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 */ 21 static unsigned int rx_refill_threshold; 22 module_param(rx_refill_threshold, uint, 0444); 23 MODULE_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. */ 41 static 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 index = rx_queue->page_remove & rx_queue->page_ptr_mask; 49 page = rx_queue->page_ring[index]; 50 if (page == NULL) 51 return NULL; 52 53 rx_queue->page_ring[index] = NULL; 54 /* page_remove cannot exceed page_add. */ 55 if (rx_queue->page_remove != rx_queue->page_add) 56 ++rx_queue->page_remove; 57 58 /* If page_count is 1 then we hold the only reference to this page. */ 59 if (page_count(page) == 1) { 60 ++rx_queue->page_recycle_count; 61 return page; 62 } else { 63 state = page_address(page); 64 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr, 65 PAGE_SIZE << efx->rx_buffer_order, 66 DMA_FROM_DEVICE); 67 put_page(page); 68 ++rx_queue->page_recycle_failed; 69 } 70 71 return NULL; 72 } 73 74 /* Attempt to recycle the page if there is an RX recycle ring; the page can 75 * only be added if this is the final RX buffer, to prevent pages being used in 76 * the descriptor ring and appearing in the recycle ring simultaneously. 77 */ 78 static void efx_recycle_rx_page(struct efx_channel *channel, 79 struct efx_rx_buffer *rx_buf) 80 { 81 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 82 struct efx_nic *efx = rx_queue->efx; 83 struct page *page = rx_buf->page; 84 unsigned int index; 85 86 /* Only recycle the page after processing the final buffer. */ 87 if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE)) 88 return; 89 90 index = rx_queue->page_add & rx_queue->page_ptr_mask; 91 if (rx_queue->page_ring[index] == NULL) { 92 unsigned int read_index = rx_queue->page_remove & 93 rx_queue->page_ptr_mask; 94 95 /* The next slot in the recycle ring is available, but 96 * increment page_remove if the read pointer currently 97 * points here. 98 */ 99 if (read_index == index) 100 ++rx_queue->page_remove; 101 rx_queue->page_ring[index] = page; 102 ++rx_queue->page_add; 103 return; 104 } 105 ++rx_queue->page_recycle_full; 106 efx_unmap_rx_buffer(efx, rx_buf); 107 put_page(rx_buf->page); 108 } 109 110 /* Recycle the pages that are used by buffers that have just been received. */ 111 void efx_recycle_rx_pages(struct efx_channel *channel, 112 struct efx_rx_buffer *rx_buf, 113 unsigned int n_frags) 114 { 115 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 116 117 do { 118 efx_recycle_rx_page(channel, rx_buf); 119 rx_buf = efx_rx_buf_next(rx_queue, rx_buf); 120 } while (--n_frags); 121 } 122 123 void efx_discard_rx_packet(struct efx_channel *channel, 124 struct efx_rx_buffer *rx_buf, 125 unsigned int n_frags) 126 { 127 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 128 129 efx_recycle_rx_pages(channel, rx_buf, n_frags); 130 131 efx_free_rx_buffers(rx_queue, rx_buf, n_frags); 132 } 133 134 static void efx_init_rx_recycle_ring(struct efx_rx_queue *rx_queue) 135 { 136 unsigned int bufs_in_recycle_ring, page_ring_size; 137 struct efx_nic *efx = rx_queue->efx; 138 139 /* Set the RX recycle ring size */ 140 #ifdef CONFIG_PPC64 141 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU; 142 #else 143 if (iommu_present(&pci_bus_type)) 144 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_IOMMU; 145 else 146 bufs_in_recycle_ring = EFX_RECYCLE_RING_SIZE_NOIOMMU; 147 #endif /* CONFIG_PPC64 */ 148 149 page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring / 150 efx->rx_bufs_per_page); 151 rx_queue->page_ring = kcalloc(page_ring_size, 152 sizeof(*rx_queue->page_ring), GFP_KERNEL); 153 rx_queue->page_ptr_mask = page_ring_size - 1; 154 } 155 156 static void efx_fini_rx_recycle_ring(struct efx_rx_queue *rx_queue) 157 { 158 struct efx_nic *efx = rx_queue->efx; 159 int i; 160 161 /* Unmap and release the pages in the recycle ring. Remove the ring. */ 162 for (i = 0; i <= rx_queue->page_ptr_mask; i++) { 163 struct page *page = rx_queue->page_ring[i]; 164 struct efx_rx_page_state *state; 165 166 if (page == NULL) 167 continue; 168 169 state = page_address(page); 170 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr, 171 PAGE_SIZE << efx->rx_buffer_order, 172 DMA_FROM_DEVICE); 173 put_page(page); 174 } 175 kfree(rx_queue->page_ring); 176 rx_queue->page_ring = NULL; 177 } 178 179 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue, 180 struct efx_rx_buffer *rx_buf) 181 { 182 /* Release the page reference we hold for the buffer. */ 183 if (rx_buf->page) 184 put_page(rx_buf->page); 185 186 /* If this is the last buffer in a page, unmap and free it. */ 187 if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) { 188 efx_unmap_rx_buffer(rx_queue->efx, rx_buf); 189 efx_free_rx_buffers(rx_queue, rx_buf, 1); 190 } 191 rx_buf->page = NULL; 192 } 193 194 int efx_probe_rx_queue(struct efx_rx_queue *rx_queue) 195 { 196 struct efx_nic *efx = rx_queue->efx; 197 unsigned int entries; 198 int rc; 199 200 /* Create the smallest power-of-two aligned ring */ 201 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE); 202 EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE); 203 rx_queue->ptr_mask = entries - 1; 204 205 netif_dbg(efx, probe, efx->net_dev, 206 "creating RX queue %d size %#x mask %#x\n", 207 efx_rx_queue_index(rx_queue), efx->rxq_entries, 208 rx_queue->ptr_mask); 209 210 /* Allocate RX buffers */ 211 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer), 212 GFP_KERNEL); 213 if (!rx_queue->buffer) 214 return -ENOMEM; 215 216 rc = efx_nic_probe_rx(rx_queue); 217 if (rc) { 218 kfree(rx_queue->buffer); 219 rx_queue->buffer = NULL; 220 } 221 222 return rc; 223 } 224 225 void efx_init_rx_queue(struct efx_rx_queue *rx_queue) 226 { 227 unsigned int max_fill, trigger, max_trigger; 228 struct efx_nic *efx = rx_queue->efx; 229 int rc = 0; 230 231 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 232 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue)); 233 234 /* Initialise ptr fields */ 235 rx_queue->added_count = 0; 236 rx_queue->notified_count = 0; 237 rx_queue->removed_count = 0; 238 rx_queue->min_fill = -1U; 239 efx_init_rx_recycle_ring(rx_queue); 240 241 rx_queue->page_remove = 0; 242 rx_queue->page_add = rx_queue->page_ptr_mask + 1; 243 rx_queue->page_recycle_count = 0; 244 rx_queue->page_recycle_failed = 0; 245 rx_queue->page_recycle_full = 0; 246 247 /* Initialise limit fields */ 248 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM; 249 max_trigger = 250 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page; 251 if (rx_refill_threshold != 0) { 252 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U; 253 if (trigger > max_trigger) 254 trigger = max_trigger; 255 } else { 256 trigger = max_trigger; 257 } 258 259 rx_queue->max_fill = max_fill; 260 rx_queue->fast_fill_trigger = trigger; 261 rx_queue->refill_enabled = true; 262 263 /* Initialise XDP queue information */ 264 rc = xdp_rxq_info_reg(&rx_queue->xdp_rxq_info, efx->net_dev, 265 rx_queue->core_index); 266 267 if (rc) { 268 netif_err(efx, rx_err, efx->net_dev, 269 "Failure to initialise XDP queue information rc=%d\n", 270 rc); 271 efx->xdp_rxq_info_failed = true; 272 } else { 273 rx_queue->xdp_rxq_info_valid = true; 274 } 275 276 /* Set up RX descriptor ring */ 277 efx_nic_init_rx(rx_queue); 278 } 279 280 void efx_fini_rx_queue(struct efx_rx_queue *rx_queue) 281 { 282 struct efx_rx_buffer *rx_buf; 283 int i; 284 285 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 286 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue)); 287 288 del_timer_sync(&rx_queue->slow_fill); 289 290 /* Release RX buffers from the current read ptr to the write ptr */ 291 if (rx_queue->buffer) { 292 for (i = rx_queue->removed_count; i < rx_queue->added_count; 293 i++) { 294 unsigned int index = i & rx_queue->ptr_mask; 295 296 rx_buf = efx_rx_buffer(rx_queue, index); 297 efx_fini_rx_buffer(rx_queue, rx_buf); 298 } 299 } 300 301 efx_fini_rx_recycle_ring(rx_queue); 302 303 if (rx_queue->xdp_rxq_info_valid) 304 xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info); 305 306 rx_queue->xdp_rxq_info_valid = false; 307 } 308 309 void efx_remove_rx_queue(struct efx_rx_queue *rx_queue) 310 { 311 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 312 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue)); 313 314 efx_nic_remove_rx(rx_queue); 315 316 kfree(rx_queue->buffer); 317 rx_queue->buffer = NULL; 318 } 319 320 /* Unmap a DMA-mapped page. This function is only called for the final RX 321 * buffer in a page. 322 */ 323 void efx_unmap_rx_buffer(struct efx_nic *efx, 324 struct efx_rx_buffer *rx_buf) 325 { 326 struct page *page = rx_buf->page; 327 328 if (page) { 329 struct efx_rx_page_state *state = page_address(page); 330 331 dma_unmap_page(&efx->pci_dev->dev, 332 state->dma_addr, 333 PAGE_SIZE << efx->rx_buffer_order, 334 DMA_FROM_DEVICE); 335 } 336 } 337 338 void efx_free_rx_buffers(struct efx_rx_queue *rx_queue, 339 struct efx_rx_buffer *rx_buf, 340 unsigned int num_bufs) 341 { 342 do { 343 if (rx_buf->page) { 344 put_page(rx_buf->page); 345 rx_buf->page = NULL; 346 } 347 rx_buf = efx_rx_buf_next(rx_queue, rx_buf); 348 } while (--num_bufs); 349 } 350 351 void efx_rx_slow_fill(struct timer_list *t) 352 { 353 struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill); 354 355 /* Post an event to cause NAPI to run and refill the queue */ 356 efx_nic_generate_fill_event(rx_queue); 357 ++rx_queue->slow_fill_count; 358 } 359 360 void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue) 361 { 362 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(10)); 363 } 364 365 /* efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers 366 * 367 * @rx_queue: Efx RX queue 368 * 369 * This allocates a batch of pages, maps them for DMA, and populates 370 * struct efx_rx_buffers for each one. Return a negative error code or 371 * 0 on success. If a single page can be used for multiple buffers, 372 * then the page will either be inserted fully, or not at all. 373 */ 374 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic) 375 { 376 unsigned int page_offset, index, count; 377 struct efx_nic *efx = rx_queue->efx; 378 struct efx_rx_page_state *state; 379 struct efx_rx_buffer *rx_buf; 380 dma_addr_t dma_addr; 381 struct page *page; 382 383 count = 0; 384 do { 385 page = efx_reuse_page(rx_queue); 386 if (page == NULL) { 387 page = alloc_pages(__GFP_COMP | 388 (atomic ? GFP_ATOMIC : GFP_KERNEL), 389 efx->rx_buffer_order); 390 if (unlikely(page == NULL)) 391 return -ENOMEM; 392 dma_addr = 393 dma_map_page(&efx->pci_dev->dev, page, 0, 394 PAGE_SIZE << efx->rx_buffer_order, 395 DMA_FROM_DEVICE); 396 if (unlikely(dma_mapping_error(&efx->pci_dev->dev, 397 dma_addr))) { 398 __free_pages(page, efx->rx_buffer_order); 399 return -EIO; 400 } 401 state = page_address(page); 402 state->dma_addr = dma_addr; 403 } else { 404 state = page_address(page); 405 dma_addr = state->dma_addr; 406 } 407 408 dma_addr += sizeof(struct efx_rx_page_state); 409 page_offset = sizeof(struct efx_rx_page_state); 410 411 do { 412 index = rx_queue->added_count & rx_queue->ptr_mask; 413 rx_buf = efx_rx_buffer(rx_queue, index); 414 rx_buf->dma_addr = dma_addr + efx->rx_ip_align + 415 EFX_XDP_HEADROOM; 416 rx_buf->page = page; 417 rx_buf->page_offset = page_offset + efx->rx_ip_align + 418 EFX_XDP_HEADROOM; 419 rx_buf->len = efx->rx_dma_len; 420 rx_buf->flags = 0; 421 ++rx_queue->added_count; 422 get_page(page); 423 dma_addr += efx->rx_page_buf_step; 424 page_offset += efx->rx_page_buf_step; 425 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE); 426 427 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE; 428 } while (++count < efx->rx_pages_per_batch); 429 430 return 0; 431 } 432 433 void efx_rx_config_page_split(struct efx_nic *efx) 434 { 435 efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align + 436 EFX_XDP_HEADROOM + EFX_XDP_TAILROOM, 437 EFX_RX_BUF_ALIGNMENT); 438 efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 : 439 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) / 440 efx->rx_page_buf_step); 441 efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) / 442 efx->rx_bufs_per_page; 443 efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH, 444 efx->rx_bufs_per_page); 445 } 446 447 /* efx_fast_push_rx_descriptors - push new RX descriptors quickly 448 * @rx_queue: RX descriptor queue 449 * 450 * This will aim to fill the RX descriptor queue up to 451 * @rx_queue->@max_fill. If there is insufficient atomic 452 * memory to do so, a slow fill will be scheduled. 453 * 454 * The caller must provide serialisation (none is used here). In practise, 455 * this means this function must run from the NAPI handler, or be called 456 * when NAPI is disabled. 457 */ 458 void efx_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, bool atomic) 459 { 460 struct efx_nic *efx = rx_queue->efx; 461 unsigned int fill_level, batch_size; 462 int space, rc = 0; 463 464 if (!rx_queue->refill_enabled) 465 return; 466 467 /* Calculate current fill level, and exit if we don't need to fill */ 468 fill_level = (rx_queue->added_count - rx_queue->removed_count); 469 EFX_WARN_ON_ONCE_PARANOID(fill_level > rx_queue->efx->rxq_entries); 470 if (fill_level >= rx_queue->fast_fill_trigger) 471 goto out; 472 473 /* Record minimum fill level */ 474 if (unlikely(fill_level < rx_queue->min_fill)) { 475 if (fill_level) 476 rx_queue->min_fill = fill_level; 477 } 478 479 batch_size = efx->rx_pages_per_batch * efx->rx_bufs_per_page; 480 space = rx_queue->max_fill - fill_level; 481 EFX_WARN_ON_ONCE_PARANOID(space < batch_size); 482 483 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev, 484 "RX queue %d fast-filling descriptor ring from" 485 " level %d to level %d\n", 486 efx_rx_queue_index(rx_queue), fill_level, 487 rx_queue->max_fill); 488 489 do { 490 rc = efx_init_rx_buffers(rx_queue, atomic); 491 if (unlikely(rc)) { 492 /* Ensure that we don't leave the rx queue empty */ 493 efx_schedule_slow_fill(rx_queue); 494 goto out; 495 } 496 } while ((space -= batch_size) >= batch_size); 497 498 netif_vdbg(rx_queue->efx, rx_status, rx_queue->efx->net_dev, 499 "RX queue %d fast-filled descriptor ring " 500 "to level %d\n", efx_rx_queue_index(rx_queue), 501 rx_queue->added_count - rx_queue->removed_count); 502 503 out: 504 if (rx_queue->notified_count != rx_queue->added_count) 505 efx_nic_notify_rx_desc(rx_queue); 506 } 507 508 /* Pass a received packet up through GRO. GRO can handle pages 509 * regardless of checksum state and skbs with a good checksum. 510 */ 511 void 512 efx_rx_packet_gro(struct efx_channel *channel, struct efx_rx_buffer *rx_buf, 513 unsigned int n_frags, u8 *eh, __wsum csum) 514 { 515 struct napi_struct *napi = &channel->napi_str; 516 struct efx_nic *efx = channel->efx; 517 struct sk_buff *skb; 518 519 skb = napi_get_frags(napi); 520 if (unlikely(!skb)) { 521 struct efx_rx_queue *rx_queue; 522 523 rx_queue = efx_channel_get_rx_queue(channel); 524 efx_free_rx_buffers(rx_queue, rx_buf, n_frags); 525 return; 526 } 527 528 if (efx->net_dev->features & NETIF_F_RXHASH && 529 efx_rx_buf_hash_valid(efx, eh)) 530 skb_set_hash(skb, efx_rx_buf_hash(efx, eh), 531 PKT_HASH_TYPE_L3); 532 if (csum) { 533 skb->csum = csum; 534 skb->ip_summed = CHECKSUM_COMPLETE; 535 } else { 536 skb->ip_summed = ((rx_buf->flags & EFX_RX_PKT_CSUMMED) ? 537 CHECKSUM_UNNECESSARY : CHECKSUM_NONE); 538 } 539 skb->csum_level = !!(rx_buf->flags & EFX_RX_PKT_CSUM_LEVEL); 540 541 for (;;) { 542 skb_fill_page_desc(skb, skb_shinfo(skb)->nr_frags, 543 rx_buf->page, rx_buf->page_offset, 544 rx_buf->len); 545 rx_buf->page = NULL; 546 skb->len += rx_buf->len; 547 if (skb_shinfo(skb)->nr_frags == n_frags) 548 break; 549 550 rx_buf = efx_rx_buf_next(&channel->rx_queue, rx_buf); 551 } 552 553 skb->data_len = skb->len; 554 skb->truesize += n_frags * efx->rx_buffer_truesize; 555 556 skb_record_rx_queue(skb, channel->rx_queue.core_index); 557 558 napi_gro_frags(napi); 559 } 560 561 /* RSS contexts. We're using linked lists and crappy O(n) algorithms, because 562 * (a) this is an infrequent control-plane operation and (b) n is small (max 64) 563 */ 564 struct efx_rss_context *efx_alloc_rss_context_entry(struct efx_nic *efx) 565 { 566 struct list_head *head = &efx->rss_context.list; 567 struct efx_rss_context *ctx, *new; 568 u32 id = 1; /* Don't use zero, that refers to the master RSS context */ 569 570 WARN_ON(!mutex_is_locked(&efx->rss_lock)); 571 572 /* Search for first gap in the numbering */ 573 list_for_each_entry(ctx, head, list) { 574 if (ctx->user_id != id) 575 break; 576 id++; 577 /* Check for wrap. If this happens, we have nearly 2^32 578 * allocated RSS contexts, which seems unlikely. 579 */ 580 if (WARN_ON_ONCE(!id)) 581 return NULL; 582 } 583 584 /* Create the new entry */ 585 new = kmalloc(sizeof(*new), GFP_KERNEL); 586 if (!new) 587 return NULL; 588 new->context_id = EFX_MCDI_RSS_CONTEXT_INVALID; 589 new->rx_hash_udp_4tuple = false; 590 591 /* Insert the new entry into the gap */ 592 new->user_id = id; 593 list_add_tail(&new->list, &ctx->list); 594 return new; 595 } 596 597 struct efx_rss_context *efx_find_rss_context_entry(struct efx_nic *efx, u32 id) 598 { 599 struct list_head *head = &efx->rss_context.list; 600 struct efx_rss_context *ctx; 601 602 WARN_ON(!mutex_is_locked(&efx->rss_lock)); 603 604 list_for_each_entry(ctx, head, list) 605 if (ctx->user_id == id) 606 return ctx; 607 return NULL; 608 } 609 610 void efx_free_rss_context_entry(struct efx_rss_context *ctx) 611 { 612 list_del(&ctx->list); 613 kfree(ctx); 614 } 615 616 void efx_set_default_rx_indir_table(struct efx_nic *efx, 617 struct efx_rss_context *ctx) 618 { 619 size_t i; 620 621 for (i = 0; i < ARRAY_SIZE(ctx->rx_indir_table); i++) 622 ctx->rx_indir_table[i] = 623 ethtool_rxfh_indir_default(i, efx->rss_spread); 624 } 625 626 /** 627 * efx_filter_is_mc_recipient - test whether spec is a multicast recipient 628 * @spec: Specification to test 629 * 630 * Return: %true if the specification is a non-drop RX filter that 631 * matches a local MAC address I/G bit value of 1 or matches a local 632 * IPv4 or IPv6 address value in the respective multicast address 633 * range. Otherwise %false. 634 */ 635 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec) 636 { 637 if (!(spec->flags & EFX_FILTER_FLAG_RX) || 638 spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP) 639 return false; 640 641 if (spec->match_flags & 642 (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) && 643 is_multicast_ether_addr(spec->loc_mac)) 644 return true; 645 646 if ((spec->match_flags & 647 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) == 648 (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) { 649 if (spec->ether_type == htons(ETH_P_IP) && 650 ipv4_is_multicast(spec->loc_host[0])) 651 return true; 652 if (spec->ether_type == htons(ETH_P_IPV6) && 653 ((const u8 *)spec->loc_host)[0] == 0xff) 654 return true; 655 } 656 657 return false; 658 } 659 660 bool efx_filter_spec_equal(const struct efx_filter_spec *left, 661 const struct efx_filter_spec *right) 662 { 663 if ((left->match_flags ^ right->match_flags) | 664 ((left->flags ^ right->flags) & 665 (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX))) 666 return false; 667 668 return memcmp(&left->outer_vid, &right->outer_vid, 669 sizeof(struct efx_filter_spec) - 670 offsetof(struct efx_filter_spec, outer_vid)) == 0; 671 } 672 673 u32 efx_filter_spec_hash(const struct efx_filter_spec *spec) 674 { 675 BUILD_BUG_ON(offsetof(struct efx_filter_spec, outer_vid) & 3); 676 return jhash2((const u32 *)&spec->outer_vid, 677 (sizeof(struct efx_filter_spec) - 678 offsetof(struct efx_filter_spec, outer_vid)) / 4, 679 0); 680 } 681 682 #ifdef CONFIG_RFS_ACCEL 683 bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx, 684 bool *force) 685 { 686 if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) { 687 /* ARFS is currently updating this entry, leave it */ 688 return false; 689 } 690 if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) { 691 /* ARFS tried and failed to update this, so it's probably out 692 * of date. Remove the filter and the ARFS rule entry. 693 */ 694 rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING; 695 *force = true; 696 return true; 697 } else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */ 698 /* ARFS has moved on, so old filter is not needed. Since we did 699 * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will 700 * not be removed by efx_rps_hash_del() subsequently. 701 */ 702 *force = true; 703 return true; 704 } 705 /* Remove it iff ARFS wants to. */ 706 return true; 707 } 708 709 static 710 struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx, 711 const struct efx_filter_spec *spec) 712 { 713 u32 hash = efx_filter_spec_hash(spec); 714 715 lockdep_assert_held(&efx->rps_hash_lock); 716 if (!efx->rps_hash_table) 717 return NULL; 718 return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE]; 719 } 720 721 struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx, 722 const struct efx_filter_spec *spec) 723 { 724 struct efx_arfs_rule *rule; 725 struct hlist_head *head; 726 struct hlist_node *node; 727 728 head = efx_rps_hash_bucket(efx, spec); 729 if (!head) 730 return NULL; 731 hlist_for_each(node, head) { 732 rule = container_of(node, struct efx_arfs_rule, node); 733 if (efx_filter_spec_equal(spec, &rule->spec)) 734 return rule; 735 } 736 return NULL; 737 } 738 739 struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx, 740 const struct efx_filter_spec *spec, 741 bool *new) 742 { 743 struct efx_arfs_rule *rule; 744 struct hlist_head *head; 745 struct hlist_node *node; 746 747 head = efx_rps_hash_bucket(efx, spec); 748 if (!head) 749 return NULL; 750 hlist_for_each(node, head) { 751 rule = container_of(node, struct efx_arfs_rule, node); 752 if (efx_filter_spec_equal(spec, &rule->spec)) { 753 *new = false; 754 return rule; 755 } 756 } 757 rule = kmalloc(sizeof(*rule), GFP_ATOMIC); 758 *new = true; 759 if (rule) { 760 memcpy(&rule->spec, spec, sizeof(rule->spec)); 761 hlist_add_head(&rule->node, head); 762 } 763 return rule; 764 } 765 766 void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec) 767 { 768 struct efx_arfs_rule *rule; 769 struct hlist_head *head; 770 struct hlist_node *node; 771 772 head = efx_rps_hash_bucket(efx, spec); 773 if (WARN_ON(!head)) 774 return; 775 hlist_for_each(node, head) { 776 rule = container_of(node, struct efx_arfs_rule, node); 777 if (efx_filter_spec_equal(spec, &rule->spec)) { 778 /* Someone already reused the entry. We know that if 779 * this check doesn't fire (i.e. filter_id == REMOVING) 780 * then the REMOVING mark was put there by our caller, 781 * because caller is holding a lock on filter table and 782 * only holders of that lock set REMOVING. 783 */ 784 if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING) 785 return; 786 hlist_del(node); 787 kfree(rule); 788 return; 789 } 790 } 791 /* We didn't find it. */ 792 WARN_ON(1); 793 } 794 #endif 795 796 int efx_probe_filters(struct efx_nic *efx) 797 { 798 int rc; 799 800 mutex_lock(&efx->mac_lock); 801 down_write(&efx->filter_sem); 802 rc = efx->type->filter_table_probe(efx); 803 if (rc) 804 goto out_unlock; 805 806 #ifdef CONFIG_RFS_ACCEL 807 if (efx->type->offload_features & NETIF_F_NTUPLE) { 808 struct efx_channel *channel; 809 int i, success = 1; 810 811 efx_for_each_channel(channel, efx) { 812 channel->rps_flow_id = 813 kcalloc(efx->type->max_rx_ip_filters, 814 sizeof(*channel->rps_flow_id), 815 GFP_KERNEL); 816 if (!channel->rps_flow_id) 817 success = 0; 818 else 819 for (i = 0; 820 i < efx->type->max_rx_ip_filters; 821 ++i) 822 channel->rps_flow_id[i] = 823 RPS_FLOW_ID_INVALID; 824 channel->rfs_expire_index = 0; 825 channel->rfs_filter_count = 0; 826 } 827 828 if (!success) { 829 efx_for_each_channel(channel, efx) 830 kfree(channel->rps_flow_id); 831 efx->type->filter_table_remove(efx); 832 rc = -ENOMEM; 833 goto out_unlock; 834 } 835 } 836 #endif 837 out_unlock: 838 up_write(&efx->filter_sem); 839 mutex_unlock(&efx->mac_lock); 840 return rc; 841 } 842 843 void efx_remove_filters(struct efx_nic *efx) 844 { 845 #ifdef CONFIG_RFS_ACCEL 846 struct efx_channel *channel; 847 848 efx_for_each_channel(channel, efx) { 849 cancel_delayed_work_sync(&channel->filter_work); 850 kfree(channel->rps_flow_id); 851 channel->rps_flow_id = NULL; 852 } 853 #endif 854 down_write(&efx->filter_sem); 855 efx->type->filter_table_remove(efx); 856 up_write(&efx->filter_sem); 857 } 858 859 #ifdef CONFIG_RFS_ACCEL 860 861 static void efx_filter_rfs_work(struct work_struct *data) 862 { 863 struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion, 864 work); 865 struct efx_nic *efx = netdev_priv(req->net_dev); 866 struct efx_channel *channel = efx_get_channel(efx, req->rxq_index); 867 int slot_idx = req - efx->rps_slot; 868 struct efx_arfs_rule *rule; 869 u16 arfs_id = 0; 870 int rc; 871 872 rc = efx->type->filter_insert(efx, &req->spec, true); 873 if (rc >= 0) 874 /* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */ 875 rc %= efx->type->max_rx_ip_filters; 876 if (efx->rps_hash_table) { 877 spin_lock_bh(&efx->rps_hash_lock); 878 rule = efx_rps_hash_find(efx, &req->spec); 879 /* The rule might have already gone, if someone else's request 880 * for the same spec was already worked and then expired before 881 * we got around to our work. In that case we have nothing 882 * tying us to an arfs_id, meaning that as soon as the filter 883 * is considered for expiry it will be removed. 884 */ 885 if (rule) { 886 if (rc < 0) 887 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR; 888 else 889 rule->filter_id = rc; 890 arfs_id = rule->arfs_id; 891 } 892 spin_unlock_bh(&efx->rps_hash_lock); 893 } 894 if (rc >= 0) { 895 /* Remember this so we can check whether to expire the filter 896 * later. 897 */ 898 mutex_lock(&efx->rps_mutex); 899 if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID) 900 channel->rfs_filter_count++; 901 channel->rps_flow_id[rc] = req->flow_id; 902 mutex_unlock(&efx->rps_mutex); 903 904 if (req->spec.ether_type == htons(ETH_P_IP)) 905 netif_info(efx, rx_status, efx->net_dev, 906 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n", 907 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 908 req->spec.rem_host, ntohs(req->spec.rem_port), 909 req->spec.loc_host, ntohs(req->spec.loc_port), 910 req->rxq_index, req->flow_id, rc, arfs_id); 911 else 912 netif_info(efx, rx_status, efx->net_dev, 913 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n", 914 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 915 req->spec.rem_host, ntohs(req->spec.rem_port), 916 req->spec.loc_host, ntohs(req->spec.loc_port), 917 req->rxq_index, req->flow_id, rc, arfs_id); 918 channel->n_rfs_succeeded++; 919 } else { 920 if (req->spec.ether_type == htons(ETH_P_IP)) 921 netif_dbg(efx, rx_status, efx->net_dev, 922 "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n", 923 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 924 req->spec.rem_host, ntohs(req->spec.rem_port), 925 req->spec.loc_host, ntohs(req->spec.loc_port), 926 req->rxq_index, req->flow_id, rc, arfs_id); 927 else 928 netif_dbg(efx, rx_status, efx->net_dev, 929 "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n", 930 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 931 req->spec.rem_host, ntohs(req->spec.rem_port), 932 req->spec.loc_host, ntohs(req->spec.loc_port), 933 req->rxq_index, req->flow_id, rc, arfs_id); 934 channel->n_rfs_failed++; 935 /* We're overloading the NIC's filter tables, so let's do a 936 * chunk of extra expiry work. 937 */ 938 __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count, 939 100u)); 940 } 941 942 /* Release references */ 943 clear_bit(slot_idx, &efx->rps_slot_map); 944 dev_put(req->net_dev); 945 } 946 947 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb, 948 u16 rxq_index, u32 flow_id) 949 { 950 struct efx_nic *efx = netdev_priv(net_dev); 951 struct efx_async_filter_insertion *req; 952 struct efx_arfs_rule *rule; 953 struct flow_keys fk; 954 int slot_idx; 955 bool new; 956 int rc; 957 958 /* find a free slot */ 959 for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++) 960 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map)) 961 break; 962 if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT) 963 return -EBUSY; 964 965 if (flow_id == RPS_FLOW_ID_INVALID) { 966 rc = -EINVAL; 967 goto out_clear; 968 } 969 970 if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) { 971 rc = -EPROTONOSUPPORT; 972 goto out_clear; 973 } 974 975 if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) { 976 rc = -EPROTONOSUPPORT; 977 goto out_clear; 978 } 979 if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) { 980 rc = -EPROTONOSUPPORT; 981 goto out_clear; 982 } 983 984 req = efx->rps_slot + slot_idx; 985 efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT, 986 efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0, 987 rxq_index); 988 req->spec.match_flags = 989 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO | 990 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT | 991 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT; 992 req->spec.ether_type = fk.basic.n_proto; 993 req->spec.ip_proto = fk.basic.ip_proto; 994 995 if (fk.basic.n_proto == htons(ETH_P_IP)) { 996 req->spec.rem_host[0] = fk.addrs.v4addrs.src; 997 req->spec.loc_host[0] = fk.addrs.v4addrs.dst; 998 } else { 999 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src, 1000 sizeof(struct in6_addr)); 1001 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst, 1002 sizeof(struct in6_addr)); 1003 } 1004 1005 req->spec.rem_port = fk.ports.src; 1006 req->spec.loc_port = fk.ports.dst; 1007 1008 if (efx->rps_hash_table) { 1009 /* Add it to ARFS hash table */ 1010 spin_lock(&efx->rps_hash_lock); 1011 rule = efx_rps_hash_add(efx, &req->spec, &new); 1012 if (!rule) { 1013 rc = -ENOMEM; 1014 goto out_unlock; 1015 } 1016 if (new) 1017 rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER; 1018 rc = rule->arfs_id; 1019 /* Skip if existing or pending filter already does the right thing */ 1020 if (!new && rule->rxq_index == rxq_index && 1021 rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING) 1022 goto out_unlock; 1023 rule->rxq_index = rxq_index; 1024 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING; 1025 spin_unlock(&efx->rps_hash_lock); 1026 } else { 1027 /* Without an ARFS hash table, we just use arfs_id 0 for all 1028 * filters. This means if multiple flows hash to the same 1029 * flow_id, all but the most recently touched will be eligible 1030 * for expiry. 1031 */ 1032 rc = 0; 1033 } 1034 1035 /* Queue the request */ 1036 dev_hold(req->net_dev = net_dev); 1037 INIT_WORK(&req->work, efx_filter_rfs_work); 1038 req->rxq_index = rxq_index; 1039 req->flow_id = flow_id; 1040 schedule_work(&req->work); 1041 return rc; 1042 out_unlock: 1043 spin_unlock(&efx->rps_hash_lock); 1044 out_clear: 1045 clear_bit(slot_idx, &efx->rps_slot_map); 1046 return rc; 1047 } 1048 1049 bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota) 1050 { 1051 bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index); 1052 struct efx_nic *efx = channel->efx; 1053 unsigned int index, size, start; 1054 u32 flow_id; 1055 1056 if (!mutex_trylock(&efx->rps_mutex)) 1057 return false; 1058 expire_one = efx->type->filter_rfs_expire_one; 1059 index = channel->rfs_expire_index; 1060 start = index; 1061 size = efx->type->max_rx_ip_filters; 1062 while (quota) { 1063 flow_id = channel->rps_flow_id[index]; 1064 1065 if (flow_id != RPS_FLOW_ID_INVALID) { 1066 quota--; 1067 if (expire_one(efx, flow_id, index)) { 1068 netif_info(efx, rx_status, efx->net_dev, 1069 "expired filter %d [channel %u flow %u]\n", 1070 index, channel->channel, flow_id); 1071 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID; 1072 channel->rfs_filter_count--; 1073 } 1074 } 1075 if (++index == size) 1076 index = 0; 1077 /* If we were called with a quota that exceeds the total number 1078 * of filters in the table (which shouldn't happen, but could 1079 * if two callers race), ensure that we don't loop forever - 1080 * stop when we've examined every row of the table. 1081 */ 1082 if (index == start) 1083 break; 1084 } 1085 1086 channel->rfs_expire_index = index; 1087 mutex_unlock(&efx->rps_mutex); 1088 return true; 1089 } 1090 1091 #endif /* CONFIG_RFS_ACCEL */ 1092