git.oblomov.eu Git - linux-2.6/blob - 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         fill_level = tx_queue->insert_count - tx_queue->old_read_count;
 166         q_space = efx->type->txd_ring_mask - 1 - fill_level;
 167
 168         /* Map for DMA.  Use pci_map_single rather than pci_map_page
 169          * since this is more efficient on machines with sparse
 170          * memory.
 171          */
 172         unmap_single = true;
 173         dma_addr = pci_map_single(pci_dev, skb->data, len, PCI_DMA_TODEVICE);
 174
 175         /* Process all fragments */
 176         while (1) {
 177                 if (unlikely(pci_dma_mapping_error(pci_dev, dma_addr)))
 178                         goto pci_err;
 179
 180                 /* Store fields for marking in the per-fragment final
 181                  * descriptor */
 182                 unmap_len = len;
 183                 unmap_addr = dma_addr;
 184
 185                 /* Add to TX queue, splitting across DMA boundaries */
 186                 do {
 187                         if (unlikely(q_space-- <= 0)) {
 188                                 /* It might be that completions have
 189                                  * happened since the xmit path last
 190                                  * checked.  Update the xmit path's
 191                                  * copy of read_count.
 192                                  */
 193                                 ++tx_queue->stopped;
 194                                 /* This memory barrier protects the
 195                                  * change of stopped from the access
 196                                  * of read_count. */
 197                                 smp_mb();
 198                                 tx_queue->old_read_count =
 199                                         *(volatile unsigned *)
 200                                         &tx_queue->read_count;
 201                                 fill_level = (tx_queue->insert_count
 202                                               - tx_queue->old_read_count);
 203                                 q_space = (efx->type->txd_ring_mask - 1 -
 204                                            fill_level);
 205                                 if (unlikely(q_space-- <= 0))
 206                                         goto stop;
 207                                 smp_mb();
 208                                 --tx_queue->stopped;
 209                         }
 210
 211                         insert_ptr = (tx_queue->insert_count &
 212                                       efx->type->txd_ring_mask);
 213                         buffer = &tx_queue->buffer[insert_ptr];
 214                         efx_tsoh_free(tx_queue, buffer);
 215                         EFX_BUG_ON_PARANOID(buffer->tsoh);
 216                         EFX_BUG_ON_PARANOID(buffer->skb);
 217                         EFX_BUG_ON_PARANOID(buffer->len);
 218                         EFX_BUG_ON_PARANOID(!buffer->continuation);
 219                         EFX_BUG_ON_PARANOID(buffer->unmap_len);
 220
 221                         dma_len = (((~dma_addr) & efx->type->tx_dma_mask) + 1);
 222                         if (likely(dma_len > len))
 223                                 dma_len = len;
 224
 225                         misalign = (unsigned)dma_addr & efx->type->bug5391_mask;
 226                         if (misalign && dma_len + misalign > 512)
 227                                 dma_len = 512 - misalign;
 228
 229                         /* Fill out per descriptor fields */
 230                         buffer->len = dma_len;
 231                         buffer->dma_addr = dma_addr;
 232                         len -= dma_len;
 233                         dma_addr += dma_len;
 234                         ++tx_queue->insert_count;
 235                 } while (len);
 236
 237                 /* Transfer ownership of the unmapping to the final buffer */
 238                 buffer->unmap_single = unmap_single;
 239                 buffer->unmap_len = unmap_len;
 240                 unmap_len = 0;
 241
 242                 /* Get address and size of next fragment */
 243                 if (i >= skb_shinfo(skb)->nr_frags)
 244                         break;
 245                 fragment = &skb_shinfo(skb)->frags[i];
 246                 len = fragment->size;
 247                 page = fragment->page;
 248                 page_offset = fragment->page_offset;
 249                 i++;
 250                 /* Map for DMA */
 251                 unmap_single = false;
 252                 dma_addr = pci_map_page(pci_dev, page, page_offset, len,
 253                                         PCI_DMA_TODEVICE);
 254         }
 255
 256         /* Transfer ownership of the skb to the final buffer */
 257         buffer->skb = skb;
 258         buffer->continuation = false;
 259
 260         /* Pass off to hardware */
 261         falcon_push_buffers(tx_queue);
 262
 263         return NETDEV_TX_OK;
 264
 265  pci_err:
 266         EFX_ERR_RL(efx, " TX queue %d could not map skb with %d bytes %d "
 267                    "fragments for DMA\n", tx_queue->queue, skb->len,
 268                    skb_shinfo(skb)->nr_frags + 1);
 269
 270         /* Mark the packet as transmitted, and free the SKB ourselves */
 271         dev_kfree_skb_any((struct sk_buff *)skb);
 272         goto unwind;
 273
 274  stop:
 275         rc = NETDEV_TX_BUSY;
 276
 277         if (tx_queue->stopped == 1)
 278                 efx_stop_queue(efx);
 279
 280  unwind:
 281         /* Work backwards until we hit the original insert pointer value */
 282         while (tx_queue->insert_count != tx_queue->write_count) {
 283                 --tx_queue->insert_count;
 284                 insert_ptr = tx_queue->insert_count & efx->type->txd_ring_mask;
 285                 buffer = &tx_queue->buffer[insert_ptr];
 286                 efx_dequeue_buffer(tx_queue, buffer);
 287                 buffer->len = 0;
 288         }
 289
 290         /* Free the fragment we were mid-way through pushing */
 291         if (unmap_len) {
 292                 if (unmap_single)
 293                         pci_unmap_single(pci_dev, unmap_addr, unmap_len,
 294                                          PCI_DMA_TODEVICE);
 295                 else
 296                         pci_unmap_page(pci_dev, unmap_addr, unmap_len,
 297                                        PCI_DMA_TODEVICE);
 298         }
 299
 300         return rc;
 301 }
 302
 303 /* Remove packets from the TX queue
 304  *
 305  * This removes packets from the TX queue, up to and including the
 306  * specified index.
 307  */
 308 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
 309                                 unsigned int index)
 310 {
 311         struct efx_nic *efx = tx_queue->efx;
 312         unsigned int stop_index, read_ptr;
 313         unsigned int mask = tx_queue->efx->type->txd_ring_mask;
 314
 315         stop_index = (index + 1) & mask;
 316         read_ptr = tx_queue->read_count & mask;
 317
 318         while (read_ptr != stop_index) {
 319                 struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
 320                 if (unlikely(buffer->len == 0)) {
 321                         EFX_ERR(tx_queue->efx, "TX queue %d spurious TX "
 322                                 "completion id %x\n", tx_queue->queue,
 323                                 read_ptr);
 324                         efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
 325                         return;
 326                 }
 327
 328                 efx_dequeue_buffer(tx_queue, buffer);
 329                 buffer->continuation = true;
 330                 buffer->len = 0;
 331
 332                 ++tx_queue->read_count;
 333                 read_ptr = tx_queue->read_count & mask;
 334         }
 335 }
 336
 337 /* Initiate a packet transmission on the specified TX queue.
 338  * Note that returning anything other than NETDEV_TX_OK will cause the
 339  * OS to free the skb.
 340  *
 341  * This function is split out from efx_hard_start_xmit to allow the
 342  * loopback test to direct packets via specific TX queues.  It is
 343  * therefore a non-static inline, so as not to penalise performance
 344  * for non-loopback transmissions.
 345  *
 346  * Context: netif_tx_lock held
 347  */
 348 inline int efx_xmit(struct efx_nic *efx,
 349                     struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 350 {
 351         int rc;
 352
 353         /* Map fragments for DMA and add to TX queue */
 354         rc = efx_enqueue_skb(tx_queue, skb);
 355         if (unlikely(rc != NETDEV_TX_OK))
 356                 goto out;
 357
 358         /* Update last TX timer */
 359         efx->net_dev->trans_start = jiffies;
 360
 361  out:
 362         return rc;
 363 }
 364
 365 /* Initiate a packet transmission.  We use one channel per CPU
 366  * (sharing when we have more CPUs than channels).  On Falcon, the TX
 367  * completion events will be directed back to the CPU that transmitted
 368  * the packet, which should be cache-efficient.
 369  *
 370  * Context: non-blocking.
 371  * Note that returning anything other than NETDEV_TX_OK will cause the
 372  * OS to free the skb.
 373  */
 374 int efx_hard_start_xmit(struct sk_buff *skb, struct net_device *net_dev)
 375 {
 376         struct efx_nic *efx = netdev_priv(net_dev);
 377         struct efx_tx_queue *tx_queue;
 378
 379         if (unlikely(efx->port_inhibited))
 380                 return NETDEV_TX_BUSY;
 381
 382         if (likely(skb->ip_summed == CHECKSUM_PARTIAL))
 383                 tx_queue = &efx->tx_queue[EFX_TX_QUEUE_OFFLOAD_CSUM];
 384         else
 385                 tx_queue = &efx->tx_queue[EFX_TX_QUEUE_NO_CSUM];
 386
 387         return efx_xmit(efx, tx_queue, skb);
 388 }
 389
 390 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 391 {
 392         unsigned fill_level;
 393         struct efx_nic *efx = tx_queue->efx;
 394
 395         EFX_BUG_ON_PARANOID(index > efx->type->txd_ring_mask);
 396
 397         efx_dequeue_buffers(tx_queue, index);
 398
 399         /* See if we need to restart the netif queue.  This barrier
 400          * separates the update of read_count from the test of
 401          * stopped. */
 402         smp_mb();
 403         if (unlikely(tx_queue->stopped) && likely(efx->port_enabled)) {
 404                 fill_level = tx_queue->insert_count - tx_queue->read_count;
 405                 if (fill_level < EFX_NETDEV_TX_THRESHOLD(tx_queue)) {
 406                         EFX_BUG_ON_PARANOID(!efx_dev_registered(efx));
 407
 408                         /* Do this under netif_tx_lock(), to avoid racing
 409                          * with efx_xmit(). */
 410                         netif_tx_lock(efx->net_dev);
 411                         if (tx_queue->stopped) {
 412                                 tx_queue->stopped = 0;
 413                                 efx_wake_queue(efx);
 414                         }
 415                         netif_tx_unlock(efx->net_dev);
 416                 }
 417         }
 418 }
 419
 420 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
 421 {
 422         struct efx_nic *efx = tx_queue->efx;
 423         unsigned int txq_size;
 424         int i, rc;
 425
 426         EFX_LOG(efx, "creating TX queue %d\n", tx_queue->queue);
 427
 428         /* Allocate software ring */
 429         txq_size = (efx->type->txd_ring_mask + 1) * sizeof(*tx_queue->buffer);
 430         tx_queue->buffer = kzalloc(txq_size, GFP_KERNEL);
 431         if (!tx_queue->buffer)
 432                 return -ENOMEM;
 433         for (i = 0; i <= efx->type->txd_ring_mask; ++i)
 434                 tx_queue->buffer[i].continuation = true;
 435
 436         /* Allocate hardware ring */
 437         rc = falcon_probe_tx(tx_queue);
 438         if (rc)
 439                 goto fail;
 440
 441         return 0;
 442
 443  fail:
 444         kfree(tx_queue->buffer);
 445         tx_queue->buffer = NULL;
 446         return rc;
 447 }
 448
 449 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
 450 {
 451         EFX_LOG(tx_queue->efx, "initialising TX queue %d\n", tx_queue->queue);
 452
 453         tx_queue->insert_count = 0;
 454         tx_queue->write_count = 0;
 455         tx_queue->read_count = 0;
 456         tx_queue->old_read_count = 0;
 457         BUG_ON(tx_queue->stopped);
 458
 459         /* Set up TX descriptor ring */
 460         falcon_init_tx(tx_queue);
 461 }
 462
 463 void efx_release_tx_buffers(struct efx_tx_queue *tx_queue)
 464 {
 465         struct efx_tx_buffer *buffer;
 466
 467         if (!tx_queue->buffer)
 468                 return;
 469
 470         /* Free any buffers left in the ring */
 471         while (tx_queue->read_count != tx_queue->write_count) {
 472                 buffer = &tx_queue->buffer[tx_queue->read_count &
 473                                            tx_queue->efx->type->txd_ring_mask];
 474                 efx_dequeue_buffer(tx_queue, buffer);
 475                 buffer->continuation = true;
 476                 buffer->len = 0;
 477
 478                 ++tx_queue->read_count;
 479         }
 480 }
 481
 482 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
 483 {
 484         EFX_LOG(tx_queue->efx, "shutting down TX queue %d\n", tx_queue->queue);
 485
 486         /* Flush TX queue, remove descriptor ring */
 487         falcon_fini_tx(tx_queue);
 488
 489         efx_release_tx_buffers(tx_queue);
 490
 491         /* Free up TSO header cache */
 492         efx_fini_tso(tx_queue);
 493
 494         /* Release queue's stop on port, if any */
 495         if (tx_queue->stopped) {
 496                 tx_queue->stopped = 0;
 497                 efx_wake_queue(tx_queue->efx);
 498         }
 499 }
 500
 501 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
 502 {
 503         EFX_LOG(tx_queue->efx, "destroying TX queue %d\n", tx_queue->queue);
 504         falcon_remove_tx(tx_queue);
 505
 506         kfree(tx_queue->buffer);
 507         tx_queue->buffer = NULL;
 508 }
 509
 510
 511 /* Efx TCP segmentation acceleration.
 512  *
 513  * Why?  Because by doing it here in the driver we can go significantly
 514  * faster than the GSO.
 515  *
 516  * Requires TX checksum offload support.
 517  */
 518
 519 /* Number of bytes inserted at the start of a TSO header buffer,
 520  * similar to NET_IP_ALIGN.
 521  */
 522 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
 523 #define TSOH_OFFSET     0
 524 #else
 525 #define TSOH_OFFSET     NET_IP_ALIGN
 526 #endif
 527
 528 #define TSOH_BUFFER(tsoh)       ((u8 *)(tsoh + 1) + TSOH_OFFSET)
 529
 530 /* Total size of struct efx_tso_header, buffer and padding */
 531 #define TSOH_SIZE(hdr_len)                                      \
 532         (sizeof(struct efx_tso_header) + TSOH_OFFSET + hdr_len)
 533
 534 /* Size of blocks on free list.  Larger blocks must be allocated from
 535  * the heap.
 536  */
 537 #define TSOH_STD_SIZE           128
 538
 539 #define PTR_DIFF(p1, p2)  ((u8 *)(p1) - (u8 *)(p2))
 540 #define ETH_HDR_LEN(skb)  (skb_network_header(skb) - (skb)->data)
 541 #define SKB_TCP_OFF(skb)  PTR_DIFF(tcp_hdr(skb), (skb)->data)
 542 #define SKB_IPV4_OFF(skb) PTR_DIFF(ip_hdr(skb), (skb)->data)
 543
 544 /**
 545  * struct tso_state - TSO state for an SKB
 546  * @out_len: Remaining length in current segment
 547  * @seqnum: Current sequence number
 548  * @ipv4_id: Current IPv4 ID, host endian
 549  * @packet_space: Remaining space in current packet
 550  * @dma_addr: DMA address of current position
 551  * @in_len: Remaining length in current SKB fragment
 552  * @unmap_len: Length of SKB fragment
 553  * @unmap_addr: DMA address of SKB fragment
 554  * @unmap_single: DMA single vs page mapping flag
 555  * @header_len: Number of bytes of header
 556  * @full_packet_size: Number of bytes to put in each outgoing segment
 557  *
 558  * The state used during segmentation.  It is put into this data structure
 559  * just to make it easy to pass into inline functions.
 560  */
 561 struct tso_state {
 562         /* Output position */
 563         unsigned out_len;
 564         unsigned seqnum;
 565         unsigned ipv4_id;
 566         unsigned packet_space;
 567
 568         /* Input position */
 569         dma_addr_t dma_addr;
 570         unsigned in_len;
 571         unsigned unmap_len;
 572         dma_addr_t unmap_addr;
 573         bool unmap_single;
 574
 575         unsigned header_len;
 576         int full_packet_size;
 577 };
 578
 579
 580 /*
 581  * Verify that our various assumptions about sk_buffs and the conditions
 582  * under which TSO will be attempted hold true.
 583  */
 584 static void efx_tso_check_safe(struct sk_buff *skb)
 585 {
 586         __be16 protocol = skb->protocol;
 587
 588         EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
 589                             protocol);
 590         if (protocol == htons(ETH_P_8021Q)) {
 591                 /* Find the encapsulated protocol; reset network header
 592                  * and transport header based on that. */
 593                 struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
 594                 protocol = veh->h_vlan_encapsulated_proto;
 595                 skb_set_network_header(skb, sizeof(*veh));
 596                 if (protocol == htons(ETH_P_IP))
 597                         skb_set_transport_header(skb, sizeof(*veh) +
 598                                                  4 * ip_hdr(skb)->ihl);
 599         }
 600
 601         EFX_BUG_ON_PARANOID(protocol != htons(ETH_P_IP));
 602         EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
 603         EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
 604                              + (tcp_hdr(skb)->doff << 2u)) >
 605                             skb_headlen(skb));
 606 }
 607
 608
 609 /*
 610  * Allocate a page worth of efx_tso_header structures, and string them
 611  * into the tx_queue->tso_headers_free linked list. Return 0 or -ENOMEM.
 612  */
 613 static int efx_tsoh_block_alloc(struct efx_tx_queue *tx_queue)
 614 {
 615
 616         struct pci_dev *pci_dev = tx_queue->efx->pci_dev;
 617         struct efx_tso_header *tsoh;
 618         dma_addr_t dma_addr;
 619         u8 *base_kva, *kva;
 620
 621         base_kva = pci_alloc_consistent(pci_dev, PAGE_SIZE, &dma_addr);
 622         if (base_kva == NULL) {
 623                 EFX_ERR(tx_queue->efx, "Unable to allocate page for TSO"
 624                         " headers\n");
 625                 return -ENOMEM;
 626         }
 627
 628         /* pci_alloc_consistent() allocates pages. */
 629         EFX_BUG_ON_PARANOID(dma_addr & (PAGE_SIZE - 1u));
 630
 631         for (kva = base_kva; kva < base_kva + PAGE_SIZE; kva += TSOH_STD_SIZE) {
 632                 tsoh = (struct efx_tso_header *)kva;
 633                 tsoh->dma_addr = dma_addr + (TSOH_BUFFER(tsoh) - base_kva);
 634                 tsoh->next = tx_queue->tso_headers_free;
 635                 tx_queue->tso_headers_free = tsoh;
 636         }
 637
 638         return 0;
 639 }
 640
 641
 642 /* Free up a TSO header, and all others in the same page. */
 643 static void efx_tsoh_block_free(struct efx_tx_queue *tx_queue,
 644                                 struct efx_tso_header *tsoh,
 645                                 struct pci_dev *pci_dev)
 646 {
 647         struct efx_tso_header **p;
 648         unsigned long base_kva;
 649         dma_addr_t base_dma;
 650
 651         base_kva = (unsigned long)tsoh & PAGE_MASK;
 652         base_dma = tsoh->dma_addr & PAGE_MASK;
 653
 654         p = &tx_queue->tso_headers_free;
 655         while (*p != NULL) {
 656                 if (((unsigned long)*p & PAGE_MASK) == base_kva)
 657                         *p = (*p)->next;
 658                 else
 659                         p = &(*p)->next;
 660         }
 661
 662         pci_free_consistent(pci_dev, PAGE_SIZE, (void *)base_kva, base_dma);
 663 }
 664
 665 static struct efx_tso_header *
 666 efx_tsoh_heap_alloc(struct efx_tx_queue *tx_queue, size_t header_len)
 667 {
 668         struct efx_tso_header *tsoh;
 669
 670         tsoh = kmalloc(TSOH_SIZE(header_len), GFP_ATOMIC | GFP_DMA);
 671         if (unlikely(!tsoh))
 672                 return NULL;
 673
 674         tsoh->dma_addr = pci_map_single(tx_queue->efx->pci_dev,
 675                                         TSOH_BUFFER(tsoh), header_len,
 676                                         PCI_DMA_TODEVICE);
 677         if (unlikely(pci_dma_mapping_error(tx_queue->efx->pci_dev,
 678                                            tsoh->dma_addr))) {
 679                 kfree(tsoh);
 680                 return NULL;
 681         }
 682
 683         tsoh->unmap_len = header_len;
 684         return tsoh;
 685 }
 686
 687 static void
 688 efx_tsoh_heap_free(struct efx_tx_queue *tx_queue, struct efx_tso_header *tsoh)
 689 {
 690         pci_unmap_single(tx_queue->efx->pci_dev,
 691                          tsoh->dma_addr, tsoh->unmap_len,
 692                          PCI_DMA_TODEVICE);
 693         kfree(tsoh);
 694 }
 695
 696 /**
 697  * efx_tx_queue_insert - push descriptors onto the TX queue
 698  * @tx_queue:           Efx TX queue
 699  * @dma_addr:           DMA address of fragment
 700  * @len:                Length of fragment
 701  * @final_buffer:       The final buffer inserted into the queue
 702  *
 703  * Push descriptors onto the TX queue.  Return 0 on success or 1 if
 704  * @tx_queue full.
 705  */
 706 static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
 707                                dma_addr_t dma_addr, unsigned len,
 708                                struct efx_tx_buffer **final_buffer)
 709 {
 710         struct efx_tx_buffer *buffer;
 711         struct efx_nic *efx = tx_queue->efx;
 712         unsigned dma_len, fill_level, insert_ptr, misalign;
 713         int q_space;
 714
 715         EFX_BUG_ON_PARANOID(len <= 0);
 716
 717         fill_level = tx_queue->insert_count - tx_queue->old_read_count;
 718         /* -1 as there is no way to represent all descriptors used */
 719         q_space = efx->type->txd_ring_mask - 1 - fill_level;
 720
 721         while (1) {
 722                 if (unlikely(q_space-- <= 0)) {
 723                         /* It might be that completions have happened
 724                          * since the xmit path last checked.  Update
 725                          * the xmit path's copy of read_count.
 726                          */
 727                         ++tx_queue->stopped;
 728                         /* This memory barrier protects the change of
 729                          * stopped from the access of read_count. */
 730                         smp_mb();
 731                         tx_queue->old_read_count =
 732                                 *(volatile unsigned *)&tx_queue->read_count;
 733                         fill_level = (tx_queue->insert_count
 734                                       - tx_queue->old_read_count);
 735                         q_space = efx->type->txd_ring_mask - 1 - fill_level;
 736                         if (unlikely(q_space-- <= 0)) {
 737                                 *final_buffer = NULL;
 738                                 return 1;
 739                         }
 740                         smp_mb();
 741                         --tx_queue->stopped;
 742                 }
 743
 744                 insert_ptr = tx_queue->insert_count & efx->type->txd_ring_mask;
 745                 buffer = &tx_queue->buffer[insert_ptr];
 746                 ++tx_queue->insert_count;
 747
 748                 EFX_BUG_ON_PARANOID(tx_queue->insert_count -
 749                                     tx_queue->read_count >
 750                                     efx->type->txd_ring_mask);
 751
 752                 efx_tsoh_free(tx_queue, buffer);
 753                 EFX_BUG_ON_PARANOID(buffer->len);
 754                 EFX_BUG_ON_PARANOID(buffer->unmap_len);
 755                 EFX_BUG_ON_PARANOID(buffer->skb);
 756                 EFX_BUG_ON_PARANOID(!buffer->continuation);
 757                 EFX_BUG_ON_PARANOID(buffer->tsoh);
 758
 759                 buffer->dma_addr = dma_addr;
 760
 761                 /* Ensure we do not cross a boundary unsupported by H/W */
 762                 dma_len = (~dma_addr & efx->type->tx_dma_mask) + 1;
 763
 764                 misalign = (unsigned)dma_addr & efx->type->bug5391_mask;
 765                 if (misalign && dma_len + misalign > 512)
 766                         dma_len = 512 - misalign;
 767
 768                 /* If there is enough space to send then do so */
 769                 if (dma_len >= len)
 770                         break;
 771
 772                 buffer->len = dma_len; /* Don't set the other members */
 773                 dma_addr += dma_len;
 774                 len -= dma_len;
 775         }
 776
 777         EFX_BUG_ON_PARANOID(!len);
 778         buffer->len = len;
 779         *final_buffer = buffer;
 780         return 0;
 781 }
 782
 783
 784 /*
 785  * Put a TSO header into the TX queue.
 786  *
 787  * This is special-cased because we know that it is small enough to fit in
 788  * a single fragment, and we know it doesn't cross a page boundary.  It
 789  * also allows us to not worry about end-of-packet etc.
 790  */
 791 static void efx_tso_put_header(struct efx_tx_queue *tx_queue,
 792                                struct efx_tso_header *tsoh, unsigned len)
 793 {
 794         struct efx_tx_buffer *buffer;
 795
 796         buffer = &tx_queue->buffer[tx_queue->insert_count &
 797                                    tx_queue->efx->type->txd_ring_mask];
 798         efx_tsoh_free(tx_queue, buffer);
 799         EFX_BUG_ON_PARANOID(buffer->len);
 800         EFX_BUG_ON_PARANOID(buffer->unmap_len);
 801         EFX_BUG_ON_PARANOID(buffer->skb);
 802         EFX_BUG_ON_PARANOID(!buffer->continuation);
 803         EFX_BUG_ON_PARANOID(buffer->tsoh);
 804         buffer->len = len;
 805         buffer->dma_addr = tsoh->dma_addr;
 806         buffer->tsoh = tsoh;
 807
 808         ++tx_queue->insert_count;
 809 }
 810
 811
 812 /* Remove descriptors put into a tx_queue. */
 813 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
 814 {
 815         struct efx_tx_buffer *buffer;
 816         dma_addr_t unmap_addr;
 817
 818         /* Work backwards until we hit the original insert pointer value */
 819         while (tx_queue->insert_count != tx_queue->write_count) {
 820                 --tx_queue->insert_count;
 821                 buffer = &tx_queue->buffer[tx_queue->insert_count &
 822                                            tx_queue->efx->type->txd_ring_mask];
 823                 efx_tsoh_free(tx_queue, buffer);
 824                 EFX_BUG_ON_PARANOID(buffer->skb);
 825                 buffer->len = 0;
 826                 buffer->continuation = true;
 827                 if (buffer->unmap_len) {
 828                         unmap_addr = (buffer->dma_addr + buffer->len -
 829                                       buffer->unmap_len);
 830                         if (buffer->unmap_single)
 831                                 pci_unmap_single(tx_queue->efx->pci_dev,
 832                                                  unmap_addr, buffer->unmap_len,
 833                                                  PCI_DMA_TODEVICE);
 834                         else
 835                                 pci_unmap_page(tx_queue->efx->pci_dev,
 836                                                unmap_addr, buffer->unmap_len,
 837                                                PCI_DMA_TODEVICE);
 838                         buffer->unmap_len = 0;
 839                 }
 840         }
 841 }
 842
 843
 844 /* Parse the SKB header and initialise state. */
 845 static void tso_start(struct tso_state *st, const struct sk_buff *skb)
 846 {
 847         /* All ethernet/IP/TCP headers combined size is TCP header size
 848          * plus offset of TCP header relative to start of packet.
 849          */
 850         st->header_len = ((tcp_hdr(skb)->doff << 2u)
 851                           + PTR_DIFF(tcp_hdr(skb), skb->data));
 852         st->full_packet_size = st->header_len + skb_shinfo(skb)->gso_size;
 853
 854         st->ipv4_id = ntohs(ip_hdr(skb)->id);
 855         st->seqnum = ntohl(tcp_hdr(skb)->seq);
 856
 857         EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
 858         EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
 859         EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
 860
 861         st->packet_space = st->full_packet_size;
 862         st->out_len = skb->len - st->header_len;
 863         st->unmap_len = 0;
 864         st->unmap_single = false;
 865 }
 866
 867 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
 868                             skb_frag_t *frag)
 869 {
 870         st->unmap_addr = pci_map_page(efx->pci_dev, frag->page,
 871                                       frag->page_offset, frag->size,
 872                                       PCI_DMA_TODEVICE);
 873         if (likely(!pci_dma_mapping_error(efx->pci_dev, st->unmap_addr))) {
 874                 st->unmap_single = false;
 875                 st->unmap_len = frag->size;
 876                 st->in_len = frag->size;
 877                 st->dma_addr = st->unmap_addr;
 878                 return 0;
 879         }
 880         return -ENOMEM;
 881 }
 882
 883 static int tso_get_head_fragment(struct tso_state *st, struct efx_nic *efx,
 884                                  const struct sk_buff *skb)
 885 {
 886         int hl = st->header_len;
 887         int len = skb_headlen(skb) - hl;
 888
 889         st->unmap_addr = pci_map_single(efx->pci_dev, skb->data + hl,
 890                                         len, PCI_DMA_TODEVICE);
 891         if (likely(!pci_dma_mapping_error(efx->pci_dev, st->unmap_addr))) {
 892                 st->unmap_single = true;
 893                 st->unmap_len = len;
 894                 st->in_len = len;
 895                 st->dma_addr = st->unmap_addr;
 896                 return 0;
 897         }
 898         return -ENOMEM;
 899 }
 900
 901
 902 /**
 903  * tso_fill_packet_with_fragment - form descriptors for the current fragment
 904  * @tx_queue:           Efx TX queue
 905  * @skb:                Socket buffer
 906  * @st:                 TSO state
 907  *
 908  * Form descriptors for the current fragment, until we reach the end
 909  * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
 910  * space in @tx_queue.
 911  */
 912 static int tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
 913                                          const struct sk_buff *skb,
 914                                          struct tso_state *st)
 915 {
 916         struct efx_tx_buffer *buffer;
 917         int n, end_of_packet, rc;
 918
 919         if (st->in_len == 0)
 920                 return 0;
 921         if (st->packet_space == 0)
 922                 return 0;
 923
 924         EFX_BUG_ON_PARANOID(st->in_len <= 0);
 925         EFX_BUG_ON_PARANOID(st->packet_space <= 0);
 926
 927         n = min(st->in_len, st->packet_space);
 928
 929         st->packet_space -= n;
 930         st->out_len -= n;
 931         st->in_len -= n;
 932
 933         rc = efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer);
 934         if (likely(rc == 0)) {
 935                 if (st->out_len == 0)
 936                         /* Transfer ownership of the skb */
 937                         buffer->skb = skb;
 938
 939                 end_of_packet = st->out_len == 0 || st->packet_space == 0;
 940                 buffer->continuation = !end_of_packet;
 941
 942                 if (st->in_len == 0) {
 943                         /* Transfer ownership of the pci mapping */
 944                         buffer->unmap_len = st->unmap_len;
 945                         buffer->unmap_single = st->unmap_single;
 946                         st->unmap_len = 0;
 947                 }
 948         }
 949
 950         st->dma_addr += n;
 951         return rc;
 952 }
 953
 954
 955 /**
 956  * tso_start_new_packet - generate a new header and prepare for the new packet
 957  * @tx_queue:           Efx TX queue
 958  * @skb:                Socket buffer
 959  * @st:                 TSO state
 960  *
 961  * Generate a new header and prepare for the new packet.  Return 0 on
 962  * success, or -1 if failed to alloc header.
 963  */
 964 static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
 965                                 const struct sk_buff *skb,
 966                                 struct tso_state *st)
 967 {
 968         struct efx_tso_header *tsoh;
 969         struct iphdr *tsoh_iph;
 970         struct tcphdr *tsoh_th;
 971         unsigned ip_length;
 972         u8 *header;
 973
 974         /* Allocate a DMA-mapped header buffer. */
 975         if (likely(TSOH_SIZE(st->header_len) <= TSOH_STD_SIZE)) {
 976                 if (tx_queue->tso_headers_free == NULL) {
 977                         if (efx_tsoh_block_alloc(tx_queue))
 978                                 return -1;
 979                 }
 980                 EFX_BUG_ON_PARANOID(!tx_queue->tso_headers_free);
 981                 tsoh = tx_queue->tso_headers_free;
 982                 tx_queue->tso_headers_free = tsoh->next;
 983                 tsoh->unmap_len = 0;
 984         } else {
 985                 tx_queue->tso_long_headers++;
 986                 tsoh = efx_tsoh_heap_alloc(tx_queue, st->header_len);
 987                 if (unlikely(!tsoh))
 988                         return -1;
 989         }
 990
 991         header = TSOH_BUFFER(tsoh);
 992         tsoh_th = (struct tcphdr *)(header + SKB_TCP_OFF(skb));
 993         tsoh_iph = (struct iphdr *)(header + SKB_IPV4_OFF(skb));
 994
 995         /* Copy and update the headers. */
 996         memcpy(header, skb->data, st->header_len);
 997
 998         tsoh_th->seq = htonl(st->seqnum);
 999         st->seqnum += skb_shinfo(skb)->gso_size;
1000         if (st->out_len > skb_shinfo(skb)->gso_size) {
1001                 /* This packet will not finish the TSO burst. */
1002                 ip_length = st->full_packet_size - ETH_HDR_LEN(skb);
1003                 tsoh_th->fin = 0;
1004                 tsoh_th->psh = 0;
1005         } else {
1006                 /* This packet will be the last in the TSO burst. */
1007                 ip_length = st->header_len - ETH_HDR_LEN(skb) + st->out_len;
1008                 tsoh_th->fin = tcp_hdr(skb)->fin;
1009                 tsoh_th->psh = tcp_hdr(skb)->psh;
1010         }
1011         tsoh_iph->tot_len = htons(ip_length);
1012
1013         /* Linux leaves suitable gaps in the IP ID space for us to fill. */
1014         tsoh_iph->id = htons(st->ipv4_id);
1015         st->ipv4_id++;
1016
1017         st->packet_space = skb_shinfo(skb)->gso_size;
1018         ++tx_queue->tso_packets;
1019
1020         /* Form a descriptor for this header. */
1021         efx_tso_put_header(tx_queue, tsoh, st->header_len);
1022
1023         return 0;
1024 }
1025
1026
1027 /**
1028  * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1029  * @tx_queue:           Efx TX queue
1030  * @skb:                Socket buffer
1031  *
1032  * Context: You must hold netif_tx_lock() to call this function.
1033  *
1034  * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1035  * @skb was not enqueued.  In all cases @skb is consumed.  Return
1036  * %NETDEV_TX_OK or %NETDEV_TX_BUSY.
1037  */
1038 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1039                                struct sk_buff *skb)
1040 {
1041         struct efx_nic *efx = tx_queue->efx;
1042         int frag_i, rc, rc2 = NETDEV_TX_OK;
1043         struct tso_state state;
1044
1045         /* Verify TSO is safe - these checks should never fail. */
1046         efx_tso_check_safe(skb);
1047
1048         EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
1049
1050         tso_start(&state, skb);
1051
1052         /* Assume that skb header area contains exactly the headers, and
1053          * all payload is in the frag list.
1054          */
1055         if (skb_headlen(skb) == state.header_len) {
1056                 /* Grab the first payload fragment. */
1057                 EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1058                 frag_i = 0;
1059                 rc = tso_get_fragment(&state, efx,
1060                                       skb_shinfo(skb)->frags + frag_i);
1061                 if (rc)
1062                         goto mem_err;
1063         } else {
1064                 rc = tso_get_head_fragment(&state, efx, skb);
1065                 if (rc)
1066                         goto mem_err;
1067                 frag_i = -1;
1068         }
1069
1070         if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1071                 goto mem_err;
1072
1073         while (1) {
1074                 rc = tso_fill_packet_with_fragment(tx_queue, skb, &state);
1075                 if (unlikely(rc))
1076                         goto stop;
1077
1078                 /* Move onto the next fragment? */
1079                 if (state.in_len == 0) {
1080                         if (++frag_i >= skb_shinfo(skb)->nr_frags)
1081                                 /* End of payload reached. */
1082                                 break;
1083                         rc = tso_get_fragment(&state, efx,
1084                                               skb_shinfo(skb)->frags + frag_i);
1085                         if (rc)
1086                                 goto mem_err;
1087                 }
1088
1089                 /* Start at new packet? */
1090                 if (state.packet_space == 0 &&
1091                     tso_start_new_packet(tx_queue, skb, &state) < 0)
1092                         goto mem_err;
1093         }
1094
1095         /* Pass off to hardware */
1096         falcon_push_buffers(tx_queue);
1097
1098         tx_queue->tso_bursts++;
1099         return NETDEV_TX_OK;
1100
1101  mem_err:
1102         EFX_ERR(efx, "Out of memory for TSO headers, or PCI mapping error\n");
1103         dev_kfree_skb_any((struct sk_buff *)skb);
1104         goto unwind;
1105
1106  stop:
1107         rc2 = NETDEV_TX_BUSY;
1108
1109         /* Stop the queue if it wasn't stopped before. */
1110         if (tx_queue->stopped == 1)
1111                 efx_stop_queue(efx);
1112
1113  unwind:
1114         /* Free the DMA mapping we were in the process of writing out */
1115         if (state.unmap_len) {
1116                 if (state.unmap_single)
1117                         pci_unmap_single(efx->pci_dev, state.unmap_addr,
1118                                          state.unmap_len, PCI_DMA_TODEVICE);
1119                 else
1120                         pci_unmap_page(efx->pci_dev, state.unmap_addr,
1121                                        state.unmap_len, PCI_DMA_TODEVICE);
1122         }
1123
1124         efx_enqueue_unwind(tx_queue);
1125         return rc2;
1126 }
1127
1128
1129 /*
1130  * Free up all TSO datastructures associated with tx_queue. This
1131  * routine should be called only once the tx_queue is both empty and
1132  * will no longer be used.
1133  */
1134 static void efx_fini_tso(struct efx_tx_queue *tx_queue)
1135 {
1136         unsigned i;
1137
1138         if (tx_queue->buffer) {
1139                 for (i = 0; i <= tx_queue->efx->type->txd_ring_mask; ++i)
1140                         efx_tsoh_free(tx_queue, &tx_queue->buffer[i]);
1141         }
1142
1143         while (tx_queue->tso_headers_free != NULL)
1144                 efx_tsoh_block_free(tx_queue, tx_queue->tso_headers_free,
1145                                     tx_queue->efx->pci_dev);
1146 }