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/pci.h>
12 #include <linux/tcp.h>
15 #include <linux/if_ether.h>
16 #include <linux/highmem.h>
17 #include "net_driver.h"
21 #include "workarounds.h"
24 * TX descriptor ring full threshold
26 * The tx_queue descriptor ring fill-level must fall below this value
27 * before we restart the netif queue
29 #define EFX_NETDEV_TX_THRESHOLD(_tx_queue) \
30 (_tx_queue->efx->type->txd_ring_mask / 2u)
32 /* We want to be able to nest calls to netif_stop_queue(), since each
33 * channel can have an individual stop on the queue.
35 void efx_stop_queue(struct efx_nic
*efx
)
37 spin_lock_bh(&efx
->netif_stop_lock
);
38 EFX_TRACE(efx
, "stop TX queue\n");
40 atomic_inc(&efx
->netif_stop_count
);
41 netif_stop_queue(efx
->net_dev
);
43 spin_unlock_bh(&efx
->netif_stop_lock
);
46 /* Wake netif's TX queue
47 * We want to be able to nest calls to netif_stop_queue(), since each
48 * channel can have an individual stop on the queue.
50 inline void efx_wake_queue(struct efx_nic
*efx
)
53 if (atomic_dec_and_lock(&efx
->netif_stop_count
,
54 &efx
->netif_stop_lock
)) {
55 EFX_TRACE(efx
, "waking TX queue\n");
56 netif_wake_queue(efx
->net_dev
);
57 spin_unlock(&efx
->netif_stop_lock
);
62 static inline void efx_dequeue_buffer(struct efx_tx_queue
*tx_queue
,
63 struct efx_tx_buffer
*buffer
)
65 if (buffer
->unmap_len
) {
66 struct pci_dev
*pci_dev
= tx_queue
->efx
->pci_dev
;
67 dma_addr_t unmap_addr
= (buffer
->dma_addr
+ buffer
->len
-
69 if (buffer
->unmap_single
)
70 pci_unmap_single(pci_dev
, unmap_addr
, buffer
->unmap_len
,
73 pci_unmap_page(pci_dev
, unmap_addr
, buffer
->unmap_len
,
75 buffer
->unmap_len
= 0;
76 buffer
->unmap_single
= false;
80 dev_kfree_skb_any((struct sk_buff
*) buffer
->skb
);
82 EFX_TRACE(tx_queue
->efx
, "TX queue %d transmission id %x "
83 "complete\n", tx_queue
->queue
, read_ptr
);
88 * struct efx_tso_header - a DMA mapped buffer for packet headers
89 * @next: Linked list of free ones.
90 * The list is protected by the TX queue lock.
91 * @dma_unmap_len: Length to unmap for an oversize buffer, or 0.
92 * @dma_addr: The DMA address of the header below.
94 * This controls the memory used for a TSO header. Use TSOH_DATA()
95 * to find the packet header data. Use TSOH_SIZE() to calculate the
96 * total size required for a given packet header length. TSO headers
97 * in the free list are exactly %TSOH_STD_SIZE bytes in size.
99 struct efx_tso_header
{
101 struct efx_tso_header
*next
;
107 static int efx_enqueue_skb_tso(struct efx_tx_queue
*tx_queue
,
108 const struct sk_buff
*skb
);
109 static void efx_fini_tso(struct efx_tx_queue
*tx_queue
);
110 static void efx_tsoh_heap_free(struct efx_tx_queue
*tx_queue
,
111 struct efx_tso_header
*tsoh
);
113 static inline void efx_tsoh_free(struct efx_tx_queue
*tx_queue
,
114 struct efx_tx_buffer
*buffer
)
117 if (likely(!buffer
->tsoh
->unmap_len
)) {
118 buffer
->tsoh
->next
= tx_queue
->tso_headers_free
;
119 tx_queue
->tso_headers_free
= buffer
->tsoh
;
121 efx_tsoh_heap_free(tx_queue
, buffer
->tsoh
);
129 * Add a socket buffer to a TX queue
131 * This maps all fragments of a socket buffer for DMA and adds them to
132 * the TX queue. The queue's insert pointer will be incremented by
133 * the number of fragments in the socket buffer.
135 * If any DMA mapping fails, any mapped fragments will be unmapped,
136 * the queue's insert pointer will be restored to its original value.
138 * Returns NETDEV_TX_OK or NETDEV_TX_BUSY
139 * You must hold netif_tx_lock() to call this function.
141 static inline int efx_enqueue_skb(struct efx_tx_queue
*tx_queue
,
142 const struct sk_buff
*skb
)
144 struct efx_nic
*efx
= tx_queue
->efx
;
145 struct pci_dev
*pci_dev
= efx
->pci_dev
;
146 struct efx_tx_buffer
*buffer
;
147 skb_frag_t
*fragment
;
150 unsigned int len
, unmap_len
= 0, fill_level
, insert_ptr
, misalign
;
151 dma_addr_t dma_addr
, unmap_addr
= 0;
152 unsigned int dma_len
;
155 int rc
= NETDEV_TX_OK
;
157 EFX_BUG_ON_PARANOID(tx_queue
->write_count
!= tx_queue
->insert_count
);
159 if (skb_shinfo((struct sk_buff
*)skb
)->gso_size
)
160 return efx_enqueue_skb_tso(tx_queue
, skb
);
162 /* Get size of the initial fragment */
163 len
= skb_headlen(skb
);
165 fill_level
= tx_queue
->insert_count
- tx_queue
->old_read_count
;
166 q_space
= efx
->type
->txd_ring_mask
- 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.
194 /* This memory barrier protects the
195 * change of stopped from the access
198 tx_queue
->old_read_count
=
199 *(volatile unsigned *)
200 &tx_queue
->read_count
;
201 fill_level
= (tx_queue
->insert_count
202 - tx_queue
->old_read_count
);
203 q_space
= (efx
->type
->txd_ring_mask
- 1 -
205 if (unlikely(q_space
-- <= 0))
211 insert_ptr
= (tx_queue
->insert_count
&
212 efx
->type
->txd_ring_mask
);
213 buffer
= &tx_queue
->buffer
[insert_ptr
];
214 efx_tsoh_free(tx_queue
, buffer
);
215 EFX_BUG_ON_PARANOID(buffer
->tsoh
);
216 EFX_BUG_ON_PARANOID(buffer
->skb
);
217 EFX_BUG_ON_PARANOID(buffer
->len
);
218 EFX_BUG_ON_PARANOID(!buffer
->continuation
);
219 EFX_BUG_ON_PARANOID(buffer
->unmap_len
);
221 dma_len
= (((~dma_addr
) & efx
->type
->tx_dma_mask
) + 1);
222 if (likely(dma_len
> len
))
225 misalign
= (unsigned)dma_addr
& efx
->type
->bug5391_mask
;
226 if (misalign
&& dma_len
+ misalign
> 512)
227 dma_len
= 512 - misalign
;
229 /* Fill out per descriptor fields */
230 buffer
->len
= dma_len
;
231 buffer
->dma_addr
= dma_addr
;
234 ++tx_queue
->insert_count
;
237 /* Transfer ownership of the unmapping to the final buffer */
238 buffer
->unmap_single
= unmap_single
;
239 buffer
->unmap_len
= unmap_len
;
242 /* Get address and size of next fragment */
243 if (i
>= skb_shinfo(skb
)->nr_frags
)
245 fragment
= &skb_shinfo(skb
)->frags
[i
];
246 len
= fragment
->size
;
247 page
= fragment
->page
;
248 page_offset
= fragment
->page_offset
;
251 unmap_single
= false;
252 dma_addr
= pci_map_page(pci_dev
, page
, page_offset
, len
,
256 /* Transfer ownership of the skb to the final buffer */
258 buffer
->continuation
= false;
260 /* Pass off to hardware */
261 falcon_push_buffers(tx_queue
);
266 EFX_ERR_RL(efx
, " TX queue %d could not map skb with %d bytes %d "
267 "fragments for DMA\n", tx_queue
->queue
, skb
->len
,
268 skb_shinfo(skb
)->nr_frags
+ 1);
270 /* Mark the packet as transmitted, and free the SKB ourselves */
271 dev_kfree_skb_any((struct sk_buff
*)skb
);
277 if (tx_queue
->stopped
== 1)
281 /* Work backwards until we hit the original insert pointer value */
282 while (tx_queue
->insert_count
!= tx_queue
->write_count
) {
283 --tx_queue
->insert_count
;
284 insert_ptr
= tx_queue
->insert_count
& efx
->type
->txd_ring_mask
;
285 buffer
= &tx_queue
->buffer
[insert_ptr
];
286 efx_dequeue_buffer(tx_queue
, buffer
);
290 /* Free the fragment we were mid-way through pushing */
293 pci_unmap_single(pci_dev
, unmap_addr
, unmap_len
,
296 pci_unmap_page(pci_dev
, unmap_addr
, unmap_len
,
303 /* Remove packets from the TX queue
305 * This removes packets from the TX queue, up to and including the
308 static inline void efx_dequeue_buffers(struct efx_tx_queue
*tx_queue
,
311 struct efx_nic
*efx
= tx_queue
->efx
;
312 unsigned int stop_index
, read_ptr
;
313 unsigned int mask
= tx_queue
->efx
->type
->txd_ring_mask
;
315 stop_index
= (index
+ 1) & mask
;
316 read_ptr
= tx_queue
->read_count
& mask
;
318 while (read_ptr
!= stop_index
) {
319 struct efx_tx_buffer
*buffer
= &tx_queue
->buffer
[read_ptr
];
320 if (unlikely(buffer
->len
== 0)) {
321 EFX_ERR(tx_queue
->efx
, "TX queue %d spurious TX "
322 "completion id %x\n", tx_queue
->queue
,
324 efx_schedule_reset(efx
, RESET_TYPE_TX_SKIP
);
328 efx_dequeue_buffer(tx_queue
, buffer
);
329 buffer
->continuation
= true;
332 ++tx_queue
->read_count
;
333 read_ptr
= tx_queue
->read_count
& mask
;
337 /* Initiate a packet transmission on the specified TX queue.
338 * Note that returning anything other than NETDEV_TX_OK will cause the
339 * OS to free the skb.
341 * This function is split out from efx_hard_start_xmit to allow the
342 * loopback test to direct packets via specific TX queues. It is
343 * therefore a non-static inline, so as not to penalise performance
344 * for non-loopback transmissions.
346 * Context: netif_tx_lock held
348 inline int efx_xmit(struct efx_nic
*efx
,
349 struct efx_tx_queue
*tx_queue
, struct sk_buff
*skb
)
353 /* Map fragments for DMA and add to TX queue */
354 rc
= efx_enqueue_skb(tx_queue
, skb
);
355 if (unlikely(rc
!= NETDEV_TX_OK
))
358 /* Update last TX timer */
359 efx
->net_dev
->trans_start
= jiffies
;
365 /* Initiate a packet transmission. We use one channel per CPU
366 * (sharing when we have more CPUs than channels). On Falcon, the TX
367 * completion events will be directed back to the CPU that transmitted
368 * the packet, which should be cache-efficient.
370 * Context: non-blocking.
371 * Note that returning anything other than NETDEV_TX_OK will cause the
372 * OS to free the skb.
374 int efx_hard_start_xmit(struct sk_buff
*skb
, struct net_device
*net_dev
)
376 struct efx_nic
*efx
= netdev_priv(net_dev
);
377 struct efx_tx_queue
*tx_queue
;
379 if (likely(skb
->ip_summed
== CHECKSUM_PARTIAL
))
380 tx_queue
= &efx
->tx_queue
[EFX_TX_QUEUE_OFFLOAD_CSUM
];
382 tx_queue
= &efx
->tx_queue
[EFX_TX_QUEUE_NO_CSUM
];
384 return efx_xmit(efx
, tx_queue
, skb
);
387 void efx_xmit_done(struct efx_tx_queue
*tx_queue
, unsigned int index
)
390 struct efx_nic
*efx
= tx_queue
->efx
;
392 EFX_BUG_ON_PARANOID(index
> efx
->type
->txd_ring_mask
);
394 efx_dequeue_buffers(tx_queue
, index
);
396 /* See if we need to restart the netif queue. This barrier
397 * separates the update of read_count from the test of
400 if (unlikely(tx_queue
->stopped
)) {
401 fill_level
= tx_queue
->insert_count
- tx_queue
->read_count
;
402 if (fill_level
< EFX_NETDEV_TX_THRESHOLD(tx_queue
)) {
403 EFX_BUG_ON_PARANOID(!efx_dev_registered(efx
));
405 /* Do this under netif_tx_lock(), to avoid racing
406 * with efx_xmit(). */
407 netif_tx_lock(efx
->net_dev
);
408 if (tx_queue
->stopped
) {
409 tx_queue
->stopped
= 0;
412 netif_tx_unlock(efx
->net_dev
);
417 int efx_probe_tx_queue(struct efx_tx_queue
*tx_queue
)
419 struct efx_nic
*efx
= tx_queue
->efx
;
420 unsigned int txq_size
;
423 EFX_LOG(efx
, "creating TX queue %d\n", tx_queue
->queue
);
425 /* Allocate software ring */
426 txq_size
= (efx
->type
->txd_ring_mask
+ 1) * sizeof(*tx_queue
->buffer
);
427 tx_queue
->buffer
= kzalloc(txq_size
, GFP_KERNEL
);
428 if (!tx_queue
->buffer
)
430 for (i
= 0; i
<= efx
->type
->txd_ring_mask
; ++i
)
431 tx_queue
->buffer
[i
].continuation
= true;
433 /* Allocate hardware ring */
434 rc
= falcon_probe_tx(tx_queue
);
441 kfree(tx_queue
->buffer
);
442 tx_queue
->buffer
= NULL
;
446 int efx_init_tx_queue(struct efx_tx_queue
*tx_queue
)
448 EFX_LOG(tx_queue
->efx
, "initialising TX queue %d\n", tx_queue
->queue
);
450 tx_queue
->insert_count
= 0;
451 tx_queue
->write_count
= 0;
452 tx_queue
->read_count
= 0;
453 tx_queue
->old_read_count
= 0;
454 BUG_ON(tx_queue
->stopped
);
456 /* Set up TX descriptor ring */
457 return falcon_init_tx(tx_queue
);
460 void efx_release_tx_buffers(struct efx_tx_queue
*tx_queue
)
462 struct efx_tx_buffer
*buffer
;
464 if (!tx_queue
->buffer
)
467 /* Free any buffers left in the ring */
468 while (tx_queue
->read_count
!= tx_queue
->write_count
) {
469 buffer
= &tx_queue
->buffer
[tx_queue
->read_count
&
470 tx_queue
->efx
->type
->txd_ring_mask
];
471 efx_dequeue_buffer(tx_queue
, buffer
);
472 buffer
->continuation
= true;
475 ++tx_queue
->read_count
;
479 void efx_fini_tx_queue(struct efx_tx_queue
*tx_queue
)
481 EFX_LOG(tx_queue
->efx
, "shutting down TX queue %d\n", tx_queue
->queue
);
483 /* Flush TX queue, remove descriptor ring */
484 falcon_fini_tx(tx_queue
);
486 efx_release_tx_buffers(tx_queue
);
488 /* Free up TSO header cache */
489 efx_fini_tso(tx_queue
);
491 /* Release queue's stop on port, if any */
492 if (tx_queue
->stopped
) {
493 tx_queue
->stopped
= 0;
494 efx_wake_queue(tx_queue
->efx
);
498 void efx_remove_tx_queue(struct efx_tx_queue
*tx_queue
)
500 EFX_LOG(tx_queue
->efx
, "destroying TX queue %d\n", tx_queue
->queue
);
501 falcon_remove_tx(tx_queue
);
503 kfree(tx_queue
->buffer
);
504 tx_queue
->buffer
= NULL
;
508 /* Efx TCP segmentation acceleration.
510 * Why? Because by doing it here in the driver we can go significantly
511 * faster than the GSO.
513 * Requires TX checksum offload support.
516 /* Number of bytes inserted at the start of a TSO header buffer,
517 * similar to NET_IP_ALIGN.
519 #if defined(__i386__) || defined(__x86_64__)
520 #define TSOH_OFFSET 0
522 #define TSOH_OFFSET NET_IP_ALIGN
525 #define TSOH_BUFFER(tsoh) ((u8 *)(tsoh + 1) + TSOH_OFFSET)
527 /* Total size of struct efx_tso_header, buffer and padding */
528 #define TSOH_SIZE(hdr_len) \
529 (sizeof(struct efx_tso_header) + TSOH_OFFSET + hdr_len)
531 /* Size of blocks on free list. Larger blocks must be allocated from
534 #define TSOH_STD_SIZE 128
536 #define PTR_DIFF(p1, p2) ((u8 *)(p1) - (u8 *)(p2))
537 #define ETH_HDR_LEN(skb) (skb_network_header(skb) - (skb)->data)
538 #define SKB_TCP_OFF(skb) PTR_DIFF(tcp_hdr(skb), (skb)->data)
539 #define SKB_IPV4_OFF(skb) PTR_DIFF(ip_hdr(skb), (skb)->data)
542 * struct tso_state - TSO state for an SKB
543 * @remaining_len: Bytes of data we've yet to segment
544 * @seqnum: Current sequence number
545 * @packet_space: Remaining space in current packet
546 * @ifc: Input fragment cursor.
547 * Where we are in the current fragment of the incoming SKB. These
548 * values get updated in place when we split a fragment over
551 * These values are set once at the start of the TSO send and do
552 * not get changed as the routine progresses.
554 * The state used during segmentation. It is put into this data structure
555 * just to make it easy to pass into inline functions.
558 unsigned remaining_len
;
560 unsigned packet_space
;
563 /* DMA address of current position */
565 /* Remaining length */
567 /* DMA address and length of the whole fragment */
568 unsigned int unmap_len
;
569 dma_addr_t unmap_addr
;
574 /* The number of bytes of header */
575 unsigned int header_length
;
577 /* The number of bytes to put in each outgoing segment. */
578 int full_packet_size
;
580 /* Current IPv4 ID, host endian. */
587 * Verify that our various assumptions about sk_buffs and the conditions
588 * under which TSO will be attempted hold true.
590 static inline void efx_tso_check_safe(const struct sk_buff
*skb
)
592 EFX_BUG_ON_PARANOID(skb
->protocol
!= htons(ETH_P_IP
));
593 EFX_BUG_ON_PARANOID(((struct ethhdr
*)skb
->data
)->h_proto
!=
595 EFX_BUG_ON_PARANOID(ip_hdr(skb
)->protocol
!= IPPROTO_TCP
);
596 EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb
), skb
->data
)
597 + (tcp_hdr(skb
)->doff
<< 2u)) >
603 * Allocate a page worth of efx_tso_header structures, and string them
604 * into the tx_queue->tso_headers_free linked list. Return 0 or -ENOMEM.
606 static int efx_tsoh_block_alloc(struct efx_tx_queue
*tx_queue
)
609 struct pci_dev
*pci_dev
= tx_queue
->efx
->pci_dev
;
610 struct efx_tso_header
*tsoh
;
614 base_kva
= pci_alloc_consistent(pci_dev
, PAGE_SIZE
, &dma_addr
);
615 if (base_kva
== NULL
) {
616 EFX_ERR(tx_queue
->efx
, "Unable to allocate page for TSO"
621 /* pci_alloc_consistent() allocates pages. */
622 EFX_BUG_ON_PARANOID(dma_addr
& (PAGE_SIZE
- 1u));
624 for (kva
= base_kva
; kva
< base_kva
+ PAGE_SIZE
; kva
+= TSOH_STD_SIZE
) {
625 tsoh
= (struct efx_tso_header
*)kva
;
626 tsoh
->dma_addr
= dma_addr
+ (TSOH_BUFFER(tsoh
) - base_kva
);
627 tsoh
->next
= tx_queue
->tso_headers_free
;
628 tx_queue
->tso_headers_free
= tsoh
;
635 /* Free up a TSO header, and all others in the same page. */
636 static void efx_tsoh_block_free(struct efx_tx_queue
*tx_queue
,
637 struct efx_tso_header
*tsoh
,
638 struct pci_dev
*pci_dev
)
640 struct efx_tso_header
**p
;
641 unsigned long base_kva
;
644 base_kva
= (unsigned long)tsoh
& PAGE_MASK
;
645 base_dma
= tsoh
->dma_addr
& PAGE_MASK
;
647 p
= &tx_queue
->tso_headers_free
;
649 if (((unsigned long)*p
& PAGE_MASK
) == base_kva
)
655 pci_free_consistent(pci_dev
, PAGE_SIZE
, (void *)base_kva
, base_dma
);
658 static struct efx_tso_header
*
659 efx_tsoh_heap_alloc(struct efx_tx_queue
*tx_queue
, size_t header_len
)
661 struct efx_tso_header
*tsoh
;
663 tsoh
= kmalloc(TSOH_SIZE(header_len
), GFP_ATOMIC
| GFP_DMA
);
667 tsoh
->dma_addr
= pci_map_single(tx_queue
->efx
->pci_dev
,
668 TSOH_BUFFER(tsoh
), header_len
,
670 if (unlikely(pci_dma_mapping_error(tx_queue
->efx
->pci_dev
,
676 tsoh
->unmap_len
= header_len
;
681 efx_tsoh_heap_free(struct efx_tx_queue
*tx_queue
, struct efx_tso_header
*tsoh
)
683 pci_unmap_single(tx_queue
->efx
->pci_dev
,
684 tsoh
->dma_addr
, tsoh
->unmap_len
,
690 * efx_tx_queue_insert - push descriptors onto the TX queue
691 * @tx_queue: Efx TX queue
692 * @dma_addr: DMA address of fragment
693 * @len: Length of fragment
694 * @final_buffer: The final buffer inserted into the queue
696 * Push descriptors onto the TX queue. Return 0 on success or 1 if
699 static int efx_tx_queue_insert(struct efx_tx_queue
*tx_queue
,
700 dma_addr_t dma_addr
, unsigned len
,
701 struct efx_tx_buffer
**final_buffer
)
703 struct efx_tx_buffer
*buffer
;
704 struct efx_nic
*efx
= tx_queue
->efx
;
705 unsigned dma_len
, fill_level
, insert_ptr
, misalign
;
708 EFX_BUG_ON_PARANOID(len
<= 0);
710 fill_level
= tx_queue
->insert_count
- tx_queue
->old_read_count
;
711 /* -1 as there is no way to represent all descriptors used */
712 q_space
= efx
->type
->txd_ring_mask
- 1 - fill_level
;
715 if (unlikely(q_space
-- <= 0)) {
716 /* It might be that completions have happened
717 * since the xmit path last checked. Update
718 * the xmit path's copy of read_count.
721 /* This memory barrier protects the change of
722 * stopped from the access of read_count. */
724 tx_queue
->old_read_count
=
725 *(volatile unsigned *)&tx_queue
->read_count
;
726 fill_level
= (tx_queue
->insert_count
727 - tx_queue
->old_read_count
);
728 q_space
= efx
->type
->txd_ring_mask
- 1 - fill_level
;
729 if (unlikely(q_space
-- <= 0)) {
730 *final_buffer
= NULL
;
737 insert_ptr
= tx_queue
->insert_count
& efx
->type
->txd_ring_mask
;
738 buffer
= &tx_queue
->buffer
[insert_ptr
];
739 ++tx_queue
->insert_count
;
741 EFX_BUG_ON_PARANOID(tx_queue
->insert_count
-
742 tx_queue
->read_count
>
743 efx
->type
->txd_ring_mask
);
745 efx_tsoh_free(tx_queue
, buffer
);
746 EFX_BUG_ON_PARANOID(buffer
->len
);
747 EFX_BUG_ON_PARANOID(buffer
->unmap_len
);
748 EFX_BUG_ON_PARANOID(buffer
->skb
);
749 EFX_BUG_ON_PARANOID(!buffer
->continuation
);
750 EFX_BUG_ON_PARANOID(buffer
->tsoh
);
752 buffer
->dma_addr
= dma_addr
;
754 /* Ensure we do not cross a boundary unsupported by H/W */
755 dma_len
= (~dma_addr
& efx
->type
->tx_dma_mask
) + 1;
757 misalign
= (unsigned)dma_addr
& efx
->type
->bug5391_mask
;
758 if (misalign
&& dma_len
+ misalign
> 512)
759 dma_len
= 512 - misalign
;
761 /* If there is enough space to send then do so */
765 buffer
->len
= dma_len
; /* Don't set the other members */
770 EFX_BUG_ON_PARANOID(!len
);
772 *final_buffer
= buffer
;
778 * Put a TSO header into the TX queue.
780 * This is special-cased because we know that it is small enough to fit in
781 * a single fragment, and we know it doesn't cross a page boundary. It
782 * also allows us to not worry about end-of-packet etc.
784 static inline void efx_tso_put_header(struct efx_tx_queue
*tx_queue
,
785 struct efx_tso_header
*tsoh
, unsigned len
)
787 struct efx_tx_buffer
*buffer
;
789 buffer
= &tx_queue
->buffer
[tx_queue
->insert_count
&
790 tx_queue
->efx
->type
->txd_ring_mask
];
791 efx_tsoh_free(tx_queue
, buffer
);
792 EFX_BUG_ON_PARANOID(buffer
->len
);
793 EFX_BUG_ON_PARANOID(buffer
->unmap_len
);
794 EFX_BUG_ON_PARANOID(buffer
->skb
);
795 EFX_BUG_ON_PARANOID(!buffer
->continuation
);
796 EFX_BUG_ON_PARANOID(buffer
->tsoh
);
798 buffer
->dma_addr
= tsoh
->dma_addr
;
801 ++tx_queue
->insert_count
;
805 /* Remove descriptors put into a tx_queue. */
806 static void efx_enqueue_unwind(struct efx_tx_queue
*tx_queue
)
808 struct efx_tx_buffer
*buffer
;
809 dma_addr_t unmap_addr
;
811 /* Work backwards until we hit the original insert pointer value */
812 while (tx_queue
->insert_count
!= tx_queue
->write_count
) {
813 --tx_queue
->insert_count
;
814 buffer
= &tx_queue
->buffer
[tx_queue
->insert_count
&
815 tx_queue
->efx
->type
->txd_ring_mask
];
816 efx_tsoh_free(tx_queue
, buffer
);
817 EFX_BUG_ON_PARANOID(buffer
->skb
);
819 buffer
->continuation
= true;
820 if (buffer
->unmap_len
) {
821 unmap_addr
= (buffer
->dma_addr
+ buffer
->len
-
823 if (buffer
->unmap_single
)
824 pci_unmap_single(tx_queue
->efx
->pci_dev
,
825 unmap_addr
, buffer
->unmap_len
,
828 pci_unmap_page(tx_queue
->efx
->pci_dev
,
829 unmap_addr
, buffer
->unmap_len
,
831 buffer
->unmap_len
= 0;
837 /* Parse the SKB header and initialise state. */
838 static inline void tso_start(struct tso_state
*st
, const struct sk_buff
*skb
)
840 /* All ethernet/IP/TCP headers combined size is TCP header size
841 * plus offset of TCP header relative to start of packet.
843 st
->p
.header_length
= ((tcp_hdr(skb
)->doff
<< 2u)
844 + PTR_DIFF(tcp_hdr(skb
), skb
->data
));
845 st
->p
.full_packet_size
= (st
->p
.header_length
846 + skb_shinfo(skb
)->gso_size
);
848 st
->p
.ipv4_id
= ntohs(ip_hdr(skb
)->id
);
849 st
->seqnum
= ntohl(tcp_hdr(skb
)->seq
);
851 EFX_BUG_ON_PARANOID(tcp_hdr(skb
)->urg
);
852 EFX_BUG_ON_PARANOID(tcp_hdr(skb
)->syn
);
853 EFX_BUG_ON_PARANOID(tcp_hdr(skb
)->rst
);
855 st
->packet_space
= st
->p
.full_packet_size
;
856 st
->remaining_len
= skb
->len
- st
->p
.header_length
;
857 st
->ifc
.unmap_len
= 0;
858 st
->ifc
.unmap_single
= false;
861 static inline int tso_get_fragment(struct tso_state
*st
, struct efx_nic
*efx
,
864 st
->ifc
.unmap_addr
= pci_map_page(efx
->pci_dev
, frag
->page
,
865 frag
->page_offset
, frag
->size
,
867 if (likely(!pci_dma_mapping_error(efx
->pci_dev
, st
->ifc
.unmap_addr
))) {
868 st
->ifc
.unmap_single
= false;
869 st
->ifc
.unmap_len
= frag
->size
;
870 st
->ifc
.len
= frag
->size
;
871 st
->ifc
.dma_addr
= st
->ifc
.unmap_addr
;
878 tso_get_head_fragment(struct tso_state
*st
, struct efx_nic
*efx
,
879 const struct sk_buff
*skb
)
881 int hl
= st
->p
.header_length
;
882 int len
= skb_headlen(skb
) - hl
;
884 st
->ifc
.unmap_addr
= pci_map_single(efx
->pci_dev
, skb
->data
+ hl
,
885 len
, PCI_DMA_TODEVICE
);
886 if (likely(!pci_dma_mapping_error(efx
->pci_dev
, st
->ifc
.unmap_addr
))) {
887 st
->ifc
.unmap_single
= true;
888 st
->ifc
.unmap_len
= len
;
890 st
->ifc
.dma_addr
= st
->ifc
.unmap_addr
;
898 * tso_fill_packet_with_fragment - form descriptors for the current fragment
899 * @tx_queue: Efx TX queue
900 * @skb: Socket buffer
903 * Form descriptors for the current fragment, until we reach the end
904 * of fragment or end-of-packet. Return 0 on success, 1 if not enough
905 * space in @tx_queue.
907 static inline int tso_fill_packet_with_fragment(struct efx_tx_queue
*tx_queue
,
908 const struct sk_buff
*skb
,
909 struct tso_state
*st
)
911 struct efx_tx_buffer
*buffer
;
912 int n
, end_of_packet
, rc
;
914 if (st
->ifc
.len
== 0)
916 if (st
->packet_space
== 0)
919 EFX_BUG_ON_PARANOID(st
->ifc
.len
<= 0);
920 EFX_BUG_ON_PARANOID(st
->packet_space
<= 0);
922 n
= min(st
->ifc
.len
, st
->packet_space
);
924 st
->packet_space
-= n
;
925 st
->remaining_len
-= n
;
928 rc
= efx_tx_queue_insert(tx_queue
, st
->ifc
.dma_addr
, n
, &buffer
);
929 if (likely(rc
== 0)) {
930 if (st
->remaining_len
== 0)
931 /* Transfer ownership of the skb */
934 end_of_packet
= st
->remaining_len
== 0 || st
->packet_space
== 0;
935 buffer
->continuation
= !end_of_packet
;
937 if (st
->ifc
.len
== 0) {
938 /* Transfer ownership of the pci mapping */
939 buffer
->unmap_len
= st
->ifc
.unmap_len
;
940 buffer
->unmap_single
= st
->ifc
.unmap_single
;
941 st
->ifc
.unmap_len
= 0;
945 st
->ifc
.dma_addr
+= n
;
951 * tso_start_new_packet - generate a new header and prepare for the new packet
952 * @tx_queue: Efx TX queue
953 * @skb: Socket buffer
956 * Generate a new header and prepare for the new packet. Return 0 on
957 * success, or -1 if failed to alloc header.
959 static inline int tso_start_new_packet(struct efx_tx_queue
*tx_queue
,
960 const struct sk_buff
*skb
,
961 struct tso_state
*st
)
963 struct efx_tso_header
*tsoh
;
964 struct iphdr
*tsoh_iph
;
965 struct tcphdr
*tsoh_th
;
969 /* Allocate a DMA-mapped header buffer. */
970 if (likely(TSOH_SIZE(st
->p
.header_length
) <= TSOH_STD_SIZE
)) {
971 if (tx_queue
->tso_headers_free
== NULL
) {
972 if (efx_tsoh_block_alloc(tx_queue
))
975 EFX_BUG_ON_PARANOID(!tx_queue
->tso_headers_free
);
976 tsoh
= tx_queue
->tso_headers_free
;
977 tx_queue
->tso_headers_free
= tsoh
->next
;
980 tx_queue
->tso_long_headers
++;
981 tsoh
= efx_tsoh_heap_alloc(tx_queue
, st
->p
.header_length
);
986 header
= TSOH_BUFFER(tsoh
);
987 tsoh_th
= (struct tcphdr
*)(header
+ SKB_TCP_OFF(skb
));
988 tsoh_iph
= (struct iphdr
*)(header
+ SKB_IPV4_OFF(skb
));
990 /* Copy and update the headers. */
991 memcpy(header
, skb
->data
, st
->p
.header_length
);
993 tsoh_th
->seq
= htonl(st
->seqnum
);
994 st
->seqnum
+= skb_shinfo(skb
)->gso_size
;
995 if (st
->remaining_len
> skb_shinfo(skb
)->gso_size
) {
996 /* This packet will not finish the TSO burst. */
997 ip_length
= st
->p
.full_packet_size
- ETH_HDR_LEN(skb
);
1001 /* This packet will be the last in the TSO burst. */
1002 ip_length
= (st
->p
.header_length
- ETH_HDR_LEN(skb
)
1003 + st
->remaining_len
);
1004 tsoh_th
->fin
= tcp_hdr(skb
)->fin
;
1005 tsoh_th
->psh
= tcp_hdr(skb
)->psh
;
1007 tsoh_iph
->tot_len
= htons(ip_length
);
1009 /* Linux leaves suitable gaps in the IP ID space for us to fill. */
1010 tsoh_iph
->id
= htons(st
->p
.ipv4_id
);
1013 st
->packet_space
= skb_shinfo(skb
)->gso_size
;
1014 ++tx_queue
->tso_packets
;
1016 /* Form a descriptor for this header. */
1017 efx_tso_put_header(tx_queue
, tsoh
, st
->p
.header_length
);
1024 * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1025 * @tx_queue: Efx TX queue
1026 * @skb: Socket buffer
1028 * Context: You must hold netif_tx_lock() to call this function.
1030 * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1031 * @skb was not enqueued. In all cases @skb is consumed. Return
1032 * %NETDEV_TX_OK or %NETDEV_TX_BUSY.
1034 static int efx_enqueue_skb_tso(struct efx_tx_queue
*tx_queue
,
1035 const struct sk_buff
*skb
)
1037 struct efx_nic
*efx
= tx_queue
->efx
;
1038 int frag_i
, rc
, rc2
= NETDEV_TX_OK
;
1039 struct tso_state state
;
1041 /* Verify TSO is safe - these checks should never fail. */
1042 efx_tso_check_safe(skb
);
1044 EFX_BUG_ON_PARANOID(tx_queue
->write_count
!= tx_queue
->insert_count
);
1046 tso_start(&state
, skb
);
1048 /* Assume that skb header area contains exactly the headers, and
1049 * all payload is in the frag list.
1051 if (skb_headlen(skb
) == state
.p
.header_length
) {
1052 /* Grab the first payload fragment. */
1053 EFX_BUG_ON_PARANOID(skb_shinfo(skb
)->nr_frags
< 1);
1055 rc
= tso_get_fragment(&state
, efx
,
1056 skb_shinfo(skb
)->frags
+ frag_i
);
1060 rc
= tso_get_head_fragment(&state
, efx
, skb
);
1066 if (tso_start_new_packet(tx_queue
, skb
, &state
) < 0)
1070 rc
= tso_fill_packet_with_fragment(tx_queue
, skb
, &state
);
1074 /* Move onto the next fragment? */
1075 if (state
.ifc
.len
== 0) {
1076 if (++frag_i
>= skb_shinfo(skb
)->nr_frags
)
1077 /* End of payload reached. */
1079 rc
= tso_get_fragment(&state
, efx
,
1080 skb_shinfo(skb
)->frags
+ frag_i
);
1085 /* Start at new packet? */
1086 if (state
.packet_space
== 0 &&
1087 tso_start_new_packet(tx_queue
, skb
, &state
) < 0)
1091 /* Pass off to hardware */
1092 falcon_push_buffers(tx_queue
);
1094 tx_queue
->tso_bursts
++;
1095 return NETDEV_TX_OK
;
1098 EFX_ERR(efx
, "Out of memory for TSO headers, or PCI mapping error\n");
1099 dev_kfree_skb_any((struct sk_buff
*)skb
);
1103 rc2
= NETDEV_TX_BUSY
;
1105 /* Stop the queue if it wasn't stopped before. */
1106 if (tx_queue
->stopped
== 1)
1107 efx_stop_queue(efx
);
1110 /* Free the DMA mapping we were in the process of writing out */
1111 if (state
.ifc
.unmap_len
) {
1112 if (state
.ifc
.unmap_single
)
1113 pci_unmap_single(efx
->pci_dev
, state
.ifc
.unmap_addr
,
1114 state
.ifc
.unmap_len
, PCI_DMA_TODEVICE
);
1116 pci_unmap_page(efx
->pci_dev
, state
.ifc
.unmap_addr
,
1117 state
.ifc
.unmap_len
, PCI_DMA_TODEVICE
);
1120 efx_enqueue_unwind(tx_queue
);
1126 * Free up all TSO datastructures associated with tx_queue. This
1127 * routine should be called only once the tx_queue is both empty and
1128 * will no longer be used.
1130 static void efx_fini_tso(struct efx_tx_queue
*tx_queue
)
1134 if (tx_queue
->buffer
) {
1135 for (i
= 0; i
<= tx_queue
->efx
->type
->txd_ring_mask
; ++i
)
1136 efx_tsoh_free(tx_queue
, &tx_queue
->buffer
[i
]);
1139 while (tx_queue
->tso_headers_free
!= NULL
)
1140 efx_tsoh_block_free(tx_queue
, tx_queue
->tso_headers_free
,
1141 tx_queue
->efx
->pci_dev
);