2 * Copyright (C) 2013 Universita` di Pisa. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * This module implements netmap support on top of standard,
28 * unmodified device drivers.
30 * A NIOCREGIF request is handled here if the device does not
31 * have native support. TX and RX rings are emulated as follows:
34 * We preallocate a block of TX mbufs (roughly as many as
35 * tx descriptors; the number is not critical) to speed up
36 * operation during transmissions. The refcount on most of
37 * these buffers is artificially bumped up so we can recycle
38 * them more easily. Also, the destructor is intercepted
39 * so we use it as an interrupt notification to wake up
40 * processes blocked on a poll().
42 * For each receive ring we allocate one "struct mbq"
43 * (an mbuf tailq plus a spinlock). We intercept packets
45 * on the receive path and put them in the mbq from which
46 * netmap receive routines can grab them.
49 * in the generic_txsync() routine, netmap buffers are copied
50 * (or linked, in a future) to the preallocated mbufs
51 * and pushed to the transmit queue. Some of these mbufs
52 * (those with NS_REPORT, or otherwise every half ring)
53 * have the refcount=1, others have refcount=2.
54 * When the destructor is invoked, we take that as
55 * a notification that all mbufs up to that one in
56 * the specific ring have been completed, and generate
57 * the equivalent of a transmit interrupt.
65 #include <sys/cdefs.h> /* prerequisite */
66 __FBSDID("$FreeBSD$");
68 #include <sys/types.h>
69 #include <sys/errno.h>
70 #include <sys/malloc.h>
71 #include <sys/lock.h> /* PROT_EXEC */
72 #include <sys/rwlock.h>
73 #include <sys/socket.h> /* sockaddrs */
74 #include <sys/selinfo.h>
76 #include <net/if_var.h>
77 #include <machine/bus.h> /* bus_dmamap_* in netmap_kern.h */
79 // XXX temporary - D() defined here
80 #include <net/netmap.h>
81 #include <dev/netmap/netmap_kern.h>
82 #include <dev/netmap/netmap_mem2.h>
84 #define rtnl_lock() D("rtnl_lock called");
85 #define rtnl_unlock() D("rtnl_lock called");
86 #define MBUF_TXQ(m) ((m)->m_pkthdr.flowid)
94 * we allocate an EXT_PACKET
96 #define netmap_get_mbuf(len) m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR|M_NOFREE)
98 /* mbuf destructor, also need to change the type to EXT_EXTREF,
99 * add an M_NOFREE flag, and then clear the flag and
100 * chain into uma_zfree(zone_pack, mf)
101 * (or reinstall the buffer ?)
103 #define SET_MBUF_DESTRUCTOR(m, fn) do { \
104 (m)->m_ext.ext_free = (void *)fn; \
105 (m)->m_ext.ext_type = EXT_EXTREF; \
109 #define GET_MBUF_REFCNT(m) ((m)->m_ext.ref_cnt ? *(m)->m_ext.ref_cnt : -1)
115 #include "bsd_glue.h"
117 #include <linux/rtnetlink.h> /* rtnl_[un]lock() */
118 #include <linux/ethtool.h> /* struct ethtool_ops, get_ringparam */
119 #include <linux/hrtimer.h>
121 //#define RATE /* Enables communication statistics. */
128 /* Common headers. */
129 #include <net/netmap.h>
130 #include <dev/netmap/netmap_kern.h>
131 #include <dev/netmap/netmap_mem2.h>
135 /* ======================== usage stats =========================== */
141 unsigned long txsync;
145 unsigned long rxsync;
148 struct rate_context {
150 struct timer_list timer;
151 struct rate_stats new;
152 struct rate_stats old;
155 #define RATE_PRINTK(_NAME_) \
156 printk( #_NAME_ " = %lu Hz\n", (cur._NAME_ - ctx->old._NAME_)/RATE_PERIOD);
157 #define RATE_PERIOD 2
158 static void rate_callback(unsigned long arg)
160 struct rate_context * ctx = (struct rate_context *)arg;
161 struct rate_stats cur = ctx->new;
173 r = mod_timer(&ctx->timer, jiffies +
174 msecs_to_jiffies(RATE_PERIOD * 1000));
176 D("[v1000] Error: mod_timer()");
179 static struct rate_context rate_ctx;
186 /* =============== GENERIC NETMAP ADAPTER SUPPORT ================= */
187 #define GENERIC_BUF_SIZE netmap_buf_size /* Size of the mbufs in the Tx pool. */
190 * Wrapper used by the generic adapter layer to notify
191 * the poller threads. Differently from netmap_rx_irq(), we check
192 * only IFCAP_NETMAP instead of NAF_NATIVE_ON to enable the irq.
195 netmap_generic_irq(struct ifnet *ifp, u_int q, u_int *work_done)
197 if (unlikely(!(ifp->if_capenable & IFCAP_NETMAP)))
200 netmap_common_irq(ifp, q, work_done);
204 /* Enable/disable netmap mode for a generic network interface. */
205 int generic_netmap_register(struct netmap_adapter *na, int enable)
207 struct ifnet *ifp = na->ifp;
208 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
217 error = ifp->netdev_ops->ndo_stop(ifp);
221 #endif /* REG_RESET */
223 if (enable) { /* Enable netmap mode. */
224 /* Initialize the rx queue, as generic_rx_handler() can
225 * be called as soon as netmap_catch_rx() returns.
227 for (r=0; r<na->num_rx_rings; r++) {
228 mbq_safe_init(&na->rx_rings[r].rx_queue);
229 na->rx_rings[r].nr_ntc = 0;
232 /* Init the mitigation timer. */
233 netmap_mitigation_init(gna);
236 * Preallocate packet buffers for the tx rings.
238 for (r=0; r<na->num_tx_rings; r++) {
239 na->tx_rings[r].nr_ntc = 0;
240 na->tx_rings[r].tx_pool = malloc(na->num_tx_desc * sizeof(struct mbuf *),
241 M_DEVBUF, M_NOWAIT | M_ZERO);
242 if (!na->tx_rings[r].tx_pool) {
243 D("tx_pool allocation failed");
247 for (i=0; i<na->num_tx_desc; i++) {
248 m = netmap_get_mbuf(GENERIC_BUF_SIZE);
250 D("tx_pool[%d] allocation failed", i);
254 na->tx_rings[r].tx_pool[i] = m;
258 /* Prepare to intercept incoming traffic. */
259 error = netmap_catch_rx(na, 1);
261 D("netdev_rx_handler_register() failed");
262 goto register_handler;
264 ifp->if_capenable |= IFCAP_NETMAP;
266 /* Make netmap control the packet steering. */
267 netmap_catch_packet_steering(gna, 1);
272 if (rate_ctx.refcount == 0) {
274 memset(&rate_ctx, 0, sizeof(rate_ctx));
275 setup_timer(&rate_ctx.timer, &rate_callback, (unsigned long)&rate_ctx);
276 if (mod_timer(&rate_ctx.timer, jiffies + msecs_to_jiffies(1500))) {
277 D("Error: mod_timer()");
283 } else { /* Disable netmap mode. */
286 ifp->if_capenable &= ~IFCAP_NETMAP;
288 /* Release packet steering control. */
289 netmap_catch_packet_steering(gna, 0);
291 /* Do not intercept packets on the rx path. */
292 netmap_catch_rx(na, 0);
296 /* Free the mbufs going to the netmap rings */
297 for (r=0; r<na->num_rx_rings; r++) {
298 mbq_safe_purge(&na->rx_rings[r].rx_queue);
299 mbq_safe_destroy(&na->rx_rings[r].rx_queue);
302 netmap_mitigation_cleanup(gna);
304 for (r=0; r<na->num_tx_rings; r++) {
305 for (i=0; i<na->num_tx_desc; i++) {
306 m_freem(na->tx_rings[r].tx_pool[i]);
308 free(na->tx_rings[r].tx_pool, M_DEVBUF);
312 if (--rate_ctx.refcount == 0) {
314 del_timer(&rate_ctx.timer);
320 error = ifp->netdev_ops->ndo_open(ifp);
332 i = na->num_tx_desc; /* Useless, but just to stay safe. */
337 m_freem(na->tx_rings[r].tx_pool[i]);
339 free(na->tx_rings[r].tx_pool, M_DEVBUF);
340 i = na->num_tx_desc - 1;
347 * Callback invoked when the device driver frees an mbuf used
348 * by netmap to transmit a packet. This usually happens when
349 * the NIC notifies the driver that transmission is completed.
352 generic_mbuf_destructor(struct mbuf *m)
355 D("Tx irq (%p) queue %d", m, MBUF_TXQ(m));
356 netmap_generic_irq(MBUF_IFP(m), MBUF_TXQ(m), NULL);
358 m->m_ext.ext_type = EXT_PACKET;
359 m->m_ext.ext_free = NULL;
360 if (*(m->m_ext.ref_cnt) == 0)
361 *(m->m_ext.ref_cnt) = 1;
362 uma_zfree(zone_pack, m);
363 #endif /* __FreeBSD__ */
364 IFRATE(rate_ctx.new.txirq++);
367 /* Record completed transmissions and update hwavail.
369 * nr_ntc is the oldest tx buffer not yet completed
370 * (same as nr_hwavail + nr_hwcur + 1),
371 * nr_hwcur is the first unsent buffer.
372 * When cleaning, we try to recover buffers between nr_ntc and nr_hwcur.
375 generic_netmap_tx_clean(struct netmap_kring *kring)
377 u_int num_slots = kring->nkr_num_slots;
378 u_int ntc = kring->nr_ntc;
379 u_int hwcur = kring->nr_hwcur;
381 struct mbuf **tx_pool = kring->tx_pool;
383 while (ntc != hwcur) { /* buffers not completed */
384 struct mbuf *m = tx_pool[ntc];
386 if (unlikely(m == NULL)) {
387 /* try to replenish the entry */
388 tx_pool[ntc] = m = netmap_get_mbuf(GENERIC_BUF_SIZE);
389 if (unlikely(m == NULL)) {
390 D("mbuf allocation failed, XXX error");
391 // XXX how do we proceed ? break ?
394 } else if (GET_MBUF_REFCNT(m) != 1) {
395 break; /* This mbuf is still busy: its refcnt is 2. */
397 if (unlikely(++ntc == num_slots)) {
403 kring->nr_hwavail += n;
404 ND("tx completed [%d] -> hwavail %d", n, kring->nr_hwavail);
411 * We have pending packets in the driver between nr_ntc and j.
412 * Compute a position in the middle, to be used to generate
416 generic_tx_event_middle(struct netmap_kring *kring, u_int hwcur)
418 u_int n = kring->nkr_num_slots;
419 u_int ntc = kring->nr_ntc;
423 e = (hwcur + ntc) / 2;
424 } else { /* wrap around */
425 e = (hwcur + n + ntc) / 2;
431 if (unlikely(e >= n)) {
432 D("This cannot happen");
440 * We have pending packets in the driver between nr_ntc and hwcur.
441 * Schedule a notification approximately in the middle of the two.
442 * There is a race but this is only called within txsync which does
446 generic_set_tx_event(struct netmap_kring *kring, u_int hwcur)
451 if (kring->nr_ntc == hwcur) {
454 e = generic_tx_event_middle(kring, hwcur);
456 m = kring->tx_pool[e];
458 /* This can happen if there is already an event on the netmap
459 slot 'e': There is nothing to do. */
462 ND("Event at %d mbuf %p refcnt %d", e, m, GET_MBUF_REFCNT(m));
463 kring->tx_pool[e] = NULL;
464 SET_MBUF_DESTRUCTOR(m, generic_mbuf_destructor);
467 /* Decrement the refcount an free it if we have the last one. */
474 * generic_netmap_txsync() transforms netmap buffers into mbufs
475 * and passes them to the standard device driver
476 * (ndo_start_xmit() or ifp->if_transmit() ).
477 * On linux this is not done directly, but using dev_queue_xmit(),
478 * since it implements the TX flow control (and takes some locks).
481 generic_netmap_txsync(struct netmap_adapter *na, u_int ring_nr, int flags)
483 struct ifnet *ifp = na->ifp;
484 struct netmap_kring *kring = &na->tx_rings[ring_nr];
485 struct netmap_ring *ring = kring->ring;
486 u_int j, k, num_slots = kring->nkr_num_slots;
489 IFRATE(rate_ctx.new.txsync++);
491 // TODO: handle the case of mbuf allocation failure
492 /* first, reclaim completed buffers */
493 generic_netmap_tx_clean(kring);
495 /* Take a copy of ring->cur now, and never read it again. */
497 if (unlikely(k >= num_slots)) {
498 return netmap_ring_reinit(kring);
504 * 'new_slots' counts how many new slots have been added:
505 * everything from hwcur to cur, excluding reserved ones, if any.
506 * nr_hwreserved start from hwcur and counts how many slots were
507 * not sent to the NIC from the previous round.
509 new_slots = k - j - kring->nr_hwreserved;
511 new_slots += num_slots;
515 /* Process new packets to send:
516 * j is the current index in the netmap ring.
519 struct netmap_slot *slot = &ring->slot[j]; /* Current slot in the netmap ring */
520 void *addr = NMB(slot);
521 u_int len = slot->len;
525 if (unlikely(addr == netmap_buffer_base || len > NETMAP_BUF_SIZE)) {
526 return netmap_ring_reinit(kring);
528 /* Tale a mbuf from the tx pool and copy in the user packet. */
529 m = kring->tx_pool[j];
531 RD(5, "This should never happen");
532 kring->tx_pool[j] = m = netmap_get_mbuf(GENERIC_BUF_SIZE);
533 if (unlikely(m == NULL)) {
534 D("mbuf allocation failed");
538 /* XXX we should ask notifications when NS_REPORT is set,
539 * or roughly every half frame. We can optimize this
540 * by lazily requesting notifications only when a
541 * transmission fails. Probably the best way is to
542 * break on failures and set notifications when
543 * ring->avail == 0 || j != k
545 tx_ret = generic_xmit_frame(ifp, m, addr, len, ring_nr);
546 if (unlikely(tx_ret)) {
547 RD(5, "start_xmit failed: err %d [%u,%u,%u,%u]",
548 tx_ret, kring->nr_ntc, j, k, kring->nr_hwavail);
550 * No room for this mbuf in the device driver.
551 * Request a notification FOR A PREVIOUS MBUF,
552 * then call generic_netmap_tx_clean(kring) to do the
553 * double check and see if we can free more buffers.
554 * If there is space continue, else break;
555 * NOTE: the double check is necessary if the problem
556 * occurs in the txsync call after selrecord().
557 * Also, we need some way to tell the caller that not
558 * all buffers were queued onto the device (this was
559 * not a problem with native netmap driver where space
560 * is preallocated). The bridge has a similar problem
561 * and we solve it there by dropping the excess packets.
563 generic_set_tx_event(kring, j);
564 if (generic_netmap_tx_clean(kring)) { /* space now available */
570 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
571 if (unlikely(++j == num_slots))
576 /* Update hwcur to the next slot to transmit. */
580 * Report all new slots as unavailable, even those not sent.
581 * We account for them with with hwreserved, so that
582 * nr_hwreserved =:= cur - nr_hwcur
584 kring->nr_hwavail -= new_slots;
585 kring->nr_hwreserved = k - j;
586 if (kring->nr_hwreserved < 0) {
587 kring->nr_hwreserved += num_slots;
590 IFRATE(rate_ctx.new.txpkt += ntx);
592 if (!kring->nr_hwavail) {
593 /* No more available slots? Set a notification event
594 * on a netmap slot that will be cleaned in the future.
595 * No doublecheck is performed, since txsync() will be
596 * called twice by netmap_poll().
598 generic_set_tx_event(kring, j);
600 ND("tx #%d, hwavail = %d", n, kring->nr_hwavail);
603 /* Synchronize the user's view to the kernel view. */
604 ring->avail = kring->nr_hwavail;
605 ring->reserved = kring->nr_hwreserved;
611 * This handler is registered (through netmap_catch_rx())
612 * within the attached network interface
613 * in the RX subsystem, so that every mbuf passed up by
614 * the driver can be stolen to the network stack.
615 * Stolen packets are put in a queue where the
616 * generic_netmap_rxsync() callback can extract them.
618 void generic_rx_handler(struct ifnet *ifp, struct mbuf *m)
620 struct netmap_adapter *na = NA(ifp);
621 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
623 u_int rr = 0; // receive ring number
626 /* limit the size of the queue */
627 if (unlikely(mbq_len(&na->rx_rings[rr].rx_queue) > 1024)) {
630 mbq_safe_enqueue(&na->rx_rings[rr].rx_queue, m);
633 if (netmap_generic_mit < 32768) {
634 /* no rx mitigation, pass notification up */
635 netmap_generic_irq(na->ifp, rr, &work_done);
636 IFRATE(rate_ctx.new.rxirq++);
638 /* same as send combining, filter notification if there is a
639 * pending timer, otherwise pass it up and start a timer.
641 if (likely(netmap_mitigation_active(gna))) {
642 /* Record that there is some pending work. */
643 gna->mit_pending = 1;
645 netmap_generic_irq(na->ifp, rr, &work_done);
646 IFRATE(rate_ctx.new.rxirq++);
647 netmap_mitigation_start(gna);
653 * generic_netmap_rxsync() extracts mbufs from the queue filled by
654 * generic_netmap_rx_handler() and puts their content in the netmap
656 * Access must be protected because the rx handler is asynchronous,
659 generic_netmap_rxsync(struct netmap_adapter *na, u_int ring_nr, int flags)
661 struct netmap_kring *kring = &na->rx_rings[ring_nr];
662 struct netmap_ring *ring = kring->ring;
663 u_int j, n, lim = kring->nkr_num_slots - 1;
664 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
665 u_int k, resvd = ring->reserved;
668 return netmap_ring_reinit(kring);
670 /* Import newly received packets into the netmap ring. */
671 if (netmap_no_pendintr || force_update) {
672 uint16_t slot_flags = kring->nkr_slot_flags;
676 j = kring->nr_ntc; /* first empty slot in the receive ring */
677 /* extract buffers from the rx queue, stop at most one
678 * slot before nr_hwcur (index k)
680 k = (kring->nr_hwcur) ? kring->nr_hwcur-1 : lim;
683 void *addr = NMB(&ring->slot[j]);
685 if (addr == netmap_buffer_base) { /* Bad buffer */
686 return netmap_ring_reinit(kring);
689 * Call the locked version of the function.
690 * XXX Ideally we could grab a batch of mbufs at once,
691 * by changing rx_queue into a ring.
693 m = mbq_safe_dequeue(&kring->rx_queue);
697 m_copydata(m, 0, len, addr);
698 ring->slot[j].len = len;
699 ring->slot[j].flags = slot_flags;
701 if (unlikely(j++ == lim))
707 kring->nr_hwavail += n;
708 IFRATE(rate_ctx.new.rxpkt += n);
710 kring->nr_kflags &= ~NKR_PENDINTR;
713 // XXX should we invert the order ?
714 /* Skip past packets that userspace has released */
718 if (resvd + ring->avail >= lim + 1) {
719 D("XXX invalid reserve/avail %d %d", resvd, ring->avail);
720 ring->reserved = resvd = 0; // XXX panic...
722 k = (k >= resvd) ? k - resvd : k + lim + 1 - resvd;
725 /* Userspace has released some packets. */
726 for (n = 0; j != k; n++) {
727 struct netmap_slot *slot = &ring->slot[j];
729 slot->flags &= ~NS_BUF_CHANGED;
730 if (unlikely(j++ == lim))
733 kring->nr_hwavail -= n;
736 /* Tell userspace that there are new packets. */
737 ring->avail = kring->nr_hwavail - resvd;
738 IFRATE(rate_ctx.new.rxsync++);
744 generic_netmap_dtor(struct netmap_adapter *na)
746 struct ifnet *ifp = na->ifp;
747 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter*)na;
748 struct netmap_adapter *prev_na = gna->prev;
750 if (prev_na != NULL) {
751 D("Released generic NA %p", gna);
753 netmap_adapter_put(prev_na);
757 D("Restored native NA %p", prev_na);
763 * generic_netmap_attach() makes it possible to use netmap on
764 * a device without native netmap support.
765 * This is less performant than native support but potentially
766 * faster than raw sockets or similar schemes.
768 * In this "emulated" mode, netmap rings do not necessarily
769 * have the same size as those in the NIC. We use a default
770 * value and possibly override it if the OS has ways to fetch the
771 * actual configuration.
774 generic_netmap_attach(struct ifnet *ifp)
776 struct netmap_adapter *na;
777 struct netmap_generic_adapter *gna;
779 u_int num_tx_desc, num_rx_desc;
781 num_tx_desc = num_rx_desc = netmap_generic_ringsize; /* starting point */
783 generic_find_num_desc(ifp, &num_tx_desc, &num_rx_desc);
784 ND("Netmap ring size: TX = %d, RX = %d", num_tx_desc, num_rx_desc);
786 gna = malloc(sizeof(*gna), M_DEVBUF, M_NOWAIT | M_ZERO);
788 D("no memory on attach, give up");
791 na = (struct netmap_adapter *)gna;
793 na->num_tx_desc = num_tx_desc;
794 na->num_rx_desc = num_rx_desc;
795 na->nm_register = &generic_netmap_register;
796 na->nm_txsync = &generic_netmap_txsync;
797 na->nm_rxsync = &generic_netmap_rxsync;
798 na->nm_dtor = &generic_netmap_dtor;
799 /* when using generic, IFCAP_NETMAP is set so we force
800 * NAF_SKIP_INTR to use the regular interrupt handler
802 na->na_flags = NAF_SKIP_INTR;
804 ND("[GNA] num_tx_queues(%d), real_num_tx_queues(%d), len(%lu)",
805 ifp->num_tx_queues, ifp->real_num_tx_queues,
807 ND("[GNA] num_rx_queues(%d), real_num_rx_queues(%d)",
808 ifp->num_rx_queues, ifp->real_num_rx_queues);
810 generic_find_num_queues(ifp, &na->num_tx_rings, &na->num_rx_rings);
812 retval = netmap_attach_common(na);