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/moduleparam.h>
49 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
50 struct mlx4_en_tx_ring **pring, u32 size,
51 u16 stride, int node, int queue_index)
53 struct mlx4_en_dev *mdev = priv->mdev;
54 struct mlx4_en_tx_ring *ring;
58 ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
60 ring = kzalloc(sizeof(*ring), GFP_KERNEL);
62 en_err(priv, "Failed allocating TX ring\n");
68 ring->size_mask = size - 1;
69 ring->stride = stride;
70 ring->full_size = ring->size - HEADROOM - MAX_DESC_TXBBS;
72 tmp = size * sizeof(struct mlx4_en_tx_info);
73 ring->tx_info = kmalloc_node(tmp, GFP_KERNEL | __GFP_NOWARN, node);
75 ring->tx_info = vmalloc(tmp);
82 en_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
85 ring->bounce_buf = kmalloc_node(MAX_DESC_SIZE, GFP_KERNEL, node);
86 if (!ring->bounce_buf) {
87 ring->bounce_buf = kmalloc(MAX_DESC_SIZE, GFP_KERNEL);
88 if (!ring->bounce_buf) {
93 ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
95 /* Allocate HW buffers on provided NUMA node */
96 set_dev_node(&mdev->dev->persist->pdev->dev, node);
97 err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
98 set_dev_node(&mdev->dev->persist->pdev->dev, mdev->dev->numa_node);
100 en_err(priv, "Failed allocating hwq resources\n");
104 ring->buf = ring->wqres.buf.direct.buf;
106 en_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d buf_size:%d dma:%llx\n",
107 ring, ring->buf, ring->size, ring->buf_size,
108 (unsigned long long) ring->wqres.buf.direct.map);
110 err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &ring->qpn,
111 MLX4_RESERVE_ETH_BF_QP);
113 en_err(priv, "failed reserving qp for TX ring\n");
117 err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp, GFP_KERNEL);
119 en_err(priv, "Failed allocating qp %d\n", ring->qpn);
122 ring->qp.event = mlx4_en_sqp_event;
124 err = mlx4_bf_alloc(mdev->dev, &ring->bf, node);
126 en_dbg(DRV, priv, "working without blueflame (%d)\n", err);
127 ring->bf.uar = &mdev->priv_uar;
128 ring->bf.uar->map = mdev->uar_map;
129 ring->bf_enabled = false;
130 ring->bf_alloced = false;
131 priv->pflags &= ~MLX4_EN_PRIV_FLAGS_BLUEFLAME;
133 ring->bf_alloced = true;
134 ring->bf_enabled = !!(priv->pflags &
135 MLX4_EN_PRIV_FLAGS_BLUEFLAME);
138 ring->hwtstamp_tx_type = priv->hwtstamp_config.tx_type;
139 ring->queue_index = queue_index;
141 if (queue_index < priv->num_tx_rings_p_up)
142 cpumask_set_cpu(cpumask_local_spread(queue_index,
143 priv->mdev->dev->numa_node),
144 &ring->affinity_mask);
150 mlx4_qp_release_range(mdev->dev, ring->qpn, 1);
152 mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
154 kfree(ring->bounce_buf);
155 ring->bounce_buf = NULL;
157 kvfree(ring->tx_info);
158 ring->tx_info = NULL;
165 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
166 struct mlx4_en_tx_ring **pring)
168 struct mlx4_en_dev *mdev = priv->mdev;
169 struct mlx4_en_tx_ring *ring = *pring;
170 en_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
172 if (ring->bf_alloced)
173 mlx4_bf_free(mdev->dev, &ring->bf);
174 mlx4_qp_remove(mdev->dev, &ring->qp);
175 mlx4_qp_free(mdev->dev, &ring->qp);
176 mlx4_qp_release_range(priv->mdev->dev, ring->qpn, 1);
177 mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
178 kfree(ring->bounce_buf);
179 ring->bounce_buf = NULL;
180 kvfree(ring->tx_info);
181 ring->tx_info = NULL;
186 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
187 struct mlx4_en_tx_ring *ring,
188 int cq, int user_prio)
190 struct mlx4_en_dev *mdev = priv->mdev;
195 ring->cons = 0xffffffff;
196 ring->last_nr_txbb = 1;
197 memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
198 memset(ring->buf, 0, ring->buf_size);
199 ring->free_tx_desc = mlx4_en_free_tx_desc;
201 ring->qp_state = MLX4_QP_STATE_RST;
202 ring->doorbell_qpn = cpu_to_be32(ring->qp.qpn << 8);
203 ring->mr_key = cpu_to_be32(mdev->mr.key);
205 mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
206 ring->cqn, user_prio, &ring->context);
207 if (ring->bf_alloced)
208 ring->context.usr_page =
209 cpu_to_be32(mlx4_to_hw_uar_index(mdev->dev,
210 ring->bf.uar->index));
212 err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
213 &ring->qp, &ring->qp_state);
214 if (!cpumask_empty(&ring->affinity_mask))
215 netif_set_xps_queue(priv->dev, &ring->affinity_mask,
221 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
222 struct mlx4_en_tx_ring *ring)
224 struct mlx4_en_dev *mdev = priv->mdev;
226 mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
227 MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
230 static inline bool mlx4_en_is_tx_ring_full(struct mlx4_en_tx_ring *ring)
232 return ring->prod - ring->cons > ring->full_size;
235 static void mlx4_en_stamp_wqe(struct mlx4_en_priv *priv,
236 struct mlx4_en_tx_ring *ring, int index,
239 __be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
240 struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
241 struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
242 void *end = ring->buf + ring->buf_size;
243 __be32 *ptr = (__be32 *)tx_desc;
246 /* Optimize the common case when there are no wraparounds */
247 if (likely((void *)tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
248 /* Stamp the freed descriptor */
249 for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
255 /* Stamp the freed descriptor */
256 for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE;
260 if ((void *)ptr >= end) {
262 stamp ^= cpu_to_be32(0x80000000);
269 u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
270 struct mlx4_en_tx_ring *ring,
271 int index, u8 owner, u64 timestamp,
274 struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
275 struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
276 struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
277 void *end = ring->buf + ring->buf_size;
278 struct sk_buff *skb = tx_info->skb;
279 int nr_maps = tx_info->nr_maps;
282 /* We do not touch skb here, so prefetch skb->users location
283 * to speedup consume_skb()
285 prefetchw(&skb->users);
287 if (unlikely(timestamp)) {
288 struct skb_shared_hwtstamps hwts;
290 mlx4_en_fill_hwtstamps(priv->mdev, &hwts, timestamp);
291 skb_tstamp_tx(skb, &hwts);
294 /* Optimize the common case when there are no wraparounds */
295 if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
298 dma_unmap_single(priv->ddev,
300 tx_info->map0_byte_count,
303 dma_unmap_page(priv->ddev,
305 tx_info->map0_byte_count,
307 for (i = 1; i < nr_maps; i++) {
309 dma_unmap_page(priv->ddev,
310 (dma_addr_t)be64_to_cpu(data->addr),
311 be32_to_cpu(data->byte_count),
317 if ((void *) data >= end) {
318 data = ring->buf + ((void *)data - end);
322 dma_unmap_single(priv->ddev,
324 tx_info->map0_byte_count,
327 dma_unmap_page(priv->ddev,
329 tx_info->map0_byte_count,
331 for (i = 1; i < nr_maps; i++) {
333 /* Check for wraparound before unmapping */
334 if ((void *) data >= end)
336 dma_unmap_page(priv->ddev,
337 (dma_addr_t)be64_to_cpu(data->addr),
338 be32_to_cpu(data->byte_count),
343 napi_consume_skb(skb, napi_mode);
345 return tx_info->nr_txbb;
348 u32 mlx4_en_recycle_tx_desc(struct mlx4_en_priv *priv,
349 struct mlx4_en_tx_ring *ring,
350 int index, u8 owner, u64 timestamp,
353 struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
354 struct mlx4_en_rx_alloc frame = {
355 .page = tx_info->page,
356 .dma = tx_info->map0_dma,
358 .page_size = PAGE_SIZE,
361 if (!mlx4_en_rx_recycle(ring->recycle_ring, &frame)) {
362 dma_unmap_page(priv->ddev, tx_info->map0_dma,
363 PAGE_SIZE, priv->frag_info[0].dma_dir);
364 put_page(tx_info->page);
367 return tx_info->nr_txbb;
370 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
372 struct mlx4_en_priv *priv = netdev_priv(dev);
375 /* Skip last polled descriptor */
376 ring->cons += ring->last_nr_txbb;
377 en_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
378 ring->cons, ring->prod);
380 if ((u32) (ring->prod - ring->cons) > ring->size) {
381 if (netif_msg_tx_err(priv))
382 en_warn(priv, "Tx consumer passed producer!\n");
386 while (ring->cons != ring->prod) {
387 ring->last_nr_txbb = ring->free_tx_desc(priv, ring,
388 ring->cons & ring->size_mask,
389 !!(ring->cons & ring->size), 0,
390 0 /* Non-NAPI caller */);
391 ring->cons += ring->last_nr_txbb;
395 netdev_tx_reset_queue(ring->tx_queue);
398 en_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
403 static bool mlx4_en_process_tx_cq(struct net_device *dev,
404 struct mlx4_en_cq *cq, int napi_budget)
406 struct mlx4_en_priv *priv = netdev_priv(dev);
407 struct mlx4_cq *mcq = &cq->mcq;
408 struct mlx4_en_tx_ring *ring = priv->tx_ring[cq->ring];
409 struct mlx4_cqe *cqe;
411 u16 new_index, ring_index, stamp_index;
412 u32 txbbs_skipped = 0;
414 u32 cons_index = mcq->cons_index;
416 u32 size_mask = ring->size_mask;
417 struct mlx4_cqe *buf = cq->buf;
420 int factor = priv->cqe_factor;
422 int budget = priv->tx_work_limit;
429 netdev_txq_bql_complete_prefetchw(ring->tx_queue);
431 index = cons_index & size_mask;
432 cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
433 last_nr_txbb = ACCESS_ONCE(ring->last_nr_txbb);
434 ring_cons = ACCESS_ONCE(ring->cons);
435 ring_index = ring_cons & size_mask;
436 stamp_index = ring_index;
438 /* Process all completed CQEs */
439 while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
440 cons_index & size) && (done < budget)) {
442 * make sure we read the CQE after we read the
447 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
448 MLX4_CQE_OPCODE_ERROR)) {
449 struct mlx4_err_cqe *cqe_err = (struct mlx4_err_cqe *)cqe;
451 en_err(priv, "CQE error - vendor syndrome: 0x%x syndrome: 0x%x\n",
452 cqe_err->vendor_err_syndrome,
456 /* Skip over last polled CQE */
457 new_index = be16_to_cpu(cqe->wqe_index) & size_mask;
462 txbbs_skipped += last_nr_txbb;
463 ring_index = (ring_index + last_nr_txbb) & size_mask;
465 if (unlikely(ring->tx_info[ring_index].ts_requested))
466 timestamp = mlx4_en_get_cqe_ts(cqe);
468 /* free next descriptor */
469 last_nr_txbb = ring->free_tx_desc(
470 priv, ring, ring_index,
471 !!((ring_cons + txbbs_skipped) &
472 ring->size), timestamp, napi_budget);
474 mlx4_en_stamp_wqe(priv, ring, stamp_index,
475 !!((ring_cons + txbbs_stamp) &
477 stamp_index = ring_index;
478 txbbs_stamp = txbbs_skipped;
480 bytes += ring->tx_info[ring_index].nr_bytes;
481 } while ((++done < budget) && (ring_index != new_index));
484 index = cons_index & size_mask;
485 cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
490 * To prevent CQ overflow we first update CQ consumer and only then
493 mcq->cons_index = cons_index;
497 /* we want to dirty this cache line once */
498 ACCESS_ONCE(ring->last_nr_txbb) = last_nr_txbb;
499 ACCESS_ONCE(ring->cons) = ring_cons + txbbs_skipped;
501 if (ring->free_tx_desc == mlx4_en_recycle_tx_desc)
502 return done < budget;
504 netdev_tx_completed_queue(ring->tx_queue, packets, bytes);
506 /* Wakeup Tx queue if this stopped, and ring is not full.
508 if (netif_tx_queue_stopped(ring->tx_queue) &&
509 !mlx4_en_is_tx_ring_full(ring)) {
510 netif_tx_wake_queue(ring->tx_queue);
513 return done < budget;
516 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
518 struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
519 struct mlx4_en_priv *priv = netdev_priv(cq->dev);
521 if (likely(priv->port_up))
522 napi_schedule_irqoff(&cq->napi);
524 mlx4_en_arm_cq(priv, cq);
527 /* TX CQ polling - called by NAPI */
528 int mlx4_en_poll_tx_cq(struct napi_struct *napi, int budget)
530 struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
531 struct net_device *dev = cq->dev;
532 struct mlx4_en_priv *priv = netdev_priv(dev);
535 clean_complete = mlx4_en_process_tx_cq(dev, cq, budget);
540 mlx4_en_arm_cq(priv, cq);
545 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
546 struct mlx4_en_tx_ring *ring,
548 unsigned int desc_size)
550 u32 copy = (ring->size - index) * TXBB_SIZE;
553 for (i = desc_size - copy - 4; i >= 0; i -= 4) {
554 if ((i & (TXBB_SIZE - 1)) == 0)
557 *((u32 *) (ring->buf + i)) =
558 *((u32 *) (ring->bounce_buf + copy + i));
561 for (i = copy - 4; i >= 4 ; i -= 4) {
562 if ((i & (TXBB_SIZE - 1)) == 0)
565 *((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
566 *((u32 *) (ring->bounce_buf + i));
569 /* Return real descriptor location */
570 return ring->buf + index * TXBB_SIZE;
573 /* Decide if skb can be inlined in tx descriptor to avoid dma mapping
575 * It seems strange we do not simply use skb_copy_bits().
576 * This would allow to inline all skbs iff skb->len <= inline_thold
578 * Note that caller already checked skb was not a gso packet
580 static bool is_inline(int inline_thold, const struct sk_buff *skb,
581 const struct skb_shared_info *shinfo,
586 if (skb->len > inline_thold || !inline_thold)
589 if (shinfo->nr_frags == 1) {
590 ptr = skb_frag_address_safe(&shinfo->frags[0]);
596 if (shinfo->nr_frags)
601 static int inline_size(const struct sk_buff *skb)
603 if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
604 <= MLX4_INLINE_ALIGN)
605 return ALIGN(skb->len + CTRL_SIZE +
606 sizeof(struct mlx4_wqe_inline_seg), 16);
608 return ALIGN(skb->len + CTRL_SIZE + 2 *
609 sizeof(struct mlx4_wqe_inline_seg), 16);
612 static int get_real_size(const struct sk_buff *skb,
613 const struct skb_shared_info *shinfo,
614 struct net_device *dev,
615 int *lso_header_size,
619 struct mlx4_en_priv *priv = netdev_priv(dev);
622 if (shinfo->gso_size) {
624 if (skb->encapsulation)
625 *lso_header_size = (skb_inner_transport_header(skb) - skb->data) + inner_tcp_hdrlen(skb);
627 *lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
628 real_size = CTRL_SIZE + shinfo->nr_frags * DS_SIZE +
629 ALIGN(*lso_header_size + 4, DS_SIZE);
630 if (unlikely(*lso_header_size != skb_headlen(skb))) {
631 /* We add a segment for the skb linear buffer only if
632 * it contains data */
633 if (*lso_header_size < skb_headlen(skb))
634 real_size += DS_SIZE;
636 if (netif_msg_tx_err(priv))
637 en_warn(priv, "Non-linear headers\n");
642 *lso_header_size = 0;
643 *inline_ok = is_inline(priv->prof->inline_thold, skb,
647 real_size = inline_size(skb);
649 real_size = CTRL_SIZE +
650 (shinfo->nr_frags + 1) * DS_SIZE;
656 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc,
657 const struct sk_buff *skb,
658 const struct skb_shared_info *shinfo,
661 struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
662 int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
663 unsigned int hlen = skb_headlen(skb);
665 if (skb->len <= spc) {
666 if (likely(skb->len >= MIN_PKT_LEN)) {
667 inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
669 inl->byte_count = cpu_to_be32(1 << 31 | MIN_PKT_LEN);
670 memset(((void *)(inl + 1)) + skb->len, 0,
671 MIN_PKT_LEN - skb->len);
673 skb_copy_from_linear_data(skb, inl + 1, hlen);
674 if (shinfo->nr_frags)
675 memcpy(((void *)(inl + 1)) + hlen, fragptr,
676 skb_frag_size(&shinfo->frags[0]));
679 inl->byte_count = cpu_to_be32(1 << 31 | spc);
681 skb_copy_from_linear_data(skb, inl + 1, hlen);
683 memcpy(((void *)(inl + 1)) + hlen,
684 fragptr, spc - hlen);
685 fragptr += spc - hlen;
687 inl = (void *) (inl + 1) + spc;
688 memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
690 skb_copy_from_linear_data(skb, inl + 1, spc);
691 inl = (void *) (inl + 1) + spc;
692 skb_copy_from_linear_data_offset(skb, spc, inl + 1,
694 if (shinfo->nr_frags)
695 memcpy(((void *)(inl + 1)) + hlen - spc,
697 skb_frag_size(&shinfo->frags[0]));
701 inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
705 u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb,
706 void *accel_priv, select_queue_fallback_t fallback)
708 struct mlx4_en_priv *priv = netdev_priv(dev);
709 u16 rings_p_up = priv->num_tx_rings_p_up;
713 return skb_tx_hash(dev, skb);
715 if (skb_vlan_tag_present(skb))
716 up = skb_vlan_tag_get(skb) >> VLAN_PRIO_SHIFT;
718 return fallback(dev, skb) % rings_p_up + up * rings_p_up;
721 static void mlx4_bf_copy(void __iomem *dst, const void *src,
722 unsigned int bytecnt)
724 __iowrite64_copy(dst, src, bytecnt / 8);
727 void mlx4_en_xmit_doorbell(struct mlx4_en_tx_ring *ring)
730 /* Since there is no iowrite*_native() that writes the
731 * value as is, without byteswapping - using the one
732 * the doesn't do byteswapping in the relevant arch
735 #if defined(__LITTLE_ENDIAN)
741 ring->bf.uar->map + MLX4_SEND_DOORBELL);
744 static void mlx4_en_tx_write_desc(struct mlx4_en_tx_ring *ring,
745 struct mlx4_en_tx_desc *tx_desc,
746 union mlx4_wqe_qpn_vlan qpn_vlan,
747 int desc_size, int bf_index,
748 __be32 op_own, bool bf_ok,
751 tx_desc->ctrl.qpn_vlan = qpn_vlan;
754 op_own |= htonl((bf_index & 0xffff) << 8);
755 /* Ensure new descriptor hits memory
756 * before setting ownership of this descriptor to HW
759 tx_desc->ctrl.owner_opcode = op_own;
763 mlx4_bf_copy(ring->bf.reg + ring->bf.offset, &tx_desc->ctrl,
768 ring->bf.offset ^= ring->bf.buf_size;
770 /* Ensure new descriptor hits memory
771 * before setting ownership of this descriptor to HW
774 tx_desc->ctrl.owner_opcode = op_own;
776 mlx4_en_xmit_doorbell(ring);
782 netdev_tx_t mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
784 struct skb_shared_info *shinfo = skb_shinfo(skb);
785 struct mlx4_en_priv *priv = netdev_priv(dev);
786 union mlx4_wqe_qpn_vlan qpn_vlan = {};
787 struct device *ddev = priv->ddev;
788 struct mlx4_en_tx_ring *ring;
789 struct mlx4_en_tx_desc *tx_desc;
790 struct mlx4_wqe_data_seg *data;
791 struct mlx4_en_tx_info *tx_info;
801 void *fragptr = NULL;
809 tx_ind = skb_get_queue_mapping(skb);
810 ring = priv->tx_ring[tx_ind];
815 /* fetch ring->cons far ahead before needing it to avoid stall */
816 ring_cons = ACCESS_ONCE(ring->cons);
818 real_size = get_real_size(skb, shinfo, dev, &lso_header_size,
819 &inline_ok, &fragptr);
820 if (unlikely(!real_size))
823 /* Align descriptor to TXBB size */
824 desc_size = ALIGN(real_size, TXBB_SIZE);
825 nr_txbb = desc_size / TXBB_SIZE;
826 if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
827 if (netif_msg_tx_err(priv))
828 en_warn(priv, "Oversized header or SG list\n");
832 bf_ok = ring->bf_enabled;
833 if (skb_vlan_tag_present(skb)) {
834 qpn_vlan.vlan_tag = cpu_to_be16(skb_vlan_tag_get(skb));
835 vlan_proto = be16_to_cpu(skb->vlan_proto);
836 if (vlan_proto == ETH_P_8021AD)
837 qpn_vlan.ins_vlan = MLX4_WQE_CTRL_INS_SVLAN;
838 else if (vlan_proto == ETH_P_8021Q)
839 qpn_vlan.ins_vlan = MLX4_WQE_CTRL_INS_CVLAN;
841 qpn_vlan.ins_vlan = 0;
845 netdev_txq_bql_enqueue_prefetchw(ring->tx_queue);
847 /* Track current inflight packets for performance analysis */
848 AVG_PERF_COUNTER(priv->pstats.inflight_avg,
849 (u32)(ring->prod - ring_cons - 1));
851 /* Packet is good - grab an index and transmit it */
852 index = ring->prod & ring->size_mask;
853 bf_index = ring->prod;
855 /* See if we have enough space for whole descriptor TXBB for setting
856 * SW ownership on next descriptor; if not, use a bounce buffer. */
857 if (likely(index + nr_txbb <= ring->size))
858 tx_desc = ring->buf + index * TXBB_SIZE;
860 tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
865 /* Save skb in tx_info ring */
866 tx_info = &ring->tx_info[index];
868 tx_info->nr_txbb = nr_txbb;
870 data = &tx_desc->data;
872 data = ((void *)&tx_desc->lso + ALIGN(lso_header_size + 4,
875 /* valid only for none inline segments */
876 tx_info->data_offset = (void *)data - (void *)tx_desc;
878 tx_info->inl = inline_ok;
880 tx_info->linear = (lso_header_size < skb_headlen(skb) &&
883 tx_info->nr_maps = shinfo->nr_frags + tx_info->linear;
884 data += tx_info->nr_maps - 1;
890 /* Map fragments if any */
891 for (i_frag = shinfo->nr_frags - 1; i_frag >= 0; i_frag--) {
892 const struct skb_frag_struct *frag;
894 frag = &shinfo->frags[i_frag];
895 byte_count = skb_frag_size(frag);
896 dma = skb_frag_dma_map(ddev, frag,
899 if (dma_mapping_error(ddev, dma))
902 data->addr = cpu_to_be64(dma);
903 data->lkey = ring->mr_key;
905 data->byte_count = cpu_to_be32(byte_count);
909 /* Map linear part if needed */
910 if (tx_info->linear) {
911 byte_count = skb_headlen(skb) - lso_header_size;
913 dma = dma_map_single(ddev, skb->data +
914 lso_header_size, byte_count,
916 if (dma_mapping_error(ddev, dma))
919 data->addr = cpu_to_be64(dma);
920 data->lkey = ring->mr_key;
922 data->byte_count = cpu_to_be32(byte_count);
924 /* tx completion can avoid cache line miss for common cases */
925 tx_info->map0_dma = dma;
926 tx_info->map0_byte_count = byte_count;
930 * For timestamping add flag to skb_shinfo and
931 * set flag for further reference
933 tx_info->ts_requested = 0;
934 if (unlikely(ring->hwtstamp_tx_type == HWTSTAMP_TX_ON &&
935 shinfo->tx_flags & SKBTX_HW_TSTAMP)) {
936 shinfo->tx_flags |= SKBTX_IN_PROGRESS;
937 tx_info->ts_requested = 1;
940 /* Prepare ctrl segement apart opcode+ownership, which depends on
941 * whether LSO is used */
942 tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
943 if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
944 if (!skb->encapsulation)
945 tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
946 MLX4_WQE_CTRL_TCP_UDP_CSUM);
948 tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM);
952 if (priv->flags & MLX4_EN_FLAG_ENABLE_HW_LOOPBACK) {
955 /* Copy dst mac address to wqe. This allows loopback in eSwitch,
956 * so that VFs and PF can communicate with each other
958 ethh = (struct ethhdr *)skb->data;
959 tx_desc->ctrl.srcrb_flags16[0] = get_unaligned((__be16 *)ethh->h_dest);
960 tx_desc->ctrl.imm = get_unaligned((__be32 *)(ethh->h_dest + 2));
963 /* Handle LSO (TSO) packets */
964 if (lso_header_size) {
967 /* Mark opcode as LSO */
968 op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
969 ((ring->prod & ring->size) ?
970 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
972 /* Fill in the LSO prefix */
973 tx_desc->lso.mss_hdr_size = cpu_to_be32(
974 shinfo->gso_size << 16 | lso_header_size);
977 * note that we already verified that it is linear */
978 memcpy(tx_desc->lso.header, skb->data, lso_header_size);
982 i = ((skb->len - lso_header_size) / shinfo->gso_size) +
983 !!((skb->len - lso_header_size) % shinfo->gso_size);
984 tx_info->nr_bytes = skb->len + (i - 1) * lso_header_size;
987 /* Normal (Non LSO) packet */
988 op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
989 ((ring->prod & ring->size) ?
990 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
991 tx_info->nr_bytes = max_t(unsigned int, skb->len, ETH_ZLEN);
994 ring->bytes += tx_info->nr_bytes;
995 netdev_tx_sent_queue(ring->tx_queue, tx_info->nr_bytes);
996 AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
999 build_inline_wqe(tx_desc, skb, shinfo, fragptr);
1001 if (skb->encapsulation) {
1009 ip.hdr = skb_inner_network_header(skb);
1010 proto = (ip.v4->version == 4) ? ip.v4->protocol :
1013 if (proto == IPPROTO_TCP || proto == IPPROTO_UDP)
1014 op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP | MLX4_WQE_CTRL_ILP);
1016 op_own |= cpu_to_be32(MLX4_WQE_CTRL_IIP);
1019 ring->prod += nr_txbb;
1021 /* If we used a bounce buffer then copy descriptor back into place */
1022 if (unlikely(bounce))
1023 tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
1025 skb_tx_timestamp(skb);
1027 /* Check available TXBBs And 2K spare for prefetch */
1028 stop_queue = mlx4_en_is_tx_ring_full(ring);
1029 if (unlikely(stop_queue)) {
1030 netif_tx_stop_queue(ring->tx_queue);
1031 ring->queue_stopped++;
1033 send_doorbell = !skb->xmit_more || netif_xmit_stopped(ring->tx_queue);
1035 real_size = (real_size / 16) & 0x3f;
1037 bf_ok &= desc_size <= MAX_BF && send_doorbell;
1040 qpn_vlan.bf_qpn = ring->doorbell_qpn | cpu_to_be32(real_size);
1042 qpn_vlan.fence_size = real_size;
1044 mlx4_en_tx_write_desc(ring, tx_desc, qpn_vlan, desc_size, bf_index,
1045 op_own, bf_ok, send_doorbell);
1047 if (unlikely(stop_queue)) {
1048 /* If queue was emptied after the if (stop_queue) , and before
1049 * the netif_tx_stop_queue() - need to wake the queue,
1050 * or else it will remain stopped forever.
1051 * Need a memory barrier to make sure ring->cons was not
1052 * updated before queue was stopped.
1056 ring_cons = ACCESS_ONCE(ring->cons);
1057 if (unlikely(!mlx4_en_is_tx_ring_full(ring))) {
1058 netif_tx_wake_queue(ring->tx_queue);
1062 return NETDEV_TX_OK;
1065 en_err(priv, "DMA mapping error\n");
1067 while (++i_frag < shinfo->nr_frags) {
1069 dma_unmap_page(ddev, (dma_addr_t) be64_to_cpu(data->addr),
1070 be32_to_cpu(data->byte_count),
1075 dev_kfree_skb_any(skb);
1077 return NETDEV_TX_OK;
1080 netdev_tx_t mlx4_en_xmit_frame(struct mlx4_en_rx_alloc *frame,
1081 struct net_device *dev, unsigned int length,
1082 int tx_ind, int *doorbell_pending)
1084 struct mlx4_en_priv *priv = netdev_priv(dev);
1085 union mlx4_wqe_qpn_vlan qpn_vlan = {};
1086 struct mlx4_en_tx_ring *ring;
1087 struct mlx4_en_tx_desc *tx_desc;
1088 struct mlx4_wqe_data_seg *data;
1089 struct mlx4_en_tx_info *tx_info;
1090 int index, bf_index;
1100 BUILD_BUG_ON_MSG(ALIGN(CTRL_SIZE + DS_SIZE, TXBB_SIZE) != TXBB_SIZE,
1101 "mlx4_en_xmit_frame requires minimum size tx desc");
1103 ring = priv->tx_ring[tx_ind];
1108 if (mlx4_en_is_tx_ring_full(ring))
1111 /* fetch ring->cons far ahead before needing it to avoid stall */
1112 ring_cons = READ_ONCE(ring->cons);
1114 index = ring->prod & ring->size_mask;
1115 tx_info = &ring->tx_info[index];
1117 bf_ok = ring->bf_enabled;
1119 /* Track current inflight packets for performance analysis */
1120 AVG_PERF_COUNTER(priv->pstats.inflight_avg,
1121 (u32)(ring->prod - ring_cons - 1));
1123 bf_index = ring->prod;
1124 tx_desc = ring->buf + index * TXBB_SIZE;
1125 data = &tx_desc->data;
1129 tx_info->page = frame->page;
1131 tx_info->map0_dma = dma;
1132 tx_info->map0_byte_count = length;
1133 tx_info->nr_txbb = nr_txbb;
1134 tx_info->nr_bytes = max_t(unsigned int, length, ETH_ZLEN);
1135 tx_info->data_offset = (void *)data - (void *)tx_desc;
1136 tx_info->ts_requested = 0;
1137 tx_info->nr_maps = 1;
1138 tx_info->linear = 1;
1141 dma_sync_single_for_device(priv->ddev, dma, length, PCI_DMA_TODEVICE);
1143 data->addr = cpu_to_be64(dma);
1144 data->lkey = ring->mr_key;
1146 data->byte_count = cpu_to_be32(length);
1148 /* tx completion can avoid cache line miss for common cases */
1149 tx_desc->ctrl.srcrb_flags = priv->ctrl_flags;
1151 op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
1152 ((ring->prod & ring->size) ?
1153 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
1156 ring->bytes += tx_info->nr_bytes;
1157 AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, length);
1159 ring->prod += nr_txbb;
1161 stop_queue = mlx4_en_is_tx_ring_full(ring);
1162 send_doorbell = stop_queue ||
1163 *doorbell_pending > MLX4_EN_DOORBELL_BUDGET;
1164 bf_ok &= send_doorbell;
1166 real_size = ((CTRL_SIZE + nr_txbb * DS_SIZE) / 16) & 0x3f;
1169 qpn_vlan.bf_qpn = ring->doorbell_qpn | cpu_to_be32(real_size);
1171 qpn_vlan.fence_size = real_size;
1173 mlx4_en_tx_write_desc(ring, tx_desc, qpn_vlan, TXBB_SIZE, bf_index,
1174 op_own, bf_ok, send_doorbell);
1175 *doorbell_pending = send_doorbell ? 0 : *doorbell_pending + 1;
1177 return NETDEV_TX_OK;
1181 return NETDEV_TX_BUSY;