1 /****************************************************************************
2 * Driver for Solarflare network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2013 Solarflare Communications Inc.
5 *
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.
9 */
10
11 #include <linux/socket.h>
12 #include <linux/in.h>
13 #include <linux/slab.h>
14 #include <linux/ip.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>
21 #include <net/ip.h>
22 #include <net/checksum.h>
23 #include "net_driver.h"
24 #include "efx.h"
25 #include "filter.h"
26 #include "nic.h"
27 #include "selftest.h"
28 #include "workarounds.h"
29
30 /* Preferred number of descriptors to fill at once */
31 #define EFX_RX_PREFERRED_BATCH 8U
32
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.
36 */
37 #define EFX_RECYCLE_RING_SIZE_IOMMU 4096
38 #define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
39
40 /* Size of buffer allocated for skb header area. */
41 #define EFX_SKB_HEADERS 128u
42
43 /* This is the percentage fill level below which new RX descriptors
44 * will be added to the RX descriptor ring.
45 */
46 static unsigned int rx_refill_threshold;
47
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), \
50 EFX_RX_USR_BUF_SIZE)
51
52 /*
53 * RX maximum head room required.
54 *
55 * This must be at least 1 to prevent overflow, plus one packet-worth
56 * to allow pipelined receives.
57 */
58 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
59
efx_rx_buf_va(struct efx_rx_buffer * buf)60 static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
61 {
62 return page_address(buf->page) + buf->page_offset;
63 }
64
efx_rx_buf_hash(struct efx_nic * efx,const u8 * eh)65 static inline u32 efx_rx_buf_hash(struct efx_nic *efx, const u8 *eh)
66 {
67 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
68 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_hash_offset));
69 #else
70 const u8 *data = eh + efx->rx_packet_hash_offset;
71 return (u32)data[0] |
72 (u32)data[1] << 8 |
73 (u32)data[2] << 16 |
74 (u32)data[3] << 24;
75 #endif
76 }
77
78 static inline struct efx_rx_buffer *
efx_rx_buf_next(struct efx_rx_queue * rx_queue,struct efx_rx_buffer * rx_buf)79 efx_rx_buf_next(struct efx_rx_queue *rx_queue, struct efx_rx_buffer *rx_buf)
80 {
81 if (unlikely(rx_buf == efx_rx_buffer(rx_queue, rx_queue->ptr_mask)))
82 return efx_rx_buffer(rx_queue, 0);
83 else
84 return rx_buf + 1;
85 }
86
efx_sync_rx_buffer(struct efx_nic * efx,struct efx_rx_buffer * rx_buf,unsigned int len)87 static inline void efx_sync_rx_buffer(struct efx_nic *efx,
88 struct efx_rx_buffer *rx_buf,
89 unsigned int len)
90 {
91 dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr, len,
92 DMA_FROM_DEVICE);
93 }
94
efx_rx_config_page_split(struct efx_nic * efx)95 void efx_rx_config_page_split(struct efx_nic *efx)
96 {
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);
106 }
107
108 /* Check the RX page recycle ring for a page that can be reused. */
efx_reuse_page(struct efx_rx_queue * rx_queue)109 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
110 {
111 struct efx_nic *efx = rx_queue->efx;
112 struct page *page;
113 struct efx_rx_page_state *state;
114 unsigned index;
115
116 index = rx_queue->page_remove & rx_queue->page_ptr_mask;
117 page = rx_queue->page_ring[index];
118 if (page == NULL)
119 return NULL;
120
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;
125
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;
129 return page;
130 } else {
131 state = page_address(page);
132 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
133 PAGE_SIZE << efx->rx_buffer_order,
134 DMA_FROM_DEVICE);
135 put_page(page);
136 ++rx_queue->page_recycle_failed;
137 }
138
139 return NULL;
140 }
141
142 /**
143 * efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
144 *
145 * @rx_queue: Efx RX queue
146 *
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.
151 */
efx_init_rx_buffers(struct efx_rx_queue * rx_queue,bool atomic)152 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
153 {
154 struct efx_nic *efx = rx_queue->efx;
155 struct efx_rx_buffer *rx_buf;
156 struct page *page;
157 unsigned int page_offset;
158 struct efx_rx_page_state *state;
159 dma_addr_t dma_addr;
160 unsigned index, count;
161
162 count = 0;
163 do {
164 page = efx_reuse_page(rx_queue);
165 if (page == NULL) {
166 page = alloc_pages(__GFP_COMP |
167 (atomic ? GFP_ATOMIC : GFP_KERNEL),
168 efx->rx_buffer_order);
169 if (unlikely(page == NULL))
170 return -ENOMEM;
171 dma_addr =
172 dma_map_page(&efx->pci_dev->dev, page, 0,
173 PAGE_SIZE << efx->rx_buffer_order,
174 DMA_FROM_DEVICE);
175 if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
176 dma_addr))) {
177 __free_pages(page, efx->rx_buffer_order);
178 return -EIO;
179 }
180 state = page_address(page);
181 state->dma_addr = dma_addr;
182 } else {
183 state = page_address(page);
184 dma_addr = state->dma_addr;
185 }
186
187 dma_addr += sizeof(struct efx_rx_page_state);
188 page_offset = sizeof(struct efx_rx_page_state);
189
190 do {
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;
194 rx_buf->page = page;
195 rx_buf->page_offset = page_offset + efx->rx_ip_align;
196 rx_buf->len = efx->rx_dma_len;
197 rx_buf->flags = 0;
198 ++rx_queue->added_count;
199 get_page(page);
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);
203
204 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
205 } while (++count < efx->rx_pages_per_batch);
206
207 return 0;
208 }
209
210 /* Unmap a DMA-mapped page. This function is only called for the final RX
211 * buffer in a page.
212 */
efx_unmap_rx_buffer(struct efx_nic * efx,struct efx_rx_buffer * rx_buf)213 static void efx_unmap_rx_buffer(struct efx_nic *efx,
214 struct efx_rx_buffer *rx_buf)
215 {
216 struct page *page = rx_buf->page;
217
218 if (page) {
219 struct efx_rx_page_state *state = page_address(page);
220 dma_unmap_page(&efx->pci_dev->dev,
221 state->dma_addr,
222 PAGE_SIZE << efx->rx_buffer_order,
223 DMA_FROM_DEVICE);
224 }
225 }
226
efx_free_rx_buffers(struct efx_rx_queue * rx_queue,struct efx_rx_buffer * rx_buf,unsigned int num_bufs)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)
230 {
231 do {
232 if (rx_buf->page) {
233 put_page(rx_buf->page);
234 rx_buf->page = NULL;
235 }
236 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
237 } while (--num_bufs);
238 }
239
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.
243 */
efx_recycle_rx_page(struct efx_channel * channel,struct efx_rx_buffer * rx_buf)244 static void efx_recycle_rx_page(struct efx_channel *channel,
245 struct efx_rx_buffer *rx_buf)
246 {
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;
250 unsigned index;
251
252 /* Only recycle the page after processing the final buffer. */
253 if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
254 return;
255
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;
260
261 /* The next slot in the recycle ring is available, but
262 * increment page_remove if the read pointer currently
263 * points here.
264 */
265 if (read_index == index)
266 ++rx_queue->page_remove;
267 rx_queue->page_ring[index] = page;
268 ++rx_queue->page_add;
269 return;
270 }
271 ++rx_queue->page_recycle_full;
272 efx_unmap_rx_buffer(efx, rx_buf);
273 put_page(rx_buf->page);
274 }
275
efx_fini_rx_buffer(struct efx_rx_queue * rx_queue,struct efx_rx_buffer * rx_buf)276 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
277 struct efx_rx_buffer *rx_buf)
278 {
279 /* Release the page reference we hold for the buffer. */
280 if (rx_buf->page)
281 put_page(rx_buf->page);
282
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);
287 }
288 rx_buf->page = NULL;
289 }
290
291 /* Recycle the pages that are used by buffers that have just been received. */
efx_recycle_rx_pages(struct efx_channel * channel,struct efx_rx_buffer * rx_buf,unsigned int n_frags)292 static void efx_recycle_rx_pages(struct efx_channel *channel,
293 struct efx_rx_buffer *rx_buf,
294 unsigned int n_frags)
295 {
296 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
297
298 do {
299 efx_recycle_rx_page(channel, rx_buf);
300 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
301 } while (--n_frags);
302 }
303
efx_discard_rx_packet(struct efx_channel * channel,struct efx_rx_buffer * rx_buf,unsigned int n_frags)304 static void efx_discard_rx_packet(struct efx_channel *channel,
305 struct efx_rx_buffer *rx_buf,
306 unsigned int n_frags)
307 {
308 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
309
310 efx_recycle_rx_pages(channel, rx_buf, n_frags);
311
312 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
313 }
314
315 /**
316 * efx_fast_push_rx_descriptors - push new RX descriptors quickly
317 * @rx_queue: RX descriptor queue
318 *
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.
322 *
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.
326 */
efx_fast_push_rx_descriptors(struct efx_rx_queue * rx_queue,bool atomic)327 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
328 {
329 struct efx_nic *efx = rx_queue->efx;
330 unsigned int fill_level, batch_size;
331 int space, rc = 0;
332
333 if (!rx_queue->refill_enabled)
334 return;
335
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)
340 goto out;
341
342 /* Record minimum fill level */
343 if (unlikely(fill_level < rx_queue->min_fill)) {
344 if (fill_level)
345 rx_queue->min_fill = fill_level;
346 }
347
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);
351
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,
356 rx_queue->max_fill);
357
358
359 do {
360 rc = efx_init_rx_buffers(rx_queue, atomic);
361 if (unlikely(rc)) {
362 /* Ensure that we don't leave the rx queue empty */
363 if (rx_queue->added_count == rx_queue->removed_count)
364 efx_schedule_slow_fill(rx_queue);
365 goto out;
366 }
367 } while ((space -= batch_size) >= batch_size);
368
369 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
370 "RX queue %d fast-filled descriptor ring "
371 "to level %d\n", efx_rx_queue_index(rx_queue),
372 rx_queue->added_count - rx_queue->removed_count);
373
374 out:
375 if (rx_queue->notified_count != rx_queue->added_count)
376 efx_nic_notify_rx_desc(rx_queue);
377 }
378
efx_rx_slow_fill(struct timer_list * t)379 void efx_rx_slow_fill(struct timer_list *t)
380 {
381 struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill);
382
383 /* Post an event to cause NAPI to run and refill the queue */
384 efx_nic_generate_fill_event(rx_queue);
385 ++rx_queue->slow_fill_count;
386 }
387
efx_rx_packet__check_len(struct efx_rx_queue * rx_queue,struct efx_rx_buffer * rx_buf,int len)388 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
389 struct efx_rx_buffer *rx_buf,
390 int len)
391 {
392 struct efx_nic *efx = rx_queue->efx;
393 unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
394
395 if (likely(len <= max_len))
396 return;
397
398 /* The packet must be discarded, but this is only a fatal error
399 * if the caller indicated it was
400 */
401 rx_buf->flags |= EFX_RX_PKT_DISCARD;
402
403 if (net_ratelimit())
404 netif_err(efx, rx_err, efx->net_dev,
405 "RX queue %d overlength RX event (%#x > %#x)\n",
406 efx_rx_queue_index(rx_queue), len, max_len);
407
408 efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
409 }
410
411 /* Pass a received packet up through GRO. GRO can handle pages
412 * regardless of checksum state and skbs with a good checksum.
413 */
414 static void
efx_rx_packet_gro(struct efx_channel * channel,struct efx_rx_buffer * rx_buf,unsigned int n_frags,u8 * eh)415 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
416 unsigned int n_frags, u8 *eh)
417 {
418 struct napi_struct *napi = &channel->napi_str;
419 gro_result_t gro_result;
420 struct efx_nic *efx = channel->efx;
421 struct sk_buff *skb;
422
423 skb = napi_get_frags(napi);
424 if (unlikely(!skb)) {
425 struct efx_rx_queue *rx_queue;
426
427 rx_queue = efx_channel_get_rx_queue(channel);
428 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
429 return;
430 }
431
432 if (efx->net_dev->features & NETIF_F_RXHASH)
433 skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
434 PKT_HASH_TYPE_L3);
435 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
436 CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
437 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
438
439 for (;;) {
440 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
441 rx_buf->page, rx_buf->page_offset,
442 rx_buf->len);
443 rx_buf->page = NULL;
444 skb->len += rx_buf->len;
445 if (skb_shinfo(skb)->nr_frags == n_frags)
446 break;
447
448 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
449 }
450
451 skb->data_len = skb->len;
452 skb->truesize += n_frags * efx->rx_buffer_truesize;
453
454 skb_record_rx_queue(skb, channel->rx_queue.core_index);
455
456 gro_result = napi_gro_frags(napi);
457 if (gro_result != GRO_DROP)
458 channel->irq_mod_score += 2;
459 }
460
461 /* Allocate and construct an SKB around page fragments */
efx_rx_mk_skb(struct efx_channel * channel,struct efx_rx_buffer * rx_buf,unsigned int n_frags,u8 * eh,int hdr_len)462 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
463 struct efx_rx_buffer *rx_buf,
464 unsigned int n_frags,
465 u8 *eh, int hdr_len)
466 {
467 struct efx_nic *efx = channel->efx;
468 struct sk_buff *skb;
469
470 /* Allocate an SKB to store the headers */
471 skb = netdev_alloc_skb(efx->net_dev,
472 efx->rx_ip_align + efx->rx_prefix_size +
473 hdr_len);
474 if (unlikely(skb == NULL)) {
475 atomic_inc(&efx->n_rx_noskb_drops);
476 return NULL;
477 }
478
479 EFX_WARN_ON_ONCE_PARANOID(rx_buf->len < hdr_len);
480
481 memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
482 efx->rx_prefix_size + hdr_len);
483 skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
484 __skb_put(skb, hdr_len);
485
486 /* Append the remaining page(s) onto the frag list */
487 if (rx_buf->len > hdr_len) {
488 rx_buf->page_offset += hdr_len;
489 rx_buf->len -= hdr_len;
490
491 for (;;) {
492 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
493 rx_buf->page, rx_buf->page_offset,
494 rx_buf->len);
495 rx_buf->page = NULL;
496 skb->len += rx_buf->len;
497 skb->data_len += rx_buf->len;
498 if (skb_shinfo(skb)->nr_frags == n_frags)
499 break;
500
501 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
502 }
503 } else {
504 __free_pages(rx_buf->page, efx->rx_buffer_order);
505 rx_buf->page = NULL;
506 n_frags = 0;
507 }
508
509 skb->truesize += n_frags * efx->rx_buffer_truesize;
510
511 /* Move past the ethernet header */
512 skb->protocol = eth_type_trans(skb, efx->net_dev);
513
514 skb_mark_napi_id(skb, &channel->napi_str);
515
516 return skb;
517 }
518
efx_rx_packet(struct efx_rx_queue * rx_queue,unsigned int index,unsigned int n_frags,unsigned int len,u16 flags)519 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
520 unsigned int n_frags, unsigned int len, u16 flags)
521 {
522 struct efx_nic *efx = rx_queue->efx;
523 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
524 struct efx_rx_buffer *rx_buf;
525
526 rx_queue->rx_packets++;
527
528 rx_buf = efx_rx_buffer(rx_queue, index);
529 rx_buf->flags |= flags;
530
531 /* Validate the number of fragments and completed length */
532 if (n_frags == 1) {
533 if (!(flags & EFX_RX_PKT_PREFIX_LEN))
534 efx_rx_packet__check_len(rx_queue, rx_buf, len);
535 } else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
536 unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
537 unlikely(len > n_frags * efx->rx_dma_len) ||
538 unlikely(!efx->rx_scatter)) {
539 /* If this isn't an explicit discard request, either
540 * the hardware or the driver is broken.
541 */
542 WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
543 rx_buf->flags |= EFX_RX_PKT_DISCARD;
544 }
545
546 netif_vdbg(efx, rx_status, efx->net_dev,
547 "RX queue %d received ids %x-%x len %d %s%s\n",
548 efx_rx_queue_index(rx_queue), index,
549 (index + n_frags - 1) & rx_queue->ptr_mask, len,
550 (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
551 (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
552
553 /* Discard packet, if instructed to do so. Process the
554 * previous receive first.
555 */
556 if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
557 efx_rx_flush_packet(channel);
558 efx_discard_rx_packet(channel, rx_buf, n_frags);
559 return;
560 }
561
562 if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
563 rx_buf->len = len;
564
565 /* Release and/or sync the DMA mapping - assumes all RX buffers
566 * consumed in-order per RX queue.
567 */
568 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
569
570 /* Prefetch nice and early so data will (hopefully) be in cache by
571 * the time we look at it.
572 */
573 prefetch(efx_rx_buf_va(rx_buf));
574
575 rx_buf->page_offset += efx->rx_prefix_size;
576 rx_buf->len -= efx->rx_prefix_size;
577
578 if (n_frags > 1) {
579 /* Release/sync DMA mapping for additional fragments.
580 * Fix length for last fragment.
581 */
582 unsigned int tail_frags = n_frags - 1;
583
584 for (;;) {
585 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
586 if (--tail_frags == 0)
587 break;
588 efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
589 }
590 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
591 efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
592 }
593
594 /* All fragments have been DMA-synced, so recycle pages. */
595 rx_buf = efx_rx_buffer(rx_queue, index);
596 efx_recycle_rx_pages(channel, rx_buf, n_frags);
597
598 /* Pipeline receives so that we give time for packet headers to be
599 * prefetched into cache.
600 */
601 efx_rx_flush_packet(channel);
602 channel->rx_pkt_n_frags = n_frags;
603 channel->rx_pkt_index = index;
604 }
605
efx_rx_deliver(struct efx_channel * channel,u8 * eh,struct efx_rx_buffer * rx_buf,unsigned int n_frags)606 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
607 struct efx_rx_buffer *rx_buf,
608 unsigned int n_frags)
609 {
610 struct sk_buff *skb;
611 u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
612
613 skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
614 if (unlikely(skb == NULL)) {
615 struct efx_rx_queue *rx_queue;
616
617 rx_queue = efx_channel_get_rx_queue(channel);
618 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
619 return;
620 }
621 skb_record_rx_queue(skb, channel->rx_queue.core_index);
622
623 /* Set the SKB flags */
624 skb_checksum_none_assert(skb);
625 if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED)) {
626 skb->ip_summed = CHECKSUM_UNNECESSARY;
627 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
628 }
629
630 efx_rx_skb_attach_timestamp(channel, skb);
631
632 if (channel->type->receive_skb)
633 if (channel->type->receive_skb(channel, skb))
634 return;
635
636 /* Pass the packet up */
637 if (channel->rx_list != NULL)
638 /* Add to list, will pass up later */
639 list_add_tail(&skb->list, channel->rx_list);
640 else
641 /* No list, so pass it up now */
642 netif_receive_skb(skb);
643 }
644
645 /* Handle a received packet. Second half: Touches packet payload. */
__efx_rx_packet(struct efx_channel * channel)646 void __efx_rx_packet(struct efx_channel *channel)
647 {
648 struct efx_nic *efx = channel->efx;
649 struct efx_rx_buffer *rx_buf =
650 efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
651 u8 *eh = efx_rx_buf_va(rx_buf);
652
653 /* Read length from the prefix if necessary. This already
654 * excludes the length of the prefix itself.
655 */
656 if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
657 rx_buf->len = le16_to_cpup((__le16 *)
658 (eh + efx->rx_packet_len_offset));
659
660 /* If we're in loopback test, then pass the packet directly to the
661 * loopback layer, and free the rx_buf here
662 */
663 if (unlikely(efx->loopback_selftest)) {
664 struct efx_rx_queue *rx_queue;
665
666 efx_loopback_rx_packet(efx, eh, rx_buf->len);
667 rx_queue = efx_channel_get_rx_queue(channel);
668 efx_free_rx_buffers(rx_queue, rx_buf,
669 channel->rx_pkt_n_frags);
670 goto out;
671 }
672
673 if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
674 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
675
676 if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb)
677 efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
678 else
679 efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
680 out:
681 channel->rx_pkt_n_frags = 0;
682 }
683
efx_probe_rx_queue(struct efx_rx_queue * rx_queue)684 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
685 {
686 struct efx_nic *efx = rx_queue->efx;
687 unsigned int entries;
688 int rc;
689
690 /* Create the smallest power-of-two aligned ring */
691 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
692 EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
693 rx_queue->ptr_mask = entries - 1;
694
695 netif_dbg(efx, probe, efx->net_dev,
696 "creating RX queue %d size %#x mask %#x\n",
697 efx_rx_queue_index(rx_queue), efx->rxq_entries,
698 rx_queue->ptr_mask);
699
700 /* Allocate RX buffers */
701 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
702 GFP_KERNEL);
703 if (!rx_queue->buffer)
704 return -ENOMEM;
705
706 rc = efx_nic_probe_rx(rx_queue);
707 if (rc) {
708 kfree(rx_queue->buffer);
709 rx_queue->buffer = NULL;
710 }
711
712 return rc;
713 }
714
efx_init_rx_recycle_ring(struct efx_nic * efx,struct efx_rx_queue * rx_queue)715 static void efx_init_rx_recycle_ring(struct efx_nic *efx,
716 struct efx_rx_queue *rx_queue)
717 {
718 unsigned int bufs_in_recycle_ring, page_ring_size;
719
720 /* Set the RX recycle ring size */
721 #ifdef CONFIG_PPC64
722 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
723 #else
724 if (iommu_present(&pci_bus_type))
725 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
726 else
727 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
728 #endif /* CONFIG_PPC64 */
729
730 page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
731 efx->rx_bufs_per_page);
732 rx_queue->page_ring = kcalloc(page_ring_size,
733 sizeof(*rx_queue->page_ring), GFP_KERNEL);
734 rx_queue->page_ptr_mask = page_ring_size - 1;
735 }
736
efx_init_rx_queue(struct efx_rx_queue * rx_queue)737 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
738 {
739 struct efx_nic *efx = rx_queue->efx;
740 unsigned int max_fill, trigger, max_trigger;
741
742 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
743 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
744
745 /* Initialise ptr fields */
746 rx_queue->added_count = 0;
747 rx_queue->notified_count = 0;
748 rx_queue->removed_count = 0;
749 rx_queue->min_fill = -1U;
750 efx_init_rx_recycle_ring(efx, rx_queue);
751
752 rx_queue->page_remove = 0;
753 rx_queue->page_add = rx_queue->page_ptr_mask + 1;
754 rx_queue->page_recycle_count = 0;
755 rx_queue->page_recycle_failed = 0;
756 rx_queue->page_recycle_full = 0;
757
758 /* Initialise limit fields */
759 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
760 max_trigger =
761 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
762 if (rx_refill_threshold != 0) {
763 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
764 if (trigger > max_trigger)
765 trigger = max_trigger;
766 } else {
767 trigger = max_trigger;
768 }
769
770 rx_queue->max_fill = max_fill;
771 rx_queue->fast_fill_trigger = trigger;
772 rx_queue->refill_enabled = true;
773
774 /* Set up RX descriptor ring */
775 efx_nic_init_rx(rx_queue);
776 }
777
efx_fini_rx_queue(struct efx_rx_queue * rx_queue)778 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
779 {
780 int i;
781 struct efx_nic *efx = rx_queue->efx;
782 struct efx_rx_buffer *rx_buf;
783
784 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
785 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
786
787 del_timer_sync(&rx_queue->slow_fill);
788
789 /* Release RX buffers from the current read ptr to the write ptr */
790 if (rx_queue->buffer) {
791 for (i = rx_queue->removed_count; i < rx_queue->added_count;
792 i++) {
793 unsigned index = i & rx_queue->ptr_mask;
794 rx_buf = efx_rx_buffer(rx_queue, index);
795 efx_fini_rx_buffer(rx_queue, rx_buf);
796 }
797 }
798
799 /* Unmap and release the pages in the recycle ring. Remove the ring. */
800 for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
801 struct page *page = rx_queue->page_ring[i];
802 struct efx_rx_page_state *state;
803
804 if (page == NULL)
805 continue;
806
807 state = page_address(page);
808 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
809 PAGE_SIZE << efx->rx_buffer_order,
810 DMA_FROM_DEVICE);
811 put_page(page);
812 }
813 kfree(rx_queue->page_ring);
814 rx_queue->page_ring = NULL;
815 }
816
efx_remove_rx_queue(struct efx_rx_queue * rx_queue)817 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
818 {
819 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
820 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
821
822 efx_nic_remove_rx(rx_queue);
823
824 kfree(rx_queue->buffer);
825 rx_queue->buffer = NULL;
826 }
827
828
829 module_param(rx_refill_threshold, uint, 0444);
830 MODULE_PARM_DESC(rx_refill_threshold,
831 "RX descriptor ring refill threshold (%)");
832
833 #ifdef CONFIG_RFS_ACCEL
834
efx_filter_rfs_work(struct work_struct * data)835 static void efx_filter_rfs_work(struct work_struct *data)
836 {
837 struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
838 work);
839 struct efx_nic *efx = netdev_priv(req->net_dev);
840 struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
841 int slot_idx = req - efx->rps_slot;
842 struct efx_arfs_rule *rule;
843 u16 arfs_id = 0;
844 int rc;
845
846 rc = efx->type->filter_insert(efx, &req->spec, true);
847 if (rc >= 0)
848 rc %= efx->type->max_rx_ip_filters;
849 if (efx->rps_hash_table) {
850 spin_lock_bh(&efx->rps_hash_lock);
851 rule = efx_rps_hash_find(efx, &req->spec);
852 /* The rule might have already gone, if someone else's request
853 * for the same spec was already worked and then expired before
854 * we got around to our work. In that case we have nothing
855 * tying us to an arfs_id, meaning that as soon as the filter
856 * is considered for expiry it will be removed.
857 */
858 if (rule) {
859 if (rc < 0)
860 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
861 else
862 rule->filter_id = rc;
863 arfs_id = rule->arfs_id;
864 }
865 spin_unlock_bh(&efx->rps_hash_lock);
866 }
867 if (rc >= 0) {
868 /* Remember this so we can check whether to expire the filter
869 * later.
870 */
871 mutex_lock(&efx->rps_mutex);
872 channel->rps_flow_id[rc] = req->flow_id;
873 ++channel->rfs_filters_added;
874 mutex_unlock(&efx->rps_mutex);
875
876 if (req->spec.ether_type == htons(ETH_P_IP))
877 netif_info(efx, rx_status, efx->net_dev,
878 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
879 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
880 req->spec.rem_host, ntohs(req->spec.rem_port),
881 req->spec.loc_host, ntohs(req->spec.loc_port),
882 req->rxq_index, req->flow_id, rc, arfs_id);
883 else
884 netif_info(efx, rx_status, efx->net_dev,
885 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
886 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
887 req->spec.rem_host, ntohs(req->spec.rem_port),
888 req->spec.loc_host, ntohs(req->spec.loc_port),
889 req->rxq_index, req->flow_id, rc, arfs_id);
890 }
891
892 /* Release references */
893 clear_bit(slot_idx, &efx->rps_slot_map);
894 dev_put(req->net_dev);
895 }
896
efx_filter_rfs(struct net_device * net_dev,const struct sk_buff * skb,u16 rxq_index,u32 flow_id)897 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
898 u16 rxq_index, u32 flow_id)
899 {
900 struct efx_nic *efx = netdev_priv(net_dev);
901 struct efx_async_filter_insertion *req;
902 struct efx_arfs_rule *rule;
903 struct flow_keys fk;
904 int slot_idx;
905 bool new;
906 int rc;
907
908 /* find a free slot */
909 for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
910 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
911 break;
912 if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
913 return -EBUSY;
914
915 if (flow_id == RPS_FLOW_ID_INVALID) {
916 rc = -EINVAL;
917 goto out_clear;
918 }
919
920 if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
921 rc = -EPROTONOSUPPORT;
922 goto out_clear;
923 }
924
925 if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
926 rc = -EPROTONOSUPPORT;
927 goto out_clear;
928 }
929 if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
930 rc = -EPROTONOSUPPORT;
931 goto out_clear;
932 }
933
934 req = efx->rps_slot + slot_idx;
935 efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
936 efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
937 rxq_index);
938 req->spec.match_flags =
939 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
940 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
941 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
942 req->spec.ether_type = fk.basic.n_proto;
943 req->spec.ip_proto = fk.basic.ip_proto;
944
945 if (fk.basic.n_proto == htons(ETH_P_IP)) {
946 req->spec.rem_host[0] = fk.addrs.v4addrs.src;
947 req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
948 } else {
949 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
950 sizeof(struct in6_addr));
951 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
952 sizeof(struct in6_addr));
953 }
954
955 req->spec.rem_port = fk.ports.src;
956 req->spec.loc_port = fk.ports.dst;
957
958 if (efx->rps_hash_table) {
959 /* Add it to ARFS hash table */
960 spin_lock(&efx->rps_hash_lock);
961 rule = efx_rps_hash_add(efx, &req->spec, &new);
962 if (!rule) {
963 rc = -ENOMEM;
964 goto out_unlock;
965 }
966 if (new)
967 rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
968 rc = rule->arfs_id;
969 /* Skip if existing or pending filter already does the right thing */
970 if (!new && rule->rxq_index == rxq_index &&
971 rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
972 goto out_unlock;
973 rule->rxq_index = rxq_index;
974 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
975 spin_unlock(&efx->rps_hash_lock);
976 } else {
977 /* Without an ARFS hash table, we just use arfs_id 0 for all
978 * filters. This means if multiple flows hash to the same
979 * flow_id, all but the most recently touched will be eligible
980 * for expiry.
981 */
982 rc = 0;
983 }
984
985 /* Queue the request */
986 dev_hold(req->net_dev = net_dev);
987 INIT_WORK(&req->work, efx_filter_rfs_work);
988 req->rxq_index = rxq_index;
989 req->flow_id = flow_id;
990 schedule_work(&req->work);
991 return rc;
992 out_unlock:
993 spin_unlock(&efx->rps_hash_lock);
994 out_clear:
995 clear_bit(slot_idx, &efx->rps_slot_map);
996 return rc;
997 }
998
__efx_filter_rfs_expire(struct efx_nic * efx,unsigned int quota)999 bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned int quota)
1000 {
1001 bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
1002 unsigned int channel_idx, index, size;
1003 u32 flow_id;
1004
1005 if (!mutex_trylock(&efx->rps_mutex))
1006 return false;
1007 expire_one = efx->type->filter_rfs_expire_one;
1008 channel_idx = efx->rps_expire_channel;
1009 index = efx->rps_expire_index;
1010 size = efx->type->max_rx_ip_filters;
1011 while (quota--) {
1012 struct efx_channel *channel = efx_get_channel(efx, channel_idx);
1013 flow_id = channel->rps_flow_id[index];
1014
1015 if (flow_id != RPS_FLOW_ID_INVALID &&
1016 expire_one(efx, flow_id, index)) {
1017 netif_info(efx, rx_status, efx->net_dev,
1018 "expired filter %d [queue %u flow %u]\n",
1019 index, channel_idx, flow_id);
1020 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1021 }
1022 if (++index == size) {
1023 if (++channel_idx == efx->n_channels)
1024 channel_idx = 0;
1025 index = 0;
1026 }
1027 }
1028 efx->rps_expire_channel = channel_idx;
1029 efx->rps_expire_index = index;
1030
1031 mutex_unlock(&efx->rps_mutex);
1032 return true;
1033 }
1034
1035 #endif /* CONFIG_RFS_ACCEL */
1036
1037 /**
1038 * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
1039 * @spec: Specification to test
1040 *
1041 * Return: %true if the specification is a non-drop RX filter that
1042 * matches a local MAC address I/G bit value of 1 or matches a local
1043 * IPv4 or IPv6 address value in the respective multicast address
1044 * range. Otherwise %false.
1045 */
efx_filter_is_mc_recipient(const struct efx_filter_spec * spec)1046 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
1047 {
1048 if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
1049 spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
1050 return false;
1051
1052 if (spec->match_flags &
1053 (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
1054 is_multicast_ether_addr(spec->loc_mac))
1055 return true;
1056
1057 if ((spec->match_flags &
1058 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
1059 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
1060 if (spec->ether_type == htons(ETH_P_IP) &&
1061 ipv4_is_multicast(spec->loc_host[0]))
1062 return true;
1063 if (spec->ether_type == htons(ETH_P_IPV6) &&
1064 ((const u8 *)spec->loc_host)[0] == 0xff)
1065 return true;
1066 }
1067
1068 return false;
1069 }
1070
