xref: /freebsd/sys/dev/cxgbe/t4_l2t.c (revision 70e0bbedef95258a4dadc996d641a9bebd3f107d)
1 /*-
2  * Copyright (c) 2011 Chelsio Communications, Inc.
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 #include <sys/cdefs.h>
27 __FBSDID("$FreeBSD$");
28 
29 #include "opt_inet.h"
30 
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/kernel.h>
34 #include <sys/module.h>
35 #include <sys/bus.h>
36 #include <sys/lock.h>
37 #include <sys/mutex.h>
38 #include <sys/rwlock.h>
39 #include <sys/socket.h>
40 #include <sys/sbuf.h>
41 #include <net/if.h>
42 #include <net/if_types.h>
43 #include <net/ethernet.h>
44 #include <net/if_vlan_var.h>
45 #include <net/if_dl.h>
46 #include <net/if_llatbl.h>
47 #include <net/route.h>
48 #include <netinet/in.h>
49 #include <netinet/in_var.h>
50 #include <netinet/if_ether.h>
51 
52 #include "common/common.h"
53 #include "common/jhash.h"
54 #include "common/t4_msg.h"
55 #include "t4_l2t.h"
56 
57 /*
58  * Module locking notes:  There is a RW lock protecting the L2 table as a
59  * whole plus a spinlock per L2T entry.  Entry lookups and allocations happen
60  * under the protection of the table lock, individual entry changes happen
61  * while holding that entry's spinlock.  The table lock nests outside the
62  * entry locks.  Allocations of new entries take the table lock as writers so
63  * no other lookups can happen while allocating new entries.  Entry updates
64  * take the table lock as readers so multiple entries can be updated in
65  * parallel.  An L2T entry can be dropped by decrementing its reference count
66  * and therefore can happen in parallel with entry allocation but no entry
67  * can change state or increment its ref count during allocation as both of
68  * these perform lookups.
69  *
70  * Note: We do not take refereces to ifnets in this module because both
71  * the TOE and the sockets already hold references to the interfaces and the
72  * lifetime of an L2T entry is fully contained in the lifetime of the TOE.
73  */
74 
75 /* identifies sync vs async L2T_WRITE_REQs */
76 #define S_SYNC_WR    12
77 #define V_SYNC_WR(x) ((x) << S_SYNC_WR)
78 #define F_SYNC_WR    V_SYNC_WR(1)
79 
80 enum {
81 	L2T_STATE_VALID,	/* entry is up to date */
82 	L2T_STATE_STALE,	/* entry may be used but needs revalidation */
83 	L2T_STATE_RESOLVING,	/* entry needs address resolution */
84 	L2T_STATE_SYNC_WRITE,	/* synchronous write of entry underway */
85 
86 	/* when state is one of the below the entry is not hashed */
87 	L2T_STATE_SWITCHING,	/* entry is being used by a switching filter */
88 	L2T_STATE_UNUSED	/* entry not in use */
89 };
90 
91 struct l2t_data {
92 	struct rwlock lock;
93 	volatile int nfree;	/* number of free entries */
94 	struct l2t_entry *rover;/* starting point for next allocation */
95 	struct l2t_entry l2tab[L2T_SIZE];
96 };
97 
98 static int do_l2t_write_rpl(struct sge_iq *, const struct rss_header *,
99     struct mbuf *);
100 
101 #define VLAN_NONE	0xfff
102 #define SA(x)           ((struct sockaddr *)(x))
103 #define SIN(x)          ((struct sockaddr_in *)(x))
104 #define SINADDR(x)      (SIN(x)->sin_addr.s_addr)
105 
106 /*
107  * Allocate a free L2T entry.  Must be called with l2t_data.lock held.
108  */
109 static struct l2t_entry *
110 alloc_l2e(struct l2t_data *d)
111 {
112 	struct l2t_entry *end, *e, **p;
113 
114 	rw_assert(&d->lock, RA_WLOCKED);
115 
116 	if (!atomic_load_acq_int(&d->nfree))
117 		return (NULL);
118 
119 	/* there's definitely a free entry */
120 	for (e = d->rover, end = &d->l2tab[L2T_SIZE]; e != end; ++e)
121 		if (atomic_load_acq_int(&e->refcnt) == 0)
122 			goto found;
123 
124 	for (e = d->l2tab; atomic_load_acq_int(&e->refcnt); ++e) ;
125 found:
126 	d->rover = e + 1;
127 	atomic_subtract_int(&d->nfree, 1);
128 
129 	/*
130 	 * The entry we found may be an inactive entry that is
131 	 * presently in the hash table.  We need to remove it.
132 	 */
133 	if (e->state < L2T_STATE_SWITCHING) {
134 		for (p = &d->l2tab[e->hash].first; *p; p = &(*p)->next) {
135 			if (*p == e) {
136 				*p = e->next;
137 				e->next = NULL;
138 				break;
139 			}
140 		}
141 	}
142 
143 	e->state = L2T_STATE_UNUSED;
144 	return (e);
145 }
146 
147 /*
148  * Write an L2T entry.  Must be called with the entry locked.
149  * The write may be synchronous or asynchronous.
150  */
151 static int
152 write_l2e(struct adapter *sc, struct l2t_entry *e, int sync)
153 {
154 	struct mbuf *m;
155 	struct cpl_l2t_write_req *req;
156 
157 	mtx_assert(&e->lock, MA_OWNED);
158 
159 	if ((m = m_gethdr(M_NOWAIT, MT_DATA)) == NULL)
160 		return (ENOMEM);
161 
162 	req = mtod(m, struct cpl_l2t_write_req *);
163 	m->m_pkthdr.len = m->m_len = sizeof(*req);
164 
165 	INIT_TP_WR(req, 0);
166 	OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_L2T_WRITE_REQ, e->idx |
167 	    V_SYNC_WR(sync) | V_TID_QID(sc->sge.fwq.abs_id)));
168 	req->params = htons(V_L2T_W_PORT(e->lport) | V_L2T_W_NOREPLY(!sync));
169 	req->l2t_idx = htons(e->idx);
170 	req->vlan = htons(e->vlan);
171 	memcpy(req->dst_mac, e->dmac, sizeof(req->dst_mac));
172 
173 	t4_mgmt_tx(sc, m);
174 
175 	if (sync && e->state != L2T_STATE_SWITCHING)
176 		e->state = L2T_STATE_SYNC_WRITE;
177 
178 	return (0);
179 }
180 
181 /*
182  * Allocate an L2T entry for use by a switching rule.  Such need to be
183  * explicitly freed and while busy they are not on any hash chain, so normal
184  * address resolution updates do not see them.
185  */
186 struct l2t_entry *
187 t4_l2t_alloc_switching(struct l2t_data *d)
188 {
189 	struct l2t_entry *e;
190 
191 	rw_rlock(&d->lock);
192 	e = alloc_l2e(d);
193 	if (e) {
194 		mtx_lock(&e->lock);          /* avoid race with t4_l2t_free */
195 		e->state = L2T_STATE_SWITCHING;
196 		atomic_store_rel_int(&e->refcnt, 1);
197 		mtx_unlock(&e->lock);
198 	}
199 	rw_runlock(&d->lock);
200 	return e;
201 }
202 
203 /*
204  * Sets/updates the contents of a switching L2T entry that has been allocated
205  * with an earlier call to @t4_l2t_alloc_switching.
206  */
207 int
208 t4_l2t_set_switching(struct adapter *sc, struct l2t_entry *e, uint16_t vlan,
209     uint8_t port, uint8_t *eth_addr)
210 {
211 	int rc;
212 
213 	e->vlan = vlan;
214 	e->lport = port;
215 	memcpy(e->dmac, eth_addr, ETHER_ADDR_LEN);
216 	mtx_lock(&e->lock);
217 	rc = write_l2e(sc, e, 0);
218 	mtx_unlock(&e->lock);
219 	return (rc);
220 }
221 
222 int
223 t4_init_l2t(struct adapter *sc, int flags)
224 {
225 	int i;
226 	struct l2t_data *d;
227 
228 	d = malloc(sizeof(*d), M_CXGBE, M_ZERO | flags);
229 	if (!d)
230 		return (ENOMEM);
231 
232 	d->rover = d->l2tab;
233 	atomic_store_rel_int(&d->nfree, L2T_SIZE);
234 	rw_init(&d->lock, "L2T");
235 
236 	for (i = 0; i < L2T_SIZE; i++) {
237 		d->l2tab[i].idx = i;
238 		d->l2tab[i].state = L2T_STATE_UNUSED;
239 		mtx_init(&d->l2tab[i].lock, "L2T_E", NULL, MTX_DEF);
240 		atomic_store_rel_int(&d->l2tab[i].refcnt, 0);
241 	}
242 
243 	sc->l2t = d;
244 	t4_register_cpl_handler(sc, CPL_L2T_WRITE_RPL, do_l2t_write_rpl);
245 
246 	return (0);
247 }
248 
249 int
250 t4_free_l2t(struct l2t_data *d)
251 {
252 	int i;
253 
254 	for (i = 0; i < L2T_SIZE; i++)
255 		mtx_destroy(&d->l2tab[i].lock);
256 	rw_destroy(&d->lock);
257 	free(d, M_CXGBE);
258 
259 	return (0);
260 }
261 
262 static inline unsigned int
263 vlan_prio(const struct l2t_entry *e)
264 {
265 	return e->vlan >> 13;
266 }
267 
268 static char
269 l2e_state(const struct l2t_entry *e)
270 {
271 	switch (e->state) {
272 	case L2T_STATE_VALID: return 'V';  /* valid, fast-path entry */
273 	case L2T_STATE_STALE: return 'S';  /* needs revalidation, but usable */
274 	case L2T_STATE_SYNC_WRITE: return 'W';
275 	case L2T_STATE_RESOLVING: return e->arpq_head ? 'A' : 'R';
276 	case L2T_STATE_SWITCHING: return 'X';
277 	default: return 'U';
278 	}
279 }
280 
281 int
282 sysctl_l2t(SYSCTL_HANDLER_ARGS)
283 {
284 	struct adapter *sc = arg1;
285 	struct l2t_data *l2t = sc->l2t;
286 	struct l2t_entry *e;
287 	struct sbuf *sb;
288 	int rc, i, header = 0;
289 	char ip[60];
290 
291 	if (l2t == NULL)
292 		return (ENXIO);
293 
294 	rc = sysctl_wire_old_buffer(req, 0);
295 	if (rc != 0)
296 		return (rc);
297 
298 	sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req);
299 	if (sb == NULL)
300 		return (ENOMEM);
301 
302 	e = &l2t->l2tab[0];
303 	for (i = 0; i < L2T_SIZE; i++, e++) {
304 		mtx_lock(&e->lock);
305 		if (e->state == L2T_STATE_UNUSED)
306 			goto skip;
307 
308 		if (header == 0) {
309 			sbuf_printf(sb, " Idx IP address      "
310 			    "Ethernet address  VLAN/P LP State Users Port");
311 			header = 1;
312 		}
313 		if (e->state == L2T_STATE_SWITCHING || e->v6)
314 			ip[0] = 0;
315 		else
316 			snprintf(ip, sizeof(ip), "%s",
317 			    inet_ntoa(*(struct in_addr *)&e->addr[0]));
318 
319 		/* XXX: accessing lle probably not safe? */
320 		sbuf_printf(sb, "\n%4u %-15s %02x:%02x:%02x:%02x:%02x:%02x %4d"
321 			   " %u %2u   %c   %5u %s",
322 			   e->idx, ip, e->dmac[0], e->dmac[1], e->dmac[2],
323 			   e->dmac[3], e->dmac[4], e->dmac[5],
324 			   e->vlan & 0xfff, vlan_prio(e), e->lport,
325 			   l2e_state(e), atomic_load_acq_int(&e->refcnt),
326 			   e->lle ? e->lle->lle_tbl->llt_ifp->if_xname : "");
327 skip:
328 		mtx_unlock(&e->lock);
329 	}
330 
331 	rc = sbuf_finish(sb);
332 	sbuf_delete(sb);
333 
334 	return (rc);
335 }
336 
337 #ifndef TCP_OFFLOAD_DISABLE
338 static inline void
339 l2t_hold(struct l2t_data *d, struct l2t_entry *e)
340 {
341 	if (atomic_fetchadd_int(&e->refcnt, 1) == 0)  /* 0 -> 1 transition */
342 		atomic_subtract_int(&d->nfree, 1);
343 }
344 
345 /*
346  * To avoid having to check address families we do not allow v4 and v6
347  * neighbors to be on the same hash chain.  We keep v4 entries in the first
348  * half of available hash buckets and v6 in the second.
349  */
350 enum {
351 	L2T_SZ_HALF = L2T_SIZE / 2,
352 	L2T_HASH_MASK = L2T_SZ_HALF - 1
353 };
354 
355 static inline unsigned int
356 arp_hash(const uint32_t *key, int ifindex)
357 {
358 	return jhash_2words(*key, ifindex, 0) & L2T_HASH_MASK;
359 }
360 
361 static inline unsigned int
362 ipv6_hash(const uint32_t *key, int ifindex)
363 {
364 	uint32_t xor = key[0] ^ key[1] ^ key[2] ^ key[3];
365 
366 	return L2T_SZ_HALF + (jhash_2words(xor, ifindex, 0) & L2T_HASH_MASK);
367 }
368 
369 static inline unsigned int
370 addr_hash(const uint32_t *addr, int addr_len, int ifindex)
371 {
372 	return addr_len == 4 ? arp_hash(addr, ifindex) :
373 			       ipv6_hash(addr, ifindex);
374 }
375 
376 /*
377  * Checks if an L2T entry is for the given IP/IPv6 address.  It does not check
378  * whether the L2T entry and the address are of the same address family.
379  * Callers ensure an address is only checked against L2T entries of the same
380  * family, something made trivial by the separation of IP and IPv6 hash chains
381  * mentioned above.  Returns 0 if there's a match,
382  */
383 static inline int
384 addreq(const struct l2t_entry *e, const uint32_t *addr)
385 {
386 	if (e->v6)
387 		return (e->addr[0] ^ addr[0]) | (e->addr[1] ^ addr[1]) |
388 		       (e->addr[2] ^ addr[2]) | (e->addr[3] ^ addr[3]);
389 	return e->addr[0] ^ addr[0];
390 }
391 
392 /*
393  * Add a packet to an L2T entry's queue of packets awaiting resolution.
394  * Must be called with the entry's lock held.
395  */
396 static inline void
397 arpq_enqueue(struct l2t_entry *e, struct mbuf *m)
398 {
399 	mtx_assert(&e->lock, MA_OWNED);
400 
401 	KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt not NULL", __func__));
402 	if (e->arpq_head)
403 		e->arpq_tail->m_nextpkt = m;
404 	else
405 		e->arpq_head = m;
406 	e->arpq_tail = m;
407 }
408 
409 static inline void
410 send_pending(struct adapter *sc, struct l2t_entry *e)
411 {
412 	struct mbuf *m, *next;
413 
414 	mtx_assert(&e->lock, MA_OWNED);
415 
416 	for (m = e->arpq_head; m; m = next) {
417 		next = m->m_nextpkt;
418 		m->m_nextpkt = NULL;
419 		t4_wrq_tx(sc, MBUF_EQ(m), m);
420 	}
421 	e->arpq_head = e->arpq_tail = NULL;
422 }
423 
424 #ifdef INET
425 /*
426  * Looks up and fills up an l2t_entry's lle.  We grab all the locks that we need
427  * ourself, and update e->state at the end if e->lle was successfully filled.
428  *
429  * The lle passed in comes from arpresolve and is ignored as it does not appear
430  * to be of much use.
431  */
432 static int
433 l2t_fill_lle(struct adapter *sc, struct l2t_entry *e, struct llentry *unused)
434 {
435         int rc = 0;
436         struct sockaddr_in sin;
437         struct ifnet *ifp = e->ifp;
438         struct llentry *lle;
439 
440         bzero(&sin, sizeof(struct sockaddr_in));
441 	if (e->v6)
442 		panic("%s: IPv6 L2 resolution not supported yet.", __func__);
443 
444 	sin.sin_family = AF_INET;
445 	sin.sin_len = sizeof(struct sockaddr_in);
446 	memcpy(&sin.sin_addr, e->addr, sizeof(struct sockaddr_in));
447 
448         mtx_assert(&e->lock, MA_NOTOWNED);
449         KASSERT(e->addr && ifp, ("%s: bad prep before call", __func__));
450 
451         IF_AFDATA_LOCK(ifp);
452         lle = lla_lookup(LLTABLE(ifp), LLE_EXCLUSIVE, SA(&sin));
453         IF_AFDATA_UNLOCK(ifp);
454         if (!LLE_IS_VALID(lle))
455                 return (ENOMEM);
456         if (!(lle->la_flags & LLE_VALID)) {
457                 rc = EINVAL;
458                 goto done;
459         }
460 
461         LLE_ADDREF(lle);
462 
463         mtx_lock(&e->lock);
464         if (e->state == L2T_STATE_RESOLVING) {
465                 KASSERT(e->lle == NULL, ("%s: lle already valid", __func__));
466                 e->lle = lle;
467                 memcpy(e->dmac, &lle->ll_addr, ETHER_ADDR_LEN);
468 		write_l2e(sc, e, 1);
469         } else {
470                 KASSERT(e->lle == lle, ("%s: lle changed", __func__));
471                 LLE_REMREF(lle);
472         }
473         mtx_unlock(&e->lock);
474 done:
475         LLE_WUNLOCK(lle);
476         return (rc);
477 }
478 #endif
479 
480 int
481 t4_l2t_send(struct adapter *sc, struct mbuf *m, struct l2t_entry *e)
482 {
483 #ifndef INET
484 	return (EINVAL);
485 #else
486 	struct llentry *lle = NULL;
487 	struct sockaddr_in sin;
488 	struct ifnet *ifp = e->ifp;
489 
490 	if (e->v6)
491 		panic("%s: IPv6 L2 resolution not supported yet.", __func__);
492 
493         bzero(&sin, sizeof(struct sockaddr_in));
494 	sin.sin_family = AF_INET;
495 	sin.sin_len = sizeof(struct sockaddr_in);
496 	memcpy(&sin.sin_addr, e->addr, sizeof(struct sockaddr_in));
497 
498 again:
499 	switch (e->state) {
500 	case L2T_STATE_STALE:     /* entry is stale, kick off revalidation */
501 		if (arpresolve(ifp, NULL, NULL, SA(&sin), e->dmac, &lle) == 0)
502 			l2t_fill_lle(sc, e, lle);
503 
504 		/* Fall through */
505 
506 	case L2T_STATE_VALID:     /* fast-path, send the packet on */
507 		return t4_wrq_tx(sc, MBUF_EQ(m), m);
508 
509 	case L2T_STATE_RESOLVING:
510 	case L2T_STATE_SYNC_WRITE:
511 		mtx_lock(&e->lock);
512 		if (e->state != L2T_STATE_SYNC_WRITE &&
513 		    e->state != L2T_STATE_RESOLVING) {
514 			/* state changed by the time we got here */
515 			mtx_unlock(&e->lock);
516 			goto again;
517 		}
518 		arpq_enqueue(e, m);
519 		mtx_unlock(&e->lock);
520 
521 		if (e->state == L2T_STATE_RESOLVING &&
522 		    arpresolve(ifp, NULL, NULL, SA(&sin), e->dmac, &lle) == 0)
523 			l2t_fill_lle(sc, e, lle);
524 	}
525 
526 	return (0);
527 #endif
528 }
529 
530 /*
531  * Called when an L2T entry has no more users.  The entry is left in the hash
532  * table since it is likely to be reused but we also bump nfree to indicate
533  * that the entry can be reallocated for a different neighbor.  We also drop
534  * the existing neighbor reference in case the neighbor is going away and is
535  * waiting on our reference.
536  *
537  * Because entries can be reallocated to other neighbors once their ref count
538  * drops to 0 we need to take the entry's lock to avoid races with a new
539  * incarnation.
540  */
541 static void
542 t4_l2e_free(struct l2t_entry *e)
543 {
544 	struct llentry *lle = NULL;
545 	struct l2t_data *d;
546 
547 	mtx_lock(&e->lock);
548 	if (atomic_load_acq_int(&e->refcnt) == 0) {  /* hasn't been recycled */
549 		lle = e->lle;
550 		e->lle = NULL;
551 		/*
552 		 * Don't need to worry about the arpq, an L2T entry can't be
553 		 * released if any packets are waiting for resolution as we
554 		 * need to be able to communicate with the device to close a
555 		 * connection.
556 		 */
557 	}
558 	mtx_unlock(&e->lock);
559 
560 	d = container_of(e, struct l2t_data, l2tab[e->idx]);
561 	atomic_add_int(&d->nfree, 1);
562 
563 	if (lle)
564 		LLE_FREE(lle);
565 }
566 
567 void
568 t4_l2t_release(struct l2t_entry *e)
569 {
570 	if (atomic_fetchadd_int(&e->refcnt, -1) == 1)
571 		t4_l2e_free(e);
572 }
573 
574 static int
575 do_l2t_write_rpl(struct sge_iq *iq, const struct rss_header *rss,
576     struct mbuf *m)
577 {
578 	struct adapter *sc = iq->adapter;
579 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
580 	unsigned int tid = GET_TID(rpl);
581 	unsigned int idx = tid & (L2T_SIZE - 1);
582 
583 	if (__predict_false(rpl->status != CPL_ERR_NONE)) {
584 		log(LOG_ERR,
585 		    "Unexpected L2T_WRITE_RPL status %u for entry %u\n",
586 		    rpl->status, idx);
587 		return (EINVAL);
588 	}
589 
590 	if (tid & F_SYNC_WR) {
591 		struct l2t_entry *e = &sc->l2t->l2tab[idx];
592 
593 		mtx_lock(&e->lock);
594 		if (e->state != L2T_STATE_SWITCHING) {
595 			send_pending(sc, e);
596 			e->state = L2T_STATE_VALID;
597 		}
598 		mtx_unlock(&e->lock);
599 	}
600 
601 	return (0);
602 }
603 
604 /*
605  * Reuse an L2T entry that was previously used for the same next hop.
606  */
607 static void
608 reuse_entry(struct l2t_entry *e)
609 {
610 	struct llentry *lle;
611 
612 	mtx_lock(&e->lock);                /* avoid race with t4_l2t_free */
613 	lle = e->lle;
614 	if (lle) {
615 		KASSERT(lle->la_flags & LLE_VALID,
616 			("%s: invalid lle stored in l2t_entry", __func__));
617 
618 		if (lle->la_expire >= time_uptime)
619 			e->state = L2T_STATE_STALE;
620 		else
621 			e->state = L2T_STATE_VALID;
622 	} else
623 		e->state = L2T_STATE_RESOLVING;
624 	mtx_unlock(&e->lock);
625 }
626 
627 /*
628  * The TOE wants an L2 table entry that it can use to reach the next hop over
629  * the specified port.  Produce such an entry - create one if needed.
630  *
631  * Note that the ifnet could be a pseudo-device like if_vlan, if_lagg, etc. on
632  * top of the real cxgbe interface.
633  */
634 struct l2t_entry *
635 t4_l2t_get(struct port_info *pi, struct ifnet *ifp, struct sockaddr *sa)
636 {
637 	struct l2t_entry *e;
638 	struct l2t_data *d = pi->adapter->l2t;
639 	int addr_len;
640 	uint32_t *addr;
641 	int hash;
642 	struct sockaddr_in6 *sin6;
643 	unsigned int smt_idx = pi->port_id;
644 
645 	if (sa->sa_family == AF_INET) {
646 		addr = (uint32_t *)&SINADDR(sa);
647 		addr_len = sizeof(SINADDR(sa));
648 	} else if (sa->sa_family == AF_INET6) {
649 		sin6 = (struct sockaddr_in6 *)sa;
650 		addr = (uint32_t *)&sin6->sin6_addr.s6_addr;
651 		addr_len = sizeof(sin6->sin6_addr.s6_addr);
652 	} else
653 		return (NULL);
654 
655 	hash = addr_hash(addr, addr_len, ifp->if_index);
656 
657 	rw_wlock(&d->lock);
658 	for (e = d->l2tab[hash].first; e; e = e->next) {
659 		if (!addreq(e, addr) && e->ifp == ifp && e->smt_idx == smt_idx){
660 			l2t_hold(d, e);
661 			if (atomic_load_acq_int(&e->refcnt) == 1)
662 				reuse_entry(e);
663 			goto done;
664 		}
665 	}
666 
667 	/* Need to allocate a new entry */
668 	e = alloc_l2e(d);
669 	if (e) {
670 		mtx_lock(&e->lock);          /* avoid race with t4_l2t_free */
671 		e->state = L2T_STATE_RESOLVING;
672 		memcpy(e->addr, addr, addr_len);
673 		e->ifindex = ifp->if_index;
674 		e->smt_idx = smt_idx;
675 		e->ifp = ifp;
676 		e->hash = hash;
677 		e->lport = pi->lport;
678 		e->v6 = (addr_len == 16);
679 		e->lle = NULL;
680 		atomic_store_rel_int(&e->refcnt, 1);
681 		if (ifp->if_type == IFT_L2VLAN)
682 			VLAN_TAG(ifp, &e->vlan);
683 		else
684 			e->vlan = VLAN_NONE;
685 		e->next = d->l2tab[hash].first;
686 		d->l2tab[hash].first = e;
687 		mtx_unlock(&e->lock);
688 	}
689 done:
690 	rw_wunlock(&d->lock);
691 	return e;
692 }
693 
694 /*
695  * Called when the host's neighbor layer makes a change to some entry that is
696  * loaded into the HW L2 table.
697  */
698 void
699 t4_l2t_update(struct adapter *sc, struct llentry *lle)
700 {
701 	struct l2t_entry *e;
702 	struct l2t_data *d = sc->l2t;
703 	struct sockaddr *sa = L3_ADDR(lle);
704 	struct llentry *old_lle = NULL;
705 	uint32_t *addr = (uint32_t *)&SINADDR(sa);
706 	struct ifnet *ifp = lle->lle_tbl->llt_ifp;
707 	int hash = addr_hash(addr, sizeof(*addr), ifp->if_index);
708 
709 	KASSERT(d != NULL, ("%s: no L2 table", __func__));
710 	LLE_WLOCK_ASSERT(lle);
711 	KASSERT(lle->la_flags & LLE_VALID || lle->la_flags & LLE_DELETED,
712 	    ("%s: entry neither valid nor deleted.", __func__));
713 
714 	rw_rlock(&d->lock);
715 	for (e = d->l2tab[hash].first; e; e = e->next) {
716 		if (!addreq(e, addr) && e->ifp == ifp) {
717 			mtx_lock(&e->lock);
718 			if (atomic_load_acq_int(&e->refcnt))
719 				goto found;
720 			e->state = L2T_STATE_STALE;
721 			mtx_unlock(&e->lock);
722 			break;
723 		}
724 	}
725 	rw_runlock(&d->lock);
726 
727 	/* The TOE has no interest in this LLE */
728 	return;
729 
730  found:
731 	rw_runlock(&d->lock);
732 
733         if (atomic_load_acq_int(&e->refcnt)) {
734 
735                 /* Entry is referenced by at least 1 offloaded connection. */
736 
737                 /* Handle deletes first */
738                 if (lle->la_flags & LLE_DELETED) {
739                         if (lle == e->lle) {
740                                 e->lle = NULL;
741                                 e->state = L2T_STATE_RESOLVING;
742                                 LLE_REMREF(lle);
743                         }
744                         goto done;
745                 }
746 
747                 if (lle != e->lle) {
748                         old_lle = e->lle;
749                         LLE_ADDREF(lle);
750                         e->lle = lle;
751                 }
752 
753                 if (e->state == L2T_STATE_RESOLVING ||
754                     memcmp(e->dmac, &lle->ll_addr, ETHER_ADDR_LEN)) {
755 
756                         /* unresolved -> resolved; or dmac changed */
757 
758                         memcpy(e->dmac, &lle->ll_addr, ETHER_ADDR_LEN);
759 			write_l2e(sc, e, 1);
760                 } else {
761 
762                         /* +ve reinforcement of a valid or stale entry */
763 
764                 }
765 
766                 e->state = L2T_STATE_VALID;
767 
768         } else {
769                 /*
770                  * Entry was used previously but is unreferenced right now.
771                  * e->lle has been released and NULL'd out by t4_l2t_free, or
772                  * l2t_release is about to call t4_l2t_free and do that.
773                  *
774                  * Either way this is of no interest to us.
775                  */
776         }
777 
778 done:
779         mtx_unlock(&e->lock);
780         if (old_lle)
781                 LLE_FREE(old_lle);
782 }
783 
784 #endif
785