xref: /linux/drivers/net/ethernet/mellanox/mlx4/en_tx.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
1 /*
2  * Copyright (c) 2007 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 #include <asm/page.h>
35 #include <linux/mlx4/cq.h>
36 #include <linux/slab.h>
37 #include <linux/mlx4/qp.h>
38 #include <linux/skbuff.h>
39 #include <linux/if_vlan.h>
40 #include <linux/prefetch.h>
41 #include <linux/vmalloc.h>
42 #include <linux/tcp.h>
43 #include <linux/ip.h>
44 #include <linux/moduleparam.h>
45 
46 #include "mlx4_en.h"
47 
48 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
49 			   struct mlx4_en_tx_ring **pring, u32 size,
50 			   u16 stride, int node, int queue_index)
51 {
52 	struct mlx4_en_dev *mdev = priv->mdev;
53 	struct mlx4_en_tx_ring *ring;
54 	int tmp;
55 	int err;
56 
57 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
58 	if (!ring) {
59 		ring = kzalloc(sizeof(*ring), GFP_KERNEL);
60 		if (!ring) {
61 			en_err(priv, "Failed allocating TX ring\n");
62 			return -ENOMEM;
63 		}
64 	}
65 
66 	ring->size = size;
67 	ring->size_mask = size - 1;
68 	ring->stride = stride;
69 	ring->full_size = ring->size - HEADROOM - MAX_DESC_TXBBS;
70 
71 	tmp = size * sizeof(struct mlx4_en_tx_info);
72 	ring->tx_info = kmalloc_node(tmp, GFP_KERNEL | __GFP_NOWARN, node);
73 	if (!ring->tx_info) {
74 		ring->tx_info = vmalloc(tmp);
75 		if (!ring->tx_info) {
76 			err = -ENOMEM;
77 			goto err_ring;
78 		}
79 	}
80 
81 	en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
82 		 ring->tx_info, tmp);
83 
84 	ring->bounce_buf = kmalloc_node(MAX_DESC_SIZE, GFP_KERNEL, node);
85 	if (!ring->bounce_buf) {
86 		ring->bounce_buf = kmalloc(MAX_DESC_SIZE, GFP_KERNEL);
87 		if (!ring->bounce_buf) {
88 			err = -ENOMEM;
89 			goto err_info;
90 		}
91 	}
92 	ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
93 
94 	/* Allocate HW buffers on provided NUMA node */
95 	set_dev_node(&mdev->dev->persist->pdev->dev, node);
96 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size,
97 				 2 * PAGE_SIZE);
98 	set_dev_node(&mdev->dev->persist->pdev->dev, mdev->dev->numa_node);
99 	if (err) {
100 		en_err(priv, "Failed allocating hwq resources\n");
101 		goto err_bounce;
102 	}
103 
104 	err = mlx4_en_map_buffer(&ring->wqres.buf);
105 	if (err) {
106 		en_err(priv, "Failed to map TX buffer\n");
107 		goto err_hwq_res;
108 	}
109 
110 	ring->buf = ring->wqres.buf.direct.buf;
111 
112 	en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d buf_size:%d dma:%llx\n",
113 	       ring, ring->buf, ring->size, ring->buf_size,
114 	       (unsigned long long) ring->wqres.buf.direct.map);
115 
116 	err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &ring->qpn,
117 				    MLX4_RESERVE_ETH_BF_QP);
118 	if (err) {
119 		en_err(priv, "failed reserving qp for TX ring\n");
120 		goto err_map;
121 	}
122 
123 	err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp, GFP_KERNEL);
124 	if (err) {
125 		en_err(priv, "Failed allocating qp %d\n", ring->qpn);
126 		goto err_reserve;
127 	}
128 	ring->qp.event = mlx4_en_sqp_event;
129 
130 	err = mlx4_bf_alloc(mdev->dev, &ring->bf, node);
131 	if (err) {
132 		en_dbg(DRV, priv, "working without blueflame (%d)\n", err);
133 		ring->bf.uar = &mdev->priv_uar;
134 		ring->bf.uar->map = mdev->uar_map;
135 		ring->bf_enabled = false;
136 		ring->bf_alloced = false;
137 		priv->pflags &= ~MLX4_EN_PRIV_FLAGS_BLUEFLAME;
138 	} else {
139 		ring->bf_alloced = true;
140 		ring->bf_enabled = !!(priv->pflags &
141 				      MLX4_EN_PRIV_FLAGS_BLUEFLAME);
142 	}
143 
144 	ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type;
145 	ring->queue_index = queue_index;
146 
147 	if (queue_index < priv->num_tx_rings_p_up)
148 		cpumask_set_cpu(cpumask_local_spread(queue_index,
149 						     priv->mdev->dev->numa_node),
150 				&ring->affinity_mask);
151 
152 	*pring = ring;
153 	return 0;
154 
155 err_reserve:
156 	mlx4_qp_release_range(mdev->dev, ring->qpn, 1);
157 err_map:
158 	mlx4_en_unmap_buffer(&ring->wqres.buf);
159 err_hwq_res:
160 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
161 err_bounce:
162 	kfree(ring->bounce_buf);
163 	ring->bounce_buf = NULL;
164 err_info:
165 	kvfree(ring->tx_info);
166 	ring->tx_info = NULL;
167 err_ring:
168 	kfree(ring);
169 	*pring = NULL;
170 	return err;
171 }
172 
173 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
174 			     struct mlx4_en_tx_ring **pring)
175 {
176 	struct mlx4_en_dev *mdev = priv->mdev;
177 	struct mlx4_en_tx_ring *ring = *pring;
178 	en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
179 
180 	if (ring->bf_alloced)
181 		mlx4_bf_free(mdev->dev, &ring->bf);
182 	mlx4_qp_remove(mdev->dev, &ring->qp);
183 	mlx4_qp_free(mdev->dev, &ring->qp);
184 	mlx4_qp_release_range(priv->mdev->dev, ring->qpn, 1);
185 	mlx4_en_unmap_buffer(&ring->wqres.buf);
186 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
187 	kfree(ring->bounce_buf);
188 	ring->bounce_buf = NULL;
189 	kvfree(ring->tx_info);
190 	ring->tx_info = NULL;
191 	kfree(ring);
192 	*pring = NULL;
193 }
194 
195 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
196 			     struct mlx4_en_tx_ring *ring,
197 			     int cq, int user_prio)
198 {
199 	struct mlx4_en_dev *mdev = priv->mdev;
200 	int err;
201 
202 	ring->cqn = cq;
203 	ring->prod = 0;
204 	ring->cons = 0xffffffff;
205 	ring->last_nr_txbb = 1;
206 	memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
207 	memset(ring->buf, 0, ring->buf_size);
208 
209 	ring->qp_state = MLX4_QP_STATE_RST;
210 	ring->doorbell_qpn = cpu_to_be32(ring->qp.qpn << 8);
211 	ring->mr_key = cpu_to_be32(mdev->mr.key);
212 
213 	mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
214 				ring->cqn, user_prio, &ring->context);
215 	if (ring->bf_alloced)
216 		ring->context.usr_page = cpu_to_be32(ring->bf.uar->index);
217 
218 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
219 			       &ring->qp, &ring->qp_state);
220 	if (!cpumask_empty(&ring->affinity_mask))
221 		netif_set_xps_queue(priv->dev, &ring->affinity_mask,
222 				    ring->queue_index);
223 
224 	return err;
225 }
226 
227 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
228 				struct mlx4_en_tx_ring *ring)
229 {
230 	struct mlx4_en_dev *mdev = priv->mdev;
231 
232 	mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
233 		       MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
234 }
235 
236 static inline bool mlx4_en_is_tx_ring_full(struct mlx4_en_tx_ring *ring)
237 {
238 	return ring->prod - ring->cons > ring->full_size;
239 }
240 
241 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
242 			      struct mlx4_en_tx_ring *ring, int index,
243 			      u8 owner)
244 {
245 	__be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
246 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
247 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
248 	void *end = ring->buf + ring->buf_size;
249 	__be32 *ptr = (__be32 *)tx_desc;
250 	int i;
251 
252 	/* Optimize the common case when there are no wraparounds */
253 	if (likely((void *)tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
254 		/* Stamp the freed descriptor */
255 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
256 		     i += STAMP_STRIDE) {
257 			*ptr = stamp;
258 			ptr += STAMP_DWORDS;
259 		}
260 	} else {
261 		/* Stamp the freed descriptor */
262 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
263 		     i += STAMP_STRIDE) {
264 			*ptr = stamp;
265 			ptr += STAMP_DWORDS;
266 			if ((void *)ptr >= end) {
267 				ptr = ring->buf;
268 				stamp ^= cpu_to_be32(0x80000000);
269 			}
270 		}
271 	}
272 }
273 
274 
275 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
276 				struct mlx4_en_tx_ring *ring,
277 				int index, u8 owner, u64 timestamp)
278 {
279 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
280 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
281 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
282 	void *end = ring->buf + ring->buf_size;
283 	struct sk_buff *skb = tx_info->skb;
284 	int nr_maps = tx_info->nr_maps;
285 	int i;
286 
287 	/* We do not touch skb here, so prefetch skb->users location
288 	 * to speedup consume_skb()
289 	 */
290 	prefetchw(&skb->users);
291 
292 	if (unlikely(timestamp)) {
293 		struct skb_shared_hwtstamps hwts;
294 
295 		mlx4_en_fill_hwtstamps(priv->mdev, &hwts, timestamp);
296 		skb_tstamp_tx(skb, &hwts);
297 	}
298 
299 	/* Optimize the common case when there are no wraparounds */
300 	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
301 		if (!tx_info->inl) {
302 			if (tx_info->linear)
303 				dma_unmap_single(priv->ddev,
304 						tx_info->map0_dma,
305 						tx_info->map0_byte_count,
306 						PCI_DMA_TODEVICE);
307 			else
308 				dma_unmap_page(priv->ddev,
309 					       tx_info->map0_dma,
310 					       tx_info->map0_byte_count,
311 					       PCI_DMA_TODEVICE);
312 			for (i = 1; i < nr_maps; i++) {
313 				data++;
314 				dma_unmap_page(priv->ddev,
315 					(dma_addr_t)be64_to_cpu(data->addr),
316 					be32_to_cpu(data->byte_count),
317 					PCI_DMA_TODEVICE);
318 			}
319 		}
320 	} else {
321 		if (!tx_info->inl) {
322 			if ((void *) data >= end) {
323 				data = ring->buf + ((void *)data - end);
324 			}
325 
326 			if (tx_info->linear)
327 				dma_unmap_single(priv->ddev,
328 						tx_info->map0_dma,
329 						tx_info->map0_byte_count,
330 						PCI_DMA_TODEVICE);
331 			else
332 				dma_unmap_page(priv->ddev,
333 					       tx_info->map0_dma,
334 					       tx_info->map0_byte_count,
335 					       PCI_DMA_TODEVICE);
336 			for (i = 1; i < nr_maps; i++) {
337 				data++;
338 				/* Check for wraparound before unmapping */
339 				if ((void *) data >= end)
340 					data = ring->buf;
341 				dma_unmap_page(priv->ddev,
342 					(dma_addr_t)be64_to_cpu(data->addr),
343 					be32_to_cpu(data->byte_count),
344 					PCI_DMA_TODEVICE);
345 			}
346 		}
347 	}
348 	dev_consume_skb_any(skb);
349 	return tx_info->nr_txbb;
350 }
351 
352 
353 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
354 {
355 	struct mlx4_en_priv *priv = netdev_priv(dev);
356 	int cnt = 0;
357 
358 	/* Skip last polled descriptor */
359 	ring->cons += ring->last_nr_txbb;
360 	en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
361 		 ring->cons, ring->prod);
362 
363 	if ((u32) (ring->prod - ring->cons) > ring->size) {
364 		if (netif_msg_tx_err(priv))
365 			en_warn(priv, "Tx consumer passed producer!\n");
366 		return 0;
367 	}
368 
369 	while (ring->cons != ring->prod) {
370 		ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
371 						ring->cons & ring->size_mask,
372 						!!(ring->cons & ring->size), 0);
373 		ring->cons += ring->last_nr_txbb;
374 		cnt++;
375 	}
376 
377 	netdev_tx_reset_queue(ring->tx_queue);
378 
379 	if (cnt)
380 		en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
381 
382 	return cnt;
383 }
384 
385 static bool mlx4_en_process_tx_cq(struct net_device *dev,
386 				 struct mlx4_en_cq *cq)
387 {
388 	struct mlx4_en_priv *priv = netdev_priv(dev);
389 	struct mlx4_cq *mcq = &cq->mcq;
390 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->ring];
391 	struct mlx4_cqe *cqe;
392 	u16 index;
393 	u16 new_index, ring_index, stamp_index;
394 	u32 txbbs_skipped = 0;
395 	u32 txbbs_stamp = 0;
396 	u32 cons_index = mcq->cons_index;
397 	int size = cq->size;
398 	u32 size_mask = ring->size_mask;
399 	struct mlx4_cqe *buf = cq->buf;
400 	u32 packets = 0;
401 	u32 bytes = 0;
402 	int factor = priv->cqe_factor;
403 	u64 timestamp = 0;
404 	int done = 0;
405 	int budget = priv->tx_work_limit;
406 	u32 last_nr_txbb;
407 	u32 ring_cons;
408 
409 	if (!priv->port_up)
410 		return true;
411 
412 	netdev_txq_bql_complete_prefetchw(ring->tx_queue);
413 
414 	index = cons_index & size_mask;
415 	cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
416 	last_nr_txbb = ACCESS_ONCE(ring->last_nr_txbb);
417 	ring_cons = ACCESS_ONCE(ring->cons);
418 	ring_index = ring_cons & size_mask;
419 	stamp_index = ring_index;
420 
421 	/* Process all completed CQEs */
422 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
423 			cons_index & size) && (done < budget)) {
424 		/*
425 		 * make sure we read the CQE after we read the
426 		 * ownership bit
427 		 */
428 		dma_rmb();
429 
430 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
431 			     MLX4_CQE_OPCODE_ERROR)) {
432 			struct mlx4_err_cqe *cqe_err = (struct mlx4_err_cqe *)cqe;
433 
434 			en_err(priv, "CQE error - vendor syndrome: 0x%x syndrome: 0x%x\n",
435 			       cqe_err->vendor_err_syndrome,
436 			       cqe_err->syndrome);
437 		}
438 
439 		/* Skip over last polled CQE */
440 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
441 
442 		do {
443 			txbbs_skipped += last_nr_txbb;
444 			ring_index = (ring_index + last_nr_txbb) & size_mask;
445 			if (ring->tx_info[ring_index].ts_requested)
446 				timestamp = mlx4_en_get_cqe_ts(cqe);
447 
448 			/* free next descriptor */
449 			last_nr_txbb = mlx4_en_free_tx_desc(
450 					priv, ring, ring_index,
451 					!!((ring_cons + txbbs_skipped) &
452 					ring->size), timestamp);
453 
454 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
455 					  !!((ring_cons + txbbs_stamp) &
456 						ring->size));
457 			stamp_index = ring_index;
458 			txbbs_stamp = txbbs_skipped;
459 			packets++;
460 			bytes += ring->tx_info[ring_index].nr_bytes;
461 		} while ((++done < budget) && (ring_index != new_index));
462 
463 		++cons_index;
464 		index = cons_index & size_mask;
465 		cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
466 	}
467 
468 
469 	/*
470 	 * To prevent CQ overflow we first update CQ consumer and only then
471 	 * the ring consumer.
472 	 */
473 	mcq->cons_index = cons_index;
474 	mlx4_cq_set_ci(mcq);
475 	wmb();
476 
477 	/* we want to dirty this cache line once */
478 	ACCESS_ONCE(ring->last_nr_txbb) = last_nr_txbb;
479 	ACCESS_ONCE(ring->cons) = ring_cons + txbbs_skipped;
480 
481 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
482 
483 	/* Wakeup Tx queue if this stopped, and ring is not full.
484 	 */
485 	if (netif_tx_queue_stopped(ring->tx_queue) &&
486 	    !mlx4_en_is_tx_ring_full(ring)) {
487 		netif_tx_wake_queue(ring->tx_queue);
488 		ring->wake_queue++;
489 	}
490 	return done < budget;
491 }
492 
493 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
494 {
495 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
496 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
497 
498 	if (likely(priv->port_up))
499 		napi_schedule_irqoff(&cq->napi);
500 	else
501 		mlx4_en_arm_cq(priv, cq);
502 }
503 
504 /* TX CQ polling - called by NAPI */
505 int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget)
506 {
507 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
508 	struct net_device *dev = cq->dev;
509 	struct mlx4_en_priv *priv = netdev_priv(dev);
510 	int clean_complete;
511 
512 	clean_complete = mlx4_en_process_tx_cq(dev, cq);
513 	if (!clean_complete)
514 		return budget;
515 
516 	napi_complete(napi);
517 	mlx4_en_arm_cq(priv, cq);
518 
519 	return 0;
520 }
521 
522 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
523 						      struct mlx4_en_tx_ring *ring,
524 						      u32 index,
525 						      unsigned int desc_size)
526 {
527 	u32 copy = (ring->size - index) * TXBB_SIZE;
528 	int i;
529 
530 	for (i = desc_size - copy - 4; i >= 0; i -= 4) {
531 		if ((i & (TXBB_SIZE - 1)) == 0)
532 			wmb();
533 
534 		*((u32 *) (ring->buf + i)) =
535 			*((u32 *) (ring->bounce_buf + copy + i));
536 	}
537 
538 	for (i = copy - 4; i >= 4 ; i -= 4) {
539 		if ((i & (TXBB_SIZE - 1)) == 0)
540 			wmb();
541 
542 		*((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
543 			*((u32 *) (ring->bounce_buf + i));
544 	}
545 
546 	/* Return real descriptor location */
547 	return ring->buf + index * TXBB_SIZE;
548 }
549 
550 /* Decide if skb can be inlined in tx descriptor to avoid dma mapping
551  *
552  * It seems strange we do not simply use skb_copy_bits().
553  * This would allow to inline all skbs iff skb->len <= inline_thold
554  *
555  * Note that caller already checked skb was not a gso packet
556  */
557 static bool is_inline(int inline_thold, const struct sk_buff *skb,
558 		      const struct skb_shared_info *shinfo,
559 		      void **pfrag)
560 {
561 	void *ptr;
562 
563 	if (skb->len > inline_thold || !inline_thold)
564 		return false;
565 
566 	if (shinfo->nr_frags == 1) {
567 		ptr = skb_frag_address_safe(&shinfo->frags[0]);
568 		if (unlikely(!ptr))
569 			return false;
570 		*pfrag = ptr;
571 		return true;
572 	}
573 	if (shinfo->nr_frags)
574 		return false;
575 	return true;
576 }
577 
578 static int inline_size(const struct sk_buff *skb)
579 {
580 	if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
581 	    <= MLX4_INLINE_ALIGN)
582 		return ALIGN(skb->len + CTRL_SIZE +
583 			     sizeof(struct mlx4_wqe_inline_seg), 16);
584 	else
585 		return ALIGN(skb->len + CTRL_SIZE + 2 *
586 			     sizeof(struct mlx4_wqe_inline_seg), 16);
587 }
588 
589 static int get_real_size(const struct sk_buff *skb,
590 			 const struct skb_shared_info *shinfo,
591 			 struct net_device *dev,
592 			 int *lso_header_size,
593 			 bool *inline_ok,
594 			 void **pfrag)
595 {
596 	struct mlx4_en_priv *priv = netdev_priv(dev);
597 	int real_size;
598 
599 	if (shinfo->gso_size) {
600 		*inline_ok = false;
601 		if (skb->encapsulation)
602 			*lso_header_size = (skb_inner_transport_header(skb) - skb->data) + inner_tcp_hdrlen(skb);
603 		else
604 			*lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
605 		real_size = CTRL_SIZE + shinfo->nr_frags * DS_SIZE +
606 			ALIGN(*lso_header_size + 4, DS_SIZE);
607 		if (unlikely(*lso_header_size != skb_headlen(skb))) {
608 			/* We add a segment for the skb linear buffer only if
609 			 * it contains data */
610 			if (*lso_header_size < skb_headlen(skb))
611 				real_size += DS_SIZE;
612 			else {
613 				if (netif_msg_tx_err(priv))
614 					en_warn(priv, "Non-linear headers\n");
615 				return 0;
616 			}
617 		}
618 	} else {
619 		*lso_header_size = 0;
620 		*inline_ok = is_inline(priv->prof->inline_thold, skb,
621 				       shinfo, pfrag);
622 
623 		if (*inline_ok)
624 			real_size = inline_size(skb);
625 		else
626 			real_size = CTRL_SIZE +
627 				    (shinfo->nr_frags + 1) * DS_SIZE;
628 	}
629 
630 	return real_size;
631 }
632 
633 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
634 			     const struct sk_buff *skb,
635 			     const struct skb_shared_info *shinfo,
636 			     int real_size, u16 *vlan_tag,
637 			     int tx_ind, void *fragptr)
638 {
639 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
640 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
641 	unsigned int hlen = skb_headlen(skb);
642 
643 	if (skb->len <= spc) {
644 		if (likely(skb->len >= MIN_PKT_LEN)) {
645 			inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
646 		} else {
647 			inl->byte_count = cpu_to_be32(1 << 31 | MIN_PKT_LEN);
648 			memset(((void *)(inl + 1)) + skb->len, 0,
649 			       MIN_PKT_LEN - skb->len);
650 		}
651 		skb_copy_from_linear_data(skb, inl + 1, hlen);
652 		if (shinfo->nr_frags)
653 			memcpy(((void *)(inl + 1)) + hlen, fragptr,
654 			       skb_frag_size(&shinfo->frags[0]));
655 
656 	} else {
657 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
658 		if (hlen <= spc) {
659 			skb_copy_from_linear_data(skb, inl + 1, hlen);
660 			if (hlen < spc) {
661 				memcpy(((void *)(inl + 1)) + hlen,
662 				       fragptr, spc - hlen);
663 				fragptr +=  spc - hlen;
664 			}
665 			inl = (void *) (inl + 1) + spc;
666 			memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
667 		} else {
668 			skb_copy_from_linear_data(skb, inl + 1, spc);
669 			inl = (void *) (inl + 1) + spc;
670 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
671 							 hlen - spc);
672 			if (shinfo->nr_frags)
673 				memcpy(((void *)(inl + 1)) + hlen - spc,
674 				       fragptr,
675 				       skb_frag_size(&shinfo->frags[0]));
676 		}
677 
678 		dma_wmb();
679 		inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
680 	}
681 }
682 
683 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb,
684 			 void *accel_priv, select_queue_fallback_t fallback)
685 {
686 	struct mlx4_en_priv *priv = netdev_priv(dev);
687 	u16 rings_p_up = priv->num_tx_rings_p_up;
688 	u8 up = 0;
689 
690 	if (dev->num_tc)
691 		return skb_tx_hash(dev, skb);
692 
693 	if (skb_vlan_tag_present(skb))
694 		up = skb_vlan_tag_get(skb) >> VLAN_PRIO_SHIFT;
695 
696 	return fallback(dev, skb) % rings_p_up + up * rings_p_up;
697 }
698 
699 static void mlx4_bf_copy(void __iomem *dst, const void *src,
700 			 unsigned int bytecnt)
701 {
702 	__iowrite64_copy(dst, src, bytecnt / 8);
703 }
704 
705 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
706 {
707 	struct skb_shared_info *shinfo = skb_shinfo(skb);
708 	struct mlx4_en_priv *priv = netdev_priv(dev);
709 	struct device *ddev = priv->ddev;
710 	struct mlx4_en_tx_ring *ring;
711 	struct mlx4_en_tx_desc *tx_desc;
712 	struct mlx4_wqe_data_seg *data;
713 	struct mlx4_en_tx_info *tx_info;
714 	int tx_ind = 0;
715 	int nr_txbb;
716 	int desc_size;
717 	int real_size;
718 	u32 index, bf_index;
719 	__be32 op_own;
720 	u16 vlan_tag = 0;
721 	u16 vlan_proto = 0;
722 	int i_frag;
723 	int lso_header_size;
724 	void *fragptr = NULL;
725 	bool bounce = false;
726 	bool send_doorbell;
727 	bool stop_queue;
728 	bool inline_ok;
729 	u32 ring_cons;
730 
731 	if (!priv->port_up)
732 		goto tx_drop;
733 
734 	tx_ind = skb_get_queue_mapping(skb);
735 	ring = priv->tx_ring[tx_ind];
736 
737 	/* fetch ring->cons far ahead before needing it to avoid stall */
738 	ring_cons = ACCESS_ONCE(ring->cons);
739 
740 	real_size = get_real_size(skb, shinfo, dev, &lso_header_size,
741 				  &inline_ok, &fragptr);
742 	if (unlikely(!real_size))
743 		goto tx_drop;
744 
745 	/* Align descriptor to TXBB size */
746 	desc_size = ALIGN(real_size, TXBB_SIZE);
747 	nr_txbb = desc_size / TXBB_SIZE;
748 	if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
749 		if (netif_msg_tx_err(priv))
750 			en_warn(priv, "Oversized header or SG list\n");
751 		goto tx_drop;
752 	}
753 
754 	if (skb_vlan_tag_present(skb)) {
755 		vlan_tag = skb_vlan_tag_get(skb);
756 		vlan_proto = be16_to_cpu(skb->vlan_proto);
757 	}
758 
759 	netdev_txq_bql_enqueue_prefetchw(ring->tx_queue);
760 
761 	/* Track current inflight packets for performance analysis */
762 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
763 			 (u32)(ring->prod - ring_cons - 1));
764 
765 	/* Packet is good - grab an index and transmit it */
766 	index = ring->prod & ring->size_mask;
767 	bf_index = ring->prod;
768 
769 	/* See if we have enough space for whole descriptor TXBB for setting
770 	 * SW ownership on next descriptor; if not, use a bounce buffer. */
771 	if (likely(index + nr_txbb <= ring->size))
772 		tx_desc = ring->buf + index * TXBB_SIZE;
773 	else {
774 		tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
775 		bounce = true;
776 	}
777 
778 	/* Save skb in tx_info ring */
779 	tx_info = &ring->tx_info[index];
780 	tx_info->skb = skb;
781 	tx_info->nr_txbb = nr_txbb;
782 
783 	data = &tx_desc->data;
784 	if (lso_header_size)
785 		data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
786 						      DS_SIZE));
787 
788 	/* valid only for none inline segments */
789 	tx_info->data_offset = (void *)data - (void *)tx_desc;
790 
791 	tx_info->inl = inline_ok;
792 
793 	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
794 			   !inline_ok) ? 1 : 0;
795 
796 	tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
797 	data += tx_info->nr_maps - 1;
798 
799 	if (!tx_info->inl) {
800 		dma_addr_t dma = 0;
801 		u32 byte_count = 0;
802 
803 		/* Map fragments if any */
804 		for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
805 			const struct skb_frag_struct *frag;
806 
807 			frag = &shinfo->frags[i_frag];
808 			byte_count = skb_frag_size(frag);
809 			dma = skb_frag_dma_map(ddev, frag,
810 					       0, byte_count,
811 					       DMA_TO_DEVICE);
812 			if (dma_mapping_error(ddev, dma))
813 				goto tx_drop_unmap;
814 
815 			data->addr = cpu_to_be64(dma);
816 			data->lkey = ring->mr_key;
817 			dma_wmb();
818 			data->byte_count = cpu_to_be32(byte_count);
819 			--data;
820 		}
821 
822 		/* Map linear part if needed */
823 		if (tx_info->linear) {
824 			byte_count = skb_headlen(skb) - lso_header_size;
825 
826 			dma = dma_map_single(ddev, skb->data +
827 					     lso_header_size, byte_count,
828 					     PCI_DMA_TODEVICE);
829 			if (dma_mapping_error(ddev, dma))
830 				goto tx_drop_unmap;
831 
832 			data->addr = cpu_to_be64(dma);
833 			data->lkey = ring->mr_key;
834 			dma_wmb();
835 			data->byte_count = cpu_to_be32(byte_count);
836 		}
837 		/* tx completion can avoid cache line miss for common cases */
838 		tx_info->map0_dma = dma;
839 		tx_info->map0_byte_count = byte_count;
840 	}
841 
842 	/*
843 	 * For timestamping add flag to skb_shinfo and
844 	 * set flag for further reference
845 	 */
846 	tx_info->ts_requested = 0;
847 	if (unlikely(ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
848 		     shinfo->tx_flags & SKBTX_HW_TSTAMP)) {
849 		shinfo->tx_flags |= SKBTX_IN_PROGRESS;
850 		tx_info->ts_requested = 1;
851 	}
852 
853 	/* Prepare ctrl segement apart opcode+ownership, which depends on
854 	 * whether LSO is used */
855 	tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
856 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
857 		if (!skb->encapsulation)
858 			tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
859 								 MLX4_WQE_CTRL_TCP_UDP_CSUM);
860 		else
861 			tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM);
862 		ring->tx_csum++;
863 	}
864 
865 	if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
866 		struct ethhdr *ethh;
867 
868 		/* Copy dst mac address to wqe. This allows loopback in eSwitch,
869 		 * so that VFs and PF can communicate with each other
870 		 */
871 		ethh = (struct ethhdr *)skb->data;
872 		tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
873 		tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
874 	}
875 
876 	/* Handle LSO (TSO) packets */
877 	if (lso_header_size) {
878 		int i;
879 
880 		/* Mark opcode as LSO */
881 		op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
882 			((ring->prod & ring->size) ?
883 				cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
884 
885 		/* Fill in the LSO prefix */
886 		tx_desc->lso.mss_hdr_size = cpu_to_be32(
887 			shinfo->gso_size << 16 | lso_header_size);
888 
889 		/* Copy headers;
890 		 * note that we already verified that it is linear */
891 		memcpy(tx_desc->lso.header, skb->data, lso_header_size);
892 
893 		ring->tso_packets++;
894 
895 		i = ((skb->len - lso_header_size) / shinfo->gso_size) +
896 			!!((skb->len - lso_header_size) % shinfo->gso_size);
897 		tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
898 		ring->packets += i;
899 	} else {
900 		/* Normal (Non LSO) packet */
901 		op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
902 			((ring->prod & ring->size) ?
903 			 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
904 		tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
905 		ring->packets++;
906 	}
907 	ring->bytes += tx_info->nr_bytes;
908 	netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
909 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
910 
911 	if (tx_info->inl)
912 		build_inline_wqe(tx_desc, skb, shinfo, real_size, &vlan_tag,
913 				 tx_ind, fragptr);
914 
915 	if (skb->encapsulation) {
916 		struct iphdr *ipv4 = (struct iphdr *)skb_inner_network_header(skb);
917 		if (ipv4->protocol == IPPROTO_TCP || ipv4->protocol == IPPROTO_UDP)
918 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP | MLX4_WQE_CTRL_ILP);
919 		else
920 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP);
921 	}
922 
923 	ring->prod += nr_txbb;
924 
925 	/* If we used a bounce buffer then copy descriptor back into place */
926 	if (unlikely(bounce))
927 		tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
928 
929 	skb_tx_timestamp(skb);
930 
931 	/* Check available TXBBs And 2K spare for prefetch */
932 	stop_queue = mlx4_en_is_tx_ring_full(ring);
933 	if (unlikely(stop_queue)) {
934 		netif_tx_stop_queue(ring->tx_queue);
935 		ring->queue_stopped++;
936 	}
937 	send_doorbell = !skb->xmit_more || netif_xmit_stopped(ring->tx_queue);
938 
939 	real_size = (real_size / 16) & 0x3f;
940 
941 	if (ring->bf_enabled && desc_size <= MAX_BF && !bounce &&
942 	    !skb_vlan_tag_present(skb) && send_doorbell) {
943 		tx_desc->ctrl.bf_qpn = ring->doorbell_qpn |
944 				       cpu_to_be32(real_size);
945 
946 		op_own |= htonl((bf_index & 0xffff) << 8);
947 		/* Ensure new descriptor hits memory
948 		 * before setting ownership of this descriptor to HW
949 		 */
950 		dma_wmb();
951 		tx_desc->ctrl.owner_opcode = op_own;
952 
953 		wmb();
954 
955 		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, &tx_desc->ctrl,
956 			     desc_size);
957 
958 		wmb();
959 
960 		ring->bf.offset ^= ring->bf.buf_size;
961 	} else {
962 		tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
963 		if (vlan_proto == ETH_P_8021AD)
964 			tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_SVLAN;
965 		else if (vlan_proto == ETH_P_8021Q)
966 			tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_CVLAN;
967 		else
968 			tx_desc->ctrl.ins_vlan = 0;
969 
970 		tx_desc->ctrl.fence_size = real_size;
971 
972 		/* Ensure new descriptor hits memory
973 		 * before setting ownership of this descriptor to HW
974 		 */
975 		dma_wmb();
976 		tx_desc->ctrl.owner_opcode = op_own;
977 		if (send_doorbell) {
978 			wmb();
979 			/* Since there is no iowrite*_native() that writes the
980 			 * value as is, without byteswapping - using the one
981 			 * the doesn't do byteswapping in the relevant arch
982 			 * endianness.
983 			 */
984 #if defined(__LITTLE_ENDIAN)
985 			iowrite32(
986 #else
987 			iowrite32be(
988 #endif
989 				  ring->doorbell_qpn,
990 				  ring->bf.uar->map + MLX4_SEND_DOORBELL);
991 		} else {
992 			ring->xmit_more++;
993 		}
994 	}
995 
996 	if (unlikely(stop_queue)) {
997 		/* If queue was emptied after the if (stop_queue) , and before
998 		 * the netif_tx_stop_queue() - need to wake the queue,
999 		 * or else it will remain stopped forever.
1000 		 * Need a memory barrier to make sure ring->cons was not
1001 		 * updated before queue was stopped.
1002 		 */
1003 		smp_rmb();
1004 
1005 		ring_cons = ACCESS_ONCE(ring->cons);
1006 		if (unlikely(!mlx4_en_is_tx_ring_full(ring))) {
1007 			netif_tx_wake_queue(ring->tx_queue);
1008 			ring->wake_queue++;
1009 		}
1010 	}
1011 	return NETDEV_TX_OK;
1012 
1013 tx_drop_unmap:
1014 	en_err(priv, "DMA mapping error\n");
1015 
1016 	while (++i_frag < shinfo->nr_frags) {
1017 		++data;
1018 		dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
1019 			       be32_to_cpu(data->byte_count),
1020 			       PCI_DMA_TODEVICE);
1021 	}
1022 
1023 tx_drop:
1024 	dev_kfree_skb_any(skb);
1025 	priv->stats.tx_dropped++;
1026 	return NETDEV_TX_OK;
1027 }
1028 
1029