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