xref: /linux/drivers/net/ethernet/sfc/falcon/rx.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
1 /****************************************************************************
2  * Driver for Solarflare network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2013 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 <linux/socket.h>
12 #include <linux/in.h>
13 #include <linux/slab.h>
14 #include <linux/ip.h>
15 #include <linux/ipv6.h>
16 #include <linux/tcp.h>
17 #include <linux/udp.h>
18 #include <linux/prefetch.h>
19 #include <linux/moduleparam.h>
20 #include <linux/iommu.h>
21 #include <net/ip.h>
22 #include <net/checksum.h>
23 #include "net_driver.h"
24 #include "efx.h"
25 #include "filter.h"
26 #include "nic.h"
27 #include "selftest.h"
28 #include "workarounds.h"
29 
30 /* Preferred number of descriptors to fill at once */
31 #define EF4_RX_PREFERRED_BATCH 8U
32 
33 /* Number of RX buffers to recycle pages for.  When creating the RX page recycle
34  * ring, this number is divided by the number of buffers per page to calculate
35  * the number of pages to store in the RX page recycle ring.
36  */
37 #define EF4_RECYCLE_RING_SIZE_IOMMU 4096
38 #define EF4_RECYCLE_RING_SIZE_NOIOMMU (2 * EF4_RX_PREFERRED_BATCH)
39 
40 /* Size of buffer allocated for skb header area. */
41 #define EF4_SKB_HEADERS  128u
42 
43 /* This is the percentage fill level below which new RX descriptors
44  * will be added to the RX descriptor ring.
45  */
46 static unsigned int rx_refill_threshold;
47 
48 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */
49 #define EF4_RX_MAX_FRAGS DIV_ROUND_UP(EF4_MAX_FRAME_LEN(EF4_MAX_MTU), \
50 				      EF4_RX_USR_BUF_SIZE)
51 
52 /*
53  * RX maximum head room required.
54  *
55  * This must be at least 1 to prevent overflow, plus one packet-worth
56  * to allow pipelined receives.
57  */
58 #define EF4_RXD_HEAD_ROOM (1 + EF4_RX_MAX_FRAGS)
59 
60 static inline u8 *ef4_rx_buf_va(struct ef4_rx_buffer *buf)
61 {
62 	return page_address(buf->page) + buf->page_offset;
63 }
64 
65 static inline u32 ef4_rx_buf_hash(struct ef4_nic *efx, const u8 *eh)
66 {
67 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
68 	return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_hash_offset));
69 #else
70 	const u8 *data = eh + efx->rx_packet_hash_offset;
71 	return (u32)data[0]	  |
72 	       (u32)data[1] << 8  |
73 	       (u32)data[2] << 16 |
74 	       (u32)data[3] << 24;
75 #endif
76 }
77 
78 static inline struct ef4_rx_buffer *
79 ef4_rx_buf_next(struct ef4_rx_queue *rx_queue, struct ef4_rx_buffer *rx_buf)
80 {
81 	if (unlikely(rx_buf == ef4_rx_buffer(rx_queue, rx_queue->ptr_mask)))
82 		return ef4_rx_buffer(rx_queue, 0);
83 	else
84 		return rx_buf + 1;
85 }
86 
87 static inline void ef4_sync_rx_buffer(struct ef4_nic *efx,
88 				      struct ef4_rx_buffer *rx_buf,
89 				      unsigned int len)
90 {
91 	dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr, len,
92 				DMA_FROM_DEVICE);
93 }
94 
95 void ef4_rx_config_page_split(struct ef4_nic *efx)
96 {
97 	efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align,
98 				      EF4_RX_BUF_ALIGNMENT);
99 	efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
100 		((PAGE_SIZE - sizeof(struct ef4_rx_page_state)) /
101 		 efx->rx_page_buf_step);
102 	efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) /
103 		efx->rx_bufs_per_page;
104 	efx->rx_pages_per_batch = DIV_ROUND_UP(EF4_RX_PREFERRED_BATCH,
105 					       efx->rx_bufs_per_page);
106 }
107 
108 /* Check the RX page recycle ring for a page that can be reused. */
109 static struct page *ef4_reuse_page(struct ef4_rx_queue *rx_queue)
110 {
111 	struct ef4_nic *efx = rx_queue->efx;
112 	struct page *page;
113 	struct ef4_rx_page_state *state;
114 	unsigned index;
115 
116 	index = rx_queue->page_remove & rx_queue->page_ptr_mask;
117 	page = rx_queue->page_ring[index];
118 	if (page == NULL)
119 		return NULL;
120 
121 	rx_queue->page_ring[index] = NULL;
122 	/* page_remove cannot exceed page_add. */
123 	if (rx_queue->page_remove != rx_queue->page_add)
124 		++rx_queue->page_remove;
125 
126 	/* If page_count is 1 then we hold the only reference to this page. */
127 	if (page_count(page) == 1) {
128 		++rx_queue->page_recycle_count;
129 		return page;
130 	} else {
131 		state = page_address(page);
132 		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
133 			       PAGE_SIZE << efx->rx_buffer_order,
134 			       DMA_FROM_DEVICE);
135 		put_page(page);
136 		++rx_queue->page_recycle_failed;
137 	}
138 
139 	return NULL;
140 }
141 
142 /**
143  * ef4_init_rx_buffers - create EF4_RX_BATCH page-based RX buffers
144  *
145  * @rx_queue:		Efx RX queue
146  *
147  * This allocates a batch of pages, maps them for DMA, and populates
148  * struct ef4_rx_buffers for each one. Return a negative error code or
149  * 0 on success. If a single page can be used for multiple buffers,
150  * then the page will either be inserted fully, or not at all.
151  */
152 static int ef4_init_rx_buffers(struct ef4_rx_queue *rx_queue, bool atomic)
153 {
154 	struct ef4_nic *efx = rx_queue->efx;
155 	struct ef4_rx_buffer *rx_buf;
156 	struct page *page;
157 	unsigned int page_offset;
158 	struct ef4_rx_page_state *state;
159 	dma_addr_t dma_addr;
160 	unsigned index, count;
161 
162 	count = 0;
163 	do {
164 		page = ef4_reuse_page(rx_queue);
165 		if (page == NULL) {
166 			page = alloc_pages(__GFP_COLD | __GFP_COMP |
167 					   (atomic ? GFP_ATOMIC : GFP_KERNEL),
168 					   efx->rx_buffer_order);
169 			if (unlikely(page == NULL))
170 				return -ENOMEM;
171 			dma_addr =
172 				dma_map_page(&efx->pci_dev->dev, page, 0,
173 					     PAGE_SIZE << efx->rx_buffer_order,
174 					     DMA_FROM_DEVICE);
175 			if (unlikely(dma_mapping_error(&efx->pci_dev->dev,
176 						       dma_addr))) {
177 				__free_pages(page, efx->rx_buffer_order);
178 				return -EIO;
179 			}
180 			state = page_address(page);
181 			state->dma_addr = dma_addr;
182 		} else {
183 			state = page_address(page);
184 			dma_addr = state->dma_addr;
185 		}
186 
187 		dma_addr += sizeof(struct ef4_rx_page_state);
188 		page_offset = sizeof(struct ef4_rx_page_state);
189 
190 		do {
191 			index = rx_queue->added_count & rx_queue->ptr_mask;
192 			rx_buf = ef4_rx_buffer(rx_queue, index);
193 			rx_buf->dma_addr = dma_addr + efx->rx_ip_align;
194 			rx_buf->page = page;
195 			rx_buf->page_offset = page_offset + efx->rx_ip_align;
196 			rx_buf->len = efx->rx_dma_len;
197 			rx_buf->flags = 0;
198 			++rx_queue->added_count;
199 			get_page(page);
200 			dma_addr += efx->rx_page_buf_step;
201 			page_offset += efx->rx_page_buf_step;
202 		} while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE);
203 
204 		rx_buf->flags = EF4_RX_BUF_LAST_IN_PAGE;
205 	} while (++count < efx->rx_pages_per_batch);
206 
207 	return 0;
208 }
209 
210 /* Unmap a DMA-mapped page.  This function is only called for the final RX
211  * buffer in a page.
212  */
213 static void ef4_unmap_rx_buffer(struct ef4_nic *efx,
214 				struct ef4_rx_buffer *rx_buf)
215 {
216 	struct page *page = rx_buf->page;
217 
218 	if (page) {
219 		struct ef4_rx_page_state *state = page_address(page);
220 		dma_unmap_page(&efx->pci_dev->dev,
221 			       state->dma_addr,
222 			       PAGE_SIZE << efx->rx_buffer_order,
223 			       DMA_FROM_DEVICE);
224 	}
225 }
226 
227 static void ef4_free_rx_buffers(struct ef4_rx_queue *rx_queue,
228 				struct ef4_rx_buffer *rx_buf,
229 				unsigned int num_bufs)
230 {
231 	do {
232 		if (rx_buf->page) {
233 			put_page(rx_buf->page);
234 			rx_buf->page = NULL;
235 		}
236 		rx_buf = ef4_rx_buf_next(rx_queue, rx_buf);
237 	} while (--num_bufs);
238 }
239 
240 /* Attempt to recycle the page if there is an RX recycle ring; the page can
241  * only be added if this is the final RX buffer, to prevent pages being used in
242  * the descriptor ring and appearing in the recycle ring simultaneously.
243  */
244 static void ef4_recycle_rx_page(struct ef4_channel *channel,
245 				struct ef4_rx_buffer *rx_buf)
246 {
247 	struct page *page = rx_buf->page;
248 	struct ef4_rx_queue *rx_queue = ef4_channel_get_rx_queue(channel);
249 	struct ef4_nic *efx = rx_queue->efx;
250 	unsigned index;
251 
252 	/* Only recycle the page after processing the final buffer. */
253 	if (!(rx_buf->flags & EF4_RX_BUF_LAST_IN_PAGE))
254 		return;
255 
256 	index = rx_queue->page_add & rx_queue->page_ptr_mask;
257 	if (rx_queue->page_ring[index] == NULL) {
258 		unsigned read_index = rx_queue->page_remove &
259 			rx_queue->page_ptr_mask;
260 
261 		/* The next slot in the recycle ring is available, but
262 		 * increment page_remove if the read pointer currently
263 		 * points here.
264 		 */
265 		if (read_index == index)
266 			++rx_queue->page_remove;
267 		rx_queue->page_ring[index] = page;
268 		++rx_queue->page_add;
269 		return;
270 	}
271 	++rx_queue->page_recycle_full;
272 	ef4_unmap_rx_buffer(efx, rx_buf);
273 	put_page(rx_buf->page);
274 }
275 
276 static void ef4_fini_rx_buffer(struct ef4_rx_queue *rx_queue,
277 			       struct ef4_rx_buffer *rx_buf)
278 {
279 	/* Release the page reference we hold for the buffer. */
280 	if (rx_buf->page)
281 		put_page(rx_buf->page);
282 
283 	/* If this is the last buffer in a page, unmap and free it. */
284 	if (rx_buf->flags & EF4_RX_BUF_LAST_IN_PAGE) {
285 		ef4_unmap_rx_buffer(rx_queue->efx, rx_buf);
286 		ef4_free_rx_buffers(rx_queue, rx_buf, 1);
287 	}
288 	rx_buf->page = NULL;
289 }
290 
291 /* Recycle the pages that are used by buffers that have just been received. */
292 static void ef4_recycle_rx_pages(struct ef4_channel *channel,
293 				 struct ef4_rx_buffer *rx_buf,
294 				 unsigned int n_frags)
295 {
296 	struct ef4_rx_queue *rx_queue = ef4_channel_get_rx_queue(channel);
297 
298 	do {
299 		ef4_recycle_rx_page(channel, rx_buf);
300 		rx_buf = ef4_rx_buf_next(rx_queue, rx_buf);
301 	} while (--n_frags);
302 }
303 
304 static void ef4_discard_rx_packet(struct ef4_channel *channel,
305 				  struct ef4_rx_buffer *rx_buf,
306 				  unsigned int n_frags)
307 {
308 	struct ef4_rx_queue *rx_queue = ef4_channel_get_rx_queue(channel);
309 
310 	ef4_recycle_rx_pages(channel, rx_buf, n_frags);
311 
312 	ef4_free_rx_buffers(rx_queue, rx_buf, n_frags);
313 }
314 
315 /**
316  * ef4_fast_push_rx_descriptors - push new RX descriptors quickly
317  * @rx_queue:		RX descriptor queue
318  *
319  * This will aim to fill the RX descriptor queue up to
320  * @rx_queue->@max_fill. If there is insufficient atomic
321  * memory to do so, a slow fill will be scheduled.
322  *
323  * The caller must provide serialisation (none is used here). In practise,
324  * this means this function must run from the NAPI handler, or be called
325  * when NAPI is disabled.
326  */
327 void ef4_fast_push_rx_descriptors(struct ef4_rx_queue *rx_queue, bool atomic)
328 {
329 	struct ef4_nic *efx = rx_queue->efx;
330 	unsigned int fill_level, batch_size;
331 	int space, rc = 0;
332 
333 	if (!rx_queue->refill_enabled)
334 		return;
335 
336 	/* Calculate current fill level, and exit if we don't need to fill */
337 	fill_level = (rx_queue->added_count - rx_queue->removed_count);
338 	EF4_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
339 	if (fill_level >= rx_queue->fast_fill_trigger)
340 		goto out;
341 
342 	/* Record minimum fill level */
343 	if (unlikely(fill_level < rx_queue->min_fill)) {
344 		if (fill_level)
345 			rx_queue->min_fill = fill_level;
346 	}
347 
348 	batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page;
349 	space = rx_queue->max_fill - fill_level;
350 	EF4_BUG_ON_PARANOID(space < batch_size);
351 
352 	netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
353 		   "RX queue %d fast-filling descriptor ring from"
354 		   " level %d to level %d\n",
355 		   ef4_rx_queue_index(rx_queue), fill_level,
356 		   rx_queue->max_fill);
357 
358 
359 	do {
360 		rc = ef4_init_rx_buffers(rx_queue, atomic);
361 		if (unlikely(rc)) {
362 			/* Ensure that we don't leave the rx queue empty */
363 			if (rx_queue->added_count == rx_queue->removed_count)
364 				ef4_schedule_slow_fill(rx_queue);
365 			goto out;
366 		}
367 	} while ((space -= batch_size) >= batch_size);
368 
369 	netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
370 		   "RX queue %d fast-filled descriptor ring "
371 		   "to level %d\n", ef4_rx_queue_index(rx_queue),
372 		   rx_queue->added_count - rx_queue->removed_count);
373 
374  out:
375 	if (rx_queue->notified_count != rx_queue->added_count)
376 		ef4_nic_notify_rx_desc(rx_queue);
377 }
378 
379 void ef4_rx_slow_fill(unsigned long context)
380 {
381 	struct ef4_rx_queue *rx_queue = (struct ef4_rx_queue *)context;
382 
383 	/* Post an event to cause NAPI to run and refill the queue */
384 	ef4_nic_generate_fill_event(rx_queue);
385 	++rx_queue->slow_fill_count;
386 }
387 
388 static void ef4_rx_packet__check_len(struct ef4_rx_queue *rx_queue,
389 				     struct ef4_rx_buffer *rx_buf,
390 				     int len)
391 {
392 	struct ef4_nic *efx = rx_queue->efx;
393 	unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
394 
395 	if (likely(len <= max_len))
396 		return;
397 
398 	/* The packet must be discarded, but this is only a fatal error
399 	 * if the caller indicated it was
400 	 */
401 	rx_buf->flags |= EF4_RX_PKT_DISCARD;
402 
403 	if ((len > rx_buf->len) && EF4_WORKAROUND_8071(efx)) {
404 		if (net_ratelimit())
405 			netif_err(efx, rx_err, efx->net_dev,
406 				  " RX queue %d seriously overlength "
407 				  "RX event (0x%x > 0x%x+0x%x). Leaking\n",
408 				  ef4_rx_queue_index(rx_queue), len, max_len,
409 				  efx->type->rx_buffer_padding);
410 		ef4_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
411 	} else {
412 		if (net_ratelimit())
413 			netif_err(efx, rx_err, efx->net_dev,
414 				  " RX queue %d overlength RX event "
415 				  "(0x%x > 0x%x)\n",
416 				  ef4_rx_queue_index(rx_queue), len, max_len);
417 	}
418 
419 	ef4_rx_queue_channel(rx_queue)->n_rx_overlength++;
420 }
421 
422 /* Pass a received packet up through GRO.  GRO can handle pages
423  * regardless of checksum state and skbs with a good checksum.
424  */
425 static void
426 ef4_rx_packet_gro(struct ef4_channel *channel, struct ef4_rx_buffer *rx_buf,
427 		  unsigned int n_frags, u8 *eh)
428 {
429 	struct napi_struct *napi = &channel->napi_str;
430 	gro_result_t gro_result;
431 	struct ef4_nic *efx = channel->efx;
432 	struct sk_buff *skb;
433 
434 	skb = napi_get_frags(napi);
435 	if (unlikely(!skb)) {
436 		struct ef4_rx_queue *rx_queue;
437 
438 		rx_queue = ef4_channel_get_rx_queue(channel);
439 		ef4_free_rx_buffers(rx_queue, rx_buf, n_frags);
440 		return;
441 	}
442 
443 	if (efx->net_dev->features & NETIF_F_RXHASH)
444 		skb_set_hash(skb, ef4_rx_buf_hash(efx, eh),
445 			     PKT_HASH_TYPE_L3);
446 	skb->ip_summed = ((rx_buf->flags & EF4_RX_PKT_CSUMMED) ?
447 			  CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
448 
449 	for (;;) {
450 		skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
451 				   rx_buf->page, rx_buf->page_offset,
452 				   rx_buf->len);
453 		rx_buf->page = NULL;
454 		skb->len += rx_buf->len;
455 		if (skb_shinfo(skb)->nr_frags == n_frags)
456 			break;
457 
458 		rx_buf = ef4_rx_buf_next(&channel->rx_queue, rx_buf);
459 	}
460 
461 	skb->data_len = skb->len;
462 	skb->truesize += n_frags * efx->rx_buffer_truesize;
463 
464 	skb_record_rx_queue(skb, channel->rx_queue.core_index);
465 
466 	gro_result = napi_gro_frags(napi);
467 	if (gro_result != GRO_DROP)
468 		channel->irq_mod_score += 2;
469 }
470 
471 /* Allocate and construct an SKB around page fragments */
472 static struct sk_buff *ef4_rx_mk_skb(struct ef4_channel *channel,
473 				     struct ef4_rx_buffer *rx_buf,
474 				     unsigned int n_frags,
475 				     u8 *eh, int hdr_len)
476 {
477 	struct ef4_nic *efx = channel->efx;
478 	struct sk_buff *skb;
479 
480 	/* Allocate an SKB to store the headers */
481 	skb = netdev_alloc_skb(efx->net_dev,
482 			       efx->rx_ip_align + efx->rx_prefix_size +
483 			       hdr_len);
484 	if (unlikely(skb == NULL)) {
485 		atomic_inc(&efx->n_rx_noskb_drops);
486 		return NULL;
487 	}
488 
489 	EF4_BUG_ON_PARANOID(rx_buf->len < hdr_len);
490 
491 	memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
492 	       efx->rx_prefix_size + hdr_len);
493 	skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
494 	__skb_put(skb, hdr_len);
495 
496 	/* Append the remaining page(s) onto the frag list */
497 	if (rx_buf->len > hdr_len) {
498 		rx_buf->page_offset += hdr_len;
499 		rx_buf->len -= hdr_len;
500 
501 		for (;;) {
502 			skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
503 					   rx_buf->page, rx_buf->page_offset,
504 					   rx_buf->len);
505 			rx_buf->page = NULL;
506 			skb->len += rx_buf->len;
507 			skb->data_len += rx_buf->len;
508 			if (skb_shinfo(skb)->nr_frags == n_frags)
509 				break;
510 
511 			rx_buf = ef4_rx_buf_next(&channel->rx_queue, rx_buf);
512 		}
513 	} else {
514 		__free_pages(rx_buf->page, efx->rx_buffer_order);
515 		rx_buf->page = NULL;
516 		n_frags = 0;
517 	}
518 
519 	skb->truesize += n_frags * efx->rx_buffer_truesize;
520 
521 	/* Move past the ethernet header */
522 	skb->protocol = eth_type_trans(skb, efx->net_dev);
523 
524 	skb_mark_napi_id(skb, &channel->napi_str);
525 
526 	return skb;
527 }
528 
529 void ef4_rx_packet(struct ef4_rx_queue *rx_queue, unsigned int index,
530 		   unsigned int n_frags, unsigned int len, u16 flags)
531 {
532 	struct ef4_nic *efx = rx_queue->efx;
533 	struct ef4_channel *channel = ef4_rx_queue_channel(rx_queue);
534 	struct ef4_rx_buffer *rx_buf;
535 
536 	rx_queue->rx_packets++;
537 
538 	rx_buf = ef4_rx_buffer(rx_queue, index);
539 	rx_buf->flags |= flags;
540 
541 	/* Validate the number of fragments and completed length */
542 	if (n_frags == 1) {
543 		if (!(flags & EF4_RX_PKT_PREFIX_LEN))
544 			ef4_rx_packet__check_len(rx_queue, rx_buf, len);
545 	} else if (unlikely(n_frags > EF4_RX_MAX_FRAGS) ||
546 		   unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
547 		   unlikely(len > n_frags * efx->rx_dma_len) ||
548 		   unlikely(!efx->rx_scatter)) {
549 		/* If this isn't an explicit discard request, either
550 		 * the hardware or the driver is broken.
551 		 */
552 		WARN_ON(!(len == 0 && rx_buf->flags & EF4_RX_PKT_DISCARD));
553 		rx_buf->flags |= EF4_RX_PKT_DISCARD;
554 	}
555 
556 	netif_vdbg(efx, rx_status, efx->net_dev,
557 		   "RX queue %d received ids %x-%x len %d %s%s\n",
558 		   ef4_rx_queue_index(rx_queue), index,
559 		   (index + n_frags - 1) & rx_queue->ptr_mask, len,
560 		   (rx_buf->flags & EF4_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
561 		   (rx_buf->flags & EF4_RX_PKT_DISCARD) ? " [DISCARD]" : "");
562 
563 	/* Discard packet, if instructed to do so.  Process the
564 	 * previous receive first.
565 	 */
566 	if (unlikely(rx_buf->flags & EF4_RX_PKT_DISCARD)) {
567 		ef4_rx_flush_packet(channel);
568 		ef4_discard_rx_packet(channel, rx_buf, n_frags);
569 		return;
570 	}
571 
572 	if (n_frags == 1 && !(flags & EF4_RX_PKT_PREFIX_LEN))
573 		rx_buf->len = len;
574 
575 	/* Release and/or sync the DMA mapping - assumes all RX buffers
576 	 * consumed in-order per RX queue.
577 	 */
578 	ef4_sync_rx_buffer(efx, rx_buf, rx_buf->len);
579 
580 	/* Prefetch nice and early so data will (hopefully) be in cache by
581 	 * the time we look at it.
582 	 */
583 	prefetch(ef4_rx_buf_va(rx_buf));
584 
585 	rx_buf->page_offset += efx->rx_prefix_size;
586 	rx_buf->len -= efx->rx_prefix_size;
587 
588 	if (n_frags > 1) {
589 		/* Release/sync DMA mapping for additional fragments.
590 		 * Fix length for last fragment.
591 		 */
592 		unsigned int tail_frags = n_frags - 1;
593 
594 		for (;;) {
595 			rx_buf = ef4_rx_buf_next(rx_queue, rx_buf);
596 			if (--tail_frags == 0)
597 				break;
598 			ef4_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
599 		}
600 		rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
601 		ef4_sync_rx_buffer(efx, rx_buf, rx_buf->len);
602 	}
603 
604 	/* All fragments have been DMA-synced, so recycle pages. */
605 	rx_buf = ef4_rx_buffer(rx_queue, index);
606 	ef4_recycle_rx_pages(channel, rx_buf, n_frags);
607 
608 	/* Pipeline receives so that we give time for packet headers to be
609 	 * prefetched into cache.
610 	 */
611 	ef4_rx_flush_packet(channel);
612 	channel->rx_pkt_n_frags = n_frags;
613 	channel->rx_pkt_index = index;
614 }
615 
616 static void ef4_rx_deliver(struct ef4_channel *channel, u8 *eh,
617 			   struct ef4_rx_buffer *rx_buf,
618 			   unsigned int n_frags)
619 {
620 	struct sk_buff *skb;
621 	u16 hdr_len = min_t(u16, rx_buf->len, EF4_SKB_HEADERS);
622 
623 	skb = ef4_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
624 	if (unlikely(skb == NULL)) {
625 		struct ef4_rx_queue *rx_queue;
626 
627 		rx_queue = ef4_channel_get_rx_queue(channel);
628 		ef4_free_rx_buffers(rx_queue, rx_buf, n_frags);
629 		return;
630 	}
631 	skb_record_rx_queue(skb, channel->rx_queue.core_index);
632 
633 	/* Set the SKB flags */
634 	skb_checksum_none_assert(skb);
635 	if (likely(rx_buf->flags & EF4_RX_PKT_CSUMMED))
636 		skb->ip_summed = CHECKSUM_UNNECESSARY;
637 
638 	if (channel->type->receive_skb)
639 		if (channel->type->receive_skb(channel, skb))
640 			return;
641 
642 	/* Pass the packet up */
643 	netif_receive_skb(skb);
644 }
645 
646 /* Handle a received packet.  Second half: Touches packet payload. */
647 void __ef4_rx_packet(struct ef4_channel *channel)
648 {
649 	struct ef4_nic *efx = channel->efx;
650 	struct ef4_rx_buffer *rx_buf =
651 		ef4_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
652 	u8 *eh = ef4_rx_buf_va(rx_buf);
653 
654 	/* Read length from the prefix if necessary.  This already
655 	 * excludes the length of the prefix itself.
656 	 */
657 	if (rx_buf->flags & EF4_RX_PKT_PREFIX_LEN)
658 		rx_buf->len = le16_to_cpup((__le16 *)
659 					   (eh + efx->rx_packet_len_offset));
660 
661 	/* If we're in loopback test, then pass the packet directly to the
662 	 * loopback layer, and free the rx_buf here
663 	 */
664 	if (unlikely(efx->loopback_selftest)) {
665 		struct ef4_rx_queue *rx_queue;
666 
667 		ef4_loopback_rx_packet(efx, eh, rx_buf->len);
668 		rx_queue = ef4_channel_get_rx_queue(channel);
669 		ef4_free_rx_buffers(rx_queue, rx_buf,
670 				    channel->rx_pkt_n_frags);
671 		goto out;
672 	}
673 
674 	if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
675 		rx_buf->flags &= ~EF4_RX_PKT_CSUMMED;
676 
677 	if ((rx_buf->flags & EF4_RX_PKT_TCP) && !channel->type->receive_skb)
678 		ef4_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
679 	else
680 		ef4_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
681 out:
682 	channel->rx_pkt_n_frags = 0;
683 }
684 
685 int ef4_probe_rx_queue(struct ef4_rx_queue *rx_queue)
686 {
687 	struct ef4_nic *efx = rx_queue->efx;
688 	unsigned int entries;
689 	int rc;
690 
691 	/* Create the smallest power-of-two aligned ring */
692 	entries = max(roundup_pow_of_two(efx->rxq_entries), EF4_MIN_DMAQ_SIZE);
693 	EF4_BUG_ON_PARANOID(entries > EF4_MAX_DMAQ_SIZE);
694 	rx_queue->ptr_mask = entries - 1;
695 
696 	netif_dbg(efx, probe, efx->net_dev,
697 		  "creating RX queue %d size %#x mask %#x\n",
698 		  ef4_rx_queue_index(rx_queue), efx->rxq_entries,
699 		  rx_queue->ptr_mask);
700 
701 	/* Allocate RX buffers */
702 	rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
703 				   GFP_KERNEL);
704 	if (!rx_queue->buffer)
705 		return -ENOMEM;
706 
707 	rc = ef4_nic_probe_rx(rx_queue);
708 	if (rc) {
709 		kfree(rx_queue->buffer);
710 		rx_queue->buffer = NULL;
711 	}
712 
713 	return rc;
714 }
715 
716 static void ef4_init_rx_recycle_ring(struct ef4_nic *efx,
717 				     struct ef4_rx_queue *rx_queue)
718 {
719 	unsigned int bufs_in_recycle_ring, page_ring_size;
720 
721 	/* Set the RX recycle ring size */
722 #ifdef CONFIG_PPC64
723 	bufs_in_recycle_ring = EF4_RECYCLE_RING_SIZE_IOMMU;
724 #else
725 	if (iommu_present(&pci_bus_type))
726 		bufs_in_recycle_ring = EF4_RECYCLE_RING_SIZE_IOMMU;
727 	else
728 		bufs_in_recycle_ring = EF4_RECYCLE_RING_SIZE_NOIOMMU;
729 #endif /* CONFIG_PPC64 */
730 
731 	page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
732 					    efx->rx_bufs_per_page);
733 	rx_queue->page_ring = kcalloc(page_ring_size,
734 				      sizeof(*rx_queue->page_ring), GFP_KERNEL);
735 	rx_queue->page_ptr_mask = page_ring_size - 1;
736 }
737 
738 void ef4_init_rx_queue(struct ef4_rx_queue *rx_queue)
739 {
740 	struct ef4_nic *efx = rx_queue->efx;
741 	unsigned int max_fill, trigger, max_trigger;
742 
743 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
744 		  "initialising RX queue %d\n", ef4_rx_queue_index(rx_queue));
745 
746 	/* Initialise ptr fields */
747 	rx_queue->added_count = 0;
748 	rx_queue->notified_count = 0;
749 	rx_queue->removed_count = 0;
750 	rx_queue->min_fill = -1U;
751 	ef4_init_rx_recycle_ring(efx, rx_queue);
752 
753 	rx_queue->page_remove = 0;
754 	rx_queue->page_add = rx_queue->page_ptr_mask + 1;
755 	rx_queue->page_recycle_count = 0;
756 	rx_queue->page_recycle_failed = 0;
757 	rx_queue->page_recycle_full = 0;
758 
759 	/* Initialise limit fields */
760 	max_fill = efx->rxq_entries - EF4_RXD_HEAD_ROOM;
761 	max_trigger =
762 		max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
763 	if (rx_refill_threshold != 0) {
764 		trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
765 		if (trigger > max_trigger)
766 			trigger = max_trigger;
767 	} else {
768 		trigger = max_trigger;
769 	}
770 
771 	rx_queue->max_fill = max_fill;
772 	rx_queue->fast_fill_trigger = trigger;
773 	rx_queue->refill_enabled = true;
774 
775 	/* Set up RX descriptor ring */
776 	ef4_nic_init_rx(rx_queue);
777 }
778 
779 void ef4_fini_rx_queue(struct ef4_rx_queue *rx_queue)
780 {
781 	int i;
782 	struct ef4_nic *efx = rx_queue->efx;
783 	struct ef4_rx_buffer *rx_buf;
784 
785 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
786 		  "shutting down RX queue %d\n", ef4_rx_queue_index(rx_queue));
787 
788 	del_timer_sync(&rx_queue->slow_fill);
789 
790 	/* Release RX buffers from the current read ptr to the write ptr */
791 	if (rx_queue->buffer) {
792 		for (i = rx_queue->removed_count; i < rx_queue->added_count;
793 		     i++) {
794 			unsigned index = i & rx_queue->ptr_mask;
795 			rx_buf = ef4_rx_buffer(rx_queue, index);
796 			ef4_fini_rx_buffer(rx_queue, rx_buf);
797 		}
798 	}
799 
800 	/* Unmap and release the pages in the recycle ring. Remove the ring. */
801 	for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
802 		struct page *page = rx_queue->page_ring[i];
803 		struct ef4_rx_page_state *state;
804 
805 		if (page == NULL)
806 			continue;
807 
808 		state = page_address(page);
809 		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
810 			       PAGE_SIZE << efx->rx_buffer_order,
811 			       DMA_FROM_DEVICE);
812 		put_page(page);
813 	}
814 	kfree(rx_queue->page_ring);
815 	rx_queue->page_ring = NULL;
816 }
817 
818 void ef4_remove_rx_queue(struct ef4_rx_queue *rx_queue)
819 {
820 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
821 		  "destroying RX queue %d\n", ef4_rx_queue_index(rx_queue));
822 
823 	ef4_nic_remove_rx(rx_queue);
824 
825 	kfree(rx_queue->buffer);
826 	rx_queue->buffer = NULL;
827 }
828 
829 
830 module_param(rx_refill_threshold, uint, 0444);
831 MODULE_PARM_DESC(rx_refill_threshold,
832 		 "RX descriptor ring refill threshold (%)");
833 
834 #ifdef CONFIG_RFS_ACCEL
835 
836 int ef4_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
837 		   u16 rxq_index, u32 flow_id)
838 {
839 	struct ef4_nic *efx = netdev_priv(net_dev);
840 	struct ef4_channel *channel;
841 	struct ef4_filter_spec spec;
842 	struct flow_keys fk;
843 	int rc;
844 
845 	if (flow_id == RPS_FLOW_ID_INVALID)
846 		return -EINVAL;
847 
848 	if (!skb_flow_dissect_flow_keys(skb, &fk, 0))
849 		return -EPROTONOSUPPORT;
850 
851 	if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6))
852 		return -EPROTONOSUPPORT;
853 	if (fk.control.flags & FLOW_DIS_IS_FRAGMENT)
854 		return -EPROTONOSUPPORT;
855 
856 	ef4_filter_init_rx(&spec, EF4_FILTER_PRI_HINT,
857 			   efx->rx_scatter ? EF4_FILTER_FLAG_RX_SCATTER : 0,
858 			   rxq_index);
859 	spec.match_flags =
860 		EF4_FILTER_MATCH_ETHER_TYPE | EF4_FILTER_MATCH_IP_PROTO |
861 		EF4_FILTER_MATCH_LOC_HOST | EF4_FILTER_MATCH_LOC_PORT |
862 		EF4_FILTER_MATCH_REM_HOST | EF4_FILTER_MATCH_REM_PORT;
863 	spec.ether_type = fk.basic.n_proto;
864 	spec.ip_proto = fk.basic.ip_proto;
865 
866 	if (fk.basic.n_proto == htons(ETH_P_IP)) {
867 		spec.rem_host[0] = fk.addrs.v4addrs.src;
868 		spec.loc_host[0] = fk.addrs.v4addrs.dst;
869 	} else {
870 		memcpy(spec.rem_host, &fk.addrs.v6addrs.src, sizeof(struct in6_addr));
871 		memcpy(spec.loc_host, &fk.addrs.v6addrs.dst, sizeof(struct in6_addr));
872 	}
873 
874 	spec.rem_port = fk.ports.src;
875 	spec.loc_port = fk.ports.dst;
876 
877 	rc = efx->type->filter_rfs_insert(efx, &spec);
878 	if (rc < 0)
879 		return rc;
880 
881 	/* Remember this so we can check whether to expire the filter later */
882 	channel = ef4_get_channel(efx, rxq_index);
883 	channel->rps_flow_id[rc] = flow_id;
884 	++channel->rfs_filters_added;
885 
886 	if (spec.ether_type == htons(ETH_P_IP))
887 		netif_info(efx, rx_status, efx->net_dev,
888 			   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n",
889 			   (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
890 			   spec.rem_host, ntohs(spec.rem_port), spec.loc_host,
891 			   ntohs(spec.loc_port), rxq_index, flow_id, rc);
892 	else
893 		netif_info(efx, rx_status, efx->net_dev,
894 			   "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d]\n",
895 			   (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
896 			   spec.rem_host, ntohs(spec.rem_port), spec.loc_host,
897 			   ntohs(spec.loc_port), rxq_index, flow_id, rc);
898 
899 	return rc;
900 }
901 
902 bool __ef4_filter_rfs_expire(struct ef4_nic *efx, unsigned int quota)
903 {
904 	bool (*expire_one)(struct ef4_nic *efx, u32 flow_id, unsigned int index);
905 	unsigned int channel_idx, index, size;
906 	u32 flow_id;
907 
908 	if (!spin_trylock_bh(&efx->filter_lock))
909 		return false;
910 
911 	expire_one = efx->type->filter_rfs_expire_one;
912 	channel_idx = efx->rps_expire_channel;
913 	index = efx->rps_expire_index;
914 	size = efx->type->max_rx_ip_filters;
915 	while (quota--) {
916 		struct ef4_channel *channel = ef4_get_channel(efx, channel_idx);
917 		flow_id = channel->rps_flow_id[index];
918 
919 		if (flow_id != RPS_FLOW_ID_INVALID &&
920 		    expire_one(efx, flow_id, index)) {
921 			netif_info(efx, rx_status, efx->net_dev,
922 				   "expired filter %d [queue %u flow %u]\n",
923 				   index, channel_idx, flow_id);
924 			channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
925 		}
926 		if (++index == size) {
927 			if (++channel_idx == efx->n_channels)
928 				channel_idx = 0;
929 			index = 0;
930 		}
931 	}
932 	efx->rps_expire_channel = channel_idx;
933 	efx->rps_expire_index = index;
934 
935 	spin_unlock_bh(&efx->filter_lock);
936 	return true;
937 }
938 
939 #endif /* CONFIG_RFS_ACCEL */
940 
941 /**
942  * ef4_filter_is_mc_recipient - test whether spec is a multicast recipient
943  * @spec: Specification to test
944  *
945  * Return: %true if the specification is a non-drop RX filter that
946  * matches a local MAC address I/G bit value of 1 or matches a local
947  * IPv4 or IPv6 address value in the respective multicast address
948  * range.  Otherwise %false.
949  */
950 bool ef4_filter_is_mc_recipient(const struct ef4_filter_spec *spec)
951 {
952 	if (!(spec->flags & EF4_FILTER_FLAG_RX) ||
953 	    spec->dmaq_id == EF4_FILTER_RX_DMAQ_ID_DROP)
954 		return false;
955 
956 	if (spec->match_flags &
957 	    (EF4_FILTER_MATCH_LOC_MAC | EF4_FILTER_MATCH_LOC_MAC_IG) &&
958 	    is_multicast_ether_addr(spec->loc_mac))
959 		return true;
960 
961 	if ((spec->match_flags &
962 	     (EF4_FILTER_MATCH_ETHER_TYPE | EF4_FILTER_MATCH_LOC_HOST)) ==
963 	    (EF4_FILTER_MATCH_ETHER_TYPE | EF4_FILTER_MATCH_LOC_HOST)) {
964 		if (spec->ether_type == htons(ETH_P_IP) &&
965 		    ipv4_is_multicast(spec->loc_host[0]))
966 			return true;
967 		if (spec->ether_type == htons(ETH_P_IPV6) &&
968 		    ((const u8 *)spec->loc_host)[0] == 0xff)
969 			return true;
970 	}
971 
972 	return false;
973 }
974