Merge commit 'v2.6.26-rc8' into x86/mce
[linux-2.6] / drivers / net / sfc / tx.c
1 /****************************************************************************
2  * Driver for Solarflare Solarstorm network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2008 Solarflare Communications Inc.
5  *
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.
9  */
10
11 #include <linux/pci.h>
12 #include <linux/tcp.h>
13 #include <linux/ip.h>
14 #include <linux/in.h>
15 #include <linux/if_ether.h>
16 #include <linux/highmem.h>
17 #include "net_driver.h"
18 #include "tx.h"
19 #include "efx.h"
20 #include "falcon.h"
21 #include "workarounds.h"
22
23 /*
24  * TX descriptor ring full threshold
25  *
26  * The tx_queue descriptor ring fill-level must fall below this value
27  * before we restart the netif queue
28  */
29 #define EFX_NETDEV_TX_THRESHOLD(_tx_queue)      \
30         (_tx_queue->efx->type->txd_ring_mask / 2u)
31
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.
34  */
35 void efx_stop_queue(struct efx_nic *efx)
36 {
37         spin_lock_bh(&efx->netif_stop_lock);
38         EFX_TRACE(efx, "stop TX queue\n");
39
40         atomic_inc(&efx->netif_stop_count);
41         netif_stop_queue(efx->net_dev);
42
43         spin_unlock_bh(&efx->netif_stop_lock);
44 }
45
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.
49  */
50 inline void efx_wake_queue(struct efx_nic *efx)
51 {
52         local_bh_disable();
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);
58         }
59         local_bh_enable();
60 }
61
62 static inline void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
63                                       struct efx_tx_buffer *buffer)
64 {
65         if (buffer->unmap_len) {
66                 struct pci_dev *pci_dev = tx_queue->efx->pci_dev;
67                 if (buffer->unmap_single)
68                         pci_unmap_single(pci_dev, buffer->unmap_addr,
69                                          buffer->unmap_len, PCI_DMA_TODEVICE);
70                 else
71                         pci_unmap_page(pci_dev, buffer->unmap_addr,
72                                        buffer->unmap_len, PCI_DMA_TODEVICE);
73                 buffer->unmap_len = 0;
74                 buffer->unmap_single = 0;
75         }
76
77         if (buffer->skb) {
78                 dev_kfree_skb_any((struct sk_buff *) buffer->skb);
79                 buffer->skb = NULL;
80                 EFX_TRACE(tx_queue->efx, "TX queue %d transmission id %x "
81                           "complete\n", tx_queue->queue, read_ptr);
82         }
83 }
84
85 /**
86  * struct efx_tso_header - a DMA mapped buffer for packet headers
87  * @next: Linked list of free ones.
88  *      The list is protected by the TX queue lock.
89  * @dma_unmap_len: Length to unmap for an oversize buffer, or 0.
90  * @dma_addr: The DMA address of the header below.
91  *
92  * This controls the memory used for a TSO header.  Use TSOH_DATA()
93  * to find the packet header data.  Use TSOH_SIZE() to calculate the
94  * total size required for a given packet header length.  TSO headers
95  * in the free list are exactly %TSOH_STD_SIZE bytes in size.
96  */
97 struct efx_tso_header {
98         union {
99                 struct efx_tso_header *next;
100                 size_t unmap_len;
101         };
102         dma_addr_t dma_addr;
103 };
104
105 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
106                                const struct sk_buff *skb);
107 static void efx_fini_tso(struct efx_tx_queue *tx_queue);
108 static void efx_tsoh_heap_free(struct efx_tx_queue *tx_queue,
109                                struct efx_tso_header *tsoh);
110
111 static inline void efx_tsoh_free(struct efx_tx_queue *tx_queue,
112                                  struct efx_tx_buffer *buffer)
113 {
114         if (buffer->tsoh) {
115                 if (likely(!buffer->tsoh->unmap_len)) {
116                         buffer->tsoh->next = tx_queue->tso_headers_free;
117                         tx_queue->tso_headers_free = buffer->tsoh;
118                 } else {
119                         efx_tsoh_heap_free(tx_queue, buffer->tsoh);
120                 }
121                 buffer->tsoh = NULL;
122         }
123 }
124
125
126 /*
127  * Add a socket buffer to a TX queue
128  *
129  * This maps all fragments of a socket buffer for DMA and adds them to
130  * the TX queue.  The queue's insert pointer will be incremented by
131  * the number of fragments in the socket buffer.
132  *
133  * If any DMA mapping fails, any mapped fragments will be unmapped,
134  * the queue's insert pointer will be restored to its original value.
135  *
136  * Returns NETDEV_TX_OK or NETDEV_TX_BUSY
137  * You must hold netif_tx_lock() to call this function.
138  */
139 static inline int efx_enqueue_skb(struct efx_tx_queue *tx_queue,
140                                   const struct sk_buff *skb)
141 {
142         struct efx_nic *efx = tx_queue->efx;
143         struct pci_dev *pci_dev = efx->pci_dev;
144         struct efx_tx_buffer *buffer;
145         skb_frag_t *fragment;
146         struct page *page;
147         int page_offset;
148         unsigned int len, unmap_len = 0, fill_level, insert_ptr, misalign;
149         dma_addr_t dma_addr, unmap_addr = 0;
150         unsigned int dma_len;
151         unsigned unmap_single;
152         int q_space, i = 0;
153         int rc = NETDEV_TX_OK;
154
155         EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
156
157         if (skb_shinfo((struct sk_buff *)skb)->gso_size)
158                 return efx_enqueue_skb_tso(tx_queue, skb);
159
160         /* Get size of the initial fragment */
161         len = skb_headlen(skb);
162
163         fill_level = tx_queue->insert_count - tx_queue->old_read_count;
164         q_space = efx->type->txd_ring_mask - 1 - fill_level;
165
166         /* Map for DMA.  Use pci_map_single rather than pci_map_page
167          * since this is more efficient on machines with sparse
168          * memory.
169          */
170         unmap_single = 1;
171         dma_addr = pci_map_single(pci_dev, skb->data, len, PCI_DMA_TODEVICE);
172
173         /* Process all fragments */
174         while (1) {
175                 if (unlikely(pci_dma_mapping_error(dma_addr)))
176                         goto pci_err;
177
178                 /* Store fields for marking in the per-fragment final
179                  * descriptor */
180                 unmap_len = len;
181                 unmap_addr = dma_addr;
182
183                 /* Add to TX queue, splitting across DMA boundaries */
184                 do {
185                         if (unlikely(q_space-- <= 0)) {
186                                 /* It might be that completions have
187                                  * happened since the xmit path last
188                                  * checked.  Update the xmit path's
189                                  * copy of read_count.
190                                  */
191                                 ++tx_queue->stopped;
192                                 /* This memory barrier protects the
193                                  * change of stopped from the access
194                                  * of read_count. */
195                                 smp_mb();
196                                 tx_queue->old_read_count =
197                                         *(volatile unsigned *)
198                                         &tx_queue->read_count;
199                                 fill_level = (tx_queue->insert_count
200                                               - tx_queue->old_read_count);
201                                 q_space = (efx->type->txd_ring_mask - 1 -
202                                            fill_level);
203                                 if (unlikely(q_space-- <= 0))
204                                         goto stop;
205                                 smp_mb();
206                                 --tx_queue->stopped;
207                         }
208
209                         insert_ptr = (tx_queue->insert_count &
210                                       efx->type->txd_ring_mask);
211                         buffer = &tx_queue->buffer[insert_ptr];
212                         efx_tsoh_free(tx_queue, buffer);
213                         EFX_BUG_ON_PARANOID(buffer->tsoh);
214                         EFX_BUG_ON_PARANOID(buffer->skb);
215                         EFX_BUG_ON_PARANOID(buffer->len);
216                         EFX_BUG_ON_PARANOID(buffer->continuation != 1);
217                         EFX_BUG_ON_PARANOID(buffer->unmap_len);
218
219                         dma_len = (((~dma_addr) & efx->type->tx_dma_mask) + 1);
220                         if (likely(dma_len > len))
221                                 dma_len = len;
222
223                         misalign = (unsigned)dma_addr & efx->type->bug5391_mask;
224                         if (misalign && dma_len + misalign > 512)
225                                 dma_len = 512 - misalign;
226
227                         /* Fill out per descriptor fields */
228                         buffer->len = dma_len;
229                         buffer->dma_addr = dma_addr;
230                         len -= dma_len;
231                         dma_addr += dma_len;
232                         ++tx_queue->insert_count;
233                 } while (len);
234
235                 /* Transfer ownership of the unmapping to the final buffer */
236                 buffer->unmap_addr = unmap_addr;
237                 buffer->unmap_single = unmap_single;
238                 buffer->unmap_len = unmap_len;
239                 unmap_len = 0;
240
241                 /* Get address and size of next fragment */
242                 if (i >= skb_shinfo(skb)->nr_frags)
243                         break;
244                 fragment = &skb_shinfo(skb)->frags[i];
245                 len = fragment->size;
246                 page = fragment->page;
247                 page_offset = fragment->page_offset;
248                 i++;
249                 /* Map for DMA */
250                 unmap_single = 0;
251                 dma_addr = pci_map_page(pci_dev, page, page_offset, len,
252                                         PCI_DMA_TODEVICE);
253         }
254
255         /* Transfer ownership of the skb to the final buffer */
256         buffer->skb = skb;
257         buffer->continuation = 0;
258
259         /* Pass off to hardware */
260         falcon_push_buffers(tx_queue);
261
262         return NETDEV_TX_OK;
263
264  pci_err:
265         EFX_ERR_RL(efx, " TX queue %d could not map skb with %d bytes %d "
266                    "fragments for DMA\n", tx_queue->queue, skb->len,
267                    skb_shinfo(skb)->nr_frags + 1);
268
269         /* Mark the packet as transmitted, and free the SKB ourselves */
270         dev_kfree_skb_any((struct sk_buff *)skb);
271         goto unwind;
272
273  stop:
274         rc = NETDEV_TX_BUSY;
275
276         if (tx_queue->stopped == 1)
277                 efx_stop_queue(efx);
278
279  unwind:
280         /* Work backwards until we hit the original insert pointer value */
281         while (tx_queue->insert_count != tx_queue->write_count) {
282                 --tx_queue->insert_count;
283                 insert_ptr = tx_queue->insert_count & efx->type->txd_ring_mask;
284                 buffer = &tx_queue->buffer[insert_ptr];
285                 efx_dequeue_buffer(tx_queue, buffer);
286                 buffer->len = 0;
287         }
288
289         /* Free the fragment we were mid-way through pushing */
290         if (unmap_len)
291                 pci_unmap_page(pci_dev, unmap_addr, unmap_len,
292                                PCI_DMA_TODEVICE);
293
294         return rc;
295 }
296
297 /* Remove packets from the TX queue
298  *
299  * This removes packets from the TX queue, up to and including the
300  * specified index.
301  */
302 static inline void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
303                                        unsigned int index)
304 {
305         struct efx_nic *efx = tx_queue->efx;
306         unsigned int stop_index, read_ptr;
307         unsigned int mask = tx_queue->efx->type->txd_ring_mask;
308
309         stop_index = (index + 1) & mask;
310         read_ptr = tx_queue->read_count & mask;
311
312         while (read_ptr != stop_index) {
313                 struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
314                 if (unlikely(buffer->len == 0)) {
315                         EFX_ERR(tx_queue->efx, "TX queue %d spurious TX "
316                                 "completion id %x\n", tx_queue->queue,
317                                 read_ptr);
318                         efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
319                         return;
320                 }
321
322                 efx_dequeue_buffer(tx_queue, buffer);
323                 buffer->continuation = 1;
324                 buffer->len = 0;
325
326                 ++tx_queue->read_count;
327                 read_ptr = tx_queue->read_count & mask;
328         }
329 }
330
331 /* Initiate a packet transmission on the specified TX queue.
332  * Note that returning anything other than NETDEV_TX_OK will cause the
333  * OS to free the skb.
334  *
335  * This function is split out from efx_hard_start_xmit to allow the
336  * loopback test to direct packets via specific TX queues.  It is
337  * therefore a non-static inline, so as not to penalise performance
338  * for non-loopback transmissions.
339  *
340  * Context: netif_tx_lock held
341  */
342 inline int efx_xmit(struct efx_nic *efx,
343                     struct efx_tx_queue *tx_queue, struct sk_buff *skb)
344 {
345         int rc;
346
347         /* Map fragments for DMA and add to TX queue */
348         rc = efx_enqueue_skb(tx_queue, skb);
349         if (unlikely(rc != NETDEV_TX_OK))
350                 goto out;
351
352         /* Update last TX timer */
353         efx->net_dev->trans_start = jiffies;
354
355  out:
356         return rc;
357 }
358
359 /* Initiate a packet transmission.  We use one channel per CPU
360  * (sharing when we have more CPUs than channels).  On Falcon, the TX
361  * completion events will be directed back to the CPU that transmitted
362  * the packet, which should be cache-efficient.
363  *
364  * Context: non-blocking.
365  * Note that returning anything other than NETDEV_TX_OK will cause the
366  * OS to free the skb.
367  */
368 int efx_hard_start_xmit(struct sk_buff *skb, struct net_device *net_dev)
369 {
370         struct efx_nic *efx = net_dev->priv;
371         return efx_xmit(efx, &efx->tx_queue[0], skb);
372 }
373
374 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
375 {
376         unsigned fill_level;
377         struct efx_nic *efx = tx_queue->efx;
378
379         EFX_BUG_ON_PARANOID(index > efx->type->txd_ring_mask);
380
381         efx_dequeue_buffers(tx_queue, index);
382
383         /* See if we need to restart the netif queue.  This barrier
384          * separates the update of read_count from the test of
385          * stopped. */
386         smp_mb();
387         if (unlikely(tx_queue->stopped)) {
388                 fill_level = tx_queue->insert_count - tx_queue->read_count;
389                 if (fill_level < EFX_NETDEV_TX_THRESHOLD(tx_queue)) {
390                         EFX_BUG_ON_PARANOID(!efx_dev_registered(efx));
391
392                         /* Do this under netif_tx_lock(), to avoid racing
393                          * with efx_xmit(). */
394                         netif_tx_lock(efx->net_dev);
395                         if (tx_queue->stopped) {
396                                 tx_queue->stopped = 0;
397                                 efx_wake_queue(efx);
398                         }
399                         netif_tx_unlock(efx->net_dev);
400                 }
401         }
402 }
403
404 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
405 {
406         struct efx_nic *efx = tx_queue->efx;
407         unsigned int txq_size;
408         int i, rc;
409
410         EFX_LOG(efx, "creating TX queue %d\n", tx_queue->queue);
411
412         /* Allocate software ring */
413         txq_size = (efx->type->txd_ring_mask + 1) * sizeof(*tx_queue->buffer);
414         tx_queue->buffer = kzalloc(txq_size, GFP_KERNEL);
415         if (!tx_queue->buffer) {
416                 rc = -ENOMEM;
417                 goto fail1;
418         }
419         for (i = 0; i <= efx->type->txd_ring_mask; ++i)
420                 tx_queue->buffer[i].continuation = 1;
421
422         /* Allocate hardware ring */
423         rc = falcon_probe_tx(tx_queue);
424         if (rc)
425                 goto fail2;
426
427         return 0;
428
429  fail2:
430         kfree(tx_queue->buffer);
431         tx_queue->buffer = NULL;
432  fail1:
433         tx_queue->used = 0;
434
435         return rc;
436 }
437
438 int efx_init_tx_queue(struct efx_tx_queue *tx_queue)
439 {
440         EFX_LOG(tx_queue->efx, "initialising TX queue %d\n", tx_queue->queue);
441
442         tx_queue->insert_count = 0;
443         tx_queue->write_count = 0;
444         tx_queue->read_count = 0;
445         tx_queue->old_read_count = 0;
446         BUG_ON(tx_queue->stopped);
447
448         /* Set up TX descriptor ring */
449         return falcon_init_tx(tx_queue);
450 }
451
452 void efx_release_tx_buffers(struct efx_tx_queue *tx_queue)
453 {
454         struct efx_tx_buffer *buffer;
455
456         if (!tx_queue->buffer)
457                 return;
458
459         /* Free any buffers left in the ring */
460         while (tx_queue->read_count != tx_queue->write_count) {
461                 buffer = &tx_queue->buffer[tx_queue->read_count &
462                                            tx_queue->efx->type->txd_ring_mask];
463                 efx_dequeue_buffer(tx_queue, buffer);
464                 buffer->continuation = 1;
465                 buffer->len = 0;
466
467                 ++tx_queue->read_count;
468         }
469 }
470
471 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
472 {
473         EFX_LOG(tx_queue->efx, "shutting down TX queue %d\n", tx_queue->queue);
474
475         /* Flush TX queue, remove descriptor ring */
476         falcon_fini_tx(tx_queue);
477
478         efx_release_tx_buffers(tx_queue);
479
480         /* Free up TSO header cache */
481         efx_fini_tso(tx_queue);
482
483         /* Release queue's stop on port, if any */
484         if (tx_queue->stopped) {
485                 tx_queue->stopped = 0;
486                 efx_wake_queue(tx_queue->efx);
487         }
488 }
489
490 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
491 {
492         EFX_LOG(tx_queue->efx, "destroying TX queue %d\n", tx_queue->queue);
493         falcon_remove_tx(tx_queue);
494
495         kfree(tx_queue->buffer);
496         tx_queue->buffer = NULL;
497         tx_queue->used = 0;
498 }
499
500
501 /* Efx TCP segmentation acceleration.
502  *
503  * Why?  Because by doing it here in the driver we can go significantly
504  * faster than the GSO.
505  *
506  * Requires TX checksum offload support.
507  */
508
509 /* Number of bytes inserted at the start of a TSO header buffer,
510  * similar to NET_IP_ALIGN.
511  */
512 #if defined(__i386__) || defined(__x86_64__)
513 #define TSOH_OFFSET     0
514 #else
515 #define TSOH_OFFSET     NET_IP_ALIGN
516 #endif
517
518 #define TSOH_BUFFER(tsoh)       ((u8 *)(tsoh + 1) + TSOH_OFFSET)
519
520 /* Total size of struct efx_tso_header, buffer and padding */
521 #define TSOH_SIZE(hdr_len)                                      \
522         (sizeof(struct efx_tso_header) + TSOH_OFFSET + hdr_len)
523
524 /* Size of blocks on free list.  Larger blocks must be allocated from
525  * the heap.
526  */
527 #define TSOH_STD_SIZE           128
528
529 #define PTR_DIFF(p1, p2)  ((u8 *)(p1) - (u8 *)(p2))
530 #define ETH_HDR_LEN(skb)  (skb_network_header(skb) - (skb)->data)
531 #define SKB_TCP_OFF(skb)  PTR_DIFF(tcp_hdr(skb), (skb)->data)
532 #define SKB_IPV4_OFF(skb) PTR_DIFF(ip_hdr(skb), (skb)->data)
533
534 /**
535  * struct tso_state - TSO state for an SKB
536  * @remaining_len: Bytes of data we've yet to segment
537  * @seqnum: Current sequence number
538  * @packet_space: Remaining space in current packet
539  * @ifc: Input fragment cursor.
540  *      Where we are in the current fragment of the incoming SKB.  These
541  *      values get updated in place when we split a fragment over
542  *      multiple packets.
543  * @p: Parameters.
544  *      These values are set once at the start of the TSO send and do
545  *      not get changed as the routine progresses.
546  *
547  * The state used during segmentation.  It is put into this data structure
548  * just to make it easy to pass into inline functions.
549  */
550 struct tso_state {
551         unsigned remaining_len;
552         unsigned seqnum;
553         unsigned packet_space;
554
555         struct {
556                 /* DMA address of current position */
557                 dma_addr_t dma_addr;
558                 /* Remaining length */
559                 unsigned int len;
560                 /* DMA address and length of the whole fragment */
561                 unsigned int unmap_len;
562                 dma_addr_t unmap_addr;
563                 struct page *page;
564                 unsigned page_off;
565         } ifc;
566
567         struct {
568                 /* The number of bytes of header */
569                 unsigned int header_length;
570
571                 /* The number of bytes to put in each outgoing segment. */
572                 int full_packet_size;
573
574                 /* Current IPv4 ID, host endian. */
575                 unsigned ipv4_id;
576         } p;
577 };
578
579
580 /*
581  * Verify that our various assumptions about sk_buffs and the conditions
582  * under which TSO will be attempted hold true.
583  */
584 static inline void efx_tso_check_safe(const struct sk_buff *skb)
585 {
586         EFX_BUG_ON_PARANOID(skb->protocol != htons(ETH_P_IP));
587         EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
588                             skb->protocol);
589         EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
590         EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
591                              + (tcp_hdr(skb)->doff << 2u)) >
592                             skb_headlen(skb));
593 }
594
595
596 /*
597  * Allocate a page worth of efx_tso_header structures, and string them
598  * into the tx_queue->tso_headers_free linked list. Return 0 or -ENOMEM.
599  */
600 static int efx_tsoh_block_alloc(struct efx_tx_queue *tx_queue)
601 {
602
603         struct pci_dev *pci_dev = tx_queue->efx->pci_dev;
604         struct efx_tso_header *tsoh;
605         dma_addr_t dma_addr;
606         u8 *base_kva, *kva;
607
608         base_kva = pci_alloc_consistent(pci_dev, PAGE_SIZE, &dma_addr);
609         if (base_kva == NULL) {
610                 EFX_ERR(tx_queue->efx, "Unable to allocate page for TSO"
611                         " headers\n");
612                 return -ENOMEM;
613         }
614
615         /* pci_alloc_consistent() allocates pages. */
616         EFX_BUG_ON_PARANOID(dma_addr & (PAGE_SIZE - 1u));
617
618         for (kva = base_kva; kva < base_kva + PAGE_SIZE; kva += TSOH_STD_SIZE) {
619                 tsoh = (struct efx_tso_header *)kva;
620                 tsoh->dma_addr = dma_addr + (TSOH_BUFFER(tsoh) - base_kva);
621                 tsoh->next = tx_queue->tso_headers_free;
622                 tx_queue->tso_headers_free = tsoh;
623         }
624
625         return 0;
626 }
627
628
629 /* Free up a TSO header, and all others in the same page. */
630 static void efx_tsoh_block_free(struct efx_tx_queue *tx_queue,
631                                 struct efx_tso_header *tsoh,
632                                 struct pci_dev *pci_dev)
633 {
634         struct efx_tso_header **p;
635         unsigned long base_kva;
636         dma_addr_t base_dma;
637
638         base_kva = (unsigned long)tsoh & PAGE_MASK;
639         base_dma = tsoh->dma_addr & PAGE_MASK;
640
641         p = &tx_queue->tso_headers_free;
642         while (*p != NULL) {
643                 if (((unsigned long)*p & PAGE_MASK) == base_kva)
644                         *p = (*p)->next;
645                 else
646                         p = &(*p)->next;
647         }
648
649         pci_free_consistent(pci_dev, PAGE_SIZE, (void *)base_kva, base_dma);
650 }
651
652 static struct efx_tso_header *
653 efx_tsoh_heap_alloc(struct efx_tx_queue *tx_queue, size_t header_len)
654 {
655         struct efx_tso_header *tsoh;
656
657         tsoh = kmalloc(TSOH_SIZE(header_len), GFP_ATOMIC | GFP_DMA);
658         if (unlikely(!tsoh))
659                 return NULL;
660
661         tsoh->dma_addr = pci_map_single(tx_queue->efx->pci_dev,
662                                         TSOH_BUFFER(tsoh), header_len,
663                                         PCI_DMA_TODEVICE);
664         if (unlikely(pci_dma_mapping_error(tsoh->dma_addr))) {
665                 kfree(tsoh);
666                 return NULL;
667         }
668
669         tsoh->unmap_len = header_len;
670         return tsoh;
671 }
672
673 static void
674 efx_tsoh_heap_free(struct efx_tx_queue *tx_queue, struct efx_tso_header *tsoh)
675 {
676         pci_unmap_single(tx_queue->efx->pci_dev,
677                          tsoh->dma_addr, tsoh->unmap_len,
678                          PCI_DMA_TODEVICE);
679         kfree(tsoh);
680 }
681
682 /**
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  * @skb:                Only non-null for end of last segment
688  * @end_of_packet:      True if last fragment in a packet
689  * @unmap_addr:         DMA address of fragment for unmapping
690  * @unmap_len:          Only set this in last segment of a fragment
691  *
692  * Push descriptors onto the TX queue.  Return 0 on success or 1 if
693  * @tx_queue full.
694  */
695 static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
696                                dma_addr_t dma_addr, unsigned len,
697                                const struct sk_buff *skb, int end_of_packet,
698                                dma_addr_t unmap_addr, unsigned unmap_len)
699 {
700         struct efx_tx_buffer *buffer;
701         struct efx_nic *efx = tx_queue->efx;
702         unsigned dma_len, fill_level, insert_ptr, misalign;
703         int q_space;
704
705         EFX_BUG_ON_PARANOID(len <= 0);
706
707         fill_level = tx_queue->insert_count - tx_queue->old_read_count;
708         /* -1 as there is no way to represent all descriptors used */
709         q_space = efx->type->txd_ring_mask - 1 - fill_level;
710
711         while (1) {
712                 if (unlikely(q_space-- <= 0)) {
713                         /* It might be that completions have happened
714                          * since the xmit path last checked.  Update
715                          * the xmit path's copy of read_count.
716                          */
717                         ++tx_queue->stopped;
718                         /* This memory barrier protects the change of
719                          * stopped from the access of read_count. */
720                         smp_mb();
721                         tx_queue->old_read_count =
722                                 *(volatile unsigned *)&tx_queue->read_count;
723                         fill_level = (tx_queue->insert_count
724                                       - tx_queue->old_read_count);
725                         q_space = efx->type->txd_ring_mask - 1 - fill_level;
726                         if (unlikely(q_space-- <= 0))
727                                 return 1;
728                         smp_mb();
729                         --tx_queue->stopped;
730                 }
731
732                 insert_ptr = tx_queue->insert_count & efx->type->txd_ring_mask;
733                 buffer = &tx_queue->buffer[insert_ptr];
734                 ++tx_queue->insert_count;
735
736                 EFX_BUG_ON_PARANOID(tx_queue->insert_count -
737                                     tx_queue->read_count >
738                                     efx->type->txd_ring_mask);
739
740                 efx_tsoh_free(tx_queue, buffer);
741                 EFX_BUG_ON_PARANOID(buffer->len);
742                 EFX_BUG_ON_PARANOID(buffer->unmap_len);
743                 EFX_BUG_ON_PARANOID(buffer->skb);
744                 EFX_BUG_ON_PARANOID(buffer->continuation != 1);
745                 EFX_BUG_ON_PARANOID(buffer->tsoh);
746
747                 buffer->dma_addr = dma_addr;
748
749                 /* Ensure we do not cross a boundary unsupported by H/W */
750                 dma_len = (~dma_addr & efx->type->tx_dma_mask) + 1;
751
752                 misalign = (unsigned)dma_addr & efx->type->bug5391_mask;
753                 if (misalign && dma_len + misalign > 512)
754                         dma_len = 512 - misalign;
755
756                 /* If there is enough space to send then do so */
757                 if (dma_len >= len)
758                         break;
759
760                 buffer->len = dma_len; /* Don't set the other members */
761                 dma_addr += dma_len;
762                 len -= dma_len;
763         }
764
765         EFX_BUG_ON_PARANOID(!len);
766         buffer->len = len;
767         buffer->skb = skb;
768         buffer->continuation = !end_of_packet;
769         buffer->unmap_addr = unmap_addr;
770         buffer->unmap_len = unmap_len;
771         return 0;
772 }
773
774
775 /*
776  * Put a TSO header into the TX queue.
777  *
778  * This is special-cased because we know that it is small enough to fit in
779  * a single fragment, and we know it doesn't cross a page boundary.  It
780  * also allows us to not worry about end-of-packet etc.
781  */
782 static inline void efx_tso_put_header(struct efx_tx_queue *tx_queue,
783                                       struct efx_tso_header *tsoh, unsigned len)
784 {
785         struct efx_tx_buffer *buffer;
786
787         buffer = &tx_queue->buffer[tx_queue->insert_count &
788                                    tx_queue->efx->type->txd_ring_mask];
789         efx_tsoh_free(tx_queue, buffer);
790         EFX_BUG_ON_PARANOID(buffer->len);
791         EFX_BUG_ON_PARANOID(buffer->unmap_len);
792         EFX_BUG_ON_PARANOID(buffer->skb);
793         EFX_BUG_ON_PARANOID(buffer->continuation != 1);
794         EFX_BUG_ON_PARANOID(buffer->tsoh);
795         buffer->len = len;
796         buffer->dma_addr = tsoh->dma_addr;
797         buffer->tsoh = tsoh;
798
799         ++tx_queue->insert_count;
800 }
801
802
803 /* Remove descriptors put into a tx_queue. */
804 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
805 {
806         struct efx_tx_buffer *buffer;
807
808         /* Work backwards until we hit the original insert pointer value */
809         while (tx_queue->insert_count != tx_queue->write_count) {
810                 --tx_queue->insert_count;
811                 buffer = &tx_queue->buffer[tx_queue->insert_count &
812                                            tx_queue->efx->type->txd_ring_mask];
813                 efx_tsoh_free(tx_queue, buffer);
814                 EFX_BUG_ON_PARANOID(buffer->skb);
815                 buffer->len = 0;
816                 buffer->continuation = 1;
817                 if (buffer->unmap_len) {
818                         pci_unmap_page(tx_queue->efx->pci_dev,
819                                        buffer->unmap_addr,
820                                        buffer->unmap_len, PCI_DMA_TODEVICE);
821                         buffer->unmap_len = 0;
822                 }
823         }
824 }
825
826
827 /* Parse the SKB header and initialise state. */
828 static inline void tso_start(struct tso_state *st, const struct sk_buff *skb)
829 {
830         /* All ethernet/IP/TCP headers combined size is TCP header size
831          * plus offset of TCP header relative to start of packet.
832          */
833         st->p.header_length = ((tcp_hdr(skb)->doff << 2u)
834                                + PTR_DIFF(tcp_hdr(skb), skb->data));
835         st->p.full_packet_size = (st->p.header_length
836                                   + skb_shinfo(skb)->gso_size);
837
838         st->p.ipv4_id = ntohs(ip_hdr(skb)->id);
839         st->seqnum = ntohl(tcp_hdr(skb)->seq);
840
841         EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
842         EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
843         EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
844
845         st->packet_space = st->p.full_packet_size;
846         st->remaining_len = skb->len - st->p.header_length;
847 }
848
849
850 /**
851  * tso_get_fragment - record fragment details and map for DMA
852  * @st:                 TSO state
853  * @efx:                Efx NIC
854  * @data:               Pointer to fragment data
855  * @len:                Length of fragment
856  *
857  * Record fragment details and map for DMA.  Return 0 on success, or
858  * -%ENOMEM if DMA mapping fails.
859  */
860 static inline int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
861                                    int len, struct page *page, int page_off)
862 {
863
864         st->ifc.unmap_addr = pci_map_page(efx->pci_dev, page, page_off,
865                                           len, PCI_DMA_TODEVICE);
866         if (likely(!pci_dma_mapping_error(st->ifc.unmap_addr))) {
867                 st->ifc.unmap_len = len;
868                 st->ifc.len = len;
869                 st->ifc.dma_addr = st->ifc.unmap_addr;
870                 st->ifc.page = page;
871                 st->ifc.page_off = page_off;
872                 return 0;
873         }
874         return -ENOMEM;
875 }
876
877
878 /**
879  * tso_fill_packet_with_fragment - form descriptors for the current fragment
880  * @tx_queue:           Efx TX queue
881  * @skb:                Socket buffer
882  * @st:                 TSO state
883  *
884  * Form descriptors for the current fragment, until we reach the end
885  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
886  * space in @tx_queue.
887  */
888 static inline int tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
889                                                 const struct sk_buff *skb,
890                                                 struct tso_state *st)
891 {
892
893         int n, end_of_packet, rc;
894
895         if (st->ifc.len == 0)
896                 return 0;
897         if (st->packet_space == 0)
898                 return 0;
899
900         EFX_BUG_ON_PARANOID(st->ifc.len <= 0);
901         EFX_BUG_ON_PARANOID(st->packet_space <= 0);
902
903         n = min(st->ifc.len, st->packet_space);
904
905         st->packet_space -= n;
906         st->remaining_len -= n;
907         st->ifc.len -= n;
908         st->ifc.page_off += n;
909         end_of_packet = st->remaining_len == 0 || st->packet_space == 0;
910
911         rc = efx_tx_queue_insert(tx_queue, st->ifc.dma_addr, n,
912                                  st->remaining_len ? NULL : skb,
913                                  end_of_packet, st->ifc.unmap_addr,
914                                  st->ifc.len ? 0 : st->ifc.unmap_len);
915
916         st->ifc.dma_addr += n;
917
918         return rc;
919 }
920
921
922 /**
923  * tso_start_new_packet - generate a new header and prepare for the new packet
924  * @tx_queue:           Efx TX queue
925  * @skb:                Socket buffer
926  * @st:                 TSO state
927  *
928  * Generate a new header and prepare for the new packet.  Return 0 on
929  * success, or -1 if failed to alloc header.
930  */
931 static inline int tso_start_new_packet(struct efx_tx_queue *tx_queue,
932                                        const struct sk_buff *skb,
933                                        struct tso_state *st)
934 {
935         struct efx_tso_header *tsoh;
936         struct iphdr *tsoh_iph;
937         struct tcphdr *tsoh_th;
938         unsigned ip_length;
939         u8 *header;
940
941         /* Allocate a DMA-mapped header buffer. */
942         if (likely(TSOH_SIZE(st->p.header_length) <= TSOH_STD_SIZE)) {
943                 if (tx_queue->tso_headers_free == NULL) {
944                         if (efx_tsoh_block_alloc(tx_queue))
945                                 return -1;
946                 }
947                 EFX_BUG_ON_PARANOID(!tx_queue->tso_headers_free);
948                 tsoh = tx_queue->tso_headers_free;
949                 tx_queue->tso_headers_free = tsoh->next;
950                 tsoh->unmap_len = 0;
951         } else {
952                 tx_queue->tso_long_headers++;
953                 tsoh = efx_tsoh_heap_alloc(tx_queue, st->p.header_length);
954                 if (unlikely(!tsoh))
955                         return -1;
956         }
957
958         header = TSOH_BUFFER(tsoh);
959         tsoh_th = (struct tcphdr *)(header + SKB_TCP_OFF(skb));
960         tsoh_iph = (struct iphdr *)(header + SKB_IPV4_OFF(skb));
961
962         /* Copy and update the headers. */
963         memcpy(header, skb->data, st->p.header_length);
964
965         tsoh_th->seq = htonl(st->seqnum);
966         st->seqnum += skb_shinfo(skb)->gso_size;
967         if (st->remaining_len > skb_shinfo(skb)->gso_size) {
968                 /* This packet will not finish the TSO burst. */
969                 ip_length = st->p.full_packet_size - ETH_HDR_LEN(skb);
970                 tsoh_th->fin = 0;
971                 tsoh_th->psh = 0;
972         } else {
973                 /* This packet will be the last in the TSO burst. */
974                 ip_length = (st->p.header_length - ETH_HDR_LEN(skb)
975                              + st->remaining_len);
976                 tsoh_th->fin = tcp_hdr(skb)->fin;
977                 tsoh_th->psh = tcp_hdr(skb)->psh;
978         }
979         tsoh_iph->tot_len = htons(ip_length);
980
981         /* Linux leaves suitable gaps in the IP ID space for us to fill. */
982         tsoh_iph->id = htons(st->p.ipv4_id);
983         st->p.ipv4_id++;
984
985         st->packet_space = skb_shinfo(skb)->gso_size;
986         ++tx_queue->tso_packets;
987
988         /* Form a descriptor for this header. */
989         efx_tso_put_header(tx_queue, tsoh, st->p.header_length);
990
991         return 0;
992 }
993
994
995 /**
996  * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
997  * @tx_queue:           Efx TX queue
998  * @skb:                Socket buffer
999  *
1000  * Context: You must hold netif_tx_lock() to call this function.
1001  *
1002  * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1003  * @skb was not enqueued.  In all cases @skb is consumed.  Return
1004  * %NETDEV_TX_OK or %NETDEV_TX_BUSY.
1005  */
1006 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1007                                const struct sk_buff *skb)
1008 {
1009         int frag_i, rc, rc2 = NETDEV_TX_OK;
1010         struct tso_state state;
1011         skb_frag_t *f;
1012
1013         /* Verify TSO is safe - these checks should never fail. */
1014         efx_tso_check_safe(skb);
1015
1016         EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
1017
1018         tso_start(&state, skb);
1019
1020         /* Assume that skb header area contains exactly the headers, and
1021          * all payload is in the frag list.
1022          */
1023         if (skb_headlen(skb) == state.p.header_length) {
1024                 /* Grab the first payload fragment. */
1025                 EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1026                 frag_i = 0;
1027                 f = &skb_shinfo(skb)->frags[frag_i];
1028                 rc = tso_get_fragment(&state, tx_queue->efx,
1029                                       f->size, f->page, f->page_offset);
1030                 if (rc)
1031                         goto mem_err;
1032         } else {
1033                 /* It may look like this code fragment assumes that the
1034                  * skb->data portion does not cross a page boundary, but
1035                  * that is not the case.  It is guaranteed to be direct
1036                  * mapped memory, and therefore is physically contiguous,
1037                  * and so DMA will work fine.  kmap_atomic() on this region
1038                  * will just return the direct mapping, so that will work
1039                  * too.
1040                  */
1041                 int page_off = (unsigned long)skb->data & (PAGE_SIZE - 1);
1042                 int hl = state.p.header_length;
1043                 rc = tso_get_fragment(&state, tx_queue->efx,
1044                                       skb_headlen(skb) - hl,
1045                                       virt_to_page(skb->data), page_off + hl);
1046                 if (rc)
1047                         goto mem_err;
1048                 frag_i = -1;
1049         }
1050
1051         if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1052                 goto mem_err;
1053
1054         while (1) {
1055                 rc = tso_fill_packet_with_fragment(tx_queue, skb, &state);
1056                 if (unlikely(rc))
1057                         goto stop;
1058
1059                 /* Move onto the next fragment? */
1060                 if (state.ifc.len == 0) {
1061                         if (++frag_i >= skb_shinfo(skb)->nr_frags)
1062                                 /* End of payload reached. */
1063                                 break;
1064                         f = &skb_shinfo(skb)->frags[frag_i];
1065                         rc = tso_get_fragment(&state, tx_queue->efx,
1066                                               f->size, f->page, f->page_offset);
1067                         if (rc)
1068                                 goto mem_err;
1069                 }
1070
1071                 /* Start at new packet? */
1072                 if (state.packet_space == 0 &&
1073                     tso_start_new_packet(tx_queue, skb, &state) < 0)
1074                         goto mem_err;
1075         }
1076
1077         /* Pass off to hardware */
1078         falcon_push_buffers(tx_queue);
1079
1080         tx_queue->tso_bursts++;
1081         return NETDEV_TX_OK;
1082
1083  mem_err:
1084         EFX_ERR(tx_queue->efx, "Out of memory for TSO headers, or PCI mapping"
1085                 " error\n");
1086         dev_kfree_skb_any((struct sk_buff *)skb);
1087         goto unwind;
1088
1089  stop:
1090         rc2 = NETDEV_TX_BUSY;
1091
1092         /* Stop the queue if it wasn't stopped before. */
1093         if (tx_queue->stopped == 1)
1094                 efx_stop_queue(tx_queue->efx);
1095
1096  unwind:
1097         efx_enqueue_unwind(tx_queue);
1098         return rc2;
1099 }
1100
1101
1102 /*
1103  * Free up all TSO datastructures associated with tx_queue. This
1104  * routine should be called only once the tx_queue is both empty and
1105  * will no longer be used.
1106  */
1107 static void efx_fini_tso(struct efx_tx_queue *tx_queue)
1108 {
1109         unsigned i;
1110
1111         if (tx_queue->buffer) {
1112                 for (i = 0; i <= tx_queue->efx->type->txd_ring_mask; ++i)
1113                         efx_tsoh_free(tx_queue, &tx_queue->buffer[i]);
1114         }
1115
1116         while (tx_queue->tso_headers_free != NULL)
1117                 efx_tsoh_block_free(tx_queue, tx_queue->tso_headers_free,
1118                                     tx_queue->efx->pci_dev);
1119 }