xref: /freebsd/sys/dev/sfxge/sfxge_tx.c (revision 3823d5e198425b4f5e5a80267d195769d1063773)
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 /* Set the block level to ensure there is space to generate a
71  * large number of descriptors for TSO.  With minimum MSS and
72  * maximum mbuf length we might need more than a ring-ful of
73  * descriptors, but this should not happen in practice except
74  * due to deliberate attack.  In that case we will truncate
75  * the output at a packet boundary.  Allow for a reasonable
76  * minimum MSS of 512.
77  */
78 #define	SFXGE_TSO_MAX_DESC ((65535 / 512) * 2 + SFXGE_TX_MAPPING_MAX_SEG - 1)
79 #define	SFXGE_TXQ_BLOCK_LEVEL(_entries)	((_entries) - SFXGE_TSO_MAX_DESC)
80 
81 #ifdef SFXGE_HAVE_MQ
82 
83 #define	SFXGE_PARAM_TX_DPL_GET_MAX	SFXGE_PARAM(tx_dpl_get_max)
84 static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT;
85 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max);
86 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN,
87 	   &sfxge_tx_dpl_get_max, 0,
88 	   "Maximum number of packets in deferred packet get-list");
89 
90 #define	SFXGE_PARAM_TX_DPL_PUT_MAX	SFXGE_PARAM(tx_dpl_put_max)
91 static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT;
92 TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max);
93 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN,
94 	   &sfxge_tx_dpl_put_max, 0,
95 	   "Maximum number of packets in deferred packet put-list");
96 
97 #endif
98 
99 
100 /* Forward declarations. */
101 static inline void sfxge_tx_qdpl_service(struct sfxge_txq *txq);
102 static void sfxge_tx_qlist_post(struct sfxge_txq *txq);
103 static void sfxge_tx_qunblock(struct sfxge_txq *txq);
104 static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
105 			      const bus_dma_segment_t *dma_seg, int n_dma_seg);
106 
107 void
108 sfxge_tx_qcomplete(struct sfxge_txq *txq)
109 {
110 	struct sfxge_softc *sc;
111 	struct sfxge_evq *evq;
112 	unsigned int completed;
113 
114 	sc = txq->sc;
115 	evq = sc->evq[txq->evq_index];
116 
117 	mtx_assert(&evq->lock, MA_OWNED);
118 
119 	completed = txq->completed;
120 	while (completed != txq->pending) {
121 		struct sfxge_tx_mapping *stmp;
122 		unsigned int id;
123 
124 		id = completed++ & txq->ptr_mask;
125 
126 		stmp = &txq->stmp[id];
127 		if (stmp->flags & TX_BUF_UNMAP) {
128 			bus_dmamap_unload(txq->packet_dma_tag, stmp->map);
129 			if (stmp->flags & TX_BUF_MBUF) {
130 				struct mbuf *m = stmp->u.mbuf;
131 				do
132 					m = m_free(m);
133 				while (m != NULL);
134 			} else {
135 				free(stmp->u.heap_buf, M_SFXGE);
136 			}
137 			stmp->flags = 0;
138 		}
139 	}
140 	txq->completed = completed;
141 
142 	/* Check whether we need to unblock the queue. */
143 	mb();
144 	if (txq->blocked) {
145 		unsigned int level;
146 
147 		level = txq->added - txq->completed;
148 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
149 			sfxge_tx_qunblock(txq);
150 	}
151 }
152 
153 #ifdef SFXGE_HAVE_MQ
154 
155 /*
156  * Reorder the put list and append it to the get list.
157  */
158 static void
159 sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq)
160 {
161 	struct sfxge_tx_dpl *stdp;
162 	struct mbuf *mbuf, *get_next, **get_tailp;
163 	volatile uintptr_t *putp;
164 	uintptr_t put;
165 	unsigned int count;
166 
167 	mtx_assert(&txq->lock, MA_OWNED);
168 
169 	stdp = &txq->dpl;
170 
171 	/* Acquire the put list. */
172 	putp = &stdp->std_put;
173 	put = atomic_readandclear_ptr(putp);
174 	mbuf = (void *)put;
175 
176 	if (mbuf == NULL)
177 		return;
178 
179 	/* Reverse the put list. */
180 	get_tailp = &mbuf->m_nextpkt;
181 	get_next = NULL;
182 
183 	count = 0;
184 	do {
185 		struct mbuf *put_next;
186 
187 		put_next = mbuf->m_nextpkt;
188 		mbuf->m_nextpkt = get_next;
189 		get_next = mbuf;
190 		mbuf = put_next;
191 
192 		count++;
193 	} while (mbuf != NULL);
194 
195 	/* Append the reversed put list to the get list. */
196 	KASSERT(*get_tailp == NULL, ("*get_tailp != NULL"));
197 	*stdp->std_getp = get_next;
198 	stdp->std_getp = get_tailp;
199 	stdp->std_get_count += count;
200 }
201 
202 #endif /* SFXGE_HAVE_MQ */
203 
204 static void
205 sfxge_tx_qreap(struct sfxge_txq *txq)
206 {
207 	mtx_assert(SFXGE_TXQ_LOCK(txq), MA_OWNED);
208 
209 	txq->reaped = txq->completed;
210 }
211 
212 static void
213 sfxge_tx_qlist_post(struct sfxge_txq *txq)
214 {
215 	unsigned int old_added;
216 	unsigned int level;
217 	int rc;
218 
219 	mtx_assert(SFXGE_TXQ_LOCK(txq), MA_OWNED);
220 
221 	KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0"));
222 	KASSERT(txq->n_pend_desc <= SFXGE_TSO_MAX_DESC,
223 		("txq->n_pend_desc too large"));
224 	KASSERT(!txq->blocked, ("txq->blocked"));
225 
226 	old_added = txq->added;
227 
228 	/* Post the fragment list. */
229 	rc = efx_tx_qpost(txq->common, txq->pend_desc, txq->n_pend_desc,
230 			  txq->reaped, &txq->added);
231 	KASSERT(rc == 0, ("efx_tx_qpost() failed"));
232 
233 	/* If efx_tx_qpost() had to refragment, our information about
234 	 * buffers to free may be associated with the wrong
235 	 * descriptors.
236 	 */
237 	KASSERT(txq->added - old_added == txq->n_pend_desc,
238 		("efx_tx_qpost() refragmented descriptors"));
239 
240 	level = txq->added - txq->reaped;
241 	KASSERT(level <= txq->entries, ("overfilled TX queue"));
242 
243 	/* Clear the fragment list. */
244 	txq->n_pend_desc = 0;
245 
246 	/* Have we reached the block level? */
247 	if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries))
248 		return;
249 
250 	/* Reap, and check again */
251 	sfxge_tx_qreap(txq);
252 	level = txq->added - txq->reaped;
253 	if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries))
254 		return;
255 
256 	txq->blocked = 1;
257 
258 	/*
259 	 * Avoid a race with completion interrupt handling that could leave
260 	 * the queue blocked.
261 	 */
262 	mb();
263 	sfxge_tx_qreap(txq);
264 	level = txq->added - txq->reaped;
265 	if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries)) {
266 		mb();
267 		txq->blocked = 0;
268 	}
269 }
270 
271 static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
272 {
273 	bus_dmamap_t *used_map;
274 	bus_dmamap_t map;
275 	bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG];
276 	unsigned int id;
277 	struct sfxge_tx_mapping *stmp;
278 	efx_buffer_t *desc;
279 	int n_dma_seg;
280 	int rc;
281 	int i;
282 
283 	KASSERT(!txq->blocked, ("txq->blocked"));
284 
285 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
286 		prefetch_read_many(mbuf->m_data);
287 
288 	if (txq->init_state != SFXGE_TXQ_STARTED) {
289 		rc = EINTR;
290 		goto reject;
291 	}
292 
293 	/* Load the packet for DMA. */
294 	id = txq->added & txq->ptr_mask;
295 	stmp = &txq->stmp[id];
296 	rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map,
297 				     mbuf, dma_seg, &n_dma_seg, 0);
298 	if (rc == EFBIG) {
299 		/* Try again. */
300 		struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT,
301 						   SFXGE_TX_MAPPING_MAX_SEG);
302 		if (new_mbuf == NULL)
303 			goto reject;
304 		++txq->collapses;
305 		mbuf = new_mbuf;
306 		rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag,
307 					     stmp->map, mbuf,
308 					     dma_seg, &n_dma_seg, 0);
309 	}
310 	if (rc != 0)
311 		goto reject;
312 
313 	/* Make the packet visible to the hardware. */
314 	bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE);
315 
316 	used_map = &stmp->map;
317 
318 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
319 		rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg);
320 		if (rc < 0)
321 			goto reject_mapped;
322 		stmp = &txq->stmp[rc];
323 	} else {
324 		/* Add the mapping to the fragment list, and set flags
325 		 * for the buffer.
326 		 */
327 		i = 0;
328 		for (;;) {
329 			desc = &txq->pend_desc[i];
330 			desc->eb_addr = dma_seg[i].ds_addr;
331 			desc->eb_size = dma_seg[i].ds_len;
332 			if (i == n_dma_seg - 1) {
333 				desc->eb_eop = 1;
334 				break;
335 			}
336 			desc->eb_eop = 0;
337 			i++;
338 
339 			stmp->flags = 0;
340 			if (__predict_false(stmp ==
341 					    &txq->stmp[txq->ptr_mask]))
342 				stmp = &txq->stmp[0];
343 			else
344 				stmp++;
345 		}
346 		txq->n_pend_desc = n_dma_seg;
347 	}
348 
349 	/*
350 	 * If the mapping required more than one descriptor
351 	 * then we need to associate the DMA map with the last
352 	 * descriptor, not the first.
353 	 */
354 	if (used_map != &stmp->map) {
355 		map = stmp->map;
356 		stmp->map = *used_map;
357 		*used_map = map;
358 	}
359 
360 	stmp->u.mbuf = mbuf;
361 	stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF;
362 
363 	/* Post the fragment list. */
364 	sfxge_tx_qlist_post(txq);
365 
366 	return (0);
367 
368 reject_mapped:
369 	bus_dmamap_unload(txq->packet_dma_tag, *used_map);
370 reject:
371 	/* Drop the packet on the floor. */
372 	m_freem(mbuf);
373 	++txq->drops;
374 
375 	return (rc);
376 }
377 
378 #ifdef SFXGE_HAVE_MQ
379 
380 /*
381  * Drain the deferred packet list into the transmit queue.
382  */
383 static void
384 sfxge_tx_qdpl_drain(struct sfxge_txq *txq)
385 {
386 	struct sfxge_softc *sc;
387 	struct sfxge_tx_dpl *stdp;
388 	struct mbuf *mbuf, *next;
389 	unsigned int count;
390 	unsigned int pushed;
391 	int rc;
392 
393 	mtx_assert(&txq->lock, MA_OWNED);
394 
395 	sc = txq->sc;
396 	stdp = &txq->dpl;
397 	pushed = txq->added;
398 
399 	prefetch_read_many(sc->enp);
400 	prefetch_read_many(txq->common);
401 
402 	mbuf = stdp->std_get;
403 	count = stdp->std_get_count;
404 
405 	while (count != 0) {
406 		KASSERT(mbuf != NULL, ("mbuf == NULL"));
407 
408 		next = mbuf->m_nextpkt;
409 		mbuf->m_nextpkt = NULL;
410 
411 		ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */
412 
413 		if (next != NULL)
414 			prefetch_read_many(next);
415 
416 		rc = sfxge_tx_queue_mbuf(txq, mbuf);
417 		--count;
418 		mbuf = next;
419 		if (rc != 0)
420 			continue;
421 
422 		if (txq->blocked)
423 			break;
424 
425 		/* Push the fragments to the hardware in batches. */
426 		if (txq->added - pushed >= SFXGE_TX_BATCH) {
427 			efx_tx_qpush(txq->common, txq->added);
428 			pushed = txq->added;
429 		}
430 	}
431 
432 	if (count == 0) {
433 		KASSERT(mbuf == NULL, ("mbuf != NULL"));
434 		stdp->std_get = NULL;
435 		stdp->std_get_count = 0;
436 		stdp->std_getp = &stdp->std_get;
437 	} else {
438 		stdp->std_get = mbuf;
439 		stdp->std_get_count = count;
440 	}
441 
442 	if (txq->added != pushed)
443 		efx_tx_qpush(txq->common, txq->added);
444 
445 	KASSERT(txq->blocked || stdp->std_get_count == 0,
446 		("queue unblocked but count is non-zero"));
447 }
448 
449 #define	SFXGE_TX_QDPL_PENDING(_txq)					\
450 	((_txq)->dpl.std_put != 0)
451 
452 /*
453  * Service the deferred packet list.
454  *
455  * NOTE: drops the txq mutex!
456  */
457 static inline void
458 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
459 {
460 	mtx_assert(&txq->lock, MA_OWNED);
461 
462 	do {
463 		if (SFXGE_TX_QDPL_PENDING(txq))
464 			sfxge_tx_qdpl_swizzle(txq);
465 
466 		if (!txq->blocked)
467 			sfxge_tx_qdpl_drain(txq);
468 
469 		mtx_unlock(&txq->lock);
470 	} while (SFXGE_TX_QDPL_PENDING(txq) &&
471 		 mtx_trylock(&txq->lock));
472 }
473 
474 /*
475  * Put a packet on the deferred packet list.
476  *
477  * If we are called with the txq lock held, we put the packet on the "get
478  * list", otherwise we atomically push it on the "put list".  The swizzle
479  * function takes care of ordering.
480  *
481  * The length of the put list is bounded by SFXGE_TX_MAX_DEFFERED.  We
482  * overload the csum_data field in the mbuf to keep track of this length
483  * because there is no cheap alternative to avoid races.
484  */
485 static inline int
486 sfxge_tx_qdpl_put(struct sfxge_txq *txq, struct mbuf *mbuf, int locked)
487 {
488 	struct sfxge_tx_dpl *stdp;
489 
490 	stdp = &txq->dpl;
491 
492 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
493 
494 	if (locked) {
495 		mtx_assert(&txq->lock, MA_OWNED);
496 
497 		sfxge_tx_qdpl_swizzle(txq);
498 
499 		if (stdp->std_get_count >= stdp->std_get_max)
500 			return (ENOBUFS);
501 
502 		*(stdp->std_getp) = mbuf;
503 		stdp->std_getp = &mbuf->m_nextpkt;
504 		stdp->std_get_count++;
505 	} else {
506 		volatile uintptr_t *putp;
507 		uintptr_t old;
508 		uintptr_t new;
509 		unsigned old_len;
510 
511 		putp = &stdp->std_put;
512 		new = (uintptr_t)mbuf;
513 
514 		do {
515 			old = *putp;
516 			if (old != 0) {
517 				struct mbuf *mp = (struct mbuf *)old;
518 				old_len = mp->m_pkthdr.csum_data;
519 			} else
520 				old_len = 0;
521 			if (old_len >= stdp->std_put_max)
522 				return (ENOBUFS);
523 			mbuf->m_pkthdr.csum_data = old_len + 1;
524 			mbuf->m_nextpkt = (void *)old;
525 		} while (atomic_cmpset_ptr(putp, old, new) == 0);
526 	}
527 
528 	return (0);
529 }
530 
531 /*
532  * Called from if_transmit - will try to grab the txq lock and enqueue to the
533  * put list if it succeeds, otherwise will push onto the defer list.
534  */
535 int
536 sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m)
537 {
538 	int locked;
539 	int rc;
540 
541 	if (!SFXGE_LINK_UP(txq->sc)) {
542 		rc = ENETDOWN;
543 		goto fail;
544 	}
545 
546 	/*
547 	 * Try to grab the txq lock.  If we are able to get the lock,
548 	 * the packet will be appended to the "get list" of the deferred
549 	 * packet list.  Otherwise, it will be pushed on the "put list".
550 	 */
551 	locked = mtx_trylock(&txq->lock);
552 
553 	if (sfxge_tx_qdpl_put(txq, m, locked) != 0) {
554 		if (locked)
555 			mtx_unlock(&txq->lock);
556 		rc = ENOBUFS;
557 		goto fail;
558 	}
559 
560 	/*
561 	 * Try to grab the lock again.
562 	 *
563 	 * If we are able to get the lock, we need to process the deferred
564 	 * packet list.  If we are not able to get the lock, another thread
565 	 * is processing the list.
566 	 */
567 	if (!locked)
568 		locked = mtx_trylock(&txq->lock);
569 
570 	if (locked) {
571 		/* Try to service the list. */
572 		sfxge_tx_qdpl_service(txq);
573 		/* Lock has been dropped. */
574 	}
575 
576 	return (0);
577 
578 fail:
579 	m_freem(m);
580 	atomic_add_long(&txq->early_drops, 1);
581 	return (rc);
582 }
583 
584 static void
585 sfxge_tx_qdpl_flush(struct sfxge_txq *txq)
586 {
587 	struct sfxge_tx_dpl *stdp = &txq->dpl;
588 	struct mbuf *mbuf, *next;
589 
590 	mtx_lock(&txq->lock);
591 
592 	sfxge_tx_qdpl_swizzle(txq);
593 	for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) {
594 		next = mbuf->m_nextpkt;
595 		m_freem(mbuf);
596 	}
597 	stdp->std_get = NULL;
598 	stdp->std_get_count = 0;
599 	stdp->std_getp = &stdp->std_get;
600 
601 	mtx_unlock(&txq->lock);
602 }
603 
604 void
605 sfxge_if_qflush(struct ifnet *ifp)
606 {
607 	struct sfxge_softc *sc;
608 	int i;
609 
610 	sc = ifp->if_softc;
611 
612 	for (i = 0; i < SFXGE_TX_SCALE(sc); i++)
613 		sfxge_tx_qdpl_flush(sc->txq[i]);
614 }
615 
616 /*
617  * TX start -- called by the stack.
618  */
619 int
620 sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m)
621 {
622 	struct sfxge_softc *sc;
623 	struct sfxge_txq *txq;
624 	int rc;
625 
626 	sc = (struct sfxge_softc *)ifp->if_softc;
627 
628 	KASSERT(ifp->if_flags & IFF_UP, ("interface not up"));
629 
630 	/* Pick the desired transmit queue. */
631 	if (m->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO)) {
632 		int index = 0;
633 
634 		if (m->m_flags & M_FLOWID) {
635 			uint32_t hash = m->m_pkthdr.flowid;
636 
637 			index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX];
638 		}
639 		txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index];
640 	} else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
641 		txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
642 	} else {
643 		txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
644 	}
645 
646 	rc = sfxge_tx_packet_add(txq, m);
647 
648 	return (rc);
649 }
650 
651 #else /* !SFXGE_HAVE_MQ */
652 
653 static void sfxge_if_start_locked(struct ifnet *ifp)
654 {
655 	struct sfxge_softc *sc = ifp->if_softc;
656 	struct sfxge_txq *txq;
657 	struct mbuf *mbuf;
658 	unsigned int pushed[SFXGE_TXQ_NTYPES];
659 	unsigned int q_index;
660 
661 	if ((ifp->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
662 	    IFF_DRV_RUNNING)
663 		return;
664 
665 	if (!sc->port.link_up)
666 		return;
667 
668 	for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) {
669 		txq = sc->txq[q_index];
670 		pushed[q_index] = txq->added;
671 	}
672 
673 	while (!IFQ_DRV_IS_EMPTY(&ifp->if_snd)) {
674 		IFQ_DRV_DEQUEUE(&ifp->if_snd, mbuf);
675 		if (mbuf == NULL)
676 			break;
677 
678 		ETHER_BPF_MTAP(ifp, mbuf); /* packet capture */
679 
680 		/* Pick the desired transmit queue. */
681 		if (mbuf->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO))
682 			q_index = SFXGE_TXQ_IP_TCP_UDP_CKSUM;
683 		else if (mbuf->m_pkthdr.csum_flags & CSUM_DELAY_IP)
684 			q_index = SFXGE_TXQ_IP_CKSUM;
685 		else
686 			q_index = SFXGE_TXQ_NON_CKSUM;
687 		txq = sc->txq[q_index];
688 
689 		if (sfxge_tx_queue_mbuf(txq, mbuf) != 0)
690 			continue;
691 
692 		if (txq->blocked) {
693 			ifp->if_drv_flags |= IFF_DRV_OACTIVE;
694 			break;
695 		}
696 
697 		/* Push the fragments to the hardware in batches. */
698 		if (txq->added - pushed[q_index] >= SFXGE_TX_BATCH) {
699 			efx_tx_qpush(txq->common, txq->added);
700 			pushed[q_index] = txq->added;
701 		}
702 	}
703 
704 	for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) {
705 		txq = sc->txq[q_index];
706 		if (txq->added != pushed[q_index])
707 			efx_tx_qpush(txq->common, txq->added);
708 	}
709 }
710 
711 void sfxge_if_start(struct ifnet *ifp)
712 {
713 	struct sfxge_softc *sc = ifp->if_softc;
714 
715 	mtx_lock(&sc->tx_lock);
716 	sfxge_if_start_locked(ifp);
717 	mtx_unlock(&sc->tx_lock);
718 }
719 
720 static inline void
721 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
722 {
723 	struct sfxge_softc *sc = txq->sc;
724 	struct ifnet *ifp = sc->ifnet;
725 
726 	mtx_assert(&sc->tx_lock, MA_OWNED);
727 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
728 	sfxge_if_start_locked(ifp);
729 	mtx_unlock(&sc->tx_lock);
730 }
731 
732 #endif /* SFXGE_HAVE_MQ */
733 
734 /*
735  * Software "TSO".  Not quite as good as doing it in hardware, but
736  * still faster than segmenting in the stack.
737  */
738 
739 struct sfxge_tso_state {
740 	/* Output position */
741 	unsigned out_len;	/* Remaining length in current segment */
742 	unsigned seqnum;	/* Current sequence number */
743 	unsigned packet_space;	/* Remaining space in current packet */
744 
745 	/* Input position */
746 	unsigned dma_seg_i;	/* Current DMA segment number */
747 	uint64_t dma_addr;	/* DMA address of current position */
748 	unsigned in_len;	/* Remaining length in current mbuf */
749 
750 	const struct mbuf *mbuf; /* Input mbuf (head of chain) */
751 	u_short protocol;	/* Network protocol (after VLAN decap) */
752 	ssize_t nh_off;		/* Offset of network header */
753 	ssize_t tcph_off;	/* Offset of TCP header */
754 	unsigned header_len;	/* Number of bytes of header */
755 	int full_packet_size;	/* Number of bytes to put in each outgoing
756 				 * segment */
757 };
758 
759 static inline const struct ip *tso_iph(const struct sfxge_tso_state *tso)
760 {
761 	KASSERT(tso->protocol == htons(ETHERTYPE_IP),
762 		("tso_iph() in non-IPv4 state"));
763 	return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
764 }
765 static inline const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
766 {
767 	KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
768 		("tso_ip6h() in non-IPv6 state"));
769 	return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
770 }
771 static inline const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
772 {
773 	return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
774 }
775 
776 /* Size of preallocated TSO header buffers.  Larger blocks must be
777  * allocated from the heap.
778  */
779 #define	TSOH_STD_SIZE	128
780 
781 /* At most half the descriptors in the queue at any time will refer to
782  * a TSO header buffer, since they must always be followed by a
783  * payload descriptor referring to an mbuf.
784  */
785 #define	TSOH_COUNT(_txq_entries)	((_txq_entries) / 2u)
786 #define	TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
787 #define	TSOH_PAGE_COUNT(_txq_entries)	\
788 	((TSOH_COUNT(_txq_entries) + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE)
789 
790 static int tso_init(struct sfxge_txq *txq)
791 {
792 	struct sfxge_softc *sc = txq->sc;
793 	unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries);
794 	int i, rc;
795 
796 	/* Allocate TSO header buffers */
797 	txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]),
798 				  M_SFXGE, M_WAITOK);
799 
800 	for (i = 0; i < tsoh_page_count; i++) {
801 		rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
802 		if (rc != 0)
803 			goto fail;
804 	}
805 
806 	return (0);
807 
808 fail:
809 	while (i-- > 0)
810 		sfxge_dma_free(&txq->tsoh_buffer[i]);
811 	free(txq->tsoh_buffer, M_SFXGE);
812 	txq->tsoh_buffer = NULL;
813 	return (rc);
814 }
815 
816 static void tso_fini(struct sfxge_txq *txq)
817 {
818 	int i;
819 
820 	if (txq->tsoh_buffer != NULL) {
821 		for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++)
822 			sfxge_dma_free(&txq->tsoh_buffer[i]);
823 		free(txq->tsoh_buffer, M_SFXGE);
824 	}
825 }
826 
827 static void tso_start(struct sfxge_tso_state *tso, struct mbuf *mbuf)
828 {
829 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
830 
831 	tso->mbuf = mbuf;
832 
833 	/* Find network protocol and header */
834 	tso->protocol = eh->ether_type;
835 	if (tso->protocol == htons(ETHERTYPE_VLAN)) {
836 		struct ether_vlan_header *veh =
837 			mtod(mbuf, struct ether_vlan_header *);
838 		tso->protocol = veh->evl_proto;
839 		tso->nh_off = sizeof(*veh);
840 	} else {
841 		tso->nh_off = sizeof(*eh);
842 	}
843 
844 	/* Find TCP header */
845 	if (tso->protocol == htons(ETHERTYPE_IP)) {
846 		KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
847 			("TSO required on non-TCP packet"));
848 		tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
849 	} else {
850 		KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
851 			("TSO required on non-IP packet"));
852 		KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
853 			("TSO required on non-TCP packet"));
854 		tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
855 	}
856 
857 	/* We assume all headers are linear in the head mbuf */
858 	tso->header_len = tso->tcph_off + 4 * tso_tcph(tso)->th_off;
859 	KASSERT(tso->header_len <= mbuf->m_len, ("packet headers fragmented"));
860 	tso->full_packet_size = tso->header_len + mbuf->m_pkthdr.tso_segsz;
861 
862 	tso->seqnum = ntohl(tso_tcph(tso)->th_seq);
863 
864 	/* These flags must not be duplicated */
865 	KASSERT(!(tso_tcph(tso)->th_flags & (TH_URG | TH_SYN | TH_RST)),
866 		("incompatible TCP flag on TSO packet"));
867 
868 	tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
869 }
870 
871 /*
872  * tso_fill_packet_with_fragment - form descriptors for the current fragment
873  *
874  * Form descriptors for the current fragment, until we reach the end
875  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
876  * space.
877  */
878 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
879 					  struct sfxge_tso_state *tso)
880 {
881 	efx_buffer_t *desc;
882 	int n;
883 
884 	if (tso->in_len == 0 || tso->packet_space == 0)
885 		return;
886 
887 	KASSERT(tso->in_len > 0, ("TSO input length went negative"));
888 	KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
889 
890 	n = min(tso->in_len, tso->packet_space);
891 
892 	tso->packet_space -= n;
893 	tso->out_len -= n;
894 	tso->in_len -= n;
895 
896 	desc = &txq->pend_desc[txq->n_pend_desc++];
897 	desc->eb_addr = tso->dma_addr;
898 	desc->eb_size = n;
899 	desc->eb_eop = tso->out_len == 0 || tso->packet_space == 0;
900 
901 	tso->dma_addr += n;
902 }
903 
904 /* Callback from bus_dmamap_load() for long TSO headers. */
905 static void tso_map_long_header(void *dma_addr_ret,
906 				bus_dma_segment_t *segs, int nseg,
907 				int error)
908 {
909 	*(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
910 				      __predict_true(nseg == 1)) ?
911 				     segs->ds_addr : 0);
912 }
913 
914 /*
915  * tso_start_new_packet - generate a new header and prepare for the new packet
916  *
917  * Generate a new header and prepare for the new packet.  Return 0 on
918  * success, or an error code if failed to alloc header.
919  */
920 static int tso_start_new_packet(struct sfxge_txq *txq,
921 				struct sfxge_tso_state *tso,
922 				unsigned int id)
923 {
924 	struct sfxge_tx_mapping *stmp = &txq->stmp[id];
925 	struct tcphdr *tsoh_th;
926 	unsigned ip_length;
927 	caddr_t header;
928 	uint64_t dma_addr;
929 	bus_dmamap_t map;
930 	efx_buffer_t *desc;
931 	int rc;
932 
933 	/* Allocate a DMA-mapped header buffer. */
934 	if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
935 		unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
936 		unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
937 
938 		header = (txq->tsoh_buffer[page_index].esm_base +
939 			  buf_index * TSOH_STD_SIZE);
940 		dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
941 			    buf_index * TSOH_STD_SIZE);
942 		map = txq->tsoh_buffer[page_index].esm_map;
943 
944 		stmp->flags = 0;
945 	} else {
946 		/* We cannot use bus_dmamem_alloc() as that may sleep */
947 		header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
948 		if (__predict_false(!header))
949 			return (ENOMEM);
950 		rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
951 				     header, tso->header_len,
952 				     tso_map_long_header, &dma_addr,
953 				     BUS_DMA_NOWAIT);
954 		if (__predict_false(dma_addr == 0)) {
955 			if (rc == 0) {
956 				/* Succeeded but got >1 segment */
957 				bus_dmamap_unload(txq->packet_dma_tag,
958 						  stmp->map);
959 				rc = EINVAL;
960 			}
961 			free(header, M_SFXGE);
962 			return (rc);
963 		}
964 		map = stmp->map;
965 
966 		txq->tso_long_headers++;
967 		stmp->u.heap_buf = header;
968 		stmp->flags = TX_BUF_UNMAP;
969 	}
970 
971 	tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
972 
973 	/* Copy and update the headers. */
974 	memcpy(header, tso->mbuf->m_data, tso->header_len);
975 
976 	tsoh_th->th_seq = htonl(tso->seqnum);
977 	tso->seqnum += tso->mbuf->m_pkthdr.tso_segsz;
978 	if (tso->out_len > tso->mbuf->m_pkthdr.tso_segsz) {
979 		/* This packet will not finish the TSO burst. */
980 		ip_length = tso->full_packet_size - tso->nh_off;
981 		tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH);
982 	} else {
983 		/* This packet will be the last in the TSO burst. */
984 		ip_length = tso->header_len - tso->nh_off + tso->out_len;
985 	}
986 
987 	if (tso->protocol == htons(ETHERTYPE_IP)) {
988 		struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
989 		tsoh_iph->ip_len = htons(ip_length);
990 		/* XXX We should increment ip_id, but FreeBSD doesn't
991 		 * currently allocate extra IDs for multiple segments.
992 		 */
993 	} else {
994 		struct ip6_hdr *tsoh_iph =
995 			(struct ip6_hdr *)(header + tso->nh_off);
996 		tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
997 	}
998 
999 	/* Make the header visible to the hardware. */
1000 	bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
1001 
1002 	tso->packet_space = tso->mbuf->m_pkthdr.tso_segsz;
1003 	txq->tso_packets++;
1004 
1005 	/* Form a descriptor for this header. */
1006 	desc = &txq->pend_desc[txq->n_pend_desc++];
1007 	desc->eb_addr = dma_addr;
1008 	desc->eb_size = tso->header_len;
1009 	desc->eb_eop = 0;
1010 
1011 	return (0);
1012 }
1013 
1014 static int
1015 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
1016 		   const bus_dma_segment_t *dma_seg, int n_dma_seg)
1017 {
1018 	struct sfxge_tso_state tso;
1019 	unsigned int id, next_id;
1020 
1021 	tso_start(&tso, mbuf);
1022 
1023 	/* Grab the first payload fragment. */
1024 	if (dma_seg->ds_len == tso.header_len) {
1025 		--n_dma_seg;
1026 		KASSERT(n_dma_seg, ("no payload found in TSO packet"));
1027 		++dma_seg;
1028 		tso.in_len = dma_seg->ds_len;
1029 		tso.dma_addr = dma_seg->ds_addr;
1030 	} else {
1031 		tso.in_len = dma_seg->ds_len - tso.header_len;
1032 		tso.dma_addr = dma_seg->ds_addr + tso.header_len;
1033 	}
1034 
1035 	id = txq->added & txq->ptr_mask;
1036 	if (__predict_false(tso_start_new_packet(txq, &tso, id)))
1037 		return (-1);
1038 
1039 	while (1) {
1040 		id = (id + 1) & txq->ptr_mask;
1041 		tso_fill_packet_with_fragment(txq, &tso);
1042 
1043 		/* Move onto the next fragment? */
1044 		if (tso.in_len == 0) {
1045 			--n_dma_seg;
1046 			if (n_dma_seg == 0)
1047 				break;
1048 			++dma_seg;
1049 			tso.in_len = dma_seg->ds_len;
1050 			tso.dma_addr = dma_seg->ds_addr;
1051 		}
1052 
1053 		/* End of packet? */
1054 		if (tso.packet_space == 0) {
1055 			/* If the queue is now full due to tiny MSS,
1056 			 * or we can't create another header, discard
1057 			 * the remainder of the input mbuf but do not
1058 			 * roll back the work we have done.
1059 			 */
1060 			if (txq->n_pend_desc >
1061 			    SFXGE_TSO_MAX_DESC - (1 + SFXGE_TX_MAPPING_MAX_SEG))
1062 				break;
1063 			next_id = (id + 1) & txq->ptr_mask;
1064 			if (__predict_false(tso_start_new_packet(txq, &tso,
1065 								 next_id)))
1066 				break;
1067 			id = next_id;
1068 		}
1069 	}
1070 
1071 	txq->tso_bursts++;
1072 	return (id);
1073 }
1074 
1075 static void
1076 sfxge_tx_qunblock(struct sfxge_txq *txq)
1077 {
1078 	struct sfxge_softc *sc;
1079 	struct sfxge_evq *evq;
1080 
1081 	sc = txq->sc;
1082 	evq = sc->evq[txq->evq_index];
1083 
1084 	mtx_assert(&evq->lock, MA_OWNED);
1085 
1086 	if (txq->init_state != SFXGE_TXQ_STARTED)
1087 		return;
1088 
1089 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1090 
1091 	if (txq->blocked) {
1092 		unsigned int level;
1093 
1094 		level = txq->added - txq->completed;
1095 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
1096 			txq->blocked = 0;
1097 	}
1098 
1099 	sfxge_tx_qdpl_service(txq);
1100 	/* note: lock has been dropped */
1101 }
1102 
1103 void
1104 sfxge_tx_qflush_done(struct sfxge_txq *txq)
1105 {
1106 
1107 	txq->flush_state = SFXGE_FLUSH_DONE;
1108 }
1109 
1110 static void
1111 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1112 {
1113 	struct sfxge_txq *txq;
1114 	struct sfxge_evq *evq;
1115 	unsigned int count;
1116 
1117 	txq = sc->txq[index];
1118 	evq = sc->evq[txq->evq_index];
1119 
1120 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1121 
1122 	KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1123 	    ("txq->init_state != SFXGE_TXQ_STARTED"));
1124 
1125 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1126 	txq->flush_state = SFXGE_FLUSH_PENDING;
1127 
1128 	/* Flush the transmit queue. */
1129 	efx_tx_qflush(txq->common);
1130 
1131 	mtx_unlock(SFXGE_TXQ_LOCK(txq));
1132 
1133 	count = 0;
1134 	do {
1135 		/* Spin for 100ms. */
1136 		DELAY(100000);
1137 
1138 		if (txq->flush_state != SFXGE_FLUSH_PENDING)
1139 			break;
1140 	} while (++count < 20);
1141 
1142 	mtx_lock(&evq->lock);
1143 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1144 
1145 	KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1146 	    ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1147 
1148 	txq->flush_state = SFXGE_FLUSH_DONE;
1149 
1150 	txq->blocked = 0;
1151 	txq->pending = txq->added;
1152 
1153 	sfxge_tx_qcomplete(txq);
1154 	KASSERT(txq->completed == txq->added,
1155 	    ("txq->completed != txq->added"));
1156 
1157 	sfxge_tx_qreap(txq);
1158 	KASSERT(txq->reaped == txq->completed,
1159 	    ("txq->reaped != txq->completed"));
1160 
1161 	txq->added = 0;
1162 	txq->pending = 0;
1163 	txq->completed = 0;
1164 	txq->reaped = 0;
1165 
1166 	/* Destroy the common code transmit queue. */
1167 	efx_tx_qdestroy(txq->common);
1168 	txq->common = NULL;
1169 
1170 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1171 	    EFX_TXQ_NBUFS(sc->txq_entries));
1172 
1173 	mtx_unlock(&evq->lock);
1174 	mtx_unlock(SFXGE_TXQ_LOCK(txq));
1175 }
1176 
1177 static int
1178 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1179 {
1180 	struct sfxge_txq *txq;
1181 	efsys_mem_t *esmp;
1182 	uint16_t flags;
1183 	struct sfxge_evq *evq;
1184 	int rc;
1185 
1186 	txq = sc->txq[index];
1187 	esmp = &txq->mem;
1188 	evq = sc->evq[txq->evq_index];
1189 
1190 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1191 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1192 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1193 	    ("evq->init_state != SFXGE_EVQ_STARTED"));
1194 
1195 	/* Program the buffer table. */
1196 	if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1197 	    EFX_TXQ_NBUFS(sc->txq_entries))) != 0)
1198 		return (rc);
1199 
1200 	/* Determine the kind of queue we are creating. */
1201 	switch (txq->type) {
1202 	case SFXGE_TXQ_NON_CKSUM:
1203 		flags = 0;
1204 		break;
1205 	case SFXGE_TXQ_IP_CKSUM:
1206 		flags = EFX_CKSUM_IPV4;
1207 		break;
1208 	case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1209 		flags = EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP;
1210 		break;
1211 	default:
1212 		KASSERT(0, ("Impossible TX queue"));
1213 		flags = 0;
1214 		break;
1215 	}
1216 
1217 	/* Create the common code transmit queue. */
1218 	if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1219 	    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1220 	    &txq->common)) != 0)
1221 		goto fail;
1222 
1223 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1224 
1225 	/* Enable the transmit queue. */
1226 	efx_tx_qenable(txq->common);
1227 
1228 	txq->init_state = SFXGE_TXQ_STARTED;
1229 
1230 	mtx_unlock(SFXGE_TXQ_LOCK(txq));
1231 
1232 	return (0);
1233 
1234 fail:
1235 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1236 	    EFX_TXQ_NBUFS(sc->txq_entries));
1237 	return (rc);
1238 }
1239 
1240 void
1241 sfxge_tx_stop(struct sfxge_softc *sc)
1242 {
1243 	const efx_nic_cfg_t *encp;
1244 	int index;
1245 
1246 	index = SFXGE_TX_SCALE(sc);
1247 	while (--index >= 0)
1248 		sfxge_tx_qstop(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1249 
1250 	sfxge_tx_qstop(sc, SFXGE_TXQ_IP_CKSUM);
1251 
1252 	encp = efx_nic_cfg_get(sc->enp);
1253 	sfxge_tx_qstop(sc, SFXGE_TXQ_NON_CKSUM);
1254 
1255 	/* Tear down the transmit module */
1256 	efx_tx_fini(sc->enp);
1257 }
1258 
1259 int
1260 sfxge_tx_start(struct sfxge_softc *sc)
1261 {
1262 	int index;
1263 	int rc;
1264 
1265 	/* Initialize the common code transmit module. */
1266 	if ((rc = efx_tx_init(sc->enp)) != 0)
1267 		return (rc);
1268 
1269 	if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_NON_CKSUM)) != 0)
1270 		goto fail;
1271 
1272 	if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_IP_CKSUM)) != 0)
1273 		goto fail2;
1274 
1275 	for (index = 0; index < SFXGE_TX_SCALE(sc); index++) {
1276 		if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM +
1277 		    index)) != 0)
1278 			goto fail3;
1279 	}
1280 
1281 	return (0);
1282 
1283 fail3:
1284 	while (--index >= 0)
1285 		sfxge_tx_qstop(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1286 
1287 	sfxge_tx_qstop(sc, SFXGE_TXQ_IP_CKSUM);
1288 
1289 fail2:
1290 	sfxge_tx_qstop(sc, SFXGE_TXQ_NON_CKSUM);
1291 
1292 fail:
1293 	efx_tx_fini(sc->enp);
1294 
1295 	return (rc);
1296 }
1297 
1298 /**
1299  * Destroy a transmit queue.
1300  */
1301 static void
1302 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1303 {
1304 	struct sfxge_txq *txq;
1305 	unsigned int nmaps;
1306 
1307 	txq = sc->txq[index];
1308 
1309 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1310 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1311 
1312 	if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1313 		tso_fini(txq);
1314 
1315 	/* Free the context arrays. */
1316 	free(txq->pend_desc, M_SFXGE);
1317 	nmaps = sc->txq_entries;
1318 	while (nmaps-- != 0)
1319 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1320 	free(txq->stmp, M_SFXGE);
1321 
1322 	/* Release DMA memory mapping. */
1323 	sfxge_dma_free(&txq->mem);
1324 
1325 	sc->txq[index] = NULL;
1326 
1327 #ifdef SFXGE_HAVE_MQ
1328 	mtx_destroy(&txq->lock);
1329 #endif
1330 
1331 	free(txq, M_SFXGE);
1332 }
1333 
1334 static int
1335 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1336     enum sfxge_txq_type type, unsigned int evq_index)
1337 {
1338 	char name[16];
1339 	struct sysctl_oid *txq_node;
1340 	struct sfxge_txq *txq;
1341 	struct sfxge_evq *evq;
1342 #ifdef SFXGE_HAVE_MQ
1343 	struct sfxge_tx_dpl *stdp;
1344 #endif
1345 	efsys_mem_t *esmp;
1346 	unsigned int nmaps;
1347 	int rc;
1348 
1349 	txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1350 	txq->sc = sc;
1351 	txq->entries = sc->txq_entries;
1352 	txq->ptr_mask = txq->entries - 1;
1353 
1354 	sc->txq[txq_index] = txq;
1355 	esmp = &txq->mem;
1356 
1357 	evq = sc->evq[evq_index];
1358 
1359 	/* Allocate and zero DMA space for the descriptor ring. */
1360 	if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0)
1361 		return (rc);
1362 	(void)memset(esmp->esm_base, 0, EFX_TXQ_SIZE(sc->txq_entries));
1363 
1364 	/* Allocate buffer table entries. */
1365 	sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries),
1366 				 &txq->buf_base_id);
1367 
1368 	/* Create a DMA tag for packet mappings. */
1369 	if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000,
1370 	    MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1371 	    NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL,
1372 	    &txq->packet_dma_tag) != 0) {
1373 		device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1374 		rc = ENOMEM;
1375 		goto fail;
1376 	}
1377 
1378 	/* Allocate pending descriptor array for batching writes. */
1379 	txq->pend_desc = malloc(sizeof(efx_buffer_t) * sc->txq_entries,
1380 				M_SFXGE, M_ZERO | M_WAITOK);
1381 
1382 	/* Allocate and initialise mbuf DMA mapping array. */
1383 	txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries,
1384 	    M_SFXGE, M_ZERO | M_WAITOK);
1385 	for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) {
1386 		rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1387 				       &txq->stmp[nmaps].map);
1388 		if (rc != 0)
1389 			goto fail2;
1390 	}
1391 
1392 	snprintf(name, sizeof(name), "%u", txq_index);
1393 	txq_node = SYSCTL_ADD_NODE(
1394 		device_get_sysctl_ctx(sc->dev),
1395 		SYSCTL_CHILDREN(sc->txqs_node),
1396 		OID_AUTO, name, CTLFLAG_RD, NULL, "");
1397 	if (txq_node == NULL) {
1398 		rc = ENOMEM;
1399 		goto fail_txq_node;
1400 	}
1401 
1402 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1403 	    (rc = tso_init(txq)) != 0)
1404 		goto fail3;
1405 
1406 #ifdef SFXGE_HAVE_MQ
1407 	if (sfxge_tx_dpl_get_max <= 0) {
1408 		log(LOG_ERR, "%s=%d must be greater than 0",
1409 		    SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max);
1410 		rc = EINVAL;
1411 		goto fail_tx_dpl_get_max;
1412 	}
1413 	if (sfxge_tx_dpl_put_max < 0) {
1414 		log(LOG_ERR, "%s=%d must be greater or equal to 0",
1415 		    SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max);
1416 		rc = EINVAL;
1417 		goto fail_tx_dpl_put_max;
1418 	}
1419 
1420 	/* Initialize the deferred packet list. */
1421 	stdp = &txq->dpl;
1422 	stdp->std_put_max = sfxge_tx_dpl_put_max;
1423 	stdp->std_get_max = sfxge_tx_dpl_get_max;
1424 	stdp->std_getp = &stdp->std_get;
1425 
1426 	mtx_init(&txq->lock, "txq", NULL, MTX_DEF);
1427 
1428 	SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev),
1429 			SYSCTL_CHILDREN(txq_node), OID_AUTO,
1430 			"dpl_get_count", CTLFLAG_RD | CTLFLAG_STATS,
1431 			&stdp->std_get_count, 0, "");
1432 #endif
1433 
1434 	txq->type = type;
1435 	txq->evq_index = evq_index;
1436 	txq->txq_index = txq_index;
1437 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1438 
1439 	return (0);
1440 
1441 fail_tx_dpl_put_max:
1442 fail_tx_dpl_get_max:
1443 fail3:
1444 fail_txq_node:
1445 	free(txq->pend_desc, M_SFXGE);
1446 fail2:
1447 	while (nmaps-- != 0)
1448 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1449 	free(txq->stmp, M_SFXGE);
1450 	bus_dma_tag_destroy(txq->packet_dma_tag);
1451 
1452 fail:
1453 	sfxge_dma_free(esmp);
1454 
1455 	return (rc);
1456 }
1457 
1458 static const struct {
1459 	const char *name;
1460 	size_t offset;
1461 } sfxge_tx_stats[] = {
1462 #define	SFXGE_TX_STAT(name, member) \
1463 	{ #name, offsetof(struct sfxge_txq, member) }
1464 	SFXGE_TX_STAT(tso_bursts, tso_bursts),
1465 	SFXGE_TX_STAT(tso_packets, tso_packets),
1466 	SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
1467 	SFXGE_TX_STAT(tx_collapses, collapses),
1468 	SFXGE_TX_STAT(tx_drops, drops),
1469 	SFXGE_TX_STAT(tx_early_drops, early_drops),
1470 };
1471 
1472 static int
1473 sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1474 {
1475 	struct sfxge_softc *sc = arg1;
1476 	unsigned int id = arg2;
1477 	unsigned long sum;
1478 	unsigned int index;
1479 
1480 	/* Sum across all TX queues */
1481 	sum = 0;
1482 	for (index = 0;
1483 	     index < SFXGE_TXQ_IP_TCP_UDP_CKSUM + SFXGE_TX_SCALE(sc);
1484 	     index++)
1485 		sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1486 					  sfxge_tx_stats[id].offset);
1487 
1488 	return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1489 }
1490 
1491 static void
1492 sfxge_tx_stat_init(struct sfxge_softc *sc)
1493 {
1494 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1495 	struct sysctl_oid_list *stat_list;
1496 	unsigned int id;
1497 
1498 	stat_list = SYSCTL_CHILDREN(sc->stats_node);
1499 
1500 	for (id = 0;
1501 	     id < sizeof(sfxge_tx_stats) / sizeof(sfxge_tx_stats[0]);
1502 	     id++) {
1503 		SYSCTL_ADD_PROC(
1504 			ctx, stat_list,
1505 			OID_AUTO, sfxge_tx_stats[id].name,
1506 			CTLTYPE_ULONG|CTLFLAG_RD,
1507 			sc, id, sfxge_tx_stat_handler, "LU",
1508 			"");
1509 	}
1510 }
1511 
1512 void
1513 sfxge_tx_fini(struct sfxge_softc *sc)
1514 {
1515 	int index;
1516 
1517 	index = SFXGE_TX_SCALE(sc);
1518 	while (--index >= 0)
1519 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1520 
1521 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1522 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1523 }
1524 
1525 
1526 int
1527 sfxge_tx_init(struct sfxge_softc *sc)
1528 {
1529 	struct sfxge_intr *intr;
1530 	int index;
1531 	int rc;
1532 
1533 	intr = &sc->intr;
1534 
1535 	KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1536 	    ("intr->state != SFXGE_INTR_INITIALIZED"));
1537 
1538 	sc->txqs_node = SYSCTL_ADD_NODE(
1539 		device_get_sysctl_ctx(sc->dev),
1540 		SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
1541 		OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues");
1542 	if (sc->txqs_node == NULL) {
1543 		rc = ENOMEM;
1544 		goto fail_txq_node;
1545 	}
1546 
1547 	/* Initialize the transmit queues */
1548 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
1549 	    SFXGE_TXQ_NON_CKSUM, 0)) != 0)
1550 		goto fail;
1551 
1552 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
1553 	    SFXGE_TXQ_IP_CKSUM, 0)) != 0)
1554 		goto fail2;
1555 
1556 	for (index = 0; index < SFXGE_TX_SCALE(sc); index++) {
1557 		if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index,
1558 		    SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
1559 			goto fail3;
1560 	}
1561 
1562 	sfxge_tx_stat_init(sc);
1563 
1564 	return (0);
1565 
1566 fail3:
1567 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1568 
1569 	while (--index >= 0)
1570 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1571 
1572 fail2:
1573 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1574 
1575 fail:
1576 fail_txq_node:
1577 	return (rc);
1578 }
1579