreno *)tp->ccv->cc_data); memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); if (ptr) { log.u_bbr.flex1 = ptr->beta; log.u_bbr.flex2 = ptr->beta_ecn; log.u_bbr.flex3 = ptr->newreno_flags; } log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta; log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn; log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags; log.u_bbr.flex7 = rack->gp_ready; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->use_fixed_rate; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->rc_pacing_cc_set; log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex8 = 3; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error, 0, &log, false, NULL, NULL, 0, &tv); } } static void rack_undo_cc_pacing(struct tcp_rack *rack) { struct newreno old, *ptr; struct tcpcb *tp; if (rack->rc_pacing_cc_set == 0) return; tp = rack->rc_tp; rack->rc_pacing_cc_set = 0; if (tp->cc_algo == NULL) /* Tcb is leaving */ return; if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { /* Not new-reno nothing to do! */ return; } ptr = ((struct newreno *)tp->ccv->cc_data); if (ptr == NULL) { /* * This happens at rack_fini() if the * cc module gets freed on us. In that * case we loose our "new" settings but * thats ok, since the tcb is going away anyway. */ return; } /* Grab out our set values */ memcpy(&old, ptr, sizeof(struct newreno)); /* Copy back in the original values */ memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno)); /* Now save back the values we had set in (for when pacing is restored) */ memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno)); if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; ptr = ((struct newreno *)tp->ccv->cc_data); memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = ptr->beta; log.u_bbr.flex2 = ptr->beta_ecn; log.u_bbr.flex3 = ptr->newreno_flags; log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta; log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn; log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags; log.u_bbr.flex7 = rack->gp_ready; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->use_fixed_rate; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->rc_pacing_cc_set; log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex8 = 4; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } } #ifdef NETFLIX_PEAKRATE static inline void rack_update_peakrate_thr(struct tcpcb *tp) { /* Keep in mind that t_maxpeakrate is in B/s. */ uint64_t peak; peak = uqmax((tp->t_maxseg * 2), (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC)); tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX); } #endif static int sysctl_rack_clear(SYSCTL_HANDLER_ARGS) { uint32_t stat; int32_t error; int i; error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t)); if (error || req->newptr == NULL) return error; error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); if (error) return (error); if (stat == 1) { #ifdef INVARIANTS printf("Clearing RACK counters\n"); #endif counter_u64_zero(rack_badfr); counter_u64_zero(rack_badfr_bytes); counter_u64_zero(rack_rtm_prr_retran); counter_u64_zero(rack_rtm_prr_newdata); counter_u64_zero(rack_timestamp_mismatch); counter_u64_zero(rack_reorder_seen); counter_u64_zero(rack_tlp_tot); counter_u64_zero(rack_tlp_newdata); counter_u64_zero(rack_tlp_retran); counter_u64_zero(rack_tlp_retran_bytes); counter_u64_zero(rack_tlp_retran_fail); counter_u64_zero(rack_to_tot); counter_u64_zero(rack_to_arm_rack); counter_u64_zero(rack_to_arm_tlp); counter_u64_zero(rack_paced_segments); counter_u64_zero(rack_calc_zero); counter_u64_zero(rack_calc_nonzero); counter_u64_zero(rack_unpaced_segments); counter_u64_zero(rack_saw_enobuf); counter_u64_zero(rack_saw_enobuf_hw); counter_u64_zero(rack_saw_enetunreach); counter_u64_zero(rack_per_timer_hole); counter_u64_zero(rack_large_ackcmp); counter_u64_zero(rack_small_ackcmp); #ifdef INVARIANTS counter_u64_zero(rack_adjust_map_bw); #endif counter_u64_zero(rack_to_alloc_hard); counter_u64_zero(rack_to_alloc_emerg); counter_u64_zero(rack_sack_proc_all); counter_u64_zero(rack_fto_send); counter_u64_zero(rack_fto_rsm_send); counter_u64_zero(rack_extended_rfo); counter_u64_zero(rack_hw_pace_init_fail); counter_u64_zero(rack_hw_pace_lost); counter_u64_zero(rack_sbsndptr_wrong); counter_u64_zero(rack_sbsndptr_right); counter_u64_zero(rack_non_fto_send); counter_u64_zero(rack_nfto_resend); counter_u64_zero(rack_sack_proc_short); counter_u64_zero(rack_sack_proc_restart); counter_u64_zero(rack_to_alloc); counter_u64_zero(rack_to_alloc_limited); counter_u64_zero(rack_alloc_limited_conns); counter_u64_zero(rack_split_limited); for (i = 0; i < MAX_NUM_OF_CNTS; i++) { counter_u64_zero(rack_proc_comp_ack[i]); } counter_u64_zero(rack_multi_single_eq); counter_u64_zero(rack_proc_non_comp_ack); counter_u64_zero(rack_find_high); counter_u64_zero(rack_sack_attacks_detected); counter_u64_zero(rack_sack_attacks_reversed); counter_u64_zero(rack_sack_used_next_merge); counter_u64_zero(rack_sack_used_prev_merge); counter_u64_zero(rack_sack_splits); counter_u64_zero(rack_sack_skipped_acked); counter_u64_zero(rack_ack_total); counter_u64_zero(rack_express_sack); counter_u64_zero(rack_sack_total); counter_u64_zero(rack_move_none); counter_u64_zero(rack_move_some); counter_u64_zero(rack_used_tlpmethod); counter_u64_zero(rack_used_tlpmethod2); counter_u64_zero(rack_enter_tlp_calc); counter_u64_zero(rack_progress_drops); counter_u64_zero(rack_tlp_does_nada); counter_u64_zero(rack_try_scwnd); counter_u64_zero(rack_collapsed_win); } rack_clear_counter = 0; return (0); } static void rack_init_sysctls(void) { int i; struct sysctl_oid *rack_counters; struct sysctl_oid *rack_attack; struct sysctl_oid *rack_pacing; struct sysctl_oid *rack_timely; struct sysctl_oid *rack_timers; struct sysctl_oid *rack_tlp; struct sysctl_oid *rack_misc; struct sysctl_oid *rack_measure; struct sysctl_oid *rack_probertt; struct sysctl_oid *rack_hw_pacing; rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "sack_attack", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack Sack Attack Counters and Controls"); rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "stats", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack Counters"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "rate_sample_method", CTLFLAG_RW, &rack_rate_sample_method , USE_RTT_LOW, "What method should we use for rate sampling 0=high, 1=low "); /* Probe rtt related controls */ rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "probertt", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "ProbeRTT related Controls"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "exit_per_hpb", CTLFLAG_RW, &rack_atexit_prtt_hbp, 130, "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW, &rack_atexit_prtt, 130, "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "gp_per_mul", CTLFLAG_RW, &rack_per_of_gp_probertt, 60, "What percentage of goodput do we pace at in probertt"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "gp_per_reduce", CTLFLAG_RW, &rack_per_of_gp_probertt_reduce, 10, "What percentage of goodput do we reduce every gp_srtt"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "gp_per_low", CTLFLAG_RW, &rack_per_of_gp_lowthresh, 40, "What percentage of goodput do we allow the multiplier to fall to"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "time_between", CTLFLAG_RW, & rack_time_between_probertt, 96000000, "How many useconds between the lowest rtt falling must past before we enter probertt"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "safety", CTLFLAG_RW, &rack_probe_rtt_safety_val, 2000000, "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "sets_cwnd", CTLFLAG_RW, &rack_probe_rtt_sets_cwnd, 0, "Do we set the cwnd too (if always_lower is on)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "maxdrainsrtts", CTLFLAG_RW, &rack_max_drain_wait, 2, "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "mustdrainsrtts", CTLFLAG_RW, &rack_must_drain, 1, "We must drain this many gp_srtt's waiting for flight to reach goal"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "goal_use_min_entry", CTLFLAG_RW, &rack_probertt_use_min_rtt_entry, 1, "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "goal_use_min_exit", CTLFLAG_RW, &rack_probertt_use_min_rtt_exit, 0, "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "length_div", CTLFLAG_RW, &rack_probertt_gpsrtt_cnt_div, 0, "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "length_mul", CTLFLAG_RW, &rack_probertt_gpsrtt_cnt_mul, 0, "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "holdtim_at_target", CTLFLAG_RW, &rack_min_probertt_hold, 200000, "What is the minimum time we hold probertt at target"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "filter_life", CTLFLAG_RW, &rack_probertt_filter_life, 10000000, "What is the time for the filters life in useconds"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "lower_within", CTLFLAG_RW, &rack_probertt_lower_within, 10, "If the rtt goes lower within this percentage of the time, go into probe-rtt"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "must_move", CTLFLAG_RW, &rack_min_rtt_movement, 250, "How much is the minimum movement in rtt to count as a drop for probertt purposes"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "clear_is_cnts", CTLFLAG_RW, &rack_probertt_clear_is, 1, "Do we clear I/S counts on exiting probe-rtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "hbp_extra_drain", CTLFLAG_RW, &rack_max_drain_hbp, 1, "How many extra drain gpsrtt's do we get in highly buffered paths"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "hbp_threshold", CTLFLAG_RW, &rack_hbp_thresh, 3, "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold"); /* Pacing related sysctls */ rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "pacing", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Pacing related Controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "max_pace_over", CTLFLAG_RW, &rack_max_per_above, 30, "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "pace_to_one", CTLFLAG_RW, &rack_pace_one_seg, 0, "Do we allow low b/w pacing of 1MSS instead of two"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "limit_wsrtt", CTLFLAG_RW, &rack_limit_time_with_srtt, 0, "Do we limit pacing time based on srtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "init_win", CTLFLAG_RW, &rack_default_init_window, 0, "Do we have a rack initial window 0 = system default"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "gp_per_ss", CTLFLAG_RW, &rack_per_of_gp_ss, 250, "If non zero, what percentage of goodput to pace at in slow start"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "gp_per_ca", CTLFLAG_RW, &rack_per_of_gp_ca, 150, "If non zero, what percentage of goodput to pace at in congestion avoidance"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "gp_per_rec", CTLFLAG_RW, &rack_per_of_gp_rec, 200, "If non zero, what percentage of goodput to pace at in recovery"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "pace_max_seg", CTLFLAG_RW, &rack_hptsi_segments, 40, "What size is the max for TSO segments in pacing and burst mitigation"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "burst_reduces", CTLFLAG_RW, &rack_slot_reduction, 4, "When doing only burst mitigation what is the reduce divisor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "use_pacing", CTLFLAG_RW, &rack_pace_every_seg, 0, "If set we use pacing, if clear we use only the original burst mitigation"); SYSCTL_ADD_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "rate_cap", CTLFLAG_RW, &rack_bw_rate_cap, 0, "If set we apply this value to the absolute rate cap used by pacing"); SYSCTL_ADD_U8(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "req_measure_cnt", CTLFLAG_RW, &rack_req_measurements, 1, "If doing dynamic pacing, how many measurements must be in before we start pacing?"); /* Hardware pacing */ rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hdwr_pacing", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Pacing related Controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rwnd_factor", CTLFLAG_RW, &rack_hw_rwnd_factor, 2, "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW, &rack_enobuf_hw_boost_mult, 2, "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "pace_enobuf_max", CTLFLAG_RW, &rack_enobuf_hw_max, 2, "What is the max boost the pacing time if we see a ENOBUFS?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "pace_enobuf_min", CTLFLAG_RW, &rack_enobuf_hw_min, 2, "What is the min boost the pacing time if we see a ENOBUFS?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "enable", CTLFLAG_RW, &rack_enable_hw_pacing, 0, "Should RACK attempt to use hw pacing?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rate_cap", CTLFLAG_RW, &rack_hw_rate_caps, 1, "Does the highest hardware pacing rate cap the rate we will send at??"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rate_min", CTLFLAG_RW, &rack_hw_rate_min, 0, "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rate_to_low", CTLFLAG_RW, &rack_hw_rate_to_low, 0, "If we fall below this rate, dis-engage hw pacing?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "up_only", CTLFLAG_RW, &rack_hw_up_only, 1, "Do we allow hw pacing to lower the rate selected?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "extra_mss_precise", CTLFLAG_RW, &rack_hw_pace_extra_slots, 2, "If the rates between software and hardware match precisely how many extra time_betweens do we get?"); rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "timely", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack Timely RTT Controls"); /* Timely based GP dynmics */ SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "upper", CTLFLAG_RW, &rack_gp_per_bw_mul_up, 2, "Rack timely upper range for equal b/w (in percentage)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "lower", CTLFLAG_RW, &rack_gp_per_bw_mul_down, 4, "Rack timely lower range for equal b/w (in percentage)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "rtt_max_mul", CTLFLAG_RW, &rack_gp_rtt_maxmul, 3, "Rack timely multipler of lowest rtt for rtt_max"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "rtt_min_div", CTLFLAG_RW, &rack_gp_rtt_mindiv, 4, "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "rtt_min_mul", CTLFLAG_RW, &rack_gp_rtt_minmul, 1, "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "decrease", CTLFLAG_RW, &rack_gp_decrease_per, 20, "Rack timely decrease percentage of our GP multiplication factor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "increase", CTLFLAG_RW, &rack_gp_increase_per, 2, "Rack timely increase perentage of our GP multiplication factor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "lowerbound", CTLFLAG_RW, &rack_per_lower_bound, 50, "Rack timely lowest percentage we allow GP multiplier to fall to"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "upperboundss", CTLFLAG_RW, &rack_per_upper_bound_ss, 0, "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "upperboundca", CTLFLAG_RW, &rack_per_upper_bound_ca, 0, "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "dynamicgp", CTLFLAG_RW, &rack_do_dyn_mul, 0, "Rack timely do we enable dynmaic timely goodput by default"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "no_rec_red", CTLFLAG_RW, &rack_gp_no_rec_chg, 1, "Rack timely do we prohibit the recovery multiplier from being lowered"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "red_clear_cnt", CTLFLAG_RW, &rack_timely_dec_clear, 6, "Rack timely what threshold do we count to before another boost during b/w decent"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "max_push_rise", CTLFLAG_RW, &rack_timely_max_push_rise, 3, "Rack timely how many times do we push up with b/w increase"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "max_push_drop", CTLFLAG_RW, &rack_timely_max_push_drop, 3, "Rack timely how many times do we push back on b/w decent"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "min_segs", CTLFLAG_RW, &rack_timely_min_segs, 4, "Rack timely when setting the cwnd what is the min num segments"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "noback_max", CTLFLAG_RW, &rack_use_max_for_nobackoff, 0, "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "interim_timely_only", CTLFLAG_RW, &rack_timely_int_timely_only, 0, "Rack timely when doing interim timely's do we only do timely (no b/w consideration)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "nonstop", CTLFLAG_RW, &rack_timely_no_stopping, 0, "Rack timely don't stop increase"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "dec_raise_thresh", CTLFLAG_RW, &rack_down_raise_thresh, 100, "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "bottom_drag_segs", CTLFLAG_RW, &rack_req_segs, 1, "Bottom dragging if not these many segments outstanding and room"); /* TLP and Rack related parameters */ rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TLP and Rack related Controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "use_rrr", CTLFLAG_RW, &use_rack_rr, 1, "Do we use Rack Rapid Recovery"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "post_rec_labc", CTLFLAG_RW, &rack_max_abc_post_recovery, 2, "Since we do early recovery, do we override the l_abc to a value, if so what?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW, &rack_non_rxt_use_cr, 0, "Do we use ss/ca rate if in recovery we are transmitting a new data chunk"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "tlpmethod", CTLFLAG_RW, &rack_tlp_threshold_use, TLP_USE_TWO_ONE, "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "limit", CTLFLAG_RW, &rack_tlp_limit, 2, "How many TLP's can be sent without sending new data"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "use_greater", CTLFLAG_RW, &rack_tlp_use_greater, 1, "Should we use the rack_rtt time if its greater than srtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "tlpminto", CTLFLAG_RW, &rack_tlp_min, 10000, "TLP minimum timeout per the specification (in microseconds)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "send_oldest", CTLFLAG_RW, &rack_always_send_oldest, 0, "Should we always send the oldest TLP and RACK-TLP"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "rack_tlimit", CTLFLAG_RW, &rack_limited_retran, 0, "How many times can a rack timeout drive out sends"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW, &rack_lower_cwnd_at_tlp, 0, "When a TLP completes a retran should we enter recovery"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "reorder_thresh", CTLFLAG_RW, &rack_reorder_thresh, 2, "What factor for rack will be added when seeing reordering (shift right)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, &rack_tlp_thresh, 1, "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "reorder_fade", CTLFLAG_RW, &rack_reorder_fade, 60000000, "Does reorder detection fade, if so how many microseconds (0 means never)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "pktdelay", CTLFLAG_RW, &rack_pkt_delay, 1000, "Extra RACK time (in microseconds) besides reordering thresh"); /* Timer related controls */ rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "timers", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Timer related controls"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "persmin", CTLFLAG_RW, &rack_persist_min, 250000, "What is the minimum time in microseconds between persists"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "persmax", CTLFLAG_RW, &rack_persist_max, 2000000, "What is the largest delay in microseconds between persists"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "delayed_ack", CTLFLAG_RW, &rack_delayed_ack_time, 40000, "Delayed ack time (40ms in microseconds)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "minrto", CTLFLAG_RW, &rack_rto_min, 30000, "Minimum RTO in microseconds -- set with caution below 1000 due to TLP"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "maxrto", CTLFLAG_RW, &rack_rto_max, 4000000, "Maxiumum RTO in microseconds -- should be at least as large as min_rto"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "minto", CTLFLAG_RW, &rack_min_to, 1000, "Minimum rack timeout in microseconds"); /* Measure controls */ rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "measure", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Measure related controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "wma_divisor", CTLFLAG_RW, &rack_wma_divisor, 8, "When doing b/w calculation what is the divisor for the WMA"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "end_cwnd", CTLFLAG_RW, &rack_cwnd_block_ends_measure, 0, "Does a cwnd just-return end the measurement window (app limited)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "end_rwnd", CTLFLAG_RW, &rack_rwnd_block_ends_measure, 0, "Does an rwnd just-return end the measurement window (app limited -- not persists)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "min_target", CTLFLAG_RW, &rack_def_data_window, 20, "What is the minimum target window (in mss) for a GP measurements"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "goal_bdp", CTLFLAG_RW, &rack_goal_bdp, 2, "What is the goal BDP to measure"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "min_srtts", CTLFLAG_RW, &rack_min_srtts, 1, "What is the goal BDP to measure"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "min_measure_tim", CTLFLAG_RW, &rack_min_measure_usec, 0, "What is the Minimum time time for a measurement if 0, this is off"); /* Misc rack controls */ rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "misc", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Misc related controls"); #ifdef TCP_ACCOUNTING SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "tcp_acct", CTLFLAG_RW, &rack_tcp_accounting, 0, "Should we turn on TCP accounting for all rack sessions?"); #endif SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "prr_addback_max", CTLFLAG_RW, &rack_prr_addbackmax, 2, "What is the maximum number of MSS we allow to be added back if prr can't send all its data?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "stats_gets_ms", CTLFLAG_RW, &rack_stats_gets_ms_rtt, 1, "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "clientlowbuf", CTLFLAG_RW, &rack_client_low_buf, 0, "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "defprofile", CTLFLAG_RW, &rack_def_profile, 0, "Should RACK use a default profile (0=no, num == profile num)?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "cmpack", CTLFLAG_RW, &rack_use_cmp_acks, 1, "Should RACK have LRO send compressed acks"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "fsb", CTLFLAG_RW, &rack_use_fsb, 1, "Should RACK use the fast send block?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "rfo", CTLFLAG_RW, &rack_use_rfo, 1, "Should RACK use rack_fast_output()?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "rsmrfo", CTLFLAG_RW, &rack_use_rsm_rfo, 1, "Should RACK use rack_fast_rsm_output()?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "shared_cwnd", CTLFLAG_RW, &rack_enable_shared_cwnd, 1, "Should RACK try to use the shared cwnd on connections where allowed"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "limits_on_scwnd", CTLFLAG_RW, &rack_limits_scwnd, 1, "Should RACK place low end time limits on the shared cwnd feature"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW, &rack_enable_mqueue_for_nonpaced, 0, "Should RACK use mbuf queuing for non-paced connections"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "iMac_dack", CTLFLAG_RW, &rack_use_imac_dack, 0, "Should RACK try to emulate iMac delayed ack"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "no_prr", CTLFLAG_RW, &rack_disable_prr, 0, "Should RACK not use prr and only pace (must have pacing on)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "bb_verbose", CTLFLAG_RW, &rack_verbose_logging, 0, "Should RACK black box logging be verbose"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "data_after_close", CTLFLAG_RW, &rack_ignore_data_after_close, 1, "Do we hold off sending a RST until all pending data is ack'd"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "no_sack_needed", CTLFLAG_RW, &rack_sack_not_required, 1, "Do we allow rack to run on connections not supporting SACK"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "prr_sendalot", CTLFLAG_RW, &rack_send_a_lot_in_prr, 1, "Send a lot in prr"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "autoscale", CTLFLAG_RW, &rack_autosndbuf_inc, 20, "What percentage should rack scale up its snd buffer by?"); /* Sack Attacker detection stuff */ SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "detect_highsackratio", CTLFLAG_RW, &rack_highest_sack_thresh_seen, 0, "Highest sack to ack ratio seen"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "detect_highmoveratio", CTLFLAG_RW, &rack_highest_move_thresh_seen, 0, "Highest move to non-move ratio seen"); rack_ack_total = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "acktotal", CTLFLAG_RD, &rack_ack_total, "Total number of Ack's"); rack_express_sack = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "exp_sacktotal", CTLFLAG_RD, &rack_express_sack, "Total expresss number of Sack's"); rack_sack_total = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "sacktotal", CTLFLAG_RD, &rack_sack_total, "Total number of SACKs"); rack_move_none = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "move_none", CTLFLAG_RD, &rack_move_none, "Total number of SACK index reuse of postions under threshold"); rack_move_some = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "move_some", CTLFLAG_RD, &rack_move_some, "Total number of SACK index reuse of postions over threshold"); rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "attacks", CTLFLAG_RD, &rack_sack_attacks_detected, "Total number of SACK attackers that had sack disabled"); rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "reversed", CTLFLAG_RD, &rack_sack_attacks_reversed, "Total number of SACK attackers that were later determined false positive"); rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "nextmerge", CTLFLAG_RD, &rack_sack_used_next_merge, "Total number of times we used the next merge"); rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "prevmerge", CTLFLAG_RD, &rack_sack_used_prev_merge, "Total number of times we used the prev merge"); /* Counters */ rack_fto_send = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "fto_send", CTLFLAG_RD, &rack_fto_send, "Total number of rack_fast_output sends"); rack_fto_rsm_send = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "fto_rsm_send", CTLFLAG_RD, &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends"); rack_nfto_resend = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "nfto_resend", CTLFLAG_RD, &rack_nfto_resend, "Total number of rack_output retransmissions"); rack_non_fto_send = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "nfto_send", CTLFLAG_RD, &rack_non_fto_send, "Total number of rack_output first sends"); rack_extended_rfo = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "rfo_extended", CTLFLAG_RD, &rack_extended_rfo, "Total number of times we extended rfo"); rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hwpace_init_fail", CTLFLAG_RD, &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing"); rack_hw_pace_lost = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hwpace_lost", CTLFLAG_RD, &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing"); rack_badfr = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "badfr", CTLFLAG_RD, &rack_badfr, "Total number of bad FRs"); rack_badfr_bytes = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "badfr_bytes", CTLFLAG_RD, &rack_badfr_bytes, "Total number of bad FRs"); rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "prrsndret", CTLFLAG_RD, &rack_rtm_prr_retran, "Total number of prr based retransmits"); rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "prrsndnew", CTLFLAG_RD, &rack_rtm_prr_newdata, "Total number of prr based new transmits"); rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tsnf", CTLFLAG_RD, &rack_timestamp_mismatch, "Total number of timestamps that we could not find the reported ts"); rack_find_high = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "findhigh", CTLFLAG_RD, &rack_find_high, "Total number of FIN causing find-high"); rack_reorder_seen = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "reordering", CTLFLAG_RD, &rack_reorder_seen, "Total number of times we added delay due to reordering"); rack_tlp_tot = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_to_total", CTLFLAG_RD, &rack_tlp_tot, "Total number of tail loss probe expirations"); rack_tlp_newdata = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_new", CTLFLAG_RD, &rack_tlp_newdata, "Total number of tail loss probe sending new data"); rack_tlp_retran = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_retran", CTLFLAG_RD, &rack_tlp_retran, "Total number of tail loss probe sending retransmitted data"); rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD, &rack_tlp_retran_bytes, "Total bytes of tail loss probe sending retransmitted data"); rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_retran_fail", CTLFLAG_RD, &rack_tlp_retran_fail, "Total number of tail loss probe sending retransmitted data that failed (wait for t3)"); rack_to_tot = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "rack_to_tot", CTLFLAG_RD, &rack_to_tot, "Total number of times the rack to expired"); rack_to_arm_rack = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "arm_rack", CTLFLAG_RD, &rack_to_arm_rack, "Total number of times the rack timer armed"); rack_to_arm_tlp = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "arm_tlp", CTLFLAG_RD, &rack_to_arm_tlp, "Total number of times the tlp timer armed"); rack_calc_zero = counter_u64_alloc(M_WAITOK); rack_calc_nonzero = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "calc_zero", CTLFLAG_RD, &rack_calc_zero, "Total number of times pacing time worked out to zero"); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "calc_nonzero", CTLFLAG_RD, &rack_calc_nonzero, "Total number of times pacing time worked out to non-zero"); rack_paced_segments = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "paced", CTLFLAG_RD, &rack_paced_segments, "Total number of times a segment send caused hptsi"); rack_unpaced_segments = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "unpaced", CTLFLAG_RD, &rack_unpaced_segments, "Total number of times a segment did not cause hptsi"); rack_saw_enobuf = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "saw_enobufs", CTLFLAG_RD, &rack_saw_enobuf, "Total number of times a sends returned enobuf for non-hdwr paced connections"); rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD, &rack_saw_enobuf_hw, "Total number of times a send returned enobuf for hdwr paced connections"); rack_saw_enetunreach = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "saw_enetunreach", CTLFLAG_RD, &rack_saw_enetunreach, "Total number of times a send received a enetunreachable"); rack_hot_alloc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "alloc_hot", CTLFLAG_RD, &rack_hot_alloc, "Total allocations from the top of our list"); rack_to_alloc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "allocs", CTLFLAG_RD, &rack_to_alloc, "Total allocations of tracking structures"); rack_to_alloc_hard = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "allochard", CTLFLAG_RD, &rack_to_alloc_hard, "Total allocations done with sleeping the hard way"); rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "allocemerg", CTLFLAG_RD, &rack_to_alloc_emerg, "Total allocations done from emergency cache"); rack_to_alloc_limited = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "alloc_limited", CTLFLAG_RD, &rack_to_alloc_limited, "Total allocations dropped due to limit"); rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "alloc_limited_conns", CTLFLAG_RD, &rack_alloc_limited_conns, "Connections with allocations dropped due to limit"); rack_split_limited = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "split_limited", CTLFLAG_RD, &rack_split_limited, "Split allocations dropped due to limit"); for (i = 0; i < MAX_NUM_OF_CNTS; i++) { char name[32]; sprintf(name, "cmp_ack_cnt_%d", i); rack_proc_comp_ack[i] = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, name, CTLFLAG_RD, &rack_proc_comp_ack[i], "Number of compressed acks we processed"); } rack_large_ackcmp = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "cmp_large_mbufs", CTLFLAG_RD, &rack_large_ackcmp, "Number of TCP connections with large mbuf's for compressed acks"); rack_small_ackcmp = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "cmp_small_mbufs", CTLFLAG_RD, &rack_small_ackcmp, "Number of TCP connections with small mbuf's for compressed acks"); #ifdef INVARIANTS rack_adjust_map_bw = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "map_adjust_req", CTLFLAG_RD, &rack_adjust_map_bw, "Number of times we hit the case where the sb went up and down on a sendmap entry"); #endif rack_multi_single_eq = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD, &rack_multi_single_eq, "Number of compressed acks total represented"); rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "cmp_ack_not", CTLFLAG_RD, &rack_proc_non_comp_ack, "Number of non compresseds acks that we processed"); rack_sack_proc_all = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sack_long", CTLFLAG_RD, &rack_sack_proc_all, "Total times we had to walk whole list for sack processing"); rack_sack_proc_restart = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sack_restart", CTLFLAG_RD, &rack_sack_proc_restart, "Total times we had to walk whole list due to a restart"); rack_sack_proc_short = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sack_short", CTLFLAG_RD, &rack_sack_proc_short, "Total times we took shortcut for sack processing"); rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_calc_entered", CTLFLAG_RD, &rack_enter_tlp_calc, "Total times we called calc-tlp"); rack_used_tlpmethod = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hit_tlp_method", CTLFLAG_RD, &rack_used_tlpmethod, "Total number of runt sacks"); rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hit_tlp_method2", CTLFLAG_RD, &rack_used_tlpmethod2, "Total number of times we hit TLP method 2"); rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "skipacked", CTLFLAG_RD, &rack_sack_skipped_acked, "Total number of times we skipped previously sacked"); rack_sack_splits = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "ofsplit", CTLFLAG_RD, &rack_sack_splits, "Total number of times we did the old fashion tree split"); rack_progress_drops = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "prog_drops", CTLFLAG_RD, &rack_progress_drops, "Total number of progress drops"); rack_input_idle_reduces = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD, &rack_input_idle_reduces, "Total number of idle reductions on input"); rack_collapsed_win = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "collapsed_win", CTLFLAG_RD, &rack_collapsed_win, "Total number of collapsed windows"); rack_tlp_does_nada = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_nada", CTLFLAG_RD, &rack_tlp_does_nada, "Total number of nada tlp calls"); rack_try_scwnd = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tried_scwnd", CTLFLAG_RD, &rack_try_scwnd, "Total number of scwnd attempts"); rack_per_timer_hole = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "timer_hole", CTLFLAG_RD, &rack_per_timer_hole, "Total persists start in timer hole"); rack_sbsndptr_wrong = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sndptr_wrong", CTLFLAG_RD, &rack_sbsndptr_wrong, "Total number of times the saved sbsndptr was incorret"); rack_sbsndptr_right = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sndptr_right", CTLFLAG_RD, &rack_sbsndptr_right, "Total number of times the saved sbsndptr was corret"); COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "outsize", CTLFLAG_RD, rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes"); COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "opts", CTLFLAG_RD, rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats"); SYSCTL_ADD_PROC(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters"); } static __inline int rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a) { if (SEQ_GEQ(b->r_start, a->r_start) && SEQ_LT(b->r_start, a->r_end)) { /* * The entry b is within the * block a. i.e.: * a -- |-------------| * b -- |----| * * b -- |------| * * b -- |-----------| */ return (0); } else if (SEQ_GEQ(b->r_start, a->r_end)) { /* * b falls as either the next * sequence block after a so a * is said to be smaller than b. * i.e: * a -- |------| * b -- |--------| * or * b -- |-----| */ return (1); } /* * Whats left is where a is * larger than b. i.e: * a -- |-------| * b -- |---| * or even possibly * b -- |--------------| */ return (-1); } RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp); RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp); static uint32_t rc_init_window(struct tcp_rack *rack) { uint32_t win; if (rack->rc_init_win == 0) { /* * Nothing set by the user, use the system stack * default. */ return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp))); } win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win; return (win); } static uint64_t rack_get_fixed_pacing_bw(struct tcp_rack *rack) { if (IN_FASTRECOVERY(rack->rc_tp->t_flags)) return (rack->r_ctl.rc_fixed_pacing_rate_rec); else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) return (rack->r_ctl.rc_fixed_pacing_rate_ss); else return (rack->r_ctl.rc_fixed_pacing_rate_ca); } static uint64_t rack_get_bw(struct tcp_rack *rack) { if (rack->use_fixed_rate) { /* Return the fixed pacing rate */ return (rack_get_fixed_pacing_bw(rack)); } if (rack->r_ctl.gp_bw == 0) { /* * We have yet no b/w measurement, * if we have a user set initial bw * return it. If we don't have that and * we have an srtt, use the tcp IW (10) to * calculate a fictional b/w over the SRTT * which is more or less a guess. Note * we don't use our IW from rack on purpose * so if we have like IW=30, we are not * calculating a "huge" b/w. */ uint64_t bw, srtt; if (rack->r_ctl.init_rate) return (rack->r_ctl.init_rate); /* Has the user set a max peak rate? */ #ifdef NETFLIX_PEAKRATE if (rack->rc_tp->t_maxpeakrate) return (rack->rc_tp->t_maxpeakrate); #endif /* Ok lets come up with the IW guess, if we have a srtt */ if (rack->rc_tp->t_srtt == 0) { /* * Go with old pacing method * i.e. burst mitigation only. */ return (0); } /* Ok lets get the initial TCP win (not racks) */ bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)); srtt = (uint64_t)rack->rc_tp->t_srtt; bw *= (uint64_t)USECS_IN_SECOND; bw /= srtt; if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap)) bw = rack->r_ctl.bw_rate_cap; return (bw); } else { uint64_t bw; if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) { /* Averaging is done, we can return the value */ bw = rack->r_ctl.gp_bw; } else { /* Still doing initial average must calculate */ bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements; } #ifdef NETFLIX_PEAKRATE if ((rack->rc_tp->t_maxpeakrate) && (bw > rack->rc_tp->t_maxpeakrate)) { /* The user has set a peak rate to pace at * don't allow us to pace faster than that. */ return (rack->rc_tp->t_maxpeakrate); } #endif if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap)) bw = rack->r_ctl.bw_rate_cap; return (bw); } } static uint16_t rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm) { if (rack->use_fixed_rate) { return (100); } else if (rack->in_probe_rtt && (rsm == NULL)) return (rack->r_ctl.rack_per_of_gp_probertt); else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) && rack->r_ctl.rack_per_of_gp_rec)) { if (rsm) { /* a retransmission always use the recovery rate */ return (rack->r_ctl.rack_per_of_gp_rec); } else if (rack->rack_rec_nonrxt_use_cr) { /* Directed to use the configured rate */ goto configured_rate; } else if (rack->rack_no_prr && (rack->r_ctl.rack_per_of_gp_rec > 100)) { /* No PRR, lets just use the b/w estimate only */ return (100); } else { /* * Here we may have a non-retransmit but we * have no overrides, so just use the recovery * rate (prr is in effect). */ return (rack->r_ctl.rack_per_of_gp_rec); } } configured_rate: /* For the configured rate we look at our cwnd vs the ssthresh */ if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) return (rack->r_ctl.rack_per_of_gp_ss); else return (rack->r_ctl.rack_per_of_gp_ca); } static void rack_log_hdwr_pacing(struct tcp_rack *rack, uint64_t rate, uint64_t hw_rate, int line, int error, uint16_t mod) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; const struct ifnet *ifp; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); if (rack->r_ctl.crte) { ifp = rack->r_ctl.crte->ptbl->rs_ifp; } else if (rack->rc_inp->inp_route.ro_nh && rack->rc_inp->inp_route.ro_nh->nh_ifp) { ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp; } else ifp = NULL; if (ifp) { log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); } log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.bw_inuse = rate; log.u_bbr.flex5 = line; log.u_bbr.flex6 = error; log.u_bbr.flex7 = mod; log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex8 = rack->use_fixed_rate; log.u_bbr.flex8 <<= 1; log.u_bbr.flex8 |= rack->rack_hdrw_pacing; log.u_bbr.pkts_out = rack->rc_tp->t_maxseg; log.u_bbr.delRate = rack->r_ctl.crte_prev_rate; if (rack->r_ctl.crte) log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate; else log.u_bbr.cur_del_rate = 0; log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_HDWR_PACE, 0, 0, &log, false, &tv); } } static uint64_t rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped) { /* * We allow rack_per_of_gp_xx to dictate our bw rate we want. */ uint64_t bw_est, high_rate; uint64_t gain; gain = (uint64_t)rack_get_output_gain(rack, rsm); bw_est = bw * gain; bw_est /= (uint64_t)100; /* Never fall below the minimum (def 64kbps) */ if (bw_est < RACK_MIN_BW) bw_est = RACK_MIN_BW; if (rack->r_rack_hw_rate_caps) { /* Rate caps are in place */ if (rack->r_ctl.crte != NULL) { /* We have a hdwr rate already */ high_rate = tcp_hw_highest_rate(rack->r_ctl.crte); if (bw_est >= high_rate) { /* We are capping bw at the highest rate table entry */ rack_log_hdwr_pacing(rack, bw_est, high_rate, __LINE__, 0, 3); bw_est = high_rate; if (capped) *capped = 1; } } else if ((rack->rack_hdrw_pacing == 0) && (rack->rack_hdw_pace_ena) && (rack->rack_attempt_hdwr_pace == 0) && (rack->rc_inp->inp_route.ro_nh != NULL) && (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) { /* * Special case, we have not yet attempted hardware * pacing, and yet we may, when we do, find out if we are * above the highest rate. We need to know the maxbw for the interface * in question (if it supports ratelimiting). We get back * a 0, if the interface is not found in the RL lists. */ high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp); if (high_rate) { /* Yep, we have a rate is it above this rate? */ if (bw_est > high_rate) { bw_est = high_rate; if (capped) *capped = 1; } } } } return (bw_est); } static void rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; if ((mod != 1) && (rack_verbose_logging == 0)) { /* * We get 3 values currently for mod * 1 - We are retransmitting and this tells the reason. * 2 - We are clearing a dup-ack count. * 3 - We are incrementing a dup-ack count. * * The clear/increment are only logged * if you have BBverbose on. */ return; } memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = tsused; log.u_bbr.flex2 = thresh; log.u_bbr.flex3 = rsm->r_flags; log.u_bbr.flex4 = rsm->r_dupack; log.u_bbr.flex5 = rsm->r_start; log.u_bbr.flex6 = rsm->r_end; log.u_bbr.flex8 = mod; log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_SETTINGS_CHG, 0, 0, &log, false, &tv); } } static void rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = rack->rc_tp->t_srtt; log.u_bbr.flex2 = to; log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = slot; log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex7 = rack->rc_in_persist; log.u_bbr.flex8 = which; if (rack->rack_no_prr) log.u_bbr.pkts_out = 0; else log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift; log.u_bbr.lost = rack_rto_min; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERSTAR, 0, 0, &log, false, &tv); } } static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex8 = to_num; log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt; log.u_bbr.flex2 = rack->rc_rack_rtt; if (rsm == NULL) log.u_bbr.flex3 = 0; else log.u_bbr.flex3 = rsm->r_end - rsm->r_start; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_RTO, 0, 0, &log, false, &tv); } } static void rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *prev, struct rack_sendmap *rsm, struct rack_sendmap *next, int flag, uint32_t th_ack, int line) { if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex8 = flag; log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.cur_del_rate = (uint64_t)prev; log.u_bbr.delRate = (uint64_t)rsm; log.u_bbr.rttProp = (uint64_t)next; log.u_bbr.flex7 = 0; if (prev) { log.u_bbr.flex1 = prev->r_start; log.u_bbr.flex2 = prev->r_end; log.u_bbr.flex7 |= 0x4; } if (rsm) { log.u_bbr.flex3 = rsm->r_start; log.u_bbr.flex4 = rsm->r_end; log.u_bbr.flex7 |= 0x2; } if (next) { log.u_bbr.flex5 = next->r_start; log.u_bbr.flex6 = next->r_end; log.u_bbr.flex7 |= 0x1; } log.u_bbr.applimited = line; log.u_bbr.pkts_out = th_ack; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); if (rack->rack_no_prr) log.u_bbr.lost = 0; else log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_MAPCHG, 0, 0, &log, false, &tv); } } static void rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len, struct rack_sendmap *rsm, int conf) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = t; log.u_bbr.flex2 = len; log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt; log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest; log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest; log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt; log.u_bbr.flex7 = conf; log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot; log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt; log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); if (rsm) { log.u_bbr.pkt_epoch = rsm->r_start; log.u_bbr.lost = rsm->r_end; log.u_bbr.cwnd_gain = rsm->r_rtr_cnt; log.u_bbr.pacing_gain = rsm->r_flags; } else { /* Its a SYN */ log.u_bbr.pkt_epoch = rack->rc_tp->iss; log.u_bbr.lost = 0; log.u_bbr.cwnd_gain = 0; log.u_bbr.pacing_gain = 0; } /* Write out general bits of interest rrs here */ log.u_bbr.use_lt_bw = rack->rc_highly_buffered; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->forced_ack; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->in_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->app_limited_needs_set; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_gp_filled; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom; log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight; log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts; log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered; log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts; log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt; log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); log.u_bbr.bw_inuse <<= 32; if (rsm) log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); TCP_LOG_EVENTP(tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRRTT, 0, 0, &log, false, &tv); } } static void rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt) { /* * Log the rtt sample we are * applying to the srtt algorithm in * useconds. */ if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; /* Convert our ms to a microsecond */ memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = rtt; log.u_bbr.flex2 = rack->r_ctl.ack_count; log.u_bbr.flex3 = rack->r_ctl.sack_count; log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move; log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex7 = 1; log.u_bbr.flex8 = rack->sack_attack_disable; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; /* * We capture in delRate the upper 32 bits as * the confidence level we had declared, and the * lower 32 bits as the actual RTT using the arrival * timestamp. */ log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence; log.u_bbr.delRate <<= 32; log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt; /* Lets capture all the things that make up t_rtxcur */ log.u_bbr.applimited = rack_rto_min; log.u_bbr.epoch = rack_rto_max; log.u_bbr.lt_epoch = rack->r_ctl.timer_slop; log.u_bbr.lost = rack_rto_min; log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop); log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp); log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec; log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC; log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_RTT, 0, 0, &log, false, &tv); } } static void rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where) { if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; /* Convert our ms to a microsecond */ memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = rtt; log.u_bbr.flex2 = send_time; log.u_bbr.flex3 = ack_time; log.u_bbr.flex4 = where; log.u_bbr.flex7 = 2; log.u_bbr.timeStamp = tcp_get_usecs(&tv); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_RTT, 0, 0, &log, false, &tv); } } static inline void rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line) { if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = line; log.u_bbr.flex2 = tick; log.u_bbr.flex3 = tp->t_maxunacktime; log.u_bbr.flex4 = tp->t_acktime; log.u_bbr.flex8 = event; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_PROGRESS, 0, 0, &log, false, &tv); } } static void rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = slot; if (rack->rack_no_prr) log.u_bbr.flex2 = 0; else log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags); log.u_bbr.flex8 = rack->rc_in_persist; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRSND, 0, 0, &log, false, tv); } } static void rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = did_out; log.u_bbr.flex2 = nxt_pkt; log.u_bbr.flex3 = way_out; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex6 = nsegs; log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex7 = rack->rc_ack_can_sendout_data; /* Do we have ack-can-send set */ log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */ log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->r_wanted_output; /* Do we want output */ log.u_bbr.flex8 = rack->rc_in_persist; log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_DOSEG_DONE, 0, 0, &log, false, &tv); } } static void rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; uint32_t cts; memset(&log, 0, sizeof(log)); cts = tcp_get_usecs(&tv); log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = arg1; log.u_bbr.flex5 = arg2; log.u_bbr.flex6 = arg3; log.u_bbr.flex8 = frm; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.applimited = rack->r_ctl.rc_sacked; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(tp, NULL, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv); } } static void rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot, uint8_t hpts_calling, int reason, uint32_t cwnd_to_use) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = slot; log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = reason; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex7 = hpts_calling; log.u_bbr.flex8 = rack->rc_in_persist; log.u_bbr.lt_epoch = cwnd_to_use; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_JUSTRET, 0, tlen, &log, false, &tv); } } static void rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts, struct timeval *tv, uint32_t flags_on_entry) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = line; log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to; log.u_bbr.flex3 = flags_on_entry; log.u_bbr.flex4 = us_cts; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex7 = hpts_removed; log.u_bbr.flex8 = 1; log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags; log.u_bbr.timeStamp = us_cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERCANC, 0, 0, &log, false, tv); } } static void rack_log_alt_to_to_cancel(struct tcp_rack *rack, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint32_t flex4, uint32_t flex5, uint32_t flex6, uint16_t flex7, uint8_t mod) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; if (mod == 1) { /* No you can't use 1, its for the real to cancel */ return; } memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = flex1; log.u_bbr.flex2 = flex2; log.u_bbr.flex3 = flex3; log.u_bbr.flex4 = flex4; log.u_bbr.flex5 = flex5; log.u_bbr.flex6 = flex6; log.u_bbr.flex7 = flex7; log.u_bbr.flex8 = mod; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERCANC, 0, 0, &log, false, &tv); } } static void rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = timers; log.u_bbr.flex2 = ret; log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex5 = cts; if (rack->rack_no_prr) log.u_bbr.flex6 = 0; else log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TO_PROCESS, 0, 0, &log, false, &tv); } } static void rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = rack->r_ctl.rc_prr_out; log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs; if (rack->rack_no_prr) log.u_bbr.flex3 = 0; else log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered; log.u_bbr.flex5 = rack->r_ctl.rc_sacked; log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt; log.u_bbr.flex8 = frm; log.u_bbr.pkts_out = orig_cwnd; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRUPD, 0, 0, &log, false, &tv); } } #ifdef NETFLIX_EXP_DETECTION static void rack_log_sad(struct tcp_rack *rack, int event) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = rack->r_ctl.sack_count; log.u_bbr.flex2 = rack->r_ctl.ack_count; log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra; log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move; log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced; log.u_bbr.flex6 = tcp_sack_to_ack_thresh; log.u_bbr.pkts_out = tcp_sack_to_move_thresh; log.u_bbr.lt_epoch = (tcp_force_detection << 8); log.u_bbr.lt_epoch |= rack->do_detection; log.u_bbr.applimited = tcp_map_minimum; log.u_bbr.flex7 = rack->sack_attack_disable; log.u_bbr.flex8 = event; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.delivered = tcp_sad_decay_val; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_SAD_DETECTION, 0, 0, &log, false, &tv); } } #endif static void rack_counter_destroy(void) { int i; counter_u64_free(rack_fto_send); counter_u64_free(rack_fto_rsm_send); counter_u64_free(rack_nfto_resend); counter_u64_free(rack_hw_pace_init_fail); counter_u64_free(rack_hw_pace_lost); counter_u64_free(rack_non_fto_send); counter_u64_free(rack_extended_rfo); counter_u64_free(rack_ack_total); counter_u64_free(rack_express_sack); counter_u64_free(rack_sack_total); counter_u64_free(rack_move_none); counter_u64_free(rack_move_some); counter_u64_free(rack_sack_attacks_detected); counter_u64_free(rack_sack_attacks_reversed); counter_u64_free(rack_sack_used_next_merge); counter_u64_free(rack_sack_used_prev_merge); counter_u64_free(rack_badfr); counter_u64_free(rack_badfr_bytes); counter_u64_free(rack_rtm_prr_retran); counter_u64_free(rack_rtm_prr_newdata); counter_u64_free(rack_timestamp_mismatch); counter_u64_free(rack_find_high); counter_u64_free(rack_reorder_seen); counter_u64_free(rack_tlp_tot); counter_u64_free(rack_tlp_newdata); counter_u64_free(rack_tlp_retran); counter_u64_free(rack_tlp_retran_bytes); counter_u64_free(rack_tlp_retran_fail); counter_u64_free(rack_to_tot); counter_u64_free(rack_to_arm_rack); counter_u64_free(rack_to_arm_tlp); counter_u64_free(rack_calc_zero); counter_u64_free(rack_calc_nonzero); counter_u64_free(rack_paced_segments); counter_u64_free(rack_unpaced_segments); counter_u64_free(rack_saw_enobuf); counter_u64_free(rack_saw_enobuf_hw); counter_u64_free(rack_saw_enetunreach); counter_u64_free(rack_hot_alloc); counter_u64_free(rack_to_alloc); counter_u64_free(rack_to_alloc_hard); counter_u64_free(rack_to_alloc_emerg); counter_u64_free(rack_to_alloc_limited); counter_u64_free(rack_alloc_limited_conns); counter_u64_free(rack_split_limited); for (i = 0; i < MAX_NUM_OF_CNTS; i++) { counter_u64_free(rack_proc_comp_ack[i]); } counter_u64_free(rack_multi_single_eq); counter_u64_free(rack_proc_non_comp_ack); counter_u64_free(rack_sack_proc_all); counter_u64_free(rack_sack_proc_restart); counter_u64_free(rack_sack_proc_short); counter_u64_free(rack_enter_tlp_calc); counter_u64_free(rack_used_tlpmethod); counter_u64_free(rack_used_tlpmethod2); counter_u64_free(rack_sack_skipped_acked); counter_u64_free(rack_sack_splits); counter_u64_free(rack_progress_drops); counter_u64_free(rack_input_idle_reduces); counter_u64_free(rack_collapsed_win); counter_u64_free(rack_tlp_does_nada); counter_u64_free(rack_try_scwnd); counter_u64_free(rack_per_timer_hole); counter_u64_free(rack_large_ackcmp); counter_u64_free(rack_small_ackcmp); #ifdef INVARIANTS counter_u64_free(rack_adjust_map_bw); #endif COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE); COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE); } static struct rack_sendmap * rack_alloc(struct tcp_rack *rack) { struct rack_sendmap *rsm; /* * First get the top of the list it in * theory is the "hottest" rsm we have, * possibly just freed by ack processing. */ if (rack->rc_free_cnt > rack_free_cache) { rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); counter_u64_add(rack_hot_alloc, 1); rack->rc_free_cnt--; return (rsm); } /* * Once we get under our free cache we probably * no longer have a "hot" one available. Lets * get one from UMA. */ rsm = uma_zalloc(rack_zone, M_NOWAIT); if (rsm) { rack->r_ctl.rc_num_maps_alloced++; counter_u64_add(rack_to_alloc, 1); return (rsm); } /* * Dig in to our aux rsm's (the last two) since * UMA failed to get us one. */ if (rack->rc_free_cnt) { counter_u64_add(rack_to_alloc_emerg, 1); rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt--; return (rsm); } return (NULL); } static struct rack_sendmap * rack_alloc_full_limit(struct tcp_rack *rack) { if ((V_tcp_map_entries_limit > 0) && (rack->do_detection == 0) && (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { counter_u64_add(rack_to_alloc_limited, 1); if (!rack->alloc_limit_reported) { rack->alloc_limit_reported = 1; counter_u64_add(rack_alloc_limited_conns, 1); } return (NULL); } return (rack_alloc(rack)); } /* wrapper to allocate a sendmap entry, subject to a specific limit */ static struct rack_sendmap * rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type) { struct rack_sendmap *rsm; if (limit_type) { /* currently there is only one limit type */ if (V_tcp_map_split_limit > 0 && (rack->do_detection == 0) && rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { counter_u64_add(rack_split_limited, 1); if (!rack->alloc_limit_reported) { rack->alloc_limit_reported = 1; counter_u64_add(rack_alloc_limited_conns, 1); } return (NULL); } } /* allocate and mark in the limit type, if set */ rsm = rack_alloc(rack); if (rsm != NULL && limit_type) { rsm->r_limit_type = limit_type; rack->r_ctl.rc_num_split_allocs++; } return (rsm); } static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm) { if (rsm->r_flags & RACK_APP_LIMITED) { if (rack->r_ctl.rc_app_limited_cnt > 0) { rack->r_ctl.rc_app_limited_cnt--; } } if (rsm->r_limit_type) { /* currently there is only one limit type */ rack->r_ctl.rc_num_split_allocs--; } if (rsm == rack->r_ctl.rc_first_appl) { if (rack->r_ctl.rc_app_limited_cnt == 0) rack->r_ctl.rc_first_appl = NULL; else { /* Follow the next one out */ struct rack_sendmap fe; fe.r_start = rsm->r_nseq_appl; rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); } } if (rsm == rack->r_ctl.rc_resend) rack->r_ctl.rc_resend = NULL; if (rsm == rack->r_ctl.rc_rsm_at_retran) rack->r_ctl.rc_rsm_at_retran = NULL; if (rsm == rack->r_ctl.rc_end_appl) rack->r_ctl.rc_end_appl = NULL; if (rack->r_ctl.rc_tlpsend == rsm) rack->r_ctl.rc_tlpsend = NULL; if (rack->r_ctl.rc_sacklast == rsm) rack->r_ctl.rc_sacklast = NULL; memset(rsm, 0, sizeof(struct rack_sendmap)); TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt++; } static void rack_free_trim(struct tcp_rack *rack) { struct rack_sendmap *rsm; /* * Free up all the tail entries until * we get our list down to the limit. */ while (rack->rc_free_cnt > rack_free_cache) { rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head); TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt--; uma_zfree(rack_zone, rsm); } } static uint32_t rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack) { uint64_t srtt, bw, len, tim; uint32_t segsiz, def_len, minl; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); def_len = rack_def_data_window * segsiz; if (rack->rc_gp_filled == 0) { /* * We have no measurement (IW is in flight?) so * we can only guess using our data_window sysctl * value (usually 20MSS). */ return (def_len); } /* * Now we have a number of factors to consider. * * 1) We have a desired BDP which is usually * at least 2. * 2) We have a minimum number of rtt's usually 1 SRTT * but we allow it too to be more. * 3) We want to make sure a measurement last N useconds (if * we have set rack_min_measure_usec. * * We handle the first concern here by trying to create a data * window of max(rack_def_data_window, DesiredBDP). The * second concern we handle in not letting the measurement * window end normally until at least the required SRTT's * have gone by which is done further below in * rack_enough_for_measurement(). Finally the third concern * we also handle here by calculating how long that time * would take at the current BW and then return the * max of our first calculation and that length. Note * that if rack_min_measure_usec is 0, we don't deal * with concern 3. Also for both Concern 1 and 3 an * application limited period could end the measurement * earlier. * * So lets calculate the BDP with the "known" b/w using * the SRTT has our rtt and then multiply it by the * goal. */ bw = rack_get_bw(rack); srtt = (uint64_t)tp->t_srtt; len = bw * srtt; len /= (uint64_t)HPTS_USEC_IN_SEC; len *= max(1, rack_goal_bdp); /* Now we need to round up to the nearest MSS */ len = roundup(len, segsiz); if (rack_min_measure_usec) { /* Now calculate our min length for this b/w */ tim = rack_min_measure_usec; minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC; if (minl == 0) minl = 1; minl = roundup(minl, segsiz); if (len < minl) len = minl; } /* * Now if we have a very small window we want * to attempt to get the window that is * as small as possible. This happens on * low b/w connections and we don't want to * span huge numbers of rtt's between measurements. * * We basically include 2 over our "MIN window" so * that the measurement can be shortened (possibly) by * an ack'ed packet. */ if (len < def_len) return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz))); else return (max((uint32_t)len, def_len)); } static int rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality) { uint32_t tim, srtts, segsiz; /* * Has enough time passed for the GP measurement to be valid? */ if ((tp->snd_max == tp->snd_una) || (th_ack == tp->snd_max)){ /* All is acked */ *quality = RACK_QUALITY_ALLACKED; return (1); } if (SEQ_LT(th_ack, tp->gput_seq)) { /* Not enough bytes yet */ return (0); } segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); if (SEQ_LT(th_ack, tp->gput_ack) && ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) { /* Not enough bytes yet */ return (0); } if (rack->r_ctl.rc_first_appl && (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) { /* * We are up to the app limited send point * we have to measure irrespective of the time.. */ *quality = RACK_QUALITY_APPLIMITED; return (1); } /* Now what about time? */ srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts); tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts; if (tim >= srtts) { *quality = RACK_QUALITY_HIGH; return (1); } /* Nope not even a full SRTT has passed */ return (0); } static void rack_log_timely(struct tcp_rack *rack, uint32_t logged, uint64_t cur_bw, uint64_t low_bnd, uint64_t up_bnd, int line, uint8_t method) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = logged; log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt; log.u_bbr.flex2 <<= 4; log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt; log.u_bbr.flex2 <<= 4; log.u_bbr.flex2 |= rack->rc_gp_incr; log.u_bbr.flex2 <<= 4; log.u_bbr.flex2 |= rack->rc_gp_bwred; log.u_bbr.flex3 = rack->rc_gp_incr; log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss; log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca; log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec; log.u_bbr.flex7 = rack->rc_gp_bwred; log.u_bbr.flex8 = method; log.u_bbr.cur_del_rate = cur_bw; log.u_bbr.delRate = low_bnd; log.u_bbr.bw_inuse = up_bnd; log.u_bbr.rttProp = rack_get_bw(rack); log.u_bbr.pkt_epoch = line; log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt; log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt; log.u_bbr.cwnd_gain = rack->rc_dragged_bottom; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca; log.u_bbr.lost = rack->r_ctl.rc_loss_count; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_TIMELY_WORK, 0, 0, &log, false, &tv); } } static int rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult) { /* * Before we increase we need to know if * the estimate just made was less than * our pacing goal (i.e. (cur_bw * mult) > last_bw_est) * * If we already are pacing at a fast enough * rate to push us faster there is no sense of * increasing. * * We first caculate our actual pacing rate (ss or ca multipler * times our cur_bw). * * Then we take the last measured rate and multipy by our * maximum pacing overage to give us a max allowable rate. * * If our act_rate is smaller than our max_allowable rate * then we should increase. Else we should hold steady. * */ uint64_t act_rate, max_allow_rate; if (rack_timely_no_stopping) return (1); if ((cur_bw == 0) || (last_bw_est == 0)) { /* * Initial startup case or * everything is acked case. */ rack_log_timely(rack, mult, cur_bw, 0, 0, __LINE__, 9); return (1); } if (mult <= 100) { /* * We can always pace at or slightly above our rate. */ rack_log_timely(rack, mult, cur_bw, 0, 0, __LINE__, 9); return (1); } act_rate = cur_bw * (uint64_t)mult; act_rate /= 100; max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100); max_allow_rate /= 100; if (act_rate < max_allow_rate) { /* * Here the rate we are actually pacing at * is smaller than 10% above our last measurement. * This means we are pacing below what we would * like to try to achieve (plus some wiggle room). */ rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate, __LINE__, 9); return (1); } else { /* * Here we are already pacing at least rack_max_per_above(10%) * what we are getting back. This indicates most likely * that we are being limited (cwnd/rwnd/app) and can't * get any more b/w. There is no sense of trying to * raise up the pacing rate its not speeding us up * and we already are pacing faster than we are getting. */ rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate, __LINE__, 8); return (0); } } static void rack_validate_multipliers_at_or_above100(struct tcp_rack *rack) { /* * When we drag bottom, we want to assure * that no multiplier is below 1.0, if so * we want to restore it to at least that. */ if (rack->r_ctl.rack_per_of_gp_rec < 100) { /* This is unlikely we usually do not touch recovery */ rack->r_ctl.rack_per_of_gp_rec = 100; } if (rack->r_ctl.rack_per_of_gp_ca < 100) { rack->r_ctl.rack_per_of_gp_ca = 100; } if (rack->r_ctl.rack_per_of_gp_ss < 100) { rack->r_ctl.rack_per_of_gp_ss = 100; } } static void rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack) { if (rack->r_ctl.rack_per_of_gp_ca > 100) { rack->r_ctl.rack_per_of_gp_ca = 100; } if (rack->r_ctl.rack_per_of_gp_ss > 100) { rack->r_ctl.rack_per_of_gp_ss = 100; } } static void rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override) { int32_t calc, logged, plus; logged = 0; if (override) { /* * override is passed when we are * loosing b/w and making one last * gasp at trying to not loose out * to a new-reno flow. */ goto extra_boost; } /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */ if (rack->rc_gp_incr && ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) { /* * Reset and get 5 strokes more before the boost. Note * that the count is 0 based so we have to add one. */ extra_boost: plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST; rack->rc_gp_timely_inc_cnt = 0; } else plus = (uint32_t)rack_gp_increase_per; /* Must be at least 1% increase for true timely increases */ if ((plus < 1) && ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0))) plus = 1; if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0) && rack_bw_can_be_raised(rack, cur_bw, last_bw_est, rack->r_ctl.rack_per_of_gp_rec)) { /* We have been in recovery ding it too */ calc = rack->r_ctl.rack_per_of_gp_rec + plus; if (calc > 0xffff) calc = 0xffff; logged |= 1; rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc; if (rack_per_upper_bound_ss && (rack->rc_dragged_bottom == 0) && (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss)) rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss; } if (rack->rc_gp_saw_ca && (rack->rc_gp_saw_ss == 0) && rack_bw_can_be_raised(rack, cur_bw, last_bw_est, rack->r_ctl.rack_per_of_gp_ca)) { /* In CA */ calc = rack->r_ctl.rack_per_of_gp_ca + plus; if (calc > 0xffff) calc = 0xffff; logged |= 2; rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc; if (rack_per_upper_bound_ca && (rack->rc_dragged_bottom == 0) && (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca)) rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca; } if (rack->rc_gp_saw_ss && rack_bw_can_be_raised(rack, cur_bw, last_bw_est, rack->r_ctl.rack_per_of_gp_ss)) { /* In SS */ calc = rack->r_ctl.rack_per_of_gp_ss + plus; if (calc > 0xffff) calc = 0xffff; rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc; if (rack_per_upper_bound_ss && (rack->rc_dragged_bottom == 0) && (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss)) rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss; logged |= 4; } if (logged && (rack->rc_gp_incr == 0)){ /* Go into increment mode */ rack->rc_gp_incr = 1; rack->rc_gp_timely_inc_cnt = 0; } if (rack->rc_gp_incr && logged && (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) { rack->rc_gp_timely_inc_cnt++; } rack_log_timely(rack, logged, plus, 0, 0, __LINE__, 1); } static uint32_t rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff) { /* * norm_grad = rtt_diff / minrtt; * new_per = curper * (1 - B * norm_grad) * * B = rack_gp_decrease_per (default 10%) * rtt_dif = input var current rtt-diff * curper = input var current percentage * minrtt = from rack filter * */ uint64_t perf; perf = (((uint64_t)curper * ((uint64_t)1000000 - ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 * (((uint64_t)rtt_diff * (uint64_t)1000000)/ (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/ (uint64_t)1000000)) / (uint64_t)1000000); if (perf > curper) { /* TSNH */ perf = curper - 1; } return ((uint32_t)perf); } static uint32_t rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt) { /* * highrttthresh * result = curper * (1 - (B * ( 1 - ------ )) * gp_srtt * * B = rack_gp_decrease_per (default 10%) * highrttthresh = filter_min * rack_gp_rtt_maxmul */ uint64_t perf; uint32_t highrttthresh; highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul; perf = (((uint64_t)curper * ((uint64_t)1000000 - ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 - ((uint64_t)highrttthresh * (uint64_t)1000000) / (uint64_t)rtt)) / 100)) /(uint64_t)1000000); return (perf); } static void rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff) { uint64_t logvar, logvar2, logvar3; uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val; if (rack->rc_gp_incr) { /* Turn off increment counting */ rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; } ss_red = ca_red = rec_red = 0; logged = 0; /* Calculate the reduction value */ if (rtt_diff < 0) { rtt_diff *= -1; } /* Must be at least 1% reduction */ if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) { /* We have been in recovery ding it too */ if (timely_says == 2) { new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt); alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (alt < new_per) val = alt; else val = new_per; } else val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (rack->r_ctl.rack_per_of_gp_rec > val) { rec_red = (rack->r_ctl.rack_per_of_gp_rec - val); rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val; } else { rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound; rec_red = 0; } if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec) rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound; logged |= 1; } if (rack->rc_gp_saw_ss) { /* Sent in SS */ if (timely_says == 2) { new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt); alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (alt < new_per) val = alt; else val = new_per; } else val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff); if (rack->r_ctl.rack_per_of_gp_ss > new_per) { ss_red = rack->r_ctl.rack_per_of_gp_ss - val; rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val; } else { ss_red = new_per; rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound; logvar = new_per; logvar <<= 32; logvar |= alt; logvar2 = (uint32_t)rtt; logvar2 <<= 32; logvar2 |= (uint32_t)rtt_diff; logvar3 = rack_gp_rtt_maxmul; logvar3 <<= 32; logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); rack_log_timely(rack, timely_says, logvar2, logvar3, logvar, __LINE__, 10); } if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss) rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound; logged |= 4; } else if (rack->rc_gp_saw_ca) { /* Sent in CA */ if (timely_says == 2) { new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt); alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (alt < new_per) val = alt; else val = new_per; } else val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff); if (rack->r_ctl.rack_per_of_gp_ca > val) { ca_red = rack->r_ctl.rack_per_of_gp_ca - val; rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val; } else { rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound; ca_red = 0; logvar = new_per; logvar <<= 32; logvar |= alt; logvar2 = (uint32_t)rtt; logvar2 <<= 32; logvar2 |= (uint32_t)rtt_diff; logvar3 = rack_gp_rtt_maxmul; logvar3 <<= 32; logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); rack_log_timely(rack, timely_says, logvar2, logvar3, logvar, __LINE__, 10); } if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca) rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound; logged |= 2; } if (rack->rc_gp_timely_dec_cnt < 0x7) { rack->rc_gp_timely_dec_cnt++; if (rack_timely_dec_clear && (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear)) rack->rc_gp_timely_dec_cnt = 0; } logvar = ss_red; logvar <<= 32; logvar |= ca_red; rack_log_timely(rack, logged, rec_red, rack_per_lower_bound, logvar, __LINE__, 2); } static void rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts, uint32_t rtt, uint32_t line, uint8_t reas) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = line; log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts; log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts; log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss; log.u_bbr.flex5 = rtt; log.u_bbr.flex6 = rack->rc_highly_buffered; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->forced_ack; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->rc_gp_dyn_mul; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->in_probe_rtt; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->measure_saw_probe_rtt; log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt; log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca; log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec; log.u_bbr.flex8 = reas; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.delRate = rack_get_bw(rack); log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt; log.u_bbr.cur_del_rate <<= 32; log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt; log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered; log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt; log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt; log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts; log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight; log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); log.u_bbr.rttProp = us_cts; log.u_bbr.rttProp <<= 32; log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_RTT_SHRINKS, 0, 0, &log, false, &rack->r_ctl.act_rcv_time); } } static void rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt) { uint64_t bwdp; bwdp = rack_get_bw(rack); bwdp *= (uint64_t)rtt; bwdp /= (uint64_t)HPTS_USEC_IN_SEC; rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz); if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) { /* * A window protocol must be able to have 4 packets * outstanding as the floor in order to function * (especially considering delayed ack :D). */ rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs); } } static void rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts) { /** * ProbeRTT is a bit different in rack_pacing than in * BBR. It is like BBR in that it uses the lowering of * the RTT as a signal that we saw something new and * counts from there for how long between. But it is * different in that its quite simple. It does not * play with the cwnd and wait until we get down * to N segments outstanding and hold that for * 200ms. Instead it just sets the pacing reduction * rate to a set percentage (70 by default) and hold * that for a number of recent GP Srtt's. */ uint32_t segsiz; if (rack->rc_gp_dyn_mul == 0) return; if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) { /* We are idle */ return; } if ((rack->rc_tp->t_flags & TF_GPUTINPROG) && SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) { /* * Stop the goodput now, the idea here is * that future measurements with in_probe_rtt * won't register if they are not greater so * we want to get what info (if any) is available * now. */ rack_do_goodput_measurement(rack->rc_tp, rack, rack->rc_tp->snd_una, __LINE__, RACK_QUALITY_PROBERTT); } rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt; rack->r_ctl.rc_time_probertt_entered = us_cts; segsiz = min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs); rack->in_probe_rtt = 1; rack->measure_saw_probe_rtt = 1; rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_probertt_starts = 0; rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt; if (rack_probertt_use_min_rtt_entry) rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)); else rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt); rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_ENTERPROBE); } static void rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts) { struct rack_sendmap *rsm; uint32_t segsiz; segsiz = min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs); rack->in_probe_rtt = 0; if ((rack->rc_tp->t_flags & TF_GPUTINPROG) && SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) { /* * Stop the goodput now, the idea here is * that future measurements with in_probe_rtt * won't register if they are not greater so * we want to get what info (if any) is available * now. */ rack_do_goodput_measurement(rack->rc_tp, rack, rack->rc_tp->snd_una, __LINE__, RACK_QUALITY_PROBERTT); } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) { /* * We don't have enough data to make a measurement. * So lets just stop and start here after exiting * probe-rtt. We probably are not interested in * the results anyway. */ rack->rc_tp->t_flags &= ~TF_GPUTINPROG; } /* * Measurements through the current snd_max are going * to be limited by the slower pacing rate. * * We need to mark these as app-limited so we * don't collapse the b/w. */ rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) { if (rack->r_ctl.rc_app_limited_cnt == 0) rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm; else { /* * Go out to the end app limited and mark * this new one as next and move the end_appl up * to this guy. */ if (rack->r_ctl.rc_end_appl) rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start; rack->r_ctl.rc_end_appl = rsm; } rsm->r_flags |= RACK_APP_LIMITED; rack->r_ctl.rc_app_limited_cnt++; } /* * Now, we need to examine our pacing rate multipliers. * If its under 100%, we need to kick it back up to * 100%. We also don't let it be over our "max" above * the actual rate i.e. 100% + rack_clamp_atexit_prtt. * Note setting clamp_atexit_prtt to 0 has the effect * of setting CA/SS to 100% always at exit (which is * the default behavior). */ if (rack_probertt_clear_is) { rack->rc_gp_incr = 0; rack->rc_gp_bwred = 0; rack->rc_gp_timely_inc_cnt = 0; rack->rc_gp_timely_dec_cnt = 0; } /* Do we do any clamping at exit? */ if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) { rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp; rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp; } if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) { rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt; rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt; } /* * Lets set rtt_diff to 0, so that we will get a "boost" * after exiting. */ rack->r_ctl.rc_rtt_diff = 0; /* Clear all flags so we start fresh */ rack->rc_tp->t_bytes_acked = 0; rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND; /* * If configured to, set the cwnd and ssthresh to * our targets. */ if (rack_probe_rtt_sets_cwnd) { uint64_t ebdp; uint32_t setto; /* Set ssthresh so we get into CA once we hit our target */ if (rack_probertt_use_min_rtt_exit == 1) { /* Set to min rtt */ rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)); } else if (rack_probertt_use_min_rtt_exit == 2) { /* Set to current gp rtt */ rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt); } else if (rack_probertt_use_min_rtt_exit == 3) { /* Set to entry gp rtt */ rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_entry_gp_rtt); } else { uint64_t sum; uint32_t setval; sum = rack->r_ctl.rc_entry_gp_rtt; sum *= 10; sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt)); if (sum >= 20) { /* * A highly buffered path needs * cwnd space for timely to work. * Lets set things up as if * we are heading back here again. */ setval = rack->r_ctl.rc_entry_gp_rtt; } else if (sum >= 15) { /* * Lets take the smaller of the * two since we are just somewhat * buffered. */ setval = rack->r_ctl.rc_gp_srtt; if (setval > rack->r_ctl.rc_entry_gp_rtt) setval = rack->r_ctl.rc_entry_gp_rtt; } else { /* * Here we are not highly buffered * and should pick the min we can to * keep from causing loss. */ setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); } rack_set_prtt_target(rack, segsiz, setval); } if (rack_probe_rtt_sets_cwnd > 1) { /* There is a percentage here to boost */ ebdp = rack->r_ctl.rc_target_probertt_flight; ebdp *= rack_probe_rtt_sets_cwnd; ebdp /= 100; setto = rack->r_ctl.rc_target_probertt_flight + ebdp; } else setto = rack->r_ctl.rc_target_probertt_flight; rack->rc_tp->snd_cwnd = roundup(setto, segsiz); if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) { /* Enforce a min */ rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs; } /* If we set in the cwnd also set the ssthresh point so we are in CA */ rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1); } rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_EXITPROBE); /* Clear times last so log has all the info */ rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max; rack->r_ctl.rc_time_probertt_entered = us_cts; rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_of_last_probertt = us_cts; } static void rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts) { /* Check in on probe-rtt */ if (rack->rc_gp_filled == 0) { /* We do not do p-rtt unless we have gp measurements */ return; } if (rack->in_probe_rtt) { uint64_t no_overflow; uint32_t endtime, must_stay; if (rack->r_ctl.rc_went_idle_time && ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) { /* * We went idle during prtt, just exit now. */ rack_exit_probertt(rack, us_cts); } else if (rack_probe_rtt_safety_val && TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) && ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) { /* * Probe RTT safety value triggered! */ rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_SAFETY); rack_exit_probertt(rack, us_cts); } /* Calculate the max we will wait */ endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait); if (rack->rc_highly_buffered) endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp); /* Calculate the min we must wait */ must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain); if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) && TSTMP_LT(us_cts, endtime)) { uint32_t calc; /* Do we lower more? */ no_exit: if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered)) calc = us_cts - rack->r_ctl.rc_time_probertt_entered; else calc = 0; calc /= max(rack->r_ctl.rc_gp_srtt, 1); if (calc) { /* Maybe */ calc *= rack_per_of_gp_probertt_reduce; rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc; /* Limit it too */ if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh) rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh; } /* We must reach target or the time set */ return; } if (rack->r_ctl.rc_time_probertt_starts == 0) { if ((TSTMP_LT(us_cts, must_stay) && rack->rc_highly_buffered) || (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight)) { /* We are not past the must_stay time */ goto no_exit; } rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_REACHTARGET); rack->r_ctl.rc_time_probertt_starts = us_cts; if (rack->r_ctl.rc_time_probertt_starts == 0) rack->r_ctl.rc_time_probertt_starts = 1; /* Restore back to our rate we want to pace at in prtt */ rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt; } /* * Setup our end time, some number of gp_srtts plus 200ms. */ no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt * (uint64_t)rack_probertt_gpsrtt_cnt_mul); if (rack_probertt_gpsrtt_cnt_div) endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div); else endtime = 0; endtime += rack_min_probertt_hold; endtime += rack->r_ctl.rc_time_probertt_starts; if (TSTMP_GEQ(us_cts, endtime)) { /* yes, exit probertt */ rack_exit_probertt(rack, us_cts); } } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) { /* Go into probertt, its been too long since we went lower */ rack_enter_probertt(rack, us_cts); } } static void rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est, uint32_t rtt, int32_t rtt_diff) { uint64_t cur_bw, up_bnd, low_bnd, subfr; uint32_t losses; if ((rack->rc_gp_dyn_mul == 0) || (rack->use_fixed_rate) || (rack->in_probe_rtt) || (rack->rc_always_pace == 0)) { /* No dynamic GP multipler in play */ return; } losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start; cur_bw = rack_get_bw(rack); /* Calculate our up and down range */ up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up; up_bnd /= 100; up_bnd += rack->r_ctl.last_gp_comp_bw; subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down; subfr /= 100; low_bnd = rack->r_ctl.last_gp_comp_bw - subfr; if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) { /* * This is the case where our RTT is above * the max target and we have been configured * to just do timely no bonus up stuff in that case. * * There are two configurations, set to 1, and we * just do timely if we are over our max. If its * set above 1 then we slam the multipliers down * to 100 and then decrement per timely. */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); if (rack->r_ctl.rc_no_push_at_mrtt > 1) rack_validate_multipliers_at_or_below_100(rack); rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff); } else if ((last_bw_est < low_bnd) && !losses) { /* * We are decreasing this is a bit complicated this * means we are loosing ground. This could be * because another flow entered and we are competing * for b/w with it. This will push the RTT up which * makes timely unusable unless we want to get shoved * into a corner and just be backed off (the age * old problem with delay based CC). * * On the other hand if it was a route change we * would like to stay somewhat contained and not * blow out the buffers. */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); rack->r_ctl.last_gp_comp_bw = cur_bw; if (rack->rc_gp_bwred == 0) { /* Go into reduction counting */ rack->rc_gp_bwred = 1; rack->rc_gp_timely_dec_cnt = 0; } if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) || (timely_says == 0)) { /* * Push another time with a faster pacing * to try to gain back (we include override to * get a full raise factor). */ if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) || (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) || (timely_says == 0) || (rack_down_raise_thresh == 0)) { /* * Do an override up in b/w if we were * below the threshold or if the threshold * is zero we always do the raise. */ rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1); } else { /* Log it stays the same */ rack_log_timely(rack, 0, last_bw_est, low_bnd, 0, __LINE__, 11); } rack->rc_gp_timely_dec_cnt++; /* We are not incrementing really no-count */ rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; } else { /* * Lets just use the RTT * information and give up * pushing. */ goto use_timely; } } else if ((timely_says != 2) && !losses && (last_bw_est > up_bnd)) { /* * We are increasing b/w lets keep going, updating * our b/w and ignoring any timely input, unless * of course we are at our max raise (if there is one). */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); rack->r_ctl.last_gp_comp_bw = cur_bw; if (rack->rc_gp_saw_ss && rack_per_upper_bound_ss && (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) { /* * In cases where we can't go higher * we should just use timely. */ goto use_timely; } if (rack->rc_gp_saw_ca && rack_per_upper_bound_ca && (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) { /* * In cases where we can't go higher * we should just use timely. */ goto use_timely; } rack->rc_gp_bwred = 0; rack->rc_gp_timely_dec_cnt = 0; /* You get a set number of pushes if timely is trying to reduce */ if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) { rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0); } else { /* Log it stays the same */ rack_log_timely(rack, 0, last_bw_est, up_bnd, 0, __LINE__, 12); } return; } else { /* * We are staying between the lower and upper range bounds * so use timely to decide. */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); use_timely: if (timely_says) { rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) && !losses && (last_bw_est < low_bnd)) { /* We are loosing ground */ rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0); rack->rc_gp_timely_dec_cnt++; /* We are not incrementing really no-count */ rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; } else rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff); } else { rack->rc_gp_bwred = 0; rack->rc_gp_timely_dec_cnt = 0; rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0); } } } static int32_t rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt) { int32_t timely_says; uint64_t log_mult, log_rtt_a_diff; log_rtt_a_diff = rtt; log_rtt_a_diff <<= 32; log_rtt_a_diff |= (uint32_t)rtt_diff; if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul)) { /* Reduce the b/w multipler */ timely_says = 2; log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul; log_mult <<= 32; log_mult |= prev_rtt; rack_log_timely(rack, timely_says, log_mult, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 4); } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) + ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) / max(rack_gp_rtt_mindiv , 1)))) { /* Increase the b/w multipler */ log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) + ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) / max(rack_gp_rtt_mindiv , 1)); log_mult <<= 32; log_mult |= prev_rtt; timely_says = 0; rack_log_timely(rack, timely_says, log_mult , get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 5); } else { /* * Use a gradient to find it the timely gradient * is: * grad = rc_rtt_diff / min_rtt; * * anything below or equal to 0 will be * a increase indication. Anything above * zero is a decrease. Note we take care * of the actual gradient calculation * in the reduction (its not needed for * increase). */ log_mult = prev_rtt; if (rtt_diff <= 0) { /* * Rttdiff is less than zero, increase the * b/w multipler (its 0 or negative) */ timely_says = 0; rack_log_timely(rack, timely_says, log_mult, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6); } else { /* Reduce the b/w multipler */ timely_says = 1; rack_log_timely(rack, timely_says, log_mult, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7); } } return (timely_says); } static void rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, int line, uint8_t quality) { uint64_t tim, bytes_ps, ltim, stim, utim; uint32_t segsiz, bytes, reqbytes, us_cts; int32_t gput, new_rtt_diff, timely_says; uint64_t resid_bw, subpart = 0, addpart = 0, srtt; int did_add = 0; us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); if (TSTMP_GEQ(us_cts, tp->gput_ts)) tim = us_cts - tp->gput_ts; else tim = 0; if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts) stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts; else stim = 0; /* * Use the larger of the send time or ack time. This prevents us * from being influenced by ack artifacts to come up with too * high of measurement. Note that since we are spanning over many more * bytes in most of our measurements hopefully that is less likely to * occur. */ if (tim > stim) utim = max(tim, 1); else utim = max(stim, 1); /* Lets get a msec time ltim too for the old stuff */ ltim = max(1, (utim / HPTS_USEC_IN_MSEC)); gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim; reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz)); if ((tim == 0) && (stim == 0)) { /* * Invalid measurement time, maybe * all on one ack/one send? */ bytes = 0; bytes_ps = 0; rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, 0, 0, 0, 10, __LINE__, NULL, quality); goto skip_measurement; } if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) { /* We never made a us_rtt measurement? */ bytes = 0; bytes_ps = 0; rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, 0, 0, 0, 10, __LINE__, NULL, quality); goto skip_measurement; } /* * Calculate the maximum possible b/w this connection * could have. We base our calculation on the lowest * rtt we have seen during the measurement and the * largest rwnd the client has given us in that time. This * forms a BDP that is the maximum that we could ever * get to the client. Anything larger is not valid. * * I originally had code here that rejected measurements * where the time was less than 1/2 the latest us_rtt. * But after thinking on that I realized its wrong since * say you had a 150Mbps or even 1Gbps link, and you * were a long way away.. example I am in Europe (100ms rtt) * talking to my 1Gbps link in S.C. Now measuring say 150,000 * bytes my time would be 1.2ms, and yet my rtt would say * the measurement was invalid the time was < 50ms. The * same thing is true for 150Mb (8ms of time). * * A better way I realized is to look at what the maximum * the connection could possibly do. This is gated on * the lowest RTT we have seen and the highest rwnd. * We should in theory never exceed that, if we are * then something on the path is storing up packets * and then feeding them all at once to our endpoint * messing up our measurement. */ rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd; rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC; rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt; if (SEQ_LT(th_ack, tp->gput_seq)) { /* No measurement can be made */ bytes = 0; bytes_ps = 0; rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, 0, 0, 0, 10, __LINE__, NULL, quality); goto skip_measurement; } else bytes = (th_ack - tp->gput_seq); bytes_ps = (uint64_t)bytes; /* * Don't measure a b/w for pacing unless we have gotten at least * an initial windows worth of data in this measurement interval. * * Small numbers of bytes get badly influenced by delayed ack and * other artifacts. Note we take the initial window or our * defined minimum GP (defaulting to 10 which hopefully is the * IW). */ if (rack->rc_gp_filled == 0) { /* * The initial estimate is special. We * have blasted out an IW worth of packets * without a real valid ack ts results. We * then setup the app_limited_needs_set flag, * this should get the first ack in (probably 2 * MSS worth) to be recorded as the timestamp. * We thus allow a smaller number of bytes i.e. * IW - 2MSS. */ reqbytes -= (2 * segsiz); /* Also lets fill previous for our first measurement to be neutral */ rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt; } if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) { rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, rack->r_ctl.rc_app_limited_cnt, 0, 0, 10, __LINE__, NULL, quality); goto skip_measurement; } /* * We now need to calculate the Timely like status so * we can update (possibly) the b/w multipliers. */ new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt; if (rack->rc_gp_filled == 0) { /* No previous reading */ rack->r_ctl.rc_rtt_diff = new_rtt_diff; } else { if (rack->measure_saw_probe_rtt == 0) { /* * We don't want a probertt to be counted * since it will be negative incorrectly. We * expect to be reducing the RTT when we * pace at a slower rate. */ rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8); rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8); } } timely_says = rack_make_timely_judgement(rack, rack->r_ctl.rc_gp_srtt, rack->r_ctl.rc_rtt_diff, rack->r_ctl.rc_prev_gp_srtt ); bytes_ps *= HPTS_USEC_IN_SEC; bytes_ps /= utim; if (bytes_ps > rack->r_ctl.last_max_bw) { /* * Something is on path playing * since this b/w is not possible based * on our BDP (highest rwnd and lowest rtt * we saw in the measurement window). * * Another option here would be to * instead skip the measurement. */ rack_log_pacing_delay_calc(rack, bytes, reqbytes, bytes_ps, rack->r_ctl.last_max_bw, 0, 11, __LINE__, NULL, quality); bytes_ps = rack->r_ctl.last_max_bw; } /* We store gp for b/w in bytes per second */ if (rack->rc_gp_filled == 0) { /* Initial measurment */ if (bytes_ps) { rack->r_ctl.gp_bw = bytes_ps; rack->rc_gp_filled = 1; rack->r_ctl.num_measurements = 1; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else { rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, rack->r_ctl.rc_app_limited_cnt, 0, 0, 10, __LINE__, NULL, quality); } if (rack->rc_inp->inp_in_hpts && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* * Ok we can't trust the pacer in this case * where we transition from un-paced to paced. * Or for that matter when the burst mitigation * was making a wild guess and got it wrong. * Stop the pacer and clear up all the aggregate * delays etc. */ tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); rack->r_ctl.rc_hpts_flags = 0; rack->r_ctl.rc_last_output_to = 0; } did_add = 2; } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) { /* Still a small number run an average */ rack->r_ctl.gp_bw += bytes_ps; addpart = rack->r_ctl.num_measurements; rack->r_ctl.num_measurements++; if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) { /* We have collected enought to move forward */ rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements; } did_add = 3; } else { /* * We want to take 1/wma of the goodput and add in to 7/8th * of the old value weighted by the srtt. So if your measurement * period is say 2 SRTT's long you would get 1/4 as the * value, if it was like 1/2 SRTT then you would get 1/16th. * * But we must be careful not to take too much i.e. if the * srtt is say 20ms and the measurement is taken over * 400ms our weight would be 400/20 i.e. 20. On the * other hand if we get a measurement over 1ms with a * 10ms rtt we only want to take a much smaller portion. */ if (rack->r_ctl.num_measurements < 0xff) { rack->r_ctl.num_measurements++; } srtt = (uint64_t)tp->t_srtt; if (srtt == 0) { /* * Strange why did t_srtt go back to zero? */ if (rack->r_ctl.rc_rack_min_rtt) srtt = rack->r_ctl.rc_rack_min_rtt; else srtt = HPTS_USEC_IN_MSEC; } /* * XXXrrs: Note for reviewers, in playing with * dynamic pacing I discovered this GP calculation * as done originally leads to some undesired results. * Basically you can get longer measurements contributing * too much to the WMA. Thus I changed it if you are doing * dynamic adjustments to only do the aportioned adjustment * if we have a very small (time wise) measurement. Longer * measurements just get there weight (defaulting to 1/8) * add to the WMA. We may want to think about changing * this to always do that for both sides i.e. dynamic * and non-dynamic... but considering lots of folks * were playing with this I did not want to change the * calculation per.se. without your thoughts.. Lawerence? * Peter?? */ if (rack->rc_gp_dyn_mul == 0) { subpart = rack->r_ctl.gp_bw * utim; subpart /= (srtt * 8); if (subpart < (rack->r_ctl.gp_bw / 2)) { /* * The b/w update takes no more * away then 1/2 our running total * so factor it in. */ addpart = bytes_ps * utim; addpart /= (srtt * 8); } else { /* * Don't allow a single measurement * to account for more than 1/2 of the * WMA. This could happen on a retransmission * where utim becomes huge compared to * srtt (multiple retransmissions when using * the sending rate which factors in all the * transmissions from the first one). */ subpart = rack->r_ctl.gp_bw / 2; addpart = bytes_ps / 2; } resid_bw = rack->r_ctl.gp_bw - subpart; rack->r_ctl.gp_bw = resid_bw + addpart; did_add = 1; } else { if ((utim / srtt) <= 1) { /* * The b/w update was over a small period * of time. The idea here is to prevent a small * measurement time period from counting * too much. So we scale it based on the * time so it attributes less than 1/rack_wma_divisor * of its measurement. */ subpart = rack->r_ctl.gp_bw * utim; subpart /= (srtt * rack_wma_divisor); addpart = bytes_ps * utim; addpart /= (srtt * rack_wma_divisor); } else { /* * The scaled measurement was long * enough so lets just add in the * portion of the measurment i.e. 1/rack_wma_divisor */ subpart = rack->r_ctl.gp_bw / rack_wma_divisor; addpart = bytes_ps / rack_wma_divisor; } if ((rack->measure_saw_probe_rtt == 0) || (bytes_ps > rack->r_ctl.gp_bw)) { /* * For probe-rtt we only add it in * if its larger, all others we just * add in. */ did_add = 1; resid_bw = rack->r_ctl.gp_bw - subpart; rack->r_ctl.gp_bw = resid_bw + addpart; } } } if ((rack->gp_ready == 0) && (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) { /* We have enough measurements now */ rack->gp_ready = 1; rack_set_cc_pacing(rack); if (rack->defer_options) rack_apply_deferred_options(rack); } rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim, rack_get_bw(rack), 22, did_add, NULL, quality); /* We do not update any multipliers if we are in or have seen a probe-rtt */ if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set) rack_update_multiplier(rack, timely_says, bytes_ps, rack->r_ctl.rc_gp_srtt, rack->r_ctl.rc_rtt_diff); rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim, rack_get_bw(rack), 3, line, NULL, quality); /* reset the gp srtt and setup the new prev */ rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt; /* Record the lost count for the next measurement */ rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count; /* * We restart our diffs based on the gpsrtt in the * measurement window. */ rack->rc_gp_rtt_set = 0; rack->rc_gp_saw_rec = 0; rack->rc_gp_saw_ca = 0; rack->rc_gp_saw_ss = 0; rack->rc_dragged_bottom = 0; skip_measurement: #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, gput); /* * XXXLAS: This is a temporary hack, and should be * chained off VOI_TCP_GPUT when stats(9) grows an * API to deal with chained VOIs. */ if (tp->t_stats_gput_prev > 0) stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_GPUT_ND, ((gput - tp->t_stats_gput_prev) * 100) / tp->t_stats_gput_prev); #endif tp->t_flags &= ~TF_GPUTINPROG; tp->t_stats_gput_prev = gput; /* * Now are we app limited now and there is space from where we * were to where we want to go? * * We don't do the other case i.e. non-applimited here since * the next send will trigger us picking up the missing data. */ if (rack->r_ctl.rc_first_appl && TCPS_HAVEESTABLISHED(tp->t_state) && rack->r_ctl.rc_app_limited_cnt && (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) && ((rack->r_ctl.rc_first_appl->r_end - th_ack) > max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) { /* * Yep there is enough outstanding to make a measurement here. */ struct rack_sendmap *rsm, fe; rack->r_ctl.rc_gp_lowrtt = 0xffffffff; rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd; tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); rack->app_limited_needs_set = 0; tp->gput_seq = th_ack; if (rack->in_probe_rtt) rack->measure_saw_probe_rtt = 1; else if ((rack->measure_saw_probe_rtt) && (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit))) rack->measure_saw_probe_rtt = 0; if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) { /* There is a full window to gain info from */ tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack); } else { /* We can only measure up to the applimited point */ tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack); if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) { /* * We don't have enough to make a measurement. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq, 0, 0, 0, 6, __LINE__, NULL, quality); return; } } if (tp->t_state >= TCPS_FIN_WAIT_1) { /* * We will get no more data into the SB * this means we need to have the data available * before we start a measurement. */ if (sbavail(&tp->t_inpcb->inp_socket->so_snd) < (tp->gput_ack - tp->gput_seq)) { /* Nope not enough data. */ return; } } tp->t_flags |= TF_GPUTINPROG; /* * Now we need to find the timestamp of the send at tp->gput_seq * for the send based measurement. */ fe.r_start = tp->gput_seq; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm) { /* Ok send-based limit is set */ if (SEQ_LT(rsm->r_start, tp->gput_seq)) { /* * Move back to include the earlier part * so our ack time lines up right (this may * make an overlapping measurement but thats * ok). */ tp->gput_seq = rsm->r_start; } if (rsm->r_flags & RACK_ACKED) tp->gput_ts = (uint32_t)rsm->r_ack_arrival; else rack->app_limited_needs_set = 1; rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } else { /* * If we don't find the rsm due to some * send-limit set the current time, which * basically disables the send-limit. */ struct timeval tv; microuptime(&tv); rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv); } rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, (uint64_t)rsm, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL, quality); } } /* * CC wrapper hook functions */ static void rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery) { uint32_t prior_cwnd, acked; struct tcp_log_buffer *lgb = NULL; uint8_t labc_to_use, quality; INP_WLOCK_ASSERT(tp->t_inpcb); tp->ccv->nsegs = nsegs; acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una); if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) { uint32_t max; max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp); if (tp->ccv->bytes_this_ack > max) { tp->ccv->bytes_this_ack = max; } } #ifdef STATS stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd); #endif quality = RACK_QUALITY_NONE; if ((tp->t_flags & TF_GPUTINPROG) && rack_enough_for_measurement(tp, rack, th_ack, &quality)) { /* Measure the Goodput */ rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality); #ifdef NETFLIX_PEAKRATE if ((type == CC_ACK) && (tp->t_maxpeakrate)) { /* * We update t_peakrate_thr. This gives us roughly * one update per round trip time. Note * it will only be used if pace_always is off i.e * we don't do this for paced flows. */ rack_update_peakrate_thr(tp); } #endif } /* Which way our we limited, if not cwnd limited no advance in CA */ if (tp->snd_cwnd <= tp->snd_wnd) tp->ccv->flags |= CCF_CWND_LIMITED; else tp->ccv->flags &= ~CCF_CWND_LIMITED; if (tp->snd_cwnd > tp->snd_ssthresh) { tp->t_bytes_acked += min(tp->ccv->bytes_this_ack, nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp)); /* For the setting of a window past use the actual scwnd we are using */ if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) { tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use; tp->ccv->flags |= CCF_ABC_SENTAWND; } } else { tp->ccv->flags &= ~CCF_ABC_SENTAWND; tp->t_bytes_acked = 0; } prior_cwnd = tp->snd_cwnd; if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec || (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf))) labc_to_use = rack->rc_labc; else labc_to_use = rack_max_abc_post_recovery; if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = th_ack; log.u_bbr.flex2 = tp->ccv->flags; log.u_bbr.flex3 = tp->ccv->bytes_this_ack; log.u_bbr.flex4 = tp->ccv->nsegs; log.u_bbr.flex5 = labc_to_use; log.u_bbr.flex6 = prior_cwnd; log.u_bbr.flex7 = V_tcp_do_newsack; log.u_bbr.flex8 = 1; lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } if (CC_ALGO(tp)->ack_received != NULL) { /* XXXLAS: Find a way to live without this */ tp->ccv->curack = th_ack; tp->ccv->labc = labc_to_use; tp->ccv->flags |= CCF_USE_LOCAL_ABC; CC_ALGO(tp)->ack_received(tp->ccv, type); } if (lgb) { lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd; } if (rack->r_must_retran) { if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) { /* * We now are beyond the rxt point so lets disable * the flag. */ rack->r_ctl.rc_out_at_rto = 0; rack->r_must_retran = 0; } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) { /* * Only decrement the rc_out_at_rto if the cwnd advances * at least a whole segment. Otherwise next time the peer * acks, we won't be able to send this generaly happens * when we are in Congestion Avoidance. */ if (acked <= rack->r_ctl.rc_out_at_rto){ rack->r_ctl.rc_out_at_rto -= acked; } else { rack->r_ctl.rc_out_at_rto = 0; } } } #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use); #endif if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) { rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use; } #ifdef NETFLIX_PEAKRATE /* we enforce max peak rate if it is set and we are not pacing */ if ((rack->rc_always_pace == 0) && tp->t_peakrate_thr && (tp->snd_cwnd > tp->t_peakrate_thr)) { tp->snd_cwnd = tp->t_peakrate_thr; } #endif } static void tcp_rack_partialack(struct tcpcb *tp) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); /* * If we are doing PRR and have enough * room to send we are pacing and prr * is disabled we will want to see if we * can send data (by setting r_wanted_output to * true). */ if ((rack->r_ctl.rc_prr_sndcnt > 0) || rack->rack_no_prr) rack->r_wanted_output = 1; } static void rack_post_recovery(struct tcpcb *tp, uint32_t th_ack) { struct tcp_rack *rack; uint32_t orig_cwnd; orig_cwnd = tp->snd_cwnd; INP_WLOCK_ASSERT(tp->t_inpcb); rack = (struct tcp_rack *)tp->t_fb_ptr; /* only alert CC if we alerted when we entered */ if (CC_ALGO(tp)->post_recovery != NULL) { tp->ccv->curack = th_ack; CC_ALGO(tp)->post_recovery(tp->ccv); if (tp->snd_cwnd < tp->snd_ssthresh) { /* * Rack has burst control and pacing * so lets not set this any lower than * snd_ssthresh per RFC-6582 (option 2). */ tp->snd_cwnd = tp->snd_ssthresh; } } if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = th_ack; log.u_bbr.flex2 = tp->ccv->flags; log.u_bbr.flex3 = tp->ccv->bytes_this_ack; log.u_bbr.flex4 = tp->ccv->nsegs; log.u_bbr.flex5 = V_tcp_abc_l_var; log.u_bbr.flex6 = orig_cwnd; log.u_bbr.flex7 = V_tcp_do_newsack; log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex8 = 2; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } if ((rack->rack_no_prr == 0) && (rack->no_prr_addback == 0) && (rack->r_ctl.rc_prr_sndcnt > 0)) { /* * Suck the next prr cnt back into cwnd, but * only do that if we are not application limited. */ if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { /* * We are allowed to add back to the cwnd the amount we did * not get out if: * a) no_prr_addback is off. * b) we are not app limited * c) we are doing prr * * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none). */ tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax), rack->r_ctl.rc_prr_sndcnt); } rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 1, 0); } rack_log_to_prr(rack, 14, orig_cwnd); tp->snd_recover = tp->snd_una; EXIT_RECOVERY(tp->t_flags); } static void rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack) { struct tcp_rack *rack; uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd; INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); #endif if (IN_RECOVERY(tp->t_flags) == 0) { in_rec_at_entry = 0; ssthresh_enter = tp->snd_ssthresh; cwnd_enter = tp->snd_cwnd; } else in_rec_at_entry = 1; rack = (struct tcp_rack *)tp->t_fb_ptr; switch (type) { case CC_NDUPACK: tp->t_flags &= ~TF_WASFRECOVERY; tp->t_flags &= ~TF_WASCRECOVERY; if (!IN_FASTRECOVERY(tp->t_flags)) { rack->r_ctl.rc_prr_delivered = 0; rack->r_ctl.rc_prr_out = 0; if (rack->rack_no_prr == 0) { rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 2, in_rec_at_entry); } rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una; tp->snd_recover = tp->snd_max; if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; } break; case CC_ECN: if (!IN_CONGRECOVERY(tp->t_flags) || /* * Allow ECN reaction on ACK to CWR, if * that data segment was also CE marked. */ SEQ_GEQ(ack, tp->snd_recover)) { EXIT_CONGRECOVERY(tp->t_flags); KMOD_TCPSTAT_INC(tcps_ecn_rcwnd); tp->snd_recover = tp->snd_max + 1; if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; } break; case CC_RTO: tp->t_dupacks = 0; tp->t_bytes_acked = 0; EXIT_RECOVERY(tp->t_flags); tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 / ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp); orig_cwnd = tp->snd_cwnd; tp->snd_cwnd = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 16, orig_cwnd); if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; break; case CC_RTO_ERR: KMOD_TCPSTAT_INC(tcps_sndrexmitbad); /* RTO was unnecessary, so reset everything. */ tp->snd_cwnd = tp->snd_cwnd_prev; tp->snd_ssthresh = tp->snd_ssthresh_prev; tp->snd_recover = tp->snd_recover_prev; if (tp->t_flags & TF_WASFRECOVERY) { ENTER_FASTRECOVERY(tp->t_flags); tp->t_flags &= ~TF_WASFRECOVERY; } if (tp->t_flags & TF_WASCRECOVERY) { ENTER_CONGRECOVERY(tp->t_flags); tp->t_flags &= ~TF_WASCRECOVERY; } tp->snd_nxt = tp->snd_max; tp->t_badrxtwin = 0; break; } if ((CC_ALGO(tp)->cong_signal != NULL) && (type != CC_RTO)){ tp->ccv->curack = ack; CC_ALGO(tp)->cong_signal(tp->ccv, type); } if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) { rack_log_to_prr(rack, 15, cwnd_enter); rack->r_ctl.dsack_byte_cnt = 0; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.rc_cwnd_at_erec = cwnd_enter; rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter; rack->r_ent_rec_ns = 1; } } static inline void rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp) { uint32_t i_cwnd; INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef NETFLIX_STATS KMOD_TCPSTAT_INC(tcps_idle_restarts); if (tp->t_state == TCPS_ESTABLISHED) KMOD_TCPSTAT_INC(tcps_idle_estrestarts); #endif if (CC_ALGO(tp)->after_idle != NULL) CC_ALGO(tp)->after_idle(tp->ccv); if (tp->snd_cwnd == 1) i_cwnd = tp->t_maxseg; /* SYN(-ACK) lost */ else i_cwnd = rc_init_window(rack); /* * Being idle is no differnt than the initial window. If the cc * clamps it down below the initial window raise it to the initial * window. */ if (tp->snd_cwnd < i_cwnd) { tp->snd_cwnd = i_cwnd; } } /* * Indicate whether this ack should be delayed. We can delay the ack if * following conditions are met: * - There is no delayed ack timer in progress. * - Our last ack wasn't a 0-sized window. We never want to delay * the ack that opens up a 0-sized window. * - LRO wasn't used for this segment. We make sure by checking that the * segment size is not larger than the MSS. * - Delayed acks are enabled or this is a half-synchronized T/TCP * connection. */ #define DELAY_ACK(tp, tlen) \ (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ ((tp->t_flags & TF_DELACK) == 0) && \ (tlen <= tp->t_maxseg) && \ (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) static struct rack_sendmap * rack_find_lowest_rsm(struct tcp_rack *rack) { struct rack_sendmap *rsm; /* * Walk the time-order transmitted list looking for an rsm that is * not acked. This will be the one that was sent the longest time * ago that is still outstanding. */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { if (rsm->r_flags & RACK_ACKED) { continue; } goto finish; } finish: return (rsm); } static struct rack_sendmap * rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm) { struct rack_sendmap *prsm; /* * Walk the sequence order list backward until we hit and arrive at * the highest seq not acked. In theory when this is called it * should be the last segment (which it was not). */ counter_u64_add(rack_find_high, 1); prsm = rsm; RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) { if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) { continue; } return (prsm); } return (NULL); } static uint32_t rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts) { int32_t lro; uint32_t thresh; /* * lro is the flag we use to determine if we have seen reordering. * If it gets set we have seen reordering. The reorder logic either * works in one of two ways: * * If reorder-fade is configured, then we track the last time we saw * re-ordering occur. If we reach the point where enough time as * passed we no longer consider reordering has occuring. * * Or if reorder-face is 0, then once we see reordering we consider * the connection to alway be subject to reordering and just set lro * to 1. * * In the end if lro is non-zero we add the extra time for * reordering in. */ if (srtt == 0) srtt = 1; if (rack->r_ctl.rc_reorder_ts) { if (rack->r_ctl.rc_reorder_fade) { if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) { lro = cts - rack->r_ctl.rc_reorder_ts; if (lro == 0) { /* * No time as passed since the last * reorder, mark it as reordering. */ lro = 1; } } else { /* Negative time? */ lro = 0; } if (lro > rack->r_ctl.rc_reorder_fade) { /* Turn off reordering seen too */ rack->r_ctl.rc_reorder_ts = 0; lro = 0; } } else { /* Reodering does not fade */ lro = 1; } } else { lro = 0; } thresh = srtt + rack->r_ctl.rc_pkt_delay; if (lro) { /* It must be set, if not you get 1/4 rtt */ if (rack->r_ctl.rc_reorder_shift) thresh += (srtt >> rack->r_ctl.rc_reorder_shift); else thresh += (srtt >> 2); } else { thresh += 1; } /* We don't let the rack timeout be above a RTO */ if (thresh > rack->rc_tp->t_rxtcur) { thresh = rack->rc_tp->t_rxtcur; } /* And we don't want it above the RTO max either */ if (thresh > rack_rto_max) { thresh = rack_rto_max; } return (thresh); } static uint32_t rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t srtt) { struct rack_sendmap *prsm; uint32_t thresh, len; int segsiz; if (srtt == 0) srtt = 1; if (rack->r_ctl.rc_tlp_threshold) thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold); else thresh = (srtt * 2); /* Get the previous sent packet, if any */ segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); counter_u64_add(rack_enter_tlp_calc, 1); len = rsm->r_end - rsm->r_start; if (rack->rack_tlp_threshold_use == TLP_USE_ID) { /* Exactly like the ID */ if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) { uint32_t alt_thresh; /* * Compensate for delayed-ack with the d-ack time. */ counter_u64_add(rack_used_tlpmethod, 1); alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) { /* 2.1 behavior */ prsm = TAILQ_PREV(rsm, rack_head, r_tnext); if (prsm && (len <= segsiz)) { /* * Two packets outstanding, thresh should be (2*srtt) + * possible inter-packet delay (if any). */ uint32_t inter_gap = 0; int idx, nidx; counter_u64_add(rack_used_tlpmethod, 1); idx = rsm->r_rtr_cnt - 1; nidx = prsm->r_rtr_cnt - 1; if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) { /* Yes it was sent later (or at the same time) */ inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; } thresh += inter_gap; } else if (len <= segsiz) { /* * Possibly compensate for delayed-ack. */ uint32_t alt_thresh; counter_u64_add(rack_used_tlpmethod2, 1); alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) { /* 2.2 behavior */ if (len <= segsiz) { uint32_t alt_thresh; /* * Compensate for delayed-ack with the d-ack time. */ counter_u64_add(rack_used_tlpmethod, 1); alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } /* Not above an RTO */ if (thresh > tp->t_rxtcur) { thresh = tp->t_rxtcur; } /* Not above a RTO max */ if (thresh > rack_rto_max) { thresh = rack_rto_max; } /* Apply user supplied min TLP */ if (thresh < rack_tlp_min) { thresh = rack_tlp_min; } return (thresh); } static uint32_t rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack) { /* * We want the rack_rtt which is the * last rtt we measured. However if that * does not exist we fallback to the srtt (which * we probably will never do) and then as a last * resort we use RACK_INITIAL_RTO if no srtt is * yet set. */ if (rack->rc_rack_rtt) return (rack->rc_rack_rtt); else if (tp->t_srtt == 0) return (RACK_INITIAL_RTO); return (tp->t_srtt); } static struct rack_sendmap * rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused) { /* * Check to see that we don't need to fall into recovery. We will * need to do so if our oldest transmit is past the time we should * have had an ack. */ struct tcp_rack *rack; struct rack_sendmap *rsm; int32_t idx; uint32_t srtt, thresh; rack = (struct tcp_rack *)tp->t_fb_ptr; if (RB_EMPTY(&rack->r_ctl.rc_mtree)) { return (NULL); } rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm == NULL) return (NULL); if (rsm->r_flags & RACK_ACKED) { rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) return (NULL); } idx = rsm->r_rtr_cnt - 1; srtt = rack_grab_rtt(tp, rack); thresh = rack_calc_thresh_rack(rack, srtt, tsused); if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) { return (NULL); } if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) { return (NULL); } /* Ok if we reach here we are over-due and this guy can be sent */ if (IN_RECOVERY(tp->t_flags) == 0) { /* * For the one that enters us into recovery record undo * info. */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; } rack_cong_signal(tp, CC_NDUPACK, tp->snd_una); return (rsm); } static uint32_t rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack) { int32_t t; int32_t tt; uint32_t ret_val; t = (tp->t_srtt + (tp->t_rttvar << 2)); RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop); if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; ret_val = (uint32_t)tt; return (ret_val); } static uint32_t rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack) { /* * Start the FR timer, we do this based on getting the first one in * the rc_tmap. Note that if its NULL we must stop the timer. in all * events we need to stop the running timer (if its running) before * starting the new one. */ uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; uint32_t srtt_cur; int32_t idx; int32_t is_tlp_timer = 0; struct rack_sendmap *rsm; if (rack->t_timers_stopped) { /* All timers have been stopped none are to run */ return (0); } if (rack->rc_in_persist) { /* We can't start any timer in persists */ return (rack_get_persists_timer_val(tp, rack)); } rack->rc_on_min_to = 0; if ((tp->t_state < TCPS_ESTABLISHED) || ((tp->t_flags & TF_SACK_PERMIT) == 0)) { goto activate_rxt; } rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if ((rsm == NULL) || sup_rack) { /* Nothing on the send map or no rack */ activate_rxt: time_since_sent = 0; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm) { /* * Should we discount the RTX timer any? * * We want to discount it the smallest amount. * If a timer (Rack/TLP or RXT) has gone off more * recently thats the discount we want to use (now - timer time). * If the retransmit of the oldest packet was more recent then * we want to use that (now - oldest-packet-last_transmit_time). * */ idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx]))) tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time; else tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx]; if (TSTMP_GT(cts, tstmp_touse)) time_since_sent = cts - tstmp_touse; } if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; to = tp->t_rxtcur; if (to > time_since_sent) to -= time_since_sent; else to = rack->r_ctl.rc_min_to; if (to == 0) to = 1; /* Special case for KEEPINIT */ if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) && (TP_KEEPINIT(tp) != 0) && rsm) { /* * We have to put a ceiling on the rxt timer * of the keep-init timeout. */ uint32_t max_time, red; max_time = TICKS_2_USEC(TP_KEEPINIT(tp)); if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) { red = (cts - (uint32_t)rsm->r_tim_lastsent[0]); if (red < max_time) max_time -= red; else max_time = 1; } /* Reduce timeout to the keep value if needed */ if (max_time < to) to = max_time; } return (to); } return (0); } if (rsm->r_flags & RACK_ACKED) { rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) { /* No lowest? */ goto activate_rxt; } } if (rack->sack_attack_disable) { /* * We don't want to do * any TLP's if you are an attacker. * Though if you are doing what * is expected you may still have * SACK-PASSED marks. */ goto activate_rxt; } /* Convert from ms to usecs */ if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & RACK_HAS_FIN)) { /* * We don't start a rack timer if all we have is a * FIN outstanding. */ goto activate_rxt; } if ((rack->use_rack_rr == 0) && (IN_FASTRECOVERY(tp->t_flags)) && (rack->rack_no_prr == 0) && (rack->r_ctl.rc_prr_sndcnt < ctf_fixed_maxseg(tp))) { /* * We are not cheating, in recovery and * not enough ack's to yet get our next * retransmission out. * * Note that classified attackers do not * get to use the rack-cheat. */ goto activate_tlp; } srtt = rack_grab_rtt(tp, rack); thresh = rack_calc_thresh_rack(rack, srtt, cts); idx = rsm->r_rtr_cnt - 1; exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh; if (SEQ_GEQ(exp, cts)) { to = exp - cts; if (to < rack->r_ctl.rc_min_to) { to = rack->r_ctl.rc_min_to; if (rack->r_rr_config == 3) rack->rc_on_min_to = 1; } } else { to = rack->r_ctl.rc_min_to; if (rack->r_rr_config == 3) rack->rc_on_min_to = 1; } } else { /* Ok we need to do a TLP not RACK */ activate_tlp: if ((rack->rc_tlp_in_progress != 0) && (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) { /* * The previous send was a TLP and we have sent * N TLP's without sending new data. */ goto activate_rxt; } rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); if (rsm == NULL) { /* We found no rsm to TLP with. */ goto activate_rxt; } if (rsm->r_flags & RACK_HAS_FIN) { /* If its a FIN we dont do TLP */ rsm = NULL; goto activate_rxt; } idx = rsm->r_rtr_cnt - 1; time_since_sent = 0; if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx]; else tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(cts, tstmp_touse)) time_since_sent = cts - tstmp_touse; is_tlp_timer = 1; if (tp->t_srtt) { if ((rack->rc_srtt_measure_made == 0) && (tp->t_srtt == 1)) { /* * If another stack as run and set srtt to 1, * then the srtt was 0, so lets use the initial. */ srtt = RACK_INITIAL_RTO; } else { srtt_cur = tp->t_srtt; srtt = srtt_cur; } } else srtt = RACK_INITIAL_RTO; /* * If the SRTT is not keeping up and the * rack RTT has spiked we want to use * the last RTT not the smoothed one. */ if (rack_tlp_use_greater && tp->t_srtt && (srtt < rack_grab_rtt(tp, rack))) { srtt = rack_grab_rtt(tp, rack); } thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt); if (thresh > time_since_sent) { to = thresh - time_since_sent; } else { to = rack->r_ctl.rc_min_to; rack_log_alt_to_to_cancel(rack, thresh, /* flex1 */ time_since_sent, /* flex2 */ tstmp_touse, /* flex3 */ rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */ (uint32_t)rsm->r_tim_lastsent[idx], srtt, idx, 99); } if (to < rack_tlp_min) { to = rack_tlp_min; } if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) { /* * If the TLP time works out to larger than the max * RTO lets not do TLP.. just RTO. */ goto activate_rxt; } } if (is_tlp_timer == 0) { rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; } else { rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; } if (to == 0) to = 1; return (to); } static void rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (rack->rc_in_persist == 0) { if (tp->t_flags & TF_GPUTINPROG) { /* * Stop the goodput now, the calling of the * measurement function clears the flag. */ rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__, RACK_QUALITY_PERSIST); } #ifdef NETFLIX_SHARED_CWND if (rack->r_ctl.rc_scw) { tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 1; } #endif rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_timer_cancel(tp, rack, cts, __LINE__); tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->rc_in_persist = 1; } } static void rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (rack->rc_inp->inp_in_hpts) { tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); rack->r_ctl.rc_hpts_flags = 0; } #ifdef NETFLIX_SHARED_CWND if (rack->r_ctl.rc_scw) { tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 0; } #endif if (rack->rc_gp_dyn_mul && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* * Do we count this as if a probe-rtt just * finished? */ uint32_t time_idle, idle_min; time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time; idle_min = rack_min_probertt_hold; if (rack_probertt_gpsrtt_cnt_div) { uint64_t extra; extra = (uint64_t)rack->r_ctl.rc_gp_srtt * (uint64_t)rack_probertt_gpsrtt_cnt_mul; extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div; idle_min += (uint32_t)extra; } if (time_idle >= idle_min) { /* Yes, we count it as a probe-rtt. */ uint32_t us_cts; us_cts = tcp_get_usecs(NULL); if (rack->in_probe_rtt == 0) { rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts; } else { rack_exit_probertt(rack, us_cts); } } } rack->rc_in_persist = 0; rack->r_ctl.rc_went_idle_time = 0; tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_late = 0; rack->r_ctl.rc_agg_early = 0; } static void rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts, struct hpts_diag *diag, struct timeval *tv) { if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = diag->p_nxt_slot; log.u_bbr.flex2 = diag->p_cur_slot; log.u_bbr.flex3 = diag->slot_req; log.u_bbr.flex4 = diag->inp_hptsslot; log.u_bbr.flex5 = diag->slot_remaining; log.u_bbr.flex6 = diag->need_new_to; log.u_bbr.flex7 = diag->p_hpts_active; log.u_bbr.flex8 = diag->p_on_min_sleep; /* Hijack other fields as needed */ log.u_bbr.epoch = diag->have_slept; log.u_bbr.lt_epoch = diag->yet_to_sleep; log.u_bbr.pkts_out = diag->co_ret; log.u_bbr.applimited = diag->hpts_sleep_time; log.u_bbr.delivered = diag->p_prev_slot; - log.u_bbr.inflight = diag->p_runningtick; - log.u_bbr.bw_inuse = diag->wheel_tick; + log.u_bbr.inflight = diag->p_runningslot; + log.u_bbr.bw_inuse = diag->wheel_slot; log.u_bbr.rttProp = diag->wheel_cts; log.u_bbr.timeStamp = cts; - log.u_bbr.delRate = diag->maxticks; + log.u_bbr.delRate = diag->maxslots; log.u_bbr.cur_del_rate = diag->p_curtick; log.u_bbr.cur_del_rate <<= 32; log.u_bbr.cur_del_rate |= diag->p_lasttick; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSDIAG, 0, 0, &log, false, tv); } } static void rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type) { if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = sb->sb_flags; log.u_bbr.flex2 = len; log.u_bbr.flex3 = sb->sb_state; log.u_bbr.flex8 = type; log.u_bbr.timeStamp = tcp_get_usecs(&tv); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_SB_WAKE, 0, len, &log, false, &tv); } } static void rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts, int32_t slot, uint32_t tot_len_this_send, int sup_rack) { struct hpts_diag diag; struct inpcb *inp; struct timeval tv; uint32_t delayed_ack = 0; uint32_t hpts_timeout; uint32_t entry_slot = slot; uint8_t stopped; uint32_t left = 0; uint32_t us_cts; inp = tp->t_inpcb; if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { return; } if (inp->inp_in_hpts) { /* Already on the pacer */ return; } stopped = rack->rc_tmr_stopped; if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) { left = rack->r_ctl.rc_timer_exp - cts; } rack->r_ctl.rc_timer_exp = 0; rack->r_ctl.rc_hpts_flags = 0; us_cts = tcp_get_usecs(&tv); /* Now early/late accounting */ rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0); if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) { /* * We have a early carry over set, * we can always add more time so we * can always make this compensation. * * Note if ack's are allowed to wake us do not * penalize the next timer for being awoke * by an ack aka the rc_agg_early (non-paced mode). */ slot += rack->r_ctl.rc_agg_early; rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; } if (rack->r_late) { /* * This is harder, we can * compensate some but it * really depends on what * the current pacing time is. */ if (rack->r_ctl.rc_agg_delayed >= slot) { /* * We can't compensate for it all. * And we have to have some time * on the clock. We always have a min * 10 slots (10 x 10 i.e. 100 usecs). */ - if (slot <= HPTS_TICKS_PER_USEC) { + if (slot <= HPTS_TICKS_PER_SLOT) { /* We gain delay */ - rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_USEC - slot); - slot = HPTS_TICKS_PER_USEC; + rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot); + slot = HPTS_TICKS_PER_SLOT; } else { /* We take off some */ - rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_USEC); - slot = HPTS_TICKS_PER_USEC; + rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT); + slot = HPTS_TICKS_PER_SLOT; } } else { slot -= rack->r_ctl.rc_agg_delayed; rack->r_ctl.rc_agg_delayed = 0; /* Make sure we have 100 useconds at minimum */ - if (slot < HPTS_TICKS_PER_USEC) { - rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_USEC - slot; - slot = HPTS_TICKS_PER_USEC; + if (slot < HPTS_TICKS_PER_SLOT) { + rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot; + slot = HPTS_TICKS_PER_SLOT; } if (rack->r_ctl.rc_agg_delayed == 0) rack->r_late = 0; } } if (slot) { /* We are pacing too */ rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; } hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack); #ifdef NETFLIX_EXP_DETECTION if (rack->sack_attack_disable && (slot < tcp_sad_pacing_interval)) { /* * We have a potential attacker on * the line. We have possibly some * (or now) pacing time set. We want to * slow down the processing of sacks by some * amount (if it is an attacker). Set the default * slot for attackers in place (unless the orginal * interval is longer). Its stored in * micro-seconds, so lets convert to msecs. */ slot = tcp_sad_pacing_interval; } #endif if (tp->t_flags & TF_DELACK) { delayed_ack = TICKS_2_USEC(tcp_delacktime); rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK; } if (delayed_ack && ((hpts_timeout == 0) || (delayed_ack < hpts_timeout))) hpts_timeout = delayed_ack; else rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; /* * If no timers are going to run and we will fall off the hptsi * wheel, we resort to a keep-alive timer if its configured. */ if ((hpts_timeout == 0) && (slot == 0)) { if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) { /* * Ok we have no timer (persists, rack, tlp, rxt or * del-ack), we don't have segments being paced. So * all that is left is the keepalive timer. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { /* Get the established keep-alive time */ hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); } else { /* * Get the initial setup keep-alive time, * note that this is probably not going to * happen, since rack will be running a rxt timer * if a SYN of some sort is outstanding. It is * actually handled in rack_timeout_rxt(). */ hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); } rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; if (rack->in_probe_rtt) { /* * We want to instead not wake up a long time from * now but to wake up about the time we would * exit probe-rtt and initiate a keep-alive ack. * This will get us out of probe-rtt and update * our min-rtt. */ hpts_timeout = rack_min_probertt_hold; } } } if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { /* * RACK, TLP, persists and RXT timers all are restartable * based on actions input .. i.e we received a packet (ack * or sack) and that changes things (rw, or snd_una etc). * Thus we can restart them with a new value. For * keep-alive, delayed_ack we keep track of what was left * and restart the timer with a smaller value. */ if (left < hpts_timeout) hpts_timeout = left; } if (hpts_timeout) { /* * Hack alert for now we can't time-out over 2,147,483 * seconds (a bit more than 596 hours), which is probably ok * :). */ if (hpts_timeout > 0x7ffffffe) hpts_timeout = 0x7ffffffe; rack->r_ctl.rc_timer_exp = cts + hpts_timeout; } rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0); if ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0) && (hpts_timeout < slot) && (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { /* * We have no good estimate yet for the * old clunky burst mitigation or the * real pacing. And the tlp or rxt is smaller * than the pacing calculation. Lets not * pace that long since we know the calculation * so far is not accurate. */ slot = hpts_timeout; } rack->r_ctl.last_pacing_time = slot; /** * Turn off all the flags for queuing by default. The * flags have important meanings to what happens when * LRO interacts with the transport. Most likely (by default now) * mbuf_queueing and ack compression are on. So the transport * has a couple of flags that control what happens (if those * are not on then these flags won't have any effect since it * won't go through the queuing LRO path). * * INP_MBUF_QUEUE_READY - This flags says that I am busy * pacing output, so don't disturb. But * it also means LRO can wake me if there * is a SACK arrival. * * INP_DONT_SACK_QUEUE - This flag is used in conjunction * with the above flag (QUEUE_READY) and * when present it says don't even wake me * if a SACK arrives. * * The idea behind these flags is that if we are pacing we * set the MBUF_QUEUE_READY and only get woken up if * a SACK arrives (which could change things) or if * our pacing timer expires. If, however, we have a rack * timer running, then we don't even want a sack to wake * us since the rack timer has to expire before we can send. * * Other cases should usually have none of the flags set * so LRO can call into us. */ inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY); if (slot) { rack->r_ctl.rc_last_output_to = us_cts + slot; /* * A pacing timer (slot) is being set, in * such a case we cannot send (we are blocked by * the timer). So lets tell LRO that it should not * wake us unless there is a SACK. Note this only * will be effective if mbuf queueing is on or * compressed acks are being processed. */ inp->inp_flags2 |= INP_MBUF_QUEUE_READY; /* * But wait if we have a Rack timer running * even a SACK should not disturb us (with * the exception of r_rr_config 3). */ if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && (rack->r_rr_config != 3)) inp->inp_flags2 |= INP_DONT_SACK_QUEUE; if (rack->rc_ack_can_sendout_data) { /* * Ahh but wait, this is that special case * where the pacing timer can be disturbed * backout the changes (used for non-paced * burst limiting). */ inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY); } if ((rack->use_rack_rr) && (rack->r_rr_config < 2) && ((hpts_timeout) && (hpts_timeout < slot))) { /* * Arrange for the hpts to kick back in after the * t-o if the t-o does not cause a send. */ (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout), __LINE__, &diag); rack_log_hpts_diag(rack, us_cts, &diag, &tv); rack_log_to_start(rack, cts, hpts_timeout, slot, 0); } else { (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot), __LINE__, &diag); rack_log_hpts_diag(rack, us_cts, &diag, &tv); rack_log_to_start(rack, cts, hpts_timeout, slot, 1); } } else if (hpts_timeout) { /* * With respect to inp_flags2 here, lets let any new acks wake * us up here. Since we are not pacing (no pacing timer), output * can happen so we should let it. If its a Rack timer, then any inbound * packet probably won't change the sending (we will be blocked) * but it may change the prr stats so letting it in (the set defaults * at the start of this block) are good enough. */ (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout), __LINE__, &diag); rack_log_hpts_diag(rack, us_cts, &diag, &tv); rack_log_to_start(rack, cts, hpts_timeout, slot, 0); } else { /* No timer starting */ #ifdef INVARIANTS if (SEQ_GT(tp->snd_max, tp->snd_una)) { panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?", tp, rack, tot_len_this_send, cts, slot, hpts_timeout); } #endif } rack->rc_tmr_stopped = 0; if (slot) rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv); } /* * RACK Timer, here we simply do logging and house keeping. * the normal rack_output() function will call the * appropriate thing to check if we need to do a RACK retransmit. * We return 1, saying don't proceed with rack_output only * when all timers have been stopped (destroyed PCB?). */ static int rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { /* * This timer simply provides an internal trigger to send out data. * The check_recovery_mode call will see if there are needed * retransmissions, if so we will enter fast-recovery. The output * call may or may not do the same thing depending on sysctl * settings. */ struct rack_sendmap *rsm; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } counter_u64_add(rack_to_tot, 1); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); rack->rc_on_min_to = 0; rsm = rack_check_recovery_mode(tp, cts); rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm); if (rsm) { rack->r_ctl.rc_resend = rsm; rack->r_timer_override = 1; if (rack->use_rack_rr) { /* * Don't accumulate extra pacing delay * we are allowing the rack timer to * over-ride pacing i.e. rrr takes precedence * if the pacing interval is longer than the rrr * time (in other words we get the min pacing * time versus rrr pacing time). */ rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; } } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; if (rsm == NULL) { /* restart a timer and return 1 */ rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); return (1); } return (0); } static void rack_adjust_orig_mlen(struct rack_sendmap *rsm) { if (rsm->m->m_len > rsm->orig_m_len) { /* * Mbuf grew, caused by sbcompress, our offset does * not change. */ rsm->orig_m_len = rsm->m->m_len; } else if (rsm->m->m_len < rsm->orig_m_len) { /* * Mbuf shrank, trimmed off the top by an ack, our * offset changes. */ rsm->soff -= (rsm->orig_m_len - rsm->m->m_len); rsm->orig_m_len = rsm->m->m_len; } } static void rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm) { struct mbuf *m; uint32_t soff; if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) { /* Fix up the orig_m_len and possibly the mbuf offset */ rack_adjust_orig_mlen(src_rsm); } m = src_rsm->m; soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start); while (soff >= m->m_len) { /* Move out past this mbuf */ soff -= m->m_len; m = m->m_next; KASSERT((m != NULL), ("rsm:%p nrsm:%p hit at soff:%u null m", src_rsm, rsm, soff)); } rsm->m = m; rsm->soff = soff; rsm->orig_m_len = m->m_len; } static __inline void rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm, struct rack_sendmap *rsm, uint32_t start) { int idx; nrsm->r_start = start; nrsm->r_end = rsm->r_end; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm->r_flags = rsm->r_flags; nrsm->r_dupack = rsm->r_dupack; nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed; nrsm->r_rtr_bytes = 0; rsm->r_end = nrsm->r_start; nrsm->r_just_ret = rsm->r_just_ret; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } /* Now if we have SYN flag we keep it on the left edge */ if (nrsm->r_flags & RACK_HAS_SYN) nrsm->r_flags &= ~RACK_HAS_SYN; /* Now if we have a FIN flag we keep it on the right edge */ if (rsm->r_flags & RACK_HAS_FIN) rsm->r_flags &= ~RACK_HAS_FIN; /* Push bit must go to the right edge as well */ if (rsm->r_flags & RACK_HAD_PUSH) rsm->r_flags &= ~RACK_HAD_PUSH; /* Clone over the state of the hw_tls flag */ nrsm->r_hw_tls = rsm->r_hw_tls; /* * Now we need to find nrsm's new location in the mbuf chain * we basically calculate a new offset, which is soff + * how much is left in original rsm. Then we walk out the mbuf * chain to find the righ postion, it may be the same mbuf * or maybe not. */ KASSERT(((rsm->m != NULL) || (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))), ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack)); if (rsm->m) rack_setup_offset_for_rsm(rsm, nrsm); } static struct rack_sendmap * rack_merge_rsm(struct tcp_rack *rack, struct rack_sendmap *l_rsm, struct rack_sendmap *r_rsm) { /* * We are merging two ack'd RSM's, * the l_rsm is on the left (lower seq * values) and the r_rsm is on the right * (higher seq value). The simplest way * to merge these is to move the right * one into the left. I don't think there * is any reason we need to try to find * the oldest (or last oldest retransmitted). */ struct rack_sendmap *rm; rack_log_map_chg(rack->rc_tp, rack, NULL, l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__); l_rsm->r_end = r_rsm->r_end; if (l_rsm->r_dupack < r_rsm->r_dupack) l_rsm->r_dupack = r_rsm->r_dupack; if (r_rsm->r_rtr_bytes) l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; if (r_rsm->r_in_tmap) { /* This really should not happen */ TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext); r_rsm->r_in_tmap = 0; } /* Now the flags */ if (r_rsm->r_flags & RACK_HAS_FIN) l_rsm->r_flags |= RACK_HAS_FIN; if (r_rsm->r_flags & RACK_TLP) l_rsm->r_flags |= RACK_TLP; if (r_rsm->r_flags & RACK_RWND_COLLAPSED) l_rsm->r_flags |= RACK_RWND_COLLAPSED; if ((r_rsm->r_flags & RACK_APP_LIMITED) && ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) { /* * If both are app-limited then let the * free lower the count. If right is app * limited and left is not, transfer. */ l_rsm->r_flags |= RACK_APP_LIMITED; r_rsm->r_flags &= ~RACK_APP_LIMITED; if (r_rsm == rack->r_ctl.rc_first_appl) rack->r_ctl.rc_first_appl = l_rsm; } rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm); #ifdef INVARIANTS if (rm != r_rsm) { panic("removing head in rack:%p rsm:%p rm:%p", rack, r_rsm, rm); } #endif if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { /* Transfer the split limit to the map we free */ r_rsm->r_limit_type = l_rsm->r_limit_type; l_rsm->r_limit_type = 0; } rack_free(rack, r_rsm); return (l_rsm); } /* * TLP Timer, here we simply setup what segment we want to * have the TLP expire on, the normal rack_output() will then * send it out. * * We return 1, saying don't proceed with rack_output only * when all timers have been stopped (destroyed PCB?). */ static int rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp) { /* * Tail Loss Probe. */ struct rack_sendmap *rsm = NULL; struct rack_sendmap *insret; struct socket *so; uint32_t amm; uint32_t out, avail; int collapsed_win = 0; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); return (1); } /* * A TLP timer has expired. We have been idle for 2 rtts. So we now * need to figure out how to force a full MSS segment out. */ rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL); rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; counter_u64_add(rack_tlp_tot, 1); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); so = tp->t_inpcb->inp_socket; avail = sbavail(&so->so_snd); out = tp->snd_max - tp->snd_una; if (out > tp->snd_wnd) { /* special case, we need a retransmission */ collapsed_win = 1; goto need_retran; } if ((tp->t_flags & TF_GPUTINPROG) && (rack->r_ctl.rc_tlp_cnt_out == 1)) { /* * If this is the second in a row * TLP and we are doing a measurement * its time to abandon the measurement. * Something is likely broken on * the clients network and measuring a * broken network does us no good. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); } /* * Check our send oldest always settings, and if * there is an oldest to send jump to the need_retran. */ if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0)) goto need_retran; if (avail > out) { /* New data is available */ amm = avail - out; if (amm > ctf_fixed_maxseg(tp)) { amm = ctf_fixed_maxseg(tp); if ((amm + out) > tp->snd_wnd) { /* We are rwnd limited */ goto need_retran; } } else if (amm < ctf_fixed_maxseg(tp)) { /* not enough to fill a MTU */ goto need_retran; } if (IN_FASTRECOVERY(tp->t_flags)) { /* Unlikely */ if (rack->rack_no_prr == 0) { if (out + amm <= tp->snd_wnd) { rack->r_ctl.rc_prr_sndcnt = amm; rack->r_ctl.rc_tlp_new_data = amm; rack_log_to_prr(rack, 4, 0); } } else goto need_retran; } else { /* Set the send-new override */ if (out + amm <= tp->snd_wnd) rack->r_ctl.rc_tlp_new_data = amm; else goto need_retran; } rack->r_ctl.rc_tlpsend = NULL; counter_u64_add(rack_tlp_newdata, 1); goto send; } need_retran: /* * Ok we need to arrange the last un-acked segment to be re-sent, or * optionally the first un-acked segment. */ if (collapsed_win == 0) { if (rack_always_send_oldest) rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); else { rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) { rsm = rack_find_high_nonack(rack, rsm); } } if (rsm == NULL) { counter_u64_add(rack_tlp_does_nada, 1); #ifdef TCP_BLACKBOX tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); #endif goto out; } } else { /* * We must find the last segment * that was acceptable by the client. */ RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) { /* Found one */ break; } } if (rsm == NULL) { /* None? if so send the first */ rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm == NULL) { counter_u64_add(rack_tlp_does_nada, 1); #ifdef TCP_BLACKBOX tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); #endif goto out; } } } if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) { /* * We need to split this the last segment in two. */ struct rack_sendmap *nrsm; nrsm = rack_alloc_full_limit(rack); if (nrsm == NULL) { /* * No memory to split, we will just exit and punt * off to the RXT timer. */ counter_u64_add(rack_tlp_does_nada, 1); goto out; } rack_clone_rsm(rack, nrsm, rsm, (rsm->r_end - ctf_fixed_maxseg(tp))); rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); rsm = nrsm; } rack->r_ctl.rc_tlpsend = rsm; send: /* Make sure output path knows we are doing a TLP */ *doing_tlp = 1; rack->r_timer_override = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); out: rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); } /* * Delayed ack Timer, here we simply need to setup the * ACK_NOW flag and remove the DELACK flag. From there * the output routine will send the ack out. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL); tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; KMOD_TCPSTAT_INC(tcps_delack); rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; return (0); } /* * Persists timer, here we simply send the * same thing as a keepalive will. * the one byte send. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { struct tcptemp *t_template; struct inpcb *inp; int32_t retval = 1; inp = tp->t_inpcb; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (rack->rc_in_persist == 0) return (0); if (ctf_progress_timeout_check(tp, false)) { tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(inp, ETIMEDOUT); return (1); } KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); /* * Persistence timer into zero window. Force a byte to be output, if * possible. */ KMOD_TCPSTAT_INC(tcps_persisttimeo); /* * Hack: if the peer is dead/unreachable, we do not time out if the * window is closed. After a full backoff, drop the connection if * the idle time (no responses to probes) reaches the maximum * backoff that we would use if retransmitting. */ if (tp->t_rxtshift == TCP_MAXRXTSHIFT && (ticks - tp->t_rcvtime >= tcp_maxpersistidle || TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) { KMOD_TCPSTAT_INC(tcps_persistdrop); retval = 1; tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); goto out; } if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) && tp->snd_una == tp->snd_max) rack_exit_persist(tp, rack, cts); rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; /* * If the user has closed the socket then drop a persisting * connection after a much reduced timeout. */ if (tp->t_state > TCPS_CLOSE_WAIT && (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { retval = 1; KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); goto out; } t_template = tcpip_maketemplate(rack->rc_inp); if (t_template) { /* only set it if we were answered */ if (rack->forced_ack == 0) { rack->forced_ack = 1; rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL); } tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); /* This sends an ack */ if (tp->t_flags & TF_DELACK) tp->t_flags &= ~TF_DELACK; free(t_template, M_TEMP); } if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; out: rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL); rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); return (retval); } /* * If a keepalive goes off, we had no other timers * happening. We always return 1 here since this * routine either drops the connection or sends * out a segment with respond. */ static int rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { struct tcptemp *t_template; struct inpcb *inp; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; inp = tp->t_inpcb; rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL); /* * Keep-alive timer went off; send something or drop connection if * idle for too long. */ KMOD_TCPSTAT_INC(tcps_keeptimeo); if (tp->t_state < TCPS_ESTABLISHED) goto dropit; if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && tp->t_state <= TCPS_CLOSING) { if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) goto dropit; /* * Send a packet designed to force a response if the peer is * up and reachable: either an ACK if the connection is * still alive, or an RST if the peer has closed the * connection due to timeout or reboot. Using sequence * number tp->snd_una-1 causes the transmitted zero-length * segment to lie outside the receive window; by the * protocol spec, this requires the correspondent TCP to * respond. */ KMOD_TCPSTAT_INC(tcps_keepprobe); t_template = tcpip_maketemplate(inp); if (t_template) { if (rack->forced_ack == 0) { rack->forced_ack = 1; rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL); } tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); free(t_template, M_TEMP); } } rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); return (1); dropit: KMOD_TCPSTAT_INC(tcps_keepdrops); tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); return (1); } /* * Retransmit helper function, clear up all the ack * flags and take care of important book keeping. */ static void rack_remxt_tmr(struct tcpcb *tp) { /* * The retransmit timer went off, all sack'd blocks must be * un-acked. */ struct rack_sendmap *rsm, *trsm = NULL; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__); rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); /* * Ideally we would like to be able to * mark SACK-PASS on anything not acked here. * * However, if we do that we would burst out * all that data 1ms apart. This would be unwise, * so for now we will just let the normal rxt timer * and tlp timer take care of it. * * Also we really need to stick them back in sequence * order. This way we send in the proper order and any * sacks that come floating in will "re-ack" the data. * To do this we zap the tmap with an INIT and then * walk through and place every rsm in the RB tree * back in its seq ordered place. */ TAILQ_INIT(&rack->r_ctl.rc_tmap); RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* We must re-add it back to the tlist */ if (trsm == NULL) { TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); } else { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext); } rsm->r_in_tmap = 1; trsm = rsm; if (rsm->r_flags & RACK_ACKED) rsm->r_flags |= RACK_WAS_ACKED; rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS); } /* Clear the count (we just un-acked them) */ rack->r_ctl.rc_last_timeout_snduna = tp->snd_una; rack->r_ctl.rc_sacked = 0; rack->r_ctl.rc_sacklast = NULL; rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; rack->r_late = 0; /* Clear the tlp rtx mark */ rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rack->r_ctl.rc_resend != NULL) rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT; rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 6, 0); rack->r_timer_override = 1; if ((((tp->t_flags & TF_SACK_PERMIT) == 0) #ifdef NETFLIX_EXP_DETECTION || (rack->sack_attack_disable != 0) #endif ) && ((tp->t_flags & TF_SENTFIN) == 0)) { /* * For non-sack customers new data * needs to go out as retransmits until * we retransmit up to snd_max. */ rack->r_must_retran = 1; rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); } rack->r_ctl.rc_snd_max_at_rto = tp->snd_max; } static void rack_convert_rtts(struct tcpcb *tp) { if (tp->t_srtt > 1) { uint32_t val, frac; val = tp->t_srtt >> TCP_RTT_SHIFT; frac = tp->t_srtt & 0x1f; tp->t_srtt = TICKS_2_USEC(val); /* * frac is the fractional part of the srtt (if any) * but its in ticks and every bit represents * 1/32nd of a hz. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE); } else { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE)); } tp->t_srtt += frac; } } if (tp->t_rttvar) { uint32_t val, frac; val = tp->t_rttvar >> TCP_RTTVAR_SHIFT; frac = tp->t_rttvar & 0x1f; tp->t_rttvar = TICKS_2_USEC(val); /* * frac is the fractional part of the srtt (if any) * but its in ticks and every bit represents * 1/32nd of a hz. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE); } else { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE)); } tp->t_rttvar += frac; } } tp->t_rxtcur = RACK_REXMTVAL(tp); if (TCPS_HAVEESTABLISHED(tp->t_state)) { tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop); } if (tp->t_rxtcur > rack_rto_max) { tp->t_rxtcur = rack_rto_max; } } static void rack_cc_conn_init(struct tcpcb *tp) { struct tcp_rack *rack; uint32_t srtt; rack = (struct tcp_rack *)tp->t_fb_ptr; srtt = tp->t_srtt; cc_conn_init(tp); /* * Now convert to rack's internal format, * if required. */ if ((srtt == 0) && (tp->t_srtt != 0)) rack_convert_rtts(tp); /* * We want a chance to stay in slowstart as * we create a connection. TCP spec says that * initially ssthresh is infinite. For our * purposes that is the snd_wnd. */ if (tp->snd_ssthresh < tp->snd_wnd) { tp->snd_ssthresh = tp->snd_wnd; } /* * We also want to assure a IW worth of * data can get inflight. */ if (rc_init_window(rack) < tp->snd_cwnd) tp->snd_cwnd = rc_init_window(rack); } /* * Re-transmit timeout! If we drop the PCB we will return 1, otherwise * we will setup to retransmit the lowest seq number outstanding. */ static int rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { int32_t rexmt; struct inpcb *inp; int32_t retval = 0; bool isipv6; inp = tp->t_inpcb; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if ((tp->t_flags & TF_GPUTINPROG) && (tp->t_rxtshift)) { /* * We have had a second timeout * measurements on successive rxt's are not profitable. * It is unlikely to be of any use (the network is * broken or the client went away). */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); } if (ctf_progress_timeout_check(tp, false)) { tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(inp, ETIMEDOUT); return (1); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; if (IN_FASTRECOVERY(tp->t_flags)) tp->t_flags |= TF_WASFRECOVERY; else tp->t_flags &= ~TF_WASFRECOVERY; if (IN_CONGRECOVERY(tp->t_flags)) tp->t_flags |= TF_WASCRECOVERY; else tp->t_flags &= ~TF_WASCRECOVERY; if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_una == tp->snd_max)) { /* Nothing outstanding .. nothing to do */ return (0); } /* * Rack can only run one timer at a time, so we cannot * run a KEEPINIT (gating SYN sending) and a retransmit * timer for the SYN. So if we are in a front state and * have a KEEPINIT timer we need to check the first transmit * against now to see if we have exceeded the KEEPINIT time * (if one is set). */ if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) && (TP_KEEPINIT(tp) != 0)) { struct rack_sendmap *rsm; rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm) { /* Ok we have something outstanding to test keepinit with */ if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) && ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) { /* We have exceeded the KEEPINIT time */ tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); goto drop_it; } } } /* * Retransmission timer went off. Message has not been acked within * retransmit interval. Back off to a longer retransmit interval * and retransmit one segment. */ rack_remxt_tmr(tp); if ((rack->r_ctl.rc_resend == NULL) || ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) { /* * If the rwnd collapsed on * the one we are retransmitting * it does not count against the * rxt count. */ tp->t_rxtshift++; } if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); drop_it: tp->t_rxtshift = TCP_MAXRXTSHIFT; KMOD_TCPSTAT_INC(tcps_timeoutdrop); retval = 1; tcp_set_inp_to_drop(rack->rc_inp, (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); goto out; } if (tp->t_state == TCPS_SYN_SENT) { /* * If the SYN was retransmitted, indicate CWND to be limited * to 1 segment in cc_conn_init(). */ tp->snd_cwnd = 1; } else if (tp->t_rxtshift == 1) { /* * first retransmit; record ssthresh and cwnd so they can be * recovered if this turns out to be a "bad" retransmit. A * retransmit is considered "bad" if an ACK for this segment * is received within RTT/2 interval; the assumption here is * that the ACK was already in flight. See "On Estimating * End-to-End Network Path Properties" by Allman and Paxson * for more details. */ tp->snd_cwnd_prev = tp->snd_cwnd; tp->snd_ssthresh_prev = tp->snd_ssthresh; tp->snd_recover_prev = tp->snd_recover; tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2); tp->t_flags |= TF_PREVVALID; } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0) tp->t_flags &= ~TF_PREVVALID; KMOD_TCPSTAT_INC(tcps_rexmttimeo); if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift]; else rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift]; RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt, max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop); /* * We enter the path for PLMTUD if connection is established or, if * connection is FIN_WAIT_1 status, reason for the last is that if * amount of data we send is very small, we could send it in couple * of packets and process straight to FIN. In that case we won't * catch ESTABLISHED state. */ #ifdef INET6 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false; #else isipv6 = false; #endif if (((V_tcp_pmtud_blackhole_detect == 1) || (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && ((tp->t_state == TCPS_ESTABLISHED) || (tp->t_state == TCPS_FIN_WAIT_1))) { /* * Idea here is that at each stage of mtu probe (usually, * 1448 -> 1188 -> 524) should be given 2 chances to recover * before further clamping down. 'tp->t_rxtshift % 2 == 0' * should take care of that. */ if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && tp->t_rxtshift % 2 == 0)) { /* * Enter Path MTU Black-hole Detection mechanism: - * Disable Path MTU Discovery (IP "DF" bit). - * Reduce MTU to lower value than what we negotiated * with peer. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { /* Record that we may have found a black hole. */ tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; /* Keep track of previous MSS. */ tp->t_pmtud_saved_maxseg = tp->t_maxseg; } /* * Reduce the MSS to blackhole value or to the * default in an attempt to retransmit. */ #ifdef INET6 if (isipv6 && tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else if (isipv6) { /* Use the default MSS. */ tp->t_maxseg = V_tcp_v6mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else { /* Use the default MSS. */ tp->t_maxseg = V_tcp_mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif } else { /* * If further retransmissions are still unsuccessful * with a lowered MTU, maybe this isn't a blackhole * and we restore the previous MSS and blackhole * detection flags. The limit '6' is determined by * giving each probe stage (1448, 1188, 524) 2 * chances to recover. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && (tp->t_rxtshift >= 6)) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; tp->t_maxseg = tp->t_pmtud_saved_maxseg; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); } } } /* * Disable RFC1323 and SACK if we haven't got any response to * our third SYN to work-around some broken terminal servers * (most of which have hopefully been retired) that have bad VJ * header compression code which trashes TCP segments containing * unknown-to-them TCP options. */ if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && (tp->t_rxtshift == 3)) tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT); /* * If we backed off this far, our srtt estimate is probably bogus. * Clobber it so we'll take the next rtt measurement as our srtt; * move the current srtt into rttvar to keep the current retransmit * times until then. */ if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { #ifdef INET6 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) in6_losing(tp->t_inpcb); else #endif in_losing(tp->t_inpcb); tp->t_rttvar += tp->t_srtt; tp->t_srtt = 0; } sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); tp->snd_recover = tp->snd_max; tp->t_flags |= TF_ACKNOW; tp->t_rtttime = 0; rack_cong_signal(tp, CC_RTO, tp->snd_una); out: return (retval); } static int rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp) { int32_t ret = 0; int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_GPUTINPROG)) { /* * We have a goodput in progress * and we have entered a late state. * Do we have enough data in the sb * to handle the GPUT request? */ uint32_t bytes; bytes = tp->gput_ack - tp->gput_seq; if (SEQ_GT(tp->gput_seq, tp->snd_una)) bytes += tp->gput_seq - tp->snd_una; if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* * There are not enough bytes in the socket * buffer that have been sent to cover this * measurement. Cancel it. */ rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); tp->t_flags &= ~TF_GPUTINPROG; } } if (timers == 0) { return (0); } if (tp->t_state == TCPS_LISTEN) { /* no timers on listen sockets */ if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) return (0); return (1); } if ((timers & PACE_TMR_RACK) && rack->rc_on_min_to) { /* * For the rack timer when we * are on a min-timeout (which means rrr_conf = 3) * we don't want to check the timer. It may * be going off for a pace and thats ok we * want to send the retransmit (if its ready). * * If its on a normal rack timer (non-min) then * we will check if its expired. */ goto skip_time_check; } if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { uint32_t left; if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { ret = -1; rack_log_to_processing(rack, cts, ret, 0); return (0); } if (hpts_calling == 0) { /* * A user send or queued mbuf (sack) has called us? We * return 0 and let the pacing guards * deal with it if they should or * should not cause a send. */ ret = -2; rack_log_to_processing(rack, cts, ret, 0); return (0); } /* * Ok our timer went off early and we are not paced false * alarm, go back to sleep. */ ret = -3; left = rack->r_ctl.rc_timer_exp - cts; tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left)); rack_log_to_processing(rack, cts, ret, left); return (1); } skip_time_check: rack->rc_tmr_stopped = 0; rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; if (timers & PACE_TMR_DELACK) { ret = rack_timeout_delack(tp, rack, cts); } else if (timers & PACE_TMR_RACK) { rack->r_ctl.rc_tlp_rxt_last_time = cts; rack->r_fast_output = 0; ret = rack_timeout_rack(tp, rack, cts); } else if (timers & PACE_TMR_TLP) { rack->r_ctl.rc_tlp_rxt_last_time = cts; ret = rack_timeout_tlp(tp, rack, cts, doing_tlp); } else if (timers & PACE_TMR_RXT) { rack->r_ctl.rc_tlp_rxt_last_time = cts; rack->r_fast_output = 0; ret = rack_timeout_rxt(tp, rack, cts); } else if (timers & PACE_TMR_PERSIT) { ret = rack_timeout_persist(tp, rack, cts); } else if (timers & PACE_TMR_KEEP) { ret = rack_timeout_keepalive(tp, rack, cts); } rack_log_to_processing(rack, cts, ret, timers); return (ret); } static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line) { struct timeval tv; uint32_t us_cts, flags_on_entry; uint8_t hpts_removed = 0; flags_on_entry = rack->r_ctl.rc_hpts_flags; us_cts = tcp_get_usecs(&tv); if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) || ((tp->snd_max - tp->snd_una) == 0))) { tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); hpts_removed = 1; /* If we were not delayed cancel out the flag. */ if ((tp->snd_max - tp->snd_una) == 0) rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry); } if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (rack->rc_inp->inp_in_hpts && ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) { /* * Canceling timer's when we have no output being * paced. We also must remove ourselves from the * hpts. */ tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); hpts_removed = 1; } rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); } if (hpts_removed == 0) rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry); } static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type) { return; } static int rack_stopall(struct tcpcb *tp) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; rack->t_timers_stopped = 1; return (0); } static void rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) { return; } static int rack_timer_active(struct tcpcb *tp, uint32_t timer_type) { return (0); } static void rack_stop_all_timers(struct tcpcb *tp) { struct tcp_rack *rack; /* * Assure no timers are running. */ if (tcp_timer_active(tp, TT_PERSIST)) { /* We enter in persists, set the flag appropriately */ rack = (struct tcp_rack *)tp->t_fb_ptr; rack->rc_in_persist = 1; } tcp_timer_suspend(tp, TT_PERSIST); tcp_timer_suspend(tp, TT_REXMT); tcp_timer_suspend(tp, TT_KEEP); tcp_timer_suspend(tp, TT_DELACK); } static void rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag) { int32_t idx; uint16_t stripped_flags; rsm->r_rtr_cnt++; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); rsm->r_dupack = 0; if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) { rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS; rsm->r_flags |= RACK_OVERMAX; } if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) { rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); } idx = rsm->r_rtr_cnt - 1; rsm->r_tim_lastsent[idx] = ts; stripped_flags = rsm->r_flags & ~(RACK_SENT_SP|RACK_SENT_FP); if (rsm->r_flags & RACK_ACKED) { /* Problably MTU discovery messing with us */ rsm->r_flags &= ~RACK_ACKED; rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); } if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (rsm->r_flags & RACK_SACK_PASSED) { /* We have retransmitted due to the SACK pass */ rsm->r_flags &= ~RACK_SACK_PASSED; rsm->r_flags |= RACK_WAS_SACKPASS; } } static uint32_t rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag) { /* * We (re-)transmitted starting at rsm->r_start for some length * (possibly less than r_end. */ struct rack_sendmap *nrsm, *insret; uint32_t c_end; int32_t len; len = *lenp; c_end = rsm->r_start + len; if (SEQ_GEQ(c_end, rsm->r_end)) { /* * We retransmitted the whole piece or more than the whole * slopping into the next rsm. */ rack_update_rsm(tp, rack, rsm, ts, add_flag); if (c_end == rsm->r_end) { *lenp = 0; return (0); } else { int32_t act_len; /* Hangs over the end return whats left */ act_len = rsm->r_end - rsm->r_start; *lenp = (len - act_len); return (rsm->r_end); } /* We don't get out of this block. */ } /* * Here we retransmitted less than the whole thing which means we * have to split this into what was transmitted and what was not. */ nrsm = rack_alloc_full_limit(rack); if (nrsm == NULL) { /* * We can't get memory, so lets not proceed. */ *lenp = 0; return (0); } /* * So here we are going to take the original rsm and make it what we * retransmitted. nrsm will be the tail portion we did not * retransmit. For example say the chunk was 1, 11 (10 bytes). And * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to * 1, 6 and the new piece will be 6, 11. */ rack_clone_rsm(rack, nrsm, rsm, c_end); nrsm->r_dupack = 0; rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); rack_update_rsm(tp, rack, rsm, ts, add_flag); /* Log a split of rsm into rsm and nrsm */ rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__); *lenp = 0; return (0); } static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t cts, struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls) { struct tcp_rack *rack; struct rack_sendmap *rsm, *nrsm, *insret, fe; register uint32_t snd_max, snd_una; /* * Add to the RACK log of packets in flight or retransmitted. If * there is a TS option we will use the TS echoed, if not we will * grab a TS. * * Retransmissions will increment the count and move the ts to its * proper place. Note that if options do not include TS's then we * won't be able to effectively use the ACK for an RTT on a retran. * * Notes about r_start and r_end. Lets consider a send starting at * sequence 1 for 10 bytes. In such an example the r_start would be * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. * This means that r_end is actually the first sequence for the next * slot (11). * */ /* * If err is set what do we do XXXrrs? should we not add the thing? * -- i.e. return if err != 0 or should we pretend we sent it? -- * i.e. proceed with add ** do this for now. */ INP_WLOCK_ASSERT(tp->t_inpcb); if (err) /* * We don't log errors -- we could but snd_max does not * advance in this case either. */ return; if (th_flags & TH_RST) { /* * We don't log resets and we return immediately from * sending */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; snd_una = tp->snd_una; snd_max = tp->snd_max; if (th_flags & (TH_SYN | TH_FIN)) { /* * The call to rack_log_output is made before bumping * snd_max. This means we can record one extra byte on a SYN * or FIN if seq_out is adding more on and a FIN is present * (and we are not resending). */ if ((th_flags & TH_SYN) && (seq_out == tp->iss)) len++; if (th_flags & TH_FIN) len++; if (SEQ_LT(snd_max, tp->snd_nxt)) { /* * The add/update as not been done for the FIN/SYN * yet. */ snd_max = tp->snd_nxt; } } if (SEQ_LEQ((seq_out + len), snd_una)) { /* Are sending an old segment to induce an ack (keep-alive)? */ return; } if (SEQ_LT(seq_out, snd_una)) { /* huh? should we panic? */ uint32_t end; end = seq_out + len; seq_out = snd_una; if (SEQ_GEQ(end, seq_out)) len = end - seq_out; else len = 0; } if (len == 0) { /* We don't log zero window probes */ return; } rack->r_ctl.rc_time_last_sent = cts; if (IN_FASTRECOVERY(tp->t_flags)) { rack->r_ctl.rc_prr_out += len; } /* First question is it a retransmission or new? */ if (seq_out == snd_max) { /* Its new */ again: rsm = rack_alloc(rack); if (rsm == NULL) { /* * Hmm out of memory and the tcb got destroyed while * we tried to wait. */ return; } if (th_flags & TH_FIN) { rsm->r_flags = RACK_HAS_FIN|add_flag; } else { rsm->r_flags = add_flag; } if (hw_tls) rsm->r_hw_tls = 1; rsm->r_tim_lastsent[0] = cts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; if (th_flags & TH_SYN) { /* The data space is one beyond snd_una */ rsm->r_flags |= RACK_HAS_SYN; } rsm->r_start = seq_out; rsm->r_end = rsm->r_start + len; rsm->r_dupack = 0; /* * save off the mbuf location that * sndmbuf_noadv returned (which is * where we started copying from).. */ rsm->m = s_mb; rsm->soff = s_moff; /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */ if (rsm->m) { if (rsm->m->m_len <= rsm->soff) { /* * XXXrrs Question, will this happen? * * If sbsndptr is set at the correct place * then s_moff should always be somewhere * within rsm->m. But if the sbsndptr was * off then that won't be true. If it occurs * we need to walkout to the correct location. */ struct mbuf *lm; lm = rsm->m; while (lm->m_len <= rsm->soff) { rsm->soff -= lm->m_len; lm = lm->m_next; KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u", __func__, rack, s_moff, s_mb, rsm->soff)); } rsm->m = lm; counter_u64_add(rack_sbsndptr_wrong, 1); } else counter_u64_add(rack_sbsndptr_right, 1); rsm->orig_m_len = rsm->m->m_len; } else rsm->orig_m_len = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* Log a new rsm */ rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; /* * Special case detection, is there just a single * packet outstanding when we are not in recovery? * * If this is true mark it so. */ if ((IN_FASTRECOVERY(tp->t_flags) == 0) && (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) { struct rack_sendmap *prsm; prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (prsm) prsm->r_one_out_nr = 1; } return; } /* * If we reach here its a retransmission and we need to find it. */ memset(&fe, 0, sizeof(fe)); more: if (hintrsm && (hintrsm->r_start == seq_out)) { rsm = hintrsm; hintrsm = NULL; } else { /* No hints sorry */ rsm = NULL; } if ((rsm) && (rsm->r_start == seq_out)) { seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag); if (len == 0) { return; } else { goto more; } } /* Ok it was not the last pointer go through it the hard way. */ refind: fe.r_start = seq_out; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm) { if (rsm->r_start == seq_out) { seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag); if (len == 0) { return; } else { goto refind; } } if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { /* Transmitted within this piece */ /* * Ok we must split off the front and then let the * update do the rest */ nrsm = rack_alloc_full_limit(rack); if (nrsm == NULL) { rack_update_rsm(tp, rack, rsm, cts, add_flag); return; } /* * copy rsm to nrsm and then trim the front of rsm * to not include this part. */ rack_clone_rsm(rack, nrsm, rsm, seq_out); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag); if (len == 0) { return; } else if (len > 0) goto refind; } } /* * Hmm not found in map did they retransmit both old and on into the * new? */ if (seq_out == tp->snd_max) { goto again; } else if (SEQ_LT(seq_out, tp->snd_max)) { #ifdef INVARIANTS printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", seq_out, len, tp->snd_una, tp->snd_max); printf("Starting Dump of all rack entries\n"); RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { printf("rsm:%p start:%u end:%u\n", rsm, rsm->r_start, rsm->r_end); } printf("Dump complete\n"); panic("seq_out not found rack:%p tp:%p", rack, tp); #endif } else { #ifdef INVARIANTS /* * Hmm beyond sndmax? (only if we are using the new rtt-pack * flag) */ panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", seq_out, len, tp->snd_max, tp); #endif } } /* * Record one of the RTT updates from an ack into * our sample structure. */ static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt) { if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) { rack->r_ctl.rack_rs.rs_rtt_lowest = rtt; } if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) { rack->r_ctl.rack_rs.rs_rtt_highest = rtt; } if (rack->rc_tp->t_flags & TF_GPUTINPROG) { if (us_rtt < rack->r_ctl.rc_gp_lowrtt) rack->r_ctl.rc_gp_lowrtt = us_rtt; if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd) rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd; } if ((confidence == 1) && ((rsm == NULL) || (rsm->r_just_ret) || (rsm->r_one_out_nr && len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) { /* * If the rsm had a just return * hit it then we can't trust the * rtt measurement for buffer deterimination * Note that a confidence of 2, indicates * SACK'd which overrides the r_just_ret or * the r_one_out_nr. If it was a CUM-ACK and * we had only two outstanding, but get an * ack for only 1. Then that also lowers our * confidence. */ confidence = 0; } if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) { if (rack->r_ctl.rack_rs.confidence == 0) { /* * We take anything with no current confidence * saved. */ rack->r_ctl.rack_rs.rs_us_rtt = us_rtt; rack->r_ctl.rack_rs.confidence = confidence; rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt; } else if (confidence || rack->r_ctl.rack_rs.confidence) { /* * Once we have a confident number, * we can update it with a smaller * value since this confident number * may include the DSACK time until * the next segment (the second one) arrived. */ rack->r_ctl.rack_rs.rs_us_rtt = us_rtt; rack->r_ctl.rack_rs.confidence = confidence; rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt; } } rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence); rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID; rack->r_ctl.rack_rs.rs_rtt_tot += rtt; rack->r_ctl.rack_rs.rs_rtt_cnt++; } /* * Collect new round-trip time estimate * and update averages and current timeout. */ static void tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp) { int32_t delta; uint32_t o_srtt, o_var; int32_t hrtt_up = 0; int32_t rtt; if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) /* No valid sample */ return; if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) { /* We are to use the lowest RTT seen in a single ack */ rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) { /* We are to use the highest RTT seen in a single ack */ rtt = rack->r_ctl.rack_rs.rs_rtt_highest; } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) { /* We are to use the average RTT seen in a single ack */ rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot / (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt); } else { #ifdef INVARIANTS panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method); #endif return; } if (rtt == 0) rtt = 1; if (rack->rc_gp_rtt_set == 0) { /* * With no RTT we have to accept * even one we are not confident of. */ rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt; rack->rc_gp_rtt_set = 1; } else if (rack->r_ctl.rack_rs.confidence) { /* update the running gp srtt */ rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8); rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8; } if (rack->r_ctl.rack_rs.confidence) { /* * record the low and high for highly buffered path computation, * we only do this if we are confident (not a retransmission). */ if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) { rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt; hrtt_up = 1; } if (rack->rc_highly_buffered == 0) { /* * Currently once we declare a path has * highly buffered there is no going * back, which may be a problem... */ if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) { rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt, rack->r_ctl.rc_highest_us_rtt, rack->r_ctl.rc_lowest_us_rtt, RACK_RTTS_SEEHBP); rack->rc_highly_buffered = 1; } } } if ((rack->r_ctl.rack_rs.confidence) || (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) { /* * If we are highly confident of it it was * never retransmitted we accept it as the last us_rtt. */ rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt; /* The lowest rtt can be set if its was not retransmited */ if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) { rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt; if (rack->r_ctl.rc_lowest_us_rtt == 0) rack->r_ctl.rc_lowest_us_rtt = 1; } } o_srtt = tp->t_srtt; o_var = tp->t_rttvar; rack = (struct tcp_rack *)tp->t_fb_ptr; if (tp->t_srtt != 0) { /* * We keep a simple srtt in microseconds, like our rtt * measurement. We don't need to do any tricks with shifting * etc. Instead we just add in 1/8th of the new measurement * and subtract out 1/8 of the old srtt. We do the same with * the variance after finding the absolute value of the * difference between this sample and the current srtt. */ delta = tp->t_srtt - rtt; /* Take off 1/8th of the current sRTT */ tp->t_srtt -= (tp->t_srtt >> 3); /* Add in 1/8th of the new RTT just measured */ tp->t_srtt += (rtt >> 3); if (tp->t_srtt <= 0) tp->t_srtt = 1; /* Now lets make the absolute value of the variance */ if (delta < 0) delta = -delta; /* Subtract out 1/8th */ tp->t_rttvar -= (tp->t_rttvar >> 3); /* Add in 1/8th of the new variance we just saw */ tp->t_rttvar += (delta >> 3); if (tp->t_rttvar <= 0) tp->t_rttvar = 1; if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } else { /* * No rtt measurement yet - use the unsmoothed rtt. Set the * variance to half the rtt (so our first retransmit happens * at 3*rtt). */ tp->t_srtt = rtt; tp->t_rttvar = rtt >> 1; tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } rack->rc_srtt_measure_made = 1; KMOD_TCPSTAT_INC(tcps_rttupdated); tp->t_rttupdated++; #ifdef STATS if (rack_stats_gets_ms_rtt == 0) { /* Send in the microsecond rtt used for rxt timeout purposes */ stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); } else if (rack_stats_gets_ms_rtt == 1) { /* Send in the millisecond rtt used for rxt timeout purposes */ int32_t ms_rtt; /* Round up */ ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC; stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt)); } else if (rack_stats_gets_ms_rtt == 2) { /* Send in the millisecond rtt has close to the path RTT as we can get */ int32_t ms_rtt; /* Round up */ ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC; stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt)); } else { /* Send in the microsecond rtt has close to the path RTT as we can get */ stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt)); } #endif /* * the retransmit should happen at rtt + 4 * rttvar. Because of the * way we do the smoothing, srtt and rttvar will each average +1/2 * tick of bias. When we compute the retransmit timer, we want 1/2 * tick of rounding and 1 extra tick because of +-1/2 tick * uncertainty in the firing of the timer. The bias will give us * exactly the 1.5 tick we need. But, because the bias is * statistical, we have to test that we don't drop below the minimum * feasible timer (which is 2 ticks). */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop); rack_log_rtt_sample(rack, rtt); tp->t_softerror = 0; } static void rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts) { /* * Apply to filter the inbound us-rtt at us_cts. */ uint32_t old_rtt; old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt, us_rtt, us_cts); if (rack->r_ctl.last_pacing_time && rack->rc_gp_dyn_mul && (rack->r_ctl.last_pacing_time > us_rtt)) rack->pacing_longer_than_rtt = 1; else rack->pacing_longer_than_rtt = 0; if (old_rtt > us_rtt) { /* We just hit a new lower rtt time */ rack_log_rtt_shrinks(rack, us_cts, old_rtt, __LINE__, RACK_RTTS_NEWRTT); /* * Only count it if its lower than what we saw within our * calculated range. */ if ((old_rtt - us_rtt) > rack_min_rtt_movement) { if (rack_probertt_lower_within && rack->rc_gp_dyn_mul && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* * We are seeing a new lower rtt very close * to the time that we would have entered probe-rtt. * This is probably due to the fact that a peer flow * has entered probe-rtt. Lets go in now too. */ uint32_t val; val = rack_probertt_lower_within * rack_time_between_probertt; val /= 100; if ((rack->in_probe_rtt == 0) && ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) { rack_enter_probertt(rack, us_cts); } } rack->r_ctl.rc_lower_rtt_us_cts = us_cts; } } } static int rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack) { int32_t i, all; uint32_t t, len_acked; if ((rsm->r_flags & RACK_ACKED) || (rsm->r_flags & RACK_WAS_ACKED)) /* Already done */ return (0); if (rsm->r_no_rtt_allowed) { /* Not allowed */ return (0); } if (ack_type == CUM_ACKED) { if (SEQ_GT(th_ack, rsm->r_end)) { len_acked = rsm->r_end - rsm->r_start; all = 1; } else { len_acked = th_ack - rsm->r_start; all = 0; } } else { len_acked = rsm->r_end - rsm->r_start; all = 0; } if (rsm->r_rtr_cnt == 1) { uint32_t us_rtt; t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; if ((int)t <= 0) t = 1; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; else us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; if (us_rtt == 0) us_rtt = 1; rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time)); if (ack_type == SACKED) { rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1); tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt); } else { /* * We need to setup what our confidence * is in this ack. * * If the rsm was app limited and it is * less than a mss in length (the end * of the send) then we have a gap. If we * were app limited but say we were sending * multiple MSS's then we are more confident * int it. * * When we are not app-limited then we see if * the rsm is being included in the current * measurement, we tell this by the app_limited_needs_set * flag. * * Note that being cwnd blocked is not applimited * as well as the pacing delay between packets which * are sending only 1 or 2 MSS's also will show up * in the RTT. We probably need to examine this algorithm * a bit more and enhance it to account for the delay * between rsm's. We could do that by saving off the * pacing delay of each rsm (in an rsm) and then * factoring that in somehow though for now I am * not sure how :) */ int calc_conf = 0; if (rsm->r_flags & RACK_APP_LIMITED) { if (all && (len_acked <= ctf_fixed_maxseg(tp))) calc_conf = 0; else calc_conf = 1; } else if (rack->app_limited_needs_set == 0) { calc_conf = 1; } else { calc_conf = 0; } rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2); tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, calc_conf, rsm, rsm->r_rtr_cnt); } if ((rsm->r_flags & RACK_TLP) && (!IN_FASTRECOVERY(tp->t_flags))) { /* Segment was a TLP and our retrans matched */ if (rack->r_ctl.rc_tlp_cwnd_reduce) { rack->r_ctl.rc_rsm_start = tp->snd_max; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; rack_cong_signal(tp, CC_NDUPACK, tp->snd_una); } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; rack->rc_rack_rtt = t; } return (1); } /* * We clear the soft/rxtshift since we got an ack. * There is no assurance we will call the commit() function * so we need to clear these to avoid incorrect handling. */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); tp->t_softerror = 0; if (to && (to->to_flags & TOF_TS) && (ack_type == CUM_ACKED) && (to->to_tsecr) && ((rsm->r_flags & RACK_OVERMAX) == 0)) { /* * Now which timestamp does it match? In this block the ACK * must be coming from a previous transmission. */ for (i = 0; i < rsm->r_rtr_cnt; i++) { if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) { t = cts - (uint32_t)rsm->r_tim_lastsent[i]; if ((int)t <= 0) t = 1; if ((i + 1) < rsm->r_rtr_cnt) { /* * The peer ack'd from our previous * transmission. We have a spurious * retransmission and thus we dont * want to update our rack_rtt. */ return (0); } if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; rack->rc_rack_rtt = t; } rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3); tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm, rsm->r_rtr_cnt); return (1); } } goto ts_not_found; } else { /* * Ok its a SACK block that we retransmitted. or a windows * machine without timestamps. We can tell nothing from the * time-stamp since its not there or the time the peer last * recieved a segment that moved forward its cum-ack point. */ ts_not_found: i = rsm->r_rtr_cnt - 1; t = cts - (uint32_t)rsm->r_tim_lastsent[i]; if ((int)t <= 0) t = 1; if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { /* * We retransmitted and the ack came back in less * than the smallest rtt we have observed. We most * likely did an improper retransmit as outlined in * 6.2 Step 2 point 2 in the rack-draft so we * don't want to update our rack_rtt. We in * theory (in future) might want to think about reverting our * cwnd state but we won't for now. */ return (0); } else if (rack->r_ctl.rc_rack_min_rtt) { /* * We retransmitted it and the retransmit did the * job. */ if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i]; rack->rc_rack_rtt = t; } return (1); } } return (0); } /* * Mark the SACK_PASSED flag on all entries prior to rsm send wise. */ static void rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm) { struct rack_sendmap *nrsm; nrsm = rsm; TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap, rack_head, r_tnext) { if (nrsm == rsm) { /* Skip orginal segment he is acked */ continue; } if (nrsm->r_flags & RACK_ACKED) { /* * Skip ack'd segments, though we * should not see these, since tmap * should not have ack'd segments. */ continue; } if (nrsm->r_flags & RACK_SACK_PASSED) { /* * We found one that is already marked * passed, we have been here before and * so all others below this are marked. */ break; } nrsm->r_flags |= RACK_SACK_PASSED; nrsm->r_flags &= ~RACK_WAS_SACKPASS; } } static void rack_need_set_test(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack, int line, int use_which) { if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GEQ(rsm->r_end, tp->gput_seq)) { /* * We were app limited, and this ack * butts up or goes beyond the point where we want * to start our next measurement. We need * to record the new gput_ts as here and * possibly update the start sequence. */ uint32_t seq, ts; if (rsm->r_rtr_cnt > 1) { /* * This is a retransmit, can we * really make any assessment at this * point? We are not really sure of * the timestamp, is it this or the * previous transmission? * * Lets wait for something better that * is not retransmitted. */ return; } seq = tp->gput_seq; ts = tp->gput_ts; rack->app_limited_needs_set = 0; tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); /* Do we start at a new end? */ if ((use_which == RACK_USE_BEG) && SEQ_GEQ(rsm->r_start, tp->gput_seq)) { /* * When we get an ACK that just eats * up some of the rsm, we set RACK_USE_BEG * since whats at r_start (i.e. th_ack) * is left unacked and thats where the * measurement not starts. */ tp->gput_seq = rsm->r_start; rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } if ((use_which == RACK_USE_END) && SEQ_GEQ(rsm->r_end, tp->gput_seq)) { /* * We use the end when the cumack * is moving forward and completely * deleting the rsm passed so basically * r_end holds th_ack. * * For SACK's we also want to use the end * since this piece just got sacked and * we want to target anything after that * in our measurement. */ tp->gput_seq = rsm->r_end; rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } if (use_which == RACK_USE_END_OR_THACK) { /* * special case for ack moving forward, * not a sack, we need to move all the * way up to where this ack cum-ack moves * to. */ if (SEQ_GT(th_ack, rsm->r_end)) tp->gput_seq = th_ack; else tp->gput_seq = rsm->r_end; rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } if (SEQ_GT(tp->gput_seq, tp->gput_ack)) { /* * We moved beyond this guy's range, re-calculate * the new end point. */ if (rack->rc_gp_filled == 0) { tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp))); } else { tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack); } } /* * We are moving the goal post, we may be able to clear the * measure_saw_probe_rtt flag. */ if ((rack->in_probe_rtt == 0) && (rack->measure_saw_probe_rtt) && (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit))) rack->measure_saw_probe_rtt = 0; rack_log_pacing_delay_calc(rack, ts, tp->gput_ts, seq, tp->gput_seq, 0, 5, line, NULL, 0); if (rack->rc_gp_filled && ((tp->gput_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp))))) { uint32_t ideal_amount; ideal_amount = rack_get_measure_window(tp, rack); if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* * There is no sense of continuing this measurement * because its too small to gain us anything we * trust. Skip it and that way we can start a new * measurement quicker. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq, 0, 0, 0, 6, __LINE__, NULL, 0); } else { /* * Reset the window further out. */ tp->gput_ack = tp->gput_seq + ideal_amount; } } } } static uint32_t rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two) { uint32_t start, end, changed = 0; struct rack_sendmap stack_map; struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next; int32_t used_ref = 1; int moved = 0; start = sack->start; end = sack->end; rsm = *prsm; memset(&fe, 0, sizeof(fe)); do_rest_ofb: if ((rsm == NULL) || (SEQ_LT(end, rsm->r_start)) || (SEQ_GEQ(start, rsm->r_end)) || (SEQ_LT(start, rsm->r_start))) { /* * We are not in the right spot, * find the correct spot in the tree. */ used_ref = 0; fe.r_start = start; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); moved++; } if (rsm == NULL) { /* TSNH */ goto out; } /* Ok we have an ACK for some piece of this rsm */ if (rsm->r_start != start) { if ((rsm->r_flags & RACK_ACKED) == 0) { /** * Need to split this in two pieces the before and after, * the before remains in the map, the after must be * added. In other words we have: * rsm |--------------| * sackblk |-------> * rsm will become * rsm |---| * and nrsm will be the sacked piece * nrsm |----------| * * But before we start down that path lets * see if the sack spans over on top of * the next guy and it is already sacked. */ next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (next && (next->r_flags & RACK_ACKED) && SEQ_GEQ(end, next->r_start)) { /** * So the next one is already acked, and * we can thus by hookery use our stack_map * to reflect the piece being sacked and * then adjust the two tree entries moving * the start and ends around. So we start like: * rsm |------------| (not-acked) * next |-----------| (acked) * sackblk |--------> * We want to end like so: * rsm |------| (not-acked) * next |-----------------| (acked) * nrsm |-----| * Where nrsm is a temporary stack piece we * use to update all the gizmos. */ /* Copy up our fudge block */ nrsm = &stack_map; memcpy(nrsm, rsm, sizeof(struct rack_sendmap)); /* Now adjust our tree blocks */ rsm->r_end = start; next->r_start = start; /* Now we must adjust back where next->m is */ rack_setup_offset_for_rsm(rsm, next); /* We don't need to adjust rsm, it did not change */ /* Clear out the dup ack count of the remainder */ rsm->r_dupack = 0; rsm->r_just_ret = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* Now lets make sure our fudge block is right */ nrsm->r_start = start; /* Now lets update all the stats and such */ rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0); if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END); changed += (nrsm->r_end - nrsm->r_start); rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start); if (nrsm->r_flags & RACK_SACK_PASSED) { counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } /* * Now we want to go up from rsm (the * one left un-acked) to the next one * in the tmap. We do this so when * we walk backwards we include marking * sack-passed on rsm (The one passed in * is skipped since it is generally called * on something sacked before removing it * from the tmap). */ if (rsm->r_in_tmap) { nrsm = TAILQ_NEXT(rsm, r_tnext); /* * Now that we have the next * one walk backwards from there. */ if (nrsm && nrsm->r_in_tmap) rack_log_sack_passed(tp, rack, nrsm); } /* Now are we done? */ if (SEQ_LT(end, next->r_end) || (end == next->r_end)) { /* Done with block */ goto out; } rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__); counter_u64_add(rack_sack_used_next_merge, 1); /* Postion for the next block */ start = next->r_end; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next); if (rsm == NULL) goto out; } else { /** * We can't use any hookery here, so we * need to split the map. We enter like * so: * rsm |--------| * sackblk |-----> * We will add the new block nrsm and * that will be the new portion, and then * fall through after reseting rsm. So we * split and look like this: * rsm |----| * sackblk |-----> * nrsm |---| * We then fall through reseting * rsm to nrsm, so the next block * picks it up. */ nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* * failed XXXrrs what can we do but loose the sack * info? */ goto out; } counter_u64_add(rack_sack_splits, 1); rack_clone_rsm(rack, nrsm, rsm, start); rsm->r_just_ret = 0; insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__); rsm->r_flags &= (~RACK_HAS_FIN); /* Position us to point to the new nrsm that starts the sack blk */ rsm = nrsm; } } else { /* Already sacked this piece */ counter_u64_add(rack_sack_skipped_acked, 1); moved++; if (end == rsm->r_end) { /* Done with block */ rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); goto out; } else if (SEQ_LT(end, rsm->r_end)) { /* A partial sack to a already sacked block */ moved++; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); goto out; } else { /* * The end goes beyond this guy * repostion the start to the * next block. */ start = rsm->r_end; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (rsm == NULL) goto out; } } } if (SEQ_GEQ(end, rsm->r_end)) { /** * The end of this block is either beyond this guy or right * at this guy. I.e.: * rsm --- |-----| * end |-----| * * end |---------| */ if ((rsm->r_flags & RACK_ACKED) == 0) { rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0); changed += (rsm->r_end - rsm->r_start); rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); if (rsm->r_in_tmap) /* should be true */ rack_log_sack_passed(tp, rack, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & RACK_SACK_PASSED) { rsm->r_flags &= ~RACK_SACK_PASSED; counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END); rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; rsm->r_flags &= ~RACK_TLP; if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__); } else { counter_u64_add(rack_sack_skipped_acked, 1); moved++; } if (end == rsm->r_end) { /* This block only - done, setup for next */ goto out; } /* * There is more not coverend by this rsm move on * to the next block in the RB tree. */ nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); start = rsm->r_end; rsm = nrsm; if (rsm == NULL) goto out; goto do_rest_ofb; } /** * The end of this sack block is smaller than * our rsm i.e.: * rsm --- |-----| * end |--| */ if ((rsm->r_flags & RACK_ACKED) == 0) { prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (prev && (prev->r_flags & RACK_ACKED)) { /** * Goal, we want the right remainder of rsm to shrink * in place and span from (rsm->r_start = end) to rsm->r_end. * We want to expand prev to go all the way * to prev->r_end <- end. * so in the tree we have before: * prev |--------| (acked) * rsm |-------| (non-acked) * sackblk |-| * We churn it so we end up with * prev |----------| (acked) * rsm |-----| (non-acked) * nrsm |-| (temporary) */ nrsm = &stack_map; memcpy(nrsm, rsm, sizeof(struct rack_sendmap)); prev->r_end = end; rsm->r_start = end; /* Now adjust nrsm (stack copy) to be * the one that is the small * piece that was "sacked". */ nrsm->r_end = end; rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* * Now that the rsm has had its start moved forward * lets go ahead and get its new place in the world. */ rack_setup_offset_for_rsm(prev, rsm); /* * Now nrsm is our new little piece * that is acked (which was merged * to prev). Update the rtt and changed * based on that. Also check for reordering. */ rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0); if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END); changed += (nrsm->r_end - nrsm->r_start); rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start); if (nrsm->r_flags & RACK_SACK_PASSED) { counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__); rsm = prev; counter_u64_add(rack_sack_used_prev_merge, 1); } else { /** * This is the case where our previous * block is not acked either, so we must * split the block in two. */ nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* failed rrs what can we do but loose the sack info? */ goto out; } /** * In this case nrsm becomes * nrsm->r_start = end; * nrsm->r_end = rsm->r_end; * which is un-acked. * * rsm->r_end = nrsm->r_start; * i.e. the remaining un-acked * piece is left on the left * hand side. * * So we start like this * rsm |----------| (not acked) * sackblk |---| * build it so we have * rsm |---| (acked) * nrsm |------| (not acked) */ counter_u64_add(rack_sack_splits, 1); rack_clone_rsm(rack, nrsm, rsm, end); rsm->r_flags &= (~RACK_HAS_FIN); rsm->r_just_ret = 0; insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } nrsm->r_dupack = 0; rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2); rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0); changed += (rsm->r_end - rsm->r_start); rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); if (rsm->r_in_tmap) /* should be true */ rack_log_sack_passed(tp, rack, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & RACK_SACK_PASSED) { rsm->r_flags &= ~RACK_SACK_PASSED; counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END); rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; rsm->r_flags &= ~RACK_TLP; rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__); if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } } } else if (start != end){ /* * The block was already acked. */ counter_u64_add(rack_sack_skipped_acked, 1); moved++; } out: if (rsm && (rsm->r_flags & RACK_ACKED)) { /* * Now can we merge where we worked * with either the previous or * next block? */ next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); while (next) { if (next->r_flags & RACK_ACKED) { /* yep this and next can be merged */ rsm = rack_merge_rsm(rack, rsm, next); next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); } else break; } /* Now what about the previous? */ prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); while (prev) { if (prev->r_flags & RACK_ACKED) { /* yep the previous and this can be merged */ rsm = rack_merge_rsm(rack, prev, rsm); prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); } else break; } } if (used_ref == 0) { counter_u64_add(rack_sack_proc_all, 1); } else { counter_u64_add(rack_sack_proc_short, 1); } /* Save off the next one for quick reference. */ if (rsm) nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); else nrsm = NULL; *prsm = rack->r_ctl.rc_sacklast = nrsm; /* Pass back the moved. */ *moved_two = moved; return (changed); } static void inline rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack) { struct rack_sendmap *tmap; tmap = NULL; while (rsm && (rsm->r_flags & RACK_ACKED)) { /* Its no longer sacked, mark it so */ rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); #ifdef INVARIANTS if (rsm->r_in_tmap) { panic("rack:%p rsm:%p flags:0x%x in tmap?", rack, rsm, rsm->r_flags); } #endif rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS); /* Rebuild it into our tmap */ if (tmap == NULL) { TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); tmap = rsm; } else { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext); tmap = rsm; } tmap->r_in_tmap = 1; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); } /* * Now lets possibly clear the sack filter so we start * recognizing sacks that cover this area. */ sack_filter_clear(&rack->r_ctl.rack_sf, th_ack); } static void rack_do_decay(struct tcp_rack *rack) { struct timeval res; #define timersub(tvp, uvp, vvp) \ do { \ (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \ (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \ if ((vvp)->tv_usec < 0) { \ (vvp)->tv_sec--; \ (vvp)->tv_usec += 1000000; \ } \ } while (0) timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res); #undef timersub rack->r_ctl.input_pkt++; if ((rack->rc_in_persist) || (res.tv_sec >= 1) || (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) { /* * Check for decay of non-SAD, * we want all SAD detection metrics to * decay 1/4 per second (or more) passed. */ uint32_t pkt_delta; pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt; /* Update our saved tracking values */ rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt; rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time; /* Now do we escape without decay? */ #ifdef NETFLIX_EXP_DETECTION if (rack->rc_in_persist || (rack->rc_tp->snd_max == rack->rc_tp->snd_una) || (pkt_delta < tcp_sad_low_pps)){ /* * We don't decay idle connections * or ones that have a low input pps. */ return; } /* Decay the counters */ rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count, tcp_sad_decay_val); rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count, tcp_sad_decay_val); rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra, tcp_sad_decay_val); rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move, tcp_sad_decay_val); #endif } } static void rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to) { struct rack_sendmap *rsm, *rm; /* * The ACK point is advancing to th_ack, we must drop off * the packets in the rack log and calculate any eligble * RTT's. */ rack->r_wanted_output = 1; more: rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm == NULL) { if ((th_ack - 1) == tp->iss) { /* * For the SYN incoming case we will not * have called tcp_output for the sending of * the SYN, so there will be no map. All * other cases should probably be a panic. */ return; } if (tp->t_flags & TF_SENTFIN) { /* if we sent a FIN we often will not have map */ return; } #ifdef INVARIANTS panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n", tp, tp->t_state, th_ack, rack, tp->snd_una, tp->snd_max, tp->snd_nxt); #endif return; } if (SEQ_LT(th_ack, rsm->r_start)) { /* Huh map is missing this */ #ifdef INVARIANTS printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n", rsm->r_start, th_ack, tp->t_state, rack->r_state); #endif return; } rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack); /* Now do we consume the whole thing? */ if (SEQ_GEQ(th_ack, rsm->r_end)) { /* Its all consumed. */ uint32_t left; uint8_t newly_acked; rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__); rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; /* Record the time of highest cumack sent */ rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS if (rm != rsm) { panic("removing head in rack:%p rsm:%p rm:%p", rack, rsm, rm); } #endif if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } newly_acked = 1; if (rsm->r_flags & RACK_ACKED) { /* * It was acked on the scoreboard -- remove * it from total */ rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); newly_acked = 0; } else if (rsm->r_flags & RACK_SACK_PASSED) { /* * There are segments ACKED on the * scoreboard further up. We are seeing * reordering. */ rsm->r_flags &= ~RACK_SACK_PASSED; counter_u64_add(rack_reorder_seen, 1); rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; rack->r_ctl.rc_reorder_ts = cts; if (rack->r_ent_rec_ns) { /* * We have sent no more, and we saw an sack * then ack arrive. */ rack->r_might_revert = 1; } } if ((rsm->r_flags & RACK_TO_REXT) && (tp->t_flags & TF_RCVD_TSTMP) && (to->to_flags & TOF_TS) && (tp->t_flags & TF_PREVVALID)) { /* * We can use the timestamp to see * if this retransmission was from the * first transmit. If so we made a mistake. */ tp->t_flags &= ~TF_PREVVALID; if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) { /* The first transmit is what this ack is for */ rack_cong_signal(tp, CC_RTO_ERR, th_ack); } } left = th_ack - rsm->r_end; if (rack->app_limited_needs_set && newly_acked) rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK); /* Free back to zone */ rack_free(rack, rsm); if (left) { goto more; } /* Check for reneging */ rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) { /* * The peer has moved snd_una up to * the edge of this send, i.e. one * that it had previously acked. The only * way that can be true if the peer threw * away data (space issues) that it had * previously sacked (else it would have * given us snd_una up to (rsm->r_end). * We need to undo the acked markings here. * * Note we have to look to make sure th_ack is * our rsm->r_start in case we get an old ack * where th_ack is behind snd_una. */ rack_peer_reneges(rack, rsm, th_ack); } return; } if (rsm->r_flags & RACK_ACKED) { /* * It was acked on the scoreboard -- remove it from * total for the part being cum-acked. */ rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start); } /* * Clear the dup ack count for * the piece that remains. */ rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); if (rsm->r_rtr_bytes) { /* * It was retransmitted adjust the * sack holes for what was acked. */ int ack_am; ack_am = (th_ack - rsm->r_start); if (ack_am >= rsm->r_rtr_bytes) { rack->r_ctl.rc_holes_rxt -= ack_am; rsm->r_rtr_bytes -= ack_am; } } /* * Update where the piece starts and record * the time of send of highest cumack sent. */ rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__); /* Now we need to move our offset forward too */ if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) { /* Fix up the orig_m_len and possibly the mbuf offset */ rack_adjust_orig_mlen(rsm); } rsm->soff += (th_ack - rsm->r_start); rsm->r_start = th_ack; /* Now do we need to move the mbuf fwd too? */ if (rsm->m) { while (rsm->soff >= rsm->m->m_len) { rsm->soff -= rsm->m->m_len; rsm->m = rsm->m->m_next; KASSERT((rsm->m != NULL), (" nrsm:%p hit at soff:%u null m", rsm, rsm->soff)); } rsm->orig_m_len = rsm->m->m_len; } if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG); } static void rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack) { struct rack_sendmap *rsm; int sack_pass_fnd = 0; if (rack->r_might_revert) { /* * Ok we have reordering, have not sent anything, we * might want to revert the congestion state if nothing * further has SACK_PASSED on it. Lets check. * * We also get here when we have DSACKs come in for * all the data that we FR'd. Note that a rxt or tlp * timer clears this from happening. */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { if (rsm->r_flags & RACK_SACK_PASSED) { sack_pass_fnd = 1; break; } } if (sack_pass_fnd == 0) { /* * We went into recovery * incorrectly due to reordering! */ int orig_cwnd; rack->r_ent_rec_ns = 0; orig_cwnd = tp->snd_cwnd; tp->snd_cwnd = rack->r_ctl.rc_cwnd_at_erec; tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec; tp->snd_recover = tp->snd_una; rack_log_to_prr(rack, 14, orig_cwnd); EXIT_RECOVERY(tp->t_flags); } rack->r_might_revert = 0; } } #ifdef NETFLIX_EXP_DETECTION static void rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack, uint32_t bytes_this_ack, uint32_t segsiz) { if ((rack->do_detection || tcp_force_detection) && tcp_sack_to_ack_thresh && tcp_sack_to_move_thresh && ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) { /* * We have thresholds set to find * possible attackers and disable sack. * Check them. */ uint64_t ackratio, moveratio, movetotal; /* Log detecting */ rack_log_sad(rack, 1); ackratio = (uint64_t)(rack->r_ctl.sack_count); ackratio *= (uint64_t)(1000); if (rack->r_ctl.ack_count) ackratio /= (uint64_t)(rack->r_ctl.ack_count); else { /* We really should not hit here */ ackratio = 1000; } if ((rack->sack_attack_disable == 0) && (ackratio > rack_highest_sack_thresh_seen)) rack_highest_sack_thresh_seen = (uint32_t)ackratio; movetotal = rack->r_ctl.sack_moved_extra; movetotal += rack->r_ctl.sack_noextra_move; moveratio = rack->r_ctl.sack_moved_extra; moveratio *= (uint64_t)1000; if (movetotal) moveratio /= movetotal; else { /* No moves, thats pretty good */ moveratio = 0; } if ((rack->sack_attack_disable == 0) && (moveratio > rack_highest_move_thresh_seen)) rack_highest_move_thresh_seen = (uint32_t)moveratio; if (rack->sack_attack_disable == 0) { if ((ackratio > tcp_sack_to_ack_thresh) && (moveratio > tcp_sack_to_move_thresh)) { /* Disable sack processing */ rack->sack_attack_disable = 1; if (rack->r_rep_attack == 0) { rack->r_rep_attack = 1; counter_u64_add(rack_sack_attacks_detected, 1); } if (tcp_attack_on_turns_on_logging) { /* * Turn on logging, used for debugging * false positives. */ rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging; } /* Clamp the cwnd at flight size */ rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd; rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); rack_log_sad(rack, 2); } } else { /* We are sack-disabled check for false positives */ if ((ackratio <= tcp_restoral_thresh) || (rack->r_ctl.rc_num_maps_alloced < tcp_map_minimum)) { rack->sack_attack_disable = 0; rack_log_sad(rack, 3); /* Restart counting */ rack->r_ctl.sack_count = 0; rack->r_ctl.sack_moved_extra = 0; rack->r_ctl.sack_noextra_move = 1; rack->r_ctl.ack_count = max(1, (bytes_this_ack / segsiz)); if (rack->r_rep_reverse == 0) { rack->r_rep_reverse = 1; counter_u64_add(rack_sack_attacks_reversed, 1); } /* Restore the cwnd */ if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd) rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd; } } } } #endif static void rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end) { uint32_t am; if (SEQ_GT(end, start)) am = end - start; else am = 0; /* * We keep track of how many DSACK blocks we get * after a recovery incident. */ rack->r_ctl.dsack_byte_cnt += am; if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) && rack->r_ctl.retran_during_recovery && (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) { /* * False recovery most likely culprit is reordering. If * nothing else is missing we need to revert. */ rack->r_might_revert = 1; rack_handle_might_revert(rack->rc_tp, rack); rack->r_might_revert = 0; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; } } static void rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack) { /* Deal with changed and PRR here (in recovery only) */ uint32_t pipe, snd_una; rack->r_ctl.rc_prr_delivered += changed; if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) { /* * It is all outstanding, we are application limited * and thus we don't need more room to send anything. * Note we use tp->snd_una here and not th_ack because * the data as yet not been cut from the sb. */ rack->r_ctl.rc_prr_sndcnt = 0; return; } /* Compute prr_sndcnt */ if (SEQ_GT(tp->snd_una, th_ack)) { snd_una = tp->snd_una; } else { snd_una = th_ack; } pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt; if (pipe > tp->snd_ssthresh) { long sndcnt; sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh; if (rack->r_ctl.rc_prr_recovery_fs > 0) sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs; else { rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 9, 0); sndcnt = 0; } sndcnt++; if (sndcnt > (long)rack->r_ctl.rc_prr_out) sndcnt -= rack->r_ctl.rc_prr_out; else sndcnt = 0; rack->r_ctl.rc_prr_sndcnt = sndcnt; rack_log_to_prr(rack, 10, 0); } else { uint32_t limit; if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out) limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out); else limit = 0; if (changed > limit) limit = changed; limit += ctf_fixed_maxseg(tp); if (tp->snd_ssthresh > pipe) { rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit); rack_log_to_prr(rack, 11, 0); } else { rack->r_ctl.rc_prr_sndcnt = min(0, limit); rack_log_to_prr(rack, 12, 0); } } } static void rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck) { uint32_t changed; struct tcp_rack *rack; struct rack_sendmap *rsm; struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; register uint32_t th_ack; int32_t i, j, k, num_sack_blks = 0; uint32_t cts, acked, ack_point, sack_changed = 0; int loop_start = 0, moved_two = 0; uint32_t tsused; INP_WLOCK_ASSERT(tp->t_inpcb); if (th->th_flags & TH_RST) { /* We don't log resets */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; cts = tcp_get_usecs(NULL); rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); changed = 0; th_ack = th->th_ack; if (rack->sack_attack_disable == 0) rack_do_decay(rack); if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) { /* * You only get credit for * MSS and greater (and you get extra * credit for larger cum-ack moves). */ int ac; ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp); rack->r_ctl.ack_count += ac; counter_u64_add(rack_ack_total, ac); } if (rack->r_ctl.ack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } if (SEQ_GT(th_ack, tp->snd_una)) { rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__); tp->t_acktime = ticks; } if (rsm && SEQ_GT(th_ack, rsm->r_start)) changed = th_ack - rsm->r_start; if (changed) { rack_process_to_cumack(tp, rack, th_ack, cts, to); } if ((to->to_flags & TOF_SACK) == 0) { /* We are done nothing left and no sack. */ rack_handle_might_revert(tp, rack); /* * For cases where we struck a dup-ack * with no SACK, add to the changes so * PRR will work right. */ if (dup_ack_struck && (changed == 0)) { changed += ctf_fixed_maxseg(rack->rc_tp); } goto out; } /* Sack block processing */ if (SEQ_GT(th_ack, tp->snd_una)) ack_point = th_ack; else ack_point = tp->snd_una; for (i = 0; i < to->to_nsacks; i++) { bcopy((to->to_sacks + i * TCPOLEN_SACK), &sack, sizeof(sack)); sack.start = ntohl(sack.start); sack.end = ntohl(sack.end); if (SEQ_GT(sack.end, sack.start) && SEQ_GT(sack.start, ack_point) && SEQ_LT(sack.start, tp->snd_max) && SEQ_GT(sack.end, ack_point) && SEQ_LEQ(sack.end, tp->snd_max)) { sack_blocks[num_sack_blks] = sack; num_sack_blks++; #ifdef NETFLIX_STATS } else if (SEQ_LEQ(sack.start, th_ack) && SEQ_LEQ(sack.end, th_ack)) { /* * Its a D-SACK block. */ tcp_record_dsack(sack.start, sack.end); #endif rack_note_dsack(rack, sack.start, sack.end); } } /* * Sort the SACK blocks so we can update the rack scoreboard with * just one pass. */ num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks, num_sack_blks, th->th_ack); ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks); if (num_sack_blks == 0) { /* Nothing to sack (DSACKs?) */ goto out_with_totals; } if (num_sack_blks < 2) { /* Only one, we don't need to sort */ goto do_sack_work; } /* Sort the sacks */ for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { sack = sack_blocks[i]; sack_blocks[i] = sack_blocks[j]; sack_blocks[j] = sack; } } } /* * Now are any of the sack block ends the same (yes some * implementations send these)? */ again: if (num_sack_blks == 0) goto out_with_totals; if (num_sack_blks > 1) { for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (sack_blocks[i].end == sack_blocks[j].end) { /* * Ok these two have the same end we * want the smallest end and then * throw away the larger and start * again. */ if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { /* * The second block covers * more area use that */ sack_blocks[i].start = sack_blocks[j].start; } /* * Now collapse out the dup-sack and * lower the count */ for (k = (j + 1); k < num_sack_blks; k++) { sack_blocks[j].start = sack_blocks[k].start; sack_blocks[j].end = sack_blocks[k].end; j++; } num_sack_blks--; goto again; } } } } do_sack_work: /* * First lets look to see if * we have retransmitted and * can use the transmit next? */ rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm && SEQ_GT(sack_blocks[0].end, rsm->r_start) && SEQ_LT(sack_blocks[0].start, rsm->r_end)) { /* * We probably did the FR and the next * SACK in continues as we would expect. */ acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two); if (acked) { rack->r_wanted_output = 1; changed += acked; sack_changed += acked; } if (num_sack_blks == 1) { /* * This is what we would expect from * a normal implementation to happen * after we have retransmitted the FR, * i.e the sack-filter pushes down * to 1 block and the next to be retransmitted * is the sequence in the sack block (has more * are acked). Count this as ACK'd data to boost * up the chances of recovering any false positives. */ rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp)); counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp))); counter_u64_add(rack_express_sack, 1); if (rack->r_ctl.ack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } goto out_with_totals; } else { /* * Start the loop through the * rest of blocks, past the first block. */ moved_two = 0; loop_start = 1; } } /* Its a sack of some sort */ rack->r_ctl.sack_count++; if (rack->r_ctl.sack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } counter_u64_add(rack_sack_total, 1); if (rack->sack_attack_disable) { /* An attacker disablement is in place */ if (num_sack_blks > 1) { rack->r_ctl.sack_count += (num_sack_blks - 1); rack->r_ctl.sack_moved_extra++; counter_u64_add(rack_move_some, 1); if (rack->r_ctl.sack_moved_extra > 0xfff00000) { rack->r_ctl.sack_moved_extra /= 2; rack->r_ctl.sack_noextra_move /= 2; } } goto out; } rsm = rack->r_ctl.rc_sacklast; for (i = loop_start; i < num_sack_blks; i++) { acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two); if (acked) { rack->r_wanted_output = 1; changed += acked; sack_changed += acked; } if (moved_two) { /* * If we did not get a SACK for at least a MSS and * had to move at all, or if we moved more than our * threshold, it counts against the "extra" move. */ rack->r_ctl.sack_moved_extra += moved_two; counter_u64_add(rack_move_some, 1); } else { /* * else we did not have to move * any more than we would expect. */ rack->r_ctl.sack_noextra_move++; counter_u64_add(rack_move_none, 1); } if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) { /* * If the SACK was not a full MSS then * we add to sack_count the number of * MSS's (or possibly more than * a MSS if its a TSO send) we had to skip by. */ rack->r_ctl.sack_count += moved_two; counter_u64_add(rack_sack_total, moved_two); } /* * Now we need to setup for the next * round. First we make sure we won't * exceed the size of our uint32_t on * the various counts, and then clear out * moved_two. */ if ((rack->r_ctl.sack_moved_extra > 0xfff00000) || (rack->r_ctl.sack_noextra_move > 0xfff00000)) { rack->r_ctl.sack_moved_extra /= 2; rack->r_ctl.sack_noextra_move /= 2; } if (rack->r_ctl.sack_count > 0xfff00000) { rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } moved_two = 0; } out_with_totals: if (num_sack_blks > 1) { /* * You get an extra stroke if * you have more than one sack-blk, this * could be where we are skipping forward * and the sack-filter is still working, or * it could be an attacker constantly * moving us. */ rack->r_ctl.sack_moved_extra++; counter_u64_add(rack_move_some, 1); } out: #ifdef NETFLIX_EXP_DETECTION rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp)); #endif if (changed) { /* Something changed cancel the rack timer */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } tsused = tcp_get_usecs(NULL); rsm = tcp_rack_output(tp, rack, tsused); if ((!IN_FASTRECOVERY(tp->t_flags)) && rsm) { /* Enter recovery */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; entered_recovery = 1; rack_cong_signal(tp, CC_NDUPACK, tp->snd_una); /* * When we enter recovery we need to assure we send * one packet. */ if (rack->rack_no_prr == 0) { rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 8, 0); } rack->r_timer_override = 1; rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; } else if (IN_FASTRECOVERY(tp->t_flags) && rsm && (rack->r_rr_config == 3)) { /* * Assure we can output and we get no * remembered pace time except the retransmit. */ rack->r_timer_override = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; rack->r_ctl.rc_resend = rsm; } if (IN_FASTRECOVERY(tp->t_flags) && (rack->rack_no_prr == 0) && (entered_recovery == 0)) { rack_update_prr(tp, rack, changed, th_ack); if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) && ((rack->rc_inp->inp_in_hpts == 0) && ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) { /* * If you are pacing output you don't want * to override. */ rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; rack->r_timer_override = 1; } } } static void rack_strike_dupack(struct tcp_rack *rack) { struct rack_sendmap *rsm; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) { rsm = TAILQ_NEXT(rsm, r_tnext); } if (rsm && (rsm->r_dupack < 0xff)) { rsm->r_dupack++; if (rsm->r_dupack >= DUP_ACK_THRESHOLD) { struct timeval tv; uint32_t cts; /* * Here we see if we need to retransmit. For * a SACK type connection if enough time has passed * we will get a return of the rsm. For a non-sack * connection we will get the rsm returned if the * dupack value is 3 or more. */ cts = tcp_get_usecs(&tv); rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts); if (rack->r_ctl.rc_resend != NULL) { if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) { rack_cong_signal(rack->rc_tp, CC_NDUPACK, rack->rc_tp->snd_una); } rack->r_wanted_output = 1; rack->r_timer_override = 1; rack_log_retran_reason(rack, rsm, __LINE__, 1, 3); } } else { rack_log_retran_reason(rack, rsm, __LINE__, 0, 3); } } } static void rack_check_bottom_drag(struct tcpcb *tp, struct tcp_rack *rack, struct socket *so, int32_t acked) { uint32_t segsiz, minseg; segsiz = ctf_fixed_maxseg(tp); minseg = segsiz; if (tp->snd_max == tp->snd_una) { /* * We are doing dynamic pacing and we are way * under. Basically everything got acked while * we were still waiting on the pacer to expire. * * This means we need to boost the b/w in * addition to any earlier boosting of * the multipler. */ rack->rc_dragged_bottom = 1; rack_validate_multipliers_at_or_above100(rack); /* * Lets use the segment bytes acked plus * the lowest RTT seen as the basis to * form a b/w estimate. This will be off * due to the fact that the true estimate * should be around 1/2 the time of the RTT * but we can settle for that. */ if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) && acked) { uint64_t bw, calc_bw, rtt; rtt = rack->r_ctl.rack_rs.rs_us_rtt; if (rtt == 0) { /* no us sample is there a ms one? */ if (rack->r_ctl.rack_rs.rs_rtt_lowest) { rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; } else { goto no_measurement; } } bw = acked; calc_bw = bw * 1000000; calc_bw /= rtt; if (rack->r_ctl.last_max_bw && (rack->r_ctl.last_max_bw < calc_bw)) { /* * If we have a last calculated max bw * enforce it. */ calc_bw = rack->r_ctl.last_max_bw; } /* now plop it in */ if (rack->rc_gp_filled == 0) { if (calc_bw > ONE_POINT_TWO_MEG) { /* * If we have no measurement * don't let us set in more than * 1.2Mbps. If we are still too * low after pacing with this we * will hopefully have a max b/w * available to sanity check things. */ calc_bw = ONE_POINT_TWO_MEG; } rack->r_ctl.rc_rtt_diff = 0; rack->r_ctl.gp_bw = calc_bw; rack->rc_gp_filled = 1; if (rack->r_ctl.num_measurements < RACK_REQ_AVG) rack->r_ctl.num_measurements = RACK_REQ_AVG; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else if (calc_bw > rack->r_ctl.gp_bw) { rack->r_ctl.rc_rtt_diff = 0; if (rack->r_ctl.num_measurements < RACK_REQ_AVG) rack->r_ctl.num_measurements = RACK_REQ_AVG; rack->r_ctl.gp_bw = calc_bw; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else rack_increase_bw_mul(rack, -1, 0, 0, 1); if ((rack->gp_ready == 0) && (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) { /* We have enough measurements now */ rack->gp_ready = 1; rack_set_cc_pacing(rack); if (rack->defer_options) rack_apply_deferred_options(rack); } /* * For acks over 1mss we do a extra boost to simulate * where we would get 2 acks (we want 110 for the mul). */ if (acked > segsiz) rack_increase_bw_mul(rack, -1, 0, 0, 1); } else { /* * zero rtt possibly?, settle for just an old increase. */ no_measurement: rack_increase_bw_mul(rack, -1, 0, 0, 1); } } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) && (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)), minseg)) && (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) && (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) && (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <= (segsiz * rack_req_segs))) { /* * We are doing dynamic GP pacing and * we have everything except 1MSS or less * bytes left out. We are still pacing away. * And there is data that could be sent, This * means we are inserting delayed ack time in * our measurements because we are pacing too slow. */ rack_validate_multipliers_at_or_above100(rack); rack->rc_dragged_bottom = 1; rack_increase_bw_mul(rack, -1, 0, 0, 1); } } static void rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount) { /* * The fast output path is enabled and we * have moved the cumack forward. Lets see if * we can expand forward the fast path length by * that amount. What we would ideally like to * do is increase the number of bytes in the * fast path block (left_to_send) by the * acked amount. However we have to gate that * by two factors: * 1) The amount outstanding and the rwnd of the peer * (i.e. we don't want to exceed the rwnd of the peer). * * 2) The amount of data left in the socket buffer (i.e. * we can't send beyond what is in the buffer). * * Note that this does not take into account any increase * in the cwnd. We will only extend the fast path by * what was acked. */ uint32_t new_total, gating_val; new_total = acked_amount + rack->r_ctl.fsb.left_to_send; gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)), (tp->snd_wnd - (tp->snd_max - tp->snd_una))); if (new_total <= gating_val) { /* We can increase left_to_send by the acked amount */ counter_u64_add(rack_extended_rfo, 1); rack->r_ctl.fsb.left_to_send = new_total; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(&rack->rc_inp->inp_socket->so_snd), (tp->snd_max - tp->snd_una))); } } static void rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una) { /* * Here any sendmap entry that points to the * beginning mbuf must be adjusted to the correct * offset. This must be called with: * 1) The socket buffer locked * 2) snd_una adjusted to its new postion. * * Note that (2) implies rack_ack_received has also * been called. * * We grab the first mbuf in the socket buffer and * then go through the front of the sendmap, recalculating * the stored offset for any sendmap entry that has * that mbuf. We must use the sb functions to do this * since its possible an add was done has well as * the subtraction we may have just completed. This should * not be a penalty though, since we just referenced the sb * to go in and trim off the mbufs that we freed (of course * there will be a penalty for the sendmap references though). */ struct mbuf *m; struct rack_sendmap *rsm; SOCKBUF_LOCK_ASSERT(sb); m = sb->sb_mb; rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if ((rsm == NULL) || (m == NULL)) { /* Nothing outstanding */ return; } while (rsm->m && (rsm->m == m)) { /* one to adjust */ #ifdef INVARIANTS struct mbuf *tm; uint32_t soff; tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff); if (rsm->orig_m_len != m->m_len) { rack_adjust_orig_mlen(rsm); } if (rsm->soff != soff) { /* * This is not a fatal error, we anticipate it * might happen (the else code), so we count it here * so that under invariant we can see that it really * does happen. */ counter_u64_add(rack_adjust_map_bw, 1); } rsm->m = tm; rsm->soff = soff; if (tm) rsm->orig_m_len = rsm->m->m_len; else rsm->orig_m_len = 0; #else rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff); if (rsm->m) rsm->orig_m_len = rsm->m->m_len; else rsm->orig_m_len = 0; #endif rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (rsm == NULL) break; } } /* * Return value of 1, we do not need to call rack_process_data(). * return value of 0, rack_process_data can be called. * For ret_val if its 0 the TCP is locked, if its non-zero * its unlocked and probably unsafe to touch the TCB. */ static int rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t *ret_val) { int32_t ourfinisacked = 0; int32_t nsegs, acked_amount; int32_t acked; struct mbuf *mfree; struct tcp_rack *rack; int32_t under_pacing = 0; int32_t recovery = 0; rack = (struct tcp_rack *)tp->t_fb_ptr; if (SEQ_GT(th->th_ack, tp->snd_max)) { __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt); rack->r_wanted_output = 1; return (1); } if (rack->gp_ready && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { under_pacing = 1; } if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { int in_rec, dup_ack_struck = 0; in_rec = IN_FASTRECOVERY(tp->t_flags); if (rack->rc_in_persist) { tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); } if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) { rack_strike_dupack(rack); dup_ack_struck = 1; } rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck); } if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* * Old ack, behind (or duplicate to) the last one rcv'd * Note: We mark reordering is occuring if its * less than and we have not closed our window. */ if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) { counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); } return (0); } /* * If we reach this point, ACK is not a duplicate, i.e., it ACKs * something we sent. */ if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our SYN has * been ACK'd (so connection is now fully synchronized). Go * to non-starred state, increment snd_una for ACK of SYN, * and check if we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; /* Send window already scaled. */ } } nsegs = max(1, m->m_pkthdr.lro_nsegs); INP_WLOCK_ASSERT(tp->t_inpcb); acked = BYTES_THIS_ACK(tp, th); KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); /* * If we just performed our first retransmit, and the ACK arrives * within our recovery window, then it was a mistake to do the * retransmit in the first place. Recover our original cwnd and * ssthresh, and proceed to transmit where we left off. */ if ((tp->t_flags & TF_PREVVALID) && ((tp->t_flags & TF_RCVD_TSTMP) == 0)) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, CC_RTO_ERR, th->th_ack); } if (acked) { /* assure we are not backed off */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; /* * If it is the RXT timer we want to * stop it, so we can restart a TLP. */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); #ifdef NETFLIX_HTTP_LOGGING tcp_http_check_for_comp(rack->rc_tp, th->th_ack); #endif } /* * If we have a timestamp reply, update smoothed round trip time. If * no timestamp is present but transmit timer is running and timed * sequence number was acked, update smoothed round trip time. Since * we now have an rtt measurement, cancel the timer backoff (cf., * Phil Karn's retransmit alg.). Recompute the initial retransmit * timer. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ /* * If all outstanding data is acked, stop retransmit timer and * remember to restart (more output or persist). If there is more * data to be acked, restart retransmit timer, using current * (possibly backed-off) value. */ if (acked == 0) { if (ofia) *ofia = ourfinisacked; return (0); } if (IN_RECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover) && (SEQ_LT(th->th_ack, tp->snd_max))) { tcp_rack_partialack(tp); } else { rack_post_recovery(tp, th->th_ack); recovery = 1; } } /* * Let the congestion control algorithm update congestion control * related information. This typically means increasing the * congestion window. */ rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery); SOCKBUF_LOCK(&so->so_snd); acked_amount = min(acked, (int)sbavail(&so->so_snd)); tp->snd_wnd -= acked_amount; mfree = sbcut_locked(&so->so_snd, acked_amount); if ((sbused(&so->so_snd) == 0) && (acked > acked_amount) && (tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_SENTFIN)) { /* * We must be sure our fin * was sent and acked (we can be * in FIN_WAIT_1 without having * sent the fin). */ ourfinisacked = 1; } tp->snd_una = th->th_ack; if (acked_amount && sbavail(&so->so_snd)) rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una); rack_log_wakeup(tp,rack, &so->so_snd, acked, 2); /* NB: sowwakeup_locked() does an implicit unlock. */ sowwakeup_locked(so); m_freem(mfree); if (SEQ_GT(tp->snd_una, tp->snd_recover)) tp->snd_recover = tp->snd_una; if (SEQ_LT(tp->snd_nxt, tp->snd_una)) { tp->snd_nxt = tp->snd_una; } if (under_pacing && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && rack->rc_gp_dyn_mul && rack->rc_always_pace) { /* Check if we are dragging bottom */ rack_check_bottom_drag(tp, rack, so, acked); } if (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ tp->t_flags &= ~TF_PREVVALID; rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) tp->t_acktime = 0; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); /* Set need output so persist might get set */ rack->r_wanted_output = 1; sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (sbavail(&so->so_snd) == 0) && (tp->t_flags2 & TF2_DROP_AF_DATA)) { /* * The socket was gone and the * peer sent data (now or in the past), time to * reset him. */ *ret_val = 1; /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); return (1); } } if (ofia) *ofia = ourfinisacked; return (0); } static void rack_collapsed_window(struct tcp_rack *rack) { /* * Now we must walk the * send map and divide the * ones left stranded. These * guys can't cause us to abort * the connection and are really * "unsent". However if a buggy * client actually did keep some * of the data i.e. collapsed the win * and refused to ack and then opened * the win and acked that data. We would * get into an ack war, the simplier * method then of just pretending we * did not send those segments something * won't work. */ struct rack_sendmap *rsm, *nrsm, fe, *insret; tcp_seq max_seq; max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd; memset(&fe, 0, sizeof(fe)); fe.r_start = max_seq; /* Find the first seq past or at maxseq */ rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm == NULL) { /* Nothing to do strange */ rack->rc_has_collapsed = 0; return; } /* * Now do we need to split at * the collapse point? */ if (SEQ_GT(max_seq, rsm->r_start)) { nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* We can't get a rsm, mark all? */ nrsm = rsm; goto no_split; } /* Clone it */ rack_clone_rsm(rack, nrsm, rsm, max_seq); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } /* * Set in the new RSM as the * collapsed starting point */ rsm = nrsm; } no_split: counter_u64_add(rack_collapsed_win, 1); RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) { nrsm->r_flags |= RACK_RWND_COLLAPSED; } rack->rc_has_collapsed = 1; } static void rack_un_collapse_window(struct tcp_rack *rack) { struct rack_sendmap *rsm; RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if (rsm->r_flags & RACK_RWND_COLLAPSED) rsm->r_flags &= ~RACK_RWND_COLLAPSED; else break; } rack->rc_has_collapsed = 0; } static void rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack, int32_t tlen, int32_t tfo_syn) { if (DELAY_ACK(tp, tlen) || tfo_syn) { if (rack->rc_dack_mode && (tlen > 500) && (rack->rc_dack_toggle == 1)) { goto no_delayed_ack; } rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { no_delayed_ack: rack->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; if (rack->rc_dack_mode) { if (tp->t_flags & TF_DELACK) rack->rc_dack_toggle = 1; else rack->rc_dack_toggle = 0; } } } static void rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack) { /* * If fast output is in progress, lets validate that * the new window did not shrink on us and make it * so fast output should end. */ if (rack->r_fast_output) { uint32_t out; /* * Calculate what we will send if left as is * and compare that to our send window. */ out = ctf_outstanding(tp); if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) { /* ok we have an issue */ if (out >= tp->snd_wnd) { /* Turn off fast output the window is met or collapsed */ rack->r_fast_output = 0; } else { /* we have some room left */ rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out; if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) { /* If not at least 1 full segment never mind */ rack->r_fast_output = 0; } } } } } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { /* * Update window information. Don't look at window if no ACK: TAC's * send garbage on first SYN. */ int32_t nsegs; int32_t tfo_syn; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); nsegs = max(1, m->m_pkthdr.lro_nsegs); if ((thflags & TH_ACK) && (SEQ_LT(tp->snd_wl1, th->th_seq) || (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if (tlen == 0 && tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) KMOD_TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; rack->r_wanted_output = 1; } else if (thflags & TH_ACK) { if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; } } if (tp->snd_wnd < ctf_outstanding(tp)) /* The peer collapsed the window */ rack_collapsed_window(rack); else if (rack->rc_has_collapsed) rack_un_collapse_window(rack); /* Was persist timer active and now we have window space? */ if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs))) { rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime); tp->snd_nxt = tp->snd_max; /* Make sure we output to start the timer */ rack->r_wanted_output = 1; } /* Do we enter persists? */ if ((rack->rc_in_persist == 0) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* * Here the rwnd is less than * the pacing size, we are established, * nothing is outstanding, and there is * data to send. Enter persists. */ rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } if (tp->t_flags2 & TF2_DROP_AF_DATA) { m_freem(m); return (0); } /* * don't process the URG bit, ignore them drag * along the up. */ tp->rcv_up = tp->rcv_nxt; INP_WLOCK_ASSERT(tp->t_inpcb); /* * Process the segment text, merging it into the TCP sequencing * queue, and arranging for acknowledgment of receipt if necessary. * This process logically involves adjusting tp->rcv_wnd as data is * presented to the user (this happens in tcp_usrreq.c, case * PRU_RCVD). If a FIN has already been received on this connection * then we just ignore the text. */ tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && IS_FASTOPEN(tp->t_flags)); if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && TCPS_HAVERCVDFIN(tp->t_state) == 0) { tcp_seq save_start = th->th_seq; tcp_seq save_rnxt = tp->rcv_nxt; int save_tlen = tlen; m_adj(m, drop_hdrlen); /* delayed header drop */ /* * Insert segment which includes th into TCP reassembly * queue with control block tp. Set thflags to whether * reassembly now includes a segment with FIN. This handles * the common case inline (segment is the next to be * received on an established connection, and the queue is * empty), avoiding linkage into and removal from the queue * and repetition of various conversions. Set DELACK for * segments received in order, but ack immediately when * segments are out of order (so fast retransmit can work). */ if (th->th_seq == tp->rcv_nxt && SEGQ_EMPTY(tp) && (TCPS_HAVEESTABLISHED(tp->t_state) || tfo_syn)) { #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (0); } } #endif rack_handle_delayed_ack(tp, rack, tlen, tfo_syn); tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } thflags = th->th_flags & TH_FIN; KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && appended != mcnt) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif } else { /* * XXX: Due to the header drop above "th" is * theoretically invalid by now. Fortunately * m_adj() doesn't actually frees any mbufs when * trimming from the head. */ tcp_seq temp = save_start; thflags = tcp_reass(tp, th, &temp, &tlen, m); tp->t_flags |= TF_ACKNOW; if (tp->t_flags & TF_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } } if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { /* * DSACK actually handled in the fastpath * above. */ RACK_OPTS_INC(tcp_sack_path_1); tcp_update_sack_list(tp, save_start, save_start + save_tlen); } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { if ((tp->rcv_numsacks >= 1) && (tp->sackblks[0].end == save_start)) { /* * Partial overlap, recorded at todrop * above. */ RACK_OPTS_INC(tcp_sack_path_2a); tcp_update_sack_list(tp, tp->sackblks[0].start, tp->sackblks[0].end); } else { RACK_OPTS_INC(tcp_sack_path_2b); tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } } else if (tlen >= save_tlen) { /* Update of sackblks. */ RACK_OPTS_INC(tcp_sack_path_3); tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } else if (tlen > 0) { RACK_OPTS_INC(tcp_sack_path_4); tcp_update_dsack_list(tp, save_start, save_start + tlen); } } } else { m_freem(m); thflags &= ~TH_FIN; } /* * If FIN is received ACK the FIN and let the user know that the * connection is closing. */ if (thflags & TH_FIN) { if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { /* The socket upcall is handled by socantrcvmore. */ socantrcvmore(so); /* * If connection is half-synchronized (ie NEEDSYN * flag on) then delay ACK, so it may be piggybacked * when SYN is sent. Otherwise, since we received a * FIN then no more input can be expected, send ACK * now. */ if (tp->t_flags & TF_NEEDSYN) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt++; } switch (tp->t_state) { /* * In SYN_RECEIVED and ESTABLISHED STATES enter the * CLOSE_WAIT state. */ case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_state_change(tp, TCPS_CLOSE_WAIT); break; /* * If still in FIN_WAIT_1 STATE FIN has not been * acked so enter the CLOSING state. */ case TCPS_FIN_WAIT_1: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_state_change(tp, TCPS_CLOSING); break; /* * In FIN_WAIT_2 state enter the TIME_WAIT state, * starting the time-wait timer, turning off the * other standard timers. */ case TCPS_FIN_WAIT_2: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_twstart(tp); return (1); } } /* * Return any desired output. */ if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) { rack->r_wanted_output = 1; } INP_WLOCK_ASSERT(tp->t_inpcb); return (0); } /* * Here nothing is really faster, its just that we * have broken out the fast-data path also just like * the fast-ack. */ static int rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) { int32_t nsegs; int32_t newsize = 0; /* automatic sockbuf scaling */ struct tcp_rack *rack; #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; #endif #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if (__predict_false(th->th_seq != tp->rcv_nxt)) { return (0); } if (__predict_false(tp->snd_nxt != tp->snd_max)) { return (0); } if (tiwin && tiwin != tp->snd_wnd) { return (0); } if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { return (0); } if (__predict_false((to->to_flags & TOF_TS) && (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { return (0); } if (__predict_false((th->th_ack != tp->snd_una))) { return (0); } if (__predict_false(tlen > sbspace(&so->so_rcv))) { return (0); } if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } rack = (struct tcp_rack *)tp->t_fb_ptr; /* * This is a pure, in-sequence data packet with nothing on the * reassembly queue and we have enough buffer space to take it. */ nsegs = max(1, m->m_pkthdr.lro_nsegs); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (1); } } #endif /* Clean receiver SACK report if present */ if (tp->rcv_numsacks) tcp_clean_sackreport(tp); KMOD_TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } /* * Pull snd_wl1 up to prevent seq wrap relative to th_seq. */ tp->snd_wl1 = th->th_seq; /* * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. */ tp->rcv_up = tp->rcv_nxt; KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else { /* * Set new socket buffer size. Give up when limit is * reached. */ if (newsize) if (!sbreserve_locked(&so->so_rcv, newsize, so, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); ctf_calc_rwin(so, tp); } rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && mcnt != appended) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif rack_handle_delayed_ack(tp, rack, tlen, 0); if (tp->snd_una == tp->snd_max) sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); return (1); } /* * This subfunction is used to try to highly optimize the * fast path. We again allow window updates that are * in sequence to remain in the fast-path. We also add * in the __predict's to attempt to help the compiler. * Note that if we return a 0, then we can *not* process * it and the caller should push the packet into the * slow-path. */ static int rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint32_t cts) { int32_t acked; int32_t nsegs; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif int32_t under_pacing = 0; struct tcp_rack *rack; if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* Old ack, behind (or duplicate to) the last one rcv'd */ return (0); } if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { /* Above what we have sent? */ return (0); } if (__predict_false(tp->snd_nxt != tp->snd_max)) { /* We are retransmitting */ return (0); } if (__predict_false(tiwin == 0)) { /* zero window */ return (0); } if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { /* We need a SYN or a FIN, unlikely.. */ return (0); } if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { /* Timestamp is behind .. old ack with seq wrap? */ return (0); } if (__predict_false(IN_RECOVERY(tp->t_flags))) { /* Still recovering */ return (0); } rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->r_ctl.rc_sacked) { /* We have sack holes on our scoreboard */ return (0); } /* Ok if we reach here, we can process a fast-ack */ if (rack->gp_ready && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { under_pacing = 1; } nsegs = max(1, m->m_pkthdr.lro_nsegs); rack_log_ack(tp, to, th, 0, 0); /* Did the window get updated? */ if (tiwin != tp->snd_wnd) { tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = th->th_seq; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; } /* Do we exit persists? */ if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs))) { rack_exit_persist(tp, rack, cts); } /* Do we enter persists? */ if ((rack->rc_in_persist == 0) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* * Here the rwnd is less than * the pacing size, we are established, * nothing is outstanding, and there is * data to send. Enter persists. */ rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * This is a pure ack for outstanding data. */ KMOD_TCPSTAT_INC(tcps_predack); /* * "bad retransmit" recovery. */ if ((tp->t_flags & TF_PREVVALID) && ((tp->t_flags & TF_RCVD_TSTMP) == 0)) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, CC_RTO_ERR, th->th_ack); } /* * Recalculate the transmit timer / rtt. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ acked = BYTES_THIS_ACK(tp, th); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, to); #endif KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); if (acked) { struct mbuf *mfree; rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0); SOCKBUF_LOCK(&so->so_snd); mfree = sbcut_locked(&so->so_snd, acked); tp->snd_una = th->th_ack; /* Note we want to hold the sb lock through the sendmap adjust */ rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una); /* Wake up the socket if we have room to write more */ rack_log_wakeup(tp,rack, &so->so_snd, acked, 2); sowwakeup_locked(so); m_freem(mfree); tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; /* * If it is the RXT timer we want to * stop it, so we can restart a TLP. */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); #ifdef NETFLIX_HTTP_LOGGING tcp_http_check_for_comp(rack->rc_tp, th->th_ack); #endif } /* * Let the congestion control algorithm update congestion control * related information. This typically means increasing the * congestion window. */ if (tp->snd_wnd < ctf_outstanding(tp)) { /* The peer collapsed the window */ rack_collapsed_window(rack); } else if (rack->rc_has_collapsed) rack_un_collapse_window(rack); /* * Pull snd_wl2 up to prevent seq wrap relative to th_ack. */ tp->snd_wl2 = th->th_ack; tp->t_dupacks = 0; m_freem(m); /* ND6_HINT(tp); *//* Some progress has been made. */ /* * If all outstanding data are acked, stop retransmit timer, * otherwise restart timer using current (possibly backed-off) * value. If process is waiting for space, wakeup/selwakeup/signal. * If data are ready to send, let tcp_output decide between more * output or persist. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif if (under_pacing && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && rack->rc_gp_dyn_mul && rack->rc_always_pace) { /* Check if we are dragging bottom */ rack_check_bottom_drag(tp, rack, so, acked); } if (tp->snd_una == tp->snd_max) { tp->t_flags &= ~TF_PREVVALID; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) tp->t_acktime = 0; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } if (acked && rack->r_fast_output) rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked); if (sbavail(&so->so_snd)) { rack->r_wanted_output = 1; } return (1); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t todrop; int32_t ourfinisacked = 0; struct tcp_rack *rack; ctf_calc_rwin(so, tp); /* * If the state is SYN_SENT: if seg contains an ACK, but not for our * SYN, drop the input. if seg contains a RST, then drop the * connection. if seg does not contain SYN, then drop it. Otherwise * this is an acceptable SYN segment initialize tp->rcv_nxt and * tp->irs if seg contains ack then advance tp->snd_una if seg * contains an ECE and ECN support is enabled, the stream is ECN * capable. if SYN has been acked change to ESTABLISHED else * SYN_RCVD state arrange for segment to be acked (eventually) * continue processing rest of data/controls. */ if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { TCP_PROBE5(connect__refused, NULL, tp, mtod(m, const char *), tp, th); tp = tcp_drop(tp, ECONNREFUSED); ctf_do_drop(m, tp); return (1); } if (thflags & TH_RST) { ctf_do_drop(m, tp); return (1); } if (!(thflags & TH_SYN)) { ctf_do_drop(m, tp); return (1); } tp->irs = th->th_seq; tcp_rcvseqinit(tp); rack = (struct tcp_rack *)tp->t_fb_ptr; if (thflags & TH_ACK) { int tfo_partial = 0; KMOD_TCPSTAT_INC(tcps_connects); soisconnected(so); #ifdef MAC mac_socketpeer_set_from_mbuf(m, so); #endif /* Do window scaling on this connection? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } tp->rcv_adv += min(tp->rcv_wnd, TCP_MAXWIN << tp->rcv_scale); /* * If not all the data that was sent in the TFO SYN * has been acked, resend the remainder right away. */ if (IS_FASTOPEN(tp->t_flags) && (tp->snd_una != tp->snd_max)) { tp->snd_nxt = th->th_ack; tfo_partial = 1; } /* * If there's data, delay ACK; if there's also a FIN ACKNOW * will be turned on later. */ if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { rack->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; rack->rc_dack_toggle = 0; } if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) && (V_tcp_do_ecn == 1)) { tp->t_flags2 |= TF2_ECN_PERMIT; KMOD_TCPSTAT_INC(tcps_ecn_shs); } if (SEQ_GT(th->th_ack, tp->snd_una)) { /* * We advance snd_una for the * fast open case. If th_ack is * acknowledging data beyond * snd_una we can't just call * ack-processing since the * data stream in our send-map * will start at snd_una + 1 (one * beyond the SYN). If its just * equal we don't need to do that * and there is no send_map. */ tp->snd_una++; } /* * Received in SYN_SENT[*] state. Transitions: * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; thflags &= ~TH_SYN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(connect__established, NULL, tp, mtod(m, const char *), tp, th); rack_cc_conn_init(tp); } } else { /* * Received initial SYN in SYN-SENT[*] state => simultaneous * open. If segment contains CC option and there is a * cached CC, apply TAO test. If it succeeds, connection is * * half-synchronized. Otherwise, do 3-way handshake: * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If * there was no CC option, clear cached CC value. */ tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); tcp_state_change(tp, TCPS_SYN_RECEIVED); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * Advance th->th_seq to correspond to first data byte. If data, * trim to stay within window, dropping FIN if necessary. */ th->th_seq++; if (tlen > tp->rcv_wnd) { todrop = tlen - tp->rcv_wnd; m_adj(m, -todrop); tlen = tp->rcv_wnd; thflags &= ~TH_FIN; KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); } tp->snd_wl1 = th->th_seq - 1; tp->rcv_up = th->th_seq; /* * Client side of transaction: already sent SYN and data. If the * remote host used T/TCP to validate the SYN, our data will be * ACK'd; if so, enter normal data segment processing in the middle * of step 5, ack processing. Otherwise, goto step 6. */ if (thflags & TH_ACK) { /* For syn-sent we need to possibly update the rtt */ if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) { uint32_t t, mcts; mcts = tcp_ts_getticks(); t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4); tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2); tcp_rack_xmit_timer_commit(rack, tp); } if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) return (ret_val); /* We may have changed to FIN_WAIT_1 above */ if (tp->t_state == TCPS_FIN_WAIT_1) { /* * In FIN_WAIT_1 STATE in addition to the processing * for the ESTABLISHED state if our FIN is now * acknowledged then enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then * closing user can proceed. Starting the * timer is contrary to the specification, * but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and * use a compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { struct tcp_rack *rack; int32_t ret_val = 0; int32_t ourfinisacked = 0; ctf_calc_rwin(so, tp); if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } rack = (struct tcp_rack *)tp->t_fb_ptr; if (IS_FASTOPEN(tp->t_flags)) { /* * When a TFO connection is in SYN_RECEIVED, the * only valid packets are the initial SYN, a * retransmit/copy of the initial SYN (possibly with * a subset of the original data), a valid ACK, a * FIN, or a RST. */ if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } else if (thflags & TH_SYN) { /* non-initial SYN is ignored */ if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { ctf_do_drop(m, NULL); return (0); } } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { ctf_do_drop(m, NULL); return (0); } } if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } /* * In the SYN-RECEIVED state, validate that the packet belongs to * this connection before trimming the data to fit the receive * window. Check the sequence number versus IRS since we know the * sequence numbers haven't wrapped. This is a partial fix for the * "LAND" DoS attack. */ if (SEQ_LT(th->th_seq, tp->irs)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (IS_FASTOPEN(tp->t_flags)) { rack_cc_conn_init(tp); } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } KMOD_TCPSTAT_INC(tcps_connects); soisconnected(so); /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } /* * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> * FIN-WAIT-1 */ tp->t_starttime = ticks; if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; } if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(accept__established, NULL, tp, mtod(m, const char *), tp, th); /* * TFO connections call cc_conn_init() during SYN * processing. Calling it again here for such connections * is not harmless as it would undo the snd_cwnd reduction * that occurs when a TFO SYN|ACK is retransmitted. */ if (!IS_FASTOPEN(tp->t_flags)) rack_cc_conn_init(tp); } /* * Account for the ACK of our SYN prior to * regular ACK processing below, except for * simultaneous SYN, which is handled later. */ if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) tp->snd_una++; /* * If segment contains data or ACK, will call tcp_reass() later; if * not, do so now to pass queued data to user. */ if (tlen == 0 && (thflags & TH_FIN) == 0) { (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0, (struct mbuf *)0); if (tp->t_flags & TF_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } } tp->snd_wl1 = th->th_seq - 1; /* For syn-recv we need to possibly update the rtt */ if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) { uint32_t t, mcts; mcts = tcp_ts_getticks(); t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5); tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2); tcp_rack_xmit_timer_commit(rack, tp); } if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (tp->t_state == TCPS_FIN_WAIT_1) { /* We could have went to FIN_WAIT_1 (or EST) above */ /* * In FIN_WAIT_1 STATE in addition to the processing for the * ESTABLISHED state if our FIN is now acknowledged then * enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then closing * user can proceed. Starting the timer is contrary * to the specification, but if we don't get a FIN * we'll hang forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; struct tcp_rack *rack; /* * Header prediction: check for the two common cases of a * uni-directional data xfer. If the packet has no control flags, * is in-sequence, the window didn't change and we're not * retransmitting, it's a candidate. If the length is zero and the * ack moved forward, we're the sender side of the xfer. Just free * the data acked & wake any higher level process that was blocked * waiting for space. If the length is non-zero and the ack didn't * move, we're the receiver side. If we're getting packets in-order * (the reassembly queue is empty), add the data toc The socket * buffer and note that we need a delayed ack. Make sure that the * hidden state-flags are also off. Since we check for * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. */ rack = (struct tcp_rack *)tp->t_fb_ptr; if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) && __predict_true(SEGQ_EMPTY(tp)) && __predict_true(th->th_seq == tp->rcv_nxt)) { if (tlen == 0) { if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) { return (0); } } else { if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt, iptos)) { return (0); } } } ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } /* State changes only happen in rack_process_data() */ return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static int rack_check_data_after_close(struct mbuf *m, struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->rc_allow_data_af_clo == 0) { close_now: tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); KMOD_TCPSTAT_INC(tcps_rcvafterclose); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); return (1); } if (sbavail(&so->so_snd) == 0) goto close_now; /* Ok we allow data that is ignored and a followup reset */ tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); tp->rcv_nxt = th->th_seq + *tlen; tp->t_flags2 |= TF2_DROP_AF_DATA; rack->r_wanted_output = 1; *tlen = 0; return (0); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { /* * If we can't receive any more data, then closing user can * proceed. Starting the timer is contrary to the * specification, but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tcp_twstart(tp); m_freem(m); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * case TCPS_LAST_ACK: Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tp = tcp_close(tp); ctf_do_drop(m, tp); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); /* Reset receive buffer auto scaling when not in bulk receive mode. */ if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static void inline rack_clear_rate_sample(struct tcp_rack *rack) { rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY; rack->r_ctl.rack_rs.rs_rtt_cnt = 0; rack->r_ctl.rack_rs.rs_rtt_tot = 0; } static void rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override) { uint64_t bw_est, rate_wanted; int chged = 0; uint32_t user_max, orig_min, orig_max; orig_min = rack->r_ctl.rc_pace_min_segs; orig_max = rack->r_ctl.rc_pace_max_segs; user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs; if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs) chged = 1; rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp); if (rack->use_fixed_rate || rack->rc_force_max_seg) { if (user_max != rack->r_ctl.rc_pace_max_segs) chged = 1; } if (rack->rc_force_max_seg) { rack->r_ctl.rc_pace_max_segs = user_max; } else if (rack->use_fixed_rate) { bw_est = rack_get_bw(rack); if ((rack->r_ctl.crte == NULL) || (bw_est != rack->r_ctl.crte->rate)) { rack->r_ctl.rc_pace_max_segs = user_max; } else { /* We are pacing right at the hardware rate */ uint32_t segsiz; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size( tp, bw_est, segsiz, 0, rack->r_ctl.crte, NULL); } } else if (rack->rc_always_pace) { if (rack->r_ctl.gp_bw || #ifdef NETFLIX_PEAKRATE rack->rc_tp->t_maxpeakrate || #endif rack->r_ctl.init_rate) { /* We have a rate of some sort set */ uint32_t orig; bw_est = rack_get_bw(rack); orig = rack->r_ctl.rc_pace_max_segs; if (fill_override) rate_wanted = *fill_override; else rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL); if (rate_wanted) { /* We have something */ rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rate_wanted, ctf_fixed_maxseg(rack->rc_tp)); } else rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs; if (orig != rack->r_ctl.rc_pace_max_segs) chged = 1; } else if ((rack->r_ctl.gp_bw == 0) && (rack->r_ctl.rc_pace_max_segs == 0)) { /* * If we have nothing limit us to bursting * out IW sized pieces. */ chged = 1; rack->r_ctl.rc_pace_max_segs = rc_init_window(rack); } } if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) { chged = 1; rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES; } if (chged) rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2); } static void rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack) { #ifdef INET6 struct ip6_hdr *ip6 = NULL; #endif #ifdef INET struct ip *ip = NULL; #endif struct udphdr *udp = NULL; /* Ok lets fill in the fast block, it can only be used with no IP options! */ #ifdef INET6 if (rack->r_is_v6) { rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; if (tp->t_port) { rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr); udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; rack->r_ctl.fsb.udp = udp; rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1); } else { rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1); rack->r_ctl.fsb.udp = NULL; } tcpip_fillheaders(rack->rc_inp, tp->t_port, ip6, rack->r_ctl.fsb.th); } else #endif /* INET6 */ { rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr); ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; if (tp->t_port) { rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr); udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; rack->r_ctl.fsb.udp = udp; rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1); } else { rack->r_ctl.fsb.udp = NULL; rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1); } tcpip_fillheaders(rack->rc_inp, tp->t_port, ip, rack->r_ctl.fsb.th); } rack->r_fsb_inited = 1; } static int rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack) { /* * Allocate the larger of spaces V6 if available else just * V4 and include udphdr (overbook) */ #ifdef INET6 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr); #else rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr); #endif rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len, M_TCPFSB, M_NOWAIT|M_ZERO); if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) { return (ENOMEM); } rack->r_fsb_inited = 0; return (0); } static int rack_init(struct tcpcb *tp) { struct tcp_rack *rack = NULL; struct rack_sendmap *insret; uint32_t iwin, snt, us_cts; int err; tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT); if (tp->t_fb_ptr == NULL) { /* * We need to allocate memory but cant. The INP and INP_INFO * locks and they are recusive (happens during setup. So a * scheme to drop the locks fails :( * */ return (ENOMEM); } memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack)); rack = (struct tcp_rack *)tp->t_fb_ptr; RB_INIT(&rack->r_ctl.rc_mtree); TAILQ_INIT(&rack->r_ctl.rc_free); TAILQ_INIT(&rack->r_ctl.rc_tmap); rack->rc_tp = tp; rack->rc_inp = tp->t_inpcb; /* Set the flag */ rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; /* Probably not needed but lets be sure */ rack_clear_rate_sample(rack); /* * Save off the default values, socket options will poke * at these if pacing is not on or we have not yet * reached where pacing is on (gp_ready/fixed enabled). * When they get set into the CC module (when gp_ready * is enabled or we enable fixed) then we will set these * values into the CC and place in here the old values * so we have a restoral. Then we will set the flag * rc_pacing_cc_set. That way whenever we turn off pacing * or switch off this stack, we will know to go restore * the saved values. */ rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn; rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn; /* We want abe like behavior as well */ rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN; rack->r_ctl.rc_reorder_fade = rack_reorder_fade; rack->rc_allow_data_af_clo = rack_ignore_data_after_close; rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh; if (use_rack_rr) rack->use_rack_rr = 1; if (V_tcp_delack_enabled) tp->t_delayed_ack = 1; else tp->t_delayed_ack = 0; #ifdef TCP_ACCOUNTING if (rack_tcp_accounting) { tp->t_flags2 |= TF2_TCP_ACCOUNTING; } #endif if (rack_enable_shared_cwnd) rack->rack_enable_scwnd = 1; rack->rc_user_set_max_segs = rack_hptsi_segments; rack->rc_force_max_seg = 0; if (rack_use_imac_dack) rack->rc_dack_mode = 1; TAILQ_INIT(&rack->r_ctl.opt_list); rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; rack->r_ctl.rc_pkt_delay = rack_pkt_delay; rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; rack->r_ctl.rc_lowest_us_rtt = 0xffffffff; rack->r_ctl.rc_highest_us_rtt = 0; rack->r_ctl.bw_rate_cap = rack_bw_rate_cap; rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop); if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; if (rack_disable_prr) rack->rack_no_prr = 1; if (rack_gp_no_rec_chg) rack->rc_gp_no_rec_chg = 1; if (rack_pace_every_seg && tcp_can_enable_pacing()) { rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); } else rack->rc_always_pace = 0; if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) rack->r_mbuf_queue = 1; else rack->r_mbuf_queue = 0; if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack) tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; else tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; rack_set_pace_segments(tp, rack, __LINE__, NULL); if (rack_limits_scwnd) rack->r_limit_scw = 1; else rack->r_limit_scw = 0; rack->rc_labc = V_tcp_abc_l_var; rack->r_ctl.rc_high_rwnd = tp->snd_wnd; rack->r_ctl.cwnd_to_use = tp->snd_cwnd; rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method; rack->rack_tlp_threshold_use = rack_tlp_threshold_use; rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr; rack->r_ctl.rc_min_to = rack_min_to; microuptime(&rack->r_ctl.act_rcv_time); rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time; rack->r_running_late = 0; rack->r_running_early = 0; rack->rc_init_win = rack_default_init_window; rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss; if (rack_hw_up_only) rack->r_up_only = 1; if (rack_do_dyn_mul) { /* When dynamic adjustment is on CA needs to start at 100% */ rack->rc_gp_dyn_mul = 1; if (rack_do_dyn_mul >= 100) rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul; } else rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca; rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec; rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt; rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time); setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN, rack_probertt_filter_life); us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_of_last_probertt = us_cts; rack->r_ctl.challenge_ack_ts = tcp_ts_getticks(); rack->r_ctl.rc_time_probertt_starts = 0; /* We require at least one measurement, even if the sysctl is 0 */ if (rack_req_measurements) rack->r_ctl.req_measurements = rack_req_measurements; else rack->r_ctl.req_measurements = 1; if (rack_enable_hw_pacing) rack->rack_hdw_pace_ena = 1; if (rack_hw_rate_caps) rack->r_rack_hw_rate_caps = 1; /* Do we force on detection? */ #ifdef NETFLIX_EXP_DETECTION if (tcp_force_detection) rack->do_detection = 1; else #endif rack->do_detection = 0; if (rack_non_rxt_use_cr) rack->rack_rec_nonrxt_use_cr = 1; err = rack_init_fsb(tp, rack); if (err) { uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; return (err); } if (tp->snd_una != tp->snd_max) { /* Create a send map for the current outstanding data */ struct rack_sendmap *rsm; rsm = rack_alloc(rack); if (rsm == NULL) { uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; return (ENOMEM); } rsm->r_no_rtt_allowed = 1; rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; if (tp->t_flags & TF_SENTFIN) { rsm->r_end = tp->snd_max - 1; rsm->r_flags |= RACK_HAS_FIN; } else { rsm->r_end = tp->snd_max; } if (tp->snd_una == tp->iss) { /* The data space is one beyond snd_una */ rsm->r_flags |= RACK_HAS_SYN; rsm->r_start = tp->iss; rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una); } else rsm->r_start = tp->snd_una; rsm->r_dupack = 0; if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) { rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff); if (rsm->m) rsm->orig_m_len = rsm->m->m_len; else rsm->orig_m_len = 0; } else { /* * This can happen if we have a stand-alone FIN or * SYN. */ rsm->m = NULL; rsm->orig_m_len = 0; rsm->soff = 0; } insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS if (insret != NULL) { panic("Insert in rb tree fails ret:%p rack:%p rsm:%p", insret, rack, rsm); } #endif TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; } /* * Timers in Rack are kept in microseconds so lets * convert any initial incoming variables * from ticks into usecs. Note that we * also change the values of t_srtt and t_rttvar, if * they are non-zero. They are kept with a 5 * bit decimal so we have to carefully convert * these to get the full precision. */ rack_convert_rtts(tp); tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow); if (rack_def_profile) rack_set_profile(rack, rack_def_profile); /* Cancel the GP measurement in progress */ tp->t_flags &= ~TF_GPUTINPROG; if (SEQ_GT(tp->snd_max, tp->iss)) snt = tp->snd_max - tp->iss; else snt = 0; iwin = rc_init_window(rack); if (snt < iwin) { /* We are not past the initial window * so we need to make sure cwnd is * correct. */ if (tp->snd_cwnd < iwin) tp->snd_cwnd = iwin; /* * If we are within the initial window * we want ssthresh to be unlimited. Setting * it to the rwnd (which the default stack does * and older racks) is not really a good idea * since we want to be in SS and grow both the * cwnd and the rwnd (via dynamic rwnd growth). If * we set it to the rwnd then as the peer grows its * rwnd we will be stuck in CA and never hit SS. * * Its far better to raise it up high (this takes the * risk that there as been a loss already, probably * we should have an indicator in all stacks of loss * but we don't), but considering the normal use this * is a risk worth taking. The consequences of not * hitting SS are far worse than going one more time * into it early on (before we have sent even a IW). * It is highly unlikely that we will have had a loss * before getting the IW out. */ tp->snd_ssthresh = 0xffffffff; } rack_stop_all_timers(tp); /* Lets setup the fsb block */ rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0); rack_log_rtt_shrinks(rack, us_cts, tp->t_rxtcur, __LINE__, RACK_RTTS_INIT); return (0); } static int rack_handoff_ok(struct tcpcb *tp) { if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { /* Sure no problem though it may not stick */ return (0); } if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) { /* * We really don't know if you support sack, * you have to get to ESTAB or beyond to tell. */ return (EAGAIN); } if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) { /* * Rack will only send a FIN after all data is acknowledged. * So in this case we have more data outstanding. We can't * switch stacks until either all data and only the FIN * is left (in which case rack_init() now knows how * to deal with that) all is acknowledged and we * are only left with incoming data, though why you * would want to switch to rack after all data is acknowledged * I have no idea (rrs)! */ return (EAGAIN); } if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){ return (0); } /* * If we reach here we don't do SACK on this connection so we can * never do rack. */ return (EINVAL); } static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged) { int ack_cmp = 0; if (tp->t_fb_ptr) { struct tcp_rack *rack; struct rack_sendmap *rsm, *nrsm, *rm; rack = (struct tcp_rack *)tp->t_fb_ptr; if (tp->t_in_pkt) { /* * It is unsafe to process the packets since a * reset may be lurking in them (its rare but it * can occur). If we were to find a RST, then we * would end up dropping the connection and the * INP lock, so when we return the caller (tcp_usrreq) * will blow up when it trys to unlock the inp. */ struct mbuf *save, *m; m = tp->t_in_pkt; tp->t_in_pkt = NULL; tp->t_tail_pkt = NULL; while (m) { save = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); m = save; } if ((tp->t_inpcb) && (tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP)) ack_cmp = 1; if (ack_cmp) { /* Total if we used large or small (if ack-cmp was used). */ if (rack->rc_inp->inp_flags2 & INP_MBUF_L_ACKS) counter_u64_add(rack_large_ackcmp, 1); else counter_u64_add(rack_small_ackcmp, 1); } } tp->t_flags &= ~TF_FORCEDATA; #ifdef NETFLIX_SHARED_CWND if (rack->r_ctl.rc_scw) { uint32_t limit; if (rack->r_limit_scw) limit = max(1, rack->r_ctl.rc_lowest_us_rtt); else limit = 0; tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index, limit); rack->r_ctl.rc_scw = NULL; } #endif if (rack->r_ctl.fsb.tcp_ip_hdr) { free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB); rack->r_ctl.fsb.tcp_ip_hdr = NULL; rack->r_ctl.fsb.th = NULL; } /* Convert back to ticks, with */ if (tp->t_srtt > 1) { uint32_t val, frac; val = USEC_2_TICKS(tp->t_srtt); frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz); tp->t_srtt = val << TCP_RTT_SHIFT; /* * frac is the fractional part here is left * over from converting to hz and shifting. * We need to convert this to the 5 bit * remainder. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC); } else { frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC); } tp->t_srtt += frac; } } if (tp->t_rttvar) { uint32_t val, frac; val = USEC_2_TICKS(tp->t_rttvar); frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz); tp->t_rttvar = val << TCP_RTTVAR_SHIFT; /* * frac is the fractional part here is left * over from converting to hz and shifting. * We need to convert this to the 5 bit * remainder. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC); } else { frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC); } tp->t_rttvar += frac; } } tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur); tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow); if (rack->rc_always_pace) { tcp_decrement_paced_conn(); rack_undo_cc_pacing(rack); rack->rc_always_pace = 0; } /* Clean up any options if they were not applied */ while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) { struct deferred_opt_list *dol; dol = TAILQ_FIRST(&rack->r_ctl.opt_list); TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next); free(dol, M_TCPDO); } /* rack does not use force data but other stacks may clear it */ if (rack->r_ctl.crte != NULL) { tcp_rel_pacing_rate(rack->r_ctl.crte, tp); rack->rack_hdrw_pacing = 0; rack->r_ctl.crte = NULL; } #ifdef TCP_BLACKBOX tcp_log_flowend(tp); #endif RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) { rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS if (rm != rsm) { panic("At fini, rack:%p rsm:%p rm:%p", rack, rsm, rm); } #endif uma_zfree(rack_zone, rsm); } rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); while (rsm) { TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); uma_zfree(rack_zone, rsm); rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); } rack->rc_free_cnt = 0; uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; } if (tp->t_inpcb) { tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY; tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP; /* Cancel the GP measurement in progress */ tp->t_flags &= ~TF_GPUTINPROG; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS; } /* Make sure snd_nxt is correctly set */ tp->snd_nxt = tp->snd_max; } static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack) { if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) { rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; } switch (tp->t_state) { case TCPS_SYN_SENT: rack->r_state = TCPS_SYN_SENT; rack->r_substate = rack_do_syn_sent; break; case TCPS_SYN_RECEIVED: rack->r_state = TCPS_SYN_RECEIVED; rack->r_substate = rack_do_syn_recv; break; case TCPS_ESTABLISHED: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_ESTABLISHED; rack->r_substate = rack_do_established; break; case TCPS_CLOSE_WAIT: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_CLOSE_WAIT; rack->r_substate = rack_do_close_wait; break; case TCPS_FIN_WAIT_1: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_FIN_WAIT_1; rack->r_substate = rack_do_fin_wait_1; break; case TCPS_CLOSING: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_CLOSING; rack->r_substate = rack_do_closing; break; case TCPS_LAST_ACK: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_LAST_ACK; rack->r_substate = rack_do_lastack; break; case TCPS_FIN_WAIT_2: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_FIN_WAIT_2; rack->r_substate = rack_do_fin_wait_2; break; case TCPS_LISTEN: case TCPS_CLOSED: case TCPS_TIME_WAIT: default: break; }; if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; } static void rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb) { /* * We received an ack, and then did not * call send or were bounced out due to the * hpts was running. Now a timer is up as well, is * it the right timer? */ struct rack_sendmap *rsm; int tmr_up; tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) return; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && (tmr_up == PACE_TMR_RXT)) { /* Should be an RXT */ return; } if (rsm == NULL) { /* Nothing outstanding? */ if (tp->t_flags & TF_DELACK) { if (tmr_up == PACE_TMR_DELACK) /* We are supposed to have delayed ack up and we do */ return; } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { /* * if we hit enobufs then we would expect the possiblity * of nothing outstanding and the RXT up (and the hptsi timer). */ return; } else if (((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) && (tmr_up == PACE_TMR_KEEP) && (tp->snd_max == tp->snd_una)) { /* We should have keep alive up and we do */ return; } } if (SEQ_GT(tp->snd_max, tp->snd_una) && ((tmr_up == PACE_TMR_TLP) || (tmr_up == PACE_TMR_RACK) || (tmr_up == PACE_TMR_RXT))) { /* * Either a Rack, TLP or RXT is fine if we * have outstanding data. */ return; } else if (tmr_up == PACE_TMR_DELACK) { /* * If the delayed ack was going to go off * before the rtx/tlp/rack timer were going to * expire, then that would be the timer in control. * Note we don't check the time here trusting the * code is correct. */ return; } /* * Ok the timer originally started is not what we want now. * We will force the hpts to be stopped if any, and restart * with the slot set to what was in the saved slot. */ if (rack->rc_inp->inp_in_hpts) { if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { uint32_t us_cts; us_cts = tcp_get_usecs(NULL); if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) { rack->r_early = 1; rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts); } rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; } tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); } rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0); } static void rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts, uint32_t high_seq) { tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = seq; tp->snd_wl2 = ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; if (tp->snd_wnd < (tp->snd_max - high_seq)) { /* The peer collapsed the window */ rack_collapsed_window(rack); } else if (rack->rc_has_collapsed) rack_un_collapse_window(rack); /* Do we exit persists? */ if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs))) { rack_exit_persist(tp, rack, cts); } /* Do we enter persists? */ if ((rack->rc_in_persist == 0) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* * Here the rwnd is less than * the pacing size, we are established, * nothing is outstanding, and there is * data to send. Enter persists. */ rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } } static void rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval ltv; char tcp_hdr_buf[60]; struct tcphdr *th; struct timespec ts; uint32_t orig_snd_una; uint8_t xx = 0; #ifdef NETFLIX_HTTP_LOGGING struct http_sendfile_track *http_req; if (SEQ_GT(ae->ack, tp->snd_una)) { http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1)); } else { http_req = tcp_http_find_req_for_seq(tp, ae->ack); } #endif memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; if (rack->rack_no_prr == 0) log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; else log.u_bbr.flex1 = 0; log.u_bbr.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced; log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex7 = 1; log.u_bbr.lost = ae->flags; log.u_bbr.cwnd_gain = ackval; log.u_bbr.pacing_gain = 0x2; if (ae->flags & TSTMP_HDWR) { /* Record the hardware timestamp if present */ log.u_bbr.flex3 = M_TSTMP; ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; ltv.tv_sec = ts.tv_sec; ltv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.lt_epoch = tcp_tv_to_usectick(<v); } else if (ae->flags & TSTMP_LRO) { /* Record the LRO the arrival timestamp */ log.u_bbr.flex3 = M_TSTMP_LRO; ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; ltv.tv_sec = ts.tv_sec; ltv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.flex5 = tcp_tv_to_usectick(<v); } log.u_bbr.timeStamp = tcp_get_usecs(<v); /* Log the rcv time */ log.u_bbr.delRate = ae->timestamp; #ifdef NETFLIX_HTTP_LOGGING log.u_bbr.applimited = tp->t_http_closed; log.u_bbr.applimited <<= 8; log.u_bbr.applimited |= tp->t_http_open; log.u_bbr.applimited <<= 8; log.u_bbr.applimited |= tp->t_http_req; if (http_req) { /* Copy out any client req info */ /* seconds */ log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC); /* useconds */ log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC); log.u_bbr.rttProp = http_req->timestamp; log.u_bbr.cur_del_rate = http_req->start; if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) { log.u_bbr.flex8 |= 1; } else { log.u_bbr.flex8 |= 2; log.u_bbr.bw_inuse = http_req->end; } log.u_bbr.flex6 = http_req->start_seq; if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) { log.u_bbr.flex8 |= 4; log.u_bbr.epoch = http_req->end_seq; } } #endif memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf)); th = (struct tcphdr *)tcp_hdr_buf; th->th_seq = ae->seq; th->th_ack = ae->ack; th->th_win = ae->win; /* Now fill in the ports */ th->th_sport = tp->t_inpcb->inp_fport; th->th_dport = tp->t_inpcb->inp_lport; th->th_flags = ae->flags & 0xff; /* Now do we have a timestamp option? */ if (ae->flags & HAS_TSTMP) { u_char *cp; uint32_t val; th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2); cp = (u_char *)(th + 1); *cp = TCPOPT_NOP; cp++; *cp = TCPOPT_NOP; cp++; *cp = TCPOPT_TIMESTAMP; cp++; *cp = TCPOLEN_TIMESTAMP; cp++; val = htonl(ae->ts_value); bcopy((char *)&val, (char *)cp, sizeof(uint32_t)); val = htonl(ae->ts_echo); bcopy((char *)&val, (char *)(cp + 4), sizeof(uint32_t)); } else th->th_off = (sizeof(struct tcphdr) >> 2); /* * For sane logging we need to play a little trick. * If the ack were fully processed we would have moved * snd_una to high_seq, but since compressed acks are * processed in two phases, at this point (logging) snd_una * won't be advanced. So we would see multiple acks showing * the advancement. We can prevent that by "pretending" that * snd_una was advanced and then un-advancing it so that the * logging code has the right value for tlb_snd_una. */ if (tp->snd_una != high_seq) { orig_snd_una = tp->snd_una; tp->snd_una = high_seq; xx = 1; } else xx = 0; TCP_LOG_EVENTP(tp, th, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0, 0, &log, true, <v); if (xx) { tp->snd_una = orig_snd_una; } } } static int rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv) { /* * Handle a "special" compressed ack mbuf. Each incoming * ack has only four possible dispositions: * * A) It moves the cum-ack forward * B) It is behind the cum-ack. * C) It is a window-update ack. * D) It is a dup-ack. * * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES * in the incoming mbuf. We also need to still pay attention * to nxt_pkt since there may be another packet after this * one. */ #ifdef TCP_ACCOUNTING uint64_t ts_val; uint64_t rdstc; #endif int segsiz; struct timespec ts; struct tcp_rack *rack; struct tcp_ackent *ae; uint32_t tiwin, us_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack; int cnt, i, did_out, ourfinisacked = 0; int win_up_req = 0; struct tcpopt to_holder, *to = NULL; int nsegs = 0; int under_pacing = 1; int recovery = 0; int idx; #ifdef TCP_ACCOUNTING sched_pin(); #endif rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->gp_ready && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) under_pacing = 0; else under_pacing = 1; if (rack->r_state != tp->t_state) rack_set_state(tp, rack); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_GPUTINPROG)) { /* * We have a goodput in progress * and we have entered a late state. * Do we have enough data in the sb * to handle the GPUT request? */ uint32_t bytes; bytes = tp->gput_ack - tp->gput_seq; if (SEQ_GT(tp->gput_seq, tp->snd_una)) bytes += tp->gput_seq - tp->snd_una; if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* * There are not enough bytes in the socket * buffer that have been sent to cover this * measurement. Cancel it. */ rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); tp->t_flags &= ~TF_GPUTINPROG; } } to = &to_holder; to->to_flags = 0; KASSERT((m->m_len >= sizeof(struct tcp_ackent)), ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len)); cnt = m->m_len / sizeof(struct tcp_ackent); idx = cnt / 5; if (idx >= MAX_NUM_OF_CNTS) idx = MAX_NUM_OF_CNTS - 1; counter_u64_add(rack_proc_comp_ack[idx], 1); counter_u64_add(rack_multi_single_eq, cnt); high_seq = tp->snd_una; the_win = tp->snd_wnd; win_seq = tp->snd_wl1; win_upd_ack = tp->snd_wl2; cts = us_cts = tcp_tv_to_usectick(tv); segsiz = ctf_fixed_maxseg(tp); if ((rack->rc_gp_dyn_mul) && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* Check in on probertt */ rack_check_probe_rtt(rack, us_cts); } for (i = 0; i < cnt; i++) { #ifdef TCP_ACCOUNTING ts_val = get_cyclecount(); #endif rack_clear_rate_sample(rack); ae = ((mtod(m, struct tcp_ackent *)) + i); /* Setup the window */ tiwin = ae->win << tp->snd_scale; /* figure out the type of ack */ if (SEQ_LT(ae->ack, high_seq)) { /* Case B*/ ae->ack_val_set = ACK_BEHIND; } else if (SEQ_GT(ae->ack, high_seq)) { /* Case A */ ae->ack_val_set = ACK_CUMACK; } else if (tiwin == the_win) { /* Case D */ ae->ack_val_set = ACK_DUPACK; } else { /* Case C */ ae->ack_val_set = ACK_RWND; } rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq); /* Validate timestamp */ if (ae->flags & HAS_TSTMP) { /* Setup for a timestamp */ to->to_flags = TOF_TS; ae->ts_echo -= tp->ts_offset; to->to_tsecr = ae->ts_echo; to->to_tsval = ae->ts_value; /* * If echoed timestamp is later than the current time, fall back to * non RFC1323 RTT calculation. Normalize timestamp if syncookies * were used when this connection was established. */ if (TSTMP_GT(ae->ts_echo, cts)) ae->ts_echo = 0; if (tp->ts_recent && TSTMP_LT(ae->ts_value, tp->ts_recent)) { if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) { #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val); } } #endif continue; } } if (SEQ_LEQ(ae->seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, ae->seq)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = ae->ts_value; } } else { /* Setup for a no options */ to->to_flags = 0; } /* Update the rcv time and perform idle reduction possibly */ if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una) && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) { counter_u64_add(rack_input_idle_reduces, 1); rack_cc_after_idle(rack, tp); } tp->t_rcvtime = ticks; /* Now what about ECN? */ if (tp->t_flags2 & TF2_ECN_PERMIT) { if (ae->flags & TH_CWR) { tp->t_flags2 &= ~TF2_ECN_SND_ECE; tp->t_flags |= TF_ACKNOW; } switch (ae->codepoint & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->t_flags2 |= TF2_ECN_SND_ECE; KMOD_TCPSTAT_INC(tcps_ecn_ce); break; case IPTOS_ECN_ECT0: KMOD_TCPSTAT_INC(tcps_ecn_ect0); break; case IPTOS_ECN_ECT1: KMOD_TCPSTAT_INC(tcps_ecn_ect1); break; } /* Process a packet differently from RFC3168. */ cc_ecnpkt_handler_flags(tp, ae->flags, ae->codepoint); /* Congestion experienced. */ if (ae->flags & TH_ECE) { rack_cong_signal(tp, CC_ECN, ae->ack); } } #ifdef TCP_ACCOUNTING /* Count for the specific type of ack in */ counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[ae->ack_val_set]++; } #endif /* * Note how we could move up these in the determination * above, but we don't so that way the timestamp checks (and ECN) * is done first before we do any processing on the ACK. * The non-compressed path through the code has this * weakness (noted by @jtl) that it actually does some * processing before verifying the timestamp information. * We don't take that path here which is why we set * the ack_val_set first, do the timestamp and ecn * processing, and then look at what we have setup. */ if (ae->ack_val_set == ACK_BEHIND) { /* * Case B flag reordering, if window is not closed * or it could be a keep-alive or persists */ if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) { counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); } } else if (ae->ack_val_set == ACK_DUPACK) { /* Case D */ rack_strike_dupack(rack); } else if (ae->ack_val_set == ACK_RWND) { /* Case C */ win_up_req = 1; win_upd_ack = ae->ack; win_seq = ae->seq; the_win = tiwin; } else { /* Case A */ if (SEQ_GT(ae->ack, tp->snd_max)) { /* * We just send an ack since the incoming * ack is beyond the largest seq we sent. */ if ((tp->t_flags & TF_ACKNOW) == 0) { ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt); if (tp->t_flags && TF_ACKNOW) rack->r_wanted_output = 1; } } else { nsegs++; /* If the window changed setup to update */ if (tiwin != tp->snd_wnd) { win_up_req = 1; win_upd_ack = ae->ack; win_seq = ae->seq; the_win = tiwin; } #ifdef TCP_ACCOUNTING /* Account for the acks */ if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN], (((ae->ack - high_seq) + segsiz - 1) / segsiz)); #endif high_seq = ae->ack; /* Setup our act_rcv_time */ if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) { ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec; rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000; } else { rack->r_ctl.act_rcv_time = *tv; } rack_process_to_cumack(tp, rack, ae->ack, cts, to); } } /* And lets be sure to commit the rtt measurements for this ack */ tcp_rack_xmit_timer_commit(rack, tp); #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val); if (ae->ack_val_set == ACK_CUMACK) tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val); } } #endif } #ifdef TCP_ACCOUNTING ts_val = get_cyclecount(); #endif acked_amount = acked = (high_seq - tp->snd_una); if (win_up_req) { rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq); } if (acked) { if (rack->sack_attack_disable == 0) rack_do_decay(rack); if (acked >= segsiz) { /* * You only get credit for * MSS and greater (and you get extra * credit for larger cum-ack moves). */ int ac; ac = acked / segsiz; rack->r_ctl.ack_count += ac; counter_u64_add(rack_ack_total, ac); } if (rack->r_ctl.ack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our SYN has * been ACK'd (so connection is now fully synchronized). Go * to non-starred state, increment snd_una for ACK of SYN, * and check if we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; acked_amount = acked = (high_seq - tp->snd_una); } if (acked > sbavail(&so->so_snd)) acked_amount = sbavail(&so->so_snd); #ifdef NETFLIX_EXP_DETECTION /* * We only care on a cum-ack move if we are in a sack-disabled * state. We have already added in to the ack_count, and we never * would disable on a cum-ack move, so we only care to do the * detection if it may "undo" it, i.e. we were in disabled already. */ if (rack->sack_attack_disable) rack_do_detection(tp, rack, acked_amount, segsiz); #endif if (IN_FASTRECOVERY(tp->t_flags) && (rack->rack_no_prr == 0)) rack_update_prr(tp, rack, acked_amount, high_seq); if (IN_RECOVERY(tp->t_flags)) { if (SEQ_LT(high_seq, tp->snd_recover) && (SEQ_LT(high_seq, tp->snd_max))) { tcp_rack_partialack(tp); } else { rack_post_recovery(tp, high_seq); recovery = 1; } } /* Handle the rack-log-ack part (sendmap) */ if ((sbused(&so->so_snd) == 0) && (acked > acked_amount) && (tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_SENTFIN)) { /* * We must be sure our fin * was sent and acked (we can be * in FIN_WAIT_1 without having * sent the fin). */ ourfinisacked = 1; /* * Lets make sure snd_una is updated * since most likely acked_amount = 0 (it * should be). */ tp->snd_una = high_seq; } /* Did we make a RTO error? */ if ((tp->t_flags & TF_PREVVALID) && ((tp->t_flags & TF_RCVD_TSTMP) == 0)) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, CC_RTO_ERR, high_seq); } /* Handle the data in the socket buffer */ KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); if (acked_amount > 0) { struct mbuf *mfree; rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery); SOCKBUF_LOCK(&so->so_snd); mfree = sbcut_locked(&so->so_snd, acked); tp->snd_una = high_seq; /* Note we want to hold the sb lock through the sendmap adjust */ rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una); /* Wake up the socket if we have room to write more */ rack_log_wakeup(tp,rack, &so->so_snd, acked, 2); sowwakeup_locked(so); m_freem(mfree); } /* update progress */ tp->t_acktime = ticks; rack_log_progress_event(rack, tp, tp->t_acktime, PROGRESS_UPDATE, __LINE__); /* Clear out shifts and such */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; /* Send recover and snd_nxt must be dragged along */ if (SEQ_GT(tp->snd_una, tp->snd_recover)) tp->snd_recover = tp->snd_una; if (SEQ_LT(tp->snd_nxt, tp->snd_una)) tp->snd_nxt = tp->snd_una; /* * If the RXT timer is running we want to * stop it, so we can restart a TLP (or new RXT). */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); #ifdef NETFLIX_HTTP_LOGGING tcp_http_check_for_comp(rack->rc_tp, high_seq); #endif tp->snd_wl2 = high_seq; tp->t_dupacks = 0; if (under_pacing && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && rack->rc_gp_dyn_mul && rack->rc_always_pace) { /* Check if we are dragging bottom */ rack_check_bottom_drag(tp, rack, so, acked); } if (tp->snd_una == tp->snd_max) { tp->t_flags &= ~TF_PREVVALID; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) tp->t_acktime = 0; /* Set so we might enter persists... */ rack->r_wanted_output = 1; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (sbavail(&so->so_snd) == 0) && (tp->t_flags2 & TF2_DROP_AF_DATA)) { /* * The socket was gone and the * peer sent data (not now in the past), time to * reset him. */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } #endif m_freem(m); tp = tcp_close(tp); if (tp == NULL) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } /* * We would normally do drop-with-reset which would * send back a reset. We can't since we don't have * all the needed bits. Instead lets arrange for * a call to tcp_output(). That way since we * are in the closed state we will generate a reset. * * Note if tcp_accounting is on we don't unpin since * we do that after the goto label. */ goto send_out_a_rst; } if ((sbused(&so->so_snd) == 0) && (tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_SENTFIN)) { /* * If we can't receive any more data, then closing user can * proceed. Starting the timer is contrary to the * specification, but if we don't get a FIN we'll hang * forever. * */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } if (ourfinisacked == 0) { /* * We don't change to fin-wait-2 if we have our fin acked * which means we are probably in TCPS_CLOSING. */ tcp_state_change(tp, TCPS_FIN_WAIT_2); } } } /* Wake up the socket if we have room to write more */ if (sbavail(&so->so_snd)) { rack->r_wanted_output = 1; if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); /* * We cheat here and don't send a RST, we should send one * when the pacer drops the connection. */ #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } sched_unpin(); #endif INP_WUNLOCK(rack->rc_inp); m_freem(m); return (1); } } if (ourfinisacked) { switch(tp->t_state) { case TCPS_CLOSING: #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } sched_unpin(); #endif tcp_twstart(tp); m_freem(m); return (1); break; case TCPS_LAST_ACK: #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } sched_unpin(); #endif tp = tcp_close(tp); ctf_do_drop(m, tp); return (1); break; case TCPS_FIN_WAIT_1: #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } #endif if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); break; default: break; } } if (rack->r_fast_output) { /* * We re doing fast output.. can we expand that? */ rack_gain_for_fastoutput(rack, tp, so, acked_amount); } #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } } else if (win_up_req) { rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val); } } #endif } /* Now is there a next packet, if so we are done */ m_freem(m); did_out = 0; if (nxt_pkt) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs); return (0); } rack_handle_might_revert(tp, rack); ctf_calc_rwin(so, tp); if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) { send_out_a_rst: (void)tp->t_fb->tfb_tcp_output(tp); did_out = 1; } rack_free_trim(rack); #ifdef TCP_ACCOUNTING sched_unpin(); #endif rack_timer_audit(tp, rack, &so->so_snd); rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs); return (0); } static int rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt, struct timeval *tv) { #ifdef TCP_ACCOUNTING uint64_t ts_val; #endif int32_t thflags, retval, did_out = 0; int32_t way_out = 0; uint32_t cts; uint32_t tiwin; struct timespec ts; struct tcpopt to; struct tcp_rack *rack; struct rack_sendmap *rsm; int32_t prev_state = 0; #ifdef TCP_ACCOUNTING int ack_val_set = 0xf; #endif int nsegs; uint32_t us_cts; /* * tv passed from common code is from either M_TSTMP_LRO or * tcp_get_usecs() if no LRO m_pkthdr timestamp is present. */ rack = (struct tcp_rack *)tp->t_fb_ptr; cts = tcp_tv_to_usectick(tv); if (m->m_flags & M_ACKCMP) { return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv)); } if (m->m_flags & M_ACKCMP) { panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp); } nsegs = m->m_pkthdr.lro_nsegs; counter_u64_add(rack_proc_non_comp_ack, 1); thflags = th->th_flags; #ifdef TCP_ACCOUNTING sched_pin(); if (thflags & TH_ACK) ts_val = get_cyclecount(); #endif if ((m->m_flags & M_TSTMP) || (m->m_flags & M_TSTMP_LRO)) { mbuf_tstmp2timespec(m, &ts); rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec; rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000; } else rack->r_ctl.act_rcv_time = *tv; kern_prefetch(rack, &prev_state); prev_state = 0; /* * Unscale the window into a 32-bit value. For the SYN_SENT state * the scale is zero. */ tiwin = th->th_win << tp->snd_scale; /* * Parse options on any incoming segment. */ memset(&to, 0, sizeof(to)); tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); #ifdef TCP_ACCOUNTING if (thflags & TH_ACK) { /* * We have a tradeoff here. We can either do what we are * doing i.e. pinning to this CPU and then doing the accounting * we could do a critical enter, setup the rdtsc and cpu * as in below, and then validate we are on the same CPU on * exit. I have choosen to not do the critical enter since * that often will gain you a context switch, and instead lock * us (line above this if) to the same CPU with sched_pin(). This * means we may be context switched out for a higher priority * interupt but we won't be moved to another CPU. * * If this occurs (which it won't very often since we most likely * are running this code in interupt context and only a higher * priority will bump us ... clock?) we will falsely add in * to the time the interupt processing time plus the ack processing * time. This is ok since its a rare event. */ ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin, ctf_fixed_maxseg(tp)); } #endif NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", __func__)); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_GPUTINPROG)) { /* * We have a goodput in progress * and we have entered a late state. * Do we have enough data in the sb * to handle the GPUT request? */ uint32_t bytes; bytes = tp->gput_ack - tp->gput_seq; if (SEQ_GT(tp->gput_seq, tp->snd_una)) bytes += tp->gput_seq - tp->snd_una; if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* * There are not enough bytes in the socket * buffer that have been sent to cover this * measurement. Cancel it. */ rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); tp->t_flags &= ~TF_GPUTINPROG; } } if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval ltv; #ifdef NETFLIX_HTTP_LOGGING struct http_sendfile_track *http_req; if (SEQ_GT(th->th_ack, tp->snd_una)) { http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1)); } else { http_req = tcp_http_find_req_for_seq(tp, th->th_ack); } #endif memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; if (rack->rack_no_prr == 0) log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; else log.u_bbr.flex1 = 0; log.u_bbr.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->rc_tp->t_maxseg; log.u_bbr.flex3 = m->m_flags; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; log.u_bbr.lost = thflags; log.u_bbr.pacing_gain = 0x1; #ifdef TCP_ACCOUNTING log.u_bbr.cwnd_gain = ack_val_set; #endif log.u_bbr.flex7 = 2; if (m->m_flags & M_TSTMP) { /* Record the hardware timestamp if present */ mbuf_tstmp2timespec(m, &ts); ltv.tv_sec = ts.tv_sec; ltv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.lt_epoch = tcp_tv_to_usectick(<v); } else if (m->m_flags & M_TSTMP_LRO) { /* Record the LRO the arrival timestamp */ mbuf_tstmp2timespec(m, &ts); ltv.tv_sec = ts.tv_sec; ltv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.flex5 = tcp_tv_to_usectick(<v); } log.u_bbr.timeStamp = tcp_get_usecs(<v); /* Log the rcv time */ log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp; #ifdef NETFLIX_HTTP_LOGGING log.u_bbr.applimited = tp->t_http_closed; log.u_bbr.applimited <<= 8; log.u_bbr.applimited |= tp->t_http_open; log.u_bbr.applimited <<= 8; log.u_bbr.applimited |= tp->t_http_req; if (http_req) { /* Copy out any client req info */ /* seconds */ log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC); /* useconds */ log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC); log.u_bbr.rttProp = http_req->timestamp; log.u_bbr.cur_del_rate = http_req->start; if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) { log.u_bbr.flex8 |= 1; } else { log.u_bbr.flex8 |= 2; log.u_bbr.bw_inuse = http_req->end; } log.u_bbr.flex6 = http_req->start_seq; if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) { log.u_bbr.flex8 |= 4; log.u_bbr.epoch = http_req->end_seq; } } #endif TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, tlen, &log, true, <v); } if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { way_out = 4; retval = 0; m_freem(m); goto done_with_input; } /* * If a segment with the ACK-bit set arrives in the SYN-SENT state * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. */ if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } /* * Parse options on any incoming segment. */ tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); /* * If timestamps were negotiated during SYN/ACK and a * segment without a timestamp is received, silently drop * the segment, unless it is a RST segment or missing timestamps are * tolerated. * See section 3.2 of RFC 7323. */ if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { way_out = 5; retval = 0; m_freem(m); goto done_with_input; } /* * Segment received on connection. Reset idle time and keep-alive * timer. XXX: This should be done after segment validation to * ignore broken/spoofed segs. */ if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una) && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) { counter_u64_add(rack_input_idle_reduces, 1); rack_cc_after_idle(rack, tp); } tp->t_rcvtime = ticks; #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); #endif if (tiwin > rack->r_ctl.rc_high_rwnd) rack->r_ctl.rc_high_rwnd = tiwin; /* * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move * this to occur after we've validated the segment. */ if (tp->t_flags2 & TF2_ECN_PERMIT) { if (thflags & TH_CWR) { tp->t_flags2 &= ~TF2_ECN_SND_ECE; tp->t_flags |= TF_ACKNOW; } switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->t_flags2 |= TF2_ECN_SND_ECE; KMOD_TCPSTAT_INC(tcps_ecn_ce); break; case IPTOS_ECN_ECT0: KMOD_TCPSTAT_INC(tcps_ecn_ect0); break; case IPTOS_ECN_ECT1: KMOD_TCPSTAT_INC(tcps_ecn_ect1); break; } /* Process a packet differently from RFC3168. */ cc_ecnpkt_handler(tp, th, iptos); /* Congestion experienced. */ if (thflags & TH_ECE) { rack_cong_signal(tp, CC_ECN, th->th_ack); } } /* * If echoed timestamp is later than the current time, fall back to * non RFC1323 RTT calculation. Normalize timestamp if syncookies * were used when this connection was established. */ if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { to.to_tsecr -= tp->ts_offset; if (TSTMP_GT(to.to_tsecr, cts)) to.to_tsecr = 0; } /* * If its the first time in we need to take care of options and * verify we can do SACK for rack! */ if (rack->r_state == 0) { /* Should be init'd by rack_init() */ KASSERT(rack->rc_inp != NULL, ("%s: rack->rc_inp unexpectedly NULL", __func__)); if (rack->rc_inp == NULL) { rack->rc_inp = tp->t_inpcb; } /* * Process options only when we get SYN/ACK back. The SYN * case for incoming connections is handled in tcp_syncache. * According to RFC1323 the window field in a SYN (i.e., a * or ) segment itself is never scaled. XXX * this is traditional behavior, may need to be cleaned up. */ if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { /* Handle parallel SYN for ECN */ if (!(thflags & TH_ACK) && ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) && ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) { tp->t_flags2 |= TF2_ECN_PERMIT; tp->t_flags2 |= TF2_ECN_SND_ECE; TCPSTAT_INC(tcps_ecn_shs); } if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) { tp->t_flags |= TF_RCVD_SCALE; tp->snd_scale = to.to_wscale; } else tp->t_flags &= ~TF_REQ_SCALE; /* * Initial send window. It will be updated with the * next incoming segment to the scaled value. */ tp->snd_wnd = th->th_win; rack_validate_fo_sendwin_up(tp, rack); if ((to.to_flags & TOF_TS) && (tp->t_flags & TF_REQ_TSTMP)) { tp->t_flags |= TF_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = cts; } else tp->t_flags &= ~TF_REQ_TSTMP; if (to.to_flags & TOF_MSS) { tcp_mss(tp, to.to_mss); } if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACKPERM) == 0) tp->t_flags &= ~TF_SACK_PERMIT; if (IS_FASTOPEN(tp->t_flags)) { if (to.to_flags & TOF_FASTOPEN) { uint16_t mss; if (to.to_flags & TOF_MSS) mss = to.to_mss; else if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) mss = TCP6_MSS; else mss = TCP_MSS; tcp_fastopen_update_cache(tp, mss, to.to_tfo_len, to.to_tfo_cookie); } else tcp_fastopen_disable_path(tp); } } /* * At this point we are at the initial call. Here we decide * if we are doing RACK or not. We do this by seeing if * TF_SACK_PERMIT is set and the sack-not-required is clear. * The code now does do dup-ack counting so if you don't * switch back you won't get rack & TLP, but you will still * get this stack. */ if ((rack_sack_not_required == 0) && ((tp->t_flags & TF_SACK_PERMIT) == 0)) { tcp_switch_back_to_default(tp); (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, tlen, iptos); #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } tcp_set_hpts(tp->t_inpcb); sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack); } if (thflags & TH_FIN) tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); if ((rack->rc_gp_dyn_mul) && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* Check in on probertt */ rack_check_probe_rtt(rack, us_cts); } rack_clear_rate_sample(rack); if (rack->forced_ack) { uint32_t us_rtt; /* * A persist or keep-alive was forced out, update our * min rtt time. Note we do not worry about lost * retransmissions since KEEP-ALIVES and persists * are usually way long on times of sending (though * if we were really paranoid or worried we could * at least use timestamps if available to validate). */ rack->forced_ack = 0; us_rtt = us_cts - rack->r_ctl.forced_ack_ts; if (us_rtt == 0) us_rtt = 1; rack_apply_updated_usrtt(rack, us_rtt, us_cts); tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1); } /* * This is the one exception case where we set the rack state * always. All other times (timers etc) we must have a rack-state * set (so we assure we have done the checks above for SACK). */ rack->r_ctl.rc_rcvtime = cts; if (rack->r_state != tp->t_state) rack_set_state(tp, rack); if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL) kern_prefetch(rsm, &prev_state); prev_state = rack->r_state; retval = (*rack->r_substate) (m, th, so, tp, &to, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt, iptos); #ifdef INVARIANTS if ((retval == 0) && (tp->t_inpcb == NULL)) { panic("retval:%d tp:%p t_inpcb:NULL state:%d", retval, tp, prev_state); } #endif if (retval == 0) { /* * If retval is 1 the tcb is unlocked and most likely the tp * is gone. */ INP_WLOCK_ASSERT(tp->t_inpcb); if ((rack->rc_gp_dyn_mul) && (rack->rc_always_pace) && (rack->use_fixed_rate == 0) && rack->in_probe_rtt && (rack->r_ctl.rc_time_probertt_starts == 0)) { /* * If we are going for target, lets recheck before * we output. */ rack_check_probe_rtt(rack, us_cts); } if (rack->set_pacing_done_a_iw == 0) { /* How much has been acked? */ if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) { /* We have enough to set in the pacing segment size */ rack->set_pacing_done_a_iw = 1; rack_set_pace_segments(tp, rack, __LINE__, NULL); } } tcp_rack_xmit_timer_commit(rack, tp); #ifdef TCP_ACCOUNTING /* * If we set the ack_val_se to what ack processing we are doing * we also want to track how many cycles we burned. Note * the bits after tcp_output we let be "free". This is because * we are also tracking the tcp_output times as well. Note the * use of 0xf here since we only have 11 counter (0 - 0xa) and * 0xf cannot be returned and is what we initialize it too to * indicate we are not doing the tabulations. */ if (ack_val_set != 0xf) { uint64_t crtsc; crtsc = get_cyclecount(); counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val); } } #endif if (nxt_pkt == 0) { if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) { do_output_now: did_out = 1; (void)tp->t_fb->tfb_tcp_output(tp); } rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); rack_free_trim(rack); } if ((nxt_pkt == 0) && ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && (SEQ_GT(tp->snd_max, tp->snd_una) || (tp->t_flags & TF_DELACK) || ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)))) { /* We could not send (probably in the hpts but stopped the timer earlier)? */ if ((tp->snd_max == tp->snd_una) && ((tp->t_flags & TF_DELACK) == 0) && (rack->rc_inp->inp_in_hpts) && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* keep alive not needed if we are hptsi output yet */ ; } else { int late = 0; if (rack->rc_inp->inp_in_hpts) { if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { us_cts = tcp_get_usecs(NULL); if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) { rack->r_early = 1; rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts); } else late = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; } tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); } if (late && (did_out == 0)) { /* * We are late in the sending * and we did not call the output * (this probably should not happen). */ goto do_output_now; } rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0); } way_out = 1; } else if (nxt_pkt == 0) { /* Do we have the correct timer running? */ rack_timer_audit(tp, rack, &so->so_snd); way_out = 2; } done_with_input: rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs)); if (did_out) rack->r_wanted_output = 0; #ifdef INVARIANTS if (tp->t_inpcb == NULL) { panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", did_out, retval, tp, prev_state); } #endif #ifdef TCP_ACCOUNTING } else { /* * Track the time (see above). */ if (ack_val_set != 0xf) { uint64_t crtsc; crtsc = get_cyclecount(); counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val)); /* * Note we *DO NOT* increment the per-tcb counters since * in the else the TP may be gone!! */ } #endif } #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (retval); } void rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) { struct timeval tv; /* First lets see if we have old packets */ if (tp->t_in_pkt) { if (ctf_do_queued_segments(so, tp, 1)) { m_freem(m); return; } } if (m->m_flags & M_TSTMP_LRO) { tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000; tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000; } else { /* Should not be should we kassert instead? */ tcp_get_usecs(&tv); } if (rack_do_segment_nounlock(m, th, so, tp, drop_hdrlen, tlen, iptos, 0, &tv) == 0) { INP_WUNLOCK(tp->t_inpcb); } } struct rack_sendmap * tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused) { struct rack_sendmap *rsm = NULL; int32_t idx; uint32_t srtt = 0, thresh = 0, ts_low = 0; /* Return the next guy to be re-transmitted */ if (RB_EMPTY(&rack->r_ctl.rc_mtree)) { return (NULL); } if (tp->t_flags & TF_SENTFIN) { /* retran the end FIN? */ return (NULL); } /* ok lets look at this one */ rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) { goto check_it; } rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) { return (NULL); } check_it: if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) { /* * No sack so we automatically do the 3 strikes and * retransmit (no rack timer would be started). */ return (rsm); } if (rsm->r_flags & RACK_ACKED) { return (NULL); } if (((rsm->r_flags & RACK_SACK_PASSED) == 0) && (rsm->r_dupack < DUP_ACK_THRESHOLD)) { /* Its not yet ready */ return (NULL); } srtt = rack_grab_rtt(tp, rack); idx = rsm->r_rtr_cnt - 1; ts_low = (uint32_t)rsm->r_tim_lastsent[idx]; thresh = rack_calc_thresh_rack(rack, srtt, tsused); if ((tsused == ts_low) || (TSTMP_LT(tsused, ts_low))) { /* No time since sending */ return (NULL); } if ((tsused - ts_low) < thresh) { /* It has not been long enough yet */ return (NULL); } if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || ((rsm->r_flags & RACK_SACK_PASSED) && (rack->sack_attack_disable == 0))) { /* * We have passed the dup-ack threshold * a SACK has indicated this is missing. * Note that if you are a declared attacker * it is only the dup-ack threshold that * will cause retransmits. */ /* log retransmit reason */ rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1); rack->r_fast_output = 0; return (rsm); } return (NULL); } static void rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm, uint8_t quality) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = slot; log.u_bbr.flex2 = len; log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss; log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca; log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_late; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_early; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->app_limited_needs_set; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_gp_filled; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->in_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->gp_ready; log.u_bbr.pkt_epoch = line; log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed; log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early; log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec; log.u_bbr.bw_inuse = bw_est; log.u_bbr.delRate = bw; if (rack->r_ctl.gp_bw == 0) log.u_bbr.cur_del_rate = 0; else log.u_bbr.cur_del_rate = rack_get_bw(rack); log.u_bbr.rttProp = len_time; log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt; log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit; log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm); if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) { /* We are in slow start */ log.u_bbr.flex7 = 1; } else { /* we are on congestion avoidance */ log.u_bbr.flex7 = 0; } log.u_bbr.flex8 = method; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca; log.u_bbr.bbr_substate = quality; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSI_CALC, 0, 0, &log, false, &tv); } } static uint32_t rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss) { uint32_t new_tso, user_max; user_max = rack->rc_user_set_max_segs * mss; if (rack->rc_force_max_seg) { return (user_max); } if (rack->use_fixed_rate && ((rack->r_ctl.crte == NULL) || (bw != rack->r_ctl.crte->rate))) { /* Use the user mss since we are not exactly matched */ return (user_max); } new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL); if (new_tso > user_max) new_tso = user_max; return (new_tso); } static int32_t pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz, int *capped, uint64_t *rate_wanted, uint8_t non_paced) { uint64_t lentim, fill_bw; /* Lets first see if we are full, if so continue with normal rate */ rack->r_via_fill_cw = 0; if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use) return (slot); if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd) return (slot); if (rack->r_ctl.rc_last_us_rtt == 0) return (slot); if (rack->rc_pace_fill_if_rttin_range && (rack->r_ctl.rc_last_us_rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) { /* The rtt is huge, N * smallest, lets not fill */ return (slot); } /* * first lets calculate the b/w based on the last us-rtt * and the sndwnd. */ fill_bw = rack->r_ctl.cwnd_to_use; /* Take the rwnd if its smaller */ if (fill_bw > rack->rc_tp->snd_wnd) fill_bw = rack->rc_tp->snd_wnd; if (rack->r_fill_less_agg) { /* * Now take away the inflight (this will reduce our * aggressiveness and yeah, if we get that much out in 1RTT * we will have had acks come back and still be behind). */ fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); } /* Now lets make it into a b/w */ fill_bw *= (uint64_t)HPTS_USEC_IN_SEC; fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt; /* We are below the min b/w */ if (non_paced) *rate_wanted = fill_bw; if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted)) return (slot); if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) fill_bw = rack->r_ctl.bw_rate_cap; rack->r_via_fill_cw = 1; if (rack->r_rack_hw_rate_caps && (rack->r_ctl.crte != NULL)) { uint64_t high_rate; high_rate = tcp_hw_highest_rate(rack->r_ctl.crte); if (fill_bw > high_rate) { /* We are capping bw at the highest rate table entry */ if (*rate_wanted > high_rate) { /* The original rate was also capped */ rack->r_via_fill_cw = 0; } rack_log_hdwr_pacing(rack, fill_bw, high_rate, __LINE__, 0, 3); fill_bw = high_rate; if (capped) *capped = 1; } } else if ((rack->r_ctl.crte == NULL) && (rack->rack_hdrw_pacing == 0) && (rack->rack_hdw_pace_ena) && rack->r_rack_hw_rate_caps && (rack->rack_attempt_hdwr_pace == 0) && (rack->rc_inp->inp_route.ro_nh != NULL) && (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) { /* * Ok we may have a first attempt that is greater than our top rate * lets check. */ uint64_t high_rate; high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp); if (high_rate) { if (fill_bw > high_rate) { fill_bw = high_rate; if (capped) *capped = 1; } } } /* * Ok fill_bw holds our mythical b/w to fill the cwnd * in a rtt, what does that time wise equate too? */ lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC; lentim /= fill_bw; *rate_wanted = fill_bw; if (non_paced || (lentim < slot)) { rack_log_pacing_delay_calc(rack, len, slot, fill_bw, 0, lentim, 12, __LINE__, NULL, 0); return ((int32_t)lentim); } else return (slot); } static int32_t rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz) { int32_t slot = 0; int can_start_hw_pacing = 1; int err; if (rack->rc_always_pace == 0) { /* * We use the most optimistic possible cwnd/srtt for * sending calculations. This will make our * calculation anticipate getting more through * quicker then possible. But thats ok we don't want * the peer to have a gap in data sending. */ uint32_t srtt, cwnd, tr_perms = 0; int32_t reduce = 0; old_method: /* * We keep no precise pacing with the old method * instead we use the pacer to mitigate bursts. */ if (rack->r_ctl.rc_rack_min_rtt) srtt = rack->r_ctl.rc_rack_min_rtt; else srtt = max(tp->t_srtt, 1); if (rack->r_ctl.rc_rack_largest_cwnd) cwnd = rack->r_ctl.rc_rack_largest_cwnd; else cwnd = rack->r_ctl.cwnd_to_use; /* Inflate cwnd by 1000 so srtt of usecs is in ms */ tr_perms = (cwnd * 1000) / srtt; if (tr_perms == 0) { tr_perms = ctf_fixed_maxseg(tp); } /* * Calculate how long this will take to drain, if * the calculation comes out to zero, thats ok we * will use send_a_lot to possibly spin around for * more increasing tot_len_this_send to the point * that its going to require a pace, or we hit the * cwnd. Which in that case we are just waiting for * a ACK. */ slot = len / tr_perms; /* Now do we reduce the time so we don't run dry? */ if (slot && rack_slot_reduction) { reduce = (slot / rack_slot_reduction); if (reduce < slot) { slot -= reduce; } else slot = 0; } slot *= HPTS_USEC_IN_MSEC; if (rack->rc_pace_to_cwnd) { uint64_t rate_wanted = 0; slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1); rack->rc_ack_can_sendout_data = 1; rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0); } else rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0); } else { uint64_t bw_est, res, lentim, rate_wanted; uint32_t orig_val, srtt, segs, oh; int capped = 0; int prev_fill; if ((rack->r_rr_config == 1) && rsm) { return (rack->r_ctl.rc_min_to); } if (rack->use_fixed_rate) { rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack); } else if ((rack->r_ctl.init_rate == 0) && #ifdef NETFLIX_PEAKRATE (rack->rc_tp->t_maxpeakrate == 0) && #endif (rack->r_ctl.gp_bw == 0)) { /* no way to yet do an estimate */ bw_est = rate_wanted = 0; } else { bw_est = rack_get_bw(rack); rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped); } if ((bw_est == 0) || (rate_wanted == 0) || ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) { /* * No way yet to make a b/w estimate or * our raise is set incorrectly. */ goto old_method; } /* We need to account for all the overheads */ segs = (len + segsiz - 1) / segsiz; /* * We need the diff between 1514 bytes (e-mtu with e-hdr) * and how much data we put in each packet. Yes this * means we may be off if we are larger than 1500 bytes * or smaller. But this just makes us more conservative. */ if (rack_hw_rate_min && (bw_est < rack_hw_rate_min)) can_start_hw_pacing = 0; if (ETHERNET_SEGMENT_SIZE > segsiz) oh = ETHERNET_SEGMENT_SIZE - segsiz; else oh = 0; segs *= oh; lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC; res = lentim / rate_wanted; slot = (uint32_t)res; orig_val = rack->r_ctl.rc_pace_max_segs; if (rack->r_ctl.crte == NULL) { /* * Only do this if we are not hardware pacing * since if we are doing hw-pacing below we will * set make a call after setting up or changing * the rate. */ rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else if (rack->rc_inp->inp_snd_tag == NULL) { /* * We lost our rate somehow, this can happen * if the interface changed underneath us. */ tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp); rack->r_ctl.crte = NULL; /* Lets re-allow attempting to setup pacing */ rack->rack_hdrw_pacing = 0; rack->rack_attempt_hdwr_pace = 0; rack_log_hdwr_pacing(rack, rate_wanted, bw_est, __LINE__, 0, 6); } /* Did we change the TSO size, if so log it */ if (rack->r_ctl.rc_pace_max_segs != orig_val) rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0); prev_fill = rack->r_via_fill_cw; if ((rack->rc_pace_to_cwnd) && (capped == 0) && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) { /* * We want to pace at our rate *or* faster to * fill the cwnd to the max if its not full. */ slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0); } if ((rack->rc_inp->inp_route.ro_nh != NULL) && (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) { if ((rack->rack_hdw_pace_ena) && (can_start_hw_pacing > 0) && (rack->rack_hdrw_pacing == 0) && (rack->rack_attempt_hdwr_pace == 0)) { /* * Lets attempt to turn on hardware pacing * if we can. */ rack->rack_attempt_hdwr_pace = 1; rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp, rack->rc_inp->inp_route.ro_nh->nh_ifp, rate_wanted, RS_PACING_GEQ, &err, &rack->r_ctl.crte_prev_rate); if (rack->r_ctl.crte) { rack->rack_hdrw_pacing = 1; rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz, 0, rack->r_ctl.crte, NULL); rack_log_hdwr_pacing(rack, rate_wanted, rack->r_ctl.crte->rate, __LINE__, err, 0); rack->r_ctl.last_hw_bw_req = rate_wanted; } else { counter_u64_add(rack_hw_pace_init_fail, 1); } } else if (rack->rack_hdrw_pacing && (rack->r_ctl.last_hw_bw_req != rate_wanted)) { /* Do we need to adjust our rate? */ const struct tcp_hwrate_limit_table *nrte; if (rack->r_up_only && (rate_wanted < rack->r_ctl.crte->rate)) { /** * We have four possible states here * having to do with the previous time * and this time. * previous | this-time * A) 0 | 0 -- fill_cw not in the picture * B) 1 | 0 -- we were doing a fill-cw but now are not * C) 1 | 1 -- all rates from fill_cw * D) 0 | 1 -- we were doing non-fill and now we are filling * * For case A, C and D we don't allow a drop. But for * case B where we now our on our steady rate we do * allow a drop. * */ if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0))) goto done_w_hdwr; } if ((rate_wanted > rack->r_ctl.crte->rate) || (rate_wanted <= rack->r_ctl.crte_prev_rate)) { if (rack_hw_rate_to_low && (bw_est < rack_hw_rate_to_low)) { /* * The pacing rate is too low for hardware, but * do allow hardware pacing to be restarted. */ rack_log_hdwr_pacing(rack, bw_est, rack->r_ctl.crte->rate, __LINE__, 0, 5); tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp); rack->r_ctl.crte = NULL; rack->rack_attempt_hdwr_pace = 0; rack->rack_hdrw_pacing = 0; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted); goto done_w_hdwr; } nrte = tcp_chg_pacing_rate(rack->r_ctl.crte, rack->rc_tp, rack->rc_inp->inp_route.ro_nh->nh_ifp, rate_wanted, RS_PACING_GEQ, &err, &rack->r_ctl.crte_prev_rate); if (nrte == NULL) { /* Lost the rate */ rack->rack_hdrw_pacing = 0; rack->r_ctl.crte = NULL; rack_log_hdwr_pacing(rack, rate_wanted, 0, __LINE__, err, 1); rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted); counter_u64_add(rack_hw_pace_lost, 1); } else if (nrte != rack->r_ctl.crte) { rack->r_ctl.crte = nrte; rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz, 0, rack->r_ctl.crte, NULL); rack_log_hdwr_pacing(rack, rate_wanted, rack->r_ctl.crte->rate, __LINE__, err, 2); rack->r_ctl.last_hw_bw_req = rate_wanted; } } else { /* We just need to adjust the segment size */ rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted); rack_log_hdwr_pacing(rack, rate_wanted, rack->r_ctl.crte->rate, __LINE__, 0, 4); rack->r_ctl.last_hw_bw_req = rate_wanted; } } } if ((rack->r_ctl.crte != NULL) && (rack->r_ctl.crte->rate == rate_wanted)) { /* * We need to add a extra if the rates * are exactly matched. The idea is * we want the software to make sure the * queue is empty before adding more, this * gives us N MSS extra pace times where * N is our sysctl */ slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots); } done_w_hdwr: if (rack_limit_time_with_srtt && (rack->use_fixed_rate == 0) && #ifdef NETFLIX_PEAKRATE (rack->rc_tp->t_maxpeakrate == 0) && #endif (rack->rack_hdrw_pacing == 0)) { /* * Sanity check, we do not allow the pacing delay * to be longer than the SRTT of the path. If it is * a slow path, then adding a packet should increase * the RTT and compensate for this i.e. the srtt will * be greater so the allowed pacing time will be greater. * * Note this restriction is not for where a peak rate * is set, we are doing fixed pacing or hardware pacing. */ if (rack->rc_tp->t_srtt) srtt = rack->rc_tp->t_srtt; else srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC; /* its in ms convert */ if (srtt < slot) { rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0); slot = srtt; } } rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0); } if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) { /* * If this rate is seeing enobufs when it * goes to send then either the nic is out * of gas or we are mis-estimating the time * somehow and not letting the queue empty * completely. Lets add to the pacing time. */ int hw_boost_delay; hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult; if (hw_boost_delay > rack_enobuf_hw_max) hw_boost_delay = rack_enobuf_hw_max; else if (hw_boost_delay < rack_enobuf_hw_min) hw_boost_delay = rack_enobuf_hw_min; slot += hw_boost_delay; } if (slot) counter_u64_add(rack_calc_nonzero, 1); else counter_u64_add(rack_calc_zero, 1); return (slot); } static void rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq startseq, uint32_t sb_offset) { struct rack_sendmap *my_rsm = NULL; struct rack_sendmap fe; if (tp->t_state < TCPS_ESTABLISHED) { /* * We don't start any measurements if we are * not at least established. */ return; } if (tp->t_state >= TCPS_FIN_WAIT_1) { /* * We will get no more data into the SB * this means we need to have the data available * before we start a measurement. */ if (sbavail(&tp->t_inpcb->inp_socket->so_snd) < max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) { /* Nope not enough data */ return; } } tp->t_flags |= TF_GPUTINPROG; rack->r_ctl.rc_gp_lowrtt = 0xffffffff; rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd; tp->gput_seq = startseq; rack->app_limited_needs_set = 0; if (rack->in_probe_rtt) rack->measure_saw_probe_rtt = 1; else if ((rack->measure_saw_probe_rtt) && (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit))) rack->measure_saw_probe_rtt = 0; if (rack->rc_gp_filled) tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); else { /* Special case initial measurement */ struct timeval tv; tp->gput_ts = tcp_get_usecs(&tv); rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv); } /* * We take a guess out into the future, * if we have no measurement and no * initial rate, we measure the first * initial-windows worth of data to * speed up getting some GP measurement and * thus start pacing. */ if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) { rack->app_limited_needs_set = 1; tp->gput_ack = startseq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp))); rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, 0, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL, 0); return; } if (sb_offset) { /* * We are out somewhere in the sb * can we use the already outstanding data? */ if (rack->r_ctl.rc_app_limited_cnt == 0) { /* * Yes first one is good and in this case * the tp->gput_ts is correctly set based on * the last ack that arrived (no need to * set things up when an ack comes in). */ my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if ((my_rsm == NULL) || (my_rsm->r_rtr_cnt != 1)) { /* retransmission? */ goto use_latest; } } else { if (rack->r_ctl.rc_first_appl == NULL) { /* * If rc_first_appl is NULL * then the cnt should be 0. * This is probably an error, maybe * a KASSERT would be approprate. */ goto use_latest; } /* * If we have a marker pointer to the last one that is * app limited we can use that, but we need to set * things up so that when it gets ack'ed we record * the ack time (if its not already acked). */ rack->app_limited_needs_set = 1; /* * We want to get to the rsm that is either * next with space i.e. over 1 MSS or the one * after that (after the app-limited). */ my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rack->r_ctl.rc_first_appl); if (my_rsm) { if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp)) /* Have to use the next one */ my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, my_rsm); else { /* Use after the first MSS of it is acked */ tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp); goto start_set; } } if ((my_rsm == NULL) || (my_rsm->r_rtr_cnt != 1)) { /* * Either its a retransmit or * the last is the app-limited one. */ goto use_latest; } } tp->gput_seq = my_rsm->r_start; start_set: if (my_rsm->r_flags & RACK_ACKED) { /* * This one has been acked use the arrival ack time */ tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival; rack->app_limited_needs_set = 0; } rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)]; tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack); rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, (uint64_t)my_rsm, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL, 0); return; } use_latest: /* * We don't know how long we may have been * idle or if this is the first-send. Lets * setup the flag so we will trim off * the first ack'd data so we get a true * measurement. */ rack->app_limited_needs_set = 1; tp->gput_ack = startseq + rack_get_measure_window(tp, rack); /* Find this guy so we can pull the send time */ fe.r_start = startseq; my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (my_rsm) { rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)]; if (my_rsm->r_flags & RACK_ACKED) { /* * Unlikely since its probably what was * just transmitted (but I am paranoid). */ tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival; rack->app_limited_needs_set = 0; } if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) { /* This also is unlikely */ tp->gput_seq = my_rsm->r_start; } } else { /* * TSNH unless we have some send-map limit, * and even at that it should not be hitting * that limit (we should have stopped sending). */ struct timeval tv; microuptime(&tv); rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv); } rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, (uint64_t)my_rsm, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL, 0); } static inline uint32_t rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cwnd_to_use, uint32_t avail, int32_t sb_offset) { uint32_t len; uint32_t sendwin; if (tp->snd_wnd > cwnd_to_use) sendwin = cwnd_to_use; else sendwin = tp->snd_wnd; if (ctf_outstanding(tp) >= tp->snd_wnd) { /* We never want to go over our peers rcv-window */ len = 0; } else { uint32_t flight; flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked); if (flight >= sendwin) { /* * We have in flight what we are allowed by cwnd (if * it was rwnd blocking it would have hit above out * >= tp->snd_wnd). */ return (0); } len = sendwin - flight; if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { /* We would send too much (beyond the rwnd) */ len = tp->snd_wnd - ctf_outstanding(tp); } if ((len + sb_offset) > avail) { /* * We don't have that much in the SB, how much is * there? */ len = avail - sb_offset; } } return (len); } static void rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags, unsigned ipoptlen, int32_t orig_len, int32_t len, int error, int rsm_is_null, int optlen, int line, uint16_t mode) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = error; log.u_bbr.flex2 = flags; log.u_bbr.flex3 = rsm_is_null; log.u_bbr.flex4 = ipoptlen; log.u_bbr.flex5 = tp->rcv_numsacks; log.u_bbr.flex6 = rack->r_ctl.rc_agg_early; log.u_bbr.flex7 = optlen; log.u_bbr.flex8 = rack->r_fsb_inited; log.u_bbr.applimited = rack->r_fast_output; log.u_bbr.bw_inuse = rack_get_bw(rack); log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL); log.u_bbr.cwnd_gain = mode; log.u_bbr.pkts_out = orig_len; log.u_bbr.lt_epoch = len; log.u_bbr.delivered = line; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0, len, &log, false, NULL, NULL, 0, &tv); } } static struct mbuf * rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen, struct rack_fast_send_blk *fsb, int32_t seglimit, int32_t segsize, int hw_tls) { #ifdef KERN_TLS struct ktls_session *tls, *ntls; struct mbuf *start; #endif struct mbuf *m, *n, **np, *smb; struct mbuf *top; int32_t off, soff; int32_t len = *plen; int32_t fragsize; int32_t len_cp = 0; uint32_t mlen, frags; soff = off = the_off; smb = m = the_m; np = ⊤ top = NULL; #ifdef KERN_TLS if (hw_tls && (m->m_flags & M_EXTPG)) tls = m->m_epg_tls; else tls = NULL; start = m; #endif while (len > 0) { if (m == NULL) { *plen = len_cp; break; } #ifdef KERN_TLS if (hw_tls) { if (m->m_flags & M_EXTPG) ntls = m->m_epg_tls; else ntls = NULL; /* * Avoid mixing TLS records with handshake * data or TLS records from different * sessions. */ if (tls != ntls) { MPASS(m != start); *plen = len_cp; break; } } #endif mlen = min(len, m->m_len - off); if (seglimit) { /* * For M_EXTPG mbufs, add 3 segments * + 1 in case we are crossing page boundaries * + 2 in case the TLS hdr/trailer are used * It is cheaper to just add the segments * than it is to take the cache miss to look * at the mbuf ext_pgs state in detail. */ if (m->m_flags & M_EXTPG) { fragsize = min(segsize, PAGE_SIZE); frags = 3; } else { fragsize = segsize; frags = 0; } /* Break if we really can't fit anymore. */ if ((frags + 1) >= seglimit) { *plen = len_cp; break; } /* * Reduce size if you can't copy the whole * mbuf. If we can't copy the whole mbuf, also * adjust len so the loop will end after this * mbuf. */ if ((frags + howmany(mlen, fragsize)) >= seglimit) { mlen = (seglimit - frags - 1) * fragsize; len = mlen; *plen = len_cp + len; } frags += howmany(mlen, fragsize); if (frags == 0) frags++; seglimit -= frags; KASSERT(seglimit > 0, ("%s: seglimit went too low", __func__)); } n = m_get(M_NOWAIT, m->m_type); *np = n; if (n == NULL) goto nospace; n->m_len = mlen; soff += mlen; len_cp += n->m_len; if (m->m_flags & (M_EXT|M_EXTPG)) { n->m_data = m->m_data + off; mb_dupcl(n, m); } else { bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), (u_int)n->m_len); } len -= n->m_len; off = 0; m = m->m_next; np = &n->m_next; if (len || (soff == smb->m_len)) { /* * We have more so we move forward or * we have consumed the entire mbuf and * len has fell to 0. */ soff = 0; smb = m; } } if (fsb != NULL) { fsb->m = smb; fsb->off = soff; if (smb) { /* * Save off the size of the mbuf. We do * this so that we can recognize when it * has been trimmed by sbcut() as acks * come in. */ fsb->o_m_len = smb->m_len; } else { /* * This is the case where the next mbuf went to NULL. This * means with this copy we have sent everything in the sb. * In theory we could clear the fast_output flag, but lets * not since its possible that we could get more added * and acks that call the extend function which would let * us send more. */ fsb->o_m_len = 0; } } return (top); nospace: if (top) m_freem(top); return (NULL); } /* * This is a copy of m_copym(), taking the TSO segment size/limit * constraints into account, and advancing the sndptr as it goes. */ static struct mbuf * rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen, int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff) { struct mbuf *m, *n; int32_t soff; soff = rack->r_ctl.fsb.off; m = rack->r_ctl.fsb.m; if (rack->r_ctl.fsb.o_m_len != m->m_len) { /* * The mbuf had the front of it chopped off by an ack * we need to adjust the soff/off by that difference. */ uint32_t delta; delta = rack->r_ctl.fsb.o_m_len - m->m_len; soff -= delta; } KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff)); KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen)); KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?", __FUNCTION__, rack, *plen, m, m->m_len)); /* Save off the right location before we copy and advance */ *s_soff = soff; *s_mb = rack->r_ctl.fsb.m; n = rack_fo_base_copym(m, soff, plen, &rack->r_ctl.fsb, seglimit, segsize, rack->r_ctl.fsb.hw_tls); return (n); } static int rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp) { /* * Enter the fast retransmit path. We are given that a sched_pin is * in place (if accounting is compliled in) and the cycle count taken * at the entry is in the ts_val. The concept her is that the rsm * now holds the mbuf offsets and such so we can directly transmit * without a lot of overhead, the len field is already set for * us to prohibit us from sending too much (usually its 1MSS). */ struct ip *ip = NULL; struct udphdr *udp = NULL; struct tcphdr *th = NULL; struct mbuf *m = NULL; struct inpcb *inp; uint8_t *cpto; struct tcp_log_buffer *lgb; #ifdef TCP_ACCOUNTING uint64_t crtsc; int cnt_thru = 1; #endif struct tcpopt to; u_char opt[TCP_MAXOLEN]; uint32_t hdrlen, optlen; int32_t slot, segsiz, max_val, tso = 0, error, flags, ulen = 0; uint32_t us_cts; uint32_t if_hw_tsomaxsegcount = 0, startseq; uint32_t if_hw_tsomaxsegsize; #ifdef INET6 struct ip6_hdr *ip6 = NULL; if (rack->r_is_v6) { ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); } else #endif /* INET6 */ { ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct tcpiphdr); } if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) { goto failed; } if (rsm->r_flags & RACK_TLP) doing_tlp = 1; else if (doing_tlp) rsm->r_flags |= RACK_TLP; startseq = rsm->r_start; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); inp = rack->rc_inp; to.to_flags = 0; flags = tcp_outflags[tp->t_state]; if (flags & (TH_SYN|TH_RST)) { goto failed; } if (rsm->r_flags & RACK_HAS_FIN) { /* We can't send a FIN here */ goto failed; } if (flags & TH_FIN) { /* We never send a FIN */ flags &= ~TH_FIN; } if (tp->t_flags & TF_RCVD_TSTMP) { to.to_tsval = ms_cts + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags = TOF_TS; } optlen = tcp_addoptions(&to, opt); hdrlen += optlen; udp = rack->r_ctl.fsb.udp; if (udp) hdrlen += sizeof(struct udphdr); if (rack->r_ctl.rc_pace_max_segs) max_val = rack->r_ctl.rc_pace_max_segs; else if (rack->rc_user_set_max_segs) max_val = rack->rc_user_set_max_segs * segsiz; else max_val = len; if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && (len > segsiz) && (tp->t_port == 0)) tso = 1; #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) goto failed; m->m_data += max_linkhdr; m->m_len = hdrlen; th = rack->r_ctl.fsb.th; /* Establish the len to send */ if (len > max_val) len = max_val; if ((tso) && (len + optlen > tp->t_maxseg)) { uint32_t if_hw_tsomax; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { goto failed; } else if (len > max_len) { len = max_len; } } if (len <= segsiz) { /* * In case there are too many small fragments don't * use TSO: */ tso = 0; } } else { tso = 0; } if ((tso == 0) && (len > segsiz)) len = segsiz; us_cts = tcp_get_usecs(tv); if ((len == 0) || (len <= MHLEN - hdrlen - max_linkhdr)) { goto failed; } th->th_seq = htonl(rsm->r_start); th->th_ack = htonl(tp->rcv_nxt); /* * The PUSH bit should only be applied * if the full retransmission is made. If * we are sending less than this is the * left hand edge and should not have * the PUSH bit. */ if ((rsm->r_flags & RACK_HAD_PUSH) && (len == (rsm->r_end - rsm->r_start))) flags |= TH_PUSH; th->th_flags = flags; th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale)); if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; if (rsm->r_flags & RACK_TLP) { /* * TLP should not count in retran count, but * in its own bin */ counter_u64_add(rack_tlp_retran, 1); counter_u64_add(rack_tlp_retran_bytes, len); } else { tp->t_sndrexmitpack++; KMOD_TCPSTAT_INC(tcps_sndrexmitpack); KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); } #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif if (rsm->m == NULL) goto failed; if (rsm->orig_m_len != rsm->m->m_len) { /* Fix up the orig_m_len and possibly the mbuf offset */ rack_adjust_orig_mlen(rsm); } m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls); if (len <= segsiz) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if ((m->m_next == NULL) || (len <= 0)){ goto failed; } if (udp) { if (rack->r_is_v6) ulen = hdrlen + len - sizeof(struct ip6_hdr); else ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); } m->m_pkthdr.rcvif = (struct ifnet *)0; m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (rack->r_is_v6) { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz tp:%p", __func__, tp)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } #ifdef INET6 if (rack->r_is_v6) { ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit; ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); ip->ip_ttl = rack->r_ctl.fsb.hoplimit; if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } } #endif /* Time to copy in our header */ cpto = mtod(m, uint8_t *); memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len); th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr)); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } else { th->th_off = sizeof(struct tcphdr) >> 2; } if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; if (rack->rack_no_prr) log.u_bbr.flex1 = 0; else log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = max_val; log.u_bbr.flex5 = 0; /* Save off the early/late values */ log.u_bbr.flex6 = rack->r_ctl.rc_agg_early; log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed; log.u_bbr.bw_inuse = rack_get_bw(rack); if (doing_tlp == 0) log.u_bbr.flex8 = 1; else log.u_bbr.flex8 = 2; log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL); log.u_bbr.flex7 = 55; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use; log.u_bbr.delivered = 0; lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, tv); } else lgb = NULL; #ifdef INET6 if (rack->r_is_v6) { error = ip6_output(m, NULL, &inp->inp_route6, 0, NULL, NULL, inp); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { error = ip_output(m, NULL, &inp->inp_route, 0, 0, inp); } #endif m = NULL; if (lgb) { lgb->tlb_errno = error; lgb = NULL; } if (error) { goto failed; } rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv), rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls); if (doing_tlp && (rack->fast_rsm_hack == 0)) { rack->rc_tlp_in_progress = 1; rack->r_ctl.rc_tlp_cnt_out++; } if (error == 0) tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls); tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); rack->forced_ack = 0; /* If we send something zap the FA flag */ if (IN_FASTRECOVERY(tp->t_flags) && rsm) rack->r_ctl.retran_during_recovery += len; { int idx; idx = (len / segsiz) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); else counter_u64_add(rack_out_size[idx], 1); } if (tp->t_rtttime == 0) { tp->t_rtttime = ticks; tp->t_rtseq = startseq; KMOD_TCPSTAT_INC(tcps_segstimed); } counter_u64_add(rack_fto_rsm_send, 1); if (error && (error == ENOBUFS)) { slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC); if (rack->rc_enobuf < 0x7f) rack->rc_enobuf++; if (slot < (10 * HPTS_USEC_IN_MSEC)) slot = 10 * HPTS_USEC_IN_MSEC; } else slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz); if ((slot == 0) || (rack->rc_always_pace == 0) || (rack->r_rr_config == 1)) { /* * We have no pacing set or we * are using old-style rack or * we are overriden to use the old 1ms pacing. */ slot = rack->r_ctl.rc_min_to; } rack_start_hpts_timer(rack, tp, cts, slot, len, 0); if (rack->r_must_retran) { rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start); if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) { /* * We have retransmitted all we need. */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; } } #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz)); sched_unpin(); #endif return (0); failed: if (m) m_free(m); return (-1); } static void rack_sndbuf_autoscale(struct tcp_rack *rack) { /* * Automatic sizing of send socket buffer. Often the send buffer * size is not optimally adjusted to the actual network conditions * at hand (delay bandwidth product). Setting the buffer size too * small limits throughput on links with high bandwidth and high * delay (eg. trans-continental/oceanic links). Setting the * buffer size too big consumes too much real kernel memory, * especially with many connections on busy servers. * * The criteria to step up the send buffer one notch are: * 1. receive window of remote host is larger than send buffer * (with a fudge factor of 5/4th); * 2. send buffer is filled to 7/8th with data (so we actually * have data to make use of it); * 3. send buffer fill has not hit maximal automatic size; * 4. our send window (slow start and cogestion controlled) is * larger than sent but unacknowledged data in send buffer. * * Note that the rack version moves things much faster since * we want to avoid hitting cache lines in the rack_fast_output() * path so this is called much less often and thus moves * the SB forward by a percentage. */ struct socket *so; struct tcpcb *tp; uint32_t sendwin, scaleup; tp = rack->rc_tp; so = rack->rc_inp->inp_socket; sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd); if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) && sbused(&so->so_snd) < V_tcp_autosndbuf_max && sendwin >= (sbused(&so->so_snd) - (tp->snd_nxt - tp->snd_una))) { if (rack_autosndbuf_inc) scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100; else scaleup = V_tcp_autosndbuf_inc; if (scaleup < V_tcp_autosndbuf_inc) scaleup = V_tcp_autosndbuf_inc; scaleup += so->so_snd.sb_hiwat; if (scaleup > V_tcp_autosndbuf_max) scaleup = V_tcp_autosndbuf_max; if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread)) so->so_snd.sb_flags &= ~SB_AUTOSIZE; } } } static int rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err) { /* * Enter to do fast output. We are given that the sched_pin is * in place (if accounting is compiled in) and the cycle count taken * at entry is in place in ts_val. The idea here is that * we know how many more bytes needs to be sent (presumably either * during pacing or to fill the cwnd and that was greater than * the max-burst). We have how much to send and all the info we * need to just send. */ struct ip *ip = NULL; struct udphdr *udp = NULL; struct tcphdr *th = NULL; struct mbuf *m, *s_mb; struct inpcb *inp; uint8_t *cpto; struct tcp_log_buffer *lgb; #ifdef TCP_ACCOUNTING uint64_t crtsc; #endif struct tcpopt to; u_char opt[TCP_MAXOLEN]; uint32_t hdrlen, optlen; int cnt_thru = 1; int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, flags, ulen = 0; uint32_t us_cts, s_soff; uint32_t if_hw_tsomaxsegcount = 0, startseq; uint32_t if_hw_tsomaxsegsize; uint16_t add_flag = RACK_SENT_FP; #ifdef INET6 struct ip6_hdr *ip6 = NULL; if (rack->r_is_v6) { ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); } else #endif /* INET6 */ { ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct tcpiphdr); } if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) { m = NULL; goto failed; } startseq = tp->snd_max; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); inp = rack->rc_inp; len = rack->r_ctl.fsb.left_to_send; to.to_flags = 0; flags = rack->r_ctl.fsb.tcp_flags; if (tp->t_flags & TF_RCVD_TSTMP) { to.to_tsval = ms_cts + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags = TOF_TS; } optlen = tcp_addoptions(&to, opt); hdrlen += optlen; udp = rack->r_ctl.fsb.udp; if (udp) hdrlen += sizeof(struct udphdr); if (rack->r_ctl.rc_pace_max_segs) max_val = rack->r_ctl.rc_pace_max_segs; else if (rack->rc_user_set_max_segs) max_val = rack->rc_user_set_max_segs * segsiz; else max_val = len; if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && (len > segsiz) && (tp->t_port == 0)) tso = 1; again: #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) goto failed; m->m_data += max_linkhdr; m->m_len = hdrlen; th = rack->r_ctl.fsb.th; /* Establish the len to send */ if (len > max_val) len = max_val; if ((tso) && (len + optlen > tp->t_maxseg)) { uint32_t if_hw_tsomax; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { goto failed; } else if (len > max_len) { len = max_len; } } if (len <= segsiz) { /* * In case there are too many small fragments don't * use TSO: */ tso = 0; } } else { tso = 0; } if ((tso == 0) && (len > segsiz)) len = segsiz; us_cts = tcp_get_usecs(tv); if ((len == 0) || (len <= MHLEN - hdrlen - max_linkhdr)) { goto failed; } sb_offset = tp->snd_max - tp->snd_una; th->th_seq = htonl(tp->snd_max); th->th_ack = htonl(tp->rcv_nxt); th->th_flags = flags; th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale)); if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; tp->snd_up = tp->snd_una; /* drag it along, its deprecated */ KMOD_TCPSTAT_INC(tcps_sndpack); KMOD_TCPSTAT_ADD(tcps_sndbyte, len); #ifdef STATS stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif if (rack->r_ctl.fsb.m == NULL) goto failed; /* s_mb and s_soff are saved for rack_log_output */ m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, &s_mb, &s_soff); if (len <= segsiz) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if (rack->r_ctl.fsb.rfo_apply_push && (len == rack->r_ctl.fsb.left_to_send)) { th->th_flags |= TH_PUSH; add_flag |= RACK_HAD_PUSH; } if ((m->m_next == NULL) || (len <= 0)){ goto failed; } if (udp) { if (rack->r_is_v6) ulen = hdrlen + len - sizeof(struct ip6_hdr); else ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); } m->m_pkthdr.rcvif = (struct ifnet *)0; if (tp->t_state == TCPS_ESTABLISHED && (tp->t_flags2 & TF2_ECN_PERMIT)) { /* * If the peer has ECN, mark data packets with ECN capable * transmission (ECT). Ignore pure ack packets, * retransmissions. */ if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max)) { #ifdef INET6 if (rack->r_is_v6) ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); else #endif ip->ip_tos |= IPTOS_ECN_ECT0; KMOD_TCPSTAT_INC(tcps_ecn_ect0); /* * Reply with proper ECN notifications. * Only set CWR on new data segments. */ if (tp->t_flags2 & TF2_ECN_SND_CWR) { flags |= TH_CWR; tp->t_flags2 &= ~TF2_ECN_SND_CWR; } } if (tp->t_flags2 & TF2_ECN_SND_ECE) flags |= TH_ECE; } m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (rack->r_is_v6) { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz tp:%p", __func__, tp)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } #ifdef INET6 if (rack->r_is_v6) { ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit; ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); ip->ip_ttl = rack->r_ctl.fsb.hoplimit; if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } } #endif /* Time to copy in our header */ cpto = mtod(m, uint8_t *); memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len); th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr)); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } else { th->th_off = sizeof(struct tcphdr) >> 2; } if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; if (rack->rack_no_prr) log.u_bbr.flex1 = 0; else log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = max_val; log.u_bbr.flex5 = 0; /* Save off the early/late values */ log.u_bbr.flex6 = rack->r_ctl.rc_agg_early; log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed; log.u_bbr.bw_inuse = rack_get_bw(rack); log.u_bbr.flex8 = 0; log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL); log.u_bbr.flex7 = 44; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use; log.u_bbr.delivered = 0; lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, tv); } else lgb = NULL; #ifdef INET6 if (rack->r_is_v6) { error = ip6_output(m, NULL, &inp->inp_route6, 0, NULL, NULL, inp); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { error = ip_output(m, NULL, &inp->inp_route, 0, 0, inp); } #endif if (lgb) { lgb->tlb_errno = error; lgb = NULL; } if (error) { *send_err = error; m = NULL; goto failed; } rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv), NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls); m = NULL; if (tp->snd_una == tp->snd_max) { rack->r_ctl.rc_tlp_rxt_last_time = cts; rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); tp->t_acktime = ticks; } if (error == 0) tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls); rack->forced_ack = 0; /* If we send something zap the FA flag */ tot_len += len; if ((tp->t_flags & TF_GPUTINPROG) == 0) rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset); tp->snd_max += len; tp->snd_nxt = tp->snd_max; { int idx; idx = (len / segsiz) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); else counter_u64_add(rack_out_size[idx], 1); } if (len <= rack->r_ctl.fsb.left_to_send) rack->r_ctl.fsb.left_to_send -= len; else rack->r_ctl.fsb.left_to_send = 0; if (rack->r_ctl.fsb.left_to_send < segsiz) { rack->r_fast_output = 0; rack->r_ctl.fsb.left_to_send = 0; /* At the end of fast_output scale up the sb */ SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd); rack_sndbuf_autoscale(rack); SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd); } if (tp->t_rtttime == 0) { tp->t_rtttime = ticks; tp->t_rtseq = startseq; KMOD_TCPSTAT_INC(tcps_segstimed); } if ((rack->r_ctl.fsb.left_to_send >= segsiz) && (max_val > len) && (tso == 0)) { max_val -= len; len = segsiz; th = rack->r_ctl.fsb.th; cnt_thru++; goto again; } tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); counter_u64_add(rack_fto_send, 1); slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz); rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz)); sched_unpin(); #endif return (0); failed: if (m) m_free(m); rack->r_fast_output = 0; return (-1); } static int rack_output(struct tcpcb *tp) { struct socket *so; uint32_t recwin; uint32_t sb_offset, s_moff = 0; int32_t len, flags, error = 0; struct mbuf *m, *s_mb = NULL; struct mbuf *mb; uint32_t if_hw_tsomaxsegcount = 0; uint32_t if_hw_tsomaxsegsize; int32_t segsiz, minseg; long tot_len_this_send = 0; #ifdef INET struct ip *ip = NULL; #endif #ifdef TCPDEBUG struct ipovly *ipov = NULL; #endif struct udphdr *udp = NULL; struct tcp_rack *rack; struct tcphdr *th; uint8_t pass = 0; uint8_t mark = 0; uint8_t wanted_cookie = 0; u_char opt[TCP_MAXOLEN]; unsigned ipoptlen, optlen, hdrlen, ulen=0; uint32_t rack_seq; #if defined(IPSEC) || defined(IPSEC_SUPPORT) unsigned ipsec_optlen = 0; #endif int32_t idle, sendalot; int32_t sub_from_prr = 0; volatile int32_t sack_rxmit; struct rack_sendmap *rsm = NULL; int32_t tso, mtu; struct tcpopt to; int32_t slot = 0; int32_t sup_rack = 0; uint32_t cts, ms_cts, delayed, early; uint16_t add_flag = RACK_SENT_SP; /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */ uint8_t hpts_calling, doing_tlp = 0; uint32_t cwnd_to_use, pace_max_seg; int32_t do_a_prefetch = 0; int32_t prefetch_rsm = 0; int32_t orig_len = 0; struct timeval tv; int32_t prefetch_so_done = 0; struct tcp_log_buffer *lgb; struct inpcb *inp; struct sockbuf *sb; uint64_t ts_val = 0; #ifdef TCP_ACCOUNTING uint64_t crtsc; #endif #ifdef INET6 struct ip6_hdr *ip6 = NULL; int32_t isipv6; #endif uint8_t filled_all = 0; bool hw_tls = false; /* setup and take the cache hits here */ rack = (struct tcp_rack *)tp->t_fb_ptr; #ifdef TCP_ACCOUNTING sched_pin(); ts_val = get_cyclecount(); #endif hpts_calling = rack->rc_inp->inp_hpts_calls; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(rack->rc_inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (tcp_offload_output(tp)); } #endif /* * For TFO connections in SYN_RECEIVED, only allow the initial * SYN|ACK and those sent by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */ (rack->r_ctl.rc_resend == NULL)) { /* not a retransmit */ #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (0); } #ifdef INET6 if (rack->r_state) { /* Use the cache line loaded if possible */ isipv6 = rack->r_is_v6; } else { isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0; } #endif early = 0; cts = tcp_get_usecs(&tv); ms_cts = tcp_tv_to_mssectick(&tv); if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && rack->rc_inp->inp_in_hpts) { /* * We are on the hpts for some timer but not hptsi output. * Remove from the hpts unconditionally. */ rack_timer_cancel(tp, rack, cts, __LINE__); } /* Are we pacing and late? */ if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) { /* We are delayed */ delayed = cts - rack->r_ctl.rc_last_output_to; } else { delayed = 0; } /* Do the timers, which may override the pacer */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { if (rack_process_timers(tp, rack, cts, hpts_calling, &doing_tlp)) { counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (0); } } if (rack->rc_in_persist) { if (rack->rc_inp->inp_in_hpts == 0) { /* Timer is not running */ rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); } #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (0); } if ((rack->r_timer_override) || (rack->rc_ack_can_sendout_data) || (delayed) || (tp->t_state < TCPS_ESTABLISHED)) { rack->rc_ack_can_sendout_data = 0; if (rack->rc_inp->inp_in_hpts) tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); } else if (rack->rc_inp->inp_in_hpts) { /* * On the hpts you can't pass even if ACKNOW is on, we will * when the hpts fires. */ #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_BLOCKED]++; } counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1); sched_unpin(); #endif counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); return (0); } rack->rc_inp->inp_hpts_calls = 0; /* Finish out both pacing early and late accounting */ if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) { early = rack->r_ctl.rc_last_output_to - cts; } else early = 0; if (delayed) { rack->r_ctl.rc_agg_delayed += delayed; rack->r_late = 1; } else if (early) { rack->r_ctl.rc_agg_early += early; rack->r_early = 1; } /* Now that early/late accounting is done turn off the flag */ rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; rack->r_wanted_output = 0; rack->r_timer_override = 0; if ((tp->t_state != rack->r_state) && TCPS_HAVEESTABLISHED(tp->t_state)) { rack_set_state(tp, rack); } if ((rack->r_fast_output) && (doing_tlp == 0) && (tp->rcv_numsacks == 0)) { int ret; error = 0; ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error); if (ret >= 0) return(ret); else if (error) { inp = rack->rc_inp; so = inp->inp_socket; sb = &so->so_snd; goto nomore; } } inp = rack->rc_inp; /* * For TFO connections in SYN_SENT or SYN_RECEIVED, * only allow the initial SYN or SYN|ACK and those sent * by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && ((tp->t_state == TCPS_SYN_RECEIVED) || (tp->t_state == TCPS_SYN_SENT)) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ (tp->t_rxtshift == 0)) { /* not a retransmit */ cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd; so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } /* * Determine length of data that should be transmitted, and flags * that will be used. If there is some data or critical controls * (SYN, RST) to send, then transmit; otherwise, investigate * further. */ idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); if (tp->t_idle_reduce) { if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) rack_cc_after_idle(rack, tp); } tp->t_flags &= ~TF_LASTIDLE; if (idle) { if (tp->t_flags & TF_MORETOCOME) { tp->t_flags |= TF_LASTIDLE; idle = 0; } } if ((tp->snd_una == tp->snd_max) && rack->r_ctl.rc_went_idle_time && TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) { idle = cts - rack->r_ctl.rc_went_idle_time; if (idle > rack_min_probertt_hold) { /* Count as a probe rtt */ if (rack->in_probe_rtt == 0) { rack->r_ctl.rc_lower_rtt_us_cts = cts; rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts; } else { rack_exit_probertt(rack, cts); } } idle = 0; } if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED)) rack_init_fsb_block(tp, rack); again: /* * If we've recently taken a timeout, snd_max will be greater than * snd_nxt. There may be SACK information that allows us to avoid * resending already delivered data. Adjust snd_nxt accordingly. */ sendalot = 0; cts = tcp_get_usecs(&tv); ms_cts = tcp_tv_to_mssectick(&tv); tso = 0; mtu = 0; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); minseg = segsiz; if (rack->r_ctl.rc_pace_max_segs == 0) pace_max_seg = rack->rc_user_set_max_segs * segsiz; else pace_max_seg = rack->r_ctl.rc_pace_max_segs; sb_offset = tp->snd_max - tp->snd_una; cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd; flags = tcp_outflags[tp->t_state]; while (rack->rc_free_cnt < rack_free_cache) { rsm = rack_alloc(rack); if (rsm == NULL) { if (inp->inp_hpts_calls) /* Retry in a ms */ slot = (1 * HPTS_USEC_IN_MSEC); so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt++; rsm = NULL; } if (inp->inp_hpts_calls) inp->inp_hpts_calls = 0; sack_rxmit = 0; len = 0; rsm = NULL; if (flags & TH_RST) { SOCKBUF_LOCK(&inp->inp_socket->so_snd); so = inp->inp_socket; sb = &so->so_snd; goto send; } if (rack->r_ctl.rc_resend) { /* Retransmit timer */ rsm = rack->r_ctl.rc_resend; rack->r_ctl.rc_resend = NULL; rsm->r_flags &= ~RACK_TLP; len = rsm->r_end - rsm->r_start; sack_rxmit = 1; sendalot = 0; KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", __func__, __LINE__, rsm->r_start, tp->snd_una, tp, rack, rsm)); sb_offset = rsm->r_start - tp->snd_una; if (len >= segsiz) len = segsiz; } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) { /* We have a retransmit that takes precedence */ rsm->r_flags &= ~RACK_TLP; if ((!IN_FASTRECOVERY(tp->t_flags)) && ((tp->t_flags & TF_WASFRECOVERY) == 0)) { /* Enter recovery if not induced by a time-out */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; rack_cong_signal(tp, CC_NDUPACK, tp->snd_una); } #ifdef INVARIANTS if (SEQ_LT(rsm->r_start, tp->snd_una)) { panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n", tp, rack, rsm, rsm->r_start, tp->snd_una); } #endif len = rsm->r_end - rsm->r_start; KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", __func__, __LINE__, rsm->r_start, tp->snd_una, tp, rack, rsm)); sb_offset = rsm->r_start - tp->snd_una; sendalot = 0; if (len >= segsiz) len = segsiz; if (len > 0) { sack_rxmit = 1; KMOD_TCPSTAT_INC(tcps_sack_rexmits); KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, min(len, segsiz)); counter_u64_add(rack_rtm_prr_retran, 1); } } else if (rack->r_ctl.rc_tlpsend) { /* Tail loss probe */ long cwin; long tlen; /* * Check if we can do a TLP with a RACK'd packet * this can happen if we are not doing the rack * cheat and we skipped to a TLP and it * went off. */ rsm = rack->r_ctl.rc_tlpsend; rsm->r_flags |= RACK_TLP; rack->r_ctl.rc_tlpsend = NULL; sack_rxmit = 1; tlen = rsm->r_end - rsm->r_start; if (tlen > segsiz) tlen = segsiz; KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", __func__, __LINE__, rsm->r_start, tp->snd_una, tp, rack, rsm)); sb_offset = rsm->r_start - tp->snd_una; cwin = min(tp->snd_wnd, tlen); len = cwin; } if (rack->r_must_retran && (rsm == NULL)) { /* * Non-Sack and we had a RTO or MTU change, we * need to retransmit until we reach * the former snd_max (rack->r_ctl.rc_snd_max_at_rto). */ if (SEQ_GT(tp->snd_max, tp->snd_una)) { int sendwin, flight; sendwin = min(tp->snd_wnd, tp->snd_cwnd); flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto); if (flight >= sendwin) { so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); KASSERT(rsm != NULL, ("rsm is NULL rack:%p r_must_retran set", rack)); if (rsm == NULL) { /* TSNH */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; rack->r_must_retran = 0; so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } sack_rxmit = 1; len = rsm->r_end - rsm->r_start; sendalot = 0; sb_offset = rsm->r_start - tp->snd_una; if (len >= segsiz) len = segsiz; } else { /* We must be done if there is nothing outstanding */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; } } /* * Enforce a connection sendmap count limit if set * as long as we are not retransmiting. */ if ((rsm == NULL) && (rack->do_detection == 0) && (V_tcp_map_entries_limit > 0) && (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { counter_u64_add(rack_to_alloc_limited, 1); if (!rack->alloc_limit_reported) { rack->alloc_limit_reported = 1; counter_u64_add(rack_alloc_limited_conns, 1); } so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } if (rsm && (rsm->r_flags & RACK_HAS_FIN)) { /* we are retransmitting the fin */ len--; if (len) { /* * When retransmitting data do *not* include the * FIN. This could happen from a TLP probe. */ flags &= ~TH_FIN; } } #ifdef INVARIANTS /* For debugging */ rack->r_ctl.rc_rsm_at_retran = rsm; #endif if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo && ((rsm->r_flags & RACK_HAS_FIN) == 0)) { int ret; ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp); if (ret == 0) return (0); } so = inp->inp_socket; sb = &so->so_snd; if (do_a_prefetch == 0) { kern_prefetch(sb, &do_a_prefetch); do_a_prefetch = 1; } #ifdef NETFLIX_SHARED_CWND if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) && rack->rack_enable_scwnd) { /* We are doing cwnd sharing */ if (rack->gp_ready && (rack->rack_attempted_scwnd == 0) && (rack->r_ctl.rc_scw == NULL) && tp->t_lib) { /* The pcbid is in, lets make an attempt */ counter_u64_add(rack_try_scwnd, 1); rack->rack_attempted_scwnd = 1; rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp, &rack->r_ctl.rc_scw_index, segsiz); } if (rack->r_ctl.rc_scw && (rack->rack_scwnd_is_idle == 1) && sbavail(&so->so_snd)) { /* we are no longer out of data */ tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 0; } if (rack->r_ctl.rc_scw) { /* First lets update and get the cwnd */ rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index, tp->snd_cwnd, tp->snd_wnd, segsiz); } } #endif /* * Get standard flags, and add SYN or FIN if requested by 'hidden' * state flags. */ if (tp->t_flags & TF_NEEDFIN) flags |= TH_FIN; if (tp->t_flags & TF_NEEDSYN) flags |= TH_SYN; if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { void *end_rsm; end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); if (end_rsm) kern_prefetch(end_rsm, &prefetch_rsm); prefetch_rsm = 1; } SOCKBUF_LOCK(sb); /* * If snd_nxt == snd_max and we have transmitted a FIN, the * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a * negative length. This can also occur when TCP opens up its * congestion window while receiving additional duplicate acks after * fast-retransmit because TCP will reset snd_nxt to snd_max after * the fast-retransmit. * * In the normal retransmit-FIN-only case, however, snd_nxt will be * set to snd_una, the sb_offset will be 0, and the length may wind * up 0. * * If sack_rxmit is true we are retransmitting from the scoreboard * in which case len is already set. */ if ((sack_rxmit == 0) && (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) { uint32_t avail; avail = sbavail(sb); if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail) sb_offset = tp->snd_nxt - tp->snd_una; else sb_offset = 0; if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) { if (rack->r_ctl.rc_tlp_new_data) { /* TLP is forcing out new data */ if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) { rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset); } if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) { if (tp->snd_wnd > sb_offset) len = tp->snd_wnd - sb_offset; else len = 0; } else { len = rack->r_ctl.rc_tlp_new_data; } } else { len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset); } if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) { /* * For prr=off, we need to send only 1 MSS * at a time. We do this because another sack could * be arriving that causes us to send retransmits and * we don't want to be on a long pace due to a larger send * that keeps us from sending out the retransmit. */ len = segsiz; } } else { uint32_t outstanding; /* * We are inside of a Fast recovery episode, this * is caused by a SACK or 3 dup acks. At this point * we have sent all the retransmissions and we rely * on PRR to dictate what we will send in the form of * new data. */ outstanding = tp->snd_max - tp->snd_una; if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) { if (tp->snd_wnd > outstanding) { len = tp->snd_wnd - outstanding; /* Check to see if we have the data */ if ((sb_offset + len) > avail) { /* It does not all fit */ if (avail > sb_offset) len = avail - sb_offset; else len = 0; } } else { len = 0; } } else if (avail > sb_offset) { len = avail - sb_offset; } else { len = 0; } if (len > 0) { if (len > rack->r_ctl.rc_prr_sndcnt) { len = rack->r_ctl.rc_prr_sndcnt; } if (len > 0) { sub_from_prr = 1; counter_u64_add(rack_rtm_prr_newdata, 1); } } if (len > segsiz) { /* * We should never send more than a MSS when * retransmitting or sending new data in prr * mode unless the override flag is on. Most * likely the PRR algorithm is not going to * let us send a lot as well :-) */ if (rack->r_ctl.rc_prr_sendalot == 0) { len = segsiz; } } else if (len < segsiz) { /* * Do we send any? The idea here is if the * send empty's the socket buffer we want to * do it. However if not then lets just wait * for our prr_sndcnt to get bigger. */ long leftinsb; leftinsb = sbavail(sb) - sb_offset; if (leftinsb > len) { /* This send does not empty the sb */ len = 0; } } } } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) { /* * If you have not established * and are not doing FAST OPEN * no data please. */ if ((sack_rxmit == 0) && (!IS_FASTOPEN(tp->t_flags))){ len = 0; sb_offset = 0; } } if (prefetch_so_done == 0) { kern_prefetch(so, &prefetch_so_done); prefetch_so_done = 1; } /* * Lop off SYN bit if it has already been sent. However, if this is * SYN-SENT state and if segment contains data and if we don't know * that foreign host supports TAO, suppress sending segment. */ if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) && ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) { /* * When sending additional segments following a TFO SYN|ACK, * do not include the SYN bit. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED)) flags &= ~TH_SYN; } /* * Be careful not to send data and/or FIN on SYN segments. This * measure is needed to prevent interoperability problems with not * fully conformant TCP implementations. */ if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { len = 0; flags &= ~TH_FIN; } /* * On TFO sockets, ensure no data is sent in the following cases: * * - When retransmitting SYN|ACK on a passively-created socket * * - When retransmitting SYN on an actively created socket * * - When sending a zero-length cookie (cookie request) on an * actively created socket * * - When the socket is in the CLOSED state (RST is being sent) */ if (IS_FASTOPEN(tp->t_flags) && (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || ((tp->t_state == TCPS_SYN_SENT) && (tp->t_tfo_client_cookie_len == 0)) || (flags & TH_RST))) { sack_rxmit = 0; len = 0; } /* Without fast-open there should never be data sent on a SYN */ if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) { tp->snd_nxt = tp->iss; len = 0; } if ((len > segsiz) && (tcp_dsack_block_exists(tp))) { /* We only send 1 MSS if we have a DSACK block */ add_flag |= RACK_SENT_W_DSACK; len = segsiz; } orig_len = len; if (len <= 0) { /* * If FIN has been sent but not acked, but we haven't been * called to retransmit, len will be < 0. Otherwise, window * shrank after we sent into it. If window shrank to 0, * cancel pending retransmit, pull snd_nxt back to (closed) * window, and set the persist timer if it isn't already * going. If the window didn't close completely, just wait * for an ACK. * * We also do a general check here to ensure that we will * set the persist timer when we have data to send, but a * 0-byte window. This makes sure the persist timer is set * even if the packet hits one of the "goto send" lines * below. */ len = 0; if ((tp->snd_wnd == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (tp->snd_una == tp->snd_max) && (sb_offset < (int)sbavail(sb))) { rack_enter_persist(tp, rack, cts); } } else if ((rsm == NULL) && (doing_tlp == 0) && (len < pace_max_seg)) { /* * We are not sending a maximum sized segment for * some reason. Should we not send anything (think * sws or persists)? */ if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (len < minseg) && (len < (int)(sbavail(sb) - sb_offset))) { /* * Here the rwnd is less than * the minimum pacing size, this is not a retransmit, * we are established and * the send is not the last in the socket buffer * we send nothing, and we may enter persists * if nothing is outstanding. */ len = 0; if (tp->snd_max == tp->snd_una) { /* * Nothing out we can * go into persists. */ rack_enter_persist(tp, rack, cts); } } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) && (len < (int)(sbavail(sb) - sb_offset)) && (len < minseg)) { /* * Here we are not retransmitting, and * the cwnd is not so small that we could * not send at least a min size (rxt timer * not having gone off), We have 2 segments or * more already in flight, its not the tail end * of the socket buffer and the cwnd is blocking * us from sending out a minimum pacing segment size. * Lets not send anything. */ len = 0; } else if (((tp->snd_wnd - ctf_outstanding(tp)) < min((rack->r_ctl.rc_high_rwnd/2), minseg)) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) && (len < (int)(sbavail(sb) - sb_offset)) && (TCPS_HAVEESTABLISHED(tp->t_state))) { /* * Here we have a send window but we have * filled it up and we can't send another pacing segment. * We also have in flight more than 2 segments * and we are not completing the sb i.e. we allow * the last bytes of the sb to go out even if * its not a full pacing segment. */ len = 0; } else if ((rack->r_ctl.crte != NULL) && (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) && (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) && (len < (int)(sbavail(sb) - sb_offset))) { /* * Here we are doing hardware pacing, this is not a TLP, * we are not sending a pace max segment size, there is rwnd * room to send at least N pace_max_seg, the cwnd is greater * than or equal to a full pacing segments plus 4 mss and we have 2 or * more segments in flight and its not the tail of the socket buffer. * * We don't want to send instead we need to get more ack's in to * allow us to send a full pacing segment. Normally, if we are pacing * about the right speed, we should have finished our pacing * send as most of the acks have come back if we are at the * right rate. This is a bit fuzzy since return path delay * can delay the acks, which is why we want to make sure we * have cwnd space to have a bit more than a max pace segments in flight. * * If we have not gotten our acks back we are pacing at too high a * rate delaying will not hurt and will bring our GP estimate down by * injecting the delay. If we don't do this we will send * 2 MSS out in response to the acks being clocked in which * defeats the point of hw-pacing (i.e. to help us get * larger TSO's out). */ len = 0; } } /* len will be >= 0 after this point. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); rack_sndbuf_autoscale(rack); /* * Decide if we can use TCP Segmentation Offloading (if supported by * hardware). * * TSO may only be used if we are in a pure bulk sending state. The * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP * options prevent using TSO. With TSO the TCP header is the same * (except for the sequence number) for all generated packets. This * makes it impossible to transmit any options which vary per * generated segment or packet. * * IPv4 handling has a clear separation of ip options and ip header * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does * the right thing below to provide length of just ip options and thus * checking for ipoptlen is enough to decide if ip options are present. */ ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* * Pre-calculate here as we save another lookup into the darknesses * of IPsec that way and can actually decide if TSO is ok. */ #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6)) ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4)) ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); #endif /* INET */ #endif #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz && (tp->t_port == 0) && ((tp->t_flags & TF_SIGNATURE) == 0) && tp->rcv_numsacks == 0 && sack_rxmit == 0 && ipoptlen == 0) tso = 1; { uint32_t outstanding; outstanding = tp->snd_max - tp->snd_una; if (tp->t_flags & TF_SENTFIN) { /* * If we sent a fin, snd_max is 1 higher than * snd_una */ outstanding--; } if (sack_rxmit) { if ((rsm->r_flags & RACK_HAS_FIN) == 0) flags &= ~TH_FIN; } else { if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + sbused(sb))) flags &= ~TH_FIN; } } recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); /* * Sender silly window avoidance. We transmit under the following * conditions when len is non-zero: * * - We have a full segment (or more with TSO) - This is the last * buffer in a write()/send() and we are either idle or running * NODELAY - we've timed out (e.g. persist timer) - we have more * then 1/2 the maximum send window's worth of data (receiver may be * limited the window size) - we need to retransmit */ if (len) { if (len >= segsiz) { goto send; } /* * NOTE! on localhost connections an 'ack' from the remote * end may occur synchronously with the output and cause us * to flush a buffer queued with moretocome. XXX * */ if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ (idle || (tp->t_flags & TF_NODELAY)) && ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) && (tp->t_flags & TF_NOPUSH) == 0) { pass = 2; goto send; } if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ pass = 22; goto send; } if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { pass = 4; goto send; } if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */ pass = 5; goto send; } if (sack_rxmit) { pass = 6; goto send; } if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) && (ctf_outstanding(tp) < (segsiz * 2))) { /* * We have less than two MSS outstanding (delayed ack) * and our rwnd will not let us send a full sized * MSS. Lets go ahead and let this small segment * out because we want to try to have at least two * packets inflight to not be caught by delayed ack. */ pass = 12; goto send; } } /* * Sending of standalone window updates. * * Window updates are important when we close our window due to a * full socket buffer and are opening it again after the application * reads data from it. Once the window has opened again and the * remote end starts to send again the ACK clock takes over and * provides the most current window information. * * We must avoid the silly window syndrome whereas every read from * the receive buffer, no matter how small, causes a window update * to be sent. We also should avoid sending a flurry of window * updates when the socket buffer had queued a lot of data and the * application is doing small reads. * * Prevent a flurry of pointless window updates by only sending an * update when we can increase the advertized window by more than * 1/4th of the socket buffer capacity. When the buffer is getting * full or is very small be more aggressive and send an update * whenever we can increase by two mss sized segments. In all other * situations the ACK's to new incoming data will carry further * window increases. * * Don't send an independent window update if a delayed ACK is * pending (it will get piggy-backed on it) or the remote side * already has done a half-close and won't send more data. Skip * this if the connection is in T/TCP half-open state. */ if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && !(tp->t_flags & TF_DELACK) && !TCPS_HAVERCVDFIN(tp->t_state)) { /* * "adv" is the amount we could increase the window, taking * into account that we are limited by TCP_MAXWIN << * tp->rcv_scale. */ int32_t adv; int oldwin; adv = recwin; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { oldwin = (tp->rcv_adv - tp->rcv_nxt); if (adv > oldwin) adv -= oldwin; else { /* We can't increase the window */ adv = 0; } } else oldwin = 0; /* * If the new window size ends up being the same as or less * than the old size when it is scaled, then don't force * a window update. */ if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) goto dontupdate; if (adv >= (int32_t)(2 * segsiz) && (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) || so->so_rcv.sb_hiwat <= 8 * segsiz)) { pass = 7; goto send; } if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) { pass = 23; goto send; } } dontupdate: /* * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW * is also a catch-all for the retransmit timer timeout case. */ if (tp->t_flags & TF_ACKNOW) { pass = 8; goto send; } if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) { pass = 9; goto send; } /* * If our state indicates that FIN should be sent and we have not * yet done so, then we need to send. */ if ((flags & TH_FIN) && (tp->snd_nxt == tp->snd_una)) { pass = 11; goto send; } /* * No reason to send a segment, just return. */ just_return: SOCKBUF_UNLOCK(sb); just_return_nolock: { int app_limited = CTF_JR_SENT_DATA; if (tot_len_this_send > 0) { /* Make sure snd_nxt is up to max */ rack->r_ctl.fsb.recwin = recwin; slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz); if ((error == 0) && rack_use_rfo && ((flags & (TH_SYN|TH_FIN)) == 0) && (ipoptlen == 0) && (tp->snd_nxt == tp->snd_max) && (tp->rcv_numsacks == 0) && rack->r_fsb_inited && TCPS_HAVEESTABLISHED(tp->t_state) && (rack->r_must_retran == 0) && ((tp->t_flags & TF_NEEDFIN) == 0) && (len > 0) && (orig_len > 0) && (orig_len > len) && ((orig_len - len) >= segsiz) && ((optlen == 0) || ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) { /* We can send at least one more MSS using our fsb */ rack->r_fast_output = 1; rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; rack->r_ctl.fsb.tcp_flags = flags; rack->r_ctl.fsb.left_to_send = orig_len - len; if (hw_tls) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(sb), (tp->snd_max - tp->snd_una))); if (rack->r_ctl.fsb.left_to_send < segsiz) rack->r_fast_output = 0; else { if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una))) rack->r_ctl.fsb.rfo_apply_push = 1; else rack->r_ctl.fsb.rfo_apply_push = 0; } } else rack->r_fast_output = 0; rack_log_fsb(rack, tp, so, flags, ipoptlen, orig_len, len, 0, 1, optlen, __LINE__, 1); if (SEQ_GT(tp->snd_max, tp->snd_nxt)) tp->snd_nxt = tp->snd_max; } else { int end_window = 0; uint32_t seq = tp->gput_ack; rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm) { /* * Mark the last sent that we just-returned (hinting * that delayed ack may play a role in any rtt measurement). */ rsm->r_just_ret = 1; } counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1); rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_late = 0; rack->r_ctl.rc_agg_early = 0; if ((ctf_outstanding(tp) + min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)), minseg)) >= tp->snd_wnd) { /* We are limited by the rwnd */ app_limited = CTF_JR_RWND_LIMITED; if (IN_FASTRECOVERY(tp->t_flags)) rack->r_ctl.rc_prr_sndcnt = 0; } else if (ctf_outstanding(tp) >= sbavail(sb)) { /* We are limited by whats available -- app limited */ app_limited = CTF_JR_APP_LIMITED; if (IN_FASTRECOVERY(tp->t_flags)) rack->r_ctl.rc_prr_sndcnt = 0; } else if ((idle == 0) && ((tp->t_flags & TF_NODELAY) == 0) && ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) && (len < segsiz)) { /* * No delay is not on and the * user is sending less than 1MSS. This * brings out SWS avoidance so we * don't send. Another app-limited case. */ app_limited = CTF_JR_APP_LIMITED; } else if (tp->t_flags & TF_NOPUSH) { /* * The user has requested no push of * the last segment and we are * at the last segment. Another app * limited case. */ app_limited = CTF_JR_APP_LIMITED; } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) { /* Its the cwnd */ app_limited = CTF_JR_CWND_LIMITED; } else if (IN_FASTRECOVERY(tp->t_flags) && (rack->rack_no_prr == 0) && (rack->r_ctl.rc_prr_sndcnt < segsiz)) { app_limited = CTF_JR_PRR; } else { /* Now why here are we not sending? */ #ifdef NOW #ifdef INVARIANTS panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use); #endif #endif app_limited = CTF_JR_ASSESSING; } /* * App limited in some fashion, for our pacing GP * measurements we don't want any gap (even cwnd). * Close down the measurement window. */ if (rack_cwnd_block_ends_measure && ((app_limited == CTF_JR_CWND_LIMITED) || (app_limited == CTF_JR_PRR))) { /* * The reason we are not sending is * the cwnd (or prr). We have been configured * to end the measurement window in * this case. */ end_window = 1; } else if (rack_rwnd_block_ends_measure && (app_limited == CTF_JR_RWND_LIMITED)) { /* * We are rwnd limited and have been * configured to end the measurement * window in this case. */ end_window = 1; } else if (app_limited == CTF_JR_APP_LIMITED) { /* * A true application limited period, we have * ran out of data. */ end_window = 1; } else if (app_limited == CTF_JR_ASSESSING) { /* * In the assessing case we hit the end of * the if/else and had no known reason * This will panic us under invariants.. * * If we get this out in logs we need to * investagate which reason we missed. */ end_window = 1; } if (end_window) { uint8_t log = 0; /* Adjust the Gput measurement */ if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GT(tp->gput_ack, tp->snd_max)) { tp->gput_ack = tp->snd_max; if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) { /* * There is not enough to measure. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); } else log = 1; } /* Mark the last packet has app limited */ rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) { if (rack->r_ctl.rc_app_limited_cnt == 0) rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm; else { /* * Go out to the end app limited and mark * this new one as next and move the end_appl up * to this guy. */ if (rack->r_ctl.rc_end_appl) rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start; rack->r_ctl.rc_end_appl = rsm; } rsm->r_flags |= RACK_APP_LIMITED; rack->r_ctl.rc_app_limited_cnt++; } if (log) rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_app_limited_cnt, seq, tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0); } } if (slot) { /* set the rack tcb into the slot N */ counter_u64_add(rack_paced_segments, 1); } else if (tot_len_this_send) { counter_u64_add(rack_unpaced_segments, 1); } /* Check if we need to go into persists or not */ if ((tp->snd_max == tp->snd_una) && TCPS_HAVEESTABLISHED(tp->t_state) && sbavail(sb) && (sbavail(sb) > tp->snd_wnd) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) { /* Yes lets make sure to move to persist before timer-start */ rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack); rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use); } #ifdef NETFLIX_SHARED_CWND if ((sbavail(sb) == 0) && rack->r_ctl.rc_scw) { tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 1; } #endif #ifdef TCP_ACCOUNTING if (tot_len_this_send > 0) { crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz)); } else { crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_LIMITED]++; } counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val)); } sched_unpin(); #endif return (0); send: if (rsm || sack_rxmit) counter_u64_add(rack_nfto_resend, 1); else counter_u64_add(rack_non_fto_send, 1); if ((flags & TH_FIN) && sbavail(sb)) { /* * We do not transmit a FIN * with data outstanding. We * need to make it so all data * is acked first. */ flags &= ~TH_FIN; } /* Enforce stack imposed max seg size if we have one */ if (rack->r_ctl.rc_pace_max_segs && (len > rack->r_ctl.rc_pace_max_segs)) { mark = 1; len = rack->r_ctl.rc_pace_max_segs; } SOCKBUF_LOCK_ASSERT(sb); if (len > 0) { if (len >= segsiz) tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; else tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; } /* * Before ESTABLISHED, force sending of initial options unless TCP * set not to do any options. NOTE: we assume that the IP/TCP header * plus TCP options always fit in a single mbuf, leaving room for a * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) * + optlen <= MCLBYTES */ optlen = 0; #ifdef INET6 if (isipv6) hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else #endif hdrlen = sizeof(struct tcpiphdr); /* * Compute options for segment. We only have to care about SYN and * established connection segments. Options for SYN-ACK segments * are handled in TCP syncache. */ to.to_flags = 0; if ((tp->t_flags & TF_NOOPT) == 0) { /* Maximum segment size. */ if (flags & TH_SYN) { tp->snd_nxt = tp->iss; to.to_mss = tcp_mssopt(&inp->inp_inc); if (tp->t_port) to.to_mss -= V_tcp_udp_tunneling_overhead; to.to_flags |= TOF_MSS; /* * On SYN or SYN|ACK transmits on TFO connections, * only include the TFO option if it is not a * retransmit, as the presence of the TFO option may * have caused the original SYN or SYN|ACK to have * been dropped by a middlebox. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_rxtshift == 0)) { if (tp->t_state == TCPS_SYN_RECEIVED) { to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; to.to_tfo_cookie = (u_int8_t *)&tp->t_tfo_cookie.server; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } else if (tp->t_state == TCPS_SYN_SENT) { to.to_tfo_len = tp->t_tfo_client_cookie_len; to.to_tfo_cookie = tp->t_tfo_cookie.client; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; /* * If we wind up having more data to * send with the SYN than can fit in * one segment, don't send any more * until the SYN|ACK comes back from * the other end. */ sendalot = 0; } } } /* Window scaling. */ if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { to.to_wscale = tp->request_r_scale; to.to_flags |= TOF_SCALE; } /* Timestamps. */ if ((tp->t_flags & TF_RCVD_TSTMP) || ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { to.to_tsval = ms_cts + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; } /* Set receive buffer autosizing timestamp. */ if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) tp->rfbuf_ts = tcp_ts_getticks(); /* Selective ACK's. */ if (tp->t_flags & TF_SACK_PERMIT) { if (flags & TH_SYN) to.to_flags |= TOF_SACKPERM; else if (TCPS_HAVEESTABLISHED(tp->t_state) && tp->rcv_numsacks > 0) { to.to_flags |= TOF_SACK; to.to_nsacks = tp->rcv_numsacks; to.to_sacks = (u_char *)tp->sackblks; } } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* TCP-MD5 (RFC2385). */ if (tp->t_flags & TF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif /* TCP_SIGNATURE */ /* Processing the options. */ hdrlen += optlen = tcp_addoptions(&to, opt); /* * If we wanted a TFO option to be added, but it was unable * to fit, ensure no data is sent. */ if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && !(to.to_flags & TOF_FASTOPEN)) len = 0; } if (tp->t_port) { if (V_tcp_udp_tunneling_port == 0) { /* The port was removed?? */ SOCKBUF_UNLOCK(&so->so_snd); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (EHOSTUNREACH); } hdrlen += sizeof(struct udphdr); } #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(tp->t_inpcb); else #endif if (tp->t_inpcb->inp_options) ipoptlen = tp->t_inpcb->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif /* * Adjust data length if insertion of options will bump the packet * length beyond the t_maxseg length. Clear the FIN bit because we * cut off the tail of the segment. */ if (len + optlen + ipoptlen > tp->t_maxseg) { if (tso) { uint32_t if_hw_tsomax; uint32_t moff; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; KASSERT(ipoptlen == 0, ("%s: TSO can't do IP options", __func__)); /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { len = 0; } else if (len > max_len) { sendalot = 1; len = max_len; mark = 2; } } /* * Prevent the last segment from being fractional * unless the send sockbuf can be emptied: */ max_len = (tp->t_maxseg - optlen); if ((sb_offset + len) < sbavail(sb)) { moff = len % (u_int)max_len; if (moff != 0) { mark = 3; len -= moff; } } /* * In case there are too many small fragments don't * use TSO: */ if (len <= segsiz) { mark = 4; tso = 0; } /* * Send the FIN in a separate segment after the bulk * sending is done. We don't trust the TSO * implementations to clear the FIN flag on all but * the last segment. */ if (tp->t_flags & TF_NEEDFIN) { sendalot = 4; } } else { mark = 5; if (optlen + ipoptlen >= tp->t_maxseg) { /* * Since we don't have enough space to put * the IP header chain and the TCP header in * one packet as required by RFC 7112, don't * send it. Also ensure that at least one * byte of the payload can be put into the * TCP segment. */ SOCKBUF_UNLOCK(&so->so_snd); error = EMSGSIZE; sack_rxmit = 0; goto out; } len = tp->t_maxseg - optlen - ipoptlen; sendalot = 5; } } else { tso = 0; mark = 6; } KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, ("%s: len > IP_MAXPACKET", __func__)); #ifdef DIAGNOSTIC #ifdef INET6 if (max_linkhdr + hdrlen > MCLBYTES) #else if (max_linkhdr + hdrlen > MHLEN) #endif panic("tcphdr too big"); #endif /* * This KASSERT is here to catch edge cases at a well defined place. * Before, those had triggered (random) panic conditions further * down. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); if ((len == 0) && (flags & TH_FIN) && (sbused(sb))) { /* * We have outstanding data, don't send a fin by itself!. */ goto just_return; } /* * Grab a header mbuf, attaching a copy of data to be transmitted, * and initialize the header from the template for sends on this * connection. */ hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0; if (len) { uint32_t max_val; uint32_t moff; if (rack->r_ctl.rc_pace_max_segs) max_val = rack->r_ctl.rc_pace_max_segs; else if (rack->rc_user_set_max_segs) max_val = rack->rc_user_set_max_segs * segsiz; else max_val = len; /* * We allow a limit on sending with hptsi. */ if (len > max_val) { mark = 7; len = max_val; } #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { SOCKBUF_UNLOCK(sb); error = ENOBUFS; sack_rxmit = 0; goto out; } m->m_data += max_linkhdr; m->m_len = hdrlen; /* * Start the m_copy functions from the closest mbuf to the * sb_offset in the socket buffer chain. */ mb = sbsndptr_noadv(sb, sb_offset, &moff); s_mb = mb; s_moff = moff; if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { m_copydata(mb, moff, (int)len, mtod(m, caddr_t)+hdrlen); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) sbsndptr_adv(sb, mb, len); m->m_len += len; } else { struct sockbuf *msb; if (SEQ_LT(tp->snd_nxt, tp->snd_max)) msb = NULL; else msb = sb; m->m_next = tcp_m_copym( mb, moff, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb, ((rsm == NULL) ? hw_tls : 0) #ifdef NETFLIX_COPY_ARGS , &filled_all #endif ); if (len <= (tp->t_maxseg - optlen)) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if (m->m_next == NULL) { SOCKBUF_UNLOCK(sb); (void)m_free(m); error = ENOBUFS; sack_rxmit = 0; goto out; } } if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { if (rsm && (rsm->r_flags & RACK_TLP)) { /* * TLP should not count in retran count, but * in its own bin */ counter_u64_add(rack_tlp_retran, 1); counter_u64_add(rack_tlp_retran_bytes, len); } else { tp->t_sndrexmitpack++; KMOD_TCPSTAT_INC(tcps_sndrexmitpack); KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); } #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif } else { KMOD_TCPSTAT_INC(tcps_sndpack); KMOD_TCPSTAT_ADD(tcps_sndbyte, len); #ifdef STATS stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif } /* * If we're sending everything we've got, set PUSH. (This * will keep happy those implementations which only give * data to the user when a buffer fills or a PUSH comes in.) */ if (sb_offset + len == sbused(sb) && sbused(sb) && !(flags & TH_SYN)) { flags |= TH_PUSH; add_flag |= RACK_HAD_PUSH; } SOCKBUF_UNLOCK(sb); } else { SOCKBUF_UNLOCK(sb); if (tp->t_flags & TF_ACKNOW) KMOD_TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN | TH_FIN | TH_RST)) KMOD_TCPSTAT_INC(tcps_sndctrl); else KMOD_TCPSTAT_INC(tcps_sndwinup); m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { error = ENOBUFS; sack_rxmit = 0; goto out; } #ifdef INET6 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && MHLEN >= hdrlen) { M_ALIGN(m, hdrlen); } else #endif m->m_data += max_linkhdr; m->m_len = hdrlen; } SOCKBUF_UNLOCK_ASSERT(sb); m->m_pkthdr.rcvif = (struct ifnet *)0; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) { #ifdef INET6 if (isipv6) ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; else #endif /* INET6 */ ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; th = rack->r_ctl.fsb.th; udp = rack->r_ctl.fsb.udp; if (udp) { #ifdef INET6 if (isipv6) ulen = hdrlen + len - sizeof(struct ip6_hdr); else #endif /* INET6 */ ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); } } else { #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip6_hdr); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else th = (struct tcphdr *)(ip6 + 1); tcpip_fillheaders(inp, tp->t_port, ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); #ifdef TCPDEBUG ipov = (struct ipovly *)ip; #endif if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else th = (struct tcphdr *)(ip + 1); tcpip_fillheaders(inp, tp->t_port, ip, th); } } /* * Fill in fields, remembering maximum advertised window for use in * delaying messages about window sizes. If resending a FIN, be sure * not to use a new sequence number. */ if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && tp->snd_nxt == tp->snd_max) tp->snd_nxt--; /* * If we are starting a connection, send ECN setup SYN packet. If we * are on a retransmit, we may resend those bits a number of times * as per RFC 3168. */ if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) { if (tp->t_rxtshift >= 1) { if (tp->t_rxtshift <= V_tcp_ecn_maxretries) flags |= TH_ECE | TH_CWR; } else flags |= TH_ECE | TH_CWR; } /* Handle parallel SYN for ECN */ if ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_flags2 & TF2_ECN_SND_ECE)) { flags |= TH_ECE; tp->t_flags2 &= ~TF2_ECN_SND_ECE; } if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags2 & TF2_ECN_PERMIT)) { /* * If the peer has ECN, mark data packets with ECN capable * transmission (ECT). Ignore pure ack packets, * retransmissions. */ if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && (sack_rxmit == 0)) { #ifdef INET6 if (isipv6) ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); else #endif ip->ip_tos |= IPTOS_ECN_ECT0; KMOD_TCPSTAT_INC(tcps_ecn_ect0); /* * Reply with proper ECN notifications. * Only set CWR on new data segments. */ if (tp->t_flags2 & TF2_ECN_SND_CWR) { flags |= TH_CWR; tp->t_flags2 &= ~TF2_ECN_SND_CWR; } } if (tp->t_flags2 & TF2_ECN_SND_ECE) flags |= TH_ECE; } /* * If we are doing retransmissions, then snd_nxt will not reflect * the first unsent octet. For ACK only packets, we do not want the * sequence number of the retransmitted packet, we want the sequence * number of the next unsent octet. So, if there is no data (and no * SYN or FIN), use snd_max instead of snd_nxt when filling in * ti_seq. But if we are in persist state, snd_max might reflect * one byte beyond the right edge of the window, so use snd_nxt in * that case, since we know we aren't doing a retransmission. * (retransmit and persist are mutually exclusive...) */ if (sack_rxmit == 0) { if (len || (flags & (TH_SYN | TH_FIN))) { th->th_seq = htonl(tp->snd_nxt); rack_seq = tp->snd_nxt; } else { th->th_seq = htonl(tp->snd_max); rack_seq = tp->snd_max; } } else { th->th_seq = htonl(rsm->r_start); rack_seq = rsm->r_start; } th->th_ack = htonl(tp->rcv_nxt); th->th_flags = flags; /* * Calculate receive window. Don't shrink window, but avoid silly * window syndrome. * If a RST segment is sent, advertise a window of zero. */ if (flags & TH_RST) { recwin = 0; } else { if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && recwin < (long)segsiz) { recwin = 0; } if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) recwin = (long)(tp->rcv_adv - tp->rcv_nxt); } /* * According to RFC1323 the window field in a SYN (i.e., a or * ) segment itself is never scaled. The case is * handled in syncache. */ if (flags & TH_SYN) th->th_win = htons((u_short) (min(sbspace(&so->so_rcv), TCP_MAXWIN))); else { /* Avoid shrinking window with window scaling. */ recwin = roundup2(recwin, 1 << tp->rcv_scale); th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); } /* * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 * window. This may cause the remote transmitter to stall. This * flag tells soreceive() to disable delayed acknowledgements when * draining the buffer. This can occur if the receiver is * attempting to read more data than can be buffered prior to * transmitting on the connection. */ if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; tp->snd_up = tp->snd_una; /* drag it along, its deprecated */ /* Now are we using fsb?, if so copy the template data to the mbuf */ if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) { uint8_t *cpto; cpto = mtod(m, uint8_t *); memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len); /* * We have just copied in: * IP/IP6 * * tcphdr (no options) * * We need to grab the correct pointers into the mbuf * for both the tcp header, and possibly the udp header (if tunneling). * We do this by using the offset in the copy buffer and adding it * to the mbuf base pointer (cpto). */ #ifdef INET6 if (isipv6) ip6 = mtod(m, struct ip6_hdr *); else #endif /* INET6 */ ip = mtod(m, struct ip *); th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr)); /* If we have a udp header lets set it into the mbuf as well */ if (udp) udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr)); } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (to.to_flags & TOF_SIGNATURE) { /* * Calculate MD5 signature and put it into the place * determined before. * NOTE: since TCP options buffer doesn't point into * mbuf's data, calculate offset and use it. */ if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { /* * Do not send segment if the calculation of MD5 * digest has failed. */ goto out; } } #endif if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } /* * Put TCP length in extended header, and then checksum extended * header and data. */ m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (isipv6) { /* * ip6_plen is not need to be filled now, and will be filled * in ip6_output. */ if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif /* * Enable TSO and specify the size of the segments. The TCP pseudo * header checksum is always provided. XXX: Fixme: This is currently * not the case for IPv6. */ if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz", __func__)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } KASSERT(len + hdrlen == m_length(m, NULL), ("%s: mbuf chain different than expected: %d + %u != %u", __func__, len, hdrlen, m_length(m, NULL))); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ hhook_run_tcp_est_out(tp, th, &to, len, tso); #endif /* We're getting ready to send; log now. */ if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; if (rack->rack_no_prr) log.u_bbr.flex1 = 0; else log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = orig_len; if (filled_all) log.u_bbr.flex5 = 0x80000000; else log.u_bbr.flex5 = 0; /* Save off the early/late values */ log.u_bbr.flex6 = rack->r_ctl.rc_agg_early; log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed; log.u_bbr.bw_inuse = rack_get_bw(rack); if (rsm || sack_rxmit) { if (doing_tlp) log.u_bbr.flex8 = 2; else log.u_bbr.flex8 = 1; } else { log.u_bbr.flex8 = 0; } log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm); log.u_bbr.flex7 = mark; log.u_bbr.flex7 <<= 8; log.u_bbr.flex7 |= pass; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.lt_epoch = cwnd_to_use; log.u_bbr.delivered = sendalot; lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, &tv); } else lgb = NULL; /* * Fill in IP length and desired time to live and send to IP level. * There should be a better way to handle ttl and tos; we could keep * them in the template, but need a way to checksum without them. */ /* * m->m_pkthdr.len should have been set before cksum calcuration, * because in6_cksum() need it. */ #ifdef INET6 if (isipv6) { /* * we separately set hoplimit for every segment, since the * user might want to change the value via setsockopt. Also, * desired default hop limit might be changed via Neighbor * Discovery. */ rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL); /* * Set the packet size here for the benefit of DTrace * probes. ip6_output() will set it properly; it's supposed * to include the option header lengths as well. */ ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); TCP_PROBE5(send, NULL, tp, ip6, tp, th); /* TODO: IPv6 IP6TOS_ECT bit on */ error = ip6_output(m, #if defined(IPSEC) || defined(IPSEC_SUPPORT) inp->in6p_outputopts, #else NULL, #endif &inp->inp_route6, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), NULL, NULL, inp); if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) mtu = inp->inp_route6.ro_nh->nh_mtu; } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) ip->ip_ttl = in6_selecthlim(inp, NULL); #endif /* INET6 */ rack->r_ctl.fsb.hoplimit = ip->ip_ttl; /* * If we do path MTU discovery, then we set DF on every * packet. This might not be the best thing to do according * to RFC3390 Section 2. However the tcp hostcache migitates * the problem so it affects only the first tcp connection * with a host. * * NB: Don't set DF on small MTU/MSS to have a safe * fallback. */ if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); TCP_PROBE5(send, NULL, tp, ip, tp, th); error = ip_output(m, #if defined(IPSEC) || defined(IPSEC_SUPPORT) inp->inp_options, #else NULL, #endif &inp->inp_route, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, inp); if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) mtu = inp->inp_route.ro_nh->nh_mtu; } #endif /* INET */ out: if (lgb) { lgb->tlb_errno = error; lgb = NULL; } /* * In transmit state, time the transmission and arrange for the * retransmit. In persist state, just set snd_max. */ if (error == 0) { tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); rack->forced_ack = 0; /* If we send something zap the FA flag */ if (rsm && (doing_tlp == 0)) { /* Set we retransmitted */ rack->rc_gp_saw_rec = 1; } else { if (cwnd_to_use > tp->snd_ssthresh) { /* Set we sent in CA */ rack->rc_gp_saw_ca = 1; } else { /* Set we sent in SS */ rack->rc_gp_saw_ss = 1; } } if (doing_tlp && (rsm == NULL)) { /* Make sure new data TLP cnt is clear */ rack->r_ctl.rc_tlp_new_data = 0; } if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0) tcp_clean_dsack_blocks(tp); tot_len_this_send += len; if (len == 0) counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1); else if (len == 1) { counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1); } else if (len > 1) { int idx; idx = (len / segsiz) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); else counter_u64_add(rack_out_size[idx], 1); } } if ((rack->rack_no_prr == 0) && sub_from_prr && (error == 0)) { if (rack->r_ctl.rc_prr_sndcnt >= len) rack->r_ctl.rc_prr_sndcnt -= len; else rack->r_ctl.rc_prr_sndcnt = 0; } sub_from_prr = 0; if (doing_tlp && (rsm == NULL)) { /* New send doing a TLP */ add_flag |= RACK_TLP; } rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, rack_to_usec_ts(&tv), rsm, add_flag, s_mb, s_moff, hw_tls); if ((error == 0) && (len > 0) && (tp->snd_una == tp->snd_max)) rack->r_ctl.rc_tlp_rxt_last_time = cts; { tcp_seq startseq = tp->snd_nxt; /* Track our lost count */ if (rsm && (doing_tlp == 0)) rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start; /* * Advance snd_nxt over sequence space of this segment. */ if (error) /* We don't log or do anything with errors */ goto nomore; if (doing_tlp == 0) { if (rsm == NULL) { /* * Not a retransmission of some * sort, new data is going out so * clear our TLP count and flag. */ rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; } } else { /* * We have just sent a TLP, mark that it is true * and make sure our in progress is set so we * continue to check the count. */ rack->rc_tlp_in_progress = 1; rack->r_ctl.rc_tlp_cnt_out++; } if (flags & (TH_SYN | TH_FIN)) { if (flags & TH_SYN) tp->snd_nxt++; if (flags & TH_FIN) { tp->snd_nxt++; tp->t_flags |= TF_SENTFIN; } } /* In the ENOBUFS case we do *not* update snd_max */ if (sack_rxmit) goto nomore; tp->snd_nxt += len; if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { if (tp->snd_una == tp->snd_max) { /* * Update the time we just added data since * none was outstanding. */ rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); tp->t_acktime = ticks; } tp->snd_max = tp->snd_nxt; /* * Time this transmission if not a retransmission and * not currently timing anything. * This is only relevant in case of switching back to * the base stack. */ if (tp->t_rtttime == 0) { tp->t_rtttime = ticks; tp->t_rtseq = startseq; KMOD_TCPSTAT_INC(tcps_segstimed); } if (len && ((tp->t_flags & TF_GPUTINPROG) == 0)) rack_start_gp_measurement(tp, rack, startseq, sb_offset); } /* * If we are doing FO we need to update the mbuf position and subtract * this happens when the peer sends us duplicate information and * we thus want to send a DSACK. * * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO * turned off? If not then we are going to echo multiple DSACK blocks * out (with the TSO), which we should not be doing. */ if (rack->r_fast_output && len) { if (rack->r_ctl.fsb.left_to_send > len) rack->r_ctl.fsb.left_to_send -= len; else rack->r_ctl.fsb.left_to_send = 0; if (rack->r_ctl.fsb.left_to_send < segsiz) rack->r_fast_output = 0; if (rack->r_fast_output) { rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; } } } nomore: if (error) { rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_late = 0; rack->r_ctl.rc_agg_early = 0; SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */ /* * Failures do not advance the seq counter above. For the * case of ENOBUFS we will fall out and retry in 1ms with * the hpts. Everything else will just have to retransmit * with the timer. * * In any case, we do not want to loop around for another * send without a good reason. */ sendalot = 0; switch (error) { case EPERM: tp->t_softerror = error; #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (error); case ENOBUFS: /* * Pace us right away to retry in a some * time */ slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC); if (rack->rc_enobuf < 0x7f) rack->rc_enobuf++; if (slot < (10 * HPTS_USEC_IN_MSEC)) slot = 10 * HPTS_USEC_IN_MSEC; if (rack->r_ctl.crte != NULL) { counter_u64_add(rack_saw_enobuf_hw, 1); tcp_rl_log_enobuf(rack->r_ctl.crte); } counter_u64_add(rack_saw_enobuf, 1); goto enobufs; case EMSGSIZE: /* * For some reason the interface we used initially * to send segments changed to another or lowered * its MTU. If TSO was active we either got an * interface without TSO capabilits or TSO was * turned off. If we obtained mtu from ip_output() * then update it and try again. */ if (tso) tp->t_flags &= ~TF_TSO; if (mtu != 0) { tcp_mss_update(tp, -1, mtu, NULL, NULL); goto again; } slot = 10 * HPTS_USEC_IN_MSEC; rack_start_hpts_timer(rack, tp, cts, slot, 0, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (error); case ENETUNREACH: counter_u64_add(rack_saw_enetunreach, 1); case EHOSTDOWN: case EHOSTUNREACH: case ENETDOWN: if (TCPS_HAVERCVDSYN(tp->t_state)) { tp->t_softerror = error; } /* FALLTHROUGH */ default: slot = 10 * HPTS_USEC_IN_MSEC; rack_start_hpts_timer(rack, tp, cts, slot, 0, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (error); } } else { rack->rc_enobuf = 0; if (IN_FASTRECOVERY(tp->t_flags) && rsm) rack->r_ctl.retran_during_recovery += len; } KMOD_TCPSTAT_INC(tcps_sndtotal); /* * Data sent (as far as we can tell). If this advertises a larger * window than any other segment, then remember the size of the * advertised window. Any pending ACK has now been sent. */ if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) tp->rcv_adv = tp->rcv_nxt + recwin; tp->last_ack_sent = tp->rcv_nxt; tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); enobufs: if (sendalot) { /* Do we need to turn off sendalot? */ if (rack->r_ctl.rc_pace_max_segs && (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) { /* We hit our max. */ sendalot = 0; } else if ((rack->rc_user_set_max_segs) && (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) { /* We hit the user defined max */ sendalot = 0; } } if ((error == 0) && (flags & TH_FIN)) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); if (flags & TH_RST) { /* * We don't send again after sending a RST. */ slot = 0; sendalot = 0; if (error == 0) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) { /* * Get our pacing rate, if an error * occurred in sending (ENOBUF) we would * hit the else if with slot preset. Other * errors return. */ slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz); } if (rsm && (rsm->r_flags & RACK_HAS_SYN) == 0 && rack->use_rack_rr) { /* Its a retransmit and we use the rack cheat? */ if ((slot == 0) || (rack->rc_always_pace == 0) || (rack->r_rr_config == 1)) { /* * We have no pacing set or we * are using old-style rack or * we are overriden to use the old 1ms pacing. */ slot = rack->r_ctl.rc_min_to; } } /* We have sent clear the flag */ rack->r_ent_rec_ns = 0; if (rack->r_must_retran) { if (rsm) { rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start); if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) { /* * We have retransmitted all. */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; } } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) { /* * Sending new data will also kill * the loop. */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; } } rack->r_ctl.fsb.recwin = recwin; if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) && SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) { /* * We hit an RTO and now have past snd_max at the RTO * clear all the WAS flags. */ tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY); } if (slot) { /* set the rack tcb into the slot N */ counter_u64_add(rack_paced_segments, 1); if ((error == 0) && rack_use_rfo && ((flags & (TH_SYN|TH_FIN)) == 0) && (rsm == NULL) && (tp->snd_nxt == tp->snd_max) && (ipoptlen == 0) && (tp->rcv_numsacks == 0) && rack->r_fsb_inited && TCPS_HAVEESTABLISHED(tp->t_state) && (rack->r_must_retran == 0) && ((tp->t_flags & TF_NEEDFIN) == 0) && (len > 0) && (orig_len > 0) && (orig_len > len) && ((orig_len - len) >= segsiz) && ((optlen == 0) || ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) { /* We can send at least one more MSS using our fsb */ rack->r_fast_output = 1; rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; rack->r_ctl.fsb.tcp_flags = flags; rack->r_ctl.fsb.left_to_send = orig_len - len; if (hw_tls) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(sb), (tp->snd_max - tp->snd_una))); if (rack->r_ctl.fsb.left_to_send < segsiz) rack->r_fast_output = 0; else { if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una))) rack->r_ctl.fsb.rfo_apply_push = 1; else rack->r_ctl.fsb.rfo_apply_push = 0; } } else rack->r_fast_output = 0; rack_log_fsb(rack, tp, so, flags, ipoptlen, orig_len, len, error, (rsm == NULL), optlen, __LINE__, 2); } else if (sendalot) { int ret; if (len) counter_u64_add(rack_unpaced_segments, 1); sack_rxmit = 0; if ((error == 0) && rack_use_rfo && ((flags & (TH_SYN|TH_FIN)) == 0) && (rsm == NULL) && (ipoptlen == 0) && (tp->rcv_numsacks == 0) && (tp->snd_nxt == tp->snd_max) && (rack->r_must_retran == 0) && rack->r_fsb_inited && TCPS_HAVEESTABLISHED(tp->t_state) && ((tp->t_flags & TF_NEEDFIN) == 0) && (len > 0) && (orig_len > 0) && (orig_len > len) && ((orig_len - len) >= segsiz) && ((optlen == 0) || ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) { /* we can use fast_output for more */ rack->r_fast_output = 1; rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; rack->r_ctl.fsb.tcp_flags = flags; rack->r_ctl.fsb.left_to_send = orig_len - len; if (hw_tls) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(sb), (tp->snd_max - tp->snd_una))); if (rack->r_ctl.fsb.left_to_send < segsiz) { rack->r_fast_output = 0; } if (rack->r_fast_output) { if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una))) rack->r_ctl.fsb.rfo_apply_push = 1; else rack->r_ctl.fsb.rfo_apply_push = 0; rack_log_fsb(rack, tp, so, flags, ipoptlen, orig_len, len, error, (rsm == NULL), optlen, __LINE__, 3); error = 0; ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error); if (ret >= 0) return (ret); else if (error) goto nomore; } } goto again; } else if (len) { counter_u64_add(rack_unpaced_segments, 1); } /* Assure when we leave that snd_nxt will point to top */ if (SEQ_GT(tp->snd_max, tp->snd_nxt)) tp->snd_nxt = tp->snd_max; rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount() - ts_val; if (tot_len_this_send) { if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += crtsc; } counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz)); } else { if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_ACK]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_ACK] += crtsc; } counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc); } sched_unpin(); #endif if (error == ENOBUFS) error = 0; return (error); } static void rack_update_seg(struct tcp_rack *rack) { uint32_t orig_val; orig_val = rack->r_ctl.rc_pace_max_segs; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); if (orig_val != rack->r_ctl.rc_pace_max_segs) rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0); } static void rack_mtu_change(struct tcpcb *tp) { /* * The MSS may have changed */ struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) { /* * The MTU has changed we need to resend everything * since all we have sent is lost. We first fix * up the mtu though. */ rack_set_pace_segments(tp, rack, __LINE__, NULL); /* We treat this like a full retransmit timeout without the cwnd adjustment */ rack_remxt_tmr(tp); rack->r_fast_output = 0; rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp, rack->r_ctl.rc_sacked); rack->r_ctl.rc_snd_max_at_rto = tp->snd_max; rack->r_must_retran = 1; } sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); /* We don't use snd_nxt to retransmit */ tp->snd_nxt = tp->snd_max; } static int rack_set_profile(struct tcp_rack *rack, int prof) { int err = EINVAL; if (prof == 1) { /* pace_always=1 */ if (rack->rc_always_pace == 0) { if (tcp_can_enable_pacing() == 0) return (EBUSY); } rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; rack->rack_attempt_hdwr_pace = 0; /* cmpack=1 */ if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) && rack->r_use_cmp_ack) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; /* scwnd=1 */ rack->rack_enable_scwnd = 1; /* dynamic=100 */ rack->rc_gp_dyn_mul = 1; /* gp_inc_ca */ rack->r_ctl.rack_per_of_gp_ca = 100; /* rrr_conf=3 */ rack->r_rr_config = 3; /* npush=2 */ rack->r_ctl.rc_no_push_at_mrtt = 2; /* fillcw=1 */ rack->rc_pace_to_cwnd = 1; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; /* noprr=1 */ rack->rack_no_prr = 1; /* lscwnd=1 */ rack->r_limit_scw = 1; /* gp_inc_rec */ rack->r_ctl.rack_per_of_gp_rec = 90; err = 0; } else if (prof == 3) { /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */ /* pace_always=1 */ if (rack->rc_always_pace == 0) { if (tcp_can_enable_pacing() == 0) return (EBUSY); } rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; rack->rack_attempt_hdwr_pace = 0; /* cmpack=1 */ if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) && rack->r_use_cmp_ack) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; /* scwnd=1 */ rack->rack_enable_scwnd = 1; /* dynamic=100 */ rack->rc_gp_dyn_mul = 1; /* gp_inc_ca */ rack->r_ctl.rack_per_of_gp_ca = 100; /* rrr_conf=3 */ rack->r_rr_config = 3; /* npush=2 */ rack->r_ctl.rc_no_push_at_mrtt = 2; /* fillcw=2 */ rack->rc_pace_to_cwnd = 1; rack->r_fill_less_agg = 1; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; /* noprr=1 */ rack->rack_no_prr = 1; /* lscwnd=1 */ rack->r_limit_scw = 1; /* gp_inc_rec */ rack->r_ctl.rack_per_of_gp_rec = 90; err = 0; } else if (prof == 2) { /* cmpack=1 */ if (rack->rc_always_pace == 0) { if (tcp_can_enable_pacing() == 0) return (EBUSY); } rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); rack->r_use_cmp_ack = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; /* pace_always=1 */ rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; /* scwnd=1 */ rack->rack_enable_scwnd = 1; /* dynamic=100 */ rack->rc_gp_dyn_mul = 1; rack->r_ctl.rack_per_of_gp_ca = 100; /* rrr_conf=3 */ rack->r_rr_config = 3; /* npush=2 */ rack->r_ctl.rc_no_push_at_mrtt = 2; /* fillcw=1 */ rack->rc_pace_to_cwnd = 1; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; /* noprr=1 */ rack->rack_no_prr = 1; /* lscwnd=0 */ rack->r_limit_scw = 0; err = 0; } else if (prof == 0) { /* This changes things back to the default settings */ err = 0; if (rack->rc_always_pace) { tcp_decrement_paced_conn(); rack_undo_cc_pacing(rack); rack->rc_always_pace = 0; } if (rack_pace_every_seg && tcp_can_enable_pacing()) { rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); } else rack->rc_always_pace = 0; if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; else rack->r_use_cmp_ack = 0; if (rack_disable_prr) rack->rack_no_prr = 1; else rack->rack_no_prr = 0; if (rack_gp_no_rec_chg) rack->rc_gp_no_rec_chg = 1; else rack->rc_gp_no_rec_chg = 0; if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) { rack->r_mbuf_queue = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; } else { rack->r_mbuf_queue = 0; rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; } if (rack_enable_shared_cwnd) rack->rack_enable_scwnd = 1; else rack->rack_enable_scwnd = 0; if (rack_do_dyn_mul) { /* When dynamic adjustment is on CA needs to start at 100% */ rack->rc_gp_dyn_mul = 1; if (rack_do_dyn_mul >= 100) rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul; } else { rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca; rack->rc_gp_dyn_mul = 0; } rack->r_rr_config = 0; rack->r_ctl.rc_no_push_at_mrtt = 0; rack->rc_pace_to_cwnd = 0; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; if (rack_enable_hw_pacing) rack->rack_hdw_pace_ena = 1; else rack->rack_hdw_pace_ena = 0; if (rack_disable_prr) rack->rack_no_prr = 1; else rack->rack_no_prr = 0; if (rack_limits_scwnd) rack->r_limit_scw = 1; else rack->r_limit_scw = 0; err = 0; } return (err); } static int rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval) { struct deferred_opt_list *dol; dol = malloc(sizeof(struct deferred_opt_list), M_TCPFSB, M_NOWAIT|M_ZERO); if (dol == NULL) { /* * No space yikes -- fail out.. */ return (0); } dol->optname = sopt_name; dol->optval = loptval; TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next); return (1); } static int rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name, uint32_t optval, uint64_t loptval) { struct epoch_tracker et; struct sockopt sopt; struct cc_newreno_opts opt; uint64_t val; int error = 0; uint16_t ca, ss; switch (sopt_name) { case TCP_RACK_PACING_BETA: RACK_OPTS_INC(tcp_rack_beta); if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { /* This only works for newreno. */ error = EINVAL; break; } if (rack->rc_pacing_cc_set) { /* * Set them into the real CC module * whats in the rack pcb is the old values * to be used on restoral/ */ sopt.sopt_dir = SOPT_SET; opt.name = CC_NEWRENO_BETA; opt.val = optval; if (CC_ALGO(tp)->ctl_output != NULL) error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); else { error = ENOENT; break; } } else { /* * Not pacing yet so set it into our local * rack pcb storage. */ rack->r_ctl.rc_saved_beta.beta = optval; } break; case TCP_RACK_TIMER_SLOP: RACK_OPTS_INC(tcp_rack_timer_slop); rack->r_ctl.timer_slop = optval; if (rack->rc_tp->t_srtt) { /* * If we have an SRTT lets update t_rxtcur * to have the new slop. */ RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); } break; case TCP_RACK_PACING_BETA_ECN: RACK_OPTS_INC(tcp_rack_beta_ecn); if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { /* This only works for newreno. */ error = EINVAL; break; } if (rack->rc_pacing_cc_set) { /* * Set them into the real CC module * whats in the rack pcb is the old values * to be used on restoral/ */ sopt.sopt_dir = SOPT_SET; opt.name = CC_NEWRENO_BETA_ECN; opt.val = optval; if (CC_ALGO(tp)->ctl_output != NULL) error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); else error = ENOENT; } else { /* * Not pacing yet so set it into our local * rack pcb storage. */ rack->r_ctl.rc_saved_beta.beta_ecn = optval; rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN; } break; case TCP_DEFER_OPTIONS: RACK_OPTS_INC(tcp_defer_opt); if (optval) { if (rack->gp_ready) { /* Too late */ error = EINVAL; break; } rack->defer_options = 1; } else rack->defer_options = 0; break; case TCP_RACK_MEASURE_CNT: RACK_OPTS_INC(tcp_rack_measure_cnt); if (optval && (optval <= 0xff)) { rack->r_ctl.req_measurements = optval; } else error = EINVAL; break; case TCP_REC_ABC_VAL: RACK_OPTS_INC(tcp_rec_abc_val); if (optval > 0) rack->r_use_labc_for_rec = 1; else rack->r_use_labc_for_rec = 0; break; case TCP_RACK_ABC_VAL: RACK_OPTS_INC(tcp_rack_abc_val); if ((optval > 0) && (optval < 255)) rack->rc_labc = optval; else error = EINVAL; break; case TCP_HDWR_UP_ONLY: RACK_OPTS_INC(tcp_pacing_up_only); if (optval) rack->r_up_only = 1; else rack->r_up_only = 0; break; case TCP_PACING_RATE_CAP: RACK_OPTS_INC(tcp_pacing_rate_cap); rack->r_ctl.bw_rate_cap = loptval; break; case TCP_RACK_PROFILE: RACK_OPTS_INC(tcp_profile); error = rack_set_profile(rack, optval); break; case TCP_USE_CMP_ACKS: RACK_OPTS_INC(tcp_use_cmp_acks); if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) { /* You can't turn it off once its on! */ error = EINVAL; } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) { rack->r_use_cmp_ack = 1; rack->r_mbuf_queue = 1; tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; } if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; break; case TCP_SHARED_CWND_TIME_LIMIT: RACK_OPTS_INC(tcp_lscwnd); if (optval) rack->r_limit_scw = 1; else rack->r_limit_scw = 0; break; case TCP_RACK_PACE_TO_FILL: RACK_OPTS_INC(tcp_fillcw); if (optval == 0) rack->rc_pace_to_cwnd = 0; else { rack->rc_pace_to_cwnd = 1; if (optval > 1) rack->r_fill_less_agg = 1; } if ((optval >= rack_gp_rtt_maxmul) && rack_gp_rtt_maxmul && (optval < 0xf)) { rack->rc_pace_fill_if_rttin_range = 1; rack->rtt_limit_mul = optval; } else { rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; } break; case TCP_RACK_NO_PUSH_AT_MAX: RACK_OPTS_INC(tcp_npush); if (optval == 0) rack->r_ctl.rc_no_push_at_mrtt = 0; else if (optval < 0xff) rack->r_ctl.rc_no_push_at_mrtt = optval; else error = EINVAL; break; case TCP_SHARED_CWND_ENABLE: RACK_OPTS_INC(tcp_rack_scwnd); if (optval == 0) rack->rack_enable_scwnd = 0; else rack->rack_enable_scwnd = 1; break; case TCP_RACK_MBUF_QUEUE: /* Now do we use the LRO mbuf-queue feature */ RACK_OPTS_INC(tcp_rack_mbufq); if (optval || rack->r_use_cmp_ack) rack->r_mbuf_queue = 1; else rack->r_mbuf_queue = 0; if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack) tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; else tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; break; case TCP_RACK_NONRXT_CFG_RATE: RACK_OPTS_INC(tcp_rack_cfg_rate); if (optval == 0) rack->rack_rec_nonrxt_use_cr = 0; else rack->rack_rec_nonrxt_use_cr = 1; break; case TCP_NO_PRR: RACK_OPTS_INC(tcp_rack_noprr); if (optval == 0) rack->rack_no_prr = 0; else if (optval == 1) rack->rack_no_prr = 1; else if (optval == 2) rack->no_prr_addback = 1; else error = EINVAL; break; case TCP_TIMELY_DYN_ADJ: RACK_OPTS_INC(tcp_timely_dyn); if (optval == 0) rack->rc_gp_dyn_mul = 0; else { rack->rc_gp_dyn_mul = 1; if (optval >= 100) { /* * If the user sets something 100 or more * its the gp_ca value. */ rack->r_ctl.rack_per_of_gp_ca = optval; } } break; case TCP_RACK_DO_DETECTION: RACK_OPTS_INC(tcp_rack_do_detection); if (optval == 0) rack->do_detection = 0; else rack->do_detection = 1; break; case TCP_RACK_TLP_USE: if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) { error = EINVAL; break; } RACK_OPTS_INC(tcp_tlp_use); rack->rack_tlp_threshold_use = optval; break; case TCP_RACK_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ RACK_OPTS_INC(tcp_rack_tlp_reduce); rack->r_ctl.rc_tlp_cwnd_reduce = optval; break; /* Pacing related ones */ case TCP_RACK_PACE_ALWAYS: /* * zero is old rack method, 1 is new * method using a pacing rate. */ RACK_OPTS_INC(tcp_rack_pace_always); if (optval > 0) { if (rack->rc_always_pace) { error = EALREADY; break; } else if (tcp_can_enable_pacing()) { rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); } else { error = ENOSPC; break; } } else { if (rack->rc_always_pace) { tcp_decrement_paced_conn(); rack->rc_always_pace = 0; rack_undo_cc_pacing(rack); } } if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack) tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; else tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; /* A rate may be set irate or other, if so set seg size */ rack_update_seg(rack); break; case TCP_BBR_RACK_INIT_RATE: RACK_OPTS_INC(tcp_initial_rate); val = optval; /* Change from kbits per second to bytes per second */ val *= 1000; val /= 8; rack->r_ctl.init_rate = val; if (rack->rc_init_win != rack_default_init_window) { uint32_t win, snt; /* * Options don't always get applied * in the order you think. So in order * to assure we update a cwnd we need * to check and see if we are still * where we should raise the cwnd. */ win = rc_init_window(rack); if (SEQ_GT(tp->snd_max, tp->iss)) snt = tp->snd_max - tp->iss; else snt = 0; if ((snt < win) && (tp->snd_cwnd < win)) tp->snd_cwnd = win; } if (rack->rc_always_pace) rack_update_seg(rack); break; case TCP_BBR_IWINTSO: RACK_OPTS_INC(tcp_initial_win); if (optval && (optval <= 0xff)) { uint32_t win, snt; rack->rc_init_win = optval; win = rc_init_window(rack); if (SEQ_GT(tp->snd_max, tp->iss)) snt = tp->snd_max - tp->iss; else snt = 0; if ((snt < win) && (tp->t_srtt | #ifdef NETFLIX_PEAKRATE tp->t_maxpeakrate | #endif rack->r_ctl.init_rate)) { /* * We are not past the initial window * and we have some bases for pacing, * so we need to possibly adjust up * the cwnd. Note even if we don't set * the cwnd, its still ok to raise the rc_init_win * which can be used coming out of idle when we * would have a rate. */ if (tp->snd_cwnd < win) tp->snd_cwnd = win; } if (rack->rc_always_pace) rack_update_seg(rack); } else error = EINVAL; break; case TCP_RACK_FORCE_MSEG: RACK_OPTS_INC(tcp_rack_force_max_seg); if (optval) rack->rc_force_max_seg = 1; else rack->rc_force_max_seg = 0; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments size in a pace in bytes */ RACK_OPTS_INC(tcp_rack_max_seg); rack->rc_user_set_max_segs = optval; rack_set_pace_segments(tp, rack, __LINE__, NULL); break; case TCP_RACK_PACE_RATE_REC: /* Set the fixed pacing rate in Bytes per second ca */ RACK_OPTS_INC(tcp_rack_pace_rate_rec); rack->r_ctl.rc_fixed_pacing_rate_rec = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0) rack->r_ctl.rc_fixed_pacing_rate_ca = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0) rack->r_ctl.rc_fixed_pacing_rate_ss = optval; rack->use_fixed_rate = 1; if (rack->rc_always_pace) rack_set_cc_pacing(rack); rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_fixed_pacing_rate_ss, rack->r_ctl.rc_fixed_pacing_rate_ca, rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8, __LINE__, NULL,0); break; case TCP_RACK_PACE_RATE_SS: /* Set the fixed pacing rate in Bytes per second ca */ RACK_OPTS_INC(tcp_rack_pace_rate_ss); rack->r_ctl.rc_fixed_pacing_rate_ss = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0) rack->r_ctl.rc_fixed_pacing_rate_ca = optval; if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0) rack->r_ctl.rc_fixed_pacing_rate_rec = optval; rack->use_fixed_rate = 1; if (rack->rc_always_pace) rack_set_cc_pacing(rack); rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_fixed_pacing_rate_ss, rack->r_ctl.rc_fixed_pacing_rate_ca, rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8, __LINE__, NULL, 0); break; case TCP_RACK_PACE_RATE_CA: /* Set the fixed pacing rate in Bytes per second ca */ RACK_OPTS_INC(tcp_rack_pace_rate_ca); rack->r_ctl.rc_fixed_pacing_rate_ca = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0) rack->r_ctl.rc_fixed_pacing_rate_ss = optval; if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0) rack->r_ctl.rc_fixed_pacing_rate_rec = optval; rack->use_fixed_rate = 1; if (rack->rc_always_pace) rack_set_cc_pacing(rack); rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_fixed_pacing_rate_ss, rack->r_ctl.rc_fixed_pacing_rate_ca, rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8, __LINE__, NULL, 0); break; case TCP_RACK_GP_INCREASE_REC: RACK_OPTS_INC(tcp_gp_inc_rec); rack->r_ctl.rack_per_of_gp_rec = optval; rack_log_pacing_delay_calc(rack, rack->r_ctl.rack_per_of_gp_ss, rack->r_ctl.rack_per_of_gp_ca, rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1, __LINE__, NULL, 0); break; case TCP_RACK_GP_INCREASE_CA: RACK_OPTS_INC(tcp_gp_inc_ca); ca = optval; if (ca < 100) { /* * We don't allow any reduction * over the GP b/w. */ error = EINVAL; break; } rack->r_ctl.rack_per_of_gp_ca = ca; rack_log_pacing_delay_calc(rack, rack->r_ctl.rack_per_of_gp_ss, rack->r_ctl.rack_per_of_gp_ca, rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1, __LINE__, NULL, 0); break; case TCP_RACK_GP_INCREASE_SS: RACK_OPTS_INC(tcp_gp_inc_ss); ss = optval; if (ss < 100) { /* * We don't allow any reduction * over the GP b/w. */ error = EINVAL; break; } rack->r_ctl.rack_per_of_gp_ss = ss; rack_log_pacing_delay_calc(rack, rack->r_ctl.rack_per_of_gp_ss, rack->r_ctl.rack_per_of_gp_ca, rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1, __LINE__, NULL, 0); break; case TCP_RACK_RR_CONF: RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate); if (optval && optval <= 3) rack->r_rr_config = optval; else rack->r_rr_config = 0; break; case TCP_HDWR_RATE_CAP: RACK_OPTS_INC(tcp_hdwr_rate_cap); if (optval) { if (rack->r_rack_hw_rate_caps == 0) rack->r_rack_hw_rate_caps = 1; else error = EALREADY; } else { rack->r_rack_hw_rate_caps = 0; } break; case TCP_BBR_HDWR_PACE: RACK_OPTS_INC(tcp_hdwr_pacing); if (optval){ if (rack->rack_hdrw_pacing == 0) { rack->rack_hdw_pace_ena = 1; rack->rack_attempt_hdwr_pace = 0; } else error = EALREADY; } else { rack->rack_hdw_pace_ena = 0; #ifdef RATELIMIT if (rack->r_ctl.crte != NULL) { rack->rack_hdrw_pacing = 0; rack->rack_attempt_hdwr_pace = 0; tcp_rel_pacing_rate(rack->r_ctl.crte, tp); rack->r_ctl.crte = NULL; } #endif } break; /* End Pacing related ones */ case TCP_RACK_PRR_SENDALOT: /* Allow PRR to send more than one seg */ RACK_OPTS_INC(tcp_rack_prr_sendalot); rack->r_ctl.rc_prr_sendalot = optval; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ RACK_OPTS_INC(tcp_rack_min_to); rack->r_ctl.rc_min_to = optval; break; case TCP_RACK_EARLY_SEG: /* If early recovery max segments */ RACK_OPTS_INC(tcp_rack_early_seg); rack->r_ctl.rc_early_recovery_segs = optval; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ RACK_OPTS_INC(tcp_rack_reord_thresh); if ((optval > 0) && (optval < 31)) rack->r_ctl.rc_reorder_shift = optval; else error = EINVAL; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ RACK_OPTS_INC(tcp_rack_reord_fade); rack->r_ctl.rc_reorder_fade = optval; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ RACK_OPTS_INC(tcp_rack_tlp_thresh); if (optval) rack->r_ctl.rc_tlp_threshold = optval; else error = EINVAL; break; case TCP_BBR_USE_RACK_RR: RACK_OPTS_INC(tcp_rack_rr); if (optval) rack->use_rack_rr = 1; else rack->use_rack_rr = 0; break; case TCP_FAST_RSM_HACK: RACK_OPTS_INC(tcp_rack_fastrsm_hack); if (optval) rack->fast_rsm_hack = 1; else rack->fast_rsm_hack = 0; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ RACK_OPTS_INC(tcp_rack_pkt_delay); rack->r_ctl.rc_pkt_delay = optval; break; case TCP_DELACK: RACK_OPTS_INC(tcp_rack_delayed_ack); if (optval == 0) tp->t_delayed_ack = 0; else tp->t_delayed_ack = 1; if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; NET_EPOCH_ENTER(et); rack_output(tp); NET_EPOCH_EXIT(et); } break; case TCP_BBR_RACK_RTT_USE: RACK_OPTS_INC(tcp_rack_rtt_use); if ((optval != USE_RTT_HIGH) && (optval != USE_RTT_LOW) && (optval != USE_RTT_AVG)) error = EINVAL; else rack->r_ctl.rc_rate_sample_method = optval; break; case TCP_DATA_AFTER_CLOSE: RACK_OPTS_INC(tcp_data_after_close); if (optval) rack->rc_allow_data_af_clo = 1; else rack->rc_allow_data_af_clo = 0; break; default: break; } #ifdef NETFLIX_STATS tcp_log_socket_option(tp, sopt_name, optval, error); #endif return (error); } static void rack_apply_deferred_options(struct tcp_rack *rack) { struct deferred_opt_list *dol, *sdol; uint32_t s_optval; TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) { TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next); /* Disadvantage of deferal is you loose the error return */ s_optval = (uint32_t)dol->optval; (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval); free(dol, M_TCPDO); } } static void rack_hw_tls_change(struct tcpcb *tp, int chg) { /* * HW tls state has changed.. fix all * rsm's in flight. */ struct tcp_rack *rack; struct rack_sendmap *rsm; rack = (struct tcp_rack *)tp->t_fb_ptr; RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if (chg) rsm->r_hw_tls = 1; else rsm->r_hw_tls = 0; } if (chg) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; } static int rack_pru_options(struct tcpcb *tp, int flags) { if (flags & PRUS_OOB) return (EOPNOTSUPP); return (0); } static struct tcp_function_block __tcp_rack = { .tfb_tcp_block_name = __XSTRING(STACKNAME), .tfb_tcp_output = rack_output, .tfb_do_queued_segments = ctf_do_queued_segments, .tfb_do_segment_nounlock = rack_do_segment_nounlock, .tfb_tcp_do_segment = rack_do_segment, .tfb_tcp_ctloutput = rack_ctloutput, .tfb_tcp_fb_init = rack_init, .tfb_tcp_fb_fini = rack_fini, .tfb_tcp_timer_stop_all = rack_stopall, .tfb_tcp_timer_activate = rack_timer_activate, .tfb_tcp_timer_active = rack_timer_active, .tfb_tcp_timer_stop = rack_timer_stop, .tfb_tcp_rexmit_tmr = rack_remxt_tmr, .tfb_tcp_handoff_ok = rack_handoff_ok, .tfb_tcp_mtu_chg = rack_mtu_change, .tfb_pru_options = rack_pru_options, .tfb_hwtls_change = rack_hw_tls_change, }; /* * rack_ctloutput() must drop the inpcb lock before performing copyin on * socket option arguments. When it re-acquires the lock after the copy, it * has to revalidate that the connection is still valid for the socket * option. */ static int rack_set_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) { uint64_t loptval; int32_t error = 0, optval; switch (sopt->sopt_name) { case TCP_RACK_TLP_REDUCE: /* URL:tlp_reduce */ /* Pacing related ones */ case TCP_RACK_PACE_ALWAYS: /* URL:pace_always */ case TCP_BBR_RACK_INIT_RATE: /* URL:irate */ case TCP_BBR_IWINTSO: /* URL:tso_iwin */ case TCP_RACK_PACE_MAX_SEG: /* URL:pace_max_seg */ case TCP_RACK_FORCE_MSEG: /* URL:force_max_seg */ case TCP_RACK_PACE_RATE_CA: /* URL:pr_ca */ case TCP_RACK_PACE_RATE_SS: /* URL:pr_ss*/ case TCP_RACK_PACE_RATE_REC: /* URL:pr_rec */ case TCP_RACK_GP_INCREASE_CA: /* URL:gp_inc_ca */ case TCP_RACK_GP_INCREASE_SS: /* URL:gp_inc_ss */ case TCP_RACK_GP_INCREASE_REC: /* URL:gp_inc_rec */ case TCP_RACK_RR_CONF: /* URL:rrr_conf */ case TCP_BBR_HDWR_PACE: /* URL:hdwrpace */ case TCP_HDWR_RATE_CAP: /* URL:hdwrcap boolean */ case TCP_PACING_RATE_CAP: /* URL:cap -- used by side-channel */ case TCP_HDWR_UP_ONLY: /* URL:uponly -- hardware pacing boolean */ /* End pacing related */ case TCP_FAST_RSM_HACK: /* URL:frsm_hack */ case TCP_DELACK: /* URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */ case TCP_RACK_PRR_SENDALOT: /* URL:prr_sendalot */ case TCP_RACK_MIN_TO: /* URL:min_to */ case TCP_RACK_EARLY_SEG: /* URL:early_seg */ case TCP_RACK_REORD_THRESH: /* URL:reord_thresh */ case TCP_RACK_REORD_FADE: /* URL:reord_fade */ case TCP_RACK_TLP_THRESH: /* URL:tlp_thresh */ case TCP_RACK_PKT_DELAY: /* URL:pkt_delay */ case TCP_RACK_TLP_USE: /* URL:tlp_use */ case TCP_BBR_RACK_RTT_USE: /* URL:rttuse */ case TCP_BBR_USE_RACK_RR: /* URL:rackrr */ case TCP_RACK_DO_DETECTION: /* URL:detect */ case TCP_NO_PRR: /* URL:noprr */ case TCP_TIMELY_DYN_ADJ: /* URL:dynamic */ case TCP_DATA_AFTER_CLOSE: /* no URL */ case TCP_RACK_NONRXT_CFG_RATE: /* URL:nonrxtcr */ case TCP_SHARED_CWND_ENABLE: /* URL:scwnd */ case TCP_RACK_MBUF_QUEUE: /* URL:mqueue */ case TCP_RACK_NO_PUSH_AT_MAX: /* URL:npush */ case TCP_RACK_PACE_TO_FILL: /* URL:fillcw */ case TCP_SHARED_CWND_TIME_LIMIT: /* URL:lscwnd */ case TCP_RACK_PROFILE: /* URL:profile */ case TCP_USE_CMP_ACKS: /* URL:cmpack */ case TCP_RACK_ABC_VAL: /* URL:labc */ case TCP_REC_ABC_VAL: /* URL:reclabc */ case TCP_RACK_MEASURE_CNT: /* URL:measurecnt */ case TCP_DEFER_OPTIONS: /* URL:defer */ case TCP_RACK_PACING_BETA: /* URL:pacing_beta */ case TCP_RACK_PACING_BETA_ECN: /* URL:pacing_beta_ecn */ case TCP_RACK_TIMER_SLOP: /* URL:timer_slop */ break; default: /* Filter off all unknown options to the base stack */ return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); if (sopt->sopt_name == TCP_PACING_RATE_CAP) { error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval)); /* * We truncate it down to 32 bits for the socket-option trace this * means rates > 34Gbps won't show right, but thats probably ok. */ optval = (uint32_t)loptval; } else { error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); /* Save it in 64 bit form too */ loptval = optval; } if (error) return (error); INP_WLOCK(inp); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { INP_WUNLOCK(inp); return (ECONNRESET); } if (tp->t_fb != &__tcp_rack) { INP_WUNLOCK(inp); return (ENOPROTOOPT); } if (rack->defer_options && (rack->gp_ready == 0) && (sopt->sopt_name != TCP_DEFER_OPTIONS) && (sopt->sopt_name != TCP_RACK_PACING_BETA) && (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) && (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) { /* Options are beind deferred */ if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) { INP_WUNLOCK(inp); return (0); } else { /* No memory to defer, fail */ INP_WUNLOCK(inp); return (ENOMEM); } } error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval); INP_WUNLOCK(inp); return (error); } static void rack_fill_info(struct tcpcb *tp, struct tcp_info *ti) { INP_WLOCK_ASSERT(tp->t_inpcb); bzero(ti, sizeof(*ti)); ti->tcpi_state = tp->t_state; if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP)) ti->tcpi_options |= TCPI_OPT_TIMESTAMPS; if (tp->t_flags & TF_SACK_PERMIT) ti->tcpi_options |= TCPI_OPT_SACK; if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) { ti->tcpi_options |= TCPI_OPT_WSCALE; ti->tcpi_snd_wscale = tp->snd_scale; ti->tcpi_rcv_wscale = tp->rcv_scale; } if (tp->t_flags2 & TF2_ECN_PERMIT) ti->tcpi_options |= TCPI_OPT_ECN; if (tp->t_flags & TF_FASTOPEN) ti->tcpi_options |= TCPI_OPT_TFO; /* still kept in ticks is t_rcvtime */ ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick; /* Since we hold everything in precise useconds this is easy */ ti->tcpi_rtt = tp->t_srtt; ti->tcpi_rttvar = tp->t_rttvar; ti->tcpi_rto = tp->t_rxtcur; ti->tcpi_snd_ssthresh = tp->snd_ssthresh; ti->tcpi_snd_cwnd = tp->snd_cwnd; /* * FreeBSD-specific extension fields for tcp_info. */ ti->tcpi_rcv_space = tp->rcv_wnd; ti->tcpi_rcv_nxt = tp->rcv_nxt; ti->tcpi_snd_wnd = tp->snd_wnd; ti->tcpi_snd_bwnd = 0; /* Unused, kept for compat. */ ti->tcpi_snd_nxt = tp->snd_nxt; ti->tcpi_snd_mss = tp->t_maxseg; ti->tcpi_rcv_mss = tp->t_maxseg; ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack; ti->tcpi_rcv_ooopack = tp->t_rcvoopack; ti->tcpi_snd_zerowin = tp->t_sndzerowin; #ifdef NETFLIX_STATS ti->tcpi_total_tlp = tp->t_sndtlppack; ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte; memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo)); #endif #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { ti->tcpi_options |= TCPI_OPT_TOE; tcp_offload_tcp_info(tp, ti); } #endif } static int rack_get_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) { int32_t error, optval; uint64_t val, loptval; struct tcp_info ti; /* * Because all our options are either boolean or an int, we can just * pull everything into optval and then unlock and copy. If we ever * add a option that is not a int, then this will have quite an * impact to this routine. */ error = 0; switch (sopt->sopt_name) { case TCP_INFO: /* First get the info filled */ rack_fill_info(tp, &ti); /* Fix up the rtt related fields if needed */ INP_WUNLOCK(inp); error = sooptcopyout(sopt, &ti, sizeof ti); return (error); /* * Beta is the congestion control value for NewReno that influences how * much of a backoff happens when loss is detected. It is normally set * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value * when you exit recovery. */ case TCP_RACK_PACING_BETA: if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) error = EINVAL; else if (rack->rc_pacing_cc_set == 0) optval = rack->r_ctl.rc_saved_beta.beta; else { /* * Reach out into the CC data and report back what * I have previously set. Yeah it looks hackish but * we don't want to report the saved values. */ if (tp->ccv->cc_data) optval = ((struct newreno *)tp->ccv->cc_data)->beta; else error = EINVAL; } break; /* * Beta_ecn is the congestion control value for NewReno that influences how * much of a backoff happens when a ECN mark is detected. It is normally set * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when * you exit recovery. Note that classic ECN has a beta of 50, it is only * ABE Ecn that uses this "less" value, but we do too with pacing :) */ case TCP_RACK_PACING_BETA_ECN: if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) error = EINVAL; else if (rack->rc_pacing_cc_set == 0) optval = rack->r_ctl.rc_saved_beta.beta_ecn; else { /* * Reach out into the CC data and report back what * I have previously set. Yeah it looks hackish but * we don't want to report the saved values. */ if (tp->ccv->cc_data) optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn; else error = EINVAL; } break; case TCP_FAST_RSM_HACK: optval = rack->fast_rsm_hack; break; case TCP_DEFER_OPTIONS: optval = rack->defer_options; break; case TCP_RACK_MEASURE_CNT: optval = rack->r_ctl.req_measurements; break; case TCP_REC_ABC_VAL: optval = rack->r_use_labc_for_rec; break; case TCP_RACK_ABC_VAL: optval = rack->rc_labc; break; case TCP_HDWR_UP_ONLY: optval= rack->r_up_only; break; case TCP_PACING_RATE_CAP: loptval = rack->r_ctl.bw_rate_cap; break; case TCP_RACK_PROFILE: /* You cannot retrieve a profile, its write only */ error = EINVAL; break; case TCP_USE_CMP_ACKS: optval = rack->r_use_cmp_ack; break; case TCP_RACK_PACE_TO_FILL: optval = rack->rc_pace_to_cwnd; if (optval && rack->r_fill_less_agg) optval++; break; case TCP_RACK_NO_PUSH_AT_MAX: optval = rack->r_ctl.rc_no_push_at_mrtt; break; case TCP_SHARED_CWND_ENABLE: optval = rack->rack_enable_scwnd; break; case TCP_RACK_NONRXT_CFG_RATE: optval = rack->rack_rec_nonrxt_use_cr; break; case TCP_NO_PRR: if (rack->rack_no_prr == 1) optval = 1; else if (rack->no_prr_addback == 1) optval = 2; else optval = 0; break; case TCP_RACK_DO_DETECTION: optval = rack->do_detection; break; case TCP_RACK_MBUF_QUEUE: /* Now do we use the LRO mbuf-queue feature */ optval = rack->r_mbuf_queue; break; case TCP_TIMELY_DYN_ADJ: optval = rack->rc_gp_dyn_mul; break; case TCP_BBR_IWINTSO: optval = rack->rc_init_win; break; case TCP_RACK_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ optval = rack->r_ctl.rc_tlp_cwnd_reduce; break; case TCP_BBR_RACK_INIT_RATE: val = rack->r_ctl.init_rate; /* convert to kbits per sec */ val *= 8; val /= 1000; optval = (uint32_t)val; break; case TCP_RACK_FORCE_MSEG: optval = rack->rc_force_max_seg; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments in a pace */ optval = rack->rc_user_set_max_segs; break; case TCP_RACK_PACE_ALWAYS: /* Use the always pace method */ optval = rack->rc_always_pace; break; case TCP_RACK_PRR_SENDALOT: /* Allow PRR to send more than one seg */ optval = rack->r_ctl.rc_prr_sendalot; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ optval = rack->r_ctl.rc_min_to; break; case TCP_RACK_EARLY_SEG: /* If early recovery max segments */ optval = rack->r_ctl.rc_early_recovery_segs; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ optval = rack->r_ctl.rc_reorder_shift; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ optval = rack->r_ctl.rc_reorder_fade; break; case TCP_BBR_USE_RACK_RR: /* Do we use the rack cheat for rxt */ optval = rack->use_rack_rr; break; case TCP_RACK_RR_CONF: optval = rack->r_rr_config; break; case TCP_HDWR_RATE_CAP: optval = rack->r_rack_hw_rate_caps; break; case TCP_BBR_HDWR_PACE: optval = rack->rack_hdw_pace_ena; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ optval = rack->r_ctl.rc_tlp_threshold; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ optval = rack->r_ctl.rc_pkt_delay; break; case TCP_RACK_TLP_USE: optval = rack->rack_tlp_threshold_use; break; case TCP_RACK_PACE_RATE_CA: optval = rack->r_ctl.rc_fixed_pacing_rate_ca; break; case TCP_RACK_PACE_RATE_SS: optval = rack->r_ctl.rc_fixed_pacing_rate_ss; break; case TCP_RACK_PACE_RATE_REC: optval = rack->r_ctl.rc_fixed_pacing_rate_rec; break; case TCP_RACK_GP_INCREASE_SS: optval = rack->r_ctl.rack_per_of_gp_ca; break; case TCP_RACK_GP_INCREASE_CA: optval = rack->r_ctl.rack_per_of_gp_ss; break; case TCP_BBR_RACK_RTT_USE: optval = rack->r_ctl.rc_rate_sample_method; break; case TCP_DELACK: optval = tp->t_delayed_ack; break; case TCP_DATA_AFTER_CLOSE: optval = rack->rc_allow_data_af_clo; break; case TCP_SHARED_CWND_TIME_LIMIT: optval = rack->r_limit_scw; break; case TCP_RACK_TIMER_SLOP: optval = rack->r_ctl.timer_slop; break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); if (error == 0) { if (TCP_PACING_RATE_CAP) error = sooptcopyout(sopt, &loptval, sizeof loptval); else error = sooptcopyout(sopt, &optval, sizeof optval); } return (error); } static int rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) { int32_t error = EINVAL; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack == NULL) { /* Huh? */ goto out; } if (sopt->sopt_dir == SOPT_SET) { return (rack_set_sockopt(so, sopt, inp, tp, rack)); } else if (sopt->sopt_dir == SOPT_GET) { return (rack_get_sockopt(so, sopt, inp, tp, rack)); } out: INP_WUNLOCK(inp); return (error); } static const char *rack_stack_names[] = { __XSTRING(STACKNAME), #ifdef STACKALIAS __XSTRING(STACKALIAS), #endif }; static int rack_ctor(void *mem, int32_t size, void *arg, int32_t how) { memset(mem, 0, size); return (0); } static void rack_dtor(void *mem, int32_t size, void *arg) { } static bool rack_mod_inited = false; static int tcp_addrack(module_t mod, int32_t type, void *data) { int32_t err = 0; int num_stacks; switch (type) { case MOD_LOAD: rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map", sizeof(struct rack_sendmap), rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0); rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", sizeof(struct tcp_rack), rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); sysctl_ctx_init(&rack_sysctl_ctx); rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp), OID_AUTO, #ifdef STACKALIAS __XSTRING(STACKALIAS), #else __XSTRING(STACKNAME), #endif CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); if (rack_sysctl_root == NULL) { printf("Failed to add sysctl node\n"); err = EFAULT; goto free_uma; } rack_init_sysctls(); num_stacks = nitems(rack_stack_names); err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK, rack_stack_names, &num_stacks); if (err) { printf("Failed to register %s stack name for " "%s module\n", rack_stack_names[num_stacks], __XSTRING(MODNAME)); sysctl_ctx_free(&rack_sysctl_ctx); free_uma: uma_zdestroy(rack_zone); uma_zdestroy(rack_pcb_zone); rack_counter_destroy(); printf("Failed to register rack module -- err:%d\n", err); return (err); } tcp_lro_reg_mbufq(); rack_mod_inited = true; break; case MOD_QUIESCE: err = deregister_tcp_functions(&__tcp_rack, true, false); break; case MOD_UNLOAD: err = deregister_tcp_functions(&__tcp_rack, false, true); if (err == EBUSY) break; if (rack_mod_inited) { uma_zdestroy(rack_zone); uma_zdestroy(rack_pcb_zone); sysctl_ctx_free(&rack_sysctl_ctx); rack_counter_destroy(); rack_mod_inited = false; } tcp_lro_dereg_mbufq(); err = 0; break; default: return (EOPNOTSUPP); } return (err); } static moduledata_t tcp_rack = { .name = __XSTRING(MODNAME), .evhand = tcp_addrack, .priv = 0 }; MODULE_VERSION(MODNAME, 1); DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);