xref: /linux/drivers/net/ethernet/sfc/rx_common.c (revision a3f5f4c2f9b6bc2aa6f5a3e8e23b7519e4f2e3e3)
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. */
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  */
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. */
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 
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 
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 
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 
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 
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 
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 	rx_queue->old_rx_packets = rx_queue->rx_packets;
245 	rx_queue->old_rx_bytes = rx_queue->rx_bytes;
246 
247 	/* Initialise limit fields */
248 	max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
249 	max_trigger =
250 		max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
251 	if (rx_refill_threshold != 0) {
252 		trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
253 		if (trigger > max_trigger)
254 			trigger = max_trigger;
255 	} else {
256 		trigger = max_trigger;
257 	}
258 
259 	rx_queue->max_fill = max_fill;
260 	rx_queue->fast_fill_trigger = trigger;
261 	rx_queue->refill_enabled = true;
262 
263 	/* Initialise XDP queue information */
264 	rc = xdp_rxq_info_reg(&rx_queue->xdp_rxq_info, efx->net_dev,
265 			      rx_queue->core_index, 0);
266 
267 	if (rc) {
268 		netif_err(efx, rx_err, efx->net_dev,
269 			  "Failure to initialise XDP queue information rc=%d\n",
270 			  rc);
271 		efx->xdp_rxq_info_failed = true;
272 	} else {
273 		rx_queue->xdp_rxq_info_valid = true;
274 	}
275 
276 	/* Set up RX descriptor ring */
277 	efx_nic_init_rx(rx_queue);
278 }
279 
280 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
281 {
282 	struct efx_rx_buffer *rx_buf;
283 	int i;
284 
285 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
286 		  "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
287 
288 	del_timer_sync(&rx_queue->slow_fill);
289 	if (rx_queue->grant_credits)
290 		flush_work(&rx_queue->grant_work);
291 
292 	/* Release RX buffers from the current read ptr to the write ptr */
293 	if (rx_queue->buffer) {
294 		for (i = rx_queue->removed_count; i < rx_queue->added_count;
295 		     i++) {
296 			unsigned int index = i & rx_queue->ptr_mask;
297 
298 			rx_buf = efx_rx_buffer(rx_queue, index);
299 			efx_fini_rx_buffer(rx_queue, rx_buf);
300 		}
301 	}
302 
303 	efx_fini_rx_recycle_ring(rx_queue);
304 
305 	if (rx_queue->xdp_rxq_info_valid)
306 		xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info);
307 
308 	rx_queue->xdp_rxq_info_valid = false;
309 }
310 
311 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
312 {
313 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
314 		  "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
315 
316 	efx_nic_remove_rx(rx_queue);
317 
318 	kfree(rx_queue->buffer);
319 	rx_queue->buffer = NULL;
320 }
321 
322 /* Unmap a DMA-mapped page.  This function is only called for the final RX
323  * buffer in a page.
324  */
325 void efx_unmap_rx_buffer(struct efx_nic *efx,
326 			 struct efx_rx_buffer *rx_buf)
327 {
328 	struct page *page = rx_buf->page;
329 
330 	if (page) {
331 		struct efx_rx_page_state *state = page_address(page);
332 
333 		dma_unmap_page(&efx->pci_dev->dev,
334 			       state->dma_addr,
335 			       PAGE_SIZE << efx->rx_buffer_order,
336 			       DMA_FROM_DEVICE);
337 	}
338 }
339 
340 void efx_free_rx_buffers(struct efx_rx_queue *rx_queue,
341 			 struct efx_rx_buffer *rx_buf,
342 			 unsigned int num_bufs)
343 {
344 	do {
345 		if (rx_buf->page) {
346 			put_page(rx_buf->page);
347 			rx_buf->page = NULL;
348 		}
349 		rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
350 	} while (--num_bufs);
351 }
352 
353 void efx_rx_slow_fill(struct timer_list *t)
354 {
355 	struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill);
356 
357 	/* Post an event to cause NAPI to run and refill the queue */
358 	efx_nic_generate_fill_event(rx_queue);
359 	++rx_queue->slow_fill_count;
360 }
361 
362 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
363 {
364 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(10));
365 }
366 
367 /* efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
368  *
369  * @rx_queue:		Efx RX queue
370  *
371  * This allocates a batch of pages, maps them for DMA, and populates
372  * struct efx_rx_buffers for each one. Return a negative error code or
373  * 0 on success. If a single page can be used for multiple buffers,
374  * then the page will either be inserted fully, or not at all.
375  */
376 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
377 {
378 	unsigned int page_offset, index, count;
379 	struct efx_nic *efx = rx_queue->efx;
380 	struct efx_rx_page_state *state;
381 	struct efx_rx_buffer *rx_buf;
382 	dma_addr_t dma_addr;
383 	struct page *page;
384 
385 	count = 0;
386 	do {
387 		page = efx_reuse_page(rx_queue);
388 		if (page == NULL) {
389 			page = alloc_pages(__GFP_COMP |
390 					   (atomic ? GFP_ATOMIC : GFP_KERNEL),
391 					   efx->rx_buffer_order);
392 			if (unlikely(page == NULL))
393 				return -ENOMEM;
394 			dma_addr =
395 				dma_map_page(&efx->pci_dev->dev, page, 0,
396 					     PAGE_SIZE << efx->rx_buffer_order,
397 					     DMA_FROM_DEVICE);
398 			if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
399 						       dma_addr))) {
400 				__free_pages(page, efx->rx_buffer_order);
401 				return -EIO;
402 			}
403 			state = page_address(page);
404 			state->dma_addr = dma_addr;
405 		} else {
406 			state = page_address(page);
407 			dma_addr = state->dma_addr;
408 		}
409 
410 		dma_addr += sizeof(struct efx_rx_page_state);
411 		page_offset = sizeof(struct efx_rx_page_state);
412 
413 		do {
414 			index = rx_queue->added_count & rx_queue->ptr_mask;
415 			rx_buf = efx_rx_buffer(rx_queue, index);
416 			rx_buf->dma_addr = dma_addr + efx->rx_ip_align +
417 					   EFX_XDP_HEADROOM;
418 			rx_buf->page = page;
419 			rx_buf->page_offset = page_offset + efx->rx_ip_align +
420 					      EFX_XDP_HEADROOM;
421 			rx_buf->len = efx->rx_dma_len;
422 			rx_buf->flags = 0;
423 			++rx_queue->added_count;
424 			get_page(page);
425 			dma_addr += efx->rx_page_buf_step;
426 			page_offset += efx->rx_page_buf_step;
427 		} while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
428 
429 		rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
430 	} while (++count < efx->rx_pages_per_batch);
431 
432 	return 0;
433 }
434 
435 void efx_rx_config_page_split(struct efx_nic *efx)
436 {
437 	efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align +
438 				      EFX_XDP_HEADROOM + EFX_XDP_TAILROOM,
439 				      EFX_RX_BUF_ALIGNMENT);
440 	efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
441 		((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
442 		efx->rx_page_buf_step);
443 	efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
444 		efx->rx_bufs_per_page;
445 	efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
446 					       efx->rx_bufs_per_page);
447 }
448 
449 /* efx_fast_push_rx_descriptors - push new RX descriptors quickly
450  * @rx_queue:		RX descriptor queue
451  *
452  * This will aim to fill the RX descriptor queue up to
453  * @rx_queue->@max_fill. If there is insufficient atomic
454  * memory to do so, a slow fill will be scheduled.
455  *
456  * The caller must provide serialisation (none is used here). In practise,
457  * this means this function must run from the NAPI handler, or be called
458  * when NAPI is disabled.
459  */
460 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
461 {
462 	struct efx_nic *efx = rx_queue->efx;
463 	unsigned int fill_level, batch_size;
464 	int space, rc = 0;
465 
466 	if (!rx_queue->refill_enabled)
467 		return;
468 
469 	/* Calculate current fill level, and exit if we don't need to fill */
470 	fill_level = (rx_queue->added_count - rx_queue->removed_count);
471 	EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries);
472 	if (fill_level >= rx_queue->fast_fill_trigger)
473 		goto out;
474 
475 	/* Record minimum fill level */
476 	if (unlikely(fill_level < rx_queue->min_fill)) {
477 		if (fill_level)
478 			rx_queue->min_fill = fill_level;
479 	}
480 
481 	batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
482 	space = rx_queue->max_fill - fill_level;
483 	EFX_WARN_ON_ONCE_PARANOID(space < batch_size);
484 
485 	netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
486 		   "RX queue %d fast-filling descriptor ring from"
487 		   " level %d to level %d\n",
488 		   efx_rx_queue_index(rx_queue), fill_level,
489 		   rx_queue->max_fill);
490 
491 	do {
492 		rc = efx_init_rx_buffers(rx_queue, atomic);
493 		if (unlikely(rc)) {
494 			/* Ensure that we don't leave the rx queue empty */
495 			efx_schedule_slow_fill(rx_queue);
496 			goto out;
497 		}
498 	} while ((space -= batch_size) >= batch_size);
499 
500 	netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
501 		   "RX queue %d fast-filled descriptor ring "
502 		   "to level %d\n", efx_rx_queue_index(rx_queue),
503 		   rx_queue->added_count - rx_queue->removed_count);
504 
505  out:
506 	if (rx_queue->notified_count != rx_queue->added_count)
507 		efx_nic_notify_rx_desc(rx_queue);
508 }
509 
510 /* Pass a received packet up through GRO.  GRO can handle pages
511  * regardless of checksum state and skbs with a good checksum.
512  */
513 void
514 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
515 		  unsigned int n_frags, u8 *eh, __wsum csum)
516 {
517 	struct napi_struct *napi = &channel->napi_str;
518 	struct efx_nic *efx = channel->efx;
519 	struct sk_buff *skb;
520 
521 	skb = napi_get_frags(napi);
522 	if (unlikely(!skb)) {
523 		struct efx_rx_queue *rx_queue;
524 
525 		rx_queue = efx_channel_get_rx_queue(channel);
526 		efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
527 		return;
528 	}
529 
530 	if (efx->net_dev->features & NETIF_F_RXHASH &&
531 	    efx_rx_buf_hash_valid(efx, eh))
532 		skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
533 			     PKT_HASH_TYPE_L3);
534 	if (csum) {
535 		skb->csum = csum;
536 		skb->ip_summed = CHECKSUM_COMPLETE;
537 	} else {
538 		skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
539 				  CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
540 	}
541 	skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
542 
543 	for (;;) {
544 		skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
545 				   rx_buf->page, rx_buf->page_offset,
546 				   rx_buf->len);
547 		rx_buf->page = NULL;
548 		skb->len += rx_buf->len;
549 		if (skb_shinfo(skb)->nr_frags == n_frags)
550 			break;
551 
552 		rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
553 	}
554 
555 	skb->data_len = skb->len;
556 	skb->truesize += n_frags * efx->rx_buffer_truesize;
557 
558 	skb_record_rx_queue(skb, channel->rx_queue.core_index);
559 
560 	napi_gro_frags(napi);
561 }
562 
563 struct efx_rss_context_priv *efx_find_rss_context_entry(struct efx_nic *efx,
564 							u32 id)
565 {
566 	struct ethtool_rxfh_context *ctx;
567 
568 	WARN_ON(!mutex_is_locked(&efx->net_dev->ethtool->rss_lock));
569 
570 	ctx = xa_load(&efx->net_dev->ethtool->rss_ctx, id);
571 	if (!ctx)
572 		return NULL;
573 	return ethtool_rxfh_context_priv(ctx);
574 }
575 
576 void efx_set_default_rx_indir_table(struct efx_nic *efx, u32 *indir)
577 {
578 	size_t i;
579 
580 	for (i = 0; i < ARRAY_SIZE(efx->rss_context.rx_indir_table); i++)
581 		indir[i] = ethtool_rxfh_indir_default(i, efx->rss_spread);
582 }
583 
584 /**
585  * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
586  * @spec: Specification to test
587  *
588  * Return: %true if the specification is a non-drop RX filter that
589  * matches a local MAC address I/G bit value of 1 or matches a local
590  * IPv4 or IPv6 address value in the respective multicast address
591  * range.  Otherwise %false.
592  */
593 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
594 {
595 	if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
596 	    spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
597 		return false;
598 
599 	if (spec->match_flags &
600 	    (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
601 	    is_multicast_ether_addr(spec->loc_mac))
602 		return true;
603 
604 	if ((spec->match_flags &
605 	     (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
606 	    (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
607 		if (spec->ether_type == htons(ETH_P_IP) &&
608 		    ipv4_is_multicast(spec->loc_host[0]))
609 			return true;
610 		if (spec->ether_type == htons(ETH_P_IPV6) &&
611 		    ((const u8 *)spec->loc_host)[0] == 0xff)
612 			return true;
613 	}
614 
615 	return false;
616 }
617 
618 bool efx_filter_spec_equal(const struct efx_filter_spec *left,
619 			   const struct efx_filter_spec *right)
620 {
621 	if ((left->match_flags ^ right->match_flags) |
622 	    ((left->flags ^ right->flags) &
623 	     (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
624 		return false;
625 
626 	return memcmp(&left->vport_id, &right->vport_id,
627 		      sizeof(struct efx_filter_spec) -
628 		      offsetof(struct efx_filter_spec, vport_id)) == 0;
629 }
630 
631 u32 efx_filter_spec_hash(const struct efx_filter_spec *spec)
632 {
633 	BUILD_BUG_ON(offsetof(struct efx_filter_spec, vport_id) & 3);
634 	return jhash2((const u32 *)&spec->vport_id,
635 		      (sizeof(struct efx_filter_spec) -
636 		       offsetof(struct efx_filter_spec, vport_id)) / 4,
637 		      0);
638 }
639 
640 #ifdef CONFIG_RFS_ACCEL
641 bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx,
642 			bool *force)
643 {
644 	if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) {
645 		/* ARFS is currently updating this entry, leave it */
646 		return false;
647 	}
648 	if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) {
649 		/* ARFS tried and failed to update this, so it's probably out
650 		 * of date.  Remove the filter and the ARFS rule entry.
651 		 */
652 		rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
653 		*force = true;
654 		return true;
655 	} else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */
656 		/* ARFS has moved on, so old filter is not needed.  Since we did
657 		 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will
658 		 * not be removed by efx_rps_hash_del() subsequently.
659 		 */
660 		*force = true;
661 		return true;
662 	}
663 	/* Remove it iff ARFS wants to. */
664 	return true;
665 }
666 
667 static
668 struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx,
669 				       const struct efx_filter_spec *spec)
670 {
671 	u32 hash = efx_filter_spec_hash(spec);
672 
673 	lockdep_assert_held(&efx->rps_hash_lock);
674 	if (!efx->rps_hash_table)
675 		return NULL;
676 	return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE];
677 }
678 
679 struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx,
680 					const struct efx_filter_spec *spec)
681 {
682 	struct efx_arfs_rule *rule;
683 	struct hlist_head *head;
684 	struct hlist_node *node;
685 
686 	head = efx_rps_hash_bucket(efx, spec);
687 	if (!head)
688 		return NULL;
689 	hlist_for_each(node, head) {
690 		rule = container_of(node, struct efx_arfs_rule, node);
691 		if (efx_filter_spec_equal(spec, &rule->spec))
692 			return rule;
693 	}
694 	return NULL;
695 }
696 
697 struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
698 				       const struct efx_filter_spec *spec,
699 				       bool *new)
700 {
701 	struct efx_arfs_rule *rule;
702 	struct hlist_head *head;
703 	struct hlist_node *node;
704 
705 	head = efx_rps_hash_bucket(efx, spec);
706 	if (!head)
707 		return NULL;
708 	hlist_for_each(node, head) {
709 		rule = container_of(node, struct efx_arfs_rule, node);
710 		if (efx_filter_spec_equal(spec, &rule->spec)) {
711 			*new = false;
712 			return rule;
713 		}
714 	}
715 	rule = kmalloc(sizeof(*rule), GFP_ATOMIC);
716 	*new = true;
717 	if (rule) {
718 		memcpy(&rule->spec, spec, sizeof(rule->spec));
719 		hlist_add_head(&rule->node, head);
720 	}
721 	return rule;
722 }
723 
724 void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec)
725 {
726 	struct efx_arfs_rule *rule;
727 	struct hlist_head *head;
728 	struct hlist_node *node;
729 
730 	head = efx_rps_hash_bucket(efx, spec);
731 	if (WARN_ON(!head))
732 		return;
733 	hlist_for_each(node, head) {
734 		rule = container_of(node, struct efx_arfs_rule, node);
735 		if (efx_filter_spec_equal(spec, &rule->spec)) {
736 			/* Someone already reused the entry.  We know that if
737 			 * this check doesn't fire (i.e. filter_id == REMOVING)
738 			 * then the REMOVING mark was put there by our caller,
739 			 * because caller is holding a lock on filter table and
740 			 * only holders of that lock set REMOVING.
741 			 */
742 			if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING)
743 				return;
744 			hlist_del(node);
745 			kfree(rule);
746 			return;
747 		}
748 	}
749 	/* We didn't find it. */
750 	WARN_ON(1);
751 }
752 #endif
753 
754 int efx_probe_filters(struct efx_nic *efx)
755 {
756 	int rc;
757 
758 	mutex_lock(&efx->mac_lock);
759 	rc = efx->type->filter_table_probe(efx);
760 	if (rc)
761 		goto out_unlock;
762 
763 #ifdef CONFIG_RFS_ACCEL
764 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
765 		struct efx_channel *channel;
766 		int i, success = 1;
767 
768 		efx_for_each_channel(channel, efx) {
769 			channel->rps_flow_id =
770 				kcalloc(efx->type->max_rx_ip_filters,
771 					sizeof(*channel->rps_flow_id),
772 					GFP_KERNEL);
773 			if (!channel->rps_flow_id)
774 				success = 0;
775 			else
776 				for (i = 0;
777 				     i < efx->type->max_rx_ip_filters;
778 				     ++i)
779 					channel->rps_flow_id[i] =
780 						RPS_FLOW_ID_INVALID;
781 			channel->rfs_expire_index = 0;
782 			channel->rfs_filter_count = 0;
783 		}
784 
785 		if (!success) {
786 			efx_for_each_channel(channel, efx) {
787 				kfree(channel->rps_flow_id);
788 				channel->rps_flow_id = NULL;
789 			}
790 			efx->type->filter_table_remove(efx);
791 			rc = -ENOMEM;
792 			goto out_unlock;
793 		}
794 	}
795 #endif
796 out_unlock:
797 	mutex_unlock(&efx->mac_lock);
798 	return rc;
799 }
800 
801 void efx_remove_filters(struct efx_nic *efx)
802 {
803 #ifdef CONFIG_RFS_ACCEL
804 	struct efx_channel *channel;
805 
806 	efx_for_each_channel(channel, efx) {
807 		cancel_delayed_work_sync(&channel->filter_work);
808 		kfree(channel->rps_flow_id);
809 		channel->rps_flow_id = NULL;
810 	}
811 #endif
812 	efx->type->filter_table_remove(efx);
813 }
814 
815 #ifdef CONFIG_RFS_ACCEL
816 
817 static void efx_filter_rfs_work(struct work_struct *data)
818 {
819 	struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
820 							      work);
821 	struct efx_nic *efx = efx_netdev_priv(req->net_dev);
822 	struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
823 	int slot_idx = req - efx->rps_slot;
824 	struct efx_arfs_rule *rule;
825 	u16 arfs_id = 0;
826 	int rc;
827 
828 	rc = efx->type->filter_insert(efx, &req->spec, true);
829 	if (rc >= 0)
830 		/* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */
831 		rc %= efx->type->max_rx_ip_filters;
832 	if (efx->rps_hash_table) {
833 		spin_lock_bh(&efx->rps_hash_lock);
834 		rule = efx_rps_hash_find(efx, &req->spec);
835 		/* The rule might have already gone, if someone else's request
836 		 * for the same spec was already worked and then expired before
837 		 * we got around to our work.  In that case we have nothing
838 		 * tying us to an arfs_id, meaning that as soon as the filter
839 		 * is considered for expiry it will be removed.
840 		 */
841 		if (rule) {
842 			if (rc < 0)
843 				rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
844 			else
845 				rule->filter_id = rc;
846 			arfs_id = rule->arfs_id;
847 		}
848 		spin_unlock_bh(&efx->rps_hash_lock);
849 	}
850 	if (rc >= 0) {
851 		/* Remember this so we can check whether to expire the filter
852 		 * later.
853 		 */
854 		mutex_lock(&efx->rps_mutex);
855 		if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID)
856 			channel->rfs_filter_count++;
857 		channel->rps_flow_id[rc] = req->flow_id;
858 		mutex_unlock(&efx->rps_mutex);
859 
860 		if (req->spec.ether_type == htons(ETH_P_IP))
861 			netif_info(efx, rx_status, efx->net_dev,
862 				   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
863 				   (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
864 				   req->spec.rem_host, ntohs(req->spec.rem_port),
865 				   req->spec.loc_host, ntohs(req->spec.loc_port),
866 				   req->rxq_index, req->flow_id, rc, arfs_id);
867 		else
868 			netif_info(efx, rx_status, efx->net_dev,
869 				   "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
870 				   (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
871 				   req->spec.rem_host, ntohs(req->spec.rem_port),
872 				   req->spec.loc_host, ntohs(req->spec.loc_port),
873 				   req->rxq_index, req->flow_id, rc, arfs_id);
874 		channel->n_rfs_succeeded++;
875 	} else {
876 		if (req->spec.ether_type == htons(ETH_P_IP))
877 			netif_dbg(efx, rx_status, efx->net_dev,
878 				  "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %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_dbg(efx, rx_status, efx->net_dev,
885 				  "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %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 		channel->n_rfs_failed++;
891 		/* We're overloading the NIC's filter tables, so let's do a
892 		 * chunk of extra expiry work.
893 		 */
894 		__efx_filter_rfs_expire(channel, min(channel->rfs_filter_count,
895 						     100u));
896 	}
897 
898 	/* Release references */
899 	clear_bit(slot_idx, &efx->rps_slot_map);
900 	dev_put(req->net_dev);
901 }
902 
903 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
904 		   u16 rxq_index, u32 flow_id)
905 {
906 	struct efx_nic *efx = efx_netdev_priv(net_dev);
907 	struct efx_async_filter_insertion *req;
908 	struct efx_arfs_rule *rule;
909 	struct flow_keys fk;
910 	int slot_idx;
911 	bool new;
912 	int rc;
913 
914 	/* find a free slot */
915 	for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
916 		if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
917 			break;
918 	if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
919 		return -EBUSY;
920 
921 	if (flow_id == RPS_FLOW_ID_INVALID) {
922 		rc = -EINVAL;
923 		goto out_clear;
924 	}
925 
926 	if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
927 		rc = -EPROTONOSUPPORT;
928 		goto out_clear;
929 	}
930 
931 	if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
932 		rc = -EPROTONOSUPPORT;
933 		goto out_clear;
934 	}
935 	if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
936 		rc = -EPROTONOSUPPORT;
937 		goto out_clear;
938 	}
939 
940 	req = efx->rps_slot + slot_idx;
941 	efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
942 			   efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
943 			   rxq_index);
944 	req->spec.match_flags =
945 		EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
946 		EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
947 		EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
948 	req->spec.ether_type = fk.basic.n_proto;
949 	req->spec.ip_proto = fk.basic.ip_proto;
950 
951 	if (fk.basic.n_proto == htons(ETH_P_IP)) {
952 		req->spec.rem_host[0] = fk.addrs.v4addrs.src;
953 		req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
954 	} else {
955 		memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
956 		       sizeof(struct in6_addr));
957 		memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
958 		       sizeof(struct in6_addr));
959 	}
960 
961 	req->spec.rem_port = fk.ports.src;
962 	req->spec.loc_port = fk.ports.dst;
963 
964 	if (efx->rps_hash_table) {
965 		/* Add it to ARFS hash table */
966 		spin_lock(&efx->rps_hash_lock);
967 		rule = efx_rps_hash_add(efx, &req->spec, &new);
968 		if (!rule) {
969 			rc = -ENOMEM;
970 			goto out_unlock;
971 		}
972 		if (new)
973 			rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
974 		rc = rule->arfs_id;
975 		/* Skip if existing or pending filter already does the right thing */
976 		if (!new && rule->rxq_index == rxq_index &&
977 		    rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
978 			goto out_unlock;
979 		rule->rxq_index = rxq_index;
980 		rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
981 		spin_unlock(&efx->rps_hash_lock);
982 	} else {
983 		/* Without an ARFS hash table, we just use arfs_id 0 for all
984 		 * filters.  This means if multiple flows hash to the same
985 		 * flow_id, all but the most recently touched will be eligible
986 		 * for expiry.
987 		 */
988 		rc = 0;
989 	}
990 
991 	/* Queue the request */
992 	dev_hold(req->net_dev = net_dev);
993 	INIT_WORK(&req->work, efx_filter_rfs_work);
994 	req->rxq_index = rxq_index;
995 	req->flow_id = flow_id;
996 	schedule_work(&req->work);
997 	return rc;
998 out_unlock:
999 	spin_unlock(&efx->rps_hash_lock);
1000 out_clear:
1001 	clear_bit(slot_idx, &efx->rps_slot_map);
1002 	return rc;
1003 }
1004 
1005 bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota)
1006 {
1007 	bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
1008 	struct efx_nic *efx = channel->efx;
1009 	unsigned int index, size, start;
1010 	u32 flow_id;
1011 
1012 	if (!mutex_trylock(&efx->rps_mutex))
1013 		return false;
1014 	expire_one = efx->type->filter_rfs_expire_one;
1015 	index = channel->rfs_expire_index;
1016 	start = index;
1017 	size = efx->type->max_rx_ip_filters;
1018 	while (quota) {
1019 		flow_id = channel->rps_flow_id[index];
1020 
1021 		if (flow_id != RPS_FLOW_ID_INVALID) {
1022 			quota--;
1023 			if (expire_one(efx, flow_id, index)) {
1024 				netif_info(efx, rx_status, efx->net_dev,
1025 					   "expired filter %d [channel %u flow %u]\n",
1026 					   index, channel->channel, flow_id);
1027 				channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1028 				channel->rfs_filter_count--;
1029 			}
1030 		}
1031 		if (++index == size)
1032 			index = 0;
1033 		/* If we were called with a quota that exceeds the total number
1034 		 * of filters in the table (which shouldn't happen, but could
1035 		 * if two callers race), ensure that we don't loop forever -
1036 		 * stop when we've examined every row of the table.
1037 		 */
1038 		if (index == start)
1039 			break;
1040 	}
1041 
1042 	channel->rfs_expire_index = index;
1043 	mutex_unlock(&efx->rps_mutex);
1044 	return true;
1045 }
1046 
1047 #endif /* CONFIG_RFS_ACCEL */
1048