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 "efx.h"
15 #include "nic.h"
16 #include "rx_common.h"
17 
18 /* This is the percentage fill level below which new RX descriptors
19  * will be added to the RX descriptor ring.
20  */
21 static unsigned int rx_refill_threshold;
22 module_param(rx_refill_threshold, uint, 0444);
23 MODULE_PARM_DESC(rx_refill_threshold,
24 		 "RX descriptor ring refill threshold (%)");
25 
26 /* RX maximum head room required.
27  *
28  * This must be at least 1 to prevent overflow, plus one packet-worth
29  * to allow pipelined receives.
30  */
31 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
32 
33 /* Check the RX page recycle ring for a page that can be reused. */
34 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
35 {
36 	struct efx_nic *efx = rx_queue->efx;
37 	struct efx_rx_page_state *state;
38 	unsigned int index;
39 	struct page *page;
40 
41 	if (unlikely(!rx_queue->page_ring))
42 		return NULL;
43 	index = rx_queue->page_remove & rx_queue->page_ptr_mask;
44 	page = rx_queue->page_ring[index];
45 	if (page == NULL)
46 		return NULL;
47 
48 	rx_queue->page_ring[index] = NULL;
49 	/* page_remove cannot exceed page_add. */
50 	if (rx_queue->page_remove != rx_queue->page_add)
51 		++rx_queue->page_remove;
52 
53 	/* If page_count is 1 then we hold the only reference to this page. */
54 	if (page_count(page) == 1) {
55 		++rx_queue->page_recycle_count;
56 		return page;
57 	} else {
58 		state = page_address(page);
59 		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
60 			       PAGE_SIZE << efx->rx_buffer_order,
61 			       DMA_FROM_DEVICE);
62 		put_page(page);
63 		++rx_queue->page_recycle_failed;
64 	}
65 
66 	return NULL;
67 }
68 
69 /* Attempt to recycle the page if there is an RX recycle ring; the page can
70  * only be added if this is the final RX buffer, to prevent pages being used in
71  * the descriptor ring and appearing in the recycle ring simultaneously.
72  */
73 static void efx_recycle_rx_page(struct efx_channel *channel,
74 				struct efx_rx_buffer *rx_buf)
75 {
76 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
77 	struct efx_nic *efx = rx_queue->efx;
78 	struct page *page = rx_buf->page;
79 	unsigned int index;
80 
81 	/* Only recycle the page after processing the final buffer. */
82 	if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
83 		return;
84 
85 	index = rx_queue->page_add & rx_queue->page_ptr_mask;
86 	if (rx_queue->page_ring[index] == NULL) {
87 		unsigned int read_index = rx_queue->page_remove &
88 			rx_queue->page_ptr_mask;
89 
90 		/* The next slot in the recycle ring is available, but
91 		 * increment page_remove if the read pointer currently
92 		 * points here.
93 		 */
94 		if (read_index == index)
95 			++rx_queue->page_remove;
96 		rx_queue->page_ring[index] = page;
97 		++rx_queue->page_add;
98 		return;
99 	}
100 	++rx_queue->page_recycle_full;
101 	efx_unmap_rx_buffer(efx, rx_buf);
102 	put_page(rx_buf->page);
103 }
104 
105 /* Recycle the pages that are used by buffers that have just been received. */
106 void efx_recycle_rx_pages(struct efx_channel *channel,
107 			  struct efx_rx_buffer *rx_buf,
108 			  unsigned int n_frags)
109 {
110 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
111 
112 	if (unlikely(!rx_queue->page_ring))
113 		return;
114 
115 	do {
116 		efx_recycle_rx_page(channel, rx_buf);
117 		rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
118 	} while (--n_frags);
119 }
120 
121 void efx_discard_rx_packet(struct efx_channel *channel,
122 			   struct efx_rx_buffer *rx_buf,
123 			   unsigned int n_frags)
124 {
125 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
126 
127 	efx_recycle_rx_pages(channel, rx_buf, n_frags);
128 
129 	efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
130 }
131 
132 static void efx_init_rx_recycle_ring(struct efx_rx_queue *rx_queue)
133 {
134 	unsigned int bufs_in_recycle_ring, page_ring_size;
135 	struct efx_nic *efx = rx_queue->efx;
136 
137 	bufs_in_recycle_ring = efx_rx_recycle_ring_size(efx);
138 	page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
139 					    efx->rx_bufs_per_page);
140 	rx_queue->page_ring = kcalloc(page_ring_size,
141 				      sizeof(*rx_queue->page_ring), GFP_KERNEL);
142 	if (!rx_queue->page_ring)
143 		rx_queue->page_ptr_mask = 0;
144 	else
145 		rx_queue->page_ptr_mask = page_ring_size - 1;
146 }
147 
148 static void efx_fini_rx_recycle_ring(struct efx_rx_queue *rx_queue)
149 {
150 	struct efx_nic *efx = rx_queue->efx;
151 	int i;
152 
153 	if (unlikely(!rx_queue->page_ring))
154 		return;
155 
156 	/* Unmap and release the pages in the recycle ring. Remove the ring. */
157 	for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
158 		struct page *page = rx_queue->page_ring[i];
159 		struct efx_rx_page_state *state;
160 
161 		if (page == NULL)
162 			continue;
163 
164 		state = page_address(page);
165 		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
166 			       PAGE_SIZE << efx->rx_buffer_order,
167 			       DMA_FROM_DEVICE);
168 		put_page(page);
169 	}
170 	kfree(rx_queue->page_ring);
171 	rx_queue->page_ring = NULL;
172 }
173 
174 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
175 			       struct efx_rx_buffer *rx_buf)
176 {
177 	/* Release the page reference we hold for the buffer. */
178 	if (rx_buf->page)
179 		put_page(rx_buf->page);
180 
181 	/* If this is the last buffer in a page, unmap and free it. */
182 	if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
183 		efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
184 		efx_free_rx_buffers(rx_queue, rx_buf, 1);
185 	}
186 	rx_buf->page = NULL;
187 }
188 
189 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
190 {
191 	struct efx_nic *efx = rx_queue->efx;
192 	unsigned int entries;
193 	int rc;
194 
195 	/* Create the smallest power-of-two aligned ring */
196 	entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
197 	EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
198 	rx_queue->ptr_mask = entries - 1;
199 
200 	netif_dbg(efx, probe, efx->net_dev,
201 		  "creating RX queue %d size %#x mask %#x\n",
202 		  efx_rx_queue_index(rx_queue), efx->rxq_entries,
203 		  rx_queue->ptr_mask);
204 
205 	/* Allocate RX buffers */
206 	rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
207 				   GFP_KERNEL);
208 	if (!rx_queue->buffer)
209 		return -ENOMEM;
210 
211 	rc = efx_nic_probe_rx(rx_queue);
212 	if (rc) {
213 		kfree(rx_queue->buffer);
214 		rx_queue->buffer = NULL;
215 	}
216 
217 	return rc;
218 }
219 
220 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
221 {
222 	unsigned int max_fill, trigger, max_trigger;
223 	struct efx_nic *efx = rx_queue->efx;
224 	int rc = 0;
225 
226 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
227 		  "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
228 
229 	/* Initialise ptr fields */
230 	rx_queue->added_count = 0;
231 	rx_queue->notified_count = 0;
232 	rx_queue->removed_count = 0;
233 	rx_queue->min_fill = -1U;
234 	efx_init_rx_recycle_ring(rx_queue);
235 
236 	rx_queue->page_remove = 0;
237 	rx_queue->page_add = rx_queue->page_ptr_mask + 1;
238 	rx_queue->page_recycle_count = 0;
239 	rx_queue->page_recycle_failed = 0;
240 	rx_queue->page_recycle_full = 0;
241 
242 	/* Initialise limit fields */
243 	max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
244 	max_trigger =
245 		max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
246 	if (rx_refill_threshold != 0) {
247 		trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
248 		if (trigger > max_trigger)
249 			trigger = max_trigger;
250 	} else {
251 		trigger = max_trigger;
252 	}
253 
254 	rx_queue->max_fill = max_fill;
255 	rx_queue->fast_fill_trigger = trigger;
256 	rx_queue->refill_enabled = true;
257 
258 	/* Initialise XDP queue information */
259 	rc = xdp_rxq_info_reg(&rx_queue->xdp_rxq_info, efx->net_dev,
260 			      rx_queue->core_index, 0);
261 
262 	if (rc) {
263 		netif_err(efx, rx_err, efx->net_dev,
264 			  "Failure to initialise XDP queue information rc=%d\n",
265 			  rc);
266 		efx->xdp_rxq_info_failed = true;
267 	} else {
268 		rx_queue->xdp_rxq_info_valid = true;
269 	}
270 
271 	/* Set up RX descriptor ring */
272 	efx_nic_init_rx(rx_queue);
273 }
274 
275 void efx_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 	del_timer_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 (rx_queue->xdp_rxq_info_valid)
299 		xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info);
300 
301 	rx_queue->xdp_rxq_info_valid = false;
302 }
303 
304 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
305 {
306 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
307 		  "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
308 
309 	efx_nic_remove_rx(rx_queue);
310 
311 	kfree(rx_queue->buffer);
312 	rx_queue->buffer = NULL;
313 }
314 
315 /* Unmap a DMA-mapped page.  This function is only called for the final RX
316  * buffer in a page.
317  */
318 void efx_unmap_rx_buffer(struct efx_nic *efx,
319 			 struct efx_rx_buffer *rx_buf)
320 {
321 	struct page *page = rx_buf->page;
322 
323 	if (page) {
324 		struct efx_rx_page_state *state = page_address(page);
325 
326 		dma_unmap_page(&efx->pci_dev->dev,
327 			       state->dma_addr,
328 			       PAGE_SIZE << efx->rx_buffer_order,
329 			       DMA_FROM_DEVICE);
330 	}
331 }
332 
333 void efx_free_rx_buffers(struct efx_rx_queue *rx_queue,
334 			 struct efx_rx_buffer *rx_buf,
335 			 unsigned int num_bufs)
336 {
337 	do {
338 		if (rx_buf->page) {
339 			put_page(rx_buf->page);
340 			rx_buf->page = NULL;
341 		}
342 		rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
343 	} while (--num_bufs);
344 }
345 
346 void efx_rx_slow_fill(struct timer_list *t)
347 {
348 	struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill);
349 
350 	/* Post an event to cause NAPI to run and refill the queue */
351 	efx_nic_generate_fill_event(rx_queue);
352 	++rx_queue->slow_fill_count;
353 }
354 
355 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
356 {
357 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(10));
358 }
359 
360 /* efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
361  *
362  * @rx_queue:		Efx RX queue
363  *
364  * This allocates a batch of pages, maps them for DMA, and populates
365  * struct efx_rx_buffers for each one. Return a negative error code or
366  * 0 on success. If a single page can be used for multiple buffers,
367  * then the page will either be inserted fully, or not at all.
368  */
369 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
370 {
371 	unsigned int page_offset, index, count;
372 	struct efx_nic *efx = rx_queue->efx;
373 	struct efx_rx_page_state *state;
374 	struct efx_rx_buffer *rx_buf;
375 	dma_addr_t dma_addr;
376 	struct page *page;
377 
378 	count = 0;
379 	do {
380 		page = efx_reuse_page(rx_queue);
381 		if (page == NULL) {
382 			page = alloc_pages(__GFP_COMP |
383 					   (atomic ? GFP_ATOMIC : GFP_KERNEL),
384 					   efx->rx_buffer_order);
385 			if (unlikely(page == NULL))
386 				return -ENOMEM;
387 			dma_addr =
388 				dma_map_page(&efx->pci_dev->dev, page, 0,
389 					     PAGE_SIZE << efx->rx_buffer_order,
390 					     DMA_FROM_DEVICE);
391 			if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
392 						       dma_addr))) {
393 				__free_pages(page, efx->rx_buffer_order);
394 				return -EIO;
395 			}
396 			state = page_address(page);
397 			state->dma_addr = dma_addr;
398 		} else {
399 			state = page_address(page);
400 			dma_addr = state->dma_addr;
401 		}
402 
403 		dma_addr += sizeof(struct efx_rx_page_state);
404 		page_offset = sizeof(struct efx_rx_page_state);
405 
406 		do {
407 			index = rx_queue->added_count & rx_queue->ptr_mask;
408 			rx_buf = efx_rx_buffer(rx_queue, index);
409 			rx_buf->dma_addr = dma_addr + efx->rx_ip_align +
410 					   EFX_XDP_HEADROOM;
411 			rx_buf->page = page;
412 			rx_buf->page_offset = page_offset + efx->rx_ip_align +
413 					      EFX_XDP_HEADROOM;
414 			rx_buf->len = efx->rx_dma_len;
415 			rx_buf->flags = 0;
416 			++rx_queue->added_count;
417 			get_page(page);
418 			dma_addr += efx->rx_page_buf_step;
419 			page_offset += efx->rx_page_buf_step;
420 		} while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
421 
422 		rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
423 	} while (++count < efx->rx_pages_per_batch);
424 
425 	return 0;
426 }
427 
428 void efx_rx_config_page_split(struct efx_nic *efx)
429 {
430 	efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align +
431 				      EFX_XDP_HEADROOM + EFX_XDP_TAILROOM,
432 				      EFX_RX_BUF_ALIGNMENT);
433 	efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
434 		((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
435 		efx->rx_page_buf_step);
436 	efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
437 		efx->rx_bufs_per_page;
438 	efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
439 					       efx->rx_bufs_per_page);
440 }
441 
442 /* efx_fast_push_rx_descriptors - push new RX descriptors quickly
443  * @rx_queue:		RX descriptor queue
444  *
445  * This will aim to fill the RX descriptor queue up to
446  * @rx_queue->@max_fill. If there is insufficient atomic
447  * memory to do so, a slow fill will be scheduled.
448  *
449  * The caller must provide serialisation (none is used here). In practise,
450  * this means this function must run from the NAPI handler, or be called
451  * when NAPI is disabled.
452  */
453 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, 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_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
508 		  unsigned int n_frags, u8 *eh, __wsum csum)
509 {
510 	struct napi_struct *napi = &channel->napi_str;
511 	struct efx_nic *efx = channel->efx;
512 	struct sk_buff *skb;
513 
514 	skb = napi_get_frags(napi);
515 	if (unlikely(!skb)) {
516 		struct efx_rx_queue *rx_queue;
517 
518 		rx_queue = efx_channel_get_rx_queue(channel);
519 		efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
520 		return;
521 	}
522 
523 	if (efx->net_dev->features & NETIF_F_RXHASH &&
524 	    efx_rx_buf_hash_valid(efx, eh))
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 /* RSS contexts.  We're using linked lists and crappy O(n) algorithms, because
557  * (a) this is an infrequent control-plane operation and (b) n is small (max 64)
558  */
559 struct efx_rss_context *efx_alloc_rss_context_entry(struct efx_nic *efx)
560 {
561 	struct list_head *head = &efx->rss_context.list;
562 	struct efx_rss_context *ctx, *new;
563 	u32 id = 1; /* Don't use zero, that refers to the master RSS context */
564 
565 	WARN_ON(!mutex_is_locked(&efx->rss_lock));
566 
567 	/* Search for first gap in the numbering */
568 	list_for_each_entry(ctx, head, list) {
569 		if (ctx->user_id != id)
570 			break;
571 		id++;
572 		/* Check for wrap.  If this happens, we have nearly 2^32
573 		 * allocated RSS contexts, which seems unlikely.
574 		 */
575 		if (WARN_ON_ONCE(!id))
576 			return NULL;
577 	}
578 
579 	/* Create the new entry */
580 	new = kmalloc(sizeof(*new), GFP_KERNEL);
581 	if (!new)
582 		return NULL;
583 	new->context_id = EFX_MCDI_RSS_CONTEXT_INVALID;
584 	new->rx_hash_udp_4tuple = false;
585 
586 	/* Insert the new entry into the gap */
587 	new->user_id = id;
588 	list_add_tail(&new->list, &ctx->list);
589 	return new;
590 }
591 
592 struct efx_rss_context *efx_find_rss_context_entry(struct efx_nic *efx, u32 id)
593 {
594 	struct list_head *head = &efx->rss_context.list;
595 	struct efx_rss_context *ctx;
596 
597 	WARN_ON(!mutex_is_locked(&efx->rss_lock));
598 
599 	list_for_each_entry(ctx, head, list)
600 		if (ctx->user_id == id)
601 			return ctx;
602 	return NULL;
603 }
604 
605 void efx_free_rss_context_entry(struct efx_rss_context *ctx)
606 {
607 	list_del(&ctx->list);
608 	kfree(ctx);
609 }
610 
611 void efx_set_default_rx_indir_table(struct efx_nic *efx,
612 				    struct efx_rss_context *ctx)
613 {
614 	size_t i;
615 
616 	for (i = 0; i < ARRAY_SIZE(ctx->rx_indir_table); i++)
617 		ctx->rx_indir_table[i] =
618 			ethtool_rxfh_indir_default(i, efx->rss_spread);
619 }
620 
621 /**
622  * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
623  * @spec: Specification to test
624  *
625  * Return: %true if the specification is a non-drop RX filter that
626  * matches a local MAC address I/G bit value of 1 or matches a local
627  * IPv4 or IPv6 address value in the respective multicast address
628  * range.  Otherwise %false.
629  */
630 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
631 {
632 	if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
633 	    spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
634 		return false;
635 
636 	if (spec->match_flags &
637 	    (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
638 	    is_multicast_ether_addr(spec->loc_mac))
639 		return true;
640 
641 	if ((spec->match_flags &
642 	     (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
643 	    (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
644 		if (spec->ether_type == htons(ETH_P_IP) &&
645 		    ipv4_is_multicast(spec->loc_host[0]))
646 			return true;
647 		if (spec->ether_type == htons(ETH_P_IPV6) &&
648 		    ((const u8 *)spec->loc_host)[0] == 0xff)
649 			return true;
650 	}
651 
652 	return false;
653 }
654 
655 bool efx_filter_spec_equal(const struct efx_filter_spec *left,
656 			   const struct efx_filter_spec *right)
657 {
658 	if ((left->match_flags ^ right->match_flags) |
659 	    ((left->flags ^ right->flags) &
660 	     (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
661 		return false;
662 
663 	return memcmp(&left->outer_vid, &right->outer_vid,
664 		      sizeof(struct efx_filter_spec) -
665 		      offsetof(struct efx_filter_spec, outer_vid)) == 0;
666 }
667 
668 u32 efx_filter_spec_hash(const struct efx_filter_spec *spec)
669 {
670 	BUILD_BUG_ON(offsetof(struct efx_filter_spec, outer_vid) & 3);
671 	return jhash2((const u32 *)&spec->outer_vid,
672 		      (sizeof(struct efx_filter_spec) -
673 		       offsetof(struct efx_filter_spec, outer_vid)) / 4,
674 		      0);
675 }
676 
677 #ifdef CONFIG_RFS_ACCEL
678 bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx,
679 			bool *force)
680 {
681 	if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) {
682 		/* ARFS is currently updating this entry, leave it */
683 		return false;
684 	}
685 	if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) {
686 		/* ARFS tried and failed to update this, so it's probably out
687 		 * of date.  Remove the filter and the ARFS rule entry.
688 		 */
689 		rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
690 		*force = true;
691 		return true;
692 	} else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */
693 		/* ARFS has moved on, so old filter is not needed.  Since we did
694 		 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will
695 		 * not be removed by efx_rps_hash_del() subsequently.
696 		 */
697 		*force = true;
698 		return true;
699 	}
700 	/* Remove it iff ARFS wants to. */
701 	return true;
702 }
703 
704 static
705 struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx,
706 				       const struct efx_filter_spec *spec)
707 {
708 	u32 hash = efx_filter_spec_hash(spec);
709 
710 	lockdep_assert_held(&efx->rps_hash_lock);
711 	if (!efx->rps_hash_table)
712 		return NULL;
713 	return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE];
714 }
715 
716 struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx,
717 					const struct efx_filter_spec *spec)
718 {
719 	struct efx_arfs_rule *rule;
720 	struct hlist_head *head;
721 	struct hlist_node *node;
722 
723 	head = efx_rps_hash_bucket(efx, spec);
724 	if (!head)
725 		return NULL;
726 	hlist_for_each(node, head) {
727 		rule = container_of(node, struct efx_arfs_rule, node);
728 		if (efx_filter_spec_equal(spec, &rule->spec))
729 			return rule;
730 	}
731 	return NULL;
732 }
733 
734 struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
735 				       const struct efx_filter_spec *spec,
736 				       bool *new)
737 {
738 	struct efx_arfs_rule *rule;
739 	struct hlist_head *head;
740 	struct hlist_node *node;
741 
742 	head = efx_rps_hash_bucket(efx, spec);
743 	if (!head)
744 		return NULL;
745 	hlist_for_each(node, head) {
746 		rule = container_of(node, struct efx_arfs_rule, node);
747 		if (efx_filter_spec_equal(spec, &rule->spec)) {
748 			*new = false;
749 			return rule;
750 		}
751 	}
752 	rule = kmalloc(sizeof(*rule), GFP_ATOMIC);
753 	*new = true;
754 	if (rule) {
755 		memcpy(&rule->spec, spec, sizeof(rule->spec));
756 		hlist_add_head(&rule->node, head);
757 	}
758 	return rule;
759 }
760 
761 void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec)
762 {
763 	struct efx_arfs_rule *rule;
764 	struct hlist_head *head;
765 	struct hlist_node *node;
766 
767 	head = efx_rps_hash_bucket(efx, spec);
768 	if (WARN_ON(!head))
769 		return;
770 	hlist_for_each(node, head) {
771 		rule = container_of(node, struct efx_arfs_rule, node);
772 		if (efx_filter_spec_equal(spec, &rule->spec)) {
773 			/* Someone already reused the entry.  We know that if
774 			 * this check doesn't fire (i.e. filter_id == REMOVING)
775 			 * then the REMOVING mark was put there by our caller,
776 			 * because caller is holding a lock on filter table and
777 			 * only holders of that lock set REMOVING.
778 			 */
779 			if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING)
780 				return;
781 			hlist_del(node);
782 			kfree(rule);
783 			return;
784 		}
785 	}
786 	/* We didn't find it. */
787 	WARN_ON(1);
788 }
789 #endif
790 
791 int efx_probe_filters(struct efx_nic *efx)
792 {
793 	int rc;
794 
795 	mutex_lock(&efx->mac_lock);
796 	rc = efx->type->filter_table_probe(efx);
797 	if (rc)
798 		goto out_unlock;
799 
800 #ifdef CONFIG_RFS_ACCEL
801 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
802 		struct efx_channel *channel;
803 		int i, success = 1;
804 
805 		efx_for_each_channel(channel, efx) {
806 			channel->rps_flow_id =
807 				kcalloc(efx->type->max_rx_ip_filters,
808 					sizeof(*channel->rps_flow_id),
809 					GFP_KERNEL);
810 			if (!channel->rps_flow_id)
811 				success = 0;
812 			else
813 				for (i = 0;
814 				     i < efx->type->max_rx_ip_filters;
815 				     ++i)
816 					channel->rps_flow_id[i] =
817 						RPS_FLOW_ID_INVALID;
818 			channel->rfs_expire_index = 0;
819 			channel->rfs_filter_count = 0;
820 		}
821 
822 		if (!success) {
823 			efx_for_each_channel(channel, efx)
824 				kfree(channel->rps_flow_id);
825 			efx->type->filter_table_remove(efx);
826 			rc = -ENOMEM;
827 			goto out_unlock;
828 		}
829 	}
830 #endif
831 out_unlock:
832 	mutex_unlock(&efx->mac_lock);
833 	return rc;
834 }
835 
836 void efx_remove_filters(struct efx_nic *efx)
837 {
838 #ifdef CONFIG_RFS_ACCEL
839 	struct efx_channel *channel;
840 
841 	efx_for_each_channel(channel, efx) {
842 		cancel_delayed_work_sync(&channel->filter_work);
843 		kfree(channel->rps_flow_id);
844 		channel->rps_flow_id = NULL;
845 	}
846 #endif
847 	efx->type->filter_table_remove(efx);
848 }
849 
850 #ifdef CONFIG_RFS_ACCEL
851 
852 static void efx_filter_rfs_work(struct work_struct *data)
853 {
854 	struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
855 							      work);
856 	struct efx_nic *efx = efx_netdev_priv(req->net_dev);
857 	struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
858 	int slot_idx = req - efx->rps_slot;
859 	struct efx_arfs_rule *rule;
860 	u16 arfs_id = 0;
861 	int rc;
862 
863 	rc = efx->type->filter_insert(efx, &req->spec, true);
864 	if (rc >= 0)
865 		/* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */
866 		rc %= efx->type->max_rx_ip_filters;
867 	if (efx->rps_hash_table) {
868 		spin_lock_bh(&efx->rps_hash_lock);
869 		rule = efx_rps_hash_find(efx, &req->spec);
870 		/* The rule might have already gone, if someone else's request
871 		 * for the same spec was already worked and then expired before
872 		 * we got around to our work.  In that case we have nothing
873 		 * tying us to an arfs_id, meaning that as soon as the filter
874 		 * is considered for expiry it will be removed.
875 		 */
876 		if (rule) {
877 			if (rc < 0)
878 				rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
879 			else
880 				rule->filter_id = rc;
881 			arfs_id = rule->arfs_id;
882 		}
883 		spin_unlock_bh(&efx->rps_hash_lock);
884 	}
885 	if (rc >= 0) {
886 		/* Remember this so we can check whether to expire the filter
887 		 * later.
888 		 */
889 		mutex_lock(&efx->rps_mutex);
890 		if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID)
891 			channel->rfs_filter_count++;
892 		channel->rps_flow_id[rc] = req->flow_id;
893 		mutex_unlock(&efx->rps_mutex);
894 
895 		if (req->spec.ether_type == htons(ETH_P_IP))
896 			netif_info(efx, rx_status, efx->net_dev,
897 				   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
898 				   (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
899 				   req->spec.rem_host, ntohs(req->spec.rem_port),
900 				   req->spec.loc_host, ntohs(req->spec.loc_port),
901 				   req->rxq_index, req->flow_id, rc, arfs_id);
902 		else
903 			netif_info(efx, rx_status, efx->net_dev,
904 				   "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
905 				   (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
906 				   req->spec.rem_host, ntohs(req->spec.rem_port),
907 				   req->spec.loc_host, ntohs(req->spec.loc_port),
908 				   req->rxq_index, req->flow_id, rc, arfs_id);
909 		channel->n_rfs_succeeded++;
910 	} else {
911 		if (req->spec.ether_type == htons(ETH_P_IP))
912 			netif_dbg(efx, rx_status, efx->net_dev,
913 				  "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n",
914 				  (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
915 				  req->spec.rem_host, ntohs(req->spec.rem_port),
916 				  req->spec.loc_host, ntohs(req->spec.loc_port),
917 				  req->rxq_index, req->flow_id, rc, arfs_id);
918 		else
919 			netif_dbg(efx, rx_status, efx->net_dev,
920 				  "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n",
921 				  (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
922 				  req->spec.rem_host, ntohs(req->spec.rem_port),
923 				  req->spec.loc_host, ntohs(req->spec.loc_port),
924 				  req->rxq_index, req->flow_id, rc, arfs_id);
925 		channel->n_rfs_failed++;
926 		/* We're overloading the NIC's filter tables, so let's do a
927 		 * chunk of extra expiry work.
928 		 */
929 		__efx_filter_rfs_expire(channel, min(channel->rfs_filter_count,
930 						     100u));
931 	}
932 
933 	/* Release references */
934 	clear_bit(slot_idx, &efx->rps_slot_map);
935 	dev_put(req->net_dev);
936 }
937 
938 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
939 		   u16 rxq_index, u32 flow_id)
940 {
941 	struct efx_nic *efx = efx_netdev_priv(net_dev);
942 	struct efx_async_filter_insertion *req;
943 	struct efx_arfs_rule *rule;
944 	struct flow_keys fk;
945 	int slot_idx;
946 	bool new;
947 	int rc;
948 
949 	/* find a free slot */
950 	for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
951 		if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
952 			break;
953 	if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
954 		return -EBUSY;
955 
956 	if (flow_id == RPS_FLOW_ID_INVALID) {
957 		rc = -EINVAL;
958 		goto out_clear;
959 	}
960 
961 	if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
962 		rc = -EPROTONOSUPPORT;
963 		goto out_clear;
964 	}
965 
966 	if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
967 		rc = -EPROTONOSUPPORT;
968 		goto out_clear;
969 	}
970 	if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
971 		rc = -EPROTONOSUPPORT;
972 		goto out_clear;
973 	}
974 
975 	req = efx->rps_slot + slot_idx;
976 	efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
977 			   efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
978 			   rxq_index);
979 	req->spec.match_flags =
980 		EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
981 		EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
982 		EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
983 	req->spec.ether_type = fk.basic.n_proto;
984 	req->spec.ip_proto = fk.basic.ip_proto;
985 
986 	if (fk.basic.n_proto == htons(ETH_P_IP)) {
987 		req->spec.rem_host[0] = fk.addrs.v4addrs.src;
988 		req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
989 	} else {
990 		memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
991 		       sizeof(struct in6_addr));
992 		memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
993 		       sizeof(struct in6_addr));
994 	}
995 
996 	req->spec.rem_port = fk.ports.src;
997 	req->spec.loc_port = fk.ports.dst;
998 
999 	if (efx->rps_hash_table) {
1000 		/* Add it to ARFS hash table */
1001 		spin_lock(&efx->rps_hash_lock);
1002 		rule = efx_rps_hash_add(efx, &req->spec, &new);
1003 		if (!rule) {
1004 			rc = -ENOMEM;
1005 			goto out_unlock;
1006 		}
1007 		if (new)
1008 			rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
1009 		rc = rule->arfs_id;
1010 		/* Skip if existing or pending filter already does the right thing */
1011 		if (!new && rule->rxq_index == rxq_index &&
1012 		    rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
1013 			goto out_unlock;
1014 		rule->rxq_index = rxq_index;
1015 		rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
1016 		spin_unlock(&efx->rps_hash_lock);
1017 	} else {
1018 		/* Without an ARFS hash table, we just use arfs_id 0 for all
1019 		 * filters.  This means if multiple flows hash to the same
1020 		 * flow_id, all but the most recently touched will be eligible
1021 		 * for expiry.
1022 		 */
1023 		rc = 0;
1024 	}
1025 
1026 	/* Queue the request */
1027 	dev_hold(req->net_dev = net_dev);
1028 	INIT_WORK(&req->work, efx_filter_rfs_work);
1029 	req->rxq_index = rxq_index;
1030 	req->flow_id = flow_id;
1031 	schedule_work(&req->work);
1032 	return rc;
1033 out_unlock:
1034 	spin_unlock(&efx->rps_hash_lock);
1035 out_clear:
1036 	clear_bit(slot_idx, &efx->rps_slot_map);
1037 	return rc;
1038 }
1039 
1040 bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota)
1041 {
1042 	bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
1043 	struct efx_nic *efx = channel->efx;
1044 	unsigned int index, size, start;
1045 	u32 flow_id;
1046 
1047 	if (!mutex_trylock(&efx->rps_mutex))
1048 		return false;
1049 	expire_one = efx->type->filter_rfs_expire_one;
1050 	index = channel->rfs_expire_index;
1051 	start = index;
1052 	size = efx->type->max_rx_ip_filters;
1053 	while (quota) {
1054 		flow_id = channel->rps_flow_id[index];
1055 
1056 		if (flow_id != RPS_FLOW_ID_INVALID) {
1057 			quota--;
1058 			if (expire_one(efx, flow_id, index)) {
1059 				netif_info(efx, rx_status, efx->net_dev,
1060 					   "expired filter %d [channel %u flow %u]\n",
1061 					   index, channel->channel, flow_id);
1062 				channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1063 				channel->rfs_filter_count--;
1064 			}
1065 		}
1066 		if (++index == size)
1067 			index = 0;
1068 		/* If we were called with a quota that exceeds the total number
1069 		 * of filters in the table (which shouldn't happen, but could
1070 		 * if two callers race), ensure that we don't loop forever -
1071 		 * stop when we've examined every row of the table.
1072 		 */
1073 		if (index == start)
1074 			break;
1075 	}
1076 
1077 	channel->rfs_expire_index = index;
1078 	mutex_unlock(&efx->rps_mutex);
1079 	return true;
1080 }
1081 
1082 #endif /* CONFIG_RFS_ACCEL */
1083