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_TXQ_MASK / 2u)
33 /* We need to be able to nest calls to netif_tx_stop_queue(), partly
34 * because of the 2 hardware queues associated with each core queue,
35 * but also so that we can inhibit TX for reasons other than a full
37 void efx_stop_queue(struct efx_channel
*channel
)
39 struct efx_nic
*efx
= channel
->efx
;
41 if (!channel
->tx_queue
)
44 spin_lock_bh(&channel
->tx_stop_lock
);
45 netif_vdbg(efx
, tx_queued
, efx
->net_dev
, "stop TX queue\n");
47 atomic_inc(&channel
->tx_stop_count
);
51 channel
->tx_queue
->queue
/ EFX_TXQ_TYPES
));
53 spin_unlock_bh(&channel
->tx_stop_lock
);
56 /* Decrement core TX queue stop count and wake it if the count is 0 */
57 void efx_wake_queue(struct efx_channel
*channel
)
59 struct efx_nic
*efx
= channel
->efx
;
61 if (!channel
->tx_queue
)
65 if (atomic_dec_and_lock(&channel
->tx_stop_count
,
66 &channel
->tx_stop_lock
)) {
67 netif_vdbg(efx
, tx_queued
, efx
->net_dev
, "waking TX queue\n");
71 channel
->tx_queue
->queue
/ EFX_TXQ_TYPES
));
72 spin_unlock(&channel
->tx_stop_lock
);
77 static void efx_dequeue_buffer(struct efx_tx_queue
*tx_queue
,
78 struct efx_tx_buffer
*buffer
)
80 if (buffer
->unmap_len
) {
81 struct pci_dev
*pci_dev
= tx_queue
->efx
->pci_dev
;
82 dma_addr_t unmap_addr
= (buffer
->dma_addr
+ buffer
->len
-
84 if (buffer
->unmap_single
)
85 pci_unmap_single(pci_dev
, unmap_addr
, buffer
->unmap_len
,
88 pci_unmap_page(pci_dev
, unmap_addr
, buffer
->unmap_len
,
90 buffer
->unmap_len
= 0;
91 buffer
->unmap_single
= false;
95 dev_kfree_skb_any((struct sk_buff
*) buffer
->skb
);
97 netif_vdbg(tx_queue
->efx
, tx_done
, tx_queue
->efx
->net_dev
,
98 "TX queue %d transmission id %x complete\n",
99 tx_queue
->queue
, tx_queue
->read_count
);
104 * struct efx_tso_header - a DMA mapped buffer for packet headers
105 * @next: Linked list of free ones.
106 * The list is protected by the TX queue lock.
107 * @dma_unmap_len: Length to unmap for an oversize buffer, or 0.
108 * @dma_addr: The DMA address of the header below.
110 * This controls the memory used for a TSO header. Use TSOH_DATA()
111 * to find the packet header data. Use TSOH_SIZE() to calculate the
112 * total size required for a given packet header length. TSO headers
113 * in the free list are exactly %TSOH_STD_SIZE bytes in size.
115 struct efx_tso_header
{
117 struct efx_tso_header
*next
;
123 static int efx_enqueue_skb_tso(struct efx_tx_queue
*tx_queue
,
124 struct sk_buff
*skb
);
125 static void efx_fini_tso(struct efx_tx_queue
*tx_queue
);
126 static void efx_tsoh_heap_free(struct efx_tx_queue
*tx_queue
,
127 struct efx_tso_header
*tsoh
);
129 static void efx_tsoh_free(struct efx_tx_queue
*tx_queue
,
130 struct efx_tx_buffer
*buffer
)
133 if (likely(!buffer
->tsoh
->unmap_len
)) {
134 buffer
->tsoh
->next
= tx_queue
->tso_headers_free
;
135 tx_queue
->tso_headers_free
= buffer
->tsoh
;
137 efx_tsoh_heap_free(tx_queue
, buffer
->tsoh
);
144 static inline unsigned
145 efx_max_tx_len(struct efx_nic
*efx
, dma_addr_t dma_addr
)
147 /* Depending on the NIC revision, we can use descriptor
148 * lengths up to 8K or 8K-1. However, since PCI Express
149 * devices must split read requests at 4K boundaries, there is
150 * little benefit from using descriptors that cross those
151 * boundaries and we keep things simple by not doing so.
153 unsigned len
= (~dma_addr
& 0xfff) + 1;
155 if (EFX_WORKAROUND_5391(efx
) && (dma_addr
& 0xf))
156 len
= min_t(unsigned, len
, 512 - (dma_addr
& 0xf));
162 * Add a socket buffer to a TX queue
164 * This maps all fragments of a socket buffer for DMA and adds them to
165 * the TX queue. The queue's insert pointer will be incremented by
166 * the number of fragments in the socket buffer.
168 * If any DMA mapping fails, any mapped fragments will be unmapped,
169 * the queue's insert pointer will be restored to its original value.
171 * This function is split out from efx_hard_start_xmit to allow the
172 * loopback test to direct packets via specific TX queues.
174 * Returns NETDEV_TX_OK or NETDEV_TX_BUSY
175 * You must hold netif_tx_lock() to call this function.
177 netdev_tx_t
efx_enqueue_skb(struct efx_tx_queue
*tx_queue
, struct sk_buff
*skb
)
179 struct efx_nic
*efx
= tx_queue
->efx
;
180 struct pci_dev
*pci_dev
= efx
->pci_dev
;
181 struct efx_tx_buffer
*buffer
;
182 skb_frag_t
*fragment
;
185 unsigned int len
, unmap_len
= 0, fill_level
, insert_ptr
;
186 dma_addr_t dma_addr
, unmap_addr
= 0;
187 unsigned int dma_len
;
190 netdev_tx_t rc
= NETDEV_TX_OK
;
192 EFX_BUG_ON_PARANOID(tx_queue
->write_count
!= tx_queue
->insert_count
);
194 if (skb_shinfo(skb
)->gso_size
)
195 return efx_enqueue_skb_tso(tx_queue
, skb
);
197 /* Get size of the initial fragment */
198 len
= skb_headlen(skb
);
200 /* Pad if necessary */
201 if (EFX_WORKAROUND_15592(efx
) && skb
->len
<= 32) {
202 EFX_BUG_ON_PARANOID(skb
->data_len
);
204 if (skb_pad(skb
, len
- skb
->len
))
208 fill_level
= tx_queue
->insert_count
- tx_queue
->old_read_count
;
209 q_space
= EFX_TXQ_MASK
- 1 - fill_level
;
211 /* Map for DMA. Use pci_map_single rather than pci_map_page
212 * since this is more efficient on machines with sparse
216 dma_addr
= pci_map_single(pci_dev
, skb
->data
, len
, PCI_DMA_TODEVICE
);
218 /* Process all fragments */
220 if (unlikely(pci_dma_mapping_error(pci_dev
, dma_addr
)))
223 /* Store fields for marking in the per-fragment final
226 unmap_addr
= dma_addr
;
228 /* Add to TX queue, splitting across DMA boundaries */
230 if (unlikely(q_space
-- <= 0)) {
231 /* It might be that completions have
232 * happened since the xmit path last
233 * checked. Update the xmit path's
234 * copy of read_count.
237 /* This memory barrier protects the
238 * change of stopped from the access
241 tx_queue
->old_read_count
=
242 *(volatile unsigned *)
243 &tx_queue
->read_count
;
244 fill_level
= (tx_queue
->insert_count
245 - tx_queue
->old_read_count
);
246 q_space
= EFX_TXQ_MASK
- 1 - fill_level
;
247 if (unlikely(q_space
-- <= 0))
253 insert_ptr
= tx_queue
->insert_count
& EFX_TXQ_MASK
;
254 buffer
= &tx_queue
->buffer
[insert_ptr
];
255 efx_tsoh_free(tx_queue
, buffer
);
256 EFX_BUG_ON_PARANOID(buffer
->tsoh
);
257 EFX_BUG_ON_PARANOID(buffer
->skb
);
258 EFX_BUG_ON_PARANOID(buffer
->len
);
259 EFX_BUG_ON_PARANOID(!buffer
->continuation
);
260 EFX_BUG_ON_PARANOID(buffer
->unmap_len
);
262 dma_len
= efx_max_tx_len(efx
, dma_addr
);
263 if (likely(dma_len
>= len
))
266 /* Fill out per descriptor fields */
267 buffer
->len
= dma_len
;
268 buffer
->dma_addr
= dma_addr
;
271 ++tx_queue
->insert_count
;
274 /* Transfer ownership of the unmapping to the final buffer */
275 buffer
->unmap_single
= unmap_single
;
276 buffer
->unmap_len
= unmap_len
;
279 /* Get address and size of next fragment */
280 if (i
>= skb_shinfo(skb
)->nr_frags
)
282 fragment
= &skb_shinfo(skb
)->frags
[i
];
283 len
= fragment
->size
;
284 page
= fragment
->page
;
285 page_offset
= fragment
->page_offset
;
288 unmap_single
= false;
289 dma_addr
= pci_map_page(pci_dev
, page
, page_offset
, len
,
293 /* Transfer ownership of the skb to the final buffer */
295 buffer
->continuation
= false;
297 /* Pass off to hardware */
298 efx_nic_push_buffers(tx_queue
);
303 netif_err(efx
, tx_err
, efx
->net_dev
,
304 " TX queue %d could not map skb with %d bytes %d "
305 "fragments for DMA\n", tx_queue
->queue
, skb
->len
,
306 skb_shinfo(skb
)->nr_frags
+ 1);
308 /* Mark the packet as transmitted, and free the SKB ourselves */
309 dev_kfree_skb_any(skb
);
315 if (tx_queue
->stopped
== 1)
316 efx_stop_queue(tx_queue
->channel
);
319 /* Work backwards until we hit the original insert pointer value */
320 while (tx_queue
->insert_count
!= tx_queue
->write_count
) {
321 --tx_queue
->insert_count
;
322 insert_ptr
= tx_queue
->insert_count
& EFX_TXQ_MASK
;
323 buffer
= &tx_queue
->buffer
[insert_ptr
];
324 efx_dequeue_buffer(tx_queue
, buffer
);
328 /* Free the fragment we were mid-way through pushing */
331 pci_unmap_single(pci_dev
, unmap_addr
, unmap_len
,
334 pci_unmap_page(pci_dev
, unmap_addr
, unmap_len
,
341 /* Remove packets from the TX queue
343 * This removes packets from the TX queue, up to and including the
346 static void efx_dequeue_buffers(struct efx_tx_queue
*tx_queue
,
349 struct efx_nic
*efx
= tx_queue
->efx
;
350 unsigned int stop_index
, read_ptr
;
352 stop_index
= (index
+ 1) & EFX_TXQ_MASK
;
353 read_ptr
= tx_queue
->read_count
& EFX_TXQ_MASK
;
355 while (read_ptr
!= stop_index
) {
356 struct efx_tx_buffer
*buffer
= &tx_queue
->buffer
[read_ptr
];
357 if (unlikely(buffer
->len
== 0)) {
358 netif_err(efx
, tx_err
, efx
->net_dev
,
359 "TX queue %d spurious TX completion id %x\n",
360 tx_queue
->queue
, read_ptr
);
361 efx_schedule_reset(efx
, RESET_TYPE_TX_SKIP
);
365 efx_dequeue_buffer(tx_queue
, buffer
);
366 buffer
->continuation
= true;
369 ++tx_queue
->read_count
;
370 read_ptr
= tx_queue
->read_count
& EFX_TXQ_MASK
;
374 /* Initiate a packet transmission. We use one channel per CPU
375 * (sharing when we have more CPUs than channels). On Falcon, the TX
376 * completion events will be directed back to the CPU that transmitted
377 * the packet, which should be cache-efficient.
379 * Context: non-blocking.
380 * Note that returning anything other than NETDEV_TX_OK will cause the
381 * OS to free the skb.
383 netdev_tx_t
efx_hard_start_xmit(struct sk_buff
*skb
,
384 struct net_device
*net_dev
)
386 struct efx_nic
*efx
= netdev_priv(net_dev
);
387 struct efx_tx_queue
*tx_queue
;
389 if (unlikely(efx
->port_inhibited
))
390 return NETDEV_TX_BUSY
;
392 tx_queue
= &efx
->tx_queue
[EFX_TXQ_TYPES
* skb_get_queue_mapping(skb
)];
393 if (likely(skb
->ip_summed
== CHECKSUM_PARTIAL
))
394 tx_queue
+= EFX_TXQ_TYPE_OFFLOAD
;
396 return efx_enqueue_skb(tx_queue
, skb
);
399 void efx_xmit_done(struct efx_tx_queue
*tx_queue
, unsigned int index
)
402 struct efx_nic
*efx
= tx_queue
->efx
;
404 EFX_BUG_ON_PARANOID(index
> EFX_TXQ_MASK
);
406 efx_dequeue_buffers(tx_queue
, index
);
408 /* See if we need to restart the netif queue. This barrier
409 * separates the update of read_count from the test of
412 if (unlikely(tx_queue
->stopped
) && likely(efx
->port_enabled
)) {
413 fill_level
= tx_queue
->insert_count
- tx_queue
->read_count
;
414 if (fill_level
< EFX_TXQ_THRESHOLD
) {
415 EFX_BUG_ON_PARANOID(!efx_dev_registered(efx
));
417 /* Do this under netif_tx_lock(), to avoid racing
418 * with efx_xmit(). */
419 netif_tx_lock(efx
->net_dev
);
420 if (tx_queue
->stopped
) {
421 tx_queue
->stopped
= 0;
422 efx_wake_queue(tx_queue
->channel
);
424 netif_tx_unlock(efx
->net_dev
);
429 int efx_probe_tx_queue(struct efx_tx_queue
*tx_queue
)
431 struct efx_nic
*efx
= tx_queue
->efx
;
432 unsigned int txq_size
;
435 netif_dbg(efx
, probe
, efx
->net_dev
, "creating TX queue %d\n",
438 /* Allocate software ring */
439 txq_size
= EFX_TXQ_SIZE
* sizeof(*tx_queue
->buffer
);
440 tx_queue
->buffer
= kzalloc(txq_size
, GFP_KERNEL
);
441 if (!tx_queue
->buffer
)
443 for (i
= 0; i
<= EFX_TXQ_MASK
; ++i
)
444 tx_queue
->buffer
[i
].continuation
= true;
446 /* Allocate hardware ring */
447 rc
= efx_nic_probe_tx(tx_queue
);
454 kfree(tx_queue
->buffer
);
455 tx_queue
->buffer
= NULL
;
459 void efx_init_tx_queue(struct efx_tx_queue
*tx_queue
)
461 netif_dbg(tx_queue
->efx
, drv
, tx_queue
->efx
->net_dev
,
462 "initialising TX queue %d\n", tx_queue
->queue
);
464 tx_queue
->insert_count
= 0;
465 tx_queue
->write_count
= 0;
466 tx_queue
->read_count
= 0;
467 tx_queue
->old_read_count
= 0;
468 BUG_ON(tx_queue
->stopped
);
470 /* Set up TX descriptor ring */
471 efx_nic_init_tx(tx_queue
);
474 void efx_release_tx_buffers(struct efx_tx_queue
*tx_queue
)
476 struct efx_tx_buffer
*buffer
;
478 if (!tx_queue
->buffer
)
481 /* Free any buffers left in the ring */
482 while (tx_queue
->read_count
!= tx_queue
->write_count
) {
483 buffer
= &tx_queue
->buffer
[tx_queue
->read_count
& EFX_TXQ_MASK
];
484 efx_dequeue_buffer(tx_queue
, buffer
);
485 buffer
->continuation
= true;
488 ++tx_queue
->read_count
;
492 void efx_fini_tx_queue(struct efx_tx_queue
*tx_queue
)
494 netif_dbg(tx_queue
->efx
, drv
, tx_queue
->efx
->net_dev
,
495 "shutting down TX queue %d\n", tx_queue
->queue
);
497 /* Flush TX queue, remove descriptor ring */
498 efx_nic_fini_tx(tx_queue
);
500 efx_release_tx_buffers(tx_queue
);
502 /* Free up TSO header cache */
503 efx_fini_tso(tx_queue
);
505 /* Release queue's stop on port, if any */
506 if (tx_queue
->stopped
) {
507 tx_queue
->stopped
= 0;
508 efx_wake_queue(tx_queue
->channel
);
512 void efx_remove_tx_queue(struct efx_tx_queue
*tx_queue
)
514 netif_dbg(tx_queue
->efx
, drv
, tx_queue
->efx
->net_dev
,
515 "destroying TX queue %d\n", tx_queue
->queue
);
516 efx_nic_remove_tx(tx_queue
);
518 kfree(tx_queue
->buffer
);
519 tx_queue
->buffer
= NULL
;
523 /* Efx TCP segmentation acceleration.
525 * Why? Because by doing it here in the driver we can go significantly
526 * faster than the GSO.
528 * Requires TX checksum offload support.
531 /* Number of bytes inserted at the start of a TSO header buffer,
532 * similar to NET_IP_ALIGN.
534 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
535 #define TSOH_OFFSET 0
537 #define TSOH_OFFSET NET_IP_ALIGN
540 #define TSOH_BUFFER(tsoh) ((u8 *)(tsoh + 1) + TSOH_OFFSET)
542 /* Total size of struct efx_tso_header, buffer and padding */
543 #define TSOH_SIZE(hdr_len) \
544 (sizeof(struct efx_tso_header) + TSOH_OFFSET + hdr_len)
546 /* Size of blocks on free list. Larger blocks must be allocated from
549 #define TSOH_STD_SIZE 128
551 #define PTR_DIFF(p1, p2) ((u8 *)(p1) - (u8 *)(p2))
552 #define ETH_HDR_LEN(skb) (skb_network_header(skb) - (skb)->data)
553 #define SKB_TCP_OFF(skb) PTR_DIFF(tcp_hdr(skb), (skb)->data)
554 #define SKB_IPV4_OFF(skb) PTR_DIFF(ip_hdr(skb), (skb)->data)
555 #define SKB_IPV6_OFF(skb) PTR_DIFF(ipv6_hdr(skb), (skb)->data)
558 * struct tso_state - TSO state for an SKB
559 * @out_len: Remaining length in current segment
560 * @seqnum: Current sequence number
561 * @ipv4_id: Current IPv4 ID, host endian
562 * @packet_space: Remaining space in current packet
563 * @dma_addr: DMA address of current position
564 * @in_len: Remaining length in current SKB fragment
565 * @unmap_len: Length of SKB fragment
566 * @unmap_addr: DMA address of SKB fragment
567 * @unmap_single: DMA single vs page mapping flag
568 * @protocol: Network protocol (after any VLAN header)
569 * @header_len: Number of bytes of header
570 * @full_packet_size: Number of bytes to put in each outgoing segment
572 * The state used during segmentation. It is put into this data structure
573 * just to make it easy to pass into inline functions.
576 /* Output position */
580 unsigned packet_space
;
586 dma_addr_t unmap_addr
;
591 int full_packet_size
;
596 * Verify that our various assumptions about sk_buffs and the conditions
597 * under which TSO will be attempted hold true. Return the protocol number.
599 static __be16
efx_tso_check_protocol(struct sk_buff
*skb
)
601 __be16 protocol
= skb
->protocol
;
603 EFX_BUG_ON_PARANOID(((struct ethhdr
*)skb
->data
)->h_proto
!=
605 if (protocol
== htons(ETH_P_8021Q
)) {
606 /* Find the encapsulated protocol; reset network header
607 * and transport header based on that. */
608 struct vlan_ethhdr
*veh
= (struct vlan_ethhdr
*)skb
->data
;
609 protocol
= veh
->h_vlan_encapsulated_proto
;
610 skb_set_network_header(skb
, sizeof(*veh
));
611 if (protocol
== htons(ETH_P_IP
))
612 skb_set_transport_header(skb
, sizeof(*veh
) +
613 4 * ip_hdr(skb
)->ihl
);
614 else if (protocol
== htons(ETH_P_IPV6
))
615 skb_set_transport_header(skb
, sizeof(*veh
) +
616 sizeof(struct ipv6hdr
));
619 if (protocol
== htons(ETH_P_IP
)) {
620 EFX_BUG_ON_PARANOID(ip_hdr(skb
)->protocol
!= IPPROTO_TCP
);
622 EFX_BUG_ON_PARANOID(protocol
!= htons(ETH_P_IPV6
));
623 EFX_BUG_ON_PARANOID(ipv6_hdr(skb
)->nexthdr
!= NEXTHDR_TCP
);
625 EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb
), skb
->data
)
626 + (tcp_hdr(skb
)->doff
<< 2u)) >
634 * Allocate a page worth of efx_tso_header structures, and string them
635 * into the tx_queue->tso_headers_free linked list. Return 0 or -ENOMEM.
637 static int efx_tsoh_block_alloc(struct efx_tx_queue
*tx_queue
)
640 struct pci_dev
*pci_dev
= tx_queue
->efx
->pci_dev
;
641 struct efx_tso_header
*tsoh
;
645 base_kva
= pci_alloc_consistent(pci_dev
, PAGE_SIZE
, &dma_addr
);
646 if (base_kva
== NULL
) {
647 netif_err(tx_queue
->efx
, tx_err
, tx_queue
->efx
->net_dev
,
648 "Unable to allocate page for TSO headers\n");
652 /* pci_alloc_consistent() allocates pages. */
653 EFX_BUG_ON_PARANOID(dma_addr
& (PAGE_SIZE
- 1u));
655 for (kva
= base_kva
; kva
< base_kva
+ PAGE_SIZE
; kva
+= TSOH_STD_SIZE
) {
656 tsoh
= (struct efx_tso_header
*)kva
;
657 tsoh
->dma_addr
= dma_addr
+ (TSOH_BUFFER(tsoh
) - base_kva
);
658 tsoh
->next
= tx_queue
->tso_headers_free
;
659 tx_queue
->tso_headers_free
= tsoh
;
666 /* Free up a TSO header, and all others in the same page. */
667 static void efx_tsoh_block_free(struct efx_tx_queue
*tx_queue
,
668 struct efx_tso_header
*tsoh
,
669 struct pci_dev
*pci_dev
)
671 struct efx_tso_header
**p
;
672 unsigned long base_kva
;
675 base_kva
= (unsigned long)tsoh
& PAGE_MASK
;
676 base_dma
= tsoh
->dma_addr
& PAGE_MASK
;
678 p
= &tx_queue
->tso_headers_free
;
680 if (((unsigned long)*p
& PAGE_MASK
) == base_kva
)
686 pci_free_consistent(pci_dev
, PAGE_SIZE
, (void *)base_kva
, base_dma
);
689 static struct efx_tso_header
*
690 efx_tsoh_heap_alloc(struct efx_tx_queue
*tx_queue
, size_t header_len
)
692 struct efx_tso_header
*tsoh
;
694 tsoh
= kmalloc(TSOH_SIZE(header_len
), GFP_ATOMIC
| GFP_DMA
);
698 tsoh
->dma_addr
= pci_map_single(tx_queue
->efx
->pci_dev
,
699 TSOH_BUFFER(tsoh
), header_len
,
701 if (unlikely(pci_dma_mapping_error(tx_queue
->efx
->pci_dev
,
707 tsoh
->unmap_len
= header_len
;
712 efx_tsoh_heap_free(struct efx_tx_queue
*tx_queue
, struct efx_tso_header
*tsoh
)
714 pci_unmap_single(tx_queue
->efx
->pci_dev
,
715 tsoh
->dma_addr
, tsoh
->unmap_len
,
721 * efx_tx_queue_insert - push descriptors onto the TX queue
722 * @tx_queue: Efx TX queue
723 * @dma_addr: DMA address of fragment
724 * @len: Length of fragment
725 * @final_buffer: The final buffer inserted into the queue
727 * Push descriptors onto the TX queue. Return 0 on success or 1 if
730 static int efx_tx_queue_insert(struct efx_tx_queue
*tx_queue
,
731 dma_addr_t dma_addr
, unsigned len
,
732 struct efx_tx_buffer
**final_buffer
)
734 struct efx_tx_buffer
*buffer
;
735 struct efx_nic
*efx
= tx_queue
->efx
;
736 unsigned dma_len
, fill_level
, insert_ptr
;
739 EFX_BUG_ON_PARANOID(len
<= 0);
741 fill_level
= tx_queue
->insert_count
- tx_queue
->old_read_count
;
742 /* -1 as there is no way to represent all descriptors used */
743 q_space
= EFX_TXQ_MASK
- 1 - fill_level
;
746 if (unlikely(q_space
-- <= 0)) {
747 /* It might be that completions have happened
748 * since the xmit path last checked. Update
749 * the xmit path's copy of read_count.
752 /* This memory barrier protects the change of
753 * stopped from the access of read_count. */
755 tx_queue
->old_read_count
=
756 *(volatile unsigned *)&tx_queue
->read_count
;
757 fill_level
= (tx_queue
->insert_count
758 - tx_queue
->old_read_count
);
759 q_space
= EFX_TXQ_MASK
- 1 - fill_level
;
760 if (unlikely(q_space
-- <= 0)) {
761 *final_buffer
= NULL
;
768 insert_ptr
= tx_queue
->insert_count
& EFX_TXQ_MASK
;
769 buffer
= &tx_queue
->buffer
[insert_ptr
];
770 ++tx_queue
->insert_count
;
772 EFX_BUG_ON_PARANOID(tx_queue
->insert_count
-
773 tx_queue
->read_count
>
776 efx_tsoh_free(tx_queue
, buffer
);
777 EFX_BUG_ON_PARANOID(buffer
->len
);
778 EFX_BUG_ON_PARANOID(buffer
->unmap_len
);
779 EFX_BUG_ON_PARANOID(buffer
->skb
);
780 EFX_BUG_ON_PARANOID(!buffer
->continuation
);
781 EFX_BUG_ON_PARANOID(buffer
->tsoh
);
783 buffer
->dma_addr
= dma_addr
;
785 dma_len
= efx_max_tx_len(efx
, dma_addr
);
787 /* If there is enough space to send then do so */
791 buffer
->len
= dma_len
; /* Don't set the other members */
796 EFX_BUG_ON_PARANOID(!len
);
798 *final_buffer
= buffer
;
804 * Put a TSO header into the TX queue.
806 * This is special-cased because we know that it is small enough to fit in
807 * a single fragment, and we know it doesn't cross a page boundary. It
808 * also allows us to not worry about end-of-packet etc.
810 static void efx_tso_put_header(struct efx_tx_queue
*tx_queue
,
811 struct efx_tso_header
*tsoh
, unsigned len
)
813 struct efx_tx_buffer
*buffer
;
815 buffer
= &tx_queue
->buffer
[tx_queue
->insert_count
& EFX_TXQ_MASK
];
816 efx_tsoh_free(tx_queue
, buffer
);
817 EFX_BUG_ON_PARANOID(buffer
->len
);
818 EFX_BUG_ON_PARANOID(buffer
->unmap_len
);
819 EFX_BUG_ON_PARANOID(buffer
->skb
);
820 EFX_BUG_ON_PARANOID(!buffer
->continuation
);
821 EFX_BUG_ON_PARANOID(buffer
->tsoh
);
823 buffer
->dma_addr
= tsoh
->dma_addr
;
826 ++tx_queue
->insert_count
;
830 /* Remove descriptors put into a tx_queue. */
831 static void efx_enqueue_unwind(struct efx_tx_queue
*tx_queue
)
833 struct efx_tx_buffer
*buffer
;
834 dma_addr_t unmap_addr
;
836 /* Work backwards until we hit the original insert pointer value */
837 while (tx_queue
->insert_count
!= tx_queue
->write_count
) {
838 --tx_queue
->insert_count
;
839 buffer
= &tx_queue
->buffer
[tx_queue
->insert_count
&
841 efx_tsoh_free(tx_queue
, buffer
);
842 EFX_BUG_ON_PARANOID(buffer
->skb
);
843 if (buffer
->unmap_len
) {
844 unmap_addr
= (buffer
->dma_addr
+ buffer
->len
-
846 if (buffer
->unmap_single
)
847 pci_unmap_single(tx_queue
->efx
->pci_dev
,
848 unmap_addr
, buffer
->unmap_len
,
851 pci_unmap_page(tx_queue
->efx
->pci_dev
,
852 unmap_addr
, buffer
->unmap_len
,
854 buffer
->unmap_len
= 0;
857 buffer
->continuation
= true;
862 /* Parse the SKB header and initialise state. */
863 static void tso_start(struct tso_state
*st
, const struct sk_buff
*skb
)
865 /* All ethernet/IP/TCP headers combined size is TCP header size
866 * plus offset of TCP header relative to start of packet.
868 st
->header_len
= ((tcp_hdr(skb
)->doff
<< 2u)
869 + PTR_DIFF(tcp_hdr(skb
), skb
->data
));
870 st
->full_packet_size
= st
->header_len
+ skb_shinfo(skb
)->gso_size
;
872 if (st
->protocol
== htons(ETH_P_IP
))
873 st
->ipv4_id
= ntohs(ip_hdr(skb
)->id
);
876 st
->seqnum
= ntohl(tcp_hdr(skb
)->seq
);
878 EFX_BUG_ON_PARANOID(tcp_hdr(skb
)->urg
);
879 EFX_BUG_ON_PARANOID(tcp_hdr(skb
)->syn
);
880 EFX_BUG_ON_PARANOID(tcp_hdr(skb
)->rst
);
882 st
->packet_space
= st
->full_packet_size
;
883 st
->out_len
= skb
->len
- st
->header_len
;
885 st
->unmap_single
= false;
888 static int tso_get_fragment(struct tso_state
*st
, struct efx_nic
*efx
,
891 st
->unmap_addr
= pci_map_page(efx
->pci_dev
, frag
->page
,
892 frag
->page_offset
, frag
->size
,
894 if (likely(!pci_dma_mapping_error(efx
->pci_dev
, st
->unmap_addr
))) {
895 st
->unmap_single
= false;
896 st
->unmap_len
= frag
->size
;
897 st
->in_len
= frag
->size
;
898 st
->dma_addr
= st
->unmap_addr
;
904 static int tso_get_head_fragment(struct tso_state
*st
, struct efx_nic
*efx
,
905 const struct sk_buff
*skb
)
907 int hl
= st
->header_len
;
908 int len
= skb_headlen(skb
) - hl
;
910 st
->unmap_addr
= pci_map_single(efx
->pci_dev
, skb
->data
+ hl
,
911 len
, PCI_DMA_TODEVICE
);
912 if (likely(!pci_dma_mapping_error(efx
->pci_dev
, st
->unmap_addr
))) {
913 st
->unmap_single
= true;
916 st
->dma_addr
= st
->unmap_addr
;
924 * tso_fill_packet_with_fragment - form descriptors for the current fragment
925 * @tx_queue: Efx TX queue
926 * @skb: Socket buffer
929 * Form descriptors for the current fragment, until we reach the end
930 * of fragment or end-of-packet. Return 0 on success, 1 if not enough
931 * space in @tx_queue.
933 static int tso_fill_packet_with_fragment(struct efx_tx_queue
*tx_queue
,
934 const struct sk_buff
*skb
,
935 struct tso_state
*st
)
937 struct efx_tx_buffer
*buffer
;
938 int n
, end_of_packet
, rc
;
942 if (st
->packet_space
== 0)
945 EFX_BUG_ON_PARANOID(st
->in_len
<= 0);
946 EFX_BUG_ON_PARANOID(st
->packet_space
<= 0);
948 n
= min(st
->in_len
, st
->packet_space
);
950 st
->packet_space
-= n
;
954 rc
= efx_tx_queue_insert(tx_queue
, st
->dma_addr
, n
, &buffer
);
955 if (likely(rc
== 0)) {
956 if (st
->out_len
== 0)
957 /* Transfer ownership of the skb */
960 end_of_packet
= st
->out_len
== 0 || st
->packet_space
== 0;
961 buffer
->continuation
= !end_of_packet
;
963 if (st
->in_len
== 0) {
964 /* Transfer ownership of the pci mapping */
965 buffer
->unmap_len
= st
->unmap_len
;
966 buffer
->unmap_single
= st
->unmap_single
;
977 * tso_start_new_packet - generate a new header and prepare for the new packet
978 * @tx_queue: Efx TX queue
979 * @skb: Socket buffer
982 * Generate a new header and prepare for the new packet. Return 0 on
983 * success, or -1 if failed to alloc header.
985 static int tso_start_new_packet(struct efx_tx_queue
*tx_queue
,
986 const struct sk_buff
*skb
,
987 struct tso_state
*st
)
989 struct efx_tso_header
*tsoh
;
990 struct tcphdr
*tsoh_th
;
994 /* Allocate a DMA-mapped header buffer. */
995 if (likely(TSOH_SIZE(st
->header_len
) <= TSOH_STD_SIZE
)) {
996 if (tx_queue
->tso_headers_free
== NULL
) {
997 if (efx_tsoh_block_alloc(tx_queue
))
1000 EFX_BUG_ON_PARANOID(!tx_queue
->tso_headers_free
);
1001 tsoh
= tx_queue
->tso_headers_free
;
1002 tx_queue
->tso_headers_free
= tsoh
->next
;
1003 tsoh
->unmap_len
= 0;
1005 tx_queue
->tso_long_headers
++;
1006 tsoh
= efx_tsoh_heap_alloc(tx_queue
, st
->header_len
);
1007 if (unlikely(!tsoh
))
1011 header
= TSOH_BUFFER(tsoh
);
1012 tsoh_th
= (struct tcphdr
*)(header
+ SKB_TCP_OFF(skb
));
1014 /* Copy and update the headers. */
1015 memcpy(header
, skb
->data
, st
->header_len
);
1017 tsoh_th
->seq
= htonl(st
->seqnum
);
1018 st
->seqnum
+= skb_shinfo(skb
)->gso_size
;
1019 if (st
->out_len
> skb_shinfo(skb
)->gso_size
) {
1020 /* This packet will not finish the TSO burst. */
1021 ip_length
= st
->full_packet_size
- ETH_HDR_LEN(skb
);
1025 /* This packet will be the last in the TSO burst. */
1026 ip_length
= st
->header_len
- ETH_HDR_LEN(skb
) + st
->out_len
;
1027 tsoh_th
->fin
= tcp_hdr(skb
)->fin
;
1028 tsoh_th
->psh
= tcp_hdr(skb
)->psh
;
1031 if (st
->protocol
== htons(ETH_P_IP
)) {
1032 struct iphdr
*tsoh_iph
=
1033 (struct iphdr
*)(header
+ SKB_IPV4_OFF(skb
));
1035 tsoh_iph
->tot_len
= htons(ip_length
);
1037 /* Linux leaves suitable gaps in the IP ID space for us to fill. */
1038 tsoh_iph
->id
= htons(st
->ipv4_id
);
1041 struct ipv6hdr
*tsoh_iph
=
1042 (struct ipv6hdr
*)(header
+ SKB_IPV6_OFF(skb
));
1044 tsoh_iph
->payload_len
= htons(ip_length
- sizeof(*tsoh_iph
));
1047 st
->packet_space
= skb_shinfo(skb
)->gso_size
;
1048 ++tx_queue
->tso_packets
;
1050 /* Form a descriptor for this header. */
1051 efx_tso_put_header(tx_queue
, tsoh
, st
->header_len
);
1058 * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1059 * @tx_queue: Efx TX queue
1060 * @skb: Socket buffer
1062 * Context: You must hold netif_tx_lock() to call this function.
1064 * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1065 * @skb was not enqueued. In all cases @skb is consumed. Return
1066 * %NETDEV_TX_OK or %NETDEV_TX_BUSY.
1068 static int efx_enqueue_skb_tso(struct efx_tx_queue
*tx_queue
,
1069 struct sk_buff
*skb
)
1071 struct efx_nic
*efx
= tx_queue
->efx
;
1072 int frag_i
, rc
, rc2
= NETDEV_TX_OK
;
1073 struct tso_state state
;
1075 /* Find the packet protocol and sanity-check it */
1076 state
.protocol
= efx_tso_check_protocol(skb
);
1078 EFX_BUG_ON_PARANOID(tx_queue
->write_count
!= tx_queue
->insert_count
);
1080 tso_start(&state
, skb
);
1082 /* Assume that skb header area contains exactly the headers, and
1083 * all payload is in the frag list.
1085 if (skb_headlen(skb
) == state
.header_len
) {
1086 /* Grab the first payload fragment. */
1087 EFX_BUG_ON_PARANOID(skb_shinfo(skb
)->nr_frags
< 1);
1089 rc
= tso_get_fragment(&state
, efx
,
1090 skb_shinfo(skb
)->frags
+ frag_i
);
1094 rc
= tso_get_head_fragment(&state
, efx
, skb
);
1100 if (tso_start_new_packet(tx_queue
, skb
, &state
) < 0)
1104 rc
= tso_fill_packet_with_fragment(tx_queue
, skb
, &state
);
1108 /* Move onto the next fragment? */
1109 if (state
.in_len
== 0) {
1110 if (++frag_i
>= skb_shinfo(skb
)->nr_frags
)
1111 /* End of payload reached. */
1113 rc
= tso_get_fragment(&state
, efx
,
1114 skb_shinfo(skb
)->frags
+ frag_i
);
1119 /* Start at new packet? */
1120 if (state
.packet_space
== 0 &&
1121 tso_start_new_packet(tx_queue
, skb
, &state
) < 0)
1125 /* Pass off to hardware */
1126 efx_nic_push_buffers(tx_queue
);
1128 tx_queue
->tso_bursts
++;
1129 return NETDEV_TX_OK
;
1132 netif_err(efx
, tx_err
, efx
->net_dev
,
1133 "Out of memory for TSO headers, or PCI mapping error\n");
1134 dev_kfree_skb_any(skb
);
1138 rc2
= NETDEV_TX_BUSY
;
1140 /* Stop the queue if it wasn't stopped before. */
1141 if (tx_queue
->stopped
== 1)
1142 efx_stop_queue(tx_queue
->channel
);
1145 /* Free the DMA mapping we were in the process of writing out */
1146 if (state
.unmap_len
) {
1147 if (state
.unmap_single
)
1148 pci_unmap_single(efx
->pci_dev
, state
.unmap_addr
,
1149 state
.unmap_len
, PCI_DMA_TODEVICE
);
1151 pci_unmap_page(efx
->pci_dev
, state
.unmap_addr
,
1152 state
.unmap_len
, PCI_DMA_TODEVICE
);
1155 efx_enqueue_unwind(tx_queue
);
1161 * Free up all TSO datastructures associated with tx_queue. This
1162 * routine should be called only once the tx_queue is both empty and
1163 * will no longer be used.
1165 static void efx_fini_tso(struct efx_tx_queue
*tx_queue
)
1169 if (tx_queue
->buffer
) {
1170 for (i
= 0; i
<= EFX_TXQ_MASK
; ++i
)
1171 efx_tsoh_free(tx_queue
, &tx_queue
->buffer
[i
]);
1174 while (tx_queue
->tso_headers_free
!= NULL
)
1175 efx_tsoh_block_free(tx_queue
, tx_queue
->tso_headers_free
,
1176 tx_queue
->efx
->pci_dev
);