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 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 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                                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 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 int efx_enqueue_skb(struct efx_tx_queue *tx_queue,
 142                            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         /* Pad if necessary */
 166         if (EFX_WORKAROUND_15592(efx) && skb->len <= 32) {
 167                 EFX_BUG_ON_PARANOID(skb->data_len);
 168                 len = 32 + 1;
 169                 if (skb_pad(skb, len - skb->len))
 170                         return NETDEV_TX_OK;
 171         }
 172
 173         fill_level = tx_queue->insert_count - tx_queue->old_read_count;
 174         q_space = efx->type->txd_ring_mask - 1 - fill_level;
 175
 176         /* Map for DMA.  Use pci_map_single rather than pci_map_page
 177          * since this is more efficient on machines with sparse
 178          * memory.
 179          */
 180         unmap_single = true;
 181         dma_addr = pci_map_single(pci_dev, skb->data, len, PCI_DMA_TODEVICE);
 182
 183         /* Process all fragments */
 184         while (1) {
 185                 if (unlikely(pci_dma_mapping_error(pci_dev, dma_addr)))
 186                         goto pci_err;
 187
 188                 /* Store fields for marking in the per-fragment final
 189                  * descriptor */
 190                 unmap_len = len;
 191                 unmap_addr = dma_addr;
 192
 193                 /* Add to TX queue, splitting across DMA boundaries */
 194                 do {
 195                         if (unlikely(q_space-- <= 0)) {
 196                                 /* It might be that completions have
 197                                  * happened since the xmit path last
 198                                  * checked.  Update the xmit path's
 199                                  * copy of read_count.
 200                                  */
 201                                 ++tx_queue->stopped;
 202                                 /* This memory barrier protects the
 203                                  * change of stopped from the access
 204                                  * of read_count. */
 205                                 smp_mb();
 206                                 tx_queue->old_read_count =
 207                                         *(volatile unsigned *)
 208                                         &tx_queue->read_count;
 209                                 fill_level = (tx_queue->insert_count
 210                                               - tx_queue->old_read_count);
 211                                 q_space = (efx->type->txd_ring_mask - 1 -
 212                                            fill_level);
 213                                 if (unlikely(q_space-- <= 0))
 214                                         goto stop;
 215                                 smp_mb();
 216                                 --tx_queue->stopped;
 217                         }
 218
 219                         insert_ptr = (tx_queue->insert_count &
 220                                       efx->type->txd_ring_mask);
 221                         buffer = &tx_queue->buffer[insert_ptr];
 222                         efx_tsoh_free(tx_queue, buffer);
 223                         EFX_BUG_ON_PARANOID(buffer->tsoh);
 224                         EFX_BUG_ON_PARANOID(buffer->skb);
 225                         EFX_BUG_ON_PARANOID(buffer->len);
 226                         EFX_BUG_ON_PARANOID(!buffer->continuation);
 227                         EFX_BUG_ON_PARANOID(buffer->unmap_len);
 228
 229                         dma_len = (((~dma_addr) & efx->type->tx_dma_mask) + 1);
 230                         if (likely(dma_len > len))
 231                                 dma_len = len;
 232
 233                         misalign = (unsigned)dma_addr & efx->type->bug5391_mask;
 234                         if (misalign && dma_len + misalign > 512)
 235                                 dma_len = 512 - misalign;
 236
 237                         /* Fill out per descriptor fields */
 238                         buffer->len = dma_len;
 239                         buffer->dma_addr = dma_addr;
 240                         len -= dma_len;
 241                         dma_addr += dma_len;
 242                         ++tx_queue->insert_count;
 243                 } while (len);
 244
 245                 /* Transfer ownership of the unmapping to the final buffer */
 246                 buffer->unmap_single = unmap_single;
 247                 buffer->unmap_len = unmap_len;
 248                 unmap_len = 0;
 249
 250                 /* Get address and size of next fragment */
 251                 if (i >= skb_shinfo(skb)->nr_frags)
 252                         break;
 253                 fragment = &skb_shinfo(skb)->frags[i];
 254                 len = fragment->size;
 255                 page = fragment->page;
 256                 page_offset = fragment->page_offset;
 257                 i++;
 258                 /* Map for DMA */
 259                 unmap_single = false;
 260                 dma_addr = pci_map_page(pci_dev, page, page_offset, len,
 261                                         PCI_DMA_TODEVICE);
 262         }
 263
 264         /* Transfer ownership of the skb to the final buffer */
 265         buffer->skb = skb;
 266         buffer->continuation = false;
 267
 268         /* Pass off to hardware */
 269         falcon_push_buffers(tx_queue);
 270
 271         return NETDEV_TX_OK;
 272
 273  pci_err:
 274         EFX_ERR_RL(efx, " TX queue %d could not map skb with %d bytes %d "
 275                    "fragments for DMA\n", tx_queue->queue, skb->len,
 276                    skb_shinfo(skb)->nr_frags + 1);
 277
 278         /* Mark the packet as transmitted, and free the SKB ourselves */
 279         dev_kfree_skb_any((struct sk_buff *)skb);
 280         goto unwind;
 281
 282  stop:
 283         rc = NETDEV_TX_BUSY;
 284
 285         if (tx_queue->stopped == 1)
 286                 efx_stop_queue(efx);
 287
 288  unwind:
 289         /* Work backwards until we hit the original insert pointer value */
 290         while (tx_queue->insert_count != tx_queue->write_count) {
 291                 --tx_queue->insert_count;
 292                 insert_ptr = tx_queue->insert_count & efx->type->txd_ring_mask;
 293                 buffer = &tx_queue->buffer[insert_ptr];
 294                 efx_dequeue_buffer(tx_queue, buffer);
 295                 buffer->len = 0;
 296         }
 297
 298         /* Free the fragment we were mid-way through pushing */
 299         if (unmap_len) {
 300                 if (unmap_single)
 301                         pci_unmap_single(pci_dev, unmap_addr, unmap_len,
 302                                          PCI_DMA_TODEVICE);
 303                 else
 304                         pci_unmap_page(pci_dev, unmap_addr, unmap_len,
 305                                        PCI_DMA_TODEVICE);
 306         }
 307
 308         return rc;
 309 }
 310
 311 /* Remove packets from the TX queue
 312  *
 313  * This removes packets from the TX queue, up to and including the
 314  * specified index.
 315  */
 316 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
 317                                 unsigned int index)
 318 {
 319         struct efx_nic *efx = tx_queue->efx;
 320         unsigned int stop_index, read_ptr;
 321         unsigned int mask = tx_queue->efx->type->txd_ring_mask;
 322
 323         stop_index = (index + 1) & mask;
 324         read_ptr = tx_queue->read_count & mask;
 325
 326         while (read_ptr != stop_index) {
 327                 struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
 328                 if (unlikely(buffer->len == 0)) {
 329                         EFX_ERR(tx_queue->efx, "TX queue %d spurious TX "
 330                                 "completion id %x\n", tx_queue->queue,
 331                                 read_ptr);
 332                         efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
 333                         return;
 334                 }
 335
 336                 efx_dequeue_buffer(tx_queue, buffer);
 337                 buffer->continuation = true;
 338                 buffer->len = 0;
 339
 340                 ++tx_queue->read_count;
 341                 read_ptr = tx_queue->read_count & mask;
 342         }
 343 }
 344
 345 /* Initiate a packet transmission on the specified TX queue.
 346  * Note that returning anything other than NETDEV_TX_OK will cause the
 347  * OS to free the skb.
 348  *
 349  * This function is split out from efx_hard_start_xmit to allow the
 350  * loopback test to direct packets via specific TX queues.  It is
 351  * therefore a non-static inline, so as not to penalise performance
 352  * for non-loopback transmissions.
 353  *
 354  * Context: netif_tx_lock held
 355  */
 356 inline int efx_xmit(struct efx_nic *efx,
 357                     struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 358 {
 359         int rc;
 360
 361         /* Map fragments for DMA and add to TX queue */
 362         rc = efx_enqueue_skb(tx_queue, skb);
 363         return rc;
 364 }
 365
 366 /* Initiate a packet transmission.  We use one channel per CPU
 367  * (sharing when we have more CPUs than channels).  On Falcon, the TX
 368  * completion events will be directed back to the CPU that transmitted
 369  * the packet, which should be cache-efficient.
 370  *
 371  * Context: non-blocking.
 372  * Note that returning anything other than NETDEV_TX_OK will cause the
 373  * OS to free the skb.
 374  */
 375 int efx_hard_start_xmit(struct sk_buff *skb, struct net_device *net_dev)
 376 {
 377         struct efx_nic *efx = netdev_priv(net_dev);
 378         struct efx_tx_queue *tx_queue;
 379
 380         if (unlikely(efx->port_inhibited))
 381                 return NETDEV_TX_BUSY;
 382
 383         if (likely(skb->ip_summed == CHECKSUM_PARTIAL))
 384                 tx_queue = &efx->tx_queue[EFX_TX_QUEUE_OFFLOAD_CSUM];
 385         else
 386                 tx_queue = &efx->tx_queue[EFX_TX_QUEUE_NO_CSUM];
 387
 388         return efx_xmit(efx, tx_queue, skb);
 389 }
 390
 391 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 392 {
 393         unsigned fill_level;
 394         struct efx_nic *efx = tx_queue->efx;
 395
 396         EFX_BUG_ON_PARANOID(index > efx->type->txd_ring_mask);
 397
 398         efx_dequeue_buffers(tx_queue, index);
 399
 400         /* See if we need to restart the netif queue.  This barrier
 401          * separates the update of read_count from the test of
 402          * stopped. */
 403         smp_mb();
 404         if (unlikely(tx_queue->stopped) && likely(efx->port_enabled)) {
 405                 fill_level = tx_queue->insert_count - tx_queue->read_count;
 406                 if (fill_level < EFX_NETDEV_TX_THRESHOLD(tx_queue)) {
 407                         EFX_BUG_ON_PARANOID(!efx_dev_registered(efx));
 408
 409                         /* Do this under netif_tx_lock(), to avoid racing
 410                          * with efx_xmit(). */
 411                         netif_tx_lock(efx->net_dev);
 412                         if (tx_queue->stopped) {
 413                                 tx_queue->stopped = 0;
 414                                 efx_wake_queue(efx);
 415                         }
 416                         netif_tx_unlock(efx->net_dev);
 417                 }
 418         }
 419 }
 420
 421 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
 422 {
 423         struct efx_nic *efx = tx_queue->efx;
 424         unsigned int txq_size;
 425         int i, rc;
 426
 427         EFX_LOG(efx, "creating TX queue %d\n", tx_queue->queue);
 428
 429         /* Allocate software ring */
 430         txq_size = (efx->type->txd_ring_mask + 1) * sizeof(*tx_queue->buffer);
 431         tx_queue->buffer = kzalloc(txq_size, GFP_KERNEL);
 432         if (!tx_queue->buffer)
 433                 return -ENOMEM;
 434         for (i = 0; i <= efx->type->txd_ring_mask; ++i)
 435                 tx_queue->buffer[i].continuation = true;
 436
 437         /* Allocate hardware ring */
 438         rc = falcon_probe_tx(tx_queue);
 439         if (rc)
 440                 goto fail;
 441
 442         return 0;
 443
 444  fail:
 445         kfree(tx_queue->buffer);
 446         tx_queue->buffer = NULL;
 447         return rc;
 448 }
 449
 450 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
 451 {
 452         EFX_LOG(tx_queue->efx, "initialising TX queue %d\n", tx_queue->queue);
 453
 454         tx_queue->insert_count = 0;
 455         tx_queue->write_count = 0;
 456         tx_queue->read_count = 0;
 457         tx_queue->old_read_count = 0;
 458         BUG_ON(tx_queue->stopped);
 459
 460         /* Set up TX descriptor ring */
 461         falcon_init_tx(tx_queue);
 462 }
 463
 464 void efx_release_tx_buffers(struct efx_tx_queue *tx_queue)
 465 {
 466         struct efx_tx_buffer *buffer;
 467
 468         if (!tx_queue->buffer)
 469                 return;
 470
 471         /* Free any buffers left in the ring */
 472         while (tx_queue->read_count != tx_queue->write_count) {
 473                 buffer = &tx_queue->buffer[tx_queue->read_count &
 474                                            tx_queue->efx->type->txd_ring_mask];
 475                 efx_dequeue_buffer(tx_queue, buffer);
 476                 buffer->continuation = true;
 477                 buffer->len = 0;
 478
 479                 ++tx_queue->read_count;
 480         }
 481 }
 482
 483 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
 484 {
 485         EFX_LOG(tx_queue->efx, "shutting down TX queue %d\n", tx_queue->queue);
 486
 487         /* Flush TX queue, remove descriptor ring */
 488         falcon_fini_tx(tx_queue);
 489
 490         efx_release_tx_buffers(tx_queue);
 491
 492         /* Free up TSO header cache */
 493         efx_fini_tso(tx_queue);
 494
 495         /* Release queue's stop on port, if any */
 496         if (tx_queue->stopped) {
 497                 tx_queue->stopped = 0;
 498                 efx_wake_queue(tx_queue->efx);
 499         }
 500 }
 501
 502 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
 503 {
 504         EFX_LOG(tx_queue->efx, "destroying TX queue %d\n", tx_queue->queue);
 505         falcon_remove_tx(tx_queue);
 506
 507         kfree(tx_queue->buffer);
 508         tx_queue->buffer = NULL;
 509 }
 510
 511
 512 /* Efx TCP segmentation acceleration.
 513  *
 514  * Why?  Because by doing it here in the driver we can go significantly
 515  * faster than the GSO.
 516  *
 517  * Requires TX checksum offload support.
 518  */
 519
 520 /* Number of bytes inserted at the start of a TSO header buffer,
 521  * similar to NET_IP_ALIGN.
 522  */
 523 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
 524 #define TSOH_OFFSET     0
 525 #else
 526 #define TSOH_OFFSET     NET_IP_ALIGN
 527 #endif
 528
 529 #define TSOH_BUFFER(tsoh)       ((u8 *)(tsoh + 1) + TSOH_OFFSET)
 530
 531 /* Total size of struct efx_tso_header, buffer and padding */
 532 #define TSOH_SIZE(hdr_len)                                      \
 533         (sizeof(struct efx_tso_header) + TSOH_OFFSET + hdr_len)
 534
 535 /* Size of blocks on free list.  Larger blocks must be allocated from
 536  * the heap.
 537  */
 538 #define TSOH_STD_SIZE           128
 539
 540 #define PTR_DIFF(p1, p2)  ((u8 *)(p1) - (u8 *)(p2))
 541 #define ETH_HDR_LEN(skb)  (skb_network_header(skb) - (skb)->data)
 542 #define SKB_TCP_OFF(skb)  PTR_DIFF(tcp_hdr(skb), (skb)->data)
 543 #define SKB_IPV4_OFF(skb) PTR_DIFF(ip_hdr(skb), (skb)->data)
 544
 545 /**
 546  * struct tso_state - TSO state for an SKB
 547  * @out_len: Remaining length in current segment
 548  * @seqnum: Current sequence number
 549  * @ipv4_id: Current IPv4 ID, host endian
 550  * @packet_space: Remaining space in current packet
 551  * @dma_addr: DMA address of current position
 552  * @in_len: Remaining length in current SKB fragment
 553  * @unmap_len: Length of SKB fragment
 554  * @unmap_addr: DMA address of SKB fragment
 555  * @unmap_single: DMA single vs page mapping flag
 556  * @header_len: Number of bytes of header
 557  * @full_packet_size: Number of bytes to put in each outgoing segment
 558  *
 559  * The state used during segmentation.  It is put into this data structure
 560  * just to make it easy to pass into inline functions.
 561  */
 562 struct tso_state {
 563         /* Output position */
 564         unsigned out_len;
 565         unsigned seqnum;
 566         unsigned ipv4_id;
 567         unsigned packet_space;
 568
 569         /* Input position */
 570         dma_addr_t dma_addr;
 571         unsigned in_len;
 572         unsigned unmap_len;
 573         dma_addr_t unmap_addr;
 574         bool unmap_single;
 575
 576         unsigned header_len;
 577         int full_packet_size;
 578 };
 579
 580
 581 /*
 582  * Verify that our various assumptions about sk_buffs and the conditions
 583  * under which TSO will be attempted hold true.
 584  */
 585 static void efx_tso_check_safe(struct sk_buff *skb)
 586 {
 587         __be16 protocol = skb->protocol;
 588
 589         EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
 590                             protocol);
 591         if (protocol == htons(ETH_P_8021Q)) {
 592                 /* Find the encapsulated protocol; reset network header
 593                  * and transport header based on that. */
 594                 struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
 595                 protocol = veh->h_vlan_encapsulated_proto;
 596                 skb_set_network_header(skb, sizeof(*veh));
 597                 if (protocol == htons(ETH_P_IP))
 598                         skb_set_transport_header(skb, sizeof(*veh) +
 599                                                  4 * ip_hdr(skb)->ihl);
 600         }
 601
 602         EFX_BUG_ON_PARANOID(protocol != htons(ETH_P_IP));
 603         EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
 604         EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
 605                              + (tcp_hdr(skb)->doff << 2u)) >
 606                             skb_headlen(skb));
 607 }
 608
 609
 610 /*
 611  * Allocate a page worth of efx_tso_header structures, and string them
 612  * into the tx_queue->tso_headers_free linked list. Return 0 or -ENOMEM.
 613  */
 614 static int efx_tsoh_block_alloc(struct efx_tx_queue *tx_queue)
 615 {
 616
 617         struct pci_dev *pci_dev = tx_queue->efx->pci_dev;
 618         struct efx_tso_header *tsoh;
 619         dma_addr_t dma_addr;
 620         u8 *base_kva, *kva;
 621
 622         base_kva = pci_alloc_consistent(pci_dev, PAGE_SIZE, &dma_addr);
 623         if (base_kva == NULL) {
 624                 EFX_ERR(tx_queue->efx, "Unable to allocate page for TSO"
 625                         " headers\n");
 626                 return -ENOMEM;
 627         }
 628
 629         /* pci_alloc_consistent() allocates pages. */
 630         EFX_BUG_ON_PARANOID(dma_addr & (PAGE_SIZE - 1u));
 631
 632         for (kva = base_kva; kva < base_kva + PAGE_SIZE; kva += TSOH_STD_SIZE) {
 633                 tsoh = (struct efx_tso_header *)kva;
 634                 tsoh->dma_addr = dma_addr + (TSOH_BUFFER(tsoh) - base_kva);
 635                 tsoh->next = tx_queue->tso_headers_free;
 636                 tx_queue->tso_headers_free = tsoh;
 637         }
 638
 639         return 0;
 640 }
 641
 642
 643 /* Free up a TSO header, and all others in the same page. */
 644 static void efx_tsoh_block_free(struct efx_tx_queue *tx_queue,
 645                                 struct efx_tso_header *tsoh,
 646                                 struct pci_dev *pci_dev)
 647 {
 648         struct efx_tso_header **p;
 649         unsigned long base_kva;
 650         dma_addr_t base_dma;
 651
 652         base_kva = (unsigned long)tsoh & PAGE_MASK;
 653         base_dma = tsoh->dma_addr & PAGE_MASK;
 654
 655         p = &tx_queue->tso_headers_free;
 656         while (*p != NULL) {
 657                 if (((unsigned long)*p & PAGE_MASK) == base_kva)
 658                         *p = (*p)->next;
 659                 else
 660                         p = &(*p)->next;
 661         }
 662
 663         pci_free_consistent(pci_dev, PAGE_SIZE, (void *)base_kva, base_dma);
 664 }
 665
 666 static struct efx_tso_header *
 667 efx_tsoh_heap_alloc(struct efx_tx_queue *tx_queue, size_t header_len)
 668 {
 669         struct efx_tso_header *tsoh;
 670
 671         tsoh = kmalloc(TSOH_SIZE(header_len), GFP_ATOMIC | GFP_DMA);
 672         if (unlikely(!tsoh))
 673                 return NULL;
 674
 675         tsoh->dma_addr = pci_map_single(tx_queue->efx->pci_dev,
 676                                         TSOH_BUFFER(tsoh), header_len,
 677                                         PCI_DMA_TODEVICE);
 678         if (unlikely(pci_dma_mapping_error(tx_queue->efx->pci_dev,
 679                                            tsoh->dma_addr))) {
 680                 kfree(tsoh);
 681                 return NULL;
 682         }
 683
 684         tsoh->unmap_len = header_len;
 685         return tsoh;
 686 }
 687
 688 static void
 689 efx_tsoh_heap_free(struct efx_tx_queue *tx_queue, struct efx_tso_header *tsoh)
 690 {
 691         pci_unmap_single(tx_queue->efx->pci_dev,
 692                          tsoh->dma_addr, tsoh->unmap_len,
 693                          PCI_DMA_TODEVICE);
 694         kfree(tsoh);
 695 }
 696
 697 /**
 698  * efx_tx_queue_insert - push descriptors onto the TX queue
 699  * @tx_queue:           Efx TX queue
 700  * @dma_addr:           DMA address of fragment
 701  * @len:                Length of fragment
 702  * @final_buffer:       The final buffer inserted into the queue
 703  *
 704  * Push descriptors onto the TX queue.  Return 0 on success or 1 if
 705  * @tx_queue full.
 706  */
 707 static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
 708                                dma_addr_t dma_addr, unsigned len,
 709                                struct efx_tx_buffer **final_buffer)
 710 {
 711         struct efx_tx_buffer *buffer;
 712         struct efx_nic *efx = tx_queue->efx;
 713         unsigned dma_len, fill_level, insert_ptr, misalign;
 714         int q_space;
 715
 716         EFX_BUG_ON_PARANOID(len <= 0);
 717
 718         fill_level = tx_queue->insert_count - tx_queue->old_read_count;
 719         /* -1 as there is no way to represent all descriptors used */
 720         q_space = efx->type->txd_ring_mask - 1 - fill_level;
 721
 722         while (1) {
 723                 if (unlikely(q_space-- <= 0)) {
 724                         /* It might be that completions have happened
 725                          * since the xmit path last checked.  Update
 726                          * the xmit path's copy of read_count.
 727                          */
 728                         ++tx_queue->stopped;
 729                         /* This memory barrier protects the change of
 730                          * stopped from the access of read_count. */
 731                         smp_mb();
 732                         tx_queue->old_read_count =
 733                                 *(volatile unsigned *)&tx_queue->read_count;
 734                         fill_level = (tx_queue->insert_count
 735                                       - tx_queue->old_read_count);
 736                         q_space = efx->type->txd_ring_mask - 1 - fill_level;
 737                         if (unlikely(q_space-- <= 0)) {
 738                                 *final_buffer = NULL;
 739                                 return 1;
 740                         }
 741                         smp_mb();
 742                         --tx_queue->stopped;
 743                 }
 744
 745                 insert_ptr = tx_queue->insert_count & efx->type->txd_ring_mask;
 746                 buffer = &tx_queue->buffer[insert_ptr];
 747                 ++tx_queue->insert_count;
 748
 749                 EFX_BUG_ON_PARANOID(tx_queue->insert_count -
 750                                     tx_queue->read_count >
 751                                     efx->type->txd_ring_mask);
 752
 753                 efx_tsoh_free(tx_queue, buffer);
 754                 EFX_BUG_ON_PARANOID(buffer->len);
 755                 EFX_BUG_ON_PARANOID(buffer->unmap_len);
 756                 EFX_BUG_ON_PARANOID(buffer->skb);
 757                 EFX_BUG_ON_PARANOID(!buffer->continuation);
 758                 EFX_BUG_ON_PARANOID(buffer->tsoh);
 759
 760                 buffer->dma_addr = dma_addr;
 761
 762                 /* Ensure we do not cross a boundary unsupported by H/W */
 763                 dma_len = (~dma_addr & efx->type->tx_dma_mask) + 1;
 764
 765                 misalign = (unsigned)dma_addr & efx->type->bug5391_mask;
 766                 if (misalign && dma_len + misalign > 512)
 767                         dma_len = 512 - misalign;
 768
 769                 /* If there is enough space to send then do so */
 770                 if (dma_len >= len)
 771                         break;
 772
 773                 buffer->len = dma_len; /* Don't set the other members */
 774                 dma_addr += dma_len;
 775                 len -= dma_len;
 776         }
 777
 778         EFX_BUG_ON_PARANOID(!len);
 779         buffer->len = len;
 780         *final_buffer = buffer;
 781         return 0;
 782 }
 783
 784
 785 /*
 786  * Put a TSO header into the TX queue.
 787  *
 788  * This is special-cased because we know that it is small enough to fit in
 789  * a single fragment, and we know it doesn't cross a page boundary.  It
 790  * also allows us to not worry about end-of-packet etc.
 791  */
 792 static void efx_tso_put_header(struct efx_tx_queue *tx_queue,
 793                                struct efx_tso_header *tsoh, unsigned len)
 794 {
 795         struct efx_tx_buffer *buffer;
 796
 797         buffer = &tx_queue->buffer[tx_queue->insert_count &
 798                                    tx_queue->efx->type->txd_ring_mask];
 799         efx_tsoh_free(tx_queue, buffer);
 800         EFX_BUG_ON_PARANOID(buffer->len);
 801         EFX_BUG_ON_PARANOID(buffer->unmap_len);
 802         EFX_BUG_ON_PARANOID(buffer->skb);
 803         EFX_BUG_ON_PARANOID(!buffer->continuation);
 804         EFX_BUG_ON_PARANOID(buffer->tsoh);
 805         buffer->len = len;
 806         buffer->dma_addr = tsoh->dma_addr;
 807         buffer->tsoh = tsoh;
 808
 809         ++tx_queue->insert_count;
 810 }
 811
 812
 813 /* Remove descriptors put into a tx_queue. */
 814 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
 815 {
 816         struct efx_tx_buffer *buffer;
 817         dma_addr_t unmap_addr;
 818
 819         /* Work backwards until we hit the original insert pointer value */
 820         while (tx_queue->insert_count != tx_queue->write_count) {
 821                 --tx_queue->insert_count;
 822                 buffer = &tx_queue->buffer[tx_queue->insert_count &
 823                                            tx_queue->efx->type->txd_ring_mask];
 824                 efx_tsoh_free(tx_queue, buffer);
 825                 EFX_BUG_ON_PARANOID(buffer->skb);
 826                 buffer->len = 0;
 827                 buffer->continuation = true;
 828                 if (buffer->unmap_len) {
 829                         unmap_addr = (buffer->dma_addr + buffer->len -
 830                                       buffer->unmap_len);
 831                         if (buffer->unmap_single)
 832                                 pci_unmap_single(tx_queue->efx->pci_dev,
 833                                                  unmap_addr, buffer->unmap_len,
 834                                                  PCI_DMA_TODEVICE);
 835                         else
 836                                 pci_unmap_page(tx_queue->efx->pci_dev,
 837                                                unmap_addr, buffer->unmap_len,
 838                                                PCI_DMA_TODEVICE);
 839                         buffer->unmap_len = 0;
 840                 }
 841         }
 842 }
 843
 844
 845 /* Parse the SKB header and initialise state. */
 846 static void tso_start(struct tso_state *st, const struct sk_buff *skb)
 847 {
 848         /* All ethernet/IP/TCP headers combined size is TCP header size
 849          * plus offset of TCP header relative to start of packet.
 850          */
 851         st->header_len = ((tcp_hdr(skb)->doff << 2u)
 852                           + PTR_DIFF(tcp_hdr(skb), skb->data));
 853         st->full_packet_size = st->header_len + skb_shinfo(skb)->gso_size;
 854
 855         st->ipv4_id = ntohs(ip_hdr(skb)->id);
 856         st->seqnum = ntohl(tcp_hdr(skb)->seq);
 857
 858         EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
 859         EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
 860         EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
 861
 862         st->packet_space = st->full_packet_size;
 863         st->out_len = skb->len - st->header_len;
 864         st->unmap_len = 0;
 865         st->unmap_single = false;
 866 }
 867
 868 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
 869                             skb_frag_t *frag)
 870 {
 871         st->unmap_addr = pci_map_page(efx->pci_dev, frag->page,
 872                                       frag->page_offset, frag->size,
 873                                       PCI_DMA_TODEVICE);
 874         if (likely(!pci_dma_mapping_error(efx->pci_dev, st->unmap_addr))) {
 875                 st->unmap_single = false;
 876                 st->unmap_len = frag->size;
 877                 st->in_len = frag->size;
 878                 st->dma_addr = st->unmap_addr;
 879                 return 0;
 880         }
 881         return -ENOMEM;
 882 }
 883
 884 static int tso_get_head_fragment(struct tso_state *st, struct efx_nic *efx,
 885                                  const struct sk_buff *skb)
 886 {
 887         int hl = st->header_len;
 888         int len = skb_headlen(skb) - hl;
 889
 890         st->unmap_addr = pci_map_single(efx->pci_dev, skb->data + hl,
 891                                         len, PCI_DMA_TODEVICE);
 892         if (likely(!pci_dma_mapping_error(efx->pci_dev, st->unmap_addr))) {
 893                 st->unmap_single = true;
 894                 st->unmap_len = len;
 895                 st->in_len = len;
 896                 st->dma_addr = st->unmap_addr;
 897                 return 0;
 898         }
 899         return -ENOMEM;
 900 }
 901
 902
 903 /**
 904  * tso_fill_packet_with_fragment - form descriptors for the current fragment
 905  * @tx_queue:           Efx TX queue
 906  * @skb:                Socket buffer
 907  * @st:                 TSO state
 908  *
 909  * Form descriptors for the current fragment, until we reach the end
 910  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
 911  * space in @tx_queue.
 912  */
 913 static int tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
 914                                          const struct sk_buff *skb,
 915                                          struct tso_state *st)
 916 {
 917         struct efx_tx_buffer *buffer;
 918         int n, end_of_packet, rc;
 919
 920         if (st->in_len == 0)
 921                 return 0;
 922         if (st->packet_space == 0)
 923                 return 0;
 924
 925         EFX_BUG_ON_PARANOID(st->in_len <= 0);
 926         EFX_BUG_ON_PARANOID(st->packet_space <= 0);
 927
 928         n = min(st->in_len, st->packet_space);
 929
 930         st->packet_space -= n;
 931         st->out_len -= n;
 932         st->in_len -= n;
 933
 934         rc = efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer);
 935         if (likely(rc == 0)) {
 936                 if (st->out_len == 0)
 937                         /* Transfer ownership of the skb */
 938                         buffer->skb = skb;
 939
 940                 end_of_packet = st->out_len == 0 || st->packet_space == 0;
 941                 buffer->continuation = !end_of_packet;
 942
 943                 if (st->in_len == 0) {
 944                         /* Transfer ownership of the pci mapping */
 945                         buffer->unmap_len = st->unmap_len;
 946                         buffer->unmap_single = st->unmap_single;
 947                         st->unmap_len = 0;
 948                 }
 949         }
 950
 951         st->dma_addr += n;
 952         return rc;
 953 }
 954
 955
 956 /**
 957  * tso_start_new_packet - generate a new header and prepare for the new packet
 958  * @tx_queue:           Efx TX queue
 959  * @skb:                Socket buffer
 960  * @st:                 TSO state
 961  *
 962  * Generate a new header and prepare for the new packet.  Return 0 on
 963  * success, or -1 if failed to alloc header.
 964  */
 965 static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
 966                                 const struct sk_buff *skb,
 967                                 struct tso_state *st)
 968 {
 969         struct efx_tso_header *tsoh;
 970         struct iphdr *tsoh_iph;
 971         struct tcphdr *tsoh_th;
 972         unsigned ip_length;
 973         u8 *header;
 974
 975         /* Allocate a DMA-mapped header buffer. */
 976         if (likely(TSOH_SIZE(st->header_len) <= TSOH_STD_SIZE)) {
 977                 if (tx_queue->tso_headers_free == NULL) {
 978                         if (efx_tsoh_block_alloc(tx_queue))
 979                                 return -1;
 980                 }
 981                 EFX_BUG_ON_PARANOID(!tx_queue->tso_headers_free);
 982                 tsoh = tx_queue->tso_headers_free;
 983                 tx_queue->tso_headers_free = tsoh->next;
 984                 tsoh->unmap_len = 0;
 985         } else {
 986                 tx_queue->tso_long_headers++;
 987                 tsoh = efx_tsoh_heap_alloc(tx_queue, st->header_len);
 988                 if (unlikely(!tsoh))
 989                         return -1;
 990         }
 991
 992         header = TSOH_BUFFER(tsoh);
 993         tsoh_th = (struct tcphdr *)(header + SKB_TCP_OFF(skb));
 994         tsoh_iph = (struct iphdr *)(header + SKB_IPV4_OFF(skb));
 995
 996         /* Copy and update the headers. */
 997         memcpy(header, skb->data, st->header_len);
 998
 999         tsoh_th->seq = htonl(st->seqnum);
1000         st->seqnum += skb_shinfo(skb)->gso_size;
1001         if (st->out_len > skb_shinfo(skb)->gso_size) {
1002                 /* This packet will not finish the TSO burst. */
1003                 ip_length = st->full_packet_size - ETH_HDR_LEN(skb);
1004                 tsoh_th->fin = 0;
1005                 tsoh_th->psh = 0;
1006         } else {
1007                 /* This packet will be the last in the TSO burst. */
1008                 ip_length = st->header_len - ETH_HDR_LEN(skb) + st->out_len;
1009                 tsoh_th->fin = tcp_hdr(skb)->fin;
1010                 tsoh_th->psh = tcp_hdr(skb)->psh;
1011         }
1012         tsoh_iph->tot_len = htons(ip_length);
1013
1014         /* Linux leaves suitable gaps in the IP ID space for us to fill. */
1015         tsoh_iph->id = htons(st->ipv4_id);
1016         st->ipv4_id++;
1017
1018         st->packet_space = skb_shinfo(skb)->gso_size;
1019         ++tx_queue->tso_packets;
1020
1021         /* Form a descriptor for this header. */
1022         efx_tso_put_header(tx_queue, tsoh, st->header_len);
1023
1024         return 0;
1025 }
1026
1027
1028 /**
1029  * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1030  * @tx_queue:           Efx TX queue
1031  * @skb:                Socket buffer
1032  *
1033  * Context: You must hold netif_tx_lock() to call this function.
1034  *
1035  * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1036  * @skb was not enqueued.  In all cases @skb is consumed.  Return
1037  * %NETDEV_TX_OK or %NETDEV_TX_BUSY.
1038  */
1039 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1040                                struct sk_buff *skb)
1041 {
1042         struct efx_nic *efx = tx_queue->efx;
1043         int frag_i, rc, rc2 = NETDEV_TX_OK;
1044         struct tso_state state;
1045
1046         /* Verify TSO is safe - these checks should never fail. */
1047         efx_tso_check_safe(skb);
1048
1049         EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
1050
1051         tso_start(&state, skb);
1052
1053         /* Assume that skb header area contains exactly the headers, and
1054          * all payload is in the frag list.
1055          */
1056         if (skb_headlen(skb) == state.header_len) {
1057                 /* Grab the first payload fragment. */
1058                 EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1059                 frag_i = 0;
1060                 rc = tso_get_fragment(&state, efx,
1061                                       skb_shinfo(skb)->frags + frag_i);
1062                 if (rc)
1063                         goto mem_err;
1064         } else {
1065                 rc = tso_get_head_fragment(&state, efx, skb);
1066                 if (rc)
1067                         goto mem_err;
1068                 frag_i = -1;
1069         }
1070
1071         if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1072                 goto mem_err;
1073
1074         while (1) {
1075                 rc = tso_fill_packet_with_fragment(tx_queue, skb, &state);
1076                 if (unlikely(rc))
1077                         goto stop;
1078
1079                 /* Move onto the next fragment? */
1080                 if (state.in_len == 0) {
1081                         if (++frag_i >= skb_shinfo(skb)->nr_frags)
1082                                 /* End of payload reached. */
1083                                 break;
1084                         rc = tso_get_fragment(&state, efx,
1085                                               skb_shinfo(skb)->frags + frag_i);
1086                         if (rc)
1087                                 goto mem_err;
1088                 }
1089
1090                 /* Start at new packet? */
1091                 if (state.packet_space == 0 &&
1092                     tso_start_new_packet(tx_queue, skb, &state) < 0)
1093                         goto mem_err;
1094         }
1095
1096         /* Pass off to hardware */
1097         falcon_push_buffers(tx_queue);
1098
1099         tx_queue->tso_bursts++;
1100         return NETDEV_TX_OK;
1101
1102  mem_err:
1103         EFX_ERR(efx, "Out of memory for TSO headers, or PCI mapping error\n");
1104         dev_kfree_skb_any((struct sk_buff *)skb);
1105         goto unwind;
1106
1107  stop:
1108         rc2 = NETDEV_TX_BUSY;
1109
1110         /* Stop the queue if it wasn't stopped before. */
1111         if (tx_queue->stopped == 1)
1112                 efx_stop_queue(efx);
1113
1114  unwind:
1115         /* Free the DMA mapping we were in the process of writing out */
1116         if (state.unmap_len) {
1117                 if (state.unmap_single)
1118                         pci_unmap_single(efx->pci_dev, state.unmap_addr,
1119                                          state.unmap_len, PCI_DMA_TODEVICE);
1120                 else
1121                         pci_unmap_page(efx->pci_dev, state.unmap_addr,
1122                                        state.unmap_len, PCI_DMA_TODEVICE);
1123         }
1124
1125         efx_enqueue_unwind(tx_queue);
1126         return rc2;
1127 }
1128
1129
1130 /*
1131  * Free up all TSO datastructures associated with tx_queue. This
1132  * routine should be called only once the tx_queue is both empty and
1133  * will no longer be used.
1134  */
1135 static void efx_fini_tso(struct efx_tx_queue *tx_queue)
1136 {
1137         unsigned i;
1138
1139         if (tx_queue->buffer) {
1140                 for (i = 0; i <= tx_queue->efx->type->txd_ring_mask; ++i)
1141                         efx_tsoh_free(tx_queue, &tx_queue->buffer[i]);
1142         }
1143
1144         while (tx_queue->tso_headers_free != NULL)
1145                 efx_tsoh_block_free(tx_queue, tx_queue->tso_headers_free,
1146                                     tx_queue->efx->pci_dev);
1147 }