xref: /linux/drivers/net/ethernet/sfc/siena/rx_common.c (revision 1fd1dc41724319406b0aff221a352a400b0ddfc5)
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 static void efx_unmap_rx_buffer(struct efx_nic *efx,
35 				struct efx_rx_buffer *rx_buf);
36 
37 /* Check the RX page recycle ring for a page that can be reused. */
38 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
39 {
40 	struct efx_nic *efx = rx_queue->efx;
41 	struct efx_rx_page_state *state;
42 	unsigned int index;
43 	struct page *page;
44 
45 	if (unlikely(!rx_queue->page_ring))
46 		return NULL;
47 	index = rx_queue->page_remove & rx_queue->page_ptr_mask;
48 	page = rx_queue->page_ring[index];
49 	if (page == NULL)
50 		return NULL;
51 
52 	rx_queue->page_ring[index] = NULL;
53 	/* page_remove cannot exceed page_add. */
54 	if (rx_queue->page_remove != rx_queue->page_add)
55 		++rx_queue->page_remove;
56 
57 	/* If page_count is 1 then we hold the only reference to this page. */
58 	if (page_count(page) == 1) {
59 		++rx_queue->page_recycle_count;
60 		return page;
61 	} else {
62 		state = page_address(page);
63 		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
64 			       PAGE_SIZE << efx->rx_buffer_order,
65 			       DMA_FROM_DEVICE);
66 		put_page(page);
67 		++rx_queue->page_recycle_failed;
68 	}
69 
70 	return NULL;
71 }
72 
73 /* Attempt to recycle the page if there is an RX recycle ring; the page can
74  * only be added if this is the final RX buffer, to prevent pages being used in
75  * the descriptor ring and appearing in the recycle ring simultaneously.
76  */
77 static void efx_recycle_rx_page(struct efx_channel *channel,
78 				struct efx_rx_buffer *rx_buf)
79 {
80 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
81 	struct efx_nic *efx = rx_queue->efx;
82 	struct page *page = rx_buf->page;
83 	unsigned int index;
84 
85 	/* Only recycle the page after processing the final buffer. */
86 	if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
87 		return;
88 
89 	index = rx_queue->page_add & rx_queue->page_ptr_mask;
90 	if (rx_queue->page_ring[index] == NULL) {
91 		unsigned int read_index = rx_queue->page_remove &
92 			rx_queue->page_ptr_mask;
93 
94 		/* The next slot in the recycle ring is available, but
95 		 * increment page_remove if the read pointer currently
96 		 * points here.
97 		 */
98 		if (read_index == index)
99 			++rx_queue->page_remove;
100 		rx_queue->page_ring[index] = page;
101 		++rx_queue->page_add;
102 		return;
103 	}
104 	++rx_queue->page_recycle_full;
105 	efx_unmap_rx_buffer(efx, rx_buf);
106 	put_page(rx_buf->page);
107 }
108 
109 /* Recycle the pages that are used by buffers that have just been received. */
110 void efx_siena_recycle_rx_pages(struct efx_channel *channel,
111 				struct efx_rx_buffer *rx_buf,
112 				unsigned int n_frags)
113 {
114 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
115 
116 	if (unlikely(!rx_queue->page_ring))
117 		return;
118 
119 	do {
120 		efx_recycle_rx_page(channel, rx_buf);
121 		rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
122 	} while (--n_frags);
123 }
124 
125 void efx_siena_discard_rx_packet(struct efx_channel *channel,
126 				 struct efx_rx_buffer *rx_buf,
127 				 unsigned int n_frags)
128 {
129 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
130 
131 	efx_siena_recycle_rx_pages(channel, rx_buf, n_frags);
132 
133 	efx_siena_free_rx_buffers(rx_queue, rx_buf, n_frags);
134 }
135 
136 static void efx_init_rx_recycle_ring(struct efx_rx_queue *rx_queue)
137 {
138 	unsigned int bufs_in_recycle_ring, page_ring_size;
139 	struct efx_nic *efx = rx_queue->efx;
140 
141 	bufs_in_recycle_ring = efx_rx_recycle_ring_size(efx);
142 	page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
143 					    efx->rx_bufs_per_page);
144 	rx_queue->page_ring = kzalloc_objs(*rx_queue->page_ring, page_ring_size);
145 	if (!rx_queue->page_ring)
146 		rx_queue->page_ptr_mask = 0;
147 	else
148 		rx_queue->page_ptr_mask = page_ring_size - 1;
149 }
150 
151 static void efx_fini_rx_recycle_ring(struct efx_rx_queue *rx_queue)
152 {
153 	struct efx_nic *efx = rx_queue->efx;
154 	int i;
155 
156 	if (unlikely(!rx_queue->page_ring))
157 		return;
158 
159 	/* Unmap and release the pages in the recycle ring. Remove the ring. */
160 	for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
161 		struct page *page = rx_queue->page_ring[i];
162 		struct efx_rx_page_state *state;
163 
164 		if (page == NULL)
165 			continue;
166 
167 		state = page_address(page);
168 		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
169 			       PAGE_SIZE << efx->rx_buffer_order,
170 			       DMA_FROM_DEVICE);
171 		put_page(page);
172 	}
173 	kfree(rx_queue->page_ring);
174 	rx_queue->page_ring = NULL;
175 }
176 
177 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
178 			       struct efx_rx_buffer *rx_buf)
179 {
180 	/* Release the page reference we hold for the buffer. */
181 	if (rx_buf->page)
182 		put_page(rx_buf->page);
183 
184 	/* If this is the last buffer in a page, unmap and free it. */
185 	if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
186 		efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
187 		efx_siena_free_rx_buffers(rx_queue, rx_buf, 1);
188 	}
189 	rx_buf->page = NULL;
190 }
191 
192 int efx_siena_probe_rx_queue(struct efx_rx_queue *rx_queue)
193 {
194 	struct efx_nic *efx = rx_queue->efx;
195 	unsigned int entries;
196 	int rc;
197 
198 	/* Create the smallest power-of-two aligned ring */
199 	entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
200 	EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
201 	rx_queue->ptr_mask = entries - 1;
202 
203 	netif_dbg(efx, probe, efx->net_dev,
204 		  "creating RX queue %d size %#x mask %#x\n",
205 		  efx_rx_queue_index(rx_queue), efx->rxq_entries,
206 		  rx_queue->ptr_mask);
207 
208 	/* Allocate RX buffers */
209 	rx_queue->buffer = kzalloc_objs(*rx_queue->buffer, entries);
210 	if (!rx_queue->buffer)
211 		return -ENOMEM;
212 
213 	rc = efx_nic_probe_rx(rx_queue);
214 	if (rc) {
215 		kfree(rx_queue->buffer);
216 		rx_queue->buffer = NULL;
217 	}
218 
219 	return rc;
220 }
221 
222 void efx_siena_init_rx_queue(struct efx_rx_queue *rx_queue)
223 {
224 	unsigned int max_fill, trigger, max_trigger;
225 	struct efx_nic *efx = rx_queue->efx;
226 	int rc = 0;
227 
228 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
229 		  "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
230 
231 	/* Initialise ptr fields */
232 	rx_queue->added_count = 0;
233 	rx_queue->notified_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 	}
270 
271 	/* Set up RX descriptor ring */
272 	efx_nic_init_rx(rx_queue);
273 }
274 
275 void efx_siena_fini_rx_queue(struct efx_rx_queue *rx_queue)
276 {
277 	struct efx_rx_buffer *rx_buf;
278 	int i;
279 
280 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
281 		  "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
282 
283 	timer_delete_sync(&rx_queue->slow_fill);
284 
285 	/* Release RX buffers from the current read ptr to the write ptr */
286 	if (rx_queue->buffer) {
287 		for (i = rx_queue->removed_count; i < rx_queue->added_count;
288 		     i++) {
289 			unsigned int index = i & rx_queue->ptr_mask;
290 
291 			rx_buf = efx_rx_buffer(rx_queue, index);
292 			efx_fini_rx_buffer(rx_queue, rx_buf);
293 		}
294 	}
295 
296 	efx_fini_rx_recycle_ring(rx_queue);
297 
298 	if (xdp_rxq_info_is_reg(&rx_queue->xdp_rxq_info))
299 		xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info);
300 }
301 
302 void efx_siena_remove_rx_queue(struct efx_rx_queue *rx_queue)
303 {
304 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
305 		  "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
306 
307 	efx_nic_remove_rx(rx_queue);
308 
309 	kfree(rx_queue->buffer);
310 	rx_queue->buffer = NULL;
311 }
312 
313 /* Unmap a DMA-mapped page.  This function is only called for the final RX
314  * buffer in a page.
315  */
316 static void efx_unmap_rx_buffer(struct efx_nic *efx,
317 				struct efx_rx_buffer *rx_buf)
318 {
319 	struct page *page = rx_buf->page;
320 
321 	if (page) {
322 		struct efx_rx_page_state *state = page_address(page);
323 
324 		dma_unmap_page(&efx->pci_dev->dev,
325 			       state->dma_addr,
326 			       PAGE_SIZE << efx->rx_buffer_order,
327 			       DMA_FROM_DEVICE);
328 	}
329 }
330 
331 void efx_siena_free_rx_buffers(struct efx_rx_queue *rx_queue,
332 			       struct efx_rx_buffer *rx_buf,
333 			       unsigned int num_bufs)
334 {
335 	do {
336 		if (rx_buf->page) {
337 			put_page(rx_buf->page);
338 			rx_buf->page = NULL;
339 		}
340 		rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
341 	} while (--num_bufs);
342 }
343 
344 void efx_siena_rx_slow_fill(struct timer_list *t)
345 {
346 	struct efx_rx_queue *rx_queue = timer_container_of(rx_queue, t,
347 							   slow_fill);
348 
349 	/* Post an event to cause NAPI to run and refill the queue */
350 	efx_nic_generate_fill_event(rx_queue);
351 	++rx_queue->slow_fill_count;
352 }
353 
354 static void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
355 {
356 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(10));
357 }
358 
359 /* efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
360  *
361  * @rx_queue:		Efx RX queue
362  *
363  * This allocates a batch of pages, maps them for DMA, and populates
364  * struct efx_rx_buffers for each one. Return a negative error code or
365  * 0 on success. If a single page can be used for multiple buffers,
366  * then the page will either be inserted fully, or not at all.
367  */
368 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
369 {
370 	unsigned int page_offset, index, count;
371 	struct efx_nic *efx = rx_queue->efx;
372 	struct efx_rx_page_state *state;
373 	struct efx_rx_buffer *rx_buf;
374 	dma_addr_t dma_addr;
375 	struct page *page;
376 
377 	count = 0;
378 	do {
379 		page = efx_reuse_page(rx_queue);
380 		if (page == NULL) {
381 			page = alloc_pages(__GFP_COMP |
382 					   (atomic ? GFP_ATOMIC : GFP_KERNEL),
383 					   efx->rx_buffer_order);
384 			if (unlikely(page == NULL))
385 				return -ENOMEM;
386 			dma_addr =
387 				dma_map_page(&efx->pci_dev->dev, page, 0,
388 					     PAGE_SIZE << efx->rx_buffer_order,
389 					     DMA_FROM_DEVICE);
390 			if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
391 						       dma_addr))) {
392 				__free_pages(page, efx->rx_buffer_order);
393 				return -EIO;
394 			}
395 			state = page_address(page);
396 			state->dma_addr = dma_addr;
397 		} else {
398 			state = page_address(page);
399 			dma_addr = state->dma_addr;
400 		}
401 
402 		dma_addr += sizeof(struct efx_rx_page_state);
403 		page_offset = sizeof(struct efx_rx_page_state);
404 
405 		do {
406 			index = rx_queue->added_count & rx_queue->ptr_mask;
407 			rx_buf = efx_rx_buffer(rx_queue, index);
408 			rx_buf->dma_addr = dma_addr + efx->rx_ip_align +
409 					   EFX_XDP_HEADROOM;
410 			rx_buf->page = page;
411 			rx_buf->page_offset = page_offset + efx->rx_ip_align +
412 					      EFX_XDP_HEADROOM;
413 			rx_buf->len = efx->rx_dma_len;
414 			rx_buf->flags = 0;
415 			++rx_queue->added_count;
416 			get_page(page);
417 			dma_addr += efx->rx_page_buf_step;
418 			page_offset += efx->rx_page_buf_step;
419 		} while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
420 
421 		rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
422 	} while (++count < efx->rx_pages_per_batch);
423 
424 	return 0;
425 }
426 
427 void efx_siena_rx_config_page_split(struct efx_nic *efx)
428 {
429 	efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align +
430 				      EFX_XDP_HEADROOM + EFX_XDP_TAILROOM,
431 				      EFX_RX_BUF_ALIGNMENT);
432 	efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
433 		((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
434 		efx->rx_page_buf_step);
435 	efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
436 		efx->rx_bufs_per_page;
437 	efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
438 					       efx->rx_bufs_per_page);
439 }
440 
441 /* efx_siena_fast_push_rx_descriptors - push new RX descriptors quickly
442  * @rx_queue:		RX descriptor queue
443  *
444  * This will aim to fill the RX descriptor queue up to
445  * @rx_queue->@max_fill. If there is insufficient atomic
446  * memory to do so, a slow fill will be scheduled.
447  *
448  * The caller must provide serialisation (none is used here). In practise,
449  * this means this function must run from the NAPI handler, or be called
450  * when NAPI is disabled.
451  */
452 void efx_siena_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue,
453 					bool atomic)
454 {
455 	struct efx_nic *efx = rx_queue->efx;
456 	unsigned int fill_level, batch_size;
457 	int space, rc = 0;
458 
459 	if (!rx_queue->refill_enabled)
460 		return;
461 
462 	/* Calculate current fill level, and exit if we don't need to fill */
463 	fill_level = (rx_queue->added_count - rx_queue->removed_count);
464 	EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries);
465 	if (fill_level >= rx_queue->fast_fill_trigger)
466 		goto out;
467 
468 	/* Record minimum fill level */
469 	if (unlikely(fill_level < rx_queue->min_fill)) {
470 		if (fill_level)
471 			rx_queue->min_fill = fill_level;
472 	}
473 
474 	batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
475 	space = rx_queue->max_fill - fill_level;
476 	EFX_WARN_ON_ONCE_PARANOID(space < batch_size);
477 
478 	netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
479 		   "RX queue %d fast-filling descriptor ring from"
480 		   " level %d to level %d\n",
481 		   efx_rx_queue_index(rx_queue), fill_level,
482 		   rx_queue->max_fill);
483 
484 	do {
485 		rc = efx_init_rx_buffers(rx_queue, atomic);
486 		if (unlikely(rc)) {
487 			/* Ensure that we don't leave the rx queue empty */
488 			efx_schedule_slow_fill(rx_queue);
489 			goto out;
490 		}
491 	} while ((space -= batch_size) >= batch_size);
492 
493 	netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
494 		   "RX queue %d fast-filled descriptor ring "
495 		   "to level %d\n", efx_rx_queue_index(rx_queue),
496 		   rx_queue->added_count - rx_queue->removed_count);
497 
498  out:
499 	if (rx_queue->notified_count != rx_queue->added_count)
500 		efx_nic_notify_rx_desc(rx_queue);
501 }
502 
503 /* Pass a received packet up through GRO.  GRO can handle pages
504  * regardless of checksum state and skbs with a good checksum.
505  */
506 void
507 efx_siena_rx_packet_gro(struct efx_channel *channel,
508 			struct efx_rx_buffer *rx_buf,
509 			unsigned int n_frags, u8 *eh, __wsum csum)
510 {
511 	struct napi_struct *napi = &channel->napi_str;
512 	struct efx_nic *efx = channel->efx;
513 	struct sk_buff *skb;
514 
515 	skb = napi_get_frags(napi);
516 	if (unlikely(!skb)) {
517 		struct efx_rx_queue *rx_queue;
518 
519 		rx_queue = efx_channel_get_rx_queue(channel);
520 		efx_siena_free_rx_buffers(rx_queue, rx_buf, n_frags);
521 		return;
522 	}
523 
524 	if (efx->net_dev->features & NETIF_F_RXHASH)
525 		skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
526 			     PKT_HASH_TYPE_L3);
527 	if (csum) {
528 		skb->csum = csum;
529 		skb->ip_summed = CHECKSUM_COMPLETE;
530 	} else {
531 		skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
532 				  CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
533 	}
534 	skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
535 
536 	for (;;) {
537 		skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
538 				   rx_buf->page, rx_buf->page_offset,
539 				   rx_buf->len);
540 		rx_buf->page = NULL;
541 		skb->len += rx_buf->len;
542 		if (skb_shinfo(skb)->nr_frags == n_frags)
543 			break;
544 
545 		rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
546 	}
547 
548 	skb->data_len = skb->len;
549 	skb->truesize += n_frags * efx->rx_buffer_truesize;
550 
551 	skb_record_rx_queue(skb, channel->rx_queue.core_index);
552 
553 	napi_gro_frags(napi);
554 }
555 
556 void efx_siena_set_default_rx_indir_table(struct efx_nic *efx,
557 					  struct efx_rss_context *ctx)
558 {
559 	size_t i;
560 
561 	for (i = 0; i < ARRAY_SIZE(ctx->rx_indir_table); i++)
562 		ctx->rx_indir_table[i] =
563 			ethtool_rxfh_indir_default(i, efx->rss_spread);
564 }
565 
566 /**
567  * efx_siena_filter_is_mc_recipient - test whether spec is a multicast recipient
568  * @spec: Specification to test
569  *
570  * Return: %true if the specification is a non-drop RX filter that
571  * matches a local MAC address I/G bit value of 1 or matches a local
572  * IPv4 or IPv6 address value in the respective multicast address
573  * range.  Otherwise %false.
574  */
575 bool efx_siena_filter_is_mc_recipient(const struct efx_filter_spec *spec)
576 {
577 	if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
578 	    spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
579 		return false;
580 
581 	if (spec->match_flags &
582 	    (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
583 	    is_multicast_ether_addr(spec->loc_mac))
584 		return true;
585 
586 	if ((spec->match_flags &
587 	     (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
588 	    (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
589 		if (spec->ether_type == htons(ETH_P_IP) &&
590 		    ipv4_is_multicast(spec->loc_host[0]))
591 			return true;
592 		if (spec->ether_type == htons(ETH_P_IPV6) &&
593 		    ((const u8 *)spec->loc_host)[0] == 0xff)
594 			return true;
595 	}
596 
597 	return false;
598 }
599 
600 bool efx_siena_filter_spec_equal(const struct efx_filter_spec *left,
601 				 const struct efx_filter_spec *right)
602 {
603 	if ((left->match_flags ^ right->match_flags) |
604 	    ((left->flags ^ right->flags) &
605 	     (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
606 		return false;
607 
608 	return memcmp(&left->outer_vid, &right->outer_vid,
609 		      sizeof(struct efx_filter_spec) -
610 		      offsetof(struct efx_filter_spec, outer_vid)) == 0;
611 }
612 
613 u32 efx_siena_filter_spec_hash(const struct efx_filter_spec *spec)
614 {
615 	BUILD_BUG_ON(offsetof(struct efx_filter_spec, outer_vid) & 3);
616 	return jhash2((const u32 *)&spec->outer_vid,
617 		      (sizeof(struct efx_filter_spec) -
618 		       offsetof(struct efx_filter_spec, outer_vid)) / 4,
619 		      0);
620 }
621 
622 #ifdef CONFIG_RFS_ACCEL
623 bool efx_siena_rps_check_rule(struct efx_arfs_rule *rule,
624 			      unsigned int filter_idx, bool *force)
625 {
626 	if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) {
627 		/* ARFS is currently updating this entry, leave it */
628 		return false;
629 	}
630 	if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) {
631 		/* ARFS tried and failed to update this, so it's probably out
632 		 * of date.  Remove the filter and the ARFS rule entry.
633 		 */
634 		rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
635 		*force = true;
636 		return true;
637 	} else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */
638 		/* ARFS has moved on, so old filter is not needed.  Since we did
639 		 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will
640 		 * not be removed by efx_siena_rps_hash_del() subsequently.
641 		 */
642 		*force = true;
643 		return true;
644 	}
645 	/* Remove it iff ARFS wants to. */
646 	return true;
647 }
648 
649 static
650 struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx,
651 				       const struct efx_filter_spec *spec)
652 {
653 	u32 hash = efx_siena_filter_spec_hash(spec);
654 
655 	lockdep_assert_held(&efx->rps_hash_lock);
656 	if (!efx->rps_hash_table)
657 		return NULL;
658 	return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE];
659 }
660 
661 struct efx_arfs_rule *efx_siena_rps_hash_find(struct efx_nic *efx,
662 					const struct efx_filter_spec *spec)
663 {
664 	struct efx_arfs_rule *rule;
665 	struct hlist_head *head;
666 	struct hlist_node *node;
667 
668 	head = efx_rps_hash_bucket(efx, spec);
669 	if (!head)
670 		return NULL;
671 	hlist_for_each(node, head) {
672 		rule = container_of(node, struct efx_arfs_rule, node);
673 		if (efx_siena_filter_spec_equal(spec, &rule->spec))
674 			return rule;
675 	}
676 	return NULL;
677 }
678 
679 static struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
680 					const struct efx_filter_spec *spec,
681 					bool *new)
682 {
683 	struct efx_arfs_rule *rule;
684 	struct hlist_head *head;
685 	struct hlist_node *node;
686 
687 	head = efx_rps_hash_bucket(efx, spec);
688 	if (!head)
689 		return NULL;
690 	hlist_for_each(node, head) {
691 		rule = container_of(node, struct efx_arfs_rule, node);
692 		if (efx_siena_filter_spec_equal(spec, &rule->spec)) {
693 			*new = false;
694 			return rule;
695 		}
696 	}
697 	rule = kmalloc_obj(*rule, GFP_ATOMIC);
698 	*new = true;
699 	if (rule) {
700 		memcpy(&rule->spec, spec, sizeof(rule->spec));
701 		hlist_add_head(&rule->node, head);
702 	}
703 	return rule;
704 }
705 
706 void efx_siena_rps_hash_del(struct efx_nic *efx,
707 			    const struct efx_filter_spec *spec)
708 {
709 	struct efx_arfs_rule *rule;
710 	struct hlist_head *head;
711 	struct hlist_node *node;
712 
713 	head = efx_rps_hash_bucket(efx, spec);
714 	if (WARN_ON(!head))
715 		return;
716 	hlist_for_each(node, head) {
717 		rule = container_of(node, struct efx_arfs_rule, node);
718 		if (efx_siena_filter_spec_equal(spec, &rule->spec)) {
719 			/* Someone already reused the entry.  We know that if
720 			 * this check doesn't fire (i.e. filter_id == REMOVING)
721 			 * then the REMOVING mark was put there by our caller,
722 			 * because caller is holding a lock on filter table and
723 			 * only holders of that lock set REMOVING.
724 			 */
725 			if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING)
726 				return;
727 			hlist_del(node);
728 			kfree(rule);
729 			return;
730 		}
731 	}
732 	/* We didn't find it. */
733 	WARN_ON(1);
734 }
735 #endif
736 
737 int efx_siena_probe_filters(struct efx_nic *efx)
738 {
739 	int rc;
740 
741 	mutex_lock(&efx->mac_lock);
742 	down_write(&efx->filter_sem);
743 	rc = efx->type->filter_table_probe(efx);
744 	if (rc)
745 		goto out_unlock;
746 
747 #ifdef CONFIG_RFS_ACCEL
748 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
749 		struct efx_channel *channel;
750 		int i, success = 1;
751 
752 		efx_for_each_channel(channel, efx) {
753 			channel->rps_flow_id =
754 				kcalloc(efx->type->max_rx_ip_filters,
755 					sizeof(*channel->rps_flow_id),
756 					GFP_KERNEL);
757 			if (!channel->rps_flow_id)
758 				success = 0;
759 			else
760 				for (i = 0;
761 				     i < efx->type->max_rx_ip_filters;
762 				     ++i)
763 					channel->rps_flow_id[i] =
764 						RPS_FLOW_ID_INVALID;
765 			channel->rfs_expire_index = 0;
766 			channel->rfs_filter_count = 0;
767 		}
768 
769 		if (!success) {
770 			efx_for_each_channel(channel, efx)
771 				kfree(channel->rps_flow_id);
772 			efx->type->filter_table_remove(efx);
773 			rc = -ENOMEM;
774 			goto out_unlock;
775 		}
776 	}
777 #endif
778 out_unlock:
779 	up_write(&efx->filter_sem);
780 	mutex_unlock(&efx->mac_lock);
781 	return rc;
782 }
783 
784 void efx_siena_remove_filters(struct efx_nic *efx)
785 {
786 #ifdef CONFIG_RFS_ACCEL
787 	struct efx_channel *channel;
788 
789 	efx_for_each_channel(channel, efx) {
790 		cancel_delayed_work_sync(&channel->filter_work);
791 		kfree(channel->rps_flow_id);
792 		channel->rps_flow_id = NULL;
793 	}
794 #endif
795 	down_write(&efx->filter_sem);
796 	efx->type->filter_table_remove(efx);
797 	up_write(&efx->filter_sem);
798 }
799 
800 #ifdef CONFIG_RFS_ACCEL
801 
802 static void efx_filter_rfs_work(struct work_struct *data)
803 {
804 	struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
805 							      work);
806 	struct efx_nic *efx = netdev_priv(req->net_dev);
807 	struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
808 	int slot_idx = req - efx->rps_slot;
809 	struct efx_arfs_rule *rule;
810 	u16 arfs_id = 0;
811 	int rc;
812 
813 	rc = efx->type->filter_insert(efx, &req->spec, true);
814 	if (rc >= 0)
815 		/* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */
816 		rc %= efx->type->max_rx_ip_filters;
817 	if (efx->rps_hash_table) {
818 		spin_lock_bh(&efx->rps_hash_lock);
819 		rule = efx_siena_rps_hash_find(efx, &req->spec);
820 		/* The rule might have already gone, if someone else's request
821 		 * for the same spec was already worked and then expired before
822 		 * we got around to our work.  In that case we have nothing
823 		 * tying us to an arfs_id, meaning that as soon as the filter
824 		 * is considered for expiry it will be removed.
825 		 */
826 		if (rule) {
827 			if (rc < 0)
828 				rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
829 			else
830 				rule->filter_id = rc;
831 			arfs_id = rule->arfs_id;
832 		}
833 		spin_unlock_bh(&efx->rps_hash_lock);
834 	}
835 	if (rc >= 0) {
836 		/* Remember this so we can check whether to expire the filter
837 		 * later.
838 		 */
839 		mutex_lock(&efx->rps_mutex);
840 		if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID)
841 			channel->rfs_filter_count++;
842 		channel->rps_flow_id[rc] = req->flow_id;
843 		mutex_unlock(&efx->rps_mutex);
844 
845 		if (req->spec.ether_type == htons(ETH_P_IP))
846 			netif_info(efx, rx_status, efx->net_dev,
847 				   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
848 				   (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
849 				   req->spec.rem_host, ntohs(req->spec.rem_port),
850 				   req->spec.loc_host, ntohs(req->spec.loc_port),
851 				   req->rxq_index, req->flow_id, rc, arfs_id);
852 		else
853 			netif_info(efx, rx_status, efx->net_dev,
854 				   "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
855 				   (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
856 				   req->spec.rem_host, ntohs(req->spec.rem_port),
857 				   req->spec.loc_host, ntohs(req->spec.loc_port),
858 				   req->rxq_index, req->flow_id, rc, arfs_id);
859 		channel->n_rfs_succeeded++;
860 	} else {
861 		if (req->spec.ether_type == htons(ETH_P_IP))
862 			netif_dbg(efx, rx_status, efx->net_dev,
863 				  "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n",
864 				  (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
865 				  req->spec.rem_host, ntohs(req->spec.rem_port),
866 				  req->spec.loc_host, ntohs(req->spec.loc_port),
867 				  req->rxq_index, req->flow_id, rc, arfs_id);
868 		else
869 			netif_dbg(efx, rx_status, efx->net_dev,
870 				  "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n",
871 				  (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
872 				  req->spec.rem_host, ntohs(req->spec.rem_port),
873 				  req->spec.loc_host, ntohs(req->spec.loc_port),
874 				  req->rxq_index, req->flow_id, rc, arfs_id);
875 		channel->n_rfs_failed++;
876 		/* We're overloading the NIC's filter tables, so let's do a
877 		 * chunk of extra expiry work.
878 		 */
879 		__efx_siena_filter_rfs_expire(channel,
880 					      min(channel->rfs_filter_count,
881 						  100u));
882 	}
883 
884 	/* Release references */
885 	clear_bit(slot_idx, &efx->rps_slot_map);
886 	netdev_put(req->net_dev, &req->net_dev_tracker);
887 }
888 
889 int efx_siena_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
890 			 u16 rxq_index, u32 flow_id)
891 {
892 	struct efx_nic *efx = netdev_priv(net_dev);
893 	struct efx_async_filter_insertion *req;
894 	struct efx_arfs_rule *rule;
895 	struct flow_keys fk;
896 	int slot_idx;
897 	bool new;
898 	int rc;
899 
900 	/* find a free slot */
901 	for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
902 		if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
903 			break;
904 	if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
905 		return -EBUSY;
906 
907 	if (flow_id == RPS_FLOW_ID_INVALID) {
908 		rc = -EINVAL;
909 		goto out_clear;
910 	}
911 
912 	if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
913 		rc = -EPROTONOSUPPORT;
914 		goto out_clear;
915 	}
916 
917 	if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
918 		rc = -EPROTONOSUPPORT;
919 		goto out_clear;
920 	}
921 	if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
922 		rc = -EPROTONOSUPPORT;
923 		goto out_clear;
924 	}
925 
926 	req = efx->rps_slot + slot_idx;
927 	efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
928 			   efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
929 			   rxq_index);
930 	req->spec.match_flags =
931 		EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
932 		EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
933 		EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
934 	req->spec.ether_type = fk.basic.n_proto;
935 	req->spec.ip_proto = fk.basic.ip_proto;
936 
937 	if (fk.basic.n_proto == htons(ETH_P_IP)) {
938 		req->spec.rem_host[0] = fk.addrs.v4addrs.src;
939 		req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
940 	} else {
941 		memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
942 		       sizeof(struct in6_addr));
943 		memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
944 		       sizeof(struct in6_addr));
945 	}
946 
947 	req->spec.rem_port = fk.ports.src;
948 	req->spec.loc_port = fk.ports.dst;
949 
950 	if (efx->rps_hash_table) {
951 		/* Add it to ARFS hash table */
952 		spin_lock(&efx->rps_hash_lock);
953 		rule = efx_rps_hash_add(efx, &req->spec, &new);
954 		if (!rule) {
955 			rc = -ENOMEM;
956 			goto out_unlock;
957 		}
958 		if (new)
959 			rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
960 		rc = rule->arfs_id;
961 		/* Skip if existing or pending filter already does the right thing */
962 		if (!new && rule->rxq_index == rxq_index &&
963 		    rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
964 			goto out_unlock;
965 		rule->rxq_index = rxq_index;
966 		rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
967 		spin_unlock(&efx->rps_hash_lock);
968 	} else {
969 		/* Without an ARFS hash table, we just use arfs_id 0 for all
970 		 * filters.  This means if multiple flows hash to the same
971 		 * flow_id, all but the most recently touched will be eligible
972 		 * for expiry.
973 		 */
974 		rc = 0;
975 	}
976 
977 	/* Queue the request */
978 	req->net_dev = net_dev;
979 	netdev_hold(req->net_dev, &req->net_dev_tracker, GFP_ATOMIC);
980 	INIT_WORK(&req->work, efx_filter_rfs_work);
981 	req->rxq_index = rxq_index;
982 	req->flow_id = flow_id;
983 	schedule_work(&req->work);
984 	return rc;
985 out_unlock:
986 	spin_unlock(&efx->rps_hash_lock);
987 out_clear:
988 	clear_bit(slot_idx, &efx->rps_slot_map);
989 	return rc;
990 }
991 
992 bool __efx_siena_filter_rfs_expire(struct efx_channel *channel,
993 				   unsigned int quota)
994 {
995 	bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
996 	struct efx_nic *efx = channel->efx;
997 	unsigned int index, size, start;
998 	u32 flow_id;
999 
1000 	if (!mutex_trylock(&efx->rps_mutex))
1001 		return false;
1002 	expire_one = efx->type->filter_rfs_expire_one;
1003 	index = channel->rfs_expire_index;
1004 	start = index;
1005 	size = efx->type->max_rx_ip_filters;
1006 	while (quota) {
1007 		flow_id = channel->rps_flow_id[index];
1008 
1009 		if (flow_id != RPS_FLOW_ID_INVALID) {
1010 			quota--;
1011 			if (expire_one(efx, flow_id, index)) {
1012 				netif_info(efx, rx_status, efx->net_dev,
1013 					   "expired filter %d [channel %u flow %u]\n",
1014 					   index, channel->channel, flow_id);
1015 				channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1016 				channel->rfs_filter_count--;
1017 			}
1018 		}
1019 		if (++index == size)
1020 			index = 0;
1021 		/* If we were called with a quota that exceeds the total number
1022 		 * of filters in the table (which shouldn't happen, but could
1023 		 * if two callers race), ensure that we don't loop forever -
1024 		 * stop when we've examined every row of the table.
1025 		 */
1026 		if (index == start)
1027 			break;
1028 	}
1029 
1030 	channel->rfs_expire_index = index;
1031 	mutex_unlock(&efx->rps_mutex);
1032 	return true;
1033 }
1034 
1035 #endif /* CONFIG_RFS_ACCEL */
1036