xref: /linux/drivers/net/ethernet/mellanox/mlx4/en_tx.c (revision 0883c2c06fb5bcf5b9e008270827e63c09a88c1e)
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/ipv6.h>
45 #include <linux/moduleparam.h>
46 
47 #include "mlx4_en.h"
48 
49 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
50 			   struct mlx4_en_tx_ring **pring, u32 size,
51 			   u16 stride, int node, int queue_index)
52 {
53 	struct mlx4_en_dev *mdev = priv->mdev;
54 	struct mlx4_en_tx_ring *ring;
55 	int tmp;
56 	int err;
57 
58 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
59 	if (!ring) {
60 		ring = kzalloc(sizeof(*ring), GFP_KERNEL);
61 		if (!ring) {
62 			en_err(priv, "Failed allocating TX ring\n");
63 			return -ENOMEM;
64 		}
65 	}
66 
67 	ring->size = size;
68 	ring->size_mask = size - 1;
69 	ring->stride = stride;
70 	ring->full_size = ring->size - HEADROOM - MAX_DESC_TXBBS;
71 
72 	tmp = size * sizeof(struct mlx4_en_tx_info);
73 	ring->tx_info = kmalloc_node(tmp, GFP_KERNEL | __GFP_NOWARN, node);
74 	if (!ring->tx_info) {
75 		ring->tx_info = vmalloc(tmp);
76 		if (!ring->tx_info) {
77 			err = -ENOMEM;
78 			goto err_ring;
79 		}
80 	}
81 
82 	en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
83 		 ring->tx_info, tmp);
84 
85 	ring->bounce_buf = kmalloc_node(MAX_DESC_SIZE, GFP_KERNEL, node);
86 	if (!ring->bounce_buf) {
87 		ring->bounce_buf = kmalloc(MAX_DESC_SIZE, GFP_KERNEL);
88 		if (!ring->bounce_buf) {
89 			err = -ENOMEM;
90 			goto err_info;
91 		}
92 	}
93 	ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
94 
95 	/* Allocate HW buffers on provided NUMA node */
96 	set_dev_node(&mdev->dev->persist->pdev->dev, node);
97 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_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 	ring->buf = ring->wqres.buf.direct.buf;
105 
106 	en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d buf_size:%d dma:%llx\n",
107 	       ring, ring->buf, ring->size, ring->buf_size,
108 	       (unsigned long long) ring->wqres.buf.direct.map);
109 
110 	err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &ring->qpn,
111 				    MLX4_RESERVE_ETH_BF_QP);
112 	if (err) {
113 		en_err(priv, "failed reserving qp for TX ring\n");
114 		goto err_hwq_res;
115 	}
116 
117 	err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp, GFP_KERNEL);
118 	if (err) {
119 		en_err(priv, "Failed allocating qp %d\n", ring->qpn);
120 		goto err_reserve;
121 	}
122 	ring->qp.event = mlx4_en_sqp_event;
123 
124 	err = mlx4_bf_alloc(mdev->dev, &ring->bf, node);
125 	if (err) {
126 		en_dbg(DRV, priv, "working without blueflame (%d)\n", err);
127 		ring->bf.uar = &mdev->priv_uar;
128 		ring->bf.uar->map = mdev->uar_map;
129 		ring->bf_enabled = false;
130 		ring->bf_alloced = false;
131 		priv->pflags &= ~MLX4_EN_PRIV_FLAGS_BLUEFLAME;
132 	} else {
133 		ring->bf_alloced = true;
134 		ring->bf_enabled = !!(priv->pflags &
135 				      MLX4_EN_PRIV_FLAGS_BLUEFLAME);
136 	}
137 
138 	ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type;
139 	ring->queue_index = queue_index;
140 
141 	if (queue_index < priv->num_tx_rings_p_up)
142 		cpumask_set_cpu(cpumask_local_spread(queue_index,
143 						     priv->mdev->dev->numa_node),
144 				&ring->affinity_mask);
145 
146 	*pring = ring;
147 	return 0;
148 
149 err_reserve:
150 	mlx4_qp_release_range(mdev->dev, ring->qpn, 1);
151 err_hwq_res:
152 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
153 err_bounce:
154 	kfree(ring->bounce_buf);
155 	ring->bounce_buf = NULL;
156 err_info:
157 	kvfree(ring->tx_info);
158 	ring->tx_info = NULL;
159 err_ring:
160 	kfree(ring);
161 	*pring = NULL;
162 	return err;
163 }
164 
165 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
166 			     struct mlx4_en_tx_ring **pring)
167 {
168 	struct mlx4_en_dev *mdev = priv->mdev;
169 	struct mlx4_en_tx_ring *ring = *pring;
170 	en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
171 
172 	if (ring->bf_alloced)
173 		mlx4_bf_free(mdev->dev, &ring->bf);
174 	mlx4_qp_remove(mdev->dev, &ring->qp);
175 	mlx4_qp_free(mdev->dev, &ring->qp);
176 	mlx4_qp_release_range(priv->mdev->dev, ring->qpn, 1);
177 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
178 	kfree(ring->bounce_buf);
179 	ring->bounce_buf = NULL;
180 	kvfree(ring->tx_info);
181 	ring->tx_info = NULL;
182 	kfree(ring);
183 	*pring = NULL;
184 }
185 
186 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
187 			     struct mlx4_en_tx_ring *ring,
188 			     int cq, int user_prio)
189 {
190 	struct mlx4_en_dev *mdev = priv->mdev;
191 	int err;
192 
193 	ring->cqn = cq;
194 	ring->prod = 0;
195 	ring->cons = 0xffffffff;
196 	ring->last_nr_txbb = 1;
197 	memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
198 	memset(ring->buf, 0, ring->buf_size);
199 
200 	ring->qp_state = MLX4_QP_STATE_RST;
201 	ring->doorbell_qpn = cpu_to_be32(ring->qp.qpn << 8);
202 	ring->mr_key = cpu_to_be32(mdev->mr.key);
203 
204 	mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
205 				ring->cqn, user_prio, &ring->context);
206 	if (ring->bf_alloced)
207 		ring->context.usr_page =
208 			cpu_to_be32(mlx4_to_hw_uar_index(mdev->dev,
209 							 ring->bf.uar->index));
210 
211 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
212 			       &ring->qp, &ring->qp_state);
213 	if (!cpumask_empty(&ring->affinity_mask))
214 		netif_set_xps_queue(priv->dev, &ring->affinity_mask,
215 				    ring->queue_index);
216 
217 	return err;
218 }
219 
220 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
221 				struct mlx4_en_tx_ring *ring)
222 {
223 	struct mlx4_en_dev *mdev = priv->mdev;
224 
225 	mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
226 		       MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
227 }
228 
229 static inline bool mlx4_en_is_tx_ring_full(struct mlx4_en_tx_ring *ring)
230 {
231 	return ring->prod - ring->cons > ring->full_size;
232 }
233 
234 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
235 			      struct mlx4_en_tx_ring *ring, int index,
236 			      u8 owner)
237 {
238 	__be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
239 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
240 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
241 	void *end = ring->buf + ring->buf_size;
242 	__be32 *ptr = (__be32 *)tx_desc;
243 	int i;
244 
245 	/* Optimize the common case when there are no wraparounds */
246 	if (likely((void *)tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
247 		/* Stamp the freed descriptor */
248 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
249 		     i += STAMP_STRIDE) {
250 			*ptr = stamp;
251 			ptr += STAMP_DWORDS;
252 		}
253 	} else {
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 			if ((void *)ptr >= end) {
260 				ptr = ring->buf;
261 				stamp ^= cpu_to_be32(0x80000000);
262 			}
263 		}
264 	}
265 }
266 
267 
268 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
269 				struct mlx4_en_tx_ring *ring,
270 				int index, u8 owner, u64 timestamp,
271 				int napi_mode)
272 {
273 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
274 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
275 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
276 	void *end = ring->buf + ring->buf_size;
277 	struct sk_buff *skb = tx_info->skb;
278 	int nr_maps = tx_info->nr_maps;
279 	int i;
280 
281 	/* We do not touch skb here, so prefetch skb->users location
282 	 * to speedup consume_skb()
283 	 */
284 	prefetchw(&skb->users);
285 
286 	if (unlikely(timestamp)) {
287 		struct skb_shared_hwtstamps hwts;
288 
289 		mlx4_en_fill_hwtstamps(priv->mdev, &hwts, timestamp);
290 		skb_tstamp_tx(skb, &hwts);
291 	}
292 
293 	/* Optimize the common case when there are no wraparounds */
294 	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
295 		if (!tx_info->inl) {
296 			if (tx_info->linear)
297 				dma_unmap_single(priv->ddev,
298 						tx_info->map0_dma,
299 						tx_info->map0_byte_count,
300 						PCI_DMA_TODEVICE);
301 			else
302 				dma_unmap_page(priv->ddev,
303 					       tx_info->map0_dma,
304 					       tx_info->map0_byte_count,
305 					       PCI_DMA_TODEVICE);
306 			for (i = 1; i < nr_maps; i++) {
307 				data++;
308 				dma_unmap_page(priv->ddev,
309 					(dma_addr_t)be64_to_cpu(data->addr),
310 					be32_to_cpu(data->byte_count),
311 					PCI_DMA_TODEVICE);
312 			}
313 		}
314 	} else {
315 		if (!tx_info->inl) {
316 			if ((void *) data >= end) {
317 				data = ring->buf + ((void *)data - end);
318 			}
319 
320 			if (tx_info->linear)
321 				dma_unmap_single(priv->ddev,
322 						tx_info->map0_dma,
323 						tx_info->map0_byte_count,
324 						PCI_DMA_TODEVICE);
325 			else
326 				dma_unmap_page(priv->ddev,
327 					       tx_info->map0_dma,
328 					       tx_info->map0_byte_count,
329 					       PCI_DMA_TODEVICE);
330 			for (i = 1; i < nr_maps; i++) {
331 				data++;
332 				/* Check for wraparound before unmapping */
333 				if ((void *) data >= end)
334 					data = ring->buf;
335 				dma_unmap_page(priv->ddev,
336 					(dma_addr_t)be64_to_cpu(data->addr),
337 					be32_to_cpu(data->byte_count),
338 					PCI_DMA_TODEVICE);
339 			}
340 		}
341 	}
342 	napi_consume_skb(skb, napi_mode);
343 
344 	return tx_info->nr_txbb;
345 }
346 
347 
348 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
349 {
350 	struct mlx4_en_priv *priv = netdev_priv(dev);
351 	int cnt = 0;
352 
353 	/* Skip last polled descriptor */
354 	ring->cons += ring->last_nr_txbb;
355 	en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
356 		 ring->cons, ring->prod);
357 
358 	if ((u32) (ring->prod - ring->cons) > ring->size) {
359 		if (netif_msg_tx_err(priv))
360 			en_warn(priv, "Tx consumer passed producer!\n");
361 		return 0;
362 	}
363 
364 	while (ring->cons != ring->prod) {
365 		ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
366 						ring->cons & ring->size_mask,
367 						!!(ring->cons & ring->size), 0,
368 						0 /* Non-NAPI caller */);
369 		ring->cons += ring->last_nr_txbb;
370 		cnt++;
371 	}
372 
373 	netdev_tx_reset_queue(ring->tx_queue);
374 
375 	if (cnt)
376 		en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
377 
378 	return cnt;
379 }
380 
381 static bool mlx4_en_process_tx_cq(struct net_device *dev,
382 				  struct mlx4_en_cq *cq, int napi_budget)
383 {
384 	struct mlx4_en_priv *priv = netdev_priv(dev);
385 	struct mlx4_cq *mcq = &cq->mcq;
386 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->ring];
387 	struct mlx4_cqe *cqe;
388 	u16 index;
389 	u16 new_index, ring_index, stamp_index;
390 	u32 txbbs_skipped = 0;
391 	u32 txbbs_stamp = 0;
392 	u32 cons_index = mcq->cons_index;
393 	int size = cq->size;
394 	u32 size_mask = ring->size_mask;
395 	struct mlx4_cqe *buf = cq->buf;
396 	u32 packets = 0;
397 	u32 bytes = 0;
398 	int factor = priv->cqe_factor;
399 	int done = 0;
400 	int budget = priv->tx_work_limit;
401 	u32 last_nr_txbb;
402 	u32 ring_cons;
403 
404 	if (!priv->port_up)
405 		return true;
406 
407 	netdev_txq_bql_complete_prefetchw(ring->tx_queue);
408 
409 	index = cons_index & size_mask;
410 	cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
411 	last_nr_txbb = ACCESS_ONCE(ring->last_nr_txbb);
412 	ring_cons = ACCESS_ONCE(ring->cons);
413 	ring_index = ring_cons & size_mask;
414 	stamp_index = ring_index;
415 
416 	/* Process all completed CQEs */
417 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
418 			cons_index & size) && (done < budget)) {
419 		/*
420 		 * make sure we read the CQE after we read the
421 		 * ownership bit
422 		 */
423 		dma_rmb();
424 
425 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
426 			     MLX4_CQE_OPCODE_ERROR)) {
427 			struct mlx4_err_cqe *cqe_err = (struct mlx4_err_cqe *)cqe;
428 
429 			en_err(priv, "CQE error - vendor syndrome: 0x%x syndrome: 0x%x\n",
430 			       cqe_err->vendor_err_syndrome,
431 			       cqe_err->syndrome);
432 		}
433 
434 		/* Skip over last polled CQE */
435 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
436 
437 		do {
438 			u64 timestamp = 0;
439 
440 			txbbs_skipped += last_nr_txbb;
441 			ring_index = (ring_index + last_nr_txbb) & size_mask;
442 
443 			if (unlikely(ring->tx_info[ring_index].ts_requested))
444 				timestamp = mlx4_en_get_cqe_ts(cqe);
445 
446 			/* free next descriptor */
447 			last_nr_txbb = mlx4_en_free_tx_desc(
448 					priv, ring, ring_index,
449 					!!((ring_cons + txbbs_skipped) &
450 					ring->size), timestamp, napi_budget);
451 
452 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
453 					  !!((ring_cons + txbbs_stamp) &
454 						ring->size));
455 			stamp_index = ring_index;
456 			txbbs_stamp = txbbs_skipped;
457 			packets++;
458 			bytes += ring->tx_info[ring_index].nr_bytes;
459 		} while ((++done < budget) && (ring_index != new_index));
460 
461 		++cons_index;
462 		index = cons_index & size_mask;
463 		cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
464 	}
465 
466 
467 	/*
468 	 * To prevent CQ overflow we first update CQ consumer and only then
469 	 * the ring consumer.
470 	 */
471 	mcq->cons_index = cons_index;
472 	mlx4_cq_set_ci(mcq);
473 	wmb();
474 
475 	/* we want to dirty this cache line once */
476 	ACCESS_ONCE(ring->last_nr_txbb) = last_nr_txbb;
477 	ACCESS_ONCE(ring->cons) = ring_cons + txbbs_skipped;
478 
479 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
480 
481 	/* Wakeup Tx queue if this stopped, and ring is not full.
482 	 */
483 	if (netif_tx_queue_stopped(ring->tx_queue) &&
484 	    !mlx4_en_is_tx_ring_full(ring)) {
485 		netif_tx_wake_queue(ring->tx_queue);
486 		ring->wake_queue++;
487 	}
488 	return done < budget;
489 }
490 
491 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
492 {
493 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
494 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
495 
496 	if (likely(priv->port_up))
497 		napi_schedule_irqoff(&cq->napi);
498 	else
499 		mlx4_en_arm_cq(priv, cq);
500 }
501 
502 /* TX CQ polling - called by NAPI */
503 int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget)
504 {
505 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
506 	struct net_device *dev = cq->dev;
507 	struct mlx4_en_priv *priv = netdev_priv(dev);
508 	int clean_complete;
509 
510 	clean_complete = mlx4_en_process_tx_cq(dev, cq, budget);
511 	if (!clean_complete)
512 		return budget;
513 
514 	napi_complete(napi);
515 	mlx4_en_arm_cq(priv, cq);
516 
517 	return 0;
518 }
519 
520 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
521 						      struct mlx4_en_tx_ring *ring,
522 						      u32 index,
523 						      unsigned int desc_size)
524 {
525 	u32 copy = (ring->size - index) * TXBB_SIZE;
526 	int i;
527 
528 	for (i = desc_size - copy - 4; i >= 0; i -= 4) {
529 		if ((i & (TXBB_SIZE - 1)) == 0)
530 			wmb();
531 
532 		*((u32 *) (ring->buf + i)) =
533 			*((u32 *) (ring->bounce_buf + copy + i));
534 	}
535 
536 	for (i = copy - 4; i >= 4 ; i -= 4) {
537 		if ((i & (TXBB_SIZE - 1)) == 0)
538 			wmb();
539 
540 		*((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
541 			*((u32 *) (ring->bounce_buf + i));
542 	}
543 
544 	/* Return real descriptor location */
545 	return ring->buf + index * TXBB_SIZE;
546 }
547 
548 /* Decide if skb can be inlined in tx descriptor to avoid dma mapping
549  *
550  * It seems strange we do not simply use skb_copy_bits().
551  * This would allow to inline all skbs iff skb->len <= inline_thold
552  *
553  * Note that caller already checked skb was not a gso packet
554  */
555 static bool is_inline(int inline_thold, const struct sk_buff *skb,
556 		      const struct skb_shared_info *shinfo,
557 		      void **pfrag)
558 {
559 	void *ptr;
560 
561 	if (skb->len > inline_thold || !inline_thold)
562 		return false;
563 
564 	if (shinfo->nr_frags == 1) {
565 		ptr = skb_frag_address_safe(&shinfo->frags[0]);
566 		if (unlikely(!ptr))
567 			return false;
568 		*pfrag = ptr;
569 		return true;
570 	}
571 	if (shinfo->nr_frags)
572 		return false;
573 	return true;
574 }
575 
576 static int inline_size(const struct sk_buff *skb)
577 {
578 	if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
579 	    <= MLX4_INLINE_ALIGN)
580 		return ALIGN(skb->len + CTRL_SIZE +
581 			     sizeof(struct mlx4_wqe_inline_seg), 16);
582 	else
583 		return ALIGN(skb->len + CTRL_SIZE + 2 *
584 			     sizeof(struct mlx4_wqe_inline_seg), 16);
585 }
586 
587 static int get_real_size(const struct sk_buff *skb,
588 			 const struct skb_shared_info *shinfo,
589 			 struct net_device *dev,
590 			 int *lso_header_size,
591 			 bool *inline_ok,
592 			 void **pfrag)
593 {
594 	struct mlx4_en_priv *priv = netdev_priv(dev);
595 	int real_size;
596 
597 	if (shinfo->gso_size) {
598 		*inline_ok = false;
599 		if (skb->encapsulation)
600 			*lso_header_size = (skb_inner_transport_header(skb) - skb->data) + inner_tcp_hdrlen(skb);
601 		else
602 			*lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
603 		real_size = CTRL_SIZE + shinfo->nr_frags * DS_SIZE +
604 			ALIGN(*lso_header_size + 4, DS_SIZE);
605 		if (unlikely(*lso_header_size != skb_headlen(skb))) {
606 			/* We add a segment for the skb linear buffer only if
607 			 * it contains data */
608 			if (*lso_header_size < skb_headlen(skb))
609 				real_size += DS_SIZE;
610 			else {
611 				if (netif_msg_tx_err(priv))
612 					en_warn(priv, "Non-linear headers\n");
613 				return 0;
614 			}
615 		}
616 	} else {
617 		*lso_header_size = 0;
618 		*inline_ok = is_inline(priv->prof->inline_thold, skb,
619 				       shinfo, pfrag);
620 
621 		if (*inline_ok)
622 			real_size = inline_size(skb);
623 		else
624 			real_size = CTRL_SIZE +
625 				    (shinfo->nr_frags + 1) * DS_SIZE;
626 	}
627 
628 	return real_size;
629 }
630 
631 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
632 			     const struct sk_buff *skb,
633 			     const struct skb_shared_info *shinfo,
634 			     int real_size, u16 *vlan_tag,
635 			     int tx_ind, void *fragptr)
636 {
637 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
638 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
639 	unsigned int hlen = skb_headlen(skb);
640 
641 	if (skb->len <= spc) {
642 		if (likely(skb->len >= MIN_PKT_LEN)) {
643 			inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
644 		} else {
645 			inl->byte_count = cpu_to_be32(1 << 31 | MIN_PKT_LEN);
646 			memset(((void *)(inl + 1)) + skb->len, 0,
647 			       MIN_PKT_LEN - skb->len);
648 		}
649 		skb_copy_from_linear_data(skb, inl + 1, hlen);
650 		if (shinfo->nr_frags)
651 			memcpy(((void *)(inl + 1)) + hlen, fragptr,
652 			       skb_frag_size(&shinfo->frags[0]));
653 
654 	} else {
655 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
656 		if (hlen <= spc) {
657 			skb_copy_from_linear_data(skb, inl + 1, hlen);
658 			if (hlen < spc) {
659 				memcpy(((void *)(inl + 1)) + hlen,
660 				       fragptr, spc - hlen);
661 				fragptr +=  spc - hlen;
662 			}
663 			inl = (void *) (inl + 1) + spc;
664 			memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
665 		} else {
666 			skb_copy_from_linear_data(skb, inl + 1, spc);
667 			inl = (void *) (inl + 1) + spc;
668 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
669 							 hlen - spc);
670 			if (shinfo->nr_frags)
671 				memcpy(((void *)(inl + 1)) + hlen - spc,
672 				       fragptr,
673 				       skb_frag_size(&shinfo->frags[0]));
674 		}
675 
676 		dma_wmb();
677 		inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
678 	}
679 }
680 
681 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb,
682 			 void *accel_priv, select_queue_fallback_t fallback)
683 {
684 	struct mlx4_en_priv *priv = netdev_priv(dev);
685 	u16 rings_p_up = priv->num_tx_rings_p_up;
686 	u8 up = 0;
687 
688 	if (dev->num_tc)
689 		return skb_tx_hash(dev, skb);
690 
691 	if (skb_vlan_tag_present(skb))
692 		up = skb_vlan_tag_get(skb) >> VLAN_PRIO_SHIFT;
693 
694 	return fallback(dev, skb) % rings_p_up + up * rings_p_up;
695 }
696 
697 static void mlx4_bf_copy(void __iomem *dst, const void *src,
698 			 unsigned int bytecnt)
699 {
700 	__iowrite64_copy(dst, src, bytecnt / 8);
701 }
702 
703 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
704 {
705 	struct skb_shared_info *shinfo = skb_shinfo(skb);
706 	struct mlx4_en_priv *priv = netdev_priv(dev);
707 	struct device *ddev = priv->ddev;
708 	struct mlx4_en_tx_ring *ring;
709 	struct mlx4_en_tx_desc *tx_desc;
710 	struct mlx4_wqe_data_seg *data;
711 	struct mlx4_en_tx_info *tx_info;
712 	int tx_ind = 0;
713 	int nr_txbb;
714 	int desc_size;
715 	int real_size;
716 	u32 index, bf_index;
717 	__be32 op_own;
718 	u16 vlan_tag = 0;
719 	u16 vlan_proto = 0;
720 	int i_frag;
721 	int lso_header_size;
722 	void *fragptr = NULL;
723 	bool bounce = false;
724 	bool send_doorbell;
725 	bool stop_queue;
726 	bool inline_ok;
727 	u32 ring_cons;
728 
729 	tx_ind = skb_get_queue_mapping(skb);
730 	ring = priv->tx_ring[tx_ind];
731 
732 	if (!priv->port_up)
733 		goto tx_drop;
734 
735 	/* fetch ring->cons far ahead before needing it to avoid stall */
736 	ring_cons = ACCESS_ONCE(ring->cons);
737 
738 	real_size = get_real_size(skb, shinfo, dev, &lso_header_size,
739 				  &inline_ok, &fragptr);
740 	if (unlikely(!real_size))
741 		goto tx_drop;
742 
743 	/* Align descriptor to TXBB size */
744 	desc_size = ALIGN(real_size, TXBB_SIZE);
745 	nr_txbb = desc_size / TXBB_SIZE;
746 	if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
747 		if (netif_msg_tx_err(priv))
748 			en_warn(priv, "Oversized header or SG list\n");
749 		goto tx_drop;
750 	}
751 
752 	if (skb_vlan_tag_present(skb)) {
753 		vlan_tag = skb_vlan_tag_get(skb);
754 		vlan_proto = be16_to_cpu(skb->vlan_proto);
755 	}
756 
757 	netdev_txq_bql_enqueue_prefetchw(ring->tx_queue);
758 
759 	/* Track current inflight packets for performance analysis */
760 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
761 			 (u32)(ring->prod - ring_cons - 1));
762 
763 	/* Packet is good - grab an index and transmit it */
764 	index = ring->prod & ring->size_mask;
765 	bf_index = ring->prod;
766 
767 	/* See if we have enough space for whole descriptor TXBB for setting
768 	 * SW ownership on next descriptor; if not, use a bounce buffer. */
769 	if (likely(index + nr_txbb <= ring->size))
770 		tx_desc = ring->buf + index * TXBB_SIZE;
771 	else {
772 		tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
773 		bounce = true;
774 	}
775 
776 	/* Save skb in tx_info ring */
777 	tx_info = &ring->tx_info[index];
778 	tx_info->skb = skb;
779 	tx_info->nr_txbb = nr_txbb;
780 
781 	data = &tx_desc->data;
782 	if (lso_header_size)
783 		data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
784 						      DS_SIZE));
785 
786 	/* valid only for none inline segments */
787 	tx_info->data_offset = (void *)data - (void *)tx_desc;
788 
789 	tx_info->inl = inline_ok;
790 
791 	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
792 			   !inline_ok) ? 1 : 0;
793 
794 	tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
795 	data += tx_info->nr_maps - 1;
796 
797 	if (!tx_info->inl) {
798 		dma_addr_t dma = 0;
799 		u32 byte_count = 0;
800 
801 		/* Map fragments if any */
802 		for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
803 			const struct skb_frag_struct *frag;
804 
805 			frag = &shinfo->frags[i_frag];
806 			byte_count = skb_frag_size(frag);
807 			dma = skb_frag_dma_map(ddev, frag,
808 					       0, byte_count,
809 					       DMA_TO_DEVICE);
810 			if (dma_mapping_error(ddev, dma))
811 				goto tx_drop_unmap;
812 
813 			data->addr = cpu_to_be64(dma);
814 			data->lkey = ring->mr_key;
815 			dma_wmb();
816 			data->byte_count = cpu_to_be32(byte_count);
817 			--data;
818 		}
819 
820 		/* Map linear part if needed */
821 		if (tx_info->linear) {
822 			byte_count = skb_headlen(skb) - lso_header_size;
823 
824 			dma = dma_map_single(ddev, skb->data +
825 					     lso_header_size, byte_count,
826 					     PCI_DMA_TODEVICE);
827 			if (dma_mapping_error(ddev, dma))
828 				goto tx_drop_unmap;
829 
830 			data->addr = cpu_to_be64(dma);
831 			data->lkey = ring->mr_key;
832 			dma_wmb();
833 			data->byte_count = cpu_to_be32(byte_count);
834 		}
835 		/* tx completion can avoid cache line miss for common cases */
836 		tx_info->map0_dma = dma;
837 		tx_info->map0_byte_count = byte_count;
838 	}
839 
840 	/*
841 	 * For timestamping add flag to skb_shinfo and
842 	 * set flag for further reference
843 	 */
844 	tx_info->ts_requested = 0;
845 	if (unlikely(ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
846 		     shinfo->tx_flags & SKBTX_HW_TSTAMP)) {
847 		shinfo->tx_flags |= SKBTX_IN_PROGRESS;
848 		tx_info->ts_requested = 1;
849 	}
850 
851 	/* Prepare ctrl segement apart opcode+ownership, which depends on
852 	 * whether LSO is used */
853 	tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
854 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
855 		if (!skb->encapsulation)
856 			tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
857 								 MLX4_WQE_CTRL_TCP_UDP_CSUM);
858 		else
859 			tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM);
860 		ring->tx_csum++;
861 	}
862 
863 	if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
864 		struct ethhdr *ethh;
865 
866 		/* Copy dst mac address to wqe. This allows loopback in eSwitch,
867 		 * so that VFs and PF can communicate with each other
868 		 */
869 		ethh = (struct ethhdr *)skb->data;
870 		tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
871 		tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
872 	}
873 
874 	/* Handle LSO (TSO) packets */
875 	if (lso_header_size) {
876 		int i;
877 
878 		/* Mark opcode as LSO */
879 		op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
880 			((ring->prod & ring->size) ?
881 				cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
882 
883 		/* Fill in the LSO prefix */
884 		tx_desc->lso.mss_hdr_size = cpu_to_be32(
885 			shinfo->gso_size << 16 | lso_header_size);
886 
887 		/* Copy headers;
888 		 * note that we already verified that it is linear */
889 		memcpy(tx_desc->lso.header, skb->data, lso_header_size);
890 
891 		ring->tso_packets++;
892 
893 		i = ((skb->len - lso_header_size) / shinfo->gso_size) +
894 			!!((skb->len - lso_header_size) % shinfo->gso_size);
895 		tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
896 		ring->packets += i;
897 	} else {
898 		/* Normal (Non LSO) packet */
899 		op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
900 			((ring->prod & ring->size) ?
901 			 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
902 		tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
903 		ring->packets++;
904 	}
905 	ring->bytes += tx_info->nr_bytes;
906 	netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
907 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
908 
909 	if (tx_info->inl)
910 		build_inline_wqe(tx_desc, skb, shinfo, real_size, &vlan_tag,
911 				 tx_ind, fragptr);
912 
913 	if (skb->encapsulation) {
914 		union {
915 			struct iphdr *v4;
916 			struct ipv6hdr *v6;
917 			unsigned char *hdr;
918 		} ip;
919 		u8 proto;
920 
921 		ip.hdr = skb_inner_network_header(skb);
922 		proto = (ip.v4->version == 4) ? ip.v4->protocol :
923 						ip.v6->nexthdr;
924 
925 		if (proto == IPPROTO_TCP || proto == IPPROTO_UDP)
926 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP | MLX4_WQE_CTRL_ILP);
927 		else
928 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP);
929 	}
930 
931 	ring->prod += nr_txbb;
932 
933 	/* If we used a bounce buffer then copy descriptor back into place */
934 	if (unlikely(bounce))
935 		tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
936 
937 	skb_tx_timestamp(skb);
938 
939 	/* Check available TXBBs And 2K spare for prefetch */
940 	stop_queue = mlx4_en_is_tx_ring_full(ring);
941 	if (unlikely(stop_queue)) {
942 		netif_tx_stop_queue(ring->tx_queue);
943 		ring->queue_stopped++;
944 	}
945 	send_doorbell = !skb->xmit_more || netif_xmit_stopped(ring->tx_queue);
946 
947 	real_size = (real_size / 16) & 0x3f;
948 
949 	if (ring->bf_enabled && desc_size <= MAX_BF && !bounce &&
950 	    !skb_vlan_tag_present(skb) && send_doorbell) {
951 		tx_desc->ctrl.bf_qpn = ring->doorbell_qpn |
952 				       cpu_to_be32(real_size);
953 
954 		op_own |= htonl((bf_index & 0xffff) << 8);
955 		/* Ensure new descriptor hits memory
956 		 * before setting ownership of this descriptor to HW
957 		 */
958 		dma_wmb();
959 		tx_desc->ctrl.owner_opcode = op_own;
960 
961 		wmb();
962 
963 		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, &tx_desc->ctrl,
964 			     desc_size);
965 
966 		wmb();
967 
968 		ring->bf.offset ^= ring->bf.buf_size;
969 	} else {
970 		tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
971 		if (vlan_proto == ETH_P_8021AD)
972 			tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_SVLAN;
973 		else if (vlan_proto == ETH_P_8021Q)
974 			tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_CVLAN;
975 		else
976 			tx_desc->ctrl.ins_vlan = 0;
977 
978 		tx_desc->ctrl.fence_size = real_size;
979 
980 		/* Ensure new descriptor hits memory
981 		 * before setting ownership of this descriptor to HW
982 		 */
983 		dma_wmb();
984 		tx_desc->ctrl.owner_opcode = op_own;
985 		if (send_doorbell) {
986 			wmb();
987 			/* Since there is no iowrite*_native() that writes the
988 			 * value as is, without byteswapping - using the one
989 			 * the doesn't do byteswapping in the relevant arch
990 			 * endianness.
991 			 */
992 #if defined(__LITTLE_ENDIAN)
993 			iowrite32(
994 #else
995 			iowrite32be(
996 #endif
997 				  ring->doorbell_qpn,
998 				  ring->bf.uar->map + MLX4_SEND_DOORBELL);
999 		} else {
1000 			ring->xmit_more++;
1001 		}
1002 	}
1003 
1004 	if (unlikely(stop_queue)) {
1005 		/* If queue was emptied after the if (stop_queue) , and before
1006 		 * the netif_tx_stop_queue() - need to wake the queue,
1007 		 * or else it will remain stopped forever.
1008 		 * Need a memory barrier to make sure ring->cons was not
1009 		 * updated before queue was stopped.
1010 		 */
1011 		smp_rmb();
1012 
1013 		ring_cons = ACCESS_ONCE(ring->cons);
1014 		if (unlikely(!mlx4_en_is_tx_ring_full(ring))) {
1015 			netif_tx_wake_queue(ring->tx_queue);
1016 			ring->wake_queue++;
1017 		}
1018 	}
1019 	return NETDEV_TX_OK;
1020 
1021 tx_drop_unmap:
1022 	en_err(priv, "DMA mapping error\n");
1023 
1024 	while (++i_frag < shinfo->nr_frags) {
1025 		++data;
1026 		dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
1027 			       be32_to_cpu(data->byte_count),
1028 			       PCI_DMA_TODEVICE);
1029 	}
1030 
1031 tx_drop:
1032 	dev_kfree_skb_any(skb);
1033 	ring->tx_dropped++;
1034 	return NETDEV_TX_OK;
1035 }
1036 
1037