sfc: Read MC firmware version when requested through ethtool
[linux-2.6.git] / drivers / net / sfc / rx.c
1 /****************************************************************************
2  * Driver for Solarflare Solarstorm network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2009 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 <net/ip.h>
18 #include <net/checksum.h>
19 #include "net_driver.h"
20 #include "efx.h"
21 #include "nic.h"
22 #include "selftest.h"
23 #include "workarounds.h"
24
25 /* Number of RX descriptors pushed at once. */
26 #define EFX_RX_BATCH  8
27
28 /* Maximum size of a buffer sharing a page */
29 #define EFX_RX_HALF_PAGE ((PAGE_SIZE >> 1) - sizeof(struct efx_rx_page_state))
30
31 /* Size of buffer allocated for skb header area. */
32 #define EFX_SKB_HEADERS  64u
33
34 /*
35  * rx_alloc_method - RX buffer allocation method
36  *
37  * This driver supports two methods for allocating and using RX buffers:
38  * each RX buffer may be backed by an skb or by an order-n page.
39  *
40  * When GRO is in use then the second method has a lower overhead,
41  * since we don't have to allocate then free skbs on reassembled frames.
42  *
43  * Values:
44  *   - RX_ALLOC_METHOD_AUTO = 0
45  *   - RX_ALLOC_METHOD_SKB  = 1
46  *   - RX_ALLOC_METHOD_PAGE = 2
47  *
48  * The heuristic for %RX_ALLOC_METHOD_AUTO is a simple hysteresis count
49  * controlled by the parameters below.
50  *
51  *   - Since pushing and popping descriptors are separated by the rx_queue
52  *     size, so the watermarks should be ~rxd_size.
53  *   - The performance win by using page-based allocation for GRO is less
54  *     than the performance hit of using page-based allocation of non-GRO,
55  *     so the watermarks should reflect this.
56  *
57  * Per channel we maintain a single variable, updated by each channel:
58  *
59  *   rx_alloc_level += (gro_performed ? RX_ALLOC_FACTOR_GRO :
60  *                      RX_ALLOC_FACTOR_SKB)
61  * Per NAPI poll interval, we constrain rx_alloc_level to 0..MAX (which
62  * limits the hysteresis), and update the allocation strategy:
63  *
64  *   rx_alloc_method = (rx_alloc_level > RX_ALLOC_LEVEL_GRO ?
65  *                      RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB)
66  */
67 static int rx_alloc_method = RX_ALLOC_METHOD_AUTO;
68
69 #define RX_ALLOC_LEVEL_GRO 0x2000
70 #define RX_ALLOC_LEVEL_MAX 0x3000
71 #define RX_ALLOC_FACTOR_GRO 1
72 #define RX_ALLOC_FACTOR_SKB (-2)
73
74 /* This is the percentage fill level below which new RX descriptors
75  * will be added to the RX descriptor ring.
76  */
77 static unsigned int rx_refill_threshold = 90;
78
79 /* This is the percentage fill level to which an RX queue will be refilled
80  * when the "RX refill threshold" is reached.
81  */
82 static unsigned int rx_refill_limit = 95;
83
84 /*
85  * RX maximum head room required.
86  *
87  * This must be at least 1 to prevent overflow and at least 2 to allow
88  * pipelined receives.
89  */
90 #define EFX_RXD_HEAD_ROOM 2
91
92 /* Offset of ethernet header within page */
93 static inline unsigned int efx_rx_buf_offset(struct efx_nic *efx,
94                                              struct efx_rx_buffer *buf)
95 {
96         /* Offset is always within one page, so we don't need to consider
97          * the page order.
98          */
99         return (((__force unsigned long) buf->dma_addr & (PAGE_SIZE - 1)) +
100                 efx->type->rx_buffer_hash_size);
101 }
102 static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
103 {
104         return PAGE_SIZE << efx->rx_buffer_order;
105 }
106
107 static u8 *efx_rx_buf_eh(struct efx_nic *efx, struct efx_rx_buffer *buf)
108 {
109         if (buf->is_page)
110                 return page_address(buf->u.page) + efx_rx_buf_offset(efx, buf);
111         else
112                 return ((u8 *)buf->u.skb->data +
113                         efx->type->rx_buffer_hash_size);
114 }
115
116 static inline u32 efx_rx_buf_hash(const u8 *eh)
117 {
118         /* The ethernet header is always directly after any hash. */
119 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) || NET_IP_ALIGN % 4 == 0
120         return __le32_to_cpup((const __le32 *)(eh - 4));
121 #else
122         const u8 *data = eh - 4;
123         return ((u32)data[0]       |
124                 (u32)data[1] << 8  |
125                 (u32)data[2] << 16 |
126                 (u32)data[3] << 24);
127 #endif
128 }
129
130 /**
131  * efx_init_rx_buffers_skb - create EFX_RX_BATCH skb-based RX buffers
132  *
133  * @rx_queue:           Efx RX queue
134  *
135  * This allocates EFX_RX_BATCH skbs, maps them for DMA, and populates a
136  * struct efx_rx_buffer for each one. Return a negative error code or 0
137  * on success. May fail having only inserted fewer than EFX_RX_BATCH
138  * buffers.
139  */
140 static int efx_init_rx_buffers_skb(struct efx_rx_queue *rx_queue)
141 {
142         struct efx_nic *efx = rx_queue->efx;
143         struct net_device *net_dev = efx->net_dev;
144         struct efx_rx_buffer *rx_buf;
145         struct sk_buff *skb;
146         int skb_len = efx->rx_buffer_len;
147         unsigned index, count;
148
149         for (count = 0; count < EFX_RX_BATCH; ++count) {
150                 index = rx_queue->added_count & rx_queue->ptr_mask;
151                 rx_buf = efx_rx_buffer(rx_queue, index);
152
153                 rx_buf->u.skb = skb = netdev_alloc_skb(net_dev, skb_len);
154                 if (unlikely(!skb))
155                         return -ENOMEM;
156
157                 /* Adjust the SKB for padding and checksum */
158                 skb_reserve(skb, NET_IP_ALIGN);
159                 rx_buf->len = skb_len - NET_IP_ALIGN;
160                 rx_buf->is_page = false;
161                 skb->ip_summed = CHECKSUM_UNNECESSARY;
162
163                 rx_buf->dma_addr = pci_map_single(efx->pci_dev,
164                                                   skb->data, rx_buf->len,
165                                                   PCI_DMA_FROMDEVICE);
166                 if (unlikely(pci_dma_mapping_error(efx->pci_dev,
167                                                    rx_buf->dma_addr))) {
168                         dev_kfree_skb_any(skb);
169                         rx_buf->u.skb = NULL;
170                         return -EIO;
171                 }
172
173                 ++rx_queue->added_count;
174                 ++rx_queue->alloc_skb_count;
175         }
176
177         return 0;
178 }
179
180 /**
181  * efx_init_rx_buffers_page - create EFX_RX_BATCH page-based RX buffers
182  *
183  * @rx_queue:           Efx RX queue
184  *
185  * This allocates memory for EFX_RX_BATCH receive buffers, maps them for DMA,
186  * and populates struct efx_rx_buffers for each one. Return a negative error
187  * code or 0 on success. If a single page can be split between two buffers,
188  * then the page will either be inserted fully, or not at at all.
189  */
190 static int efx_init_rx_buffers_page(struct efx_rx_queue *rx_queue)
191 {
192         struct efx_nic *efx = rx_queue->efx;
193         struct efx_rx_buffer *rx_buf;
194         struct page *page;
195         void *page_addr;
196         struct efx_rx_page_state *state;
197         dma_addr_t dma_addr;
198         unsigned index, count;
199
200         /* We can split a page between two buffers */
201         BUILD_BUG_ON(EFX_RX_BATCH & 1);
202
203         for (count = 0; count < EFX_RX_BATCH; ++count) {
204                 page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
205                                    efx->rx_buffer_order);
206                 if (unlikely(page == NULL))
207                         return -ENOMEM;
208                 dma_addr = pci_map_page(efx->pci_dev, page, 0,
209                                         efx_rx_buf_size(efx),
210                                         PCI_DMA_FROMDEVICE);
211                 if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
212                         __free_pages(page, efx->rx_buffer_order);
213                         return -EIO;
214                 }
215                 page_addr = page_address(page);
216                 state = page_addr;
217                 state->refcnt = 0;
218                 state->dma_addr = dma_addr;
219
220                 page_addr += sizeof(struct efx_rx_page_state);
221                 dma_addr += sizeof(struct efx_rx_page_state);
222
223         split:
224                 index = rx_queue->added_count & rx_queue->ptr_mask;
225                 rx_buf = efx_rx_buffer(rx_queue, index);
226                 rx_buf->dma_addr = dma_addr + EFX_PAGE_IP_ALIGN;
227                 rx_buf->u.page = page;
228                 rx_buf->len = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
229                 rx_buf->is_page = true;
230                 ++rx_queue->added_count;
231                 ++rx_queue->alloc_page_count;
232                 ++state->refcnt;
233
234                 if ((~count & 1) && (efx->rx_buffer_len <= EFX_RX_HALF_PAGE)) {
235                         /* Use the second half of the page */
236                         get_page(page);
237                         dma_addr += (PAGE_SIZE >> 1);
238                         page_addr += (PAGE_SIZE >> 1);
239                         ++count;
240                         goto split;
241                 }
242         }
243
244         return 0;
245 }
246
247 static void efx_unmap_rx_buffer(struct efx_nic *efx,
248                                 struct efx_rx_buffer *rx_buf)
249 {
250         if (rx_buf->is_page && rx_buf->u.page) {
251                 struct efx_rx_page_state *state;
252
253                 state = page_address(rx_buf->u.page);
254                 if (--state->refcnt == 0) {
255                         pci_unmap_page(efx->pci_dev,
256                                        state->dma_addr,
257                                        efx_rx_buf_size(efx),
258                                        PCI_DMA_FROMDEVICE);
259                 }
260         } else if (!rx_buf->is_page && rx_buf->u.skb) {
261                 pci_unmap_single(efx->pci_dev, rx_buf->dma_addr,
262                                  rx_buf->len, PCI_DMA_FROMDEVICE);
263         }
264 }
265
266 static void efx_free_rx_buffer(struct efx_nic *efx,
267                                struct efx_rx_buffer *rx_buf)
268 {
269         if (rx_buf->is_page && rx_buf->u.page) {
270                 __free_pages(rx_buf->u.page, efx->rx_buffer_order);
271                 rx_buf->u.page = NULL;
272         } else if (!rx_buf->is_page && rx_buf->u.skb) {
273                 dev_kfree_skb_any(rx_buf->u.skb);
274                 rx_buf->u.skb = NULL;
275         }
276 }
277
278 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
279                                struct efx_rx_buffer *rx_buf)
280 {
281         efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
282         efx_free_rx_buffer(rx_queue->efx, rx_buf);
283 }
284
285 /* Attempt to resurrect the other receive buffer that used to share this page,
286  * which had previously been passed up to the kernel and freed. */
287 static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
288                                     struct efx_rx_buffer *rx_buf)
289 {
290         struct efx_rx_page_state *state = page_address(rx_buf->u.page);
291         struct efx_rx_buffer *new_buf;
292         unsigned fill_level, index;
293
294         /* +1 because efx_rx_packet() incremented removed_count. +1 because
295          * we'd like to insert an additional descriptor whilst leaving
296          * EFX_RXD_HEAD_ROOM for the non-recycle path */
297         fill_level = (rx_queue->added_count - rx_queue->removed_count + 2);
298         if (unlikely(fill_level > rx_queue->max_fill)) {
299                 /* We could place "state" on a list, and drain the list in
300                  * efx_fast_push_rx_descriptors(). For now, this will do. */
301                 return;
302         }
303
304         ++state->refcnt;
305         get_page(rx_buf->u.page);
306
307         index = rx_queue->added_count & rx_queue->ptr_mask;
308         new_buf = efx_rx_buffer(rx_queue, index);
309         new_buf->dma_addr = rx_buf->dma_addr ^ (PAGE_SIZE >> 1);
310         new_buf->u.page = rx_buf->u.page;
311         new_buf->len = rx_buf->len;
312         new_buf->is_page = true;
313         ++rx_queue->added_count;
314 }
315
316 /* Recycle the given rx buffer directly back into the rx_queue. There is
317  * always room to add this buffer, because we've just popped a buffer. */
318 static void efx_recycle_rx_buffer(struct efx_channel *channel,
319                                   struct efx_rx_buffer *rx_buf)
320 {
321         struct efx_nic *efx = channel->efx;
322         struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
323         struct efx_rx_buffer *new_buf;
324         unsigned index;
325
326         if (rx_buf->is_page && efx->rx_buffer_len <= EFX_RX_HALF_PAGE &&
327             page_count(rx_buf->u.page) == 1)
328                 efx_resurrect_rx_buffer(rx_queue, rx_buf);
329
330         index = rx_queue->added_count & rx_queue->ptr_mask;
331         new_buf = efx_rx_buffer(rx_queue, index);
332
333         memcpy(new_buf, rx_buf, sizeof(*new_buf));
334         rx_buf->u.page = NULL;
335         ++rx_queue->added_count;
336 }
337
338 /**
339  * efx_fast_push_rx_descriptors - push new RX descriptors quickly
340  * @rx_queue:           RX descriptor queue
341  * This will aim to fill the RX descriptor queue up to
342  * @rx_queue->@fast_fill_limit. If there is insufficient atomic
343  * memory to do so, a slow fill will be scheduled.
344  *
345  * The caller must provide serialisation (none is used here). In practise,
346  * this means this function must run from the NAPI handler, or be called
347  * when NAPI is disabled.
348  */
349 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
350 {
351         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
352         unsigned fill_level;
353         int space, rc = 0;
354
355         /* Calculate current fill level, and exit if we don't need to fill */
356         fill_level = (rx_queue->added_count - rx_queue->removed_count);
357         EFX_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
358         if (fill_level >= rx_queue->fast_fill_trigger)
359                 goto out;
360
361         /* Record minimum fill level */
362         if (unlikely(fill_level < rx_queue->min_fill)) {
363                 if (fill_level)
364                         rx_queue->min_fill = fill_level;
365         }
366
367         space = rx_queue->fast_fill_limit - fill_level;
368         if (space < EFX_RX_BATCH)
369                 goto out;
370
371         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
372                    "RX queue %d fast-filling descriptor ring from"
373                    " level %d to level %d using %s allocation\n",
374                    efx_rx_queue_index(rx_queue), fill_level,
375                    rx_queue->fast_fill_limit,
376                    channel->rx_alloc_push_pages ? "page" : "skb");
377
378         do {
379                 if (channel->rx_alloc_push_pages)
380                         rc = efx_init_rx_buffers_page(rx_queue);
381                 else
382                         rc = efx_init_rx_buffers_skb(rx_queue);
383                 if (unlikely(rc)) {
384                         /* Ensure that we don't leave the rx queue empty */
385                         if (rx_queue->added_count == rx_queue->removed_count)
386                                 efx_schedule_slow_fill(rx_queue);
387                         goto out;
388                 }
389         } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
390
391         netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
392                    "RX queue %d fast-filled descriptor ring "
393                    "to level %d\n", efx_rx_queue_index(rx_queue),
394                    rx_queue->added_count - rx_queue->removed_count);
395
396  out:
397         if (rx_queue->notified_count != rx_queue->added_count)
398                 efx_nic_notify_rx_desc(rx_queue);
399 }
400
401 void efx_rx_slow_fill(unsigned long context)
402 {
403         struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
404         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
405
406         /* Post an event to cause NAPI to run and refill the queue */
407         efx_nic_generate_fill_event(channel);
408         ++rx_queue->slow_fill_count;
409 }
410
411 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
412                                      struct efx_rx_buffer *rx_buf,
413                                      int len, bool *discard,
414                                      bool *leak_packet)
415 {
416         struct efx_nic *efx = rx_queue->efx;
417         unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
418
419         if (likely(len <= max_len))
420                 return;
421
422         /* The packet must be discarded, but this is only a fatal error
423          * if the caller indicated it was
424          */
425         *discard = true;
426
427         if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
428                 if (net_ratelimit())
429                         netif_err(efx, rx_err, efx->net_dev,
430                                   " RX queue %d seriously overlength "
431                                   "RX event (0x%x > 0x%x+0x%x). Leaking\n",
432                                   efx_rx_queue_index(rx_queue), len, max_len,
433                                   efx->type->rx_buffer_padding);
434                 /* If this buffer was skb-allocated, then the meta
435                  * data at the end of the skb will be trashed. So
436                  * we have no choice but to leak the fragment.
437                  */
438                 *leak_packet = !rx_buf->is_page;
439                 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
440         } else {
441                 if (net_ratelimit())
442                         netif_err(efx, rx_err, efx->net_dev,
443                                   " RX queue %d overlength RX event "
444                                   "(0x%x > 0x%x)\n",
445                                   efx_rx_queue_index(rx_queue), len, max_len);
446         }
447
448         efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
449 }
450
451 /* Pass a received packet up through the generic GRO stack
452  *
453  * Handles driverlink veto, and passes the fragment up via
454  * the appropriate GRO method
455  */
456 static void efx_rx_packet_gro(struct efx_channel *channel,
457                               struct efx_rx_buffer *rx_buf,
458                               const u8 *eh, bool checksummed)
459 {
460         struct napi_struct *napi = &channel->napi_str;
461         gro_result_t gro_result;
462
463         /* Pass the skb/page into the GRO engine */
464         if (rx_buf->is_page) {
465                 struct efx_nic *efx = channel->efx;
466                 struct page *page = rx_buf->u.page;
467                 struct sk_buff *skb;
468
469                 rx_buf->u.page = NULL;
470
471                 skb = napi_get_frags(napi);
472                 if (!skb) {
473                         put_page(page);
474                         return;
475                 }
476
477                 if (efx->net_dev->features & NETIF_F_RXHASH)
478                         skb->rxhash = efx_rx_buf_hash(eh);
479
480                 skb_shinfo(skb)->frags[0].page = page;
481                 skb_shinfo(skb)->frags[0].page_offset =
482                         efx_rx_buf_offset(efx, rx_buf);
483                 skb_shinfo(skb)->frags[0].size = rx_buf->len;
484                 skb_shinfo(skb)->nr_frags = 1;
485
486                 skb->len = rx_buf->len;
487                 skb->data_len = rx_buf->len;
488                 skb->truesize += rx_buf->len;
489                 skb->ip_summed =
490                         checksummed ? CHECKSUM_UNNECESSARY : CHECKSUM_NONE;
491
492                 skb_record_rx_queue(skb, channel->channel);
493
494                 gro_result = napi_gro_frags(napi);
495         } else {
496                 struct sk_buff *skb = rx_buf->u.skb;
497
498                 EFX_BUG_ON_PARANOID(!checksummed);
499                 rx_buf->u.skb = NULL;
500
501                 gro_result = napi_gro_receive(napi, skb);
502         }
503
504         if (gro_result == GRO_NORMAL) {
505                 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
506         } else if (gro_result != GRO_DROP) {
507                 channel->rx_alloc_level += RX_ALLOC_FACTOR_GRO;
508                 channel->irq_mod_score += 2;
509         }
510 }
511
512 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
513                    unsigned int len, bool checksummed, bool discard)
514 {
515         struct efx_nic *efx = rx_queue->efx;
516         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
517         struct efx_rx_buffer *rx_buf;
518         bool leak_packet = false;
519
520         rx_buf = efx_rx_buffer(rx_queue, index);
521
522         /* This allows the refill path to post another buffer.
523          * EFX_RXD_HEAD_ROOM ensures that the slot we are using
524          * isn't overwritten yet.
525          */
526         rx_queue->removed_count++;
527
528         /* Validate the length encoded in the event vs the descriptor pushed */
529         efx_rx_packet__check_len(rx_queue, rx_buf, len,
530                                  &discard, &leak_packet);
531
532         netif_vdbg(efx, rx_status, efx->net_dev,
533                    "RX queue %d received id %x at %llx+%x %s%s\n",
534                    efx_rx_queue_index(rx_queue), index,
535                    (unsigned long long)rx_buf->dma_addr, len,
536                    (checksummed ? " [SUMMED]" : ""),
537                    (discard ? " [DISCARD]" : ""));
538
539         /* Discard packet, if instructed to do so */
540         if (unlikely(discard)) {
541                 if (unlikely(leak_packet))
542                         channel->n_skbuff_leaks++;
543                 else
544                         efx_recycle_rx_buffer(channel, rx_buf);
545
546                 /* Don't hold off the previous receive */
547                 rx_buf = NULL;
548                 goto out;
549         }
550
551         /* Release card resources - assumes all RX buffers consumed in-order
552          * per RX queue
553          */
554         efx_unmap_rx_buffer(efx, rx_buf);
555
556         /* Prefetch nice and early so data will (hopefully) be in cache by
557          * the time we look at it.
558          */
559         prefetch(efx_rx_buf_eh(efx, rx_buf));
560
561         /* Pipeline receives so that we give time for packet headers to be
562          * prefetched into cache.
563          */
564         rx_buf->len = len - efx->type->rx_buffer_hash_size;
565 out:
566         if (channel->rx_pkt)
567                 __efx_rx_packet(channel,
568                                 channel->rx_pkt, channel->rx_pkt_csummed);
569         channel->rx_pkt = rx_buf;
570         channel->rx_pkt_csummed = checksummed;
571 }
572
573 /* Handle a received packet.  Second half: Touches packet payload. */
574 void __efx_rx_packet(struct efx_channel *channel,
575                      struct efx_rx_buffer *rx_buf, bool checksummed)
576 {
577         struct efx_nic *efx = channel->efx;
578         struct sk_buff *skb;
579         u8 *eh = efx_rx_buf_eh(efx, rx_buf);
580
581         /* If we're in loopback test, then pass the packet directly to the
582          * loopback layer, and free the rx_buf here
583          */
584         if (unlikely(efx->loopback_selftest)) {
585                 efx_loopback_rx_packet(efx, eh, rx_buf->len);
586                 efx_free_rx_buffer(efx, rx_buf);
587                 return;
588         }
589
590         if (!rx_buf->is_page) {
591                 skb = rx_buf->u.skb;
592
593                 prefetch(skb_shinfo(skb));
594
595                 skb_reserve(skb, efx->type->rx_buffer_hash_size);
596                 skb_put(skb, rx_buf->len);
597
598                 if (efx->net_dev->features & NETIF_F_RXHASH)
599                         skb->rxhash = efx_rx_buf_hash(eh);
600
601                 /* Move past the ethernet header. rx_buf->data still points
602                  * at the ethernet header */
603                 skb->protocol = eth_type_trans(skb, efx->net_dev);
604
605                 skb_record_rx_queue(skb, channel->channel);
606         }
607
608         if (likely(checksummed || rx_buf->is_page)) {
609                 efx_rx_packet_gro(channel, rx_buf, eh, checksummed);
610                 return;
611         }
612
613         /* We now own the SKB */
614         skb = rx_buf->u.skb;
615         rx_buf->u.skb = NULL;
616
617         /* Set the SKB flags */
618         skb_checksum_none_assert(skb);
619
620         /* Pass the packet up */
621         netif_receive_skb(skb);
622
623         /* Update allocation strategy method */
624         channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
625 }
626
627 void efx_rx_strategy(struct efx_channel *channel)
628 {
629         enum efx_rx_alloc_method method = rx_alloc_method;
630
631         /* Only makes sense to use page based allocation if GRO is enabled */
632         if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
633                 method = RX_ALLOC_METHOD_SKB;
634         } else if (method == RX_ALLOC_METHOD_AUTO) {
635                 /* Constrain the rx_alloc_level */
636                 if (channel->rx_alloc_level < 0)
637                         channel->rx_alloc_level = 0;
638                 else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
639                         channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
640
641                 /* Decide on the allocation method */
642                 method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_GRO) ?
643                           RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
644         }
645
646         /* Push the option */
647         channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
648 }
649
650 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
651 {
652         struct efx_nic *efx = rx_queue->efx;
653         unsigned int entries;
654         int rc;
655
656         /* Create the smallest power-of-two aligned ring */
657         entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
658         EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
659         rx_queue->ptr_mask = entries - 1;
660
661         netif_dbg(efx, probe, efx->net_dev,
662                   "creating RX queue %d size %#x mask %#x\n",
663                   efx_rx_queue_index(rx_queue), efx->rxq_entries,
664                   rx_queue->ptr_mask);
665
666         /* Allocate RX buffers */
667         rx_queue->buffer = kzalloc(entries * sizeof(*rx_queue->buffer),
668                                    GFP_KERNEL);
669         if (!rx_queue->buffer)
670                 return -ENOMEM;
671
672         rc = efx_nic_probe_rx(rx_queue);
673         if (rc) {
674                 kfree(rx_queue->buffer);
675                 rx_queue->buffer = NULL;
676         }
677         return rc;
678 }
679
680 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
681 {
682         struct efx_nic *efx = rx_queue->efx;
683         unsigned int max_fill, trigger, limit;
684
685         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
686                   "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
687
688         /* Initialise ptr fields */
689         rx_queue->added_count = 0;
690         rx_queue->notified_count = 0;
691         rx_queue->removed_count = 0;
692         rx_queue->min_fill = -1U;
693
694         /* Initialise limit fields */
695         max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
696         trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
697         limit = max_fill * min(rx_refill_limit, 100U) / 100U;
698
699         rx_queue->max_fill = max_fill;
700         rx_queue->fast_fill_trigger = trigger;
701         rx_queue->fast_fill_limit = limit;
702
703         /* Set up RX descriptor ring */
704         efx_nic_init_rx(rx_queue);
705 }
706
707 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
708 {
709         int i;
710         struct efx_rx_buffer *rx_buf;
711
712         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
713                   "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
714
715         del_timer_sync(&rx_queue->slow_fill);
716         efx_nic_fini_rx(rx_queue);
717
718         /* Release RX buffers NB start at index 0 not current HW ptr */
719         if (rx_queue->buffer) {
720                 for (i = 0; i <= rx_queue->ptr_mask; i++) {
721                         rx_buf = efx_rx_buffer(rx_queue, i);
722                         efx_fini_rx_buffer(rx_queue, rx_buf);
723                 }
724         }
725 }
726
727 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
728 {
729         netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
730                   "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
731
732         efx_nic_remove_rx(rx_queue);
733
734         kfree(rx_queue->buffer);
735         rx_queue->buffer = NULL;
736 }
737
738
739 module_param(rx_alloc_method, int, 0644);
740 MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
741
742 module_param(rx_refill_threshold, uint, 0444);
743 MODULE_PARM_DESC(rx_refill_threshold,
744                  "RX descriptor ring fast/slow fill threshold (%)");
745