xref: /freebsd/sys/dev/sfxge/sfxge_tx.c (revision 98e0ffaefb0f241cda3a72395d3be04192ae0d47)
1 /*-
2  * Copyright (c) 2010-2011 Solarflare Communications, Inc.
3  * All rights reserved.
4  *
5  * This software was developed in part by Philip Paeps under contract for
6  * Solarflare Communications, Inc.
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 /* Theory of operation:
31  *
32  * Tx queues allocation and mapping
33  *
34  * One Tx queue with enabled checksum offload is allocated per Rx channel
35  * (event queue).  Also 2 Tx queues (one without checksum offload and one
36  * with IP checksum offload only) are allocated and bound to event queue 0.
37  * sfxge_txq_type is used as Tx queue label.
38  *
39  * So, event queue plus label mapping to Tx queue index is:
40  *	if event queue index is 0, TxQ-index = TxQ-label * [0..SFXGE_TXQ_NTYPES)
41  *	else TxQ-index = SFXGE_TXQ_NTYPES + EvQ-index - 1
42  * See sfxge_get_txq_by_label() sfxge_ev.c
43  */
44 
45 #include <sys/cdefs.h>
46 __FBSDID("$FreeBSD$");
47 
48 #include <sys/types.h>
49 #include <sys/mbuf.h>
50 #include <sys/smp.h>
51 #include <sys/socket.h>
52 #include <sys/sysctl.h>
53 #include <sys/syslog.h>
54 
55 #include <net/bpf.h>
56 #include <net/ethernet.h>
57 #include <net/if.h>
58 #include <net/if_vlan_var.h>
59 
60 #include <netinet/in.h>
61 #include <netinet/ip.h>
62 #include <netinet/ip6.h>
63 #include <netinet/tcp.h>
64 
65 #include "common/efx.h"
66 
67 #include "sfxge.h"
68 #include "sfxge_tx.h"
69 
70 /*
71  * Estimate maximum number of Tx descriptors required for TSO packet.
72  * With minimum MSS and maximum mbuf length we might need more (even
73  * than a ring-ful of descriptors), but this should not happen in
74  * practice except due to deliberate attack.  In that case we will
75  * truncate the output at a packet boundary.
76  */
77 #define	SFXGE_TSO_MAX_DESC						\
78 	(SFXGE_TSO_MAX_SEGS * 2 + SFXGE_TX_MAPPING_MAX_SEG - 1)
79 
80 /*
81  * Set the block level to ensure there is space to generate a
82  * large number of descriptors for TSO.
83  */
84 #define	SFXGE_TXQ_BLOCK_LEVEL(_entries)					\
85 	(EFX_TXQ_LIMIT(_entries) - SFXGE_TSO_MAX_DESC)
86 
87 
88 #define	SFXGE_PARAM_TX_DPL_GET_MAX	SFXGE_PARAM(tx_dpl_get_max)
89 static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT;
90 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max);
91 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN,
92 	   &sfxge_tx_dpl_get_max, 0,
93 	   "Maximum number of any packets in deferred packet get-list");
94 
95 #define	SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \
96 	SFXGE_PARAM(tx_dpl_get_non_tcp_max)
97 static int sfxge_tx_dpl_get_non_tcp_max =
98 	SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT;
99 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max);
100 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN,
101 	   &sfxge_tx_dpl_get_non_tcp_max, 0,
102 	   "Maximum number of non-TCP packets in deferred packet get-list");
103 
104 #define	SFXGE_PARAM_TX_DPL_PUT_MAX	SFXGE_PARAM(tx_dpl_put_max)
105 static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT;
106 TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max);
107 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN,
108 	   &sfxge_tx_dpl_put_max, 0,
109 	   "Maximum number of any packets in deferred packet put-list");
110 
111 
112 static const struct {
113 	const char *name;
114 	size_t offset;
115 } sfxge_tx_stats[] = {
116 #define	SFXGE_TX_STAT(name, member) \
117 	{ #name, offsetof(struct sfxge_txq, member) }
118 	SFXGE_TX_STAT(tso_bursts, tso_bursts),
119 	SFXGE_TX_STAT(tso_packets, tso_packets),
120 	SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
121 	SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many),
122 	SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc),
123 	SFXGE_TX_STAT(tx_collapses, collapses),
124 	SFXGE_TX_STAT(tx_drops, drops),
125 	SFXGE_TX_STAT(tx_get_overflow, get_overflow),
126 	SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow),
127 	SFXGE_TX_STAT(tx_put_overflow, put_overflow),
128 	SFXGE_TX_STAT(tx_netdown_drops, netdown_drops),
129 };
130 
131 
132 /* Forward declarations. */
133 static void sfxge_tx_qdpl_service(struct sfxge_txq *txq);
134 static void sfxge_tx_qlist_post(struct sfxge_txq *txq);
135 static void sfxge_tx_qunblock(struct sfxge_txq *txq);
136 static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
137 			      const bus_dma_segment_t *dma_seg, int n_dma_seg);
138 
139 void
140 sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq)
141 {
142 	unsigned int completed;
143 
144 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
145 
146 	completed = txq->completed;
147 	while (completed != txq->pending) {
148 		struct sfxge_tx_mapping *stmp;
149 		unsigned int id;
150 
151 		id = completed++ & txq->ptr_mask;
152 
153 		stmp = &txq->stmp[id];
154 		if (stmp->flags & TX_BUF_UNMAP) {
155 			bus_dmamap_unload(txq->packet_dma_tag, stmp->map);
156 			if (stmp->flags & TX_BUF_MBUF) {
157 				struct mbuf *m = stmp->u.mbuf;
158 				do
159 					m = m_free(m);
160 				while (m != NULL);
161 			} else {
162 				free(stmp->u.heap_buf, M_SFXGE);
163 			}
164 			stmp->flags = 0;
165 		}
166 	}
167 	txq->completed = completed;
168 
169 	/* Check whether we need to unblock the queue. */
170 	mb();
171 	if (txq->blocked) {
172 		unsigned int level;
173 
174 		level = txq->added - txq->completed;
175 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
176 			sfxge_tx_qunblock(txq);
177 	}
178 }
179 
180 static unsigned int
181 sfxge_is_mbuf_non_tcp(struct mbuf *mbuf)
182 {
183 	/* Absense of TCP checksum flags does not mean that it is non-TCP
184 	 * but it should be true if user wants to achieve high throughput.
185 	 */
186 	return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP)));
187 }
188 
189 /*
190  * Reorder the put list and append it to the get list.
191  */
192 static void
193 sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq)
194 {
195 	struct sfxge_tx_dpl *stdp;
196 	struct mbuf *mbuf, *get_next, **get_tailp;
197 	volatile uintptr_t *putp;
198 	uintptr_t put;
199 	unsigned int count;
200 	unsigned int non_tcp_count;
201 
202 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
203 
204 	stdp = &txq->dpl;
205 
206 	/* Acquire the put list. */
207 	putp = &stdp->std_put;
208 	put = atomic_readandclear_ptr(putp);
209 	mbuf = (void *)put;
210 
211 	if (mbuf == NULL)
212 		return;
213 
214 	/* Reverse the put list. */
215 	get_tailp = &mbuf->m_nextpkt;
216 	get_next = NULL;
217 
218 	count = 0;
219 	non_tcp_count = 0;
220 	do {
221 		struct mbuf *put_next;
222 
223 		non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf);
224 		put_next = mbuf->m_nextpkt;
225 		mbuf->m_nextpkt = get_next;
226 		get_next = mbuf;
227 		mbuf = put_next;
228 
229 		count++;
230 	} while (mbuf != NULL);
231 
232 	if (count > stdp->std_put_hiwat)
233 		stdp->std_put_hiwat = count;
234 
235 	/* Append the reversed put list to the get list. */
236 	KASSERT(*get_tailp == NULL, ("*get_tailp != NULL"));
237 	*stdp->std_getp = get_next;
238 	stdp->std_getp = get_tailp;
239 	stdp->std_get_count += count;
240 	stdp->std_get_non_tcp_count += non_tcp_count;
241 }
242 
243 static void
244 sfxge_tx_qreap(struct sfxge_txq *txq)
245 {
246 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
247 
248 	txq->reaped = txq->completed;
249 }
250 
251 static void
252 sfxge_tx_qlist_post(struct sfxge_txq *txq)
253 {
254 	unsigned int old_added;
255 	unsigned int level;
256 	int rc;
257 
258 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
259 
260 	KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0"));
261 	KASSERT(txq->n_pend_desc <= SFXGE_TSO_MAX_DESC,
262 		("txq->n_pend_desc too large"));
263 	KASSERT(!txq->blocked, ("txq->blocked"));
264 
265 	old_added = txq->added;
266 
267 	/* Post the fragment list. */
268 	rc = efx_tx_qpost(txq->common, txq->pend_desc, txq->n_pend_desc,
269 			  txq->reaped, &txq->added);
270 	KASSERT(rc == 0, ("efx_tx_qpost() failed"));
271 
272 	/* If efx_tx_qpost() had to refragment, our information about
273 	 * buffers to free may be associated with the wrong
274 	 * descriptors.
275 	 */
276 	KASSERT(txq->added - old_added == txq->n_pend_desc,
277 		("efx_tx_qpost() refragmented descriptors"));
278 
279 	level = txq->added - txq->reaped;
280 	KASSERT(level <= txq->entries, ("overfilled TX queue"));
281 
282 	/* Clear the fragment list. */
283 	txq->n_pend_desc = 0;
284 
285 	/* Have we reached the block level? */
286 	if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries))
287 		return;
288 
289 	/* Reap, and check again */
290 	sfxge_tx_qreap(txq);
291 	level = txq->added - txq->reaped;
292 	if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries))
293 		return;
294 
295 	txq->blocked = 1;
296 
297 	/*
298 	 * Avoid a race with completion interrupt handling that could leave
299 	 * the queue blocked.
300 	 */
301 	mb();
302 	sfxge_tx_qreap(txq);
303 	level = txq->added - txq->reaped;
304 	if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries)) {
305 		mb();
306 		txq->blocked = 0;
307 	}
308 }
309 
310 static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
311 {
312 	bus_dmamap_t *used_map;
313 	bus_dmamap_t map;
314 	bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG];
315 	unsigned int id;
316 	struct sfxge_tx_mapping *stmp;
317 	efx_buffer_t *desc;
318 	int n_dma_seg;
319 	int rc;
320 	int i;
321 
322 	KASSERT(!txq->blocked, ("txq->blocked"));
323 
324 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
325 		prefetch_read_many(mbuf->m_data);
326 
327 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) {
328 		rc = EINTR;
329 		goto reject;
330 	}
331 
332 	/* Load the packet for DMA. */
333 	id = txq->added & txq->ptr_mask;
334 	stmp = &txq->stmp[id];
335 	rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map,
336 				     mbuf, dma_seg, &n_dma_seg, 0);
337 	if (rc == EFBIG) {
338 		/* Try again. */
339 		struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT,
340 						   SFXGE_TX_MAPPING_MAX_SEG);
341 		if (new_mbuf == NULL)
342 			goto reject;
343 		++txq->collapses;
344 		mbuf = new_mbuf;
345 		rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag,
346 					     stmp->map, mbuf,
347 					     dma_seg, &n_dma_seg, 0);
348 	}
349 	if (rc != 0)
350 		goto reject;
351 
352 	/* Make the packet visible to the hardware. */
353 	bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE);
354 
355 	used_map = &stmp->map;
356 
357 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
358 		rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg);
359 		if (rc < 0)
360 			goto reject_mapped;
361 		stmp = &txq->stmp[rc];
362 	} else {
363 		/* Add the mapping to the fragment list, and set flags
364 		 * for the buffer.
365 		 */
366 		i = 0;
367 		for (;;) {
368 			desc = &txq->pend_desc[i];
369 			desc->eb_addr = dma_seg[i].ds_addr;
370 			desc->eb_size = dma_seg[i].ds_len;
371 			if (i == n_dma_seg - 1) {
372 				desc->eb_eop = 1;
373 				break;
374 			}
375 			desc->eb_eop = 0;
376 			i++;
377 
378 			stmp->flags = 0;
379 			if (__predict_false(stmp ==
380 					    &txq->stmp[txq->ptr_mask]))
381 				stmp = &txq->stmp[0];
382 			else
383 				stmp++;
384 		}
385 		txq->n_pend_desc = n_dma_seg;
386 	}
387 
388 	/*
389 	 * If the mapping required more than one descriptor
390 	 * then we need to associate the DMA map with the last
391 	 * descriptor, not the first.
392 	 */
393 	if (used_map != &stmp->map) {
394 		map = stmp->map;
395 		stmp->map = *used_map;
396 		*used_map = map;
397 	}
398 
399 	stmp->u.mbuf = mbuf;
400 	stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF;
401 
402 	/* Post the fragment list. */
403 	sfxge_tx_qlist_post(txq);
404 
405 	return (0);
406 
407 reject_mapped:
408 	bus_dmamap_unload(txq->packet_dma_tag, *used_map);
409 reject:
410 	/* Drop the packet on the floor. */
411 	m_freem(mbuf);
412 	++txq->drops;
413 
414 	return (rc);
415 }
416 
417 /*
418  * Drain the deferred packet list into the transmit queue.
419  */
420 static void
421 sfxge_tx_qdpl_drain(struct sfxge_txq *txq)
422 {
423 	struct sfxge_softc *sc;
424 	struct sfxge_tx_dpl *stdp;
425 	struct mbuf *mbuf, *next;
426 	unsigned int count;
427 	unsigned int non_tcp_count;
428 	unsigned int pushed;
429 	int rc;
430 
431 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
432 
433 	sc = txq->sc;
434 	stdp = &txq->dpl;
435 	pushed = txq->added;
436 
437 	if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) {
438 		prefetch_read_many(sc->enp);
439 		prefetch_read_many(txq->common);
440 	}
441 
442 	mbuf = stdp->std_get;
443 	count = stdp->std_get_count;
444 	non_tcp_count = stdp->std_get_non_tcp_count;
445 
446 	if (count > stdp->std_get_hiwat)
447 		stdp->std_get_hiwat = count;
448 
449 	while (count != 0) {
450 		KASSERT(mbuf != NULL, ("mbuf == NULL"));
451 
452 		next = mbuf->m_nextpkt;
453 		mbuf->m_nextpkt = NULL;
454 
455 		ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */
456 
457 		if (next != NULL)
458 			prefetch_read_many(next);
459 
460 		rc = sfxge_tx_queue_mbuf(txq, mbuf);
461 		--count;
462 		non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf);
463 		mbuf = next;
464 		if (rc != 0)
465 			continue;
466 
467 		if (txq->blocked)
468 			break;
469 
470 		/* Push the fragments to the hardware in batches. */
471 		if (txq->added - pushed >= SFXGE_TX_BATCH) {
472 			efx_tx_qpush(txq->common, txq->added);
473 			pushed = txq->added;
474 		}
475 	}
476 
477 	if (count == 0) {
478 		KASSERT(mbuf == NULL, ("mbuf != NULL"));
479 		KASSERT(non_tcp_count == 0,
480 			("inconsistent TCP/non-TCP detection"));
481 		stdp->std_get = NULL;
482 		stdp->std_get_count = 0;
483 		stdp->std_get_non_tcp_count = 0;
484 		stdp->std_getp = &stdp->std_get;
485 	} else {
486 		stdp->std_get = mbuf;
487 		stdp->std_get_count = count;
488 		stdp->std_get_non_tcp_count = non_tcp_count;
489 	}
490 
491 	if (txq->added != pushed)
492 		efx_tx_qpush(txq->common, txq->added);
493 
494 	KASSERT(txq->blocked || stdp->std_get_count == 0,
495 		("queue unblocked but count is non-zero"));
496 }
497 
498 #define	SFXGE_TX_QDPL_PENDING(_txq)					\
499 	((_txq)->dpl.std_put != 0)
500 
501 /*
502  * Service the deferred packet list.
503  *
504  * NOTE: drops the txq mutex!
505  */
506 static void
507 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
508 {
509 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
510 
511 	do {
512 		if (SFXGE_TX_QDPL_PENDING(txq))
513 			sfxge_tx_qdpl_swizzle(txq);
514 
515 		if (!txq->blocked)
516 			sfxge_tx_qdpl_drain(txq);
517 
518 		SFXGE_TXQ_UNLOCK(txq);
519 	} while (SFXGE_TX_QDPL_PENDING(txq) &&
520 		 SFXGE_TXQ_TRYLOCK(txq));
521 }
522 
523 /*
524  * Put a packet on the deferred packet get-list.
525  */
526 static int
527 sfxge_tx_qdpl_put_locked(struct sfxge_txq *txq, struct mbuf *mbuf)
528 {
529 	struct sfxge_tx_dpl *stdp;
530 
531 	stdp = &txq->dpl;
532 
533 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
534 
535 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
536 
537 	if (stdp->std_get_count >= stdp->std_get_max) {
538 		txq->get_overflow++;
539 		return (ENOBUFS);
540 	}
541 	if (sfxge_is_mbuf_non_tcp(mbuf)) {
542 		if (stdp->std_get_non_tcp_count >=
543 		    stdp->std_get_non_tcp_max) {
544 			txq->get_non_tcp_overflow++;
545 			return (ENOBUFS);
546 		}
547 		stdp->std_get_non_tcp_count++;
548 	}
549 
550 	*(stdp->std_getp) = mbuf;
551 	stdp->std_getp = &mbuf->m_nextpkt;
552 	stdp->std_get_count++;
553 
554 	return (0);
555 }
556 
557 /*
558  * Put a packet on the deferred packet put-list.
559  *
560  * We overload the csum_data field in the mbuf to keep track of this length
561  * because there is no cheap alternative to avoid races.
562  */
563 static int
564 sfxge_tx_qdpl_put_unlocked(struct sfxge_txq *txq, struct mbuf *mbuf)
565 {
566 	struct sfxge_tx_dpl *stdp;
567 	volatile uintptr_t *putp;
568 	uintptr_t old;
569 	uintptr_t new;
570 	unsigned old_len;
571 
572 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
573 
574 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
575 
576 	stdp = &txq->dpl;
577 	putp = &stdp->std_put;
578 	new = (uintptr_t)mbuf;
579 
580 	do {
581 		old = *putp;
582 		if (old != 0) {
583 			struct mbuf *mp = (struct mbuf *)old;
584 			old_len = mp->m_pkthdr.csum_data;
585 		} else
586 			old_len = 0;
587 		if (old_len >= stdp->std_put_max) {
588 			atomic_add_long(&txq->put_overflow, 1);
589 			return (ENOBUFS);
590 		}
591 		mbuf->m_pkthdr.csum_data = old_len + 1;
592 		mbuf->m_nextpkt = (void *)old;
593 	} while (atomic_cmpset_ptr(putp, old, new) == 0);
594 
595 	return (0);
596 }
597 
598 /*
599  * Called from if_transmit - will try to grab the txq lock and enqueue to the
600  * put list if it succeeds, otherwise try to push onto the defer list if space.
601  */
602 int
603 sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m)
604 {
605 	int rc;
606 
607 	if (!SFXGE_LINK_UP(txq->sc)) {
608 		atomic_add_long(&txq->netdown_drops, 1);
609 		return (ENETDOWN);
610 	}
611 
612 	/*
613 	 * Try to grab the txq lock.  If we are able to get the lock,
614 	 * the packet will be appended to the "get list" of the deferred
615 	 * packet list.  Otherwise, it will be pushed on the "put list".
616 	 */
617 	if (SFXGE_TXQ_TRYLOCK(txq)) {
618 		/* First swizzle put-list to get-list to keep order */
619 		sfxge_tx_qdpl_swizzle(txq);
620 
621 		rc = sfxge_tx_qdpl_put_locked(txq, m);
622 
623 		/* Try to service the list. */
624 		sfxge_tx_qdpl_service(txq);
625 		/* Lock has been dropped. */
626 	} else {
627 		rc = sfxge_tx_qdpl_put_unlocked(txq, m);
628 
629 		/*
630 		 * Try to grab the lock again.
631 		 *
632 		 * If we are able to get the lock, we need to process
633 		 * the deferred packet list.  If we are not able to get
634 		 * the lock, another thread is processing the list.
635 		 */
636 		if ((rc == 0) && SFXGE_TXQ_TRYLOCK(txq)) {
637 			sfxge_tx_qdpl_service(txq);
638 			/* Lock has been dropped. */
639 		}
640 	}
641 
642 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
643 
644 	return (rc);
645 }
646 
647 static void
648 sfxge_tx_qdpl_flush(struct sfxge_txq *txq)
649 {
650 	struct sfxge_tx_dpl *stdp = &txq->dpl;
651 	struct mbuf *mbuf, *next;
652 
653 	SFXGE_TXQ_LOCK(txq);
654 
655 	sfxge_tx_qdpl_swizzle(txq);
656 	for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) {
657 		next = mbuf->m_nextpkt;
658 		m_freem(mbuf);
659 	}
660 	stdp->std_get = NULL;
661 	stdp->std_get_count = 0;
662 	stdp->std_get_non_tcp_count = 0;
663 	stdp->std_getp = &stdp->std_get;
664 
665 	SFXGE_TXQ_UNLOCK(txq);
666 }
667 
668 void
669 sfxge_if_qflush(struct ifnet *ifp)
670 {
671 	struct sfxge_softc *sc;
672 	unsigned int i;
673 
674 	sc = ifp->if_softc;
675 
676 	for (i = 0; i < sc->txq_count; i++)
677 		sfxge_tx_qdpl_flush(sc->txq[i]);
678 }
679 
680 /*
681  * TX start -- called by the stack.
682  */
683 int
684 sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m)
685 {
686 	struct sfxge_softc *sc;
687 	struct sfxge_txq *txq;
688 	int rc;
689 
690 	sc = (struct sfxge_softc *)ifp->if_softc;
691 
692 	/*
693 	 * Transmit may be called when interface is up from the kernel
694 	 * point of view, but not yet up (in progress) from the driver
695 	 * point of view. I.e. link aggregation bring up.
696 	 * Transmit may be called when interface is up from the driver
697 	 * point of view, but already down from the kernel point of
698 	 * view. I.e. Rx when interface shutdown is in progress.
699 	 */
700 	KASSERT((ifp->if_flags & IFF_UP) || (sc->if_flags & IFF_UP),
701 		("interface not up"));
702 
703 	/* Pick the desired transmit queue. */
704 	if (m->m_pkthdr.csum_flags &
705 	    (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO)) {
706 		int index = 0;
707 
708 		/* check if flowid is set */
709 		if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
710 			uint32_t hash = m->m_pkthdr.flowid;
711 
712 			index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX];
713 		}
714 		txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index];
715 	} else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
716 		txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
717 	} else {
718 		txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
719 	}
720 
721 	rc = sfxge_tx_packet_add(txq, m);
722 	if (rc != 0)
723 		m_freem(m);
724 
725 	return (rc);
726 }
727 
728 /*
729  * Software "TSO".  Not quite as good as doing it in hardware, but
730  * still faster than segmenting in the stack.
731  */
732 
733 struct sfxge_tso_state {
734 	/* Output position */
735 	unsigned out_len;	/* Remaining length in current segment */
736 	unsigned seqnum;	/* Current sequence number */
737 	unsigned packet_space;	/* Remaining space in current packet */
738 
739 	/* Input position */
740 	uint64_t dma_addr;	/* DMA address of current position */
741 	unsigned in_len;	/* Remaining length in current mbuf */
742 
743 	const struct mbuf *mbuf; /* Input mbuf (head of chain) */
744 	u_short protocol;	/* Network protocol (after VLAN decap) */
745 	ssize_t nh_off;		/* Offset of network header */
746 	ssize_t tcph_off;	/* Offset of TCP header */
747 	unsigned header_len;	/* Number of bytes of header */
748 	unsigned seg_size;	/* TCP segment size */
749 };
750 
751 static const struct ip *tso_iph(const struct sfxge_tso_state *tso)
752 {
753 	KASSERT(tso->protocol == htons(ETHERTYPE_IP),
754 		("tso_iph() in non-IPv4 state"));
755 	return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
756 }
757 static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
758 {
759 	KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
760 		("tso_ip6h() in non-IPv6 state"));
761 	return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
762 }
763 static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
764 {
765 	return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
766 }
767 
768 /* Size of preallocated TSO header buffers.  Larger blocks must be
769  * allocated from the heap.
770  */
771 #define	TSOH_STD_SIZE	128
772 
773 /* At most half the descriptors in the queue at any time will refer to
774  * a TSO header buffer, since they must always be followed by a
775  * payload descriptor referring to an mbuf.
776  */
777 #define	TSOH_COUNT(_txq_entries)	((_txq_entries) / 2u)
778 #define	TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
779 #define	TSOH_PAGE_COUNT(_txq_entries)	\
780 	((TSOH_COUNT(_txq_entries) + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE)
781 
782 static int tso_init(struct sfxge_txq *txq)
783 {
784 	struct sfxge_softc *sc = txq->sc;
785 	unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries);
786 	int i, rc;
787 
788 	/* Allocate TSO header buffers */
789 	txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]),
790 				  M_SFXGE, M_WAITOK);
791 
792 	for (i = 0; i < tsoh_page_count; i++) {
793 		rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
794 		if (rc != 0)
795 			goto fail;
796 	}
797 
798 	return (0);
799 
800 fail:
801 	while (i-- > 0)
802 		sfxge_dma_free(&txq->tsoh_buffer[i]);
803 	free(txq->tsoh_buffer, M_SFXGE);
804 	txq->tsoh_buffer = NULL;
805 	return (rc);
806 }
807 
808 static void tso_fini(struct sfxge_txq *txq)
809 {
810 	int i;
811 
812 	if (txq->tsoh_buffer != NULL) {
813 		for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++)
814 			sfxge_dma_free(&txq->tsoh_buffer[i]);
815 		free(txq->tsoh_buffer, M_SFXGE);
816 	}
817 }
818 
819 static void tso_start(struct sfxge_tso_state *tso, struct mbuf *mbuf)
820 {
821 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
822 	const struct tcphdr *th;
823 	struct tcphdr th_copy;
824 
825 	tso->mbuf = mbuf;
826 
827 	/* Find network protocol and header */
828 	tso->protocol = eh->ether_type;
829 	if (tso->protocol == htons(ETHERTYPE_VLAN)) {
830 		struct ether_vlan_header *veh =
831 			mtod(mbuf, struct ether_vlan_header *);
832 		tso->protocol = veh->evl_proto;
833 		tso->nh_off = sizeof(*veh);
834 	} else {
835 		tso->nh_off = sizeof(*eh);
836 	}
837 
838 	/* Find TCP header */
839 	if (tso->protocol == htons(ETHERTYPE_IP)) {
840 		KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
841 			("TSO required on non-TCP packet"));
842 		tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
843 	} else {
844 		KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
845 			("TSO required on non-IP packet"));
846 		KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
847 			("TSO required on non-TCP packet"));
848 		tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
849 	}
850 
851 	KASSERT(mbuf->m_len >= tso->tcph_off,
852 		("network header is fragmented in mbuf"));
853 	/* We need TCP header including flags (window is the next) */
854 	if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) {
855 		m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy),
856 			   (caddr_t)&th_copy);
857 		th = &th_copy;
858 	} else {
859 		th = tso_tcph(tso);
860 	}
861 
862 	tso->header_len = tso->tcph_off + 4 * th->th_off;
863 	tso->seg_size = mbuf->m_pkthdr.tso_segsz;
864 
865 	tso->seqnum = ntohl(th->th_seq);
866 
867 	/* These flags must not be duplicated */
868 	/*
869 	 * RST should not be duplicated as well, but FreeBSD kernel
870 	 * generates TSO packets with RST flag. So, do not assert
871 	 * its absence.
872 	 */
873 	KASSERT(!(th->th_flags & (TH_URG | TH_SYN)),
874 		("incompatible TCP flag 0x%x on TSO packet",
875 		 th->th_flags & (TH_URG | TH_SYN)));
876 
877 	tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
878 }
879 
880 /*
881  * tso_fill_packet_with_fragment - form descriptors for the current fragment
882  *
883  * Form descriptors for the current fragment, until we reach the end
884  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
885  * space.
886  */
887 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
888 					  struct sfxge_tso_state *tso)
889 {
890 	efx_buffer_t *desc;
891 	int n;
892 
893 	if (tso->in_len == 0 || tso->packet_space == 0)
894 		return;
895 
896 	KASSERT(tso->in_len > 0, ("TSO input length went negative"));
897 	KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
898 
899 	n = min(tso->in_len, tso->packet_space);
900 
901 	tso->packet_space -= n;
902 	tso->out_len -= n;
903 	tso->in_len -= n;
904 
905 	desc = &txq->pend_desc[txq->n_pend_desc++];
906 	desc->eb_addr = tso->dma_addr;
907 	desc->eb_size = n;
908 	desc->eb_eop = tso->out_len == 0 || tso->packet_space == 0;
909 
910 	tso->dma_addr += n;
911 }
912 
913 /* Callback from bus_dmamap_load() for long TSO headers. */
914 static void tso_map_long_header(void *dma_addr_ret,
915 				bus_dma_segment_t *segs, int nseg,
916 				int error)
917 {
918 	*(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
919 				      __predict_true(nseg == 1)) ?
920 				     segs->ds_addr : 0);
921 }
922 
923 /*
924  * tso_start_new_packet - generate a new header and prepare for the new packet
925  *
926  * Generate a new header and prepare for the new packet.  Return 0 on
927  * success, or an error code if failed to alloc header.
928  */
929 static int tso_start_new_packet(struct sfxge_txq *txq,
930 				struct sfxge_tso_state *tso,
931 				unsigned int id)
932 {
933 	struct sfxge_tx_mapping *stmp = &txq->stmp[id];
934 	struct tcphdr *tsoh_th;
935 	unsigned ip_length;
936 	caddr_t header;
937 	uint64_t dma_addr;
938 	bus_dmamap_t map;
939 	efx_buffer_t *desc;
940 	int rc;
941 
942 	/* Allocate a DMA-mapped header buffer. */
943 	if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
944 		unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
945 		unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
946 
947 		header = (txq->tsoh_buffer[page_index].esm_base +
948 			  buf_index * TSOH_STD_SIZE);
949 		dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
950 			    buf_index * TSOH_STD_SIZE);
951 		map = txq->tsoh_buffer[page_index].esm_map;
952 
953 		stmp->flags = 0;
954 	} else {
955 		/* We cannot use bus_dmamem_alloc() as that may sleep */
956 		header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
957 		if (__predict_false(!header))
958 			return (ENOMEM);
959 		rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
960 				     header, tso->header_len,
961 				     tso_map_long_header, &dma_addr,
962 				     BUS_DMA_NOWAIT);
963 		if (__predict_false(dma_addr == 0)) {
964 			if (rc == 0) {
965 				/* Succeeded but got >1 segment */
966 				bus_dmamap_unload(txq->packet_dma_tag,
967 						  stmp->map);
968 				rc = EINVAL;
969 			}
970 			free(header, M_SFXGE);
971 			return (rc);
972 		}
973 		map = stmp->map;
974 
975 		txq->tso_long_headers++;
976 		stmp->u.heap_buf = header;
977 		stmp->flags = TX_BUF_UNMAP;
978 	}
979 
980 	tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
981 
982 	/* Copy and update the headers. */
983 	m_copydata(tso->mbuf, 0, tso->header_len, header);
984 
985 	tsoh_th->th_seq = htonl(tso->seqnum);
986 	tso->seqnum += tso->seg_size;
987 	if (tso->out_len > tso->seg_size) {
988 		/* This packet will not finish the TSO burst. */
989 		ip_length = tso->header_len - tso->nh_off + tso->seg_size;
990 		tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH);
991 	} else {
992 		/* This packet will be the last in the TSO burst. */
993 		ip_length = tso->header_len - tso->nh_off + tso->out_len;
994 	}
995 
996 	if (tso->protocol == htons(ETHERTYPE_IP)) {
997 		struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
998 		tsoh_iph->ip_len = htons(ip_length);
999 		/* XXX We should increment ip_id, but FreeBSD doesn't
1000 		 * currently allocate extra IDs for multiple segments.
1001 		 */
1002 	} else {
1003 		struct ip6_hdr *tsoh_iph =
1004 			(struct ip6_hdr *)(header + tso->nh_off);
1005 		tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
1006 	}
1007 
1008 	/* Make the header visible to the hardware. */
1009 	bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
1010 
1011 	tso->packet_space = tso->seg_size;
1012 	txq->tso_packets++;
1013 
1014 	/* Form a descriptor for this header. */
1015 	desc = &txq->pend_desc[txq->n_pend_desc++];
1016 	desc->eb_addr = dma_addr;
1017 	desc->eb_size = tso->header_len;
1018 	desc->eb_eop = 0;
1019 
1020 	return (0);
1021 }
1022 
1023 static int
1024 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
1025 		   const bus_dma_segment_t *dma_seg, int n_dma_seg)
1026 {
1027 	struct sfxge_tso_state tso;
1028 	unsigned int id, next_id;
1029 	unsigned skipped = 0;
1030 
1031 	tso_start(&tso, mbuf);
1032 
1033 	while (dma_seg->ds_len + skipped <= tso.header_len) {
1034 		skipped += dma_seg->ds_len;
1035 		--n_dma_seg;
1036 		KASSERT(n_dma_seg, ("no payload found in TSO packet"));
1037 		++dma_seg;
1038 	}
1039 	tso.in_len = dma_seg->ds_len - (tso.header_len - skipped);
1040 	tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped);
1041 
1042 	id = txq->added & txq->ptr_mask;
1043 	if (__predict_false(tso_start_new_packet(txq, &tso, id)))
1044 		return (-1);
1045 
1046 	while (1) {
1047 		id = (id + 1) & txq->ptr_mask;
1048 		tso_fill_packet_with_fragment(txq, &tso);
1049 
1050 		/* Move onto the next fragment? */
1051 		if (tso.in_len == 0) {
1052 			--n_dma_seg;
1053 			if (n_dma_seg == 0)
1054 				break;
1055 			++dma_seg;
1056 			tso.in_len = dma_seg->ds_len;
1057 			tso.dma_addr = dma_seg->ds_addr;
1058 		}
1059 
1060 		/* End of packet? */
1061 		if (tso.packet_space == 0) {
1062 			/* If the queue is now full due to tiny MSS,
1063 			 * or we can't create another header, discard
1064 			 * the remainder of the input mbuf but do not
1065 			 * roll back the work we have done.
1066 			 */
1067 			if (txq->n_pend_desc + 1 /* header */ + n_dma_seg >
1068 			    SFXGE_TSO_MAX_DESC) {
1069 				txq->tso_pdrop_too_many++;
1070 				break;
1071 			}
1072 			next_id = (id + 1) & txq->ptr_mask;
1073 			if (__predict_false(tso_start_new_packet(txq, &tso,
1074 								 next_id))) {
1075 				txq->tso_pdrop_no_rsrc++;
1076 				break;
1077 			}
1078 			id = next_id;
1079 		}
1080 	}
1081 
1082 	txq->tso_bursts++;
1083 	return (id);
1084 }
1085 
1086 static void
1087 sfxge_tx_qunblock(struct sfxge_txq *txq)
1088 {
1089 	struct sfxge_softc *sc;
1090 	struct sfxge_evq *evq;
1091 
1092 	sc = txq->sc;
1093 	evq = sc->evq[txq->evq_index];
1094 
1095 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
1096 
1097 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED))
1098 		return;
1099 
1100 	SFXGE_TXQ_LOCK(txq);
1101 
1102 	if (txq->blocked) {
1103 		unsigned int level;
1104 
1105 		level = txq->added - txq->completed;
1106 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) {
1107 			/* reaped must be in sync with blocked */
1108 			sfxge_tx_qreap(txq);
1109 			txq->blocked = 0;
1110 		}
1111 	}
1112 
1113 	sfxge_tx_qdpl_service(txq);
1114 	/* note: lock has been dropped */
1115 }
1116 
1117 void
1118 sfxge_tx_qflush_done(struct sfxge_txq *txq)
1119 {
1120 
1121 	txq->flush_state = SFXGE_FLUSH_DONE;
1122 }
1123 
1124 static void
1125 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1126 {
1127 	struct sfxge_txq *txq;
1128 	struct sfxge_evq *evq;
1129 	unsigned int count;
1130 
1131 	txq = sc->txq[index];
1132 	evq = sc->evq[txq->evq_index];
1133 
1134 	SFXGE_TXQ_LOCK(txq);
1135 
1136 	KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1137 	    ("txq->init_state != SFXGE_TXQ_STARTED"));
1138 
1139 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1140 	txq->flush_state = SFXGE_FLUSH_PENDING;
1141 
1142 	/* Flush the transmit queue. */
1143 	efx_tx_qflush(txq->common);
1144 
1145 	SFXGE_TXQ_UNLOCK(txq);
1146 
1147 	count = 0;
1148 	do {
1149 		/* Spin for 100ms. */
1150 		DELAY(100000);
1151 
1152 		if (txq->flush_state != SFXGE_FLUSH_PENDING)
1153 			break;
1154 	} while (++count < 20);
1155 
1156 	SFXGE_EVQ_LOCK(evq);
1157 	SFXGE_TXQ_LOCK(txq);
1158 
1159 	KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1160 	    ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1161 
1162 	txq->flush_state = SFXGE_FLUSH_DONE;
1163 
1164 	txq->blocked = 0;
1165 	txq->pending = txq->added;
1166 
1167 	sfxge_tx_qcomplete(txq, evq);
1168 	KASSERT(txq->completed == txq->added,
1169 	    ("txq->completed != txq->added"));
1170 
1171 	sfxge_tx_qreap(txq);
1172 	KASSERT(txq->reaped == txq->completed,
1173 	    ("txq->reaped != txq->completed"));
1174 
1175 	txq->added = 0;
1176 	txq->pending = 0;
1177 	txq->completed = 0;
1178 	txq->reaped = 0;
1179 
1180 	/* Destroy the common code transmit queue. */
1181 	efx_tx_qdestroy(txq->common);
1182 	txq->common = NULL;
1183 
1184 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1185 	    EFX_TXQ_NBUFS(sc->txq_entries));
1186 
1187 	SFXGE_EVQ_UNLOCK(evq);
1188 	SFXGE_TXQ_UNLOCK(txq);
1189 }
1190 
1191 static int
1192 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1193 {
1194 	struct sfxge_txq *txq;
1195 	efsys_mem_t *esmp;
1196 	uint16_t flags;
1197 	struct sfxge_evq *evq;
1198 	int rc;
1199 
1200 	txq = sc->txq[index];
1201 	esmp = &txq->mem;
1202 	evq = sc->evq[txq->evq_index];
1203 
1204 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1205 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1206 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1207 	    ("evq->init_state != SFXGE_EVQ_STARTED"));
1208 
1209 	/* Program the buffer table. */
1210 	if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1211 	    EFX_TXQ_NBUFS(sc->txq_entries))) != 0)
1212 		return (rc);
1213 
1214 	/* Determine the kind of queue we are creating. */
1215 	switch (txq->type) {
1216 	case SFXGE_TXQ_NON_CKSUM:
1217 		flags = 0;
1218 		break;
1219 	case SFXGE_TXQ_IP_CKSUM:
1220 		flags = EFX_CKSUM_IPV4;
1221 		break;
1222 	case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1223 		flags = EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP;
1224 		break;
1225 	default:
1226 		KASSERT(0, ("Impossible TX queue"));
1227 		flags = 0;
1228 		break;
1229 	}
1230 
1231 	/* Create the common code transmit queue. */
1232 	if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1233 	    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1234 	    &txq->common)) != 0)
1235 		goto fail;
1236 
1237 	SFXGE_TXQ_LOCK(txq);
1238 
1239 	/* Enable the transmit queue. */
1240 	efx_tx_qenable(txq->common);
1241 
1242 	txq->init_state = SFXGE_TXQ_STARTED;
1243 
1244 	SFXGE_TXQ_UNLOCK(txq);
1245 
1246 	return (0);
1247 
1248 fail:
1249 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1250 	    EFX_TXQ_NBUFS(sc->txq_entries));
1251 	return (rc);
1252 }
1253 
1254 void
1255 sfxge_tx_stop(struct sfxge_softc *sc)
1256 {
1257 	int index;
1258 
1259 	index = sc->txq_count;
1260 	while (--index >= 0)
1261 		sfxge_tx_qstop(sc, index);
1262 
1263 	/* Tear down the transmit module */
1264 	efx_tx_fini(sc->enp);
1265 }
1266 
1267 int
1268 sfxge_tx_start(struct sfxge_softc *sc)
1269 {
1270 	int index;
1271 	int rc;
1272 
1273 	/* Initialize the common code transmit module. */
1274 	if ((rc = efx_tx_init(sc->enp)) != 0)
1275 		return (rc);
1276 
1277 	for (index = 0; index < sc->txq_count; index++) {
1278 		if ((rc = sfxge_tx_qstart(sc, index)) != 0)
1279 			goto fail;
1280 	}
1281 
1282 	return (0);
1283 
1284 fail:
1285 	while (--index >= 0)
1286 		sfxge_tx_qstop(sc, index);
1287 
1288 	efx_tx_fini(sc->enp);
1289 
1290 	return (rc);
1291 }
1292 
1293 static int
1294 sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node)
1295 {
1296 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev);
1297 	struct sysctl_oid *stat_node;
1298 	unsigned int id;
1299 
1300 	stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1301 				    "stats", CTLFLAG_RD, NULL,
1302 				    "Tx queue statistics");
1303 	if (stat_node == NULL)
1304 		return (ENOMEM);
1305 
1306 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1307 		SYSCTL_ADD_ULONG(
1308 		    ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO,
1309 		    sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS,
1310 		    (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset),
1311 		    "");
1312 	}
1313 
1314 	return (0);
1315 }
1316 
1317 /**
1318  * Destroy a transmit queue.
1319  */
1320 static void
1321 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1322 {
1323 	struct sfxge_txq *txq;
1324 	unsigned int nmaps;
1325 
1326 	txq = sc->txq[index];
1327 
1328 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1329 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1330 
1331 	if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1332 		tso_fini(txq);
1333 
1334 	/* Free the context arrays. */
1335 	free(txq->pend_desc, M_SFXGE);
1336 	nmaps = sc->txq_entries;
1337 	while (nmaps-- != 0)
1338 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1339 	free(txq->stmp, M_SFXGE);
1340 
1341 	/* Release DMA memory mapping. */
1342 	sfxge_dma_free(&txq->mem);
1343 
1344 	sc->txq[index] = NULL;
1345 
1346 	SFXGE_TXQ_LOCK_DESTROY(txq);
1347 
1348 	free(txq, M_SFXGE);
1349 }
1350 
1351 static int
1352 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1353     enum sfxge_txq_type type, unsigned int evq_index)
1354 {
1355 	char name[16];
1356 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1357 	struct sysctl_oid *txq_node;
1358 	struct sfxge_txq *txq;
1359 	struct sfxge_evq *evq;
1360 	struct sfxge_tx_dpl *stdp;
1361 	struct sysctl_oid *dpl_node;
1362 	efsys_mem_t *esmp;
1363 	unsigned int nmaps;
1364 	int rc;
1365 
1366 	txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1367 	txq->sc = sc;
1368 	txq->entries = sc->txq_entries;
1369 	txq->ptr_mask = txq->entries - 1;
1370 
1371 	sc->txq[txq_index] = txq;
1372 	esmp = &txq->mem;
1373 
1374 	evq = sc->evq[evq_index];
1375 
1376 	/* Allocate and zero DMA space for the descriptor ring. */
1377 	if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0)
1378 		return (rc);
1379 
1380 	/* Allocate buffer table entries. */
1381 	sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries),
1382 				 &txq->buf_base_id);
1383 
1384 	/* Create a DMA tag for packet mappings. */
1385 	if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000,
1386 	    MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1387 	    NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL,
1388 	    &txq->packet_dma_tag) != 0) {
1389 		device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1390 		rc = ENOMEM;
1391 		goto fail;
1392 	}
1393 
1394 	/* Allocate pending descriptor array for batching writes. */
1395 	txq->pend_desc = malloc(sizeof(efx_buffer_t) * sc->txq_entries,
1396 				M_SFXGE, M_ZERO | M_WAITOK);
1397 
1398 	/* Allocate and initialise mbuf DMA mapping array. */
1399 	txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries,
1400 	    M_SFXGE, M_ZERO | M_WAITOK);
1401 	for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) {
1402 		rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1403 				       &txq->stmp[nmaps].map);
1404 		if (rc != 0)
1405 			goto fail2;
1406 	}
1407 
1408 	snprintf(name, sizeof(name), "%u", txq_index);
1409 	txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node),
1410 				   OID_AUTO, name, CTLFLAG_RD, NULL, "");
1411 	if (txq_node == NULL) {
1412 		rc = ENOMEM;
1413 		goto fail_txq_node;
1414 	}
1415 
1416 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1417 	    (rc = tso_init(txq)) != 0)
1418 		goto fail3;
1419 
1420 	if (sfxge_tx_dpl_get_max <= 0) {
1421 		log(LOG_ERR, "%s=%d must be greater than 0",
1422 		    SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max);
1423 		rc = EINVAL;
1424 		goto fail_tx_dpl_get_max;
1425 	}
1426 	if (sfxge_tx_dpl_get_non_tcp_max <= 0) {
1427 		log(LOG_ERR, "%s=%d must be greater than 0",
1428 		    SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX,
1429 		    sfxge_tx_dpl_get_non_tcp_max);
1430 		rc = EINVAL;
1431 		goto fail_tx_dpl_get_max;
1432 	}
1433 	if (sfxge_tx_dpl_put_max < 0) {
1434 		log(LOG_ERR, "%s=%d must be greater or equal to 0",
1435 		    SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max);
1436 		rc = EINVAL;
1437 		goto fail_tx_dpl_put_max;
1438 	}
1439 
1440 	/* Initialize the deferred packet list. */
1441 	stdp = &txq->dpl;
1442 	stdp->std_put_max = sfxge_tx_dpl_put_max;
1443 	stdp->std_get_max = sfxge_tx_dpl_get_max;
1444 	stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max;
1445 	stdp->std_getp = &stdp->std_get;
1446 
1447 	SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index);
1448 
1449 	dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1450 				   "dpl", CTLFLAG_RD, NULL,
1451 				   "Deferred packet list statistics");
1452 	if (dpl_node == NULL) {
1453 		rc = ENOMEM;
1454 		goto fail_dpl_node;
1455 	}
1456 
1457 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1458 			"get_count", CTLFLAG_RD | CTLFLAG_STATS,
1459 			&stdp->std_get_count, 0, "");
1460 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1461 			"get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS,
1462 			&stdp->std_get_non_tcp_count, 0, "");
1463 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1464 			"get_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1465 			&stdp->std_get_hiwat, 0, "");
1466 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1467 			"put_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1468 			&stdp->std_put_hiwat, 0, "");
1469 
1470 	rc = sfxge_txq_stat_init(txq, txq_node);
1471 	if (rc != 0)
1472 		goto fail_txq_stat_init;
1473 
1474 	txq->type = type;
1475 	txq->evq_index = evq_index;
1476 	txq->txq_index = txq_index;
1477 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1478 
1479 	return (0);
1480 
1481 fail_txq_stat_init:
1482 fail_dpl_node:
1483 fail_tx_dpl_put_max:
1484 fail_tx_dpl_get_max:
1485 fail3:
1486 fail_txq_node:
1487 	free(txq->pend_desc, M_SFXGE);
1488 fail2:
1489 	while (nmaps-- != 0)
1490 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1491 	free(txq->stmp, M_SFXGE);
1492 	bus_dma_tag_destroy(txq->packet_dma_tag);
1493 
1494 fail:
1495 	sfxge_dma_free(esmp);
1496 
1497 	return (rc);
1498 }
1499 
1500 static int
1501 sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1502 {
1503 	struct sfxge_softc *sc = arg1;
1504 	unsigned int id = arg2;
1505 	unsigned long sum;
1506 	unsigned int index;
1507 
1508 	/* Sum across all TX queues */
1509 	sum = 0;
1510 	for (index = 0; index < sc->txq_count; index++)
1511 		sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1512 					  sfxge_tx_stats[id].offset);
1513 
1514 	return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1515 }
1516 
1517 static void
1518 sfxge_tx_stat_init(struct sfxge_softc *sc)
1519 {
1520 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1521 	struct sysctl_oid_list *stat_list;
1522 	unsigned int id;
1523 
1524 	stat_list = SYSCTL_CHILDREN(sc->stats_node);
1525 
1526 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1527 		SYSCTL_ADD_PROC(
1528 			ctx, stat_list,
1529 			OID_AUTO, sfxge_tx_stats[id].name,
1530 			CTLTYPE_ULONG|CTLFLAG_RD,
1531 			sc, id, sfxge_tx_stat_handler, "LU",
1532 			"");
1533 	}
1534 }
1535 
1536 uint64_t
1537 sfxge_tx_get_drops(struct sfxge_softc *sc)
1538 {
1539 	unsigned int index;
1540 	uint64_t drops = 0;
1541 	struct sfxge_txq *txq;
1542 
1543 	/* Sum across all TX queues */
1544 	for (index = 0; index < sc->txq_count; index++) {
1545 		txq = sc->txq[index];
1546 		/*
1547 		 * In theory, txq->put_overflow and txq->netdown_drops
1548 		 * should use atomic operation and other should be
1549 		 * obtained under txq lock, but it is just statistics.
1550 		 */
1551 		drops += txq->drops + txq->get_overflow +
1552 			 txq->get_non_tcp_overflow +
1553 			 txq->put_overflow + txq->netdown_drops +
1554 			 txq->tso_pdrop_too_many + txq->tso_pdrop_no_rsrc;
1555 	}
1556 	return (drops);
1557 }
1558 
1559 void
1560 sfxge_tx_fini(struct sfxge_softc *sc)
1561 {
1562 	int index;
1563 
1564 	index = sc->txq_count;
1565 	while (--index >= 0)
1566 		sfxge_tx_qfini(sc, index);
1567 
1568 	sc->txq_count = 0;
1569 }
1570 
1571 
1572 int
1573 sfxge_tx_init(struct sfxge_softc *sc)
1574 {
1575 	struct sfxge_intr *intr;
1576 	int index;
1577 	int rc;
1578 
1579 	intr = &sc->intr;
1580 
1581 	KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1582 	    ("intr->state != SFXGE_INTR_INITIALIZED"));
1583 
1584 	sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc;
1585 
1586 	sc->txqs_node = SYSCTL_ADD_NODE(
1587 		device_get_sysctl_ctx(sc->dev),
1588 		SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
1589 		OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues");
1590 	if (sc->txqs_node == NULL) {
1591 		rc = ENOMEM;
1592 		goto fail_txq_node;
1593 	}
1594 
1595 	/* Initialize the transmit queues */
1596 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
1597 	    SFXGE_TXQ_NON_CKSUM, 0)) != 0)
1598 		goto fail;
1599 
1600 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
1601 	    SFXGE_TXQ_IP_CKSUM, 0)) != 0)
1602 		goto fail2;
1603 
1604 	for (index = 0;
1605 	     index < sc->txq_count - SFXGE_TXQ_NTYPES + 1;
1606 	     index++) {
1607 		if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index,
1608 		    SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
1609 			goto fail3;
1610 	}
1611 
1612 	sfxge_tx_stat_init(sc);
1613 
1614 	return (0);
1615 
1616 fail3:
1617 	while (--index >= 0)
1618 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1619 
1620 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1621 
1622 fail2:
1623 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1624 
1625 fail:
1626 fail_txq_node:
1627 	sc->txq_count = 0;
1628 	return (rc);
1629 }
1630