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 /* RX maximum head room required. 27 * 28 * This must be at least 1 to prevent overflow, plus one packet-worth 29 * to allow pipelined receives. 30 */ 31 #define EFX_RXD_HEAD_ROOM (1 + EFX_RX_MAX_FRAGS) 32 33 static void efx_unmap_rx_buffer(struct efx_nic *efx, 34 struct efx_rx_buffer *rx_buf); 35 36 /* Check the RX page recycle ring for a page that can be reused. */ 37 static struct page *efx_reuse_page(struct efx_rx_queue *rx_queue) 38 { 39 struct efx_nic *efx = rx_queue->efx; 40 struct efx_rx_page_state *state; 41 unsigned int index; 42 struct page *page; 43 44 if (unlikely(!rx_queue->page_ring)) 45 return NULL; 46 index = rx_queue->page_remove & rx_queue->page_ptr_mask; 47 page = rx_queue->page_ring[index]; 48 if (page == NULL) 49 return NULL; 50 51 rx_queue->page_ring[index] = NULL; 52 /* page_remove cannot exceed page_add. */ 53 if (rx_queue->page_remove != rx_queue->page_add) 54 ++rx_queue->page_remove; 55 56 /* If page_count is 1 then we hold the only reference to this page. */ 57 if (page_count(page) == 1) { 58 ++rx_queue->page_recycle_count; 59 return page; 60 } else { 61 state = page_address(page); 62 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr, 63 PAGE_SIZE << efx->rx_buffer_order, 64 DMA_FROM_DEVICE); 65 put_page(page); 66 ++rx_queue->page_recycle_failed; 67 } 68 69 return NULL; 70 } 71 72 /* Attempt to recycle the page if there is an RX recycle ring; the page can 73 * only be added if this is the final RX buffer, to prevent pages being used in 74 * the descriptor ring and appearing in the recycle ring simultaneously. 75 */ 76 static void efx_recycle_rx_page(struct efx_channel *channel, 77 struct efx_rx_buffer *rx_buf) 78 { 79 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 80 struct efx_nic *efx = rx_queue->efx; 81 struct page *page = rx_buf->page; 82 unsigned int index; 83 84 /* Only recycle the page after processing the final buffer. */ 85 if (!(rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE)) 86 return; 87 88 index = rx_queue->page_add & rx_queue->page_ptr_mask; 89 if (rx_queue->page_ring[index] == NULL) { 90 unsigned int read_index = rx_queue->page_remove & 91 rx_queue->page_ptr_mask; 92 93 /* The next slot in the recycle ring is available, but 94 * increment page_remove if the read pointer currently 95 * points here. 96 */ 97 if (read_index == index) 98 ++rx_queue->page_remove; 99 rx_queue->page_ring[index] = page; 100 ++rx_queue->page_add; 101 return; 102 } 103 ++rx_queue->page_recycle_full; 104 efx_unmap_rx_buffer(efx, rx_buf); 105 put_page(rx_buf->page); 106 } 107 108 /* Recycle the pages that are used by buffers that have just been received. */ 109 void efx_siena_recycle_rx_pages(struct efx_channel *channel, 110 struct efx_rx_buffer *rx_buf, 111 unsigned int n_frags) 112 { 113 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 114 115 if (unlikely(!rx_queue->page_ring)) 116 return; 117 118 do { 119 efx_recycle_rx_page(channel, rx_buf); 120 rx_buf = efx_rx_buf_next(rx_queue, rx_buf); 121 } while (--n_frags); 122 } 123 124 void efx_siena_discard_rx_packet(struct efx_channel *channel, 125 struct efx_rx_buffer *rx_buf, 126 unsigned int n_frags) 127 { 128 struct efx_rx_queue *rx_queue = efx_channel_get_rx_queue(channel); 129 130 efx_siena_recycle_rx_pages(channel, rx_buf, n_frags); 131 132 efx_siena_free_rx_buffers(rx_queue, rx_buf, n_frags); 133 } 134 135 static void efx_init_rx_recycle_ring(struct efx_rx_queue *rx_queue) 136 { 137 unsigned int bufs_in_recycle_ring, page_ring_size; 138 struct efx_nic *efx = rx_queue->efx; 139 140 bufs_in_recycle_ring = efx_rx_recycle_ring_size(efx); 141 page_ring_size = roundup_pow_of_two(bufs_in_recycle_ring / 142 efx->rx_bufs_per_page); 143 rx_queue->page_ring = kcalloc(page_ring_size, 144 sizeof(*rx_queue->page_ring), GFP_KERNEL); 145 if (!rx_queue->page_ring) 146 rx_queue->page_ptr_mask = 0; 147 else 148 rx_queue->page_ptr_mask = page_ring_size - 1; 149 } 150 151 static void efx_fini_rx_recycle_ring(struct efx_rx_queue *rx_queue) 152 { 153 struct efx_nic *efx = rx_queue->efx; 154 int i; 155 156 if (unlikely(!rx_queue->page_ring)) 157 return; 158 159 /* Unmap and release the pages in the recycle ring. Remove the ring. */ 160 for (i = 0; i <= rx_queue->page_ptr_mask; i++) { 161 struct page *page = rx_queue->page_ring[i]; 162 struct efx_rx_page_state *state; 163 164 if (page == NULL) 165 continue; 166 167 state = page_address(page); 168 dma_unmap_page(&efx->pci_dev->dev, state->dma_addr, 169 PAGE_SIZE << efx->rx_buffer_order, 170 DMA_FROM_DEVICE); 171 put_page(page); 172 } 173 kfree(rx_queue->page_ring); 174 rx_queue->page_ring = NULL; 175 } 176 177 static void efx_fini_rx_buffer(struct efx_rx_queue *rx_queue, 178 struct efx_rx_buffer *rx_buf) 179 { 180 /* Release the page reference we hold for the buffer. */ 181 if (rx_buf->page) 182 put_page(rx_buf->page); 183 184 /* If this is the last buffer in a page, unmap and free it. */ 185 if (rx_buf->flags & EFX_RX_BUF_LAST_IN_PAGE) { 186 efx_unmap_rx_buffer(rx_queue->efx, rx_buf); 187 efx_siena_free_rx_buffers(rx_queue, rx_buf, 1); 188 } 189 rx_buf->page = NULL; 190 } 191 192 int efx_siena_probe_rx_queue(struct efx_rx_queue *rx_queue) 193 { 194 struct efx_nic *efx = rx_queue->efx; 195 unsigned int entries; 196 int rc; 197 198 /* Create the smallest power-of-two aligned ring */ 199 entries = max(roundup_pow_of_two(efx->rxq_entries), EFX_MIN_DMAQ_SIZE); 200 EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE); 201 rx_queue->ptr_mask = entries - 1; 202 203 netif_dbg(efx, probe, efx->net_dev, 204 "creating RX queue %d size %#x mask %#x\n", 205 efx_rx_queue_index(rx_queue), efx->rxq_entries, 206 rx_queue->ptr_mask); 207 208 /* Allocate RX buffers */ 209 rx_queue->buffer = kcalloc(entries, sizeof(*rx_queue->buffer), 210 GFP_KERNEL); 211 if (!rx_queue->buffer) 212 return -ENOMEM; 213 214 rc = efx_nic_probe_rx(rx_queue); 215 if (rc) { 216 kfree(rx_queue->buffer); 217 rx_queue->buffer = NULL; 218 } 219 220 return rc; 221 } 222 223 void efx_siena_init_rx_queue(struct efx_rx_queue *rx_queue) 224 { 225 unsigned int max_fill, trigger, max_trigger; 226 struct efx_nic *efx = rx_queue->efx; 227 int rc = 0; 228 229 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 230 "initialising RX queue %d\n", efx_rx_queue_index(rx_queue)); 231 232 /* Initialise ptr fields */ 233 rx_queue->added_count = 0; 234 rx_queue->notified_count = 0; 235 rx_queue->removed_count = 0; 236 rx_queue->min_fill = -1U; 237 efx_init_rx_recycle_ring(rx_queue); 238 239 rx_queue->page_remove = 0; 240 rx_queue->page_add = rx_queue->page_ptr_mask + 1; 241 rx_queue->page_recycle_count = 0; 242 rx_queue->page_recycle_failed = 0; 243 rx_queue->page_recycle_full = 0; 244 245 /* Initialise limit fields */ 246 max_fill = efx->rxq_entries - EFX_RXD_HEAD_ROOM; 247 max_trigger = 248 max_fill - efx->rx_pages_per_batch * efx->rx_bufs_per_page; 249 if (rx_refill_threshold != 0) { 250 trigger = max_fill * min(rx_refill_threshold, 100U) / 100U; 251 if (trigger > max_trigger) 252 trigger = max_trigger; 253 } else { 254 trigger = max_trigger; 255 } 256 257 rx_queue->max_fill = max_fill; 258 rx_queue->fast_fill_trigger = trigger; 259 rx_queue->refill_enabled = true; 260 261 /* Initialise XDP queue information */ 262 rc = xdp_rxq_info_reg(&rx_queue->xdp_rxq_info, efx->net_dev, 263 rx_queue->core_index, 0); 264 265 if (rc) { 266 netif_err(efx, rx_err, efx->net_dev, 267 "Failure to initialise XDP queue information rc=%d\n", 268 rc); 269 efx->xdp_rxq_info_failed = true; 270 } else { 271 rx_queue->xdp_rxq_info_valid = true; 272 } 273 274 /* Set up RX descriptor ring */ 275 efx_nic_init_rx(rx_queue); 276 } 277 278 void efx_siena_fini_rx_queue(struct efx_rx_queue *rx_queue) 279 { 280 struct efx_rx_buffer *rx_buf; 281 int i; 282 283 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 284 "shutting down RX queue %d\n", efx_rx_queue_index(rx_queue)); 285 286 del_timer_sync(&rx_queue->slow_fill); 287 288 /* Release RX buffers from the current read ptr to the write ptr */ 289 if (rx_queue->buffer) { 290 for (i = rx_queue->removed_count; i < rx_queue->added_count; 291 i++) { 292 unsigned int index = i & rx_queue->ptr_mask; 293 294 rx_buf = efx_rx_buffer(rx_queue, index); 295 efx_fini_rx_buffer(rx_queue, rx_buf); 296 } 297 } 298 299 efx_fini_rx_recycle_ring(rx_queue); 300 301 if (rx_queue->xdp_rxq_info_valid) 302 xdp_rxq_info_unreg(&rx_queue->xdp_rxq_info); 303 304 rx_queue->xdp_rxq_info_valid = false; 305 } 306 307 void efx_siena_remove_rx_queue(struct efx_rx_queue *rx_queue) 308 { 309 netif_dbg(rx_queue->efx, drv, rx_queue->efx->net_dev, 310 "destroying RX queue %d\n", efx_rx_queue_index(rx_queue)); 311 312 efx_nic_remove_rx(rx_queue); 313 314 kfree(rx_queue->buffer); 315 rx_queue->buffer = NULL; 316 } 317 318 /* Unmap a DMA-mapped page. This function is only called for the final RX 319 * buffer in a page. 320 */ 321 static void efx_unmap_rx_buffer(struct efx_nic *efx, 322 struct efx_rx_buffer *rx_buf) 323 { 324 struct page *page = rx_buf->page; 325 326 if (page) { 327 struct efx_rx_page_state *state = page_address(page); 328 329 dma_unmap_page(&efx->pci_dev->dev, 330 state->dma_addr, 331 PAGE_SIZE << efx->rx_buffer_order, 332 DMA_FROM_DEVICE); 333 } 334 } 335 336 void efx_siena_free_rx_buffers(struct efx_rx_queue *rx_queue, 337 struct efx_rx_buffer *rx_buf, 338 unsigned int num_bufs) 339 { 340 do { 341 if (rx_buf->page) { 342 put_page(rx_buf->page); 343 rx_buf->page = NULL; 344 } 345 rx_buf = efx_rx_buf_next(rx_queue, rx_buf); 346 } while (--num_bufs); 347 } 348 349 void efx_siena_rx_slow_fill(struct timer_list *t) 350 { 351 struct efx_rx_queue *rx_queue = from_timer(rx_queue, t, slow_fill); 352 353 /* Post an event to cause NAPI to run and refill the queue */ 354 efx_nic_generate_fill_event(rx_queue); 355 ++rx_queue->slow_fill_count; 356 } 357 358 static void efx_schedule_slow_fill(struct efx_rx_queue *rx_queue) 359 { 360 mod_timer(&rx_queue->slow_fill, jiffies + msecs_to_jiffies(10)); 361 } 362 363 /* efx_init_rx_buffers - create EFX_RX_BATCH page-based RX buffers 364 * 365 * @rx_queue: Efx RX queue 366 * 367 * This allocates a batch of pages, maps them for DMA, and populates 368 * struct efx_rx_buffers for each one. Return a negative error code or 369 * 0 on success. If a single page can be used for multiple buffers, 370 * then the page will either be inserted fully, or not at all. 371 */ 372 static int efx_init_rx_buffers(struct efx_rx_queue *rx_queue, bool atomic) 373 { 374 unsigned int page_offset, index, count; 375 struct efx_nic *efx = rx_queue->efx; 376 struct efx_rx_page_state *state; 377 struct efx_rx_buffer *rx_buf; 378 dma_addr_t dma_addr; 379 struct page *page; 380 381 count = 0; 382 do { 383 page = efx_reuse_page(rx_queue); 384 if (page == NULL) { 385 page = alloc_pages(__GFP_COMP | 386 (atomic ? GFP_ATOMIC : GFP_KERNEL), 387 efx->rx_buffer_order); 388 if (unlikely(page == NULL)) 389 return -ENOMEM; 390 dma_addr = 391 dma_map_page(&efx->pci_dev->dev, page, 0, 392 PAGE_SIZE << efx->rx_buffer_order, 393 DMA_FROM_DEVICE); 394 if (unlikely(dma_mapping_error(&efx->pci_dev->dev, 395 dma_addr))) { 396 __free_pages(page, efx->rx_buffer_order); 397 return -EIO; 398 } 399 state = page_address(page); 400 state->dma_addr = dma_addr; 401 } else { 402 state = page_address(page); 403 dma_addr = state->dma_addr; 404 } 405 406 dma_addr += sizeof(struct efx_rx_page_state); 407 page_offset = sizeof(struct efx_rx_page_state); 408 409 do { 410 index = rx_queue->added_count & rx_queue->ptr_mask; 411 rx_buf = efx_rx_buffer(rx_queue, index); 412 rx_buf->dma_addr = dma_addr + efx->rx_ip_align + 413 EFX_XDP_HEADROOM; 414 rx_buf->page = page; 415 rx_buf->page_offset = page_offset + efx->rx_ip_align + 416 EFX_XDP_HEADROOM; 417 rx_buf->len = efx->rx_dma_len; 418 rx_buf->flags = 0; 419 ++rx_queue->added_count; 420 get_page(page); 421 dma_addr += efx->rx_page_buf_step; 422 page_offset += efx->rx_page_buf_step; 423 } while (page_offset + efx->rx_page_buf_step <= PAGE_SIZE); 424 425 rx_buf->flags = EFX_RX_BUF_LAST_IN_PAGE; 426 } while (++count < efx->rx_pages_per_batch); 427 428 return 0; 429 } 430 431 void efx_siena_rx_config_page_split(struct efx_nic *efx) 432 { 433 efx->rx_page_buf_step = ALIGN(efx->rx_dma_len + efx->rx_ip_align + 434 EFX_XDP_HEADROOM + EFX_XDP_TAILROOM, 435 EFX_RX_BUF_ALIGNMENT); 436 efx->rx_bufs_per_page = efx->rx_buffer_order ? 1 : 437 ((PAGE_SIZE - sizeof(struct efx_rx_page_state)) / 438 efx->rx_page_buf_step); 439 efx->rx_buffer_truesize = (PAGE_SIZE << efx->rx_buffer_order) / 440 efx->rx_bufs_per_page; 441 efx->rx_pages_per_batch = DIV_ROUND_UP(EFX_RX_PREFERRED_BATCH, 442 efx->rx_bufs_per_page); 443 } 444 445 /* efx_siena_fast_push_rx_descriptors - push new RX descriptors quickly 446 * @rx_queue: RX descriptor queue 447 * 448 * This will aim to fill the RX descriptor queue up to 449 * @rx_queue->@max_fill. If there is insufficient atomic 450 * memory to do so, a slow fill will be scheduled. 451 * 452 * The caller must provide serialisation (none is used here). In practise, 453 * this means this function must run from the NAPI handler, or be called 454 * when NAPI is disabled. 455 */ 456 void efx_siena_fast_push_rx_descriptors(struct efx_rx_queue *rx_queue, 457 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 */ 510 void 511 efx_siena_rx_packet_gro(struct efx_channel *channel, 512 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_siena_free_rx_buffers(rx_queue, rx_buf, n_frags); 525 return; 526 } 527 528 if (efx->net_dev->features & NETIF_F_RXHASH) 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 */ 563 struct efx_rss_context *efx_siena_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 } 595 596 struct efx_rss_context *efx_siena_find_rss_context_entry(struct efx_nic *efx, 597 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_siena_free_rss_context_entry(struct efx_rss_context *ctx) 611 { 612 list_del(&ctx->list); 613 kfree(ctx); 614 } 615 616 void efx_siena_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_siena_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_siena_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_siena_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_siena_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_siena_rps_check_rule(struct efx_arfs_rule *rule, 684 unsigned int filter_idx, 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_siena_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_siena_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_siena_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_siena_filter_spec_equal(spec, &rule->spec)) 734 return rule; 735 } 736 return NULL; 737 } 738 739 static 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_siena_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_siena_rps_hash_del(struct efx_nic *efx, 767 const struct efx_filter_spec *spec) 768 { 769 struct efx_arfs_rule *rule; 770 struct hlist_head *head; 771 struct hlist_node *node; 772 773 head = efx_rps_hash_bucket(efx, spec); 774 if (WARN_ON(!head)) 775 return; 776 hlist_for_each(node, head) { 777 rule = container_of(node, struct efx_arfs_rule, node); 778 if (efx_siena_filter_spec_equal(spec, &rule->spec)) { 779 /* Someone already reused the entry. We know that if 780 * this check doesn't fire (i.e. filter_id == REMOVING) 781 * then the REMOVING mark was put there by our caller, 782 * because caller is holding a lock on filter table and 783 * only holders of that lock set REMOVING. 784 */ 785 if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING) 786 return; 787 hlist_del(node); 788 kfree(rule); 789 return; 790 } 791 } 792 /* We didn't find it. */ 793 WARN_ON(1); 794 } 795 #endif 796 797 int efx_siena_probe_filters(struct efx_nic *efx) 798 { 799 int rc; 800 801 mutex_lock(&efx->mac_lock); 802 down_write(&efx->filter_sem); 803 rc = efx->type->filter_table_probe(efx); 804 if (rc) 805 goto out_unlock; 806 807 #ifdef CONFIG_RFS_ACCEL 808 if (efx->type->offload_features & NETIF_F_NTUPLE) { 809 struct efx_channel *channel; 810 int i, success = 1; 811 812 efx_for_each_channel(channel, efx) { 813 channel->rps_flow_id = 814 kcalloc(efx->type->max_rx_ip_filters, 815 sizeof(*channel->rps_flow_id), 816 GFP_KERNEL); 817 if (!channel->rps_flow_id) 818 success = 0; 819 else 820 for (i = 0; 821 i < efx->type->max_rx_ip_filters; 822 ++i) 823 channel->rps_flow_id[i] = 824 RPS_FLOW_ID_INVALID; 825 channel->rfs_expire_index = 0; 826 channel->rfs_filter_count = 0; 827 } 828 829 if (!success) { 830 efx_for_each_channel(channel, efx) 831 kfree(channel->rps_flow_id); 832 efx->type->filter_table_remove(efx); 833 rc = -ENOMEM; 834 goto out_unlock; 835 } 836 } 837 #endif 838 out_unlock: 839 up_write(&efx->filter_sem); 840 mutex_unlock(&efx->mac_lock); 841 return rc; 842 } 843 844 void efx_siena_remove_filters(struct efx_nic *efx) 845 { 846 #ifdef CONFIG_RFS_ACCEL 847 struct efx_channel *channel; 848 849 efx_for_each_channel(channel, efx) { 850 cancel_delayed_work_sync(&channel->filter_work); 851 kfree(channel->rps_flow_id); 852 channel->rps_flow_id = NULL; 853 } 854 #endif 855 down_write(&efx->filter_sem); 856 efx->type->filter_table_remove(efx); 857 up_write(&efx->filter_sem); 858 } 859 860 #ifdef CONFIG_RFS_ACCEL 861 862 static void efx_filter_rfs_work(struct work_struct *data) 863 { 864 struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion, 865 work); 866 struct efx_nic *efx = netdev_priv(req->net_dev); 867 struct efx_channel *channel = efx_get_channel(efx, req->rxq_index); 868 int slot_idx = req - efx->rps_slot; 869 struct efx_arfs_rule *rule; 870 u16 arfs_id = 0; 871 int rc; 872 873 rc = efx->type->filter_insert(efx, &req->spec, true); 874 if (rc >= 0) 875 /* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */ 876 rc %= efx->type->max_rx_ip_filters; 877 if (efx->rps_hash_table) { 878 spin_lock_bh(&efx->rps_hash_lock); 879 rule = efx_siena_rps_hash_find(efx, &req->spec); 880 /* The rule might have already gone, if someone else's request 881 * for the same spec was already worked and then expired before 882 * we got around to our work. In that case we have nothing 883 * tying us to an arfs_id, meaning that as soon as the filter 884 * is considered for expiry it will be removed. 885 */ 886 if (rule) { 887 if (rc < 0) 888 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR; 889 else 890 rule->filter_id = rc; 891 arfs_id = rule->arfs_id; 892 } 893 spin_unlock_bh(&efx->rps_hash_lock); 894 } 895 if (rc >= 0) { 896 /* Remember this so we can check whether to expire the filter 897 * later. 898 */ 899 mutex_lock(&efx->rps_mutex); 900 if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID) 901 channel->rfs_filter_count++; 902 channel->rps_flow_id[rc] = req->flow_id; 903 mutex_unlock(&efx->rps_mutex); 904 905 if (req->spec.ether_type == htons(ETH_P_IP)) 906 netif_info(efx, rx_status, efx->net_dev, 907 "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n", 908 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 909 req->spec.rem_host, ntohs(req->spec.rem_port), 910 req->spec.loc_host, ntohs(req->spec.loc_port), 911 req->rxq_index, req->flow_id, rc, arfs_id); 912 else 913 netif_info(efx, rx_status, efx->net_dev, 914 "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n", 915 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 916 req->spec.rem_host, ntohs(req->spec.rem_port), 917 req->spec.loc_host, ntohs(req->spec.loc_port), 918 req->rxq_index, req->flow_id, rc, arfs_id); 919 channel->n_rfs_succeeded++; 920 } else { 921 if (req->spec.ether_type == htons(ETH_P_IP)) 922 netif_dbg(efx, rx_status, efx->net_dev, 923 "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n", 924 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 925 req->spec.rem_host, ntohs(req->spec.rem_port), 926 req->spec.loc_host, ntohs(req->spec.loc_port), 927 req->rxq_index, req->flow_id, rc, arfs_id); 928 else 929 netif_dbg(efx, rx_status, efx->net_dev, 930 "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n", 931 (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP", 932 req->spec.rem_host, ntohs(req->spec.rem_port), 933 req->spec.loc_host, ntohs(req->spec.loc_port), 934 req->rxq_index, req->flow_id, rc, arfs_id); 935 channel->n_rfs_failed++; 936 /* We're overloading the NIC's filter tables, so let's do a 937 * chunk of extra expiry work. 938 */ 939 __efx_siena_filter_rfs_expire(channel, 940 min(channel->rfs_filter_count, 941 100u)); 942 } 943 944 /* Release references */ 945 clear_bit(slot_idx, &efx->rps_slot_map); 946 dev_put(req->net_dev); 947 } 948 949 int efx_siena_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb, 950 u16 rxq_index, u32 flow_id) 951 { 952 struct efx_nic *efx = netdev_priv(net_dev); 953 struct efx_async_filter_insertion *req; 954 struct efx_arfs_rule *rule; 955 struct flow_keys fk; 956 int slot_idx; 957 bool new; 958 int rc; 959 960 /* find a free slot */ 961 for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++) 962 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map)) 963 break; 964 if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT) 965 return -EBUSY; 966 967 if (flow_id == RPS_FLOW_ID_INVALID) { 968 rc = -EINVAL; 969 goto out_clear; 970 } 971 972 if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) { 973 rc = -EPROTONOSUPPORT; 974 goto out_clear; 975 } 976 977 if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) { 978 rc = -EPROTONOSUPPORT; 979 goto out_clear; 980 } 981 if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) { 982 rc = -EPROTONOSUPPORT; 983 goto out_clear; 984 } 985 986 req = efx->rps_slot + slot_idx; 987 efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT, 988 efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0, 989 rxq_index); 990 req->spec.match_flags = 991 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO | 992 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT | 993 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT; 994 req->spec.ether_type = fk.basic.n_proto; 995 req->spec.ip_proto = fk.basic.ip_proto; 996 997 if (fk.basic.n_proto == htons(ETH_P_IP)) { 998 req->spec.rem_host[0] = fk.addrs.v4addrs.src; 999 req->spec.loc_host[0] = fk.addrs.v4addrs.dst; 1000 } else { 1001 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src, 1002 sizeof(struct in6_addr)); 1003 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst, 1004 sizeof(struct in6_addr)); 1005 } 1006 1007 req->spec.rem_port = fk.ports.src; 1008 req->spec.loc_port = fk.ports.dst; 1009 1010 if (efx->rps_hash_table) { 1011 /* Add it to ARFS hash table */ 1012 spin_lock(&efx->rps_hash_lock); 1013 rule = efx_rps_hash_add(efx, &req->spec, &new); 1014 if (!rule) { 1015 rc = -ENOMEM; 1016 goto out_unlock; 1017 } 1018 if (new) 1019 rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER; 1020 rc = rule->arfs_id; 1021 /* Skip if existing or pending filter already does the right thing */ 1022 if (!new && rule->rxq_index == rxq_index && 1023 rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING) 1024 goto out_unlock; 1025 rule->rxq_index = rxq_index; 1026 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING; 1027 spin_unlock(&efx->rps_hash_lock); 1028 } else { 1029 /* Without an ARFS hash table, we just use arfs_id 0 for all 1030 * filters. This means if multiple flows hash to the same 1031 * flow_id, all but the most recently touched will be eligible 1032 * for expiry. 1033 */ 1034 rc = 0; 1035 } 1036 1037 /* Queue the request */ 1038 dev_hold(req->net_dev = net_dev); 1039 INIT_WORK(&req->work, efx_filter_rfs_work); 1040 req->rxq_index = rxq_index; 1041 req->flow_id = flow_id; 1042 schedule_work(&req->work); 1043 return rc; 1044 out_unlock: 1045 spin_unlock(&efx->rps_hash_lock); 1046 out_clear: 1047 clear_bit(slot_idx, &efx->rps_slot_map); 1048 return rc; 1049 } 1050 1051 bool __efx_siena_filter_rfs_expire(struct efx_channel *channel, 1052 unsigned int quota) 1053 { 1054 bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index); 1055 struct efx_nic *efx = channel->efx; 1056 unsigned int index, size, start; 1057 u32 flow_id; 1058 1059 if (!mutex_trylock(&efx->rps_mutex)) 1060 return false; 1061 expire_one = efx->type->filter_rfs_expire_one; 1062 index = channel->rfs_expire_index; 1063 start = index; 1064 size = efx->type->max_rx_ip_filters; 1065 while (quota) { 1066 flow_id = channel->rps_flow_id[index]; 1067 1068 if (flow_id != RPS_FLOW_ID_INVALID) { 1069 quota--; 1070 if (expire_one(efx, flow_id, index)) { 1071 netif_info(efx, rx_status, efx->net_dev, 1072 "expired filter %d [channel %u flow %u]\n", 1073 index, channel->channel, flow_id); 1074 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID; 1075 channel->rfs_filter_count--; 1076 } 1077 } 1078 if (++index == size) 1079 index = 0; 1080 /* If we were called with a quota that exceeds the total number 1081 * of filters in the table (which shouldn't happen, but could 1082 * if two callers race), ensure that we don't loop forever - 1083 * stop when we've examined every row of the table. 1084 */ 1085 if (index == start) 1086 break; 1087 } 1088 1089 channel->rfs_expire_index = index; 1090 mutex_unlock(&efx->rps_mutex); 1091 return true; 1092 } 1093 1094 #endif /* CONFIG_RFS_ACCEL */ 1095