1 /****************************************************************************
2 * Driver for Solarflare network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2013 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 <linux/socket.h>
13 #include <linux/slab.h>
15 #include <linux/ipv6.h>
16 #include <linux/tcp.h>
17 #include <linux/udp.h>
18 #include <linux/prefetch.h>
19 #include <linux/moduleparam.h>
20 #include <linux/iommu.h>
22 #include <net/checksum.h>
23 #include "net_driver.h"
28 #include "workarounds.h"
30 /* Preferred number of descriptors to fill at once */
31 #define EFX_RX_PREFERRED_BATCH 8U
33 /* Number of RX buffers to recycle pages for. When creating the RX page recycle
34 * ring, this number is divided by the number of buffers per page to calculate
35 * the number of pages to store in the RX page recycle ring.
37 #define EFX_RECYCLE_RING_SIZE_IOMMU 4096
38 #define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
40 /* Size of buffer allocated for skb header area. */
41 #define EFX_SKB_HEADERS 128u
43 /* This is the percentage fill level below which new RX descriptors
44 * will be added to the RX descriptor ring.
46 static unsigned int rx_refill_threshold
;
48 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
49 #define EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
53 * RX maximum head room required.
55 * This must be at least 1 to prevent overflow, plus one packet-worth
56 * to allow pipelined receives.
58 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
60 static inline u8
*efx_rx_buf_va(struct efx_rx_buffer
*buf
)
62 return page_address(buf
->page
) + buf
->page_offset
;
65 static inline u32
efx_rx_buf_hash(struct efx_nic
*efx
, const u8
*eh
)
67 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
68 return __le32_to_cpup((const __le32
*)(eh
+ efx
->rx_packet_hash_offset
));
70 const u8
*data
= eh
+ efx
->rx_packet_hash_offset
;
78 static inline struct efx_rx_buffer
*
79 efx_rx_buf_next(struct efx_rx_queue
*rx_queue
, struct efx_rx_buffer
*rx_buf
)
81 if (unlikely(rx_buf
== efx_rx_buffer(rx_queue
, rx_queue
->ptr_mask
)))
82 return efx_rx_buffer(rx_queue
, 0);
87 static inline void efx_sync_rx_buffer(struct efx_nic
*efx
,
88 struct efx_rx_buffer
*rx_buf
,
91 dma_sync_single_for_cpu(&efx
->pci_dev
->dev
, rx_buf
->dma_addr
, len
,
95 void efx_rx_config_page_split(struct efx_nic
*efx
)
97 efx
->rx_page_buf_step
= ALIGN(efx
->rx_dma_len
+ efx
->rx_ip_align
,
98 EFX_RX_BUF_ALIGNMENT
);
99 efx
->rx_bufs_per_page
= efx
->rx_buffer_order
? 1 :
100 ((PAGE_SIZE
- sizeof(struct efx_rx_page_state
)) /
101 efx
->rx_page_buf_step
);
102 efx
->rx_buffer_truesize
= (PAGE_SIZE
<< efx
->rx_buffer_order
) /
103 efx
->rx_bufs_per_page
;
104 efx
->rx_pages_per_batch
= DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH
,
105 efx
->rx_bufs_per_page
);
108 /* Check the RX page recycle ring for a page that can be reused. */
109 static struct page
*efx_reuse_page(struct efx_rx_queue
*rx_queue
)
111 struct efx_nic
*efx
= rx_queue
->efx
;
113 struct efx_rx_page_state
*state
;
116 index
= rx_queue
->page_remove
& rx_queue
->page_ptr_mask
;
117 page
= rx_queue
->page_ring
[index
];
121 rx_queue
->page_ring
[index
] = NULL
;
122 /* page_remove cannot exceed page_add. */
123 if (rx_queue
->page_remove
!= rx_queue
->page_add
)
124 ++rx_queue
->page_remove
;
126 /* If page_count is 1 then we hold the only reference to this page. */
127 if (page_count(page
) == 1) {
128 ++rx_queue
->page_recycle_count
;
131 state
= page_address(page
);
132 dma_unmap_page(&efx
->pci_dev
->dev
, state
->dma_addr
,
133 PAGE_SIZE
<< efx
->rx_buffer_order
,
136 ++rx_queue
->page_recycle_failed
;
143 * efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
145 * @rx_queue: Efx RX queue
147 * This allocates a batch of pages, maps them for DMA, and populates
148 * struct efx_rx_buffers for each one. Return a negative error code or
149 * 0 on success. If a single page can be used for multiple buffers,
150 * then the page will either be inserted fully, or not at all.
152 static int efx_init_rx_buffers(struct efx_rx_queue
*rx_queue
, bool atomic
)
154 struct efx_nic
*efx
= rx_queue
->efx
;
155 struct efx_rx_buffer
*rx_buf
;
157 unsigned int page_offset
;
158 struct efx_rx_page_state
*state
;
160 unsigned index
, count
;
164 page
= efx_reuse_page(rx_queue
);
166 page
= alloc_pages(__GFP_COMP
|
167 (atomic
? GFP_ATOMIC
: GFP_KERNEL
),
168 efx
->rx_buffer_order
);
169 if (unlikely(page
== NULL
))
172 dma_map_page(&efx
->pci_dev
->dev
, page
, 0,
173 PAGE_SIZE
<< efx
->rx_buffer_order
,
175 if (unlikely(dma_mapping_error(&efx
->pci_dev
->dev
,
177 __free_pages(page
, efx
->rx_buffer_order
);
180 state
= page_address(page
);
181 state
->dma_addr
= dma_addr
;
183 state
= page_address(page
);
184 dma_addr
= state
->dma_addr
;
187 dma_addr
+= sizeof(struct efx_rx_page_state
);
188 page_offset
= sizeof(struct efx_rx_page_state
);
191 index
= rx_queue
->added_count
& rx_queue
->ptr_mask
;
192 rx_buf
= efx_rx_buffer(rx_queue
, index
);
193 rx_buf
->dma_addr
= dma_addr
+ efx
->rx_ip_align
;
195 rx_buf
->page_offset
= page_offset
+ efx
->rx_ip_align
;
196 rx_buf
->len
= efx
->rx_dma_len
;
198 ++rx_queue
->added_count
;
200 dma_addr
+= efx
->rx_page_buf_step
;
201 page_offset
+= efx
->rx_page_buf_step
;
202 } while (page_offset
+ efx
->rx_page_buf_step
<= PAGE_SIZE
);
204 rx_buf
->flags
= EFX_RX_BUF_LAST_IN_PAGE
;
205 } while (++count
< efx
->rx_pages_per_batch
);
210 /* Unmap a DMA-mapped page. This function is only called for the final RX
213 static void efx_unmap_rx_buffer(struct efx_nic
*efx
,
214 struct efx_rx_buffer
*rx_buf
)
216 struct page
*page
= rx_buf
->page
;
219 struct efx_rx_page_state
*state
= page_address(page
);
220 dma_unmap_page(&efx
->pci_dev
->dev
,
222 PAGE_SIZE
<< efx
->rx_buffer_order
,
227 static void efx_free_rx_buffers(struct efx_rx_queue
*rx_queue
,
228 struct efx_rx_buffer
*rx_buf
,
229 unsigned int num_bufs
)
233 put_page(rx_buf
->page
);
236 rx_buf
= efx_rx_buf_next(rx_queue
, rx_buf
);
237 } while (--num_bufs
);
240 /* Attempt to recycle the page if there is an RX recycle ring; the page can
241 * only be added if this is the final RX buffer, to prevent pages being used in
242 * the descriptor ring and appearing in the recycle ring simultaneously.
244 static void efx_recycle_rx_page(struct efx_channel
*channel
,
245 struct efx_rx_buffer
*rx_buf
)
247 struct page
*page
= rx_buf
->page
;
248 struct efx_rx_queue
*rx_queue
= efx_channel_get_rx_queue(channel
);
249 struct efx_nic
*efx
= rx_queue
->efx
;
252 /* Only recycle the page after processing the final buffer. */
253 if (!(rx_buf
->flags
& EFX_RX_BUF_LAST_IN_PAGE
))
256 index
= rx_queue
->page_add
& rx_queue
->page_ptr_mask
;
257 if (rx_queue
->page_ring
[index
] == NULL
) {
258 unsigned read_index
= rx_queue
->page_remove
&
259 rx_queue
->page_ptr_mask
;
261 /* The next slot in the recycle ring is available, but
262 * increment page_remove if the read pointer currently
265 if (read_index
== index
)
266 ++rx_queue
->page_remove
;
267 rx_queue
->page_ring
[index
] = page
;
268 ++rx_queue
->page_add
;
271 ++rx_queue
->page_recycle_full
;
272 efx_unmap_rx_buffer(efx
, rx_buf
);
273 put_page(rx_buf
->page
);
276 static void efx_fini_rx_buffer(struct efx_rx_queue
*rx_queue
,
277 struct efx_rx_buffer
*rx_buf
)
279 /* Release the page reference we hold for the buffer. */
281 put_page(rx_buf
->page
);
283 /* If this is the last buffer in a page, unmap and free it. */
284 if (rx_buf
->flags
& EFX_RX_BUF_LAST_IN_PAGE
) {
285 efx_unmap_rx_buffer(rx_queue
->efx
, rx_buf
);
286 efx_free_rx_buffers(rx_queue
, rx_buf
, 1);
291 /* Recycle the pages that are used by buffers that have just been received. */
292 static void efx_recycle_rx_pages(struct efx_channel
*channel
,
293 struct efx_rx_buffer
*rx_buf
,
294 unsigned int n_frags
)
296 struct efx_rx_queue
*rx_queue
= efx_channel_get_rx_queue(channel
);
299 efx_recycle_rx_page(channel
, rx_buf
);
300 rx_buf
= efx_rx_buf_next(rx_queue
, rx_buf
);
304 static void efx_discard_rx_packet(struct efx_channel
*channel
,
305 struct efx_rx_buffer
*rx_buf
,
306 unsigned int n_frags
)
308 struct efx_rx_queue
*rx_queue
= efx_channel_get_rx_queue(channel
);
310 efx_recycle_rx_pages(channel
, rx_buf
, n_frags
);
312 efx_free_rx_buffers(rx_queue
, rx_buf
, n_frags
);
316 * efx_fast_push_rx_descriptors - push new RX descriptors quickly
317 * @rx_queue: RX descriptor queue
319 * This will aim to fill the RX descriptor queue up to
320 * @rx_queue->@max_fill. If there is insufficient atomic
321 * memory to do so, a slow fill will be scheduled.
323 * The caller must provide serialisation (none is used here). In practise,
324 * this means this function must run from the NAPI handler, or be called
325 * when NAPI is disabled.
327 void efx_fast_push_rx_descriptors(struct efx_rx_queue
*rx_queue
, bool atomic
)
329 struct efx_nic
*efx
= rx_queue
->efx
;
330 unsigned int fill_level
, batch_size
;
333 if (!rx_queue
->refill_enabled
)
336 /* Calculate current fill level, and exit if we don't need to fill */
337 fill_level
= (rx_queue
->added_count
- rx_queue
->removed_count
);
338 EFX_WARN_ON_ONCE_PARANOID(fill_level
> rx_queue
->efx
->rxq_entries
);
339 if (fill_level
>= rx_queue
->fast_fill_trigger
)
342 /* Record minimum fill level */
343 if (unlikely(fill_level
< rx_queue
->min_fill
)) {
345 rx_queue
->min_fill
= fill_level
;
348 batch_size
= efx
->rx_pages_per_batch
* efx
->rx_bufs_per_page
;
349 space
= rx_queue
->max_fill
- fill_level
;
350 EFX_WARN_ON_ONCE_PARANOID(space
< batch_size
);
352 netif_vdbg(rx_queue
->efx
, rx_status
, rx_queue
->efx
->net_dev
,
353 "RX queue %d fast-filling descriptor ring from"
354 " level %d to level %d\n",
355 efx_rx_queue_index(rx_queue
), fill_level
,
360 rc
= efx_init_rx_buffers(rx_queue
, atomic
);
362 /* Ensure that we don't leave the rx queue empty */
363 efx_schedule_slow_fill(rx_queue
);
366 } while ((space
-= batch_size
) >= batch_size
);
368 netif_vdbg(rx_queue
->efx
, rx_status
, rx_queue
->efx
->net_dev
,
369 "RX queue %d fast-filled descriptor ring "
370 "to level %d\n", efx_rx_queue_index(rx_queue
),
371 rx_queue
->added_count
- rx_queue
->removed_count
);
374 if (rx_queue
->notified_count
!= rx_queue
->added_count
)
375 efx_nic_notify_rx_desc(rx_queue
);
378 void efx_rx_slow_fill(struct timer_list
*t
)
380 struct efx_rx_queue
*rx_queue
= from_timer(rx_queue
, t
, slow_fill
);
382 /* Post an event to cause NAPI to run and refill the queue */
383 efx_nic_generate_fill_event(rx_queue
);
384 ++rx_queue
->slow_fill_count
;
387 static void efx_rx_packet__check_len(struct efx_rx_queue
*rx_queue
,
388 struct efx_rx_buffer
*rx_buf
,
391 struct efx_nic
*efx
= rx_queue
->efx
;
392 unsigned max_len
= rx_buf
->len
- efx
->type
->rx_buffer_padding
;
394 if (likely(len
<= max_len
))
397 /* The packet must be discarded, but this is only a fatal error
398 * if the caller indicated it was
400 rx_buf
->flags
|= EFX_RX_PKT_DISCARD
;
403 netif_err(efx
, rx_err
, efx
->net_dev
,
404 "RX queue %d overlength RX event (%#x > %#x)\n",
405 efx_rx_queue_index(rx_queue
), len
, max_len
);
407 efx_rx_queue_channel(rx_queue
)->n_rx_overlength
++;
410 /* Pass a received packet up through GRO. GRO can handle pages
411 * regardless of checksum state and skbs with a good checksum.
414 efx_rx_packet_gro(struct efx_channel
*channel
, struct efx_rx_buffer
*rx_buf
,
415 unsigned int n_frags
, u8
*eh
)
417 struct napi_struct
*napi
= &channel
->napi_str
;
418 gro_result_t gro_result
;
419 struct efx_nic
*efx
= channel
->efx
;
422 skb
= napi_get_frags(napi
);
423 if (unlikely(!skb
)) {
424 struct efx_rx_queue
*rx_queue
;
426 rx_queue
= efx_channel_get_rx_queue(channel
);
427 efx_free_rx_buffers(rx_queue
, rx_buf
, n_frags
);
431 if (efx
->net_dev
->features
& NETIF_F_RXHASH
)
432 skb_set_hash(skb
, efx_rx_buf_hash(efx
, eh
),
434 skb
->ip_summed
= ((rx_buf
->flags
& EFX_RX_PKT_CSUMMED
) ?
435 CHECKSUM_UNNECESSARY
: CHECKSUM_NONE
);
436 skb
->csum_level
= !!(rx_buf
->flags
& EFX_RX_PKT_CSUM_LEVEL
);
439 skb_fill_page_desc(skb
, skb_shinfo(skb
)->nr_frags
,
440 rx_buf
->page
, rx_buf
->page_offset
,
443 skb
->len
+= rx_buf
->len
;
444 if (skb_shinfo(skb
)->nr_frags
== n_frags
)
447 rx_buf
= efx_rx_buf_next(&channel
->rx_queue
, rx_buf
);
450 skb
->data_len
= skb
->len
;
451 skb
->truesize
+= n_frags
* efx
->rx_buffer_truesize
;
453 skb_record_rx_queue(skb
, channel
->rx_queue
.core_index
);
455 gro_result
= napi_gro_frags(napi
);
456 if (gro_result
!= GRO_DROP
)
457 channel
->irq_mod_score
+= 2;
460 /* Allocate and construct an SKB around page fragments */
461 static struct sk_buff
*efx_rx_mk_skb(struct efx_channel
*channel
,
462 struct efx_rx_buffer
*rx_buf
,
463 unsigned int n_frags
,
466 struct efx_nic
*efx
= channel
->efx
;
469 /* Allocate an SKB to store the headers */
470 skb
= netdev_alloc_skb(efx
->net_dev
,
471 efx
->rx_ip_align
+ efx
->rx_prefix_size
+
473 if (unlikely(skb
== NULL
)) {
474 atomic_inc(&efx
->n_rx_noskb_drops
);
478 EFX_WARN_ON_ONCE_PARANOID(rx_buf
->len
< hdr_len
);
480 memcpy(skb
->data
+ efx
->rx_ip_align
, eh
- efx
->rx_prefix_size
,
481 efx
->rx_prefix_size
+ hdr_len
);
482 skb_reserve(skb
, efx
->rx_ip_align
+ efx
->rx_prefix_size
);
483 __skb_put(skb
, hdr_len
);
485 /* Append the remaining page(s) onto the frag list */
486 if (rx_buf
->len
> hdr_len
) {
487 rx_buf
->page_offset
+= hdr_len
;
488 rx_buf
->len
-= hdr_len
;
491 skb_fill_page_desc(skb
, skb_shinfo(skb
)->nr_frags
,
492 rx_buf
->page
, rx_buf
->page_offset
,
495 skb
->len
+= rx_buf
->len
;
496 skb
->data_len
+= rx_buf
->len
;
497 if (skb_shinfo(skb
)->nr_frags
== n_frags
)
500 rx_buf
= efx_rx_buf_next(&channel
->rx_queue
, rx_buf
);
503 __free_pages(rx_buf
->page
, efx
->rx_buffer_order
);
508 skb
->truesize
+= n_frags
* efx
->rx_buffer_truesize
;
510 /* Move past the ethernet header */
511 skb
->protocol
= eth_type_trans(skb
, efx
->net_dev
);
513 skb_mark_napi_id(skb
, &channel
->napi_str
);
518 void efx_rx_packet(struct efx_rx_queue
*rx_queue
, unsigned int index
,
519 unsigned int n_frags
, unsigned int len
, u16 flags
)
521 struct efx_nic
*efx
= rx_queue
->efx
;
522 struct efx_channel
*channel
= efx_rx_queue_channel(rx_queue
);
523 struct efx_rx_buffer
*rx_buf
;
525 rx_queue
->rx_packets
++;
527 rx_buf
= efx_rx_buffer(rx_queue
, index
);
528 rx_buf
->flags
|= flags
;
530 /* Validate the number of fragments and completed length */
532 if (!(flags
& EFX_RX_PKT_PREFIX_LEN
))
533 efx_rx_packet__check_len(rx_queue
, rx_buf
, len
);
534 } else if (unlikely(n_frags
> EFX_RX_MAX_FRAGS
) ||
535 unlikely(len
<= (n_frags
- 1) * efx
->rx_dma_len
) ||
536 unlikely(len
> n_frags
* efx
->rx_dma_len
) ||
537 unlikely(!efx
->rx_scatter
)) {
538 /* If this isn't an explicit discard request, either
539 * the hardware or the driver is broken.
541 WARN_ON(!(len
== 0 && rx_buf
->flags
& EFX_RX_PKT_DISCARD
));
542 rx_buf
->flags
|= EFX_RX_PKT_DISCARD
;
545 netif_vdbg(efx
, rx_status
, efx
->net_dev
,
546 "RX queue %d received ids %x-%x len %d %s%s\n",
547 efx_rx_queue_index(rx_queue
), index
,
548 (index
+ n_frags
- 1) & rx_queue
->ptr_mask
, len
,
549 (rx_buf
->flags
& EFX_RX_PKT_CSUMMED
) ? " [SUMMED]" : "",
550 (rx_buf
->flags
& EFX_RX_PKT_DISCARD
) ? " [DISCARD]" : "");
552 /* Discard packet, if instructed to do so. Process the
553 * previous receive first.
555 if (unlikely(rx_buf
->flags
& EFX_RX_PKT_DISCARD
)) {
556 efx_rx_flush_packet(channel
);
557 efx_discard_rx_packet(channel
, rx_buf
, n_frags
);
561 if (n_frags
== 1 && !(flags
& EFX_RX_PKT_PREFIX_LEN
))
564 /* Release and/or sync the DMA mapping - assumes all RX buffers
565 * consumed in-order per RX queue.
567 efx_sync_rx_buffer(efx
, rx_buf
, rx_buf
->len
);
569 /* Prefetch nice and early so data will (hopefully) be in cache by
570 * the time we look at it.
572 prefetch(efx_rx_buf_va(rx_buf
));
574 rx_buf
->page_offset
+= efx
->rx_prefix_size
;
575 rx_buf
->len
-= efx
->rx_prefix_size
;
578 /* Release/sync DMA mapping for additional fragments.
579 * Fix length for last fragment.
581 unsigned int tail_frags
= n_frags
- 1;
584 rx_buf
= efx_rx_buf_next(rx_queue
, rx_buf
);
585 if (--tail_frags
== 0)
587 efx_sync_rx_buffer(efx
, rx_buf
, efx
->rx_dma_len
);
589 rx_buf
->len
= len
- (n_frags
- 1) * efx
->rx_dma_len
;
590 efx_sync_rx_buffer(efx
, rx_buf
, rx_buf
->len
);
593 /* All fragments have been DMA-synced, so recycle pages. */
594 rx_buf
= efx_rx_buffer(rx_queue
, index
);
595 efx_recycle_rx_pages(channel
, rx_buf
, n_frags
);
597 /* Pipeline receives so that we give time for packet headers to be
598 * prefetched into cache.
600 efx_rx_flush_packet(channel
);
601 channel
->rx_pkt_n_frags
= n_frags
;
602 channel
->rx_pkt_index
= index
;
605 static void efx_rx_deliver(struct efx_channel
*channel
, u8
*eh
,
606 struct efx_rx_buffer
*rx_buf
,
607 unsigned int n_frags
)
610 u16 hdr_len
= min_t(u16
, rx_buf
->len
, EFX_SKB_HEADERS
);
612 skb
= efx_rx_mk_skb(channel
, rx_buf
, n_frags
, eh
, hdr_len
);
613 if (unlikely(skb
== NULL
)) {
614 struct efx_rx_queue
*rx_queue
;
616 rx_queue
= efx_channel_get_rx_queue(channel
);
617 efx_free_rx_buffers(rx_queue
, rx_buf
, n_frags
);
620 skb_record_rx_queue(skb
, channel
->rx_queue
.core_index
);
622 /* Set the SKB flags */
623 skb_checksum_none_assert(skb
);
624 if (likely(rx_buf
->flags
& EFX_RX_PKT_CSUMMED
)) {
625 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
626 skb
->csum_level
= !!(rx_buf
->flags
& EFX_RX_PKT_CSUM_LEVEL
);
629 efx_rx_skb_attach_timestamp(channel
, skb
);
631 if (channel
->type
->receive_skb
)
632 if (channel
->type
->receive_skb(channel
, skb
))
635 /* Pass the packet up */
636 if (channel
->rx_list
!= NULL
)
637 /* Add to list, will pass up later */
638 list_add_tail(&skb
->list
, channel
->rx_list
);
640 /* No list, so pass it up now */
641 netif_receive_skb(skb
);
644 /* Handle a received packet. Second half: Touches packet payload. */
645 void __efx_rx_packet(struct efx_channel
*channel
)
647 struct efx_nic
*efx
= channel
->efx
;
648 struct efx_rx_buffer
*rx_buf
=
649 efx_rx_buffer(&channel
->rx_queue
, channel
->rx_pkt_index
);
650 u8
*eh
= efx_rx_buf_va(rx_buf
);
652 /* Read length from the prefix if necessary. This already
653 * excludes the length of the prefix itself.
655 if (rx_buf
->flags
& EFX_RX_PKT_PREFIX_LEN
)
656 rx_buf
->len
= le16_to_cpup((__le16
*)
657 (eh
+ efx
->rx_packet_len_offset
));
659 /* If we're in loopback test, then pass the packet directly to the
660 * loopback layer, and free the rx_buf here
662 if (unlikely(efx
->loopback_selftest
)) {
663 struct efx_rx_queue
*rx_queue
;
665 efx_loopback_rx_packet(efx
, eh
, rx_buf
->len
);
666 rx_queue
= efx_channel_get_rx_queue(channel
);
667 efx_free_rx_buffers(rx_queue
, rx_buf
,
668 channel
->rx_pkt_n_frags
);
672 if (unlikely(!(efx
->net_dev
->features
& NETIF_F_RXCSUM
)))
673 rx_buf
->flags
&= ~EFX_RX_PKT_CSUMMED
;
675 if ((rx_buf
->flags
& EFX_RX_PKT_TCP
) && !channel
->type
->receive_skb
)
676 efx_rx_packet_gro(channel
, rx_buf
, channel
->rx_pkt_n_frags
, eh
);
678 efx_rx_deliver(channel
, eh
, rx_buf
, channel
->rx_pkt_n_frags
);
680 channel
->rx_pkt_n_frags
= 0;
683 int efx_probe_rx_queue(struct efx_rx_queue
*rx_queue
)
685 struct efx_nic
*efx
= rx_queue
->efx
;
686 unsigned int entries
;
689 /* Create the smallest power-of-two aligned ring */
690 entries
= max(roundup_pow_of_two(efx
->rxq_entries
), EFX_MIN_DMAQ_SIZE
);
691 EFX_WARN_ON_PARANOID(entries
> EFX_MAX_DMAQ_SIZE
);
692 rx_queue
->ptr_mask
= entries
- 1;
694 netif_dbg(efx
, probe
, efx
->net_dev
,
695 "creating RX queue %d size %#x mask %#x\n",
696 efx_rx_queue_index(rx_queue
), efx
->rxq_entries
,
699 /* Allocate RX buffers */
700 rx_queue
->buffer
= kcalloc(entries
, sizeof(*rx_queue
->buffer
),
702 if (!rx_queue
->buffer
)
705 rc
= efx_nic_probe_rx(rx_queue
);
707 kfree(rx_queue
->buffer
);
708 rx_queue
->buffer
= NULL
;
714 static void efx_init_rx_recycle_ring(struct efx_nic
*efx
,
715 struct efx_rx_queue
*rx_queue
)
717 unsigned int bufs_in_recycle_ring
, page_ring_size
;
719 /* Set the RX recycle ring size */
721 bufs_in_recycle_ring
= EFX_RECYCLE_RING_SIZE_IOMMU
;
723 if (iommu_present(&pci_bus_type
))
724 bufs_in_recycle_ring
= EFX_RECYCLE_RING_SIZE_IOMMU
;
726 bufs_in_recycle_ring
= EFX_RECYCLE_RING_SIZE_NOIOMMU
;
727 #endif /* CONFIG_PPC64 */
729 page_ring_size
= roundup_pow_of_two(bufs_in_recycle_ring
/
730 efx
->rx_bufs_per_page
);
731 rx_queue
->page_ring
= kcalloc(page_ring_size
,
732 sizeof(*rx_queue
->page_ring
), GFP_KERNEL
);
733 rx_queue
->page_ptr_mask
= page_ring_size
- 1;
736 void efx_init_rx_queue(struct efx_rx_queue
*rx_queue
)
738 struct efx_nic
*efx
= rx_queue
->efx
;
739 unsigned int max_fill
, trigger
, max_trigger
;
741 netif_dbg(rx_queue
->efx
, drv
, rx_queue
->efx
->net_dev
,
742 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue
));
744 /* Initialise ptr fields */
745 rx_queue
->added_count
= 0;
746 rx_queue
->notified_count
= 0;
747 rx_queue
->removed_count
= 0;
748 rx_queue
->min_fill
= -1U;
749 efx_init_rx_recycle_ring(efx
, rx_queue
);
751 rx_queue
->page_remove
= 0;
752 rx_queue
->page_add
= rx_queue
->page_ptr_mask
+ 1;
753 rx_queue
->page_recycle_count
= 0;
754 rx_queue
->page_recycle_failed
= 0;
755 rx_queue
->page_recycle_full
= 0;
757 /* Initialise limit fields */
758 max_fill
= efx
->rxq_entries
- EFX_RXD_HEAD_ROOM
;
760 max_fill
- efx
->rx_pages_per_batch
* efx
->rx_bufs_per_page
;
761 if (rx_refill_threshold
!= 0) {
762 trigger
= max_fill
* min(rx_refill_threshold
, 100U) / 100U;
763 if (trigger
> max_trigger
)
764 trigger
= max_trigger
;
766 trigger
= max_trigger
;
769 rx_queue
->max_fill
= max_fill
;
770 rx_queue
->fast_fill_trigger
= trigger
;
771 rx_queue
->refill_enabled
= true;
773 /* Set up RX descriptor ring */
774 efx_nic_init_rx(rx_queue
);
777 void efx_fini_rx_queue(struct efx_rx_queue
*rx_queue
)
780 struct efx_nic
*efx
= rx_queue
->efx
;
781 struct efx_rx_buffer
*rx_buf
;
783 netif_dbg(rx_queue
->efx
, drv
, rx_queue
->efx
->net_dev
,
784 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue
));
786 del_timer_sync(&rx_queue
->slow_fill
);
788 /* Release RX buffers from the current read ptr to the write ptr */
789 if (rx_queue
->buffer
) {
790 for (i
= rx_queue
->removed_count
; i
< rx_queue
->added_count
;
792 unsigned index
= i
& rx_queue
->ptr_mask
;
793 rx_buf
= efx_rx_buffer(rx_queue
, index
);
794 efx_fini_rx_buffer(rx_queue
, rx_buf
);
798 /* Unmap and release the pages in the recycle ring. Remove the ring. */
799 for (i
= 0; i
<= rx_queue
->page_ptr_mask
; i
++) {
800 struct page
*page
= rx_queue
->page_ring
[i
];
801 struct efx_rx_page_state
*state
;
806 state
= page_address(page
);
807 dma_unmap_page(&efx
->pci_dev
->dev
, state
->dma_addr
,
808 PAGE_SIZE
<< efx
->rx_buffer_order
,
812 kfree(rx_queue
->page_ring
);
813 rx_queue
->page_ring
= NULL
;
816 void efx_remove_rx_queue(struct efx_rx_queue
*rx_queue
)
818 netif_dbg(rx_queue
->efx
, drv
, rx_queue
->efx
->net_dev
,
819 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue
));
821 efx_nic_remove_rx(rx_queue
);
823 kfree(rx_queue
->buffer
);
824 rx_queue
->buffer
= NULL
;
828 module_param(rx_refill_threshold
, uint
, 0444);
829 MODULE_PARM_DESC(rx_refill_threshold
,
830 "RX descriptor ring refill threshold (%)");
832 #ifdef CONFIG_RFS_ACCEL
834 static void efx_filter_rfs_work(struct work_struct
*data
)
836 struct efx_async_filter_insertion
*req
= container_of(data
, struct efx_async_filter_insertion
,
838 struct efx_nic
*efx
= netdev_priv(req
->net_dev
);
839 struct efx_channel
*channel
= efx_get_channel(efx
, req
->rxq_index
);
840 int slot_idx
= req
- efx
->rps_slot
;
841 struct efx_arfs_rule
*rule
;
845 rc
= efx
->type
->filter_insert(efx
, &req
->spec
, true);
847 rc
%= efx
->type
->max_rx_ip_filters
;
848 if (efx
->rps_hash_table
) {
849 spin_lock_bh(&efx
->rps_hash_lock
);
850 rule
= efx_rps_hash_find(efx
, &req
->spec
);
851 /* The rule might have already gone, if someone else's request
852 * for the same spec was already worked and then expired before
853 * we got around to our work. In that case we have nothing
854 * tying us to an arfs_id, meaning that as soon as the filter
855 * is considered for expiry it will be removed.
859 rule
->filter_id
= EFX_ARFS_FILTER_ID_ERROR
;
861 rule
->filter_id
= rc
;
862 arfs_id
= rule
->arfs_id
;
864 spin_unlock_bh(&efx
->rps_hash_lock
);
867 /* Remember this so we can check whether to expire the filter
870 mutex_lock(&efx
->rps_mutex
);
871 channel
->rps_flow_id
[rc
] = req
->flow_id
;
872 ++channel
->rfs_filters_added
;
873 mutex_unlock(&efx
->rps_mutex
);
875 if (req
->spec
.ether_type
== htons(ETH_P_IP
))
876 netif_info(efx
, rx_status
, efx
->net_dev
,
877 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
878 (req
->spec
.ip_proto
== IPPROTO_TCP
) ? "TCP" : "UDP",
879 req
->spec
.rem_host
, ntohs(req
->spec
.rem_port
),
880 req
->spec
.loc_host
, ntohs(req
->spec
.loc_port
),
881 req
->rxq_index
, req
->flow_id
, rc
, arfs_id
);
883 netif_info(efx
, rx_status
, efx
->net_dev
,
884 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
885 (req
->spec
.ip_proto
== IPPROTO_TCP
) ? "TCP" : "UDP",
886 req
->spec
.rem_host
, ntohs(req
->spec
.rem_port
),
887 req
->spec
.loc_host
, ntohs(req
->spec
.loc_port
),
888 req
->rxq_index
, req
->flow_id
, rc
, arfs_id
);
891 /* Release references */
892 clear_bit(slot_idx
, &efx
->rps_slot_map
);
893 dev_put(req
->net_dev
);
896 int efx_filter_rfs(struct net_device
*net_dev
, const struct sk_buff
*skb
,
897 u16 rxq_index
, u32 flow_id
)
899 struct efx_nic
*efx
= netdev_priv(net_dev
);
900 struct efx_async_filter_insertion
*req
;
901 struct efx_arfs_rule
*rule
;
907 /* find a free slot */
908 for (slot_idx
= 0; slot_idx
< EFX_RPS_MAX_IN_FLIGHT
; slot_idx
++)
909 if (!test_and_set_bit(slot_idx
, &efx
->rps_slot_map
))
911 if (slot_idx
>= EFX_RPS_MAX_IN_FLIGHT
)
914 if (flow_id
== RPS_FLOW_ID_INVALID
) {
919 if (!skb_flow_dissect_flow_keys(skb
, &fk
, 0)) {
920 rc
= -EPROTONOSUPPORT
;
924 if (fk
.basic
.n_proto
!= htons(ETH_P_IP
) && fk
.basic
.n_proto
!= htons(ETH_P_IPV6
)) {
925 rc
= -EPROTONOSUPPORT
;
928 if (fk
.control
.flags
& FLOW_DIS_IS_FRAGMENT
) {
929 rc
= -EPROTONOSUPPORT
;
933 req
= efx
->rps_slot
+ slot_idx
;
934 efx_filter_init_rx(&req
->spec
, EFX_FILTER_PRI_HINT
,
935 efx
->rx_scatter
? EFX_FILTER_FLAG_RX_SCATTER
: 0,
937 req
->spec
.match_flags
=
938 EFX_FILTER_MATCH_ETHER_TYPE
| EFX_FILTER_MATCH_IP_PROTO
|
939 EFX_FILTER_MATCH_LOC_HOST
| EFX_FILTER_MATCH_LOC_PORT
|
940 EFX_FILTER_MATCH_REM_HOST
| EFX_FILTER_MATCH_REM_PORT
;
941 req
->spec
.ether_type
= fk
.basic
.n_proto
;
942 req
->spec
.ip_proto
= fk
.basic
.ip_proto
;
944 if (fk
.basic
.n_proto
== htons(ETH_P_IP
)) {
945 req
->spec
.rem_host
[0] = fk
.addrs
.v4addrs
.src
;
946 req
->spec
.loc_host
[0] = fk
.addrs
.v4addrs
.dst
;
948 memcpy(req
->spec
.rem_host
, &fk
.addrs
.v6addrs
.src
,
949 sizeof(struct in6_addr
));
950 memcpy(req
->spec
.loc_host
, &fk
.addrs
.v6addrs
.dst
,
951 sizeof(struct in6_addr
));
954 req
->spec
.rem_port
= fk
.ports
.src
;
955 req
->spec
.loc_port
= fk
.ports
.dst
;
957 if (efx
->rps_hash_table
) {
958 /* Add it to ARFS hash table */
959 spin_lock(&efx
->rps_hash_lock
);
960 rule
= efx_rps_hash_add(efx
, &req
->spec
, &new);
966 rule
->arfs_id
= efx
->rps_next_id
++ % RPS_NO_FILTER
;
968 /* Skip if existing or pending filter already does the right thing */
969 if (!new && rule
->rxq_index
== rxq_index
&&
970 rule
->filter_id
>= EFX_ARFS_FILTER_ID_PENDING
)
972 rule
->rxq_index
= rxq_index
;
973 rule
->filter_id
= EFX_ARFS_FILTER_ID_PENDING
;
974 spin_unlock(&efx
->rps_hash_lock
);
976 /* Without an ARFS hash table, we just use arfs_id 0 for all
977 * filters. This means if multiple flows hash to the same
978 * flow_id, all but the most recently touched will be eligible
984 /* Queue the request */
985 dev_hold(req
->net_dev
= net_dev
);
986 INIT_WORK(&req
->work
, efx_filter_rfs_work
);
987 req
->rxq_index
= rxq_index
;
988 req
->flow_id
= flow_id
;
989 schedule_work(&req
->work
);
992 spin_unlock(&efx
->rps_hash_lock
);
994 clear_bit(slot_idx
, &efx
->rps_slot_map
);
998 bool __efx_filter_rfs_expire(struct efx_nic
*efx
, unsigned int quota
)
1000 bool (*expire_one
)(struct efx_nic
*efx
, u32 flow_id
, unsigned int index
);
1001 unsigned int channel_idx
, index
, size
;
1004 if (!mutex_trylock(&efx
->rps_mutex
))
1006 expire_one
= efx
->type
->filter_rfs_expire_one
;
1007 channel_idx
= efx
->rps_expire_channel
;
1008 index
= efx
->rps_expire_index
;
1009 size
= efx
->type
->max_rx_ip_filters
;
1011 struct efx_channel
*channel
= efx_get_channel(efx
, channel_idx
);
1012 flow_id
= channel
->rps_flow_id
[index
];
1014 if (flow_id
!= RPS_FLOW_ID_INVALID
&&
1015 expire_one(efx
, flow_id
, index
)) {
1016 netif_info(efx
, rx_status
, efx
->net_dev
,
1017 "expired filter %d [queue %u flow %u]\n",
1018 index
, channel_idx
, flow_id
);
1019 channel
->rps_flow_id
[index
] = RPS_FLOW_ID_INVALID
;
1021 if (++index
== size
) {
1022 if (++channel_idx
== efx
->n_channels
)
1027 efx
->rps_expire_channel
= channel_idx
;
1028 efx
->rps_expire_index
= index
;
1030 mutex_unlock(&efx
->rps_mutex
);
1034 #endif /* CONFIG_RFS_ACCEL */
1037 * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
1038 * @spec: Specification to test
1040 * Return: %true if the specification is a non-drop RX filter that
1041 * matches a local MAC address I/G bit value of 1 or matches a local
1042 * IPv4 or IPv6 address value in the respective multicast address
1043 * range. Otherwise %false.
1045 bool efx_filter_is_mc_recipient(const struct efx_filter_spec
*spec
)
1047 if (!(spec
->flags
& EFX_FILTER_FLAG_RX
) ||
1048 spec
->dmaq_id
== EFX_FILTER_RX_DMAQ_ID_DROP
)
1051 if (spec
->match_flags
&
1052 (EFX_FILTER_MATCH_LOC_MAC
| EFX_FILTER_MATCH_LOC_MAC_IG
) &&
1053 is_multicast_ether_addr(spec
->loc_mac
))
1056 if ((spec
->match_flags
&
1057 (EFX_FILTER_MATCH_ETHER_TYPE
| EFX_FILTER_MATCH_LOC_HOST
)) ==
1058 (EFX_FILTER_MATCH_ETHER_TYPE
| EFX_FILTER_MATCH_LOC_HOST
)) {
1059 if (spec
->ether_type
== htons(ETH_P_IP
) &&
1060 ipv4_is_multicast(spec
->loc_host
[0]))
1062 if (spec
->ether_type
== htons(ETH_P_IPV6
) &&
1063 ((const u8
*)spec
->loc_host
)[0] == 0xff)