Index: head/sys/dev/hwpmc/hwpmc_mod.c =================================================================== --- head/sys/dev/hwpmc/hwpmc_mod.c (revision 359437) +++ head/sys/dev/hwpmc/hwpmc_mod.c (revision 359438) @@ -1,5972 +1,5969 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2003-2008 Joseph Koshy * Copyright (c) 2007 The FreeBSD Foundation * Copyright (c) 2018 Matthew Macy * All rights reserved. * * 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. * * 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 +#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 #include /* needs to be after */ #include #include #include #include #include #include #include #include "hwpmc_soft.h" #define PMC_EPOCH_ENTER() struct epoch_tracker pmc_et; epoch_enter_preempt(global_epoch_preempt, &pmc_et) #define PMC_EPOCH_EXIT() epoch_exit_preempt(global_epoch_preempt, &pmc_et) /* * Types */ enum pmc_flags { PMC_FLAG_NONE = 0x00, /* do nothing */ PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */ PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */ PMC_FLAG_NOWAIT = 0x04, /* do not wait for mallocs */ }; /* * The offset in sysent where the syscall is allocated. */ static int pmc_syscall_num = NO_SYSCALL; struct pmc_cpu **pmc_pcpu; /* per-cpu state */ pmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */ #define PMC_PCPU_SAVED(C,R) pmc_pcpu_saved[(R) + md->pmd_npmc*(C)] struct mtx_pool *pmc_mtxpool; static int *pmc_pmcdisp; /* PMC row dispositions */ #define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0) #define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0) #define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0) #define PMC_MARK_ROW_FREE(R) do { \ pmc_pmcdisp[(R)] = 0; \ } while (0) #define PMC_MARK_ROW_STANDALONE(R) do { \ KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ __LINE__)); \ atomic_add_int(&pmc_pmcdisp[(R)], -1); \ KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()), \ ("[pmc,%d] row disposition error", __LINE__)); \ } while (0) #define PMC_UNMARK_ROW_STANDALONE(R) do { \ atomic_add_int(&pmc_pmcdisp[(R)], 1); \ KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ __LINE__)); \ } while (0) #define PMC_MARK_ROW_THREAD(R) do { \ KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ __LINE__)); \ atomic_add_int(&pmc_pmcdisp[(R)], 1); \ } while (0) #define PMC_UNMARK_ROW_THREAD(R) do { \ atomic_add_int(&pmc_pmcdisp[(R)], -1); \ KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ __LINE__)); \ } while (0) /* various event handlers */ static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag, pmc_kld_unload_tag; /* Module statistics */ struct pmc_driverstats pmc_stats; /* Machine/processor dependent operations */ static struct pmc_mdep *md; /* * Hash tables mapping owner processes and target threads to PMCs. */ struct mtx pmc_processhash_mtx; /* spin mutex */ static u_long pmc_processhashmask; static LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash; /* * Hash table of PMC owner descriptors. This table is protected by * the shared PMC "sx" lock. */ static u_long pmc_ownerhashmask; static LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash; /* * List of PMC owners with system-wide sampling PMCs. */ static CK_LIST_HEAD(, pmc_owner) pmc_ss_owners; /* * List of free thread entries. This is protected by the spin * mutex. */ static struct mtx pmc_threadfreelist_mtx; /* spin mutex */ static LIST_HEAD(, pmc_thread) pmc_threadfreelist; static int pmc_threadfreelist_entries=0; #define THREADENTRY_SIZE \ (sizeof(struct pmc_thread) + (md->pmd_npmc * sizeof(struct pmc_threadpmcstate))) /* * Task to free thread descriptors */ -static struct grouptask free_gtask; +static struct task free_task; /* * A map of row indices to classdep structures. */ static struct pmc_classdep **pmc_rowindex_to_classdep; /* * Prototypes */ #ifdef HWPMC_DEBUG static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS); static int pmc_debugflags_parse(char *newstr, char *fence); #endif static int load(struct module *module, int cmd, void *arg); static int pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf); static void pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp); static int pmc_attach_process(struct proc *p, struct pmc *pm); static struct pmc *pmc_allocate_pmc_descriptor(void); static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p); static int pmc_attach_one_process(struct proc *p, struct pmc *pm); static int pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu); static int pmc_can_attach(struct pmc *pm, struct proc *p); static void pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf); static void pmc_cleanup(void); static int pmc_detach_process(struct proc *p, struct pmc *pm); static int pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags); static void pmc_destroy_owner_descriptor(struct pmc_owner *po); static void pmc_destroy_pmc_descriptor(struct pmc *pm); static void pmc_destroy_process_descriptor(struct pmc_process *pp); static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p); static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm); static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmc); static struct pmc_process *pmc_find_process_descriptor(struct proc *p, uint32_t mode); static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td, uint32_t mode); static void pmc_force_context_switch(void); static void pmc_link_target_process(struct pmc *pm, struct pmc_process *pp); static void pmc_log_all_process_mappings(struct pmc_owner *po); static void pmc_log_kernel_mappings(struct pmc *pm); static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p); static void pmc_maybe_remove_owner(struct pmc_owner *po); static void pmc_process_csw_in(struct thread *td); static void pmc_process_csw_out(struct thread *td); static void pmc_process_exit(void *arg, struct proc *p); static void pmc_process_fork(void *arg, struct proc *p1, struct proc *p2, int n); static void pmc_process_samples(int cpu, ring_type_t soft); static void pmc_release_pmc_descriptor(struct pmc *pmc); static void pmc_process_thread_add(struct thread *td); static void pmc_process_thread_delete(struct thread *td); static void pmc_process_thread_userret(struct thread *td); static void pmc_remove_owner(struct pmc_owner *po); static void pmc_remove_process_descriptor(struct pmc_process *pp); static void pmc_restore_cpu_binding(struct pmc_binding *pb); static void pmc_save_cpu_binding(struct pmc_binding *pb); static void pmc_select_cpu(int cpu); static int pmc_start(struct pmc *pm); static int pmc_stop(struct pmc *pm); static int pmc_syscall_handler(struct thread *td, void *syscall_args); static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void); static void pmc_thread_descriptor_pool_drain(void); static void pmc_thread_descriptor_pool_free(struct pmc_thread *pt); static void pmc_unlink_target_process(struct pmc *pmc, struct pmc_process *pp); static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp); static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp); static struct pmc_mdep *pmc_generic_cpu_initialize(void); static void pmc_generic_cpu_finalize(struct pmc_mdep *md); static void pmc_post_callchain_callback(void); static void pmc_process_threadcreate(struct thread *td); static void pmc_process_threadexit(struct thread *td); static void pmc_process_proccreate(struct proc *p); static void pmc_process_allproc(struct pmc *pm); /* * Kernel tunables and sysctl(8) interface. */ SYSCTL_DECL(_kern_hwpmc); SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "HWPMC stats"); /* Stats. */ SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW, &pmc_stats.pm_intr_ignored, "# of interrupts ignored"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW, &pmc_stats.pm_intr_processed, "# of interrupts processed"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW, &pmc_stats.pm_intr_bufferfull, "# of interrupts where buffer was full"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW, &pmc_stats.pm_syscalls, "# of syscalls"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW, &pmc_stats.pm_syscall_errors, "# of syscall_errors"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW, &pmc_stats.pm_buffer_requests, "# of buffer requests"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, CTLFLAG_RW, &pmc_stats.pm_buffer_requests_failed, "# of buffer requests which failed"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW, &pmc_stats.pm_log_sweeps, "# of ?"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW, &pmc_stats.pm_merges, "# of times kernel stack was found for user trace"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW, &pmc_stats.pm_overwrites, "# of times a sample was overwritten before being logged"); static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH; SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN, &pmc_callchaindepth, 0, "depth of call chain records"); char pmc_cpuid[64]; SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD, pmc_cpuid, 0, "cpu version string"); #ifdef HWPMC_DEBUG struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS; char pmc_debugstr[PMC_DEBUG_STRSIZE]; TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, sizeof(pmc_debugstr)); SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags, CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_NEEDGIANT, 0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags"); #endif /* * kern.hwpmc.hashrows -- determines the number of rows in the * of the hash table used to look up threads */ static int pmc_hashsize = PMC_HASH_SIZE; SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN, &pmc_hashsize, 0, "rows in hash tables"); /* * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU */ static int pmc_nsamples = PMC_NSAMPLES; SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN, &pmc_nsamples, 0, "number of PC samples per CPU"); static uint64_t pmc_sample_mask = PMC_NSAMPLES-1; /* * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool. */ static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE; SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN, &pmc_mtxpool_size, 0, "size of spin mutex pool"); /* * kern.hwpmc.threadfreelist_entries -- number of free entries */ SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD, &pmc_threadfreelist_entries, 0, "number of avalable thread entries"); /* * kern.hwpmc.threadfreelist_max -- maximum number of free entries */ static int pmc_threadfreelist_max = PMC_THREADLIST_MAX; SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW, &pmc_threadfreelist_max, 0, "maximum number of available thread entries before freeing some"); /* * security.bsd.unprivileged_syspmcs -- allow non-root processes to * allocate system-wide PMCs. * * Allowing unprivileged processes to allocate system PMCs is convenient * if system-wide measurements need to be taken concurrently with other * per-process measurements. This feature is turned off by default. */ static int pmc_unprivileged_syspmcs = 0; SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN, &pmc_unprivileged_syspmcs, 0, "allow unprivileged process to allocate system PMCs"); /* * Hash function. Discard the lower 2 bits of the pointer since * these are always zero for our uses. The hash multiplier is * round((2^LONG_BIT) * ((sqrt(5)-1)/2)). */ #if LONG_BIT == 64 #define _PMC_HM 11400714819323198486u #elif LONG_BIT == 32 #define _PMC_HM 2654435769u #else #error Must know the size of 'long' to compile #endif #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M)) /* * Syscall structures */ /* The `sysent' for the new syscall */ static struct sysent pmc_sysent = { .sy_narg = 2, .sy_call = pmc_syscall_handler, }; static struct syscall_module_data pmc_syscall_mod = { .chainevh = load, .chainarg = NULL, .offset = &pmc_syscall_num, .new_sysent = &pmc_sysent, .old_sysent = { .sy_narg = 0, .sy_call = NULL }, .flags = SY_THR_STATIC_KLD, }; static moduledata_t pmc_mod = { .name = PMC_MODULE_NAME, .evhand = syscall_module_handler, .priv = &pmc_syscall_mod, }; #ifdef EARLY_AP_STARTUP DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY); #else DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY); #endif MODULE_VERSION(pmc, PMC_VERSION); #ifdef HWPMC_DEBUG enum pmc_dbgparse_state { PMCDS_WS, /* in whitespace */ PMCDS_MAJOR, /* seen a major keyword */ PMCDS_MINOR }; static int pmc_debugflags_parse(char *newstr, char *fence) { char c, *p, *q; struct pmc_debugflags *tmpflags; int error, found, *newbits, tmp; size_t kwlen; tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO); p = newstr; error = 0; for (; p < fence && (c = *p); p++) { /* skip white space */ if (c == ' ' || c == '\t') continue; /* look for a keyword followed by "=" */ for (q = p; p < fence && (c = *p) && c != '='; p++) ; if (c != '=') { error = EINVAL; goto done; } kwlen = p - q; newbits = NULL; /* lookup flag group name */ #define DBG_SET_FLAG_MAJ(S,F) \ if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ newbits = &tmpflags->pdb_ ## F; DBG_SET_FLAG_MAJ("cpu", CPU); DBG_SET_FLAG_MAJ("csw", CSW); DBG_SET_FLAG_MAJ("logging", LOG); DBG_SET_FLAG_MAJ("module", MOD); DBG_SET_FLAG_MAJ("md", MDP); DBG_SET_FLAG_MAJ("owner", OWN); DBG_SET_FLAG_MAJ("pmc", PMC); DBG_SET_FLAG_MAJ("process", PRC); DBG_SET_FLAG_MAJ("sampling", SAM); if (newbits == NULL) { error = EINVAL; goto done; } p++; /* skip the '=' */ /* Now parse the individual flags */ tmp = 0; newflag: for (q = p; p < fence && (c = *p); p++) if (c == ' ' || c == '\t' || c == ',') break; /* p == fence or c == ws or c == "," or c == 0 */ if ((kwlen = p - q) == 0) { *newbits = tmp; continue; } found = 0; #define DBG_SET_FLAG_MIN(S,F) \ if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ tmp |= found = (1 << PMC_DEBUG_MIN_ ## F) /* a '*' denotes all possible flags in the group */ if (kwlen == 1 && *q == '*') tmp = found = ~0; /* look for individual flag names */ DBG_SET_FLAG_MIN("allocaterow", ALR); DBG_SET_FLAG_MIN("allocate", ALL); DBG_SET_FLAG_MIN("attach", ATT); DBG_SET_FLAG_MIN("bind", BND); DBG_SET_FLAG_MIN("config", CFG); DBG_SET_FLAG_MIN("exec", EXC); DBG_SET_FLAG_MIN("exit", EXT); DBG_SET_FLAG_MIN("find", FND); DBG_SET_FLAG_MIN("flush", FLS); DBG_SET_FLAG_MIN("fork", FRK); DBG_SET_FLAG_MIN("getbuf", GTB); DBG_SET_FLAG_MIN("hook", PMH); DBG_SET_FLAG_MIN("init", INI); DBG_SET_FLAG_MIN("intr", INT); DBG_SET_FLAG_MIN("linktarget", TLK); DBG_SET_FLAG_MIN("mayberemove", OMR); DBG_SET_FLAG_MIN("ops", OPS); DBG_SET_FLAG_MIN("read", REA); DBG_SET_FLAG_MIN("register", REG); DBG_SET_FLAG_MIN("release", REL); DBG_SET_FLAG_MIN("remove", ORM); DBG_SET_FLAG_MIN("sample", SAM); DBG_SET_FLAG_MIN("scheduleio", SIO); DBG_SET_FLAG_MIN("select", SEL); DBG_SET_FLAG_MIN("signal", SIG); DBG_SET_FLAG_MIN("swi", SWI); DBG_SET_FLAG_MIN("swo", SWO); DBG_SET_FLAG_MIN("start", STA); DBG_SET_FLAG_MIN("stop", STO); DBG_SET_FLAG_MIN("syscall", PMS); DBG_SET_FLAG_MIN("unlinktarget", TUL); DBG_SET_FLAG_MIN("write", WRI); if (found == 0) { /* unrecognized flag name */ error = EINVAL; goto done; } if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */ *newbits = tmp; continue; } p++; goto newflag; } /* save the new flag set */ bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags)); done: free(tmpflags, M_PMC); return error; } static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS) { char *fence, *newstr; int error; unsigned int n; (void) arg1; (void) arg2; /* unused parameters */ n = sizeof(pmc_debugstr); newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO); (void) strlcpy(newstr, pmc_debugstr, n); error = sysctl_handle_string(oidp, newstr, n, req); /* if there is a new string, parse and copy it */ if (error == 0 && req->newptr != NULL) { fence = newstr + (n < req->newlen ? n : req->newlen + 1); if ((error = pmc_debugflags_parse(newstr, fence)) == 0) (void) strlcpy(pmc_debugstr, newstr, sizeof(pmc_debugstr)); } free(newstr, M_PMC); return error; } #endif /* * Map a row index to a classdep structure and return the adjusted row * index for the PMC class index. */ static struct pmc_classdep * pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri) { struct pmc_classdep *pcd; (void) md; KASSERT(ri >= 0 && ri < md->pmd_npmc, ("[pmc,%d] illegal row-index %d", __LINE__, ri)); pcd = pmc_rowindex_to_classdep[ri]; KASSERT(pcd != NULL, ("[pmc,%d] ri %d null pcd", __LINE__, ri)); *adjri = ri - pcd->pcd_ri; KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num, ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri)); return (pcd); } /* * Concurrency Control * * The driver manages the following data structures: * * - target process descriptors, one per target process * - owner process descriptors (and attached lists), one per owner process * - lookup hash tables for owner and target processes * - PMC descriptors (and attached lists) * - per-cpu hardware state * - the 'hook' variable through which the kernel calls into * this module * - the machine hardware state (managed by the MD layer) * * These data structures are accessed from: * * - thread context-switch code * - interrupt handlers (possibly on multiple cpus) * - kernel threads on multiple cpus running on behalf of user * processes doing system calls * - this driver's private kernel threads * * = Locks and Locking strategy = * * The driver uses four locking strategies for its operation: * * - The global SX lock "pmc_sx" is used to protect internal * data structures. * * Calls into the module by syscall() start with this lock being * held in exclusive mode. Depending on the requested operation, * the lock may be downgraded to 'shared' mode to allow more * concurrent readers into the module. Calls into the module from * other parts of the kernel acquire the lock in shared mode. * * This SX lock is held in exclusive mode for any operations that * modify the linkages between the driver's internal data structures. * * The 'pmc_hook' function pointer is also protected by this lock. * It is only examined with the sx lock held in exclusive mode. The * kernel module is allowed to be unloaded only with the sx lock held * in exclusive mode. In normal syscall handling, after acquiring the * pmc_sx lock we first check that 'pmc_hook' is non-null before * proceeding. This prevents races between the thread unloading the module * and other threads seeking to use the module. * * - Lookups of target process structures and owner process structures * cannot use the global "pmc_sx" SX lock because these lookups need * to happen during context switches and in other critical sections * where sleeping is not allowed. We protect these lookup tables * with their own private spin-mutexes, "pmc_processhash_mtx" and * "pmc_ownerhash_mtx". * * - Interrupt handlers work in a lock free manner. At interrupt * time, handlers look at the PMC pointer (phw->phw_pmc) configured * when the PMC was started. If this pointer is NULL, the interrupt * is ignored after updating driver statistics. We ensure that this * pointer is set (using an atomic operation if necessary) before the * PMC hardware is started. Conversely, this pointer is unset atomically * only after the PMC hardware is stopped. * * We ensure that everything needed for the operation of an * interrupt handler is available without it needing to acquire any * locks. We also ensure that a PMC's software state is destroyed only * after the PMC is taken off hardware (on all CPUs). * * - Context-switch handling with process-private PMCs needs more * care. * * A given process may be the target of multiple PMCs. For example, * PMCATTACH and PMCDETACH may be requested by a process on one CPU * while the target process is running on another. A PMC could also * be getting released because its owner is exiting. We tackle * these situations in the following manner: * * - each target process structure 'pmc_process' has an array * of 'struct pmc *' pointers, one for each hardware PMC. * * - At context switch IN time, each "target" PMC in RUNNING state * gets started on hardware and a pointer to each PMC is copied into * the per-cpu phw array. The 'runcount' for the PMC is * incremented. * * - At context switch OUT time, all process-virtual PMCs are stopped * on hardware. The saved value is added to the PMCs value field * only if the PMC is in a non-deleted state (the PMCs state could * have changed during the current time slice). * * Note that since in-between a switch IN on a processor and a switch * OUT, the PMC could have been released on another CPU. Therefore * context switch OUT always looks at the hardware state to turn * OFF PMCs and will update a PMC's saved value only if reachable * from the target process record. * * - OP PMCRELEASE could be called on a PMC at any time (the PMC could * be attached to many processes at the time of the call and could * be active on multiple CPUs). * * We prevent further scheduling of the PMC by marking it as in * state 'DELETED'. If the runcount of the PMC is non-zero then * this PMC is currently running on a CPU somewhere. The thread * doing the PMCRELEASE operation waits by repeatedly doing a * pause() till the runcount comes to zero. * * The contents of a PMC descriptor (struct pmc) are protected using * a spin-mutex. In order to save space, we use a mutex pool. * * In terms of lock types used by witness(4), we use: * - Type "pmc-sx", used by the global SX lock. * - Type "pmc-sleep", for sleep mutexes used by logger threads. * - Type "pmc-per-proc", for protecting PMC owner descriptors. * - Type "pmc-leaf", used for all other spin mutexes. */ /* * save the cpu binding of the current kthread */ static void pmc_save_cpu_binding(struct pmc_binding *pb) { PMCDBG0(CPU,BND,2, "save-cpu"); thread_lock(curthread); pb->pb_bound = sched_is_bound(curthread); pb->pb_cpu = curthread->td_oncpu; thread_unlock(curthread); PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu); } /* * restore the cpu binding of the current thread */ static void pmc_restore_cpu_binding(struct pmc_binding *pb) { PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d", curthread->td_oncpu, pb->pb_cpu); thread_lock(curthread); if (pb->pb_bound) sched_bind(curthread, pb->pb_cpu); else sched_unbind(curthread); thread_unlock(curthread); PMCDBG0(CPU,BND,2, "restore-cpu done"); } /* * move execution over the specified cpu and bind it there. */ static void pmc_select_cpu(int cpu) { KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[pmc,%d] bad cpu number %d", __LINE__, cpu)); /* Never move to an inactive CPU. */ KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive " "CPU %d", __LINE__, cpu)); PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu); thread_lock(curthread); sched_bind(curthread, cpu); thread_unlock(curthread); KASSERT(curthread->td_oncpu == cpu, ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__, cpu, curthread->td_oncpu)); PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu); } /* * Force a context switch. * * We do this by pause'ing for 1 tick -- invoking mi_switch() is not * guaranteed to force a context switch. */ static void pmc_force_context_switch(void) { pause("pmcctx", 1); } uint64_t pmc_rdtsc(void) { #if defined(__i386__) || defined(__amd64__) if (__predict_true(amd_feature & AMDID_RDTSCP)) return rdtscp(); else return rdtsc(); #else return get_cyclecount(); #endif } /* * Get the file name for an executable. This is a simple wrapper * around vn_fullpath(9). */ static void pmc_getfilename(struct vnode *v, char **fullpath, char **freepath) { *fullpath = "unknown"; *freepath = NULL; vn_fullpath(curthread, v, fullpath, freepath); } /* * remove an process owning PMCs */ void pmc_remove_owner(struct pmc_owner *po) { struct pmc *pm, *tmp; sx_assert(&pmc_sx, SX_XLOCKED); PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po); /* Remove descriptor from the owner hash table */ LIST_REMOVE(po, po_next); /* release all owned PMC descriptors */ LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) { PMCDBG1(OWN,ORM,2, "pmc=%p", pm); KASSERT(pm->pm_owner == po, ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po)); pmc_release_pmc_descriptor(pm); /* will unlink from the list */ pmc_destroy_pmc_descriptor(pm); } KASSERT(po->po_sscount == 0, ("[pmc,%d] SS count not zero", __LINE__)); KASSERT(LIST_EMPTY(&po->po_pmcs), ("[pmc,%d] PMC list not empty", __LINE__)); /* de-configure the log file if present */ if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_deconfigure_log(po); } /* * remove an owner process record if all conditions are met. */ static void pmc_maybe_remove_owner(struct pmc_owner *po) { PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po); /* * Remove owner record if * - this process does not own any PMCs * - this process has not allocated a system-wide sampling buffer */ if (LIST_EMPTY(&po->po_pmcs) && ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) { pmc_remove_owner(po); pmc_destroy_owner_descriptor(po); } } /* * Add an association between a target process and a PMC. */ static void pmc_link_target_process(struct pmc *pm, struct pmc_process *pp) { int ri; struct pmc_target *pt; #ifdef INVARIANTS struct pmc_thread *pt_td; #endif sx_assert(&pmc_sx, SX_XLOCKED); KASSERT(pm != NULL && pp != NULL, ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d", __LINE__, pm, pp->pp_proc->p_pid)); KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1), ("[pmc,%d] Illegal reference count %d for process record %p", __LINE__, pp->pp_refcnt, (void *) pp)); ri = PMC_TO_ROWINDEX(pm); PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p", pm, ri, pp); #ifdef HWPMC_DEBUG LIST_FOREACH(pt, &pm->pm_targets, pt_next) if (pt->pt_process == pp) KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets", __LINE__, pp, pm)); #endif pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO); pt->pt_process = pp; LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next); atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc, (uintptr_t)pm); if (pm->pm_owner->po_owner == pp->pp_proc) pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER; /* * Initialize the per-process values at this row index. */ pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ? pm->pm_sc.pm_reloadcount : 0; pp->pp_refcnt++; #ifdef INVARIANTS /* Confirm that the per-thread values at this row index are cleared. */ if (PMC_TO_MODE(pm) == PMC_MODE_TS) { mtx_lock_spin(pp->pp_tdslock); LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) { KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0, ("[pmc,%d] pt_pmcval not cleared for pid=%d at " "ri=%d", __LINE__, pp->pp_proc->p_pid, ri)); } mtx_unlock_spin(pp->pp_tdslock); } #endif } /* * Removes the association between a target process and a PMC. */ static void pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp) { int ri; struct proc *p; struct pmc_target *ptgt; struct pmc_thread *pt; sx_assert(&pmc_sx, SX_XLOCKED); KASSERT(pm != NULL && pp != NULL, ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc, ("[pmc,%d] Illegal ref count %d on process record %p", __LINE__, pp->pp_refcnt, (void *) pp)); ri = PMC_TO_ROWINDEX(pm); PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p", pm, ri, pp); KASSERT(pp->pp_pmcs[ri].pp_pmc == pm, ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__, ri, pm, pp->pp_pmcs[ri].pp_pmc)); pp->pp_pmcs[ri].pp_pmc = NULL; pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0; /* Clear the per-thread values at this row index. */ if (PMC_TO_MODE(pm) == PMC_MODE_TS) { mtx_lock_spin(pp->pp_tdslock); LIST_FOREACH(pt, &pp->pp_tds, pt_next) pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t) 0; mtx_unlock_spin(pp->pp_tdslock); } /* Remove owner-specific flags */ if (pm->pm_owner->po_owner == pp->pp_proc) { pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS; pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER; } pp->pp_refcnt--; /* Remove the target process from the PMC structure */ LIST_FOREACH(ptgt, &pm->pm_targets, pt_next) if (ptgt->pt_process == pp) break; KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found " "in pmc %p", __LINE__, pp->pp_proc, pp, pm)); LIST_REMOVE(ptgt, pt_next); free(ptgt, M_PMC); /* if the PMC now lacks targets, send the owner a SIGIO */ if (LIST_EMPTY(&pm->pm_targets)) { p = pm->pm_owner->po_owner; PROC_LOCK(p); kern_psignal(p, SIGIO); PROC_UNLOCK(p); PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p, SIGIO); } } /* * Check if PMC 'pm' may be attached to target process 't'. */ static int pmc_can_attach(struct pmc *pm, struct proc *t) { struct proc *o; /* pmc owner */ struct ucred *oc, *tc; /* owner, target credentials */ int decline_attach, i; /* * A PMC's owner can always attach that PMC to itself. */ if ((o = pm->pm_owner->po_owner) == t) return 0; PROC_LOCK(o); oc = o->p_ucred; crhold(oc); PROC_UNLOCK(o); PROC_LOCK(t); tc = t->p_ucred; crhold(tc); PROC_UNLOCK(t); /* * The effective uid of the PMC owner should match at least one * of the {effective,real,saved} uids of the target process. */ decline_attach = oc->cr_uid != tc->cr_uid && oc->cr_uid != tc->cr_svuid && oc->cr_uid != tc->cr_ruid; /* * Every one of the target's group ids, must be in the owner's * group list. */ for (i = 0; !decline_attach && i < tc->cr_ngroups; i++) decline_attach = !groupmember(tc->cr_groups[i], oc); /* check the read and saved gids too */ if (decline_attach == 0) decline_attach = !groupmember(tc->cr_rgid, oc) || !groupmember(tc->cr_svgid, oc); crfree(tc); crfree(oc); return !decline_attach; } /* * Attach a process to a PMC. */ static int pmc_attach_one_process(struct proc *p, struct pmc *pm) { int ri, error; char *fullpath, *freepath; struct pmc_process *pp; sx_assert(&pmc_sx, SX_XLOCKED); PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm, PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); /* * Locate the process descriptor corresponding to process 'p', * allocating space as needed. * * Verify that rowindex 'pm_rowindex' is free in the process * descriptor. * * If not, allocate space for a descriptor and link the * process descriptor and PMC. */ ri = PMC_TO_ROWINDEX(pm); /* mark process as using HWPMCs */ PROC_LOCK(p); p->p_flag |= P_HWPMC; PROC_UNLOCK(p); if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) { error = ENOMEM; goto fail; } if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */ error = EEXIST; goto fail; } if (pp->pp_pmcs[ri].pp_pmc != NULL) { error = EBUSY; goto fail; } pmc_link_target_process(pm, pp); if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) && (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0) pm->pm_flags |= PMC_F_NEEDS_LOGFILE; pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */ /* issue an attach event to a configured log file */ if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) { if (p->p_flag & P_KPROC) { fullpath = kernelname; freepath = NULL; } else { pmc_getfilename(p->p_textvp, &fullpath, &freepath); pmclog_process_pmcattach(pm, p->p_pid, fullpath); } free(freepath, M_TEMP); if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) pmc_log_process_mappings(pm->pm_owner, p); } return (0); fail: PROC_LOCK(p); p->p_flag &= ~P_HWPMC; PROC_UNLOCK(p); return (error); } /* * Attach a process and optionally its children */ static int pmc_attach_process(struct proc *p, struct pmc *pm) { int error; struct proc *top; sx_assert(&pmc_sx, SX_XLOCKED); PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm, PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); /* * If this PMC successfully allowed a GETMSR operation * in the past, disallow further ATTACHes. */ if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0) return EPERM; if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) return pmc_attach_one_process(p, pm); /* * Traverse all child processes, attaching them to * this PMC. */ sx_slock(&proctree_lock); top = p; for (;;) { if ((error = pmc_attach_one_process(p, pm)) != 0) break; if (!LIST_EMPTY(&p->p_children)) p = LIST_FIRST(&p->p_children); else for (;;) { if (p == top) goto done; if (LIST_NEXT(p, p_sibling)) { p = LIST_NEXT(p, p_sibling); break; } p = p->p_pptr; } } if (error) (void) pmc_detach_process(top, pm); done: sx_sunlock(&proctree_lock); return error; } /* * Detach a process from a PMC. If there are no other PMCs tracking * this process, remove the process structure from its hash table. If * 'flags' contains PMC_FLAG_REMOVE, then free the process structure. */ static int pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags) { int ri; struct pmc_process *pp; sx_assert(&pmc_sx, SX_XLOCKED); KASSERT(pm != NULL, ("[pmc,%d] null pm pointer", __LINE__)); ri = PMC_TO_ROWINDEX(pm); PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x", pm, ri, p, p->p_pid, p->p_comm, flags); if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) return ESRCH; if (pp->pp_pmcs[ri].pp_pmc != pm) return EINVAL; pmc_unlink_target_process(pm, pp); /* Issue a detach entry if a log file is configured */ if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_pmcdetach(pm, p->p_pid); /* * If there are no PMCs targeting this process, we remove its * descriptor from the target hash table and unset the P_HWPMC * flag in the struct proc. */ KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, ("[pmc,%d] Illegal refcnt %d for process struct %p", __LINE__, pp->pp_refcnt, pp)); if (pp->pp_refcnt != 0) /* still a target of some PMC */ return 0; pmc_remove_process_descriptor(pp); if (flags & PMC_FLAG_REMOVE) pmc_destroy_process_descriptor(pp); PROC_LOCK(p); p->p_flag &= ~P_HWPMC; PROC_UNLOCK(p); return 0; } /* * Detach a process and optionally its descendants from a PMC. */ static int pmc_detach_process(struct proc *p, struct pmc *pm) { struct proc *top; sx_assert(&pmc_sx, SX_XLOCKED); PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm, PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); /* * Traverse all children, detaching them from this PMC. We * ignore errors since we could be detaching a PMC from a * partially attached proc tree. */ sx_slock(&proctree_lock); top = p; for (;;) { (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); if (!LIST_EMPTY(&p->p_children)) p = LIST_FIRST(&p->p_children); else for (;;) { if (p == top) goto done; if (LIST_NEXT(p, p_sibling)) { p = LIST_NEXT(p, p_sibling); break; } p = p->p_pptr; } } done: sx_sunlock(&proctree_lock); if (LIST_EMPTY(&pm->pm_targets)) pm->pm_flags &= ~PMC_F_ATTACH_DONE; return 0; } /* * Thread context switch IN */ static void pmc_process_csw_in(struct thread *td) { int cpu; unsigned int adjri, ri; struct pmc *pm; struct proc *p; struct pmc_cpu *pc; struct pmc_hw *phw; pmc_value_t newvalue; struct pmc_process *pp; struct pmc_thread *pt; struct pmc_classdep *pcd; p = td->td_proc; pt = NULL; if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL) return; KASSERT(pp->pp_proc == td->td_proc, ("[pmc,%d] not my thread state", __LINE__)); critical_enter(); /* no preemption from this point */ cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, p->p_pid, p->p_comm, pp); KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[pmc,%d] weird CPU id %d", __LINE__, cpu)); pc = pmc_pcpu[cpu]; for (ri = 0; ri < md->pmd_npmc; ri++) { if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) continue; KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), ("[pmc,%d] Target PMC in non-virtual mode (%d)", __LINE__, PMC_TO_MODE(pm))); KASSERT(PMC_TO_ROWINDEX(pm) == ri, ("[pmc,%d] Row index mismatch pmc %d != ri %d", __LINE__, PMC_TO_ROWINDEX(pm), ri)); /* * Only PMCs that are marked as 'RUNNING' need * be placed on hardware. */ if (pm->pm_state != PMC_STATE_RUNNING) continue; KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm, (unsigned long)counter_u64_fetch(pm->pm_runcount))); /* increment PMC runcount */ counter_u64_add(pm->pm_runcount, 1); /* configure the HWPMC we are going to use. */ pcd = pmc_ri_to_classdep(md, ri, &adjri); pcd->pcd_config_pmc(cpu, adjri, pm); phw = pc->pc_hwpmcs[ri]; KASSERT(phw != NULL, ("[pmc,%d] null hw pointer", __LINE__)); KASSERT(phw->phw_pmc == pm, ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__, phw->phw_pmc, pm)); /* * Write out saved value and start the PMC. * * Sampling PMCs use a per-thread value, while * counting mode PMCs use a per-pmc value that is * inherited across descendants. */ if (PMC_TO_MODE(pm) == PMC_MODE_TS) { if (pt == NULL) pt = pmc_find_thread_descriptor(pp, td, PMC_FLAG_NONE); KASSERT(pt != NULL, ("[pmc,%d] No thread found for td=%p", __LINE__, td)); mtx_pool_lock_spin(pmc_mtxpool, pm); /* * If we have a thread descriptor, use the per-thread * counter in the descriptor. If not, we will use * a per-process counter. * * TODO: Remove the per-process "safety net" once * we have thoroughly tested that we don't hit the * above assert. */ if (pt != NULL) { if (pt->pt_pmcs[ri].pt_pmcval > 0) newvalue = pt->pt_pmcs[ri].pt_pmcval; else newvalue = pm->pm_sc.pm_reloadcount; } else { /* * Use the saved value calculated after the most * recent time a thread using the shared counter * switched out. Reset the saved count in case * another thread from this process switches in * before any threads switch out. */ newvalue = pp->pp_pmcs[ri].pp_pmcval; pp->pp_pmcs[ri].pp_pmcval = pm->pm_sc.pm_reloadcount; } mtx_pool_unlock_spin(pmc_mtxpool, pm); KASSERT(newvalue > 0 && newvalue <= pm->pm_sc.pm_reloadcount, ("[pmc,%d] pmcval outside of expected range cpu=%d " "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__, cpu, ri, newvalue, pm->pm_sc.pm_reloadcount)); } else { KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC, ("[pmc,%d] illegal mode=%d", __LINE__, PMC_TO_MODE(pm))); mtx_pool_lock_spin(pmc_mtxpool, pm); newvalue = PMC_PCPU_SAVED(cpu, ri) = pm->pm_gv.pm_savedvalue; mtx_pool_unlock_spin(pmc_mtxpool, pm); } PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue); pcd->pcd_write_pmc(cpu, adjri, newvalue); /* If a sampling mode PMC, reset stalled state. */ if (PMC_TO_MODE(pm) == PMC_MODE_TS) pm->pm_pcpu_state[cpu].pps_stalled = 0; /* Indicate that we desire this to run. */ pm->pm_pcpu_state[cpu].pps_cpustate = 1; /* Start the PMC. */ pcd->pcd_start_pmc(cpu, adjri); } /* * perform any other architecture/cpu dependent thread * switch-in actions. */ (void) (*md->pmd_switch_in)(pc, pp); critical_exit(); } /* * Thread context switch OUT. */ static void pmc_process_csw_out(struct thread *td) { int cpu; int64_t tmp; struct pmc *pm; struct proc *p; enum pmc_mode mode; struct pmc_cpu *pc; pmc_value_t newvalue; unsigned int adjri, ri; struct pmc_process *pp; struct pmc_thread *pt = NULL; struct pmc_classdep *pcd; /* * Locate our process descriptor; this may be NULL if * this process is exiting and we have already removed * the process from the target process table. * * Note that due to kernel preemption, multiple * context switches may happen while the process is * exiting. * * Note also that if the target process cannot be * found we still need to deconfigure any PMCs that * are currently running on hardware. */ p = td->td_proc; pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE); /* * save PMCs */ critical_enter(); cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, p->p_pid, p->p_comm, pp); KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[pmc,%d weird CPU id %d", __LINE__, cpu)); pc = pmc_pcpu[cpu]; /* * When a PMC gets unlinked from a target PMC, it will * be removed from the target's pp_pmc[] array. * * However, on a MP system, the target could have been * executing on another CPU at the time of the unlink. * So, at context switch OUT time, we need to look at * the hardware to determine if a PMC is scheduled on * it. */ for (ri = 0; ri < md->pmd_npmc; ri++) { pcd = pmc_ri_to_classdep(md, ri, &adjri); pm = NULL; (void) (*pcd->pcd_get_config)(cpu, adjri, &pm); if (pm == NULL) /* nothing at this row index */ continue; mode = PMC_TO_MODE(pm); if (!PMC_IS_VIRTUAL_MODE(mode)) continue; /* not a process virtual PMC */ KASSERT(PMC_TO_ROWINDEX(pm) == ri, ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", __LINE__, PMC_TO_ROWINDEX(pm), ri)); /* * Change desired state, and then stop if not stalled. * This two-step dance should avoid race conditions where * an interrupt re-enables the PMC after this code has * already checked the pm_stalled flag. */ pm->pm_pcpu_state[cpu].pps_cpustate = 0; if (pm->pm_pcpu_state[cpu].pps_stalled == 0) pcd->pcd_stop_pmc(cpu, adjri); KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm, (unsigned long)counter_u64_fetch(pm->pm_runcount))); /* reduce this PMC's runcount */ counter_u64_add(pm->pm_runcount, -1); /* * If this PMC is associated with this process, * save the reading. */ if (pm->pm_state != PMC_STATE_DELETED && pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) { KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc)); KASSERT(pp->pp_refcnt > 0, ("[pmc,%d] pp refcnt = %d", __LINE__, pp->pp_refcnt)); pcd->pcd_read_pmc(cpu, adjri, &newvalue); if (mode == PMC_MODE_TS) { PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)", cpu, ri, newvalue); if (pt == NULL) pt = pmc_find_thread_descriptor(pp, td, PMC_FLAG_NONE); KASSERT(pt != NULL, ("[pmc,%d] No thread found for td=%p", __LINE__, td)); mtx_pool_lock_spin(pmc_mtxpool, pm); /* * If we have a thread descriptor, save the * per-thread counter in the descriptor. If not, * we will update the per-process counter. * * TODO: Remove the per-process "safety net" * once we have thoroughly tested that we * don't hit the above assert. */ if (pt != NULL) pt->pt_pmcs[ri].pt_pmcval = newvalue; else { /* * For sampling process-virtual PMCs, * newvalue is the number of events to * be seen until the next sampling * interrupt. We can just add the events * left from this invocation to the * counter, then adjust in case we * overflow our range. * * (Recall that we reload the counter * every time we use it.) */ pp->pp_pmcs[ri].pp_pmcval += newvalue; if (pp->pp_pmcs[ri].pp_pmcval > pm->pm_sc.pm_reloadcount) pp->pp_pmcs[ri].pp_pmcval -= pm->pm_sc.pm_reloadcount; } mtx_pool_unlock_spin(pmc_mtxpool, pm); } else { tmp = newvalue - PMC_PCPU_SAVED(cpu,ri); PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)", cpu, ri, tmp); /* * For counting process-virtual PMCs, * we expect the count to be * increasing monotonically, modulo a 64 * bit wraparound. */ KASSERT(tmp >= 0, ("[pmc,%d] negative increment cpu=%d " "ri=%d newvalue=%jx saved=%jx " "incr=%jx", __LINE__, cpu, ri, newvalue, PMC_PCPU_SAVED(cpu,ri), tmp)); mtx_pool_lock_spin(pmc_mtxpool, pm); pm->pm_gv.pm_savedvalue += tmp; pp->pp_pmcs[ri].pp_pmcval += tmp; mtx_pool_unlock_spin(pmc_mtxpool, pm); if (pm->pm_flags & PMC_F_LOG_PROCCSW) pmclog_process_proccsw(pm, pp, tmp, td); } } /* mark hardware as free */ pcd->pcd_config_pmc(cpu, adjri, NULL); } /* * perform any other architecture/cpu dependent thread * switch out functions. */ (void) (*md->pmd_switch_out)(pc, pp); critical_exit(); } /* * A new thread for a process. */ static void pmc_process_thread_add(struct thread *td) { struct pmc_process *pmc; pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); if (pmc != NULL) pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE); } /* * A thread delete for a process. */ static void pmc_process_thread_delete(struct thread *td) { struct pmc_process *pmc; pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); if (pmc != NULL) pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc, td, PMC_FLAG_REMOVE)); } /* * A userret() call for a thread. */ static void pmc_process_thread_userret(struct thread *td) { sched_pin(); pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame); sched_unpin(); } /* * A mapping change for a process. */ static void pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm) { int ri; pid_t pid; char *fullpath, *freepath; const struct pmc *pm; struct pmc_owner *po; const struct pmc_process *pp; freepath = fullpath = NULL; MPASS(!in_epoch(global_epoch_preempt)); pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath); pid = td->td_proc->p_pid; PMC_EPOCH_ENTER(); /* Inform owners of all system-wide sampling PMCs. */ CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_map_in(po, pid, pkm->pm_address, fullpath); if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) goto done; /* * Inform sampling PMC owners tracking this process. */ for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) pmclog_process_map_in(pm->pm_owner, pid, pkm->pm_address, fullpath); done: if (freepath) free(freepath, M_TEMP); PMC_EPOCH_EXIT(); } /* * Log an munmap request. */ static void pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm) { int ri; pid_t pid; struct pmc_owner *po; const struct pmc *pm; const struct pmc_process *pp; pid = td->td_proc->p_pid; PMC_EPOCH_ENTER(); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_map_out(po, pid, pkm->pm_address, pkm->pm_address + pkm->pm_size); PMC_EPOCH_EXIT(); if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) return; for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) pmclog_process_map_out(pm->pm_owner, pid, pkm->pm_address, pkm->pm_address + pkm->pm_size); } /* * Log mapping information about the kernel. */ static void pmc_log_kernel_mappings(struct pmc *pm) { struct pmc_owner *po; struct pmckern_map_in *km, *kmbase; MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx)); KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), ("[pmc,%d] non-sampling PMC (%p) desires mapping information", __LINE__, (void *) pm)); po = pm->pm_owner; if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE) return; if (PMC_TO_MODE(pm) == PMC_MODE_SS) pmc_process_allproc(pm); /* * Log the current set of kernel modules. */ kmbase = linker_hwpmc_list_objects(); for (km = kmbase; km->pm_file != NULL; km++) { PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file, (void *) km->pm_address); pmclog_process_map_in(po, (pid_t) -1, km->pm_address, km->pm_file); } free(kmbase, M_LINKER); po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE; } /* * Log the mappings for a single process. */ static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p) { vm_map_t map; struct vnode *vp; struct vmspace *vm; vm_map_entry_t entry; vm_offset_t last_end; u_int last_timestamp; struct vnode *last_vp; vm_offset_t start_addr; vm_object_t obj, lobj, tobj; char *fullpath, *freepath; last_vp = NULL; last_end = (vm_offset_t) 0; fullpath = freepath = NULL; if ((vm = vmspace_acquire_ref(p)) == NULL) return; map = &vm->vm_map; vm_map_lock_read(map); VM_MAP_ENTRY_FOREACH(entry, map) { if (entry == NULL) { PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly " "NULL! pid=%d vm_map=%p\n", p->p_pid, map); break; } /* * We only care about executable map entries. */ if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) || !(entry->protection & VM_PROT_EXECUTE) || (entry->object.vm_object == NULL)) { continue; } obj = entry->object.vm_object; VM_OBJECT_RLOCK(obj); /* * Walk the backing_object list to find the base * (non-shadowed) vm_object. */ for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) { if (tobj != obj) VM_OBJECT_RLOCK(tobj); if (lobj != obj) VM_OBJECT_RUNLOCK(lobj); lobj = tobj; } /* * At this point lobj is the base vm_object and it is locked. */ if (lobj == NULL) { PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d " "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj); VM_OBJECT_RUNLOCK(obj); continue; } vp = vm_object_vnode(lobj); if (vp == NULL) { if (lobj != obj) VM_OBJECT_RUNLOCK(lobj); VM_OBJECT_RUNLOCK(obj); continue; } /* * Skip contiguous regions that point to the same * vnode, so we don't emit redundant MAP-IN * directives. */ if (entry->start == last_end && vp == last_vp) { last_end = entry->end; if (lobj != obj) VM_OBJECT_RUNLOCK(lobj); VM_OBJECT_RUNLOCK(obj); continue; } /* * We don't want to keep the proc's vm_map or this * vm_object locked while we walk the pathname, since * vn_fullpath() can sleep. However, if we drop the * lock, it's possible for concurrent activity to * modify the vm_map list. To protect against this, * we save the vm_map timestamp before we release the * lock, and check it after we reacquire the lock * below. */ start_addr = entry->start; last_end = entry->end; last_timestamp = map->timestamp; vm_map_unlock_read(map); vref(vp); if (lobj != obj) VM_OBJECT_RUNLOCK(lobj); VM_OBJECT_RUNLOCK(obj); freepath = NULL; pmc_getfilename(vp, &fullpath, &freepath); last_vp = vp; vrele(vp); vp = NULL; pmclog_process_map_in(po, p->p_pid, start_addr, fullpath); if (freepath) free(freepath, M_TEMP); vm_map_lock_read(map); /* * If our saved timestamp doesn't match, this means * that the vm_map was modified out from under us and * we can't trust our current "entry" pointer. Do a * new lookup for this entry. If there is no entry * for this address range, vm_map_lookup_entry() will * return the previous one, so we always want to go to * the next entry on the next loop iteration. * * There is an edge condition here that can occur if * there is no entry at or before this address. In * this situation, vm_map_lookup_entry returns * &map->header, which would cause our loop to abort * without processing the rest of the map. However, * in practice this will never happen for process * vm_map. This is because the executable's text * segment is the first mapping in the proc's address * space, and this mapping is never removed until the * process exits, so there will always be a non-header * entry at or before the requested address for * vm_map_lookup_entry to return. */ if (map->timestamp != last_timestamp) vm_map_lookup_entry(map, last_end - 1, &entry); } vm_map_unlock_read(map); vmspace_free(vm); return; } /* * Log mappings for all processes in the system. */ static void pmc_log_all_process_mappings(struct pmc_owner *po) { struct proc *p, *top; sx_assert(&pmc_sx, SX_XLOCKED); if ((p = pfind(1)) == NULL) panic("[pmc,%d] Cannot find init", __LINE__); PROC_UNLOCK(p); sx_slock(&proctree_lock); top = p; for (;;) { pmc_log_process_mappings(po, p); if (!LIST_EMPTY(&p->p_children)) p = LIST_FIRST(&p->p_children); else for (;;) { if (p == top) goto done; if (LIST_NEXT(p, p_sibling)) { p = LIST_NEXT(p, p_sibling); break; } p = p->p_pptr; } } done: sx_sunlock(&proctree_lock); } /* * The 'hook' invoked from the kernel proper */ #ifdef HWPMC_DEBUG const char *pmc_hooknames[] = { /* these strings correspond to PMC_FN_* in */ "", "EXEC", "CSW-IN", "CSW-OUT", "SAMPLE", "UNUSED1", "UNUSED2", "MMAP", "MUNMAP", "CALLCHAIN-NMI", "CALLCHAIN-SOFT", "SOFTSAMPLING", "THR-CREATE", "THR-EXIT", "THR-USERRET", "THR-CREATE-LOG", "THR-EXIT-LOG", "PROC-CREATE-LOG" }; #endif static int pmc_hook_handler(struct thread *td, int function, void *arg) { int cpu; PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function, pmc_hooknames[function], arg); switch (function) { /* * Process exec() */ case PMC_FN_PROCESS_EXEC: { char *fullpath, *freepath; unsigned int ri; int is_using_hwpmcs; struct pmc *pm; struct proc *p; struct pmc_owner *po; struct pmc_process *pp; struct pmckern_procexec *pk; sx_assert(&pmc_sx, SX_XLOCKED); p = td->td_proc; pmc_getfilename(p->p_textvp, &fullpath, &freepath); pk = (struct pmckern_procexec *) arg; PMC_EPOCH_ENTER(); /* Inform owners of SS mode PMCs of the exec event. */ CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_procexec(po, PMC_ID_INVALID, p->p_pid, pk->pm_entryaddr, fullpath); PMC_EPOCH_EXIT(); PROC_LOCK(p); is_using_hwpmcs = p->p_flag & P_HWPMC; PROC_UNLOCK(p); if (!is_using_hwpmcs) { if (freepath) free(freepath, M_TEMP); break; } /* * PMCs are not inherited across an exec(): remove any * PMCs that this process is the owner of. */ if ((po = pmc_find_owner_descriptor(p)) != NULL) { pmc_remove_owner(po); pmc_destroy_owner_descriptor(po); } /* * If the process being exec'ed is not the target of any * PMC, we are done. */ if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) { if (freepath) free(freepath, M_TEMP); break; } /* * Log the exec event to all monitoring owners. Skip * owners who have already received the event because * they had system sampling PMCs active. */ for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { po = pm->pm_owner; if (po->po_sscount == 0 && po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_procexec(po, pm->pm_id, p->p_pid, pk->pm_entryaddr, fullpath); } if (freepath) free(freepath, M_TEMP); PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d", p, p->p_pid, p->p_comm, pk->pm_credentialschanged); if (pk->pm_credentialschanged == 0) /* no change */ break; /* * If the newly exec()'ed process has a different credential * than before, allow it to be the target of a PMC only if * the PMC's owner has sufficient privilege. */ for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) if (pmc_can_attach(pm, td->td_proc) != 0) pmc_detach_one_process(td->td_proc, pm, PMC_FLAG_NONE); KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__, pp->pp_refcnt, pp)); /* * If this process is no longer the target of any * PMCs, we can remove the process entry and free * up space. */ if (pp->pp_refcnt == 0) { pmc_remove_process_descriptor(pp); pmc_destroy_process_descriptor(pp); break; } } break; case PMC_FN_CSW_IN: pmc_process_csw_in(td); break; case PMC_FN_CSW_OUT: pmc_process_csw_out(td); break; /* * Process accumulated PC samples. * * This function is expected to be called by hardclock() for * each CPU that has accumulated PC samples. * * This function is to be executed on the CPU whose samples * are being processed. */ case PMC_FN_DO_SAMPLES: /* * Clear the cpu specific bit in the CPU mask before * do the rest of the processing. If the NMI handler * gets invoked after the "atomic_clear_int()" call * below but before "pmc_process_samples()" gets * around to processing the interrupt, then we will * come back here at the next hardclock() tick (and * may find nothing to do if "pmc_process_samples()" * had already processed the interrupt). We don't * lose the interrupt sample. */ DPCPU_SET(pmc_sampled, 0); cpu = PCPU_GET(cpuid); pmc_process_samples(cpu, PMC_HR); pmc_process_samples(cpu, PMC_SR); pmc_process_samples(cpu, PMC_UR); break; case PMC_FN_MMAP: pmc_process_mmap(td, (struct pmckern_map_in *) arg); break; case PMC_FN_MUNMAP: MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx)); pmc_process_munmap(td, (struct pmckern_map_out *) arg); break; case PMC_FN_PROC_CREATE_LOG: pmc_process_proccreate((struct proc *)arg); break; case PMC_FN_USER_CALLCHAIN: /* * Record a call chain. */ KASSERT(td == curthread, ("[pmc,%d] td != curthread", __LINE__)); pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR, (struct trapframe *) arg); KASSERT(td->td_pinned == 1, ("[pmc,%d] invalid td_pinned value", __LINE__)); sched_unpin(); /* Can migrate safely now. */ td->td_pflags &= ~TDP_CALLCHAIN; break; case PMC_FN_USER_CALLCHAIN_SOFT: /* * Record a call chain. */ KASSERT(td == curthread, ("[pmc,%d] td != curthread", __LINE__)); cpu = PCPU_GET(cpuid); pmc_capture_user_callchain(cpu, PMC_SR, (struct trapframe *) arg); KASSERT(td->td_pinned == 1, ("[pmc,%d] invalid td_pinned value", __LINE__)); sched_unpin(); /* Can migrate safely now. */ td->td_pflags &= ~TDP_CALLCHAIN; break; case PMC_FN_SOFT_SAMPLING: /* * Call soft PMC sampling intr. */ pmc_soft_intr((struct pmckern_soft *) arg); break; case PMC_FN_THR_CREATE: pmc_process_thread_add(td); pmc_process_threadcreate(td); break; case PMC_FN_THR_CREATE_LOG: pmc_process_threadcreate(td); break; case PMC_FN_THR_EXIT: KASSERT(td == curthread, ("[pmc,%d] td != curthread", __LINE__)); pmc_process_thread_delete(td); pmc_process_threadexit(td); break; case PMC_FN_THR_EXIT_LOG: pmc_process_threadexit(td); break; case PMC_FN_THR_USERRET: KASSERT(td == curthread, ("[pmc,%d] td != curthread", __LINE__)); pmc_process_thread_userret(td); break; default: #ifdef HWPMC_DEBUG KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function)); #endif break; } return 0; } /* * allocate a 'struct pmc_owner' descriptor in the owner hash table. */ static struct pmc_owner * pmc_allocate_owner_descriptor(struct proc *p) { uint32_t hindex; struct pmc_owner *po; struct pmc_ownerhash *poh; hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); poh = &pmc_ownerhash[hindex]; /* allocate space for N pointers and one descriptor struct */ po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO); po->po_owner = p; LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */ TAILQ_INIT(&po->po_logbuffers); mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN); PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p", p, p->p_pid, p->p_comm, po); return po; } static void pmc_destroy_owner_descriptor(struct pmc_owner *po) { PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)", po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); mtx_destroy(&po->po_mtx); free(po, M_PMC); } /* * Allocate a thread descriptor from the free pool. * * NOTE: This *can* return NULL. */ static struct pmc_thread * pmc_thread_descriptor_pool_alloc(void) { struct pmc_thread *pt; mtx_lock_spin(&pmc_threadfreelist_mtx); if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { LIST_REMOVE(pt, pt_next); pmc_threadfreelist_entries--; } mtx_unlock_spin(&pmc_threadfreelist_mtx); return (pt); } /* * Add a thread descriptor to the free pool. We use this instead of free() * to maintain a cache of free entries. Additionally, we can safely call * this function when we cannot call free(), such as in a critical section. * */ static void pmc_thread_descriptor_pool_free(struct pmc_thread *pt) { if (pt == NULL) return; memset(pt, 0, THREADENTRY_SIZE); mtx_lock_spin(&pmc_threadfreelist_mtx); LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next); pmc_threadfreelist_entries++; if (pmc_threadfreelist_entries > pmc_threadfreelist_max) - GROUPTASK_ENQUEUE(&free_gtask); + taskqueue_enqueue(taskqueue_fast, &free_task); mtx_unlock_spin(&pmc_threadfreelist_mtx); } /* - * A callout to manage the free list. + * An asynchronous task to manage the free list. */ static void -pmc_thread_descriptor_pool_free_task(void *arg __unused) +pmc_thread_descriptor_pool_free_task(void *arg __unused, int pending __unused) { struct pmc_thread *pt; LIST_HEAD(, pmc_thread) tmplist; int delta; LIST_INIT(&tmplist); /* Determine what changes, if any, we need to make. */ mtx_lock_spin(&pmc_threadfreelist_mtx); delta = pmc_threadfreelist_entries - pmc_threadfreelist_max; while (delta > 0 && (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { delta--; pmc_threadfreelist_entries--; LIST_REMOVE(pt, pt_next); LIST_INSERT_HEAD(&tmplist, pt, pt_next); } mtx_unlock_spin(&pmc_threadfreelist_mtx); /* If there are entries to free, free them. */ while (!LIST_EMPTY(&tmplist)) { pt = LIST_FIRST(&tmplist); LIST_REMOVE(pt, pt_next); free(pt, M_PMC); } } /* * Drain the thread free pool, freeing all allocations. */ static void pmc_thread_descriptor_pool_drain() { struct pmc_thread *pt, *next; LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) { LIST_REMOVE(pt, pt_next); free(pt, M_PMC); } } /* * find the descriptor corresponding to thread 'td', adding or removing it * as specified by 'mode'. * * Note that this supports additional mode flags in addition to those * supported by pmc_find_process_descriptor(): * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs. * This makes it safe to call while holding certain other locks. */ static struct pmc_thread * pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td, uint32_t mode) { struct pmc_thread *pt = NULL, *ptnew = NULL; int wait_flag; KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__)); /* * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to * acquiring the lock. */ if (mode & PMC_FLAG_ALLOCATE) { if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) { wait_flag = M_WAITOK; if ((mode & PMC_FLAG_NOWAIT) || in_epoch(global_epoch_preempt)) wait_flag = M_NOWAIT; ptnew = malloc(THREADENTRY_SIZE, M_PMC, wait_flag|M_ZERO); } } mtx_lock_spin(pp->pp_tdslock); LIST_FOREACH(pt, &pp->pp_tds, pt_next) if (pt->pt_td == td) break; if ((mode & PMC_FLAG_REMOVE) && pt != NULL) LIST_REMOVE(pt, pt_next); if ((mode & PMC_FLAG_ALLOCATE) && pt == NULL && ptnew != NULL) { pt = ptnew; ptnew = NULL; pt->pt_td = td; LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next); } mtx_unlock_spin(pp->pp_tdslock); if (ptnew != NULL) { free(ptnew, M_PMC); } return pt; } /* * Try to add thread descriptors for each thread in a process. */ static void pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp) { struct thread *curtd; struct pmc_thread **tdlist; int i, tdcnt, tdlistsz; KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked", __LINE__)); tdcnt = 32; restart: tdlistsz = roundup2(tdcnt, 32); tdcnt = 0; tdlist = malloc(sizeof(struct pmc_thread*) * tdlistsz, M_TEMP, M_WAITOK); PROC_LOCK(p); FOREACH_THREAD_IN_PROC(p, curtd) tdcnt++; if (tdcnt >= tdlistsz) { PROC_UNLOCK(p); free(tdlist, M_TEMP); goto restart; } /* * Try to add each thread to the list without sleeping. If unable, * add to a queue to retry after dropping the process lock. */ tdcnt = 0; FOREACH_THREAD_IN_PROC(p, curtd) { tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd, PMC_FLAG_ALLOCATE|PMC_FLAG_NOWAIT); if (tdlist[tdcnt] == NULL) { PROC_UNLOCK(p); for (i = 0; i <= tdcnt; i++) pmc_thread_descriptor_pool_free(tdlist[i]); free(tdlist, M_TEMP); goto restart; } tdcnt++; } PROC_UNLOCK(p); free(tdlist, M_TEMP); } /* * find the descriptor corresponding to process 'p', adding or removing it * as specified by 'mode'. */ static struct pmc_process * pmc_find_process_descriptor(struct proc *p, uint32_t mode) { uint32_t hindex; struct pmc_process *pp, *ppnew; struct pmc_processhash *pph; hindex = PMC_HASH_PTR(p, pmc_processhashmask); pph = &pmc_processhash[hindex]; ppnew = NULL; /* * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we * cannot call malloc(9) once we hold a spin lock. */ if (mode & PMC_FLAG_ALLOCATE) ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc * sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO); mtx_lock_spin(&pmc_processhash_mtx); LIST_FOREACH(pp, pph, pp_next) if (pp->pp_proc == p) break; if ((mode & PMC_FLAG_REMOVE) && pp != NULL) LIST_REMOVE(pp, pp_next); if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL && ppnew != NULL) { ppnew->pp_proc = p; LIST_INIT(&ppnew->pp_tds); ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew); LIST_INSERT_HEAD(pph, ppnew, pp_next); mtx_unlock_spin(&pmc_processhash_mtx); pp = ppnew; ppnew = NULL; /* Add thread descriptors for this process' current threads. */ pmc_add_thread_descriptors_from_proc(p, pp); } else mtx_unlock_spin(&pmc_processhash_mtx); if (ppnew != NULL) free(ppnew, M_PMC); return pp; } /* * remove a process descriptor from the process hash table. */ static void pmc_remove_process_descriptor(struct pmc_process *pp) { KASSERT(pp->pp_refcnt == 0, ("[pmc,%d] Removing process descriptor %p with count %d", __LINE__, pp, pp->pp_refcnt)); mtx_lock_spin(&pmc_processhash_mtx); LIST_REMOVE(pp, pp_next); mtx_unlock_spin(&pmc_processhash_mtx); } /* * destroy a process descriptor. */ static void pmc_destroy_process_descriptor(struct pmc_process *pp) { struct pmc_thread *pmc_td; while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) { LIST_REMOVE(pmc_td, pt_next); pmc_thread_descriptor_pool_free(pmc_td); } free(pp, M_PMC); } /* * find an owner descriptor corresponding to proc 'p' */ static struct pmc_owner * pmc_find_owner_descriptor(struct proc *p) { uint32_t hindex; struct pmc_owner *po; struct pmc_ownerhash *poh; hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); poh = &pmc_ownerhash[hindex]; po = NULL; LIST_FOREACH(po, poh, po_next) if (po->po_owner == p) break; PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> " "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po); return po; } /* * pmc_allocate_pmc_descriptor * * Allocate a pmc descriptor and initialize its * fields. */ static struct pmc * pmc_allocate_pmc_descriptor(void) { struct pmc *pmc; pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO); pmc->pm_runcount = counter_u64_alloc(M_WAITOK); pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state)*mp_ncpus, M_PMC, M_WAITOK|M_ZERO); PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc); return pmc; } /* * Destroy a pmc descriptor. */ static void pmc_destroy_pmc_descriptor(struct pmc *pm) { KASSERT(pm->pm_state == PMC_STATE_DELETED || pm->pm_state == PMC_STATE_FREE, ("[pmc,%d] destroying non-deleted PMC", __LINE__)); KASSERT(LIST_EMPTY(&pm->pm_targets), ("[pmc,%d] destroying pmc with targets", __LINE__)); KASSERT(pm->pm_owner == NULL, ("[pmc,%d] destroying pmc attached to an owner", __LINE__)); KASSERT(counter_u64_fetch(pm->pm_runcount) == 0, ("[pmc,%d] pmc has non-zero run count %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount))); counter_u64_free(pm->pm_runcount); free(pm->pm_pcpu_state, M_PMC); free(pm, M_PMC); } static void pmc_wait_for_pmc_idle(struct pmc *pm) { #ifdef INVARIANTS volatile int maxloop; maxloop = 100 * pmc_cpu_max(); #endif /* * Loop (with a forced context switch) till the PMC's runcount * comes down to zero. */ pmclog_flush(pm->pm_owner, 1); while (counter_u64_fetch(pm->pm_runcount) > 0) { pmclog_flush(pm->pm_owner, 1); #ifdef INVARIANTS maxloop--; KASSERT(maxloop > 0, ("[pmc,%d] (ri%d, rc%ld) waiting too long for " "pmc to be free", __LINE__, PMC_TO_ROWINDEX(pm), (unsigned long)counter_u64_fetch(pm->pm_runcount))); #endif pmc_force_context_switch(); } } /* * This function does the following things: * * - detaches the PMC from hardware * - unlinks all target threads that were attached to it * - removes the PMC from its owner's list * - destroys the PMC private mutex * * Once this function completes, the given pmc pointer can be freed by * calling pmc_destroy_pmc_descriptor(). */ static void pmc_release_pmc_descriptor(struct pmc *pm) { enum pmc_mode mode; struct pmc_hw *phw; u_int adjri, ri, cpu; struct pmc_owner *po; struct pmc_binding pb; struct pmc_process *pp; struct pmc_classdep *pcd; struct pmc_target *ptgt, *tmp; sx_assert(&pmc_sx, SX_XLOCKED); KASSERT(pm, ("[pmc,%d] null pmc", __LINE__)); ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); mode = PMC_TO_MODE(pm); PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri, mode); /* * First, we take the PMC off hardware. */ cpu = 0; if (PMC_IS_SYSTEM_MODE(mode)) { /* * A system mode PMC runs on a specific CPU. Switch * to this CPU and turn hardware off. */ pmc_save_cpu_binding(&pb); cpu = PMC_TO_CPU(pm); pmc_select_cpu(cpu); /* switch off non-stalled CPUs */ pm->pm_pcpu_state[cpu].pps_cpustate = 0; if (pm->pm_state == PMC_STATE_RUNNING && pm->pm_pcpu_state[cpu].pps_stalled == 0) { phw = pmc_pcpu[cpu]->pc_hwpmcs[ri]; KASSERT(phw->phw_pmc == pm, ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)", __LINE__, ri, phw->phw_pmc, pm)); PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri); critical_enter(); pcd->pcd_stop_pmc(cpu, adjri); critical_exit(); } PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri); critical_enter(); pcd->pcd_config_pmc(cpu, adjri, NULL); critical_exit(); /* adjust the global and process count of SS mode PMCs */ if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) { po = pm->pm_owner; po->po_sscount--; if (po->po_sscount == 0) { atomic_subtract_rel_int(&pmc_ss_count, 1); CK_LIST_REMOVE(po, po_ssnext); epoch_wait_preempt(global_epoch_preempt); } } pm->pm_state = PMC_STATE_DELETED; pmc_restore_cpu_binding(&pb); /* * We could have references to this PMC structure in * the per-cpu sample queues. Wait for the queue to * drain. */ pmc_wait_for_pmc_idle(pm); } else if (PMC_IS_VIRTUAL_MODE(mode)) { /* * A virtual PMC could be running on multiple CPUs at * a given instant. * * By marking its state as DELETED, we ensure that * this PMC is never further scheduled on hardware. * * Then we wait till all CPUs are done with this PMC. */ pm->pm_state = PMC_STATE_DELETED; /* Wait for the PMCs runcount to come to zero. */ pmc_wait_for_pmc_idle(pm); /* * At this point the PMC is off all CPUs and cannot be * freshly scheduled onto a CPU. It is now safe to * unlink all targets from this PMC. If a * process-record's refcount falls to zero, we remove * it from the hash table. The module-wide SX lock * protects us from races. */ LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) { pp = ptgt->pt_process; pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */ PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt); /* * If the target process record shows that no * PMCs are attached to it, reclaim its space. */ if (pp->pp_refcnt == 0) { pmc_remove_process_descriptor(pp); pmc_destroy_process_descriptor(pp); } } cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */ } /* * Release any MD resources */ (void) pcd->pcd_release_pmc(cpu, adjri, pm); /* * Update row disposition */ if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) PMC_UNMARK_ROW_STANDALONE(ri); else PMC_UNMARK_ROW_THREAD(ri); /* unlink from the owner's list */ if (pm->pm_owner) { LIST_REMOVE(pm, pm_next); pm->pm_owner = NULL; } } /* * Register an owner and a pmc. */ static int pmc_register_owner(struct proc *p, struct pmc *pmc) { struct pmc_owner *po; sx_assert(&pmc_sx, SX_XLOCKED); if ((po = pmc_find_owner_descriptor(p)) == NULL) if ((po = pmc_allocate_owner_descriptor(p)) == NULL) return ENOMEM; KASSERT(pmc->pm_owner == NULL, ("[pmc,%d] attempting to own an initialized PMC", __LINE__)); pmc->pm_owner = po; LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next); PROC_LOCK(p); p->p_flag |= P_HWPMC; PROC_UNLOCK(p); if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_pmcallocate(pmc); PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p", po, pmc); return 0; } /* * Return the current row disposition: * == 0 => FREE * > 0 => PROCESS MODE * < 0 => SYSTEM MODE */ int pmc_getrowdisp(int ri) { return pmc_pmcdisp[ri]; } /* * Check if a PMC at row index 'ri' can be allocated to the current * process. * * Allocation can fail if: * - the current process is already being profiled by a PMC at index 'ri', * attached to it via OP_PMCATTACH. * - the current process has already allocated a PMC at index 'ri' * via OP_ALLOCATE. */ static int pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu) { enum pmc_mode mode; struct pmc *pm; struct pmc_owner *po; struct pmc_process *pp; PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d " "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu); /* * We shouldn't have already allocated a process-mode PMC at * row index 'ri'. * * We shouldn't have allocated a system-wide PMC on the same * CPU and same RI. */ if ((po = pmc_find_owner_descriptor(p)) != NULL) LIST_FOREACH(pm, &po->po_pmcs, pm_next) { if (PMC_TO_ROWINDEX(pm) == ri) { mode = PMC_TO_MODE(pm); if (PMC_IS_VIRTUAL_MODE(mode)) return EEXIST; if (PMC_IS_SYSTEM_MODE(mode) && (int) PMC_TO_CPU(pm) == cpu) return EEXIST; } } /* * We also shouldn't be the target of any PMC at this index * since otherwise a PMC_ATTACH to ourselves will fail. */ if ((pp = pmc_find_process_descriptor(p, 0)) != NULL) if (pp->pp_pmcs[ri].pp_pmc) return EEXIST; PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok", p, p->p_pid, p->p_comm, ri); return 0; } /* * Check if a given PMC at row index 'ri' can be currently used in * mode 'mode'. */ static int pmc_can_allocate_row(int ri, enum pmc_mode mode) { enum pmc_disp disp; sx_assert(&pmc_sx, SX_XLOCKED); PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode); if (PMC_IS_SYSTEM_MODE(mode)) disp = PMC_DISP_STANDALONE; else disp = PMC_DISP_THREAD; /* * check disposition for PMC row 'ri': * * Expected disposition Row-disposition Result * * STANDALONE STANDALONE or FREE proceed * STANDALONE THREAD fail * THREAD THREAD or FREE proceed * THREAD STANDALONE fail */ if (!PMC_ROW_DISP_IS_FREE(ri) && !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) && !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri))) return EBUSY; /* * All OK */ PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode); return 0; } /* * Find a PMC descriptor with user handle 'pmcid' for thread 'td'. */ static struct pmc * pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid) { struct pmc *pm; KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc, ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__, PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc)); LIST_FOREACH(pm, &po->po_pmcs, pm_next) if (pm->pm_id == pmcid) return pm; return NULL; } static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc) { struct pmc *pm, *opm; struct pmc_owner *po; struct pmc_process *pp; PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid); if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc) return (EINVAL); if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) { /* * In case of PMC_F_DESCENDANTS child processes we will not find * the current process in the owners hash list. Find the owner * process first and from there lookup the po. */ if ((pp = pmc_find_process_descriptor(curthread->td_proc, PMC_FLAG_NONE)) == NULL) { return ESRCH; } else { opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc; if (opm == NULL) return ESRCH; if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER| PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER| PMC_F_DESCENDANTS)) return ESRCH; po = opm->pm_owner; } } if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL) return EINVAL; PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm); *pmc = pm; return 0; } /* * Start a PMC. */ static int pmc_start(struct pmc *pm) { enum pmc_mode mode; struct pmc_owner *po; struct pmc_binding pb; struct pmc_classdep *pcd; int adjri, error, cpu, ri; KASSERT(pm != NULL, ("[pmc,%d] null pm", __LINE__)); mode = PMC_TO_MODE(pm); ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); error = 0; PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri); po = pm->pm_owner; /* * Disallow PMCSTART if a logfile is required but has not been * configured yet. */ if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) && (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) return (EDOOFUS); /* programming error */ /* * If this is a sampling mode PMC, log mapping information for * the kernel modules that are currently loaded. */ if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) pmc_log_kernel_mappings(pm); if (PMC_IS_VIRTUAL_MODE(mode)) { /* * If a PMCATTACH has never been done on this PMC, * attach it to its owner process. */ if (LIST_EMPTY(&pm->pm_targets)) error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH : pmc_attach_process(po->po_owner, pm); /* * If the PMC is attached to its owner, then force a context * switch to ensure that the MD state gets set correctly. */ if (error == 0) { pm->pm_state = PMC_STATE_RUNNING; if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) pmc_force_context_switch(); } return (error); } /* * A system-wide PMC. * * Add the owner to the global list if this is a system-wide * sampling PMC. */ if (mode == PMC_MODE_SS) { /* * Log mapping information for all existing processes in the * system. Subsequent mappings are logged as they happen; * see pmc_process_mmap(). */ if (po->po_logprocmaps == 0) { pmc_log_all_process_mappings(po); po->po_logprocmaps = 1; } po->po_sscount++; if (po->po_sscount == 1) { atomic_add_rel_int(&pmc_ss_count, 1); CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext); PMCDBG1(PMC,OPS,1, "po=%p in global list", po); } } /* * Move to the CPU associated with this * PMC, and start the hardware. */ pmc_save_cpu_binding(&pb); cpu = PMC_TO_CPU(pm); if (!pmc_cpu_is_active(cpu)) return (ENXIO); pmc_select_cpu(cpu); /* * global PMCs are configured at allocation time * so write out the initial value and start the PMC. */ pm->pm_state = PMC_STATE_RUNNING; critical_enter(); if ((error = pcd->pcd_write_pmc(cpu, adjri, PMC_IS_SAMPLING_MODE(mode) ? pm->pm_sc.pm_reloadcount : pm->pm_sc.pm_initial)) == 0) { /* If a sampling mode PMC, reset stalled state. */ if (PMC_IS_SAMPLING_MODE(mode)) pm->pm_pcpu_state[cpu].pps_stalled = 0; /* Indicate that we desire this to run. Start it. */ pm->pm_pcpu_state[cpu].pps_cpustate = 1; error = pcd->pcd_start_pmc(cpu, adjri); } critical_exit(); pmc_restore_cpu_binding(&pb); return (error); } /* * Stop a PMC. */ static int pmc_stop(struct pmc *pm) { struct pmc_owner *po; struct pmc_binding pb; struct pmc_classdep *pcd; int adjri, cpu, error, ri; KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__)); PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm, PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm)); pm->pm_state = PMC_STATE_STOPPED; /* * If the PMC is a virtual mode one, changing the state to * non-RUNNING is enough to ensure that the PMC never gets * scheduled. * * If this PMC is current running on a CPU, then it will * handled correctly at the time its target process is context * switched out. */ if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) return 0; /* * A system-mode PMC. Move to the CPU associated with * this PMC, and stop the hardware. We update the * 'initial count' so that a subsequent PMCSTART will * resume counting from the current hardware count. */ pmc_save_cpu_binding(&pb); cpu = PMC_TO_CPU(pm); KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[pmc,%d] illegal cpu=%d", __LINE__, cpu)); if (!pmc_cpu_is_active(cpu)) return ENXIO; pmc_select_cpu(cpu); ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); pm->pm_pcpu_state[cpu].pps_cpustate = 0; critical_enter(); if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0) error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial); critical_exit(); pmc_restore_cpu_binding(&pb); po = pm->pm_owner; /* remove this owner from the global list of SS PMC owners */ if (PMC_TO_MODE(pm) == PMC_MODE_SS) { po->po_sscount--; if (po->po_sscount == 0) { atomic_subtract_rel_int(&pmc_ss_count, 1); CK_LIST_REMOVE(po, po_ssnext); epoch_wait_preempt(global_epoch_preempt); PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po); } } return (error); } static struct pmc_classdep * pmc_class_to_classdep(enum pmc_class class) { int n; for (n = 0; n < md->pmd_nclass; n++) if (md->pmd_classdep[n].pcd_class == class) return (&md->pmd_classdep[n]); return (NULL); } #if defined(HWPMC_DEBUG) && defined(KTR) static const char *pmc_op_to_name[] = { #undef __PMC_OP #define __PMC_OP(N, D) #N , __PMC_OPS() NULL }; #endif /* * The syscall interface */ #define PMC_GET_SX_XLOCK(...) do { \ sx_xlock(&pmc_sx); \ if (pmc_hook == NULL) { \ sx_xunlock(&pmc_sx); \ return __VA_ARGS__; \ } \ } while (0) #define PMC_DOWNGRADE_SX() do { \ sx_downgrade(&pmc_sx); \ is_sx_downgraded = 1; \ } while (0) static int pmc_syscall_handler(struct thread *td, void *syscall_args) { int error, is_sx_downgraded, op; struct pmc_syscall_args *c; void *pmclog_proc_handle; void *arg; c = (struct pmc_syscall_args *)syscall_args; op = c->pmop_code; arg = c->pmop_data; /* PMC isn't set up yet */ if (pmc_hook == NULL) return (EINVAL); if (op == PMC_OP_CONFIGURELOG) { /* * We cannot create the logging process inside * pmclog_configure_log() because there is a LOR * between pmc_sx and process structure locks. * Instead, pre-create the process and ignite the loop * if everything is fine, otherwise direct the process * to exit. */ error = pmclog_proc_create(td, &pmclog_proc_handle); if (error != 0) goto done_syscall; } PMC_GET_SX_XLOCK(ENOSYS); is_sx_downgraded = 0; PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op, pmc_op_to_name[op], arg); error = 0; counter_u64_add(pmc_stats.pm_syscalls, 1); switch (op) { /* * Configure a log file. * * XXX This OP will be reworked. */ case PMC_OP_CONFIGURELOG: { struct proc *p; struct pmc *pm; struct pmc_owner *po; struct pmc_op_configurelog cl; if ((error = copyin(arg, &cl, sizeof(cl))) != 0) { pmclog_proc_ignite(pmclog_proc_handle, NULL); break; } /* mark this process as owning a log file */ p = td->td_proc; if ((po = pmc_find_owner_descriptor(p)) == NULL) if ((po = pmc_allocate_owner_descriptor(p)) == NULL) { pmclog_proc_ignite(pmclog_proc_handle, NULL); error = ENOMEM; break; } /* * If a valid fd was passed in, try to configure that, * otherwise if 'fd' was less than zero and there was * a log file configured, flush its buffers and * de-configure it. */ if (cl.pm_logfd >= 0) { error = pmclog_configure_log(md, po, cl.pm_logfd); pmclog_proc_ignite(pmclog_proc_handle, error == 0 ? po : NULL); } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) { pmclog_proc_ignite(pmclog_proc_handle, NULL); error = pmclog_close(po); if (error == 0) { LIST_FOREACH(pm, &po->po_pmcs, pm_next) if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && pm->pm_state == PMC_STATE_RUNNING) pmc_stop(pm); error = pmclog_deconfigure_log(po); } } else { pmclog_proc_ignite(pmclog_proc_handle, NULL); error = EINVAL; } } break; /* * Flush a log file. */ case PMC_OP_FLUSHLOG: { struct pmc_owner *po; sx_assert(&pmc_sx, SX_XLOCKED); if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { error = EINVAL; break; } error = pmclog_flush(po, 0); } break; /* * Close a log file. */ case PMC_OP_CLOSELOG: { struct pmc_owner *po; sx_assert(&pmc_sx, SX_XLOCKED); if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { error = EINVAL; break; } error = pmclog_close(po); } break; /* * Retrieve hardware configuration. */ case PMC_OP_GETCPUINFO: /* CPU information */ { struct pmc_op_getcpuinfo gci; struct pmc_classinfo *pci; struct pmc_classdep *pcd; int cl; memset(&gci, 0, sizeof(gci)); gci.pm_cputype = md->pmd_cputype; gci.pm_ncpu = pmc_cpu_max(); gci.pm_npmc = md->pmd_npmc; gci.pm_nclass = md->pmd_nclass; pci = gci.pm_classes; pcd = md->pmd_classdep; for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) { pci->pm_caps = pcd->pcd_caps; pci->pm_class = pcd->pcd_class; pci->pm_width = pcd->pcd_width; pci->pm_num = pcd->pcd_num; } error = copyout(&gci, arg, sizeof(gci)); } break; /* * Retrieve soft events list. */ case PMC_OP_GETDYNEVENTINFO: { enum pmc_class cl; enum pmc_event ev; struct pmc_op_getdyneventinfo *gei; struct pmc_dyn_event_descr dev; struct pmc_soft *ps; uint32_t nevent; sx_assert(&pmc_sx, SX_LOCKED); gei = (struct pmc_op_getdyneventinfo *) arg; if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0) break; /* Only SOFT class is dynamic. */ if (cl != PMC_CLASS_SOFT) { error = EINVAL; break; } nevent = 0; for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) { ps = pmc_soft_ev_acquire(ev); if (ps == NULL) continue; bcopy(&ps->ps_ev, &dev, sizeof(dev)); pmc_soft_ev_release(ps); error = copyout(&dev, &gei->pm_events[nevent], sizeof(struct pmc_dyn_event_descr)); if (error != 0) break; nevent++; } if (error != 0) break; error = copyout(&nevent, &gei->pm_nevent, sizeof(nevent)); } break; /* * Get module statistics */ case PMC_OP_GETDRIVERSTATS: { struct pmc_op_getdriverstats gms; #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field) CFETCH(gms, pmc_stats, pm_intr_ignored); CFETCH(gms, pmc_stats, pm_intr_processed); CFETCH(gms, pmc_stats, pm_intr_bufferfull); CFETCH(gms, pmc_stats, pm_syscalls); CFETCH(gms, pmc_stats, pm_syscall_errors); CFETCH(gms, pmc_stats, pm_buffer_requests); CFETCH(gms, pmc_stats, pm_buffer_requests_failed); CFETCH(gms, pmc_stats, pm_log_sweeps); #undef CFETCH error = copyout(&gms, arg, sizeof(gms)); } break; /* * Retrieve module version number */ case PMC_OP_GETMODULEVERSION: { uint32_t cv, modv; /* retrieve the client's idea of the ABI version */ if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0) break; /* don't service clients newer than our driver */ modv = PMC_VERSION; if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) { error = EPROGMISMATCH; break; } error = copyout(&modv, arg, sizeof(int)); } break; /* * Retrieve the state of all the PMCs on a given * CPU. */ case PMC_OP_GETPMCINFO: { int ari; struct pmc *pm; size_t pmcinfo_size; uint32_t cpu, n, npmc; struct pmc_owner *po; struct pmc_binding pb; struct pmc_classdep *pcd; struct pmc_info *p, *pmcinfo; struct pmc_op_getpmcinfo *gpi; PMC_DOWNGRADE_SX(); gpi = (struct pmc_op_getpmcinfo *) arg; if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0) break; if (cpu >= pmc_cpu_max()) { error = EINVAL; break; } if (!pmc_cpu_is_active(cpu)) { error = ENXIO; break; } /* switch to CPU 'cpu' */ pmc_save_cpu_binding(&pb); pmc_select_cpu(cpu); npmc = md->pmd_npmc; pmcinfo_size = npmc * sizeof(struct pmc_info); pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK | M_ZERO); p = pmcinfo; for (n = 0; n < md->pmd_npmc; n++, p++) { pcd = pmc_ri_to_classdep(md, n, &ari); KASSERT(pcd != NULL, ("[pmc,%d] null pcd ri=%d", __LINE__, n)); if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0) break; if (PMC_ROW_DISP_IS_STANDALONE(n)) p->pm_rowdisp = PMC_DISP_STANDALONE; else if (PMC_ROW_DISP_IS_THREAD(n)) p->pm_rowdisp = PMC_DISP_THREAD; else p->pm_rowdisp = PMC_DISP_FREE; p->pm_ownerpid = -1; if (pm == NULL) /* no PMC associated */ continue; po = pm->pm_owner; KASSERT(po->po_owner != NULL, ("[pmc,%d] pmc_owner had a null proc pointer", __LINE__)); p->pm_ownerpid = po->po_owner->p_pid; p->pm_mode = PMC_TO_MODE(pm); p->pm_event = pm->pm_event; p->pm_flags = pm->pm_flags; if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) p->pm_reloadcount = pm->pm_sc.pm_reloadcount; } pmc_restore_cpu_binding(&pb); /* now copy out the PMC info collected */ if (error == 0) error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size); free(pmcinfo, M_PMC); } break; /* * Set the administrative state of a PMC. I.e. whether * the PMC is to be used or not. */ case PMC_OP_PMCADMIN: { int cpu, ri; enum pmc_state request; struct pmc_cpu *pc; struct pmc_hw *phw; struct pmc_op_pmcadmin pma; struct pmc_binding pb; sx_assert(&pmc_sx, SX_XLOCKED); KASSERT(td == curthread, ("[pmc,%d] td != curthread", __LINE__)); error = priv_check(td, PRIV_PMC_MANAGE); if (error) break; if ((error = copyin(arg, &pma, sizeof(pma))) != 0) break; cpu = pma.pm_cpu; if (cpu < 0 || cpu >= (int) pmc_cpu_max()) { error = EINVAL; break; } if (!pmc_cpu_is_active(cpu)) { error = ENXIO; break; } request = pma.pm_state; if (request != PMC_STATE_DISABLED && request != PMC_STATE_FREE) { error = EINVAL; break; } ri = pma.pm_pmc; /* pmc id == row index */ if (ri < 0 || ri >= (int) md->pmd_npmc) { error = EINVAL; break; } /* * We can't disable a PMC with a row-index allocated * for process virtual PMCs. */ if (PMC_ROW_DISP_IS_THREAD(ri) && request == PMC_STATE_DISABLED) { error = EBUSY; break; } /* * otherwise, this PMC on this CPU is either free or * in system-wide mode. */ pmc_save_cpu_binding(&pb); pmc_select_cpu(cpu); pc = pmc_pcpu[cpu]; phw = pc->pc_hwpmcs[ri]; /* * XXX do we need some kind of 'forced' disable? */ if (phw->phw_pmc == NULL) { if (request == PMC_STATE_DISABLED && (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) { phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED; PMC_MARK_ROW_STANDALONE(ri); } else if (request == PMC_STATE_FREE && (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED; PMC_UNMARK_ROW_STANDALONE(ri); } /* other cases are a no-op */ } else error = EBUSY; pmc_restore_cpu_binding(&pb); } break; /* * Allocate a PMC. */ case PMC_OP_PMCALLOCATE: { int adjri, n; u_int cpu; uint32_t caps; struct pmc *pmc; enum pmc_mode mode; struct pmc_hw *phw; struct pmc_binding pb; struct pmc_classdep *pcd; struct pmc_op_pmcallocate pa; if ((error = copyin(arg, &pa, sizeof(pa))) != 0) break; caps = pa.pm_caps; mode = pa.pm_mode; cpu = pa.pm_cpu; if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC && mode != PMC_MODE_TS && mode != PMC_MODE_TC) || (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) { error = EINVAL; break; } /* * Virtual PMCs should only ask for a default CPU. * System mode PMCs need to specify a non-default CPU. */ if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) || (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) { error = EINVAL; break; } /* * Check that an inactive CPU is not being asked for. */ if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) { error = ENXIO; break; } /* * Refuse an allocation for a system-wide PMC if this * process has been jailed, or if this process lacks * super-user credentials and the sysctl tunable * 'security.bsd.unprivileged_syspmcs' is zero. */ if (PMC_IS_SYSTEM_MODE(mode)) { if (jailed(curthread->td_ucred)) { error = EPERM; break; } if (!pmc_unprivileged_syspmcs) { error = priv_check(curthread, PRIV_PMC_SYSTEM); if (error) break; } } /* * Look for valid values for 'pm_flags' */ if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) != 0) { error = EINVAL; break; } /* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN */ if ((pa.pm_flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) == PMC_F_USERCALLCHAIN) { error = EINVAL; break; } /* PMC_F_USERCALLCHAIN is only valid for sampling mode */ if (pa.pm_flags & PMC_F_USERCALLCHAIN && mode != PMC_MODE_TS && mode != PMC_MODE_SS) { error = EINVAL; break; } /* process logging options are not allowed for system PMCs */ if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags & (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) { error = EINVAL; break; } /* * All sampling mode PMCs need to be able to interrupt the * CPU. */ if (PMC_IS_SAMPLING_MODE(mode)) caps |= PMC_CAP_INTERRUPT; /* A valid class specifier should have been passed in. */ pcd = pmc_class_to_classdep(pa.pm_class); if (pcd == NULL) { error = EINVAL; break; } /* The requested PMC capabilities should be feasible. */ if ((pcd->pcd_caps & caps) != caps) { error = EOPNOTSUPP; break; } PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d", pa.pm_ev, caps, mode, cpu); pmc = pmc_allocate_pmc_descriptor(); pmc->pm_id = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class, PMC_ID_INVALID); pmc->pm_event = pa.pm_ev; pmc->pm_state = PMC_STATE_FREE; pmc->pm_caps = caps; pmc->pm_flags = pa.pm_flags; /* XXX set lower bound on sampling for process counters */ if (PMC_IS_SAMPLING_MODE(mode)) { /* * Don't permit requested sample rate to be less than 1000 */ if (pa.pm_count < 1000) log(LOG_WARNING, "pmcallocate: passed sample rate %ju - setting to 1000\n", (uintmax_t)pa.pm_count); pmc->pm_sc.pm_reloadcount = MAX(1000, pa.pm_count); } else pmc->pm_sc.pm_initial = pa.pm_count; /* switch thread to CPU 'cpu' */ pmc_save_cpu_binding(&pb); #define PMC_IS_SHAREABLE_PMC(cpu, n) \ (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \ PMC_PHW_FLAG_IS_SHAREABLE) #define PMC_IS_UNALLOCATED(cpu, n) \ (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL) if (PMC_IS_SYSTEM_MODE(mode)) { pmc_select_cpu(cpu); for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) { pcd = pmc_ri_to_classdep(md, n, &adjri); if (pmc_can_allocate_row(n, mode) == 0 && pmc_can_allocate_rowindex( curthread->td_proc, n, cpu) == 0 && (PMC_IS_UNALLOCATED(cpu, n) || PMC_IS_SHAREABLE_PMC(cpu, n)) && pcd->pcd_allocate_pmc(cpu, adjri, pmc, &pa) == 0) break; } } else { /* Process virtual mode */ for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) { pcd = pmc_ri_to_classdep(md, n, &adjri); if (pmc_can_allocate_row(n, mode) == 0 && pmc_can_allocate_rowindex( curthread->td_proc, n, PMC_CPU_ANY) == 0 && pcd->pcd_allocate_pmc(curthread->td_oncpu, adjri, pmc, &pa) == 0) break; } } #undef PMC_IS_UNALLOCATED #undef PMC_IS_SHAREABLE_PMC pmc_restore_cpu_binding(&pb); if (n == (int) md->pmd_npmc) { pmc_destroy_pmc_descriptor(pmc); pmc = NULL; error = EINVAL; break; } /* Fill in the correct value in the ID field */ pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n); PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x", pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id); /* Process mode PMCs with logging enabled need log files */ if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW)) pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; /* All system mode sampling PMCs require a log file */ if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode)) pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; /* * Configure global pmc's immediately */ if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) { pmc_save_cpu_binding(&pb); pmc_select_cpu(cpu); phw = pmc_pcpu[cpu]->pc_hwpmcs[n]; pcd = pmc_ri_to_classdep(md, n, &adjri); if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 || (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) { (void) pcd->pcd_release_pmc(cpu, adjri, pmc); pmc_destroy_pmc_descriptor(pmc); pmc = NULL; pmc_restore_cpu_binding(&pb); error = EPERM; break; } pmc_restore_cpu_binding(&pb); } pmc->pm_state = PMC_STATE_ALLOCATED; pmc->pm_class = pa.pm_class; /* * mark row disposition */ if (PMC_IS_SYSTEM_MODE(mode)) PMC_MARK_ROW_STANDALONE(n); else PMC_MARK_ROW_THREAD(n); /* * Register this PMC with the current thread as its owner. */ if ((error = pmc_register_owner(curthread->td_proc, pmc)) != 0) { pmc_release_pmc_descriptor(pmc); pmc_destroy_pmc_descriptor(pmc); pmc = NULL; break; } /* * Return the allocated index. */ pa.pm_pmcid = pmc->pm_id; error = copyout(&pa, arg, sizeof(pa)); } break; /* * Attach a PMC to a process. */ case PMC_OP_PMCATTACH: { struct pmc *pm; struct proc *p; struct pmc_op_pmcattach a; sx_assert(&pmc_sx, SX_XLOCKED); if ((error = copyin(arg, &a, sizeof(a))) != 0) break; if (a.pm_pid < 0) { error = EINVAL; break; } else if (a.pm_pid == 0) a.pm_pid = td->td_proc->p_pid; if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) break; if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { error = EINVAL; break; } /* PMCs may be (re)attached only when allocated or stopped */ if (pm->pm_state == PMC_STATE_RUNNING) { error = EBUSY; break; } else if (pm->pm_state != PMC_STATE_ALLOCATED && pm->pm_state != PMC_STATE_STOPPED) { error = EINVAL; break; } /* lookup pid */ if ((p = pfind(a.pm_pid)) == NULL) { error = ESRCH; break; } /* * Ignore processes that are working on exiting. */ if (p->p_flag & P_WEXIT) { error = ESRCH; PROC_UNLOCK(p); /* pfind() returns a locked process */ break; } /* * we are allowed to attach a PMC to a process if * we can debug it. */ error = p_candebug(curthread, p); PROC_UNLOCK(p); if (error == 0) error = pmc_attach_process(p, pm); } break; /* * Detach an attached PMC from a process. */ case PMC_OP_PMCDETACH: { struct pmc *pm; struct proc *p; struct pmc_op_pmcattach a; if ((error = copyin(arg, &a, sizeof(a))) != 0) break; if (a.pm_pid < 0) { error = EINVAL; break; } else if (a.pm_pid == 0) a.pm_pid = td->td_proc->p_pid; if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) break; if ((p = pfind(a.pm_pid)) == NULL) { error = ESRCH; break; } /* * Treat processes that are in the process of exiting * as if they were not present. */ if (p->p_flag & P_WEXIT) error = ESRCH; PROC_UNLOCK(p); /* pfind() returns a locked process */ if (error == 0) error = pmc_detach_process(p, pm); } break; /* * Retrieve the MSR number associated with the counter * 'pmc_id'. This allows processes to directly use RDPMC * instructions to read their PMCs, without the overhead of a * system call. */ case PMC_OP_PMCGETMSR: { int adjri, ri; struct pmc *pm; struct pmc_target *pt; struct pmc_op_getmsr gm; struct pmc_classdep *pcd; PMC_DOWNGRADE_SX(); if ((error = copyin(arg, &gm, sizeof(gm))) != 0) break; if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0) break; /* * The allocated PMC has to be a process virtual PMC, * i.e., of type MODE_T[CS]. Global PMCs can only be * read using the PMCREAD operation since they may be * allocated on a different CPU than the one we could * be running on at the time of the RDPMC instruction. * * The GETMSR operation is not allowed for PMCs that * are inherited across processes. */ if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) || (pm->pm_flags & PMC_F_DESCENDANTS)) { error = EINVAL; break; } /* * It only makes sense to use a RDPMC (or its * equivalent instruction on non-x86 architectures) on * a process that has allocated and attached a PMC to * itself. Conversely the PMC is only allowed to have * one process attached to it -- its owner. */ if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL || LIST_NEXT(pt, pt_next) != NULL || pt->pt_process->pp_proc != pm->pm_owner->po_owner) { error = EINVAL; break; } ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); /* PMC class has no 'GETMSR' support */ if (pcd->pcd_get_msr == NULL) { error = ENOSYS; break; } if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0) break; if ((error = copyout(&gm, arg, sizeof(gm))) < 0) break; /* * Mark our process as using MSRs. Update machine * state using a forced context switch. */ pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS; pmc_force_context_switch(); } break; /* * Release an allocated PMC */ case PMC_OP_PMCRELEASE: { pmc_id_t pmcid; struct pmc *pm; struct pmc_owner *po; struct pmc_op_simple sp; /* * Find PMC pointer for the named PMC. * * Use pmc_release_pmc_descriptor() to switch off the * PMC, remove all its target threads, and remove the * PMC from its owner's list. * * Remove the owner record if this is the last PMC * owned. * * Free up space. */ if ((error = copyin(arg, &sp, sizeof(sp))) != 0) break; pmcid = sp.pm_pmcid; if ((error = pmc_find_pmc(pmcid, &pm)) != 0) break; po = pm->pm_owner; pmc_release_pmc_descriptor(pm); pmc_maybe_remove_owner(po); pmc_destroy_pmc_descriptor(pm); } break; /* * Read and/or write a PMC. */ case PMC_OP_PMCRW: { int adjri; struct pmc *pm; uint32_t cpu, ri; pmc_value_t oldvalue; struct pmc_binding pb; struct pmc_op_pmcrw prw; struct pmc_classdep *pcd; struct pmc_op_pmcrw *pprw; PMC_DOWNGRADE_SX(); if ((error = copyin(arg, &prw, sizeof(prw))) != 0) break; ri = 0; PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid, prw.pm_flags); /* must have at least one flag set */ if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) { error = EINVAL; break; } /* locate pmc descriptor */ if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0) break; /* Can't read a PMC that hasn't been started. */ if (pm->pm_state != PMC_STATE_ALLOCATED && pm->pm_state != PMC_STATE_STOPPED && pm->pm_state != PMC_STATE_RUNNING) { error = EINVAL; break; } /* writing a new value is allowed only for 'STOPPED' pmcs */ if (pm->pm_state == PMC_STATE_RUNNING && (prw.pm_flags & PMC_F_NEWVALUE)) { error = EBUSY; break; } if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { /* * If this PMC is attached to its owner (i.e., * the process requesting this operation) and * is running, then attempt to get an * upto-date reading from hardware for a READ. * Writes are only allowed when the PMC is * stopped, so only update the saved value * field. * * If the PMC is not running, or is not * attached to its owner, read/write to the * savedvalue field. */ ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); mtx_pool_lock_spin(pmc_mtxpool, pm); cpu = curthread->td_oncpu; if (prw.pm_flags & PMC_F_OLDVALUE) { if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) && (pm->pm_state == PMC_STATE_RUNNING)) error = (*pcd->pcd_read_pmc)(cpu, adjri, &oldvalue); else oldvalue = pm->pm_gv.pm_savedvalue; } if (prw.pm_flags & PMC_F_NEWVALUE) pm->pm_gv.pm_savedvalue = prw.pm_value; mtx_pool_unlock_spin(pmc_mtxpool, pm); } else { /* System mode PMCs */ cpu = PMC_TO_CPU(pm); ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); if (!pmc_cpu_is_active(cpu)) { error = ENXIO; break; } /* move this thread to CPU 'cpu' */ pmc_save_cpu_binding(&pb); pmc_select_cpu(cpu); critical_enter(); /* save old value */ if (prw.pm_flags & PMC_F_OLDVALUE) if ((error = (*pcd->pcd_read_pmc)(cpu, adjri, &oldvalue))) goto error; /* write out new value */ if (prw.pm_flags & PMC_F_NEWVALUE) error = (*pcd->pcd_write_pmc)(cpu, adjri, prw.pm_value); error: critical_exit(); pmc_restore_cpu_binding(&pb); if (error) break; } pprw = (struct pmc_op_pmcrw *) arg; #ifdef HWPMC_DEBUG if (prw.pm_flags & PMC_F_NEWVALUE) PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx", ri, prw.pm_value, oldvalue); else if (prw.pm_flags & PMC_F_OLDVALUE) PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue); #endif /* return old value if requested */ if (prw.pm_flags & PMC_F_OLDVALUE) if ((error = copyout(&oldvalue, &pprw->pm_value, sizeof(prw.pm_value)))) break; } break; /* * Set the sampling rate for a sampling mode PMC and the * initial count for a counting mode PMC. */ case PMC_OP_PMCSETCOUNT: { struct pmc *pm; struct pmc_op_pmcsetcount sc; PMC_DOWNGRADE_SX(); if ((error = copyin(arg, &sc, sizeof(sc))) != 0) break; if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0) break; if (pm->pm_state == PMC_STATE_RUNNING) { error = EBUSY; break; } if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { /* * Don't permit requested sample rate to be less than 1000 */ if (sc.pm_count < 1000) log(LOG_WARNING, "pmcsetcount: passed sample rate %ju - setting to 1000\n", (uintmax_t)sc.pm_count); pm->pm_sc.pm_reloadcount = MAX(1000, sc.pm_count); } else pm->pm_sc.pm_initial = sc.pm_count; } break; /* * Start a PMC. */ case PMC_OP_PMCSTART: { pmc_id_t pmcid; struct pmc *pm; struct pmc_op_simple sp; sx_assert(&pmc_sx, SX_XLOCKED); if ((error = copyin(arg, &sp, sizeof(sp))) != 0) break; pmcid = sp.pm_pmcid; if ((error = pmc_find_pmc(pmcid, &pm)) != 0) break; KASSERT(pmcid == pm->pm_id, ("[pmc,%d] pmcid %x != id %x", __LINE__, pm->pm_id, pmcid)); if (pm->pm_state == PMC_STATE_RUNNING) /* already running */ break; else if (pm->pm_state != PMC_STATE_STOPPED && pm->pm_state != PMC_STATE_ALLOCATED) { error = EINVAL; break; } error = pmc_start(pm); } break; /* * Stop a PMC. */ case PMC_OP_PMCSTOP: { pmc_id_t pmcid; struct pmc *pm; struct pmc_op_simple sp; PMC_DOWNGRADE_SX(); if ((error = copyin(arg, &sp, sizeof(sp))) != 0) break; pmcid = sp.pm_pmcid; /* * Mark the PMC as inactive and invoke the MD stop * routines if needed. */ if ((error = pmc_find_pmc(pmcid, &pm)) != 0) break; KASSERT(pmcid == pm->pm_id, ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, pm->pm_id, pmcid)); if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */ break; else if (pm->pm_state != PMC_STATE_RUNNING) { error = EINVAL; break; } error = pmc_stop(pm); } break; /* * Write a user supplied value to the log file. */ case PMC_OP_WRITELOG: { struct pmc_op_writelog wl; struct pmc_owner *po; PMC_DOWNGRADE_SX(); if ((error = copyin(arg, &wl, sizeof(wl))) != 0) break; if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { error = EINVAL; break; } if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { error = EINVAL; break; } error = pmclog_process_userlog(po, &wl); } break; default: error = EINVAL; break; } if (is_sx_downgraded) sx_sunlock(&pmc_sx); else sx_xunlock(&pmc_sx); done_syscall: if (error) counter_u64_add(pmc_stats.pm_syscall_errors, 1); return (error); } /* * Helper functions */ /* * Mark the thread as needing callchain capture and post an AST. The * actual callchain capture will be done in a context where it is safe * to take page faults. */ static void pmc_post_callchain_callback(void) { struct thread *td; td = curthread; /* * If there is multiple PMCs for the same interrupt ignore new post */ if (td->td_pflags & TDP_CALLCHAIN) return; /* * Mark this thread as needing callchain capture. * `td->td_pflags' will be safe to touch because this thread * was in user space when it was interrupted. */ td->td_pflags |= TDP_CALLCHAIN; /* * Don't let this thread migrate between CPUs until callchain * capture completes. */ sched_pin(); return; } /* * Find a free slot in the per-cpu array of samples and capture the * current callchain there. If a sample was successfully added, a bit * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook * needs to be invoked from the clock handler. * * This function is meant to be called from an NMI handler. It cannot * use any of the locking primitives supplied by the OS. */ static int pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf) { int error, cpu, callchaindepth, inuserspace; struct thread *td; struct pmc_sample *ps; struct pmc_samplebuffer *psb; error = 0; /* * Allocate space for a sample buffer. */ cpu = curcpu; psb = pmc_pcpu[cpu]->pc_sb[ring]; inuserspace = TRAPF_USERMODE(tf); ps = PMC_PROD_SAMPLE(psb); if (psb->ps_considx != psb->ps_prodidx && ps->ps_nsamples) { /* in use, reader hasn't caught up */ pm->pm_pcpu_state[cpu].pps_stalled = 1; counter_u64_add(pmc_stats.pm_intr_bufferfull, 1); PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm, (void *) tf, inuserspace, (int) (psb->ps_prodidx & pmc_sample_mask), (int) (psb->ps_considx & pmc_sample_mask)); callchaindepth = 1; error = ENOMEM; goto done; } /* Fill in entry. */ PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm, (void *) tf, inuserspace, (int) (psb->ps_prodidx & pmc_sample_mask), (int) (psb->ps_considx & pmc_sample_mask)); td = curthread; ps->ps_pmc = pm; ps->ps_td = td; ps->ps_pid = td->td_proc->p_pid; ps->ps_tid = td->td_tid; ps->ps_tsc = pmc_rdtsc(); ps->ps_ticks = ticks; ps->ps_cpu = cpu; ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0; callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ? pmc_callchaindepth : 1; MPASS(ps->ps_pc != NULL); if (callchaindepth == 1) ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf); else { /* * Kernel stack traversals can be done immediately, * while we defer to an AST for user space traversals. */ if (!inuserspace) { callchaindepth = pmc_save_kernel_callchain(ps->ps_pc, callchaindepth, tf); } else { pmc_post_callchain_callback(); callchaindepth = PMC_USER_CALLCHAIN_PENDING; } } ps->ps_nsamples = callchaindepth; /* mark entry as in use */ if (ring == PMC_UR) { ps->ps_nsamples_actual = callchaindepth; /* mark entry as in use */ ps->ps_nsamples = PMC_USER_CALLCHAIN_PENDING; } else ps->ps_nsamples = callchaindepth; /* mark entry as in use */ KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm, (unsigned long)counter_u64_fetch(pm->pm_runcount))); counter_u64_add(pm->pm_runcount, 1); /* hold onto PMC */ /* increment write pointer */ psb->ps_prodidx++; done: /* mark CPU as needing processing */ if (callchaindepth != PMC_USER_CALLCHAIN_PENDING) DPCPU_SET(pmc_sampled, 1); return (error); } /* * Interrupt processing. * * This function is meant to be called from an NMI handler. It cannot * use any of the locking primitives supplied by the OS. */ int pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf) { struct thread *td; td = curthread; if ((pm->pm_flags & PMC_F_USERCALLCHAIN) && (td->td_proc->p_flag & P_KPROC) == 0 && !TRAPF_USERMODE(tf)) { atomic_add_int(&td->td_pmcpend, 1); return (pmc_add_sample(PMC_UR, pm, tf)); } return (pmc_add_sample(ring, pm, tf)); } /* * Capture a user call chain. This function will be called from ast() * before control returns to userland and before the process gets * rescheduled. */ static void pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf) { struct pmc *pm; struct thread *td; struct pmc_sample *ps; struct pmc_samplebuffer *psb; uint64_t considx, prodidx; int nsamples, nrecords, pass, iter; #ifdef INVARIANTS int ncallchains; int nfree; int start_ticks = ticks; #endif psb = pmc_pcpu[cpu]->pc_sb[ring]; td = curthread; KASSERT(td->td_pflags & TDP_CALLCHAIN, ("[pmc,%d] Retrieving callchain for thread that doesn't want it", __LINE__)); #ifdef INVARIANTS ncallchains = 0; nfree = 0; #endif nrecords = INT_MAX; pass = 0; restart: if (ring == PMC_UR) nrecords = atomic_readandclear_32(&td->td_pmcpend); for (iter = 0, considx = psb->ps_considx, prodidx = psb->ps_prodidx; considx < prodidx && iter < pmc_nsamples; considx++, iter++) { ps = PMC_CONS_SAMPLE_OFF(psb, considx); /* * Iterate through all deferred callchain requests. * Walk from the current read pointer to the current * write pointer. */ #ifdef INVARIANTS if (ps->ps_nsamples == PMC_SAMPLE_FREE) { nfree++; continue; } if ((ps->ps_pmc == NULL) || (ps->ps_pmc->pm_state != PMC_STATE_RUNNING)) nfree++; #endif if (ps->ps_td != td || ps->ps_nsamples != PMC_USER_CALLCHAIN_PENDING || ps->ps_pmc->pm_state != PMC_STATE_RUNNING) continue; KASSERT(ps->ps_cpu == cpu, ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__, ps->ps_cpu, PCPU_GET(cpuid))); pm = ps->ps_pmc; KASSERT(pm->pm_flags & PMC_F_CALLCHAIN, ("[pmc,%d] Retrieving callchain for PMC that doesn't " "want it", __LINE__)); KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, ("[pmc,%d] runcount %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount))); if (ring == PMC_UR) { nsamples = ps->ps_nsamples_actual; counter_u64_add(pmc_stats.pm_merges, 1); } else nsamples = 0; /* * Retrieve the callchain and mark the sample buffer * as 'processable' by the timer tick sweep code. */ #ifdef INVARIANTS ncallchains++; #endif if (__predict_true(nsamples < pmc_callchaindepth - 1)) nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples, pmc_callchaindepth - nsamples - 1, tf); /* * We have to prevent hardclock from potentially overwriting * this sample between when we read the value and when we set * it */ spinlock_enter(); /* * Verify that the sample hasn't been dropped in the meantime */ if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) { ps->ps_nsamples = nsamples; /* * If we couldn't get a sample, simply drop the reference */ if (nsamples == 0) counter_u64_add(pm->pm_runcount, -1); } spinlock_exit(); if (nrecords-- == 1) break; } if (__predict_false(ring == PMC_UR && td->td_pmcpend)) { if (pass == 0) { pass = 1; goto restart; } /* only collect samples for this part once */ td->td_pmcpend = 0; } #ifdef INVARIANTS if ((ticks - start_ticks) > hz) log(LOG_ERR, "%s took %d ticks\n", __func__, (ticks - start_ticks)); #endif /* mark CPU as needing processing */ DPCPU_SET(pmc_sampled, 1); } /* * Process saved PC samples. */ static void pmc_process_samples(int cpu, ring_type_t ring) { struct pmc *pm; int adjri, n; struct thread *td; struct pmc_owner *po; struct pmc_sample *ps; struct pmc_classdep *pcd; struct pmc_samplebuffer *psb; uint64_t delta; KASSERT(PCPU_GET(cpuid) == cpu, ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__, PCPU_GET(cpuid), cpu)); psb = pmc_pcpu[cpu]->pc_sb[ring]; delta = psb->ps_prodidx - psb->ps_considx; MPASS(delta <= pmc_nsamples); MPASS(psb->ps_considx <= psb->ps_prodidx); for (n = 0; psb->ps_considx < psb->ps_prodidx; psb->ps_considx++, n++) { ps = PMC_CONS_SAMPLE(psb); if (__predict_false(ps->ps_nsamples == PMC_SAMPLE_FREE)) continue; pm = ps->ps_pmc; /* skip non-running samples */ if (pm->pm_state != PMC_STATE_RUNNING) goto entrydone; KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm, (unsigned long)counter_u64_fetch(pm->pm_runcount))); po = pm->pm_owner; KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__, pm, PMC_TO_MODE(pm))); /* If there is a pending AST wait for completion */ if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) { /* if we've been waiting more than 1 tick to * collect a callchain for this record then * drop it and move on. */ if (ticks - ps->ps_ticks > 1) { /* * track how often we hit this as it will * preferentially lose user samples * for long running system calls */ counter_u64_add(pmc_stats.pm_overwrites, 1); goto entrydone; } /* Need a rescan at a later time. */ DPCPU_SET(pmc_sampled, 1); break; } PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu, pm, ps->ps_nsamples, ps->ps_flags, (int) (psb->ps_prodidx & pmc_sample_mask), (int) (psb->ps_considx & pmc_sample_mask)); /* * If this is a process-mode PMC that is attached to * its owner, and if the PC is in user mode, update * profiling statistics like timer-based profiling * would have done. * * Otherwise, this is either a sampling-mode PMC that * is attached to a different process than its owner, * or a system-wide sampling PMC. Dispatch a log * entry to the PMC's owner process. */ if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) { if (ps->ps_flags & PMC_CC_F_USERSPACE) { td = FIRST_THREAD_IN_PROC(po->po_owner); addupc_intr(td, ps->ps_pc[0], 1); } } else pmclog_process_callchain(pm, ps); entrydone: ps->ps_nsamples = 0; /* mark entry as free */ KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm, (unsigned long)counter_u64_fetch(pm->pm_runcount))); counter_u64_add(pm->pm_runcount, -1); } counter_u64_add(pmc_stats.pm_log_sweeps, 1); /* Do not re-enable stalled PMCs if we failed to process any samples */ if (n == 0) return; /* * Restart any stalled sampling PMCs on this CPU. * * If the NMI handler sets the pm_stalled field of a PMC after * the check below, we'll end up processing the stalled PMC at * the next hardclock tick. */ for (n = 0; n < md->pmd_npmc; n++) { pcd = pmc_ri_to_classdep(md, n, &adjri); KASSERT(pcd != NULL, ("[pmc,%d] null pcd ri=%d", __LINE__, n)); (void) (*pcd->pcd_get_config)(cpu,adjri,&pm); if (pm == NULL || /* !cfg'ed */ pm->pm_state != PMC_STATE_RUNNING || /* !active */ !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */ !pm->pm_pcpu_state[cpu].pps_cpustate || /* !desired */ !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */ continue; pm->pm_pcpu_state[cpu].pps_stalled = 0; (*pcd->pcd_start_pmc)(cpu, adjri); } } /* * Event handlers. */ /* * Handle a process exit. * * Remove this process from all hash tables. If this process * owned any PMCs, turn off those PMCs and deallocate them, * removing any associations with target processes. * * This function will be called by the last 'thread' of a * process. * * XXX This eventhandler gets called early in the exit process. * Consider using a 'hook' invocation from thread_exit() or equivalent * spot. Another negative is that kse_exit doesn't seem to call * exit1() [??]. * */ static void pmc_process_exit(void *arg __unused, struct proc *p) { struct pmc *pm; int adjri, cpu; unsigned int ri; int is_using_hwpmcs; struct pmc_owner *po; struct pmc_process *pp; struct pmc_classdep *pcd; pmc_value_t newvalue, tmp; PROC_LOCK(p); is_using_hwpmcs = p->p_flag & P_HWPMC; PROC_UNLOCK(p); /* * Log a sysexit event to all SS PMC owners. */ PMC_EPOCH_ENTER(); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_sysexit(po, p->p_pid); PMC_EPOCH_EXIT(); if (!is_using_hwpmcs) return; PMC_GET_SX_XLOCK(); PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid, p->p_comm); /* * Since this code is invoked by the last thread in an exiting * process, we would have context switched IN at some prior * point. However, with PREEMPTION, kernel mode context * switches may happen any time, so we want to disable a * context switch OUT till we get any PMCs targeting this * process off the hardware. * * We also need to atomically remove this process' * entry from our target process hash table, using * PMC_FLAG_REMOVE. */ PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid, p->p_comm); critical_enter(); /* no preemption */ cpu = curthread->td_oncpu; if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_REMOVE)) != NULL) { PMCDBG2(PRC,EXT,2, "process-exit proc=%p pmc-process=%p", p, pp); /* * The exiting process could the target of * some PMCs which will be running on * currently executing CPU. * * We need to turn these PMCs off like we * would do at context switch OUT time. */ for (ri = 0; ri < md->pmd_npmc; ri++) { /* * Pick up the pmc pointer from hardware * state similar to the CSW_OUT code. */ pm = NULL; pcd = pmc_ri_to_classdep(md, ri, &adjri); (void) (*pcd->pcd_get_config)(cpu, adjri, &pm); PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm); if (pm == NULL || !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) continue; PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p " "state=%d", ri, pp->pp_pmcs[ri].pp_pmc, pm, pm->pm_state); KASSERT(PMC_TO_ROWINDEX(pm) == ri, ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", __LINE__, PMC_TO_ROWINDEX(pm), ri)); KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc)); KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, ("[pmc,%d] bad runcount ri %d rc %ld", __LINE__, ri, (unsigned long)counter_u64_fetch(pm->pm_runcount))); /* * Change desired state, and then stop if not * stalled. This two-step dance should avoid * race conditions where an interrupt re-enables * the PMC after this code has already checked * the pm_stalled flag. */ if (pm->pm_pcpu_state[cpu].pps_cpustate) { pm->pm_pcpu_state[cpu].pps_cpustate = 0; if (!pm->pm_pcpu_state[cpu].pps_stalled) { (void) pcd->pcd_stop_pmc(cpu, adjri); if (PMC_TO_MODE(pm) == PMC_MODE_TC) { pcd->pcd_read_pmc(cpu, adjri, &newvalue); tmp = newvalue - PMC_PCPU_SAVED(cpu,ri); mtx_pool_lock_spin(pmc_mtxpool, pm); pm->pm_gv.pm_savedvalue += tmp; pp->pp_pmcs[ri].pp_pmcval += tmp; mtx_pool_unlock_spin( pmc_mtxpool, pm); } } } KASSERT((int64_t) counter_u64_fetch(pm->pm_runcount) > 0, ("[pmc,%d] runcount is %d", __LINE__, ri)); counter_u64_add(pm->pm_runcount, -1); (void) pcd->pcd_config_pmc(cpu, adjri, NULL); } /* * Inform the MD layer of this pseudo "context switch * out" */ (void) md->pmd_switch_out(pmc_pcpu[cpu], pp); critical_exit(); /* ok to be pre-empted now */ /* * Unlink this process from the PMCs that are * targeting it. This will send a signal to * all PMC owner's whose PMCs are orphaned. * * Log PMC value at exit time if requested. */ for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm))) pmclog_process_procexit(pm, pp); pmc_unlink_target_process(pm, pp); } free(pp, M_PMC); } else critical_exit(); /* pp == NULL */ /* * If the process owned PMCs, free them up and free up * memory. */ if ((po = pmc_find_owner_descriptor(p)) != NULL) { pmc_remove_owner(po); pmc_destroy_owner_descriptor(po); } sx_xunlock(&pmc_sx); } /* * Handle a process fork. * * If the parent process 'p1' is under HWPMC monitoring, then copy * over any attached PMCs that have 'do_descendants' semantics. */ static void pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc, int flags) { int is_using_hwpmcs; unsigned int ri; uint32_t do_descendants; struct pmc *pm; struct pmc_owner *po; struct pmc_process *ppnew, *ppold; (void) flags; /* unused parameter */ PROC_LOCK(p1); is_using_hwpmcs = p1->p_flag & P_HWPMC; PROC_UNLOCK(p1); /* * If there are system-wide sampling PMCs active, we need to * log all fork events to their owner's logs. */ PMC_EPOCH_ENTER(); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) { pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); pmclog_process_proccreate(po, newproc, 1); } PMC_EPOCH_EXIT(); if (!is_using_hwpmcs) return; PMC_GET_SX_XLOCK(); PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1, p1->p_pid, p1->p_comm, newproc); /* * If the parent process (curthread->td_proc) is a * target of any PMCs, look for PMCs that are to be * inherited, and link these into the new process * descriptor. */ if ((ppold = pmc_find_process_descriptor(curthread->td_proc, PMC_FLAG_NONE)) == NULL) goto done; /* nothing to do */ do_descendants = 0; for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL) do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS; if (do_descendants == 0) /* nothing to do */ goto done; /* * Now mark the new process as being tracked by this driver. */ PROC_LOCK(newproc); newproc->p_flag |= P_HWPMC; PROC_UNLOCK(newproc); /* allocate a descriptor for the new process */ if ((ppnew = pmc_find_process_descriptor(newproc, PMC_FLAG_ALLOCATE)) == NULL) goto done; /* * Run through all PMCs that were targeting the old process * and which specified F_DESCENDANTS and attach them to the * new process. * * Log the fork event to all owners of PMCs attached to this * process, if not already logged. */ for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && (pm->pm_flags & PMC_F_DESCENDANTS)) { pmc_link_target_process(pm, ppnew); po = pm->pm_owner; if (po->po_sscount == 0 && po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); } done: sx_xunlock(&pmc_sx); } static void pmc_process_threadcreate(struct thread *td) { struct pmc_owner *po; PMC_EPOCH_ENTER(); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_threadcreate(po, td, 1); PMC_EPOCH_EXIT(); } static void pmc_process_threadexit(struct thread *td) { struct pmc_owner *po; PMC_EPOCH_ENTER(); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_threadexit(po, td); PMC_EPOCH_EXIT(); } static void pmc_process_proccreate(struct proc *p) { struct pmc_owner *po; PMC_EPOCH_ENTER(); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_proccreate(po, p, 1 /* sync */); PMC_EPOCH_EXIT(); } static void pmc_process_allproc(struct pmc *pm) { struct pmc_owner *po; struct thread *td; struct proc *p; po = pm->pm_owner; if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) return; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { pmclog_process_proccreate(po, p, 0 /* sync */); PROC_LOCK(p); FOREACH_THREAD_IN_PROC(p, td) pmclog_process_threadcreate(po, td, 0 /* sync */); PROC_UNLOCK(p); } sx_sunlock(&allproc_lock); pmclog_flush(po, 0); } static void pmc_kld_load(void *arg __unused, linker_file_t lf) { struct pmc_owner *po; /* * Notify owners of system sampling PMCs about KLD operations. */ PMC_EPOCH_ENTER(); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_map_in(po, (pid_t) -1, (uintfptr_t) lf->address, lf->filename); PMC_EPOCH_EXIT(); /* * TODO: Notify owners of (all) process-sampling PMCs too. */ } static void pmc_kld_unload(void *arg __unused, const char *filename __unused, caddr_t address, size_t size) { struct pmc_owner *po; PMC_EPOCH_ENTER(); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_map_out(po, (pid_t) -1, (uintfptr_t) address, (uintfptr_t) address + size); PMC_EPOCH_EXIT(); /* * TODO: Notify owners of process-sampling PMCs. */ } /* * initialization */ static const char * pmc_name_of_pmcclass(enum pmc_class class) { switch (class) { #undef __PMC_CLASS #define __PMC_CLASS(S,V,D) \ case PMC_CLASS_##S: \ return #S; __PMC_CLASSES(); default: return (""); } } /* * Base class initializer: allocate structure and set default classes. */ struct pmc_mdep * pmc_mdep_alloc(int nclasses) { struct pmc_mdep *md; int n; /* SOFT + md classes */ n = 1 + nclasses; md = malloc(sizeof(struct pmc_mdep) + n * sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO); md->pmd_nclass = n; /* Add base class. */ pmc_soft_initialize(md); return md; } void pmc_mdep_free(struct pmc_mdep *md) { pmc_soft_finalize(md); free(md, M_PMC); } static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp) { (void) pc; (void) pp; return (0); } static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp) { (void) pc; (void) pp; return (0); } static struct pmc_mdep * pmc_generic_cpu_initialize(void) { struct pmc_mdep *md; md = pmc_mdep_alloc(0); md->pmd_cputype = PMC_CPU_GENERIC; md->pmd_pcpu_init = NULL; md->pmd_pcpu_fini = NULL; md->pmd_switch_in = generic_switch_in; md->pmd_switch_out = generic_switch_out; return (md); } static void pmc_generic_cpu_finalize(struct pmc_mdep *md) { (void) md; } static int pmc_initialize(void) { int c, cpu, error, n, ri; unsigned int maxcpu, domain; struct pcpu *pc; struct pmc_binding pb; struct pmc_sample *ps; struct pmc_classdep *pcd; struct pmc_samplebuffer *sb; md = NULL; error = 0; pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK); pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK); pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK); pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK); pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK); pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK); pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK); pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK); pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK); pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK); #ifdef HWPMC_DEBUG /* parse debug flags first */ if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, sizeof(pmc_debugstr))) pmc_debugflags_parse(pmc_debugstr, pmc_debugstr+strlen(pmc_debugstr)); #endif PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION); /* check kernel version */ if (pmc_kernel_version != PMC_VERSION) { if (pmc_kernel_version == 0) printf("hwpmc: this kernel has not been compiled with " "'options HWPMC_HOOKS'.\n"); else printf("hwpmc: kernel version (0x%x) does not match " "module version (0x%x).\n", pmc_kernel_version, PMC_VERSION); return EPROGMISMATCH; } /* * check sysctl parameters */ if (pmc_hashsize <= 0) { (void) printf("hwpmc: tunable \"hashsize\"=%d must be " "greater than zero.\n", pmc_hashsize); pmc_hashsize = PMC_HASH_SIZE; } if (pmc_nsamples <= 0 || pmc_nsamples > 65535) { (void) printf("hwpmc: tunable \"nsamples\"=%d out of " "range.\n", pmc_nsamples); pmc_nsamples = PMC_NSAMPLES; } pmc_sample_mask = pmc_nsamples-1; if (pmc_callchaindepth <= 0 || pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) { (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of " "range - using %d.\n", pmc_callchaindepth, PMC_CALLCHAIN_DEPTH_MAX); pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX; } md = pmc_md_initialize(); if (md == NULL) { /* Default to generic CPU. */ md = pmc_generic_cpu_initialize(); if (md == NULL) return (ENOSYS); } KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1, ("[pmc,%d] no classes or pmcs", __LINE__)); /* Compute the map from row-indices to classdep pointers. */ pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) * md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO); for (n = 0; n < md->pmd_npmc; n++) pmc_rowindex_to_classdep[n] = NULL; for (ri = c = 0; c < md->pmd_nclass; c++) { pcd = &md->pmd_classdep[c]; for (n = 0; n < pcd->pcd_num; n++, ri++) pmc_rowindex_to_classdep[ri] = pcd; } KASSERT(ri == md->pmd_npmc, ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__, ri, md->pmd_npmc)); maxcpu = pmc_cpu_max(); /* allocate space for the per-cpu array */ pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC, M_WAITOK|M_ZERO); /* per-cpu 'saved values' for managing process-mode PMCs */ pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc, M_PMC, M_WAITOK); /* Perform CPU-dependent initialization. */ pmc_save_cpu_binding(&pb); error = 0; for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) { if (!pmc_cpu_is_active(cpu)) continue; pmc_select_cpu(cpu); pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) + md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC, M_WAITOK|M_ZERO); if (md->pmd_pcpu_init) error = md->pmd_pcpu_init(md, cpu); for (n = 0; error == 0 && n < md->pmd_nclass; n++) error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu); } pmc_restore_cpu_binding(&pb); if (error) return (error); /* allocate space for the sample array */ for (cpu = 0; cpu < maxcpu; cpu++) { if (!pmc_cpu_is_active(cpu)) continue; pc = pcpu_find(cpu); domain = pc->pc_domain; sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + pmc_nsamples * sizeof(struct pmc_sample), M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); KASSERT(pmc_pcpu[cpu] != NULL, ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); sb->ps_callchains = malloc_domainset(pmc_callchaindepth * pmc_nsamples * sizeof(uintptr_t), M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) ps->ps_pc = sb->ps_callchains + (n * pmc_callchaindepth); pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb; sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + pmc_nsamples * sizeof(struct pmc_sample), M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); sb->ps_callchains = malloc_domainset(pmc_callchaindepth * pmc_nsamples * sizeof(uintptr_t), M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) ps->ps_pc = sb->ps_callchains + (n * pmc_callchaindepth); pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb; sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + pmc_nsamples * sizeof(struct pmc_sample), M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); sb->ps_callchains = malloc_domainset(pmc_callchaindepth * pmc_nsamples * sizeof(uintptr_t), M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) ps->ps_pc = sb->ps_callchains + n * pmc_callchaindepth; pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb; } /* allocate space for the row disposition array */ pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO); /* mark all PMCs as available */ for (n = 0; n < (int) md->pmd_npmc; n++) PMC_MARK_ROW_FREE(n); /* allocate thread hash tables */ pmc_ownerhash = hashinit(pmc_hashsize, M_PMC, &pmc_ownerhashmask); pmc_processhash = hashinit(pmc_hashsize, M_PMC, &pmc_processhashmask); mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf", MTX_SPIN); CK_LIST_INIT(&pmc_ss_owners); pmc_ss_count = 0; /* allocate a pool of spin mutexes */ pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size, MTX_SPIN); PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx " "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask, pmc_processhash, pmc_processhashmask); /* Initialize a spin mutex for the thread free list. */ mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf", MTX_SPIN); - /* - * Initialize the callout to monitor the thread free list. - * This callout will also handle the initial population of the list. - */ - taskqgroup_config_gtask_init(NULL, &free_gtask, pmc_thread_descriptor_pool_free_task, "thread descriptor pool free task"); + /* Initialize the task to prune the thread free list. */ + TASK_INIT(&free_task, 0, pmc_thread_descriptor_pool_free_task, NULL); /* register process {exit,fork,exec} handlers */ pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit, pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY); pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork, pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY); /* register kld event handlers */ pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load, NULL, EVENTHANDLER_PRI_ANY); pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload, NULL, EVENTHANDLER_PRI_ANY); /* initialize logging */ pmclog_initialize(); /* set hook functions */ pmc_intr = md->pmd_intr; wmb(); pmc_hook = pmc_hook_handler; if (error == 0) { printf(PMC_MODULE_NAME ":"); for (n = 0; n < (int) md->pmd_nclass; n++) { pcd = &md->pmd_classdep[n]; printf(" %s/%d/%d/0x%b", pmc_name_of_pmcclass(pcd->pcd_class), pcd->pcd_num, pcd->pcd_width, pcd->pcd_caps, "\20" "\1INT\2USR\3SYS\4EDG\5THR" "\6REA\7WRI\10INV\11QUA\12PRC" "\13TAG\14CSC"); } printf("\n"); } return (error); } /* prepare to be unloaded */ static void pmc_cleanup(void) { int c, cpu; unsigned int maxcpu; struct pmc_ownerhash *ph; struct pmc_owner *po, *tmp; struct pmc_binding pb; #ifdef HWPMC_DEBUG struct pmc_processhash *prh; #endif PMCDBG0(MOD,INI,0, "cleanup"); /* switch off sampling */ CPU_FOREACH(cpu) DPCPU_ID_SET(cpu, pmc_sampled, 0); pmc_intr = NULL; sx_xlock(&pmc_sx); if (pmc_hook == NULL) { /* being unloaded already */ sx_xunlock(&pmc_sx); return; } pmc_hook = NULL; /* prevent new threads from entering module */ /* deregister event handlers */ EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag); EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag); EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag); EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag); /* send SIGBUS to all owner threads, free up allocations */ if (pmc_ownerhash) for (ph = pmc_ownerhash; ph <= &pmc_ownerhash[pmc_ownerhashmask]; ph++) { LIST_FOREACH_SAFE(po, ph, po_next, tmp) { pmc_remove_owner(po); /* send SIGBUS to owner processes */ PMCDBG3(MOD,INI,2, "cleanup signal proc=%p " "(%d, %s)", po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); PROC_LOCK(po->po_owner); kern_psignal(po->po_owner, SIGBUS); PROC_UNLOCK(po->po_owner); pmc_destroy_owner_descriptor(po); } } /* reclaim allocated data structures */ + taskqueue_drain(taskqueue_fast, &free_task); mtx_destroy(&pmc_threadfreelist_mtx); pmc_thread_descriptor_pool_drain(); if (pmc_mtxpool) mtx_pool_destroy(&pmc_mtxpool); mtx_destroy(&pmc_processhash_mtx); - taskqgroup_config_gtask_deinit(&free_gtask); if (pmc_processhash) { #ifdef HWPMC_DEBUG struct pmc_process *pp; PMCDBG0(MOD,INI,3, "destroy process hash"); for (prh = pmc_processhash; prh <= &pmc_processhash[pmc_processhashmask]; prh++) LIST_FOREACH(pp, prh, pp_next) PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid); #endif hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask); pmc_processhash = NULL; } if (pmc_ownerhash) { PMCDBG0(MOD,INI,3, "destroy owner hash"); hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask); pmc_ownerhash = NULL; } KASSERT(CK_LIST_EMPTY(&pmc_ss_owners), ("[pmc,%d] Global SS owner list not empty", __LINE__)); KASSERT(pmc_ss_count == 0, ("[pmc,%d] Global SS count not empty", __LINE__)); /* do processor and pmc-class dependent cleanup */ maxcpu = pmc_cpu_max(); PMCDBG0(MOD,INI,3, "md cleanup"); if (md) { pmc_save_cpu_binding(&pb); for (cpu = 0; cpu < maxcpu; cpu++) { PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p", cpu, pmc_pcpu[cpu]); if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL) continue; pmc_select_cpu(cpu); for (c = 0; c < md->pmd_nclass; c++) md->pmd_classdep[c].pcd_pcpu_fini(md, cpu); if (md->pmd_pcpu_fini) md->pmd_pcpu_fini(md, cpu); } if (md->pmd_cputype == PMC_CPU_GENERIC) pmc_generic_cpu_finalize(md); else pmc_md_finalize(md); pmc_mdep_free(md); md = NULL; pmc_restore_cpu_binding(&pb); } /* Free per-cpu descriptors. */ for (cpu = 0; cpu < maxcpu; cpu++) { if (!pmc_cpu_is_active(cpu)) continue; KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL, ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__, cpu)); KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL, ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__, cpu)); KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL, ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__, cpu)); free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC); free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC); free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC); free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC); free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC); free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC); free_domain(pmc_pcpu[cpu], M_PMC); } free(pmc_pcpu, M_PMC); pmc_pcpu = NULL; free(pmc_pcpu_saved, M_PMC); pmc_pcpu_saved = NULL; if (pmc_pmcdisp) { free(pmc_pmcdisp, M_PMC); pmc_pmcdisp = NULL; } if (pmc_rowindex_to_classdep) { free(pmc_rowindex_to_classdep, M_PMC); pmc_rowindex_to_classdep = NULL; } pmclog_shutdown(); counter_u64_free(pmc_stats.pm_intr_ignored); counter_u64_free(pmc_stats.pm_intr_processed); counter_u64_free(pmc_stats.pm_intr_bufferfull); counter_u64_free(pmc_stats.pm_syscalls); counter_u64_free(pmc_stats.pm_syscall_errors); counter_u64_free(pmc_stats.pm_buffer_requests); counter_u64_free(pmc_stats.pm_buffer_requests_failed); counter_u64_free(pmc_stats.pm_log_sweeps); counter_u64_free(pmc_stats.pm_merges); counter_u64_free(pmc_stats.pm_overwrites); sx_xunlock(&pmc_sx); /* we are done */ } /* * The function called at load/unload. */ static int load (struct module *module __unused, int cmd, void *arg __unused) { int error; error = 0; switch (cmd) { case MOD_LOAD : /* initialize the subsystem */ error = pmc_initialize(); if (error != 0) break; PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d", pmc_syscall_num, pmc_cpu_max()); break; case MOD_UNLOAD : case MOD_SHUTDOWN: pmc_cleanup(); PMCDBG0(MOD,INI,1, "unloaded"); break; default : error = EINVAL; /* XXX should panic(9) */ break; } return error; } Index: head/sys/kern/subr_gtaskqueue.c =================================================================== --- head/sys/kern/subr_gtaskqueue.c (revision 359437) +++ head/sys/kern/subr_gtaskqueue.c (revision 359438) @@ -1,837 +1,819 @@ /*- * Copyright (c) 2000 Doug Rabson * Copyright (c) 2014 Jeff Roberson * Copyright (c) 2016 Matthew Macy * 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. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_GTASKQUEUE, "gtaskqueue", "Group Task Queues"); static void gtaskqueue_thread_enqueue(void *); static void gtaskqueue_thread_loop(void *arg); static int task_is_running(struct gtaskqueue *queue, struct gtask *gtask); static void gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask); TASKQGROUP_DEFINE(softirq, mp_ncpus, 1); -TASKQGROUP_DEFINE(config, 1, 1); struct gtaskqueue_busy { struct gtask *tb_running; u_int tb_seq; LIST_ENTRY(gtaskqueue_busy) tb_link; }; typedef void (*gtaskqueue_enqueue_fn)(void *context); struct gtaskqueue { STAILQ_HEAD(, gtask) tq_queue; LIST_HEAD(, gtaskqueue_busy) tq_active; u_int tq_seq; int tq_callouts; struct mtx_padalign tq_mutex; gtaskqueue_enqueue_fn tq_enqueue; void *tq_context; char *tq_name; struct thread **tq_threads; int tq_tcount; int tq_spin; int tq_flags; taskqueue_callback_fn tq_callbacks[TASKQUEUE_NUM_CALLBACKS]; void *tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS]; }; #define TQ_FLAGS_ACTIVE (1 << 0) #define TQ_FLAGS_BLOCKED (1 << 1) #define TQ_FLAGS_UNLOCKED_ENQUEUE (1 << 2) #define DT_CALLOUT_ARMED (1 << 0) #define TQ_LOCK(tq) \ do { \ if ((tq)->tq_spin) \ mtx_lock_spin(&(tq)->tq_mutex); \ else \ mtx_lock(&(tq)->tq_mutex); \ } while (0) #define TQ_ASSERT_LOCKED(tq) mtx_assert(&(tq)->tq_mutex, MA_OWNED) #define TQ_UNLOCK(tq) \ do { \ if ((tq)->tq_spin) \ mtx_unlock_spin(&(tq)->tq_mutex); \ else \ mtx_unlock(&(tq)->tq_mutex); \ } while (0) #define TQ_ASSERT_UNLOCKED(tq) mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED) #ifdef INVARIANTS static void gtask_dump(struct gtask *gtask) { printf("gtask: %p ta_flags=%x ta_priority=%d ta_func=%p ta_context=%p\n", gtask, gtask->ta_flags, gtask->ta_priority, gtask->ta_func, gtask->ta_context); } #endif static __inline int TQ_SLEEP(struct gtaskqueue *tq, void *p, const char *wm) { if (tq->tq_spin) return (msleep_spin(p, (struct mtx *)&tq->tq_mutex, wm, 0)); return (msleep(p, &tq->tq_mutex, 0, wm, 0)); } static struct gtaskqueue * _gtaskqueue_create(const char *name, int mflags, taskqueue_enqueue_fn enqueue, void *context, int mtxflags, const char *mtxname __unused) { struct gtaskqueue *queue; char *tq_name; tq_name = malloc(TASKQUEUE_NAMELEN, M_GTASKQUEUE, mflags | M_ZERO); if (!tq_name) return (NULL); snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue"); queue = malloc(sizeof(struct gtaskqueue), M_GTASKQUEUE, mflags | M_ZERO); if (!queue) { free(tq_name, M_GTASKQUEUE); return (NULL); } STAILQ_INIT(&queue->tq_queue); LIST_INIT(&queue->tq_active); queue->tq_enqueue = enqueue; queue->tq_context = context; queue->tq_name = tq_name; queue->tq_spin = (mtxflags & MTX_SPIN) != 0; queue->tq_flags |= TQ_FLAGS_ACTIVE; if (enqueue == gtaskqueue_thread_enqueue) queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE; mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags); return (queue); } /* * Signal a taskqueue thread to terminate. */ static void gtaskqueue_terminate(struct thread **pp, struct gtaskqueue *tq) { while (tq->tq_tcount > 0 || tq->tq_callouts > 0) { wakeup(tq); TQ_SLEEP(tq, pp, "gtq_destroy"); } } static void __unused gtaskqueue_free(struct gtaskqueue *queue) { TQ_LOCK(queue); queue->tq_flags &= ~TQ_FLAGS_ACTIVE; gtaskqueue_terminate(queue->tq_threads, queue); KASSERT(LIST_EMPTY(&queue->tq_active), ("Tasks still running?")); KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks")); mtx_destroy(&queue->tq_mutex); free(queue->tq_threads, M_GTASKQUEUE); free(queue->tq_name, M_GTASKQUEUE); free(queue, M_GTASKQUEUE); } /* * Wait for all to complete, then prevent it from being enqueued */ void grouptask_block(struct grouptask *grouptask) { struct gtaskqueue *queue = grouptask->gt_taskqueue; struct gtask *gtask = &grouptask->gt_task; #ifdef INVARIANTS if (queue == NULL) { gtask_dump(gtask); panic("queue == NULL"); } #endif TQ_LOCK(queue); gtask->ta_flags |= TASK_NOENQUEUE; gtaskqueue_drain_locked(queue, gtask); TQ_UNLOCK(queue); } void grouptask_unblock(struct grouptask *grouptask) { struct gtaskqueue *queue = grouptask->gt_taskqueue; struct gtask *gtask = &grouptask->gt_task; #ifdef INVARIANTS if (queue == NULL) { gtask_dump(gtask); panic("queue == NULL"); } #endif TQ_LOCK(queue); gtask->ta_flags &= ~TASK_NOENQUEUE; TQ_UNLOCK(queue); } int grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *gtask) { #ifdef INVARIANTS if (queue == NULL) { gtask_dump(gtask); panic("queue == NULL"); } #endif TQ_LOCK(queue); if (gtask->ta_flags & TASK_ENQUEUED) { TQ_UNLOCK(queue); return (0); } if (gtask->ta_flags & TASK_NOENQUEUE) { TQ_UNLOCK(queue); return (EAGAIN); } STAILQ_INSERT_TAIL(&queue->tq_queue, gtask, ta_link); gtask->ta_flags |= TASK_ENQUEUED; TQ_UNLOCK(queue); if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0) queue->tq_enqueue(queue->tq_context); return (0); } static void gtaskqueue_task_nop_fn(void *context) { } /* * Block until all currently queued tasks in this taskqueue * have begun execution. Tasks queued during execution of * this function are ignored. */ static void gtaskqueue_drain_tq_queue(struct gtaskqueue *queue) { struct gtask t_barrier; if (STAILQ_EMPTY(&queue->tq_queue)) return; /* * Enqueue our barrier after all current tasks, but with * the highest priority so that newly queued tasks cannot * pass it. Because of the high priority, we can not use * taskqueue_enqueue_locked directly (which drops the lock * anyway) so just insert it at tail while we have the * queue lock. */ GTASK_INIT(&t_barrier, 0, USHRT_MAX, gtaskqueue_task_nop_fn, &t_barrier); STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link); t_barrier.ta_flags |= TASK_ENQUEUED; /* * Once the barrier has executed, all previously queued tasks * have completed or are currently executing. */ while (t_barrier.ta_flags & TASK_ENQUEUED) TQ_SLEEP(queue, &t_barrier, "gtq_qdrain"); } /* * Block until all currently executing tasks for this taskqueue * complete. Tasks that begin execution during the execution * of this function are ignored. */ static void gtaskqueue_drain_tq_active(struct gtaskqueue *queue) { struct gtaskqueue_busy *tb; u_int seq; if (LIST_EMPTY(&queue->tq_active)) return; /* Block taskq_terminate().*/ queue->tq_callouts++; /* Wait for any active task with sequence from the past. */ seq = queue->tq_seq; restart: LIST_FOREACH(tb, &queue->tq_active, tb_link) { if ((int)(tb->tb_seq - seq) <= 0) { TQ_SLEEP(queue, tb->tb_running, "gtq_adrain"); goto restart; } } /* Release taskqueue_terminate(). */ queue->tq_callouts--; if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0) wakeup_one(queue->tq_threads); } void gtaskqueue_block(struct gtaskqueue *queue) { TQ_LOCK(queue); queue->tq_flags |= TQ_FLAGS_BLOCKED; TQ_UNLOCK(queue); } void gtaskqueue_unblock(struct gtaskqueue *queue) { TQ_LOCK(queue); queue->tq_flags &= ~TQ_FLAGS_BLOCKED; if (!STAILQ_EMPTY(&queue->tq_queue)) queue->tq_enqueue(queue->tq_context); TQ_UNLOCK(queue); } static void gtaskqueue_run_locked(struct gtaskqueue *queue) { struct epoch_tracker et; struct gtaskqueue_busy tb; struct gtask *gtask; bool in_net_epoch; KASSERT(queue != NULL, ("tq is NULL")); TQ_ASSERT_LOCKED(queue); tb.tb_running = NULL; LIST_INSERT_HEAD(&queue->tq_active, &tb, tb_link); in_net_epoch = false; while ((gtask = STAILQ_FIRST(&queue->tq_queue)) != NULL) { STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link); gtask->ta_flags &= ~TASK_ENQUEUED; tb.tb_running = gtask; tb.tb_seq = ++queue->tq_seq; TQ_UNLOCK(queue); KASSERT(gtask->ta_func != NULL, ("task->ta_func is NULL")); if (!in_net_epoch && TASK_IS_NET(gtask)) { in_net_epoch = true; NET_EPOCH_ENTER(et); } else if (in_net_epoch && !TASK_IS_NET(gtask)) { NET_EPOCH_EXIT(et); in_net_epoch = false; } gtask->ta_func(gtask->ta_context); TQ_LOCK(queue); wakeup(gtask); } if (in_net_epoch) NET_EPOCH_EXIT(et); LIST_REMOVE(&tb, tb_link); } static int task_is_running(struct gtaskqueue *queue, struct gtask *gtask) { struct gtaskqueue_busy *tb; TQ_ASSERT_LOCKED(queue); LIST_FOREACH(tb, &queue->tq_active, tb_link) { if (tb->tb_running == gtask) return (1); } return (0); } static int gtaskqueue_cancel_locked(struct gtaskqueue *queue, struct gtask *gtask) { if (gtask->ta_flags & TASK_ENQUEUED) STAILQ_REMOVE(&queue->tq_queue, gtask, gtask, ta_link); gtask->ta_flags &= ~TASK_ENQUEUED; return (task_is_running(queue, gtask) ? EBUSY : 0); } int gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask) { int error; TQ_LOCK(queue); error = gtaskqueue_cancel_locked(queue, gtask); TQ_UNLOCK(queue); return (error); } static void gtaskqueue_drain_locked(struct gtaskqueue *queue, struct gtask *gtask) { while ((gtask->ta_flags & TASK_ENQUEUED) || task_is_running(queue, gtask)) TQ_SLEEP(queue, gtask, "gtq_drain"); } void gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *gtask) { if (!queue->tq_spin) WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); TQ_LOCK(queue); gtaskqueue_drain_locked(queue, gtask); TQ_UNLOCK(queue); } void gtaskqueue_drain_all(struct gtaskqueue *queue) { if (!queue->tq_spin) WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); TQ_LOCK(queue); gtaskqueue_drain_tq_queue(queue); gtaskqueue_drain_tq_active(queue); TQ_UNLOCK(queue); } static int _gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri, cpuset_t *mask, const char *name, va_list ap) { char ktname[MAXCOMLEN + 1]; struct thread *td; struct gtaskqueue *tq; int i, error; if (count <= 0) return (EINVAL); vsnprintf(ktname, sizeof(ktname), name, ap); tq = *tqp; tq->tq_threads = malloc(sizeof(struct thread *) * count, M_GTASKQUEUE, M_NOWAIT | M_ZERO); if (tq->tq_threads == NULL) { printf("%s: no memory for %s threads\n", __func__, ktname); return (ENOMEM); } for (i = 0; i < count; i++) { if (count == 1) error = kthread_add(gtaskqueue_thread_loop, tqp, NULL, &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname); else error = kthread_add(gtaskqueue_thread_loop, tqp, NULL, &tq->tq_threads[i], RFSTOPPED, 0, "%s_%d", ktname, i); if (error) { /* should be ok to continue, taskqueue_free will dtrt */ printf("%s: kthread_add(%s): error %d", __func__, ktname, error); tq->tq_threads[i] = NULL; /* paranoid */ } else tq->tq_tcount++; } for (i = 0; i < count; i++) { if (tq->tq_threads[i] == NULL) continue; td = tq->tq_threads[i]; if (mask) { error = cpuset_setthread(td->td_tid, mask); /* * Failing to pin is rarely an actual fatal error; * it'll just affect performance. */ if (error) printf("%s: curthread=%llu: can't pin; " "error=%d\n", __func__, (unsigned long long) td->td_tid, error); } thread_lock(td); sched_prio(td, pri); sched_add(td, SRQ_BORING); } return (0); } static int gtaskqueue_start_threads(struct gtaskqueue **tqp, int count, int pri, const char *name, ...) { va_list ap; int error; va_start(ap, name); error = _gtaskqueue_start_threads(tqp, count, pri, NULL, name, ap); va_end(ap); return (error); } static inline void gtaskqueue_run_callback(struct gtaskqueue *tq, enum taskqueue_callback_type cb_type) { taskqueue_callback_fn tq_callback; TQ_ASSERT_UNLOCKED(tq); tq_callback = tq->tq_callbacks[cb_type]; if (tq_callback != NULL) tq_callback(tq->tq_cb_contexts[cb_type]); } static void gtaskqueue_thread_loop(void *arg) { struct gtaskqueue **tqp, *tq; tqp = arg; tq = *tqp; gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT); TQ_LOCK(tq); while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) { /* XXX ? */ gtaskqueue_run_locked(tq); /* * Because taskqueue_run() can drop tq_mutex, we need to * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the * meantime, which means we missed a wakeup. */ if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0) break; TQ_SLEEP(tq, tq, "-"); } gtaskqueue_run_locked(tq); /* * This thread is on its way out, so just drop the lock temporarily * in order to call the shutdown callback. This allows the callback * to look at the taskqueue, even just before it dies. */ TQ_UNLOCK(tq); gtaskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN); TQ_LOCK(tq); /* rendezvous with thread that asked us to terminate */ tq->tq_tcount--; wakeup_one(tq->tq_threads); TQ_UNLOCK(tq); kthread_exit(); } static void gtaskqueue_thread_enqueue(void *context) { struct gtaskqueue **tqp, *tq; tqp = context; tq = *tqp; wakeup_any(tq); } static struct gtaskqueue * gtaskqueue_create_fast(const char *name, int mflags, taskqueue_enqueue_fn enqueue, void *context) { return _gtaskqueue_create(name, mflags, enqueue, context, MTX_SPIN, "fast_taskqueue"); } struct taskqgroup_cpu { LIST_HEAD(, grouptask) tgc_tasks; struct gtaskqueue *tgc_taskq; int tgc_cnt; int tgc_cpu; }; struct taskqgroup { struct taskqgroup_cpu tqg_queue[MAXCPU]; struct mtx tqg_lock; const char * tqg_name; int tqg_cnt; }; struct taskq_bind_task { struct gtask bt_task; int bt_cpuid; }; static void taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx, int cpu) { struct taskqgroup_cpu *qcpu; qcpu = &qgroup->tqg_queue[idx]; LIST_INIT(&qcpu->tgc_tasks); qcpu->tgc_taskq = gtaskqueue_create_fast(NULL, M_WAITOK, taskqueue_thread_enqueue, &qcpu->tgc_taskq); gtaskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT, "%s_%d", qgroup->tqg_name, idx); qcpu->tgc_cpu = cpu; } /* * Find the taskq with least # of tasks that doesn't currently have any * other queues from the uniq identifier. */ static int taskqgroup_find(struct taskqgroup *qgroup, void *uniq) { struct grouptask *n; int i, idx, mincnt; int strict; mtx_assert(&qgroup->tqg_lock, MA_OWNED); KASSERT(qgroup->tqg_cnt != 0, ("qgroup %s has no queues", qgroup->tqg_name)); /* * Two passes: first scan for a queue with the least tasks that * does not already service this uniq id. If that fails simply find * the queue with the least total tasks. */ for (idx = -1, mincnt = INT_MAX, strict = 1; mincnt == INT_MAX; strict = 0) { for (i = 0; i < qgroup->tqg_cnt; i++) { if (qgroup->tqg_queue[i].tgc_cnt > mincnt) continue; if (strict) { LIST_FOREACH(n, &qgroup->tqg_queue[i].tgc_tasks, gt_list) if (n->gt_uniq == uniq) break; if (n != NULL) continue; } mincnt = qgroup->tqg_queue[i].tgc_cnt; idx = i; } } if (idx == -1) panic("%s: failed to pick a qid.", __func__); return (idx); } void taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask, void *uniq, device_t dev, struct resource *irq, const char *name) { int cpu, qid, error; KASSERT(qgroup->tqg_cnt > 0, ("qgroup %s has no queues", qgroup->tqg_name)); gtask->gt_uniq = uniq; snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask"); gtask->gt_dev = dev; gtask->gt_irq = irq; gtask->gt_cpu = -1; mtx_lock(&qgroup->tqg_lock); qid = taskqgroup_find(qgroup, uniq); qgroup->tqg_queue[qid].tgc_cnt++; LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; if (dev != NULL && irq != NULL) { cpu = qgroup->tqg_queue[qid].tgc_cpu; gtask->gt_cpu = cpu; mtx_unlock(&qgroup->tqg_lock); error = bus_bind_intr(dev, irq, cpu); if (error) printf("%s: binding interrupt failed for %s: %d\n", __func__, gtask->gt_name, error); } else mtx_unlock(&qgroup->tqg_lock); } int taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask, void *uniq, int cpu, device_t dev, struct resource *irq, const char *name) { int i, qid, error; gtask->gt_uniq = uniq; snprintf(gtask->gt_name, GROUPTASK_NAMELEN, "%s", name ? name : "grouptask"); gtask->gt_dev = dev; gtask->gt_irq = irq; gtask->gt_cpu = cpu; mtx_lock(&qgroup->tqg_lock); for (i = 0, qid = -1; i < qgroup->tqg_cnt; i++) if (qgroup->tqg_queue[i].tgc_cpu == cpu) { qid = i; break; } if (qid == -1) { mtx_unlock(&qgroup->tqg_lock); printf("%s: qid not found for %s cpu=%d\n", __func__, gtask->gt_name, cpu); return (EINVAL); } qgroup->tqg_queue[qid].tgc_cnt++; LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; cpu = qgroup->tqg_queue[qid].tgc_cpu; mtx_unlock(&qgroup->tqg_lock); if (dev != NULL && irq != NULL) { error = bus_bind_intr(dev, irq, cpu); if (error) printf("%s: binding interrupt failed for %s: %d\n", __func__, gtask->gt_name, error); } return (0); } void taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask) { int i; grouptask_block(gtask); mtx_lock(&qgroup->tqg_lock); for (i = 0; i < qgroup->tqg_cnt; i++) if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue) break; if (i == qgroup->tqg_cnt) panic("%s: task %s not in group", __func__, gtask->gt_name); qgroup->tqg_queue[i].tgc_cnt--; LIST_REMOVE(gtask, gt_list); mtx_unlock(&qgroup->tqg_lock); gtask->gt_taskqueue = NULL; gtask->gt_task.ta_flags &= ~TASK_NOENQUEUE; } static void taskqgroup_binder(void *ctx) { struct taskq_bind_task *gtask; cpuset_t mask; int error; gtask = ctx; CPU_ZERO(&mask); CPU_SET(gtask->bt_cpuid, &mask); error = cpuset_setthread(curthread->td_tid, &mask); thread_lock(curthread); sched_bind(curthread, gtask->bt_cpuid); thread_unlock(curthread); if (error) printf("%s: binding curthread failed: %d\n", __func__, error); free(gtask, M_DEVBUF); } void taskqgroup_bind(struct taskqgroup *qgroup) { struct taskq_bind_task *gtask; int i; /* * Bind taskqueue threads to specific CPUs, if they have been assigned * one. */ if (qgroup->tqg_cnt == 1) return; for (i = 0; i < qgroup->tqg_cnt; i++) { gtask = malloc(sizeof(*gtask), M_DEVBUF, M_WAITOK); GTASK_INIT(>ask->bt_task, 0, 0, taskqgroup_binder, gtask); gtask->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu; grouptaskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq, >ask->bt_task); } } struct taskqgroup * taskqgroup_create(const char *name, int cnt, int stride) { struct taskqgroup *qgroup; int cpu, i, j; qgroup = malloc(sizeof(*qgroup), M_GTASKQUEUE, M_WAITOK | M_ZERO); mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF); qgroup->tqg_name = name; qgroup->tqg_cnt = cnt; for (cpu = i = 0; i < cnt; i++) { taskqgroup_cpu_create(qgroup, i, cpu); for (j = 0; j < stride; j++) cpu = CPU_NEXT(cpu); } return (qgroup); } void taskqgroup_destroy(struct taskqgroup *qgroup) { - -} - -void -taskqgroup_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn, - const char *name) -{ - - GROUPTASK_INIT(gtask, 0, fn, ctx); - taskqgroup_attach(qgroup_config, gtask, gtask, NULL, NULL, name); -} - -void -taskqgroup_config_gtask_deinit(struct grouptask *gtask) -{ - - taskqgroup_detach(qgroup_config, gtask); } Index: head/sys/netinet/in_mcast.c =================================================================== --- head/sys/netinet/in_mcast.c (revision 359437) +++ head/sys/netinet/in_mcast.c (revision 359438) @@ -1,3066 +1,3058 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 2007-2009 Bruce Simpson. * Copyright (c) 2005 Robert N. M. Watson. * 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. * 3. The name of the author may not be used to endorse or promote * products derived from this software without specific prior written * permission. * * 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. */ /* * IPv4 multicast socket, group, and socket option processing module. */ #include __FBSDID("$FreeBSD$"); #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 #ifndef KTR_IGMPV3 #define KTR_IGMPV3 KTR_INET #endif #ifndef __SOCKUNION_DECLARED union sockunion { struct sockaddr_storage ss; struct sockaddr sa; struct sockaddr_dl sdl; struct sockaddr_in sin; }; typedef union sockunion sockunion_t; #define __SOCKUNION_DECLARED #endif /* __SOCKUNION_DECLARED */ static MALLOC_DEFINE(M_INMFILTER, "in_mfilter", "IPv4 multicast PCB-layer source filter"); static MALLOC_DEFINE(M_IPMADDR, "in_multi", "IPv4 multicast group"); static MALLOC_DEFINE(M_IPMOPTS, "ip_moptions", "IPv4 multicast options"); static MALLOC_DEFINE(M_IPMSOURCE, "ip_msource", "IPv4 multicast IGMP-layer source filter"); /* * Locking: * * - Lock order is: Giant, IN_MULTI_LOCK, INP_WLOCK, * IN_MULTI_LIST_LOCK, IGMP_LOCK, IF_ADDR_LOCK. * - The IF_ADDR_LOCK is implicitly taken by inm_lookup() earlier, however * it can be taken by code in net/if.c also. * - ip_moptions and in_mfilter are covered by the INP_WLOCK. * * struct in_multi is covered by IN_MULTI_LIST_LOCK. There isn't strictly * any need for in_multi itself to be virtualized -- it is bound to an ifp * anyway no matter what happens. */ struct mtx in_multi_list_mtx; MTX_SYSINIT(in_multi_mtx, &in_multi_list_mtx, "in_multi_list_mtx", MTX_DEF); struct mtx in_multi_free_mtx; MTX_SYSINIT(in_multi_free_mtx, &in_multi_free_mtx, "in_multi_free_mtx", MTX_DEF); struct sx in_multi_sx; SX_SYSINIT(in_multi_sx, &in_multi_sx, "in_multi_sx"); int ifma_restart; /* * Functions with non-static linkage defined in this file should be * declared in in_var.h: * imo_multi_filter() * in_addmulti() * in_delmulti() * in_joingroup() * in_joingroup_locked() * in_leavegroup() * in_leavegroup_locked() * and ip_var.h: * inp_freemoptions() * inp_getmoptions() * inp_setmoptions() * * XXX: Both carp and pf need to use the legacy (*,G) KPIs in_addmulti() * and in_delmulti(). */ static void imf_commit(struct in_mfilter *); static int imf_get_source(struct in_mfilter *imf, const struct sockaddr_in *psin, struct in_msource **); static struct in_msource * imf_graft(struct in_mfilter *, const uint8_t, const struct sockaddr_in *); static void imf_leave(struct in_mfilter *); static int imf_prune(struct in_mfilter *, const struct sockaddr_in *); static void imf_purge(struct in_mfilter *); static void imf_rollback(struct in_mfilter *); static void imf_reap(struct in_mfilter *); static struct in_mfilter * imo_match_group(const struct ip_moptions *, const struct ifnet *, const struct sockaddr *); static struct in_msource * imo_match_source(struct in_mfilter *, const struct sockaddr *); static void ims_merge(struct ip_msource *ims, const struct in_msource *lims, const int rollback); static int in_getmulti(struct ifnet *, const struct in_addr *, struct in_multi **); static int inm_get_source(struct in_multi *inm, const in_addr_t haddr, const int noalloc, struct ip_msource **pims); #ifdef KTR static int inm_is_ifp_detached(const struct in_multi *); #endif static int inm_merge(struct in_multi *, /*const*/ struct in_mfilter *); static void inm_purge(struct in_multi *); static void inm_reap(struct in_multi *); static void inm_release(struct in_multi *); static struct ip_moptions * inp_findmoptions(struct inpcb *); static int inp_get_source_filters(struct inpcb *, struct sockopt *); static int inp_join_group(struct inpcb *, struct sockopt *); static int inp_leave_group(struct inpcb *, struct sockopt *); static struct ifnet * inp_lookup_mcast_ifp(const struct inpcb *, const struct sockaddr_in *, const struct in_addr); static int inp_block_unblock_source(struct inpcb *, struct sockopt *); static int inp_set_multicast_if(struct inpcb *, struct sockopt *); static int inp_set_source_filters(struct inpcb *, struct sockopt *); static int sysctl_ip_mcast_filters(SYSCTL_HANDLER_ARGS); static SYSCTL_NODE(_net_inet_ip, OID_AUTO, mcast, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "IPv4 multicast"); static u_long in_mcast_maxgrpsrc = IP_MAX_GROUP_SRC_FILTER; SYSCTL_ULONG(_net_inet_ip_mcast, OID_AUTO, maxgrpsrc, CTLFLAG_RWTUN, &in_mcast_maxgrpsrc, 0, "Max source filters per group"); static u_long in_mcast_maxsocksrc = IP_MAX_SOCK_SRC_FILTER; SYSCTL_ULONG(_net_inet_ip_mcast, OID_AUTO, maxsocksrc, CTLFLAG_RWTUN, &in_mcast_maxsocksrc, 0, "Max source filters per socket"); int in_mcast_loop = IP_DEFAULT_MULTICAST_LOOP; SYSCTL_INT(_net_inet_ip_mcast, OID_AUTO, loop, CTLFLAG_RWTUN, &in_mcast_loop, 0, "Loopback multicast datagrams by default"); static SYSCTL_NODE(_net_inet_ip_mcast, OID_AUTO, filters, CTLFLAG_RD | CTLFLAG_MPSAFE, sysctl_ip_mcast_filters, "Per-interface stack-wide source filters"); #ifdef KTR /* * Inline function which wraps assertions for a valid ifp. * The ifnet layer will set the ifma's ifp pointer to NULL if the ifp * is detached. */ static int __inline inm_is_ifp_detached(const struct in_multi *inm) { struct ifnet *ifp; KASSERT(inm->inm_ifma != NULL, ("%s: no ifma", __func__)); ifp = inm->inm_ifma->ifma_ifp; if (ifp != NULL) { /* * Sanity check that netinet's notion of ifp is the * same as net's. */ KASSERT(inm->inm_ifp == ifp, ("%s: bad ifp", __func__)); } return (ifp == NULL); } #endif -static struct grouptask free_gtask; -static struct in_multi_head inm_free_list; -static void inm_release_task(void *arg __unused); -static void inm_init(void) +static struct task free_task; +static struct in_multi_head inm_free_list = SLIST_HEAD_INITIALIZER(); +static void inm_release_task(void *arg __unused, int pending __unused); + +static void +inm_init(void) { - SLIST_INIT(&inm_free_list); - taskqgroup_config_gtask_init(NULL, &free_gtask, inm_release_task, "inm release task"); + TASK_INIT(&free_task, 0, inm_release_task, NULL); } +SYSINIT(inm_init, SI_SUB_TASKQ, SI_ORDER_ANY, inm_init, NULL); -#ifdef EARLY_AP_STARTUP -SYSINIT(inm_init, SI_SUB_SMP + 1, SI_ORDER_FIRST, - inm_init, NULL); -#else -SYSINIT(inm_init, SI_SUB_ROOT_CONF - 1, SI_ORDER_FIRST, - inm_init, NULL); -#endif - - void inm_release_list_deferred(struct in_multi_head *inmh) { if (SLIST_EMPTY(inmh)) return; mtx_lock(&in_multi_free_mtx); SLIST_CONCAT(&inm_free_list, inmh, in_multi, inm_nrele); mtx_unlock(&in_multi_free_mtx); - GROUPTASK_ENQUEUE(&free_gtask); + taskqueue_enqueue(taskqueue_thread, &free_task); } void inm_disconnect(struct in_multi *inm) { struct ifnet *ifp; struct ifmultiaddr *ifma, *ll_ifma; ifp = inm->inm_ifp; IF_ADDR_WLOCK_ASSERT(ifp); ifma = inm->inm_ifma; if_ref(ifp); if (ifma->ifma_flags & IFMA_F_ENQUEUED) { CK_STAILQ_REMOVE(&ifp->if_multiaddrs, ifma, ifmultiaddr, ifma_link); ifma->ifma_flags &= ~IFMA_F_ENQUEUED; } MCDPRINTF("removed ifma: %p from %s\n", ifma, ifp->if_xname); if ((ll_ifma = ifma->ifma_llifma) != NULL) { MPASS(ifma != ll_ifma); ifma->ifma_llifma = NULL; MPASS(ll_ifma->ifma_llifma == NULL); MPASS(ll_ifma->ifma_ifp == ifp); if (--ll_ifma->ifma_refcount == 0) { if (ll_ifma->ifma_flags & IFMA_F_ENQUEUED) { CK_STAILQ_REMOVE(&ifp->if_multiaddrs, ll_ifma, ifmultiaddr, ifma_link); ll_ifma->ifma_flags &= ~IFMA_F_ENQUEUED; } MCDPRINTF("removed ll_ifma: %p from %s\n", ll_ifma, ifp->if_xname); if_freemulti(ll_ifma); ifma_restart = true; } } } void inm_release_deferred(struct in_multi *inm) { struct in_multi_head tmp; IN_MULTI_LIST_LOCK_ASSERT(); MPASS(inm->inm_refcount > 0); if (--inm->inm_refcount == 0) { SLIST_INIT(&tmp); inm_disconnect(inm); inm->inm_ifma->ifma_protospec = NULL; SLIST_INSERT_HEAD(&tmp, inm, inm_nrele); inm_release_list_deferred(&tmp); } } static void -inm_release_task(void *arg __unused) +inm_release_task(void *arg __unused, int pending __unused) { struct in_multi_head inm_free_tmp; struct in_multi *inm, *tinm; SLIST_INIT(&inm_free_tmp); mtx_lock(&in_multi_free_mtx); SLIST_CONCAT(&inm_free_tmp, &inm_free_list, in_multi, inm_nrele); mtx_unlock(&in_multi_free_mtx); IN_MULTI_LOCK(); SLIST_FOREACH_SAFE(inm, &inm_free_tmp, inm_nrele, tinm) { SLIST_REMOVE_HEAD(&inm_free_tmp, inm_nrele); MPASS(inm); inm_release(inm); } IN_MULTI_UNLOCK(); } /* * Initialize an in_mfilter structure to a known state at t0, t1 * with an empty source filter list. */ static __inline void imf_init(struct in_mfilter *imf, const int st0, const int st1) { memset(imf, 0, sizeof(struct in_mfilter)); RB_INIT(&imf->imf_sources); imf->imf_st[0] = st0; imf->imf_st[1] = st1; } struct in_mfilter * ip_mfilter_alloc(const int mflags, const int st0, const int st1) { struct in_mfilter *imf; imf = malloc(sizeof(*imf), M_INMFILTER, mflags); if (imf != NULL) imf_init(imf, st0, st1); return (imf); } void ip_mfilter_free(struct in_mfilter *imf) { imf_purge(imf); free(imf, M_INMFILTER); } /* * Function for looking up an in_multi record for an IPv4 multicast address * on a given interface. ifp must be valid. If no record found, return NULL. * The IN_MULTI_LIST_LOCK and IF_ADDR_LOCK on ifp must be held. */ struct in_multi * inm_lookup_locked(struct ifnet *ifp, const struct in_addr ina) { struct ifmultiaddr *ifma; struct in_multi *inm; IN_MULTI_LIST_LOCK_ASSERT(); IF_ADDR_LOCK_ASSERT(ifp); inm = NULL; CK_STAILQ_FOREACH(ifma, &((ifp)->if_multiaddrs), ifma_link) { if (ifma->ifma_addr->sa_family != AF_INET || ifma->ifma_protospec == NULL) continue; inm = (struct in_multi *)ifma->ifma_protospec; if (inm->inm_addr.s_addr == ina.s_addr) break; inm = NULL; } return (inm); } /* * Wrapper for inm_lookup_locked(). * The IF_ADDR_LOCK will be taken on ifp and released on return. */ struct in_multi * inm_lookup(struct ifnet *ifp, const struct in_addr ina) { struct epoch_tracker et; struct in_multi *inm; IN_MULTI_LIST_LOCK_ASSERT(); NET_EPOCH_ENTER(et); inm = inm_lookup_locked(ifp, ina); NET_EPOCH_EXIT(et); return (inm); } /* * Find an IPv4 multicast group entry for this ip_moptions instance * which matches the specified group, and optionally an interface. * Return its index into the array, or -1 if not found. */ static struct in_mfilter * imo_match_group(const struct ip_moptions *imo, const struct ifnet *ifp, const struct sockaddr *group) { const struct sockaddr_in *gsin; struct in_mfilter *imf; struct in_multi *inm; gsin = (const struct sockaddr_in *)group; IP_MFILTER_FOREACH(imf, &imo->imo_head) { inm = imf->imf_inm; if (inm == NULL) continue; if ((ifp == NULL || (inm->inm_ifp == ifp)) && in_hosteq(inm->inm_addr, gsin->sin_addr)) { break; } } return (imf); } /* * Find an IPv4 multicast source entry for this imo which matches * the given group index for this socket, and source address. * * NOTE: This does not check if the entry is in-mode, merely if * it exists, which may not be the desired behaviour. */ static struct in_msource * imo_match_source(struct in_mfilter *imf, const struct sockaddr *src) { struct ip_msource find; struct ip_msource *ims; const sockunion_t *psa; KASSERT(src->sa_family == AF_INET, ("%s: !AF_INET", __func__)); /* Source trees are keyed in host byte order. */ psa = (const sockunion_t *)src; find.ims_haddr = ntohl(psa->sin.sin_addr.s_addr); ims = RB_FIND(ip_msource_tree, &imf->imf_sources, &find); return ((struct in_msource *)ims); } /* * Perform filtering for multicast datagrams on a socket by group and source. * * Returns 0 if a datagram should be allowed through, or various error codes * if the socket was not a member of the group, or the source was muted, etc. */ int imo_multi_filter(const struct ip_moptions *imo, const struct ifnet *ifp, const struct sockaddr *group, const struct sockaddr *src) { struct in_mfilter *imf; struct in_msource *ims; int mode; KASSERT(ifp != NULL, ("%s: null ifp", __func__)); imf = imo_match_group(imo, ifp, group); if (imf == NULL) return (MCAST_NOTGMEMBER); /* * Check if the source was included in an (S,G) join. * Allow reception on exclusive memberships by default, * reject reception on inclusive memberships by default. * Exclude source only if an in-mode exclude filter exists. * Include source only if an in-mode include filter exists. * NOTE: We are comparing group state here at IGMP t1 (now) * with socket-layer t0 (since last downcall). */ mode = imf->imf_st[1]; ims = imo_match_source(imf, src); if ((ims == NULL && mode == MCAST_INCLUDE) || (ims != NULL && ims->imsl_st[0] != mode)) return (MCAST_NOTSMEMBER); return (MCAST_PASS); } /* * Find and return a reference to an in_multi record for (ifp, group), * and bump its reference count. * If one does not exist, try to allocate it, and update link-layer multicast * filters on ifp to listen for group. * Assumes the IN_MULTI lock is held across the call. * Return 0 if successful, otherwise return an appropriate error code. */ static int in_getmulti(struct ifnet *ifp, const struct in_addr *group, struct in_multi **pinm) { struct sockaddr_in gsin; struct ifmultiaddr *ifma; struct in_ifinfo *ii; struct in_multi *inm; int error; IN_MULTI_LOCK_ASSERT(); ii = (struct in_ifinfo *)ifp->if_afdata[AF_INET]; IN_MULTI_LIST_LOCK(); inm = inm_lookup(ifp, *group); if (inm != NULL) { /* * If we already joined this group, just bump the * refcount and return it. */ KASSERT(inm->inm_refcount >= 1, ("%s: bad refcount %d", __func__, inm->inm_refcount)); inm_acquire_locked(inm); *pinm = inm; } IN_MULTI_LIST_UNLOCK(); if (inm != NULL) return (0); memset(&gsin, 0, sizeof(gsin)); gsin.sin_family = AF_INET; gsin.sin_len = sizeof(struct sockaddr_in); gsin.sin_addr = *group; /* * Check if a link-layer group is already associated * with this network-layer group on the given ifnet. */ error = if_addmulti(ifp, (struct sockaddr *)&gsin, &ifma); if (error != 0) return (error); /* XXX ifma_protospec must be covered by IF_ADDR_LOCK */ IN_MULTI_LIST_LOCK(); IF_ADDR_WLOCK(ifp); /* * If something other than netinet is occupying the link-layer * group, print a meaningful error message and back out of * the allocation. * Otherwise, bump the refcount on the existing network-layer * group association and return it. */ if (ifma->ifma_protospec != NULL) { inm = (struct in_multi *)ifma->ifma_protospec; #ifdef INVARIANTS KASSERT(ifma->ifma_addr != NULL, ("%s: no ifma_addr", __func__)); KASSERT(ifma->ifma_addr->sa_family == AF_INET, ("%s: ifma not AF_INET", __func__)); KASSERT(inm != NULL, ("%s: no ifma_protospec", __func__)); if (inm->inm_ifma != ifma || inm->inm_ifp != ifp || !in_hosteq(inm->inm_addr, *group)) { char addrbuf[INET_ADDRSTRLEN]; panic("%s: ifma %p is inconsistent with %p (%s)", __func__, ifma, inm, inet_ntoa_r(*group, addrbuf)); } #endif inm_acquire_locked(inm); *pinm = inm; goto out_locked; } IF_ADDR_WLOCK_ASSERT(ifp); /* * A new in_multi record is needed; allocate and initialize it. * We DO NOT perform an IGMP join as the in_ layer may need to * push an initial source list down to IGMP to support SSM. * * The initial source filter state is INCLUDE, {} as per the RFC. */ inm = malloc(sizeof(*inm), M_IPMADDR, M_NOWAIT | M_ZERO); if (inm == NULL) { IF_ADDR_WUNLOCK(ifp); IN_MULTI_LIST_UNLOCK(); if_delmulti_ifma(ifma); return (ENOMEM); } inm->inm_addr = *group; inm->inm_ifp = ifp; inm->inm_igi = ii->ii_igmp; inm->inm_ifma = ifma; inm->inm_refcount = 1; inm->inm_state = IGMP_NOT_MEMBER; mbufq_init(&inm->inm_scq, IGMP_MAX_STATE_CHANGES); inm->inm_st[0].iss_fmode = MCAST_UNDEFINED; inm->inm_st[1].iss_fmode = MCAST_UNDEFINED; RB_INIT(&inm->inm_srcs); ifma->ifma_protospec = inm; *pinm = inm; out_locked: IF_ADDR_WUNLOCK(ifp); IN_MULTI_LIST_UNLOCK(); return (0); } /* * Drop a reference to an in_multi record. * * If the refcount drops to 0, free the in_multi record and * delete the underlying link-layer membership. */ static void inm_release(struct in_multi *inm) { struct ifmultiaddr *ifma; struct ifnet *ifp; CTR2(KTR_IGMPV3, "%s: refcount is %d", __func__, inm->inm_refcount); MPASS(inm->inm_refcount == 0); CTR2(KTR_IGMPV3, "%s: freeing inm %p", __func__, inm); ifma = inm->inm_ifma; ifp = inm->inm_ifp; /* XXX this access is not covered by IF_ADDR_LOCK */ CTR2(KTR_IGMPV3, "%s: purging ifma %p", __func__, ifma); if (ifp != NULL) { CURVNET_SET(ifp->if_vnet); inm_purge(inm); free(inm, M_IPMADDR); if_delmulti_ifma_flags(ifma, 1); CURVNET_RESTORE(); if_rele(ifp); } else { inm_purge(inm); free(inm, M_IPMADDR); if_delmulti_ifma_flags(ifma, 1); } } /* * Clear recorded source entries for a group. * Used by the IGMP code. Caller must hold the IN_MULTI lock. * FIXME: Should reap. */ void inm_clear_recorded(struct in_multi *inm) { struct ip_msource *ims; IN_MULTI_LIST_LOCK_ASSERT(); RB_FOREACH(ims, ip_msource_tree, &inm->inm_srcs) { if (ims->ims_stp) { ims->ims_stp = 0; --inm->inm_st[1].iss_rec; } } KASSERT(inm->inm_st[1].iss_rec == 0, ("%s: iss_rec %d not 0", __func__, inm->inm_st[1].iss_rec)); } /* * Record a source as pending for a Source-Group IGMPv3 query. * This lives here as it modifies the shared tree. * * inm is the group descriptor. * naddr is the address of the source to record in network-byte order. * * If the net.inet.igmp.sgalloc sysctl is non-zero, we will * lazy-allocate a source node in response to an SG query. * Otherwise, no allocation is performed. This saves some memory * with the trade-off that the source will not be reported to the * router if joined in the window between the query response and * the group actually being joined on the local host. * * VIMAGE: XXX: Currently the igmp_sgalloc feature has been removed. * This turns off the allocation of a recorded source entry if * the group has not been joined. * * Return 0 if the source didn't exist or was already marked as recorded. * Return 1 if the source was marked as recorded by this function. * Return <0 if any error occurred (negated errno code). */ int inm_record_source(struct in_multi *inm, const in_addr_t naddr) { struct ip_msource find; struct ip_msource *ims, *nims; IN_MULTI_LIST_LOCK_ASSERT(); find.ims_haddr = ntohl(naddr); ims = RB_FIND(ip_msource_tree, &inm->inm_srcs, &find); if (ims && ims->ims_stp) return (0); if (ims == NULL) { if (inm->inm_nsrc == in_mcast_maxgrpsrc) return (-ENOSPC); nims = malloc(sizeof(struct ip_msource), M_IPMSOURCE, M_NOWAIT | M_ZERO); if (nims == NULL) return (-ENOMEM); nims->ims_haddr = find.ims_haddr; RB_INSERT(ip_msource_tree, &inm->inm_srcs, nims); ++inm->inm_nsrc; ims = nims; } /* * Mark the source as recorded and update the recorded * source count. */ ++ims->ims_stp; ++inm->inm_st[1].iss_rec; return (1); } /* * Return a pointer to an in_msource owned by an in_mfilter, * given its source address. * Lazy-allocate if needed. If this is a new entry its filter state is * undefined at t0. * * imf is the filter set being modified. * haddr is the source address in *host* byte-order. * * SMPng: May be called with locks held; malloc must not block. */ static int imf_get_source(struct in_mfilter *imf, const struct sockaddr_in *psin, struct in_msource **plims) { struct ip_msource find; struct ip_msource *ims, *nims; struct in_msource *lims; int error; error = 0; ims = NULL; lims = NULL; /* key is host byte order */ find.ims_haddr = ntohl(psin->sin_addr.s_addr); ims = RB_FIND(ip_msource_tree, &imf->imf_sources, &find); lims = (struct in_msource *)ims; if (lims == NULL) { if (imf->imf_nsrc == in_mcast_maxsocksrc) return (ENOSPC); nims = malloc(sizeof(struct in_msource), M_INMFILTER, M_NOWAIT | M_ZERO); if (nims == NULL) return (ENOMEM); lims = (struct in_msource *)nims; lims->ims_haddr = find.ims_haddr; lims->imsl_st[0] = MCAST_UNDEFINED; RB_INSERT(ip_msource_tree, &imf->imf_sources, nims); ++imf->imf_nsrc; } *plims = lims; return (error); } /* * Graft a source entry into an existing socket-layer filter set, * maintaining any required invariants and checking allocations. * * The source is marked as being in the new filter mode at t1. * * Return the pointer to the new node, otherwise return NULL. */ static struct in_msource * imf_graft(struct in_mfilter *imf, const uint8_t st1, const struct sockaddr_in *psin) { struct ip_msource *nims; struct in_msource *lims; nims = malloc(sizeof(struct in_msource), M_INMFILTER, M_NOWAIT | M_ZERO); if (nims == NULL) return (NULL); lims = (struct in_msource *)nims; lims->ims_haddr = ntohl(psin->sin_addr.s_addr); lims->imsl_st[0] = MCAST_UNDEFINED; lims->imsl_st[1] = st1; RB_INSERT(ip_msource_tree, &imf->imf_sources, nims); ++imf->imf_nsrc; return (lims); } /* * Prune a source entry from an existing socket-layer filter set, * maintaining any required invariants and checking allocations. * * The source is marked as being left at t1, it is not freed. * * Return 0 if no error occurred, otherwise return an errno value. */ static int imf_prune(struct in_mfilter *imf, const struct sockaddr_in *psin) { struct ip_msource find; struct ip_msource *ims; struct in_msource *lims; /* key is host byte order */ find.ims_haddr = ntohl(psin->sin_addr.s_addr); ims = RB_FIND(ip_msource_tree, &imf->imf_sources, &find); if (ims == NULL) return (ENOENT); lims = (struct in_msource *)ims; lims->imsl_st[1] = MCAST_UNDEFINED; return (0); } /* * Revert socket-layer filter set deltas at t1 to t0 state. */ static void imf_rollback(struct in_mfilter *imf) { struct ip_msource *ims, *tims; struct in_msource *lims; RB_FOREACH_SAFE(ims, ip_msource_tree, &imf->imf_sources, tims) { lims = (struct in_msource *)ims; if (lims->imsl_st[0] == lims->imsl_st[1]) { /* no change at t1 */ continue; } else if (lims->imsl_st[0] != MCAST_UNDEFINED) { /* revert change to existing source at t1 */ lims->imsl_st[1] = lims->imsl_st[0]; } else { /* revert source added t1 */ CTR2(KTR_IGMPV3, "%s: free ims %p", __func__, ims); RB_REMOVE(ip_msource_tree, &imf->imf_sources, ims); free(ims, M_INMFILTER); imf->imf_nsrc--; } } imf->imf_st[1] = imf->imf_st[0]; } /* * Mark socket-layer filter set as INCLUDE {} at t1. */ static void imf_leave(struct in_mfilter *imf) { struct ip_msource *ims; struct in_msource *lims; RB_FOREACH(ims, ip_msource_tree, &imf->imf_sources) { lims = (struct in_msource *)ims; lims->imsl_st[1] = MCAST_UNDEFINED; } imf->imf_st[1] = MCAST_INCLUDE; } /* * Mark socket-layer filter set deltas as committed. */ static void imf_commit(struct in_mfilter *imf) { struct ip_msource *ims; struct in_msource *lims; RB_FOREACH(ims, ip_msource_tree, &imf->imf_sources) { lims = (struct in_msource *)ims; lims->imsl_st[0] = lims->imsl_st[1]; } imf->imf_st[0] = imf->imf_st[1]; } /* * Reap unreferenced sources from socket-layer filter set. */ static void imf_reap(struct in_mfilter *imf) { struct ip_msource *ims, *tims; struct in_msource *lims; RB_FOREACH_SAFE(ims, ip_msource_tree, &imf->imf_sources, tims) { lims = (struct in_msource *)ims; if ((lims->imsl_st[0] == MCAST_UNDEFINED) && (lims->imsl_st[1] == MCAST_UNDEFINED)) { CTR2(KTR_IGMPV3, "%s: free lims %p", __func__, ims); RB_REMOVE(ip_msource_tree, &imf->imf_sources, ims); free(ims, M_INMFILTER); imf->imf_nsrc--; } } } /* * Purge socket-layer filter set. */ static void imf_purge(struct in_mfilter *imf) { struct ip_msource *ims, *tims; RB_FOREACH_SAFE(ims, ip_msource_tree, &imf->imf_sources, tims) { CTR2(KTR_IGMPV3, "%s: free ims %p", __func__, ims); RB_REMOVE(ip_msource_tree, &imf->imf_sources, ims); free(ims, M_INMFILTER); imf->imf_nsrc--; } imf->imf_st[0] = imf->imf_st[1] = MCAST_UNDEFINED; KASSERT(RB_EMPTY(&imf->imf_sources), ("%s: imf_sources not empty", __func__)); } /* * Look up a source filter entry for a multicast group. * * inm is the group descriptor to work with. * haddr is the host-byte-order IPv4 address to look up. * noalloc may be non-zero to suppress allocation of sources. * *pims will be set to the address of the retrieved or allocated source. * * SMPng: NOTE: may be called with locks held. * Return 0 if successful, otherwise return a non-zero error code. */ static int inm_get_source(struct in_multi *inm, const in_addr_t haddr, const int noalloc, struct ip_msource **pims) { struct ip_msource find; struct ip_msource *ims, *nims; find.ims_haddr = haddr; ims = RB_FIND(ip_msource_tree, &inm->inm_srcs, &find); if (ims == NULL && !noalloc) { if (inm->inm_nsrc == in_mcast_maxgrpsrc) return (ENOSPC); nims = malloc(sizeof(struct ip_msource), M_IPMSOURCE, M_NOWAIT | M_ZERO); if (nims == NULL) return (ENOMEM); nims->ims_haddr = haddr; RB_INSERT(ip_msource_tree, &inm->inm_srcs, nims); ++inm->inm_nsrc; ims = nims; #ifdef KTR CTR3(KTR_IGMPV3, "%s: allocated 0x%08x as %p", __func__, haddr, ims); #endif } *pims = ims; return (0); } /* * Merge socket-layer source into IGMP-layer source. * If rollback is non-zero, perform the inverse of the merge. */ static void ims_merge(struct ip_msource *ims, const struct in_msource *lims, const int rollback) { int n = rollback ? -1 : 1; if (lims->imsl_st[0] == MCAST_EXCLUDE) { CTR3(KTR_IGMPV3, "%s: t1 ex -= %d on 0x%08x", __func__, n, ims->ims_haddr); ims->ims_st[1].ex -= n; } else if (lims->imsl_st[0] == MCAST_INCLUDE) { CTR3(KTR_IGMPV3, "%s: t1 in -= %d on 0x%08x", __func__, n, ims->ims_haddr); ims->ims_st[1].in -= n; } if (lims->imsl_st[1] == MCAST_EXCLUDE) { CTR3(KTR_IGMPV3, "%s: t1 ex += %d on 0x%08x", __func__, n, ims->ims_haddr); ims->ims_st[1].ex += n; } else if (lims->imsl_st[1] == MCAST_INCLUDE) { CTR3(KTR_IGMPV3, "%s: t1 in += %d on 0x%08x", __func__, n, ims->ims_haddr); ims->ims_st[1].in += n; } } /* * Atomically update the global in_multi state, when a membership's * filter list is being updated in any way. * * imf is the per-inpcb-membership group filter pointer. * A fake imf may be passed for in-kernel consumers. * * XXX This is a candidate for a set-symmetric-difference style loop * which would eliminate the repeated lookup from root of ims nodes, * as they share the same key space. * * If any error occurred this function will back out of refcounts * and return a non-zero value. */ static int inm_merge(struct in_multi *inm, /*const*/ struct in_mfilter *imf) { struct ip_msource *ims, *nims; struct in_msource *lims; int schanged, error; int nsrc0, nsrc1; schanged = 0; error = 0; nsrc1 = nsrc0 = 0; IN_MULTI_LIST_LOCK_ASSERT(); /* * Update the source filters first, as this may fail. * Maintain count of in-mode filters at t0, t1. These are * used to work out if we transition into ASM mode or not. * Maintain a count of source filters whose state was * actually modified by this operation. */ RB_FOREACH(ims, ip_msource_tree, &imf->imf_sources) { lims = (struct in_msource *)ims; if (lims->imsl_st[0] == imf->imf_st[0]) nsrc0++; if (lims->imsl_st[1] == imf->imf_st[1]) nsrc1++; if (lims->imsl_st[0] == lims->imsl_st[1]) continue; error = inm_get_source(inm, lims->ims_haddr, 0, &nims); ++schanged; if (error) break; ims_merge(nims, lims, 0); } if (error) { struct ip_msource *bims; RB_FOREACH_REVERSE_FROM(ims, ip_msource_tree, nims) { lims = (struct in_msource *)ims; if (lims->imsl_st[0] == lims->imsl_st[1]) continue; (void)inm_get_source(inm, lims->ims_haddr, 1, &bims); if (bims == NULL) continue; ims_merge(bims, lims, 1); } goto out_reap; } CTR3(KTR_IGMPV3, "%s: imf filters in-mode: %d at t0, %d at t1", __func__, nsrc0, nsrc1); /* Handle transition between INCLUDE {n} and INCLUDE {} on socket. */ if (imf->imf_st[0] == imf->imf_st[1] && imf->imf_st[1] == MCAST_INCLUDE) { if (nsrc1 == 0) { CTR1(KTR_IGMPV3, "%s: --in on inm at t1", __func__); --inm->inm_st[1].iss_in; } } /* Handle filter mode transition on socket. */ if (imf->imf_st[0] != imf->imf_st[1]) { CTR3(KTR_IGMPV3, "%s: imf transition %d to %d", __func__, imf->imf_st[0], imf->imf_st[1]); if (imf->imf_st[0] == MCAST_EXCLUDE) { CTR1(KTR_IGMPV3, "%s: --ex on inm at t1", __func__); --inm->inm_st[1].iss_ex; } else if (imf->imf_st[0] == MCAST_INCLUDE) { CTR1(KTR_IGMPV3, "%s: --in on inm at t1", __func__); --inm->inm_st[1].iss_in; } if (imf->imf_st[1] == MCAST_EXCLUDE) { CTR1(KTR_IGMPV3, "%s: ex++ on inm at t1", __func__); inm->inm_st[1].iss_ex++; } else if (imf->imf_st[1] == MCAST_INCLUDE && nsrc1 > 0) { CTR1(KTR_IGMPV3, "%s: in++ on inm at t1", __func__); inm->inm_st[1].iss_in++; } } /* * Track inm filter state in terms of listener counts. * If there are any exclusive listeners, stack-wide * membership is exclusive. * Otherwise, if only inclusive listeners, stack-wide is inclusive. * If no listeners remain, state is undefined at t1, * and the IGMP lifecycle for this group should finish. */ if (inm->inm_st[1].iss_ex > 0) { CTR1(KTR_IGMPV3, "%s: transition to EX", __func__); inm->inm_st[1].iss_fmode = MCAST_EXCLUDE; } else if (inm->inm_st[1].iss_in > 0) { CTR1(KTR_IGMPV3, "%s: transition to IN", __func__); inm->inm_st[1].iss_fmode = MCAST_INCLUDE; } else { CTR1(KTR_IGMPV3, "%s: transition to UNDEF", __func__); inm->inm_st[1].iss_fmode = MCAST_UNDEFINED; } /* Decrement ASM listener count on transition out of ASM mode. */ if (imf->imf_st[0] == MCAST_EXCLUDE && nsrc0 == 0) { if ((imf->imf_st[1] != MCAST_EXCLUDE) || (imf->imf_st[1] == MCAST_EXCLUDE && nsrc1 > 0)) { CTR1(KTR_IGMPV3, "%s: --asm on inm at t1", __func__); --inm->inm_st[1].iss_asm; } } /* Increment ASM listener count on transition to ASM mode. */ if (imf->imf_st[1] == MCAST_EXCLUDE && nsrc1 == 0) { CTR1(KTR_IGMPV3, "%s: asm++ on inm at t1", __func__); inm->inm_st[1].iss_asm++; } CTR3(KTR_IGMPV3, "%s: merged imf %p to inm %p", __func__, imf, inm); inm_print(inm); out_reap: if (schanged > 0) { CTR1(KTR_IGMPV3, "%s: sources changed; reaping", __func__); inm_reap(inm); } return (error); } /* * Mark an in_multi's filter set deltas as committed. * Called by IGMP after a state change has been enqueued. */ void inm_commit(struct in_multi *inm) { struct ip_msource *ims; CTR2(KTR_IGMPV3, "%s: commit inm %p", __func__, inm); CTR1(KTR_IGMPV3, "%s: pre commit:", __func__); inm_print(inm); RB_FOREACH(ims, ip_msource_tree, &inm->inm_srcs) { ims->ims_st[0] = ims->ims_st[1]; } inm->inm_st[0] = inm->inm_st[1]; } /* * Reap unreferenced nodes from an in_multi's filter set. */ static void inm_reap(struct in_multi *inm) { struct ip_msource *ims, *tims; RB_FOREACH_SAFE(ims, ip_msource_tree, &inm->inm_srcs, tims) { if (ims->ims_st[0].ex > 0 || ims->ims_st[0].in > 0 || ims->ims_st[1].ex > 0 || ims->ims_st[1].in > 0 || ims->ims_stp != 0) continue; CTR2(KTR_IGMPV3, "%s: free ims %p", __func__, ims); RB_REMOVE(ip_msource_tree, &inm->inm_srcs, ims); free(ims, M_IPMSOURCE); inm->inm_nsrc--; } } /* * Purge all source nodes from an in_multi's filter set. */ static void inm_purge(struct in_multi *inm) { struct ip_msource *ims, *tims; RB_FOREACH_SAFE(ims, ip_msource_tree, &inm->inm_srcs, tims) { CTR2(KTR_IGMPV3, "%s: free ims %p", __func__, ims); RB_REMOVE(ip_msource_tree, &inm->inm_srcs, ims); free(ims, M_IPMSOURCE); inm->inm_nsrc--; } } /* * Join a multicast group; unlocked entry point. * * SMPng: XXX: in_joingroup() is called from in_control() when Giant * is not held. Fortunately, ifp is unlikely to have been detached * at this point, so we assume it's OK to recurse. */ int in_joingroup(struct ifnet *ifp, const struct in_addr *gina, /*const*/ struct in_mfilter *imf, struct in_multi **pinm) { int error; IN_MULTI_LOCK(); error = in_joingroup_locked(ifp, gina, imf, pinm); IN_MULTI_UNLOCK(); return (error); } /* * Join a multicast group; real entry point. * * Only preserves atomicity at inm level. * NOTE: imf argument cannot be const due to sys/tree.h limitations. * * If the IGMP downcall fails, the group is not joined, and an error * code is returned. */ int in_joingroup_locked(struct ifnet *ifp, const struct in_addr *gina, /*const*/ struct in_mfilter *imf, struct in_multi **pinm) { struct in_mfilter timf; struct in_multi *inm; int error; IN_MULTI_LOCK_ASSERT(); IN_MULTI_LIST_UNLOCK_ASSERT(); CTR4(KTR_IGMPV3, "%s: join 0x%08x on %p(%s))", __func__, ntohl(gina->s_addr), ifp, ifp->if_xname); error = 0; inm = NULL; /* * If no imf was specified (i.e. kernel consumer), * fake one up and assume it is an ASM join. */ if (imf == NULL) { imf_init(&timf, MCAST_UNDEFINED, MCAST_EXCLUDE); imf = &timf; } error = in_getmulti(ifp, gina, &inm); if (error) { CTR1(KTR_IGMPV3, "%s: in_getmulti() failure", __func__); return (error); } IN_MULTI_LIST_LOCK(); CTR1(KTR_IGMPV3, "%s: merge inm state", __func__); error = inm_merge(inm, imf); if (error) { CTR1(KTR_IGMPV3, "%s: failed to merge inm state", __func__); goto out_inm_release; } CTR1(KTR_IGMPV3, "%s: doing igmp downcall", __func__); error = igmp_change_state(inm); if (error) { CTR1(KTR_IGMPV3, "%s: failed to update source", __func__); goto out_inm_release; } out_inm_release: if (error) { CTR2(KTR_IGMPV3, "%s: dropping ref on %p", __func__, inm); IF_ADDR_WLOCK(ifp); inm_release_deferred(inm); IF_ADDR_WUNLOCK(ifp); } else { *pinm = inm; } IN_MULTI_LIST_UNLOCK(); return (error); } /* * Leave a multicast group; unlocked entry point. */ int in_leavegroup(struct in_multi *inm, /*const*/ struct in_mfilter *imf) { int error; IN_MULTI_LOCK(); error = in_leavegroup_locked(inm, imf); IN_MULTI_UNLOCK(); return (error); } /* * Leave a multicast group; real entry point. * All source filters will be expunged. * * Only preserves atomicity at inm level. * * Holding the write lock for the INP which contains imf * is highly advisable. We can't assert for it as imf does not * contain a back-pointer to the owning inp. * * Note: This is not the same as inm_release(*) as this function also * makes a state change downcall into IGMP. */ int in_leavegroup_locked(struct in_multi *inm, /*const*/ struct in_mfilter *imf) { struct in_mfilter timf; int error; IN_MULTI_LOCK_ASSERT(); IN_MULTI_LIST_UNLOCK_ASSERT(); error = 0; CTR5(KTR_IGMPV3, "%s: leave inm %p, 0x%08x/%s, imf %p", __func__, inm, ntohl(inm->inm_addr.s_addr), (inm_is_ifp_detached(inm) ? "null" : inm->inm_ifp->if_xname), imf); /* * If no imf was specified (i.e. kernel consumer), * fake one up and assume it is an ASM join. */ if (imf == NULL) { imf_init(&timf, MCAST_EXCLUDE, MCAST_UNDEFINED); imf = &timf; } /* * Begin state merge transaction at IGMP layer. * * As this particular invocation should not cause any memory * to be allocated, and there is no opportunity to roll back * the transaction, it MUST NOT fail. */ CTR1(KTR_IGMPV3, "%s: merge inm state", __func__); IN_MULTI_LIST_LOCK(); error = inm_merge(inm, imf); KASSERT(error == 0, ("%s: failed to merge inm state", __func__)); CTR1(KTR_IGMPV3, "%s: doing igmp downcall", __func__); CURVNET_SET(inm->inm_ifp->if_vnet); error = igmp_change_state(inm); IF_ADDR_WLOCK(inm->inm_ifp); inm_release_deferred(inm); IF_ADDR_WUNLOCK(inm->inm_ifp); IN_MULTI_LIST_UNLOCK(); CURVNET_RESTORE(); if (error) CTR1(KTR_IGMPV3, "%s: failed igmp downcall", __func__); CTR2(KTR_IGMPV3, "%s: dropping ref on %p", __func__, inm); return (error); } /*#ifndef BURN_BRIDGES*/ /* * Join an IPv4 multicast group in (*,G) exclusive mode. * The group must be a 224.0.0.0/24 link-scope group. * This KPI is for legacy kernel consumers only. */ struct in_multi * in_addmulti(struct in_addr *ap, struct ifnet *ifp) { struct in_multi *pinm; int error; #ifdef INVARIANTS char addrbuf[INET_ADDRSTRLEN]; #endif KASSERT(IN_LOCAL_GROUP(ntohl(ap->s_addr)), ("%s: %s not in 224.0.0.0/24", __func__, inet_ntoa_r(*ap, addrbuf))); error = in_joingroup(ifp, ap, NULL, &pinm); if (error != 0) pinm = NULL; return (pinm); } /* * Block or unblock an ASM multicast source on an inpcb. * This implements the delta-based API described in RFC 3678. * * The delta-based API applies only to exclusive-mode memberships. * An IGMP downcall will be performed. * * SMPng: NOTE: Must take Giant as a join may create a new ifma. * * Return 0 if successful, otherwise return an appropriate error code. */ static int inp_block_unblock_source(struct inpcb *inp, struct sockopt *sopt) { struct group_source_req gsr; struct rm_priotracker in_ifa_tracker; sockunion_t *gsa, *ssa; struct ifnet *ifp; struct in_mfilter *imf; struct ip_moptions *imo; struct in_msource *ims; struct in_multi *inm; uint16_t fmode; int error, doblock; ifp = NULL; error = 0; doblock = 0; memset(&gsr, 0, sizeof(struct group_source_req)); gsa = (sockunion_t *)&gsr.gsr_group; ssa = (sockunion_t *)&gsr.gsr_source; switch (sopt->sopt_name) { case IP_BLOCK_SOURCE: case IP_UNBLOCK_SOURCE: { struct ip_mreq_source mreqs; error = sooptcopyin(sopt, &mreqs, sizeof(struct ip_mreq_source), sizeof(struct ip_mreq_source)); if (error) return (error); gsa->sin.sin_family = AF_INET; gsa->sin.sin_len = sizeof(struct sockaddr_in); gsa->sin.sin_addr = mreqs.imr_multiaddr; ssa->sin.sin_family = AF_INET; ssa->sin.sin_len = sizeof(struct sockaddr_in); ssa->sin.sin_addr = mreqs.imr_sourceaddr; if (!in_nullhost(mreqs.imr_interface)) { IN_IFADDR_RLOCK(&in_ifa_tracker); INADDR_TO_IFP(mreqs.imr_interface, ifp); IN_IFADDR_RUNLOCK(&in_ifa_tracker); } if (sopt->sopt_name == IP_BLOCK_SOURCE) doblock = 1; CTR3(KTR_IGMPV3, "%s: imr_interface = 0x%08x, ifp = %p", __func__, ntohl(mreqs.imr_interface.s_addr), ifp); break; } case MCAST_BLOCK_SOURCE: case MCAST_UNBLOCK_SOURCE: error = sooptcopyin(sopt, &gsr, sizeof(struct group_source_req), sizeof(struct group_source_req)); if (error) return (error); if (gsa->sin.sin_family != AF_INET || gsa->sin.sin_len != sizeof(struct sockaddr_in)) return (EINVAL); if (ssa->sin.sin_family != AF_INET || ssa->sin.sin_len != sizeof(struct sockaddr_in)) return (EINVAL); if (gsr.gsr_interface == 0 || V_if_index < gsr.gsr_interface) return (EADDRNOTAVAIL); ifp = ifnet_byindex(gsr.gsr_interface); if (sopt->sopt_name == MCAST_BLOCK_SOURCE) doblock = 1; break; default: CTR2(KTR_IGMPV3, "%s: unknown sopt_name %d", __func__, sopt->sopt_name); return (EOPNOTSUPP); break; } if (!IN_MULTICAST(ntohl(gsa->sin.sin_addr.s_addr))) return (EINVAL); IN_MULTI_LOCK(); /* * Check if we are actually a member of this group. */ imo = inp_findmoptions(inp); imf = imo_match_group(imo, ifp, &gsa->sa); if (imf == NULL) { error = EADDRNOTAVAIL; goto out_inp_locked; } inm = imf->imf_inm; /* * Attempting to use the delta-based API on an * non exclusive-mode membership is an error. */ fmode = imf->imf_st[0]; if (fmode != MCAST_EXCLUDE) { error = EINVAL; goto out_inp_locked; } /* * Deal with error cases up-front: * Asked to block, but already blocked; or * Asked to unblock, but nothing to unblock. * If adding a new block entry, allocate it. */ ims = imo_match_source(imf, &ssa->sa); if ((ims != NULL && doblock) || (ims == NULL && !doblock)) { CTR3(KTR_IGMPV3, "%s: source 0x%08x %spresent", __func__, ntohl(ssa->sin.sin_addr.s_addr), doblock ? "" : "not "); error = EADDRNOTAVAIL; goto out_inp_locked; } INP_WLOCK_ASSERT(inp); /* * Begin state merge transaction at socket layer. */ if (doblock) { CTR2(KTR_IGMPV3, "%s: %s source", __func__, "block"); ims = imf_graft(imf, fmode, &ssa->sin); if (ims == NULL) error = ENOMEM; } else { CTR2(KTR_IGMPV3, "%s: %s source", __func__, "allow"); error = imf_prune(imf, &ssa->sin); } if (error) { CTR1(KTR_IGMPV3, "%s: merge imf state failed", __func__); goto out_imf_rollback; } /* * Begin state merge transaction at IGMP layer. */ CTR1(KTR_IGMPV3, "%s: merge inm state", __func__); IN_MULTI_LIST_LOCK(); error = inm_merge(inm, imf); if (error) { CTR1(KTR_IGMPV3, "%s: failed to merge inm state", __func__); IN_MULTI_LIST_UNLOCK(); goto out_imf_rollback; } CTR1(KTR_IGMPV3, "%s: doing igmp downcall", __func__); error = igmp_change_state(inm); IN_MULTI_LIST_UNLOCK(); if (error) CTR1(KTR_IGMPV3, "%s: failed igmp downcall", __func__); out_imf_rollback: if (error) imf_rollback(imf); else imf_commit(imf); imf_reap(imf); out_inp_locked: INP_WUNLOCK(inp); IN_MULTI_UNLOCK(); return (error); } /* * Given an inpcb, return its multicast options structure pointer. Accepts * an unlocked inpcb pointer, but will return it locked. May sleep. * * SMPng: NOTE: Potentially calls malloc(M_WAITOK) with Giant held. * SMPng: NOTE: Returns with the INP write lock held. */ static struct ip_moptions * inp_findmoptions(struct inpcb *inp) { struct ip_moptions *imo; INP_WLOCK(inp); if (inp->inp_moptions != NULL) return (inp->inp_moptions); INP_WUNLOCK(inp); imo = malloc(sizeof(*imo), M_IPMOPTS, M_WAITOK); imo->imo_multicast_ifp = NULL; imo->imo_multicast_addr.s_addr = INADDR_ANY; imo->imo_multicast_vif = -1; imo->imo_multicast_ttl = IP_DEFAULT_MULTICAST_TTL; imo->imo_multicast_loop = in_mcast_loop; STAILQ_INIT(&imo->imo_head); INP_WLOCK(inp); if (inp->inp_moptions != NULL) { free(imo, M_IPMOPTS); return (inp->inp_moptions); } inp->inp_moptions = imo; return (imo); } static void inp_gcmoptions(struct ip_moptions *imo) { struct in_mfilter *imf; struct in_multi *inm; struct ifnet *ifp; while ((imf = ip_mfilter_first(&imo->imo_head)) != NULL) { ip_mfilter_remove(&imo->imo_head, imf); imf_leave(imf); if ((inm = imf->imf_inm) != NULL) { if ((ifp = inm->inm_ifp) != NULL) { CURVNET_SET(ifp->if_vnet); (void)in_leavegroup(inm, imf); CURVNET_RESTORE(); } else { (void)in_leavegroup(inm, imf); } } ip_mfilter_free(imf); } free(imo, M_IPMOPTS); } /* * Discard the IP multicast options (and source filters). To minimize * the amount of work done while holding locks such as the INP's * pcbinfo lock (which is used in the receive path), the free * operation is deferred to the epoch callback task. */ void inp_freemoptions(struct ip_moptions *imo) { if (imo == NULL) return; inp_gcmoptions(imo); } /* * Atomically get source filters on a socket for an IPv4 multicast group. * Called with INP lock held; returns with lock released. */ static int inp_get_source_filters(struct inpcb *inp, struct sockopt *sopt) { struct __msfilterreq msfr; sockunion_t *gsa; struct ifnet *ifp; struct ip_moptions *imo; struct in_mfilter *imf; struct ip_msource *ims; struct in_msource *lims; struct sockaddr_in *psin; struct sockaddr_storage *ptss; struct sockaddr_storage *tss; int error; size_t nsrcs, ncsrcs; INP_WLOCK_ASSERT(inp); imo = inp->inp_moptions; KASSERT(imo != NULL, ("%s: null ip_moptions", __func__)); INP_WUNLOCK(inp); error = sooptcopyin(sopt, &msfr, sizeof(struct __msfilterreq), sizeof(struct __msfilterreq)); if (error) return (error); if (msfr.msfr_ifindex == 0 || V_if_index < msfr.msfr_ifindex) return (EINVAL); ifp = ifnet_byindex(msfr.msfr_ifindex); if (ifp == NULL) return (EINVAL); INP_WLOCK(inp); /* * Lookup group on the socket. */ gsa = (sockunion_t *)&msfr.msfr_group; imf = imo_match_group(imo, ifp, &gsa->sa); if (imf == NULL) { INP_WUNLOCK(inp); return (EADDRNOTAVAIL); } /* * Ignore memberships which are in limbo. */ if (imf->imf_st[1] == MCAST_UNDEFINED) { INP_WUNLOCK(inp); return (EAGAIN); } msfr.msfr_fmode = imf->imf_st[1]; /* * If the user specified a buffer, copy out the source filter * entries to userland gracefully. * We only copy out the number of entries which userland * has asked for, but we always tell userland how big the * buffer really needs to be. */ if (msfr.msfr_nsrcs > in_mcast_maxsocksrc) msfr.msfr_nsrcs = in_mcast_maxsocksrc; tss = NULL; if (msfr.msfr_srcs != NULL && msfr.msfr_nsrcs > 0) { tss = malloc(sizeof(struct sockaddr_storage) * msfr.msfr_nsrcs, M_TEMP, M_NOWAIT | M_ZERO); if (tss == NULL) { INP_WUNLOCK(inp); return (ENOBUFS); } } /* * Count number of sources in-mode at t0. * If buffer space exists and remains, copy out source entries. */ nsrcs = msfr.msfr_nsrcs; ncsrcs = 0; ptss = tss; RB_FOREACH(ims, ip_msource_tree, &imf->imf_sources) { lims = (struct in_msource *)ims; if (lims->imsl_st[0] == MCAST_UNDEFINED || lims->imsl_st[0] != imf->imf_st[0]) continue; ++ncsrcs; if (tss != NULL && nsrcs > 0) { psin = (struct sockaddr_in *)ptss; psin->sin_family = AF_INET; psin->sin_len = sizeof(struct sockaddr_in); psin->sin_addr.s_addr = htonl(lims->ims_haddr); psin->sin_port = 0; ++ptss; --nsrcs; } } INP_WUNLOCK(inp); if (tss != NULL) { error = copyout(tss, msfr.msfr_srcs, sizeof(struct sockaddr_storage) * msfr.msfr_nsrcs); free(tss, M_TEMP); if (error) return (error); } msfr.msfr_nsrcs = ncsrcs; error = sooptcopyout(sopt, &msfr, sizeof(struct __msfilterreq)); return (error); } /* * Return the IP multicast options in response to user getsockopt(). */ int inp_getmoptions(struct inpcb *inp, struct sockopt *sopt) { struct rm_priotracker in_ifa_tracker; struct ip_mreqn mreqn; struct ip_moptions *imo; struct ifnet *ifp; struct in_ifaddr *ia; int error, optval; u_char coptval; INP_WLOCK(inp); imo = inp->inp_moptions; /* * If socket is neither of type SOCK_RAW or SOCK_DGRAM, * or is a divert socket, reject it. */ if (inp->inp_socket->so_proto->pr_protocol == IPPROTO_DIVERT || (inp->inp_socket->so_proto->pr_type != SOCK_RAW && inp->inp_socket->so_proto->pr_type != SOCK_DGRAM)) { INP_WUNLOCK(inp); return (EOPNOTSUPP); } error = 0; switch (sopt->sopt_name) { case IP_MULTICAST_VIF: if (imo != NULL) optval = imo->imo_multicast_vif; else optval = -1; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof(int)); break; case IP_MULTICAST_IF: memset(&mreqn, 0, sizeof(struct ip_mreqn)); if (imo != NULL) { ifp = imo->imo_multicast_ifp; if (!in_nullhost(imo->imo_multicast_addr)) { mreqn.imr_address = imo->imo_multicast_addr; } else if (ifp != NULL) { struct epoch_tracker et; mreqn.imr_ifindex = ifp->if_index; NET_EPOCH_ENTER(et); IFP_TO_IA(ifp, ia, &in_ifa_tracker); if (ia != NULL) mreqn.imr_address = IA_SIN(ia)->sin_addr; NET_EPOCH_EXIT(et); } } INP_WUNLOCK(inp); if (sopt->sopt_valsize == sizeof(struct ip_mreqn)) { error = sooptcopyout(sopt, &mreqn, sizeof(struct ip_mreqn)); } else { error = sooptcopyout(sopt, &mreqn.imr_address, sizeof(struct in_addr)); } break; case IP_MULTICAST_TTL: if (imo == NULL) optval = coptval = IP_DEFAULT_MULTICAST_TTL; else optval = coptval = imo->imo_multicast_ttl; INP_WUNLOCK(inp); if (sopt->sopt_valsize == sizeof(u_char)) error = sooptcopyout(sopt, &coptval, sizeof(u_char)); else error = sooptcopyout(sopt, &optval, sizeof(int)); break; case IP_MULTICAST_LOOP: if (imo == NULL) optval = coptval = IP_DEFAULT_MULTICAST_LOOP; else optval = coptval = imo->imo_multicast_loop; INP_WUNLOCK(inp); if (sopt->sopt_valsize == sizeof(u_char)) error = sooptcopyout(sopt, &coptval, sizeof(u_char)); else error = sooptcopyout(sopt, &optval, sizeof(int)); break; case IP_MSFILTER: if (imo == NULL) { error = EADDRNOTAVAIL; INP_WUNLOCK(inp); } else { error = inp_get_source_filters(inp, sopt); } break; default: INP_WUNLOCK(inp); error = ENOPROTOOPT; break; } INP_UNLOCK_ASSERT(inp); return (error); } /* * Look up the ifnet to use for a multicast group membership, * given the IPv4 address of an interface, and the IPv4 group address. * * This routine exists to support legacy multicast applications * which do not understand that multicast memberships are scoped to * specific physical links in the networking stack, or which need * to join link-scope groups before IPv4 addresses are configured. * * If inp is non-NULL, use this socket's current FIB number for any * required FIB lookup. * If ina is INADDR_ANY, look up the group address in the unicast FIB, * and use its ifp; usually, this points to the default next-hop. * * If the FIB lookup fails, attempt to use the first non-loopback * interface with multicast capability in the system as a * last resort. The legacy IPv4 ASM API requires that we do * this in order to allow groups to be joined when the routing * table has not yet been populated during boot. * * Returns NULL if no ifp could be found. * * FUTURE: Implement IPv4 source-address selection. */ static struct ifnet * inp_lookup_mcast_ifp(const struct inpcb *inp, const struct sockaddr_in *gsin, const struct in_addr ina) { struct rm_priotracker in_ifa_tracker; struct ifnet *ifp; struct nhop4_basic nh4; uint32_t fibnum; KASSERT(gsin->sin_family == AF_INET, ("%s: not AF_INET", __func__)); KASSERT(IN_MULTICAST(ntohl(gsin->sin_addr.s_addr)), ("%s: not multicast", __func__)); ifp = NULL; if (!in_nullhost(ina)) { IN_IFADDR_RLOCK(&in_ifa_tracker); INADDR_TO_IFP(ina, ifp); IN_IFADDR_RUNLOCK(&in_ifa_tracker); } else { fibnum = inp ? inp->inp_inc.inc_fibnum : 0; if (fib4_lookup_nh_basic(fibnum, gsin->sin_addr, 0, 0, &nh4)==0) ifp = nh4.nh_ifp; else { struct in_ifaddr *ia; struct ifnet *mifp; mifp = NULL; IN_IFADDR_RLOCK(&in_ifa_tracker); CK_STAILQ_FOREACH(ia, &V_in_ifaddrhead, ia_link) { mifp = ia->ia_ifp; if (!(mifp->if_flags & IFF_LOOPBACK) && (mifp->if_flags & IFF_MULTICAST)) { ifp = mifp; break; } } IN_IFADDR_RUNLOCK(&in_ifa_tracker); } } return (ifp); } /* * Join an IPv4 multicast group, possibly with a source. */ static int inp_join_group(struct inpcb *inp, struct sockopt *sopt) { struct group_source_req gsr; sockunion_t *gsa, *ssa; struct ifnet *ifp; struct in_mfilter *imf; struct ip_moptions *imo; struct in_multi *inm; struct in_msource *lims; int error, is_new; ifp = NULL; lims = NULL; error = 0; memset(&gsr, 0, sizeof(struct group_source_req)); gsa = (sockunion_t *)&gsr.gsr_group; gsa->ss.ss_family = AF_UNSPEC; ssa = (sockunion_t *)&gsr.gsr_source; ssa->ss.ss_family = AF_UNSPEC; switch (sopt->sopt_name) { case IP_ADD_MEMBERSHIP: { struct ip_mreqn mreqn; if (sopt->sopt_valsize == sizeof(struct ip_mreqn)) error = sooptcopyin(sopt, &mreqn, sizeof(struct ip_mreqn), sizeof(struct ip_mreqn)); else error = sooptcopyin(sopt, &mreqn, sizeof(struct ip_mreq), sizeof(struct ip_mreq)); if (error) return (error); gsa->sin.sin_family = AF_INET; gsa->sin.sin_len = sizeof(struct sockaddr_in); gsa->sin.sin_addr = mreqn.imr_multiaddr; if (!IN_MULTICAST(ntohl(gsa->sin.sin_addr.s_addr))) return (EINVAL); if (sopt->sopt_valsize == sizeof(struct ip_mreqn) && mreqn.imr_ifindex != 0) ifp = ifnet_byindex(mreqn.imr_ifindex); else ifp = inp_lookup_mcast_ifp(inp, &gsa->sin, mreqn.imr_address); break; } case IP_ADD_SOURCE_MEMBERSHIP: { struct ip_mreq_source mreqs; error = sooptcopyin(sopt, &mreqs, sizeof(struct ip_mreq_source), sizeof(struct ip_mreq_source)); if (error) return (error); gsa->sin.sin_family = ssa->sin.sin_family = AF_INET; gsa->sin.sin_len = ssa->sin.sin_len = sizeof(struct sockaddr_in); gsa->sin.sin_addr = mreqs.imr_multiaddr; if (!IN_MULTICAST(ntohl(gsa->sin.sin_addr.s_addr))) return (EINVAL); ssa->sin.sin_addr = mreqs.imr_sourceaddr; ifp = inp_lookup_mcast_ifp(inp, &gsa->sin, mreqs.imr_interface); CTR3(KTR_IGMPV3, "%s: imr_interface = 0x%08x, ifp = %p", __func__, ntohl(mreqs.imr_interface.s_addr), ifp); break; } case MCAST_JOIN_GROUP: case MCAST_JOIN_SOURCE_GROUP: if (sopt->sopt_name == MCAST_JOIN_GROUP) { error = sooptcopyin(sopt, &gsr, sizeof(struct group_req), sizeof(struct group_req)); } else if (sopt->sopt_name == MCAST_JOIN_SOURCE_GROUP) { error = sooptcopyin(sopt, &gsr, sizeof(struct group_source_req), sizeof(struct group_source_req)); } if (error) return (error); if (gsa->sin.sin_family != AF_INET || gsa->sin.sin_len != sizeof(struct sockaddr_in)) return (EINVAL); /* * Overwrite the port field if present, as the sockaddr * being copied in may be matched with a binary comparison. */ gsa->sin.sin_port = 0; if (sopt->sopt_name == MCAST_JOIN_SOURCE_GROUP) { if (ssa->sin.sin_family != AF_INET || ssa->sin.sin_len != sizeof(struct sockaddr_in)) return (EINVAL); ssa->sin.sin_port = 0; } if (!IN_MULTICAST(ntohl(gsa->sin.sin_addr.s_addr))) return (EINVAL); if (gsr.gsr_interface == 0 || V_if_index < gsr.gsr_interface) return (EADDRNOTAVAIL); ifp = ifnet_byindex(gsr.gsr_interface); break; default: CTR2(KTR_IGMPV3, "%s: unknown sopt_name %d", __func__, sopt->sopt_name); return (EOPNOTSUPP); break; } if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0) return (EADDRNOTAVAIL); IN_MULTI_LOCK(); /* * Find the membership in the membership list. */ imo = inp_findmoptions(inp); imf = imo_match_group(imo, ifp, &gsa->sa); if (imf == NULL) { is_new = 1; inm = NULL; if (ip_mfilter_count(&imo->imo_head) >= IP_MAX_MEMBERSHIPS) { error = ENOMEM; goto out_inp_locked; } } else { is_new = 0; inm = imf->imf_inm; if (ssa->ss.ss_family != AF_UNSPEC) { /* * MCAST_JOIN_SOURCE_GROUP on an exclusive membership * is an error. On an existing inclusive membership, * it just adds the source to the filter list. */ if (imf->imf_st[1] != MCAST_INCLUDE) { error = EINVAL; goto out_inp_locked; } /* * Throw out duplicates. * * XXX FIXME: This makes a naive assumption that * even if entries exist for *ssa in this imf, * they will be rejected as dupes, even if they * are not valid in the current mode (in-mode). * * in_msource is transactioned just as for anything * else in SSM -- but note naive use of inm_graft() * below for allocating new filter entries. * * This is only an issue if someone mixes the * full-state SSM API with the delta-based API, * which is discouraged in the relevant RFCs. */ lims = imo_match_source(imf, &ssa->sa); if (lims != NULL /*&& lims->imsl_st[1] == MCAST_INCLUDE*/) { error = EADDRNOTAVAIL; goto out_inp_locked; } } else { /* * MCAST_JOIN_GROUP on an existing exclusive * membership is an error; return EADDRINUSE * to preserve 4.4BSD API idempotence, and * avoid tedious detour to code below. * NOTE: This is bending RFC 3678 a bit. * * On an existing inclusive membership, this is also * an error; if you want to change filter mode, * you must use the userland API setsourcefilter(). * XXX We don't reject this for imf in UNDEFINED * state at t1, because allocation of a filter * is atomic with allocation of a membership. */ error = EINVAL; if (imf->imf_st[1] == MCAST_EXCLUDE) error = EADDRINUSE; goto out_inp_locked; } } /* * Begin state merge transaction at socket layer. */ INP_WLOCK_ASSERT(inp); /* * Graft new source into filter list for this inpcb's * membership of the group. The in_multi may not have * been allocated yet if this is a new membership, however, * the in_mfilter slot will be allocated and must be initialized. * * Note: Grafting of exclusive mode filters doesn't happen * in this path. * XXX: Should check for non-NULL lims (node exists but may * not be in-mode) for interop with full-state API. */ if (ssa->ss.ss_family != AF_UNSPEC) { /* Membership starts in IN mode */ if (is_new) { CTR1(KTR_IGMPV3, "%s: new join w/source", __func__); imf = ip_mfilter_alloc(M_NOWAIT, MCAST_UNDEFINED, MCAST_INCLUDE); if (imf == NULL) { error = ENOMEM; goto out_inp_locked; } } else { CTR2(KTR_IGMPV3, "%s: %s source", __func__, "allow"); } lims = imf_graft(imf, MCAST_INCLUDE, &ssa->sin); if (lims == NULL) { CTR1(KTR_IGMPV3, "%s: merge imf state failed", __func__); error = ENOMEM; goto out_inp_locked; } } else { /* No address specified; Membership starts in EX mode */ if (is_new) { CTR1(KTR_IGMPV3, "%s: new join w/o source", __func__); imf = ip_mfilter_alloc(M_NOWAIT, MCAST_UNDEFINED, MCAST_EXCLUDE); if (imf == NULL) { error = ENOMEM; goto out_inp_locked; } } } /* * Begin state merge transaction at IGMP layer. */ if (is_new) { in_pcbref(inp); INP_WUNLOCK(inp); error = in_joingroup_locked(ifp, &gsa->sin.sin_addr, imf, &imf->imf_inm); INP_WLOCK(inp); if (in_pcbrele_wlocked(inp)) { error = ENXIO; goto out_inp_unlocked; } if (error) { CTR1(KTR_IGMPV3, "%s: in_joingroup_locked failed", __func__); goto out_inp_locked; } /* * NOTE: Refcount from in_joingroup_locked() * is protecting membership. */ ip_mfilter_insert(&imo->imo_head, imf); } else { CTR1(KTR_IGMPV3, "%s: merge inm state", __func__); IN_MULTI_LIST_LOCK(); error = inm_merge(inm, imf); if (error) { CTR1(KTR_IGMPV3, "%s: failed to merge inm state", __func__); IN_MULTI_LIST_UNLOCK(); imf_rollback(imf); imf_reap(imf); goto out_inp_locked; } CTR1(KTR_IGMPV3, "%s: doing igmp downcall", __func__); error = igmp_change_state(inm); IN_MULTI_LIST_UNLOCK(); if (error) { CTR1(KTR_IGMPV3, "%s: failed igmp downcall", __func__); imf_rollback(imf); imf_reap(imf); goto out_inp_locked; } } imf_commit(imf); imf = NULL; out_inp_locked: INP_WUNLOCK(inp); out_inp_unlocked: IN_MULTI_UNLOCK(); if (is_new && imf) { if (imf->imf_inm != NULL) { IN_MULTI_LIST_LOCK(); IF_ADDR_WLOCK(ifp); inm_release_deferred(imf->imf_inm); IF_ADDR_WUNLOCK(ifp); IN_MULTI_LIST_UNLOCK(); } ip_mfilter_free(imf); } return (error); } /* * Leave an IPv4 multicast group on an inpcb, possibly with a source. */ static int inp_leave_group(struct inpcb *inp, struct sockopt *sopt) { struct group_source_req gsr; struct ip_mreq_source mreqs; struct rm_priotracker in_ifa_tracker; sockunion_t *gsa, *ssa; struct ifnet *ifp; struct in_mfilter *imf; struct ip_moptions *imo; struct in_msource *ims; struct in_multi *inm; int error; bool is_final; ifp = NULL; error = 0; is_final = true; memset(&gsr, 0, sizeof(struct group_source_req)); gsa = (sockunion_t *)&gsr.gsr_group; gsa->ss.ss_family = AF_UNSPEC; ssa = (sockunion_t *)&gsr.gsr_source; ssa->ss.ss_family = AF_UNSPEC; switch (sopt->sopt_name) { case IP_DROP_MEMBERSHIP: case IP_DROP_SOURCE_MEMBERSHIP: if (sopt->sopt_name == IP_DROP_MEMBERSHIP) { error = sooptcopyin(sopt, &mreqs, sizeof(struct ip_mreq), sizeof(struct ip_mreq)); /* * Swap interface and sourceaddr arguments, * as ip_mreq and ip_mreq_source are laid * out differently. */ mreqs.imr_interface = mreqs.imr_sourceaddr; mreqs.imr_sourceaddr.s_addr = INADDR_ANY; } else if (sopt->sopt_name == IP_DROP_SOURCE_MEMBERSHIP) { error = sooptcopyin(sopt, &mreqs, sizeof(struct ip_mreq_source), sizeof(struct ip_mreq_source)); } if (error) return (error); gsa->sin.sin_family = AF_INET; gsa->sin.sin_len = sizeof(struct sockaddr_in); gsa->sin.sin_addr = mreqs.imr_multiaddr; if (sopt->sopt_name == IP_DROP_SOURCE_MEMBERSHIP) { ssa->sin.sin_family = AF_INET; ssa->sin.sin_len = sizeof(struct sockaddr_in); ssa->sin.sin_addr = mreqs.imr_sourceaddr; } /* * Attempt to look up hinted ifp from interface address. * Fallthrough with null ifp iff lookup fails, to * preserve 4.4BSD mcast API idempotence. * XXX NOTE WELL: The RFC 3678 API is preferred because * using an IPv4 address as a key is racy. */ if (!in_nullhost(mreqs.imr_interface)) { IN_IFADDR_RLOCK(&in_ifa_tracker); INADDR_TO_IFP(mreqs.imr_interface, ifp); IN_IFADDR_RUNLOCK(&in_ifa_tracker); } CTR3(KTR_IGMPV3, "%s: imr_interface = 0x%08x, ifp = %p", __func__, ntohl(mreqs.imr_interface.s_addr), ifp); break; case MCAST_LEAVE_GROUP: case MCAST_LEAVE_SOURCE_GROUP: if (sopt->sopt_name == MCAST_LEAVE_GROUP) { error = sooptcopyin(sopt, &gsr, sizeof(struct group_req), sizeof(struct group_req)); } else if (sopt->sopt_name == MCAST_LEAVE_SOURCE_GROUP) { error = sooptcopyin(sopt, &gsr, sizeof(struct group_source_req), sizeof(struct group_source_req)); } if (error) return (error); if (gsa->sin.sin_family != AF_INET || gsa->sin.sin_len != sizeof(struct sockaddr_in)) return (EINVAL); if (sopt->sopt_name == MCAST_LEAVE_SOURCE_GROUP) { if (ssa->sin.sin_family != AF_INET || ssa->sin.sin_len != sizeof(struct sockaddr_in)) return (EINVAL); } if (gsr.gsr_interface == 0 || V_if_index < gsr.gsr_interface) return (EADDRNOTAVAIL); ifp = ifnet_byindex(gsr.gsr_interface); if (ifp == NULL) return (EADDRNOTAVAIL); break; default: CTR2(KTR_IGMPV3, "%s: unknown sopt_name %d", __func__, sopt->sopt_name); return (EOPNOTSUPP); break; } if (!IN_MULTICAST(ntohl(gsa->sin.sin_addr.s_addr))) return (EINVAL); IN_MULTI_LOCK(); /* * Find the membership in the membership list. */ imo = inp_findmoptions(inp); imf = imo_match_group(imo, ifp, &gsa->sa); if (imf == NULL) { error = EADDRNOTAVAIL; goto out_inp_locked; } inm = imf->imf_inm; if (ssa->ss.ss_family != AF_UNSPEC) is_final = false; /* * Begin state merge transaction at socket layer. */ INP_WLOCK_ASSERT(inp); /* * If we were instructed only to leave a given source, do so. * MCAST_LEAVE_SOURCE_GROUP is only valid for inclusive memberships. */ if (is_final) { ip_mfilter_remove(&imo->imo_head, imf); imf_leave(imf); /* * Give up the multicast address record to which * the membership points. */ (void) in_leavegroup_locked(imf->imf_inm, imf); } else { if (imf->imf_st[0] == MCAST_EXCLUDE) { error = EADDRNOTAVAIL; goto out_inp_locked; } ims = imo_match_source(imf, &ssa->sa); if (ims == NULL) { CTR3(KTR_IGMPV3, "%s: source 0x%08x %spresent", __func__, ntohl(ssa->sin.sin_addr.s_addr), "not "); error = EADDRNOTAVAIL; goto out_inp_locked; } CTR2(KTR_IGMPV3, "%s: %s source", __func__, "block"); error = imf_prune(imf, &ssa->sin); if (error) { CTR1(KTR_IGMPV3, "%s: merge imf state failed", __func__); goto out_inp_locked; } } /* * Begin state merge transaction at IGMP layer. */ if (!is_final) { CTR1(KTR_IGMPV3, "%s: merge inm state", __func__); IN_MULTI_LIST_LOCK(); error = inm_merge(inm, imf); if (error) { CTR1(KTR_IGMPV3, "%s: failed to merge inm state", __func__); IN_MULTI_LIST_UNLOCK(); imf_rollback(imf); imf_reap(imf); goto out_inp_locked; } CTR1(KTR_IGMPV3, "%s: doing igmp downcall", __func__); error = igmp_change_state(inm); IN_MULTI_LIST_UNLOCK(); if (error) { CTR1(KTR_IGMPV3, "%s: failed igmp downcall", __func__); imf_rollback(imf); imf_reap(imf); goto out_inp_locked; } } imf_commit(imf); imf_reap(imf); out_inp_locked: INP_WUNLOCK(inp); if (is_final && imf) ip_mfilter_free(imf); IN_MULTI_UNLOCK(); return (error); } /* * Select the interface for transmitting IPv4 multicast datagrams. * * Either an instance of struct in_addr or an instance of struct ip_mreqn * may be passed to this socket option. An address of INADDR_ANY or an * interface index of 0 is used to remove a previous selection. * When no interface is selected, one is chosen for every send. */ static int inp_set_multicast_if(struct inpcb *inp, struct sockopt *sopt) { struct rm_priotracker in_ifa_tracker; struct in_addr addr; struct ip_mreqn mreqn; struct ifnet *ifp; struct ip_moptions *imo; int error; if (sopt->sopt_valsize == sizeof(struct ip_mreqn)) { /* * An interface index was specified using the * Linux-derived ip_mreqn structure. */ error = sooptcopyin(sopt, &mreqn, sizeof(struct ip_mreqn), sizeof(struct ip_mreqn)); if (error) return (error); if (mreqn.imr_ifindex < 0 || V_if_index < mreqn.imr_ifindex) return (EINVAL); if (mreqn.imr_ifindex == 0) { ifp = NULL; } else { ifp = ifnet_byindex(mreqn.imr_ifindex); if (ifp == NULL) return (EADDRNOTAVAIL); } } else { /* * An interface was specified by IPv4 address. * This is the traditional BSD usage. */ error = sooptcopyin(sopt, &addr, sizeof(struct in_addr), sizeof(struct in_addr)); if (error) return (error); if (in_nullhost(addr)) { ifp = NULL; } else { IN_IFADDR_RLOCK(&in_ifa_tracker); INADDR_TO_IFP(addr, ifp); IN_IFADDR_RUNLOCK(&in_ifa_tracker); if (ifp == NULL) return (EADDRNOTAVAIL); } CTR3(KTR_IGMPV3, "%s: ifp = %p, addr = 0x%08x", __func__, ifp, ntohl(addr.s_addr)); } /* Reject interfaces which do not support multicast. */ if (ifp != NULL && (ifp->if_flags & IFF_MULTICAST) == 0) return (EOPNOTSUPP); imo = inp_findmoptions(inp); imo->imo_multicast_ifp = ifp; imo->imo_multicast_addr.s_addr = INADDR_ANY; INP_WUNLOCK(inp); return (0); } /* * Atomically set source filters on a socket for an IPv4 multicast group. * * SMPng: NOTE: Potentially calls malloc(M_WAITOK) with Giant held. */ static int inp_set_source_filters(struct inpcb *inp, struct sockopt *sopt) { struct __msfilterreq msfr; sockunion_t *gsa; struct ifnet *ifp; struct in_mfilter *imf; struct ip_moptions *imo; struct in_multi *inm; int error; error = sooptcopyin(sopt, &msfr, sizeof(struct __msfilterreq), sizeof(struct __msfilterreq)); if (error) return (error); if (msfr.msfr_nsrcs > in_mcast_maxsocksrc) return (ENOBUFS); if ((msfr.msfr_fmode != MCAST_EXCLUDE && msfr.msfr_fmode != MCAST_INCLUDE)) return (EINVAL); if (msfr.msfr_group.ss_family != AF_INET || msfr.msfr_group.ss_len != sizeof(struct sockaddr_in)) return (EINVAL); gsa = (sockunion_t *)&msfr.msfr_group; if (!IN_MULTICAST(ntohl(gsa->sin.sin_addr.s_addr))) return (EINVAL); gsa->sin.sin_port = 0; /* ignore port */ if (msfr.msfr_ifindex == 0 || V_if_index < msfr.msfr_ifindex) return (EADDRNOTAVAIL); ifp = ifnet_byindex(msfr.msfr_ifindex); if (ifp == NULL) return (EADDRNOTAVAIL); IN_MULTI_LOCK(); /* * Take the INP write lock. * Check if this socket is a member of this group. */ imo = inp_findmoptions(inp); imf = imo_match_group(imo, ifp, &gsa->sa); if (imf == NULL) { error = EADDRNOTAVAIL; goto out_inp_locked; } inm = imf->imf_inm; /* * Begin state merge transaction at socket layer. */ INP_WLOCK_ASSERT(inp); imf->imf_st[1] = msfr.msfr_fmode; /* * Apply any new source filters, if present. * Make a copy of the user-space source vector so * that we may copy them with a single copyin. This * allows us to deal with page faults up-front. */ if (msfr.msfr_nsrcs > 0) { struct in_msource *lims; struct sockaddr_in *psin; struct sockaddr_storage *kss, *pkss; int i; INP_WUNLOCK(inp); CTR2(KTR_IGMPV3, "%s: loading %lu source list entries", __func__, (unsigned long)msfr.msfr_nsrcs); kss = malloc(sizeof(struct sockaddr_storage) * msfr.msfr_nsrcs, M_TEMP, M_WAITOK); error = copyin(msfr.msfr_srcs, kss, sizeof(struct sockaddr_storage) * msfr.msfr_nsrcs); if (error) { free(kss, M_TEMP); return (error); } INP_WLOCK(inp); /* * Mark all source filters as UNDEFINED at t1. * Restore new group filter mode, as imf_leave() * will set it to INCLUDE. */ imf_leave(imf); imf->imf_st[1] = msfr.msfr_fmode; /* * Update socket layer filters at t1, lazy-allocating * new entries. This saves a bunch of memory at the * cost of one RB_FIND() per source entry; duplicate * entries in the msfr_nsrcs vector are ignored. * If we encounter an error, rollback transaction. * * XXX This too could be replaced with a set-symmetric * difference like loop to avoid walking from root * every time, as the key space is common. */ for (i = 0, pkss = kss; i < msfr.msfr_nsrcs; i++, pkss++) { psin = (struct sockaddr_in *)pkss; if (psin->sin_family != AF_INET) { error = EAFNOSUPPORT; break; } if (psin->sin_len != sizeof(struct sockaddr_in)) { error = EINVAL; break; } error = imf_get_source(imf, psin, &lims); if (error) break; lims->imsl_st[1] = imf->imf_st[1]; } free(kss, M_TEMP); } if (error) goto out_imf_rollback; INP_WLOCK_ASSERT(inp); /* * Begin state merge transaction at IGMP layer. */ CTR1(KTR_IGMPV3, "%s: merge inm state", __func__); IN_MULTI_LIST_LOCK(); error = inm_merge(inm, imf); if (error) { CTR1(KTR_IGMPV3, "%s: failed to merge inm state", __func__); IN_MULTI_LIST_UNLOCK(); goto out_imf_rollback; } CTR1(KTR_IGMPV3, "%s: doing igmp downcall", __func__); error = igmp_change_state(inm); IN_MULTI_LIST_UNLOCK(); if (error) CTR1(KTR_IGMPV3, "%s: failed igmp downcall", __func__); out_imf_rollback: if (error) imf_rollback(imf); else imf_commit(imf); imf_reap(imf); out_inp_locked: INP_WUNLOCK(inp); IN_MULTI_UNLOCK(); return (error); } /* * Set the IP multicast options in response to user setsockopt(). * * Many of the socket options handled in this function duplicate the * functionality of socket options in the regular unicast API. However, * it is not possible to merge the duplicate code, because the idempotence * of the IPv4 multicast part of the BSD Sockets API must be preserved; * the effects of these options must be treated as separate and distinct. * * SMPng: XXX: Unlocked read of inp_socket believed OK. * FUTURE: The IP_MULTICAST_VIF option may be eliminated if MROUTING * is refactored to no longer use vifs. */ int inp_setmoptions(struct inpcb *inp, struct sockopt *sopt) { struct ip_moptions *imo; int error; error = 0; /* * If socket is neither of type SOCK_RAW or SOCK_DGRAM, * or is a divert socket, reject it. */ if (inp->inp_socket->so_proto->pr_protocol == IPPROTO_DIVERT || (inp->inp_socket->so_proto->pr_type != SOCK_RAW && inp->inp_socket->so_proto->pr_type != SOCK_DGRAM)) return (EOPNOTSUPP); switch (sopt->sopt_name) { case IP_MULTICAST_VIF: { int vifi; /* * Select a multicast VIF for transmission. * Only useful if multicast forwarding is active. */ if (legal_vif_num == NULL) { error = EOPNOTSUPP; break; } error = sooptcopyin(sopt, &vifi, sizeof(int), sizeof(int)); if (error) break; if (!legal_vif_num(vifi) && (vifi != -1)) { error = EINVAL; break; } imo = inp_findmoptions(inp); imo->imo_multicast_vif = vifi; INP_WUNLOCK(inp); break; } case IP_MULTICAST_IF: error = inp_set_multicast_if(inp, sopt); break; case IP_MULTICAST_TTL: { u_char ttl; /* * Set the IP time-to-live for outgoing multicast packets. * The original multicast API required a char argument, * which is inconsistent with the rest of the socket API. * We allow either a char or an int. */ if (sopt->sopt_valsize == sizeof(u_char)) { error = sooptcopyin(sopt, &ttl, sizeof(u_char), sizeof(u_char)); if (error) break; } else { u_int ittl; error = sooptcopyin(sopt, &ittl, sizeof(u_int), sizeof(u_int)); if (error) break; if (ittl > 255) { error = EINVAL; break; } ttl = (u_char)ittl; } imo = inp_findmoptions(inp); imo->imo_multicast_ttl = ttl; INP_WUNLOCK(inp); break; } case IP_MULTICAST_LOOP: { u_char loop; /* * Set the loopback flag for outgoing multicast packets. * Must be zero or one. The original multicast API required a * char argument, which is inconsistent with the rest * of the socket API. We allow either a char or an int. */ if (sopt->sopt_valsize == sizeof(u_char)) { error = sooptcopyin(sopt, &loop, sizeof(u_char), sizeof(u_char)); if (error) break; } else { u_int iloop; error = sooptcopyin(sopt, &iloop, sizeof(u_int), sizeof(u_int)); if (error) break; loop = (u_char)iloop; } imo = inp_findmoptions(inp); imo->imo_multicast_loop = !!loop; INP_WUNLOCK(inp); break; } case IP_ADD_MEMBERSHIP: case IP_ADD_SOURCE_MEMBERSHIP: case MCAST_JOIN_GROUP: case MCAST_JOIN_SOURCE_GROUP: error = inp_join_group(inp, sopt); break; case IP_DROP_MEMBERSHIP: case IP_DROP_SOURCE_MEMBERSHIP: case MCAST_LEAVE_GROUP: case MCAST_LEAVE_SOURCE_GROUP: error = inp_leave_group(inp, sopt); break; case IP_BLOCK_SOURCE: case IP_UNBLOCK_SOURCE: case MCAST_BLOCK_SOURCE: case MCAST_UNBLOCK_SOURCE: error = inp_block_unblock_source(inp, sopt); break; case IP_MSFILTER: error = inp_set_source_filters(inp, sopt); break; default: error = EOPNOTSUPP; break; } INP_UNLOCK_ASSERT(inp); return (error); } /* * Expose IGMP's multicast filter mode and source list(s) to userland, * keyed by (ifindex, group). * The filter mode is written out as a uint32_t, followed by * 0..n of struct in_addr. * For use by ifmcstat(8). * SMPng: NOTE: unlocked read of ifindex space. */ static int sysctl_ip_mcast_filters(SYSCTL_HANDLER_ARGS) { struct in_addr src, group; struct epoch_tracker et; struct ifnet *ifp; struct ifmultiaddr *ifma; struct in_multi *inm; struct ip_msource *ims; int *name; int retval; u_int namelen; uint32_t fmode, ifindex; name = (int *)arg1; namelen = arg2; if (req->newptr != NULL) return (EPERM); if (namelen != 2) return (EINVAL); ifindex = name[0]; if (ifindex <= 0 || ifindex > V_if_index) { CTR2(KTR_IGMPV3, "%s: ifindex %u out of range", __func__, ifindex); return (ENOENT); } group.s_addr = name[1]; if (!IN_MULTICAST(ntohl(group.s_addr))) { CTR2(KTR_IGMPV3, "%s: group 0x%08x is not multicast", __func__, ntohl(group.s_addr)); return (EINVAL); } NET_EPOCH_ENTER(et); ifp = ifnet_byindex(ifindex); if (ifp == NULL) { NET_EPOCH_EXIT(et); CTR2(KTR_IGMPV3, "%s: no ifp for ifindex %u", __func__, ifindex); return (ENOENT); } retval = sysctl_wire_old_buffer(req, sizeof(uint32_t) + (in_mcast_maxgrpsrc * sizeof(struct in_addr))); if (retval) { NET_EPOCH_EXIT(et); return (retval); } IN_MULTI_LIST_LOCK(); CK_STAILQ_FOREACH(ifma, &ifp->if_multiaddrs, ifma_link) { if (ifma->ifma_addr->sa_family != AF_INET || ifma->ifma_protospec == NULL) continue; inm = (struct in_multi *)ifma->ifma_protospec; if (!in_hosteq(inm->inm_addr, group)) continue; fmode = inm->inm_st[1].iss_fmode; retval = SYSCTL_OUT(req, &fmode, sizeof(uint32_t)); if (retval != 0) break; RB_FOREACH(ims, ip_msource_tree, &inm->inm_srcs) { CTR2(KTR_IGMPV3, "%s: visit node 0x%08x", __func__, ims->ims_haddr); /* * Only copy-out sources which are in-mode. */ if (fmode != ims_get_mode(inm, ims, 1)) { CTR1(KTR_IGMPV3, "%s: skip non-in-mode", __func__); continue; } src.s_addr = htonl(ims->ims_haddr); retval = SYSCTL_OUT(req, &src, sizeof(struct in_addr)); if (retval != 0) break; } } IN_MULTI_LIST_UNLOCK(); NET_EPOCH_EXIT(et); return (retval); } #if defined(KTR) && (KTR_COMPILE & KTR_IGMPV3) static const char *inm_modestrs[] = { [MCAST_UNDEFINED] = "un", [MCAST_INCLUDE] = "in", [MCAST_EXCLUDE] = "ex", }; _Static_assert(MCAST_UNDEFINED == 0 && MCAST_EXCLUDE + 1 == nitems(inm_modestrs), "inm_modestrs: no longer matches #defines"); static const char * inm_mode_str(const int mode) { if (mode >= MCAST_UNDEFINED && mode <= MCAST_EXCLUDE) return (inm_modestrs[mode]); return ("??"); } static const char *inm_statestrs[] = { [IGMP_NOT_MEMBER] = "not-member", [IGMP_SILENT_MEMBER] = "silent", [IGMP_REPORTING_MEMBER] = "reporting", [IGMP_IDLE_MEMBER] = "idle", [IGMP_LAZY_MEMBER] = "lazy", [IGMP_SLEEPING_MEMBER] = "sleeping", [IGMP_AWAKENING_MEMBER] = "awakening", [IGMP_G_QUERY_PENDING_MEMBER] = "query-pending", [IGMP_SG_QUERY_PENDING_MEMBER] = "sg-query-pending", [IGMP_LEAVING_MEMBER] = "leaving", }; _Static_assert(IGMP_NOT_MEMBER == 0 && IGMP_LEAVING_MEMBER + 1 == nitems(inm_statestrs), "inm_statetrs: no longer matches #defines"); static const char * inm_state_str(const int state) { if (state >= IGMP_NOT_MEMBER && state <= IGMP_LEAVING_MEMBER) return (inm_statestrs[state]); return ("??"); } /* * Dump an in_multi structure to the console. */ void inm_print(const struct in_multi *inm) { int t; char addrbuf[INET_ADDRSTRLEN]; if ((ktr_mask & KTR_IGMPV3) == 0) return; printf("%s: --- begin inm %p ---\n", __func__, inm); printf("addr %s ifp %p(%s) ifma %p\n", inet_ntoa_r(inm->inm_addr, addrbuf), inm->inm_ifp, inm->inm_ifp->if_xname, inm->inm_ifma); printf("timer %u state %s refcount %u scq.len %u\n", inm->inm_timer, inm_state_str(inm->inm_state), inm->inm_refcount, inm->inm_scq.mq_len); printf("igi %p nsrc %lu sctimer %u scrv %u\n", inm->inm_igi, inm->inm_nsrc, inm->inm_sctimer, inm->inm_scrv); for (t = 0; t < 2; t++) { printf("t%d: fmode %s asm %u ex %u in %u rec %u\n", t, inm_mode_str(inm->inm_st[t].iss_fmode), inm->inm_st[t].iss_asm, inm->inm_st[t].iss_ex, inm->inm_st[t].iss_in, inm->inm_st[t].iss_rec); } printf("%s: --- end inm %p ---\n", __func__, inm); } #else /* !KTR || !(KTR_COMPILE & KTR_IGMPV3) */ void inm_print(const struct in_multi *inm) { } #endif /* KTR && (KTR_COMPILE & KTR_IGMPV3) */ RB_GENERATE(ip_msource_tree, ip_msource, ims_link, ip_msource_cmp); Index: head/sys/netinet6/in6_mcast.c =================================================================== --- head/sys/netinet6/in6_mcast.c (revision 359437) +++ head/sys/netinet6/in6_mcast.c (revision 359438) @@ -1,2911 +1,2901 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 2009 Bruce Simpson. * 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. * 3. The name of the author may not be used to endorse or promote * products derived from this software without specific prior written * permission. * * 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. */ /* * IPv6 multicast socket, group, and socket option processing module. * Normative references: RFC 2292, RFC 3492, RFC 3542, RFC 3678, RFC 3810. */ #include __FBSDID("$FreeBSD$"); #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 #include #include #include #include #ifndef KTR_MLD #define KTR_MLD KTR_INET6 #endif #ifndef __SOCKUNION_DECLARED union sockunion { struct sockaddr_storage ss; struct sockaddr sa; struct sockaddr_dl sdl; struct sockaddr_in6 sin6; }; typedef union sockunion sockunion_t; #define __SOCKUNION_DECLARED #endif /* __SOCKUNION_DECLARED */ static MALLOC_DEFINE(M_IN6MFILTER, "in6_mfilter", "IPv6 multicast PCB-layer source filter"); MALLOC_DEFINE(M_IP6MADDR, "in6_multi", "IPv6 multicast group"); static MALLOC_DEFINE(M_IP6MOPTS, "ip6_moptions", "IPv6 multicast options"); static MALLOC_DEFINE(M_IP6MSOURCE, "ip6_msource", "IPv6 multicast MLD-layer source filter"); RB_GENERATE(ip6_msource_tree, ip6_msource, im6s_link, ip6_msource_cmp); /* * Locking: * - Lock order is: Giant, IN6_MULTI_LOCK, INP_WLOCK, * IN6_MULTI_LIST_LOCK, MLD_LOCK, IF_ADDR_LOCK. * - The IF_ADDR_LOCK is implicitly taken by in6m_lookup() earlier, however * it can be taken by code in net/if.c also. * - ip6_moptions and in6_mfilter are covered by the INP_WLOCK. * * struct in6_multi is covered by IN6_MULTI_LOCK. There isn't strictly * any need for in6_multi itself to be virtualized -- it is bound to an ifp * anyway no matter what happens. */ struct mtx in6_multi_list_mtx; MTX_SYSINIT(in6_multi_mtx, &in6_multi_list_mtx, "in6_multi_list_mtx", MTX_DEF); struct mtx in6_multi_free_mtx; MTX_SYSINIT(in6_multi_free_mtx, &in6_multi_free_mtx, "in6_multi_free_mtx", MTX_DEF); struct sx in6_multi_sx; SX_SYSINIT(in6_multi_sx, &in6_multi_sx, "in6_multi_sx"); static void im6f_commit(struct in6_mfilter *); static int im6f_get_source(struct in6_mfilter *imf, const struct sockaddr_in6 *psin, struct in6_msource **); static struct in6_msource * im6f_graft(struct in6_mfilter *, const uint8_t, const struct sockaddr_in6 *); static void im6f_leave(struct in6_mfilter *); static int im6f_prune(struct in6_mfilter *, const struct sockaddr_in6 *); static void im6f_purge(struct in6_mfilter *); static void im6f_rollback(struct in6_mfilter *); static void im6f_reap(struct in6_mfilter *); static struct in6_mfilter * im6o_match_group(const struct ip6_moptions *, const struct ifnet *, const struct sockaddr *); static struct in6_msource * im6o_match_source(struct in6_mfilter *, const struct sockaddr *); static void im6s_merge(struct ip6_msource *ims, const struct in6_msource *lims, const int rollback); static int in6_getmulti(struct ifnet *, const struct in6_addr *, struct in6_multi **); static int in6_joingroup_locked(struct ifnet *, const struct in6_addr *, struct in6_mfilter *, struct in6_multi **, int); static int in6m_get_source(struct in6_multi *inm, const struct in6_addr *addr, const int noalloc, struct ip6_msource **pims); #ifdef KTR static int in6m_is_ifp_detached(const struct in6_multi *); #endif static int in6m_merge(struct in6_multi *, /*const*/ struct in6_mfilter *); static void in6m_purge(struct in6_multi *); static void in6m_reap(struct in6_multi *); static struct ip6_moptions * in6p_findmoptions(struct inpcb *); static int in6p_get_source_filters(struct inpcb *, struct sockopt *); static int in6p_join_group(struct inpcb *, struct sockopt *); static int in6p_leave_group(struct inpcb *, struct sockopt *); static struct ifnet * in6p_lookup_mcast_ifp(const struct inpcb *, const struct sockaddr_in6 *); static int in6p_block_unblock_source(struct inpcb *, struct sockopt *); static int in6p_set_multicast_if(struct inpcb *, struct sockopt *); static int in6p_set_source_filters(struct inpcb *, struct sockopt *); static int sysctl_ip6_mcast_filters(SYSCTL_HANDLER_ARGS); SYSCTL_DECL(_net_inet6_ip6); /* XXX Not in any common header. */ static SYSCTL_NODE(_net_inet6_ip6, OID_AUTO, mcast, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "IPv6 multicast"); static u_long in6_mcast_maxgrpsrc = IPV6_MAX_GROUP_SRC_FILTER; SYSCTL_ULONG(_net_inet6_ip6_mcast, OID_AUTO, maxgrpsrc, CTLFLAG_RWTUN, &in6_mcast_maxgrpsrc, 0, "Max source filters per group"); static u_long in6_mcast_maxsocksrc = IPV6_MAX_SOCK_SRC_FILTER; SYSCTL_ULONG(_net_inet6_ip6_mcast, OID_AUTO, maxsocksrc, CTLFLAG_RWTUN, &in6_mcast_maxsocksrc, 0, "Max source filters per socket"); /* TODO Virtualize this switch. */ int in6_mcast_loop = IPV6_DEFAULT_MULTICAST_LOOP; SYSCTL_INT(_net_inet6_ip6_mcast, OID_AUTO, loop, CTLFLAG_RWTUN, &in6_mcast_loop, 0, "Loopback multicast datagrams by default"); static SYSCTL_NODE(_net_inet6_ip6_mcast, OID_AUTO, filters, CTLFLAG_RD | CTLFLAG_MPSAFE, sysctl_ip6_mcast_filters, "Per-interface stack-wide source filters"); #ifdef KTR /* * Inline function which wraps assertions for a valid ifp. * The ifnet layer will set the ifma's ifp pointer to NULL if the ifp * is detached. */ static int __inline in6m_is_ifp_detached(const struct in6_multi *inm) { struct ifnet *ifp; KASSERT(inm->in6m_ifma != NULL, ("%s: no ifma", __func__)); ifp = inm->in6m_ifma->ifma_ifp; if (ifp != NULL) { /* * Sanity check that network-layer notion of ifp is the * same as that of link-layer. */ KASSERT(inm->in6m_ifp == ifp, ("%s: bad ifp", __func__)); } return (ifp == NULL); } #endif /* * Initialize an in6_mfilter structure to a known state at t0, t1 * with an empty source filter list. */ static __inline void im6f_init(struct in6_mfilter *imf, const int st0, const int st1) { memset(imf, 0, sizeof(struct in6_mfilter)); RB_INIT(&imf->im6f_sources); imf->im6f_st[0] = st0; imf->im6f_st[1] = st1; } struct in6_mfilter * ip6_mfilter_alloc(const int mflags, const int st0, const int st1) { struct in6_mfilter *imf; imf = malloc(sizeof(*imf), M_IN6MFILTER, mflags); if (imf != NULL) im6f_init(imf, st0, st1); return (imf); } void ip6_mfilter_free(struct in6_mfilter *imf) { im6f_purge(imf); free(imf, M_IN6MFILTER); } /* * Find an IPv6 multicast group entry for this ip6_moptions instance * which matches the specified group, and optionally an interface. * Return its index into the array, or -1 if not found. */ static struct in6_mfilter * im6o_match_group(const struct ip6_moptions *imo, const struct ifnet *ifp, const struct sockaddr *group) { const struct sockaddr_in6 *gsin6; struct in6_mfilter *imf; struct in6_multi *inm; gsin6 = (const struct sockaddr_in6 *)group; IP6_MFILTER_FOREACH(imf, &imo->im6o_head) { inm = imf->im6f_in6m; if (inm == NULL) continue; if ((ifp == NULL || (inm->in6m_ifp == ifp)) && IN6_ARE_ADDR_EQUAL(&inm->in6m_addr, &gsin6->sin6_addr)) { break; } } return (imf); } /* * Find an IPv6 multicast source entry for this imo which matches * the given group index for this socket, and source address. * * XXX TODO: The scope ID, if present in src, is stripped before * any comparison. We SHOULD enforce scope/zone checks where the source * filter entry has a link scope. * * NOTE: This does not check if the entry is in-mode, merely if * it exists, which may not be the desired behaviour. */ static struct in6_msource * im6o_match_source(struct in6_mfilter *imf, const struct sockaddr *src) { struct ip6_msource find; struct ip6_msource *ims; const sockunion_t *psa; KASSERT(src->sa_family == AF_INET6, ("%s: !AF_INET6", __func__)); psa = (const sockunion_t *)src; find.im6s_addr = psa->sin6.sin6_addr; in6_clearscope(&find.im6s_addr); /* XXX */ ims = RB_FIND(ip6_msource_tree, &imf->im6f_sources, &find); return ((struct in6_msource *)ims); } /* * Perform filtering for multicast datagrams on a socket by group and source. * * Returns 0 if a datagram should be allowed through, or various error codes * if the socket was not a member of the group, or the source was muted, etc. */ int im6o_mc_filter(const struct ip6_moptions *imo, const struct ifnet *ifp, const struct sockaddr *group, const struct sockaddr *src) { struct in6_mfilter *imf; struct in6_msource *ims; int mode; KASSERT(ifp != NULL, ("%s: null ifp", __func__)); imf = im6o_match_group(imo, ifp, group); if (imf == NULL) return (MCAST_NOTGMEMBER); /* * Check if the source was included in an (S,G) join. * Allow reception on exclusive memberships by default, * reject reception on inclusive memberships by default. * Exclude source only if an in-mode exclude filter exists. * Include source only if an in-mode include filter exists. * NOTE: We are comparing group state here at MLD t1 (now) * with socket-layer t0 (since last downcall). */ mode = imf->im6f_st[1]; ims = im6o_match_source(imf, src); if ((ims == NULL && mode == MCAST_INCLUDE) || (ims != NULL && ims->im6sl_st[0] != mode)) return (MCAST_NOTSMEMBER); return (MCAST_PASS); } /* * Find and return a reference to an in6_multi record for (ifp, group), * and bump its reference count. * If one does not exist, try to allocate it, and update link-layer multicast * filters on ifp to listen for group. * Assumes the IN6_MULTI lock is held across the call. * Return 0 if successful, otherwise return an appropriate error code. */ static int in6_getmulti(struct ifnet *ifp, const struct in6_addr *group, struct in6_multi **pinm) { struct epoch_tracker et; struct sockaddr_in6 gsin6; struct ifmultiaddr *ifma; struct in6_multi *inm; int error; error = 0; /* * XXX: Accesses to ifma_protospec must be covered by IF_ADDR_LOCK; * if_addmulti() takes this mutex itself, so we must drop and * re-acquire around the call. */ IN6_MULTI_LOCK_ASSERT(); IN6_MULTI_LIST_LOCK(); IF_ADDR_WLOCK(ifp); NET_EPOCH_ENTER(et); inm = in6m_lookup_locked(ifp, group); NET_EPOCH_EXIT(et); if (inm != NULL) { /* * If we already joined this group, just bump the * refcount and return it. */ KASSERT(inm->in6m_refcount >= 1, ("%s: bad refcount %d", __func__, inm->in6m_refcount)); in6m_acquire_locked(inm); *pinm = inm; goto out_locked; } memset(&gsin6, 0, sizeof(gsin6)); gsin6.sin6_family = AF_INET6; gsin6.sin6_len = sizeof(struct sockaddr_in6); gsin6.sin6_addr = *group; /* * Check if a link-layer group is already associated * with this network-layer group on the given ifnet. */ IN6_MULTI_LIST_UNLOCK(); IF_ADDR_WUNLOCK(ifp); error = if_addmulti(ifp, (struct sockaddr *)&gsin6, &ifma); if (error != 0) return (error); IN6_MULTI_LIST_LOCK(); IF_ADDR_WLOCK(ifp); /* * If something other than netinet6 is occupying the link-layer * group, print a meaningful error message and back out of * the allocation. * Otherwise, bump the refcount on the existing network-layer * group association and return it. */ if (ifma->ifma_protospec != NULL) { inm = (struct in6_multi *)ifma->ifma_protospec; #ifdef INVARIANTS KASSERT(ifma->ifma_addr != NULL, ("%s: no ifma_addr", __func__)); KASSERT(ifma->ifma_addr->sa_family == AF_INET6, ("%s: ifma not AF_INET6", __func__)); KASSERT(inm != NULL, ("%s: no ifma_protospec", __func__)); if (inm->in6m_ifma != ifma || inm->in6m_ifp != ifp || !IN6_ARE_ADDR_EQUAL(&inm->in6m_addr, group)) panic("%s: ifma %p is inconsistent with %p (%p)", __func__, ifma, inm, group); #endif in6m_acquire_locked(inm); *pinm = inm; goto out_locked; } IF_ADDR_WLOCK_ASSERT(ifp); /* * A new in6_multi record is needed; allocate and initialize it. * We DO NOT perform an MLD join as the in6_ layer may need to * push an initial source list down to MLD to support SSM. * * The initial source filter state is INCLUDE, {} as per the RFC. * Pending state-changes per group are subject to a bounds check. */ inm = malloc(sizeof(*inm), M_IP6MADDR, M_NOWAIT | M_ZERO); if (inm == NULL) { IN6_MULTI_LIST_UNLOCK(); IF_ADDR_WUNLOCK(ifp); if_delmulti_ifma(ifma); return (ENOMEM); } inm->in6m_addr = *group; inm->in6m_ifp = ifp; inm->in6m_mli = MLD_IFINFO(ifp); inm->in6m_ifma = ifma; inm->in6m_refcount = 1; inm->in6m_state = MLD_NOT_MEMBER; mbufq_init(&inm->in6m_scq, MLD_MAX_STATE_CHANGES); inm->in6m_st[0].iss_fmode = MCAST_UNDEFINED; inm->in6m_st[1].iss_fmode = MCAST_UNDEFINED; RB_INIT(&inm->in6m_srcs); ifma->ifma_protospec = inm; *pinm = inm; out_locked: IN6_MULTI_LIST_UNLOCK(); IF_ADDR_WUNLOCK(ifp); return (error); } /* * Drop a reference to an in6_multi record. * * If the refcount drops to 0, free the in6_multi record and * delete the underlying link-layer membership. */ static void in6m_release(struct in6_multi *inm) { struct ifmultiaddr *ifma; struct ifnet *ifp; CTR2(KTR_MLD, "%s: refcount is %d", __func__, inm->in6m_refcount); MPASS(inm->in6m_refcount == 0); CTR2(KTR_MLD, "%s: freeing inm %p", __func__, inm); ifma = inm->in6m_ifma; ifp = inm->in6m_ifp; MPASS(ifma->ifma_llifma == NULL); /* XXX this access is not covered by IF_ADDR_LOCK */ CTR2(KTR_MLD, "%s: purging ifma %p", __func__, ifma); KASSERT(ifma->ifma_protospec == NULL, ("%s: ifma_protospec != NULL", __func__)); if (ifp == NULL) ifp = ifma->ifma_ifp; if (ifp != NULL) { CURVNET_SET(ifp->if_vnet); in6m_purge(inm); free(inm, M_IP6MADDR); if_delmulti_ifma_flags(ifma, 1); CURVNET_RESTORE(); if_rele(ifp); } else { in6m_purge(inm); free(inm, M_IP6MADDR); if_delmulti_ifma_flags(ifma, 1); } } -static struct grouptask free_gtask; -static struct in6_multi_head in6m_free_list; -static void in6m_release_task(void *arg __unused); -static void in6m_init(void) +static struct task free_task; +static struct in6_multi_head in6m_free_list = SLIST_HEAD_INITIALIZER(); +static void in6m_release_task(void *arg __unused, int pending __unused); + +static void +in6m_init(void) { - SLIST_INIT(&in6m_free_list); - taskqgroup_config_gtask_init(NULL, &free_gtask, in6m_release_task, "in6m release task"); + TASK_INIT(&free_task, 0, in6m_release_task, NULL); } +SYSINIT(in6m_init, SI_SUB_TASKQ, SI_ORDER_ANY, in6m_init, NULL); -#ifdef EARLY_AP_STARTUP -SYSINIT(in6m_init, SI_SUB_SMP + 1, SI_ORDER_FIRST, - in6m_init, NULL); -#else -SYSINIT(in6m_init, SI_SUB_ROOT_CONF - 1, SI_ORDER_SECOND, - in6m_init, NULL); -#endif - - void in6m_release_list_deferred(struct in6_multi_head *inmh) { if (SLIST_EMPTY(inmh)) return; mtx_lock(&in6_multi_free_mtx); SLIST_CONCAT(&in6m_free_list, inmh, in6_multi, in6m_nrele); mtx_unlock(&in6_multi_free_mtx); - GROUPTASK_ENQUEUE(&free_gtask); + taskqueue_enqueue(taskqueue_thread, &free_task); } void in6m_release_wait(void) { - - /* Wait for all jobs to complete. */ - gtaskqueue_drain_all(free_gtask.gt_taskqueue); + taskqueue_drain_all(taskqueue_thread); } void in6m_disconnect_locked(struct in6_multi_head *inmh, struct in6_multi *inm) { struct ifnet *ifp; struct ifaddr *ifa; struct in6_ifaddr *ifa6; struct in6_multi_mship *imm, *imm_tmp; struct ifmultiaddr *ifma, *ll_ifma; IN6_MULTI_LIST_LOCK_ASSERT(); ifp = inm->in6m_ifp; if (ifp == NULL) return; /* already called */ inm->in6m_ifp = NULL; IF_ADDR_WLOCK_ASSERT(ifp); ifma = inm->in6m_ifma; if (ifma == NULL) return; if_ref(ifp); if (ifma->ifma_flags & IFMA_F_ENQUEUED) { CK_STAILQ_REMOVE(&ifp->if_multiaddrs, ifma, ifmultiaddr, ifma_link); ifma->ifma_flags &= ~IFMA_F_ENQUEUED; } MCDPRINTF("removed ifma: %p from %s\n", ifma, ifp->if_xname); if ((ll_ifma = ifma->ifma_llifma) != NULL) { MPASS(ifma != ll_ifma); ifma->ifma_llifma = NULL; MPASS(ll_ifma->ifma_llifma == NULL); MPASS(ll_ifma->ifma_ifp == ifp); if (--ll_ifma->ifma_refcount == 0) { if (ll_ifma->ifma_flags & IFMA_F_ENQUEUED) { CK_STAILQ_REMOVE(&ifp->if_multiaddrs, ll_ifma, ifmultiaddr, ifma_link); ll_ifma->ifma_flags &= ~IFMA_F_ENQUEUED; } MCDPRINTF("removed ll_ifma: %p from %s\n", ll_ifma, ifp->if_xname); if_freemulti(ll_ifma); } } CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) { if (ifa->ifa_addr->sa_family != AF_INET6) continue; ifa6 = (void *)ifa; LIST_FOREACH_SAFE(imm, &ifa6->ia6_memberships, i6mm_chain, imm_tmp) { if (inm == imm->i6mm_maddr) { LIST_REMOVE(imm, i6mm_chain); free(imm, M_IP6MADDR); in6m_rele_locked(inmh, inm); } } } } static void -in6m_release_task(void *arg __unused) +in6m_release_task(void *arg __unused, int pending __unused) { struct in6_multi_head in6m_free_tmp; struct in6_multi *inm, *tinm; SLIST_INIT(&in6m_free_tmp); mtx_lock(&in6_multi_free_mtx); SLIST_CONCAT(&in6m_free_tmp, &in6m_free_list, in6_multi, in6m_nrele); mtx_unlock(&in6_multi_free_mtx); IN6_MULTI_LOCK(); SLIST_FOREACH_SAFE(inm, &in6m_free_tmp, in6m_nrele, tinm) { SLIST_REMOVE_HEAD(&in6m_free_tmp, in6m_nrele); in6m_release(inm); } IN6_MULTI_UNLOCK(); } /* * Clear recorded source entries for a group. * Used by the MLD code. Caller must hold the IN6_MULTI lock. * FIXME: Should reap. */ void in6m_clear_recorded(struct in6_multi *inm) { struct ip6_msource *ims; IN6_MULTI_LIST_LOCK_ASSERT(); RB_FOREACH(ims, ip6_msource_tree, &inm->in6m_srcs) { if (ims->im6s_stp) { ims->im6s_stp = 0; --inm->in6m_st[1].iss_rec; } } KASSERT(inm->in6m_st[1].iss_rec == 0, ("%s: iss_rec %d not 0", __func__, inm->in6m_st[1].iss_rec)); } /* * Record a source as pending for a Source-Group MLDv2 query. * This lives here as it modifies the shared tree. * * inm is the group descriptor. * naddr is the address of the source to record in network-byte order. * * If the net.inet6.mld.sgalloc sysctl is non-zero, we will * lazy-allocate a source node in response to an SG query. * Otherwise, no allocation is performed. This saves some memory * with the trade-off that the source will not be reported to the * router if joined in the window between the query response and * the group actually being joined on the local host. * * VIMAGE: XXX: Currently the mld_sgalloc feature has been removed. * This turns off the allocation of a recorded source entry if * the group has not been joined. * * Return 0 if the source didn't exist or was already marked as recorded. * Return 1 if the source was marked as recorded by this function. * Return <0 if any error occurred (negated errno code). */ int in6m_record_source(struct in6_multi *inm, const struct in6_addr *addr) { struct ip6_msource find; struct ip6_msource *ims, *nims; IN6_MULTI_LIST_LOCK_ASSERT(); find.im6s_addr = *addr; ims = RB_FIND(ip6_msource_tree, &inm->in6m_srcs, &find); if (ims && ims->im6s_stp) return (0); if (ims == NULL) { if (inm->in6m_nsrc == in6_mcast_maxgrpsrc) return (-ENOSPC); nims = malloc(sizeof(struct ip6_msource), M_IP6MSOURCE, M_NOWAIT | M_ZERO); if (nims == NULL) return (-ENOMEM); nims->im6s_addr = find.im6s_addr; RB_INSERT(ip6_msource_tree, &inm->in6m_srcs, nims); ++inm->in6m_nsrc; ims = nims; } /* * Mark the source as recorded and update the recorded * source count. */ ++ims->im6s_stp; ++inm->in6m_st[1].iss_rec; return (1); } /* * Return a pointer to an in6_msource owned by an in6_mfilter, * given its source address. * Lazy-allocate if needed. If this is a new entry its filter state is * undefined at t0. * * imf is the filter set being modified. * addr is the source address. * * SMPng: May be called with locks held; malloc must not block. */ static int im6f_get_source(struct in6_mfilter *imf, const struct sockaddr_in6 *psin, struct in6_msource **plims) { struct ip6_msource find; struct ip6_msource *ims, *nims; struct in6_msource *lims; int error; error = 0; ims = NULL; lims = NULL; find.im6s_addr = psin->sin6_addr; ims = RB_FIND(ip6_msource_tree, &imf->im6f_sources, &find); lims = (struct in6_msource *)ims; if (lims == NULL) { if (imf->im6f_nsrc == in6_mcast_maxsocksrc) return (ENOSPC); nims = malloc(sizeof(struct in6_msource), M_IN6MFILTER, M_NOWAIT | M_ZERO); if (nims == NULL) return (ENOMEM); lims = (struct in6_msource *)nims; lims->im6s_addr = find.im6s_addr; lims->im6sl_st[0] = MCAST_UNDEFINED; RB_INSERT(ip6_msource_tree, &imf->im6f_sources, nims); ++imf->im6f_nsrc; } *plims = lims; return (error); } /* * Graft a source entry into an existing socket-layer filter set, * maintaining any required invariants and checking allocations. * * The source is marked as being in the new filter mode at t1. * * Return the pointer to the new node, otherwise return NULL. */ static struct in6_msource * im6f_graft(struct in6_mfilter *imf, const uint8_t st1, const struct sockaddr_in6 *psin) { struct ip6_msource *nims; struct in6_msource *lims; nims = malloc(sizeof(struct in6_msource), M_IN6MFILTER, M_NOWAIT | M_ZERO); if (nims == NULL) return (NULL); lims = (struct in6_msource *)nims; lims->im6s_addr = psin->sin6_addr; lims->im6sl_st[0] = MCAST_UNDEFINED; lims->im6sl_st[1] = st1; RB_INSERT(ip6_msource_tree, &imf->im6f_sources, nims); ++imf->im6f_nsrc; return (lims); } /* * Prune a source entry from an existing socket-layer filter set, * maintaining any required invariants and checking allocations. * * The source is marked as being left at t1, it is not freed. * * Return 0 if no error occurred, otherwise return an errno value. */ static int im6f_prune(struct in6_mfilter *imf, const struct sockaddr_in6 *psin) { struct ip6_msource find; struct ip6_msource *ims; struct in6_msource *lims; find.im6s_addr = psin->sin6_addr; ims = RB_FIND(ip6_msource_tree, &imf->im6f_sources, &find); if (ims == NULL) return (ENOENT); lims = (struct in6_msource *)ims; lims->im6sl_st[1] = MCAST_UNDEFINED; return (0); } /* * Revert socket-layer filter set deltas at t1 to t0 state. */ static void im6f_rollback(struct in6_mfilter *imf) { struct ip6_msource *ims, *tims; struct in6_msource *lims; RB_FOREACH_SAFE(ims, ip6_msource_tree, &imf->im6f_sources, tims) { lims = (struct in6_msource *)ims; if (lims->im6sl_st[0] == lims->im6sl_st[1]) { /* no change at t1 */ continue; } else if (lims->im6sl_st[0] != MCAST_UNDEFINED) { /* revert change to existing source at t1 */ lims->im6sl_st[1] = lims->im6sl_st[0]; } else { /* revert source added t1 */ CTR2(KTR_MLD, "%s: free ims %p", __func__, ims); RB_REMOVE(ip6_msource_tree, &imf->im6f_sources, ims); free(ims, M_IN6MFILTER); imf->im6f_nsrc--; } } imf->im6f_st[1] = imf->im6f_st[0]; } /* * Mark socket-layer filter set as INCLUDE {} at t1. */ static void im6f_leave(struct in6_mfilter *imf) { struct ip6_msource *ims; struct in6_msource *lims; RB_FOREACH(ims, ip6_msource_tree, &imf->im6f_sources) { lims = (struct in6_msource *)ims; lims->im6sl_st[1] = MCAST_UNDEFINED; } imf->im6f_st[1] = MCAST_INCLUDE; } /* * Mark socket-layer filter set deltas as committed. */ static void im6f_commit(struct in6_mfilter *imf) { struct ip6_msource *ims; struct in6_msource *lims; RB_FOREACH(ims, ip6_msource_tree, &imf->im6f_sources) { lims = (struct in6_msource *)ims; lims->im6sl_st[0] = lims->im6sl_st[1]; } imf->im6f_st[0] = imf->im6f_st[1]; } /* * Reap unreferenced sources from socket-layer filter set. */ static void im6f_reap(struct in6_mfilter *imf) { struct ip6_msource *ims, *tims; struct in6_msource *lims; RB_FOREACH_SAFE(ims, ip6_msource_tree, &imf->im6f_sources, tims) { lims = (struct in6_msource *)ims; if ((lims->im6sl_st[0] == MCAST_UNDEFINED) && (lims->im6sl_st[1] == MCAST_UNDEFINED)) { CTR2(KTR_MLD, "%s: free lims %p", __func__, ims); RB_REMOVE(ip6_msource_tree, &imf->im6f_sources, ims); free(ims, M_IN6MFILTER); imf->im6f_nsrc--; } } } /* * Purge socket-layer filter set. */ static void im6f_purge(struct in6_mfilter *imf) { struct ip6_msource *ims, *tims; RB_FOREACH_SAFE(ims, ip6_msource_tree, &imf->im6f_sources, tims) { CTR2(KTR_MLD, "%s: free ims %p", __func__, ims); RB_REMOVE(ip6_msource_tree, &imf->im6f_sources, ims); free(ims, M_IN6MFILTER); imf->im6f_nsrc--; } imf->im6f_st[0] = imf->im6f_st[1] = MCAST_UNDEFINED; KASSERT(RB_EMPTY(&imf->im6f_sources), ("%s: im6f_sources not empty", __func__)); } /* * Look up a source filter entry for a multicast group. * * inm is the group descriptor to work with. * addr is the IPv6 address to look up. * noalloc may be non-zero to suppress allocation of sources. * *pims will be set to the address of the retrieved or allocated source. * * SMPng: NOTE: may be called with locks held. * Return 0 if successful, otherwise return a non-zero error code. */ static int in6m_get_source(struct in6_multi *inm, const struct in6_addr *addr, const int noalloc, struct ip6_msource **pims) { struct ip6_msource find; struct ip6_msource *ims, *nims; #ifdef KTR char ip6tbuf[INET6_ADDRSTRLEN]; #endif find.im6s_addr = *addr; ims = RB_FIND(ip6_msource_tree, &inm->in6m_srcs, &find); if (ims == NULL && !noalloc) { if (inm->in6m_nsrc == in6_mcast_maxgrpsrc) return (ENOSPC); nims = malloc(sizeof(struct ip6_msource), M_IP6MSOURCE, M_NOWAIT | M_ZERO); if (nims == NULL) return (ENOMEM); nims->im6s_addr = *addr; RB_INSERT(ip6_msource_tree, &inm->in6m_srcs, nims); ++inm->in6m_nsrc; ims = nims; CTR3(KTR_MLD, "%s: allocated %s as %p", __func__, ip6_sprintf(ip6tbuf, addr), ims); } *pims = ims; return (0); } /* * Merge socket-layer source into MLD-layer source. * If rollback is non-zero, perform the inverse of the merge. */ static void im6s_merge(struct ip6_msource *ims, const struct in6_msource *lims, const int rollback) { int n = rollback ? -1 : 1; #ifdef KTR char ip6tbuf[INET6_ADDRSTRLEN]; ip6_sprintf(ip6tbuf, &lims->im6s_addr); #endif if (lims->im6sl_st[0] == MCAST_EXCLUDE) { CTR3(KTR_MLD, "%s: t1 ex -= %d on %s", __func__, n, ip6tbuf); ims->im6s_st[1].ex -= n; } else if (lims->im6sl_st[0] == MCAST_INCLUDE) { CTR3(KTR_MLD, "%s: t1 in -= %d on %s", __func__, n, ip6tbuf); ims->im6s_st[1].in -= n; } if (lims->im6sl_st[1] == MCAST_EXCLUDE) { CTR3(KTR_MLD, "%s: t1 ex += %d on %s", __func__, n, ip6tbuf); ims->im6s_st[1].ex += n; } else if (lims->im6sl_st[1] == MCAST_INCLUDE) { CTR3(KTR_MLD, "%s: t1 in += %d on %s", __func__, n, ip6tbuf); ims->im6s_st[1].in += n; } } /* * Atomically update the global in6_multi state, when a membership's * filter list is being updated in any way. * * imf is the per-inpcb-membership group filter pointer. * A fake imf may be passed for in-kernel consumers. * * XXX This is a candidate for a set-symmetric-difference style loop * which would eliminate the repeated lookup from root of ims nodes, * as they share the same key space. * * If any error occurred this function will back out of refcounts * and return a non-zero value. */ static int in6m_merge(struct in6_multi *inm, /*const*/ struct in6_mfilter *imf) { struct ip6_msource *ims, *nims; struct in6_msource *lims; int schanged, error; int nsrc0, nsrc1; schanged = 0; error = 0; nsrc1 = nsrc0 = 0; IN6_MULTI_LIST_LOCK_ASSERT(); /* * Update the source filters first, as this may fail. * Maintain count of in-mode filters at t0, t1. These are * used to work out if we transition into ASM mode or not. * Maintain a count of source filters whose state was * actually modified by this operation. */ RB_FOREACH(ims, ip6_msource_tree, &imf->im6f_sources) { lims = (struct in6_msource *)ims; if (lims->im6sl_st[0] == imf->im6f_st[0]) nsrc0++; if (lims->im6sl_st[1] == imf->im6f_st[1]) nsrc1++; if (lims->im6sl_st[0] == lims->im6sl_st[1]) continue; error = in6m_get_source(inm, &lims->im6s_addr, 0, &nims); ++schanged; if (error) break; im6s_merge(nims, lims, 0); } if (error) { struct ip6_msource *bims; RB_FOREACH_REVERSE_FROM(ims, ip6_msource_tree, nims) { lims = (struct in6_msource *)ims; if (lims->im6sl_st[0] == lims->im6sl_st[1]) continue; (void)in6m_get_source(inm, &lims->im6s_addr, 1, &bims); if (bims == NULL) continue; im6s_merge(bims, lims, 1); } goto out_reap; } CTR3(KTR_MLD, "%s: imf filters in-mode: %d at t0, %d at t1", __func__, nsrc0, nsrc1); /* Handle transition between INCLUDE {n} and INCLUDE {} on socket. */ if (imf->im6f_st[0] == imf->im6f_st[1] && imf->im6f_st[1] == MCAST_INCLUDE) { if (nsrc1 == 0) { CTR1(KTR_MLD, "%s: --in on inm at t1", __func__); --inm->in6m_st[1].iss_in; } } /* Handle filter mode transition on socket. */ if (imf->im6f_st[0] != imf->im6f_st[1]) { CTR3(KTR_MLD, "%s: imf transition %d to %d", __func__, imf->im6f_st[0], imf->im6f_st[1]); if (imf->im6f_st[0] == MCAST_EXCLUDE) { CTR1(KTR_MLD, "%s: --ex on inm at t1", __func__); --inm->in6m_st[1].iss_ex; } else if (imf->im6f_st[0] == MCAST_INCLUDE) { CTR1(KTR_MLD, "%s: --in on inm at t1", __func__); --inm->in6m_st[1].iss_in; } if (imf->im6f_st[1] == MCAST_EXCLUDE) { CTR1(KTR_MLD, "%s: ex++ on inm at t1", __func__); inm->in6m_st[1].iss_ex++; } else if (imf->im6f_st[1] == MCAST_INCLUDE && nsrc1 > 0) { CTR1(KTR_MLD, "%s: in++ on inm at t1", __func__); inm->in6m_st[1].iss_in++; } } /* * Track inm filter state in terms of listener counts. * If there are any exclusive listeners, stack-wide * membership is exclusive. * Otherwise, if only inclusive listeners, stack-wide is inclusive. * If no listeners remain, state is undefined at t1, * and the MLD lifecycle for this group should finish. */ if (inm->in6m_st[1].iss_ex > 0) { CTR1(KTR_MLD, "%s: transition to EX", __func__); inm->in6m_st[1].iss_fmode = MCAST_EXCLUDE; } else if (inm->in6m_st[1].iss_in > 0) { CTR1(KTR_MLD, "%s: transition to IN", __func__); inm->in6m_st[1].iss_fmode = MCAST_INCLUDE; } else { CTR1(KTR_MLD, "%s: transition to UNDEF", __func__); inm->in6m_st[1].iss_fmode = MCAST_UNDEFINED; } /* Decrement ASM listener count on transition out of ASM mode. */ if (imf->im6f_st[0] == MCAST_EXCLUDE && nsrc0 == 0) { if ((imf->im6f_st[1] != MCAST_EXCLUDE) || (imf->im6f_st[1] == MCAST_EXCLUDE && nsrc1 > 0)) { CTR1(KTR_MLD, "%s: --asm on inm at t1", __func__); --inm->in6m_st[1].iss_asm; } } /* Increment ASM listener count on transition to ASM mode. */ if (imf->im6f_st[1] == MCAST_EXCLUDE && nsrc1 == 0) { CTR1(KTR_MLD, "%s: asm++ on inm at t1", __func__); inm->in6m_st[1].iss_asm++; } CTR3(KTR_MLD, "%s: merged imf %p to inm %p", __func__, imf, inm); in6m_print(inm); out_reap: if (schanged > 0) { CTR1(KTR_MLD, "%s: sources changed; reaping", __func__); in6m_reap(inm); } return (error); } /* * Mark an in6_multi's filter set deltas as committed. * Called by MLD after a state change has been enqueued. */ void in6m_commit(struct in6_multi *inm) { struct ip6_msource *ims; CTR2(KTR_MLD, "%s: commit inm %p", __func__, inm); CTR1(KTR_MLD, "%s: pre commit:", __func__); in6m_print(inm); RB_FOREACH(ims, ip6_msource_tree, &inm->in6m_srcs) { ims->im6s_st[0] = ims->im6s_st[1]; } inm->in6m_st[0] = inm->in6m_st[1]; } /* * Reap unreferenced nodes from an in6_multi's filter set. */ static void in6m_reap(struct in6_multi *inm) { struct ip6_msource *ims, *tims; RB_FOREACH_SAFE(ims, ip6_msource_tree, &inm->in6m_srcs, tims) { if (ims->im6s_st[0].ex > 0 || ims->im6s_st[0].in > 0 || ims->im6s_st[1].ex > 0 || ims->im6s_st[1].in > 0 || ims->im6s_stp != 0) continue; CTR2(KTR_MLD, "%s: free ims %p", __func__, ims); RB_REMOVE(ip6_msource_tree, &inm->in6m_srcs, ims); free(ims, M_IP6MSOURCE); inm->in6m_nsrc--; } } /* * Purge all source nodes from an in6_multi's filter set. */ static void in6m_purge(struct in6_multi *inm) { struct ip6_msource *ims, *tims; RB_FOREACH_SAFE(ims, ip6_msource_tree, &inm->in6m_srcs, tims) { CTR2(KTR_MLD, "%s: free ims %p", __func__, ims); RB_REMOVE(ip6_msource_tree, &inm->in6m_srcs, ims); free(ims, M_IP6MSOURCE); inm->in6m_nsrc--; } /* Free state-change requests that might be queued. */ mbufq_drain(&inm->in6m_scq); } /* * Join a multicast address w/o sources. * KAME compatibility entry point. * * SMPng: Assume no mc locks held by caller. */ int in6_joingroup(struct ifnet *ifp, const struct in6_addr *mcaddr, /*const*/ struct in6_mfilter *imf, struct in6_multi **pinm, const int delay) { int error; IN6_MULTI_LOCK(); error = in6_joingroup_locked(ifp, mcaddr, NULL, pinm, delay); IN6_MULTI_UNLOCK(); return (error); } /* * Join a multicast group; real entry point. * * Only preserves atomicity at inm level. * NOTE: imf argument cannot be const due to sys/tree.h limitations. * * If the MLD downcall fails, the group is not joined, and an error * code is returned. */ static int in6_joingroup_locked(struct ifnet *ifp, const struct in6_addr *mcaddr, /*const*/ struct in6_mfilter *imf, struct in6_multi **pinm, const int delay) { struct in6_multi_head inmh; struct in6_mfilter timf; struct in6_multi *inm; struct ifmultiaddr *ifma; int error; #ifdef KTR char ip6tbuf[INET6_ADDRSTRLEN]; #endif /* * Sanity: Check scope zone ID was set for ifp, if and * only if group is scoped to an interface. */ KASSERT(IN6_IS_ADDR_MULTICAST(mcaddr), ("%s: not a multicast address", __func__)); if (IN6_IS_ADDR_MC_LINKLOCAL(mcaddr) || IN6_IS_ADDR_MC_INTFACELOCAL(mcaddr)) { KASSERT(mcaddr->s6_addr16[1] != 0, ("%s: scope zone ID not set", __func__)); } IN6_MULTI_LOCK_ASSERT(); IN6_MULTI_LIST_UNLOCK_ASSERT(); CTR4(KTR_MLD, "%s: join %s on %p(%s))", __func__, ip6_sprintf(ip6tbuf, mcaddr), ifp, if_name(ifp)); error = 0; inm = NULL; /* * If no imf was specified (i.e. kernel consumer), * fake one up and assume it is an ASM join. */ if (imf == NULL) { im6f_init(&timf, MCAST_UNDEFINED, MCAST_EXCLUDE); imf = &timf; } error = in6_getmulti(ifp, mcaddr, &inm); if (error) { CTR1(KTR_MLD, "%s: in6_getmulti() failure", __func__); return (error); } IN6_MULTI_LIST_LOCK(); CTR1(KTR_MLD, "%s: merge inm state", __func__); error = in6m_merge(inm, imf); if (error) { CTR1(KTR_MLD, "%s: failed to merge inm state", __func__); goto out_in6m_release; } CTR1(KTR_MLD, "%s: doing mld downcall", __func__); error = mld_change_state(inm, delay); if (error) { CTR1(KTR_MLD, "%s: failed to update source", __func__); goto out_in6m_release; } out_in6m_release: SLIST_INIT(&inmh); if (error) { struct epoch_tracker et; CTR2(KTR_MLD, "%s: dropping ref on %p", __func__, inm); IF_ADDR_WLOCK(ifp); NET_EPOCH_ENTER(et); CK_STAILQ_FOREACH(ifma, &ifp->if_multiaddrs, ifma_link) { if (ifma->ifma_protospec == inm) { ifma->ifma_protospec = NULL; break; } } in6m_disconnect_locked(&inmh, inm); in6m_rele_locked(&inmh, inm); NET_EPOCH_EXIT(et); IF_ADDR_WUNLOCK(ifp); } else { *pinm = inm; } IN6_MULTI_LIST_UNLOCK(); in6m_release_list_deferred(&inmh); return (error); } /* * Leave a multicast group; unlocked entry point. */ int in6_leavegroup(struct in6_multi *inm, /*const*/ struct in6_mfilter *imf) { int error; IN6_MULTI_LOCK(); error = in6_leavegroup_locked(inm, imf); IN6_MULTI_UNLOCK(); return (error); } /* * Leave a multicast group; real entry point. * All source filters will be expunged. * * Only preserves atomicity at inm level. * * Holding the write lock for the INP which contains imf * is highly advisable. We can't assert for it as imf does not * contain a back-pointer to the owning inp. * * Note: This is not the same as in6m_release(*) as this function also * makes a state change downcall into MLD. */ int in6_leavegroup_locked(struct in6_multi *inm, /*const*/ struct in6_mfilter *imf) { struct in6_multi_head inmh; struct in6_mfilter timf; struct ifnet *ifp; int error; #ifdef KTR char ip6tbuf[INET6_ADDRSTRLEN]; #endif error = 0; IN6_MULTI_LOCK_ASSERT(); CTR5(KTR_MLD, "%s: leave inm %p, %s/%s, imf %p", __func__, inm, ip6_sprintf(ip6tbuf, &inm->in6m_addr), (in6m_is_ifp_detached(inm) ? "null" : if_name(inm->in6m_ifp)), imf); /* * If no imf was specified (i.e. kernel consumer), * fake one up and assume it is an ASM join. */ if (imf == NULL) { im6f_init(&timf, MCAST_EXCLUDE, MCAST_UNDEFINED); imf = &timf; } /* * Begin state merge transaction at MLD layer. * * As this particular invocation should not cause any memory * to be allocated, and there is no opportunity to roll back * the transaction, it MUST NOT fail. */ ifp = inm->in6m_ifp; IN6_MULTI_LIST_LOCK(); CTR1(KTR_MLD, "%s: merge inm state", __func__); error = in6m_merge(inm, imf); KASSERT(error == 0, ("%s: failed to merge inm state", __func__)); CTR1(KTR_MLD, "%s: doing mld downcall", __func__); error = 0; if (ifp) error = mld_change_state(inm, 0); if (error) CTR1(KTR_MLD, "%s: failed mld downcall", __func__); CTR2(KTR_MLD, "%s: dropping ref on %p", __func__, inm); if (ifp) IF_ADDR_WLOCK(ifp); SLIST_INIT(&inmh); if (inm->in6m_refcount == 1) in6m_disconnect_locked(&inmh, inm); in6m_rele_locked(&inmh, inm); if (ifp) IF_ADDR_WUNLOCK(ifp); IN6_MULTI_LIST_UNLOCK(); in6m_release_list_deferred(&inmh); return (error); } /* * Block or unblock an ASM multicast source on an inpcb. * This implements the delta-based API described in RFC 3678. * * The delta-based API applies only to exclusive-mode memberships. * An MLD downcall will be performed. * * SMPng: NOTE: Must take Giant as a join may create a new ifma. * * Return 0 if successful, otherwise return an appropriate error code. */ static int in6p_block_unblock_source(struct inpcb *inp, struct sockopt *sopt) { struct group_source_req gsr; sockunion_t *gsa, *ssa; struct ifnet *ifp; struct in6_mfilter *imf; struct ip6_moptions *imo; struct in6_msource *ims; struct in6_multi *inm; uint16_t fmode; int error, doblock; #ifdef KTR char ip6tbuf[INET6_ADDRSTRLEN]; #endif ifp = NULL; error = 0; doblock = 0; memset(&gsr, 0, sizeof(struct group_source_req)); gsa = (sockunion_t *)&gsr.gsr_group; ssa = (sockunion_t *)&gsr.gsr_source; switch (sopt->sopt_name) { case MCAST_BLOCK_SOURCE: case MCAST_UNBLOCK_SOURCE: error = sooptcopyin(sopt, &gsr, sizeof(struct group_source_req), sizeof(struct group_source_req)); if (error) return (error); if (gsa->sin6.sin6_family != AF_INET6 || gsa->sin6.sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (ssa->sin6.sin6_family != AF_INET6 || ssa->sin6.sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (gsr.gsr_interface == 0 || V_if_index < gsr.gsr_interface) return (EADDRNOTAVAIL); ifp = ifnet_byindex(gsr.gsr_interface); if (sopt->sopt_name == MCAST_BLOCK_SOURCE) doblock = 1; break; default: CTR2(KTR_MLD, "%s: unknown sopt_name %d", __func__, sopt->sopt_name); return (EOPNOTSUPP); break; } if (!IN6_IS_ADDR_MULTICAST(&gsa->sin6.sin6_addr)) return (EINVAL); (void)in6_setscope(&gsa->sin6.sin6_addr, ifp, NULL); /* * Check if we are actually a member of this group. */ imo = in6p_findmoptions(inp); imf = im6o_match_group(imo, ifp, &gsa->sa); if (imf == NULL) { error = EADDRNOTAVAIL; goto out_in6p_locked; } inm = imf->im6f_in6m; /* * Attempting to use the delta-based API on an * non exclusive-mode membership is an error. */ fmode = imf->im6f_st[0]; if (fmode != MCAST_EXCLUDE) { error = EINVAL; goto out_in6p_locked; } /* * Deal with error cases up-front: * Asked to block, but already blocked; or * Asked to unblock, but nothing to unblock. * If adding a new block entry, allocate it. */ ims = im6o_match_source(imf, &ssa->sa); if ((ims != NULL && doblock) || (ims == NULL && !doblock)) { CTR3(KTR_MLD, "%s: source %s %spresent", __func__, ip6_sprintf(ip6tbuf, &ssa->sin6.sin6_addr), doblock ? "" : "not "); error = EADDRNOTAVAIL; goto out_in6p_locked; } INP_WLOCK_ASSERT(inp); /* * Begin state merge transaction at socket layer. */ if (doblock) { CTR2(KTR_MLD, "%s: %s source", __func__, "block"); ims = im6f_graft(imf, fmode, &ssa->sin6); if (ims == NULL) error = ENOMEM; } else { CTR2(KTR_MLD, "%s: %s source", __func__, "allow"); error = im6f_prune(imf, &ssa->sin6); } if (error) { CTR1(KTR_MLD, "%s: merge imf state failed", __func__); goto out_im6f_rollback; } /* * Begin state merge transaction at MLD layer. */ IN6_MULTI_LIST_LOCK(); CTR1(KTR_MLD, "%s: merge inm state", __func__); error = in6m_merge(inm, imf); if (error) CTR1(KTR_MLD, "%s: failed to merge inm state", __func__); else { CTR1(KTR_MLD, "%s: doing mld downcall", __func__); error = mld_change_state(inm, 0); if (error) CTR1(KTR_MLD, "%s: failed mld downcall", __func__); } IN6_MULTI_LIST_UNLOCK(); out_im6f_rollback: if (error) im6f_rollback(imf); else im6f_commit(imf); im6f_reap(imf); out_in6p_locked: INP_WUNLOCK(inp); return (error); } /* * Given an inpcb, return its multicast options structure pointer. Accepts * an unlocked inpcb pointer, but will return it locked. May sleep. * * SMPng: NOTE: Potentially calls malloc(M_WAITOK) with Giant held. * SMPng: NOTE: Returns with the INP write lock held. */ static struct ip6_moptions * in6p_findmoptions(struct inpcb *inp) { struct ip6_moptions *imo; INP_WLOCK(inp); if (inp->in6p_moptions != NULL) return (inp->in6p_moptions); INP_WUNLOCK(inp); imo = malloc(sizeof(*imo), M_IP6MOPTS, M_WAITOK); imo->im6o_multicast_ifp = NULL; imo->im6o_multicast_hlim = V_ip6_defmcasthlim; imo->im6o_multicast_loop = in6_mcast_loop; STAILQ_INIT(&imo->im6o_head); INP_WLOCK(inp); if (inp->in6p_moptions != NULL) { free(imo, M_IP6MOPTS); return (inp->in6p_moptions); } inp->in6p_moptions = imo; return (imo); } /* * Discard the IPv6 multicast options (and source filters). * * SMPng: NOTE: assumes INP write lock is held. * * XXX can all be safely deferred to epoch_call * */ static void inp_gcmoptions(struct ip6_moptions *imo) { struct in6_mfilter *imf; struct in6_multi *inm; struct ifnet *ifp; while ((imf = ip6_mfilter_first(&imo->im6o_head)) != NULL) { ip6_mfilter_remove(&imo->im6o_head, imf); im6f_leave(imf); if ((inm = imf->im6f_in6m) != NULL) { if ((ifp = inm->in6m_ifp) != NULL) { CURVNET_SET(ifp->if_vnet); (void)in6_leavegroup(inm, imf); CURVNET_RESTORE(); } else { (void)in6_leavegroup(inm, imf); } } ip6_mfilter_free(imf); } free(imo, M_IP6MOPTS); } void ip6_freemoptions(struct ip6_moptions *imo) { if (imo == NULL) return; inp_gcmoptions(imo); } /* * Atomically get source filters on a socket for an IPv6 multicast group. * Called with INP lock held; returns with lock released. */ static int in6p_get_source_filters(struct inpcb *inp, struct sockopt *sopt) { struct __msfilterreq msfr; sockunion_t *gsa; struct ifnet *ifp; struct ip6_moptions *imo; struct in6_mfilter *imf; struct ip6_msource *ims; struct in6_msource *lims; struct sockaddr_in6 *psin; struct sockaddr_storage *ptss; struct sockaddr_storage *tss; int error; size_t nsrcs, ncsrcs; INP_WLOCK_ASSERT(inp); imo = inp->in6p_moptions; KASSERT(imo != NULL, ("%s: null ip6_moptions", __func__)); INP_WUNLOCK(inp); error = sooptcopyin(sopt, &msfr, sizeof(struct __msfilterreq), sizeof(struct __msfilterreq)); if (error) return (error); if (msfr.msfr_group.ss_family != AF_INET6 || msfr.msfr_group.ss_len != sizeof(struct sockaddr_in6)) return (EINVAL); gsa = (sockunion_t *)&msfr.msfr_group; if (!IN6_IS_ADDR_MULTICAST(&gsa->sin6.sin6_addr)) return (EINVAL); if (msfr.msfr_ifindex == 0 || V_if_index < msfr.msfr_ifindex) return (EADDRNOTAVAIL); ifp = ifnet_byindex(msfr.msfr_ifindex); if (ifp == NULL) return (EADDRNOTAVAIL); (void)in6_setscope(&gsa->sin6.sin6_addr, ifp, NULL); INP_WLOCK(inp); /* * Lookup group on the socket. */ imf = im6o_match_group(imo, ifp, &gsa->sa); if (imf == NULL) { INP_WUNLOCK(inp); return (EADDRNOTAVAIL); } /* * Ignore memberships which are in limbo. */ if (imf->im6f_st[1] == MCAST_UNDEFINED) { INP_WUNLOCK(inp); return (EAGAIN); } msfr.msfr_fmode = imf->im6f_st[1]; /* * If the user specified a buffer, copy out the source filter * entries to userland gracefully. * We only copy out the number of entries which userland * has asked for, but we always tell userland how big the * buffer really needs to be. */ if (msfr.msfr_nsrcs > in6_mcast_maxsocksrc) msfr.msfr_nsrcs = in6_mcast_maxsocksrc; tss = NULL; if (msfr.msfr_srcs != NULL && msfr.msfr_nsrcs > 0) { tss = malloc(sizeof(struct sockaddr_storage) * msfr.msfr_nsrcs, M_TEMP, M_NOWAIT | M_ZERO); if (tss == NULL) { INP_WUNLOCK(inp); return (ENOBUFS); } } /* * Count number of sources in-mode at t0. * If buffer space exists and remains, copy out source entries. */ nsrcs = msfr.msfr_nsrcs; ncsrcs = 0; ptss = tss; RB_FOREACH(ims, ip6_msource_tree, &imf->im6f_sources) { lims = (struct in6_msource *)ims; if (lims->im6sl_st[0] == MCAST_UNDEFINED || lims->im6sl_st[0] != imf->im6f_st[0]) continue; ++ncsrcs; if (tss != NULL && nsrcs > 0) { psin = (struct sockaddr_in6 *)ptss; psin->sin6_family = AF_INET6; psin->sin6_len = sizeof(struct sockaddr_in6); psin->sin6_addr = lims->im6s_addr; psin->sin6_port = 0; --nsrcs; ++ptss; } } INP_WUNLOCK(inp); if (tss != NULL) { error = copyout(tss, msfr.msfr_srcs, sizeof(struct sockaddr_storage) * msfr.msfr_nsrcs); free(tss, M_TEMP); if (error) return (error); } msfr.msfr_nsrcs = ncsrcs; error = sooptcopyout(sopt, &msfr, sizeof(struct __msfilterreq)); return (error); } /* * Return the IP multicast options in response to user getsockopt(). */ int ip6_getmoptions(struct inpcb *inp, struct sockopt *sopt) { struct ip6_moptions *im6o; int error; u_int optval; INP_WLOCK(inp); im6o = inp->in6p_moptions; /* * If socket is neither of type SOCK_RAW or SOCK_DGRAM, * or is a divert socket, reject it. */ if (inp->inp_socket->so_proto->pr_protocol == IPPROTO_DIVERT || (inp->inp_socket->so_proto->pr_type != SOCK_RAW && inp->inp_socket->so_proto->pr_type != SOCK_DGRAM)) { INP_WUNLOCK(inp); return (EOPNOTSUPP); } error = 0; switch (sopt->sopt_name) { case IPV6_MULTICAST_IF: if (im6o == NULL || im6o->im6o_multicast_ifp == NULL) { optval = 0; } else { optval = im6o->im6o_multicast_ifp->if_index; } INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof(u_int)); break; case IPV6_MULTICAST_HOPS: if (im6o == NULL) optval = V_ip6_defmcasthlim; else optval = im6o->im6o_multicast_hlim; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof(u_int)); break; case IPV6_MULTICAST_LOOP: if (im6o == NULL) optval = in6_mcast_loop; /* XXX VIMAGE */ else optval = im6o->im6o_multicast_loop; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof(u_int)); break; case IPV6_MSFILTER: if (im6o == NULL) { error = EADDRNOTAVAIL; INP_WUNLOCK(inp); } else { error = in6p_get_source_filters(inp, sopt); } break; default: INP_WUNLOCK(inp); error = ENOPROTOOPT; break; } INP_UNLOCK_ASSERT(inp); return (error); } /* * Look up the ifnet to use for a multicast group membership, * given the address of an IPv6 group. * * This routine exists to support legacy IPv6 multicast applications. * * If inp is non-NULL, use this socket's current FIB number for any * required FIB lookup. Look up the group address in the unicast FIB, * and use its ifp; usually, this points to the default next-hop. * If the FIB lookup fails, return NULL. * * FUTURE: Support multiple forwarding tables for IPv6. * * Returns NULL if no ifp could be found. */ static struct ifnet * in6p_lookup_mcast_ifp(const struct inpcb *inp, const struct sockaddr_in6 *gsin6) { struct nhop6_basic nh6; struct in6_addr dst; uint32_t scopeid; uint32_t fibnum; KASSERT(inp->inp_vflag & INP_IPV6, ("%s: not INP_IPV6 inpcb", __func__)); KASSERT(gsin6->sin6_family == AF_INET6, ("%s: not AF_INET6 group", __func__)); in6_splitscope(&gsin6->sin6_addr, &dst, &scopeid); fibnum = inp ? inp->inp_inc.inc_fibnum : RT_DEFAULT_FIB; if (fib6_lookup_nh_basic(fibnum, &dst, scopeid, 0, 0, &nh6) != 0) return (NULL); return (nh6.nh_ifp); } /* * Join an IPv6 multicast group, possibly with a source. * * FIXME: The KAME use of the unspecified address (::) * to join *all* multicast groups is currently unsupported. */ static int in6p_join_group(struct inpcb *inp, struct sockopt *sopt) { struct in6_multi_head inmh; struct group_source_req gsr; sockunion_t *gsa, *ssa; struct ifnet *ifp; struct in6_mfilter *imf; struct ip6_moptions *imo; struct in6_multi *inm; struct in6_msource *lims; int error, is_new; SLIST_INIT(&inmh); ifp = NULL; lims = NULL; error = 0; memset(&gsr, 0, sizeof(struct group_source_req)); gsa = (sockunion_t *)&gsr.gsr_group; gsa->ss.ss_family = AF_UNSPEC; ssa = (sockunion_t *)&gsr.gsr_source; ssa->ss.ss_family = AF_UNSPEC; /* * Chew everything into struct group_source_req. * Overwrite the port field if present, as the sockaddr * being copied in may be matched with a binary comparison. * Ignore passed-in scope ID. */ switch (sopt->sopt_name) { case IPV6_JOIN_GROUP: { struct ipv6_mreq mreq; error = sooptcopyin(sopt, &mreq, sizeof(struct ipv6_mreq), sizeof(struct ipv6_mreq)); if (error) return (error); gsa->sin6.sin6_family = AF_INET6; gsa->sin6.sin6_len = sizeof(struct sockaddr_in6); gsa->sin6.sin6_addr = mreq.ipv6mr_multiaddr; if (mreq.ipv6mr_interface == 0) { ifp = in6p_lookup_mcast_ifp(inp, &gsa->sin6); } else { if (V_if_index < mreq.ipv6mr_interface) return (EADDRNOTAVAIL); ifp = ifnet_byindex(mreq.ipv6mr_interface); } CTR3(KTR_MLD, "%s: ipv6mr_interface = %d, ifp = %p", __func__, mreq.ipv6mr_interface, ifp); } break; case MCAST_JOIN_GROUP: case MCAST_JOIN_SOURCE_GROUP: if (sopt->sopt_name == MCAST_JOIN_GROUP) { error = sooptcopyin(sopt, &gsr, sizeof(struct group_req), sizeof(struct group_req)); } else if (sopt->sopt_name == MCAST_JOIN_SOURCE_GROUP) { error = sooptcopyin(sopt, &gsr, sizeof(struct group_source_req), sizeof(struct group_source_req)); } if (error) return (error); if (gsa->sin6.sin6_family != AF_INET6 || gsa->sin6.sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (sopt->sopt_name == MCAST_JOIN_SOURCE_GROUP) { if (ssa->sin6.sin6_family != AF_INET6 || ssa->sin6.sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (IN6_IS_ADDR_MULTICAST(&ssa->sin6.sin6_addr)) return (EINVAL); /* * TODO: Validate embedded scope ID in source * list entry against passed-in ifp, if and only * if source list filter entry is iface or node local. */ in6_clearscope(&ssa->sin6.sin6_addr); ssa->sin6.sin6_port = 0; ssa->sin6.sin6_scope_id = 0; } if (gsr.gsr_interface == 0 || V_if_index < gsr.gsr_interface) return (EADDRNOTAVAIL); ifp = ifnet_byindex(gsr.gsr_interface); break; default: CTR2(KTR_MLD, "%s: unknown sopt_name %d", __func__, sopt->sopt_name); return (EOPNOTSUPP); break; } if (!IN6_IS_ADDR_MULTICAST(&gsa->sin6.sin6_addr)) return (EINVAL); if (ifp == NULL || (ifp->if_flags & IFF_MULTICAST) == 0) return (EADDRNOTAVAIL); gsa->sin6.sin6_port = 0; gsa->sin6.sin6_scope_id = 0; /* * Always set the scope zone ID on memberships created from userland. * Use the passed-in ifp to do this. * XXX The in6_setscope() return value is meaningless. * XXX SCOPE6_LOCK() is taken by in6_setscope(). */ (void)in6_setscope(&gsa->sin6.sin6_addr, ifp, NULL); IN6_MULTI_LOCK(); /* * Find the membership in the membership list. */ imo = in6p_findmoptions(inp); imf = im6o_match_group(imo, ifp, &gsa->sa); if (imf == NULL) { is_new = 1; inm = NULL; if (ip6_mfilter_count(&imo->im6o_head) >= IPV6_MAX_MEMBERSHIPS) { error = ENOMEM; goto out_in6p_locked; } } else { is_new = 0; inm = imf->im6f_in6m; if (ssa->ss.ss_family != AF_UNSPEC) { /* * MCAST_JOIN_SOURCE_GROUP on an exclusive membership * is an error. On an existing inclusive membership, * it just adds the source to the filter list. */ if (imf->im6f_st[1] != MCAST_INCLUDE) { error = EINVAL; goto out_in6p_locked; } /* * Throw out duplicates. * * XXX FIXME: This makes a naive assumption that * even if entries exist for *ssa in this imf, * they will be rejected as dupes, even if they * are not valid in the current mode (in-mode). * * in6_msource is transactioned just as for anything * else in SSM -- but note naive use of in6m_graft() * below for allocating new filter entries. * * This is only an issue if someone mixes the * full-state SSM API with the delta-based API, * which is discouraged in the relevant RFCs. */ lims = im6o_match_source(imf, &ssa->sa); if (lims != NULL /*&& lims->im6sl_st[1] == MCAST_INCLUDE*/) { error = EADDRNOTAVAIL; goto out_in6p_locked; } } else { /* * MCAST_JOIN_GROUP alone, on any existing membership, * is rejected, to stop the same inpcb tying up * multiple refs to the in_multi. * On an existing inclusive membership, this is also * an error; if you want to change filter mode, * you must use the userland API setsourcefilter(). * XXX We don't reject this for imf in UNDEFINED * state at t1, because allocation of a filter * is atomic with allocation of a membership. */ error = EINVAL; goto out_in6p_locked; } } /* * Begin state merge transaction at socket layer. */ INP_WLOCK_ASSERT(inp); /* * Graft new source into filter list for this inpcb's * membership of the group. The in6_multi may not have * been allocated yet if this is a new membership, however, * the in_mfilter slot will be allocated and must be initialized. * * Note: Grafting of exclusive mode filters doesn't happen * in this path. * XXX: Should check for non-NULL lims (node exists but may * not be in-mode) for interop with full-state API. */ if (ssa->ss.ss_family != AF_UNSPEC) { /* Membership starts in IN mode */ if (is_new) { CTR1(KTR_MLD, "%s: new join w/source", __func__); imf = ip6_mfilter_alloc(M_NOWAIT, MCAST_UNDEFINED, MCAST_INCLUDE); if (imf == NULL) { error = ENOMEM; goto out_in6p_locked; } } else { CTR2(KTR_MLD, "%s: %s source", __func__, "allow"); } lims = im6f_graft(imf, MCAST_INCLUDE, &ssa->sin6); if (lims == NULL) { CTR1(KTR_MLD, "%s: merge imf state failed", __func__); error = ENOMEM; goto out_in6p_locked; } } else { /* No address specified; Membership starts in EX mode */ if (is_new) { CTR1(KTR_MLD, "%s: new join w/o source", __func__); imf = ip6_mfilter_alloc(M_NOWAIT, MCAST_UNDEFINED, MCAST_EXCLUDE); if (imf == NULL) { error = ENOMEM; goto out_in6p_locked; } } } /* * Begin state merge transaction at MLD layer. */ if (is_new) { in_pcbref(inp); INP_WUNLOCK(inp); error = in6_joingroup_locked(ifp, &gsa->sin6.sin6_addr, imf, &imf->im6f_in6m, 0); INP_WLOCK(inp); if (in_pcbrele_wlocked(inp)) { error = ENXIO; goto out_in6p_unlocked; } if (error) { goto out_in6p_locked; } /* * NOTE: Refcount from in6_joingroup_locked() * is protecting membership. */ ip6_mfilter_insert(&imo->im6o_head, imf); } else { CTR1(KTR_MLD, "%s: merge inm state", __func__); IN6_MULTI_LIST_LOCK(); error = in6m_merge(inm, imf); if (error) { CTR1(KTR_MLD, "%s: failed to merge inm state", __func__); IN6_MULTI_LIST_UNLOCK(); im6f_rollback(imf); im6f_reap(imf); goto out_in6p_locked; } CTR1(KTR_MLD, "%s: doing mld downcall", __func__); error = mld_change_state(inm, 0); IN6_MULTI_LIST_UNLOCK(); if (error) { CTR1(KTR_MLD, "%s: failed mld downcall", __func__); im6f_rollback(imf); im6f_reap(imf); goto out_in6p_locked; } } im6f_commit(imf); imf = NULL; out_in6p_locked: INP_WUNLOCK(inp); out_in6p_unlocked: IN6_MULTI_UNLOCK(); if (is_new && imf) { if (imf->im6f_in6m != NULL) { struct in6_multi_head inmh; SLIST_INIT(&inmh); SLIST_INSERT_HEAD(&inmh, imf->im6f_in6m, in6m_defer); in6m_release_list_deferred(&inmh); } ip6_mfilter_free(imf); } return (error); } /* * Leave an IPv6 multicast group on an inpcb, possibly with a source. */ static int in6p_leave_group(struct inpcb *inp, struct sockopt *sopt) { struct ipv6_mreq mreq; struct group_source_req gsr; sockunion_t *gsa, *ssa; struct ifnet *ifp; struct in6_mfilter *imf; struct ip6_moptions *imo; struct in6_msource *ims; struct in6_multi *inm; uint32_t ifindex; int error; bool is_final; #ifdef KTR char ip6tbuf[INET6_ADDRSTRLEN]; #endif ifp = NULL; ifindex = 0; error = 0; is_final = true; memset(&gsr, 0, sizeof(struct group_source_req)); gsa = (sockunion_t *)&gsr.gsr_group; gsa->ss.ss_family = AF_UNSPEC; ssa = (sockunion_t *)&gsr.gsr_source; ssa->ss.ss_family = AF_UNSPEC; /* * Chew everything passed in up into a struct group_source_req * as that is easier to process. * Note: Any embedded scope ID in the multicast group passed * in by userland is ignored, the interface index is the recommended * mechanism to specify an interface; see below. */ switch (sopt->sopt_name) { case IPV6_LEAVE_GROUP: error = sooptcopyin(sopt, &mreq, sizeof(struct ipv6_mreq), sizeof(struct ipv6_mreq)); if (error) return (error); gsa->sin6.sin6_family = AF_INET6; gsa->sin6.sin6_len = sizeof(struct sockaddr_in6); gsa->sin6.sin6_addr = mreq.ipv6mr_multiaddr; gsa->sin6.sin6_port = 0; gsa->sin6.sin6_scope_id = 0; ifindex = mreq.ipv6mr_interface; break; case MCAST_LEAVE_GROUP: case MCAST_LEAVE_SOURCE_GROUP: if (sopt->sopt_name == MCAST_LEAVE_GROUP) { error = sooptcopyin(sopt, &gsr, sizeof(struct group_req), sizeof(struct group_req)); } else if (sopt->sopt_name == MCAST_LEAVE_SOURCE_GROUP) { error = sooptcopyin(sopt, &gsr, sizeof(struct group_source_req), sizeof(struct group_source_req)); } if (error) return (error); if (gsa->sin6.sin6_family != AF_INET6 || gsa->sin6.sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (sopt->sopt_name == MCAST_LEAVE_SOURCE_GROUP) { if (ssa->sin6.sin6_family != AF_INET6 || ssa->sin6.sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (IN6_IS_ADDR_MULTICAST(&ssa->sin6.sin6_addr)) return (EINVAL); /* * TODO: Validate embedded scope ID in source * list entry against passed-in ifp, if and only * if source list filter entry is iface or node local. */ in6_clearscope(&ssa->sin6.sin6_addr); } gsa->sin6.sin6_port = 0; gsa->sin6.sin6_scope_id = 0; ifindex = gsr.gsr_interface; break; default: CTR2(KTR_MLD, "%s: unknown sopt_name %d", __func__, sopt->sopt_name); return (EOPNOTSUPP); break; } if (!IN6_IS_ADDR_MULTICAST(&gsa->sin6.sin6_addr)) return (EINVAL); /* * Validate interface index if provided. If no interface index * was provided separately, attempt to look the membership up * from the default scope as a last resort to disambiguate * the membership we are being asked to leave. * XXX SCOPE6 lock potentially taken here. */ if (ifindex != 0) { if (V_if_index < ifindex) return (EADDRNOTAVAIL); ifp = ifnet_byindex(ifindex); if (ifp == NULL) return (EADDRNOTAVAIL); (void)in6_setscope(&gsa->sin6.sin6_addr, ifp, NULL); } else { error = sa6_embedscope(&gsa->sin6, V_ip6_use_defzone); if (error) return (EADDRNOTAVAIL); /* * Some badly behaved applications don't pass an ifindex * or a scope ID, which is an API violation. In this case, * perform a lookup as per a v6 join. * * XXX For now, stomp on zone ID for the corner case. * This is not the 'KAME way', but we need to see the ifp * directly until such time as this implementation is * refactored, assuming the scope IDs are the way to go. */ ifindex = ntohs(gsa->sin6.sin6_addr.s6_addr16[1]); if (ifindex == 0) { CTR2(KTR_MLD, "%s: warning: no ifindex, looking up " "ifp for group %s.", __func__, ip6_sprintf(ip6tbuf, &gsa->sin6.sin6_addr)); ifp = in6p_lookup_mcast_ifp(inp, &gsa->sin6); } else { ifp = ifnet_byindex(ifindex); } if (ifp == NULL) return (EADDRNOTAVAIL); } CTR2(KTR_MLD, "%s: ifp = %p", __func__, ifp); KASSERT(ifp != NULL, ("%s: ifp did not resolve", __func__)); IN6_MULTI_LOCK(); /* * Find the membership in the membership list. */ imo = in6p_findmoptions(inp); imf = im6o_match_group(imo, ifp, &gsa->sa); if (imf == NULL) { error = EADDRNOTAVAIL; goto out_in6p_locked; } inm = imf->im6f_in6m; if (ssa->ss.ss_family != AF_UNSPEC) is_final = false; /* * Begin state merge transaction at socket layer. */ INP_WLOCK_ASSERT(inp); /* * If we were instructed only to leave a given source, do so. * MCAST_LEAVE_SOURCE_GROUP is only valid for inclusive memberships. */ if (is_final) { ip6_mfilter_remove(&imo->im6o_head, imf); im6f_leave(imf); /* * Give up the multicast address record to which * the membership points. */ (void)in6_leavegroup_locked(inm, imf); } else { if (imf->im6f_st[0] == MCAST_EXCLUDE) { error = EADDRNOTAVAIL; goto out_in6p_locked; } ims = im6o_match_source(imf, &ssa->sa); if (ims == NULL) { CTR3(KTR_MLD, "%s: source %p %spresent", __func__, ip6_sprintf(ip6tbuf, &ssa->sin6.sin6_addr), "not "); error = EADDRNOTAVAIL; goto out_in6p_locked; } CTR2(KTR_MLD, "%s: %s source", __func__, "block"); error = im6f_prune(imf, &ssa->sin6); if (error) { CTR1(KTR_MLD, "%s: merge imf state failed", __func__); goto out_in6p_locked; } } /* * Begin state merge transaction at MLD layer. */ if (!is_final) { CTR1(KTR_MLD, "%s: merge inm state", __func__); IN6_MULTI_LIST_LOCK(); error = in6m_merge(inm, imf); if (error) { CTR1(KTR_MLD, "%s: failed to merge inm state", __func__); IN6_MULTI_LIST_UNLOCK(); im6f_rollback(imf); im6f_reap(imf); goto out_in6p_locked; } CTR1(KTR_MLD, "%s: doing mld downcall", __func__); error = mld_change_state(inm, 0); IN6_MULTI_LIST_UNLOCK(); if (error) { CTR1(KTR_MLD, "%s: failed mld downcall", __func__); im6f_rollback(imf); im6f_reap(imf); goto out_in6p_locked; } } im6f_commit(imf); im6f_reap(imf); out_in6p_locked: INP_WUNLOCK(inp); if (is_final && imf) ip6_mfilter_free(imf); IN6_MULTI_UNLOCK(); return (error); } /* * Select the interface for transmitting IPv6 multicast datagrams. * * Either an instance of struct in6_addr or an instance of struct ipv6_mreqn * may be passed to this socket option. An address of in6addr_any or an * interface index of 0 is used to remove a previous selection. * When no interface is selected, one is chosen for every send. */ static int in6p_set_multicast_if(struct inpcb *inp, struct sockopt *sopt) { struct ifnet *ifp; struct ip6_moptions *imo; u_int ifindex; int error; if (sopt->sopt_valsize != sizeof(u_int)) return (EINVAL); error = sooptcopyin(sopt, &ifindex, sizeof(u_int), sizeof(u_int)); if (error) return (error); if (V_if_index < ifindex) return (EINVAL); if (ifindex == 0) ifp = NULL; else { ifp = ifnet_byindex(ifindex); if (ifp == NULL) return (EINVAL); if ((ifp->if_flags & IFF_MULTICAST) == 0) return (EADDRNOTAVAIL); } imo = in6p_findmoptions(inp); imo->im6o_multicast_ifp = ifp; INP_WUNLOCK(inp); return (0); } /* * Atomically set source filters on a socket for an IPv6 multicast group. * * SMPng: NOTE: Potentially calls malloc(M_WAITOK) with Giant held. */ static int in6p_set_source_filters(struct inpcb *inp, struct sockopt *sopt) { struct __msfilterreq msfr; sockunion_t *gsa; struct ifnet *ifp; struct in6_mfilter *imf; struct ip6_moptions *imo; struct in6_multi *inm; int error; error = sooptcopyin(sopt, &msfr, sizeof(struct __msfilterreq), sizeof(struct __msfilterreq)); if (error) return (error); if (msfr.msfr_nsrcs > in6_mcast_maxsocksrc) return (ENOBUFS); if (msfr.msfr_fmode != MCAST_EXCLUDE && msfr.msfr_fmode != MCAST_INCLUDE) return (EINVAL); if (msfr.msfr_group.ss_family != AF_INET6 || msfr.msfr_group.ss_len != sizeof(struct sockaddr_in6)) return (EINVAL); gsa = (sockunion_t *)&msfr.msfr_group; if (!IN6_IS_ADDR_MULTICAST(&gsa->sin6.sin6_addr)) return (EINVAL); gsa->sin6.sin6_port = 0; /* ignore port */ if (msfr.msfr_ifindex == 0 || V_if_index < msfr.msfr_ifindex) return (EADDRNOTAVAIL); ifp = ifnet_byindex(msfr.msfr_ifindex); if (ifp == NULL) return (EADDRNOTAVAIL); (void)in6_setscope(&gsa->sin6.sin6_addr, ifp, NULL); /* * Take the INP write lock. * Check if this socket is a member of this group. */ imo = in6p_findmoptions(inp); imf = im6o_match_group(imo, ifp, &gsa->sa); if (imf == NULL) { error = EADDRNOTAVAIL; goto out_in6p_locked; } inm = imf->im6f_in6m; /* * Begin state merge transaction at socket layer. */ INP_WLOCK_ASSERT(inp); imf->im6f_st[1] = msfr.msfr_fmode; /* * Apply any new source filters, if present. * Make a copy of the user-space source vector so * that we may copy them with a single copyin. This * allows us to deal with page faults up-front. */ if (msfr.msfr_nsrcs > 0) { struct in6_msource *lims; struct sockaddr_in6 *psin; struct sockaddr_storage *kss, *pkss; int i; INP_WUNLOCK(inp); CTR2(KTR_MLD, "%s: loading %lu source list entries", __func__, (unsigned long)msfr.msfr_nsrcs); kss = malloc(sizeof(struct sockaddr_storage) * msfr.msfr_nsrcs, M_TEMP, M_WAITOK); error = copyin(msfr.msfr_srcs, kss, sizeof(struct sockaddr_storage) * msfr.msfr_nsrcs); if (error) { free(kss, M_TEMP); return (error); } INP_WLOCK(inp); /* * Mark all source filters as UNDEFINED at t1. * Restore new group filter mode, as im6f_leave() * will set it to INCLUDE. */ im6f_leave(imf); imf->im6f_st[1] = msfr.msfr_fmode; /* * Update socket layer filters at t1, lazy-allocating * new entries. This saves a bunch of memory at the * cost of one RB_FIND() per source entry; duplicate * entries in the msfr_nsrcs vector are ignored. * If we encounter an error, rollback transaction. * * XXX This too could be replaced with a set-symmetric * difference like loop to avoid walking from root * every time, as the key space is common. */ for (i = 0, pkss = kss; i < msfr.msfr_nsrcs; i++, pkss++) { psin = (struct sockaddr_in6 *)pkss; if (psin->sin6_family != AF_INET6) { error = EAFNOSUPPORT; break; } if (psin->sin6_len != sizeof(struct sockaddr_in6)) { error = EINVAL; break; } if (IN6_IS_ADDR_MULTICAST(&psin->sin6_addr)) { error = EINVAL; break; } /* * TODO: Validate embedded scope ID in source * list entry against passed-in ifp, if and only * if source list filter entry is iface or node local. */ in6_clearscope(&psin->sin6_addr); error = im6f_get_source(imf, psin, &lims); if (error) break; lims->im6sl_st[1] = imf->im6f_st[1]; } free(kss, M_TEMP); } if (error) goto out_im6f_rollback; INP_WLOCK_ASSERT(inp); IN6_MULTI_LIST_LOCK(); /* * Begin state merge transaction at MLD layer. */ CTR1(KTR_MLD, "%s: merge inm state", __func__); error = in6m_merge(inm, imf); if (error) CTR1(KTR_MLD, "%s: failed to merge inm state", __func__); else { CTR1(KTR_MLD, "%s: doing mld downcall", __func__); error = mld_change_state(inm, 0); if (error) CTR1(KTR_MLD, "%s: failed mld downcall", __func__); } IN6_MULTI_LIST_UNLOCK(); out_im6f_rollback: if (error) im6f_rollback(imf); else im6f_commit(imf); im6f_reap(imf); out_in6p_locked: INP_WUNLOCK(inp); return (error); } /* * Set the IP multicast options in response to user setsockopt(). * * Many of the socket options handled in this function duplicate the * functionality of socket options in the regular unicast API. However, * it is not possible to merge the duplicate code, because the idempotence * of the IPv6 multicast part of the BSD Sockets API must be preserved; * the effects of these options must be treated as separate and distinct. * * SMPng: XXX: Unlocked read of inp_socket believed OK. */ int ip6_setmoptions(struct inpcb *inp, struct sockopt *sopt) { struct ip6_moptions *im6o; int error; error = 0; /* * If socket is neither of type SOCK_RAW or SOCK_DGRAM, * or is a divert socket, reject it. */ if (inp->inp_socket->so_proto->pr_protocol == IPPROTO_DIVERT || (inp->inp_socket->so_proto->pr_type != SOCK_RAW && inp->inp_socket->so_proto->pr_type != SOCK_DGRAM)) return (EOPNOTSUPP); switch (sopt->sopt_name) { case IPV6_MULTICAST_IF: error = in6p_set_multicast_if(inp, sopt); break; case IPV6_MULTICAST_HOPS: { int hlim; if (sopt->sopt_valsize != sizeof(int)) { error = EINVAL; break; } error = sooptcopyin(sopt, &hlim, sizeof(hlim), sizeof(int)); if (error) break; if (hlim < -1 || hlim > 255) { error = EINVAL; break; } else if (hlim == -1) { hlim = V_ip6_defmcasthlim; } im6o = in6p_findmoptions(inp); im6o->im6o_multicast_hlim = hlim; INP_WUNLOCK(inp); break; } case IPV6_MULTICAST_LOOP: { u_int loop; /* * Set the loopback flag for outgoing multicast packets. * Must be zero or one. */ if (sopt->sopt_valsize != sizeof(u_int)) { error = EINVAL; break; } error = sooptcopyin(sopt, &loop, sizeof(u_int), sizeof(u_int)); if (error) break; if (loop > 1) { error = EINVAL; break; } im6o = in6p_findmoptions(inp); im6o->im6o_multicast_loop = loop; INP_WUNLOCK(inp); break; } case IPV6_JOIN_GROUP: case MCAST_JOIN_GROUP: case MCAST_JOIN_SOURCE_GROUP: error = in6p_join_group(inp, sopt); break; case IPV6_LEAVE_GROUP: case MCAST_LEAVE_GROUP: case MCAST_LEAVE_SOURCE_GROUP: error = in6p_leave_group(inp, sopt); break; case MCAST_BLOCK_SOURCE: case MCAST_UNBLOCK_SOURCE: error = in6p_block_unblock_source(inp, sopt); break; case IPV6_MSFILTER: error = in6p_set_source_filters(inp, sopt); break; default: error = EOPNOTSUPP; break; } INP_UNLOCK_ASSERT(inp); return (error); } /* * Expose MLD's multicast filter mode and source list(s) to userland, * keyed by (ifindex, group). * The filter mode is written out as a uint32_t, followed by * 0..n of struct in6_addr. * For use by ifmcstat(8). * SMPng: NOTE: unlocked read of ifindex space. */ static int sysctl_ip6_mcast_filters(SYSCTL_HANDLER_ARGS) { struct in6_addr mcaddr; struct in6_addr src; struct epoch_tracker et; struct ifnet *ifp; struct ifmultiaddr *ifma; struct in6_multi *inm; struct ip6_msource *ims; int *name; int retval; u_int namelen; uint32_t fmode, ifindex; #ifdef KTR char ip6tbuf[INET6_ADDRSTRLEN]; #endif name = (int *)arg1; namelen = arg2; if (req->newptr != NULL) return (EPERM); /* int: ifindex + 4 * 32 bits of IPv6 address */ if (namelen != 5) return (EINVAL); ifindex = name[0]; if (ifindex <= 0 || ifindex > V_if_index) { CTR2(KTR_MLD, "%s: ifindex %u out of range", __func__, ifindex); return (ENOENT); } memcpy(&mcaddr, &name[1], sizeof(struct in6_addr)); if (!IN6_IS_ADDR_MULTICAST(&mcaddr)) { CTR2(KTR_MLD, "%s: group %s is not multicast", __func__, ip6_sprintf(ip6tbuf, &mcaddr)); return (EINVAL); } NET_EPOCH_ENTER(et); ifp = ifnet_byindex(ifindex); if (ifp == NULL) { NET_EPOCH_EXIT(et); CTR2(KTR_MLD, "%s: no ifp for ifindex %u", __func__, ifindex); return (ENOENT); } /* * Internal MLD lookups require that scope/zone ID is set. */ (void)in6_setscope(&mcaddr, ifp, NULL); retval = sysctl_wire_old_buffer(req, sizeof(uint32_t) + (in6_mcast_maxgrpsrc * sizeof(struct in6_addr))); if (retval) { NET_EPOCH_EXIT(et); return (retval); } IN6_MULTI_LOCK(); IN6_MULTI_LIST_LOCK(); CK_STAILQ_FOREACH(ifma, &ifp->if_multiaddrs, ifma_link) { inm = in6m_ifmultiaddr_get_inm(ifma); if (inm == NULL) continue; if (!IN6_ARE_ADDR_EQUAL(&inm->in6m_addr, &mcaddr)) continue; fmode = inm->in6m_st[1].iss_fmode; retval = SYSCTL_OUT(req, &fmode, sizeof(uint32_t)); if (retval != 0) break; RB_FOREACH(ims, ip6_msource_tree, &inm->in6m_srcs) { CTR2(KTR_MLD, "%s: visit node %p", __func__, ims); /* * Only copy-out sources which are in-mode. */ if (fmode != im6s_get_mode(inm, ims, 1)) { CTR1(KTR_MLD, "%s: skip non-in-mode", __func__); continue; } src = ims->im6s_addr; retval = SYSCTL_OUT(req, &src, sizeof(struct in6_addr)); if (retval != 0) break; } } IN6_MULTI_LIST_UNLOCK(); IN6_MULTI_UNLOCK(); NET_EPOCH_EXIT(et); return (retval); } #ifdef KTR static const char *in6m_modestrs[] = { "un", "in", "ex" }; static const char * in6m_mode_str(const int mode) { if (mode >= MCAST_UNDEFINED && mode <= MCAST_EXCLUDE) return (in6m_modestrs[mode]); return ("??"); } static const char *in6m_statestrs[] = { "not-member", "silent", "idle", "lazy", "sleeping", "awakening", "query-pending", "sg-query-pending", "leaving" }; static const char * in6m_state_str(const int state) { if (state >= MLD_NOT_MEMBER && state <= MLD_LEAVING_MEMBER) return (in6m_statestrs[state]); return ("??"); } /* * Dump an in6_multi structure to the console. */ void in6m_print(const struct in6_multi *inm) { int t; char ip6tbuf[INET6_ADDRSTRLEN]; if ((ktr_mask & KTR_MLD) == 0) return; printf("%s: --- begin in6m %p ---\n", __func__, inm); printf("addr %s ifp %p(%s) ifma %p\n", ip6_sprintf(ip6tbuf, &inm->in6m_addr), inm->in6m_ifp, if_name(inm->in6m_ifp), inm->in6m_ifma); printf("timer %u state %s refcount %u scq.len %u\n", inm->in6m_timer, in6m_state_str(inm->in6m_state), inm->in6m_refcount, mbufq_len(&inm->in6m_scq)); printf("mli %p nsrc %lu sctimer %u scrv %u\n", inm->in6m_mli, inm->in6m_nsrc, inm->in6m_sctimer, inm->in6m_scrv); for (t = 0; t < 2; t++) { printf("t%d: fmode %s asm %u ex %u in %u rec %u\n", t, in6m_mode_str(inm->in6m_st[t].iss_fmode), inm->in6m_st[t].iss_asm, inm->in6m_st[t].iss_ex, inm->in6m_st[t].iss_in, inm->in6m_st[t].iss_rec); } printf("%s: --- end in6m %p ---\n", __func__, inm); } #else /* !KTR */ void in6m_print(const struct in6_multi *inm) { } #endif /* KTR */ Index: head/sys/sys/gtaskqueue.h =================================================================== --- head/sys/sys/gtaskqueue.h (revision 359437) +++ head/sys/sys/gtaskqueue.h (revision 359438) @@ -1,125 +1,122 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2014 Jeffrey Roberson * Copyright (c) 2016 Matthew Macy * 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 _SYS_GTASKQUEUE_H_ #define _SYS_GTASKQUEUE_H_ #ifndef _KERNEL #error "no user-serviceable parts inside" #endif #include #include #include #include struct gtaskqueue; /* * Taskqueue groups. Manages dynamic thread groups and irq binding for * device and other tasks. */ struct grouptask { struct gtask gt_task; void *gt_taskqueue; LIST_ENTRY(grouptask) gt_list; void *gt_uniq; #define GROUPTASK_NAMELEN 32 char gt_name[GROUPTASK_NAMELEN]; device_t gt_dev; struct resource *gt_irq; int gt_cpu; }; void gtaskqueue_block(struct gtaskqueue *queue); void gtaskqueue_unblock(struct gtaskqueue *queue); int gtaskqueue_cancel(struct gtaskqueue *queue, struct gtask *gtask); void gtaskqueue_drain(struct gtaskqueue *queue, struct gtask *task); void gtaskqueue_drain_all(struct gtaskqueue *queue); void grouptask_block(struct grouptask *grouptask); void grouptask_unblock(struct grouptask *grouptask); int grouptaskqueue_enqueue(struct gtaskqueue *queue, struct gtask *task); void taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *grptask, void *uniq, device_t dev, struct resource *irq, const char *name); int taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *grptask, void *uniq, int cpu, device_t dev, struct resource *irq, const char *name); void taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask); struct taskqgroup *taskqgroup_create(const char *name, int cnt, int stride); void taskqgroup_destroy(struct taskqgroup *qgroup); void taskqgroup_bind(struct taskqgroup *qgroup); -void taskqgroup_config_gtask_init(void *ctx, struct grouptask *gtask, - gtask_fn_t *fn, const char *name); -void taskqgroup_config_gtask_deinit(struct grouptask *gtask); #define GTASK_INIT(gtask, flags, priority, func, context) do { \ (gtask)->ta_flags = flags; \ (gtask)->ta_priority = (priority); \ (gtask)->ta_func = (func); \ (gtask)->ta_context = (context); \ } while (0) #define GROUPTASK_INIT(gtask, priority, func, context) \ GTASK_INIT(&(gtask)->gt_task, 0, priority, func, context) #define GROUPTASK_ENQUEUE(gtask) \ grouptaskqueue_enqueue((gtask)->gt_taskqueue, &(gtask)->gt_task) #define TASKQGROUP_DECLARE(name) \ extern struct taskqgroup *qgroup_##name #define TASKQGROUP_DEFINE(name, cnt, stride) \ \ struct taskqgroup *qgroup_##name; \ \ static void \ taskqgroup_define_##name(void *arg) \ { \ qgroup_##name = taskqgroup_create(#name, (cnt), (stride)); \ } \ SYSINIT(taskqgroup_##name, SI_SUB_TASKQ, SI_ORDER_FIRST, \ taskqgroup_define_##name, NULL); \ \ static void \ taskqgroup_bind_##name(void *arg) \ { \ taskqgroup_bind(qgroup_##name); \ } \ SYSINIT(taskqgroup_bind_##name, SI_SUB_SMP, SI_ORDER_ANY, \ taskqgroup_bind_##name, NULL) TASKQGROUP_DECLARE(softirq); #endif /* !_SYS_GTASKQUEUE_H_ */