2 * Copyright (c) 2007 Mellanox Technologies. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
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>
44 #include <linux/ipv6.h>
45 #include <linux/indirect_call_wrapper.h>
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)
54 struct mlx4_en_dev *mdev = priv->mdev;
55 struct mlx4_en_tx_ring *ring;
59 ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
61 en_err(priv, "Failed allocating TX ring\n");
66 ring->size_mask = size - 1;
67 ring->sp_stride = stride;
68 ring->full_size = ring->size - HEADROOM - MLX4_MAX_DESC_TXBBS;
70 tmp = size * sizeof(struct mlx4_en_tx_info);
71 ring->tx_info = kvmalloc_node(tmp, GFP_KERNEL, node);
77 en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
80 ring->bounce_buf = kmalloc_node(MLX4_TX_BOUNCE_BUFFER_SIZE,
82 if (!ring->bounce_buf) {
83 ring->bounce_buf = kmalloc(MLX4_TX_BOUNCE_BUFFER_SIZE,
85 if (!ring->bounce_buf) {
90 ring->buf_size = ALIGN(size * ring->sp_stride, MLX4_EN_PAGE_SIZE);
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);
97 en_err(priv, "Failed allocating hwq resources\n");
101 ring->buf = ring->sp_wqres.buf.direct.buf;
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);
107 err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &ring->qpn,
108 MLX4_RESERVE_ETH_BF_QP,
109 MLX4_RES_USAGE_DRIVER);
111 en_err(priv, "failed reserving qp for TX ring\n");
115 err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->sp_qp);
117 en_err(priv, "Failed allocating qp %d\n", ring->qpn);
120 ring->sp_qp.event = mlx4_en_sqp_event;
122 err = mlx4_bf_alloc(mdev->dev, &ring->bf, node);
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;
131 ring->bf_alloced = true;
132 ring->bf_enabled = !!(priv->pflags &
133 MLX4_EN_PRIV_FLAGS_BLUEFLAME);
135 ring->doorbell_address = ring->bf.uar->map + MLX4_SEND_DOORBELL;
137 ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type;
138 ring->queue_index = queue_index;
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);
149 mlx4_qp_release_range(mdev->dev, ring->qpn, 1);
151 mlx4_free_hwq_res(mdev->dev, &ring->sp_wqres, ring->buf_size);
153 kfree(ring->bounce_buf);
154 ring->bounce_buf = NULL;
156 kvfree(ring->tx_info);
157 ring->tx_info = NULL;
164 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
165 struct mlx4_en_tx_ring **pring)
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);
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;
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)
189 struct mlx4_en_dev *mdev = priv->mdev;
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;
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);
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));
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,
220 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
221 struct mlx4_en_tx_ring *ring)
223 struct mlx4_en_dev *mdev = priv->mdev;
225 mlx4_qp_modify(mdev->dev, NULL, ring->sp_qp_state,
226 MLX4_QP_STATE_RST, NULL, 0, 0, &ring->sp_qp);
229 static inline bool mlx4_en_is_tx_ring_full(struct mlx4_en_tx_ring *ring)
231 u32 used = READ_ONCE(ring->prod) - READ_ONCE(ring->cons);
233 return used > ring->full_size;
236 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
237 struct mlx4_en_tx_ring *ring, int index,
240 __be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
241 struct mlx4_en_tx_desc *tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
242 struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
243 void *end = ring->buf + ring->buf_size;
244 __be32 *ptr = (__be32 *)tx_desc;
247 /* Optimize the common case when there are no wraparounds */
248 if (likely((void *)tx_desc +
249 (tx_info->nr_txbb << LOG_TXBB_SIZE) <= end)) {
250 /* Stamp the freed descriptor */
251 for (i = 0; i < tx_info->nr_txbb << LOG_TXBB_SIZE;
257 /* Stamp the freed descriptor */
258 for (i = 0; i < tx_info->nr_txbb << LOG_TXBB_SIZE;
262 if ((void *)ptr >= end) {
264 stamp ^= cpu_to_be32(0x80000000);
270 INDIRECT_CALLABLE_DECLARE(u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
271 struct mlx4_en_tx_ring *ring,
272 int index, u64 timestamp,
275 u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
276 struct mlx4_en_tx_ring *ring,
277 int index, u64 timestamp,
280 struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
281 struct mlx4_en_tx_desc *tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
282 struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
283 void *end = ring->buf + ring->buf_size;
284 struct sk_buff *skb = tx_info->skb;
285 int nr_maps = tx_info->nr_maps;
288 /* We do not touch skb here, so prefetch skb->users location
289 * to speedup consume_skb()
291 prefetchw(&skb->users);
293 if (unlikely(timestamp)) {
294 struct skb_shared_hwtstamps hwts;
296 mlx4_en_fill_hwtstamps(priv->mdev, &hwts, timestamp);
297 skb_tstamp_tx(skb, &hwts);
302 dma_unmap_single(priv->ddev,
304 tx_info->map0_byte_count,
307 dma_unmap_page(priv->ddev,
309 tx_info->map0_byte_count,
311 /* Optimize the common case when there are no wraparounds */
312 if (likely((void *)tx_desc +
313 (tx_info->nr_txbb << LOG_TXBB_SIZE) <= end)) {
314 for (i = 1; i < nr_maps; i++) {
316 dma_unmap_page(priv->ddev,
317 (dma_addr_t)be64_to_cpu(data->addr),
318 be32_to_cpu(data->byte_count),
322 if ((void *)data >= end)
323 data = ring->buf + ((void *)data - end);
325 for (i = 1; i < nr_maps; i++) {
327 /* Check for wraparound before unmapping */
328 if ((void *) data >= end)
330 dma_unmap_page(priv->ddev,
331 (dma_addr_t)be64_to_cpu(data->addr),
332 be32_to_cpu(data->byte_count),
337 napi_consume_skb(skb, napi_mode);
339 return tx_info->nr_txbb;
342 INDIRECT_CALLABLE_DECLARE(u32 mlx4_en_recycle_tx_desc(struct mlx4_en_priv *priv,
343 struct mlx4_en_tx_ring *ring,
344 int index, u64 timestamp,
347 u32 mlx4_en_recycle_tx_desc(struct mlx4_en_priv *priv,
348 struct mlx4_en_tx_ring *ring,
349 int index, u64 timestamp,
352 struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
353 struct mlx4_en_rx_alloc frame = {
354 .page = tx_info->page,
355 .dma = tx_info->map0_dma,
358 if (!napi_mode || !mlx4_en_rx_recycle(ring->recycle_ring, &frame)) {
359 dma_unmap_page(priv->ddev, tx_info->map0_dma,
360 PAGE_SIZE, priv->dma_dir);
361 put_page(tx_info->page);
364 return tx_info->nr_txbb;
367 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
369 struct mlx4_en_priv *priv = netdev_priv(dev);
372 /* Skip last polled descriptor */
373 ring->cons += ring->last_nr_txbb;
374 en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
375 ring->cons, ring->prod);
377 if ((u32) (ring->prod - ring->cons) > ring->size) {
378 if (netif_msg_tx_err(priv))
379 en_warn(priv, "Tx consumer passed producer!\n");
383 while (ring->cons != ring->prod) {
384 ring->last_nr_txbb = ring->free_tx_desc(priv, ring,
385 ring->cons & ring->size_mask,
386 0, 0 /* Non-NAPI caller */);
387 ring->cons += ring->last_nr_txbb;
392 netdev_tx_reset_queue(ring->tx_queue);
395 en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
400 static void mlx4_en_handle_err_cqe(struct mlx4_en_priv *priv, struct mlx4_err_cqe *err_cqe,
401 u16 cqe_index, struct mlx4_en_tx_ring *ring)
403 struct mlx4_en_dev *mdev = priv->mdev;
404 struct mlx4_en_tx_info *tx_info;
405 struct mlx4_en_tx_desc *tx_desc;
409 en_err(priv, "CQE error - cqn 0x%x, ci 0x%x, vendor syndrome: 0x%x syndrome: 0x%x\n",
410 ring->sp_cqn, cqe_index, err_cqe->vendor_err_syndrome, err_cqe->syndrome);
411 print_hex_dump(KERN_WARNING, "", DUMP_PREFIX_OFFSET, 16, 1, err_cqe, sizeof(*err_cqe),
414 wqe_index = be16_to_cpu(err_cqe->wqe_index) & ring->size_mask;
415 tx_info = &ring->tx_info[wqe_index];
416 desc_size = tx_info->nr_txbb << LOG_TXBB_SIZE;
417 en_err(priv, "Related WQE - qpn 0x%x, wqe index 0x%x, wqe size 0x%x\n", ring->qpn,
418 wqe_index, desc_size);
419 tx_desc = ring->buf + (wqe_index << LOG_TXBB_SIZE);
420 print_hex_dump(KERN_WARNING, "", DUMP_PREFIX_OFFSET, 16, 1, tx_desc, desc_size, false);
422 if (test_and_set_bit(MLX4_EN_STATE_FLAG_RESTARTING, &priv->state))
425 en_err(priv, "Scheduling port restart\n");
426 queue_work(mdev->workqueue, &priv->restart_task);
429 int mlx4_en_process_tx_cq(struct net_device *dev,
430 struct mlx4_en_cq *cq, int napi_budget)
432 struct mlx4_en_priv *priv = netdev_priv(dev);
433 struct mlx4_cq *mcq = &cq->mcq;
434 struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->type][cq->ring];
435 struct mlx4_cqe *cqe;
436 u16 index, ring_index, stamp_index;
437 u32 txbbs_skipped = 0;
439 u32 cons_index = mcq->cons_index;
441 u32 size_mask = ring->size_mask;
442 struct mlx4_cqe *buf = cq->buf;
445 int factor = priv->cqe_factor;
447 int budget = priv->tx_work_limit;
451 if (unlikely(!priv->port_up))
454 netdev_txq_bql_complete_prefetchw(ring->tx_queue);
456 index = cons_index & size_mask;
457 cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
458 last_nr_txbb = READ_ONCE(ring->last_nr_txbb);
459 ring_cons = READ_ONCE(ring->cons);
460 ring_index = ring_cons & size_mask;
461 stamp_index = ring_index;
463 /* Process all completed CQEs */
464 while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
465 cons_index & size) && (done < budget)) {
469 * make sure we read the CQE after we read the
474 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
475 MLX4_CQE_OPCODE_ERROR))
476 if (!test_and_set_bit(MLX4_EN_TX_RING_STATE_RECOVERING, &ring->state))
477 mlx4_en_handle_err_cqe(priv, (struct mlx4_err_cqe *)cqe, index,
480 /* Skip over last polled CQE */
481 new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
486 txbbs_skipped += last_nr_txbb;
487 ring_index = (ring_index + last_nr_txbb) & size_mask;
489 if (unlikely(ring->tx_info[ring_index].ts_requested))
490 timestamp = mlx4_en_get_cqe_ts(cqe);
492 /* free next descriptor */
493 last_nr_txbb = INDIRECT_CALL_2(ring->free_tx_desc,
494 mlx4_en_free_tx_desc,
495 mlx4_en_recycle_tx_desc,
496 priv, ring, ring_index,
497 timestamp, napi_budget);
499 mlx4_en_stamp_wqe(priv, ring, stamp_index,
500 !!((ring_cons + txbbs_stamp) &
502 stamp_index = ring_index;
503 txbbs_stamp = txbbs_skipped;
505 bytes += ring->tx_info[ring_index].nr_bytes;
506 } while ((++done < budget) && (ring_index != new_index));
509 index = cons_index & size_mask;
510 cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
514 * To prevent CQ overflow we first update CQ consumer and only then
517 mcq->cons_index = cons_index;
521 /* we want to dirty this cache line once */
522 WRITE_ONCE(ring->last_nr_txbb, last_nr_txbb);
523 WRITE_ONCE(ring->cons, ring_cons + txbbs_skipped);
525 if (cq->type == TX_XDP)
528 netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
530 /* Wakeup Tx queue if this stopped, and ring is not full.
532 if (netif_tx_queue_stopped(ring->tx_queue) &&
533 !mlx4_en_is_tx_ring_full(ring)) {
534 netif_tx_wake_queue(ring->tx_queue);
541 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
543 struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
544 struct mlx4_en_priv *priv = netdev_priv(cq->dev);
546 if (likely(priv->port_up))
547 napi_schedule_irqoff(&cq->napi);
549 mlx4_en_arm_cq(priv, cq);
552 /* TX CQ polling - called by NAPI */
553 int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget)
555 struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
556 struct net_device *dev = cq->dev;
557 struct mlx4_en_priv *priv = netdev_priv(dev);
560 work_done = mlx4_en_process_tx_cq(dev, cq, budget);
561 if (work_done >= budget)
564 if (napi_complete_done(napi, work_done))
565 mlx4_en_arm_cq(priv, cq);
570 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
571 struct mlx4_en_tx_ring *ring,
573 unsigned int desc_size)
575 u32 copy = (ring->size - index) << LOG_TXBB_SIZE;
578 for (i = desc_size - copy - 4; i >= 0; i -= 4) {
579 if ((i & (TXBB_SIZE - 1)) == 0)
582 *((u32 *) (ring->buf + i)) =
583 *((u32 *) (ring->bounce_buf + copy + i));
586 for (i = copy - 4; i >= 4 ; i -= 4) {
587 if ((i & (TXBB_SIZE - 1)) == 0)
590 *((u32 *)(ring->buf + (index << LOG_TXBB_SIZE) + i)) =
591 *((u32 *) (ring->bounce_buf + i));
594 /* Return real descriptor location */
595 return ring->buf + (index << LOG_TXBB_SIZE);
598 /* Decide if skb can be inlined in tx descriptor to avoid dma mapping
600 * It seems strange we do not simply use skb_copy_bits().
601 * This would allow to inline all skbs iff skb->len <= inline_thold
603 * Note that caller already checked skb was not a gso packet
605 static bool is_inline(int inline_thold, const struct sk_buff *skb,
606 const struct skb_shared_info *shinfo,
611 if (skb->len > inline_thold || !inline_thold)
614 if (shinfo->nr_frags == 1) {
615 ptr = skb_frag_address_safe(&shinfo->frags[0]);
621 if (shinfo->nr_frags)
626 static int inline_size(const struct sk_buff *skb)
628 if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
629 <= MLX4_INLINE_ALIGN)
630 return ALIGN(skb->len + CTRL_SIZE +
631 sizeof(struct mlx4_wqe_inline_seg), 16);
633 return ALIGN(skb->len + CTRL_SIZE + 2 *
634 sizeof(struct mlx4_wqe_inline_seg), 16);
637 static int get_real_size(const struct sk_buff *skb,
638 const struct skb_shared_info *shinfo,
639 struct net_device *dev,
640 int *lso_header_size,
645 struct mlx4_en_priv *priv = netdev_priv(dev);
648 if (shinfo->gso_size) {
651 if (skb->encapsulation) {
652 *lso_header_size = skb_inner_tcp_all_headers(skb);
654 /* Detects large IPV6 TCP packets and prepares for removal of
655 * HBH header that has been pushed by ip6_xmit(),
656 * mainly so that tcpdump can dissect them.
658 if (ipv6_has_hopopt_jumbo(skb))
659 *hopbyhop = sizeof(struct hop_jumbo_hdr);
660 *lso_header_size = skb_tcp_all_headers(skb);
662 real_size = CTRL_SIZE + shinfo->nr_frags * DS_SIZE +
663 ALIGN(*lso_header_size - *hopbyhop + 4, DS_SIZE);
664 if (unlikely(*lso_header_size != skb_headlen(skb))) {
665 /* We add a segment for the skb linear buffer only if
666 * it contains data */
667 if (*lso_header_size < skb_headlen(skb))
668 real_size += DS_SIZE;
670 if (netif_msg_tx_err(priv))
671 en_warn(priv, "Non-linear headers\n");
676 *lso_header_size = 0;
677 *inline_ok = is_inline(priv->prof->inline_thold, skb,
681 real_size = inline_size(skb);
683 real_size = CTRL_SIZE +
684 (shinfo->nr_frags + 1) * DS_SIZE;
690 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
691 const struct sk_buff *skb,
692 const struct skb_shared_info *shinfo,
695 struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
696 int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof(*inl);
697 unsigned int hlen = skb_headlen(skb);
699 if (skb->len <= spc) {
700 if (likely(skb->len >= MIN_PKT_LEN)) {
701 inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
703 inl->byte_count = cpu_to_be32(1 << 31 | MIN_PKT_LEN);
704 memset(inl->data + skb->len, 0,
705 MIN_PKT_LEN - skb->len);
707 skb_copy_from_linear_data(skb, inl->data, hlen);
708 if (shinfo->nr_frags)
709 memcpy(inl->data + hlen, fragptr,
710 skb_frag_size(&shinfo->frags[0]));
713 inl->byte_count = cpu_to_be32(1 << 31 | spc);
715 skb_copy_from_linear_data(skb, inl->data, hlen);
717 memcpy(inl->data + hlen,
718 fragptr, spc - hlen);
719 fragptr += spc - hlen;
721 inl = (void *)inl->data + spc;
722 memcpy(inl->data, fragptr, skb->len - spc);
724 skb_copy_from_linear_data(skb, inl->data, spc);
725 inl = (void *)inl->data + spc;
726 skb_copy_from_linear_data_offset(skb, spc, inl->data,
728 if (shinfo->nr_frags)
729 memcpy(inl->data + hlen - spc,
731 skb_frag_size(&shinfo->frags[0]));
735 inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
739 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb,
740 struct net_device *sb_dev)
742 struct mlx4_en_priv *priv = netdev_priv(dev);
743 u16 rings_p_up = priv->num_tx_rings_p_up;
745 if (netdev_get_num_tc(dev))
746 return netdev_pick_tx(dev, skb, NULL);
748 return netdev_pick_tx(dev, skb, NULL) % rings_p_up;
751 static void mlx4_bf_copy(void __iomem *dst, const void *src,
752 unsigned int bytecnt)
754 __iowrite64_copy(dst, src, bytecnt / 8);
757 void mlx4_en_xmit_doorbell(struct mlx4_en_tx_ring *ring)
760 /* Since there is no iowrite*_native() that writes the
761 * value as is, without byteswapping - using the one
762 * the doesn't do byteswapping in the relevant arch
765 #if defined(__LITTLE_ENDIAN)
770 (__force u32)ring->doorbell_qpn, ring->doorbell_address);
773 static void mlx4_en_tx_write_desc(struct mlx4_en_tx_ring *ring,
774 struct mlx4_en_tx_desc *tx_desc,
775 union mlx4_wqe_qpn_vlan qpn_vlan,
776 int desc_size, int bf_index,
777 __be32 op_own, bool bf_ok,
780 tx_desc->ctrl.qpn_vlan = qpn_vlan;
783 op_own |= htonl((bf_index & 0xffff) << 8);
784 /* Ensure new descriptor hits memory
785 * before setting ownership of this descriptor to HW
788 tx_desc->ctrl.owner_opcode = op_own;
792 mlx4_bf_copy(ring->bf.reg + ring->bf.offset, &tx_desc->ctrl,
797 ring->bf.offset ^= ring->bf.buf_size;
799 /* Ensure new descriptor hits memory
800 * before setting ownership of this descriptor to HW
803 tx_desc->ctrl.owner_opcode = op_own;
805 mlx4_en_xmit_doorbell(ring);
811 static bool mlx4_en_build_dma_wqe(struct mlx4_en_priv *priv,
812 struct skb_shared_info *shinfo,
813 struct mlx4_wqe_data_seg *data,
817 struct mlx4_en_tx_info *tx_info)
819 struct device *ddev = priv->ddev;
824 /* Map fragments if any */
825 for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
826 const skb_frag_t *frag = &shinfo->frags[i_frag];
827 byte_count = skb_frag_size(frag);
828 dma = skb_frag_dma_map(ddev, frag,
831 if (dma_mapping_error(ddev, dma))
834 data->addr = cpu_to_be64(dma);
837 data->byte_count = cpu_to_be32(byte_count);
841 /* Map linear part if needed */
842 if (tx_info->linear) {
843 byte_count = skb_headlen(skb) - lso_header_size;
845 dma = dma_map_single(ddev, skb->data +
846 lso_header_size, byte_count,
848 if (dma_mapping_error(ddev, dma))
851 data->addr = cpu_to_be64(dma);
854 data->byte_count = cpu_to_be32(byte_count);
856 /* tx completion can avoid cache line miss for common cases */
857 tx_info->map0_dma = dma;
858 tx_info->map0_byte_count = byte_count;
863 en_err(priv, "DMA mapping error\n");
865 while (++i_frag < shinfo->nr_frags) {
867 dma_unmap_page(ddev, (dma_addr_t)be64_to_cpu(data->addr),
868 be32_to_cpu(data->byte_count),
875 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
877 struct skb_shared_info *shinfo = skb_shinfo(skb);
878 struct mlx4_en_priv *priv = netdev_priv(dev);
879 union mlx4_wqe_qpn_vlan qpn_vlan = {};
880 struct mlx4_en_tx_ring *ring;
881 struct mlx4_en_tx_desc *tx_desc;
882 struct mlx4_wqe_data_seg *data;
883 struct mlx4_en_tx_info *tx_info;
884 u32 __maybe_unused ring_cons;
893 void *fragptr = NULL;
902 tx_ind = skb_get_queue_mapping(skb);
903 ring = priv->tx_ring[TX][tx_ind];
905 if (unlikely(!priv->port_up))
908 real_size = get_real_size(skb, shinfo, dev, &lso_header_size,
909 &inline_ok, &fragptr, &hopbyhop);
910 if (unlikely(!real_size))
913 /* Align descriptor to TXBB size */
914 desc_size = ALIGN(real_size, TXBB_SIZE);
915 nr_txbb = desc_size >> LOG_TXBB_SIZE;
917 bf_ok = ring->bf_enabled;
918 if (skb_vlan_tag_present(skb)) {
921 qpn_vlan.vlan_tag = cpu_to_be16(skb_vlan_tag_get(skb));
922 vlan_proto = be16_to_cpu(skb->vlan_proto);
923 if (vlan_proto == ETH_P_8021AD)
924 qpn_vlan.ins_vlan = MLX4_WQE_CTRL_INS_SVLAN;
925 else if (vlan_proto == ETH_P_8021Q)
926 qpn_vlan.ins_vlan = MLX4_WQE_CTRL_INS_CVLAN;
928 qpn_vlan.ins_vlan = 0;
932 netdev_txq_bql_enqueue_prefetchw(ring->tx_queue);
934 /* Packet is good - grab an index and transmit it */
935 index = ring->prod & ring->size_mask;
936 bf_index = ring->prod;
938 /* See if we have enough space for whole descriptor TXBB for setting
939 * SW ownership on next descriptor; if not, use a bounce buffer. */
940 if (likely(index + nr_txbb <= ring->size))
941 tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
943 if (unlikely(nr_txbb > MLX4_MAX_DESC_TXBBS)) {
944 if (netif_msg_tx_err(priv))
945 en_warn(priv, "Oversized header or SG list\n");
948 tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
953 /* Save skb in tx_info ring */
954 tx_info = &ring->tx_info[index];
956 tx_info->nr_txbb = nr_txbb;
958 if (!lso_header_size) {
959 data = &tx_desc->data;
960 data_offset = offsetof(struct mlx4_en_tx_desc, data);
962 int lso_align = ALIGN(lso_header_size - hopbyhop + 4, DS_SIZE);
964 data = (void *)&tx_desc->lso + lso_align;
965 data_offset = offsetof(struct mlx4_en_tx_desc, lso) + lso_align;
968 /* valid only for none inline segments */
969 tx_info->data_offset = data_offset;
971 tx_info->inl = inline_ok;
973 tx_info->linear = lso_header_size < skb_headlen(skb) && !inline_ok;
975 tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
976 data += tx_info->nr_maps - 1;
979 if (!mlx4_en_build_dma_wqe(priv, shinfo, data, skb,
980 lso_header_size, ring->mr_key,
985 * For timestamping add flag to skb_shinfo and
986 * set flag for further reference
988 tx_info->ts_requested = 0;
989 if (unlikely(ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
990 shinfo->tx_flags & SKBTX_HW_TSTAMP)) {
991 shinfo->tx_flags |= SKBTX_IN_PROGRESS;
992 tx_info->ts_requested = 1;
995 /* Prepare ctrl segement apart opcode+ownership, which depends on
996 * whether LSO is used */
997 tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
998 if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
999 if (!skb->encapsulation)
1000 tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
1001 MLX4_WQE_CTRL_TCP_UDP_CSUM);
1003 tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM);
1007 if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
1008 struct ethhdr *ethh;
1010 /* Copy dst mac address to wqe. This allows loopback in eSwitch,
1011 * so that VFs and PF can communicate with each other
1013 ethh = (struct ethhdr *)skb->data;
1014 tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
1015 tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
1018 /* Handle LSO (TSO) packets */
1019 if (lso_header_size) {
1022 /* Mark opcode as LSO */
1023 op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
1024 ((ring->prod & ring->size) ?
1025 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
1027 lso_header_size -= hopbyhop;
1028 /* Fill in the LSO prefix */
1029 tx_desc->lso.mss_hdr_size = cpu_to_be32(
1030 shinfo->gso_size << 16 | lso_header_size);
1033 if (unlikely(hopbyhop)) {
1034 /* remove the HBH header.
1035 * Layout: [Ethernet header][IPv6 header][HBH][TCP header]
1037 memcpy(tx_desc->lso.header, skb->data, ETH_HLEN + sizeof(*h6));
1038 h6 = (struct ipv6hdr *)((char *)tx_desc->lso.header + ETH_HLEN);
1039 h6->nexthdr = IPPROTO_TCP;
1040 /* Copy the TCP header after the IPv6 one */
1042 skb->data + ETH_HLEN + sizeof(*h6) +
1043 sizeof(struct hop_jumbo_hdr),
1045 /* Leave ipv6 payload_len set to 0, as LSO v2 specs request. */
1048 * note that we already verified that it is linear
1050 memcpy(tx_desc->lso.header, skb->data, lso_header_size);
1052 ring->tso_packets++;
1054 i = shinfo->gso_segs;
1055 tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
1058 /* Normal (Non LSO) packet */
1059 op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
1060 ((ring->prod & ring->size) ?
1061 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
1062 tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
1065 ring->bytes += tx_info->nr_bytes;
1068 build_inline_wqe(tx_desc, skb, shinfo, fragptr);
1070 if (skb->encapsulation) {
1078 ip.hdr = skb_inner_network_header(skb);
1079 proto = (ip.v4->version == 4) ? ip.v4->protocol :
1082 if (proto == IPPROTO_TCP || proto == IPPROTO_UDP)
1083 op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP | MLX4_WQE_CTRL_ILP);
1085 op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP);
1088 WRITE_ONCE(ring->prod, ring->prod + nr_txbb);
1090 /* If we used a bounce buffer then copy descriptor back into place */
1091 if (unlikely(bounce))
1092 tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
1094 skb_tx_timestamp(skb);
1096 /* Check available TXBBs And 2K spare for prefetch */
1097 stop_queue = mlx4_en_is_tx_ring_full(ring);
1098 if (unlikely(stop_queue)) {
1099 netif_tx_stop_queue(ring->tx_queue);
1100 ring->queue_stopped++;
1103 send_doorbell = __netdev_tx_sent_queue(ring->tx_queue,
1105 netdev_xmit_more());
1107 real_size = (real_size / 16) & 0x3f;
1109 bf_ok &= desc_size <= MAX_BF && send_doorbell;
1112 qpn_vlan.bf_qpn = ring->doorbell_qpn | cpu_to_be32(real_size);
1114 qpn_vlan.fence_size = real_size;
1116 mlx4_en_tx_write_desc(ring, tx_desc, qpn_vlan, desc_size, bf_index,
1117 op_own, bf_ok, send_doorbell);
1119 if (unlikely(stop_queue)) {
1120 /* If queue was emptied after the if (stop_queue) , and before
1121 * the netif_tx_stop_queue() - need to wake the queue,
1122 * or else it will remain stopped forever.
1123 * Need a memory barrier to make sure ring->cons was not
1124 * updated before queue was stopped.
1128 if (unlikely(!mlx4_en_is_tx_ring_full(ring))) {
1129 netif_tx_wake_queue(ring->tx_queue);
1133 return NETDEV_TX_OK;
1138 dev_kfree_skb_any(skb);
1139 return NETDEV_TX_OK;
1142 #define MLX4_EN_XDP_TX_NRTXBB 1
1143 #define MLX4_EN_XDP_TX_REAL_SZ (((CTRL_SIZE + MLX4_EN_XDP_TX_NRTXBB * DS_SIZE) \
1146 void mlx4_en_init_tx_xdp_ring_descs(struct mlx4_en_priv *priv,
1147 struct mlx4_en_tx_ring *ring)
1151 for (i = 0; i < ring->size; i++) {
1152 struct mlx4_en_tx_info *tx_info = &ring->tx_info[i];
1153 struct mlx4_en_tx_desc *tx_desc = ring->buf +
1154 (i << LOG_TXBB_SIZE);
1156 tx_info->map0_byte_count = PAGE_SIZE;
1157 tx_info->nr_txbb = MLX4_EN_XDP_TX_NRTXBB;
1158 tx_info->data_offset = offsetof(struct mlx4_en_tx_desc, data);
1159 tx_info->ts_requested = 0;
1160 tx_info->nr_maps = 1;
1161 tx_info->linear = 1;
1164 tx_desc->data.lkey = ring->mr_key;
1165 tx_desc->ctrl.qpn_vlan.fence_size = MLX4_EN_XDP_TX_REAL_SZ;
1166 tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
1170 netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_ring *rx_ring,
1171 struct mlx4_en_rx_alloc *frame,
1172 struct mlx4_en_priv *priv, unsigned int length,
1173 int tx_ind, bool *doorbell_pending)
1175 struct mlx4_en_tx_desc *tx_desc;
1176 struct mlx4_en_tx_info *tx_info;
1177 struct mlx4_wqe_data_seg *data;
1178 struct mlx4_en_tx_ring *ring;
1183 if (unlikely(!priv->port_up))
1186 ring = priv->tx_ring[TX_XDP][tx_ind];
1188 if (unlikely(mlx4_en_is_tx_ring_full(ring)))
1191 index = ring->prod & ring->size_mask;
1192 tx_info = &ring->tx_info[index];
1194 tx_desc = ring->buf + (index << LOG_TXBB_SIZE);
1195 data = &tx_desc->data;
1199 tx_info->page = frame->page;
1201 tx_info->map0_dma = dma;
1202 tx_info->nr_bytes = max_t(unsigned int, length, ETH_ZLEN);
1204 dma_sync_single_range_for_device(priv->ddev, dma, frame->page_offset,
1205 length, DMA_TO_DEVICE);
1207 data->addr = cpu_to_be64(dma + frame->page_offset);
1209 data->byte_count = cpu_to_be32(length);
1211 /* tx completion can avoid cache line miss for common cases */
1213 op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
1214 ((ring->prod & ring->size) ?
1215 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
1219 WRITE_ONCE(ring->prod, ring->prod + MLX4_EN_XDP_TX_NRTXBB);
1221 /* Ensure new descriptor hits memory
1222 * before setting ownership of this descriptor to HW
1225 tx_desc->ctrl.owner_opcode = op_own;
1228 *doorbell_pending = true;
1230 return NETDEV_TX_OK;
1233 rx_ring->xdp_tx_full++;
1234 *doorbell_pending = true;
1236 return NETDEV_TX_BUSY;