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 "efx.h" 15#include "nic.h" 16#include "rx_common.h" 17 18/* This is the percentage fill level below which new RX descriptors 19 * will be added to the RX descriptor ring. 20 */ 21static unsigned int rx_refill_threshold; 22module_param(rx_refill_threshold, uint, 0444); 23MODULE_PARM_DESC(rx_refill_threshold, 24 "RX descriptor ring refill threshold (%)"); 25 26/* Number of RX buffers to recycle pages for. When creating the RX page recycle 27 * ring, this number is divided by the number of buffers per page to calculate 28 * the number of pages to store in the RX page recycle ring. 29 */ 30#define EFX_RECYCLE_RING_SIZE_IOMMU 4096 31#define EFX_RECYCLE_RING_SIZE_NOIOMMU (2 * EFX_RX_PREFERRED_BATCH) 32 33/* RX maximum head room required. 34 * 35 * This must be at least 1 to prevent overflow, plus one packet-worth 36 * to allow pipelined receives. 37 */ 38#define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS) 39 40/* Check the RX page recycle ring for a page that can be reused. */ 41static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue) 42{ 43 struct efx_nic *efx = rx_queue->efx; 44 struct efx_rx_page_state *state; 45 unsigned int index; 46 struct page *page; 47 48 if (unlikely(!rx_queue->page_ring)) 49 return NULL; 50 index = rx_queue->page_remove & rx_queue->page_ptr_mask; 51 page = rx_queue->page_ring[index]; 52 if (page == NULL) 53 return NULL; 54 55 rx_queue->page_ring[index] = NULL; 56 /* page_remove cannot exceed page_add. */ 57 if (rx_queue->page_remove != rx_queue->page_add) 58 ++rx_queue->page_remove; 59 60 /* If page_count is 1 then we hold the only reference to this page. */ 61 if (page_count(page) == 1) { 62 ++rx_queue->page_recycle_count; 63 return page; 64 } else { 65 state = page_address(page); 66 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr, 67 PAGE_SIZE << efx->rx_buffer_order, 68 DMA_FROM_DEVICE); 69 put_page(page); 70 ++rx_queue->page_recycle_failed; 71 } 72 73 return NULL; 74} 75 76/* Attempt to recycle the page if there is an RX recycle ring; the page can 77 * only be added if this is the final RX buffer, to prevent pages being used in 78 * the descriptor ring and appearing in the recycle ring simultaneously. 79 */ 80static void efx_recycle_rx_page(struct efx_channel *channel, 81 struct efx_rx_buffer *rx_buf) 82{ 83 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 84 struct efx_nic *efx = rx_queue->efx; 85 struct page *page = rx_buf->page; 86 unsigned int index; 87 88 /* Only recycle the page after processing the final buffer. */ 89 if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE)) 90 return; 91 92 index = rx_queue->page_add & rx_queue->page_ptr_mask; 93 if (rx_queue->page_ring[index] == NULL) { 94 unsigned int read_index = rx_queue->page_remove & 95 rx_queue->page_ptr_mask; 96 97 /* The next slot in the recycle ring is available, but 98 * increment page_remove if the read pointer currently 99 * points here. 100 */ 101 if (read_index == index) 102 ++rx_queue->page_remove; 103 rx_queue->page_ring[index] = page; 104 ++rx_queue->page_add; 105 return; 106 } 107 ++rx_queue->page_recycle_full; 108 efx_unmap_rx_buffer(efx, rx_buf); 109 put_page(rx_buf->page); 110} 111 112/* Recycle the pages that are used by buffers that have just been received. */ 113void efx_recycle_rx_pages(struct efx_channel *channel, 114 struct efx_rx_buffer *rx_buf, 115 unsigned int n_frags) 116{ 117 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 118 119 if (unlikely(!rx_queue->page_ring)) 120 return; 121 122 do { 123 efx_recycle_rx_page(channel, rx_buf); 124 rx_buf = efx_rx_buf_next(rx_queue, rx_buf); 125 } while (--n_frags); 126} 127 128void efx_discard_rx_packet(struct efx_channel *channel, 129 struct efx_rx_buffer *rx_buf, 130 unsigned int n_frags) 131{ 132 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 133 134 efx_recycle_rx_pages(channel, rx_buf, n_frags); 135 136 efx_free_rx_buffers(rx_queue, rx_buf, n_frags); 137} 138 139static void efx_init_rx_recycle_ring(struct efx_rx_queue *rx_queue) 140{ 141 unsigned int bufs_in_recycle_ring, page_ring_size; 142 struct efx_nic *efx = rx_queue->efx; 143 144 /* Set the RX recycle ring size */ 145#ifdef CONFIG_PPC64 146 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU; 147#else 148 if (iommu_present(&pci_bus_type)) 149 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU; 150 else 151 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU; 152#endif /* CONFIG_PPC64 */ 153 154 page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring / 155 efx->rx_bufs_per_page); 156 rx_queue->page_ring = kcalloc(page_ring_size, 157 sizeof(*rx_queue->page_ring), GFP_KERNEL); 158 if (!rx_queue->page_ring) 159 rx_queue->page_ptr_mask = 0; 160 else 161 rx_queue->page_ptr_mask = page_ring_size - 1; 162} 163 164static void efx_fini_rx_recycle_ring(struct efx_rx_queue *rx_queue) 165{ 166 struct efx_nic *efx = rx_queue->efx; 167 int i; 168 169 if (unlikely(!rx_queue->page_ring)) 170 return; 171 172 /* Unmap and release the pages in the recycle ring. Remove the ring. */ 173 for (i = 0; i <= rx_queue->page_ptr_mask; i++) { 174 struct page *page = rx_queue->page_ring[i]; 175 struct efx_rx_page_state *state; 176 177 if (page == NULL) 178 continue; 179 180 state = page_address(page); 181 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr, 182 PAGE_SIZE << efx->rx_buffer_order, 183 DMA_FROM_DEVICE); 184 put_page(page); 185 } 186 kfree(rx_queue->page_ring); 187 rx_queue->page_ring = NULL; 188} 189 190static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue, 191 struct efx_rx_buffer *rx_buf) 192{ 193 /* Release the page reference we hold for the buffer. */ 194 if (rx_buf->page) 195 put_page(rx_buf->page); 196 197 /* If this is the last buffer in a page, unmap and free it. */ 198 if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) { 199 efx_unmap_rx_buffer(rx_queue->efx, rx_buf); 200 efx_free_rx_buffers(rx_queue, rx_buf, 1); 201 } 202 rx_buf->page = NULL; 203} 204 205int efx_probe_rx_queue(struct efx_rx_queue *rx_queue) 206{ 207 struct efx_nic *efx = rx_queue->efx; 208 unsigned int entries; 209 int rc; 210 211 /* Create the smallest power-of-two aligned ring */ 212 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE); 213 EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE); 214 rx_queue->ptr_mask = entries - 1; 215 216 netif_dbg(efx, probe, efx->net_dev, 217 "creating RX queue %d size %#x mask %#x\n", 218 efx_rx_queue_index(rx_queue), efx->rxq_entries, 219 rx_queue->ptr_mask); 220 221 /* Allocate RX buffers */ 222 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer), 223 GFP_KERNEL); 224 if (!rx_queue->buffer) 225 return -ENOMEM; 226 227 rc = efx_nic_probe_rx(rx_queue); 228 if (rc) { 229 kfree(rx_queue->buffer); 230 rx_queue->buffer = NULL; 231 } 232 233 return rc; 234} 235 236void efx_init_rx_queue(struct efx_rx_queue *rx_queue) 237{ 238 unsigned int max_fill, trigger, max_trigger; 239 struct efx_nic *efx = rx_queue->efx; 240 int rc = 0; 241 242 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 243 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue)); 244 245 /* Initialise ptr fields */ 246 rx_queue->added_count = 0; 247 rx_queue->notified_count = 0; 248 rx_queue->removed_count = 0; 249 rx_queue->min_fill = -1U; 250 efx_init_rx_recycle_ring(rx_queue); 251 252 rx_queue->page_remove = 0; 253 rx_queue->page_add = rx_queue->page_ptr_mask + 1; 254 rx_queue->page_recycle_count = 0; 255 rx_queue->page_recycle_failed = 0; 256 rx_queue->page_recycle_full = 0; 257 258 /* Initialise limit fields */ 259 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM; 260 max_trigger = 261 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page; 262 if (rx_refill_threshold != 0) { 263 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U; 264 if (trigger > max_trigger) 265 trigger = max_trigger; 266 } else { 267 trigger = max_trigger; 268 } 269 270 rx_queue->max_fill = max_fill; 271 rx_queue->fast_fill_trigger = trigger; 272 rx_queue->refill_enabled = true; 273 274 /* Initialise XDP queue information */ 275 rc = xdp_rxq_info_reg(&rx_queue->xdp_rxq_info, efx->net_dev, 276 rx_queue->core_index); 277 278 if (rc) { 279 netif_err(efx, rx_err, efx->net_dev, 280 "Failure to initialise XDP queue information rc=%d\n", 281 rc); 282 efx->xdp_rxq_info_failed = true; 283 } else { 284 rx_queue->xdp_rxq_info_valid = true; 285 } 286 287 /* Set up RX descriptor ring */ 288 efx_nic_init_rx(rx_queue); 289} 290 291void efx_fini_rx_queue(struct efx_rx_queue *rx_queue) 292{ 293 struct efx_rx_buffer *rx_buf; 294 int i; 295 296 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 297 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue)); 298 299 del_timer_sync(&rx_queue->slow_fill); 300 301 /* Release RX buffers from the current read ptr to the write ptr */ 302 if (rx_queue->buffer) { 303 for (i = rx_queue->removed_count; i < rx_queue->added_count; 304 i++) { 305 unsigned int index = i & rx_queue->ptr_mask; 306 307 rx_buf = efx_rx_buffer(rx_queue, index); 308 efx_fini_rx_buffer(rx_queue, rx_buf); 309 } 310 } 311 312 efx_fini_rx_recycle_ring(rx_queue); 313 314 if (rx_queue->xdp_rxq_info_valid) 315 xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info); 316 317 rx_queue->xdp_rxq_info_valid = false; 318} 319 320void efx_remove_rx_queue(struct efx_rx_queue *rx_queue) 321{ 322 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 323 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue)); 324 325 efx_nic_remove_rx(rx_queue); 326 327 kfree(rx_queue->buffer); 328 rx_queue->buffer = NULL; 329} 330 331/* Unmap a DMA-mapped page. This function is only called for the final RX 332 * buffer in a page. 333 */ 334void efx_unmap_rx_buffer(struct efx_nic *efx, 335 struct efx_rx_buffer *rx_buf) 336{ 337 struct page *page = rx_buf->page; 338 339 if (page) { 340 struct efx_rx_page_state *state = page_address(page); 341 342 dma_unmap_page(&efx->pci_dev->dev, 343 state->dma_addr, 344 PAGE_SIZE << efx->rx_buffer_order, 345 DMA_FROM_DEVICE); 346 } 347} 348 349void efx_free_rx_buffers(struct efx_rx_queue *rx_queue, 350 struct efx_rx_buffer *rx_buf, 351 unsigned int num_bufs) 352{ 353 do { 354 if (rx_buf->page) { 355 put_page(rx_buf->page); 356 rx_buf->page = NULL; 357 } 358 rx_buf = efx_rx_buf_next(rx_queue, rx_buf); 359 } while (--num_bufs); 360} 361 362void efx_rx_slow_fill(struct timer_list *t) 363{ 364 struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill); 365 366 /* Post an event to cause NAPI to run and refill the queue */ 367 efx_nic_generate_fill_event(rx_queue); 368 ++rx_queue->slow_fill_count; 369} 370 371void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue) 372{ 373 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(10)); 374} 375 376/* efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers 377 * 378 * @rx_queue: Efx RX queue 379 * 380 * This allocates a batch of pages, maps them for DMA, and populates 381 * struct efx_rx_buffers for each one. Return a negative error code or 382 * 0 on success. If a single page can be used for multiple buffers, 383 * then the page will either be inserted fully, or not at all. 384 */ 385static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic) 386{ 387 unsigned int page_offset, index, count; 388 struct efx_nic *efx = rx_queue->efx; 389 struct efx_rx_page_state *state; 390 struct efx_rx_buffer *rx_buf; 391 dma_addr_t dma_addr; 392 struct page *page; 393 394 count = 0; 395 do { 396 page = efx_reuse_page(rx_queue); 397 if (page == NULL) { 398 page = alloc_pages(__GFP_COMP | 399 (atomic ? GFP_ATOMIC : GFP_KERNEL), 400 efx->rx_buffer_order); 401 if (unlikely(page == NULL)) 402 return -ENOMEM; 403 dma_addr = 404 dma_map_page(&efx->pci_dev->dev, page, 0, 405 PAGE_SIZE << efx->rx_buffer_order, 406 DMA_FROM_DEVICE); 407 if (unlikely(dma_mapping_error(&efx->pci_dev->dev, 408 dma_addr))) { 409 __free_pages(page, efx->rx_buffer_order); 410 return -EIO; 411 } 412 state = page_address(page); 413 state->dma_addr = dma_addr; 414 } else { 415 state = page_address(page); 416 dma_addr = state->dma_addr; 417 } 418 419 dma_addr += sizeof(struct efx_rx_page_state); 420 page_offset = sizeof(struct efx_rx_page_state); 421 422 do { 423 index = rx_queue->added_count & rx_queue->ptr_mask; 424 rx_buf = efx_rx_buffer(rx_queue, index); 425 rx_buf->dma_addr = dma_addr + efx->rx_ip_align + 426 EFX_XDP_HEADROOM; 427 rx_buf->page = page; 428 rx_buf->page_offset = page_offset + efx->rx_ip_align + 429 EFX_XDP_HEADROOM; 430 rx_buf->len = efx->rx_dma_len; 431 rx_buf->flags = 0; 432 ++rx_queue->added_count; 433 get_page(page); 434 dma_addr += efx->rx_page_buf_step; 435 page_offset += efx->rx_page_buf_step; 436 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE); 437 438 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE; 439 } while (++count < efx->rx_pages_per_batch); 440 441 return 0; 442} 443 444void efx_rx_config_page_split(struct efx_nic *efx) 445{ 446 efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align + 447 EFX_XDP_HEADROOM + EFX_XDP_TAILROOM, 448 EFX_RX_BUF_ALIGNMENT); 449 efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 : 450 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) / 451 efx->rx_page_buf_step); 452 efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) / 453 efx->rx_bufs_per_page; 454 efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH, 455 efx->rx_bufs_per_page); 456} 457 458/* efx_fast_push_rx_descriptors - push new RX descriptors quickly 459 * @rx_queue: RX descriptor queue 460 * 461 * This will aim to fill the RX descriptor queue up to 462 * @rx_queue->@max_fill. If there is insufficient atomic 463 * memory to do so, a slow fill will be scheduled. 464 * 465 * The caller must provide serialisation (none is used here). In practise, 466 * this means this function must run from the NAPI handler, or be called 467 * when NAPI is disabled. 468 */ 469void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic) 470{ 471 struct efx_nic *efx = rx_queue->efx; 472 unsigned int fill_level, batch_size; 473 int space, rc = 0; 474 475 if (!rx_queue->refill_enabled) 476 return; 477 478 /* Calculate current fill level, and exit if we don't need to fill */ 479 fill_level = (rx_queue->added_count - rx_queue->removed_count); 480 EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries); 481 if (fill_level >= rx_queue->fast_fill_trigger) 482 goto out; 483 484 /* Record minimum fill level */ 485 if (unlikely(fill_level < rx_queue->min_fill)) { 486 if (fill_level) 487 rx_queue->min_fill = fill_level; 488 } 489 490 batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page; 491 space = rx_queue->max_fill - fill_level; 492 EFX_WARN_ON_ONCE_PARANOID(space < batch_size); 493 494 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev, 495 "RX queue %d fast-filling descriptor ring from" 496 " level %d to level %d\n", 497 efx_rx_queue_index(rx_queue), fill_level, 498 rx_queue->max_fill); 499 500 do { 501 rc = efx_init_rx_buffers(rx_queue, atomic); 502 if (unlikely(rc)) { 503 /* Ensure that we don't leave the rx queue empty */ 504 efx_schedule_slow_fill(rx_queue); 505 goto out; 506 } 507 } while ((space -= batch_size) >= batch_size); 508 509 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev, 510 "RX queue %d fast-filled descriptor ring " 511 "to level %d\n", efx_rx_queue_index(rx_queue), 512 rx_queue->added_count - rx_queue->removed_count); 513 514 out: 515 if (rx_queue->notified_count != rx_queue->added_count) 516 efx_nic_notify_rx_desc(rx_queue); 517} 518 519/* Pass a received packet up through GRO. GRO can handle pages 520 * regardless of checksum state and skbs with a good checksum. 521 */ 522void 523efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf, 524 unsigned int n_frags, u8 *eh, __wsum csum) 525{ 526 struct napi_struct *napi = &channel->napi_str; 527 struct efx_nic *efx = channel->efx; 528 struct sk_buff *skb; 529 530 skb = napi_get_frags(napi); 531 if (unlikely(!skb)) { 532 struct efx_rx_queue *rx_queue; 533 534 rx_queue = efx_channel_get_rx_queue(channel); 535 efx_free_rx_buffers(rx_queue, rx_buf, n_frags); 536 return; 537 } 538 539 if (efx->net_dev->features & NETIF_F_RXHASH && 540 efx_rx_buf_hash_valid(efx, eh)) 541 skb_set_hash(skb, efx_rx_buf_hash(efx, eh), 542 PKT_HASH_TYPE_L3); 543 if (csum) { 544 skb->csum = csum; 545 skb->ip_summed = CHECKSUM_COMPLETE; 546 } else { 547 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ? 548 CHECKSUM_UNNECESSARY : CHECKSUM_NONE); 549 } 550 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL); 551 552 for (;;) { 553 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags, 554 rx_buf->page, rx_buf->page_offset, 555 rx_buf->len); 556 rx_buf->page = NULL; 557 skb->len += rx_buf->len; 558 if (skb_shinfo(skb)->nr_frags == n_frags) 559 break; 560 561 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf); 562 } 563 564 skb->data_len = skb->len; 565 skb->truesize += n_frags * efx->rx_buffer_truesize; 566 567 skb_record_rx_queue(skb, channel->rx_queue.core_index); 568 569 napi_gro_frags(napi); 570} 571 572/* RSS contexts. We're using linked lists and crappy O(n) algorithms, because 573 * (a) this is an infrequent control-plane operation and (b) n is small (max 64) 574 */ 575struct efx_rss_context *efx_alloc_rss_context_entry(struct efx_nic *efx) 576{ 577 struct list_head *head = &efx->rss_context.list; 578 struct efx_rss_context *ctx, *new; 579 u32 id = 1; /* Don't use zero, that refers to the master RSS context */ 580 581 WARN_ON(!mutex_is_locked(&efx->rss_lock)); 582 583 /* Search for first gap in the numbering */ 584 list_for_each_entry(ctx, head, list) { 585 if (ctx->user_id != id) 586 break; 587 id++; 588 /* Check for wrap. If this happens, we have nearly 2^32 589 * allocated RSS contexts, which seems unlikely. 590 */ 591 if (WARN_ON_ONCE(!id)) 592 return NULL; 593 } 594 595 /* Create the new entry */ 596 new = kmalloc(sizeof(*new), GFP_KERNEL); 597 if (!new) 598 return NULL; 599 new->context_id = EFX_MCDI_RSS_CONTEXT_INVALID; 600 new->rx_hash_udp_4tuple = false; 601 602 /* Insert the new entry into the gap */ 603 new->user_id = id; 604 list_add_tail(&new->list, &ctx->list); 605 return new; 606} 607 608struct efx_rss_context *efx_find_rss_context_entry(struct efx_nic *efx, u32 id) 609{ 610 struct list_head *head = &efx->rss_context.list; 611 struct efx_rss_context *ctx; 612 613 WARN_ON(!mutex_is_locked(&efx->rss_lock)); 614 615 list_for_each_entry(ctx, head, list) 616 if (ctx->user_id == id) 617 return ctx; 618 return NULL; 619} 620 621void efx_free_rss_context_entry(struct efx_rss_context *ctx) 622{ 623 list_del(&ctx->list); 624 kfree(ctx); 625} 626 627void efx_set_default_rx_indir_table(struct efx_nic *efx, 628 struct efx_rss_context *ctx) 629{ 630 size_t i; 631 632 for (i = 0; i < ARRAY_SIZE(ctx->rx_indir_table); i++) 633 ctx->rx_indir_table[i] = 634 ethtool_rxfh_indir_default(i, efx->rss_spread); 635} 636 637/** 638 * efx_filter_is_mc_recipient - test whether spec is a multicast recipient 639 * @spec: Specification to test 640 * 641 * Return: %true if the specification is a non-drop RX filter that 642 * matches a local MAC address I/G bit value of 1 or matches a local 643 * IPv4 or IPv6 address value in the respective multicast address 644 * range. Otherwise %false. 645 */ 646bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec) 647{ 648 if (!(spec->flags & EFX_FILTER_FLAG_RX) || 649 spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP) 650 return false; 651 652 if (spec->match_flags & 653 (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) && 654 is_multicast_ether_addr(spec->loc_mac)) 655 return true; 656 657 if ((spec->match_flags & 658 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) == 659 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) { 660 if (spec->ether_type == htons(ETH_P_IP) && 661 ipv4_is_multicast(spec->loc_host[0])) 662 return true; 663 if (spec->ether_type == htons(ETH_P_IPV6) && 664 ((const u8 *)spec->loc_host)[0] == 0xff) 665 return true; 666 } 667 668 return false; 669} 670 671bool efx_filter_spec_equal(const struct efx_filter_spec *left, 672 const struct efx_filter_spec *right) 673{ 674 if ((left->match_flags ^ right->match_flags) | 675 ((left->flags ^ right->flags) & 676 (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX))) 677 return false; 678 679 return memcmp(&left->vport_id, &right->vport_id, 680 sizeof(struct efx_filter_spec) - 681 offsetof(struct efx_filter_spec, vport_id)) == 0; 682} 683 684u32 efx_filter_spec_hash(const struct efx_filter_spec *spec) 685{ 686 BUILD_BUG_ON(offsetof(struct efx_filter_spec, vport_id) & 3); 687 return jhash2((const u32 *)&spec->vport_id, 688 (sizeof(struct efx_filter_spec) - 689 offsetof(struct efx_filter_spec, vport_id)) / 4, 690 0); 691} 692 693#ifdef CONFIG_RFS_ACCEL 694bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx, 695 bool *force) 696{ 697 if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) { 698 /* ARFS is currently updating this entry, leave it */ 699 return false; 700 } 701 if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) { 702 /* ARFS tried and failed to update this, so it's probably out 703 * of date. Remove the filter and the ARFS rule entry. 704 */ 705 rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING; 706 *force = true; 707 return true; 708 } else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */ 709 /* ARFS has moved on, so old filter is not needed. Since we did 710 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will 711 * not be removed by efx_rps_hash_del() subsequently. 712 */ 713 *force = true; 714 return true; 715 } 716 /* Remove it iff ARFS wants to. */ 717 return true; 718} 719 720static 721struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx, 722 const struct efx_filter_spec *spec) 723{ 724 u32 hash = efx_filter_spec_hash(spec); 725 726 lockdep_assert_held(&efx->rps_hash_lock); 727 if (!efx->rps_hash_table) 728 return NULL; 729 return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE]; 730} 731 732struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx, 733 const struct efx_filter_spec *spec) 734{ 735 struct efx_arfs_rule *rule; 736 struct hlist_head *head; 737 struct hlist_node *node; 738 739 head = efx_rps_hash_bucket(efx, spec); 740 if (!head) 741 return NULL; 742 hlist_for_each(node, head) { 743 rule = container_of(node, struct efx_arfs_rule, node); 744 if (efx_filter_spec_equal(spec, &rule->spec)) 745 return rule; 746 } 747 return NULL; 748} 749 750struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx, 751 const struct efx_filter_spec *spec, 752 bool *new) 753{ 754 struct efx_arfs_rule *rule; 755 struct hlist_head *head; 756 struct hlist_node *node; 757 758 head = efx_rps_hash_bucket(efx, spec); 759 if (!head) 760 return NULL; 761 hlist_for_each(node, head) { 762 rule = container_of(node, struct efx_arfs_rule, node); 763 if (efx_filter_spec_equal(spec, &rule->spec)) { 764 *new = false; 765 return rule; 766 } 767 } 768 rule = kmalloc(sizeof(*rule), GFP_ATOMIC); 769 *new = true; 770 if (rule) { 771 memcpy(&rule->spec, spec, sizeof(rule->spec)); 772 hlist_add_head(&rule->node, head); 773 } 774 return rule; 775} 776 777void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec) 778{ 779 struct efx_arfs_rule *rule; 780 struct hlist_head *head; 781 struct hlist_node *node; 782 783 head = efx_rps_hash_bucket(efx, spec); 784 if (WARN_ON(!head)) 785 return; 786 hlist_for_each(node, head) { 787 rule = container_of(node, struct efx_arfs_rule, node); 788 if (efx_filter_spec_equal(spec, &rule->spec)) { 789 /* Someone already reused the entry. We know that if 790 * this check doesn't fire (i.e. filter_id == REMOVING) 791 * then the REMOVING mark was put there by our caller, 792 * because caller is holding a lock on filter table and 793 * only holders of that lock set REMOVING. 794 */ 795 if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING) 796 return; 797 hlist_del(node); 798 kfree(rule); 799 return; 800 } 801 } 802 /* We didn't find it. */ 803 WARN_ON(1); 804} 805#endif 806 807int efx_probe_filters(struct efx_nic *efx) 808{ 809 int rc; 810 811 mutex_lock(&efx->mac_lock); 812 down_write(&efx->filter_sem); 813 rc = efx->type->filter_table_probe(efx); 814 if (rc) 815 goto out_unlock; 816 817#ifdef CONFIG_RFS_ACCEL 818 if (efx->type->offload_features & NETIF_F_NTUPLE) { 819 struct efx_channel *channel; 820 int i, success = 1; 821 822 efx_for_each_channel(channel, efx) { 823 channel->rps_flow_id = 824 kcalloc(efx->type->max_rx_ip_filters, 825 sizeof(*channel->rps_flow_id), 826 GFP_KERNEL); 827 if (!channel->rps_flow_id) 828 success = 0; 829 else 830 for (i = 0; 831 i < efx->type->max_rx_ip_filters; 832 ++i) 833 channel->rps_flow_id[i] = 834 RPS_FLOW_ID_INVALID; 835 channel->rfs_expire_index = 0; 836 channel->rfs_filter_count = 0; 837 } 838 839 if (!success) { 840 efx_for_each_channel(channel, efx) { 841 kfree(channel->rps_flow_id); 842 channel->rps_flow_id = NULL; 843 } 844 efx->type->filter_table_remove(efx); 845 rc = -ENOMEM; 846 goto out_unlock; 847 } 848 } 849#endif 850out_unlock: 851 up_write(&efx->filter_sem); 852 mutex_unlock(&efx->mac_lock); 853 return rc; 854} 855 856void efx_remove_filters(struct efx_nic *efx) 857{ 858#ifdef CONFIG_RFS_ACCEL 859 struct efx_channel *channel; 860 861 efx_for_each_channel(channel, efx) { 862 cancel_delayed_work_sync(&channel->filter_work); 863 kfree(channel->rps_flow_id); 864 channel->rps_flow_id = NULL; 865 } 866#endif 867 down_write(&efx->filter_sem); 868 efx->type->filter_table_remove(efx); 869 up_write(&efx->filter_sem); 870} 871 872#ifdef CONFIG_RFS_ACCEL 873 874static void efx_filter_rfs_work(struct work_struct *data) 875{ 876 struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion, 877 work); 878 struct efx_nic *efx = netdev_priv(req->net_dev); 879 struct efx_channel *channel = efx_get_channel(efx, req->rxq_index); 880 int slot_idx = req - efx->rps_slot; 881 struct efx_arfs_rule *rule; 882 u16 arfs_id = 0; 883 int rc; 884 885 rc = efx->type->filter_insert(efx, &req->spec, true); 886 if (rc >= 0) 887 /* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */ 888 rc %= efx->type->max_rx_ip_filters; 889 if (efx->rps_hash_table) { 890 spin_lock_bh(&efx->rps_hash_lock); 891 rule = efx_rps_hash_find(efx, &req->spec); 892 /* The rule might have already gone, if someone else's request 893 * for the same spec was already worked and then expired before 894 * we got around to our work. In that case we have nothing 895 * tying us to an arfs_id, meaning that as soon as the filter 896 * is considered for expiry it will be removed. 897 */ 898 if (rule) { 899 if (rc < 0) 900 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR; 901 else 902 rule->filter_id = rc; 903 arfs_id = rule->arfs_id; 904 } 905 spin_unlock_bh(&efx->rps_hash_lock); 906 } 907 if (rc >= 0) { 908 /* Remember this so we can check whether to expire the filter 909 * later. 910 */ 911 mutex_lock(&efx->rps_mutex); 912 if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID) 913 channel->rfs_filter_count++; 914 channel->rps_flow_id[rc] = req->flow_id; 915 mutex_unlock(&efx->rps_mutex); 916 917 if (req->spec.ether_type == htons(ETH_P_IP)) 918 netif_info(efx, rx_status, efx->net_dev, 919 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %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_info(efx, rx_status, efx->net_dev, 926 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %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_succeeded++; 932 } else { 933 if (req->spec.ether_type == htons(ETH_P_IP)) 934 netif_dbg(efx, rx_status, efx->net_dev, 935 "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n", 936 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 937 req->spec.rem_host, ntohs(req->spec.rem_port), 938 req->spec.loc_host, ntohs(req->spec.loc_port), 939 req->rxq_index, req->flow_id, rc, arfs_id); 940 else 941 netif_dbg(efx, rx_status, efx->net_dev, 942 "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n", 943 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 944 req->spec.rem_host, ntohs(req->spec.rem_port), 945 req->spec.loc_host, ntohs(req->spec.loc_port), 946 req->rxq_index, req->flow_id, rc, arfs_id); 947 channel->n_rfs_failed++; 948 /* We're overloading the NIC's filter tables, so let's do a 949 * chunk of extra expiry work. 950 */ 951 __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count, 952 100u)); 953 } 954 955 /* Release references */ 956 clear_bit(slot_idx, &efx->rps_slot_map); 957 dev_put(req->net_dev); 958} 959 960int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb, 961 u16 rxq_index, u32 flow_id) 962{ 963 struct efx_nic *efx = netdev_priv(net_dev); 964 struct efx_async_filter_insertion *req; 965 struct efx_arfs_rule *rule; 966 struct flow_keys fk; 967 int slot_idx; 968 bool new; 969 int rc; 970 971 /* find a free slot */ 972 for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++) 973 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map)) 974 break; 975 if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT) 976 return -EBUSY; 977 978 if (flow_id == RPS_FLOW_ID_INVALID) { 979 rc = -EINVAL; 980 goto out_clear; 981 } 982 983 if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) { 984 rc = -EPROTONOSUPPORT; 985 goto out_clear; 986 } 987 988 if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) { 989 rc = -EPROTONOSUPPORT; 990 goto out_clear; 991 } 992 if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) { 993 rc = -EPROTONOSUPPORT; 994 goto out_clear; 995 } 996 997 req = efx->rps_slot + slot_idx; 998 efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT, 999 efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0, 1000 rxq_index); 1001 req->spec.match_flags = 1002 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO | 1003 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT | 1004 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT; 1005 req->spec.ether_type = fk.basic.n_proto; 1006 req->spec.ip_proto = fk.basic.ip_proto; 1007 1008 if (fk.basic.n_proto == htons(ETH_P_IP)) { 1009 req->spec.rem_host[0] = fk.addrs.v4addrs.src; 1010 req->spec.loc_host[0] = fk.addrs.v4addrs.dst; 1011 } else { 1012 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src, 1013 sizeof(struct in6_addr)); 1014 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst, 1015 sizeof(struct in6_addr)); 1016 } 1017 1018 req->spec.rem_port = fk.ports.src; 1019 req->spec.loc_port = fk.ports.dst; 1020 1021 if (efx->rps_hash_table) { 1022 /* Add it to ARFS hash table */ 1023 spin_lock(&efx->rps_hash_lock); 1024 rule = efx_rps_hash_add(efx, &req->spec, &new); 1025 if (!rule) { 1026 rc = -ENOMEM; 1027 goto out_unlock; 1028 } 1029 if (new) 1030 rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER; 1031 rc = rule->arfs_id; 1032 /* Skip if existing or pending filter already does the right thing */ 1033 if (!new && rule->rxq_index == rxq_index && 1034 rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING) 1035 goto out_unlock; 1036 rule->rxq_index = rxq_index; 1037 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING; 1038 spin_unlock(&efx->rps_hash_lock); 1039 } else { 1040 /* Without an ARFS hash table, we just use arfs_id 0 for all 1041 * filters. This means if multiple flows hash to the same 1042 * flow_id, all but the most recently touched will be eligible 1043 * for expiry. 1044 */ 1045 rc = 0; 1046 } 1047 1048 /* Queue the request */ 1049 dev_hold(req->net_dev = net_dev); 1050 INIT_WORK(&req->work, efx_filter_rfs_work); 1051 req->rxq_index = rxq_index; 1052 req->flow_id = flow_id; 1053 schedule_work(&req->work); 1054 return rc; 1055out_unlock: 1056 spin_unlock(&efx->rps_hash_lock); 1057out_clear: 1058 clear_bit(slot_idx, &efx->rps_slot_map); 1059 return rc; 1060} 1061 1062bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota) 1063{ 1064 bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index); 1065 struct efx_nic *efx = channel->efx; 1066 unsigned int index, size, start; 1067 u32 flow_id; 1068 1069 if (!mutex_trylock(&efx->rps_mutex)) 1070 return false; 1071 expire_one = efx->type->filter_rfs_expire_one; 1072 index = channel->rfs_expire_index; 1073 start = index; 1074 size = efx->type->max_rx_ip_filters; 1075 while (quota) { 1076 flow_id = channel->rps_flow_id[index]; 1077 1078 if (flow_id != RPS_FLOW_ID_INVALID) { 1079 quota--; 1080 if (expire_one(efx, flow_id, index)) { 1081 netif_info(efx, rx_status, efx->net_dev, 1082 "expired filter %d [channel %u flow %u]\n", 1083 index, channel->channel, flow_id); 1084 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID; 1085 channel->rfs_filter_count--; 1086 } 1087 } 1088 if (++index == size) 1089 index = 0; 1090 /* If we were called with a quota that exceeds the total number 1091 * of filters in the table (which shouldn't happen, but could 1092 * if two callers race), ensure that we don't loop forever - 1093 * stop when we've examined every row of the table. 1094 */ 1095 if (index == start) 1096 break; 1097 } 1098 1099 channel->rfs_expire_index = index; 1100 mutex_unlock(&efx->rps_mutex); 1101 return true; 1102} 1103 1104#endif /* CONFIG_RFS_ACCEL */ 1105