Index: head/sys/dev/netmap/netmap.c =================================================================== --- head/sys/dev/netmap/netmap.c (revision 285358) +++ head/sys/dev/netmap/netmap.c (revision 285359) @@ -1,3162 +1,3162 @@ /* * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. 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. */ /* * $FreeBSD$ * * This module supports memory mapped access to network devices, * see netmap(4). * * The module uses a large, memory pool allocated by the kernel * and accessible as mmapped memory by multiple userspace threads/processes. * The memory pool contains packet buffers and "netmap rings", * i.e. user-accessible copies of the interface's queues. * * Access to the network card works like this: * 1. a process/thread issues one or more open() on /dev/netmap, to create * select()able file descriptor on which events are reported. * 2. on each descriptor, the process issues an ioctl() to identify * the interface that should report events to the file descriptor. * 3. on each descriptor, the process issues an mmap() request to * map the shared memory region within the process' address space. * The list of interesting queues is indicated by a location in * the shared memory region. * 4. using the functions in the netmap(4) userspace API, a process * can look up the occupation state of a queue, access memory buffers, * and retrieve received packets or enqueue packets to transmit. * 5. using some ioctl()s the process can synchronize the userspace view * of the queue with the actual status in the kernel. This includes both * receiving the notification of new packets, and transmitting new * packets on the output interface. * 6. select() or poll() can be used to wait for events on individual * transmit or receive queues (or all queues for a given interface). * SYNCHRONIZATION (USER) The netmap rings and data structures may be shared among multiple user threads or even independent processes. Any synchronization among those threads/processes is delegated to the threads themselves. Only one thread at a time can be in a system call on the same netmap ring. The OS does not enforce this and only guarantees against system crashes in case of invalid usage. LOCKING (INTERNAL) Within the kernel, access to the netmap rings is protected as follows: - a spinlock on each ring, to handle producer/consumer races on RX rings attached to the host stack (against multiple host threads writing from the host stack to the same ring), and on 'destination' rings attached to a VALE switch (i.e. RX rings in VALE ports, and TX rings in NIC/host ports) protecting multiple active senders for the same destination) - an atomic variable to guarantee that there is at most one instance of *_*xsync() on the ring at any time. For rings connected to user file descriptors, an atomic_test_and_set() protects this, and the lock on the ring is not actually used. For NIC RX rings connected to a VALE switch, an atomic_test_and_set() is also used to prevent multiple executions (the driver might indeed already guarantee this). For NIC TX rings connected to a VALE switch, the lock arbitrates access to the queue (both when allocating buffers and when pushing them out). - *xsync() should be protected against initializations of the card. On FreeBSD most devices have the reset routine protected by a RING lock (ixgbe, igb, em) or core lock (re). lem is missing the RING protection on rx_reset(), this should be added. On linux there is an external lock on the tx path, which probably also arbitrates access to the reset routine. XXX to be revised - a per-interface core_lock protecting access from the host stack while interfaces may be detached from netmap mode. XXX there should be no need for this lock if we detach the interfaces only while they are down. --- VALE SWITCH --- NMG_LOCK() serializes all modifications to switches and ports. A switch cannot be deleted until all ports are gone. For each switch, an SX lock (RWlock on linux) protects deletion of ports. When configuring or deleting a new port, the lock is acquired in exclusive mode (after holding NMG_LOCK). When forwarding, the lock is acquired in shared mode (without NMG_LOCK). The lock is held throughout the entire forwarding cycle, during which the thread may incur in a page fault. Hence it is important that sleepable shared locks are used. On the rx ring, the per-port lock is grabbed initially to reserve a number of slot in the ring, then the lock is released, packets are copied from source to destination, and then the lock is acquired again and the receive ring is updated. (A similar thing is done on the tx ring for NIC and host stack ports attached to the switch) */ /* --- internals ---- * * Roadmap to the code that implements the above. * * > 1. a process/thread issues one or more open() on /dev/netmap, to create * > select()able file descriptor on which events are reported. * * Internally, we allocate a netmap_priv_d structure, that will be * initialized on ioctl(NIOCREGIF). * * os-specific: * FreeBSD: netmap_open (netmap_freebsd.c). The priv is * per-thread. * linux: linux_netmap_open (netmap_linux.c). The priv is * per-open. * * > 2. on each descriptor, the process issues an ioctl() to identify * > the interface that should report events to the file descriptor. * * Implemented by netmap_ioctl(), NIOCREGIF case, with nmr->nr_cmd==0. * Most important things happen in netmap_get_na() and * netmap_do_regif(), called from there. Additional details can be * found in the comments above those functions. * * In all cases, this action creates/takes-a-reference-to a * netmap_*_adapter describing the port, and allocates a netmap_if * and all necessary netmap rings, filling them with netmap buffers. * * In this phase, the sync callbacks for each ring are set (these are used * in steps 5 and 6 below). The callbacks depend on the type of adapter. * The adapter creation/initialization code puts them in the * netmap_adapter (fields na->nm_txsync and na->nm_rxsync). Then, they * are copied from there to the netmap_kring's during netmap_do_regif(), by * the nm_krings_create() callback. All the nm_krings_create callbacks * actually call netmap_krings_create() to perform this and the other * common stuff. netmap_krings_create() also takes care of the host rings, * if needed, by setting their sync callbacks appropriately. * * Additional actions depend on the kind of netmap_adapter that has been * registered: * * - netmap_hw_adapter: [netmap.c] * This is a system netdev/ifp with native netmap support. * The ifp is detached from the host stack by redirecting: * - transmissions (from the network stack) to netmap_transmit() * - receive notifications to the nm_notify() callback for * this adapter. The callback is normally netmap_notify(), unless * the ifp is attached to a bridge using bwrap, in which case it * is netmap_bwrap_intr_notify(). * * - netmap_generic_adapter: [netmap_generic.c] * A system netdev/ifp without native netmap support. * * (the decision about native/non native support is taken in * netmap_get_hw_na(), called by netmap_get_na()) * * - netmap_vp_adapter [netmap_vale.c] * Returned by netmap_get_bdg_na(). * This is a persistent or ephemeral VALE port. Ephemeral ports * are created on the fly if they don't already exist, and are * always attached to a bridge. * Persistent VALE ports must must be created seperately, and i * then attached like normal NICs. The NIOCREGIF we are examining * will find them only if they had previosly been created and * attached (see VALE_CTL below). * * - netmap_pipe_adapter [netmap_pipe.c] * Returned by netmap_get_pipe_na(). * Both pipe ends are created, if they didn't already exist. * * - netmap_monitor_adapter [netmap_monitor.c] * Returned by netmap_get_monitor_na(). * If successful, the nm_sync callbacks of the monitored adapter * will be intercepted by the returned monitor. * * - netmap_bwrap_adapter [netmap_vale.c] * Cannot be obtained in this way, see VALE_CTL below * * * os-specific: * linux: we first go through linux_netmap_ioctl() to * adapt the FreeBSD interface to the linux one. * * * > 3. on each descriptor, the process issues an mmap() request to * > map the shared memory region within the process' address space. * > The list of interesting queues is indicated by a location in * > the shared memory region. * * os-specific: * FreeBSD: netmap_mmap_single (netmap_freebsd.c). * linux: linux_netmap_mmap (netmap_linux.c). * * > 4. using the functions in the netmap(4) userspace API, a process * > can look up the occupation state of a queue, access memory buffers, * > and retrieve received packets or enqueue packets to transmit. * * these actions do not involve the kernel. * * > 5. using some ioctl()s the process can synchronize the userspace view * > of the queue with the actual status in the kernel. This includes both * > receiving the notification of new packets, and transmitting new * > packets on the output interface. * * These are implemented in netmap_ioctl(), NIOCTXSYNC and NIOCRXSYNC * cases. They invoke the nm_sync callbacks on the netmap_kring * structures, as initialized in step 2 and maybe later modified * by a monitor. Monitors, however, will always call the original * callback before doing anything else. * * * > 6. select() or poll() can be used to wait for events on individual * > transmit or receive queues (or all queues for a given interface). * * Implemented in netmap_poll(). This will call the same nm_sync() * callbacks as in step 5 above. * * os-specific: * linux: we first go through linux_netmap_poll() to adapt * the FreeBSD interface to the linux one. * * * ---- VALE_CTL ----- * * VALE switches are controlled by issuing a NIOCREGIF with a non-null * nr_cmd in the nmreq structure. These subcommands are handled by * netmap_bdg_ctl() in netmap_vale.c. Persistent VALE ports are created * and destroyed by issuing the NETMAP_BDG_NEWIF and NETMAP_BDG_DELIF * subcommands, respectively. * * Any network interface known to the system (including a persistent VALE * port) can be attached to a VALE switch by issuing the * NETMAP_BDG_ATTACH subcommand. After the attachment, persistent VALE ports * look exactly like ephemeral VALE ports (as created in step 2 above). The * attachment of other interfaces, instead, requires the creation of a * netmap_bwrap_adapter. Moreover, the attached interface must be put in * netmap mode. This may require the creation of a netmap_generic_adapter if * we have no native support for the interface, or if generic adapters have * been forced by sysctl. * * Both persistent VALE ports and bwraps are handled by netmap_get_bdg_na(), * called by nm_bdg_ctl_attach(), and discriminated by the nm_bdg_attach() * callback. In the case of the bwrap, the callback creates the * netmap_bwrap_adapter. The initialization of the bwrap is then * completed by calling netmap_do_regif() on it, in the nm_bdg_ctl() * callback (netmap_bwrap_bdg_ctl in netmap_vale.c). * A generic adapter for the wrapped ifp will be created if needed, when * netmap_get_bdg_na() calls netmap_get_hw_na(). * * * ---- DATAPATHS ----- * * -= SYSTEM DEVICE WITH NATIVE SUPPORT =- * * na == NA(ifp) == netmap_hw_adapter created in DEVICE_netmap_attach() * * - tx from netmap userspace: * concurrently: * 1) ioctl(NIOCTXSYNC)/netmap_poll() in process context * kring->nm_sync() == DEVICE_netmap_txsync() * 2) device interrupt handler * na->nm_notify() == netmap_notify() * - rx from netmap userspace: * concurrently: * 1) ioctl(NIOCRXSYNC)/netmap_poll() in process context * kring->nm_sync() == DEVICE_netmap_rxsync() * 2) device interrupt handler * na->nm_notify() == netmap_notify() * - rx from host stack * concurrently: * 1) host stack * netmap_transmit() * na->nm_notify == netmap_notify() * 2) ioctl(NIOCRXSYNC)/netmap_poll() in process context * kring->nm_sync() == netmap_rxsync_from_host_compat * netmap_rxsync_from_host(na, NULL, NULL) * - tx to host stack * ioctl(NIOCTXSYNC)/netmap_poll() in process context * kring->nm_sync() == netmap_txsync_to_host_compat * netmap_txsync_to_host(na) * NM_SEND_UP() * FreeBSD: na->if_input() == ?? XXX * linux: netif_rx() with NM_MAGIC_PRIORITY_RX * * * * -= SYSTEM DEVICE WITH GENERIC SUPPORT =- * * na == NA(ifp) == generic_netmap_adapter created in generic_netmap_attach() * * - tx from netmap userspace: * concurrently: * 1) ioctl(NIOCTXSYNC)/netmap_poll() in process context * kring->nm_sync() == generic_netmap_txsync() * linux: dev_queue_xmit() with NM_MAGIC_PRIORITY_TX * generic_ndo_start_xmit() * orig. dev. start_xmit * FreeBSD: na->if_transmit() == orig. dev if_transmit * 2) generic_mbuf_destructor() * na->nm_notify() == netmap_notify() * - rx from netmap userspace: * 1) ioctl(NIOCRXSYNC)/netmap_poll() in process context * kring->nm_sync() == generic_netmap_rxsync() * mbq_safe_dequeue() * 2) device driver * generic_rx_handler() * mbq_safe_enqueue() * na->nm_notify() == netmap_notify() * - rx from host stack: * concurrently: * 1) host stack * linux: generic_ndo_start_xmit() * netmap_transmit() * FreeBSD: ifp->if_input() == netmap_transmit * both: * na->nm_notify() == netmap_notify() * 2) ioctl(NIOCRXSYNC)/netmap_poll() in process context * kring->nm_sync() == netmap_rxsync_from_host_compat * netmap_rxsync_from_host(na, NULL, NULL) * - tx to host stack: * ioctl(NIOCTXSYNC)/netmap_poll() in process context * kring->nm_sync() == netmap_txsync_to_host_compat * netmap_txsync_to_host(na) * NM_SEND_UP() * FreeBSD: na->if_input() == ??? XXX * linux: netif_rx() with NM_MAGIC_PRIORITY_RX * * * -= VALE =- * * INCOMING: * * - VALE ports: * ioctl(NIOCTXSYNC)/netmap_poll() in process context * kring->nm_sync() == netmap_vp_txsync() * * - system device with native support: * from cable: * interrupt * na->nm_notify() == netmap_bwrap_intr_notify(ring_nr != host ring) * kring->nm_sync() == DEVICE_netmap_rxsync() * netmap_vp_txsync() * kring->nm_sync() == DEVICE_netmap_rxsync() * from host stack: * netmap_transmit() * na->nm_notify() == netmap_bwrap_intr_notify(ring_nr == host ring) * kring->nm_sync() == netmap_rxsync_from_host_compat() * netmap_vp_txsync() * * - system device with generic support: * from device driver: * generic_rx_handler() * na->nm_notify() == netmap_bwrap_intr_notify(ring_nr != host ring) * kring->nm_sync() == generic_netmap_rxsync() * netmap_vp_txsync() * kring->nm_sync() == generic_netmap_rxsync() * from host stack: * netmap_transmit() * na->nm_notify() == netmap_bwrap_intr_notify(ring_nr == host ring) * kring->nm_sync() == netmap_rxsync_from_host_compat() * netmap_vp_txsync() * * (all cases) --> nm_bdg_flush() * dest_na->nm_notify() == (see below) * * OUTGOING: * * - VALE ports: * concurrently: * 1) ioctlNIOCRXSYNC)/netmap_poll() in process context * kring->nm_sync() == netmap_vp_rxsync() * 2) from nm_bdg_flush() * na->nm_notify() == netmap_notify() * * - system device with native support: * to cable: * na->nm_notify() == netmap_bwrap_notify() * netmap_vp_rxsync() * kring->nm_sync() == DEVICE_netmap_txsync() * netmap_vp_rxsync() * to host stack: * netmap_vp_rxsync() * kring->nm_sync() == netmap_txsync_to_host_compat * netmap_vp_rxsync_locked() * * - system device with generic adapter: * to device driver: * na->nm_notify() == netmap_bwrap_notify() * netmap_vp_rxsync() * kring->nm_sync() == generic_netmap_txsync() * netmap_vp_rxsync() * to host stack: * netmap_vp_rxsync() * kring->nm_sync() == netmap_txsync_to_host_compat * netmap_vp_rxsync() * */ /* * OS-specific code that is used only within this file. * Other OS-specific code that must be accessed by drivers * is present in netmap_kern.h */ #if defined(__FreeBSD__) #include /* prerequisite */ #include #include #include /* defines used in kernel.h */ #include /* types used in module initialization */ #include /* cdevsw struct, UID, GID */ #include /* FIONBIO */ #include #include /* struct socket */ #include #include #include #include /* sockaddrs */ #include #include #include #include #include #include #include /* BIOCIMMEDIATE */ #include /* bus_dmamap_* */ #include #include /* reduce conditional code */ // linux API, use for the knlist in FreeBSD /* use a private mutex for the knlist */ #define init_waitqueue_head(x) do { \ struct mtx *m = &(x)->m; \ mtx_init(m, "nm_kn_lock", NULL, MTX_DEF); \ knlist_init_mtx(&(x)->si.si_note, m); \ } while (0) #define OS_selrecord(a, b) selrecord(a, &((b)->si)) #define OS_selwakeup(a, b) freebsd_selwakeup(a, b) #elif defined(linux) #include "bsd_glue.h" #elif defined(__APPLE__) #warning OSX support is only partial #include "osx_glue.h" #else #error Unsupported platform #endif /* unsupported */ /* * common headers */ #include #include #include MALLOC_DEFINE(M_NETMAP, "netmap", "Network memory map"); /* user-controlled variables */ int netmap_verbose; static int netmap_no_timestamp; /* don't timestamp on rxsync */ SYSCTL_NODE(_dev, OID_AUTO, netmap, CTLFLAG_RW, 0, "Netmap args"); SYSCTL_INT(_dev_netmap, OID_AUTO, verbose, CTLFLAG_RW, &netmap_verbose, 0, "Verbose mode"); SYSCTL_INT(_dev_netmap, OID_AUTO, no_timestamp, CTLFLAG_RW, &netmap_no_timestamp, 0, "no_timestamp"); int netmap_mitigate = 1; SYSCTL_INT(_dev_netmap, OID_AUTO, mitigate, CTLFLAG_RW, &netmap_mitigate, 0, ""); int netmap_no_pendintr = 1; SYSCTL_INT(_dev_netmap, OID_AUTO, no_pendintr, CTLFLAG_RW, &netmap_no_pendintr, 0, "Always look for new received packets."); int netmap_txsync_retry = 2; SYSCTL_INT(_dev_netmap, OID_AUTO, txsync_retry, CTLFLAG_RW, &netmap_txsync_retry, 0 , "Number of txsync loops in bridge's flush."); int netmap_adaptive_io = 0; SYSCTL_INT(_dev_netmap, OID_AUTO, adaptive_io, CTLFLAG_RW, &netmap_adaptive_io, 0 , "Adaptive I/O on paravirt"); int netmap_flags = 0; /* debug flags */ int netmap_fwd = 0; /* force transparent mode */ /* * netmap_admode selects the netmap mode to use. * Invalid values are reset to NETMAP_ADMODE_BEST */ enum { NETMAP_ADMODE_BEST = 0, /* use native, fallback to generic */ NETMAP_ADMODE_NATIVE, /* either native or none */ NETMAP_ADMODE_GENERIC, /* force generic */ NETMAP_ADMODE_LAST }; static int netmap_admode = NETMAP_ADMODE_BEST; int netmap_generic_mit = 100*1000; /* Generic mitigation interval in nanoseconds. */ int netmap_generic_ringsize = 1024; /* Generic ringsize. */ int netmap_generic_rings = 1; /* number of queues in generic. */ SYSCTL_INT(_dev_netmap, OID_AUTO, flags, CTLFLAG_RW, &netmap_flags, 0 , ""); SYSCTL_INT(_dev_netmap, OID_AUTO, fwd, CTLFLAG_RW, &netmap_fwd, 0 , ""); SYSCTL_INT(_dev_netmap, OID_AUTO, admode, CTLFLAG_RW, &netmap_admode, 0 , ""); SYSCTL_INT(_dev_netmap, OID_AUTO, generic_mit, CTLFLAG_RW, &netmap_generic_mit, 0 , ""); SYSCTL_INT(_dev_netmap, OID_AUTO, generic_ringsize, CTLFLAG_RW, &netmap_generic_ringsize, 0 , ""); SYSCTL_INT(_dev_netmap, OID_AUTO, generic_rings, CTLFLAG_RW, &netmap_generic_rings, 0 , ""); NMG_LOCK_T netmap_global_lock; /* * mark the ring as stopped, and run through the locks * to make sure other users get to see it. */ static void netmap_disable_ring(struct netmap_kring *kr) { kr->nkr_stopped = 1; nm_kr_get(kr); mtx_lock(&kr->q_lock); mtx_unlock(&kr->q_lock); nm_kr_put(kr); } /* stop or enable a single ring */ void netmap_set_ring(struct netmap_adapter *na, u_int ring_id, enum txrx t, int stopped) { if (stopped) netmap_disable_ring(NMR(na, t) + ring_id); else NMR(na, t)[ring_id].nkr_stopped = 0; } /* stop or enable all the rings of na */ void netmap_set_all_rings(struct netmap_adapter *na, int stopped) { int i; enum txrx t; if (!nm_netmap_on(na)) return; for_rx_tx(t) { for (i = 0; i < netmap_real_rings(na, t); i++) { netmap_set_ring(na, i, t, stopped); } } } /* * Convenience function used in drivers. Waits for current txsync()s/rxsync()s * to finish and prevents any new one from starting. Call this before turning * netmap mode off, or before removing the harware rings (e.g., on module * onload). As a rule of thumb for linux drivers, this should be placed near * each napi_disable(). */ void netmap_disable_all_rings(struct ifnet *ifp) { netmap_set_all_rings(NA(ifp), 1 /* stopped */); } /* * Convenience function used in drivers. Re-enables rxsync and txsync on the * adapter's rings In linux drivers, this should be placed near each * napi_enable(). */ void netmap_enable_all_rings(struct ifnet *ifp) { netmap_set_all_rings(NA(ifp), 0 /* enabled */); } /* * generic bound_checking function */ u_int nm_bound_var(u_int *v, u_int dflt, u_int lo, u_int hi, const char *msg) { u_int oldv = *v; const char *op = NULL; if (dflt < lo) dflt = lo; if (dflt > hi) dflt = hi; if (oldv < lo) { *v = dflt; op = "Bump"; } else if (oldv > hi) { *v = hi; op = "Clamp"; } if (op && msg) printf("%s %s to %d (was %d)\n", op, msg, *v, oldv); return *v; } /* * packet-dump function, user-supplied or static buffer. * The destination buffer must be at least 30+4*len */ const char * nm_dump_buf(char *p, int len, int lim, char *dst) { static char _dst[8192]; int i, j, i0; static char hex[] ="0123456789abcdef"; char *o; /* output position */ #define P_HI(x) hex[((x) & 0xf0)>>4] #define P_LO(x) hex[((x) & 0xf)] #define P_C(x) ((x) >= 0x20 && (x) <= 0x7e ? (x) : '.') if (!dst) dst = _dst; if (lim <= 0 || lim > len) lim = len; o = dst; sprintf(o, "buf 0x%p len %d lim %d\n", p, len, lim); o += strlen(o); /* hexdump routine */ for (i = 0; i < lim; ) { sprintf(o, "%5d: ", i); o += strlen(o); memset(o, ' ', 48); i0 = i; for (j=0; j < 16 && i < lim; i++, j++) { o[j*3] = P_HI(p[i]); o[j*3+1] = P_LO(p[i]); } i = i0; for (j=0; j < 16 && i < lim; i++, j++) o[j + 48] = P_C(p[i]); o[j+48] = '\n'; o += j+49; } *o = '\0'; #undef P_HI #undef P_LO #undef P_C return dst; } /* * Fetch configuration from the device, to cope with dynamic * reconfigurations after loading the module. */ /* call with NMG_LOCK held */ int netmap_update_config(struct netmap_adapter *na) { u_int txr, txd, rxr, rxd; txr = txd = rxr = rxd = 0; if (na->nm_config == NULL || na->nm_config(na, &txr, &txd, &rxr, &rxd)) { /* take whatever we had at init time */ txr = na->num_tx_rings; txd = na->num_tx_desc; rxr = na->num_rx_rings; rxd = na->num_rx_desc; } if (na->num_tx_rings == txr && na->num_tx_desc == txd && na->num_rx_rings == rxr && na->num_rx_desc == rxd) return 0; /* nothing changed */ if (netmap_verbose || na->active_fds > 0) { D("stored config %s: txring %d x %d, rxring %d x %d", na->name, na->num_tx_rings, na->num_tx_desc, na->num_rx_rings, na->num_rx_desc); D("new config %s: txring %d x %d, rxring %d x %d", na->name, txr, txd, rxr, rxd); } if (na->active_fds == 0) { D("configuration changed (but fine)"); na->num_tx_rings = txr; na->num_tx_desc = txd; na->num_rx_rings = rxr; na->num_rx_desc = rxd; return 0; } D("configuration changed while active, this is bad..."); return 1; } +static void netmap_txsync_to_host(struct netmap_adapter *na); +static int netmap_rxsync_from_host(struct netmap_adapter *na, struct thread *td, void *pwait); + /* kring->nm_sync callback for the host tx ring */ static int netmap_txsync_to_host_compat(struct netmap_kring *kring, int flags) { (void)flags; /* unused */ netmap_txsync_to_host(kring->na); return 0; } /* kring->nm_sync callback for the host rx ring */ static int netmap_rxsync_from_host_compat(struct netmap_kring *kring, int flags) { (void)flags; /* unused */ netmap_rxsync_from_host(kring->na, NULL, NULL); return 0; } /* create the krings array and initialize the fields common to all adapters. * The array layout is this: * * +----------+ * na->tx_rings ----->| | \ * | | } na->num_tx_ring * | | / * +----------+ * | | host tx kring * na->rx_rings ----> +----------+ * | | \ * | | } na->num_rx_rings * | | / * +----------+ * | | host rx kring * +----------+ * na->tailroom ----->| | \ * | | } tailroom bytes * | | / * +----------+ * * Note: for compatibility, host krings are created even when not needed. * The tailroom space is currently used by vale ports for allocating leases. */ /* call with NMG_LOCK held */ int netmap_krings_create(struct netmap_adapter *na, u_int tailroom) { u_int i, len, ndesc; struct netmap_kring *kring; u_int n[NR_TXRX]; enum txrx t; /* account for the (possibly fake) host rings */ n[NR_TX] = na->num_tx_rings + 1; n[NR_RX] = na->num_rx_rings + 1; len = (n[NR_TX] + n[NR_RX]) * sizeof(struct netmap_kring) + tailroom; na->tx_rings = malloc((size_t)len, M_DEVBUF, M_NOWAIT | M_ZERO); if (na->tx_rings == NULL) { D("Cannot allocate krings"); return ENOMEM; } na->rx_rings = na->tx_rings + n[NR_TX]; /* * All fields in krings are 0 except the one initialized below. * but better be explicit on important kring fields. */ for_rx_tx(t) { ndesc = nma_get_ndesc(na, t); for (i = 0; i < n[t]; i++) { kring = &NMR(na, t)[i]; bzero(kring, sizeof(*kring)); kring->na = na; kring->ring_id = i; kring->tx = t; kring->nkr_num_slots = ndesc; if (i < nma_get_nrings(na, t)) { kring->nm_sync = (t == NR_TX ? na->nm_txsync : na->nm_rxsync); } else if (i == na->num_tx_rings) { kring->nm_sync = (t == NR_TX ? netmap_txsync_to_host_compat : netmap_rxsync_from_host_compat); } kring->nm_notify = na->nm_notify; kring->rhead = kring->rcur = kring->nr_hwcur = 0; /* * IMPORTANT: Always keep one slot empty. */ kring->rtail = kring->nr_hwtail = (t == NR_TX ? ndesc - 1 : 0); snprintf(kring->name, sizeof(kring->name) - 1, "%s %s%d", na->name, nm_txrx2str(t), i); ND("ktx %s h %d c %d t %d", kring->name, kring->rhead, kring->rcur, kring->rtail); mtx_init(&kring->q_lock, (t == NR_TX ? "nm_txq_lock" : "nm_rxq_lock"), NULL, MTX_DEF); init_waitqueue_head(&kring->si); } init_waitqueue_head(&na->si[t]); } na->tailroom = na->rx_rings + n[NR_RX]; return 0; } #ifdef __FreeBSD__ static void netmap_knlist_destroy(NM_SELINFO_T *si) { /* XXX kqueue(9) needed; these will mirror knlist_init. */ knlist_delete(&si->si.si_note, curthread, 0 /* not locked */ ); knlist_destroy(&si->si.si_note); /* now we don't need the mutex anymore */ mtx_destroy(&si->m); } #endif /* __FreeBSD__ */ /* undo the actions performed by netmap_krings_create */ /* call with NMG_LOCK held */ void netmap_krings_delete(struct netmap_adapter *na) { struct netmap_kring *kring = na->tx_rings; enum txrx t; for_rx_tx(t) netmap_knlist_destroy(&na->si[t]); /* we rely on the krings layout described above */ for ( ; kring != na->tailroom; kring++) { mtx_destroy(&kring->q_lock); netmap_knlist_destroy(&kring->si); } free(na->tx_rings, M_DEVBUF); na->tx_rings = na->rx_rings = na->tailroom = NULL; } /* * Destructor for NIC ports. They also have an mbuf queue * on the rings connected to the host so we need to purge * them first. */ /* call with NMG_LOCK held */ static void netmap_hw_krings_delete(struct netmap_adapter *na) { struct mbq *q = &na->rx_rings[na->num_rx_rings].rx_queue; ND("destroy sw mbq with len %d", mbq_len(q)); mbq_purge(q); mbq_safe_destroy(q); netmap_krings_delete(na); } /* * Undo everything that was done in netmap_do_regif(). In particular, * call nm_register(ifp,0) to stop netmap mode on the interface and * revert to normal operation. */ /* call with NMG_LOCK held */ static void netmap_unset_ringid(struct netmap_priv_d *); static void netmap_rel_exclusive(struct netmap_priv_d *); static void netmap_do_unregif(struct netmap_priv_d *priv) { struct netmap_adapter *na = priv->np_na; NMG_LOCK_ASSERT(); na->active_fds--; /* release exclusive use if it was requested on regif */ netmap_rel_exclusive(priv); if (na->active_fds <= 0) { /* last instance */ if (netmap_verbose) D("deleting last instance for %s", na->name); #ifdef WITH_MONITOR /* walk through all the rings and tell any monitor * that the port is going to exit netmap mode */ netmap_monitor_stop(na); #endif /* * (TO CHECK) This function is only called * when the last reference to this file descriptor goes * away. This means we cannot have any pending poll() * or interrupt routine operating on the structure. * XXX The file may be closed in a thread while * another thread is using it. * Linux keeps the file opened until the last reference * by any outstanding ioctl/poll or mmap is gone. * FreeBSD does not track mmap()s (but we do) and * wakes up any sleeping poll(). Need to check what * happens if the close() occurs while a concurrent * syscall is running. */ na->nm_register(na, 0); /* off, clear flags */ /* Wake up any sleeping threads. netmap_poll will * then return POLLERR * XXX The wake up now must happen during *_down(), when * we order all activities to stop. -gl */ /* delete rings and buffers */ netmap_mem_rings_delete(na); na->nm_krings_delete(na); } /* possibily decrement counter of tx_si/rx_si users */ netmap_unset_ringid(priv); /* delete the nifp */ netmap_mem_if_delete(na, priv->np_nifp); /* drop the allocator */ netmap_mem_deref(na->nm_mem, na); /* mark the priv as unregistered */ priv->np_na = NULL; priv->np_nifp = NULL; } /* call with NMG_LOCK held */ static __inline int nm_si_user(struct netmap_priv_d *priv, enum txrx t) { return (priv->np_na != NULL && (priv->np_qlast[t] - priv->np_qfirst[t] > 1)); } /* - * Destructor of the netmap_priv_d, called when the fd has - * no active open() and mmap(). - * Undo all the things done by NIOCREGIF. + * Destructor of the netmap_priv_d, called when the fd is closed + * Action: undo all the things done by NIOCREGIF, + * On FreeBSD we need to track whether there are active mmap()s, + * and we use np_active_mmaps for that. On linux, the field is always 0. + * Return: 1 if we can free priv, 0 otherwise. * - * returns 1 if this is the last instance and we can free priv */ /* call with NMG_LOCK held */ int netmap_dtor_locked(struct netmap_priv_d *priv) { struct netmap_adapter *na = priv->np_na; -#ifdef __FreeBSD__ - /* - * np_refcount is the number of active mmaps on - * this file descriptor - */ - if (--priv->np_refcount > 0) { + /* number of active mmaps on this fd (FreeBSD only) */ + if (--priv->np_refs > 0) { return 0; } -#endif /* __FreeBSD__ */ + if (!na) { - return 1; //XXX is it correct? + return 1; //XXX is it correct? } netmap_do_unregif(priv); netmap_adapter_put(na); return 1; } /* call with NMG_LOCK *not* held */ void netmap_dtor(void *data) { struct netmap_priv_d *priv = data; int last_instance; NMG_LOCK(); last_instance = netmap_dtor_locked(priv); NMG_UNLOCK(); if (last_instance) { bzero(priv, sizeof(*priv)); /* for safety */ free(priv, M_DEVBUF); } } /* * Handlers for synchronization of the queues from/to the host. * Netmap has two operating modes: * - in the default mode, the rings connected to the host stack are * just another ring pair managed by userspace; * - in transparent mode (XXX to be defined) incoming packets * (from the host or the NIC) are marked as NS_FORWARD upon * arrival, and the user application has a chance to reset the * flag for packets that should be dropped. * On the RXSYNC or poll(), packets in RX rings between * kring->nr_kcur and ring->cur with NS_FORWARD still set are moved * to the other side. * The transfer NIC --> host is relatively easy, just encapsulate * into mbufs and we are done. The host --> NIC side is slightly * harder because there might not be room in the tx ring so it * might take a while before releasing the buffer. */ /* * pass a chain of buffers to the host stack as coming from 'dst' * We do not need to lock because the queue is private. */ static void netmap_send_up(struct ifnet *dst, struct mbq *q) { struct mbuf *m; /* send packets up, outside the lock */ while ((m = mbq_dequeue(q)) != NULL) { if (netmap_verbose & NM_VERB_HOST) D("sending up pkt %p size %d", m, MBUF_LEN(m)); NM_SEND_UP(dst, m); } mbq_destroy(q); } /* * put a copy of the buffers marked NS_FORWARD into an mbuf chain. * Take packets from hwcur to ring->head marked NS_FORWARD (or forced) * and pass them up. Drop remaining packets in the unlikely event * of an mbuf shortage. */ static void netmap_grab_packets(struct netmap_kring *kring, struct mbq *q, int force) { u_int const lim = kring->nkr_num_slots - 1; u_int const head = kring->rhead; u_int n; struct netmap_adapter *na = kring->na; for (n = kring->nr_hwcur; n != head; n = nm_next(n, lim)) { struct mbuf *m; struct netmap_slot *slot = &kring->ring->slot[n]; if ((slot->flags & NS_FORWARD) == 0 && !force) continue; if (slot->len < 14 || slot->len > NETMAP_BUF_SIZE(na)) { RD(5, "bad pkt at %d len %d", n, slot->len); continue; } slot->flags &= ~NS_FORWARD; // XXX needed ? /* XXX TODO: adapt to the case of a multisegment packet */ m = m_devget(NMB(na, slot), slot->len, 0, na->ifp, NULL); if (m == NULL) break; mbq_enqueue(q, m); } } /* * Send to the NIC rings packets marked NS_FORWARD between * kring->nr_hwcur and kring->rhead * Called under kring->rx_queue.lock on the sw rx ring, */ static u_int netmap_sw_to_nic(struct netmap_adapter *na) { struct netmap_kring *kring = &na->rx_rings[na->num_rx_rings]; struct netmap_slot *rxslot = kring->ring->slot; u_int i, rxcur = kring->nr_hwcur; u_int const head = kring->rhead; u_int const src_lim = kring->nkr_num_slots - 1; u_int sent = 0; /* scan rings to find space, then fill as much as possible */ for (i = 0; i < na->num_tx_rings; i++) { struct netmap_kring *kdst = &na->tx_rings[i]; struct netmap_ring *rdst = kdst->ring; u_int const dst_lim = kdst->nkr_num_slots - 1; /* XXX do we trust ring or kring->rcur,rtail ? */ for (; rxcur != head && !nm_ring_empty(rdst); rxcur = nm_next(rxcur, src_lim) ) { struct netmap_slot *src, *dst, tmp; u_int dst_cur = rdst->cur; src = &rxslot[rxcur]; if ((src->flags & NS_FORWARD) == 0 && !netmap_fwd) continue; sent++; dst = &rdst->slot[dst_cur]; tmp = *src; src->buf_idx = dst->buf_idx; src->flags = NS_BUF_CHANGED; dst->buf_idx = tmp.buf_idx; dst->len = tmp.len; dst->flags = NS_BUF_CHANGED; rdst->cur = nm_next(dst_cur, dst_lim); } /* if (sent) XXX txsync ? */ } return sent; } /* * netmap_txsync_to_host() passes packets up. We are called from a * system call in user process context, and the only contention * can be among multiple user threads erroneously calling * this routine concurrently. */ -void +static void netmap_txsync_to_host(struct netmap_adapter *na) { struct netmap_kring *kring = &na->tx_rings[na->num_tx_rings]; u_int const lim = kring->nkr_num_slots - 1; u_int const head = kring->rhead; struct mbq q; /* Take packets from hwcur to head and pass them up. * force head = cur since netmap_grab_packets() stops at head * In case of no buffers we give up. At the end of the loop, * the queue is drained in all cases. */ mbq_init(&q); netmap_grab_packets(kring, &q, 1 /* force */); ND("have %d pkts in queue", mbq_len(&q)); kring->nr_hwcur = head; kring->nr_hwtail = head + lim; if (kring->nr_hwtail > lim) kring->nr_hwtail -= lim + 1; netmap_send_up(na->ifp, &q); } /* * rxsync backend for packets coming from the host stack. * They have been put in kring->rx_queue by netmap_transmit(). * We protect access to the kring using kring->rx_queue.lock * * This routine also does the selrecord if called from the poll handler * (we know because td != NULL). * * NOTE: on linux, selrecord() is defined as a macro and uses pwait * as an additional hidden argument. * returns the number of packets delivered to tx queues in * transparent mode, or a negative value if error */ -int +static int netmap_rxsync_from_host(struct netmap_adapter *na, struct thread *td, void *pwait) { struct netmap_kring *kring = &na->rx_rings[na->num_rx_rings]; struct netmap_ring *ring = kring->ring; u_int nm_i, n; u_int const lim = kring->nkr_num_slots - 1; u_int const head = kring->rhead; int ret = 0; struct mbq *q = &kring->rx_queue, fq; (void)pwait; /* disable unused warnings */ (void)td; mbq_init(&fq); /* fq holds packets to be freed */ mbq_lock(q); /* First part: import newly received packets */ n = mbq_len(q); if (n) { /* grab packets from the queue */ struct mbuf *m; uint32_t stop_i; nm_i = kring->nr_hwtail; stop_i = nm_prev(nm_i, lim); while ( nm_i != stop_i && (m = mbq_dequeue(q)) != NULL ) { int len = MBUF_LEN(m); struct netmap_slot *slot = &ring->slot[nm_i]; m_copydata(m, 0, len, NMB(na, slot)); ND("nm %d len %d", nm_i, len); if (netmap_verbose) D("%s", nm_dump_buf(NMB(na, slot),len, 128, NULL)); slot->len = len; slot->flags = kring->nkr_slot_flags; nm_i = nm_next(nm_i, lim); mbq_enqueue(&fq, m); } kring->nr_hwtail = nm_i; } /* * Second part: skip past packets that userspace has released. */ nm_i = kring->nr_hwcur; if (nm_i != head) { /* something was released */ if (netmap_fwd || kring->ring->flags & NR_FORWARD) ret = netmap_sw_to_nic(na); kring->nr_hwcur = head; } /* access copies of cur,tail in the kring */ if (kring->rcur == kring->rtail && td) /* no bufs available */ OS_selrecord(td, &kring->si); mbq_unlock(q); mbq_purge(&fq); mbq_destroy(&fq); return ret; } /* Get a netmap adapter for the port. * * If it is possible to satisfy the request, return 0 * with *na containing the netmap adapter found. * Otherwise return an error code, with *na containing NULL. * * When the port is attached to a bridge, we always return * EBUSY. * Otherwise, if the port is already bound to a file descriptor, * then we unconditionally return the existing adapter into *na. * In all the other cases, we return (into *na) either native, * generic or NULL, according to the following table: * * native_support * active_fds dev.netmap.admode YES NO * ------------------------------------------------------- * >0 * NA(ifp) NA(ifp) * * 0 NETMAP_ADMODE_BEST NATIVE GENERIC * 0 NETMAP_ADMODE_NATIVE NATIVE NULL * 0 NETMAP_ADMODE_GENERIC GENERIC GENERIC * */ int netmap_get_hw_na(struct ifnet *ifp, struct netmap_adapter **na) { /* generic support */ int i = netmap_admode; /* Take a snapshot. */ struct netmap_adapter *prev_na; #ifdef WITH_GENERIC struct netmap_generic_adapter *gna; int error = 0; #endif *na = NULL; /* default */ /* reset in case of invalid value */ if (i < NETMAP_ADMODE_BEST || i >= NETMAP_ADMODE_LAST) i = netmap_admode = NETMAP_ADMODE_BEST; if (NETMAP_CAPABLE(ifp)) { prev_na = NA(ifp); /* If an adapter already exists, return it if * there are active file descriptors or if * netmap is not forced to use generic * adapters. */ if (NETMAP_OWNED_BY_ANY(prev_na) || i != NETMAP_ADMODE_GENERIC || prev_na->na_flags & NAF_FORCE_NATIVE #ifdef WITH_PIPES /* ugly, but we cannot allow an adapter switch * if some pipe is referring to this one */ || prev_na->na_next_pipe > 0 #endif ) { *na = prev_na; return 0; } } /* If there isn't native support and netmap is not allowed * to use generic adapters, we cannot satisfy the request. */ if (!NETMAP_CAPABLE(ifp) && i == NETMAP_ADMODE_NATIVE) return EOPNOTSUPP; #ifdef WITH_GENERIC /* Otherwise, create a generic adapter and return it, * saving the previously used netmap adapter, if any. * * Note that here 'prev_na', if not NULL, MUST be a * native adapter, and CANNOT be a generic one. This is * true because generic adapters are created on demand, and * destroyed when not used anymore. Therefore, if the adapter * currently attached to an interface 'ifp' is generic, it * must be that * (NA(ifp)->active_fds > 0 || NETMAP_OWNED_BY_KERN(NA(ifp))). * Consequently, if NA(ifp) is generic, we will enter one of * the branches above. This ensures that we never override * a generic adapter with another generic adapter. */ prev_na = NA(ifp); error = generic_netmap_attach(ifp); if (error) return error; *na = NA(ifp); gna = (struct netmap_generic_adapter*)NA(ifp); gna->prev = prev_na; /* save old na */ if (prev_na != NULL) { ifunit_ref(ifp->if_xname); // XXX add a refcount ? netmap_adapter_get(prev_na); } ND("Created generic NA %p (prev %p)", gna, gna->prev); return 0; #else /* !WITH_GENERIC */ return EOPNOTSUPP; #endif } /* * MUST BE CALLED UNDER NMG_LOCK() * * Get a refcounted reference to a netmap adapter attached * to the interface specified by nmr. * This is always called in the execution of an ioctl(). * * Return ENXIO if the interface specified by the request does * not exist, ENOTSUP if netmap is not supported by the interface, * EBUSY if the interface is already attached to a bridge, * EINVAL if parameters are invalid, ENOMEM if needed resources * could not be allocated. * If successful, hold a reference to the netmap adapter. * * No reference is kept on the real interface, which may then * disappear at any time. */ int netmap_get_na(struct nmreq *nmr, struct netmap_adapter **na, int create) { struct ifnet *ifp = NULL; int error = 0; struct netmap_adapter *ret = NULL; *na = NULL; /* default return value */ NMG_LOCK_ASSERT(); /* we cascade through all possibile types of netmap adapter. * All netmap_get_*_na() functions return an error and an na, * with the following combinations: * * error na * 0 NULL type doesn't match * !0 NULL type matches, but na creation/lookup failed * 0 !NULL type matches and na created/found * !0 !NULL impossible */ /* try to see if this is a monitor port */ error = netmap_get_monitor_na(nmr, na, create); if (error || *na != NULL) return error; /* try to see if this is a pipe port */ error = netmap_get_pipe_na(nmr, na, create); if (error || *na != NULL) return error; /* try to see if this is a bridge port */ error = netmap_get_bdg_na(nmr, na, create); if (error) return error; if (*na != NULL) /* valid match in netmap_get_bdg_na() */ goto out; /* * This must be a hardware na, lookup the name in the system. * Note that by hardware we actually mean "it shows up in ifconfig". * This may still be a tap, a veth/epair, or even a * persistent VALE port. */ ifp = ifunit_ref(nmr->nr_name); if (ifp == NULL) { return ENXIO; } error = netmap_get_hw_na(ifp, &ret); if (error) goto out; *na = ret; netmap_adapter_get(ret); out: if (error && ret != NULL) netmap_adapter_put(ret); if (ifp) if_rele(ifp); /* allow live unloading of drivers modules */ return error; } /* * validate parameters on entry for *_txsync() * Returns ring->cur if ok, or something >= kring->nkr_num_slots * in case of error. * * rhead, rcur and rtail=hwtail are stored from previous round. * hwcur is the next packet to send to the ring. * * We want * hwcur <= *rhead <= head <= cur <= tail = *rtail <= hwtail * * hwcur, rhead, rtail and hwtail are reliable */ static u_int nm_txsync_prologue(struct netmap_kring *kring) { #define NM_ASSERT(t) if (t) { D("fail " #t); goto error; } struct netmap_ring *ring = kring->ring; u_int head = ring->head; /* read only once */ u_int cur = ring->cur; /* read only once */ u_int n = kring->nkr_num_slots; ND(5, "%s kcur %d ktail %d head %d cur %d tail %d", kring->name, kring->nr_hwcur, kring->nr_hwtail, ring->head, ring->cur, ring->tail); #if 1 /* kernel sanity checks; but we can trust the kring. */ if (kring->nr_hwcur >= n || kring->rhead >= n || kring->rtail >= n || kring->nr_hwtail >= n) goto error; #endif /* kernel sanity checks */ /* * user sanity checks. We only use 'cur', * A, B, ... are possible positions for cur: * * 0 A cur B tail C n-1 * 0 D tail E cur F n-1 * * B, F, D are valid. A, C, E are wrong */ if (kring->rtail >= kring->rhead) { /* want rhead <= head <= rtail */ NM_ASSERT(head < kring->rhead || head > kring->rtail); /* and also head <= cur <= rtail */ NM_ASSERT(cur < head || cur > kring->rtail); } else { /* here rtail < rhead */ /* we need head outside rtail .. rhead */ NM_ASSERT(head > kring->rtail && head < kring->rhead); /* two cases now: head <= rtail or head >= rhead */ if (head <= kring->rtail) { /* want head <= cur <= rtail */ NM_ASSERT(cur < head || cur > kring->rtail); } else { /* head >= rhead */ /* cur must be outside rtail..head */ NM_ASSERT(cur > kring->rtail && cur < head); } } if (ring->tail != kring->rtail) { RD(5, "tail overwritten was %d need %d", ring->tail, kring->rtail); ring->tail = kring->rtail; } kring->rhead = head; kring->rcur = cur; return head; error: RD(5, "%s kring error: head %d cur %d tail %d rhead %d rcur %d rtail %d hwcur %d hwtail %d", kring->name, head, cur, ring->tail, kring->rhead, kring->rcur, kring->rtail, kring->nr_hwcur, kring->nr_hwtail); return n; #undef NM_ASSERT } /* * validate parameters on entry for *_rxsync() * Returns ring->head if ok, kring->nkr_num_slots on error. * * For a valid configuration, * hwcur <= head <= cur <= tail <= hwtail * * We only consider head and cur. * hwcur and hwtail are reliable. * */ static u_int nm_rxsync_prologue(struct netmap_kring *kring) { struct netmap_ring *ring = kring->ring; uint32_t const n = kring->nkr_num_slots; uint32_t head, cur; ND(5,"%s kc %d kt %d h %d c %d t %d", kring->name, kring->nr_hwcur, kring->nr_hwtail, ring->head, ring->cur, ring->tail); /* * Before storing the new values, we should check they do not * move backwards. However: * - head is not an issue because the previous value is hwcur; * - cur could in principle go back, however it does not matter * because we are processing a brand new rxsync() */ cur = kring->rcur = ring->cur; /* read only once */ head = kring->rhead = ring->head; /* read only once */ #if 1 /* kernel sanity checks */ if (kring->nr_hwcur >= n || kring->nr_hwtail >= n) goto error; #endif /* kernel sanity checks */ /* user sanity checks */ if (kring->nr_hwtail >= kring->nr_hwcur) { /* want hwcur <= rhead <= hwtail */ if (head < kring->nr_hwcur || head > kring->nr_hwtail) goto error; /* and also rhead <= rcur <= hwtail */ if (cur < head || cur > kring->nr_hwtail) goto error; } else { /* we need rhead outside hwtail..hwcur */ if (head < kring->nr_hwcur && head > kring->nr_hwtail) goto error; /* two cases now: head <= hwtail or head >= hwcur */ if (head <= kring->nr_hwtail) { /* want head <= cur <= hwtail */ if (cur < head || cur > kring->nr_hwtail) goto error; } else { /* cur must be outside hwtail..head */ if (cur < head && cur > kring->nr_hwtail) goto error; } } if (ring->tail != kring->rtail) { RD(5, "%s tail overwritten was %d need %d", kring->name, ring->tail, kring->rtail); ring->tail = kring->rtail; } return head; error: RD(5, "kring error: hwcur %d rcur %d hwtail %d head %d cur %d tail %d", kring->nr_hwcur, kring->rcur, kring->nr_hwtail, kring->rhead, kring->rcur, ring->tail); return n; } /* * Error routine called when txsync/rxsync detects an error. * Can't do much more than resetting head =cur = hwcur, tail = hwtail * Return 1 on reinit. * * This routine is only called by the upper half of the kernel. * It only reads hwcur (which is changed only by the upper half, too) * and hwtail (which may be changed by the lower half, but only on * a tx ring and only to increase it, so any error will be recovered * on the next call). For the above, we don't strictly need to call * it under lock. */ int netmap_ring_reinit(struct netmap_kring *kring) { struct netmap_ring *ring = kring->ring; u_int i, lim = kring->nkr_num_slots - 1; int errors = 0; // XXX KASSERT nm_kr_tryget RD(10, "called for %s", kring->name); // XXX probably wrong to trust userspace kring->rhead = ring->head; kring->rcur = ring->cur; kring->rtail = ring->tail; if (ring->cur > lim) errors++; if (ring->head > lim) errors++; if (ring->tail > lim) errors++; for (i = 0; i <= lim; i++) { u_int idx = ring->slot[i].buf_idx; u_int len = ring->slot[i].len; if (idx < 2 || idx >= kring->na->na_lut.objtotal) { RD(5, "bad index at slot %d idx %d len %d ", i, idx, len); ring->slot[i].buf_idx = 0; ring->slot[i].len = 0; } else if (len > NETMAP_BUF_SIZE(kring->na)) { ring->slot[i].len = 0; RD(5, "bad len at slot %d idx %d len %d", i, idx, len); } } if (errors) { RD(10, "total %d errors", errors); RD(10, "%s reinit, cur %d -> %d tail %d -> %d", kring->name, ring->cur, kring->nr_hwcur, ring->tail, kring->nr_hwtail); ring->head = kring->rhead = kring->nr_hwcur; ring->cur = kring->rcur = kring->nr_hwcur; ring->tail = kring->rtail = kring->nr_hwtail; } return (errors ? 1 : 0); } /* interpret the ringid and flags fields of an nmreq, by translating them * into a pair of intervals of ring indices: * * [priv->np_txqfirst, priv->np_txqlast) and * [priv->np_rxqfirst, priv->np_rxqlast) * */ int netmap_interp_ringid(struct netmap_priv_d *priv, uint16_t ringid, uint32_t flags) { struct netmap_adapter *na = priv->np_na; u_int j, i = ringid & NETMAP_RING_MASK; u_int reg = flags & NR_REG_MASK; enum txrx t; if (reg == NR_REG_DEFAULT) { /* convert from old ringid to flags */ if (ringid & NETMAP_SW_RING) { reg = NR_REG_SW; } else if (ringid & NETMAP_HW_RING) { reg = NR_REG_ONE_NIC; } else { reg = NR_REG_ALL_NIC; } D("deprecated API, old ringid 0x%x -> ringid %x reg %d", ringid, i, reg); } switch (reg) { case NR_REG_ALL_NIC: case NR_REG_PIPE_MASTER: case NR_REG_PIPE_SLAVE: for_rx_tx(t) { priv->np_qfirst[t] = 0; priv->np_qlast[t] = nma_get_nrings(na, t); } ND("%s %d %d", "ALL/PIPE", priv->np_qfirst[NR_RX], priv->np_qlast[NR_RX]); break; case NR_REG_SW: case NR_REG_NIC_SW: if (!(na->na_flags & NAF_HOST_RINGS)) { D("host rings not supported"); return EINVAL; } for_rx_tx(t) { priv->np_qfirst[t] = (reg == NR_REG_SW ? nma_get_nrings(na, t) : 0); priv->np_qlast[t] = nma_get_nrings(na, t) + 1; } ND("%s %d %d", reg == NR_REG_SW ? "SW" : "NIC+SW", priv->np_qfirst[NR_RX], priv->np_qlast[NR_RX]); break; case NR_REG_ONE_NIC: if (i >= na->num_tx_rings && i >= na->num_rx_rings) { D("invalid ring id %d", i); return EINVAL; } for_rx_tx(t) { /* if not enough rings, use the first one */ j = i; if (j >= nma_get_nrings(na, t)) j = 0; priv->np_qfirst[t] = j; priv->np_qlast[t] = j + 1; } break; default: D("invalid regif type %d", reg); return EINVAL; } priv->np_flags = (flags & ~NR_REG_MASK) | reg; if (netmap_verbose) { D("%s: tx [%d,%d) rx [%d,%d) id %d", na->name, priv->np_qfirst[NR_TX], priv->np_qlast[NR_TX], priv->np_qfirst[NR_RX], priv->np_qlast[NR_RX], i); } return 0; } /* * Set the ring ID. For devices with a single queue, a request * for all rings is the same as a single ring. */ static int netmap_set_ringid(struct netmap_priv_d *priv, uint16_t ringid, uint32_t flags) { struct netmap_adapter *na = priv->np_na; int error; enum txrx t; error = netmap_interp_ringid(priv, ringid, flags); if (error) { return error; } priv->np_txpoll = (ringid & NETMAP_NO_TX_POLL) ? 0 : 1; /* optimization: count the users registered for more than * one ring, which are the ones sleeping on the global queue. * The default netmap_notify() callback will then * avoid signaling the global queue if nobody is using it */ for_rx_tx(t) { if (nm_si_user(priv, t)) na->si_users[t]++; } return 0; } static void netmap_unset_ringid(struct netmap_priv_d *priv) { struct netmap_adapter *na = priv->np_na; enum txrx t; for_rx_tx(t) { if (nm_si_user(priv, t)) na->si_users[t]--; priv->np_qfirst[t] = priv->np_qlast[t] = 0; } priv->np_flags = 0; priv->np_txpoll = 0; } /* check that the rings we want to bind are not exclusively owned by a previous * bind. If exclusive ownership has been requested, we also mark the rings. */ static int netmap_get_exclusive(struct netmap_priv_d *priv) { struct netmap_adapter *na = priv->np_na; u_int i; struct netmap_kring *kring; int excl = (priv->np_flags & NR_EXCLUSIVE); enum txrx t; ND("%s: grabbing tx [%d, %d) rx [%d, %d)", na->name, priv->np_qfirst[NR_TX], priv->np_qlast[NR_TX], priv->np_qfirst[NR_RX], priv->np_qlast[NR_RX]); /* first round: check that all the requested rings * are neither alread exclusively owned, nor we * want exclusive ownership when they are already in use */ for_rx_tx(t) { for (i = priv->np_qfirst[t]; i < priv->np_qlast[t]; i++) { kring = &NMR(na, t)[i]; if ((kring->nr_kflags & NKR_EXCLUSIVE) || (kring->users && excl)) { ND("ring %s busy", kring->name); return EBUSY; } } } /* second round: increment usage cound and possibly * mark as exclusive */ for_rx_tx(t) { for (i = priv->np_qfirst[t]; i < priv->np_qlast[t]; i++) { kring = &NMR(na, t)[i]; kring->users++; if (excl) kring->nr_kflags |= NKR_EXCLUSIVE; } } return 0; } /* undo netmap_get_ownership() */ static void netmap_rel_exclusive(struct netmap_priv_d *priv) { struct netmap_adapter *na = priv->np_na; u_int i; struct netmap_kring *kring; int excl = (priv->np_flags & NR_EXCLUSIVE); enum txrx t; ND("%s: releasing tx [%d, %d) rx [%d, %d)", na->name, priv->np_qfirst[NR_TX], priv->np_qlast[NR_TX], priv->np_qfirst[NR_RX], priv->np_qlast[MR_RX]); for_rx_tx(t) { for (i = priv->np_qfirst[t]; i < priv->np_qlast[t]; i++) { kring = &NMR(na, t)[i]; if (excl) kring->nr_kflags &= ~NKR_EXCLUSIVE; kring->users--; } } } /* * possibly move the interface to netmap-mode. * If success it returns a pointer to netmap_if, otherwise NULL. * This must be called with NMG_LOCK held. * * The following na callbacks are called in the process: * * na->nm_config() [by netmap_update_config] * (get current number and size of rings) * * We have a generic one for linux (netmap_linux_config). * The bwrap has to override this, since it has to forward * the request to the wrapped adapter (netmap_bwrap_config). * * * na->nm_krings_create() * (create and init the krings array) * * One of the following: * * * netmap_hw_krings_create, (hw ports) * creates the standard layout for the krings * and adds the mbq (used for the host rings). * * * netmap_vp_krings_create (VALE ports) * add leases and scratchpads * * * netmap_pipe_krings_create (pipes) * create the krings and rings of both ends and * cross-link them * * * netmap_monitor_krings_create (monitors) * avoid allocating the mbq * * * netmap_bwrap_krings_create (bwraps) * create both the brap krings array, * the krings array of the wrapped adapter, and * (if needed) the fake array for the host adapter * * na->nm_register(, 1) * (put the adapter in netmap mode) * * This may be one of the following: * (XXX these should be either all *_register or all *_reg 2014-03-15) * * * netmap_hw_register (hw ports) * checks that the ifp is still there, then calls * the hardware specific callback; * * * netmap_vp_reg (VALE ports) * If the port is connected to a bridge, * set the NAF_NETMAP_ON flag under the * bridge write lock. * * * netmap_pipe_reg (pipes) * inform the other pipe end that it is no * longer responsibile for the lifetime of this * pipe end * * * netmap_monitor_reg (monitors) * intercept the sync callbacks of the monitored * rings * * * netmap_bwrap_register (bwraps) * cross-link the bwrap and hwna rings, * forward the request to the hwna, override * the hwna notify callback (to get the frames * coming from outside go through the bridge). * * */ int netmap_do_regif(struct netmap_priv_d *priv, struct netmap_adapter *na, uint16_t ringid, uint32_t flags) { struct netmap_if *nifp = NULL; int error; NMG_LOCK_ASSERT(); /* ring configuration may have changed, fetch from the card */ netmap_update_config(na); priv->np_na = na; /* store the reference */ error = netmap_set_ringid(priv, ringid, flags); if (error) goto err; error = netmap_mem_finalize(na->nm_mem, na); if (error) goto err; if (na->active_fds == 0) { /* * If this is the first registration of the adapter, * also create the netmap rings and their in-kernel view, * the netmap krings. */ /* * Depending on the adapter, this may also create * the netmap rings themselves */ error = na->nm_krings_create(na); if (error) goto err_drop_mem; /* create all missing netmap rings */ error = netmap_mem_rings_create(na); if (error) goto err_del_krings; } /* now the kring must exist and we can check whether some * previous bind has exclusive ownership on them */ error = netmap_get_exclusive(priv); if (error) goto err_del_rings; /* in all cases, create a new netmap if */ nifp = netmap_mem_if_new(na); if (nifp == NULL) { error = ENOMEM; goto err_rel_excl; } na->active_fds++; if (!nm_netmap_on(na)) { /* Netmap not active, set the card in netmap mode * and make it use the shared buffers. */ /* cache the allocator info in the na */ netmap_mem_get_lut(na->nm_mem, &na->na_lut); ND("%p->na_lut == %p", na, na->na_lut.lut); error = na->nm_register(na, 1); /* mode on */ if (error) goto err_del_if; } /* * advertise that the interface is ready by setting np_nifp. * The barrier is needed because readers (poll, *SYNC and mmap) * check for priv->np_nifp != NULL without locking */ mb(); /* make sure previous writes are visible to all CPUs */ priv->np_nifp = nifp; return 0; err_del_if: memset(&na->na_lut, 0, sizeof(na->na_lut)); na->active_fds--; netmap_mem_if_delete(na, nifp); err_rel_excl: netmap_rel_exclusive(priv); err_del_rings: if (na->active_fds == 0) netmap_mem_rings_delete(na); err_del_krings: if (na->active_fds == 0) na->nm_krings_delete(na); err_drop_mem: netmap_mem_deref(na->nm_mem, na); err: priv->np_na = NULL; return error; } /* * update kring and ring at the end of txsync. */ static inline void nm_txsync_finalize(struct netmap_kring *kring) { /* update ring tail to what the kernel knows */ kring->ring->tail = kring->rtail = kring->nr_hwtail; /* note, head/rhead/hwcur might be behind cur/rcur * if no carrier */ ND(5, "%s now hwcur %d hwtail %d head %d cur %d tail %d", kring->name, kring->nr_hwcur, kring->nr_hwtail, kring->rhead, kring->rcur, kring->rtail); } /* * update kring and ring at the end of rxsync */ static inline void nm_rxsync_finalize(struct netmap_kring *kring) { /* tell userspace that there might be new packets */ //struct netmap_ring *ring = kring->ring; ND("head %d cur %d tail %d -> %d", ring->head, ring->cur, ring->tail, kring->nr_hwtail); kring->ring->tail = kring->rtail = kring->nr_hwtail; /* make a copy of the state for next round */ kring->rhead = kring->ring->head; kring->rcur = kring->ring->cur; } /* * ioctl(2) support for the "netmap" device. * * Following a list of accepted commands: * - NIOCGINFO * - SIOCGIFADDR just for convenience * - NIOCREGIF * - NIOCTXSYNC * - NIOCRXSYNC * * Return 0 on success, errno otherwise. */ int netmap_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag, struct thread *td) { struct netmap_priv_d *priv = NULL; struct nmreq *nmr = (struct nmreq *) data; struct netmap_adapter *na = NULL; int error; u_int i, qfirst, qlast; struct netmap_if *nifp; struct netmap_kring *krings; enum txrx t; (void)dev; /* UNUSED */ (void)fflag; /* UNUSED */ if (cmd == NIOCGINFO || cmd == NIOCREGIF) { /* truncate name */ nmr->nr_name[sizeof(nmr->nr_name) - 1] = '\0'; if (nmr->nr_version != NETMAP_API) { D("API mismatch for %s got %d need %d", nmr->nr_name, nmr->nr_version, NETMAP_API); nmr->nr_version = NETMAP_API; } if (nmr->nr_version < NETMAP_MIN_API || nmr->nr_version > NETMAP_MAX_API) { return EINVAL; } } CURVNET_SET(TD_TO_VNET(td)); error = devfs_get_cdevpriv((void **)&priv); if (error) { CURVNET_RESTORE(); /* XXX ENOENT should be impossible, since the priv * is now created in the open */ return (error == ENOENT ? ENXIO : error); } switch (cmd) { case NIOCGINFO: /* return capabilities etc */ if (nmr->nr_cmd == NETMAP_BDG_LIST) { error = netmap_bdg_ctl(nmr, NULL); break; } NMG_LOCK(); do { /* memsize is always valid */ struct netmap_mem_d *nmd = &nm_mem; u_int memflags; if (nmr->nr_name[0] != '\0') { /* get a refcount */ error = netmap_get_na(nmr, &na, 1 /* create */); if (error) break; nmd = na->nm_mem; /* get memory allocator */ } error = netmap_mem_get_info(nmd, &nmr->nr_memsize, &memflags, &nmr->nr_arg2); if (error) break; if (na == NULL) /* only memory info */ break; nmr->nr_offset = 0; nmr->nr_rx_slots = nmr->nr_tx_slots = 0; netmap_update_config(na); nmr->nr_rx_rings = na->num_rx_rings; nmr->nr_tx_rings = na->num_tx_rings; nmr->nr_rx_slots = na->num_rx_desc; nmr->nr_tx_slots = na->num_tx_desc; netmap_adapter_put(na); } while (0); NMG_UNLOCK(); break; case NIOCREGIF: /* possibly attach/detach NIC and VALE switch */ i = nmr->nr_cmd; if (i == NETMAP_BDG_ATTACH || i == NETMAP_BDG_DETACH || i == NETMAP_BDG_VNET_HDR || i == NETMAP_BDG_NEWIF || i == NETMAP_BDG_DELIF) { error = netmap_bdg_ctl(nmr, NULL); break; } else if (i != 0) { D("nr_cmd must be 0 not %d", i); error = EINVAL; break; } /* protect access to priv from concurrent NIOCREGIF */ NMG_LOCK(); do { u_int memflags; if (priv->np_nifp != NULL) { /* thread already registered */ error = EBUSY; break; } /* find the interface and a reference */ error = netmap_get_na(nmr, &na, 1 /* create */); /* keep reference */ if (error) break; if (NETMAP_OWNED_BY_KERN(na)) { netmap_adapter_put(na); error = EBUSY; break; } error = netmap_do_regif(priv, na, nmr->nr_ringid, nmr->nr_flags); if (error) { /* reg. failed, release priv and ref */ netmap_adapter_put(na); break; } nifp = priv->np_nifp; priv->np_td = td; // XXX kqueue, debugging only /* return the offset of the netmap_if object */ nmr->nr_rx_rings = na->num_rx_rings; nmr->nr_tx_rings = na->num_tx_rings; nmr->nr_rx_slots = na->num_rx_desc; nmr->nr_tx_slots = na->num_tx_desc; error = netmap_mem_get_info(na->nm_mem, &nmr->nr_memsize, &memflags, &nmr->nr_arg2); if (error) { netmap_do_unregif(priv); netmap_adapter_put(na); break; } if (memflags & NETMAP_MEM_PRIVATE) { *(uint32_t *)(uintptr_t)&nifp->ni_flags |= NI_PRIV_MEM; } for_rx_tx(t) { priv->np_si[t] = nm_si_user(priv, t) ? &na->si[t] : &NMR(na, t)[priv->np_qfirst[t]].si; } if (nmr->nr_arg3) { D("requested %d extra buffers", nmr->nr_arg3); nmr->nr_arg3 = netmap_extra_alloc(na, &nifp->ni_bufs_head, nmr->nr_arg3); D("got %d extra buffers", nmr->nr_arg3); } nmr->nr_offset = netmap_mem_if_offset(na->nm_mem, nifp); } while (0); NMG_UNLOCK(); break; case NIOCTXSYNC: case NIOCRXSYNC: nifp = priv->np_nifp; if (nifp == NULL) { error = ENXIO; break; } mb(); /* make sure following reads are not from cache */ na = priv->np_na; /* we have a reference */ if (na == NULL) { D("Internal error: nifp != NULL && na == NULL"); error = ENXIO; break; } if (!nm_netmap_on(na)) { error = ENXIO; break; } t = (cmd == NIOCTXSYNC ? NR_TX : NR_RX); krings = NMR(na, t); qfirst = priv->np_qfirst[t]; qlast = priv->np_qlast[t]; for (i = qfirst; i < qlast; i++) { struct netmap_kring *kring = krings + i; if (nm_kr_tryget(kring)) { error = EBUSY; goto out; } if (cmd == NIOCTXSYNC) { if (netmap_verbose & NM_VERB_TXSYNC) D("pre txsync ring %d cur %d hwcur %d", i, kring->ring->cur, kring->nr_hwcur); if (nm_txsync_prologue(kring) >= kring->nkr_num_slots) { netmap_ring_reinit(kring); } else if (kring->nm_sync(kring, NAF_FORCE_RECLAIM) == 0) { nm_txsync_finalize(kring); } if (netmap_verbose & NM_VERB_TXSYNC) D("post txsync ring %d cur %d hwcur %d", i, kring->ring->cur, kring->nr_hwcur); } else { if (nm_rxsync_prologue(kring) >= kring->nkr_num_slots) { netmap_ring_reinit(kring); } else if (kring->nm_sync(kring, NAF_FORCE_READ) == 0) { nm_rxsync_finalize(kring); } microtime(&na->rx_rings[i].ring->ts); } nm_kr_put(kring); } break; #ifdef WITH_VALE case NIOCCONFIG: error = netmap_bdg_config(nmr); break; #endif #ifdef __FreeBSD__ case FIONBIO: case FIOASYNC: ND("FIONBIO/FIOASYNC are no-ops"); break; case BIOCIMMEDIATE: case BIOCGHDRCMPLT: case BIOCSHDRCMPLT: case BIOCSSEESENT: D("ignore BIOCIMMEDIATE/BIOCSHDRCMPLT/BIOCSHDRCMPLT/BIOCSSEESENT"); break; default: /* allow device-specific ioctls */ { struct ifnet *ifp = ifunit_ref(nmr->nr_name); if (ifp == NULL) { error = ENXIO; } else { struct socket so; bzero(&so, sizeof(so)); so.so_vnet = ifp->if_vnet; // so->so_proto not null. error = ifioctl(&so, cmd, data, td); if_rele(ifp); } break; } #else /* linux */ default: error = EOPNOTSUPP; #endif /* linux */ } out: CURVNET_RESTORE(); return (error); } /* * select(2) and poll(2) handlers for the "netmap" device. * * Can be called for one or more queues. * Return true the event mask corresponding to ready events. * If there are no ready events, do a selrecord on either individual * selinfo or on the global one. * Device-dependent parts (locking and sync of tx/rx rings) * are done through callbacks. * * On linux, arguments are really pwait, the poll table, and 'td' is struct file * * The first one is remapped to pwait as selrecord() uses the name as an * hidden argument. */ int netmap_poll(struct cdev *dev, int events, struct thread *td) { struct netmap_priv_d *priv = NULL; struct netmap_adapter *na; struct netmap_kring *kring; u_int i, check_all_tx, check_all_rx, want[NR_TXRX], revents = 0; #define want_tx want[NR_TX] #define want_rx want[NR_RX] struct mbq q; /* packets from hw queues to host stack */ void *pwait = dev; /* linux compatibility */ int is_kevent = 0; enum txrx t; /* * In order to avoid nested locks, we need to "double check" * txsync and rxsync if we decide to do a selrecord(). * retry_tx (and retry_rx, later) prevent looping forever. */ int retry_tx = 1, retry_rx = 1; (void)pwait; mbq_init(&q); /* * XXX kevent has curthread->tp_fop == NULL, * so devfs_get_cdevpriv() fails. We circumvent this by passing * priv as the first argument, which is also useful to avoid * the selrecord() which are not necessary in that case. */ if (devfs_get_cdevpriv((void **)&priv) != 0) { is_kevent = 1; if (netmap_verbose) D("called from kevent"); priv = (struct netmap_priv_d *)dev; } if (priv == NULL) return POLLERR; if (priv->np_nifp == NULL) { D("No if registered"); return POLLERR; } mb(); /* make sure following reads are not from cache */ na = priv->np_na; if (!nm_netmap_on(na)) return POLLERR; if (netmap_verbose & 0x8000) D("device %s events 0x%x", na->name, events); want_tx = events & (POLLOUT | POLLWRNORM); want_rx = events & (POLLIN | POLLRDNORM); /* * check_all_{tx|rx} are set if the card has more than one queue AND * the file descriptor is bound to all of them. If so, we sleep on * the "global" selinfo, otherwise we sleep on individual selinfo * (FreeBSD only allows two selinfo's per file descriptor). * The interrupt routine in the driver wake one or the other * (or both) depending on which clients are active. * * rxsync() is only called if we run out of buffers on a POLLIN. * txsync() is called if we run out of buffers on POLLOUT, or * there are pending packets to send. The latter can be disabled * passing NETMAP_NO_TX_POLL in the NIOCREG call. */ check_all_tx = nm_si_user(priv, NR_TX); check_all_rx = nm_si_user(priv, NR_RX); /* * We start with a lock free round which is cheap if we have * slots available. If this fails, then lock and call the sync * routines. */ for_rx_tx(t) { for (i = priv->np_qfirst[t]; want[t] && i < priv->np_qlast[t]; i++) { kring = &NMR(na, t)[i]; /* XXX compare ring->cur and kring->tail */ if (!nm_ring_empty(kring->ring)) { revents |= want[t]; want[t] = 0; /* also breaks the loop */ } } } /* * If we want to push packets out (priv->np_txpoll) or * want_tx is still set, we must issue txsync calls * (on all rings, to avoid that the tx rings stall). * XXX should also check cur != hwcur on the tx rings. * Fortunately, normal tx mode has np_txpoll set. */ if (priv->np_txpoll || want_tx) { /* * The first round checks if anyone is ready, if not * do a selrecord and another round to handle races. * want_tx goes to 0 if any space is found, and is * used to skip rings with no pending transmissions. */ flush_tx: for (i = priv->np_qfirst[NR_TX]; i < priv->np_qlast[NR_RX]; i++) { int found = 0; kring = &na->tx_rings[i]; if (!want_tx && kring->ring->cur == kring->nr_hwcur) continue; /* only one thread does txsync */ if (nm_kr_tryget(kring)) { /* either busy or stopped * XXX if the ring is stopped, sleeping would * be better. In current code, however, we only * stop the rings for brief intervals (2014-03-14) */ if (netmap_verbose) RD(2, "%p lost race on txring %d, ok", priv, i); continue; } if (nm_txsync_prologue(kring) >= kring->nkr_num_slots) { netmap_ring_reinit(kring); revents |= POLLERR; } else { if (kring->nm_sync(kring, 0)) revents |= POLLERR; else nm_txsync_finalize(kring); } /* * If we found new slots, notify potential * listeners on the same ring. * Since we just did a txsync, look at the copies * of cur,tail in the kring. */ found = kring->rcur != kring->rtail; nm_kr_put(kring); if (found) { /* notify other listeners */ revents |= want_tx; want_tx = 0; kring->nm_notify(kring, 0); } } if (want_tx && retry_tx && !is_kevent) { OS_selrecord(td, check_all_tx ? &na->si[NR_TX] : &na->tx_rings[priv->np_qfirst[NR_TX]].si); retry_tx = 0; goto flush_tx; } } /* * If want_rx is still set scan receive rings. * Do it on all rings because otherwise we starve. */ if (want_rx) { int send_down = 0; /* transparent mode */ /* two rounds here for race avoidance */ do_retry_rx: for (i = priv->np_qfirst[NR_RX]; i < priv->np_qlast[NR_RX]; i++) { int found = 0; kring = &na->rx_rings[i]; if (nm_kr_tryget(kring)) { if (netmap_verbose) RD(2, "%p lost race on rxring %d, ok", priv, i); continue; } if (nm_rxsync_prologue(kring) >= kring->nkr_num_slots) { netmap_ring_reinit(kring); revents |= POLLERR; } /* now we can use kring->rcur, rtail */ /* * transparent mode support: collect packets * from the rxring(s). * XXX NR_FORWARD should only be read on * physical or NIC ports */ if (netmap_fwd ||kring->ring->flags & NR_FORWARD) { ND(10, "forwarding some buffers up %d to %d", kring->nr_hwcur, kring->ring->cur); netmap_grab_packets(kring, &q, netmap_fwd); } if (kring->nm_sync(kring, 0)) revents |= POLLERR; else nm_rxsync_finalize(kring); if (netmap_no_timestamp == 0 || kring->ring->flags & NR_TIMESTAMP) { microtime(&kring->ring->ts); } found = kring->rcur != kring->rtail; nm_kr_put(kring); if (found) { revents |= want_rx; retry_rx = 0; kring->nm_notify(kring, 0); } } /* transparent mode XXX only during first pass ? */ if (na->na_flags & NAF_HOST_RINGS) { kring = &na->rx_rings[na->num_rx_rings]; if (check_all_rx && (netmap_fwd || kring->ring->flags & NR_FORWARD)) { /* XXX fix to use kring fields */ if (nm_ring_empty(kring->ring)) send_down = netmap_rxsync_from_host(na, td, dev); if (!nm_ring_empty(kring->ring)) revents |= want_rx; } } if (retry_rx && !is_kevent) OS_selrecord(td, check_all_rx ? &na->si[NR_RX] : &na->rx_rings[priv->np_qfirst[NR_RX]].si); if (send_down > 0 || retry_rx) { retry_rx = 0; if (send_down) goto flush_tx; /* and retry_rx */ else goto do_retry_rx; } } /* * Transparent mode: marked bufs on rx rings between * kring->nr_hwcur and ring->head * are passed to the other endpoint. * * In this mode we also scan the sw rxring, which in * turn passes packets up. * * XXX Transparent mode at the moment requires to bind all * rings to a single file descriptor. */ if (q.head && na->ifp != NULL) netmap_send_up(na->ifp, &q); return (revents); #undef want_tx #undef want_rx } /*-------------------- driver support routines -------------------*/ static int netmap_hw_krings_create(struct netmap_adapter *); /* default notify callback */ static int netmap_notify(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; enum txrx t = kring->tx; OS_selwakeup(&kring->si, PI_NET); /* optimization: avoid a wake up on the global * queue if nobody has registered for more * than one ring */ if (na->si_users[t] > 0) OS_selwakeup(&na->si[t], PI_NET); return 0; } /* called by all routines that create netmap_adapters. * Attach na to the ifp (if any) and provide defaults * for optional callbacks. Defaults assume that we * are creating an hardware netmap_adapter. */ int netmap_attach_common(struct netmap_adapter *na) { struct ifnet *ifp = na->ifp; if (na->num_tx_rings == 0 || na->num_rx_rings == 0) { D("%s: invalid rings tx %d rx %d", na->name, na->num_tx_rings, na->num_rx_rings); return EINVAL; } /* ifp is NULL for virtual adapters (bwrap, non-persistent VALE ports, * pipes, monitors). For bwrap we actually have a non-null ifp for * use by the external modules, but that is set after this * function has been called. * XXX this is ugly, maybe split this function in two (2014-03-14) */ if (ifp != NULL) { WNA(ifp) = na; /* the following is only needed for na that use the host port. * XXX do we have something similar for linux ? */ #ifdef __FreeBSD__ na->if_input = ifp->if_input; /* for netmap_send_up */ #endif /* __FreeBSD__ */ NETMAP_SET_CAPABLE(ifp); } if (na->nm_krings_create == NULL) { /* we assume that we have been called by a driver, * since other port types all provide their own * nm_krings_create */ na->nm_krings_create = netmap_hw_krings_create; na->nm_krings_delete = netmap_hw_krings_delete; } if (na->nm_notify == NULL) na->nm_notify = netmap_notify; na->active_fds = 0; if (na->nm_mem == NULL) /* use the global allocator */ na->nm_mem = &nm_mem; netmap_mem_get(na->nm_mem); #ifdef WITH_VALE if (na->nm_bdg_attach == NULL) /* no special nm_bdg_attach callback. On VALE * attach, we need to interpose a bwrap */ na->nm_bdg_attach = netmap_bwrap_attach; #endif return 0; } /* standard cleanup, called by all destructors */ void netmap_detach_common(struct netmap_adapter *na) { if (na->ifp != NULL) WNA(na->ifp) = NULL; /* XXX do we need this? */ if (na->tx_rings) { /* XXX should not happen */ D("freeing leftover tx_rings"); na->nm_krings_delete(na); } netmap_pipe_dealloc(na); if (na->nm_mem) netmap_mem_put(na->nm_mem); bzero(na, sizeof(*na)); free(na, M_DEVBUF); } /* Wrapper for the register callback provided hardware drivers. * na->ifp == NULL means the the driver module has been * unloaded, so we cannot call into it. * Note that module unloading, in our patched linux drivers, * happens under NMG_LOCK and after having stopped all the * nic rings (see netmap_detach). This provides sufficient * protection for the other driver-provied callbacks * (i.e., nm_config and nm_*xsync), that therefore don't need * to wrapped. */ static int netmap_hw_register(struct netmap_adapter *na, int onoff) { struct netmap_hw_adapter *hwna = (struct netmap_hw_adapter*)na; if (na->ifp == NULL) return onoff ? ENXIO : 0; return hwna->nm_hw_register(na, onoff); } /* * Initialize a ``netmap_adapter`` object created by driver on attach. * We allocate a block of memory with room for a struct netmap_adapter * plus two sets of N+2 struct netmap_kring (where N is the number * of hardware rings): * krings 0..N-1 are for the hardware queues. * kring N is for the host stack queue * kring N+1 is only used for the selinfo for all queues. // XXX still true ? * Return 0 on success, ENOMEM otherwise. */ int netmap_attach(struct netmap_adapter *arg) { struct netmap_hw_adapter *hwna = NULL; // XXX when is arg == NULL ? struct ifnet *ifp = arg ? arg->ifp : NULL; if (arg == NULL || ifp == NULL) goto fail; hwna = malloc(sizeof(*hwna), M_DEVBUF, M_NOWAIT | M_ZERO); if (hwna == NULL) goto fail; hwna->up = *arg; hwna->up.na_flags |= NAF_HOST_RINGS | NAF_NATIVE; strncpy(hwna->up.name, ifp->if_xname, sizeof(hwna->up.name)); hwna->nm_hw_register = hwna->up.nm_register; hwna->up.nm_register = netmap_hw_register; if (netmap_attach_common(&hwna->up)) { free(hwna, M_DEVBUF); goto fail; } netmap_adapter_get(&hwna->up); #ifdef linux if (ifp->netdev_ops) { /* prepare a clone of the netdev ops */ #ifndef NETMAP_LINUX_HAVE_NETDEV_OPS hwna->nm_ndo.ndo_start_xmit = ifp->netdev_ops; #else hwna->nm_ndo = *ifp->netdev_ops; #endif } hwna->nm_ndo.ndo_start_xmit = linux_netmap_start_xmit; if (ifp->ethtool_ops) { hwna->nm_eto = *ifp->ethtool_ops; } hwna->nm_eto.set_ringparam = linux_netmap_set_ringparam; #ifdef NETMAP_LINUX_HAVE_SET_CHANNELS hwna->nm_eto.set_channels = linux_netmap_set_channels; #endif if (arg->nm_config == NULL) { hwna->up.nm_config = netmap_linux_config; } #endif /* linux */ if_printf(ifp, "netmap queues/slots: TX %d/%d, RX %d/%d\n", hwna->up.num_tx_rings, hwna->up.num_tx_desc, hwna->up.num_rx_rings, hwna->up.num_rx_desc); return 0; fail: D("fail, arg %p ifp %p na %p", arg, ifp, hwna); if (ifp) netmap_detach(ifp); return (hwna ? EINVAL : ENOMEM); } void NM_DBG(netmap_adapter_get)(struct netmap_adapter *na) { if (!na) { return; } refcount_acquire(&na->na_refcount); } /* returns 1 iff the netmap_adapter is destroyed */ int NM_DBG(netmap_adapter_put)(struct netmap_adapter *na) { if (!na) return 1; if (!refcount_release(&na->na_refcount)) return 0; if (na->nm_dtor) na->nm_dtor(na); netmap_detach_common(na); return 1; } /* nm_krings_create callback for all hardware native adapters */ int netmap_hw_krings_create(struct netmap_adapter *na) { int ret = netmap_krings_create(na, 0); if (ret == 0) { /* initialize the mbq for the sw rx ring */ mbq_safe_init(&na->rx_rings[na->num_rx_rings].rx_queue); ND("initialized sw rx queue %d", na->num_rx_rings); } return ret; } /* * Called on module unload by the netmap-enabled drivers */ void netmap_detach(struct ifnet *ifp) { struct netmap_adapter *na = NA(ifp); if (!na) return; NMG_LOCK(); netmap_disable_all_rings(ifp); na->ifp = NULL; na->na_flags &= ~NAF_NETMAP_ON; /* * if the netmap adapter is not native, somebody * changed it, so we can not release it here. * The NULL na->ifp will notify the new owner that * the driver is gone. */ if (na->na_flags & NAF_NATIVE) { netmap_adapter_put(na); } /* give them a chance to notice */ netmap_enable_all_rings(ifp); NMG_UNLOCK(); } /* * Intercept packets from the network stack and pass them * to netmap as incoming packets on the 'software' ring. * * We only store packets in a bounded mbq and then copy them * in the relevant rxsync routine. * * We rely on the OS to make sure that the ifp and na do not go * away (typically the caller checks for IFF_DRV_RUNNING or the like). * In nm_register() or whenever there is a reinitialization, * we make sure to make the mode change visible here. */ int netmap_transmit(struct ifnet *ifp, struct mbuf *m) { struct netmap_adapter *na = NA(ifp); struct netmap_kring *kring; u_int len = MBUF_LEN(m); u_int error = ENOBUFS; struct mbq *q; int space; kring = &na->rx_rings[na->num_rx_rings]; // XXX [Linux] we do not need this lock // if we follow the down/configure/up protocol -gl // mtx_lock(&na->core_lock); if (!nm_netmap_on(na)) { D("%s not in netmap mode anymore", na->name); error = ENXIO; goto done; } q = &kring->rx_queue; // XXX reconsider long packets if we handle fragments if (len > NETMAP_BUF_SIZE(na)) { /* too long for us */ D("%s from_host, drop packet size %d > %d", na->name, len, NETMAP_BUF_SIZE(na)); goto done; } /* protect against rxsync_from_host(), netmap_sw_to_nic() * and maybe other instances of netmap_transmit (the latter * not possible on Linux). * Also avoid overflowing the queue. */ mbq_lock(q); space = kring->nr_hwtail - kring->nr_hwcur; if (space < 0) space += kring->nkr_num_slots; if (space + mbq_len(q) >= kring->nkr_num_slots - 1) { // XXX RD(10, "%s full hwcur %d hwtail %d qlen %d len %d m %p", na->name, kring->nr_hwcur, kring->nr_hwtail, mbq_len(q), len, m); } else { mbq_enqueue(q, m); ND(10, "%s %d bufs in queue len %d m %p", na->name, mbq_len(q), len, m); /* notify outside the lock */ m = NULL; error = 0; } mbq_unlock(q); done: if (m) m_freem(m); /* unconditionally wake up listeners */ kring->nm_notify(kring, 0); /* this is normally netmap_notify(), but for nics * connected to a bridge it is netmap_bwrap_intr_notify(), * that possibly forwards the frames through the switch */ return (error); } /* * netmap_reset() is called by the driver routines when reinitializing * a ring. The driver is in charge of locking to protect the kring. * If native netmap mode is not set just return NULL. */ struct netmap_slot * netmap_reset(struct netmap_adapter *na, enum txrx tx, u_int n, u_int new_cur) { struct netmap_kring *kring; int new_hwofs, lim; if (!nm_native_on(na)) { ND("interface not in native netmap mode"); return NULL; /* nothing to reinitialize */ } /* XXX note- in the new scheme, we are not guaranteed to be * under lock (e.g. when called on a device reset). * In this case, we should set a flag and do not trust too * much the values. In practice: TODO * - set a RESET flag somewhere in the kring * - do the processing in a conservative way * - let the *sync() fixup at the end. */ if (tx == NR_TX) { if (n >= na->num_tx_rings) return NULL; kring = na->tx_rings + n; // XXX check whether we should use hwcur or rcur new_hwofs = kring->nr_hwcur - new_cur; } else { if (n >= na->num_rx_rings) return NULL; kring = na->rx_rings + n; new_hwofs = kring->nr_hwtail - new_cur; } lim = kring->nkr_num_slots - 1; if (new_hwofs > lim) new_hwofs -= lim + 1; /* Always set the new offset value and realign the ring. */ if (netmap_verbose) D("%s %s%d hwofs %d -> %d, hwtail %d -> %d", na->name, tx == NR_TX ? "TX" : "RX", n, kring->nkr_hwofs, new_hwofs, kring->nr_hwtail, tx == NR_TX ? lim : kring->nr_hwtail); kring->nkr_hwofs = new_hwofs; if (tx == NR_TX) { kring->nr_hwtail = kring->nr_hwcur + lim; if (kring->nr_hwtail > lim) kring->nr_hwtail -= lim + 1; } #if 0 // def linux /* XXX check that the mappings are correct */ /* need ring_nr, adapter->pdev, direction */ buffer_info->dma = dma_map_single(&pdev->dev, addr, adapter->rx_buffer_len, DMA_FROM_DEVICE); if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) { D("error mapping rx netmap buffer %d", i); // XXX fix error handling } #endif /* linux */ /* * Wakeup on the individual and global selwait * We do the wakeup here, but the ring is not yet reconfigured. * However, we are under lock so there are no races. */ kring->nm_notify(kring, 0); return kring->ring->slot; } /* * Dispatch rx/tx interrupts to the netmap rings. * * "work_done" is non-null on the RX path, NULL for the TX path. * We rely on the OS to make sure that there is only one active * instance per queue, and that there is appropriate locking. * * The 'notify' routine depends on what the ring is attached to. * - for a netmap file descriptor, do a selwakeup on the individual * waitqueue, plus one on the global one if needed * (see netmap_notify) * - for a nic connected to a switch, call the proper forwarding routine * (see netmap_bwrap_intr_notify) */ void netmap_common_irq(struct ifnet *ifp, u_int q, u_int *work_done) { struct netmap_adapter *na = NA(ifp); struct netmap_kring *kring; enum txrx t = (work_done ? NR_RX : NR_TX); q &= NETMAP_RING_MASK; if (netmap_verbose) { RD(5, "received %s queue %d", work_done ? "RX" : "TX" , q); } if (q >= nma_get_nrings(na, t)) return; // not a physical queue kring = NMR(na, t) + q; if (t == NR_RX) { kring->nr_kflags |= NKR_PENDINTR; // XXX atomic ? *work_done = 1; /* do not fire napi again */ } kring->nm_notify(kring, 0); } /* * Default functions to handle rx/tx interrupts from a physical device. * "work_done" is non-null on the RX path, NULL for the TX path. * * If the card is not in netmap mode, simply return 0, * so that the caller proceeds with regular processing. * Otherwise call netmap_common_irq() and return 1. * * If the card is connected to a netmap file descriptor, * do a selwakeup on the individual queue, plus one on the global one * if needed (multiqueue card _and_ there are multiqueue listeners), * and return 1. * * Finally, if called on rx from an interface connected to a switch, * calls the proper forwarding routine, and return 1. */ int netmap_rx_irq(struct ifnet *ifp, u_int q, u_int *work_done) { struct netmap_adapter *na = NA(ifp); /* * XXX emulated netmap mode sets NAF_SKIP_INTR so * we still use the regular driver even though the previous * check fails. It is unclear whether we should use * nm_native_on() here. */ if (!nm_netmap_on(na)) return 0; if (na->na_flags & NAF_SKIP_INTR) { ND("use regular interrupt"); return 0; } netmap_common_irq(ifp, q, work_done); return 1; } /* * Module loader and unloader * * netmap_init() creates the /dev/netmap device and initializes * all global variables. Returns 0 on success, errno on failure * (but there is no chance) * * netmap_fini() destroys everything. */ static struct cdev *netmap_dev; /* /dev/netmap character device. */ extern struct cdevsw netmap_cdevsw; void netmap_fini(void) { netmap_uninit_bridges(); if (netmap_dev) destroy_dev(netmap_dev); netmap_mem_fini(); NMG_LOCK_DESTROY(); printf("netmap: unloaded module.\n"); } int netmap_init(void) { int error; NMG_LOCK_INIT(); error = netmap_mem_init(); if (error != 0) goto fail; /* * MAKEDEV_ETERNAL_KLD avoids an expensive check on syscalls * when the module is compiled in. * XXX could use make_dev_credv() to get error number */ netmap_dev = make_dev_credf(MAKEDEV_ETERNAL_KLD, &netmap_cdevsw, 0, NULL, UID_ROOT, GID_WHEEL, 0600, "netmap"); if (!netmap_dev) goto fail; error = netmap_init_bridges(); if (error) goto fail; #ifdef __FreeBSD__ nm_vi_init_index(); #endif printf("netmap: loaded module\n"); return (0); fail: netmap_fini(); return (EINVAL); /* may be incorrect */ } Index: head/sys/dev/netmap/netmap_freebsd.c =================================================================== --- head/sys/dev/netmap/netmap_freebsd.c (revision 285358) +++ head/sys/dev/netmap/netmap_freebsd.c (revision 285359) @@ -1,846 +1,842 @@ /* * Copyright (C) 2013-2014 Universita` di Pisa. 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. */ /* $FreeBSD$ */ #include "opt_inet.h" #include "opt_inet6.h" #include #include #include #include /* defines used in kernel.h */ #include /* POLLIN, POLLOUT */ #include /* types used in module initialization */ #include /* DEV_MODULE */ #include #include #include /* vtophys */ #include /* vtophys */ #include #include #include #include #include #include #include /* sockaddrs */ #include #include #include #include /* IFT_ETHER */ #include /* ether_ifdetach */ #include /* LLADDR */ #include /* bus_dmamap_* */ #include /* in6_cksum_pseudo() */ #include /* in_pseudo(), in_cksum_hdr() */ #include #include #include /* ======================== FREEBSD-SPECIFIC ROUTINES ================== */ rawsum_t nm_csum_raw(uint8_t *data, size_t len, rawsum_t cur_sum) { /* TODO XXX please use the FreeBSD implementation for this. */ uint16_t *words = (uint16_t *)data; int nw = len / 2; int i; for (i = 0; i < nw; i++) cur_sum += be16toh(words[i]); if (len & 1) cur_sum += (data[len-1] << 8); return cur_sum; } /* Fold a raw checksum: 'cur_sum' is in host byte order, while the * return value is in network byte order. */ uint16_t nm_csum_fold(rawsum_t cur_sum) { /* TODO XXX please use the FreeBSD implementation for this. */ while (cur_sum >> 16) cur_sum = (cur_sum & 0xFFFF) + (cur_sum >> 16); return htobe16((~cur_sum) & 0xFFFF); } uint16_t nm_csum_ipv4(struct nm_iphdr *iph) { #if 0 return in_cksum_hdr((void *)iph); #else return nm_csum_fold(nm_csum_raw((uint8_t*)iph, sizeof(struct nm_iphdr), 0)); #endif } void nm_csum_tcpudp_ipv4(struct nm_iphdr *iph, void *data, size_t datalen, uint16_t *check) { #ifdef INET uint16_t pseudolen = datalen + iph->protocol; /* Compute and insert the pseudo-header cheksum. */ *check = in_pseudo(iph->saddr, iph->daddr, htobe16(pseudolen)); /* Compute the checksum on TCP/UDP header + payload * (includes the pseudo-header). */ *check = nm_csum_fold(nm_csum_raw(data, datalen, 0)); #else static int notsupported = 0; if (!notsupported) { notsupported = 1; D("inet4 segmentation not supported"); } #endif } void nm_csum_tcpudp_ipv6(struct nm_ipv6hdr *ip6h, void *data, size_t datalen, uint16_t *check) { #ifdef INET6 *check = in6_cksum_pseudo((void*)ip6h, datalen, ip6h->nexthdr, 0); *check = nm_csum_fold(nm_csum_raw(data, datalen, 0)); #else static int notsupported = 0; if (!notsupported) { notsupported = 1; D("inet6 segmentation not supported"); } #endif } /* * Intercept the rx routine in the standard device driver. * Second argument is non-zero to intercept, 0 to restore */ int netmap_catch_rx(struct netmap_generic_adapter *gna, int intercept) { struct netmap_adapter *na = &gna->up.up; struct ifnet *ifp = na->ifp; if (intercept) { if (gna->save_if_input) { D("cannot intercept again"); return EINVAL; /* already set */ } gna->save_if_input = ifp->if_input; ifp->if_input = generic_rx_handler; } else { if (!gna->save_if_input){ D("cannot restore"); return EINVAL; /* not saved */ } ifp->if_input = gna->save_if_input; gna->save_if_input = NULL; } return 0; } /* * Intercept the packet steering routine in the tx path, * so that we can decide which queue is used for an mbuf. * Second argument is non-zero to intercept, 0 to restore. * On freebsd we just intercept if_transmit. */ void netmap_catch_tx(struct netmap_generic_adapter *gna, int enable) { struct netmap_adapter *na = &gna->up.up; struct ifnet *ifp = netmap_generic_getifp(gna); if (enable) { na->if_transmit = ifp->if_transmit; ifp->if_transmit = netmap_transmit; } else { ifp->if_transmit = na->if_transmit; } } /* * Transmit routine used by generic_netmap_txsync(). Returns 0 on success * and non-zero on error (which may be packet drops or other errors). * addr and len identify the netmap buffer, m is the (preallocated) * mbuf to use for transmissions. * * We should add a reference to the mbuf so the m_freem() at the end * of the transmission does not consume resources. * * On FreeBSD, and on multiqueue cards, we can force the queue using * if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) * i = m->m_pkthdr.flowid % adapter->num_queues; * else * i = curcpu % adapter->num_queues; * */ int generic_xmit_frame(struct ifnet *ifp, struct mbuf *m, void *addr, u_int len, u_int ring_nr) { int ret; /* * The mbuf should be a cluster from our special pool, * so we do not need to do an m_copyback but just copy * (and eventually, just reference the netmap buffer) */ if (GET_MBUF_REFCNT(m) != 1) { D("invalid refcnt %d for %p", GET_MBUF_REFCNT(m), m); panic("in generic_xmit_frame"); } // XXX the ext_size check is unnecessary if we link the netmap buf if (m->m_ext.ext_size < len) { RD(5, "size %d < len %d", m->m_ext.ext_size, len); len = m->m_ext.ext_size; } if (0) { /* XXX seems to have negligible benefits */ m->m_ext.ext_buf = m->m_data = addr; } else { bcopy(addr, m->m_data, len); } m->m_len = m->m_pkthdr.len = len; // inc refcount. All ours, we could skip the atomic atomic_fetchadd_int(PNT_MBUF_REFCNT(m), 1); M_HASHTYPE_SET(m, M_HASHTYPE_OPAQUE); m->m_pkthdr.flowid = ring_nr; m->m_pkthdr.rcvif = ifp; /* used for tx notification */ ret = NA(ifp)->if_transmit(ifp, m); return ret; } #if __FreeBSD_version >= 1100005 struct netmap_adapter * netmap_getna(if_t ifp) { return (NA((struct ifnet *)ifp)); } #endif /* __FreeBSD_version >= 1100005 */ /* * The following two functions are empty until we have a generic * way to extract the info from the ifp */ int generic_find_num_desc(struct ifnet *ifp, unsigned int *tx, unsigned int *rx) { D("called, in tx %d rx %d", *tx, *rx); return 0; } void generic_find_num_queues(struct ifnet *ifp, u_int *txq, u_int *rxq) { D("called, in txq %d rxq %d", *txq, *rxq); *txq = netmap_generic_rings; *rxq = netmap_generic_rings; } void netmap_mitigation_init(struct nm_generic_mit *mit, int idx, struct netmap_adapter *na) { ND("called"); mit->mit_pending = 0; mit->mit_ring_idx = idx; mit->mit_na = na; } void netmap_mitigation_start(struct nm_generic_mit *mit) { ND("called"); } void netmap_mitigation_restart(struct nm_generic_mit *mit) { ND("called"); } int netmap_mitigation_active(struct nm_generic_mit *mit) { ND("called"); return 0; } void netmap_mitigation_cleanup(struct nm_generic_mit *mit) { ND("called"); } static int nm_vi_dummy(struct ifnet *ifp, u_long cmd, caddr_t addr) { return EINVAL; } static void nm_vi_start(struct ifnet *ifp) { panic("nm_vi_start() must not be called"); } /* * Index manager of persistent virtual interfaces. * It is used to decide the lowest byte of the MAC address. * We use the same algorithm with management of bridge port index. */ #define NM_VI_MAX 255 static struct { uint8_t index[NM_VI_MAX]; /* XXX just for a reasonable number */ uint8_t active; struct mtx lock; } nm_vi_indices; void nm_vi_init_index(void) { int i; for (i = 0; i < NM_VI_MAX; i++) nm_vi_indices.index[i] = i; nm_vi_indices.active = 0; mtx_init(&nm_vi_indices.lock, "nm_vi_indices_lock", NULL, MTX_DEF); } /* return -1 if no index available */ static int nm_vi_get_index(void) { int ret; mtx_lock(&nm_vi_indices.lock); ret = nm_vi_indices.active == NM_VI_MAX ? -1 : nm_vi_indices.index[nm_vi_indices.active++]; mtx_unlock(&nm_vi_indices.lock); return ret; } static void nm_vi_free_index(uint8_t val) { int i, lim; mtx_lock(&nm_vi_indices.lock); lim = nm_vi_indices.active; for (i = 0; i < lim; i++) { if (nm_vi_indices.index[i] == val) { /* swap index[lim-1] and j */ int tmp = nm_vi_indices.index[lim-1]; nm_vi_indices.index[lim-1] = val; nm_vi_indices.index[i] = tmp; nm_vi_indices.active--; break; } } if (lim == nm_vi_indices.active) D("funny, index %u didn't found", val); mtx_unlock(&nm_vi_indices.lock); } #undef NM_VI_MAX /* * Implementation of a netmap-capable virtual interface that * registered to the system. * It is based on if_tap.c and ip_fw_log.c in FreeBSD 9. * * Note: Linux sets refcount to 0 on allocation of net_device, * then increments it on registration to the system. * FreeBSD sets refcount to 1 on if_alloc(), and does not * increment this refcount on if_attach(). */ int nm_vi_persist(const char *name, struct ifnet **ret) { struct ifnet *ifp; u_short macaddr_hi; uint32_t macaddr_mid; u_char eaddr[6]; int unit = nm_vi_get_index(); /* just to decide MAC address */ if (unit < 0) return EBUSY; /* * We use the same MAC address generation method with tap * except for the highest octet is 00:be instead of 00:bd */ macaddr_hi = htons(0x00be); /* XXX tap + 1 */ macaddr_mid = (uint32_t) ticks; bcopy(&macaddr_hi, eaddr, sizeof(short)); bcopy(&macaddr_mid, &eaddr[2], sizeof(uint32_t)); eaddr[5] = (uint8_t)unit; ifp = if_alloc(IFT_ETHER); if (ifp == NULL) { D("if_alloc failed"); return ENOMEM; } if_initname(ifp, name, IF_DUNIT_NONE); ifp->if_mtu = 65536; ifp->if_flags = IFF_UP | IFF_SIMPLEX | IFF_MULTICAST; ifp->if_init = (void *)nm_vi_dummy; ifp->if_ioctl = nm_vi_dummy; ifp->if_start = nm_vi_start; ifp->if_mtu = ETHERMTU; IFQ_SET_MAXLEN(&ifp->if_snd, ifqmaxlen); ifp->if_capabilities |= IFCAP_LINKSTATE; ifp->if_capenable |= IFCAP_LINKSTATE; ether_ifattach(ifp, eaddr); *ret = ifp; return 0; } /* unregister from the system and drop the final refcount */ void nm_vi_detach(struct ifnet *ifp) { nm_vi_free_index(((char *)IF_LLADDR(ifp))[5]); ether_ifdetach(ifp); if_free(ifp); } /* * In order to track whether pages are still mapped, we hook into * the standard cdev_pager and intercept the constructor and * destructor. */ struct netmap_vm_handle_t { struct cdev *dev; struct netmap_priv_d *priv; }; static int netmap_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot, vm_ooffset_t foff, struct ucred *cred, u_short *color) { struct netmap_vm_handle_t *vmh = handle; if (netmap_verbose) D("handle %p size %jd prot %d foff %jd", handle, (intmax_t)size, prot, (intmax_t)foff); if (color) *color = 0; dev_ref(vmh->dev); return 0; } static void netmap_dev_pager_dtor(void *handle) { struct netmap_vm_handle_t *vmh = handle; struct cdev *dev = vmh->dev; struct netmap_priv_d *priv = vmh->priv; if (netmap_verbose) D("handle %p", handle); netmap_dtor(priv); free(vmh, M_DEVBUF); dev_rel(dev); } static int netmap_dev_pager_fault(vm_object_t object, vm_ooffset_t offset, int prot, vm_page_t *mres) { struct netmap_vm_handle_t *vmh = object->handle; struct netmap_priv_d *priv = vmh->priv; struct netmap_adapter *na = priv->np_na; vm_paddr_t paddr; vm_page_t page; vm_memattr_t memattr; vm_pindex_t pidx; ND("object %p offset %jd prot %d mres %p", object, (intmax_t)offset, prot, mres); memattr = object->memattr; pidx = OFF_TO_IDX(offset); paddr = netmap_mem_ofstophys(na->nm_mem, offset); if (paddr == 0) return VM_PAGER_FAIL; if (((*mres)->flags & PG_FICTITIOUS) != 0) { /* * If the passed in result page is a fake page, update it with * the new physical address. */ page = *mres; vm_page_updatefake(page, paddr, memattr); } else { /* * Replace the passed in reqpage page with our own fake page and * free up the all of the original pages. */ #ifndef VM_OBJECT_WUNLOCK /* FreeBSD < 10.x */ #define VM_OBJECT_WUNLOCK VM_OBJECT_UNLOCK #define VM_OBJECT_WLOCK VM_OBJECT_LOCK #endif /* VM_OBJECT_WUNLOCK */ VM_OBJECT_WUNLOCK(object); page = vm_page_getfake(paddr, memattr); VM_OBJECT_WLOCK(object); vm_page_lock(*mres); vm_page_free(*mres); vm_page_unlock(*mres); *mres = page; vm_page_insert(page, object, pidx); } page->valid = VM_PAGE_BITS_ALL; return (VM_PAGER_OK); } static struct cdev_pager_ops netmap_cdev_pager_ops = { .cdev_pg_ctor = netmap_dev_pager_ctor, .cdev_pg_dtor = netmap_dev_pager_dtor, .cdev_pg_fault = netmap_dev_pager_fault, }; static int netmap_mmap_single(struct cdev *cdev, vm_ooffset_t *foff, vm_size_t objsize, vm_object_t *objp, int prot) { int error; struct netmap_vm_handle_t *vmh; struct netmap_priv_d *priv; vm_object_t obj; if (netmap_verbose) D("cdev %p foff %jd size %jd objp %p prot %d", cdev, (intmax_t )*foff, (intmax_t )objsize, objp, prot); vmh = malloc(sizeof(struct netmap_vm_handle_t), M_DEVBUF, M_NOWAIT | M_ZERO); if (vmh == NULL) return ENOMEM; vmh->dev = cdev; NMG_LOCK(); error = devfs_get_cdevpriv((void**)&priv); if (error) goto err_unlock; if (priv->np_nifp == NULL) { error = EINVAL; goto err_unlock; } vmh->priv = priv; - priv->np_refcount++; + priv->np_refs++; NMG_UNLOCK(); obj = cdev_pager_allocate(vmh, OBJT_DEVICE, &netmap_cdev_pager_ops, objsize, prot, *foff, NULL); if (obj == NULL) { D("cdev_pager_allocate failed"); error = EINVAL; goto err_deref; } *objp = obj; return 0; err_deref: NMG_LOCK(); - priv->np_refcount--; + priv->np_refs--; err_unlock: NMG_UNLOCK(); // err: free(vmh, M_DEVBUF); return error; } /* - * netmap_close() is called on every close(), but we do not need to do - * anything at that moment, since the process may have other open file - * descriptors for /dev/netmap. Instead, we pass netmap_dtor() to + * On FreeBSD the close routine is only called on the last close on + * the device (/dev/netmap) so we cannot do anything useful. + * To track close() on individual file descriptors we pass netmap_dtor() to * devfs_set_cdevpriv() on open(). The FreeBSD kernel will call the destructor * when the last fd pointing to the device is closed. * - * Unfortunately, FreeBSD does not automatically track active mmap()s on an fd, - * so we have to track them by ourselvesi (see above). The result is that + * Note that FreeBSD does not even munmap() on close() so we also have + * to track mmap() ourselves, and postpone the call to * netmap_dtor() is called when the process has no open fds and no active * memory maps on /dev/netmap, as in linux. */ static int netmap_close(struct cdev *dev, int fflag, int devtype, struct thread *td) { if (netmap_verbose) D("dev %p fflag 0x%x devtype %d td %p", dev, fflag, devtype, td); return 0; } static int netmap_open(struct cdev *dev, int oflags, int devtype, struct thread *td) { struct netmap_priv_d *priv; int error; (void)dev; (void)oflags; (void)devtype; (void)td; - // XXX wait or nowait ? priv = malloc(sizeof(struct netmap_priv_d), M_DEVBUF, M_NOWAIT | M_ZERO); if (priv == NULL) return ENOMEM; - error = devfs_set_cdevpriv(priv, netmap_dtor); - if (error) - return error; - - priv->np_refcount = 1; - - return 0; + if (error) { + free(priv, M_DEVBUF); + } + return error; } /******************** kqueue support ****************/ /* * The OS_selwakeup also needs to issue a KNOTE_UNLOCKED. * We use a non-zero argument to distinguish the call from the one * in kevent_scan() which instead also needs to run netmap_poll(). * The knote uses a global mutex for the time being. We might * try to reuse the one in the si, but it is not allocated * permanently so it might be a bit tricky. * * The *kqfilter function registers one or another f_event * depending on read or write mode. * In the call to f_event() td_fpop is NULL so any child function * calling devfs_get_cdevpriv() would fail - and we need it in * netmap_poll(). As a workaround we store priv into kn->kn_hook * and pass it as first argument to netmap_poll(), which then * uses the failure to tell that we are called from f_event() * and do not need the selrecord(). */ void freebsd_selwakeup(struct nm_selinfo *si, int pri) { if (netmap_verbose) D("on knote %p", &si->si.si_note); selwakeuppri(&si->si, pri); /* use a non-zero hint to tell the notification from the * call done in kqueue_scan() which uses 0 */ KNOTE_UNLOCKED(&si->si.si_note, 0x100 /* notification */); } static void netmap_knrdetach(struct knote *kn) { struct netmap_priv_d *priv = (struct netmap_priv_d *)kn->kn_hook; struct selinfo *si = &priv->np_si[NR_RX]->si; D("remove selinfo %p", si); knlist_remove(&si->si_note, kn, 0); } static void netmap_knwdetach(struct knote *kn) { struct netmap_priv_d *priv = (struct netmap_priv_d *)kn->kn_hook; struct selinfo *si = &priv->np_si[NR_TX]->si; D("remove selinfo %p", si); knlist_remove(&si->si_note, kn, 0); } /* * callback from notifies (generated externally) and our * calls to kevent(). The former we just return 1 (ready) * since we do not know better. * In the latter we call netmap_poll and return 0/1 accordingly. */ static int netmap_knrw(struct knote *kn, long hint, int events) { struct netmap_priv_d *priv; int revents; if (hint != 0) { ND(5, "call from notify"); return 1; /* assume we are ready */ } priv = kn->kn_hook; /* the notification may come from an external thread, * in which case we do not want to run the netmap_poll * This should be filtered above, but check just in case. */ if (curthread != priv->np_td) { /* should not happen */ RD(5, "curthread changed %p %p", curthread, priv->np_td); return 1; } else { revents = netmap_poll((void *)priv, events, curthread); return (events & revents) ? 1 : 0; } } static int netmap_knread(struct knote *kn, long hint) { return netmap_knrw(kn, hint, POLLIN); } static int netmap_knwrite(struct knote *kn, long hint) { return netmap_knrw(kn, hint, POLLOUT); } static struct filterops netmap_rfiltops = { .f_isfd = 1, .f_detach = netmap_knrdetach, .f_event = netmap_knread, }; static struct filterops netmap_wfiltops = { .f_isfd = 1, .f_detach = netmap_knwdetach, .f_event = netmap_knwrite, }; /* * This is called when a thread invokes kevent() to record * a change in the configuration of the kqueue(). * The 'priv' should be the same as in the netmap device. */ static int netmap_kqfilter(struct cdev *dev, struct knote *kn) { struct netmap_priv_d *priv; int error; struct netmap_adapter *na; struct nm_selinfo *si; int ev = kn->kn_filter; if (ev != EVFILT_READ && ev != EVFILT_WRITE) { D("bad filter request %d", ev); return 1; } error = devfs_get_cdevpriv((void**)&priv); if (error) { D("device not yet setup"); return 1; } na = priv->np_na; if (na == NULL) { D("no netmap adapter for this file descriptor"); return 1; } /* the si is indicated in the priv */ si = priv->np_si[(ev == EVFILT_WRITE) ? NR_TX : NR_RX]; // XXX lock(priv) ? kn->kn_fop = (ev == EVFILT_WRITE) ? &netmap_wfiltops : &netmap_rfiltops; kn->kn_hook = priv; knlist_add(&si->si.si_note, kn, 1); // XXX unlock(priv) ND("register %p %s td %p priv %p kn %p np_nifp %p kn_fp/fpop %s", na, na->ifp->if_xname, curthread, priv, kn, priv->np_nifp, kn->kn_fp == curthread->td_fpop ? "match" : "MISMATCH"); return 0; } struct cdevsw netmap_cdevsw = { .d_version = D_VERSION, .d_name = "netmap", .d_open = netmap_open, .d_mmap_single = netmap_mmap_single, .d_ioctl = netmap_ioctl, .d_poll = netmap_poll, .d_kqfilter = netmap_kqfilter, .d_close = netmap_close, }; /*--- end of kqueue support ----*/ /* * Kernel entry point. * * Initialize/finalize the module and return. * * Return 0 on success, errno on failure. */ static int netmap_loader(__unused struct module *module, int event, __unused void *arg) { int error = 0; switch (event) { case MOD_LOAD: error = netmap_init(); break; case MOD_UNLOAD: netmap_fini(); break; default: error = EOPNOTSUPP; break; } return (error); } DEV_MODULE(netmap, netmap_loader, NULL); MODULE_VERSION(netmap, 1); Index: head/sys/dev/netmap/netmap_kern.h =================================================================== --- head/sys/dev/netmap/netmap_kern.h (revision 285358) +++ head/sys/dev/netmap/netmap_kern.h (revision 285359) @@ -1,1699 +1,1674 @@ /* * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. All rights reserved. * Copyright (C) 2013-2014 Universita` di Pisa. 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. */ /* * $FreeBSD$ * * The header contains the definitions of constants and function * prototypes used only in kernelspace. */ #ifndef _NET_NETMAP_KERN_H_ #define _NET_NETMAP_KERN_H_ #if defined(linux) #if defined(CONFIG_NETMAP_VALE) #define WITH_VALE #endif #if defined(CONFIG_NETMAP_PIPE) #define WITH_PIPES #endif #if defined(CONFIG_NETMAP_MONITOR) #define WITH_MONITOR #endif #if defined(CONFIG_NETMAP_GENERIC) #define WITH_GENERIC #endif #if defined(CONFIG_NETMAP_V1000) #define WITH_V1000 #endif #else /* not linux */ #define WITH_VALE // comment out to disable VALE support #define WITH_PIPES #define WITH_MONITOR #define WITH_GENERIC #endif #if defined(__FreeBSD__) #define likely(x) __builtin_expect((long)!!(x), 1L) #define unlikely(x) __builtin_expect((long)!!(x), 0L) #define NM_LOCK_T struct mtx /* low level spinlock, used to protect queues */ #define NM_MTX_T struct sx /* OS-specific mutex (sleepable) */ #define NM_MTX_INIT(m) sx_init(&(m), #m) #define NM_MTX_DESTROY(m) sx_destroy(&(m)) #define NM_MTX_LOCK(m) sx_xlock(&(m)) #define NM_MTX_UNLOCK(m) sx_xunlock(&(m)) #define NM_MTX_ASSERT(m) sx_assert(&(m), SA_XLOCKED) #define NM_SELINFO_T struct nm_selinfo #define MBUF_LEN(m) ((m)->m_pkthdr.len) #define MBUF_IFP(m) ((m)->m_pkthdr.rcvif) #define NM_SEND_UP(ifp, m) ((NA(ifp))->if_input)(ifp, m) #define NM_ATOMIC_T volatile int // XXX ? /* atomic operations */ #include #define NM_ATOMIC_TEST_AND_SET(p) (!atomic_cmpset_acq_int((p), 0, 1)) #define NM_ATOMIC_CLEAR(p) atomic_store_rel_int((p), 0) #if __FreeBSD_version >= 1100030 #define WNA(_ifp) (_ifp)->if_netmap #else /* older FreeBSD */ #define WNA(_ifp) (_ifp)->if_pspare[0] #endif /* older FreeBSD */ #if __FreeBSD_version >= 1100005 struct netmap_adapter *netmap_getna(if_t ifp); #endif #if __FreeBSD_version >= 1100027 #define GET_MBUF_REFCNT(m) ((m)->m_ext.ext_cnt ? *((m)->m_ext.ext_cnt) : -1) #define SET_MBUF_REFCNT(m, x) *((m)->m_ext.ext_cnt) = x #define PNT_MBUF_REFCNT(m) ((m)->m_ext.ext_cnt) #else #define GET_MBUF_REFCNT(m) ((m)->m_ext.ref_cnt ? *((m)->m_ext.ref_cnt) : -1) #define SET_MBUF_REFCNT(m, x) *((m)->m_ext.ref_cnt) = x #define PNT_MBUF_REFCNT(m) ((m)->m_ext.ref_cnt) #endif MALLOC_DECLARE(M_NETMAP); struct nm_selinfo { struct selinfo si; struct mtx m; }; void freebsd_selwakeup(struct nm_selinfo *si, int pri); // XXX linux struct, not used in FreeBSD struct net_device_ops { }; struct ethtool_ops { }; struct hrtimer { }; #define NM_BNS_GET(b) #define NM_BNS_PUT(b) #elif defined (linux) #define NM_LOCK_T safe_spinlock_t // see bsd_glue.h #define NM_SELINFO_T wait_queue_head_t #define MBUF_LEN(m) ((m)->len) #define MBUF_IFP(m) ((m)->dev) #define NM_SEND_UP(ifp, m) \ do { \ m->priority = NM_MAGIC_PRIORITY_RX; \ netif_rx(m); \ } while (0) #define NM_ATOMIC_T volatile long unsigned int #define NM_MTX_T struct mutex /* OS-specific sleepable lock */ #define NM_MTX_INIT(m) mutex_init(&(m)) #define NM_MTX_DESTROY(m) do { (void)(m); } while (0) #define NM_MTX_LOCK(m) mutex_lock(&(m)) #define NM_MTX_UNLOCK(m) mutex_unlock(&(m)) #define NM_MTX_ASSERT(m) mutex_is_locked(&(m)) #ifndef DEV_NETMAP #define DEV_NETMAP #endif /* DEV_NETMAP */ #elif defined (__APPLE__) #warning apple support is incomplete. #define likely(x) __builtin_expect(!!(x), 1) #define unlikely(x) __builtin_expect(!!(x), 0) #define NM_LOCK_T IOLock * #define NM_SELINFO_T struct selinfo #define MBUF_LEN(m) ((m)->m_pkthdr.len) #define NM_SEND_UP(ifp, m) ((ifp)->if_input)(ifp, m) #else #error unsupported platform #endif /* end - platform-specific code */ #define NMG_LOCK_T NM_MTX_T #define NMG_LOCK_INIT() NM_MTX_INIT(netmap_global_lock) #define NMG_LOCK_DESTROY() NM_MTX_DESTROY(netmap_global_lock) #define NMG_LOCK() NM_MTX_LOCK(netmap_global_lock) #define NMG_UNLOCK() NM_MTX_UNLOCK(netmap_global_lock) #define NMG_LOCK_ASSERT() NM_MTX_ASSERT(netmap_global_lock) #define ND(format, ...) #define D(format, ...) \ do { \ struct timeval __xxts; \ microtime(&__xxts); \ printf("%03d.%06d [%4d] %-25s " format "\n", \ (int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec, \ __LINE__, __FUNCTION__, ##__VA_ARGS__); \ } while (0) /* rate limited, lps indicates how many per second */ #define RD(lps, format, ...) \ do { \ static int t0, __cnt; \ if (t0 != time_second) { \ t0 = time_second; \ __cnt = 0; \ } \ if (__cnt++ < lps) \ D(format, ##__VA_ARGS__); \ } while (0) struct netmap_adapter; struct nm_bdg_fwd; struct nm_bridge; struct netmap_priv_d; const char *nm_dump_buf(char *p, int len, int lim, char *dst); #include "netmap_mbq.h" extern NMG_LOCK_T netmap_global_lock; enum txrx { NR_RX = 0, NR_TX = 1, NR_TXRX }; static __inline const char* nm_txrx2str(enum txrx t) { return (t== NR_RX ? "RX" : "TX"); } static __inline enum txrx nm_txrx_swap(enum txrx t) { return (t== NR_RX ? NR_TX : NR_RX); } #define for_rx_tx(t) for ((t) = 0; (t) < NR_TXRX; (t)++) /* * private, kernel view of a ring. Keeps track of the status of * a ring across system calls. * * nr_hwcur index of the next buffer to refill. * It corresponds to ring->head * at the time the system call returns. * * nr_hwtail index of the first buffer owned by the kernel. * On RX, hwcur->hwtail are receive buffers * not yet released. hwcur is advanced following * ring->head, hwtail is advanced on incoming packets, * and a wakeup is generated when hwtail passes ring->cur * On TX, hwcur->rcur have been filled by the sender * but not sent yet to the NIC; rcur->hwtail are available * for new transmissions, and hwtail->hwcur-1 are pending * transmissions not yet acknowledged. * * The indexes in the NIC and netmap rings are offset by nkr_hwofs slots. * This is so that, on a reset, buffers owned by userspace are not * modified by the kernel. In particular: * RX rings: the next empty buffer (hwtail + hwofs) coincides with * the next empty buffer as known by the hardware (next_to_check or so). * TX rings: hwcur + hwofs coincides with next_to_send * * For received packets, slot->flags is set to nkr_slot_flags * so we can provide a proper initial value (e.g. set NS_FORWARD * when operating in 'transparent' mode). * * The following fields are used to implement lock-free copy of packets * from input to output ports in VALE switch: * nkr_hwlease buffer after the last one being copied. * A writer in nm_bdg_flush reserves N buffers * from nr_hwlease, advances it, then does the * copy outside the lock. * In RX rings (used for VALE ports), * nkr_hwtail <= nkr_hwlease < nkr_hwcur+N-1 * In TX rings (used for NIC or host stack ports) * nkr_hwcur <= nkr_hwlease < nkr_hwtail * nkr_leases array of nkr_num_slots where writers can report * completion of their block. NR_NOSLOT (~0) indicates * that the writer has not finished yet * nkr_lease_idx index of next free slot in nr_leases, to be assigned * * The kring is manipulated by txsync/rxsync and generic netmap function. * * Concurrent rxsync or txsync on the same ring are prevented through * by nm_kr_(try)lock() which in turn uses nr_busy. This is all we need * for NIC rings, and for TX rings attached to the host stack. * * RX rings attached to the host stack use an mbq (rx_queue) on both * rxsync_from_host() and netmap_transmit(). The mbq is protected * by its internal lock. * * RX rings attached to the VALE switch are accessed by both senders * and receiver. They are protected through the q_lock on the RX ring. */ struct netmap_kring { struct netmap_ring *ring; uint32_t nr_hwcur; uint32_t nr_hwtail; /* * Copies of values in user rings, so we do not need to look * at the ring (which could be modified). These are set in the * *sync_prologue()/finalize() routines. */ uint32_t rhead; uint32_t rcur; uint32_t rtail; uint32_t nr_kflags; /* private driver flags */ #define NKR_PENDINTR 0x1 // Pending interrupt. #define NKR_EXCLUSIVE 0x2 /* exclusive binding */ uint32_t nkr_num_slots; /* * On a NIC reset, the NIC ring indexes may be reset but the * indexes in the netmap rings remain the same. nkr_hwofs * keeps track of the offset between the two. */ int32_t nkr_hwofs; uint16_t nkr_slot_flags; /* initial value for flags */ /* last_reclaim is opaque marker to help reduce the frequency * of operations such as reclaiming tx buffers. A possible use * is set it to ticks and do the reclaim only once per tick. */ uint64_t last_reclaim; NM_SELINFO_T si; /* poll/select wait queue */ NM_LOCK_T q_lock; /* protects kring and ring. */ NM_ATOMIC_T nr_busy; /* prevent concurrent syscalls */ struct netmap_adapter *na; /* The following fields are for VALE switch support */ struct nm_bdg_fwd *nkr_ft; uint32_t *nkr_leases; #define NR_NOSLOT ((uint32_t)~0) /* used in nkr_*lease* */ uint32_t nkr_hwlease; uint32_t nkr_lease_idx; /* while nkr_stopped is set, no new [tr]xsync operations can * be started on this kring. * This is used by netmap_disable_all_rings() * to find a synchronization point where critical data * structures pointed to by the kring can be added or removed */ volatile int nkr_stopped; /* Support for adapters without native netmap support. * On tx rings we preallocate an array of tx buffers * (same size as the netmap ring), on rx rings we * store incoming mbufs in a queue that is drained by * a rxsync. */ struct mbuf **tx_pool; // u_int nr_ntc; /* Emulation of a next-to-clean RX ring pointer. */ struct mbq rx_queue; /* intercepted rx mbufs. */ uint32_t users; /* existing bindings for this ring */ uint32_t ring_id; /* debugging */ enum txrx tx; /* kind of ring (tx or rx) */ char name[64]; /* diagnostic */ /* [tx]sync callback for this kring. * The default nm_kring_create callback (netmap_krings_create) * sets the nm_sync callback of each hardware tx(rx) kring to * the corresponding nm_txsync(nm_rxsync) taken from the * netmap_adapter; moreover, it sets the sync callback * of the host tx(rx) ring to netmap_txsync_to_host * (netmap_rxsync_from_host). * * Overrides: the above configuration is not changed by * any of the nm_krings_create callbacks. */ int (*nm_sync)(struct netmap_kring *kring, int flags); int (*nm_notify)(struct netmap_kring *kring, int flags); #ifdef WITH_PIPES struct netmap_kring *pipe; /* if this is a pipe ring, * pointer to the other end */ struct netmap_ring *save_ring; /* pointer to hidden rings * (see netmap_pipe.c for details) */ #endif /* WITH_PIPES */ #ifdef WITH_VALE int (*save_notify)(struct netmap_kring *kring, int flags); #endif #ifdef WITH_MONITOR /* array of krings that are monitoring this kring */ struct netmap_kring **monitors; uint32_t max_monitors; /* current size of the monitors array */ uint32_t n_monitors; /* next unused entry in the monitor array */ /* * Monitors work by intercepting the sync and notify callbacks of the * monitored krings. This is implemented by replacing the pointers * above and saving the previous ones in mon_* pointers below */ int (*mon_sync)(struct netmap_kring *kring, int flags); int (*mon_notify)(struct netmap_kring *kring, int flags); uint32_t mon_tail; /* last seen slot on rx */ uint32_t mon_pos; /* index of this ring in the monitored ring array */ #endif } __attribute__((__aligned__(64))); /* return the next index, with wraparound */ static inline uint32_t nm_next(uint32_t i, uint32_t lim) { return unlikely (i == lim) ? 0 : i + 1; } /* return the previous index, with wraparound */ static inline uint32_t nm_prev(uint32_t i, uint32_t lim) { return unlikely (i == 0) ? lim : i - 1; } /* * * Here is the layout for the Rx and Tx rings. RxRING TxRING +-----------------+ +-----------------+ | | | | |XXX free slot XXX| |XXX free slot XXX| +-----------------+ +-----------------+ head->| owned by user |<-hwcur | not sent to nic |<-hwcur | | | yet | +-----------------+ | | cur->| available to | | | | user, not read | +-----------------+ | yet | cur->| (being | | | | prepared) | | | | | +-----------------+ + ------ + tail->| |<-hwtail | |<-hwlease | (being | ... | | ... | prepared) | ... | | ... +-----------------+ ... | | ... | |<-hwlease +-----------------+ | | tail->| |<-hwtail | | | | | | | | | | | | +-----------------+ +-----------------+ * The cur/tail (user view) and hwcur/hwtail (kernel view) * are used in the normal operation of the card. * * When a ring is the output of a switch port (Rx ring for * a VALE port, Tx ring for the host stack or NIC), slots * are reserved in blocks through 'hwlease' which points * to the next unused slot. * On an Rx ring, hwlease is always after hwtail, * and completions cause hwtail to advance. * On a Tx ring, hwlease is always between cur and hwtail, * and completions cause cur to advance. * * nm_kr_space() returns the maximum number of slots that * can be assigned. * nm_kr_lease() reserves the required number of buffers, * advances nkr_hwlease and also returns an entry in * a circular array where completions should be reported. */ struct netmap_lut { struct lut_entry *lut; uint32_t objtotal; /* max buffer index */ uint32_t objsize; /* buffer size */ }; struct netmap_vp_adapter; // forward /* * The "struct netmap_adapter" extends the "struct adapter" * (or equivalent) device descriptor. * It contains all base fields needed to support netmap operation. * There are in fact different types of netmap adapters * (native, generic, VALE switch...) so a netmap_adapter is * just the first field in the derived type. */ struct netmap_adapter { /* * On linux we do not have a good way to tell if an interface * is netmap-capable. So we always use the following trick: * NA(ifp) points here, and the first entry (which hopefully * always exists and is at least 32 bits) contains a magic * value which we can use to detect that the interface is good. */ uint32_t magic; uint32_t na_flags; /* enabled, and other flags */ #define NAF_SKIP_INTR 1 /* use the regular interrupt handler. * useful during initialization */ #define NAF_SW_ONLY 2 /* forward packets only to sw adapter */ #define NAF_BDG_MAYSLEEP 4 /* the bridge is allowed to sleep when * forwarding packets coming from this * interface */ #define NAF_MEM_OWNER 8 /* the adapter uses its own memory area * that cannot be changed */ #define NAF_NATIVE 16 /* the adapter is native. * Virtual ports (vale, pipe, monitor...) * should never use this flag. */ #define NAF_NETMAP_ON 32 /* netmap is active (either native or * emulated). Where possible (e.g. FreeBSD) * IFCAP_NETMAP also mirrors this flag. */ #define NAF_HOST_RINGS 64 /* the adapter supports the host rings */ #define NAF_FORCE_NATIVE 128 /* the adapter is always NATIVE */ #define NAF_BUSY (1U<<31) /* the adapter is used internally and * cannot be registered from userspace */ int active_fds; /* number of user-space descriptors using this interface, which is equal to the number of struct netmap_if objs in the mapped region. */ u_int num_rx_rings; /* number of adapter receive rings */ u_int num_tx_rings; /* number of adapter transmit rings */ u_int num_tx_desc; /* number of descriptor in each queue */ u_int num_rx_desc; /* tx_rings and rx_rings are private but allocated * as a contiguous chunk of memory. Each array has * N+1 entries, for the adapter queues and for the host queue. */ struct netmap_kring *tx_rings; /* array of TX rings. */ struct netmap_kring *rx_rings; /* array of RX rings. */ void *tailroom; /* space below the rings array */ /* (used for leases) */ NM_SELINFO_T si[NR_TXRX]; /* global wait queues */ /* count users of the global wait queues */ int si_users[NR_TXRX]; void *pdev; /* used to store pci device */ /* copy of if_qflush and if_transmit pointers, to intercept * packets from the network stack when netmap is active. */ int (*if_transmit)(struct ifnet *, struct mbuf *); /* copy of if_input for netmap_send_up() */ void (*if_input)(struct ifnet *, struct mbuf *); /* references to the ifnet and device routines, used by * the generic netmap functions. */ struct ifnet *ifp; /* adapter is ifp->if_softc */ /*---- callbacks for this netmap adapter -----*/ /* * nm_dtor() is the cleanup routine called when destroying * the adapter. * Called with NMG_LOCK held. * * nm_register() is called on NIOCREGIF and close() to enter * or exit netmap mode on the NIC * Called with NNG_LOCK held. * * nm_txsync() pushes packets to the underlying hw/switch * * nm_rxsync() collects packets from the underlying hw/switch * * nm_config() returns configuration information from the OS * Called with NMG_LOCK held. * * nm_krings_create() create and init the tx_rings and * rx_rings arrays of kring structures. In particular, * set the nm_sync callbacks for each ring. * There is no need to also allocate the corresponding * netmap_rings, since netmap_mem_rings_create() will always * be called to provide the missing ones. * Called with NNG_LOCK held. * * nm_krings_delete() cleanup and delete the tx_rings and rx_rings * arrays * Called with NMG_LOCK held. * * nm_notify() is used to act after data have become available * (or the stopped state of the ring has changed) * For hw devices this is typically a selwakeup(), * but for NIC/host ports attached to a switch (or vice-versa) * we also need to invoke the 'txsync' code downstream. */ void (*nm_dtor)(struct netmap_adapter *); int (*nm_register)(struct netmap_adapter *, int onoff); int (*nm_txsync)(struct netmap_kring *kring, int flags); int (*nm_rxsync)(struct netmap_kring *kring, int flags); int (*nm_notify)(struct netmap_kring *kring, int flags); #define NAF_FORCE_READ 1 #define NAF_FORCE_RECLAIM 2 /* return configuration information */ int (*nm_config)(struct netmap_adapter *, u_int *txr, u_int *txd, u_int *rxr, u_int *rxd); int (*nm_krings_create)(struct netmap_adapter *); void (*nm_krings_delete)(struct netmap_adapter *); #ifdef WITH_VALE /* * nm_bdg_attach() initializes the na_vp field to point * to an adapter that can be attached to a VALE switch. If the * current adapter is already a VALE port, na_vp is simply a cast; * otherwise, na_vp points to a netmap_bwrap_adapter. * If applicable, this callback also initializes na_hostvp, * that can be used to connect the adapter host rings to the * switch. * Called with NMG_LOCK held. * * nm_bdg_ctl() is called on the actual attach/detach to/from * to/from the switch, to perform adapter-specific * initializations * Called with NMG_LOCK held. */ int (*nm_bdg_attach)(const char *bdg_name, struct netmap_adapter *); int (*nm_bdg_ctl)(struct netmap_adapter *, struct nmreq *, int); /* adapter used to attach this adapter to a VALE switch (if any) */ struct netmap_vp_adapter *na_vp; /* adapter used to attach the host rings of this adapter * to a VALE switch (if any) */ struct netmap_vp_adapter *na_hostvp; #endif /* standard refcount to control the lifetime of the adapter * (it should be equal to the lifetime of the corresponding ifp) */ int na_refcount; /* memory allocator (opaque) * We also cache a pointer to the lut_entry for translating * buffer addresses, and the total number of buffers. */ struct netmap_mem_d *nm_mem; struct netmap_lut na_lut; /* additional information attached to this adapter * by other netmap subsystems. Currently used by * bwrap and LINUX/v1000. */ void *na_private; /* array of pipes that have this adapter as a parent */ struct netmap_pipe_adapter **na_pipes; int na_next_pipe; /* next free slot in the array */ int na_max_pipes; /* size of the array */ char name[64]; }; static __inline u_int nma_get_ndesc(struct netmap_adapter *na, enum txrx t) { return (t == NR_TX ? na->num_tx_desc : na->num_rx_desc); } static __inline void nma_set_ndesc(struct netmap_adapter *na, enum txrx t, u_int v) { if (t == NR_TX) na->num_tx_desc = v; else na->num_rx_desc = v; } static __inline u_int nma_get_nrings(struct netmap_adapter *na, enum txrx t) { return (t == NR_TX ? na->num_tx_rings : na->num_rx_rings); } static __inline void nma_set_nrings(struct netmap_adapter *na, enum txrx t, u_int v) { if (t == NR_TX) na->num_tx_rings = v; else na->num_rx_rings = v; } static __inline struct netmap_kring* NMR(struct netmap_adapter *na, enum txrx t) { return (t == NR_TX ? na->tx_rings : na->rx_rings); } /* * If the NIC is owned by the kernel * (i.e., bridge), neither another bridge nor user can use it; * if the NIC is owned by a user, only users can share it. * Evaluation must be done under NMG_LOCK(). */ #define NETMAP_OWNED_BY_KERN(na) ((na)->na_flags & NAF_BUSY) #define NETMAP_OWNED_BY_ANY(na) \ (NETMAP_OWNED_BY_KERN(na) || ((na)->active_fds > 0)) /* * derived netmap adapters for various types of ports */ struct netmap_vp_adapter { /* VALE software port */ struct netmap_adapter up; /* * Bridge support: * * bdg_port is the port number used in the bridge; * na_bdg points to the bridge this NA is attached to. */ int bdg_port; struct nm_bridge *na_bdg; int retry; /* Offset of ethernet header for each packet. */ u_int virt_hdr_len; /* Maximum Frame Size, used in bdg_mismatch_datapath() */ u_int mfs; /* Last source MAC on this port */ uint64_t last_smac; }; struct netmap_hw_adapter { /* physical device */ struct netmap_adapter up; struct net_device_ops nm_ndo; // XXX linux only struct ethtool_ops nm_eto; // XXX linux only const struct ethtool_ops* save_ethtool; int (*nm_hw_register)(struct netmap_adapter *, int onoff); }; #ifdef WITH_GENERIC /* Mitigation support. */ struct nm_generic_mit { struct hrtimer mit_timer; int mit_pending; int mit_ring_idx; /* index of the ring being mitigated */ struct netmap_adapter *mit_na; /* backpointer */ }; struct netmap_generic_adapter { /* emulated device */ struct netmap_hw_adapter up; /* Pointer to a previously used netmap adapter. */ struct netmap_adapter *prev; /* generic netmap adapters support: * a net_device_ops struct overrides ndo_select_queue(), * save_if_input saves the if_input hook (FreeBSD), * mit implements rx interrupt mitigation, */ struct net_device_ops generic_ndo; void (*save_if_input)(struct ifnet *, struct mbuf *); struct nm_generic_mit *mit; #ifdef linux netdev_tx_t (*save_start_xmit)(struct mbuf *, struct ifnet *); #endif }; #endif /* WITH_GENERIC */ static __inline int netmap_real_rings(struct netmap_adapter *na, enum txrx t) { return nma_get_nrings(na, t) + !!(na->na_flags & NAF_HOST_RINGS); } #ifdef WITH_VALE /* * Bridge wrapper for non VALE ports attached to a VALE switch. * * The real device must already have its own netmap adapter (hwna). * The bridge wrapper and the hwna adapter share the same set of * netmap rings and buffers, but they have two separate sets of * krings descriptors, with tx/rx meanings swapped: * * netmap * bwrap krings rings krings hwna * +------+ +------+ +-----+ +------+ +------+ * |tx_rings->| |\ /| |----| |<-tx_rings| * | | +------+ \ / +-----+ +------+ | | * | | X | | * | | / \ | | * | | +------+/ \+-----+ +------+ | | * |rx_rings->| | | |----| |<-rx_rings| * | | +------+ +-----+ +------+ | | * +------+ +------+ * * - packets coming from the bridge go to the brwap rx rings, * which are also the hwna tx rings. The bwrap notify callback * will then complete the hwna tx (see netmap_bwrap_notify). * * - packets coming from the outside go to the hwna rx rings, * which are also the bwrap tx rings. The (overwritten) hwna * notify method will then complete the bridge tx * (see netmap_bwrap_intr_notify). * * The bridge wrapper may optionally connect the hwna 'host' rings * to the bridge. This is done by using a second port in the * bridge and connecting it to the 'host' netmap_vp_adapter * contained in the netmap_bwrap_adapter. The brwap host adapter * cross-links the hwna host rings in the same way as shown above. * * - packets coming from the bridge and directed to the host stack * are handled by the bwrap host notify callback * (see netmap_bwrap_host_notify) * * - packets coming from the host stack are still handled by the * overwritten hwna notify callback (netmap_bwrap_intr_notify), * but are diverted to the host adapter depending on the ring number. * */ struct netmap_bwrap_adapter { struct netmap_vp_adapter up; struct netmap_vp_adapter host; /* for host rings */ struct netmap_adapter *hwna; /* the underlying device */ /* backup of the hwna memory allocator */ struct netmap_mem_d *save_nmd; /* * When we attach a physical interface to the bridge, we * allow the controlling process to terminate, so we need * a place to store the n_detmap_priv_d data structure. * This is only done when physical interfaces * are attached to a bridge. */ struct netmap_priv_d *na_kpriv; }; int netmap_bwrap_attach(const char *name, struct netmap_adapter *); #endif /* WITH_VALE */ #ifdef WITH_PIPES #define NM_MAXPIPES 64 /* max number of pipes per adapter */ struct netmap_pipe_adapter { struct netmap_adapter up; u_int id; /* pipe identifier */ int role; /* either NR_REG_PIPE_MASTER or NR_REG_PIPE_SLAVE */ struct netmap_adapter *parent; /* adapter that owns the memory */ struct netmap_pipe_adapter *peer; /* the other end of the pipe */ int peer_ref; /* 1 iff we are holding a ref to the peer */ u_int parent_slot; /* index in the parent pipe array */ }; #endif /* WITH_PIPES */ /* return slots reserved to rx clients; used in drivers */ static inline uint32_t nm_kr_rxspace(struct netmap_kring *k) { int space = k->nr_hwtail - k->nr_hwcur; if (space < 0) space += k->nkr_num_slots; ND("preserving %d rx slots %d -> %d", space, k->nr_hwcur, k->nr_hwtail); return space; } /* True if no space in the tx ring. only valid after txsync_prologue */ static inline int nm_kr_txempty(struct netmap_kring *kring) { return kring->rcur == kring->nr_hwtail; } /* * protect against multiple threads using the same ring. * also check that the ring has not been stopped. * We only care for 0 or !=0 as a return code. */ #define NM_KR_BUSY 1 #define NM_KR_STOPPED 2 static __inline void nm_kr_put(struct netmap_kring *kr) { NM_ATOMIC_CLEAR(&kr->nr_busy); } static __inline int nm_kr_tryget(struct netmap_kring *kr) { /* check a first time without taking the lock * to avoid starvation for nm_kr_get() */ if (unlikely(kr->nkr_stopped)) { ND("ring %p stopped (%d)", kr, kr->nkr_stopped); return NM_KR_STOPPED; } if (unlikely(NM_ATOMIC_TEST_AND_SET(&kr->nr_busy))) return NM_KR_BUSY; /* check a second time with lock held */ if (unlikely(kr->nkr_stopped)) { ND("ring %p stopped (%d)", kr, kr->nkr_stopped); nm_kr_put(kr); return NM_KR_STOPPED; } return 0; } static __inline void nm_kr_get(struct netmap_kring *kr) { while (NM_ATOMIC_TEST_AND_SET(&kr->nr_busy)) tsleep(kr, 0, "NM_KR_GET", 4); } - - /* * The following functions are used by individual drivers to * support netmap operation. * * netmap_attach() initializes a struct netmap_adapter, allocating the * struct netmap_ring's and the struct selinfo. * * netmap_detach() frees the memory allocated by netmap_attach(). * * netmap_transmit() replaces the if_transmit routine of the interface, * and is used to intercept packets coming from the stack. * * netmap_load_map/netmap_reload_map are helper routines to set/reset * the dmamap for a packet buffer * * netmap_reset() is a helper routine to be called in the hw driver * when reinitializing a ring. It should not be called by * virtual ports (vale, pipes, monitor) */ int netmap_attach(struct netmap_adapter *); void netmap_detach(struct ifnet *); int netmap_transmit(struct ifnet *, struct mbuf *); struct netmap_slot *netmap_reset(struct netmap_adapter *na, enum txrx tx, u_int n, u_int new_cur); int netmap_ring_reinit(struct netmap_kring *); /* default functions to handle rx/tx interrupts */ int netmap_rx_irq(struct ifnet *, u_int, u_int *); #define netmap_tx_irq(_n, _q) netmap_rx_irq(_n, _q, NULL) void netmap_common_irq(struct ifnet *, u_int, u_int *work_done); #ifdef WITH_VALE /* functions used by external modules to interface with VALE */ #define netmap_vp_to_ifp(_vp) ((_vp)->up.ifp) #define netmap_ifp_to_vp(_ifp) (NA(_ifp)->na_vp) #define netmap_ifp_to_host_vp(_ifp) (NA(_ifp)->na_hostvp) #define netmap_bdg_idx(_vp) ((_vp)->bdg_port) const char *netmap_bdg_name(struct netmap_vp_adapter *); #else /* !WITH_VALE */ #define netmap_vp_to_ifp(_vp) NULL #define netmap_ifp_to_vp(_ifp) NULL #define netmap_ifp_to_host_vp(_ifp) NULL #define netmap_bdg_idx(_vp) -1 #define netmap_bdg_name(_vp) NULL #endif /* WITH_VALE */ static inline int nm_netmap_on(struct netmap_adapter *na) { return na && na->na_flags & NAF_NETMAP_ON; } static inline int nm_native_on(struct netmap_adapter *na) { return nm_netmap_on(na) && (na->na_flags & NAF_NATIVE); } /* set/clear native flags and if_transmit/netdev_ops */ static inline void nm_set_native_flags(struct netmap_adapter *na) { struct ifnet *ifp = na->ifp; na->na_flags |= NAF_NETMAP_ON; #ifdef IFCAP_NETMAP /* or FreeBSD ? */ ifp->if_capenable |= IFCAP_NETMAP; #endif #ifdef __FreeBSD__ na->if_transmit = ifp->if_transmit; ifp->if_transmit = netmap_transmit; #else na->if_transmit = (void *)ifp->netdev_ops; ifp->netdev_ops = &((struct netmap_hw_adapter *)na)->nm_ndo; ((struct netmap_hw_adapter *)na)->save_ethtool = ifp->ethtool_ops; ifp->ethtool_ops = &((struct netmap_hw_adapter*)na)->nm_eto; #endif } static inline void nm_clear_native_flags(struct netmap_adapter *na) { struct ifnet *ifp = na->ifp; #ifdef __FreeBSD__ ifp->if_transmit = na->if_transmit; #else ifp->netdev_ops = (void *)na->if_transmit; ifp->ethtool_ops = ((struct netmap_hw_adapter*)na)->save_ethtool; #endif na->na_flags &= ~NAF_NETMAP_ON; #ifdef IFCAP_NETMAP /* or FreeBSD ? */ ifp->if_capenable &= ~IFCAP_NETMAP; #endif } /* check/fix address and len in tx rings */ #if 1 /* debug version */ #define NM_CHECK_ADDR_LEN(_na, _a, _l) do { \ if (_a == NETMAP_BUF_BASE(_na) || _l > NETMAP_BUF_SIZE(_na)) { \ RD(5, "bad addr/len ring %d slot %d idx %d len %d", \ kring->ring_id, nm_i, slot->buf_idx, len); \ if (_l > NETMAP_BUF_SIZE(_na)) \ _l = NETMAP_BUF_SIZE(_na); \ } } while (0) #else /* no debug version */ #define NM_CHECK_ADDR_LEN(_na, _a, _l) do { \ if (_l > NETMAP_BUF_SIZE(_na)) \ _l = NETMAP_BUF_SIZE(_na); \ } while (0) #endif /*---------------------------------------------------------------*/ /* * Support routines used by netmap subsystems * (native drivers, VALE, generic, pipes, monitors, ...) */ /* common routine for all functions that create a netmap adapter. It performs * two main tasks: * - if the na points to an ifp, mark the ifp as netmap capable * using na as its native adapter; * - provide defaults for the setup callbacks and the memory allocator */ int netmap_attach_common(struct netmap_adapter *); /* common actions to be performed on netmap adapter destruction */ void netmap_detach_common(struct netmap_adapter *); /* fill priv->np_[tr]xq{first,last} using the ringid and flags information * coming from a struct nmreq */ int netmap_interp_ringid(struct netmap_priv_d *priv, uint16_t ringid, uint32_t flags); /* update the ring parameters (number and size of tx and rx rings). * It calls the nm_config callback, if available. */ int netmap_update_config(struct netmap_adapter *na); /* create and initialize the common fields of the krings array. * using the information that must be already available in the na. * tailroom can be used to request the allocation of additional * tailroom bytes after the krings array. This is used by * netmap_vp_adapter's (i.e., VALE ports) to make room for * leasing-related data structures */ int netmap_krings_create(struct netmap_adapter *na, u_int tailroom); /* deletes the kring array of the adapter. The array must have * been created using netmap_krings_create */ void netmap_krings_delete(struct netmap_adapter *na); -int netmap_rxsync_from_host(struct netmap_adapter *na, struct thread *td, void *pwait); - /* set the stopped/enabled status of ring * When stopping, they also wait for all current activity on the ring to * terminate. The status change is then notified using the na nm_notify * callback. */ void netmap_set_ring(struct netmap_adapter *, u_int ring_id, enum txrx, int stopped); /* set the stopped/enabled status of all rings of the adapter. */ void netmap_set_all_rings(struct netmap_adapter *, int stopped); /* convenience wrappers for netmap_set_all_rings, used in drivers */ void netmap_disable_all_rings(struct ifnet *); void netmap_enable_all_rings(struct ifnet *); -int netmap_rxsync_from_host(struct netmap_adapter *na, struct thread *td, void *pwait); - -int -netmap_do_regif(struct netmap_priv_d *priv, struct netmap_adapter *na, +int netmap_do_regif(struct netmap_priv_d *priv, struct netmap_adapter *na, uint16_t ringid, uint32_t flags); - u_int nm_bound_var(u_int *v, u_int dflt, u_int lo, u_int hi, const char *msg); int netmap_get_na(struct nmreq *nmr, struct netmap_adapter **na, int create); int netmap_get_hw_na(struct ifnet *ifp, struct netmap_adapter **na); #ifdef WITH_VALE /* * The following bridge-related functions are used by other * kernel modules. * * VALE only supports unicast or broadcast. The lookup * function can return 0 .. NM_BDG_MAXPORTS-1 for regular ports, * NM_BDG_MAXPORTS for broadcast, NM_BDG_MAXPORTS+1 for unknown. * XXX in practice "unknown" might be handled same as broadcast. */ typedef u_int (*bdg_lookup_fn_t)(struct nm_bdg_fwd *ft, uint8_t *ring_nr, struct netmap_vp_adapter *); typedef int (*bdg_config_fn_t)(struct nm_ifreq *); typedef void (*bdg_dtor_fn_t)(const struct netmap_vp_adapter *); struct netmap_bdg_ops { bdg_lookup_fn_t lookup; bdg_config_fn_t config; bdg_dtor_fn_t dtor; }; u_int netmap_bdg_learning(struct nm_bdg_fwd *ft, uint8_t *dst_ring, struct netmap_vp_adapter *); #define NM_BDG_MAXPORTS 254 /* up to 254 */ #define NM_BDG_BROADCAST NM_BDG_MAXPORTS #define NM_BDG_NOPORT (NM_BDG_MAXPORTS+1) #define NM_NAME "vale" /* prefix for bridge port name */ /* these are redefined in case of no VALE support */ int netmap_get_bdg_na(struct nmreq *nmr, struct netmap_adapter **na, int create); struct nm_bridge *netmap_init_bridges2(u_int); void netmap_uninit_bridges2(struct nm_bridge *, u_int); int netmap_init_bridges(void); void netmap_uninit_bridges(void); int netmap_bdg_ctl(struct nmreq *nmr, struct netmap_bdg_ops *bdg_ops); int netmap_bdg_config(struct nmreq *nmr); #else /* !WITH_VALE */ #define netmap_get_bdg_na(_1, _2, _3) 0 #define netmap_init_bridges(_1) 0 #define netmap_uninit_bridges() #define netmap_bdg_ctl(_1, _2) EINVAL #endif /* !WITH_VALE */ #ifdef WITH_PIPES /* max number of pipes per device */ #define NM_MAXPIPES 64 /* XXX how many? */ void netmap_pipe_dealloc(struct netmap_adapter *); int netmap_get_pipe_na(struct nmreq *nmr, struct netmap_adapter **na, int create); #else /* !WITH_PIPES */ #define NM_MAXPIPES 0 #define netmap_pipe_alloc(_1, _2) 0 #define netmap_pipe_dealloc(_1) #define netmap_get_pipe_na(nmr, _2, _3) \ ({ int role__ = (nmr)->nr_flags & NR_REG_MASK; \ (role__ == NR_REG_PIPE_MASTER || \ role__ == NR_REG_PIPE_SLAVE) ? EOPNOTSUPP : 0; }) #endif #ifdef WITH_MONITOR int netmap_get_monitor_na(struct nmreq *nmr, struct netmap_adapter **na, int create); void netmap_monitor_stop(struct netmap_adapter *na); #else #define netmap_get_monitor_na(nmr, _2, _3) \ ((nmr)->nr_flags & (NR_MONITOR_TX | NR_MONITOR_RX) ? EOPNOTSUPP : 0) #endif #ifdef CONFIG_NET_NS struct net *netmap_bns_get(void); void netmap_bns_put(struct net *); void netmap_bns_getbridges(struct nm_bridge **, u_int *); #else #define netmap_bns_get() #define netmap_bns_put(_1) #define netmap_bns_getbridges(b, n) \ do { *b = nm_bridges; *n = NM_BRIDGES; } while (0) #endif /* Various prototypes */ int netmap_poll(struct cdev *dev, int events, struct thread *td); int netmap_init(void); void netmap_fini(void); int netmap_get_memory(struct netmap_priv_d* p); void netmap_dtor(void *data); int netmap_dtor_locked(struct netmap_priv_d *priv); int netmap_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag, struct thread *td); /* netmap_adapter creation/destruction */ // #define NM_DEBUG_PUTGET 1 #ifdef NM_DEBUG_PUTGET #define NM_DBG(f) __##f void __netmap_adapter_get(struct netmap_adapter *na); #define netmap_adapter_get(na) \ do { \ struct netmap_adapter *__na = na; \ D("getting %p:%s (%d)", __na, (__na)->name, (__na)->na_refcount); \ __netmap_adapter_get(__na); \ } while (0) int __netmap_adapter_put(struct netmap_adapter *na); #define netmap_adapter_put(na) \ ({ \ struct netmap_adapter *__na = na; \ D("putting %p:%s (%d)", __na, (__na)->name, (__na)->na_refcount); \ __netmap_adapter_put(__na); \ }) #else /* !NM_DEBUG_PUTGET */ #define NM_DBG(f) f void netmap_adapter_get(struct netmap_adapter *na); int netmap_adapter_put(struct netmap_adapter *na); #endif /* !NM_DEBUG_PUTGET */ /* * module variables */ #define NETMAP_BUF_BASE(na) ((na)->na_lut.lut[0].vaddr) #define NETMAP_BUF_SIZE(na) ((na)->na_lut.objsize) extern int netmap_mitigate; // XXX not really used extern int netmap_no_pendintr; extern int netmap_verbose; // XXX debugging enum { /* verbose flags */ NM_VERB_ON = 1, /* generic verbose */ NM_VERB_HOST = 0x2, /* verbose host stack */ NM_VERB_RXSYNC = 0x10, /* verbose on rxsync/txsync */ NM_VERB_TXSYNC = 0x20, NM_VERB_RXINTR = 0x100, /* verbose on rx/tx intr (driver) */ NM_VERB_TXINTR = 0x200, NM_VERB_NIC_RXSYNC = 0x1000, /* verbose on rx/tx intr (driver) */ NM_VERB_NIC_TXSYNC = 0x2000, }; extern int netmap_txsync_retry; extern int netmap_generic_mit; extern int netmap_generic_ringsize; extern int netmap_generic_rings; /* * NA returns a pointer to the struct netmap adapter from the ifp, * WNA is used to write it. */ #define NA(_ifp) ((struct netmap_adapter *)WNA(_ifp)) /* * Macros to determine if an interface is netmap capable or netmap enabled. * See the magic field in struct netmap_adapter. */ #ifdef __FreeBSD__ /* * on FreeBSD just use if_capabilities and if_capenable. */ #define NETMAP_CAPABLE(ifp) (NA(ifp) && \ (ifp)->if_capabilities & IFCAP_NETMAP ) #define NETMAP_SET_CAPABLE(ifp) \ (ifp)->if_capabilities |= IFCAP_NETMAP #else /* linux */ /* * on linux: * we check if NA(ifp) is set and its first element has a related * magic value. The capenable is within the struct netmap_adapter. */ #define NETMAP_MAGIC 0x52697a7a #define NETMAP_CAPABLE(ifp) (NA(ifp) && \ ((uint32_t)(uintptr_t)NA(ifp) ^ NA(ifp)->magic) == NETMAP_MAGIC ) #define NETMAP_SET_CAPABLE(ifp) \ NA(ifp)->magic = ((uint32_t)(uintptr_t)NA(ifp)) ^ NETMAP_MAGIC #endif /* linux */ #ifdef __FreeBSD__ /* Assigns the device IOMMU domain to an allocator. * Returns -ENOMEM in case the domain is different */ #define nm_iommu_group_id(dev) (0) /* Callback invoked by the dma machinery after a successful dmamap_load */ static void netmap_dmamap_cb(__unused void *arg, __unused bus_dma_segment_t * segs, __unused int nseg, __unused int error) { } /* bus_dmamap_load wrapper: call aforementioned function if map != NULL. * XXX can we do it without a callback ? */ static inline void netmap_load_map(struct netmap_adapter *na, bus_dma_tag_t tag, bus_dmamap_t map, void *buf) { if (map) bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE(na), netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT); } static inline void netmap_unload_map(struct netmap_adapter *na, bus_dma_tag_t tag, bus_dmamap_t map) { if (map) bus_dmamap_unload(tag, map); } /* update the map when a buffer changes. */ static inline void netmap_reload_map(struct netmap_adapter *na, bus_dma_tag_t tag, bus_dmamap_t map, void *buf) { if (map) { bus_dmamap_unload(tag, map); bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE(na), netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT); } } #else /* linux */ int nm_iommu_group_id(bus_dma_tag_t dev); #include static inline void netmap_load_map(struct netmap_adapter *na, bus_dma_tag_t tag, bus_dmamap_t map, void *buf) { if (0 && map) { *map = dma_map_single(na->pdev, buf, na->na_lut.objsize, DMA_BIDIRECTIONAL); } } static inline void netmap_unload_map(struct netmap_adapter *na, bus_dma_tag_t tag, bus_dmamap_t map) { u_int sz = na->na_lut.objsize; if (*map) { dma_unmap_single(na->pdev, *map, sz, DMA_BIDIRECTIONAL); } } static inline void netmap_reload_map(struct netmap_adapter *na, bus_dma_tag_t tag, bus_dmamap_t map, void *buf) { u_int sz = na->na_lut.objsize; if (*map) { dma_unmap_single(na->pdev, *map, sz, DMA_BIDIRECTIONAL); } *map = dma_map_single(na->pdev, buf, sz, DMA_BIDIRECTIONAL); } /* * XXX How do we redefine these functions: * * on linux we need * dma_map_single(&pdev->dev, virt_addr, len, direction) * dma_unmap_single(&adapter->pdev->dev, phys_addr, len, direction * The len can be implicit (on netmap it is NETMAP_BUF_SIZE) * unfortunately the direction is not, so we need to change * something to have a cross API */ #if 0 struct e1000_buffer *buffer_info = &tx_ring->buffer_info[l]; /* set time_stamp *before* dma to help avoid a possible race */ buffer_info->time_stamp = jiffies; buffer_info->mapped_as_page = false; buffer_info->length = len; //buffer_info->next_to_watch = l; /* reload dma map */ dma_unmap_single(&adapter->pdev->dev, buffer_info->dma, NETMAP_BUF_SIZE, DMA_TO_DEVICE); buffer_info->dma = dma_map_single(&adapter->pdev->dev, addr, NETMAP_BUF_SIZE, DMA_TO_DEVICE); if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) { D("dma mapping error"); /* goto dma_error; See e1000_put_txbuf() */ /* XXX reset */ } tx_desc->buffer_addr = htole64(buffer_info->dma); //XXX #endif /* * The bus_dmamap_sync() can be one of wmb() or rmb() depending on direction. */ #define bus_dmamap_sync(_a, _b, _c) #endif /* linux */ /* * functions to map NIC to KRING indexes (n2k) and vice versa (k2n) */ static inline int netmap_idx_n2k(struct netmap_kring *kr, int idx) { int n = kr->nkr_num_slots; idx += kr->nkr_hwofs; if (idx < 0) return idx + n; else if (idx < n) return idx; else return idx - n; } static inline int netmap_idx_k2n(struct netmap_kring *kr, int idx) { int n = kr->nkr_num_slots; idx -= kr->nkr_hwofs; if (idx < 0) return idx + n; else if (idx < n) return idx; else return idx - n; } /* Entries of the look-up table. */ struct lut_entry { void *vaddr; /* virtual address. */ vm_paddr_t paddr; /* physical address. */ }; struct netmap_obj_pool; /* * NMB return the virtual address of a buffer (buffer 0 on bad index) * PNMB also fills the physical address */ static inline void * NMB(struct netmap_adapter *na, struct netmap_slot *slot) { struct lut_entry *lut = na->na_lut.lut; uint32_t i = slot->buf_idx; return (unlikely(i >= na->na_lut.objtotal)) ? lut[0].vaddr : lut[i].vaddr; } static inline void * PNMB(struct netmap_adapter *na, struct netmap_slot *slot, uint64_t *pp) { uint32_t i = slot->buf_idx; struct lut_entry *lut = na->na_lut.lut; void *ret = (i >= na->na_lut.objtotal) ? lut[0].vaddr : lut[i].vaddr; *pp = (i >= na->na_lut.objtotal) ? lut[0].paddr : lut[i].paddr; return ret; } -/* Generic version of NMB, which uses device-specific memory. */ - - -void netmap_txsync_to_host(struct netmap_adapter *na); - - /* - * Structure associated to each thread which registered an interface. + * Structure associated to each netmap file descriptor. + * It is created on open and left unbound (np_nifp == NULL). + * A successful NIOCREGIF will set np_nifp and the first few fields; + * this is protected by a global lock (NMG_LOCK) due to low contention. * - * The first 4 fields of this structure are written by NIOCREGIF and - * read by poll() and NIOC?XSYNC. + * np_refs counts the number of references to the structure: one for the fd, + * plus (on FreeBSD) one for each active mmap which we track ourselves + * (they are not unmapped on close(), unlike linux). + * np_refs is protected by NMG_LOCK. * - * There is low contention among writers (a correct user program - * should have none) and among writers and readers, so we use a - * single global lock to protect the structure initialization; - * since initialization involves the allocation of memory, - * we reuse the memory allocator lock. - * - * Read access to the structure is lock free. Readers must check that - * np_nifp is not NULL before using the other fields. - * If np_nifp is NULL initialization has not been performed, - * so they should return an error to userspace. - * - * The ref_done field (XXX ?) is used to regulate access to the refcount in the - * memory allocator. The refcount must be incremented at most once for - * each open("/dev/netmap"). The increment is performed by the first - * function that calls netmap_get_memory() (currently called by - * mmap(), NIOCGINFO and NIOCREGIF). - * If the refcount is incremented, it is then decremented when the - * private structure is destroyed. + * Read access to the structure is lock free, because ni_nifp once set + * can only go to 0 when nobody is using the entry anymore. Readers + * must check that np_nifp != NULL before using the other fields. */ struct netmap_priv_d { struct netmap_if * volatile np_nifp; /* netmap if descriptor. */ struct netmap_adapter *np_na; uint32_t np_flags; /* from the ioctl */ u_int np_qfirst[NR_TXRX], np_qlast[NR_TXRX]; /* range of tx/rx rings to scan */ uint16_t np_txpoll; /* XXX and also np_rxpoll ? */ - /* np_refcount is only used on FreeBSD */ - int np_refcount; /* use with NMG_LOCK held */ + int np_refs; /* use with NMG_LOCK held */ /* pointers to the selinfo to be used for selrecord. * Either the local or the global one depending on the * number of rings. */ NM_SELINFO_T *np_si[NR_TXRX]; struct thread *np_td; /* kqueue, just debugging */ }; #ifdef WITH_MONITOR struct netmap_monitor_adapter { struct netmap_adapter up; struct netmap_priv_d priv; uint32_t flags; }; #endif /* WITH_MONITOR */ #ifdef WITH_GENERIC /* * generic netmap emulation for devices that do not have * native netmap support. */ int generic_netmap_attach(struct ifnet *ifp); int netmap_catch_rx(struct netmap_generic_adapter *na, int intercept); void generic_rx_handler(struct ifnet *ifp, struct mbuf *m);; void netmap_catch_tx(struct netmap_generic_adapter *na, int enable); int generic_xmit_frame(struct ifnet *ifp, struct mbuf *m, void *addr, u_int len, u_int ring_nr); int generic_find_num_desc(struct ifnet *ifp, u_int *tx, u_int *rx); void generic_find_num_queues(struct ifnet *ifp, u_int *txq, u_int *rxq); static inline struct ifnet* netmap_generic_getifp(struct netmap_generic_adapter *gna) { if (gna->prev) return gna->prev->ifp; return gna->up.up.ifp; } //#define RATE_GENERIC /* Enables communication statistics for generic. */ #ifdef RATE_GENERIC void generic_rate(int txp, int txs, int txi, int rxp, int rxs, int rxi); #else #define generic_rate(txp, txs, txi, rxp, rxs, rxi) #endif /* * netmap_mitigation API. This is used by the generic adapter * to reduce the number of interrupt requests/selwakeup * to clients on incoming packets. */ void netmap_mitigation_init(struct nm_generic_mit *mit, int idx, struct netmap_adapter *na); void netmap_mitigation_start(struct nm_generic_mit *mit); void netmap_mitigation_restart(struct nm_generic_mit *mit); int netmap_mitigation_active(struct nm_generic_mit *mit); void netmap_mitigation_cleanup(struct nm_generic_mit *mit); #endif /* WITH_GENERIC */ /* Shared declarations for the VALE switch. */ /* * Each transmit queue accumulates a batch of packets into * a structure before forwarding. Packets to the same * destination are put in a list using ft_next as a link field. * ft_frags and ft_next are valid only on the first fragment. */ struct nm_bdg_fwd { /* forwarding entry for a bridge */ void *ft_buf; /* netmap or indirect buffer */ uint8_t ft_frags; /* how many fragments (only on 1st frag) */ uint8_t _ft_port; /* dst port (unused) */ uint16_t ft_flags; /* flags, e.g. indirect */ uint16_t ft_len; /* src fragment len */ uint16_t ft_next; /* next packet to same destination */ }; /* struct 'virtio_net_hdr' from linux. */ struct nm_vnet_hdr { #define VIRTIO_NET_HDR_F_NEEDS_CSUM 1 /* Use csum_start, csum_offset */ #define VIRTIO_NET_HDR_F_DATA_VALID 2 /* Csum is valid */ uint8_t flags; #define VIRTIO_NET_HDR_GSO_NONE 0 /* Not a GSO frame */ #define VIRTIO_NET_HDR_GSO_TCPV4 1 /* GSO frame, IPv4 TCP (TSO) */ #define VIRTIO_NET_HDR_GSO_UDP 3 /* GSO frame, IPv4 UDP (UFO) */ #define VIRTIO_NET_HDR_GSO_TCPV6 4 /* GSO frame, IPv6 TCP */ #define VIRTIO_NET_HDR_GSO_ECN 0x80 /* TCP has ECN set */ uint8_t gso_type; uint16_t hdr_len; uint16_t gso_size; uint16_t csum_start; uint16_t csum_offset; }; #define WORST_CASE_GSO_HEADER (14+40+60) /* IPv6 + TCP */ /* Private definitions for IPv4, IPv6, UDP and TCP headers. */ struct nm_iphdr { uint8_t version_ihl; uint8_t tos; uint16_t tot_len; uint16_t id; uint16_t frag_off; uint8_t ttl; uint8_t protocol; uint16_t check; uint32_t saddr; uint32_t daddr; /*The options start here. */ }; struct nm_tcphdr { uint16_t source; uint16_t dest; uint32_t seq; uint32_t ack_seq; uint8_t doff; /* Data offset + Reserved */ uint8_t flags; uint16_t window; uint16_t check; uint16_t urg_ptr; }; struct nm_udphdr { uint16_t source; uint16_t dest; uint16_t len; uint16_t check; }; struct nm_ipv6hdr { uint8_t priority_version; uint8_t flow_lbl[3]; uint16_t payload_len; uint8_t nexthdr; uint8_t hop_limit; uint8_t saddr[16]; uint8_t daddr[16]; }; /* Type used to store a checksum (in host byte order) that hasn't been * folded yet. */ #define rawsum_t uint32_t rawsum_t nm_csum_raw(uint8_t *data, size_t len, rawsum_t cur_sum); uint16_t nm_csum_ipv4(struct nm_iphdr *iph); void nm_csum_tcpudp_ipv4(struct nm_iphdr *iph, void *data, size_t datalen, uint16_t *check); void nm_csum_tcpudp_ipv6(struct nm_ipv6hdr *ip6h, void *data, size_t datalen, uint16_t *check); uint16_t nm_csum_fold(rawsum_t cur_sum); void bdg_mismatch_datapath(struct netmap_vp_adapter *na, struct netmap_vp_adapter *dst_na, struct nm_bdg_fwd *ft_p, struct netmap_ring *ring, u_int *j, u_int lim, u_int *howmany); /* persistent virtual port routines */ int nm_vi_persist(const char *, struct ifnet **); void nm_vi_detach(struct ifnet *); void nm_vi_init_index(void); #endif /* _NET_NETMAP_KERN_H_ */ Index: head/sys/dev/netmap/netmap_vale.c =================================================================== --- head/sys/dev/netmap/netmap_vale.c (revision 285358) +++ head/sys/dev/netmap/netmap_vale.c (revision 285359) @@ -1,2435 +1,2433 @@ /* * Copyright (C) 2013-2014 Universita` di Pisa. 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. */ /* * This module implements the VALE switch for netmap --- VALE SWITCH --- NMG_LOCK() serializes all modifications to switches and ports. A switch cannot be deleted until all ports are gone. For each switch, an SX lock (RWlock on linux) protects deletion of ports. When configuring or deleting a new port, the lock is acquired in exclusive mode (after holding NMG_LOCK). When forwarding, the lock is acquired in shared mode (without NMG_LOCK). The lock is held throughout the entire forwarding cycle, during which the thread may incur in a page fault. Hence it is important that sleepable shared locks are used. On the rx ring, the per-port lock is grabbed initially to reserve a number of slot in the ring, then the lock is released, packets are copied from source to destination, and then the lock is acquired again and the receive ring is updated. (A similar thing is done on the tx ring for NIC and host stack ports attached to the switch) */ /* * OS-specific code that is used only within this file. * Other OS-specific code that must be accessed by drivers * is present in netmap_kern.h */ #if defined(__FreeBSD__) #include /* prerequisite */ __FBSDID("$FreeBSD$"); #include #include #include /* defines used in kernel.h */ #include /* types used in module initialization */ #include /* cdevsw struct, UID, GID */ #include #include /* struct socket */ #include #include #include #include /* sockaddrs */ #include #include #include #include #include /* BIOCIMMEDIATE */ #include /* bus_dmamap_* */ #include #include #define BDG_RWLOCK_T struct rwlock // struct rwlock #define BDG_RWINIT(b) \ rw_init_flags(&(b)->bdg_lock, "bdg lock", RW_NOWITNESS) #define BDG_WLOCK(b) rw_wlock(&(b)->bdg_lock) #define BDG_WUNLOCK(b) rw_wunlock(&(b)->bdg_lock) #define BDG_RLOCK(b) rw_rlock(&(b)->bdg_lock) #define BDG_RTRYLOCK(b) rw_try_rlock(&(b)->bdg_lock) #define BDG_RUNLOCK(b) rw_runlock(&(b)->bdg_lock) #define BDG_RWDESTROY(b) rw_destroy(&(b)->bdg_lock) #elif defined(linux) #include "bsd_glue.h" #elif defined(__APPLE__) #warning OSX support is only partial #include "osx_glue.h" #else #error Unsupported platform #endif /* unsupported */ /* * common headers */ #include #include #include #ifdef WITH_VALE /* * system parameters (most of them in netmap_kern.h) * NM_NAME prefix for switch port names, default "vale" * NM_BDG_MAXPORTS number of ports * NM_BRIDGES max number of switches in the system. * XXX should become a sysctl or tunable * * Switch ports are named valeX:Y where X is the switch name and Y * is the port. If Y matches a physical interface name, the port is * connected to a physical device. * * Unlike physical interfaces, switch ports use their own memory region * for rings and buffers. * The virtual interfaces use per-queue lock instead of core lock. * In the tx loop, we aggregate traffic in batches to make all operations * faster. The batch size is bridge_batch. */ #define NM_BDG_MAXRINGS 16 /* XXX unclear how many. */ #define NM_BDG_MAXSLOTS 4096 /* XXX same as above */ #define NM_BRIDGE_RINGSIZE 1024 /* in the device */ #define NM_BDG_HASH 1024 /* forwarding table entries */ #define NM_BDG_BATCH 1024 /* entries in the forwarding buffer */ #define NM_MULTISEG 64 /* max size of a chain of bufs */ /* actual size of the tables */ #define NM_BDG_BATCH_MAX (NM_BDG_BATCH + NM_MULTISEG) /* NM_FT_NULL terminates a list of slots in the ft */ #define NM_FT_NULL NM_BDG_BATCH_MAX #define NM_BRIDGES 8 /* number of bridges */ /* * bridge_batch is set via sysctl to the max batch size to be * used in the bridge. The actual value may be larger as the * last packet in the block may overflow the size. */ int bridge_batch = NM_BDG_BATCH; /* bridge batch size */ SYSCTL_DECL(_dev_netmap); SYSCTL_INT(_dev_netmap, OID_AUTO, bridge_batch, CTLFLAG_RW, &bridge_batch, 0 , ""); static int netmap_vp_create(struct nmreq *, struct ifnet *, struct netmap_vp_adapter **); static int netmap_vp_reg(struct netmap_adapter *na, int onoff); static int netmap_bwrap_register(struct netmap_adapter *, int onoff); /* * For each output interface, nm_bdg_q is used to construct a list. * bq_len is the number of output buffers (we can have coalescing * during the copy). */ struct nm_bdg_q { uint16_t bq_head; uint16_t bq_tail; uint32_t bq_len; /* number of buffers */ }; /* XXX revise this */ struct nm_hash_ent { uint64_t mac; /* the top 2 bytes are the epoch */ uint64_t ports; }; /* * nm_bridge is a descriptor for a VALE switch. * Interfaces for a bridge are all in bdg_ports[]. * The array has fixed size, an empty entry does not terminate * the search, but lookups only occur on attach/detach so we * don't mind if they are slow. * * The bridge is non blocking on the transmit ports: excess * packets are dropped if there is no room on the output port. * * bdg_lock protects accesses to the bdg_ports array. * This is a rw lock (or equivalent). */ struct nm_bridge { /* XXX what is the proper alignment/layout ? */ BDG_RWLOCK_T bdg_lock; /* protects bdg_ports */ int bdg_namelen; uint32_t bdg_active_ports; /* 0 means free */ char bdg_basename[IFNAMSIZ]; /* Indexes of active ports (up to active_ports) * and all other remaining ports. */ uint8_t bdg_port_index[NM_BDG_MAXPORTS]; struct netmap_vp_adapter *bdg_ports[NM_BDG_MAXPORTS]; /* * The function to decide the destination port. * It returns either of an index of the destination port, * NM_BDG_BROADCAST to broadcast this packet, or NM_BDG_NOPORT not to * forward this packet. ring_nr is the source ring index, and the * function may overwrite this value to forward this packet to a * different ring index. * This function must be set by netmap_bdgctl(). */ struct netmap_bdg_ops bdg_ops; /* the forwarding table, MAC+ports. * XXX should be changed to an argument to be passed to * the lookup function, and allocated on attach */ struct nm_hash_ent ht[NM_BDG_HASH]; #ifdef CONFIG_NET_NS struct net *ns; #endif /* CONFIG_NET_NS */ }; const char* netmap_bdg_name(struct netmap_vp_adapter *vp) { struct nm_bridge *b = vp->na_bdg; if (b == NULL) return NULL; return b->bdg_basename; } #ifndef CONFIG_NET_NS /* * XXX in principle nm_bridges could be created dynamically * Right now we have a static array and deletions are protected * by an exclusive lock. */ struct nm_bridge *nm_bridges; #endif /* !CONFIG_NET_NS */ /* * this is a slightly optimized copy routine which rounds * to multiple of 64 bytes and is often faster than dealing * with other odd sizes. We assume there is enough room * in the source and destination buffers. * * XXX only for multiples of 64 bytes, non overlapped. */ static inline void pkt_copy(void *_src, void *_dst, int l) { uint64_t *src = _src; uint64_t *dst = _dst; if (unlikely(l >= 1024)) { memcpy(dst, src, l); return; } for (; likely(l > 0); l-=64) { *dst++ = *src++; *dst++ = *src++; *dst++ = *src++; *dst++ = *src++; *dst++ = *src++; *dst++ = *src++; *dst++ = *src++; *dst++ = *src++; } } /* * locate a bridge among the existing ones. * MUST BE CALLED WITH NMG_LOCK() * * a ':' in the name terminates the bridge name. Otherwise, just NM_NAME. * We assume that this is called with a name of at least NM_NAME chars. */ static struct nm_bridge * nm_find_bridge(const char *name, int create) { int i, l, namelen; struct nm_bridge *b = NULL, *bridges; u_int num_bridges; NMG_LOCK_ASSERT(); netmap_bns_getbridges(&bridges, &num_bridges); namelen = strlen(NM_NAME); /* base length */ l = name ? strlen(name) : 0; /* actual length */ if (l < namelen) { D("invalid bridge name %s", name ? name : NULL); return NULL; } for (i = namelen + 1; i < l; i++) { if (name[i] == ':') { namelen = i; break; } } if (namelen >= IFNAMSIZ) namelen = IFNAMSIZ; ND("--- prefix is '%.*s' ---", namelen, name); /* lookup the name, remember empty slot if there is one */ for (i = 0; i < num_bridges; i++) { struct nm_bridge *x = bridges + i; if (x->bdg_active_ports == 0) { if (create && b == NULL) b = x; /* record empty slot */ } else if (x->bdg_namelen != namelen) { continue; } else if (strncmp(name, x->bdg_basename, namelen) == 0) { ND("found '%.*s' at %d", namelen, name, i); b = x; break; } } if (i == num_bridges && b) { /* name not found, can create entry */ /* initialize the bridge */ strncpy(b->bdg_basename, name, namelen); ND("create new bridge %s with ports %d", b->bdg_basename, b->bdg_active_ports); b->bdg_namelen = namelen; b->bdg_active_ports = 0; for (i = 0; i < NM_BDG_MAXPORTS; i++) b->bdg_port_index[i] = i; /* set the default function */ b->bdg_ops.lookup = netmap_bdg_learning; /* reset the MAC address table */ bzero(b->ht, sizeof(struct nm_hash_ent) * NM_BDG_HASH); NM_BNS_GET(b); } return b; } /* * Free the forwarding tables for rings attached to switch ports. */ static void nm_free_bdgfwd(struct netmap_adapter *na) { int nrings, i; struct netmap_kring *kring; NMG_LOCK_ASSERT(); nrings = na->num_tx_rings; kring = na->tx_rings; for (i = 0; i < nrings; i++) { if (kring[i].nkr_ft) { free(kring[i].nkr_ft, M_DEVBUF); kring[i].nkr_ft = NULL; /* protect from freeing twice */ } } } /* * Allocate the forwarding tables for the rings attached to the bridge ports. */ static int nm_alloc_bdgfwd(struct netmap_adapter *na) { int nrings, l, i, num_dstq; struct netmap_kring *kring; NMG_LOCK_ASSERT(); /* all port:rings + broadcast */ num_dstq = NM_BDG_MAXPORTS * NM_BDG_MAXRINGS + 1; l = sizeof(struct nm_bdg_fwd) * NM_BDG_BATCH_MAX; l += sizeof(struct nm_bdg_q) * num_dstq; l += sizeof(uint16_t) * NM_BDG_BATCH_MAX; nrings = netmap_real_rings(na, NR_TX); kring = na->tx_rings; for (i = 0; i < nrings; i++) { struct nm_bdg_fwd *ft; struct nm_bdg_q *dstq; int j; ft = malloc(l, M_DEVBUF, M_NOWAIT | M_ZERO); if (!ft) { nm_free_bdgfwd(na); return ENOMEM; } dstq = (struct nm_bdg_q *)(ft + NM_BDG_BATCH_MAX); for (j = 0; j < num_dstq; j++) { dstq[j].bq_head = dstq[j].bq_tail = NM_FT_NULL; dstq[j].bq_len = 0; } kring[i].nkr_ft = ft; } return 0; } /* remove from bridge b the ports in slots hw and sw * (sw can be -1 if not needed) */ static void netmap_bdg_detach_common(struct nm_bridge *b, int hw, int sw) { int s_hw = hw, s_sw = sw; int i, lim =b->bdg_active_ports; uint8_t tmp[NM_BDG_MAXPORTS]; /* New algorithm: make a copy of bdg_port_index; lookup NA(ifp)->bdg_port and SWNA(ifp)->bdg_port in the array of bdg_port_index, replacing them with entries from the bottom of the array; decrement bdg_active_ports; acquire BDG_WLOCK() and copy back the array. */ if (netmap_verbose) D("detach %d and %d (lim %d)", hw, sw, lim); /* make a copy of the list of active ports, update it, * and then copy back within BDG_WLOCK(). */ memcpy(tmp, b->bdg_port_index, sizeof(tmp)); for (i = 0; (hw >= 0 || sw >= 0) && i < lim; ) { if (hw >= 0 && tmp[i] == hw) { ND("detach hw %d at %d", hw, i); lim--; /* point to last active port */ tmp[i] = tmp[lim]; /* swap with i */ tmp[lim] = hw; /* now this is inactive */ hw = -1; } else if (sw >= 0 && tmp[i] == sw) { ND("detach sw %d at %d", sw, i); lim--; tmp[i] = tmp[lim]; tmp[lim] = sw; sw = -1; } else { i++; } } if (hw >= 0 || sw >= 0) { D("XXX delete failed hw %d sw %d, should panic...", hw, sw); } BDG_WLOCK(b); if (b->bdg_ops.dtor) b->bdg_ops.dtor(b->bdg_ports[s_hw]); b->bdg_ports[s_hw] = NULL; if (s_sw >= 0) { b->bdg_ports[s_sw] = NULL; } memcpy(b->bdg_port_index, tmp, sizeof(tmp)); b->bdg_active_ports = lim; BDG_WUNLOCK(b); ND("now %d active ports", lim); if (lim == 0) { ND("marking bridge %s as free", b->bdg_basename); bzero(&b->bdg_ops, sizeof(b->bdg_ops)); NM_BNS_PUT(b); } } /* nm_bdg_ctl callback for VALE ports */ static int netmap_vp_bdg_ctl(struct netmap_adapter *na, struct nmreq *nmr, int attach) { struct netmap_vp_adapter *vpna = (struct netmap_vp_adapter *)na; struct nm_bridge *b = vpna->na_bdg; if (attach) return 0; /* nothing to do */ if (b) { netmap_set_all_rings(na, 0 /* disable */); netmap_bdg_detach_common(b, vpna->bdg_port, -1); vpna->na_bdg = NULL; netmap_set_all_rings(na, 1 /* enable */); } /* I have took reference just for attach */ netmap_adapter_put(na); return 0; } /* nm_dtor callback for ephemeral VALE ports */ static void netmap_vp_dtor(struct netmap_adapter *na) { struct netmap_vp_adapter *vpna = (struct netmap_vp_adapter*)na; struct nm_bridge *b = vpna->na_bdg; ND("%s has %d references", na->name, na->na_refcount); if (b) { netmap_bdg_detach_common(b, vpna->bdg_port, -1); } } /* nm_dtor callback for persistent VALE ports */ static void netmap_persist_vp_dtor(struct netmap_adapter *na) { struct ifnet *ifp = na->ifp; netmap_vp_dtor(na); na->ifp = NULL; nm_vi_detach(ifp); } /* remove a persistent VALE port from the system */ static int nm_vi_destroy(const char *name) { struct ifnet *ifp; int error; ifp = ifunit_ref(name); if (!ifp) return ENXIO; NMG_LOCK(); /* make sure this is actually a VALE port */ if (!NETMAP_CAPABLE(ifp) || NA(ifp)->nm_register != netmap_vp_reg) { error = EINVAL; goto err; } if (NA(ifp)->na_refcount > 1) { error = EBUSY; goto err; } NMG_UNLOCK(); D("destroying a persistent vale interface %s", ifp->if_xname); /* Linux requires all the references are released * before unregister */ if_rele(ifp); netmap_detach(ifp); return 0; err: NMG_UNLOCK(); if_rele(ifp); return error; } /* * Create a virtual interface registered to the system. * The interface will be attached to a bridge later. */ static int nm_vi_create(struct nmreq *nmr) { struct ifnet *ifp; struct netmap_vp_adapter *vpna; int error; /* don't include VALE prefix */ if (!strncmp(nmr->nr_name, NM_NAME, strlen(NM_NAME))) return EINVAL; ifp = ifunit_ref(nmr->nr_name); if (ifp) { /* already exist, cannot create new one */ if_rele(ifp); return EEXIST; } error = nm_vi_persist(nmr->nr_name, &ifp); if (error) return error; NMG_LOCK(); /* netmap_vp_create creates a struct netmap_vp_adapter */ error = netmap_vp_create(nmr, ifp, &vpna); if (error) { D("error %d", error); nm_vi_detach(ifp); return error; } /* persist-specific routines */ vpna->up.nm_bdg_ctl = netmap_vp_bdg_ctl; vpna->up.nm_dtor = netmap_persist_vp_dtor; netmap_adapter_get(&vpna->up); NMG_UNLOCK(); D("created %s", ifp->if_xname); return 0; } /* Try to get a reference to a netmap adapter attached to a VALE switch. * If the adapter is found (or is created), this function returns 0, a * non NULL pointer is returned into *na, and the caller holds a * reference to the adapter. * If an adapter is not found, then no reference is grabbed and the * function returns an error code, or 0 if there is just a VALE prefix * mismatch. Therefore the caller holds a reference when * (*na != NULL && return == 0). */ int netmap_get_bdg_na(struct nmreq *nmr, struct netmap_adapter **na, int create) { char *nr_name = nmr->nr_name; const char *ifname; struct ifnet *ifp; int error = 0; struct netmap_vp_adapter *vpna, *hostna = NULL; struct nm_bridge *b; int i, j, cand = -1, cand2 = -1; int needed; *na = NULL; /* default return value */ /* first try to see if this is a bridge port. */ NMG_LOCK_ASSERT(); if (strncmp(nr_name, NM_NAME, sizeof(NM_NAME) - 1)) { return 0; /* no error, but no VALE prefix */ } b = nm_find_bridge(nr_name, create); if (b == NULL) { D("no bridges available for '%s'", nr_name); return (create ? ENOMEM : ENXIO); } if (strlen(nr_name) < b->bdg_namelen) /* impossible */ panic("x"); /* Now we are sure that name starts with the bridge's name, * lookup the port in the bridge. We need to scan the entire * list. It is not important to hold a WLOCK on the bridge * during the search because NMG_LOCK already guarantees * that there are no other possible writers. */ /* lookup in the local list of ports */ for (j = 0; j < b->bdg_active_ports; j++) { i = b->bdg_port_index[j]; vpna = b->bdg_ports[i]; // KASSERT(na != NULL); ND("checking %s", vpna->up.name); if (!strcmp(vpna->up.name, nr_name)) { netmap_adapter_get(&vpna->up); ND("found existing if %s refs %d", nr_name) *na = &vpna->up; return 0; } } /* not found, should we create it? */ if (!create) return ENXIO; /* yes we should, see if we have space to attach entries */ needed = 2; /* in some cases we only need 1 */ if (b->bdg_active_ports + needed >= NM_BDG_MAXPORTS) { D("bridge full %d, cannot create new port", b->bdg_active_ports); return ENOMEM; } /* record the next two ports available, but do not allocate yet */ cand = b->bdg_port_index[b->bdg_active_ports]; cand2 = b->bdg_port_index[b->bdg_active_ports + 1]; ND("+++ bridge %s port %s used %d avail %d %d", b->bdg_basename, ifname, b->bdg_active_ports, cand, cand2); /* * try see if there is a matching NIC with this name * (after the bridge's name) */ ifname = nr_name + b->bdg_namelen + 1; ifp = ifunit_ref(ifname); if (!ifp) { /* Create an ephemeral virtual port * This block contains all the ephemeral-specific logics */ if (nmr->nr_cmd) { /* nr_cmd must be 0 for a virtual port */ return EINVAL; } /* bdg_netmap_attach creates a struct netmap_adapter */ error = netmap_vp_create(nmr, NULL, &vpna); if (error) { D("error %d", error); free(ifp, M_DEVBUF); return error; } /* shortcut - we can skip get_hw_na(), * ownership check and nm_bdg_attach() */ } else { struct netmap_adapter *hw; error = netmap_get_hw_na(ifp, &hw); if (error || hw == NULL) goto out; /* host adapter might not be created */ error = hw->nm_bdg_attach(nr_name, hw); if (error) goto out; vpna = hw->na_vp; hostna = hw->na_hostvp; if_rele(ifp); if (nmr->nr_arg1 != NETMAP_BDG_HOST) hostna = NULL; } BDG_WLOCK(b); vpna->bdg_port = cand; ND("NIC %p to bridge port %d", vpna, cand); /* bind the port to the bridge (virtual ports are not active) */ b->bdg_ports[cand] = vpna; vpna->na_bdg = b; b->bdg_active_ports++; if (hostna != NULL) { /* also bind the host stack to the bridge */ b->bdg_ports[cand2] = hostna; hostna->bdg_port = cand2; hostna->na_bdg = b; b->bdg_active_ports++; ND("host %p to bridge port %d", hostna, cand2); } ND("if %s refs %d", ifname, vpna->up.na_refcount); BDG_WUNLOCK(b); *na = &vpna->up; netmap_adapter_get(*na); return 0; out: if_rele(ifp); return error; } /* Process NETMAP_BDG_ATTACH */ static int nm_bdg_ctl_attach(struct nmreq *nmr) { struct netmap_adapter *na; int error; NMG_LOCK(); error = netmap_get_bdg_na(nmr, &na, 1 /* create if not exists */); if (error) /* no device */ goto unlock_exit; if (na == NULL) { /* VALE prefix missing */ error = EINVAL; goto unlock_exit; } if (NETMAP_OWNED_BY_ANY(na)) { error = EBUSY; goto unref_exit; } if (na->nm_bdg_ctl) { /* nop for VALE ports. The bwrap needs to put the hwna * in netmap mode (see netmap_bwrap_bdg_ctl) */ error = na->nm_bdg_ctl(na, nmr, 1); if (error) goto unref_exit; ND("registered %s to netmap-mode", na->name); } NMG_UNLOCK(); return 0; unref_exit: netmap_adapter_put(na); unlock_exit: NMG_UNLOCK(); return error; } /* process NETMAP_BDG_DETACH */ static int nm_bdg_ctl_detach(struct nmreq *nmr) { struct netmap_adapter *na; int error; NMG_LOCK(); error = netmap_get_bdg_na(nmr, &na, 0 /* don't create */); if (error) { /* no device, or another bridge or user owns the device */ goto unlock_exit; } if (na == NULL) { /* VALE prefix missing */ error = EINVAL; goto unlock_exit; } if (na->nm_bdg_ctl) { /* remove the port from bridge. The bwrap * also needs to put the hwna in normal mode */ error = na->nm_bdg_ctl(na, nmr, 0); } netmap_adapter_put(na); unlock_exit: NMG_UNLOCK(); return error; } /* Called by either user's context (netmap_ioctl()) * or external kernel modules (e.g., Openvswitch). * Operation is indicated in nmr->nr_cmd. * NETMAP_BDG_OPS that sets configure/lookup/dtor functions to the bridge * requires bdg_ops argument; the other commands ignore this argument. * * Called without NMG_LOCK. */ int netmap_bdg_ctl(struct nmreq *nmr, struct netmap_bdg_ops *bdg_ops) { struct nm_bridge *b, *bridges; struct netmap_adapter *na; struct netmap_vp_adapter *vpna; char *name = nmr->nr_name; int cmd = nmr->nr_cmd, namelen = strlen(name); int error = 0, i, j; u_int num_bridges; netmap_bns_getbridges(&bridges, &num_bridges); switch (cmd) { case NETMAP_BDG_NEWIF: error = nm_vi_create(nmr); break; case NETMAP_BDG_DELIF: error = nm_vi_destroy(nmr->nr_name); break; case NETMAP_BDG_ATTACH: error = nm_bdg_ctl_attach(nmr); break; case NETMAP_BDG_DETACH: error = nm_bdg_ctl_detach(nmr); break; case NETMAP_BDG_LIST: /* this is used to enumerate bridges and ports */ if (namelen) { /* look up indexes of bridge and port */ if (strncmp(name, NM_NAME, strlen(NM_NAME))) { error = EINVAL; break; } NMG_LOCK(); b = nm_find_bridge(name, 0 /* don't create */); if (!b) { error = ENOENT; NMG_UNLOCK(); break; } error = ENOENT; for (j = 0; j < b->bdg_active_ports; j++) { i = b->bdg_port_index[j]; vpna = b->bdg_ports[i]; if (vpna == NULL) { D("---AAAAAAAAARGH-------"); continue; } /* the former and the latter identify a * virtual port and a NIC, respectively */ if (!strcmp(vpna->up.name, name)) { /* bridge index */ nmr->nr_arg1 = b - bridges; nmr->nr_arg2 = i; /* port index */ error = 0; break; } } NMG_UNLOCK(); } else { /* return the first non-empty entry starting from * bridge nr_arg1 and port nr_arg2. * * Users can detect the end of the same bridge by * seeing the new and old value of nr_arg1, and can * detect the end of all the bridge by error != 0 */ i = nmr->nr_arg1; j = nmr->nr_arg2; NMG_LOCK(); for (error = ENOENT; i < NM_BRIDGES; i++) { b = bridges + i; if (j >= b->bdg_active_ports) { j = 0; /* following bridges scan from 0 */ continue; } nmr->nr_arg1 = i; nmr->nr_arg2 = j; j = b->bdg_port_index[j]; vpna = b->bdg_ports[j]; strncpy(name, vpna->up.name, (size_t)IFNAMSIZ); error = 0; break; } NMG_UNLOCK(); } break; case NETMAP_BDG_REGOPS: /* XXX this should not be available from userspace */ /* register callbacks to the given bridge. * nmr->nr_name may be just bridge's name (including ':' * if it is not just NM_NAME). */ if (!bdg_ops) { error = EINVAL; break; } NMG_LOCK(); b = nm_find_bridge(name, 0 /* don't create */); if (!b) { error = EINVAL; } else { b->bdg_ops = *bdg_ops; } NMG_UNLOCK(); break; case NETMAP_BDG_VNET_HDR: /* Valid lengths for the virtio-net header are 0 (no header), 10 and 12. */ if (nmr->nr_arg1 != 0 && nmr->nr_arg1 != sizeof(struct nm_vnet_hdr) && nmr->nr_arg1 != 12) { error = EINVAL; break; } NMG_LOCK(); error = netmap_get_bdg_na(nmr, &na, 0); if (na && !error) { vpna = (struct netmap_vp_adapter *)na; vpna->virt_hdr_len = nmr->nr_arg1; if (vpna->virt_hdr_len) vpna->mfs = NETMAP_BUF_SIZE(na); D("Using vnet_hdr_len %d for %p", vpna->virt_hdr_len, vpna); netmap_adapter_put(na); } NMG_UNLOCK(); break; default: D("invalid cmd (nmr->nr_cmd) (0x%x)", cmd); error = EINVAL; break; } return error; } int netmap_bdg_config(struct nmreq *nmr) { struct nm_bridge *b; int error = EINVAL; NMG_LOCK(); b = nm_find_bridge(nmr->nr_name, 0); if (!b) { NMG_UNLOCK(); return error; } NMG_UNLOCK(); /* Don't call config() with NMG_LOCK() held */ BDG_RLOCK(b); if (b->bdg_ops.config != NULL) error = b->bdg_ops.config((struct nm_ifreq *)nmr); BDG_RUNLOCK(b); return error; } /* nm_krings_create callback for VALE ports. * Calls the standard netmap_krings_create, then adds leases on rx * rings and bdgfwd on tx rings. */ static int netmap_vp_krings_create(struct netmap_adapter *na) { u_int tailroom; int error, i; uint32_t *leases; u_int nrx = netmap_real_rings(na, NR_RX); /* * Leases are attached to RX rings on vale ports */ tailroom = sizeof(uint32_t) * na->num_rx_desc * nrx; error = netmap_krings_create(na, tailroom); if (error) return error; leases = na->tailroom; for (i = 0; i < nrx; i++) { /* Receive rings */ na->rx_rings[i].nkr_leases = leases; leases += na->num_rx_desc; } error = nm_alloc_bdgfwd(na); if (error) { netmap_krings_delete(na); return error; } return 0; } /* nm_krings_delete callback for VALE ports. */ static void netmap_vp_krings_delete(struct netmap_adapter *na) { nm_free_bdgfwd(na); netmap_krings_delete(na); } static int nm_bdg_flush(struct nm_bdg_fwd *ft, u_int n, struct netmap_vp_adapter *na, u_int ring_nr); /* * main dispatch routine for the bridge. * Grab packets from a kring, move them into the ft structure * associated to the tx (input) port. Max one instance per port, * filtered on input (ioctl, poll or XXX). * Returns the next position in the ring. */ static int nm_bdg_preflush(struct netmap_kring *kring, u_int end) { struct netmap_vp_adapter *na = (struct netmap_vp_adapter*)kring->na; struct netmap_ring *ring = kring->ring; struct nm_bdg_fwd *ft; u_int ring_nr = kring->ring_id; u_int j = kring->nr_hwcur, lim = kring->nkr_num_slots - 1; u_int ft_i = 0; /* start from 0 */ u_int frags = 1; /* how many frags ? */ struct nm_bridge *b = na->na_bdg; /* To protect against modifications to the bridge we acquire a * shared lock, waiting if we can sleep (if the source port is * attached to a user process) or with a trylock otherwise (NICs). */ ND("wait rlock for %d packets", ((j > end ? lim+1 : 0) + end) - j); if (na->up.na_flags & NAF_BDG_MAYSLEEP) BDG_RLOCK(b); else if (!BDG_RTRYLOCK(b)) return 0; ND(5, "rlock acquired for %d packets", ((j > end ? lim+1 : 0) + end) - j); ft = kring->nkr_ft; for (; likely(j != end); j = nm_next(j, lim)) { struct netmap_slot *slot = &ring->slot[j]; char *buf; ft[ft_i].ft_len = slot->len; ft[ft_i].ft_flags = slot->flags; ND("flags is 0x%x", slot->flags); /* we do not use the buf changed flag, but we still need to reset it */ slot->flags &= ~NS_BUF_CHANGED; /* this slot goes into a list so initialize the link field */ ft[ft_i].ft_next = NM_FT_NULL; buf = ft[ft_i].ft_buf = (slot->flags & NS_INDIRECT) ? (void *)(uintptr_t)slot->ptr : NMB(&na->up, slot); if (unlikely(buf == NULL)) { RD(5, "NULL %s buffer pointer from %s slot %d len %d", (slot->flags & NS_INDIRECT) ? "INDIRECT" : "DIRECT", kring->name, j, ft[ft_i].ft_len); buf = ft[ft_i].ft_buf = NETMAP_BUF_BASE(&na->up); ft[ft_i].ft_len = 0; ft[ft_i].ft_flags = 0; } __builtin_prefetch(buf); ++ft_i; if (slot->flags & NS_MOREFRAG) { frags++; continue; } if (unlikely(netmap_verbose && frags > 1)) RD(5, "%d frags at %d", frags, ft_i - frags); ft[ft_i - frags].ft_frags = frags; frags = 1; if (unlikely((int)ft_i >= bridge_batch)) ft_i = nm_bdg_flush(ft, ft_i, na, ring_nr); } if (frags > 1) { D("truncate incomplete fragment at %d (%d frags)", ft_i, frags); // ft_i > 0, ft[ft_i-1].flags has NS_MOREFRAG ft[ft_i - 1].ft_frags &= ~NS_MOREFRAG; ft[ft_i - frags].ft_frags = frags - 1; } if (ft_i) ft_i = nm_bdg_flush(ft, ft_i, na, ring_nr); BDG_RUNLOCK(b); return j; } /* ----- FreeBSD if_bridge hash function ------- */ /* * The following hash function is adapted from "Hash Functions" by Bob Jenkins * ("Algorithm Alley", Dr. Dobbs Journal, September 1997). * * http://www.burtleburtle.net/bob/hash/spooky.html */ #define mix(a, b, c) \ do { \ a -= b; a -= c; a ^= (c >> 13); \ b -= c; b -= a; b ^= (a << 8); \ c -= a; c -= b; c ^= (b >> 13); \ a -= b; a -= c; a ^= (c >> 12); \ b -= c; b -= a; b ^= (a << 16); \ c -= a; c -= b; c ^= (b >> 5); \ a -= b; a -= c; a ^= (c >> 3); \ b -= c; b -= a; b ^= (a << 10); \ c -= a; c -= b; c ^= (b >> 15); \ } while (/*CONSTCOND*/0) static __inline uint32_t nm_bridge_rthash(const uint8_t *addr) { uint32_t a = 0x9e3779b9, b = 0x9e3779b9, c = 0; // hask key b += addr[5] << 8; b += addr[4]; a += addr[3] << 24; a += addr[2] << 16; a += addr[1] << 8; a += addr[0]; mix(a, b, c); #define BRIDGE_RTHASH_MASK (NM_BDG_HASH-1) return (c & BRIDGE_RTHASH_MASK); } #undef mix /* nm_register callback for VALE ports */ static int netmap_vp_reg(struct netmap_adapter *na, int onoff) { struct netmap_vp_adapter *vpna = (struct netmap_vp_adapter*)na; /* persistent ports may be put in netmap mode * before being attached to a bridge */ if (vpna->na_bdg) BDG_WLOCK(vpna->na_bdg); if (onoff) { na->na_flags |= NAF_NETMAP_ON; /* XXX on FreeBSD, persistent VALE ports should also * toggle IFCAP_NETMAP in na->ifp (2014-03-16) */ } else { na->na_flags &= ~NAF_NETMAP_ON; } if (vpna->na_bdg) BDG_WUNLOCK(vpna->na_bdg); return 0; } /* * Lookup function for a learning bridge. * Update the hash table with the source address, * and then returns the destination port index, and the * ring in *dst_ring (at the moment, always use ring 0) */ u_int netmap_bdg_learning(struct nm_bdg_fwd *ft, uint8_t *dst_ring, struct netmap_vp_adapter *na) { uint8_t *buf = ft->ft_buf; u_int buf_len = ft->ft_len; struct nm_hash_ent *ht = na->na_bdg->ht; uint32_t sh, dh; u_int dst, mysrc = na->bdg_port; uint64_t smac, dmac; /* safety check, unfortunately we have many cases */ if (buf_len >= 14 + na->virt_hdr_len) { /* virthdr + mac_hdr in the same slot */ buf += na->virt_hdr_len; buf_len -= na->virt_hdr_len; } else if (buf_len == na->virt_hdr_len && ft->ft_flags & NS_MOREFRAG) { /* only header in first fragment */ ft++; buf = ft->ft_buf; buf_len = ft->ft_len; } else { RD(5, "invalid buf format, length %d", buf_len); return NM_BDG_NOPORT; } dmac = le64toh(*(uint64_t *)(buf)) & 0xffffffffffff; smac = le64toh(*(uint64_t *)(buf + 4)); smac >>= 16; /* * The hash is somewhat expensive, there might be some * worthwhile optimizations here. */ if (((buf[6] & 1) == 0) && (na->last_smac != smac)) { /* valid src */ uint8_t *s = buf+6; sh = nm_bridge_rthash(s); // XXX hash of source /* update source port forwarding entry */ na->last_smac = ht[sh].mac = smac; /* XXX expire ? */ ht[sh].ports = mysrc; if (netmap_verbose) D("src %02x:%02x:%02x:%02x:%02x:%02x on port %d", s[0], s[1], s[2], s[3], s[4], s[5], mysrc); } dst = NM_BDG_BROADCAST; if ((buf[0] & 1) == 0) { /* unicast */ dh = nm_bridge_rthash(buf); // XXX hash of dst if (ht[dh].mac == dmac) { /* found dst */ dst = ht[dh].ports; } /* XXX otherwise return NM_BDG_UNKNOWN ? */ } return dst; } /* * Available space in the ring. Only used in VALE code * and only with is_rx = 1 */ static inline uint32_t nm_kr_space(struct netmap_kring *k, int is_rx) { int space; if (is_rx) { int busy = k->nkr_hwlease - k->nr_hwcur; if (busy < 0) busy += k->nkr_num_slots; space = k->nkr_num_slots - 1 - busy; } else { /* XXX never used in this branch */ space = k->nr_hwtail - k->nkr_hwlease; if (space < 0) space += k->nkr_num_slots; } #if 0 // sanity check if (k->nkr_hwlease >= k->nkr_num_slots || k->nr_hwcur >= k->nkr_num_slots || k->nr_tail >= k->nkr_num_slots || busy < 0 || busy >= k->nkr_num_slots) { D("invalid kring, cur %d tail %d lease %d lease_idx %d lim %d", k->nr_hwcur, k->nr_hwtail, k->nkr_hwlease, k->nkr_lease_idx, k->nkr_num_slots); } #endif return space; } /* make a lease on the kring for N positions. return the * lease index * XXX only used in VALE code and with is_rx = 1 */ static inline uint32_t nm_kr_lease(struct netmap_kring *k, u_int n, int is_rx) { uint32_t lim = k->nkr_num_slots - 1; uint32_t lease_idx = k->nkr_lease_idx; k->nkr_leases[lease_idx] = NR_NOSLOT; k->nkr_lease_idx = nm_next(lease_idx, lim); if (n > nm_kr_space(k, is_rx)) { D("invalid request for %d slots", n); panic("x"); } /* XXX verify that there are n slots */ k->nkr_hwlease += n; if (k->nkr_hwlease > lim) k->nkr_hwlease -= lim + 1; if (k->nkr_hwlease >= k->nkr_num_slots || k->nr_hwcur >= k->nkr_num_slots || k->nr_hwtail >= k->nkr_num_slots || k->nkr_lease_idx >= k->nkr_num_slots) { D("invalid kring %s, cur %d tail %d lease %d lease_idx %d lim %d", k->na->name, k->nr_hwcur, k->nr_hwtail, k->nkr_hwlease, k->nkr_lease_idx, k->nkr_num_slots); } return lease_idx; } /* * * This flush routine supports only unicast and broadcast but a large * number of ports, and lets us replace the learn and dispatch functions. */ int nm_bdg_flush(struct nm_bdg_fwd *ft, u_int n, struct netmap_vp_adapter *na, u_int ring_nr) { struct nm_bdg_q *dst_ents, *brddst; uint16_t num_dsts = 0, *dsts; struct nm_bridge *b = na->na_bdg; u_int i, j, me = na->bdg_port; /* * The work area (pointed by ft) is followed by an array of * pointers to queues , dst_ents; there are NM_BDG_MAXRINGS * queues per port plus one for the broadcast traffic. * Then we have an array of destination indexes. */ dst_ents = (struct nm_bdg_q *)(ft + NM_BDG_BATCH_MAX); dsts = (uint16_t *)(dst_ents + NM_BDG_MAXPORTS * NM_BDG_MAXRINGS + 1); /* first pass: find a destination for each packet in the batch */ for (i = 0; likely(i < n); i += ft[i].ft_frags) { uint8_t dst_ring = ring_nr; /* default, same ring as origin */ uint16_t dst_port, d_i; struct nm_bdg_q *d; ND("slot %d frags %d", i, ft[i].ft_frags); /* Drop the packet if the virtio-net header is not into the first fragment nor at the very beginning of the second. */ if (unlikely(na->virt_hdr_len > ft[i].ft_len)) continue; dst_port = b->bdg_ops.lookup(&ft[i], &dst_ring, na); if (netmap_verbose > 255) RD(5, "slot %d port %d -> %d", i, me, dst_port); if (dst_port == NM_BDG_NOPORT) continue; /* this packet is identified to be dropped */ else if (unlikely(dst_port > NM_BDG_MAXPORTS)) continue; else if (dst_port == NM_BDG_BROADCAST) dst_ring = 0; /* broadcasts always go to ring 0 */ else if (unlikely(dst_port == me || !b->bdg_ports[dst_port])) continue; /* get a position in the scratch pad */ d_i = dst_port * NM_BDG_MAXRINGS + dst_ring; d = dst_ents + d_i; /* append the first fragment to the list */ if (d->bq_head == NM_FT_NULL) { /* new destination */ d->bq_head = d->bq_tail = i; /* remember this position to be scanned later */ if (dst_port != NM_BDG_BROADCAST) dsts[num_dsts++] = d_i; } else { ft[d->bq_tail].ft_next = i; d->bq_tail = i; } d->bq_len += ft[i].ft_frags; } /* * Broadcast traffic goes to ring 0 on all destinations. * So we need to add these rings to the list of ports to scan. * XXX at the moment we scan all NM_BDG_MAXPORTS ports, which is * expensive. We should keep a compact list of active destinations * so we could shorten this loop. */ brddst = dst_ents + NM_BDG_BROADCAST * NM_BDG_MAXRINGS; if (brddst->bq_head != NM_FT_NULL) { for (j = 0; likely(j < b->bdg_active_ports); j++) { uint16_t d_i; i = b->bdg_port_index[j]; if (unlikely(i == me)) continue; d_i = i * NM_BDG_MAXRINGS; if (dst_ents[d_i].bq_head == NM_FT_NULL) dsts[num_dsts++] = d_i; } } ND(5, "pass 1 done %d pkts %d dsts", n, num_dsts); /* second pass: scan destinations */ for (i = 0; i < num_dsts; i++) { struct netmap_vp_adapter *dst_na; struct netmap_kring *kring; struct netmap_ring *ring; u_int dst_nr, lim, j, d_i, next, brd_next; u_int needed, howmany; int retry = netmap_txsync_retry; struct nm_bdg_q *d; uint32_t my_start = 0, lease_idx = 0; int nrings; int virt_hdr_mismatch = 0; d_i = dsts[i]; ND("second pass %d port %d", i, d_i); d = dst_ents + d_i; // XXX fix the division dst_na = b->bdg_ports[d_i/NM_BDG_MAXRINGS]; /* protect from the lookup function returning an inactive * destination port */ if (unlikely(dst_na == NULL)) goto cleanup; if (dst_na->up.na_flags & NAF_SW_ONLY) goto cleanup; /* * The interface may be in !netmap mode in two cases: * - when na is attached but not activated yet; * - when na is being deactivated but is still attached. */ if (unlikely(!nm_netmap_on(&dst_na->up))) { ND("not in netmap mode!"); goto cleanup; } /* there is at least one either unicast or broadcast packet */ brd_next = brddst->bq_head; next = d->bq_head; /* we need to reserve this many slots. If fewer are * available, some packets will be dropped. * Packets may have multiple fragments, so we may not use * there is a chance that we may not use all of the slots * we have claimed, so we will need to handle the leftover * ones when we regain the lock. */ needed = d->bq_len + brddst->bq_len; if (unlikely(dst_na->virt_hdr_len != na->virt_hdr_len)) { RD(3, "virt_hdr_mismatch, src %d dst %d", na->virt_hdr_len, dst_na->virt_hdr_len); /* There is a virtio-net header/offloadings mismatch between * source and destination. The slower mismatch datapath will * be used to cope with all the mismatches. */ virt_hdr_mismatch = 1; if (dst_na->mfs < na->mfs) { /* We may need to do segmentation offloadings, and so * we may need a number of destination slots greater * than the number of input slots ('needed'). * We look for the smallest integer 'x' which satisfies: * needed * na->mfs + x * H <= x * na->mfs * where 'H' is the length of the longest header that may * be replicated in the segmentation process (e.g. for * TCPv4 we must account for ethernet header, IP header * and TCPv4 header). */ needed = (needed * na->mfs) / (dst_na->mfs - WORST_CASE_GSO_HEADER) + 1; ND(3, "srcmtu=%u, dstmtu=%u, x=%u", na->mfs, dst_na->mfs, needed); } } ND(5, "pass 2 dst %d is %x %s", i, d_i, is_vp ? "virtual" : "nic/host"); dst_nr = d_i & (NM_BDG_MAXRINGS-1); nrings = dst_na->up.num_rx_rings; if (dst_nr >= nrings) dst_nr = dst_nr % nrings; kring = &dst_na->up.rx_rings[dst_nr]; ring = kring->ring; lim = kring->nkr_num_slots - 1; retry: if (dst_na->retry && retry) { /* try to get some free slot from the previous run */ kring->nm_notify(kring, 0); /* actually useful only for bwraps, since there * the notify will trigger a txsync on the hwna. VALE ports * have dst_na->retry == 0 */ } /* reserve the buffers in the queue and an entry * to report completion, and drop lock. * XXX this might become a helper function. */ mtx_lock(&kring->q_lock); if (kring->nkr_stopped) { mtx_unlock(&kring->q_lock); goto cleanup; } my_start = j = kring->nkr_hwlease; howmany = nm_kr_space(kring, 1); if (needed < howmany) howmany = needed; lease_idx = nm_kr_lease(kring, howmany, 1); mtx_unlock(&kring->q_lock); /* only retry if we need more than available slots */ if (retry && needed <= howmany) retry = 0; /* copy to the destination queue */ while (howmany > 0) { struct netmap_slot *slot; struct nm_bdg_fwd *ft_p, *ft_end; u_int cnt; /* find the queue from which we pick next packet. * NM_FT_NULL is always higher than valid indexes * so we never dereference it if the other list * has packets (and if both are empty we never * get here). */ if (next < brd_next) { ft_p = ft + next; next = ft_p->ft_next; } else { /* insert broadcast */ ft_p = ft + brd_next; brd_next = ft_p->ft_next; } cnt = ft_p->ft_frags; // cnt > 0 if (unlikely(cnt > howmany)) break; /* no more space */ if (netmap_verbose && cnt > 1) RD(5, "rx %d frags to %d", cnt, j); ft_end = ft_p + cnt; if (unlikely(virt_hdr_mismatch)) { bdg_mismatch_datapath(na, dst_na, ft_p, ring, &j, lim, &howmany); } else { howmany -= cnt; do { char *dst, *src = ft_p->ft_buf; size_t copy_len = ft_p->ft_len, dst_len = copy_len; slot = &ring->slot[j]; dst = NMB(&dst_na->up, slot); ND("send [%d] %d(%d) bytes at %s:%d", i, (int)copy_len, (int)dst_len, NM_IFPNAME(dst_ifp), j); /* round to a multiple of 64 */ copy_len = (copy_len + 63) & ~63; if (unlikely(copy_len > NETMAP_BUF_SIZE(&dst_na->up) || copy_len > NETMAP_BUF_SIZE(&na->up))) { RD(5, "invalid len %d, down to 64", (int)copy_len); copy_len = dst_len = 64; // XXX } if (ft_p->ft_flags & NS_INDIRECT) { if (copyin(src, dst, copy_len)) { // invalid user pointer, pretend len is 0 dst_len = 0; } } else { //memcpy(dst, src, copy_len); pkt_copy(src, dst, (int)copy_len); } slot->len = dst_len; slot->flags = (cnt << 8)| NS_MOREFRAG; j = nm_next(j, lim); needed--; ft_p++; } while (ft_p != ft_end); slot->flags = (cnt << 8); /* clear flag on last entry */ } /* are we done ? */ if (next == NM_FT_NULL && brd_next == NM_FT_NULL) break; } { /* current position */ uint32_t *p = kring->nkr_leases; /* shorthand */ uint32_t update_pos; int still_locked = 1; mtx_lock(&kring->q_lock); if (unlikely(howmany > 0)) { /* not used all bufs. If i am the last one * i can recover the slots, otherwise must * fill them with 0 to mark empty packets. */ ND("leftover %d bufs", howmany); if (nm_next(lease_idx, lim) == kring->nkr_lease_idx) { /* yes i am the last one */ ND("roll back nkr_hwlease to %d", j); kring->nkr_hwlease = j; } else { while (howmany-- > 0) { ring->slot[j].len = 0; ring->slot[j].flags = 0; j = nm_next(j, lim); } } } p[lease_idx] = j; /* report I am done */ update_pos = kring->nr_hwtail; if (my_start == update_pos) { /* all slots before my_start have been reported, * so scan subsequent leases to see if other ranges * have been completed, and to a selwakeup or txsync. */ while (lease_idx != kring->nkr_lease_idx && p[lease_idx] != NR_NOSLOT) { j = p[lease_idx]; p[lease_idx] = NR_NOSLOT; lease_idx = nm_next(lease_idx, lim); } /* j is the new 'write' position. j != my_start * means there are new buffers to report */ if (likely(j != my_start)) { kring->nr_hwtail = j; still_locked = 0; mtx_unlock(&kring->q_lock); kring->nm_notify(kring, 0); /* this is netmap_notify for VALE ports and * netmap_bwrap_notify for bwrap. The latter will * trigger a txsync on the underlying hwna */ if (dst_na->retry && retry--) { /* XXX this is going to call nm_notify again. * Only useful for bwrap in virtual machines */ goto retry; } } } if (still_locked) mtx_unlock(&kring->q_lock); } cleanup: d->bq_head = d->bq_tail = NM_FT_NULL; /* cleanup */ d->bq_len = 0; } brddst->bq_head = brddst->bq_tail = NM_FT_NULL; /* cleanup */ brddst->bq_len = 0; return 0; } /* nm_txsync callback for VALE ports */ static int netmap_vp_txsync(struct netmap_kring *kring, int flags) { struct netmap_vp_adapter *na = (struct netmap_vp_adapter *)kring->na; u_int done; u_int const lim = kring->nkr_num_slots - 1; u_int const head = kring->rhead; if (bridge_batch <= 0) { /* testing only */ done = head; // used all goto done; } if (!na->na_bdg) { done = head; goto done; } if (bridge_batch > NM_BDG_BATCH) bridge_batch = NM_BDG_BATCH; done = nm_bdg_preflush(kring, head); done: if (done != head) D("early break at %d/ %d, tail %d", done, head, kring->nr_hwtail); /* * packets between 'done' and 'cur' are left unsent. */ kring->nr_hwcur = done; kring->nr_hwtail = nm_prev(done, lim); if (netmap_verbose) D("%s ring %d flags %d", na->up.name, kring->ring_id, flags); return 0; } /* rxsync code used by VALE ports nm_rxsync callback and also * internally by the brwap */ static int netmap_vp_rxsync_locked(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; struct netmap_ring *ring = kring->ring; u_int nm_i, lim = kring->nkr_num_slots - 1; u_int head = kring->rhead; int n; if (head > lim) { D("ouch dangerous reset!!!"); n = netmap_ring_reinit(kring); goto done; } /* First part, import newly received packets. */ /* actually nothing to do here, they are already in the kring */ /* Second part, skip past packets that userspace has released. */ nm_i = kring->nr_hwcur; if (nm_i != head) { /* consistency check, but nothing really important here */ for (n = 0; likely(nm_i != head); n++) { struct netmap_slot *slot = &ring->slot[nm_i]; void *addr = NMB(na, slot); if (addr == NETMAP_BUF_BASE(kring->na)) { /* bad buf */ D("bad buffer index %d, ignore ?", slot->buf_idx); } slot->flags &= ~NS_BUF_CHANGED; nm_i = nm_next(nm_i, lim); } kring->nr_hwcur = head; } n = 0; done: return n; } /* * nm_rxsync callback for VALE ports * user process reading from a VALE switch. * Already protected against concurrent calls from userspace, * but we must acquire the queue's lock to protect against * writers on the same queue. */ static int netmap_vp_rxsync(struct netmap_kring *kring, int flags) { int n; mtx_lock(&kring->q_lock); n = netmap_vp_rxsync_locked(kring, flags); mtx_unlock(&kring->q_lock); return n; } /* nm_bdg_attach callback for VALE ports * The na_vp port is this same netmap_adapter. There is no host port. */ static int netmap_vp_bdg_attach(const char *name, struct netmap_adapter *na) { struct netmap_vp_adapter *vpna = (struct netmap_vp_adapter *)na; if (vpna->na_bdg) return EBUSY; na->na_vp = vpna; strncpy(na->name, name, sizeof(na->name)); na->na_hostvp = NULL; return 0; } /* create a netmap_vp_adapter that describes a VALE port. * Only persistent VALE ports have a non-null ifp. */ static int netmap_vp_create(struct nmreq *nmr, struct ifnet *ifp, struct netmap_vp_adapter **ret) { struct netmap_vp_adapter *vpna; struct netmap_adapter *na; int error; u_int npipes = 0; vpna = malloc(sizeof(*vpna), M_DEVBUF, M_NOWAIT | M_ZERO); if (vpna == NULL) return ENOMEM; na = &vpna->up; na->ifp = ifp; strncpy(na->name, nmr->nr_name, sizeof(na->name)); /* bound checking */ na->num_tx_rings = nmr->nr_tx_rings; nm_bound_var(&na->num_tx_rings, 1, 1, NM_BDG_MAXRINGS, NULL); nmr->nr_tx_rings = na->num_tx_rings; // write back na->num_rx_rings = nmr->nr_rx_rings; nm_bound_var(&na->num_rx_rings, 1, 1, NM_BDG_MAXRINGS, NULL); nmr->nr_rx_rings = na->num_rx_rings; // write back nm_bound_var(&nmr->nr_tx_slots, NM_BRIDGE_RINGSIZE, 1, NM_BDG_MAXSLOTS, NULL); na->num_tx_desc = nmr->nr_tx_slots; nm_bound_var(&nmr->nr_rx_slots, NM_BRIDGE_RINGSIZE, 1, NM_BDG_MAXSLOTS, NULL); /* validate number of pipes. We want at least 1, * but probably can do with some more. * So let's use 2 as default (when 0 is supplied) */ npipes = nmr->nr_arg1; nm_bound_var(&npipes, 2, 1, NM_MAXPIPES, NULL); nmr->nr_arg1 = npipes; /* write back */ /* validate extra bufs */ nm_bound_var(&nmr->nr_arg3, 0, 0, 128*NM_BDG_MAXSLOTS, NULL); na->num_rx_desc = nmr->nr_rx_slots; vpna->virt_hdr_len = 0; vpna->mfs = 1514; vpna->last_smac = ~0llu; /*if (vpna->mfs > netmap_buf_size) TODO netmap_buf_size is zero?? vpna->mfs = netmap_buf_size; */ if (netmap_verbose) D("max frame size %u", vpna->mfs); na->na_flags |= NAF_BDG_MAYSLEEP; na->nm_txsync = netmap_vp_txsync; na->nm_rxsync = netmap_vp_rxsync; na->nm_register = netmap_vp_reg; na->nm_krings_create = netmap_vp_krings_create; na->nm_krings_delete = netmap_vp_krings_delete; na->nm_dtor = netmap_vp_dtor; na->nm_mem = netmap_mem_private_new(na->name, na->num_tx_rings, na->num_tx_desc, na->num_rx_rings, na->num_rx_desc, nmr->nr_arg3, npipes, &error); if (na->nm_mem == NULL) goto err; na->nm_bdg_attach = netmap_vp_bdg_attach; /* other nmd fields are set in the common routine */ error = netmap_attach_common(na); if (error) goto err; *ret = vpna; return 0; err: if (na->nm_mem != NULL) netmap_mem_delete(na->nm_mem); free(vpna, M_DEVBUF); return error; } /* Bridge wrapper code (bwrap). * This is used to connect a non-VALE-port netmap_adapter (hwna) to a * VALE switch. * The main task is to swap the meaning of tx and rx rings to match the * expectations of the VALE switch code (see nm_bdg_flush). * * The bwrap works by interposing a netmap_bwrap_adapter between the * rest of the system and the hwna. The netmap_bwrap_adapter looks like * a netmap_vp_adapter to the rest the system, but, internally, it * translates all callbacks to what the hwna expects. * * Note that we have to intercept callbacks coming from two sides: * * - callbacks coming from the netmap module are intercepted by * passing around the netmap_bwrap_adapter instead of the hwna * * - callbacks coming from outside of the netmap module only know * about the hwna. This, however, only happens in interrupt * handlers, where only the hwna->nm_notify callback is called. * What the bwrap does is to overwrite the hwna->nm_notify callback * with its own netmap_bwrap_intr_notify. * XXX This assumes that the hwna->nm_notify callback was the * standard netmap_notify(), as it is the case for nic adapters. * Any additional action performed by hwna->nm_notify will not be * performed by netmap_bwrap_intr_notify. * * Additionally, the bwrap can optionally attach the host rings pair * of the wrapped adapter to a different port of the switch. */ static void netmap_bwrap_dtor(struct netmap_adapter *na) { struct netmap_bwrap_adapter *bna = (struct netmap_bwrap_adapter*)na; struct netmap_adapter *hwna = bna->hwna; ND("na %p", na); /* drop reference to hwna->ifp. * If we don't do this, netmap_detach_common(na) * will think it has set NA(na->ifp) to NULL */ na->ifp = NULL; /* for safety, also drop the possible reference * in the hostna */ bna->host.up.ifp = NULL; hwna->nm_mem = bna->save_nmd; hwna->na_private = NULL; hwna->na_vp = hwna->na_hostvp = NULL; hwna->na_flags &= ~NAF_BUSY; netmap_adapter_put(hwna); } /* * Intr callback for NICs connected to a bridge. * Simply ignore tx interrupts (maybe we could try to recover space ?) * and pass received packets from nic to the bridge. * * XXX TODO check locking: this is called from the interrupt * handler so we should make sure that the interface is not * disconnected while passing down an interrupt. * * Note, no user process can access this NIC or the host stack. * The only part of the ring that is significant are the slots, * and head/cur/tail are set from the kring as needed * (part as a receive ring, part as a transmit ring). * * callback that overwrites the hwna notify callback. * Packets come from the outside or from the host stack and are put on an hwna rx ring. * The bridge wrapper then sends the packets through the bridge. */ static int netmap_bwrap_intr_notify(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; struct netmap_bwrap_adapter *bna = na->na_private; struct netmap_kring *bkring; - struct netmap_ring *ring; struct netmap_vp_adapter *vpna = &bna->up; u_int ring_nr = kring->ring_id; int error = 0; if (netmap_verbose) D("%s %s 0x%x", na->name, kring->name, flags); if (!nm_netmap_on(na)) return 0; bkring = &vpna->up.tx_rings[ring_nr]; - ring = kring->ring; /* == kbkring->ring */ /* make sure the ring is not disabled */ if (nm_kr_tryget(kring)) return 0; if (netmap_verbose) D("%s head %d cur %d tail %d", na->name, kring->rhead, kring->rcur, kring->rtail); /* simulate a user wakeup on the rx ring * fetch packets that have arrived. */ error = kring->nm_sync(kring, 0); if (error) goto put_out; if (kring->nr_hwcur == kring->nr_hwtail && netmap_verbose) { D("how strange, interrupt with no packets on %s", na->name); goto put_out; } /* new packets are kring->rcur to kring->nr_hwtail, and the bkring * had hwcur == bkring->rhead. So advance bkring->rhead to kring->nr_hwtail * to push all packets out. */ bkring->rhead = bkring->rcur = kring->nr_hwtail; netmap_vp_txsync(bkring, flags); /* mark all buffers as released on this ring */ kring->rhead = kring->rcur = kring->rtail = kring->nr_hwtail; /* another call to actually release the buffers */ error = kring->nm_sync(kring, 0); put_out: nm_kr_put(kring); return error; } /* nm_register callback for bwrap */ static int netmap_bwrap_register(struct netmap_adapter *na, int onoff) { struct netmap_bwrap_adapter *bna = (struct netmap_bwrap_adapter *)na; struct netmap_adapter *hwna = bna->hwna; struct netmap_vp_adapter *hostna = &bna->host; int error; enum txrx t; ND("%s %s", na->name, onoff ? "on" : "off"); if (onoff) { int i; /* netmap_do_regif has been called on the bwrap na. * We need to pass the information about the * memory allocator down to the hwna before * putting it in netmap mode */ hwna->na_lut = na->na_lut; if (hostna->na_bdg) { /* if the host rings have been attached to switch, * we need to copy the memory allocator information * in the hostna also */ hostna->up.na_lut = na->na_lut; } /* cross-link the netmap rings * The original number of rings comes from hwna, * rx rings on one side equals tx rings on the other. * We need to do this now, after the initialization * of the kring->ring pointers */ for_rx_tx(t) { enum txrx r= nm_txrx_swap(t); /* swap NR_TX <-> NR_RX */ for (i = 0; i < nma_get_nrings(na, r) + 1; i++) { NMR(hwna, t)[i].nkr_num_slots = NMR(na, r)[i].nkr_num_slots; NMR(hwna, t)[i].ring = NMR(na, r)[i].ring; } } } /* forward the request to the hwna */ error = hwna->nm_register(hwna, onoff); if (error) return error; /* impersonate a netmap_vp_adapter */ netmap_vp_reg(na, onoff); if (hostna->na_bdg) netmap_vp_reg(&hostna->up, onoff); if (onoff) { u_int i; /* intercept the hwna nm_nofify callback on the hw rings */ for (i = 0; i < hwna->num_rx_rings; i++) { hwna->rx_rings[i].save_notify = hwna->rx_rings[i].nm_notify; hwna->rx_rings[i].nm_notify = netmap_bwrap_intr_notify; } i = hwna->num_rx_rings; /* for safety */ /* save the host ring notify unconditionally */ hwna->rx_rings[i].save_notify = hwna->rx_rings[i].nm_notify; if (hostna->na_bdg) { /* also intercept the host ring notify */ hwna->rx_rings[i].nm_notify = netmap_bwrap_intr_notify; } } else { u_int i; /* reset all notify callbacks (including host ring) */ for (i = 0; i <= hwna->num_rx_rings; i++) { hwna->rx_rings[i].nm_notify = hwna->rx_rings[i].save_notify; hwna->rx_rings[i].save_notify = NULL; } hwna->na_lut.lut = NULL; hwna->na_lut.objtotal = 0; hwna->na_lut.objsize = 0; } return 0; } /* nm_config callback for bwrap */ static int netmap_bwrap_config(struct netmap_adapter *na, u_int *txr, u_int *txd, u_int *rxr, u_int *rxd) { struct netmap_bwrap_adapter *bna = (struct netmap_bwrap_adapter *)na; struct netmap_adapter *hwna = bna->hwna; /* forward the request */ netmap_update_config(hwna); /* swap the results */ *txr = hwna->num_rx_rings; *txd = hwna->num_rx_desc; *rxr = hwna->num_tx_rings; *rxd = hwna->num_rx_desc; return 0; } /* nm_krings_create callback for bwrap */ static int netmap_bwrap_krings_create(struct netmap_adapter *na) { struct netmap_bwrap_adapter *bna = (struct netmap_bwrap_adapter *)na; struct netmap_adapter *hwna = bna->hwna; struct netmap_adapter *hostna = &bna->host.up; int error; ND("%s", na->name); /* impersonate a netmap_vp_adapter */ error = netmap_vp_krings_create(na); if (error) return error; /* also create the hwna krings */ error = hwna->nm_krings_create(hwna); if (error) { netmap_vp_krings_delete(na); return error; } /* the connection between the bwrap krings and the hwna krings * will be perfomed later, in the nm_register callback, since * now the kring->ring pointers have not been initialized yet */ if (na->na_flags & NAF_HOST_RINGS) { /* the hostna rings are the host rings of the bwrap. * The corresponding krings must point back to the * hostna */ hostna->tx_rings = &na->tx_rings[na->num_tx_rings]; hostna->tx_rings[0].na = hostna; hostna->rx_rings = &na->rx_rings[na->num_rx_rings]; hostna->rx_rings[0].na = hostna; } return 0; } static void netmap_bwrap_krings_delete(struct netmap_adapter *na) { struct netmap_bwrap_adapter *bna = (struct netmap_bwrap_adapter *)na; struct netmap_adapter *hwna = bna->hwna; ND("%s", na->name); hwna->nm_krings_delete(hwna); netmap_vp_krings_delete(na); } /* notify method for the bridge-->hwna direction */ static int netmap_bwrap_notify(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; struct netmap_bwrap_adapter *bna = na->na_private; struct netmap_adapter *hwna = bna->hwna; u_int ring_n = kring->ring_id; u_int lim = kring->nkr_num_slots - 1; struct netmap_kring *hw_kring; int error = 0; ND("%s: na %s hwna %s", (kring ? kring->name : "NULL!"), (na ? na->name : "NULL!"), (hwna ? hwna->name : "NULL!")); hw_kring = &hwna->tx_rings[ring_n]; if (nm_kr_tryget(hw_kring)) return 0; if (!nm_netmap_on(hwna)) return 0; /* first step: simulate a user wakeup on the rx ring */ netmap_vp_rxsync(kring, flags); ND("%s[%d] PRE rx(c%3d t%3d l%3d) ring(h%3d c%3d t%3d) tx(c%3d ht%3d t%3d)", na->name, ring_n, kring->nr_hwcur, kring->nr_hwtail, kring->nkr_hwlease, ring->head, ring->cur, ring->tail, hw_kring->nr_hwcur, hw_kring->nr_hwtail, hw_ring->rtail); /* second step: the new packets are sent on the tx ring * (which is actually the same ring) */ hw_kring->rhead = hw_kring->rcur = kring->nr_hwtail; error = hw_kring->nm_sync(hw_kring, flags); if (error) goto out; /* third step: now we are back the rx ring */ /* claim ownership on all hw owned bufs */ kring->rhead = kring->rcur = nm_next(hw_kring->nr_hwtail, lim); /* skip past reserved slot */ /* fourth step: the user goes to sleep again, causing another rxsync */ netmap_vp_rxsync(kring, flags); ND("%s[%d] PST rx(c%3d t%3d l%3d) ring(h%3d c%3d t%3d) tx(c%3d ht%3d t%3d)", na->name, ring_n, kring->nr_hwcur, kring->nr_hwtail, kring->nkr_hwlease, ring->head, ring->cur, ring->tail, hw_kring->nr_hwcur, hw_kring->nr_hwtail, hw_kring->rtail); out: nm_kr_put(hw_kring); return error; } /* nm_bdg_ctl callback for the bwrap. * Called on bridge-attach and detach, as an effect of vale-ctl -[ahd]. * On attach, it needs to provide a fake netmap_priv_d structure and * perform a netmap_do_regif() on the bwrap. This will put both the * bwrap and the hwna in netmap mode, with the netmap rings shared * and cross linked. Moroever, it will start intercepting interrupts * directed to hwna. */ static int netmap_bwrap_bdg_ctl(struct netmap_adapter *na, struct nmreq *nmr, int attach) { struct netmap_priv_d *npriv; struct netmap_bwrap_adapter *bna = (struct netmap_bwrap_adapter*)na; int error = 0; if (attach) { if (NETMAP_OWNED_BY_ANY(na)) { return EBUSY; } if (bna->na_kpriv) { /* nothing to do */ return 0; } npriv = malloc(sizeof(*npriv), M_DEVBUF, M_NOWAIT|M_ZERO); if (npriv == NULL) return ENOMEM; error = netmap_do_regif(npriv, na, nmr->nr_ringid, nmr->nr_flags); if (error) { bzero(npriv, sizeof(*npriv)); free(npriv, M_DEVBUF); return error; } bna->na_kpriv = npriv; na->na_flags |= NAF_BUSY; } else { int last_instance; if (na->active_fds == 0) /* not registered */ return EINVAL; last_instance = netmap_dtor_locked(bna->na_kpriv); if (!last_instance) { D("--- error, trying to detach an entry with active mmaps"); error = EINVAL; } else { struct nm_bridge *b = bna->up.na_bdg, *bh = bna->host.na_bdg; npriv = bna->na_kpriv; bna->na_kpriv = NULL; D("deleting priv"); bzero(npriv, sizeof(*npriv)); free(npriv, M_DEVBUF); if (b) { /* XXX the bwrap dtor should take care * of this (2014-06-16) */ netmap_bdg_detach_common(b, bna->up.bdg_port, (bh ? bna->host.bdg_port : -1)); } na->na_flags &= ~NAF_BUSY; } } return error; } /* attach a bridge wrapper to the 'real' device */ int netmap_bwrap_attach(const char *nr_name, struct netmap_adapter *hwna) { struct netmap_bwrap_adapter *bna; struct netmap_adapter *na = NULL; struct netmap_adapter *hostna = NULL; int error = 0; enum txrx t; /* make sure the NIC is not already in use */ if (NETMAP_OWNED_BY_ANY(hwna)) { D("NIC %s busy, cannot attach to bridge", hwna->name); return EBUSY; } bna = malloc(sizeof(*bna), M_DEVBUF, M_NOWAIT | M_ZERO); if (bna == NULL) { return ENOMEM; } na = &bna->up.up; na->na_private = bna; strncpy(na->name, nr_name, sizeof(na->name)); /* fill the ring data for the bwrap adapter with rx/tx meanings * swapped. The real cross-linking will be done during register, * when all the krings will have been created. */ for_rx_tx(t) { enum txrx r = nm_txrx_swap(t); /* swap NR_TX <-> NR_RX */ nma_set_nrings(na, t, nma_get_nrings(hwna, r)); nma_set_ndesc(na, t, nma_get_ndesc(hwna, r)); } na->nm_dtor = netmap_bwrap_dtor; na->nm_register = netmap_bwrap_register; // na->nm_txsync = netmap_bwrap_txsync; // na->nm_rxsync = netmap_bwrap_rxsync; na->nm_config = netmap_bwrap_config; na->nm_krings_create = netmap_bwrap_krings_create; na->nm_krings_delete = netmap_bwrap_krings_delete; na->nm_notify = netmap_bwrap_notify; na->nm_bdg_ctl = netmap_bwrap_bdg_ctl; na->pdev = hwna->pdev; na->nm_mem = netmap_mem_private_new(na->name, na->num_tx_rings, na->num_tx_desc, na->num_rx_rings, na->num_rx_desc, 0, 0, &error); na->na_flags |= NAF_MEM_OWNER; if (na->nm_mem == NULL) goto err_put; bna->up.retry = 1; /* XXX maybe this should depend on the hwna */ bna->hwna = hwna; netmap_adapter_get(hwna); hwna->na_private = bna; /* weak reference */ hwna->na_vp = &bna->up; if (hwna->na_flags & NAF_HOST_RINGS) { if (hwna->na_flags & NAF_SW_ONLY) na->na_flags |= NAF_SW_ONLY; na->na_flags |= NAF_HOST_RINGS; hostna = &bna->host.up; snprintf(hostna->name, sizeof(hostna->name), "%s^", nr_name); hostna->ifp = hwna->ifp; for_rx_tx(t) { enum txrx r = nm_txrx_swap(t); nma_set_nrings(hostna, t, 1); nma_set_ndesc(hostna, t, nma_get_ndesc(hwna, r)); } // hostna->nm_txsync = netmap_bwrap_host_txsync; // hostna->nm_rxsync = netmap_bwrap_host_rxsync; hostna->nm_notify = netmap_bwrap_notify; hostna->nm_mem = na->nm_mem; hostna->na_private = bna; hostna->na_vp = &bna->up; na->na_hostvp = hwna->na_hostvp = hostna->na_hostvp = &bna->host; hostna->na_flags = NAF_BUSY; /* prevent NIOCREGIF */ } ND("%s<->%s txr %d txd %d rxr %d rxd %d", na->name, ifp->if_xname, na->num_tx_rings, na->num_tx_desc, na->num_rx_rings, na->num_rx_desc); error = netmap_attach_common(na); if (error) { goto err_free; } /* make bwrap ifp point to the real ifp * NOTE: netmap_attach_common() interprets a non-NULL na->ifp * as a request to make the ifp point to the na. Since we * do not want to change the na already pointed to by hwna->ifp, * the following assignment has to be delayed until now */ na->ifp = hwna->ifp; hwna->na_flags |= NAF_BUSY; /* make hwna point to the allocator we are actually using, * so that monitors will be able to find it */ bna->save_nmd = hwna->nm_mem; hwna->nm_mem = na->nm_mem; return 0; err_free: netmap_mem_delete(na->nm_mem); err_put: hwna->na_vp = hwna->na_hostvp = NULL; netmap_adapter_put(hwna); free(bna, M_DEVBUF); return error; } struct nm_bridge * netmap_init_bridges2(u_int n) { int i; struct nm_bridge *b; b = malloc(sizeof(struct nm_bridge) * n, M_DEVBUF, M_NOWAIT | M_ZERO); if (b == NULL) return NULL; for (i = 0; i < n; i++) BDG_RWINIT(&b[i]); return b; } void netmap_uninit_bridges2(struct nm_bridge *b, u_int n) { int i; if (b == NULL) return; for (i = 0; i < n; i++) BDG_RWDESTROY(&b[i]); free(b, M_DEVBUF); } int netmap_init_bridges(void) { #ifdef CONFIG_NET_NS return netmap_bns_register(); #else nm_bridges = netmap_init_bridges2(NM_BRIDGES); if (nm_bridges == NULL) return ENOMEM; return 0; #endif } void netmap_uninit_bridges(void) { #ifdef CONFIG_NET_NS netmap_bns_unregister(); #else netmap_uninit_bridges2(nm_bridges, NM_BRIDGES); #endif } #endif /* WITH_VALE */