sfc: Correct reporting and validation of TX interrupt coalescing
[pandora-kernel.git] / drivers / net / ethernet / sfc / rx.c
1 /****************************************************************************
2  * Driver for Solarflare Solarstorm network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2011 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/tcp.h>
16 #include <linux/udp.h>
17 #include <linux/prefetch.h>
18 #include <net/ip.h>
19 #include <net/checksum.h>
20 #include "net_driver.h"
21 #include "efx.h"
22 #include "nic.h"
23 #include "selftest.h"
24 #include "workarounds.h"
25
26 /* Number of RX descriptors pushed at once. */
27 #define EFX_RX_BATCH  8
28
29 /* Maximum size of a buffer sharing a page */
30 #define EFX_RX_HALF_PAGE ((PAGE_SIZE >> 1) - sizeof(struct efx_rx_page_state))
31
32 /* Size of buffer allocated for skb header area. */
33 #define EFX_SKB_HEADERS  64u
34
35 /*
36  * rx_alloc_method - RX buffer allocation method
37  *
38  * This driver supports two methods for allocating and using RX buffers:
39  * each RX buffer may be backed by an skb or by an order-n page.
40  *
41  * When GRO is in use then the second method has a lower overhead,
42  * since we don't have to allocate then free skbs on reassembled frames.
43  *
44  * Values:
45  *   - RX_ALLOC_METHOD_AUTO = 0
46  *   - RX_ALLOC_METHOD_SKB  = 1
47  *   - RX_ALLOC_METHOD_PAGE = 2
48  *
49  * The heuristic for %RX_ALLOC_METHOD_AUTO is a simple hysteresis count
50  * controlled by the parameters below.
51  *
52  *   - Since pushing and popping descriptors are separated by the rx_queue
53  *     size, so the watermarks should be ~rxd_size.
54  *   - The performance win by using page-based allocation for GRO is less
55  *     than the performance hit of using page-based allocation of non-GRO,
56  *     so the watermarks should reflect this.
57  *
58  * Per channel we maintain a single variable, updated by each channel:
59  *
60  *   rx_alloc_level += (gro_performed ? RX_ALLOC_FACTOR_GRO :
61  *                      RX_ALLOC_FACTOR_SKB)
62  * Per NAPI poll interval, we constrain rx_alloc_level to 0..MAX (which
63  * limits the hysteresis), and update the allocation strategy:
64  *
65  *   rx_alloc_method = (rx_alloc_level > RX_ALLOC_LEVEL_GRO ?
66  *                      RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB)
67  */
68 static int rx_alloc_method = RX_ALLOC_METHOD_AUTO;
69
70 #define RX_ALLOC_LEVEL_GRO 0x2000
71 #define RX_ALLOC_LEVEL_MAX 0x3000
72 #define RX_ALLOC_FACTOR_GRO 1
73 #define RX_ALLOC_FACTOR_SKB (-2)
74
75 /* This is the percentage fill level below which new RX descriptors
76  * will be added to the RX descriptor ring.
77  */
78 static unsigned int rx_refill_threshold = 90;
79
80 /* This is the percentage fill level to which an RX queue will be refilled
81  * when the "RX refill threshold" is reached.
82  */
83 static unsigned int rx_refill_limit = 95;
84
85 /*
86  * RX maximum head room required.
87  *
88  * This must be at least 1 to prevent overflow and at least 2 to allow
89  * pipelined receives.
90  */
91 #define EFX_RXD_HEAD_ROOM 2
92
93 /* Offset of ethernet header within page */
94 static inline unsigned int efx_rx_buf_offset(struct efx_nic *efx,
95                                              struct efx_rx_buffer *buf)
96 {
97         /* Offset is always within one page, so we don't need to consider
98          * the page order.
99          */
100         return (((__force unsigned long) buf->dma_addr & (PAGE_SIZE - 1)) +
101                 efx->type->rx_buffer_hash_size);
102 }
103 static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
104 {
105         return PAGE_SIZE << efx->rx_buffer_order;
106 }
107
108 static u8 *efx_rx_buf_eh(struct efx_nic *efx, struct efx_rx_buffer *buf)
109 {
110         if (buf->is_page)
111                 return page_address(buf->u.page) + efx_rx_buf_offset(efx, buf);
112         else
113                 return ((u8 *)buf->u.skb->data +
114                         efx->type->rx_buffer_hash_size);
115 }
116
117 static inline u32 efx_rx_buf_hash(const u8 *eh)
118 {
119         /* The ethernet header is always directly after any hash. */
120 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) || NET_IP_ALIGN % 4 == 0
121         return __le32_to_cpup((const __le32 *)(eh - 4));
122 #else
123         const u8 *data = eh - 4;
124         return ((u32)data[0]       |
125                 (u32)data[1] << 8  |
126                 (u32)data[2] << 16 |
127                 (u32)data[3] << 24);
128 #endif
129 }
130
131 /**
132  * efx_init_rx_buffers_skb - create EFX_RX_BATCH skb-based RX buffers
133  *
134  * @rx_queue:           Efx RX queue
135  *
136  * This allocates EFX_RX_BATCH skbs, maps them for DMA, and populates a
137  * struct efx_rx_buffer for each one. Return a negative error code or 0
138  * on success. May fail having only inserted fewer than EFX_RX_BATCH
139  * buffers.
140  */
141 static int efx_init_rx_buffers_skb(struct efx_rx_queue *rx_queue)
142 {
143         struct efx_nic *efx = rx_queue->efx;
144         struct net_device *net_dev = efx->net_dev;
145         struct efx_rx_buffer *rx_buf;
146         struct sk_buff *skb;
147         int skb_len = efx->rx_buffer_len;
148         unsigned index, count;
149
150         for (count = 0; count < EFX_RX_BATCH; ++count) {
151                 index = rx_queue->added_count & rx_queue->ptr_mask;
152                 rx_buf = efx_rx_buffer(rx_queue, index);
153
154                 rx_buf->u.skb = skb = netdev_alloc_skb(net_dev, skb_len);
155                 if (unlikely(!skb))
156                         return -ENOMEM;
157
158                 /* Adjust the SKB for padding and checksum */
159                 skb_reserve(skb, NET_IP_ALIGN);
160                 rx_buf->len = skb_len - NET_IP_ALIGN;
161                 rx_buf->is_page = false;
162                 skb->ip_summed = CHECKSUM_UNNECESSARY;
163
164                 rx_buf->dma_addr = pci_map_single(efx->pci_dev,
165                                                   skb->data, rx_buf->len,
166                                                   PCI_DMA_FROMDEVICE);
167                 if (unlikely(pci_dma_mapping_error(efx->pci_dev,
168                                                    rx_buf->dma_addr))) {
169                         dev_kfree_skb_any(skb);
170                         rx_buf->u.skb = NULL;
171                         return -EIO;
172                 }
173
174                 ++rx_queue->added_count;
175                 ++rx_queue->alloc_skb_count;
176         }
177
178         return 0;
179 }
180
181 /**
182  * efx_init_rx_buffers_page - create EFX_RX_BATCH page-based RX buffers
183  *
184  * @rx_queue:           Efx RX queue
185  *
186  * This allocates memory for EFX_RX_BATCH receive buffers, maps them for DMA,
187  * and populates struct efx_rx_buffers for each one. Return a negative error
188  * code or 0 on success. If a single page can be split between two buffers,
189  * then the page will either be inserted fully, or not at at all.
190  */
191 static int efx_init_rx_buffers_page(struct efx_rx_queue *rx_queue)
192 {
193         struct efx_nic *efx = rx_queue->efx;
194         struct efx_rx_buffer *rx_buf;
195         struct page *page;
196         void *page_addr;
197         struct efx_rx_page_state *state;
198         dma_addr_t dma_addr;
199         unsigned index, count;
200
201         /* We can split a page between two buffers */
202         BUILD_BUG_ON(EFX_RX_BATCH & 1);
203
204         for (count = 0; count < EFX_RX_BATCH; ++count) {
205                 page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
206                                    efx->rx_buffer_order);
207                 if (unlikely(page == NULL))
208                         return -ENOMEM;
209                 dma_addr = pci_map_page(efx->pci_dev, page, 0,
210                                         efx_rx_buf_size(efx),
211                                         PCI_DMA_FROMDEVICE);
212                 if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
213                         __free_pages(page, efx->rx_buffer_order);
214                         return -EIO;
215                 }
216                 page_addr = page_address(page);
217                 state = page_addr;
218                 state->refcnt = 0;
219                 state->dma_addr = dma_addr;
220
221                 page_addr += sizeof(struct efx_rx_page_state);
222                 dma_addr += sizeof(struct efx_rx_page_state);
223
224         split:
225                 index = rx_queue->added_count & rx_queue->ptr_mask;
226                 rx_buf = efx_rx_buffer(rx_queue, index);
227                 rx_buf->dma_addr = dma_addr + EFX_PAGE_IP_ALIGN;
228                 rx_buf->u.page = page;
229                 rx_buf->len = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
230                 rx_buf->is_page = true;
231                 ++rx_queue->added_count;
232                 ++rx_queue->alloc_page_count;
233                 ++state->refcnt;
234
235                 if ((~count & 1) && (efx->rx_buffer_len <= EFX_RX_HALF_PAGE)) {
236                         /* Use the second half of the page */
237                         get_page(page);
238                         dma_addr += (PAGE_SIZE >> 1);
239                         page_addr += (PAGE_SIZE >> 1);
240                         ++count;
241                         goto split;
242                 }
243         }
244
245         return 0;
246 }
247
248 static void efx_unmap_rx_buffer(struct efx_nic *efx,
249                                 struct efx_rx_buffer *rx_buf)
250 {
251         if (rx_buf->is_page && rx_buf->u.page) {
252                 struct efx_rx_page_state *state;
253
254                 state = page_address(rx_buf->u.page);
255                 if (--state->refcnt == 0) {
256                         pci_unmap_page(efx->pci_dev,
257                                        state->dma_addr,
258                                        efx_rx_buf_size(efx),
259                                        PCI_DMA_FROMDEVICE);
260                 }
261         } else if (!rx_buf->is_page && rx_buf->u.skb) {
262                 pci_unmap_single(efx->pci_dev, rx_buf->dma_addr,
263                                  rx_buf->len, PCI_DMA_FROMDEVICE);
264         }
265 }
266
267 static void efx_free_rx_buffer(struct efx_nic *efx,
268                                struct efx_rx_buffer *rx_buf)
269 {
270         if (rx_buf->is_page && rx_buf->u.page) {
271                 __free_pages(rx_buf->u.page, efx->rx_buffer_order);
272                 rx_buf->u.page = NULL;
273         } else if (!rx_buf->is_page && rx_buf->u.skb) {
274                 dev_kfree_skb_any(rx_buf->u.skb);
275                 rx_buf->u.skb = NULL;
276         }
277 }
278
279 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
280                                struct efx_rx_buffer *rx_buf)
281 {
282         efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
283         efx_free_rx_buffer(rx_queue->efx, rx_buf);
284 }
285
286 /* Attempt to resurrect the other receive buffer that used to share this page,
287  * which had previously been passed up to the kernel and freed. */
288 static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
289                                     struct efx_rx_buffer *rx_buf)
290 {
291         struct efx_rx_page_state *state = page_address(rx_buf->u.page);
292         struct efx_rx_buffer *new_buf;
293         unsigned fill_level, index;
294
295         /* +1 because efx_rx_packet() incremented removed_count. +1 because
296          * we'd like to insert an additional descriptor whilst leaving
297          * EFX_RXD_HEAD_ROOM for the non-recycle path */
298         fill_level = (rx_queue->added_count - rx_queue->removed_count + 2);
299         if (unlikely(fill_level > rx_queue->max_fill)) {
300                 /* We could place "state" on a list, and drain the list in
301                  * efx_fast_push_rx_descriptors(). For now, this will do. */
302                 return;
303         }
304
305         ++state->refcnt;
306         get_page(rx_buf->u.page);
307
308         index = rx_queue->added_count & rx_queue->ptr_mask;
309         new_buf = efx_rx_buffer(rx_queue, index);
310         new_buf->dma_addr = rx_buf->dma_addr ^ (PAGE_SIZE >> 1);
311         new_buf->u.page = rx_buf->u.page;
312         new_buf->len = rx_buf->len;
313         new_buf->is_page = true;
314         ++rx_queue->added_count;
315 }
316
317 /* Recycle the given rx buffer directly back into the rx_queue. There is
318  * always room to add this buffer, because we've just popped a buffer. */
319 static void efx_recycle_rx_buffer(struct efx_channel *channel,
320                                   struct efx_rx_buffer *rx_buf)
321 {
322         struct efx_nic *efx = channel->efx;
323         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
324         struct efx_rx_buffer *new_buf;
325         unsigned index;
326
327         if (rx_buf->is_page && efx->rx_buffer_len <= EFX_RX_HALF_PAGE &&
328             page_count(rx_buf->u.page) == 1)
329                 efx_resurrect_rx_buffer(rx_queue, rx_buf);
330
331         index = rx_queue->added_count & rx_queue->ptr_mask;
332         new_buf = efx_rx_buffer(rx_queue, index);
333
334         memcpy(new_buf, rx_buf, sizeof(*new_buf));
335         rx_buf->u.page = NULL;
336         ++rx_queue->added_count;
337 }
338
339 /**
340  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
341  * @rx_queue:           RX descriptor queue
342  * This will aim to fill the RX descriptor queue up to
343  * @rx_queue->@fast_fill_limit. If there is insufficient atomic
344  * memory to do so, a slow fill will be scheduled.
345  *
346  * The caller must provide serialisation (none is used here). In practise,
347  * this means this function must run from the NAPI handler, or be called
348  * when NAPI is disabled.
349  */
350 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
351 {
352         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
353         unsigned fill_level;
354         int space, rc = 0;
355
356         /* Calculate current fill level, and exit if we don't need to fill */
357         fill_level = (rx_queue->added_count - rx_queue->removed_count);
358         EFX_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
359         if (fill_level >= rx_queue->fast_fill_trigger)
360                 goto out;
361
362         /* Record minimum fill level */
363         if (unlikely(fill_level < rx_queue->min_fill)) {
364                 if (fill_level)
365                         rx_queue->min_fill = fill_level;
366         }
367
368         space = rx_queue->fast_fill_limit - fill_level;
369         if (space < EFX_RX_BATCH)
370                 goto out;
371
372         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
373                    "RX queue %d fast-filling descriptor ring from"
374                    " level %d to level %d using %s allocation\n",
375                    efx_rx_queue_index(rx_queue), fill_level,
376                    rx_queue->fast_fill_limit,
377                    channel->rx_alloc_push_pages ? "page" : "skb");
378
379         do {
380                 if (channel->rx_alloc_push_pages)
381                         rc = efx_init_rx_buffers_page(rx_queue);
382                 else
383                         rc = efx_init_rx_buffers_skb(rx_queue);
384                 if (unlikely(rc)) {
385                         /* Ensure that we don't leave the rx queue empty */
386                         if (rx_queue->added_count == rx_queue->removed_count)
387                                 efx_schedule_slow_fill(rx_queue);
388                         goto out;
389                 }
390         } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
391
392         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
393                    "RX queue %d fast-filled descriptor ring "
394                    "to level %d\n", efx_rx_queue_index(rx_queue),
395                    rx_queue->added_count - rx_queue->removed_count);
396
397  out:
398         if (rx_queue->notified_count != rx_queue->added_count)
399                 efx_nic_notify_rx_desc(rx_queue);
400 }
401
402 void efx_rx_slow_fill(unsigned long context)
403 {
404         struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
405         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
406
407         /* Post an event to cause NAPI to run and refill the queue */
408         efx_nic_generate_fill_event(channel);
409         ++rx_queue->slow_fill_count;
410 }
411
412 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
413                                      struct efx_rx_buffer *rx_buf,
414                                      int len, bool *discard,
415                                      bool *leak_packet)
416 {
417         struct efx_nic *efx = rx_queue->efx;
418         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
419
420         if (likely(len <= max_len))
421                 return;
422
423         /* The packet must be discarded, but this is only a fatal error
424          * if the caller indicated it was
425          */
426         *discard = true;
427
428         if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
429                 if (net_ratelimit())
430                         netif_err(efx, rx_err, efx->net_dev,
431                                   " RX queue %d seriously overlength "
432                                   "RX event (0x%x > 0x%x+0x%x). Leaking\n",
433                                   efx_rx_queue_index(rx_queue), len, max_len,
434                                   efx->type->rx_buffer_padding);
435                 /* If this buffer was skb-allocated, then the meta
436                  * data at the end of the skb will be trashed. So
437                  * we have no choice but to leak the fragment.
438                  */
439                 *leak_packet = !rx_buf->is_page;
440                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
441         } else {
442                 if (net_ratelimit())
443                         netif_err(efx, rx_err, efx->net_dev,
444                                   " RX queue %d overlength RX event "
445                                   "(0x%x > 0x%x)\n",
446                                   efx_rx_queue_index(rx_queue), len, max_len);
447         }
448
449         efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
450 }
451
452 /* Pass a received packet up through the generic GRO stack
453  *
454  * Handles driverlink veto, and passes the fragment up via
455  * the appropriate GRO method
456  */
457 static void efx_rx_packet_gro(struct efx_channel *channel,
458                               struct efx_rx_buffer *rx_buf,
459                               const u8 *eh, bool checksummed)
460 {
461         struct napi_struct *napi = &channel->napi_str;
462         gro_result_t gro_result;
463
464         /* Pass the skb/page into the GRO engine */
465         if (rx_buf->is_page) {
466                 struct efx_nic *efx = channel->efx;
467                 struct page *page = rx_buf->u.page;
468                 struct sk_buff *skb;
469
470                 rx_buf->u.page = NULL;
471
472                 skb = napi_get_frags(napi);
473                 if (!skb) {
474                         put_page(page);
475                         return;
476                 }
477
478                 if (efx->net_dev->features & NETIF_F_RXHASH)
479                         skb->rxhash = efx_rx_buf_hash(eh);
480
481                 skb_shinfo(skb)->frags[0].page = page;
482                 skb_shinfo(skb)->frags[0].page_offset =
483                         efx_rx_buf_offset(efx, rx_buf);
484                 skb_shinfo(skb)->frags[0].size = rx_buf->len;
485                 skb_shinfo(skb)->nr_frags = 1;
486
487                 skb->len = rx_buf->len;
488                 skb->data_len = rx_buf->len;
489                 skb->truesize += rx_buf->len;
490                 skb->ip_summed =
491                         checksummed ? CHECKSUM_UNNECESSARY : CHECKSUM_NONE;
492
493                 skb_record_rx_queue(skb, channel->channel);
494
495                 gro_result = napi_gro_frags(napi);
496         } else {
497                 struct sk_buff *skb = rx_buf->u.skb;
498
499                 EFX_BUG_ON_PARANOID(!checksummed);
500                 rx_buf->u.skb = NULL;
501
502                 gro_result = napi_gro_receive(napi, skb);
503         }
504
505         if (gro_result == GRO_NORMAL) {
506                 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
507         } else if (gro_result != GRO_DROP) {
508                 channel->rx_alloc_level += RX_ALLOC_FACTOR_GRO;
509                 channel->irq_mod_score += 2;
510         }
511 }
512
513 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
514                    unsigned int len, bool checksummed, bool discard)
515 {
516         struct efx_nic *efx = rx_queue->efx;
517         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
518         struct efx_rx_buffer *rx_buf;
519         bool leak_packet = false;
520
521         rx_buf = efx_rx_buffer(rx_queue, index);
522
523         /* This allows the refill path to post another buffer.
524          * EFX_RXD_HEAD_ROOM ensures that the slot we are using
525          * isn't overwritten yet.
526          */
527         rx_queue->removed_count++;
528
529         /* Validate the length encoded in the event vs the descriptor pushed */
530         efx_rx_packet__check_len(rx_queue, rx_buf, len,
531                                  &discard, &leak_packet);
532
533         netif_vdbg(efx, rx_status, efx->net_dev,
534                    "RX queue %d received id %x at %llx+%x %s%s\n",
535                    efx_rx_queue_index(rx_queue), index,
536                    (unsigned long long)rx_buf->dma_addr, len,
537                    (checksummed ? " [SUMMED]" : ""),
538                    (discard ? " [DISCARD]" : ""));
539
540         /* Discard packet, if instructed to do so */
541         if (unlikely(discard)) {
542                 if (unlikely(leak_packet))
543                         channel->n_skbuff_leaks++;
544                 else
545                         efx_recycle_rx_buffer(channel, rx_buf);
546
547                 /* Don't hold off the previous receive */
548                 rx_buf = NULL;
549                 goto out;
550         }
551
552         /* Release card resources - assumes all RX buffers consumed in-order
553          * per RX queue
554          */
555         efx_unmap_rx_buffer(efx, rx_buf);
556
557         /* Prefetch nice and early so data will (hopefully) be in cache by
558          * the time we look at it.
559          */
560         prefetch(efx_rx_buf_eh(efx, rx_buf));
561
562         /* Pipeline receives so that we give time for packet headers to be
563          * prefetched into cache.
564          */
565         rx_buf->len = len - efx->type->rx_buffer_hash_size;
566 out:
567         if (channel->rx_pkt)
568                 __efx_rx_packet(channel,
569                                 channel->rx_pkt, channel->rx_pkt_csummed);
570         channel->rx_pkt = rx_buf;
571         channel->rx_pkt_csummed = checksummed;
572 }
573
574 /* Handle a received packet.  Second half: Touches packet payload. */
575 void __efx_rx_packet(struct efx_channel *channel,
576                      struct efx_rx_buffer *rx_buf, bool checksummed)
577 {
578         struct efx_nic *efx = channel->efx;
579         struct sk_buff *skb;
580         u8 *eh = efx_rx_buf_eh(efx, rx_buf);
581
582         /* If we're in loopback test, then pass the packet directly to the
583          * loopback layer, and free the rx_buf here
584          */
585         if (unlikely(efx->loopback_selftest)) {
586                 efx_loopback_rx_packet(efx, eh, rx_buf->len);
587                 efx_free_rx_buffer(efx, rx_buf);
588                 return;
589         }
590
591         if (!rx_buf->is_page) {
592                 skb = rx_buf->u.skb;
593
594                 prefetch(skb_shinfo(skb));
595
596                 skb_reserve(skb, efx->type->rx_buffer_hash_size);
597                 skb_put(skb, rx_buf->len);
598
599                 if (efx->net_dev->features & NETIF_F_RXHASH)
600                         skb->rxhash = efx_rx_buf_hash(eh);
601
602                 /* Move past the ethernet header. rx_buf->data still points
603                  * at the ethernet header */
604                 skb->protocol = eth_type_trans(skb, efx->net_dev);
605
606                 skb_record_rx_queue(skb, channel->channel);
607         }
608
609         if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
610                 checksummed = false;
611
612         if (likely(checksummed || rx_buf->is_page)) {
613                 efx_rx_packet_gro(channel, rx_buf, eh, checksummed);
614                 return;
615         }
616
617         /* We now own the SKB */
618         skb = rx_buf->u.skb;
619         rx_buf->u.skb = NULL;
620
621         /* Set the SKB flags */
622         skb_checksum_none_assert(skb);
623
624         /* Pass the packet up */
625         netif_receive_skb(skb);
626
627         /* Update allocation strategy method */
628         channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
629 }
630
631 void efx_rx_strategy(struct efx_channel *channel)
632 {
633         enum efx_rx_alloc_method method = rx_alloc_method;
634
635         /* Only makes sense to use page based allocation if GRO is enabled */
636         if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
637                 method = RX_ALLOC_METHOD_SKB;
638         } else if (method == RX_ALLOC_METHOD_AUTO) {
639                 /* Constrain the rx_alloc_level */
640                 if (channel->rx_alloc_level < 0)
641                         channel->rx_alloc_level = 0;
642                 else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
643                         channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
644
645                 /* Decide on the allocation method */
646                 method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_GRO) ?
647                           RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
648         }
649
650         /* Push the option */
651         channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
652 }
653
654 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
655 {
656         struct efx_nic *efx = rx_queue->efx;
657         unsigned int entries;
658         int rc;
659
660         /* Create the smallest power-of-two aligned ring */
661         entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
662         EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
663         rx_queue->ptr_mask = entries - 1;
664
665         netif_dbg(efx, probe, efx->net_dev,
666                   "creating RX queue %d size %#x mask %#x\n",
667                   efx_rx_queue_index(rx_queue), efx->rxq_entries,
668                   rx_queue->ptr_mask);
669
670         /* Allocate RX buffers */
671         rx_queue->buffer = kzalloc(entries * sizeof(*rx_queue->buffer),
672                                    GFP_KERNEL);
673         if (!rx_queue->buffer)
674                 return -ENOMEM;
675
676         rc = efx_nic_probe_rx(rx_queue);
677         if (rc) {
678                 kfree(rx_queue->buffer);
679                 rx_queue->buffer = NULL;
680         }
681         return rc;
682 }
683
684 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
685 {
686         struct efx_nic *efx = rx_queue->efx;
687         unsigned int max_fill, trigger, limit;
688
689         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
690                   "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
691
692         /* Initialise ptr fields */
693         rx_queue->added_count = 0;
694         rx_queue->notified_count = 0;
695         rx_queue->removed_count = 0;
696         rx_queue->min_fill = -1U;
697
698         /* Initialise limit fields */
699         max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
700         trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
701         limit = max_fill * min(rx_refill_limit, 100U) / 100U;
702
703         rx_queue->max_fill = max_fill;
704         rx_queue->fast_fill_trigger = trigger;
705         rx_queue->fast_fill_limit = limit;
706
707         /* Set up RX descriptor ring */
708         efx_nic_init_rx(rx_queue);
709 }
710
711 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
712 {
713         int i;
714         struct efx_rx_buffer *rx_buf;
715
716         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
717                   "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
718
719         del_timer_sync(&rx_queue->slow_fill);
720         efx_nic_fini_rx(rx_queue);
721
722         /* Release RX buffers NB start at index 0 not current HW ptr */
723         if (rx_queue->buffer) {
724                 for (i = 0; i <= rx_queue->ptr_mask; i++) {
725                         rx_buf = efx_rx_buffer(rx_queue, i);
726                         efx_fini_rx_buffer(rx_queue, rx_buf);
727                 }
728         }
729 }
730
731 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
732 {
733         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
734                   "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
735
736         efx_nic_remove_rx(rx_queue);
737
738         kfree(rx_queue->buffer);
739         rx_queue->buffer = NULL;
740 }
741
742
743 module_param(rx_alloc_method, int, 0644);
744 MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
745
746 module_param(rx_refill_threshold, uint, 0444);
747 MODULE_PARM_DESC(rx_refill_threshold,
748                  "RX descriptor ring fast/slow fill threshold (%)");
749