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/prefetch.h>
41 #include <linux/vmalloc.h>
42 #include <linux/tcp.h>
43 #include <linux/ip.h>
44 #include <linux/ipv6.h>
45 #include <linux/indirect_call_wrapper.h>
46 #include <net/ipv6.h>
47 
48 #include "mlx4_en.h"
49 
50 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
51 			   struct mlx4_en_tx_ring **pring, u32 size,
52 			   u16 stride, int node, int queue_index)
53 {
54 	struct mlx4_en_dev *mdev = priv->mdev;
55 	struct mlx4_en_tx_ring *ring;
56 	int tmp;
57 	int err;
58 
59 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
60 	if (!ring) {
61 		en_err(priv, "Failed allocating TX ring\n");
62 		return -ENOMEM;
63 	}
64 
65 	ring->size = size;
66 	ring->size_mask = size - 1;
67 	ring->sp_stride = stride;
68 	ring->full_size = ring->size - HEADROOM - MLX4_MAX_DESC_TXBBS;
69 
70 	tmp = size * sizeof(struct mlx4_en_tx_info);
71 	ring->tx_info = kvmalloc_node(tmp, GFP_KERNEL, node);
72 	if (!ring->tx_info) {
73 		err = -ENOMEM;
74 		goto err_ring;
75 	}
76 
77 	en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
78 		 ring->tx_info, tmp);
79 
80 	ring->bounce_buf = kmalloc_node(MLX4_TX_BOUNCE_BUFFER_SIZE,
81 					GFP_KERNEL, node);
82 	if (!ring->bounce_buf) {
83 		ring->bounce_buf = kmalloc(MLX4_TX_BOUNCE_BUFFER_SIZE,
84 					   GFP_KERNEL);
85 		if (!ring->bounce_buf) {
86 			err = -ENOMEM;
87 			goto err_info;
88 		}
89 	}
90 	ring->buf_size = ALIGN(size * ring->sp_stride, MLX4_EN_PAGE_SIZE);
91 
92 	/* Allocate HW buffers on provided NUMA node */
93 	set_dev_node(&mdev->dev->persist->pdev->dev, node);
94 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->sp_wqres, ring->buf_size);
95 	set_dev_node(&mdev->dev->persist->pdev->dev, mdev->dev->numa_node);
96 	if (err) {
97 		en_err(priv, "Failed allocating hwq resources\n");
98 		goto err_bounce;
99 	}
100 
101 	ring->buf = ring->sp_wqres.buf.direct.buf;
102 
103 	en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d buf_size:%d dma:%llx\n",
104 	       ring, ring->buf, ring->size, ring->buf_size,
105 	       (unsigned long long) ring->sp_wqres.buf.direct.map);
106 
107 	err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &ring->qpn,
108 				    MLX4_RESERVE_ETH_BF_QP,
109 				    MLX4_RES_USAGE_DRIVER);
110 	if (err) {
111 		en_err(priv, "failed reserving qp for TX ring\n");
112 		goto err_hwq_res;
113 	}
114 
115 	err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->sp_qp);
116 	if (err) {
117 		en_err(priv, "Failed allocating qp %d\n", ring->qpn);
118 		goto err_reserve;
119 	}
120 	ring->sp_qp.event = mlx4_en_sqp_event;
121 
122 	err = mlx4_bf_alloc(mdev->dev, &ring->bf, node);
123 	if (err) {
124 		en_dbg(DRV, priv, "working without blueflame (%d)\n", err);
125 		ring->bf.uar = &mdev->priv_uar;
126 		ring->bf.uar->map = mdev->uar_map;
127 		ring->bf_enabled = false;
128 		ring->bf_alloced = false;
129 		priv->pflags &= ~MLX4_EN_PRIV_FLAGS_BLUEFLAME;
130 	} else {
131 		ring->bf_alloced = true;
132 		ring->bf_enabled = !!(priv->pflags &
133 				      MLX4_EN_PRIV_FLAGS_BLUEFLAME);
134 	}
135 	ring->doorbell_address = ring->bf.uar->map + MLX4_SEND_DOORBELL;
136 
137 	ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type;
138 	ring->queue_index = queue_index;
139 
140 	if (queue_index < priv->num_tx_rings_p_up)
141 		cpumask_set_cpu(cpumask_local_spread(queue_index,
142 						     priv->mdev->dev->numa_node),
143 				&ring->sp_affinity_mask);
144 
145 	*pring = ring;
146 	return 0;
147 
148 err_reserve:
149 	mlx4_qp_release_range(mdev->dev, ring->qpn, 1);
150 err_hwq_res:
151 	mlx4_free_hwq_res(mdev->dev, &ring->sp_wqres, ring->buf_size);
152 err_bounce:
153 	kfree(ring->bounce_buf);
154 	ring->bounce_buf = NULL;
155 err_info:
156 	kvfree(ring->tx_info);
157 	ring->tx_info = NULL;
158 err_ring:
159 	kfree(ring);
160 	*pring = NULL;
161 	return err;
162 }
163 
164 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
165 			     struct mlx4_en_tx_ring **pring)
166 {
167 	struct mlx4_en_dev *mdev = priv->mdev;
168 	struct mlx4_en_tx_ring *ring = *pring;
169 	en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
170 
171 	if (ring->bf_alloced)
172 		mlx4_bf_free(mdev->dev, &ring->bf);
173 	mlx4_qp_remove(mdev->dev, &ring->sp_qp);
174 	mlx4_qp_free(mdev->dev, &ring->sp_qp);
175 	mlx4_qp_release_range(priv->mdev->dev, ring->qpn, 1);
176 	mlx4_free_hwq_res(mdev->dev, &ring->sp_wqres, ring->buf_size);
177 	kfree(ring->bounce_buf);
178 	ring->bounce_buf = NULL;
179 	kvfree(ring->tx_info);
180 	ring->tx_info = NULL;
181 	kfree(ring);
182 	*pring = NULL;
183 }
184 
185 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
186 			     struct mlx4_en_tx_ring *ring,
187 			     int cq, int user_prio)
188 {
189 	struct mlx4_en_dev *mdev = priv->mdev;
190 	int err;
191 
192 	ring->sp_cqn = cq;
193 	ring->prod = 0;
194 	ring->cons = 0xffffffff;
195 	ring->last_nr_txbb = 1;
196 	memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
197 	memset(ring->buf, 0, ring->buf_size);
198 	ring->free_tx_desc = mlx4_en_free_tx_desc;
199 
200 	ring->sp_qp_state = MLX4_QP_STATE_RST;
201 	ring->doorbell_qpn = cpu_to_be32(ring->sp_qp.qpn << 8);
202 	ring->mr_key = cpu_to_be32(mdev->mr.key);
203 
204 	mlx4_en_fill_qp_context(priv, ring->size, ring->sp_stride, 1, 0, ring->qpn,
205 				ring->sp_cqn, user_prio, &ring->sp_context);
206 	if (ring->bf_alloced)
207 		ring->sp_context.usr_page =
208 			cpu_to_be32(mlx4_to_hw_uar_index(mdev->dev,
209 							 ring->bf.uar->index));
210 
211 	err = mlx4_qp_to_ready(mdev->dev, &ring->sp_wqres.mtt, &ring->sp_context,
212 			       &ring->sp_qp, &ring->sp_qp_state);
213 	if (!cpumask_empty(&ring->sp_affinity_mask))
214 		netif_set_xps_queue(priv->dev, &ring->sp_affinity_mask,
215 				    ring->queue_index);
216 
217 	return err;
218 }
219 
220 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
221 				struct mlx4_en_tx_ring *ring)
222 {
223 	struct mlx4_en_dev *mdev = priv->mdev;
224 
225 	mlx4_qp_modify(mdev->dev, NULL, ring->sp_qp_state,
226 		       MLX4_QP_STATE_RST, NULL, 0, 0, &ring->sp_qp);
227 }
228 
229 static inline bool mlx4_en_is_tx_ring_full(struct mlx4_en_tx_ring *ring)
230 {
231 	return ring->prod - ring->cons > ring->full_size;
232 }
233 
234 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
235 			      struct mlx4_en_tx_ring *ring, int index,
236 			      u8 owner)
237 {
238 	__be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
239 	struct mlx4_en_tx_desc *tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
240 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
241 	void *end = ring->buf + ring->buf_size;
242 	__be32 *ptr = (__be32 *)tx_desc;
243 	int i;
244 
245 	/* Optimize the common case when there are no wraparounds */
246 	if (likely((void *)tx_desc +
247 		   (tx_info->nr_txbb << LOG_TXBB_SIZE) <= end)) {
248 		/* Stamp the freed descriptor */
249 		for (i = 0; i < tx_info->nr_txbb << LOG_TXBB_SIZE;
250 		     i += STAMP_STRIDE) {
251 			*ptr = stamp;
252 			ptr += STAMP_DWORDS;
253 		}
254 	} else {
255 		/* Stamp the freed descriptor */
256 		for (i = 0; i < tx_info->nr_txbb << LOG_TXBB_SIZE;
257 		     i += STAMP_STRIDE) {
258 			*ptr = stamp;
259 			ptr += STAMP_DWORDS;
260 			if ((void *)ptr >= end) {
261 				ptr = ring->buf;
262 				stamp ^= cpu_to_be32(0x80000000);
263 			}
264 		}
265 	}
266 }
267 
268 INDIRECT_CALLABLE_DECLARE(u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
269 						   struct mlx4_en_tx_ring *ring,
270 						   int index, u64 timestamp,
271 						   int napi_mode));
272 
273 u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
274 			 struct mlx4_en_tx_ring *ring,
275 			 int index, u64 timestamp,
276 			 int napi_mode)
277 {
278 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
279 	struct mlx4_en_tx_desc *tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
280 	struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
281 	void *end = ring->buf + ring->buf_size;
282 	struct sk_buff *skb = tx_info->skb;
283 	int nr_maps = tx_info->nr_maps;
284 	int i;
285 
286 	/* We do not touch skb here, so prefetch skb->users location
287 	 * to speedup consume_skb()
288 	 */
289 	prefetchw(&skb->users);
290 
291 	if (unlikely(timestamp)) {
292 		struct skb_shared_hwtstamps hwts;
293 
294 		mlx4_en_fill_hwtstamps(priv->mdev, &hwts, timestamp);
295 		skb_tstamp_tx(skb, &hwts);
296 	}
297 
298 	if (!tx_info->inl) {
299 		if (tx_info->linear)
300 			dma_unmap_single(priv->ddev,
301 					 tx_info->map0_dma,
302 					 tx_info->map0_byte_count,
303 					 DMA_TO_DEVICE);
304 		else
305 			dma_unmap_page(priv->ddev,
306 				       tx_info->map0_dma,
307 				       tx_info->map0_byte_count,
308 				       DMA_TO_DEVICE);
309 		/* Optimize the common case when there are no wraparounds */
310 		if (likely((void *)tx_desc +
311 			   (tx_info->nr_txbb << LOG_TXBB_SIZE) <= end)) {
312 			for (i = 1; i < nr_maps; i++) {
313 				data++;
314 				dma_unmap_page(priv->ddev,
315 					(dma_addr_t)be64_to_cpu(data->addr),
316 					be32_to_cpu(data->byte_count),
317 					DMA_TO_DEVICE);
318 			}
319 		} else {
320 			if ((void *)data >= end)
321 				data = ring->buf + ((void *)data - end);
322 
323 			for (i = 1; i < nr_maps; i++) {
324 				data++;
325 				/* Check for wraparound before unmapping */
326 				if ((void *) data >= end)
327 					data = ring->buf;
328 				dma_unmap_page(priv->ddev,
329 					(dma_addr_t)be64_to_cpu(data->addr),
330 					be32_to_cpu(data->byte_count),
331 					DMA_TO_DEVICE);
332 			}
333 		}
334 	}
335 	napi_consume_skb(skb, napi_mode);
336 
337 	return tx_info->nr_txbb;
338 }
339 
340 INDIRECT_CALLABLE_DECLARE(u32 mlx4_en_recycle_tx_desc(struct mlx4_en_priv *priv,
341 						      struct mlx4_en_tx_ring *ring,
342 						      int index, u64 timestamp,
343 						      int napi_mode));
344 
345 u32 mlx4_en_recycle_tx_desc(struct mlx4_en_priv *priv,
346 			    struct mlx4_en_tx_ring *ring,
347 			    int index, u64 timestamp,
348 			    int napi_mode)
349 {
350 	struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
351 	struct mlx4_en_rx_alloc frame = {
352 		.page = tx_info->page,
353 		.dma = tx_info->map0_dma,
354 	};
355 
356 	if (!napi_mode || !mlx4_en_rx_recycle(ring->recycle_ring, &frame)) {
357 		dma_unmap_page(priv->ddev, tx_info->map0_dma,
358 			       PAGE_SIZE, priv->dma_dir);
359 		put_page(tx_info->page);
360 	}
361 
362 	return tx_info->nr_txbb;
363 }
364 
365 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
366 {
367 	struct mlx4_en_priv *priv = netdev_priv(dev);
368 	int cnt = 0;
369 
370 	/* Skip last polled descriptor */
371 	ring->cons += ring->last_nr_txbb;
372 	en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
373 		 ring->cons, ring->prod);
374 
375 	if ((u32) (ring->prod - ring->cons) > ring->size) {
376 		if (netif_msg_tx_err(priv))
377 			en_warn(priv, "Tx consumer passed producer!\n");
378 		return 0;
379 	}
380 
381 	while (ring->cons != ring->prod) {
382 		ring->last_nr_txbb = ring->free_tx_desc(priv, ring,
383 						ring->cons & ring->size_mask,
384 						0, 0 /* Non-NAPI caller */);
385 		ring->cons += ring->last_nr_txbb;
386 		cnt++;
387 	}
388 
389 	if (ring->tx_queue)
390 		netdev_tx_reset_queue(ring->tx_queue);
391 
392 	if (cnt)
393 		en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
394 
395 	return cnt;
396 }
397 
398 static void mlx4_en_handle_err_cqe(struct mlx4_en_priv *priv, struct mlx4_err_cqe *err_cqe,
399 				   u16 cqe_index, struct mlx4_en_tx_ring *ring)
400 {
401 	struct mlx4_en_dev *mdev = priv->mdev;
402 	struct mlx4_en_tx_info *tx_info;
403 	struct mlx4_en_tx_desc *tx_desc;
404 	u16 wqe_index;
405 	int desc_size;
406 
407 	en_err(priv, "CQE error - cqn 0x%x, ci 0x%x, vendor syndrome: 0x%x syndrome: 0x%x\n",
408 	       ring->sp_cqn, cqe_index, err_cqe->vendor_err_syndrome, err_cqe->syndrome);
409 	print_hex_dump(KERN_WARNING, "", DUMP_PREFIX_OFFSET, 16, 1, err_cqe, sizeof(*err_cqe),
410 		       false);
411 
412 	wqe_index = be16_to_cpu(err_cqe->wqe_index) & ring->size_mask;
413 	tx_info = &ring->tx_info[wqe_index];
414 	desc_size = tx_info->nr_txbb << LOG_TXBB_SIZE;
415 	en_err(priv, "Related WQE - qpn 0x%x, wqe index 0x%x, wqe size 0x%x\n", ring->qpn,
416 	       wqe_index, desc_size);
417 	tx_desc = ring->buf + (wqe_index << LOG_TXBB_SIZE);
418 	print_hex_dump(KERN_WARNING, "", DUMP_PREFIX_OFFSET, 16, 1, tx_desc, desc_size, false);
419 
420 	if (test_and_set_bit(MLX4_EN_STATE_FLAG_RESTARTING, &priv->state))
421 		return;
422 
423 	en_err(priv, "Scheduling port restart\n");
424 	queue_work(mdev->workqueue, &priv->restart_task);
425 }
426 
427 int mlx4_en_process_tx_cq(struct net_device *dev,
428 			  struct mlx4_en_cq *cq, int napi_budget)
429 {
430 	struct mlx4_en_priv *priv = netdev_priv(dev);
431 	struct mlx4_cq *mcq = &cq->mcq;
432 	struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->type][cq->ring];
433 	struct mlx4_cqe *cqe;
434 	u16 index, ring_index, stamp_index;
435 	u32 txbbs_skipped = 0;
436 	u32 txbbs_stamp = 0;
437 	u32 cons_index = mcq->cons_index;
438 	int size = cq->size;
439 	u32 size_mask = ring->size_mask;
440 	struct mlx4_cqe *buf = cq->buf;
441 	u32 packets = 0;
442 	u32 bytes = 0;
443 	int factor = priv->cqe_factor;
444 	int done = 0;
445 	int budget = priv->tx_work_limit;
446 	u32 last_nr_txbb;
447 	u32 ring_cons;
448 
449 	if (unlikely(!priv->port_up))
450 		return 0;
451 
452 	netdev_txq_bql_complete_prefetchw(ring->tx_queue);
453 
454 	index = cons_index & size_mask;
455 	cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
456 	last_nr_txbb = READ_ONCE(ring->last_nr_txbb);
457 	ring_cons = READ_ONCE(ring->cons);
458 	ring_index = ring_cons & size_mask;
459 	stamp_index = ring_index;
460 
461 	/* Process all completed CQEs */
462 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
463 			cons_index & size) && (done < budget)) {
464 		u16 new_index;
465 
466 		/*
467 		 * make sure we read the CQE after we read the
468 		 * ownership bit
469 		 */
470 		dma_rmb();
471 
472 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
473 			     MLX4_CQE_OPCODE_ERROR))
474 			if (!test_and_set_bit(MLX4_EN_TX_RING_STATE_RECOVERING, &ring->state))
475 				mlx4_en_handle_err_cqe(priv, (struct mlx4_err_cqe *)cqe, index,
476 						       ring);
477 
478 		/* Skip over last polled CQE */
479 		new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
480 
481 		do {
482 			u64 timestamp = 0;
483 
484 			txbbs_skipped += last_nr_txbb;
485 			ring_index = (ring_index + last_nr_txbb) & size_mask;
486 
487 			if (unlikely(ring->tx_info[ring_index].ts_requested))
488 				timestamp = mlx4_en_get_cqe_ts(cqe);
489 
490 			/* free next descriptor */
491 			last_nr_txbb = INDIRECT_CALL_2(ring->free_tx_desc,
492 						       mlx4_en_free_tx_desc,
493 						       mlx4_en_recycle_tx_desc,
494 					priv, ring, ring_index,
495 					timestamp, napi_budget);
496 
497 			mlx4_en_stamp_wqe(priv, ring, stamp_index,
498 					  !!((ring_cons + txbbs_stamp) &
499 						ring->size));
500 			stamp_index = ring_index;
501 			txbbs_stamp = txbbs_skipped;
502 			packets++;
503 			bytes += ring->tx_info[ring_index].nr_bytes;
504 		} while ((++done < budget) && (ring_index != new_index));
505 
506 		++cons_index;
507 		index = cons_index & size_mask;
508 		cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
509 	}
510 
511 	/*
512 	 * To prevent CQ overflow we first update CQ consumer and only then
513 	 * the ring consumer.
514 	 */
515 	mcq->cons_index = cons_index;
516 	mlx4_cq_set_ci(mcq);
517 	wmb();
518 
519 	/* we want to dirty this cache line once */
520 	WRITE_ONCE(ring->last_nr_txbb, last_nr_txbb);
521 	WRITE_ONCE(ring->cons, ring_cons + txbbs_skipped);
522 
523 	if (cq->type == TX_XDP)
524 		return done;
525 
526 	netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
527 
528 	/* Wakeup Tx queue if this stopped, and ring is not full.
529 	 */
530 	if (netif_tx_queue_stopped(ring->tx_queue) &&
531 	    !mlx4_en_is_tx_ring_full(ring)) {
532 		netif_tx_wake_queue(ring->tx_queue);
533 		ring->wake_queue++;
534 	}
535 
536 	return done;
537 }
538 
539 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
540 {
541 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
542 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
543 
544 	if (likely(priv->port_up))
545 		napi_schedule_irqoff(&cq->napi);
546 	else
547 		mlx4_en_arm_cq(priv, cq);
548 }
549 
550 /* TX CQ polling - called by NAPI */
551 int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget)
552 {
553 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
554 	struct net_device *dev = cq->dev;
555 	struct mlx4_en_priv *priv = netdev_priv(dev);
556 	int work_done;
557 
558 	work_done = mlx4_en_process_tx_cq(dev, cq, budget);
559 	if (work_done >= budget)
560 		return budget;
561 
562 	if (napi_complete_done(napi, work_done))
563 		mlx4_en_arm_cq(priv, cq);
564 
565 	return 0;
566 }
567 
568 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
569 						      struct mlx4_en_tx_ring *ring,
570 						      u32 index,
571 						      unsigned int desc_size)
572 {
573 	u32 copy = (ring->size - index) << LOG_TXBB_SIZE;
574 	int i;
575 
576 	for (i = desc_size - copy - 4; i >= 0; i -= 4) {
577 		if ((i & (TXBB_SIZE - 1)) == 0)
578 			wmb();
579 
580 		*((u32 *) (ring->buf + i)) =
581 			*((u32 *) (ring->bounce_buf + copy + i));
582 	}
583 
584 	for (i = copy - 4; i >= 4 ; i -= 4) {
585 		if ((i & (TXBB_SIZE - 1)) == 0)
586 			wmb();
587 
588 		*((u32 *)(ring->buf + (index << LOG_TXBB_SIZE) + i)) =
589 			*((u32 *) (ring->bounce_buf + i));
590 	}
591 
592 	/* Return real descriptor location */
593 	return ring->buf + (index << LOG_TXBB_SIZE);
594 }
595 
596 /* Decide if skb can be inlined in tx descriptor to avoid dma mapping
597  *
598  * It seems strange we do not simply use skb_copy_bits().
599  * This would allow to inline all skbs iff skb->len <= inline_thold
600  *
601  * Note that caller already checked skb was not a gso packet
602  */
603 static bool is_inline(int inline_thold, const struct sk_buff *skb,
604 		      const struct skb_shared_info *shinfo,
605 		      void **pfrag)
606 {
607 	void *ptr;
608 
609 	if (skb->len > inline_thold || !inline_thold)
610 		return false;
611 
612 	if (shinfo->nr_frags == 1) {
613 		ptr = skb_frag_address_safe(&shinfo->frags[0]);
614 		if (unlikely(!ptr))
615 			return false;
616 		*pfrag = ptr;
617 		return true;
618 	}
619 	if (shinfo->nr_frags)
620 		return false;
621 	return true;
622 }
623 
624 static int inline_size(const struct sk_buff *skb)
625 {
626 	if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
627 	    <= MLX4_INLINE_ALIGN)
628 		return ALIGN(skb->len + CTRL_SIZE +
629 			     sizeof(struct mlx4_wqe_inline_seg), 16);
630 	else
631 		return ALIGN(skb->len + CTRL_SIZE + 2 *
632 			     sizeof(struct mlx4_wqe_inline_seg), 16);
633 }
634 
635 static int get_real_size(const struct sk_buff *skb,
636 			 const struct skb_shared_info *shinfo,
637 			 struct net_device *dev,
638 			 int *lso_header_size,
639 			 bool *inline_ok,
640 			 void **pfrag,
641 			 int *hopbyhop)
642 {
643 	struct mlx4_en_priv *priv = netdev_priv(dev);
644 	int real_size;
645 
646 	if (shinfo->gso_size) {
647 		*inline_ok = false;
648 		*hopbyhop = 0;
649 		if (skb->encapsulation) {
650 			*lso_header_size = skb_inner_tcp_all_headers(skb);
651 		} else {
652 			/* Detects large IPV6 TCP packets and prepares for removal of
653 			 * HBH header that has been pushed by ip6_xmit(),
654 			 * mainly so that tcpdump can dissect them.
655 			 */
656 			if (ipv6_has_hopopt_jumbo(skb))
657 				*hopbyhop = sizeof(struct hop_jumbo_hdr);
658 			*lso_header_size = skb_tcp_all_headers(skb);
659 		}
660 		real_size = CTRL_SIZE + shinfo->nr_frags * DS_SIZE +
661 			ALIGN(*lso_header_size - *hopbyhop + 4, DS_SIZE);
662 		if (unlikely(*lso_header_size != skb_headlen(skb))) {
663 			/* We add a segment for the skb linear buffer only if
664 			 * it contains data */
665 			if (*lso_header_size < skb_headlen(skb))
666 				real_size += DS_SIZE;
667 			else {
668 				if (netif_msg_tx_err(priv))
669 					en_warn(priv, "Non-linear headers\n");
670 				return 0;
671 			}
672 		}
673 	} else {
674 		*lso_header_size = 0;
675 		*inline_ok = is_inline(priv->prof->inline_thold, skb,
676 				       shinfo, pfrag);
677 
678 		if (*inline_ok)
679 			real_size = inline_size(skb);
680 		else
681 			real_size = CTRL_SIZE +
682 				    (shinfo->nr_frags + 1) * DS_SIZE;
683 	}
684 
685 	return real_size;
686 }
687 
688 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
689 			     const struct sk_buff *skb,
690 			     const struct skb_shared_info *shinfo,
691 			     void *fragptr)
692 {
693 	struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
694 	int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof(*inl);
695 	unsigned int hlen = skb_headlen(skb);
696 
697 	if (skb->len <= spc) {
698 		if (likely(skb->len >= MIN_PKT_LEN)) {
699 			inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
700 		} else {
701 			inl->byte_count = cpu_to_be32(1 << 31 | MIN_PKT_LEN);
702 			memset(((void *)(inl + 1)) + skb->len, 0,
703 			       MIN_PKT_LEN - skb->len);
704 		}
705 		skb_copy_from_linear_data(skb, inl + 1, hlen);
706 		if (shinfo->nr_frags)
707 			memcpy(((void *)(inl + 1)) + hlen, fragptr,
708 			       skb_frag_size(&shinfo->frags[0]));
709 
710 	} else {
711 		inl->byte_count = cpu_to_be32(1 << 31 | spc);
712 		if (hlen <= spc) {
713 			skb_copy_from_linear_data(skb, inl + 1, hlen);
714 			if (hlen < spc) {
715 				memcpy(((void *)(inl + 1)) + hlen,
716 				       fragptr, spc - hlen);
717 				fragptr +=  spc - hlen;
718 			}
719 			inl = (void *) (inl + 1) + spc;
720 			memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
721 		} else {
722 			skb_copy_from_linear_data(skb, inl + 1, spc);
723 			inl = (void *) (inl + 1) + spc;
724 			skb_copy_from_linear_data_offset(skb, spc, inl + 1,
725 							 hlen - spc);
726 			if (shinfo->nr_frags)
727 				memcpy(((void *)(inl + 1)) + hlen - spc,
728 				       fragptr,
729 				       skb_frag_size(&shinfo->frags[0]));
730 		}
731 
732 		dma_wmb();
733 		inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
734 	}
735 }
736 
737 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb,
738 			 struct net_device *sb_dev)
739 {
740 	struct mlx4_en_priv *priv = netdev_priv(dev);
741 	u16 rings_p_up = priv->num_tx_rings_p_up;
742 
743 	if (netdev_get_num_tc(dev))
744 		return netdev_pick_tx(dev, skb, NULL);
745 
746 	return netdev_pick_tx(dev, skb, NULL) % rings_p_up;
747 }
748 
749 static void mlx4_bf_copy(void __iomem *dst, const void *src,
750 			 unsigned int bytecnt)
751 {
752 	__iowrite64_copy(dst, src, bytecnt / 8);
753 }
754 
755 void mlx4_en_xmit_doorbell(struct mlx4_en_tx_ring *ring)
756 {
757 	wmb();
758 	/* Since there is no iowrite*_native() that writes the
759 	 * value as is, without byteswapping - using the one
760 	 * the doesn't do byteswapping in the relevant arch
761 	 * endianness.
762 	 */
763 #if defined(__LITTLE_ENDIAN)
764 	iowrite32(
765 #else
766 	iowrite32be(
767 #endif
768 		  (__force u32)ring->doorbell_qpn, ring->doorbell_address);
769 }
770 
771 static void mlx4_en_tx_write_desc(struct mlx4_en_tx_ring *ring,
772 				  struct mlx4_en_tx_desc *tx_desc,
773 				  union mlx4_wqe_qpn_vlan qpn_vlan,
774 				  int desc_size, int bf_index,
775 				  __be32 op_own, bool bf_ok,
776 				  bool send_doorbell)
777 {
778 	tx_desc->ctrl.qpn_vlan = qpn_vlan;
779 
780 	if (bf_ok) {
781 		op_own |= htonl((bf_index & 0xffff) << 8);
782 		/* Ensure new descriptor hits memory
783 		 * before setting ownership of this descriptor to HW
784 		 */
785 		dma_wmb();
786 		tx_desc->ctrl.owner_opcode = op_own;
787 
788 		wmb();
789 
790 		mlx4_bf_copy(ring->bf.reg + ring->bf.offset, &tx_desc->ctrl,
791 			     desc_size);
792 
793 		wmb();
794 
795 		ring->bf.offset ^= ring->bf.buf_size;
796 	} else {
797 		/* Ensure new descriptor hits memory
798 		 * before setting ownership of this descriptor to HW
799 		 */
800 		dma_wmb();
801 		tx_desc->ctrl.owner_opcode = op_own;
802 		if (send_doorbell)
803 			mlx4_en_xmit_doorbell(ring);
804 		else
805 			ring->xmit_more++;
806 	}
807 }
808 
809 static bool mlx4_en_build_dma_wqe(struct mlx4_en_priv *priv,
810 				  struct skb_shared_info *shinfo,
811 				  struct mlx4_wqe_data_seg *data,
812 				  struct sk_buff *skb,
813 				  int lso_header_size,
814 				  __be32 mr_key,
815 				  struct mlx4_en_tx_info *tx_info)
816 {
817 	struct device *ddev = priv->ddev;
818 	dma_addr_t dma = 0;
819 	u32 byte_count = 0;
820 	int i_frag;
821 
822 	/* Map fragments if any */
823 	for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
824 		const skb_frag_t *frag = &shinfo->frags[i_frag];
825 		byte_count = skb_frag_size(frag);
826 		dma = skb_frag_dma_map(ddev, frag,
827 				       0, byte_count,
828 				       DMA_TO_DEVICE);
829 		if (dma_mapping_error(ddev, dma))
830 			goto tx_drop_unmap;
831 
832 		data->addr = cpu_to_be64(dma);
833 		data->lkey = mr_key;
834 		dma_wmb();
835 		data->byte_count = cpu_to_be32(byte_count);
836 		--data;
837 	}
838 
839 	/* Map linear part if needed */
840 	if (tx_info->linear) {
841 		byte_count = skb_headlen(skb) - lso_header_size;
842 
843 		dma = dma_map_single(ddev, skb->data +
844 				     lso_header_size, byte_count,
845 				     DMA_TO_DEVICE);
846 		if (dma_mapping_error(ddev, dma))
847 			goto tx_drop_unmap;
848 
849 		data->addr = cpu_to_be64(dma);
850 		data->lkey = mr_key;
851 		dma_wmb();
852 		data->byte_count = cpu_to_be32(byte_count);
853 	}
854 	/* tx completion can avoid cache line miss for common cases */
855 	tx_info->map0_dma = dma;
856 	tx_info->map0_byte_count = byte_count;
857 
858 	return true;
859 
860 tx_drop_unmap:
861 	en_err(priv, "DMA mapping error\n");
862 
863 	while (++i_frag < shinfo->nr_frags) {
864 		++data;
865 		dma_unmap_page(ddev, (dma_addr_t)be64_to_cpu(data->addr),
866 			       be32_to_cpu(data->byte_count),
867 			       DMA_TO_DEVICE);
868 	}
869 
870 	return false;
871 }
872 
873 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
874 {
875 	struct skb_shared_info *shinfo = skb_shinfo(skb);
876 	struct mlx4_en_priv *priv = netdev_priv(dev);
877 	union mlx4_wqe_qpn_vlan	qpn_vlan = {};
878 	struct mlx4_en_tx_ring *ring;
879 	struct mlx4_en_tx_desc *tx_desc;
880 	struct mlx4_wqe_data_seg *data;
881 	struct mlx4_en_tx_info *tx_info;
882 	u32 __maybe_unused ring_cons;
883 	int tx_ind;
884 	int nr_txbb;
885 	int desc_size;
886 	int real_size;
887 	u32 index, bf_index;
888 	struct ipv6hdr *h6;
889 	__be32 op_own;
890 	int lso_header_size;
891 	void *fragptr = NULL;
892 	bool bounce = false;
893 	bool send_doorbell;
894 	bool stop_queue;
895 	bool inline_ok;
896 	u8 data_offset;
897 	int hopbyhop;
898 	bool bf_ok;
899 
900 	tx_ind = skb_get_queue_mapping(skb);
901 	ring = priv->tx_ring[TX][tx_ind];
902 
903 	if (unlikely(!priv->port_up))
904 		goto tx_drop;
905 
906 	real_size = get_real_size(skb, shinfo, dev, &lso_header_size,
907 				  &inline_ok, &fragptr, &hopbyhop);
908 	if (unlikely(!real_size))
909 		goto tx_drop_count;
910 
911 	/* Align descriptor to TXBB size */
912 	desc_size = ALIGN(real_size, TXBB_SIZE);
913 	nr_txbb = desc_size >> LOG_TXBB_SIZE;
914 
915 	bf_ok = ring->bf_enabled;
916 	if (skb_vlan_tag_present(skb)) {
917 		u16 vlan_proto;
918 
919 		qpn_vlan.vlan_tag = cpu_to_be16(skb_vlan_tag_get(skb));
920 		vlan_proto = be16_to_cpu(skb->vlan_proto);
921 		if (vlan_proto == ETH_P_8021AD)
922 			qpn_vlan.ins_vlan = MLX4_WQE_CTRL_INS_SVLAN;
923 		else if (vlan_proto == ETH_P_8021Q)
924 			qpn_vlan.ins_vlan = MLX4_WQE_CTRL_INS_CVLAN;
925 		else
926 			qpn_vlan.ins_vlan = 0;
927 		bf_ok = false;
928 	}
929 
930 	netdev_txq_bql_enqueue_prefetchw(ring->tx_queue);
931 
932 	/* Packet is good - grab an index and transmit it */
933 	index = ring->prod & ring->size_mask;
934 	bf_index = ring->prod;
935 
936 	/* See if we have enough space for whole descriptor TXBB for setting
937 	 * SW ownership on next descriptor; if not, use a bounce buffer. */
938 	if (likely(index + nr_txbb <= ring->size))
939 		tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
940 	else {
941 		if (unlikely(nr_txbb > MLX4_MAX_DESC_TXBBS)) {
942 			if (netif_msg_tx_err(priv))
943 				en_warn(priv, "Oversized header or SG list\n");
944 			goto tx_drop_count;
945 		}
946 		tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
947 		bounce = true;
948 		bf_ok = false;
949 	}
950 
951 	/* Save skb in tx_info ring */
952 	tx_info = &ring->tx_info[index];
953 	tx_info->skb = skb;
954 	tx_info->nr_txbb = nr_txbb;
955 
956 	if (!lso_header_size) {
957 		data = &tx_desc->data;
958 		data_offset = offsetof(struct mlx4_en_tx_desc, data);
959 	} else {
960 		int lso_align = ALIGN(lso_header_size - hopbyhop + 4, DS_SIZE);
961 
962 		data = (void *)&tx_desc->lso + lso_align;
963 		data_offset = offsetof(struct mlx4_en_tx_desc, lso) + lso_align;
964 	}
965 
966 	/* valid only for none inline segments */
967 	tx_info->data_offset = data_offset;
968 
969 	tx_info->inl = inline_ok;
970 
971 	tx_info->linear = lso_header_size < skb_headlen(skb) && !inline_ok;
972 
973 	tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
974 	data += tx_info->nr_maps - 1;
975 
976 	if (!tx_info->inl)
977 		if (!mlx4_en_build_dma_wqe(priv, shinfo, data, skb,
978 					   lso_header_size, ring->mr_key,
979 					   tx_info))
980 			goto tx_drop_count;
981 
982 	/*
983 	 * For timestamping add flag to skb_shinfo and
984 	 * set flag for further reference
985 	 */
986 	tx_info->ts_requested = 0;
987 	if (unlikely(ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
988 		     shinfo->tx_flags & SKBTX_HW_TSTAMP)) {
989 		shinfo->tx_flags |= SKBTX_IN_PROGRESS;
990 		tx_info->ts_requested = 1;
991 	}
992 
993 	/* Prepare ctrl segement apart opcode+ownership, which depends on
994 	 * whether LSO is used */
995 	tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
996 	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
997 		if (!skb->encapsulation)
998 			tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
999 								 MLX4_WQE_CTRL_TCP_UDP_CSUM);
1000 		else
1001 			tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM);
1002 		ring->tx_csum++;
1003 	}
1004 
1005 	if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
1006 		struct ethhdr *ethh;
1007 
1008 		/* Copy dst mac address to wqe. This allows loopback in eSwitch,
1009 		 * so that VFs and PF can communicate with each other
1010 		 */
1011 		ethh = (struct ethhdr *)skb->data;
1012 		tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
1013 		tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
1014 	}
1015 
1016 	/* Handle LSO (TSO) packets */
1017 	if (lso_header_size) {
1018 		int i;
1019 
1020 		/* Mark opcode as LSO */
1021 		op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
1022 			((ring->prod & ring->size) ?
1023 				cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
1024 
1025 		lso_header_size -= hopbyhop;
1026 		/* Fill in the LSO prefix */
1027 		tx_desc->lso.mss_hdr_size = cpu_to_be32(
1028 			shinfo->gso_size << 16 | lso_header_size);
1029 
1030 
1031 		if (unlikely(hopbyhop)) {
1032 			/* remove the HBH header.
1033 			 * Layout: [Ethernet header][IPv6 header][HBH][TCP header]
1034 			 */
1035 			memcpy(tx_desc->lso.header, skb->data, ETH_HLEN + sizeof(*h6));
1036 			h6 = (struct ipv6hdr *)((char *)tx_desc->lso.header + ETH_HLEN);
1037 			h6->nexthdr = IPPROTO_TCP;
1038 			/* Copy the TCP header after the IPv6 one */
1039 			memcpy(h6 + 1,
1040 			       skb->data + ETH_HLEN + sizeof(*h6) +
1041 					sizeof(struct hop_jumbo_hdr),
1042 			       tcp_hdrlen(skb));
1043 			/* Leave ipv6 payload_len set to 0, as LSO v2 specs request. */
1044 		} else {
1045 			/* Copy headers;
1046 			 * note that we already verified that it is linear
1047 			 */
1048 			memcpy(tx_desc->lso.header, skb->data, lso_header_size);
1049 		}
1050 		ring->tso_packets++;
1051 
1052 		i = shinfo->gso_segs;
1053 		tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
1054 		ring->packets += i;
1055 	} else {
1056 		/* Normal (Non LSO) packet */
1057 		op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
1058 			((ring->prod & ring->size) ?
1059 			 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
1060 		tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
1061 		ring->packets++;
1062 	}
1063 	ring->bytes += tx_info->nr_bytes;
1064 
1065 	if (tx_info->inl)
1066 		build_inline_wqe(tx_desc, skb, shinfo, fragptr);
1067 
1068 	if (skb->encapsulation) {
1069 		union {
1070 			struct iphdr *v4;
1071 			struct ipv6hdr *v6;
1072 			unsigned char *hdr;
1073 		} ip;
1074 		u8 proto;
1075 
1076 		ip.hdr = skb_inner_network_header(skb);
1077 		proto = (ip.v4->version == 4) ? ip.v4->protocol :
1078 						ip.v6->nexthdr;
1079 
1080 		if (proto == IPPROTO_TCP || proto == IPPROTO_UDP)
1081 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP | MLX4_WQE_CTRL_ILP);
1082 		else
1083 			op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP);
1084 	}
1085 
1086 	ring->prod += nr_txbb;
1087 
1088 	/* If we used a bounce buffer then copy descriptor back into place */
1089 	if (unlikely(bounce))
1090 		tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
1091 
1092 	skb_tx_timestamp(skb);
1093 
1094 	/* Check available TXBBs And 2K spare for prefetch */
1095 	stop_queue = mlx4_en_is_tx_ring_full(ring);
1096 	if (unlikely(stop_queue)) {
1097 		netif_tx_stop_queue(ring->tx_queue);
1098 		ring->queue_stopped++;
1099 	}
1100 
1101 	send_doorbell = __netdev_tx_sent_queue(ring->tx_queue,
1102 					       tx_info->nr_bytes,
1103 					       netdev_xmit_more());
1104 
1105 	real_size = (real_size / 16) & 0x3f;
1106 
1107 	bf_ok &= desc_size <= MAX_BF && send_doorbell;
1108 
1109 	if (bf_ok)
1110 		qpn_vlan.bf_qpn = ring->doorbell_qpn | cpu_to_be32(real_size);
1111 	else
1112 		qpn_vlan.fence_size = real_size;
1113 
1114 	mlx4_en_tx_write_desc(ring, tx_desc, qpn_vlan, desc_size, bf_index,
1115 			      op_own, bf_ok, send_doorbell);
1116 
1117 	if (unlikely(stop_queue)) {
1118 		/* If queue was emptied after the if (stop_queue) , and before
1119 		 * the netif_tx_stop_queue() - need to wake the queue,
1120 		 * or else it will remain stopped forever.
1121 		 * Need a memory barrier to make sure ring->cons was not
1122 		 * updated before queue was stopped.
1123 		 */
1124 		smp_rmb();
1125 
1126 		if (unlikely(!mlx4_en_is_tx_ring_full(ring))) {
1127 			netif_tx_wake_queue(ring->tx_queue);
1128 			ring->wake_queue++;
1129 		}
1130 	}
1131 	return NETDEV_TX_OK;
1132 
1133 tx_drop_count:
1134 	ring->tx_dropped++;
1135 tx_drop:
1136 	dev_kfree_skb_any(skb);
1137 	return NETDEV_TX_OK;
1138 }
1139 
1140 #define MLX4_EN_XDP_TX_NRTXBB  1
1141 #define MLX4_EN_XDP_TX_REAL_SZ (((CTRL_SIZE + MLX4_EN_XDP_TX_NRTXBB * DS_SIZE) \
1142 				 / 16) & 0x3f)
1143 
1144 void mlx4_en_init_tx_xdp_ring_descs(struct mlx4_en_priv *priv,
1145 				    struct mlx4_en_tx_ring *ring)
1146 {
1147 	int i;
1148 
1149 	for (i = 0; i < ring->size; i++) {
1150 		struct mlx4_en_tx_info *tx_info = &ring->tx_info[i];
1151 		struct mlx4_en_tx_desc *tx_desc = ring->buf +
1152 			(i << LOG_TXBB_SIZE);
1153 
1154 		tx_info->map0_byte_count = PAGE_SIZE;
1155 		tx_info->nr_txbb = MLX4_EN_XDP_TX_NRTXBB;
1156 		tx_info->data_offset = offsetof(struct mlx4_en_tx_desc, data);
1157 		tx_info->ts_requested = 0;
1158 		tx_info->nr_maps = 1;
1159 		tx_info->linear = 1;
1160 		tx_info->inl = 0;
1161 
1162 		tx_desc->data.lkey = ring->mr_key;
1163 		tx_desc->ctrl.qpn_vlan.fence_size = MLX4_EN_XDP_TX_REAL_SZ;
1164 		tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
1165 	}
1166 }
1167 
1168 netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_ring *rx_ring,
1169 			       struct mlx4_en_rx_alloc *frame,
1170 			       struct mlx4_en_priv *priv, unsigned int length,
1171 			       int tx_ind, bool *doorbell_pending)
1172 {
1173 	struct mlx4_en_tx_desc *tx_desc;
1174 	struct mlx4_en_tx_info *tx_info;
1175 	struct mlx4_wqe_data_seg *data;
1176 	struct mlx4_en_tx_ring *ring;
1177 	dma_addr_t dma;
1178 	__be32 op_own;
1179 	int index;
1180 
1181 	if (unlikely(!priv->port_up))
1182 		goto tx_drop;
1183 
1184 	ring = priv->tx_ring[TX_XDP][tx_ind];
1185 
1186 	if (unlikely(mlx4_en_is_tx_ring_full(ring)))
1187 		goto tx_drop_count;
1188 
1189 	index = ring->prod & ring->size_mask;
1190 	tx_info = &ring->tx_info[index];
1191 
1192 	tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
1193 	data = &tx_desc->data;
1194 
1195 	dma = frame->dma;
1196 
1197 	tx_info->page = frame->page;
1198 	frame->page = NULL;
1199 	tx_info->map0_dma = dma;
1200 	tx_info->nr_bytes = max_t(unsigned int, length, ETH_ZLEN);
1201 
1202 	dma_sync_single_range_for_device(priv->ddev, dma, frame->page_offset,
1203 					 length, DMA_TO_DEVICE);
1204 
1205 	data->addr = cpu_to_be64(dma + frame->page_offset);
1206 	dma_wmb();
1207 	data->byte_count = cpu_to_be32(length);
1208 
1209 	/* tx completion can avoid cache line miss for common cases */
1210 
1211 	op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
1212 		((ring->prod & ring->size) ?
1213 		 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
1214 
1215 	rx_ring->xdp_tx++;
1216 
1217 	ring->prod += MLX4_EN_XDP_TX_NRTXBB;
1218 
1219 	/* Ensure new descriptor hits memory
1220 	 * before setting ownership of this descriptor to HW
1221 	 */
1222 	dma_wmb();
1223 	tx_desc->ctrl.owner_opcode = op_own;
1224 	ring->xmit_more++;
1225 
1226 	*doorbell_pending = true;
1227 
1228 	return NETDEV_TX_OK;
1229 
1230 tx_drop_count:
1231 	rx_ring->xdp_tx_full++;
1232 	*doorbell_pending = true;
1233 tx_drop:
1234 	return NETDEV_TX_BUSY;
1235 }
1236