1 /**************************************************************************** 2 * Driver for Solarflare network controllers and boards 3 * Copyright 2005-2006 Fen Systems Ltd. 4 * Copyright 2005-2013 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/ipv6.h> 16 #include <linux/slab.h> 17 #include <net/ipv6.h> 18 #include <linux/if_ether.h> 19 #include <linux/highmem.h> 20 #include <linux/cache.h> 21 #include "net_driver.h" 22 #include "efx.h" 23 #include "io.h" 24 #include "nic.h" 25 #include "workarounds.h" 26 #include "ef10_regs.h" 27 28 #ifdef EFX_USE_PIO 29 30 #define EFX_PIOBUF_SIZE_MAX ER_DZ_TX_PIOBUF_SIZE 31 #define EFX_PIOBUF_SIZE_DEF ALIGN(256, L1_CACHE_BYTES) 32 unsigned int efx_piobuf_size __read_mostly = EFX_PIOBUF_SIZE_DEF; 33 34 #endif /* EFX_USE_PIO */ 35 36 static inline unsigned int 37 efx_tx_queue_get_insert_index(const struct efx_tx_queue *tx_queue) 38 { 39 return tx_queue->insert_count & tx_queue->ptr_mask; 40 } 41 42 static inline struct efx_tx_buffer * 43 __efx_tx_queue_get_insert_buffer(const struct efx_tx_queue *tx_queue) 44 { 45 return &tx_queue->buffer[efx_tx_queue_get_insert_index(tx_queue)]; 46 } 47 48 static inline struct efx_tx_buffer * 49 efx_tx_queue_get_insert_buffer(const struct efx_tx_queue *tx_queue) 50 { 51 struct efx_tx_buffer *buffer = 52 __efx_tx_queue_get_insert_buffer(tx_queue); 53 54 EFX_BUG_ON_PARANOID(buffer->len); 55 EFX_BUG_ON_PARANOID(buffer->flags); 56 EFX_BUG_ON_PARANOID(buffer->unmap_len); 57 58 return buffer; 59 } 60 61 static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue, 62 struct efx_tx_buffer *buffer, 63 unsigned int *pkts_compl, 64 unsigned int *bytes_compl) 65 { 66 if (buffer->unmap_len) { 67 struct device *dma_dev = &tx_queue->efx->pci_dev->dev; 68 dma_addr_t unmap_addr = buffer->dma_addr - buffer->dma_offset; 69 if (buffer->flags & EFX_TX_BUF_MAP_SINGLE) 70 dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len, 71 DMA_TO_DEVICE); 72 else 73 dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len, 74 DMA_TO_DEVICE); 75 buffer->unmap_len = 0; 76 } 77 78 if (buffer->flags & EFX_TX_BUF_SKB) { 79 (*pkts_compl)++; 80 (*bytes_compl) += buffer->skb->len; 81 dev_kfree_skb_any((struct sk_buff *) buffer->skb); 82 netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev, 83 "TX queue %d transmission id %x complete\n", 84 tx_queue->queue, tx_queue->read_count); 85 } else if (buffer->flags & EFX_TX_BUF_HEAP) { 86 kfree(buffer->heap_buf); 87 } 88 89 buffer->len = 0; 90 buffer->flags = 0; 91 } 92 93 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue, 94 struct sk_buff *skb); 95 96 static inline unsigned 97 efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr) 98 { 99 /* Depending on the NIC revision, we can use descriptor 100 * lengths up to 8K or 8K-1. However, since PCI Express 101 * devices must split read requests at 4K boundaries, there is 102 * little benefit from using descriptors that cross those 103 * boundaries and we keep things simple by not doing so. 104 */ 105 unsigned len = (~dma_addr & (EFX_PAGE_SIZE - 1)) + 1; 106 107 /* Work around hardware bug for unaligned buffers. */ 108 if (EFX_WORKAROUND_5391(efx) && (dma_addr & 0xf)) 109 len = min_t(unsigned, len, 512 - (dma_addr & 0xf)); 110 111 return len; 112 } 113 114 unsigned int efx_tx_max_skb_descs(struct efx_nic *efx) 115 { 116 /* Header and payload descriptor for each output segment, plus 117 * one for every input fragment boundary within a segment 118 */ 119 unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS; 120 121 /* Possibly one more per segment for the alignment workaround, 122 * or for option descriptors 123 */ 124 if (EFX_WORKAROUND_5391(efx) || efx_nic_rev(efx) >= EFX_REV_HUNT_A0) 125 max_descs += EFX_TSO_MAX_SEGS; 126 127 /* Possibly more for PCIe page boundaries within input fragments */ 128 if (PAGE_SIZE > EFX_PAGE_SIZE) 129 max_descs += max_t(unsigned int, MAX_SKB_FRAGS, 130 DIV_ROUND_UP(GSO_MAX_SIZE, EFX_PAGE_SIZE)); 131 132 return max_descs; 133 } 134 135 /* Get partner of a TX queue, seen as part of the same net core queue */ 136 static struct efx_tx_queue *efx_tx_queue_partner(struct efx_tx_queue *tx_queue) 137 { 138 if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD) 139 return tx_queue - EFX_TXQ_TYPE_OFFLOAD; 140 else 141 return tx_queue + EFX_TXQ_TYPE_OFFLOAD; 142 } 143 144 static void efx_tx_maybe_stop_queue(struct efx_tx_queue *txq1) 145 { 146 /* We need to consider both queues that the net core sees as one */ 147 struct efx_tx_queue *txq2 = efx_tx_queue_partner(txq1); 148 struct efx_nic *efx = txq1->efx; 149 unsigned int fill_level; 150 151 fill_level = max(txq1->insert_count - txq1->old_read_count, 152 txq2->insert_count - txq2->old_read_count); 153 if (likely(fill_level < efx->txq_stop_thresh)) 154 return; 155 156 /* We used the stale old_read_count above, which gives us a 157 * pessimistic estimate of the fill level (which may even 158 * validly be >= efx->txq_entries). Now try again using 159 * read_count (more likely to be a cache miss). 160 * 161 * If we read read_count and then conditionally stop the 162 * queue, it is possible for the completion path to race with 163 * us and complete all outstanding descriptors in the middle, 164 * after which there will be no more completions to wake it. 165 * Therefore we stop the queue first, then read read_count 166 * (with a memory barrier to ensure the ordering), then 167 * restart the queue if the fill level turns out to be low 168 * enough. 169 */ 170 netif_tx_stop_queue(txq1->core_txq); 171 smp_mb(); 172 txq1->old_read_count = ACCESS_ONCE(txq1->read_count); 173 txq2->old_read_count = ACCESS_ONCE(txq2->read_count); 174 175 fill_level = max(txq1->insert_count - txq1->old_read_count, 176 txq2->insert_count - txq2->old_read_count); 177 EFX_BUG_ON_PARANOID(fill_level >= efx->txq_entries); 178 if (likely(fill_level < efx->txq_stop_thresh)) { 179 smp_mb(); 180 if (likely(!efx->loopback_selftest)) 181 netif_tx_start_queue(txq1->core_txq); 182 } 183 } 184 185 #ifdef EFX_USE_PIO 186 187 struct efx_short_copy_buffer { 188 int used; 189 u8 buf[L1_CACHE_BYTES]; 190 }; 191 192 /* Copy to PIO, respecting that writes to PIO buffers must be dword aligned. 193 * Advances piobuf pointer. Leaves additional data in the copy buffer. 194 */ 195 static void efx_memcpy_toio_aligned(struct efx_nic *efx, u8 __iomem **piobuf, 196 u8 *data, int len, 197 struct efx_short_copy_buffer *copy_buf) 198 { 199 int block_len = len & ~(sizeof(copy_buf->buf) - 1); 200 201 memcpy_toio(*piobuf, data, block_len); 202 *piobuf += block_len; 203 len -= block_len; 204 205 if (len) { 206 data += block_len; 207 BUG_ON(copy_buf->used); 208 BUG_ON(len > sizeof(copy_buf->buf)); 209 memcpy(copy_buf->buf, data, len); 210 copy_buf->used = len; 211 } 212 } 213 214 /* Copy to PIO, respecting dword alignment, popping data from copy buffer first. 215 * Advances piobuf pointer. Leaves additional data in the copy buffer. 216 */ 217 static void efx_memcpy_toio_aligned_cb(struct efx_nic *efx, u8 __iomem **piobuf, 218 u8 *data, int len, 219 struct efx_short_copy_buffer *copy_buf) 220 { 221 if (copy_buf->used) { 222 /* if the copy buffer is partially full, fill it up and write */ 223 int copy_to_buf = 224 min_t(int, sizeof(copy_buf->buf) - copy_buf->used, len); 225 226 memcpy(copy_buf->buf + copy_buf->used, data, copy_to_buf); 227 copy_buf->used += copy_to_buf; 228 229 /* if we didn't fill it up then we're done for now */ 230 if (copy_buf->used < sizeof(copy_buf->buf)) 231 return; 232 233 memcpy_toio(*piobuf, copy_buf->buf, sizeof(copy_buf->buf)); 234 *piobuf += sizeof(copy_buf->buf); 235 data += copy_to_buf; 236 len -= copy_to_buf; 237 copy_buf->used = 0; 238 } 239 240 efx_memcpy_toio_aligned(efx, piobuf, data, len, copy_buf); 241 } 242 243 static void efx_flush_copy_buffer(struct efx_nic *efx, u8 __iomem *piobuf, 244 struct efx_short_copy_buffer *copy_buf) 245 { 246 /* if there's anything in it, write the whole buffer, including junk */ 247 if (copy_buf->used) 248 memcpy_toio(piobuf, copy_buf->buf, sizeof(copy_buf->buf)); 249 } 250 251 /* Traverse skb structure and copy fragments in to PIO buffer. 252 * Advances piobuf pointer. 253 */ 254 static void efx_skb_copy_bits_to_pio(struct efx_nic *efx, struct sk_buff *skb, 255 u8 __iomem **piobuf, 256 struct efx_short_copy_buffer *copy_buf) 257 { 258 int i; 259 260 efx_memcpy_toio_aligned(efx, piobuf, skb->data, skb_headlen(skb), 261 copy_buf); 262 263 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) { 264 skb_frag_t *f = &skb_shinfo(skb)->frags[i]; 265 u8 *vaddr; 266 267 vaddr = kmap_atomic(skb_frag_page(f)); 268 269 efx_memcpy_toio_aligned_cb(efx, piobuf, vaddr + f->page_offset, 270 skb_frag_size(f), copy_buf); 271 kunmap_atomic(vaddr); 272 } 273 274 EFX_BUG_ON_PARANOID(skb_shinfo(skb)->frag_list); 275 } 276 277 static struct efx_tx_buffer * 278 efx_enqueue_skb_pio(struct efx_tx_queue *tx_queue, struct sk_buff *skb) 279 { 280 struct efx_tx_buffer *buffer = 281 efx_tx_queue_get_insert_buffer(tx_queue); 282 u8 __iomem *piobuf = tx_queue->piobuf; 283 284 /* Copy to PIO buffer. Ensure the writes are padded to the end 285 * of a cache line, as this is required for write-combining to be 286 * effective on at least x86. 287 */ 288 289 if (skb_shinfo(skb)->nr_frags) { 290 /* The size of the copy buffer will ensure all writes 291 * are the size of a cache line. 292 */ 293 struct efx_short_copy_buffer copy_buf; 294 295 copy_buf.used = 0; 296 297 efx_skb_copy_bits_to_pio(tx_queue->efx, skb, 298 &piobuf, ©_buf); 299 efx_flush_copy_buffer(tx_queue->efx, piobuf, ©_buf); 300 } else { 301 /* Pad the write to the size of a cache line. 302 * We can do this because we know the skb_shared_info sruct is 303 * after the source, and the destination buffer is big enough. 304 */ 305 BUILD_BUG_ON(L1_CACHE_BYTES > 306 SKB_DATA_ALIGN(sizeof(struct skb_shared_info))); 307 memcpy_toio(tx_queue->piobuf, skb->data, 308 ALIGN(skb->len, L1_CACHE_BYTES)); 309 } 310 311 EFX_POPULATE_QWORD_5(buffer->option, 312 ESF_DZ_TX_DESC_IS_OPT, 1, 313 ESF_DZ_TX_OPTION_TYPE, ESE_DZ_TX_OPTION_DESC_PIO, 314 ESF_DZ_TX_PIO_CONT, 0, 315 ESF_DZ_TX_PIO_BYTE_CNT, skb->len, 316 ESF_DZ_TX_PIO_BUF_ADDR, 317 tx_queue->piobuf_offset); 318 ++tx_queue->pio_packets; 319 ++tx_queue->insert_count; 320 return buffer; 321 } 322 #endif /* EFX_USE_PIO */ 323 324 /* 325 * Add a socket buffer to a TX queue 326 * 327 * This maps all fragments of a socket buffer for DMA and adds them to 328 * the TX queue. The queue's insert pointer will be incremented by 329 * the number of fragments in the socket buffer. 330 * 331 * If any DMA mapping fails, any mapped fragments will be unmapped, 332 * the queue's insert pointer will be restored to its original value. 333 * 334 * This function is split out from efx_hard_start_xmit to allow the 335 * loopback test to direct packets via specific TX queues. 336 * 337 * Returns NETDEV_TX_OK. 338 * You must hold netif_tx_lock() to call this function. 339 */ 340 netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb) 341 { 342 struct efx_nic *efx = tx_queue->efx; 343 struct device *dma_dev = &efx->pci_dev->dev; 344 struct efx_tx_buffer *buffer; 345 skb_frag_t *fragment; 346 unsigned int len, unmap_len = 0; 347 dma_addr_t dma_addr, unmap_addr = 0; 348 unsigned int dma_len; 349 unsigned short dma_flags; 350 int i = 0; 351 352 EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count); 353 354 if (skb_shinfo(skb)->gso_size) 355 return efx_enqueue_skb_tso(tx_queue, skb); 356 357 /* Get size of the initial fragment */ 358 len = skb_headlen(skb); 359 360 /* Pad if necessary */ 361 if (EFX_WORKAROUND_15592(efx) && skb->len <= 32) { 362 EFX_BUG_ON_PARANOID(skb->data_len); 363 len = 32 + 1; 364 if (skb_pad(skb, len - skb->len)) 365 return NETDEV_TX_OK; 366 } 367 368 /* Consider using PIO for short packets */ 369 #ifdef EFX_USE_PIO 370 if (skb->len <= efx_piobuf_size && tx_queue->piobuf && 371 efx_nic_tx_is_empty(tx_queue) && 372 efx_nic_tx_is_empty(efx_tx_queue_partner(tx_queue))) { 373 buffer = efx_enqueue_skb_pio(tx_queue, skb); 374 dma_flags = EFX_TX_BUF_OPTION; 375 goto finish_packet; 376 } 377 #endif 378 379 /* Map for DMA. Use dma_map_single rather than dma_map_page 380 * since this is more efficient on machines with sparse 381 * memory. 382 */ 383 dma_flags = EFX_TX_BUF_MAP_SINGLE; 384 dma_addr = dma_map_single(dma_dev, skb->data, len, PCI_DMA_TODEVICE); 385 386 /* Process all fragments */ 387 while (1) { 388 if (unlikely(dma_mapping_error(dma_dev, dma_addr))) 389 goto dma_err; 390 391 /* Store fields for marking in the per-fragment final 392 * descriptor */ 393 unmap_len = len; 394 unmap_addr = dma_addr; 395 396 /* Add to TX queue, splitting across DMA boundaries */ 397 do { 398 buffer = efx_tx_queue_get_insert_buffer(tx_queue); 399 400 dma_len = efx_max_tx_len(efx, dma_addr); 401 if (likely(dma_len >= len)) 402 dma_len = len; 403 404 /* Fill out per descriptor fields */ 405 buffer->len = dma_len; 406 buffer->dma_addr = dma_addr; 407 buffer->flags = EFX_TX_BUF_CONT; 408 len -= dma_len; 409 dma_addr += dma_len; 410 ++tx_queue->insert_count; 411 } while (len); 412 413 /* Transfer ownership of the unmapping to the final buffer */ 414 buffer->flags = EFX_TX_BUF_CONT | dma_flags; 415 buffer->unmap_len = unmap_len; 416 buffer->dma_offset = buffer->dma_addr - unmap_addr; 417 unmap_len = 0; 418 419 /* Get address and size of next fragment */ 420 if (i >= skb_shinfo(skb)->nr_frags) 421 break; 422 fragment = &skb_shinfo(skb)->frags[i]; 423 len = skb_frag_size(fragment); 424 i++; 425 /* Map for DMA */ 426 dma_flags = 0; 427 dma_addr = skb_frag_dma_map(dma_dev, fragment, 0, len, 428 DMA_TO_DEVICE); 429 } 430 431 /* Transfer ownership of the skb to the final buffer */ 432 #ifdef EFX_USE_PIO 433 finish_packet: 434 #endif 435 buffer->skb = skb; 436 buffer->flags = EFX_TX_BUF_SKB | dma_flags; 437 438 netdev_tx_sent_queue(tx_queue->core_txq, skb->len); 439 440 /* Pass off to hardware */ 441 efx_nic_push_buffers(tx_queue); 442 443 efx_tx_maybe_stop_queue(tx_queue); 444 445 return NETDEV_TX_OK; 446 447 dma_err: 448 netif_err(efx, tx_err, efx->net_dev, 449 " TX queue %d could not map skb with %d bytes %d " 450 "fragments for DMA\n", tx_queue->queue, skb->len, 451 skb_shinfo(skb)->nr_frags + 1); 452 453 /* Mark the packet as transmitted, and free the SKB ourselves */ 454 dev_kfree_skb_any(skb); 455 456 /* Work backwards until we hit the original insert pointer value */ 457 while (tx_queue->insert_count != tx_queue->write_count) { 458 unsigned int pkts_compl = 0, bytes_compl = 0; 459 --tx_queue->insert_count; 460 buffer = __efx_tx_queue_get_insert_buffer(tx_queue); 461 efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl); 462 } 463 464 /* Free the fragment we were mid-way through pushing */ 465 if (unmap_len) { 466 if (dma_flags & EFX_TX_BUF_MAP_SINGLE) 467 dma_unmap_single(dma_dev, unmap_addr, unmap_len, 468 DMA_TO_DEVICE); 469 else 470 dma_unmap_page(dma_dev, unmap_addr, unmap_len, 471 DMA_TO_DEVICE); 472 } 473 474 return NETDEV_TX_OK; 475 } 476 477 /* Remove packets from the TX queue 478 * 479 * This removes packets from the TX queue, up to and including the 480 * specified index. 481 */ 482 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue, 483 unsigned int index, 484 unsigned int *pkts_compl, 485 unsigned int *bytes_compl) 486 { 487 struct efx_nic *efx = tx_queue->efx; 488 unsigned int stop_index, read_ptr; 489 490 stop_index = (index + 1) & tx_queue->ptr_mask; 491 read_ptr = tx_queue->read_count & tx_queue->ptr_mask; 492 493 while (read_ptr != stop_index) { 494 struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr]; 495 496 if (!(buffer->flags & EFX_TX_BUF_OPTION) && 497 unlikely(buffer->len == 0)) { 498 netif_err(efx, tx_err, efx->net_dev, 499 "TX queue %d spurious TX completion id %x\n", 500 tx_queue->queue, read_ptr); 501 efx_schedule_reset(efx, RESET_TYPE_TX_SKIP); 502 return; 503 } 504 505 efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl); 506 507 ++tx_queue->read_count; 508 read_ptr = tx_queue->read_count & tx_queue->ptr_mask; 509 } 510 } 511 512 /* Initiate a packet transmission. We use one channel per CPU 513 * (sharing when we have more CPUs than channels). On Falcon, the TX 514 * completion events will be directed back to the CPU that transmitted 515 * the packet, which should be cache-efficient. 516 * 517 * Context: non-blocking. 518 * Note that returning anything other than NETDEV_TX_OK will cause the 519 * OS to free the skb. 520 */ 521 netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb, 522 struct net_device *net_dev) 523 { 524 struct efx_nic *efx = netdev_priv(net_dev); 525 struct efx_tx_queue *tx_queue; 526 unsigned index, type; 527 528 EFX_WARN_ON_PARANOID(!netif_device_present(net_dev)); 529 530 /* PTP "event" packet */ 531 if (unlikely(efx_xmit_with_hwtstamp(skb)) && 532 unlikely(efx_ptp_is_ptp_tx(efx, skb))) { 533 return efx_ptp_tx(efx, skb); 534 } 535 536 index = skb_get_queue_mapping(skb); 537 type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0; 538 if (index >= efx->n_tx_channels) { 539 index -= efx->n_tx_channels; 540 type |= EFX_TXQ_TYPE_HIGHPRI; 541 } 542 tx_queue = efx_get_tx_queue(efx, index, type); 543 544 return efx_enqueue_skb(tx_queue, skb); 545 } 546 547 void efx_init_tx_queue_core_txq(struct efx_tx_queue *tx_queue) 548 { 549 struct efx_nic *efx = tx_queue->efx; 550 551 /* Must be inverse of queue lookup in efx_hard_start_xmit() */ 552 tx_queue->core_txq = 553 netdev_get_tx_queue(efx->net_dev, 554 tx_queue->queue / EFX_TXQ_TYPES + 555 ((tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ? 556 efx->n_tx_channels : 0)); 557 } 558 559 int efx_setup_tc(struct net_device *net_dev, u8 num_tc) 560 { 561 struct efx_nic *efx = netdev_priv(net_dev); 562 struct efx_channel *channel; 563 struct efx_tx_queue *tx_queue; 564 unsigned tc; 565 int rc; 566 567 if (efx_nic_rev(efx) < EFX_REV_FALCON_B0 || num_tc > EFX_MAX_TX_TC) 568 return -EINVAL; 569 570 if (num_tc == net_dev->num_tc) 571 return 0; 572 573 for (tc = 0; tc < num_tc; tc++) { 574 net_dev->tc_to_txq[tc].offset = tc * efx->n_tx_channels; 575 net_dev->tc_to_txq[tc].count = efx->n_tx_channels; 576 } 577 578 if (num_tc > net_dev->num_tc) { 579 /* Initialise high-priority queues as necessary */ 580 efx_for_each_channel(channel, efx) { 581 efx_for_each_possible_channel_tx_queue(tx_queue, 582 channel) { 583 if (!(tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI)) 584 continue; 585 if (!tx_queue->buffer) { 586 rc = efx_probe_tx_queue(tx_queue); 587 if (rc) 588 return rc; 589 } 590 if (!tx_queue->initialised) 591 efx_init_tx_queue(tx_queue); 592 efx_init_tx_queue_core_txq(tx_queue); 593 } 594 } 595 } else { 596 /* Reduce number of classes before number of queues */ 597 net_dev->num_tc = num_tc; 598 } 599 600 rc = netif_set_real_num_tx_queues(net_dev, 601 max_t(int, num_tc, 1) * 602 efx->n_tx_channels); 603 if (rc) 604 return rc; 605 606 /* Do not destroy high-priority queues when they become 607 * unused. We would have to flush them first, and it is 608 * fairly difficult to flush a subset of TX queues. Leave 609 * it to efx_fini_channels(). 610 */ 611 612 net_dev->num_tc = num_tc; 613 return 0; 614 } 615 616 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index) 617 { 618 unsigned fill_level; 619 struct efx_nic *efx = tx_queue->efx; 620 struct efx_tx_queue *txq2; 621 unsigned int pkts_compl = 0, bytes_compl = 0; 622 623 EFX_BUG_ON_PARANOID(index > tx_queue->ptr_mask); 624 625 efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl); 626 netdev_tx_completed_queue(tx_queue->core_txq, pkts_compl, bytes_compl); 627 628 if (pkts_compl > 1) 629 ++tx_queue->merge_events; 630 631 /* See if we need to restart the netif queue. This memory 632 * barrier ensures that we write read_count (inside 633 * efx_dequeue_buffers()) before reading the queue status. 634 */ 635 smp_mb(); 636 if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) && 637 likely(efx->port_enabled) && 638 likely(netif_device_present(efx->net_dev))) { 639 txq2 = efx_tx_queue_partner(tx_queue); 640 fill_level = max(tx_queue->insert_count - tx_queue->read_count, 641 txq2->insert_count - txq2->read_count); 642 if (fill_level <= efx->txq_wake_thresh) 643 netif_tx_wake_queue(tx_queue->core_txq); 644 } 645 646 /* Check whether the hardware queue is now empty */ 647 if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) { 648 tx_queue->old_write_count = ACCESS_ONCE(tx_queue->write_count); 649 if (tx_queue->read_count == tx_queue->old_write_count) { 650 smp_mb(); 651 tx_queue->empty_read_count = 652 tx_queue->read_count | EFX_EMPTY_COUNT_VALID; 653 } 654 } 655 } 656 657 /* Size of page-based TSO header buffers. Larger blocks must be 658 * allocated from the heap. 659 */ 660 #define TSOH_STD_SIZE 128 661 #define TSOH_PER_PAGE (PAGE_SIZE / TSOH_STD_SIZE) 662 663 /* At most half the descriptors in the queue at any time will refer to 664 * a TSO header buffer, since they must always be followed by a 665 * payload descriptor referring to an skb. 666 */ 667 static unsigned int efx_tsoh_page_count(struct efx_tx_queue *tx_queue) 668 { 669 return DIV_ROUND_UP(tx_queue->ptr_mask + 1, 2 * TSOH_PER_PAGE); 670 } 671 672 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue) 673 { 674 struct efx_nic *efx = tx_queue->efx; 675 unsigned int entries; 676 int rc; 677 678 /* Create the smallest power-of-two aligned ring */ 679 entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE); 680 EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE); 681 tx_queue->ptr_mask = entries - 1; 682 683 netif_dbg(efx, probe, efx->net_dev, 684 "creating TX queue %d size %#x mask %#x\n", 685 tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask); 686 687 /* Allocate software ring */ 688 tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer), 689 GFP_KERNEL); 690 if (!tx_queue->buffer) 691 return -ENOMEM; 692 693 if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD) { 694 tx_queue->tsoh_page = 695 kcalloc(efx_tsoh_page_count(tx_queue), 696 sizeof(tx_queue->tsoh_page[0]), GFP_KERNEL); 697 if (!tx_queue->tsoh_page) { 698 rc = -ENOMEM; 699 goto fail1; 700 } 701 } 702 703 /* Allocate hardware ring */ 704 rc = efx_nic_probe_tx(tx_queue); 705 if (rc) 706 goto fail2; 707 708 return 0; 709 710 fail2: 711 kfree(tx_queue->tsoh_page); 712 tx_queue->tsoh_page = NULL; 713 fail1: 714 kfree(tx_queue->buffer); 715 tx_queue->buffer = NULL; 716 return rc; 717 } 718 719 void efx_init_tx_queue(struct efx_tx_queue *tx_queue) 720 { 721 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev, 722 "initialising TX queue %d\n", tx_queue->queue); 723 724 tx_queue->insert_count = 0; 725 tx_queue->write_count = 0; 726 tx_queue->old_write_count = 0; 727 tx_queue->read_count = 0; 728 tx_queue->old_read_count = 0; 729 tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID; 730 731 /* Set up TX descriptor ring */ 732 efx_nic_init_tx(tx_queue); 733 734 tx_queue->initialised = true; 735 } 736 737 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue) 738 { 739 struct efx_tx_buffer *buffer; 740 741 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev, 742 "shutting down TX queue %d\n", tx_queue->queue); 743 744 if (!tx_queue->buffer) 745 return; 746 747 /* Free any buffers left in the ring */ 748 while (tx_queue->read_count != tx_queue->write_count) { 749 unsigned int pkts_compl = 0, bytes_compl = 0; 750 buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask]; 751 efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl); 752 753 ++tx_queue->read_count; 754 } 755 netdev_tx_reset_queue(tx_queue->core_txq); 756 } 757 758 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue) 759 { 760 int i; 761 762 if (!tx_queue->buffer) 763 return; 764 765 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev, 766 "destroying TX queue %d\n", tx_queue->queue); 767 efx_nic_remove_tx(tx_queue); 768 769 if (tx_queue->tsoh_page) { 770 for (i = 0; i < efx_tsoh_page_count(tx_queue); i++) 771 efx_nic_free_buffer(tx_queue->efx, 772 &tx_queue->tsoh_page[i]); 773 kfree(tx_queue->tsoh_page); 774 tx_queue->tsoh_page = NULL; 775 } 776 777 kfree(tx_queue->buffer); 778 tx_queue->buffer = NULL; 779 } 780 781 782 /* Efx TCP segmentation acceleration. 783 * 784 * Why? Because by doing it here in the driver we can go significantly 785 * faster than the GSO. 786 * 787 * Requires TX checksum offload support. 788 */ 789 790 /* Number of bytes inserted at the start of a TSO header buffer, 791 * similar to NET_IP_ALIGN. 792 */ 793 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS 794 #define TSOH_OFFSET 0 795 #else 796 #define TSOH_OFFSET NET_IP_ALIGN 797 #endif 798 799 #define PTR_DIFF(p1, p2) ((u8 *)(p1) - (u8 *)(p2)) 800 801 /** 802 * struct tso_state - TSO state for an SKB 803 * @out_len: Remaining length in current segment 804 * @seqnum: Current sequence number 805 * @ipv4_id: Current IPv4 ID, host endian 806 * @packet_space: Remaining space in current packet 807 * @dma_addr: DMA address of current position 808 * @in_len: Remaining length in current SKB fragment 809 * @unmap_len: Length of SKB fragment 810 * @unmap_addr: DMA address of SKB fragment 811 * @dma_flags: TX buffer flags for DMA mapping - %EFX_TX_BUF_MAP_SINGLE or 0 812 * @protocol: Network protocol (after any VLAN header) 813 * @ip_off: Offset of IP header 814 * @tcp_off: Offset of TCP header 815 * @header_len: Number of bytes of header 816 * @ip_base_len: IPv4 tot_len or IPv6 payload_len, before TCP payload 817 * @header_dma_addr: Header DMA address, when using option descriptors 818 * @header_unmap_len: Header DMA mapped length, or 0 if not using option 819 * descriptors 820 * 821 * The state used during segmentation. It is put into this data structure 822 * just to make it easy to pass into inline functions. 823 */ 824 struct tso_state { 825 /* Output position */ 826 unsigned out_len; 827 unsigned seqnum; 828 u16 ipv4_id; 829 unsigned packet_space; 830 831 /* Input position */ 832 dma_addr_t dma_addr; 833 unsigned in_len; 834 unsigned unmap_len; 835 dma_addr_t unmap_addr; 836 unsigned short dma_flags; 837 838 __be16 protocol; 839 unsigned int ip_off; 840 unsigned int tcp_off; 841 unsigned header_len; 842 unsigned int ip_base_len; 843 dma_addr_t header_dma_addr; 844 unsigned int header_unmap_len; 845 }; 846 847 848 /* 849 * Verify that our various assumptions about sk_buffs and the conditions 850 * under which TSO will be attempted hold true. Return the protocol number. 851 */ 852 static __be16 efx_tso_check_protocol(struct sk_buff *skb) 853 { 854 __be16 protocol = skb->protocol; 855 856 EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto != 857 protocol); 858 if (protocol == htons(ETH_P_8021Q)) { 859 struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data; 860 protocol = veh->h_vlan_encapsulated_proto; 861 } 862 863 if (protocol == htons(ETH_P_IP)) { 864 EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP); 865 } else { 866 EFX_BUG_ON_PARANOID(protocol != htons(ETH_P_IPV6)); 867 EFX_BUG_ON_PARANOID(ipv6_hdr(skb)->nexthdr != NEXTHDR_TCP); 868 } 869 EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data) 870 + (tcp_hdr(skb)->doff << 2u)) > 871 skb_headlen(skb)); 872 873 return protocol; 874 } 875 876 static u8 *efx_tsoh_get_buffer(struct efx_tx_queue *tx_queue, 877 struct efx_tx_buffer *buffer, unsigned int len) 878 { 879 u8 *result; 880 881 EFX_BUG_ON_PARANOID(buffer->len); 882 EFX_BUG_ON_PARANOID(buffer->flags); 883 EFX_BUG_ON_PARANOID(buffer->unmap_len); 884 885 if (likely(len <= TSOH_STD_SIZE - TSOH_OFFSET)) { 886 unsigned index = 887 (tx_queue->insert_count & tx_queue->ptr_mask) / 2; 888 struct efx_buffer *page_buf = 889 &tx_queue->tsoh_page[index / TSOH_PER_PAGE]; 890 unsigned offset = 891 TSOH_STD_SIZE * (index % TSOH_PER_PAGE) + TSOH_OFFSET; 892 893 if (unlikely(!page_buf->addr) && 894 efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE, 895 GFP_ATOMIC)) 896 return NULL; 897 898 result = (u8 *)page_buf->addr + offset; 899 buffer->dma_addr = page_buf->dma_addr + offset; 900 buffer->flags = EFX_TX_BUF_CONT; 901 } else { 902 tx_queue->tso_long_headers++; 903 904 buffer->heap_buf = kmalloc(TSOH_OFFSET + len, GFP_ATOMIC); 905 if (unlikely(!buffer->heap_buf)) 906 return NULL; 907 result = (u8 *)buffer->heap_buf + TSOH_OFFSET; 908 buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_HEAP; 909 } 910 911 buffer->len = len; 912 913 return result; 914 } 915 916 /** 917 * efx_tx_queue_insert - push descriptors onto the TX queue 918 * @tx_queue: Efx TX queue 919 * @dma_addr: DMA address of fragment 920 * @len: Length of fragment 921 * @final_buffer: The final buffer inserted into the queue 922 * 923 * Push descriptors onto the TX queue. 924 */ 925 static void efx_tx_queue_insert(struct efx_tx_queue *tx_queue, 926 dma_addr_t dma_addr, unsigned len, 927 struct efx_tx_buffer **final_buffer) 928 { 929 struct efx_tx_buffer *buffer; 930 struct efx_nic *efx = tx_queue->efx; 931 unsigned dma_len; 932 933 EFX_BUG_ON_PARANOID(len <= 0); 934 935 while (1) { 936 buffer = efx_tx_queue_get_insert_buffer(tx_queue); 937 ++tx_queue->insert_count; 938 939 EFX_BUG_ON_PARANOID(tx_queue->insert_count - 940 tx_queue->read_count >= 941 efx->txq_entries); 942 943 buffer->dma_addr = dma_addr; 944 945 dma_len = efx_max_tx_len(efx, dma_addr); 946 947 /* If there is enough space to send then do so */ 948 if (dma_len >= len) 949 break; 950 951 buffer->len = dma_len; 952 buffer->flags = EFX_TX_BUF_CONT; 953 dma_addr += dma_len; 954 len -= dma_len; 955 } 956 957 EFX_BUG_ON_PARANOID(!len); 958 buffer->len = len; 959 *final_buffer = buffer; 960 } 961 962 963 /* 964 * Put a TSO header into the TX queue. 965 * 966 * This is special-cased because we know that it is small enough to fit in 967 * a single fragment, and we know it doesn't cross a page boundary. It 968 * also allows us to not worry about end-of-packet etc. 969 */ 970 static int efx_tso_put_header(struct efx_tx_queue *tx_queue, 971 struct efx_tx_buffer *buffer, u8 *header) 972 { 973 if (unlikely(buffer->flags & EFX_TX_BUF_HEAP)) { 974 buffer->dma_addr = dma_map_single(&tx_queue->efx->pci_dev->dev, 975 header, buffer->len, 976 DMA_TO_DEVICE); 977 if (unlikely(dma_mapping_error(&tx_queue->efx->pci_dev->dev, 978 buffer->dma_addr))) { 979 kfree(buffer->heap_buf); 980 buffer->len = 0; 981 buffer->flags = 0; 982 return -ENOMEM; 983 } 984 buffer->unmap_len = buffer->len; 985 buffer->dma_offset = 0; 986 buffer->flags |= EFX_TX_BUF_MAP_SINGLE; 987 } 988 989 ++tx_queue->insert_count; 990 return 0; 991 } 992 993 994 /* Remove buffers put into a tx_queue. None of the buffers must have 995 * an skb attached. 996 */ 997 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue) 998 { 999 struct efx_tx_buffer *buffer; 1000 1001 /* Work backwards until we hit the original insert pointer value */ 1002 while (tx_queue->insert_count != tx_queue->write_count) { 1003 --tx_queue->insert_count; 1004 buffer = __efx_tx_queue_get_insert_buffer(tx_queue); 1005 efx_dequeue_buffer(tx_queue, buffer, NULL, NULL); 1006 } 1007 } 1008 1009 1010 /* Parse the SKB header and initialise state. */ 1011 static int tso_start(struct tso_state *st, struct efx_nic *efx, 1012 const struct sk_buff *skb) 1013 { 1014 bool use_options = efx_nic_rev(efx) >= EFX_REV_HUNT_A0; 1015 struct device *dma_dev = &efx->pci_dev->dev; 1016 unsigned int header_len, in_len; 1017 dma_addr_t dma_addr; 1018 1019 st->ip_off = skb_network_header(skb) - skb->data; 1020 st->tcp_off = skb_transport_header(skb) - skb->data; 1021 header_len = st->tcp_off + (tcp_hdr(skb)->doff << 2u); 1022 in_len = skb_headlen(skb) - header_len; 1023 st->header_len = header_len; 1024 st->in_len = in_len; 1025 if (st->protocol == htons(ETH_P_IP)) { 1026 st->ip_base_len = st->header_len - st->ip_off; 1027 st->ipv4_id = ntohs(ip_hdr(skb)->id); 1028 } else { 1029 st->ip_base_len = st->header_len - st->tcp_off; 1030 st->ipv4_id = 0; 1031 } 1032 st->seqnum = ntohl(tcp_hdr(skb)->seq); 1033 1034 EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg); 1035 EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn); 1036 EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst); 1037 1038 st->out_len = skb->len - header_len; 1039 1040 if (!use_options) { 1041 st->header_unmap_len = 0; 1042 1043 if (likely(in_len == 0)) { 1044 st->dma_flags = 0; 1045 st->unmap_len = 0; 1046 return 0; 1047 } 1048 1049 dma_addr = dma_map_single(dma_dev, skb->data + header_len, 1050 in_len, DMA_TO_DEVICE); 1051 st->dma_flags = EFX_TX_BUF_MAP_SINGLE; 1052 st->dma_addr = dma_addr; 1053 st->unmap_addr = dma_addr; 1054 st->unmap_len = in_len; 1055 } else { 1056 dma_addr = dma_map_single(dma_dev, skb->data, 1057 skb_headlen(skb), DMA_TO_DEVICE); 1058 st->header_dma_addr = dma_addr; 1059 st->header_unmap_len = skb_headlen(skb); 1060 st->dma_flags = 0; 1061 st->dma_addr = dma_addr + header_len; 1062 st->unmap_len = 0; 1063 } 1064 1065 return unlikely(dma_mapping_error(dma_dev, dma_addr)) ? -ENOMEM : 0; 1066 } 1067 1068 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx, 1069 skb_frag_t *frag) 1070 { 1071 st->unmap_addr = skb_frag_dma_map(&efx->pci_dev->dev, frag, 0, 1072 skb_frag_size(frag), DMA_TO_DEVICE); 1073 if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) { 1074 st->dma_flags = 0; 1075 st->unmap_len = skb_frag_size(frag); 1076 st->in_len = skb_frag_size(frag); 1077 st->dma_addr = st->unmap_addr; 1078 return 0; 1079 } 1080 return -ENOMEM; 1081 } 1082 1083 1084 /** 1085 * tso_fill_packet_with_fragment - form descriptors for the current fragment 1086 * @tx_queue: Efx TX queue 1087 * @skb: Socket buffer 1088 * @st: TSO state 1089 * 1090 * Form descriptors for the current fragment, until we reach the end 1091 * of fragment or end-of-packet. 1092 */ 1093 static void tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue, 1094 const struct sk_buff *skb, 1095 struct tso_state *st) 1096 { 1097 struct efx_tx_buffer *buffer; 1098 int n; 1099 1100 if (st->in_len == 0) 1101 return; 1102 if (st->packet_space == 0) 1103 return; 1104 1105 EFX_BUG_ON_PARANOID(st->in_len <= 0); 1106 EFX_BUG_ON_PARANOID(st->packet_space <= 0); 1107 1108 n = min(st->in_len, st->packet_space); 1109 1110 st->packet_space -= n; 1111 st->out_len -= n; 1112 st->in_len -= n; 1113 1114 efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer); 1115 1116 if (st->out_len == 0) { 1117 /* Transfer ownership of the skb */ 1118 buffer->skb = skb; 1119 buffer->flags = EFX_TX_BUF_SKB; 1120 } else if (st->packet_space != 0) { 1121 buffer->flags = EFX_TX_BUF_CONT; 1122 } 1123 1124 if (st->in_len == 0) { 1125 /* Transfer ownership of the DMA mapping */ 1126 buffer->unmap_len = st->unmap_len; 1127 buffer->dma_offset = buffer->unmap_len - buffer->len; 1128 buffer->flags |= st->dma_flags; 1129 st->unmap_len = 0; 1130 } 1131 1132 st->dma_addr += n; 1133 } 1134 1135 1136 /** 1137 * tso_start_new_packet - generate a new header and prepare for the new packet 1138 * @tx_queue: Efx TX queue 1139 * @skb: Socket buffer 1140 * @st: TSO state 1141 * 1142 * Generate a new header and prepare for the new packet. Return 0 on 1143 * success, or -%ENOMEM if failed to alloc header. 1144 */ 1145 static int tso_start_new_packet(struct efx_tx_queue *tx_queue, 1146 const struct sk_buff *skb, 1147 struct tso_state *st) 1148 { 1149 struct efx_tx_buffer *buffer = 1150 efx_tx_queue_get_insert_buffer(tx_queue); 1151 bool is_last = st->out_len <= skb_shinfo(skb)->gso_size; 1152 u8 tcp_flags_clear; 1153 1154 if (!is_last) { 1155 st->packet_space = skb_shinfo(skb)->gso_size; 1156 tcp_flags_clear = 0x09; /* mask out FIN and PSH */ 1157 } else { 1158 st->packet_space = st->out_len; 1159 tcp_flags_clear = 0x00; 1160 } 1161 1162 if (!st->header_unmap_len) { 1163 /* Allocate and insert a DMA-mapped header buffer. */ 1164 struct tcphdr *tsoh_th; 1165 unsigned ip_length; 1166 u8 *header; 1167 int rc; 1168 1169 header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len); 1170 if (!header) 1171 return -ENOMEM; 1172 1173 tsoh_th = (struct tcphdr *)(header + st->tcp_off); 1174 1175 /* Copy and update the headers. */ 1176 memcpy(header, skb->data, st->header_len); 1177 1178 tsoh_th->seq = htonl(st->seqnum); 1179 ((u8 *)tsoh_th)[13] &= ~tcp_flags_clear; 1180 1181 ip_length = st->ip_base_len + st->packet_space; 1182 1183 if (st->protocol == htons(ETH_P_IP)) { 1184 struct iphdr *tsoh_iph = 1185 (struct iphdr *)(header + st->ip_off); 1186 1187 tsoh_iph->tot_len = htons(ip_length); 1188 tsoh_iph->id = htons(st->ipv4_id); 1189 } else { 1190 struct ipv6hdr *tsoh_iph = 1191 (struct ipv6hdr *)(header + st->ip_off); 1192 1193 tsoh_iph->payload_len = htons(ip_length); 1194 } 1195 1196 rc = efx_tso_put_header(tx_queue, buffer, header); 1197 if (unlikely(rc)) 1198 return rc; 1199 } else { 1200 /* Send the original headers with a TSO option descriptor 1201 * in front 1202 */ 1203 u8 tcp_flags = ((u8 *)tcp_hdr(skb))[13] & ~tcp_flags_clear; 1204 1205 buffer->flags = EFX_TX_BUF_OPTION; 1206 buffer->len = 0; 1207 buffer->unmap_len = 0; 1208 EFX_POPULATE_QWORD_5(buffer->option, 1209 ESF_DZ_TX_DESC_IS_OPT, 1, 1210 ESF_DZ_TX_OPTION_TYPE, 1211 ESE_DZ_TX_OPTION_DESC_TSO, 1212 ESF_DZ_TX_TSO_TCP_FLAGS, tcp_flags, 1213 ESF_DZ_TX_TSO_IP_ID, st->ipv4_id, 1214 ESF_DZ_TX_TSO_TCP_SEQNO, st->seqnum); 1215 ++tx_queue->insert_count; 1216 1217 /* We mapped the headers in tso_start(). Unmap them 1218 * when the last segment is completed. 1219 */ 1220 buffer = efx_tx_queue_get_insert_buffer(tx_queue); 1221 buffer->dma_addr = st->header_dma_addr; 1222 buffer->len = st->header_len; 1223 if (is_last) { 1224 buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_MAP_SINGLE; 1225 buffer->unmap_len = st->header_unmap_len; 1226 buffer->dma_offset = 0; 1227 /* Ensure we only unmap them once in case of a 1228 * later DMA mapping error and rollback 1229 */ 1230 st->header_unmap_len = 0; 1231 } else { 1232 buffer->flags = EFX_TX_BUF_CONT; 1233 buffer->unmap_len = 0; 1234 } 1235 ++tx_queue->insert_count; 1236 } 1237 1238 st->seqnum += skb_shinfo(skb)->gso_size; 1239 1240 /* Linux leaves suitable gaps in the IP ID space for us to fill. */ 1241 ++st->ipv4_id; 1242 1243 ++tx_queue->tso_packets; 1244 1245 return 0; 1246 } 1247 1248 1249 /** 1250 * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer 1251 * @tx_queue: Efx TX queue 1252 * @skb: Socket buffer 1253 * 1254 * Context: You must hold netif_tx_lock() to call this function. 1255 * 1256 * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if 1257 * @skb was not enqueued. In all cases @skb is consumed. Return 1258 * %NETDEV_TX_OK. 1259 */ 1260 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue, 1261 struct sk_buff *skb) 1262 { 1263 struct efx_nic *efx = tx_queue->efx; 1264 int frag_i, rc; 1265 struct tso_state state; 1266 1267 /* Find the packet protocol and sanity-check it */ 1268 state.protocol = efx_tso_check_protocol(skb); 1269 1270 EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count); 1271 1272 rc = tso_start(&state, efx, skb); 1273 if (rc) 1274 goto mem_err; 1275 1276 if (likely(state.in_len == 0)) { 1277 /* Grab the first payload fragment. */ 1278 EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1); 1279 frag_i = 0; 1280 rc = tso_get_fragment(&state, efx, 1281 skb_shinfo(skb)->frags + frag_i); 1282 if (rc) 1283 goto mem_err; 1284 } else { 1285 /* Payload starts in the header area. */ 1286 frag_i = -1; 1287 } 1288 1289 if (tso_start_new_packet(tx_queue, skb, &state) < 0) 1290 goto mem_err; 1291 1292 while (1) { 1293 tso_fill_packet_with_fragment(tx_queue, skb, &state); 1294 1295 /* Move onto the next fragment? */ 1296 if (state.in_len == 0) { 1297 if (++frag_i >= skb_shinfo(skb)->nr_frags) 1298 /* End of payload reached. */ 1299 break; 1300 rc = tso_get_fragment(&state, efx, 1301 skb_shinfo(skb)->frags + frag_i); 1302 if (rc) 1303 goto mem_err; 1304 } 1305 1306 /* Start at new packet? */ 1307 if (state.packet_space == 0 && 1308 tso_start_new_packet(tx_queue, skb, &state) < 0) 1309 goto mem_err; 1310 } 1311 1312 netdev_tx_sent_queue(tx_queue->core_txq, skb->len); 1313 1314 /* Pass off to hardware */ 1315 efx_nic_push_buffers(tx_queue); 1316 1317 efx_tx_maybe_stop_queue(tx_queue); 1318 1319 tx_queue->tso_bursts++; 1320 return NETDEV_TX_OK; 1321 1322 mem_err: 1323 netif_err(efx, tx_err, efx->net_dev, 1324 "Out of memory for TSO headers, or DMA mapping error\n"); 1325 dev_kfree_skb_any(skb); 1326 1327 /* Free the DMA mapping we were in the process of writing out */ 1328 if (state.unmap_len) { 1329 if (state.dma_flags & EFX_TX_BUF_MAP_SINGLE) 1330 dma_unmap_single(&efx->pci_dev->dev, state.unmap_addr, 1331 state.unmap_len, DMA_TO_DEVICE); 1332 else 1333 dma_unmap_page(&efx->pci_dev->dev, state.unmap_addr, 1334 state.unmap_len, DMA_TO_DEVICE); 1335 } 1336 1337 /* Free the header DMA mapping, if using option descriptors */ 1338 if (state.header_unmap_len) 1339 dma_unmap_single(&efx->pci_dev->dev, state.header_dma_addr, 1340 state.header_unmap_len, DMA_TO_DEVICE); 1341 1342 efx_enqueue_unwind(tx_queue); 1343 return NETDEV_TX_OK; 1344 } 1345