1 /****************************************************************************
2 * Driver for Solarflare Solarstorm network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2009 Solarflare Communications Inc.
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.
11 #include <linux/socket.h>
13 #include <linux/slab.h>
15 #include <linux/tcp.h>
16 #include <linux/udp.h>
18 #include <net/checksum.h>
19 #include "net_driver.h"
23 #include "workarounds.h"
25 /* Number of RX descriptors pushed at once. */
26 #define EFX_RX_BATCH 8
28 /* Maximum size of a buffer sharing a page */
29 #define EFX_RX_HALF_PAGE ((PAGE_SIZE >> 1) - sizeof(struct efx_rx_page_state))
31 /* Size of buffer allocated for skb header area. */
32 #define EFX_SKB_HEADERS 64u
35 * rx_alloc_method - RX buffer allocation method
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.
40 * When LRO 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.
44 * - RX_ALLOC_METHOD_AUTO = 0
45 * - RX_ALLOC_METHOD_SKB = 1
46 * - RX_ALLOC_METHOD_PAGE = 2
48 * The heuristic for %RX_ALLOC_METHOD_AUTO is a simple hysteresis count
49 * controlled by the parameters below.
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 LRO is less
54 * than the performance hit of using page-based allocation of non-LRO,
55 * so the watermarks should reflect this.
57 * Per channel we maintain a single variable, updated by each channel:
59 * rx_alloc_level += (lro_performed ? RX_ALLOC_FACTOR_LRO :
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:
64 * rx_alloc_method = (rx_alloc_level > RX_ALLOC_LEVEL_LRO ?
65 * RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB)
67 static int rx_alloc_method = RX_ALLOC_METHOD_AUTO;
69 #define RX_ALLOC_LEVEL_LRO 0x2000
70 #define RX_ALLOC_LEVEL_MAX 0x3000
71 #define RX_ALLOC_FACTOR_LRO 1
72 #define RX_ALLOC_FACTOR_SKB (-2)
74 /* This is the percentage fill level below which new RX descriptors
75 * will be added to the RX descriptor ring.
77 static unsigned int rx_refill_threshold = 90;
79 /* This is the percentage fill level to which an RX queue will be refilled
80 * when the "RX refill threshold" is reached.
82 static unsigned int rx_refill_limit = 95;
85 * RX maximum head room required.
87 * This must be at least 1 to prevent overflow and at least 2 to allow
90 #define EFX_RXD_HEAD_ROOM 2
92 static inline unsigned int efx_rx_buf_offset(struct efx_rx_buffer *buf)
94 /* Offset is always within one page, so we don't need to consider
97 return (__force unsigned long) buf->data & (PAGE_SIZE - 1);
99 static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
101 return PAGE_SIZE << efx->rx_buffer_order;
105 * efx_init_rx_buffers_skb - create EFX_RX_BATCH skb-based RX buffers
107 * @rx_queue: Efx RX queue
109 * This allocates EFX_RX_BATCH skbs, maps them for DMA, and populates a
110 * struct efx_rx_buffer for each one. Return a negative error code or 0
111 * on success. May fail having only inserted fewer than EFX_RX_BATCH
114 static int efx_init_rx_buffers_skb(struct efx_rx_queue *rx_queue)
116 struct efx_nic *efx = rx_queue->efx;
117 struct net_device *net_dev = efx->net_dev;
118 struct efx_rx_buffer *rx_buf;
119 int skb_len = efx->rx_buffer_len;
120 unsigned index, count;
122 for (count = 0; count < EFX_RX_BATCH; ++count) {
123 index = rx_queue->added_count & EFX_RXQ_MASK;
124 rx_buf = efx_rx_buffer(rx_queue, index);
126 rx_buf->skb = netdev_alloc_skb(net_dev, skb_len);
127 if (unlikely(!rx_buf->skb))
131 /* Adjust the SKB for padding and checksum */
132 skb_reserve(rx_buf->skb, NET_IP_ALIGN);
133 rx_buf->len = skb_len - NET_IP_ALIGN;
134 rx_buf->data = (char *)rx_buf->skb->data;
135 rx_buf->skb->ip_summed = CHECKSUM_UNNECESSARY;
137 rx_buf->dma_addr = pci_map_single(efx->pci_dev,
138 rx_buf->data, rx_buf->len,
140 if (unlikely(pci_dma_mapping_error(efx->pci_dev,
141 rx_buf->dma_addr))) {
142 dev_kfree_skb_any(rx_buf->skb);
147 ++rx_queue->added_count;
148 ++rx_queue->alloc_skb_count;
155 * efx_init_rx_buffers_page - create EFX_RX_BATCH page-based RX buffers
157 * @rx_queue: Efx RX queue
159 * This allocates memory for EFX_RX_BATCH receive buffers, maps them for DMA,
160 * and populates struct efx_rx_buffers for each one. Return a negative error
161 * code or 0 on success. If a single page can be split between two buffers,
162 * then the page will either be inserted fully, or not at at all.
164 static int efx_init_rx_buffers_page(struct efx_rx_queue *rx_queue)
166 struct efx_nic *efx = rx_queue->efx;
167 struct efx_rx_buffer *rx_buf;
170 struct efx_rx_page_state *state;
172 unsigned index, count;
174 /* We can split a page between two buffers */
175 BUILD_BUG_ON(EFX_RX_BATCH & 1);
177 for (count = 0; count < EFX_RX_BATCH; ++count) {
178 page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
179 efx->rx_buffer_order);
180 if (unlikely(page == NULL))
182 dma_addr = pci_map_page(efx->pci_dev, page, 0,
183 efx_rx_buf_size(efx),
185 if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
186 __free_pages(page, efx->rx_buffer_order);
189 page_addr = page_address(page);
192 state->dma_addr = dma_addr;
194 page_addr += sizeof(struct efx_rx_page_state);
195 dma_addr += sizeof(struct efx_rx_page_state);
198 index = rx_queue->added_count & EFX_RXQ_MASK;
199 rx_buf = efx_rx_buffer(rx_queue, index);
200 rx_buf->dma_addr = dma_addr + EFX_PAGE_IP_ALIGN;
203 rx_buf->data = page_addr + EFX_PAGE_IP_ALIGN;
204 rx_buf->len = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
205 ++rx_queue->added_count;
206 ++rx_queue->alloc_page_count;
209 if ((~count & 1) && (efx->rx_buffer_len <= EFX_RX_HALF_PAGE)) {
210 /* Use the second half of the page */
212 dma_addr += (PAGE_SIZE >> 1);
213 page_addr += (PAGE_SIZE >> 1);
222 static void efx_unmap_rx_buffer(struct efx_nic *efx,
223 struct efx_rx_buffer *rx_buf)
226 struct efx_rx_page_state *state;
228 EFX_BUG_ON_PARANOID(rx_buf->skb);
230 state = page_address(rx_buf->page);
231 if (--state->refcnt == 0) {
232 pci_unmap_page(efx->pci_dev,
234 efx_rx_buf_size(efx),
237 } else if (likely(rx_buf->skb)) {
238 pci_unmap_single(efx->pci_dev, rx_buf->dma_addr,
239 rx_buf->len, PCI_DMA_FROMDEVICE);
243 static void efx_free_rx_buffer(struct efx_nic *efx,
244 struct efx_rx_buffer *rx_buf)
247 __free_pages(rx_buf->page, efx->rx_buffer_order);
249 } else if (likely(rx_buf->skb)) {
250 dev_kfree_skb_any(rx_buf->skb);
255 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
256 struct efx_rx_buffer *rx_buf)
258 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
259 efx_free_rx_buffer(rx_queue->efx, rx_buf);
262 /* Attempt to resurrect the other receive buffer that used to share this page,
263 * which had previously been passed up to the kernel and freed. */
264 static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
265 struct efx_rx_buffer *rx_buf)
267 struct efx_rx_page_state *state = page_address(rx_buf->page);
268 struct efx_rx_buffer *new_buf;
269 unsigned fill_level, index;
271 /* +1 because efx_rx_packet() incremented removed_count. +1 because
272 * we'd like to insert an additional descriptor whilst leaving
273 * EFX_RXD_HEAD_ROOM for the non-recycle path */
274 fill_level = (rx_queue->added_count - rx_queue->removed_count + 2);
275 if (unlikely(fill_level >= EFX_RXQ_SIZE - EFX_RXD_HEAD_ROOM)) {
276 /* We could place "state" on a list, and drain the list in
277 * efx_fast_push_rx_descriptors(). For now, this will do. */
282 get_page(rx_buf->page);
284 index = rx_queue->added_count & EFX_RXQ_MASK;
285 new_buf = efx_rx_buffer(rx_queue, index);
286 new_buf->dma_addr = rx_buf->dma_addr ^ (PAGE_SIZE >> 1);
288 new_buf->page = rx_buf->page;
289 new_buf->data = (void *)
290 ((__force unsigned long)rx_buf->data ^ (PAGE_SIZE >> 1));
291 new_buf->len = rx_buf->len;
292 ++rx_queue->added_count;
295 /* Recycle the given rx buffer directly back into the rx_queue. There is
296 * always room to add this buffer, because we've just popped a buffer. */
297 static void efx_recycle_rx_buffer(struct efx_channel *channel,
298 struct efx_rx_buffer *rx_buf)
300 struct efx_nic *efx = channel->efx;
301 struct efx_rx_queue *rx_queue = &efx->rx_queue[channel->channel];
302 struct efx_rx_buffer *new_buf;
305 if (rx_buf->page != NULL && efx->rx_buffer_len <= EFX_RX_HALF_PAGE &&
306 page_count(rx_buf->page) == 1)
307 efx_resurrect_rx_buffer(rx_queue, rx_buf);
309 index = rx_queue->added_count & EFX_RXQ_MASK;
310 new_buf = efx_rx_buffer(rx_queue, index);
312 memcpy(new_buf, rx_buf, sizeof(*new_buf));
315 ++rx_queue->added_count;
319 * efx_fast_push_rx_descriptors - push new RX descriptors quickly
320 * @rx_queue: RX descriptor queue
321 * This will aim to fill the RX descriptor queue up to
322 * @rx_queue->@fast_fill_limit. If there is insufficient atomic
323 * memory to do so, a slow fill will be scheduled.
325 * The caller must provide serialisation (none is used here). In practise,
326 * this means this function must run from the NAPI handler, or be called
327 * when NAPI is disabled.
329 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
331 struct efx_channel *channel = rx_queue->channel;
335 /* Calculate current fill level, and exit if we don't need to fill */
336 fill_level = (rx_queue->added_count - rx_queue->removed_count);
337 EFX_BUG_ON_PARANOID(fill_level > EFX_RXQ_SIZE);
338 if (fill_level >= rx_queue->fast_fill_trigger)
341 /* Record minimum fill level */
342 if (unlikely(fill_level < rx_queue->min_fill)) {
344 rx_queue->min_fill = fill_level;
347 space = rx_queue->fast_fill_limit - fill_level;
348 if (space < EFX_RX_BATCH)
351 EFX_TRACE(rx_queue->efx, "RX queue %d fast-filling descriptor ring from"
352 " level %d to level %d using %s allocation\n",
353 rx_queue->queue, fill_level, rx_queue->fast_fill_limit,
354 channel->rx_alloc_push_pages ? "page" : "skb");
357 if (channel->rx_alloc_push_pages)
358 rc = efx_init_rx_buffers_page(rx_queue);
360 rc = efx_init_rx_buffers_skb(rx_queue);
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);
367 } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
369 EFX_TRACE(rx_queue->efx, "RX queue %d fast-filled descriptor ring "
370 "to level %d\n", rx_queue->queue,
371 rx_queue->added_count - rx_queue->removed_count);
374 if (rx_queue->notified_count != rx_queue->added_count)
375 efx_nic_notify_rx_desc(rx_queue);
378 void efx_rx_slow_fill(unsigned long context)
380 struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
381 struct efx_channel *channel = rx_queue->channel;
383 /* Post an event to cause NAPI to run and refill the queue */
384 efx_nic_generate_fill_event(channel);
385 ++rx_queue->slow_fill_count;
388 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
389 struct efx_rx_buffer *rx_buf,
390 int len, bool *discard,
393 struct efx_nic *efx = rx_queue->efx;
394 unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
396 if (likely(len <= max_len))
399 /* The packet must be discarded, but this is only a fatal error
400 * if the caller indicated it was
404 if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
405 EFX_ERR_RL(efx, " RX queue %d seriously overlength "
406 "RX event (0x%x > 0x%x+0x%x). Leaking\n",
407 rx_queue->queue, len, max_len,
408 efx->type->rx_buffer_padding);
409 /* If this buffer was skb-allocated, then the meta
410 * data at the end of the skb will be trashed. So
411 * we have no choice but to leak the fragment.
413 *leak_packet = (rx_buf->skb != NULL);
414 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
416 EFX_ERR_RL(efx, " RX queue %d overlength RX event "
417 "(0x%x > 0x%x)\n", rx_queue->queue, len, max_len);
420 rx_queue->channel->n_rx_overlength++;
423 /* Pass a received packet up through the generic LRO stack
425 * Handles driverlink veto, and passes the fragment up via
426 * the appropriate LRO method
428 static void efx_rx_packet_lro(struct efx_channel *channel,
429 struct efx_rx_buffer *rx_buf,
432 struct napi_struct *napi = &channel->napi_str;
433 gro_result_t gro_result;
435 /* Pass the skb/page into the LRO engine */
437 struct page *page = rx_buf->page;
440 EFX_BUG_ON_PARANOID(rx_buf->skb);
443 skb = napi_get_frags(napi);
449 skb_shinfo(skb)->frags[0].page = page;
450 skb_shinfo(skb)->frags[0].page_offset =
451 efx_rx_buf_offset(rx_buf);
452 skb_shinfo(skb)->frags[0].size = rx_buf->len;
453 skb_shinfo(skb)->nr_frags = 1;
455 skb->len = rx_buf->len;
456 skb->data_len = rx_buf->len;
457 skb->truesize += rx_buf->len;
459 checksummed ? CHECKSUM_UNNECESSARY : CHECKSUM_NONE;
461 skb_record_rx_queue(skb, channel->channel);
463 gro_result = napi_gro_frags(napi);
465 struct sk_buff *skb = rx_buf->skb;
467 EFX_BUG_ON_PARANOID(!skb);
468 EFX_BUG_ON_PARANOID(!checksummed);
471 gro_result = napi_gro_receive(napi, skb);
474 if (gro_result == GRO_NORMAL) {
475 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
476 } else if (gro_result != GRO_DROP) {
477 channel->rx_alloc_level += RX_ALLOC_FACTOR_LRO;
478 channel->irq_mod_score += 2;
482 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
483 unsigned int len, bool checksummed, bool discard)
485 struct efx_nic *efx = rx_queue->efx;
486 struct efx_channel *channel = rx_queue->channel;
487 struct efx_rx_buffer *rx_buf;
488 bool leak_packet = false;
490 rx_buf = efx_rx_buffer(rx_queue, index);
491 EFX_BUG_ON_PARANOID(!rx_buf->data);
492 EFX_BUG_ON_PARANOID(rx_buf->skb && rx_buf->page);
493 EFX_BUG_ON_PARANOID(!(rx_buf->skb || rx_buf->page));
495 /* This allows the refill path to post another buffer.
496 * EFX_RXD_HEAD_ROOM ensures that the slot we are using
497 * isn't overwritten yet.
499 rx_queue->removed_count++;
501 /* Validate the length encoded in the event vs the descriptor pushed */
502 efx_rx_packet__check_len(rx_queue, rx_buf, len,
503 &discard, &leak_packet);
505 EFX_TRACE(efx, "RX queue %d received id %x at %llx+%x %s%s\n",
506 rx_queue->queue, index,
507 (unsigned long long)rx_buf->dma_addr, len,
508 (checksummed ? " [SUMMED]" : ""),
509 (discard ? " [DISCARD]" : ""));
511 /* Discard packet, if instructed to do so */
512 if (unlikely(discard)) {
513 if (unlikely(leak_packet))
514 channel->n_skbuff_leaks++;
516 efx_recycle_rx_buffer(channel, rx_buf);
518 /* Don't hold off the previous receive */
523 /* Release card resources - assumes all RX buffers consumed in-order
526 efx_unmap_rx_buffer(efx, rx_buf);
528 /* Prefetch nice and early so data will (hopefully) be in cache by
529 * the time we look at it.
531 prefetch(rx_buf->data);
533 /* Pipeline receives so that we give time for packet headers to be
534 * prefetched into cache.
538 if (rx_queue->channel->rx_pkt)
539 __efx_rx_packet(rx_queue->channel,
540 rx_queue->channel->rx_pkt,
541 rx_queue->channel->rx_pkt_csummed);
542 rx_queue->channel->rx_pkt = rx_buf;
543 rx_queue->channel->rx_pkt_csummed = checksummed;
546 /* Handle a received packet. Second half: Touches packet payload. */
547 void __efx_rx_packet(struct efx_channel *channel,
548 struct efx_rx_buffer *rx_buf, bool checksummed)
550 struct efx_nic *efx = channel->efx;
553 /* If we're in loopback test, then pass the packet directly to the
554 * loopback layer, and free the rx_buf here
556 if (unlikely(efx->loopback_selftest)) {
557 efx_loopback_rx_packet(efx, rx_buf->data, rx_buf->len);
558 efx_free_rx_buffer(efx, rx_buf);
563 prefetch(skb_shinfo(rx_buf->skb));
565 skb_put(rx_buf->skb, rx_buf->len);
567 /* Move past the ethernet header. rx_buf->data still points
568 * at the ethernet header */
569 rx_buf->skb->protocol = eth_type_trans(rx_buf->skb,
572 skb_record_rx_queue(rx_buf->skb, channel->channel);
575 if (likely(checksummed || rx_buf->page)) {
576 efx_rx_packet_lro(channel, rx_buf, checksummed);
580 /* We now own the SKB */
583 EFX_BUG_ON_PARANOID(!skb);
585 /* Set the SKB flags */
586 skb->ip_summed = CHECKSUM_NONE;
588 /* Pass the packet up */
589 netif_receive_skb(skb);
591 /* Update allocation strategy method */
592 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
595 void efx_rx_strategy(struct efx_channel *channel)
597 enum efx_rx_alloc_method method = rx_alloc_method;
599 /* Only makes sense to use page based allocation if LRO is enabled */
600 if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
601 method = RX_ALLOC_METHOD_SKB;
602 } else if (method == RX_ALLOC_METHOD_AUTO) {
603 /* Constrain the rx_alloc_level */
604 if (channel->rx_alloc_level < 0)
605 channel->rx_alloc_level = 0;
606 else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
607 channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
609 /* Decide on the allocation method */
610 method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_LRO) ?
611 RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
614 /* Push the option */
615 channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
618 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
620 struct efx_nic *efx = rx_queue->efx;
621 unsigned int rxq_size;
624 EFX_LOG(efx, "creating RX queue %d\n", rx_queue->queue);
626 /* Allocate RX buffers */
627 rxq_size = EFX_RXQ_SIZE * sizeof(*rx_queue->buffer);
628 rx_queue->buffer = kzalloc(rxq_size, GFP_KERNEL);
629 if (!rx_queue->buffer)
632 rc = efx_nic_probe_rx(rx_queue);
634 kfree(rx_queue->buffer);
635 rx_queue->buffer = NULL;
640 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
642 unsigned int max_fill, trigger, limit;
644 EFX_LOG(rx_queue->efx, "initialising RX queue %d\n", rx_queue->queue);
646 /* Initialise ptr fields */
647 rx_queue->added_count = 0;
648 rx_queue->notified_count = 0;
649 rx_queue->removed_count = 0;
650 rx_queue->min_fill = -1U;
651 rx_queue->min_overfill = -1U;
653 /* Initialise limit fields */
654 max_fill = EFX_RXQ_SIZE - EFX_RXD_HEAD_ROOM;
655 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
656 limit = max_fill * min(rx_refill_limit, 100U) / 100U;
658 rx_queue->max_fill = max_fill;
659 rx_queue->fast_fill_trigger = trigger;
660 rx_queue->fast_fill_limit = limit;
662 /* Set up RX descriptor ring */
663 efx_nic_init_rx(rx_queue);
666 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
669 struct efx_rx_buffer *rx_buf;
671 EFX_LOG(rx_queue->efx, "shutting down RX queue %d\n", rx_queue->queue);
673 del_timer_sync(&rx_queue->slow_fill);
674 efx_nic_fini_rx(rx_queue);
676 /* Release RX buffers NB start at index 0 not current HW ptr */
677 if (rx_queue->buffer) {
678 for (i = 0; i <= EFX_RXQ_MASK; i++) {
679 rx_buf = efx_rx_buffer(rx_queue, i);
680 efx_fini_rx_buffer(rx_queue, rx_buf);
685 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
687 EFX_LOG(rx_queue->efx, "destroying RX queue %d\n", rx_queue->queue);
689 efx_nic_remove_rx(rx_queue);
691 kfree(rx_queue->buffer);
692 rx_queue->buffer = NULL;
696 module_param(rx_alloc_method, int, 0644);
697 MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
699 module_param(rx_refill_threshold, uint, 0444);
700 MODULE_PARM_DESC(rx_refill_threshold,
701 "RX descriptor ring fast/slow fill threshold (%)");