xref: /freebsd/sys/dev/sfxge/sfxge_tx.c (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2010-2016 Solarflare Communications Inc.
5  * All rights reserved.
6  *
7  * This software was developed in part by Philip Paeps under contract for
8  * Solarflare Communications, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions are met:
12  *
13  * 1. Redistributions of source code must retain the above copyright notice,
14  *    this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright notice,
16  *    this list of conditions and the following disclaimer in the documentation
17  *    and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
20  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
21  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
23  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
24  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
25  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
26  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
27  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
28  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
29  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30  *
31  * The views and conclusions contained in the software and documentation are
32  * those of the authors and should not be interpreted as representing official
33  * policies, either expressed or implied, of the FreeBSD Project.
34  */
35 
36 /* Theory of operation:
37  *
38  * Tx queues allocation and mapping on Siena
39  *
40  * One Tx queue with enabled checksum offload is allocated per Rx channel
41  * (event queue).  Also 2 Tx queues (one without checksum offload and one
42  * with IP checksum offload only) are allocated and bound to event queue 0.
43  * sfxge_txq_type is used as Tx queue label.
44  *
45  * So, event queue plus label mapping to Tx queue index is:
46  *	if event queue index is 0, TxQ-index = TxQ-label * [0..SFXGE_TXQ_NTYPES)
47  *	else TxQ-index = SFXGE_TXQ_NTYPES + EvQ-index - 1
48  * See sfxge_get_txq_by_label() sfxge_ev.c
49  *
50  * Tx queue allocation and mapping on EF10
51  *
52  * One Tx queue with enabled checksum offload is allocated per Rx
53  * channel (event queue). Checksum offload on all Tx queues is enabled or
54  * disabled dynamically by inserting option descriptors, so the additional
55  * queues used on Siena are not required.
56  *
57  * TxQ label is always set to zero on EF10 hardware.
58  * So, event queue to Tx queue mapping is simple:
59  * TxQ-index = EvQ-index
60  */
61 
62 #include <sys/cdefs.h>
63 #include "opt_rss.h"
64 
65 #include <sys/param.h>
66 #include <sys/malloc.h>
67 #include <sys/mbuf.h>
68 #include <sys/smp.h>
69 #include <sys/socket.h>
70 #include <sys/sysctl.h>
71 #include <sys/syslog.h>
72 #include <sys/limits.h>
73 
74 #include <net/bpf.h>
75 #include <net/ethernet.h>
76 #include <net/if.h>
77 #include <net/if_vlan_var.h>
78 
79 #include <netinet/in.h>
80 #include <netinet/ip.h>
81 #include <netinet/ip6.h>
82 #include <netinet/tcp.h>
83 
84 #ifdef RSS
85 #include <net/rss_config.h>
86 #endif
87 
88 #include "common/efx.h"
89 
90 #include "sfxge.h"
91 #include "sfxge_tx.h"
92 
93 #define	SFXGE_PARAM_TX_DPL_GET_MAX	SFXGE_PARAM(tx_dpl_get_max)
94 static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT;
95 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max);
96 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN,
97 	   &sfxge_tx_dpl_get_max, 0,
98 	   "Maximum number of any packets in deferred packet get-list");
99 
100 #define	SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \
101 	SFXGE_PARAM(tx_dpl_get_non_tcp_max)
102 static int sfxge_tx_dpl_get_non_tcp_max =
103 	SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT;
104 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max);
105 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN,
106 	   &sfxge_tx_dpl_get_non_tcp_max, 0,
107 	   "Maximum number of non-TCP packets in deferred packet get-list");
108 
109 #define	SFXGE_PARAM_TX_DPL_PUT_MAX	SFXGE_PARAM(tx_dpl_put_max)
110 static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT;
111 TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max);
112 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN,
113 	   &sfxge_tx_dpl_put_max, 0,
114 	   "Maximum number of any packets in deferred packet put-list");
115 
116 #define	SFXGE_PARAM_TSO_FW_ASSISTED	SFXGE_PARAM(tso_fw_assisted)
117 static int sfxge_tso_fw_assisted = (SFXGE_FATSOV1 | SFXGE_FATSOV2);
118 TUNABLE_INT(SFXGE_PARAM_TSO_FW_ASSISTED, &sfxge_tso_fw_assisted);
119 SYSCTL_INT(_hw_sfxge, OID_AUTO, tso_fw_assisted, CTLFLAG_RDTUN,
120 	   &sfxge_tso_fw_assisted, 0,
121 	   "Bitmask of FW-assisted TSO allowed to use if supported by NIC firmware");
122 
123 static const struct {
124 	const char *name;
125 	size_t offset;
126 } sfxge_tx_stats[] = {
127 #define	SFXGE_TX_STAT(name, member) \
128 	{ #name, offsetof(struct sfxge_txq, member) }
129 	SFXGE_TX_STAT(tso_bursts, tso_bursts),
130 	SFXGE_TX_STAT(tso_packets, tso_packets),
131 	SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
132 	SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many),
133 	SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc),
134 	SFXGE_TX_STAT(tx_collapses, collapses),
135 	SFXGE_TX_STAT(tx_drops, drops),
136 	SFXGE_TX_STAT(tx_get_overflow, get_overflow),
137 	SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow),
138 	SFXGE_TX_STAT(tx_put_overflow, put_overflow),
139 	SFXGE_TX_STAT(tx_netdown_drops, netdown_drops),
140 };
141 
142 /* Forward declarations. */
143 static void sfxge_tx_qdpl_service(struct sfxge_txq *txq);
144 static void sfxge_tx_qlist_post(struct sfxge_txq *txq);
145 static void sfxge_tx_qunblock(struct sfxge_txq *txq);
146 static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
147 			      const bus_dma_segment_t *dma_seg, int n_dma_seg,
148 			      int n_extra_descs);
149 
150 static inline void
151 sfxge_next_stmp(struct sfxge_txq *txq, struct sfxge_tx_mapping **pstmp)
152 {
153 	KASSERT((*pstmp)->flags == 0, ("stmp flags are not 0"));
154 	if (__predict_false(*pstmp ==
155 			    &txq->stmp[txq->ptr_mask]))
156 		*pstmp = &txq->stmp[0];
157 	else
158 		(*pstmp)++;
159 }
160 
161 static int
162 sfxge_tx_maybe_toggle_cksum_offload(struct sfxge_txq *txq, struct mbuf *mbuf,
163 				    struct sfxge_tx_mapping **pstmp)
164 {
165 	uint16_t new_hw_cksum_flags;
166 	efx_desc_t *desc;
167 
168 	if (mbuf->m_pkthdr.csum_flags &
169 	    (CSUM_DELAY_DATA | CSUM_DELAY_DATA_IPV6 | CSUM_TSO)) {
170 		/*
171 		 * We always set EFX_TXQ_CKSUM_IPV4 here because this
172 		 * configuration is the most useful, and this won't
173 		 * cause any trouble in case of IPv6 traffic anyway.
174 		 */
175 		new_hw_cksum_flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP;
176 	} else if (mbuf->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
177 		new_hw_cksum_flags = EFX_TXQ_CKSUM_IPV4;
178 	} else {
179 		new_hw_cksum_flags = 0;
180 	}
181 
182 	if (new_hw_cksum_flags == txq->hw_cksum_flags)
183 		return (0);
184 
185 	desc = &txq->pend_desc[txq->n_pend_desc];
186 	efx_tx_qdesc_checksum_create(txq->common, new_hw_cksum_flags, desc);
187 	txq->hw_cksum_flags = new_hw_cksum_flags;
188 	txq->n_pend_desc++;
189 
190 	sfxge_next_stmp(txq, pstmp);
191 
192 	return (1);
193 }
194 
195 static int
196 sfxge_tx_maybe_insert_tag(struct sfxge_txq *txq, struct mbuf *mbuf,
197 			  struct sfxge_tx_mapping **pstmp)
198 {
199 	uint16_t this_tag = ((mbuf->m_flags & M_VLANTAG) ?
200 			     mbuf->m_pkthdr.ether_vtag :
201 			     0);
202 	efx_desc_t *desc;
203 
204 	if (this_tag == txq->hw_vlan_tci)
205 		return (0);
206 
207 	desc = &txq->pend_desc[txq->n_pend_desc];
208 	efx_tx_qdesc_vlantci_create(txq->common, bswap16(this_tag), desc);
209 	txq->hw_vlan_tci = this_tag;
210 	txq->n_pend_desc++;
211 
212 	sfxge_next_stmp(txq, pstmp);
213 
214 	return (1);
215 }
216 
217 void
218 sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq)
219 {
220 	unsigned int completed;
221 
222 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
223 
224 	completed = txq->completed;
225 	while (completed != txq->pending) {
226 		struct sfxge_tx_mapping *stmp;
227 		unsigned int id;
228 
229 		id = completed++ & txq->ptr_mask;
230 
231 		stmp = &txq->stmp[id];
232 		if (stmp->flags & TX_BUF_UNMAP) {
233 			bus_dmamap_unload(txq->packet_dma_tag, stmp->map);
234 			if (stmp->flags & TX_BUF_MBUF) {
235 				struct mbuf *m = stmp->u.mbuf;
236 				do
237 					m = m_free(m);
238 				while (m != NULL);
239 			} else {
240 				free(stmp->u.heap_buf, M_SFXGE);
241 			}
242 			stmp->flags = 0;
243 		}
244 	}
245 	txq->completed = completed;
246 
247 	/* Check whether we need to unblock the queue. */
248 	mb();
249 	if (txq->blocked) {
250 		unsigned int level;
251 
252 		level = txq->added - txq->completed;
253 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
254 			sfxge_tx_qunblock(txq);
255 	}
256 }
257 
258 static unsigned int
259 sfxge_is_mbuf_non_tcp(struct mbuf *mbuf)
260 {
261 	/* Absence of TCP checksum flags does not mean that it is non-TCP
262 	 * but it should be true if user wants to achieve high throughput.
263 	 */
264 	return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP)));
265 }
266 
267 /*
268  * Reorder the put list and append it to the get list.
269  */
270 static void
271 sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq)
272 {
273 	struct sfxge_tx_dpl *stdp;
274 	struct mbuf *mbuf, *get_next, **get_tailp;
275 	volatile uintptr_t *putp;
276 	uintptr_t put;
277 	unsigned int count;
278 	unsigned int non_tcp_count;
279 
280 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
281 
282 	stdp = &txq->dpl;
283 
284 	/* Acquire the put list. */
285 	putp = &stdp->std_put;
286 	put = atomic_readandclear_ptr(putp);
287 	mbuf = (void *)put;
288 
289 	if (mbuf == NULL)
290 		return;
291 
292 	/* Reverse the put list. */
293 	get_tailp = &mbuf->m_nextpkt;
294 	get_next = NULL;
295 
296 	count = 0;
297 	non_tcp_count = 0;
298 	do {
299 		struct mbuf *put_next;
300 
301 		non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf);
302 		put_next = mbuf->m_nextpkt;
303 		mbuf->m_nextpkt = get_next;
304 		get_next = mbuf;
305 		mbuf = put_next;
306 
307 		count++;
308 	} while (mbuf != NULL);
309 
310 	if (count > stdp->std_put_hiwat)
311 		stdp->std_put_hiwat = count;
312 
313 	/* Append the reversed put list to the get list. */
314 	KASSERT(*get_tailp == NULL, ("*get_tailp != NULL"));
315 	*stdp->std_getp = get_next;
316 	stdp->std_getp = get_tailp;
317 	stdp->std_get_count += count;
318 	stdp->std_get_non_tcp_count += non_tcp_count;
319 }
320 
321 static void
322 sfxge_tx_qreap(struct sfxge_txq *txq)
323 {
324 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
325 
326 	txq->reaped = txq->completed;
327 }
328 
329 static void
330 sfxge_tx_qlist_post(struct sfxge_txq *txq)
331 {
332 	unsigned int old_added __diagused;
333 	unsigned int block_level;
334 	unsigned int level;
335 	int rc __diagused;
336 
337 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
338 
339 	KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0"));
340 	KASSERT(txq->n_pend_desc <= txq->max_pkt_desc,
341 		("txq->n_pend_desc too large"));
342 	KASSERT(!txq->blocked, ("txq->blocked"));
343 
344 	old_added = txq->added;
345 
346 	/* Post the fragment list. */
347 	rc = efx_tx_qdesc_post(txq->common, txq->pend_desc, txq->n_pend_desc,
348 			  txq->reaped, &txq->added);
349 	KASSERT(rc == 0, ("efx_tx_qdesc_post() failed"));
350 
351 	/* If efx_tx_qdesc_post() had to refragment, our information about
352 	 * buffers to free may be associated with the wrong
353 	 * descriptors.
354 	 */
355 	KASSERT(txq->added - old_added == txq->n_pend_desc,
356 		("efx_tx_qdesc_post() refragmented descriptors"));
357 
358 	level = txq->added - txq->reaped;
359 	KASSERT(level <= txq->entries, ("overfilled TX queue"));
360 
361 	/* Clear the fragment list. */
362 	txq->n_pend_desc = 0;
363 
364 	/*
365 	 * Set the block level to ensure there is space to generate a
366 	 * large number of descriptors for TSO.
367 	 */
368 	block_level = EFX_TXQ_LIMIT(txq->entries) - txq->max_pkt_desc;
369 
370 	/* Have we reached the block level? */
371 	if (level < block_level)
372 		return;
373 
374 	/* Reap, and check again */
375 	sfxge_tx_qreap(txq);
376 	level = txq->added - txq->reaped;
377 	if (level < block_level)
378 		return;
379 
380 	txq->blocked = 1;
381 
382 	/*
383 	 * Avoid a race with completion interrupt handling that could leave
384 	 * the queue blocked.
385 	 */
386 	mb();
387 	sfxge_tx_qreap(txq);
388 	level = txq->added - txq->reaped;
389 	if (level < block_level) {
390 		mb();
391 		txq->blocked = 0;
392 	}
393 }
394 
395 static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
396 {
397 	bus_dmamap_t *used_map;
398 	bus_dmamap_t map;
399 	bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG];
400 	unsigned int id;
401 	struct sfxge_tx_mapping *stmp;
402 	efx_desc_t *desc;
403 	int n_dma_seg;
404 	int rc;
405 	int i;
406 	int eop;
407 	uint16_t hw_cksum_flags_prev;
408 	uint16_t hw_vlan_tci_prev;
409 	int n_extra_descs;
410 
411 	KASSERT(!txq->blocked, ("txq->blocked"));
412 
413 #if SFXGE_TX_PARSE_EARLY
414 	/*
415 	 * If software TSO is used, we still need to copy packet header,
416 	 * even if we have already parsed it early before enqueue.
417 	 */
418 	if ((mbuf->m_pkthdr.csum_flags & CSUM_TSO) &&
419 	    (txq->tso_fw_assisted == 0))
420 		prefetch_read_many(mbuf->m_data);
421 #else
422 	/*
423 	 * Prefetch packet header since we need to parse it and extract
424 	 * IP ID, TCP sequence number and flags.
425 	 */
426 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
427 		prefetch_read_many(mbuf->m_data);
428 #endif
429 
430 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) {
431 		rc = EINTR;
432 		goto reject;
433 	}
434 
435 	/* Load the packet for DMA. */
436 	id = txq->added & txq->ptr_mask;
437 	stmp = &txq->stmp[id];
438 	rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map,
439 				     mbuf, dma_seg, &n_dma_seg, 0);
440 	if (rc == EFBIG) {
441 		/* Try again. */
442 		struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT,
443 						   SFXGE_TX_MAPPING_MAX_SEG);
444 		if (new_mbuf == NULL)
445 			goto reject;
446 		++txq->collapses;
447 		mbuf = new_mbuf;
448 		rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag,
449 					     stmp->map, mbuf,
450 					     dma_seg, &n_dma_seg, 0);
451 	}
452 	if (rc != 0)
453 		goto reject;
454 
455 	/* Make the packet visible to the hardware. */
456 	bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE);
457 
458 	used_map = &stmp->map;
459 
460 	hw_cksum_flags_prev = txq->hw_cksum_flags;
461 	hw_vlan_tci_prev = txq->hw_vlan_tci;
462 
463 	/*
464 	 * The order of option descriptors, which are used to leverage VLAN tag
465 	 * and checksum offloads, might be important. Changing checksum offload
466 	 * between VLAN option and packet descriptors probably does not work.
467 	 */
468 	n_extra_descs = sfxge_tx_maybe_toggle_cksum_offload(txq, mbuf, &stmp);
469 	n_extra_descs += sfxge_tx_maybe_insert_tag(txq, mbuf, &stmp);
470 
471 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
472 		rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg,
473 					n_extra_descs);
474 		if (rc < 0)
475 			goto reject_mapped;
476 		stmp = &txq->stmp[(rc - 1) & txq->ptr_mask];
477 	} else {
478 		/* Add the mapping to the fragment list, and set flags
479 		 * for the buffer.
480 		 */
481 
482 		i = 0;
483 		for (;;) {
484 			desc = &txq->pend_desc[i + n_extra_descs];
485 			eop = (i == n_dma_seg - 1);
486 			efx_tx_qdesc_dma_create(txq->common,
487 						dma_seg[i].ds_addr,
488 						dma_seg[i].ds_len,
489 						eop,
490 						desc);
491 			if (eop)
492 				break;
493 			i++;
494 			sfxge_next_stmp(txq, &stmp);
495 		}
496 		txq->n_pend_desc = n_dma_seg + n_extra_descs;
497 	}
498 
499 	/*
500 	 * If the mapping required more than one descriptor
501 	 * then we need to associate the DMA map with the last
502 	 * descriptor, not the first.
503 	 */
504 	if (used_map != &stmp->map) {
505 		map = stmp->map;
506 		stmp->map = *used_map;
507 		*used_map = map;
508 	}
509 
510 	stmp->u.mbuf = mbuf;
511 	stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF;
512 
513 	/* Post the fragment list. */
514 	sfxge_tx_qlist_post(txq);
515 
516 	return (0);
517 
518 reject_mapped:
519 	txq->hw_vlan_tci = hw_vlan_tci_prev;
520 	txq->hw_cksum_flags = hw_cksum_flags_prev;
521 	bus_dmamap_unload(txq->packet_dma_tag, *used_map);
522 reject:
523 	/* Drop the packet on the floor. */
524 	m_freem(mbuf);
525 	++txq->drops;
526 
527 	return (rc);
528 }
529 
530 /*
531  * Drain the deferred packet list into the transmit queue.
532  */
533 static void
534 sfxge_tx_qdpl_drain(struct sfxge_txq *txq)
535 {
536 	struct sfxge_softc *sc;
537 	struct sfxge_tx_dpl *stdp;
538 	struct mbuf *mbuf, *next;
539 	unsigned int count;
540 	unsigned int non_tcp_count;
541 	unsigned int pushed;
542 	int rc;
543 
544 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
545 
546 	sc = txq->sc;
547 	stdp = &txq->dpl;
548 	pushed = txq->added;
549 
550 	if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) {
551 		prefetch_read_many(sc->enp);
552 		prefetch_read_many(txq->common);
553 	}
554 
555 	mbuf = stdp->std_get;
556 	count = stdp->std_get_count;
557 	non_tcp_count = stdp->std_get_non_tcp_count;
558 
559 	if (count > stdp->std_get_hiwat)
560 		stdp->std_get_hiwat = count;
561 
562 	while (count != 0) {
563 		KASSERT(mbuf != NULL, ("mbuf == NULL"));
564 
565 		next = mbuf->m_nextpkt;
566 		mbuf->m_nextpkt = NULL;
567 
568 		ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */
569 
570 		if (next != NULL)
571 			prefetch_read_many(next);
572 
573 		rc = sfxge_tx_queue_mbuf(txq, mbuf);
574 		--count;
575 		non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf);
576 		mbuf = next;
577 		if (rc != 0)
578 			continue;
579 
580 		if (txq->blocked)
581 			break;
582 
583 		/* Push the fragments to the hardware in batches. */
584 		if (txq->added - pushed >= SFXGE_TX_BATCH) {
585 			efx_tx_qpush(txq->common, txq->added, pushed);
586 			pushed = txq->added;
587 		}
588 	}
589 
590 	if (count == 0) {
591 		KASSERT(mbuf == NULL, ("mbuf != NULL"));
592 		KASSERT(non_tcp_count == 0,
593 			("inconsistent TCP/non-TCP detection"));
594 		stdp->std_get = NULL;
595 		stdp->std_get_count = 0;
596 		stdp->std_get_non_tcp_count = 0;
597 		stdp->std_getp = &stdp->std_get;
598 	} else {
599 		stdp->std_get = mbuf;
600 		stdp->std_get_count = count;
601 		stdp->std_get_non_tcp_count = non_tcp_count;
602 	}
603 
604 	if (txq->added != pushed)
605 		efx_tx_qpush(txq->common, txq->added, pushed);
606 
607 	KASSERT(txq->blocked || stdp->std_get_count == 0,
608 		("queue unblocked but count is non-zero"));
609 }
610 
611 #define	SFXGE_TX_QDPL_PENDING(_txq)	((_txq)->dpl.std_put != 0)
612 
613 /*
614  * Service the deferred packet list.
615  *
616  * NOTE: drops the txq mutex!
617  */
618 static void
619 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
620 {
621 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
622 
623 	do {
624 		if (SFXGE_TX_QDPL_PENDING(txq))
625 			sfxge_tx_qdpl_swizzle(txq);
626 
627 		if (!txq->blocked)
628 			sfxge_tx_qdpl_drain(txq);
629 
630 		SFXGE_TXQ_UNLOCK(txq);
631 	} while (SFXGE_TX_QDPL_PENDING(txq) &&
632 		 SFXGE_TXQ_TRYLOCK(txq));
633 }
634 
635 /*
636  * Put a packet on the deferred packet get-list.
637  */
638 static int
639 sfxge_tx_qdpl_put_locked(struct sfxge_txq *txq, struct mbuf *mbuf)
640 {
641 	struct sfxge_tx_dpl *stdp;
642 
643 	stdp = &txq->dpl;
644 
645 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
646 
647 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
648 
649 	if (stdp->std_get_count >= stdp->std_get_max) {
650 		txq->get_overflow++;
651 		return (ENOBUFS);
652 	}
653 	if (sfxge_is_mbuf_non_tcp(mbuf)) {
654 		if (stdp->std_get_non_tcp_count >=
655 		    stdp->std_get_non_tcp_max) {
656 			txq->get_non_tcp_overflow++;
657 			return (ENOBUFS);
658 		}
659 		stdp->std_get_non_tcp_count++;
660 	}
661 
662 	*(stdp->std_getp) = mbuf;
663 	stdp->std_getp = &mbuf->m_nextpkt;
664 	stdp->std_get_count++;
665 
666 	return (0);
667 }
668 
669 /*
670  * Put a packet on the deferred packet put-list.
671  *
672  * We overload the csum_data field in the mbuf to keep track of this length
673  * because there is no cheap alternative to avoid races.
674  */
675 static int
676 sfxge_tx_qdpl_put_unlocked(struct sfxge_txq *txq, struct mbuf *mbuf)
677 {
678 	struct sfxge_tx_dpl *stdp;
679 	volatile uintptr_t *putp;
680 	uintptr_t old;
681 	uintptr_t new;
682 	unsigned int put_count;
683 
684 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
685 
686 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
687 
688 	stdp = &txq->dpl;
689 	putp = &stdp->std_put;
690 	new = (uintptr_t)mbuf;
691 
692 	do {
693 		old = *putp;
694 		if (old != 0) {
695 			struct mbuf *mp = (struct mbuf *)old;
696 			put_count = mp->m_pkthdr.csum_data;
697 		} else
698 			put_count = 0;
699 		if (put_count >= stdp->std_put_max) {
700 			atomic_add_long(&txq->put_overflow, 1);
701 			return (ENOBUFS);
702 		}
703 		mbuf->m_pkthdr.csum_data = put_count + 1;
704 		mbuf->m_nextpkt = (void *)old;
705 	} while (atomic_cmpset_ptr(putp, old, new) == 0);
706 
707 	return (0);
708 }
709 
710 /*
711  * Called from if_transmit - will try to grab the txq lock and enqueue to the
712  * put list if it succeeds, otherwise try to push onto the defer list if space.
713  */
714 static int
715 sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m)
716 {
717 	int rc;
718 
719 	if (!SFXGE_LINK_UP(txq->sc)) {
720 		atomic_add_long(&txq->netdown_drops, 1);
721 		return (ENETDOWN);
722 	}
723 
724 	/*
725 	 * Try to grab the txq lock.  If we are able to get the lock,
726 	 * the packet will be appended to the "get list" of the deferred
727 	 * packet list.  Otherwise, it will be pushed on the "put list".
728 	 */
729 	if (SFXGE_TXQ_TRYLOCK(txq)) {
730 		/* First swizzle put-list to get-list to keep order */
731 		sfxge_tx_qdpl_swizzle(txq);
732 
733 		rc = sfxge_tx_qdpl_put_locked(txq, m);
734 
735 		/* Try to service the list. */
736 		sfxge_tx_qdpl_service(txq);
737 		/* Lock has been dropped. */
738 	} else {
739 		rc = sfxge_tx_qdpl_put_unlocked(txq, m);
740 
741 		/*
742 		 * Try to grab the lock again.
743 		 *
744 		 * If we are able to get the lock, we need to process
745 		 * the deferred packet list.  If we are not able to get
746 		 * the lock, another thread is processing the list.
747 		 */
748 		if ((rc == 0) && SFXGE_TXQ_TRYLOCK(txq)) {
749 			sfxge_tx_qdpl_service(txq);
750 			/* Lock has been dropped. */
751 		}
752 	}
753 
754 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
755 
756 	return (rc);
757 }
758 
759 static void
760 sfxge_tx_qdpl_flush(struct sfxge_txq *txq)
761 {
762 	struct sfxge_tx_dpl *stdp = &txq->dpl;
763 	struct mbuf *mbuf, *next;
764 
765 	SFXGE_TXQ_LOCK(txq);
766 
767 	sfxge_tx_qdpl_swizzle(txq);
768 	for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) {
769 		next = mbuf->m_nextpkt;
770 		m_freem(mbuf);
771 	}
772 	stdp->std_get = NULL;
773 	stdp->std_get_count = 0;
774 	stdp->std_get_non_tcp_count = 0;
775 	stdp->std_getp = &stdp->std_get;
776 
777 	SFXGE_TXQ_UNLOCK(txq);
778 }
779 
780 void
781 sfxge_if_qflush(if_t ifp)
782 {
783 	struct sfxge_softc *sc;
784 	unsigned int i;
785 
786 	sc = if_getsoftc(ifp);
787 
788 	for (i = 0; i < sc->txq_count; i++)
789 		sfxge_tx_qdpl_flush(sc->txq[i]);
790 }
791 
792 #if SFXGE_TX_PARSE_EARLY
793 
794 /* There is little space for user data in mbuf pkthdr, so we
795  * use l*hlen fields which are not used by the driver otherwise
796  * to store header offsets.
797  * The fields are 8-bit, but it's ok, no header may be longer than 255 bytes.
798  */
799 
800 #define TSO_MBUF_PROTO(_mbuf)    ((_mbuf)->m_pkthdr.PH_loc.sixteen[0])
801 /* We abuse l5hlen here because PH_loc can hold only 64 bits of data */
802 #define TSO_MBUF_FLAGS(_mbuf)    ((_mbuf)->m_pkthdr.l5hlen)
803 #define TSO_MBUF_PACKETID(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[1])
804 #define TSO_MBUF_SEQNUM(_mbuf)   ((_mbuf)->m_pkthdr.PH_loc.thirtytwo[1])
805 
806 static void sfxge_parse_tx_packet(struct mbuf *mbuf)
807 {
808 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
809 	const struct tcphdr *th;
810 	struct tcphdr th_copy;
811 
812 	/* Find network protocol and header */
813 	TSO_MBUF_PROTO(mbuf) = eh->ether_type;
814 	if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_VLAN)) {
815 		struct ether_vlan_header *veh =
816 			mtod(mbuf, struct ether_vlan_header *);
817 		TSO_MBUF_PROTO(mbuf) = veh->evl_proto;
818 		mbuf->m_pkthdr.l2hlen = sizeof(*veh);
819 	} else {
820 		mbuf->m_pkthdr.l2hlen = sizeof(*eh);
821 	}
822 
823 	/* Find TCP header */
824 	if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IP)) {
825 		const struct ip *iph = (const struct ip *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen);
826 
827 		KASSERT(iph->ip_p == IPPROTO_TCP,
828 			("TSO required on non-TCP packet"));
829 		mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + 4 * iph->ip_hl;
830 		TSO_MBUF_PACKETID(mbuf) = iph->ip_id;
831 	} else {
832 		KASSERT(TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IPV6),
833 			("TSO required on non-IP packet"));
834 		KASSERT(((const struct ip6_hdr *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen))->ip6_nxt ==
835 			IPPROTO_TCP,
836 			("TSO required on non-TCP packet"));
837 		mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + sizeof(struct ip6_hdr);
838 		TSO_MBUF_PACKETID(mbuf) = 0;
839 	}
840 
841 	KASSERT(mbuf->m_len >= mbuf->m_pkthdr.l3hlen,
842 		("network header is fragmented in mbuf"));
843 
844 	/* We need TCP header including flags (window is the next) */
845 	if (mbuf->m_len < mbuf->m_pkthdr.l3hlen + offsetof(struct tcphdr, th_win)) {
846 		m_copydata(mbuf, mbuf->m_pkthdr.l3hlen, sizeof(th_copy),
847 			   (caddr_t)&th_copy);
848 		th = &th_copy;
849 	} else {
850 		th = (const struct tcphdr *)mtodo(mbuf, mbuf->m_pkthdr.l3hlen);
851 	}
852 
853 	mbuf->m_pkthdr.l4hlen = mbuf->m_pkthdr.l3hlen + 4 * th->th_off;
854 	TSO_MBUF_SEQNUM(mbuf) = ntohl(th->th_seq);
855 
856 	/* These flags must not be duplicated */
857 	/*
858 	 * RST should not be duplicated as well, but FreeBSD kernel
859 	 * generates TSO packets with RST flag. So, do not assert
860 	 * its absence.
861 	 */
862 	KASSERT(!(tcp_get_flags(th) & (TH_URG | TH_SYN)),
863 		("incompatible TCP flag 0x%x on TSO packet",
864 		 tcp_get_flags(th) & (TH_URG | TH_SYN)));
865 	TSO_MBUF_FLAGS(mbuf) = tcp_get_flags(th);
866 }
867 #endif
868 
869 /*
870  * TX start -- called by the stack.
871  */
872 int
873 sfxge_if_transmit(if_t ifp, struct mbuf *m)
874 {
875 	struct sfxge_softc *sc;
876 	struct sfxge_txq *txq;
877 	int rc;
878 
879 	sc = (struct sfxge_softc *)if_getsoftc(ifp);
880 
881 	/*
882 	 * Transmit may be called when interface is up from the kernel
883 	 * point of view, but not yet up (in progress) from the driver
884 	 * point of view. I.e. link aggregation bring up.
885 	 * Transmit may be called when interface is up from the driver
886 	 * point of view, but already down from the kernel point of
887 	 * view. I.e. Rx when interface shutdown is in progress.
888 	 */
889 	KASSERT((if_getflags(ifp) & IFF_UP) || (sc->if_flags & IFF_UP),
890 		("interface not up"));
891 
892 	/* Pick the desired transmit queue. */
893 	if (sc->txq_dynamic_cksum_toggle_supported |
894 	    (m->m_pkthdr.csum_flags &
895 	     (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO))) {
896 		int index = 0;
897 
898 #ifdef RSS
899 		uint32_t bucket_id;
900 
901 		/*
902 		 * Select a TX queue which matches the corresponding
903 		 * RX queue for the hash in order to assign both
904 		 * TX and RX parts of the flow to the same CPU
905 		 */
906 		if (rss_m2bucket(m, &bucket_id) == 0)
907 			index = bucket_id % (sc->txq_count - (SFXGE_TXQ_NTYPES - 1));
908 #else
909 		/* check if flowid is set */
910 		if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
911 			uint32_t hash = m->m_pkthdr.flowid;
912 			uint32_t idx = hash % nitems(sc->rx_indir_table);
913 
914 			index = sc->rx_indir_table[idx];
915 		}
916 #endif
917 #if SFXGE_TX_PARSE_EARLY
918 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
919 			sfxge_parse_tx_packet(m);
920 #endif
921 		index += (sc->txq_dynamic_cksum_toggle_supported == B_FALSE) ?
922 			 SFXGE_TXQ_IP_TCP_UDP_CKSUM : 0;
923 		txq = sc->txq[index];
924 	} else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
925 		txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
926 	} else {
927 		txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
928 	}
929 
930 	rc = sfxge_tx_packet_add(txq, m);
931 	if (rc != 0)
932 		m_freem(m);
933 
934 	return (rc);
935 }
936 
937 /*
938  * Software "TSO".  Not quite as good as doing it in hardware, but
939  * still faster than segmenting in the stack.
940  */
941 
942 struct sfxge_tso_state {
943 	/* Output position */
944 	unsigned out_len;	/* Remaining length in current segment */
945 	unsigned seqnum;	/* Current sequence number */
946 	unsigned packet_space;	/* Remaining space in current packet */
947 	unsigned segs_space;	/* Remaining number of DMA segments
948 				   for the packet (FATSOv2 only) */
949 
950 	/* Input position */
951 	uint64_t dma_addr;	/* DMA address of current position */
952 	unsigned in_len;	/* Remaining length in current mbuf */
953 
954 	const struct mbuf *mbuf; /* Input mbuf (head of chain) */
955 	u_short protocol;	/* Network protocol (after VLAN decap) */
956 	ssize_t nh_off;		/* Offset of network header */
957 	ssize_t tcph_off;	/* Offset of TCP header */
958 	unsigned header_len;	/* Number of bytes of header */
959 	unsigned seg_size;	/* TCP segment size */
960 	int fw_assisted;	/* Use FW-assisted TSO */
961 	u_short packet_id;	/* IPv4 packet ID from the original packet */
962 	uint8_t tcp_flags;	/* TCP flags */
963 	efx_desc_t header_desc; /* Precomputed header descriptor for
964 				 * FW-assisted TSO */
965 };
966 
967 #if !SFXGE_TX_PARSE_EARLY
968 static const struct ip *tso_iph(const struct sfxge_tso_state *tso)
969 {
970 	KASSERT(tso->protocol == htons(ETHERTYPE_IP),
971 		("tso_iph() in non-IPv4 state"));
972 	return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
973 }
974 
975 static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
976 {
977 	KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
978 		("tso_ip6h() in non-IPv6 state"));
979 	return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
980 }
981 
982 static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
983 {
984 	return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
985 }
986 #endif
987 
988 /* Size of preallocated TSO header buffers.  Larger blocks must be
989  * allocated from the heap.
990  */
991 #define	TSOH_STD_SIZE	128
992 
993 /* At most half the descriptors in the queue at any time will refer to
994  * a TSO header buffer, since they must always be followed by a
995  * payload descriptor referring to an mbuf.
996  */
997 #define	TSOH_COUNT(_txq_entries)	((_txq_entries) / 2u)
998 #define	TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
999 #define	TSOH_PAGE_COUNT(_txq_entries)	\
1000 	howmany(TSOH_COUNT(_txq_entries), TSOH_PER_PAGE)
1001 
1002 static int tso_init(struct sfxge_txq *txq)
1003 {
1004 	struct sfxge_softc *sc = txq->sc;
1005 	unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries);
1006 	int i, rc;
1007 
1008 	/* Allocate TSO header buffers */
1009 	txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]),
1010 				  M_SFXGE, M_WAITOK);
1011 
1012 	for (i = 0; i < tsoh_page_count; i++) {
1013 		rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
1014 		if (rc != 0)
1015 			goto fail;
1016 	}
1017 
1018 	return (0);
1019 
1020 fail:
1021 	while (i-- > 0)
1022 		sfxge_dma_free(&txq->tsoh_buffer[i]);
1023 	free(txq->tsoh_buffer, M_SFXGE);
1024 	txq->tsoh_buffer = NULL;
1025 	return (rc);
1026 }
1027 
1028 static void tso_fini(struct sfxge_txq *txq)
1029 {
1030 	int i;
1031 
1032 	if (txq->tsoh_buffer != NULL) {
1033 		for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++)
1034 			sfxge_dma_free(&txq->tsoh_buffer[i]);
1035 		free(txq->tsoh_buffer, M_SFXGE);
1036 	}
1037 }
1038 
1039 static void tso_start(struct sfxge_txq *txq, struct sfxge_tso_state *tso,
1040 		      const bus_dma_segment_t *hdr_dma_seg,
1041 		      struct mbuf *mbuf)
1042 {
1043 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->sc->enp);
1044 #if !SFXGE_TX_PARSE_EARLY
1045 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
1046 	const struct tcphdr *th;
1047 	struct tcphdr th_copy;
1048 #endif
1049 
1050 	tso->fw_assisted = txq->tso_fw_assisted;
1051 	tso->mbuf = mbuf;
1052 
1053 	/* Find network protocol and header */
1054 #if !SFXGE_TX_PARSE_EARLY
1055 	tso->protocol = eh->ether_type;
1056 	if (tso->protocol == htons(ETHERTYPE_VLAN)) {
1057 		struct ether_vlan_header *veh =
1058 			mtod(mbuf, struct ether_vlan_header *);
1059 		tso->protocol = veh->evl_proto;
1060 		tso->nh_off = sizeof(*veh);
1061 	} else {
1062 		tso->nh_off = sizeof(*eh);
1063 	}
1064 #else
1065 	tso->protocol = TSO_MBUF_PROTO(mbuf);
1066 	tso->nh_off = mbuf->m_pkthdr.l2hlen;
1067 	tso->tcph_off = mbuf->m_pkthdr.l3hlen;
1068 	tso->packet_id = ntohs(TSO_MBUF_PACKETID(mbuf));
1069 #endif
1070 
1071 #if !SFXGE_TX_PARSE_EARLY
1072 	/* Find TCP header */
1073 	if (tso->protocol == htons(ETHERTYPE_IP)) {
1074 		KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
1075 			("TSO required on non-TCP packet"));
1076 		tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
1077 		tso->packet_id = ntohs(tso_iph(tso)->ip_id);
1078 	} else {
1079 		KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
1080 			("TSO required on non-IP packet"));
1081 		KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
1082 			("TSO required on non-TCP packet"));
1083 		tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
1084 		tso->packet_id = 0;
1085 	}
1086 #endif
1087 
1088 	if (tso->fw_assisted &&
1089 	    __predict_false(tso->tcph_off >
1090 			    encp->enc_tx_tso_tcp_header_offset_limit)) {
1091 		tso->fw_assisted = 0;
1092 	}
1093 
1094 #if !SFXGE_TX_PARSE_EARLY
1095 	KASSERT(mbuf->m_len >= tso->tcph_off,
1096 		("network header is fragmented in mbuf"));
1097 	/* We need TCP header including flags (window is the next) */
1098 	if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) {
1099 		m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy),
1100 			   (caddr_t)&th_copy);
1101 		th = &th_copy;
1102 	} else {
1103 		th = tso_tcph(tso);
1104 	}
1105 	tso->header_len = tso->tcph_off + 4 * th->th_off;
1106 #else
1107 	tso->header_len = mbuf->m_pkthdr.l4hlen;
1108 #endif
1109 	tso->seg_size = mbuf->m_pkthdr.tso_segsz;
1110 
1111 #if !SFXGE_TX_PARSE_EARLY
1112 	tso->seqnum = ntohl(th->th_seq);
1113 
1114 	/* These flags must not be duplicated */
1115 	/*
1116 	 * RST should not be duplicated as well, but FreeBSD kernel
1117 	 * generates TSO packets with RST flag. So, do not assert
1118 	 * its absence.
1119 	 */
1120 	KASSERT(!(tcp_get_flags(th) & (TH_URG | TH_SYN)),
1121 		("incompatible TCP flag 0x%x on TSO packet",
1122 		 tcp_get_flags(th) & (TH_URG | TH_SYN)));
1123 	tso->tcp_flags = tcp_get_flags(th);
1124 #else
1125 	tso->seqnum = TSO_MBUF_SEQNUM(mbuf);
1126 	tso->tcp_flags = TSO_MBUF_FLAGS(mbuf);
1127 #endif
1128 
1129 	tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
1130 
1131 	if (tso->fw_assisted) {
1132 		if (hdr_dma_seg->ds_len >= tso->header_len)
1133 			efx_tx_qdesc_dma_create(txq->common,
1134 						hdr_dma_seg->ds_addr,
1135 						tso->header_len,
1136 						B_FALSE,
1137 						&tso->header_desc);
1138 		else
1139 			tso->fw_assisted = 0;
1140 	}
1141 }
1142 
1143 /*
1144  * tso_fill_packet_with_fragment - form descriptors for the current fragment
1145  *
1146  * Form descriptors for the current fragment, until we reach the end
1147  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
1148  * space.
1149  */
1150 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
1151 					  struct sfxge_tso_state *tso)
1152 {
1153 	efx_desc_t *desc;
1154 	int n;
1155 	uint64_t dma_addr = tso->dma_addr;
1156 	boolean_t eop;
1157 
1158 	if (tso->in_len == 0 || tso->packet_space == 0)
1159 		return;
1160 
1161 	KASSERT(tso->in_len > 0, ("TSO input length went negative"));
1162 	KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
1163 
1164 	if (tso->fw_assisted & SFXGE_FATSOV2) {
1165 		n = tso->in_len;
1166 		tso->out_len -= n;
1167 		tso->seqnum += n;
1168 		tso->in_len = 0;
1169 		if (n < tso->packet_space) {
1170 			tso->packet_space -= n;
1171 			tso->segs_space--;
1172 		} else {
1173 			tso->packet_space = tso->seg_size -
1174 			    (n - tso->packet_space) % tso->seg_size;
1175 			tso->segs_space =
1176 			    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1 -
1177 			    (tso->packet_space != tso->seg_size);
1178 		}
1179 	} else {
1180 		n = min(tso->in_len, tso->packet_space);
1181 		tso->packet_space -= n;
1182 		tso->out_len -= n;
1183 		tso->dma_addr += n;
1184 		tso->in_len -= n;
1185 	}
1186 
1187 	/*
1188 	 * It is OK to use binary OR below to avoid extra branching
1189 	 * since all conditions may always be checked.
1190 	 */
1191 	eop = (tso->out_len == 0) | (tso->packet_space == 0) |
1192 	    (tso->segs_space == 0);
1193 
1194 	desc = &txq->pend_desc[txq->n_pend_desc++];
1195 	efx_tx_qdesc_dma_create(txq->common, dma_addr, n, eop, desc);
1196 }
1197 
1198 /* Callback from bus_dmamap_load() for long TSO headers. */
1199 static void tso_map_long_header(void *dma_addr_ret,
1200 				bus_dma_segment_t *segs, int nseg,
1201 				int error)
1202 {
1203 	*(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
1204 				      __predict_true(nseg == 1)) ?
1205 				     segs->ds_addr : 0);
1206 }
1207 
1208 /*
1209  * tso_start_new_packet - generate a new header and prepare for the new packet
1210  *
1211  * Generate a new header and prepare for the new packet.  Return 0 on
1212  * success, or an error code if failed to alloc header.
1213  */
1214 static int tso_start_new_packet(struct sfxge_txq *txq,
1215 				struct sfxge_tso_state *tso,
1216 				unsigned int *idp)
1217 {
1218 	unsigned int id = *idp;
1219 	struct tcphdr *tsoh_th;
1220 	unsigned ip_length;
1221 	caddr_t header;
1222 	uint64_t dma_addr;
1223 	bus_dmamap_t map;
1224 	efx_desc_t *desc;
1225 	int rc;
1226 
1227 	if (tso->fw_assisted) {
1228 		if (tso->fw_assisted & SFXGE_FATSOV2) {
1229 			/* Add 2 FATSOv2 option descriptors */
1230 			desc = &txq->pend_desc[txq->n_pend_desc];
1231 			efx_tx_qdesc_tso2_create(txq->common,
1232 						 tso->packet_id,
1233 						 0,
1234 						 tso->seqnum,
1235 						 tso->seg_size,
1236 						 desc,
1237 						 EFX_TX_FATSOV2_OPT_NDESCS);
1238 			desc += EFX_TX_FATSOV2_OPT_NDESCS;
1239 			txq->n_pend_desc += EFX_TX_FATSOV2_OPT_NDESCS;
1240 			KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1241 			id = (id + EFX_TX_FATSOV2_OPT_NDESCS) & txq->ptr_mask;
1242 
1243 			tso->segs_space =
1244 			    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1;
1245 		} else {
1246 			uint8_t tcp_flags = tso->tcp_flags;
1247 
1248 			if (tso->out_len > tso->seg_size)
1249 				tcp_flags &= ~(TH_FIN | TH_PUSH);
1250 
1251 			/* Add FATSOv1 option descriptor */
1252 			desc = &txq->pend_desc[txq->n_pend_desc++];
1253 			efx_tx_qdesc_tso_create(txq->common,
1254 						tso->packet_id,
1255 						tso->seqnum,
1256 						tcp_flags,
1257 						desc++);
1258 			KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1259 			id = (id + 1) & txq->ptr_mask;
1260 
1261 			tso->seqnum += tso->seg_size;
1262 			tso->segs_space = UINT_MAX;
1263 		}
1264 
1265 		/* Header DMA descriptor */
1266 		*desc = tso->header_desc;
1267 		txq->n_pend_desc++;
1268 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1269 		id = (id + 1) & txq->ptr_mask;
1270 	} else {
1271 		/* Allocate a DMA-mapped header buffer. */
1272 		if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
1273 			unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
1274 			unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
1275 
1276 			header = (txq->tsoh_buffer[page_index].esm_base +
1277 				  buf_index * TSOH_STD_SIZE);
1278 			dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
1279 				    buf_index * TSOH_STD_SIZE);
1280 			map = txq->tsoh_buffer[page_index].esm_map;
1281 
1282 			KASSERT(txq->stmp[id].flags == 0,
1283 				("stmp flags are not 0"));
1284 		} else {
1285 			struct sfxge_tx_mapping *stmp = &txq->stmp[id];
1286 
1287 			/* We cannot use bus_dmamem_alloc() as that may sleep */
1288 			header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
1289 			if (__predict_false(!header))
1290 				return (ENOMEM);
1291 			rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
1292 					     header, tso->header_len,
1293 					     tso_map_long_header, &dma_addr,
1294 					     BUS_DMA_NOWAIT);
1295 			if (__predict_false(dma_addr == 0)) {
1296 				if (rc == 0) {
1297 					/* Succeeded but got >1 segment */
1298 					bus_dmamap_unload(txq->packet_dma_tag,
1299 							  stmp->map);
1300 					rc = EINVAL;
1301 				}
1302 				free(header, M_SFXGE);
1303 				return (rc);
1304 			}
1305 			map = stmp->map;
1306 
1307 			txq->tso_long_headers++;
1308 			stmp->u.heap_buf = header;
1309 			stmp->flags = TX_BUF_UNMAP;
1310 		}
1311 
1312 		tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
1313 
1314 		/* Copy and update the headers. */
1315 		m_copydata(tso->mbuf, 0, tso->header_len, header);
1316 
1317 		tsoh_th->th_seq = htonl(tso->seqnum);
1318 		tso->seqnum += tso->seg_size;
1319 		if (tso->out_len > tso->seg_size) {
1320 			/* This packet will not finish the TSO burst. */
1321 			ip_length = tso->header_len - tso->nh_off + tso->seg_size;
1322 			tcp_set_flags(tsoh_th, tcp_get_flags(tsoh_th) & ~(TH_FIN | TH_PUSH));
1323 		} else {
1324 			/* This packet will be the last in the TSO burst. */
1325 			ip_length = tso->header_len - tso->nh_off + tso->out_len;
1326 		}
1327 
1328 		if (tso->protocol == htons(ETHERTYPE_IP)) {
1329 			struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
1330 			tsoh_iph->ip_len = htons(ip_length);
1331 			/* XXX We should increment ip_id, but FreeBSD doesn't
1332 			 * currently allocate extra IDs for multiple segments.
1333 			 */
1334 		} else {
1335 			struct ip6_hdr *tsoh_iph =
1336 				(struct ip6_hdr *)(header + tso->nh_off);
1337 			tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
1338 		}
1339 
1340 		/* Make the header visible to the hardware. */
1341 		bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
1342 
1343 		/* Form a descriptor for this header. */
1344 		desc = &txq->pend_desc[txq->n_pend_desc++];
1345 		efx_tx_qdesc_dma_create(txq->common,
1346 					dma_addr,
1347 					tso->header_len,
1348 					0,
1349 					desc);
1350 		id = (id + 1) & txq->ptr_mask;
1351 
1352 		tso->segs_space = UINT_MAX;
1353 	}
1354 	tso->packet_space = tso->seg_size;
1355 	txq->tso_packets++;
1356 	*idp = id;
1357 
1358 	return (0);
1359 }
1360 
1361 static int
1362 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
1363 		   const bus_dma_segment_t *dma_seg, int n_dma_seg,
1364 		   int n_extra_descs)
1365 {
1366 	struct sfxge_tso_state tso;
1367 	unsigned int id;
1368 	unsigned skipped = 0;
1369 
1370 	tso_start(txq, &tso, dma_seg, mbuf);
1371 
1372 	while (dma_seg->ds_len + skipped <= tso.header_len) {
1373 		skipped += dma_seg->ds_len;
1374 		--n_dma_seg;
1375 		KASSERT(n_dma_seg, ("no payload found in TSO packet"));
1376 		++dma_seg;
1377 	}
1378 	tso.in_len = dma_seg->ds_len - (tso.header_len - skipped);
1379 	tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped);
1380 
1381 	id = (txq->added + n_extra_descs) & txq->ptr_mask;
1382 	if (__predict_false(tso_start_new_packet(txq, &tso, &id)))
1383 		return (-1);
1384 
1385 	while (1) {
1386 		tso_fill_packet_with_fragment(txq, &tso);
1387 		/* Exactly one DMA descriptor is added */
1388 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1389 		id = (id + 1) & txq->ptr_mask;
1390 
1391 		/* Move onto the next fragment? */
1392 		if (tso.in_len == 0) {
1393 			--n_dma_seg;
1394 			if (n_dma_seg == 0)
1395 				break;
1396 			++dma_seg;
1397 			tso.in_len = dma_seg->ds_len;
1398 			tso.dma_addr = dma_seg->ds_addr;
1399 		}
1400 
1401 		/* End of packet? */
1402 		if ((tso.packet_space == 0) | (tso.segs_space == 0)) {
1403 			unsigned int n_fatso_opt_desc =
1404 			    (tso.fw_assisted & SFXGE_FATSOV2) ?
1405 			    EFX_TX_FATSOV2_OPT_NDESCS :
1406 			    (tso.fw_assisted & SFXGE_FATSOV1) ? 1 : 0;
1407 
1408 			/* If the queue is now full due to tiny MSS,
1409 			 * or we can't create another header, discard
1410 			 * the remainder of the input mbuf but do not
1411 			 * roll back the work we have done.
1412 			 */
1413 			if (txq->n_pend_desc + n_fatso_opt_desc +
1414 			    1 /* header */ + n_dma_seg > txq->max_pkt_desc) {
1415 				txq->tso_pdrop_too_many++;
1416 				break;
1417 			}
1418 			if (__predict_false(tso_start_new_packet(txq, &tso,
1419 								 &id))) {
1420 				txq->tso_pdrop_no_rsrc++;
1421 				break;
1422 			}
1423 		}
1424 	}
1425 
1426 	txq->tso_bursts++;
1427 	return (id);
1428 }
1429 
1430 static void
1431 sfxge_tx_qunblock(struct sfxge_txq *txq)
1432 {
1433 	struct sfxge_softc *sc;
1434 	struct sfxge_evq *evq __diagused;
1435 
1436 	sc = txq->sc;
1437 	evq = sc->evq[txq->evq_index];
1438 
1439 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
1440 
1441 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED))
1442 		return;
1443 
1444 	SFXGE_TXQ_LOCK(txq);
1445 
1446 	if (txq->blocked) {
1447 		unsigned int level;
1448 
1449 		level = txq->added - txq->completed;
1450 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) {
1451 			/* reaped must be in sync with blocked */
1452 			sfxge_tx_qreap(txq);
1453 			txq->blocked = 0;
1454 		}
1455 	}
1456 
1457 	sfxge_tx_qdpl_service(txq);
1458 	/* note: lock has been dropped */
1459 }
1460 
1461 void
1462 sfxge_tx_qflush_done(struct sfxge_txq *txq)
1463 {
1464 
1465 	txq->flush_state = SFXGE_FLUSH_DONE;
1466 }
1467 
1468 static void
1469 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1470 {
1471 	struct sfxge_txq *txq;
1472 	struct sfxge_evq *evq;
1473 	unsigned int count;
1474 
1475 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1476 
1477 	txq = sc->txq[index];
1478 	evq = sc->evq[txq->evq_index];
1479 
1480 	SFXGE_EVQ_LOCK(evq);
1481 	SFXGE_TXQ_LOCK(txq);
1482 
1483 	KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1484 	    ("txq->init_state != SFXGE_TXQ_STARTED"));
1485 
1486 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1487 
1488 	if (txq->flush_state != SFXGE_FLUSH_DONE) {
1489 		txq->flush_state = SFXGE_FLUSH_PENDING;
1490 
1491 		SFXGE_EVQ_UNLOCK(evq);
1492 		SFXGE_TXQ_UNLOCK(txq);
1493 
1494 		/* Flush the transmit queue. */
1495 		if (efx_tx_qflush(txq->common) != 0) {
1496 			log(LOG_ERR, "%s: Flushing Tx queue %u failed\n",
1497 			    device_get_nameunit(sc->dev), index);
1498 			txq->flush_state = SFXGE_FLUSH_DONE;
1499 		} else {
1500 			count = 0;
1501 			do {
1502 				/* Spin for 100ms. */
1503 				DELAY(100000);
1504 				if (txq->flush_state != SFXGE_FLUSH_PENDING)
1505 					break;
1506 			} while (++count < 20);
1507 		}
1508 		SFXGE_EVQ_LOCK(evq);
1509 		SFXGE_TXQ_LOCK(txq);
1510 
1511 		KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1512 		    ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1513 
1514 		if (txq->flush_state != SFXGE_FLUSH_DONE) {
1515 			/* Flush timeout */
1516 			log(LOG_ERR, "%s: Cannot flush Tx queue %u\n",
1517 			    device_get_nameunit(sc->dev), index);
1518 			txq->flush_state = SFXGE_FLUSH_DONE;
1519 		}
1520 	}
1521 
1522 	txq->blocked = 0;
1523 	txq->pending = txq->added;
1524 
1525 	sfxge_tx_qcomplete(txq, evq);
1526 	KASSERT(txq->completed == txq->added,
1527 	    ("txq->completed != txq->added"));
1528 
1529 	sfxge_tx_qreap(txq);
1530 	KASSERT(txq->reaped == txq->completed,
1531 	    ("txq->reaped != txq->completed"));
1532 
1533 	txq->added = 0;
1534 	txq->pending = 0;
1535 	txq->completed = 0;
1536 	txq->reaped = 0;
1537 
1538 	/* Destroy the common code transmit queue. */
1539 	efx_tx_qdestroy(txq->common);
1540 	txq->common = NULL;
1541 
1542 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1543 	    EFX_TXQ_NBUFS(sc->txq_entries));
1544 
1545 	txq->hw_cksum_flags = 0;
1546 
1547 	SFXGE_EVQ_UNLOCK(evq);
1548 	SFXGE_TXQ_UNLOCK(txq);
1549 }
1550 
1551 /*
1552  * Estimate maximum number of Tx descriptors required for TSO packet.
1553  * With minimum MSS and maximum mbuf length we might need more (even
1554  * than a ring-ful of descriptors), but this should not happen in
1555  * practice except due to deliberate attack.  In that case we will
1556  * truncate the output at a packet boundary.
1557  */
1558 static unsigned int
1559 sfxge_tx_max_pkt_desc(const struct sfxge_softc *sc, enum sfxge_txq_type type,
1560 		      unsigned int tso_fw_assisted)
1561 {
1562 	/* One descriptor for every input fragment */
1563 	unsigned int max_descs = SFXGE_TX_MAPPING_MAX_SEG;
1564 	unsigned int sw_tso_max_descs;
1565 	unsigned int fa_tso_v1_max_descs = 0;
1566 	unsigned int fa_tso_v2_max_descs = 0;
1567 
1568 	/* Checksum offload Tx option descriptor may be required */
1569 	if (sc->txq_dynamic_cksum_toggle_supported)
1570 		max_descs++;
1571 
1572 	/* VLAN tagging Tx option descriptor may be required */
1573 	if (efx_nic_cfg_get(sc->enp)->enc_hw_tx_insert_vlan_enabled)
1574 		max_descs++;
1575 
1576 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) {
1577 		/*
1578 		 * Plus header and payload descriptor for each output segment.
1579 		 * Minus one since header fragment is already counted.
1580 		 * Even if FATSO is used, we should be ready to fallback
1581 		 * to do it in the driver.
1582 		 */
1583 		sw_tso_max_descs = SFXGE_TSO_MAX_SEGS * 2 - 1;
1584 
1585 		/* FW assisted TSOv1 requires one more descriptor per segment
1586 		 * in comparison to SW TSO */
1587 		if (tso_fw_assisted & SFXGE_FATSOV1)
1588 			fa_tso_v1_max_descs =
1589 			    sw_tso_max_descs + SFXGE_TSO_MAX_SEGS;
1590 
1591 		/* FW assisted TSOv2 requires 3 (2 FATSO plus header) extra
1592 		 * descriptors per superframe limited by number of DMA fetches
1593 		 * per packet. The first packet header is already counted.
1594 		 */
1595 		if (tso_fw_assisted & SFXGE_FATSOV2) {
1596 			fa_tso_v2_max_descs =
1597 			    howmany(SFXGE_TX_MAPPING_MAX_SEG,
1598 				    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1) *
1599 			    (EFX_TX_FATSOV2_OPT_NDESCS + 1) - 1;
1600 		}
1601 
1602 		max_descs += MAX(sw_tso_max_descs,
1603 				 MAX(fa_tso_v1_max_descs, fa_tso_v2_max_descs));
1604 	}
1605 
1606 	return (max_descs);
1607 }
1608 
1609 static int
1610 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1611 {
1612 	struct sfxge_txq *txq;
1613 	efsys_mem_t *esmp;
1614 	uint16_t flags;
1615 	unsigned int tso_fw_assisted;
1616 	unsigned int label;
1617 	struct sfxge_evq *evq;
1618 	unsigned int desc_index;
1619 	int rc;
1620 
1621 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1622 
1623 	txq = sc->txq[index];
1624 	esmp = &txq->mem;
1625 	evq = sc->evq[txq->evq_index];
1626 
1627 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1628 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1629 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1630 	    ("evq->init_state != SFXGE_EVQ_STARTED"));
1631 
1632 	/* Program the buffer table. */
1633 	if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1634 	    EFX_TXQ_NBUFS(sc->txq_entries))) != 0)
1635 		return (rc);
1636 
1637 	/* Determine the kind of queue we are creating. */
1638 	tso_fw_assisted = 0;
1639 	switch (txq->type) {
1640 	case SFXGE_TXQ_NON_CKSUM:
1641 		flags = 0;
1642 		break;
1643 	case SFXGE_TXQ_IP_CKSUM:
1644 		flags = EFX_TXQ_CKSUM_IPV4;
1645 		break;
1646 	case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1647 		flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP;
1648 		tso_fw_assisted = sc->tso_fw_assisted;
1649 		if (tso_fw_assisted & SFXGE_FATSOV2)
1650 			flags |= EFX_TXQ_FATSOV2;
1651 		break;
1652 	default:
1653 		KASSERT(0, ("Impossible TX queue"));
1654 		flags = 0;
1655 		break;
1656 	}
1657 
1658 	label = (sc->txq_dynamic_cksum_toggle_supported) ? 0 : txq->type;
1659 
1660 	/* Create the common code transmit queue. */
1661 	if ((rc = efx_tx_qcreate(sc->enp, index, label, esmp,
1662 	    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1663 	    &txq->common, &desc_index)) != 0) {
1664 		/* Retry if no FATSOv2 resources, otherwise fail */
1665 		if ((rc != ENOSPC) || (~flags & EFX_TXQ_FATSOV2))
1666 			goto fail;
1667 
1668 		/* Looks like all FATSOv2 contexts are used */
1669 		flags &= ~EFX_TXQ_FATSOV2;
1670 		tso_fw_assisted &= ~SFXGE_FATSOV2;
1671 		if ((rc = efx_tx_qcreate(sc->enp, index, label, esmp,
1672 		    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1673 		    &txq->common, &desc_index)) != 0)
1674 			goto fail;
1675 	}
1676 
1677 	/* Initialise queue descriptor indexes */
1678 	txq->added = txq->pending = txq->completed = txq->reaped = desc_index;
1679 
1680 	SFXGE_TXQ_LOCK(txq);
1681 
1682 	/* Enable the transmit queue. */
1683 	efx_tx_qenable(txq->common);
1684 
1685 	txq->init_state = SFXGE_TXQ_STARTED;
1686 	txq->flush_state = SFXGE_FLUSH_REQUIRED;
1687 	txq->tso_fw_assisted = tso_fw_assisted;
1688 
1689 	txq->max_pkt_desc = sfxge_tx_max_pkt_desc(sc, txq->type,
1690 						  tso_fw_assisted);
1691 
1692 	txq->hw_vlan_tci = 0;
1693 
1694 	txq->hw_cksum_flags = flags &
1695 			      (EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP);
1696 
1697 	SFXGE_TXQ_UNLOCK(txq);
1698 
1699 	return (0);
1700 
1701 fail:
1702 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1703 	    EFX_TXQ_NBUFS(sc->txq_entries));
1704 	return (rc);
1705 }
1706 
1707 void
1708 sfxge_tx_stop(struct sfxge_softc *sc)
1709 {
1710 	int index;
1711 
1712 	index = sc->txq_count;
1713 	while (--index >= 0)
1714 		sfxge_tx_qstop(sc, index);
1715 
1716 	/* Tear down the transmit module */
1717 	efx_tx_fini(sc->enp);
1718 }
1719 
1720 int
1721 sfxge_tx_start(struct sfxge_softc *sc)
1722 {
1723 	int index;
1724 	int rc;
1725 
1726 	/* Initialize the common code transmit module. */
1727 	if ((rc = efx_tx_init(sc->enp)) != 0)
1728 		return (rc);
1729 
1730 	for (index = 0; index < sc->txq_count; index++) {
1731 		if ((rc = sfxge_tx_qstart(sc, index)) != 0)
1732 			goto fail;
1733 	}
1734 
1735 	return (0);
1736 
1737 fail:
1738 	while (--index >= 0)
1739 		sfxge_tx_qstop(sc, index);
1740 
1741 	efx_tx_fini(sc->enp);
1742 
1743 	return (rc);
1744 }
1745 
1746 static int
1747 sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node)
1748 {
1749 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev);
1750 	struct sysctl_oid *stat_node;
1751 	unsigned int id;
1752 
1753 	stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1754 	    "stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Tx queue statistics");
1755 	if (stat_node == NULL)
1756 		return (ENOMEM);
1757 
1758 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1759 		SYSCTL_ADD_ULONG(
1760 		    ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO,
1761 		    sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS,
1762 		    (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset),
1763 		    "");
1764 	}
1765 
1766 	return (0);
1767 }
1768 
1769 /**
1770  * Destroy a transmit queue.
1771  */
1772 static void
1773 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1774 {
1775 	struct sfxge_txq *txq;
1776 	unsigned int nmaps;
1777 
1778 	txq = sc->txq[index];
1779 
1780 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1781 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1782 
1783 	if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1784 		tso_fini(txq);
1785 
1786 	/* Free the context arrays. */
1787 	free(txq->pend_desc, M_SFXGE);
1788 	nmaps = sc->txq_entries;
1789 	while (nmaps-- != 0)
1790 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1791 	free(txq->stmp, M_SFXGE);
1792 
1793 	/* Release DMA memory mapping. */
1794 	sfxge_dma_free(&txq->mem);
1795 
1796 	sc->txq[index] = NULL;
1797 
1798 	SFXGE_TXQ_LOCK_DESTROY(txq);
1799 
1800 	free(txq, M_SFXGE);
1801 }
1802 
1803 static int
1804 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1805 	       enum sfxge_txq_type type, unsigned int evq_index)
1806 {
1807 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
1808 	char name[16];
1809 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1810 	struct sysctl_oid *txq_node;
1811 	struct sfxge_txq *txq;
1812 	struct sfxge_tx_dpl *stdp;
1813 	struct sysctl_oid *dpl_node;
1814 	efsys_mem_t *esmp;
1815 	unsigned int nmaps;
1816 	int rc;
1817 
1818 	txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1819 	txq->sc = sc;
1820 	txq->entries = sc->txq_entries;
1821 	txq->ptr_mask = txq->entries - 1;
1822 
1823 	sc->txq[txq_index] = txq;
1824 	esmp = &txq->mem;
1825 
1826 	/* Allocate and zero DMA space for the descriptor ring. */
1827 	if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0)
1828 		return (rc);
1829 
1830 	/* Allocate buffer table entries. */
1831 	sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries),
1832 				 &txq->buf_base_id);
1833 
1834 	/* Create a DMA tag for packet mappings. */
1835 	if (bus_dma_tag_create(sc->parent_dma_tag, 1,
1836 	    encp->enc_tx_dma_desc_boundary,
1837 	    MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1838 	    NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG,
1839 	    encp->enc_tx_dma_desc_size_max, 0, NULL, NULL,
1840 	    &txq->packet_dma_tag) != 0) {
1841 		device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1842 		rc = ENOMEM;
1843 		goto fail;
1844 	}
1845 
1846 	/* Allocate pending descriptor array for batching writes. */
1847 	txq->pend_desc = malloc(sizeof(efx_desc_t) * sc->txq_entries,
1848 				M_SFXGE, M_ZERO | M_WAITOK);
1849 
1850 	/* Allocate and initialise mbuf DMA mapping array. */
1851 	txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries,
1852 	    M_SFXGE, M_ZERO | M_WAITOK);
1853 	for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) {
1854 		rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1855 				       &txq->stmp[nmaps].map);
1856 		if (rc != 0)
1857 			goto fail2;
1858 	}
1859 
1860 	snprintf(name, sizeof(name), "%u", txq_index);
1861 	txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node),
1862 	    OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
1863 	if (txq_node == NULL) {
1864 		rc = ENOMEM;
1865 		goto fail_txq_node;
1866 	}
1867 
1868 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1869 	    (rc = tso_init(txq)) != 0)
1870 		goto fail3;
1871 
1872 	/* Initialize the deferred packet list. */
1873 	stdp = &txq->dpl;
1874 	stdp->std_put_max = sfxge_tx_dpl_put_max;
1875 	stdp->std_get_max = sfxge_tx_dpl_get_max;
1876 	stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max;
1877 	stdp->std_getp = &stdp->std_get;
1878 
1879 	SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index);
1880 
1881 	dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1882 	    "dpl", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
1883 	    "Deferred packet list statistics");
1884 	if (dpl_node == NULL) {
1885 		rc = ENOMEM;
1886 		goto fail_dpl_node;
1887 	}
1888 
1889 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1890 			"get_count", CTLFLAG_RD | CTLFLAG_STATS,
1891 			&stdp->std_get_count, 0, "");
1892 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1893 			"get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS,
1894 			&stdp->std_get_non_tcp_count, 0, "");
1895 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1896 			"get_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1897 			&stdp->std_get_hiwat, 0, "");
1898 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1899 			"put_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1900 			&stdp->std_put_hiwat, 0, "");
1901 
1902 	rc = sfxge_txq_stat_init(txq, txq_node);
1903 	if (rc != 0)
1904 		goto fail_txq_stat_init;
1905 
1906 	txq->type = type;
1907 	txq->evq_index = evq_index;
1908 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1909 
1910 	return (0);
1911 
1912 fail_txq_stat_init:
1913 fail_dpl_node:
1914 fail3:
1915 fail_txq_node:
1916 	free(txq->pend_desc, M_SFXGE);
1917 fail2:
1918 	while (nmaps-- != 0)
1919 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1920 	free(txq->stmp, M_SFXGE);
1921 	bus_dma_tag_destroy(txq->packet_dma_tag);
1922 
1923 fail:
1924 	sfxge_dma_free(esmp);
1925 
1926 	return (rc);
1927 }
1928 
1929 static int
1930 sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1931 {
1932 	struct sfxge_softc *sc = arg1;
1933 	unsigned int id = arg2;
1934 	unsigned long sum;
1935 	unsigned int index;
1936 
1937 	/* Sum across all TX queues */
1938 	sum = 0;
1939 	for (index = 0; index < sc->txq_count; index++)
1940 		sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1941 					  sfxge_tx_stats[id].offset);
1942 
1943 	return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1944 }
1945 
1946 static void
1947 sfxge_tx_stat_init(struct sfxge_softc *sc)
1948 {
1949 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1950 	struct sysctl_oid_list *stat_list;
1951 	unsigned int id;
1952 
1953 	stat_list = SYSCTL_CHILDREN(sc->stats_node);
1954 
1955 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1956 		SYSCTL_ADD_PROC(ctx, stat_list, OID_AUTO,
1957 		    sfxge_tx_stats[id].name,
1958 		    CTLTYPE_ULONG | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
1959 		    sc, id, sfxge_tx_stat_handler, "LU", "");
1960 	}
1961 }
1962 
1963 uint64_t
1964 sfxge_tx_get_drops(struct sfxge_softc *sc)
1965 {
1966 	unsigned int index;
1967 	uint64_t drops = 0;
1968 	struct sfxge_txq *txq;
1969 
1970 	/* Sum across all TX queues */
1971 	for (index = 0; index < sc->txq_count; index++) {
1972 		txq = sc->txq[index];
1973 		/*
1974 		 * In theory, txq->put_overflow and txq->netdown_drops
1975 		 * should use atomic operation and other should be
1976 		 * obtained under txq lock, but it is just statistics.
1977 		 */
1978 		drops += txq->drops + txq->get_overflow +
1979 			 txq->get_non_tcp_overflow +
1980 			 txq->put_overflow + txq->netdown_drops +
1981 			 txq->tso_pdrop_too_many + txq->tso_pdrop_no_rsrc;
1982 	}
1983 	return (drops);
1984 }
1985 
1986 void
1987 sfxge_tx_fini(struct sfxge_softc *sc)
1988 {
1989 	int index;
1990 
1991 	index = sc->txq_count;
1992 	while (--index >= 0)
1993 		sfxge_tx_qfini(sc, index);
1994 
1995 	sc->txq_count = 0;
1996 }
1997 
1998 int
1999 sfxge_tx_init(struct sfxge_softc *sc)
2000 {
2001 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
2002 	struct sfxge_intr *intr __diagused;
2003 	int index;
2004 	int rc;
2005 
2006 	intr = &sc->intr;
2007 
2008 	KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
2009 	    ("intr->state != SFXGE_INTR_INITIALIZED"));
2010 
2011 	if (sfxge_tx_dpl_get_max <= 0) {
2012 		log(LOG_ERR, "%s=%d must be greater than 0",
2013 		    SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max);
2014 		rc = EINVAL;
2015 		goto fail_tx_dpl_get_max;
2016 	}
2017 	if (sfxge_tx_dpl_get_non_tcp_max <= 0) {
2018 		log(LOG_ERR, "%s=%d must be greater than 0",
2019 		    SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX,
2020 		    sfxge_tx_dpl_get_non_tcp_max);
2021 		rc = EINVAL;
2022 		goto fail_tx_dpl_get_non_tcp_max;
2023 	}
2024 	if (sfxge_tx_dpl_put_max < 0) {
2025 		log(LOG_ERR, "%s=%d must be greater or equal to 0",
2026 		    SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max);
2027 		rc = EINVAL;
2028 		goto fail_tx_dpl_put_max;
2029 	}
2030 
2031 	sc->txq_count = SFXGE_EVQ0_N_TXQ(sc) - 1 + sc->intr.n_alloc;
2032 
2033 	sc->tso_fw_assisted = sfxge_tso_fw_assisted;
2034 	if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO) ||
2035 	    (!encp->enc_fw_assisted_tso_enabled))
2036 		sc->tso_fw_assisted &= ~SFXGE_FATSOV1;
2037 	if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO_V2) ||
2038 	    (!encp->enc_fw_assisted_tso_v2_enabled))
2039 		sc->tso_fw_assisted &= ~SFXGE_FATSOV2;
2040 
2041 	sc->txqs_node = SYSCTL_ADD_NODE(device_get_sysctl_ctx(sc->dev),
2042 	    SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)), OID_AUTO,
2043 	    "txq", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Tx queues");
2044 	if (sc->txqs_node == NULL) {
2045 		rc = ENOMEM;
2046 		goto fail_txq_node;
2047 	}
2048 
2049 	/* Initialize the transmit queues */
2050 	if (sc->txq_dynamic_cksum_toggle_supported == B_FALSE) {
2051 		if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
2052 		    SFXGE_TXQ_NON_CKSUM, 0)) != 0)
2053 			goto fail;
2054 
2055 		if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
2056 		    SFXGE_TXQ_IP_CKSUM, 0)) != 0)
2057 			goto fail2;
2058 	}
2059 
2060 	for (index = 0;
2061 	     index < sc->txq_count - SFXGE_EVQ0_N_TXQ(sc) + 1;
2062 	     index++) {
2063 		if ((rc = sfxge_tx_qinit(sc, SFXGE_EVQ0_N_TXQ(sc) - 1 + index,
2064 		    SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
2065 			goto fail3;
2066 	}
2067 
2068 	sfxge_tx_stat_init(sc);
2069 
2070 	return (0);
2071 
2072 fail3:
2073 	while (--index >= 0)
2074 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
2075 
2076 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
2077 
2078 fail2:
2079 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
2080 
2081 fail:
2082 fail_txq_node:
2083 	sc->txq_count = 0;
2084 fail_tx_dpl_put_max:
2085 fail_tx_dpl_get_non_tcp_max:
2086 fail_tx_dpl_get_max:
2087 	return (rc);
2088 }
2089