xref: /freebsd/sys/dev/mlx4/mlx4_en/mlx4_en_tx.c (revision af23369a6deaaeb612ab266eb88b8bb8d560c322)
1 /*
2  * Copyright (c) 2007, 2014 Mellanox Technologies. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  */
33 
34 #define	LINUXKPI_PARAM_PREFIX mlx4_
35 
36 #include <linux/page.h>
37 #include <dev/mlx4/cq.h>
38 #include <linux/slab.h>
39 #include <dev/mlx4/qp.h>
40 #include <linux/if_vlan.h>
41 #include <linux/vmalloc.h>
42 #include <linux/moduleparam.h>
43 
44 #include <netinet/in_systm.h>
45 #include <netinet/in.h>
46 #include <netinet/if_ether.h>
47 #include <netinet/ip.h>
48 #include <netinet/ip6.h>
49 #include <netinet/tcp.h>
50 #include <netinet/tcp_lro.h>
51 #include <netinet/udp.h>
52 
53 #include "en.h"
54 
55 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
56 			   struct mlx4_en_tx_ring **pring, u32 size,
57 			   u16 stride, int node, int queue_idx)
58 {
59 	struct mlx4_en_dev *mdev = priv->mdev;
60 	struct mlx4_en_tx_ring *ring;
61 	uint32_t x;
62 	int tmp;
63 	int err;
64 
65 	ring = kzalloc_node(sizeof(struct mlx4_en_tx_ring), GFP_KERNEL, node);
66 	if (!ring) {
67 		ring = kzalloc(sizeof(struct mlx4_en_tx_ring), GFP_KERNEL);
68 		if (!ring) {
69 			en_err(priv, "Failed allocating TX ring\n");
70 			return -ENOMEM;
71 		}
72 	}
73 
74 	/* Create DMA descriptor TAG */
75 	if ((err = -bus_dma_tag_create(
76 	    bus_get_dma_tag(mdev->pdev->dev.bsddev),
77 	    1,					/* any alignment */
78 	    0,					/* no boundary */
79 	    BUS_SPACE_MAXADDR,			/* lowaddr */
80 	    BUS_SPACE_MAXADDR,			/* highaddr */
81 	    NULL, NULL,				/* filter, filterarg */
82 	    MLX4_EN_TX_MAX_PAYLOAD_SIZE,	/* maxsize */
83 	    MLX4_EN_TX_MAX_MBUF_FRAGS,		/* nsegments */
84 	    MLX4_EN_TX_MAX_MBUF_SIZE,		/* maxsegsize */
85 	    0,					/* flags */
86 	    NULL, NULL,				/* lockfunc, lockfuncarg */
87 	    &ring->dma_tag)))
88 		goto done;
89 
90 	ring->size = size;
91 	ring->size_mask = size - 1;
92 	ring->stride = stride;
93 	ring->inline_thold = MAX(MIN_PKT_LEN, MIN(priv->prof->inline_thold, MAX_INLINE));
94 	mtx_init(&ring->tx_lock.m, "mlx4 tx", NULL, MTX_DEF);
95 	mtx_init(&ring->comp_lock.m, "mlx4 comp", NULL, MTX_DEF);
96 
97 	tmp = size * sizeof(struct mlx4_en_tx_info);
98 	ring->tx_info = kzalloc_node(tmp, GFP_KERNEL, node);
99 	if (!ring->tx_info) {
100 		ring->tx_info = kzalloc(tmp, GFP_KERNEL);
101 		if (!ring->tx_info) {
102 			err = -ENOMEM;
103 			goto err_ring;
104 		}
105 	}
106 
107 	/* Create DMA descriptor MAPs */
108 	for (x = 0; x != size; x++) {
109 		err = -bus_dmamap_create(ring->dma_tag, 0,
110 		    &ring->tx_info[x].dma_map);
111 		if (err != 0) {
112 			while (x--) {
113 				bus_dmamap_destroy(ring->dma_tag,
114 				    ring->tx_info[x].dma_map);
115 			}
116 			goto err_info;
117 		}
118 	}
119 
120 	en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
121 		 ring->tx_info, tmp);
122 
123 	ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
124 
125 	/* Allocate HW buffers on provided NUMA node */
126 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size,
127 				 2 * PAGE_SIZE);
128 	if (err) {
129 		en_err(priv, "Failed allocating hwq resources\n");
130 		goto err_dma_map;
131 	}
132 
133 	err = mlx4_en_map_buffer(&ring->wqres.buf);
134 	if (err) {
135 		en_err(priv, "Failed to map TX buffer\n");
136 		goto err_hwq_res;
137 	}
138 
139 	ring->buf = ring->wqres.buf.direct.buf;
140 
141 	en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d "
142 	       "buf_size:%d dma:%llx\n", ring, ring->buf, ring->size,
143 	       ring->buf_size, (unsigned long long) ring->wqres.buf.direct.map);
144 
145 	err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &ring->qpn,
146 				    MLX4_RESERVE_ETH_BF_QP);
147 	if (err) {
148 		en_err(priv, "failed reserving qp for TX ring\n");
149 		goto err_map;
150 	}
151 
152 	err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp, GFP_KERNEL);
153 	if (err) {
154 		en_err(priv, "Failed allocating qp %d\n", ring->qpn);
155 		goto err_reserve;
156 	}
157 	ring->qp.event = mlx4_en_sqp_event;
158 
159 	err = mlx4_bf_alloc(mdev->dev, &ring->bf, node);
160 	if (err) {
161 		en_dbg(DRV, priv, "working without blueflame (%d)", err);
162 		ring->bf.uar = &mdev->priv_uar;
163 		ring->bf.uar->map = mdev->uar_map;
164 		ring->bf_enabled = false;
165 	} else
166 		ring->bf_enabled = true;
167 	ring->queue_index = queue_idx;
168 
169 	*pring = ring;
170 	return 0;
171 
172 err_reserve:
173 	mlx4_qp_release_range(mdev->dev, ring->qpn, 1);
174 err_map:
175 	mlx4_en_unmap_buffer(&ring->wqres.buf);
176 err_hwq_res:
177 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
178 err_dma_map:
179 	for (x = 0; x != size; x++)
180 		bus_dmamap_destroy(ring->dma_tag, ring->tx_info[x].dma_map);
181 err_info:
182 	vfree(ring->tx_info);
183 err_ring:
184 	bus_dma_tag_destroy(ring->dma_tag);
185 done:
186 	kfree(ring);
187 	return err;
188 }
189 
190 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
191 			     struct mlx4_en_tx_ring **pring)
192 {
193 	struct mlx4_en_dev *mdev = priv->mdev;
194 	struct mlx4_en_tx_ring *ring = *pring;
195 	uint32_t x;
196 	en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
197 
198 	if (ring->bf_enabled)
199 		mlx4_bf_free(mdev->dev, &ring->bf);
200 	mlx4_qp_remove(mdev->dev, &ring->qp);
201 	mlx4_qp_free(mdev->dev, &ring->qp);
202 	mlx4_qp_release_range(priv->mdev->dev, ring->qpn, 1);
203 	mlx4_en_unmap_buffer(&ring->wqres.buf);
204 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
205 	for (x = 0; x != ring->size; x++)
206 		bus_dmamap_destroy(ring->dma_tag, ring->tx_info[x].dma_map);
207 	vfree(ring->tx_info);
208 	mtx_destroy(&ring->tx_lock.m);
209 	mtx_destroy(&ring->comp_lock.m);
210 	bus_dma_tag_destroy(ring->dma_tag);
211 	kfree(ring);
212 	*pring = NULL;
213 }
214 
215 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
216 			     struct mlx4_en_tx_ring *ring,
217 			     int cq, int user_prio)
218 {
219 	struct mlx4_en_dev *mdev = priv->mdev;
220 	int err;
221 
222 	ring->cqn = cq;
223 	ring->prod = 0;
224 	ring->cons = 0xffffffff;
225 	ring->last_nr_txbb = 1;
226 	ring->poll_cnt = 0;
227 	memset(ring->buf, 0, ring->buf_size);
228 	ring->watchdog_time = 0;
229 
230 	ring->qp_state = MLX4_QP_STATE_RST;
231 	ring->doorbell_qpn = ring->qp.qpn << 8;
232 
233 	mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
234 				ring->cqn, user_prio, &ring->context);
235 	if (ring->bf_enabled)
236 		ring->context.usr_page = cpu_to_be32(ring->bf.uar->index);
237 
238 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
239 			       &ring->qp, &ring->qp_state);
240 	return err;
241 }
242 
243 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
244 				struct mlx4_en_tx_ring *ring)
245 {
246 	struct mlx4_en_dev *mdev = priv->mdev;
247 
248 	mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
249 		       MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
250 }
251 
252 static volatile struct mlx4_wqe_data_seg *
253 mlx4_en_store_inline_lso_data(volatile struct mlx4_wqe_data_seg *dseg,
254     struct mbuf *mb, int len, __be32 owner_bit)
255 {
256 	uint8_t *inl = __DEVOLATILE(uint8_t *, dseg);
257 
258 	/* copy data into place */
259 	m_copydata(mb, 0, len, inl + 4);
260 	dseg += DIV_ROUND_UP(4 + len, DS_SIZE_ALIGNMENT);
261 	return (dseg);
262 }
263 
264 static void
265 mlx4_en_store_inline_lso_header(volatile struct mlx4_wqe_data_seg *dseg,
266     int len, __be32 owner_bit)
267 {
268 }
269 
270 static void
271 mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
272     struct mlx4_en_tx_ring *ring, u32 index, u8 owner)
273 {
274 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
275 	struct mlx4_en_tx_desc *tx_desc = (struct mlx4_en_tx_desc *)
276 	    (ring->buf + (index * TXBB_SIZE));
277 	volatile __be32 *ptr = (__be32 *)tx_desc;
278 	const __be32 stamp = cpu_to_be32(STAMP_VAL |
279 	    ((u32)owner << STAMP_SHIFT));
280 	u32 i;
281 
282 	/* Stamp the freed descriptor */
283 	for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE; i += STAMP_STRIDE) {
284 		*ptr = stamp;
285 		ptr += STAMP_DWORDS;
286 	}
287 }
288 
289 static u32
290 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
291     struct mlx4_en_tx_ring *ring, u32 index)
292 {
293 	struct mlx4_en_tx_info *tx_info;
294 	struct mbuf *mb;
295 
296 	tx_info = &ring->tx_info[index];
297 	mb = tx_info->mb;
298 
299 	if (mb == NULL)
300 		goto done;
301 
302 	bus_dmamap_sync(ring->dma_tag, tx_info->dma_map,
303 	    BUS_DMASYNC_POSTWRITE);
304 	bus_dmamap_unload(ring->dma_tag, tx_info->dma_map);
305 
306         m_freem(mb);
307 done:
308 	return (tx_info->nr_txbb);
309 }
310 
311 int mlx4_en_free_tx_buf(if_t dev, struct mlx4_en_tx_ring *ring)
312 {
313 	struct mlx4_en_priv *priv = mlx4_netdev_priv(dev);
314 	int cnt = 0;
315 
316 	/* Skip last polled descriptor */
317 	ring->cons += ring->last_nr_txbb;
318 	en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
319 		 ring->cons, ring->prod);
320 
321 	if ((u32) (ring->prod - ring->cons) > ring->size) {
322                 en_warn(priv, "Tx consumer passed producer!\n");
323 		return 0;
324 	}
325 
326 	while (ring->cons != ring->prod) {
327 		ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
328 		    ring->cons & ring->size_mask);
329 		ring->cons += ring->last_nr_txbb;
330 		cnt++;
331 	}
332 
333 	if (cnt)
334 		en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
335 
336 	return cnt;
337 }
338 
339 static bool
340 mlx4_en_tx_ring_is_full(struct mlx4_en_tx_ring *ring)
341 {
342 	int wqs;
343 	wqs = ring->size - (ring->prod - ring->cons);
344 	return (wqs < (HEADROOM + (2 * MLX4_EN_TX_WQE_MAX_WQEBBS)));
345 }
346 
347 static int mlx4_en_process_tx_cq(if_t dev,
348 				 struct mlx4_en_cq *cq)
349 {
350 	struct mlx4_en_priv *priv = mlx4_netdev_priv(dev);
351 	struct mlx4_cq *mcq = &cq->mcq;
352 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->ring];
353 	struct mlx4_cqe *cqe;
354 	u16 index;
355 	u16 new_index, ring_index, stamp_index;
356 	u32 txbbs_skipped = 0;
357 	u32 txbbs_stamp = 0;
358 	u32 cons_index = mcq->cons_index;
359 	int size = cq->size;
360 	u32 size_mask = ring->size_mask;
361 	struct mlx4_cqe *buf = cq->buf;
362 	int factor = priv->cqe_factor;
363 
364 	if (!priv->port_up)
365 		return 0;
366 
367 	index = cons_index & size_mask;
368 	cqe = &buf[(index << factor) + factor];
369 	ring_index = ring->cons & size_mask;
370 	stamp_index = ring_index;
371 
372 	/* Process all completed CQEs */
373 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
374 			cons_index & size)) {
375 		/*
376 		 * make sure we read the CQE after we read the
377 		 * ownership bit
378 		 */
379 		rmb();
380 
381 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
382 			     MLX4_CQE_OPCODE_ERROR)) {
383 			en_err(priv, "CQE completed in error - vendor syndrom: 0x%x syndrom: 0x%x\n",
384 			       ((struct mlx4_err_cqe *)cqe)->
385 				       vendor_err_syndrome,
386 			       ((struct mlx4_err_cqe *)cqe)->syndrome);
387 		}
388 
389 		/* Skip over last polled CQE */
390 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
391 
392 		do {
393 			txbbs_skipped += ring->last_nr_txbb;
394 			ring_index = (ring_index + ring->last_nr_txbb) & size_mask;
395 			/* free next descriptor */
396 			ring->last_nr_txbb = mlx4_en_free_tx_desc(
397 			    priv, ring, ring_index);
398 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
399 					  !!((ring->cons + txbbs_stamp) &
400 						ring->size));
401 			stamp_index = ring_index;
402 			txbbs_stamp = txbbs_skipped;
403 		} while (ring_index != new_index);
404 
405 		++cons_index;
406 		index = cons_index & size_mask;
407 		cqe = &buf[(index << factor) + factor];
408 	}
409 
410 
411 	/*
412 	 * To prevent CQ overflow we first update CQ consumer and only then
413 	 * the ring consumer.
414 	 */
415 	mcq->cons_index = cons_index;
416 	mlx4_cq_set_ci(mcq);
417 	wmb();
418 	ring->cons += txbbs_skipped;
419 
420 	return (0);
421 }
422 
423 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
424 {
425 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
426 	struct mlx4_en_priv *priv = mlx4_netdev_priv(cq->dev);
427 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->ring];
428 
429 	if (priv->port_up == 0 || !spin_trylock(&ring->comp_lock))
430 		return;
431 	mlx4_en_process_tx_cq(cq->dev, cq);
432 	mod_timer(&cq->timer, jiffies + 1);
433 	spin_unlock(&ring->comp_lock);
434 }
435 
436 void mlx4_en_poll_tx_cq(unsigned long data)
437 {
438 	struct mlx4_en_cq *cq = (struct mlx4_en_cq *) data;
439 	struct mlx4_en_priv *priv = mlx4_netdev_priv(cq->dev);
440 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->ring];
441 	u32 inflight;
442 
443 	INC_PERF_COUNTER(priv->pstats.tx_poll);
444 
445 	if (priv->port_up == 0)
446 		return;
447 	if (!spin_trylock(&ring->comp_lock)) {
448 		mod_timer(&cq->timer, jiffies + MLX4_EN_TX_POLL_TIMEOUT);
449 		return;
450 	}
451 	mlx4_en_process_tx_cq(cq->dev, cq);
452 	inflight = (u32) (ring->prod - ring->cons - ring->last_nr_txbb);
453 
454 	/* If there are still packets in flight and the timer has not already
455 	 * been scheduled by the Tx routine then schedule it here to guarantee
456 	 * completion processing of these packets */
457 	if (inflight && priv->port_up)
458 		mod_timer(&cq->timer, jiffies + MLX4_EN_TX_POLL_TIMEOUT);
459 
460 	spin_unlock(&ring->comp_lock);
461 }
462 
463 static inline void mlx4_en_xmit_poll(struct mlx4_en_priv *priv, int tx_ind)
464 {
465 	struct mlx4_en_cq *cq = priv->tx_cq[tx_ind];
466 	struct mlx4_en_tx_ring *ring = priv->tx_ring[tx_ind];
467 
468 	if (priv->port_up == 0)
469 		return;
470 
471 	/* If we don't have a pending timer, set one up to catch our recent
472 	   post in case the interface becomes idle */
473 	if (!timer_pending(&cq->timer))
474 		mod_timer(&cq->timer, jiffies + MLX4_EN_TX_POLL_TIMEOUT);
475 
476 	/* Poll the CQ every mlx4_en_TX_MODER_POLL packets */
477 	if ((++ring->poll_cnt & (MLX4_EN_TX_POLL_MODER - 1)) == 0)
478 		if (spin_trylock(&ring->comp_lock)) {
479 			mlx4_en_process_tx_cq(priv->dev, cq);
480 			spin_unlock(&ring->comp_lock);
481 		}
482 }
483 
484 static u16
485 mlx4_en_get_inline_hdr_size(struct mlx4_en_tx_ring *ring, struct mbuf *mb)
486 {
487 	u16 retval;
488 
489 	/* only copy from first fragment, if possible */
490 	retval = MIN(ring->inline_thold, mb->m_len);
491 
492 	/* check for too little data */
493 	if (unlikely(retval < MIN_PKT_LEN))
494 		retval = MIN(ring->inline_thold, mb->m_pkthdr.len);
495 	return (retval);
496 }
497 
498 static int
499 mlx4_en_get_header_size(struct mbuf *mb)
500 {
501 	struct ether_vlan_header *eh;
502         struct tcphdr *th;
503         struct ip *ip;
504         int ip_hlen, tcp_hlen;
505 	struct ip6_hdr *ip6;
506 	uint16_t eth_type;
507 	int eth_hdr_len;
508 
509 	eh = mtod(mb, struct ether_vlan_header *);
510 	if (mb->m_len < ETHER_HDR_LEN)
511 		return (0);
512 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
513 		eth_type = ntohs(eh->evl_proto);
514 		eth_hdr_len = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
515 	} else {
516 		eth_type = ntohs(eh->evl_encap_proto);
517 		eth_hdr_len = ETHER_HDR_LEN;
518 	}
519 	if (mb->m_len < eth_hdr_len)
520 		return (0);
521 	switch (eth_type) {
522 	case ETHERTYPE_IP:
523 		ip = (struct ip *)(mb->m_data + eth_hdr_len);
524 		if (mb->m_len < eth_hdr_len + sizeof(*ip))
525 			return (0);
526 		if (ip->ip_p != IPPROTO_TCP)
527 			return (0);
528 		ip_hlen = ip->ip_hl << 2;
529 		eth_hdr_len += ip_hlen;
530 		break;
531 	case ETHERTYPE_IPV6:
532 		ip6 = (struct ip6_hdr *)(mb->m_data + eth_hdr_len);
533 		if (mb->m_len < eth_hdr_len + sizeof(*ip6))
534 			return (0);
535 		if (ip6->ip6_nxt != IPPROTO_TCP)
536 			return (0);
537 		eth_hdr_len += sizeof(*ip6);
538 		break;
539 	default:
540 		return (0);
541 	}
542 	if (mb->m_len < eth_hdr_len + sizeof(*th))
543 		return (0);
544 	th = (struct tcphdr *)(mb->m_data + eth_hdr_len);
545 	tcp_hlen = th->th_off << 2;
546 	eth_hdr_len += tcp_hlen;
547 	if (mb->m_len < eth_hdr_len)
548 		return (0);
549 	return (eth_hdr_len);
550 }
551 
552 static volatile struct mlx4_wqe_data_seg *
553 mlx4_en_store_inline_data(volatile struct mlx4_wqe_data_seg *dseg,
554     struct mbuf *mb, int len, __be32 owner_bit)
555 {
556 	uint8_t *inl = __DEVOLATILE(uint8_t *, dseg);
557 	const int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - 4;
558 
559 	if (unlikely(len < MIN_PKT_LEN)) {
560 		m_copydata(mb, 0, len, inl + 4);
561 		memset(inl + 4 + len, 0, MIN_PKT_LEN - len);
562 		dseg += DIV_ROUND_UP(4 + MIN_PKT_LEN, DS_SIZE_ALIGNMENT);
563 	} else if (len <= spc) {
564 		m_copydata(mb, 0, len, inl + 4);
565 		dseg += DIV_ROUND_UP(4 + len, DS_SIZE_ALIGNMENT);
566 	} else {
567 		m_copydata(mb, 0, spc, inl + 4);
568 		m_copydata(mb, spc, len - spc, inl + 8 + spc);
569 		dseg += DIV_ROUND_UP(8 + len, DS_SIZE_ALIGNMENT);
570 	}
571 	return (dseg);
572 }
573 
574 static void
575 mlx4_en_store_inline_header(volatile struct mlx4_wqe_data_seg *dseg,
576     int len, __be32 owner_bit)
577 {
578 	uint8_t *inl = __DEVOLATILE(uint8_t *, dseg);
579 	const int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - 4;
580 
581 	if (unlikely(len < MIN_PKT_LEN)) {
582 		*(volatile uint32_t *)inl =
583 		    SET_BYTE_COUNT((1U << 31) | MIN_PKT_LEN);
584 	} else if (len <= spc) {
585 		*(volatile uint32_t *)inl =
586 		    SET_BYTE_COUNT((1U << 31) | len);
587 	} else {
588 		*(volatile uint32_t *)(inl + 4 + spc) =
589 		    SET_BYTE_COUNT((1U << 31) | (len - spc));
590 		wmb();
591 		*(volatile uint32_t *)inl =
592 		    SET_BYTE_COUNT((1U << 31) | spc);
593 	}
594 }
595 
596 static uint32_t hashrandom;
597 static void hashrandom_init(void *arg)
598 {
599 	/*
600 	 * It is assumed that the random subsystem has been
601 	 * initialized when this function is called:
602 	 */
603 	hashrandom = m_ether_tcpip_hash_init();
604 }
605 SYSINIT(hashrandom_init, SI_SUB_RANDOM, SI_ORDER_ANY, &hashrandom_init, NULL);
606 
607 u16 mlx4_en_select_queue(if_t dev, struct mbuf *mb)
608 {
609 	struct mlx4_en_priv *priv = mlx4_netdev_priv(dev);
610 	u32 rings_p_up = priv->num_tx_rings_p_up;
611 	u32 up = 0;
612 	u32 queue_index;
613 
614 #if (MLX4_EN_NUM_UP > 1)
615 	/* Obtain VLAN information if present */
616 	if (mb->m_flags & M_VLANTAG) {
617 		u32 vlan_tag = mb->m_pkthdr.ether_vtag;
618 	        up = (vlan_tag >> 13) % MLX4_EN_NUM_UP;
619 	}
620 #endif
621 	queue_index = m_ether_tcpip_hash(MBUF_HASHFLAG_L3 | MBUF_HASHFLAG_L4, mb, hashrandom);
622 
623 	return ((queue_index % rings_p_up) + (up * rings_p_up));
624 }
625 
626 static void mlx4_bf_copy(void __iomem *dst, volatile unsigned long *src, unsigned bytecnt)
627 {
628 	__iowrite64_copy(dst, __DEVOLATILE(void *, src), bytecnt / 8);
629 }
630 
631 int mlx4_en_xmit(struct mlx4_en_priv *priv, int tx_ind, struct mbuf **mbp)
632 {
633 	enum {
634 		DS_FACT = TXBB_SIZE / DS_SIZE_ALIGNMENT,
635 		CTRL_FLAGS = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE |
636 		    MLX4_WQE_CTRL_SOLICITED),
637 	};
638 	bus_dma_segment_t segs[MLX4_EN_TX_MAX_MBUF_FRAGS];
639 	volatile struct mlx4_wqe_data_seg *dseg;
640 	volatile struct mlx4_wqe_data_seg *dseg_inline;
641 	volatile struct mlx4_en_tx_desc *tx_desc;
642 	struct mlx4_en_tx_ring *ring = priv->tx_ring[tx_ind];
643 	if_t ifp = priv->dev;
644 	struct mlx4_en_tx_info *tx_info;
645 	struct mbuf *mb = *mbp;
646 	struct mbuf *m;
647 	__be32 owner_bit;
648 	int nr_segs;
649 	int pad;
650 	int err;
651 	u32 bf_size;
652 	u32 bf_prod;
653 	u32 opcode;
654 	u16 index;
655 	u16 ds_cnt;
656 	u16 ihs;
657 
658 	if (unlikely(!priv->port_up)) {
659 		err = EINVAL;
660 		goto tx_drop;
661 	}
662 
663 	/* check if TX ring is full */
664 	if (unlikely(mlx4_en_tx_ring_is_full(ring))) {
665 		/* Use interrupts to find out when queue opened */
666 		mlx4_en_arm_cq(priv, priv->tx_cq[tx_ind]);
667 		return (ENOBUFS);
668 	}
669 
670 	/* sanity check we are not wrapping around */
671 	KASSERT(((~ring->prod) & ring->size_mask) >=
672 	    (MLX4_EN_TX_WQE_MAX_WQEBBS - 1), ("Wrapping around TX ring"));
673 
674 	/* Track current inflight packets for performance analysis */
675 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
676 			 (u32) (ring->prod - ring->cons - 1));
677 
678 	/* Track current mbuf packet header length */
679 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, mb->m_pkthdr.len);
680 
681 	/* Grab an index and try to transmit packet */
682 	owner_bit = (ring->prod & ring->size) ?
683 		cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0;
684 	index = ring->prod & ring->size_mask;
685 	tx_desc = (volatile struct mlx4_en_tx_desc *)
686 	    (ring->buf + index * TXBB_SIZE);
687 	tx_info = &ring->tx_info[index];
688 	dseg = &tx_desc->data;
689 
690 	/* send a copy of the frame to the BPF listener, if any */
691 	if (ifp != NULL && if_getbpf(ifp) != NULL)
692 		ETHER_BPF_MTAP(ifp, mb);
693 
694 	/* get default flags */
695 	tx_desc->ctrl.srcrb_flags = CTRL_FLAGS;
696 
697 	if (mb->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))
698 		tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM);
699 
700 	if (mb->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP |
701 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
702 		tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_TCP_UDP_CSUM);
703 
704 	/* do statistics */
705 	if (likely(tx_desc->ctrl.srcrb_flags != CTRL_FLAGS)) {
706 		priv->port_stats.tx_chksum_offload++;
707 		ring->tx_csum++;
708 	}
709 
710 	/* check for VLAN tag */
711 	if (mb->m_flags & M_VLANTAG) {
712 		tx_desc->ctrl.vlan_tag = cpu_to_be16(mb->m_pkthdr.ether_vtag);
713 		tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_CVLAN;
714 	} else {
715 		tx_desc->ctrl.vlan_tag = 0;
716 		tx_desc->ctrl.ins_vlan = 0;
717 	}
718 
719 	if (unlikely(mlx4_is_mfunc(priv->mdev->dev) || priv->validate_loopback)) {
720 		/*
721 		 * Copy destination MAC address to WQE. This allows
722 		 * loopback in eSwitch, so that VFs and PF can
723 		 * communicate with each other:
724 		 */
725 		m_copydata(mb, 0, 2, __DEVOLATILE(void *, &tx_desc->ctrl.srcrb_flags16[0]));
726 		m_copydata(mb, 2, 4, __DEVOLATILE(void *, &tx_desc->ctrl.imm));
727 	} else {
728 		/* clear immediate field */
729 		tx_desc->ctrl.imm = 0;
730 	}
731 
732 	/* Handle LSO (TSO) packets */
733 	if (mb->m_pkthdr.csum_flags & CSUM_TSO) {
734 		u32 payload_len;
735 		u32 mss = mb->m_pkthdr.tso_segsz;
736 		u32 num_pkts;
737 
738 		opcode = cpu_to_be32(MLX4_OPCODE_LSO | MLX4_WQE_CTRL_RR) |
739 		    owner_bit;
740 		ihs = mlx4_en_get_header_size(mb);
741 		if (unlikely(ihs > MAX_INLINE)) {
742 			ring->oversized_packets++;
743 			err = EINVAL;
744 			goto tx_drop;
745 		}
746 		tx_desc->lso.mss_hdr_size = cpu_to_be32((mss << 16) | ihs);
747 		payload_len = mb->m_pkthdr.len - ihs;
748 		if (unlikely(payload_len == 0))
749 			num_pkts = 1;
750 		else
751 			num_pkts = DIV_ROUND_UP(payload_len, mss);
752 		ring->bytes += payload_len + (num_pkts * ihs);
753 		ring->packets += num_pkts;
754 		ring->tso_packets++;
755 		/* store pointer to inline header */
756 		dseg_inline = dseg;
757 		/* copy data inline */
758 		dseg = mlx4_en_store_inline_lso_data(dseg,
759 		    mb, ihs, owner_bit);
760 	} else {
761 		opcode = cpu_to_be32(MLX4_OPCODE_SEND) |
762 		    owner_bit;
763 		ihs = mlx4_en_get_inline_hdr_size(ring, mb);
764 		ring->bytes += max_t (unsigned int,
765 		    mb->m_pkthdr.len, ETHER_MIN_LEN - ETHER_CRC_LEN);
766 		ring->packets++;
767 		/* store pointer to inline header */
768 		dseg_inline = dseg;
769 		/* copy data inline */
770 		dseg = mlx4_en_store_inline_data(dseg,
771 		    mb, ihs, owner_bit);
772 	}
773 	m_adj(mb, ihs);
774 
775 	err = bus_dmamap_load_mbuf_sg(ring->dma_tag, tx_info->dma_map,
776 	    mb, segs, &nr_segs, BUS_DMA_NOWAIT);
777 	if (unlikely(err == EFBIG)) {
778 		/* Too many mbuf fragments */
779 		ring->defrag_attempts++;
780 		m = m_defrag(mb, M_NOWAIT);
781 		if (m == NULL) {
782 			ring->oversized_packets++;
783 			goto tx_drop;
784 		}
785 		mb = m;
786 		/* Try again */
787 		err = bus_dmamap_load_mbuf_sg(ring->dma_tag, tx_info->dma_map,
788 		    mb, segs, &nr_segs, BUS_DMA_NOWAIT);
789 	}
790 	/* catch errors */
791 	if (unlikely(err != 0)) {
792 		ring->oversized_packets++;
793 		goto tx_drop;
794 	}
795 	/* If there were no errors and we didn't load anything, don't sync. */
796 	if (nr_segs != 0) {
797 		/* make sure all mbuf data is written to RAM */
798 		bus_dmamap_sync(ring->dma_tag, tx_info->dma_map,
799 		    BUS_DMASYNC_PREWRITE);
800 	} else {
801 		/* All data was inlined, free the mbuf. */
802 		bus_dmamap_unload(ring->dma_tag, tx_info->dma_map);
803 		m_freem(mb);
804 		mb = NULL;
805 	}
806 
807 	/* compute number of DS needed */
808 	ds_cnt = (dseg - ((volatile struct mlx4_wqe_data_seg *)tx_desc)) + nr_segs;
809 
810 	/*
811 	 * Check if the next request can wrap around and fill the end
812 	 * of the current request with zero immediate data:
813 	 */
814 	pad = DIV_ROUND_UP(ds_cnt, DS_FACT);
815 	pad = (~(ring->prod + pad)) & ring->size_mask;
816 
817 	if (unlikely(pad < (MLX4_EN_TX_WQE_MAX_WQEBBS - 1))) {
818 		/*
819 		 * Compute the least number of DS blocks we need to
820 		 * pad in order to achieve a TX ring wraparound:
821 		 */
822 		pad = (DS_FACT * (pad + 1));
823 	} else {
824 		/*
825 		 * The hardware will automatically jump to the next
826 		 * TXBB. No need for padding.
827 		 */
828 		pad = 0;
829 	}
830 
831 	/* compute total number of DS blocks */
832 	ds_cnt += pad;
833 	/*
834 	 * When modifying this code, please ensure that the following
835 	 * computation is always less than or equal to 0x3F:
836 	 *
837 	 * ((MLX4_EN_TX_WQE_MAX_WQEBBS - 1) * DS_FACT) +
838 	 * (MLX4_EN_TX_WQE_MAX_WQEBBS * DS_FACT)
839 	 *
840 	 * Else the "ds_cnt" variable can become too big.
841 	 */
842 	tx_desc->ctrl.fence_size = (ds_cnt & 0x3f);
843 
844 	/* store pointer to mbuf */
845 	tx_info->mb = mb;
846 	tx_info->nr_txbb = DIV_ROUND_UP(ds_cnt, DS_FACT);
847 	bf_size = ds_cnt * DS_SIZE_ALIGNMENT;
848 	bf_prod = ring->prod;
849 
850 	/* compute end of "dseg" array */
851 	dseg += nr_segs + pad;
852 
853 	/* pad using zero immediate dseg */
854 	while (pad--) {
855 		dseg--;
856 		dseg->addr = 0;
857 		dseg->lkey = 0;
858 		wmb();
859 		dseg->byte_count = SET_BYTE_COUNT((1U << 31)|0);
860 	}
861 
862 	/* fill segment list */
863 	while (nr_segs--) {
864 		if (unlikely(segs[nr_segs].ds_len == 0)) {
865 			dseg--;
866 			dseg->addr = 0;
867 			dseg->lkey = 0;
868 			wmb();
869 			dseg->byte_count = SET_BYTE_COUNT((1U << 31)|0);
870 		} else {
871 			dseg--;
872 			dseg->addr = cpu_to_be64((uint64_t)segs[nr_segs].ds_addr);
873 			dseg->lkey = cpu_to_be32(priv->mdev->mr.key);
874 			wmb();
875 			dseg->byte_count = SET_BYTE_COUNT((uint32_t)segs[nr_segs].ds_len);
876 		}
877 	}
878 
879 	wmb();
880 
881 	/* write owner bits in reverse order */
882 	if ((opcode & cpu_to_be32(0x1F)) == cpu_to_be32(MLX4_OPCODE_LSO))
883 		mlx4_en_store_inline_lso_header(dseg_inline, ihs, owner_bit);
884 	else
885 		mlx4_en_store_inline_header(dseg_inline, ihs, owner_bit);
886 
887 	/* update producer counter */
888 	ring->prod += tx_info->nr_txbb;
889 
890 	if (ring->bf_enabled && bf_size <= MAX_BF &&
891 	    (tx_desc->ctrl.ins_vlan != MLX4_WQE_CTRL_INS_CVLAN)) {
892 
893 		/* store doorbell number */
894 		*(volatile __be32 *) (&tx_desc->ctrl.vlan_tag) |= cpu_to_be32(ring->doorbell_qpn);
895 
896 		/* or in producer number for this WQE */
897 		opcode |= cpu_to_be32((bf_prod & 0xffff) << 8);
898 
899 		/*
900 		 * Ensure the new descriptor hits memory before
901 		 * setting ownership of this descriptor to HW:
902 		 */
903 		wmb();
904 		tx_desc->ctrl.owner_opcode = opcode;
905 		wmb();
906 		mlx4_bf_copy(((u8 *)ring->bf.reg) + ring->bf.offset,
907 		     (volatile unsigned long *) &tx_desc->ctrl, bf_size);
908 		wmb();
909 		ring->bf.offset ^= ring->bf.buf_size;
910 	} else {
911 		/*
912 		 * Ensure the new descriptor hits memory before
913 		 * setting ownership of this descriptor to HW:
914 		 */
915 		wmb();
916 		tx_desc->ctrl.owner_opcode = opcode;
917 		wmb();
918 		writel(cpu_to_be32(ring->doorbell_qpn),
919 		    ((u8 *)ring->bf.uar->map) + MLX4_SEND_DOORBELL);
920 	}
921 
922 	return (0);
923 tx_drop:
924 	*mbp = NULL;
925 	m_freem(mb);
926 	return (err);
927 }
928 
929 static int
930 mlx4_en_transmit_locked(if_t ifp, int tx_ind, struct mbuf *mb)
931 {
932 	struct mlx4_en_priv *priv = mlx4_netdev_priv(ifp);
933 	struct mlx4_en_tx_ring *ring = priv->tx_ring[tx_ind];
934 	int err = 0;
935 
936 	if (unlikely((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 ||
937 	    READ_ONCE(priv->port_up) == 0)) {
938 		m_freem(mb);
939 		return (ENETDOWN);
940 	}
941 
942 	if (mlx4_en_xmit(priv, tx_ind, &mb) != 0) {
943 		/* NOTE: m_freem() is NULL safe */
944 		m_freem(mb);
945 		err = ENOBUFS;
946 		if (ring->watchdog_time == 0)
947 			ring->watchdog_time = ticks + MLX4_EN_WATCHDOG_TIMEOUT;
948 	} else {
949 		ring->watchdog_time = 0;
950 	}
951 	return (err);
952 }
953 
954 int
955 mlx4_en_transmit(if_t dev, struct mbuf *m)
956 {
957 	struct mlx4_en_priv *priv = mlx4_netdev_priv(dev);
958 	struct mlx4_en_tx_ring *ring;
959 	int i, err = 0;
960 
961 	if (priv->port_up == 0) {
962 		m_freem(m);
963 		return (ENETDOWN);
964 	}
965 
966 	/* Compute which queue to use */
967 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
968 		i = (m->m_pkthdr.flowid % 128) % priv->tx_ring_num;
969 	}
970 	else {
971 		i = mlx4_en_select_queue(dev, m);
972 	}
973 
974 	ring = priv->tx_ring[i];
975 
976 	spin_lock(&ring->tx_lock);
977 
978 	err = mlx4_en_transmit_locked(dev, i, m);
979 	spin_unlock(&ring->tx_lock);
980 
981 	/* Poll CQ here */
982 	mlx4_en_xmit_poll(priv, i);
983 
984 	if (unlikely(err != 0))
985 		if_inc_counter(dev, IFCOUNTER_IQDROPS, 1);
986 
987 	return (err);
988 }
989 
990 /*
991  * Flush ring buffers.
992  */
993 void
994 mlx4_en_qflush(if_t dev)
995 {
996 	struct mlx4_en_priv *priv = mlx4_netdev_priv(dev);
997 
998 	if (priv->port_up == 0)
999 		return;
1000 
1001 	if_qflush(dev);
1002 }
1003