2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (C) 2013-2016 Vincenzo Maffione
5 * Copyright (C) 2013-2016 Luigi Rizzo
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * This module implements netmap support on top of standard,
32 * unmodified device drivers.
34 * A NIOCREGIF request is handled here if the device does not
35 * have native support. TX and RX rings are emulated as follows:
38 * We preallocate a block of TX mbufs (roughly as many as
39 * tx descriptors; the number is not critical) to speed up
40 * operation during transmissions. The refcount on most of
41 * these buffers is artificially bumped up so we can recycle
42 * them more easily. Also, the destructor is intercepted
43 * so we use it as an interrupt notification to wake up
44 * processes blocked on a poll().
46 * For each receive ring we allocate one "struct mbq"
47 * (an mbuf tailq plus a spinlock). We intercept packets
49 * on the receive path and put them in the mbq from which
50 * netmap receive routines can grab them.
53 * in the generic_txsync() routine, netmap buffers are copied
54 * (or linked, in a future) to the preallocated mbufs
55 * and pushed to the transmit queue. Some of these mbufs
56 * (those with NS_REPORT, or otherwise every half ring)
57 * have the refcount=1, others have refcount=2.
58 * When the destructor is invoked, we take that as
59 * a notification that all mbufs up to that one in
60 * the specific ring have been completed, and generate
61 * the equivalent of a transmit interrupt.
69 #include <sys/cdefs.h> /* prerequisite */
70 __FBSDID("$FreeBSD$");
72 #include <sys/types.h>
73 #include <sys/errno.h>
74 #include <sys/malloc.h>
75 #include <sys/lock.h> /* PROT_EXEC */
76 #include <sys/rwlock.h>
77 #include <sys/socket.h> /* sockaddrs */
78 #include <sys/selinfo.h>
80 #include <net/if_types.h>
81 #include <net/if_var.h>
82 #include <machine/bus.h> /* bus_dmamap_* in netmap_kern.h */
84 #include <net/netmap.h>
85 #include <dev/netmap/netmap_kern.h>
86 #include <dev/netmap/netmap_mem2.h>
88 #define MBUF_RXQ(m) ((m)->m_pkthdr.flowid)
95 #define MBUF_TXQ(m) 0//((m)->m_pkthdr.flowid)
96 #define MBUF_RXQ(m) 0//((m)->m_pkthdr.flowid)
97 #define smp_mb() //XXX: to be correctly defined
101 #include "bsd_glue.h"
103 #include <linux/ethtool.h> /* struct ethtool_ops, get_ringparam */
104 #include <linux/hrtimer.h>
106 static inline struct mbuf *
107 nm_os_get_mbuf(struct ifnet *ifp, int len)
109 return alloc_skb(ifp->needed_headroom + len +
110 ifp->needed_tailroom, GFP_ATOMIC);
116 /* Common headers. */
117 #include <net/netmap.h>
118 #include <dev/netmap/netmap_kern.h>
119 #include <dev/netmap/netmap_mem2.h>
122 #define for_each_kring_n(_i, _k, _karr, _n) \
123 for ((_k)=*(_karr), (_i) = 0; (_i) < (_n); (_i)++, (_k) = (_karr)[(_i)])
125 #define for_each_tx_kring(_i, _k, _na) \
126 for_each_kring_n(_i, _k, (_na)->tx_rings, (_na)->num_tx_rings)
127 #define for_each_tx_kring_h(_i, _k, _na) \
128 for_each_kring_n(_i, _k, (_na)->tx_rings, (_na)->num_tx_rings + 1)
130 #define for_each_rx_kring(_i, _k, _na) \
131 for_each_kring_n(_i, _k, (_na)->rx_rings, (_na)->num_rx_rings)
132 #define for_each_rx_kring_h(_i, _k, _na) \
133 for_each_kring_n(_i, _k, (_na)->rx_rings, (_na)->num_rx_rings + 1)
136 /* ======================== PERFORMANCE STATISTICS =========================== */
142 unsigned long txsync;
144 unsigned long txrepl;
145 unsigned long txdrop;
148 unsigned long rxsync;
151 struct rate_context {
153 struct timer_list timer;
154 struct rate_stats new;
155 struct rate_stats old;
158 #define RATE_PRINTK(_NAME_) \
159 printk( #_NAME_ " = %lu Hz\n", (cur._NAME_ - ctx->old._NAME_)/RATE_PERIOD);
160 #define RATE_PERIOD 2
161 static void rate_callback(unsigned long arg)
163 struct rate_context * ctx = (struct rate_context *)arg;
164 struct rate_stats cur = ctx->new;
178 r = mod_timer(&ctx->timer, jiffies +
179 msecs_to_jiffies(RATE_PERIOD * 1000));
181 nm_prerr("mod_timer() failed");
184 static struct rate_context rate_ctx;
186 void generic_rate(int txp, int txs, int txi, int rxp, int rxs, int rxi)
188 if (txp) rate_ctx.new.txpkt++;
189 if (txs) rate_ctx.new.txsync++;
190 if (txi) rate_ctx.new.txirq++;
191 if (rxp) rate_ctx.new.rxpkt++;
192 if (rxs) rate_ctx.new.rxsync++;
193 if (rxi) rate_ctx.new.rxirq++;
201 /* ========== GENERIC (EMULATED) NETMAP ADAPTER SUPPORT ============= */
204 * Wrapper used by the generic adapter layer to notify
205 * the poller threads. Differently from netmap_rx_irq(), we check
206 * only NAF_NETMAP_ON instead of NAF_NATIVE_ON to enable the irq.
209 netmap_generic_irq(struct netmap_adapter *na, u_int q, u_int *work_done)
211 if (unlikely(!nm_netmap_on(na)))
214 netmap_common_irq(na, q, work_done);
217 rate_ctx.new.rxirq++;
219 rate_ctx.new.txirq++;
220 #endif /* RATE_GENERIC */
224 generic_netmap_unregister(struct netmap_adapter *na)
226 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
227 struct netmap_kring *kring = NULL;
230 if (na->active_fds == 0) {
231 na->na_flags &= ~NAF_NETMAP_ON;
233 /* Stop intercepting packets on the RX path. */
234 nm_os_catch_rx(gna, 0);
236 /* Release packet steering control. */
237 nm_os_catch_tx(gna, 0);
240 netmap_krings_mode_commit(na, /*onoff=*/0);
242 for_each_rx_kring(r, kring, na) {
243 /* Free the mbufs still pending in the RX queues,
244 * that did not end up into the corresponding netmap
246 mbq_safe_purge(&kring->rx_queue);
247 nm_os_mitigation_cleanup(&gna->mit[r]);
250 /* Decrement reference counter for the mbufs in the
251 * TX pools. These mbufs can be still pending in drivers,
252 * (e.g. this happens with virtio-net driver, which
253 * does lazy reclaiming of transmitted mbufs). */
254 for_each_tx_kring(r, kring, na) {
255 /* We must remove the destructor on the TX event,
256 * because the destructor invokes netmap code, and
257 * the netmap module may disappear before the
258 * TX event is consumed. */
259 mtx_lock_spin(&kring->tx_event_lock);
260 if (kring->tx_event) {
261 SET_MBUF_DESTRUCTOR(kring->tx_event, NULL);
263 kring->tx_event = NULL;
264 mtx_unlock_spin(&kring->tx_event_lock);
267 if (na->active_fds == 0) {
268 nm_os_free(gna->mit);
270 for_each_rx_kring(r, kring, na) {
271 mbq_safe_fini(&kring->rx_queue);
274 for_each_tx_kring(r, kring, na) {
275 mtx_destroy(&kring->tx_event_lock);
276 if (kring->tx_pool == NULL) {
280 for (i=0; i<na->num_tx_desc; i++) {
281 if (kring->tx_pool[i]) {
282 m_freem(kring->tx_pool[i]);
285 nm_os_free(kring->tx_pool);
286 kring->tx_pool = NULL;
290 if (--rate_ctx.refcount == 0) {
291 nm_prinf("del_timer()");
292 del_timer(&rate_ctx.timer);
295 nm_prinf("Emulated adapter for %s deactivated", na->name);
301 /* Enable/disable netmap mode for a generic network interface. */
303 generic_netmap_register(struct netmap_adapter *na, int enable)
305 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
306 struct netmap_kring *kring = NULL;
315 /* This is actually an unregif. */
316 return generic_netmap_unregister(na);
319 if (na->active_fds == 0) {
320 nm_prinf("Emulated adapter for %s activated", na->name);
321 /* Do all memory allocations when (na->active_fds == 0), to
322 * simplify error management. */
324 /* Allocate memory for mitigation support on all the rx queues. */
325 gna->mit = nm_os_malloc(na->num_rx_rings * sizeof(struct nm_generic_mit));
327 nm_prerr("mitigation allocation failed");
332 for_each_rx_kring(r, kring, na) {
333 /* Init mitigation support. */
334 nm_os_mitigation_init(&gna->mit[r], r, na);
336 /* Initialize the rx queue, as generic_rx_handler() can
337 * be called as soon as nm_os_catch_rx() returns.
339 mbq_safe_init(&kring->rx_queue);
343 * Prepare mbuf pools (parallel to the tx rings), for packet
344 * transmission. Don't preallocate the mbufs here, it's simpler
345 * to leave this task to txsync.
347 for_each_tx_kring(r, kring, na) {
348 kring->tx_pool = NULL;
350 for_each_tx_kring(r, kring, na) {
352 nm_os_malloc(na->num_tx_desc * sizeof(struct mbuf *));
353 if (!kring->tx_pool) {
354 nm_prerr("tx_pool allocation failed");
358 mtx_init(&kring->tx_event_lock, "tx_event_lock",
363 netmap_krings_mode_commit(na, /*onoff=*/1);
365 for_each_tx_kring(r, kring, na) {
366 /* Initialize tx_pool and tx_event. */
367 for (i=0; i<na->num_tx_desc; i++) {
368 kring->tx_pool[i] = NULL;
371 kring->tx_event = NULL;
374 if (na->active_fds == 0) {
375 /* Prepare to intercept incoming traffic. */
376 error = nm_os_catch_rx(gna, 1);
378 nm_prerr("nm_os_catch_rx(1) failed (%d)", error);
382 /* Let netmap control the packet steering. */
383 error = nm_os_catch_tx(gna, 1);
385 nm_prerr("nm_os_catch_tx(1) failed (%d)", error);
389 na->na_flags |= NAF_NETMAP_ON;
392 if (rate_ctx.refcount == 0) {
393 nm_prinf("setup_timer()");
394 memset(&rate_ctx, 0, sizeof(rate_ctx));
395 setup_timer(&rate_ctx.timer, &rate_callback, (unsigned long)&rate_ctx);
396 if (mod_timer(&rate_ctx.timer, jiffies + msecs_to_jiffies(1500))) {
397 nm_prerr("Error: mod_timer()");
406 /* Here (na->active_fds == 0) holds. */
408 nm_os_catch_rx(gna, 0);
410 for_each_tx_kring(r, kring, na) {
411 mtx_destroy(&kring->tx_event_lock);
412 if (kring->tx_pool == NULL) {
415 nm_os_free(kring->tx_pool);
416 kring->tx_pool = NULL;
418 for_each_rx_kring(r, kring, na) {
419 mbq_safe_fini(&kring->rx_queue);
421 nm_os_free(gna->mit);
428 * Callback invoked when the device driver frees an mbuf used
429 * by netmap to transmit a packet. This usually happens when
430 * the NIC notifies the driver that transmission is completed.
433 generic_mbuf_destructor(struct mbuf *m)
435 struct netmap_adapter *na = NA(GEN_TX_MBUF_IFP(m));
436 struct netmap_kring *kring;
437 unsigned int r = MBUF_TXQ(m);
438 unsigned int r_orig = r;
440 if (unlikely(!nm_netmap_on(na) || r >= na->num_tx_rings)) {
441 nm_prerr("Error: no netmap adapter on device %p",
447 * First, clear the event mbuf.
448 * In principle, the event 'm' should match the one stored
449 * on ring 'r'. However we check it explicitely to stay
450 * safe against lower layers (qdisc, driver, etc.) changing
451 * MBUF_TXQ(m) under our feet. If the match is not found
452 * on 'r', we try to see if it belongs to some other ring.
457 kring = na->tx_rings[r];
458 mtx_lock_spin(&kring->tx_event_lock);
459 if (kring->tx_event == m) {
460 kring->tx_event = NULL;
463 mtx_unlock_spin(&kring->tx_event_lock);
467 nm_prlim(1, "event %p migrated: ring %u --> %u",
473 if (++r == na->num_tx_rings) r = 0;
476 nm_prlim(1, "Cannot match event %p", m);
481 /* Second, wake up clients. They will reclaim the event through
483 netmap_generic_irq(na, r, NULL);
485 #if __FreeBSD_version <= 1200050
486 void_mbuf_dtor(m, NULL, NULL);
487 #else /* __FreeBSD_version >= 1200051 */
489 #endif /* __FreeBSD_version >= 1200051 */
493 /* Record completed transmissions and update hwtail.
495 * The oldest tx buffer not yet completed is at nr_hwtail + 1,
496 * nr_hwcur is the first unsent buffer.
499 generic_netmap_tx_clean(struct netmap_kring *kring, int txqdisc)
501 u_int const lim = kring->nkr_num_slots - 1;
502 u_int nm_i = nm_next(kring->nr_hwtail, lim);
503 u_int hwcur = kring->nr_hwcur;
505 struct mbuf **tx_pool = kring->tx_pool;
507 nm_prdis("hwcur = %d, hwtail = %d", kring->nr_hwcur, kring->nr_hwtail);
509 while (nm_i != hwcur) { /* buffers not completed */
510 struct mbuf *m = tx_pool[nm_i];
514 /* Nothing to do, this is going
515 * to be replenished. */
516 nm_prlim(3, "Is this happening?");
518 } else if (MBUF_QUEUED(m)) {
519 break; /* Not dequeued yet. */
521 } else if (MBUF_REFCNT(m) != 1) {
522 /* This mbuf has been dequeued but is still busy
524 * Leave it to the driver and replenish. */
526 tx_pool[nm_i] = NULL;
530 if (unlikely(m == NULL)) {
533 /* This slot was used to place an event. */
534 mtx_lock_spin(&kring->tx_event_lock);
535 event_consumed = (kring->tx_event == NULL);
536 mtx_unlock_spin(&kring->tx_event_lock);
537 if (!event_consumed) {
538 /* The event has not been consumed yet,
539 * still busy in the driver. */
542 /* The event has been consumed, we can go
545 } else if (MBUF_REFCNT(m) != 1) {
546 /* This mbuf is still busy: its refcnt is 2. */
552 nm_i = nm_next(nm_i, lim);
554 kring->nr_hwtail = nm_prev(nm_i, lim);
555 nm_prdis("tx completed [%d] -> hwtail %d", n, kring->nr_hwtail);
560 /* Compute a slot index in the middle between inf and sup. */
562 ring_middle(u_int inf, u_int sup, u_int lim)
569 } else { /* wrap around */
570 e = (sup + n + inf) / 2;
576 if (unlikely(e >= n)) {
577 nm_prerr("This cannot happen");
585 generic_set_tx_event(struct netmap_kring *kring, u_int hwcur)
587 u_int lim = kring->nkr_num_slots - 1;
590 u_int ntc = nm_next(kring->nr_hwtail, lim); /* next to clean */
593 return; /* all buffers are free */
597 * We have pending packets in the driver between hwtail+1
598 * and hwcur, and we have to chose one of these slot to
599 * generate a notification.
600 * There is a race but this is only called within txsync which
601 * does a double check.
604 /* Choose a slot in the middle, so that we don't risk ending
605 * up in a situation where the client continuously wake up,
606 * fills one or a few TX slots and go to sleep again. */
607 e = ring_middle(ntc, hwcur, lim);
609 /* Choose the first pending slot, to be safe against driver
610 * reordering mbuf transmissions. */
614 m = kring->tx_pool[e];
616 /* An event is already in place. */
620 mtx_lock_spin(&kring->tx_event_lock);
621 if (kring->tx_event) {
622 /* An event is already in place. */
623 mtx_unlock_spin(&kring->tx_event_lock);
627 SET_MBUF_DESTRUCTOR(m, generic_mbuf_destructor);
629 mtx_unlock_spin(&kring->tx_event_lock);
631 kring->tx_pool[e] = NULL;
633 nm_prdis("Request Event at %d mbuf %p refcnt %d", e, m, m ? MBUF_REFCNT(m) : -2 );
635 /* Decrement the refcount. This will free it if we lose the race
636 * with the driver. */
643 * generic_netmap_txsync() transforms netmap buffers into mbufs
644 * and passes them to the standard device driver
645 * (ndo_start_xmit() or ifp->if_transmit() ).
646 * On linux this is not done directly, but using dev_queue_xmit(),
647 * since it implements the TX flow control (and takes some locks).
650 generic_netmap_txsync(struct netmap_kring *kring, int flags)
652 struct netmap_adapter *na = kring->na;
653 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
654 struct ifnet *ifp = na->ifp;
655 struct netmap_ring *ring = kring->ring;
656 u_int nm_i; /* index into the netmap ring */ // j
657 u_int const lim = kring->nkr_num_slots - 1;
658 u_int const head = kring->rhead;
659 u_int ring_nr = kring->ring_id;
661 IFRATE(rate_ctx.new.txsync++);
666 * First part: process new packets to send.
668 nm_i = kring->nr_hwcur;
669 if (nm_i != head) { /* we have new packets to send */
670 struct nm_os_gen_arg a;
673 if (gna->txqdisc && nm_kr_txempty(kring)) {
674 /* In txqdisc mode, we ask for a delayed notification,
675 * but only when cur == hwtail, which means that the
676 * client is going to block. */
677 event = ring_middle(nm_i, head, lim);
678 nm_prdis("Place txqdisc event (hwcur=%u,event=%u,"
679 "head=%u,hwtail=%u)", nm_i, event, head,
685 a.head = a.tail = NULL;
687 while (nm_i != head) {
688 struct netmap_slot *slot = &ring->slot[nm_i];
689 u_int len = slot->len;
690 void *addr = NMB(na, slot);
691 /* device-specific */
695 NM_CHECK_ADDR_LEN(na, addr, len);
697 /* Tale a mbuf from the tx pool (replenishing the pool
698 * entry if necessary) and copy in the user packet. */
699 m = kring->tx_pool[nm_i];
700 if (unlikely(m == NULL)) {
701 kring->tx_pool[nm_i] = m =
702 nm_os_get_mbuf(ifp, NETMAP_BUF_SIZE(na));
704 nm_prlim(2, "Failed to replenish mbuf");
705 /* Here we could schedule a timer which
706 * retries to replenish after a while,
707 * and notifies the client when it
708 * manages to replenish some slots. In
709 * any case we break early to avoid
713 IFRATE(rate_ctx.new.txrepl++);
719 a.qevent = (nm_i == event);
720 /* When not in txqdisc mode, we should ask
721 * notifications when NS_REPORT is set, or roughly
722 * every half ring. To optimize this, we set a
723 * notification event when the client runs out of
724 * TX ring space, or when transmission fails. In
725 * the latter case we also break early.
727 tx_ret = nm_os_generic_xmit_frame(&a);
728 if (unlikely(tx_ret)) {
731 * No room for this mbuf in the device driver.
732 * Request a notification FOR A PREVIOUS MBUF,
733 * then call generic_netmap_tx_clean(kring) to do the
734 * double check and see if we can free more buffers.
735 * If there is space continue, else break;
736 * NOTE: the double check is necessary if the problem
737 * occurs in the txsync call after selrecord().
738 * Also, we need some way to tell the caller that not
739 * all buffers were queued onto the device (this was
740 * not a problem with native netmap driver where space
741 * is preallocated). The bridge has a similar problem
742 * and we solve it there by dropping the excess packets.
744 generic_set_tx_event(kring, nm_i);
745 if (generic_netmap_tx_clean(kring, gna->txqdisc)) {
746 /* space now available */
753 /* In txqdisc mode, the netmap-aware qdisc
754 * queue has the same length as the number of
755 * netmap slots (N). Since tail is advanced
756 * only when packets are dequeued, qdisc
757 * queue overrun cannot happen, so
758 * nm_os_generic_xmit_frame() did not fail
760 * However, packets can be dropped because
761 * carrier is off, or because our qdisc is
762 * being deactivated, or possibly for other
763 * reasons. In these cases, we just let the
764 * packet to be dropped. */
765 IFRATE(rate_ctx.new.txdrop++);
768 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
769 nm_i = nm_next(nm_i, lim);
770 IFRATE(rate_ctx.new.txpkt++);
772 if (a.head != NULL) {
774 nm_os_generic_xmit_frame(&a);
776 /* Update hwcur to the next slot to transmit. Here nm_i
777 * is not necessarily head, we could break early. */
778 kring->nr_hwcur = nm_i;
782 * Second, reclaim completed buffers
784 if (!gna->txqdisc && (flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring))) {
785 /* No more available slots? Set a notification event
786 * on a netmap slot that will be cleaned in the future.
787 * No doublecheck is performed, since txsync() will be
788 * called twice by netmap_poll().
790 generic_set_tx_event(kring, nm_i);
793 generic_netmap_tx_clean(kring, gna->txqdisc);
800 * This handler is registered (through nm_os_catch_rx())
801 * within the attached network interface
802 * in the RX subsystem, so that every mbuf passed up by
803 * the driver can be stolen to the network stack.
804 * Stolen packets are put in a queue where the
805 * generic_netmap_rxsync() callback can extract them.
806 * Returns 1 if the packet was stolen, 0 otherwise.
809 generic_rx_handler(struct ifnet *ifp, struct mbuf *m)
811 struct netmap_adapter *na = NA(ifp);
812 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
813 struct netmap_kring *kring;
815 u_int r = MBUF_RXQ(m); /* receive ring number */
817 if (r >= na->num_rx_rings) {
818 r = r % na->num_rx_rings;
821 kring = na->rx_rings[r];
823 if (kring->nr_mode == NKR_NETMAP_OFF) {
824 /* We must not intercept this mbuf. */
828 /* limit the size of the queue */
829 if (unlikely(!gna->rxsg && MBUF_LEN(m) > NETMAP_BUF_SIZE(na))) {
830 /* This may happen when GRO/LRO features are enabled for
831 * the NIC driver when the generic adapter does not
832 * support RX scatter-gather. */
833 nm_prlim(2, "Warning: driver pushed up big packet "
834 "(size=%d)", (int)MBUF_LEN(m));
836 } else if (unlikely(mbq_len(&kring->rx_queue) > 1024)) {
839 mbq_safe_enqueue(&kring->rx_queue, m);
842 if (netmap_generic_mit < 32768) {
843 /* no rx mitigation, pass notification up */
844 netmap_generic_irq(na, r, &work_done);
846 /* same as send combining, filter notification if there is a
847 * pending timer, otherwise pass it up and start a timer.
849 if (likely(nm_os_mitigation_active(&gna->mit[r]))) {
850 /* Record that there is some pending work. */
851 gna->mit[r].mit_pending = 1;
853 netmap_generic_irq(na, r, &work_done);
854 nm_os_mitigation_start(&gna->mit[r]);
858 /* We have intercepted the mbuf. */
863 * generic_netmap_rxsync() extracts mbufs from the queue filled by
864 * generic_netmap_rx_handler() and puts their content in the netmap
866 * Access must be protected because the rx handler is asynchronous,
869 generic_netmap_rxsync(struct netmap_kring *kring, int flags)
871 struct netmap_ring *ring = kring->ring;
872 struct netmap_adapter *na = kring->na;
873 u_int nm_i; /* index into the netmap ring */ //j,
875 u_int const lim = kring->nkr_num_slots - 1;
876 u_int const head = kring->rhead;
877 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
879 /* Adapter-specific variables. */
880 u_int nm_buf_len = NETMAP_BUF_SIZE(na);
883 int avail; /* in bytes */
888 return netmap_ring_reinit(kring);
890 IFRATE(rate_ctx.new.rxsync++);
893 * First part: skip past packets that userspace has released.
894 * This can possibly make room for the second part.
896 nm_i = kring->nr_hwcur;
898 /* Userspace has released some packets. */
899 for (n = 0; nm_i != head; n++) {
900 struct netmap_slot *slot = &ring->slot[nm_i];
902 slot->flags &= ~NS_BUF_CHANGED;
903 nm_i = nm_next(nm_i, lim);
905 kring->nr_hwcur = head;
909 * Second part: import newly received packets.
911 if (!netmap_no_pendintr && !force_update) {
915 nm_i = kring->nr_hwtail; /* First empty slot in the receive ring. */
917 /* Compute the available space (in bytes) in this netmap ring.
918 * The first slot that is not considered in is the one before
921 avail = nm_prev(kring->nr_hwcur, lim) - nm_i;
926 /* First pass: While holding the lock on the RX mbuf queue,
927 * extract as many mbufs as they fit the available space,
928 * and put them in a temporary queue.
929 * To avoid performing a per-mbuf division (mlen / nm_buf_len) to
930 * to update avail, we do the update in a while loop that we
931 * also use to set the RX slots, but without performing the copy. */
933 mbq_lock(&kring->rx_queue);
935 m = mbq_peek(&kring->rx_queue);
937 /* No more packets from the driver. */
943 /* No more space in the ring. */
947 mbq_dequeue(&kring->rx_queue);
957 ring->slot[nm_i].len = copy;
958 ring->slot[nm_i].flags = (mlen ? NS_MOREFRAG : 0);
959 nm_i = nm_next(nm_i, lim);
962 mbq_enqueue(&tmpq, m);
964 mbq_unlock(&kring->rx_queue);
966 /* Second pass: Drain the temporary queue, going over the used RX slots,
967 * and perform the copy out of the RX queue lock. */
968 nm_i = kring->nr_hwtail;
975 m = mbq_dequeue(&tmpq);
981 nmaddr = NMB(na, &ring->slot[nm_i]);
982 /* We only check the address here on generic rx rings. */
983 if (nmaddr == NETMAP_BUF_BASE(na)) { /* Bad buffer */
987 return netmap_ring_reinit(kring);
990 copy = ring->slot[nm_i].len;
991 m_copydata(m, ofs, copy, nmaddr);
993 morefrag = ring->slot[nm_i].flags & NS_MOREFRAG;
994 nm_i = nm_next(nm_i, lim);
1003 kring->nr_hwtail = nm_i;
1004 IFRATE(rate_ctx.new.rxpkt += n);
1006 kring->nr_kflags &= ~NKR_PENDINTR;
1012 generic_netmap_dtor(struct netmap_adapter *na)
1014 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter*)na;
1015 struct ifnet *ifp = netmap_generic_getifp(gna);
1016 struct netmap_adapter *prev_na = gna->prev;
1018 if (prev_na != NULL) {
1019 netmap_adapter_put(prev_na);
1020 if (nm_iszombie(na)) {
1022 * The driver has been removed without releasing
1023 * the reference so we need to do it here.
1025 netmap_adapter_put(prev_na);
1027 nm_prinf("Native netmap adapter for %s restored", prev_na->name);
1029 NM_RESTORE_NA(ifp, prev_na);
1031 * netmap_detach_common(), that it's called after this function,
1032 * overrides WNA(ifp) if na->ifp is not NULL.
1035 nm_prinf("Emulated netmap adapter for %s destroyed", na->name);
1039 na_is_generic(struct netmap_adapter *na)
1041 return na->nm_register == generic_netmap_register;
1045 * generic_netmap_attach() makes it possible to use netmap on
1046 * a device without native netmap support.
1047 * This is less performant than native support but potentially
1048 * faster than raw sockets or similar schemes.
1050 * In this "emulated" mode, netmap rings do not necessarily
1051 * have the same size as those in the NIC. We use a default
1052 * value and possibly override it if the OS has ways to fetch the
1053 * actual configuration.
1056 generic_netmap_attach(struct ifnet *ifp)
1058 struct netmap_adapter *na;
1059 struct netmap_generic_adapter *gna;
1061 u_int num_tx_desc, num_rx_desc;
1064 if (ifp->if_type == IFT_LOOP) {
1065 nm_prerr("if_loop is not supported by %s", __func__);
1070 if (NM_NA_CLASH(ifp)) {
1071 /* If NA(ifp) is not null but there is no valid netmap
1072 * adapter it means that someone else is using the same
1073 * pointer (e.g. ax25_ptr on linux). This happens for
1074 * instance when also PF_RING is in use. */
1075 nm_prerr("Error: netmap adapter hook is busy");
1079 num_tx_desc = num_rx_desc = netmap_generic_ringsize; /* starting point */
1081 nm_os_generic_find_num_desc(ifp, &num_tx_desc, &num_rx_desc); /* ignore errors */
1082 if (num_tx_desc == 0 || num_rx_desc == 0) {
1083 nm_prerr("Device has no hw slots (tx %u, rx %u)", num_tx_desc, num_rx_desc);
1087 gna = nm_os_malloc(sizeof(*gna));
1089 nm_prerr("no memory on attach, give up");
1092 na = (struct netmap_adapter *)gna;
1093 strlcpy(na->name, ifp->if_xname, sizeof(na->name));
1095 na->num_tx_desc = num_tx_desc;
1096 na->num_rx_desc = num_rx_desc;
1097 na->rx_buf_maxsize = 32768;
1098 na->nm_register = &generic_netmap_register;
1099 na->nm_txsync = &generic_netmap_txsync;
1100 na->nm_rxsync = &generic_netmap_rxsync;
1101 na->nm_dtor = &generic_netmap_dtor;
1102 /* when using generic, NAF_NETMAP_ON is set so we force
1103 * NAF_SKIP_INTR to use the regular interrupt handler
1105 na->na_flags = NAF_SKIP_INTR | NAF_HOST_RINGS;
1107 nm_prdis("[GNA] num_tx_queues(%d), real_num_tx_queues(%d), len(%lu)",
1108 ifp->num_tx_queues, ifp->real_num_tx_queues,
1110 nm_prdis("[GNA] num_rx_queues(%d), real_num_rx_queues(%d)",
1111 ifp->num_rx_queues, ifp->real_num_rx_queues);
1113 nm_os_generic_find_num_queues(ifp, &na->num_tx_rings, &na->num_rx_rings);
1115 retval = netmap_attach_common(na);
1121 if (NM_NA_VALID(ifp)) {
1122 gna->prev = NA(ifp); /* save old na */
1123 netmap_adapter_get(gna->prev);
1125 NM_ATTACH_NA(ifp, na);
1127 nm_os_generic_set_features(gna);
1129 nm_prinf("Emulated adapter for %s created (prev was %s)", na->name,
1130 gna->prev ? gna->prev->name : "NULL");