xref: /freebsd/sys/dev/ena/ena_datapath.c (revision da73e3a7)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2015-2024 Amazon.com, Inc. or its affiliates.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  *
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 #include <sys/cdefs.h>
31 #include "opt_rss.h"
32 #include "ena.h"
33 #include "ena_datapath.h"
34 #ifdef DEV_NETMAP
35 #include "ena_netmap.h"
36 #endif /* DEV_NETMAP */
37 #ifdef RSS
38 #include <net/rss_config.h>
39 #endif /* RSS */
40 
41 #include <netinet6/ip6_var.h>
42 
43 /*********************************************************************
44  *  Static functions prototypes
45  *********************************************************************/
46 
47 static int ena_tx_cleanup(struct ena_ring *);
48 static int ena_rx_cleanup(struct ena_ring *);
49 static inline int ena_get_tx_req_id(struct ena_ring *tx_ring,
50     struct ena_com_io_cq *io_cq, uint16_t *req_id);
51 static void ena_rx_hash_mbuf(struct ena_ring *, struct ena_com_rx_ctx *,
52     struct mbuf *);
53 static struct mbuf *ena_rx_mbuf(struct ena_ring *, struct ena_com_rx_buf_info *,
54     struct ena_com_rx_ctx *, uint16_t *);
55 static inline void ena_rx_checksum(struct ena_ring *, struct ena_com_rx_ctx *,
56     struct mbuf *);
57 static void ena_tx_csum(struct ena_com_tx_ctx *, struct mbuf *, bool);
58 static int ena_check_and_collapse_mbuf(struct ena_ring *tx_ring,
59     struct mbuf **mbuf);
60 static int ena_xmit_mbuf(struct ena_ring *, struct mbuf **);
61 static void ena_start_xmit(struct ena_ring *);
62 
63 /*********************************************************************
64  *  Global functions
65  *********************************************************************/
66 
67 void
ena_cleanup(void * arg,int pending)68 ena_cleanup(void *arg, int pending)
69 {
70 	struct ena_que *que = arg;
71 	struct ena_adapter *adapter = que->adapter;
72 	if_t ifp = adapter->ifp;
73 	struct ena_ring *tx_ring;
74 	struct ena_ring *rx_ring;
75 	struct ena_com_io_cq *io_cq;
76 	struct ena_eth_io_intr_reg intr_reg;
77 	int qid, ena_qid;
78 	int txc, rxc, i;
79 
80 	tx_ring = que->tx_ring;
81 	rx_ring = que->rx_ring;
82 	qid = que->id;
83 	ena_qid = ENA_IO_TXQ_IDX(qid);
84 	io_cq = &adapter->ena_dev->io_cq_queues[ena_qid];
85 
86 	atomic_store_8(&tx_ring->cleanup_running, 1);
87 	/* Need to make sure that ENA_FLAG_TRIGGER_RESET is visible to ena_cleanup() and
88 	 * that cleanup_running is visible to check_missing_comp_in_tx_queue() to
89 	 * prevent the case of accessing CQ concurrently with check_cdesc_in_tx_cq()
90 	 */
91 	mb();
92 	if (unlikely(((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0) ||
93 	    (ENA_FLAG_ISSET(ENA_FLAG_TRIGGER_RESET, adapter))))
94 		return;
95 
96 	ena_log_io(adapter->pdev, DBG, "MSI-X TX/RX routine\n");
97 
98 	atomic_store_8(&tx_ring->first_interrupt, 1);
99 	atomic_store_8(&rx_ring->first_interrupt, 1);
100 
101 	for (i = 0; i < ENA_CLEAN_BUDGET; ++i) {
102 		rxc = ena_rx_cleanup(rx_ring);
103 		txc = ena_tx_cleanup(tx_ring);
104 
105 		if (unlikely(((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0) ||
106 		    (ENA_FLAG_ISSET(ENA_FLAG_TRIGGER_RESET, adapter))))
107 			return;
108 
109 		if ((txc != ENA_TX_BUDGET) && (rxc != ENA_RX_BUDGET))
110 			break;
111 	}
112 
113 	/* Signal that work is done and unmask interrupt */
114 	ena_com_update_intr_reg(&intr_reg, ENA_RX_IRQ_INTERVAL,
115 	    ENA_TX_IRQ_INTERVAL, true, false);
116 	counter_u64_add(tx_ring->tx_stats.unmask_interrupt_num, 1);
117 	ena_com_unmask_intr(io_cq, &intr_reg);
118 	atomic_store_8(&tx_ring->cleanup_running, 0);
119 }
120 
121 void
ena_deferred_mq_start(void * arg,int pending)122 ena_deferred_mq_start(void *arg, int pending)
123 {
124 	struct ena_ring *tx_ring = (struct ena_ring *)arg;
125 	if_t ifp = tx_ring->adapter->ifp;
126 
127 	while (!drbr_empty(ifp, tx_ring->br) && tx_ring->running &&
128 	    (if_getdrvflags(ifp) & IFF_DRV_RUNNING) != 0) {
129 		ENA_RING_MTX_LOCK(tx_ring);
130 		ena_start_xmit(tx_ring);
131 		ENA_RING_MTX_UNLOCK(tx_ring);
132 	}
133 }
134 
135 int
ena_mq_start(if_t ifp,struct mbuf * m)136 ena_mq_start(if_t ifp, struct mbuf *m)
137 {
138 	struct ena_adapter *adapter = if_getsoftc(ifp);
139 	struct ena_ring *tx_ring;
140 	int ret, is_drbr_empty;
141 	uint32_t i;
142 #ifdef RSS
143 	uint32_t bucket_id;
144 #endif
145 
146 	if (unlikely((if_getdrvflags(adapter->ifp) & IFF_DRV_RUNNING) == 0))
147 		return (ENODEV);
148 
149 	/* Which queue to use */
150 	/*
151 	 * If everything is setup correctly, it should be the
152 	 * same bucket that the current CPU we're on is.
153 	 * It should improve performance.
154 	 */
155 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
156 #ifdef RSS
157 		if (rss_hash2bucket(m->m_pkthdr.flowid, M_HASHTYPE_GET(m),
158 		    &bucket_id) == 0)
159 			i = bucket_id % adapter->num_io_queues;
160 		else
161 #endif
162 			i = m->m_pkthdr.flowid % adapter->num_io_queues;
163 	} else {
164 		i = curcpu % adapter->num_io_queues;
165 	}
166 	tx_ring = &adapter->tx_ring[i];
167 
168 	/* Check if drbr is empty before putting packet */
169 	is_drbr_empty = drbr_empty(ifp, tx_ring->br);
170 	ret = drbr_enqueue(ifp, tx_ring->br, m);
171 	if (unlikely(ret != 0)) {
172 		taskqueue_enqueue(tx_ring->enqueue_tq, &tx_ring->enqueue_task);
173 		return (ret);
174 	}
175 
176 	if (is_drbr_empty && (ENA_RING_MTX_TRYLOCK(tx_ring) != 0)) {
177 		ena_start_xmit(tx_ring);
178 		ENA_RING_MTX_UNLOCK(tx_ring);
179 	} else {
180 		taskqueue_enqueue(tx_ring->enqueue_tq, &tx_ring->enqueue_task);
181 	}
182 
183 	return (0);
184 }
185 
186 void
ena_qflush(if_t ifp)187 ena_qflush(if_t ifp)
188 {
189 	struct ena_adapter *adapter = if_getsoftc(ifp);
190 	struct ena_ring *tx_ring = adapter->tx_ring;
191 	int i;
192 
193 	for (i = 0; i < adapter->num_io_queues; ++i, ++tx_ring)
194 		if (!drbr_empty(ifp, tx_ring->br)) {
195 			ENA_RING_MTX_LOCK(tx_ring);
196 			drbr_flush(ifp, tx_ring->br);
197 			ENA_RING_MTX_UNLOCK(tx_ring);
198 		}
199 
200 	if_qflush(ifp);
201 }
202 
203 /*********************************************************************
204  *  Static functions
205  *********************************************************************/
206 
207 static inline int
ena_get_tx_req_id(struct ena_ring * tx_ring,struct ena_com_io_cq * io_cq,uint16_t * req_id)208 ena_get_tx_req_id(struct ena_ring *tx_ring, struct ena_com_io_cq *io_cq,
209     uint16_t *req_id)
210 {
211 	struct ena_adapter *adapter = tx_ring->adapter;
212 	int rc = ena_com_tx_comp_req_id_get(io_cq, req_id);
213 
214 	if (unlikely(rc == ENA_COM_TRY_AGAIN))
215 		return (EAGAIN);
216 
217 	rc = validate_tx_req_id(tx_ring, *req_id, rc);
218 
219 	if (unlikely(tx_ring->tx_buffer_info[*req_id].mbuf == NULL)) {
220 		ena_log(adapter->pdev, ERR,
221 		    "tx_info doesn't have valid mbuf. req_id %hu qid %hu\n",
222 		    *req_id, tx_ring->qid);
223 		ena_trigger_reset(adapter, ENA_REGS_RESET_INV_TX_REQ_ID);
224 		rc = EFAULT;
225 	}
226 
227 	return (rc);
228 }
229 
230 /**
231  * ena_tx_cleanup - clear sent packets and corresponding descriptors
232  * @tx_ring: ring for which we want to clean packets
233  *
234  * Once packets are sent, we ask the device in a loop for no longer used
235  * descriptors. We find the related mbuf chain in a map (index in an array)
236  * and free it, then update ring state.
237  * This is performed in "endless" loop, updating ring pointers every
238  * TX_COMMIT. The first check of free descriptor is performed before the actual
239  * loop, then repeated at the loop end.
240  **/
241 static int
ena_tx_cleanup(struct ena_ring * tx_ring)242 ena_tx_cleanup(struct ena_ring *tx_ring)
243 {
244 	struct ena_adapter *adapter;
245 	struct ena_com_io_cq *io_cq;
246 	uint16_t next_to_clean;
247 	uint16_t req_id;
248 	uint16_t ena_qid;
249 	unsigned int total_done = 0;
250 	int rc;
251 	int commit = ENA_TX_COMMIT;
252 	int budget = ENA_TX_BUDGET;
253 	int work_done;
254 	bool above_thresh;
255 
256 	adapter = tx_ring->que->adapter;
257 	ena_qid = ENA_IO_TXQ_IDX(tx_ring->que->id);
258 	io_cq = &adapter->ena_dev->io_cq_queues[ena_qid];
259 	next_to_clean = tx_ring->next_to_clean;
260 
261 #ifdef DEV_NETMAP
262 	if (netmap_tx_irq(adapter->ifp, tx_ring->qid) != NM_IRQ_PASS)
263 		return (0);
264 #endif /* DEV_NETMAP */
265 
266 	do {
267 		struct ena_tx_buffer *tx_info;
268 		struct mbuf *mbuf;
269 
270 		rc = ena_get_tx_req_id(tx_ring, io_cq, &req_id);
271 		if (unlikely(rc != 0))
272 			break;
273 
274 		tx_info = &tx_ring->tx_buffer_info[req_id];
275 
276 		mbuf = tx_info->mbuf;
277 
278 		tx_info->mbuf = NULL;
279 		bintime_clear(&tx_info->timestamp);
280 
281 		bus_dmamap_sync(adapter->tx_buf_tag, tx_info->dmamap,
282 		    BUS_DMASYNC_POSTWRITE);
283 		bus_dmamap_unload(adapter->tx_buf_tag, tx_info->dmamap);
284 
285 		ena_log_io(adapter->pdev, DBG, "tx: q %d mbuf %p completed\n",
286 		    tx_ring->qid, mbuf);
287 
288 		m_freem(mbuf);
289 
290 		total_done += tx_info->tx_descs;
291 
292 		tx_ring->free_tx_ids[next_to_clean] = req_id;
293 		next_to_clean = ENA_TX_RING_IDX_NEXT(next_to_clean,
294 		    tx_ring->ring_size);
295 
296 		if (unlikely(--commit == 0)) {
297 			commit = ENA_TX_COMMIT;
298 			/* update ring state every ENA_TX_COMMIT descriptor */
299 			tx_ring->next_to_clean = next_to_clean;
300 			ena_com_comp_ack(
301 			    &adapter->ena_dev->io_sq_queues[ena_qid],
302 			    total_done);
303 			total_done = 0;
304 		}
305 	} while (likely(--budget));
306 
307 	work_done = ENA_TX_BUDGET - budget;
308 
309 	ena_log_io(adapter->pdev, DBG, "tx: q %d done. total pkts: %d\n",
310 	    tx_ring->qid, work_done);
311 
312 	/* If there is still something to commit update ring state */
313 	if (likely(commit != ENA_TX_COMMIT)) {
314 		tx_ring->next_to_clean = next_to_clean;
315 		ena_com_comp_ack(&adapter->ena_dev->io_sq_queues[ena_qid],
316 		    total_done);
317 	}
318 
319 	/*
320 	 * Need to make the rings circular update visible to
321 	 * ena_xmit_mbuf() before checking for tx_ring->running.
322 	 */
323 	mb();
324 
325 	above_thresh = ena_com_sq_have_enough_space(tx_ring->ena_com_io_sq,
326 	    ENA_TX_RESUME_THRESH);
327 	if (unlikely(!tx_ring->running && above_thresh)) {
328 		ENA_RING_MTX_LOCK(tx_ring);
329 		above_thresh = ena_com_sq_have_enough_space(
330 		    tx_ring->ena_com_io_sq, ENA_TX_RESUME_THRESH);
331 		if (!tx_ring->running && above_thresh) {
332 			tx_ring->running = true;
333 			counter_u64_add(tx_ring->tx_stats.queue_wakeup, 1);
334 			taskqueue_enqueue(tx_ring->enqueue_tq,
335 			    &tx_ring->enqueue_task);
336 		}
337 		ENA_RING_MTX_UNLOCK(tx_ring);
338 	}
339 
340 	tx_ring->tx_last_cleanup_ticks = ticks;
341 
342 	return (work_done);
343 }
344 
345 static void
ena_rx_hash_mbuf(struct ena_ring * rx_ring,struct ena_com_rx_ctx * ena_rx_ctx,struct mbuf * mbuf)346 ena_rx_hash_mbuf(struct ena_ring *rx_ring, struct ena_com_rx_ctx *ena_rx_ctx,
347     struct mbuf *mbuf)
348 {
349 	struct ena_adapter *adapter = rx_ring->adapter;
350 
351 	if (likely(ENA_FLAG_ISSET(ENA_FLAG_RSS_ACTIVE, adapter))) {
352 		mbuf->m_pkthdr.flowid = ena_rx_ctx->hash;
353 
354 #ifdef RSS
355 		/*
356 		 * Hardware and software RSS are in agreement only when both are
357 		 * configured to Toeplitz algorithm.  This driver configures
358 		 * that algorithm only when software RSS is enabled and uses it.
359 		 */
360 		if (adapter->ena_dev->rss.hash_func != ENA_ADMIN_TOEPLITZ &&
361 		    ena_rx_ctx->l3_proto != ENA_ETH_IO_L3_PROTO_UNKNOWN) {
362 			M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE_HASH);
363 			return;
364 		}
365 #endif
366 
367 		if (ena_rx_ctx->frag &&
368 		    (ena_rx_ctx->l3_proto != ENA_ETH_IO_L3_PROTO_UNKNOWN)) {
369 			M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE_HASH);
370 			return;
371 		}
372 
373 		switch (ena_rx_ctx->l3_proto) {
374 		case ENA_ETH_IO_L3_PROTO_IPV4:
375 			switch (ena_rx_ctx->l4_proto) {
376 			case ENA_ETH_IO_L4_PROTO_TCP:
377 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_TCP_IPV4);
378 				break;
379 			case ENA_ETH_IO_L4_PROTO_UDP:
380 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_UDP_IPV4);
381 				break;
382 			default:
383 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_IPV4);
384 			}
385 			break;
386 		case ENA_ETH_IO_L3_PROTO_IPV6:
387 			switch (ena_rx_ctx->l4_proto) {
388 			case ENA_ETH_IO_L4_PROTO_TCP:
389 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_TCP_IPV6);
390 				break;
391 			case ENA_ETH_IO_L4_PROTO_UDP:
392 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_UDP_IPV6);
393 				break;
394 			default:
395 				M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_IPV6);
396 			}
397 			break;
398 		case ENA_ETH_IO_L3_PROTO_UNKNOWN:
399 			M_HASHTYPE_SET(mbuf, M_HASHTYPE_NONE);
400 			break;
401 		default:
402 			M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE_HASH);
403 		}
404 	} else {
405 		mbuf->m_pkthdr.flowid = rx_ring->qid;
406 		M_HASHTYPE_SET(mbuf, M_HASHTYPE_NONE);
407 	}
408 }
409 
410 /**
411  * ena_rx_mbuf - assemble mbuf from descriptors
412  * @rx_ring: ring for which we want to clean packets
413  * @ena_bufs: buffer info
414  * @ena_rx_ctx: metadata for this packet(s)
415  * @next_to_clean: ring pointer, will be updated only upon success
416  *
417  **/
418 static struct mbuf *
ena_rx_mbuf(struct ena_ring * rx_ring,struct ena_com_rx_buf_info * ena_bufs,struct ena_com_rx_ctx * ena_rx_ctx,uint16_t * next_to_clean)419 ena_rx_mbuf(struct ena_ring *rx_ring, struct ena_com_rx_buf_info *ena_bufs,
420     struct ena_com_rx_ctx *ena_rx_ctx, uint16_t *next_to_clean)
421 {
422 	struct mbuf *mbuf;
423 	struct ena_rx_buffer *rx_info;
424 	struct ena_adapter *adapter;
425 	device_t pdev;
426 	unsigned int descs = ena_rx_ctx->descs;
427 	uint16_t ntc, len, req_id, buf = 0;
428 
429 	ntc = *next_to_clean;
430 	adapter = rx_ring->adapter;
431 	pdev = adapter->pdev;
432 
433 	len = ena_bufs[buf].len;
434 	req_id = ena_bufs[buf].req_id;
435 	rx_info = &rx_ring->rx_buffer_info[req_id];
436 	if (unlikely(rx_info->mbuf == NULL)) {
437 		ena_log(pdev, ERR, "NULL mbuf in rx_info. qid %u req_id %u\n",
438 		    rx_ring->qid, req_id);
439 		ena_trigger_reset(adapter, ENA_REGS_RESET_INV_RX_REQ_ID);
440 		return (NULL);
441 	}
442 
443 	ena_log_io(pdev, DBG, "rx_info %p, mbuf %p, paddr %jx\n", rx_info,
444 	    rx_info->mbuf, (uintmax_t)rx_info->ena_buf.paddr);
445 
446 	bus_dmamap_sync(adapter->rx_buf_tag, rx_info->map,
447 	    BUS_DMASYNC_POSTREAD);
448 	mbuf = rx_info->mbuf;
449 	mbuf->m_flags |= M_PKTHDR;
450 	mbuf->m_pkthdr.len = len;
451 	mbuf->m_len = len;
452 	/* Only for the first segment the data starts at specific offset */
453 	mbuf->m_data = mtodo(mbuf, ena_rx_ctx->pkt_offset);
454 	ena_log_io(pdev, DBG, "Mbuf data offset=%u\n", ena_rx_ctx->pkt_offset);
455 	mbuf->m_pkthdr.rcvif = rx_ring->que->adapter->ifp;
456 
457 	/* Fill mbuf with hash key and it's interpretation for optimization */
458 	ena_rx_hash_mbuf(rx_ring, ena_rx_ctx, mbuf);
459 
460 	ena_log_io(pdev, DBG, "rx mbuf 0x%p, flags=0x%x, len: %d\n", mbuf,
461 	    mbuf->m_flags, mbuf->m_pkthdr.len);
462 
463 	/* DMA address is not needed anymore, unmap it */
464 	bus_dmamap_unload(rx_ring->adapter->rx_buf_tag, rx_info->map);
465 
466 	rx_info->mbuf = NULL;
467 	rx_ring->free_rx_ids[ntc] = req_id;
468 	ntc = ENA_RX_RING_IDX_NEXT(ntc, rx_ring->ring_size);
469 
470 	/*
471 	 * While we have more than 1 descriptors for one rcvd packet, append
472 	 * other mbufs to the main one
473 	 */
474 	while (--descs) {
475 		++buf;
476 		len = ena_bufs[buf].len;
477 		req_id = ena_bufs[buf].req_id;
478 		rx_info = &rx_ring->rx_buffer_info[req_id];
479 
480 		if (unlikely(rx_info->mbuf == NULL)) {
481 			ena_log(pdev, ERR, "NULL mbuf in rx_info. qid %u req_id %u\n",
482 			    rx_ring->qid, req_id);
483 			/*
484 			 * If one of the required mbufs was not allocated yet,
485 			 * we can break there.
486 			 * All earlier used descriptors will be reallocated
487 			 * later and not used mbufs can be reused.
488 			 * The next_to_clean pointer will not be updated in case
489 			 * of an error, so caller should advance it manually
490 			 * in error handling routine to keep it up to date
491 			 * with hw ring.
492 			 */
493 			m_freem(mbuf);
494 			ena_trigger_reset(adapter, ENA_REGS_RESET_INV_RX_REQ_ID);
495 			return (NULL);
496 		}
497 
498 		bus_dmamap_sync(adapter->rx_buf_tag, rx_info->map,
499 		    BUS_DMASYNC_POSTREAD);
500 		if (unlikely(m_append(mbuf, len, rx_info->mbuf->m_data) == 0)) {
501 			counter_u64_add(rx_ring->rx_stats.mbuf_alloc_fail, 1);
502 			ena_log_io(pdev, WARN, "Failed to append Rx mbuf %p\n",
503 			    mbuf);
504 		}
505 
506 		ena_log_io(pdev, DBG, "rx mbuf updated. len %d\n",
507 		    mbuf->m_pkthdr.len);
508 
509 		/* Free already appended mbuf, it won't be useful anymore */
510 		bus_dmamap_unload(rx_ring->adapter->rx_buf_tag, rx_info->map);
511 		m_freem(rx_info->mbuf);
512 		rx_info->mbuf = NULL;
513 
514 		rx_ring->free_rx_ids[ntc] = req_id;
515 		ntc = ENA_RX_RING_IDX_NEXT(ntc, rx_ring->ring_size);
516 	}
517 
518 	*next_to_clean = ntc;
519 
520 	return (mbuf);
521 }
522 
523 /**
524  * ena_rx_checksum - indicate in mbuf if hw indicated a good cksum
525  **/
526 static inline void
ena_rx_checksum(struct ena_ring * rx_ring,struct ena_com_rx_ctx * ena_rx_ctx,struct mbuf * mbuf)527 ena_rx_checksum(struct ena_ring *rx_ring, struct ena_com_rx_ctx *ena_rx_ctx,
528     struct mbuf *mbuf)
529 {
530 	device_t pdev = rx_ring->adapter->pdev;
531 
532 	/* if IP and error */
533 	if (unlikely((ena_rx_ctx->l3_proto == ENA_ETH_IO_L3_PROTO_IPV4) &&
534 	    ena_rx_ctx->l3_csum_err)) {
535 		/* ipv4 checksum error */
536 		mbuf->m_pkthdr.csum_flags = 0;
537 		counter_u64_add(rx_ring->rx_stats.csum_bad, 1);
538 		ena_log_io(pdev, DBG, "RX IPv4 header checksum error\n");
539 		return;
540 	}
541 
542 	/* if TCP/UDP */
543 	if ((ena_rx_ctx->l4_proto == ENA_ETH_IO_L4_PROTO_TCP) ||
544 	    (ena_rx_ctx->l4_proto == ENA_ETH_IO_L4_PROTO_UDP)) {
545 		if (ena_rx_ctx->l4_csum_err) {
546 			/* TCP/UDP checksum error */
547 			mbuf->m_pkthdr.csum_flags = 0;
548 			counter_u64_add(rx_ring->rx_stats.csum_bad, 1);
549 			ena_log_io(pdev, DBG, "RX L4 checksum error\n");
550 		} else {
551 			mbuf->m_pkthdr.csum_flags = CSUM_IP_CHECKED;
552 			mbuf->m_pkthdr.csum_flags |= CSUM_IP_VALID;
553 			counter_u64_add(rx_ring->rx_stats.csum_good, 1);
554 		}
555 	}
556 }
557 
558 /**
559  * ena_rx_cleanup - handle rx irq
560  * @arg: ring for which irq is being handled
561  **/
562 static int
ena_rx_cleanup(struct ena_ring * rx_ring)563 ena_rx_cleanup(struct ena_ring *rx_ring)
564 {
565 	struct ena_adapter *adapter;
566 	device_t pdev;
567 	struct mbuf *mbuf;
568 	struct ena_com_rx_ctx ena_rx_ctx;
569 	struct ena_com_io_cq *io_cq;
570 	struct ena_com_io_sq *io_sq;
571 	enum ena_regs_reset_reason_types reset_reason;
572 	if_t ifp;
573 	uint16_t ena_qid;
574 	uint16_t next_to_clean;
575 	uint32_t refill_required;
576 	uint32_t refill_threshold;
577 	uint32_t do_if_input = 0;
578 	unsigned int qid;
579 	int rc, i;
580 	int budget = ENA_RX_BUDGET;
581 #ifdef DEV_NETMAP
582 	int done;
583 #endif /* DEV_NETMAP */
584 
585 	adapter = rx_ring->que->adapter;
586 	pdev = adapter->pdev;
587 	ifp = adapter->ifp;
588 	qid = rx_ring->que->id;
589 	ena_qid = ENA_IO_RXQ_IDX(qid);
590 	io_cq = &adapter->ena_dev->io_cq_queues[ena_qid];
591 	io_sq = &adapter->ena_dev->io_sq_queues[ena_qid];
592 	next_to_clean = rx_ring->next_to_clean;
593 
594 #ifdef DEV_NETMAP
595 	if (netmap_rx_irq(adapter->ifp, rx_ring->qid, &done) != NM_IRQ_PASS)
596 		return (0);
597 #endif /* DEV_NETMAP */
598 
599 	ena_log_io(pdev, DBG, "rx: qid %d\n", qid);
600 
601 	do {
602 		ena_rx_ctx.ena_bufs = rx_ring->ena_bufs;
603 		ena_rx_ctx.max_bufs = adapter->max_rx_sgl_size;
604 		ena_rx_ctx.descs = 0;
605 		ena_rx_ctx.pkt_offset = 0;
606 
607 		bus_dmamap_sync(io_cq->cdesc_addr.mem_handle.tag,
608 		    io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_POSTREAD);
609 		rc = ena_com_rx_pkt(io_cq, io_sq, &ena_rx_ctx);
610 		if (unlikely(rc != 0)) {
611 			if (rc == ENA_COM_NO_SPACE) {
612 				counter_u64_add(rx_ring->rx_stats.bad_desc_num,
613 				    1);
614 				reset_reason = ENA_REGS_RESET_TOO_MANY_RX_DESCS;
615 			} else if (rc == ENA_COM_FAULT) {
616 				reset_reason = ENA_REGS_RESET_RX_DESCRIPTOR_MALFORMED;
617 			} else {
618 				counter_u64_add(rx_ring->rx_stats.bad_req_id,
619 				    1);
620 				reset_reason = ENA_REGS_RESET_INV_RX_REQ_ID;
621 			}
622 			ena_trigger_reset(adapter, reset_reason);
623 			return (0);
624 		}
625 
626 		if (unlikely(ena_rx_ctx.descs == 0))
627 			break;
628 
629 		ena_log_io(pdev, DBG,
630 		    "rx: q %d got packet from ena. descs #: %d l3 proto %d l4 proto %d hash: %x\n",
631 		    rx_ring->qid, ena_rx_ctx.descs, ena_rx_ctx.l3_proto,
632 		    ena_rx_ctx.l4_proto, ena_rx_ctx.hash);
633 
634 		/* Receive mbuf from the ring */
635 		mbuf = ena_rx_mbuf(rx_ring, rx_ring->ena_bufs, &ena_rx_ctx,
636 		    &next_to_clean);
637 		bus_dmamap_sync(io_cq->cdesc_addr.mem_handle.tag,
638 		    io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_PREREAD);
639 		/* Exit if we failed to retrieve a buffer */
640 		if (unlikely(mbuf == NULL)) {
641 			for (i = 0; i < ena_rx_ctx.descs; ++i) {
642 				rx_ring->free_rx_ids[next_to_clean] =
643 				    rx_ring->ena_bufs[i].req_id;
644 				next_to_clean = ENA_RX_RING_IDX_NEXT(
645 				    next_to_clean, rx_ring->ring_size);
646 			}
647 			break;
648 		}
649 
650 		if (((if_getcapenable(ifp) & IFCAP_RXCSUM) != 0) ||
651 		    ((if_getcapenable(ifp) & IFCAP_RXCSUM_IPV6) != 0)) {
652 			ena_rx_checksum(rx_ring, &ena_rx_ctx, mbuf);
653 		}
654 
655 		counter_enter();
656 		counter_u64_add_protected(rx_ring->rx_stats.bytes,
657 		    mbuf->m_pkthdr.len);
658 		counter_u64_add_protected(adapter->hw_stats.rx_bytes,
659 		    mbuf->m_pkthdr.len);
660 		counter_exit();
661 		/*
662 		 * LRO is only for IP/TCP packets and TCP checksum of the packet
663 		 * should be computed by hardware.
664 		 */
665 		do_if_input = 1;
666 		if (((if_getcapenable(ifp) & IFCAP_LRO) != 0)  &&
667 		    ((mbuf->m_pkthdr.csum_flags & CSUM_IP_VALID) != 0) &&
668 		    (ena_rx_ctx.l4_proto == ENA_ETH_IO_L4_PROTO_TCP)) {
669 			/*
670 			 * Send to the stack if:
671 			 *  - LRO not enabled, or
672 			 *  - no LRO resources, or
673 			 *  - lro enqueue fails
674 			 */
675 			if ((rx_ring->lro.lro_cnt != 0) &&
676 			    (tcp_lro_rx(&rx_ring->lro, mbuf, 0) == 0))
677 				do_if_input = 0;
678 		}
679 		if (do_if_input != 0) {
680 			ena_log_io(pdev, DBG,
681 			    "calling if_input() with mbuf %p\n", mbuf);
682 			if_input(ifp, mbuf);
683 		}
684 
685 		counter_enter();
686 		counter_u64_add_protected(rx_ring->rx_stats.cnt, 1);
687 		counter_u64_add_protected(adapter->hw_stats.rx_packets, 1);
688 		counter_exit();
689 	} while (--budget);
690 
691 	rx_ring->next_to_clean = next_to_clean;
692 
693 	refill_required = ena_com_free_q_entries(io_sq);
694 	refill_threshold = min_t(int,
695 	    rx_ring->ring_size / ENA_RX_REFILL_THRESH_DIVIDER,
696 	    ENA_RX_REFILL_THRESH_PACKET);
697 
698 	if (refill_required > refill_threshold) {
699 		ena_refill_rx_bufs(rx_ring, refill_required);
700 	}
701 
702 	tcp_lro_flush_all(&rx_ring->lro);
703 
704 	return (ENA_RX_BUDGET - budget);
705 }
706 
707 static void
ena_tx_csum(struct ena_com_tx_ctx * ena_tx_ctx,struct mbuf * mbuf,bool disable_meta_caching)708 ena_tx_csum(struct ena_com_tx_ctx *ena_tx_ctx, struct mbuf *mbuf,
709     bool disable_meta_caching)
710 {
711 	struct ena_com_tx_meta *ena_meta;
712 	struct ether_vlan_header *eh;
713 	struct mbuf *mbuf_next;
714 	u32 mss;
715 	bool offload;
716 	uint16_t etype;
717 	int ehdrlen;
718 	struct ip *ip;
719 	int ipproto;
720 	int iphlen;
721 	struct tcphdr *th;
722 	int offset;
723 
724 	offload = false;
725 	ena_meta = &ena_tx_ctx->ena_meta;
726 	mss = mbuf->m_pkthdr.tso_segsz;
727 
728 	if (mss != 0)
729 		offload = true;
730 
731 	if ((mbuf->m_pkthdr.csum_flags & CSUM_TSO) != 0)
732 		offload = true;
733 
734 	if ((mbuf->m_pkthdr.csum_flags & CSUM_OFFLOAD) != 0)
735 		offload = true;
736 
737 	if ((mbuf->m_pkthdr.csum_flags & CSUM6_OFFLOAD) != 0)
738 		offload = true;
739 
740 	if (!offload) {
741 		if (disable_meta_caching) {
742 			memset(ena_meta, 0, sizeof(*ena_meta));
743 			ena_tx_ctx->meta_valid = 1;
744 		} else {
745 			ena_tx_ctx->meta_valid = 0;
746 		}
747 		return;
748 	}
749 
750 	/* Determine where frame payload starts. */
751 	eh = mtod(mbuf, struct ether_vlan_header *);
752 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
753 		etype = ntohs(eh->evl_proto);
754 		ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
755 	} else {
756 		etype = ntohs(eh->evl_encap_proto);
757 		ehdrlen = ETHER_HDR_LEN;
758 	}
759 
760 	mbuf_next = m_getptr(mbuf, ehdrlen, &offset);
761 
762 	switch (etype) {
763 	case ETHERTYPE_IP:
764 		ip = (struct ip *)(mtodo(mbuf_next, offset));
765 		iphlen = ip->ip_hl << 2;
766 		ipproto = ip->ip_p;
767 		ena_tx_ctx->l3_proto = ENA_ETH_IO_L3_PROTO_IPV4;
768 		if ((ip->ip_off & htons(IP_DF)) != 0)
769 			ena_tx_ctx->df = 1;
770 		break;
771 	case ETHERTYPE_IPV6:
772 		ena_tx_ctx->l3_proto = ENA_ETH_IO_L3_PROTO_IPV6;
773 		iphlen = ip6_lasthdr(mbuf, ehdrlen, IPPROTO_IPV6, &ipproto);
774 		iphlen -= ehdrlen;
775 		ena_tx_ctx->df = 1;
776 		break;
777 	default:
778 		iphlen = 0;
779 		ipproto = 0;
780 		break;
781 	}
782 
783 	mbuf_next = m_getptr(mbuf, iphlen + ehdrlen, &offset);
784 	th = (struct tcphdr *)(mtodo(mbuf_next, offset));
785 
786 	if ((mbuf->m_pkthdr.csum_flags & CSUM_IP) != 0) {
787 		ena_tx_ctx->l3_csum_enable = 1;
788 	}
789 	if ((mbuf->m_pkthdr.csum_flags & CSUM_TSO) != 0) {
790 		ena_tx_ctx->tso_enable = 1;
791 		ena_meta->l4_hdr_len = (th->th_off);
792 	}
793 
794 	if (ipproto == IPPROTO_TCP) {
795 		ena_tx_ctx->l4_proto = ENA_ETH_IO_L4_PROTO_TCP;
796 		if ((mbuf->m_pkthdr.csum_flags &
797 		    (CSUM_IP_TCP | CSUM_IP6_TCP)) != 0)
798 			ena_tx_ctx->l4_csum_enable = 1;
799 		else
800 			ena_tx_ctx->l4_csum_enable = 0;
801 	} else if (ipproto == IPPROTO_UDP) {
802 		ena_tx_ctx->l4_proto = ENA_ETH_IO_L4_PROTO_UDP;
803 		if ((mbuf->m_pkthdr.csum_flags &
804 		    (CSUM_IP_UDP | CSUM_IP6_UDP)) != 0)
805 			ena_tx_ctx->l4_csum_enable = 1;
806 		else
807 			ena_tx_ctx->l4_csum_enable = 0;
808 	} else {
809 		ena_tx_ctx->l4_proto = ENA_ETH_IO_L4_PROTO_UNKNOWN;
810 		ena_tx_ctx->l4_csum_enable = 0;
811 	}
812 
813 	ena_meta->mss = mss;
814 	ena_meta->l3_hdr_len = iphlen;
815 	ena_meta->l3_hdr_offset = ehdrlen;
816 	ena_tx_ctx->meta_valid = 1;
817 }
818 
819 static int
ena_check_and_collapse_mbuf(struct ena_ring * tx_ring,struct mbuf ** mbuf)820 ena_check_and_collapse_mbuf(struct ena_ring *tx_ring, struct mbuf **mbuf)
821 {
822 	struct ena_adapter *adapter;
823 	struct mbuf *collapsed_mbuf;
824 	int num_frags;
825 
826 	adapter = tx_ring->adapter;
827 	num_frags = ena_mbuf_count(*mbuf);
828 
829 	/* One segment must be reserved for configuration descriptor. */
830 	if (num_frags < adapter->max_tx_sgl_size)
831 		return (0);
832 
833 	if ((num_frags == adapter->max_tx_sgl_size) &&
834 	    ((*mbuf)->m_pkthdr.len < tx_ring->tx_max_header_size))
835 		return (0);
836 
837 	counter_u64_add(tx_ring->tx_stats.collapse, 1);
838 
839 	collapsed_mbuf = m_collapse(*mbuf, M_NOWAIT,
840 	    adapter->max_tx_sgl_size - 1);
841 	if (unlikely(collapsed_mbuf == NULL)) {
842 		counter_u64_add(tx_ring->tx_stats.collapse_err, 1);
843 		return (ENOMEM);
844 	}
845 
846 	/* If mbuf was collapsed succesfully, original mbuf is released. */
847 	*mbuf = collapsed_mbuf;
848 
849 	return (0);
850 }
851 
852 static int
ena_tx_map_mbuf(struct ena_ring * tx_ring,struct ena_tx_buffer * tx_info,struct mbuf * mbuf,void ** push_hdr,u16 * header_len)853 ena_tx_map_mbuf(struct ena_ring *tx_ring, struct ena_tx_buffer *tx_info,
854     struct mbuf *mbuf, void **push_hdr, u16 *header_len)
855 {
856 	struct ena_adapter *adapter = tx_ring->adapter;
857 	struct ena_com_buf *ena_buf;
858 	bus_dma_segment_t segs[ENA_BUS_DMA_SEGS];
859 	size_t iseg = 0;
860 	uint32_t mbuf_head_len;
861 	uint16_t offset;
862 	int rc, nsegs;
863 
864 	mbuf_head_len = mbuf->m_len;
865 	tx_info->mbuf = mbuf;
866 	ena_buf = tx_info->bufs;
867 
868 	/*
869 	 * For easier maintaining of the DMA map, map the whole mbuf even if
870 	 * the LLQ is used. The descriptors will be filled using the segments.
871 	 */
872 	rc = bus_dmamap_load_mbuf_sg(adapter->tx_buf_tag,
873 	    tx_info->dmamap, mbuf, segs, &nsegs, BUS_DMA_NOWAIT);
874 	if (unlikely((rc != 0) || (nsegs == 0))) {
875 		ena_log_io(adapter->pdev, WARN,
876 		    "dmamap load failed! err: %d nsegs: %d\n", rc, nsegs);
877 		goto dma_error;
878 	}
879 
880 	if (tx_ring->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) {
881 		/*
882 		 * When the device is LLQ mode, the driver will copy
883 		 * the header into the device memory space.
884 		 * the ena_com layer assumes the header is in a linear
885 		 * memory space.
886 		 * This assumption might be wrong since part of the header
887 		 * can be in the fragmented buffers.
888 		 * First check if header fits in the mbuf. If not, copy it to
889 		 * separate buffer that will be holding linearized data.
890 		 */
891 		*header_len = min_t(uint32_t, mbuf->m_pkthdr.len,
892 		    tx_ring->tx_max_header_size);
893 
894 		/* If header is in linear space, just point into mbuf's data. */
895 		if (likely(*header_len <= mbuf_head_len)) {
896 			*push_hdr = mbuf->m_data;
897 		/*
898 		 * Otherwise, copy whole portion of header from multiple
899 		 * mbufs to intermediate buffer.
900 		 */
901 		} else {
902 			m_copydata(mbuf, 0, *header_len,
903 			    tx_ring->push_buf_intermediate_buf);
904 			*push_hdr = tx_ring->push_buf_intermediate_buf;
905 
906 			counter_u64_add(tx_ring->tx_stats.llq_buffer_copy, 1);
907 		}
908 
909 		ena_log_io(adapter->pdev, DBG,
910 		    "mbuf: %p header_buf->vaddr: %p push_len: %d\n",
911 		    mbuf, *push_hdr, *header_len);
912 
913 		/* If packet is fitted in LLQ header, no need for DMA segments. */
914 		if (mbuf->m_pkthdr.len <= tx_ring->tx_max_header_size) {
915 			return (0);
916 		} else {
917 			offset = tx_ring->tx_max_header_size;
918 			/*
919 			 * As Header part is mapped to LLQ header, we can skip
920 			 * it and just map the residuum of the mbuf to DMA
921 			 * Segments.
922 			 */
923 			while (offset > 0) {
924 				if (offset >= segs[iseg].ds_len) {
925 					offset -= segs[iseg].ds_len;
926 				} else {
927 					ena_buf->paddr = segs[iseg].ds_addr +
928 					    offset;
929 					ena_buf->len = segs[iseg].ds_len -
930 					    offset;
931 					ena_buf++;
932 					tx_info->num_of_bufs++;
933 					offset = 0;
934 				}
935 				iseg++;
936 			}
937 		}
938 	} else {
939 		*push_hdr = NULL;
940 		/*
941 		 * header_len is just a hint for the device. Because FreeBSD is
942 		 * not giving us information about packet header length and it
943 		 * is not guaranteed that all packet headers will be in the 1st
944 		 * mbuf, setting header_len to 0 is making the device ignore
945 		 * this value and resolve header on it's own.
946 		 */
947 		*header_len = 0;
948 	}
949 
950 	/* Map rest of the mbuf */
951 	while (iseg < nsegs) {
952 		ena_buf->paddr = segs[iseg].ds_addr;
953 		ena_buf->len = segs[iseg].ds_len;
954 		ena_buf++;
955 		iseg++;
956 		tx_info->num_of_bufs++;
957 	}
958 
959 	return (0);
960 
961 dma_error:
962 	counter_u64_add(tx_ring->tx_stats.dma_mapping_err, 1);
963 	tx_info->mbuf = NULL;
964 	return (rc);
965 }
966 
967 static int
ena_xmit_mbuf(struct ena_ring * tx_ring,struct mbuf ** mbuf)968 ena_xmit_mbuf(struct ena_ring *tx_ring, struct mbuf **mbuf)
969 {
970 	struct ena_adapter *adapter;
971 	device_t pdev;
972 	struct ena_tx_buffer *tx_info;
973 	struct ena_com_tx_ctx ena_tx_ctx;
974 	struct ena_com_dev *ena_dev;
975 	struct ena_com_io_sq *io_sq;
976 	void *push_hdr;
977 	uint16_t next_to_use;
978 	uint16_t req_id;
979 	uint16_t ena_qid;
980 	uint16_t header_len;
981 	int rc;
982 	int nb_hw_desc;
983 
984 	ena_qid = ENA_IO_TXQ_IDX(tx_ring->que->id);
985 	adapter = tx_ring->que->adapter;
986 	pdev = adapter->pdev;
987 	ena_dev = adapter->ena_dev;
988 	io_sq = &ena_dev->io_sq_queues[ena_qid];
989 
990 	rc = ena_check_and_collapse_mbuf(tx_ring, mbuf);
991 	if (unlikely(rc != 0)) {
992 		ena_log_io(pdev, WARN, "Failed to collapse mbuf! err: %d\n",
993 		    rc);
994 		return (rc);
995 	}
996 
997 	ena_log_io(pdev, DBG, "Tx: %d bytes\n", (*mbuf)->m_pkthdr.len);
998 
999 	next_to_use = tx_ring->next_to_use;
1000 	req_id = tx_ring->free_tx_ids[next_to_use];
1001 	tx_info = &tx_ring->tx_buffer_info[req_id];
1002 	tx_info->num_of_bufs = 0;
1003 
1004 	ENA_WARN(tx_info->mbuf != NULL, adapter->ena_dev,
1005 	    "mbuf isn't NULL for req_id %d\n", req_id);
1006 
1007 	rc = ena_tx_map_mbuf(tx_ring, tx_info, *mbuf, &push_hdr, &header_len);
1008 	if (unlikely(rc != 0)) {
1009 		ena_log_io(pdev, WARN, "Failed to map TX mbuf\n");
1010 		return (rc);
1011 	}
1012 	memset(&ena_tx_ctx, 0x0, sizeof(struct ena_com_tx_ctx));
1013 	ena_tx_ctx.ena_bufs = tx_info->bufs;
1014 	ena_tx_ctx.push_header = push_hdr;
1015 	ena_tx_ctx.num_bufs = tx_info->num_of_bufs;
1016 	ena_tx_ctx.req_id = req_id;
1017 	ena_tx_ctx.header_len = header_len;
1018 
1019 	/* Set flags and meta data */
1020 	ena_tx_csum(&ena_tx_ctx, *mbuf, adapter->disable_meta_caching);
1021 
1022 	if (tx_ring->acum_pkts == ENA_DB_THRESHOLD ||
1023 	    ena_com_is_doorbell_needed(tx_ring->ena_com_io_sq, &ena_tx_ctx)) {
1024 		ena_log_io(pdev, DBG,
1025 		    "llq tx max burst size of queue %d achieved, writing doorbell to send burst\n",
1026 		    tx_ring->que->id);
1027 		ena_ring_tx_doorbell(tx_ring);
1028 	}
1029 
1030 	/* Prepare the packet's descriptors and send them to device */
1031 	rc = ena_com_prepare_tx(io_sq, &ena_tx_ctx, &nb_hw_desc);
1032 	if (unlikely(rc != 0)) {
1033 		if (likely(rc == ENA_COM_NO_MEM)) {
1034 			ena_log_io(pdev, DBG, "tx ring[%d] is out of space\n",
1035 			    tx_ring->que->id);
1036 		} else {
1037 			ena_log(pdev, ERR, "failed to prepare tx bufs\n");
1038 			ena_trigger_reset(adapter,
1039 			    ENA_REGS_RESET_DRIVER_INVALID_STATE);
1040 		}
1041 		counter_u64_add(tx_ring->tx_stats.prepare_ctx_err, 1);
1042 		goto dma_error;
1043 	}
1044 
1045 	counter_enter();
1046 	counter_u64_add_protected(tx_ring->tx_stats.cnt, 1);
1047 	counter_u64_add_protected(tx_ring->tx_stats.bytes,
1048 	    (*mbuf)->m_pkthdr.len);
1049 
1050 	counter_u64_add_protected(adapter->hw_stats.tx_packets, 1);
1051 	counter_u64_add_protected(adapter->hw_stats.tx_bytes,
1052 	    (*mbuf)->m_pkthdr.len);
1053 	counter_exit();
1054 
1055 	tx_info->tx_descs = nb_hw_desc;
1056 	getbinuptime(&tx_info->timestamp);
1057 	tx_info->print_once = true;
1058 
1059 	tx_ring->next_to_use = ENA_TX_RING_IDX_NEXT(next_to_use,
1060 	    tx_ring->ring_size);
1061 
1062 	/* stop the queue when no more space available, the packet can have up
1063 	 * to sgl_size + 2. one for the meta descriptor and one for header
1064 	 * (if the header is larger than tx_max_header_size).
1065 	 */
1066 	if (unlikely(!ena_com_sq_have_enough_space(tx_ring->ena_com_io_sq,
1067 	    adapter->max_tx_sgl_size + 2))) {
1068 		ena_log_io(pdev, DBG, "Stop queue %d\n", tx_ring->que->id);
1069 
1070 		tx_ring->running = false;
1071 		counter_u64_add(tx_ring->tx_stats.queue_stop, 1);
1072 
1073 		/* There is a rare condition where this function decides to
1074 		 * stop the queue but meanwhile tx_cleanup() updates
1075 		 * next_to_completion and terminates.
1076 		 * The queue will remain stopped forever.
1077 		 * To solve this issue this function performs mb(), checks
1078 		 * the wakeup condition and wakes up the queue if needed.
1079 		 */
1080 		mb();
1081 
1082 		if (ena_com_sq_have_enough_space(tx_ring->ena_com_io_sq,
1083 		    ENA_TX_RESUME_THRESH)) {
1084 			tx_ring->running = true;
1085 			counter_u64_add(tx_ring->tx_stats.queue_wakeup, 1);
1086 		}
1087 	}
1088 
1089 	bus_dmamap_sync(adapter->tx_buf_tag, tx_info->dmamap,
1090 	    BUS_DMASYNC_PREWRITE);
1091 
1092 	return (0);
1093 
1094 dma_error:
1095 	tx_info->mbuf = NULL;
1096 	bus_dmamap_unload(adapter->tx_buf_tag, tx_info->dmamap);
1097 
1098 	return (rc);
1099 }
1100 
1101 static void
ena_start_xmit(struct ena_ring * tx_ring)1102 ena_start_xmit(struct ena_ring *tx_ring)
1103 {
1104 	struct mbuf *mbuf;
1105 	struct ena_adapter *adapter = tx_ring->adapter;
1106 	int ret = 0;
1107 
1108 	ENA_RING_MTX_ASSERT(tx_ring);
1109 
1110 	if (unlikely((if_getdrvflags(adapter->ifp) & IFF_DRV_RUNNING) == 0))
1111 		return;
1112 
1113 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, adapter)))
1114 		return;
1115 
1116 	while ((mbuf = drbr_peek(adapter->ifp, tx_ring->br)) != NULL) {
1117 		ena_log_io(adapter->pdev, DBG,
1118 		    "\ndequeued mbuf %p with flags %#x and header csum flags %#jx\n",
1119 		    mbuf, mbuf->m_flags, (uint64_t)mbuf->m_pkthdr.csum_flags);
1120 
1121 		if (unlikely(!tx_ring->running)) {
1122 			drbr_putback(adapter->ifp, tx_ring->br, mbuf);
1123 			break;
1124 		}
1125 
1126 		if (unlikely((ret = ena_xmit_mbuf(tx_ring, &mbuf)) != 0)) {
1127 			if (ret == ENA_COM_NO_MEM) {
1128 				drbr_putback(adapter->ifp, tx_ring->br, mbuf);
1129 			} else if (ret == ENA_COM_NO_SPACE) {
1130 				drbr_putback(adapter->ifp, tx_ring->br, mbuf);
1131 			} else {
1132 				m_freem(mbuf);
1133 				drbr_advance(adapter->ifp, tx_ring->br);
1134 			}
1135 
1136 			break;
1137 		}
1138 
1139 		drbr_advance(adapter->ifp, tx_ring->br);
1140 
1141 		if (unlikely((if_getdrvflags(adapter->ifp) & IFF_DRV_RUNNING) == 0))
1142 			return;
1143 
1144 		tx_ring->acum_pkts++;
1145 
1146 		BPF_MTAP(adapter->ifp, mbuf);
1147 	}
1148 
1149 	if (likely(tx_ring->acum_pkts != 0)) {
1150 		/* Trigger the dma engine */
1151 		ena_ring_tx_doorbell(tx_ring);
1152 	}
1153 
1154 	if (unlikely(!tx_ring->running))
1155 		taskqueue_enqueue(tx_ring->que->cleanup_tq,
1156 		    &tx_ring->que->cleanup_task);
1157 }
1158