xref: /linux/drivers/net/ethernet/sfc/rx.c (revision c0c914eca7f251c70facc37dfebeaf176601918d)
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 EFX_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 EFX_RECYCLE_RING_SIZE_IOMMU 4096
38 #define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH)
39 
40 /* Size of buffer allocated for skb header area. */
41 #define EFX_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 EFX_RX_MAX_FRAGS DIV_ROUND_UP(EFX_MAX_FRAME_LEN(EFX_MAX_MTU), \
50 				      EFX_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 EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS)
59 
60 static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
61 {
62 	return page_address(buf->page) + buf->page_offset;
63 }
64 
65 static inline u32 efx_rx_buf_hash(struct efx_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 efx_rx_buffer *
79 efx_rx_buf_next(struct efx_rx_queue *rx_queue, struct efx_rx_buffer *rx_buf)
80 {
81 	if (unlikely(rx_buf == efx_rx_buffer(rx_queue, rx_queue->ptr_mask)))
82 		return efx_rx_buffer(rx_queue, 0);
83 	else
84 		return rx_buf + 1;
85 }
86 
87 static inline void efx_sync_rx_buffer(struct efx_nic *efx,
88 				      struct efx_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 efx_rx_config_page_split(struct efx_nic *efx)
96 {
97 	efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align,
98 				      EFX_RX_BUF_ALIGNMENT);
99 	efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 :
100 		((PAGE_SIZE - sizeof(struct efx_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(EFX_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 *efx_reuse_page(struct efx_rx_queue *rx_queue)
110 {
111 	struct efx_nic *efx = rx_queue->efx;
112 	struct page *page;
113 	struct efx_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  * efx_init_rx_buffers - create EFX_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 efx_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 efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic)
153 {
154 	struct efx_nic *efx = rx_queue->efx;
155 	struct efx_rx_buffer *rx_buf;
156 	struct page *page;
157 	unsigned int page_offset;
158 	struct efx_rx_page_state *state;
159 	dma_addr_t dma_addr;
160 	unsigned index, count;
161 
162 	count = 0;
163 	do {
164 		page = efx_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 efx_rx_page_state);
188 		page_offset = sizeof(struct efx_rx_page_state);
189 
190 		do {
191 			index = rx_queue->added_count & rx_queue->ptr_mask;
192 			rx_buf = efx_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 = EFX_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 efx_unmap_rx_buffer(struct efx_nic *efx,
214 				struct efx_rx_buffer *rx_buf)
215 {
216 	struct page *page = rx_buf->page;
217 
218 	if (page) {
219 		struct efx_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 efx_free_rx_buffers(struct efx_rx_queue *rx_queue,
228 				struct efx_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 = efx_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 efx_recycle_rx_page(struct efx_channel *channel,
245 				struct efx_rx_buffer *rx_buf)
246 {
247 	struct page *page = rx_buf->page;
248 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
249 	struct efx_nic *efx = rx_queue->efx;
250 	unsigned index;
251 
252 	/* Only recycle the page after processing the final buffer. */
253 	if (!(rx_buf->flags & EFX_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 	efx_unmap_rx_buffer(efx, rx_buf);
273 	put_page(rx_buf->page);
274 }
275 
276 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
277 			       struct efx_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 & EFX_RX_BUF_LAST_IN_PAGE) {
285 		efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
286 		efx_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 efx_recycle_rx_pages(struct efx_channel *channel,
293 				 struct efx_rx_buffer *rx_buf,
294 				 unsigned int n_frags)
295 {
296 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
297 
298 	do {
299 		efx_recycle_rx_page(channel, rx_buf);
300 		rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
301 	} while (--n_frags);
302 }
303 
304 static void efx_discard_rx_packet(struct efx_channel *channel,
305 				  struct efx_rx_buffer *rx_buf,
306 				  unsigned int n_frags)
307 {
308 	struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
309 
310 	efx_recycle_rx_pages(channel, rx_buf, n_frags);
311 
312 	efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
313 }
314 
315 /**
316  * efx_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 efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic)
328 {
329 	struct efx_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 	EFX_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 	EFX_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 		   efx_rx_queue_index(rx_queue), fill_level,
356 		   rx_queue->max_fill);
357 
358 
359 	do {
360 		rc = efx_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 				efx_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", efx_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 		efx_nic_notify_rx_desc(rx_queue);
377 }
378 
379 void efx_rx_slow_fill(unsigned long context)
380 {
381 	struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
382 
383 	/* Post an event to cause NAPI to run and refill the queue */
384 	efx_nic_generate_fill_event(rx_queue);
385 	++rx_queue->slow_fill_count;
386 }
387 
388 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
389 				     struct efx_rx_buffer *rx_buf,
390 				     int len)
391 {
392 	struct efx_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 |= EFX_RX_PKT_DISCARD;
402 
403 	if ((len > rx_buf->len) && EFX_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 				  efx_rx_queue_index(rx_queue), len, max_len,
409 				  efx->type->rx_buffer_padding);
410 		efx_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 				  efx_rx_queue_index(rx_queue), len, max_len);
417 	}
418 
419 	efx_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 efx_rx_packet_gro(struct efx_channel *channel, struct efx_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 efx_nic *efx = channel->efx;
432 	struct sk_buff *skb;
433 
434 	skb = napi_get_frags(napi);
435 	if (unlikely(!skb)) {
436 		struct efx_rx_queue *rx_queue;
437 
438 		rx_queue = efx_channel_get_rx_queue(channel);
439 		efx_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, efx_rx_buf_hash(efx, eh),
445 			     PKT_HASH_TYPE_L3);
446 	skb->ip_summed = ((rx_buf->flags & EFX_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 = efx_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 *efx_rx_mk_skb(struct efx_channel *channel,
473 				     struct efx_rx_buffer *rx_buf,
474 				     unsigned int n_frags,
475 				     u8 *eh, int hdr_len)
476 {
477 	struct efx_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 	EFX_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 = efx_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 efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
530 		   unsigned int n_frags, unsigned int len, u16 flags)
531 {
532 	struct efx_nic *efx = rx_queue->efx;
533 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
534 	struct efx_rx_buffer *rx_buf;
535 
536 	rx_queue->rx_packets++;
537 
538 	rx_buf = efx_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 & EFX_RX_PKT_PREFIX_LEN))
544 			efx_rx_packet__check_len(rx_queue, rx_buf, len);
545 	} else if (unlikely(n_frags > EFX_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 & EFX_RX_PKT_DISCARD));
553 		rx_buf->flags |= EFX_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 		   efx_rx_queue_index(rx_queue), index,
559 		   (index + n_frags - 1) & rx_queue->ptr_mask, len,
560 		   (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
561 		   (rx_buf->flags & EFX_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 & EFX_RX_PKT_DISCARD)) {
567 		efx_rx_flush_packet(channel);
568 		efx_discard_rx_packet(channel, rx_buf, n_frags);
569 		return;
570 	}
571 
572 	if (n_frags == 1 && !(flags & EFX_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 	efx_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(efx_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 = efx_rx_buf_next(rx_queue, rx_buf);
596 			if (--tail_frags == 0)
597 				break;
598 			efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
599 		}
600 		rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
601 		efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
602 	}
603 
604 	/* All fragments have been DMA-synced, so recycle pages. */
605 	rx_buf = efx_rx_buffer(rx_queue, index);
606 	efx_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 	efx_rx_flush_packet(channel);
612 	channel->rx_pkt_n_frags = n_frags;
613 	channel->rx_pkt_index = index;
614 }
615 
616 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
617 			   struct efx_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, EFX_SKB_HEADERS);
622 
623 	skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
624 	if (unlikely(skb == NULL)) {
625 		struct efx_rx_queue *rx_queue;
626 
627 		rx_queue = efx_channel_get_rx_queue(channel);
628 		efx_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 & EFX_RX_PKT_CSUMMED))
636 		skb->ip_summed = CHECKSUM_UNNECESSARY;
637 
638 	efx_rx_skb_attach_timestamp(channel, skb);
639 
640 	if (channel->type->receive_skb)
641 		if (channel->type->receive_skb(channel, skb))
642 			return;
643 
644 	/* Pass the packet up */
645 	netif_receive_skb(skb);
646 }
647 
648 /* Handle a received packet.  Second half: Touches packet payload. */
649 void __efx_rx_packet(struct efx_channel *channel)
650 {
651 	struct efx_nic *efx = channel->efx;
652 	struct efx_rx_buffer *rx_buf =
653 		efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
654 	u8 *eh = efx_rx_buf_va(rx_buf);
655 
656 	/* Read length from the prefix if necessary.  This already
657 	 * excludes the length of the prefix itself.
658 	 */
659 	if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
660 		rx_buf->len = le16_to_cpup((__le16 *)
661 					   (eh + efx->rx_packet_len_offset));
662 
663 	/* If we're in loopback test, then pass the packet directly to the
664 	 * loopback layer, and free the rx_buf here
665 	 */
666 	if (unlikely(efx->loopback_selftest)) {
667 		struct efx_rx_queue *rx_queue;
668 
669 		efx_loopback_rx_packet(efx, eh, rx_buf->len);
670 		rx_queue = efx_channel_get_rx_queue(channel);
671 		efx_free_rx_buffers(rx_queue, rx_buf,
672 				    channel->rx_pkt_n_frags);
673 		goto out;
674 	}
675 
676 	if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
677 		rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
678 
679 	if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb &&
680 	    !efx_channel_busy_polling(channel))
681 		efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
682 	else
683 		efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
684 out:
685 	channel->rx_pkt_n_frags = 0;
686 }
687 
688 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
689 {
690 	struct efx_nic *efx = rx_queue->efx;
691 	unsigned int entries;
692 	int rc;
693 
694 	/* Create the smallest power-of-two aligned ring */
695 	entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
696 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
697 	rx_queue->ptr_mask = entries - 1;
698 
699 	netif_dbg(efx, probe, efx->net_dev,
700 		  "creating RX queue %d size %#x mask %#x\n",
701 		  efx_rx_queue_index(rx_queue), efx->rxq_entries,
702 		  rx_queue->ptr_mask);
703 
704 	/* Allocate RX buffers */
705 	rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
706 				   GFP_KERNEL);
707 	if (!rx_queue->buffer)
708 		return -ENOMEM;
709 
710 	rc = efx_nic_probe_rx(rx_queue);
711 	if (rc) {
712 		kfree(rx_queue->buffer);
713 		rx_queue->buffer = NULL;
714 	}
715 
716 	return rc;
717 }
718 
719 static void efx_init_rx_recycle_ring(struct efx_nic *efx,
720 				     struct efx_rx_queue *rx_queue)
721 {
722 	unsigned int bufs_in_recycle_ring, page_ring_size;
723 
724 	/* Set the RX recycle ring size */
725 #ifdef CONFIG_PPC64
726 	bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
727 #else
728 	if (iommu_present(&pci_bus_type))
729 		bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
730 	else
731 		bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
732 #endif /* CONFIG_PPC64 */
733 
734 	page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
735 					    efx->rx_bufs_per_page);
736 	rx_queue->page_ring = kcalloc(page_ring_size,
737 				      sizeof(*rx_queue->page_ring), GFP_KERNEL);
738 	rx_queue->page_ptr_mask = page_ring_size - 1;
739 }
740 
741 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
742 {
743 	struct efx_nic *efx = rx_queue->efx;
744 	unsigned int max_fill, trigger, max_trigger;
745 
746 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
747 		  "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
748 
749 	/* Initialise ptr fields */
750 	rx_queue->added_count = 0;
751 	rx_queue->notified_count = 0;
752 	rx_queue->removed_count = 0;
753 	rx_queue->min_fill = -1U;
754 	efx_init_rx_recycle_ring(efx, rx_queue);
755 
756 	rx_queue->page_remove = 0;
757 	rx_queue->page_add = rx_queue->page_ptr_mask + 1;
758 	rx_queue->page_recycle_count = 0;
759 	rx_queue->page_recycle_failed = 0;
760 	rx_queue->page_recycle_full = 0;
761 
762 	/* Initialise limit fields */
763 	max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
764 	max_trigger =
765 		max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
766 	if (rx_refill_threshold != 0) {
767 		trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
768 		if (trigger > max_trigger)
769 			trigger = max_trigger;
770 	} else {
771 		trigger = max_trigger;
772 	}
773 
774 	rx_queue->max_fill = max_fill;
775 	rx_queue->fast_fill_trigger = trigger;
776 	rx_queue->refill_enabled = true;
777 
778 	/* Set up RX descriptor ring */
779 	efx_nic_init_rx(rx_queue);
780 }
781 
782 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
783 {
784 	int i;
785 	struct efx_nic *efx = rx_queue->efx;
786 	struct efx_rx_buffer *rx_buf;
787 
788 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
789 		  "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
790 
791 	del_timer_sync(&rx_queue->slow_fill);
792 
793 	/* Release RX buffers from the current read ptr to the write ptr */
794 	if (rx_queue->buffer) {
795 		for (i = rx_queue->removed_count; i < rx_queue->added_count;
796 		     i++) {
797 			unsigned index = i & rx_queue->ptr_mask;
798 			rx_buf = efx_rx_buffer(rx_queue, index);
799 			efx_fini_rx_buffer(rx_queue, rx_buf);
800 		}
801 	}
802 
803 	/* Unmap and release the pages in the recycle ring. Remove the ring. */
804 	for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
805 		struct page *page = rx_queue->page_ring[i];
806 		struct efx_rx_page_state *state;
807 
808 		if (page == NULL)
809 			continue;
810 
811 		state = page_address(page);
812 		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
813 			       PAGE_SIZE << efx->rx_buffer_order,
814 			       DMA_FROM_DEVICE);
815 		put_page(page);
816 	}
817 	kfree(rx_queue->page_ring);
818 	rx_queue->page_ring = NULL;
819 }
820 
821 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
822 {
823 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
824 		  "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
825 
826 	efx_nic_remove_rx(rx_queue);
827 
828 	kfree(rx_queue->buffer);
829 	rx_queue->buffer = NULL;
830 }
831 
832 
833 module_param(rx_refill_threshold, uint, 0444);
834 MODULE_PARM_DESC(rx_refill_threshold,
835 		 "RX descriptor ring refill threshold (%)");
836 
837 #ifdef CONFIG_RFS_ACCEL
838 
839 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
840 		   u16 rxq_index, u32 flow_id)
841 {
842 	struct efx_nic *efx = netdev_priv(net_dev);
843 	struct efx_channel *channel;
844 	struct efx_filter_spec spec;
845 	const __be16 *ports;
846 	__be16 ether_type;
847 	int nhoff;
848 	int rc;
849 
850 	/* The core RPS/RFS code has already parsed and validated
851 	 * VLAN, IP and transport headers.  We assume they are in the
852 	 * header area.
853 	 */
854 
855 	if (skb->protocol == htons(ETH_P_8021Q)) {
856 		const struct vlan_hdr *vh =
857 			(const struct vlan_hdr *)skb->data;
858 
859 		/* We can't filter on the IP 5-tuple and the vlan
860 		 * together, so just strip the vlan header and filter
861 		 * on the IP part.
862 		 */
863 		EFX_BUG_ON_PARANOID(skb_headlen(skb) < sizeof(*vh));
864 		ether_type = vh->h_vlan_encapsulated_proto;
865 		nhoff = sizeof(struct vlan_hdr);
866 	} else {
867 		ether_type = skb->protocol;
868 		nhoff = 0;
869 	}
870 
871 	if (ether_type != htons(ETH_P_IP) && ether_type != htons(ETH_P_IPV6))
872 		return -EPROTONOSUPPORT;
873 
874 	efx_filter_init_rx(&spec, EFX_FILTER_PRI_HINT,
875 			   efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
876 			   rxq_index);
877 	spec.match_flags =
878 		EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
879 		EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
880 		EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
881 	spec.ether_type = ether_type;
882 
883 	if (ether_type == htons(ETH_P_IP)) {
884 		const struct iphdr *ip =
885 			(const struct iphdr *)(skb->data + nhoff);
886 
887 		EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + sizeof(*ip));
888 		if (ip_is_fragment(ip))
889 			return -EPROTONOSUPPORT;
890 		spec.ip_proto = ip->protocol;
891 		spec.rem_host[0] = ip->saddr;
892 		spec.loc_host[0] = ip->daddr;
893 		EFX_BUG_ON_PARANOID(skb_headlen(skb) < nhoff + 4 * ip->ihl + 4);
894 		ports = (const __be16 *)(skb->data + nhoff + 4 * ip->ihl);
895 	} else {
896 		const struct ipv6hdr *ip6 =
897 			(const struct ipv6hdr *)(skb->data + nhoff);
898 
899 		EFX_BUG_ON_PARANOID(skb_headlen(skb) <
900 				    nhoff + sizeof(*ip6) + 4);
901 		spec.ip_proto = ip6->nexthdr;
902 		memcpy(spec.rem_host, &ip6->saddr, sizeof(ip6->saddr));
903 		memcpy(spec.loc_host, &ip6->daddr, sizeof(ip6->daddr));
904 		ports = (const __be16 *)(ip6 + 1);
905 	}
906 
907 	spec.rem_port = ports[0];
908 	spec.loc_port = ports[1];
909 
910 	rc = efx->type->filter_rfs_insert(efx, &spec);
911 	if (rc < 0)
912 		return rc;
913 
914 	/* Remember this so we can check whether to expire the filter later */
915 	efx->rps_flow_id[rc] = flow_id;
916 	channel = efx_get_channel(efx, skb_get_rx_queue(skb));
917 	++channel->rfs_filters_added;
918 
919 	if (ether_type == htons(ETH_P_IP))
920 		netif_info(efx, rx_status, efx->net_dev,
921 			   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n",
922 			   (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
923 			   spec.rem_host, ntohs(ports[0]), spec.loc_host,
924 			   ntohs(ports[1]), rxq_index, flow_id, rc);
925 	else
926 		netif_info(efx, rx_status, efx->net_dev,
927 			   "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d]\n",
928 			   (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
929 			   spec.rem_host, ntohs(ports[0]), spec.loc_host,
930 			   ntohs(ports[1]), rxq_index, flow_id, rc);
931 
932 	return rc;
933 }
934 
935 bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned int quota)
936 {
937 	bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
938 	unsigned int index, size;
939 	u32 flow_id;
940 
941 	if (!spin_trylock_bh(&efx->filter_lock))
942 		return false;
943 
944 	expire_one = efx->type->filter_rfs_expire_one;
945 	index = efx->rps_expire_index;
946 	size = efx->type->max_rx_ip_filters;
947 	while (quota--) {
948 		flow_id = efx->rps_flow_id[index];
949 		if (expire_one(efx, flow_id, index))
950 			netif_info(efx, rx_status, efx->net_dev,
951 				   "expired filter %d [flow %u]\n",
952 				   index, flow_id);
953 		if (++index == size)
954 			index = 0;
955 	}
956 	efx->rps_expire_index = index;
957 
958 	spin_unlock_bh(&efx->filter_lock);
959 	return true;
960 }
961 
962 #endif /* CONFIG_RFS_ACCEL */
963 
964 /**
965  * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
966  * @spec: Specification to test
967  *
968  * Return: %true if the specification is a non-drop RX filter that
969  * matches a local MAC address I/G bit value of 1 or matches a local
970  * IPv4 or IPv6 address value in the respective multicast address
971  * range.  Otherwise %false.
972  */
973 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
974 {
975 	if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
976 	    spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
977 		return false;
978 
979 	if (spec->match_flags &
980 	    (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
981 	    is_multicast_ether_addr(spec->loc_mac))
982 		return true;
983 
984 	if ((spec->match_flags &
985 	     (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
986 	    (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
987 		if (spec->ether_type == htons(ETH_P_IP) &&
988 		    ipv4_is_multicast(spec->loc_host[0]))
989 			return true;
990 		if (spec->ether_type == htons(ETH_P_IPV6) &&
991 		    ((const u8 *)spec->loc_host)[0] == 0xff)
992 			return true;
993 	}
994 
995 	return false;
996 }
997