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/vmalloc.h>
41 #include <linux/tcp.h>
42 #include <linux/moduleparam.h>
43 
44 #include "mlx4_en.h"
45 
46 enum {
47 	MAX_INLINE = 104, /* 128 - 16 - 4 - 4 */
48 	MAX_BF = 256,
49 };
50 
51 static int inline_thold __read_mostly = MAX_INLINE;
52 
53 module_param_named(inline_thold, inline_thold, int, 0444);
54 MODULE_PARM_DESC(inline_thold, "threshold for using inline data");
55 
56 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
57 			   struct mlx4_en_tx_ring *ring, int qpn, u32 size,
58 			   u16 stride)
59 {
60 	struct mlx4_en_dev *mdev = priv->mdev;
61 	int tmp;
62 	int err;
63 
64 	ring->size = size;
65 	ring->size_mask = size - 1;
66 	ring->stride = stride;
67 
68 	inline_thold = min(inline_thold, MAX_INLINE);
69 
70 	tmp = size * sizeof(struct mlx4_en_tx_info);
71 	ring->tx_info = vmalloc(tmp);
72 	if (!ring->tx_info)
73 		return -ENOMEM;
74 
75 	en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
76 		 ring->tx_info, tmp);
77 
78 	ring->bounce_buf = kmalloc(MAX_DESC_SIZE, GFP_KERNEL);
79 	if (!ring->bounce_buf) {
80 		err = -ENOMEM;
81 		goto err_tx;
82 	}
83 	ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
84 
85 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size,
86 				 2 * PAGE_SIZE);
87 	if (err) {
88 		en_err(priv, "Failed allocating hwq resources\n");
89 		goto err_bounce;
90 	}
91 
92 	err = mlx4_en_map_buffer(&ring->wqres.buf);
93 	if (err) {
94 		en_err(priv, "Failed to map TX buffer\n");
95 		goto err_hwq_res;
96 	}
97 
98 	ring->buf = ring->wqres.buf.direct.buf;
99 
100 	en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d "
101 	       "buf_size:%d dma:%llx\n", ring, ring->buf, ring->size,
102 	       ring->buf_size, (unsigned long long) ring->wqres.buf.direct.map);
103 
104 	ring->qpn = qpn;
105 	err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp);
106 	if (err) {
107 		en_err(priv, "Failed allocating qp %d\n", ring->qpn);
108 		goto err_map;
109 	}
110 	ring->qp.event = mlx4_en_sqp_event;
111 
112 	err = mlx4_bf_alloc(mdev->dev, &ring->bf);
113 	if (err) {
114 		en_dbg(DRV, priv, "working without blueflame (%d)", err);
115 		ring->bf.uar = &mdev->priv_uar;
116 		ring->bf.uar->map = mdev->uar_map;
117 		ring->bf_enabled = false;
118 	} else
119 		ring->bf_enabled = true;
120 
121 	ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type;
122 
123 	return 0;
124 
125 err_map:
126 	mlx4_en_unmap_buffer(&ring->wqres.buf);
127 err_hwq_res:
128 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
129 err_bounce:
130 	kfree(ring->bounce_buf);
131 	ring->bounce_buf = NULL;
132 err_tx:
133 	vfree(ring->tx_info);
134 	ring->tx_info = NULL;
135 	return err;
136 }
137 
138 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
139 			     struct mlx4_en_tx_ring *ring)
140 {
141 	struct mlx4_en_dev *mdev = priv->mdev;
142 	en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
143 
144 	if (ring->bf_enabled)
145 		mlx4_bf_free(mdev->dev, &ring->bf);
146 	mlx4_qp_remove(mdev->dev, &ring->qp);
147 	mlx4_qp_free(mdev->dev, &ring->qp);
148 	mlx4_en_unmap_buffer(&ring->wqres.buf);
149 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
150 	kfree(ring->bounce_buf);
151 	ring->bounce_buf = NULL;
152 	vfree(ring->tx_info);
153 	ring->tx_info = NULL;
154 }
155 
156 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
157 			     struct mlx4_en_tx_ring *ring,
158 			     int cq, int user_prio)
159 {
160 	struct mlx4_en_dev *mdev = priv->mdev;
161 	int err;
162 
163 	ring->cqn = cq;
164 	ring->prod = 0;
165 	ring->cons = 0xffffffff;
166 	ring->last_nr_txbb = 1;
167 	ring->poll_cnt = 0;
168 	memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
169 	memset(ring->buf, 0, ring->buf_size);
170 
171 	ring->qp_state = MLX4_QP_STATE_RST;
172 	ring->doorbell_qpn = ring->qp.qpn << 8;
173 
174 	mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
175 				ring->cqn, user_prio, &ring->context);
176 	if (ring->bf_enabled)
177 		ring->context.usr_page = cpu_to_be32(ring->bf.uar->index);
178 
179 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
180 			       &ring->qp, &ring->qp_state);
181 
182 	return err;
183 }
184 
185 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
186 				struct mlx4_en_tx_ring *ring)
187 {
188 	struct mlx4_en_dev *mdev = priv->mdev;
189 
190 	mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
191 		       MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
192 }
193 
194 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
195 			      struct mlx4_en_tx_ring *ring, int index,
196 			      u8 owner)
197 {
198 	__be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
199 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
200 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
201 	void *end = ring->buf + ring->buf_size;
202 	__be32 *ptr = (__be32 *)tx_desc;
203 	int i;
204 
205 	/* Optimize the common case when there are no wraparounds */
206 	if (likely((void *)tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
207 		/* Stamp the freed descriptor */
208 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
209 		     i += STAMP_STRIDE) {
210 			*ptr = stamp;
211 			ptr += STAMP_DWORDS;
212 		}
213 	} else {
214 		/* Stamp the freed descriptor */
215 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
216 		     i += STAMP_STRIDE) {
217 			*ptr = stamp;
218 			ptr += STAMP_DWORDS;
219 			if ((void *)ptr >= end) {
220 				ptr = ring->buf;
221 				stamp ^= cpu_to_be32(0x80000000);
222 			}
223 		}
224 	}
225 }
226 
227 
228 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
229 				struct mlx4_en_tx_ring *ring,
230 				int index, u8 owner, u64 timestamp)
231 {
232 	struct mlx4_en_dev *mdev = priv->mdev;
233 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
234 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
235 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
236 	struct sk_buff *skb = tx_info->skb;
237 	struct skb_frag_struct *frag;
238 	void *end = ring->buf + ring->buf_size;
239 	int frags = skb_shinfo(skb)->nr_frags;
240 	int i;
241 	struct skb_shared_hwtstamps hwts;
242 
243 	if (timestamp) {
244 		mlx4_en_fill_hwtstamps(mdev, &hwts, timestamp);
245 		skb_tstamp_tx(skb, &hwts);
246 	}
247 
248 	/* Optimize the common case when there are no wraparounds */
249 	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
250 		if (!tx_info->inl) {
251 			if (tx_info->linear) {
252 				dma_unmap_single(priv->ddev,
253 					(dma_addr_t) be64_to_cpu(data->addr),
254 					 be32_to_cpu(data->byte_count),
255 					 PCI_DMA_TODEVICE);
256 				++data;
257 			}
258 
259 			for (i = 0; i < frags; i++) {
260 				frag = &skb_shinfo(skb)->frags[i];
261 				dma_unmap_page(priv->ddev,
262 					(dma_addr_t) be64_to_cpu(data[i].addr),
263 					skb_frag_size(frag), PCI_DMA_TODEVICE);
264 			}
265 		}
266 	} else {
267 		if (!tx_info->inl) {
268 			if ((void *) data >= end) {
269 				data = ring->buf + ((void *)data - end);
270 			}
271 
272 			if (tx_info->linear) {
273 				dma_unmap_single(priv->ddev,
274 					(dma_addr_t) be64_to_cpu(data->addr),
275 					 be32_to_cpu(data->byte_count),
276 					 PCI_DMA_TODEVICE);
277 				++data;
278 			}
279 
280 			for (i = 0; i < frags; i++) {
281 				/* Check for wraparound before unmapping */
282 				if ((void *) data >= end)
283 					data = ring->buf;
284 				frag = &skb_shinfo(skb)->frags[i];
285 				dma_unmap_page(priv->ddev,
286 					(dma_addr_t) be64_to_cpu(data->addr),
287 					 skb_frag_size(frag), PCI_DMA_TODEVICE);
288 				++data;
289 			}
290 		}
291 	}
292 	dev_kfree_skb_any(skb);
293 	return tx_info->nr_txbb;
294 }
295 
296 
297 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
298 {
299 	struct mlx4_en_priv *priv = netdev_priv(dev);
300 	int cnt = 0;
301 
302 	/* Skip last polled descriptor */
303 	ring->cons += ring->last_nr_txbb;
304 	en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
305 		 ring->cons, ring->prod);
306 
307 	if ((u32) (ring->prod - ring->cons) > ring->size) {
308 		if (netif_msg_tx_err(priv))
309 			en_warn(priv, "Tx consumer passed producer!\n");
310 		return 0;
311 	}
312 
313 	while (ring->cons != ring->prod) {
314 		ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
315 						ring->cons & ring->size_mask,
316 						!!(ring->cons & ring->size), 0);
317 		ring->cons += ring->last_nr_txbb;
318 		cnt++;
319 	}
320 
321 	netdev_tx_reset_queue(ring->tx_queue);
322 
323 	if (cnt)
324 		en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
325 
326 	return cnt;
327 }
328 
329 static void mlx4_en_process_tx_cq(struct net_device *dev, struct mlx4_en_cq *cq)
330 {
331 	struct mlx4_en_priv *priv = netdev_priv(dev);
332 	struct mlx4_cq *mcq = &cq->mcq;
333 	struct mlx4_en_tx_ring *ring = &priv->tx_ring[cq->ring];
334 	struct mlx4_cqe *cqe;
335 	u16 index;
336 	u16 new_index, ring_index, stamp_index;
337 	u32 txbbs_skipped = 0;
338 	u32 txbbs_stamp = 0;
339 	u32 cons_index = mcq->cons_index;
340 	int size = cq->size;
341 	u32 size_mask = ring->size_mask;
342 	struct mlx4_cqe *buf = cq->buf;
343 	u32 packets = 0;
344 	u32 bytes = 0;
345 	int factor = priv->cqe_factor;
346 	u64 timestamp = 0;
347 
348 	if (!priv->port_up)
349 		return;
350 
351 	index = cons_index & size_mask;
352 	cqe = &buf[(index << factor) + factor];
353 	ring_index = ring->cons & size_mask;
354 	stamp_index = ring_index;
355 
356 	/* Process all completed CQEs */
357 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
358 			cons_index & size)) {
359 		/*
360 		 * make sure we read the CQE after we read the
361 		 * ownership bit
362 		 */
363 		rmb();
364 
365 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
366 			     MLX4_CQE_OPCODE_ERROR)) {
367 			struct mlx4_err_cqe *cqe_err = (struct mlx4_err_cqe *)cqe;
368 
369 			en_err(priv, "CQE error - vendor syndrome: 0x%x syndrome: 0x%x\n",
370 			       cqe_err->vendor_err_syndrome,
371 			       cqe_err->syndrome);
372 		}
373 
374 		/* Skip over last polled CQE */
375 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
376 
377 		do {
378 			txbbs_skipped += ring->last_nr_txbb;
379 			ring_index = (ring_index + ring->last_nr_txbb) & size_mask;
380 			if (ring->tx_info[ring_index].ts_requested)
381 				timestamp = mlx4_en_get_cqe_ts(cqe);
382 
383 			/* free next descriptor */
384 			ring->last_nr_txbb = mlx4_en_free_tx_desc(
385 					priv, ring, ring_index,
386 					!!((ring->cons + txbbs_skipped) &
387 					ring->size), timestamp);
388 
389 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
390 					  !!((ring->cons + txbbs_stamp) &
391 						ring->size));
392 			stamp_index = ring_index;
393 			txbbs_stamp = txbbs_skipped;
394 			packets++;
395 			bytes += ring->tx_info[ring_index].nr_bytes;
396 		} while (ring_index != new_index);
397 
398 		++cons_index;
399 		index = cons_index & size_mask;
400 		cqe = &buf[(index << factor) + factor];
401 	}
402 
403 
404 	/*
405 	 * To prevent CQ overflow we first update CQ consumer and only then
406 	 * the ring consumer.
407 	 */
408 	mcq->cons_index = cons_index;
409 	mlx4_cq_set_ci(mcq);
410 	wmb();
411 	ring->cons += txbbs_skipped;
412 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
413 
414 	/*
415 	 * Wakeup Tx queue if this stopped, and at least 1 packet
416 	 * was completed
417 	 */
418 	if (netif_tx_queue_stopped(ring->tx_queue) && txbbs_skipped > 0) {
419 		netif_tx_wake_queue(ring->tx_queue);
420 		priv->port_stats.wake_queue++;
421 	}
422 }
423 
424 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
425 {
426 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
427 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
428 
429 	mlx4_en_process_tx_cq(cq->dev, cq);
430 	mlx4_en_arm_cq(priv, cq);
431 }
432 
433 
434 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
435 						      struct mlx4_en_tx_ring *ring,
436 						      u32 index,
437 						      unsigned int desc_size)
438 {
439 	u32 copy = (ring->size - index) * TXBB_SIZE;
440 	int i;
441 
442 	for (i = desc_size - copy - 4; i >= 0; i -= 4) {
443 		if ((i & (TXBB_SIZE - 1)) == 0)
444 			wmb();
445 
446 		*((u32 *) (ring->buf + i)) =
447 			*((u32 *) (ring->bounce_buf + copy + i));
448 	}
449 
450 	for (i = copy - 4; i >= 4 ; i -= 4) {
451 		if ((i & (TXBB_SIZE - 1)) == 0)
452 			wmb();
453 
454 		*((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
455 			*((u32 *) (ring->bounce_buf + i));
456 	}
457 
458 	/* Return real descriptor location */
459 	return ring->buf + index * TXBB_SIZE;
460 }
461 
462 static int is_inline(struct sk_buff *skb, void **pfrag)
463 {
464 	void *ptr;
465 
466 	if (inline_thold && !skb_is_gso(skb) && skb->len <= inline_thold) {
467 		if (skb_shinfo(skb)->nr_frags == 1) {
468 			ptr = skb_frag_address_safe(&skb_shinfo(skb)->frags[0]);
469 			if (unlikely(!ptr))
470 				return 0;
471 
472 			if (pfrag)
473 				*pfrag = ptr;
474 
475 			return 1;
476 		} else if (unlikely(skb_shinfo(skb)->nr_frags))
477 			return 0;
478 		else
479 			return 1;
480 	}
481 
482 	return 0;
483 }
484 
485 static int inline_size(struct sk_buff *skb)
486 {
487 	if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
488 	    <= MLX4_INLINE_ALIGN)
489 		return ALIGN(skb->len + CTRL_SIZE +
490 			     sizeof(struct mlx4_wqe_inline_seg), 16);
491 	else
492 		return ALIGN(skb->len + CTRL_SIZE + 2 *
493 			     sizeof(struct mlx4_wqe_inline_seg), 16);
494 }
495 
496 static int get_real_size(struct sk_buff *skb, struct net_device *dev,
497 			 int *lso_header_size)
498 {
499 	struct mlx4_en_priv *priv = netdev_priv(dev);
500 	int real_size;
501 
502 	if (skb_is_gso(skb)) {
503 		*lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
504 		real_size = CTRL_SIZE + skb_shinfo(skb)->nr_frags * DS_SIZE +
505 			ALIGN(*lso_header_size + 4, DS_SIZE);
506 		if (unlikely(*lso_header_size != skb_headlen(skb))) {
507 			/* We add a segment for the skb linear buffer only if
508 			 * it contains data */
509 			if (*lso_header_size < skb_headlen(skb))
510 				real_size += DS_SIZE;
511 			else {
512 				if (netif_msg_tx_err(priv))
513 					en_warn(priv, "Non-linear headers\n");
514 				return 0;
515 			}
516 		}
517 	} else {
518 		*lso_header_size = 0;
519 		if (!is_inline(skb, NULL))
520 			real_size = CTRL_SIZE + (skb_shinfo(skb)->nr_frags + 1) * DS_SIZE;
521 		else
522 			real_size = inline_size(skb);
523 	}
524 
525 	return real_size;
526 }
527 
528 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc, struct sk_buff *skb,
529 			     int real_size, u16 *vlan_tag, int tx_ind, void *fragptr)
530 {
531 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
532 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
533 
534 	if (skb->len <= spc) {
535 		inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
536 		skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
537 		if (skb_shinfo(skb)->nr_frags)
538 			memcpy(((void *)(inl + 1)) + skb_headlen(skb), fragptr,
539 			       skb_frag_size(&skb_shinfo(skb)->frags[0]));
540 
541 	} else {
542 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
543 		if (skb_headlen(skb) <= spc) {
544 			skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
545 			if (skb_headlen(skb) < spc) {
546 				memcpy(((void *)(inl + 1)) + skb_headlen(skb),
547 					fragptr, spc - skb_headlen(skb));
548 				fragptr +=  spc - skb_headlen(skb);
549 			}
550 			inl = (void *) (inl + 1) + spc;
551 			memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
552 		} else {
553 			skb_copy_from_linear_data(skb, inl + 1, spc);
554 			inl = (void *) (inl + 1) + spc;
555 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
556 					skb_headlen(skb) - spc);
557 			if (skb_shinfo(skb)->nr_frags)
558 				memcpy(((void *)(inl + 1)) + skb_headlen(skb) - spc,
559 					fragptr, skb_frag_size(&skb_shinfo(skb)->frags[0]));
560 		}
561 
562 		wmb();
563 		inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
564 	}
565 }
566 
567 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb)
568 {
569 	struct mlx4_en_priv *priv = netdev_priv(dev);
570 	u16 rings_p_up = priv->num_tx_rings_p_up;
571 	u8 up = 0;
572 
573 	if (dev->num_tc)
574 		return skb_tx_hash(dev, skb);
575 
576 	if (vlan_tx_tag_present(skb))
577 		up = vlan_tx_tag_get(skb) >> VLAN_PRIO_SHIFT;
578 
579 	return __netdev_pick_tx(dev, skb) % rings_p_up + up * rings_p_up;
580 }
581 
582 static void mlx4_bf_copy(void __iomem *dst, unsigned long *src, unsigned bytecnt)
583 {
584 	__iowrite64_copy(dst, src, bytecnt / 8);
585 }
586 
587 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
588 {
589 	struct mlx4_en_priv *priv = netdev_priv(dev);
590 	struct mlx4_en_dev *mdev = priv->mdev;
591 	struct device *ddev = priv->ddev;
592 	struct mlx4_en_tx_ring *ring;
593 	struct mlx4_en_tx_desc *tx_desc;
594 	struct mlx4_wqe_data_seg *data;
595 	struct mlx4_en_tx_info *tx_info;
596 	int tx_ind = 0;
597 	int nr_txbb;
598 	int desc_size;
599 	int real_size;
600 	u32 index, bf_index;
601 	__be32 op_own;
602 	u16 vlan_tag = 0;
603 	int i;
604 	int lso_header_size;
605 	void *fragptr;
606 	bool bounce = false;
607 
608 	if (!priv->port_up)
609 		goto tx_drop;
610 
611 	real_size = get_real_size(skb, dev, &lso_header_size);
612 	if (unlikely(!real_size))
613 		goto tx_drop;
614 
615 	/* Align descriptor to TXBB size */
616 	desc_size = ALIGN(real_size, TXBB_SIZE);
617 	nr_txbb = desc_size / TXBB_SIZE;
618 	if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
619 		if (netif_msg_tx_err(priv))
620 			en_warn(priv, "Oversized header or SG list\n");
621 		goto tx_drop;
622 	}
623 
624 	tx_ind = skb->queue_mapping;
625 	ring = &priv->tx_ring[tx_ind];
626 	if (vlan_tx_tag_present(skb))
627 		vlan_tag = vlan_tx_tag_get(skb);
628 
629 	/* Check available TXBBs And 2K spare for prefetch */
630 	if (unlikely(((int)(ring->prod - ring->cons)) >
631 		     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
632 		/* every full Tx ring stops queue */
633 		netif_tx_stop_queue(ring->tx_queue);
634 		priv->port_stats.queue_stopped++;
635 
636 		/* If queue was emptied after the if, and before the
637 		 * stop_queue - need to wake the queue, or else it will remain
638 		 * stopped forever.
639 		 * Need a memory barrier to make sure ring->cons was not
640 		 * updated before queue was stopped.
641 		 */
642 		wmb();
643 
644 		if (unlikely(((int)(ring->prod - ring->cons)) <=
645 			     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
646 			netif_tx_wake_queue(ring->tx_queue);
647 			priv->port_stats.wake_queue++;
648 		} else {
649 			return NETDEV_TX_BUSY;
650 		}
651 	}
652 
653 	/* Track current inflight packets for performance analysis */
654 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
655 			 (u32) (ring->prod - ring->cons - 1));
656 
657 	/* Packet is good - grab an index and transmit it */
658 	index = ring->prod & ring->size_mask;
659 	bf_index = ring->prod;
660 
661 	/* See if we have enough space for whole descriptor TXBB for setting
662 	 * SW ownership on next descriptor; if not, use a bounce buffer. */
663 	if (likely(index + nr_txbb <= ring->size))
664 		tx_desc = ring->buf + index * TXBB_SIZE;
665 	else {
666 		tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
667 		bounce = true;
668 	}
669 
670 	/* Save skb in tx_info ring */
671 	tx_info = &ring->tx_info[index];
672 	tx_info->skb = skb;
673 	tx_info->nr_txbb = nr_txbb;
674 
675 	if (lso_header_size)
676 		data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
677 						      DS_SIZE));
678 	else
679 		data = &tx_desc->data;
680 
681 	/* valid only for none inline segments */
682 	tx_info->data_offset = (void *)data - (void *)tx_desc;
683 
684 	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
685 			   !is_inline(skb, NULL)) ? 1 : 0;
686 
687 	data += skb_shinfo(skb)->nr_frags + tx_info->linear - 1;
688 
689 	if (is_inline(skb, &fragptr)) {
690 		tx_info->inl = 1;
691 	} else {
692 		/* Map fragments */
693 		for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
694 			struct skb_frag_struct *frag;
695 			dma_addr_t dma;
696 
697 			frag = &skb_shinfo(skb)->frags[i];
698 			dma = skb_frag_dma_map(ddev, frag,
699 					       0, skb_frag_size(frag),
700 					       DMA_TO_DEVICE);
701 			if (dma_mapping_error(ddev, dma))
702 				goto tx_drop_unmap;
703 
704 			data->addr = cpu_to_be64(dma);
705 			data->lkey = cpu_to_be32(mdev->mr.key);
706 			wmb();
707 			data->byte_count = cpu_to_be32(skb_frag_size(frag));
708 			--data;
709 		}
710 
711 		/* Map linear part */
712 		if (tx_info->linear) {
713 			u32 byte_count = skb_headlen(skb) - lso_header_size;
714 			dma_addr_t dma;
715 
716 			dma = dma_map_single(ddev, skb->data +
717 					     lso_header_size, byte_count,
718 					     PCI_DMA_TODEVICE);
719 			if (dma_mapping_error(ddev, dma))
720 				goto tx_drop_unmap;
721 
722 			data->addr = cpu_to_be64(dma);
723 			data->lkey = cpu_to_be32(mdev->mr.key);
724 			wmb();
725 			data->byte_count = cpu_to_be32(byte_count);
726 		}
727 		tx_info->inl = 0;
728 	}
729 
730 	/*
731 	 * For timestamping add flag to skb_shinfo and
732 	 * set flag for further reference
733 	 */
734 	if (ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
735 	    skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
736 		skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
737 		tx_info->ts_requested = 1;
738 	}
739 
740 	/* Prepare ctrl segement apart opcode+ownership, which depends on
741 	 * whether LSO is used */
742 	tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
743 	tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN *
744 		!!vlan_tx_tag_present(skb);
745 	tx_desc->ctrl.fence_size = (real_size / 16) & 0x3f;
746 	tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
747 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
748 		tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
749 							 MLX4_WQE_CTRL_TCP_UDP_CSUM);
750 		ring->tx_csum++;
751 	}
752 
753 	if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
754 		struct ethhdr *ethh;
755 
756 		/* Copy dst mac address to wqe. This allows loopback in eSwitch,
757 		 * so that VFs and PF can communicate with each other
758 		 */
759 		ethh = (struct ethhdr *)skb->data;
760 		tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
761 		tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
762 	}
763 
764 	/* Handle LSO (TSO) packets */
765 	if (lso_header_size) {
766 		/* Mark opcode as LSO */
767 		op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
768 			((ring->prod & ring->size) ?
769 				cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
770 
771 		/* Fill in the LSO prefix */
772 		tx_desc->lso.mss_hdr_size = cpu_to_be32(
773 			skb_shinfo(skb)->gso_size << 16 | lso_header_size);
774 
775 		/* Copy headers;
776 		 * note that we already verified that it is linear */
777 		memcpy(tx_desc->lso.header, skb->data, lso_header_size);
778 
779 		priv->port_stats.tso_packets++;
780 		i = ((skb->len - lso_header_size) / skb_shinfo(skb)->gso_size) +
781 			!!((skb->len - lso_header_size) % skb_shinfo(skb)->gso_size);
782 		tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
783 		ring->packets += i;
784 	} else {
785 		/* Normal (Non LSO) packet */
786 		op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
787 			((ring->prod & ring->size) ?
788 			 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
789 		tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
790 		ring->packets++;
791 
792 	}
793 	ring->bytes += tx_info->nr_bytes;
794 	netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
795 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
796 
797 	if (tx_info->inl) {
798 		build_inline_wqe(tx_desc, skb, real_size, &vlan_tag, tx_ind, fragptr);
799 		tx_info->inl = 1;
800 	}
801 
802 	ring->prod += nr_txbb;
803 
804 	/* If we used a bounce buffer then copy descriptor back into place */
805 	if (bounce)
806 		tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
807 
808 	skb_tx_timestamp(skb);
809 
810 	if (ring->bf_enabled && desc_size <= MAX_BF && !bounce && !vlan_tx_tag_present(skb)) {
811 		*(__be32 *) (&tx_desc->ctrl.vlan_tag) |= cpu_to_be32(ring->doorbell_qpn);
812 		op_own |= htonl((bf_index & 0xffff) << 8);
813 		/* Ensure new descirptor hits memory
814 		* before setting ownership of this descriptor to HW */
815 		wmb();
816 		tx_desc->ctrl.owner_opcode = op_own;
817 
818 		wmb();
819 
820 		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, (unsigned long *) &tx_desc->ctrl,
821 		     desc_size);
822 
823 		wmb();
824 
825 		ring->bf.offset ^= ring->bf.buf_size;
826 	} else {
827 		/* Ensure new descirptor hits memory
828 		* before setting ownership of this descriptor to HW */
829 		wmb();
830 		tx_desc->ctrl.owner_opcode = op_own;
831 		wmb();
832 		iowrite32be(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL);
833 	}
834 
835 	return NETDEV_TX_OK;
836 
837 tx_drop_unmap:
838 	en_err(priv, "DMA mapping error\n");
839 
840 	for (i++; i < skb_shinfo(skb)->nr_frags; i++) {
841 		data++;
842 		dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
843 			       be32_to_cpu(data->byte_count),
844 			       PCI_DMA_TODEVICE);
845 	}
846 
847 tx_drop:
848 	dev_kfree_skb_any(skb);
849 	priv->stats.tx_dropped++;
850 	return NETDEV_TX_OK;
851 }
852 
853