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
34 #include <net/ll_poll.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/rculist.h>
40 #include <linux/if_ether.h>
41 #include <linux/if_vlan.h>
42 #include <linux/vmalloc.h>
46 static int mlx4_alloc_pages(struct mlx4_en_priv *priv,
47 struct mlx4_en_rx_alloc *page_alloc,
48 const struct mlx4_en_frag_info *frag_info,
55 for (order = MLX4_EN_ALLOC_PREFER_ORDER; ;) {
59 gfp |= __GFP_COMP | __GFP_NOWARN;
60 page = alloc_pages(gfp, order);
64 ((PAGE_SIZE << order) < frag_info->frag_size))
67 dma = dma_map_page(priv->ddev, page, 0, PAGE_SIZE << order,
69 if (dma_mapping_error(priv->ddev, dma)) {
73 page_alloc->size = PAGE_SIZE << order;
74 page_alloc->page = page;
75 page_alloc->dma = dma;
76 page_alloc->offset = frag_info->frag_align;
77 /* Not doing get_page() for each frag is a big win
78 * on asymetric workloads.
80 atomic_set(&page->_count, page_alloc->size / frag_info->frag_stride);
84 static int mlx4_en_alloc_frags(struct mlx4_en_priv *priv,
85 struct mlx4_en_rx_desc *rx_desc,
86 struct mlx4_en_rx_alloc *frags,
87 struct mlx4_en_rx_alloc *ring_alloc,
90 struct mlx4_en_rx_alloc page_alloc[MLX4_EN_MAX_RX_FRAGS];
91 const struct mlx4_en_frag_info *frag_info;
96 for (i = 0; i < priv->num_frags; i++) {
97 frag_info = &priv->frag_info[i];
98 page_alloc[i] = ring_alloc[i];
99 page_alloc[i].offset += frag_info->frag_stride;
100 if (page_alloc[i].offset + frag_info->frag_stride <= ring_alloc[i].size)
102 if (mlx4_alloc_pages(priv, &page_alloc[i], frag_info, gfp))
106 for (i = 0; i < priv->num_frags; i++) {
107 frags[i] = ring_alloc[i];
108 dma = ring_alloc[i].dma + ring_alloc[i].offset;
109 ring_alloc[i] = page_alloc[i];
110 rx_desc->data[i].addr = cpu_to_be64(dma);
117 frag_info = &priv->frag_info[i];
118 if (page_alloc[i].page != ring_alloc[i].page) {
119 dma_unmap_page(priv->ddev, page_alloc[i].dma,
120 page_alloc[i].size, PCI_DMA_FROMDEVICE);
121 page = page_alloc[i].page;
122 atomic_set(&page->_count, 1);
129 static void mlx4_en_free_frag(struct mlx4_en_priv *priv,
130 struct mlx4_en_rx_alloc *frags,
133 const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
135 if (frags[i].offset + frag_info->frag_stride > frags[i].size)
136 dma_unmap_page(priv->ddev, frags[i].dma, frags[i].size,
140 put_page(frags[i].page);
143 static int mlx4_en_init_allocator(struct mlx4_en_priv *priv,
144 struct mlx4_en_rx_ring *ring)
147 struct mlx4_en_rx_alloc *page_alloc;
149 for (i = 0; i < priv->num_frags; i++) {
150 const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
152 if (mlx4_alloc_pages(priv, &ring->page_alloc[i],
153 frag_info, GFP_KERNEL))
162 page_alloc = &ring->page_alloc[i];
163 dma_unmap_page(priv->ddev, page_alloc->dma,
164 page_alloc->size, PCI_DMA_FROMDEVICE);
165 page = page_alloc->page;
166 atomic_set(&page->_count, 1);
168 page_alloc->page = NULL;
173 static void mlx4_en_destroy_allocator(struct mlx4_en_priv *priv,
174 struct mlx4_en_rx_ring *ring)
176 struct mlx4_en_rx_alloc *page_alloc;
179 for (i = 0; i < priv->num_frags; i++) {
180 const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
182 page_alloc = &ring->page_alloc[i];
183 en_dbg(DRV, priv, "Freeing allocator:%d count:%d\n",
184 i, page_count(page_alloc->page));
186 dma_unmap_page(priv->ddev, page_alloc->dma,
187 page_alloc->size, PCI_DMA_FROMDEVICE);
188 while (page_alloc->offset + frag_info->frag_stride < page_alloc->size) {
189 put_page(page_alloc->page);
190 page_alloc->offset += frag_info->frag_stride;
192 page_alloc->page = NULL;
196 static void mlx4_en_init_rx_desc(struct mlx4_en_priv *priv,
197 struct mlx4_en_rx_ring *ring, int index)
199 struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
203 /* Set size and memtype fields */
204 for (i = 0; i < priv->num_frags; i++) {
205 rx_desc->data[i].byte_count =
206 cpu_to_be32(priv->frag_info[i].frag_size);
207 rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
210 /* If the number of used fragments does not fill up the ring stride,
211 * remaining (unused) fragments must be padded with null address/size
212 * and a special memory key */
213 possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
214 for (i = priv->num_frags; i < possible_frags; i++) {
215 rx_desc->data[i].byte_count = 0;
216 rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
217 rx_desc->data[i].addr = 0;
221 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
222 struct mlx4_en_rx_ring *ring, int index,
225 struct mlx4_en_rx_desc *rx_desc = ring->buf + (index * ring->stride);
226 struct mlx4_en_rx_alloc *frags = ring->rx_info +
227 (index << priv->log_rx_info);
229 return mlx4_en_alloc_frags(priv, rx_desc, frags, ring->page_alloc, gfp);
232 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
234 *ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
237 static void mlx4_en_free_rx_desc(struct mlx4_en_priv *priv,
238 struct mlx4_en_rx_ring *ring,
241 struct mlx4_en_rx_alloc *frags;
244 frags = ring->rx_info + (index << priv->log_rx_info);
245 for (nr = 0; nr < priv->num_frags; nr++) {
246 en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
247 mlx4_en_free_frag(priv, frags, nr);
251 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
253 struct mlx4_en_rx_ring *ring;
258 for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
259 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
260 ring = &priv->rx_ring[ring_ind];
262 if (mlx4_en_prepare_rx_desc(priv, ring,
265 if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
266 en_err(priv, "Failed to allocate "
267 "enough rx buffers\n");
270 new_size = rounddown_pow_of_two(ring->actual_size);
271 en_warn(priv, "Only %d buffers allocated "
272 "reducing ring size to %d",
273 ring->actual_size, new_size);
284 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
285 ring = &priv->rx_ring[ring_ind];
286 while (ring->actual_size > new_size) {
289 mlx4_en_free_rx_desc(priv, ring, ring->actual_size);
296 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
297 struct mlx4_en_rx_ring *ring)
301 en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
302 ring->cons, ring->prod);
304 /* Unmap and free Rx buffers */
305 BUG_ON((u32) (ring->prod - ring->cons) > ring->actual_size);
306 while (ring->cons != ring->prod) {
307 index = ring->cons & ring->size_mask;
308 en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
309 mlx4_en_free_rx_desc(priv, ring, index);
314 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
315 struct mlx4_en_rx_ring *ring, u32 size, u16 stride)
317 struct mlx4_en_dev *mdev = priv->mdev;
324 ring->size_mask = size - 1;
325 ring->stride = stride;
326 ring->log_stride = ffs(ring->stride) - 1;
327 ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
329 tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
330 sizeof(struct mlx4_en_rx_alloc));
331 ring->rx_info = vmalloc(tmp);
335 en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
338 err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres,
339 ring->buf_size, 2 * PAGE_SIZE);
343 err = mlx4_en_map_buffer(&ring->wqres.buf);
345 en_err(priv, "Failed to map RX buffer\n");
348 ring->buf = ring->wqres.buf.direct.buf;
350 ring->hwtstamp_rx_filter = priv->hwtstamp_config.rx_filter;
355 mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
357 vfree(ring->rx_info);
358 ring->rx_info = NULL;
362 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
364 struct mlx4_en_rx_ring *ring;
368 int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
369 DS_SIZE * priv->num_frags);
371 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
372 ring = &priv->rx_ring[ring_ind];
376 ring->actual_size = 0;
377 ring->cqn = priv->rx_cq[ring_ind].mcq.cqn;
379 ring->stride = stride;
380 if (ring->stride <= TXBB_SIZE)
381 ring->buf += TXBB_SIZE;
383 ring->log_stride = ffs(ring->stride) - 1;
384 ring->buf_size = ring->size * ring->stride;
386 memset(ring->buf, 0, ring->buf_size);
387 mlx4_en_update_rx_prod_db(ring);
389 /* Initialize all descriptors */
390 for (i = 0; i < ring->size; i++)
391 mlx4_en_init_rx_desc(priv, ring, i);
393 /* Initialize page allocators */
394 err = mlx4_en_init_allocator(priv, ring);
396 en_err(priv, "Failed initializing ring allocator\n");
397 if (ring->stride <= TXBB_SIZE)
398 ring->buf -= TXBB_SIZE;
403 err = mlx4_en_fill_rx_buffers(priv);
407 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
408 ring = &priv->rx_ring[ring_ind];
410 ring->size_mask = ring->actual_size - 1;
411 mlx4_en_update_rx_prod_db(ring);
417 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
418 mlx4_en_free_rx_buf(priv, &priv->rx_ring[ring_ind]);
420 ring_ind = priv->rx_ring_num - 1;
422 while (ring_ind >= 0) {
423 if (priv->rx_ring[ring_ind].stride <= TXBB_SIZE)
424 priv->rx_ring[ring_ind].buf -= TXBB_SIZE;
425 mlx4_en_destroy_allocator(priv, &priv->rx_ring[ring_ind]);
431 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
432 struct mlx4_en_rx_ring *ring, u32 size, u16 stride)
434 struct mlx4_en_dev *mdev = priv->mdev;
436 mlx4_en_unmap_buffer(&ring->wqres.buf);
437 mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
438 vfree(ring->rx_info);
439 ring->rx_info = NULL;
440 #ifdef CONFIG_RFS_ACCEL
441 mlx4_en_cleanup_filters(priv, ring);
445 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
446 struct mlx4_en_rx_ring *ring)
448 mlx4_en_free_rx_buf(priv, ring);
449 if (ring->stride <= TXBB_SIZE)
450 ring->buf -= TXBB_SIZE;
451 mlx4_en_destroy_allocator(priv, ring);
455 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
456 struct mlx4_en_rx_desc *rx_desc,
457 struct mlx4_en_rx_alloc *frags,
461 struct skb_frag_struct *skb_frags_rx = skb_shinfo(skb)->frags;
462 struct mlx4_en_frag_info *frag_info;
466 /* Collect used fragments while replacing them in the HW descriptors */
467 for (nr = 0; nr < priv->num_frags; nr++) {
468 frag_info = &priv->frag_info[nr];
469 if (length <= frag_info->frag_prefix_size)
474 dma = be64_to_cpu(rx_desc->data[nr].addr);
475 dma_sync_single_for_cpu(priv->ddev, dma, frag_info->frag_size,
478 /* Save page reference in skb */
479 __skb_frag_set_page(&skb_frags_rx[nr], frags[nr].page);
480 skb_frag_size_set(&skb_frags_rx[nr], frag_info->frag_size);
481 skb_frags_rx[nr].page_offset = frags[nr].offset;
482 skb->truesize += frag_info->frag_stride;
483 frags[nr].page = NULL;
485 /* Adjust size of last fragment to match actual length */
487 skb_frag_size_set(&skb_frags_rx[nr - 1],
488 length - priv->frag_info[nr - 1].frag_prefix_size);
494 __skb_frag_unref(&skb_frags_rx[nr]);
500 static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
501 struct mlx4_en_rx_desc *rx_desc,
502 struct mlx4_en_rx_alloc *frags,
510 skb = netdev_alloc_skb(priv->dev, SMALL_PACKET_SIZE + NET_IP_ALIGN);
512 en_dbg(RX_ERR, priv, "Failed allocating skb\n");
515 skb_reserve(skb, NET_IP_ALIGN);
518 /* Get pointer to first fragment so we could copy the headers into the
519 * (linear part of the) skb */
520 va = page_address(frags[0].page) + frags[0].offset;
522 if (length <= SMALL_PACKET_SIZE) {
523 /* We are copying all relevant data to the skb - temporarily
524 * sync buffers for the copy */
525 dma = be64_to_cpu(rx_desc->data[0].addr);
526 dma_sync_single_for_cpu(priv->ddev, dma, length,
528 skb_copy_to_linear_data(skb, va, length);
531 /* Move relevant fragments to skb */
532 used_frags = mlx4_en_complete_rx_desc(priv, rx_desc, frags,
534 if (unlikely(!used_frags)) {
538 skb_shinfo(skb)->nr_frags = used_frags;
540 /* Copy headers into the skb linear buffer */
541 memcpy(skb->data, va, HEADER_COPY_SIZE);
542 skb->tail += HEADER_COPY_SIZE;
544 /* Skip headers in first fragment */
545 skb_shinfo(skb)->frags[0].page_offset += HEADER_COPY_SIZE;
547 /* Adjust size of first fragment */
548 skb_frag_size_sub(&skb_shinfo(skb)->frags[0], HEADER_COPY_SIZE);
549 skb->data_len = length - HEADER_COPY_SIZE;
554 static void validate_loopback(struct mlx4_en_priv *priv, struct sk_buff *skb)
557 int offset = ETH_HLEN;
559 for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++, offset++) {
560 if (*(skb->data + offset) != (unsigned char) (i & 0xff))
564 priv->loopback_ok = 1;
567 dev_kfree_skb_any(skb);
570 static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
571 struct mlx4_en_rx_ring *ring)
573 int index = ring->prod & ring->size_mask;
575 while ((u32) (ring->prod - ring->cons) < ring->actual_size) {
576 if (mlx4_en_prepare_rx_desc(priv, ring, index, GFP_ATOMIC))
579 index = ring->prod & ring->size_mask;
583 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
585 struct mlx4_en_priv *priv = netdev_priv(dev);
586 struct mlx4_en_dev *mdev = priv->mdev;
587 struct mlx4_cqe *cqe;
588 struct mlx4_en_rx_ring *ring = &priv->rx_ring[cq->ring];
589 struct mlx4_en_rx_alloc *frags;
590 struct mlx4_en_rx_desc *rx_desc;
597 int factor = priv->cqe_factor;
603 /* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
604 * descriptor offset can be deduced from the CQE index instead of
605 * reading 'cqe->index' */
606 index = cq->mcq.cons_index & ring->size_mask;
607 cqe = &cq->buf[(index << factor) + factor];
609 /* Process all completed CQEs */
610 while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
611 cq->mcq.cons_index & cq->size)) {
613 frags = ring->rx_info + (index << priv->log_rx_info);
614 rx_desc = ring->buf + (index << ring->log_stride);
617 * make sure we read the CQE after we read the ownership bit
621 /* Drop packet on bad receive or bad checksum */
622 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
623 MLX4_CQE_OPCODE_ERROR)) {
624 en_err(priv, "CQE completed in error - vendor "
625 "syndrom:%d syndrom:%d\n",
626 ((struct mlx4_err_cqe *) cqe)->vendor_err_syndrome,
627 ((struct mlx4_err_cqe *) cqe)->syndrome);
630 if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
631 en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
635 /* Check if we need to drop the packet if SRIOV is not enabled
636 * and not performing the selftest or flb disabled
638 if (priv->flags & MLX4_EN_FLAG_RX_FILTER_NEEDED) {
641 /* Get pointer to first fragment since we haven't
642 * skb yet and cast it to ethhdr struct
644 dma = be64_to_cpu(rx_desc->data[0].addr);
645 dma_sync_single_for_cpu(priv->ddev, dma, sizeof(*ethh),
647 ethh = (struct ethhdr *)(page_address(frags[0].page) +
650 if (is_multicast_ether_addr(ethh->h_dest)) {
651 struct mlx4_mac_entry *entry;
652 struct hlist_head *bucket;
653 unsigned int mac_hash;
655 /* Drop the packet, since HW loopback-ed it */
656 mac_hash = ethh->h_source[MLX4_EN_MAC_HASH_IDX];
657 bucket = &priv->mac_hash[mac_hash];
659 hlist_for_each_entry_rcu(entry, bucket, hlist) {
660 if (ether_addr_equal_64bits(entry->mac,
671 * Packet is OK - process it.
673 length = be32_to_cpu(cqe->byte_cnt);
674 length -= ring->fcs_del;
675 ring->bytes += length;
678 if (likely(dev->features & NETIF_F_RXCSUM)) {
679 if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
680 (cqe->checksum == cpu_to_be16(0xffff))) {
682 /* This packet is eligible for GRO if it is:
683 * - DIX Ethernet (type interpretation)
685 * - without IP options
686 * - not an IP fragment
687 * - no LLS polling in progress
689 if (!mlx4_en_cq_ll_polling(cq) &&
690 (dev->features & NETIF_F_GRO)) {
691 struct sk_buff *gro_skb = napi_get_frags(&cq->napi);
695 nr = mlx4_en_complete_rx_desc(priv,
696 rx_desc, frags, gro_skb,
701 skb_shinfo(gro_skb)->nr_frags = nr;
702 gro_skb->len = length;
703 gro_skb->data_len = length;
704 gro_skb->ip_summed = CHECKSUM_UNNECESSARY;
706 if ((cqe->vlan_my_qpn &
707 cpu_to_be32(MLX4_CQE_VLAN_PRESENT_MASK)) &&
708 (dev->features & NETIF_F_HW_VLAN_CTAG_RX)) {
709 u16 vid = be16_to_cpu(cqe->sl_vid);
711 __vlan_hwaccel_put_tag(gro_skb, htons(ETH_P_8021Q), vid);
714 if (dev->features & NETIF_F_RXHASH)
715 gro_skb->rxhash = be32_to_cpu(cqe->immed_rss_invalid);
717 skb_record_rx_queue(gro_skb, cq->ring);
719 if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
720 timestamp = mlx4_en_get_cqe_ts(cqe);
721 mlx4_en_fill_hwtstamps(mdev,
722 skb_hwtstamps(gro_skb),
726 napi_gro_frags(&cq->napi);
730 /* GRO not possible, complete processing here */
731 ip_summed = CHECKSUM_UNNECESSARY;
733 ip_summed = CHECKSUM_NONE;
737 ip_summed = CHECKSUM_NONE;
741 skb = mlx4_en_rx_skb(priv, rx_desc, frags, length);
743 priv->stats.rx_dropped++;
747 if (unlikely(priv->validate_loopback)) {
748 validate_loopback(priv, skb);
752 skb->ip_summed = ip_summed;
753 skb->protocol = eth_type_trans(skb, dev);
754 skb_record_rx_queue(skb, cq->ring);
756 if (dev->features & NETIF_F_RXHASH)
757 skb->rxhash = be32_to_cpu(cqe->immed_rss_invalid);
759 if ((be32_to_cpu(cqe->vlan_my_qpn) &
760 MLX4_CQE_VLAN_PRESENT_MASK) &&
761 (dev->features & NETIF_F_HW_VLAN_CTAG_RX))
762 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), be16_to_cpu(cqe->sl_vid));
764 if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
765 timestamp = mlx4_en_get_cqe_ts(cqe);
766 mlx4_en_fill_hwtstamps(mdev, skb_hwtstamps(skb),
770 skb_mark_ll(skb, &cq->napi);
772 /* Push it up the stack */
773 netif_receive_skb(skb);
776 for (nr = 0; nr < priv->num_frags; nr++)
777 mlx4_en_free_frag(priv, frags, nr);
779 ++cq->mcq.cons_index;
780 index = (cq->mcq.cons_index) & ring->size_mask;
781 cqe = &cq->buf[(index << factor) + factor];
782 if (++polled == budget)
787 AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
788 mlx4_cq_set_ci(&cq->mcq);
789 wmb(); /* ensure HW sees CQ consumer before we post new buffers */
790 ring->cons = cq->mcq.cons_index;
791 mlx4_en_refill_rx_buffers(priv, ring);
792 mlx4_en_update_rx_prod_db(ring);
797 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
799 struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
800 struct mlx4_en_priv *priv = netdev_priv(cq->dev);
803 napi_schedule(&cq->napi);
805 mlx4_en_arm_cq(priv, cq);
808 /* Rx CQ polling - called by NAPI */
809 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
811 struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
812 struct net_device *dev = cq->dev;
813 struct mlx4_en_priv *priv = netdev_priv(dev);
816 if (!mlx4_en_cq_lock_napi(cq))
819 done = mlx4_en_process_rx_cq(dev, cq, budget);
821 mlx4_en_cq_unlock_napi(cq);
823 /* If we used up all the quota - we're probably not done yet... */
825 INC_PERF_COUNTER(priv->pstats.napi_quota);
829 mlx4_en_arm_cq(priv, cq);
834 static const int frag_sizes[] = {
841 void mlx4_en_calc_rx_buf(struct net_device *dev)
843 struct mlx4_en_priv *priv = netdev_priv(dev);
844 int eff_mtu = dev->mtu + ETH_HLEN + VLAN_HLEN + ETH_LLC_SNAP_SIZE;
848 while (buf_size < eff_mtu) {
849 priv->frag_info[i].frag_size =
850 (eff_mtu > buf_size + frag_sizes[i]) ?
851 frag_sizes[i] : eff_mtu - buf_size;
852 priv->frag_info[i].frag_prefix_size = buf_size;
854 priv->frag_info[i].frag_align = NET_IP_ALIGN;
855 priv->frag_info[i].frag_stride =
856 ALIGN(frag_sizes[i] + NET_IP_ALIGN, SMP_CACHE_BYTES);
858 priv->frag_info[i].frag_align = 0;
859 priv->frag_info[i].frag_stride =
860 ALIGN(frag_sizes[i], SMP_CACHE_BYTES);
862 buf_size += priv->frag_info[i].frag_size;
867 priv->rx_skb_size = eff_mtu;
868 priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct mlx4_en_rx_alloc));
870 en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d "
871 "num_frags:%d):\n", eff_mtu, priv->num_frags);
872 for (i = 0; i < priv->num_frags; i++) {
874 " frag:%d - size:%d prefix:%d align:%d stride:%d\n",
876 priv->frag_info[i].frag_size,
877 priv->frag_info[i].frag_prefix_size,
878 priv->frag_info[i].frag_align,
879 priv->frag_info[i].frag_stride);
883 /* RSS related functions */
885 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
886 struct mlx4_en_rx_ring *ring,
887 enum mlx4_qp_state *state,
890 struct mlx4_en_dev *mdev = priv->mdev;
891 struct mlx4_qp_context *context;
894 context = kmalloc(sizeof(*context), GFP_KERNEL);
898 err = mlx4_qp_alloc(mdev->dev, qpn, qp);
900 en_err(priv, "Failed to allocate qp #%x\n", qpn);
903 qp->event = mlx4_en_sqp_event;
905 memset(context, 0, sizeof *context);
906 mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
907 qpn, ring->cqn, -1, context);
908 context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
910 /* Cancel FCS removal if FW allows */
911 if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
912 context->param3 |= cpu_to_be32(1 << 29);
913 ring->fcs_del = ETH_FCS_LEN;
917 err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
919 mlx4_qp_remove(mdev->dev, qp);
920 mlx4_qp_free(mdev->dev, qp);
922 mlx4_en_update_rx_prod_db(ring);
928 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
933 err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn);
935 en_err(priv, "Failed reserving drop qpn\n");
938 err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp);
940 en_err(priv, "Failed allocating drop qp\n");
941 mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
948 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
952 qpn = priv->drop_qp.qpn;
953 mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
954 mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
955 mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
958 /* Allocate rx qp's and configure them according to rss map */
959 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
961 struct mlx4_en_dev *mdev = priv->mdev;
962 struct mlx4_en_rss_map *rss_map = &priv->rss_map;
963 struct mlx4_qp_context context;
964 struct mlx4_rss_context *rss_context;
967 u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
972 static const u32 rsskey[10] = { 0xD181C62C, 0xF7F4DB5B, 0x1983A2FC,
973 0x943E1ADB, 0xD9389E6B, 0xD1039C2C, 0xA74499AD,
974 0x593D56D9, 0xF3253C06, 0x2ADC1FFC};
976 en_dbg(DRV, priv, "Configuring rss steering\n");
977 err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
981 en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
985 for (i = 0; i < priv->rx_ring_num; i++) {
986 qpn = rss_map->base_qpn + i;
987 err = mlx4_en_config_rss_qp(priv, qpn, &priv->rx_ring[i],
996 /* Configure RSS indirection qp */
997 err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp);
999 en_err(priv, "Failed to allocate RSS indirection QP\n");
1002 rss_map->indir_qp.event = mlx4_en_sqp_event;
1003 mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1004 priv->rx_ring[0].cqn, -1, &context);
1006 if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
1007 rss_rings = priv->rx_ring_num;
1009 rss_rings = priv->prof->rss_rings;
1011 ptr = ((void *) &context) + offsetof(struct mlx4_qp_context, pri_path)
1012 + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
1014 rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
1015 (rss_map->base_qpn));
1016 rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1017 if (priv->mdev->profile.udp_rss) {
1018 rss_mask |= MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
1019 rss_context->base_qpn_udp = rss_context->default_qpn;
1021 rss_context->flags = rss_mask;
1022 rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1023 for (i = 0; i < 10; i++)
1024 rss_context->rss_key[i] = cpu_to_be32(rsskey[i]);
1026 err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1027 &rss_map->indir_qp, &rss_map->indir_state);
1034 mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1035 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1036 mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1037 mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1039 for (i = 0; i < good_qps; i++) {
1040 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1041 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1042 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1043 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1045 mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1049 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1051 struct mlx4_en_dev *mdev = priv->mdev;
1052 struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1055 mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1056 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1057 mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1058 mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1060 for (i = 0; i < priv->rx_ring_num; i++) {
1061 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1062 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1063 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1064 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1066 mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);