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                 dma_addr_t unmap_addr = (buffer->dma_addr + buffer->len -
  68                                          buffer->unmap_len);
  69                 if (buffer->unmap_single)
  70                         pci_unmap_single(pci_dev, unmap_addr, buffer->unmap_len,
  71                                          PCI_DMA_TODEVICE);
  72                 else
  73                         pci_unmap_page(pci_dev, unmap_addr, buffer->unmap_len,
  74                                        PCI_DMA_TODEVICE);
  75                 buffer->unmap_len = 0;
  76                 buffer->unmap_single = false;
  77         }
  78
  79         if (buffer->skb) {
  80                 dev_kfree_skb_any((struct sk_buff *) buffer->skb);
  81                 buffer->skb = NULL;
  82                 EFX_TRACE(tx_queue->efx, "TX queue %d transmission id %x "
  83                           "complete\n", tx_queue->queue, read_ptr);
  84         }
  85 }
  86
  87 /**
  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.
  93  *
  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.
  98  */
  99 struct efx_tso_header {
 100         union {
 101                 struct efx_tso_header *next;
 102                 size_t unmap_len;
 103         };
 104         dma_addr_t dma_addr;
 105 };
 106
 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);
 112
 113 static inline void efx_tsoh_free(struct efx_tx_queue *tx_queue,
 114                                  struct efx_tx_buffer *buffer)
 115 {
 116         if (buffer->tsoh) {
 117                 if (likely(!buffer->tsoh->unmap_len)) {
 118                         buffer->tsoh->next = tx_queue->tso_headers_free;
 119                         tx_queue->tso_headers_free = buffer->tsoh;
 120                 } else {
 121                         efx_tsoh_heap_free(tx_queue, buffer->tsoh);
 122                 }
 123                 buffer->tsoh = NULL;
 124         }
 125 }
 126
 127
 128 /*
 129  * Add a socket buffer to a TX queue
 130  *
 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.
 134  *
 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.
 137  *
 138  * Returns NETDEV_TX_OK or NETDEV_TX_BUSY
 139  * You must hold netif_tx_lock() to call this function.
 140  */
 141 static inline int efx_enqueue_skb(struct efx_tx_queue *tx_queue,
 142                                   const struct sk_buff *skb)
 143 {
 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;
 148         struct page *page;
 149         int page_offset;
 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;
 153         bool unmap_single;
 154         int q_space, i = 0;
 155         int rc = NETDEV_TX_OK;
 156
 157         EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
 158
 159         if (skb_shinfo((struct sk_buff *)skb)->gso_size)
 160                 return efx_enqueue_skb_tso(tx_queue, skb);
 161
 162         /* Get size of the initial fragment */
 163         len = skb_headlen(skb);
 164
 165         fill_level = tx_queue->insert_count - tx_queue->old_read_count;
 166         q_space = efx->type->txd_ring_mask - 1 - fill_level;
 167
 168         /* Map for DMA.  Use pci_map_single rather than pci_map_page
 169          * since this is more efficient on machines with sparse
 170          * memory.
 171          */
 172         unmap_single = true;
 173         dma_addr = pci_map_single(pci_dev, skb->data, len, PCI_DMA_TODEVICE);
 174
 175         /* Process all fragments */
 176         while (1) {
 177                 if (unlikely(pci_dma_mapping_error(pci_dev, dma_addr)))
 178                         goto pci_err;
 179
 180                 /* Store fields for marking in the per-fragment final
 181                  * descriptor */
 182                 unmap_len = len;
 183                 unmap_addr = dma_addr;
 184
 185                 /* Add to TX queue, splitting across DMA boundaries */
 186                 do {
 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.
 192                                  */
 193                                 ++tx_queue->stopped;
 194                                 /* This memory barrier protects the
 195                                  * change of stopped from the access
 196                                  * of read_count. */
 197                                 smp_mb();
 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 -
 204                                            fill_level);
 205                                 if (unlikely(q_space-- <= 0))
 206                                         goto stop;
 207                                 smp_mb();
 208                                 --tx_queue->stopped;
 209                         }
 210
 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);
 220
 221                         dma_len = (((~dma_addr) & efx->type->tx_dma_mask) + 1);
 222                         if (likely(dma_len > len))
 223                                 dma_len = len;
 224
 225                         misalign = (unsigned)dma_addr & efx->type->bug5391_mask;
 226                         if (misalign && dma_len + misalign > 512)
 227                                 dma_len = 512 - misalign;
 228
 229                         /* Fill out per descriptor fields */
 230                         buffer->len = dma_len;
 231                         buffer->dma_addr = dma_addr;
 232                         len -= dma_len;
 233                         dma_addr += dma_len;
 234                         ++tx_queue->insert_count;
 235                 } while (len);
 236
 237                 /* Transfer ownership of the unmapping to the final buffer */
 238                 buffer->unmap_single = unmap_single;
 239                 buffer->unmap_len = unmap_len;
 240                 unmap_len = 0;
 241
 242                 /* Get address and size of next fragment */
 243                 if (i >= skb_shinfo(skb)->nr_frags)
 244                         break;
 245                 fragment = &skb_shinfo(skb)->frags[i];
 246                 len = fragment->size;
 247                 page = fragment->page;
 248                 page_offset = fragment->page_offset;
 249                 i++;
 250                 /* Map for DMA */
 251                 unmap_single = false;
 252                 dma_addr = pci_map_page(pci_dev, page, page_offset, len,
 253                                         PCI_DMA_TODEVICE);
 254         }
 255
 256         /* Transfer ownership of the skb to the final buffer */
 257         buffer->skb = skb;
 258         buffer->continuation = false;
 259
 260         /* Pass off to hardware */
 261         falcon_push_buffers(tx_queue);
 262
 263         return NETDEV_TX_OK;
 264
 265  pci_err:
 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);
 269
 270         /* Mark the packet as transmitted, and free the SKB ourselves */
 271         dev_kfree_skb_any((struct sk_buff *)skb);
 272         goto unwind;
 273
 274  stop:
 275         rc = NETDEV_TX_BUSY;
 276
 277         if (tx_queue->stopped == 1)
 278                 efx_stop_queue(efx);
 279
 280  unwind:
 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);
 287                 buffer->len = 0;
 288         }
 289
 290         /* Free the fragment we were mid-way through pushing */
 291         if (unmap_len) {
 292                 if (unmap_single)
 293                         pci_unmap_single(pci_dev, unmap_addr, unmap_len,
 294                                          PCI_DMA_TODEVICE);
 295                 else
 296                         pci_unmap_page(pci_dev, unmap_addr, unmap_len,
 297                                        PCI_DMA_TODEVICE);
 298         }
 299
 300         return rc;
 301 }
 302
 303 /* Remove packets from the TX queue
 304  *
 305  * This removes packets from the TX queue, up to and including the
 306  * specified index.
 307  */
 308 static inline void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
 309                                        unsigned int index)
 310 {
 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;
 314
 315         stop_index = (index + 1) & mask;
 316         read_ptr = tx_queue->read_count & mask;
 317
 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,
 323                                 read_ptr);
 324                         efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
 325                         return;
 326                 }
 327
 328                 efx_dequeue_buffer(tx_queue, buffer);
 329                 buffer->continuation = true;
 330                 buffer->len = 0;
 331
 332                 ++tx_queue->read_count;
 333                 read_ptr = tx_queue->read_count & mask;
 334         }
 335 }
 336
 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.
 340  *
 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.
 345  *
 346  * Context: netif_tx_lock held
 347  */
 348 inline int efx_xmit(struct efx_nic *efx,
 349                     struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 350 {
 351         int rc;
 352
 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))
 356                 goto out;
 357
 358         /* Update last TX timer */
 359         efx->net_dev->trans_start = jiffies;
 360
 361  out:
 362         return rc;
 363 }
 364
 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.
 369  *
 370  * Context: non-blocking.
 371  * Note that returning anything other than NETDEV_TX_OK will cause the
 372  * OS to free the skb.
 373  */
 374 int efx_hard_start_xmit(struct sk_buff *skb, struct net_device *net_dev)
 375 {
 376         struct efx_nic *efx = netdev_priv(net_dev);
 377         struct efx_tx_queue *tx_queue;
 378
 379         if (likely(skb->ip_summed == CHECKSUM_PARTIAL))
 380                 tx_queue = &efx->tx_queue[EFX_TX_QUEUE_OFFLOAD_CSUM];
 381         else
 382                 tx_queue = &efx->tx_queue[EFX_TX_QUEUE_NO_CSUM];
 383
 384         return efx_xmit(efx, tx_queue, skb);
 385 }
 386
 387 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 388 {
 389         unsigned fill_level;
 390         struct efx_nic *efx = tx_queue->efx;
 391
 392         EFX_BUG_ON_PARANOID(index > efx->type->txd_ring_mask);
 393
 394         efx_dequeue_buffers(tx_queue, index);
 395
 396         /* See if we need to restart the netif queue.  This barrier
 397          * separates the update of read_count from the test of
 398          * stopped. */
 399         smp_mb();
 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));
 404
 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;
 410                                 efx_wake_queue(efx);
 411                         }
 412                         netif_tx_unlock(efx->net_dev);
 413                 }
 414         }
 415 }
 416
 417 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
 418 {
 419         struct efx_nic *efx = tx_queue->efx;
 420         unsigned int txq_size;
 421         int i, rc;
 422
 423         EFX_LOG(efx, "creating TX queue %d\n", tx_queue->queue);
 424
 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)
 429                 return -ENOMEM;
 430         for (i = 0; i <= efx->type->txd_ring_mask; ++i)
 431                 tx_queue->buffer[i].continuation = true;
 432
 433         /* Allocate hardware ring */
 434         rc = falcon_probe_tx(tx_queue);
 435         if (rc)
 436                 goto fail;
 437
 438         return 0;
 439
 440  fail:
 441         kfree(tx_queue->buffer);
 442         tx_queue->buffer = NULL;
 443         return rc;
 444 }
 445
 446 int efx_init_tx_queue(struct efx_tx_queue *tx_queue)
 447 {
 448         EFX_LOG(tx_queue->efx, "initialising TX queue %d\n", tx_queue->queue);
 449
 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);
 455
 456         /* Set up TX descriptor ring */
 457         return falcon_init_tx(tx_queue);
 458 }
 459
 460 void efx_release_tx_buffers(struct efx_tx_queue *tx_queue)
 461 {
 462         struct efx_tx_buffer *buffer;
 463
 464         if (!tx_queue->buffer)
 465                 return;
 466
 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;
 473                 buffer->len = 0;
 474
 475                 ++tx_queue->read_count;
 476         }
 477 }
 478
 479 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
 480 {
 481         EFX_LOG(tx_queue->efx, "shutting down TX queue %d\n", tx_queue->queue);
 482
 483         /* Flush TX queue, remove descriptor ring */
 484         falcon_fini_tx(tx_queue);
 485
 486         efx_release_tx_buffers(tx_queue);
 487
 488         /* Free up TSO header cache */
 489         efx_fini_tso(tx_queue);
 490
 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);
 495         }
 496 }
 497
 498 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
 499 {
 500         EFX_LOG(tx_queue->efx, "destroying TX queue %d\n", tx_queue->queue);
 501         falcon_remove_tx(tx_queue);
 502
 503         kfree(tx_queue->buffer);
 504         tx_queue->buffer = NULL;
 505 }
 506
 507
 508 /* Efx TCP segmentation acceleration.
 509  *
 510  * Why?  Because by doing it here in the driver we can go significantly
 511  * faster than the GSO.
 512  *
 513  * Requires TX checksum offload support.
 514  */
 515
 516 /* Number of bytes inserted at the start of a TSO header buffer,
 517  * similar to NET_IP_ALIGN.
 518  */
 519 #if defined(__i386__) || defined(__x86_64__)
 520 #define TSOH_OFFSET     0
 521 #else
 522 #define TSOH_OFFSET     NET_IP_ALIGN
 523 #endif
 524
 525 #define TSOH_BUFFER(tsoh)       ((u8 *)(tsoh + 1) + TSOH_OFFSET)
 526
 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)
 530
 531 /* Size of blocks on free list.  Larger blocks must be allocated from
 532  * the heap.
 533  */
 534 #define TSOH_STD_SIZE           128
 535
 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)
 540
 541 /**
 542  * struct tso_state - TSO state for an SKB
 543  * @out_len: Remaining length in current segment
 544  * @seqnum: Current sequence number
 545  * @ipv4_id: Current IPv4 ID, host endian
 546  * @packet_space: Remaining space in current packet
 547  * @dma_addr: DMA address of current position
 548  * @in_len: Remaining length in current SKB fragment
 549  * @unmap_len: Length of SKB fragment
 550  * @unmap_addr: DMA address of SKB fragment
 551  * @unmap_single: DMA single vs page mapping flag
 552  * @header_len: Number of bytes of header
 553  * @full_packet_size: Number of bytes to put in each outgoing segment
 554  *
 555  * The state used during segmentation.  It is put into this data structure
 556  * just to make it easy to pass into inline functions.
 557  */
 558 struct tso_state {
 559         /* Output position */
 560         unsigned out_len;
 561         unsigned seqnum;
 562         unsigned ipv4_id;
 563         unsigned packet_space;
 564
 565         /* Input position */
 566         dma_addr_t dma_addr;
 567         unsigned in_len;
 568         unsigned unmap_len;
 569         dma_addr_t unmap_addr;
 570         bool unmap_single;
 571
 572         unsigned header_len;
 573         int full_packet_size;
 574 };
 575
 576
 577 /*
 578  * Verify that our various assumptions about sk_buffs and the conditions
 579  * under which TSO will be attempted hold true.
 580  */
 581 static inline void efx_tso_check_safe(const struct sk_buff *skb)
 582 {
 583         EFX_BUG_ON_PARANOID(skb->protocol != htons(ETH_P_IP));
 584         EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
 585                             skb->protocol);
 586         EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
 587         EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
 588                              + (tcp_hdr(skb)->doff << 2u)) >
 589                             skb_headlen(skb));
 590 }
 591
 592
 593 /*
 594  * Allocate a page worth of efx_tso_header structures, and string them
 595  * into the tx_queue->tso_headers_free linked list. Return 0 or -ENOMEM.
 596  */
 597 static int efx_tsoh_block_alloc(struct efx_tx_queue *tx_queue)
 598 {
 599
 600         struct pci_dev *pci_dev = tx_queue->efx->pci_dev;
 601         struct efx_tso_header *tsoh;
 602         dma_addr_t dma_addr;
 603         u8 *base_kva, *kva;
 604
 605         base_kva = pci_alloc_consistent(pci_dev, PAGE_SIZE, &dma_addr);
 606         if (base_kva == NULL) {
 607                 EFX_ERR(tx_queue->efx, "Unable to allocate page for TSO"
 608                         " headers\n");
 609                 return -ENOMEM;
 610         }
 611
 612         /* pci_alloc_consistent() allocates pages. */
 613         EFX_BUG_ON_PARANOID(dma_addr & (PAGE_SIZE - 1u));
 614
 615         for (kva = base_kva; kva < base_kva + PAGE_SIZE; kva += TSOH_STD_SIZE) {
 616                 tsoh = (struct efx_tso_header *)kva;
 617                 tsoh->dma_addr = dma_addr + (TSOH_BUFFER(tsoh) - base_kva);
 618                 tsoh->next = tx_queue->tso_headers_free;
 619                 tx_queue->tso_headers_free = tsoh;
 620         }
 621
 622         return 0;
 623 }
 624
 625
 626 /* Free up a TSO header, and all others in the same page. */
 627 static void efx_tsoh_block_free(struct efx_tx_queue *tx_queue,
 628                                 struct efx_tso_header *tsoh,
 629                                 struct pci_dev *pci_dev)
 630 {
 631         struct efx_tso_header **p;
 632         unsigned long base_kva;
 633         dma_addr_t base_dma;
 634
 635         base_kva = (unsigned long)tsoh & PAGE_MASK;
 636         base_dma = tsoh->dma_addr & PAGE_MASK;
 637
 638         p = &tx_queue->tso_headers_free;
 639         while (*p != NULL) {
 640                 if (((unsigned long)*p & PAGE_MASK) == base_kva)
 641                         *p = (*p)->next;
 642                 else
 643                         p = &(*p)->next;
 644         }
 645
 646         pci_free_consistent(pci_dev, PAGE_SIZE, (void *)base_kva, base_dma);
 647 }
 648
 649 static struct efx_tso_header *
 650 efx_tsoh_heap_alloc(struct efx_tx_queue *tx_queue, size_t header_len)
 651 {
 652         struct efx_tso_header *tsoh;
 653
 654         tsoh = kmalloc(TSOH_SIZE(header_len), GFP_ATOMIC | GFP_DMA);
 655         if (unlikely(!tsoh))
 656                 return NULL;
 657
 658         tsoh->dma_addr = pci_map_single(tx_queue->efx->pci_dev,
 659                                         TSOH_BUFFER(tsoh), header_len,
 660                                         PCI_DMA_TODEVICE);
 661         if (unlikely(pci_dma_mapping_error(tx_queue->efx->pci_dev,
 662                                            tsoh->dma_addr))) {
 663                 kfree(tsoh);
 664                 return NULL;
 665         }
 666
 667         tsoh->unmap_len = header_len;
 668         return tsoh;
 669 }
 670
 671 static void
 672 efx_tsoh_heap_free(struct efx_tx_queue *tx_queue, struct efx_tso_header *tsoh)
 673 {
 674         pci_unmap_single(tx_queue->efx->pci_dev,
 675                          tsoh->dma_addr, tsoh->unmap_len,
 676                          PCI_DMA_TODEVICE);
 677         kfree(tsoh);
 678 }
 679
 680 /**
 681  * efx_tx_queue_insert - push descriptors onto the TX queue
 682  * @tx_queue:           Efx TX queue
 683  * @dma_addr:           DMA address of fragment
 684  * @len:                Length of fragment
 685  * @final_buffer:       The final buffer inserted into the queue
 686  *
 687  * Push descriptors onto the TX queue.  Return 0 on success or 1 if
 688  * @tx_queue full.
 689  */
 690 static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
 691                                dma_addr_t dma_addr, unsigned len,
 692                                struct efx_tx_buffer **final_buffer)
 693 {
 694         struct efx_tx_buffer *buffer;
 695         struct efx_nic *efx = tx_queue->efx;
 696         unsigned dma_len, fill_level, insert_ptr, misalign;
 697         int q_space;
 698
 699         EFX_BUG_ON_PARANOID(len <= 0);
 700
 701         fill_level = tx_queue->insert_count - tx_queue->old_read_count;
 702         /* -1 as there is no way to represent all descriptors used */
 703         q_space = efx->type->txd_ring_mask - 1 - fill_level;
 704
 705         while (1) {
 706                 if (unlikely(q_space-- <= 0)) {
 707                         /* It might be that completions have happened
 708                          * since the xmit path last checked.  Update
 709                          * the xmit path's copy of read_count.
 710                          */
 711                         ++tx_queue->stopped;
 712                         /* This memory barrier protects the change of
 713                          * stopped from the access of read_count. */
 714                         smp_mb();
 715                         tx_queue->old_read_count =
 716                                 *(volatile unsigned *)&tx_queue->read_count;
 717                         fill_level = (tx_queue->insert_count
 718                                       - tx_queue->old_read_count);
 719                         q_space = efx->type->txd_ring_mask - 1 - fill_level;
 720                         if (unlikely(q_space-- <= 0)) {
 721                                 *final_buffer = NULL;
 722                                 return 1;
 723                         }
 724                         smp_mb();
 725                         --tx_queue->stopped;
 726                 }
 727
 728                 insert_ptr = tx_queue->insert_count & efx->type->txd_ring_mask;
 729                 buffer = &tx_queue->buffer[insert_ptr];
 730                 ++tx_queue->insert_count;
 731
 732                 EFX_BUG_ON_PARANOID(tx_queue->insert_count -
 733                                     tx_queue->read_count >
 734                                     efx->type->txd_ring_mask);
 735
 736                 efx_tsoh_free(tx_queue, buffer);
 737                 EFX_BUG_ON_PARANOID(buffer->len);
 738                 EFX_BUG_ON_PARANOID(buffer->unmap_len);
 739                 EFX_BUG_ON_PARANOID(buffer->skb);
 740                 EFX_BUG_ON_PARANOID(!buffer->continuation);
 741                 EFX_BUG_ON_PARANOID(buffer->tsoh);
 742
 743                 buffer->dma_addr = dma_addr;
 744
 745                 /* Ensure we do not cross a boundary unsupported by H/W */
 746                 dma_len = (~dma_addr & efx->type->tx_dma_mask) + 1;
 747
 748                 misalign = (unsigned)dma_addr & efx->type->bug5391_mask;
 749                 if (misalign && dma_len + misalign > 512)
 750                         dma_len = 512 - misalign;
 751
 752                 /* If there is enough space to send then do so */
 753                 if (dma_len >= len)
 754                         break;
 755
 756                 buffer->len = dma_len; /* Don't set the other members */
 757                 dma_addr += dma_len;
 758                 len -= dma_len;
 759         }
 760
 761         EFX_BUG_ON_PARANOID(!len);
 762         buffer->len = len;
 763         *final_buffer = buffer;
 764         return 0;
 765 }
 766
 767
 768 /*
 769  * Put a TSO header into the TX queue.
 770  *
 771  * This is special-cased because we know that it is small enough to fit in
 772  * a single fragment, and we know it doesn't cross a page boundary.  It
 773  * also allows us to not worry about end-of-packet etc.
 774  */
 775 static inline void efx_tso_put_header(struct efx_tx_queue *tx_queue,
 776                                       struct efx_tso_header *tsoh, unsigned len)
 777 {
 778         struct efx_tx_buffer *buffer;
 779
 780         buffer = &tx_queue->buffer[tx_queue->insert_count &
 781                                    tx_queue->efx->type->txd_ring_mask];
 782         efx_tsoh_free(tx_queue, buffer);
 783         EFX_BUG_ON_PARANOID(buffer->len);
 784         EFX_BUG_ON_PARANOID(buffer->unmap_len);
 785         EFX_BUG_ON_PARANOID(buffer->skb);
 786         EFX_BUG_ON_PARANOID(!buffer->continuation);
 787         EFX_BUG_ON_PARANOID(buffer->tsoh);
 788         buffer->len = len;
 789         buffer->dma_addr = tsoh->dma_addr;
 790         buffer->tsoh = tsoh;
 791
 792         ++tx_queue->insert_count;
 793 }
 794
 795
 796 /* Remove descriptors put into a tx_queue. */
 797 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
 798 {
 799         struct efx_tx_buffer *buffer;
 800         dma_addr_t unmap_addr;
 801
 802         /* Work backwards until we hit the original insert pointer value */
 803         while (tx_queue->insert_count != tx_queue->write_count) {
 804                 --tx_queue->insert_count;
 805                 buffer = &tx_queue->buffer[tx_queue->insert_count &
 806                                            tx_queue->efx->type->txd_ring_mask];
 807                 efx_tsoh_free(tx_queue, buffer);
 808                 EFX_BUG_ON_PARANOID(buffer->skb);
 809                 buffer->len = 0;
 810                 buffer->continuation = true;
 811                 if (buffer->unmap_len) {
 812                         unmap_addr = (buffer->dma_addr + buffer->len -
 813                                       buffer->unmap_len);
 814                         if (buffer->unmap_single)
 815                                 pci_unmap_single(tx_queue->efx->pci_dev,
 816                                                  unmap_addr, buffer->unmap_len,
 817                                                  PCI_DMA_TODEVICE);
 818                         else
 819                                 pci_unmap_page(tx_queue->efx->pci_dev,
 820                                                unmap_addr, buffer->unmap_len,
 821                                                PCI_DMA_TODEVICE);
 822                         buffer->unmap_len = 0;
 823                 }
 824         }
 825 }
 826
 827
 828 /* Parse the SKB header and initialise state. */
 829 static inline void tso_start(struct tso_state *st, const struct sk_buff *skb)
 830 {
 831         /* All ethernet/IP/TCP headers combined size is TCP header size
 832          * plus offset of TCP header relative to start of packet.
 833          */
 834         st->header_len = ((tcp_hdr(skb)->doff << 2u)
 835                           + PTR_DIFF(tcp_hdr(skb), skb->data));
 836         st->full_packet_size = st->header_len + skb_shinfo(skb)->gso_size;
 837
 838         st->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->full_packet_size;
 846         st->out_len = skb->len - st->header_len;
 847         st->unmap_len = 0;
 848         st->unmap_single = false;
 849 }
 850
 851 static inline int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
 852                                    skb_frag_t *frag)
 853 {
 854         st->unmap_addr = pci_map_page(efx->pci_dev, frag->page,
 855                                       frag->page_offset, frag->size,
 856                                       PCI_DMA_TODEVICE);
 857         if (likely(!pci_dma_mapping_error(efx->pci_dev, st->unmap_addr))) {
 858                 st->unmap_single = false;
 859                 st->unmap_len = frag->size;
 860                 st->in_len = frag->size;
 861                 st->dma_addr = st->unmap_addr;
 862                 return 0;
 863         }
 864         return -ENOMEM;
 865 }
 866
 867 static inline int
 868 tso_get_head_fragment(struct tso_state *st, struct efx_nic *efx,
 869                       const struct sk_buff *skb)
 870 {
 871         int hl = st->header_len;
 872         int len = skb_headlen(skb) - hl;
 873
 874         st->unmap_addr = pci_map_single(efx->pci_dev, skb->data + hl,
 875                                         len, PCI_DMA_TODEVICE);
 876         if (likely(!pci_dma_mapping_error(efx->pci_dev, st->unmap_addr))) {
 877                 st->unmap_single = true;
 878                 st->unmap_len = len;
 879                 st->in_len = len;
 880                 st->dma_addr = st->unmap_addr;
 881                 return 0;
 882         }
 883         return -ENOMEM;
 884 }
 885
 886
 887 /**
 888  * tso_fill_packet_with_fragment - form descriptors for the current fragment
 889  * @tx_queue:           Efx TX queue
 890  * @skb:                Socket buffer
 891  * @st:                 TSO state
 892  *
 893  * Form descriptors for the current fragment, until we reach the end
 894  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
 895  * space in @tx_queue.
 896  */
 897 static inline int tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
 898                                                 const struct sk_buff *skb,
 899                                                 struct tso_state *st)
 900 {
 901         struct efx_tx_buffer *buffer;
 902         int n, end_of_packet, rc;
 903
 904         if (st->in_len == 0)
 905                 return 0;
 906         if (st->packet_space == 0)
 907                 return 0;
 908
 909         EFX_BUG_ON_PARANOID(st->in_len <= 0);
 910         EFX_BUG_ON_PARANOID(st->packet_space <= 0);
 911
 912         n = min(st->in_len, st->packet_space);
 913
 914         st->packet_space -= n;
 915         st->out_len -= n;
 916         st->in_len -= n;
 917
 918         rc = efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer);
 919         if (likely(rc == 0)) {
 920                 if (st->out_len == 0)
 921                         /* Transfer ownership of the skb */
 922                         buffer->skb = skb;
 923
 924                 end_of_packet = st->out_len == 0 || st->packet_space == 0;
 925                 buffer->continuation = !end_of_packet;
 926
 927                 if (st->in_len == 0) {
 928                         /* Transfer ownership of the pci mapping */
 929                         buffer->unmap_len = st->unmap_len;
 930                         buffer->unmap_single = st->unmap_single;
 931                         st->unmap_len = 0;
 932                 }
 933         }
 934
 935         st->dma_addr += n;
 936         return rc;
 937 }
 938
 939
 940 /**
 941  * tso_start_new_packet - generate a new header and prepare for the new packet
 942  * @tx_queue:           Efx TX queue
 943  * @skb:                Socket buffer
 944  * @st:                 TSO state
 945  *
 946  * Generate a new header and prepare for the new packet.  Return 0 on
 947  * success, or -1 if failed to alloc header.
 948  */
 949 static inline int tso_start_new_packet(struct efx_tx_queue *tx_queue,
 950                                        const struct sk_buff *skb,
 951                                        struct tso_state *st)
 952 {
 953         struct efx_tso_header *tsoh;
 954         struct iphdr *tsoh_iph;
 955         struct tcphdr *tsoh_th;
 956         unsigned ip_length;
 957         u8 *header;
 958
 959         /* Allocate a DMA-mapped header buffer. */
 960         if (likely(TSOH_SIZE(st->header_len) <= TSOH_STD_SIZE)) {
 961                 if (tx_queue->tso_headers_free == NULL) {
 962                         if (efx_tsoh_block_alloc(tx_queue))
 963                                 return -1;
 964                 }
 965                 EFX_BUG_ON_PARANOID(!tx_queue->tso_headers_free);
 966                 tsoh = tx_queue->tso_headers_free;
 967                 tx_queue->tso_headers_free = tsoh->next;
 968                 tsoh->unmap_len = 0;
 969         } else {
 970                 tx_queue->tso_long_headers++;
 971                 tsoh = efx_tsoh_heap_alloc(tx_queue, st->header_len);
 972                 if (unlikely(!tsoh))
 973                         return -1;
 974         }
 975
 976         header = TSOH_BUFFER(tsoh);
 977         tsoh_th = (struct tcphdr *)(header + SKB_TCP_OFF(skb));
 978         tsoh_iph = (struct iphdr *)(header + SKB_IPV4_OFF(skb));
 979
 980         /* Copy and update the headers. */
 981         memcpy(header, skb->data, st->header_len);
 982
 983         tsoh_th->seq = htonl(st->seqnum);
 984         st->seqnum += skb_shinfo(skb)->gso_size;
 985         if (st->out_len > skb_shinfo(skb)->gso_size) {
 986                 /* This packet will not finish the TSO burst. */
 987                 ip_length = st->full_packet_size - ETH_HDR_LEN(skb);
 988                 tsoh_th->fin = 0;
 989                 tsoh_th->psh = 0;
 990         } else {
 991                 /* This packet will be the last in the TSO burst. */
 992                 ip_length = st->header_len - ETH_HDR_LEN(skb) + st->out_len;
 993                 tsoh_th->fin = tcp_hdr(skb)->fin;
 994                 tsoh_th->psh = tcp_hdr(skb)->psh;
 995         }
 996         tsoh_iph->tot_len = htons(ip_length);
 997
 998         /* Linux leaves suitable gaps in the IP ID space for us to fill. */
 999         tsoh_iph->id = htons(st->ipv4_id);
1000         st->ipv4_id++;
1001
1002         st->packet_space = skb_shinfo(skb)->gso_size;
1003         ++tx_queue->tso_packets;
1004
1005         /* Form a descriptor for this header. */
1006         efx_tso_put_header(tx_queue, tsoh, st->header_len);
1007
1008         return 0;
1009 }
1010
1011
1012 /**
1013  * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1014  * @tx_queue:           Efx TX queue
1015  * @skb:                Socket buffer
1016  *
1017  * Context: You must hold netif_tx_lock() to call this function.
1018  *
1019  * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1020  * @skb was not enqueued.  In all cases @skb is consumed.  Return
1021  * %NETDEV_TX_OK or %NETDEV_TX_BUSY.
1022  */
1023 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1024                                const struct sk_buff *skb)
1025 {
1026         struct efx_nic *efx = tx_queue->efx;
1027         int frag_i, rc, rc2 = NETDEV_TX_OK;
1028         struct tso_state state;
1029
1030         /* Verify TSO is safe - these checks should never fail. */
1031         efx_tso_check_safe(skb);
1032
1033         EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
1034
1035         tso_start(&state, skb);
1036
1037         /* Assume that skb header area contains exactly the headers, and
1038          * all payload is in the frag list.
1039          */
1040         if (skb_headlen(skb) == state.header_len) {
1041                 /* Grab the first payload fragment. */
1042                 EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1043                 frag_i = 0;
1044                 rc = tso_get_fragment(&state, efx,
1045                                       skb_shinfo(skb)->frags + frag_i);
1046                 if (rc)
1047                         goto mem_err;
1048         } else {
1049                 rc = tso_get_head_fragment(&state, efx, skb);
1050                 if (rc)
1051                         goto mem_err;
1052                 frag_i = -1;
1053         }
1054
1055         if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1056                 goto mem_err;
1057
1058         while (1) {
1059                 rc = tso_fill_packet_with_fragment(tx_queue, skb, &state);
1060                 if (unlikely(rc))
1061                         goto stop;
1062
1063                 /* Move onto the next fragment? */
1064                 if (state.in_len == 0) {
1065                         if (++frag_i >= skb_shinfo(skb)->nr_frags)
1066                                 /* End of payload reached. */
1067                                 break;
1068                         rc = tso_get_fragment(&state, efx,
1069                                               skb_shinfo(skb)->frags + frag_i);
1070                         if (rc)
1071                                 goto mem_err;
1072                 }
1073
1074                 /* Start at new packet? */
1075                 if (state.packet_space == 0 &&
1076                     tso_start_new_packet(tx_queue, skb, &state) < 0)
1077                         goto mem_err;
1078         }
1079
1080         /* Pass off to hardware */
1081         falcon_push_buffers(tx_queue);
1082
1083         tx_queue->tso_bursts++;
1084         return NETDEV_TX_OK;
1085
1086  mem_err:
1087         EFX_ERR(efx, "Out of memory for TSO headers, or PCI mapping error\n");
1088         dev_kfree_skb_any((struct sk_buff *)skb);
1089         goto unwind;
1090
1091  stop:
1092         rc2 = NETDEV_TX_BUSY;
1093
1094         /* Stop the queue if it wasn't stopped before. */
1095         if (tx_queue->stopped == 1)
1096                 efx_stop_queue(efx);
1097
1098  unwind:
1099         /* Free the DMA mapping we were in the process of writing out */
1100         if (state.unmap_len) {
1101                 if (state.unmap_single)
1102                         pci_unmap_single(efx->pci_dev, state.unmap_addr,
1103                                          state.unmap_len, PCI_DMA_TODEVICE);
1104                 else
1105                         pci_unmap_page(efx->pci_dev, state.unmap_addr,
1106                                        state.unmap_len, PCI_DMA_TODEVICE);
1107         }
1108
1109         efx_enqueue_unwind(tx_queue);
1110         return rc2;
1111 }
1112
1113
1114 /*
1115  * Free up all TSO datastructures associated with tx_queue. This
1116  * routine should be called only once the tx_queue is both empty and
1117  * will no longer be used.
1118  */
1119 static void efx_fini_tso(struct efx_tx_queue *tx_queue)
1120 {
1121         unsigned i;
1122
1123         if (tx_queue->buffer) {
1124                 for (i = 0; i <= tx_queue->efx->type->txd_ring_mask; ++i)
1125                         efx_tsoh_free(tx_queue, &tx_queue->buffer[i]);
1126         }
1127
1128         while (tx_queue->tso_headers_free != NULL)
1129                 efx_tsoh_block_free(tx_queue, tx_queue->tso_headers_free,
1130                                     tx_queue->efx->pci_dev);
1131 }