1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2018 Solarflare Communications Inc.
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation, incorporated herein by reference.
11 #include "net_driver.h"
12 #include <linux/module.h>
13 #include <linux/iommu.h>
16 #include "rx_common.h"
18 /* This is the percentage fill level below which new RX descriptors
19 * will be added to the RX descriptor ring.
21 static unsigned int rx_refill_threshold;
22 module_param(rx_refill_threshold, uint, 0444);
23 MODULE_PARM_DESC(rx_refill_threshold,
24 "RX descriptor ring refill threshold (%)");
26 /* RX maximum head room required.
28 * This must be at least 1 to prevent overflow, plus one packet-worth
29 * to allow pipelined receives.
31 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
33 /* Check the RX page recycle ring for a page that can be reused. */
34 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
36 struct efx_nic *efx = rx_queue->efx;
37 struct efx_rx_page_state *state;
41 if (unlikely(!rx_queue->page_ring))
43 index = rx_queue->page_remove & rx_queue->page_ptr_mask;
44 page = rx_queue->page_ring[index];
48 rx_queue->page_ring[index] = NULL;
49 /* page_remove cannot exceed page_add. */
50 if (rx_queue->page_remove != rx_queue->page_add)
51 ++rx_queue->page_remove;
53 /* If page_count is 1 then we hold the only reference to this page. */
54 if (page_count(page) == 1) {
55 ++rx_queue->page_recycle_count;
58 state = page_address(page);
59 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
60 PAGE_SIZE << efx->rx_buffer_order,
63 ++rx_queue->page_recycle_failed;
69 /* Attempt to recycle the page if there is an RX recycle ring; the page can
70 * only be added if this is the final RX buffer, to prevent pages being used in
71 * the descriptor ring and appearing in the recycle ring simultaneously.
73 static void efx_recycle_rx_page(struct efx_channel *channel,
74 struct efx_rx_buffer *rx_buf)
76 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
77 struct efx_nic *efx = rx_queue->efx;
78 struct page *page = rx_buf->page;
81 /* Only recycle the page after processing the final buffer. */
82 if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
85 index = rx_queue->page_add & rx_queue->page_ptr_mask;
86 if (rx_queue->page_ring[index] == NULL) {
87 unsigned int read_index = rx_queue->page_remove &
88 rx_queue->page_ptr_mask;
90 /* The next slot in the recycle ring is available, but
91 * increment page_remove if the read pointer currently
94 if (read_index == index)
95 ++rx_queue->page_remove;
96 rx_queue->page_ring[index] = page;
100 ++rx_queue->page_recycle_full;
101 efx_unmap_rx_buffer(efx, rx_buf);
102 put_page(rx_buf->page);
105 /* Recycle the pages that are used by buffers that have just been received. */
106 void efx_recycle_rx_pages(struct efx_channel *channel,
107 struct efx_rx_buffer *rx_buf,
108 unsigned int n_frags)
110 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
112 if (unlikely(!rx_queue->page_ring))
116 efx_recycle_rx_page(channel, rx_buf);
117 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
121 void efx_discard_rx_packet(struct efx_channel *channel,
122 struct efx_rx_buffer *rx_buf,
123 unsigned int n_frags)
125 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
127 efx_recycle_rx_pages(channel, rx_buf, n_frags);
129 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
132 static void efx_init_rx_recycle_ring(struct efx_rx_queue *rx_queue)
134 unsigned int bufs_in_recycle_ring, page_ring_size;
135 struct efx_nic *efx = rx_queue->efx;
137 bufs_in_recycle_ring = efx_rx_recycle_ring_size(efx);
138 page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
139 efx->rx_bufs_per_page);
140 rx_queue->page_ring = kcalloc(page_ring_size,
141 sizeof(*rx_queue->page_ring), GFP_KERNEL);
142 if (!rx_queue->page_ring)
143 rx_queue->page_ptr_mask = 0;
145 rx_queue->page_ptr_mask = page_ring_size - 1;
148 static void efx_fini_rx_recycle_ring(struct efx_rx_queue *rx_queue)
150 struct efx_nic *efx = rx_queue->efx;
153 if (unlikely(!rx_queue->page_ring))
156 /* Unmap and release the pages in the recycle ring. Remove the ring. */
157 for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
158 struct page *page = rx_queue->page_ring[i];
159 struct efx_rx_page_state *state;
164 state = page_address(page);
165 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
166 PAGE_SIZE << efx->rx_buffer_order,
170 kfree(rx_queue->page_ring);
171 rx_queue->page_ring = NULL;
174 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
175 struct efx_rx_buffer *rx_buf)
177 /* Release the page reference we hold for the buffer. */
179 put_page(rx_buf->page);
181 /* If this is the last buffer in a page, unmap and free it. */
182 if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
183 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
184 efx_free_rx_buffers(rx_queue, rx_buf, 1);
189 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
191 struct efx_nic *efx = rx_queue->efx;
192 unsigned int entries;
195 /* Create the smallest power-of-two aligned ring */
196 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
197 EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
198 rx_queue->ptr_mask = entries - 1;
200 netif_dbg(efx, probe, efx->net_dev,
201 "creating RX queue %d size %#x mask %#x\n",
202 efx_rx_queue_index(rx_queue), efx->rxq_entries,
205 /* Allocate RX buffers */
206 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
208 if (!rx_queue->buffer)
211 rc = efx_nic_probe_rx(rx_queue);
213 kfree(rx_queue->buffer);
214 rx_queue->buffer = NULL;
220 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
222 unsigned int max_fill, trigger, max_trigger;
223 struct efx_nic *efx = rx_queue->efx;
226 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
227 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
229 /* Initialise ptr fields */
230 rx_queue->added_count = 0;
231 rx_queue->notified_count = 0;
232 rx_queue->granted_count = 0;
233 rx_queue->removed_count = 0;
234 rx_queue->min_fill = -1U;
235 efx_init_rx_recycle_ring(rx_queue);
237 rx_queue->page_remove = 0;
238 rx_queue->page_add = rx_queue->page_ptr_mask + 1;
239 rx_queue->page_recycle_count = 0;
240 rx_queue->page_recycle_failed = 0;
241 rx_queue->page_recycle_full = 0;
243 /* Initialise limit fields */
244 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
246 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
247 if (rx_refill_threshold != 0) {
248 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
249 if (trigger > max_trigger)
250 trigger = max_trigger;
252 trigger = max_trigger;
255 rx_queue->max_fill = max_fill;
256 rx_queue->fast_fill_trigger = trigger;
257 rx_queue->refill_enabled = true;
259 /* Initialise XDP queue information */
260 rc = xdp_rxq_info_reg(&rx_queue->xdp_rxq_info, efx->net_dev,
261 rx_queue->core_index, 0);
264 netif_err(efx, rx_err, efx->net_dev,
265 "Failure to initialise XDP queue information rc=%d\n",
267 efx->xdp_rxq_info_failed = true;
269 rx_queue->xdp_rxq_info_valid = true;
272 /* Set up RX descriptor ring */
273 efx_nic_init_rx(rx_queue);
276 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
278 struct efx_rx_buffer *rx_buf;
281 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
282 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
284 del_timer_sync(&rx_queue->slow_fill);
285 if (rx_queue->grant_credits)
286 flush_work(&rx_queue->grant_work);
288 /* Release RX buffers from the current read ptr to the write ptr */
289 if (rx_queue->buffer) {
290 for (i = rx_queue->removed_count; i < rx_queue->added_count;
292 unsigned int index = i & rx_queue->ptr_mask;
294 rx_buf = efx_rx_buffer(rx_queue, index);
295 efx_fini_rx_buffer(rx_queue, rx_buf);
299 efx_fini_rx_recycle_ring(rx_queue);
301 if (rx_queue->xdp_rxq_info_valid)
302 xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info);
304 rx_queue->xdp_rxq_info_valid = false;
307 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
309 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
310 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
312 efx_nic_remove_rx(rx_queue);
314 kfree(rx_queue->buffer);
315 rx_queue->buffer = NULL;
318 /* Unmap a DMA-mapped page. This function is only called for the final RX
321 void efx_unmap_rx_buffer(struct efx_nic *efx,
322 struct efx_rx_buffer *rx_buf)
324 struct page *page = rx_buf->page;
327 struct efx_rx_page_state *state = page_address(page);
329 dma_unmap_page(&efx->pci_dev->dev,
331 PAGE_SIZE << efx->rx_buffer_order,
336 void efx_free_rx_buffers(struct efx_rx_queue *rx_queue,
337 struct efx_rx_buffer *rx_buf,
338 unsigned int num_bufs)
342 put_page(rx_buf->page);
345 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
346 } while (--num_bufs);
349 void efx_rx_slow_fill(struct timer_list *t)
351 struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill);
353 /* Post an event to cause NAPI to run and refill the queue */
354 efx_nic_generate_fill_event(rx_queue);
355 ++rx_queue->slow_fill_count;
358 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
360 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(10));
363 /* efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
365 * @rx_queue: Efx RX queue
367 * This allocates a batch of pages, maps them for DMA, and populates
368 * struct efx_rx_buffers for each one. Return a negative error code or
369 * 0 on success. If a single page can be used for multiple buffers,
370 * then the page will either be inserted fully, or not at all.
372 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
374 unsigned int page_offset, index, count;
375 struct efx_nic *efx = rx_queue->efx;
376 struct efx_rx_page_state *state;
377 struct efx_rx_buffer *rx_buf;
383 page = efx_reuse_page(rx_queue);
385 page = alloc_pages(__GFP_COMP |
386 (atomic ? GFP_ATOMIC : GFP_KERNEL),
387 efx->rx_buffer_order);
388 if (unlikely(page == NULL))
391 dma_map_page(&efx->pci_dev->dev, page, 0,
392 PAGE_SIZE << efx->rx_buffer_order,
394 if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
396 __free_pages(page, efx->rx_buffer_order);
399 state = page_address(page);
400 state->dma_addr = dma_addr;
402 state = page_address(page);
403 dma_addr = state->dma_addr;
406 dma_addr += sizeof(struct efx_rx_page_state);
407 page_offset = sizeof(struct efx_rx_page_state);
410 index = rx_queue->added_count & rx_queue->ptr_mask;
411 rx_buf = efx_rx_buffer(rx_queue, index);
412 rx_buf->dma_addr = dma_addr + efx->rx_ip_align +
415 rx_buf->page_offset = page_offset + efx->rx_ip_align +
417 rx_buf->len = efx->rx_dma_len;
419 ++rx_queue->added_count;
421 dma_addr += efx->rx_page_buf_step;
422 page_offset += efx->rx_page_buf_step;
423 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
425 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
426 } while (++count < efx->rx_pages_per_batch);
431 void efx_rx_config_page_split(struct efx_nic *efx)
433 efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align +
434 EFX_XDP_HEADROOM + EFX_XDP_TAILROOM,
435 EFX_RX_BUF_ALIGNMENT);
436 efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
437 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
438 efx->rx_page_buf_step);
439 efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
440 efx->rx_bufs_per_page;
441 efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
442 efx->rx_bufs_per_page);
445 /* efx_fast_push_rx_descriptors - push new RX descriptors quickly
446 * @rx_queue: RX descriptor queue
448 * This will aim to fill the RX descriptor queue up to
449 * @rx_queue->@max_fill. If there is insufficient atomic
450 * memory to do so, a slow fill will be scheduled.
452 * The caller must provide serialisation (none is used here). In practise,
453 * this means this function must run from the NAPI handler, or be called
454 * when NAPI is disabled.
456 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
458 struct efx_nic *efx = rx_queue->efx;
459 unsigned int fill_level, batch_size;
462 if (!rx_queue->refill_enabled)
465 /* Calculate current fill level, and exit if we don't need to fill */
466 fill_level = (rx_queue->added_count - rx_queue->removed_count);
467 EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries);
468 if (fill_level >= rx_queue->fast_fill_trigger)
471 /* Record minimum fill level */
472 if (unlikely(fill_level < rx_queue->min_fill)) {
474 rx_queue->min_fill = fill_level;
477 batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
478 space = rx_queue->max_fill - fill_level;
479 EFX_WARN_ON_ONCE_PARANOID(space < batch_size);
481 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
482 "RX queue %d fast-filling descriptor ring from"
483 " level %d to level %d\n",
484 efx_rx_queue_index(rx_queue), fill_level,
488 rc = efx_init_rx_buffers(rx_queue, atomic);
490 /* Ensure that we don't leave the rx queue empty */
491 efx_schedule_slow_fill(rx_queue);
494 } while ((space -= batch_size) >= batch_size);
496 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
497 "RX queue %d fast-filled descriptor ring "
498 "to level %d\n", efx_rx_queue_index(rx_queue),
499 rx_queue->added_count - rx_queue->removed_count);
502 if (rx_queue->notified_count != rx_queue->added_count)
503 efx_nic_notify_rx_desc(rx_queue);
506 /* Pass a received packet up through GRO. GRO can handle pages
507 * regardless of checksum state and skbs with a good checksum.
510 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
511 unsigned int n_frags, u8 *eh, __wsum csum)
513 struct napi_struct *napi = &channel->napi_str;
514 struct efx_nic *efx = channel->efx;
517 skb = napi_get_frags(napi);
518 if (unlikely(!skb)) {
519 struct efx_rx_queue *rx_queue;
521 rx_queue = efx_channel_get_rx_queue(channel);
522 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
526 if (efx->net_dev->features & NETIF_F_RXHASH &&
527 efx_rx_buf_hash_valid(efx, eh))
528 skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
532 skb->ip_summed = CHECKSUM_COMPLETE;
534 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
535 CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
537 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
540 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
541 rx_buf->page, rx_buf->page_offset,
544 skb->len += rx_buf->len;
545 if (skb_shinfo(skb)->nr_frags == n_frags)
548 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
551 skb->data_len = skb->len;
552 skb->truesize += n_frags * efx->rx_buffer_truesize;
554 skb_record_rx_queue(skb, channel->rx_queue.core_index);
556 napi_gro_frags(napi);
559 /* RSS contexts. We're using linked lists and crappy O(n) algorithms, because
560 * (a) this is an infrequent control-plane operation and (b) n is small (max 64)
562 struct efx_rss_context *efx_alloc_rss_context_entry(struct efx_nic *efx)
564 struct list_head *head = &efx->rss_context.list;
565 struct efx_rss_context *ctx, *new;
566 u32 id = 1; /* Don't use zero, that refers to the master RSS context */
568 WARN_ON(!mutex_is_locked(&efx->rss_lock));
570 /* Search for first gap in the numbering */
571 list_for_each_entry(ctx, head, list) {
572 if (ctx->user_id != id)
575 /* Check for wrap. If this happens, we have nearly 2^32
576 * allocated RSS contexts, which seems unlikely.
578 if (WARN_ON_ONCE(!id))
582 /* Create the new entry */
583 new = kmalloc(sizeof(*new), GFP_KERNEL);
586 new->context_id = EFX_MCDI_RSS_CONTEXT_INVALID;
587 new->rx_hash_udp_4tuple = false;
589 /* Insert the new entry into the gap */
591 list_add_tail(&new->list, &ctx->list);
595 struct efx_rss_context *efx_find_rss_context_entry(struct efx_nic *efx, u32 id)
597 struct list_head *head = &efx->rss_context.list;
598 struct efx_rss_context *ctx;
600 WARN_ON(!mutex_is_locked(&efx->rss_lock));
602 list_for_each_entry(ctx, head, list)
603 if (ctx->user_id == id)
608 void efx_free_rss_context_entry(struct efx_rss_context *ctx)
610 list_del(&ctx->list);
614 void efx_set_default_rx_indir_table(struct efx_nic *efx,
615 struct efx_rss_context *ctx)
619 for (i = 0; i < ARRAY_SIZE(ctx->rx_indir_table); i++)
620 ctx->rx_indir_table[i] =
621 ethtool_rxfh_indir_default(i, efx->rss_spread);
625 * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
626 * @spec: Specification to test
628 * Return: %true if the specification is a non-drop RX filter that
629 * matches a local MAC address I/G bit value of 1 or matches a local
630 * IPv4 or IPv6 address value in the respective multicast address
631 * range. Otherwise %false.
633 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
635 if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
636 spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
639 if (spec->match_flags &
640 (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
641 is_multicast_ether_addr(spec->loc_mac))
644 if ((spec->match_flags &
645 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
646 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
647 if (spec->ether_type == htons(ETH_P_IP) &&
648 ipv4_is_multicast(spec->loc_host[0]))
650 if (spec->ether_type == htons(ETH_P_IPV6) &&
651 ((const u8 *)spec->loc_host)[0] == 0xff)
658 bool efx_filter_spec_equal(const struct efx_filter_spec *left,
659 const struct efx_filter_spec *right)
661 if ((left->match_flags ^ right->match_flags) |
662 ((left->flags ^ right->flags) &
663 (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
666 return memcmp(&left->vport_id, &right->vport_id,
667 sizeof(struct efx_filter_spec) -
668 offsetof(struct efx_filter_spec, vport_id)) == 0;
671 u32 efx_filter_spec_hash(const struct efx_filter_spec *spec)
673 BUILD_BUG_ON(offsetof(struct efx_filter_spec, vport_id) & 3);
674 return jhash2((const u32 *)&spec->vport_id,
675 (sizeof(struct efx_filter_spec) -
676 offsetof(struct efx_filter_spec, vport_id)) / 4,
680 #ifdef CONFIG_RFS_ACCEL
681 bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx,
684 if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) {
685 /* ARFS is currently updating this entry, leave it */
688 if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) {
689 /* ARFS tried and failed to update this, so it's probably out
690 * of date. Remove the filter and the ARFS rule entry.
692 rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
695 } else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */
696 /* ARFS has moved on, so old filter is not needed. Since we did
697 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will
698 * not be removed by efx_rps_hash_del() subsequently.
703 /* Remove it iff ARFS wants to. */
708 struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx,
709 const struct efx_filter_spec *spec)
711 u32 hash = efx_filter_spec_hash(spec);
713 lockdep_assert_held(&efx->rps_hash_lock);
714 if (!efx->rps_hash_table)
716 return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE];
719 struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx,
720 const struct efx_filter_spec *spec)
722 struct efx_arfs_rule *rule;
723 struct hlist_head *head;
724 struct hlist_node *node;
726 head = efx_rps_hash_bucket(efx, spec);
729 hlist_for_each(node, head) {
730 rule = container_of(node, struct efx_arfs_rule, node);
731 if (efx_filter_spec_equal(spec, &rule->spec))
737 struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
738 const struct efx_filter_spec *spec,
741 struct efx_arfs_rule *rule;
742 struct hlist_head *head;
743 struct hlist_node *node;
745 head = efx_rps_hash_bucket(efx, spec);
748 hlist_for_each(node, head) {
749 rule = container_of(node, struct efx_arfs_rule, node);
750 if (efx_filter_spec_equal(spec, &rule->spec)) {
755 rule = kmalloc(sizeof(*rule), GFP_ATOMIC);
758 memcpy(&rule->spec, spec, sizeof(rule->spec));
759 hlist_add_head(&rule->node, head);
764 void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec)
766 struct efx_arfs_rule *rule;
767 struct hlist_head *head;
768 struct hlist_node *node;
770 head = efx_rps_hash_bucket(efx, spec);
773 hlist_for_each(node, head) {
774 rule = container_of(node, struct efx_arfs_rule, node);
775 if (efx_filter_spec_equal(spec, &rule->spec)) {
776 /* Someone already reused the entry. We know that if
777 * this check doesn't fire (i.e. filter_id == REMOVING)
778 * then the REMOVING mark was put there by our caller,
779 * because caller is holding a lock on filter table and
780 * only holders of that lock set REMOVING.
782 if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING)
789 /* We didn't find it. */
794 int efx_probe_filters(struct efx_nic *efx)
798 mutex_lock(&efx->mac_lock);
799 rc = efx->type->filter_table_probe(efx);
803 #ifdef CONFIG_RFS_ACCEL
804 if (efx->type->offload_features & NETIF_F_NTUPLE) {
805 struct efx_channel *channel;
808 efx_for_each_channel(channel, efx) {
809 channel->rps_flow_id =
810 kcalloc(efx->type->max_rx_ip_filters,
811 sizeof(*channel->rps_flow_id),
813 if (!channel->rps_flow_id)
817 i < efx->type->max_rx_ip_filters;
819 channel->rps_flow_id[i] =
821 channel->rfs_expire_index = 0;
822 channel->rfs_filter_count = 0;
826 efx_for_each_channel(channel, efx) {
827 kfree(channel->rps_flow_id);
828 channel->rps_flow_id = NULL;
830 efx->type->filter_table_remove(efx);
837 mutex_unlock(&efx->mac_lock);
841 void efx_remove_filters(struct efx_nic *efx)
843 #ifdef CONFIG_RFS_ACCEL
844 struct efx_channel *channel;
846 efx_for_each_channel(channel, efx) {
847 cancel_delayed_work_sync(&channel->filter_work);
848 kfree(channel->rps_flow_id);
849 channel->rps_flow_id = NULL;
852 efx->type->filter_table_remove(efx);
855 #ifdef CONFIG_RFS_ACCEL
857 static void efx_filter_rfs_work(struct work_struct *data)
859 struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
861 struct efx_nic *efx = efx_netdev_priv(req->net_dev);
862 struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
863 int slot_idx = req - efx->rps_slot;
864 struct efx_arfs_rule *rule;
868 rc = efx->type->filter_insert(efx, &req->spec, true);
870 /* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */
871 rc %= efx->type->max_rx_ip_filters;
872 if (efx->rps_hash_table) {
873 spin_lock_bh(&efx->rps_hash_lock);
874 rule = efx_rps_hash_find(efx, &req->spec);
875 /* The rule might have already gone, if someone else's request
876 * for the same spec was already worked and then expired before
877 * we got around to our work. In that case we have nothing
878 * tying us to an arfs_id, meaning that as soon as the filter
879 * is considered for expiry it will be removed.
883 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
885 rule->filter_id = rc;
886 arfs_id = rule->arfs_id;
888 spin_unlock_bh(&efx->rps_hash_lock);
891 /* Remember this so we can check whether to expire the filter
894 mutex_lock(&efx->rps_mutex);
895 if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID)
896 channel->rfs_filter_count++;
897 channel->rps_flow_id[rc] = req->flow_id;
898 mutex_unlock(&efx->rps_mutex);
900 if (req->spec.ether_type == htons(ETH_P_IP))
901 netif_info(efx, rx_status, efx->net_dev,
902 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
903 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
904 req->spec.rem_host, ntohs(req->spec.rem_port),
905 req->spec.loc_host, ntohs(req->spec.loc_port),
906 req->rxq_index, req->flow_id, rc, arfs_id);
908 netif_info(efx, rx_status, efx->net_dev,
909 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
910 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
911 req->spec.rem_host, ntohs(req->spec.rem_port),
912 req->spec.loc_host, ntohs(req->spec.loc_port),
913 req->rxq_index, req->flow_id, rc, arfs_id);
914 channel->n_rfs_succeeded++;
916 if (req->spec.ether_type == htons(ETH_P_IP))
917 netif_dbg(efx, rx_status, efx->net_dev,
918 "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n",
919 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
920 req->spec.rem_host, ntohs(req->spec.rem_port),
921 req->spec.loc_host, ntohs(req->spec.loc_port),
922 req->rxq_index, req->flow_id, rc, arfs_id);
924 netif_dbg(efx, rx_status, efx->net_dev,
925 "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n",
926 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
927 req->spec.rem_host, ntohs(req->spec.rem_port),
928 req->spec.loc_host, ntohs(req->spec.loc_port),
929 req->rxq_index, req->flow_id, rc, arfs_id);
930 channel->n_rfs_failed++;
931 /* We're overloading the NIC's filter tables, so let's do a
932 * chunk of extra expiry work.
934 __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count,
938 /* Release references */
939 clear_bit(slot_idx, &efx->rps_slot_map);
940 dev_put(req->net_dev);
943 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
944 u16 rxq_index, u32 flow_id)
946 struct efx_nic *efx = efx_netdev_priv(net_dev);
947 struct efx_async_filter_insertion *req;
948 struct efx_arfs_rule *rule;
954 /* find a free slot */
955 for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
956 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
958 if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
961 if (flow_id == RPS_FLOW_ID_INVALID) {
966 if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
967 rc = -EPROTONOSUPPORT;
971 if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
972 rc = -EPROTONOSUPPORT;
975 if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
976 rc = -EPROTONOSUPPORT;
980 req = efx->rps_slot + slot_idx;
981 efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
982 efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
984 req->spec.match_flags =
985 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
986 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
987 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
988 req->spec.ether_type = fk.basic.n_proto;
989 req->spec.ip_proto = fk.basic.ip_proto;
991 if (fk.basic.n_proto == htons(ETH_P_IP)) {
992 req->spec.rem_host[0] = fk.addrs.v4addrs.src;
993 req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
995 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
996 sizeof(struct in6_addr));
997 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
998 sizeof(struct in6_addr));
1001 req->spec.rem_port = fk.ports.src;
1002 req->spec.loc_port = fk.ports.dst;
1004 if (efx->rps_hash_table) {
1005 /* Add it to ARFS hash table */
1006 spin_lock(&efx->rps_hash_lock);
1007 rule = efx_rps_hash_add(efx, &req->spec, &new);
1013 rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
1015 /* Skip if existing or pending filter already does the right thing */
1016 if (!new && rule->rxq_index == rxq_index &&
1017 rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
1019 rule->rxq_index = rxq_index;
1020 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
1021 spin_unlock(&efx->rps_hash_lock);
1023 /* Without an ARFS hash table, we just use arfs_id 0 for all
1024 * filters. This means if multiple flows hash to the same
1025 * flow_id, all but the most recently touched will be eligible
1031 /* Queue the request */
1032 dev_hold(req->net_dev = net_dev);
1033 INIT_WORK(&req->work, efx_filter_rfs_work);
1034 req->rxq_index = rxq_index;
1035 req->flow_id = flow_id;
1036 schedule_work(&req->work);
1039 spin_unlock(&efx->rps_hash_lock);
1041 clear_bit(slot_idx, &efx->rps_slot_map);
1045 bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota)
1047 bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
1048 struct efx_nic *efx = channel->efx;
1049 unsigned int index, size, start;
1052 if (!mutex_trylock(&efx->rps_mutex))
1054 expire_one = efx->type->filter_rfs_expire_one;
1055 index = channel->rfs_expire_index;
1057 size = efx->type->max_rx_ip_filters;
1059 flow_id = channel->rps_flow_id[index];
1061 if (flow_id != RPS_FLOW_ID_INVALID) {
1063 if (expire_one(efx, flow_id, index)) {
1064 netif_info(efx, rx_status, efx->net_dev,
1065 "expired filter %d [channel %u flow %u]\n",
1066 index, channel->channel, flow_id);
1067 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1068 channel->rfs_filter_count--;
1071 if (++index == size)
1073 /* If we were called with a quota that exceeds the total number
1074 * of filters in the table (which shouldn't happen, but could
1075 * if two callers race), ensure that we don't loop forever -
1076 * stop when we've examined every row of the table.
1082 channel->rfs_expire_index = index;
1083 mutex_unlock(&efx->rps_mutex);
1087 #endif /* CONFIG_RFS_ACCEL */