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 /* Number of RX buffers to recycle pages for.  When creating the RX page recycle
27  * ring, this number is divided by the number of buffers per page to calculate
28  * the number of pages to store in the RX page recycle ring.
29  */
30 #define EFX_RECYCLE_RING_SIZE_IOMMU 4096
31 #define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
32 
33 /* RX maximum head room required.
34  *
35  * This must be at least 1 to prevent overflow, plus one packet-worth
36  * to allow pipelined receives.
37  */
38 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
39 
40 /* Check the RX page recycle ring for a page that can be reused. */
41 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue)
42 {
43 	struct efx_nic *efx = rx_queue->efx;
44 	struct efx_rx_page_state *state;
45 	unsigned int index;
46 	struct page *page;
47 
48 	if (unlikely(!rx_queue->page_ring))
49 		return NULL;
50 	index = rx_queue->page_remove & rx_queue->page_ptr_mask;
51 	page = rx_queue->page_ring[index];
52 	if (page == NULL)
53 		return NULL;
54 
55 	rx_queue->page_ring[index] = NULL;
56 	/* page_remove cannot exceed page_add. */
57 	if (rx_queue->page_remove != rx_queue->page_add)
58 		++rx_queue->page_remove;
59 
60 	/* If page_count is 1 then we hold the only reference to this page. */
61 	if (page_count(page) == 1) {
62 		++rx_queue->page_recycle_count;
63 		return page;
64 	} else {
65 		state = page_address(page);
66 		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
67 			       PAGE_SIZE << efx->rx_buffer_order,
68 			       DMA_FROM_DEVICE);
69 		put_page(page);
70 		++rx_queue->page_recycle_failed;
71 	}
72 
73 	return NULL;
74 }
75 
76 /* Attempt to recycle the page if there is an RX recycle ring; the page can
77  * only be added if this is the final RX buffer, to prevent pages being used in
78  * the descriptor ring and appearing in the recycle ring simultaneously.
79  */
80 static void efx_recycle_rx_page(struct efx_channel *channel,
81 				struct efx_rx_buffer *rx_buf)
82 {
83 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
84 	struct efx_nic *efx = rx_queue->efx;
85 	struct page *page = rx_buf->page;
86 	unsigned int index;
87 
88 	/* Only recycle the page after processing the final buffer. */
89 	if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE))
90 		return;
91 
92 	index = rx_queue->page_add & rx_queue->page_ptr_mask;
93 	if (rx_queue->page_ring[index] == NULL) {
94 		unsigned int read_index = rx_queue->page_remove &
95 			rx_queue->page_ptr_mask;
96 
97 		/* The next slot in the recycle ring is available, but
98 		 * increment page_remove if the read pointer currently
99 		 * points here.
100 		 */
101 		if (read_index == index)
102 			++rx_queue->page_remove;
103 		rx_queue->page_ring[index] = page;
104 		++rx_queue->page_add;
105 		return;
106 	}
107 	++rx_queue->page_recycle_full;
108 	efx_unmap_rx_buffer(efx, rx_buf);
109 	put_page(rx_buf->page);
110 }
111 
112 /* Recycle the pages that are used by buffers that have just been received. */
113 void efx_recycle_rx_pages(struct efx_channel *channel,
114 			  struct efx_rx_buffer *rx_buf,
115 			  unsigned int n_frags)
116 {
117 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
118 
119 	if (unlikely(!rx_queue->page_ring))
120 		return;
121 
122 	do {
123 		efx_recycle_rx_page(channel, rx_buf);
124 		rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
125 	} while (--n_frags);
126 }
127 
128 void efx_discard_rx_packet(struct efx_channel *channel,
129 			   struct efx_rx_buffer *rx_buf,
130 			   unsigned int n_frags)
131 {
132 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
133 
134 	efx_recycle_rx_pages(channel, rx_buf, n_frags);
135 
136 	efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
137 }
138 
139 static void efx_init_rx_recycle_ring(struct efx_rx_queue *rx_queue)
140 {
141 	unsigned int bufs_in_recycle_ring, page_ring_size;
142 	struct efx_nic *efx = rx_queue->efx;
143 
144 	/* Set the RX recycle ring size */
145 #ifdef CONFIG_PPC64
146 	bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
147 #else
148 	if (iommu_present(&pci_bus_type))
149 		bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
150 	else
151 		bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
152 #endif /* CONFIG_PPC64 */
153 
154 	page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
155 					    efx->rx_bufs_per_page);
156 	rx_queue->page_ring = kcalloc(page_ring_size,
157 				      sizeof(*rx_queue->page_ring), GFP_KERNEL);
158 	if (!rx_queue->page_ring)
159 		rx_queue->page_ptr_mask = 0;
160 	else
161 		rx_queue->page_ptr_mask = page_ring_size - 1;
162 }
163 
164 static void efx_fini_rx_recycle_ring(struct efx_rx_queue *rx_queue)
165 {
166 	struct efx_nic *efx = rx_queue->efx;
167 	int i;
168 
169 	/* Unmap and release the pages in the recycle ring. Remove the ring. */
170 	for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
171 		struct page *page = rx_queue->page_ring[i];
172 		struct efx_rx_page_state *state;
173 
174 		if (page == NULL)
175 			continue;
176 
177 		state = page_address(page);
178 		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
179 			       PAGE_SIZE << efx->rx_buffer_order,
180 			       DMA_FROM_DEVICE);
181 		put_page(page);
182 	}
183 	kfree(rx_queue->page_ring);
184 	rx_queue->page_ring = NULL;
185 }
186 
187 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
188 			       struct efx_rx_buffer *rx_buf)
189 {
190 	/* Release the page reference we hold for the buffer. */
191 	if (rx_buf->page)
192 		put_page(rx_buf->page);
193 
194 	/* If this is the last buffer in a page, unmap and free it. */
195 	if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) {
196 		efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
197 		efx_free_rx_buffers(rx_queue, rx_buf, 1);
198 	}
199 	rx_buf->page = NULL;
200 }
201 
202 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
203 {
204 	struct efx_nic *efx = rx_queue->efx;
205 	unsigned int entries;
206 	int rc;
207 
208 	/* Create the smallest power-of-two aligned ring */
209 	entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
210 	EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
211 	rx_queue->ptr_mask = entries - 1;
212 
213 	netif_dbg(efx, probe, efx->net_dev,
214 		  "creating RX queue %d size %#x mask %#x\n",
215 		  efx_rx_queue_index(rx_queue), efx->rxq_entries,
216 		  rx_queue->ptr_mask);
217 
218 	/* Allocate RX buffers */
219 	rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
220 				   GFP_KERNEL);
221 	if (!rx_queue->buffer)
222 		return -ENOMEM;
223 
224 	rc = efx_nic_probe_rx(rx_queue);
225 	if (rc) {
226 		kfree(rx_queue->buffer);
227 		rx_queue->buffer = NULL;
228 	}
229 
230 	return rc;
231 }
232 
233 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
234 {
235 	unsigned int max_fill, trigger, max_trigger;
236 	struct efx_nic *efx = rx_queue->efx;
237 	int rc = 0;
238 
239 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
240 		  "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
241 
242 	/* Initialise ptr fields */
243 	rx_queue->added_count = 0;
244 	rx_queue->notified_count = 0;
245 	rx_queue->removed_count = 0;
246 	rx_queue->min_fill = -1U;
247 	efx_init_rx_recycle_ring(rx_queue);
248 
249 	rx_queue->page_remove = 0;
250 	rx_queue->page_add = rx_queue->page_ptr_mask + 1;
251 	rx_queue->page_recycle_count = 0;
252 	rx_queue->page_recycle_failed = 0;
253 	rx_queue->page_recycle_full = 0;
254 
255 	/* Initialise limit fields */
256 	max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
257 	max_trigger =
258 		max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
259 	if (rx_refill_threshold != 0) {
260 		trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
261 		if (trigger > max_trigger)
262 			trigger = max_trigger;
263 	} else {
264 		trigger = max_trigger;
265 	}
266 
267 	rx_queue->max_fill = max_fill;
268 	rx_queue->fast_fill_trigger = trigger;
269 	rx_queue->refill_enabled = true;
270 
271 	/* Initialise XDP queue information */
272 	rc = xdp_rxq_info_reg(&rx_queue->xdp_rxq_info, efx->net_dev,
273 			      rx_queue->core_index, 0);
274 
275 	if (rc) {
276 		netif_err(efx, rx_err, efx->net_dev,
277 			  "Failure to initialise XDP queue information rc=%d\n",
278 			  rc);
279 		efx->xdp_rxq_info_failed = true;
280 	} else {
281 		rx_queue->xdp_rxq_info_valid = true;
282 	}
283 
284 	/* Set up RX descriptor ring */
285 	efx_nic_init_rx(rx_queue);
286 }
287 
288 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
289 {
290 	struct efx_rx_buffer *rx_buf;
291 	int i;
292 
293 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
294 		  "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
295 
296 	del_timer_sync(&rx_queue->slow_fill);
297 
298 	/* Release RX buffers from the current read ptr to the write ptr */
299 	if (rx_queue->buffer) {
300 		for (i = rx_queue->removed_count; i < rx_queue->added_count;
301 		     i++) {
302 			unsigned int index = i & rx_queue->ptr_mask;
303 
304 			rx_buf = efx_rx_buffer(rx_queue, index);
305 			efx_fini_rx_buffer(rx_queue, rx_buf);
306 		}
307 	}
308 
309 	efx_fini_rx_recycle_ring(rx_queue);
310 
311 	if (rx_queue->xdp_rxq_info_valid)
312 		xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info);
313 
314 	rx_queue->xdp_rxq_info_valid = false;
315 }
316 
317 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
318 {
319 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
320 		  "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
321 
322 	efx_nic_remove_rx(rx_queue);
323 
324 	kfree(rx_queue->buffer);
325 	rx_queue->buffer = NULL;
326 }
327 
328 /* Unmap a DMA-mapped page.  This function is only called for the final RX
329  * buffer in a page.
330  */
331 void efx_unmap_rx_buffer(struct efx_nic *efx,
332 			 struct efx_rx_buffer *rx_buf)
333 {
334 	struct page *page = rx_buf->page;
335 
336 	if (page) {
337 		struct efx_rx_page_state *state = page_address(page);
338 
339 		dma_unmap_page(&efx->pci_dev->dev,
340 			       state->dma_addr,
341 			       PAGE_SIZE << efx->rx_buffer_order,
342 			       DMA_FROM_DEVICE);
343 	}
344 }
345 
346 void efx_free_rx_buffers(struct efx_rx_queue *rx_queue,
347 			 struct efx_rx_buffer *rx_buf,
348 			 unsigned int num_bufs)
349 {
350 	do {
351 		if (rx_buf->page) {
352 			put_page(rx_buf->page);
353 			rx_buf->page = NULL;
354 		}
355 		rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
356 	} while (--num_bufs);
357 }
358 
359 void efx_rx_slow_fill(struct timer_list *t)
360 {
361 	struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill);
362 
363 	/* Post an event to cause NAPI to run and refill the queue */
364 	efx_nic_generate_fill_event(rx_queue);
365 	++rx_queue->slow_fill_count;
366 }
367 
368 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue)
369 {
370 	mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(10));
371 }
372 
373 /* efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
374  *
375  * @rx_queue:		Efx RX queue
376  *
377  * This allocates a batch of pages, maps them for DMA, and populates
378  * struct efx_rx_buffers for each one. Return a negative error code or
379  * 0 on success. If a single page can be used for multiple buffers,
380  * then the page will either be inserted fully, or not at all.
381  */
382 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
383 {
384 	unsigned int page_offset, index, count;
385 	struct efx_nic *efx = rx_queue->efx;
386 	struct efx_rx_page_state *state;
387 	struct efx_rx_buffer *rx_buf;
388 	dma_addr_t dma_addr;
389 	struct page *page;
390 
391 	count = 0;
392 	do {
393 		page = efx_reuse_page(rx_queue);
394 		if (page == NULL) {
395 			page = alloc_pages(__GFP_COMP |
396 					   (atomic ? GFP_ATOMIC : GFP_KERNEL),
397 					   efx->rx_buffer_order);
398 			if (unlikely(page == NULL))
399 				return -ENOMEM;
400 			dma_addr =
401 				dma_map_page(&efx->pci_dev->dev, page, 0,
402 					     PAGE_SIZE << efx->rx_buffer_order,
403 					     DMA_FROM_DEVICE);
404 			if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
405 						       dma_addr))) {
406 				__free_pages(page, efx->rx_buffer_order);
407 				return -EIO;
408 			}
409 			state = page_address(page);
410 			state->dma_addr = dma_addr;
411 		} else {
412 			state = page_address(page);
413 			dma_addr = state->dma_addr;
414 		}
415 
416 		dma_addr += sizeof(struct efx_rx_page_state);
417 		page_offset = sizeof(struct efx_rx_page_state);
418 
419 		do {
420 			index = rx_queue->added_count & rx_queue->ptr_mask;
421 			rx_buf = efx_rx_buffer(rx_queue, index);
422 			rx_buf->dma_addr = dma_addr + efx->rx_ip_align +
423 					   EFX_XDP_HEADROOM;
424 			rx_buf->page = page;
425 			rx_buf->page_offset = page_offset + efx->rx_ip_align +
426 					      EFX_XDP_HEADROOM;
427 			rx_buf->len = efx->rx_dma_len;
428 			rx_buf->flags = 0;
429 			++rx_queue->added_count;
430 			get_page(page);
431 			dma_addr += efx->rx_page_buf_step;
432 			page_offset += efx->rx_page_buf_step;
433 		} while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
434 
435 		rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE;
436 	} while (++count < efx->rx_pages_per_batch);
437 
438 	return 0;
439 }
440 
441 void efx_rx_config_page_split(struct efx_nic *efx)
442 {
443 	efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align +
444 				      EFX_XDP_HEADROOM + EFX_XDP_TAILROOM,
445 				      EFX_RX_BUF_ALIGNMENT);
446 	efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
447 		((PAGE_SIZE - sizeof(struct efx_rx_page_state)) /
448 		efx->rx_page_buf_step);
449 	efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
450 		efx->rx_bufs_per_page;
451 	efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH,
452 					       efx->rx_bufs_per_page);
453 }
454 
455 /* efx_fast_push_rx_descriptors - push new RX descriptors quickly
456  * @rx_queue:		RX descriptor queue
457  *
458  * This will aim to fill the RX descriptor queue up to
459  * @rx_queue->@max_fill. If there is insufficient atomic
460  * memory to do so, a slow fill will be scheduled.
461  *
462  * The caller must provide serialisation (none is used here). In practise,
463  * this means this function must run from the NAPI handler, or be called
464  * when NAPI is disabled.
465  */
466 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
467 {
468 	struct efx_nic *efx = rx_queue->efx;
469 	unsigned int fill_level, batch_size;
470 	int space, rc = 0;
471 
472 	if (!rx_queue->refill_enabled)
473 		return;
474 
475 	/* Calculate current fill level, and exit if we don't need to fill */
476 	fill_level = (rx_queue->added_count - rx_queue->removed_count);
477 	EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries);
478 	if (fill_level >= rx_queue->fast_fill_trigger)
479 		goto out;
480 
481 	/* Record minimum fill level */
482 	if (unlikely(fill_level < rx_queue->min_fill)) {
483 		if (fill_level)
484 			rx_queue->min_fill = fill_level;
485 	}
486 
487 	batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
488 	space = rx_queue->max_fill - fill_level;
489 	EFX_WARN_ON_ONCE_PARANOID(space < batch_size);
490 
491 	netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
492 		   "RX queue %d fast-filling descriptor ring from"
493 		   " level %d to level %d\n",
494 		   efx_rx_queue_index(rx_queue), fill_level,
495 		   rx_queue->max_fill);
496 
497 	do {
498 		rc = efx_init_rx_buffers(rx_queue, atomic);
499 		if (unlikely(rc)) {
500 			/* Ensure that we don't leave the rx queue empty */
501 			efx_schedule_slow_fill(rx_queue);
502 			goto out;
503 		}
504 	} while ((space -= batch_size) >= batch_size);
505 
506 	netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
507 		   "RX queue %d fast-filled descriptor ring "
508 		   "to level %d\n", efx_rx_queue_index(rx_queue),
509 		   rx_queue->added_count - rx_queue->removed_count);
510 
511  out:
512 	if (rx_queue->notified_count != rx_queue->added_count)
513 		efx_nic_notify_rx_desc(rx_queue);
514 }
515 
516 /* Pass a received packet up through GRO.  GRO can handle pages
517  * regardless of checksum state and skbs with a good checksum.
518  */
519 void
520 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
521 		  unsigned int n_frags, u8 *eh, __wsum csum)
522 {
523 	struct napi_struct *napi = &channel->napi_str;
524 	struct efx_nic *efx = channel->efx;
525 	struct sk_buff *skb;
526 
527 	skb = napi_get_frags(napi);
528 	if (unlikely(!skb)) {
529 		struct efx_rx_queue *rx_queue;
530 
531 		rx_queue = efx_channel_get_rx_queue(channel);
532 		efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
533 		return;
534 	}
535 
536 	if (efx->net_dev->features & NETIF_F_RXHASH &&
537 	    efx_rx_buf_hash_valid(efx, eh))
538 		skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
539 			     PKT_HASH_TYPE_L3);
540 	if (csum) {
541 		skb->csum = csum;
542 		skb->ip_summed = CHECKSUM_COMPLETE;
543 	} else {
544 		skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
545 				  CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
546 	}
547 	skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL);
548 
549 	for (;;) {
550 		skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
551 				   rx_buf->page, rx_buf->page_offset,
552 				   rx_buf->len);
553 		rx_buf->page = NULL;
554 		skb->len += rx_buf->len;
555 		if (skb_shinfo(skb)->nr_frags == n_frags)
556 			break;
557 
558 		rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
559 	}
560 
561 	skb->data_len = skb->len;
562 	skb->truesize += n_frags * efx->rx_buffer_truesize;
563 
564 	skb_record_rx_queue(skb, channel->rx_queue.core_index);
565 
566 	napi_gro_frags(napi);
567 }
568 
569 /* RSS contexts.  We're using linked lists and crappy O(n) algorithms, because
570  * (a) this is an infrequent control-plane operation and (b) n is small (max 64)
571  */
572 struct efx_rss_context *efx_alloc_rss_context_entry(struct efx_nic *efx)
573 {
574 	struct list_head *head = &efx->rss_context.list;
575 	struct efx_rss_context *ctx, *new;
576 	u32 id = 1; /* Don't use zero, that refers to the master RSS context */
577 
578 	WARN_ON(!mutex_is_locked(&efx->rss_lock));
579 
580 	/* Search for first gap in the numbering */
581 	list_for_each_entry(ctx, head, list) {
582 		if (ctx->user_id != id)
583 			break;
584 		id++;
585 		/* Check for wrap.  If this happens, we have nearly 2^32
586 		 * allocated RSS contexts, which seems unlikely.
587 		 */
588 		if (WARN_ON_ONCE(!id))
589 			return NULL;
590 	}
591 
592 	/* Create the new entry */
593 	new = kmalloc(sizeof(*new), GFP_KERNEL);
594 	if (!new)
595 		return NULL;
596 	new->context_id = EFX_MCDI_RSS_CONTEXT_INVALID;
597 	new->rx_hash_udp_4tuple = false;
598 
599 	/* Insert the new entry into the gap */
600 	new->user_id = id;
601 	list_add_tail(&new->list, &ctx->list);
602 	return new;
603 }
604 
605 struct efx_rss_context *efx_find_rss_context_entry(struct efx_nic *efx, u32 id)
606 {
607 	struct list_head *head = &efx->rss_context.list;
608 	struct efx_rss_context *ctx;
609 
610 	WARN_ON(!mutex_is_locked(&efx->rss_lock));
611 
612 	list_for_each_entry(ctx, head, list)
613 		if (ctx->user_id == id)
614 			return ctx;
615 	return NULL;
616 }
617 
618 void efx_free_rss_context_entry(struct efx_rss_context *ctx)
619 {
620 	list_del(&ctx->list);
621 	kfree(ctx);
622 }
623 
624 void efx_set_default_rx_indir_table(struct efx_nic *efx,
625 				    struct efx_rss_context *ctx)
626 {
627 	size_t i;
628 
629 	for (i = 0; i < ARRAY_SIZE(ctx->rx_indir_table); i++)
630 		ctx->rx_indir_table[i] =
631 			ethtool_rxfh_indir_default(i, efx->rss_spread);
632 }
633 
634 /**
635  * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
636  * @spec: Specification to test
637  *
638  * Return: %true if the specification is a non-drop RX filter that
639  * matches a local MAC address I/G bit value of 1 or matches a local
640  * IPv4 or IPv6 address value in the respective multicast address
641  * range.  Otherwise %false.
642  */
643 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
644 {
645 	if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
646 	    spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
647 		return false;
648 
649 	if (spec->match_flags &
650 	    (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
651 	    is_multicast_ether_addr(spec->loc_mac))
652 		return true;
653 
654 	if ((spec->match_flags &
655 	     (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
656 	    (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
657 		if (spec->ether_type == htons(ETH_P_IP) &&
658 		    ipv4_is_multicast(spec->loc_host[0]))
659 			return true;
660 		if (spec->ether_type == htons(ETH_P_IPV6) &&
661 		    ((const u8 *)spec->loc_host)[0] == 0xff)
662 			return true;
663 	}
664 
665 	return false;
666 }
667 
668 bool efx_filter_spec_equal(const struct efx_filter_spec *left,
669 			   const struct efx_filter_spec *right)
670 {
671 	if ((left->match_flags ^ right->match_flags) |
672 	    ((left->flags ^ right->flags) &
673 	     (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
674 		return false;
675 
676 	return memcmp(&left->outer_vid, &right->outer_vid,
677 		      sizeof(struct efx_filter_spec) -
678 		      offsetof(struct efx_filter_spec, outer_vid)) == 0;
679 }
680 
681 u32 efx_filter_spec_hash(const struct efx_filter_spec *spec)
682 {
683 	BUILD_BUG_ON(offsetof(struct efx_filter_spec, outer_vid) & 3);
684 	return jhash2((const u32 *)&spec->outer_vid,
685 		      (sizeof(struct efx_filter_spec) -
686 		       offsetof(struct efx_filter_spec, outer_vid)) / 4,
687 		      0);
688 }
689 
690 #ifdef CONFIG_RFS_ACCEL
691 bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx,
692 			bool *force)
693 {
694 	if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) {
695 		/* ARFS is currently updating this entry, leave it */
696 		return false;
697 	}
698 	if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) {
699 		/* ARFS tried and failed to update this, so it's probably out
700 		 * of date.  Remove the filter and the ARFS rule entry.
701 		 */
702 		rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
703 		*force = true;
704 		return true;
705 	} else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */
706 		/* ARFS has moved on, so old filter is not needed.  Since we did
707 		 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will
708 		 * not be removed by efx_rps_hash_del() subsequently.
709 		 */
710 		*force = true;
711 		return true;
712 	}
713 	/* Remove it iff ARFS wants to. */
714 	return true;
715 }
716 
717 static
718 struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx,
719 				       const struct efx_filter_spec *spec)
720 {
721 	u32 hash = efx_filter_spec_hash(spec);
722 
723 	lockdep_assert_held(&efx->rps_hash_lock);
724 	if (!efx->rps_hash_table)
725 		return NULL;
726 	return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE];
727 }
728 
729 struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx,
730 					const struct efx_filter_spec *spec)
731 {
732 	struct efx_arfs_rule *rule;
733 	struct hlist_head *head;
734 	struct hlist_node *node;
735 
736 	head = efx_rps_hash_bucket(efx, spec);
737 	if (!head)
738 		return NULL;
739 	hlist_for_each(node, head) {
740 		rule = container_of(node, struct efx_arfs_rule, node);
741 		if (efx_filter_spec_equal(spec, &rule->spec))
742 			return rule;
743 	}
744 	return NULL;
745 }
746 
747 struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
748 				       const struct efx_filter_spec *spec,
749 				       bool *new)
750 {
751 	struct efx_arfs_rule *rule;
752 	struct hlist_head *head;
753 	struct hlist_node *node;
754 
755 	head = efx_rps_hash_bucket(efx, spec);
756 	if (!head)
757 		return NULL;
758 	hlist_for_each(node, head) {
759 		rule = container_of(node, struct efx_arfs_rule, node);
760 		if (efx_filter_spec_equal(spec, &rule->spec)) {
761 			*new = false;
762 			return rule;
763 		}
764 	}
765 	rule = kmalloc(sizeof(*rule), GFP_ATOMIC);
766 	*new = true;
767 	if (rule) {
768 		memcpy(&rule->spec, spec, sizeof(rule->spec));
769 		hlist_add_head(&rule->node, head);
770 	}
771 	return rule;
772 }
773 
774 void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec)
775 {
776 	struct efx_arfs_rule *rule;
777 	struct hlist_head *head;
778 	struct hlist_node *node;
779 
780 	head = efx_rps_hash_bucket(efx, spec);
781 	if (WARN_ON(!head))
782 		return;
783 	hlist_for_each(node, head) {
784 		rule = container_of(node, struct efx_arfs_rule, node);
785 		if (efx_filter_spec_equal(spec, &rule->spec)) {
786 			/* Someone already reused the entry.  We know that if
787 			 * this check doesn't fire (i.e. filter_id == REMOVING)
788 			 * then the REMOVING mark was put there by our caller,
789 			 * because caller is holding a lock on filter table and
790 			 * only holders of that lock set REMOVING.
791 			 */
792 			if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING)
793 				return;
794 			hlist_del(node);
795 			kfree(rule);
796 			return;
797 		}
798 	}
799 	/* We didn't find it. */
800 	WARN_ON(1);
801 }
802 #endif
803 
804 int efx_probe_filters(struct efx_nic *efx)
805 {
806 	int rc;
807 
808 	mutex_lock(&efx->mac_lock);
809 	down_write(&efx->filter_sem);
810 	rc = efx->type->filter_table_probe(efx);
811 	if (rc)
812 		goto out_unlock;
813 
814 #ifdef CONFIG_RFS_ACCEL
815 	if (efx->type->offload_features & NETIF_F_NTUPLE) {
816 		struct efx_channel *channel;
817 		int i, success = 1;
818 
819 		efx_for_each_channel(channel, efx) {
820 			channel->rps_flow_id =
821 				kcalloc(efx->type->max_rx_ip_filters,
822 					sizeof(*channel->rps_flow_id),
823 					GFP_KERNEL);
824 			if (!channel->rps_flow_id)
825 				success = 0;
826 			else
827 				for (i = 0;
828 				     i < efx->type->max_rx_ip_filters;
829 				     ++i)
830 					channel->rps_flow_id[i] =
831 						RPS_FLOW_ID_INVALID;
832 			channel->rfs_expire_index = 0;
833 			channel->rfs_filter_count = 0;
834 		}
835 
836 		if (!success) {
837 			efx_for_each_channel(channel, efx)
838 				kfree(channel->rps_flow_id);
839 			efx->type->filter_table_remove(efx);
840 			rc = -ENOMEM;
841 			goto out_unlock;
842 		}
843 	}
844 #endif
845 out_unlock:
846 	up_write(&efx->filter_sem);
847 	mutex_unlock(&efx->mac_lock);
848 	return rc;
849 }
850 
851 void efx_remove_filters(struct efx_nic *efx)
852 {
853 #ifdef CONFIG_RFS_ACCEL
854 	struct efx_channel *channel;
855 
856 	efx_for_each_channel(channel, efx) {
857 		cancel_delayed_work_sync(&channel->filter_work);
858 		kfree(channel->rps_flow_id);
859 		channel->rps_flow_id = NULL;
860 	}
861 #endif
862 	down_write(&efx->filter_sem);
863 	efx->type->filter_table_remove(efx);
864 	up_write(&efx->filter_sem);
865 }
866 
867 #ifdef CONFIG_RFS_ACCEL
868 
869 static void efx_filter_rfs_work(struct work_struct *data)
870 {
871 	struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
872 							      work);
873 	struct efx_nic *efx = netdev_priv(req->net_dev);
874 	struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
875 	int slot_idx = req - efx->rps_slot;
876 	struct efx_arfs_rule *rule;
877 	u16 arfs_id = 0;
878 	int rc;
879 
880 	rc = efx->type->filter_insert(efx, &req->spec, true);
881 	if (rc >= 0)
882 		/* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */
883 		rc %= efx->type->max_rx_ip_filters;
884 	if (efx->rps_hash_table) {
885 		spin_lock_bh(&efx->rps_hash_lock);
886 		rule = efx_rps_hash_find(efx, &req->spec);
887 		/* The rule might have already gone, if someone else's request
888 		 * for the same spec was already worked and then expired before
889 		 * we got around to our work.  In that case we have nothing
890 		 * tying us to an arfs_id, meaning that as soon as the filter
891 		 * is considered for expiry it will be removed.
892 		 */
893 		if (rule) {
894 			if (rc < 0)
895 				rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
896 			else
897 				rule->filter_id = rc;
898 			arfs_id = rule->arfs_id;
899 		}
900 		spin_unlock_bh(&efx->rps_hash_lock);
901 	}
902 	if (rc >= 0) {
903 		/* Remember this so we can check whether to expire the filter
904 		 * later.
905 		 */
906 		mutex_lock(&efx->rps_mutex);
907 		if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID)
908 			channel->rfs_filter_count++;
909 		channel->rps_flow_id[rc] = req->flow_id;
910 		mutex_unlock(&efx->rps_mutex);
911 
912 		if (req->spec.ether_type == htons(ETH_P_IP))
913 			netif_info(efx, rx_status, efx->net_dev,
914 				   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
915 				   (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
916 				   req->spec.rem_host, ntohs(req->spec.rem_port),
917 				   req->spec.loc_host, ntohs(req->spec.loc_port),
918 				   req->rxq_index, req->flow_id, rc, arfs_id);
919 		else
920 			netif_info(efx, rx_status, efx->net_dev,
921 				   "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
922 				   (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
923 				   req->spec.rem_host, ntohs(req->spec.rem_port),
924 				   req->spec.loc_host, ntohs(req->spec.loc_port),
925 				   req->rxq_index, req->flow_id, rc, arfs_id);
926 		channel->n_rfs_succeeded++;
927 	} else {
928 		if (req->spec.ether_type == htons(ETH_P_IP))
929 			netif_dbg(efx, rx_status, efx->net_dev,
930 				  "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n",
931 				  (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
932 				  req->spec.rem_host, ntohs(req->spec.rem_port),
933 				  req->spec.loc_host, ntohs(req->spec.loc_port),
934 				  req->rxq_index, req->flow_id, rc, arfs_id);
935 		else
936 			netif_dbg(efx, rx_status, efx->net_dev,
937 				  "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n",
938 				  (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
939 				  req->spec.rem_host, ntohs(req->spec.rem_port),
940 				  req->spec.loc_host, ntohs(req->spec.loc_port),
941 				  req->rxq_index, req->flow_id, rc, arfs_id);
942 		channel->n_rfs_failed++;
943 		/* We're overloading the NIC's filter tables, so let's do a
944 		 * chunk of extra expiry work.
945 		 */
946 		__efx_filter_rfs_expire(channel, min(channel->rfs_filter_count,
947 						     100u));
948 	}
949 
950 	/* Release references */
951 	clear_bit(slot_idx, &efx->rps_slot_map);
952 	dev_put(req->net_dev);
953 }
954 
955 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
956 		   u16 rxq_index, u32 flow_id)
957 {
958 	struct efx_nic *efx = netdev_priv(net_dev);
959 	struct efx_async_filter_insertion *req;
960 	struct efx_arfs_rule *rule;
961 	struct flow_keys fk;
962 	int slot_idx;
963 	bool new;
964 	int rc;
965 
966 	/* find a free slot */
967 	for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
968 		if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
969 			break;
970 	if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
971 		return -EBUSY;
972 
973 	if (flow_id == RPS_FLOW_ID_INVALID) {
974 		rc = -EINVAL;
975 		goto out_clear;
976 	}
977 
978 	if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
979 		rc = -EPROTONOSUPPORT;
980 		goto out_clear;
981 	}
982 
983 	if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
984 		rc = -EPROTONOSUPPORT;
985 		goto out_clear;
986 	}
987 	if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
988 		rc = -EPROTONOSUPPORT;
989 		goto out_clear;
990 	}
991 
992 	req = efx->rps_slot + slot_idx;
993 	efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
994 			   efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
995 			   rxq_index);
996 	req->spec.match_flags =
997 		EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
998 		EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
999 		EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
1000 	req->spec.ether_type = fk.basic.n_proto;
1001 	req->spec.ip_proto = fk.basic.ip_proto;
1002 
1003 	if (fk.basic.n_proto == htons(ETH_P_IP)) {
1004 		req->spec.rem_host[0] = fk.addrs.v4addrs.src;
1005 		req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
1006 	} else {
1007 		memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
1008 		       sizeof(struct in6_addr));
1009 		memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
1010 		       sizeof(struct in6_addr));
1011 	}
1012 
1013 	req->spec.rem_port = fk.ports.src;
1014 	req->spec.loc_port = fk.ports.dst;
1015 
1016 	if (efx->rps_hash_table) {
1017 		/* Add it to ARFS hash table */
1018 		spin_lock(&efx->rps_hash_lock);
1019 		rule = efx_rps_hash_add(efx, &req->spec, &new);
1020 		if (!rule) {
1021 			rc = -ENOMEM;
1022 			goto out_unlock;
1023 		}
1024 		if (new)
1025 			rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
1026 		rc = rule->arfs_id;
1027 		/* Skip if existing or pending filter already does the right thing */
1028 		if (!new && rule->rxq_index == rxq_index &&
1029 		    rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
1030 			goto out_unlock;
1031 		rule->rxq_index = rxq_index;
1032 		rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
1033 		spin_unlock(&efx->rps_hash_lock);
1034 	} else {
1035 		/* Without an ARFS hash table, we just use arfs_id 0 for all
1036 		 * filters.  This means if multiple flows hash to the same
1037 		 * flow_id, all but the most recently touched will be eligible
1038 		 * for expiry.
1039 		 */
1040 		rc = 0;
1041 	}
1042 
1043 	/* Queue the request */
1044 	dev_hold(req->net_dev = net_dev);
1045 	INIT_WORK(&req->work, efx_filter_rfs_work);
1046 	req->rxq_index = rxq_index;
1047 	req->flow_id = flow_id;
1048 	schedule_work(&req->work);
1049 	return rc;
1050 out_unlock:
1051 	spin_unlock(&efx->rps_hash_lock);
1052 out_clear:
1053 	clear_bit(slot_idx, &efx->rps_slot_map);
1054 	return rc;
1055 }
1056 
1057 bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota)
1058 {
1059 	bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
1060 	struct efx_nic *efx = channel->efx;
1061 	unsigned int index, size, start;
1062 	u32 flow_id;
1063 
1064 	if (!mutex_trylock(&efx->rps_mutex))
1065 		return false;
1066 	expire_one = efx->type->filter_rfs_expire_one;
1067 	index = channel->rfs_expire_index;
1068 	start = index;
1069 	size = efx->type->max_rx_ip_filters;
1070 	while (quota) {
1071 		flow_id = channel->rps_flow_id[index];
1072 
1073 		if (flow_id != RPS_FLOW_ID_INVALID) {
1074 			quota--;
1075 			if (expire_one(efx, flow_id, index)) {
1076 				netif_info(efx, rx_status, efx->net_dev,
1077 					   "expired filter %d [channel %u flow %u]\n",
1078 					   index, channel->channel, flow_id);
1079 				channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1080 				channel->rfs_filter_count--;
1081 			}
1082 		}
1083 		if (++index == size)
1084 			index = 0;
1085 		/* If we were called with a quota that exceeds the total number
1086 		 * of filters in the table (which shouldn't happen, but could
1087 		 * if two callers race), ensure that we don't loop forever -
1088 		 * stop when we've examined every row of the table.
1089 		 */
1090 		if (index == start)
1091 			break;
1092 	}
1093 
1094 	channel->rfs_expire_index = index;
1095 	mutex_unlock(&efx->rps_mutex);
1096 	return true;
1097 }
1098 
1099 #endif /* CONFIG_RFS_ACCEL */
1100