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 
195 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
196 				struct mlx4_en_tx_ring *ring,
197 				int index, u8 owner, u64 timestamp)
198 {
199 	struct mlx4_en_dev *mdev = priv->mdev;
200 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
201 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
202 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
203 	struct sk_buff *skb = tx_info->skb;
204 	struct skb_frag_struct *frag;
205 	void *end = ring->buf + ring->buf_size;
206 	int frags = skb_shinfo(skb)->nr_frags;
207 	int i;
208 	__be32 *ptr = (__be32 *)tx_desc;
209 	__be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
210 	struct skb_shared_hwtstamps hwts;
211 
212 	if (timestamp) {
213 		mlx4_en_fill_hwtstamps(mdev, &hwts, timestamp);
214 		skb_tstamp_tx(skb, &hwts);
215 	}
216 
217 	/* Optimize the common case when there are no wraparounds */
218 	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
219 		if (!tx_info->inl) {
220 			if (tx_info->linear) {
221 				dma_unmap_single(priv->ddev,
222 					(dma_addr_t) be64_to_cpu(data->addr),
223 					 be32_to_cpu(data->byte_count),
224 					 PCI_DMA_TODEVICE);
225 				++data;
226 			}
227 
228 			for (i = 0; i < frags; i++) {
229 				frag = &skb_shinfo(skb)->frags[i];
230 				dma_unmap_page(priv->ddev,
231 					(dma_addr_t) be64_to_cpu(data[i].addr),
232 					skb_frag_size(frag), PCI_DMA_TODEVICE);
233 			}
234 		}
235 		/* Stamp the freed descriptor */
236 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE; i += STAMP_STRIDE) {
237 			*ptr = stamp;
238 			ptr += STAMP_DWORDS;
239 		}
240 
241 	} else {
242 		if (!tx_info->inl) {
243 			if ((void *) data >= end) {
244 				data = ring->buf + ((void *)data - end);
245 			}
246 
247 			if (tx_info->linear) {
248 				dma_unmap_single(priv->ddev,
249 					(dma_addr_t) be64_to_cpu(data->addr),
250 					 be32_to_cpu(data->byte_count),
251 					 PCI_DMA_TODEVICE);
252 				++data;
253 			}
254 
255 			for (i = 0; i < frags; i++) {
256 				/* Check for wraparound before unmapping */
257 				if ((void *) data >= end)
258 					data = ring->buf;
259 				frag = &skb_shinfo(skb)->frags[i];
260 				dma_unmap_page(priv->ddev,
261 					(dma_addr_t) be64_to_cpu(data->addr),
262 					 skb_frag_size(frag), PCI_DMA_TODEVICE);
263 				++data;
264 			}
265 		}
266 		/* Stamp the freed descriptor */
267 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE; i += STAMP_STRIDE) {
268 			*ptr = stamp;
269 			ptr += STAMP_DWORDS;
270 			if ((void *) ptr >= end) {
271 				ptr = ring->buf;
272 				stamp ^= cpu_to_be32(0x80000000);
273 			}
274 		}
275 
276 	}
277 	dev_kfree_skb_any(skb);
278 	return tx_info->nr_txbb;
279 }
280 
281 
282 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
283 {
284 	struct mlx4_en_priv *priv = netdev_priv(dev);
285 	int cnt = 0;
286 
287 	/* Skip last polled descriptor */
288 	ring->cons += ring->last_nr_txbb;
289 	en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
290 		 ring->cons, ring->prod);
291 
292 	if ((u32) (ring->prod - ring->cons) > ring->size) {
293 		if (netif_msg_tx_err(priv))
294 			en_warn(priv, "Tx consumer passed producer!\n");
295 		return 0;
296 	}
297 
298 	while (ring->cons != ring->prod) {
299 		ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
300 						ring->cons & ring->size_mask,
301 						!!(ring->cons & ring->size), 0);
302 		ring->cons += ring->last_nr_txbb;
303 		cnt++;
304 	}
305 
306 	netdev_tx_reset_queue(ring->tx_queue);
307 
308 	if (cnt)
309 		en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
310 
311 	return cnt;
312 }
313 
314 static void mlx4_en_process_tx_cq(struct net_device *dev, struct mlx4_en_cq *cq)
315 {
316 	struct mlx4_en_priv *priv = netdev_priv(dev);
317 	struct mlx4_cq *mcq = &cq->mcq;
318 	struct mlx4_en_tx_ring *ring = &priv->tx_ring[cq->ring];
319 	struct mlx4_cqe *cqe;
320 	u16 index;
321 	u16 new_index, ring_index;
322 	u32 txbbs_skipped = 0;
323 	u32 cons_index = mcq->cons_index;
324 	int size = cq->size;
325 	u32 size_mask = ring->size_mask;
326 	struct mlx4_cqe *buf = cq->buf;
327 	u32 packets = 0;
328 	u32 bytes = 0;
329 	int factor = priv->cqe_factor;
330 	u64 timestamp = 0;
331 
332 	if (!priv->port_up)
333 		return;
334 
335 	index = cons_index & size_mask;
336 	cqe = &buf[(index << factor) + factor];
337 	ring_index = ring->cons & size_mask;
338 
339 	/* Process all completed CQEs */
340 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
341 			cons_index & size)) {
342 		/*
343 		 * make sure we read the CQE after we read the
344 		 * ownership bit
345 		 */
346 		rmb();
347 
348 		/* Skip over last polled CQE */
349 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
350 
351 		do {
352 			txbbs_skipped += ring->last_nr_txbb;
353 			ring_index = (ring_index + ring->last_nr_txbb) & size_mask;
354 			if (ring->tx_info[ring_index].ts_requested)
355 				timestamp = mlx4_en_get_cqe_ts(cqe);
356 
357 			/* free next descriptor */
358 			ring->last_nr_txbb = mlx4_en_free_tx_desc(
359 					priv, ring, ring_index,
360 					!!((ring->cons + txbbs_skipped) &
361 					ring->size), timestamp);
362 			packets++;
363 			bytes += ring->tx_info[ring_index].nr_bytes;
364 		} while (ring_index != new_index);
365 
366 		++cons_index;
367 		index = cons_index & size_mask;
368 		cqe = &buf[(index << factor) + factor];
369 	}
370 
371 
372 	/*
373 	 * To prevent CQ overflow we first update CQ consumer and only then
374 	 * the ring consumer.
375 	 */
376 	mcq->cons_index = cons_index;
377 	mlx4_cq_set_ci(mcq);
378 	wmb();
379 	ring->cons += txbbs_skipped;
380 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
381 
382 	/*
383 	 * Wakeup Tx queue if this stopped, and at least 1 packet
384 	 * was completed
385 	 */
386 	if (netif_tx_queue_stopped(ring->tx_queue) && txbbs_skipped > 0) {
387 		netif_tx_wake_queue(ring->tx_queue);
388 		priv->port_stats.wake_queue++;
389 	}
390 }
391 
392 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
393 {
394 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
395 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
396 
397 	mlx4_en_process_tx_cq(cq->dev, cq);
398 	mlx4_en_arm_cq(priv, cq);
399 }
400 
401 
402 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
403 						      struct mlx4_en_tx_ring *ring,
404 						      u32 index,
405 						      unsigned int desc_size)
406 {
407 	u32 copy = (ring->size - index) * TXBB_SIZE;
408 	int i;
409 
410 	for (i = desc_size - copy - 4; i >= 0; i -= 4) {
411 		if ((i & (TXBB_SIZE - 1)) == 0)
412 			wmb();
413 
414 		*((u32 *) (ring->buf + i)) =
415 			*((u32 *) (ring->bounce_buf + copy + i));
416 	}
417 
418 	for (i = copy - 4; i >= 4 ; i -= 4) {
419 		if ((i & (TXBB_SIZE - 1)) == 0)
420 			wmb();
421 
422 		*((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
423 			*((u32 *) (ring->bounce_buf + i));
424 	}
425 
426 	/* Return real descriptor location */
427 	return ring->buf + index * TXBB_SIZE;
428 }
429 
430 static int is_inline(struct sk_buff *skb, void **pfrag)
431 {
432 	void *ptr;
433 
434 	if (inline_thold && !skb_is_gso(skb) && skb->len <= inline_thold) {
435 		if (skb_shinfo(skb)->nr_frags == 1) {
436 			ptr = skb_frag_address_safe(&skb_shinfo(skb)->frags[0]);
437 			if (unlikely(!ptr))
438 				return 0;
439 
440 			if (pfrag)
441 				*pfrag = ptr;
442 
443 			return 1;
444 		} else if (unlikely(skb_shinfo(skb)->nr_frags))
445 			return 0;
446 		else
447 			return 1;
448 	}
449 
450 	return 0;
451 }
452 
453 static int inline_size(struct sk_buff *skb)
454 {
455 	if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
456 	    <= MLX4_INLINE_ALIGN)
457 		return ALIGN(skb->len + CTRL_SIZE +
458 			     sizeof(struct mlx4_wqe_inline_seg), 16);
459 	else
460 		return ALIGN(skb->len + CTRL_SIZE + 2 *
461 			     sizeof(struct mlx4_wqe_inline_seg), 16);
462 }
463 
464 static int get_real_size(struct sk_buff *skb, struct net_device *dev,
465 			 int *lso_header_size)
466 {
467 	struct mlx4_en_priv *priv = netdev_priv(dev);
468 	int real_size;
469 
470 	if (skb_is_gso(skb)) {
471 		*lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
472 		real_size = CTRL_SIZE + skb_shinfo(skb)->nr_frags * DS_SIZE +
473 			ALIGN(*lso_header_size + 4, DS_SIZE);
474 		if (unlikely(*lso_header_size != skb_headlen(skb))) {
475 			/* We add a segment for the skb linear buffer only if
476 			 * it contains data */
477 			if (*lso_header_size < skb_headlen(skb))
478 				real_size += DS_SIZE;
479 			else {
480 				if (netif_msg_tx_err(priv))
481 					en_warn(priv, "Non-linear headers\n");
482 				return 0;
483 			}
484 		}
485 	} else {
486 		*lso_header_size = 0;
487 		if (!is_inline(skb, NULL))
488 			real_size = CTRL_SIZE + (skb_shinfo(skb)->nr_frags + 1) * DS_SIZE;
489 		else
490 			real_size = inline_size(skb);
491 	}
492 
493 	return real_size;
494 }
495 
496 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc, struct sk_buff *skb,
497 			     int real_size, u16 *vlan_tag, int tx_ind, void *fragptr)
498 {
499 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
500 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
501 
502 	if (skb->len <= spc) {
503 		inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
504 		skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
505 		if (skb_shinfo(skb)->nr_frags)
506 			memcpy(((void *)(inl + 1)) + skb_headlen(skb), fragptr,
507 			       skb_frag_size(&skb_shinfo(skb)->frags[0]));
508 
509 	} else {
510 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
511 		if (skb_headlen(skb) <= spc) {
512 			skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
513 			if (skb_headlen(skb) < spc) {
514 				memcpy(((void *)(inl + 1)) + skb_headlen(skb),
515 					fragptr, spc - skb_headlen(skb));
516 				fragptr +=  spc - skb_headlen(skb);
517 			}
518 			inl = (void *) (inl + 1) + spc;
519 			memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
520 		} else {
521 			skb_copy_from_linear_data(skb, inl + 1, spc);
522 			inl = (void *) (inl + 1) + spc;
523 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
524 					skb_headlen(skb) - spc);
525 			if (skb_shinfo(skb)->nr_frags)
526 				memcpy(((void *)(inl + 1)) + skb_headlen(skb) - spc,
527 					fragptr, skb_frag_size(&skb_shinfo(skb)->frags[0]));
528 		}
529 
530 		wmb();
531 		inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
532 	}
533 }
534 
535 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb)
536 {
537 	struct mlx4_en_priv *priv = netdev_priv(dev);
538 	u16 rings_p_up = priv->num_tx_rings_p_up;
539 	u8 up = 0;
540 
541 	if (dev->num_tc)
542 		return skb_tx_hash(dev, skb);
543 
544 	if (vlan_tx_tag_present(skb))
545 		up = vlan_tx_tag_get(skb) >> VLAN_PRIO_SHIFT;
546 
547 	return __skb_tx_hash(dev, skb, rings_p_up) + up * rings_p_up;
548 }
549 
550 static void mlx4_bf_copy(void __iomem *dst, unsigned long *src, unsigned bytecnt)
551 {
552 	__iowrite64_copy(dst, src, bytecnt / 8);
553 }
554 
555 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
556 {
557 	struct mlx4_en_priv *priv = netdev_priv(dev);
558 	struct mlx4_en_dev *mdev = priv->mdev;
559 	struct mlx4_en_tx_ring *ring;
560 	struct mlx4_en_tx_desc *tx_desc;
561 	struct mlx4_wqe_data_seg *data;
562 	struct skb_frag_struct *frag;
563 	struct mlx4_en_tx_info *tx_info;
564 	struct ethhdr *ethh;
565 	int tx_ind = 0;
566 	int nr_txbb;
567 	int desc_size;
568 	int real_size;
569 	dma_addr_t dma;
570 	u32 index, bf_index;
571 	__be32 op_own;
572 	u16 vlan_tag = 0;
573 	int i;
574 	int lso_header_size;
575 	void *fragptr;
576 	bool bounce = false;
577 
578 	if (!priv->port_up)
579 		goto tx_drop;
580 
581 	real_size = get_real_size(skb, dev, &lso_header_size);
582 	if (unlikely(!real_size))
583 		goto tx_drop;
584 
585 	/* Align descriptor to TXBB size */
586 	desc_size = ALIGN(real_size, TXBB_SIZE);
587 	nr_txbb = desc_size / TXBB_SIZE;
588 	if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
589 		if (netif_msg_tx_err(priv))
590 			en_warn(priv, "Oversized header or SG list\n");
591 		goto tx_drop;
592 	}
593 
594 	tx_ind = skb->queue_mapping;
595 	ring = &priv->tx_ring[tx_ind];
596 	if (vlan_tx_tag_present(skb))
597 		vlan_tag = vlan_tx_tag_get(skb);
598 
599 	/* Check available TXBBs And 2K spare for prefetch */
600 	if (unlikely(((int)(ring->prod - ring->cons)) >
601 		     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
602 		/* every full Tx ring stops queue */
603 		netif_tx_stop_queue(ring->tx_queue);
604 		priv->port_stats.queue_stopped++;
605 
606 		/* If queue was emptied after the if, and before the
607 		 * stop_queue - need to wake the queue, or else it will remain
608 		 * stopped forever.
609 		 * Need a memory barrier to make sure ring->cons was not
610 		 * updated before queue was stopped.
611 		 */
612 		wmb();
613 
614 		if (unlikely(((int)(ring->prod - ring->cons)) <=
615 			     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
616 			netif_tx_wake_queue(ring->tx_queue);
617 			priv->port_stats.wake_queue++;
618 		} else {
619 			return NETDEV_TX_BUSY;
620 		}
621 	}
622 
623 	/* Track current inflight packets for performance analysis */
624 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
625 			 (u32) (ring->prod - ring->cons - 1));
626 
627 	/* Packet is good - grab an index and transmit it */
628 	index = ring->prod & ring->size_mask;
629 	bf_index = ring->prod;
630 
631 	/* See if we have enough space for whole descriptor TXBB for setting
632 	 * SW ownership on next descriptor; if not, use a bounce buffer. */
633 	if (likely(index + nr_txbb <= ring->size))
634 		tx_desc = ring->buf + index * TXBB_SIZE;
635 	else {
636 		tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
637 		bounce = true;
638 	}
639 
640 	/* Save skb in tx_info ring */
641 	tx_info = &ring->tx_info[index];
642 	tx_info->skb = skb;
643 	tx_info->nr_txbb = nr_txbb;
644 
645 	/*
646 	 * For timestamping add flag to skb_shinfo and
647 	 * set flag for further reference
648 	 */
649 	if (ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
650 	    skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
651 		skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
652 		tx_info->ts_requested = 1;
653 	}
654 
655 	/* Prepare ctrl segement apart opcode+ownership, which depends on
656 	 * whether LSO is used */
657 	tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
658 	tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN *
659 		!!vlan_tx_tag_present(skb);
660 	tx_desc->ctrl.fence_size = (real_size / 16) & 0x3f;
661 	tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
662 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
663 		tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
664 							 MLX4_WQE_CTRL_TCP_UDP_CSUM);
665 		ring->tx_csum++;
666 	}
667 
668 	if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
669 		/* Copy dst mac address to wqe. This allows loopback in eSwitch,
670 		 * so that VFs and PF can communicate with each other
671 		 */
672 		ethh = (struct ethhdr *)skb->data;
673 		tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
674 		tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
675 	}
676 
677 	/* Handle LSO (TSO) packets */
678 	if (lso_header_size) {
679 		/* Mark opcode as LSO */
680 		op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
681 			((ring->prod & ring->size) ?
682 				cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
683 
684 		/* Fill in the LSO prefix */
685 		tx_desc->lso.mss_hdr_size = cpu_to_be32(
686 			skb_shinfo(skb)->gso_size << 16 | lso_header_size);
687 
688 		/* Copy headers;
689 		 * note that we already verified that it is linear */
690 		memcpy(tx_desc->lso.header, skb->data, lso_header_size);
691 		data = ((void *) &tx_desc->lso +
692 			ALIGN(lso_header_size + 4, DS_SIZE));
693 
694 		priv->port_stats.tso_packets++;
695 		i = ((skb->len - lso_header_size) / skb_shinfo(skb)->gso_size) +
696 			!!((skb->len - lso_header_size) % skb_shinfo(skb)->gso_size);
697 		tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
698 		ring->packets += i;
699 	} else {
700 		/* Normal (Non LSO) packet */
701 		op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
702 			((ring->prod & ring->size) ?
703 			 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
704 		data = &tx_desc->data;
705 		tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
706 		ring->packets++;
707 
708 	}
709 	ring->bytes += tx_info->nr_bytes;
710 	netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
711 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
712 
713 
714 	/* valid only for none inline segments */
715 	tx_info->data_offset = (void *) data - (void *) tx_desc;
716 
717 	tx_info->linear = (lso_header_size < skb_headlen(skb) && !is_inline(skb, NULL)) ? 1 : 0;
718 	data += skb_shinfo(skb)->nr_frags + tx_info->linear - 1;
719 
720 	if (!is_inline(skb, &fragptr)) {
721 		/* Map fragments */
722 		for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
723 			frag = &skb_shinfo(skb)->frags[i];
724 			dma = skb_frag_dma_map(priv->ddev, frag,
725 					       0, skb_frag_size(frag),
726 					       DMA_TO_DEVICE);
727 			data->addr = cpu_to_be64(dma);
728 			data->lkey = cpu_to_be32(mdev->mr.key);
729 			wmb();
730 			data->byte_count = cpu_to_be32(skb_frag_size(frag));
731 			--data;
732 		}
733 
734 		/* Map linear part */
735 		if (tx_info->linear) {
736 			dma = dma_map_single(priv->ddev, skb->data + lso_header_size,
737 					     skb_headlen(skb) - lso_header_size, PCI_DMA_TODEVICE);
738 			data->addr = cpu_to_be64(dma);
739 			data->lkey = cpu_to_be32(mdev->mr.key);
740 			wmb();
741 			data->byte_count = cpu_to_be32(skb_headlen(skb) - lso_header_size);
742 		}
743 		tx_info->inl = 0;
744 	} else {
745 		build_inline_wqe(tx_desc, skb, real_size, &vlan_tag, tx_ind, fragptr);
746 		tx_info->inl = 1;
747 	}
748 
749 	ring->prod += nr_txbb;
750 
751 	/* If we used a bounce buffer then copy descriptor back into place */
752 	if (bounce)
753 		tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
754 
755 	skb_tx_timestamp(skb);
756 
757 	if (ring->bf_enabled && desc_size <= MAX_BF && !bounce && !vlan_tx_tag_present(skb)) {
758 		*(__be32 *) (&tx_desc->ctrl.vlan_tag) |= cpu_to_be32(ring->doorbell_qpn);
759 		op_own |= htonl((bf_index & 0xffff) << 8);
760 		/* Ensure new descirptor hits memory
761 		* before setting ownership of this descriptor to HW */
762 		wmb();
763 		tx_desc->ctrl.owner_opcode = op_own;
764 
765 		wmb();
766 
767 		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, (unsigned long *) &tx_desc->ctrl,
768 		     desc_size);
769 
770 		wmb();
771 
772 		ring->bf.offset ^= ring->bf.buf_size;
773 	} else {
774 		/* Ensure new descirptor hits memory
775 		* before setting ownership of this descriptor to HW */
776 		wmb();
777 		tx_desc->ctrl.owner_opcode = op_own;
778 		wmb();
779 		iowrite32be(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL);
780 	}
781 
782 	return NETDEV_TX_OK;
783 
784 tx_drop:
785 	dev_kfree_skb_any(skb);
786 	priv->stats.tx_dropped++;
787 	return NETDEV_TX_OK;
788 }
789 
790