1 /*-
2  * Copyright (c) 2015 Mellanox Technologies. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  * $FreeBSD: stable/10/sys/dev/mlx5/mlx5_en/mlx5_en_tx.c 362313 2020-06-18 10:41:51Z hselasky $
26  */
27 
28 #include "en.h"
29 #include <machine/atomic.h>
30 
31 static inline bool
mlx5e_do_send_cqe(struct mlx5e_sq * sq)32 mlx5e_do_send_cqe(struct mlx5e_sq *sq)
33 {
34 	sq->cev_counter++;
35 	/* interleave the CQEs */
36 	if (sq->cev_counter >= sq->cev_factor) {
37 		sq->cev_counter = 0;
38 		return (1);
39 	}
40 	return (0);
41 }
42 
43 void
mlx5e_send_nop(struct mlx5e_sq * sq,u32 ds_cnt)44 mlx5e_send_nop(struct mlx5e_sq *sq, u32 ds_cnt)
45 {
46 	u16 pi = sq->pc & sq->wq.sz_m1;
47 	struct mlx5e_tx_wqe *wqe = mlx5_wq_cyc_get_wqe(&sq->wq, pi);
48 
49 	memset(&wqe->ctrl, 0, sizeof(wqe->ctrl));
50 
51 	wqe->ctrl.opmod_idx_opcode = cpu_to_be32((sq->pc << 8) | MLX5_OPCODE_NOP);
52 	wqe->ctrl.qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_cnt);
53 	if (mlx5e_do_send_cqe(sq))
54 		wqe->ctrl.fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE;
55 	else
56 		wqe->ctrl.fm_ce_se = 0;
57 
58 	/* Copy data for doorbell */
59 	memcpy(sq->doorbell.d32, &wqe->ctrl, sizeof(sq->doorbell.d32));
60 
61 	sq->mbuf[pi].mbuf = NULL;
62 	sq->mbuf[pi].num_bytes = 0;
63 	sq->mbuf[pi].num_wqebbs = DIV_ROUND_UP(ds_cnt, MLX5_SEND_WQEBB_NUM_DS);
64 	sq->pc += sq->mbuf[pi].num_wqebbs;
65 }
66 
67 #if (__FreeBSD_version >= 1100000)
68 static uint32_t mlx5e_hash_value;
69 
70 static void
mlx5e_hash_init(void * arg)71 mlx5e_hash_init(void *arg)
72 {
73 	mlx5e_hash_value = m_ether_tcpip_hash_init();
74 }
75 
76 /* Make kernel call mlx5e_hash_init after the random stack finished initializing */
77 SYSINIT(mlx5e_hash_init, SI_SUB_RANDOM, SI_ORDER_ANY, &mlx5e_hash_init, NULL);
78 #endif
79 
80 static struct mlx5e_sq *
mlx5e_select_queue(struct ifnet * ifp,struct mbuf * mb)81 mlx5e_select_queue(struct ifnet *ifp, struct mbuf *mb)
82 {
83 	struct mlx5e_priv *priv = ifp->if_softc;
84 	struct mlx5e_channel * volatile *ppch;
85 	struct mlx5e_channel *pch;
86 	u32 ch;
87 	u32 tc;
88 
89 	ppch = priv->channel;
90 
91 	/* check if channels are successfully opened */
92 	if (unlikely(ppch == NULL))
93 		return (NULL);
94 
95 	/* obtain VLAN information if present */
96 	if (mb->m_flags & M_VLANTAG) {
97 		tc = (mb->m_pkthdr.ether_vtag >> 13);
98 		if (tc >= priv->num_tc)
99 			tc = priv->default_vlan_prio;
100 	} else {
101 		tc = priv->default_vlan_prio;
102 	}
103 
104 	ch = priv->params.num_channels;
105 
106 	/* check if flowid is set */
107 	if (M_HASHTYPE_GET(mb) != M_HASHTYPE_NONE) {
108 #ifdef RSS
109 		u32 temp;
110 
111 		if (rss_hash2bucket(mb->m_pkthdr.flowid,
112 		    M_HASHTYPE_GET(mb), &temp) == 0)
113 			ch = temp % ch;
114 		else
115 #endif
116 			ch = (mb->m_pkthdr.flowid % 128) % ch;
117 	} else {
118 #if (__FreeBSD_version >= 1100000)
119 		ch = m_ether_tcpip_hash(MBUF_HASHFLAG_L3 |
120 		    MBUF_HASHFLAG_L4, mb, mlx5e_hash_value) % ch;
121 #else
122 		/*
123 		 * m_ether_tcpip_hash not present in stable, so just
124 		 * throw unhashed mbufs on queue 0
125 		 */
126 		ch = 0;
127 #endif
128 	}
129 
130 	/* check if channel is allocated and not stopped */
131 	pch = ppch[ch];
132 	if (likely(pch != NULL && pch->sq[tc].stopped == 0))
133 		return (&pch->sq[tc]);
134 	return (NULL);
135 }
136 
137 static inline u16
mlx5e_get_inline_hdr_size(struct mlx5e_sq * sq,struct mbuf * mb)138 mlx5e_get_inline_hdr_size(struct mlx5e_sq *sq, struct mbuf *mb)
139 {
140 
141 	switch(sq->min_inline_mode) {
142 	case MLX5_INLINE_MODE_NONE:
143 		/*
144 		 * When inline mode is NONE, we do not need to copy
145 		 * headers into WQEs, except when vlan tag framing is
146 		 * requested. Hardware might offload vlan tagging on
147 		 * transmit. This is a separate capability, which is
148 		 * known to be disabled on ConnectX-5 due to a hardware
149 		 * bug RM 931383. If vlan_inline_cap is not present and
150 		 * the packet has vlan tag, fall back to inlining.
151 		 */
152 		if ((mb->m_flags & M_VLANTAG) != 0 &&
153 		    sq->vlan_inline_cap == 0)
154 			break;
155 		return (0);
156 	case MLX5_INLINE_MODE_L2:
157 		/*
158 		 * Due to hardware limitations, when trust mode is
159 		 * DSCP, the hardware may request MLX5_INLINE_MODE_L2
160 		 * while it really needs all L2 headers and the 4 first
161 		 * bytes of the IP header (which include the
162 		 * TOS/traffic-class).
163 		 *
164 		 * To avoid doing a firmware command for querying the
165 		 * trust state and parsing the mbuf for doing
166 		 * unnecessary checks (VLAN/eth_type) in the fast path,
167 		 * we are going for the worth case (22 Bytes) if
168 		 * the mb->m_pkthdr.len allows it.
169 		 */
170 		if (mb->m_pkthdr.len > ETHER_HDR_LEN +
171 		    ETHER_VLAN_ENCAP_LEN + 4)
172 			return (MIN(sq->max_inline, ETHER_HDR_LEN +
173 			    ETHER_VLAN_ENCAP_LEN + 4));
174 		break;
175 	}
176 	return (MIN(sq->max_inline, mb->m_pkthdr.len));
177 }
178 
179 /*
180  * This function parse IPv4 and IPv6 packets looking for TCP and UDP
181  * headers.
182  *
183  * The return value indicates the number of bytes from the beginning
184  * of the packet until the first byte after the TCP or UDP header. If
185  * this function returns zero, the parsing failed.
186  */
187 static int
mlx5e_get_header_size(const struct mbuf * mb)188 mlx5e_get_header_size(const struct mbuf *mb)
189 {
190 	const struct ether_vlan_header *eh;
191 	const struct tcphdr *th;
192 	const struct ip *ip;
193 	int ip_hlen, tcp_hlen;
194 	const struct ip6_hdr *ip6;
195 	uint16_t eth_type;
196 	int eth_hdr_len;
197 
198 	eh = mtod(mb, const struct ether_vlan_header *);
199 	if (unlikely(mb->m_len < ETHER_HDR_LEN))
200 		return (0);
201 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
202 		if (unlikely(mb->m_len < (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN)))
203 			return (0);
204 		eth_type = ntohs(eh->evl_proto);
205 		eth_hdr_len = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
206 	} else {
207 		eth_type = ntohs(eh->evl_encap_proto);
208 		eth_hdr_len = ETHER_HDR_LEN;
209 	}
210 	switch (eth_type) {
211 	case ETHERTYPE_IP:
212 		ip = (const struct ip *)(mb->m_data + eth_hdr_len);
213 		if (unlikely(mb->m_len < eth_hdr_len + sizeof(*ip)))
214 			return (0);
215 		if (ip->ip_p != IPPROTO_TCP)
216 			return (0);
217 		ip_hlen = ip->ip_hl << 2;
218 		eth_hdr_len += ip_hlen;
219 		break;
220 	case ETHERTYPE_IPV6:
221 		ip6 = (const struct ip6_hdr *)(mb->m_data + eth_hdr_len);
222 		if (unlikely(mb->m_len < eth_hdr_len + sizeof(*ip6)))
223 			return (0);
224 		if (ip6->ip6_nxt != IPPROTO_TCP)
225 			return (0);
226 		eth_hdr_len += sizeof(*ip6);
227 		break;
228 	default:
229 		return (0);
230 	}
231 	if (unlikely(mb->m_len < eth_hdr_len + sizeof(*th))) {
232 		const struct mbuf *m_th = mb->m_next;
233 		if (unlikely(mb->m_len != eth_hdr_len ||
234 		    m_th == NULL || m_th->m_len < sizeof(*th)))
235 			return (0);
236 		th = (const struct tcphdr *)(m_th->m_data);
237 	} else {
238 		th = (const struct tcphdr *)(mb->m_data + eth_hdr_len);
239 	}
240 	tcp_hlen = th->th_off << 2;
241 	eth_hdr_len += tcp_hlen;
242 	/*
243 	 * m_copydata() will be used on the remaining header which
244 	 * does not need to reside within the first m_len bytes of
245 	 * data:
246 	 */
247 	if (unlikely(mb->m_pkthdr.len < eth_hdr_len))
248 		return (0);
249 	return (eth_hdr_len);
250 }
251 
252 /*
253  * The return value is not going back to the stack because of
254  * the drbr
255  */
256 static int
mlx5e_sq_xmit(struct mlx5e_sq * sq,struct mbuf ** mbp)257 mlx5e_sq_xmit(struct mlx5e_sq *sq, struct mbuf **mbp)
258 {
259 	bus_dma_segment_t segs[MLX5E_MAX_TX_MBUF_FRAGS];
260 	struct mlx5_wqe_data_seg *dseg;
261 	struct mlx5e_tx_wqe *wqe;
262 	struct ifnet *ifp;
263 	int nsegs;
264 	int err;
265 	int x;
266 	struct mbuf *mb = *mbp;
267 	u16 ds_cnt;
268 	u16 ihs;
269 	u16 pi;
270 	u8 opcode;
271 
272 	/*
273 	 * Return ENOBUFS if the queue is full, this may trigger reinsertion
274 	 * of the mbuf into the drbr (see mlx5e_xmit_locked)
275 	 */
276 	if (unlikely(!mlx5e_sq_has_room_for(sq, 2 * MLX5_SEND_WQE_MAX_WQEBBS))) {
277 		sq->stats.enobuf++;
278 		return (ENOBUFS);
279 	}
280 
281 	/* Align SQ edge with NOPs to avoid WQE wrap around */
282 	pi = ((~sq->pc) & sq->wq.sz_m1);
283 	if (pi < (MLX5_SEND_WQE_MAX_WQEBBS - 1)) {
284 		/* Send one multi NOP message instead of many */
285 		mlx5e_send_nop(sq, (pi + 1) * MLX5_SEND_WQEBB_NUM_DS);
286 		pi = ((~sq->pc) & sq->wq.sz_m1);
287 		if (pi < (MLX5_SEND_WQE_MAX_WQEBBS - 1)) {
288 			sq->stats.enobuf++;
289 			return (ENOMEM);
290 		}
291 	}
292 
293 	/* Setup local variables */
294 	pi = sq->pc & sq->wq.sz_m1;
295 	wqe = mlx5_wq_cyc_get_wqe(&sq->wq, pi);
296 	ifp = sq->ifp;
297 
298 	memset(wqe, 0, sizeof(*wqe));
299 
300 	/* Send a copy of the frame to the BPF listener, if any */
301 	if (ifp != NULL && ifp->if_bpf != NULL)
302 		ETHER_BPF_MTAP(ifp, mb);
303 
304 	if (mb->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO)) {
305 		wqe->eth.cs_flags |= MLX5_ETH_WQE_L3_CSUM;
306 	}
307 	if (mb->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO)) {
308 		wqe->eth.cs_flags |= MLX5_ETH_WQE_L4_CSUM;
309 	}
310 	if (wqe->eth.cs_flags == 0) {
311 		sq->stats.csum_offload_none++;
312 	}
313 	if (mb->m_pkthdr.csum_flags & CSUM_TSO) {
314 		u32 payload_len;
315 		u32 mss = mb->m_pkthdr.tso_segsz;
316 		u32 num_pkts;
317 
318 		wqe->eth.mss = cpu_to_be16(mss);
319 		opcode = MLX5_OPCODE_LSO;
320 		ihs = mlx5e_get_header_size(mb);
321 		payload_len = mb->m_pkthdr.len - ihs;
322 		if (payload_len == 0)
323 			num_pkts = 1;
324 		else
325 			num_pkts = DIV_ROUND_UP(payload_len, mss);
326 		sq->mbuf[pi].num_bytes = payload_len + (num_pkts * ihs);
327 
328 		sq->stats.tso_packets++;
329 		sq->stats.tso_bytes += payload_len;
330 	} else {
331 		opcode = MLX5_OPCODE_SEND;
332 		ihs = mlx5e_get_inline_hdr_size(sq, mb);
333 		sq->mbuf[pi].num_bytes = max_t (unsigned int,
334 		    mb->m_pkthdr.len, ETHER_MIN_LEN - ETHER_CRC_LEN);
335 	}
336 	if (ihs == 0) {
337 		if ((mb->m_flags & M_VLANTAG) != 0) {
338 			wqe->eth.vlan_cmd = htons(0x8000); /* bit 0 CVLAN */
339 			wqe->eth.vlan_hdr = htons(mb->m_pkthdr.ether_vtag);
340 		} else {
341 			wqe->eth.inline_hdr_sz = 0;
342 		}
343 	} else {
344 		if ((mb->m_flags & M_VLANTAG) != 0) {
345 			struct ether_vlan_header *eh = (struct ether_vlan_header
346 			    *)wqe->eth.inline_hdr_start;
347 
348 			/* Range checks */
349 			if (ihs > (MLX5E_MAX_TX_INLINE - ETHER_VLAN_ENCAP_LEN))
350 				ihs = (MLX5E_MAX_TX_INLINE -
351 				    ETHER_VLAN_ENCAP_LEN);
352 			else if (ihs < ETHER_HDR_LEN) {
353 				err = EINVAL;
354 				goto tx_drop;
355 			}
356 			m_copydata(mb, 0, ETHER_HDR_LEN, (caddr_t)eh);
357 			m_adj(mb, ETHER_HDR_LEN);
358 			/* Insert 4 bytes VLAN tag into data stream */
359 			eh->evl_proto = eh->evl_encap_proto;
360 			eh->evl_encap_proto = htons(ETHERTYPE_VLAN);
361 			eh->evl_tag = htons(mb->m_pkthdr.ether_vtag);
362 			/* Copy rest of header data, if any */
363 			m_copydata(mb, 0, ihs - ETHER_HDR_LEN, (caddr_t)(eh +
364 			    1));
365 			m_adj(mb, ihs - ETHER_HDR_LEN);
366 			/* Extend header by 4 bytes */
367 			ihs += ETHER_VLAN_ENCAP_LEN;
368 		} else {
369 			m_copydata(mb, 0, ihs, wqe->eth.inline_hdr_start);
370 			m_adj(mb, ihs);
371 		}
372 		wqe->eth.inline_hdr_sz = cpu_to_be16(ihs);
373 	}
374 
375 	ds_cnt = sizeof(*wqe) / MLX5_SEND_WQE_DS;
376 	if (ihs > sizeof(wqe->eth.inline_hdr_start)) {
377 		ds_cnt += DIV_ROUND_UP(ihs - sizeof(wqe->eth.inline_hdr_start),
378 		    MLX5_SEND_WQE_DS);
379 	}
380 	dseg = ((struct mlx5_wqe_data_seg *)&wqe->ctrl) + ds_cnt;
381 
382 	/* Trim off empty mbufs */
383 	while (mb->m_len == 0) {
384 		mb = m_free(mb);
385 		/* Check if all data has been inlined */
386 		if (mb == NULL)
387 			goto skip_dma;
388 	}
389 
390 	err = bus_dmamap_load_mbuf_sg(sq->dma_tag, sq->mbuf[pi].dma_map,
391 	    mb, segs, &nsegs, BUS_DMA_NOWAIT);
392 	if (err == EFBIG) {
393 		/*
394 		 * Update *mbp before defrag in case it was trimmed in the
395 		 * loop above
396 		 */
397 		*mbp = mb;
398 		/* Update statistics */
399 		sq->stats.defragged++;
400 		/* Too many mbuf fragments */
401 		mb = m_defrag(*mbp, M_NOWAIT);
402 		if (mb == NULL) {
403 			mb = *mbp;
404 			goto tx_drop;
405 		}
406 		/* Try again */
407 		err = bus_dmamap_load_mbuf_sg(sq->dma_tag, sq->mbuf[pi].dma_map,
408 		    mb, segs, &nsegs, BUS_DMA_NOWAIT);
409 	}
410 	/* Catch errors */
411 	if (err != 0)
412 		goto tx_drop;
413 
414 	for (x = 0; x != nsegs; x++) {
415 		if (segs[x].ds_len == 0)
416 			continue;
417 		dseg->addr = cpu_to_be64((uint64_t)segs[x].ds_addr);
418 		dseg->lkey = sq->mkey_be;
419 		dseg->byte_count = cpu_to_be32((uint32_t)segs[x].ds_len);
420 		dseg++;
421 	}
422 skip_dma:
423 	ds_cnt = (dseg - ((struct mlx5_wqe_data_seg *)&wqe->ctrl));
424 
425 	wqe->ctrl.opmod_idx_opcode = cpu_to_be32((sq->pc << 8) | opcode);
426 	wqe->ctrl.qpn_ds = cpu_to_be32((sq->sqn << 8) | ds_cnt);
427 	if (mlx5e_do_send_cqe(sq))
428 		wqe->ctrl.fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE;
429 	else
430 		wqe->ctrl.fm_ce_se = 0;
431 
432 	/* Copy data for doorbell */
433 	memcpy(sq->doorbell.d32, &wqe->ctrl, sizeof(sq->doorbell.d32));
434 
435 	/* Store pointer to mbuf */
436 	sq->mbuf[pi].mbuf = mb;
437 	sq->mbuf[pi].num_wqebbs = DIV_ROUND_UP(ds_cnt, MLX5_SEND_WQEBB_NUM_DS);
438 	sq->pc += sq->mbuf[pi].num_wqebbs;
439 
440 	/* Make sure all mbuf data is written to RAM */
441 	if (mb != NULL)
442 		bus_dmamap_sync(sq->dma_tag, sq->mbuf[pi].dma_map, BUS_DMASYNC_PREWRITE);
443 
444 	sq->stats.packets++;
445 	*mbp = NULL;	/* safety clear */
446 	return (0);
447 
448 tx_drop:
449 	sq->stats.dropped++;
450 	*mbp = NULL;
451 	m_freem(mb);
452 	return err;
453 }
454 
455 static void
mlx5e_poll_tx_cq(struct mlx5e_sq * sq,int budget)456 mlx5e_poll_tx_cq(struct mlx5e_sq *sq, int budget)
457 {
458 	u16 sqcc;
459 
460 	/*
461 	 * sq->cc must be updated only after mlx5_cqwq_update_db_record(),
462 	 * otherwise a cq overrun may occur
463 	 */
464 	sqcc = sq->cc;
465 
466 	while (budget > 0) {
467 		struct mlx5_cqe64 *cqe;
468 		struct mbuf *mb;
469 		u16 x;
470 		u16 ci;
471 
472 		cqe = mlx5e_get_cqe(&sq->cq);
473 		if (!cqe)
474 			break;
475 
476 		mlx5_cqwq_pop(&sq->cq.wq);
477 
478 		/* update budget according to the event factor */
479 		budget -= sq->cev_factor;
480 
481 		for (x = 0; x != sq->cev_factor; x++) {
482 			ci = sqcc & sq->wq.sz_m1;
483 			mb = sq->mbuf[ci].mbuf;
484 			sq->mbuf[ci].mbuf = NULL;	/* Safety clear */
485 
486 			if (mb == NULL) {
487 				if (sq->mbuf[ci].num_bytes == 0) {
488 					/* NOP */
489 					sq->stats.nop++;
490 				}
491 			} else {
492 				bus_dmamap_sync(sq->dma_tag, sq->mbuf[ci].dma_map,
493 				    BUS_DMASYNC_POSTWRITE);
494 				bus_dmamap_unload(sq->dma_tag, sq->mbuf[ci].dma_map);
495 
496 				/* Free transmitted mbuf */
497 				m_freem(mb);
498 			}
499 			sqcc += sq->mbuf[ci].num_wqebbs;
500 		}
501 	}
502 
503 	mlx5_cqwq_update_db_record(&sq->cq.wq);
504 
505 	/* Ensure cq space is freed before enabling more cqes */
506 	wmb();
507 
508 	sq->cc = sqcc;
509 
510 	if (sq->sq_tq != NULL &&
511 	    atomic_cmpset_int(&sq->queue_state, MLX5E_SQ_FULL, MLX5E_SQ_READY))
512 		taskqueue_enqueue(sq->sq_tq, &sq->sq_task);
513 }
514 
515 static int
mlx5e_xmit_locked(struct ifnet * ifp,struct mlx5e_sq * sq,struct mbuf * mb)516 mlx5e_xmit_locked(struct ifnet *ifp, struct mlx5e_sq *sq, struct mbuf *mb)
517 {
518 	struct mbuf *next;
519 	int err = 0;
520 
521 	if (likely(mb != NULL)) {
522 		/*
523 		 * If we can't insert mbuf into drbr, try to xmit anyway.
524 		 * We keep the error we got so we could return that after xmit.
525 		 */
526 		err = drbr_enqueue(ifp, sq->br, mb);
527 	}
528 
529 	/*
530 	 * Check if the network interface is closed or if the SQ is
531 	 * being stopped:
532 	 */
533 	if (unlikely((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
534 	    sq->stopped != 0))
535 		return (err);
536 
537 	/* Process the queue */
538 	while ((next = drbr_peek(ifp, sq->br)) != NULL) {
539 		if (mlx5e_sq_xmit(sq, &next) != 0) {
540 			if (next != NULL) {
541 				drbr_putback(ifp, sq->br, next);
542 				atomic_store_rel_int(&sq->queue_state, MLX5E_SQ_FULL);
543 				break;
544 			}
545 		}
546 		drbr_advance(ifp, sq->br);
547 	}
548 	/* Check if we need to write the doorbell */
549 	if (likely(sq->doorbell.d64 != 0)) {
550 		mlx5e_tx_notify_hw(sq, sq->doorbell.d32, 0);
551 		sq->doorbell.d64 = 0;
552 	}
553 	/*
554 	 * Check if we need to start the event timer which flushes the
555 	 * transmit ring on timeout:
556 	 */
557 	if (unlikely(sq->cev_next_state == MLX5E_CEV_STATE_INITIAL &&
558 	    sq->cev_factor != 1)) {
559 		/* start the timer */
560 		mlx5e_sq_cev_timeout(sq);
561 	} else {
562 		/* don't send NOPs yet */
563 		sq->cev_next_state = MLX5E_CEV_STATE_HOLD_NOPS;
564 	}
565 	return (err);
566 }
567 
568 static int
mlx5e_xmit_locked_no_br(struct ifnet * ifp,struct mlx5e_sq * sq,struct mbuf * mb)569 mlx5e_xmit_locked_no_br(struct ifnet *ifp, struct mlx5e_sq *sq, struct mbuf *mb)
570 {
571 	int err = 0;
572 
573 	if (unlikely((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
574 	    sq->stopped != 0)) {
575 		m_freem(mb);
576 		return (ENETDOWN);
577 	}
578 
579 	/* Do transmit */
580 	if (mlx5e_sq_xmit(sq, &mb) != 0) {
581 		/* NOTE: m_freem() is NULL safe */
582 		m_freem(mb);
583 		err = ENOBUFS;
584 	}
585 
586 	/* Check if we need to write the doorbell */
587 	if (likely(sq->doorbell.d64 != 0)) {
588 		mlx5e_tx_notify_hw(sq, sq->doorbell.d32, 0);
589 		sq->doorbell.d64 = 0;
590 	}
591 
592 	/*
593 	 * Check if we need to start the event timer which flushes the
594 	 * transmit ring on timeout:
595 	 */
596 	if (unlikely(sq->cev_next_state == MLX5E_CEV_STATE_INITIAL &&
597 	    sq->cev_factor != 1)) {
598 		/* start the timer */
599 		mlx5e_sq_cev_timeout(sq);
600 	} else {
601 		/* don't send NOPs yet */
602 		sq->cev_next_state = MLX5E_CEV_STATE_HOLD_NOPS;
603 	}
604 	return (err);
605 }
606 
607 int
mlx5e_xmit(struct ifnet * ifp,struct mbuf * mb)608 mlx5e_xmit(struct ifnet *ifp, struct mbuf *mb)
609 {
610 	struct mlx5e_sq *sq;
611 	int ret;
612 
613 	sq = mlx5e_select_queue(ifp, mb);
614 	if (unlikely(sq == NULL)) {
615 		/* Invalid send queue */
616 		m_freem(mb);
617 		return (ENXIO);
618 	}
619 
620 	if (unlikely(sq->br == NULL)) {
621 		/* rate limited traffic */
622 		mtx_lock(&sq->lock);
623 		ret = mlx5e_xmit_locked_no_br(ifp, sq, mb);
624 		mtx_unlock(&sq->lock);
625 	} else if (mtx_trylock(&sq->lock)) {
626 		ret = mlx5e_xmit_locked(ifp, sq, mb);
627 		mtx_unlock(&sq->lock);
628 	} else {
629 		ret = drbr_enqueue(ifp, sq->br, mb);
630 		taskqueue_enqueue(sq->sq_tq, &sq->sq_task);
631 	}
632 
633 	return (ret);
634 }
635 
636 void
mlx5e_tx_cq_comp(struct mlx5_core_cq * mcq)637 mlx5e_tx_cq_comp(struct mlx5_core_cq *mcq)
638 {
639 	struct mlx5e_sq *sq = container_of(mcq, struct mlx5e_sq, cq.mcq);
640 
641 	mtx_lock(&sq->comp_lock);
642 	mlx5e_poll_tx_cq(sq, MLX5E_BUDGET_MAX);
643 	mlx5e_cq_arm(&sq->cq, MLX5_GET_DOORBELL_LOCK(&sq->priv->doorbell_lock));
644 	mtx_unlock(&sq->comp_lock);
645 }
646 
647 void
mlx5e_tx_que(void * context,int pending)648 mlx5e_tx_que(void *context, int pending)
649 {
650 	struct mlx5e_sq *sq = context;
651 	struct ifnet *ifp = sq->ifp;
652 
653 	if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
654 		mtx_lock(&sq->lock);
655 		if (!drbr_empty(ifp, sq->br))
656 			mlx5e_xmit_locked(ifp, sq, NULL);
657 		mtx_unlock(&sq->lock);
658 	}
659 }
660