xref: /freebsd/sys/dev/sfxge/sfxge_tx.c (revision 641a6cfb86023499caafe26a4d821a0b885cf00b)
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 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include <sys/types.h>
34 #include <sys/mbuf.h>
35 #include <sys/smp.h>
36 #include <sys/socket.h>
37 #include <sys/sysctl.h>
38 
39 #include <net/bpf.h>
40 #include <net/ethernet.h>
41 #include <net/if.h>
42 #include <net/if_vlan_var.h>
43 
44 #include <netinet/in.h>
45 #include <netinet/ip.h>
46 #include <netinet/ip6.h>
47 #include <netinet/tcp.h>
48 
49 #include "common/efx.h"
50 
51 #include "sfxge.h"
52 #include "sfxge_tx.h"
53 
54 /* Set the block level to ensure there is space to generate a
55  * large number of descriptors for TSO.  With minimum MSS and
56  * maximum mbuf length we might need more than a ring-ful of
57  * descriptors, but this should not happen in practice except
58  * due to deliberate attack.  In that case we will truncate
59  * the output at a packet boundary.  Allow for a reasonable
60  * minimum MSS of 512.
61  */
62 #define SFXGE_TSO_MAX_DESC ((65535 / 512) * 2 + SFXGE_TX_MAPPING_MAX_SEG - 1)
63 #define SFXGE_TXQ_BLOCK_LEVEL (SFXGE_NDESCS - SFXGE_TSO_MAX_DESC)
64 
65 /* Forward declarations. */
66 static inline void sfxge_tx_qdpl_service(struct sfxge_txq *txq);
67 static void sfxge_tx_qlist_post(struct sfxge_txq *txq);
68 static void sfxge_tx_qunblock(struct sfxge_txq *txq);
69 static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
70 			      const bus_dma_segment_t *dma_seg, int n_dma_seg);
71 
72 void
73 sfxge_tx_qcomplete(struct sfxge_txq *txq)
74 {
75 	struct sfxge_softc *sc;
76 	struct sfxge_evq *evq;
77 	unsigned int completed;
78 
79 	sc = txq->sc;
80 	evq = sc->evq[txq->evq_index];
81 
82 	mtx_assert(&evq->lock, MA_OWNED);
83 
84 	completed = txq->completed;
85 	while (completed != txq->pending) {
86 		struct sfxge_tx_mapping *stmp;
87 		unsigned int id;
88 
89 		id = completed++ & (SFXGE_NDESCS - 1);
90 
91 		stmp = &txq->stmp[id];
92 		if (stmp->flags & TX_BUF_UNMAP) {
93 			bus_dmamap_unload(txq->packet_dma_tag, stmp->map);
94 			if (stmp->flags & TX_BUF_MBUF) {
95 				struct mbuf *m = stmp->u.mbuf;
96 				do
97 					m = m_free(m);
98 				while (m != NULL);
99 			} else {
100 				free(stmp->u.heap_buf, M_SFXGE);
101 			}
102 			stmp->flags = 0;
103 		}
104 	}
105 	txq->completed = completed;
106 
107 	/* Check whether we need to unblock the queue. */
108 	mb();
109 	if (txq->blocked) {
110 		unsigned int level;
111 
112 		level = txq->added - txq->completed;
113 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL)
114 			sfxge_tx_qunblock(txq);
115 	}
116 }
117 
118 #ifdef SFXGE_HAVE_MQ
119 
120 /*
121  * Reorder the put list and append it to the get list.
122  */
123 static void
124 sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq)
125 {
126 	struct sfxge_tx_dpl *stdp;
127 	struct mbuf *mbuf, *get_next, **get_tailp;
128 	volatile uintptr_t *putp;
129 	uintptr_t put;
130 	unsigned int count;
131 
132 	mtx_assert(&txq->lock, MA_OWNED);
133 
134 	stdp = &txq->dpl;
135 
136 	/* Acquire the put list. */
137 	putp = &stdp->std_put;
138 	put = atomic_readandclear_ptr(putp);
139 	mbuf = (void *)put;
140 
141 	if (mbuf == NULL)
142 		return;
143 
144 	/* Reverse the put list. */
145 	get_tailp = &mbuf->m_nextpkt;
146 	get_next = NULL;
147 
148 	count = 0;
149 	do {
150 		struct mbuf *put_next;
151 
152 		put_next = mbuf->m_nextpkt;
153 		mbuf->m_nextpkt = get_next;
154 		get_next = mbuf;
155 		mbuf = put_next;
156 
157 		count++;
158 	} while (mbuf != NULL);
159 
160 	/* Append the reversed put list to the get list. */
161 	KASSERT(*get_tailp == NULL, ("*get_tailp != NULL"));
162 	*stdp->std_getp = get_next;
163 	stdp->std_getp = get_tailp;
164 	stdp->std_count += count;
165 }
166 
167 #endif /* SFXGE_HAVE_MQ */
168 
169 static void
170 sfxge_tx_qreap(struct sfxge_txq *txq)
171 {
172 	mtx_assert(SFXGE_TXQ_LOCK(txq), MA_OWNED);
173 
174 	txq->reaped = txq->completed;
175 }
176 
177 static void
178 sfxge_tx_qlist_post(struct sfxge_txq *txq)
179 {
180 	unsigned int old_added;
181 	unsigned int level;
182 	int rc;
183 
184 	mtx_assert(SFXGE_TXQ_LOCK(txq), MA_OWNED);
185 
186 	KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0"));
187 	KASSERT(txq->n_pend_desc <= SFXGE_TSO_MAX_DESC,
188 		("txq->n_pend_desc too large"));
189 	KASSERT(!txq->blocked, ("txq->blocked"));
190 
191 	old_added = txq->added;
192 
193 	/* Post the fragment list. */
194 	rc = efx_tx_qpost(txq->common, txq->pend_desc, txq->n_pend_desc,
195 			  txq->reaped, &txq->added);
196 	KASSERT(rc == 0, ("efx_tx_qpost() failed"));
197 
198 	/* If efx_tx_qpost() had to refragment, our information about
199 	 * buffers to free may be associated with the wrong
200 	 * descriptors.
201 	 */
202 	KASSERT(txq->added - old_added == txq->n_pend_desc,
203 		("efx_tx_qpost() refragmented descriptors"));
204 
205 	level = txq->added - txq->reaped;
206 	KASSERT(level <= SFXGE_NDESCS, ("overfilled TX queue"));
207 
208 	/* Clear the fragment list. */
209 	txq->n_pend_desc = 0;
210 
211 	/* Have we reached the block level? */
212 	if (level < SFXGE_TXQ_BLOCK_LEVEL)
213 		return;
214 
215 	/* Reap, and check again */
216 	sfxge_tx_qreap(txq);
217 	level = txq->added - txq->reaped;
218 	if (level < SFXGE_TXQ_BLOCK_LEVEL)
219 		return;
220 
221 	txq->blocked = 1;
222 
223 	/*
224 	 * Avoid a race with completion interrupt handling that could leave
225 	 * the queue blocked.
226 	 */
227 	mb();
228 	sfxge_tx_qreap(txq);
229 	level = txq->added - txq->reaped;
230 	if (level < SFXGE_TXQ_BLOCK_LEVEL) {
231 		mb();
232 		txq->blocked = 0;
233 	}
234 }
235 
236 static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
237 {
238 	bus_dmamap_t *used_map;
239 	bus_dmamap_t map;
240 	bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG];
241 	unsigned int id;
242 	struct sfxge_tx_mapping *stmp;
243 	efx_buffer_t *desc;
244 	int n_dma_seg;
245 	int rc;
246 	int i;
247 
248 	KASSERT(!txq->blocked, ("txq->blocked"));
249 
250 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
251 		prefetch_read_many(mbuf->m_data);
252 
253 	if (txq->init_state != SFXGE_TXQ_STARTED) {
254 		rc = EINTR;
255 		goto reject;
256 	}
257 
258 	/* Load the packet for DMA. */
259 	id = txq->added & (SFXGE_NDESCS - 1);
260 	stmp = &txq->stmp[id];
261 	rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map,
262 				     mbuf, dma_seg, &n_dma_seg, 0);
263 	if (rc == EFBIG) {
264 		/* Try again. */
265 		struct mbuf *new_mbuf = m_collapse(mbuf, M_DONTWAIT,
266 						   SFXGE_TX_MAPPING_MAX_SEG);
267 		if (new_mbuf == NULL)
268 			goto reject;
269 		++txq->collapses;
270 		mbuf = new_mbuf;
271 		rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag,
272 					     stmp->map, mbuf,
273 					     dma_seg, &n_dma_seg, 0);
274 	}
275 	if (rc != 0)
276 		goto reject;
277 
278 	/* Make the packet visible to the hardware. */
279 	bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE);
280 
281 	used_map = &stmp->map;
282 
283 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
284 		rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg);
285 		if (rc < 0)
286 			goto reject_mapped;
287 		stmp = &txq->stmp[rc];
288 	} else {
289 		/* Add the mapping to the fragment list, and set flags
290 		 * for the buffer.
291 		 */
292 		i = 0;
293 		for (;;) {
294 			desc = &txq->pend_desc[i];
295 			desc->eb_addr = dma_seg[i].ds_addr;
296 			desc->eb_size = dma_seg[i].ds_len;
297 			if (i == n_dma_seg - 1) {
298 				desc->eb_eop = 1;
299 				break;
300 			}
301 			desc->eb_eop = 0;
302 			i++;
303 
304 			stmp->flags = 0;
305 			if (__predict_false(stmp ==
306 					    &txq->stmp[SFXGE_NDESCS - 1]))
307 				stmp = &txq->stmp[0];
308 			else
309 				stmp++;
310 		}
311 		txq->n_pend_desc = n_dma_seg;
312 	}
313 
314 	/*
315 	 * If the mapping required more than one descriptor
316 	 * then we need to associate the DMA map with the last
317 	 * descriptor, not the first.
318 	 */
319 	if (used_map != &stmp->map) {
320 		map = stmp->map;
321 		stmp->map = *used_map;
322 		*used_map = map;
323 	}
324 
325 	stmp->u.mbuf = mbuf;
326 	stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF;
327 
328 	/* Post the fragment list. */
329 	sfxge_tx_qlist_post(txq);
330 
331 	return 0;
332 
333 reject_mapped:
334 	bus_dmamap_unload(txq->packet_dma_tag, *used_map);
335 reject:
336 	/* Drop the packet on the floor. */
337 	m_freem(mbuf);
338 	++txq->drops;
339 
340 	return rc;
341 }
342 
343 #ifdef SFXGE_HAVE_MQ
344 
345 /*
346  * Drain the deferred packet list into the transmit queue.
347  */
348 static void
349 sfxge_tx_qdpl_drain(struct sfxge_txq *txq)
350 {
351 	struct sfxge_softc *sc;
352 	struct sfxge_tx_dpl *stdp;
353 	struct mbuf *mbuf, *next;
354 	unsigned int count;
355 	unsigned int pushed;
356 	int rc;
357 
358 	mtx_assert(&txq->lock, MA_OWNED);
359 
360 	sc = txq->sc;
361 	stdp = &txq->dpl;
362 	pushed = txq->added;
363 
364 	prefetch_read_many(sc->enp);
365 	prefetch_read_many(txq->common);
366 
367 	mbuf = stdp->std_get;
368 	count = stdp->std_count;
369 
370 	while (count != 0) {
371 		KASSERT(mbuf != NULL, ("mbuf == NULL"));
372 
373 		next = mbuf->m_nextpkt;
374 		mbuf->m_nextpkt = NULL;
375 
376 		ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */
377 
378 		if (next != NULL)
379 			prefetch_read_many(next);
380 
381 		rc = sfxge_tx_queue_mbuf(txq, mbuf);
382 		--count;
383 		mbuf = next;
384 		if (rc != 0)
385 			continue;
386 
387 		if (txq->blocked)
388 			break;
389 
390 		/* Push the fragments to the hardware in batches. */
391 		if (txq->added - pushed >= SFXGE_TX_BATCH) {
392 			efx_tx_qpush(txq->common, txq->added);
393 			pushed = txq->added;
394 		}
395 	}
396 
397 	if (count == 0) {
398 		KASSERT(mbuf == NULL, ("mbuf != NULL"));
399 		stdp->std_get = NULL;
400 		stdp->std_count = 0;
401 		stdp->std_getp = &stdp->std_get;
402 	} else {
403 		stdp->std_get = mbuf;
404 		stdp->std_count = count;
405 	}
406 
407 	if (txq->added != pushed)
408 		efx_tx_qpush(txq->common, txq->added);
409 
410 	KASSERT(txq->blocked || stdp->std_count == 0,
411 		("queue unblocked but count is non-zero"));
412 }
413 
414 #define SFXGE_TX_QDPL_PENDING(_txq)					\
415     ((_txq)->dpl.std_put != 0)
416 
417 /*
418  * Service the deferred packet list.
419  *
420  * NOTE: drops the txq mutex!
421  */
422 static inline void
423 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
424 {
425 	mtx_assert(&txq->lock, MA_OWNED);
426 
427 	do {
428 		if (SFXGE_TX_QDPL_PENDING(txq))
429 			sfxge_tx_qdpl_swizzle(txq);
430 
431 		if (!txq->blocked)
432 			sfxge_tx_qdpl_drain(txq);
433 
434 		mtx_unlock(&txq->lock);
435 	} while (SFXGE_TX_QDPL_PENDING(txq) &&
436 		 mtx_trylock(&txq->lock));
437 }
438 
439 /*
440  * Put a packet on the deferred packet list.
441  *
442  * If we are called with the txq lock held, we put the packet on the "get
443  * list", otherwise we atomically push it on the "put list".  The swizzle
444  * function takes care of ordering.
445  *
446  * The length of the put list is bounded by SFXGE_TX_MAX_DEFFERED.  We
447  * overload the csum_data field in the mbuf to keep track of this length
448  * because there is no cheap alternative to avoid races.
449  */
450 static inline int
451 sfxge_tx_qdpl_put(struct sfxge_txq *txq, struct mbuf *mbuf, int locked)
452 {
453 	struct sfxge_tx_dpl *stdp;
454 
455 	stdp = &txq->dpl;
456 
457 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
458 
459 	if (locked) {
460 		mtx_assert(&txq->lock, MA_OWNED);
461 
462 		sfxge_tx_qdpl_swizzle(txq);
463 
464 		*(stdp->std_getp) = mbuf;
465 		stdp->std_getp = &mbuf->m_nextpkt;
466 		stdp->std_count++;
467 	} else {
468 		volatile uintptr_t *putp;
469 		uintptr_t old;
470 		uintptr_t new;
471 		unsigned old_len;
472 
473 		putp = &stdp->std_put;
474 		new = (uintptr_t)mbuf;
475 
476 		do {
477 			old = *putp;
478 			if (old) {
479 				struct mbuf *mp = (struct mbuf *)old;
480 				old_len = mp->m_pkthdr.csum_data;
481 			} else
482 				old_len = 0;
483 			if (old_len >= SFXGE_TX_MAX_DEFERRED)
484 				return ENOBUFS;
485 			mbuf->m_pkthdr.csum_data = old_len + 1;
486 			mbuf->m_nextpkt = (void *)old;
487 		} while (atomic_cmpset_ptr(putp, old, new) == 0);
488 	}
489 
490 	return (0);
491 }
492 
493 /*
494  * Called from if_transmit - will try to grab the txq lock and enqueue to the
495  * put list if it succeeds, otherwise will push onto the defer list.
496  */
497 int
498 sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m)
499 {
500 	int locked;
501 	int rc;
502 
503 	/*
504 	 * Try to grab the txq lock.  If we are able to get the lock,
505 	 * the packet will be appended to the "get list" of the deferred
506 	 * packet list.  Otherwise, it will be pushed on the "put list".
507 	 */
508 	locked = mtx_trylock(&txq->lock);
509 
510 	/*
511 	 * Can only fail if we weren't able to get the lock.
512 	 */
513 	if (sfxge_tx_qdpl_put(txq, m, locked) != 0) {
514 		KASSERT(!locked,
515 		    ("sfxge_tx_qdpl_put() failed locked"));
516 		rc = ENOBUFS;
517 		goto fail;
518 	}
519 
520 	/*
521 	 * Try to grab the lock again.
522 	 *
523 	 * If we are able to get the lock, we need to process the deferred
524 	 * packet list.  If we are not able to get the lock, another thread
525 	 * is processing the list.
526 	 */
527 	if (!locked)
528 		locked = mtx_trylock(&txq->lock);
529 
530 	if (locked) {
531 		/* Try to service the list. */
532 		sfxge_tx_qdpl_service(txq);
533 		/* Lock has been dropped. */
534 	}
535 
536 	return (0);
537 
538 fail:
539 	return (rc);
540 
541 }
542 
543 static void
544 sfxge_tx_qdpl_flush(struct sfxge_txq *txq)
545 {
546 	struct sfxge_tx_dpl *stdp = &txq->dpl;
547 	struct mbuf *mbuf, *next;
548 
549 	mtx_lock(&txq->lock);
550 
551 	sfxge_tx_qdpl_swizzle(txq);
552 	for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) {
553 		next = mbuf->m_nextpkt;
554 		m_freem(mbuf);
555 	}
556 	stdp->std_get = NULL;
557 	stdp->std_count = 0;
558 	stdp->std_getp = &stdp->std_get;
559 
560 	mtx_unlock(&txq->lock);
561 }
562 
563 void
564 sfxge_if_qflush(struct ifnet *ifp)
565 {
566 	struct sfxge_softc *sc;
567 	int i;
568 
569 	sc = ifp->if_softc;
570 
571 	for (i = 0; i < SFXGE_TX_SCALE(sc); i++)
572 		sfxge_tx_qdpl_flush(sc->txq[i]);
573 }
574 
575 /*
576  * TX start -- called by the stack.
577  */
578 int
579 sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m)
580 {
581 	struct sfxge_softc *sc;
582 	struct sfxge_txq *txq;
583 	int rc;
584 
585 	sc = (struct sfxge_softc *)ifp->if_softc;
586 
587 	KASSERT(ifp->if_flags & IFF_UP, ("interface not up"));
588 
589 	if (!SFXGE_LINK_UP(sc)) {
590 		m_freem(m);
591 		return (0);
592 	}
593 
594 	/* Pick the desired transmit queue. */
595 	if (m->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO)) {
596 		int index = 0;
597 
598 		if (m->m_flags & M_FLOWID) {
599 			uint32_t hash = m->m_pkthdr.flowid;
600 
601 			index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX];
602 		}
603 		txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index];
604 	} else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
605 		txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
606 	} else {
607 		txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
608 	}
609 
610 	rc = sfxge_tx_packet_add(txq, m);
611 
612 	return (rc);
613 }
614 
615 #else /* !SFXGE_HAVE_MQ */
616 
617 static void sfxge_if_start_locked(struct ifnet *ifp)
618 {
619 	struct sfxge_softc *sc = ifp->if_softc;
620 	struct sfxge_txq *txq;
621 	struct mbuf *mbuf;
622 	unsigned int pushed[SFXGE_TXQ_NTYPES];
623 	unsigned int q_index;
624 
625 	if ((ifp->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
626 	    IFF_DRV_RUNNING)
627 		return;
628 
629 	if (!sc->port.link_up)
630 		return;
631 
632 	for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) {
633 		txq = sc->txq[q_index];
634 		pushed[q_index] = txq->added;
635 	}
636 
637 	while (!IFQ_DRV_IS_EMPTY(&ifp->if_snd)) {
638                 IFQ_DRV_DEQUEUE(&ifp->if_snd, mbuf);
639 		if (mbuf == NULL)
640 			break;
641 
642 		ETHER_BPF_MTAP(ifp, mbuf); /* packet capture */
643 
644 		/* Pick the desired transmit queue. */
645 		if (mbuf->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO))
646 			q_index = SFXGE_TXQ_IP_TCP_UDP_CKSUM;
647 		else if (mbuf->m_pkthdr.csum_flags & CSUM_DELAY_IP)
648 			q_index = SFXGE_TXQ_IP_CKSUM;
649 		else
650 			q_index = SFXGE_TXQ_NON_CKSUM;
651 		txq = sc->txq[q_index];
652 
653 		if (sfxge_tx_queue_mbuf(txq, mbuf) != 0)
654 			continue;
655 
656 		if (txq->blocked) {
657 			ifp->if_drv_flags |= IFF_DRV_OACTIVE;
658 			break;
659 		}
660 
661 		/* Push the fragments to the hardware in batches. */
662 		if (txq->added - pushed[q_index] >= SFXGE_TX_BATCH) {
663 			efx_tx_qpush(txq->common, txq->added);
664 			pushed[q_index] = txq->added;
665 		}
666 	}
667 
668 	for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) {
669 		txq = sc->txq[q_index];
670 		if (txq->added != pushed[q_index])
671 			efx_tx_qpush(txq->common, txq->added);
672 	}
673 }
674 
675 void sfxge_if_start(struct ifnet *ifp)
676 {
677 	struct sfxge_softc *sc = ifp->if_softc;
678 
679 	mtx_lock(&sc->tx_lock);
680 	sfxge_if_start_locked(ifp);
681 	mtx_unlock(&sc->tx_lock);
682 }
683 
684 static inline void
685 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
686 {
687 	struct sfxge_softc *sc = txq->sc;
688 	struct ifnet *ifp = sc->ifnet;
689 
690 	mtx_assert(&sc->tx_lock, MA_OWNED);
691 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
692 	sfxge_if_start_locked(ifp);
693 	mtx_unlock(&sc->tx_lock);
694 }
695 
696 #endif /* SFXGE_HAVE_MQ */
697 
698 /*
699  * Software "TSO".  Not quite as good as doing it in hardware, but
700  * still faster than segmenting in the stack.
701  */
702 
703 struct sfxge_tso_state {
704 	/* Output position */
705 	unsigned out_len;	/* Remaining length in current segment */
706 	unsigned seqnum;	/* Current sequence number */
707 	unsigned packet_space;	/* Remaining space in current packet */
708 
709 	/* Input position */
710 	unsigned dma_seg_i;	/* Current DMA segment number */
711 	uint64_t dma_addr;	/* DMA address of current position */
712 	unsigned in_len;	/* Remaining length in current mbuf */
713 
714 	const struct mbuf *mbuf; /* Input mbuf (head of chain) */
715 	u_short protocol;	/* Network protocol (after VLAN decap) */
716 	ssize_t nh_off;		/* Offset of network header */
717 	ssize_t tcph_off;	/* Offset of TCP header */
718 	unsigned header_len;	/* Number of bytes of header */
719 	int full_packet_size;	/* Number of bytes to put in each outgoing
720 				 * segment */
721 };
722 
723 static inline const struct ip *tso_iph(const struct sfxge_tso_state *tso)
724 {
725 	KASSERT(tso->protocol == htons(ETHERTYPE_IP),
726 		("tso_iph() in non-IPv4 state"));
727 	return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
728 }
729 static inline const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
730 {
731 	KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
732 		("tso_ip6h() in non-IPv6 state"));
733 	return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
734 }
735 static inline const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
736 {
737 	return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
738 }
739 
740 /* Size of preallocated TSO header buffers.  Larger blocks must be
741  * allocated from the heap.
742  */
743 #define TSOH_STD_SIZE	128
744 
745 /* At most half the descriptors in the queue at any time will refer to
746  * a TSO header buffer, since they must always be followed by a
747  * payload descriptor referring to an mbuf.
748  */
749 #define TSOH_COUNT	(SFXGE_NDESCS / 2u)
750 #define TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
751 #define TSOH_PAGE_COUNT	((TSOH_COUNT + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE)
752 
753 static int tso_init(struct sfxge_txq *txq)
754 {
755 	struct sfxge_softc *sc = txq->sc;
756 	int i, rc;
757 
758 	/* Allocate TSO header buffers */
759 	txq->tsoh_buffer = malloc(TSOH_PAGE_COUNT * sizeof(txq->tsoh_buffer[0]),
760 				  M_SFXGE, M_WAITOK);
761 
762 	for (i = 0; i < TSOH_PAGE_COUNT; i++) {
763 		rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
764 		if (rc)
765 			goto fail;
766 	}
767 
768 	return 0;
769 
770 fail:
771 	while (i-- > 0)
772 		sfxge_dma_free(&txq->tsoh_buffer[i]);
773 	free(txq->tsoh_buffer, M_SFXGE);
774 	txq->tsoh_buffer = NULL;
775 	return rc;
776 }
777 
778 static void tso_fini(struct sfxge_txq *txq)
779 {
780 	int i;
781 
782 	if (txq->tsoh_buffer) {
783 		for (i = 0; i < TSOH_PAGE_COUNT; i++)
784 			sfxge_dma_free(&txq->tsoh_buffer[i]);
785 		free(txq->tsoh_buffer, M_SFXGE);
786 	}
787 }
788 
789 static void tso_start(struct sfxge_tso_state *tso, struct mbuf *mbuf)
790 {
791 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
792 
793 	tso->mbuf = mbuf;
794 
795 	/* Find network protocol and header */
796 	tso->protocol = eh->ether_type;
797 	if (tso->protocol == htons(ETHERTYPE_VLAN)) {
798 		struct ether_vlan_header *veh =
799 			mtod(mbuf, struct ether_vlan_header *);
800 		tso->protocol = veh->evl_proto;
801 		tso->nh_off = sizeof(*veh);
802 	} else {
803 		tso->nh_off = sizeof(*eh);
804 	}
805 
806 	/* Find TCP header */
807 	if (tso->protocol == htons(ETHERTYPE_IP)) {
808 		KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
809 			("TSO required on non-TCP packet"));
810 		tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
811 	} else {
812 		KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
813 			("TSO required on non-IP packet"));
814 		KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
815 			("TSO required on non-TCP packet"));
816 		tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
817 	}
818 
819 	/* We assume all headers are linear in the head mbuf */
820 	tso->header_len = tso->tcph_off + 4 * tso_tcph(tso)->th_off;
821 	KASSERT(tso->header_len <= mbuf->m_len, ("packet headers fragmented"));
822 	tso->full_packet_size = tso->header_len + mbuf->m_pkthdr.tso_segsz;
823 
824 	tso->seqnum = ntohl(tso_tcph(tso)->th_seq);
825 
826 	/* These flags must not be duplicated */
827 	KASSERT(!(tso_tcph(tso)->th_flags & (TH_URG | TH_SYN | TH_RST)),
828 		("incompatible TCP flag on TSO packet"));
829 
830 	tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
831 }
832 
833 /*
834  * tso_fill_packet_with_fragment - form descriptors for the current fragment
835  *
836  * Form descriptors for the current fragment, until we reach the end
837  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
838  * space.
839  */
840 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
841 					  struct sfxge_tso_state *tso)
842 {
843 	efx_buffer_t *desc;
844 	int n;
845 
846 	if (tso->in_len == 0 || tso->packet_space == 0)
847 		return;
848 
849 	KASSERT(tso->in_len > 0, ("TSO input length went negative"));
850 	KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
851 
852 	n = min(tso->in_len, tso->packet_space);
853 
854 	tso->packet_space -= n;
855 	tso->out_len -= n;
856 	tso->in_len -= n;
857 
858 	desc = &txq->pend_desc[txq->n_pend_desc++];
859 	desc->eb_addr = tso->dma_addr;
860 	desc->eb_size = n;
861 	desc->eb_eop = tso->out_len == 0 || tso->packet_space == 0;
862 
863 	tso->dma_addr += n;
864 }
865 
866 /* Callback from bus_dmamap_load() for long TSO headers. */
867 static void tso_map_long_header(void *dma_addr_ret,
868 				bus_dma_segment_t *segs, int nseg,
869 				int error)
870 {
871 	*(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
872 				      __predict_true(nseg == 1)) ?
873 				     segs->ds_addr : 0);
874 }
875 
876 /*
877  * tso_start_new_packet - generate a new header and prepare for the new packet
878  *
879  * Generate a new header and prepare for the new packet.  Return 0 on
880  * success, or an error code if failed to alloc header.
881  */
882 static int tso_start_new_packet(struct sfxge_txq *txq,
883 				struct sfxge_tso_state *tso,
884 				unsigned int id)
885 {
886 	struct sfxge_tx_mapping *stmp = &txq->stmp[id];
887 	struct tcphdr *tsoh_th;
888 	unsigned ip_length;
889 	caddr_t header;
890 	uint64_t dma_addr;
891 	bus_dmamap_t map;
892 	efx_buffer_t *desc;
893 	int rc;
894 
895 	/* Allocate a DMA-mapped header buffer. */
896 	if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
897 		unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
898 		unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
899 
900 		header = (txq->tsoh_buffer[page_index].esm_base +
901 			  buf_index * TSOH_STD_SIZE);
902 		dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
903 			    buf_index * TSOH_STD_SIZE);
904 		map = txq->tsoh_buffer[page_index].esm_map;
905 
906 		stmp->flags = 0;
907 	} else {
908 		/* We cannot use bus_dmamem_alloc() as that may sleep */
909 		header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
910 		if (__predict_false(!header))
911 			return ENOMEM;
912 		rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
913 				     header, tso->header_len,
914 				     tso_map_long_header, &dma_addr,
915 				     BUS_DMA_NOWAIT);
916 		if (__predict_false(dma_addr == 0)) {
917 			if (rc == 0) {
918 				/* Succeeded but got >1 segment */
919 				bus_dmamap_unload(txq->packet_dma_tag,
920 						  stmp->map);
921 				rc = EINVAL;
922 			}
923 			free(header, M_SFXGE);
924 			return rc;
925 		}
926 		map = stmp->map;
927 
928 		txq->tso_long_headers++;
929 		stmp->u.heap_buf = header;
930 		stmp->flags = TX_BUF_UNMAP;
931 	}
932 
933 	tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
934 
935 	/* Copy and update the headers. */
936 	memcpy(header, tso->mbuf->m_data, tso->header_len);
937 
938 	tsoh_th->th_seq = htonl(tso->seqnum);
939 	tso->seqnum += tso->mbuf->m_pkthdr.tso_segsz;
940 	if (tso->out_len > tso->mbuf->m_pkthdr.tso_segsz) {
941 		/* This packet will not finish the TSO burst. */
942 		ip_length = tso->full_packet_size - tso->nh_off;
943 		tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH);
944 	} else {
945 		/* This packet will be the last in the TSO burst. */
946 		ip_length = tso->header_len - tso->nh_off + tso->out_len;
947 	}
948 
949 	if (tso->protocol == htons(ETHERTYPE_IP)) {
950 		struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
951 		tsoh_iph->ip_len = htons(ip_length);
952 		/* XXX We should increment ip_id, but FreeBSD doesn't
953 		 * currently allocate extra IDs for multiple segments.
954 		 */
955 	} else {
956 		struct ip6_hdr *tsoh_iph =
957 			(struct ip6_hdr *)(header + tso->nh_off);
958 		tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
959 	}
960 
961 	/* Make the header visible to the hardware. */
962 	bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
963 
964 	tso->packet_space = tso->mbuf->m_pkthdr.tso_segsz;
965 	txq->tso_packets++;
966 
967 	/* Form a descriptor for this header. */
968 	desc = &txq->pend_desc[txq->n_pend_desc++];
969 	desc->eb_addr = dma_addr;
970 	desc->eb_size = tso->header_len;
971 	desc->eb_eop = 0;
972 
973 	return 0;
974 }
975 
976 static int
977 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
978 		   const bus_dma_segment_t *dma_seg, int n_dma_seg)
979 {
980 	struct sfxge_tso_state tso;
981 	unsigned int id, next_id;
982 
983 	tso_start(&tso, mbuf);
984 
985 	/* Grab the first payload fragment. */
986 	if (dma_seg->ds_len == tso.header_len) {
987 		--n_dma_seg;
988 		KASSERT(n_dma_seg, ("no payload found in TSO packet"));
989 		++dma_seg;
990 		tso.in_len = dma_seg->ds_len;
991 		tso.dma_addr = dma_seg->ds_addr;
992 	} else {
993 		tso.in_len = dma_seg->ds_len - tso.header_len;
994 		tso.dma_addr = dma_seg->ds_addr + tso.header_len;
995 	}
996 
997 	id = txq->added & (SFXGE_NDESCS - 1);
998 	if (__predict_false(tso_start_new_packet(txq, &tso, id)))
999 		return -1;
1000 
1001 	while (1) {
1002 		id = (id + 1) & (SFXGE_NDESCS - 1);
1003 		tso_fill_packet_with_fragment(txq, &tso);
1004 
1005 		/* Move onto the next fragment? */
1006 		if (tso.in_len == 0) {
1007 			--n_dma_seg;
1008 			if (n_dma_seg == 0)
1009 				break;
1010 			++dma_seg;
1011 			tso.in_len = dma_seg->ds_len;
1012 			tso.dma_addr = dma_seg->ds_addr;
1013 		}
1014 
1015 		/* End of packet? */
1016 		if (tso.packet_space == 0) {
1017 			/* If the queue is now full due to tiny MSS,
1018 			 * or we can't create another header, discard
1019 			 * the remainder of the input mbuf but do not
1020 			 * roll back the work we have done.
1021 			 */
1022 			if (txq->n_pend_desc >
1023 			    SFXGE_TSO_MAX_DESC - (1 + SFXGE_TX_MAPPING_MAX_SEG))
1024 				break;
1025 			next_id = (id + 1) & (SFXGE_NDESCS - 1);
1026 			if (__predict_false(tso_start_new_packet(txq, &tso,
1027 								 next_id)))
1028 				break;
1029 			id = next_id;
1030 		}
1031 	}
1032 
1033 	txq->tso_bursts++;
1034 	return id;
1035 }
1036 
1037 static void
1038 sfxge_tx_qunblock(struct sfxge_txq *txq)
1039 {
1040 	struct sfxge_softc *sc;
1041 	struct sfxge_evq *evq;
1042 
1043 	sc = txq->sc;
1044 	evq = sc->evq[txq->evq_index];
1045 
1046 	mtx_assert(&evq->lock, MA_OWNED);
1047 
1048 	if (txq->init_state != SFXGE_TXQ_STARTED)
1049 		return;
1050 
1051 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1052 
1053 	if (txq->blocked) {
1054 		unsigned int level;
1055 
1056 		level = txq->added - txq->completed;
1057 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL)
1058 			txq->blocked = 0;
1059 	}
1060 
1061 	sfxge_tx_qdpl_service(txq);
1062 	/* note: lock has been dropped */
1063 }
1064 
1065 void
1066 sfxge_tx_qflush_done(struct sfxge_txq *txq)
1067 {
1068 
1069 	txq->flush_state = SFXGE_FLUSH_DONE;
1070 }
1071 
1072 static void
1073 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1074 {
1075 	struct sfxge_txq *txq;
1076 	struct sfxge_evq *evq;
1077 	unsigned int count;
1078 
1079 	txq = sc->txq[index];
1080 	evq = sc->evq[txq->evq_index];
1081 
1082 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1083 
1084 	KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1085 	    ("txq->init_state != SFXGE_TXQ_STARTED"));
1086 
1087 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1088 	txq->flush_state = SFXGE_FLUSH_PENDING;
1089 
1090 	/* Flush the transmit queue. */
1091 	efx_tx_qflush(txq->common);
1092 
1093 	mtx_unlock(SFXGE_TXQ_LOCK(txq));
1094 
1095 	count = 0;
1096 	do {
1097 		/* Spin for 100ms. */
1098 		DELAY(100000);
1099 
1100 		if (txq->flush_state != SFXGE_FLUSH_PENDING)
1101 			break;
1102 	} while (++count < 20);
1103 
1104 	mtx_lock(&evq->lock);
1105 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1106 
1107 	KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1108 	    ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1109 
1110 	txq->flush_state = SFXGE_FLUSH_DONE;
1111 
1112 	txq->blocked = 0;
1113 	txq->pending = txq->added;
1114 
1115 	sfxge_tx_qcomplete(txq);
1116 	KASSERT(txq->completed == txq->added,
1117 	    ("txq->completed != txq->added"));
1118 
1119 	sfxge_tx_qreap(txq);
1120 	KASSERT(txq->reaped == txq->completed,
1121 	    ("txq->reaped != txq->completed"));
1122 
1123 	txq->added = 0;
1124 	txq->pending = 0;
1125 	txq->completed = 0;
1126 	txq->reaped = 0;
1127 
1128 	/* Destroy the common code transmit queue. */
1129 	efx_tx_qdestroy(txq->common);
1130 	txq->common = NULL;
1131 
1132 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1133 	    EFX_TXQ_NBUFS(SFXGE_NDESCS));
1134 
1135 	mtx_unlock(&evq->lock);
1136 	mtx_unlock(SFXGE_TXQ_LOCK(txq));
1137 }
1138 
1139 static int
1140 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1141 {
1142 	struct sfxge_txq *txq;
1143 	efsys_mem_t *esmp;
1144 	uint16_t flags;
1145 	struct sfxge_evq *evq;
1146 	int rc;
1147 
1148 	txq = sc->txq[index];
1149 	esmp = &txq->mem;
1150 	evq = sc->evq[txq->evq_index];
1151 
1152 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1153 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1154 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1155 	    ("evq->init_state != SFXGE_EVQ_STARTED"));
1156 
1157 	/* Program the buffer table. */
1158 	if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1159 	    EFX_TXQ_NBUFS(SFXGE_NDESCS))) != 0)
1160 		return rc;
1161 
1162 	/* Determine the kind of queue we are creating. */
1163 	switch (txq->type) {
1164 	case SFXGE_TXQ_NON_CKSUM:
1165 		flags = 0;
1166 		break;
1167 	case SFXGE_TXQ_IP_CKSUM:
1168 		flags = EFX_CKSUM_IPV4;
1169 		break;
1170 	case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1171 		flags = EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP;
1172 		break;
1173 	default:
1174 		KASSERT(0, ("Impossible TX queue"));
1175 		flags = 0;
1176 		break;
1177 	}
1178 
1179 	/* Create the common code transmit queue. */
1180 	if ((rc = efx_tx_qcreate(sc->enp, index, index, esmp,
1181 	    SFXGE_NDESCS, txq->buf_base_id, flags, evq->common,
1182 	    &txq->common)) != 0)
1183 		goto fail;
1184 
1185 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1186 
1187 	/* Enable the transmit queue. */
1188 	efx_tx_qenable(txq->common);
1189 
1190 	txq->init_state = SFXGE_TXQ_STARTED;
1191 
1192 	mtx_unlock(SFXGE_TXQ_LOCK(txq));
1193 
1194 	return (0);
1195 
1196 fail:
1197 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1198 	    EFX_TXQ_NBUFS(SFXGE_NDESCS));
1199 	return rc;
1200 }
1201 
1202 void
1203 sfxge_tx_stop(struct sfxge_softc *sc)
1204 {
1205 	const efx_nic_cfg_t *encp;
1206 	int index;
1207 
1208 	index = SFXGE_TX_SCALE(sc);
1209 	while (--index >= 0)
1210 		sfxge_tx_qstop(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1211 
1212 	sfxge_tx_qstop(sc, SFXGE_TXQ_IP_CKSUM);
1213 
1214 	encp = efx_nic_cfg_get(sc->enp);
1215         sfxge_tx_qstop(sc, SFXGE_TXQ_NON_CKSUM);
1216 
1217 	/* Tear down the transmit module */
1218 	efx_tx_fini(sc->enp);
1219 }
1220 
1221 int
1222 sfxge_tx_start(struct sfxge_softc *sc)
1223 {
1224 	int index;
1225 	int rc;
1226 
1227 	/* Initialize the common code transmit module. */
1228 	if ((rc = efx_tx_init(sc->enp)) != 0)
1229 		return (rc);
1230 
1231 	if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_NON_CKSUM)) != 0)
1232 		goto fail;
1233 
1234 	if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_IP_CKSUM)) != 0)
1235 		goto fail2;
1236 
1237 	for (index = 0; index < SFXGE_TX_SCALE(sc); index++) {
1238 		if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM +
1239 		    index)) != 0)
1240 			goto fail3;
1241 	}
1242 
1243 	return (0);
1244 
1245 fail3:
1246 	while (--index >= 0)
1247 		sfxge_tx_qstop(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1248 
1249 	sfxge_tx_qstop(sc, SFXGE_TXQ_IP_CKSUM);
1250 
1251 fail2:
1252         sfxge_tx_qstop(sc, SFXGE_TXQ_NON_CKSUM);
1253 
1254 fail:
1255 	efx_tx_fini(sc->enp);
1256 
1257 	return (rc);
1258 }
1259 
1260 /**
1261  * Destroy a transmit queue.
1262  */
1263 static void
1264 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1265 {
1266 	struct sfxge_txq *txq;
1267 	unsigned int nmaps = SFXGE_NDESCS;
1268 
1269 	txq = sc->txq[index];
1270 
1271 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1272 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1273 
1274 	if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1275 		tso_fini(txq);
1276 
1277 	/* Free the context arrays. */
1278 	free(txq->pend_desc, M_SFXGE);
1279 	while (nmaps--)
1280 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1281 	free(txq->stmp, M_SFXGE);
1282 
1283 	/* Release DMA memory mapping. */
1284 	sfxge_dma_free(&txq->mem);
1285 
1286 	sc->txq[index] = NULL;
1287 
1288 #ifdef SFXGE_HAVE_MQ
1289 	mtx_destroy(&txq->lock);
1290 #endif
1291 
1292 	free(txq, M_SFXGE);
1293 }
1294 
1295 static int
1296 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1297     enum sfxge_txq_type type, unsigned int evq_index)
1298 {
1299 	struct sfxge_txq *txq;
1300 	struct sfxge_evq *evq;
1301 #ifdef SFXGE_HAVE_MQ
1302 	struct sfxge_tx_dpl *stdp;
1303 #endif
1304 	efsys_mem_t *esmp;
1305 	unsigned int nmaps;
1306 	int rc;
1307 
1308 	txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1309 	txq->sc = sc;
1310 
1311 	sc->txq[txq_index] = txq;
1312 	esmp = &txq->mem;
1313 
1314 	evq = sc->evq[evq_index];
1315 
1316 	/* Allocate and zero DMA space for the descriptor ring. */
1317 	if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(SFXGE_NDESCS), esmp)) != 0)
1318 		return (rc);
1319 	(void)memset(esmp->esm_base, 0, EFX_TXQ_SIZE(SFXGE_NDESCS));
1320 
1321 	/* Allocate buffer table entries. */
1322 	sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(SFXGE_NDESCS),
1323 				 &txq->buf_base_id);
1324 
1325 	/* Create a DMA tag for packet mappings. */
1326 	if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000,
1327 	    MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1328 	    NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL,
1329 	    &txq->packet_dma_tag) != 0) {
1330 		device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1331 		rc = ENOMEM;
1332 		goto fail;
1333 	}
1334 
1335 	/* Allocate pending descriptor array for batching writes. */
1336 	txq->pend_desc = malloc(sizeof(efx_buffer_t) * SFXGE_NDESCS,
1337 				M_SFXGE, M_ZERO | M_WAITOK);
1338 
1339 	/* Allocate and initialise mbuf DMA mapping array. */
1340 	txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * SFXGE_NDESCS,
1341 	    M_SFXGE, M_ZERO | M_WAITOK);
1342 	for (nmaps = 0; nmaps < SFXGE_NDESCS; nmaps++) {
1343 		rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1344 				       &txq->stmp[nmaps].map);
1345 		if (rc != 0)
1346 			goto fail2;
1347 	}
1348 
1349 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1350 	    (rc = tso_init(txq)) != 0)
1351 		goto fail3;
1352 
1353 #ifdef SFXGE_HAVE_MQ
1354 	/* Initialize the deferred packet list. */
1355 	stdp = &txq->dpl;
1356 	stdp->std_getp = &stdp->std_get;
1357 
1358 	mtx_init(&txq->lock, "txq", NULL, MTX_DEF);
1359 #endif
1360 
1361 	txq->type = type;
1362 	txq->evq_index = evq_index;
1363 	txq->txq_index = txq_index;
1364 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1365 
1366 	return (0);
1367 
1368 fail3:
1369 	free(txq->pend_desc, M_SFXGE);
1370 fail2:
1371 	while (nmaps--)
1372 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1373 	free(txq->stmp, M_SFXGE);
1374 	bus_dma_tag_destroy(txq->packet_dma_tag);
1375 
1376 fail:
1377 	sfxge_dma_free(esmp);
1378 
1379 	return (rc);
1380 }
1381 
1382 static const struct {
1383 	const char *name;
1384 	size_t offset;
1385 } sfxge_tx_stats[] = {
1386 #define SFXGE_TX_STAT(name, member) \
1387 	{ #name, offsetof(struct sfxge_txq, member) }
1388 	SFXGE_TX_STAT(tso_bursts, tso_bursts),
1389 	SFXGE_TX_STAT(tso_packets, tso_packets),
1390 	SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
1391 	SFXGE_TX_STAT(tx_collapses, collapses),
1392 	SFXGE_TX_STAT(tx_drops, drops),
1393 };
1394 
1395 static int
1396 sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1397 {
1398 	struct sfxge_softc *sc = arg1;
1399 	unsigned int id = arg2;
1400 	unsigned long sum;
1401 	unsigned int index;
1402 
1403 	/* Sum across all TX queues */
1404 	sum = 0;
1405 	for (index = 0;
1406 	     index < SFXGE_TXQ_IP_TCP_UDP_CKSUM + SFXGE_TX_SCALE(sc);
1407 	     index++)
1408 		sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1409 					  sfxge_tx_stats[id].offset);
1410 
1411 	return SYSCTL_OUT(req, &sum, sizeof(sum));
1412 }
1413 
1414 static void
1415 sfxge_tx_stat_init(struct sfxge_softc *sc)
1416 {
1417 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1418 	struct sysctl_oid_list *stat_list;
1419 	unsigned int id;
1420 
1421 	stat_list = SYSCTL_CHILDREN(sc->stats_node);
1422 
1423 	for (id = 0;
1424 	     id < sizeof(sfxge_tx_stats) / sizeof(sfxge_tx_stats[0]);
1425 	     id++) {
1426 		SYSCTL_ADD_PROC(
1427 			ctx, stat_list,
1428 			OID_AUTO, sfxge_tx_stats[id].name,
1429 			CTLTYPE_ULONG|CTLFLAG_RD,
1430 			sc, id, sfxge_tx_stat_handler, "LU",
1431 			"");
1432 	}
1433 }
1434 
1435 void
1436 sfxge_tx_fini(struct sfxge_softc *sc)
1437 {
1438 	int index;
1439 
1440 	index = SFXGE_TX_SCALE(sc);
1441 	while (--index >= 0)
1442 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1443 
1444 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1445         sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1446 }
1447 
1448 
1449 int
1450 sfxge_tx_init(struct sfxge_softc *sc)
1451 {
1452 	struct sfxge_intr *intr;
1453 	int index;
1454 	int rc;
1455 
1456 	intr = &sc->intr;
1457 
1458 	KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1459 	    ("intr->state != SFXGE_INTR_INITIALIZED"));
1460 
1461 	/* Initialize the transmit queues */
1462 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
1463 	    SFXGE_TXQ_NON_CKSUM, 0)) != 0)
1464 		goto fail;
1465 
1466 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
1467 	    SFXGE_TXQ_IP_CKSUM, 0)) != 0)
1468 		goto fail2;
1469 
1470 	for (index = 0; index < SFXGE_TX_SCALE(sc); index++) {
1471 		if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index,
1472 		    SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
1473 			goto fail3;
1474 	}
1475 
1476 	sfxge_tx_stat_init(sc);
1477 
1478 	return (0);
1479 
1480 fail3:
1481 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1482 
1483 	while (--index >= 0)
1484 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1485 
1486 fail2:
1487 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1488 
1489 fail:
1490 	return (rc);
1491 }
1492