xref: /freebsd/sys/netinet/toecore.c (revision 148a8da8) 
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2012 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 
36 #include <sys/param.h>
37 #include <sys/kernel.h>
38 #include <sys/systm.h>
39 #include <sys/malloc.h>
40 #include <sys/mbuf.h>
41 #include <sys/module.h>
42 #include <sys/types.h>
43 #include <sys/sockopt.h>
44 #include <sys/sysctl.h>
45 #include <sys/socket.h>
46 
47 #include <net/ethernet.h>
48 #include <net/if.h>
49 #include <net/if_var.h>
50 #include <net/if_types.h>
51 #include <net/if_vlan_var.h>
52 #include <net/if_llatbl.h>
53 #include <net/route.h>
54 
55 #include <netinet/if_ether.h>
56 #include <netinet/in.h>
57 #include <netinet/in_pcb.h>
58 #include <netinet/in_var.h>
59 #include <netinet6/in6_var.h>
60 #include <netinet6/in6_pcb.h>
61 #include <netinet6/nd6.h>
62 #define TCPSTATES
63 #include <netinet/tcp.h>
64 #include <netinet/tcp_fsm.h>
65 #include <netinet/tcp_timer.h>
66 #include <netinet/tcp_var.h>
67 #include <netinet/tcp_syncache.h>
68 #include <netinet/tcp_offload.h>
69 #include <netinet/toecore.h>
70 
71 static struct mtx toedev_lock;
72 static TAILQ_HEAD(, toedev) toedev_list;
73 static eventhandler_tag listen_start_eh;
74 static eventhandler_tag listen_stop_eh;
75 static eventhandler_tag lle_event_eh;
76 
77 static int
78 toedev_connect(struct toedev *tod __unused, struct socket *so __unused,
79     struct rtentry *rt __unused, struct sockaddr *nam __unused)
80 {
81 
82 	return (ENOTSUP);
83 }
84 
85 static int
86 toedev_listen_start(struct toedev *tod __unused, struct tcpcb *tp __unused)
87 {
88 
89 	return (ENOTSUP);
90 }
91 
92 static int
93 toedev_listen_stop(struct toedev *tod __unused, struct tcpcb *tp __unused)
94 {
95 
96 	return (ENOTSUP);
97 }
98 
99 static void
100 toedev_input(struct toedev *tod __unused, struct tcpcb *tp __unused,
101     struct mbuf *m)
102 {
103 
104 	m_freem(m);
105 	return;
106 }
107 
108 static void
109 toedev_rcvd(struct toedev *tod __unused, struct tcpcb *tp __unused)
110 {
111 
112 	return;
113 }
114 
115 static int
116 toedev_output(struct toedev *tod __unused, struct tcpcb *tp __unused)
117 {
118 
119 	return (ENOTSUP);
120 }
121 
122 static void
123 toedev_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp __unused)
124 {
125 
126 	return;
127 }
128 
129 static void
130 toedev_l2_update(struct toedev *tod __unused, struct ifnet *ifp __unused,
131     struct sockaddr *sa __unused, uint8_t *lladdr __unused,
132     uint16_t vtag __unused)
133 {
134 
135 	return;
136 }
137 
138 static void
139 toedev_route_redirect(struct toedev *tod __unused, struct ifnet *ifp __unused,
140     struct rtentry *rt0 __unused, struct rtentry *rt1 __unused)
141 {
142 
143 	return;
144 }
145 
146 static void
147 toedev_syncache_added(struct toedev *tod __unused, void *ctx __unused)
148 {
149 
150 	return;
151 }
152 
153 static void
154 toedev_syncache_removed(struct toedev *tod __unused, void *ctx __unused)
155 {
156 
157 	return;
158 }
159 
160 static int
161 toedev_syncache_respond(struct toedev *tod __unused, void *ctx __unused,
162     struct mbuf *m)
163 {
164 
165 	m_freem(m);
166 	return (0);
167 }
168 
169 static void
170 toedev_offload_socket(struct toedev *tod __unused, void *ctx __unused,
171     struct socket *so __unused)
172 {
173 
174 	return;
175 }
176 
177 static void
178 toedev_ctloutput(struct toedev *tod __unused, struct tcpcb *tp __unused,
179     int sopt_dir __unused, int sopt_name __unused)
180 {
181 
182 	return;
183 }
184 
185 static void
186 toedev_tcp_info(struct toedev *tod __unused, struct tcpcb *tp __unused,
187     struct tcp_info *ti __unused)
188 {
189 
190 	return;
191 }
192 
193 /*
194  * Inform one or more TOE devices about a listening socket.
195  */
196 static void
197 toe_listen_start(struct inpcb *inp, void *arg)
198 {
199 	struct toedev *t, *tod;
200 	struct tcpcb *tp;
201 
202 	INP_WLOCK_ASSERT(inp);
203 	KASSERT(inp->inp_pcbinfo == &V_tcbinfo,
204 	    ("%s: inp is not a TCP inp", __func__));
205 
206 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))
207 		return;
208 
209 	tp = intotcpcb(inp);
210 	if (tp->t_state != TCPS_LISTEN)
211 		return;
212 
213 	t = arg;
214 	mtx_lock(&toedev_lock);
215 	TAILQ_FOREACH(tod, &toedev_list, link) {
216 		if (t == NULL || t == tod)
217 			tod->tod_listen_start(tod, tp);
218 	}
219 	mtx_unlock(&toedev_lock);
220 }
221 
222 static void
223 toe_listen_start_event(void *arg __unused, struct tcpcb *tp)
224 {
225 	struct inpcb *inp = tp->t_inpcb;
226 
227 	INP_WLOCK_ASSERT(inp);
228 	KASSERT(tp->t_state == TCPS_LISTEN,
229 	    ("%s: t_state %s", __func__, tcpstates[tp->t_state]));
230 
231 	toe_listen_start(inp, NULL);
232 }
233 
234 static void
235 toe_listen_stop_event(void *arg __unused, struct tcpcb *tp)
236 {
237 	struct toedev *tod;
238 #ifdef INVARIANTS
239 	struct inpcb *inp = tp->t_inpcb;
240 #endif
241 
242 	INP_WLOCK_ASSERT(inp);
243 	KASSERT(tp->t_state == TCPS_LISTEN,
244 	    ("%s: t_state %s", __func__, tcpstates[tp->t_state]));
245 
246 	mtx_lock(&toedev_lock);
247 	TAILQ_FOREACH(tod, &toedev_list, link)
248 	    tod->tod_listen_stop(tod, tp);
249 	mtx_unlock(&toedev_lock);
250 }
251 
252 /*
253  * Fill up a freshly allocated toedev struct with reasonable defaults.
254  */
255 void
256 init_toedev(struct toedev *tod)
257 {
258 
259 	tod->tod_softc = NULL;
260 
261 	/*
262 	 * Provide no-op defaults so that the kernel can call any toedev
263 	 * function without having to check whether the TOE driver supplied one
264 	 * or not.
265 	 */
266 	tod->tod_connect = toedev_connect;
267 	tod->tod_listen_start = toedev_listen_start;
268 	tod->tod_listen_stop = toedev_listen_stop;
269 	tod->tod_input = toedev_input;
270 	tod->tod_rcvd = toedev_rcvd;
271 	tod->tod_output = toedev_output;
272 	tod->tod_send_rst = toedev_output;
273 	tod->tod_send_fin = toedev_output;
274 	tod->tod_pcb_detach = toedev_pcb_detach;
275 	tod->tod_l2_update = toedev_l2_update;
276 	tod->tod_route_redirect = toedev_route_redirect;
277 	tod->tod_syncache_added = toedev_syncache_added;
278 	tod->tod_syncache_removed = toedev_syncache_removed;
279 	tod->tod_syncache_respond = toedev_syncache_respond;
280 	tod->tod_offload_socket = toedev_offload_socket;
281 	tod->tod_ctloutput = toedev_ctloutput;
282 	tod->tod_tcp_info = toedev_tcp_info;
283 }
284 
285 /*
286  * Register an active TOE device with the system.  This allows it to receive
287  * notifications from the kernel.
288  */
289 int
290 register_toedev(struct toedev *tod)
291 {
292 	struct toedev *t;
293 
294 	mtx_lock(&toedev_lock);
295 	TAILQ_FOREACH(t, &toedev_list, link) {
296 		if (t == tod) {
297 			mtx_unlock(&toedev_lock);
298 			return (EEXIST);
299 		}
300 	}
301 
302 	TAILQ_INSERT_TAIL(&toedev_list, tod, link);
303 	registered_toedevs++;
304 	mtx_unlock(&toedev_lock);
305 
306 	inp_apply_all(toe_listen_start, tod);
307 
308 	return (0);
309 }
310 
311 /*
312  * Remove the TOE device from the global list of active TOE devices.  It is the
313  * caller's responsibility to ensure that the TOE device is quiesced prior to
314  * this call.
315  */
316 int
317 unregister_toedev(struct toedev *tod)
318 {
319 	struct toedev *t, *t2;
320 	int rc = ENODEV;
321 
322 	mtx_lock(&toedev_lock);
323 	TAILQ_FOREACH_SAFE(t, &toedev_list, link, t2) {
324 		if (t == tod) {
325 			TAILQ_REMOVE(&toedev_list, tod, link);
326 			registered_toedevs--;
327 			rc = 0;
328 			break;
329 		}
330 	}
331 	KASSERT(registered_toedevs >= 0,
332 	    ("%s: registered_toedevs (%d) < 0", __func__, registered_toedevs));
333 	mtx_unlock(&toedev_lock);
334 	return (rc);
335 }
336 
337 void
338 toe_syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th,
339     struct inpcb *inp, void *tod, void *todctx)
340 {
341 	struct socket *lso = inp->inp_socket;
342 
343 	INP_WLOCK_ASSERT(inp);
344 
345 	syncache_add(inc, to, th, inp, &lso, NULL, tod, todctx);
346 }
347 
348 int
349 toe_syncache_expand(struct in_conninfo *inc, struct tcpopt *to,
350     struct tcphdr *th, struct socket **lsop)
351 {
352 
353 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
354 
355 	return (syncache_expand(inc, to, th, lsop, NULL));
356 }
357 
358 /*
359  * General purpose check to see if a 4-tuple is in use by the kernel.  If a TCP
360  * header (presumably for an incoming SYN) is also provided, an existing 4-tuple
361  * in TIME_WAIT may be assassinated freeing it up for re-use.
362  *
363  * Note that the TCP header must have been run through tcp_fields_to_host() or
364  * equivalent.
365  */
366 int
367 toe_4tuple_check(struct in_conninfo *inc, struct tcphdr *th, struct ifnet *ifp)
368 {
369 	struct inpcb *inp;
370 
371 	if (inc->inc_flags & INC_ISIPV6) {
372 		inp = in6_pcblookup(&V_tcbinfo, &inc->inc6_faddr,
373 		    inc->inc_fport, &inc->inc6_laddr, inc->inc_lport,
374 		    INPLOOKUP_WLOCKPCB, ifp);
375 	} else {
376 		inp = in_pcblookup(&V_tcbinfo, inc->inc_faddr, inc->inc_fport,
377 		    inc->inc_laddr, inc->inc_lport, INPLOOKUP_WLOCKPCB, ifp);
378 	}
379 	if (inp != NULL) {
380 		INP_WLOCK_ASSERT(inp);
381 
382 		if ((inp->inp_flags & INP_TIMEWAIT) && th != NULL) {
383 
384 			INP_INFO_RLOCK_ASSERT(&V_tcbinfo); /* for twcheck */
385 			if (!tcp_twcheck(inp, NULL, th, NULL, 0))
386 				return (EADDRINUSE);
387 		} else {
388 			INP_WUNLOCK(inp);
389 			return (EADDRINUSE);
390 		}
391 	}
392 
393 	return (0);
394 }
395 
396 static void
397 toe_lle_event(void *arg __unused, struct llentry *lle, int evt)
398 {
399 	struct toedev *tod;
400 	struct ifnet *ifp;
401 	struct sockaddr *sa;
402 	uint8_t *lladdr;
403 	uint16_t vid, pcp;
404 	int family;
405 	struct sockaddr_in6 sin6;
406 
407 	LLE_WLOCK_ASSERT(lle);
408 
409 	ifp = lltable_get_ifp(lle->lle_tbl);
410 	family = lltable_get_af(lle->lle_tbl);
411 
412 	if (family != AF_INET && family != AF_INET6)
413 		return;
414 	/*
415 	 * Not interested if the interface's TOE capability is not enabled.
416 	 */
417 	if ((family == AF_INET && !(ifp->if_capenable & IFCAP_TOE4)) ||
418 	    (family == AF_INET6 && !(ifp->if_capenable & IFCAP_TOE6)))
419 		return;
420 
421 	tod = TOEDEV(ifp);
422 	if (tod == NULL)
423 		return;
424 
425 	sa = (struct sockaddr *)&sin6;
426 	lltable_fill_sa_entry(lle, sa);
427 
428 	vid = 0xfff;
429 	pcp = 0;
430 	if (evt != LLENTRY_RESOLVED) {
431 
432 		/*
433 		 * LLENTRY_TIMEDOUT, LLENTRY_DELETED, LLENTRY_EXPIRED all mean
434 		 * this entry is going to be deleted.
435 		 */
436 
437 		lladdr = NULL;
438 	} else {
439 
440 		KASSERT(lle->la_flags & LLE_VALID,
441 		    ("%s: %p resolved but not valid?", __func__, lle));
442 
443 		lladdr = (uint8_t *)lle->ll_addr;
444 		VLAN_TAG(ifp, &vid);
445 		VLAN_PCP(ifp, &pcp);
446 	}
447 
448 	tod->tod_l2_update(tod, ifp, sa, lladdr, EVL_MAKETAG(vid, pcp, 0));
449 }
450 
451 /*
452  * Returns 0 or EWOULDBLOCK on success (any other value is an error).  0 means
453  * lladdr and vtag are valid on return, EWOULDBLOCK means the TOE driver's
454  * tod_l2_update will be called later, when the entry is resolved or times out.
455  */
456 int
457 toe_l2_resolve(struct toedev *tod, struct ifnet *ifp, struct sockaddr *sa,
458     uint8_t *lladdr, uint16_t *vtag)
459 {
460 	int rc;
461 	uint16_t vid, pcp;
462 
463 	switch (sa->sa_family) {
464 #ifdef INET
465 	case AF_INET:
466 		rc = arpresolve(ifp, 0, NULL, sa, lladdr, NULL, NULL);
467 		break;
468 #endif
469 #ifdef INET6
470 	case AF_INET6:
471 		rc = nd6_resolve(ifp, 0, NULL, sa, lladdr, NULL, NULL);
472 		break;
473 #endif
474 	default:
475 		return (EPROTONOSUPPORT);
476 	}
477 
478 	if (rc == 0) {
479 		vid = 0xfff;
480 		pcp = 0;
481 		if (ifp->if_type == IFT_L2VLAN) {
482 			VLAN_TAG(ifp, &vid);
483 			VLAN_PCP(ifp, &pcp);
484 		} else if (ifp->if_pcp != IFNET_PCP_NONE) {
485 			vid = 0;
486 			pcp = ifp->if_pcp;
487 		}
488 		*vtag = EVL_MAKETAG(vid, pcp, 0);
489 	}
490 
491 	return (rc);
492 }
493 
494 void
495 toe_connect_failed(struct toedev *tod, struct inpcb *inp, int err)
496 {
497 
498 	INP_WLOCK_ASSERT(inp);
499 
500 	if (!(inp->inp_flags & INP_DROPPED)) {
501 		struct tcpcb *tp = intotcpcb(inp);
502 
503 		KASSERT(tp->t_flags & TF_TOE,
504 		    ("%s: tp %p not offloaded.", __func__, tp));
505 
506 		if (err == EAGAIN) {
507 
508 			/*
509 			 * Temporary failure during offload, take this PCB back.
510 			 * Detach from the TOE driver and do the rest of what
511 			 * TCP's pru_connect would have done if the connection
512 			 * wasn't offloaded.
513 			 */
514 
515 			tod->tod_pcb_detach(tod, tp);
516 			KASSERT(!(tp->t_flags & TF_TOE),
517 			    ("%s: tp %p still offloaded.", __func__, tp));
518 			tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp));
519 			(void) tp->t_fb->tfb_tcp_output(tp);
520 		} else {
521 
522 			INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
523 			tp = tcp_drop(tp, err);
524 			if (tp == NULL)
525 				INP_WLOCK(inp);	/* re-acquire */
526 		}
527 	}
528 	INP_WLOCK_ASSERT(inp);
529 }
530 
531 static int
532 toecore_load(void)
533 {
534 
535 	mtx_init(&toedev_lock, "toedev lock", NULL, MTX_DEF);
536 	TAILQ_INIT(&toedev_list);
537 
538 	listen_start_eh = EVENTHANDLER_REGISTER(tcp_offload_listen_start,
539 	    toe_listen_start_event, NULL, EVENTHANDLER_PRI_ANY);
540 	listen_stop_eh = EVENTHANDLER_REGISTER(tcp_offload_listen_stop,
541 	    toe_listen_stop_event, NULL, EVENTHANDLER_PRI_ANY);
542 	lle_event_eh = EVENTHANDLER_REGISTER(lle_event, toe_lle_event, NULL,
543 	    EVENTHANDLER_PRI_ANY);
544 
545 	return (0);
546 }
547 
548 static int
549 toecore_unload(void)
550 {
551 
552 	mtx_lock(&toedev_lock);
553 	if (!TAILQ_EMPTY(&toedev_list)) {
554 		mtx_unlock(&toedev_lock);
555 		return (EBUSY);
556 	}
557 
558 	EVENTHANDLER_DEREGISTER(tcp_offload_listen_start, listen_start_eh);
559 	EVENTHANDLER_DEREGISTER(tcp_offload_listen_stop, listen_stop_eh);
560 	EVENTHANDLER_DEREGISTER(lle_event, lle_event_eh);
561 
562 	mtx_unlock(&toedev_lock);
563 	mtx_destroy(&toedev_lock);
564 
565 	return (0);
566 }
567 
568 static int
569 toecore_mod_handler(module_t mod, int cmd, void *arg)
570 {
571 
572 	if (cmd == MOD_LOAD)
573 		return (toecore_load());
574 
575 	if (cmd == MOD_UNLOAD)
576 		return (toecore_unload());
577 
578 	return (EOPNOTSUPP);
579 }
580 
581 static moduledata_t mod_data= {
582 	"toecore",
583 	toecore_mod_handler,
584 	0
585 };
586 
587 MODULE_VERSION(toecore, 1);
588 DECLARE_MODULE(toecore, mod_data, SI_SUB_EXEC, SI_ORDER_ANY);
589