1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause
3 *
4 * Copyright (C) 2013-2014 Universita` di Pisa. All rights reserved.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 *
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25 * SUCH DAMAGE.
26 */
27
28 #include "opt_inet.h"
29 #include "opt_inet6.h"
30
31 #include <sys/param.h>
32 #include <sys/module.h>
33 #include <sys/errno.h>
34 #include <sys/eventhandler.h>
35 #include <sys/jail.h>
36 #include <sys/poll.h> /* POLLIN, POLLOUT */
37 #include <sys/kernel.h> /* types used in module initialization */
38 #include <sys/conf.h> /* DEV_MODULE_ORDERED */
39 #include <sys/endian.h>
40 #include <sys/syscallsubr.h> /* kern_ioctl() */
41
42 #include <sys/rwlock.h>
43
44 #include <vm/vm.h> /* vtophys */
45 #include <vm/pmap.h> /* vtophys */
46 #include <vm/vm_param.h>
47 #include <vm/vm_object.h>
48 #include <vm/vm_page.h>
49 #include <vm/vm_pager.h>
50 #include <vm/uma.h>
51
52
53 #include <sys/malloc.h>
54 #include <sys/socket.h> /* sockaddrs */
55 #include <sys/selinfo.h>
56 #include <sys/kthread.h> /* kthread_add() */
57 #include <sys/proc.h> /* PROC_LOCK() */
58 #include <sys/unistd.h> /* RFNOWAIT */
59 #include <sys/sched.h> /* sched_bind() */
60 #include <sys/smp.h> /* mp_maxid */
61 #include <sys/taskqueue.h> /* taskqueue_enqueue(), taskqueue_create(), ... */
62 #include <net/if.h>
63 #include <net/if_var.h>
64 #include <net/if_types.h> /* IFT_ETHER */
65 #include <net/ethernet.h> /* ether_ifdetach */
66 #include <net/if_dl.h> /* LLADDR */
67 #include <machine/bus.h> /* bus_dmamap_* */
68 #include <netinet/in.h> /* in6_cksum_pseudo() */
69 #include <machine/in_cksum.h> /* in_pseudo(), in_cksum_hdr() */
70
71 #include <net/netmap.h>
72 #include <dev/netmap/netmap_kern.h>
73 #include <net/netmap_virt.h>
74 #include <dev/netmap/netmap_mem2.h>
75
76
77 /* ======================== FREEBSD-SPECIFIC ROUTINES ================== */
78
79 static void
nm_kqueue_notify(void * opaque,int pending)80 nm_kqueue_notify(void *opaque, int pending)
81 {
82 struct nm_selinfo *si = opaque;
83
84 /* We use a non-zero hint to distinguish this notification call
85 * from the call done in kqueue_scan(), which uses hint=0.
86 */
87 KNOTE_UNLOCKED(&si->si.si_note, /*hint=*/0x100);
88 }
89
nm_os_selinfo_init(NM_SELINFO_T * si,const char * name)90 int nm_os_selinfo_init(NM_SELINFO_T *si, const char *name) {
91 int err;
92
93 TASK_INIT(&si->ntfytask, 0, nm_kqueue_notify, si);
94 si->ntfytq = taskqueue_create(name, M_NOWAIT,
95 taskqueue_thread_enqueue, &si->ntfytq);
96 if (si->ntfytq == NULL)
97 return -ENOMEM;
98 err = taskqueue_start_threads(&si->ntfytq, 1, PI_NET, "tq %s", name);
99 if (err) {
100 taskqueue_free(si->ntfytq);
101 si->ntfytq = NULL;
102 return err;
103 }
104
105 snprintf(si->mtxname, sizeof(si->mtxname), "nmkl%s", name);
106 mtx_init(&si->m, si->mtxname, NULL, MTX_DEF);
107 knlist_init_mtx(&si->si.si_note, &si->m);
108 si->kqueue_users = 0;
109
110 return (0);
111 }
112
113 void
nm_os_selinfo_uninit(NM_SELINFO_T * si)114 nm_os_selinfo_uninit(NM_SELINFO_T *si)
115 {
116 if (si->ntfytq == NULL) {
117 return; /* si was not initialized */
118 }
119 taskqueue_drain(si->ntfytq, &si->ntfytask);
120 taskqueue_free(si->ntfytq);
121 si->ntfytq = NULL;
122 knlist_delete(&si->si.si_note, curthread, /*islocked=*/0);
123 knlist_destroy(&si->si.si_note);
124 /* now we don't need the mutex anymore */
125 mtx_destroy(&si->m);
126 }
127
128 void *
nm_os_malloc(size_t size)129 nm_os_malloc(size_t size)
130 {
131 return malloc(size, M_DEVBUF, M_NOWAIT | M_ZERO);
132 }
133
134 void *
nm_os_realloc(void * addr,size_t new_size,size_t old_size __unused)135 nm_os_realloc(void *addr, size_t new_size, size_t old_size __unused)
136 {
137 return realloc(addr, new_size, M_DEVBUF, M_NOWAIT | M_ZERO);
138 }
139
140 void
nm_os_free(void * addr)141 nm_os_free(void *addr)
142 {
143 free(addr, M_DEVBUF);
144 }
145
146 void
nm_os_ifnet_lock(void)147 nm_os_ifnet_lock(void)
148 {
149 IFNET_RLOCK();
150 }
151
152 void
nm_os_ifnet_unlock(void)153 nm_os_ifnet_unlock(void)
154 {
155 IFNET_RUNLOCK();
156 }
157
158 static int netmap_use_count = 0;
159
160 void
nm_os_get_module(void)161 nm_os_get_module(void)
162 {
163 netmap_use_count++;
164 }
165
166 void
nm_os_put_module(void)167 nm_os_put_module(void)
168 {
169 netmap_use_count--;
170 }
171
172 static void
netmap_ifnet_arrival_handler(void * arg __unused,if_t ifp)173 netmap_ifnet_arrival_handler(void *arg __unused, if_t ifp)
174 {
175 netmap_undo_zombie(ifp);
176 }
177
178 static void
netmap_ifnet_departure_handler(void * arg __unused,if_t ifp)179 netmap_ifnet_departure_handler(void *arg __unused, if_t ifp)
180 {
181 netmap_make_zombie(ifp);
182 }
183
184 static eventhandler_tag nm_ifnet_ah_tag;
185 static eventhandler_tag nm_ifnet_dh_tag;
186
187 int
nm_os_ifnet_init(void)188 nm_os_ifnet_init(void)
189 {
190 nm_ifnet_ah_tag =
191 EVENTHANDLER_REGISTER(ifnet_arrival_event,
192 netmap_ifnet_arrival_handler,
193 NULL, EVENTHANDLER_PRI_ANY);
194 nm_ifnet_dh_tag =
195 EVENTHANDLER_REGISTER(ifnet_departure_event,
196 netmap_ifnet_departure_handler,
197 NULL, EVENTHANDLER_PRI_ANY);
198 return 0;
199 }
200
201 void
nm_os_ifnet_fini(void)202 nm_os_ifnet_fini(void)
203 {
204 EVENTHANDLER_DEREGISTER(ifnet_arrival_event,
205 nm_ifnet_ah_tag);
206 EVENTHANDLER_DEREGISTER(ifnet_departure_event,
207 nm_ifnet_dh_tag);
208 }
209
210 unsigned
nm_os_ifnet_mtu(if_t ifp)211 nm_os_ifnet_mtu(if_t ifp)
212 {
213 return if_getmtu(ifp);
214 }
215
216 rawsum_t
nm_os_csum_raw(uint8_t * data,size_t len,rawsum_t cur_sum)217 nm_os_csum_raw(uint8_t *data, size_t len, rawsum_t cur_sum)
218 {
219 /* TODO XXX please use the FreeBSD implementation for this. */
220 uint16_t *words = (uint16_t *)data;
221 int nw = len / 2;
222 int i;
223
224 for (i = 0; i < nw; i++)
225 cur_sum += be16toh(words[i]);
226
227 if (len & 1)
228 cur_sum += (data[len-1] << 8);
229
230 return cur_sum;
231 }
232
233 /* Fold a raw checksum: 'cur_sum' is in host byte order, while the
234 * return value is in network byte order.
235 */
236 uint16_t
nm_os_csum_fold(rawsum_t cur_sum)237 nm_os_csum_fold(rawsum_t cur_sum)
238 {
239 /* TODO XXX please use the FreeBSD implementation for this. */
240 while (cur_sum >> 16)
241 cur_sum = (cur_sum & 0xFFFF) + (cur_sum >> 16);
242
243 return htobe16((~cur_sum) & 0xFFFF);
244 }
245
nm_os_csum_ipv4(struct nm_iphdr * iph)246 uint16_t nm_os_csum_ipv4(struct nm_iphdr *iph)
247 {
248 #if 0
249 return in_cksum_hdr((void *)iph);
250 #else
251 return nm_os_csum_fold(nm_os_csum_raw((uint8_t*)iph, sizeof(struct nm_iphdr), 0));
252 #endif
253 }
254
255 void
nm_os_csum_tcpudp_ipv4(struct nm_iphdr * iph,void * data,size_t datalen,uint16_t * check)256 nm_os_csum_tcpudp_ipv4(struct nm_iphdr *iph, void *data,
257 size_t datalen, uint16_t *check)
258 {
259 #ifdef INET
260 uint16_t pseudolen = datalen + iph->protocol;
261
262 /* Compute and insert the pseudo-header checksum. */
263 *check = in_pseudo(iph->saddr, iph->daddr,
264 htobe16(pseudolen));
265 /* Compute the checksum on TCP/UDP header + payload
266 * (includes the pseudo-header).
267 */
268 *check = nm_os_csum_fold(nm_os_csum_raw(data, datalen, 0));
269 #else
270 static int notsupported = 0;
271 if (!notsupported) {
272 notsupported = 1;
273 nm_prerr("inet4 segmentation not supported");
274 }
275 #endif
276 }
277
278 void
nm_os_csum_tcpudp_ipv6(struct nm_ipv6hdr * ip6h,void * data,size_t datalen,uint16_t * check)279 nm_os_csum_tcpudp_ipv6(struct nm_ipv6hdr *ip6h, void *data,
280 size_t datalen, uint16_t *check)
281 {
282 #ifdef INET6
283 *check = in6_cksum_pseudo((void*)ip6h, datalen, ip6h->nexthdr, 0);
284 *check = nm_os_csum_fold(nm_os_csum_raw(data, datalen, 0));
285 #else
286 static int notsupported = 0;
287 if (!notsupported) {
288 notsupported = 1;
289 nm_prerr("inet6 segmentation not supported");
290 }
291 #endif
292 }
293
294 /* on FreeBSD we send up one packet at a time */
295 void *
nm_os_send_up(if_t ifp,struct mbuf * m,struct mbuf * prev)296 nm_os_send_up(if_t ifp, struct mbuf *m, struct mbuf *prev)
297 {
298 NA(ifp)->if_input(ifp, m);
299 return NULL;
300 }
301
302 int
nm_os_mbuf_has_csum_offld(struct mbuf * m)303 nm_os_mbuf_has_csum_offld(struct mbuf *m)
304 {
305 return m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_SCTP |
306 CSUM_TCP_IPV6 | CSUM_UDP_IPV6 |
307 CSUM_SCTP_IPV6);
308 }
309
310 int
nm_os_mbuf_has_seg_offld(struct mbuf * m)311 nm_os_mbuf_has_seg_offld(struct mbuf *m)
312 {
313 return m->m_pkthdr.csum_flags & CSUM_TSO;
314 }
315
316 static void
freebsd_generic_rx_handler(if_t ifp,struct mbuf * m)317 freebsd_generic_rx_handler(if_t ifp, struct mbuf *m)
318 {
319 int stolen;
320
321 if (unlikely(!NM_NA_VALID(ifp))) {
322 nm_prlim(1, "Warning: RX packet intercepted, but no"
323 " emulated adapter");
324 return;
325 }
326
327 do {
328 struct mbuf *n;
329
330 n = m->m_nextpkt;
331 m->m_nextpkt = NULL;
332 stolen = generic_rx_handler(ifp, m);
333 if (!stolen) {
334 NA(ifp)->if_input(ifp, m);
335 }
336 m = n;
337 } while (m != NULL);
338 }
339
340 /*
341 * Intercept the rx routine in the standard device driver.
342 * Second argument is non-zero to intercept, 0 to restore
343 */
344 int
nm_os_catch_rx(struct netmap_generic_adapter * gna,int intercept)345 nm_os_catch_rx(struct netmap_generic_adapter *gna, int intercept)
346 {
347 struct netmap_adapter *na = &gna->up.up;
348 if_t ifp = na->ifp;
349 int ret = 0;
350
351 nm_os_ifnet_lock();
352 if (intercept) {
353 if_setcapenablebit(ifp, IFCAP_NETMAP, 0);
354 if_setinputfn(ifp, freebsd_generic_rx_handler);
355 } else {
356 if_setcapenablebit(ifp, 0, IFCAP_NETMAP);
357 if_setinputfn(ifp, na->if_input);
358 }
359 nm_os_ifnet_unlock();
360
361 return ret;
362 }
363
364
365 /*
366 * Intercept the packet steering routine in the tx path,
367 * so that we can decide which queue is used for an mbuf.
368 * Second argument is non-zero to intercept, 0 to restore.
369 * On freebsd we just intercept if_transmit.
370 */
371 int
nm_os_catch_tx(struct netmap_generic_adapter * gna,int intercept)372 nm_os_catch_tx(struct netmap_generic_adapter *gna, int intercept)
373 {
374 struct netmap_adapter *na = &gna->up.up;
375 if_t ifp = netmap_generic_getifp(gna);
376
377 nm_os_ifnet_lock();
378 if (intercept) {
379 na->if_transmit = if_gettransmitfn(ifp);
380 if_settransmitfn(ifp, netmap_transmit);
381 } else {
382 if_settransmitfn(ifp, na->if_transmit);
383 }
384 nm_os_ifnet_unlock();
385
386 return 0;
387 }
388
389
390 /*
391 * Transmit routine used by generic_netmap_txsync(). Returns 0 on success
392 * and non-zero on error (which may be packet drops or other errors).
393 * addr and len identify the netmap buffer, m is the (preallocated)
394 * mbuf to use for transmissions.
395 *
396 * Zero-copy transmission is possible if netmap is attached directly to a
397 * hardware interface: when cleaning we simply wait for the mbuf cluster
398 * refcount to decrement to 1, indicating that the driver has completed
399 * transmission and is done with the buffer. However, this approach can
400 * lead to queue deadlocks when attaching to software interfaces (e.g.,
401 * if_bridge) since we cannot rely on member ports to promptly reclaim
402 * transmitted mbufs. Since there is no easy way to distinguish these
403 * cases, we currently always copy the buffer.
404 *
405 * On multiqueue cards, we can force the queue using
406 * if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
407 * i = m->m_pkthdr.flowid % adapter->num_queues;
408 * else
409 * i = curcpu % adapter->num_queues;
410 */
411 int
nm_os_generic_xmit_frame(struct nm_os_gen_arg * a)412 nm_os_generic_xmit_frame(struct nm_os_gen_arg *a)
413 {
414 int ret;
415 u_int len = a->len;
416 if_t ifp = a->ifp;
417 struct mbuf *m = a->m;
418
419 M_ASSERTPKTHDR(m);
420 KASSERT((m->m_flags & M_EXT) != 0,
421 ("%s: mbuf %p has no cluster", __func__, m));
422
423 if (MBUF_REFCNT(m) != 1) {
424 nm_prerr("invalid refcnt %d for %p", MBUF_REFCNT(m), m);
425 panic("in generic_xmit_frame");
426 }
427 if (unlikely(m->m_ext.ext_size < len)) {
428 nm_prlim(2, "size %d < len %d", m->m_ext.ext_size, len);
429 len = m->m_ext.ext_size;
430 }
431
432 m_copyback(m, 0, len, a->addr);
433 m->m_len = m->m_pkthdr.len = len;
434 SET_MBUF_REFCNT(m, 2);
435 M_HASHTYPE_SET(m, M_HASHTYPE_OPAQUE);
436 m->m_pkthdr.flowid = a->ring_nr;
437 m->m_pkthdr.rcvif = ifp; /* used for tx notification */
438 CURVNET_SET(if_getvnet(ifp));
439 ret = NA(ifp)->if_transmit(ifp, m);
440 CURVNET_RESTORE();
441 return ret ? -1 : 0;
442 }
443
444 struct netmap_adapter *
netmap_getna(if_t ifp)445 netmap_getna(if_t ifp)
446 {
447 return (NA(ifp));
448 }
449
450 /*
451 * The following two functions are empty until we have a generic
452 * way to extract the info from the ifp
453 */
454 int
nm_os_generic_find_num_desc(if_t ifp,unsigned int * tx,unsigned int * rx)455 nm_os_generic_find_num_desc(if_t ifp, unsigned int *tx, unsigned int *rx)
456 {
457 return 0;
458 }
459
460
461 void
nm_os_generic_find_num_queues(if_t ifp,u_int * txq,u_int * rxq)462 nm_os_generic_find_num_queues(if_t ifp, u_int *txq, u_int *rxq)
463 {
464 unsigned num_rings = netmap_generic_rings ? netmap_generic_rings : 1;
465
466 *txq = num_rings;
467 *rxq = num_rings;
468 }
469
470 void
nm_os_generic_set_features(struct netmap_generic_adapter * gna)471 nm_os_generic_set_features(struct netmap_generic_adapter *gna)
472 {
473
474 gna->rxsg = 1; /* Supported through m_copydata. */
475 gna->txqdisc = 0; /* Not supported. */
476 }
477
478 void
nm_os_mitigation_init(struct nm_generic_mit * mit,int idx,struct netmap_adapter * na)479 nm_os_mitigation_init(struct nm_generic_mit *mit, int idx, struct netmap_adapter *na)
480 {
481 mit->mit_pending = 0;
482 mit->mit_ring_idx = idx;
483 mit->mit_na = na;
484 }
485
486
487 void
nm_os_mitigation_start(struct nm_generic_mit * mit)488 nm_os_mitigation_start(struct nm_generic_mit *mit)
489 {
490 }
491
492
493 void
nm_os_mitigation_restart(struct nm_generic_mit * mit)494 nm_os_mitigation_restart(struct nm_generic_mit *mit)
495 {
496 }
497
498
499 int
nm_os_mitigation_active(struct nm_generic_mit * mit)500 nm_os_mitigation_active(struct nm_generic_mit *mit)
501 {
502
503 return 0;
504 }
505
506
507 void
nm_os_mitigation_cleanup(struct nm_generic_mit * mit)508 nm_os_mitigation_cleanup(struct nm_generic_mit *mit)
509 {
510 }
511
512 static int
nm_vi_dummy(if_t ifp,u_long cmd,caddr_t addr)513 nm_vi_dummy(if_t ifp, u_long cmd, caddr_t addr)
514 {
515
516 return EINVAL;
517 }
518
519 static void
nm_vi_start(if_t ifp)520 nm_vi_start(if_t ifp)
521 {
522 panic("nm_vi_start() must not be called");
523 }
524
525 /*
526 * Index manager of persistent virtual interfaces.
527 * It is used to decide the lowest byte of the MAC address.
528 * We use the same algorithm with management of bridge port index.
529 */
530 #define NM_VI_MAX 255
531 static struct {
532 uint8_t index[NM_VI_MAX]; /* XXX just for a reasonable number */
533 uint8_t active;
534 struct mtx lock;
535 } nm_vi_indices;
536
537 void
nm_os_vi_init_index(void)538 nm_os_vi_init_index(void)
539 {
540 int i;
541 for (i = 0; i < NM_VI_MAX; i++)
542 nm_vi_indices.index[i] = i;
543 nm_vi_indices.active = 0;
544 mtx_init(&nm_vi_indices.lock, "nm_vi_indices_lock", NULL, MTX_DEF);
545 }
546
547 /* return -1 if no index available */
548 static int
nm_vi_get_index(void)549 nm_vi_get_index(void)
550 {
551 int ret;
552
553 mtx_lock(&nm_vi_indices.lock);
554 ret = nm_vi_indices.active == NM_VI_MAX ? -1 :
555 nm_vi_indices.index[nm_vi_indices.active++];
556 mtx_unlock(&nm_vi_indices.lock);
557 return ret;
558 }
559
560 static void
nm_vi_free_index(uint8_t val)561 nm_vi_free_index(uint8_t val)
562 {
563 int i, lim;
564
565 mtx_lock(&nm_vi_indices.lock);
566 lim = nm_vi_indices.active;
567 for (i = 0; i < lim; i++) {
568 if (nm_vi_indices.index[i] == val) {
569 /* swap index[lim-1] and j */
570 int tmp = nm_vi_indices.index[lim-1];
571 nm_vi_indices.index[lim-1] = val;
572 nm_vi_indices.index[i] = tmp;
573 nm_vi_indices.active--;
574 break;
575 }
576 }
577 if (lim == nm_vi_indices.active)
578 nm_prerr("Index %u not found", val);
579 mtx_unlock(&nm_vi_indices.lock);
580 }
581 #undef NM_VI_MAX
582
583 /*
584 * Implementation of a netmap-capable virtual interface that
585 * registered to the system.
586 * It is based on if_tap.c and ip_fw_log.c in FreeBSD 9.
587 *
588 * Note: Linux sets refcount to 0 on allocation of net_device,
589 * then increments it on registration to the system.
590 * FreeBSD sets refcount to 1 on if_alloc(), and does not
591 * increment this refcount on if_attach().
592 */
593 int
nm_os_vi_persist(const char * name,if_t * ret)594 nm_os_vi_persist(const char *name, if_t *ret)
595 {
596 if_t ifp;
597 u_short macaddr_hi;
598 uint32_t macaddr_mid;
599 u_char eaddr[6];
600 int unit = nm_vi_get_index(); /* just to decide MAC address */
601
602 if (unit < 0)
603 return EBUSY;
604 /*
605 * We use the same MAC address generation method with tap
606 * except for the highest octet is 00:be instead of 00:bd
607 */
608 macaddr_hi = htons(0x00be); /* XXX tap + 1 */
609 macaddr_mid = (uint32_t) ticks;
610 bcopy(&macaddr_hi, eaddr, sizeof(short));
611 bcopy(&macaddr_mid, &eaddr[2], sizeof(uint32_t));
612 eaddr[5] = (uint8_t)unit;
613
614 ifp = if_alloc(IFT_ETHER);
615 if (ifp == NULL) {
616 nm_prerr("if_alloc failed");
617 return ENOMEM;
618 }
619 if_initname(ifp, name, IF_DUNIT_NONE);
620 if_setflags(ifp, IFF_UP | IFF_SIMPLEX | IFF_MULTICAST);
621 if_setinitfn(ifp, (void *)nm_vi_dummy);
622 if_setioctlfn(ifp, nm_vi_dummy);
623 if_setstartfn(ifp, nm_vi_start);
624 if_setmtu(ifp, ETHERMTU);
625 if_setsendqlen(ifp, ifqmaxlen);
626 if_setcapabilitiesbit(ifp, IFCAP_LINKSTATE, 0);
627 if_setcapenablebit(ifp, IFCAP_LINKSTATE, 0);
628
629 ether_ifattach(ifp, eaddr);
630 *ret = ifp;
631 return 0;
632 }
633
634 /* unregister from the system and drop the final refcount */
635 void
nm_os_vi_detach(if_t ifp)636 nm_os_vi_detach(if_t ifp)
637 {
638 nm_vi_free_index(((char *)if_getlladdr(ifp))[5]);
639 ether_ifdetach(ifp);
640 if_free(ifp);
641 }
642
643 #ifdef WITH_EXTMEM
644 #include <vm/vm_map.h>
645 #include <vm/vm_extern.h>
646 #include <vm/vm_kern.h>
647 struct nm_os_extmem {
648 vm_object_t obj;
649 vm_offset_t kva;
650 vm_offset_t size;
651 uintptr_t scan;
652 };
653
654 void
nm_os_extmem_delete(struct nm_os_extmem * e)655 nm_os_extmem_delete(struct nm_os_extmem *e)
656 {
657 nm_prinf("freeing %zx bytes", (size_t)e->size);
658 vm_map_remove(kernel_map, e->kva, e->kva + e->size);
659 nm_os_free(e);
660 }
661
662 char *
nm_os_extmem_nextpage(struct nm_os_extmem * e)663 nm_os_extmem_nextpage(struct nm_os_extmem *e)
664 {
665 char *rv = NULL;
666 if (e->scan < e->kva + e->size) {
667 rv = (char *)e->scan;
668 e->scan += PAGE_SIZE;
669 }
670 return rv;
671 }
672
673 int
nm_os_extmem_isequal(struct nm_os_extmem * e1,struct nm_os_extmem * e2)674 nm_os_extmem_isequal(struct nm_os_extmem *e1, struct nm_os_extmem *e2)
675 {
676 return (e1->obj == e2->obj);
677 }
678
679 int
nm_os_extmem_nr_pages(struct nm_os_extmem * e)680 nm_os_extmem_nr_pages(struct nm_os_extmem *e)
681 {
682 return e->size >> PAGE_SHIFT;
683 }
684
685 struct nm_os_extmem *
nm_os_extmem_create(unsigned long p,struct nmreq_pools_info * pi,int * perror)686 nm_os_extmem_create(unsigned long p, struct nmreq_pools_info *pi, int *perror)
687 {
688 vm_map_t map;
689 vm_map_entry_t entry;
690 vm_object_t obj;
691 vm_prot_t prot;
692 vm_pindex_t index;
693 boolean_t wired;
694 struct nm_os_extmem *e = NULL;
695 int rv, error = 0;
696
697 e = nm_os_malloc(sizeof(*e));
698 if (e == NULL) {
699 error = ENOMEM;
700 goto out;
701 }
702
703 map = &curthread->td_proc->p_vmspace->vm_map;
704 rv = vm_map_lookup(&map, p, VM_PROT_RW, &entry,
705 &obj, &index, &prot, &wired);
706 if (rv != KERN_SUCCESS) {
707 nm_prerr("address %lx not found", p);
708 error = vm_mmap_to_errno(rv);
709 goto out_free;
710 }
711 vm_object_reference(obj);
712
713 /* check that we are given the whole vm_object ? */
714 vm_map_lookup_done(map, entry);
715
716 e->obj = obj;
717 /* Wire the memory and add the vm_object to the kernel map,
718 * to make sure that it is not freed even if all the processes
719 * that are mmap()ing should munmap() it.
720 */
721 e->kva = vm_map_min(kernel_map);
722 e->size = obj->size << PAGE_SHIFT;
723 rv = vm_map_find(kernel_map, obj, 0, &e->kva, e->size, 0,
724 VMFS_OPTIMAL_SPACE, VM_PROT_READ | VM_PROT_WRITE,
725 VM_PROT_READ | VM_PROT_WRITE, 0);
726 if (rv != KERN_SUCCESS) {
727 nm_prerr("vm_map_find(%zx) failed", (size_t)e->size);
728 error = vm_mmap_to_errno(rv);
729 goto out_rel;
730 }
731 rv = vm_map_wire(kernel_map, e->kva, e->kva + e->size,
732 VM_MAP_WIRE_SYSTEM | VM_MAP_WIRE_NOHOLES);
733 if (rv != KERN_SUCCESS) {
734 nm_prerr("vm_map_wire failed");
735 error = vm_mmap_to_errno(rv);
736 goto out_rem;
737 }
738
739 e->scan = e->kva;
740
741 return e;
742
743 out_rem:
744 vm_map_remove(kernel_map, e->kva, e->kva + e->size);
745 out_rel:
746 vm_object_deallocate(e->obj);
747 e->obj = NULL;
748 out_free:
749 nm_os_free(e);
750 out:
751 if (perror)
752 *perror = error;
753 return NULL;
754 }
755 #endif /* WITH_EXTMEM */
756
757 /* ================== PTNETMAP GUEST SUPPORT ==================== */
758
759 #ifdef WITH_PTNETMAP
760 #include <sys/bus.h>
761 #include <sys/rman.h>
762 #include <machine/bus.h> /* bus_dmamap_* */
763 #include <machine/resource.h>
764 #include <dev/pci/pcivar.h>
765 #include <dev/pci/pcireg.h>
766 /*
767 * ptnetmap memory device (memdev) for freebsd guest,
768 * ssed to expose host netmap memory to the guest through a PCI BAR.
769 */
770
771 /*
772 * ptnetmap memdev private data structure
773 */
774 struct ptnetmap_memdev {
775 device_t dev;
776 struct resource *pci_io;
777 struct resource *pci_mem;
778 struct netmap_mem_d *nm_mem;
779 };
780
781 static int ptn_memdev_probe(device_t);
782 static int ptn_memdev_attach(device_t);
783 static int ptn_memdev_detach(device_t);
784 static int ptn_memdev_shutdown(device_t);
785
786 static device_method_t ptn_memdev_methods[] = {
787 DEVMETHOD(device_probe, ptn_memdev_probe),
788 DEVMETHOD(device_attach, ptn_memdev_attach),
789 DEVMETHOD(device_detach, ptn_memdev_detach),
790 DEVMETHOD(device_shutdown, ptn_memdev_shutdown),
791 DEVMETHOD_END
792 };
793
794 static driver_t ptn_memdev_driver = {
795 PTNETMAP_MEMDEV_NAME,
796 ptn_memdev_methods,
797 sizeof(struct ptnetmap_memdev),
798 };
799
800 /* We use (SI_ORDER_MIDDLE+1) here, see DEV_MODULE_ORDERED() invocation
801 * below. */
802 DRIVER_MODULE_ORDERED(ptn_memdev, pci, ptn_memdev_driver, NULL, NULL,
803 SI_ORDER_MIDDLE + 1);
804
805 /*
806 * Map host netmap memory through PCI-BAR in the guest OS,
807 * returning physical (nm_paddr) and virtual (nm_addr) addresses
808 * of the netmap memory mapped in the guest.
809 */
810 int
nm_os_pt_memdev_iomap(struct ptnetmap_memdev * ptn_dev,vm_paddr_t * nm_paddr,void ** nm_addr,uint64_t * mem_size)811 nm_os_pt_memdev_iomap(struct ptnetmap_memdev *ptn_dev, vm_paddr_t *nm_paddr,
812 void **nm_addr, uint64_t *mem_size)
813 {
814 int rid;
815
816 nm_prinf("ptn_memdev_driver iomap");
817
818 rid = PCIR_BAR(PTNETMAP_MEM_PCI_BAR);
819 *mem_size = bus_read_4(ptn_dev->pci_io, PTNET_MDEV_IO_MEMSIZE_HI);
820 *mem_size = bus_read_4(ptn_dev->pci_io, PTNET_MDEV_IO_MEMSIZE_LO) |
821 (*mem_size << 32);
822
823 /* map memory allocator */
824 ptn_dev->pci_mem = bus_alloc_resource(ptn_dev->dev, SYS_RES_MEMORY,
825 &rid, 0, ~0, *mem_size, RF_ACTIVE);
826 if (ptn_dev->pci_mem == NULL) {
827 *nm_paddr = 0;
828 *nm_addr = NULL;
829 return ENOMEM;
830 }
831
832 *nm_paddr = rman_get_start(ptn_dev->pci_mem);
833 *nm_addr = rman_get_virtual(ptn_dev->pci_mem);
834
835 nm_prinf("=== BAR %d start %lx len %lx mem_size %lx ===",
836 PTNETMAP_MEM_PCI_BAR,
837 (unsigned long)(*nm_paddr),
838 (unsigned long)rman_get_size(ptn_dev->pci_mem),
839 (unsigned long)*mem_size);
840 return (0);
841 }
842
843 uint32_t
nm_os_pt_memdev_ioread(struct ptnetmap_memdev * ptn_dev,unsigned int reg)844 nm_os_pt_memdev_ioread(struct ptnetmap_memdev *ptn_dev, unsigned int reg)
845 {
846 return bus_read_4(ptn_dev->pci_io, reg);
847 }
848
849 /* Unmap host netmap memory. */
850 void
nm_os_pt_memdev_iounmap(struct ptnetmap_memdev * ptn_dev)851 nm_os_pt_memdev_iounmap(struct ptnetmap_memdev *ptn_dev)
852 {
853 nm_prinf("ptn_memdev_driver iounmap");
854
855 if (ptn_dev->pci_mem) {
856 bus_release_resource(ptn_dev->dev, SYS_RES_MEMORY,
857 PCIR_BAR(PTNETMAP_MEM_PCI_BAR), ptn_dev->pci_mem);
858 ptn_dev->pci_mem = NULL;
859 }
860 }
861
862 /* Device identification routine, return BUS_PROBE_DEFAULT on success,
863 * positive on failure */
864 static int
ptn_memdev_probe(device_t dev)865 ptn_memdev_probe(device_t dev)
866 {
867 char desc[256];
868
869 if (pci_get_vendor(dev) != PTNETMAP_PCI_VENDOR_ID)
870 return (ENXIO);
871 if (pci_get_device(dev) != PTNETMAP_PCI_DEVICE_ID)
872 return (ENXIO);
873
874 snprintf(desc, sizeof(desc), "%s PCI adapter",
875 PTNETMAP_MEMDEV_NAME);
876 device_set_desc_copy(dev, desc);
877
878 return (BUS_PROBE_DEFAULT);
879 }
880
881 /* Device initialization routine. */
882 static int
ptn_memdev_attach(device_t dev)883 ptn_memdev_attach(device_t dev)
884 {
885 struct ptnetmap_memdev *ptn_dev;
886 int rid;
887 uint16_t mem_id;
888
889 ptn_dev = device_get_softc(dev);
890 ptn_dev->dev = dev;
891
892 pci_enable_busmaster(dev);
893
894 rid = PCIR_BAR(PTNETMAP_IO_PCI_BAR);
895 ptn_dev->pci_io = bus_alloc_resource_any(dev, SYS_RES_IOPORT, &rid,
896 RF_ACTIVE);
897 if (ptn_dev->pci_io == NULL) {
898 device_printf(dev, "cannot map I/O space\n");
899 return (ENXIO);
900 }
901
902 mem_id = bus_read_4(ptn_dev->pci_io, PTNET_MDEV_IO_MEMID);
903
904 /* create guest allocator */
905 ptn_dev->nm_mem = netmap_mem_pt_guest_attach(ptn_dev, mem_id);
906 if (ptn_dev->nm_mem == NULL) {
907 ptn_memdev_detach(dev);
908 return (ENOMEM);
909 }
910 netmap_mem_get(ptn_dev->nm_mem);
911
912 nm_prinf("ptnetmap memdev attached, host memid: %u", mem_id);
913
914 return (0);
915 }
916
917 /* Device removal routine. */
918 static int
ptn_memdev_detach(device_t dev)919 ptn_memdev_detach(device_t dev)
920 {
921 struct ptnetmap_memdev *ptn_dev;
922
923 ptn_dev = device_get_softc(dev);
924
925 if (ptn_dev->nm_mem) {
926 nm_prinf("ptnetmap memdev detached, host memid %u",
927 netmap_mem_get_id(ptn_dev->nm_mem));
928 netmap_mem_put(ptn_dev->nm_mem);
929 ptn_dev->nm_mem = NULL;
930 }
931 if (ptn_dev->pci_mem) {
932 bus_release_resource(dev, SYS_RES_MEMORY,
933 PCIR_BAR(PTNETMAP_MEM_PCI_BAR), ptn_dev->pci_mem);
934 ptn_dev->pci_mem = NULL;
935 }
936 if (ptn_dev->pci_io) {
937 bus_release_resource(dev, SYS_RES_IOPORT,
938 PCIR_BAR(PTNETMAP_IO_PCI_BAR), ptn_dev->pci_io);
939 ptn_dev->pci_io = NULL;
940 }
941
942 return (0);
943 }
944
945 static int
ptn_memdev_shutdown(device_t dev)946 ptn_memdev_shutdown(device_t dev)
947 {
948 return bus_generic_shutdown(dev);
949 }
950
951 #endif /* WITH_PTNETMAP */
952
953 /*
954 * In order to track whether pages are still mapped, we hook into
955 * the standard cdev_pager and intercept the constructor and
956 * destructor.
957 */
958
959 struct netmap_vm_handle_t {
960 struct cdev *dev;
961 struct netmap_priv_d *priv;
962 };
963
964
965 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)966 netmap_dev_pager_ctor(void *handle, vm_ooffset_t size, vm_prot_t prot,
967 vm_ooffset_t foff, struct ucred *cred, u_short *color)
968 {
969 struct netmap_vm_handle_t *vmh = handle;
970
971 if (netmap_verbose)
972 nm_prinf("handle %p size %jd prot %d foff %jd",
973 handle, (intmax_t)size, prot, (intmax_t)foff);
974 if (color)
975 *color = 0;
976 dev_ref(vmh->dev);
977 return 0;
978 }
979
980
981 static void
netmap_dev_pager_dtor(void * handle)982 netmap_dev_pager_dtor(void *handle)
983 {
984 struct netmap_vm_handle_t *vmh = handle;
985 struct cdev *dev = vmh->dev;
986 struct netmap_priv_d *priv = vmh->priv;
987
988 if (netmap_verbose)
989 nm_prinf("handle %p", handle);
990 netmap_dtor(priv);
991 free(vmh, M_DEVBUF);
992 dev_rel(dev);
993 }
994
995
996 static int
netmap_dev_pager_fault(vm_object_t object,vm_ooffset_t offset,int prot,vm_page_t * mres)997 netmap_dev_pager_fault(vm_object_t object, vm_ooffset_t offset,
998 int prot, vm_page_t *mres)
999 {
1000 struct netmap_vm_handle_t *vmh = object->handle;
1001 struct netmap_priv_d *priv = vmh->priv;
1002 struct netmap_adapter *na = priv->np_na;
1003 vm_paddr_t paddr;
1004 vm_page_t page;
1005 vm_memattr_t memattr;
1006
1007 nm_prdis("object %p offset %jd prot %d mres %p",
1008 object, (intmax_t)offset, prot, mres);
1009 memattr = object->memattr;
1010 paddr = netmap_mem_ofstophys(na->nm_mem, offset);
1011 if (paddr == 0)
1012 return VM_PAGER_FAIL;
1013
1014 if (((*mres)->flags & PG_FICTITIOUS) != 0) {
1015 /*
1016 * If the passed in result page is a fake page, update it with
1017 * the new physical address.
1018 */
1019 page = *mres;
1020 vm_page_updatefake(page, paddr, memattr);
1021 } else {
1022 /*
1023 * Replace the passed in reqpage page with our own fake page and
1024 * free up the all of the original pages.
1025 */
1026 VM_OBJECT_WUNLOCK(object);
1027 page = vm_page_getfake(paddr, memattr);
1028 VM_OBJECT_WLOCK(object);
1029 vm_page_replace(page, object, (*mres)->pindex, *mres);
1030 *mres = page;
1031 }
1032 page->valid = VM_PAGE_BITS_ALL;
1033 return (VM_PAGER_OK);
1034 }
1035
1036
1037 static struct cdev_pager_ops netmap_cdev_pager_ops = {
1038 .cdev_pg_ctor = netmap_dev_pager_ctor,
1039 .cdev_pg_dtor = netmap_dev_pager_dtor,
1040 .cdev_pg_fault = netmap_dev_pager_fault,
1041 };
1042
1043
1044 static int
netmap_mmap_single(struct cdev * cdev,vm_ooffset_t * foff,vm_size_t objsize,vm_object_t * objp,int prot)1045 netmap_mmap_single(struct cdev *cdev, vm_ooffset_t *foff,
1046 vm_size_t objsize, vm_object_t *objp, int prot)
1047 {
1048 int error;
1049 struct netmap_vm_handle_t *vmh;
1050 struct netmap_priv_d *priv;
1051 vm_object_t obj;
1052
1053 if (netmap_verbose)
1054 nm_prinf("cdev %p foff %jd size %jd objp %p prot %d", cdev,
1055 (intmax_t )*foff, (intmax_t )objsize, objp, prot);
1056
1057 vmh = malloc(sizeof(struct netmap_vm_handle_t), M_DEVBUF,
1058 M_NOWAIT | M_ZERO);
1059 if (vmh == NULL)
1060 return ENOMEM;
1061 vmh->dev = cdev;
1062
1063 NMG_LOCK();
1064 error = devfs_get_cdevpriv((void**)&priv);
1065 if (error)
1066 goto err_unlock;
1067 if (priv->np_nifp == NULL) {
1068 error = EINVAL;
1069 goto err_unlock;
1070 }
1071 vmh->priv = priv;
1072 priv->np_refs++;
1073 NMG_UNLOCK();
1074
1075 obj = cdev_pager_allocate(vmh, OBJT_DEVICE,
1076 &netmap_cdev_pager_ops, objsize, prot,
1077 *foff, NULL);
1078 if (obj == NULL) {
1079 nm_prerr("cdev_pager_allocate failed");
1080 error = EINVAL;
1081 goto err_deref;
1082 }
1083
1084 *objp = obj;
1085 return 0;
1086
1087 err_deref:
1088 NMG_LOCK();
1089 priv->np_refs--;
1090 err_unlock:
1091 NMG_UNLOCK();
1092 // err:
1093 free(vmh, M_DEVBUF);
1094 return error;
1095 }
1096
1097 /*
1098 * On FreeBSD the close routine is only called on the last close on
1099 * the device (/dev/netmap) so we cannot do anything useful.
1100 * To track close() on individual file descriptors we pass netmap_dtor() to
1101 * devfs_set_cdevpriv() on open(). The FreeBSD kernel will call the destructor
1102 * when the last fd pointing to the device is closed.
1103 *
1104 * Note that FreeBSD does not even munmap() on close() so we also have
1105 * to track mmap() ourselves, and postpone the call to
1106 * netmap_dtor() is called when the process has no open fds and no active
1107 * memory maps on /dev/netmap, as in linux.
1108 */
1109 static int
netmap_close(struct cdev * dev,int fflag,int devtype,struct thread * td)1110 netmap_close(struct cdev *dev, int fflag, int devtype, struct thread *td)
1111 {
1112 if (netmap_verbose)
1113 nm_prinf("dev %p fflag 0x%x devtype %d td %p",
1114 dev, fflag, devtype, td);
1115 return 0;
1116 }
1117
1118
1119 static int
netmap_open(struct cdev * dev,int oflags,int devtype,struct thread * td)1120 netmap_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
1121 {
1122 struct netmap_priv_d *priv;
1123 int error;
1124
1125 (void)dev;
1126 (void)oflags;
1127 (void)devtype;
1128 (void)td;
1129
1130 NMG_LOCK();
1131 priv = netmap_priv_new();
1132 if (priv == NULL) {
1133 error = ENOMEM;
1134 goto out;
1135 }
1136 error = devfs_set_cdevpriv(priv, netmap_dtor);
1137 if (error) {
1138 netmap_priv_delete(priv);
1139 }
1140 out:
1141 NMG_UNLOCK();
1142 return error;
1143 }
1144
1145 /******************** kthread wrapper ****************/
1146 #include <sys/sysproto.h>
1147 u_int
nm_os_ncpus(void)1148 nm_os_ncpus(void)
1149 {
1150 return mp_maxid + 1;
1151 }
1152
1153 struct nm_kctx_ctx {
1154 /* Userspace thread (kthread creator). */
1155 struct thread *user_td;
1156
1157 /* worker function and parameter */
1158 nm_kctx_worker_fn_t worker_fn;
1159 void *worker_private;
1160
1161 struct nm_kctx *nmk;
1162
1163 /* integer to manage multiple worker contexts (e.g., RX or TX on ptnetmap) */
1164 long type;
1165 };
1166
1167 struct nm_kctx {
1168 struct thread *worker;
1169 struct mtx worker_lock;
1170 struct nm_kctx_ctx worker_ctx;
1171 int run; /* used to stop kthread */
1172 int attach_user; /* kthread attached to user_process */
1173 int affinity;
1174 };
1175
1176 static void
nm_kctx_worker(void * data)1177 nm_kctx_worker(void *data)
1178 {
1179 struct nm_kctx *nmk = data;
1180 struct nm_kctx_ctx *ctx = &nmk->worker_ctx;
1181
1182 if (nmk->affinity >= 0) {
1183 thread_lock(curthread);
1184 sched_bind(curthread, nmk->affinity);
1185 thread_unlock(curthread);
1186 }
1187
1188 while (nmk->run) {
1189 /*
1190 * check if the parent process dies
1191 * (when kthread is attached to user process)
1192 */
1193 if (ctx->user_td) {
1194 PROC_LOCK(curproc);
1195 thread_suspend_check(0);
1196 PROC_UNLOCK(curproc);
1197 } else {
1198 kthread_suspend_check();
1199 }
1200
1201 /* Continuously execute worker process. */
1202 ctx->worker_fn(ctx->worker_private); /* worker body */
1203 }
1204
1205 kthread_exit();
1206 }
1207
1208 void
nm_os_kctx_worker_setaff(struct nm_kctx * nmk,int affinity)1209 nm_os_kctx_worker_setaff(struct nm_kctx *nmk, int affinity)
1210 {
1211 nmk->affinity = affinity;
1212 }
1213
1214 struct nm_kctx *
nm_os_kctx_create(struct nm_kctx_cfg * cfg,void * opaque)1215 nm_os_kctx_create(struct nm_kctx_cfg *cfg, void *opaque)
1216 {
1217 struct nm_kctx *nmk = NULL;
1218
1219 nmk = malloc(sizeof(*nmk), M_DEVBUF, M_NOWAIT | M_ZERO);
1220 if (!nmk)
1221 return NULL;
1222
1223 mtx_init(&nmk->worker_lock, "nm_kthread lock", NULL, MTX_DEF);
1224 nmk->worker_ctx.worker_fn = cfg->worker_fn;
1225 nmk->worker_ctx.worker_private = cfg->worker_private;
1226 nmk->worker_ctx.type = cfg->type;
1227 nmk->affinity = -1;
1228
1229 /* attach kthread to user process (ptnetmap) */
1230 nmk->attach_user = cfg->attach_user;
1231
1232 return nmk;
1233 }
1234
1235 int
nm_os_kctx_worker_start(struct nm_kctx * nmk)1236 nm_os_kctx_worker_start(struct nm_kctx *nmk)
1237 {
1238 struct proc *p = NULL;
1239 int error = 0;
1240
1241 /* Temporarily disable this function as it is currently broken
1242 * and causes kernel crashes. The failure can be triggered by
1243 * the "vale_polling_enable_disable" test in ctrl-api-test.c. */
1244 return EOPNOTSUPP;
1245
1246 if (nmk->worker)
1247 return EBUSY;
1248
1249 /* check if we want to attach kthread to user process */
1250 if (nmk->attach_user) {
1251 nmk->worker_ctx.user_td = curthread;
1252 p = curthread->td_proc;
1253 }
1254
1255 /* enable kthread main loop */
1256 nmk->run = 1;
1257 /* create kthread */
1258 if((error = kthread_add(nm_kctx_worker, nmk, p,
1259 &nmk->worker, RFNOWAIT /* to be checked */, 0, "nm-kthread-%ld",
1260 nmk->worker_ctx.type))) {
1261 goto err;
1262 }
1263
1264 nm_prinf("nm_kthread started td %p", nmk->worker);
1265
1266 return 0;
1267 err:
1268 nm_prerr("nm_kthread start failed err %d", error);
1269 nmk->worker = NULL;
1270 return error;
1271 }
1272
1273 void
nm_os_kctx_worker_stop(struct nm_kctx * nmk)1274 nm_os_kctx_worker_stop(struct nm_kctx *nmk)
1275 {
1276 if (!nmk->worker)
1277 return;
1278
1279 /* tell to kthread to exit from main loop */
1280 nmk->run = 0;
1281
1282 /* wake up kthread if it sleeps */
1283 kthread_resume(nmk->worker);
1284
1285 nmk->worker = NULL;
1286 }
1287
1288 void
nm_os_kctx_destroy(struct nm_kctx * nmk)1289 nm_os_kctx_destroy(struct nm_kctx *nmk)
1290 {
1291 if (!nmk)
1292 return;
1293
1294 if (nmk->worker)
1295 nm_os_kctx_worker_stop(nmk);
1296
1297 free(nmk, M_DEVBUF);
1298 }
1299
1300 /******************** kqueue support ****************/
1301
1302 /*
1303 * In addition to calling selwakeuppri(), nm_os_selwakeup() also
1304 * needs to call knote() to wake up kqueue listeners.
1305 * This operation is deferred to a taskqueue in order to avoid possible
1306 * lock order reversals; these may happen because knote() grabs a
1307 * private lock associated to the 'si' (see struct selinfo,
1308 * struct nm_selinfo, and nm_os_selinfo_init), and nm_os_selwakeup()
1309 * can be called while holding the lock associated to a different
1310 * 'si'.
1311 * When calling knote() we use a non-zero 'hint' argument to inform
1312 * the netmap_knrw() function that it is being called from
1313 * 'nm_os_selwakeup'; this is necessary because when netmap_knrw() is
1314 * called by the kevent subsystem (i.e. kevent_scan()) we also need to
1315 * call netmap_poll().
1316 *
1317 * The netmap_kqfilter() function registers one or another f_event
1318 * depending on read or write mode. A pointer to the struct
1319 * 'netmap_priv_d' is stored into kn->kn_hook, so that it can later
1320 * be passed to netmap_poll(). We pass NULL as a third argument to
1321 * netmap_poll(), so that the latter only runs the txsync/rxsync
1322 * (if necessary), and skips the nm_os_selrecord() calls.
1323 */
1324
1325
1326 void
nm_os_selwakeup(struct nm_selinfo * si)1327 nm_os_selwakeup(struct nm_selinfo *si)
1328 {
1329 selwakeuppri(&si->si, PI_NET);
1330 if (si->kqueue_users > 0) {
1331 taskqueue_enqueue(si->ntfytq, &si->ntfytask);
1332 }
1333 }
1334
1335 void
nm_os_selrecord(struct thread * td,struct nm_selinfo * si)1336 nm_os_selrecord(struct thread *td, struct nm_selinfo *si)
1337 {
1338 selrecord(td, &si->si);
1339 }
1340
1341 static void
netmap_knrdetach(struct knote * kn)1342 netmap_knrdetach(struct knote *kn)
1343 {
1344 struct netmap_priv_d *priv = (struct netmap_priv_d *)kn->kn_hook;
1345 struct nm_selinfo *si = priv->np_si[NR_RX];
1346
1347 knlist_remove(&si->si.si_note, kn, /*islocked=*/0);
1348 NMG_LOCK();
1349 KASSERT(si->kqueue_users > 0, ("kqueue_user underflow on %s",
1350 si->mtxname));
1351 si->kqueue_users--;
1352 nm_prinf("kqueue users for %s: %d", si->mtxname, si->kqueue_users);
1353 NMG_UNLOCK();
1354 }
1355
1356 static void
netmap_knwdetach(struct knote * kn)1357 netmap_knwdetach(struct knote *kn)
1358 {
1359 struct netmap_priv_d *priv = (struct netmap_priv_d *)kn->kn_hook;
1360 struct nm_selinfo *si = priv->np_si[NR_TX];
1361
1362 knlist_remove(&si->si.si_note, kn, /*islocked=*/0);
1363 NMG_LOCK();
1364 si->kqueue_users--;
1365 nm_prinf("kqueue users for %s: %d", si->mtxname, si->kqueue_users);
1366 NMG_UNLOCK();
1367 }
1368
1369 /*
1370 * Callback triggered by netmap notifications (see netmap_notify()),
1371 * and by the application calling kevent(). In the former case we
1372 * just return 1 (events ready), since we are not able to do better.
1373 * In the latter case we use netmap_poll() to see which events are
1374 * ready.
1375 */
1376 static int
netmap_knrw(struct knote * kn,long hint,int events)1377 netmap_knrw(struct knote *kn, long hint, int events)
1378 {
1379 struct netmap_priv_d *priv;
1380 int revents;
1381
1382 if (hint != 0) {
1383 /* Called from netmap_notify(), typically from a
1384 * thread different from the one issuing kevent().
1385 * Assume we are ready. */
1386 return 1;
1387 }
1388
1389 /* Called from kevent(). */
1390 priv = kn->kn_hook;
1391 revents = netmap_poll(priv, events, /*thread=*/NULL);
1392
1393 return (events & revents) ? 1 : 0;
1394 }
1395
1396 static int
netmap_knread(struct knote * kn,long hint)1397 netmap_knread(struct knote *kn, long hint)
1398 {
1399 return netmap_knrw(kn, hint, POLLIN);
1400 }
1401
1402 static int
netmap_knwrite(struct knote * kn,long hint)1403 netmap_knwrite(struct knote *kn, long hint)
1404 {
1405 return netmap_knrw(kn, hint, POLLOUT);
1406 }
1407
1408 static struct filterops netmap_rfiltops = {
1409 .f_isfd = 1,
1410 .f_detach = netmap_knrdetach,
1411 .f_event = netmap_knread,
1412 };
1413
1414 static struct filterops netmap_wfiltops = {
1415 .f_isfd = 1,
1416 .f_detach = netmap_knwdetach,
1417 .f_event = netmap_knwrite,
1418 };
1419
1420
1421 /*
1422 * This is called when a thread invokes kevent() to record
1423 * a change in the configuration of the kqueue().
1424 * The 'priv' is the one associated to the open netmap device.
1425 */
1426 static int
netmap_kqfilter(struct cdev * dev,struct knote * kn)1427 netmap_kqfilter(struct cdev *dev, struct knote *kn)
1428 {
1429 struct netmap_priv_d *priv;
1430 int error;
1431 struct netmap_adapter *na;
1432 struct nm_selinfo *si;
1433 int ev = kn->kn_filter;
1434
1435 if (ev != EVFILT_READ && ev != EVFILT_WRITE) {
1436 nm_prerr("bad filter request %d", ev);
1437 return 1;
1438 }
1439 error = devfs_get_cdevpriv((void**)&priv);
1440 if (error) {
1441 nm_prerr("device not yet setup");
1442 return 1;
1443 }
1444 na = priv->np_na;
1445 if (na == NULL) {
1446 nm_prerr("no netmap adapter for this file descriptor");
1447 return 1;
1448 }
1449 /* the si is indicated in the priv */
1450 si = priv->np_si[(ev == EVFILT_WRITE) ? NR_TX : NR_RX];
1451 kn->kn_fop = (ev == EVFILT_WRITE) ?
1452 &netmap_wfiltops : &netmap_rfiltops;
1453 kn->kn_hook = priv;
1454 NMG_LOCK();
1455 si->kqueue_users++;
1456 nm_prinf("kqueue users for %s: %d", si->mtxname, si->kqueue_users);
1457 NMG_UNLOCK();
1458 knlist_add(&si->si.si_note, kn, /*islocked=*/0);
1459
1460 return 0;
1461 }
1462
1463 static int
freebsd_netmap_poll(struct cdev * cdevi __unused,int events,struct thread * td)1464 freebsd_netmap_poll(struct cdev *cdevi __unused, int events, struct thread *td)
1465 {
1466 struct netmap_priv_d *priv;
1467 if (devfs_get_cdevpriv((void **)&priv)) {
1468 return POLLERR;
1469 }
1470 return netmap_poll(priv, events, td);
1471 }
1472
1473 static int
freebsd_netmap_ioctl(struct cdev * dev __unused,u_long cmd,caddr_t data,int ffla __unused,struct thread * td)1474 freebsd_netmap_ioctl(struct cdev *dev __unused, u_long cmd, caddr_t data,
1475 int ffla __unused, struct thread *td)
1476 {
1477 int error;
1478 struct netmap_priv_d *priv;
1479
1480 CURVNET_SET(TD_TO_VNET(td));
1481 error = devfs_get_cdevpriv((void **)&priv);
1482 if (error) {
1483 /* XXX ENOENT should be impossible, since the priv
1484 * is now created in the open */
1485 if (error == ENOENT)
1486 error = ENXIO;
1487 goto out;
1488 }
1489 error = netmap_ioctl(priv, cmd, data, td, /*nr_body_is_user=*/1);
1490 out:
1491 CURVNET_RESTORE();
1492
1493 return error;
1494 }
1495
1496 void
nm_os_onattach(if_t ifp)1497 nm_os_onattach(if_t ifp)
1498 {
1499 if_setcapabilitiesbit(ifp, IFCAP_NETMAP, 0);
1500 }
1501
1502 void
nm_os_onenter(if_t ifp)1503 nm_os_onenter(if_t ifp)
1504 {
1505 struct netmap_adapter *na = NA(ifp);
1506
1507 na->if_transmit = if_gettransmitfn(ifp);
1508 if_settransmitfn(ifp, netmap_transmit);
1509 if_setcapenablebit(ifp, IFCAP_NETMAP, 0);
1510 }
1511
1512 void
nm_os_onexit(if_t ifp)1513 nm_os_onexit(if_t ifp)
1514 {
1515 struct netmap_adapter *na = NA(ifp);
1516
1517 if_settransmitfn(ifp, na->if_transmit);
1518 if_setcapenablebit(ifp, 0, IFCAP_NETMAP);
1519 }
1520
1521 extern struct cdevsw netmap_cdevsw; /* XXX used in netmap.c, should go elsewhere */
1522 struct cdevsw netmap_cdevsw = {
1523 .d_version = D_VERSION,
1524 .d_name = "netmap",
1525 .d_open = netmap_open,
1526 .d_mmap_single = netmap_mmap_single,
1527 .d_ioctl = freebsd_netmap_ioctl,
1528 .d_poll = freebsd_netmap_poll,
1529 .d_kqfilter = netmap_kqfilter,
1530 .d_close = netmap_close,
1531 };
1532 /*--- end of kqueue support ----*/
1533
1534 /*
1535 * Kernel entry point.
1536 *
1537 * Initialize/finalize the module and return.
1538 *
1539 * Return 0 on success, errno on failure.
1540 */
1541 static int
netmap_loader(__unused struct module * module,int event,__unused void * arg)1542 netmap_loader(__unused struct module *module, int event, __unused void *arg)
1543 {
1544 int error = 0;
1545
1546 switch (event) {
1547 case MOD_LOAD:
1548 error = netmap_init();
1549 break;
1550
1551 case MOD_UNLOAD:
1552 /*
1553 * if some one is still using netmap,
1554 * then the module can not be unloaded.
1555 */
1556 if (netmap_use_count) {
1557 nm_prerr("netmap module can not be unloaded - netmap_use_count: %d",
1558 netmap_use_count);
1559 error = EBUSY;
1560 break;
1561 }
1562 netmap_fini();
1563 break;
1564
1565 default:
1566 error = EOPNOTSUPP;
1567 break;
1568 }
1569
1570 return (error);
1571 }
1572
1573 #ifdef DEV_MODULE_ORDERED
1574 /*
1575 * The netmap module contains three drivers: (i) the netmap character device
1576 * driver; (ii) the ptnetmap memdev PCI device driver, (iii) the ptnet PCI
1577 * device driver. The attach() routines of both (ii) and (iii) need the
1578 * lock of the global allocator, and such lock is initialized in netmap_init(),
1579 * which is part of (i).
1580 * Therefore, we make sure that (i) is loaded before (ii) and (iii), using
1581 * the 'order' parameter of driver declaration macros. For (i), we specify
1582 * SI_ORDER_MIDDLE, while higher orders are used with the DRIVER_MODULE_ORDERED
1583 * macros for (ii) and (iii).
1584 */
1585 DEV_MODULE_ORDERED(netmap, netmap_loader, NULL, SI_ORDER_MIDDLE);
1586 #else /* !DEV_MODULE_ORDERED */
1587 DEV_MODULE(netmap, netmap_loader, NULL);
1588 #endif /* DEV_MODULE_ORDERED */
1589 MODULE_DEPEND(netmap, pci, 1, 1, 1);
1590 MODULE_VERSION(netmap, 1);
1591 /* reduce conditional code */
1592 // linux API, use for the knlist in FreeBSD
1593 /* use a private mutex for the knlist */
1594