2 * Copyright (C) 2013-2014 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_unlock 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. */
206 generic_netmap_register(struct netmap_adapter *na, int enable)
208 struct ifnet *ifp = na->ifp;
209 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
218 error = ifp->netdev_ops->ndo_stop(ifp);
222 #endif /* REG_RESET */
224 if (enable) { /* Enable netmap mode. */
225 /* Initialize the rx queue, as generic_rx_handler() can
226 * be called as soon as netmap_catch_rx() returns.
228 for (r=0; r<na->num_rx_rings; r++) {
229 mbq_safe_init(&na->rx_rings[r].rx_queue);
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].tx_pool = NULL;
240 for (r=0; r<na->num_tx_rings; r++) {
241 na->tx_rings[r].tx_pool = malloc(na->num_tx_desc * sizeof(struct mbuf *),
242 M_DEVBUF, M_NOWAIT | M_ZERO);
243 if (!na->tx_rings[r].tx_pool) {
244 D("tx_pool allocation failed");
248 for (i=0; i<na->num_tx_desc; i++)
249 na->tx_rings[r].tx_pool[i] = NULL;
250 for (i=0; i<na->num_tx_desc; i++) {
251 m = netmap_get_mbuf(GENERIC_BUF_SIZE);
253 D("tx_pool[%d] allocation failed", i);
257 na->tx_rings[r].tx_pool[i] = m;
261 /* Prepare to intercept incoming traffic. */
262 error = netmap_catch_rx(na, 1);
264 D("netdev_rx_handler_register() failed (%d)", error);
265 goto register_handler;
267 ifp->if_capenable |= IFCAP_NETMAP;
269 /* Make netmap control the packet steering. */
270 netmap_catch_tx(gna, 1);
275 if (rate_ctx.refcount == 0) {
277 memset(&rate_ctx, 0, sizeof(rate_ctx));
278 setup_timer(&rate_ctx.timer, &rate_callback, (unsigned long)&rate_ctx);
279 if (mod_timer(&rate_ctx.timer, jiffies + msecs_to_jiffies(1500))) {
280 D("Error: mod_timer()");
286 } else if (na->tx_rings[0].tx_pool) {
287 /* Disable netmap mode. We enter here only if the previous
288 generic_netmap_register(na, 1) was successfull.
289 If it was not, na->tx_rings[0].tx_pool was set to NULL by the
290 error handling code below. */
293 ifp->if_capenable &= ~IFCAP_NETMAP;
295 /* Release packet steering control. */
296 netmap_catch_tx(gna, 0);
298 /* Do not intercept packets on the rx path. */
299 netmap_catch_rx(na, 0);
303 /* Free the mbufs going to the netmap rings */
304 for (r=0; r<na->num_rx_rings; r++) {
305 mbq_safe_purge(&na->rx_rings[r].rx_queue);
306 mbq_safe_destroy(&na->rx_rings[r].rx_queue);
309 netmap_mitigation_cleanup(gna);
311 for (r=0; r<na->num_tx_rings; r++) {
312 for (i=0; i<na->num_tx_desc; i++) {
313 m_freem(na->tx_rings[r].tx_pool[i]);
315 free(na->tx_rings[r].tx_pool, M_DEVBUF);
319 if (--rate_ctx.refcount == 0) {
321 del_timer(&rate_ctx.timer);
327 error = ifp->netdev_ops->ndo_open(ifp);
338 for (r=0; r<na->num_tx_rings; r++) {
339 if (na->tx_rings[r].tx_pool == NULL)
341 for (i=0; i<na->num_tx_desc; i++)
342 if (na->tx_rings[r].tx_pool[i])
343 m_freem(na->tx_rings[r].tx_pool[i]);
344 free(na->tx_rings[r].tx_pool, M_DEVBUF);
345 na->tx_rings[r].tx_pool = NULL;
347 netmap_mitigation_cleanup(gna);
348 for (r=0; r<na->num_rx_rings; r++) {
349 mbq_safe_destroy(&na->rx_rings[r].rx_queue);
356 * Callback invoked when the device driver frees an mbuf used
357 * by netmap to transmit a packet. This usually happens when
358 * the NIC notifies the driver that transmission is completed.
361 generic_mbuf_destructor(struct mbuf *m)
364 D("Tx irq (%p) queue %d", m, MBUF_TXQ(m));
365 netmap_generic_irq(MBUF_IFP(m), MBUF_TXQ(m), NULL);
367 m->m_ext.ext_type = EXT_PACKET;
368 m->m_ext.ext_free = NULL;
369 if (*(m->m_ext.ref_cnt) == 0)
370 *(m->m_ext.ref_cnt) = 1;
371 uma_zfree(zone_pack, m);
372 #endif /* __FreeBSD__ */
373 IFRATE(rate_ctx.new.txirq++);
376 /* Record completed transmissions and update hwtail.
378 * The oldest tx buffer not yet completed is at nr_hwtail + 1,
379 * nr_hwcur is the first unsent buffer.
382 generic_netmap_tx_clean(struct netmap_kring *kring)
384 u_int const lim = kring->nkr_num_slots - 1;
385 u_int nm_i = nm_next(kring->nr_hwtail, lim);
386 u_int hwcur = kring->nr_hwcur;
388 struct mbuf **tx_pool = kring->tx_pool;
390 while (nm_i != hwcur) { /* buffers not completed */
391 struct mbuf *m = tx_pool[nm_i];
393 if (unlikely(m == NULL)) {
394 /* this is done, try to replenish the entry */
395 tx_pool[nm_i] = m = netmap_get_mbuf(GENERIC_BUF_SIZE);
396 if (unlikely(m == NULL)) {
397 D("mbuf allocation failed, XXX error");
398 // XXX how do we proceed ? break ?
401 } else if (GET_MBUF_REFCNT(m) != 1) {
402 break; /* This mbuf is still busy: its refcnt is 2. */
405 nm_i = nm_next(nm_i, lim);
407 kring->nr_hwtail = nm_prev(nm_i, lim);
408 ND("tx completed [%d] -> hwtail %d", n, kring->nr_hwtail);
415 * We have pending packets in the driver between nr_hwtail +1 and hwcur.
416 * Compute a position in the middle, to be used to generate
420 generic_tx_event_middle(struct netmap_kring *kring, u_int hwcur)
422 u_int n = kring->nkr_num_slots;
423 u_int ntc = nm_next(kring->nr_hwtail, n-1);
427 e = (hwcur + ntc) / 2;
428 } else { /* wrap around */
429 e = (hwcur + n + ntc) / 2;
435 if (unlikely(e >= n)) {
436 D("This cannot happen");
444 * We have pending packets in the driver between nr_hwtail+1 and hwcur.
445 * Schedule a notification approximately in the middle of the two.
446 * There is a race but this is only called within txsync which does
450 generic_set_tx_event(struct netmap_kring *kring, u_int hwcur)
455 if (nm_next(kring->nr_hwtail, kring->nkr_num_slots -1) == hwcur) {
456 return; /* all buffers are free */
458 e = generic_tx_event_middle(kring, hwcur);
460 m = kring->tx_pool[e];
462 /* This can happen if there is already an event on the netmap
463 slot 'e': There is nothing to do. */
466 ND("Event at %d mbuf %p refcnt %d", e, m, GET_MBUF_REFCNT(m));
467 kring->tx_pool[e] = NULL;
468 SET_MBUF_DESTRUCTOR(m, generic_mbuf_destructor);
471 /* Decrement the refcount an free it if we have the last one. */
478 * generic_netmap_txsync() transforms netmap buffers into mbufs
479 * and passes them to the standard device driver
480 * (ndo_start_xmit() or ifp->if_transmit() ).
481 * On linux this is not done directly, but using dev_queue_xmit(),
482 * since it implements the TX flow control (and takes some locks).
485 generic_netmap_txsync(struct netmap_adapter *na, u_int ring_nr, int flags)
487 struct ifnet *ifp = na->ifp;
488 struct netmap_kring *kring = &na->tx_rings[ring_nr];
489 struct netmap_ring *ring = kring->ring;
490 u_int nm_i; /* index into the netmap ring */ // j
491 u_int const lim = kring->nkr_num_slots - 1;
492 u_int const head = kring->rhead;
494 IFRATE(rate_ctx.new.txsync++);
496 // TODO: handle the case of mbuf allocation failure
501 * First part: process new packets to send.
503 nm_i = kring->nr_hwcur;
504 if (nm_i != head) { /* we have new packets to send */
505 while (nm_i != head) {
506 struct netmap_slot *slot = &ring->slot[nm_i];
507 u_int len = slot->len;
508 void *addr = NMB(slot);
510 /* device-specific */
514 NM_CHECK_ADDR_LEN(addr, len);
516 /* Tale a mbuf from the tx pool and copy in the user packet. */
517 m = kring->tx_pool[nm_i];
519 RD(5, "This should never happen");
520 kring->tx_pool[nm_i] = m = netmap_get_mbuf(GENERIC_BUF_SIZE);
521 if (unlikely(m == NULL)) {
522 D("mbuf allocation failed");
526 /* XXX we should ask notifications when NS_REPORT is set,
527 * or roughly every half frame. We can optimize this
528 * by lazily requesting notifications only when a
529 * transmission fails. Probably the best way is to
530 * break on failures and set notifications when
531 * ring->cur == ring->tail || nm_i != cur
533 tx_ret = generic_xmit_frame(ifp, m, addr, len, ring_nr);
534 if (unlikely(tx_ret)) {
535 RD(5, "start_xmit failed: err %d [nm_i %u, head %u, hwtail %u]",
536 tx_ret, nm_i, head, kring->nr_hwtail);
538 * No room for this mbuf in the device driver.
539 * Request a notification FOR A PREVIOUS MBUF,
540 * then call generic_netmap_tx_clean(kring) to do the
541 * double check and see if we can free more buffers.
542 * If there is space continue, else break;
543 * NOTE: the double check is necessary if the problem
544 * occurs in the txsync call after selrecord().
545 * Also, we need some way to tell the caller that not
546 * all buffers were queued onto the device (this was
547 * not a problem with native netmap driver where space
548 * is preallocated). The bridge has a similar problem
549 * and we solve it there by dropping the excess packets.
551 generic_set_tx_event(kring, nm_i);
552 if (generic_netmap_tx_clean(kring)) { /* space now available */
558 slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
559 nm_i = nm_next(nm_i, lim);
562 /* Update hwcur to the next slot to transmit. */
563 kring->nr_hwcur = nm_i; /* not head, we could break early */
565 IFRATE(rate_ctx.new.txpkt += ntx);
569 * Second, reclaim completed buffers
571 if (flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) {
572 /* No more available slots? Set a notification event
573 * on a netmap slot that will be cleaned in the future.
574 * No doublecheck is performed, since txsync() will be
575 * called twice by netmap_poll().
577 generic_set_tx_event(kring, nm_i);
579 ND("tx #%d, hwtail = %d", n, kring->nr_hwtail);
581 generic_netmap_tx_clean(kring);
583 nm_txsync_finalize(kring);
590 * This handler is registered (through netmap_catch_rx())
591 * within the attached network interface
592 * in the RX subsystem, so that every mbuf passed up by
593 * the driver can be stolen to the network stack.
594 * Stolen packets are put in a queue where the
595 * generic_netmap_rxsync() callback can extract them.
598 generic_rx_handler(struct ifnet *ifp, struct mbuf *m)
600 struct netmap_adapter *na = NA(ifp);
601 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter *)na;
603 u_int rr = 0; // receive ring number
605 /* limit the size of the queue */
606 if (unlikely(mbq_len(&na->rx_rings[rr].rx_queue) > 1024)) {
609 mbq_safe_enqueue(&na->rx_rings[rr].rx_queue, m);
612 if (netmap_generic_mit < 32768) {
613 /* no rx mitigation, pass notification up */
614 netmap_generic_irq(na->ifp, rr, &work_done);
615 IFRATE(rate_ctx.new.rxirq++);
617 /* same as send combining, filter notification if there is a
618 * pending timer, otherwise pass it up and start a timer.
620 if (likely(netmap_mitigation_active(gna))) {
621 /* Record that there is some pending work. */
622 gna->mit_pending = 1;
624 netmap_generic_irq(na->ifp, rr, &work_done);
625 IFRATE(rate_ctx.new.rxirq++);
626 netmap_mitigation_start(gna);
632 * generic_netmap_rxsync() extracts mbufs from the queue filled by
633 * generic_netmap_rx_handler() and puts their content in the netmap
635 * Access must be protected because the rx handler is asynchronous,
638 generic_netmap_rxsync(struct netmap_adapter *na, u_int ring_nr, int flags)
640 struct netmap_kring *kring = &na->rx_rings[ring_nr];
641 struct netmap_ring *ring = kring->ring;
642 u_int nm_i; /* index into the netmap ring */ //j,
644 u_int const lim = kring->nkr_num_slots - 1;
645 u_int const head = nm_rxsync_prologue(kring);
646 int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
649 return netmap_ring_reinit(kring);
652 * First part: import newly received packets.
654 if (netmap_no_pendintr || force_update) {
655 /* extract buffers from the rx queue, stop at most one
656 * slot before nr_hwcur (stop_i)
658 uint16_t slot_flags = kring->nkr_slot_flags;
659 u_int stop_i = nm_prev(kring->nr_hwcur, lim);
661 nm_i = kring->nr_hwtail; /* first empty slot in the receive ring */
662 for (n = 0; nm_i != stop_i; n++) {
664 void *addr = NMB(&ring->slot[nm_i]);
667 /* we only check the address here on generic rx rings */
668 if (addr == netmap_buffer_base) { /* Bad buffer */
669 return netmap_ring_reinit(kring);
672 * Call the locked version of the function.
673 * XXX Ideally we could grab a batch of mbufs at once
674 * and save some locking overhead.
676 m = mbq_safe_dequeue(&kring->rx_queue);
677 if (!m) /* no more data */
680 m_copydata(m, 0, len, addr);
681 ring->slot[nm_i].len = len;
682 ring->slot[nm_i].flags = slot_flags;
684 nm_i = nm_next(nm_i, lim);
688 kring->nr_hwtail = nm_i;
689 IFRATE(rate_ctx.new.rxpkt += n);
691 kring->nr_kflags &= ~NKR_PENDINTR;
694 // XXX should we invert the order ?
696 * Second part: skip past packets that userspace has released.
698 nm_i = kring->nr_hwcur;
700 /* Userspace has released some packets. */
701 for (n = 0; nm_i != head; n++) {
702 struct netmap_slot *slot = &ring->slot[nm_i];
704 slot->flags &= ~NS_BUF_CHANGED;
705 nm_i = nm_next(nm_i, lim);
707 kring->nr_hwcur = head;
709 /* tell userspace that there might be new packets. */
710 nm_rxsync_finalize(kring);
711 IFRATE(rate_ctx.new.rxsync++);
717 generic_netmap_dtor(struct netmap_adapter *na)
719 struct ifnet *ifp = na->ifp;
720 struct netmap_generic_adapter *gna = (struct netmap_generic_adapter*)na;
721 struct netmap_adapter *prev_na = gna->prev;
723 if (prev_na != NULL) {
724 D("Released generic NA %p", gna);
726 netmap_adapter_put(prev_na);
730 D("Restored native NA %p", prev_na);
736 * generic_netmap_attach() makes it possible to use netmap on
737 * a device without native netmap support.
738 * This is less performant than native support but potentially
739 * faster than raw sockets or similar schemes.
741 * In this "emulated" mode, netmap rings do not necessarily
742 * have the same size as those in the NIC. We use a default
743 * value and possibly override it if the OS has ways to fetch the
744 * actual configuration.
747 generic_netmap_attach(struct ifnet *ifp)
749 struct netmap_adapter *na;
750 struct netmap_generic_adapter *gna;
752 u_int num_tx_desc, num_rx_desc;
754 num_tx_desc = num_rx_desc = netmap_generic_ringsize; /* starting point */
756 generic_find_num_desc(ifp, &num_tx_desc, &num_rx_desc);
757 ND("Netmap ring size: TX = %d, RX = %d", num_tx_desc, num_rx_desc);
759 gna = malloc(sizeof(*gna), M_DEVBUF, M_NOWAIT | M_ZERO);
761 D("no memory on attach, give up");
764 na = (struct netmap_adapter *)gna;
766 na->num_tx_desc = num_tx_desc;
767 na->num_rx_desc = num_rx_desc;
768 na->nm_register = &generic_netmap_register;
769 na->nm_txsync = &generic_netmap_txsync;
770 na->nm_rxsync = &generic_netmap_rxsync;
771 na->nm_dtor = &generic_netmap_dtor;
772 /* when using generic, IFCAP_NETMAP is set so we force
773 * NAF_SKIP_INTR to use the regular interrupt handler
775 na->na_flags = NAF_SKIP_INTR;
777 ND("[GNA] num_tx_queues(%d), real_num_tx_queues(%d), len(%lu)",
778 ifp->num_tx_queues, ifp->real_num_tx_queues,
780 ND("[GNA] num_rx_queues(%d), real_num_rx_queues(%d)",
781 ifp->num_rx_queues, ifp->real_num_rx_queues);
783 generic_find_num_queues(ifp, &na->num_tx_rings, &na->num_rx_rings);
785 retval = netmap_attach_common(na);