xref: /freebsd/sys/dev/sfxge/sfxge_tx.c (revision 895f86f15fbf6540071feb9328c3c50ed1f027b8)
1 /*-
2  * Copyright (c) 2010-2015 Solarflare Communications Inc.
3  * All rights reserved.
4  *
5  * This software was developed in part by Philip Paeps under contract for
6  * Solarflare Communications, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright notice,
12  *    this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright notice,
14  *    this list of conditions and the following disclaimer in the documentation
15  *    and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
18  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
19  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
21  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
24  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
25  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
26  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
27  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28  *
29  * The views and conclusions contained in the software and documentation are
30  * those of the authors and should not be interpreted as representing official
31  * policies, either expressed or implied, of the FreeBSD Project.
32  */
33 
34 /* Theory of operation:
35  *
36  * Tx queues allocation and mapping
37  *
38  * One Tx queue with enabled checksum offload is allocated per Rx channel
39  * (event queue).  Also 2 Tx queues (one without checksum offload and one
40  * with IP checksum offload only) are allocated and bound to event queue 0.
41  * sfxge_txq_type is used as Tx queue label.
42  *
43  * So, event queue plus label mapping to Tx queue index is:
44  *	if event queue index is 0, TxQ-index = TxQ-label * [0..SFXGE_TXQ_NTYPES)
45  *	else TxQ-index = SFXGE_TXQ_NTYPES + EvQ-index - 1
46  * See sfxge_get_txq_by_label() sfxge_ev.c
47  */
48 
49 #include <sys/cdefs.h>
50 __FBSDID("$FreeBSD$");
51 
52 #include <sys/types.h>
53 #include <sys/mbuf.h>
54 #include <sys/smp.h>
55 #include <sys/socket.h>
56 #include <sys/sysctl.h>
57 #include <sys/syslog.h>
58 
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/if.h>
62 #include <net/if_vlan_var.h>
63 
64 #include <netinet/in.h>
65 #include <netinet/ip.h>
66 #include <netinet/ip6.h>
67 #include <netinet/tcp.h>
68 
69 #include "common/efx.h"
70 
71 #include "sfxge.h"
72 #include "sfxge_tx.h"
73 
74 
75 #define	SFXGE_PARAM_TX_DPL_GET_MAX	SFXGE_PARAM(tx_dpl_get_max)
76 static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT;
77 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max);
78 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN,
79 	   &sfxge_tx_dpl_get_max, 0,
80 	   "Maximum number of any packets in deferred packet get-list");
81 
82 #define	SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \
83 	SFXGE_PARAM(tx_dpl_get_non_tcp_max)
84 static int sfxge_tx_dpl_get_non_tcp_max =
85 	SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT;
86 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max);
87 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN,
88 	   &sfxge_tx_dpl_get_non_tcp_max, 0,
89 	   "Maximum number of non-TCP packets in deferred packet get-list");
90 
91 #define	SFXGE_PARAM_TX_DPL_PUT_MAX	SFXGE_PARAM(tx_dpl_put_max)
92 static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT;
93 TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max);
94 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN,
95 	   &sfxge_tx_dpl_put_max, 0,
96 	   "Maximum number of any packets in deferred packet put-list");
97 
98 #define	SFXGE_PARAM_TSO_FW_ASSISTED	SFXGE_PARAM(tso_fw_assisted)
99 static int sfxge_tso_fw_assisted = 1;
100 TUNABLE_INT(SFXGE_PARAM_TSO_FW_ASSISTED, &sfxge_tso_fw_assisted);
101 SYSCTL_INT(_hw_sfxge, OID_AUTO, tso_fw_assisted, CTLFLAG_RDTUN,
102 	   &sfxge_tso_fw_assisted, 0,
103 	   "Use FW-assisted TSO if supported by NIC firmware");
104 
105 
106 static const struct {
107 	const char *name;
108 	size_t offset;
109 } sfxge_tx_stats[] = {
110 #define	SFXGE_TX_STAT(name, member) \
111 	{ #name, offsetof(struct sfxge_txq, member) }
112 	SFXGE_TX_STAT(tso_bursts, tso_bursts),
113 	SFXGE_TX_STAT(tso_packets, tso_packets),
114 	SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
115 	SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many),
116 	SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc),
117 	SFXGE_TX_STAT(tx_collapses, collapses),
118 	SFXGE_TX_STAT(tx_drops, drops),
119 	SFXGE_TX_STAT(tx_get_overflow, get_overflow),
120 	SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow),
121 	SFXGE_TX_STAT(tx_put_overflow, put_overflow),
122 	SFXGE_TX_STAT(tx_netdown_drops, netdown_drops),
123 };
124 
125 
126 /* Forward declarations. */
127 static void sfxge_tx_qdpl_service(struct sfxge_txq *txq);
128 static void sfxge_tx_qlist_post(struct sfxge_txq *txq);
129 static void sfxge_tx_qunblock(struct sfxge_txq *txq);
130 static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
131 			      const bus_dma_segment_t *dma_seg, int n_dma_seg,
132 			      int vlan_tagged);
133 
134 static int
135 sfxge_tx_maybe_insert_tag(struct sfxge_txq *txq, struct mbuf *mbuf)
136 {
137 	uint16_t this_tag = ((mbuf->m_flags & M_VLANTAG) ?
138 			     mbuf->m_pkthdr.ether_vtag :
139 			     0);
140 
141 	if (this_tag == txq->hw_vlan_tci)
142 		return (0);
143 
144 	efx_tx_qdesc_vlantci_create(txq->common,
145 				    bswap16(this_tag),
146 				    &txq->pend_desc[0]);
147 	txq->n_pend_desc = 1;
148 	txq->hw_vlan_tci = this_tag;
149 	return (1);
150 }
151 
152 static inline void
153 sfxge_next_stmp(struct sfxge_txq *txq, struct sfxge_tx_mapping **pstmp)
154 {
155 	KASSERT((*pstmp)->flags == 0, ("stmp flags are not 0"));
156 	if (__predict_false(*pstmp ==
157 			    &txq->stmp[txq->ptr_mask]))
158 		*pstmp = &txq->stmp[0];
159 	else
160 		(*pstmp)++;
161 }
162 
163 
164 void
165 sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq)
166 {
167 	unsigned int completed;
168 
169 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
170 
171 	completed = txq->completed;
172 	while (completed != txq->pending) {
173 		struct sfxge_tx_mapping *stmp;
174 		unsigned int id;
175 
176 		id = completed++ & txq->ptr_mask;
177 
178 		stmp = &txq->stmp[id];
179 		if (stmp->flags & TX_BUF_UNMAP) {
180 			bus_dmamap_unload(txq->packet_dma_tag, stmp->map);
181 			if (stmp->flags & TX_BUF_MBUF) {
182 				struct mbuf *m = stmp->u.mbuf;
183 				do
184 					m = m_free(m);
185 				while (m != NULL);
186 			} else {
187 				free(stmp->u.heap_buf, M_SFXGE);
188 			}
189 			stmp->flags = 0;
190 		}
191 	}
192 	txq->completed = completed;
193 
194 	/* Check whether we need to unblock the queue. */
195 	mb();
196 	if (txq->blocked) {
197 		unsigned int level;
198 
199 		level = txq->added - txq->completed;
200 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
201 			sfxge_tx_qunblock(txq);
202 	}
203 }
204 
205 static unsigned int
206 sfxge_is_mbuf_non_tcp(struct mbuf *mbuf)
207 {
208 	/* Absense of TCP checksum flags does not mean that it is non-TCP
209 	 * but it should be true if user wants to achieve high throughput.
210 	 */
211 	return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP)));
212 }
213 
214 /*
215  * Reorder the put list and append it to the get list.
216  */
217 static void
218 sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq)
219 {
220 	struct sfxge_tx_dpl *stdp;
221 	struct mbuf *mbuf, *get_next, **get_tailp;
222 	volatile uintptr_t *putp;
223 	uintptr_t put;
224 	unsigned int count;
225 	unsigned int non_tcp_count;
226 
227 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
228 
229 	stdp = &txq->dpl;
230 
231 	/* Acquire the put list. */
232 	putp = &stdp->std_put;
233 	put = atomic_readandclear_ptr(putp);
234 	mbuf = (void *)put;
235 
236 	if (mbuf == NULL)
237 		return;
238 
239 	/* Reverse the put list. */
240 	get_tailp = &mbuf->m_nextpkt;
241 	get_next = NULL;
242 
243 	count = 0;
244 	non_tcp_count = 0;
245 	do {
246 		struct mbuf *put_next;
247 
248 		non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf);
249 		put_next = mbuf->m_nextpkt;
250 		mbuf->m_nextpkt = get_next;
251 		get_next = mbuf;
252 		mbuf = put_next;
253 
254 		count++;
255 	} while (mbuf != NULL);
256 
257 	if (count > stdp->std_put_hiwat)
258 		stdp->std_put_hiwat = count;
259 
260 	/* Append the reversed put list to the get list. */
261 	KASSERT(*get_tailp == NULL, ("*get_tailp != NULL"));
262 	*stdp->std_getp = get_next;
263 	stdp->std_getp = get_tailp;
264 	stdp->std_get_count += count;
265 	stdp->std_get_non_tcp_count += non_tcp_count;
266 }
267 
268 static void
269 sfxge_tx_qreap(struct sfxge_txq *txq)
270 {
271 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
272 
273 	txq->reaped = txq->completed;
274 }
275 
276 static void
277 sfxge_tx_qlist_post(struct sfxge_txq *txq)
278 {
279 	unsigned int old_added;
280 	unsigned int block_level;
281 	unsigned int level;
282 	int rc;
283 
284 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
285 
286 	KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0"));
287 	KASSERT(txq->n_pend_desc <= txq->max_pkt_desc,
288 		("txq->n_pend_desc too large"));
289 	KASSERT(!txq->blocked, ("txq->blocked"));
290 
291 	old_added = txq->added;
292 
293 	/* Post the fragment list. */
294 	rc = efx_tx_qdesc_post(txq->common, txq->pend_desc, txq->n_pend_desc,
295 			  txq->reaped, &txq->added);
296 	KASSERT(rc == 0, ("efx_tx_qdesc_post() failed"));
297 
298 	/* If efx_tx_qdesc_post() had to refragment, our information about
299 	 * buffers to free may be associated with the wrong
300 	 * descriptors.
301 	 */
302 	KASSERT(txq->added - old_added == txq->n_pend_desc,
303 		("efx_tx_qdesc_post() refragmented descriptors"));
304 
305 	level = txq->added - txq->reaped;
306 	KASSERT(level <= txq->entries, ("overfilled TX queue"));
307 
308 	/* Clear the fragment list. */
309 	txq->n_pend_desc = 0;
310 
311 	/*
312 	 * Set the block level to ensure there is space to generate a
313 	 * large number of descriptors for TSO.
314 	 */
315 	block_level = EFX_TXQ_LIMIT(txq->entries) - txq->max_pkt_desc;
316 
317 	/* Have we reached the block level? */
318 	if (level < block_level)
319 		return;
320 
321 	/* Reap, and check again */
322 	sfxge_tx_qreap(txq);
323 	level = txq->added - txq->reaped;
324 	if (level < block_level)
325 		return;
326 
327 	txq->blocked = 1;
328 
329 	/*
330 	 * Avoid a race with completion interrupt handling that could leave
331 	 * the queue blocked.
332 	 */
333 	mb();
334 	sfxge_tx_qreap(txq);
335 	level = txq->added - txq->reaped;
336 	if (level < block_level) {
337 		mb();
338 		txq->blocked = 0;
339 	}
340 }
341 
342 static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
343 {
344 	bus_dmamap_t *used_map;
345 	bus_dmamap_t map;
346 	bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG];
347 	unsigned int id;
348 	struct sfxge_tx_mapping *stmp;
349 	efx_desc_t *desc;
350 	int n_dma_seg;
351 	int rc;
352 	int i;
353 	int eop;
354 	int vlan_tagged;
355 
356 	KASSERT(!txq->blocked, ("txq->blocked"));
357 
358 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
359 		prefetch_read_many(mbuf->m_data);
360 
361 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) {
362 		rc = EINTR;
363 		goto reject;
364 	}
365 
366 	/* Load the packet for DMA. */
367 	id = txq->added & txq->ptr_mask;
368 	stmp = &txq->stmp[id];
369 	rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map,
370 				     mbuf, dma_seg, &n_dma_seg, 0);
371 	if (rc == EFBIG) {
372 		/* Try again. */
373 		struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT,
374 						   SFXGE_TX_MAPPING_MAX_SEG);
375 		if (new_mbuf == NULL)
376 			goto reject;
377 		++txq->collapses;
378 		mbuf = new_mbuf;
379 		rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag,
380 					     stmp->map, mbuf,
381 					     dma_seg, &n_dma_seg, 0);
382 	}
383 	if (rc != 0)
384 		goto reject;
385 
386 	/* Make the packet visible to the hardware. */
387 	bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE);
388 
389 	used_map = &stmp->map;
390 
391 	vlan_tagged = sfxge_tx_maybe_insert_tag(txq, mbuf);
392 	if (vlan_tagged) {
393 		sfxge_next_stmp(txq, &stmp);
394 	}
395 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
396 		rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg, vlan_tagged);
397 		if (rc < 0)
398 			goto reject_mapped;
399 		stmp = &txq->stmp[(rc - 1) & txq->ptr_mask];
400 	} else {
401 		/* Add the mapping to the fragment list, and set flags
402 		 * for the buffer.
403 		 */
404 
405 		i = 0;
406 		for (;;) {
407 			desc = &txq->pend_desc[i + vlan_tagged];
408 			eop = (i == n_dma_seg - 1);
409 			efx_tx_qdesc_dma_create(txq->common,
410 						dma_seg[i].ds_addr,
411 						dma_seg[i].ds_len,
412 						eop,
413 						desc);
414 			if (eop)
415 				break;
416 			i++;
417 			sfxge_next_stmp(txq, &stmp);
418 		}
419 		txq->n_pend_desc = n_dma_seg + vlan_tagged;
420 	}
421 
422 	/*
423 	 * If the mapping required more than one descriptor
424 	 * then we need to associate the DMA map with the last
425 	 * descriptor, not the first.
426 	 */
427 	if (used_map != &stmp->map) {
428 		map = stmp->map;
429 		stmp->map = *used_map;
430 		*used_map = map;
431 	}
432 
433 	stmp->u.mbuf = mbuf;
434 	stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF;
435 
436 	/* Post the fragment list. */
437 	sfxge_tx_qlist_post(txq);
438 
439 	return (0);
440 
441 reject_mapped:
442 	bus_dmamap_unload(txq->packet_dma_tag, *used_map);
443 reject:
444 	/* Drop the packet on the floor. */
445 	m_freem(mbuf);
446 	++txq->drops;
447 
448 	return (rc);
449 }
450 
451 /*
452  * Drain the deferred packet list into the transmit queue.
453  */
454 static void
455 sfxge_tx_qdpl_drain(struct sfxge_txq *txq)
456 {
457 	struct sfxge_softc *sc;
458 	struct sfxge_tx_dpl *stdp;
459 	struct mbuf *mbuf, *next;
460 	unsigned int count;
461 	unsigned int non_tcp_count;
462 	unsigned int pushed;
463 	int rc;
464 
465 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
466 
467 	sc = txq->sc;
468 	stdp = &txq->dpl;
469 	pushed = txq->added;
470 
471 	if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) {
472 		prefetch_read_many(sc->enp);
473 		prefetch_read_many(txq->common);
474 	}
475 
476 	mbuf = stdp->std_get;
477 	count = stdp->std_get_count;
478 	non_tcp_count = stdp->std_get_non_tcp_count;
479 
480 	if (count > stdp->std_get_hiwat)
481 		stdp->std_get_hiwat = count;
482 
483 	while (count != 0) {
484 		KASSERT(mbuf != NULL, ("mbuf == NULL"));
485 
486 		next = mbuf->m_nextpkt;
487 		mbuf->m_nextpkt = NULL;
488 
489 		ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */
490 
491 		if (next != NULL)
492 			prefetch_read_many(next);
493 
494 		rc = sfxge_tx_queue_mbuf(txq, mbuf);
495 		--count;
496 		non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf);
497 		mbuf = next;
498 		if (rc != 0)
499 			continue;
500 
501 		if (txq->blocked)
502 			break;
503 
504 		/* Push the fragments to the hardware in batches. */
505 		if (txq->added - pushed >= SFXGE_TX_BATCH) {
506 			efx_tx_qpush(txq->common, txq->added, pushed);
507 			pushed = txq->added;
508 		}
509 	}
510 
511 	if (count == 0) {
512 		KASSERT(mbuf == NULL, ("mbuf != NULL"));
513 		KASSERT(non_tcp_count == 0,
514 			("inconsistent TCP/non-TCP detection"));
515 		stdp->std_get = NULL;
516 		stdp->std_get_count = 0;
517 		stdp->std_get_non_tcp_count = 0;
518 		stdp->std_getp = &stdp->std_get;
519 	} else {
520 		stdp->std_get = mbuf;
521 		stdp->std_get_count = count;
522 		stdp->std_get_non_tcp_count = non_tcp_count;
523 	}
524 
525 	if (txq->added != pushed)
526 		efx_tx_qpush(txq->common, txq->added, pushed);
527 
528 	KASSERT(txq->blocked || stdp->std_get_count == 0,
529 		("queue unblocked but count is non-zero"));
530 }
531 
532 #define	SFXGE_TX_QDPL_PENDING(_txq)	((_txq)->dpl.std_put != 0)
533 
534 /*
535  * Service the deferred packet list.
536  *
537  * NOTE: drops the txq mutex!
538  */
539 static void
540 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
541 {
542 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
543 
544 	do {
545 		if (SFXGE_TX_QDPL_PENDING(txq))
546 			sfxge_tx_qdpl_swizzle(txq);
547 
548 		if (!txq->blocked)
549 			sfxge_tx_qdpl_drain(txq);
550 
551 		SFXGE_TXQ_UNLOCK(txq);
552 	} while (SFXGE_TX_QDPL_PENDING(txq) &&
553 		 SFXGE_TXQ_TRYLOCK(txq));
554 }
555 
556 /*
557  * Put a packet on the deferred packet get-list.
558  */
559 static int
560 sfxge_tx_qdpl_put_locked(struct sfxge_txq *txq, struct mbuf *mbuf)
561 {
562 	struct sfxge_tx_dpl *stdp;
563 
564 	stdp = &txq->dpl;
565 
566 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
567 
568 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
569 
570 	if (stdp->std_get_count >= stdp->std_get_max) {
571 		txq->get_overflow++;
572 		return (ENOBUFS);
573 	}
574 	if (sfxge_is_mbuf_non_tcp(mbuf)) {
575 		if (stdp->std_get_non_tcp_count >=
576 		    stdp->std_get_non_tcp_max) {
577 			txq->get_non_tcp_overflow++;
578 			return (ENOBUFS);
579 		}
580 		stdp->std_get_non_tcp_count++;
581 	}
582 
583 	*(stdp->std_getp) = mbuf;
584 	stdp->std_getp = &mbuf->m_nextpkt;
585 	stdp->std_get_count++;
586 
587 	return (0);
588 }
589 
590 /*
591  * Put a packet on the deferred packet put-list.
592  *
593  * We overload the csum_data field in the mbuf to keep track of this length
594  * because there is no cheap alternative to avoid races.
595  */
596 static int
597 sfxge_tx_qdpl_put_unlocked(struct sfxge_txq *txq, struct mbuf *mbuf)
598 {
599 	struct sfxge_tx_dpl *stdp;
600 	volatile uintptr_t *putp;
601 	uintptr_t old;
602 	uintptr_t new;
603 	unsigned old_len;
604 
605 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
606 
607 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
608 
609 	stdp = &txq->dpl;
610 	putp = &stdp->std_put;
611 	new = (uintptr_t)mbuf;
612 
613 	do {
614 		old = *putp;
615 		if (old != 0) {
616 			struct mbuf *mp = (struct mbuf *)old;
617 			old_len = mp->m_pkthdr.csum_data;
618 		} else
619 			old_len = 0;
620 		if (old_len >= stdp->std_put_max) {
621 			atomic_add_long(&txq->put_overflow, 1);
622 			return (ENOBUFS);
623 		}
624 		mbuf->m_pkthdr.csum_data = old_len + 1;
625 		mbuf->m_nextpkt = (void *)old;
626 	} while (atomic_cmpset_ptr(putp, old, new) == 0);
627 
628 	return (0);
629 }
630 
631 /*
632  * Called from if_transmit - will try to grab the txq lock and enqueue to the
633  * put list if it succeeds, otherwise try to push onto the defer list if space.
634  */
635 static int
636 sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m)
637 {
638 	int rc;
639 
640 	if (!SFXGE_LINK_UP(txq->sc)) {
641 		atomic_add_long(&txq->netdown_drops, 1);
642 		return (ENETDOWN);
643 	}
644 
645 	/*
646 	 * Try to grab the txq lock.  If we are able to get the lock,
647 	 * the packet will be appended to the "get list" of the deferred
648 	 * packet list.  Otherwise, it will be pushed on the "put list".
649 	 */
650 	if (SFXGE_TXQ_TRYLOCK(txq)) {
651 		/* First swizzle put-list to get-list to keep order */
652 		sfxge_tx_qdpl_swizzle(txq);
653 
654 		rc = sfxge_tx_qdpl_put_locked(txq, m);
655 
656 		/* Try to service the list. */
657 		sfxge_tx_qdpl_service(txq);
658 		/* Lock has been dropped. */
659 	} else {
660 		rc = sfxge_tx_qdpl_put_unlocked(txq, m);
661 
662 		/*
663 		 * Try to grab the lock again.
664 		 *
665 		 * If we are able to get the lock, we need to process
666 		 * the deferred packet list.  If we are not able to get
667 		 * the lock, another thread is processing the list.
668 		 */
669 		if ((rc == 0) && SFXGE_TXQ_TRYLOCK(txq)) {
670 			sfxge_tx_qdpl_service(txq);
671 			/* Lock has been dropped. */
672 		}
673 	}
674 
675 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
676 
677 	return (rc);
678 }
679 
680 static void
681 sfxge_tx_qdpl_flush(struct sfxge_txq *txq)
682 {
683 	struct sfxge_tx_dpl *stdp = &txq->dpl;
684 	struct mbuf *mbuf, *next;
685 
686 	SFXGE_TXQ_LOCK(txq);
687 
688 	sfxge_tx_qdpl_swizzle(txq);
689 	for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) {
690 		next = mbuf->m_nextpkt;
691 		m_freem(mbuf);
692 	}
693 	stdp->std_get = NULL;
694 	stdp->std_get_count = 0;
695 	stdp->std_get_non_tcp_count = 0;
696 	stdp->std_getp = &stdp->std_get;
697 
698 	SFXGE_TXQ_UNLOCK(txq);
699 }
700 
701 void
702 sfxge_if_qflush(struct ifnet *ifp)
703 {
704 	struct sfxge_softc *sc;
705 	unsigned int i;
706 
707 	sc = ifp->if_softc;
708 
709 	for (i = 0; i < sc->txq_count; i++)
710 		sfxge_tx_qdpl_flush(sc->txq[i]);
711 }
712 
713 #if SFXGE_TX_PARSE_EARLY
714 
715 /* There is little space for user data in mbuf pkthdr, so we
716  * use l*hlen fields which are not used by the driver otherwise
717  * to store header offsets.
718  * The fields are 8-bit, but it's ok, no header may be longer than 255 bytes.
719  */
720 
721 
722 #define TSO_MBUF_PROTO(_mbuf)    ((_mbuf)->m_pkthdr.PH_loc.sixteen[0])
723 /* We abuse l5hlen here because PH_loc can hold only 64 bits of data */
724 #define TSO_MBUF_FLAGS(_mbuf)    ((_mbuf)->m_pkthdr.l5hlen)
725 #define TSO_MBUF_PACKETID(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[1])
726 #define TSO_MBUF_SEQNUM(_mbuf)   ((_mbuf)->m_pkthdr.PH_loc.thirtytwo[1])
727 
728 static void sfxge_parse_tx_packet(struct mbuf *mbuf)
729 {
730 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
731 	const struct tcphdr *th;
732 	struct tcphdr th_copy;
733 
734 	/* Find network protocol and header */
735 	TSO_MBUF_PROTO(mbuf) = eh->ether_type;
736 	if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_VLAN)) {
737 		struct ether_vlan_header *veh =
738 			mtod(mbuf, struct ether_vlan_header *);
739 		TSO_MBUF_PROTO(mbuf) = veh->evl_proto;
740 		mbuf->m_pkthdr.l2hlen = sizeof(*veh);
741 	} else {
742 		mbuf->m_pkthdr.l2hlen = sizeof(*eh);
743 	}
744 
745 	/* Find TCP header */
746 	if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IP)) {
747 		const struct ip *iph = (const struct ip *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen);
748 
749 		KASSERT(iph->ip_p == IPPROTO_TCP,
750 			("TSO required on non-TCP packet"));
751 		mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + 4 * iph->ip_hl;
752 		TSO_MBUF_PACKETID(mbuf) = iph->ip_id;
753 	} else {
754 		KASSERT(TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IPV6),
755 			("TSO required on non-IP packet"));
756 		KASSERT(((const struct ip6_hdr *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen))->ip6_nxt ==
757 			IPPROTO_TCP,
758 			("TSO required on non-TCP packet"));
759 		mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + sizeof(struct ip6_hdr);
760 		TSO_MBUF_PACKETID(mbuf) = 0;
761 	}
762 
763 	KASSERT(mbuf->m_len >= mbuf->m_pkthdr.l3hlen,
764 		("network header is fragmented in mbuf"));
765 
766 	/* We need TCP header including flags (window is the next) */
767 	if (mbuf->m_len < mbuf->m_pkthdr.l3hlen + offsetof(struct tcphdr, th_win)) {
768 		m_copydata(mbuf, mbuf->m_pkthdr.l3hlen, sizeof(th_copy),
769 			   (caddr_t)&th_copy);
770 		th = &th_copy;
771 	} else {
772 		th = (const struct tcphdr *)mtodo(mbuf, mbuf->m_pkthdr.l3hlen);
773 	}
774 
775 	mbuf->m_pkthdr.l4hlen = mbuf->m_pkthdr.l3hlen + 4 * th->th_off;
776 	TSO_MBUF_SEQNUM(mbuf) = ntohl(th->th_seq);
777 
778 	/* These flags must not be duplicated */
779 	/*
780 	 * RST should not be duplicated as well, but FreeBSD kernel
781 	 * generates TSO packets with RST flag. So, do not assert
782 	 * its absence.
783 	 */
784 	KASSERT(!(th->th_flags & (TH_URG | TH_SYN)),
785 		("incompatible TCP flag 0x%x on TSO packet",
786 		 th->th_flags & (TH_URG | TH_SYN)));
787 	TSO_MBUF_FLAGS(mbuf) = th->th_flags;
788 }
789 #endif
790 
791 /*
792  * TX start -- called by the stack.
793  */
794 int
795 sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m)
796 {
797 	struct sfxge_softc *sc;
798 	struct sfxge_txq *txq;
799 	int rc;
800 
801 	sc = (struct sfxge_softc *)ifp->if_softc;
802 
803 	/*
804 	 * Transmit may be called when interface is up from the kernel
805 	 * point of view, but not yet up (in progress) from the driver
806 	 * point of view. I.e. link aggregation bring up.
807 	 * Transmit may be called when interface is up from the driver
808 	 * point of view, but already down from the kernel point of
809 	 * view. I.e. Rx when interface shutdown is in progress.
810 	 */
811 	KASSERT((ifp->if_flags & IFF_UP) || (sc->if_flags & IFF_UP),
812 		("interface not up"));
813 
814 	/* Pick the desired transmit queue. */
815 	if (m->m_pkthdr.csum_flags &
816 	    (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO)) {
817 		int index = 0;
818 
819 		/* check if flowid is set */
820 		if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
821 			uint32_t hash = m->m_pkthdr.flowid;
822 
823 			index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX];
824 		}
825 #if SFXGE_TX_PARSE_EARLY
826 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
827 			sfxge_parse_tx_packet(m);
828 #endif
829 		txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index];
830 	} else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
831 		txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
832 	} else {
833 		txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
834 	}
835 
836 	rc = sfxge_tx_packet_add(txq, m);
837 	if (rc != 0)
838 		m_freem(m);
839 
840 	return (rc);
841 }
842 
843 /*
844  * Software "TSO".  Not quite as good as doing it in hardware, but
845  * still faster than segmenting in the stack.
846  */
847 
848 struct sfxge_tso_state {
849 	/* Output position */
850 	unsigned out_len;	/* Remaining length in current segment */
851 	unsigned seqnum;	/* Current sequence number */
852 	unsigned packet_space;	/* Remaining space in current packet */
853 
854 	/* Input position */
855 	uint64_t dma_addr;	/* DMA address of current position */
856 	unsigned in_len;	/* Remaining length in current mbuf */
857 
858 	const struct mbuf *mbuf; /* Input mbuf (head of chain) */
859 	u_short protocol;	/* Network protocol (after VLAN decap) */
860 	ssize_t nh_off;		/* Offset of network header */
861 	ssize_t tcph_off;	/* Offset of TCP header */
862 	unsigned header_len;	/* Number of bytes of header */
863 	unsigned seg_size;	/* TCP segment size */
864 	int fw_assisted;	/* Use FW-assisted TSO */
865 	u_short packet_id;	/* IPv4 packet ID from the original packet */
866 	uint8_t tcp_flags;	/* TCP flags */
867 	efx_desc_t header_desc; /* Precomputed header descriptor for
868 				 * FW-assisted TSO */
869 };
870 
871 #if !SFXGE_TX_PARSE_EARLY
872 static const struct ip *tso_iph(const struct sfxge_tso_state *tso)
873 {
874 	KASSERT(tso->protocol == htons(ETHERTYPE_IP),
875 		("tso_iph() in non-IPv4 state"));
876 	return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
877 }
878 
879 static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
880 {
881 	KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
882 		("tso_ip6h() in non-IPv6 state"));
883 	return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
884 }
885 
886 static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
887 {
888 	return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
889 }
890 #endif
891 
892 
893 /* Size of preallocated TSO header buffers.  Larger blocks must be
894  * allocated from the heap.
895  */
896 #define	TSOH_STD_SIZE	128
897 
898 /* At most half the descriptors in the queue at any time will refer to
899  * a TSO header buffer, since they must always be followed by a
900  * payload descriptor referring to an mbuf.
901  */
902 #define	TSOH_COUNT(_txq_entries)	((_txq_entries) / 2u)
903 #define	TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
904 #define	TSOH_PAGE_COUNT(_txq_entries)	\
905 	((TSOH_COUNT(_txq_entries) + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE)
906 
907 static int tso_init(struct sfxge_txq *txq)
908 {
909 	struct sfxge_softc *sc = txq->sc;
910 	unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries);
911 	int i, rc;
912 
913 	/* Allocate TSO header buffers */
914 	txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]),
915 				  M_SFXGE, M_WAITOK);
916 
917 	for (i = 0; i < tsoh_page_count; i++) {
918 		rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
919 		if (rc != 0)
920 			goto fail;
921 	}
922 
923 	return (0);
924 
925 fail:
926 	while (i-- > 0)
927 		sfxge_dma_free(&txq->tsoh_buffer[i]);
928 	free(txq->tsoh_buffer, M_SFXGE);
929 	txq->tsoh_buffer = NULL;
930 	return (rc);
931 }
932 
933 static void tso_fini(struct sfxge_txq *txq)
934 {
935 	int i;
936 
937 	if (txq->tsoh_buffer != NULL) {
938 		for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++)
939 			sfxge_dma_free(&txq->tsoh_buffer[i]);
940 		free(txq->tsoh_buffer, M_SFXGE);
941 	}
942 }
943 
944 static void tso_start(struct sfxge_txq *txq, struct sfxge_tso_state *tso,
945 		      const bus_dma_segment_t *hdr_dma_seg,
946 		      struct mbuf *mbuf)
947 {
948 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->sc->enp);
949 #if !SFXGE_TX_PARSE_EARLY
950 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
951 	const struct tcphdr *th;
952 	struct tcphdr th_copy;
953 #endif
954 
955 	tso->fw_assisted = txq->sc->tso_fw_assisted;
956 	tso->mbuf = mbuf;
957 
958 	/* Find network protocol and header */
959 #if !SFXGE_TX_PARSE_EARLY
960 	tso->protocol = eh->ether_type;
961 	if (tso->protocol == htons(ETHERTYPE_VLAN)) {
962 		struct ether_vlan_header *veh =
963 			mtod(mbuf, struct ether_vlan_header *);
964 		tso->protocol = veh->evl_proto;
965 		tso->nh_off = sizeof(*veh);
966 	} else {
967 		tso->nh_off = sizeof(*eh);
968 	}
969 #else
970 	tso->protocol = TSO_MBUF_PROTO(mbuf);
971 	tso->nh_off = mbuf->m_pkthdr.l2hlen;
972 	tso->tcph_off = mbuf->m_pkthdr.l3hlen;
973 	tso->packet_id = TSO_MBUF_PACKETID(mbuf);
974 #endif
975 
976 #if !SFXGE_TX_PARSE_EARLY
977 	/* Find TCP header */
978 	if (tso->protocol == htons(ETHERTYPE_IP)) {
979 		KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
980 			("TSO required on non-TCP packet"));
981 		tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
982 		tso->packet_id = tso_iph(tso)->ip_id;
983 	} else {
984 		KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
985 			("TSO required on non-IP packet"));
986 		KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
987 			("TSO required on non-TCP packet"));
988 		tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
989 		tso->packet_id = 0;
990 	}
991 #endif
992 
993 
994 	if (tso->fw_assisted &&
995 	    __predict_false(tso->tcph_off >
996 			    encp->enc_tx_tso_tcp_header_offset_limit)) {
997 		tso->fw_assisted = 0;
998 	}
999 
1000 
1001 #if !SFXGE_TX_PARSE_EARLY
1002 	KASSERT(mbuf->m_len >= tso->tcph_off,
1003 		("network header is fragmented in mbuf"));
1004 	/* We need TCP header including flags (window is the next) */
1005 	if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) {
1006 		m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy),
1007 			   (caddr_t)&th_copy);
1008 		th = &th_copy;
1009 	} else {
1010 		th = tso_tcph(tso);
1011 	}
1012 	tso->header_len = tso->tcph_off + 4 * th->th_off;
1013 #else
1014 	tso->header_len = mbuf->m_pkthdr.l4hlen;
1015 #endif
1016 	tso->seg_size = mbuf->m_pkthdr.tso_segsz;
1017 
1018 #if !SFXGE_TX_PARSE_EARLY
1019 	tso->seqnum = ntohl(th->th_seq);
1020 
1021 	/* These flags must not be duplicated */
1022 	/*
1023 	 * RST should not be duplicated as well, but FreeBSD kernel
1024 	 * generates TSO packets with RST flag. So, do not assert
1025 	 * its absence.
1026 	 */
1027 	KASSERT(!(th->th_flags & (TH_URG | TH_SYN)),
1028 		("incompatible TCP flag 0x%x on TSO packet",
1029 		 th->th_flags & (TH_URG | TH_SYN)));
1030 	tso->tcp_flags = th->th_flags;
1031 #else
1032 	tso->seqnum = TSO_MBUF_SEQNUM(mbuf);
1033 	tso->tcp_flags = TSO_MBUF_FLAGS(mbuf);
1034 #endif
1035 
1036 	tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
1037 
1038 	if (tso->fw_assisted) {
1039 		if (hdr_dma_seg->ds_len >= tso->header_len)
1040 			efx_tx_qdesc_dma_create(txq->common,
1041 						hdr_dma_seg->ds_addr,
1042 						tso->header_len,
1043 						B_FALSE,
1044 						&tso->header_desc);
1045 		else
1046 			tso->fw_assisted = 0;
1047 	}
1048 }
1049 
1050 /*
1051  * tso_fill_packet_with_fragment - form descriptors for the current fragment
1052  *
1053  * Form descriptors for the current fragment, until we reach the end
1054  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
1055  * space.
1056  */
1057 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
1058 					  struct sfxge_tso_state *tso)
1059 {
1060 	efx_desc_t *desc;
1061 	int n;
1062 
1063 	if (tso->in_len == 0 || tso->packet_space == 0)
1064 		return;
1065 
1066 	KASSERT(tso->in_len > 0, ("TSO input length went negative"));
1067 	KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
1068 
1069 	n = min(tso->in_len, tso->packet_space);
1070 
1071 	tso->packet_space -= n;
1072 	tso->out_len -= n;
1073 	tso->in_len -= n;
1074 
1075 	desc = &txq->pend_desc[txq->n_pend_desc++];
1076 	efx_tx_qdesc_dma_create(txq->common,
1077 				tso->dma_addr,
1078 				n,
1079 				tso->out_len == 0 || tso->packet_space == 0,
1080 				desc);
1081 
1082 	tso->dma_addr += n;
1083 }
1084 
1085 /* Callback from bus_dmamap_load() for long TSO headers. */
1086 static void tso_map_long_header(void *dma_addr_ret,
1087 				bus_dma_segment_t *segs, int nseg,
1088 				int error)
1089 {
1090 	*(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
1091 				      __predict_true(nseg == 1)) ?
1092 				     segs->ds_addr : 0);
1093 }
1094 
1095 /*
1096  * tso_start_new_packet - generate a new header and prepare for the new packet
1097  *
1098  * Generate a new header and prepare for the new packet.  Return 0 on
1099  * success, or an error code if failed to alloc header.
1100  */
1101 static int tso_start_new_packet(struct sfxge_txq *txq,
1102 				struct sfxge_tso_state *tso,
1103 				unsigned int *idp)
1104 {
1105 	unsigned int id = *idp;
1106 	struct tcphdr *tsoh_th;
1107 	unsigned ip_length;
1108 	caddr_t header;
1109 	uint64_t dma_addr;
1110 	bus_dmamap_t map;
1111 	efx_desc_t *desc;
1112 	int rc;
1113 
1114 	if (tso->fw_assisted) {
1115 		uint8_t tcp_flags = tso->tcp_flags;
1116 
1117 		if (tso->out_len > tso->seg_size)
1118 			tcp_flags &= ~(TH_FIN | TH_PUSH);
1119 
1120 		/* TSO option descriptor */
1121 		desc = &txq->pend_desc[txq->n_pend_desc++];
1122 		efx_tx_qdesc_tso_create(txq->common,
1123 					tso->packet_id,
1124 					tso->seqnum,
1125 					tcp_flags,
1126 					desc++);
1127 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1128 		id = (id + 1) & txq->ptr_mask;
1129 
1130 		/* Header DMA descriptor */
1131 		*desc = tso->header_desc;
1132 		txq->n_pend_desc++;
1133 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1134 		id = (id + 1) & txq->ptr_mask;
1135 
1136 		tso->seqnum += tso->seg_size;
1137 	} else {
1138 		/* Allocate a DMA-mapped header buffer. */
1139 		if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
1140 			unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
1141 			unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
1142 
1143 			header = (txq->tsoh_buffer[page_index].esm_base +
1144 				  buf_index * TSOH_STD_SIZE);
1145 			dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
1146 				    buf_index * TSOH_STD_SIZE);
1147 			map = txq->tsoh_buffer[page_index].esm_map;
1148 
1149 			KASSERT(txq->stmp[id].flags == 0,
1150 				("stmp flags are not 0"));
1151 		} else {
1152 			struct sfxge_tx_mapping *stmp = &txq->stmp[id];
1153 
1154 			/* We cannot use bus_dmamem_alloc() as that may sleep */
1155 			header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
1156 			if (__predict_false(!header))
1157 				return (ENOMEM);
1158 			rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
1159 					     header, tso->header_len,
1160 					     tso_map_long_header, &dma_addr,
1161 					     BUS_DMA_NOWAIT);
1162 			if (__predict_false(dma_addr == 0)) {
1163 				if (rc == 0) {
1164 					/* Succeeded but got >1 segment */
1165 					bus_dmamap_unload(txq->packet_dma_tag,
1166 							  stmp->map);
1167 					rc = EINVAL;
1168 				}
1169 				free(header, M_SFXGE);
1170 				return (rc);
1171 			}
1172 			map = stmp->map;
1173 
1174 			txq->tso_long_headers++;
1175 			stmp->u.heap_buf = header;
1176 			stmp->flags = TX_BUF_UNMAP;
1177 		}
1178 
1179 		tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
1180 
1181 		/* Copy and update the headers. */
1182 		m_copydata(tso->mbuf, 0, tso->header_len, header);
1183 
1184 		tsoh_th->th_seq = htonl(tso->seqnum);
1185 		tso->seqnum += tso->seg_size;
1186 		if (tso->out_len > tso->seg_size) {
1187 			/* This packet will not finish the TSO burst. */
1188 			ip_length = tso->header_len - tso->nh_off + tso->seg_size;
1189 			tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH);
1190 		} else {
1191 			/* This packet will be the last in the TSO burst. */
1192 			ip_length = tso->header_len - tso->nh_off + tso->out_len;
1193 		}
1194 
1195 		if (tso->protocol == htons(ETHERTYPE_IP)) {
1196 			struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
1197 			tsoh_iph->ip_len = htons(ip_length);
1198 			/* XXX We should increment ip_id, but FreeBSD doesn't
1199 			 * currently allocate extra IDs for multiple segments.
1200 			 */
1201 		} else {
1202 			struct ip6_hdr *tsoh_iph =
1203 				(struct ip6_hdr *)(header + tso->nh_off);
1204 			tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
1205 		}
1206 
1207 		/* Make the header visible to the hardware. */
1208 		bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
1209 
1210 		/* Form a descriptor for this header. */
1211 		desc = &txq->pend_desc[txq->n_pend_desc++];
1212 		efx_tx_qdesc_dma_create(txq->common,
1213 					dma_addr,
1214 					tso->header_len,
1215 					0,
1216 					desc);
1217 		id = (id + 1) & txq->ptr_mask;
1218 	}
1219 	tso->packet_space = tso->seg_size;
1220 	txq->tso_packets++;
1221 	*idp = id;
1222 
1223 	return (0);
1224 }
1225 
1226 static int
1227 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
1228 		   const bus_dma_segment_t *dma_seg, int n_dma_seg,
1229 		   int vlan_tagged)
1230 {
1231 	struct sfxge_tso_state tso;
1232 	unsigned int id;
1233 	unsigned skipped = 0;
1234 
1235 	tso_start(txq, &tso, dma_seg, mbuf);
1236 
1237 	while (dma_seg->ds_len + skipped <= tso.header_len) {
1238 		skipped += dma_seg->ds_len;
1239 		--n_dma_seg;
1240 		KASSERT(n_dma_seg, ("no payload found in TSO packet"));
1241 		++dma_seg;
1242 	}
1243 	tso.in_len = dma_seg->ds_len - (tso.header_len - skipped);
1244 	tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped);
1245 
1246 	id = (txq->added + vlan_tagged) & txq->ptr_mask;
1247 	if (__predict_false(tso_start_new_packet(txq, &tso, &id)))
1248 		return (-1);
1249 
1250 	while (1) {
1251 		tso_fill_packet_with_fragment(txq, &tso);
1252 		/* Exactly one DMA descriptor is added */
1253 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1254 		id = (id + 1) & txq->ptr_mask;
1255 
1256 		/* Move onto the next fragment? */
1257 		if (tso.in_len == 0) {
1258 			--n_dma_seg;
1259 			if (n_dma_seg == 0)
1260 				break;
1261 			++dma_seg;
1262 			tso.in_len = dma_seg->ds_len;
1263 			tso.dma_addr = dma_seg->ds_addr;
1264 		}
1265 
1266 		/* End of packet? */
1267 		if (tso.packet_space == 0) {
1268 			/* If the queue is now full due to tiny MSS,
1269 			 * or we can't create another header, discard
1270 			 * the remainder of the input mbuf but do not
1271 			 * roll back the work we have done.
1272 			 */
1273 			if (txq->n_pend_desc + tso.fw_assisted +
1274 			    1 /* header */ + n_dma_seg >
1275 			    txq->max_pkt_desc) {
1276 				txq->tso_pdrop_too_many++;
1277 				break;
1278 			}
1279 			if (__predict_false(tso_start_new_packet(txq, &tso,
1280 								 &id))) {
1281 				txq->tso_pdrop_no_rsrc++;
1282 				break;
1283 			}
1284 		}
1285 	}
1286 
1287 	txq->tso_bursts++;
1288 	return (id);
1289 }
1290 
1291 static void
1292 sfxge_tx_qunblock(struct sfxge_txq *txq)
1293 {
1294 	struct sfxge_softc *sc;
1295 	struct sfxge_evq *evq;
1296 
1297 	sc = txq->sc;
1298 	evq = sc->evq[txq->evq_index];
1299 
1300 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
1301 
1302 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED))
1303 		return;
1304 
1305 	SFXGE_TXQ_LOCK(txq);
1306 
1307 	if (txq->blocked) {
1308 		unsigned int level;
1309 
1310 		level = txq->added - txq->completed;
1311 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) {
1312 			/* reaped must be in sync with blocked */
1313 			sfxge_tx_qreap(txq);
1314 			txq->blocked = 0;
1315 		}
1316 	}
1317 
1318 	sfxge_tx_qdpl_service(txq);
1319 	/* note: lock has been dropped */
1320 }
1321 
1322 void
1323 sfxge_tx_qflush_done(struct sfxge_txq *txq)
1324 {
1325 
1326 	txq->flush_state = SFXGE_FLUSH_DONE;
1327 }
1328 
1329 static void
1330 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1331 {
1332 	struct sfxge_txq *txq;
1333 	struct sfxge_evq *evq;
1334 	unsigned int count;
1335 
1336 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1337 
1338 	txq = sc->txq[index];
1339 	evq = sc->evq[txq->evq_index];
1340 
1341 	SFXGE_EVQ_LOCK(evq);
1342 	SFXGE_TXQ_LOCK(txq);
1343 
1344 	KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1345 	    ("txq->init_state != SFXGE_TXQ_STARTED"));
1346 
1347 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1348 
1349 	if (txq->flush_state != SFXGE_FLUSH_DONE) {
1350 		txq->flush_state = SFXGE_FLUSH_PENDING;
1351 
1352 		SFXGE_EVQ_UNLOCK(evq);
1353 		SFXGE_TXQ_UNLOCK(txq);
1354 
1355 		/* Flush the transmit queue. */
1356 		if (efx_tx_qflush(txq->common) != 0) {
1357 			log(LOG_ERR, "%s: Flushing Tx queue %u failed\n",
1358 			    device_get_nameunit(sc->dev), index);
1359 			txq->flush_state = SFXGE_FLUSH_DONE;
1360 		} else {
1361 			count = 0;
1362 			do {
1363 				/* Spin for 100ms. */
1364 				DELAY(100000);
1365 				if (txq->flush_state != SFXGE_FLUSH_PENDING)
1366 					break;
1367 			} while (++count < 20);
1368 		}
1369 		SFXGE_EVQ_LOCK(evq);
1370 		SFXGE_TXQ_LOCK(txq);
1371 
1372 		KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1373 		    ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1374 
1375 		if (txq->flush_state != SFXGE_FLUSH_DONE) {
1376 			/* Flush timeout */
1377 			log(LOG_ERR, "%s: Cannot flush Tx queue %u\n",
1378 			    device_get_nameunit(sc->dev), index);
1379 			txq->flush_state = SFXGE_FLUSH_DONE;
1380 		}
1381 	}
1382 
1383 	txq->blocked = 0;
1384 	txq->pending = txq->added;
1385 
1386 	sfxge_tx_qcomplete(txq, evq);
1387 	KASSERT(txq->completed == txq->added,
1388 	    ("txq->completed != txq->added"));
1389 
1390 	sfxge_tx_qreap(txq);
1391 	KASSERT(txq->reaped == txq->completed,
1392 	    ("txq->reaped != txq->completed"));
1393 
1394 	txq->added = 0;
1395 	txq->pending = 0;
1396 	txq->completed = 0;
1397 	txq->reaped = 0;
1398 
1399 	/* Destroy the common code transmit queue. */
1400 	efx_tx_qdestroy(txq->common);
1401 	txq->common = NULL;
1402 
1403 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1404 	    EFX_TXQ_NBUFS(sc->txq_entries));
1405 
1406 	SFXGE_EVQ_UNLOCK(evq);
1407 	SFXGE_TXQ_UNLOCK(txq);
1408 }
1409 
1410 static int
1411 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1412 {
1413 	struct sfxge_txq *txq;
1414 	efsys_mem_t *esmp;
1415 	uint16_t flags;
1416 	struct sfxge_evq *evq;
1417 	unsigned int desc_index;
1418 	int rc;
1419 
1420 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1421 
1422 	txq = sc->txq[index];
1423 	esmp = &txq->mem;
1424 	evq = sc->evq[txq->evq_index];
1425 
1426 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1427 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1428 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1429 	    ("evq->init_state != SFXGE_EVQ_STARTED"));
1430 
1431 	/* Program the buffer table. */
1432 	if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1433 	    EFX_TXQ_NBUFS(sc->txq_entries))) != 0)
1434 		return (rc);
1435 
1436 	/* Determine the kind of queue we are creating. */
1437 	switch (txq->type) {
1438 	case SFXGE_TXQ_NON_CKSUM:
1439 		flags = 0;
1440 		break;
1441 	case SFXGE_TXQ_IP_CKSUM:
1442 		flags = EFX_CKSUM_IPV4;
1443 		break;
1444 	case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1445 		flags = EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP;
1446 		break;
1447 	default:
1448 		KASSERT(0, ("Impossible TX queue"));
1449 		flags = 0;
1450 		break;
1451 	}
1452 
1453 	/* Create the common code transmit queue. */
1454 	if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1455 	    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1456 	    &txq->common, &desc_index)) != 0)
1457 		goto fail;
1458 
1459 	/* Initialise queue descriptor indexes */
1460 	txq->added = txq->pending = txq->completed = txq->reaped = desc_index;
1461 
1462 	SFXGE_TXQ_LOCK(txq);
1463 
1464 	/* Enable the transmit queue. */
1465 	efx_tx_qenable(txq->common);
1466 
1467 	txq->init_state = SFXGE_TXQ_STARTED;
1468 	txq->flush_state = SFXGE_FLUSH_REQUIRED;
1469 
1470 	SFXGE_TXQ_UNLOCK(txq);
1471 
1472 	return (0);
1473 
1474 fail:
1475 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1476 	    EFX_TXQ_NBUFS(sc->txq_entries));
1477 	return (rc);
1478 }
1479 
1480 void
1481 sfxge_tx_stop(struct sfxge_softc *sc)
1482 {
1483 	int index;
1484 
1485 	index = sc->txq_count;
1486 	while (--index >= 0)
1487 		sfxge_tx_qstop(sc, index);
1488 
1489 	/* Tear down the transmit module */
1490 	efx_tx_fini(sc->enp);
1491 }
1492 
1493 int
1494 sfxge_tx_start(struct sfxge_softc *sc)
1495 {
1496 	int index;
1497 	int rc;
1498 
1499 	/* Initialize the common code transmit module. */
1500 	if ((rc = efx_tx_init(sc->enp)) != 0)
1501 		return (rc);
1502 
1503 	for (index = 0; index < sc->txq_count; index++) {
1504 		if ((rc = sfxge_tx_qstart(sc, index)) != 0)
1505 			goto fail;
1506 	}
1507 
1508 	return (0);
1509 
1510 fail:
1511 	while (--index >= 0)
1512 		sfxge_tx_qstop(sc, index);
1513 
1514 	efx_tx_fini(sc->enp);
1515 
1516 	return (rc);
1517 }
1518 
1519 static int
1520 sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node)
1521 {
1522 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev);
1523 	struct sysctl_oid *stat_node;
1524 	unsigned int id;
1525 
1526 	stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1527 				    "stats", CTLFLAG_RD, NULL,
1528 				    "Tx queue statistics");
1529 	if (stat_node == NULL)
1530 		return (ENOMEM);
1531 
1532 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1533 		SYSCTL_ADD_ULONG(
1534 		    ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO,
1535 		    sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS,
1536 		    (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset),
1537 		    "");
1538 	}
1539 
1540 	return (0);
1541 }
1542 
1543 /**
1544  * Destroy a transmit queue.
1545  */
1546 static void
1547 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1548 {
1549 	struct sfxge_txq *txq;
1550 	unsigned int nmaps;
1551 
1552 	txq = sc->txq[index];
1553 
1554 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1555 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1556 
1557 	if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1558 		tso_fini(txq);
1559 
1560 	/* Free the context arrays. */
1561 	free(txq->pend_desc, M_SFXGE);
1562 	nmaps = sc->txq_entries;
1563 	while (nmaps-- != 0)
1564 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1565 	free(txq->stmp, M_SFXGE);
1566 
1567 	/* Release DMA memory mapping. */
1568 	sfxge_dma_free(&txq->mem);
1569 
1570 	sc->txq[index] = NULL;
1571 
1572 	SFXGE_TXQ_LOCK_DESTROY(txq);
1573 
1574 	free(txq, M_SFXGE);
1575 }
1576 
1577 /*
1578  * Estimate maximum number of Tx descriptors required for TSO packet.
1579  * With minimum MSS and maximum mbuf length we might need more (even
1580  * than a ring-ful of descriptors), but this should not happen in
1581  * practice except due to deliberate attack.  In that case we will
1582  * truncate the output at a packet boundary.
1583  */
1584 static unsigned int
1585 sfxge_tx_max_pkt_desc(const struct sfxge_softc *sc, enum sfxge_txq_type type)
1586 {
1587 	/* One descriptor for every input fragment */
1588 	unsigned int max_descs = SFXGE_TX_MAPPING_MAX_SEG;
1589 
1590 	/* VLAN tagging Tx option descriptor may be required */
1591 	if (efx_nic_cfg_get(sc->enp)->enc_hw_tx_insert_vlan_enabled)
1592 		max_descs++;
1593 
1594 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) {
1595 		/*
1596 		 * Plus header and payload descriptor for each output segment.
1597 		 * Minus one since header fragment is already counted.
1598 		 */
1599 		max_descs += SFXGE_TSO_MAX_SEGS * 2 - 1;
1600 
1601 		/* FW assisted TSO requires one more descriptor per segment */
1602 		if (sc->tso_fw_assisted)
1603 			max_descs += SFXGE_TSO_MAX_SEGS;
1604 	}
1605 
1606 	return (max_descs);
1607 }
1608 
1609 static int
1610 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1611 	       enum sfxge_txq_type type, unsigned int evq_index)
1612 {
1613 	char name[16];
1614 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1615 	struct sysctl_oid *txq_node;
1616 	struct sfxge_txq *txq;
1617 	struct sfxge_evq *evq;
1618 	struct sfxge_tx_dpl *stdp;
1619 	struct sysctl_oid *dpl_node;
1620 	efsys_mem_t *esmp;
1621 	unsigned int nmaps;
1622 	int rc;
1623 
1624 	txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1625 	txq->sc = sc;
1626 	txq->entries = sc->txq_entries;
1627 	txq->ptr_mask = txq->entries - 1;
1628 
1629 	sc->txq[txq_index] = txq;
1630 	esmp = &txq->mem;
1631 
1632 	evq = sc->evq[evq_index];
1633 
1634 	/* Allocate and zero DMA space for the descriptor ring. */
1635 	if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0)
1636 		return (rc);
1637 
1638 	/* Allocate buffer table entries. */
1639 	sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries),
1640 				 &txq->buf_base_id);
1641 
1642 	/* Create a DMA tag for packet mappings. */
1643 	if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000,
1644 	    MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1645 	    NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL,
1646 	    &txq->packet_dma_tag) != 0) {
1647 		device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1648 		rc = ENOMEM;
1649 		goto fail;
1650 	}
1651 
1652 	/* Allocate pending descriptor array for batching writes. */
1653 	txq->pend_desc = malloc(sizeof(efx_desc_t) * sc->txq_entries,
1654 				M_SFXGE, M_ZERO | M_WAITOK);
1655 
1656 	/* Allocate and initialise mbuf DMA mapping array. */
1657 	txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries,
1658 	    M_SFXGE, M_ZERO | M_WAITOK);
1659 	for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) {
1660 		rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1661 				       &txq->stmp[nmaps].map);
1662 		if (rc != 0)
1663 			goto fail2;
1664 	}
1665 
1666 	snprintf(name, sizeof(name), "%u", txq_index);
1667 	txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node),
1668 				   OID_AUTO, name, CTLFLAG_RD, NULL, "");
1669 	if (txq_node == NULL) {
1670 		rc = ENOMEM;
1671 		goto fail_txq_node;
1672 	}
1673 
1674 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1675 	    (rc = tso_init(txq)) != 0)
1676 		goto fail3;
1677 
1678 	if (sfxge_tx_dpl_get_max <= 0) {
1679 		log(LOG_ERR, "%s=%d must be greater than 0",
1680 		    SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max);
1681 		rc = EINVAL;
1682 		goto fail_tx_dpl_get_max;
1683 	}
1684 	if (sfxge_tx_dpl_get_non_tcp_max <= 0) {
1685 		log(LOG_ERR, "%s=%d must be greater than 0",
1686 		    SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX,
1687 		    sfxge_tx_dpl_get_non_tcp_max);
1688 		rc = EINVAL;
1689 		goto fail_tx_dpl_get_max;
1690 	}
1691 	if (sfxge_tx_dpl_put_max < 0) {
1692 		log(LOG_ERR, "%s=%d must be greater or equal to 0",
1693 		    SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max);
1694 		rc = EINVAL;
1695 		goto fail_tx_dpl_put_max;
1696 	}
1697 
1698 	/* Initialize the deferred packet list. */
1699 	stdp = &txq->dpl;
1700 	stdp->std_put_max = sfxge_tx_dpl_put_max;
1701 	stdp->std_get_max = sfxge_tx_dpl_get_max;
1702 	stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max;
1703 	stdp->std_getp = &stdp->std_get;
1704 
1705 	SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index);
1706 
1707 	dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1708 				   "dpl", CTLFLAG_RD, NULL,
1709 				   "Deferred packet list statistics");
1710 	if (dpl_node == NULL) {
1711 		rc = ENOMEM;
1712 		goto fail_dpl_node;
1713 	}
1714 
1715 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1716 			"get_count", CTLFLAG_RD | CTLFLAG_STATS,
1717 			&stdp->std_get_count, 0, "");
1718 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1719 			"get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS,
1720 			&stdp->std_get_non_tcp_count, 0, "");
1721 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1722 			"get_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1723 			&stdp->std_get_hiwat, 0, "");
1724 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1725 			"put_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1726 			&stdp->std_put_hiwat, 0, "");
1727 
1728 	rc = sfxge_txq_stat_init(txq, txq_node);
1729 	if (rc != 0)
1730 		goto fail_txq_stat_init;
1731 
1732 	txq->type = type;
1733 	txq->evq_index = evq_index;
1734 	txq->txq_index = txq_index;
1735 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1736 	txq->hw_vlan_tci = 0;
1737 
1738 	txq->max_pkt_desc = sfxge_tx_max_pkt_desc(sc, type);
1739 
1740 	return (0);
1741 
1742 fail_txq_stat_init:
1743 fail_dpl_node:
1744 fail_tx_dpl_put_max:
1745 fail_tx_dpl_get_max:
1746 fail3:
1747 fail_txq_node:
1748 	free(txq->pend_desc, M_SFXGE);
1749 fail2:
1750 	while (nmaps-- != 0)
1751 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1752 	free(txq->stmp, M_SFXGE);
1753 	bus_dma_tag_destroy(txq->packet_dma_tag);
1754 
1755 fail:
1756 	sfxge_dma_free(esmp);
1757 
1758 	return (rc);
1759 }
1760 
1761 static int
1762 sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1763 {
1764 	struct sfxge_softc *sc = arg1;
1765 	unsigned int id = arg2;
1766 	unsigned long sum;
1767 	unsigned int index;
1768 
1769 	/* Sum across all TX queues */
1770 	sum = 0;
1771 	for (index = 0; index < sc->txq_count; index++)
1772 		sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1773 					  sfxge_tx_stats[id].offset);
1774 
1775 	return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1776 }
1777 
1778 static void
1779 sfxge_tx_stat_init(struct sfxge_softc *sc)
1780 {
1781 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1782 	struct sysctl_oid_list *stat_list;
1783 	unsigned int id;
1784 
1785 	stat_list = SYSCTL_CHILDREN(sc->stats_node);
1786 
1787 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1788 		SYSCTL_ADD_PROC(
1789 			ctx, stat_list,
1790 			OID_AUTO, sfxge_tx_stats[id].name,
1791 			CTLTYPE_ULONG|CTLFLAG_RD,
1792 			sc, id, sfxge_tx_stat_handler, "LU",
1793 			"");
1794 	}
1795 }
1796 
1797 uint64_t
1798 sfxge_tx_get_drops(struct sfxge_softc *sc)
1799 {
1800 	unsigned int index;
1801 	uint64_t drops = 0;
1802 	struct sfxge_txq *txq;
1803 
1804 	/* Sum across all TX queues */
1805 	for (index = 0; index < sc->txq_count; index++) {
1806 		txq = sc->txq[index];
1807 		/*
1808 		 * In theory, txq->put_overflow and txq->netdown_drops
1809 		 * should use atomic operation and other should be
1810 		 * obtained under txq lock, but it is just statistics.
1811 		 */
1812 		drops += txq->drops + txq->get_overflow +
1813 			 txq->get_non_tcp_overflow +
1814 			 txq->put_overflow + txq->netdown_drops +
1815 			 txq->tso_pdrop_too_many + txq->tso_pdrop_no_rsrc;
1816 	}
1817 	return (drops);
1818 }
1819 
1820 void
1821 sfxge_tx_fini(struct sfxge_softc *sc)
1822 {
1823 	int index;
1824 
1825 	index = sc->txq_count;
1826 	while (--index >= 0)
1827 		sfxge_tx_qfini(sc, index);
1828 
1829 	sc->txq_count = 0;
1830 }
1831 
1832 
1833 int
1834 sfxge_tx_init(struct sfxge_softc *sc)
1835 {
1836 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
1837 	struct sfxge_intr *intr;
1838 	int index;
1839 	int rc;
1840 
1841 	intr = &sc->intr;
1842 
1843 	KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1844 	    ("intr->state != SFXGE_INTR_INITIALIZED"));
1845 
1846 	sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc;
1847 
1848 	sc->tso_fw_assisted = sfxge_tso_fw_assisted;
1849 	if (sc->tso_fw_assisted)
1850 		sc->tso_fw_assisted =
1851 		    (encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO) &&
1852 		    (encp->enc_fw_assisted_tso_enabled);
1853 
1854 	sc->txqs_node = SYSCTL_ADD_NODE(
1855 		device_get_sysctl_ctx(sc->dev),
1856 		SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
1857 		OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues");
1858 	if (sc->txqs_node == NULL) {
1859 		rc = ENOMEM;
1860 		goto fail_txq_node;
1861 	}
1862 
1863 	/* Initialize the transmit queues */
1864 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
1865 	    SFXGE_TXQ_NON_CKSUM, 0)) != 0)
1866 		goto fail;
1867 
1868 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
1869 	    SFXGE_TXQ_IP_CKSUM, 0)) != 0)
1870 		goto fail2;
1871 
1872 	for (index = 0;
1873 	     index < sc->txq_count - SFXGE_TXQ_NTYPES + 1;
1874 	     index++) {
1875 		if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index,
1876 		    SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
1877 			goto fail3;
1878 	}
1879 
1880 	sfxge_tx_stat_init(sc);
1881 
1882 	return (0);
1883 
1884 fail3:
1885 	while (--index >= 0)
1886 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1887 
1888 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1889 
1890 fail2:
1891 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1892 
1893 fail:
1894 fail_txq_node:
1895 	sc->txq_count = 0;
1896 	return (rc);
1897 }
1898