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/busy_poll.h>
35 #include <linux/bpf.h>
36 #include <linux/mlx4/cq.h>
37 #include <linux/slab.h>
38 #include <linux/mlx4/qp.h>
39 #include <linux/skbuff.h>
40 #include <linux/rculist.h>
41 #include <linux/if_ether.h>
42 #include <linux/if_vlan.h>
43 #include <linux/vmalloc.h>
44 #include <linux/irq.h>
46 #if IS_ENABLED(CONFIG_IPV6)
47 #include <net/ip6_checksum.h>
52 static int mlx4_alloc_pages(struct mlx4_en_priv *priv,
53 struct mlx4_en_rx_alloc *page_alloc,
54 const struct mlx4_en_frag_info *frag_info,
61 for (order = frag_info->order; ;) {
65 gfp |= __GFP_COMP | __GFP_NOWARN | __GFP_NOMEMALLOC;
66 page = alloc_pages(gfp, order);
70 ((PAGE_SIZE << order) < frag_info->frag_size))
73 dma = dma_map_page(priv->ddev, page, 0, PAGE_SIZE << order,
75 if (unlikely(dma_mapping_error(priv->ddev, dma))) {
79 page_alloc->page_size = PAGE_SIZE << order;
80 page_alloc->page = page;
81 page_alloc->dma = dma;
82 page_alloc->page_offset = 0;
83 /* Not doing get_page() for each frag is a big win
84 * on asymetric workloads. Note we can not use atomic_set().
86 page_ref_add(page, page_alloc->page_size / frag_info->frag_stride - 1);
90 static int mlx4_en_alloc_frags(struct mlx4_en_priv *priv,
91 struct mlx4_en_rx_desc *rx_desc,
92 struct mlx4_en_rx_alloc *frags,
93 struct mlx4_en_rx_alloc *ring_alloc,
96 struct mlx4_en_rx_alloc page_alloc[MLX4_EN_MAX_RX_FRAGS];
97 const struct mlx4_en_frag_info *frag_info;
101 for (i = 0; i < priv->num_frags; i++) {
102 frag_info = &priv->frag_info[i];
103 page_alloc[i] = ring_alloc[i];
104 page_alloc[i].page_offset += frag_info->frag_stride;
106 if (page_alloc[i].page_offset + frag_info->frag_stride <=
107 ring_alloc[i].page_size)
110 if (unlikely(mlx4_alloc_pages(priv, &page_alloc[i],
115 for (i = 0; i < priv->num_frags; i++) {
116 frags[i] = ring_alloc[i];
117 frags[i].page_offset += priv->frag_info[i].rx_headroom;
118 rx_desc->data[i].addr = cpu_to_be64(frags[i].dma +
119 frags[i].page_offset);
120 ring_alloc[i] = page_alloc[i];
127 if (page_alloc[i].page != ring_alloc[i].page) {
128 dma_unmap_page(priv->ddev, page_alloc[i].dma,
129 page_alloc[i].page_size,
130 priv->frag_info[i].dma_dir);
131 page = page_alloc[i].page;
132 /* Revert changes done by mlx4_alloc_pages */
133 page_ref_sub(page, page_alloc[i].page_size /
134 priv->frag_info[i].frag_stride - 1);
141 static void mlx4_en_free_frag(struct mlx4_en_priv *priv,
142 struct mlx4_en_rx_alloc *frags,
145 const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
146 u32 next_frag_end = frags[i].page_offset + 2 * frag_info->frag_stride;
149 if (next_frag_end > frags[i].page_size)
150 dma_unmap_page(priv->ddev, frags[i].dma, frags[i].page_size,
154 put_page(frags[i].page);
157 static int mlx4_en_init_allocator(struct mlx4_en_priv *priv,
158 struct mlx4_en_rx_ring *ring)
161 struct mlx4_en_rx_alloc *page_alloc;
163 for (i = 0; i < priv->num_frags; i++) {
164 const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
166 if (mlx4_alloc_pages(priv, &ring->page_alloc[i],
167 frag_info, GFP_KERNEL | __GFP_COLD))
170 en_dbg(DRV, priv, " frag %d allocator: - size:%d frags:%d\n",
171 i, ring->page_alloc[i].page_size,
172 page_ref_count(ring->page_alloc[i].page));
180 page_alloc = &ring->page_alloc[i];
181 dma_unmap_page(priv->ddev, page_alloc->dma,
182 page_alloc->page_size,
183 priv->frag_info[i].dma_dir);
184 page = page_alloc->page;
185 /* Revert changes done by mlx4_alloc_pages */
186 page_ref_sub(page, page_alloc->page_size /
187 priv->frag_info[i].frag_stride - 1);
189 page_alloc->page = NULL;
194 static void mlx4_en_destroy_allocator(struct mlx4_en_priv *priv,
195 struct mlx4_en_rx_ring *ring)
197 struct mlx4_en_rx_alloc *page_alloc;
200 for (i = 0; i < priv->num_frags; i++) {
201 const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
203 page_alloc = &ring->page_alloc[i];
204 en_dbg(DRV, priv, "Freeing allocator:%d count:%d\n",
205 i, page_count(page_alloc->page));
207 dma_unmap_page(priv->ddev, page_alloc->dma,
208 page_alloc->page_size, frag_info->dma_dir);
209 while (page_alloc->page_offset + frag_info->frag_stride <
210 page_alloc->page_size) {
211 put_page(page_alloc->page);
212 page_alloc->page_offset += frag_info->frag_stride;
214 page_alloc->page = NULL;
218 static void mlx4_en_init_rx_desc(struct mlx4_en_priv *priv,
219 struct mlx4_en_rx_ring *ring, int index)
221 struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
225 /* Set size and memtype fields */
226 for (i = 0; i < priv->num_frags; i++) {
227 rx_desc->data[i].byte_count =
228 cpu_to_be32(priv->frag_info[i].frag_size);
229 rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
232 /* If the number of used fragments does not fill up the ring stride,
233 * remaining (unused) fragments must be padded with null address/size
234 * and a special memory key */
235 possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
236 for (i = priv->num_frags; i < possible_frags; i++) {
237 rx_desc->data[i].byte_count = 0;
238 rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
239 rx_desc->data[i].addr = 0;
243 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
244 struct mlx4_en_rx_ring *ring, int index,
247 struct mlx4_en_rx_desc *rx_desc = ring->buf + (index * ring->stride);
248 struct mlx4_en_rx_alloc *frags = ring->rx_info +
249 (index << priv->log_rx_info);
251 if (ring->page_cache.index > 0) {
252 frags[0] = ring->page_cache.buf[--ring->page_cache.index];
253 rx_desc->data[0].addr = cpu_to_be64(frags[0].dma +
254 frags[0].page_offset);
258 return mlx4_en_alloc_frags(priv, rx_desc, frags, ring->page_alloc, gfp);
261 static inline bool mlx4_en_is_ring_empty(struct mlx4_en_rx_ring *ring)
263 return ring->prod == ring->cons;
266 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
268 *ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
271 static void mlx4_en_free_rx_desc(struct mlx4_en_priv *priv,
272 struct mlx4_en_rx_ring *ring,
275 struct mlx4_en_rx_alloc *frags;
278 frags = ring->rx_info + (index << priv->log_rx_info);
279 for (nr = 0; nr < priv->num_frags; nr++) {
280 en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
281 mlx4_en_free_frag(priv, frags, nr);
285 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
287 struct mlx4_en_rx_ring *ring;
292 for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
293 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
294 ring = priv->rx_ring[ring_ind];
296 if (mlx4_en_prepare_rx_desc(priv, ring,
298 GFP_KERNEL | __GFP_COLD)) {
299 if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
300 en_err(priv, "Failed to allocate enough rx buffers\n");
303 new_size = rounddown_pow_of_two(ring->actual_size);
304 en_warn(priv, "Only %d buffers allocated reducing ring size to %d\n",
305 ring->actual_size, new_size);
316 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
317 ring = priv->rx_ring[ring_ind];
318 while (ring->actual_size > new_size) {
321 mlx4_en_free_rx_desc(priv, ring, ring->actual_size);
328 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
329 struct mlx4_en_rx_ring *ring)
333 en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
334 ring->cons, ring->prod);
336 /* Unmap and free Rx buffers */
337 while (!mlx4_en_is_ring_empty(ring)) {
338 index = ring->cons & ring->size_mask;
339 en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
340 mlx4_en_free_rx_desc(priv, ring, index);
345 void mlx4_en_set_num_rx_rings(struct mlx4_en_dev *mdev)
350 struct mlx4_dev *dev = mdev->dev;
352 mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_ETH) {
353 num_of_eqs = max_t(int, MIN_RX_RINGS,
355 mlx4_get_eqs_per_port(mdev->dev, i),
358 num_rx_rings = mlx4_low_memory_profile() ? MIN_RX_RINGS :
359 min_t(int, num_of_eqs,
360 netif_get_num_default_rss_queues());
361 mdev->profile.prof[i].rx_ring_num =
362 rounddown_pow_of_two(num_rx_rings);
366 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
367 struct mlx4_en_rx_ring **pring,
368 u32 size, u16 stride, int node)
370 struct mlx4_en_dev *mdev = priv->mdev;
371 struct mlx4_en_rx_ring *ring;
375 ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
377 ring = kzalloc(sizeof(*ring), GFP_KERNEL);
379 en_err(priv, "Failed to allocate RX ring structure\n");
387 ring->size_mask = size - 1;
388 ring->stride = stride;
389 ring->log_stride = ffs(ring->stride) - 1;
390 ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
392 tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
393 sizeof(struct mlx4_en_rx_alloc));
394 ring->rx_info = vmalloc_node(tmp, node);
395 if (!ring->rx_info) {
396 ring->rx_info = vmalloc(tmp);
397 if (!ring->rx_info) {
403 en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
406 /* Allocate HW buffers on provided NUMA node */
407 set_dev_node(&mdev->dev->persist->pdev->dev, node);
408 err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
409 set_dev_node(&mdev->dev->persist->pdev->dev, mdev->dev->numa_node);
413 ring->buf = ring->wqres.buf.direct.buf;
415 ring->hwtstamp_rx_filter = priv->hwtstamp_config.rx_filter;
421 vfree(ring->rx_info);
422 ring->rx_info = NULL;
430 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
432 struct mlx4_en_rx_ring *ring;
436 int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
437 DS_SIZE * priv->num_frags);
439 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
440 ring = priv->rx_ring[ring_ind];
444 ring->actual_size = 0;
445 ring->cqn = priv->rx_cq[ring_ind]->mcq.cqn;
447 ring->stride = stride;
448 if (ring->stride <= TXBB_SIZE) {
449 /* Stamp first unused send wqe */
450 __be32 *ptr = (__be32 *)ring->buf;
451 __be32 stamp = cpu_to_be32(1 << STAMP_SHIFT);
453 /* Move pointer to start of rx section */
454 ring->buf += TXBB_SIZE;
457 ring->log_stride = ffs(ring->stride) - 1;
458 ring->buf_size = ring->size * ring->stride;
460 memset(ring->buf, 0, ring->buf_size);
461 mlx4_en_update_rx_prod_db(ring);
463 /* Initialize all descriptors */
464 for (i = 0; i < ring->size; i++)
465 mlx4_en_init_rx_desc(priv, ring, i);
467 /* Initialize page allocators */
468 err = mlx4_en_init_allocator(priv, ring);
470 en_err(priv, "Failed initializing ring allocator\n");
471 if (ring->stride <= TXBB_SIZE)
472 ring->buf -= TXBB_SIZE;
477 err = mlx4_en_fill_rx_buffers(priv);
481 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
482 ring = priv->rx_ring[ring_ind];
484 ring->size_mask = ring->actual_size - 1;
485 mlx4_en_update_rx_prod_db(ring);
491 for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
492 mlx4_en_free_rx_buf(priv, priv->rx_ring[ring_ind]);
494 ring_ind = priv->rx_ring_num - 1;
496 while (ring_ind >= 0) {
497 if (priv->rx_ring[ring_ind]->stride <= TXBB_SIZE)
498 priv->rx_ring[ring_ind]->buf -= TXBB_SIZE;
499 mlx4_en_destroy_allocator(priv, priv->rx_ring[ring_ind]);
505 /* We recover from out of memory by scheduling our napi poll
506 * function (mlx4_en_process_cq), which tries to allocate
507 * all missing RX buffers (call to mlx4_en_refill_rx_buffers).
509 void mlx4_en_recover_from_oom(struct mlx4_en_priv *priv)
516 for (ring = 0; ring < priv->rx_ring_num; ring++) {
517 if (mlx4_en_is_ring_empty(priv->rx_ring[ring])) {
519 napi_reschedule(&priv->rx_cq[ring]->napi);
525 /* When the rx ring is running in page-per-packet mode, a released frame can go
526 * directly into a small cache, to avoid unmapping or touching the page
527 * allocator. In bpf prog performance scenarios, buffers are either forwarded
528 * or dropped, never converted to skbs, so every page can come directly from
529 * this cache when it is sized to be a multiple of the napi budget.
531 bool mlx4_en_rx_recycle(struct mlx4_en_rx_ring *ring,
532 struct mlx4_en_rx_alloc *frame)
534 struct mlx4_en_page_cache *cache = &ring->page_cache;
536 if (cache->index >= MLX4_EN_CACHE_SIZE)
539 cache->buf[cache->index++] = *frame;
543 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
544 struct mlx4_en_rx_ring **pring,
545 u32 size, u16 stride)
547 struct mlx4_en_dev *mdev = priv->mdev;
548 struct mlx4_en_rx_ring *ring = *pring;
549 struct bpf_prog *old_prog;
551 old_prog = rcu_dereference_protected(
553 lockdep_is_held(&mdev->state_lock));
555 bpf_prog_put(old_prog);
556 mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
557 vfree(ring->rx_info);
558 ring->rx_info = NULL;
563 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
564 struct mlx4_en_rx_ring *ring)
568 for (i = 0; i < ring->page_cache.index; i++) {
569 struct mlx4_en_rx_alloc *frame = &ring->page_cache.buf[i];
571 dma_unmap_page(priv->ddev, frame->dma, frame->page_size,
572 priv->frag_info[0].dma_dir);
573 put_page(frame->page);
575 ring->page_cache.index = 0;
576 mlx4_en_free_rx_buf(priv, ring);
577 if (ring->stride <= TXBB_SIZE)
578 ring->buf -= TXBB_SIZE;
579 mlx4_en_destroy_allocator(priv, ring);
583 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
584 struct mlx4_en_rx_desc *rx_desc,
585 struct mlx4_en_rx_alloc *frags,
589 struct skb_frag_struct *skb_frags_rx = skb_shinfo(skb)->frags;
590 struct mlx4_en_frag_info *frag_info;
594 /* Collect used fragments while replacing them in the HW descriptors */
595 for (nr = 0; nr < priv->num_frags; nr++) {
596 frag_info = &priv->frag_info[nr];
597 if (length <= frag_info->frag_prefix_size)
599 if (unlikely(!frags[nr].page))
602 dma = be64_to_cpu(rx_desc->data[nr].addr);
603 dma_sync_single_for_cpu(priv->ddev, dma, frag_info->frag_size,
606 /* Save page reference in skb */
607 __skb_frag_set_page(&skb_frags_rx[nr], frags[nr].page);
608 skb_frag_size_set(&skb_frags_rx[nr], frag_info->frag_size);
609 skb_frags_rx[nr].page_offset = frags[nr].page_offset;
610 skb->truesize += frag_info->frag_stride;
611 frags[nr].page = NULL;
613 /* Adjust size of last fragment to match actual length */
615 skb_frag_size_set(&skb_frags_rx[nr - 1],
616 length - priv->frag_info[nr - 1].frag_prefix_size);
622 __skb_frag_unref(&skb_frags_rx[nr]);
628 static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
629 struct mlx4_en_rx_desc *rx_desc,
630 struct mlx4_en_rx_alloc *frags,
638 skb = netdev_alloc_skb(priv->dev, SMALL_PACKET_SIZE + NET_IP_ALIGN);
639 if (unlikely(!skb)) {
640 en_dbg(RX_ERR, priv, "Failed allocating skb\n");
643 skb_reserve(skb, NET_IP_ALIGN);
646 /* Get pointer to first fragment so we could copy the headers into the
647 * (linear part of the) skb */
648 va = page_address(frags[0].page) + frags[0].page_offset;
650 if (length <= SMALL_PACKET_SIZE) {
651 /* We are copying all relevant data to the skb - temporarily
652 * sync buffers for the copy */
653 dma = be64_to_cpu(rx_desc->data[0].addr);
654 dma_sync_single_for_cpu(priv->ddev, dma, length,
656 skb_copy_to_linear_data(skb, va, length);
659 unsigned int pull_len;
661 /* Move relevant fragments to skb */
662 used_frags = mlx4_en_complete_rx_desc(priv, rx_desc, frags,
664 if (unlikely(!used_frags)) {
668 skb_shinfo(skb)->nr_frags = used_frags;
670 pull_len = eth_get_headlen(va, SMALL_PACKET_SIZE);
671 /* Copy headers into the skb linear buffer */
672 memcpy(skb->data, va, pull_len);
673 skb->tail += pull_len;
675 /* Skip headers in first fragment */
676 skb_shinfo(skb)->frags[0].page_offset += pull_len;
678 /* Adjust size of first fragment */
679 skb_frag_size_sub(&skb_shinfo(skb)->frags[0], pull_len);
680 skb->data_len = length - pull_len;
685 static void validate_loopback(struct mlx4_en_priv *priv, struct sk_buff *skb)
688 int offset = ETH_HLEN;
690 for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++, offset++) {
691 if (*(skb->data + offset) != (unsigned char) (i & 0xff))
695 priv->loopback_ok = 1;
698 dev_kfree_skb_any(skb);
701 static bool mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
702 struct mlx4_en_rx_ring *ring)
704 u32 missing = ring->actual_size - (ring->prod - ring->cons);
706 /* Try to batch allocations, but not too much. */
710 if (mlx4_en_prepare_rx_desc(priv, ring,
711 ring->prod & ring->size_mask,
712 GFP_ATOMIC | __GFP_COLD))
720 /* When hardware doesn't strip the vlan, we need to calculate the checksum
721 * over it and add it to the hardware's checksum calculation
723 static inline __wsum get_fixed_vlan_csum(__wsum hw_checksum,
724 struct vlan_hdr *vlanh)
726 return csum_add(hw_checksum, *(__wsum *)vlanh);
729 /* Although the stack expects checksum which doesn't include the pseudo
730 * header, the HW adds it. To address that, we are subtracting the pseudo
731 * header checksum from the checksum value provided by the HW.
733 static void get_fixed_ipv4_csum(__wsum hw_checksum, struct sk_buff *skb,
736 __u16 length_for_csum = 0;
737 __wsum csum_pseudo_header = 0;
739 length_for_csum = (be16_to_cpu(iph->tot_len) - (iph->ihl << 2));
740 csum_pseudo_header = csum_tcpudp_nofold(iph->saddr, iph->daddr,
741 length_for_csum, iph->protocol, 0);
742 skb->csum = csum_sub(hw_checksum, csum_pseudo_header);
745 #if IS_ENABLED(CONFIG_IPV6)
746 /* In IPv6 packets, besides subtracting the pseudo header checksum,
747 * we also compute/add the IP header checksum which
748 * is not added by the HW.
750 static int get_fixed_ipv6_csum(__wsum hw_checksum, struct sk_buff *skb,
751 struct ipv6hdr *ipv6h)
753 __wsum csum_pseudo_hdr = 0;
755 if (unlikely(ipv6h->nexthdr == IPPROTO_FRAGMENT ||
756 ipv6h->nexthdr == IPPROTO_HOPOPTS))
758 hw_checksum = csum_add(hw_checksum, (__force __wsum)htons(ipv6h->nexthdr));
760 csum_pseudo_hdr = csum_partial(&ipv6h->saddr,
761 sizeof(ipv6h->saddr) + sizeof(ipv6h->daddr), 0);
762 csum_pseudo_hdr = csum_add(csum_pseudo_hdr, (__force __wsum)ipv6h->payload_len);
763 csum_pseudo_hdr = csum_add(csum_pseudo_hdr, (__force __wsum)ntohs(ipv6h->nexthdr));
765 skb->csum = csum_sub(hw_checksum, csum_pseudo_hdr);
766 skb->csum = csum_add(skb->csum, csum_partial(ipv6h, sizeof(struct ipv6hdr), 0));
770 static int check_csum(struct mlx4_cqe *cqe, struct sk_buff *skb, void *va,
771 netdev_features_t dev_features)
773 __wsum hw_checksum = 0;
775 void *hdr = (u8 *)va + sizeof(struct ethhdr);
777 hw_checksum = csum_unfold((__force __sum16)cqe->checksum);
779 if (cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK) &&
780 !(dev_features & NETIF_F_HW_VLAN_CTAG_RX)) {
781 hw_checksum = get_fixed_vlan_csum(hw_checksum, hdr);
782 hdr += sizeof(struct vlan_hdr);
785 if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV4))
786 get_fixed_ipv4_csum(hw_checksum, skb, hdr);
787 #if IS_ENABLED(CONFIG_IPV6)
788 else if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV6))
789 if (unlikely(get_fixed_ipv6_csum(hw_checksum, skb, hdr)))
795 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
797 struct mlx4_en_priv *priv = netdev_priv(dev);
798 struct mlx4_en_dev *mdev = priv->mdev;
799 struct mlx4_cqe *cqe;
800 struct mlx4_en_rx_ring *ring = priv->rx_ring[cq->ring];
801 struct mlx4_en_rx_alloc *frags;
802 struct mlx4_en_rx_desc *rx_desc;
803 struct bpf_prog *xdp_prog;
804 int doorbell_pending;
811 int factor = priv->cqe_factor;
815 if (unlikely(!priv->port_up))
818 if (unlikely(budget <= 0))
821 /* Protect accesses to: ring->xdp_prog, priv->mac_hash list */
823 xdp_prog = rcu_dereference(ring->xdp_prog);
824 doorbell_pending = 0;
826 /* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
827 * descriptor offset can be deduced from the CQE index instead of
828 * reading 'cqe->index' */
829 index = cq->mcq.cons_index & ring->size_mask;
830 cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
832 /* Process all completed CQEs */
833 while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
834 cq->mcq.cons_index & cq->size)) {
836 frags = ring->rx_info + (index << priv->log_rx_info);
837 rx_desc = ring->buf + (index << ring->log_stride);
840 * make sure we read the CQE after we read the ownership bit
844 /* Drop packet on bad receive or bad checksum */
845 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
846 MLX4_CQE_OPCODE_ERROR)) {
847 en_err(priv, "CQE completed in error - vendor syndrom:%d syndrom:%d\n",
848 ((struct mlx4_err_cqe *)cqe)->vendor_err_syndrome,
849 ((struct mlx4_err_cqe *)cqe)->syndrome);
852 if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
853 en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
857 /* Check if we need to drop the packet if SRIOV is not enabled
858 * and not performing the selftest or flb disabled
860 if (priv->flags & MLX4_EN_FLAG_RX_FILTER_NEEDED) {
863 /* Get pointer to first fragment since we haven't
864 * skb yet and cast it to ethhdr struct
866 dma = be64_to_cpu(rx_desc->data[0].addr);
867 dma_sync_single_for_cpu(priv->ddev, dma, sizeof(*ethh),
869 ethh = (struct ethhdr *)(page_address(frags[0].page) +
870 frags[0].page_offset);
872 if (is_multicast_ether_addr(ethh->h_dest)) {
873 struct mlx4_mac_entry *entry;
874 struct hlist_head *bucket;
875 unsigned int mac_hash;
877 /* Drop the packet, since HW loopback-ed it */
878 mac_hash = ethh->h_source[MLX4_EN_MAC_HASH_IDX];
879 bucket = &priv->mac_hash[mac_hash];
880 hlist_for_each_entry_rcu(entry, bucket, hlist) {
881 if (ether_addr_equal_64bits(entry->mac,
889 * Packet is OK - process it.
891 length = be32_to_cpu(cqe->byte_cnt);
892 length -= ring->fcs_del;
893 l2_tunnel = (dev->hw_enc_features & NETIF_F_RXCSUM) &&
894 (cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_L2_TUNNEL));
896 /* A bpf program gets first chance to drop the packet. It may
897 * read bytes but not past the end of the frag.
905 dma = be64_to_cpu(rx_desc->data[0].addr);
906 dma_sync_single_for_cpu(priv->ddev, dma,
907 priv->frag_info[0].frag_size,
910 xdp.data_hard_start = page_address(frags[0].page);
911 xdp.data = xdp.data_hard_start + frags[0].page_offset;
912 xdp.data_end = xdp.data + length;
913 orig_data = xdp.data;
915 act = bpf_prog_run_xdp(xdp_prog, &xdp);
917 if (xdp.data != orig_data) {
918 length = xdp.data_end - xdp.data;
919 frags[0].page_offset = xdp.data -
927 if (likely(!mlx4_en_xmit_frame(ring, frags, dev,
931 goto xdp_drop_no_cnt; /* Drop on xmit failure */
933 bpf_warn_invalid_xdp_action(act);
938 if (likely(mlx4_en_rx_recycle(ring, frags)))
944 ring->bytes += length;
947 if (likely(dev->features & NETIF_F_RXCSUM)) {
948 if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_TCP |
949 MLX4_CQE_STATUS_UDP)) {
950 if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
951 cqe->checksum == cpu_to_be16(0xffff)) {
952 ip_summed = CHECKSUM_UNNECESSARY;
955 ip_summed = CHECKSUM_NONE;
959 if (priv->flags & MLX4_EN_FLAG_RX_CSUM_NON_TCP_UDP &&
960 (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
961 MLX4_CQE_STATUS_IPV6))) {
962 ip_summed = CHECKSUM_COMPLETE;
963 ring->csum_complete++;
965 ip_summed = CHECKSUM_NONE;
970 ip_summed = CHECKSUM_NONE;
974 /* This packet is eligible for GRO if it is:
975 * - DIX Ethernet (type interpretation)
977 * - without IP options
978 * - not an IP fragment
980 if (dev->features & NETIF_F_GRO) {
981 struct sk_buff *gro_skb = napi_get_frags(&cq->napi);
985 nr = mlx4_en_complete_rx_desc(priv,
986 rx_desc, frags, gro_skb,
991 if (ip_summed == CHECKSUM_COMPLETE) {
992 void *va = skb_frag_address(skb_shinfo(gro_skb)->frags);
993 if (check_csum(cqe, gro_skb, va,
995 ip_summed = CHECKSUM_NONE;
997 ring->csum_complete--;
1001 skb_shinfo(gro_skb)->nr_frags = nr;
1002 gro_skb->len = length;
1003 gro_skb->data_len = length;
1004 gro_skb->ip_summed = ip_summed;
1006 if (l2_tunnel && ip_summed == CHECKSUM_UNNECESSARY)
1007 gro_skb->csum_level = 1;
1009 if ((cqe->vlan_my_qpn &
1010 cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK)) &&
1011 (dev->features & NETIF_F_HW_VLAN_CTAG_RX)) {
1012 u16 vid = be16_to_cpu(cqe->sl_vid);
1014 __vlan_hwaccel_put_tag(gro_skb, htons(ETH_P_8021Q), vid);
1015 } else if ((be32_to_cpu(cqe->vlan_my_qpn) &
1016 MLX4_CQE_SVLAN_PRESENT_MASK) &&
1017 (dev->features & NETIF_F_HW_VLAN_STAG_RX)) {
1018 __vlan_hwaccel_put_tag(gro_skb,
1019 htons(ETH_P_8021AD),
1020 be16_to_cpu(cqe->sl_vid));
1023 if (dev->features & NETIF_F_RXHASH)
1024 skb_set_hash(gro_skb,
1025 be32_to_cpu(cqe->immed_rss_invalid),
1026 (ip_summed == CHECKSUM_UNNECESSARY) ?
1030 skb_record_rx_queue(gro_skb, cq->ring);
1032 if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
1033 timestamp = mlx4_en_get_cqe_ts(cqe);
1034 mlx4_en_fill_hwtstamps(mdev,
1035 skb_hwtstamps(gro_skb),
1039 napi_gro_frags(&cq->napi);
1043 /* GRO not possible, complete processing here */
1044 skb = mlx4_en_rx_skb(priv, rx_desc, frags, length);
1045 if (unlikely(!skb)) {
1050 if (unlikely(priv->validate_loopback)) {
1051 validate_loopback(priv, skb);
1055 if (ip_summed == CHECKSUM_COMPLETE) {
1056 if (check_csum(cqe, skb, skb->data, dev->features)) {
1057 ip_summed = CHECKSUM_NONE;
1058 ring->csum_complete--;
1063 skb->ip_summed = ip_summed;
1064 skb->protocol = eth_type_trans(skb, dev);
1065 skb_record_rx_queue(skb, cq->ring);
1067 if (l2_tunnel && ip_summed == CHECKSUM_UNNECESSARY)
1068 skb->csum_level = 1;
1070 if (dev->features & NETIF_F_RXHASH)
1072 be32_to_cpu(cqe->immed_rss_invalid),
1073 (ip_summed == CHECKSUM_UNNECESSARY) ?
1077 if ((be32_to_cpu(cqe->vlan_my_qpn) &
1078 MLX4_CQE_CVLAN_PRESENT_MASK) &&
1079 (dev->features & NETIF_F_HW_VLAN_CTAG_RX))
1080 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), be16_to_cpu(cqe->sl_vid));
1081 else if ((be32_to_cpu(cqe->vlan_my_qpn) &
1082 MLX4_CQE_SVLAN_PRESENT_MASK) &&
1083 (dev->features & NETIF_F_HW_VLAN_STAG_RX))
1084 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021AD),
1085 be16_to_cpu(cqe->sl_vid));
1087 if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
1088 timestamp = mlx4_en_get_cqe_ts(cqe);
1089 mlx4_en_fill_hwtstamps(mdev, skb_hwtstamps(skb),
1093 napi_gro_receive(&cq->napi, skb);
1095 for (nr = 0; nr < priv->num_frags; nr++)
1096 mlx4_en_free_frag(priv, frags, nr);
1099 ++cq->mcq.cons_index;
1100 index = (cq->mcq.cons_index) & ring->size_mask;
1101 cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
1102 if (++polled == budget)
1110 if (doorbell_pending)
1111 mlx4_en_xmit_doorbell(priv->tx_ring[TX_XDP][cq->ring]);
1113 mlx4_cq_set_ci(&cq->mcq);
1114 wmb(); /* ensure HW sees CQ consumer before we post new buffers */
1115 ring->cons = cq->mcq.cons_index;
1117 AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
1119 if (mlx4_en_refill_rx_buffers(priv, ring))
1120 mlx4_en_update_rx_prod_db(ring);
1126 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
1128 struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
1129 struct mlx4_en_priv *priv = netdev_priv(cq->dev);
1131 if (likely(priv->port_up))
1132 napi_schedule_irqoff(&cq->napi);
1134 mlx4_en_arm_cq(priv, cq);
1137 /* Rx CQ polling - called by NAPI */
1138 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
1140 struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
1141 struct net_device *dev = cq->dev;
1142 struct mlx4_en_priv *priv = netdev_priv(dev);
1145 done = mlx4_en_process_rx_cq(dev, cq, budget);
1147 /* If we used up all the quota - we're probably not done yet... */
1148 if (done == budget) {
1149 const struct cpumask *aff;
1150 struct irq_data *idata;
1153 INC_PERF_COUNTER(priv->pstats.napi_quota);
1155 cpu_curr = smp_processor_id();
1156 idata = irq_desc_get_irq_data(cq->irq_desc);
1157 aff = irq_data_get_affinity_mask(idata);
1159 if (likely(cpumask_test_cpu(cpu_curr, aff)))
1162 /* Current cpu is not according to smp_irq_affinity -
1163 * probably affinity changed. Need to stop this NAPI
1164 * poll, and restart it on the right CPU.
1165 * Try to avoid returning a too small value (like 0),
1166 * to not fool net_rx_action() and its netdev_budget
1172 if (napi_complete_done(napi, done))
1173 mlx4_en_arm_cq(priv, cq);
1177 static const int frag_sizes[] = {
1184 void mlx4_en_calc_rx_buf(struct net_device *dev)
1186 struct mlx4_en_priv *priv = netdev_priv(dev);
1187 int eff_mtu = MLX4_EN_EFF_MTU(dev->mtu);
1190 /* bpf requires buffers to be set up as 1 packet per page.
1191 * This only works when num_frags == 1.
1193 if (priv->tx_ring_num[TX_XDP]) {
1194 priv->frag_info[0].order = 0;
1195 priv->frag_info[0].frag_size = eff_mtu;
1196 priv->frag_info[0].frag_prefix_size = 0;
1197 /* This will gain efficient xdp frame recycling at the
1198 * expense of more costly truesize accounting
1200 priv->frag_info[0].frag_stride = PAGE_SIZE;
1201 priv->frag_info[0].dma_dir = PCI_DMA_BIDIRECTIONAL;
1202 priv->frag_info[0].rx_headroom = XDP_PACKET_HEADROOM;
1207 while (buf_size < eff_mtu) {
1208 priv->frag_info[i].order = MLX4_EN_ALLOC_PREFER_ORDER;
1209 priv->frag_info[i].frag_size =
1210 (eff_mtu > buf_size + frag_sizes[i]) ?
1211 frag_sizes[i] : eff_mtu - buf_size;
1212 priv->frag_info[i].frag_prefix_size = buf_size;
1213 priv->frag_info[i].frag_stride =
1214 ALIGN(priv->frag_info[i].frag_size,
1216 priv->frag_info[i].dma_dir = PCI_DMA_FROMDEVICE;
1217 priv->frag_info[i].rx_headroom = 0;
1218 buf_size += priv->frag_info[i].frag_size;
1223 priv->num_frags = i;
1224 priv->rx_skb_size = eff_mtu;
1225 priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct mlx4_en_rx_alloc));
1227 en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d num_frags:%d):\n",
1228 eff_mtu, priv->num_frags);
1229 for (i = 0; i < priv->num_frags; i++) {
1231 " frag:%d - size:%d prefix:%d stride:%d\n",
1233 priv->frag_info[i].frag_size,
1234 priv->frag_info[i].frag_prefix_size,
1235 priv->frag_info[i].frag_stride);
1239 /* RSS related functions */
1241 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
1242 struct mlx4_en_rx_ring *ring,
1243 enum mlx4_qp_state *state,
1246 struct mlx4_en_dev *mdev = priv->mdev;
1247 struct mlx4_qp_context *context;
1250 context = kmalloc(sizeof(*context), GFP_KERNEL);
1254 err = mlx4_qp_alloc(mdev->dev, qpn, qp, GFP_KERNEL);
1256 en_err(priv, "Failed to allocate qp #%x\n", qpn);
1259 qp->event = mlx4_en_sqp_event;
1261 memset(context, 0, sizeof *context);
1262 mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
1263 qpn, ring->cqn, -1, context);
1264 context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
1266 /* Cancel FCS removal if FW allows */
1267 if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
1268 context->param3 |= cpu_to_be32(1 << 29);
1269 if (priv->dev->features & NETIF_F_RXFCS)
1272 ring->fcs_del = ETH_FCS_LEN;
1276 err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
1278 mlx4_qp_remove(mdev->dev, qp);
1279 mlx4_qp_free(mdev->dev, qp);
1281 mlx4_en_update_rx_prod_db(ring);
1287 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
1292 err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn,
1293 MLX4_RESERVE_A0_QP);
1295 en_err(priv, "Failed reserving drop qpn\n");
1298 err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp, GFP_KERNEL);
1300 en_err(priv, "Failed allocating drop qp\n");
1301 mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1308 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
1312 qpn = priv->drop_qp.qpn;
1313 mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
1314 mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
1315 mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1318 /* Allocate rx qp's and configure them according to rss map */
1319 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
1321 struct mlx4_en_dev *mdev = priv->mdev;
1322 struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1323 struct mlx4_qp_context context;
1324 struct mlx4_rss_context *rss_context;
1327 u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
1333 en_dbg(DRV, priv, "Configuring rss steering\n");
1334 err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
1336 &rss_map->base_qpn, 0);
1338 en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
1342 for (i = 0; i < priv->rx_ring_num; i++) {
1343 qpn = rss_map->base_qpn + i;
1344 err = mlx4_en_config_rss_qp(priv, qpn, priv->rx_ring[i],
1353 /* Configure RSS indirection qp */
1354 err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp, GFP_KERNEL);
1356 en_err(priv, "Failed to allocate RSS indirection QP\n");
1359 rss_map->indir_qp.event = mlx4_en_sqp_event;
1360 mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1361 priv->rx_ring[0]->cqn, -1, &context);
1363 if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
1364 rss_rings = priv->rx_ring_num;
1366 rss_rings = priv->prof->rss_rings;
1368 ptr = ((void *) &context) + offsetof(struct mlx4_qp_context, pri_path)
1369 + MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
1371 rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
1372 (rss_map->base_qpn));
1373 rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1374 if (priv->mdev->profile.udp_rss) {
1375 rss_mask |= MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
1376 rss_context->base_qpn_udp = rss_context->default_qpn;
1379 if (mdev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1380 en_info(priv, "Setting RSS context tunnel type to RSS on inner headers\n");
1381 rss_mask |= MLX4_RSS_BY_INNER_HEADERS;
1384 rss_context->flags = rss_mask;
1385 rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1386 if (priv->rss_hash_fn == ETH_RSS_HASH_XOR) {
1387 rss_context->hash_fn = MLX4_RSS_HASH_XOR;
1388 } else if (priv->rss_hash_fn == ETH_RSS_HASH_TOP) {
1389 rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1390 memcpy(rss_context->rss_key, priv->rss_key,
1391 MLX4_EN_RSS_KEY_SIZE);
1393 en_err(priv, "Unknown RSS hash function requested\n");
1397 err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1398 &rss_map->indir_qp, &rss_map->indir_state);
1405 mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1406 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1407 mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1408 mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1410 for (i = 0; i < good_qps; i++) {
1411 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1412 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1413 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1414 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1416 mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1420 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1422 struct mlx4_en_dev *mdev = priv->mdev;
1423 struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1426 mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1427 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1428 mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1429 mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1431 for (i = 0; i < priv->rx_ring_num; i++) {
1432 mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1433 MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1434 mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1435 mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1437 mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);