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/mlx4/qp.h>
37 #include <linux/skbuff.h>
38 #include <linux/if_vlan.h>
39 #include <linux/vmalloc.h>
44 MAX_INLINE = 104, /* 128 - 16 - 4 - 4 */
47 static int inline_thold __read_mostly = MAX_INLINE;
49 module_param_named(inline_thold, inline_thold, int, 0444);
50 MODULE_PARM_DESC(inline_thold, "treshold for using inline data");
52 int mlx4_en_create_tx_ring(struct mlx4_en_priv *priv,
53 struct mlx4_en_tx_ring *ring, u32 size,
56 struct mlx4_en_dev *mdev = priv->mdev;
61 ring->size_mask = size - 1;
62 ring->stride = stride;
64 inline_thold = min(inline_thold, MAX_INLINE);
66 spin_lock_init(&ring->comp_lock);
68 tmp = size * sizeof(struct mlx4_en_tx_info);
69 ring->tx_info = vmalloc(tmp);
71 mlx4_err(mdev, "Failed allocating tx_info ring\n");
74 mlx4_dbg(DRV, priv, "Allocated tx_info ring at addr:%p size:%d\n",
77 ring->bounce_buf = kmalloc(MAX_DESC_SIZE, GFP_KERNEL);
78 if (!ring->bounce_buf) {
79 mlx4_err(mdev, "Failed allocating bounce buffer\n");
83 ring->buf_size = ALIGN(size * ring->stride, MLX4_EN_PAGE_SIZE);
85 err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size,
88 mlx4_err(mdev, "Failed allocating hwq resources\n");
92 err = mlx4_en_map_buffer(&ring->wqres.buf);
94 mlx4_err(mdev, "Failed to map TX buffer\n");
98 ring->buf = ring->wqres.buf.direct.buf;
100 mlx4_dbg(DRV, priv, "Allocated TX ring (addr:%p) - buf:%p size:%d "
101 "buf_size:%d dma:%llx\n", ring, ring->buf, ring->size,
102 ring->buf_size, (unsigned long long) ring->wqres.buf.direct.map);
104 err = mlx4_qp_reserve_range(mdev->dev, 1, 1, &ring->qpn);
106 mlx4_err(mdev, "Failed reserving qp for tx ring.\n");
110 err = mlx4_qp_alloc(mdev->dev, ring->qpn, &ring->qp);
112 mlx4_err(mdev, "Failed allocating qp %d\n", ring->qpn);
119 mlx4_qp_release_range(mdev->dev, ring->qpn, 1);
121 mlx4_en_unmap_buffer(&ring->wqres.buf);
123 mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
125 kfree(ring->bounce_buf);
126 ring->bounce_buf = NULL;
128 vfree(ring->tx_info);
129 ring->tx_info = NULL;
133 void mlx4_en_destroy_tx_ring(struct mlx4_en_priv *priv,
134 struct mlx4_en_tx_ring *ring)
136 struct mlx4_en_dev *mdev = priv->mdev;
137 mlx4_dbg(DRV, priv, "Destroying tx ring, qpn: %d\n", ring->qpn);
139 mlx4_qp_remove(mdev->dev, &ring->qp);
140 mlx4_qp_free(mdev->dev, &ring->qp);
141 mlx4_qp_release_range(mdev->dev, ring->qpn, 1);
142 mlx4_en_unmap_buffer(&ring->wqres.buf);
143 mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
144 kfree(ring->bounce_buf);
145 ring->bounce_buf = NULL;
146 vfree(ring->tx_info);
147 ring->tx_info = NULL;
150 int mlx4_en_activate_tx_ring(struct mlx4_en_priv *priv,
151 struct mlx4_en_tx_ring *ring,
154 struct mlx4_en_dev *mdev = priv->mdev;
159 ring->cons = 0xffffffff;
160 ring->last_nr_txbb = 1;
163 memset(ring->tx_info, 0, ring->size * sizeof(struct mlx4_en_tx_info));
164 memset(ring->buf, 0, ring->buf_size);
166 ring->qp_state = MLX4_QP_STATE_RST;
167 ring->doorbell_qpn = swab32(ring->qp.qpn << 8);
169 mlx4_en_fill_qp_context(priv, ring->size, ring->stride, 1, 0, ring->qpn,
170 ring->cqn, srqn, &ring->context);
172 err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, &ring->context,
173 &ring->qp, &ring->qp_state);
178 void mlx4_en_deactivate_tx_ring(struct mlx4_en_priv *priv,
179 struct mlx4_en_tx_ring *ring)
181 struct mlx4_en_dev *mdev = priv->mdev;
183 mlx4_qp_modify(mdev->dev, NULL, ring->qp_state,
184 MLX4_QP_STATE_RST, NULL, 0, 0, &ring->qp);
188 static u32 mlx4_en_free_tx_desc(struct mlx4_en_priv *priv,
189 struct mlx4_en_tx_ring *ring,
192 struct mlx4_en_dev *mdev = priv->mdev;
193 struct mlx4_en_tx_info *tx_info = &ring->tx_info[index];
194 struct mlx4_en_tx_desc *tx_desc = ring->buf + index * TXBB_SIZE;
195 struct mlx4_wqe_data_seg *data = (void *) tx_desc + tx_info->data_offset;
196 struct sk_buff *skb = tx_info->skb;
197 struct skb_frag_struct *frag;
198 void *end = ring->buf + ring->buf_size;
199 int frags = skb_shinfo(skb)->nr_frags;
201 __be32 *ptr = (__be32 *)tx_desc;
202 __be32 stamp = cpu_to_be32(STAMP_VAL | (!!owner << STAMP_SHIFT));
204 /* Optimize the common case when there are no wraparounds */
205 if (likely((void *) tx_desc + tx_info->nr_txbb * TXBB_SIZE <= end)) {
206 if (tx_info->linear) {
207 pci_unmap_single(mdev->pdev,
208 (dma_addr_t) be64_to_cpu(data->addr),
209 be32_to_cpu(data->byte_count),
214 for (i = 0; i < frags; i++) {
215 frag = &skb_shinfo(skb)->frags[i];
216 pci_unmap_page(mdev->pdev,
217 (dma_addr_t) be64_to_cpu(data[i].addr),
218 frag->size, PCI_DMA_TODEVICE);
220 /* Stamp the freed descriptor */
221 for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE; i += STAMP_STRIDE) {
227 if ((void *) data >= end) {
228 data = (struct mlx4_wqe_data_seg *)
229 (ring->buf + ((void *) data - end));
232 if (tx_info->linear) {
233 pci_unmap_single(mdev->pdev,
234 (dma_addr_t) be64_to_cpu(data->addr),
235 be32_to_cpu(data->byte_count),
240 for (i = 0; i < frags; i++) {
241 /* Check for wraparound before unmapping */
242 if ((void *) data >= end)
243 data = (struct mlx4_wqe_data_seg *) ring->buf;
244 frag = &skb_shinfo(skb)->frags[i];
245 pci_unmap_page(mdev->pdev,
246 (dma_addr_t) be64_to_cpu(data->addr),
247 frag->size, PCI_DMA_TODEVICE);
249 /* Stamp the freed descriptor */
250 for (i = 0; i < tx_info->nr_txbb * TXBB_SIZE; i += STAMP_STRIDE) {
253 if ((void *) ptr >= end) {
255 stamp ^= cpu_to_be32(0x80000000);
260 dev_kfree_skb_any(skb);
261 return tx_info->nr_txbb;
265 int mlx4_en_free_tx_buf(struct net_device *dev, struct mlx4_en_tx_ring *ring)
267 struct mlx4_en_priv *priv = netdev_priv(dev);
270 /* Skip last polled descriptor */
271 ring->cons += ring->last_nr_txbb;
272 mlx4_dbg(DRV, priv, "Freeing Tx buf - cons:0x%x prod:0x%x\n",
273 ring->cons, ring->prod);
275 if ((u32) (ring->prod - ring->cons) > ring->size) {
276 if (netif_msg_tx_err(priv))
277 mlx4_warn(priv->mdev, "Tx consumer passed producer!\n");
281 while (ring->cons != ring->prod) {
282 ring->last_nr_txbb = mlx4_en_free_tx_desc(priv, ring,
283 ring->cons & ring->size_mask,
284 !!(ring->cons & ring->size));
285 ring->cons += ring->last_nr_txbb;
290 mlx4_dbg(DRV, priv, "Freed %d uncompleted tx descriptors\n", cnt);
295 void mlx4_en_set_prio_map(struct mlx4_en_priv *priv, u16 *prio_map, u32 ring_num)
297 int block = 8 / ring_num;
298 int extra = 8 - (block * ring_num);
304 for (prio = 0; prio < 8; prio++)
309 for (prio = 0; prio < 8; prio++) {
310 if (extra && (num == block + 1)) {
314 } else if (!extra && (num == block)) {
318 prio_map[prio] = ring;
319 mlx4_dbg(DRV, priv, " prio:%d --> ring:%d\n", prio, ring);
324 static void mlx4_en_process_tx_cq(struct net_device *dev, struct mlx4_en_cq *cq)
326 struct mlx4_en_priv *priv = netdev_priv(dev);
327 struct mlx4_cq *mcq = &cq->mcq;
328 struct mlx4_en_tx_ring *ring = &priv->tx_ring[cq->ring];
329 struct mlx4_cqe *cqe = cq->buf;
332 u32 txbbs_skipped = 0;
335 /* index always points to the first TXBB of the last polled descriptor */
336 index = ring->cons & ring->size_mask;
337 new_index = be16_to_cpu(cqe->wqe_index) & ring->size_mask;
338 if (index == new_index)
345 * We use a two-stage loop:
346 * - the first samples the HW-updated CQE
347 * - the second frees TXBBs until the last sample
348 * This lets us amortize CQE cache misses, while still polling the CQ
349 * until is quiescent.
351 cq_last_sav = mcq->cons_index;
354 /* Skip over last polled CQE */
355 index = (index + ring->last_nr_txbb) & ring->size_mask;
356 txbbs_skipped += ring->last_nr_txbb;
359 ring->last_nr_txbb = mlx4_en_free_tx_desc(
361 !!((ring->cons + txbbs_skipped) &
365 } while (index != new_index);
367 new_index = be16_to_cpu(cqe->wqe_index) & ring->size_mask;
368 } while (index != new_index);
369 AVG_PERF_COUNTER(priv->pstats.tx_coal_avg,
370 (u32) (mcq->cons_index - cq_last_sav));
373 * To prevent CQ overflow we first update CQ consumer and only then
378 ring->cons += txbbs_skipped;
380 /* Wakeup Tx queue if this ring stopped it */
381 if (unlikely(ring->blocked)) {
382 if ((u32) (ring->prod - ring->cons) <=
383 ring->size - HEADROOM - MAX_DESC_TXBBS) {
385 /* TODO: support multiqueue netdevs. Currently, we block
386 * when *any* ring is full. Note that:
387 * - 2 Tx rings can unblock at the same time and call
388 * netif_wake_queue(), which is OK since this
389 * operation is idempotent.
390 * - We might wake the queue just after another ring
391 * stopped it. This is no big deal because the next
392 * transmission on that ring would stop the queue.
395 netif_wake_queue(dev);
396 priv->port_stats.wake_queue++;
401 void mlx4_en_tx_irq(struct mlx4_cq *mcq)
403 struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
404 struct mlx4_en_priv *priv = netdev_priv(cq->dev);
405 struct mlx4_en_tx_ring *ring = &priv->tx_ring[cq->ring];
407 if (!spin_trylock(&ring->comp_lock))
409 mlx4_en_process_tx_cq(cq->dev, cq);
410 mod_timer(&cq->timer, jiffies + 1);
411 spin_unlock(&ring->comp_lock);
415 void mlx4_en_poll_tx_cq(unsigned long data)
417 struct mlx4_en_cq *cq = (struct mlx4_en_cq *) data;
418 struct mlx4_en_priv *priv = netdev_priv(cq->dev);
419 struct mlx4_en_tx_ring *ring = &priv->tx_ring[cq->ring];
422 INC_PERF_COUNTER(priv->pstats.tx_poll);
424 if (!spin_trylock(&ring->comp_lock)) {
425 mod_timer(&cq->timer, jiffies + MLX4_EN_TX_POLL_TIMEOUT);
428 mlx4_en_process_tx_cq(cq->dev, cq);
429 inflight = (u32) (ring->prod - ring->cons - ring->last_nr_txbb);
431 /* If there are still packets in flight and the timer has not already
432 * been scheduled by the Tx routine then schedule it here to guarantee
433 * completion processing of these packets */
434 if (inflight && priv->port_up)
435 mod_timer(&cq->timer, jiffies + MLX4_EN_TX_POLL_TIMEOUT);
437 spin_unlock(&ring->comp_lock);
440 static struct mlx4_en_tx_desc *mlx4_en_bounce_to_desc(struct mlx4_en_priv *priv,
441 struct mlx4_en_tx_ring *ring,
443 unsigned int desc_size)
445 u32 copy = (ring->size - index) * TXBB_SIZE;
448 for (i = desc_size - copy - 4; i >= 0; i -= 4) {
449 if ((i & (TXBB_SIZE - 1)) == 0)
452 *((u32 *) (ring->buf + i)) =
453 *((u32 *) (ring->bounce_buf + copy + i));
456 for (i = copy - 4; i >= 4 ; i -= 4) {
457 if ((i & (TXBB_SIZE - 1)) == 0)
460 *((u32 *) (ring->buf + index * TXBB_SIZE + i)) =
461 *((u32 *) (ring->bounce_buf + i));
464 /* Return real descriptor location */
465 return ring->buf + index * TXBB_SIZE;
468 static inline void mlx4_en_xmit_poll(struct mlx4_en_priv *priv, int tx_ind)
470 struct mlx4_en_cq *cq = &priv->tx_cq[tx_ind];
471 struct mlx4_en_tx_ring *ring = &priv->tx_ring[tx_ind];
473 /* If we don't have a pending timer, set one up to catch our recent
474 post in case the interface becomes idle */
475 if (!timer_pending(&cq->timer))
476 mod_timer(&cq->timer, jiffies + MLX4_EN_TX_POLL_TIMEOUT);
478 /* Poll the CQ every mlx4_en_TX_MODER_POLL packets */
479 if ((++ring->poll_cnt & (MLX4_EN_TX_POLL_MODER - 1)) == 0)
480 if (spin_trylock(&ring->comp_lock)) {
481 mlx4_en_process_tx_cq(priv->dev, cq);
482 spin_unlock(&ring->comp_lock);
486 static void *get_frag_ptr(struct sk_buff *skb)
488 struct skb_frag_struct *frag = &skb_shinfo(skb)->frags[0];
489 struct page *page = frag->page;
492 ptr = page_address(page);
496 return ptr + frag->page_offset;
499 static int is_inline(struct sk_buff *skb, void **pfrag)
503 if (inline_thold && !skb_is_gso(skb) && skb->len <= inline_thold) {
504 if (skb_shinfo(skb)->nr_frags == 1) {
505 ptr = get_frag_ptr(skb);
513 } else if (unlikely(skb_shinfo(skb)->nr_frags))
522 static int inline_size(struct sk_buff *skb)
524 if (skb->len + CTRL_SIZE + sizeof(struct mlx4_wqe_inline_seg)
525 <= MLX4_INLINE_ALIGN)
526 return ALIGN(skb->len + CTRL_SIZE +
527 sizeof(struct mlx4_wqe_inline_seg), 16);
529 return ALIGN(skb->len + CTRL_SIZE + 2 *
530 sizeof(struct mlx4_wqe_inline_seg), 16);
533 static int get_real_size(struct sk_buff *skb, struct net_device *dev,
534 int *lso_header_size)
536 struct mlx4_en_priv *priv = netdev_priv(dev);
537 struct mlx4_en_dev *mdev = priv->mdev;
540 if (skb_is_gso(skb)) {
541 *lso_header_size = skb_transport_offset(skb) + tcp_hdrlen(skb);
542 real_size = CTRL_SIZE + skb_shinfo(skb)->nr_frags * DS_SIZE +
543 ALIGN(*lso_header_size + 4, DS_SIZE);
544 if (unlikely(*lso_header_size != skb_headlen(skb))) {
545 /* We add a segment for the skb linear buffer only if
546 * it contains data */
547 if (*lso_header_size < skb_headlen(skb))
548 real_size += DS_SIZE;
550 if (netif_msg_tx_err(priv))
551 mlx4_warn(mdev, "Non-linear headers\n");
552 dev_kfree_skb_any(skb);
556 if (unlikely(*lso_header_size > MAX_LSO_HDR_SIZE)) {
557 if (netif_msg_tx_err(priv))
558 mlx4_warn(mdev, "LSO header size too big\n");
559 dev_kfree_skb_any(skb);
563 *lso_header_size = 0;
564 if (!is_inline(skb, NULL))
565 real_size = CTRL_SIZE + (skb_shinfo(skb)->nr_frags + 1) * DS_SIZE;
567 real_size = inline_size(skb);
573 static void build_inline_wqe(struct mlx4_en_tx_desc *tx_desc, struct sk_buff *skb,
574 int real_size, u16 *vlan_tag, int tx_ind, void *fragptr)
576 struct mlx4_wqe_inline_seg *inl = &tx_desc->inl;
577 int spc = MLX4_INLINE_ALIGN - CTRL_SIZE - sizeof *inl;
579 if (skb->len <= spc) {
580 inl->byte_count = cpu_to_be32(1 << 31 | skb->len);
581 skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
582 if (skb_shinfo(skb)->nr_frags)
583 memcpy(((void *)(inl + 1)) + skb_headlen(skb), fragptr,
584 skb_shinfo(skb)->frags[0].size);
587 inl->byte_count = cpu_to_be32(1 << 31 | spc);
588 if (skb_headlen(skb) <= spc) {
589 skb_copy_from_linear_data(skb, inl + 1, skb_headlen(skb));
590 if (skb_headlen(skb) < spc) {
591 memcpy(((void *)(inl + 1)) + skb_headlen(skb),
592 fragptr, spc - skb_headlen(skb));
593 fragptr += spc - skb_headlen(skb);
595 inl = (void *) (inl + 1) + spc;
596 memcpy(((void *)(inl + 1)), fragptr, skb->len - spc);
598 skb_copy_from_linear_data(skb, inl + 1, spc);
599 inl = (void *) (inl + 1) + spc;
600 skb_copy_from_linear_data_offset(skb, spc, inl + 1,
601 skb_headlen(skb) - spc);
602 if (skb_shinfo(skb)->nr_frags)
603 memcpy(((void *)(inl + 1)) + skb_headlen(skb) - spc,
604 fragptr, skb_shinfo(skb)->frags[0].size);
608 inl->byte_count = cpu_to_be32(1 << 31 | (skb->len - spc));
610 tx_desc->ctrl.vlan_tag = cpu_to_be16(*vlan_tag);
611 tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN * !!(*vlan_tag);
612 tx_desc->ctrl.fence_size = (real_size / 16) & 0x3f;
615 static int get_vlan_info(struct mlx4_en_priv *priv, struct sk_buff *skb,
620 /* Obtain VLAN information if present */
621 if (priv->vlgrp && vlan_tx_tag_present(skb)) {
622 *vlan_tag = vlan_tx_tag_get(skb);
623 /* Set the Tx ring to use according to vlan priority */
624 tx_ind = priv->tx_prio_map[*vlan_tag >> 13];
632 int mlx4_en_xmit(struct sk_buff *skb, struct net_device *dev)
634 struct mlx4_en_priv *priv = netdev_priv(dev);
635 struct mlx4_en_dev *mdev = priv->mdev;
636 struct mlx4_en_tx_ring *ring;
637 struct mlx4_en_cq *cq;
638 struct mlx4_en_tx_desc *tx_desc;
639 struct mlx4_wqe_data_seg *data;
640 struct skb_frag_struct *frag;
641 struct mlx4_en_tx_info *tx_info;
654 if (unlikely(!skb->len)) {
655 dev_kfree_skb_any(skb);
658 real_size = get_real_size(skb, dev, &lso_header_size);
659 if (unlikely(!real_size))
662 /* Allign descriptor to TXBB size */
663 desc_size = ALIGN(real_size, TXBB_SIZE);
664 nr_txbb = desc_size / TXBB_SIZE;
665 if (unlikely(nr_txbb > MAX_DESC_TXBBS)) {
666 if (netif_msg_tx_err(priv))
667 mlx4_warn(mdev, "Oversized header or SG list\n");
668 dev_kfree_skb_any(skb);
672 tx_ind = get_vlan_info(priv, skb, &vlan_tag);
673 ring = &priv->tx_ring[tx_ind];
675 /* Check available TXBBs And 2K spare for prefetch */
676 if (unlikely(((int)(ring->prod - ring->cons)) >
677 ring->size - HEADROOM - MAX_DESC_TXBBS)) {
678 /* every full Tx ring stops queue.
679 * TODO: implement multi-queue support (per-queue stop) */
680 netif_stop_queue(dev);
682 priv->port_stats.queue_stopped++;
684 /* Use interrupts to find out when queue opened */
685 cq = &priv->tx_cq[tx_ind];
686 mlx4_en_arm_cq(priv, cq);
687 return NETDEV_TX_BUSY;
690 /* Now that we know what Tx ring to use */
691 if (unlikely(!priv->port_up)) {
692 if (netif_msg_tx_err(priv))
693 mlx4_warn(mdev, "xmit: port down!\n");
694 dev_kfree_skb_any(skb);
698 /* Track current inflight packets for performance analysis */
699 AVG_PERF_COUNTER(priv->pstats.inflight_avg,
700 (u32) (ring->prod - ring->cons - 1));
702 /* Packet is good - grab an index and transmit it */
703 index = ring->prod & ring->size_mask;
705 /* See if we have enough space for whole descriptor TXBB for setting
706 * SW ownership on next descriptor; if not, use a bounce buffer. */
707 if (likely(index + nr_txbb <= ring->size))
708 tx_desc = ring->buf + index * TXBB_SIZE;
710 tx_desc = (struct mlx4_en_tx_desc *) ring->bounce_buf;
712 /* Save skb in tx_info ring */
713 tx_info = &ring->tx_info[index];
715 tx_info->nr_txbb = nr_txbb;
717 /* Prepare ctrl segement apart opcode+ownership, which depends on
718 * whether LSO is used */
719 tx_desc->ctrl.vlan_tag = cpu_to_be16(vlan_tag);
720 tx_desc->ctrl.ins_vlan = MLX4_WQE_CTRL_INS_VLAN * !!vlan_tag;
721 tx_desc->ctrl.fence_size = (real_size / 16) & 0x3f;
722 tx_desc->ctrl.srcrb_flags = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE |
723 MLX4_WQE_CTRL_SOLICITED);
724 if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
725 tx_desc->ctrl.srcrb_flags |= cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
726 MLX4_WQE_CTRL_TCP_UDP_CSUM);
727 priv->port_stats.tx_chksum_offload++;
730 /* Handle LSO (TSO) packets */
731 if (lso_header_size) {
732 /* Mark opcode as LSO */
733 op_own = cpu_to_be32(MLX4_OPCODE_LSO | (1 << 6)) |
734 ((ring->prod & ring->size) ?
735 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
737 /* Fill in the LSO prefix */
738 tx_desc->lso.mss_hdr_size = cpu_to_be32(
739 skb_shinfo(skb)->gso_size << 16 | lso_header_size);
742 * note that we already verified that it is linear */
743 memcpy(tx_desc->lso.header, skb->data, lso_header_size);
744 data = ((void *) &tx_desc->lso +
745 ALIGN(lso_header_size + 4, DS_SIZE));
747 priv->port_stats.tso_packets++;
748 i = ((skb->len - lso_header_size) / skb_shinfo(skb)->gso_size) +
749 !!((skb->len - lso_header_size) % skb_shinfo(skb)->gso_size);
750 ring->bytes += skb->len + (i - 1) * lso_header_size;
753 /* Normal (Non LSO) packet */
754 op_own = cpu_to_be32(MLX4_OPCODE_SEND) |
755 ((ring->prod & ring->size) ?
756 cpu_to_be32(MLX4_EN_BIT_DESC_OWN) : 0);
757 data = &tx_desc->data;
758 ring->bytes += max(skb->len, (unsigned int) ETH_ZLEN);
762 AVG_PERF_COUNTER(priv->pstats.tx_pktsz_avg, skb->len);
765 /* valid only for none inline segments */
766 tx_info->data_offset = (void *) data - (void *) tx_desc;
768 tx_info->linear = (lso_header_size < skb_headlen(skb) && !is_inline(skb, NULL)) ? 1 : 0;
769 data += skb_shinfo(skb)->nr_frags + tx_info->linear - 1;
771 if (!is_inline(skb, &fragptr)) {
773 for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
774 frag = &skb_shinfo(skb)->frags[i];
775 dma = pci_map_page(mdev->dev->pdev, frag->page, frag->page_offset,
776 frag->size, PCI_DMA_TODEVICE);
777 data->addr = cpu_to_be64(dma);
778 data->lkey = cpu_to_be32(mdev->mr.key);
780 data->byte_count = cpu_to_be32(frag->size);
784 /* Map linear part */
785 if (tx_info->linear) {
786 dma = pci_map_single(mdev->dev->pdev, skb->data + lso_header_size,
787 skb_headlen(skb) - lso_header_size, PCI_DMA_TODEVICE);
788 data->addr = cpu_to_be64(dma);
789 data->lkey = cpu_to_be32(mdev->mr.key);
791 data->byte_count = cpu_to_be32(skb_headlen(skb) - lso_header_size);
794 build_inline_wqe(tx_desc, skb, real_size, &vlan_tag, tx_ind, fragptr);
796 ring->prod += nr_txbb;
798 /* If we used a bounce buffer then copy descriptor back into place */
799 if (tx_desc == (struct mlx4_en_tx_desc *) ring->bounce_buf)
800 tx_desc = mlx4_en_bounce_to_desc(priv, ring, index, desc_size);
802 /* Run destructor before passing skb to HW */
803 if (likely(!skb_shared(skb)))
806 /* Ensure new descirptor hits memory
807 * before setting ownership of this descriptor to HW */
809 tx_desc->ctrl.owner_opcode = op_own;
813 writel(ring->doorbell_qpn, mdev->uar_map + MLX4_SEND_DOORBELL);
814 dev->trans_start = jiffies;
817 mlx4_en_xmit_poll(priv, tx_ind);