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, int qpn, 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 
70 	tmp = size * sizeof(struct mlx4_en_tx_info);
71 	ring->tx_info = kmalloc_node(tmp, GFP_KERNEL | __GFP_NOWARN, node);
72 	if (!ring->tx_info) {
73 		ring->tx_info = vmalloc(tmp);
74 		if (!ring->tx_info) {
75 			err = -ENOMEM;
76 			goto err_ring;
77 		}
78 	}
79 
80 	en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
81 		 ring->tx_info, tmp);
82 
83 	ring->bounce_buf = kmalloc_node(MAX_DESC_SIZE, GFP_KERNEL, node);
84 	if (!ring->bounce_buf) {
85 		ring->bounce_buf = kmalloc(MAX_DESC_SIZE, GFP_KERNEL);
86 		if (!ring->bounce_buf) {
87 			err = -ENOMEM;
88 			goto err_info;
89 		}
90 	}
91 	ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
92 
93 	/* Allocate HW buffers on provided NUMA node */
94 	set_dev_node(&mdev->dev->pdev->dev, node);
95 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size,
96 				 2 * PAGE_SIZE);
97 	set_dev_node(&mdev->dev->pdev->dev, mdev->dev->numa_node);
98 	if (err) {
99 		en_err(priv, "Failed allocating hwq resources\n");
100 		goto err_bounce;
101 	}
102 
103 	err = mlx4_en_map_buffer(&ring->wqres.buf);
104 	if (err) {
105 		en_err(priv, "Failed to map TX buffer\n");
106 		goto err_hwq_res;
107 	}
108 
109 	ring->buf = ring->wqres.buf.direct.buf;
110 
111 	en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d buf_size:%d dma:%llx\n",
112 	       ring, ring->buf, ring->size, ring->buf_size,
113 	       (unsigned long long) ring->wqres.buf.direct.map);
114 
115 	ring->qpn = qpn;
116 	err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp, GFP_KERNEL);
117 	if (err) {
118 		en_err(priv, "Failed allocating qp %d\n", ring->qpn);
119 		goto err_map;
120 	}
121 	ring->qp.event = mlx4_en_sqp_event;
122 
123 	err = mlx4_bf_alloc(mdev->dev, &ring->bf, node);
124 	if (err) {
125 		en_dbg(DRV, priv, "working without blueflame (%d)\n", err);
126 		ring->bf.uar = &mdev->priv_uar;
127 		ring->bf.uar->map = mdev->uar_map;
128 		ring->bf_enabled = false;
129 		ring->bf_alloced = false;
130 		priv->pflags &= ~MLX4_EN_PRIV_FLAGS_BLUEFLAME;
131 	} else {
132 		ring->bf_alloced = true;
133 		ring->bf_enabled = !!(priv->pflags &
134 				      MLX4_EN_PRIV_FLAGS_BLUEFLAME);
135 	}
136 
137 	ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type;
138 	ring->queue_index = queue_index;
139 
140 	if (queue_index < priv->num_tx_rings_p_up && cpu_online(queue_index))
141 		cpumask_set_cpu(queue_index, &ring->affinity_mask);
142 
143 	*pring = ring;
144 	return 0;
145 
146 err_map:
147 	mlx4_en_unmap_buffer(&ring->wqres.buf);
148 err_hwq_res:
149 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
150 err_bounce:
151 	kfree(ring->bounce_buf);
152 	ring->bounce_buf = NULL;
153 err_info:
154 	kvfree(ring->tx_info);
155 	ring->tx_info = NULL;
156 err_ring:
157 	kfree(ring);
158 	*pring = NULL;
159 	return err;
160 }
161 
162 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
163 			     struct mlx4_en_tx_ring **pring)
164 {
165 	struct mlx4_en_dev *mdev = priv->mdev;
166 	struct mlx4_en_tx_ring *ring = *pring;
167 	en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
168 
169 	if (ring->bf_alloced)
170 		mlx4_bf_free(mdev->dev, &ring->bf);
171 	mlx4_qp_remove(mdev->dev, &ring->qp);
172 	mlx4_qp_free(mdev->dev, &ring->qp);
173 	mlx4_en_unmap_buffer(&ring->wqres.buf);
174 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
175 	kfree(ring->bounce_buf);
176 	ring->bounce_buf = NULL;
177 	kvfree(ring->tx_info);
178 	ring->tx_info = NULL;
179 	kfree(ring);
180 	*pring = NULL;
181 }
182 
183 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
184 			     struct mlx4_en_tx_ring *ring,
185 			     int cq, int user_prio)
186 {
187 	struct mlx4_en_dev *mdev = priv->mdev;
188 	int err;
189 
190 	ring->cqn = cq;
191 	ring->prod = 0;
192 	ring->cons = 0xffffffff;
193 	ring->last_nr_txbb = 1;
194 	memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
195 	memset(ring->buf, 0, ring->buf_size);
196 
197 	ring->qp_state = MLX4_QP_STATE_RST;
198 	ring->doorbell_qpn = cpu_to_be32(ring->qp.qpn << 8);
199 	ring->mr_key = cpu_to_be32(mdev->mr.key);
200 
201 	mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
202 				ring->cqn, user_prio, &ring->context);
203 	if (ring->bf_alloced)
204 		ring->context.usr_page = cpu_to_be32(ring->bf.uar->index);
205 
206 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
207 			       &ring->qp, &ring->qp_state);
208 	if (!user_prio && cpu_online(ring->queue_index))
209 		netif_set_xps_queue(priv->dev, &ring->affinity_mask,
210 				    ring->queue_index);
211 
212 	return err;
213 }
214 
215 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
216 				struct mlx4_en_tx_ring *ring)
217 {
218 	struct mlx4_en_dev *mdev = priv->mdev;
219 
220 	mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
221 		       MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
222 }
223 
224 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
225 			      struct mlx4_en_tx_ring *ring, int index,
226 			      u8 owner)
227 {
228 	__be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
229 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
230 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
231 	void *end = ring->buf + ring->buf_size;
232 	__be32 *ptr = (__be32 *)tx_desc;
233 	int i;
234 
235 	/* Optimize the common case when there are no wraparounds */
236 	if (likely((void *)tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
237 		/* Stamp the freed descriptor */
238 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
239 		     i += STAMP_STRIDE) {
240 			*ptr = stamp;
241 			ptr += STAMP_DWORDS;
242 		}
243 	} else {
244 		/* Stamp the freed descriptor */
245 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
246 		     i += STAMP_STRIDE) {
247 			*ptr = stamp;
248 			ptr += STAMP_DWORDS;
249 			if ((void *)ptr >= end) {
250 				ptr = ring->buf;
251 				stamp ^= cpu_to_be32(0x80000000);
252 			}
253 		}
254 	}
255 }
256 
257 
258 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
259 				struct mlx4_en_tx_ring *ring,
260 				int index, u8 owner, u64 timestamp)
261 {
262 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
263 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
264 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
265 	void *end = ring->buf + ring->buf_size;
266 	struct sk_buff *skb = tx_info->skb;
267 	int nr_maps = tx_info->nr_maps;
268 	int i;
269 
270 	/* We do not touch skb here, so prefetch skb->users location
271 	 * to speedup consume_skb()
272 	 */
273 	prefetchw(&skb->users);
274 
275 	if (unlikely(timestamp)) {
276 		struct skb_shared_hwtstamps hwts;
277 
278 		mlx4_en_fill_hwtstamps(priv->mdev, &hwts, timestamp);
279 		skb_tstamp_tx(skb, &hwts);
280 	}
281 
282 	/* Optimize the common case when there are no wraparounds */
283 	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
284 		if (!tx_info->inl) {
285 			if (tx_info->linear)
286 				dma_unmap_single(priv->ddev,
287 						tx_info->map0_dma,
288 						tx_info->map0_byte_count,
289 						PCI_DMA_TODEVICE);
290 			else
291 				dma_unmap_page(priv->ddev,
292 					       tx_info->map0_dma,
293 					       tx_info->map0_byte_count,
294 					       PCI_DMA_TODEVICE);
295 			for (i = 1; i < nr_maps; i++) {
296 				data++;
297 				dma_unmap_page(priv->ddev,
298 					(dma_addr_t)be64_to_cpu(data->addr),
299 					be32_to_cpu(data->byte_count),
300 					PCI_DMA_TODEVICE);
301 			}
302 		}
303 	} else {
304 		if (!tx_info->inl) {
305 			if ((void *) data >= end) {
306 				data = ring->buf + ((void *)data - end);
307 			}
308 
309 			if (tx_info->linear)
310 				dma_unmap_single(priv->ddev,
311 						tx_info->map0_dma,
312 						tx_info->map0_byte_count,
313 						PCI_DMA_TODEVICE);
314 			else
315 				dma_unmap_page(priv->ddev,
316 					       tx_info->map0_dma,
317 					       tx_info->map0_byte_count,
318 					       PCI_DMA_TODEVICE);
319 			for (i = 1; i < nr_maps; i++) {
320 				data++;
321 				/* Check for wraparound before unmapping */
322 				if ((void *) data >= end)
323 					data = ring->buf;
324 				dma_unmap_page(priv->ddev,
325 					(dma_addr_t)be64_to_cpu(data->addr),
326 					be32_to_cpu(data->byte_count),
327 					PCI_DMA_TODEVICE);
328 			}
329 		}
330 	}
331 	dev_consume_skb_any(skb);
332 	return tx_info->nr_txbb;
333 }
334 
335 
336 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
337 {
338 	struct mlx4_en_priv *priv = netdev_priv(dev);
339 	int cnt = 0;
340 
341 	/* Skip last polled descriptor */
342 	ring->cons += ring->last_nr_txbb;
343 	en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
344 		 ring->cons, ring->prod);
345 
346 	if ((u32) (ring->prod - ring->cons) > ring->size) {
347 		if (netif_msg_tx_err(priv))
348 			en_warn(priv, "Tx consumer passed producer!\n");
349 		return 0;
350 	}
351 
352 	while (ring->cons != ring->prod) {
353 		ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
354 						ring->cons & ring->size_mask,
355 						!!(ring->cons & ring->size), 0);
356 		ring->cons += ring->last_nr_txbb;
357 		cnt++;
358 	}
359 
360 	netdev_tx_reset_queue(ring->tx_queue);
361 
362 	if (cnt)
363 		en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
364 
365 	return cnt;
366 }
367 
368 static bool mlx4_en_process_tx_cq(struct net_device *dev,
369 				 struct mlx4_en_cq *cq)
370 {
371 	struct mlx4_en_priv *priv = netdev_priv(dev);
372 	struct mlx4_cq *mcq = &cq->mcq;
373 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->ring];
374 	struct mlx4_cqe *cqe;
375 	u16 index;
376 	u16 new_index, ring_index, stamp_index;
377 	u32 txbbs_skipped = 0;
378 	u32 txbbs_stamp = 0;
379 	u32 cons_index = mcq->cons_index;
380 	int size = cq->size;
381 	u32 size_mask = ring->size_mask;
382 	struct mlx4_cqe *buf = cq->buf;
383 	u32 packets = 0;
384 	u32 bytes = 0;
385 	int factor = priv->cqe_factor;
386 	u64 timestamp = 0;
387 	int done = 0;
388 	int budget = priv->tx_work_limit;
389 	u32 last_nr_txbb;
390 	u32 ring_cons;
391 
392 	if (!priv->port_up)
393 		return true;
394 
395 	netdev_txq_bql_complete_prefetchw(ring->tx_queue);
396 
397 	index = cons_index & size_mask;
398 	cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
399 	last_nr_txbb = ACCESS_ONCE(ring->last_nr_txbb);
400 	ring_cons = ACCESS_ONCE(ring->cons);
401 	ring_index = ring_cons & size_mask;
402 	stamp_index = ring_index;
403 
404 	/* Process all completed CQEs */
405 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
406 			cons_index & size) && (done < budget)) {
407 		/*
408 		 * make sure we read the CQE after we read the
409 		 * ownership bit
410 		 */
411 		rmb();
412 
413 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
414 			     MLX4_CQE_OPCODE_ERROR)) {
415 			struct mlx4_err_cqe *cqe_err = (struct mlx4_err_cqe *)cqe;
416 
417 			en_err(priv, "CQE error - vendor syndrome: 0x%x syndrome: 0x%x\n",
418 			       cqe_err->vendor_err_syndrome,
419 			       cqe_err->syndrome);
420 		}
421 
422 		/* Skip over last polled CQE */
423 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
424 
425 		do {
426 			txbbs_skipped += last_nr_txbb;
427 			ring_index = (ring_index + last_nr_txbb) & size_mask;
428 			if (ring->tx_info[ring_index].ts_requested)
429 				timestamp = mlx4_en_get_cqe_ts(cqe);
430 
431 			/* free next descriptor */
432 			last_nr_txbb = mlx4_en_free_tx_desc(
433 					priv, ring, ring_index,
434 					!!((ring_cons + txbbs_skipped) &
435 					ring->size), timestamp);
436 
437 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
438 					  !!((ring_cons + txbbs_stamp) &
439 						ring->size));
440 			stamp_index = ring_index;
441 			txbbs_stamp = txbbs_skipped;
442 			packets++;
443 			bytes += ring->tx_info[ring_index].nr_bytes;
444 		} while ((++done < budget) && (ring_index != new_index));
445 
446 		++cons_index;
447 		index = cons_index & size_mask;
448 		cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
449 	}
450 
451 
452 	/*
453 	 * To prevent CQ overflow we first update CQ consumer and only then
454 	 * the ring consumer.
455 	 */
456 	mcq->cons_index = cons_index;
457 	mlx4_cq_set_ci(mcq);
458 	wmb();
459 
460 	/* we want to dirty this cache line once */
461 	ACCESS_ONCE(ring->last_nr_txbb) = last_nr_txbb;
462 	ACCESS_ONCE(ring->cons) = ring_cons + txbbs_skipped;
463 
464 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
465 
466 	/*
467 	 * Wakeup Tx queue if this stopped, and at least 1 packet
468 	 * was completed
469 	 */
470 	if (netif_tx_queue_stopped(ring->tx_queue) && txbbs_skipped > 0) {
471 		netif_tx_wake_queue(ring->tx_queue);
472 		ring->wake_queue++;
473 	}
474 	return done < budget;
475 }
476 
477 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
478 {
479 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
480 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
481 
482 	if (priv->port_up)
483 		napi_schedule(&cq->napi);
484 	else
485 		mlx4_en_arm_cq(priv, cq);
486 }
487 
488 /* TX CQ polling - called by NAPI */
489 int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget)
490 {
491 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
492 	struct net_device *dev = cq->dev;
493 	struct mlx4_en_priv *priv = netdev_priv(dev);
494 	int clean_complete;
495 
496 	clean_complete = mlx4_en_process_tx_cq(dev, cq);
497 	if (!clean_complete)
498 		return budget;
499 
500 	napi_complete(napi);
501 	mlx4_en_arm_cq(priv, cq);
502 
503 	return 0;
504 }
505 
506 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
507 						      struct mlx4_en_tx_ring *ring,
508 						      u32 index,
509 						      unsigned int desc_size)
510 {
511 	u32 copy = (ring->size - index) * TXBB_SIZE;
512 	int i;
513 
514 	for (i = desc_size - copy - 4; i >= 0; i -= 4) {
515 		if ((i & (TXBB_SIZE - 1)) == 0)
516 			wmb();
517 
518 		*((u32 *) (ring->buf + i)) =
519 			*((u32 *) (ring->bounce_buf + copy + i));
520 	}
521 
522 	for (i = copy - 4; i >= 4 ; i -= 4) {
523 		if ((i & (TXBB_SIZE - 1)) == 0)
524 			wmb();
525 
526 		*((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
527 			*((u32 *) (ring->bounce_buf + i));
528 	}
529 
530 	/* Return real descriptor location */
531 	return ring->buf + index * TXBB_SIZE;
532 }
533 
534 /* Decide if skb can be inlined in tx descriptor to avoid dma mapping
535  *
536  * It seems strange we do not simply use skb_copy_bits().
537  * This would allow to inline all skbs iff skb->len <= inline_thold
538  *
539  * Note that caller already checked skb was not a gso packet
540  */
541 static bool is_inline(int inline_thold, const struct sk_buff *skb,
542 		      const struct skb_shared_info *shinfo,
543 		      void **pfrag)
544 {
545 	void *ptr;
546 
547 	if (skb->len > inline_thold || !inline_thold)
548 		return false;
549 
550 	if (shinfo->nr_frags == 1) {
551 		ptr = skb_frag_address_safe(&shinfo->frags[0]);
552 		if (unlikely(!ptr))
553 			return false;
554 		*pfrag = ptr;
555 		return true;
556 	}
557 	if (shinfo->nr_frags)
558 		return false;
559 	return true;
560 }
561 
562 static int inline_size(const struct sk_buff *skb)
563 {
564 	if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
565 	    <= MLX4_INLINE_ALIGN)
566 		return ALIGN(skb->len + CTRL_SIZE +
567 			     sizeof(struct mlx4_wqe_inline_seg), 16);
568 	else
569 		return ALIGN(skb->len + CTRL_SIZE + 2 *
570 			     sizeof(struct mlx4_wqe_inline_seg), 16);
571 }
572 
573 static int get_real_size(const struct sk_buff *skb,
574 			 const struct skb_shared_info *shinfo,
575 			 struct net_device *dev,
576 			 int *lso_header_size,
577 			 bool *inline_ok,
578 			 void **pfrag)
579 {
580 	struct mlx4_en_priv *priv = netdev_priv(dev);
581 	int real_size;
582 
583 	if (shinfo->gso_size) {
584 		*inline_ok = false;
585 		if (skb->encapsulation)
586 			*lso_header_size = (skb_inner_transport_header(skb) - skb->data) + inner_tcp_hdrlen(skb);
587 		else
588 			*lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
589 		real_size = CTRL_SIZE + shinfo->nr_frags * DS_SIZE +
590 			ALIGN(*lso_header_size + 4, DS_SIZE);
591 		if (unlikely(*lso_header_size != skb_headlen(skb))) {
592 			/* We add a segment for the skb linear buffer only if
593 			 * it contains data */
594 			if (*lso_header_size < skb_headlen(skb))
595 				real_size += DS_SIZE;
596 			else {
597 				if (netif_msg_tx_err(priv))
598 					en_warn(priv, "Non-linear headers\n");
599 				return 0;
600 			}
601 		}
602 	} else {
603 		*lso_header_size = 0;
604 		*inline_ok = is_inline(priv->prof->inline_thold, skb,
605 				       shinfo, pfrag);
606 
607 		if (*inline_ok)
608 			real_size = inline_size(skb);
609 		else
610 			real_size = CTRL_SIZE +
611 				    (shinfo->nr_frags + 1) * DS_SIZE;
612 	}
613 
614 	return real_size;
615 }
616 
617 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
618 			     const struct sk_buff *skb,
619 			     const struct skb_shared_info *shinfo,
620 			     int real_size, u16 *vlan_tag,
621 			     int tx_ind, void *fragptr)
622 {
623 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
624 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
625 	unsigned int hlen = skb_headlen(skb);
626 
627 	if (skb->len <= spc) {
628 		if (likely(skb->len >= MIN_PKT_LEN)) {
629 			inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
630 		} else {
631 			inl->byte_count = cpu_to_be32(1 << 31 | MIN_PKT_LEN);
632 			memset(((void *)(inl + 1)) + skb->len, 0,
633 			       MIN_PKT_LEN - skb->len);
634 		}
635 		skb_copy_from_linear_data(skb, inl + 1, hlen);
636 		if (shinfo->nr_frags)
637 			memcpy(((void *)(inl + 1)) + hlen, fragptr,
638 			       skb_frag_size(&shinfo->frags[0]));
639 
640 	} else {
641 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
642 		if (hlen <= spc) {
643 			skb_copy_from_linear_data(skb, inl + 1, hlen);
644 			if (hlen < spc) {
645 				memcpy(((void *)(inl + 1)) + hlen,
646 				       fragptr, spc - hlen);
647 				fragptr +=  spc - hlen;
648 			}
649 			inl = (void *) (inl + 1) + spc;
650 			memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
651 		} else {
652 			skb_copy_from_linear_data(skb, inl + 1, spc);
653 			inl = (void *) (inl + 1) + spc;
654 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
655 							 hlen - spc);
656 			if (shinfo->nr_frags)
657 				memcpy(((void *)(inl + 1)) + hlen - spc,
658 				       fragptr,
659 				       skb_frag_size(&shinfo->frags[0]));
660 		}
661 
662 		wmb();
663 		inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
664 	}
665 }
666 
667 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb,
668 			 void *accel_priv, select_queue_fallback_t fallback)
669 {
670 	struct mlx4_en_priv *priv = netdev_priv(dev);
671 	u16 rings_p_up = priv->num_tx_rings_p_up;
672 	u8 up = 0;
673 
674 	if (dev->num_tc)
675 		return skb_tx_hash(dev, skb);
676 
677 	if (vlan_tx_tag_present(skb))
678 		up = vlan_tx_tag_get(skb) >> VLAN_PRIO_SHIFT;
679 
680 	return fallback(dev, skb) % rings_p_up + up * rings_p_up;
681 }
682 
683 static void mlx4_bf_copy(void __iomem *dst, const void *src,
684 			 unsigned int bytecnt)
685 {
686 	__iowrite64_copy(dst, src, bytecnt / 8);
687 }
688 
689 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
690 {
691 	struct skb_shared_info *shinfo = skb_shinfo(skb);
692 	struct mlx4_en_priv *priv = netdev_priv(dev);
693 	struct device *ddev = priv->ddev;
694 	struct mlx4_en_tx_ring *ring;
695 	struct mlx4_en_tx_desc *tx_desc;
696 	struct mlx4_wqe_data_seg *data;
697 	struct mlx4_en_tx_info *tx_info;
698 	int tx_ind = 0;
699 	int nr_txbb;
700 	int desc_size;
701 	int real_size;
702 	u32 index, bf_index;
703 	__be32 op_own;
704 	u16 vlan_tag = 0;
705 	int i_frag;
706 	int lso_header_size;
707 	void *fragptr = NULL;
708 	bool bounce = false;
709 	bool send_doorbell;
710 	bool stop_queue;
711 	bool inline_ok;
712 	u32 ring_cons;
713 
714 	if (!priv->port_up)
715 		goto tx_drop;
716 
717 	tx_ind = skb_get_queue_mapping(skb);
718 	ring = priv->tx_ring[tx_ind];
719 
720 	/* fetch ring->cons far ahead before needing it to avoid stall */
721 	ring_cons = ACCESS_ONCE(ring->cons);
722 
723 	real_size = get_real_size(skb, shinfo, dev, &lso_header_size,
724 				  &inline_ok, &fragptr);
725 	if (unlikely(!real_size))
726 		goto tx_drop;
727 
728 	/* Align descriptor to TXBB size */
729 	desc_size = ALIGN(real_size, TXBB_SIZE);
730 	nr_txbb = desc_size / TXBB_SIZE;
731 	if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
732 		if (netif_msg_tx_err(priv))
733 			en_warn(priv, "Oversized header or SG list\n");
734 		goto tx_drop;
735 	}
736 
737 	if (vlan_tx_tag_present(skb))
738 		vlan_tag = vlan_tx_tag_get(skb);
739 
740 
741 	netdev_txq_bql_enqueue_prefetchw(ring->tx_queue);
742 
743 	/* Track current inflight packets for performance analysis */
744 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
745 			 (u32)(ring->prod - ring_cons - 1));
746 
747 	/* Packet is good - grab an index and transmit it */
748 	index = ring->prod & ring->size_mask;
749 	bf_index = ring->prod;
750 
751 	/* See if we have enough space for whole descriptor TXBB for setting
752 	 * SW ownership on next descriptor; if not, use a bounce buffer. */
753 	if (likely(index + nr_txbb <= ring->size))
754 		tx_desc = ring->buf + index * TXBB_SIZE;
755 	else {
756 		tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
757 		bounce = true;
758 	}
759 
760 	/* Save skb in tx_info ring */
761 	tx_info = &ring->tx_info[index];
762 	tx_info->skb = skb;
763 	tx_info->nr_txbb = nr_txbb;
764 
765 	data = &tx_desc->data;
766 	if (lso_header_size)
767 		data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
768 						      DS_SIZE));
769 
770 	/* valid only for none inline segments */
771 	tx_info->data_offset = (void *)data - (void *)tx_desc;
772 
773 	tx_info->inl = inline_ok;
774 
775 	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
776 			   !inline_ok) ? 1 : 0;
777 
778 	tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
779 	data += tx_info->nr_maps - 1;
780 
781 	if (!tx_info->inl) {
782 		dma_addr_t dma = 0;
783 		u32 byte_count = 0;
784 
785 		/* Map fragments if any */
786 		for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
787 			const struct skb_frag_struct *frag;
788 
789 			frag = &shinfo->frags[i_frag];
790 			byte_count = skb_frag_size(frag);
791 			dma = skb_frag_dma_map(ddev, frag,
792 					       0, byte_count,
793 					       DMA_TO_DEVICE);
794 			if (dma_mapping_error(ddev, dma))
795 				goto tx_drop_unmap;
796 
797 			data->addr = cpu_to_be64(dma);
798 			data->lkey = ring->mr_key;
799 			wmb();
800 			data->byte_count = cpu_to_be32(byte_count);
801 			--data;
802 		}
803 
804 		/* Map linear part if needed */
805 		if (tx_info->linear) {
806 			byte_count = skb_headlen(skb) - lso_header_size;
807 
808 			dma = dma_map_single(ddev, skb->data +
809 					     lso_header_size, byte_count,
810 					     PCI_DMA_TODEVICE);
811 			if (dma_mapping_error(ddev, dma))
812 				goto tx_drop_unmap;
813 
814 			data->addr = cpu_to_be64(dma);
815 			data->lkey = ring->mr_key;
816 			wmb();
817 			data->byte_count = cpu_to_be32(byte_count);
818 		}
819 		/* tx completion can avoid cache line miss for common cases */
820 		tx_info->map0_dma = dma;
821 		tx_info->map0_byte_count = byte_count;
822 	}
823 
824 	/*
825 	 * For timestamping add flag to skb_shinfo and
826 	 * set flag for further reference
827 	 */
828 	tx_info->ts_requested = 0;
829 	if (unlikely(ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
830 		     shinfo->tx_flags & SKBTX_HW_TSTAMP)) {
831 		shinfo->tx_flags |= SKBTX_IN_PROGRESS;
832 		tx_info->ts_requested = 1;
833 	}
834 
835 	/* Prepare ctrl segement apart opcode+ownership, which depends on
836 	 * whether LSO is used */
837 	tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
838 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
839 		if (!skb->encapsulation)
840 			tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
841 								 MLX4_WQE_CTRL_TCP_UDP_CSUM);
842 		else
843 			tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM);
844 		ring->tx_csum++;
845 	}
846 
847 	if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
848 		struct ethhdr *ethh;
849 
850 		/* Copy dst mac address to wqe. This allows loopback in eSwitch,
851 		 * so that VFs and PF can communicate with each other
852 		 */
853 		ethh = (struct ethhdr *)skb->data;
854 		tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
855 		tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
856 	}
857 
858 	/* Handle LSO (TSO) packets */
859 	if (lso_header_size) {
860 		int i;
861 
862 		/* Mark opcode as LSO */
863 		op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
864 			((ring->prod & ring->size) ?
865 				cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
866 
867 		/* Fill in the LSO prefix */
868 		tx_desc->lso.mss_hdr_size = cpu_to_be32(
869 			shinfo->gso_size << 16 | lso_header_size);
870 
871 		/* Copy headers;
872 		 * note that we already verified that it is linear */
873 		memcpy(tx_desc->lso.header, skb->data, lso_header_size);
874 
875 		ring->tso_packets++;
876 
877 		i = ((skb->len - lso_header_size) / shinfo->gso_size) +
878 			!!((skb->len - lso_header_size) % shinfo->gso_size);
879 		tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
880 		ring->packets += i;
881 	} else {
882 		/* Normal (Non LSO) packet */
883 		op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
884 			((ring->prod & ring->size) ?
885 			 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
886 		tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
887 		ring->packets++;
888 	}
889 	ring->bytes += tx_info->nr_bytes;
890 	netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
891 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
892 
893 	if (tx_info->inl)
894 		build_inline_wqe(tx_desc, skb, shinfo, real_size, &vlan_tag,
895 				 tx_ind, fragptr);
896 
897 	if (skb->encapsulation) {
898 		struct iphdr *ipv4 = (struct iphdr *)skb_inner_network_header(skb);
899 		if (ipv4->protocol == IPPROTO_TCP || ipv4->protocol == IPPROTO_UDP)
900 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP | MLX4_WQE_CTRL_ILP);
901 		else
902 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP);
903 	}
904 
905 	ring->prod += nr_txbb;
906 
907 	/* If we used a bounce buffer then copy descriptor back into place */
908 	if (unlikely(bounce))
909 		tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
910 
911 	skb_tx_timestamp(skb);
912 
913 	/* Check available TXBBs And 2K spare for prefetch */
914 	stop_queue = (int)(ring->prod - ring_cons) >
915 		      ring->size - HEADROOM - MAX_DESC_TXBBS;
916 	if (unlikely(stop_queue)) {
917 		netif_tx_stop_queue(ring->tx_queue);
918 		ring->queue_stopped++;
919 	}
920 	send_doorbell = !skb->xmit_more || netif_xmit_stopped(ring->tx_queue);
921 
922 	real_size = (real_size / 16) & 0x3f;
923 
924 	if (ring->bf_enabled && desc_size <= MAX_BF && !bounce &&
925 	    !vlan_tx_tag_present(skb) && send_doorbell) {
926 		tx_desc->ctrl.bf_qpn = ring->doorbell_qpn |
927 				       cpu_to_be32(real_size);
928 
929 		op_own |= htonl((bf_index & 0xffff) << 8);
930 		/* Ensure new descriptor hits memory
931 		 * before setting ownership of this descriptor to HW
932 		 */
933 		wmb();
934 		tx_desc->ctrl.owner_opcode = op_own;
935 
936 		wmb();
937 
938 		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, &tx_desc->ctrl,
939 			     desc_size);
940 
941 		wmb();
942 
943 		ring->bf.offset ^= ring->bf.buf_size;
944 	} else {
945 		tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
946 		tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN *
947 			!!vlan_tx_tag_present(skb);
948 		tx_desc->ctrl.fence_size = real_size;
949 
950 		/* Ensure new descriptor hits memory
951 		 * before setting ownership of this descriptor to HW
952 		 */
953 		wmb();
954 		tx_desc->ctrl.owner_opcode = op_own;
955 		if (send_doorbell) {
956 			wmb();
957 			iowrite32(ring->doorbell_qpn,
958 				  ring->bf.uar->map + MLX4_SEND_DOORBELL);
959 		} else {
960 			ring->xmit_more++;
961 		}
962 	}
963 
964 	if (unlikely(stop_queue)) {
965 		/* If queue was emptied after the if (stop_queue) , and before
966 		 * the netif_tx_stop_queue() - need to wake the queue,
967 		 * or else it will remain stopped forever.
968 		 * Need a memory barrier to make sure ring->cons was not
969 		 * updated before queue was stopped.
970 		 */
971 		smp_rmb();
972 
973 		ring_cons = ACCESS_ONCE(ring->cons);
974 		if (unlikely(((int)(ring->prod - ring_cons)) <=
975 			     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
976 			netif_tx_wake_queue(ring->tx_queue);
977 			ring->wake_queue++;
978 		}
979 	}
980 	return NETDEV_TX_OK;
981 
982 tx_drop_unmap:
983 	en_err(priv, "DMA mapping error\n");
984 
985 	while (++i_frag < shinfo->nr_frags) {
986 		++data;
987 		dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
988 			       be32_to_cpu(data->byte_count),
989 			       PCI_DMA_TODEVICE);
990 	}
991 
992 tx_drop:
993 	dev_kfree_skb_any(skb);
994 	priv->stats.tx_dropped++;
995 	return NETDEV_TX_OK;
996 }
997 
998