xref: /freebsd/sys/dev/sfxge/sfxge_rx.c (revision 98fc0505155575df4170c863107153df2a751a13)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2010-2016 Solarflare Communications Inc.
5  * All rights reserved.
6  *
7  * This software was developed in part by Philip Paeps under contract for
8  * Solarflare Communications, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions are met:
12  *
13  * 1. Redistributions of source code must retain the above copyright notice,
14  *    this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright notice,
16  *    this list of conditions and the following disclaimer in the documentation
17  *    and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
20  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
21  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
23  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
24  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
25  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
26  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
27  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
28  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
29  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30  *
31  * The views and conclusions contained in the software and documentation are
32  * those of the authors and should not be interpreted as representing official
33  * policies, either expressed or implied, of the FreeBSD Project.
34  */
35 
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 #include "opt_rss.h"
40 
41 #include <sys/param.h>
42 #include <sys/malloc.h>
43 #include <sys/mbuf.h>
44 #include <sys/smp.h>
45 #include <sys/socket.h>
46 #include <sys/sysctl.h>
47 #include <sys/syslog.h>
48 #include <sys/limits.h>
49 #include <sys/syslog.h>
50 
51 #include <net/ethernet.h>
52 #include <net/if.h>
53 #include <net/if_vlan_var.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/ip6.h>
58 #include <netinet/tcp.h>
59 
60 #include <machine/in_cksum.h>
61 
62 #ifdef RSS
63 #include <net/rss_config.h>
64 #endif
65 
66 #include "common/efx.h"
67 
68 
69 #include "sfxge.h"
70 #include "sfxge_rx.h"
71 
72 #define	RX_REFILL_THRESHOLD(_entries)	(EFX_RXQ_LIMIT(_entries) * 9 / 10)
73 
74 #ifdef SFXGE_LRO
75 
76 SYSCTL_NODE(_hw_sfxge, OID_AUTO, lro, CTLFLAG_RD, NULL,
77 	    "Large receive offload (LRO) parameters");
78 
79 #define	SFXGE_LRO_PARAM(_param)	SFXGE_PARAM(lro._param)
80 
81 /* Size of the LRO hash table.  Must be a power of 2.  A larger table
82  * means we can accelerate a larger number of streams.
83  */
84 static unsigned lro_table_size = 128;
85 TUNABLE_INT(SFXGE_LRO_PARAM(table_size), &lro_table_size);
86 SYSCTL_UINT(_hw_sfxge_lro, OID_AUTO, table_size, CTLFLAG_RDTUN,
87 	    &lro_table_size, 0,
88 	    "Size of the LRO hash table (must be a power of 2)");
89 
90 /* Maximum length of a hash chain.  If chains get too long then the lookup
91  * time increases and may exceed the benefit of LRO.
92  */
93 static unsigned lro_chain_max = 20;
94 TUNABLE_INT(SFXGE_LRO_PARAM(chain_max), &lro_chain_max);
95 SYSCTL_UINT(_hw_sfxge_lro, OID_AUTO, chain_max, CTLFLAG_RDTUN,
96 	    &lro_chain_max, 0,
97 	    "The maximum length of a hash chain");
98 
99 /* Maximum time (in ticks) that a connection can be idle before it's LRO
100  * state is discarded.
101  */
102 static unsigned lro_idle_ticks; /* initialised in sfxge_rx_init() */
103 TUNABLE_INT(SFXGE_LRO_PARAM(idle_ticks), &lro_idle_ticks);
104 SYSCTL_UINT(_hw_sfxge_lro, OID_AUTO, idle_ticks, CTLFLAG_RDTUN,
105 	    &lro_idle_ticks, 0,
106 	    "The maximum time (in ticks) that a connection can be idle "
107 	    "before it's LRO state is discarded");
108 
109 /* Number of packets with payload that must arrive in-order before a
110  * connection is eligible for LRO.  The idea is we should avoid coalescing
111  * segments when the sender is in slow-start because reducing the ACK rate
112  * can damage performance.
113  */
114 static int lro_slow_start_packets = 2000;
115 TUNABLE_INT(SFXGE_LRO_PARAM(slow_start_packets), &lro_slow_start_packets);
116 SYSCTL_UINT(_hw_sfxge_lro, OID_AUTO, slow_start_packets, CTLFLAG_RDTUN,
117 	    &lro_slow_start_packets, 0,
118 	    "Number of packets with payload that must arrive in-order before "
119 	    "a connection is eligible for LRO");
120 
121 /* Number of packets with payload that must arrive in-order following loss
122  * before a connection is eligible for LRO.  The idea is we should avoid
123  * coalescing segments when the sender is recovering from loss, because
124  * reducing the ACK rate can damage performance.
125  */
126 static int lro_loss_packets = 20;
127 TUNABLE_INT(SFXGE_LRO_PARAM(loss_packets), &lro_loss_packets);
128 SYSCTL_UINT(_hw_sfxge_lro, OID_AUTO, loss_packets, CTLFLAG_RDTUN,
129 	    &lro_loss_packets, 0,
130 	    "Number of packets with payload that must arrive in-order "
131 	    "following loss before a connection is eligible for LRO");
132 
133 /* Flags for sfxge_lro_conn::l2_id; must not collide with EVL_VLID_MASK */
134 #define	SFXGE_LRO_L2_ID_VLAN 0x4000
135 #define	SFXGE_LRO_L2_ID_IPV6 0x8000
136 #define	SFXGE_LRO_CONN_IS_VLAN_ENCAP(c) ((c)->l2_id & SFXGE_LRO_L2_ID_VLAN)
137 #define	SFXGE_LRO_CONN_IS_TCPIPV4(c) (!((c)->l2_id & SFXGE_LRO_L2_ID_IPV6))
138 
139 /* Compare IPv6 addresses, avoiding conditional branches */
140 static unsigned long ipv6_addr_cmp(const struct in6_addr *left,
141 				   const struct in6_addr *right)
142 {
143 #if LONG_BIT == 64
144 	const uint64_t *left64 = (const uint64_t *)left;
145 	const uint64_t *right64 = (const uint64_t *)right;
146 	return (left64[0] - right64[0]) | (left64[1] - right64[1]);
147 #else
148 	return (left->s6_addr32[0] - right->s6_addr32[0]) |
149 	       (left->s6_addr32[1] - right->s6_addr32[1]) |
150 	       (left->s6_addr32[2] - right->s6_addr32[2]) |
151 	       (left->s6_addr32[3] - right->s6_addr32[3]);
152 #endif
153 }
154 
155 #endif	/* SFXGE_LRO */
156 
157 void
158 sfxge_rx_qflush_done(struct sfxge_rxq *rxq)
159 {
160 
161 	rxq->flush_state = SFXGE_FLUSH_DONE;
162 }
163 
164 void
165 sfxge_rx_qflush_failed(struct sfxge_rxq *rxq)
166 {
167 
168 	rxq->flush_state = SFXGE_FLUSH_FAILED;
169 }
170 
171 #ifdef RSS
172 static uint8_t toep_key[RSS_KEYSIZE];
173 #else
174 static uint8_t toep_key[] = {
175 	0x6d, 0x5a, 0x56, 0xda, 0x25, 0x5b, 0x0e, 0xc2,
176 	0x41, 0x67, 0x25, 0x3d, 0x43, 0xa3, 0x8f, 0xb0,
177 	0xd0, 0xca, 0x2b, 0xcb, 0xae, 0x7b, 0x30, 0xb4,
178 	0x77, 0xcb, 0x2d, 0xa3, 0x80, 0x30, 0xf2, 0x0c,
179 	0x6a, 0x42, 0xb7, 0x3b, 0xbe, 0xac, 0x01, 0xfa
180 };
181 #endif
182 
183 static void
184 sfxge_rx_post_refill(void *arg)
185 {
186 	struct sfxge_rxq *rxq = arg;
187 	struct sfxge_softc *sc;
188 	unsigned int index;
189 	struct sfxge_evq *evq;
190 	uint16_t magic;
191 
192 	sc = rxq->sc;
193 	index = rxq->index;
194 	evq = sc->evq[index];
195 	magic = sfxge_sw_ev_rxq_magic(SFXGE_SW_EV_RX_QREFILL, rxq);
196 
197 	/* This is guaranteed due to the start/stop order of rx and ev */
198 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
199 	    ("evq not started"));
200 	KASSERT(rxq->init_state == SFXGE_RXQ_STARTED,
201 	    ("rxq not started"));
202 	efx_ev_qpost(evq->common, magic);
203 }
204 
205 static void
206 sfxge_rx_schedule_refill(struct sfxge_rxq *rxq, boolean_t retrying)
207 {
208 	/* Initially retry after 100 ms, but back off in case of
209 	 * repeated failures as we probably have to wait for the
210 	 * administrator to raise the pool limit. */
211 	if (retrying)
212 		rxq->refill_delay = min(rxq->refill_delay * 2, 10 * hz);
213 	else
214 		rxq->refill_delay = hz / 10;
215 
216 	callout_reset_curcpu(&rxq->refill_callout, rxq->refill_delay,
217 			     sfxge_rx_post_refill, rxq);
218 }
219 
220 #define	SFXGE_REFILL_BATCH  64
221 
222 static void
223 sfxge_rx_qfill(struct sfxge_rxq *rxq, unsigned int target, boolean_t retrying)
224 {
225 	struct sfxge_softc *sc;
226 	unsigned int index;
227 	struct sfxge_evq *evq;
228 	unsigned int batch;
229 	unsigned int rxfill;
230 	unsigned int mblksize;
231 	int ntodo;
232 	efsys_dma_addr_t addr[SFXGE_REFILL_BATCH];
233 
234 	sc = rxq->sc;
235 	index = rxq->index;
236 	evq = sc->evq[index];
237 
238 	prefetch_read_many(sc->enp);
239 	prefetch_read_many(rxq->common);
240 
241 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
242 
243 	if (__predict_false(rxq->init_state != SFXGE_RXQ_STARTED))
244 		return;
245 
246 	rxfill = rxq->added - rxq->completed;
247 	KASSERT(rxfill <= EFX_RXQ_LIMIT(rxq->entries),
248 	    ("rxfill > EFX_RXQ_LIMIT(rxq->entries)"));
249 	ntodo = min(EFX_RXQ_LIMIT(rxq->entries) - rxfill, target);
250 	KASSERT(ntodo <= EFX_RXQ_LIMIT(rxq->entries),
251 	    ("ntodo > EFX_RQX_LIMIT(rxq->entries)"));
252 
253 	if (ntodo == 0)
254 		return;
255 
256 	batch = 0;
257 	mblksize = sc->rx_buffer_size - sc->rx_buffer_align;
258 	while (ntodo-- > 0) {
259 		unsigned int id;
260 		struct sfxge_rx_sw_desc *rx_desc;
261 		bus_dma_segment_t seg;
262 		struct mbuf *m;
263 
264 		id = (rxq->added + batch) & rxq->ptr_mask;
265 		rx_desc = &rxq->queue[id];
266 		KASSERT(rx_desc->mbuf == NULL, ("rx_desc->mbuf != NULL"));
267 
268 		rx_desc->flags = EFX_DISCARD;
269 		m = rx_desc->mbuf = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR,
270 		    sc->rx_cluster_size);
271 		if (m == NULL)
272 			break;
273 
274 		/* m_len specifies length of area to be mapped for DMA */
275 		m->m_len  = mblksize;
276 		m->m_data = (caddr_t)P2ROUNDUP((uintptr_t)m->m_data, CACHE_LINE_SIZE);
277 		m->m_data += sc->rx_buffer_align;
278 
279 		sfxge_map_mbuf_fast(rxq->mem.esm_tag, rxq->mem.esm_map, m, &seg);
280 		addr[batch++] = seg.ds_addr;
281 
282 		if (batch == SFXGE_REFILL_BATCH) {
283 			efx_rx_qpost(rxq->common, addr, mblksize, batch,
284 			    rxq->completed, rxq->added);
285 			rxq->added += batch;
286 			batch = 0;
287 		}
288 	}
289 
290 	if (ntodo != 0)
291 		sfxge_rx_schedule_refill(rxq, retrying);
292 
293 	if (batch != 0) {
294 		efx_rx_qpost(rxq->common, addr, mblksize, batch,
295 		    rxq->completed, rxq->added);
296 		rxq->added += batch;
297 	}
298 
299 	/* Make the descriptors visible to the hardware */
300 	bus_dmamap_sync(rxq->mem.esm_tag, rxq->mem.esm_map,
301 			BUS_DMASYNC_PREWRITE);
302 
303 	efx_rx_qpush(rxq->common, rxq->added, &rxq->pushed);
304 
305 	/* The queue could still be empty if no descriptors were actually
306 	 * pushed, in which case there will be no event to cause the next
307 	 * refill, so we must schedule a refill ourselves.
308 	 */
309 	if(rxq->pushed == rxq->completed) {
310 		sfxge_rx_schedule_refill(rxq, retrying);
311 	}
312 }
313 
314 void
315 sfxge_rx_qrefill(struct sfxge_rxq *rxq)
316 {
317 
318 	if (__predict_false(rxq->init_state != SFXGE_RXQ_STARTED))
319 		return;
320 
321 	/* Make sure the queue is full */
322 	sfxge_rx_qfill(rxq, EFX_RXQ_LIMIT(rxq->entries), B_TRUE);
323 }
324 
325 static void __sfxge_rx_deliver(struct sfxge_softc *sc, struct mbuf *m)
326 {
327 	struct ifnet *ifp = sc->ifnet;
328 
329 	m->m_pkthdr.rcvif = ifp;
330 	m->m_pkthdr.csum_data = 0xffff;
331 	ifp->if_input(ifp, m);
332 }
333 
334 static void
335 sfxge_rx_deliver(struct sfxge_rxq *rxq, struct sfxge_rx_sw_desc *rx_desc)
336 {
337 	struct sfxge_softc *sc = rxq->sc;
338 	struct mbuf *m = rx_desc->mbuf;
339 	int flags = rx_desc->flags;
340 	int csum_flags;
341 
342 	/* Convert checksum flags */
343 	csum_flags = (flags & EFX_CKSUM_IPV4) ?
344 		(CSUM_IP_CHECKED | CSUM_IP_VALID) : 0;
345 	if (flags & EFX_CKSUM_TCPUDP)
346 		csum_flags |= CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
347 
348 	if (flags & (EFX_PKT_IPV4 | EFX_PKT_IPV6)) {
349 		m->m_pkthdr.flowid =
350 			efx_pseudo_hdr_hash_get(rxq->common,
351 						EFX_RX_HASHALG_TOEPLITZ,
352 						mtod(m, uint8_t *));
353 		/* The hash covers a 4-tuple for TCP only */
354 		M_HASHTYPE_SET(m,
355 		    (flags & EFX_PKT_IPV4) ?
356 			((flags & EFX_PKT_TCP) ?
357 			    M_HASHTYPE_RSS_TCP_IPV4 : M_HASHTYPE_RSS_IPV4) :
358 			((flags & EFX_PKT_TCP) ?
359 			    M_HASHTYPE_RSS_TCP_IPV6 : M_HASHTYPE_RSS_IPV6));
360 	}
361 	m->m_data += sc->rx_prefix_size;
362 	m->m_len = rx_desc->size - sc->rx_prefix_size;
363 	m->m_pkthdr.len = m->m_len;
364 	m->m_pkthdr.csum_flags = csum_flags;
365 	__sfxge_rx_deliver(sc, rx_desc->mbuf);
366 
367 	rx_desc->flags = EFX_DISCARD;
368 	rx_desc->mbuf = NULL;
369 }
370 
371 #ifdef SFXGE_LRO
372 
373 static void
374 sfxge_lro_deliver(struct sfxge_lro_state *st, struct sfxge_lro_conn *c)
375 {
376 	struct sfxge_softc *sc = st->sc;
377 	struct mbuf *m = c->mbuf;
378 	struct tcphdr *c_th;
379 	int csum_flags;
380 
381 	KASSERT(m, ("no mbuf to deliver"));
382 
383 	++st->n_bursts;
384 
385 	/* Finish off packet munging and recalculate IP header checksum. */
386 	if (SFXGE_LRO_CONN_IS_TCPIPV4(c)) {
387 		struct ip *iph = c->nh;
388 		iph->ip_len = htons(iph->ip_len);
389 		iph->ip_sum = 0;
390 		iph->ip_sum = in_cksum_hdr(iph);
391 		c_th = (struct tcphdr *)(iph + 1);
392 		csum_flags = (CSUM_DATA_VALID | CSUM_PSEUDO_HDR |
393 			      CSUM_IP_CHECKED | CSUM_IP_VALID);
394 	} else {
395 		struct ip6_hdr *iph = c->nh;
396 		iph->ip6_plen = htons(iph->ip6_plen);
397 		c_th = (struct tcphdr *)(iph + 1);
398 		csum_flags = CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
399 	}
400 
401 	c_th->th_win = c->th_last->th_win;
402 	c_th->th_ack = c->th_last->th_ack;
403 	if (c_th->th_off == c->th_last->th_off) {
404 		/* Copy TCP options (take care to avoid going negative). */
405 		int optlen = ((c_th->th_off - 5) & 0xf) << 2u;
406 		memcpy(c_th + 1, c->th_last + 1, optlen);
407 	}
408 
409 	m->m_pkthdr.flowid = c->conn_hash;
410 	M_HASHTYPE_SET(m,
411 	    SFXGE_LRO_CONN_IS_TCPIPV4(c) ?
412 		M_HASHTYPE_RSS_TCP_IPV4 : M_HASHTYPE_RSS_TCP_IPV6);
413 
414 	m->m_pkthdr.csum_flags = csum_flags;
415 	__sfxge_rx_deliver(sc, m);
416 
417 	c->mbuf = NULL;
418 	c->delivered = 1;
419 }
420 
421 /* Drop the given connection, and add it to the free list. */
422 static void sfxge_lro_drop(struct sfxge_rxq *rxq, struct sfxge_lro_conn *c)
423 {
424 	unsigned bucket;
425 
426 	KASSERT(!c->mbuf, ("found orphaned mbuf"));
427 
428 	if (c->next_buf.mbuf != NULL) {
429 		sfxge_rx_deliver(rxq, &c->next_buf);
430 		LIST_REMOVE(c, active_link);
431 	}
432 
433 	bucket = c->conn_hash & rxq->lro.conns_mask;
434 	KASSERT(rxq->lro.conns_n[bucket] > 0, ("LRO: bucket fill level wrong"));
435 	--rxq->lro.conns_n[bucket];
436 	TAILQ_REMOVE(&rxq->lro.conns[bucket], c, link);
437 	TAILQ_INSERT_HEAD(&rxq->lro.free_conns, c, link);
438 }
439 
440 /* Stop tracking connections that have gone idle in order to keep hash
441  * chains short.
442  */
443 static void sfxge_lro_purge_idle(struct sfxge_rxq *rxq, unsigned now)
444 {
445 	struct sfxge_lro_conn *c;
446 	unsigned i;
447 
448 	KASSERT(LIST_EMPTY(&rxq->lro.active_conns),
449 		("found active connections"));
450 
451 	rxq->lro.last_purge_ticks = now;
452 	for (i = 0; i <= rxq->lro.conns_mask; ++i) {
453 		if (TAILQ_EMPTY(&rxq->lro.conns[i]))
454 			continue;
455 
456 		c = TAILQ_LAST(&rxq->lro.conns[i], sfxge_lro_tailq);
457 		if (now - c->last_pkt_ticks > lro_idle_ticks) {
458 			++rxq->lro.n_drop_idle;
459 			sfxge_lro_drop(rxq, c);
460 		}
461 	}
462 }
463 
464 static void
465 sfxge_lro_merge(struct sfxge_lro_state *st, struct sfxge_lro_conn *c,
466 		struct mbuf *mbuf, struct tcphdr *th)
467 {
468 	struct tcphdr *c_th;
469 
470 	/* Tack the new mbuf onto the chain. */
471 	KASSERT(!mbuf->m_next, ("mbuf already chained"));
472 	c->mbuf_tail->m_next = mbuf;
473 	c->mbuf_tail = mbuf;
474 
475 	/* Increase length appropriately */
476 	c->mbuf->m_pkthdr.len += mbuf->m_len;
477 
478 	/* Update the connection state flags */
479 	if (SFXGE_LRO_CONN_IS_TCPIPV4(c)) {
480 		struct ip *iph = c->nh;
481 		iph->ip_len += mbuf->m_len;
482 		c_th = (struct tcphdr *)(iph + 1);
483 	} else {
484 		struct ip6_hdr *iph = c->nh;
485 		iph->ip6_plen += mbuf->m_len;
486 		c_th = (struct tcphdr *)(iph + 1);
487 	}
488 	c_th->th_flags |= (th->th_flags & TH_PUSH);
489 	c->th_last = th;
490 	++st->n_merges;
491 
492 	/* Pass packet up now if another segment could overflow the IP
493 	 * length.
494 	 */
495 	if (c->mbuf->m_pkthdr.len > 65536 - 9200)
496 		sfxge_lro_deliver(st, c);
497 }
498 
499 static void
500 sfxge_lro_start(struct sfxge_lro_state *st, struct sfxge_lro_conn *c,
501 		struct mbuf *mbuf, void *nh, struct tcphdr *th)
502 {
503 	/* Start the chain */
504 	c->mbuf = mbuf;
505 	c->mbuf_tail = c->mbuf;
506 	c->nh = nh;
507 	c->th_last = th;
508 
509 	mbuf->m_pkthdr.len = mbuf->m_len;
510 
511 	/* Mangle header fields for later processing */
512 	if (SFXGE_LRO_CONN_IS_TCPIPV4(c)) {
513 		struct ip *iph = nh;
514 		iph->ip_len = ntohs(iph->ip_len);
515 	} else {
516 		struct ip6_hdr *iph = nh;
517 		iph->ip6_plen = ntohs(iph->ip6_plen);
518 	}
519 }
520 
521 /* Try to merge or otherwise hold or deliver (as appropriate) the
522  * packet buffered for this connection (c->next_buf).  Return a flag
523  * indicating whether the connection is still active for LRO purposes.
524  */
525 static int
526 sfxge_lro_try_merge(struct sfxge_rxq *rxq, struct sfxge_lro_conn *c)
527 {
528 	struct sfxge_rx_sw_desc *rx_buf = &c->next_buf;
529 	char *eh = c->next_eh;
530 	int data_length, hdr_length, dont_merge;
531 	unsigned th_seq, pkt_length;
532 	struct tcphdr *th;
533 	unsigned now;
534 
535 	if (SFXGE_LRO_CONN_IS_TCPIPV4(c)) {
536 		struct ip *iph = c->next_nh;
537 		th = (struct tcphdr *)(iph + 1);
538 		pkt_length = ntohs(iph->ip_len) + (char *) iph - eh;
539 	} else {
540 		struct ip6_hdr *iph = c->next_nh;
541 		th = (struct tcphdr *)(iph + 1);
542 		pkt_length = ntohs(iph->ip6_plen) + (char *) th - eh;
543 	}
544 
545 	hdr_length = (char *) th + th->th_off * 4 - eh;
546 	data_length = (min(pkt_length, rx_buf->size - rxq->sc->rx_prefix_size) -
547 		       hdr_length);
548 	th_seq = ntohl(th->th_seq);
549 	dont_merge = ((data_length <= 0)
550 		      | (th->th_flags & (TH_URG | TH_SYN | TH_RST | TH_FIN)));
551 
552 	/* Check for options other than aligned timestamp. */
553 	if (th->th_off != 5) {
554 		const uint32_t *opt_ptr = (const uint32_t *) (th + 1);
555 		if (th->th_off == 8 &&
556 		    opt_ptr[0] == ntohl((TCPOPT_NOP << 24) |
557 					(TCPOPT_NOP << 16) |
558 					(TCPOPT_TIMESTAMP << 8) |
559 					TCPOLEN_TIMESTAMP)) {
560 			/* timestamp option -- okay */
561 		} else {
562 			dont_merge = 1;
563 		}
564 	}
565 
566 	if (__predict_false(th_seq != c->next_seq)) {
567 		/* Out-of-order, so start counting again. */
568 		if (c->mbuf != NULL)
569 			sfxge_lro_deliver(&rxq->lro, c);
570 		c->n_in_order_pkts -= lro_loss_packets;
571 		c->next_seq = th_seq + data_length;
572 		++rxq->lro.n_misorder;
573 		goto deliver_buf_out;
574 	}
575 	c->next_seq = th_seq + data_length;
576 
577 	now = ticks;
578 	if (now - c->last_pkt_ticks > lro_idle_ticks) {
579 		++rxq->lro.n_drop_idle;
580 		if (c->mbuf != NULL)
581 			sfxge_lro_deliver(&rxq->lro, c);
582 		sfxge_lro_drop(rxq, c);
583 		return (0);
584 	}
585 	c->last_pkt_ticks = ticks;
586 
587 	if (c->n_in_order_pkts < lro_slow_start_packets) {
588 		/* May be in slow-start, so don't merge. */
589 		++rxq->lro.n_slow_start;
590 		++c->n_in_order_pkts;
591 		goto deliver_buf_out;
592 	}
593 
594 	if (__predict_false(dont_merge)) {
595 		if (c->mbuf != NULL)
596 			sfxge_lro_deliver(&rxq->lro, c);
597 		if (th->th_flags & (TH_FIN | TH_RST)) {
598 			++rxq->lro.n_drop_closed;
599 			sfxge_lro_drop(rxq, c);
600 			return (0);
601 		}
602 		goto deliver_buf_out;
603 	}
604 
605 	rx_buf->mbuf->m_data += rxq->sc->rx_prefix_size;
606 
607 	if (__predict_true(c->mbuf != NULL)) {
608 		/* Remove headers and any padding */
609 		rx_buf->mbuf->m_data += hdr_length;
610 		rx_buf->mbuf->m_len = data_length;
611 
612 		sfxge_lro_merge(&rxq->lro, c, rx_buf->mbuf, th);
613 	} else {
614 		/* Remove any padding */
615 		rx_buf->mbuf->m_len = pkt_length;
616 
617 		sfxge_lro_start(&rxq->lro, c, rx_buf->mbuf, c->next_nh, th);
618 	}
619 
620 	rx_buf->mbuf = NULL;
621 	return (1);
622 
623  deliver_buf_out:
624 	sfxge_rx_deliver(rxq, rx_buf);
625 	return (1);
626 }
627 
628 static void sfxge_lro_new_conn(struct sfxge_lro_state *st, uint32_t conn_hash,
629 			       uint16_t l2_id, void *nh, struct tcphdr *th)
630 {
631 	unsigned bucket = conn_hash & st->conns_mask;
632 	struct sfxge_lro_conn *c;
633 
634 	if (st->conns_n[bucket] >= lro_chain_max) {
635 		++st->n_too_many;
636 		return;
637 	}
638 
639 	if (!TAILQ_EMPTY(&st->free_conns)) {
640 		c = TAILQ_FIRST(&st->free_conns);
641 		TAILQ_REMOVE(&st->free_conns, c, link);
642 	} else {
643 		c = malloc(sizeof(*c), M_SFXGE, M_NOWAIT);
644 		if (c == NULL)
645 			return;
646 		c->mbuf = NULL;
647 		c->next_buf.mbuf = NULL;
648 	}
649 
650 	/* Create the connection tracking data */
651 	++st->conns_n[bucket];
652 	TAILQ_INSERT_HEAD(&st->conns[bucket], c, link);
653 	c->l2_id = l2_id;
654 	c->conn_hash = conn_hash;
655 	c->source = th->th_sport;
656 	c->dest = th->th_dport;
657 	c->n_in_order_pkts = 0;
658 	c->last_pkt_ticks = *(volatile int *)&ticks;
659 	c->delivered = 0;
660 	++st->n_new_stream;
661 	/* NB. We don't initialise c->next_seq, and it doesn't matter what
662 	 * value it has.  Most likely the next packet received for this
663 	 * connection will not match -- no harm done.
664 	 */
665 }
666 
667 /* Process mbuf and decide whether to dispatch it to the stack now or
668  * later.
669  */
670 static void
671 sfxge_lro(struct sfxge_rxq *rxq, struct sfxge_rx_sw_desc *rx_buf)
672 {
673 	struct sfxge_softc *sc = rxq->sc;
674 	struct mbuf *m = rx_buf->mbuf;
675 	struct ether_header *eh;
676 	struct sfxge_lro_conn *c;
677 	uint16_t l2_id;
678 	uint16_t l3_proto;
679 	void *nh;
680 	struct tcphdr *th;
681 	uint32_t conn_hash;
682 	unsigned bucket;
683 
684 	/* Get the hardware hash */
685 	conn_hash = efx_pseudo_hdr_hash_get(rxq->common,
686 					    EFX_RX_HASHALG_TOEPLITZ,
687 					    mtod(m, uint8_t *));
688 
689 	eh = (struct ether_header *)(m->m_data + sc->rx_prefix_size);
690 	if (eh->ether_type == htons(ETHERTYPE_VLAN)) {
691 		struct ether_vlan_header *veh = (struct ether_vlan_header *)eh;
692 		l2_id = EVL_VLANOFTAG(ntohs(veh->evl_tag)) |
693 			SFXGE_LRO_L2_ID_VLAN;
694 		l3_proto = veh->evl_proto;
695 		nh = veh + 1;
696 	} else {
697 		l2_id = 0;
698 		l3_proto = eh->ether_type;
699 		nh = eh + 1;
700 	}
701 
702 	/* Check whether this is a suitable packet (unfragmented
703 	 * TCP/IPv4 or TCP/IPv6).  If so, find the TCP header and
704 	 * length, and compute a hash if necessary.  If not, return.
705 	 */
706 	if (l3_proto == htons(ETHERTYPE_IP)) {
707 		struct ip *iph = nh;
708 
709 		KASSERT(iph->ip_p == IPPROTO_TCP,
710 		    ("IPv4 protocol is not TCP, but packet marker is set"));
711 		if ((iph->ip_hl - (sizeof(*iph) >> 2u)) |
712 		    (iph->ip_off & htons(IP_MF | IP_OFFMASK)))
713 			goto deliver_now;
714 		th = (struct tcphdr *)(iph + 1);
715 	} else if (l3_proto == htons(ETHERTYPE_IPV6)) {
716 		struct ip6_hdr *iph = nh;
717 
718 		KASSERT(iph->ip6_nxt == IPPROTO_TCP,
719 		    ("IPv6 next header is not TCP, but packet marker is set"));
720 		l2_id |= SFXGE_LRO_L2_ID_IPV6;
721 		th = (struct tcphdr *)(iph + 1);
722 	} else {
723 		goto deliver_now;
724 	}
725 
726 	bucket = conn_hash & rxq->lro.conns_mask;
727 
728 	TAILQ_FOREACH(c, &rxq->lro.conns[bucket], link) {
729 		if ((c->l2_id - l2_id) | (c->conn_hash - conn_hash))
730 			continue;
731 		if ((c->source - th->th_sport) | (c->dest - th->th_dport))
732 			continue;
733 		if (c->mbuf != NULL) {
734 			if (SFXGE_LRO_CONN_IS_TCPIPV4(c)) {
735 				struct ip *c_iph, *iph = nh;
736 				c_iph = c->nh;
737 				if ((c_iph->ip_src.s_addr - iph->ip_src.s_addr) |
738 				    (c_iph->ip_dst.s_addr - iph->ip_dst.s_addr))
739 					continue;
740 			} else {
741 				struct ip6_hdr *c_iph, *iph = nh;
742 				c_iph = c->nh;
743 				if (ipv6_addr_cmp(&c_iph->ip6_src, &iph->ip6_src) |
744 				    ipv6_addr_cmp(&c_iph->ip6_dst, &iph->ip6_dst))
745 					continue;
746 			}
747 		}
748 
749 		/* Re-insert at head of list to reduce lookup time. */
750 		TAILQ_REMOVE(&rxq->lro.conns[bucket], c, link);
751 		TAILQ_INSERT_HEAD(&rxq->lro.conns[bucket], c, link);
752 
753 		if (c->next_buf.mbuf != NULL) {
754 			if (!sfxge_lro_try_merge(rxq, c))
755 				goto deliver_now;
756 		} else {
757 			LIST_INSERT_HEAD(&rxq->lro.active_conns, c,
758 			    active_link);
759 		}
760 		c->next_buf = *rx_buf;
761 		c->next_eh = eh;
762 		c->next_nh = nh;
763 
764 		rx_buf->mbuf = NULL;
765 		rx_buf->flags = EFX_DISCARD;
766 		return;
767 	}
768 
769 	sfxge_lro_new_conn(&rxq->lro, conn_hash, l2_id, nh, th);
770  deliver_now:
771 	sfxge_rx_deliver(rxq, rx_buf);
772 }
773 
774 static void sfxge_lro_end_of_burst(struct sfxge_rxq *rxq)
775 {
776 	struct sfxge_lro_state *st = &rxq->lro;
777 	struct sfxge_lro_conn *c;
778 	unsigned t;
779 
780 	while (!LIST_EMPTY(&st->active_conns)) {
781 		c = LIST_FIRST(&st->active_conns);
782 		if (!c->delivered && c->mbuf != NULL)
783 			sfxge_lro_deliver(st, c);
784 		if (sfxge_lro_try_merge(rxq, c)) {
785 			if (c->mbuf != NULL)
786 				sfxge_lro_deliver(st, c);
787 			LIST_REMOVE(c, active_link);
788 		}
789 		c->delivered = 0;
790 	}
791 
792 	t = *(volatile int *)&ticks;
793 	if (__predict_false(t != st->last_purge_ticks))
794 		sfxge_lro_purge_idle(rxq, t);
795 }
796 
797 #else	/* !SFXGE_LRO */
798 
799 static void
800 sfxge_lro(struct sfxge_rxq *rxq, struct sfxge_rx_sw_desc *rx_buf)
801 {
802 }
803 
804 static void
805 sfxge_lro_end_of_burst(struct sfxge_rxq *rxq)
806 {
807 }
808 
809 #endif	/* SFXGE_LRO */
810 
811 void
812 sfxge_rx_qcomplete(struct sfxge_rxq *rxq, boolean_t eop)
813 {
814 	struct sfxge_softc *sc = rxq->sc;
815 	int if_capenable = sc->ifnet->if_capenable;
816 	int lro_enabled = if_capenable & IFCAP_LRO;
817 	unsigned int index;
818 	struct sfxge_evq *evq;
819 	unsigned int completed;
820 	unsigned int level;
821 	struct mbuf *m;
822 	struct sfxge_rx_sw_desc *prev = NULL;
823 
824 	index = rxq->index;
825 	evq = sc->evq[index];
826 
827 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
828 
829 	completed = rxq->completed;
830 	while (completed != rxq->pending) {
831 		unsigned int id;
832 		struct sfxge_rx_sw_desc *rx_desc;
833 
834 		id = completed++ & rxq->ptr_mask;
835 		rx_desc = &rxq->queue[id];
836 		m = rx_desc->mbuf;
837 
838 		if (__predict_false(rxq->init_state != SFXGE_RXQ_STARTED))
839 			goto discard;
840 
841 		if (rx_desc->flags & (EFX_ADDR_MISMATCH | EFX_DISCARD))
842 			goto discard;
843 
844 		/* Read the length from the pseudo header if required */
845 		if (rx_desc->flags & EFX_PKT_PREFIX_LEN) {
846 			uint16_t tmp_size;
847 			int rc;
848 			rc = efx_pseudo_hdr_pkt_length_get(rxq->common,
849 							   mtod(m, uint8_t *),
850 							   &tmp_size);
851 			KASSERT(rc == 0, ("cannot get packet length: %d", rc));
852 			rx_desc->size = (int)tmp_size + sc->rx_prefix_size;
853 		}
854 
855 		prefetch_read_many(mtod(m, caddr_t));
856 
857 		switch (rx_desc->flags & (EFX_PKT_IPV4 | EFX_PKT_IPV6)) {
858 		case EFX_PKT_IPV4:
859 			if (~if_capenable & IFCAP_RXCSUM)
860 				rx_desc->flags &=
861 				    ~(EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP);
862 			break;
863 		case EFX_PKT_IPV6:
864 			if (~if_capenable & IFCAP_RXCSUM_IPV6)
865 				rx_desc->flags &= ~EFX_CKSUM_TCPUDP;
866 			break;
867 		case 0:
868 			/* Check for loopback packets */
869 			{
870 				struct ether_header *etherhp;
871 
872 				/*LINTED*/
873 				etherhp = mtod(m, struct ether_header *);
874 
875 				if (etherhp->ether_type ==
876 				    htons(SFXGE_ETHERTYPE_LOOPBACK)) {
877 					EFSYS_PROBE(loopback);
878 
879 					rxq->loopback++;
880 					goto discard;
881 				}
882 			}
883 			break;
884 		default:
885 			KASSERT(B_FALSE,
886 			    ("Rx descriptor with both IPv4 and IPv6 flags"));
887 			goto discard;
888 		}
889 
890 		/* Pass packet up the stack or into LRO (pipelined) */
891 		if (prev != NULL) {
892 			if (lro_enabled &&
893 			    ((prev->flags & (EFX_PKT_TCP | EFX_CKSUM_TCPUDP)) ==
894 			     (EFX_PKT_TCP | EFX_CKSUM_TCPUDP)))
895 				sfxge_lro(rxq, prev);
896 			else
897 				sfxge_rx_deliver(rxq, prev);
898 		}
899 		prev = rx_desc;
900 		continue;
901 
902 discard:
903 		/* Return the packet to the pool */
904 		m_free(m);
905 		rx_desc->mbuf = NULL;
906 	}
907 	rxq->completed = completed;
908 
909 	level = rxq->added - rxq->completed;
910 
911 	/* Pass last packet up the stack or into LRO */
912 	if (prev != NULL) {
913 		if (lro_enabled &&
914 		    ((prev->flags & (EFX_PKT_TCP | EFX_CKSUM_TCPUDP)) ==
915 		     (EFX_PKT_TCP | EFX_CKSUM_TCPUDP)))
916 			sfxge_lro(rxq, prev);
917 		else
918 			sfxge_rx_deliver(rxq, prev);
919 	}
920 
921 	/*
922 	 * If there are any pending flows and this is the end of the
923 	 * poll then they must be completed.
924 	 */
925 	if (eop)
926 		sfxge_lro_end_of_burst(rxq);
927 
928 	/* Top up the queue if necessary */
929 	if (level < rxq->refill_threshold)
930 		sfxge_rx_qfill(rxq, EFX_RXQ_LIMIT(rxq->entries), B_FALSE);
931 }
932 
933 static void
934 sfxge_rx_qstop(struct sfxge_softc *sc, unsigned int index)
935 {
936 	struct sfxge_rxq *rxq;
937 	struct sfxge_evq *evq;
938 	unsigned int count;
939 	unsigned int retry = 3;
940 
941 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
942 
943 	rxq = sc->rxq[index];
944 	evq = sc->evq[index];
945 
946 	SFXGE_EVQ_LOCK(evq);
947 
948 	KASSERT(rxq->init_state == SFXGE_RXQ_STARTED,
949 	    ("rxq not started"));
950 
951 	rxq->init_state = SFXGE_RXQ_INITIALIZED;
952 
953 	callout_stop(&rxq->refill_callout);
954 
955 	while (rxq->flush_state != SFXGE_FLUSH_DONE && retry != 0) {
956 		rxq->flush_state = SFXGE_FLUSH_PENDING;
957 
958 		SFXGE_EVQ_UNLOCK(evq);
959 
960 		/* Flush the receive queue */
961 		if (efx_rx_qflush(rxq->common) != 0) {
962 			SFXGE_EVQ_LOCK(evq);
963 			rxq->flush_state = SFXGE_FLUSH_FAILED;
964 			break;
965 		}
966 
967 		count = 0;
968 		do {
969 			/* Spin for 100 ms */
970 			DELAY(100000);
971 
972 			if (rxq->flush_state != SFXGE_FLUSH_PENDING)
973 				break;
974 
975 		} while (++count < 20);
976 
977 		SFXGE_EVQ_LOCK(evq);
978 
979 		if (rxq->flush_state == SFXGE_FLUSH_PENDING) {
980 			/* Flush timeout - neither done nor failed */
981 			log(LOG_ERR, "%s: Cannot flush Rx queue %u\n",
982 			    device_get_nameunit(sc->dev), index);
983 			rxq->flush_state = SFXGE_FLUSH_DONE;
984 		}
985 		retry--;
986 	}
987 	if (rxq->flush_state == SFXGE_FLUSH_FAILED) {
988 		log(LOG_ERR, "%s: Flushing Rx queue %u failed\n",
989 		    device_get_nameunit(sc->dev), index);
990 		rxq->flush_state = SFXGE_FLUSH_DONE;
991 	}
992 
993 	rxq->pending = rxq->added;
994 	sfxge_rx_qcomplete(rxq, B_TRUE);
995 
996 	KASSERT(rxq->completed == rxq->pending,
997 	    ("rxq->completed != rxq->pending"));
998 
999 	rxq->added = 0;
1000 	rxq->pushed = 0;
1001 	rxq->pending = 0;
1002 	rxq->completed = 0;
1003 	rxq->loopback = 0;
1004 
1005 	/* Destroy the common code receive queue. */
1006 	efx_rx_qdestroy(rxq->common);
1007 
1008 	efx_sram_buf_tbl_clear(sc->enp, rxq->buf_base_id,
1009 	    EFX_RXQ_NBUFS(sc->rxq_entries));
1010 
1011 	SFXGE_EVQ_UNLOCK(evq);
1012 }
1013 
1014 static int
1015 sfxge_rx_qstart(struct sfxge_softc *sc, unsigned int index)
1016 {
1017 	struct sfxge_rxq *rxq;
1018 	efsys_mem_t *esmp;
1019 	struct sfxge_evq *evq;
1020 	int rc;
1021 
1022 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1023 
1024 	rxq = sc->rxq[index];
1025 	esmp = &rxq->mem;
1026 	evq = sc->evq[index];
1027 
1028 	KASSERT(rxq->init_state == SFXGE_RXQ_INITIALIZED,
1029 	    ("rxq->init_state != SFXGE_RXQ_INITIALIZED"));
1030 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1031 	    ("evq->init_state != SFXGE_EVQ_STARTED"));
1032 
1033 	/* Program the buffer table. */
1034 	if ((rc = efx_sram_buf_tbl_set(sc->enp, rxq->buf_base_id, esmp,
1035 	    EFX_RXQ_NBUFS(sc->rxq_entries))) != 0)
1036 		return (rc);
1037 
1038 	/* Create the common code receive queue. */
1039 	if ((rc = efx_rx_qcreate(sc->enp, index, 0, EFX_RXQ_TYPE_DEFAULT,
1040 	    esmp, sc->rxq_entries, rxq->buf_base_id, evq->common,
1041 	    &rxq->common)) != 0)
1042 		goto fail;
1043 
1044 	SFXGE_EVQ_LOCK(evq);
1045 
1046 	/* Enable the receive queue. */
1047 	efx_rx_qenable(rxq->common);
1048 
1049 	rxq->init_state = SFXGE_RXQ_STARTED;
1050 	rxq->flush_state = SFXGE_FLUSH_REQUIRED;
1051 
1052 	/* Try to fill the queue from the pool. */
1053 	sfxge_rx_qfill(rxq, EFX_RXQ_LIMIT(sc->rxq_entries), B_FALSE);
1054 
1055 	SFXGE_EVQ_UNLOCK(evq);
1056 
1057 	return (0);
1058 
1059 fail:
1060 	efx_sram_buf_tbl_clear(sc->enp, rxq->buf_base_id,
1061 	    EFX_RXQ_NBUFS(sc->rxq_entries));
1062 	return (rc);
1063 }
1064 
1065 void
1066 sfxge_rx_stop(struct sfxge_softc *sc)
1067 {
1068 	int index;
1069 
1070 	efx_mac_filter_default_rxq_clear(sc->enp);
1071 
1072 	/* Stop the receive queue(s) */
1073 	index = sc->rxq_count;
1074 	while (--index >= 0)
1075 		sfxge_rx_qstop(sc, index);
1076 
1077 	sc->rx_prefix_size = 0;
1078 	sc->rx_buffer_size = 0;
1079 
1080 	efx_rx_fini(sc->enp);
1081 }
1082 
1083 int
1084 sfxge_rx_start(struct sfxge_softc *sc)
1085 {
1086 	struct sfxge_intr *intr;
1087 	const efx_nic_cfg_t *encp;
1088 	size_t hdrlen, align, reserved;
1089 	int index;
1090 	int rc;
1091 
1092 	intr = &sc->intr;
1093 
1094 	/* Initialize the common code receive module. */
1095 	if ((rc = efx_rx_init(sc->enp)) != 0)
1096 		return (rc);
1097 
1098 	encp = efx_nic_cfg_get(sc->enp);
1099 	sc->rx_buffer_size = EFX_MAC_PDU(sc->ifnet->if_mtu);
1100 
1101 	/* Calculate the receive packet buffer size. */
1102 	sc->rx_prefix_size = encp->enc_rx_prefix_size;
1103 
1104 	/* Ensure IP headers are 32bit aligned */
1105 	hdrlen = sc->rx_prefix_size + sizeof (struct ether_header);
1106 	sc->rx_buffer_align = P2ROUNDUP(hdrlen, 4) - hdrlen;
1107 
1108 	sc->rx_buffer_size += sc->rx_buffer_align;
1109 
1110 	/* Align end of packet buffer for RX DMA end padding */
1111 	align = MAX(1, encp->enc_rx_buf_align_end);
1112 	EFSYS_ASSERT(ISP2(align));
1113 	sc->rx_buffer_size = P2ROUNDUP(sc->rx_buffer_size, align);
1114 
1115 	/*
1116 	 * Standard mbuf zones only guarantee pointer-size alignment;
1117 	 * we need extra space to align to the cache line
1118 	 */
1119 	reserved = sc->rx_buffer_size + CACHE_LINE_SIZE;
1120 
1121 	/* Select zone for packet buffers */
1122 	if (reserved <= MCLBYTES)
1123 		sc->rx_cluster_size = MCLBYTES;
1124 	else if (reserved <= MJUMPAGESIZE)
1125 		sc->rx_cluster_size = MJUMPAGESIZE;
1126 	else if (reserved <= MJUM9BYTES)
1127 		sc->rx_cluster_size = MJUM9BYTES;
1128 	else
1129 		sc->rx_cluster_size = MJUM16BYTES;
1130 
1131 	/*
1132 	 * Set up the scale table.  Enable all hash types and hash insertion.
1133 	 */
1134 	for (index = 0; index < nitems(sc->rx_indir_table); index++)
1135 #ifdef RSS
1136 		sc->rx_indir_table[index] =
1137 			rss_get_indirection_to_bucket(index) % sc->rxq_count;
1138 #else
1139 		sc->rx_indir_table[index] = index % sc->rxq_count;
1140 #endif
1141 	if ((rc = efx_rx_scale_tbl_set(sc->enp, sc->rx_indir_table,
1142 				       nitems(sc->rx_indir_table))) != 0)
1143 		goto fail;
1144 	(void)efx_rx_scale_mode_set(sc->enp, EFX_RX_HASHALG_TOEPLITZ,
1145 	    EFX_RX_HASH_IPV4 | EFX_RX_HASH_TCPIPV4 |
1146 	    EFX_RX_HASH_IPV6 | EFX_RX_HASH_TCPIPV6, B_TRUE);
1147 
1148 #ifdef RSS
1149 	rss_getkey(toep_key);
1150 #endif
1151 	if ((rc = efx_rx_scale_key_set(sc->enp, toep_key,
1152 				       sizeof(toep_key))) != 0)
1153 		goto fail;
1154 
1155 	/* Start the receive queue(s). */
1156 	for (index = 0; index < sc->rxq_count; index++) {
1157 		if ((rc = sfxge_rx_qstart(sc, index)) != 0)
1158 			goto fail2;
1159 	}
1160 
1161 	rc = efx_mac_filter_default_rxq_set(sc->enp, sc->rxq[0]->common,
1162 					    sc->intr.n_alloc > 1);
1163 	if (rc != 0)
1164 		goto fail3;
1165 
1166 	return (0);
1167 
1168 fail3:
1169 fail2:
1170 	while (--index >= 0)
1171 		sfxge_rx_qstop(sc, index);
1172 
1173 fail:
1174 	efx_rx_fini(sc->enp);
1175 
1176 	return (rc);
1177 }
1178 
1179 #ifdef SFXGE_LRO
1180 
1181 static void sfxge_lro_init(struct sfxge_rxq *rxq)
1182 {
1183 	struct sfxge_lro_state *st = &rxq->lro;
1184 	unsigned i;
1185 
1186 	st->conns_mask = lro_table_size - 1;
1187 	KASSERT(!((st->conns_mask + 1) & st->conns_mask),
1188 		("lro_table_size must be a power of 2"));
1189 	st->sc = rxq->sc;
1190 	st->conns = malloc((st->conns_mask + 1) * sizeof(st->conns[0]),
1191 			   M_SFXGE, M_WAITOK);
1192 	st->conns_n = malloc((st->conns_mask + 1) * sizeof(st->conns_n[0]),
1193 			     M_SFXGE, M_WAITOK);
1194 	for (i = 0; i <= st->conns_mask; ++i) {
1195 		TAILQ_INIT(&st->conns[i]);
1196 		st->conns_n[i] = 0;
1197 	}
1198 	LIST_INIT(&st->active_conns);
1199 	TAILQ_INIT(&st->free_conns);
1200 }
1201 
1202 static void sfxge_lro_fini(struct sfxge_rxq *rxq)
1203 {
1204 	struct sfxge_lro_state *st = &rxq->lro;
1205 	struct sfxge_lro_conn *c;
1206 	unsigned i;
1207 
1208 	/* Return cleanly if sfxge_lro_init() has not been called. */
1209 	if (st->conns == NULL)
1210 		return;
1211 
1212 	KASSERT(LIST_EMPTY(&st->active_conns), ("found active connections"));
1213 
1214 	for (i = 0; i <= st->conns_mask; ++i) {
1215 		while (!TAILQ_EMPTY(&st->conns[i])) {
1216 			c = TAILQ_LAST(&st->conns[i], sfxge_lro_tailq);
1217 			sfxge_lro_drop(rxq, c);
1218 		}
1219 	}
1220 
1221 	while (!TAILQ_EMPTY(&st->free_conns)) {
1222 		c = TAILQ_FIRST(&st->free_conns);
1223 		TAILQ_REMOVE(&st->free_conns, c, link);
1224 		KASSERT(!c->mbuf, ("found orphaned mbuf"));
1225 		free(c, M_SFXGE);
1226 	}
1227 
1228 	free(st->conns_n, M_SFXGE);
1229 	free(st->conns, M_SFXGE);
1230 	st->conns = NULL;
1231 }
1232 
1233 #else
1234 
1235 static void
1236 sfxge_lro_init(struct sfxge_rxq *rxq)
1237 {
1238 }
1239 
1240 static void
1241 sfxge_lro_fini(struct sfxge_rxq *rxq)
1242 {
1243 }
1244 
1245 #endif	/* SFXGE_LRO */
1246 
1247 static void
1248 sfxge_rx_qfini(struct sfxge_softc *sc, unsigned int index)
1249 {
1250 	struct sfxge_rxq *rxq;
1251 
1252 	rxq = sc->rxq[index];
1253 
1254 	KASSERT(rxq->init_state == SFXGE_RXQ_INITIALIZED,
1255 	    ("rxq->init_state != SFXGE_RXQ_INITIALIZED"));
1256 
1257 	/* Free the context array and the flow table. */
1258 	free(rxq->queue, M_SFXGE);
1259 	sfxge_lro_fini(rxq);
1260 
1261 	/* Release DMA memory. */
1262 	sfxge_dma_free(&rxq->mem);
1263 
1264 	sc->rxq[index] = NULL;
1265 
1266 	free(rxq, M_SFXGE);
1267 }
1268 
1269 static int
1270 sfxge_rx_qinit(struct sfxge_softc *sc, unsigned int index)
1271 {
1272 	struct sfxge_rxq *rxq;
1273 	struct sfxge_evq *evq;
1274 	efsys_mem_t *esmp;
1275 	int rc;
1276 
1277 	KASSERT(index < sc->rxq_count, ("index >= %d", sc->rxq_count));
1278 
1279 	rxq = malloc(sizeof(struct sfxge_rxq), M_SFXGE, M_ZERO | M_WAITOK);
1280 	rxq->sc = sc;
1281 	rxq->index = index;
1282 	rxq->entries = sc->rxq_entries;
1283 	rxq->ptr_mask = rxq->entries - 1;
1284 	rxq->refill_threshold = RX_REFILL_THRESHOLD(rxq->entries);
1285 
1286 	sc->rxq[index] = rxq;
1287 	esmp = &rxq->mem;
1288 
1289 	evq = sc->evq[index];
1290 
1291 	/* Allocate and zero DMA space. */
1292 	if ((rc = sfxge_dma_alloc(sc, EFX_RXQ_SIZE(sc->rxq_entries), esmp)) != 0)
1293 		return (rc);
1294 
1295 	/* Allocate buffer table entries. */
1296 	sfxge_sram_buf_tbl_alloc(sc, EFX_RXQ_NBUFS(sc->rxq_entries),
1297 				 &rxq->buf_base_id);
1298 
1299 	/* Allocate the context array and the flow table. */
1300 	rxq->queue = malloc(sizeof(struct sfxge_rx_sw_desc) * sc->rxq_entries,
1301 	    M_SFXGE, M_WAITOK | M_ZERO);
1302 	sfxge_lro_init(rxq);
1303 
1304 	callout_init(&rxq->refill_callout, 1);
1305 
1306 	rxq->init_state = SFXGE_RXQ_INITIALIZED;
1307 
1308 	return (0);
1309 }
1310 
1311 static const struct {
1312 	const char *name;
1313 	size_t offset;
1314 } sfxge_rx_stats[] = {
1315 #define	SFXGE_RX_STAT(name, member) \
1316 	{ #name, offsetof(struct sfxge_rxq, member) }
1317 #ifdef SFXGE_LRO
1318 	SFXGE_RX_STAT(lro_merges, lro.n_merges),
1319 	SFXGE_RX_STAT(lro_bursts, lro.n_bursts),
1320 	SFXGE_RX_STAT(lro_slow_start, lro.n_slow_start),
1321 	SFXGE_RX_STAT(lro_misorder, lro.n_misorder),
1322 	SFXGE_RX_STAT(lro_too_many, lro.n_too_many),
1323 	SFXGE_RX_STAT(lro_new_stream, lro.n_new_stream),
1324 	SFXGE_RX_STAT(lro_drop_idle, lro.n_drop_idle),
1325 	SFXGE_RX_STAT(lro_drop_closed, lro.n_drop_closed)
1326 #endif
1327 };
1328 
1329 static int
1330 sfxge_rx_stat_handler(SYSCTL_HANDLER_ARGS)
1331 {
1332 	struct sfxge_softc *sc = arg1;
1333 	unsigned int id = arg2;
1334 	unsigned int sum, index;
1335 
1336 	/* Sum across all RX queues */
1337 	sum = 0;
1338 	for (index = 0; index < sc->rxq_count; index++)
1339 		sum += *(unsigned int *)((caddr_t)sc->rxq[index] +
1340 					 sfxge_rx_stats[id].offset);
1341 
1342 	return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1343 }
1344 
1345 static void
1346 sfxge_rx_stat_init(struct sfxge_softc *sc)
1347 {
1348 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1349 	struct sysctl_oid_list *stat_list;
1350 	unsigned int id;
1351 
1352 	stat_list = SYSCTL_CHILDREN(sc->stats_node);
1353 
1354 	for (id = 0; id < nitems(sfxge_rx_stats); id++) {
1355 		SYSCTL_ADD_PROC(
1356 			ctx, stat_list,
1357 			OID_AUTO, sfxge_rx_stats[id].name,
1358 			CTLTYPE_UINT|CTLFLAG_RD,
1359 			sc, id, sfxge_rx_stat_handler, "IU",
1360 			"");
1361 	}
1362 }
1363 
1364 void
1365 sfxge_rx_fini(struct sfxge_softc *sc)
1366 {
1367 	int index;
1368 
1369 	index = sc->rxq_count;
1370 	while (--index >= 0)
1371 		sfxge_rx_qfini(sc, index);
1372 
1373 	sc->rxq_count = 0;
1374 }
1375 
1376 int
1377 sfxge_rx_init(struct sfxge_softc *sc)
1378 {
1379 	struct sfxge_intr *intr;
1380 	int index;
1381 	int rc;
1382 
1383 #ifdef SFXGE_LRO
1384 	if (!ISP2(lro_table_size)) {
1385 		log(LOG_ERR, "%s=%u must be power of 2",
1386 		    SFXGE_LRO_PARAM(table_size), lro_table_size);
1387 		rc = EINVAL;
1388 		goto fail_lro_table_size;
1389 	}
1390 
1391 	if (lro_idle_ticks == 0)
1392 		lro_idle_ticks = hz / 10 + 1; /* 100 ms */
1393 #endif
1394 
1395 	intr = &sc->intr;
1396 
1397 	sc->rxq_count = intr->n_alloc;
1398 
1399 	KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1400 	    ("intr->state != SFXGE_INTR_INITIALIZED"));
1401 
1402 	/* Initialize the receive queue(s) - one per interrupt. */
1403 	for (index = 0; index < sc->rxq_count; index++) {
1404 		if ((rc = sfxge_rx_qinit(sc, index)) != 0)
1405 			goto fail;
1406 	}
1407 
1408 	sfxge_rx_stat_init(sc);
1409 
1410 	return (0);
1411 
1412 fail:
1413 	/* Tear down the receive queue(s). */
1414 	while (--index >= 0)
1415 		sfxge_rx_qfini(sc, index);
1416 
1417 	sc->rxq_count = 0;
1418 
1419 #ifdef SFXGE_LRO
1420 fail_lro_table_size:
1421 #endif
1422 	return (rc);
1423 }
1424