1 /****************************************************************************
2 * Driver for Solarflare Solarstorm network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2008 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>
14 #include <linux/tcp.h>
15 #include <linux/udp.h>
17 #include <net/checksum.h>
18 #include "net_driver.h"
23 #include "workarounds.h"
25 /* Number of RX descriptors pushed at once. */
26 #define EFX_RX_BATCH 8
28 /* Size of buffer allocated for skb header area. */
29 #define EFX_SKB_HEADERS 64u
32 * rx_alloc_method - RX buffer allocation method
34 * This driver supports two methods for allocating and using RX buffers:
35 * each RX buffer may be backed by an skb or by an order-n page.
37 * When LRO is in use then the second method has a lower overhead,
38 * since we don't have to allocate then free skbs on reassembled frames.
41 * - RX_ALLOC_METHOD_AUTO = 0
42 * - RX_ALLOC_METHOD_SKB = 1
43 * - RX_ALLOC_METHOD_PAGE = 2
45 * The heuristic for %RX_ALLOC_METHOD_AUTO is a simple hysteresis count
46 * controlled by the parameters below.
48 * - Since pushing and popping descriptors are separated by the rx_queue
49 * size, so the watermarks should be ~rxd_size.
50 * - The performance win by using page-based allocation for LRO is less
51 * than the performance hit of using page-based allocation of non-LRO,
52 * so the watermarks should reflect this.
54 * Per channel we maintain a single variable, updated by each channel:
56 * rx_alloc_level += (lro_performed ? RX_ALLOC_FACTOR_LRO :
57 * RX_ALLOC_FACTOR_SKB)
58 * Per NAPI poll interval, we constrain rx_alloc_level to 0..MAX (which
59 * limits the hysteresis), and update the allocation strategy:
61 * rx_alloc_method = (rx_alloc_level > RX_ALLOC_LEVEL_LRO ?
62 * RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB)
64 static int rx_alloc_method = RX_ALLOC_METHOD_PAGE;
66 #define RX_ALLOC_LEVEL_LRO 0x2000
67 #define RX_ALLOC_LEVEL_MAX 0x3000
68 #define RX_ALLOC_FACTOR_LRO 1
69 #define RX_ALLOC_FACTOR_SKB (-2)
71 /* This is the percentage fill level below which new RX descriptors
72 * will be added to the RX descriptor ring.
74 static unsigned int rx_refill_threshold = 90;
76 /* This is the percentage fill level to which an RX queue will be refilled
77 * when the "RX refill threshold" is reached.
79 static unsigned int rx_refill_limit = 95;
82 * RX maximum head room required.
84 * This must be at least 1 to prevent overflow and at least 2 to allow
87 #define EFX_RXD_HEAD_ROOM 2
89 static inline unsigned int efx_rx_buf_offset(struct efx_rx_buffer *buf)
91 /* Offset is always within one page, so we don't need to consider
94 return (__force unsigned long) buf->data & (PAGE_SIZE - 1);
96 static inline unsigned int efx_rx_buf_size(struct efx_nic *efx)
98 return PAGE_SIZE << efx->rx_buffer_order;
103 * efx_init_rx_buffer_skb - create new RX buffer using skb-based allocation
105 * @rx_queue: Efx RX queue
106 * @rx_buf: RX buffer structure to populate
108 * This allocates memory for a new receive buffer, maps it for DMA,
109 * and populates a struct efx_rx_buffer with the relevant
110 * information. Return a negative error code or 0 on success.
112 static int efx_init_rx_buffer_skb(struct efx_rx_queue *rx_queue,
113 struct efx_rx_buffer *rx_buf)
115 struct efx_nic *efx = rx_queue->efx;
116 struct net_device *net_dev = efx->net_dev;
117 int skb_len = efx->rx_buffer_len;
119 rx_buf->skb = netdev_alloc_skb(net_dev, skb_len);
120 if (unlikely(!rx_buf->skb))
123 /* Adjust the SKB for padding and checksum */
124 skb_reserve(rx_buf->skb, NET_IP_ALIGN);
125 rx_buf->len = skb_len - NET_IP_ALIGN;
126 rx_buf->data = (char *)rx_buf->skb->data;
127 rx_buf->skb->ip_summed = CHECKSUM_UNNECESSARY;
129 rx_buf->dma_addr = pci_map_single(efx->pci_dev,
130 rx_buf->data, rx_buf->len,
133 if (unlikely(pci_dma_mapping_error(efx->pci_dev, rx_buf->dma_addr))) {
134 dev_kfree_skb_any(rx_buf->skb);
143 * efx_init_rx_buffer_page - create new RX buffer using page-based allocation
145 * @rx_queue: Efx RX queue
146 * @rx_buf: RX buffer structure to populate
148 * This allocates memory for a new receive buffer, maps it for DMA,
149 * and populates a struct efx_rx_buffer with the relevant
150 * information. Return a negative error code or 0 on success.
152 static int efx_init_rx_buffer_page(struct efx_rx_queue *rx_queue,
153 struct efx_rx_buffer *rx_buf)
155 struct efx_nic *efx = rx_queue->efx;
156 int bytes, space, offset;
158 bytes = efx->rx_buffer_len - EFX_PAGE_IP_ALIGN;
160 /* If there is space left in the previously allocated page,
161 * then use it. Otherwise allocate a new one */
162 rx_buf->page = rx_queue->buf_page;
163 if (rx_buf->page == NULL) {
166 rx_buf->page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
167 efx->rx_buffer_order);
168 if (unlikely(rx_buf->page == NULL))
171 dma_addr = pci_map_page(efx->pci_dev, rx_buf->page,
172 0, efx_rx_buf_size(efx),
175 if (unlikely(pci_dma_mapping_error(efx->pci_dev, dma_addr))) {
176 __free_pages(rx_buf->page, efx->rx_buffer_order);
181 rx_queue->buf_page = rx_buf->page;
182 rx_queue->buf_dma_addr = dma_addr;
183 rx_queue->buf_data = (page_address(rx_buf->page) +
188 rx_buf->data = rx_queue->buf_data;
189 offset = efx_rx_buf_offset(rx_buf);
190 rx_buf->dma_addr = rx_queue->buf_dma_addr + offset;
192 /* Try to pack multiple buffers per page */
193 if (efx->rx_buffer_order == 0) {
194 /* The next buffer starts on the next 512 byte boundary */
195 rx_queue->buf_data += ((bytes + 0x1ff) & ~0x1ff);
196 offset += ((bytes + 0x1ff) & ~0x1ff);
198 space = efx_rx_buf_size(efx) - offset;
199 if (space >= bytes) {
200 /* Refs dropped on kernel releasing each skb */
201 get_page(rx_queue->buf_page);
206 /* This is the final RX buffer for this page, so mark it for
208 rx_queue->buf_page = NULL;
209 rx_buf->unmap_addr = rx_queue->buf_dma_addr;
215 /* This allocates memory for a new receive buffer, maps it for DMA,
216 * and populates a struct efx_rx_buffer with the relevant
219 static int efx_init_rx_buffer(struct efx_rx_queue *rx_queue,
220 struct efx_rx_buffer *new_rx_buf)
224 if (rx_queue->channel->rx_alloc_push_pages) {
225 new_rx_buf->skb = NULL;
226 rc = efx_init_rx_buffer_page(rx_queue, new_rx_buf);
227 rx_queue->alloc_page_count++;
229 new_rx_buf->page = NULL;
230 rc = efx_init_rx_buffer_skb(rx_queue, new_rx_buf);
231 rx_queue->alloc_skb_count++;
234 if (unlikely(rc < 0))
235 EFX_LOG_RL(rx_queue->efx, "%s RXQ[%d] =%d\n", __func__,
236 rx_queue->queue, rc);
240 static void efx_unmap_rx_buffer(struct efx_nic *efx,
241 struct efx_rx_buffer *rx_buf)
244 EFX_BUG_ON_PARANOID(rx_buf->skb);
245 if (rx_buf->unmap_addr) {
246 pci_unmap_page(efx->pci_dev, rx_buf->unmap_addr,
247 efx_rx_buf_size(efx),
249 rx_buf->unmap_addr = 0;
251 } else if (likely(rx_buf->skb)) {
252 pci_unmap_single(efx->pci_dev, rx_buf->dma_addr,
253 rx_buf->len, PCI_DMA_FROMDEVICE);
257 static void efx_free_rx_buffer(struct efx_nic *efx,
258 struct efx_rx_buffer *rx_buf)
261 __free_pages(rx_buf->page, efx->rx_buffer_order);
263 } else if (likely(rx_buf->skb)) {
264 dev_kfree_skb_any(rx_buf->skb);
269 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
270 struct efx_rx_buffer *rx_buf)
272 efx_unmap_rx_buffer(rx_queue->efx, rx_buf);
273 efx_free_rx_buffer(rx_queue->efx, rx_buf);
277 * efx_fast_push_rx_descriptors - push new RX descriptors quickly
278 * @rx_queue: RX descriptor queue
279 * @retry: Recheck the fill level
280 * This will aim to fill the RX descriptor queue up to
281 * @rx_queue->@fast_fill_limit. If there is insufficient atomic
282 * memory to do so, the caller should retry.
284 static int __efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue,
287 struct efx_rx_buffer *rx_buf;
288 unsigned fill_level, index;
289 int i, space, rc = 0;
291 /* Calculate current fill level. Do this outside the lock,
292 * because most of the time we'll end up not wanting to do the
295 fill_level = (rx_queue->added_count - rx_queue->removed_count);
296 EFX_BUG_ON_PARANOID(fill_level >
297 rx_queue->efx->type->rxd_ring_mask + 1);
299 /* Don't fill if we don't need to */
300 if (fill_level >= rx_queue->fast_fill_trigger)
303 /* Record minimum fill level */
304 if (unlikely(fill_level < rx_queue->min_fill)) {
306 rx_queue->min_fill = fill_level;
309 /* Acquire RX add lock. If this lock is contended, then a fast
310 * fill must already be in progress (e.g. in the refill
311 * tasklet), so we don't need to do anything
313 if (!spin_trylock_bh(&rx_queue->add_lock))
317 /* Recalculate current fill level now that we have the lock */
318 fill_level = (rx_queue->added_count - rx_queue->removed_count);
319 EFX_BUG_ON_PARANOID(fill_level >
320 rx_queue->efx->type->rxd_ring_mask + 1);
321 space = rx_queue->fast_fill_limit - fill_level;
322 if (space < EFX_RX_BATCH)
325 EFX_TRACE(rx_queue->efx, "RX queue %d fast-filling descriptor ring from"
326 " level %d to level %d using %s allocation\n",
327 rx_queue->queue, fill_level, rx_queue->fast_fill_limit,
328 rx_queue->channel->rx_alloc_push_pages ? "page" : "skb");
331 for (i = 0; i < EFX_RX_BATCH; ++i) {
332 index = (rx_queue->added_count &
333 rx_queue->efx->type->rxd_ring_mask);
334 rx_buf = efx_rx_buffer(rx_queue, index);
335 rc = efx_init_rx_buffer(rx_queue, rx_buf);
338 ++rx_queue->added_count;
340 } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
342 EFX_TRACE(rx_queue->efx, "RX queue %d fast-filled descriptor ring "
343 "to level %d\n", rx_queue->queue,
344 rx_queue->added_count - rx_queue->removed_count);
347 /* Send write pointer to card. */
348 falcon_notify_rx_desc(rx_queue);
350 /* If the fast fill is running inside from the refill tasklet, then
351 * for SMP systems it may be running on a different CPU to
352 * RX event processing, which means that the fill level may now be
354 if (unlikely(retry && (rc == 0)))
358 spin_unlock_bh(&rx_queue->add_lock);
364 * efx_fast_push_rx_descriptors - push new RX descriptors quickly
365 * @rx_queue: RX descriptor queue
367 * This will aim to fill the RX descriptor queue up to
368 * @rx_queue->@fast_fill_limit. If there is insufficient memory to do so,
369 * it will schedule a work item to immediately continue the fast fill
371 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
375 rc = __efx_fast_push_rx_descriptors(rx_queue, 0);
377 /* Schedule the work item to run immediately. The hope is
378 * that work is immediately pending to free some memory
379 * (e.g. an RX event or TX completion)
381 efx_schedule_slow_fill(rx_queue, 0);
385 void efx_rx_work(struct work_struct *data)
387 struct efx_rx_queue *rx_queue;
390 rx_queue = container_of(data, struct efx_rx_queue, work.work);
392 if (unlikely(!rx_queue->channel->enabled))
395 EFX_TRACE(rx_queue->efx, "RX queue %d worker thread executing on CPU "
396 "%d\n", rx_queue->queue, raw_smp_processor_id());
398 ++rx_queue->slow_fill_count;
399 /* Push new RX descriptors, allowing at least 1 jiffy for
400 * the kernel to free some more memory. */
401 rc = __efx_fast_push_rx_descriptors(rx_queue, 1);
403 efx_schedule_slow_fill(rx_queue, 1);
406 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
407 struct efx_rx_buffer *rx_buf,
408 int len, bool *discard,
411 struct efx_nic *efx = rx_queue->efx;
412 unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
414 if (likely(len <= max_len))
417 /* The packet must be discarded, but this is only a fatal error
418 * if the caller indicated it was
422 if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
423 EFX_ERR_RL(efx, " RX queue %d seriously overlength "
424 "RX event (0x%x > 0x%x+0x%x). Leaking\n",
425 rx_queue->queue, len, max_len,
426 efx->type->rx_buffer_padding);
427 /* If this buffer was skb-allocated, then the meta
428 * data at the end of the skb will be trashed. So
429 * we have no choice but to leak the fragment.
431 *leak_packet = (rx_buf->skb != NULL);
432 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
434 EFX_ERR_RL(efx, " RX queue %d overlength RX event "
435 "(0x%x > 0x%x)\n", rx_queue->queue, len, max_len);
438 rx_queue->channel->n_rx_overlength++;
441 /* Pass a received packet up through the generic LRO stack
443 * Handles driverlink veto, and passes the fragment up via
444 * the appropriate LRO method
446 static void efx_rx_packet_lro(struct efx_channel *channel,
447 struct efx_rx_buffer *rx_buf)
449 struct napi_struct *napi = &channel->napi_str;
451 /* Pass the skb/page into the LRO engine */
453 struct napi_gro_fraginfo info;
455 info.frags[0].page = rx_buf->page;
456 info.frags[0].page_offset = efx_rx_buf_offset(rx_buf);
457 info.frags[0].size = rx_buf->len;
459 info.ip_summed = CHECKSUM_UNNECESSARY;
460 info.len = rx_buf->len;
462 napi_gro_frags(napi, &info);
464 EFX_BUG_ON_PARANOID(rx_buf->skb);
467 EFX_BUG_ON_PARANOID(!rx_buf->skb);
469 napi_gro_receive(napi, rx_buf->skb);
474 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
475 unsigned int len, bool checksummed, bool discard)
477 struct efx_nic *efx = rx_queue->efx;
478 struct efx_rx_buffer *rx_buf;
479 bool leak_packet = false;
481 rx_buf = efx_rx_buffer(rx_queue, index);
482 EFX_BUG_ON_PARANOID(!rx_buf->data);
483 EFX_BUG_ON_PARANOID(rx_buf->skb && rx_buf->page);
484 EFX_BUG_ON_PARANOID(!(rx_buf->skb || rx_buf->page));
486 /* This allows the refill path to post another buffer.
487 * EFX_RXD_HEAD_ROOM ensures that the slot we are using
488 * isn't overwritten yet.
490 rx_queue->removed_count++;
492 /* Validate the length encoded in the event vs the descriptor pushed */
493 efx_rx_packet__check_len(rx_queue, rx_buf, len,
494 &discard, &leak_packet);
496 EFX_TRACE(efx, "RX queue %d received id %x at %llx+%x %s%s\n",
497 rx_queue->queue, index,
498 (unsigned long long)rx_buf->dma_addr, len,
499 (checksummed ? " [SUMMED]" : ""),
500 (discard ? " [DISCARD]" : ""));
502 /* Discard packet, if instructed to do so */
503 if (unlikely(discard)) {
504 if (unlikely(leak_packet))
505 rx_queue->channel->n_skbuff_leaks++;
507 /* We haven't called efx_unmap_rx_buffer yet,
508 * so fini the entire rx_buffer here */
509 efx_fini_rx_buffer(rx_queue, rx_buf);
513 /* Release card resources - assumes all RX buffers consumed in-order
516 efx_unmap_rx_buffer(efx, rx_buf);
518 /* Prefetch nice and early so data will (hopefully) be in cache by
519 * the time we look at it.
521 prefetch(rx_buf->data);
523 /* Pipeline receives so that we give time for packet headers to be
524 * prefetched into cache.
527 if (rx_queue->channel->rx_pkt)
528 __efx_rx_packet(rx_queue->channel,
529 rx_queue->channel->rx_pkt,
530 rx_queue->channel->rx_pkt_csummed);
531 rx_queue->channel->rx_pkt = rx_buf;
532 rx_queue->channel->rx_pkt_csummed = checksummed;
535 /* Handle a received packet. Second half: Touches packet payload. */
536 void __efx_rx_packet(struct efx_channel *channel,
537 struct efx_rx_buffer *rx_buf, bool checksummed)
539 struct efx_nic *efx = channel->efx;
542 /* If we're in loopback test, then pass the packet directly to the
543 * loopback layer, and free the rx_buf here
545 if (unlikely(efx->loopback_selftest)) {
546 efx_loopback_rx_packet(efx, rx_buf->data, rx_buf->len);
547 efx_free_rx_buffer(efx, rx_buf);
552 prefetch(skb_shinfo(rx_buf->skb));
554 skb_put(rx_buf->skb, rx_buf->len);
556 /* Move past the ethernet header. rx_buf->data still points
557 * at the ethernet header */
558 rx_buf->skb->protocol = eth_type_trans(rx_buf->skb,
562 if (likely(checksummed || rx_buf->page)) {
563 efx_rx_packet_lro(channel, rx_buf);
567 /* We now own the SKB */
571 EFX_BUG_ON_PARANOID(rx_buf->page);
572 EFX_BUG_ON_PARANOID(rx_buf->skb);
573 EFX_BUG_ON_PARANOID(!skb);
575 /* Set the SKB flags */
576 skb->ip_summed = CHECKSUM_NONE;
578 skb_record_rx_queue(skb, channel->channel);
580 /* Pass the packet up */
581 netif_receive_skb(skb);
583 /* Update allocation strategy method */
584 channel->rx_alloc_level += RX_ALLOC_FACTOR_SKB;
590 void efx_rx_strategy(struct efx_channel *channel)
592 enum efx_rx_alloc_method method = rx_alloc_method;
594 /* Only makes sense to use page based allocation if LRO is enabled */
595 if (!(channel->efx->net_dev->features & NETIF_F_GRO)) {
596 method = RX_ALLOC_METHOD_SKB;
597 } else if (method == RX_ALLOC_METHOD_AUTO) {
598 /* Constrain the rx_alloc_level */
599 if (channel->rx_alloc_level < 0)
600 channel->rx_alloc_level = 0;
601 else if (channel->rx_alloc_level > RX_ALLOC_LEVEL_MAX)
602 channel->rx_alloc_level = RX_ALLOC_LEVEL_MAX;
604 /* Decide on the allocation method */
605 method = ((channel->rx_alloc_level > RX_ALLOC_LEVEL_LRO) ?
606 RX_ALLOC_METHOD_PAGE : RX_ALLOC_METHOD_SKB);
609 /* Push the option */
610 channel->rx_alloc_push_pages = (method == RX_ALLOC_METHOD_PAGE);
613 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
615 struct efx_nic *efx = rx_queue->efx;
616 unsigned int rxq_size;
619 EFX_LOG(efx, "creating RX queue %d\n", rx_queue->queue);
621 /* Allocate RX buffers */
622 rxq_size = (efx->type->rxd_ring_mask + 1) * sizeof(*rx_queue->buffer);
623 rx_queue->buffer = kzalloc(rxq_size, GFP_KERNEL);
624 if (!rx_queue->buffer)
627 rc = falcon_probe_rx(rx_queue);
629 kfree(rx_queue->buffer);
630 rx_queue->buffer = NULL;
635 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
637 struct efx_nic *efx = rx_queue->efx;
638 unsigned int max_fill, trigger, limit;
640 EFX_LOG(rx_queue->efx, "initialising RX queue %d\n", rx_queue->queue);
642 /* Initialise ptr fields */
643 rx_queue->added_count = 0;
644 rx_queue->notified_count = 0;
645 rx_queue->removed_count = 0;
646 rx_queue->min_fill = -1U;
647 rx_queue->min_overfill = -1U;
649 /* Initialise limit fields */
650 max_fill = efx->type->rxd_ring_mask + 1 - EFX_RXD_HEAD_ROOM;
651 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
652 limit = max_fill * min(rx_refill_limit, 100U) / 100U;
654 rx_queue->max_fill = max_fill;
655 rx_queue->fast_fill_trigger = trigger;
656 rx_queue->fast_fill_limit = limit;
658 /* Set up RX descriptor ring */
659 falcon_init_rx(rx_queue);
662 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
665 struct efx_rx_buffer *rx_buf;
667 EFX_LOG(rx_queue->efx, "shutting down RX queue %d\n", rx_queue->queue);
669 falcon_fini_rx(rx_queue);
671 /* Release RX buffers NB start at index 0 not current HW ptr */
672 if (rx_queue->buffer) {
673 for (i = 0; i <= rx_queue->efx->type->rxd_ring_mask; i++) {
674 rx_buf = efx_rx_buffer(rx_queue, i);
675 efx_fini_rx_buffer(rx_queue, rx_buf);
679 /* For a page that is part-way through splitting into RX buffers */
680 if (rx_queue->buf_page != NULL) {
681 pci_unmap_page(rx_queue->efx->pci_dev, rx_queue->buf_dma_addr,
682 efx_rx_buf_size(rx_queue->efx),
684 __free_pages(rx_queue->buf_page,
685 rx_queue->efx->rx_buffer_order);
686 rx_queue->buf_page = NULL;
690 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
692 EFX_LOG(rx_queue->efx, "destroying RX queue %d\n", rx_queue->queue);
694 falcon_remove_rx(rx_queue);
696 kfree(rx_queue->buffer);
697 rx_queue->buffer = NULL;
701 module_param(rx_alloc_method, int, 0644);
702 MODULE_PARM_DESC(rx_alloc_method, "Allocation method used for RX buffers");
704 module_param(rx_refill_threshold, uint, 0444);
705 MODULE_PARM_DESC(rx_refill_threshold,
706 "RX descriptor ring fast/slow fill threshold (%)");