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/pci.h>
12 #include <linux/tcp.h>
15 #include <linux/ipv6.h>
16 #include <linux/slab.h>
18 #include <linux/if_ether.h>
19 #include <linux/highmem.h>
20 #include "net_driver.h"
23 #include "workarounds.h"
26 * TX descriptor ring full threshold
28 * The tx_queue descriptor ring fill-level must fall below this value
29 * before we restart the netif queue
31 #define EFX_TXQ_THRESHOLD(_efx) ((_efx)->txq_entries / 2u)
33 static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
34 struct efx_tx_buffer *buffer)
36 if (buffer->unmap_len) {
37 struct pci_dev *pci_dev = tx_queue->efx->pci_dev;
38 dma_addr_t unmap_addr = (buffer->dma_addr + buffer->len -
40 if (buffer->unmap_single)
41 pci_unmap_single(pci_dev, unmap_addr, buffer->unmap_len,
44 pci_unmap_page(pci_dev, unmap_addr, buffer->unmap_len,
46 buffer->unmap_len = 0;
47 buffer->unmap_single = false;
51 dev_kfree_skb_any((struct sk_buff *) buffer->skb);
53 netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
54 "TX queue %d transmission id %x complete\n",
55 tx_queue->queue, tx_queue->read_count);
60 * struct efx_tso_header - a DMA mapped buffer for packet headers
61 * @next: Linked list of free ones.
62 * The list is protected by the TX queue lock.
63 * @dma_unmap_len: Length to unmap for an oversize buffer, or 0.
64 * @dma_addr: The DMA address of the header below.
66 * This controls the memory used for a TSO header. Use TSOH_DATA()
67 * to find the packet header data. Use TSOH_SIZE() to calculate the
68 * total size required for a given packet header length. TSO headers
69 * in the free list are exactly %TSOH_STD_SIZE bytes in size.
71 struct efx_tso_header {
73 struct efx_tso_header *next;
79 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
81 static void efx_fini_tso(struct efx_tx_queue *tx_queue);
82 static void efx_tsoh_heap_free(struct efx_tx_queue *tx_queue,
83 struct efx_tso_header *tsoh);
85 static void efx_tsoh_free(struct efx_tx_queue *tx_queue,
86 struct efx_tx_buffer *buffer)
89 if (likely(!buffer->tsoh->unmap_len)) {
90 buffer->tsoh->next = tx_queue->tso_headers_free;
91 tx_queue->tso_headers_free = buffer->tsoh;
93 efx_tsoh_heap_free(tx_queue, buffer->tsoh);
100 static inline unsigned
101 efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr)
103 /* Depending on the NIC revision, we can use descriptor
104 * lengths up to 8K or 8K-1. However, since PCI Express
105 * devices must split read requests at 4K boundaries, there is
106 * little benefit from using descriptors that cross those
107 * boundaries and we keep things simple by not doing so.
109 unsigned len = (~dma_addr & 0xfff) + 1;
111 /* Work around hardware bug for unaligned buffers. */
112 if (EFX_WORKAROUND_5391(efx) && (dma_addr & 0xf))
113 len = min_t(unsigned, len, 512 - (dma_addr & 0xf));
119 * Add a socket buffer to a TX queue
121 * This maps all fragments of a socket buffer for DMA and adds them to
122 * the TX queue. The queue's insert pointer will be incremented by
123 * the number of fragments in the socket buffer.
125 * If any DMA mapping fails, any mapped fragments will be unmapped,
126 * the queue's insert pointer will be restored to its original value.
128 * This function is split out from efx_hard_start_xmit to allow the
129 * loopback test to direct packets via specific TX queues.
131 * Returns NETDEV_TX_OK or NETDEV_TX_BUSY
132 * You must hold netif_tx_lock() to call this function.
134 netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
136 struct efx_nic *efx = tx_queue->efx;
137 struct pci_dev *pci_dev = efx->pci_dev;
138 struct efx_tx_buffer *buffer;
139 skb_frag_t *fragment;
142 unsigned int len, unmap_len = 0, fill_level, insert_ptr;
143 dma_addr_t dma_addr, unmap_addr = 0;
144 unsigned int dma_len;
147 netdev_tx_t rc = NETDEV_TX_OK;
149 EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
151 if (skb_shinfo(skb)->gso_size)
152 return efx_enqueue_skb_tso(tx_queue, skb);
154 /* Get size of the initial fragment */
155 len = skb_headlen(skb);
157 /* Pad if necessary */
158 if (EFX_WORKAROUND_15592(efx) && skb->len <= 32) {
159 EFX_BUG_ON_PARANOID(skb->data_len);
161 if (skb_pad(skb, len - skb->len))
165 fill_level = tx_queue->insert_count - tx_queue->old_read_count;
166 q_space = efx->txq_entries - 1 - fill_level;
168 /* Map for DMA. Use pci_map_single rather than pci_map_page
169 * since this is more efficient on machines with sparse
173 dma_addr = pci_map_single(pci_dev, skb->data, len, PCI_DMA_TODEVICE);
175 /* Process all fragments */
177 if (unlikely(pci_dma_mapping_error(pci_dev, dma_addr)))
180 /* Store fields for marking in the per-fragment final
183 unmap_addr = dma_addr;
185 /* Add to TX queue, splitting across DMA boundaries */
187 if (unlikely(q_space-- <= 0)) {
188 /* It might be that completions have
189 * happened since the xmit path last
190 * checked. Update the xmit path's
191 * copy of read_count.
193 netif_tx_stop_queue(tx_queue->core_txq);
194 /* This memory barrier protects the
195 * change of queue state from the access
198 tx_queue->old_read_count =
199 ACCESS_ONCE(tx_queue->read_count);
200 fill_level = (tx_queue->insert_count
201 - tx_queue->old_read_count);
202 q_space = efx->txq_entries - 1 - fill_level;
203 if (unlikely(q_space-- <= 0)) {
208 netif_tx_start_queue(tx_queue->core_txq);
211 insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
212 buffer = &tx_queue->buffer[insert_ptr];
213 efx_tsoh_free(tx_queue, buffer);
214 EFX_BUG_ON_PARANOID(buffer->tsoh);
215 EFX_BUG_ON_PARANOID(buffer->skb);
216 EFX_BUG_ON_PARANOID(buffer->len);
217 EFX_BUG_ON_PARANOID(!buffer->continuation);
218 EFX_BUG_ON_PARANOID(buffer->unmap_len);
220 dma_len = efx_max_tx_len(efx, dma_addr);
221 if (likely(dma_len >= len))
224 /* Fill out per descriptor fields */
225 buffer->len = dma_len;
226 buffer->dma_addr = dma_addr;
229 ++tx_queue->insert_count;
232 /* Transfer ownership of the unmapping to the final buffer */
233 buffer->unmap_single = unmap_single;
234 buffer->unmap_len = unmap_len;
237 /* Get address and size of next fragment */
238 if (i >= skb_shinfo(skb)->nr_frags)
240 fragment = &skb_shinfo(skb)->frags[i];
241 len = fragment->size;
242 page = fragment->page;
243 page_offset = fragment->page_offset;
246 unmap_single = false;
247 dma_addr = pci_map_page(pci_dev, page, page_offset, len,
251 /* Transfer ownership of the skb to the final buffer */
253 buffer->continuation = false;
255 /* Pass off to hardware */
256 efx_nic_push_buffers(tx_queue);
261 netif_err(efx, tx_err, efx->net_dev,
262 " TX queue %d could not map skb with %d bytes %d "
263 "fragments for DMA\n", tx_queue->queue, skb->len,
264 skb_shinfo(skb)->nr_frags + 1);
266 /* Mark the packet as transmitted, and free the SKB ourselves */
267 dev_kfree_skb_any(skb);
270 /* Work backwards until we hit the original insert pointer value */
271 while (tx_queue->insert_count != tx_queue->write_count) {
272 --tx_queue->insert_count;
273 insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
274 buffer = &tx_queue->buffer[insert_ptr];
275 efx_dequeue_buffer(tx_queue, buffer);
279 /* Free the fragment we were mid-way through pushing */
282 pci_unmap_single(pci_dev, unmap_addr, unmap_len,
285 pci_unmap_page(pci_dev, unmap_addr, unmap_len,
292 /* Remove packets from the TX queue
294 * This removes packets from the TX queue, up to and including the
297 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
300 struct efx_nic *efx = tx_queue->efx;
301 unsigned int stop_index, read_ptr;
303 stop_index = (index + 1) & tx_queue->ptr_mask;
304 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
306 while (read_ptr != stop_index) {
307 struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
308 if (unlikely(buffer->len == 0)) {
309 netif_err(efx, tx_err, efx->net_dev,
310 "TX queue %d spurious TX completion id %x\n",
311 tx_queue->queue, read_ptr);
312 efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
316 efx_dequeue_buffer(tx_queue, buffer);
317 buffer->continuation = true;
320 ++tx_queue->read_count;
321 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
325 /* Initiate a packet transmission. We use one channel per CPU
326 * (sharing when we have more CPUs than channels). On Falcon, the TX
327 * completion events will be directed back to the CPU that transmitted
328 * the packet, which should be cache-efficient.
330 * Context: non-blocking.
331 * Note that returning anything other than NETDEV_TX_OK will cause the
332 * OS to free the skb.
334 netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
335 struct net_device *net_dev)
337 struct efx_nic *efx = netdev_priv(net_dev);
338 struct efx_tx_queue *tx_queue;
340 if (unlikely(efx->port_inhibited))
341 return NETDEV_TX_BUSY;
343 tx_queue = efx_get_tx_queue(efx, skb_get_queue_mapping(skb),
344 skb->ip_summed == CHECKSUM_PARTIAL ?
345 EFX_TXQ_TYPE_OFFLOAD : 0);
347 return efx_enqueue_skb(tx_queue, skb);
350 void efx_init_tx_queue_core_txq(struct efx_tx_queue *tx_queue)
352 /* Must be inverse of queue lookup in efx_hard_start_xmit() */
353 tx_queue->core_txq = netdev_get_tx_queue(
354 tx_queue->efx->net_dev, tx_queue->queue / EFX_TXQ_TYPES);
357 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
360 struct efx_nic *efx = tx_queue->efx;
362 EFX_BUG_ON_PARANOID(index > tx_queue->ptr_mask);
364 efx_dequeue_buffers(tx_queue, index);
366 /* See if we need to restart the netif queue. This barrier
367 * separates the update of read_count from the test of the
370 if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
371 likely(efx->port_enabled)) {
372 fill_level = tx_queue->insert_count - tx_queue->read_count;
373 if (fill_level < EFX_TXQ_THRESHOLD(efx)) {
374 EFX_BUG_ON_PARANOID(!efx_dev_registered(efx));
375 netif_tx_wake_queue(tx_queue->core_txq);
379 /* Check whether the hardware queue is now empty */
380 if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
381 tx_queue->old_write_count = ACCESS_ONCE(tx_queue->write_count);
382 if (tx_queue->read_count == tx_queue->old_write_count) {
384 tx_queue->empty_read_count =
385 tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
390 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
392 struct efx_nic *efx = tx_queue->efx;
393 unsigned int entries;
396 /* Create the smallest power-of-two aligned ring */
397 entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
398 EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
399 tx_queue->ptr_mask = entries - 1;
401 netif_dbg(efx, probe, efx->net_dev,
402 "creating TX queue %d size %#x mask %#x\n",
403 tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
405 /* Allocate software ring */
406 tx_queue->buffer = kzalloc(entries * sizeof(*tx_queue->buffer),
408 if (!tx_queue->buffer)
410 for (i = 0; i <= tx_queue->ptr_mask; ++i)
411 tx_queue->buffer[i].continuation = true;
413 /* Allocate hardware ring */
414 rc = efx_nic_probe_tx(tx_queue);
421 kfree(tx_queue->buffer);
422 tx_queue->buffer = NULL;
426 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
428 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
429 "initialising TX queue %d\n", tx_queue->queue);
431 tx_queue->insert_count = 0;
432 tx_queue->write_count = 0;
433 tx_queue->old_write_count = 0;
434 tx_queue->read_count = 0;
435 tx_queue->old_read_count = 0;
436 tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
438 /* Set up TX descriptor ring */
439 efx_nic_init_tx(tx_queue);
442 void efx_release_tx_buffers(struct efx_tx_queue *tx_queue)
444 struct efx_tx_buffer *buffer;
446 if (!tx_queue->buffer)
449 /* Free any buffers left in the ring */
450 while (tx_queue->read_count != tx_queue->write_count) {
451 buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
452 efx_dequeue_buffer(tx_queue, buffer);
453 buffer->continuation = true;
456 ++tx_queue->read_count;
460 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
462 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
463 "shutting down TX queue %d\n", tx_queue->queue);
465 /* Flush TX queue, remove descriptor ring */
466 efx_nic_fini_tx(tx_queue);
468 efx_release_tx_buffers(tx_queue);
470 /* Free up TSO header cache */
471 efx_fini_tso(tx_queue);
474 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
476 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
477 "destroying TX queue %d\n", tx_queue->queue);
478 efx_nic_remove_tx(tx_queue);
480 kfree(tx_queue->buffer);
481 tx_queue->buffer = NULL;
485 /* Efx TCP segmentation acceleration.
487 * Why? Because by doing it here in the driver we can go significantly
488 * faster than the GSO.
490 * Requires TX checksum offload support.
493 /* Number of bytes inserted at the start of a TSO header buffer,
494 * similar to NET_IP_ALIGN.
496 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
497 #define TSOH_OFFSET 0
499 #define TSOH_OFFSET NET_IP_ALIGN
502 #define TSOH_BUFFER(tsoh) ((u8 *)(tsoh + 1) + TSOH_OFFSET)
504 /* Total size of struct efx_tso_header, buffer and padding */
505 #define TSOH_SIZE(hdr_len) \
506 (sizeof(struct efx_tso_header) + TSOH_OFFSET + hdr_len)
508 /* Size of blocks on free list. Larger blocks must be allocated from
511 #define TSOH_STD_SIZE 128
513 #define PTR_DIFF(p1, p2) ((u8 *)(p1) - (u8 *)(p2))
514 #define ETH_HDR_LEN(skb) (skb_network_header(skb) - (skb)->data)
515 #define SKB_TCP_OFF(skb) PTR_DIFF(tcp_hdr(skb), (skb)->data)
516 #define SKB_IPV4_OFF(skb) PTR_DIFF(ip_hdr(skb), (skb)->data)
517 #define SKB_IPV6_OFF(skb) PTR_DIFF(ipv6_hdr(skb), (skb)->data)
520 * struct tso_state - TSO state for an SKB
521 * @out_len: Remaining length in current segment
522 * @seqnum: Current sequence number
523 * @ipv4_id: Current IPv4 ID, host endian
524 * @packet_space: Remaining space in current packet
525 * @dma_addr: DMA address of current position
526 * @in_len: Remaining length in current SKB fragment
527 * @unmap_len: Length of SKB fragment
528 * @unmap_addr: DMA address of SKB fragment
529 * @unmap_single: DMA single vs page mapping flag
530 * @protocol: Network protocol (after any VLAN header)
531 * @header_len: Number of bytes of header
532 * @full_packet_size: Number of bytes to put in each outgoing segment
534 * The state used during segmentation. It is put into this data structure
535 * just to make it easy to pass into inline functions.
538 /* Output position */
542 unsigned packet_space;
548 dma_addr_t unmap_addr;
553 int full_packet_size;
558 * Verify that our various assumptions about sk_buffs and the conditions
559 * under which TSO will be attempted hold true. Return the protocol number.
561 static __be16 efx_tso_check_protocol(struct sk_buff *skb)
563 __be16 protocol = skb->protocol;
565 EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
567 if (protocol == htons(ETH_P_8021Q)) {
568 /* Find the encapsulated protocol; reset network header
569 * and transport header based on that. */
570 struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
571 protocol = veh->h_vlan_encapsulated_proto;
572 skb_set_network_header(skb, sizeof(*veh));
573 if (protocol == htons(ETH_P_IP))
574 skb_set_transport_header(skb, sizeof(*veh) +
575 4 * ip_hdr(skb)->ihl);
576 else if (protocol == htons(ETH_P_IPV6))
577 skb_set_transport_header(skb, sizeof(*veh) +
578 sizeof(struct ipv6hdr));
581 if (protocol == htons(ETH_P_IP)) {
582 EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
584 EFX_BUG_ON_PARANOID(protocol != htons(ETH_P_IPV6));
585 EFX_BUG_ON_PARANOID(ipv6_hdr(skb)->nexthdr != NEXTHDR_TCP);
587 EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
588 + (tcp_hdr(skb)->doff << 2u)) >
596 * Allocate a page worth of efx_tso_header structures, and string them
597 * into the tx_queue->tso_headers_free linked list. Return 0 or -ENOMEM.
599 static int efx_tsoh_block_alloc(struct efx_tx_queue *tx_queue)
602 struct pci_dev *pci_dev = tx_queue->efx->pci_dev;
603 struct efx_tso_header *tsoh;
607 base_kva = pci_alloc_consistent(pci_dev, PAGE_SIZE, &dma_addr);
608 if (base_kva == NULL) {
609 netif_err(tx_queue->efx, tx_err, tx_queue->efx->net_dev,
610 "Unable to allocate page for TSO headers\n");
614 /* pci_alloc_consistent() allocates pages. */
615 EFX_BUG_ON_PARANOID(dma_addr & (PAGE_SIZE - 1u));
617 for (kva = base_kva; kva < base_kva + PAGE_SIZE; kva += TSOH_STD_SIZE) {
618 tsoh = (struct efx_tso_header *)kva;
619 tsoh->dma_addr = dma_addr + (TSOH_BUFFER(tsoh) - base_kva);
620 tsoh->next = tx_queue->tso_headers_free;
621 tx_queue->tso_headers_free = tsoh;
628 /* Free up a TSO header, and all others in the same page. */
629 static void efx_tsoh_block_free(struct efx_tx_queue *tx_queue,
630 struct efx_tso_header *tsoh,
631 struct pci_dev *pci_dev)
633 struct efx_tso_header **p;
634 unsigned long base_kva;
637 base_kva = (unsigned long)tsoh & PAGE_MASK;
638 base_dma = tsoh->dma_addr & PAGE_MASK;
640 p = &tx_queue->tso_headers_free;
642 if (((unsigned long)*p & PAGE_MASK) == base_kva)
648 pci_free_consistent(pci_dev, PAGE_SIZE, (void *)base_kva, base_dma);
651 static struct efx_tso_header *
652 efx_tsoh_heap_alloc(struct efx_tx_queue *tx_queue, size_t header_len)
654 struct efx_tso_header *tsoh;
656 tsoh = kmalloc(TSOH_SIZE(header_len), GFP_ATOMIC | GFP_DMA);
660 tsoh->dma_addr = pci_map_single(tx_queue->efx->pci_dev,
661 TSOH_BUFFER(tsoh), header_len,
663 if (unlikely(pci_dma_mapping_error(tx_queue->efx->pci_dev,
669 tsoh->unmap_len = header_len;
674 efx_tsoh_heap_free(struct efx_tx_queue *tx_queue, struct efx_tso_header *tsoh)
676 pci_unmap_single(tx_queue->efx->pci_dev,
677 tsoh->dma_addr, tsoh->unmap_len,
683 * efx_tx_queue_insert - push descriptors onto the TX queue
684 * @tx_queue: Efx TX queue
685 * @dma_addr: DMA address of fragment
686 * @len: Length of fragment
687 * @final_buffer: The final buffer inserted into the queue
689 * Push descriptors onto the TX queue. Return 0 on success or 1 if
692 static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
693 dma_addr_t dma_addr, unsigned len,
694 struct efx_tx_buffer **final_buffer)
696 struct efx_tx_buffer *buffer;
697 struct efx_nic *efx = tx_queue->efx;
698 unsigned dma_len, fill_level, insert_ptr;
701 EFX_BUG_ON_PARANOID(len <= 0);
703 fill_level = tx_queue->insert_count - tx_queue->old_read_count;
704 /* -1 as there is no way to represent all descriptors used */
705 q_space = efx->txq_entries - 1 - fill_level;
708 if (unlikely(q_space-- <= 0)) {
709 /* It might be that completions have happened
710 * since the xmit path last checked. Update
711 * the xmit path's copy of read_count.
713 netif_tx_stop_queue(tx_queue->core_txq);
714 /* This memory barrier protects the change of
715 * queue state from the access of read_count. */
717 tx_queue->old_read_count =
718 ACCESS_ONCE(tx_queue->read_count);
719 fill_level = (tx_queue->insert_count
720 - tx_queue->old_read_count);
721 q_space = efx->txq_entries - 1 - fill_level;
722 if (unlikely(q_space-- <= 0)) {
723 *final_buffer = NULL;
727 netif_tx_start_queue(tx_queue->core_txq);
730 insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
731 buffer = &tx_queue->buffer[insert_ptr];
732 ++tx_queue->insert_count;
734 EFX_BUG_ON_PARANOID(tx_queue->insert_count -
735 tx_queue->read_count >=
738 efx_tsoh_free(tx_queue, buffer);
739 EFX_BUG_ON_PARANOID(buffer->len);
740 EFX_BUG_ON_PARANOID(buffer->unmap_len);
741 EFX_BUG_ON_PARANOID(buffer->skb);
742 EFX_BUG_ON_PARANOID(!buffer->continuation);
743 EFX_BUG_ON_PARANOID(buffer->tsoh);
745 buffer->dma_addr = dma_addr;
747 dma_len = efx_max_tx_len(efx, dma_addr);
749 /* If there is enough space to send then do so */
753 buffer->len = dma_len; /* Don't set the other members */
758 EFX_BUG_ON_PARANOID(!len);
760 *final_buffer = buffer;
766 * Put a TSO header into the TX queue.
768 * This is special-cased because we know that it is small enough to fit in
769 * a single fragment, and we know it doesn't cross a page boundary. It
770 * also allows us to not worry about end-of-packet etc.
772 static void efx_tso_put_header(struct efx_tx_queue *tx_queue,
773 struct efx_tso_header *tsoh, unsigned len)
775 struct efx_tx_buffer *buffer;
777 buffer = &tx_queue->buffer[tx_queue->insert_count & tx_queue->ptr_mask];
778 efx_tsoh_free(tx_queue, buffer);
779 EFX_BUG_ON_PARANOID(buffer->len);
780 EFX_BUG_ON_PARANOID(buffer->unmap_len);
781 EFX_BUG_ON_PARANOID(buffer->skb);
782 EFX_BUG_ON_PARANOID(!buffer->continuation);
783 EFX_BUG_ON_PARANOID(buffer->tsoh);
785 buffer->dma_addr = tsoh->dma_addr;
788 ++tx_queue->insert_count;
792 /* Remove descriptors put into a tx_queue. */
793 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
795 struct efx_tx_buffer *buffer;
796 dma_addr_t unmap_addr;
798 /* Work backwards until we hit the original insert pointer value */
799 while (tx_queue->insert_count != tx_queue->write_count) {
800 --tx_queue->insert_count;
801 buffer = &tx_queue->buffer[tx_queue->insert_count &
803 efx_tsoh_free(tx_queue, buffer);
804 EFX_BUG_ON_PARANOID(buffer->skb);
805 if (buffer->unmap_len) {
806 unmap_addr = (buffer->dma_addr + buffer->len -
808 if (buffer->unmap_single)
809 pci_unmap_single(tx_queue->efx->pci_dev,
810 unmap_addr, buffer->unmap_len,
813 pci_unmap_page(tx_queue->efx->pci_dev,
814 unmap_addr, buffer->unmap_len,
816 buffer->unmap_len = 0;
819 buffer->continuation = true;
824 /* Parse the SKB header and initialise state. */
825 static void tso_start(struct tso_state *st, const struct sk_buff *skb)
827 /* All ethernet/IP/TCP headers combined size is TCP header size
828 * plus offset of TCP header relative to start of packet.
830 st->header_len = ((tcp_hdr(skb)->doff << 2u)
831 + PTR_DIFF(tcp_hdr(skb), skb->data));
832 st->full_packet_size = st->header_len + skb_shinfo(skb)->gso_size;
834 if (st->protocol == htons(ETH_P_IP))
835 st->ipv4_id = ntohs(ip_hdr(skb)->id);
838 st->seqnum = ntohl(tcp_hdr(skb)->seq);
840 EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
841 EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
842 EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
844 st->packet_space = st->full_packet_size;
845 st->out_len = skb->len - st->header_len;
847 st->unmap_single = false;
850 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
853 st->unmap_addr = pci_map_page(efx->pci_dev, frag->page,
854 frag->page_offset, frag->size,
856 if (likely(!pci_dma_mapping_error(efx->pci_dev, st->unmap_addr))) {
857 st->unmap_single = false;
858 st->unmap_len = frag->size;
859 st->in_len = frag->size;
860 st->dma_addr = st->unmap_addr;
866 static int tso_get_head_fragment(struct tso_state *st, struct efx_nic *efx,
867 const struct sk_buff *skb)
869 int hl = st->header_len;
870 int len = skb_headlen(skb) - hl;
872 st->unmap_addr = pci_map_single(efx->pci_dev, skb->data + hl,
873 len, PCI_DMA_TODEVICE);
874 if (likely(!pci_dma_mapping_error(efx->pci_dev, st->unmap_addr))) {
875 st->unmap_single = true;
878 st->dma_addr = st->unmap_addr;
886 * tso_fill_packet_with_fragment - form descriptors for the current fragment
887 * @tx_queue: Efx TX queue
888 * @skb: Socket buffer
891 * Form descriptors for the current fragment, until we reach the end
892 * of fragment or end-of-packet. Return 0 on success, 1 if not enough
893 * space in @tx_queue.
895 static int tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
896 const struct sk_buff *skb,
897 struct tso_state *st)
899 struct efx_tx_buffer *buffer;
900 int n, end_of_packet, rc;
904 if (st->packet_space == 0)
907 EFX_BUG_ON_PARANOID(st->in_len <= 0);
908 EFX_BUG_ON_PARANOID(st->packet_space <= 0);
910 n = min(st->in_len, st->packet_space);
912 st->packet_space -= n;
916 rc = efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer);
917 if (likely(rc == 0)) {
918 if (st->out_len == 0)
919 /* Transfer ownership of the skb */
922 end_of_packet = st->out_len == 0 || st->packet_space == 0;
923 buffer->continuation = !end_of_packet;
925 if (st->in_len == 0) {
926 /* Transfer ownership of the pci mapping */
927 buffer->unmap_len = st->unmap_len;
928 buffer->unmap_single = st->unmap_single;
939 * tso_start_new_packet - generate a new header and prepare for the new packet
940 * @tx_queue: Efx TX queue
941 * @skb: Socket buffer
944 * Generate a new header and prepare for the new packet. Return 0 on
945 * success, or -1 if failed to alloc header.
947 static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
948 const struct sk_buff *skb,
949 struct tso_state *st)
951 struct efx_tso_header *tsoh;
952 struct tcphdr *tsoh_th;
956 /* Allocate a DMA-mapped header buffer. */
957 if (likely(TSOH_SIZE(st->header_len) <= TSOH_STD_SIZE)) {
958 if (tx_queue->tso_headers_free == NULL) {
959 if (efx_tsoh_block_alloc(tx_queue))
962 EFX_BUG_ON_PARANOID(!tx_queue->tso_headers_free);
963 tsoh = tx_queue->tso_headers_free;
964 tx_queue->tso_headers_free = tsoh->next;
967 tx_queue->tso_long_headers++;
968 tsoh = efx_tsoh_heap_alloc(tx_queue, st->header_len);
973 header = TSOH_BUFFER(tsoh);
974 tsoh_th = (struct tcphdr *)(header + SKB_TCP_OFF(skb));
976 /* Copy and update the headers. */
977 memcpy(header, skb->data, st->header_len);
979 tsoh_th->seq = htonl(st->seqnum);
980 st->seqnum += skb_shinfo(skb)->gso_size;
981 if (st->out_len > skb_shinfo(skb)->gso_size) {
982 /* This packet will not finish the TSO burst. */
983 ip_length = st->full_packet_size - ETH_HDR_LEN(skb);
987 /* This packet will be the last in the TSO burst. */
988 ip_length = st->header_len - ETH_HDR_LEN(skb) + st->out_len;
989 tsoh_th->fin = tcp_hdr(skb)->fin;
990 tsoh_th->psh = tcp_hdr(skb)->psh;
993 if (st->protocol == htons(ETH_P_IP)) {
994 struct iphdr *tsoh_iph =
995 (struct iphdr *)(header + SKB_IPV4_OFF(skb));
997 tsoh_iph->tot_len = htons(ip_length);
999 /* Linux leaves suitable gaps in the IP ID space for us to fill. */
1000 tsoh_iph->id = htons(st->ipv4_id);
1003 struct ipv6hdr *tsoh_iph =
1004 (struct ipv6hdr *)(header + SKB_IPV6_OFF(skb));
1006 tsoh_iph->payload_len = htons(ip_length - sizeof(*tsoh_iph));
1009 st->packet_space = skb_shinfo(skb)->gso_size;
1010 ++tx_queue->tso_packets;
1012 /* Form a descriptor for this header. */
1013 efx_tso_put_header(tx_queue, tsoh, st->header_len);
1020 * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1021 * @tx_queue: Efx TX queue
1022 * @skb: Socket buffer
1024 * Context: You must hold netif_tx_lock() to call this function.
1026 * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1027 * @skb was not enqueued. In all cases @skb is consumed. Return
1028 * %NETDEV_TX_OK or %NETDEV_TX_BUSY.
1030 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1031 struct sk_buff *skb)
1033 struct efx_nic *efx = tx_queue->efx;
1034 int frag_i, rc, rc2 = NETDEV_TX_OK;
1035 struct tso_state state;
1037 /* Find the packet protocol and sanity-check it */
1038 state.protocol = efx_tso_check_protocol(skb);
1040 EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
1042 tso_start(&state, skb);
1044 /* Assume that skb header area contains exactly the headers, and
1045 * all payload is in the frag list.
1047 if (skb_headlen(skb) == state.header_len) {
1048 /* Grab the first payload fragment. */
1049 EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1051 rc = tso_get_fragment(&state, efx,
1052 skb_shinfo(skb)->frags + frag_i);
1056 rc = tso_get_head_fragment(&state, efx, skb);
1062 if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1066 rc = tso_fill_packet_with_fragment(tx_queue, skb, &state);
1068 rc2 = NETDEV_TX_BUSY;
1072 /* Move onto the next fragment? */
1073 if (state.in_len == 0) {
1074 if (++frag_i >= skb_shinfo(skb)->nr_frags)
1075 /* End of payload reached. */
1077 rc = tso_get_fragment(&state, efx,
1078 skb_shinfo(skb)->frags + frag_i);
1083 /* Start at new packet? */
1084 if (state.packet_space == 0 &&
1085 tso_start_new_packet(tx_queue, skb, &state) < 0)
1089 /* Pass off to hardware */
1090 efx_nic_push_buffers(tx_queue);
1092 tx_queue->tso_bursts++;
1093 return NETDEV_TX_OK;
1096 netif_err(efx, tx_err, efx->net_dev,
1097 "Out of memory for TSO headers, or PCI mapping error\n");
1098 dev_kfree_skb_any(skb);
1101 /* Free the DMA mapping we were in the process of writing out */
1102 if (state.unmap_len) {
1103 if (state.unmap_single)
1104 pci_unmap_single(efx->pci_dev, state.unmap_addr,
1105 state.unmap_len, PCI_DMA_TODEVICE);
1107 pci_unmap_page(efx->pci_dev, state.unmap_addr,
1108 state.unmap_len, PCI_DMA_TODEVICE);
1111 efx_enqueue_unwind(tx_queue);
1117 * Free up all TSO datastructures associated with tx_queue. This
1118 * routine should be called only once the tx_queue is both empty and
1119 * will no longer be used.
1121 static void efx_fini_tso(struct efx_tx_queue *tx_queue)
1125 if (tx_queue->buffer) {
1126 for (i = 0; i <= tx_queue->ptr_mask; ++i)
1127 efx_tsoh_free(tx_queue, &tx_queue->buffer[i]);
1130 while (tx_queue->tso_headers_free != NULL)
1131 efx_tsoh_block_free(tx_queue, tx_queue->tso_headers_free,
1132 tx_queue->efx->pci_dev);