1// SPDX-License-Identifier: GPL-2.0-only 2/**************************************************************************** 3 * Driver for Solarflare network controllers and boards 4 * Copyright 2018 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 "net_driver.h" 12#include <linux/module.h> 13#include <linux/iommu.h> 14#include <net/rps.h> 15#include "efx.h" 16#include "nic.h" 17#include "rx_common.h" 18 19/* This is the percentage fill level below which new RX descriptors 20 * will be added to the RX descriptor ring. 21 */ 22static unsigned int rx_refill_threshold; 23module_param(rx_refill_threshold, uint, 0444); 24MODULE_PARM_DESC(rx_refill_threshold, 25 "RX descriptor ring refill threshold (%)"); 26 27/* RX maximum head room required. 28 * 29 * This must be at least 1 to prevent overflow, plus one packet-worth 30 * to allow pipelined receives. 31 */ 32#define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS) 33 34/* Check the RX page recycle ring for a page that can be reused. */ 35static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue) 36{ 37 struct efx_nic *efx = rx_queue->efx; 38 struct efx_rx_page_state *state; 39 unsigned int index; 40 struct page *page; 41 42 if (unlikely(!rx_queue->page_ring)) 43 return NULL; 44 index = rx_queue->page_remove & rx_queue->page_ptr_mask; 45 page = rx_queue->page_ring[index]; 46 if (page == NULL) 47 return NULL; 48 49 rx_queue->page_ring[index] = NULL; 50 /* page_remove cannot exceed page_add. */ 51 if (rx_queue->page_remove != rx_queue->page_add) 52 ++rx_queue->page_remove; 53 54 /* If page_count is 1 then we hold the only reference to this page. */ 55 if (page_count(page) == 1) { 56 ++rx_queue->page_recycle_count; 57 return page; 58 } else { 59 state = page_address(page); 60 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr, 61 PAGE_SIZE << efx->rx_buffer_order, 62 DMA_FROM_DEVICE); 63 put_page(page); 64 ++rx_queue->page_recycle_failed; 65 } 66 67 return NULL; 68} 69 70/* Attempt to recycle the page if there is an RX recycle ring; the page can 71 * only be added if this is the final RX buffer, to prevent pages being used in 72 * the descriptor ring and appearing in the recycle ring simultaneously. 73 */ 74static void efx_recycle_rx_page(struct efx_channel *channel, 75 struct efx_rx_buffer *rx_buf) 76{ 77 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 78 struct efx_nic *efx = rx_queue->efx; 79 struct page *page = rx_buf->page; 80 unsigned int index; 81 82 /* Only recycle the page after processing the final buffer. */ 83 if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE)) 84 return; 85 86 index = rx_queue->page_add & rx_queue->page_ptr_mask; 87 if (rx_queue->page_ring[index] == NULL) { 88 unsigned int read_index = rx_queue->page_remove & 89 rx_queue->page_ptr_mask; 90 91 /* The next slot in the recycle ring is available, but 92 * increment page_remove if the read pointer currently 93 * points here. 94 */ 95 if (read_index == index) 96 ++rx_queue->page_remove; 97 rx_queue->page_ring[index] = page; 98 ++rx_queue->page_add; 99 return; 100 } 101 ++rx_queue->page_recycle_full; 102 efx_unmap_rx_buffer(efx, rx_buf); 103 put_page(rx_buf->page); 104} 105 106/* Recycle the pages that are used by buffers that have just been received. */ 107void efx_recycle_rx_pages(struct efx_channel *channel, 108 struct efx_rx_buffer *rx_buf, 109 unsigned int n_frags) 110{ 111 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 112 113 if (unlikely(!rx_queue->page_ring)) 114 return; 115 116 do { 117 efx_recycle_rx_page(channel, rx_buf); 118 rx_buf = efx_rx_buf_next(rx_queue, rx_buf); 119 } while (--n_frags); 120} 121 122void efx_discard_rx_packet(struct efx_channel *channel, 123 struct efx_rx_buffer *rx_buf, 124 unsigned int n_frags) 125{ 126 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 127 128 efx_recycle_rx_pages(channel, rx_buf, n_frags); 129 130 efx_free_rx_buffers(rx_queue, rx_buf, n_frags); 131} 132 133static void efx_init_rx_recycle_ring(struct efx_rx_queue *rx_queue) 134{ 135 unsigned int bufs_in_recycle_ring, page_ring_size; 136 struct efx_nic *efx = rx_queue->efx; 137 138 bufs_in_recycle_ring = efx_rx_recycle_ring_size(efx); 139 page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring / 140 efx->rx_bufs_per_page); 141 rx_queue->page_ring = kcalloc(page_ring_size, 142 sizeof(*rx_queue->page_ring), GFP_KERNEL); 143 if (!rx_queue->page_ring) 144 rx_queue->page_ptr_mask = 0; 145 else 146 rx_queue->page_ptr_mask = page_ring_size - 1; 147} 148 149static void efx_fini_rx_recycle_ring(struct efx_rx_queue *rx_queue) 150{ 151 struct efx_nic *efx = rx_queue->efx; 152 int i; 153 154 if (unlikely(!rx_queue->page_ring)) 155 return; 156 157 /* Unmap and release the pages in the recycle ring. Remove the ring. */ 158 for (i = 0; i <= rx_queue->page_ptr_mask; i++) { 159 struct page *page = rx_queue->page_ring[i]; 160 struct efx_rx_page_state *state; 161 162 if (page == NULL) 163 continue; 164 165 state = page_address(page); 166 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr, 167 PAGE_SIZE << efx->rx_buffer_order, 168 DMA_FROM_DEVICE); 169 put_page(page); 170 } 171 kfree(rx_queue->page_ring); 172 rx_queue->page_ring = NULL; 173} 174 175static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue, 176 struct efx_rx_buffer *rx_buf) 177{ 178 /* Release the page reference we hold for the buffer. */ 179 if (rx_buf->page) 180 put_page(rx_buf->page); 181 182 /* If this is the last buffer in a page, unmap and free it. */ 183 if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) { 184 efx_unmap_rx_buffer(rx_queue->efx, rx_buf); 185 efx_free_rx_buffers(rx_queue, rx_buf, 1); 186 } 187 rx_buf->page = NULL; 188} 189 190int efx_probe_rx_queue(struct efx_rx_queue *rx_queue) 191{ 192 struct efx_nic *efx = rx_queue->efx; 193 unsigned int entries; 194 int rc; 195 196 /* Create the smallest power-of-two aligned ring */ 197 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE); 198 EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE); 199 rx_queue->ptr_mask = entries - 1; 200 201 netif_dbg(efx, probe, efx->net_dev, 202 "creating RX queue %d size %#x mask %#x\n", 203 efx_rx_queue_index(rx_queue), efx->rxq_entries, 204 rx_queue->ptr_mask); 205 206 /* Allocate RX buffers */ 207 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer), 208 GFP_KERNEL); 209 if (!rx_queue->buffer) 210 return -ENOMEM; 211 212 rc = efx_nic_probe_rx(rx_queue); 213 if (rc) { 214 kfree(rx_queue->buffer); 215 rx_queue->buffer = NULL; 216 } 217 218 return rc; 219} 220 221void efx_init_rx_queue(struct efx_rx_queue *rx_queue) 222{ 223 unsigned int max_fill, trigger, max_trigger; 224 struct efx_nic *efx = rx_queue->efx; 225 int rc = 0; 226 227 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 228 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue)); 229 230 /* Initialise ptr fields */ 231 rx_queue->added_count = 0; 232 rx_queue->notified_count = 0; 233 rx_queue->granted_count = 0; 234 rx_queue->removed_count = 0; 235 rx_queue->min_fill = -1U; 236 efx_init_rx_recycle_ring(rx_queue); 237 238 rx_queue->page_remove = 0; 239 rx_queue->page_add = rx_queue->page_ptr_mask + 1; 240 rx_queue->page_recycle_count = 0; 241 rx_queue->page_recycle_failed = 0; 242 rx_queue->page_recycle_full = 0; 243 244 /* Initialise limit fields */ 245 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM; 246 max_trigger = 247 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page; 248 if (rx_refill_threshold != 0) { 249 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U; 250 if (trigger > max_trigger) 251 trigger = max_trigger; 252 } else { 253 trigger = max_trigger; 254 } 255 256 rx_queue->max_fill = max_fill; 257 rx_queue->fast_fill_trigger = trigger; 258 rx_queue->refill_enabled = true; 259 260 /* Initialise XDP queue information */ 261 rc = xdp_rxq_info_reg(&rx_queue->xdp_rxq_info, efx->net_dev, 262 rx_queue->core_index, 0); 263 264 if (rc) { 265 netif_err(efx, rx_err, efx->net_dev, 266 "Failure to initialise XDP queue information rc=%d\n", 267 rc); 268 efx->xdp_rxq_info_failed = true; 269 } else { 270 rx_queue->xdp_rxq_info_valid = true; 271 } 272 273 /* Set up RX descriptor ring */ 274 efx_nic_init_rx(rx_queue); 275} 276 277void efx_fini_rx_queue(struct efx_rx_queue *rx_queue) 278{ 279 struct efx_rx_buffer *rx_buf; 280 int i; 281 282 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 283 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue)); 284 285 del_timer_sync(&rx_queue->slow_fill); 286 if (rx_queue->grant_credits) 287 flush_work(&rx_queue->grant_work); 288 289 /* Release RX buffers from the current read ptr to the write ptr */ 290 if (rx_queue->buffer) { 291 for (i = rx_queue->removed_count; i < rx_queue->added_count; 292 i++) { 293 unsigned int index = i & rx_queue->ptr_mask; 294 295 rx_buf = efx_rx_buffer(rx_queue, index); 296 efx_fini_rx_buffer(rx_queue, rx_buf); 297 } 298 } 299 300 efx_fini_rx_recycle_ring(rx_queue); 301 302 if (rx_queue->xdp_rxq_info_valid) 303 xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info); 304 305 rx_queue->xdp_rxq_info_valid = false; 306} 307 308void efx_remove_rx_queue(struct efx_rx_queue *rx_queue) 309{ 310 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 311 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue)); 312 313 efx_nic_remove_rx(rx_queue); 314 315 kfree(rx_queue->buffer); 316 rx_queue->buffer = NULL; 317} 318 319/* Unmap a DMA-mapped page. This function is only called for the final RX 320 * buffer in a page. 321 */ 322void efx_unmap_rx_buffer(struct efx_nic *efx, 323 struct efx_rx_buffer *rx_buf) 324{ 325 struct page *page = rx_buf->page; 326 327 if (page) { 328 struct efx_rx_page_state *state = page_address(page); 329 330 dma_unmap_page(&efx->pci_dev->dev, 331 state->dma_addr, 332 PAGE_SIZE << efx->rx_buffer_order, 333 DMA_FROM_DEVICE); 334 } 335} 336 337void efx_free_rx_buffers(struct efx_rx_queue *rx_queue, 338 struct efx_rx_buffer *rx_buf, 339 unsigned int num_bufs) 340{ 341 do { 342 if (rx_buf->page) { 343 put_page(rx_buf->page); 344 rx_buf->page = NULL; 345 } 346 rx_buf = efx_rx_buf_next(rx_queue, rx_buf); 347 } while (--num_bufs); 348} 349 350void efx_rx_slow_fill(struct timer_list *t) 351{ 352 struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill); 353 354 /* Post an event to cause NAPI to run and refill the queue */ 355 efx_nic_generate_fill_event(rx_queue); 356 ++rx_queue->slow_fill_count; 357} 358 359void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue) 360{ 361 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(10)); 362} 363 364/* efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers 365 * 366 * @rx_queue: Efx RX queue 367 * 368 * This allocates a batch of pages, maps them for DMA, and populates 369 * struct efx_rx_buffers for each one. Return a negative error code or 370 * 0 on success. If a single page can be used for multiple buffers, 371 * then the page will either be inserted fully, or not at all. 372 */ 373static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic) 374{ 375 unsigned int page_offset, index, count; 376 struct efx_nic *efx = rx_queue->efx; 377 struct efx_rx_page_state *state; 378 struct efx_rx_buffer *rx_buf; 379 dma_addr_t dma_addr; 380 struct page *page; 381 382 count = 0; 383 do { 384 page = efx_reuse_page(rx_queue); 385 if (page == NULL) { 386 page = alloc_pages(__GFP_COMP | 387 (atomic ? GFP_ATOMIC : GFP_KERNEL), 388 efx->rx_buffer_order); 389 if (unlikely(page == NULL)) 390 return -ENOMEM; 391 dma_addr = 392 dma_map_page(&efx->pci_dev->dev, page, 0, 393 PAGE_SIZE << efx->rx_buffer_order, 394 DMA_FROM_DEVICE); 395 if (unlikely(dma_mapping_error(&efx->pci_dev->dev, 396 dma_addr))) { 397 __free_pages(page, efx->rx_buffer_order); 398 return -EIO; 399 } 400 state = page_address(page); 401 state->dma_addr = dma_addr; 402 } else { 403 state = page_address(page); 404 dma_addr = state->dma_addr; 405 } 406 407 dma_addr += sizeof(struct efx_rx_page_state); 408 page_offset = sizeof(struct efx_rx_page_state); 409 410 do { 411 index = rx_queue->added_count & rx_queue->ptr_mask; 412 rx_buf = efx_rx_buffer(rx_queue, index); 413 rx_buf->dma_addr = dma_addr + efx->rx_ip_align + 414 EFX_XDP_HEADROOM; 415 rx_buf->page = page; 416 rx_buf->page_offset = page_offset + efx->rx_ip_align + 417 EFX_XDP_HEADROOM; 418 rx_buf->len = efx->rx_dma_len; 419 rx_buf->flags = 0; 420 ++rx_queue->added_count; 421 get_page(page); 422 dma_addr += efx->rx_page_buf_step; 423 page_offset += efx->rx_page_buf_step; 424 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE); 425 426 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE; 427 } while (++count < efx->rx_pages_per_batch); 428 429 return 0; 430} 431 432void efx_rx_config_page_split(struct efx_nic *efx) 433{ 434 efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align + 435 EFX_XDP_HEADROOM + EFX_XDP_TAILROOM, 436 EFX_RX_BUF_ALIGNMENT); 437 efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 : 438 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) / 439 efx->rx_page_buf_step); 440 efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) / 441 efx->rx_bufs_per_page; 442 efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH, 443 efx->rx_bufs_per_page); 444} 445 446/* efx_fast_push_rx_descriptors - push new RX descriptors quickly 447 * @rx_queue: RX descriptor queue 448 * 449 * This will aim to fill the RX descriptor queue up to 450 * @rx_queue->@max_fill. If there is insufficient atomic 451 * memory to do so, a slow fill will be scheduled. 452 * 453 * The caller must provide serialisation (none is used here). In practise, 454 * this means this function must run from the NAPI handler, or be called 455 * when NAPI is disabled. 456 */ 457void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic) 458{ 459 struct efx_nic *efx = rx_queue->efx; 460 unsigned int fill_level, batch_size; 461 int space, rc = 0; 462 463 if (!rx_queue->refill_enabled) 464 return; 465 466 /* Calculate current fill level, and exit if we don't need to fill */ 467 fill_level = (rx_queue->added_count - rx_queue->removed_count); 468 EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries); 469 if (fill_level >= rx_queue->fast_fill_trigger) 470 goto out; 471 472 /* Record minimum fill level */ 473 if (unlikely(fill_level < rx_queue->min_fill)) { 474 if (fill_level) 475 rx_queue->min_fill = fill_level; 476 } 477 478 batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page; 479 space = rx_queue->max_fill - fill_level; 480 EFX_WARN_ON_ONCE_PARANOID(space < batch_size); 481 482 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev, 483 "RX queue %d fast-filling descriptor ring from" 484 " level %d to level %d\n", 485 efx_rx_queue_index(rx_queue), fill_level, 486 rx_queue->max_fill); 487 488 do { 489 rc = efx_init_rx_buffers(rx_queue, atomic); 490 if (unlikely(rc)) { 491 /* Ensure that we don't leave the rx queue empty */ 492 efx_schedule_slow_fill(rx_queue); 493 goto out; 494 } 495 } while ((space -= batch_size) >= batch_size); 496 497 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev, 498 "RX queue %d fast-filled descriptor ring " 499 "to level %d\n", efx_rx_queue_index(rx_queue), 500 rx_queue->added_count - rx_queue->removed_count); 501 502 out: 503 if (rx_queue->notified_count != rx_queue->added_count) 504 efx_nic_notify_rx_desc(rx_queue); 505} 506 507/* Pass a received packet up through GRO. GRO can handle pages 508 * regardless of checksum state and skbs with a good checksum. 509 */ 510void 511efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf, 512 unsigned int n_frags, u8 *eh, __wsum csum) 513{ 514 struct napi_struct *napi = &channel->napi_str; 515 struct efx_nic *efx = channel->efx; 516 struct sk_buff *skb; 517 518 skb = napi_get_frags(napi); 519 if (unlikely(!skb)) { 520 struct efx_rx_queue *rx_queue; 521 522 rx_queue = efx_channel_get_rx_queue(channel); 523 efx_free_rx_buffers(rx_queue, rx_buf, n_frags); 524 return; 525 } 526 527 if (efx->net_dev->features & NETIF_F_RXHASH && 528 efx_rx_buf_hash_valid(efx, eh)) 529 skb_set_hash(skb, efx_rx_buf_hash(efx, eh), 530 PKT_HASH_TYPE_L3); 531 if (csum) { 532 skb->csum = csum; 533 skb->ip_summed = CHECKSUM_COMPLETE; 534 } else { 535 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ? 536 CHECKSUM_UNNECESSARY : CHECKSUM_NONE); 537 } 538 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL); 539 540 for (;;) { 541 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags, 542 rx_buf->page, rx_buf->page_offset, 543 rx_buf->len); 544 rx_buf->page = NULL; 545 skb->len += rx_buf->len; 546 if (skb_shinfo(skb)->nr_frags == n_frags) 547 break; 548 549 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf); 550 } 551 552 skb->data_len = skb->len; 553 skb->truesize += n_frags * efx->rx_buffer_truesize; 554 555 skb_record_rx_queue(skb, channel->rx_queue.core_index); 556 557 napi_gro_frags(napi); 558} 559 560/* RSS contexts. We're using linked lists and crappy O(n) algorithms, because 561 * (a) this is an infrequent control-plane operation and (b) n is small (max 64) 562 */ 563struct efx_rss_context *efx_alloc_rss_context_entry(struct efx_nic *efx) 564{ 565 struct list_head *head = &efx->rss_context.list; 566 struct efx_rss_context *ctx, *new; 567 u32 id = 1; /* Don't use zero, that refers to the master RSS context */ 568 569 WARN_ON(!mutex_is_locked(&efx->rss_lock)); 570 571 /* Search for first gap in the numbering */ 572 list_for_each_entry(ctx, head, list) { 573 if (ctx->user_id != id) 574 break; 575 id++; 576 /* Check for wrap. If this happens, we have nearly 2^32 577 * allocated RSS contexts, which seems unlikely. 578 */ 579 if (WARN_ON_ONCE(!id)) 580 return NULL; 581 } 582 583 /* Create the new entry */ 584 new = kmalloc(sizeof(*new), GFP_KERNEL); 585 if (!new) 586 return NULL; 587 new->context_id = EFX_MCDI_RSS_CONTEXT_INVALID; 588 new->rx_hash_udp_4tuple = false; 589 590 /* Insert the new entry into the gap */ 591 new->user_id = id; 592 list_add_tail(&new->list, &ctx->list); 593 return new; 594} 595 596struct efx_rss_context *efx_find_rss_context_entry(struct efx_nic *efx, u32 id) 597{ 598 struct list_head *head = &efx->rss_context.list; 599 struct efx_rss_context *ctx; 600 601 WARN_ON(!mutex_is_locked(&efx->rss_lock)); 602 603 list_for_each_entry(ctx, head, list) 604 if (ctx->user_id == id) 605 return ctx; 606 return NULL; 607} 608 609void efx_free_rss_context_entry(struct efx_rss_context *ctx) 610{ 611 list_del(&ctx->list); 612 kfree(ctx); 613} 614 615void efx_set_default_rx_indir_table(struct efx_nic *efx, 616 struct efx_rss_context *ctx) 617{ 618 size_t i; 619 620 for (i = 0; i < ARRAY_SIZE(ctx->rx_indir_table); i++) 621 ctx->rx_indir_table[i] = 622 ethtool_rxfh_indir_default(i, efx->rss_spread); 623} 624 625/** 626 * efx_filter_is_mc_recipient - test whether spec is a multicast recipient 627 * @spec: Specification to test 628 * 629 * Return: %true if the specification is a non-drop RX filter that 630 * matches a local MAC address I/G bit value of 1 or matches a local 631 * IPv4 or IPv6 address value in the respective multicast address 632 * range. Otherwise %false. 633 */ 634bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec) 635{ 636 if (!(spec->flags & EFX_FILTER_FLAG_RX) || 637 spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP) 638 return false; 639 640 if (spec->match_flags & 641 (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) && 642 is_multicast_ether_addr(spec->loc_mac)) 643 return true; 644 645 if ((spec->match_flags & 646 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) == 647 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) { 648 if (spec->ether_type == htons(ETH_P_IP) && 649 ipv4_is_multicast(spec->loc_host[0])) 650 return true; 651 if (spec->ether_type == htons(ETH_P_IPV6) && 652 ((const u8 *)spec->loc_host)[0] == 0xff) 653 return true; 654 } 655 656 return false; 657} 658 659bool efx_filter_spec_equal(const struct efx_filter_spec *left, 660 const struct efx_filter_spec *right) 661{ 662 if ((left->match_flags ^ right->match_flags) | 663 ((left->flags ^ right->flags) & 664 (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX))) 665 return false; 666 667 return memcmp(&left->vport_id, &right->vport_id, 668 sizeof(struct efx_filter_spec) - 669 offsetof(struct efx_filter_spec, vport_id)) == 0; 670} 671 672u32 efx_filter_spec_hash(const struct efx_filter_spec *spec) 673{ 674 BUILD_BUG_ON(offsetof(struct efx_filter_spec, vport_id) & 3); 675 return jhash2((const u32 *)&spec->vport_id, 676 (sizeof(struct efx_filter_spec) - 677 offsetof(struct efx_filter_spec, vport_id)) / 4, 678 0); 679} 680 681#ifdef CONFIG_RFS_ACCEL 682bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx, 683 bool *force) 684{ 685 if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) { 686 /* ARFS is currently updating this entry, leave it */ 687 return false; 688 } 689 if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) { 690 /* ARFS tried and failed to update this, so it's probably out 691 * of date. Remove the filter and the ARFS rule entry. 692 */ 693 rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING; 694 *force = true; 695 return true; 696 } else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */ 697 /* ARFS has moved on, so old filter is not needed. Since we did 698 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will 699 * not be removed by efx_rps_hash_del() subsequently. 700 */ 701 *force = true; 702 return true; 703 } 704 /* Remove it iff ARFS wants to. */ 705 return true; 706} 707 708static 709struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx, 710 const struct efx_filter_spec *spec) 711{ 712 u32 hash = efx_filter_spec_hash(spec); 713 714 lockdep_assert_held(&efx->rps_hash_lock); 715 if (!efx->rps_hash_table) 716 return NULL; 717 return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE]; 718} 719 720struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx, 721 const struct efx_filter_spec *spec) 722{ 723 struct efx_arfs_rule *rule; 724 struct hlist_head *head; 725 struct hlist_node *node; 726 727 head = efx_rps_hash_bucket(efx, spec); 728 if (!head) 729 return NULL; 730 hlist_for_each(node, head) { 731 rule = container_of(node, struct efx_arfs_rule, node); 732 if (efx_filter_spec_equal(spec, &rule->spec)) 733 return rule; 734 } 735 return NULL; 736} 737 738struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx, 739 const struct efx_filter_spec *spec, 740 bool *new) 741{ 742 struct efx_arfs_rule *rule; 743 struct hlist_head *head; 744 struct hlist_node *node; 745 746 head = efx_rps_hash_bucket(efx, spec); 747 if (!head) 748 return NULL; 749 hlist_for_each(node, head) { 750 rule = container_of(node, struct efx_arfs_rule, node); 751 if (efx_filter_spec_equal(spec, &rule->spec)) { 752 *new = false; 753 return rule; 754 } 755 } 756 rule = kmalloc(sizeof(*rule), GFP_ATOMIC); 757 *new = true; 758 if (rule) { 759 memcpy(&rule->spec, spec, sizeof(rule->spec)); 760 hlist_add_head(&rule->node, head); 761 } 762 return rule; 763} 764 765void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec) 766{ 767 struct efx_arfs_rule *rule; 768 struct hlist_head *head; 769 struct hlist_node *node; 770 771 head = efx_rps_hash_bucket(efx, spec); 772 if (WARN_ON(!head)) 773 return; 774 hlist_for_each(node, head) { 775 rule = container_of(node, struct efx_arfs_rule, node); 776 if (efx_filter_spec_equal(spec, &rule->spec)) { 777 /* Someone already reused the entry. We know that if 778 * this check doesn't fire (i.e. filter_id == REMOVING) 779 * then the REMOVING mark was put there by our caller, 780 * because caller is holding a lock on filter table and 781 * only holders of that lock set REMOVING. 782 */ 783 if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING) 784 return; 785 hlist_del(node); 786 kfree(rule); 787 return; 788 } 789 } 790 /* We didn't find it. */ 791 WARN_ON(1); 792} 793#endif 794 795int efx_probe_filters(struct efx_nic *efx) 796{ 797 int rc; 798 799 mutex_lock(&efx->mac_lock); 800 rc = efx->type->filter_table_probe(efx); 801 if (rc) 802 goto out_unlock; 803 804#ifdef CONFIG_RFS_ACCEL 805 if (efx->type->offload_features & NETIF_F_NTUPLE) { 806 struct efx_channel *channel; 807 int i, success = 1; 808 809 efx_for_each_channel(channel, efx) { 810 channel->rps_flow_id = 811 kcalloc(efx->type->max_rx_ip_filters, 812 sizeof(*channel->rps_flow_id), 813 GFP_KERNEL); 814 if (!channel->rps_flow_id) 815 success = 0; 816 else 817 for (i = 0; 818 i < efx->type->max_rx_ip_filters; 819 ++i) 820 channel->rps_flow_id[i] = 821 RPS_FLOW_ID_INVALID; 822 channel->rfs_expire_index = 0; 823 channel->rfs_filter_count = 0; 824 } 825 826 if (!success) { 827 efx_for_each_channel(channel, efx) { 828 kfree(channel->rps_flow_id); 829 channel->rps_flow_id = NULL; 830 } 831 efx->type->filter_table_remove(efx); 832 rc = -ENOMEM; 833 goto out_unlock; 834 } 835 } 836#endif 837out_unlock: 838 mutex_unlock(&efx->mac_lock); 839 return rc; 840} 841 842void efx_remove_filters(struct efx_nic *efx) 843{ 844#ifdef CONFIG_RFS_ACCEL 845 struct efx_channel *channel; 846 847 efx_for_each_channel(channel, efx) { 848 cancel_delayed_work_sync(&channel->filter_work); 849 kfree(channel->rps_flow_id); 850 channel->rps_flow_id = NULL; 851 } 852#endif 853 efx->type->filter_table_remove(efx); 854} 855 856#ifdef CONFIG_RFS_ACCEL 857 858static void efx_filter_rfs_work(struct work_struct *data) 859{ 860 struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion, 861 work); 862 struct efx_nic *efx = efx_netdev_priv(req->net_dev); 863 struct efx_channel *channel = efx_get_channel(efx, req->rxq_index); 864 int slot_idx = req - efx->rps_slot; 865 struct efx_arfs_rule *rule; 866 u16 arfs_id = 0; 867 int rc; 868 869 rc = efx->type->filter_insert(efx, &req->spec, true); 870 if (rc >= 0) 871 /* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */ 872 rc %= efx->type->max_rx_ip_filters; 873 if (efx->rps_hash_table) { 874 spin_lock_bh(&efx->rps_hash_lock); 875 rule = efx_rps_hash_find(efx, &req->spec); 876 /* The rule might have already gone, if someone else's request 877 * for the same spec was already worked and then expired before 878 * we got around to our work. In that case we have nothing 879 * tying us to an arfs_id, meaning that as soon as the filter 880 * is considered for expiry it will be removed. 881 */ 882 if (rule) { 883 if (rc < 0) 884 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR; 885 else 886 rule->filter_id = rc; 887 arfs_id = rule->arfs_id; 888 } 889 spin_unlock_bh(&efx->rps_hash_lock); 890 } 891 if (rc >= 0) { 892 /* Remember this so we can check whether to expire the filter 893 * later. 894 */ 895 mutex_lock(&efx->rps_mutex); 896 if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID) 897 channel->rfs_filter_count++; 898 channel->rps_flow_id[rc] = req->flow_id; 899 mutex_unlock(&efx->rps_mutex); 900 901 if (req->spec.ether_type == htons(ETH_P_IP)) 902 netif_info(efx, rx_status, efx->net_dev, 903 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n", 904 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 905 req->spec.rem_host, ntohs(req->spec.rem_port), 906 req->spec.loc_host, ntohs(req->spec.loc_port), 907 req->rxq_index, req->flow_id, rc, arfs_id); 908 else 909 netif_info(efx, rx_status, efx->net_dev, 910 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n", 911 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 912 req->spec.rem_host, ntohs(req->spec.rem_port), 913 req->spec.loc_host, ntohs(req->spec.loc_port), 914 req->rxq_index, req->flow_id, rc, arfs_id); 915 channel->n_rfs_succeeded++; 916 } else { 917 if (req->spec.ether_type == htons(ETH_P_IP)) 918 netif_dbg(efx, rx_status, efx->net_dev, 919 "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n", 920 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 921 req->spec.rem_host, ntohs(req->spec.rem_port), 922 req->spec.loc_host, ntohs(req->spec.loc_port), 923 req->rxq_index, req->flow_id, rc, arfs_id); 924 else 925 netif_dbg(efx, rx_status, efx->net_dev, 926 "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n", 927 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 928 req->spec.rem_host, ntohs(req->spec.rem_port), 929 req->spec.loc_host, ntohs(req->spec.loc_port), 930 req->rxq_index, req->flow_id, rc, arfs_id); 931 channel->n_rfs_failed++; 932 /* We're overloading the NIC's filter tables, so let's do a 933 * chunk of extra expiry work. 934 */ 935 __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count, 936 100u)); 937 } 938 939 /* Release references */ 940 clear_bit(slot_idx, &efx->rps_slot_map); 941 dev_put(req->net_dev); 942} 943 944int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb, 945 u16 rxq_index, u32 flow_id) 946{ 947 struct efx_nic *efx = efx_netdev_priv(net_dev); 948 struct efx_async_filter_insertion *req; 949 struct efx_arfs_rule *rule; 950 struct flow_keys fk; 951 int slot_idx; 952 bool new; 953 int rc; 954 955 /* find a free slot */ 956 for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++) 957 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map)) 958 break; 959 if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT) 960 return -EBUSY; 961 962 if (flow_id == RPS_FLOW_ID_INVALID) { 963 rc = -EINVAL; 964 goto out_clear; 965 } 966 967 if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) { 968 rc = -EPROTONOSUPPORT; 969 goto out_clear; 970 } 971 972 if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) { 973 rc = -EPROTONOSUPPORT; 974 goto out_clear; 975 } 976 if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) { 977 rc = -EPROTONOSUPPORT; 978 goto out_clear; 979 } 980 981 req = efx->rps_slot + slot_idx; 982 efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT, 983 efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0, 984 rxq_index); 985 req->spec.match_flags = 986 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO | 987 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT | 988 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT; 989 req->spec.ether_type = fk.basic.n_proto; 990 req->spec.ip_proto = fk.basic.ip_proto; 991 992 if (fk.basic.n_proto == htons(ETH_P_IP)) { 993 req->spec.rem_host[0] = fk.addrs.v4addrs.src; 994 req->spec.loc_host[0] = fk.addrs.v4addrs.dst; 995 } else { 996 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src, 997 sizeof(struct in6_addr)); 998 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst, 999 sizeof(struct in6_addr)); 1000 } 1001 1002 req->spec.rem_port = fk.ports.src; 1003 req->spec.loc_port = fk.ports.dst; 1004 1005 if (efx->rps_hash_table) { 1006 /* Add it to ARFS hash table */ 1007 spin_lock(&efx->rps_hash_lock); 1008 rule = efx_rps_hash_add(efx, &req->spec, &new); 1009 if (!rule) { 1010 rc = -ENOMEM; 1011 goto out_unlock; 1012 } 1013 if (new) 1014 rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER; 1015 rc = rule->arfs_id; 1016 /* Skip if existing or pending filter already does the right thing */ 1017 if (!new && rule->rxq_index == rxq_index && 1018 rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING) 1019 goto out_unlock; 1020 rule->rxq_index = rxq_index; 1021 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING; 1022 spin_unlock(&efx->rps_hash_lock); 1023 } else { 1024 /* Without an ARFS hash table, we just use arfs_id 0 for all 1025 * filters. This means if multiple flows hash to the same 1026 * flow_id, all but the most recently touched will be eligible 1027 * for expiry. 1028 */ 1029 rc = 0; 1030 } 1031 1032 /* Queue the request */ 1033 dev_hold(req->net_dev = net_dev); 1034 INIT_WORK(&req->work, efx_filter_rfs_work); 1035 req->rxq_index = rxq_index; 1036 req->flow_id = flow_id; 1037 schedule_work(&req->work); 1038 return rc; 1039out_unlock: 1040 spin_unlock(&efx->rps_hash_lock); 1041out_clear: 1042 clear_bit(slot_idx, &efx->rps_slot_map); 1043 return rc; 1044} 1045 1046bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota) 1047{ 1048 bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index); 1049 struct efx_nic *efx = channel->efx; 1050 unsigned int index, size, start; 1051 u32 flow_id; 1052 1053 if (!mutex_trylock(&efx->rps_mutex)) 1054 return false; 1055 expire_one = efx->type->filter_rfs_expire_one; 1056 index = channel->rfs_expire_index; 1057 start = index; 1058 size = efx->type->max_rx_ip_filters; 1059 while (quota) { 1060 flow_id = channel->rps_flow_id[index]; 1061 1062 if (flow_id != RPS_FLOW_ID_INVALID) { 1063 quota--; 1064 if (expire_one(efx, flow_id, index)) { 1065 netif_info(efx, rx_status, efx->net_dev, 1066 "expired filter %d [channel %u flow %u]\n", 1067 index, channel->channel, flow_id); 1068 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID; 1069 channel->rfs_filter_count--; 1070 } 1071 } 1072 if (++index == size) 1073 index = 0; 1074 /* If we were called with a quota that exceeds the total number 1075 * of filters in the table (which shouldn't happen, but could 1076 * if two callers race), ensure that we don't loop forever - 1077 * stop when we've examined every row of the table. 1078 */ 1079 if (index == start) 1080 break; 1081 } 1082 1083 channel->rfs_expire_index = index; 1084 mutex_unlock(&efx->rps_mutex); 1085 return true; 1086} 1087 1088#endif /* CONFIG_RFS_ACCEL */ 1089