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/ip.h>
43 #include <linux/moduleparam.h>
44 
45 #include "mlx4_en.h"
46 
47 enum {
48 	MAX_INLINE = 104, /* 128 - 16 - 4 - 4 */
49 	MAX_BF = 256,
50 };
51 
52 static int inline_thold __read_mostly = MAX_INLINE;
53 
54 module_param_named(inline_thold, inline_thold, int, 0444);
55 MODULE_PARM_DESC(inline_thold, "threshold for using inline data");
56 
57 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
58 			   struct mlx4_en_tx_ring **pring, int qpn, u32 size,
59 			   u16 stride, int node, int queue_index)
60 {
61 	struct mlx4_en_dev *mdev = priv->mdev;
62 	struct mlx4_en_tx_ring *ring;
63 	int tmp;
64 	int err;
65 
66 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
67 	if (!ring) {
68 		ring = kzalloc(sizeof(*ring), GFP_KERNEL);
69 		if (!ring) {
70 			en_err(priv, "Failed allocating TX ring\n");
71 			return -ENOMEM;
72 		}
73 	}
74 
75 	ring->size = size;
76 	ring->size_mask = size - 1;
77 	ring->stride = stride;
78 
79 	inline_thold = min(inline_thold, MAX_INLINE);
80 
81 	tmp = size * sizeof(struct mlx4_en_tx_info);
82 	ring->tx_info = vmalloc_node(tmp, node);
83 	if (!ring->tx_info) {
84 		ring->tx_info = vmalloc(tmp);
85 		if (!ring->tx_info) {
86 			err = -ENOMEM;
87 			goto err_ring;
88 		}
89 	}
90 
91 	en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
92 		 ring->tx_info, tmp);
93 
94 	ring->bounce_buf = kmalloc_node(MAX_DESC_SIZE, GFP_KERNEL, node);
95 	if (!ring->bounce_buf) {
96 		ring->bounce_buf = kmalloc(MAX_DESC_SIZE, GFP_KERNEL);
97 		if (!ring->bounce_buf) {
98 			err = -ENOMEM;
99 			goto err_info;
100 		}
101 	}
102 	ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
103 
104 	/* Allocate HW buffers on provided NUMA node */
105 	set_dev_node(&mdev->dev->pdev->dev, node);
106 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size,
107 				 2 * PAGE_SIZE);
108 	set_dev_node(&mdev->dev->pdev->dev, mdev->dev->numa_node);
109 	if (err) {
110 		en_err(priv, "Failed allocating hwq resources\n");
111 		goto err_bounce;
112 	}
113 
114 	err = mlx4_en_map_buffer(&ring->wqres.buf);
115 	if (err) {
116 		en_err(priv, "Failed to map TX buffer\n");
117 		goto err_hwq_res;
118 	}
119 
120 	ring->buf = ring->wqres.buf.direct.buf;
121 
122 	en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d "
123 	       "buf_size:%d dma:%llx\n", ring, ring->buf, ring->size,
124 	       ring->buf_size, (unsigned long long) ring->wqres.buf.direct.map);
125 
126 	ring->qpn = qpn;
127 	err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp);
128 	if (err) {
129 		en_err(priv, "Failed allocating qp %d\n", ring->qpn);
130 		goto err_map;
131 	}
132 	ring->qp.event = mlx4_en_sqp_event;
133 
134 	err = mlx4_bf_alloc(mdev->dev, &ring->bf, node);
135 	if (err) {
136 		en_dbg(DRV, priv, "working without blueflame (%d)", err);
137 		ring->bf.uar = &mdev->priv_uar;
138 		ring->bf.uar->map = mdev->uar_map;
139 		ring->bf_enabled = false;
140 	} else
141 		ring->bf_enabled = true;
142 
143 	ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type;
144 	ring->queue_index = queue_index;
145 
146 	if (queue_index < priv->num_tx_rings_p_up && cpu_online(queue_index))
147 		cpumask_set_cpu(queue_index, &ring->affinity_mask);
148 
149 	*pring = ring;
150 	return 0;
151 
152 err_map:
153 	mlx4_en_unmap_buffer(&ring->wqres.buf);
154 err_hwq_res:
155 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
156 err_bounce:
157 	kfree(ring->bounce_buf);
158 	ring->bounce_buf = NULL;
159 err_info:
160 	vfree(ring->tx_info);
161 	ring->tx_info = NULL;
162 err_ring:
163 	kfree(ring);
164 	*pring = NULL;
165 	return err;
166 }
167 
168 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
169 			     struct mlx4_en_tx_ring **pring)
170 {
171 	struct mlx4_en_dev *mdev = priv->mdev;
172 	struct mlx4_en_tx_ring *ring = *pring;
173 	en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
174 
175 	if (ring->bf_enabled)
176 		mlx4_bf_free(mdev->dev, &ring->bf);
177 	mlx4_qp_remove(mdev->dev, &ring->qp);
178 	mlx4_qp_free(mdev->dev, &ring->qp);
179 	mlx4_en_unmap_buffer(&ring->wqres.buf);
180 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
181 	kfree(ring->bounce_buf);
182 	ring->bounce_buf = NULL;
183 	vfree(ring->tx_info);
184 	ring->tx_info = NULL;
185 	kfree(ring);
186 	*pring = NULL;
187 }
188 
189 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
190 			     struct mlx4_en_tx_ring *ring,
191 			     int cq, int user_prio)
192 {
193 	struct mlx4_en_dev *mdev = priv->mdev;
194 	int err;
195 
196 	ring->cqn = cq;
197 	ring->prod = 0;
198 	ring->cons = 0xffffffff;
199 	ring->last_nr_txbb = 1;
200 	ring->poll_cnt = 0;
201 	memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
202 	memset(ring->buf, 0, ring->buf_size);
203 
204 	ring->qp_state = MLX4_QP_STATE_RST;
205 	ring->doorbell_qpn = ring->qp.qpn << 8;
206 
207 	mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
208 				ring->cqn, user_prio, &ring->context);
209 	if (ring->bf_enabled)
210 		ring->context.usr_page = cpu_to_be32(ring->bf.uar->index);
211 
212 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
213 			       &ring->qp, &ring->qp_state);
214 	if (!user_prio && cpu_online(ring->queue_index))
215 		netif_set_xps_queue(priv->dev, &ring->affinity_mask,
216 				    ring->queue_index);
217 
218 	return err;
219 }
220 
221 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
222 				struct mlx4_en_tx_ring *ring)
223 {
224 	struct mlx4_en_dev *mdev = priv->mdev;
225 
226 	mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
227 		       MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
228 }
229 
230 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
231 			      struct mlx4_en_tx_ring *ring, int index,
232 			      u8 owner)
233 {
234 	__be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
235 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
236 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
237 	void *end = ring->buf + ring->buf_size;
238 	__be32 *ptr = (__be32 *)tx_desc;
239 	int i;
240 
241 	/* Optimize the common case when there are no wraparounds */
242 	if (likely((void *)tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
243 		/* Stamp the freed descriptor */
244 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
245 		     i += STAMP_STRIDE) {
246 			*ptr = stamp;
247 			ptr += STAMP_DWORDS;
248 		}
249 	} else {
250 		/* Stamp the freed descriptor */
251 		for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
252 		     i += STAMP_STRIDE) {
253 			*ptr = stamp;
254 			ptr += STAMP_DWORDS;
255 			if ((void *)ptr >= end) {
256 				ptr = ring->buf;
257 				stamp ^= cpu_to_be32(0x80000000);
258 			}
259 		}
260 	}
261 }
262 
263 
264 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
265 				struct mlx4_en_tx_ring *ring,
266 				int index, u8 owner, u64 timestamp)
267 {
268 	struct mlx4_en_dev *mdev = priv->mdev;
269 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
270 	struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
271 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
272 	struct sk_buff *skb = tx_info->skb;
273 	struct skb_frag_struct *frag;
274 	void *end = ring->buf + ring->buf_size;
275 	int frags = skb_shinfo(skb)->nr_frags;
276 	int i;
277 	struct skb_shared_hwtstamps hwts;
278 
279 	if (timestamp) {
280 		mlx4_en_fill_hwtstamps(mdev, &hwts, timestamp);
281 		skb_tstamp_tx(skb, &hwts);
282 	}
283 
284 	/* Optimize the common case when there are no wraparounds */
285 	if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
286 		if (!tx_info->inl) {
287 			if (tx_info->linear) {
288 				dma_unmap_single(priv->ddev,
289 					(dma_addr_t) be64_to_cpu(data->addr),
290 					 be32_to_cpu(data->byte_count),
291 					 PCI_DMA_TODEVICE);
292 				++data;
293 			}
294 
295 			for (i = 0; i < frags; i++) {
296 				frag = &skb_shinfo(skb)->frags[i];
297 				dma_unmap_page(priv->ddev,
298 					(dma_addr_t) be64_to_cpu(data[i].addr),
299 					skb_frag_size(frag), PCI_DMA_TODEVICE);
300 			}
301 		}
302 	} else {
303 		if (!tx_info->inl) {
304 			if ((void *) data >= end) {
305 				data = ring->buf + ((void *)data - end);
306 			}
307 
308 			if (tx_info->linear) {
309 				dma_unmap_single(priv->ddev,
310 					(dma_addr_t) be64_to_cpu(data->addr),
311 					 be32_to_cpu(data->byte_count),
312 					 PCI_DMA_TODEVICE);
313 				++data;
314 			}
315 
316 			for (i = 0; i < frags; i++) {
317 				/* Check for wraparound before unmapping */
318 				if ((void *) data >= end)
319 					data = ring->buf;
320 				frag = &skb_shinfo(skb)->frags[i];
321 				dma_unmap_page(priv->ddev,
322 					(dma_addr_t) be64_to_cpu(data->addr),
323 					 skb_frag_size(frag), PCI_DMA_TODEVICE);
324 				++data;
325 			}
326 		}
327 	}
328 	dev_kfree_skb(skb);
329 	return tx_info->nr_txbb;
330 }
331 
332 
333 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
334 {
335 	struct mlx4_en_priv *priv = netdev_priv(dev);
336 	int cnt = 0;
337 
338 	/* Skip last polled descriptor */
339 	ring->cons += ring->last_nr_txbb;
340 	en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
341 		 ring->cons, ring->prod);
342 
343 	if ((u32) (ring->prod - ring->cons) > ring->size) {
344 		if (netif_msg_tx_err(priv))
345 			en_warn(priv, "Tx consumer passed producer!\n");
346 		return 0;
347 	}
348 
349 	while (ring->cons != ring->prod) {
350 		ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
351 						ring->cons & ring->size_mask,
352 						!!(ring->cons & ring->size), 0);
353 		ring->cons += ring->last_nr_txbb;
354 		cnt++;
355 	}
356 
357 	netdev_tx_reset_queue(ring->tx_queue);
358 
359 	if (cnt)
360 		en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
361 
362 	return cnt;
363 }
364 
365 static int mlx4_en_process_tx_cq(struct net_device *dev,
366 				 struct mlx4_en_cq *cq,
367 				 int budget)
368 {
369 	struct mlx4_en_priv *priv = netdev_priv(dev);
370 	struct mlx4_cq *mcq = &cq->mcq;
371 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->ring];
372 	struct mlx4_cqe *cqe;
373 	u16 index;
374 	u16 new_index, ring_index, stamp_index;
375 	u32 txbbs_skipped = 0;
376 	u32 txbbs_stamp = 0;
377 	u32 cons_index = mcq->cons_index;
378 	int size = cq->size;
379 	u32 size_mask = ring->size_mask;
380 	struct mlx4_cqe *buf = cq->buf;
381 	u32 packets = 0;
382 	u32 bytes = 0;
383 	int factor = priv->cqe_factor;
384 	u64 timestamp = 0;
385 	int done = 0;
386 
387 	if (!priv->port_up)
388 		return 0;
389 
390 	index = cons_index & size_mask;
391 	cqe = &buf[(index << factor) + factor];
392 	ring_index = ring->cons & size_mask;
393 	stamp_index = ring_index;
394 
395 	/* Process all completed CQEs */
396 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
397 			cons_index & size) && (done < budget)) {
398 		/*
399 		 * make sure we read the CQE after we read the
400 		 * ownership bit
401 		 */
402 		rmb();
403 
404 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
405 			     MLX4_CQE_OPCODE_ERROR)) {
406 			struct mlx4_err_cqe *cqe_err = (struct mlx4_err_cqe *)cqe;
407 
408 			en_err(priv, "CQE error - vendor syndrome: 0x%x syndrome: 0x%x\n",
409 			       cqe_err->vendor_err_syndrome,
410 			       cqe_err->syndrome);
411 		}
412 
413 		/* Skip over last polled CQE */
414 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
415 
416 		do {
417 			txbbs_skipped += ring->last_nr_txbb;
418 			ring_index = (ring_index + ring->last_nr_txbb) & size_mask;
419 			if (ring->tx_info[ring_index].ts_requested)
420 				timestamp = mlx4_en_get_cqe_ts(cqe);
421 
422 			/* free next descriptor */
423 			ring->last_nr_txbb = mlx4_en_free_tx_desc(
424 					priv, ring, ring_index,
425 					!!((ring->cons + txbbs_skipped) &
426 					ring->size), timestamp);
427 
428 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
429 					  !!((ring->cons + txbbs_stamp) &
430 						ring->size));
431 			stamp_index = ring_index;
432 			txbbs_stamp = txbbs_skipped;
433 			packets++;
434 			bytes += ring->tx_info[ring_index].nr_bytes;
435 		} while ((++done < budget) && (ring_index != new_index));
436 
437 		++cons_index;
438 		index = cons_index & size_mask;
439 		cqe = &buf[(index << factor) + factor];
440 	}
441 
442 
443 	/*
444 	 * To prevent CQ overflow we first update CQ consumer and only then
445 	 * the ring consumer.
446 	 */
447 	mcq->cons_index = cons_index;
448 	mlx4_cq_set_ci(mcq);
449 	wmb();
450 	ring->cons += txbbs_skipped;
451 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
452 
453 	/*
454 	 * Wakeup Tx queue if this stopped, and at least 1 packet
455 	 * was completed
456 	 */
457 	if (netif_tx_queue_stopped(ring->tx_queue) && txbbs_skipped > 0) {
458 		netif_tx_wake_queue(ring->tx_queue);
459 		priv->port_stats.wake_queue++;
460 	}
461 	return done;
462 }
463 
464 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
465 {
466 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
467 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
468 
469 	if (priv->port_up)
470 		napi_schedule(&cq->napi);
471 	else
472 		mlx4_en_arm_cq(priv, cq);
473 }
474 
475 /* TX CQ polling - called by NAPI */
476 int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget)
477 {
478 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
479 	struct net_device *dev = cq->dev;
480 	struct mlx4_en_priv *priv = netdev_priv(dev);
481 	int done;
482 
483 	done = mlx4_en_process_tx_cq(dev, cq, budget);
484 
485 	/* If we used up all the quota - we're probably not done yet... */
486 	if (done < budget) {
487 		/* Done for now */
488 		napi_complete(napi);
489 		mlx4_en_arm_cq(priv, cq);
490 		return done;
491 	}
492 	return budget;
493 }
494 
495 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
496 						      struct mlx4_en_tx_ring *ring,
497 						      u32 index,
498 						      unsigned int desc_size)
499 {
500 	u32 copy = (ring->size - index) * TXBB_SIZE;
501 	int i;
502 
503 	for (i = desc_size - copy - 4; i >= 0; i -= 4) {
504 		if ((i & (TXBB_SIZE - 1)) == 0)
505 			wmb();
506 
507 		*((u32 *) (ring->buf + i)) =
508 			*((u32 *) (ring->bounce_buf + copy + i));
509 	}
510 
511 	for (i = copy - 4; i >= 4 ; i -= 4) {
512 		if ((i & (TXBB_SIZE - 1)) == 0)
513 			wmb();
514 
515 		*((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
516 			*((u32 *) (ring->bounce_buf + i));
517 	}
518 
519 	/* Return real descriptor location */
520 	return ring->buf + index * TXBB_SIZE;
521 }
522 
523 static int is_inline(struct sk_buff *skb, void **pfrag)
524 {
525 	void *ptr;
526 
527 	if (inline_thold && !skb_is_gso(skb) && skb->len <= inline_thold) {
528 		if (skb_shinfo(skb)->nr_frags == 1) {
529 			ptr = skb_frag_address_safe(&skb_shinfo(skb)->frags[0]);
530 			if (unlikely(!ptr))
531 				return 0;
532 
533 			if (pfrag)
534 				*pfrag = ptr;
535 
536 			return 1;
537 		} else if (unlikely(skb_shinfo(skb)->nr_frags))
538 			return 0;
539 		else
540 			return 1;
541 	}
542 
543 	return 0;
544 }
545 
546 static int inline_size(struct sk_buff *skb)
547 {
548 	if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
549 	    <= MLX4_INLINE_ALIGN)
550 		return ALIGN(skb->len + CTRL_SIZE +
551 			     sizeof(struct mlx4_wqe_inline_seg), 16);
552 	else
553 		return ALIGN(skb->len + CTRL_SIZE + 2 *
554 			     sizeof(struct mlx4_wqe_inline_seg), 16);
555 }
556 
557 static int get_real_size(struct sk_buff *skb, struct net_device *dev,
558 			 int *lso_header_size)
559 {
560 	struct mlx4_en_priv *priv = netdev_priv(dev);
561 	int real_size;
562 
563 	if (skb_is_gso(skb)) {
564 		if (skb->encapsulation)
565 			*lso_header_size = (skb_inner_transport_header(skb) - skb->data) + inner_tcp_hdrlen(skb);
566 		else
567 			*lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
568 		real_size = CTRL_SIZE + skb_shinfo(skb)->nr_frags * DS_SIZE +
569 			ALIGN(*lso_header_size + 4, DS_SIZE);
570 		if (unlikely(*lso_header_size != skb_headlen(skb))) {
571 			/* We add a segment for the skb linear buffer only if
572 			 * it contains data */
573 			if (*lso_header_size < skb_headlen(skb))
574 				real_size += DS_SIZE;
575 			else {
576 				if (netif_msg_tx_err(priv))
577 					en_warn(priv, "Non-linear headers\n");
578 				return 0;
579 			}
580 		}
581 	} else {
582 		*lso_header_size = 0;
583 		if (!is_inline(skb, NULL))
584 			real_size = CTRL_SIZE + (skb_shinfo(skb)->nr_frags + 1) * DS_SIZE;
585 		else
586 			real_size = inline_size(skb);
587 	}
588 
589 	return real_size;
590 }
591 
592 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc, struct sk_buff *skb,
593 			     int real_size, u16 *vlan_tag, int tx_ind, void *fragptr)
594 {
595 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
596 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
597 
598 	if (skb->len <= spc) {
599 		inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
600 		skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
601 		if (skb_shinfo(skb)->nr_frags)
602 			memcpy(((void *)(inl + 1)) + skb_headlen(skb), fragptr,
603 			       skb_frag_size(&skb_shinfo(skb)->frags[0]));
604 
605 	} else {
606 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
607 		if (skb_headlen(skb) <= spc) {
608 			skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
609 			if (skb_headlen(skb) < spc) {
610 				memcpy(((void *)(inl + 1)) + skb_headlen(skb),
611 					fragptr, spc - skb_headlen(skb));
612 				fragptr +=  spc - skb_headlen(skb);
613 			}
614 			inl = (void *) (inl + 1) + spc;
615 			memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
616 		} else {
617 			skb_copy_from_linear_data(skb, inl + 1, spc);
618 			inl = (void *) (inl + 1) + spc;
619 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
620 					skb_headlen(skb) - spc);
621 			if (skb_shinfo(skb)->nr_frags)
622 				memcpy(((void *)(inl + 1)) + skb_headlen(skb) - spc,
623 					fragptr, skb_frag_size(&skb_shinfo(skb)->frags[0]));
624 		}
625 
626 		wmb();
627 		inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
628 	}
629 }
630 
631 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb,
632 			 void *accel_priv)
633 {
634 	struct mlx4_en_priv *priv = netdev_priv(dev);
635 	u16 rings_p_up = priv->num_tx_rings_p_up;
636 	u8 up = 0;
637 
638 	if (dev->num_tc)
639 		return skb_tx_hash(dev, skb);
640 
641 	if (vlan_tx_tag_present(skb))
642 		up = vlan_tx_tag_get(skb) >> VLAN_PRIO_SHIFT;
643 
644 	return __netdev_pick_tx(dev, skb) % rings_p_up + up * rings_p_up;
645 }
646 
647 static void mlx4_bf_copy(void __iomem *dst, unsigned long *src, unsigned bytecnt)
648 {
649 	__iowrite64_copy(dst, src, bytecnt / 8);
650 }
651 
652 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
653 {
654 	struct mlx4_en_priv *priv = netdev_priv(dev);
655 	struct mlx4_en_dev *mdev = priv->mdev;
656 	struct device *ddev = priv->ddev;
657 	struct mlx4_en_tx_ring *ring;
658 	struct mlx4_en_tx_desc *tx_desc;
659 	struct mlx4_wqe_data_seg *data;
660 	struct mlx4_en_tx_info *tx_info;
661 	int tx_ind = 0;
662 	int nr_txbb;
663 	int desc_size;
664 	int real_size;
665 	u32 index, bf_index;
666 	__be32 op_own;
667 	u16 vlan_tag = 0;
668 	int i;
669 	int lso_header_size;
670 	void *fragptr;
671 	bool bounce = false;
672 
673 	if (!priv->port_up)
674 		goto tx_drop;
675 
676 	real_size = get_real_size(skb, dev, &lso_header_size);
677 	if (unlikely(!real_size))
678 		goto tx_drop;
679 
680 	/* Align descriptor to TXBB size */
681 	desc_size = ALIGN(real_size, TXBB_SIZE);
682 	nr_txbb = desc_size / TXBB_SIZE;
683 	if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
684 		if (netif_msg_tx_err(priv))
685 			en_warn(priv, "Oversized header or SG list\n");
686 		goto tx_drop;
687 	}
688 
689 	tx_ind = skb->queue_mapping;
690 	ring = priv->tx_ring[tx_ind];
691 	if (vlan_tx_tag_present(skb))
692 		vlan_tag = vlan_tx_tag_get(skb);
693 
694 	/* Check available TXBBs And 2K spare for prefetch */
695 	if (unlikely(((int)(ring->prod - ring->cons)) >
696 		     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
697 		/* every full Tx ring stops queue */
698 		netif_tx_stop_queue(ring->tx_queue);
699 		priv->port_stats.queue_stopped++;
700 
701 		/* If queue was emptied after the if, and before the
702 		 * stop_queue - need to wake the queue, or else it will remain
703 		 * stopped forever.
704 		 * Need a memory barrier to make sure ring->cons was not
705 		 * updated before queue was stopped.
706 		 */
707 		wmb();
708 
709 		if (unlikely(((int)(ring->prod - ring->cons)) <=
710 			     ring->size - HEADROOM - MAX_DESC_TXBBS)) {
711 			netif_tx_wake_queue(ring->tx_queue);
712 			priv->port_stats.wake_queue++;
713 		} else {
714 			return NETDEV_TX_BUSY;
715 		}
716 	}
717 
718 	/* Track current inflight packets for performance analysis */
719 	AVG_PERF_COUNTER(priv->pstats.inflight_avg,
720 			 (u32) (ring->prod - ring->cons - 1));
721 
722 	/* Packet is good - grab an index and transmit it */
723 	index = ring->prod & ring->size_mask;
724 	bf_index = ring->prod;
725 
726 	/* See if we have enough space for whole descriptor TXBB for setting
727 	 * SW ownership on next descriptor; if not, use a bounce buffer. */
728 	if (likely(index + nr_txbb <= ring->size))
729 		tx_desc = ring->buf + index * TXBB_SIZE;
730 	else {
731 		tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
732 		bounce = true;
733 	}
734 
735 	/* Save skb in tx_info ring */
736 	tx_info = &ring->tx_info[index];
737 	tx_info->skb = skb;
738 	tx_info->nr_txbb = nr_txbb;
739 
740 	if (lso_header_size)
741 		data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
742 						      DS_SIZE));
743 	else
744 		data = &tx_desc->data;
745 
746 	/* valid only for none inline segments */
747 	tx_info->data_offset = (void *)data - (void *)tx_desc;
748 
749 	tx_info->linear = (lso_header_size < skb_headlen(skb) &&
750 			   !is_inline(skb, NULL)) ? 1 : 0;
751 
752 	data += skb_shinfo(skb)->nr_frags + tx_info->linear - 1;
753 
754 	if (is_inline(skb, &fragptr)) {
755 		tx_info->inl = 1;
756 	} else {
757 		/* Map fragments */
758 		for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
759 			struct skb_frag_struct *frag;
760 			dma_addr_t dma;
761 
762 			frag = &skb_shinfo(skb)->frags[i];
763 			dma = skb_frag_dma_map(ddev, frag,
764 					       0, skb_frag_size(frag),
765 					       DMA_TO_DEVICE);
766 			if (dma_mapping_error(ddev, dma))
767 				goto tx_drop_unmap;
768 
769 			data->addr = cpu_to_be64(dma);
770 			data->lkey = cpu_to_be32(mdev->mr.key);
771 			wmb();
772 			data->byte_count = cpu_to_be32(skb_frag_size(frag));
773 			--data;
774 		}
775 
776 		/* Map linear part */
777 		if (tx_info->linear) {
778 			u32 byte_count = skb_headlen(skb) - lso_header_size;
779 			dma_addr_t dma;
780 
781 			dma = dma_map_single(ddev, skb->data +
782 					     lso_header_size, byte_count,
783 					     PCI_DMA_TODEVICE);
784 			if (dma_mapping_error(ddev, dma))
785 				goto tx_drop_unmap;
786 
787 			data->addr = cpu_to_be64(dma);
788 			data->lkey = cpu_to_be32(mdev->mr.key);
789 			wmb();
790 			data->byte_count = cpu_to_be32(byte_count);
791 		}
792 		tx_info->inl = 0;
793 	}
794 
795 	/*
796 	 * For timestamping add flag to skb_shinfo and
797 	 * set flag for further reference
798 	 */
799 	if (ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
800 	    skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
801 		skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
802 		tx_info->ts_requested = 1;
803 	}
804 
805 	/* Prepare ctrl segement apart opcode+ownership, which depends on
806 	 * whether LSO is used */
807 	tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
808 	tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN *
809 		!!vlan_tx_tag_present(skb);
810 	tx_desc->ctrl.fence_size = (real_size / 16) & 0x3f;
811 	tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
812 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
813 		tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
814 							 MLX4_WQE_CTRL_TCP_UDP_CSUM);
815 		ring->tx_csum++;
816 	}
817 
818 	if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
819 		struct ethhdr *ethh;
820 
821 		/* Copy dst mac address to wqe. This allows loopback in eSwitch,
822 		 * so that VFs and PF can communicate with each other
823 		 */
824 		ethh = (struct ethhdr *)skb->data;
825 		tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
826 		tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
827 	}
828 
829 	/* Handle LSO (TSO) packets */
830 	if (lso_header_size) {
831 		/* Mark opcode as LSO */
832 		op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
833 			((ring->prod & ring->size) ?
834 				cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
835 
836 		/* Fill in the LSO prefix */
837 		tx_desc->lso.mss_hdr_size = cpu_to_be32(
838 			skb_shinfo(skb)->gso_size << 16 | lso_header_size);
839 
840 		/* Copy headers;
841 		 * note that we already verified that it is linear */
842 		memcpy(tx_desc->lso.header, skb->data, lso_header_size);
843 
844 		priv->port_stats.tso_packets++;
845 		i = ((skb->len - lso_header_size) / skb_shinfo(skb)->gso_size) +
846 			!!((skb->len - lso_header_size) % skb_shinfo(skb)->gso_size);
847 		tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
848 		ring->packets += i;
849 	} else {
850 		/* Normal (Non LSO) packet */
851 		op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
852 			((ring->prod & ring->size) ?
853 			 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
854 		tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
855 		ring->packets++;
856 
857 	}
858 	ring->bytes += tx_info->nr_bytes;
859 	netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
860 	AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
861 
862 	if (tx_info->inl) {
863 		build_inline_wqe(tx_desc, skb, real_size, &vlan_tag, tx_ind, fragptr);
864 		tx_info->inl = 1;
865 	}
866 
867 	if (skb->encapsulation) {
868 		struct iphdr *ipv4 = (struct iphdr *)skb_inner_network_header(skb);
869 		if (ipv4->protocol == IPPROTO_TCP || ipv4->protocol == IPPROTO_UDP)
870 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP | MLX4_WQE_CTRL_ILP);
871 		else
872 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP);
873 	}
874 
875 	ring->prod += nr_txbb;
876 
877 	/* If we used a bounce buffer then copy descriptor back into place */
878 	if (bounce)
879 		tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
880 
881 	skb_tx_timestamp(skb);
882 
883 	if (ring->bf_enabled && desc_size <= MAX_BF && !bounce && !vlan_tx_tag_present(skb)) {
884 		*(__be32 *) (&tx_desc->ctrl.vlan_tag) |= cpu_to_be32(ring->doorbell_qpn);
885 		op_own |= htonl((bf_index & 0xffff) << 8);
886 		/* Ensure new descirptor hits memory
887 		* before setting ownership of this descriptor to HW */
888 		wmb();
889 		tx_desc->ctrl.owner_opcode = op_own;
890 
891 		wmb();
892 
893 		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, (unsigned long *) &tx_desc->ctrl,
894 		     desc_size);
895 
896 		wmb();
897 
898 		ring->bf.offset ^= ring->bf.buf_size;
899 	} else {
900 		/* Ensure new descirptor hits memory
901 		* before setting ownership of this descriptor to HW */
902 		wmb();
903 		tx_desc->ctrl.owner_opcode = op_own;
904 		wmb();
905 		iowrite32be(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL);
906 	}
907 
908 	return NETDEV_TX_OK;
909 
910 tx_drop_unmap:
911 	en_err(priv, "DMA mapping error\n");
912 
913 	for (i++; i < skb_shinfo(skb)->nr_frags; i++) {
914 		data++;
915 		dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
916 			       be32_to_cpu(data->byte_count),
917 			       PCI_DMA_TODEVICE);
918 	}
919 
920 tx_drop:
921 	dev_kfree_skb_any(skb);
922 	priv->stats.tx_dropped++;
923 	return NETDEV_TX_OK;
924 }
925 
926