xref: /openbmc/linux/drivers/net/ethernet/sfc/rx.c (revision 8730046c)
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_WARN_ON_ONCE_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_WARN_ON_ONCE_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 (net_ratelimit())
404 		netif_err(efx, rx_err, efx->net_dev,
405 			  "RX queue %d overlength RX event (%#x > %#x)\n",
406 			  efx_rx_queue_index(rx_queue), len, max_len);
407 
408 	efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
409 }
410 
411 /* Pass a received packet up through GRO.  GRO can handle pages
412  * regardless of checksum state and skbs with a good checksum.
413  */
414 static void
415 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf,
416 		  unsigned int n_frags, u8 *eh)
417 {
418 	struct napi_struct *napi = &channel->napi_str;
419 	gro_result_t gro_result;
420 	struct efx_nic *efx = channel->efx;
421 	struct sk_buff *skb;
422 
423 	skb = napi_get_frags(napi);
424 	if (unlikely(!skb)) {
425 		struct efx_rx_queue *rx_queue;
426 
427 		rx_queue = efx_channel_get_rx_queue(channel);
428 		efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
429 		return;
430 	}
431 
432 	if (efx->net_dev->features & NETIF_F_RXHASH)
433 		skb_set_hash(skb, efx_rx_buf_hash(efx, eh),
434 			     PKT_HASH_TYPE_L3);
435 	skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
436 			  CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
437 
438 	for (;;) {
439 		skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
440 				   rx_buf->page, rx_buf->page_offset,
441 				   rx_buf->len);
442 		rx_buf->page = NULL;
443 		skb->len += rx_buf->len;
444 		if (skb_shinfo(skb)->nr_frags == n_frags)
445 			break;
446 
447 		rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
448 	}
449 
450 	skb->data_len = skb->len;
451 	skb->truesize += n_frags * efx->rx_buffer_truesize;
452 
453 	skb_record_rx_queue(skb, channel->rx_queue.core_index);
454 
455 	gro_result = napi_gro_frags(napi);
456 	if (gro_result != GRO_DROP)
457 		channel->irq_mod_score += 2;
458 }
459 
460 /* Allocate and construct an SKB around page fragments */
461 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
462 				     struct efx_rx_buffer *rx_buf,
463 				     unsigned int n_frags,
464 				     u8 *eh, int hdr_len)
465 {
466 	struct efx_nic *efx = channel->efx;
467 	struct sk_buff *skb;
468 
469 	/* Allocate an SKB to store the headers */
470 	skb = netdev_alloc_skb(efx->net_dev,
471 			       efx->rx_ip_align + efx->rx_prefix_size +
472 			       hdr_len);
473 	if (unlikely(skb == NULL)) {
474 		atomic_inc(&efx->n_rx_noskb_drops);
475 		return NULL;
476 	}
477 
478 	EFX_WARN_ON_ONCE_PARANOID(rx_buf->len < hdr_len);
479 
480 	memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size,
481 	       efx->rx_prefix_size + hdr_len);
482 	skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size);
483 	__skb_put(skb, hdr_len);
484 
485 	/* Append the remaining page(s) onto the frag list */
486 	if (rx_buf->len > hdr_len) {
487 		rx_buf->page_offset += hdr_len;
488 		rx_buf->len -= hdr_len;
489 
490 		for (;;) {
491 			skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags,
492 					   rx_buf->page, rx_buf->page_offset,
493 					   rx_buf->len);
494 			rx_buf->page = NULL;
495 			skb->len += rx_buf->len;
496 			skb->data_len += rx_buf->len;
497 			if (skb_shinfo(skb)->nr_frags == n_frags)
498 				break;
499 
500 			rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf);
501 		}
502 	} else {
503 		__free_pages(rx_buf->page, efx->rx_buffer_order);
504 		rx_buf->page = NULL;
505 		n_frags = 0;
506 	}
507 
508 	skb->truesize += n_frags * efx->rx_buffer_truesize;
509 
510 	/* Move past the ethernet header */
511 	skb->protocol = eth_type_trans(skb, efx->net_dev);
512 
513 	skb_mark_napi_id(skb, &channel->napi_str);
514 
515 	return skb;
516 }
517 
518 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
519 		   unsigned int n_frags, unsigned int len, u16 flags)
520 {
521 	struct efx_nic *efx = rx_queue->efx;
522 	struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
523 	struct efx_rx_buffer *rx_buf;
524 
525 	rx_queue->rx_packets++;
526 
527 	rx_buf = efx_rx_buffer(rx_queue, index);
528 	rx_buf->flags |= flags;
529 
530 	/* Validate the number of fragments and completed length */
531 	if (n_frags == 1) {
532 		if (!(flags & EFX_RX_PKT_PREFIX_LEN))
533 			efx_rx_packet__check_len(rx_queue, rx_buf, len);
534 	} else if (unlikely(n_frags > EFX_RX_MAX_FRAGS) ||
535 		   unlikely(len <= (n_frags - 1) * efx->rx_dma_len) ||
536 		   unlikely(len > n_frags * efx->rx_dma_len) ||
537 		   unlikely(!efx->rx_scatter)) {
538 		/* If this isn't an explicit discard request, either
539 		 * the hardware or the driver is broken.
540 		 */
541 		WARN_ON(!(len == 0 && rx_buf->flags & EFX_RX_PKT_DISCARD));
542 		rx_buf->flags |= EFX_RX_PKT_DISCARD;
543 	}
544 
545 	netif_vdbg(efx, rx_status, efx->net_dev,
546 		   "RX queue %d received ids %x-%x len %d %s%s\n",
547 		   efx_rx_queue_index(rx_queue), index,
548 		   (index + n_frags - 1) & rx_queue->ptr_mask, len,
549 		   (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
550 		   (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
551 
552 	/* Discard packet, if instructed to do so.  Process the
553 	 * previous receive first.
554 	 */
555 	if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
556 		efx_rx_flush_packet(channel);
557 		efx_discard_rx_packet(channel, rx_buf, n_frags);
558 		return;
559 	}
560 
561 	if (n_frags == 1 && !(flags & EFX_RX_PKT_PREFIX_LEN))
562 		rx_buf->len = len;
563 
564 	/* Release and/or sync the DMA mapping - assumes all RX buffers
565 	 * consumed in-order per RX queue.
566 	 */
567 	efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
568 
569 	/* Prefetch nice and early so data will (hopefully) be in cache by
570 	 * the time we look at it.
571 	 */
572 	prefetch(efx_rx_buf_va(rx_buf));
573 
574 	rx_buf->page_offset += efx->rx_prefix_size;
575 	rx_buf->len -= efx->rx_prefix_size;
576 
577 	if (n_frags > 1) {
578 		/* Release/sync DMA mapping for additional fragments.
579 		 * Fix length for last fragment.
580 		 */
581 		unsigned int tail_frags = n_frags - 1;
582 
583 		for (;;) {
584 			rx_buf = efx_rx_buf_next(rx_queue, rx_buf);
585 			if (--tail_frags == 0)
586 				break;
587 			efx_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len);
588 		}
589 		rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len;
590 		efx_sync_rx_buffer(efx, rx_buf, rx_buf->len);
591 	}
592 
593 	/* All fragments have been DMA-synced, so recycle pages. */
594 	rx_buf = efx_rx_buffer(rx_queue, index);
595 	efx_recycle_rx_pages(channel, rx_buf, n_frags);
596 
597 	/* Pipeline receives so that we give time for packet headers to be
598 	 * prefetched into cache.
599 	 */
600 	efx_rx_flush_packet(channel);
601 	channel->rx_pkt_n_frags = n_frags;
602 	channel->rx_pkt_index = index;
603 }
604 
605 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
606 			   struct efx_rx_buffer *rx_buf,
607 			   unsigned int n_frags)
608 {
609 	struct sk_buff *skb;
610 	u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
611 
612 	skb = efx_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len);
613 	if (unlikely(skb == NULL)) {
614 		struct efx_rx_queue *rx_queue;
615 
616 		rx_queue = efx_channel_get_rx_queue(channel);
617 		efx_free_rx_buffers(rx_queue, rx_buf, n_frags);
618 		return;
619 	}
620 	skb_record_rx_queue(skb, channel->rx_queue.core_index);
621 
622 	/* Set the SKB flags */
623 	skb_checksum_none_assert(skb);
624 	if (likely(rx_buf->flags & EFX_RX_PKT_CSUMMED))
625 		skb->ip_summed = CHECKSUM_UNNECESSARY;
626 
627 	efx_rx_skb_attach_timestamp(channel, skb);
628 
629 	if (channel->type->receive_skb)
630 		if (channel->type->receive_skb(channel, skb))
631 			return;
632 
633 	/* Pass the packet up */
634 	netif_receive_skb(skb);
635 }
636 
637 /* Handle a received packet.  Second half: Touches packet payload. */
638 void __efx_rx_packet(struct efx_channel *channel)
639 {
640 	struct efx_nic *efx = channel->efx;
641 	struct efx_rx_buffer *rx_buf =
642 		efx_rx_buffer(&channel->rx_queue, channel->rx_pkt_index);
643 	u8 *eh = efx_rx_buf_va(rx_buf);
644 
645 	/* Read length from the prefix if necessary.  This already
646 	 * excludes the length of the prefix itself.
647 	 */
648 	if (rx_buf->flags & EFX_RX_PKT_PREFIX_LEN)
649 		rx_buf->len = le16_to_cpup((__le16 *)
650 					   (eh + efx->rx_packet_len_offset));
651 
652 	/* If we're in loopback test, then pass the packet directly to the
653 	 * loopback layer, and free the rx_buf here
654 	 */
655 	if (unlikely(efx->loopback_selftest)) {
656 		struct efx_rx_queue *rx_queue;
657 
658 		efx_loopback_rx_packet(efx, eh, rx_buf->len);
659 		rx_queue = efx_channel_get_rx_queue(channel);
660 		efx_free_rx_buffers(rx_queue, rx_buf,
661 				    channel->rx_pkt_n_frags);
662 		goto out;
663 	}
664 
665 	if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
666 		rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
667 
668 	if ((rx_buf->flags & EFX_RX_PKT_TCP) && !channel->type->receive_skb &&
669 	    !efx_channel_busy_polling(channel))
670 		efx_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh);
671 	else
672 		efx_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags);
673 out:
674 	channel->rx_pkt_n_frags = 0;
675 }
676 
677 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
678 {
679 	struct efx_nic *efx = rx_queue->efx;
680 	unsigned int entries;
681 	int rc;
682 
683 	/* Create the smallest power-of-two aligned ring */
684 	entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
685 	EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
686 	rx_queue->ptr_mask = entries - 1;
687 
688 	netif_dbg(efx, probe, efx->net_dev,
689 		  "creating RX queue %d size %#x mask %#x\n",
690 		  efx_rx_queue_index(rx_queue), efx->rxq_entries,
691 		  rx_queue->ptr_mask);
692 
693 	/* Allocate RX buffers */
694 	rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
695 				   GFP_KERNEL);
696 	if (!rx_queue->buffer)
697 		return -ENOMEM;
698 
699 	rc = efx_nic_probe_rx(rx_queue);
700 	if (rc) {
701 		kfree(rx_queue->buffer);
702 		rx_queue->buffer = NULL;
703 	}
704 
705 	return rc;
706 }
707 
708 static void efx_init_rx_recycle_ring(struct efx_nic *efx,
709 				     struct efx_rx_queue *rx_queue)
710 {
711 	unsigned int bufs_in_recycle_ring, page_ring_size;
712 
713 	/* Set the RX recycle ring size */
714 #ifdef CONFIG_PPC64
715 	bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
716 #else
717 	if (iommu_present(&pci_bus_type))
718 		bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU;
719 	else
720 		bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU;
721 #endif /* CONFIG_PPC64 */
722 
723 	page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring /
724 					    efx->rx_bufs_per_page);
725 	rx_queue->page_ring = kcalloc(page_ring_size,
726 				      sizeof(*rx_queue->page_ring), GFP_KERNEL);
727 	rx_queue->page_ptr_mask = page_ring_size - 1;
728 }
729 
730 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
731 {
732 	struct efx_nic *efx = rx_queue->efx;
733 	unsigned int max_fill, trigger, max_trigger;
734 
735 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
736 		  "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
737 
738 	/* Initialise ptr fields */
739 	rx_queue->added_count = 0;
740 	rx_queue->notified_count = 0;
741 	rx_queue->removed_count = 0;
742 	rx_queue->min_fill = -1U;
743 	efx_init_rx_recycle_ring(efx, rx_queue);
744 
745 	rx_queue->page_remove = 0;
746 	rx_queue->page_add = rx_queue->page_ptr_mask + 1;
747 	rx_queue->page_recycle_count = 0;
748 	rx_queue->page_recycle_failed = 0;
749 	rx_queue->page_recycle_full = 0;
750 
751 	/* Initialise limit fields */
752 	max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
753 	max_trigger =
754 		max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page;
755 	if (rx_refill_threshold != 0) {
756 		trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
757 		if (trigger > max_trigger)
758 			trigger = max_trigger;
759 	} else {
760 		trigger = max_trigger;
761 	}
762 
763 	rx_queue->max_fill = max_fill;
764 	rx_queue->fast_fill_trigger = trigger;
765 	rx_queue->refill_enabled = true;
766 
767 	/* Set up RX descriptor ring */
768 	efx_nic_init_rx(rx_queue);
769 }
770 
771 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
772 {
773 	int i;
774 	struct efx_nic *efx = rx_queue->efx;
775 	struct efx_rx_buffer *rx_buf;
776 
777 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
778 		  "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
779 
780 	del_timer_sync(&rx_queue->slow_fill);
781 
782 	/* Release RX buffers from the current read ptr to the write ptr */
783 	if (rx_queue->buffer) {
784 		for (i = rx_queue->removed_count; i < rx_queue->added_count;
785 		     i++) {
786 			unsigned index = i & rx_queue->ptr_mask;
787 			rx_buf = efx_rx_buffer(rx_queue, index);
788 			efx_fini_rx_buffer(rx_queue, rx_buf);
789 		}
790 	}
791 
792 	/* Unmap and release the pages in the recycle ring. Remove the ring. */
793 	for (i = 0; i <= rx_queue->page_ptr_mask; i++) {
794 		struct page *page = rx_queue->page_ring[i];
795 		struct efx_rx_page_state *state;
796 
797 		if (page == NULL)
798 			continue;
799 
800 		state = page_address(page);
801 		dma_unmap_page(&efx->pci_dev->dev, state->dma_addr,
802 			       PAGE_SIZE << efx->rx_buffer_order,
803 			       DMA_FROM_DEVICE);
804 		put_page(page);
805 	}
806 	kfree(rx_queue->page_ring);
807 	rx_queue->page_ring = NULL;
808 }
809 
810 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
811 {
812 	netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
813 		  "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
814 
815 	efx_nic_remove_rx(rx_queue);
816 
817 	kfree(rx_queue->buffer);
818 	rx_queue->buffer = NULL;
819 }
820 
821 
822 module_param(rx_refill_threshold, uint, 0444);
823 MODULE_PARM_DESC(rx_refill_threshold,
824 		 "RX descriptor ring refill threshold (%)");
825 
826 #ifdef CONFIG_RFS_ACCEL
827 
828 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
829 		   u16 rxq_index, u32 flow_id)
830 {
831 	struct efx_nic *efx = netdev_priv(net_dev);
832 	struct efx_channel *channel;
833 	struct efx_filter_spec spec;
834 	struct flow_keys fk;
835 	int rc;
836 
837 	if (flow_id == RPS_FLOW_ID_INVALID)
838 		return -EINVAL;
839 
840 	if (!skb_flow_dissect_flow_keys(skb, &fk, 0))
841 		return -EPROTONOSUPPORT;
842 
843 	if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6))
844 		return -EPROTONOSUPPORT;
845 	if (fk.control.flags & FLOW_DIS_IS_FRAGMENT)
846 		return -EPROTONOSUPPORT;
847 
848 	efx_filter_init_rx(&spec, EFX_FILTER_PRI_HINT,
849 			   efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
850 			   rxq_index);
851 	spec.match_flags =
852 		EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
853 		EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
854 		EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
855 	spec.ether_type = fk.basic.n_proto;
856 	spec.ip_proto = fk.basic.ip_proto;
857 
858 	if (fk.basic.n_proto == htons(ETH_P_IP)) {
859 		spec.rem_host[0] = fk.addrs.v4addrs.src;
860 		spec.loc_host[0] = fk.addrs.v4addrs.dst;
861 	} else {
862 		memcpy(spec.rem_host, &fk.addrs.v6addrs.src, sizeof(struct in6_addr));
863 		memcpy(spec.loc_host, &fk.addrs.v6addrs.dst, sizeof(struct in6_addr));
864 	}
865 
866 	spec.rem_port = fk.ports.src;
867 	spec.loc_port = fk.ports.dst;
868 
869 	rc = efx->type->filter_rfs_insert(efx, &spec);
870 	if (rc < 0)
871 		return rc;
872 
873 	/* Remember this so we can check whether to expire the filter later */
874 	channel = efx_get_channel(efx, rxq_index);
875 	channel->rps_flow_id[rc] = flow_id;
876 	++channel->rfs_filters_added;
877 
878 	if (spec.ether_type == htons(ETH_P_IP))
879 		netif_info(efx, rx_status, efx->net_dev,
880 			   "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n",
881 			   (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
882 			   spec.rem_host, ntohs(spec.rem_port), spec.loc_host,
883 			   ntohs(spec.loc_port), rxq_index, flow_id, rc);
884 	else
885 		netif_info(efx, rx_status, efx->net_dev,
886 			   "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d]\n",
887 			   (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
888 			   spec.rem_host, ntohs(spec.rem_port), spec.loc_host,
889 			   ntohs(spec.loc_port), rxq_index, flow_id, rc);
890 
891 	return rc;
892 }
893 
894 bool __efx_filter_rfs_expire(struct efx_nic *efx, unsigned int quota)
895 {
896 	bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
897 	unsigned int channel_idx, index, size;
898 	u32 flow_id;
899 
900 	if (!spin_trylock_bh(&efx->filter_lock))
901 		return false;
902 
903 	expire_one = efx->type->filter_rfs_expire_one;
904 	channel_idx = efx->rps_expire_channel;
905 	index = efx->rps_expire_index;
906 	size = efx->type->max_rx_ip_filters;
907 	while (quota--) {
908 		struct efx_channel *channel = efx_get_channel(efx, channel_idx);
909 		flow_id = channel->rps_flow_id[index];
910 
911 		if (flow_id != RPS_FLOW_ID_INVALID &&
912 		    expire_one(efx, flow_id, index)) {
913 			netif_info(efx, rx_status, efx->net_dev,
914 				   "expired filter %d [queue %u flow %u]\n",
915 				   index, channel_idx, flow_id);
916 			channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
917 		}
918 		if (++index == size) {
919 			if (++channel_idx == efx->n_channels)
920 				channel_idx = 0;
921 			index = 0;
922 		}
923 	}
924 	efx->rps_expire_channel = channel_idx;
925 	efx->rps_expire_index = index;
926 
927 	spin_unlock_bh(&efx->filter_lock);
928 	return true;
929 }
930 
931 #endif /* CONFIG_RFS_ACCEL */
932 
933 /**
934  * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
935  * @spec: Specification to test
936  *
937  * Return: %true if the specification is a non-drop RX filter that
938  * matches a local MAC address I/G bit value of 1 or matches a local
939  * IPv4 or IPv6 address value in the respective multicast address
940  * range.  Otherwise %false.
941  */
942 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
943 {
944 	if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
945 	    spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
946 		return false;
947 
948 	if (spec->match_flags &
949 	    (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
950 	    is_multicast_ether_addr(spec->loc_mac))
951 		return true;
952 
953 	if ((spec->match_flags &
954 	     (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
955 	    (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
956 		if (spec->ether_type == htons(ETH_P_IP) &&
957 		    ipv4_is_multicast(spec->loc_host[0]))
958 			return true;
959 		if (spec->ether_type == htons(ETH_P_IPV6) &&
960 		    ((const u8 *)spec->loc_host)[0] == 0xff)
961 			return true;
962 	}
963 
964 	return false;
965 }
966