1 /****************************************************************************
2 * Driver for Solarflare Solarstorm network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2011 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>
17 #include <linux/prefetch.h>
18 #include <linux/moduleparam.h>
20 #include <net/checksum.h>
21 #include "net_driver.h"
25 #include "workarounds.h"
27 /* Number of RX descriptors pushed at once. */
28 #define EFX_RX_BATCH 8
30 /* Maximum length for an RX descriptor sharing a page */
31 #define EFX_RX_HALF_PAGE ((PAGE_SIZE >> 1) - sizeof(struct efx_rx_page_state) \
34 /* Size of buffer allocated for skb header area. */
35 #define EFX_SKB_HEADERS 64u
37 /* This is the percentage fill level below which new RX descriptors
38 * will be added to the RX descriptor ring.
40 static unsigned int rx_refill_threshold;
43 * RX maximum head room required.
45 * This must be at least 1 to prevent overflow and at least 2 to allow
48 #define EFX_RXD_HEAD_ROOM 2
50 /* Offset of ethernet header within page */
51 static inline unsigned int efx_rx_buf_offset(struct efx_nic *efx,
52 struct efx_rx_buffer *buf)
54 return buf->page_offset + efx->type->rx_buffer_hash_size;
57 static inline u8 *efx_rx_buf_va(struct efx_rx_buffer *buf)
59 return page_address(buf->page) + buf->page_offset;
62 static inline u32 efx_rx_buf_hash(const u8 *eh)
64 /* The ethernet header is always directly after any hash. */
65 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) || NET_IP_ALIGN % 4 == 0
66 return __le32_to_cpup((const __le32 *)(eh - 4));
68 const u8 *data = eh - 4;
77 * efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers
79 * @rx_queue: Efx RX queue
81 * This allocates memory for EFX_RX_BATCH receive buffers, maps them for DMA,
82 * and populates struct efx_rx_buffers for each one. Return a negative error
83 * code or 0 on success. If a single page can be split between two buffers,
84 * then the page will either be inserted fully, or not at at all.
86 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue)
88 struct efx_nic *efx = rx_queue->efx;
89 struct efx_rx_buffer *rx_buf;
91 unsigned int page_offset;
92 struct efx_rx_page_state *state;
94 unsigned index, count;
96 /* We can split a page between two buffers */
97 BUILD_BUG_ON(EFX_RX_BATCH & 1);
99 for (count = 0; count < EFX_RX_BATCH; ++count) {
100 page = alloc_pages(__GFP_COLD | __GFP_COMP | GFP_ATOMIC,
101 efx->rx_buffer_order);
102 if (unlikely(page == NULL))
104 dma_addr = dma_map_page(&efx->pci_dev->dev, page, 0,
105 PAGE_SIZE << efx->rx_buffer_order,
107 if (unlikely(dma_mapping_error(&efx->pci_dev->dev, dma_addr))) {
108 __free_pages(page, efx->rx_buffer_order);
111 state = page_address(page);
113 state->dma_addr = dma_addr;
115 dma_addr += sizeof(struct efx_rx_page_state);
116 page_offset = sizeof(struct efx_rx_page_state);
119 index = rx_queue->added_count & rx_queue->ptr_mask;
120 rx_buf = efx_rx_buffer(rx_queue, index);
121 rx_buf->dma_addr = dma_addr + EFX_PAGE_IP_ALIGN;
123 rx_buf->page_offset = page_offset + EFX_PAGE_IP_ALIGN;
124 rx_buf->len = efx->rx_dma_len;
126 ++rx_queue->added_count;
129 if ((~count & 1) && (efx->rx_dma_len <= EFX_RX_HALF_PAGE)) {
130 /* Use the second half of the page */
132 dma_addr += (PAGE_SIZE >> 1);
133 page_offset += (PAGE_SIZE >> 1);
142 static void efx_unmap_rx_buffer(struct efx_nic *efx,
143 struct efx_rx_buffer *rx_buf,
144 unsigned int used_len)
147 struct efx_rx_page_state *state;
149 state = page_address(rx_buf->page);
150 if (--state->refcnt == 0) {
151 dma_unmap_page(&efx->pci_dev->dev,
153 PAGE_SIZE << efx->rx_buffer_order,
155 } else if (used_len) {
156 dma_sync_single_for_cpu(&efx->pci_dev->dev,
157 rx_buf->dma_addr, used_len,
163 static void efx_free_rx_buffer(struct efx_nic *efx,
164 struct efx_rx_buffer *rx_buf)
167 __free_pages(rx_buf->page, efx->rx_buffer_order);
172 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue,
173 struct efx_rx_buffer *rx_buf)
175 efx_unmap_rx_buffer(rx_queue->efx, rx_buf, 0);
176 efx_free_rx_buffer(rx_queue->efx, rx_buf);
179 /* Attempt to resurrect the other receive buffer that used to share this page,
180 * which had previously been passed up to the kernel and freed. */
181 static void efx_resurrect_rx_buffer(struct efx_rx_queue *rx_queue,
182 struct efx_rx_buffer *rx_buf)
184 struct efx_rx_page_state *state = page_address(rx_buf->page);
185 struct efx_rx_buffer *new_buf;
186 unsigned fill_level, index;
188 /* +1 because efx_rx_packet() incremented removed_count. +1 because
189 * we'd like to insert an additional descriptor whilst leaving
190 * EFX_RXD_HEAD_ROOM for the non-recycle path */
191 fill_level = (rx_queue->added_count - rx_queue->removed_count + 2);
192 if (unlikely(fill_level > rx_queue->max_fill)) {
193 /* We could place "state" on a list, and drain the list in
194 * efx_fast_push_rx_descriptors(). For now, this will do. */
199 get_page(rx_buf->page);
201 index = rx_queue->added_count & rx_queue->ptr_mask;
202 new_buf = efx_rx_buffer(rx_queue, index);
203 new_buf->dma_addr = rx_buf->dma_addr ^ (PAGE_SIZE >> 1);
204 new_buf->page = rx_buf->page;
205 new_buf->len = rx_buf->len;
206 ++rx_queue->added_count;
209 /* Recycle the given rx buffer directly back into the rx_queue. There is
210 * always room to add this buffer, because we've just popped a buffer. */
211 static void efx_recycle_rx_buffer(struct efx_channel *channel,
212 struct efx_rx_buffer *rx_buf)
214 struct efx_nic *efx = channel->efx;
215 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel);
216 struct efx_rx_buffer *new_buf;
221 if (efx->rx_dma_len <= EFX_RX_HALF_PAGE &&
222 page_count(rx_buf->page) == 1)
223 efx_resurrect_rx_buffer(rx_queue, rx_buf);
225 index = rx_queue->added_count & rx_queue->ptr_mask;
226 new_buf = efx_rx_buffer(rx_queue, index);
228 memcpy(new_buf, rx_buf, sizeof(*new_buf));
230 ++rx_queue->added_count;
234 * efx_fast_push_rx_descriptors - push new RX descriptors quickly
235 * @rx_queue: RX descriptor queue
237 * This will aim to fill the RX descriptor queue up to
238 * @rx_queue->@max_fill. If there is insufficient atomic
239 * memory to do so, a slow fill will be scheduled.
241 * The caller must provide serialisation (none is used here). In practise,
242 * this means this function must run from the NAPI handler, or be called
243 * when NAPI is disabled.
245 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue)
250 /* Calculate current fill level, and exit if we don't need to fill */
251 fill_level = (rx_queue->added_count - rx_queue->removed_count);
252 EFX_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries);
253 if (fill_level >= rx_queue->fast_fill_trigger)
256 /* Record minimum fill level */
257 if (unlikely(fill_level < rx_queue->min_fill)) {
259 rx_queue->min_fill = fill_level;
262 space = rx_queue->max_fill - fill_level;
263 EFX_BUG_ON_PARANOID(space < EFX_RX_BATCH);
265 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
266 "RX queue %d fast-filling descriptor ring from"
267 " level %d to level %d\n",
268 efx_rx_queue_index(rx_queue), fill_level,
273 rc = efx_init_rx_buffers(rx_queue);
275 /* Ensure that we don't leave the rx queue empty */
276 if (rx_queue->added_count == rx_queue->removed_count)
277 efx_schedule_slow_fill(rx_queue);
280 } while ((space -= EFX_RX_BATCH) >= EFX_RX_BATCH);
282 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev,
283 "RX queue %d fast-filled descriptor ring "
284 "to level %d\n", efx_rx_queue_index(rx_queue),
285 rx_queue->added_count - rx_queue->removed_count);
288 if (rx_queue->notified_count != rx_queue->added_count)
289 efx_nic_notify_rx_desc(rx_queue);
292 void efx_rx_slow_fill(unsigned long context)
294 struct efx_rx_queue *rx_queue = (struct efx_rx_queue *)context;
296 /* Post an event to cause NAPI to run and refill the queue */
297 efx_nic_generate_fill_event(rx_queue);
298 ++rx_queue->slow_fill_count;
301 static void efx_rx_packet__check_len(struct efx_rx_queue *rx_queue,
302 struct efx_rx_buffer *rx_buf,
305 struct efx_nic *efx = rx_queue->efx;
306 unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding;
308 if (likely(len <= max_len))
311 /* The packet must be discarded, but this is only a fatal error
312 * if the caller indicated it was
314 rx_buf->flags |= EFX_RX_PKT_DISCARD;
316 if ((len > rx_buf->len) && EFX_WORKAROUND_8071(efx)) {
318 netif_err(efx, rx_err, efx->net_dev,
319 " RX queue %d seriously overlength "
320 "RX event (0x%x > 0x%x+0x%x). Leaking\n",
321 efx_rx_queue_index(rx_queue), len, max_len,
322 efx->type->rx_buffer_padding);
323 efx_schedule_reset(efx, RESET_TYPE_RX_RECOVERY);
326 netif_err(efx, rx_err, efx->net_dev,
327 " RX queue %d overlength RX event "
329 efx_rx_queue_index(rx_queue), len, max_len);
332 efx_rx_queue_channel(rx_queue)->n_rx_overlength++;
335 /* Pass a received packet up through GRO. GRO can handle pages
336 * regardless of checksum state and skbs with a good checksum.
338 static void efx_rx_packet_gro(struct efx_channel *channel,
339 struct efx_rx_buffer *rx_buf,
342 struct napi_struct *napi = &channel->napi_str;
343 gro_result_t gro_result;
344 struct efx_nic *efx = channel->efx;
345 struct page *page = rx_buf->page;
350 skb = napi_get_frags(napi);
356 if (efx->net_dev->features & NETIF_F_RXHASH)
357 skb->rxhash = efx_rx_buf_hash(eh);
359 skb_fill_page_desc(skb, 0, page,
360 efx_rx_buf_offset(efx, rx_buf), rx_buf->len);
362 skb->len = rx_buf->len;
363 skb->data_len = rx_buf->len;
364 skb->truesize += rx_buf->len;
365 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ?
366 CHECKSUM_UNNECESSARY : CHECKSUM_NONE);
368 skb_record_rx_queue(skb, channel->rx_queue.core_index);
370 gro_result = napi_gro_frags(napi);
372 if (gro_result != GRO_DROP)
373 channel->irq_mod_score += 2;
376 /* Allocate and construct an SKB around a struct page.*/
377 static struct sk_buff *efx_rx_mk_skb(struct efx_channel *channel,
378 struct efx_rx_buffer *rx_buf,
381 struct efx_nic *efx = channel->efx;
384 /* Allocate an SKB to store the headers */
385 skb = netdev_alloc_skb(efx->net_dev, hdr_len + EFX_PAGE_SKB_ALIGN);
386 if (unlikely(skb == NULL))
389 EFX_BUG_ON_PARANOID(rx_buf->len < hdr_len);
391 skb_reserve(skb, EFX_PAGE_SKB_ALIGN);
393 skb->len = rx_buf->len;
394 skb->truesize = rx_buf->len + sizeof(struct sk_buff);
395 memcpy(skb->data, eh, hdr_len);
396 skb->tail += hdr_len;
398 /* Append the remaining page onto the frag list */
399 if (rx_buf->len > hdr_len) {
400 skb->data_len = skb->len - hdr_len;
401 skb_fill_page_desc(skb, 0, rx_buf->page,
402 efx_rx_buf_offset(efx, rx_buf) + hdr_len,
405 __free_pages(rx_buf->page, efx->rx_buffer_order);
409 /* Ownership has transferred from the rx_buf to skb */
412 /* Move past the ethernet header */
413 skb->protocol = eth_type_trans(skb, efx->net_dev);
418 void efx_rx_packet(struct efx_rx_queue *rx_queue, unsigned int index,
419 unsigned int len, u16 flags)
421 struct efx_nic *efx = rx_queue->efx;
422 struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
423 struct efx_rx_buffer *rx_buf;
425 rx_buf = efx_rx_buffer(rx_queue, index);
426 rx_buf->flags |= flags;
428 /* This allows the refill path to post another buffer.
429 * EFX_RXD_HEAD_ROOM ensures that the slot we are using
430 * isn't overwritten yet.
432 rx_queue->removed_count++;
434 /* Validate the length encoded in the event vs the descriptor pushed */
435 efx_rx_packet__check_len(rx_queue, rx_buf, len);
437 netif_vdbg(efx, rx_status, efx->net_dev,
438 "RX queue %d received id %x at %llx+%x %s%s\n",
439 efx_rx_queue_index(rx_queue), index,
440 (unsigned long long)rx_buf->dma_addr, len,
441 (rx_buf->flags & EFX_RX_PKT_CSUMMED) ? " [SUMMED]" : "",
442 (rx_buf->flags & EFX_RX_PKT_DISCARD) ? " [DISCARD]" : "");
444 /* Discard packet, if instructed to do so */
445 if (unlikely(rx_buf->flags & EFX_RX_PKT_DISCARD)) {
446 efx_recycle_rx_buffer(channel, rx_buf);
448 /* Don't hold off the previous receive */
453 /* Release and/or sync DMA mapping - assumes all RX buffers
454 * consumed in-order per RX queue
456 efx_unmap_rx_buffer(efx, rx_buf, len);
458 /* Prefetch nice and early so data will (hopefully) be in cache by
459 * the time we look at it.
461 prefetch(efx_rx_buf_va(rx_buf) + efx->type->rx_buffer_hash_size);
463 /* Pipeline receives so that we give time for packet headers to be
464 * prefetched into cache.
466 rx_buf->len = len - efx->type->rx_buffer_hash_size;
468 efx_rx_flush_packet(channel);
469 channel->rx_pkt = rx_buf;
472 static void efx_rx_deliver(struct efx_channel *channel, u8 *eh,
473 struct efx_rx_buffer *rx_buf)
476 u16 hdr_len = min_t(u16, rx_buf->len, EFX_SKB_HEADERS);
478 skb = efx_rx_mk_skb(channel, rx_buf, eh, hdr_len);
479 if (unlikely(skb == NULL)) {
480 efx_free_rx_buffer(channel->efx, rx_buf);
483 skb_record_rx_queue(skb, channel->rx_queue.core_index);
485 /* Set the SKB flags */
486 skb_checksum_none_assert(skb);
488 if (channel->type->receive_skb)
489 if (channel->type->receive_skb(channel, skb))
492 /* Pass the packet up */
493 netif_receive_skb(skb);
496 /* Handle a received packet. Second half: Touches packet payload. */
497 void __efx_rx_packet(struct efx_channel *channel, struct efx_rx_buffer *rx_buf)
499 struct efx_nic *efx = channel->efx;
500 u8 *eh = efx_rx_buf_va(rx_buf) + efx->type->rx_buffer_hash_size;
502 /* If we're in loopback test, then pass the packet directly to the
503 * loopback layer, and free the rx_buf here
505 if (unlikely(efx->loopback_selftest)) {
506 efx_loopback_rx_packet(efx, eh, rx_buf->len);
507 efx_free_rx_buffer(efx, rx_buf);
511 if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM)))
512 rx_buf->flags &= ~EFX_RX_PKT_CSUMMED;
514 if (!channel->type->receive_skb)
515 efx_rx_packet_gro(channel, rx_buf, eh);
517 efx_rx_deliver(channel, eh, rx_buf);
520 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue)
522 struct efx_nic *efx = rx_queue->efx;
523 unsigned int entries;
526 /* Create the smallest power-of-two aligned ring */
527 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE);
528 EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
529 rx_queue->ptr_mask = entries - 1;
531 netif_dbg(efx, probe, efx->net_dev,
532 "creating RX queue %d size %#x mask %#x\n",
533 efx_rx_queue_index(rx_queue), efx->rxq_entries,
536 /* Allocate RX buffers */
537 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer),
539 if (!rx_queue->buffer)
542 rc = efx_nic_probe_rx(rx_queue);
544 kfree(rx_queue->buffer);
545 rx_queue->buffer = NULL;
550 void efx_init_rx_queue(struct efx_rx_queue *rx_queue)
552 struct efx_nic *efx = rx_queue->efx;
553 unsigned int max_fill, trigger, max_trigger;
555 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
556 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue));
558 /* Initialise ptr fields */
559 rx_queue->added_count = 0;
560 rx_queue->notified_count = 0;
561 rx_queue->removed_count = 0;
562 rx_queue->min_fill = -1U;
564 /* Initialise limit fields */
565 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM;
566 max_trigger = max_fill - EFX_RX_BATCH;
567 if (rx_refill_threshold != 0) {
568 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U;
569 if (trigger > max_trigger)
570 trigger = max_trigger;
572 trigger = max_trigger;
575 rx_queue->max_fill = max_fill;
576 rx_queue->fast_fill_trigger = trigger;
578 /* Set up RX descriptor ring */
579 rx_queue->enabled = true;
580 efx_nic_init_rx(rx_queue);
583 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue)
586 struct efx_rx_buffer *rx_buf;
588 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
589 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue));
591 /* A flush failure might have left rx_queue->enabled */
592 rx_queue->enabled = false;
594 del_timer_sync(&rx_queue->slow_fill);
595 efx_nic_fini_rx(rx_queue);
597 /* Release RX buffers NB start at index 0 not current HW ptr */
598 if (rx_queue->buffer) {
599 for (i = 0; i <= rx_queue->ptr_mask; i++) {
600 rx_buf = efx_rx_buffer(rx_queue, i);
601 efx_fini_rx_buffer(rx_queue, rx_buf);
606 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue)
608 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev,
609 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue));
611 efx_nic_remove_rx(rx_queue);
613 kfree(rx_queue->buffer);
614 rx_queue->buffer = NULL;
618 module_param(rx_refill_threshold, uint, 0444);
619 MODULE_PARM_DESC(rx_refill_threshold,
620 "RX descriptor ring refill threshold (%)");