2 * SPDX-License-Identifier: BSD-2-Clause
4 * Copyright (C) 2014-2018 Giuseppe Lettieri
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 #if defined(__FreeBSD__)
31 #include <sys/cdefs.h> /* prerequisite */
33 #include <sys/types.h>
34 #include <sys/errno.h>
35 #include <sys/param.h> /* defines used in kernel.h */
36 #include <sys/kernel.h> /* types used in module initialization */
37 #include <sys/malloc.h>
40 #include <sys/rwlock.h>
41 #include <sys/selinfo.h>
42 #include <sys/sysctl.h>
43 #include <sys/socket.h> /* sockaddrs */
45 #include <net/if_var.h>
46 #include <machine/bus.h> /* bus_dmamap_* */
47 #include <sys/refcount.h>
54 #elif defined(__APPLE__)
56 #warning OSX support is only partial
64 #error Unsupported platform
66 #endif /* unsupported */
72 #include <net/netmap.h>
73 #include <dev/netmap/netmap_kern.h>
74 #include <dev/netmap/netmap_mem2.h>
78 #define NM_PIPE_MAXSLOTS 4096
79 #define NM_PIPE_MAXRINGS 256
81 static int netmap_default_pipes = 0; /* ignored, kept for compatibility */
83 SYSCTL_DECL(_dev_netmap);
84 SYSCTL_INT(_dev_netmap, OID_AUTO, default_pipes, CTLFLAG_RW,
85 &netmap_default_pipes, 0, "For compatibility only");
88 /* allocate the pipe array in the parent adapter */
90 nm_pipe_alloc(struct netmap_adapter *na, u_int npipes)
93 struct netmap_pipe_adapter **npa;
95 if (npipes <= na->na_max_pipes)
96 /* we already have more entries that requested */
99 if (npipes < na->na_next_pipe || npipes > NM_MAXPIPES)
102 old_len = sizeof(struct netmap_pipe_adapter *)*na->na_max_pipes;
103 len = sizeof(struct netmap_pipe_adapter *) * npipes;
104 npa = nm_os_realloc(na->na_pipes, len, old_len);
109 na->na_max_pipes = npipes;
114 /* deallocate the parent array in the parent adapter */
116 netmap_pipe_dealloc(struct netmap_adapter *na)
119 if (na->na_next_pipe > 0) {
120 nm_prerr("freeing not empty pipe array for %s (%d dangling pipes)!",
121 na->name, na->na_next_pipe);
123 nm_os_free(na->na_pipes);
125 na->na_max_pipes = 0;
126 na->na_next_pipe = 0;
130 /* find a pipe endpoint with the given id among the parent's pipes */
131 static struct netmap_pipe_adapter *
132 netmap_pipe_find(struct netmap_adapter *parent, const char *pipe_id)
135 struct netmap_pipe_adapter *na;
137 for (i = 0; i < parent->na_next_pipe; i++) {
138 const char *na_pipe_id;
139 na = parent->na_pipes[i];
140 na_pipe_id = strrchr(na->up.name,
141 na->role == NM_PIPE_ROLE_MASTER ? '{' : '}');
142 KASSERT(na_pipe_id != NULL, ("Invalid pipe name"));
144 if (!strcmp(na_pipe_id, pipe_id)) {
151 /* add a new pipe endpoint to the parent array */
153 netmap_pipe_add(struct netmap_adapter *parent, struct netmap_pipe_adapter *na)
155 if (parent->na_next_pipe >= parent->na_max_pipes) {
156 u_int npipes = parent->na_max_pipes ? 2*parent->na_max_pipes : 2;
157 int error = nm_pipe_alloc(parent, npipes);
162 parent->na_pipes[parent->na_next_pipe] = na;
163 na->parent_slot = parent->na_next_pipe;
164 parent->na_next_pipe++;
168 /* remove the given pipe endpoint from the parent array */
170 netmap_pipe_remove(struct netmap_adapter *parent, struct netmap_pipe_adapter *na)
173 n = --parent->na_next_pipe;
174 if (n != na->parent_slot) {
175 struct netmap_pipe_adapter **p =
176 &parent->na_pipes[na->parent_slot];
177 *p = parent->na_pipes[n];
178 (*p)->parent_slot = na->parent_slot;
180 parent->na_pipes[n] = NULL;
184 netmap_pipe_txsync(struct netmap_kring *txkring, int flags)
186 struct netmap_kring *rxkring = txkring->pipe;
187 u_int k, lim = txkring->nkr_num_slots - 1, nk;
188 int m; /* slots to transfer */
189 int complete; /* did we see a complete packet ? */
190 struct netmap_ring *txring = txkring->ring, *rxring = rxkring->ring;
192 nm_prdis("%p: %s %x -> %s", txkring, txkring->name, flags, rxkring->name);
193 nm_prdis(20, "TX before: hwcur %d hwtail %d cur %d head %d tail %d",
194 txkring->nr_hwcur, txkring->nr_hwtail,
195 txkring->rcur, txkring->rhead, txkring->rtail);
197 /* update the hwtail */
198 txkring->nr_hwtail = txkring->pipe_tail;
200 m = txkring->rhead - txkring->nr_hwcur; /* new slots */
202 m += txkring->nkr_num_slots;
205 /* nothing to send */
209 for (k = txkring->nr_hwcur, nk = lim + 1, complete = 0; m;
210 m--, k = nm_next(k, lim), nk = (complete ? k : nk)) {
211 struct netmap_slot *rs = &rxring->slot[k];
212 struct netmap_slot *ts = &txring->slot[k];
213 uint64_t off = nm_get_offset(rxkring, rs);
216 if (nm_get_offset(rxkring, rs) < off) {
217 nm_write_offset(rxkring, rs, off);
219 if (ts->flags & NS_BUF_CHANGED) {
220 ts->flags &= ~NS_BUF_CHANGED;
222 complete = !(ts->flags & NS_MOREFRAG);
225 txkring->nr_hwcur = k;
227 nm_prdis(20, "TX after : hwcur %d hwtail %d cur %d head %d tail %d k %d",
228 txkring->nr_hwcur, txkring->nr_hwtail,
229 txkring->rcur, txkring->rhead, txkring->rtail, k);
231 if (likely(nk <= lim)) {
232 mb(); /* make sure the slots are updated before publishing them */
233 rxkring->pipe_tail = nk; /* only publish complete packets */
234 rxkring->nm_notify(rxkring, 0);
241 netmap_pipe_rxsync(struct netmap_kring *rxkring, int flags)
243 struct netmap_kring *txkring = rxkring->pipe;
244 u_int k, lim = rxkring->nkr_num_slots - 1;
245 int m; /* slots to release */
246 struct netmap_ring *txring = txkring->ring, *rxring = rxkring->ring;
248 nm_prdis("%p: %s %x -> %s", txkring, txkring->name, flags, rxkring->name);
249 nm_prdis(20, "RX before: hwcur %d hwtail %d cur %d head %d tail %d",
250 rxkring->nr_hwcur, rxkring->nr_hwtail,
251 rxkring->rcur, rxkring->rhead, rxkring->rtail);
253 /* update the hwtail */
254 rxkring->nr_hwtail = rxkring->pipe_tail;
256 m = rxkring->rhead - rxkring->nr_hwcur; /* released slots */
258 m += rxkring->nkr_num_slots;
261 /* nothing to release */
265 for (k = rxkring->nr_hwcur; m; m--, k = nm_next(k, lim)) {
266 struct netmap_slot *rs = &rxring->slot[k];
267 struct netmap_slot *ts = &txring->slot[k];
269 /* copy the slot. This also propagates any offset */
271 if (rs->flags & NS_BUF_CHANGED) {
272 rs->flags &= ~NS_BUF_CHANGED;
276 mb(); /* make sure the slots are updated before publishing them */
277 txkring->pipe_tail = nm_prev(k, lim);
278 rxkring->nr_hwcur = k;
280 nm_prdis(20, "RX after : hwcur %d hwtail %d cur %d head %d tail %d k %d",
281 rxkring->nr_hwcur, rxkring->nr_hwtail,
282 rxkring->rcur, rxkring->rhead, rxkring->rtail, k);
284 txkring->nm_notify(txkring, 0);
289 /* Pipe endpoints are created and destroyed together, so that endopoints do not
290 * have to check for the existence of their peer at each ?xsync.
292 * To play well with the existing netmap infrastructure (refcounts etc.), we
293 * adopt the following strategy:
295 * 1) The first endpoint that is created also creates the other endpoint and
296 * grabs a reference to it.
298 * state A) user1 --> endpoint1 --> endpoint2
300 * 2) If, starting from state A, endpoint2 is then registered, endpoint1 gives
301 * its reference to the user:
303 * state B) user1 --> endpoint1 endpoint2 <--- user2
305 * 3) Assume that, starting from state B endpoint2 is closed. In the unregister
306 * callback endpoint2 notes that endpoint1 is still active and adds a reference
307 * from endpoint1 to itself. When user2 then releases her own reference,
308 * endpoint2 is not destroyed and we are back to state A. A symmetrical state
309 * would be reached if endpoint1 were released instead.
311 * 4) If, starting from state A, endpoint1 is closed, the destructor notes that
312 * it owns a reference to endpoint2 and releases it.
314 * Something similar goes on for the creation and destruction of the krings.
318 int netmap_pipe_krings_create_both(struct netmap_adapter *na,
319 struct netmap_adapter *ona)
326 nm_prdis("%p: case 1, create both ends", na);
327 error = netmap_krings_create(na, 0);
331 /* create the krings of the other end */
332 error = netmap_krings_create(ona, 0);
336 /* cross link the krings and initialize the pipe_tails */
338 enum txrx r = nm_txrx_swap(t); /* swap NR_TX <-> NR_RX */
339 for (i = 0; i < nma_get_nrings(na, t); i++) {
340 struct netmap_kring *k1 = NMR(na, t)[i],
341 *k2 = NMR(ona, r)[i];
344 /* mark all peer-adapter rings as fake */
345 k2->nr_kflags |= NKR_FAKERING;
347 k1->pipe_tail = k1->nr_hwtail;
348 k2->pipe_tail = k2->nr_hwtail;
355 netmap_krings_delete(na);
359 /* netmap_pipe_krings_create.
361 * There are two cases:
367 * and we are e1. We have to create both sets
374 * and we are e2. e1 is certainly registered and our
375 * krings already exist. Nothing to do.
378 netmap_pipe_krings_create(struct netmap_adapter *na)
380 struct netmap_pipe_adapter *pna =
381 (struct netmap_pipe_adapter *)na;
382 struct netmap_adapter *ona = &pna->peer->up;
385 return netmap_pipe_krings_create_both(na, ona);
391 netmap_pipe_reg_both(struct netmap_adapter *na, struct netmap_adapter *ona)
397 for (i = 0; i < nma_get_nrings(na, t); i++) {
398 struct netmap_kring *kring = NMR(na, t)[i];
400 if (nm_kring_pending_on(kring)) {
401 /* mark the peer ring as needed */
402 kring->pipe->nr_kflags |= NKR_NEEDRING;
407 /* create all missing needed rings on the other end.
408 * Either our end, or the other, has been marked as
409 * fake, so the allocation will not be done twice.
411 error = netmap_mem_rings_create(ona);
415 /* In case of no error we put our rings in netmap mode */
417 for (i = 0; i < nma_get_nrings(na, t); i++) {
418 struct netmap_kring *kring = NMR(na, t)[i];
419 if (nm_kring_pending_on(kring)) {
421 kring->nr_mode = NKR_NETMAP_ON;
422 if ((kring->nr_kflags & NKR_FAKERING) &&
423 (kring->pipe->nr_kflags & NKR_FAKERING)) {
424 /* this is a re-open of a pipe
425 * end-point kept alive by the other end.
426 * We need to leave everything as it is
431 /* copy the buffers from the non-fake ring
432 * (this also propagates any initial offset)
434 memcpy(kring->pipe->ring->slot,
436 sizeof(struct netmap_slot) *
437 kring->nkr_num_slots);
438 /* copy the offset-related fields */
439 *(uint64_t *)(uintptr_t)&kring->pipe->ring->offset_mask =
440 kring->ring->offset_mask;
441 *(uint64_t *)(uintptr_t)&kring->pipe->ring->buf_align =
442 kring->ring->buf_align;
443 /* mark both rings as fake and needed,
444 * so that buffers will not be
445 * deleted by the standard machinery
446 * (we will delete them by ourselves in
447 * netmap_pipe_krings_delete)
450 (NKR_FAKERING | NKR_NEEDRING);
451 kring->nr_mode = NKR_NETMAP_ON;
461 * There are two cases on registration (onoff==1)
467 * and we are e1. Create the needed rings of the
472 * usr1 --> e1 --> e2 <-- usr2
474 * and we are e2. Drop the ref e1 is holding.
476 * There are two additional cases on unregister (onoff==0)
482 * and we are e1. Nothing special to do, e2 will
483 * be cleaned up by the destructor of e1.
487 * usr1 --> e1 e2 <-- usr2
489 * and we are either e1 or e2. Add a ref from the
493 netmap_pipe_reg(struct netmap_adapter *na, int onoff)
495 struct netmap_pipe_adapter *pna =
496 (struct netmap_pipe_adapter *)na;
497 struct netmap_adapter *ona = &pna->peer->up;
500 nm_prdis("%p: onoff %d", na, onoff);
502 error = netmap_pipe_reg_both(na, ona);
506 if (na->active_fds == 0)
507 na->na_flags |= NAF_NETMAP_ON;
509 if (na->active_fds == 0)
510 na->na_flags &= ~NAF_NETMAP_ON;
511 netmap_krings_mode_commit(na, onoff);
514 if (na->active_fds) {
515 nm_prdis("active_fds %d", na->active_fds);
520 nm_prdis("%p: case 1.a or 2.a, nothing to do", na);
524 nm_prdis("%p: case 1.b, drop peer", na);
525 pna->peer->peer_ref = 0;
526 netmap_adapter_put(na);
528 nm_prdis("%p: case 2.b, grab peer", na);
529 netmap_adapter_get(na);
530 pna->peer->peer_ref = 1;
536 netmap_pipe_krings_delete_both(struct netmap_adapter *na,
537 struct netmap_adapter *ona)
539 struct netmap_adapter *sna;
544 nm_prdis("%p: case 1, deleting everything", na);
545 /* To avoid double-frees we zero-out all the buffers in the kernel part
546 * of each ring. The reason is this: If the user is behaving correctly,
547 * all buffers are found in exactly one slot in the userspace part of
548 * some ring. If the user is not behaving correctly, we cannot release
549 * buffers cleanly anyway. In the latter case, the allocator will
550 * return to a clean state only when all its users will close.
555 for (i = 0; i < nma_get_nrings(sna, t); i++) {
556 struct netmap_kring *kring = NMR(sna, t)[i];
557 struct netmap_ring *ring = kring->ring;
558 uint32_t j, lim = kring->nkr_num_slots - 1;
560 nm_prdis("%s ring %p hwtail %u hwcur %u",
561 kring->name, ring, kring->nr_hwtail, kring->nr_hwcur);
566 if (kring->tx == NR_RX)
567 ring->slot[kring->pipe_tail].buf_idx = 0;
569 for (j = nm_next(kring->pipe_tail, lim);
570 j != kring->nr_hwcur;
573 nm_prdis("%s[%d] %u", kring->name, j, ring->slot[j].buf_idx);
574 ring->slot[j].buf_idx = 0;
576 kring->nr_kflags &= ~(NKR_FAKERING | NKR_NEEDRING);
580 if (sna != ona && ona->tx_rings) {
585 netmap_mem_rings_delete(na);
586 netmap_krings_delete(na); /* also zeroes tx_rings etc. */
588 if (ona->tx_rings == NULL) {
589 /* already deleted, we must be on an
590 * cleanup-after-error path */
593 netmap_mem_rings_delete(ona);
594 netmap_krings_delete(ona);
597 /* netmap_pipe_krings_delete.
599 * There are two cases:
605 * and we are e1 (e2 is not registered, so krings_delete cannot be
610 * usr1 --> e1 e2 <-- usr2
612 * and we are either e1 or e2.
614 * In the former case we have to also delete the krings of e2;
615 * in the latter case we do nothing.
618 netmap_pipe_krings_delete(struct netmap_adapter *na)
620 struct netmap_pipe_adapter *pna =
621 (struct netmap_pipe_adapter *)na;
622 struct netmap_adapter *ona; /* na of the other end */
624 if (!pna->peer_ref) {
625 nm_prdis("%p: case 2, kept alive by peer", na);
628 ona = &pna->peer->up;
629 netmap_pipe_krings_delete_both(na, ona);
634 netmap_pipe_dtor(struct netmap_adapter *na)
636 struct netmap_pipe_adapter *pna =
637 (struct netmap_pipe_adapter *)na;
638 nm_prdis("%p %p", na, pna->parent_ifp);
640 nm_prdis("%p: clean up peer", na);
642 netmap_adapter_put(&pna->peer->up);
644 if (pna->role == NM_PIPE_ROLE_MASTER)
645 netmap_pipe_remove(pna->parent, pna);
647 if_rele(pna->parent_ifp);
648 netmap_adapter_put(pna->parent);
653 netmap_get_pipe_na(struct nmreq_header *hdr, struct netmap_adapter **na,
654 struct netmap_mem_d *nmd, int create)
656 struct nmreq_register *req = (struct nmreq_register *)(uintptr_t)hdr->nr_body;
657 struct netmap_adapter *pna; /* parent adapter */
658 struct netmap_pipe_adapter *mna, *sna, *reqna;
660 const char *pipe_id = NULL;
662 int error, retries = 0;
663 char *cbra, pipe_char;
665 /* Try to parse the pipe syntax 'xx{yy' or 'xx}yy'. */
666 cbra = strrchr(hdr->nr_name, '{');
668 role = NM_PIPE_ROLE_MASTER;
670 cbra = strrchr(hdr->nr_name, '}');
672 role = NM_PIPE_ROLE_SLAVE;
674 nm_prdis("not a pipe");
680 if (*pipe_id == '\0' || cbra == hdr->nr_name) {
681 /* Bracket is the last character, so pipe name is missing;
682 * or bracket is the first character, so base port name
687 if (req->nr_mode != NR_REG_ALL_NIC && req->nr_mode != NR_REG_ONE_NIC) {
688 /* We only accept modes involving hardware rings. */
692 /* first, try to find the parent adapter */
696 /* Temporarily remove the pipe suffix. */
698 error = netmap_get_na(hdr, &pna, &ifp, nmd, create);
699 /* Restore the pipe suffix. */
703 if (error != ENXIO || retries++) {
704 nm_prdis("parent lookup failed: %d", error);
707 nm_prdis("try to create a persistent vale port");
708 /* create a persistent vale port and try again */
711 create_error = netmap_vi_create(hdr, 1 /* autodelete */);
714 if (create_error && create_error != EEXIST) {
715 if (create_error != EOPNOTSUPP) {
716 nm_prerr("failed to create a persistent vale port: %d",
723 if (NETMAP_OWNED_BY_KERN(pna)) {
724 nm_prdis("parent busy");
729 /* next, lookup the pipe id in the parent list */
731 mna = netmap_pipe_find(pna, pipe_id);
733 if (mna->role == role) {
734 nm_prdis("found %s directly at %d", pipe_id, mna->parent_slot);
737 nm_prdis("found %s indirectly at %d", pipe_id, mna->parent_slot);
740 /* the pipe we have found already holds a ref to the parent,
741 * so we need to drop the one we got from netmap_get_na()
743 netmap_unget_na(pna, ifp);
746 nm_prdis("pipe %s not found, create %d", pipe_id, create);
751 /* we create both master and slave.
752 * The endpoint we were asked for holds a reference to
755 mna = nm_os_malloc(sizeof(*mna));
760 snprintf(mna->up.name, sizeof(mna->up.name), "%s{%s", pna->name, pipe_id);
762 mna->role = NM_PIPE_ROLE_MASTER;
764 mna->parent_ifp = ifp;
766 mna->up.nm_txsync = netmap_pipe_txsync;
767 mna->up.nm_rxsync = netmap_pipe_rxsync;
768 mna->up.nm_register = netmap_pipe_reg;
769 mna->up.nm_dtor = netmap_pipe_dtor;
770 mna->up.nm_krings_create = netmap_pipe_krings_create;
771 mna->up.nm_krings_delete = netmap_pipe_krings_delete;
772 mna->up.nm_mem = netmap_mem_get(pna->nm_mem);
773 mna->up.na_flags |= NAF_MEM_OWNER | NAF_OFFSETS;
774 mna->up.na_lut = pna->na_lut;
776 mna->up.num_tx_rings = req->nr_tx_rings;
777 nm_bound_var(&mna->up.num_tx_rings, 1,
778 1, NM_PIPE_MAXRINGS, NULL);
779 mna->up.num_rx_rings = req->nr_rx_rings;
780 nm_bound_var(&mna->up.num_rx_rings, 1,
781 1, NM_PIPE_MAXRINGS, NULL);
782 mna->up.num_tx_desc = req->nr_tx_slots;
783 nm_bound_var(&mna->up.num_tx_desc, pna->num_tx_desc,
784 1, NM_PIPE_MAXSLOTS, NULL);
785 mna->up.num_rx_desc = req->nr_rx_slots;
786 nm_bound_var(&mna->up.num_rx_desc, pna->num_rx_desc,
787 1, NM_PIPE_MAXSLOTS, NULL);
788 error = netmap_attach_common(&mna->up);
791 /* register the master with the parent */
792 error = netmap_pipe_add(pna, mna);
796 /* create the slave */
797 sna = nm_os_malloc(sizeof(*mna));
802 /* most fields are the same, copy from master and then fix */
804 sna->up.nm_mem = netmap_mem_get(mna->up.nm_mem);
805 /* swap the number of tx/rx rings and slots */
806 sna->up.num_tx_rings = mna->up.num_rx_rings;
807 sna->up.num_tx_desc = mna->up.num_rx_desc;
808 sna->up.num_rx_rings = mna->up.num_tx_rings;
809 sna->up.num_rx_desc = mna->up.num_tx_desc;
810 snprintf(sna->up.name, sizeof(sna->up.name), "%s}%s", pna->name, pipe_id);
811 sna->role = NM_PIPE_ROLE_SLAVE;
812 error = netmap_attach_common(&sna->up);
816 /* join the two endpoints */
820 /* we already have a reference to the parent, but we
821 * need another one for the other endpoint we created
823 netmap_adapter_get(pna);
824 /* likewise for the ifp, if any */
828 if (role == NM_PIPE_ROLE_MASTER) {
831 netmap_adapter_get(&sna->up);
835 netmap_adapter_get(&mna->up);
837 nm_prdis("created master %p and slave %p", mna, sna);
840 nm_prdis("pipe %s %s at %p", pipe_id,
841 (reqna->role == NM_PIPE_ROLE_MASTER ? "master" : "slave"), reqna);
843 netmap_adapter_get(*na);
845 /* keep the reference to the parent.
846 * It will be released by the req destructor
854 netmap_pipe_remove(pna, mna);
858 netmap_unget_na(pna, ifp);
863 #endif /* WITH_PIPES */