xref: /linux/drivers/net/ethernet/mellanox/mlx4/en_tx.c (revision 110e6f26af80dfd90b6e5c645b1aed7228aa580d)
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 =
217 			cpu_to_be32(mlx4_to_hw_uar_index(mdev->dev,
218 							 ring->bf.uar->index));
219 
220 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
221 			       &ring->qp, &ring->qp_state);
222 	if (!cpumask_empty(&ring->affinity_mask))
223 		netif_set_xps_queue(priv->dev, &ring->affinity_mask,
224 				    ring->queue_index);
225 
226 	return err;
227 }
228 
229 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
230 				struct mlx4_en_tx_ring *ring)
231 {
232 	struct mlx4_en_dev *mdev = priv->mdev;
233 
234 	mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
235 		       MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
236 }
237 
238 static inline bool mlx4_en_is_tx_ring_full(struct mlx4_en_tx_ring *ring)
239 {
240 	return ring->prod - ring->cons > ring->full_size;
241 }
242 
243 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
244 			      struct mlx4_en_tx_ring *ring, int index,
245 			      u8 owner)
246 {
247 	__be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
248 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
249 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
250 	void *end = ring->buf + ring->buf_size;
251 	__be32 *ptr = (__be32 *)tx_desc;
252 	int i;
253 
254 	/* Optimize the common case when there are no wraparounds */
255 	if (likely((void *)tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
256 		/* Stamp the freed descriptor */
257 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
258 		     i += STAMP_STRIDE) {
259 			*ptr = stamp;
260 			ptr += STAMP_DWORDS;
261 		}
262 	} else {
263 		/* Stamp the freed descriptor */
264 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
265 		     i += STAMP_STRIDE) {
266 			*ptr = stamp;
267 			ptr += STAMP_DWORDS;
268 			if ((void *)ptr >= end) {
269 				ptr = ring->buf;
270 				stamp ^= cpu_to_be32(0x80000000);
271 			}
272 		}
273 	}
274 }
275 
276 
277 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
278 				struct mlx4_en_tx_ring *ring,
279 				int index, u8 owner, u64 timestamp,
280 				int napi_mode)
281 {
282 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
283 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
284 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
285 	void *end = ring->buf + ring->buf_size;
286 	struct sk_buff *skb = tx_info->skb;
287 	int nr_maps = tx_info->nr_maps;
288 	int i;
289 
290 	/* We do not touch skb here, so prefetch skb->users location
291 	 * to speedup consume_skb()
292 	 */
293 	prefetchw(&skb->users);
294 
295 	if (unlikely(timestamp)) {
296 		struct skb_shared_hwtstamps hwts;
297 
298 		mlx4_en_fill_hwtstamps(priv->mdev, &hwts, timestamp);
299 		skb_tstamp_tx(skb, &hwts);
300 	}
301 
302 	/* Optimize the common case when there are no wraparounds */
303 	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
304 		if (!tx_info->inl) {
305 			if (tx_info->linear)
306 				dma_unmap_single(priv->ddev,
307 						tx_info->map0_dma,
308 						tx_info->map0_byte_count,
309 						PCI_DMA_TODEVICE);
310 			else
311 				dma_unmap_page(priv->ddev,
312 					       tx_info->map0_dma,
313 					       tx_info->map0_byte_count,
314 					       PCI_DMA_TODEVICE);
315 			for (i = 1; i < nr_maps; i++) {
316 				data++;
317 				dma_unmap_page(priv->ddev,
318 					(dma_addr_t)be64_to_cpu(data->addr),
319 					be32_to_cpu(data->byte_count),
320 					PCI_DMA_TODEVICE);
321 			}
322 		}
323 	} else {
324 		if (!tx_info->inl) {
325 			if ((void *) data >= end) {
326 				data = ring->buf + ((void *)data - end);
327 			}
328 
329 			if (tx_info->linear)
330 				dma_unmap_single(priv->ddev,
331 						tx_info->map0_dma,
332 						tx_info->map0_byte_count,
333 						PCI_DMA_TODEVICE);
334 			else
335 				dma_unmap_page(priv->ddev,
336 					       tx_info->map0_dma,
337 					       tx_info->map0_byte_count,
338 					       PCI_DMA_TODEVICE);
339 			for (i = 1; i < nr_maps; i++) {
340 				data++;
341 				/* Check for wraparound before unmapping */
342 				if ((void *) data >= end)
343 					data = ring->buf;
344 				dma_unmap_page(priv->ddev,
345 					(dma_addr_t)be64_to_cpu(data->addr),
346 					be32_to_cpu(data->byte_count),
347 					PCI_DMA_TODEVICE);
348 			}
349 		}
350 	}
351 	napi_consume_skb(skb, napi_mode);
352 
353 	return tx_info->nr_txbb;
354 }
355 
356 
357 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
358 {
359 	struct mlx4_en_priv *priv = netdev_priv(dev);
360 	int cnt = 0;
361 
362 	/* Skip last polled descriptor */
363 	ring->cons += ring->last_nr_txbb;
364 	en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
365 		 ring->cons, ring->prod);
366 
367 	if ((u32) (ring->prod - ring->cons) > ring->size) {
368 		if (netif_msg_tx_err(priv))
369 			en_warn(priv, "Tx consumer passed producer!\n");
370 		return 0;
371 	}
372 
373 	while (ring->cons != ring->prod) {
374 		ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
375 						ring->cons & ring->size_mask,
376 						!!(ring->cons & ring->size), 0,
377 						0 /* Non-NAPI caller */);
378 		ring->cons += ring->last_nr_txbb;
379 		cnt++;
380 	}
381 
382 	netdev_tx_reset_queue(ring->tx_queue);
383 
384 	if (cnt)
385 		en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
386 
387 	return cnt;
388 }
389 
390 static bool mlx4_en_process_tx_cq(struct net_device *dev,
391 				  struct mlx4_en_cq *cq, int napi_budget)
392 {
393 	struct mlx4_en_priv *priv = netdev_priv(dev);
394 	struct mlx4_cq *mcq = &cq->mcq;
395 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->ring];
396 	struct mlx4_cqe *cqe;
397 	u16 index;
398 	u16 new_index, ring_index, stamp_index;
399 	u32 txbbs_skipped = 0;
400 	u32 txbbs_stamp = 0;
401 	u32 cons_index = mcq->cons_index;
402 	int size = cq->size;
403 	u32 size_mask = ring->size_mask;
404 	struct mlx4_cqe *buf = cq->buf;
405 	u32 packets = 0;
406 	u32 bytes = 0;
407 	int factor = priv->cqe_factor;
408 	u64 timestamp = 0;
409 	int done = 0;
410 	int budget = priv->tx_work_limit;
411 	u32 last_nr_txbb;
412 	u32 ring_cons;
413 
414 	if (!priv->port_up)
415 		return true;
416 
417 	netdev_txq_bql_complete_prefetchw(ring->tx_queue);
418 
419 	index = cons_index & size_mask;
420 	cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
421 	last_nr_txbb = ACCESS_ONCE(ring->last_nr_txbb);
422 	ring_cons = ACCESS_ONCE(ring->cons);
423 	ring_index = ring_cons & size_mask;
424 	stamp_index = ring_index;
425 
426 	/* Process all completed CQEs */
427 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
428 			cons_index & size) && (done < budget)) {
429 		/*
430 		 * make sure we read the CQE after we read the
431 		 * ownership bit
432 		 */
433 		dma_rmb();
434 
435 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
436 			     MLX4_CQE_OPCODE_ERROR)) {
437 			struct mlx4_err_cqe *cqe_err = (struct mlx4_err_cqe *)cqe;
438 
439 			en_err(priv, "CQE error - vendor syndrome: 0x%x syndrome: 0x%x\n",
440 			       cqe_err->vendor_err_syndrome,
441 			       cqe_err->syndrome);
442 		}
443 
444 		/* Skip over last polled CQE */
445 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
446 
447 		do {
448 			txbbs_skipped += last_nr_txbb;
449 			ring_index = (ring_index + last_nr_txbb) & size_mask;
450 			if (ring->tx_info[ring_index].ts_requested)
451 				timestamp = mlx4_en_get_cqe_ts(cqe);
452 
453 			/* free next descriptor */
454 			last_nr_txbb = mlx4_en_free_tx_desc(
455 					priv, ring, ring_index,
456 					!!((ring_cons + txbbs_skipped) &
457 					ring->size), timestamp, napi_budget);
458 
459 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
460 					  !!((ring_cons + txbbs_stamp) &
461 						ring->size));
462 			stamp_index = ring_index;
463 			txbbs_stamp = txbbs_skipped;
464 			packets++;
465 			bytes += ring->tx_info[ring_index].nr_bytes;
466 		} while ((++done < budget) && (ring_index != new_index));
467 
468 		++cons_index;
469 		index = cons_index & size_mask;
470 		cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
471 	}
472 
473 
474 	/*
475 	 * To prevent CQ overflow we first update CQ consumer and only then
476 	 * the ring consumer.
477 	 */
478 	mcq->cons_index = cons_index;
479 	mlx4_cq_set_ci(mcq);
480 	wmb();
481 
482 	/* we want to dirty this cache line once */
483 	ACCESS_ONCE(ring->last_nr_txbb) = last_nr_txbb;
484 	ACCESS_ONCE(ring->cons) = ring_cons + txbbs_skipped;
485 
486 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
487 
488 	/* Wakeup Tx queue if this stopped, and ring is not full.
489 	 */
490 	if (netif_tx_queue_stopped(ring->tx_queue) &&
491 	    !mlx4_en_is_tx_ring_full(ring)) {
492 		netif_tx_wake_queue(ring->tx_queue);
493 		ring->wake_queue++;
494 	}
495 	return done < budget;
496 }
497 
498 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
499 {
500 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
501 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
502 
503 	if (likely(priv->port_up))
504 		napi_schedule_irqoff(&cq->napi);
505 	else
506 		mlx4_en_arm_cq(priv, cq);
507 }
508 
509 /* TX CQ polling - called by NAPI */
510 int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget)
511 {
512 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
513 	struct net_device *dev = cq->dev;
514 	struct mlx4_en_priv *priv = netdev_priv(dev);
515 	int clean_complete;
516 
517 	clean_complete = mlx4_en_process_tx_cq(dev, cq, budget);
518 	if (!clean_complete)
519 		return budget;
520 
521 	napi_complete(napi);
522 	mlx4_en_arm_cq(priv, cq);
523 
524 	return 0;
525 }
526 
527 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
528 						      struct mlx4_en_tx_ring *ring,
529 						      u32 index,
530 						      unsigned int desc_size)
531 {
532 	u32 copy = (ring->size - index) * TXBB_SIZE;
533 	int i;
534 
535 	for (i = desc_size - copy - 4; i >= 0; i -= 4) {
536 		if ((i & (TXBB_SIZE - 1)) == 0)
537 			wmb();
538 
539 		*((u32 *) (ring->buf + i)) =
540 			*((u32 *) (ring->bounce_buf + copy + i));
541 	}
542 
543 	for (i = copy - 4; i >= 4 ; i -= 4) {
544 		if ((i & (TXBB_SIZE - 1)) == 0)
545 			wmb();
546 
547 		*((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
548 			*((u32 *) (ring->bounce_buf + i));
549 	}
550 
551 	/* Return real descriptor location */
552 	return ring->buf + index * TXBB_SIZE;
553 }
554 
555 /* Decide if skb can be inlined in tx descriptor to avoid dma mapping
556  *
557  * It seems strange we do not simply use skb_copy_bits().
558  * This would allow to inline all skbs iff skb->len <= inline_thold
559  *
560  * Note that caller already checked skb was not a gso packet
561  */
562 static bool is_inline(int inline_thold, const struct sk_buff *skb,
563 		      const struct skb_shared_info *shinfo,
564 		      void **pfrag)
565 {
566 	void *ptr;
567 
568 	if (skb->len > inline_thold || !inline_thold)
569 		return false;
570 
571 	if (shinfo->nr_frags == 1) {
572 		ptr = skb_frag_address_safe(&shinfo->frags[0]);
573 		if (unlikely(!ptr))
574 			return false;
575 		*pfrag = ptr;
576 		return true;
577 	}
578 	if (shinfo->nr_frags)
579 		return false;
580 	return true;
581 }
582 
583 static int inline_size(const struct sk_buff *skb)
584 {
585 	if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
586 	    <= MLX4_INLINE_ALIGN)
587 		return ALIGN(skb->len + CTRL_SIZE +
588 			     sizeof(struct mlx4_wqe_inline_seg), 16);
589 	else
590 		return ALIGN(skb->len + CTRL_SIZE + 2 *
591 			     sizeof(struct mlx4_wqe_inline_seg), 16);
592 }
593 
594 static int get_real_size(const struct sk_buff *skb,
595 			 const struct skb_shared_info *shinfo,
596 			 struct net_device *dev,
597 			 int *lso_header_size,
598 			 bool *inline_ok,
599 			 void **pfrag)
600 {
601 	struct mlx4_en_priv *priv = netdev_priv(dev);
602 	int real_size;
603 
604 	if (shinfo->gso_size) {
605 		*inline_ok = false;
606 		if (skb->encapsulation)
607 			*lso_header_size = (skb_inner_transport_header(skb) - skb->data) + inner_tcp_hdrlen(skb);
608 		else
609 			*lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
610 		real_size = CTRL_SIZE + shinfo->nr_frags * DS_SIZE +
611 			ALIGN(*lso_header_size + 4, DS_SIZE);
612 		if (unlikely(*lso_header_size != skb_headlen(skb))) {
613 			/* We add a segment for the skb linear buffer only if
614 			 * it contains data */
615 			if (*lso_header_size < skb_headlen(skb))
616 				real_size += DS_SIZE;
617 			else {
618 				if (netif_msg_tx_err(priv))
619 					en_warn(priv, "Non-linear headers\n");
620 				return 0;
621 			}
622 		}
623 	} else {
624 		*lso_header_size = 0;
625 		*inline_ok = is_inline(priv->prof->inline_thold, skb,
626 				       shinfo, pfrag);
627 
628 		if (*inline_ok)
629 			real_size = inline_size(skb);
630 		else
631 			real_size = CTRL_SIZE +
632 				    (shinfo->nr_frags + 1) * DS_SIZE;
633 	}
634 
635 	return real_size;
636 }
637 
638 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
639 			     const struct sk_buff *skb,
640 			     const struct skb_shared_info *shinfo,
641 			     int real_size, u16 *vlan_tag,
642 			     int tx_ind, void *fragptr)
643 {
644 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
645 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
646 	unsigned int hlen = skb_headlen(skb);
647 
648 	if (skb->len <= spc) {
649 		if (likely(skb->len >= MIN_PKT_LEN)) {
650 			inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
651 		} else {
652 			inl->byte_count = cpu_to_be32(1 << 31 | MIN_PKT_LEN);
653 			memset(((void *)(inl + 1)) + skb->len, 0,
654 			       MIN_PKT_LEN - skb->len);
655 		}
656 		skb_copy_from_linear_data(skb, inl + 1, hlen);
657 		if (shinfo->nr_frags)
658 			memcpy(((void *)(inl + 1)) + hlen, fragptr,
659 			       skb_frag_size(&shinfo->frags[0]));
660 
661 	} else {
662 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
663 		if (hlen <= spc) {
664 			skb_copy_from_linear_data(skb, inl + 1, hlen);
665 			if (hlen < spc) {
666 				memcpy(((void *)(inl + 1)) + hlen,
667 				       fragptr, spc - hlen);
668 				fragptr +=  spc - hlen;
669 			}
670 			inl = (void *) (inl + 1) + spc;
671 			memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
672 		} else {
673 			skb_copy_from_linear_data(skb, inl + 1, spc);
674 			inl = (void *) (inl + 1) + spc;
675 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
676 							 hlen - spc);
677 			if (shinfo->nr_frags)
678 				memcpy(((void *)(inl + 1)) + hlen - spc,
679 				       fragptr,
680 				       skb_frag_size(&shinfo->frags[0]));
681 		}
682 
683 		dma_wmb();
684 		inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
685 	}
686 }
687 
688 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb,
689 			 void *accel_priv, select_queue_fallback_t fallback)
690 {
691 	struct mlx4_en_priv *priv = netdev_priv(dev);
692 	u16 rings_p_up = priv->num_tx_rings_p_up;
693 	u8 up = 0;
694 
695 	if (dev->num_tc)
696 		return skb_tx_hash(dev, skb);
697 
698 	if (skb_vlan_tag_present(skb))
699 		up = skb_vlan_tag_get(skb) >> VLAN_PRIO_SHIFT;
700 
701 	return fallback(dev, skb) % rings_p_up + up * rings_p_up;
702 }
703 
704 static void mlx4_bf_copy(void __iomem *dst, const void *src,
705 			 unsigned int bytecnt)
706 {
707 	__iowrite64_copy(dst, src, bytecnt / 8);
708 }
709 
710 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
711 {
712 	struct skb_shared_info *shinfo = skb_shinfo(skb);
713 	struct mlx4_en_priv *priv = netdev_priv(dev);
714 	struct device *ddev = priv->ddev;
715 	struct mlx4_en_tx_ring *ring;
716 	struct mlx4_en_tx_desc *tx_desc;
717 	struct mlx4_wqe_data_seg *data;
718 	struct mlx4_en_tx_info *tx_info;
719 	int tx_ind = 0;
720 	int nr_txbb;
721 	int desc_size;
722 	int real_size;
723 	u32 index, bf_index;
724 	__be32 op_own;
725 	u16 vlan_tag = 0;
726 	u16 vlan_proto = 0;
727 	int i_frag;
728 	int lso_header_size;
729 	void *fragptr = NULL;
730 	bool bounce = false;
731 	bool send_doorbell;
732 	bool stop_queue;
733 	bool inline_ok;
734 	u32 ring_cons;
735 
736 	if (!priv->port_up)
737 		goto tx_drop;
738 
739 	tx_ind = skb_get_queue_mapping(skb);
740 	ring = priv->tx_ring[tx_ind];
741 
742 	/* fetch ring->cons far ahead before needing it to avoid stall */
743 	ring_cons = ACCESS_ONCE(ring->cons);
744 
745 	real_size = get_real_size(skb, shinfo, dev, &lso_header_size,
746 				  &inline_ok, &fragptr);
747 	if (unlikely(!real_size))
748 		goto tx_drop;
749 
750 	/* Align descriptor to TXBB size */
751 	desc_size = ALIGN(real_size, TXBB_SIZE);
752 	nr_txbb = desc_size / TXBB_SIZE;
753 	if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
754 		if (netif_msg_tx_err(priv))
755 			en_warn(priv, "Oversized header or SG list\n");
756 		goto tx_drop;
757 	}
758 
759 	if (skb_vlan_tag_present(skb)) {
760 		vlan_tag = skb_vlan_tag_get(skb);
761 		vlan_proto = be16_to_cpu(skb->vlan_proto);
762 	}
763 
764 	netdev_txq_bql_enqueue_prefetchw(ring->tx_queue);
765 
766 	/* Track current inflight packets for performance analysis */
767 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
768 			 (u32)(ring->prod - ring_cons - 1));
769 
770 	/* Packet is good - grab an index and transmit it */
771 	index = ring->prod & ring->size_mask;
772 	bf_index = ring->prod;
773 
774 	/* See if we have enough space for whole descriptor TXBB for setting
775 	 * SW ownership on next descriptor; if not, use a bounce buffer. */
776 	if (likely(index + nr_txbb <= ring->size))
777 		tx_desc = ring->buf + index * TXBB_SIZE;
778 	else {
779 		tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
780 		bounce = true;
781 	}
782 
783 	/* Save skb in tx_info ring */
784 	tx_info = &ring->tx_info[index];
785 	tx_info->skb = skb;
786 	tx_info->nr_txbb = nr_txbb;
787 
788 	data = &tx_desc->data;
789 	if (lso_header_size)
790 		data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
791 						      DS_SIZE));
792 
793 	/* valid only for none inline segments */
794 	tx_info->data_offset = (void *)data - (void *)tx_desc;
795 
796 	tx_info->inl = inline_ok;
797 
798 	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
799 			   !inline_ok) ? 1 : 0;
800 
801 	tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
802 	data += tx_info->nr_maps - 1;
803 
804 	if (!tx_info->inl) {
805 		dma_addr_t dma = 0;
806 		u32 byte_count = 0;
807 
808 		/* Map fragments if any */
809 		for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
810 			const struct skb_frag_struct *frag;
811 
812 			frag = &shinfo->frags[i_frag];
813 			byte_count = skb_frag_size(frag);
814 			dma = skb_frag_dma_map(ddev, frag,
815 					       0, byte_count,
816 					       DMA_TO_DEVICE);
817 			if (dma_mapping_error(ddev, dma))
818 				goto tx_drop_unmap;
819 
820 			data->addr = cpu_to_be64(dma);
821 			data->lkey = ring->mr_key;
822 			dma_wmb();
823 			data->byte_count = cpu_to_be32(byte_count);
824 			--data;
825 		}
826 
827 		/* Map linear part if needed */
828 		if (tx_info->linear) {
829 			byte_count = skb_headlen(skb) - lso_header_size;
830 
831 			dma = dma_map_single(ddev, skb->data +
832 					     lso_header_size, byte_count,
833 					     PCI_DMA_TODEVICE);
834 			if (dma_mapping_error(ddev, dma))
835 				goto tx_drop_unmap;
836 
837 			data->addr = cpu_to_be64(dma);
838 			data->lkey = ring->mr_key;
839 			dma_wmb();
840 			data->byte_count = cpu_to_be32(byte_count);
841 		}
842 		/* tx completion can avoid cache line miss for common cases */
843 		tx_info->map0_dma = dma;
844 		tx_info->map0_byte_count = byte_count;
845 	}
846 
847 	/*
848 	 * For timestamping add flag to skb_shinfo and
849 	 * set flag for further reference
850 	 */
851 	tx_info->ts_requested = 0;
852 	if (unlikely(ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
853 		     shinfo->tx_flags & SKBTX_HW_TSTAMP)) {
854 		shinfo->tx_flags |= SKBTX_IN_PROGRESS;
855 		tx_info->ts_requested = 1;
856 	}
857 
858 	/* Prepare ctrl segement apart opcode+ownership, which depends on
859 	 * whether LSO is used */
860 	tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
861 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
862 		if (!skb->encapsulation)
863 			tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
864 								 MLX4_WQE_CTRL_TCP_UDP_CSUM);
865 		else
866 			tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM);
867 		ring->tx_csum++;
868 	}
869 
870 	if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
871 		struct ethhdr *ethh;
872 
873 		/* Copy dst mac address to wqe. This allows loopback in eSwitch,
874 		 * so that VFs and PF can communicate with each other
875 		 */
876 		ethh = (struct ethhdr *)skb->data;
877 		tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
878 		tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
879 	}
880 
881 	/* Handle LSO (TSO) packets */
882 	if (lso_header_size) {
883 		int i;
884 
885 		/* Mark opcode as LSO */
886 		op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
887 			((ring->prod & ring->size) ?
888 				cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
889 
890 		/* Fill in the LSO prefix */
891 		tx_desc->lso.mss_hdr_size = cpu_to_be32(
892 			shinfo->gso_size << 16 | lso_header_size);
893 
894 		/* Copy headers;
895 		 * note that we already verified that it is linear */
896 		memcpy(tx_desc->lso.header, skb->data, lso_header_size);
897 
898 		ring->tso_packets++;
899 
900 		i = ((skb->len - lso_header_size) / shinfo->gso_size) +
901 			!!((skb->len - lso_header_size) % shinfo->gso_size);
902 		tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
903 		ring->packets += i;
904 	} else {
905 		/* Normal (Non LSO) packet */
906 		op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
907 			((ring->prod & ring->size) ?
908 			 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
909 		tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
910 		ring->packets++;
911 	}
912 	ring->bytes += tx_info->nr_bytes;
913 	netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
914 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
915 
916 	if (tx_info->inl)
917 		build_inline_wqe(tx_desc, skb, shinfo, real_size, &vlan_tag,
918 				 tx_ind, fragptr);
919 
920 	if (skb->encapsulation) {
921 		struct iphdr *ipv4 = (struct iphdr *)skb_inner_network_header(skb);
922 		if (ipv4->protocol == IPPROTO_TCP || ipv4->protocol == IPPROTO_UDP)
923 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP | MLX4_WQE_CTRL_ILP);
924 		else
925 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP);
926 	}
927 
928 	ring->prod += nr_txbb;
929 
930 	/* If we used a bounce buffer then copy descriptor back into place */
931 	if (unlikely(bounce))
932 		tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
933 
934 	skb_tx_timestamp(skb);
935 
936 	/* Check available TXBBs And 2K spare for prefetch */
937 	stop_queue = mlx4_en_is_tx_ring_full(ring);
938 	if (unlikely(stop_queue)) {
939 		netif_tx_stop_queue(ring->tx_queue);
940 		ring->queue_stopped++;
941 	}
942 	send_doorbell = !skb->xmit_more || netif_xmit_stopped(ring->tx_queue);
943 
944 	real_size = (real_size / 16) & 0x3f;
945 
946 	if (ring->bf_enabled && desc_size <= MAX_BF && !bounce &&
947 	    !skb_vlan_tag_present(skb) && send_doorbell) {
948 		tx_desc->ctrl.bf_qpn = ring->doorbell_qpn |
949 				       cpu_to_be32(real_size);
950 
951 		op_own |= htonl((bf_index & 0xffff) << 8);
952 		/* Ensure new descriptor hits memory
953 		 * before setting ownership of this descriptor to HW
954 		 */
955 		dma_wmb();
956 		tx_desc->ctrl.owner_opcode = op_own;
957 
958 		wmb();
959 
960 		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, &tx_desc->ctrl,
961 			     desc_size);
962 
963 		wmb();
964 
965 		ring->bf.offset ^= ring->bf.buf_size;
966 	} else {
967 		tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
968 		if (vlan_proto == ETH_P_8021AD)
969 			tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_SVLAN;
970 		else if (vlan_proto == ETH_P_8021Q)
971 			tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_CVLAN;
972 		else
973 			tx_desc->ctrl.ins_vlan = 0;
974 
975 		tx_desc->ctrl.fence_size = real_size;
976 
977 		/* Ensure new descriptor hits memory
978 		 * before setting ownership of this descriptor to HW
979 		 */
980 		dma_wmb();
981 		tx_desc->ctrl.owner_opcode = op_own;
982 		if (send_doorbell) {
983 			wmb();
984 			/* Since there is no iowrite*_native() that writes the
985 			 * value as is, without byteswapping - using the one
986 			 * the doesn't do byteswapping in the relevant arch
987 			 * endianness.
988 			 */
989 #if defined(__LITTLE_ENDIAN)
990 			iowrite32(
991 #else
992 			iowrite32be(
993 #endif
994 				  ring->doorbell_qpn,
995 				  ring->bf.uar->map + MLX4_SEND_DOORBELL);
996 		} else {
997 			ring->xmit_more++;
998 		}
999 	}
1000 
1001 	if (unlikely(stop_queue)) {
1002 		/* If queue was emptied after the if (stop_queue) , and before
1003 		 * the netif_tx_stop_queue() - need to wake the queue,
1004 		 * or else it will remain stopped forever.
1005 		 * Need a memory barrier to make sure ring->cons was not
1006 		 * updated before queue was stopped.
1007 		 */
1008 		smp_rmb();
1009 
1010 		ring_cons = ACCESS_ONCE(ring->cons);
1011 		if (unlikely(!mlx4_en_is_tx_ring_full(ring))) {
1012 			netif_tx_wake_queue(ring->tx_queue);
1013 			ring->wake_queue++;
1014 		}
1015 	}
1016 	return NETDEV_TX_OK;
1017 
1018 tx_drop_unmap:
1019 	en_err(priv, "DMA mapping error\n");
1020 
1021 	while (++i_frag < shinfo->nr_frags) {
1022 		++data;
1023 		dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
1024 			       be32_to_cpu(data->byte_count),
1025 			       PCI_DMA_TODEVICE);
1026 	}
1027 
1028 tx_drop:
1029 	dev_kfree_skb_any(skb);
1030 	priv->stats.tx_dropped++;
1031 	return NETDEV_TX_OK;
1032 }
1033 
1034