1 // SPDX-License-Identifier: GPL-2.0-only 2 /**************************************************************************** 3 * Driver for Solarflare network controllers and boards 4 * Copyright 2005-2006 Fen Systems Ltd. 5 * Copyright 2005-2013 Solarflare Communications Inc. 6 */ 7 8 #include <linux/socket.h> 9 #include <linux/in.h> 10 #include <linux/slab.h> 11 #include <linux/ip.h> 12 #include <linux/ipv6.h> 13 #include <linux/tcp.h> 14 #include <linux/udp.h> 15 #include <linux/prefetch.h> 16 #include <linux/moduleparam.h> 17 #include <linux/iommu.h> 18 #include <net/ip.h> 19 #include <net/checksum.h> 20 #include "net_driver.h" 21 #include "efx.h" 22 #include "filter.h" 23 #include "nic.h" 24 #include "selftest.h" 25 #include "workarounds.h" 26 27 /* Preferred number of descriptors to fill at once */ 28 #define EF4_RX_PREFERRED_BATCH 8U 29 30 /* Number of RX buffers to recycle pages for. When creating the RX page recycle 31 * ring, this number is divided by the number of buffers per page to calculate 32 * the number of pages to store in the RX page recycle ring. 33 */ 34 #define EF4_RECYCLE_RING_SIZE_IOMMU 4096 35 #define EF4_RECYCLE_RING_SIZE_NOIOMMU (2 * EF4_RX_PREFERRED_BATCH) 36 37 /* Size of buffer allocated for skb header area. */ 38 #define EF4_SKB_HEADERS 128u 39 40 /* This is the percentage fill level below which new RX descriptors 41 * will be added to the RX descriptor ring. 42 */ 43 static unsigned int rx_refill_threshold; 44 45 /* Each packet can consume up to ceil(max_frame_len / buffer_size) buffers */ 46 #define EF4_RX_MAX_FRAGS DIV_ROUND_UP(EF4_MAX_FRAME_LEN(EF4_MAX_MTU), \ 47 EF4_RX_USR_BUF_SIZE) 48 49 /* 50 * RX maximum head room required. 51 * 52 * This must be at least 1 to prevent overflow, plus one packet-worth 53 * to allow pipelined receives. 54 */ 55 #define EF4_RXD_HEAD_ROOM (1 + EF4_RX_MAX_FRAGS) 56 57 static inline u8 *ef4_rx_buf_va(struct ef4_rx_buffer *buf) 58 { 59 return page_address(buf->page) + buf->page_offset; 60 } 61 62 static inline u32 ef4_rx_buf_hash(struct ef4_nic *efx, const u8 *eh) 63 { 64 #if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) 65 return __le32_to_cpup((const __le32 *)(eh + efx->rx_packet_hash_offset)); 66 #else 67 const u8 *data = eh + efx->rx_packet_hash_offset; 68 return (u32)data[0] | 69 (u32)data[1] << 8 | 70 (u32)data[2] << 16 | 71 (u32)data[3] << 24; 72 #endif 73 } 74 75 static inline struct ef4_rx_buffer * 76 ef4_rx_buf_next(struct ef4_rx_queue *rx_queue, struct ef4_rx_buffer *rx_buf) 77 { 78 if (unlikely(rx_buf == ef4_rx_buffer(rx_queue, rx_queue->ptr_mask))) 79 return ef4_rx_buffer(rx_queue, 0); 80 else 81 return rx_buf + 1; 82 } 83 84 static inline void ef4_sync_rx_buffer(struct ef4_nic *efx, 85 struct ef4_rx_buffer *rx_buf, 86 unsigned int len) 87 { 88 dma_sync_single_for_cpu(&efx->pci_dev->dev, rx_buf->dma_addr, len, 89 DMA_FROM_DEVICE); 90 } 91 92 void ef4_rx_config_page_split(struct ef4_nic *efx) 93 { 94 efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align, 95 EF4_RX_BUF_ALIGNMENT); 96 efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 : 97 ((PAGE_SIZE - sizeof(struct ef4_rx_page_state)) / 98 efx->rx_page_buf_step); 99 efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) / 100 efx->rx_bufs_per_page; 101 efx->rx_pages_per_batch = DIV_ROUND_UP(EF4_RX_PREFERRED_BATCH, 102 efx->rx_bufs_per_page); 103 } 104 105 /* Check the RX page recycle ring for a page that can be reused. */ 106 static struct page *ef4_reuse_page(struct ef4_rx_queue *rx_queue) 107 { 108 struct ef4_nic *efx = rx_queue->efx; 109 struct page *page; 110 struct ef4_rx_page_state *state; 111 unsigned index; 112 113 if (unlikely(!rx_queue->page_ring)) 114 return NULL; 115 index = rx_queue->page_remove & rx_queue->page_ptr_mask; 116 page = rx_queue->page_ring[index]; 117 if (page == NULL) 118 return NULL; 119 120 rx_queue->page_ring[index] = NULL; 121 /* page_remove cannot exceed page_add. */ 122 if (rx_queue->page_remove != rx_queue->page_add) 123 ++rx_queue->page_remove; 124 125 /* If page_count is 1 then we hold the only reference to this page. */ 126 if (page_count(page) == 1) { 127 ++rx_queue->page_recycle_count; 128 return page; 129 } else { 130 state = page_address(page); 131 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr, 132 PAGE_SIZE << efx->rx_buffer_order, 133 DMA_FROM_DEVICE); 134 put_page(page); 135 ++rx_queue->page_recycle_failed; 136 } 137 138 return NULL; 139 } 140 141 /** 142 * ef4_init_rx_buffers - create EF4_RX_BATCH page-based RX buffers 143 * 144 * @rx_queue: Efx RX queue 145 * @atomic: control memory allocation flags 146 * 147 * This allocates a batch of pages, maps them for DMA, and populates 148 * struct ef4_rx_buffers for each one. Return a negative error code or 149 * 0 on success. If a single page can be used for multiple buffers, 150 * then the page will either be inserted fully, or not at all. 151 */ 152 static int ef4_init_rx_buffers(struct ef4_rx_queue *rx_queue, bool atomic) 153 { 154 struct ef4_nic *efx = rx_queue->efx; 155 struct ef4_rx_buffer *rx_buf; 156 struct page *page; 157 unsigned int page_offset; 158 struct ef4_rx_page_state *state; 159 dma_addr_t dma_addr; 160 unsigned index, count; 161 162 count = 0; 163 do { 164 page = ef4_reuse_page(rx_queue); 165 if (page == NULL) { 166 page = alloc_pages(__GFP_COMP | 167 (atomic ? GFP_ATOMIC : GFP_KERNEL), 168 efx->rx_buffer_order); 169 if (unlikely(page == NULL)) 170 return -ENOMEM; 171 dma_addr = 172 dma_map_page(&efx->pci_dev->dev, page, 0, 173 PAGE_SIZE << efx->rx_buffer_order, 174 DMA_FROM_DEVICE); 175 if (unlikely(dma_mapping_error(&efx->pci_dev->dev, 176 dma_addr))) { 177 __free_pages(page, efx->rx_buffer_order); 178 return -EIO; 179 } 180 state = page_address(page); 181 state->dma_addr = dma_addr; 182 } else { 183 state = page_address(page); 184 dma_addr = state->dma_addr; 185 } 186 187 dma_addr += sizeof(struct ef4_rx_page_state); 188 page_offset = sizeof(struct ef4_rx_page_state); 189 190 do { 191 index = rx_queue->added_count & rx_queue->ptr_mask; 192 rx_buf = ef4_rx_buffer(rx_queue, index); 193 rx_buf->dma_addr = dma_addr + efx->rx_ip_align; 194 rx_buf->page = page; 195 rx_buf->page_offset = page_offset + efx->rx_ip_align; 196 rx_buf->len = efx->rx_dma_len; 197 rx_buf->flags = 0; 198 ++rx_queue->added_count; 199 get_page(page); 200 dma_addr += efx->rx_page_buf_step; 201 page_offset += efx->rx_page_buf_step; 202 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE); 203 204 rx_buf->flags = EF4_RX_BUF_LAST_IN_PAGE; 205 } while (++count < efx->rx_pages_per_batch); 206 207 return 0; 208 } 209 210 /* Unmap a DMA-mapped page. This function is only called for the final RX 211 * buffer in a page. 212 */ 213 static void ef4_unmap_rx_buffer(struct ef4_nic *efx, 214 struct ef4_rx_buffer *rx_buf) 215 { 216 struct page *page = rx_buf->page; 217 218 if (page) { 219 struct ef4_rx_page_state *state = page_address(page); 220 dma_unmap_page(&efx->pci_dev->dev, 221 state->dma_addr, 222 PAGE_SIZE << efx->rx_buffer_order, 223 DMA_FROM_DEVICE); 224 } 225 } 226 227 static void ef4_free_rx_buffers(struct ef4_rx_queue *rx_queue, 228 struct ef4_rx_buffer *rx_buf, 229 unsigned int num_bufs) 230 { 231 do { 232 if (rx_buf->page) { 233 put_page(rx_buf->page); 234 rx_buf->page = NULL; 235 } 236 rx_buf = ef4_rx_buf_next(rx_queue, rx_buf); 237 } while (--num_bufs); 238 } 239 240 /* Attempt to recycle the page if there is an RX recycle ring; the page can 241 * only be added if this is the final RX buffer, to prevent pages being used in 242 * the descriptor ring and appearing in the recycle ring simultaneously. 243 */ 244 static void ef4_recycle_rx_page(struct ef4_channel *channel, 245 struct ef4_rx_buffer *rx_buf) 246 { 247 struct page *page = rx_buf->page; 248 struct ef4_rx_queue *rx_queue = ef4_channel_get_rx_queue(channel); 249 struct ef4_nic *efx = rx_queue->efx; 250 unsigned index; 251 252 /* Only recycle the page after processing the final buffer. */ 253 if (!(rx_buf->flags & EF4_RX_BUF_LAST_IN_PAGE)) 254 return; 255 256 index = rx_queue->page_add & rx_queue->page_ptr_mask; 257 if (rx_queue->page_ring[index] == NULL) { 258 unsigned read_index = rx_queue->page_remove & 259 rx_queue->page_ptr_mask; 260 261 /* The next slot in the recycle ring is available, but 262 * increment page_remove if the read pointer currently 263 * points here. 264 */ 265 if (read_index == index) 266 ++rx_queue->page_remove; 267 rx_queue->page_ring[index] = page; 268 ++rx_queue->page_add; 269 return; 270 } 271 ++rx_queue->page_recycle_full; 272 ef4_unmap_rx_buffer(efx, rx_buf); 273 put_page(rx_buf->page); 274 } 275 276 static void ef4_fini_rx_buffer(struct ef4_rx_queue *rx_queue, 277 struct ef4_rx_buffer *rx_buf) 278 { 279 /* Release the page reference we hold for the buffer. */ 280 if (rx_buf->page) 281 put_page(rx_buf->page); 282 283 /* If this is the last buffer in a page, unmap and free it. */ 284 if (rx_buf->flags & EF4_RX_BUF_LAST_IN_PAGE) { 285 ef4_unmap_rx_buffer(rx_queue->efx, rx_buf); 286 ef4_free_rx_buffers(rx_queue, rx_buf, 1); 287 } 288 rx_buf->page = NULL; 289 } 290 291 /* Recycle the pages that are used by buffers that have just been received. */ 292 static void ef4_recycle_rx_pages(struct ef4_channel *channel, 293 struct ef4_rx_buffer *rx_buf, 294 unsigned int n_frags) 295 { 296 struct ef4_rx_queue *rx_queue = ef4_channel_get_rx_queue(channel); 297 298 if (unlikely(!rx_queue->page_ring)) 299 return; 300 301 do { 302 ef4_recycle_rx_page(channel, rx_buf); 303 rx_buf = ef4_rx_buf_next(rx_queue, rx_buf); 304 } while (--n_frags); 305 } 306 307 static void ef4_discard_rx_packet(struct ef4_channel *channel, 308 struct ef4_rx_buffer *rx_buf, 309 unsigned int n_frags) 310 { 311 struct ef4_rx_queue *rx_queue = ef4_channel_get_rx_queue(channel); 312 313 ef4_recycle_rx_pages(channel, rx_buf, n_frags); 314 315 ef4_free_rx_buffers(rx_queue, rx_buf, n_frags); 316 } 317 318 /** 319 * ef4_fast_push_rx_descriptors - push new RX descriptors quickly 320 * @rx_queue: RX descriptor queue 321 * 322 * This will aim to fill the RX descriptor queue up to 323 * @rx_queue->@max_fill. If there is insufficient atomic 324 * memory to do so, a slow fill will be scheduled. 325 * @atomic: control memory allocation flags 326 * 327 * The caller must provide serialisation (none is used here). In practise, 328 * this means this function must run from the NAPI handler, or be called 329 * when NAPI is disabled. 330 */ 331 void ef4_fast_push_rx_descriptors(struct ef4_rx_queue *rx_queue, bool atomic) 332 { 333 struct ef4_nic *efx = rx_queue->efx; 334 unsigned int fill_level, batch_size; 335 int space, rc = 0; 336 337 if (!rx_queue->refill_enabled) 338 return; 339 340 /* Calculate current fill level, and exit if we don't need to fill */ 341 fill_level = (rx_queue->added_count - rx_queue->removed_count); 342 EF4_BUG_ON_PARANOID(fill_level > rx_queue->efx->rxq_entries); 343 if (fill_level >= rx_queue->fast_fill_trigger) 344 goto out; 345 346 /* Record minimum fill level */ 347 if (unlikely(fill_level < rx_queue->min_fill)) { 348 if (fill_level) 349 rx_queue->min_fill = fill_level; 350 } 351 352 batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page; 353 space = rx_queue->max_fill - fill_level; 354 EF4_BUG_ON_PARANOID(space < batch_size); 355 356 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev, 357 "RX queue %d fast-filling descriptor ring from" 358 " level %d to level %d\n", 359 ef4_rx_queue_index(rx_queue), fill_level, 360 rx_queue->max_fill); 361 362 363 do { 364 rc = ef4_init_rx_buffers(rx_queue, atomic); 365 if (unlikely(rc)) { 366 /* Ensure that we don't leave the rx queue empty */ 367 if (rx_queue->added_count == rx_queue->removed_count) 368 ef4_schedule_slow_fill(rx_queue); 369 goto out; 370 } 371 } while ((space -= batch_size) >= batch_size); 372 373 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev, 374 "RX queue %d fast-filled descriptor ring " 375 "to level %d\n", ef4_rx_queue_index(rx_queue), 376 rx_queue->added_count - rx_queue->removed_count); 377 378 out: 379 if (rx_queue->notified_count != rx_queue->added_count) 380 ef4_nic_notify_rx_desc(rx_queue); 381 } 382 383 void ef4_rx_slow_fill(struct timer_list *t) 384 { 385 struct ef4_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill); 386 387 /* Post an event to cause NAPI to run and refill the queue */ 388 ef4_nic_generate_fill_event(rx_queue); 389 ++rx_queue->slow_fill_count; 390 } 391 392 static void ef4_rx_packet__check_len(struct ef4_rx_queue *rx_queue, 393 struct ef4_rx_buffer *rx_buf, 394 int len) 395 { 396 struct ef4_nic *efx = rx_queue->efx; 397 unsigned max_len = rx_buf->len - efx->type->rx_buffer_padding; 398 399 if (likely(len <= max_len)) 400 return; 401 402 /* The packet must be discarded, but this is only a fatal error 403 * if the caller indicated it was 404 */ 405 rx_buf->flags |= EF4_RX_PKT_DISCARD; 406 407 if ((len > rx_buf->len) && EF4_WORKAROUND_8071(efx)) { 408 if (net_ratelimit()) 409 netif_err(efx, rx_err, efx->net_dev, 410 " RX queue %d seriously overlength " 411 "RX event (0x%x > 0x%x+0x%x). Leaking\n", 412 ef4_rx_queue_index(rx_queue), len, max_len, 413 efx->type->rx_buffer_padding); 414 ef4_schedule_reset(efx, RESET_TYPE_RX_RECOVERY); 415 } else { 416 if (net_ratelimit()) 417 netif_err(efx, rx_err, efx->net_dev, 418 " RX queue %d overlength RX event " 419 "(0x%x > 0x%x)\n", 420 ef4_rx_queue_index(rx_queue), len, max_len); 421 } 422 423 ef4_rx_queue_channel(rx_queue)->n_rx_overlength++; 424 } 425 426 /* Pass a received packet up through GRO. GRO can handle pages 427 * regardless of checksum state and skbs with a good checksum. 428 */ 429 static void 430 ef4_rx_packet_gro(struct ef4_channel *channel, struct ef4_rx_buffer *rx_buf, 431 unsigned int n_frags, u8 *eh) 432 { 433 struct napi_struct *napi = &channel->napi_str; 434 struct ef4_nic *efx = channel->efx; 435 struct sk_buff *skb; 436 437 skb = napi_get_frags(napi); 438 if (unlikely(!skb)) { 439 struct ef4_rx_queue *rx_queue; 440 441 rx_queue = ef4_channel_get_rx_queue(channel); 442 ef4_free_rx_buffers(rx_queue, rx_buf, n_frags); 443 return; 444 } 445 446 if (efx->net_dev->features & NETIF_F_RXHASH) 447 skb_set_hash(skb, ef4_rx_buf_hash(efx, eh), 448 PKT_HASH_TYPE_L3); 449 skb->ip_summed = ((rx_buf->flags & EF4_RX_PKT_CSUMMED) ? 450 CHECKSUM_UNNECESSARY : CHECKSUM_NONE); 451 452 for (;;) { 453 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags, 454 rx_buf->page, rx_buf->page_offset, 455 rx_buf->len); 456 rx_buf->page = NULL; 457 skb->len += rx_buf->len; 458 if (skb_shinfo(skb)->nr_frags == n_frags) 459 break; 460 461 rx_buf = ef4_rx_buf_next(&channel->rx_queue, rx_buf); 462 } 463 464 skb->data_len = skb->len; 465 skb->truesize += n_frags * efx->rx_buffer_truesize; 466 467 skb_record_rx_queue(skb, channel->rx_queue.core_index); 468 469 napi_gro_frags(napi); 470 } 471 472 /* Allocate and construct an SKB around page fragments */ 473 static struct sk_buff *ef4_rx_mk_skb(struct ef4_channel *channel, 474 struct ef4_rx_buffer *rx_buf, 475 unsigned int n_frags, 476 u8 *eh, int hdr_len) 477 { 478 struct ef4_nic *efx = channel->efx; 479 struct sk_buff *skb; 480 481 /* Allocate an SKB to store the headers */ 482 skb = netdev_alloc_skb(efx->net_dev, 483 efx->rx_ip_align + efx->rx_prefix_size + 484 hdr_len); 485 if (unlikely(skb == NULL)) { 486 atomic_inc(&efx->n_rx_noskb_drops); 487 return NULL; 488 } 489 490 EF4_BUG_ON_PARANOID(rx_buf->len < hdr_len); 491 492 memcpy(skb->data + efx->rx_ip_align, eh - efx->rx_prefix_size, 493 efx->rx_prefix_size + hdr_len); 494 skb_reserve(skb, efx->rx_ip_align + efx->rx_prefix_size); 495 __skb_put(skb, hdr_len); 496 497 /* Append the remaining page(s) onto the frag list */ 498 if (rx_buf->len > hdr_len) { 499 rx_buf->page_offset += hdr_len; 500 rx_buf->len -= hdr_len; 501 502 for (;;) { 503 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags, 504 rx_buf->page, rx_buf->page_offset, 505 rx_buf->len); 506 rx_buf->page = NULL; 507 skb->len += rx_buf->len; 508 skb->data_len += rx_buf->len; 509 if (skb_shinfo(skb)->nr_frags == n_frags) 510 break; 511 512 rx_buf = ef4_rx_buf_next(&channel->rx_queue, rx_buf); 513 } 514 } else { 515 __free_pages(rx_buf->page, efx->rx_buffer_order); 516 rx_buf->page = NULL; 517 n_frags = 0; 518 } 519 520 skb->truesize += n_frags * efx->rx_buffer_truesize; 521 522 /* Move past the ethernet header */ 523 skb->protocol = eth_type_trans(skb, efx->net_dev); 524 525 skb_mark_napi_id(skb, &channel->napi_str); 526 527 return skb; 528 } 529 530 void ef4_rx_packet(struct ef4_rx_queue *rx_queue, unsigned int index, 531 unsigned int n_frags, unsigned int len, u16 flags) 532 { 533 struct ef4_nic *efx = rx_queue->efx; 534 struct ef4_channel *channel = ef4_rx_queue_channel(rx_queue); 535 struct ef4_rx_buffer *rx_buf; 536 537 rx_queue->rx_packets++; 538 539 rx_buf = ef4_rx_buffer(rx_queue, index); 540 rx_buf->flags |= flags; 541 542 /* Validate the number of fragments and completed length */ 543 if (n_frags == 1) { 544 if (!(flags & EF4_RX_PKT_PREFIX_LEN)) 545 ef4_rx_packet__check_len(rx_queue, rx_buf, len); 546 } else if (unlikely(n_frags > EF4_RX_MAX_FRAGS) || 547 unlikely(len <= (n_frags - 1) * efx->rx_dma_len) || 548 unlikely(len > n_frags * efx->rx_dma_len) || 549 unlikely(!efx->rx_scatter)) { 550 /* If this isn't an explicit discard request, either 551 * the hardware or the driver is broken. 552 */ 553 WARN_ON(!(len == 0 && rx_buf->flags & EF4_RX_PKT_DISCARD)); 554 rx_buf->flags |= EF4_RX_PKT_DISCARD; 555 } 556 557 netif_vdbg(efx, rx_status, efx->net_dev, 558 "RX queue %d received ids %x-%x len %d %s%s\n", 559 ef4_rx_queue_index(rx_queue), index, 560 (index + n_frags - 1) & rx_queue->ptr_mask, len, 561 (rx_buf->flags & EF4_RX_PKT_CSUMMED) ? " [SUMMED]" : "", 562 (rx_buf->flags & EF4_RX_PKT_DISCARD) ? " [DISCARD]" : ""); 563 564 /* Discard packet, if instructed to do so. Process the 565 * previous receive first. 566 */ 567 if (unlikely(rx_buf->flags & EF4_RX_PKT_DISCARD)) { 568 ef4_rx_flush_packet(channel); 569 ef4_discard_rx_packet(channel, rx_buf, n_frags); 570 return; 571 } 572 573 if (n_frags == 1 && !(flags & EF4_RX_PKT_PREFIX_LEN)) 574 rx_buf->len = len; 575 576 /* Release and/or sync the DMA mapping - assumes all RX buffers 577 * consumed in-order per RX queue. 578 */ 579 ef4_sync_rx_buffer(efx, rx_buf, rx_buf->len); 580 581 /* Prefetch nice and early so data will (hopefully) be in cache by 582 * the time we look at it. 583 */ 584 prefetch(ef4_rx_buf_va(rx_buf)); 585 586 rx_buf->page_offset += efx->rx_prefix_size; 587 rx_buf->len -= efx->rx_prefix_size; 588 589 if (n_frags > 1) { 590 /* Release/sync DMA mapping for additional fragments. 591 * Fix length for last fragment. 592 */ 593 unsigned int tail_frags = n_frags - 1; 594 595 for (;;) { 596 rx_buf = ef4_rx_buf_next(rx_queue, rx_buf); 597 if (--tail_frags == 0) 598 break; 599 ef4_sync_rx_buffer(efx, rx_buf, efx->rx_dma_len); 600 } 601 rx_buf->len = len - (n_frags - 1) * efx->rx_dma_len; 602 ef4_sync_rx_buffer(efx, rx_buf, rx_buf->len); 603 } 604 605 /* All fragments have been DMA-synced, so recycle pages. */ 606 rx_buf = ef4_rx_buffer(rx_queue, index); 607 ef4_recycle_rx_pages(channel, rx_buf, n_frags); 608 609 /* Pipeline receives so that we give time for packet headers to be 610 * prefetched into cache. 611 */ 612 ef4_rx_flush_packet(channel); 613 channel->rx_pkt_n_frags = n_frags; 614 channel->rx_pkt_index = index; 615 } 616 617 static void ef4_rx_deliver(struct ef4_channel *channel, u8 *eh, 618 struct ef4_rx_buffer *rx_buf, 619 unsigned int n_frags) 620 { 621 struct sk_buff *skb; 622 u16 hdr_len = min_t(u16, rx_buf->len, EF4_SKB_HEADERS); 623 624 skb = ef4_rx_mk_skb(channel, rx_buf, n_frags, eh, hdr_len); 625 if (unlikely(skb == NULL)) { 626 struct ef4_rx_queue *rx_queue; 627 628 rx_queue = ef4_channel_get_rx_queue(channel); 629 ef4_free_rx_buffers(rx_queue, rx_buf, n_frags); 630 return; 631 } 632 skb_record_rx_queue(skb, channel->rx_queue.core_index); 633 634 /* Set the SKB flags */ 635 skb_checksum_none_assert(skb); 636 if (likely(rx_buf->flags & EF4_RX_PKT_CSUMMED)) 637 skb->ip_summed = CHECKSUM_UNNECESSARY; 638 639 if (channel->type->receive_skb) 640 if (channel->type->receive_skb(channel, skb)) 641 return; 642 643 /* Pass the packet up */ 644 netif_receive_skb(skb); 645 } 646 647 /* Handle a received packet. Second half: Touches packet payload. */ 648 void __ef4_rx_packet(struct ef4_channel *channel) 649 { 650 struct ef4_nic *efx = channel->efx; 651 struct ef4_rx_buffer *rx_buf = 652 ef4_rx_buffer(&channel->rx_queue, channel->rx_pkt_index); 653 u8 *eh = ef4_rx_buf_va(rx_buf); 654 655 /* Read length from the prefix if necessary. This already 656 * excludes the length of the prefix itself. 657 */ 658 if (rx_buf->flags & EF4_RX_PKT_PREFIX_LEN) 659 rx_buf->len = le16_to_cpup((__le16 *) 660 (eh + efx->rx_packet_len_offset)); 661 662 /* If we're in loopback test, then pass the packet directly to the 663 * loopback layer, and free the rx_buf here 664 */ 665 if (unlikely(efx->loopback_selftest)) { 666 struct ef4_rx_queue *rx_queue; 667 668 ef4_loopback_rx_packet(efx, eh, rx_buf->len); 669 rx_queue = ef4_channel_get_rx_queue(channel); 670 ef4_free_rx_buffers(rx_queue, rx_buf, 671 channel->rx_pkt_n_frags); 672 goto out; 673 } 674 675 if (unlikely(!(efx->net_dev->features & NETIF_F_RXCSUM))) 676 rx_buf->flags &= ~EF4_RX_PKT_CSUMMED; 677 678 if ((rx_buf->flags & EF4_RX_PKT_TCP) && !channel->type->receive_skb) 679 ef4_rx_packet_gro(channel, rx_buf, channel->rx_pkt_n_frags, eh); 680 else 681 ef4_rx_deliver(channel, eh, rx_buf, channel->rx_pkt_n_frags); 682 out: 683 channel->rx_pkt_n_frags = 0; 684 } 685 686 int ef4_probe_rx_queue(struct ef4_rx_queue *rx_queue) 687 { 688 struct ef4_nic *efx = rx_queue->efx; 689 unsigned int entries; 690 int rc; 691 692 /* Create the smallest power-of-two aligned ring */ 693 entries = max(roundup_pow_of_two(efx->rxq_entries), EF4_MIN_DMAQ_SIZE); 694 EF4_BUG_ON_PARANOID(entries > EF4_MAX_DMAQ_SIZE); 695 rx_queue->ptr_mask = entries - 1; 696 697 netif_dbg(efx, probe, efx->net_dev, 698 "creating RX queue %d size %#x mask %#x\n", 699 ef4_rx_queue_index(rx_queue), efx->rxq_entries, 700 rx_queue->ptr_mask); 701 702 /* Allocate RX buffers */ 703 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer), 704 GFP_KERNEL); 705 if (!rx_queue->buffer) 706 return -ENOMEM; 707 708 rc = ef4_nic_probe_rx(rx_queue); 709 if (rc) { 710 kfree(rx_queue->buffer); 711 rx_queue->buffer = NULL; 712 } 713 714 return rc; 715 } 716 717 static void ef4_init_rx_recycle_ring(struct ef4_nic *efx, 718 struct ef4_rx_queue *rx_queue) 719 { 720 unsigned int bufs_in_recycle_ring, page_ring_size; 721 struct iommu_domain __maybe_unused *domain; 722 723 /* Set the RX recycle ring size */ 724 #ifdef CONFIG_PPC64 725 bufs_in_recycle_ring = EF4_RECYCLE_RING_SIZE_IOMMU; 726 #else 727 domain = iommu_get_domain_for_dev(&efx->pci_dev->dev); 728 if (domain && domain->type != IOMMU_DOMAIN_IDENTITY) 729 bufs_in_recycle_ring = EF4_RECYCLE_RING_SIZE_IOMMU; 730 else 731 bufs_in_recycle_ring = EF4_RECYCLE_RING_SIZE_NOIOMMU; 732 #endif /* CONFIG_PPC64 */ 733 734 page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring / 735 efx->rx_bufs_per_page); 736 rx_queue->page_ring = kcalloc(page_ring_size, 737 sizeof(*rx_queue->page_ring), GFP_KERNEL); 738 if (!rx_queue->page_ring) 739 rx_queue->page_ptr_mask = 0; 740 else 741 rx_queue->page_ptr_mask = page_ring_size - 1; 742 } 743 744 void ef4_init_rx_queue(struct ef4_rx_queue *rx_queue) 745 { 746 struct ef4_nic *efx = rx_queue->efx; 747 unsigned int max_fill, trigger, max_trigger; 748 749 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 750 "initialising RX queue %d\n", ef4_rx_queue_index(rx_queue)); 751 752 /* Initialise ptr fields */ 753 rx_queue->added_count = 0; 754 rx_queue->notified_count = 0; 755 rx_queue->removed_count = 0; 756 rx_queue->min_fill = -1U; 757 ef4_init_rx_recycle_ring(efx, rx_queue); 758 759 rx_queue->page_remove = 0; 760 rx_queue->page_add = rx_queue->page_ptr_mask + 1; 761 rx_queue->page_recycle_count = 0; 762 rx_queue->page_recycle_failed = 0; 763 rx_queue->page_recycle_full = 0; 764 765 /* Initialise limit fields */ 766 max_fill = efx->rxq_entries - EF4_RXD_HEAD_ROOM; 767 max_trigger = 768 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page; 769 if (rx_refill_threshold != 0) { 770 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U; 771 if (trigger > max_trigger) 772 trigger = max_trigger; 773 } else { 774 trigger = max_trigger; 775 } 776 777 rx_queue->max_fill = max_fill; 778 rx_queue->fast_fill_trigger = trigger; 779 rx_queue->refill_enabled = true; 780 781 /* Set up RX descriptor ring */ 782 ef4_nic_init_rx(rx_queue); 783 } 784 785 void ef4_fini_rx_queue(struct ef4_rx_queue *rx_queue) 786 { 787 int i; 788 struct ef4_nic *efx = rx_queue->efx; 789 struct ef4_rx_buffer *rx_buf; 790 791 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 792 "shutting down RX queue %d\n", ef4_rx_queue_index(rx_queue)); 793 794 del_timer_sync(&rx_queue->slow_fill); 795 796 /* Release RX buffers from the current read ptr to the write ptr */ 797 if (rx_queue->buffer) { 798 for (i = rx_queue->removed_count; i < rx_queue->added_count; 799 i++) { 800 unsigned index = i & rx_queue->ptr_mask; 801 rx_buf = ef4_rx_buffer(rx_queue, index); 802 ef4_fini_rx_buffer(rx_queue, rx_buf); 803 } 804 } 805 806 /* Unmap and release the pages in the recycle ring. Remove the ring. */ 807 for (i = 0; i <= rx_queue->page_ptr_mask; i++) { 808 struct page *page = rx_queue->page_ring[i]; 809 struct ef4_rx_page_state *state; 810 811 if (page == NULL) 812 continue; 813 814 state = page_address(page); 815 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr, 816 PAGE_SIZE << efx->rx_buffer_order, 817 DMA_FROM_DEVICE); 818 put_page(page); 819 } 820 kfree(rx_queue->page_ring); 821 rx_queue->page_ring = NULL; 822 } 823 824 void ef4_remove_rx_queue(struct ef4_rx_queue *rx_queue) 825 { 826 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 827 "destroying RX queue %d\n", ef4_rx_queue_index(rx_queue)); 828 829 ef4_nic_remove_rx(rx_queue); 830 831 kfree(rx_queue->buffer); 832 rx_queue->buffer = NULL; 833 } 834 835 836 module_param(rx_refill_threshold, uint, 0444); 837 MODULE_PARM_DESC(rx_refill_threshold, 838 "RX descriptor ring refill threshold (%)"); 839 840 #ifdef CONFIG_RFS_ACCEL 841 842 int ef4_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb, 843 u16 rxq_index, u32 flow_id) 844 { 845 struct ef4_nic *efx = netdev_priv(net_dev); 846 struct ef4_channel *channel; 847 struct ef4_filter_spec spec; 848 struct flow_keys fk; 849 int rc; 850 851 if (flow_id == RPS_FLOW_ID_INVALID) 852 return -EINVAL; 853 854 if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) 855 return -EPROTONOSUPPORT; 856 857 if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) 858 return -EPROTONOSUPPORT; 859 if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) 860 return -EPROTONOSUPPORT; 861 862 ef4_filter_init_rx(&spec, EF4_FILTER_PRI_HINT, 863 efx->rx_scatter ? EF4_FILTER_FLAG_RX_SCATTER : 0, 864 rxq_index); 865 spec.match_flags = 866 EF4_FILTER_MATCH_ETHER_TYPE | EF4_FILTER_MATCH_IP_PROTO | 867 EF4_FILTER_MATCH_LOC_HOST | EF4_FILTER_MATCH_LOC_PORT | 868 EF4_FILTER_MATCH_REM_HOST | EF4_FILTER_MATCH_REM_PORT; 869 spec.ether_type = fk.basic.n_proto; 870 spec.ip_proto = fk.basic.ip_proto; 871 872 if (fk.basic.n_proto == htons(ETH_P_IP)) { 873 spec.rem_host[0] = fk.addrs.v4addrs.src; 874 spec.loc_host[0] = fk.addrs.v4addrs.dst; 875 } else { 876 memcpy(spec.rem_host, &fk.addrs.v6addrs.src, sizeof(struct in6_addr)); 877 memcpy(spec.loc_host, &fk.addrs.v6addrs.dst, sizeof(struct in6_addr)); 878 } 879 880 spec.rem_port = fk.ports.src; 881 spec.loc_port = fk.ports.dst; 882 883 rc = efx->type->filter_rfs_insert(efx, &spec); 884 if (rc < 0) 885 return rc; 886 887 /* Remember this so we can check whether to expire the filter later */ 888 channel = ef4_get_channel(efx, rxq_index); 889 channel->rps_flow_id[rc] = flow_id; 890 ++channel->rfs_filters_added; 891 892 if (spec.ether_type == htons(ETH_P_IP)) 893 netif_info(efx, rx_status, efx->net_dev, 894 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d]\n", 895 (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 896 spec.rem_host, ntohs(spec.rem_port), spec.loc_host, 897 ntohs(spec.loc_port), rxq_index, flow_id, rc); 898 else 899 netif_info(efx, rx_status, efx->net_dev, 900 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d]\n", 901 (spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 902 spec.rem_host, ntohs(spec.rem_port), spec.loc_host, 903 ntohs(spec.loc_port), rxq_index, flow_id, rc); 904 905 return rc; 906 } 907 908 bool __ef4_filter_rfs_expire(struct ef4_nic *efx, unsigned int quota) 909 { 910 bool (*expire_one)(struct ef4_nic *efx, u32 flow_id, unsigned int index); 911 unsigned int channel_idx, index, size; 912 u32 flow_id; 913 914 if (!spin_trylock_bh(&efx->filter_lock)) 915 return false; 916 917 expire_one = efx->type->filter_rfs_expire_one; 918 channel_idx = efx->rps_expire_channel; 919 index = efx->rps_expire_index; 920 size = efx->type->max_rx_ip_filters; 921 while (quota--) { 922 struct ef4_channel *channel = ef4_get_channel(efx, channel_idx); 923 flow_id = channel->rps_flow_id[index]; 924 925 if (flow_id != RPS_FLOW_ID_INVALID && 926 expire_one(efx, flow_id, index)) { 927 netif_info(efx, rx_status, efx->net_dev, 928 "expired filter %d [queue %u flow %u]\n", 929 index, channel_idx, flow_id); 930 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID; 931 } 932 if (++index == size) { 933 if (++channel_idx == efx->n_channels) 934 channel_idx = 0; 935 index = 0; 936 } 937 } 938 efx->rps_expire_channel = channel_idx; 939 efx->rps_expire_index = index; 940 941 spin_unlock_bh(&efx->filter_lock); 942 return true; 943 } 944 945 #endif /* CONFIG_RFS_ACCEL */ 946 947 /** 948 * ef4_filter_is_mc_recipient - test whether spec is a multicast recipient 949 * @spec: Specification to test 950 * 951 * Return: %true if the specification is a non-drop RX filter that 952 * matches a local MAC address I/G bit value of 1 or matches a local 953 * IPv4 or IPv6 address value in the respective multicast address 954 * range. Otherwise %false. 955 */ 956 bool ef4_filter_is_mc_recipient(const struct ef4_filter_spec *spec) 957 { 958 if (!(spec->flags & EF4_FILTER_FLAG_RX) || 959 spec->dmaq_id == EF4_FILTER_RX_DMAQ_ID_DROP) 960 return false; 961 962 if (spec->match_flags & 963 (EF4_FILTER_MATCH_LOC_MAC | EF4_FILTER_MATCH_LOC_MAC_IG) && 964 is_multicast_ether_addr(spec->loc_mac)) 965 return true; 966 967 if ((spec->match_flags & 968 (EF4_FILTER_MATCH_ETHER_TYPE | EF4_FILTER_MATCH_LOC_HOST)) == 969 (EF4_FILTER_MATCH_ETHER_TYPE | EF4_FILTER_MATCH_LOC_HOST)) { 970 if (spec->ether_type == htons(ETH_P_IP) && 971 ipv4_is_multicast(spec->loc_host[0])) 972 return true; 973 if (spec->ether_type == htons(ETH_P_IPV6) && 974 ((const u8 *)spec->loc_host)[0] == 0xff) 975 return true; 976 } 977 978 return false; 979 } 980