1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2005-2006 Fen Systems Ltd.
5 * Copyright 2005-2013 Solarflare Communications Inc.
8 #include <linux/socket.h>
10 #include <linux/slab.h>
12 #include <linux/ipv6.h>
13 #include <linux/tcp.h>
14 #include <linux/udp.h>
15 #include <linux/prefetch.h>
16 #include <linux/moduleparam.h>
17 #include <linux/iommu.h>
19 #include <net/checksum.h>
20 #include "net_driver.h"
25 #include "workarounds.h"
27 /* Preferred number of descriptors to fill at once */
28 #define EFX_RX_PREFERRED_BATCH 8U
30 /* Number of RX buffers to recycle pages for. When creating the RX page recycle
31 * ring, this number is divided by the number of buffers per page to calculate
32 * the number of pages to store in the RX page recycle ring.
34 #define EFX_RECYCLE_RING_SIZE_IOMMU 4096
35 #define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
37 /* Size of buffer allocated for skb header area. */
38 #define EFX_SKB_HEADERS 128u
40 /* This is the percentage fill level below which new RX descriptors
41 * will be added to the RX descriptor ring.
43 static unsigned int rx_refill_threshold;
45 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
46 #define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
50 * RX maximum head room required.
52 * This must be at least 1 to prevent overflow, plus one packet-worth
53 * to allow pipelined receives.
55 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
57 static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
59 return page_address(buf->page) + buf->page_offset;
62 static inline u32 efx_rx_buf_hash(struct efx_nic *efx, const u8 *eh)
64 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
65 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_hash_offset));
67 const u8 *data = eh + efx->rx_packet_hash_offset;
75 static inline struct efx_rx_buffer *
76 efx_rx_buf_next(struct efx_rx_queue *rx_queue, struct efx_rx_buffer *rx_buf)
78 if (unlikely(rx_buf == efx_rx_buffer(rx_queue, rx_queue->ptr_mask)))
79 return efx_rx_buffer(rx_queue, 0);
84 static inline void efx_sync_rx_buffer(struct efx_nic *efx,
85 struct efx_rx_buffer *rx_buf,
88 dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr, len,
92 void efx_rx_config_page_split(struct efx_nic *efx)
94 efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align,
95 EFX_RX_BUF_ALIGNMENT);
96 efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
97 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
98 efx->rx_page_buf_step);
99 efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
100 efx->rx_bufs_per_page;
101 efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
102 efx->rx_bufs_per_page);
105 /* Check the RX page recycle ring for a page that can be reused. */
106 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
108 struct efx_nic *efx = rx_queue->efx;
110 struct efx_rx_page_state *state;
113 index = rx_queue->page_remove & rx_queue->page_ptr_mask;
114 page = rx_queue->page_ring[index];
118 rx_queue->page_ring[index] = NULL;
119 /* page_remove cannot exceed page_add. */
120 if (rx_queue->page_remove != rx_queue->page_add)
121 ++rx_queue->page_remove;
123 /* If page_count is 1 then we hold the only reference to this page. */
124 if (page_count(page) == 1) {
125 ++rx_queue->page_recycle_count;
128 state = page_address(page);
129 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
130 PAGE_SIZE << efx->rx_buffer_order,
133 ++rx_queue->page_recycle_failed;
140 * efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
142 * @rx_queue: Efx RX queue
144 * This allocates a batch of pages, maps them for DMA, and populates
145 * struct efx_rx_buffers for each one. Return a negative error code or
146 * 0 on success. If a single page can be used for multiple buffers,
147 * then the page will either be inserted fully, or not at all.
149 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
151 struct efx_nic *efx = rx_queue->efx;
152 struct efx_rx_buffer *rx_buf;
154 unsigned int page_offset;
155 struct efx_rx_page_state *state;
157 unsigned index, count;
161 page = efx_reuse_page(rx_queue);
163 page = alloc_pages(__GFP_COMP |
164 (atomic ? GFP_ATOMIC : GFP_KERNEL),
165 efx->rx_buffer_order);
166 if (unlikely(page == NULL))
169 dma_map_page(&efx->pci_dev->dev, page, 0,
170 PAGE_SIZE << efx->rx_buffer_order,
172 if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
174 __free_pages(page, efx->rx_buffer_order);
177 state = page_address(page);
178 state->dma_addr = dma_addr;
180 state = page_address(page);
181 dma_addr = state->dma_addr;
184 dma_addr += sizeof(struct efx_rx_page_state);
185 page_offset = sizeof(struct efx_rx_page_state);
188 index = rx_queue->added_count & rx_queue->ptr_mask;
189 rx_buf = efx_rx_buffer(rx_queue, index);
190 rx_buf->dma_addr = dma_addr + efx->rx_ip_align;
192 rx_buf->page_offset = page_offset + efx->rx_ip_align;
193 rx_buf->len = efx->rx_dma_len;
195 ++rx_queue->added_count;
197 dma_addr += efx->rx_page_buf_step;
198 page_offset += efx->rx_page_buf_step;
199 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
201 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
202 } while (++count < efx->rx_pages_per_batch);
207 /* Unmap a DMA-mapped page. This function is only called for the final RX
210 static void efx_unmap_rx_buffer(struct efx_nic *efx,
211 struct efx_rx_buffer *rx_buf)
213 struct page *page = rx_buf->page;
216 struct efx_rx_page_state *state = page_address(page);
217 dma_unmap_page(&efx->pci_dev->dev,
219 PAGE_SIZE << efx->rx_buffer_order,
224 static void efx_free_rx_buffers(struct efx_rx_queue *rx_queue,
225 struct efx_rx_buffer *rx_buf,
226 unsigned int num_bufs)
230 put_page(rx_buf->page);
233 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
234 } while (--num_bufs);
237 /* Attempt to recycle the page if there is an RX recycle ring; the page can
238 * only be added if this is the final RX buffer, to prevent pages being used in
239 * the descriptor ring and appearing in the recycle ring simultaneously.
241 static void efx_recycle_rx_page(struct efx_channel *channel,
242 struct efx_rx_buffer *rx_buf)
244 struct page *page = rx_buf->page;
245 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
246 struct efx_nic *efx = rx_queue->efx;
249 /* Only recycle the page after processing the final buffer. */
250 if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
253 index = rx_queue->page_add & rx_queue->page_ptr_mask;
254 if (rx_queue->page_ring[index] == NULL) {
255 unsigned read_index = rx_queue->page_remove &
256 rx_queue->page_ptr_mask;
258 /* The next slot in the recycle ring is available, but
259 * increment page_remove if the read pointer currently
262 if (read_index == index)
263 ++rx_queue->page_remove;
264 rx_queue->page_ring[index] = page;
265 ++rx_queue->page_add;
268 ++rx_queue->page_recycle_full;
269 efx_unmap_rx_buffer(efx, rx_buf);
270 put_page(rx_buf->page);
273 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
274 struct efx_rx_buffer *rx_buf)
276 /* Release the page reference we hold for the buffer. */
278 put_page(rx_buf->page);
280 /* If this is the last buffer in a page, unmap and free it. */
281 if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
282 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
283 efx_free_rx_buffers(rx_queue, rx_buf, 1);
288 /* Recycle the pages that are used by buffers that have just been received. */
289 static void efx_recycle_rx_pages(struct efx_channel *channel,
290 struct efx_rx_buffer *rx_buf,
291 unsigned int n_frags)
293 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
296 efx_recycle_rx_page(channel, rx_buf);
297 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
301 static void efx_discard_rx_packet(struct efx_channel *channel,
302 struct efx_rx_buffer *rx_buf,
303 unsigned int n_frags)
305 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
307 efx_recycle_rx_pages(channel, rx_buf, n_frags);
309 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
313 * efx_fast_push_rx_descriptors - push new RX descriptors quickly
314 * @rx_queue: RX descriptor queue
316 * This will aim to fill the RX descriptor queue up to
317 * @rx_queue->@max_fill. If there is insufficient atomic
318 * memory to do so, a slow fill will be scheduled.
320 * The caller must provide serialisation (none is used here). In practise,
321 * this means this function must run from the NAPI handler, or be called
322 * when NAPI is disabled.
324 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
326 struct efx_nic *efx = rx_queue->efx;
327 unsigned int fill_level, batch_size;
330 if (!rx_queue->refill_enabled)
333 /* Calculate current fill level, and exit if we don't need to fill */
334 fill_level = (rx_queue->added_count - rx_queue->removed_count);
335 EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries);
336 if (fill_level >= rx_queue->fast_fill_trigger)
339 /* Record minimum fill level */
340 if (unlikely(fill_level < rx_queue->min_fill)) {
342 rx_queue->min_fill = fill_level;
345 batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
346 space = rx_queue->max_fill - fill_level;
347 EFX_WARN_ON_ONCE_PARANOID(space < batch_size);
349 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
350 "RX queue %d fast-filling descriptor ring from"
351 " level %d to level %d\n",
352 efx_rx_queue_index(rx_queue), fill_level,
357 rc = efx_init_rx_buffers(rx_queue, atomic);
359 /* Ensure that we don't leave the rx queue empty */
360 efx_schedule_slow_fill(rx_queue);
363 } while ((space -= batch_size) >= batch_size);
365 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
366 "RX queue %d fast-filled descriptor ring "
367 "to level %d\n", efx_rx_queue_index(rx_queue),
368 rx_queue->added_count - rx_queue->removed_count);
371 if (rx_queue->notified_count != rx_queue->added_count)
372 efx_nic_notify_rx_desc(rx_queue);
375 void efx_rx_slow_fill(struct timer_list *t)
377 struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill);
379 /* Post an event to cause NAPI to run and refill the queue */
380 efx_nic_generate_fill_event(rx_queue);
381 ++rx_queue->slow_fill_count;
384 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
385 struct efx_rx_buffer *rx_buf,
388 struct efx_nic *efx = rx_queue->efx;
389 unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
391 if (likely(len <= max_len))
394 /* The packet must be discarded, but this is only a fatal error
395 * if the caller indicated it was
397 rx_buf->flags |= EFX_RX_PKT_DISCARD;
400 netif_err(efx, rx_err, efx->net_dev,
401 "RX queue %d overlength RX event (%#x > %#x)\n",
402 efx_rx_queue_index(rx_queue), len, max_len);
404 efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
407 /* Pass a received packet up through GRO. GRO can handle pages
408 * regardless of checksum state and skbs with a good checksum.
411 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
412 unsigned int n_frags, u8 *eh)
414 struct napi_struct *napi = &channel->napi_str;
415 gro_result_t gro_result;
416 struct efx_nic *efx = channel->efx;
419 skb = napi_get_frags(napi);
420 if (unlikely(!skb)) {
421 struct efx_rx_queue *rx_queue;
423 rx_queue = efx_channel_get_rx_queue(channel);
424 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
428 if (efx->net_dev->features & NETIF_F_RXHASH)
429 skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
431 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
432 CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
433 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
436 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
437 rx_buf->page, rx_buf->page_offset,
440 skb->len += rx_buf->len;
441 if (skb_shinfo(skb)->nr_frags == n_frags)
444 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
447 skb->data_len = skb->len;
448 skb->truesize += n_frags * efx->rx_buffer_truesize;
450 skb_record_rx_queue(skb, channel->rx_queue.core_index);
452 gro_result = napi_gro_frags(napi);
453 if (gro_result != GRO_DROP)
454 channel->irq_mod_score += 2;
457 /* Allocate and construct an SKB around page fragments */
458 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
459 struct efx_rx_buffer *rx_buf,
460 unsigned int n_frags,
463 struct efx_nic *efx = channel->efx;
466 /* Allocate an SKB to store the headers */
467 skb = netdev_alloc_skb(efx->net_dev,
468 efx->rx_ip_align + efx->rx_prefix_size +
470 if (unlikely(skb == NULL)) {
471 atomic_inc(&efx->n_rx_noskb_drops);
475 EFX_WARN_ON_ONCE_PARANOID(rx_buf->len < hdr_len);
477 memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
478 efx->rx_prefix_size + hdr_len);
479 skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
480 __skb_put(skb, hdr_len);
482 /* Append the remaining page(s) onto the frag list */
483 if (rx_buf->len > hdr_len) {
484 rx_buf->page_offset += hdr_len;
485 rx_buf->len -= hdr_len;
488 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
489 rx_buf->page, rx_buf->page_offset,
492 skb->len += rx_buf->len;
493 skb->data_len += rx_buf->len;
494 if (skb_shinfo(skb)->nr_frags == n_frags)
497 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
500 __free_pages(rx_buf->page, efx->rx_buffer_order);
505 skb->truesize += n_frags * efx->rx_buffer_truesize;
507 /* Move past the ethernet header */
508 skb->protocol = eth_type_trans(skb, efx->net_dev);
510 skb_mark_napi_id(skb, &channel->napi_str);
515 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
516 unsigned int n_frags, unsigned int len, u16 flags)
518 struct efx_nic *efx = rx_queue->efx;
519 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
520 struct efx_rx_buffer *rx_buf;
522 rx_queue->rx_packets++;
524 rx_buf = efx_rx_buffer(rx_queue, index);
525 rx_buf->flags |= flags;
527 /* Validate the number of fragments and completed length */
529 if (!(flags & EFX_RX_PKT_PREFIX_LEN))
530 efx_rx_packet__check_len(rx_queue, rx_buf, len);
531 } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
532 unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
533 unlikely(len > n_frags * efx->rx_dma_len) ||
534 unlikely(!efx->rx_scatter)) {
535 /* If this isn't an explicit discard request, either
536 * the hardware or the driver is broken.
538 WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
539 rx_buf->flags |= EFX_RX_PKT_DISCARD;
542 netif_vdbg(efx, rx_status, efx->net_dev,
543 "RX queue %d received ids %x-%x len %d %s%s\n",
544 efx_rx_queue_index(rx_queue), index,
545 (index + n_frags - 1) & rx_queue->ptr_mask, len,
546 (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
547 (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
549 /* Discard packet, if instructed to do so. Process the
550 * previous receive first.
552 if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
553 efx_rx_flush_packet(channel);
554 efx_discard_rx_packet(channel, rx_buf, n_frags);
558 if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
561 /* Release and/or sync the DMA mapping - assumes all RX buffers
562 * consumed in-order per RX queue.
564 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
566 /* Prefetch nice and early so data will (hopefully) be in cache by
567 * the time we look at it.
569 prefetch(efx_rx_buf_va(rx_buf));
571 rx_buf->page_offset += efx->rx_prefix_size;
572 rx_buf->len -= efx->rx_prefix_size;
575 /* Release/sync DMA mapping for additional fragments.
576 * Fix length for last fragment.
578 unsigned int tail_frags = n_frags - 1;
581 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
582 if (--tail_frags == 0)
584 efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
586 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
587 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
590 /* All fragments have been DMA-synced, so recycle pages. */
591 rx_buf = efx_rx_buffer(rx_queue, index);
592 efx_recycle_rx_pages(channel, rx_buf, n_frags);
594 /* Pipeline receives so that we give time for packet headers to be
595 * prefetched into cache.
597 efx_rx_flush_packet(channel);
598 channel->rx_pkt_n_frags = n_frags;
599 channel->rx_pkt_index = index;
602 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
603 struct efx_rx_buffer *rx_buf,
604 unsigned int n_frags)
607 u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
609 skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
610 if (unlikely(skb == NULL)) {
611 struct efx_rx_queue *rx_queue;
613 rx_queue = efx_channel_get_rx_queue(channel);
614 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
617 skb_record_rx_queue(skb, channel->rx_queue.core_index);
619 /* Set the SKB flags */
620 skb_checksum_none_assert(skb);
621 if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED)) {
622 skb->ip_summed = CHECKSUM_UNNECESSARY;
623 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
626 efx_rx_skb_attach_timestamp(channel, skb);
628 if (channel->type->receive_skb)
629 if (channel->type->receive_skb(channel, skb))
632 /* Pass the packet up */
633 if (channel->rx_list != NULL)
634 /* Add to list, will pass up later */
635 list_add_tail(&skb->list, channel->rx_list);
637 /* No list, so pass it up now */
638 netif_receive_skb(skb);
641 /* Handle a received packet. Second half: Touches packet payload. */
642 void __efx_rx_packet(struct efx_channel *channel)
644 struct efx_nic *efx = channel->efx;
645 struct efx_rx_buffer *rx_buf =
646 efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
647 u8 *eh = efx_rx_buf_va(rx_buf);
649 /* Read length from the prefix if necessary. This already
650 * excludes the length of the prefix itself.
652 if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
653 rx_buf->len = le16_to_cpup((__le16 *)
654 (eh + efx->rx_packet_len_offset));
656 /* If we're in loopback test, then pass the packet directly to the
657 * loopback layer, and free the rx_buf here
659 if (unlikely(efx->loopback_selftest)) {
660 struct efx_rx_queue *rx_queue;
662 efx_loopback_rx_packet(efx, eh, rx_buf->len);
663 rx_queue = efx_channel_get_rx_queue(channel);
664 efx_free_rx_buffers(rx_queue, rx_buf,
665 channel->rx_pkt_n_frags);
669 if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
670 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
672 if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb)
673 efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
675 efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
677 channel->rx_pkt_n_frags = 0;
680 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
682 struct efx_nic *efx = rx_queue->efx;
683 unsigned int entries;
686 /* Create the smallest power-of-two aligned ring */
687 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
688 EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
689 rx_queue->ptr_mask = entries - 1;
691 netif_dbg(efx, probe, efx->net_dev,
692 "creating RX queue %d size %#x mask %#x\n",
693 efx_rx_queue_index(rx_queue), efx->rxq_entries,
696 /* Allocate RX buffers */
697 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
699 if (!rx_queue->buffer)
702 rc = efx_nic_probe_rx(rx_queue);
704 kfree(rx_queue->buffer);
705 rx_queue->buffer = NULL;
711 static void efx_init_rx_recycle_ring(struct efx_nic *efx,
712 struct efx_rx_queue *rx_queue)
714 unsigned int bufs_in_recycle_ring, page_ring_size;
716 /* Set the RX recycle ring size */
718 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
720 if (iommu_present(&pci_bus_type))
721 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
723 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
724 #endif /* CONFIG_PPC64 */
726 page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
727 efx->rx_bufs_per_page);
728 rx_queue->page_ring = kcalloc(page_ring_size,
729 sizeof(*rx_queue->page_ring), GFP_KERNEL);
730 rx_queue->page_ptr_mask = page_ring_size - 1;
733 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
735 struct efx_nic *efx = rx_queue->efx;
736 unsigned int max_fill, trigger, max_trigger;
738 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
739 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
741 /* Initialise ptr fields */
742 rx_queue->added_count = 0;
743 rx_queue->notified_count = 0;
744 rx_queue->removed_count = 0;
745 rx_queue->min_fill = -1U;
746 efx_init_rx_recycle_ring(efx, rx_queue);
748 rx_queue->page_remove = 0;
749 rx_queue->page_add = rx_queue->page_ptr_mask + 1;
750 rx_queue->page_recycle_count = 0;
751 rx_queue->page_recycle_failed = 0;
752 rx_queue->page_recycle_full = 0;
754 /* Initialise limit fields */
755 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
757 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
758 if (rx_refill_threshold != 0) {
759 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
760 if (trigger > max_trigger)
761 trigger = max_trigger;
763 trigger = max_trigger;
766 rx_queue->max_fill = max_fill;
767 rx_queue->fast_fill_trigger = trigger;
768 rx_queue->refill_enabled = true;
770 /* Set up RX descriptor ring */
771 efx_nic_init_rx(rx_queue);
774 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
777 struct efx_nic *efx = rx_queue->efx;
778 struct efx_rx_buffer *rx_buf;
780 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
781 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
783 del_timer_sync(&rx_queue->slow_fill);
785 /* Release RX buffers from the current read ptr to the write ptr */
786 if (rx_queue->buffer) {
787 for (i = rx_queue->removed_count; i < rx_queue->added_count;
789 unsigned index = i & rx_queue->ptr_mask;
790 rx_buf = efx_rx_buffer(rx_queue, index);
791 efx_fini_rx_buffer(rx_queue, rx_buf);
795 /* Unmap and release the pages in the recycle ring. Remove the ring. */
796 for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
797 struct page *page = rx_queue->page_ring[i];
798 struct efx_rx_page_state *state;
803 state = page_address(page);
804 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
805 PAGE_SIZE << efx->rx_buffer_order,
809 kfree(rx_queue->page_ring);
810 rx_queue->page_ring = NULL;
813 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
815 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
816 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
818 efx_nic_remove_rx(rx_queue);
820 kfree(rx_queue->buffer);
821 rx_queue->buffer = NULL;
825 module_param(rx_refill_threshold, uint, 0444);
826 MODULE_PARM_DESC(rx_refill_threshold,
827 "RX descriptor ring refill threshold (%)");
829 #ifdef CONFIG_RFS_ACCEL
831 static void efx_filter_rfs_work(struct work_struct *data)
833 struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
835 struct efx_nic *efx = netdev_priv(req->net_dev);
836 struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
837 int slot_idx = req - efx->rps_slot;
838 struct efx_arfs_rule *rule;
842 rc = efx->type->filter_insert(efx, &req->spec, true);
844 rc %= efx->type->max_rx_ip_filters;
845 if (efx->rps_hash_table) {
846 spin_lock_bh(&efx->rps_hash_lock);
847 rule = efx_rps_hash_find(efx, &req->spec);
848 /* The rule might have already gone, if someone else's request
849 * for the same spec was already worked and then expired before
850 * we got around to our work. In that case we have nothing
851 * tying us to an arfs_id, meaning that as soon as the filter
852 * is considered for expiry it will be removed.
856 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
858 rule->filter_id = rc;
859 arfs_id = rule->arfs_id;
861 spin_unlock_bh(&efx->rps_hash_lock);
864 /* Remember this so we can check whether to expire the filter
867 mutex_lock(&efx->rps_mutex);
868 channel->rps_flow_id[rc] = req->flow_id;
869 ++channel->rfs_filters_added;
870 mutex_unlock(&efx->rps_mutex);
872 if (req->spec.ether_type == htons(ETH_P_IP))
873 netif_info(efx, rx_status, efx->net_dev,
874 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
875 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
876 req->spec.rem_host, ntohs(req->spec.rem_port),
877 req->spec.loc_host, ntohs(req->spec.loc_port),
878 req->rxq_index, req->flow_id, rc, arfs_id);
880 netif_info(efx, rx_status, efx->net_dev,
881 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
882 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
883 req->spec.rem_host, ntohs(req->spec.rem_port),
884 req->spec.loc_host, ntohs(req->spec.loc_port),
885 req->rxq_index, req->flow_id, rc, arfs_id);
888 /* Release references */
889 clear_bit(slot_idx, &efx->rps_slot_map);
890 dev_put(req->net_dev);
893 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
894 u16 rxq_index, u32 flow_id)
896 struct efx_nic *efx = netdev_priv(net_dev);
897 struct efx_async_filter_insertion *req;
898 struct efx_arfs_rule *rule;
904 /* find a free slot */
905 for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
906 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
908 if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
911 if (flow_id == RPS_FLOW_ID_INVALID) {
916 if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
917 rc = -EPROTONOSUPPORT;
921 if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
922 rc = -EPROTONOSUPPORT;
925 if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
926 rc = -EPROTONOSUPPORT;
930 req = efx->rps_slot + slot_idx;
931 efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
932 efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
934 req->spec.match_flags =
935 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
936 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
937 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
938 req->spec.ether_type = fk.basic.n_proto;
939 req->spec.ip_proto = fk.basic.ip_proto;
941 if (fk.basic.n_proto == htons(ETH_P_IP)) {
942 req->spec.rem_host[0] = fk.addrs.v4addrs.src;
943 req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
945 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
946 sizeof(struct in6_addr));
947 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
948 sizeof(struct in6_addr));
951 req->spec.rem_port = fk.ports.src;
952 req->spec.loc_port = fk.ports.dst;
954 if (efx->rps_hash_table) {
955 /* Add it to ARFS hash table */
956 spin_lock(&efx->rps_hash_lock);
957 rule = efx_rps_hash_add(efx, &req->spec, &new);
963 rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
965 /* Skip if existing or pending filter already does the right thing */
966 if (!new && rule->rxq_index == rxq_index &&
967 rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
969 rule->rxq_index = rxq_index;
970 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
971 spin_unlock(&efx->rps_hash_lock);
973 /* Without an ARFS hash table, we just use arfs_id 0 for all
974 * filters. This means if multiple flows hash to the same
975 * flow_id, all but the most recently touched will be eligible
981 /* Queue the request */
982 dev_hold(req->net_dev = net_dev);
983 INIT_WORK(&req->work, efx_filter_rfs_work);
984 req->rxq_index = rxq_index;
985 req->flow_id = flow_id;
986 schedule_work(&req->work);
989 spin_unlock(&efx->rps_hash_lock);
991 clear_bit(slot_idx, &efx->rps_slot_map);
995 bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned int quota)
997 bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
998 unsigned int channel_idx, index, size;
1001 if (!mutex_trylock(&efx->rps_mutex))
1003 expire_one = efx->type->filter_rfs_expire_one;
1004 channel_idx = efx->rps_expire_channel;
1005 index = efx->rps_expire_index;
1006 size = efx->type->max_rx_ip_filters;
1008 struct efx_channel *channel = efx_get_channel(efx, channel_idx);
1009 flow_id = channel->rps_flow_id[index];
1011 if (flow_id != RPS_FLOW_ID_INVALID &&
1012 expire_one(efx, flow_id, index)) {
1013 netif_info(efx, rx_status, efx->net_dev,
1014 "expired filter %d [queue %u flow %u]\n",
1015 index, channel_idx, flow_id);
1016 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1018 if (++index == size) {
1019 if (++channel_idx == efx->n_channels)
1024 efx->rps_expire_channel = channel_idx;
1025 efx->rps_expire_index = index;
1027 mutex_unlock(&efx->rps_mutex);
1031 #endif /* CONFIG_RFS_ACCEL */
1034 * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
1035 * @spec: Specification to test
1037 * Return: %true if the specification is a non-drop RX filter that
1038 * matches a local MAC address I/G bit value of 1 or matches a local
1039 * IPv4 or IPv6 address value in the respective multicast address
1040 * range. Otherwise %false.
1042 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
1044 if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
1045 spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
1048 if (spec->match_flags &
1049 (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
1050 is_multicast_ether_addr(spec->loc_mac))
1053 if ((spec->match_flags &
1054 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
1055 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
1056 if (spec->ether_type == htons(ETH_P_IP) &&
1057 ipv4_is_multicast(spec->loc_host[0]))
1059 if (spec->ether_type == htons(ETH_P_IPV6) &&
1060 ((const u8 *)spec->loc_host)[0] == 0xff)