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