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