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 **pring, int qpn, u32 size,
58 			   u16 stride, int node)
59 {
60 	struct mlx4_en_dev *mdev = priv->mdev;
61 	struct mlx4_en_tx_ring *ring;
62 	int tmp;
63 	int err;
64 
65 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
66 	if (!ring) {
67 		ring = kzalloc(sizeof(*ring), GFP_KERNEL);
68 		if (!ring) {
69 			en_err(priv, "Failed allocating TX ring\n");
70 			return -ENOMEM;
71 		}
72 	}
73 
74 	ring->size = size;
75 	ring->size_mask = size - 1;
76 	ring->stride = stride;
77 
78 	inline_thold = min(inline_thold, MAX_INLINE);
79 
80 	tmp = size * sizeof(struct mlx4_en_tx_info);
81 	ring->tx_info = vmalloc_node(tmp, node);
82 	if (!ring->tx_info) {
83 		ring->tx_info = vmalloc(tmp);
84 		if (!ring->tx_info) {
85 			err = -ENOMEM;
86 			goto err_ring;
87 		}
88 	}
89 
90 	en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
91 		 ring->tx_info, tmp);
92 
93 	ring->bounce_buf = kmalloc_node(MAX_DESC_SIZE, GFP_KERNEL, node);
94 	if (!ring->bounce_buf) {
95 		ring->bounce_buf = kmalloc(MAX_DESC_SIZE, GFP_KERNEL);
96 		if (!ring->bounce_buf) {
97 			err = -ENOMEM;
98 			goto err_info;
99 		}
100 	}
101 	ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
102 
103 	/* Allocate HW buffers on provided NUMA node */
104 	set_dev_node(&mdev->dev->pdev->dev, node);
105 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size,
106 				 2 * PAGE_SIZE);
107 	set_dev_node(&mdev->dev->pdev->dev, mdev->dev->numa_node);
108 	if (err) {
109 		en_err(priv, "Failed allocating hwq resources\n");
110 		goto err_bounce;
111 	}
112 
113 	err = mlx4_en_map_buffer(&ring->wqres.buf);
114 	if (err) {
115 		en_err(priv, "Failed to map TX buffer\n");
116 		goto err_hwq_res;
117 	}
118 
119 	ring->buf = ring->wqres.buf.direct.buf;
120 
121 	en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d "
122 	       "buf_size:%d dma:%llx\n", ring, ring->buf, ring->size,
123 	       ring->buf_size, (unsigned long long) ring->wqres.buf.direct.map);
124 
125 	ring->qpn = qpn;
126 	err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp);
127 	if (err) {
128 		en_err(priv, "Failed allocating qp %d\n", ring->qpn);
129 		goto err_map;
130 	}
131 	ring->qp.event = mlx4_en_sqp_event;
132 
133 	err = mlx4_bf_alloc(mdev->dev, &ring->bf, node);
134 	if (err) {
135 		en_dbg(DRV, priv, "working without blueflame (%d)", err);
136 		ring->bf.uar = &mdev->priv_uar;
137 		ring->bf.uar->map = mdev->uar_map;
138 		ring->bf_enabled = false;
139 	} else
140 		ring->bf_enabled = true;
141 
142 	ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type;
143 
144 	*pring = ring;
145 	return 0;
146 
147 err_map:
148 	mlx4_en_unmap_buffer(&ring->wqres.buf);
149 err_hwq_res:
150 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
151 err_bounce:
152 	kfree(ring->bounce_buf);
153 	ring->bounce_buf = NULL;
154 err_info:
155 	vfree(ring->tx_info);
156 	ring->tx_info = NULL;
157 err_ring:
158 	kfree(ring);
159 	*pring = NULL;
160 	return err;
161 }
162 
163 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
164 			     struct mlx4_en_tx_ring **pring)
165 {
166 	struct mlx4_en_dev *mdev = priv->mdev;
167 	struct mlx4_en_tx_ring *ring = *pring;
168 	en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
169 
170 	if (ring->bf_enabled)
171 		mlx4_bf_free(mdev->dev, &ring->bf);
172 	mlx4_qp_remove(mdev->dev, &ring->qp);
173 	mlx4_qp_free(mdev->dev, &ring->qp);
174 	mlx4_en_unmap_buffer(&ring->wqres.buf);
175 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
176 	kfree(ring->bounce_buf);
177 	ring->bounce_buf = NULL;
178 	vfree(ring->tx_info);
179 	ring->tx_info = NULL;
180 	kfree(ring);
181 	*pring = NULL;
182 }
183 
184 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
185 			     struct mlx4_en_tx_ring *ring,
186 			     int cq, int user_prio)
187 {
188 	struct mlx4_en_dev *mdev = priv->mdev;
189 	int err;
190 
191 	ring->cqn = cq;
192 	ring->prod = 0;
193 	ring->cons = 0xffffffff;
194 	ring->last_nr_txbb = 1;
195 	ring->poll_cnt = 0;
196 	memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
197 	memset(ring->buf, 0, ring->buf_size);
198 
199 	ring->qp_state = MLX4_QP_STATE_RST;
200 	ring->doorbell_qpn = ring->qp.qpn << 8;
201 
202 	mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
203 				ring->cqn, user_prio, &ring->context);
204 	if (ring->bf_enabled)
205 		ring->context.usr_page = cpu_to_be32(ring->bf.uar->index);
206 
207 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
208 			       &ring->qp, &ring->qp_state);
209 
210 	return err;
211 }
212 
213 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
214 				struct mlx4_en_tx_ring *ring)
215 {
216 	struct mlx4_en_dev *mdev = priv->mdev;
217 
218 	mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
219 		       MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
220 }
221 
222 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
223 			      struct mlx4_en_tx_ring *ring, int index,
224 			      u8 owner)
225 {
226 	__be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
227 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
228 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
229 	void *end = ring->buf + ring->buf_size;
230 	__be32 *ptr = (__be32 *)tx_desc;
231 	int i;
232 
233 	/* Optimize the common case when there are no wraparounds */
234 	if (likely((void *)tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
235 		/* Stamp the freed descriptor */
236 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
237 		     i += STAMP_STRIDE) {
238 			*ptr = stamp;
239 			ptr += STAMP_DWORDS;
240 		}
241 	} else {
242 		/* Stamp the freed descriptor */
243 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
244 		     i += STAMP_STRIDE) {
245 			*ptr = stamp;
246 			ptr += STAMP_DWORDS;
247 			if ((void *)ptr >= end) {
248 				ptr = ring->buf;
249 				stamp ^= cpu_to_be32(0x80000000);
250 			}
251 		}
252 	}
253 }
254 
255 
256 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
257 				struct mlx4_en_tx_ring *ring,
258 				int index, u8 owner, u64 timestamp)
259 {
260 	struct mlx4_en_dev *mdev = priv->mdev;
261 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
262 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
263 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
264 	struct sk_buff *skb = tx_info->skb;
265 	struct skb_frag_struct *frag;
266 	void *end = ring->buf + ring->buf_size;
267 	int frags = skb_shinfo(skb)->nr_frags;
268 	int i;
269 	struct skb_shared_hwtstamps hwts;
270 
271 	if (timestamp) {
272 		mlx4_en_fill_hwtstamps(mdev, &hwts, timestamp);
273 		skb_tstamp_tx(skb, &hwts);
274 	}
275 
276 	/* Optimize the common case when there are no wraparounds */
277 	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
278 		if (!tx_info->inl) {
279 			if (tx_info->linear) {
280 				dma_unmap_single(priv->ddev,
281 					(dma_addr_t) be64_to_cpu(data->addr),
282 					 be32_to_cpu(data->byte_count),
283 					 PCI_DMA_TODEVICE);
284 				++data;
285 			}
286 
287 			for (i = 0; i < frags; i++) {
288 				frag = &skb_shinfo(skb)->frags[i];
289 				dma_unmap_page(priv->ddev,
290 					(dma_addr_t) be64_to_cpu(data[i].addr),
291 					skb_frag_size(frag), PCI_DMA_TODEVICE);
292 			}
293 		}
294 	} else {
295 		if (!tx_info->inl) {
296 			if ((void *) data >= end) {
297 				data = ring->buf + ((void *)data - end);
298 			}
299 
300 			if (tx_info->linear) {
301 				dma_unmap_single(priv->ddev,
302 					(dma_addr_t) be64_to_cpu(data->addr),
303 					 be32_to_cpu(data->byte_count),
304 					 PCI_DMA_TODEVICE);
305 				++data;
306 			}
307 
308 			for (i = 0; i < frags; i++) {
309 				/* Check for wraparound before unmapping */
310 				if ((void *) data >= end)
311 					data = ring->buf;
312 				frag = &skb_shinfo(skb)->frags[i];
313 				dma_unmap_page(priv->ddev,
314 					(dma_addr_t) be64_to_cpu(data->addr),
315 					 skb_frag_size(frag), PCI_DMA_TODEVICE);
316 				++data;
317 			}
318 		}
319 	}
320 	dev_kfree_skb_any(skb);
321 	return tx_info->nr_txbb;
322 }
323 
324 
325 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
326 {
327 	struct mlx4_en_priv *priv = netdev_priv(dev);
328 	int cnt = 0;
329 
330 	/* Skip last polled descriptor */
331 	ring->cons += ring->last_nr_txbb;
332 	en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
333 		 ring->cons, ring->prod);
334 
335 	if ((u32) (ring->prod - ring->cons) > ring->size) {
336 		if (netif_msg_tx_err(priv))
337 			en_warn(priv, "Tx consumer passed producer!\n");
338 		return 0;
339 	}
340 
341 	while (ring->cons != ring->prod) {
342 		ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
343 						ring->cons & ring->size_mask,
344 						!!(ring->cons & ring->size), 0);
345 		ring->cons += ring->last_nr_txbb;
346 		cnt++;
347 	}
348 
349 	netdev_tx_reset_queue(ring->tx_queue);
350 
351 	if (cnt)
352 		en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
353 
354 	return cnt;
355 }
356 
357 static void mlx4_en_process_tx_cq(struct net_device *dev, struct mlx4_en_cq *cq)
358 {
359 	struct mlx4_en_priv *priv = netdev_priv(dev);
360 	struct mlx4_cq *mcq = &cq->mcq;
361 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->ring];
362 	struct mlx4_cqe *cqe;
363 	u16 index;
364 	u16 new_index, ring_index, stamp_index;
365 	u32 txbbs_skipped = 0;
366 	u32 txbbs_stamp = 0;
367 	u32 cons_index = mcq->cons_index;
368 	int size = cq->size;
369 	u32 size_mask = ring->size_mask;
370 	struct mlx4_cqe *buf = cq->buf;
371 	u32 packets = 0;
372 	u32 bytes = 0;
373 	int factor = priv->cqe_factor;
374 	u64 timestamp = 0;
375 
376 	if (!priv->port_up)
377 		return;
378 
379 	index = cons_index & size_mask;
380 	cqe = &buf[(index << factor) + factor];
381 	ring_index = ring->cons & size_mask;
382 	stamp_index = ring_index;
383 
384 	/* Process all completed CQEs */
385 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
386 			cons_index & size)) {
387 		/*
388 		 * make sure we read the CQE after we read the
389 		 * ownership bit
390 		 */
391 		rmb();
392 
393 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
394 			     MLX4_CQE_OPCODE_ERROR)) {
395 			struct mlx4_err_cqe *cqe_err = (struct mlx4_err_cqe *)cqe;
396 
397 			en_err(priv, "CQE error - vendor syndrome: 0x%x syndrome: 0x%x\n",
398 			       cqe_err->vendor_err_syndrome,
399 			       cqe_err->syndrome);
400 		}
401 
402 		/* Skip over last polled CQE */
403 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
404 
405 		do {
406 			txbbs_skipped += ring->last_nr_txbb;
407 			ring_index = (ring_index + ring->last_nr_txbb) & size_mask;
408 			if (ring->tx_info[ring_index].ts_requested)
409 				timestamp = mlx4_en_get_cqe_ts(cqe);
410 
411 			/* free next descriptor */
412 			ring->last_nr_txbb = mlx4_en_free_tx_desc(
413 					priv, ring, ring_index,
414 					!!((ring->cons + txbbs_skipped) &
415 					ring->size), timestamp);
416 
417 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
418 					  !!((ring->cons + txbbs_stamp) &
419 						ring->size));
420 			stamp_index = ring_index;
421 			txbbs_stamp = txbbs_skipped;
422 			packets++;
423 			bytes += ring->tx_info[ring_index].nr_bytes;
424 		} while (ring_index != new_index);
425 
426 		++cons_index;
427 		index = cons_index & size_mask;
428 		cqe = &buf[(index << factor) + factor];
429 	}
430 
431 
432 	/*
433 	 * To prevent CQ overflow we first update CQ consumer and only then
434 	 * the ring consumer.
435 	 */
436 	mcq->cons_index = cons_index;
437 	mlx4_cq_set_ci(mcq);
438 	wmb();
439 	ring->cons += txbbs_skipped;
440 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
441 
442 	/*
443 	 * Wakeup Tx queue if this stopped, and at least 1 packet
444 	 * was completed
445 	 */
446 	if (netif_tx_queue_stopped(ring->tx_queue) && txbbs_skipped > 0) {
447 		netif_tx_wake_queue(ring->tx_queue);
448 		priv->port_stats.wake_queue++;
449 	}
450 }
451 
452 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
453 {
454 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
455 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
456 
457 	mlx4_en_process_tx_cq(cq->dev, cq);
458 	mlx4_en_arm_cq(priv, cq);
459 }
460 
461 
462 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
463 						      struct mlx4_en_tx_ring *ring,
464 						      u32 index,
465 						      unsigned int desc_size)
466 {
467 	u32 copy = (ring->size - index) * TXBB_SIZE;
468 	int i;
469 
470 	for (i = desc_size - copy - 4; i >= 0; i -= 4) {
471 		if ((i & (TXBB_SIZE - 1)) == 0)
472 			wmb();
473 
474 		*((u32 *) (ring->buf + i)) =
475 			*((u32 *) (ring->bounce_buf + copy + i));
476 	}
477 
478 	for (i = copy - 4; i >= 4 ; i -= 4) {
479 		if ((i & (TXBB_SIZE - 1)) == 0)
480 			wmb();
481 
482 		*((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
483 			*((u32 *) (ring->bounce_buf + i));
484 	}
485 
486 	/* Return real descriptor location */
487 	return ring->buf + index * TXBB_SIZE;
488 }
489 
490 static int is_inline(struct sk_buff *skb, void **pfrag)
491 {
492 	void *ptr;
493 
494 	if (inline_thold && !skb_is_gso(skb) && skb->len <= inline_thold) {
495 		if (skb_shinfo(skb)->nr_frags == 1) {
496 			ptr = skb_frag_address_safe(&skb_shinfo(skb)->frags[0]);
497 			if (unlikely(!ptr))
498 				return 0;
499 
500 			if (pfrag)
501 				*pfrag = ptr;
502 
503 			return 1;
504 		} else if (unlikely(skb_shinfo(skb)->nr_frags))
505 			return 0;
506 		else
507 			return 1;
508 	}
509 
510 	return 0;
511 }
512 
513 static int inline_size(struct sk_buff *skb)
514 {
515 	if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
516 	    <= MLX4_INLINE_ALIGN)
517 		return ALIGN(skb->len + CTRL_SIZE +
518 			     sizeof(struct mlx4_wqe_inline_seg), 16);
519 	else
520 		return ALIGN(skb->len + CTRL_SIZE + 2 *
521 			     sizeof(struct mlx4_wqe_inline_seg), 16);
522 }
523 
524 static int get_real_size(struct sk_buff *skb, struct net_device *dev,
525 			 int *lso_header_size)
526 {
527 	struct mlx4_en_priv *priv = netdev_priv(dev);
528 	int real_size;
529 
530 	if (skb_is_gso(skb)) {
531 		*lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
532 		real_size = CTRL_SIZE + skb_shinfo(skb)->nr_frags * DS_SIZE +
533 			ALIGN(*lso_header_size + 4, DS_SIZE);
534 		if (unlikely(*lso_header_size != skb_headlen(skb))) {
535 			/* We add a segment for the skb linear buffer only if
536 			 * it contains data */
537 			if (*lso_header_size < skb_headlen(skb))
538 				real_size += DS_SIZE;
539 			else {
540 				if (netif_msg_tx_err(priv))
541 					en_warn(priv, "Non-linear headers\n");
542 				return 0;
543 			}
544 		}
545 	} else {
546 		*lso_header_size = 0;
547 		if (!is_inline(skb, NULL))
548 			real_size = CTRL_SIZE + (skb_shinfo(skb)->nr_frags + 1) * DS_SIZE;
549 		else
550 			real_size = inline_size(skb);
551 	}
552 
553 	return real_size;
554 }
555 
556 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc, struct sk_buff *skb,
557 			     int real_size, u16 *vlan_tag, int tx_ind, void *fragptr)
558 {
559 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
560 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
561 
562 	if (skb->len <= spc) {
563 		inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
564 		skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
565 		if (skb_shinfo(skb)->nr_frags)
566 			memcpy(((void *)(inl + 1)) + skb_headlen(skb), fragptr,
567 			       skb_frag_size(&skb_shinfo(skb)->frags[0]));
568 
569 	} else {
570 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
571 		if (skb_headlen(skb) <= spc) {
572 			skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
573 			if (skb_headlen(skb) < spc) {
574 				memcpy(((void *)(inl + 1)) + skb_headlen(skb),
575 					fragptr, spc - skb_headlen(skb));
576 				fragptr +=  spc - skb_headlen(skb);
577 			}
578 			inl = (void *) (inl + 1) + spc;
579 			memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
580 		} else {
581 			skb_copy_from_linear_data(skb, inl + 1, spc);
582 			inl = (void *) (inl + 1) + spc;
583 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
584 					skb_headlen(skb) - spc);
585 			if (skb_shinfo(skb)->nr_frags)
586 				memcpy(((void *)(inl + 1)) + skb_headlen(skb) - spc,
587 					fragptr, skb_frag_size(&skb_shinfo(skb)->frags[0]));
588 		}
589 
590 		wmb();
591 		inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
592 	}
593 }
594 
595 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb)
596 {
597 	struct mlx4_en_priv *priv = netdev_priv(dev);
598 	u16 rings_p_up = priv->num_tx_rings_p_up;
599 	u8 up = 0;
600 
601 	if (dev->num_tc)
602 		return skb_tx_hash(dev, skb);
603 
604 	if (vlan_tx_tag_present(skb))
605 		up = vlan_tx_tag_get(skb) >> VLAN_PRIO_SHIFT;
606 
607 	return __netdev_pick_tx(dev, skb) % rings_p_up + up * rings_p_up;
608 }
609 
610 static void mlx4_bf_copy(void __iomem *dst, unsigned long *src, unsigned bytecnt)
611 {
612 	__iowrite64_copy(dst, src, bytecnt / 8);
613 }
614 
615 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
616 {
617 	struct mlx4_en_priv *priv = netdev_priv(dev);
618 	struct mlx4_en_dev *mdev = priv->mdev;
619 	struct device *ddev = priv->ddev;
620 	struct mlx4_en_tx_ring *ring;
621 	struct mlx4_en_tx_desc *tx_desc;
622 	struct mlx4_wqe_data_seg *data;
623 	struct mlx4_en_tx_info *tx_info;
624 	int tx_ind = 0;
625 	int nr_txbb;
626 	int desc_size;
627 	int real_size;
628 	u32 index, bf_index;
629 	__be32 op_own;
630 	u16 vlan_tag = 0;
631 	int i;
632 	int lso_header_size;
633 	void *fragptr;
634 	bool bounce = false;
635 
636 	if (!priv->port_up)
637 		goto tx_drop;
638 
639 	real_size = get_real_size(skb, dev, &lso_header_size);
640 	if (unlikely(!real_size))
641 		goto tx_drop;
642 
643 	/* Align descriptor to TXBB size */
644 	desc_size = ALIGN(real_size, TXBB_SIZE);
645 	nr_txbb = desc_size / TXBB_SIZE;
646 	if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
647 		if (netif_msg_tx_err(priv))
648 			en_warn(priv, "Oversized header or SG list\n");
649 		goto tx_drop;
650 	}
651 
652 	tx_ind = skb->queue_mapping;
653 	ring = priv->tx_ring[tx_ind];
654 	if (vlan_tx_tag_present(skb))
655 		vlan_tag = vlan_tx_tag_get(skb);
656 
657 	/* Check available TXBBs And 2K spare for prefetch */
658 	if (unlikely(((int)(ring->prod - ring->cons)) >
659 		     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
660 		/* every full Tx ring stops queue */
661 		netif_tx_stop_queue(ring->tx_queue);
662 		priv->port_stats.queue_stopped++;
663 
664 		/* If queue was emptied after the if, and before the
665 		 * stop_queue - need to wake the queue, or else it will remain
666 		 * stopped forever.
667 		 * Need a memory barrier to make sure ring->cons was not
668 		 * updated before queue was stopped.
669 		 */
670 		wmb();
671 
672 		if (unlikely(((int)(ring->prod - ring->cons)) <=
673 			     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
674 			netif_tx_wake_queue(ring->tx_queue);
675 			priv->port_stats.wake_queue++;
676 		} else {
677 			return NETDEV_TX_BUSY;
678 		}
679 	}
680 
681 	/* Track current inflight packets for performance analysis */
682 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
683 			 (u32) (ring->prod - ring->cons - 1));
684 
685 	/* Packet is good - grab an index and transmit it */
686 	index = ring->prod & ring->size_mask;
687 	bf_index = ring->prod;
688 
689 	/* See if we have enough space for whole descriptor TXBB for setting
690 	 * SW ownership on next descriptor; if not, use a bounce buffer. */
691 	if (likely(index + nr_txbb <= ring->size))
692 		tx_desc = ring->buf + index * TXBB_SIZE;
693 	else {
694 		tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
695 		bounce = true;
696 	}
697 
698 	/* Save skb in tx_info ring */
699 	tx_info = &ring->tx_info[index];
700 	tx_info->skb = skb;
701 	tx_info->nr_txbb = nr_txbb;
702 
703 	if (lso_header_size)
704 		data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
705 						      DS_SIZE));
706 	else
707 		data = &tx_desc->data;
708 
709 	/* valid only for none inline segments */
710 	tx_info->data_offset = (void *)data - (void *)tx_desc;
711 
712 	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
713 			   !is_inline(skb, NULL)) ? 1 : 0;
714 
715 	data += skb_shinfo(skb)->nr_frags + tx_info->linear - 1;
716 
717 	if (is_inline(skb, &fragptr)) {
718 		tx_info->inl = 1;
719 	} else {
720 		/* Map fragments */
721 		for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
722 			struct skb_frag_struct *frag;
723 			dma_addr_t dma;
724 
725 			frag = &skb_shinfo(skb)->frags[i];
726 			dma = skb_frag_dma_map(ddev, frag,
727 					       0, skb_frag_size(frag),
728 					       DMA_TO_DEVICE);
729 			if (dma_mapping_error(ddev, dma))
730 				goto tx_drop_unmap;
731 
732 			data->addr = cpu_to_be64(dma);
733 			data->lkey = cpu_to_be32(mdev->mr.key);
734 			wmb();
735 			data->byte_count = cpu_to_be32(skb_frag_size(frag));
736 			--data;
737 		}
738 
739 		/* Map linear part */
740 		if (tx_info->linear) {
741 			u32 byte_count = skb_headlen(skb) - lso_header_size;
742 			dma_addr_t dma;
743 
744 			dma = dma_map_single(ddev, skb->data +
745 					     lso_header_size, byte_count,
746 					     PCI_DMA_TODEVICE);
747 			if (dma_mapping_error(ddev, dma))
748 				goto tx_drop_unmap;
749 
750 			data->addr = cpu_to_be64(dma);
751 			data->lkey = cpu_to_be32(mdev->mr.key);
752 			wmb();
753 			data->byte_count = cpu_to_be32(byte_count);
754 		}
755 		tx_info->inl = 0;
756 	}
757 
758 	/*
759 	 * For timestamping add flag to skb_shinfo and
760 	 * set flag for further reference
761 	 */
762 	if (ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
763 	    skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
764 		skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
765 		tx_info->ts_requested = 1;
766 	}
767 
768 	/* Prepare ctrl segement apart opcode+ownership, which depends on
769 	 * whether LSO is used */
770 	tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
771 	tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN *
772 		!!vlan_tx_tag_present(skb);
773 	tx_desc->ctrl.fence_size = (real_size / 16) & 0x3f;
774 	tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
775 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
776 		tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
777 							 MLX4_WQE_CTRL_TCP_UDP_CSUM);
778 		ring->tx_csum++;
779 	}
780 
781 	if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
782 		struct ethhdr *ethh;
783 
784 		/* Copy dst mac address to wqe. This allows loopback in eSwitch,
785 		 * so that VFs and PF can communicate with each other
786 		 */
787 		ethh = (struct ethhdr *)skb->data;
788 		tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
789 		tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
790 	}
791 
792 	/* Handle LSO (TSO) packets */
793 	if (lso_header_size) {
794 		/* Mark opcode as LSO */
795 		op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
796 			((ring->prod & ring->size) ?
797 				cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
798 
799 		/* Fill in the LSO prefix */
800 		tx_desc->lso.mss_hdr_size = cpu_to_be32(
801 			skb_shinfo(skb)->gso_size << 16 | lso_header_size);
802 
803 		/* Copy headers;
804 		 * note that we already verified that it is linear */
805 		memcpy(tx_desc->lso.header, skb->data, lso_header_size);
806 
807 		priv->port_stats.tso_packets++;
808 		i = ((skb->len - lso_header_size) / skb_shinfo(skb)->gso_size) +
809 			!!((skb->len - lso_header_size) % skb_shinfo(skb)->gso_size);
810 		tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
811 		ring->packets += i;
812 	} else {
813 		/* Normal (Non LSO) packet */
814 		op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
815 			((ring->prod & ring->size) ?
816 			 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
817 		tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
818 		ring->packets++;
819 
820 	}
821 	ring->bytes += tx_info->nr_bytes;
822 	netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
823 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
824 
825 	if (tx_info->inl) {
826 		build_inline_wqe(tx_desc, skb, real_size, &vlan_tag, tx_ind, fragptr);
827 		tx_info->inl = 1;
828 	}
829 
830 	ring->prod += nr_txbb;
831 
832 	/* If we used a bounce buffer then copy descriptor back into place */
833 	if (bounce)
834 		tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
835 
836 	skb_tx_timestamp(skb);
837 
838 	if (ring->bf_enabled && desc_size <= MAX_BF && !bounce && !vlan_tx_tag_present(skb)) {
839 		*(__be32 *) (&tx_desc->ctrl.vlan_tag) |= cpu_to_be32(ring->doorbell_qpn);
840 		op_own |= htonl((bf_index & 0xffff) << 8);
841 		/* Ensure new descirptor hits memory
842 		* before setting ownership of this descriptor to HW */
843 		wmb();
844 		tx_desc->ctrl.owner_opcode = op_own;
845 
846 		wmb();
847 
848 		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, (unsigned long *) &tx_desc->ctrl,
849 		     desc_size);
850 
851 		wmb();
852 
853 		ring->bf.offset ^= ring->bf.buf_size;
854 	} else {
855 		/* Ensure new descirptor hits memory
856 		* before setting ownership of this descriptor to HW */
857 		wmb();
858 		tx_desc->ctrl.owner_opcode = op_own;
859 		wmb();
860 		iowrite32be(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL);
861 	}
862 
863 	return NETDEV_TX_OK;
864 
865 tx_drop_unmap:
866 	en_err(priv, "DMA mapping error\n");
867 
868 	for (i++; i < skb_shinfo(skb)->nr_frags; i++) {
869 		data++;
870 		dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
871 			       be32_to_cpu(data->byte_count),
872 			       PCI_DMA_TODEVICE);
873 	}
874 
875 tx_drop:
876 	dev_kfree_skb_any(skb);
877 	priv->stats.tx_dropped++;
878 	return NETDEV_TX_OK;
879 }
880 
881