xref: /freebsd/sys/dev/sfxge/sfxge_tx.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
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/param.h>
53 #include <sys/malloc.h>
54 #include <sys/mbuf.h>
55 #include <sys/smp.h>
56 #include <sys/socket.h>
57 #include <sys/sysctl.h>
58 #include <sys/syslog.h>
59 #include <sys/limits.h>
60 
61 #include <net/bpf.h>
62 #include <net/ethernet.h>
63 #include <net/if.h>
64 #include <net/if_vlan_var.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/ip.h>
68 #include <netinet/ip6.h>
69 #include <netinet/tcp.h>
70 
71 #include "common/efx.h"
72 
73 #include "sfxge.h"
74 #include "sfxge_tx.h"
75 
76 
77 #define	SFXGE_PARAM_TX_DPL_GET_MAX	SFXGE_PARAM(tx_dpl_get_max)
78 static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT;
79 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max);
80 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN,
81 	   &sfxge_tx_dpl_get_max, 0,
82 	   "Maximum number of any packets in deferred packet get-list");
83 
84 #define	SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \
85 	SFXGE_PARAM(tx_dpl_get_non_tcp_max)
86 static int sfxge_tx_dpl_get_non_tcp_max =
87 	SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT;
88 TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max);
89 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN,
90 	   &sfxge_tx_dpl_get_non_tcp_max, 0,
91 	   "Maximum number of non-TCP packets in deferred packet get-list");
92 
93 #define	SFXGE_PARAM_TX_DPL_PUT_MAX	SFXGE_PARAM(tx_dpl_put_max)
94 static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT;
95 TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max);
96 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN,
97 	   &sfxge_tx_dpl_put_max, 0,
98 	   "Maximum number of any packets in deferred packet put-list");
99 
100 #define	SFXGE_PARAM_TSO_FW_ASSISTED	SFXGE_PARAM(tso_fw_assisted)
101 static int sfxge_tso_fw_assisted = (SFXGE_FATSOV1 | SFXGE_FATSOV2);
102 TUNABLE_INT(SFXGE_PARAM_TSO_FW_ASSISTED, &sfxge_tso_fw_assisted);
103 SYSCTL_INT(_hw_sfxge, OID_AUTO, tso_fw_assisted, CTLFLAG_RDTUN,
104 	   &sfxge_tso_fw_assisted, 0,
105 	   "Bitmask of FW-assisted TSO allowed to use if supported by NIC firmware");
106 
107 
108 static const struct {
109 	const char *name;
110 	size_t offset;
111 } sfxge_tx_stats[] = {
112 #define	SFXGE_TX_STAT(name, member) \
113 	{ #name, offsetof(struct sfxge_txq, member) }
114 	SFXGE_TX_STAT(tso_bursts, tso_bursts),
115 	SFXGE_TX_STAT(tso_packets, tso_packets),
116 	SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
117 	SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many),
118 	SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc),
119 	SFXGE_TX_STAT(tx_collapses, collapses),
120 	SFXGE_TX_STAT(tx_drops, drops),
121 	SFXGE_TX_STAT(tx_get_overflow, get_overflow),
122 	SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow),
123 	SFXGE_TX_STAT(tx_put_overflow, put_overflow),
124 	SFXGE_TX_STAT(tx_netdown_drops, netdown_drops),
125 };
126 
127 
128 /* Forward declarations. */
129 static void sfxge_tx_qdpl_service(struct sfxge_txq *txq);
130 static void sfxge_tx_qlist_post(struct sfxge_txq *txq);
131 static void sfxge_tx_qunblock(struct sfxge_txq *txq);
132 static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
133 			      const bus_dma_segment_t *dma_seg, int n_dma_seg,
134 			      int vlan_tagged);
135 
136 static int
137 sfxge_tx_maybe_insert_tag(struct sfxge_txq *txq, struct mbuf *mbuf)
138 {
139 	uint16_t this_tag = ((mbuf->m_flags & M_VLANTAG) ?
140 			     mbuf->m_pkthdr.ether_vtag :
141 			     0);
142 
143 	if (this_tag == txq->hw_vlan_tci)
144 		return (0);
145 
146 	efx_tx_qdesc_vlantci_create(txq->common,
147 				    bswap16(this_tag),
148 				    &txq->pend_desc[0]);
149 	txq->n_pend_desc = 1;
150 	txq->hw_vlan_tci = this_tag;
151 	return (1);
152 }
153 
154 static inline void
155 sfxge_next_stmp(struct sfxge_txq *txq, struct sfxge_tx_mapping **pstmp)
156 {
157 	KASSERT((*pstmp)->flags == 0, ("stmp flags are not 0"));
158 	if (__predict_false(*pstmp ==
159 			    &txq->stmp[txq->ptr_mask]))
160 		*pstmp = &txq->stmp[0];
161 	else
162 		(*pstmp)++;
163 }
164 
165 
166 void
167 sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq)
168 {
169 	unsigned int completed;
170 
171 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
172 
173 	completed = txq->completed;
174 	while (completed != txq->pending) {
175 		struct sfxge_tx_mapping *stmp;
176 		unsigned int id;
177 
178 		id = completed++ & txq->ptr_mask;
179 
180 		stmp = &txq->stmp[id];
181 		if (stmp->flags & TX_BUF_UNMAP) {
182 			bus_dmamap_unload(txq->packet_dma_tag, stmp->map);
183 			if (stmp->flags & TX_BUF_MBUF) {
184 				struct mbuf *m = stmp->u.mbuf;
185 				do
186 					m = m_free(m);
187 				while (m != NULL);
188 			} else {
189 				free(stmp->u.heap_buf, M_SFXGE);
190 			}
191 			stmp->flags = 0;
192 		}
193 	}
194 	txq->completed = completed;
195 
196 	/* Check whether we need to unblock the queue. */
197 	mb();
198 	if (txq->blocked) {
199 		unsigned int level;
200 
201 		level = txq->added - txq->completed;
202 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
203 			sfxge_tx_qunblock(txq);
204 	}
205 }
206 
207 static unsigned int
208 sfxge_is_mbuf_non_tcp(struct mbuf *mbuf)
209 {
210 	/* Absense of TCP checksum flags does not mean that it is non-TCP
211 	 * but it should be true if user wants to achieve high throughput.
212 	 */
213 	return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP)));
214 }
215 
216 /*
217  * Reorder the put list and append it to the get list.
218  */
219 static void
220 sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq)
221 {
222 	struct sfxge_tx_dpl *stdp;
223 	struct mbuf *mbuf, *get_next, **get_tailp;
224 	volatile uintptr_t *putp;
225 	uintptr_t put;
226 	unsigned int count;
227 	unsigned int non_tcp_count;
228 
229 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
230 
231 	stdp = &txq->dpl;
232 
233 	/* Acquire the put list. */
234 	putp = &stdp->std_put;
235 	put = atomic_readandclear_ptr(putp);
236 	mbuf = (void *)put;
237 
238 	if (mbuf == NULL)
239 		return;
240 
241 	/* Reverse the put list. */
242 	get_tailp = &mbuf->m_nextpkt;
243 	get_next = NULL;
244 
245 	count = 0;
246 	non_tcp_count = 0;
247 	do {
248 		struct mbuf *put_next;
249 
250 		non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf);
251 		put_next = mbuf->m_nextpkt;
252 		mbuf->m_nextpkt = get_next;
253 		get_next = mbuf;
254 		mbuf = put_next;
255 
256 		count++;
257 	} while (mbuf != NULL);
258 
259 	if (count > stdp->std_put_hiwat)
260 		stdp->std_put_hiwat = count;
261 
262 	/* Append the reversed put list to the get list. */
263 	KASSERT(*get_tailp == NULL, ("*get_tailp != NULL"));
264 	*stdp->std_getp = get_next;
265 	stdp->std_getp = get_tailp;
266 	stdp->std_get_count += count;
267 	stdp->std_get_non_tcp_count += non_tcp_count;
268 }
269 
270 static void
271 sfxge_tx_qreap(struct sfxge_txq *txq)
272 {
273 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
274 
275 	txq->reaped = txq->completed;
276 }
277 
278 static void
279 sfxge_tx_qlist_post(struct sfxge_txq *txq)
280 {
281 	unsigned int old_added;
282 	unsigned int block_level;
283 	unsigned int level;
284 	int rc;
285 
286 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
287 
288 	KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0"));
289 	KASSERT(txq->n_pend_desc <= txq->max_pkt_desc,
290 		("txq->n_pend_desc too large"));
291 	KASSERT(!txq->blocked, ("txq->blocked"));
292 
293 	old_added = txq->added;
294 
295 	/* Post the fragment list. */
296 	rc = efx_tx_qdesc_post(txq->common, txq->pend_desc, txq->n_pend_desc,
297 			  txq->reaped, &txq->added);
298 	KASSERT(rc == 0, ("efx_tx_qdesc_post() failed"));
299 
300 	/* If efx_tx_qdesc_post() had to refragment, our information about
301 	 * buffers to free may be associated with the wrong
302 	 * descriptors.
303 	 */
304 	KASSERT(txq->added - old_added == txq->n_pend_desc,
305 		("efx_tx_qdesc_post() refragmented descriptors"));
306 
307 	level = txq->added - txq->reaped;
308 	KASSERT(level <= txq->entries, ("overfilled TX queue"));
309 
310 	/* Clear the fragment list. */
311 	txq->n_pend_desc = 0;
312 
313 	/*
314 	 * Set the block level to ensure there is space to generate a
315 	 * large number of descriptors for TSO.
316 	 */
317 	block_level = EFX_TXQ_LIMIT(txq->entries) - txq->max_pkt_desc;
318 
319 	/* Have we reached the block level? */
320 	if (level < block_level)
321 		return;
322 
323 	/* Reap, and check again */
324 	sfxge_tx_qreap(txq);
325 	level = txq->added - txq->reaped;
326 	if (level < block_level)
327 		return;
328 
329 	txq->blocked = 1;
330 
331 	/*
332 	 * Avoid a race with completion interrupt handling that could leave
333 	 * the queue blocked.
334 	 */
335 	mb();
336 	sfxge_tx_qreap(txq);
337 	level = txq->added - txq->reaped;
338 	if (level < block_level) {
339 		mb();
340 		txq->blocked = 0;
341 	}
342 }
343 
344 static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
345 {
346 	bus_dmamap_t *used_map;
347 	bus_dmamap_t map;
348 	bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG];
349 	unsigned int id;
350 	struct sfxge_tx_mapping *stmp;
351 	efx_desc_t *desc;
352 	int n_dma_seg;
353 	int rc;
354 	int i;
355 	int eop;
356 	int vlan_tagged;
357 
358 	KASSERT(!txq->blocked, ("txq->blocked"));
359 
360 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
361 		prefetch_read_many(mbuf->m_data);
362 
363 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) {
364 		rc = EINTR;
365 		goto reject;
366 	}
367 
368 	/* Load the packet for DMA. */
369 	id = txq->added & txq->ptr_mask;
370 	stmp = &txq->stmp[id];
371 	rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map,
372 				     mbuf, dma_seg, &n_dma_seg, 0);
373 	if (rc == EFBIG) {
374 		/* Try again. */
375 		struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT,
376 						   SFXGE_TX_MAPPING_MAX_SEG);
377 		if (new_mbuf == NULL)
378 			goto reject;
379 		++txq->collapses;
380 		mbuf = new_mbuf;
381 		rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag,
382 					     stmp->map, mbuf,
383 					     dma_seg, &n_dma_seg, 0);
384 	}
385 	if (rc != 0)
386 		goto reject;
387 
388 	/* Make the packet visible to the hardware. */
389 	bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE);
390 
391 	used_map = &stmp->map;
392 
393 	vlan_tagged = sfxge_tx_maybe_insert_tag(txq, mbuf);
394 	if (vlan_tagged) {
395 		sfxge_next_stmp(txq, &stmp);
396 	}
397 	if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
398 		rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg, vlan_tagged);
399 		if (rc < 0)
400 			goto reject_mapped;
401 		stmp = &txq->stmp[(rc - 1) & txq->ptr_mask];
402 	} else {
403 		/* Add the mapping to the fragment list, and set flags
404 		 * for the buffer.
405 		 */
406 
407 		i = 0;
408 		for (;;) {
409 			desc = &txq->pend_desc[i + vlan_tagged];
410 			eop = (i == n_dma_seg - 1);
411 			efx_tx_qdesc_dma_create(txq->common,
412 						dma_seg[i].ds_addr,
413 						dma_seg[i].ds_len,
414 						eop,
415 						desc);
416 			if (eop)
417 				break;
418 			i++;
419 			sfxge_next_stmp(txq, &stmp);
420 		}
421 		txq->n_pend_desc = n_dma_seg + vlan_tagged;
422 	}
423 
424 	/*
425 	 * If the mapping required more than one descriptor
426 	 * then we need to associate the DMA map with the last
427 	 * descriptor, not the first.
428 	 */
429 	if (used_map != &stmp->map) {
430 		map = stmp->map;
431 		stmp->map = *used_map;
432 		*used_map = map;
433 	}
434 
435 	stmp->u.mbuf = mbuf;
436 	stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF;
437 
438 	/* Post the fragment list. */
439 	sfxge_tx_qlist_post(txq);
440 
441 	return (0);
442 
443 reject_mapped:
444 	bus_dmamap_unload(txq->packet_dma_tag, *used_map);
445 reject:
446 	/* Drop the packet on the floor. */
447 	m_freem(mbuf);
448 	++txq->drops;
449 
450 	return (rc);
451 }
452 
453 /*
454  * Drain the deferred packet list into the transmit queue.
455  */
456 static void
457 sfxge_tx_qdpl_drain(struct sfxge_txq *txq)
458 {
459 	struct sfxge_softc *sc;
460 	struct sfxge_tx_dpl *stdp;
461 	struct mbuf *mbuf, *next;
462 	unsigned int count;
463 	unsigned int non_tcp_count;
464 	unsigned int pushed;
465 	int rc;
466 
467 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
468 
469 	sc = txq->sc;
470 	stdp = &txq->dpl;
471 	pushed = txq->added;
472 
473 	if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) {
474 		prefetch_read_many(sc->enp);
475 		prefetch_read_many(txq->common);
476 	}
477 
478 	mbuf = stdp->std_get;
479 	count = stdp->std_get_count;
480 	non_tcp_count = stdp->std_get_non_tcp_count;
481 
482 	if (count > stdp->std_get_hiwat)
483 		stdp->std_get_hiwat = count;
484 
485 	while (count != 0) {
486 		KASSERT(mbuf != NULL, ("mbuf == NULL"));
487 
488 		next = mbuf->m_nextpkt;
489 		mbuf->m_nextpkt = NULL;
490 
491 		ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */
492 
493 		if (next != NULL)
494 			prefetch_read_many(next);
495 
496 		rc = sfxge_tx_queue_mbuf(txq, mbuf);
497 		--count;
498 		non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf);
499 		mbuf = next;
500 		if (rc != 0)
501 			continue;
502 
503 		if (txq->blocked)
504 			break;
505 
506 		/* Push the fragments to the hardware in batches. */
507 		if (txq->added - pushed >= SFXGE_TX_BATCH) {
508 			efx_tx_qpush(txq->common, txq->added, pushed);
509 			pushed = txq->added;
510 		}
511 	}
512 
513 	if (count == 0) {
514 		KASSERT(mbuf == NULL, ("mbuf != NULL"));
515 		KASSERT(non_tcp_count == 0,
516 			("inconsistent TCP/non-TCP detection"));
517 		stdp->std_get = NULL;
518 		stdp->std_get_count = 0;
519 		stdp->std_get_non_tcp_count = 0;
520 		stdp->std_getp = &stdp->std_get;
521 	} else {
522 		stdp->std_get = mbuf;
523 		stdp->std_get_count = count;
524 		stdp->std_get_non_tcp_count = non_tcp_count;
525 	}
526 
527 	if (txq->added != pushed)
528 		efx_tx_qpush(txq->common, txq->added, pushed);
529 
530 	KASSERT(txq->blocked || stdp->std_get_count == 0,
531 		("queue unblocked but count is non-zero"));
532 }
533 
534 #define	SFXGE_TX_QDPL_PENDING(_txq)	((_txq)->dpl.std_put != 0)
535 
536 /*
537  * Service the deferred packet list.
538  *
539  * NOTE: drops the txq mutex!
540  */
541 static void
542 sfxge_tx_qdpl_service(struct sfxge_txq *txq)
543 {
544 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
545 
546 	do {
547 		if (SFXGE_TX_QDPL_PENDING(txq))
548 			sfxge_tx_qdpl_swizzle(txq);
549 
550 		if (!txq->blocked)
551 			sfxge_tx_qdpl_drain(txq);
552 
553 		SFXGE_TXQ_UNLOCK(txq);
554 	} while (SFXGE_TX_QDPL_PENDING(txq) &&
555 		 SFXGE_TXQ_TRYLOCK(txq));
556 }
557 
558 /*
559  * Put a packet on the deferred packet get-list.
560  */
561 static int
562 sfxge_tx_qdpl_put_locked(struct sfxge_txq *txq, struct mbuf *mbuf)
563 {
564 	struct sfxge_tx_dpl *stdp;
565 
566 	stdp = &txq->dpl;
567 
568 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
569 
570 	SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
571 
572 	if (stdp->std_get_count >= stdp->std_get_max) {
573 		txq->get_overflow++;
574 		return (ENOBUFS);
575 	}
576 	if (sfxge_is_mbuf_non_tcp(mbuf)) {
577 		if (stdp->std_get_non_tcp_count >=
578 		    stdp->std_get_non_tcp_max) {
579 			txq->get_non_tcp_overflow++;
580 			return (ENOBUFS);
581 		}
582 		stdp->std_get_non_tcp_count++;
583 	}
584 
585 	*(stdp->std_getp) = mbuf;
586 	stdp->std_getp = &mbuf->m_nextpkt;
587 	stdp->std_get_count++;
588 
589 	return (0);
590 }
591 
592 /*
593  * Put a packet on the deferred packet put-list.
594  *
595  * We overload the csum_data field in the mbuf to keep track of this length
596  * because there is no cheap alternative to avoid races.
597  */
598 static int
599 sfxge_tx_qdpl_put_unlocked(struct sfxge_txq *txq, struct mbuf *mbuf)
600 {
601 	struct sfxge_tx_dpl *stdp;
602 	volatile uintptr_t *putp;
603 	uintptr_t old;
604 	uintptr_t new;
605 	unsigned old_len;
606 
607 	KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL"));
608 
609 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
610 
611 	stdp = &txq->dpl;
612 	putp = &stdp->std_put;
613 	new = (uintptr_t)mbuf;
614 
615 	do {
616 		old = *putp;
617 		if (old != 0) {
618 			struct mbuf *mp = (struct mbuf *)old;
619 			old_len = mp->m_pkthdr.csum_data;
620 		} else
621 			old_len = 0;
622 		if (old_len >= stdp->std_put_max) {
623 			atomic_add_long(&txq->put_overflow, 1);
624 			return (ENOBUFS);
625 		}
626 		mbuf->m_pkthdr.csum_data = old_len + 1;
627 		mbuf->m_nextpkt = (void *)old;
628 	} while (atomic_cmpset_ptr(putp, old, new) == 0);
629 
630 	return (0);
631 }
632 
633 /*
634  * Called from if_transmit - will try to grab the txq lock and enqueue to the
635  * put list if it succeeds, otherwise try to push onto the defer list if space.
636  */
637 static int
638 sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m)
639 {
640 	int rc;
641 
642 	if (!SFXGE_LINK_UP(txq->sc)) {
643 		atomic_add_long(&txq->netdown_drops, 1);
644 		return (ENETDOWN);
645 	}
646 
647 	/*
648 	 * Try to grab the txq lock.  If we are able to get the lock,
649 	 * the packet will be appended to the "get list" of the deferred
650 	 * packet list.  Otherwise, it will be pushed on the "put list".
651 	 */
652 	if (SFXGE_TXQ_TRYLOCK(txq)) {
653 		/* First swizzle put-list to get-list to keep order */
654 		sfxge_tx_qdpl_swizzle(txq);
655 
656 		rc = sfxge_tx_qdpl_put_locked(txq, m);
657 
658 		/* Try to service the list. */
659 		sfxge_tx_qdpl_service(txq);
660 		/* Lock has been dropped. */
661 	} else {
662 		rc = sfxge_tx_qdpl_put_unlocked(txq, m);
663 
664 		/*
665 		 * Try to grab the lock again.
666 		 *
667 		 * If we are able to get the lock, we need to process
668 		 * the deferred packet list.  If we are not able to get
669 		 * the lock, another thread is processing the list.
670 		 */
671 		if ((rc == 0) && SFXGE_TXQ_TRYLOCK(txq)) {
672 			sfxge_tx_qdpl_service(txq);
673 			/* Lock has been dropped. */
674 		}
675 	}
676 
677 	SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq);
678 
679 	return (rc);
680 }
681 
682 static void
683 sfxge_tx_qdpl_flush(struct sfxge_txq *txq)
684 {
685 	struct sfxge_tx_dpl *stdp = &txq->dpl;
686 	struct mbuf *mbuf, *next;
687 
688 	SFXGE_TXQ_LOCK(txq);
689 
690 	sfxge_tx_qdpl_swizzle(txq);
691 	for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) {
692 		next = mbuf->m_nextpkt;
693 		m_freem(mbuf);
694 	}
695 	stdp->std_get = NULL;
696 	stdp->std_get_count = 0;
697 	stdp->std_get_non_tcp_count = 0;
698 	stdp->std_getp = &stdp->std_get;
699 
700 	SFXGE_TXQ_UNLOCK(txq);
701 }
702 
703 void
704 sfxge_if_qflush(struct ifnet *ifp)
705 {
706 	struct sfxge_softc *sc;
707 	unsigned int i;
708 
709 	sc = ifp->if_softc;
710 
711 	for (i = 0; i < sc->txq_count; i++)
712 		sfxge_tx_qdpl_flush(sc->txq[i]);
713 }
714 
715 #if SFXGE_TX_PARSE_EARLY
716 
717 /* There is little space for user data in mbuf pkthdr, so we
718  * use l*hlen fields which are not used by the driver otherwise
719  * to store header offsets.
720  * The fields are 8-bit, but it's ok, no header may be longer than 255 bytes.
721  */
722 
723 
724 #define TSO_MBUF_PROTO(_mbuf)    ((_mbuf)->m_pkthdr.PH_loc.sixteen[0])
725 /* We abuse l5hlen here because PH_loc can hold only 64 bits of data */
726 #define TSO_MBUF_FLAGS(_mbuf)    ((_mbuf)->m_pkthdr.l5hlen)
727 #define TSO_MBUF_PACKETID(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[1])
728 #define TSO_MBUF_SEQNUM(_mbuf)   ((_mbuf)->m_pkthdr.PH_loc.thirtytwo[1])
729 
730 static void sfxge_parse_tx_packet(struct mbuf *mbuf)
731 {
732 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
733 	const struct tcphdr *th;
734 	struct tcphdr th_copy;
735 
736 	/* Find network protocol and header */
737 	TSO_MBUF_PROTO(mbuf) = eh->ether_type;
738 	if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_VLAN)) {
739 		struct ether_vlan_header *veh =
740 			mtod(mbuf, struct ether_vlan_header *);
741 		TSO_MBUF_PROTO(mbuf) = veh->evl_proto;
742 		mbuf->m_pkthdr.l2hlen = sizeof(*veh);
743 	} else {
744 		mbuf->m_pkthdr.l2hlen = sizeof(*eh);
745 	}
746 
747 	/* Find TCP header */
748 	if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IP)) {
749 		const struct ip *iph = (const struct ip *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen);
750 
751 		KASSERT(iph->ip_p == IPPROTO_TCP,
752 			("TSO required on non-TCP packet"));
753 		mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + 4 * iph->ip_hl;
754 		TSO_MBUF_PACKETID(mbuf) = iph->ip_id;
755 	} else {
756 		KASSERT(TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IPV6),
757 			("TSO required on non-IP packet"));
758 		KASSERT(((const struct ip6_hdr *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen))->ip6_nxt ==
759 			IPPROTO_TCP,
760 			("TSO required on non-TCP packet"));
761 		mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + sizeof(struct ip6_hdr);
762 		TSO_MBUF_PACKETID(mbuf) = 0;
763 	}
764 
765 	KASSERT(mbuf->m_len >= mbuf->m_pkthdr.l3hlen,
766 		("network header is fragmented in mbuf"));
767 
768 	/* We need TCP header including flags (window is the next) */
769 	if (mbuf->m_len < mbuf->m_pkthdr.l3hlen + offsetof(struct tcphdr, th_win)) {
770 		m_copydata(mbuf, mbuf->m_pkthdr.l3hlen, sizeof(th_copy),
771 			   (caddr_t)&th_copy);
772 		th = &th_copy;
773 	} else {
774 		th = (const struct tcphdr *)mtodo(mbuf, mbuf->m_pkthdr.l3hlen);
775 	}
776 
777 	mbuf->m_pkthdr.l4hlen = mbuf->m_pkthdr.l3hlen + 4 * th->th_off;
778 	TSO_MBUF_SEQNUM(mbuf) = ntohl(th->th_seq);
779 
780 	/* These flags must not be duplicated */
781 	/*
782 	 * RST should not be duplicated as well, but FreeBSD kernel
783 	 * generates TSO packets with RST flag. So, do not assert
784 	 * its absence.
785 	 */
786 	KASSERT(!(th->th_flags & (TH_URG | TH_SYN)),
787 		("incompatible TCP flag 0x%x on TSO packet",
788 		 th->th_flags & (TH_URG | TH_SYN)));
789 	TSO_MBUF_FLAGS(mbuf) = th->th_flags;
790 }
791 #endif
792 
793 /*
794  * TX start -- called by the stack.
795  */
796 int
797 sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m)
798 {
799 	struct sfxge_softc *sc;
800 	struct sfxge_txq *txq;
801 	int rc;
802 
803 	sc = (struct sfxge_softc *)ifp->if_softc;
804 
805 	/*
806 	 * Transmit may be called when interface is up from the kernel
807 	 * point of view, but not yet up (in progress) from the driver
808 	 * point of view. I.e. link aggregation bring up.
809 	 * Transmit may be called when interface is up from the driver
810 	 * point of view, but already down from the kernel point of
811 	 * view. I.e. Rx when interface shutdown is in progress.
812 	 */
813 	KASSERT((ifp->if_flags & IFF_UP) || (sc->if_flags & IFF_UP),
814 		("interface not up"));
815 
816 	/* Pick the desired transmit queue. */
817 	if (m->m_pkthdr.csum_flags &
818 	    (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO)) {
819 		int index = 0;
820 
821 		/* check if flowid is set */
822 		if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
823 			uint32_t hash = m->m_pkthdr.flowid;
824 
825 			index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX];
826 		}
827 #if SFXGE_TX_PARSE_EARLY
828 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
829 			sfxge_parse_tx_packet(m);
830 #endif
831 		txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index];
832 	} else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
833 		txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
834 	} else {
835 		txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
836 	}
837 
838 	rc = sfxge_tx_packet_add(txq, m);
839 	if (rc != 0)
840 		m_freem(m);
841 
842 	return (rc);
843 }
844 
845 /*
846  * Software "TSO".  Not quite as good as doing it in hardware, but
847  * still faster than segmenting in the stack.
848  */
849 
850 struct sfxge_tso_state {
851 	/* Output position */
852 	unsigned out_len;	/* Remaining length in current segment */
853 	unsigned seqnum;	/* Current sequence number */
854 	unsigned packet_space;	/* Remaining space in current packet */
855 	unsigned segs_space;	/* Remaining number of DMA segments
856 				   for the packet (FATSOv2 only) */
857 
858 	/* Input position */
859 	uint64_t dma_addr;	/* DMA address of current position */
860 	unsigned in_len;	/* Remaining length in current mbuf */
861 
862 	const struct mbuf *mbuf; /* Input mbuf (head of chain) */
863 	u_short protocol;	/* Network protocol (after VLAN decap) */
864 	ssize_t nh_off;		/* Offset of network header */
865 	ssize_t tcph_off;	/* Offset of TCP header */
866 	unsigned header_len;	/* Number of bytes of header */
867 	unsigned seg_size;	/* TCP segment size */
868 	int fw_assisted;	/* Use FW-assisted TSO */
869 	u_short packet_id;	/* IPv4 packet ID from the original packet */
870 	uint8_t tcp_flags;	/* TCP flags */
871 	efx_desc_t header_desc; /* Precomputed header descriptor for
872 				 * FW-assisted TSO */
873 };
874 
875 #if !SFXGE_TX_PARSE_EARLY
876 static const struct ip *tso_iph(const struct sfxge_tso_state *tso)
877 {
878 	KASSERT(tso->protocol == htons(ETHERTYPE_IP),
879 		("tso_iph() in non-IPv4 state"));
880 	return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
881 }
882 
883 static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
884 {
885 	KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
886 		("tso_ip6h() in non-IPv6 state"));
887 	return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
888 }
889 
890 static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
891 {
892 	return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
893 }
894 #endif
895 
896 
897 /* Size of preallocated TSO header buffers.  Larger blocks must be
898  * allocated from the heap.
899  */
900 #define	TSOH_STD_SIZE	128
901 
902 /* At most half the descriptors in the queue at any time will refer to
903  * a TSO header buffer, since they must always be followed by a
904  * payload descriptor referring to an mbuf.
905  */
906 #define	TSOH_COUNT(_txq_entries)	((_txq_entries) / 2u)
907 #define	TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
908 #define	TSOH_PAGE_COUNT(_txq_entries)	\
909 	((TSOH_COUNT(_txq_entries) + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE)
910 
911 static int tso_init(struct sfxge_txq *txq)
912 {
913 	struct sfxge_softc *sc = txq->sc;
914 	unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries);
915 	int i, rc;
916 
917 	/* Allocate TSO header buffers */
918 	txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]),
919 				  M_SFXGE, M_WAITOK);
920 
921 	for (i = 0; i < tsoh_page_count; i++) {
922 		rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
923 		if (rc != 0)
924 			goto fail;
925 	}
926 
927 	return (0);
928 
929 fail:
930 	while (i-- > 0)
931 		sfxge_dma_free(&txq->tsoh_buffer[i]);
932 	free(txq->tsoh_buffer, M_SFXGE);
933 	txq->tsoh_buffer = NULL;
934 	return (rc);
935 }
936 
937 static void tso_fini(struct sfxge_txq *txq)
938 {
939 	int i;
940 
941 	if (txq->tsoh_buffer != NULL) {
942 		for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++)
943 			sfxge_dma_free(&txq->tsoh_buffer[i]);
944 		free(txq->tsoh_buffer, M_SFXGE);
945 	}
946 }
947 
948 static void tso_start(struct sfxge_txq *txq, struct sfxge_tso_state *tso,
949 		      const bus_dma_segment_t *hdr_dma_seg,
950 		      struct mbuf *mbuf)
951 {
952 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->sc->enp);
953 #if !SFXGE_TX_PARSE_EARLY
954 	struct ether_header *eh = mtod(mbuf, struct ether_header *);
955 	const struct tcphdr *th;
956 	struct tcphdr th_copy;
957 #endif
958 
959 	tso->fw_assisted = txq->tso_fw_assisted;
960 	tso->mbuf = mbuf;
961 
962 	/* Find network protocol and header */
963 #if !SFXGE_TX_PARSE_EARLY
964 	tso->protocol = eh->ether_type;
965 	if (tso->protocol == htons(ETHERTYPE_VLAN)) {
966 		struct ether_vlan_header *veh =
967 			mtod(mbuf, struct ether_vlan_header *);
968 		tso->protocol = veh->evl_proto;
969 		tso->nh_off = sizeof(*veh);
970 	} else {
971 		tso->nh_off = sizeof(*eh);
972 	}
973 #else
974 	tso->protocol = TSO_MBUF_PROTO(mbuf);
975 	tso->nh_off = mbuf->m_pkthdr.l2hlen;
976 	tso->tcph_off = mbuf->m_pkthdr.l3hlen;
977 	tso->packet_id = TSO_MBUF_PACKETID(mbuf);
978 #endif
979 
980 #if !SFXGE_TX_PARSE_EARLY
981 	/* Find TCP header */
982 	if (tso->protocol == htons(ETHERTYPE_IP)) {
983 		KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
984 			("TSO required on non-TCP packet"));
985 		tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
986 		tso->packet_id = tso_iph(tso)->ip_id;
987 	} else {
988 		KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
989 			("TSO required on non-IP packet"));
990 		KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
991 			("TSO required on non-TCP packet"));
992 		tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
993 		tso->packet_id = 0;
994 	}
995 #endif
996 
997 
998 	if (tso->fw_assisted &&
999 	    __predict_false(tso->tcph_off >
1000 			    encp->enc_tx_tso_tcp_header_offset_limit)) {
1001 		tso->fw_assisted = 0;
1002 	}
1003 
1004 
1005 #if !SFXGE_TX_PARSE_EARLY
1006 	KASSERT(mbuf->m_len >= tso->tcph_off,
1007 		("network header is fragmented in mbuf"));
1008 	/* We need TCP header including flags (window is the next) */
1009 	if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) {
1010 		m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy),
1011 			   (caddr_t)&th_copy);
1012 		th = &th_copy;
1013 	} else {
1014 		th = tso_tcph(tso);
1015 	}
1016 	tso->header_len = tso->tcph_off + 4 * th->th_off;
1017 #else
1018 	tso->header_len = mbuf->m_pkthdr.l4hlen;
1019 #endif
1020 	tso->seg_size = mbuf->m_pkthdr.tso_segsz;
1021 
1022 #if !SFXGE_TX_PARSE_EARLY
1023 	tso->seqnum = ntohl(th->th_seq);
1024 
1025 	/* These flags must not be duplicated */
1026 	/*
1027 	 * RST should not be duplicated as well, but FreeBSD kernel
1028 	 * generates TSO packets with RST flag. So, do not assert
1029 	 * its absence.
1030 	 */
1031 	KASSERT(!(th->th_flags & (TH_URG | TH_SYN)),
1032 		("incompatible TCP flag 0x%x on TSO packet",
1033 		 th->th_flags & (TH_URG | TH_SYN)));
1034 	tso->tcp_flags = th->th_flags;
1035 #else
1036 	tso->seqnum = TSO_MBUF_SEQNUM(mbuf);
1037 	tso->tcp_flags = TSO_MBUF_FLAGS(mbuf);
1038 #endif
1039 
1040 	tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
1041 
1042 	if (tso->fw_assisted) {
1043 		if (hdr_dma_seg->ds_len >= tso->header_len)
1044 			efx_tx_qdesc_dma_create(txq->common,
1045 						hdr_dma_seg->ds_addr,
1046 						tso->header_len,
1047 						B_FALSE,
1048 						&tso->header_desc);
1049 		else
1050 			tso->fw_assisted = 0;
1051 	}
1052 }
1053 
1054 /*
1055  * tso_fill_packet_with_fragment - form descriptors for the current fragment
1056  *
1057  * Form descriptors for the current fragment, until we reach the end
1058  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
1059  * space.
1060  */
1061 static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
1062 					  struct sfxge_tso_state *tso)
1063 {
1064 	efx_desc_t *desc;
1065 	int n;
1066 	uint64_t dma_addr = tso->dma_addr;
1067 	boolean_t eop;
1068 
1069 	if (tso->in_len == 0 || tso->packet_space == 0)
1070 		return;
1071 
1072 	KASSERT(tso->in_len > 0, ("TSO input length went negative"));
1073 	KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
1074 
1075 	if (tso->fw_assisted & SFXGE_FATSOV2) {
1076 		n = tso->in_len;
1077 		tso->out_len -= n;
1078 		tso->seqnum += n;
1079 		tso->in_len = 0;
1080 		if (n < tso->packet_space) {
1081 			tso->packet_space -= n;
1082 			tso->segs_space--;
1083 		} else {
1084 			tso->packet_space = tso->seg_size -
1085 			    (n - tso->packet_space) % tso->seg_size;
1086 			tso->segs_space =
1087 			    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1 -
1088 			    (tso->packet_space != tso->seg_size);
1089 		}
1090 	} else {
1091 		n = min(tso->in_len, tso->packet_space);
1092 		tso->packet_space -= n;
1093 		tso->out_len -= n;
1094 		tso->dma_addr += n;
1095 		tso->in_len -= n;
1096 	}
1097 
1098 	/*
1099 	 * It is OK to use binary OR below to avoid extra branching
1100 	 * since all conditions may always be checked.
1101 	 */
1102 	eop = (tso->out_len == 0) | (tso->packet_space == 0) |
1103 	    (tso->segs_space == 0);
1104 
1105 	desc = &txq->pend_desc[txq->n_pend_desc++];
1106 	efx_tx_qdesc_dma_create(txq->common, dma_addr, n, eop, desc);
1107 }
1108 
1109 /* Callback from bus_dmamap_load() for long TSO headers. */
1110 static void tso_map_long_header(void *dma_addr_ret,
1111 				bus_dma_segment_t *segs, int nseg,
1112 				int error)
1113 {
1114 	*(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
1115 				      __predict_true(nseg == 1)) ?
1116 				     segs->ds_addr : 0);
1117 }
1118 
1119 /*
1120  * tso_start_new_packet - generate a new header and prepare for the new packet
1121  *
1122  * Generate a new header and prepare for the new packet.  Return 0 on
1123  * success, or an error code if failed to alloc header.
1124  */
1125 static int tso_start_new_packet(struct sfxge_txq *txq,
1126 				struct sfxge_tso_state *tso,
1127 				unsigned int *idp)
1128 {
1129 	unsigned int id = *idp;
1130 	struct tcphdr *tsoh_th;
1131 	unsigned ip_length;
1132 	caddr_t header;
1133 	uint64_t dma_addr;
1134 	bus_dmamap_t map;
1135 	efx_desc_t *desc;
1136 	int rc;
1137 
1138 	if (tso->fw_assisted) {
1139 		if (tso->fw_assisted & SFXGE_FATSOV2) {
1140 			/* Add 2 FATSOv2 option descriptors */
1141 			desc = &txq->pend_desc[txq->n_pend_desc];
1142 			efx_tx_qdesc_tso2_create(txq->common,
1143 						 tso->packet_id,
1144 						 tso->seqnum,
1145 						 tso->seg_size,
1146 						 desc,
1147 						 EFX_TX_FATSOV2_OPT_NDESCS);
1148 			desc += EFX_TX_FATSOV2_OPT_NDESCS;
1149 			txq->n_pend_desc += EFX_TX_FATSOV2_OPT_NDESCS;
1150 			KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1151 			id = (id + EFX_TX_FATSOV2_OPT_NDESCS) & txq->ptr_mask;
1152 
1153 			tso->segs_space =
1154 			    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1;
1155 		} else {
1156 			uint8_t tcp_flags = tso->tcp_flags;
1157 
1158 			if (tso->out_len > tso->seg_size)
1159 				tcp_flags &= ~(TH_FIN | TH_PUSH);
1160 
1161 			/* Add FATSOv1 option descriptor */
1162 			desc = &txq->pend_desc[txq->n_pend_desc++];
1163 			efx_tx_qdesc_tso_create(txq->common,
1164 						tso->packet_id,
1165 						tso->seqnum,
1166 						tcp_flags,
1167 						desc++);
1168 			KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1169 			id = (id + 1) & txq->ptr_mask;
1170 
1171 			tso->seqnum += tso->seg_size;
1172 			tso->segs_space = UINT_MAX;
1173 		}
1174 
1175 		/* Header DMA descriptor */
1176 		*desc = tso->header_desc;
1177 		txq->n_pend_desc++;
1178 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1179 		id = (id + 1) & txq->ptr_mask;
1180 	} else {
1181 		/* Allocate a DMA-mapped header buffer. */
1182 		if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
1183 			unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
1184 			unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
1185 
1186 			header = (txq->tsoh_buffer[page_index].esm_base +
1187 				  buf_index * TSOH_STD_SIZE);
1188 			dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
1189 				    buf_index * TSOH_STD_SIZE);
1190 			map = txq->tsoh_buffer[page_index].esm_map;
1191 
1192 			KASSERT(txq->stmp[id].flags == 0,
1193 				("stmp flags are not 0"));
1194 		} else {
1195 			struct sfxge_tx_mapping *stmp = &txq->stmp[id];
1196 
1197 			/* We cannot use bus_dmamem_alloc() as that may sleep */
1198 			header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
1199 			if (__predict_false(!header))
1200 				return (ENOMEM);
1201 			rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
1202 					     header, tso->header_len,
1203 					     tso_map_long_header, &dma_addr,
1204 					     BUS_DMA_NOWAIT);
1205 			if (__predict_false(dma_addr == 0)) {
1206 				if (rc == 0) {
1207 					/* Succeeded but got >1 segment */
1208 					bus_dmamap_unload(txq->packet_dma_tag,
1209 							  stmp->map);
1210 					rc = EINVAL;
1211 				}
1212 				free(header, M_SFXGE);
1213 				return (rc);
1214 			}
1215 			map = stmp->map;
1216 
1217 			txq->tso_long_headers++;
1218 			stmp->u.heap_buf = header;
1219 			stmp->flags = TX_BUF_UNMAP;
1220 		}
1221 
1222 		tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
1223 
1224 		/* Copy and update the headers. */
1225 		m_copydata(tso->mbuf, 0, tso->header_len, header);
1226 
1227 		tsoh_th->th_seq = htonl(tso->seqnum);
1228 		tso->seqnum += tso->seg_size;
1229 		if (tso->out_len > tso->seg_size) {
1230 			/* This packet will not finish the TSO burst. */
1231 			ip_length = tso->header_len - tso->nh_off + tso->seg_size;
1232 			tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH);
1233 		} else {
1234 			/* This packet will be the last in the TSO burst. */
1235 			ip_length = tso->header_len - tso->nh_off + tso->out_len;
1236 		}
1237 
1238 		if (tso->protocol == htons(ETHERTYPE_IP)) {
1239 			struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
1240 			tsoh_iph->ip_len = htons(ip_length);
1241 			/* XXX We should increment ip_id, but FreeBSD doesn't
1242 			 * currently allocate extra IDs for multiple segments.
1243 			 */
1244 		} else {
1245 			struct ip6_hdr *tsoh_iph =
1246 				(struct ip6_hdr *)(header + tso->nh_off);
1247 			tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
1248 		}
1249 
1250 		/* Make the header visible to the hardware. */
1251 		bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
1252 
1253 		/* Form a descriptor for this header. */
1254 		desc = &txq->pend_desc[txq->n_pend_desc++];
1255 		efx_tx_qdesc_dma_create(txq->common,
1256 					dma_addr,
1257 					tso->header_len,
1258 					0,
1259 					desc);
1260 		id = (id + 1) & txq->ptr_mask;
1261 
1262 		tso->segs_space = UINT_MAX;
1263 	}
1264 	tso->packet_space = tso->seg_size;
1265 	txq->tso_packets++;
1266 	*idp = id;
1267 
1268 	return (0);
1269 }
1270 
1271 static int
1272 sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
1273 		   const bus_dma_segment_t *dma_seg, int n_dma_seg,
1274 		   int vlan_tagged)
1275 {
1276 	struct sfxge_tso_state tso;
1277 	unsigned int id;
1278 	unsigned skipped = 0;
1279 
1280 	tso_start(txq, &tso, dma_seg, mbuf);
1281 
1282 	while (dma_seg->ds_len + skipped <= tso.header_len) {
1283 		skipped += dma_seg->ds_len;
1284 		--n_dma_seg;
1285 		KASSERT(n_dma_seg, ("no payload found in TSO packet"));
1286 		++dma_seg;
1287 	}
1288 	tso.in_len = dma_seg->ds_len - (tso.header_len - skipped);
1289 	tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped);
1290 
1291 	id = (txq->added + vlan_tagged) & txq->ptr_mask;
1292 	if (__predict_false(tso_start_new_packet(txq, &tso, &id)))
1293 		return (-1);
1294 
1295 	while (1) {
1296 		tso_fill_packet_with_fragment(txq, &tso);
1297 		/* Exactly one DMA descriptor is added */
1298 		KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0"));
1299 		id = (id + 1) & txq->ptr_mask;
1300 
1301 		/* Move onto the next fragment? */
1302 		if (tso.in_len == 0) {
1303 			--n_dma_seg;
1304 			if (n_dma_seg == 0)
1305 				break;
1306 			++dma_seg;
1307 			tso.in_len = dma_seg->ds_len;
1308 			tso.dma_addr = dma_seg->ds_addr;
1309 		}
1310 
1311 		/* End of packet? */
1312 		if ((tso.packet_space == 0) | (tso.segs_space == 0)) {
1313 			unsigned int n_fatso_opt_desc =
1314 			    (tso.fw_assisted & SFXGE_FATSOV2) ?
1315 			    EFX_TX_FATSOV2_OPT_NDESCS :
1316 			    (tso.fw_assisted & SFXGE_FATSOV1) ? 1 : 0;
1317 
1318 			/* If the queue is now full due to tiny MSS,
1319 			 * or we can't create another header, discard
1320 			 * the remainder of the input mbuf but do not
1321 			 * roll back the work we have done.
1322 			 */
1323 			if (txq->n_pend_desc + n_fatso_opt_desc +
1324 			    1 /* header */ + n_dma_seg > txq->max_pkt_desc) {
1325 				txq->tso_pdrop_too_many++;
1326 				break;
1327 			}
1328 			if (__predict_false(tso_start_new_packet(txq, &tso,
1329 								 &id))) {
1330 				txq->tso_pdrop_no_rsrc++;
1331 				break;
1332 			}
1333 		}
1334 	}
1335 
1336 	txq->tso_bursts++;
1337 	return (id);
1338 }
1339 
1340 static void
1341 sfxge_tx_qunblock(struct sfxge_txq *txq)
1342 {
1343 	struct sfxge_softc *sc;
1344 	struct sfxge_evq *evq;
1345 
1346 	sc = txq->sc;
1347 	evq = sc->evq[txq->evq_index];
1348 
1349 	SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
1350 
1351 	if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED))
1352 		return;
1353 
1354 	SFXGE_TXQ_LOCK(txq);
1355 
1356 	if (txq->blocked) {
1357 		unsigned int level;
1358 
1359 		level = txq->added - txq->completed;
1360 		if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) {
1361 			/* reaped must be in sync with blocked */
1362 			sfxge_tx_qreap(txq);
1363 			txq->blocked = 0;
1364 		}
1365 	}
1366 
1367 	sfxge_tx_qdpl_service(txq);
1368 	/* note: lock has been dropped */
1369 }
1370 
1371 void
1372 sfxge_tx_qflush_done(struct sfxge_txq *txq)
1373 {
1374 
1375 	txq->flush_state = SFXGE_FLUSH_DONE;
1376 }
1377 
1378 static void
1379 sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1380 {
1381 	struct sfxge_txq *txq;
1382 	struct sfxge_evq *evq;
1383 	unsigned int count;
1384 
1385 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1386 
1387 	txq = sc->txq[index];
1388 	evq = sc->evq[txq->evq_index];
1389 
1390 	SFXGE_EVQ_LOCK(evq);
1391 	SFXGE_TXQ_LOCK(txq);
1392 
1393 	KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1394 	    ("txq->init_state != SFXGE_TXQ_STARTED"));
1395 
1396 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1397 
1398 	if (txq->flush_state != SFXGE_FLUSH_DONE) {
1399 		txq->flush_state = SFXGE_FLUSH_PENDING;
1400 
1401 		SFXGE_EVQ_UNLOCK(evq);
1402 		SFXGE_TXQ_UNLOCK(txq);
1403 
1404 		/* Flush the transmit queue. */
1405 		if (efx_tx_qflush(txq->common) != 0) {
1406 			log(LOG_ERR, "%s: Flushing Tx queue %u failed\n",
1407 			    device_get_nameunit(sc->dev), index);
1408 			txq->flush_state = SFXGE_FLUSH_DONE;
1409 		} else {
1410 			count = 0;
1411 			do {
1412 				/* Spin for 100ms. */
1413 				DELAY(100000);
1414 				if (txq->flush_state != SFXGE_FLUSH_PENDING)
1415 					break;
1416 			} while (++count < 20);
1417 		}
1418 		SFXGE_EVQ_LOCK(evq);
1419 		SFXGE_TXQ_LOCK(txq);
1420 
1421 		KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1422 		    ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1423 
1424 		if (txq->flush_state != SFXGE_FLUSH_DONE) {
1425 			/* Flush timeout */
1426 			log(LOG_ERR, "%s: Cannot flush Tx queue %u\n",
1427 			    device_get_nameunit(sc->dev), index);
1428 			txq->flush_state = SFXGE_FLUSH_DONE;
1429 		}
1430 	}
1431 
1432 	txq->blocked = 0;
1433 	txq->pending = txq->added;
1434 
1435 	sfxge_tx_qcomplete(txq, evq);
1436 	KASSERT(txq->completed == txq->added,
1437 	    ("txq->completed != txq->added"));
1438 
1439 	sfxge_tx_qreap(txq);
1440 	KASSERT(txq->reaped == txq->completed,
1441 	    ("txq->reaped != txq->completed"));
1442 
1443 	txq->added = 0;
1444 	txq->pending = 0;
1445 	txq->completed = 0;
1446 	txq->reaped = 0;
1447 
1448 	/* Destroy the common code transmit queue. */
1449 	efx_tx_qdestroy(txq->common);
1450 	txq->common = NULL;
1451 
1452 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1453 	    EFX_TXQ_NBUFS(sc->txq_entries));
1454 
1455 	SFXGE_EVQ_UNLOCK(evq);
1456 	SFXGE_TXQ_UNLOCK(txq);
1457 }
1458 
1459 /*
1460  * Estimate maximum number of Tx descriptors required for TSO packet.
1461  * With minimum MSS and maximum mbuf length we might need more (even
1462  * than a ring-ful of descriptors), but this should not happen in
1463  * practice except due to deliberate attack.  In that case we will
1464  * truncate the output at a packet boundary.
1465  */
1466 static unsigned int
1467 sfxge_tx_max_pkt_desc(const struct sfxge_softc *sc, enum sfxge_txq_type type,
1468 		      unsigned int tso_fw_assisted)
1469 {
1470 	/* One descriptor for every input fragment */
1471 	unsigned int max_descs = SFXGE_TX_MAPPING_MAX_SEG;
1472 	unsigned int sw_tso_max_descs;
1473 	unsigned int fa_tso_v1_max_descs = 0;
1474 	unsigned int fa_tso_v2_max_descs = 0;
1475 
1476 	/* VLAN tagging Tx option descriptor may be required */
1477 	if (efx_nic_cfg_get(sc->enp)->enc_hw_tx_insert_vlan_enabled)
1478 		max_descs++;
1479 
1480 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) {
1481 		/*
1482 		 * Plus header and payload descriptor for each output segment.
1483 		 * Minus one since header fragment is already counted.
1484 		 * Even if FATSO is used, we should be ready to fallback
1485 		 * to do it in the driver.
1486 		 */
1487 		sw_tso_max_descs = SFXGE_TSO_MAX_SEGS * 2 - 1;
1488 
1489 		/* FW assisted TSOv1 requires one more descriptor per segment
1490 		 * in comparison to SW TSO */
1491 		if (tso_fw_assisted & SFXGE_FATSOV1)
1492 			fa_tso_v1_max_descs =
1493 			    sw_tso_max_descs + SFXGE_TSO_MAX_SEGS;
1494 
1495 		/* FW assisted TSOv2 requires 3 (2 FATSO plus header) extra
1496 		 * descriptors per superframe limited by number of DMA fetches
1497 		 * per packet. The first packet header is already counted.
1498 		 */
1499 		if (tso_fw_assisted & SFXGE_FATSOV2) {
1500 			fa_tso_v2_max_descs =
1501 			    howmany(SFXGE_TX_MAPPING_MAX_SEG,
1502 				    EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1) *
1503 			    (EFX_TX_FATSOV2_OPT_NDESCS + 1) - 1;
1504 		}
1505 
1506 		max_descs += MAX(sw_tso_max_descs,
1507 				 MAX(fa_tso_v1_max_descs, fa_tso_v2_max_descs));
1508 	}
1509 
1510 	return (max_descs);
1511 }
1512 
1513 static int
1514 sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1515 {
1516 	struct sfxge_txq *txq;
1517 	efsys_mem_t *esmp;
1518 	uint16_t flags;
1519 	unsigned int tso_fw_assisted;
1520 	struct sfxge_evq *evq;
1521 	unsigned int desc_index;
1522 	int rc;
1523 
1524 	SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
1525 
1526 	txq = sc->txq[index];
1527 	esmp = &txq->mem;
1528 	evq = sc->evq[txq->evq_index];
1529 
1530 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1531 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1532 	KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1533 	    ("evq->init_state != SFXGE_EVQ_STARTED"));
1534 
1535 	/* Program the buffer table. */
1536 	if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1537 	    EFX_TXQ_NBUFS(sc->txq_entries))) != 0)
1538 		return (rc);
1539 
1540 	/* Determine the kind of queue we are creating. */
1541 	tso_fw_assisted = 0;
1542 	switch (txq->type) {
1543 	case SFXGE_TXQ_NON_CKSUM:
1544 		flags = 0;
1545 		break;
1546 	case SFXGE_TXQ_IP_CKSUM:
1547 		flags = EFX_TXQ_CKSUM_IPV4;
1548 		break;
1549 	case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1550 		flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP;
1551 		tso_fw_assisted = sc->tso_fw_assisted;
1552 		if (tso_fw_assisted & SFXGE_FATSOV2)
1553 			flags |= EFX_TXQ_FATSOV2;
1554 		break;
1555 	default:
1556 		KASSERT(0, ("Impossible TX queue"));
1557 		flags = 0;
1558 		break;
1559 	}
1560 
1561 	/* Create the common code transmit queue. */
1562 	if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1563 	    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1564 	    &txq->common, &desc_index)) != 0) {
1565 		/* Retry if no FATSOv2 resources, otherwise fail */
1566 		if ((rc != ENOSPC) || (~flags & EFX_TXQ_FATSOV2))
1567 			goto fail;
1568 
1569 		/* Looks like all FATSOv2 contexts are used */
1570 		flags &= ~EFX_TXQ_FATSOV2;
1571 		tso_fw_assisted &= ~SFXGE_FATSOV2;
1572 		if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1573 		    sc->txq_entries, txq->buf_base_id, flags, evq->common,
1574 		    &txq->common, &desc_index)) != 0)
1575 			goto fail;
1576 	}
1577 
1578 	/* Initialise queue descriptor indexes */
1579 	txq->added = txq->pending = txq->completed = txq->reaped = desc_index;
1580 
1581 	SFXGE_TXQ_LOCK(txq);
1582 
1583 	/* Enable the transmit queue. */
1584 	efx_tx_qenable(txq->common);
1585 
1586 	txq->init_state = SFXGE_TXQ_STARTED;
1587 	txq->flush_state = SFXGE_FLUSH_REQUIRED;
1588 	txq->tso_fw_assisted = tso_fw_assisted;
1589 
1590 	txq->max_pkt_desc = sfxge_tx_max_pkt_desc(sc, txq->type,
1591 						  tso_fw_assisted);
1592 
1593 	SFXGE_TXQ_UNLOCK(txq);
1594 
1595 	return (0);
1596 
1597 fail:
1598 	efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1599 	    EFX_TXQ_NBUFS(sc->txq_entries));
1600 	return (rc);
1601 }
1602 
1603 void
1604 sfxge_tx_stop(struct sfxge_softc *sc)
1605 {
1606 	int index;
1607 
1608 	index = sc->txq_count;
1609 	while (--index >= 0)
1610 		sfxge_tx_qstop(sc, index);
1611 
1612 	/* Tear down the transmit module */
1613 	efx_tx_fini(sc->enp);
1614 }
1615 
1616 int
1617 sfxge_tx_start(struct sfxge_softc *sc)
1618 {
1619 	int index;
1620 	int rc;
1621 
1622 	/* Initialize the common code transmit module. */
1623 	if ((rc = efx_tx_init(sc->enp)) != 0)
1624 		return (rc);
1625 
1626 	for (index = 0; index < sc->txq_count; index++) {
1627 		if ((rc = sfxge_tx_qstart(sc, index)) != 0)
1628 			goto fail;
1629 	}
1630 
1631 	return (0);
1632 
1633 fail:
1634 	while (--index >= 0)
1635 		sfxge_tx_qstop(sc, index);
1636 
1637 	efx_tx_fini(sc->enp);
1638 
1639 	return (rc);
1640 }
1641 
1642 static int
1643 sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node)
1644 {
1645 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev);
1646 	struct sysctl_oid *stat_node;
1647 	unsigned int id;
1648 
1649 	stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1650 				    "stats", CTLFLAG_RD, NULL,
1651 				    "Tx queue statistics");
1652 	if (stat_node == NULL)
1653 		return (ENOMEM);
1654 
1655 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1656 		SYSCTL_ADD_ULONG(
1657 		    ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO,
1658 		    sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS,
1659 		    (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset),
1660 		    "");
1661 	}
1662 
1663 	return (0);
1664 }
1665 
1666 /**
1667  * Destroy a transmit queue.
1668  */
1669 static void
1670 sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1671 {
1672 	struct sfxge_txq *txq;
1673 	unsigned int nmaps;
1674 
1675 	txq = sc->txq[index];
1676 
1677 	KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1678 	    ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1679 
1680 	if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1681 		tso_fini(txq);
1682 
1683 	/* Free the context arrays. */
1684 	free(txq->pend_desc, M_SFXGE);
1685 	nmaps = sc->txq_entries;
1686 	while (nmaps-- != 0)
1687 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1688 	free(txq->stmp, M_SFXGE);
1689 
1690 	/* Release DMA memory mapping. */
1691 	sfxge_dma_free(&txq->mem);
1692 
1693 	sc->txq[index] = NULL;
1694 
1695 	SFXGE_TXQ_LOCK_DESTROY(txq);
1696 
1697 	free(txq, M_SFXGE);
1698 }
1699 
1700 static int
1701 sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1702 	       enum sfxge_txq_type type, unsigned int evq_index)
1703 {
1704 	char name[16];
1705 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1706 	struct sysctl_oid *txq_node;
1707 	struct sfxge_txq *txq;
1708 	struct sfxge_evq *evq;
1709 	struct sfxge_tx_dpl *stdp;
1710 	struct sysctl_oid *dpl_node;
1711 	efsys_mem_t *esmp;
1712 	unsigned int nmaps;
1713 	int rc;
1714 
1715 	txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1716 	txq->sc = sc;
1717 	txq->entries = sc->txq_entries;
1718 	txq->ptr_mask = txq->entries - 1;
1719 
1720 	sc->txq[txq_index] = txq;
1721 	esmp = &txq->mem;
1722 
1723 	evq = sc->evq[evq_index];
1724 
1725 	/* Allocate and zero DMA space for the descriptor ring. */
1726 	if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0)
1727 		return (rc);
1728 
1729 	/* Allocate buffer table entries. */
1730 	sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries),
1731 				 &txq->buf_base_id);
1732 
1733 	/* Create a DMA tag for packet mappings. */
1734 	if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000,
1735 	    MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1736 	    NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL,
1737 	    &txq->packet_dma_tag) != 0) {
1738 		device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1739 		rc = ENOMEM;
1740 		goto fail;
1741 	}
1742 
1743 	/* Allocate pending descriptor array for batching writes. */
1744 	txq->pend_desc = malloc(sizeof(efx_desc_t) * sc->txq_entries,
1745 				M_SFXGE, M_ZERO | M_WAITOK);
1746 
1747 	/* Allocate and initialise mbuf DMA mapping array. */
1748 	txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries,
1749 	    M_SFXGE, M_ZERO | M_WAITOK);
1750 	for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) {
1751 		rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1752 				       &txq->stmp[nmaps].map);
1753 		if (rc != 0)
1754 			goto fail2;
1755 	}
1756 
1757 	snprintf(name, sizeof(name), "%u", txq_index);
1758 	txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node),
1759 				   OID_AUTO, name, CTLFLAG_RD, NULL, "");
1760 	if (txq_node == NULL) {
1761 		rc = ENOMEM;
1762 		goto fail_txq_node;
1763 	}
1764 
1765 	if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1766 	    (rc = tso_init(txq)) != 0)
1767 		goto fail3;
1768 
1769 	if (sfxge_tx_dpl_get_max <= 0) {
1770 		log(LOG_ERR, "%s=%d must be greater than 0",
1771 		    SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max);
1772 		rc = EINVAL;
1773 		goto fail_tx_dpl_get_max;
1774 	}
1775 	if (sfxge_tx_dpl_get_non_tcp_max <= 0) {
1776 		log(LOG_ERR, "%s=%d must be greater than 0",
1777 		    SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX,
1778 		    sfxge_tx_dpl_get_non_tcp_max);
1779 		rc = EINVAL;
1780 		goto fail_tx_dpl_get_max;
1781 	}
1782 	if (sfxge_tx_dpl_put_max < 0) {
1783 		log(LOG_ERR, "%s=%d must be greater or equal to 0",
1784 		    SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max);
1785 		rc = EINVAL;
1786 		goto fail_tx_dpl_put_max;
1787 	}
1788 
1789 	/* Initialize the deferred packet list. */
1790 	stdp = &txq->dpl;
1791 	stdp->std_put_max = sfxge_tx_dpl_put_max;
1792 	stdp->std_get_max = sfxge_tx_dpl_get_max;
1793 	stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max;
1794 	stdp->std_getp = &stdp->std_get;
1795 
1796 	SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index);
1797 
1798 	dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO,
1799 				   "dpl", CTLFLAG_RD, NULL,
1800 				   "Deferred packet list statistics");
1801 	if (dpl_node == NULL) {
1802 		rc = ENOMEM;
1803 		goto fail_dpl_node;
1804 	}
1805 
1806 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1807 			"get_count", CTLFLAG_RD | CTLFLAG_STATS,
1808 			&stdp->std_get_count, 0, "");
1809 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1810 			"get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS,
1811 			&stdp->std_get_non_tcp_count, 0, "");
1812 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1813 			"get_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1814 			&stdp->std_get_hiwat, 0, "");
1815 	SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO,
1816 			"put_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1817 			&stdp->std_put_hiwat, 0, "");
1818 
1819 	rc = sfxge_txq_stat_init(txq, txq_node);
1820 	if (rc != 0)
1821 		goto fail_txq_stat_init;
1822 
1823 	txq->type = type;
1824 	txq->evq_index = evq_index;
1825 	txq->txq_index = txq_index;
1826 	txq->init_state = SFXGE_TXQ_INITIALIZED;
1827 	txq->hw_vlan_tci = 0;
1828 
1829 	return (0);
1830 
1831 fail_txq_stat_init:
1832 fail_dpl_node:
1833 fail_tx_dpl_put_max:
1834 fail_tx_dpl_get_max:
1835 fail3:
1836 fail_txq_node:
1837 	free(txq->pend_desc, M_SFXGE);
1838 fail2:
1839 	while (nmaps-- != 0)
1840 		bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1841 	free(txq->stmp, M_SFXGE);
1842 	bus_dma_tag_destroy(txq->packet_dma_tag);
1843 
1844 fail:
1845 	sfxge_dma_free(esmp);
1846 
1847 	return (rc);
1848 }
1849 
1850 static int
1851 sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1852 {
1853 	struct sfxge_softc *sc = arg1;
1854 	unsigned int id = arg2;
1855 	unsigned long sum;
1856 	unsigned int index;
1857 
1858 	/* Sum across all TX queues */
1859 	sum = 0;
1860 	for (index = 0; index < sc->txq_count; index++)
1861 		sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1862 					  sfxge_tx_stats[id].offset);
1863 
1864 	return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1865 }
1866 
1867 static void
1868 sfxge_tx_stat_init(struct sfxge_softc *sc)
1869 {
1870 	struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1871 	struct sysctl_oid_list *stat_list;
1872 	unsigned int id;
1873 
1874 	stat_list = SYSCTL_CHILDREN(sc->stats_node);
1875 
1876 	for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1877 		SYSCTL_ADD_PROC(
1878 			ctx, stat_list,
1879 			OID_AUTO, sfxge_tx_stats[id].name,
1880 			CTLTYPE_ULONG|CTLFLAG_RD,
1881 			sc, id, sfxge_tx_stat_handler, "LU",
1882 			"");
1883 	}
1884 }
1885 
1886 uint64_t
1887 sfxge_tx_get_drops(struct sfxge_softc *sc)
1888 {
1889 	unsigned int index;
1890 	uint64_t drops = 0;
1891 	struct sfxge_txq *txq;
1892 
1893 	/* Sum across all TX queues */
1894 	for (index = 0; index < sc->txq_count; index++) {
1895 		txq = sc->txq[index];
1896 		/*
1897 		 * In theory, txq->put_overflow and txq->netdown_drops
1898 		 * should use atomic operation and other should be
1899 		 * obtained under txq lock, but it is just statistics.
1900 		 */
1901 		drops += txq->drops + txq->get_overflow +
1902 			 txq->get_non_tcp_overflow +
1903 			 txq->put_overflow + txq->netdown_drops +
1904 			 txq->tso_pdrop_too_many + txq->tso_pdrop_no_rsrc;
1905 	}
1906 	return (drops);
1907 }
1908 
1909 void
1910 sfxge_tx_fini(struct sfxge_softc *sc)
1911 {
1912 	int index;
1913 
1914 	index = sc->txq_count;
1915 	while (--index >= 0)
1916 		sfxge_tx_qfini(sc, index);
1917 
1918 	sc->txq_count = 0;
1919 }
1920 
1921 
1922 int
1923 sfxge_tx_init(struct sfxge_softc *sc)
1924 {
1925 	const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
1926 	struct sfxge_intr *intr;
1927 	int index;
1928 	int rc;
1929 
1930 	intr = &sc->intr;
1931 
1932 	KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1933 	    ("intr->state != SFXGE_INTR_INITIALIZED"));
1934 
1935 	sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc;
1936 
1937 	sc->tso_fw_assisted = sfxge_tso_fw_assisted;
1938 	if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO) ||
1939 	    (!encp->enc_fw_assisted_tso_enabled))
1940 		sc->tso_fw_assisted &= ~SFXGE_FATSOV1;
1941 	if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO_V2) ||
1942 	    (!encp->enc_fw_assisted_tso_v2_enabled))
1943 		sc->tso_fw_assisted &= ~SFXGE_FATSOV2;
1944 
1945 	sc->txqs_node = SYSCTL_ADD_NODE(
1946 		device_get_sysctl_ctx(sc->dev),
1947 		SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
1948 		OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues");
1949 	if (sc->txqs_node == NULL) {
1950 		rc = ENOMEM;
1951 		goto fail_txq_node;
1952 	}
1953 
1954 	/* Initialize the transmit queues */
1955 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
1956 	    SFXGE_TXQ_NON_CKSUM, 0)) != 0)
1957 		goto fail;
1958 
1959 	if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
1960 	    SFXGE_TXQ_IP_CKSUM, 0)) != 0)
1961 		goto fail2;
1962 
1963 	for (index = 0;
1964 	     index < sc->txq_count - SFXGE_TXQ_NTYPES + 1;
1965 	     index++) {
1966 		if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index,
1967 		    SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
1968 			goto fail3;
1969 	}
1970 
1971 	sfxge_tx_stat_init(sc);
1972 
1973 	return (0);
1974 
1975 fail3:
1976 	while (--index >= 0)
1977 		sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1978 
1979 	sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1980 
1981 fail2:
1982 	sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1983 
1984 fail:
1985 fail_txq_node:
1986 	sc->txq_count = 0;
1987 	return (rc);
1988 }
1989