1 // SPDX-License-Identifier: GPL-2.0-only
2 /****************************************************************************
3 * Driver for Solarflare network controllers and boards
4 * Copyright 2018 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 "net_driver.h"
12 #include <linux/module.h>
13 #include <linux/iommu.h>
14 #include <net/rps.h>
15 #include "efx.h"
16 #include "nic.h"
17 #include "rx_common.h"
18
19 /* This is the percentage fill level below which new RX descriptors
20 * will be added to the RX descriptor ring.
21 */
22 static unsigned int rx_refill_threshold;
23 module_param(rx_refill_threshold, uint, 0444);
24 MODULE_PARM_DESC(rx_refill_threshold,
25 "RX descriptor ring refill threshold (%)");
26
27 /* RX maximum head room required.
28 *
29 * This must be at least 1 to prevent overflow, plus one packet-worth
30 * to allow pipelined receives.
31 */
32 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
33
34 /* Check the RX page recycle ring for a page that can be reused. */
efx_reuse_page(struct efx_rx_queue * rx_queue)35 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
36 {
37 struct efx_nic *efx = rx_queue->efx;
38 struct efx_rx_page_state *state;
39 unsigned int index;
40 struct page *page;
41
42 if (unlikely(!rx_queue->page_ring))
43 return NULL;
44 index = rx_queue->page_remove & rx_queue->page_ptr_mask;
45 page = rx_queue->page_ring[index];
46 if (page == NULL)
47 return NULL;
48
49 rx_queue->page_ring[index] = NULL;
50 /* page_remove cannot exceed page_add. */
51 if (rx_queue->page_remove != rx_queue->page_add)
52 ++rx_queue->page_remove;
53
54 /* If page_count is 1 then we hold the only reference to this page. */
55 if (page_count(page) == 1) {
56 ++rx_queue->page_recycle_count;
57 return page;
58 } else {
59 state = page_address(page);
60 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
61 PAGE_SIZE << efx->rx_buffer_order,
62 DMA_FROM_DEVICE);
63 put_page(page);
64 ++rx_queue->page_recycle_failed;
65 }
66
67 return NULL;
68 }
69
70 /* Attempt to recycle the page if there is an RX recycle ring; the page can
71 * only be added if this is the final RX buffer, to prevent pages being used in
72 * the descriptor ring and appearing in the recycle ring simultaneously.
73 */
efx_recycle_rx_page(struct efx_channel * channel,struct efx_rx_buffer * rx_buf)74 static void efx_recycle_rx_page(struct efx_channel *channel,
75 struct efx_rx_buffer *rx_buf)
76 {
77 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
78 struct efx_nic *efx = rx_queue->efx;
79 struct page *page = rx_buf->page;
80 unsigned int index;
81
82 /* Only recycle the page after processing the final buffer. */
83 if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
84 return;
85
86 index = rx_queue->page_add & rx_queue->page_ptr_mask;
87 if (rx_queue->page_ring[index] == NULL) {
88 unsigned int read_index = rx_queue->page_remove &
89 rx_queue->page_ptr_mask;
90
91 /* The next slot in the recycle ring is available, but
92 * increment page_remove if the read pointer currently
93 * points here.
94 */
95 if (read_index == index)
96 ++rx_queue->page_remove;
97 rx_queue->page_ring[index] = page;
98 ++rx_queue->page_add;
99 return;
100 }
101 ++rx_queue->page_recycle_full;
102 efx_unmap_rx_buffer(efx, rx_buf);
103 put_page(rx_buf->page);
104 }
105
106 /* 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)107 void efx_recycle_rx_pages(struct efx_channel *channel,
108 struct efx_rx_buffer *rx_buf,
109 unsigned int n_frags)
110 {
111 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
112
113 if (unlikely(!rx_queue->page_ring))
114 return;
115
116 do {
117 efx_recycle_rx_page(channel, rx_buf);
118 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
119 } while (--n_frags);
120 }
121
efx_discard_rx_packet(struct efx_channel * channel,struct efx_rx_buffer * rx_buf,unsigned int n_frags)122 void efx_discard_rx_packet(struct efx_channel *channel,
123 struct efx_rx_buffer *rx_buf,
124 unsigned int n_frags)
125 {
126 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
127
128 efx_recycle_rx_pages(channel, rx_buf, n_frags);
129
130 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
131 }
132
efx_init_rx_recycle_ring(struct efx_rx_queue * rx_queue)133 static void efx_init_rx_recycle_ring(struct efx_rx_queue *rx_queue)
134 {
135 unsigned int bufs_in_recycle_ring, page_ring_size;
136 struct efx_nic *efx = rx_queue->efx;
137
138 bufs_in_recycle_ring = efx_rx_recycle_ring_size(efx);
139 page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
140 efx->rx_bufs_per_page);
141 rx_queue->page_ring = kcalloc(page_ring_size,
142 sizeof(*rx_queue->page_ring), GFP_KERNEL);
143 if (!rx_queue->page_ring)
144 rx_queue->page_ptr_mask = 0;
145 else
146 rx_queue->page_ptr_mask = page_ring_size - 1;
147 }
148
efx_fini_rx_recycle_ring(struct efx_rx_queue * rx_queue)149 static void efx_fini_rx_recycle_ring(struct efx_rx_queue *rx_queue)
150 {
151 struct efx_nic *efx = rx_queue->efx;
152 int i;
153
154 if (unlikely(!rx_queue->page_ring))
155 return;
156
157 /* Unmap and release the pages in the recycle ring. Remove the ring. */
158 for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
159 struct page *page = rx_queue->page_ring[i];
160 struct efx_rx_page_state *state;
161
162 if (page == NULL)
163 continue;
164
165 state = page_address(page);
166 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
167 PAGE_SIZE << efx->rx_buffer_order,
168 DMA_FROM_DEVICE);
169 put_page(page);
170 }
171 kfree(rx_queue->page_ring);
172 rx_queue->page_ring = NULL;
173 }
174
efx_fini_rx_buffer(struct efx_rx_queue * rx_queue,struct efx_rx_buffer * rx_buf)175 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
176 struct efx_rx_buffer *rx_buf)
177 {
178 /* Release the page reference we hold for the buffer. */
179 if (rx_buf->page)
180 put_page(rx_buf->page);
181
182 /* If this is the last buffer in a page, unmap and free it. */
183 if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
184 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
185 efx_free_rx_buffers(rx_queue, rx_buf, 1);
186 }
187 rx_buf->page = NULL;
188 }
189
efx_probe_rx_queue(struct efx_rx_queue * rx_queue)190 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
191 {
192 struct efx_nic *efx = rx_queue->efx;
193 unsigned int entries;
194 int rc;
195
196 /* Create the smallest power-of-two aligned ring */
197 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
198 EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
199 rx_queue->ptr_mask = entries - 1;
200
201 netif_dbg(efx, probe, efx->net_dev,
202 "creating RX queue %d size %#x mask %#x\n",
203 efx_rx_queue_index(rx_queue), efx->rxq_entries,
204 rx_queue->ptr_mask);
205
206 /* Allocate RX buffers */
207 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
208 GFP_KERNEL);
209 if (!rx_queue->buffer)
210 return -ENOMEM;
211
212 rc = efx_nic_probe_rx(rx_queue);
213 if (rc) {
214 kfree(rx_queue->buffer);
215 rx_queue->buffer = NULL;
216 }
217
218 return rc;
219 }
220
efx_init_rx_queue(struct efx_rx_queue * rx_queue)221 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
222 {
223 unsigned int max_fill, trigger, max_trigger;
224 struct efx_nic *efx = rx_queue->efx;
225 int rc = 0;
226
227 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
228 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
229
230 /* Initialise ptr fields */
231 rx_queue->added_count = 0;
232 rx_queue->notified_count = 0;
233 rx_queue->granted_count = 0;
234 rx_queue->removed_count = 0;
235 rx_queue->min_fill = -1U;
236 efx_init_rx_recycle_ring(rx_queue);
237
238 rx_queue->page_remove = 0;
239 rx_queue->page_add = rx_queue->page_ptr_mask + 1;
240 rx_queue->page_recycle_count = 0;
241 rx_queue->page_recycle_failed = 0;
242 rx_queue->page_recycle_full = 0;
243
244 /* Initialise limit fields */
245 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
246 max_trigger =
247 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
248 if (rx_refill_threshold != 0) {
249 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
250 if (trigger > max_trigger)
251 trigger = max_trigger;
252 } else {
253 trigger = max_trigger;
254 }
255
256 rx_queue->max_fill = max_fill;
257 rx_queue->fast_fill_trigger = trigger;
258 rx_queue->refill_enabled = true;
259
260 /* Initialise XDP queue information */
261 rc = xdp_rxq_info_reg(&rx_queue->xdp_rxq_info, efx->net_dev,
262 rx_queue->core_index, 0);
263
264 if (rc) {
265 netif_err(efx, rx_err, efx->net_dev,
266 "Failure to initialise XDP queue information rc=%d\n",
267 rc);
268 efx->xdp_rxq_info_failed = true;
269 } else {
270 rx_queue->xdp_rxq_info_valid = true;
271 }
272
273 /* Set up RX descriptor ring */
274 efx_nic_init_rx(rx_queue);
275 }
276
efx_fini_rx_queue(struct efx_rx_queue * rx_queue)277 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
278 {
279 struct efx_rx_buffer *rx_buf;
280 int i;
281
282 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
283 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
284
285 del_timer_sync(&rx_queue->slow_fill);
286 if (rx_queue->grant_credits)
287 flush_work(&rx_queue->grant_work);
288
289 /* Release RX buffers from the current read ptr to the write ptr */
290 if (rx_queue->buffer) {
291 for (i = rx_queue->removed_count; i < rx_queue->added_count;
292 i++) {
293 unsigned int index = i & rx_queue->ptr_mask;
294
295 rx_buf = efx_rx_buffer(rx_queue, index);
296 efx_fini_rx_buffer(rx_queue, rx_buf);
297 }
298 }
299
300 efx_fini_rx_recycle_ring(rx_queue);
301
302 if (rx_queue->xdp_rxq_info_valid)
303 xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info);
304
305 rx_queue->xdp_rxq_info_valid = false;
306 }
307
efx_remove_rx_queue(struct efx_rx_queue * rx_queue)308 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
309 {
310 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
311 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
312
313 efx_nic_remove_rx(rx_queue);
314
315 kfree(rx_queue->buffer);
316 rx_queue->buffer = NULL;
317 }
318
319 /* Unmap a DMA-mapped page. This function is only called for the final RX
320 * buffer in a page.
321 */
efx_unmap_rx_buffer(struct efx_nic * efx,struct efx_rx_buffer * rx_buf)322 void efx_unmap_rx_buffer(struct efx_nic *efx,
323 struct efx_rx_buffer *rx_buf)
324 {
325 struct page *page = rx_buf->page;
326
327 if (page) {
328 struct efx_rx_page_state *state = page_address(page);
329
330 dma_unmap_page(&efx->pci_dev->dev,
331 state->dma_addr,
332 PAGE_SIZE << efx->rx_buffer_order,
333 DMA_FROM_DEVICE);
334 }
335 }
336
efx_free_rx_buffers(struct efx_rx_queue * rx_queue,struct efx_rx_buffer * rx_buf,unsigned int num_bufs)337 void efx_free_rx_buffers(struct efx_rx_queue *rx_queue,
338 struct efx_rx_buffer *rx_buf,
339 unsigned int num_bufs)
340 {
341 do {
342 if (rx_buf->page) {
343 put_page(rx_buf->page);
344 rx_buf->page = NULL;
345 }
346 rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
347 } while (--num_bufs);
348 }
349
efx_rx_slow_fill(struct timer_list * t)350 void efx_rx_slow_fill(struct timer_list *t)
351 {
352 struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill);
353
354 /* Post an event to cause NAPI to run and refill the queue */
355 efx_nic_generate_fill_event(rx_queue);
356 ++rx_queue->slow_fill_count;
357 }
358
efx_schedule_slow_fill(struct efx_rx_queue * rx_queue)359 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
360 {
361 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(10));
362 }
363
364 /* efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
365 *
366 * @rx_queue: Efx RX queue
367 *
368 * This allocates a batch of pages, maps them for DMA, and populates
369 * struct efx_rx_buffers for each one. Return a negative error code or
370 * 0 on success. If a single page can be used for multiple buffers,
371 * then the page will either be inserted fully, or not at all.
372 */
efx_init_rx_buffers(struct efx_rx_queue * rx_queue,bool atomic)373 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
374 {
375 unsigned int page_offset, index, count;
376 struct efx_nic *efx = rx_queue->efx;
377 struct efx_rx_page_state *state;
378 struct efx_rx_buffer *rx_buf;
379 dma_addr_t dma_addr;
380 struct page *page;
381
382 count = 0;
383 do {
384 page = efx_reuse_page(rx_queue);
385 if (page == NULL) {
386 page = alloc_pages(__GFP_COMP |
387 (atomic ? GFP_ATOMIC : GFP_KERNEL),
388 efx->rx_buffer_order);
389 if (unlikely(page == NULL))
390 return -ENOMEM;
391 dma_addr =
392 dma_map_page(&efx->pci_dev->dev, page, 0,
393 PAGE_SIZE << efx->rx_buffer_order,
394 DMA_FROM_DEVICE);
395 if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
396 dma_addr))) {
397 __free_pages(page, efx->rx_buffer_order);
398 return -EIO;
399 }
400 state = page_address(page);
401 state->dma_addr = dma_addr;
402 } else {
403 state = page_address(page);
404 dma_addr = state->dma_addr;
405 }
406
407 dma_addr += sizeof(struct efx_rx_page_state);
408 page_offset = sizeof(struct efx_rx_page_state);
409
410 do {
411 index = rx_queue->added_count & rx_queue->ptr_mask;
412 rx_buf = efx_rx_buffer(rx_queue, index);
413 rx_buf->dma_addr = dma_addr + efx->rx_ip_align +
414 EFX_XDP_HEADROOM;
415 rx_buf->page = page;
416 rx_buf->page_offset = page_offset + efx->rx_ip_align +
417 EFX_XDP_HEADROOM;
418 rx_buf->len = efx->rx_dma_len;
419 rx_buf->flags = 0;
420 ++rx_queue->added_count;
421 get_page(page);
422 dma_addr += efx->rx_page_buf_step;
423 page_offset += efx->rx_page_buf_step;
424 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
425
426 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
427 } while (++count < efx->rx_pages_per_batch);
428
429 return 0;
430 }
431
efx_rx_config_page_split(struct efx_nic * efx)432 void efx_rx_config_page_split(struct efx_nic *efx)
433 {
434 efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align +
435 EFX_XDP_HEADROOM + EFX_XDP_TAILROOM,
436 EFX_RX_BUF_ALIGNMENT);
437 efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
438 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
439 efx->rx_page_buf_step);
440 efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
441 efx->rx_bufs_per_page;
442 efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
443 efx->rx_bufs_per_page);
444 }
445
446 /* efx_fast_push_rx_descriptors - push new RX descriptors quickly
447 * @rx_queue: RX descriptor queue
448 *
449 * This will aim to fill the RX descriptor queue up to
450 * @rx_queue->@max_fill. If there is insufficient atomic
451 * memory to do so, a slow fill will be scheduled.
452 *
453 * The caller must provide serialisation (none is used here). In practise,
454 * this means this function must run from the NAPI handler, or be called
455 * when NAPI is disabled.
456 */
efx_fast_push_rx_descriptors(struct efx_rx_queue * rx_queue,bool atomic)457 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
458 {
459 struct efx_nic *efx = rx_queue->efx;
460 unsigned int fill_level, batch_size;
461 int space, rc = 0;
462
463 if (!rx_queue->refill_enabled)
464 return;
465
466 /* Calculate current fill level, and exit if we don't need to fill */
467 fill_level = (rx_queue->added_count - rx_queue->removed_count);
468 EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries);
469 if (fill_level >= rx_queue->fast_fill_trigger)
470 goto out;
471
472 /* Record minimum fill level */
473 if (unlikely(fill_level < rx_queue->min_fill)) {
474 if (fill_level)
475 rx_queue->min_fill = fill_level;
476 }
477
478 batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
479 space = rx_queue->max_fill - fill_level;
480 EFX_WARN_ON_ONCE_PARANOID(space < batch_size);
481
482 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
483 "RX queue %d fast-filling descriptor ring from"
484 " level %d to level %d\n",
485 efx_rx_queue_index(rx_queue), fill_level,
486 rx_queue->max_fill);
487
488 do {
489 rc = efx_init_rx_buffers(rx_queue, atomic);
490 if (unlikely(rc)) {
491 /* Ensure that we don't leave the rx queue empty */
492 efx_schedule_slow_fill(rx_queue);
493 goto out;
494 }
495 } while ((space -= batch_size) >= batch_size);
496
497 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
498 "RX queue %d fast-filled descriptor ring "
499 "to level %d\n", efx_rx_queue_index(rx_queue),
500 rx_queue->added_count - rx_queue->removed_count);
501
502 out:
503 if (rx_queue->notified_count != rx_queue->added_count)
504 efx_nic_notify_rx_desc(rx_queue);
505 }
506
507 /* Pass a received packet up through GRO. GRO can handle pages
508 * regardless of checksum state and skbs with a good checksum.
509 */
510 void
efx_rx_packet_gro(struct efx_channel * channel,struct efx_rx_buffer * rx_buf,unsigned int n_frags,u8 * eh,__wsum csum)511 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
512 unsigned int n_frags, u8 *eh, __wsum csum)
513 {
514 struct napi_struct *napi = &channel->napi_str;
515 struct efx_nic *efx = channel->efx;
516 struct sk_buff *skb;
517
518 skb = napi_get_frags(napi);
519 if (unlikely(!skb)) {
520 struct efx_rx_queue *rx_queue;
521
522 rx_queue = efx_channel_get_rx_queue(channel);
523 efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
524 return;
525 }
526
527 if (efx->net_dev->features & NETIF_F_RXHASH &&
528 efx_rx_buf_hash_valid(efx, eh))
529 skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
530 PKT_HASH_TYPE_L3);
531 if (csum) {
532 skb->csum = csum;
533 skb->ip_summed = CHECKSUM_COMPLETE;
534 } else {
535 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
536 CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
537 }
538 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
539
540 for (;;) {
541 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
542 rx_buf->page, rx_buf->page_offset,
543 rx_buf->len);
544 rx_buf->page = NULL;
545 skb->len += rx_buf->len;
546 if (skb_shinfo(skb)->nr_frags == n_frags)
547 break;
548
549 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
550 }
551
552 skb->data_len = skb->len;
553 skb->truesize += n_frags * efx->rx_buffer_truesize;
554
555 skb_record_rx_queue(skb, channel->rx_queue.core_index);
556
557 napi_gro_frags(napi);
558 }
559
efx_find_rss_context_entry(struct efx_nic * efx,u32 id)560 struct efx_rss_context_priv *efx_find_rss_context_entry(struct efx_nic *efx,
561 u32 id)
562 {
563 struct ethtool_rxfh_context *ctx;
564
565 WARN_ON(!mutex_is_locked(&efx->net_dev->ethtool->rss_lock));
566
567 ctx = xa_load(&efx->net_dev->ethtool->rss_ctx, id);
568 if (!ctx)
569 return NULL;
570 return ethtool_rxfh_context_priv(ctx);
571 }
572
efx_set_default_rx_indir_table(struct efx_nic * efx,u32 * indir)573 void efx_set_default_rx_indir_table(struct efx_nic *efx, u32 *indir)
574 {
575 size_t i;
576
577 for (i = 0; i < ARRAY_SIZE(efx->rss_context.rx_indir_table); i++)
578 indir[i] = ethtool_rxfh_indir_default(i, efx->rss_spread);
579 }
580
581 /**
582 * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
583 * @spec: Specification to test
584 *
585 * Return: %true if the specification is a non-drop RX filter that
586 * matches a local MAC address I/G bit value of 1 or matches a local
587 * IPv4 or IPv6 address value in the respective multicast address
588 * range. Otherwise %false.
589 */
efx_filter_is_mc_recipient(const struct efx_filter_spec * spec)590 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
591 {
592 if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
593 spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
594 return false;
595
596 if (spec->match_flags &
597 (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
598 is_multicast_ether_addr(spec->loc_mac))
599 return true;
600
601 if ((spec->match_flags &
602 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
603 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
604 if (spec->ether_type == htons(ETH_P_IP) &&
605 ipv4_is_multicast(spec->loc_host[0]))
606 return true;
607 if (spec->ether_type == htons(ETH_P_IPV6) &&
608 ((const u8 *)spec->loc_host)[0] == 0xff)
609 return true;
610 }
611
612 return false;
613 }
614
efx_filter_spec_equal(const struct efx_filter_spec * left,const struct efx_filter_spec * right)615 bool efx_filter_spec_equal(const struct efx_filter_spec *left,
616 const struct efx_filter_spec *right)
617 {
618 if ((left->match_flags ^ right->match_flags) |
619 ((left->flags ^ right->flags) &
620 (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
621 return false;
622
623 return memcmp(&left->vport_id, &right->vport_id,
624 sizeof(struct efx_filter_spec) -
625 offsetof(struct efx_filter_spec, vport_id)) == 0;
626 }
627
efx_filter_spec_hash(const struct efx_filter_spec * spec)628 u32 efx_filter_spec_hash(const struct efx_filter_spec *spec)
629 {
630 BUILD_BUG_ON(offsetof(struct efx_filter_spec, vport_id) & 3);
631 return jhash2((const u32 *)&spec->vport_id,
632 (sizeof(struct efx_filter_spec) -
633 offsetof(struct efx_filter_spec, vport_id)) / 4,
634 0);
635 }
636
637 #ifdef CONFIG_RFS_ACCEL
efx_rps_check_rule(struct efx_arfs_rule * rule,unsigned int filter_idx,bool * force)638 bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx,
639 bool *force)
640 {
641 if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) {
642 /* ARFS is currently updating this entry, leave it */
643 return false;
644 }
645 if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) {
646 /* ARFS tried and failed to update this, so it's probably out
647 * of date. Remove the filter and the ARFS rule entry.
648 */
649 rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
650 *force = true;
651 return true;
652 } else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */
653 /* ARFS has moved on, so old filter is not needed. Since we did
654 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will
655 * not be removed by efx_rps_hash_del() subsequently.
656 */
657 *force = true;
658 return true;
659 }
660 /* Remove it iff ARFS wants to. */
661 return true;
662 }
663
664 static
efx_rps_hash_bucket(struct efx_nic * efx,const struct efx_filter_spec * spec)665 struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx,
666 const struct efx_filter_spec *spec)
667 {
668 u32 hash = efx_filter_spec_hash(spec);
669
670 lockdep_assert_held(&efx->rps_hash_lock);
671 if (!efx->rps_hash_table)
672 return NULL;
673 return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE];
674 }
675
efx_rps_hash_find(struct efx_nic * efx,const struct efx_filter_spec * spec)676 struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx,
677 const struct efx_filter_spec *spec)
678 {
679 struct efx_arfs_rule *rule;
680 struct hlist_head *head;
681 struct hlist_node *node;
682
683 head = efx_rps_hash_bucket(efx, spec);
684 if (!head)
685 return NULL;
686 hlist_for_each(node, head) {
687 rule = container_of(node, struct efx_arfs_rule, node);
688 if (efx_filter_spec_equal(spec, &rule->spec))
689 return rule;
690 }
691 return NULL;
692 }
693
efx_rps_hash_add(struct efx_nic * efx,const struct efx_filter_spec * spec,bool * new)694 struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
695 const struct efx_filter_spec *spec,
696 bool *new)
697 {
698 struct efx_arfs_rule *rule;
699 struct hlist_head *head;
700 struct hlist_node *node;
701
702 head = efx_rps_hash_bucket(efx, spec);
703 if (!head)
704 return NULL;
705 hlist_for_each(node, head) {
706 rule = container_of(node, struct efx_arfs_rule, node);
707 if (efx_filter_spec_equal(spec, &rule->spec)) {
708 *new = false;
709 return rule;
710 }
711 }
712 rule = kmalloc(sizeof(*rule), GFP_ATOMIC);
713 *new = true;
714 if (rule) {
715 memcpy(&rule->spec, spec, sizeof(rule->spec));
716 hlist_add_head(&rule->node, head);
717 }
718 return rule;
719 }
720
efx_rps_hash_del(struct efx_nic * efx,const struct efx_filter_spec * spec)721 void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec)
722 {
723 struct efx_arfs_rule *rule;
724 struct hlist_head *head;
725 struct hlist_node *node;
726
727 head = efx_rps_hash_bucket(efx, spec);
728 if (WARN_ON(!head))
729 return;
730 hlist_for_each(node, head) {
731 rule = container_of(node, struct efx_arfs_rule, node);
732 if (efx_filter_spec_equal(spec, &rule->spec)) {
733 /* Someone already reused the entry. We know that if
734 * this check doesn't fire (i.e. filter_id == REMOVING)
735 * then the REMOVING mark was put there by our caller,
736 * because caller is holding a lock on filter table and
737 * only holders of that lock set REMOVING.
738 */
739 if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING)
740 return;
741 hlist_del(node);
742 kfree(rule);
743 return;
744 }
745 }
746 /* We didn't find it. */
747 WARN_ON(1);
748 }
749 #endif
750
efx_probe_filters(struct efx_nic * efx)751 int efx_probe_filters(struct efx_nic *efx)
752 {
753 int rc;
754
755 mutex_lock(&efx->mac_lock);
756 rc = efx->type->filter_table_probe(efx);
757 if (rc)
758 goto out_unlock;
759
760 #ifdef CONFIG_RFS_ACCEL
761 if (efx->type->offload_features & NETIF_F_NTUPLE) {
762 struct efx_channel *channel;
763 int i, success = 1;
764
765 efx_for_each_channel(channel, efx) {
766 channel->rps_flow_id =
767 kcalloc(efx->type->max_rx_ip_filters,
768 sizeof(*channel->rps_flow_id),
769 GFP_KERNEL);
770 if (!channel->rps_flow_id)
771 success = 0;
772 else
773 for (i = 0;
774 i < efx->type->max_rx_ip_filters;
775 ++i)
776 channel->rps_flow_id[i] =
777 RPS_FLOW_ID_INVALID;
778 channel->rfs_expire_index = 0;
779 channel->rfs_filter_count = 0;
780 }
781
782 if (!success) {
783 efx_for_each_channel(channel, efx) {
784 kfree(channel->rps_flow_id);
785 channel->rps_flow_id = NULL;
786 }
787 efx->type->filter_table_remove(efx);
788 rc = -ENOMEM;
789 goto out_unlock;
790 }
791 }
792 #endif
793 out_unlock:
794 mutex_unlock(&efx->mac_lock);
795 return rc;
796 }
797
efx_remove_filters(struct efx_nic * efx)798 void efx_remove_filters(struct efx_nic *efx)
799 {
800 #ifdef CONFIG_RFS_ACCEL
801 struct efx_channel *channel;
802
803 efx_for_each_channel(channel, efx) {
804 cancel_delayed_work_sync(&channel->filter_work);
805 kfree(channel->rps_flow_id);
806 channel->rps_flow_id = NULL;
807 }
808 #endif
809 efx->type->filter_table_remove(efx);
810 }
811
812 #ifdef CONFIG_RFS_ACCEL
813
efx_filter_rfs_work(struct work_struct * data)814 static void efx_filter_rfs_work(struct work_struct *data)
815 {
816 struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
817 work);
818 struct efx_nic *efx = efx_netdev_priv(req->net_dev);
819 struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
820 int slot_idx = req - efx->rps_slot;
821 struct efx_arfs_rule *rule;
822 u16 arfs_id = 0;
823 int rc;
824
825 rc = efx->type->filter_insert(efx, &req->spec, true);
826 if (rc >= 0)
827 /* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */
828 rc %= efx->type->max_rx_ip_filters;
829 if (efx->rps_hash_table) {
830 spin_lock_bh(&efx->rps_hash_lock);
831 rule = efx_rps_hash_find(efx, &req->spec);
832 /* The rule might have already gone, if someone else's request
833 * for the same spec was already worked and then expired before
834 * we got around to our work. In that case we have nothing
835 * tying us to an arfs_id, meaning that as soon as the filter
836 * is considered for expiry it will be removed.
837 */
838 if (rule) {
839 if (rc < 0)
840 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
841 else
842 rule->filter_id = rc;
843 arfs_id = rule->arfs_id;
844 }
845 spin_unlock_bh(&efx->rps_hash_lock);
846 }
847 if (rc >= 0) {
848 /* Remember this so we can check whether to expire the filter
849 * later.
850 */
851 mutex_lock(&efx->rps_mutex);
852 if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID)
853 channel->rfs_filter_count++;
854 channel->rps_flow_id[rc] = req->flow_id;
855 mutex_unlock(&efx->rps_mutex);
856
857 if (req->spec.ether_type == htons(ETH_P_IP))
858 netif_info(efx, rx_status, efx->net_dev,
859 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
860 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
861 req->spec.rem_host, ntohs(req->spec.rem_port),
862 req->spec.loc_host, ntohs(req->spec.loc_port),
863 req->rxq_index, req->flow_id, rc, arfs_id);
864 else
865 netif_info(efx, rx_status, efx->net_dev,
866 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
867 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
868 req->spec.rem_host, ntohs(req->spec.rem_port),
869 req->spec.loc_host, ntohs(req->spec.loc_port),
870 req->rxq_index, req->flow_id, rc, arfs_id);
871 channel->n_rfs_succeeded++;
872 } else {
873 if (req->spec.ether_type == htons(ETH_P_IP))
874 netif_dbg(efx, rx_status, efx->net_dev,
875 "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n",
876 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
877 req->spec.rem_host, ntohs(req->spec.rem_port),
878 req->spec.loc_host, ntohs(req->spec.loc_port),
879 req->rxq_index, req->flow_id, rc, arfs_id);
880 else
881 netif_dbg(efx, rx_status, efx->net_dev,
882 "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n",
883 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
884 req->spec.rem_host, ntohs(req->spec.rem_port),
885 req->spec.loc_host, ntohs(req->spec.loc_port),
886 req->rxq_index, req->flow_id, rc, arfs_id);
887 channel->n_rfs_failed++;
888 /* We're overloading the NIC's filter tables, so let's do a
889 * chunk of extra expiry work.
890 */
891 __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count,
892 100u));
893 }
894
895 /* Release references */
896 clear_bit(slot_idx, &efx->rps_slot_map);
897 dev_put(req->net_dev);
898 }
899
efx_filter_rfs(struct net_device * net_dev,const struct sk_buff * skb,u16 rxq_index,u32 flow_id)900 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
901 u16 rxq_index, u32 flow_id)
902 {
903 struct efx_nic *efx = efx_netdev_priv(net_dev);
904 struct efx_async_filter_insertion *req;
905 struct efx_arfs_rule *rule;
906 struct flow_keys fk;
907 int slot_idx;
908 bool new;
909 int rc;
910
911 /* find a free slot */
912 for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
913 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
914 break;
915 if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
916 return -EBUSY;
917
918 if (flow_id == RPS_FLOW_ID_INVALID) {
919 rc = -EINVAL;
920 goto out_clear;
921 }
922
923 if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
924 rc = -EPROTONOSUPPORT;
925 goto out_clear;
926 }
927
928 if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
929 rc = -EPROTONOSUPPORT;
930 goto out_clear;
931 }
932 if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
933 rc = -EPROTONOSUPPORT;
934 goto out_clear;
935 }
936
937 req = efx->rps_slot + slot_idx;
938 efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
939 efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
940 rxq_index);
941 req->spec.match_flags =
942 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
943 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
944 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
945 req->spec.ether_type = fk.basic.n_proto;
946 req->spec.ip_proto = fk.basic.ip_proto;
947
948 if (fk.basic.n_proto == htons(ETH_P_IP)) {
949 req->spec.rem_host[0] = fk.addrs.v4addrs.src;
950 req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
951 } else {
952 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
953 sizeof(struct in6_addr));
954 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
955 sizeof(struct in6_addr));
956 }
957
958 req->spec.rem_port = fk.ports.src;
959 req->spec.loc_port = fk.ports.dst;
960
961 if (efx->rps_hash_table) {
962 /* Add it to ARFS hash table */
963 spin_lock(&efx->rps_hash_lock);
964 rule = efx_rps_hash_add(efx, &req->spec, &new);
965 if (!rule) {
966 rc = -ENOMEM;
967 goto out_unlock;
968 }
969 if (new)
970 rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
971 rc = rule->arfs_id;
972 /* Skip if existing or pending filter already does the right thing */
973 if (!new && rule->rxq_index == rxq_index &&
974 rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
975 goto out_unlock;
976 rule->rxq_index = rxq_index;
977 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
978 spin_unlock(&efx->rps_hash_lock);
979 } else {
980 /* Without an ARFS hash table, we just use arfs_id 0 for all
981 * filters. This means if multiple flows hash to the same
982 * flow_id, all but the most recently touched will be eligible
983 * for expiry.
984 */
985 rc = 0;
986 }
987
988 /* Queue the request */
989 dev_hold(req->net_dev = net_dev);
990 INIT_WORK(&req->work, efx_filter_rfs_work);
991 req->rxq_index = rxq_index;
992 req->flow_id = flow_id;
993 schedule_work(&req->work);
994 return rc;
995 out_unlock:
996 spin_unlock(&efx->rps_hash_lock);
997 out_clear:
998 clear_bit(slot_idx, &efx->rps_slot_map);
999 return rc;
1000 }
1001
__efx_filter_rfs_expire(struct efx_channel * channel,unsigned int quota)1002 bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota)
1003 {
1004 bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
1005 struct efx_nic *efx = channel->efx;
1006 unsigned int index, size, start;
1007 u32 flow_id;
1008
1009 if (!mutex_trylock(&efx->rps_mutex))
1010 return false;
1011 expire_one = efx->type->filter_rfs_expire_one;
1012 index = channel->rfs_expire_index;
1013 start = index;
1014 size = efx->type->max_rx_ip_filters;
1015 while (quota) {
1016 flow_id = channel->rps_flow_id[index];
1017
1018 if (flow_id != RPS_FLOW_ID_INVALID) {
1019 quota--;
1020 if (expire_one(efx, flow_id, index)) {
1021 netif_info(efx, rx_status, efx->net_dev,
1022 "expired filter %d [channel %u flow %u]\n",
1023 index, channel->channel, flow_id);
1024 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1025 channel->rfs_filter_count--;
1026 }
1027 }
1028 if (++index == size)
1029 index = 0;
1030 /* If we were called with a quota that exceeds the total number
1031 * of filters in the table (which shouldn't happen, but could
1032 * if two callers race), ensure that we don't loop forever -
1033 * stop when we've examined every row of the table.
1034 */
1035 if (index == start)
1036 break;
1037 }
1038
1039 channel->rfs_expire_index = index;
1040 mutex_unlock(&efx->rps_mutex);
1041 return true;
1042 }
1043
1044 #endif /* CONFIG_RFS_ACCEL */
1045