xref: /freebsd/sys/dev/sfxge/sfxge_tx.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
1 /*-
2  * Copyright (c) 2010-2016 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 are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright notice,
12  *    this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright notice,
14  *    this list of conditions and the following disclaimer in the documentation
15  *    and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
18  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
19  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
21  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
24  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
25  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
26  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
27  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28  *
29  * The views and conclusions contained in the software and documentation are
30  * those of the authors and should not be interpreted as representing official
31  * policies, either expressed or implied, of the FreeBSD Project.
32  */
33 
34 /* Theory of operation:
35  *
36  * Tx queues allocation and mapping
37  *
38  * One Tx queue with enabled checksum offload is allocated per Rx channel
39  * (event queue).  Also 2 Tx queues (one without checksum offload and one
40  * with IP checksum offload only) are allocated and bound to event queue 0.
41  * sfxge_txq_type is used as Tx queue label.
42  *
43  * So, event queue plus label mapping to Tx queue index is:
44  *	if event queue index is 0, TxQ-index = TxQ-label * [0..SFXGE_TXQ_NTYPES)
45  *	else TxQ-index = SFXGE_TXQ_NTYPES + EvQ-index - 1
46  * See sfxge_get_txq_by_label() sfxge_ev.c
47  */
48 
49 #include <sys/cdefs.h>
50 __FBSDID("$FreeBSD$");
51 
52 #include "opt_rss.h"
53 
54 #include <sys/param.h>
55 #include <sys/malloc.h>
56 #include <sys/mbuf.h>
57 #include <sys/smp.h>
58 #include <sys/socket.h>
59 #include <sys/sysctl.h>
60 #include <sys/syslog.h>
61 #include <sys/limits.h>
62 
63 #include <net/bpf.h>
64 #include <net/ethernet.h>
65 #include <net/if.h>
66 #include <net/if_vlan_var.h>
67 
68 #include <netinet/in.h>
69 #include <netinet/ip.h>
70 #include <netinet/ip6.h>
71 #include <netinet/tcp.h>
72 
73 #ifdef RSS
74 #include <net/rss_config.h>
75 #endif
76 
77 #include "common/efx.h"
78 
79 #include "sfxge.h"
80 #include "sfxge_tx.h"
81 
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 any packets in deferred packet get-list");
89 
90 #define	SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \
91 	SFXGE_PARAM(tx_dpl_get_non_tcp_max)
92 static int sfxge_tx_dpl_get_non_tcp_max =
93 	SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT;
94 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max);
95 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN,
96 	   &sfxge_tx_dpl_get_non_tcp_max, 0,
97 	   "Maximum number of non-TCP packets in deferred packet get-list");
98 
99 #define	SFXGE_PARAM_TX_DPL_PUT_MAX	SFXGE_PARAM(tx_dpl_put_max)
100 static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT;
101 TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max);
102 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN,
103 	   &sfxge_tx_dpl_put_max, 0,
104 	   "Maximum number of any packets in deferred packet put-list");
105 
106 #define	SFXGE_PARAM_TSO_FW_ASSISTED	SFXGE_PARAM(tso_fw_assisted)
107 static int sfxge_tso_fw_assisted = (SFXGE_FATSOV1 | SFXGE_FATSOV2);
108 TUNABLE_INT(SFXGE_PARAM_TSO_FW_ASSISTED, &sfxge_tso_fw_assisted);
109 SYSCTL_INT(_hw_sfxge, OID_AUTO, tso_fw_assisted, CTLFLAG_RDTUN,
110 	   &sfxge_tso_fw_assisted, 0,
111 	   "Bitmask of FW-assisted TSO allowed to use if supported by NIC firmware");
112 
113 
114 static const struct {
115 	const char *name;
116 	size_t offset;
117 } sfxge_tx_stats[] = {
118 #define	SFXGE_TX_STAT(name, member) \
119 	{ #name, offsetof(struct sfxge_txq, member) }
120 	SFXGE_TX_STAT(tso_bursts, tso_bursts),
121 	SFXGE_TX_STAT(tso_packets, tso_packets),
122 	SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
123 	SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many),
124 	SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc),
125 	SFXGE_TX_STAT(tx_collapses, collapses),
126 	SFXGE_TX_STAT(tx_drops, drops),
127 	SFXGE_TX_STAT(tx_get_overflow, get_overflow),
128 	SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow),
129 	SFXGE_TX_STAT(tx_put_overflow, put_overflow),
130 	SFXGE_TX_STAT(tx_netdown_drops, netdown_drops),
131 };
132 
133 
134 /* Forward declarations. */
135 static void sfxge_tx_qdpl_service(struct sfxge_txq *txq);
136 static void sfxge_tx_qlist_post(struct sfxge_txq *txq);
137 static void sfxge_tx_qunblock(struct sfxge_txq *txq);
138 static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
139 			      const bus_dma_segment_t *dma_seg, int n_dma_seg,
140 			      int vlan_tagged);
141 
142 static int
143 sfxge_tx_maybe_insert_tag(struct sfxge_txq *txq, struct mbuf *mbuf)
144 {
145 	uint16_t this_tag = ((mbuf->m_flags & M_VLANTAG) ?
146 			     mbuf->m_pkthdr.ether_vtag :
147 			     0);
148 
149 	if (this_tag == txq->hw_vlan_tci)
150 		return (0);
151 
152 	efx_tx_qdesc_vlantci_create(txq->common,
153 				    bswap16(this_tag),
154 				    &txq->pend_desc[0]);
155 	txq->n_pend_desc = 1;
156 	txq->hw_vlan_tci = this_tag;
157 	return (1);
158 }
159 
160 static inline void
161 sfxge_next_stmp(struct sfxge_txq *txq, struct sfxge_tx_mapping **pstmp)
162 {
163 	KASSERT((*pstmp)->flags == 0, ("stmp flags are not 0"));
164 	if (__predict_false(*pstmp ==
165 			    &txq->stmp[txq->ptr_mask]))
166 		*pstmp = &txq->stmp[0];
167 	else
168 		(*pstmp)++;
169 }
170 
171 
172 void
173 sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq)
174 {
175 	unsigned int completed;
176 
177 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
178 
179 	completed = txq->completed;
180 	while (completed != txq->pending) {
181 		struct sfxge_tx_mapping *stmp;
182 		unsigned int id;
183 
184 		id = completed++ & txq->ptr_mask;
185 
186 		stmp = &txq->stmp[id];
187 		if (stmp->flags & TX_BUF_UNMAP) {
188 			bus_dmamap_unload(txq->packet_dma_tag, stmp->map);
189 			if (stmp->flags & TX_BUF_MBUF) {
190 				struct mbuf *m = stmp->u.mbuf;
191 				do
192 					m = m_free(m);
193 				while (m != NULL);
194 			} else {
195 				free(stmp->u.heap_buf, M_SFXGE);
196 			}
197 			stmp->flags = 0;
198 		}
199 	}
200 	txq->completed = completed;
201 
202 	/* Check whether we need to unblock the queue. */
203 	mb();
204 	if (txq->blocked) {
205 		unsigned int level;
206 
207 		level = txq->added - txq->completed;
208 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
209 			sfxge_tx_qunblock(txq);
210 	}
211 }
212 
213 static unsigned int
214 sfxge_is_mbuf_non_tcp(struct mbuf *mbuf)
215 {
216 	/* Absence of TCP checksum flags does not mean that it is non-TCP
217 	 * but it should be true if user wants to achieve high throughput.
218 	 */
219 	return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP)));
220 }
221 
222 /*
223  * Reorder the put list and append it to the get list.
224  */
225 static void
226 sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq)
227 {
228 	struct sfxge_tx_dpl *stdp;
229 	struct mbuf *mbuf, *get_next, **get_tailp;
230 	volatile uintptr_t *putp;
231 	uintptr_t put;
232 	unsigned int count;
233 	unsigned int non_tcp_count;
234 
235 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
236 
237 	stdp = &txq->dpl;
238 
239 	/* Acquire the put list. */
240 	putp = &stdp->std_put;
241 	put = atomic_readandclear_ptr(putp);
242 	mbuf = (void *)put;
243 
244 	if (mbuf == NULL)
245 		return;
246 
247 	/* Reverse the put list. */
248 	get_tailp = &mbuf->m_nextpkt;
249 	get_next = NULL;
250 
251 	count = 0;
252 	non_tcp_count = 0;
253 	do {
254 		struct mbuf *put_next;
255 
256 		non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf);
257 		put_next = mbuf->m_nextpkt;
258 		mbuf->m_nextpkt = get_next;
259 		get_next = mbuf;
260 		mbuf = put_next;
261 
262 		count++;
263 	} while (mbuf != NULL);
264 
265 	if (count > stdp->std_put_hiwat)
266 		stdp->std_put_hiwat = count;
267 
268 	/* Append the reversed put list to the get list. */
269 	KASSERT(*get_tailp == NULL, ("*get_tailp != NULL"));
270 	*stdp->std_getp = get_next;
271 	stdp->std_getp = get_tailp;
272 	stdp->std_get_count += count;
273 	stdp->std_get_non_tcp_count += non_tcp_count;
274 }
275 
276 static void
277 sfxge_tx_qreap(struct sfxge_txq *txq)
278 {
279 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
280 
281 	txq->reaped = txq->completed;
282 }
283 
284 static void
285 sfxge_tx_qlist_post(struct sfxge_txq *txq)
286 {
287 	unsigned int old_added;
288 	unsigned int block_level;
289 	unsigned int level;
290 	int rc;
291 
292 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
293 
294 	KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0"));
295 	KASSERT(txq->n_pend_desc <= txq->max_pkt_desc,
296 		("txq->n_pend_desc too large"));
297 	KASSERT(!txq->blocked, ("txq->blocked"));
298 
299 	old_added = txq->added;
300 
301 	/* Post the fragment list. */
302 	rc = efx_tx_qdesc_post(txq->common, txq->pend_desc, txq->n_pend_desc,
303 			  txq->reaped, &txq->added);
304 	KASSERT(rc == 0, ("efx_tx_qdesc_post() failed"));
305 
306 	/* If efx_tx_qdesc_post() had to refragment, our information about
307 	 * buffers to free may be associated with the wrong
308 	 * descriptors.
309 	 */
310 	KASSERT(txq->added - old_added == txq->n_pend_desc,
311 		("efx_tx_qdesc_post() refragmented descriptors"));
312 
313 	level = txq->added - txq->reaped;
314 	KASSERT(level <= txq->entries, ("overfilled TX queue"));
315 
316 	/* Clear the fragment list. */
317 	txq->n_pend_desc = 0;
318 
319 	/*
320 	 * Set the block level to ensure there is space to generate a
321 	 * large number of descriptors for TSO.
322 	 */
323 	block_level = EFX_TXQ_LIMIT(txq->entries) - txq->max_pkt_desc;
324 
325 	/* Have we reached the block level? */
326 	if (level < block_level)
327 		return;
328 
329 	/* Reap, and check again */
330 	sfxge_tx_qreap(txq);
331 	level = txq->added - txq->reaped;
332 	if (level < block_level)
333 		return;
334 
335 	txq->blocked = 1;
336 
337 	/*
338 	 * Avoid a race with completion interrupt handling that could leave
339 	 * the queue blocked.
340 	 */
341 	mb();
342 	sfxge_tx_qreap(txq);
343 	level = txq->added - txq->reaped;
344 	if (level < block_level) {
345 		mb();
346 		txq->blocked = 0;
347 	}
348 }
349 
350 static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
351 {
352 	bus_dmamap_t *used_map;
353 	bus_dmamap_t map;
354 	bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG];
355 	unsigned int id;
356 	struct sfxge_tx_mapping *stmp;
357 	efx_desc_t *desc;
358 	int n_dma_seg;
359 	int rc;
360 	int i;
361 	int eop;
362 	int vlan_tagged;
363 
364 	KASSERT(!txq->blocked, ("txq->blocked"));
365 
366 #if SFXGE_TX_PARSE_EARLY
367 	/*
368 	 * If software TSO is used, we still need to copy packet header,
369 	 * even if we have already parsed it early before enqueue.
370 	 */
371 	if ((mbuf->m_pkthdr.csum_flags & CSUM_TSO) &&
372 	    (txq->tso_fw_assisted == 0))
373 		prefetch_read_many(mbuf->m_data);
374 #else
375 	/*
376 	 * Prefetch packet header since we need to parse it and extract
377 	 * IP ID, TCP sequence number and flags.
378 	 */
379 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
380 		prefetch_read_many(mbuf->m_data);
381 #endif
382 
383 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) {
384 		rc = EINTR;
385 		goto reject;
386 	}
387 
388 	/* Load the packet for DMA. */
389 	id = txq->added & txq->ptr_mask;
390 	stmp = &txq->stmp[id];
391 	rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map,
392 				     mbuf, dma_seg, &n_dma_seg, 0);
393 	if (rc == EFBIG) {
394 		/* Try again. */
395 		struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT,
396 						   SFXGE_TX_MAPPING_MAX_SEG);
397 		if (new_mbuf == NULL)
398 			goto reject;
399 		++txq->collapses;
400 		mbuf = new_mbuf;
401 		rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag,
402 					     stmp->map, mbuf,
403 					     dma_seg, &n_dma_seg, 0);
404 	}
405 	if (rc != 0)
406 		goto reject;
407 
408 	/* Make the packet visible to the hardware. */
409 	bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE);
410 
411 	used_map = &stmp->map;
412 
413 	vlan_tagged = sfxge_tx_maybe_insert_tag(txq, mbuf);
414 	if (vlan_tagged) {
415 		sfxge_next_stmp(txq, &stmp);
416 	}
417 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
418 		rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg, vlan_tagged);
419 		if (rc < 0)
420 			goto reject_mapped;
421 		stmp = &txq->stmp[(rc - 1) & txq->ptr_mask];
422 	} else {
423 		/* Add the mapping to the fragment list, and set flags
424 		 * for the buffer.
425 		 */
426 
427 		i = 0;
428 		for (;;) {
429 			desc = &txq->pend_desc[i + vlan_tagged];
430 			eop = (i == n_dma_seg - 1);
431 			efx_tx_qdesc_dma_create(txq->common,
432 						dma_seg[i].ds_addr,
433 						dma_seg[i].ds_len,
434 						eop,
435 						desc);
436 			if (eop)
437 				break;
438 			i++;
439 			sfxge_next_stmp(txq, &stmp);
440 		}
441 		txq->n_pend_desc = n_dma_seg + vlan_tagged;
442 	}
443 
444 	/*
445 	 * If the mapping required more than one descriptor
446 	 * then we need to associate the DMA map with the last
447 	 * descriptor, not the first.
448 	 */
449 	if (used_map != &stmp->map) {
450 		map = stmp->map;
451 		stmp->map = *used_map;
452 		*used_map = map;
453 	}
454 
455 	stmp->u.mbuf = mbuf;
456 	stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF;
457 
458 	/* Post the fragment list. */
459 	sfxge_tx_qlist_post(txq);
460 
461 	return (0);
462 
463 reject_mapped:
464 	bus_dmamap_unload(txq->packet_dma_tag, *used_map);
465 reject:
466 	/* Drop the packet on the floor. */
467 	m_freem(mbuf);
468 	++txq->drops;
469 
470 	return (rc);
471 }
472 
473 /*
474  * Drain the deferred packet list into the transmit queue.
475  */
476 static void
477 sfxge_tx_qdpl_drain(struct sfxge_txq *txq)
478 {
479 	struct sfxge_softc *sc;
480 	struct sfxge_tx_dpl *stdp;
481 	struct mbuf *mbuf, *next;
482 	unsigned int count;
483 	unsigned int non_tcp_count;
484 	unsigned int pushed;
485 	int rc;
486 
487 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
488 
489 	sc = txq->sc;
490 	stdp = &txq->dpl;
491 	pushed = txq->added;
492 
493 	if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) {
494 		prefetch_read_many(sc->enp);
495 		prefetch_read_many(txq->common);
496 	}
497 
498 	mbuf = stdp->std_get;
499 	count = stdp->std_get_count;
500 	non_tcp_count = stdp->std_get_non_tcp_count;
501 
502 	if (count > stdp->std_get_hiwat)
503 		stdp->std_get_hiwat = count;
504 
505 	while (count != 0) {
506 		KASSERT(mbuf != NULL, ("mbuf == NULL"));
507 
508 		next = mbuf->m_nextpkt;
509 		mbuf->m_nextpkt = NULL;
510 
511 		ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */
512 
513 		if (next != NULL)
514 			prefetch_read_many(next);
515 
516 		rc = sfxge_tx_queue_mbuf(txq, mbuf);
517 		--count;
518 		non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf);
519 		mbuf = next;
520 		if (rc != 0)
521 			continue;
522 
523 		if (txq->blocked)
524 			break;
525 
526 		/* Push the fragments to the hardware in batches. */
527 		if (txq->added - pushed >= SFXGE_TX_BATCH) {
528 			efx_tx_qpush(txq->common, txq->added, pushed);
529 			pushed = txq->added;
530 		}
531 	}
532 
533 	if (count == 0) {
534 		KASSERT(mbuf == NULL, ("mbuf != NULL"));
535 		KASSERT(non_tcp_count == 0,
536 			("inconsistent TCP/non-TCP detection"));
537 		stdp->std_get = NULL;
538 		stdp->std_get_count = 0;
539 		stdp->std_get_non_tcp_count = 0;
540 		stdp->std_getp = &stdp->std_get;
541 	} else {
542 		stdp->std_get = mbuf;
543 		stdp->std_get_count = count;
544 		stdp->std_get_non_tcp_count = non_tcp_count;
545 	}
546 
547 	if (txq->added != pushed)
548 		efx_tx_qpush(txq->common, txq->added, pushed);
549 
550 	KASSERT(txq->blocked || stdp->std_get_count == 0,
551 		("queue unblocked but count is non-zero"));
552 }
553 
554 #define	SFXGE_TX_QDPL_PENDING(_txq)	((_txq)->dpl.std_put != 0)
555 
556 /*
557  * Service the deferred packet list.
558  *
559  * NOTE: drops the txq mutex!
560  */
561 static void
562 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
563 {
564 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
565 
566 	do {
567 		if (SFXGE_TX_QDPL_PENDING(txq))
568 			sfxge_tx_qdpl_swizzle(txq);
569 
570 		if (!txq->blocked)
571 			sfxge_tx_qdpl_drain(txq);
572 
573 		SFXGE_TXQ_UNLOCK(txq);
574 	} while (SFXGE_TX_QDPL_PENDING(txq) &&
575 		 SFXGE_TXQ_TRYLOCK(txq));
576 }
577 
578 /*
579  * Put a packet on the deferred packet get-list.
580  */
581 static int
582 sfxge_tx_qdpl_put_locked(struct sfxge_txq *txq, struct mbuf *mbuf)
583 {
584 	struct sfxge_tx_dpl *stdp;
585 
586 	stdp = &txq->dpl;
587 
588 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
589 
590 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
591 
592 	if (stdp->std_get_count >= stdp->std_get_max) {
593 		txq->get_overflow++;
594 		return (ENOBUFS);
595 	}
596 	if (sfxge_is_mbuf_non_tcp(mbuf)) {
597 		if (stdp->std_get_non_tcp_count >=
598 		    stdp->std_get_non_tcp_max) {
599 			txq->get_non_tcp_overflow++;
600 			return (ENOBUFS);
601 		}
602 		stdp->std_get_non_tcp_count++;
603 	}
604 
605 	*(stdp->std_getp) = mbuf;
606 	stdp->std_getp = &mbuf->m_nextpkt;
607 	stdp->std_get_count++;
608 
609 	return (0);
610 }
611 
612 /*
613  * Put a packet on the deferred packet put-list.
614  *
615  * We overload the csum_data field in the mbuf to keep track of this length
616  * because there is no cheap alternative to avoid races.
617  */
618 static int
619 sfxge_tx_qdpl_put_unlocked(struct sfxge_txq *txq, struct mbuf *mbuf)
620 {
621 	struct sfxge_tx_dpl *stdp;
622 	volatile uintptr_t *putp;
623 	uintptr_t old;
624 	uintptr_t new;
625 	unsigned int put_count;
626 
627 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
628 
629 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
630 
631 	stdp = &txq->dpl;
632 	putp = &stdp->std_put;
633 	new = (uintptr_t)mbuf;
634 
635 	do {
636 		old = *putp;
637 		if (old != 0) {
638 			struct mbuf *mp = (struct mbuf *)old;
639 			put_count = mp->m_pkthdr.csum_data;
640 		} else
641 			put_count = 0;
642 		if (put_count >= stdp->std_put_max) {
643 			atomic_add_long(&txq->put_overflow, 1);
644 			return (ENOBUFS);
645 		}
646 		mbuf->m_pkthdr.csum_data = put_count + 1;
647 		mbuf->m_nextpkt = (void *)old;
648 	} while (atomic_cmpset_ptr(putp, old, new) == 0);
649 
650 	return (0);
651 }
652 
653 /*
654  * Called from if_transmit - will try to grab the txq lock and enqueue to the
655  * put list if it succeeds, otherwise try to push onto the defer list if space.
656  */
657 static int
658 sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m)
659 {
660 	int rc;
661 
662 	if (!SFXGE_LINK_UP(txq->sc)) {
663 		atomic_add_long(&txq->netdown_drops, 1);
664 		return (ENETDOWN);
665 	}
666 
667 	/*
668 	 * Try to grab the txq lock.  If we are able to get the lock,
669 	 * the packet will be appended to the "get list" of the deferred
670 	 * packet list.  Otherwise, it will be pushed on the "put list".
671 	 */
672 	if (SFXGE_TXQ_TRYLOCK(txq)) {
673 		/* First swizzle put-list to get-list to keep order */
674 		sfxge_tx_qdpl_swizzle(txq);
675 
676 		rc = sfxge_tx_qdpl_put_locked(txq, m);
677 
678 		/* Try to service the list. */
679 		sfxge_tx_qdpl_service(txq);
680 		/* Lock has been dropped. */
681 	} else {
682 		rc = sfxge_tx_qdpl_put_unlocked(txq, m);
683 
684 		/*
685 		 * Try to grab the lock again.
686 		 *
687 		 * If we are able to get the lock, we need to process
688 		 * the deferred packet list.  If we are not able to get
689 		 * the lock, another thread is processing the list.
690 		 */
691 		if ((rc == 0) && SFXGE_TXQ_TRYLOCK(txq)) {
692 			sfxge_tx_qdpl_service(txq);
693 			/* Lock has been dropped. */
694 		}
695 	}
696 
697 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
698 
699 	return (rc);
700 }
701 
702 static void
703 sfxge_tx_qdpl_flush(struct sfxge_txq *txq)
704 {
705 	struct sfxge_tx_dpl *stdp = &txq->dpl;
706 	struct mbuf *mbuf, *next;
707 
708 	SFXGE_TXQ_LOCK(txq);
709 
710 	sfxge_tx_qdpl_swizzle(txq);
711 	for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) {
712 		next = mbuf->m_nextpkt;
713 		m_freem(mbuf);
714 	}
715 	stdp->std_get = NULL;
716 	stdp->std_get_count = 0;
717 	stdp->std_get_non_tcp_count = 0;
718 	stdp->std_getp = &stdp->std_get;
719 
720 	SFXGE_TXQ_UNLOCK(txq);
721 }
722 
723 void
724 sfxge_if_qflush(struct ifnet *ifp)
725 {
726 	struct sfxge_softc *sc;
727 	unsigned int i;
728 
729 	sc = ifp->if_softc;
730 
731 	for (i = 0; i < sc->txq_count; i++)
732 		sfxge_tx_qdpl_flush(sc->txq[i]);
733 }
734 
735 #if SFXGE_TX_PARSE_EARLY
736 
737 /* There is little space for user data in mbuf pkthdr, so we
738  * use l*hlen fields which are not used by the driver otherwise
739  * to store header offsets.
740  * The fields are 8-bit, but it's ok, no header may be longer than 255 bytes.
741  */
742 
743 
744 #define TSO_MBUF_PROTO(_mbuf)    ((_mbuf)->m_pkthdr.PH_loc.sixteen[0])
745 /* We abuse l5hlen here because PH_loc can hold only 64 bits of data */
746 #define TSO_MBUF_FLAGS(_mbuf)    ((_mbuf)->m_pkthdr.l5hlen)
747 #define TSO_MBUF_PACKETID(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[1])
748 #define TSO_MBUF_SEQNUM(_mbuf)   ((_mbuf)->m_pkthdr.PH_loc.thirtytwo[1])
749 
750 static void sfxge_parse_tx_packet(struct mbuf *mbuf)
751 {
752 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
753 	const struct tcphdr *th;
754 	struct tcphdr th_copy;
755 
756 	/* Find network protocol and header */
757 	TSO_MBUF_PROTO(mbuf) = eh->ether_type;
758 	if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_VLAN)) {
759 		struct ether_vlan_header *veh =
760 			mtod(mbuf, struct ether_vlan_header *);
761 		TSO_MBUF_PROTO(mbuf) = veh->evl_proto;
762 		mbuf->m_pkthdr.l2hlen = sizeof(*veh);
763 	} else {
764 		mbuf->m_pkthdr.l2hlen = sizeof(*eh);
765 	}
766 
767 	/* Find TCP header */
768 	if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IP)) {
769 		const struct ip *iph = (const struct ip *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen);
770 
771 		KASSERT(iph->ip_p == IPPROTO_TCP,
772 			("TSO required on non-TCP packet"));
773 		mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + 4 * iph->ip_hl;
774 		TSO_MBUF_PACKETID(mbuf) = iph->ip_id;
775 	} else {
776 		KASSERT(TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IPV6),
777 			("TSO required on non-IP packet"));
778 		KASSERT(((const struct ip6_hdr *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen))->ip6_nxt ==
779 			IPPROTO_TCP,
780 			("TSO required on non-TCP packet"));
781 		mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + sizeof(struct ip6_hdr);
782 		TSO_MBUF_PACKETID(mbuf) = 0;
783 	}
784 
785 	KASSERT(mbuf->m_len >= mbuf->m_pkthdr.l3hlen,
786 		("network header is fragmented in mbuf"));
787 
788 	/* We need TCP header including flags (window is the next) */
789 	if (mbuf->m_len < mbuf->m_pkthdr.l3hlen + offsetof(struct tcphdr, th_win)) {
790 		m_copydata(mbuf, mbuf->m_pkthdr.l3hlen, sizeof(th_copy),
791 			   (caddr_t)&th_copy);
792 		th = &th_copy;
793 	} else {
794 		th = (const struct tcphdr *)mtodo(mbuf, mbuf->m_pkthdr.l3hlen);
795 	}
796 
797 	mbuf->m_pkthdr.l4hlen = mbuf->m_pkthdr.l3hlen + 4 * th->th_off;
798 	TSO_MBUF_SEQNUM(mbuf) = ntohl(th->th_seq);
799 
800 	/* These flags must not be duplicated */
801 	/*
802 	 * RST should not be duplicated as well, but FreeBSD kernel
803 	 * generates TSO packets with RST flag. So, do not assert
804 	 * its absence.
805 	 */
806 	KASSERT(!(th->th_flags & (TH_URG | TH_SYN)),
807 		("incompatible TCP flag 0x%x on TSO packet",
808 		 th->th_flags & (TH_URG | TH_SYN)));
809 	TSO_MBUF_FLAGS(mbuf) = th->th_flags;
810 }
811 #endif
812 
813 /*
814  * TX start -- called by the stack.
815  */
816 int
817 sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m)
818 {
819 	struct sfxge_softc *sc;
820 	struct sfxge_txq *txq;
821 	int rc;
822 
823 	sc = (struct sfxge_softc *)ifp->if_softc;
824 
825 	/*
826 	 * Transmit may be called when interface is up from the kernel
827 	 * point of view, but not yet up (in progress) from the driver
828 	 * point of view. I.e. link aggregation bring up.
829 	 * Transmit may be called when interface is up from the driver
830 	 * point of view, but already down from the kernel point of
831 	 * view. I.e. Rx when interface shutdown is in progress.
832 	 */
833 	KASSERT((ifp->if_flags & IFF_UP) || (sc->if_flags & IFF_UP),
834 		("interface not up"));
835 
836 	/* Pick the desired transmit queue. */
837 	if (m->m_pkthdr.csum_flags &
838 	    (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO)) {
839 		int index = 0;
840 
841 #ifdef RSS
842 		uint32_t bucket_id;
843 
844 		/*
845 		 * Select a TX queue which matches the corresponding
846 		 * RX queue for the hash in order to assign both
847 		 * TX and RX parts of the flow to the same CPU
848 		 */
849 		if (rss_m2bucket(m, &bucket_id) == 0)
850 			index = bucket_id % (sc->txq_count - (SFXGE_TXQ_NTYPES - 1));
851 #else
852 		/* check if flowid is set */
853 		if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
854 			uint32_t hash = m->m_pkthdr.flowid;
855 			uint32_t idx = hash % nitems(sc->rx_indir_table);
856 
857 			index = sc->rx_indir_table[idx];
858 		}
859 #endif
860 #if SFXGE_TX_PARSE_EARLY
861 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
862 			sfxge_parse_tx_packet(m);
863 #endif
864 		txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index];
865 	} else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
866 		txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
867 	} else {
868 		txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
869 	}
870 
871 	rc = sfxge_tx_packet_add(txq, m);
872 	if (rc != 0)
873 		m_freem(m);
874 
875 	return (rc);
876 }
877 
878 /*
879  * Software "TSO".  Not quite as good as doing it in hardware, but
880  * still faster than segmenting in the stack.
881  */
882 
883 struct sfxge_tso_state {
884 	/* Output position */
885 	unsigned out_len;	/* Remaining length in current segment */
886 	unsigned seqnum;	/* Current sequence number */
887 	unsigned packet_space;	/* Remaining space in current packet */
888 	unsigned segs_space;	/* Remaining number of DMA segments
889 				   for the packet (FATSOv2 only) */
890 
891 	/* Input position */
892 	uint64_t dma_addr;	/* DMA address of current position */
893 	unsigned in_len;	/* Remaining length in current mbuf */
894 
895 	const struct mbuf *mbuf; /* Input mbuf (head of chain) */
896 	u_short protocol;	/* Network protocol (after VLAN decap) */
897 	ssize_t nh_off;		/* Offset of network header */
898 	ssize_t tcph_off;	/* Offset of TCP header */
899 	unsigned header_len;	/* Number of bytes of header */
900 	unsigned seg_size;	/* TCP segment size */
901 	int fw_assisted;	/* Use FW-assisted TSO */
902 	u_short packet_id;	/* IPv4 packet ID from the original packet */
903 	uint8_t tcp_flags;	/* TCP flags */
904 	efx_desc_t header_desc; /* Precomputed header descriptor for
905 				 * FW-assisted TSO */
906 };
907 
908 #if !SFXGE_TX_PARSE_EARLY
909 static const struct ip *tso_iph(const struct sfxge_tso_state *tso)
910 {
911 	KASSERT(tso->protocol == htons(ETHERTYPE_IP),
912 		("tso_iph() in non-IPv4 state"));
913 	return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
914 }
915 
916 static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
917 {
918 	KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
919 		("tso_ip6h() in non-IPv6 state"));
920 	return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
921 }
922 
923 static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
924 {
925 	return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
926 }
927 #endif
928 
929 
930 /* Size of preallocated TSO header buffers.  Larger blocks must be
931  * allocated from the heap.
932  */
933 #define	TSOH_STD_SIZE	128
934 
935 /* At most half the descriptors in the queue at any time will refer to
936  * a TSO header buffer, since they must always be followed by a
937  * payload descriptor referring to an mbuf.
938  */
939 #define	TSOH_COUNT(_txq_entries)	((_txq_entries) / 2u)
940 #define	TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
941 #define	TSOH_PAGE_COUNT(_txq_entries)	\
942 	howmany(TSOH_COUNT(_txq_entries), TSOH_PER_PAGE)
943 
944 static int tso_init(struct sfxge_txq *txq)
945 {
946 	struct sfxge_softc *sc = txq->sc;
947 	unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries);
948 	int i, rc;
949 
950 	/* Allocate TSO header buffers */
951 	txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]),
952 				  M_SFXGE, M_WAITOK);
953 
954 	for (i = 0; i < tsoh_page_count; i++) {
955 		rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
956 		if (rc != 0)
957 			goto fail;
958 	}
959 
960 	return (0);
961 
962 fail:
963 	while (i-- > 0)
964 		sfxge_dma_free(&txq->tsoh_buffer[i]);
965 	free(txq->tsoh_buffer, M_SFXGE);
966 	txq->tsoh_buffer = NULL;
967 	return (rc);
968 }
969 
970 static void tso_fini(struct sfxge_txq *txq)
971 {
972 	int i;
973 
974 	if (txq->tsoh_buffer != NULL) {
975 		for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++)
976 			sfxge_dma_free(&txq->tsoh_buffer[i]);
977 		free(txq->tsoh_buffer, M_SFXGE);
978 	}
979 }
980 
981 static void tso_start(struct sfxge_txq *txq, struct sfxge_tso_state *tso,
982 		      const bus_dma_segment_t *hdr_dma_seg,
983 		      struct mbuf *mbuf)
984 {
985 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->sc->enp);
986 #if !SFXGE_TX_PARSE_EARLY
987 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
988 	const struct tcphdr *th;
989 	struct tcphdr th_copy;
990 #endif
991 
992 	tso->fw_assisted = txq->tso_fw_assisted;
993 	tso->mbuf = mbuf;
994 
995 	/* Find network protocol and header */
996 #if !SFXGE_TX_PARSE_EARLY
997 	tso->protocol = eh->ether_type;
998 	if (tso->protocol == htons(ETHERTYPE_VLAN)) {
999 		struct ether_vlan_header *veh =
1000 			mtod(mbuf, struct ether_vlan_header *);
1001 		tso->protocol = veh->evl_proto;
1002 		tso->nh_off = sizeof(*veh);
1003 	} else {
1004 		tso->nh_off = sizeof(*eh);
1005 	}
1006 #else
1007 	tso->protocol = TSO_MBUF_PROTO(mbuf);
1008 	tso->nh_off = mbuf->m_pkthdr.l2hlen;
1009 	tso->tcph_off = mbuf->m_pkthdr.l3hlen;
1010 	tso->packet_id = ntohs(TSO_MBUF_PACKETID(mbuf));
1011 #endif
1012 
1013 #if !SFXGE_TX_PARSE_EARLY
1014 	/* Find TCP header */
1015 	if (tso->protocol == htons(ETHERTYPE_IP)) {
1016 		KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
1017 			("TSO required on non-TCP packet"));
1018 		tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
1019 		tso->packet_id = ntohs(tso_iph(tso)->ip_id);
1020 	} else {
1021 		KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
1022 			("TSO required on non-IP packet"));
1023 		KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
1024 			("TSO required on non-TCP packet"));
1025 		tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
1026 		tso->packet_id = 0;
1027 	}
1028 #endif
1029 
1030 
1031 	if (tso->fw_assisted &&
1032 	    __predict_false(tso->tcph_off >
1033 			    encp->enc_tx_tso_tcp_header_offset_limit)) {
1034 		tso->fw_assisted = 0;
1035 	}
1036 
1037 
1038 #if !SFXGE_TX_PARSE_EARLY
1039 	KASSERT(mbuf->m_len >= tso->tcph_off,
1040 		("network header is fragmented in mbuf"));
1041 	/* We need TCP header including flags (window is the next) */
1042 	if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) {
1043 		m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy),
1044 			   (caddr_t)&th_copy);
1045 		th = &th_copy;
1046 	} else {
1047 		th = tso_tcph(tso);
1048 	}
1049 	tso->header_len = tso->tcph_off + 4 * th->th_off;
1050 #else
1051 	tso->header_len = mbuf->m_pkthdr.l4hlen;
1052 #endif
1053 	tso->seg_size = mbuf->m_pkthdr.tso_segsz;
1054 
1055 #if !SFXGE_TX_PARSE_EARLY
1056 	tso->seqnum = ntohl(th->th_seq);
1057 
1058 	/* These flags must not be duplicated */
1059 	/*
1060 	 * RST should not be duplicated as well, but FreeBSD kernel
1061 	 * generates TSO packets with RST flag. So, do not assert
1062 	 * its absence.
1063 	 */
1064 	KASSERT(!(th->th_flags & (TH_URG | TH_SYN)),
1065 		("incompatible TCP flag 0x%x on TSO packet",
1066 		 th->th_flags & (TH_URG | TH_SYN)));
1067 	tso->tcp_flags = th->th_flags;
1068 #else
1069 	tso->seqnum = TSO_MBUF_SEQNUM(mbuf);
1070 	tso->tcp_flags = TSO_MBUF_FLAGS(mbuf);
1071 #endif
1072 
1073 	tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
1074 
1075 	if (tso->fw_assisted) {
1076 		if (hdr_dma_seg->ds_len >= tso->header_len)
1077 			efx_tx_qdesc_dma_create(txq->common,
1078 						hdr_dma_seg->ds_addr,
1079 						tso->header_len,
1080 						B_FALSE,
1081 						&tso->header_desc);
1082 		else
1083 			tso->fw_assisted = 0;
1084 	}
1085 }
1086 
1087 /*
1088  * tso_fill_packet_with_fragment - form descriptors for the current fragment
1089  *
1090  * Form descriptors for the current fragment, until we reach the end
1091  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
1092  * space.
1093  */
1094 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
1095 					  struct sfxge_tso_state *tso)
1096 {
1097 	efx_desc_t *desc;
1098 	int n;
1099 	uint64_t dma_addr = tso->dma_addr;
1100 	boolean_t eop;
1101 
1102 	if (tso->in_len == 0 || tso->packet_space == 0)
1103 		return;
1104 
1105 	KASSERT(tso->in_len > 0, ("TSO input length went negative"));
1106 	KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
1107 
1108 	if (tso->fw_assisted & SFXGE_FATSOV2) {
1109 		n = tso->in_len;
1110 		tso->out_len -= n;
1111 		tso->seqnum += n;
1112 		tso->in_len = 0;
1113 		if (n < tso->packet_space) {
1114 			tso->packet_space -= n;
1115 			tso->segs_space--;
1116 		} else {
1117 			tso->packet_space = tso->seg_size -
1118 			    (n - tso->packet_space) % tso->seg_size;
1119 			tso->segs_space =
1120 			    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1 -
1121 			    (tso->packet_space != tso->seg_size);
1122 		}
1123 	} else {
1124 		n = min(tso->in_len, tso->packet_space);
1125 		tso->packet_space -= n;
1126 		tso->out_len -= n;
1127 		tso->dma_addr += n;
1128 		tso->in_len -= n;
1129 	}
1130 
1131 	/*
1132 	 * It is OK to use binary OR below to avoid extra branching
1133 	 * since all conditions may always be checked.
1134 	 */
1135 	eop = (tso->out_len == 0) | (tso->packet_space == 0) |
1136 	    (tso->segs_space == 0);
1137 
1138 	desc = &txq->pend_desc[txq->n_pend_desc++];
1139 	efx_tx_qdesc_dma_create(txq->common, dma_addr, n, eop, desc);
1140 }
1141 
1142 /* Callback from bus_dmamap_load() for long TSO headers. */
1143 static void tso_map_long_header(void *dma_addr_ret,
1144 				bus_dma_segment_t *segs, int nseg,
1145 				int error)
1146 {
1147 	*(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
1148 				      __predict_true(nseg == 1)) ?
1149 				     segs->ds_addr : 0);
1150 }
1151 
1152 /*
1153  * tso_start_new_packet - generate a new header and prepare for the new packet
1154  *
1155  * Generate a new header and prepare for the new packet.  Return 0 on
1156  * success, or an error code if failed to alloc header.
1157  */
1158 static int tso_start_new_packet(struct sfxge_txq *txq,
1159 				struct sfxge_tso_state *tso,
1160 				unsigned int *idp)
1161 {
1162 	unsigned int id = *idp;
1163 	struct tcphdr *tsoh_th;
1164 	unsigned ip_length;
1165 	caddr_t header;
1166 	uint64_t dma_addr;
1167 	bus_dmamap_t map;
1168 	efx_desc_t *desc;
1169 	int rc;
1170 
1171 	if (tso->fw_assisted) {
1172 		if (tso->fw_assisted & SFXGE_FATSOV2) {
1173 			/* Add 2 FATSOv2 option descriptors */
1174 			desc = &txq->pend_desc[txq->n_pend_desc];
1175 			efx_tx_qdesc_tso2_create(txq->common,
1176 						 tso->packet_id,
1177 						 tso->seqnum,
1178 						 tso->seg_size,
1179 						 desc,
1180 						 EFX_TX_FATSOV2_OPT_NDESCS);
1181 			desc += EFX_TX_FATSOV2_OPT_NDESCS;
1182 			txq->n_pend_desc += EFX_TX_FATSOV2_OPT_NDESCS;
1183 			KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1184 			id = (id + EFX_TX_FATSOV2_OPT_NDESCS) & txq->ptr_mask;
1185 
1186 			tso->segs_space =
1187 			    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1;
1188 		} else {
1189 			uint8_t tcp_flags = tso->tcp_flags;
1190 
1191 			if (tso->out_len > tso->seg_size)
1192 				tcp_flags &= ~(TH_FIN | TH_PUSH);
1193 
1194 			/* Add FATSOv1 option descriptor */
1195 			desc = &txq->pend_desc[txq->n_pend_desc++];
1196 			efx_tx_qdesc_tso_create(txq->common,
1197 						tso->packet_id,
1198 						tso->seqnum,
1199 						tcp_flags,
1200 						desc++);
1201 			KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1202 			id = (id + 1) & txq->ptr_mask;
1203 
1204 			tso->seqnum += tso->seg_size;
1205 			tso->segs_space = UINT_MAX;
1206 		}
1207 
1208 		/* Header DMA descriptor */
1209 		*desc = tso->header_desc;
1210 		txq->n_pend_desc++;
1211 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1212 		id = (id + 1) & txq->ptr_mask;
1213 	} else {
1214 		/* Allocate a DMA-mapped header buffer. */
1215 		if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
1216 			unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
1217 			unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
1218 
1219 			header = (txq->tsoh_buffer[page_index].esm_base +
1220 				  buf_index * TSOH_STD_SIZE);
1221 			dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
1222 				    buf_index * TSOH_STD_SIZE);
1223 			map = txq->tsoh_buffer[page_index].esm_map;
1224 
1225 			KASSERT(txq->stmp[id].flags == 0,
1226 				("stmp flags are not 0"));
1227 		} else {
1228 			struct sfxge_tx_mapping *stmp = &txq->stmp[id];
1229 
1230 			/* We cannot use bus_dmamem_alloc() as that may sleep */
1231 			header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
1232 			if (__predict_false(!header))
1233 				return (ENOMEM);
1234 			rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
1235 					     header, tso->header_len,
1236 					     tso_map_long_header, &dma_addr,
1237 					     BUS_DMA_NOWAIT);
1238 			if (__predict_false(dma_addr == 0)) {
1239 				if (rc == 0) {
1240 					/* Succeeded but got >1 segment */
1241 					bus_dmamap_unload(txq->packet_dma_tag,
1242 							  stmp->map);
1243 					rc = EINVAL;
1244 				}
1245 				free(header, M_SFXGE);
1246 				return (rc);
1247 			}
1248 			map = stmp->map;
1249 
1250 			txq->tso_long_headers++;
1251 			stmp->u.heap_buf = header;
1252 			stmp->flags = TX_BUF_UNMAP;
1253 		}
1254 
1255 		tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
1256 
1257 		/* Copy and update the headers. */
1258 		m_copydata(tso->mbuf, 0, tso->header_len, header);
1259 
1260 		tsoh_th->th_seq = htonl(tso->seqnum);
1261 		tso->seqnum += tso->seg_size;
1262 		if (tso->out_len > tso->seg_size) {
1263 			/* This packet will not finish the TSO burst. */
1264 			ip_length = tso->header_len - tso->nh_off + tso->seg_size;
1265 			tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH);
1266 		} else {
1267 			/* This packet will be the last in the TSO burst. */
1268 			ip_length = tso->header_len - tso->nh_off + tso->out_len;
1269 		}
1270 
1271 		if (tso->protocol == htons(ETHERTYPE_IP)) {
1272 			struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
1273 			tsoh_iph->ip_len = htons(ip_length);
1274 			/* XXX We should increment ip_id, but FreeBSD doesn't
1275 			 * currently allocate extra IDs for multiple segments.
1276 			 */
1277 		} else {
1278 			struct ip6_hdr *tsoh_iph =
1279 				(struct ip6_hdr *)(header + tso->nh_off);
1280 			tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
1281 		}
1282 
1283 		/* Make the header visible to the hardware. */
1284 		bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
1285 
1286 		/* Form a descriptor for this header. */
1287 		desc = &txq->pend_desc[txq->n_pend_desc++];
1288 		efx_tx_qdesc_dma_create(txq->common,
1289 					dma_addr,
1290 					tso->header_len,
1291 					0,
1292 					desc);
1293 		id = (id + 1) & txq->ptr_mask;
1294 
1295 		tso->segs_space = UINT_MAX;
1296 	}
1297 	tso->packet_space = tso->seg_size;
1298 	txq->tso_packets++;
1299 	*idp = id;
1300 
1301 	return (0);
1302 }
1303 
1304 static int
1305 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
1306 		   const bus_dma_segment_t *dma_seg, int n_dma_seg,
1307 		   int vlan_tagged)
1308 {
1309 	struct sfxge_tso_state tso;
1310 	unsigned int id;
1311 	unsigned skipped = 0;
1312 
1313 	tso_start(txq, &tso, dma_seg, mbuf);
1314 
1315 	while (dma_seg->ds_len + skipped <= tso.header_len) {
1316 		skipped += dma_seg->ds_len;
1317 		--n_dma_seg;
1318 		KASSERT(n_dma_seg, ("no payload found in TSO packet"));
1319 		++dma_seg;
1320 	}
1321 	tso.in_len = dma_seg->ds_len - (tso.header_len - skipped);
1322 	tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped);
1323 
1324 	id = (txq->added + vlan_tagged) & txq->ptr_mask;
1325 	if (__predict_false(tso_start_new_packet(txq, &tso, &id)))
1326 		return (-1);
1327 
1328 	while (1) {
1329 		tso_fill_packet_with_fragment(txq, &tso);
1330 		/* Exactly one DMA descriptor is added */
1331 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1332 		id = (id + 1) & txq->ptr_mask;
1333 
1334 		/* Move onto the next fragment? */
1335 		if (tso.in_len == 0) {
1336 			--n_dma_seg;
1337 			if (n_dma_seg == 0)
1338 				break;
1339 			++dma_seg;
1340 			tso.in_len = dma_seg->ds_len;
1341 			tso.dma_addr = dma_seg->ds_addr;
1342 		}
1343 
1344 		/* End of packet? */
1345 		if ((tso.packet_space == 0) | (tso.segs_space == 0)) {
1346 			unsigned int n_fatso_opt_desc =
1347 			    (tso.fw_assisted & SFXGE_FATSOV2) ?
1348 			    EFX_TX_FATSOV2_OPT_NDESCS :
1349 			    (tso.fw_assisted & SFXGE_FATSOV1) ? 1 : 0;
1350 
1351 			/* If the queue is now full due to tiny MSS,
1352 			 * or we can't create another header, discard
1353 			 * the remainder of the input mbuf but do not
1354 			 * roll back the work we have done.
1355 			 */
1356 			if (txq->n_pend_desc + n_fatso_opt_desc +
1357 			    1 /* header */ + n_dma_seg > txq->max_pkt_desc) {
1358 				txq->tso_pdrop_too_many++;
1359 				break;
1360 			}
1361 			if (__predict_false(tso_start_new_packet(txq, &tso,
1362 								 &id))) {
1363 				txq->tso_pdrop_no_rsrc++;
1364 				break;
1365 			}
1366 		}
1367 	}
1368 
1369 	txq->tso_bursts++;
1370 	return (id);
1371 }
1372 
1373 static void
1374 sfxge_tx_qunblock(struct sfxge_txq *txq)
1375 {
1376 	struct sfxge_softc *sc;
1377 	struct sfxge_evq *evq;
1378 
1379 	sc = txq->sc;
1380 	evq = sc->evq[txq->evq_index];
1381 
1382 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
1383 
1384 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED))
1385 		return;
1386 
1387 	SFXGE_TXQ_LOCK(txq);
1388 
1389 	if (txq->blocked) {
1390 		unsigned int level;
1391 
1392 		level = txq->added - txq->completed;
1393 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) {
1394 			/* reaped must be in sync with blocked */
1395 			sfxge_tx_qreap(txq);
1396 			txq->blocked = 0;
1397 		}
1398 	}
1399 
1400 	sfxge_tx_qdpl_service(txq);
1401 	/* note: lock has been dropped */
1402 }
1403 
1404 void
1405 sfxge_tx_qflush_done(struct sfxge_txq *txq)
1406 {
1407 
1408 	txq->flush_state = SFXGE_FLUSH_DONE;
1409 }
1410 
1411 static void
1412 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1413 {
1414 	struct sfxge_txq *txq;
1415 	struct sfxge_evq *evq;
1416 	unsigned int count;
1417 
1418 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1419 
1420 	txq = sc->txq[index];
1421 	evq = sc->evq[txq->evq_index];
1422 
1423 	SFXGE_EVQ_LOCK(evq);
1424 	SFXGE_TXQ_LOCK(txq);
1425 
1426 	KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1427 	    ("txq->init_state != SFXGE_TXQ_STARTED"));
1428 
1429 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1430 
1431 	if (txq->flush_state != SFXGE_FLUSH_DONE) {
1432 		txq->flush_state = SFXGE_FLUSH_PENDING;
1433 
1434 		SFXGE_EVQ_UNLOCK(evq);
1435 		SFXGE_TXQ_UNLOCK(txq);
1436 
1437 		/* Flush the transmit queue. */
1438 		if (efx_tx_qflush(txq->common) != 0) {
1439 			log(LOG_ERR, "%s: Flushing Tx queue %u failed\n",
1440 			    device_get_nameunit(sc->dev), index);
1441 			txq->flush_state = SFXGE_FLUSH_DONE;
1442 		} else {
1443 			count = 0;
1444 			do {
1445 				/* Spin for 100ms. */
1446 				DELAY(100000);
1447 				if (txq->flush_state != SFXGE_FLUSH_PENDING)
1448 					break;
1449 			} while (++count < 20);
1450 		}
1451 		SFXGE_EVQ_LOCK(evq);
1452 		SFXGE_TXQ_LOCK(txq);
1453 
1454 		KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1455 		    ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1456 
1457 		if (txq->flush_state != SFXGE_FLUSH_DONE) {
1458 			/* Flush timeout */
1459 			log(LOG_ERR, "%s: Cannot flush Tx queue %u\n",
1460 			    device_get_nameunit(sc->dev), index);
1461 			txq->flush_state = SFXGE_FLUSH_DONE;
1462 		}
1463 	}
1464 
1465 	txq->blocked = 0;
1466 	txq->pending = txq->added;
1467 
1468 	sfxge_tx_qcomplete(txq, evq);
1469 	KASSERT(txq->completed == txq->added,
1470 	    ("txq->completed != txq->added"));
1471 
1472 	sfxge_tx_qreap(txq);
1473 	KASSERT(txq->reaped == txq->completed,
1474 	    ("txq->reaped != txq->completed"));
1475 
1476 	txq->added = 0;
1477 	txq->pending = 0;
1478 	txq->completed = 0;
1479 	txq->reaped = 0;
1480 
1481 	/* Destroy the common code transmit queue. */
1482 	efx_tx_qdestroy(txq->common);
1483 	txq->common = NULL;
1484 
1485 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1486 	    EFX_TXQ_NBUFS(sc->txq_entries));
1487 
1488 	SFXGE_EVQ_UNLOCK(evq);
1489 	SFXGE_TXQ_UNLOCK(txq);
1490 }
1491 
1492 /*
1493  * Estimate maximum number of Tx descriptors required for TSO packet.
1494  * With minimum MSS and maximum mbuf length we might need more (even
1495  * than a ring-ful of descriptors), but this should not happen in
1496  * practice except due to deliberate attack.  In that case we will
1497  * truncate the output at a packet boundary.
1498  */
1499 static unsigned int
1500 sfxge_tx_max_pkt_desc(const struct sfxge_softc *sc, enum sfxge_txq_type type,
1501 		      unsigned int tso_fw_assisted)
1502 {
1503 	/* One descriptor for every input fragment */
1504 	unsigned int max_descs = SFXGE_TX_MAPPING_MAX_SEG;
1505 	unsigned int sw_tso_max_descs;
1506 	unsigned int fa_tso_v1_max_descs = 0;
1507 	unsigned int fa_tso_v2_max_descs = 0;
1508 
1509 	/* VLAN tagging Tx option descriptor may be required */
1510 	if (efx_nic_cfg_get(sc->enp)->enc_hw_tx_insert_vlan_enabled)
1511 		max_descs++;
1512 
1513 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) {
1514 		/*
1515 		 * Plus header and payload descriptor for each output segment.
1516 		 * Minus one since header fragment is already counted.
1517 		 * Even if FATSO is used, we should be ready to fallback
1518 		 * to do it in the driver.
1519 		 */
1520 		sw_tso_max_descs = SFXGE_TSO_MAX_SEGS * 2 - 1;
1521 
1522 		/* FW assisted TSOv1 requires one more descriptor per segment
1523 		 * in comparison to SW TSO */
1524 		if (tso_fw_assisted & SFXGE_FATSOV1)
1525 			fa_tso_v1_max_descs =
1526 			    sw_tso_max_descs + SFXGE_TSO_MAX_SEGS;
1527 
1528 		/* FW assisted TSOv2 requires 3 (2 FATSO plus header) extra
1529 		 * descriptors per superframe limited by number of DMA fetches
1530 		 * per packet. The first packet header is already counted.
1531 		 */
1532 		if (tso_fw_assisted & SFXGE_FATSOV2) {
1533 			fa_tso_v2_max_descs =
1534 			    howmany(SFXGE_TX_MAPPING_MAX_SEG,
1535 				    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1) *
1536 			    (EFX_TX_FATSOV2_OPT_NDESCS + 1) - 1;
1537 		}
1538 
1539 		max_descs += MAX(sw_tso_max_descs,
1540 				 MAX(fa_tso_v1_max_descs, fa_tso_v2_max_descs));
1541 	}
1542 
1543 	return (max_descs);
1544 }
1545 
1546 static int
1547 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1548 {
1549 	struct sfxge_txq *txq;
1550 	efsys_mem_t *esmp;
1551 	uint16_t flags;
1552 	unsigned int tso_fw_assisted;
1553 	struct sfxge_evq *evq;
1554 	unsigned int desc_index;
1555 	int rc;
1556 
1557 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1558 
1559 	txq = sc->txq[index];
1560 	esmp = &txq->mem;
1561 	evq = sc->evq[txq->evq_index];
1562 
1563 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1564 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1565 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1566 	    ("evq->init_state != SFXGE_EVQ_STARTED"));
1567 
1568 	/* Program the buffer table. */
1569 	if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1570 	    EFX_TXQ_NBUFS(sc->txq_entries))) != 0)
1571 		return (rc);
1572 
1573 	/* Determine the kind of queue we are creating. */
1574 	tso_fw_assisted = 0;
1575 	switch (txq->type) {
1576 	case SFXGE_TXQ_NON_CKSUM:
1577 		flags = 0;
1578 		break;
1579 	case SFXGE_TXQ_IP_CKSUM:
1580 		flags = EFX_TXQ_CKSUM_IPV4;
1581 		break;
1582 	case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1583 		flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP;
1584 		tso_fw_assisted = sc->tso_fw_assisted;
1585 		if (tso_fw_assisted & SFXGE_FATSOV2)
1586 			flags |= EFX_TXQ_FATSOV2;
1587 		break;
1588 	default:
1589 		KASSERT(0, ("Impossible TX queue"));
1590 		flags = 0;
1591 		break;
1592 	}
1593 
1594 	/* Create the common code transmit queue. */
1595 	if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1596 	    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1597 	    &txq->common, &desc_index)) != 0) {
1598 		/* Retry if no FATSOv2 resources, otherwise fail */
1599 		if ((rc != ENOSPC) || (~flags & EFX_TXQ_FATSOV2))
1600 			goto fail;
1601 
1602 		/* Looks like all FATSOv2 contexts are used */
1603 		flags &= ~EFX_TXQ_FATSOV2;
1604 		tso_fw_assisted &= ~SFXGE_FATSOV2;
1605 		if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1606 		    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1607 		    &txq->common, &desc_index)) != 0)
1608 			goto fail;
1609 	}
1610 
1611 	/* Initialise queue descriptor indexes */
1612 	txq->added = txq->pending = txq->completed = txq->reaped = desc_index;
1613 
1614 	SFXGE_TXQ_LOCK(txq);
1615 
1616 	/* Enable the transmit queue. */
1617 	efx_tx_qenable(txq->common);
1618 
1619 	txq->init_state = SFXGE_TXQ_STARTED;
1620 	txq->flush_state = SFXGE_FLUSH_REQUIRED;
1621 	txq->tso_fw_assisted = tso_fw_assisted;
1622 
1623 	txq->max_pkt_desc = sfxge_tx_max_pkt_desc(sc, txq->type,
1624 						  tso_fw_assisted);
1625 
1626 	txq->hw_vlan_tci = 0;
1627 
1628 	SFXGE_TXQ_UNLOCK(txq);
1629 
1630 	return (0);
1631 
1632 fail:
1633 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1634 	    EFX_TXQ_NBUFS(sc->txq_entries));
1635 	return (rc);
1636 }
1637 
1638 void
1639 sfxge_tx_stop(struct sfxge_softc *sc)
1640 {
1641 	int index;
1642 
1643 	index = sc->txq_count;
1644 	while (--index >= 0)
1645 		sfxge_tx_qstop(sc, index);
1646 
1647 	/* Tear down the transmit module */
1648 	efx_tx_fini(sc->enp);
1649 }
1650 
1651 int
1652 sfxge_tx_start(struct sfxge_softc *sc)
1653 {
1654 	int index;
1655 	int rc;
1656 
1657 	/* Initialize the common code transmit module. */
1658 	if ((rc = efx_tx_init(sc->enp)) != 0)
1659 		return (rc);
1660 
1661 	for (index = 0; index < sc->txq_count; index++) {
1662 		if ((rc = sfxge_tx_qstart(sc, index)) != 0)
1663 			goto fail;
1664 	}
1665 
1666 	return (0);
1667 
1668 fail:
1669 	while (--index >= 0)
1670 		sfxge_tx_qstop(sc, index);
1671 
1672 	efx_tx_fini(sc->enp);
1673 
1674 	return (rc);
1675 }
1676 
1677 static int
1678 sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node)
1679 {
1680 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev);
1681 	struct sysctl_oid *stat_node;
1682 	unsigned int id;
1683 
1684 	stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1685 				    "stats", CTLFLAG_RD, NULL,
1686 				    "Tx queue statistics");
1687 	if (stat_node == NULL)
1688 		return (ENOMEM);
1689 
1690 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1691 		SYSCTL_ADD_ULONG(
1692 		    ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO,
1693 		    sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS,
1694 		    (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset),
1695 		    "");
1696 	}
1697 
1698 	return (0);
1699 }
1700 
1701 /**
1702  * Destroy a transmit queue.
1703  */
1704 static void
1705 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1706 {
1707 	struct sfxge_txq *txq;
1708 	unsigned int nmaps;
1709 
1710 	txq = sc->txq[index];
1711 
1712 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1713 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1714 
1715 	if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1716 		tso_fini(txq);
1717 
1718 	/* Free the context arrays. */
1719 	free(txq->pend_desc, M_SFXGE);
1720 	nmaps = sc->txq_entries;
1721 	while (nmaps-- != 0)
1722 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1723 	free(txq->stmp, M_SFXGE);
1724 
1725 	/* Release DMA memory mapping. */
1726 	sfxge_dma_free(&txq->mem);
1727 
1728 	sc->txq[index] = NULL;
1729 
1730 	SFXGE_TXQ_LOCK_DESTROY(txq);
1731 
1732 	free(txq, M_SFXGE);
1733 }
1734 
1735 static int
1736 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1737 	       enum sfxge_txq_type type, unsigned int evq_index)
1738 {
1739 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
1740 	char name[16];
1741 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1742 	struct sysctl_oid *txq_node;
1743 	struct sfxge_txq *txq;
1744 	struct sfxge_evq *evq;
1745 	struct sfxge_tx_dpl *stdp;
1746 	struct sysctl_oid *dpl_node;
1747 	efsys_mem_t *esmp;
1748 	unsigned int nmaps;
1749 	int rc;
1750 
1751 	txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1752 	txq->sc = sc;
1753 	txq->entries = sc->txq_entries;
1754 	txq->ptr_mask = txq->entries - 1;
1755 
1756 	sc->txq[txq_index] = txq;
1757 	esmp = &txq->mem;
1758 
1759 	evq = sc->evq[evq_index];
1760 
1761 	/* Allocate and zero DMA space for the descriptor ring. */
1762 	if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0)
1763 		return (rc);
1764 
1765 	/* Allocate buffer table entries. */
1766 	sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries),
1767 				 &txq->buf_base_id);
1768 
1769 	/* Create a DMA tag for packet mappings. */
1770 	if (bus_dma_tag_create(sc->parent_dma_tag, 1,
1771 	    encp->enc_tx_dma_desc_boundary,
1772 	    MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1773 	    NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG,
1774 	    encp->enc_tx_dma_desc_size_max, 0, NULL, NULL,
1775 	    &txq->packet_dma_tag) != 0) {
1776 		device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1777 		rc = ENOMEM;
1778 		goto fail;
1779 	}
1780 
1781 	/* Allocate pending descriptor array for batching writes. */
1782 	txq->pend_desc = malloc(sizeof(efx_desc_t) * sc->txq_entries,
1783 				M_SFXGE, M_ZERO | M_WAITOK);
1784 
1785 	/* Allocate and initialise mbuf DMA mapping array. */
1786 	txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries,
1787 	    M_SFXGE, M_ZERO | M_WAITOK);
1788 	for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) {
1789 		rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1790 				       &txq->stmp[nmaps].map);
1791 		if (rc != 0)
1792 			goto fail2;
1793 	}
1794 
1795 	snprintf(name, sizeof(name), "%u", txq_index);
1796 	txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node),
1797 				   OID_AUTO, name, CTLFLAG_RD, NULL, "");
1798 	if (txq_node == NULL) {
1799 		rc = ENOMEM;
1800 		goto fail_txq_node;
1801 	}
1802 
1803 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1804 	    (rc = tso_init(txq)) != 0)
1805 		goto fail3;
1806 
1807 	/* Initialize the deferred packet list. */
1808 	stdp = &txq->dpl;
1809 	stdp->std_put_max = sfxge_tx_dpl_put_max;
1810 	stdp->std_get_max = sfxge_tx_dpl_get_max;
1811 	stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max;
1812 	stdp->std_getp = &stdp->std_get;
1813 
1814 	SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index);
1815 
1816 	dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1817 				   "dpl", CTLFLAG_RD, NULL,
1818 				   "Deferred packet list statistics");
1819 	if (dpl_node == NULL) {
1820 		rc = ENOMEM;
1821 		goto fail_dpl_node;
1822 	}
1823 
1824 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1825 			"get_count", CTLFLAG_RD | CTLFLAG_STATS,
1826 			&stdp->std_get_count, 0, "");
1827 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1828 			"get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS,
1829 			&stdp->std_get_non_tcp_count, 0, "");
1830 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1831 			"get_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1832 			&stdp->std_get_hiwat, 0, "");
1833 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1834 			"put_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1835 			&stdp->std_put_hiwat, 0, "");
1836 
1837 	rc = sfxge_txq_stat_init(txq, txq_node);
1838 	if (rc != 0)
1839 		goto fail_txq_stat_init;
1840 
1841 	txq->type = type;
1842 	txq->evq_index = evq_index;
1843 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1844 
1845 	return (0);
1846 
1847 fail_txq_stat_init:
1848 fail_dpl_node:
1849 fail3:
1850 fail_txq_node:
1851 	free(txq->pend_desc, M_SFXGE);
1852 fail2:
1853 	while (nmaps-- != 0)
1854 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1855 	free(txq->stmp, M_SFXGE);
1856 	bus_dma_tag_destroy(txq->packet_dma_tag);
1857 
1858 fail:
1859 	sfxge_dma_free(esmp);
1860 
1861 	return (rc);
1862 }
1863 
1864 static int
1865 sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1866 {
1867 	struct sfxge_softc *sc = arg1;
1868 	unsigned int id = arg2;
1869 	unsigned long sum;
1870 	unsigned int index;
1871 
1872 	/* Sum across all TX queues */
1873 	sum = 0;
1874 	for (index = 0; index < sc->txq_count; index++)
1875 		sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1876 					  sfxge_tx_stats[id].offset);
1877 
1878 	return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1879 }
1880 
1881 static void
1882 sfxge_tx_stat_init(struct sfxge_softc *sc)
1883 {
1884 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1885 	struct sysctl_oid_list *stat_list;
1886 	unsigned int id;
1887 
1888 	stat_list = SYSCTL_CHILDREN(sc->stats_node);
1889 
1890 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1891 		SYSCTL_ADD_PROC(
1892 			ctx, stat_list,
1893 			OID_AUTO, sfxge_tx_stats[id].name,
1894 			CTLTYPE_ULONG|CTLFLAG_RD,
1895 			sc, id, sfxge_tx_stat_handler, "LU",
1896 			"");
1897 	}
1898 }
1899 
1900 uint64_t
1901 sfxge_tx_get_drops(struct sfxge_softc *sc)
1902 {
1903 	unsigned int index;
1904 	uint64_t drops = 0;
1905 	struct sfxge_txq *txq;
1906 
1907 	/* Sum across all TX queues */
1908 	for (index = 0; index < sc->txq_count; index++) {
1909 		txq = sc->txq[index];
1910 		/*
1911 		 * In theory, txq->put_overflow and txq->netdown_drops
1912 		 * should use atomic operation and other should be
1913 		 * obtained under txq lock, but it is just statistics.
1914 		 */
1915 		drops += txq->drops + txq->get_overflow +
1916 			 txq->get_non_tcp_overflow +
1917 			 txq->put_overflow + txq->netdown_drops +
1918 			 txq->tso_pdrop_too_many + txq->tso_pdrop_no_rsrc;
1919 	}
1920 	return (drops);
1921 }
1922 
1923 void
1924 sfxge_tx_fini(struct sfxge_softc *sc)
1925 {
1926 	int index;
1927 
1928 	index = sc->txq_count;
1929 	while (--index >= 0)
1930 		sfxge_tx_qfini(sc, index);
1931 
1932 	sc->txq_count = 0;
1933 }
1934 
1935 
1936 int
1937 sfxge_tx_init(struct sfxge_softc *sc)
1938 {
1939 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
1940 	struct sfxge_intr *intr;
1941 	int index;
1942 	int rc;
1943 
1944 	intr = &sc->intr;
1945 
1946 	KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1947 	    ("intr->state != SFXGE_INTR_INITIALIZED"));
1948 
1949 	if (sfxge_tx_dpl_get_max <= 0) {
1950 		log(LOG_ERR, "%s=%d must be greater than 0",
1951 		    SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max);
1952 		rc = EINVAL;
1953 		goto fail_tx_dpl_get_max;
1954 	}
1955 	if (sfxge_tx_dpl_get_non_tcp_max <= 0) {
1956 		log(LOG_ERR, "%s=%d must be greater than 0",
1957 		    SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX,
1958 		    sfxge_tx_dpl_get_non_tcp_max);
1959 		rc = EINVAL;
1960 		goto fail_tx_dpl_get_non_tcp_max;
1961 	}
1962 	if (sfxge_tx_dpl_put_max < 0) {
1963 		log(LOG_ERR, "%s=%d must be greater or equal to 0",
1964 		    SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max);
1965 		rc = EINVAL;
1966 		goto fail_tx_dpl_put_max;
1967 	}
1968 
1969 	sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc;
1970 
1971 	sc->tso_fw_assisted = sfxge_tso_fw_assisted;
1972 	if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO) ||
1973 	    (!encp->enc_fw_assisted_tso_enabled))
1974 		sc->tso_fw_assisted &= ~SFXGE_FATSOV1;
1975 	if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO_V2) ||
1976 	    (!encp->enc_fw_assisted_tso_v2_enabled))
1977 		sc->tso_fw_assisted &= ~SFXGE_FATSOV2;
1978 
1979 	sc->txqs_node = SYSCTL_ADD_NODE(
1980 		device_get_sysctl_ctx(sc->dev),
1981 		SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
1982 		OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues");
1983 	if (sc->txqs_node == NULL) {
1984 		rc = ENOMEM;
1985 		goto fail_txq_node;
1986 	}
1987 
1988 	/* Initialize the transmit queues */
1989 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
1990 	    SFXGE_TXQ_NON_CKSUM, 0)) != 0)
1991 		goto fail;
1992 
1993 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
1994 	    SFXGE_TXQ_IP_CKSUM, 0)) != 0)
1995 		goto fail2;
1996 
1997 	for (index = 0;
1998 	     index < sc->txq_count - SFXGE_TXQ_NTYPES + 1;
1999 	     index++) {
2000 		if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index,
2001 		    SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
2002 			goto fail3;
2003 	}
2004 
2005 	sfxge_tx_stat_init(sc);
2006 
2007 	return (0);
2008 
2009 fail3:
2010 	while (--index >= 0)
2011 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
2012 
2013 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
2014 
2015 fail2:
2016 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
2017 
2018 fail:
2019 fail_txq_node:
2020 	sc->txq_count = 0;
2021 fail_tx_dpl_put_max:
2022 fail_tx_dpl_get_non_tcp_max:
2023 fail_tx_dpl_get_max:
2024 	return (rc);
2025 }
2026