xref: /freebsd/sys/dev/sfxge/sfxge_tx.c (revision bc96366c864c07ef352edb92017357917c75b36c)
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 		/* check if flowid is set */
635 		if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
636 			uint32_t hash = m->m_pkthdr.flowid;
637 
638 			index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX];
639 		}
640 		txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index];
641 	} else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
642 		txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
643 	} else {
644 		txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
645 	}
646 
647 	rc = sfxge_tx_packet_add(txq, m);
648 
649 	return (rc);
650 }
651 
652 #else /* !SFXGE_HAVE_MQ */
653 
654 static void sfxge_if_start_locked(struct ifnet *ifp)
655 {
656 	struct sfxge_softc *sc = ifp->if_softc;
657 	struct sfxge_txq *txq;
658 	struct mbuf *mbuf;
659 	unsigned int pushed[SFXGE_TXQ_NTYPES];
660 	unsigned int q_index;
661 
662 	if ((ifp->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
663 	    IFF_DRV_RUNNING)
664 		return;
665 
666 	if (!sc->port.link_up)
667 		return;
668 
669 	for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) {
670 		txq = sc->txq[q_index];
671 		pushed[q_index] = txq->added;
672 	}
673 
674 	while (!IFQ_DRV_IS_EMPTY(&ifp->if_snd)) {
675 		IFQ_DRV_DEQUEUE(&ifp->if_snd, mbuf);
676 		if (mbuf == NULL)
677 			break;
678 
679 		ETHER_BPF_MTAP(ifp, mbuf); /* packet capture */
680 
681 		/* Pick the desired transmit queue. */
682 		if (mbuf->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO))
683 			q_index = SFXGE_TXQ_IP_TCP_UDP_CKSUM;
684 		else if (mbuf->m_pkthdr.csum_flags & CSUM_DELAY_IP)
685 			q_index = SFXGE_TXQ_IP_CKSUM;
686 		else
687 			q_index = SFXGE_TXQ_NON_CKSUM;
688 		txq = sc->txq[q_index];
689 
690 		if (sfxge_tx_queue_mbuf(txq, mbuf) != 0)
691 			continue;
692 
693 		if (txq->blocked) {
694 			ifp->if_drv_flags |= IFF_DRV_OACTIVE;
695 			break;
696 		}
697 
698 		/* Push the fragments to the hardware in batches. */
699 		if (txq->added - pushed[q_index] >= SFXGE_TX_BATCH) {
700 			efx_tx_qpush(txq->common, txq->added);
701 			pushed[q_index] = txq->added;
702 		}
703 	}
704 
705 	for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) {
706 		txq = sc->txq[q_index];
707 		if (txq->added != pushed[q_index])
708 			efx_tx_qpush(txq->common, txq->added);
709 	}
710 }
711 
712 void sfxge_if_start(struct ifnet *ifp)
713 {
714 	struct sfxge_softc *sc = ifp->if_softc;
715 
716 	mtx_lock(&sc->tx_lock);
717 	sfxge_if_start_locked(ifp);
718 	mtx_unlock(&sc->tx_lock);
719 }
720 
721 static inline void
722 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
723 {
724 	struct sfxge_softc *sc = txq->sc;
725 	struct ifnet *ifp = sc->ifnet;
726 
727 	mtx_assert(&sc->tx_lock, MA_OWNED);
728 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
729 	sfxge_if_start_locked(ifp);
730 	mtx_unlock(&sc->tx_lock);
731 }
732 
733 #endif /* SFXGE_HAVE_MQ */
734 
735 /*
736  * Software "TSO".  Not quite as good as doing it in hardware, but
737  * still faster than segmenting in the stack.
738  */
739 
740 struct sfxge_tso_state {
741 	/* Output position */
742 	unsigned out_len;	/* Remaining length in current segment */
743 	unsigned seqnum;	/* Current sequence number */
744 	unsigned packet_space;	/* Remaining space in current packet */
745 
746 	/* Input position */
747 	unsigned dma_seg_i;	/* Current DMA segment number */
748 	uint64_t dma_addr;	/* DMA address of current position */
749 	unsigned in_len;	/* Remaining length in current mbuf */
750 
751 	const struct mbuf *mbuf; /* Input mbuf (head of chain) */
752 	u_short protocol;	/* Network protocol (after VLAN decap) */
753 	ssize_t nh_off;		/* Offset of network header */
754 	ssize_t tcph_off;	/* Offset of TCP header */
755 	unsigned header_len;	/* Number of bytes of header */
756 	int full_packet_size;	/* Number of bytes to put in each outgoing
757 				 * segment */
758 };
759 
760 static inline const struct ip *tso_iph(const struct sfxge_tso_state *tso)
761 {
762 	KASSERT(tso->protocol == htons(ETHERTYPE_IP),
763 		("tso_iph() in non-IPv4 state"));
764 	return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
765 }
766 static inline const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
767 {
768 	KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
769 		("tso_ip6h() in non-IPv6 state"));
770 	return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
771 }
772 static inline const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
773 {
774 	return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
775 }
776 
777 /* Size of preallocated TSO header buffers.  Larger blocks must be
778  * allocated from the heap.
779  */
780 #define	TSOH_STD_SIZE	128
781 
782 /* At most half the descriptors in the queue at any time will refer to
783  * a TSO header buffer, since they must always be followed by a
784  * payload descriptor referring to an mbuf.
785  */
786 #define	TSOH_COUNT(_txq_entries)	((_txq_entries) / 2u)
787 #define	TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
788 #define	TSOH_PAGE_COUNT(_txq_entries)	\
789 	((TSOH_COUNT(_txq_entries) + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE)
790 
791 static int tso_init(struct sfxge_txq *txq)
792 {
793 	struct sfxge_softc *sc = txq->sc;
794 	unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries);
795 	int i, rc;
796 
797 	/* Allocate TSO header buffers */
798 	txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]),
799 				  M_SFXGE, M_WAITOK);
800 
801 	for (i = 0; i < tsoh_page_count; i++) {
802 		rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
803 		if (rc != 0)
804 			goto fail;
805 	}
806 
807 	return (0);
808 
809 fail:
810 	while (i-- > 0)
811 		sfxge_dma_free(&txq->tsoh_buffer[i]);
812 	free(txq->tsoh_buffer, M_SFXGE);
813 	txq->tsoh_buffer = NULL;
814 	return (rc);
815 }
816 
817 static void tso_fini(struct sfxge_txq *txq)
818 {
819 	int i;
820 
821 	if (txq->tsoh_buffer != NULL) {
822 		for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++)
823 			sfxge_dma_free(&txq->tsoh_buffer[i]);
824 		free(txq->tsoh_buffer, M_SFXGE);
825 	}
826 }
827 
828 static void tso_start(struct sfxge_tso_state *tso, struct mbuf *mbuf)
829 {
830 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
831 
832 	tso->mbuf = mbuf;
833 
834 	/* Find network protocol and header */
835 	tso->protocol = eh->ether_type;
836 	if (tso->protocol == htons(ETHERTYPE_VLAN)) {
837 		struct ether_vlan_header *veh =
838 			mtod(mbuf, struct ether_vlan_header *);
839 		tso->protocol = veh->evl_proto;
840 		tso->nh_off = sizeof(*veh);
841 	} else {
842 		tso->nh_off = sizeof(*eh);
843 	}
844 
845 	/* Find TCP header */
846 	if (tso->protocol == htons(ETHERTYPE_IP)) {
847 		KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
848 			("TSO required on non-TCP packet"));
849 		tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
850 	} else {
851 		KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
852 			("TSO required on non-IP packet"));
853 		KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
854 			("TSO required on non-TCP packet"));
855 		tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
856 	}
857 
858 	/* We assume all headers are linear in the head mbuf */
859 	tso->header_len = tso->tcph_off + 4 * tso_tcph(tso)->th_off;
860 	KASSERT(tso->header_len <= mbuf->m_len, ("packet headers fragmented"));
861 	tso->full_packet_size = tso->header_len + mbuf->m_pkthdr.tso_segsz;
862 
863 	tso->seqnum = ntohl(tso_tcph(tso)->th_seq);
864 
865 	/* These flags must not be duplicated */
866 	KASSERT(!(tso_tcph(tso)->th_flags & (TH_URG | TH_SYN | TH_RST)),
867 		("incompatible TCP flag on TSO packet"));
868 
869 	tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
870 }
871 
872 /*
873  * tso_fill_packet_with_fragment - form descriptors for the current fragment
874  *
875  * Form descriptors for the current fragment, until we reach the end
876  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
877  * space.
878  */
879 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
880 					  struct sfxge_tso_state *tso)
881 {
882 	efx_buffer_t *desc;
883 	int n;
884 
885 	if (tso->in_len == 0 || tso->packet_space == 0)
886 		return;
887 
888 	KASSERT(tso->in_len > 0, ("TSO input length went negative"));
889 	KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
890 
891 	n = min(tso->in_len, tso->packet_space);
892 
893 	tso->packet_space -= n;
894 	tso->out_len -= n;
895 	tso->in_len -= n;
896 
897 	desc = &txq->pend_desc[txq->n_pend_desc++];
898 	desc->eb_addr = tso->dma_addr;
899 	desc->eb_size = n;
900 	desc->eb_eop = tso->out_len == 0 || tso->packet_space == 0;
901 
902 	tso->dma_addr += n;
903 }
904 
905 /* Callback from bus_dmamap_load() for long TSO headers. */
906 static void tso_map_long_header(void *dma_addr_ret,
907 				bus_dma_segment_t *segs, int nseg,
908 				int error)
909 {
910 	*(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
911 				      __predict_true(nseg == 1)) ?
912 				     segs->ds_addr : 0);
913 }
914 
915 /*
916  * tso_start_new_packet - generate a new header and prepare for the new packet
917  *
918  * Generate a new header and prepare for the new packet.  Return 0 on
919  * success, or an error code if failed to alloc header.
920  */
921 static int tso_start_new_packet(struct sfxge_txq *txq,
922 				struct sfxge_tso_state *tso,
923 				unsigned int id)
924 {
925 	struct sfxge_tx_mapping *stmp = &txq->stmp[id];
926 	struct tcphdr *tsoh_th;
927 	unsigned ip_length;
928 	caddr_t header;
929 	uint64_t dma_addr;
930 	bus_dmamap_t map;
931 	efx_buffer_t *desc;
932 	int rc;
933 
934 	/* Allocate a DMA-mapped header buffer. */
935 	if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
936 		unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
937 		unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
938 
939 		header = (txq->tsoh_buffer[page_index].esm_base +
940 			  buf_index * TSOH_STD_SIZE);
941 		dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
942 			    buf_index * TSOH_STD_SIZE);
943 		map = txq->tsoh_buffer[page_index].esm_map;
944 
945 		stmp->flags = 0;
946 	} else {
947 		/* We cannot use bus_dmamem_alloc() as that may sleep */
948 		header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
949 		if (__predict_false(!header))
950 			return (ENOMEM);
951 		rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
952 				     header, tso->header_len,
953 				     tso_map_long_header, &dma_addr,
954 				     BUS_DMA_NOWAIT);
955 		if (__predict_false(dma_addr == 0)) {
956 			if (rc == 0) {
957 				/* Succeeded but got >1 segment */
958 				bus_dmamap_unload(txq->packet_dma_tag,
959 						  stmp->map);
960 				rc = EINVAL;
961 			}
962 			free(header, M_SFXGE);
963 			return (rc);
964 		}
965 		map = stmp->map;
966 
967 		txq->tso_long_headers++;
968 		stmp->u.heap_buf = header;
969 		stmp->flags = TX_BUF_UNMAP;
970 	}
971 
972 	tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
973 
974 	/* Copy and update the headers. */
975 	memcpy(header, tso->mbuf->m_data, tso->header_len);
976 
977 	tsoh_th->th_seq = htonl(tso->seqnum);
978 	tso->seqnum += tso->mbuf->m_pkthdr.tso_segsz;
979 	if (tso->out_len > tso->mbuf->m_pkthdr.tso_segsz) {
980 		/* This packet will not finish the TSO burst. */
981 		ip_length = tso->full_packet_size - tso->nh_off;
982 		tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH);
983 	} else {
984 		/* This packet will be the last in the TSO burst. */
985 		ip_length = tso->header_len - tso->nh_off + tso->out_len;
986 	}
987 
988 	if (tso->protocol == htons(ETHERTYPE_IP)) {
989 		struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
990 		tsoh_iph->ip_len = htons(ip_length);
991 		/* XXX We should increment ip_id, but FreeBSD doesn't
992 		 * currently allocate extra IDs for multiple segments.
993 		 */
994 	} else {
995 		struct ip6_hdr *tsoh_iph =
996 			(struct ip6_hdr *)(header + tso->nh_off);
997 		tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
998 	}
999 
1000 	/* Make the header visible to the hardware. */
1001 	bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
1002 
1003 	tso->packet_space = tso->mbuf->m_pkthdr.tso_segsz;
1004 	txq->tso_packets++;
1005 
1006 	/* Form a descriptor for this header. */
1007 	desc = &txq->pend_desc[txq->n_pend_desc++];
1008 	desc->eb_addr = dma_addr;
1009 	desc->eb_size = tso->header_len;
1010 	desc->eb_eop = 0;
1011 
1012 	return (0);
1013 }
1014 
1015 static int
1016 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
1017 		   const bus_dma_segment_t *dma_seg, int n_dma_seg)
1018 {
1019 	struct sfxge_tso_state tso;
1020 	unsigned int id, next_id;
1021 
1022 	tso_start(&tso, mbuf);
1023 
1024 	/* Grab the first payload fragment. */
1025 	if (dma_seg->ds_len == tso.header_len) {
1026 		--n_dma_seg;
1027 		KASSERT(n_dma_seg, ("no payload found in TSO packet"));
1028 		++dma_seg;
1029 		tso.in_len = dma_seg->ds_len;
1030 		tso.dma_addr = dma_seg->ds_addr;
1031 	} else {
1032 		tso.in_len = dma_seg->ds_len - tso.header_len;
1033 		tso.dma_addr = dma_seg->ds_addr + tso.header_len;
1034 	}
1035 
1036 	id = txq->added & txq->ptr_mask;
1037 	if (__predict_false(tso_start_new_packet(txq, &tso, id)))
1038 		return (-1);
1039 
1040 	while (1) {
1041 		id = (id + 1) & txq->ptr_mask;
1042 		tso_fill_packet_with_fragment(txq, &tso);
1043 
1044 		/* Move onto the next fragment? */
1045 		if (tso.in_len == 0) {
1046 			--n_dma_seg;
1047 			if (n_dma_seg == 0)
1048 				break;
1049 			++dma_seg;
1050 			tso.in_len = dma_seg->ds_len;
1051 			tso.dma_addr = dma_seg->ds_addr;
1052 		}
1053 
1054 		/* End of packet? */
1055 		if (tso.packet_space == 0) {
1056 			/* If the queue is now full due to tiny MSS,
1057 			 * or we can't create another header, discard
1058 			 * the remainder of the input mbuf but do not
1059 			 * roll back the work we have done.
1060 			 */
1061 			if (txq->n_pend_desc >
1062 			    SFXGE_TSO_MAX_DESC - (1 + SFXGE_TX_MAPPING_MAX_SEG))
1063 				break;
1064 			next_id = (id + 1) & txq->ptr_mask;
1065 			if (__predict_false(tso_start_new_packet(txq, &tso,
1066 								 next_id)))
1067 				break;
1068 			id = next_id;
1069 		}
1070 	}
1071 
1072 	txq->tso_bursts++;
1073 	return (id);
1074 }
1075 
1076 static void
1077 sfxge_tx_qunblock(struct sfxge_txq *txq)
1078 {
1079 	struct sfxge_softc *sc;
1080 	struct sfxge_evq *evq;
1081 
1082 	sc = txq->sc;
1083 	evq = sc->evq[txq->evq_index];
1084 
1085 	mtx_assert(&evq->lock, MA_OWNED);
1086 
1087 	if (txq->init_state != SFXGE_TXQ_STARTED)
1088 		return;
1089 
1090 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1091 
1092 	if (txq->blocked) {
1093 		unsigned int level;
1094 
1095 		level = txq->added - txq->completed;
1096 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
1097 			txq->blocked = 0;
1098 	}
1099 
1100 	sfxge_tx_qdpl_service(txq);
1101 	/* note: lock has been dropped */
1102 }
1103 
1104 void
1105 sfxge_tx_qflush_done(struct sfxge_txq *txq)
1106 {
1107 
1108 	txq->flush_state = SFXGE_FLUSH_DONE;
1109 }
1110 
1111 static void
1112 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1113 {
1114 	struct sfxge_txq *txq;
1115 	struct sfxge_evq *evq;
1116 	unsigned int count;
1117 
1118 	txq = sc->txq[index];
1119 	evq = sc->evq[txq->evq_index];
1120 
1121 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1122 
1123 	KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1124 	    ("txq->init_state != SFXGE_TXQ_STARTED"));
1125 
1126 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1127 	txq->flush_state = SFXGE_FLUSH_PENDING;
1128 
1129 	/* Flush the transmit queue. */
1130 	efx_tx_qflush(txq->common);
1131 
1132 	mtx_unlock(SFXGE_TXQ_LOCK(txq));
1133 
1134 	count = 0;
1135 	do {
1136 		/* Spin for 100ms. */
1137 		DELAY(100000);
1138 
1139 		if (txq->flush_state != SFXGE_FLUSH_PENDING)
1140 			break;
1141 	} while (++count < 20);
1142 
1143 	mtx_lock(&evq->lock);
1144 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1145 
1146 	KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1147 	    ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1148 
1149 	txq->flush_state = SFXGE_FLUSH_DONE;
1150 
1151 	txq->blocked = 0;
1152 	txq->pending = txq->added;
1153 
1154 	sfxge_tx_qcomplete(txq);
1155 	KASSERT(txq->completed == txq->added,
1156 	    ("txq->completed != txq->added"));
1157 
1158 	sfxge_tx_qreap(txq);
1159 	KASSERT(txq->reaped == txq->completed,
1160 	    ("txq->reaped != txq->completed"));
1161 
1162 	txq->added = 0;
1163 	txq->pending = 0;
1164 	txq->completed = 0;
1165 	txq->reaped = 0;
1166 
1167 	/* Destroy the common code transmit queue. */
1168 	efx_tx_qdestroy(txq->common);
1169 	txq->common = NULL;
1170 
1171 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1172 	    EFX_TXQ_NBUFS(sc->txq_entries));
1173 
1174 	mtx_unlock(&evq->lock);
1175 	mtx_unlock(SFXGE_TXQ_LOCK(txq));
1176 }
1177 
1178 static int
1179 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1180 {
1181 	struct sfxge_txq *txq;
1182 	efsys_mem_t *esmp;
1183 	uint16_t flags;
1184 	struct sfxge_evq *evq;
1185 	int rc;
1186 
1187 	txq = sc->txq[index];
1188 	esmp = &txq->mem;
1189 	evq = sc->evq[txq->evq_index];
1190 
1191 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1192 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1193 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1194 	    ("evq->init_state != SFXGE_EVQ_STARTED"));
1195 
1196 	/* Program the buffer table. */
1197 	if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1198 	    EFX_TXQ_NBUFS(sc->txq_entries))) != 0)
1199 		return (rc);
1200 
1201 	/* Determine the kind of queue we are creating. */
1202 	switch (txq->type) {
1203 	case SFXGE_TXQ_NON_CKSUM:
1204 		flags = 0;
1205 		break;
1206 	case SFXGE_TXQ_IP_CKSUM:
1207 		flags = EFX_CKSUM_IPV4;
1208 		break;
1209 	case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1210 		flags = EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP;
1211 		break;
1212 	default:
1213 		KASSERT(0, ("Impossible TX queue"));
1214 		flags = 0;
1215 		break;
1216 	}
1217 
1218 	/* Create the common code transmit queue. */
1219 	if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1220 	    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1221 	    &txq->common)) != 0)
1222 		goto fail;
1223 
1224 	mtx_lock(SFXGE_TXQ_LOCK(txq));
1225 
1226 	/* Enable the transmit queue. */
1227 	efx_tx_qenable(txq->common);
1228 
1229 	txq->init_state = SFXGE_TXQ_STARTED;
1230 
1231 	mtx_unlock(SFXGE_TXQ_LOCK(txq));
1232 
1233 	return (0);
1234 
1235 fail:
1236 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1237 	    EFX_TXQ_NBUFS(sc->txq_entries));
1238 	return (rc);
1239 }
1240 
1241 void
1242 sfxge_tx_stop(struct sfxge_softc *sc)
1243 {
1244 	const efx_nic_cfg_t *encp;
1245 	int index;
1246 
1247 	index = SFXGE_TX_SCALE(sc);
1248 	while (--index >= 0)
1249 		sfxge_tx_qstop(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1250 
1251 	sfxge_tx_qstop(sc, SFXGE_TXQ_IP_CKSUM);
1252 
1253 	encp = efx_nic_cfg_get(sc->enp);
1254 	sfxge_tx_qstop(sc, SFXGE_TXQ_NON_CKSUM);
1255 
1256 	/* Tear down the transmit module */
1257 	efx_tx_fini(sc->enp);
1258 }
1259 
1260 int
1261 sfxge_tx_start(struct sfxge_softc *sc)
1262 {
1263 	int index;
1264 	int rc;
1265 
1266 	/* Initialize the common code transmit module. */
1267 	if ((rc = efx_tx_init(sc->enp)) != 0)
1268 		return (rc);
1269 
1270 	if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_NON_CKSUM)) != 0)
1271 		goto fail;
1272 
1273 	if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_IP_CKSUM)) != 0)
1274 		goto fail2;
1275 
1276 	for (index = 0; index < SFXGE_TX_SCALE(sc); index++) {
1277 		if ((rc = sfxge_tx_qstart(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM +
1278 		    index)) != 0)
1279 			goto fail3;
1280 	}
1281 
1282 	return (0);
1283 
1284 fail3:
1285 	while (--index >= 0)
1286 		sfxge_tx_qstop(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1287 
1288 	sfxge_tx_qstop(sc, SFXGE_TXQ_IP_CKSUM);
1289 
1290 fail2:
1291 	sfxge_tx_qstop(sc, SFXGE_TXQ_NON_CKSUM);
1292 
1293 fail:
1294 	efx_tx_fini(sc->enp);
1295 
1296 	return (rc);
1297 }
1298 
1299 /**
1300  * Destroy a transmit queue.
1301  */
1302 static void
1303 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1304 {
1305 	struct sfxge_txq *txq;
1306 	unsigned int nmaps;
1307 
1308 	txq = sc->txq[index];
1309 
1310 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1311 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1312 
1313 	if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1314 		tso_fini(txq);
1315 
1316 	/* Free the context arrays. */
1317 	free(txq->pend_desc, M_SFXGE);
1318 	nmaps = sc->txq_entries;
1319 	while (nmaps-- != 0)
1320 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1321 	free(txq->stmp, M_SFXGE);
1322 
1323 	/* Release DMA memory mapping. */
1324 	sfxge_dma_free(&txq->mem);
1325 
1326 	sc->txq[index] = NULL;
1327 
1328 #ifdef SFXGE_HAVE_MQ
1329 	mtx_destroy(&txq->lock);
1330 #endif
1331 
1332 	free(txq, M_SFXGE);
1333 }
1334 
1335 static int
1336 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1337     enum sfxge_txq_type type, unsigned int evq_index)
1338 {
1339 	char name[16];
1340 	struct sysctl_oid *txq_node;
1341 	struct sfxge_txq *txq;
1342 	struct sfxge_evq *evq;
1343 #ifdef SFXGE_HAVE_MQ
1344 	struct sfxge_tx_dpl *stdp;
1345 #endif
1346 	efsys_mem_t *esmp;
1347 	unsigned int nmaps;
1348 	int rc;
1349 
1350 	txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1351 	txq->sc = sc;
1352 	txq->entries = sc->txq_entries;
1353 	txq->ptr_mask = txq->entries - 1;
1354 
1355 	sc->txq[txq_index] = txq;
1356 	esmp = &txq->mem;
1357 
1358 	evq = sc->evq[evq_index];
1359 
1360 	/* Allocate and zero DMA space for the descriptor ring. */
1361 	if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0)
1362 		return (rc);
1363 	(void)memset(esmp->esm_base, 0, EFX_TXQ_SIZE(sc->txq_entries));
1364 
1365 	/* Allocate buffer table entries. */
1366 	sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries),
1367 				 &txq->buf_base_id);
1368 
1369 	/* Create a DMA tag for packet mappings. */
1370 	if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000,
1371 	    MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1372 	    NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL,
1373 	    &txq->packet_dma_tag) != 0) {
1374 		device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1375 		rc = ENOMEM;
1376 		goto fail;
1377 	}
1378 
1379 	/* Allocate pending descriptor array for batching writes. */
1380 	txq->pend_desc = malloc(sizeof(efx_buffer_t) * sc->txq_entries,
1381 				M_SFXGE, M_ZERO | M_WAITOK);
1382 
1383 	/* Allocate and initialise mbuf DMA mapping array. */
1384 	txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries,
1385 	    M_SFXGE, M_ZERO | M_WAITOK);
1386 	for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) {
1387 		rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1388 				       &txq->stmp[nmaps].map);
1389 		if (rc != 0)
1390 			goto fail2;
1391 	}
1392 
1393 	snprintf(name, sizeof(name), "%u", txq_index);
1394 	txq_node = SYSCTL_ADD_NODE(
1395 		device_get_sysctl_ctx(sc->dev),
1396 		SYSCTL_CHILDREN(sc->txqs_node),
1397 		OID_AUTO, name, CTLFLAG_RD, NULL, "");
1398 	if (txq_node == NULL) {
1399 		rc = ENOMEM;
1400 		goto fail_txq_node;
1401 	}
1402 
1403 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1404 	    (rc = tso_init(txq)) != 0)
1405 		goto fail3;
1406 
1407 #ifdef SFXGE_HAVE_MQ
1408 	if (sfxge_tx_dpl_get_max <= 0) {
1409 		log(LOG_ERR, "%s=%d must be greater than 0",
1410 		    SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max);
1411 		rc = EINVAL;
1412 		goto fail_tx_dpl_get_max;
1413 	}
1414 	if (sfxge_tx_dpl_put_max < 0) {
1415 		log(LOG_ERR, "%s=%d must be greater or equal to 0",
1416 		    SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max);
1417 		rc = EINVAL;
1418 		goto fail_tx_dpl_put_max;
1419 	}
1420 
1421 	/* Initialize the deferred packet list. */
1422 	stdp = &txq->dpl;
1423 	stdp->std_put_max = sfxge_tx_dpl_put_max;
1424 	stdp->std_get_max = sfxge_tx_dpl_get_max;
1425 	stdp->std_getp = &stdp->std_get;
1426 
1427 	mtx_init(&txq->lock, "txq", NULL, MTX_DEF);
1428 
1429 	SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev),
1430 			SYSCTL_CHILDREN(txq_node), OID_AUTO,
1431 			"dpl_get_count", CTLFLAG_RD | CTLFLAG_STATS,
1432 			&stdp->std_get_count, 0, "");
1433 #endif
1434 
1435 	txq->type = type;
1436 	txq->evq_index = evq_index;
1437 	txq->txq_index = txq_index;
1438 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1439 
1440 	return (0);
1441 
1442 fail_tx_dpl_put_max:
1443 fail_tx_dpl_get_max:
1444 fail3:
1445 fail_txq_node:
1446 	free(txq->pend_desc, M_SFXGE);
1447 fail2:
1448 	while (nmaps-- != 0)
1449 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1450 	free(txq->stmp, M_SFXGE);
1451 	bus_dma_tag_destroy(txq->packet_dma_tag);
1452 
1453 fail:
1454 	sfxge_dma_free(esmp);
1455 
1456 	return (rc);
1457 }
1458 
1459 static const struct {
1460 	const char *name;
1461 	size_t offset;
1462 } sfxge_tx_stats[] = {
1463 #define	SFXGE_TX_STAT(name, member) \
1464 	{ #name, offsetof(struct sfxge_txq, member) }
1465 	SFXGE_TX_STAT(tso_bursts, tso_bursts),
1466 	SFXGE_TX_STAT(tso_packets, tso_packets),
1467 	SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
1468 	SFXGE_TX_STAT(tx_collapses, collapses),
1469 	SFXGE_TX_STAT(tx_drops, drops),
1470 	SFXGE_TX_STAT(tx_early_drops, early_drops),
1471 };
1472 
1473 static int
1474 sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1475 {
1476 	struct sfxge_softc *sc = arg1;
1477 	unsigned int id = arg2;
1478 	unsigned long sum;
1479 	unsigned int index;
1480 
1481 	/* Sum across all TX queues */
1482 	sum = 0;
1483 	for (index = 0;
1484 	     index < SFXGE_TXQ_IP_TCP_UDP_CKSUM + SFXGE_TX_SCALE(sc);
1485 	     index++)
1486 		sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1487 					  sfxge_tx_stats[id].offset);
1488 
1489 	return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1490 }
1491 
1492 static void
1493 sfxge_tx_stat_init(struct sfxge_softc *sc)
1494 {
1495 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1496 	struct sysctl_oid_list *stat_list;
1497 	unsigned int id;
1498 
1499 	stat_list = SYSCTL_CHILDREN(sc->stats_node);
1500 
1501 	for (id = 0;
1502 	     id < sizeof(sfxge_tx_stats) / sizeof(sfxge_tx_stats[0]);
1503 	     id++) {
1504 		SYSCTL_ADD_PROC(
1505 			ctx, stat_list,
1506 			OID_AUTO, sfxge_tx_stats[id].name,
1507 			CTLTYPE_ULONG|CTLFLAG_RD,
1508 			sc, id, sfxge_tx_stat_handler, "LU",
1509 			"");
1510 	}
1511 }
1512 
1513 void
1514 sfxge_tx_fini(struct sfxge_softc *sc)
1515 {
1516 	int index;
1517 
1518 	index = SFXGE_TX_SCALE(sc);
1519 	while (--index >= 0)
1520 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1521 
1522 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1523 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1524 }
1525 
1526 
1527 int
1528 sfxge_tx_init(struct sfxge_softc *sc)
1529 {
1530 	struct sfxge_intr *intr;
1531 	int index;
1532 	int rc;
1533 
1534 	intr = &sc->intr;
1535 
1536 	KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1537 	    ("intr->state != SFXGE_INTR_INITIALIZED"));
1538 
1539 	sc->txqs_node = SYSCTL_ADD_NODE(
1540 		device_get_sysctl_ctx(sc->dev),
1541 		SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
1542 		OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues");
1543 	if (sc->txqs_node == NULL) {
1544 		rc = ENOMEM;
1545 		goto fail_txq_node;
1546 	}
1547 
1548 	/* Initialize the transmit queues */
1549 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
1550 	    SFXGE_TXQ_NON_CKSUM, 0)) != 0)
1551 		goto fail;
1552 
1553 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
1554 	    SFXGE_TXQ_IP_CKSUM, 0)) != 0)
1555 		goto fail2;
1556 
1557 	for (index = 0; index < SFXGE_TX_SCALE(sc); index++) {
1558 		if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index,
1559 		    SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
1560 			goto fail3;
1561 	}
1562 
1563 	sfxge_tx_stat_init(sc);
1564 
1565 	return (0);
1566 
1567 fail3:
1568 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1569 
1570 	while (--index >= 0)
1571 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1572 
1573 fail2:
1574 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1575 
1576 fail:
1577 fail_txq_node:
1578 	return (rc);
1579 }
1580