xref: /freebsd/sys/dev/sfxge/sfxge_tx.c (revision 037479ff5ee18977b1c48e1e59770aad2f200a5a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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
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 
51 #include <sys/cdefs.h>
52 __FBSDID("$FreeBSD$");
53 
54 #include "opt_rss.h"
55 
56 #include <sys/param.h>
57 #include <sys/malloc.h>
58 #include <sys/mbuf.h>
59 #include <sys/smp.h>
60 #include <sys/socket.h>
61 #include <sys/sysctl.h>
62 #include <sys/syslog.h>
63 #include <sys/limits.h>
64 
65 #include <net/bpf.h>
66 #include <net/ethernet.h>
67 #include <net/if.h>
68 #include <net/if_vlan_var.h>
69 
70 #include <netinet/in.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 
75 #ifdef RSS
76 #include <net/rss_config.h>
77 #endif
78 
79 #include "common/efx.h"
80 
81 #include "sfxge.h"
82 #include "sfxge_tx.h"
83 
84 
85 #define	SFXGE_PARAM_TX_DPL_GET_MAX	SFXGE_PARAM(tx_dpl_get_max)
86 static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT;
87 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max);
88 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN,
89 	   &sfxge_tx_dpl_get_max, 0,
90 	   "Maximum number of any packets in deferred packet get-list");
91 
92 #define	SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \
93 	SFXGE_PARAM(tx_dpl_get_non_tcp_max)
94 static int sfxge_tx_dpl_get_non_tcp_max =
95 	SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT;
96 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max);
97 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN,
98 	   &sfxge_tx_dpl_get_non_tcp_max, 0,
99 	   "Maximum number of non-TCP packets in deferred packet get-list");
100 
101 #define	SFXGE_PARAM_TX_DPL_PUT_MAX	SFXGE_PARAM(tx_dpl_put_max)
102 static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT;
103 TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max);
104 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN,
105 	   &sfxge_tx_dpl_put_max, 0,
106 	   "Maximum number of any packets in deferred packet put-list");
107 
108 #define	SFXGE_PARAM_TSO_FW_ASSISTED	SFXGE_PARAM(tso_fw_assisted)
109 static int sfxge_tso_fw_assisted = (SFXGE_FATSOV1 | SFXGE_FATSOV2);
110 TUNABLE_INT(SFXGE_PARAM_TSO_FW_ASSISTED, &sfxge_tso_fw_assisted);
111 SYSCTL_INT(_hw_sfxge, OID_AUTO, tso_fw_assisted, CTLFLAG_RDTUN,
112 	   &sfxge_tso_fw_assisted, 0,
113 	   "Bitmask of FW-assisted TSO allowed to use if supported by NIC firmware");
114 
115 
116 static const struct {
117 	const char *name;
118 	size_t offset;
119 } sfxge_tx_stats[] = {
120 #define	SFXGE_TX_STAT(name, member) \
121 	{ #name, offsetof(struct sfxge_txq, member) }
122 	SFXGE_TX_STAT(tso_bursts, tso_bursts),
123 	SFXGE_TX_STAT(tso_packets, tso_packets),
124 	SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
125 	SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many),
126 	SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc),
127 	SFXGE_TX_STAT(tx_collapses, collapses),
128 	SFXGE_TX_STAT(tx_drops, drops),
129 	SFXGE_TX_STAT(tx_get_overflow, get_overflow),
130 	SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow),
131 	SFXGE_TX_STAT(tx_put_overflow, put_overflow),
132 	SFXGE_TX_STAT(tx_netdown_drops, netdown_drops),
133 };
134 
135 
136 /* Forward declarations. */
137 static void sfxge_tx_qdpl_service(struct sfxge_txq *txq);
138 static void sfxge_tx_qlist_post(struct sfxge_txq *txq);
139 static void sfxge_tx_qunblock(struct sfxge_txq *txq);
140 static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
141 			      const bus_dma_segment_t *dma_seg, int n_dma_seg,
142 			      int vlan_tagged);
143 
144 static int
145 sfxge_tx_maybe_insert_tag(struct sfxge_txq *txq, struct mbuf *mbuf)
146 {
147 	uint16_t this_tag = ((mbuf->m_flags & M_VLANTAG) ?
148 			     mbuf->m_pkthdr.ether_vtag :
149 			     0);
150 
151 	if (this_tag == txq->hw_vlan_tci)
152 		return (0);
153 
154 	efx_tx_qdesc_vlantci_create(txq->common,
155 				    bswap16(this_tag),
156 				    &txq->pend_desc[0]);
157 	txq->n_pend_desc = 1;
158 	txq->hw_vlan_tci = this_tag;
159 	return (1);
160 }
161 
162 static inline void
163 sfxge_next_stmp(struct sfxge_txq *txq, struct sfxge_tx_mapping **pstmp)
164 {
165 	KASSERT((*pstmp)->flags == 0, ("stmp flags are not 0"));
166 	if (__predict_false(*pstmp ==
167 			    &txq->stmp[txq->ptr_mask]))
168 		*pstmp = &txq->stmp[0];
169 	else
170 		(*pstmp)++;
171 }
172 
173 
174 void
175 sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq)
176 {
177 	unsigned int completed;
178 
179 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
180 
181 	completed = txq->completed;
182 	while (completed != txq->pending) {
183 		struct sfxge_tx_mapping *stmp;
184 		unsigned int id;
185 
186 		id = completed++ & txq->ptr_mask;
187 
188 		stmp = &txq->stmp[id];
189 		if (stmp->flags & TX_BUF_UNMAP) {
190 			bus_dmamap_unload(txq->packet_dma_tag, stmp->map);
191 			if (stmp->flags & TX_BUF_MBUF) {
192 				struct mbuf *m = stmp->u.mbuf;
193 				do
194 					m = m_free(m);
195 				while (m != NULL);
196 			} else {
197 				free(stmp->u.heap_buf, M_SFXGE);
198 			}
199 			stmp->flags = 0;
200 		}
201 	}
202 	txq->completed = completed;
203 
204 	/* Check whether we need to unblock the queue. */
205 	mb();
206 	if (txq->blocked) {
207 		unsigned int level;
208 
209 		level = txq->added - txq->completed;
210 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
211 			sfxge_tx_qunblock(txq);
212 	}
213 }
214 
215 static unsigned int
216 sfxge_is_mbuf_non_tcp(struct mbuf *mbuf)
217 {
218 	/* Absence of TCP checksum flags does not mean that it is non-TCP
219 	 * but it should be true if user wants to achieve high throughput.
220 	 */
221 	return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP)));
222 }
223 
224 /*
225  * Reorder the put list and append it to the get list.
226  */
227 static void
228 sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq)
229 {
230 	struct sfxge_tx_dpl *stdp;
231 	struct mbuf *mbuf, *get_next, **get_tailp;
232 	volatile uintptr_t *putp;
233 	uintptr_t put;
234 	unsigned int count;
235 	unsigned int non_tcp_count;
236 
237 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
238 
239 	stdp = &txq->dpl;
240 
241 	/* Acquire the put list. */
242 	putp = &stdp->std_put;
243 	put = atomic_readandclear_ptr(putp);
244 	mbuf = (void *)put;
245 
246 	if (mbuf == NULL)
247 		return;
248 
249 	/* Reverse the put list. */
250 	get_tailp = &mbuf->m_nextpkt;
251 	get_next = NULL;
252 
253 	count = 0;
254 	non_tcp_count = 0;
255 	do {
256 		struct mbuf *put_next;
257 
258 		non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf);
259 		put_next = mbuf->m_nextpkt;
260 		mbuf->m_nextpkt = get_next;
261 		get_next = mbuf;
262 		mbuf = put_next;
263 
264 		count++;
265 	} while (mbuf != NULL);
266 
267 	if (count > stdp->std_put_hiwat)
268 		stdp->std_put_hiwat = count;
269 
270 	/* Append the reversed put list to the get list. */
271 	KASSERT(*get_tailp == NULL, ("*get_tailp != NULL"));
272 	*stdp->std_getp = get_next;
273 	stdp->std_getp = get_tailp;
274 	stdp->std_get_count += count;
275 	stdp->std_get_non_tcp_count += non_tcp_count;
276 }
277 
278 static void
279 sfxge_tx_qreap(struct sfxge_txq *txq)
280 {
281 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
282 
283 	txq->reaped = txq->completed;
284 }
285 
286 static void
287 sfxge_tx_qlist_post(struct sfxge_txq *txq)
288 {
289 	unsigned int old_added;
290 	unsigned int block_level;
291 	unsigned int level;
292 	int rc;
293 
294 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
295 
296 	KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0"));
297 	KASSERT(txq->n_pend_desc <= txq->max_pkt_desc,
298 		("txq->n_pend_desc too large"));
299 	KASSERT(!txq->blocked, ("txq->blocked"));
300 
301 	old_added = txq->added;
302 
303 	/* Post the fragment list. */
304 	rc = efx_tx_qdesc_post(txq->common, txq->pend_desc, txq->n_pend_desc,
305 			  txq->reaped, &txq->added);
306 	KASSERT(rc == 0, ("efx_tx_qdesc_post() failed"));
307 
308 	/* If efx_tx_qdesc_post() had to refragment, our information about
309 	 * buffers to free may be associated with the wrong
310 	 * descriptors.
311 	 */
312 	KASSERT(txq->added - old_added == txq->n_pend_desc,
313 		("efx_tx_qdesc_post() refragmented descriptors"));
314 
315 	level = txq->added - txq->reaped;
316 	KASSERT(level <= txq->entries, ("overfilled TX queue"));
317 
318 	/* Clear the fragment list. */
319 	txq->n_pend_desc = 0;
320 
321 	/*
322 	 * Set the block level to ensure there is space to generate a
323 	 * large number of descriptors for TSO.
324 	 */
325 	block_level = EFX_TXQ_LIMIT(txq->entries) - txq->max_pkt_desc;
326 
327 	/* Have we reached the block level? */
328 	if (level < block_level)
329 		return;
330 
331 	/* Reap, and check again */
332 	sfxge_tx_qreap(txq);
333 	level = txq->added - txq->reaped;
334 	if (level < block_level)
335 		return;
336 
337 	txq->blocked = 1;
338 
339 	/*
340 	 * Avoid a race with completion interrupt handling that could leave
341 	 * the queue blocked.
342 	 */
343 	mb();
344 	sfxge_tx_qreap(txq);
345 	level = txq->added - txq->reaped;
346 	if (level < block_level) {
347 		mb();
348 		txq->blocked = 0;
349 	}
350 }
351 
352 static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
353 {
354 	bus_dmamap_t *used_map;
355 	bus_dmamap_t map;
356 	bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG];
357 	unsigned int id;
358 	struct sfxge_tx_mapping *stmp;
359 	efx_desc_t *desc;
360 	int n_dma_seg;
361 	int rc;
362 	int i;
363 	int eop;
364 	uint16_t hw_vlan_tci_prev;
365 	int vlan_tagged;
366 
367 	KASSERT(!txq->blocked, ("txq->blocked"));
368 
369 #if SFXGE_TX_PARSE_EARLY
370 	/*
371 	 * If software TSO is used, we still need to copy packet header,
372 	 * even if we have already parsed it early before enqueue.
373 	 */
374 	if ((mbuf->m_pkthdr.csum_flags & CSUM_TSO) &&
375 	    (txq->tso_fw_assisted == 0))
376 		prefetch_read_many(mbuf->m_data);
377 #else
378 	/*
379 	 * Prefetch packet header since we need to parse it and extract
380 	 * IP ID, TCP sequence number and flags.
381 	 */
382 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
383 		prefetch_read_many(mbuf->m_data);
384 #endif
385 
386 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) {
387 		rc = EINTR;
388 		goto reject;
389 	}
390 
391 	/* Load the packet for DMA. */
392 	id = txq->added & txq->ptr_mask;
393 	stmp = &txq->stmp[id];
394 	rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map,
395 				     mbuf, dma_seg, &n_dma_seg, 0);
396 	if (rc == EFBIG) {
397 		/* Try again. */
398 		struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT,
399 						   SFXGE_TX_MAPPING_MAX_SEG);
400 		if (new_mbuf == NULL)
401 			goto reject;
402 		++txq->collapses;
403 		mbuf = new_mbuf;
404 		rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag,
405 					     stmp->map, mbuf,
406 					     dma_seg, &n_dma_seg, 0);
407 	}
408 	if (rc != 0)
409 		goto reject;
410 
411 	/* Make the packet visible to the hardware. */
412 	bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE);
413 
414 	used_map = &stmp->map;
415 
416 	hw_vlan_tci_prev = txq->hw_vlan_tci;
417 
418 	vlan_tagged = sfxge_tx_maybe_insert_tag(txq, mbuf);
419 	if (vlan_tagged) {
420 		sfxge_next_stmp(txq, &stmp);
421 	}
422 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
423 		rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg, vlan_tagged);
424 		if (rc < 0)
425 			goto reject_mapped;
426 		stmp = &txq->stmp[(rc - 1) & txq->ptr_mask];
427 	} else {
428 		/* Add the mapping to the fragment list, and set flags
429 		 * for the buffer.
430 		 */
431 
432 		i = 0;
433 		for (;;) {
434 			desc = &txq->pend_desc[i + vlan_tagged];
435 			eop = (i == n_dma_seg - 1);
436 			efx_tx_qdesc_dma_create(txq->common,
437 						dma_seg[i].ds_addr,
438 						dma_seg[i].ds_len,
439 						eop,
440 						desc);
441 			if (eop)
442 				break;
443 			i++;
444 			sfxge_next_stmp(txq, &stmp);
445 		}
446 		txq->n_pend_desc = n_dma_seg + vlan_tagged;
447 	}
448 
449 	/*
450 	 * If the mapping required more than one descriptor
451 	 * then we need to associate the DMA map with the last
452 	 * descriptor, not the first.
453 	 */
454 	if (used_map != &stmp->map) {
455 		map = stmp->map;
456 		stmp->map = *used_map;
457 		*used_map = map;
458 	}
459 
460 	stmp->u.mbuf = mbuf;
461 	stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF;
462 
463 	/* Post the fragment list. */
464 	sfxge_tx_qlist_post(txq);
465 
466 	return (0);
467 
468 reject_mapped:
469 	txq->hw_vlan_tci = hw_vlan_tci_prev;
470 	bus_dmamap_unload(txq->packet_dma_tag, *used_map);
471 reject:
472 	/* Drop the packet on the floor. */
473 	m_freem(mbuf);
474 	++txq->drops;
475 
476 	return (rc);
477 }
478 
479 /*
480  * Drain the deferred packet list into the transmit queue.
481  */
482 static void
483 sfxge_tx_qdpl_drain(struct sfxge_txq *txq)
484 {
485 	struct sfxge_softc *sc;
486 	struct sfxge_tx_dpl *stdp;
487 	struct mbuf *mbuf, *next;
488 	unsigned int count;
489 	unsigned int non_tcp_count;
490 	unsigned int pushed;
491 	int rc;
492 
493 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
494 
495 	sc = txq->sc;
496 	stdp = &txq->dpl;
497 	pushed = txq->added;
498 
499 	if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) {
500 		prefetch_read_many(sc->enp);
501 		prefetch_read_many(txq->common);
502 	}
503 
504 	mbuf = stdp->std_get;
505 	count = stdp->std_get_count;
506 	non_tcp_count = stdp->std_get_non_tcp_count;
507 
508 	if (count > stdp->std_get_hiwat)
509 		stdp->std_get_hiwat = count;
510 
511 	while (count != 0) {
512 		KASSERT(mbuf != NULL, ("mbuf == NULL"));
513 
514 		next = mbuf->m_nextpkt;
515 		mbuf->m_nextpkt = NULL;
516 
517 		ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */
518 
519 		if (next != NULL)
520 			prefetch_read_many(next);
521 
522 		rc = sfxge_tx_queue_mbuf(txq, mbuf);
523 		--count;
524 		non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf);
525 		mbuf = next;
526 		if (rc != 0)
527 			continue;
528 
529 		if (txq->blocked)
530 			break;
531 
532 		/* Push the fragments to the hardware in batches. */
533 		if (txq->added - pushed >= SFXGE_TX_BATCH) {
534 			efx_tx_qpush(txq->common, txq->added, pushed);
535 			pushed = txq->added;
536 		}
537 	}
538 
539 	if (count == 0) {
540 		KASSERT(mbuf == NULL, ("mbuf != NULL"));
541 		KASSERT(non_tcp_count == 0,
542 			("inconsistent TCP/non-TCP detection"));
543 		stdp->std_get = NULL;
544 		stdp->std_get_count = 0;
545 		stdp->std_get_non_tcp_count = 0;
546 		stdp->std_getp = &stdp->std_get;
547 	} else {
548 		stdp->std_get = mbuf;
549 		stdp->std_get_count = count;
550 		stdp->std_get_non_tcp_count = non_tcp_count;
551 	}
552 
553 	if (txq->added != pushed)
554 		efx_tx_qpush(txq->common, txq->added, pushed);
555 
556 	KASSERT(txq->blocked || stdp->std_get_count == 0,
557 		("queue unblocked but count is non-zero"));
558 }
559 
560 #define	SFXGE_TX_QDPL_PENDING(_txq)	((_txq)->dpl.std_put != 0)
561 
562 /*
563  * Service the deferred packet list.
564  *
565  * NOTE: drops the txq mutex!
566  */
567 static void
568 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
569 {
570 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
571 
572 	do {
573 		if (SFXGE_TX_QDPL_PENDING(txq))
574 			sfxge_tx_qdpl_swizzle(txq);
575 
576 		if (!txq->blocked)
577 			sfxge_tx_qdpl_drain(txq);
578 
579 		SFXGE_TXQ_UNLOCK(txq);
580 	} while (SFXGE_TX_QDPL_PENDING(txq) &&
581 		 SFXGE_TXQ_TRYLOCK(txq));
582 }
583 
584 /*
585  * Put a packet on the deferred packet get-list.
586  */
587 static int
588 sfxge_tx_qdpl_put_locked(struct sfxge_txq *txq, struct mbuf *mbuf)
589 {
590 	struct sfxge_tx_dpl *stdp;
591 
592 	stdp = &txq->dpl;
593 
594 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
595 
596 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
597 
598 	if (stdp->std_get_count >= stdp->std_get_max) {
599 		txq->get_overflow++;
600 		return (ENOBUFS);
601 	}
602 	if (sfxge_is_mbuf_non_tcp(mbuf)) {
603 		if (stdp->std_get_non_tcp_count >=
604 		    stdp->std_get_non_tcp_max) {
605 			txq->get_non_tcp_overflow++;
606 			return (ENOBUFS);
607 		}
608 		stdp->std_get_non_tcp_count++;
609 	}
610 
611 	*(stdp->std_getp) = mbuf;
612 	stdp->std_getp = &mbuf->m_nextpkt;
613 	stdp->std_get_count++;
614 
615 	return (0);
616 }
617 
618 /*
619  * Put a packet on the deferred packet put-list.
620  *
621  * We overload the csum_data field in the mbuf to keep track of this length
622  * because there is no cheap alternative to avoid races.
623  */
624 static int
625 sfxge_tx_qdpl_put_unlocked(struct sfxge_txq *txq, struct mbuf *mbuf)
626 {
627 	struct sfxge_tx_dpl *stdp;
628 	volatile uintptr_t *putp;
629 	uintptr_t old;
630 	uintptr_t new;
631 	unsigned int put_count;
632 
633 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
634 
635 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
636 
637 	stdp = &txq->dpl;
638 	putp = &stdp->std_put;
639 	new = (uintptr_t)mbuf;
640 
641 	do {
642 		old = *putp;
643 		if (old != 0) {
644 			struct mbuf *mp = (struct mbuf *)old;
645 			put_count = mp->m_pkthdr.csum_data;
646 		} else
647 			put_count = 0;
648 		if (put_count >= stdp->std_put_max) {
649 			atomic_add_long(&txq->put_overflow, 1);
650 			return (ENOBUFS);
651 		}
652 		mbuf->m_pkthdr.csum_data = put_count + 1;
653 		mbuf->m_nextpkt = (void *)old;
654 	} while (atomic_cmpset_ptr(putp, old, new) == 0);
655 
656 	return (0);
657 }
658 
659 /*
660  * Called from if_transmit - will try to grab the txq lock and enqueue to the
661  * put list if it succeeds, otherwise try to push onto the defer list if space.
662  */
663 static int
664 sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m)
665 {
666 	int rc;
667 
668 	if (!SFXGE_LINK_UP(txq->sc)) {
669 		atomic_add_long(&txq->netdown_drops, 1);
670 		return (ENETDOWN);
671 	}
672 
673 	/*
674 	 * Try to grab the txq lock.  If we are able to get the lock,
675 	 * the packet will be appended to the "get list" of the deferred
676 	 * packet list.  Otherwise, it will be pushed on the "put list".
677 	 */
678 	if (SFXGE_TXQ_TRYLOCK(txq)) {
679 		/* First swizzle put-list to get-list to keep order */
680 		sfxge_tx_qdpl_swizzle(txq);
681 
682 		rc = sfxge_tx_qdpl_put_locked(txq, m);
683 
684 		/* Try to service the list. */
685 		sfxge_tx_qdpl_service(txq);
686 		/* Lock has been dropped. */
687 	} else {
688 		rc = sfxge_tx_qdpl_put_unlocked(txq, m);
689 
690 		/*
691 		 * Try to grab the lock again.
692 		 *
693 		 * If we are able to get the lock, we need to process
694 		 * the deferred packet list.  If we are not able to get
695 		 * the lock, another thread is processing the list.
696 		 */
697 		if ((rc == 0) && SFXGE_TXQ_TRYLOCK(txq)) {
698 			sfxge_tx_qdpl_service(txq);
699 			/* Lock has been dropped. */
700 		}
701 	}
702 
703 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
704 
705 	return (rc);
706 }
707 
708 static void
709 sfxge_tx_qdpl_flush(struct sfxge_txq *txq)
710 {
711 	struct sfxge_tx_dpl *stdp = &txq->dpl;
712 	struct mbuf *mbuf, *next;
713 
714 	SFXGE_TXQ_LOCK(txq);
715 
716 	sfxge_tx_qdpl_swizzle(txq);
717 	for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) {
718 		next = mbuf->m_nextpkt;
719 		m_freem(mbuf);
720 	}
721 	stdp->std_get = NULL;
722 	stdp->std_get_count = 0;
723 	stdp->std_get_non_tcp_count = 0;
724 	stdp->std_getp = &stdp->std_get;
725 
726 	SFXGE_TXQ_UNLOCK(txq);
727 }
728 
729 void
730 sfxge_if_qflush(struct ifnet *ifp)
731 {
732 	struct sfxge_softc *sc;
733 	unsigned int i;
734 
735 	sc = ifp->if_softc;
736 
737 	for (i = 0; i < sc->txq_count; i++)
738 		sfxge_tx_qdpl_flush(sc->txq[i]);
739 }
740 
741 #if SFXGE_TX_PARSE_EARLY
742 
743 /* There is little space for user data in mbuf pkthdr, so we
744  * use l*hlen fields which are not used by the driver otherwise
745  * to store header offsets.
746  * The fields are 8-bit, but it's ok, no header may be longer than 255 bytes.
747  */
748 
749 
750 #define TSO_MBUF_PROTO(_mbuf)    ((_mbuf)->m_pkthdr.PH_loc.sixteen[0])
751 /* We abuse l5hlen here because PH_loc can hold only 64 bits of data */
752 #define TSO_MBUF_FLAGS(_mbuf)    ((_mbuf)->m_pkthdr.l5hlen)
753 #define TSO_MBUF_PACKETID(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[1])
754 #define TSO_MBUF_SEQNUM(_mbuf)   ((_mbuf)->m_pkthdr.PH_loc.thirtytwo[1])
755 
756 static void sfxge_parse_tx_packet(struct mbuf *mbuf)
757 {
758 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
759 	const struct tcphdr *th;
760 	struct tcphdr th_copy;
761 
762 	/* Find network protocol and header */
763 	TSO_MBUF_PROTO(mbuf) = eh->ether_type;
764 	if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_VLAN)) {
765 		struct ether_vlan_header *veh =
766 			mtod(mbuf, struct ether_vlan_header *);
767 		TSO_MBUF_PROTO(mbuf) = veh->evl_proto;
768 		mbuf->m_pkthdr.l2hlen = sizeof(*veh);
769 	} else {
770 		mbuf->m_pkthdr.l2hlen = sizeof(*eh);
771 	}
772 
773 	/* Find TCP header */
774 	if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IP)) {
775 		const struct ip *iph = (const struct ip *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen);
776 
777 		KASSERT(iph->ip_p == IPPROTO_TCP,
778 			("TSO required on non-TCP packet"));
779 		mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + 4 * iph->ip_hl;
780 		TSO_MBUF_PACKETID(mbuf) = iph->ip_id;
781 	} else {
782 		KASSERT(TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IPV6),
783 			("TSO required on non-IP packet"));
784 		KASSERT(((const struct ip6_hdr *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen))->ip6_nxt ==
785 			IPPROTO_TCP,
786 			("TSO required on non-TCP packet"));
787 		mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + sizeof(struct ip6_hdr);
788 		TSO_MBUF_PACKETID(mbuf) = 0;
789 	}
790 
791 	KASSERT(mbuf->m_len >= mbuf->m_pkthdr.l3hlen,
792 		("network header is fragmented in mbuf"));
793 
794 	/* We need TCP header including flags (window is the next) */
795 	if (mbuf->m_len < mbuf->m_pkthdr.l3hlen + offsetof(struct tcphdr, th_win)) {
796 		m_copydata(mbuf, mbuf->m_pkthdr.l3hlen, sizeof(th_copy),
797 			   (caddr_t)&th_copy);
798 		th = &th_copy;
799 	} else {
800 		th = (const struct tcphdr *)mtodo(mbuf, mbuf->m_pkthdr.l3hlen);
801 	}
802 
803 	mbuf->m_pkthdr.l4hlen = mbuf->m_pkthdr.l3hlen + 4 * th->th_off;
804 	TSO_MBUF_SEQNUM(mbuf) = ntohl(th->th_seq);
805 
806 	/* These flags must not be duplicated */
807 	/*
808 	 * RST should not be duplicated as well, but FreeBSD kernel
809 	 * generates TSO packets with RST flag. So, do not assert
810 	 * its absence.
811 	 */
812 	KASSERT(!(th->th_flags & (TH_URG | TH_SYN)),
813 		("incompatible TCP flag 0x%x on TSO packet",
814 		 th->th_flags & (TH_URG | TH_SYN)));
815 	TSO_MBUF_FLAGS(mbuf) = th->th_flags;
816 }
817 #endif
818 
819 /*
820  * TX start -- called by the stack.
821  */
822 int
823 sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m)
824 {
825 	struct sfxge_softc *sc;
826 	struct sfxge_txq *txq;
827 	int rc;
828 
829 	sc = (struct sfxge_softc *)ifp->if_softc;
830 
831 	/*
832 	 * Transmit may be called when interface is up from the kernel
833 	 * point of view, but not yet up (in progress) from the driver
834 	 * point of view. I.e. link aggregation bring up.
835 	 * Transmit may be called when interface is up from the driver
836 	 * point of view, but already down from the kernel point of
837 	 * view. I.e. Rx when interface shutdown is in progress.
838 	 */
839 	KASSERT((ifp->if_flags & IFF_UP) || (sc->if_flags & IFF_UP),
840 		("interface not up"));
841 
842 	/* Pick the desired transmit queue. */
843 	if (m->m_pkthdr.csum_flags &
844 	    (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO)) {
845 		int index = 0;
846 
847 #ifdef RSS
848 		uint32_t bucket_id;
849 
850 		/*
851 		 * Select a TX queue which matches the corresponding
852 		 * RX queue for the hash in order to assign both
853 		 * TX and RX parts of the flow to the same CPU
854 		 */
855 		if (rss_m2bucket(m, &bucket_id) == 0)
856 			index = bucket_id % (sc->txq_count - (SFXGE_TXQ_NTYPES - 1));
857 #else
858 		/* check if flowid is set */
859 		if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
860 			uint32_t hash = m->m_pkthdr.flowid;
861 			uint32_t idx = hash % nitems(sc->rx_indir_table);
862 
863 			index = sc->rx_indir_table[idx];
864 		}
865 #endif
866 #if SFXGE_TX_PARSE_EARLY
867 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
868 			sfxge_parse_tx_packet(m);
869 #endif
870 		txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index];
871 	} else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
872 		txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
873 	} else {
874 		txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
875 	}
876 
877 	rc = sfxge_tx_packet_add(txq, m);
878 	if (rc != 0)
879 		m_freem(m);
880 
881 	return (rc);
882 }
883 
884 /*
885  * Software "TSO".  Not quite as good as doing it in hardware, but
886  * still faster than segmenting in the stack.
887  */
888 
889 struct sfxge_tso_state {
890 	/* Output position */
891 	unsigned out_len;	/* Remaining length in current segment */
892 	unsigned seqnum;	/* Current sequence number */
893 	unsigned packet_space;	/* Remaining space in current packet */
894 	unsigned segs_space;	/* Remaining number of DMA segments
895 				   for the packet (FATSOv2 only) */
896 
897 	/* Input position */
898 	uint64_t dma_addr;	/* DMA address of current position */
899 	unsigned in_len;	/* Remaining length in current mbuf */
900 
901 	const struct mbuf *mbuf; /* Input mbuf (head of chain) */
902 	u_short protocol;	/* Network protocol (after VLAN decap) */
903 	ssize_t nh_off;		/* Offset of network header */
904 	ssize_t tcph_off;	/* Offset of TCP header */
905 	unsigned header_len;	/* Number of bytes of header */
906 	unsigned seg_size;	/* TCP segment size */
907 	int fw_assisted;	/* Use FW-assisted TSO */
908 	u_short packet_id;	/* IPv4 packet ID from the original packet */
909 	uint8_t tcp_flags;	/* TCP flags */
910 	efx_desc_t header_desc; /* Precomputed header descriptor for
911 				 * FW-assisted TSO */
912 };
913 
914 #if !SFXGE_TX_PARSE_EARLY
915 static const struct ip *tso_iph(const struct sfxge_tso_state *tso)
916 {
917 	KASSERT(tso->protocol == htons(ETHERTYPE_IP),
918 		("tso_iph() in non-IPv4 state"));
919 	return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
920 }
921 
922 static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
923 {
924 	KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
925 		("tso_ip6h() in non-IPv6 state"));
926 	return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
927 }
928 
929 static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
930 {
931 	return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
932 }
933 #endif
934 
935 
936 /* Size of preallocated TSO header buffers.  Larger blocks must be
937  * allocated from the heap.
938  */
939 #define	TSOH_STD_SIZE	128
940 
941 /* At most half the descriptors in the queue at any time will refer to
942  * a TSO header buffer, since they must always be followed by a
943  * payload descriptor referring to an mbuf.
944  */
945 #define	TSOH_COUNT(_txq_entries)	((_txq_entries) / 2u)
946 #define	TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
947 #define	TSOH_PAGE_COUNT(_txq_entries)	\
948 	howmany(TSOH_COUNT(_txq_entries), TSOH_PER_PAGE)
949 
950 static int tso_init(struct sfxge_txq *txq)
951 {
952 	struct sfxge_softc *sc = txq->sc;
953 	unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries);
954 	int i, rc;
955 
956 	/* Allocate TSO header buffers */
957 	txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]),
958 				  M_SFXGE, M_WAITOK);
959 
960 	for (i = 0; i < tsoh_page_count; i++) {
961 		rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
962 		if (rc != 0)
963 			goto fail;
964 	}
965 
966 	return (0);
967 
968 fail:
969 	while (i-- > 0)
970 		sfxge_dma_free(&txq->tsoh_buffer[i]);
971 	free(txq->tsoh_buffer, M_SFXGE);
972 	txq->tsoh_buffer = NULL;
973 	return (rc);
974 }
975 
976 static void tso_fini(struct sfxge_txq *txq)
977 {
978 	int i;
979 
980 	if (txq->tsoh_buffer != NULL) {
981 		for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++)
982 			sfxge_dma_free(&txq->tsoh_buffer[i]);
983 		free(txq->tsoh_buffer, M_SFXGE);
984 	}
985 }
986 
987 static void tso_start(struct sfxge_txq *txq, struct sfxge_tso_state *tso,
988 		      const bus_dma_segment_t *hdr_dma_seg,
989 		      struct mbuf *mbuf)
990 {
991 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->sc->enp);
992 #if !SFXGE_TX_PARSE_EARLY
993 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
994 	const struct tcphdr *th;
995 	struct tcphdr th_copy;
996 #endif
997 
998 	tso->fw_assisted = txq->tso_fw_assisted;
999 	tso->mbuf = mbuf;
1000 
1001 	/* Find network protocol and header */
1002 #if !SFXGE_TX_PARSE_EARLY
1003 	tso->protocol = eh->ether_type;
1004 	if (tso->protocol == htons(ETHERTYPE_VLAN)) {
1005 		struct ether_vlan_header *veh =
1006 			mtod(mbuf, struct ether_vlan_header *);
1007 		tso->protocol = veh->evl_proto;
1008 		tso->nh_off = sizeof(*veh);
1009 	} else {
1010 		tso->nh_off = sizeof(*eh);
1011 	}
1012 #else
1013 	tso->protocol = TSO_MBUF_PROTO(mbuf);
1014 	tso->nh_off = mbuf->m_pkthdr.l2hlen;
1015 	tso->tcph_off = mbuf->m_pkthdr.l3hlen;
1016 	tso->packet_id = ntohs(TSO_MBUF_PACKETID(mbuf));
1017 #endif
1018 
1019 #if !SFXGE_TX_PARSE_EARLY
1020 	/* Find TCP header */
1021 	if (tso->protocol == htons(ETHERTYPE_IP)) {
1022 		KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
1023 			("TSO required on non-TCP packet"));
1024 		tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
1025 		tso->packet_id = ntohs(tso_iph(tso)->ip_id);
1026 	} else {
1027 		KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
1028 			("TSO required on non-IP packet"));
1029 		KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
1030 			("TSO required on non-TCP packet"));
1031 		tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
1032 		tso->packet_id = 0;
1033 	}
1034 #endif
1035 
1036 
1037 	if (tso->fw_assisted &&
1038 	    __predict_false(tso->tcph_off >
1039 			    encp->enc_tx_tso_tcp_header_offset_limit)) {
1040 		tso->fw_assisted = 0;
1041 	}
1042 
1043 
1044 #if !SFXGE_TX_PARSE_EARLY
1045 	KASSERT(mbuf->m_len >= tso->tcph_off,
1046 		("network header is fragmented in mbuf"));
1047 	/* We need TCP header including flags (window is the next) */
1048 	if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) {
1049 		m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy),
1050 			   (caddr_t)&th_copy);
1051 		th = &th_copy;
1052 	} else {
1053 		th = tso_tcph(tso);
1054 	}
1055 	tso->header_len = tso->tcph_off + 4 * th->th_off;
1056 #else
1057 	tso->header_len = mbuf->m_pkthdr.l4hlen;
1058 #endif
1059 	tso->seg_size = mbuf->m_pkthdr.tso_segsz;
1060 
1061 #if !SFXGE_TX_PARSE_EARLY
1062 	tso->seqnum = ntohl(th->th_seq);
1063 
1064 	/* These flags must not be duplicated */
1065 	/*
1066 	 * RST should not be duplicated as well, but FreeBSD kernel
1067 	 * generates TSO packets with RST flag. So, do not assert
1068 	 * its absence.
1069 	 */
1070 	KASSERT(!(th->th_flags & (TH_URG | TH_SYN)),
1071 		("incompatible TCP flag 0x%x on TSO packet",
1072 		 th->th_flags & (TH_URG | TH_SYN)));
1073 	tso->tcp_flags = th->th_flags;
1074 #else
1075 	tso->seqnum = TSO_MBUF_SEQNUM(mbuf);
1076 	tso->tcp_flags = TSO_MBUF_FLAGS(mbuf);
1077 #endif
1078 
1079 	tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
1080 
1081 	if (tso->fw_assisted) {
1082 		if (hdr_dma_seg->ds_len >= tso->header_len)
1083 			efx_tx_qdesc_dma_create(txq->common,
1084 						hdr_dma_seg->ds_addr,
1085 						tso->header_len,
1086 						B_FALSE,
1087 						&tso->header_desc);
1088 		else
1089 			tso->fw_assisted = 0;
1090 	}
1091 }
1092 
1093 /*
1094  * tso_fill_packet_with_fragment - form descriptors for the current fragment
1095  *
1096  * Form descriptors for the current fragment, until we reach the end
1097  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
1098  * space.
1099  */
1100 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
1101 					  struct sfxge_tso_state *tso)
1102 {
1103 	efx_desc_t *desc;
1104 	int n;
1105 	uint64_t dma_addr = tso->dma_addr;
1106 	boolean_t eop;
1107 
1108 	if (tso->in_len == 0 || tso->packet_space == 0)
1109 		return;
1110 
1111 	KASSERT(tso->in_len > 0, ("TSO input length went negative"));
1112 	KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
1113 
1114 	if (tso->fw_assisted & SFXGE_FATSOV2) {
1115 		n = tso->in_len;
1116 		tso->out_len -= n;
1117 		tso->seqnum += n;
1118 		tso->in_len = 0;
1119 		if (n < tso->packet_space) {
1120 			tso->packet_space -= n;
1121 			tso->segs_space--;
1122 		} else {
1123 			tso->packet_space = tso->seg_size -
1124 			    (n - tso->packet_space) % tso->seg_size;
1125 			tso->segs_space =
1126 			    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1 -
1127 			    (tso->packet_space != tso->seg_size);
1128 		}
1129 	} else {
1130 		n = min(tso->in_len, tso->packet_space);
1131 		tso->packet_space -= n;
1132 		tso->out_len -= n;
1133 		tso->dma_addr += n;
1134 		tso->in_len -= n;
1135 	}
1136 
1137 	/*
1138 	 * It is OK to use binary OR below to avoid extra branching
1139 	 * since all conditions may always be checked.
1140 	 */
1141 	eop = (tso->out_len == 0) | (tso->packet_space == 0) |
1142 	    (tso->segs_space == 0);
1143 
1144 	desc = &txq->pend_desc[txq->n_pend_desc++];
1145 	efx_tx_qdesc_dma_create(txq->common, dma_addr, n, eop, desc);
1146 }
1147 
1148 /* Callback from bus_dmamap_load() for long TSO headers. */
1149 static void tso_map_long_header(void *dma_addr_ret,
1150 				bus_dma_segment_t *segs, int nseg,
1151 				int error)
1152 {
1153 	*(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
1154 				      __predict_true(nseg == 1)) ?
1155 				     segs->ds_addr : 0);
1156 }
1157 
1158 /*
1159  * tso_start_new_packet - generate a new header and prepare for the new packet
1160  *
1161  * Generate a new header and prepare for the new packet.  Return 0 on
1162  * success, or an error code if failed to alloc header.
1163  */
1164 static int tso_start_new_packet(struct sfxge_txq *txq,
1165 				struct sfxge_tso_state *tso,
1166 				unsigned int *idp)
1167 {
1168 	unsigned int id = *idp;
1169 	struct tcphdr *tsoh_th;
1170 	unsigned ip_length;
1171 	caddr_t header;
1172 	uint64_t dma_addr;
1173 	bus_dmamap_t map;
1174 	efx_desc_t *desc;
1175 	int rc;
1176 
1177 	if (tso->fw_assisted) {
1178 		if (tso->fw_assisted & SFXGE_FATSOV2) {
1179 			/* Add 2 FATSOv2 option descriptors */
1180 			desc = &txq->pend_desc[txq->n_pend_desc];
1181 			efx_tx_qdesc_tso2_create(txq->common,
1182 						 tso->packet_id,
1183 						 0,
1184 						 tso->seqnum,
1185 						 tso->seg_size,
1186 						 desc,
1187 						 EFX_TX_FATSOV2_OPT_NDESCS);
1188 			desc += EFX_TX_FATSOV2_OPT_NDESCS;
1189 			txq->n_pend_desc += EFX_TX_FATSOV2_OPT_NDESCS;
1190 			KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1191 			id = (id + EFX_TX_FATSOV2_OPT_NDESCS) & txq->ptr_mask;
1192 
1193 			tso->segs_space =
1194 			    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1;
1195 		} else {
1196 			uint8_t tcp_flags = tso->tcp_flags;
1197 
1198 			if (tso->out_len > tso->seg_size)
1199 				tcp_flags &= ~(TH_FIN | TH_PUSH);
1200 
1201 			/* Add FATSOv1 option descriptor */
1202 			desc = &txq->pend_desc[txq->n_pend_desc++];
1203 			efx_tx_qdesc_tso_create(txq->common,
1204 						tso->packet_id,
1205 						tso->seqnum,
1206 						tcp_flags,
1207 						desc++);
1208 			KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1209 			id = (id + 1) & txq->ptr_mask;
1210 
1211 			tso->seqnum += tso->seg_size;
1212 			tso->segs_space = UINT_MAX;
1213 		}
1214 
1215 		/* Header DMA descriptor */
1216 		*desc = tso->header_desc;
1217 		txq->n_pend_desc++;
1218 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1219 		id = (id + 1) & txq->ptr_mask;
1220 	} else {
1221 		/* Allocate a DMA-mapped header buffer. */
1222 		if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
1223 			unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
1224 			unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
1225 
1226 			header = (txq->tsoh_buffer[page_index].esm_base +
1227 				  buf_index * TSOH_STD_SIZE);
1228 			dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
1229 				    buf_index * TSOH_STD_SIZE);
1230 			map = txq->tsoh_buffer[page_index].esm_map;
1231 
1232 			KASSERT(txq->stmp[id].flags == 0,
1233 				("stmp flags are not 0"));
1234 		} else {
1235 			struct sfxge_tx_mapping *stmp = &txq->stmp[id];
1236 
1237 			/* We cannot use bus_dmamem_alloc() as that may sleep */
1238 			header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
1239 			if (__predict_false(!header))
1240 				return (ENOMEM);
1241 			rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
1242 					     header, tso->header_len,
1243 					     tso_map_long_header, &dma_addr,
1244 					     BUS_DMA_NOWAIT);
1245 			if (__predict_false(dma_addr == 0)) {
1246 				if (rc == 0) {
1247 					/* Succeeded but got >1 segment */
1248 					bus_dmamap_unload(txq->packet_dma_tag,
1249 							  stmp->map);
1250 					rc = EINVAL;
1251 				}
1252 				free(header, M_SFXGE);
1253 				return (rc);
1254 			}
1255 			map = stmp->map;
1256 
1257 			txq->tso_long_headers++;
1258 			stmp->u.heap_buf = header;
1259 			stmp->flags = TX_BUF_UNMAP;
1260 		}
1261 
1262 		tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
1263 
1264 		/* Copy and update the headers. */
1265 		m_copydata(tso->mbuf, 0, tso->header_len, header);
1266 
1267 		tsoh_th->th_seq = htonl(tso->seqnum);
1268 		tso->seqnum += tso->seg_size;
1269 		if (tso->out_len > tso->seg_size) {
1270 			/* This packet will not finish the TSO burst. */
1271 			ip_length = tso->header_len - tso->nh_off + tso->seg_size;
1272 			tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH);
1273 		} else {
1274 			/* This packet will be the last in the TSO burst. */
1275 			ip_length = tso->header_len - tso->nh_off + tso->out_len;
1276 		}
1277 
1278 		if (tso->protocol == htons(ETHERTYPE_IP)) {
1279 			struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
1280 			tsoh_iph->ip_len = htons(ip_length);
1281 			/* XXX We should increment ip_id, but FreeBSD doesn't
1282 			 * currently allocate extra IDs for multiple segments.
1283 			 */
1284 		} else {
1285 			struct ip6_hdr *tsoh_iph =
1286 				(struct ip6_hdr *)(header + tso->nh_off);
1287 			tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
1288 		}
1289 
1290 		/* Make the header visible to the hardware. */
1291 		bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
1292 
1293 		/* Form a descriptor for this header. */
1294 		desc = &txq->pend_desc[txq->n_pend_desc++];
1295 		efx_tx_qdesc_dma_create(txq->common,
1296 					dma_addr,
1297 					tso->header_len,
1298 					0,
1299 					desc);
1300 		id = (id + 1) & txq->ptr_mask;
1301 
1302 		tso->segs_space = UINT_MAX;
1303 	}
1304 	tso->packet_space = tso->seg_size;
1305 	txq->tso_packets++;
1306 	*idp = id;
1307 
1308 	return (0);
1309 }
1310 
1311 static int
1312 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
1313 		   const bus_dma_segment_t *dma_seg, int n_dma_seg,
1314 		   int vlan_tagged)
1315 {
1316 	struct sfxge_tso_state tso;
1317 	unsigned int id;
1318 	unsigned skipped = 0;
1319 
1320 	tso_start(txq, &tso, dma_seg, mbuf);
1321 
1322 	while (dma_seg->ds_len + skipped <= tso.header_len) {
1323 		skipped += dma_seg->ds_len;
1324 		--n_dma_seg;
1325 		KASSERT(n_dma_seg, ("no payload found in TSO packet"));
1326 		++dma_seg;
1327 	}
1328 	tso.in_len = dma_seg->ds_len - (tso.header_len - skipped);
1329 	tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped);
1330 
1331 	id = (txq->added + vlan_tagged) & txq->ptr_mask;
1332 	if (__predict_false(tso_start_new_packet(txq, &tso, &id)))
1333 		return (-1);
1334 
1335 	while (1) {
1336 		tso_fill_packet_with_fragment(txq, &tso);
1337 		/* Exactly one DMA descriptor is added */
1338 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1339 		id = (id + 1) & txq->ptr_mask;
1340 
1341 		/* Move onto the next fragment? */
1342 		if (tso.in_len == 0) {
1343 			--n_dma_seg;
1344 			if (n_dma_seg == 0)
1345 				break;
1346 			++dma_seg;
1347 			tso.in_len = dma_seg->ds_len;
1348 			tso.dma_addr = dma_seg->ds_addr;
1349 		}
1350 
1351 		/* End of packet? */
1352 		if ((tso.packet_space == 0) | (tso.segs_space == 0)) {
1353 			unsigned int n_fatso_opt_desc =
1354 			    (tso.fw_assisted & SFXGE_FATSOV2) ?
1355 			    EFX_TX_FATSOV2_OPT_NDESCS :
1356 			    (tso.fw_assisted & SFXGE_FATSOV1) ? 1 : 0;
1357 
1358 			/* If the queue is now full due to tiny MSS,
1359 			 * or we can't create another header, discard
1360 			 * the remainder of the input mbuf but do not
1361 			 * roll back the work we have done.
1362 			 */
1363 			if (txq->n_pend_desc + n_fatso_opt_desc +
1364 			    1 /* header */ + n_dma_seg > txq->max_pkt_desc) {
1365 				txq->tso_pdrop_too_many++;
1366 				break;
1367 			}
1368 			if (__predict_false(tso_start_new_packet(txq, &tso,
1369 								 &id))) {
1370 				txq->tso_pdrop_no_rsrc++;
1371 				break;
1372 			}
1373 		}
1374 	}
1375 
1376 	txq->tso_bursts++;
1377 	return (id);
1378 }
1379 
1380 static void
1381 sfxge_tx_qunblock(struct sfxge_txq *txq)
1382 {
1383 	struct sfxge_softc *sc;
1384 	struct sfxge_evq *evq;
1385 
1386 	sc = txq->sc;
1387 	evq = sc->evq[txq->evq_index];
1388 
1389 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
1390 
1391 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED))
1392 		return;
1393 
1394 	SFXGE_TXQ_LOCK(txq);
1395 
1396 	if (txq->blocked) {
1397 		unsigned int level;
1398 
1399 		level = txq->added - txq->completed;
1400 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) {
1401 			/* reaped must be in sync with blocked */
1402 			sfxge_tx_qreap(txq);
1403 			txq->blocked = 0;
1404 		}
1405 	}
1406 
1407 	sfxge_tx_qdpl_service(txq);
1408 	/* note: lock has been dropped */
1409 }
1410 
1411 void
1412 sfxge_tx_qflush_done(struct sfxge_txq *txq)
1413 {
1414 
1415 	txq->flush_state = SFXGE_FLUSH_DONE;
1416 }
1417 
1418 static void
1419 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1420 {
1421 	struct sfxge_txq *txq;
1422 	struct sfxge_evq *evq;
1423 	unsigned int count;
1424 
1425 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1426 
1427 	txq = sc->txq[index];
1428 	evq = sc->evq[txq->evq_index];
1429 
1430 	SFXGE_EVQ_LOCK(evq);
1431 	SFXGE_TXQ_LOCK(txq);
1432 
1433 	KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1434 	    ("txq->init_state != SFXGE_TXQ_STARTED"));
1435 
1436 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1437 
1438 	if (txq->flush_state != SFXGE_FLUSH_DONE) {
1439 		txq->flush_state = SFXGE_FLUSH_PENDING;
1440 
1441 		SFXGE_EVQ_UNLOCK(evq);
1442 		SFXGE_TXQ_UNLOCK(txq);
1443 
1444 		/* Flush the transmit queue. */
1445 		if (efx_tx_qflush(txq->common) != 0) {
1446 			log(LOG_ERR, "%s: Flushing Tx queue %u failed\n",
1447 			    device_get_nameunit(sc->dev), index);
1448 			txq->flush_state = SFXGE_FLUSH_DONE;
1449 		} else {
1450 			count = 0;
1451 			do {
1452 				/* Spin for 100ms. */
1453 				DELAY(100000);
1454 				if (txq->flush_state != SFXGE_FLUSH_PENDING)
1455 					break;
1456 			} while (++count < 20);
1457 		}
1458 		SFXGE_EVQ_LOCK(evq);
1459 		SFXGE_TXQ_LOCK(txq);
1460 
1461 		KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1462 		    ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1463 
1464 		if (txq->flush_state != SFXGE_FLUSH_DONE) {
1465 			/* Flush timeout */
1466 			log(LOG_ERR, "%s: Cannot flush Tx queue %u\n",
1467 			    device_get_nameunit(sc->dev), index);
1468 			txq->flush_state = SFXGE_FLUSH_DONE;
1469 		}
1470 	}
1471 
1472 	txq->blocked = 0;
1473 	txq->pending = txq->added;
1474 
1475 	sfxge_tx_qcomplete(txq, evq);
1476 	KASSERT(txq->completed == txq->added,
1477 	    ("txq->completed != txq->added"));
1478 
1479 	sfxge_tx_qreap(txq);
1480 	KASSERT(txq->reaped == txq->completed,
1481 	    ("txq->reaped != txq->completed"));
1482 
1483 	txq->added = 0;
1484 	txq->pending = 0;
1485 	txq->completed = 0;
1486 	txq->reaped = 0;
1487 
1488 	/* Destroy the common code transmit queue. */
1489 	efx_tx_qdestroy(txq->common);
1490 	txq->common = NULL;
1491 
1492 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1493 	    EFX_TXQ_NBUFS(sc->txq_entries));
1494 
1495 	SFXGE_EVQ_UNLOCK(evq);
1496 	SFXGE_TXQ_UNLOCK(txq);
1497 }
1498 
1499 /*
1500  * Estimate maximum number of Tx descriptors required for TSO packet.
1501  * With minimum MSS and maximum mbuf length we might need more (even
1502  * than a ring-ful of descriptors), but this should not happen in
1503  * practice except due to deliberate attack.  In that case we will
1504  * truncate the output at a packet boundary.
1505  */
1506 static unsigned int
1507 sfxge_tx_max_pkt_desc(const struct sfxge_softc *sc, enum sfxge_txq_type type,
1508 		      unsigned int tso_fw_assisted)
1509 {
1510 	/* One descriptor for every input fragment */
1511 	unsigned int max_descs = SFXGE_TX_MAPPING_MAX_SEG;
1512 	unsigned int sw_tso_max_descs;
1513 	unsigned int fa_tso_v1_max_descs = 0;
1514 	unsigned int fa_tso_v2_max_descs = 0;
1515 
1516 	/* VLAN tagging Tx option descriptor may be required */
1517 	if (efx_nic_cfg_get(sc->enp)->enc_hw_tx_insert_vlan_enabled)
1518 		max_descs++;
1519 
1520 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) {
1521 		/*
1522 		 * Plus header and payload descriptor for each output segment.
1523 		 * Minus one since header fragment is already counted.
1524 		 * Even if FATSO is used, we should be ready to fallback
1525 		 * to do it in the driver.
1526 		 */
1527 		sw_tso_max_descs = SFXGE_TSO_MAX_SEGS * 2 - 1;
1528 
1529 		/* FW assisted TSOv1 requires one more descriptor per segment
1530 		 * in comparison to SW TSO */
1531 		if (tso_fw_assisted & SFXGE_FATSOV1)
1532 			fa_tso_v1_max_descs =
1533 			    sw_tso_max_descs + SFXGE_TSO_MAX_SEGS;
1534 
1535 		/* FW assisted TSOv2 requires 3 (2 FATSO plus header) extra
1536 		 * descriptors per superframe limited by number of DMA fetches
1537 		 * per packet. The first packet header is already counted.
1538 		 */
1539 		if (tso_fw_assisted & SFXGE_FATSOV2) {
1540 			fa_tso_v2_max_descs =
1541 			    howmany(SFXGE_TX_MAPPING_MAX_SEG,
1542 				    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1) *
1543 			    (EFX_TX_FATSOV2_OPT_NDESCS + 1) - 1;
1544 		}
1545 
1546 		max_descs += MAX(sw_tso_max_descs,
1547 				 MAX(fa_tso_v1_max_descs, fa_tso_v2_max_descs));
1548 	}
1549 
1550 	return (max_descs);
1551 }
1552 
1553 static int
1554 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1555 {
1556 	struct sfxge_txq *txq;
1557 	efsys_mem_t *esmp;
1558 	uint16_t flags;
1559 	unsigned int tso_fw_assisted;
1560 	struct sfxge_evq *evq;
1561 	unsigned int desc_index;
1562 	int rc;
1563 
1564 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1565 
1566 	txq = sc->txq[index];
1567 	esmp = &txq->mem;
1568 	evq = sc->evq[txq->evq_index];
1569 
1570 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1571 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1572 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1573 	    ("evq->init_state != SFXGE_EVQ_STARTED"));
1574 
1575 	/* Program the buffer table. */
1576 	if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1577 	    EFX_TXQ_NBUFS(sc->txq_entries))) != 0)
1578 		return (rc);
1579 
1580 	/* Determine the kind of queue we are creating. */
1581 	tso_fw_assisted = 0;
1582 	switch (txq->type) {
1583 	case SFXGE_TXQ_NON_CKSUM:
1584 		flags = 0;
1585 		break;
1586 	case SFXGE_TXQ_IP_CKSUM:
1587 		flags = EFX_TXQ_CKSUM_IPV4;
1588 		break;
1589 	case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1590 		flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP;
1591 		tso_fw_assisted = sc->tso_fw_assisted;
1592 		if (tso_fw_assisted & SFXGE_FATSOV2)
1593 			flags |= EFX_TXQ_FATSOV2;
1594 		break;
1595 	default:
1596 		KASSERT(0, ("Impossible TX queue"));
1597 		flags = 0;
1598 		break;
1599 	}
1600 
1601 	/* Create the common code transmit queue. */
1602 	if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1603 	    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1604 	    &txq->common, &desc_index)) != 0) {
1605 		/* Retry if no FATSOv2 resources, otherwise fail */
1606 		if ((rc != ENOSPC) || (~flags & EFX_TXQ_FATSOV2))
1607 			goto fail;
1608 
1609 		/* Looks like all FATSOv2 contexts are used */
1610 		flags &= ~EFX_TXQ_FATSOV2;
1611 		tso_fw_assisted &= ~SFXGE_FATSOV2;
1612 		if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1613 		    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1614 		    &txq->common, &desc_index)) != 0)
1615 			goto fail;
1616 	}
1617 
1618 	/* Initialise queue descriptor indexes */
1619 	txq->added = txq->pending = txq->completed = txq->reaped = desc_index;
1620 
1621 	SFXGE_TXQ_LOCK(txq);
1622 
1623 	/* Enable the transmit queue. */
1624 	efx_tx_qenable(txq->common);
1625 
1626 	txq->init_state = SFXGE_TXQ_STARTED;
1627 	txq->flush_state = SFXGE_FLUSH_REQUIRED;
1628 	txq->tso_fw_assisted = tso_fw_assisted;
1629 
1630 	txq->max_pkt_desc = sfxge_tx_max_pkt_desc(sc, txq->type,
1631 						  tso_fw_assisted);
1632 
1633 	txq->hw_vlan_tci = 0;
1634 
1635 	SFXGE_TXQ_UNLOCK(txq);
1636 
1637 	return (0);
1638 
1639 fail:
1640 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1641 	    EFX_TXQ_NBUFS(sc->txq_entries));
1642 	return (rc);
1643 }
1644 
1645 void
1646 sfxge_tx_stop(struct sfxge_softc *sc)
1647 {
1648 	int index;
1649 
1650 	index = sc->txq_count;
1651 	while (--index >= 0)
1652 		sfxge_tx_qstop(sc, index);
1653 
1654 	/* Tear down the transmit module */
1655 	efx_tx_fini(sc->enp);
1656 }
1657 
1658 int
1659 sfxge_tx_start(struct sfxge_softc *sc)
1660 {
1661 	int index;
1662 	int rc;
1663 
1664 	/* Initialize the common code transmit module. */
1665 	if ((rc = efx_tx_init(sc->enp)) != 0)
1666 		return (rc);
1667 
1668 	for (index = 0; index < sc->txq_count; index++) {
1669 		if ((rc = sfxge_tx_qstart(sc, index)) != 0)
1670 			goto fail;
1671 	}
1672 
1673 	return (0);
1674 
1675 fail:
1676 	while (--index >= 0)
1677 		sfxge_tx_qstop(sc, index);
1678 
1679 	efx_tx_fini(sc->enp);
1680 
1681 	return (rc);
1682 }
1683 
1684 static int
1685 sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node)
1686 {
1687 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev);
1688 	struct sysctl_oid *stat_node;
1689 	unsigned int id;
1690 
1691 	stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1692 				    "stats", CTLFLAG_RD, NULL,
1693 				    "Tx queue statistics");
1694 	if (stat_node == NULL)
1695 		return (ENOMEM);
1696 
1697 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1698 		SYSCTL_ADD_ULONG(
1699 		    ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO,
1700 		    sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS,
1701 		    (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset),
1702 		    "");
1703 	}
1704 
1705 	return (0);
1706 }
1707 
1708 /**
1709  * Destroy a transmit queue.
1710  */
1711 static void
1712 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1713 {
1714 	struct sfxge_txq *txq;
1715 	unsigned int nmaps;
1716 
1717 	txq = sc->txq[index];
1718 
1719 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1720 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1721 
1722 	if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1723 		tso_fini(txq);
1724 
1725 	/* Free the context arrays. */
1726 	free(txq->pend_desc, M_SFXGE);
1727 	nmaps = sc->txq_entries;
1728 	while (nmaps-- != 0)
1729 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1730 	free(txq->stmp, M_SFXGE);
1731 
1732 	/* Release DMA memory mapping. */
1733 	sfxge_dma_free(&txq->mem);
1734 
1735 	sc->txq[index] = NULL;
1736 
1737 	SFXGE_TXQ_LOCK_DESTROY(txq);
1738 
1739 	free(txq, M_SFXGE);
1740 }
1741 
1742 static int
1743 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1744 	       enum sfxge_txq_type type, unsigned int evq_index)
1745 {
1746 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
1747 	char name[16];
1748 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1749 	struct sysctl_oid *txq_node;
1750 	struct sfxge_txq *txq;
1751 	struct sfxge_evq *evq;
1752 	struct sfxge_tx_dpl *stdp;
1753 	struct sysctl_oid *dpl_node;
1754 	efsys_mem_t *esmp;
1755 	unsigned int nmaps;
1756 	int rc;
1757 
1758 	txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1759 	txq->sc = sc;
1760 	txq->entries = sc->txq_entries;
1761 	txq->ptr_mask = txq->entries - 1;
1762 
1763 	sc->txq[txq_index] = txq;
1764 	esmp = &txq->mem;
1765 
1766 	evq = sc->evq[evq_index];
1767 
1768 	/* Allocate and zero DMA space for the descriptor ring. */
1769 	if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0)
1770 		return (rc);
1771 
1772 	/* Allocate buffer table entries. */
1773 	sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries),
1774 				 &txq->buf_base_id);
1775 
1776 	/* Create a DMA tag for packet mappings. */
1777 	if (bus_dma_tag_create(sc->parent_dma_tag, 1,
1778 	    encp->enc_tx_dma_desc_boundary,
1779 	    MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1780 	    NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG,
1781 	    encp->enc_tx_dma_desc_size_max, 0, NULL, NULL,
1782 	    &txq->packet_dma_tag) != 0) {
1783 		device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1784 		rc = ENOMEM;
1785 		goto fail;
1786 	}
1787 
1788 	/* Allocate pending descriptor array for batching writes. */
1789 	txq->pend_desc = malloc(sizeof(efx_desc_t) * sc->txq_entries,
1790 				M_SFXGE, M_ZERO | M_WAITOK);
1791 
1792 	/* Allocate and initialise mbuf DMA mapping array. */
1793 	txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries,
1794 	    M_SFXGE, M_ZERO | M_WAITOK);
1795 	for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) {
1796 		rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1797 				       &txq->stmp[nmaps].map);
1798 		if (rc != 0)
1799 			goto fail2;
1800 	}
1801 
1802 	snprintf(name, sizeof(name), "%u", txq_index);
1803 	txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node),
1804 				   OID_AUTO, name, CTLFLAG_RD, NULL, "");
1805 	if (txq_node == NULL) {
1806 		rc = ENOMEM;
1807 		goto fail_txq_node;
1808 	}
1809 
1810 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1811 	    (rc = tso_init(txq)) != 0)
1812 		goto fail3;
1813 
1814 	/* Initialize the deferred packet list. */
1815 	stdp = &txq->dpl;
1816 	stdp->std_put_max = sfxge_tx_dpl_put_max;
1817 	stdp->std_get_max = sfxge_tx_dpl_get_max;
1818 	stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max;
1819 	stdp->std_getp = &stdp->std_get;
1820 
1821 	SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index);
1822 
1823 	dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1824 				   "dpl", CTLFLAG_RD, NULL,
1825 				   "Deferred packet list statistics");
1826 	if (dpl_node == NULL) {
1827 		rc = ENOMEM;
1828 		goto fail_dpl_node;
1829 	}
1830 
1831 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1832 			"get_count", CTLFLAG_RD | CTLFLAG_STATS,
1833 			&stdp->std_get_count, 0, "");
1834 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1835 			"get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS,
1836 			&stdp->std_get_non_tcp_count, 0, "");
1837 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1838 			"get_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1839 			&stdp->std_get_hiwat, 0, "");
1840 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1841 			"put_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1842 			&stdp->std_put_hiwat, 0, "");
1843 
1844 	rc = sfxge_txq_stat_init(txq, txq_node);
1845 	if (rc != 0)
1846 		goto fail_txq_stat_init;
1847 
1848 	txq->type = type;
1849 	txq->evq_index = evq_index;
1850 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1851 
1852 	return (0);
1853 
1854 fail_txq_stat_init:
1855 fail_dpl_node:
1856 fail3:
1857 fail_txq_node:
1858 	free(txq->pend_desc, M_SFXGE);
1859 fail2:
1860 	while (nmaps-- != 0)
1861 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1862 	free(txq->stmp, M_SFXGE);
1863 	bus_dma_tag_destroy(txq->packet_dma_tag);
1864 
1865 fail:
1866 	sfxge_dma_free(esmp);
1867 
1868 	return (rc);
1869 }
1870 
1871 static int
1872 sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1873 {
1874 	struct sfxge_softc *sc = arg1;
1875 	unsigned int id = arg2;
1876 	unsigned long sum;
1877 	unsigned int index;
1878 
1879 	/* Sum across all TX queues */
1880 	sum = 0;
1881 	for (index = 0; index < sc->txq_count; index++)
1882 		sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1883 					  sfxge_tx_stats[id].offset);
1884 
1885 	return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1886 }
1887 
1888 static void
1889 sfxge_tx_stat_init(struct sfxge_softc *sc)
1890 {
1891 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1892 	struct sysctl_oid_list *stat_list;
1893 	unsigned int id;
1894 
1895 	stat_list = SYSCTL_CHILDREN(sc->stats_node);
1896 
1897 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1898 		SYSCTL_ADD_PROC(
1899 			ctx, stat_list,
1900 			OID_AUTO, sfxge_tx_stats[id].name,
1901 			CTLTYPE_ULONG|CTLFLAG_RD,
1902 			sc, id, sfxge_tx_stat_handler, "LU",
1903 			"");
1904 	}
1905 }
1906 
1907 uint64_t
1908 sfxge_tx_get_drops(struct sfxge_softc *sc)
1909 {
1910 	unsigned int index;
1911 	uint64_t drops = 0;
1912 	struct sfxge_txq *txq;
1913 
1914 	/* Sum across all TX queues */
1915 	for (index = 0; index < sc->txq_count; index++) {
1916 		txq = sc->txq[index];
1917 		/*
1918 		 * In theory, txq->put_overflow and txq->netdown_drops
1919 		 * should use atomic operation and other should be
1920 		 * obtained under txq lock, but it is just statistics.
1921 		 */
1922 		drops += txq->drops + txq->get_overflow +
1923 			 txq->get_non_tcp_overflow +
1924 			 txq->put_overflow + txq->netdown_drops +
1925 			 txq->tso_pdrop_too_many + txq->tso_pdrop_no_rsrc;
1926 	}
1927 	return (drops);
1928 }
1929 
1930 void
1931 sfxge_tx_fini(struct sfxge_softc *sc)
1932 {
1933 	int index;
1934 
1935 	index = sc->txq_count;
1936 	while (--index >= 0)
1937 		sfxge_tx_qfini(sc, index);
1938 
1939 	sc->txq_count = 0;
1940 }
1941 
1942 
1943 int
1944 sfxge_tx_init(struct sfxge_softc *sc)
1945 {
1946 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
1947 	struct sfxge_intr *intr;
1948 	int index;
1949 	int rc;
1950 
1951 	intr = &sc->intr;
1952 
1953 	KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1954 	    ("intr->state != SFXGE_INTR_INITIALIZED"));
1955 
1956 	if (sfxge_tx_dpl_get_max <= 0) {
1957 		log(LOG_ERR, "%s=%d must be greater than 0",
1958 		    SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max);
1959 		rc = EINVAL;
1960 		goto fail_tx_dpl_get_max;
1961 	}
1962 	if (sfxge_tx_dpl_get_non_tcp_max <= 0) {
1963 		log(LOG_ERR, "%s=%d must be greater than 0",
1964 		    SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX,
1965 		    sfxge_tx_dpl_get_non_tcp_max);
1966 		rc = EINVAL;
1967 		goto fail_tx_dpl_get_non_tcp_max;
1968 	}
1969 	if (sfxge_tx_dpl_put_max < 0) {
1970 		log(LOG_ERR, "%s=%d must be greater or equal to 0",
1971 		    SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max);
1972 		rc = EINVAL;
1973 		goto fail_tx_dpl_put_max;
1974 	}
1975 
1976 	sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc;
1977 
1978 	sc->tso_fw_assisted = sfxge_tso_fw_assisted;
1979 	if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO) ||
1980 	    (!encp->enc_fw_assisted_tso_enabled))
1981 		sc->tso_fw_assisted &= ~SFXGE_FATSOV1;
1982 	if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO_V2) ||
1983 	    (!encp->enc_fw_assisted_tso_v2_enabled))
1984 		sc->tso_fw_assisted &= ~SFXGE_FATSOV2;
1985 
1986 	sc->txqs_node = SYSCTL_ADD_NODE(
1987 		device_get_sysctl_ctx(sc->dev),
1988 		SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
1989 		OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues");
1990 	if (sc->txqs_node == NULL) {
1991 		rc = ENOMEM;
1992 		goto fail_txq_node;
1993 	}
1994 
1995 	/* Initialize the transmit queues */
1996 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
1997 	    SFXGE_TXQ_NON_CKSUM, 0)) != 0)
1998 		goto fail;
1999 
2000 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
2001 	    SFXGE_TXQ_IP_CKSUM, 0)) != 0)
2002 		goto fail2;
2003 
2004 	for (index = 0;
2005 	     index < sc->txq_count - SFXGE_TXQ_NTYPES + 1;
2006 	     index++) {
2007 		if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index,
2008 		    SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
2009 			goto fail3;
2010 	}
2011 
2012 	sfxge_tx_stat_init(sc);
2013 
2014 	return (0);
2015 
2016 fail3:
2017 	while (--index >= 0)
2018 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
2019 
2020 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
2021 
2022 fail2:
2023 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
2024 
2025 fail:
2026 fail_txq_node:
2027 	sc->txq_count = 0;
2028 fail_tx_dpl_put_max:
2029 fail_tx_dpl_get_non_tcp_max:
2030 fail_tx_dpl_get_max:
2031 	return (rc);
2032 }
2033