drivers/net/ethernet/sfc/rx_common.c

   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                 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_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_find_rss_context_entry(struct efx_nic *efx, u32 id)
 597 {
 598         struct list_head *head = &efx->rss_context.list;
 599         struct efx_rss_context *ctx;
 600
 601         WARN_ON(!mutex_is_locked(&efx->rss_lock));
 602
 603         list_for_each_entry(ctx, head, list)
 604                 if (ctx->user_id == id)
 605                         return ctx;
 606         return NULL;
 607 }
 608
 609 void efx_free_rss_context_entry(struct efx_rss_context *ctx)
 610 {
 611         list_del(&ctx->list);
 612         kfree(ctx);
 613 }
 614
 615 void efx_set_default_rx_indir_table(struct efx_nic *efx,
 616                                     struct efx_rss_context *ctx)
 617 {
 618         size_t i;
 619
 620         for (i = 0; i < ARRAY_SIZE(ctx->rx_indir_table); i++)
 621                 ctx->rx_indir_table[i] =
 622                         ethtool_rxfh_indir_default(i, efx->rss_spread);
 623 }
 624
 625 /**
 626  * efx_filter_is_mc_recipient - test whether spec is a multicast recipient
 627  * @spec: Specification to test
 628  *
 629  * Return: %true if the specification is a non-drop RX filter that
 630  * matches a local MAC address I/G bit value of 1 or matches a local
 631  * IPv4 or IPv6 address value in the respective multicast address
 632  * range.  Otherwise %false.
 633  */
 634 bool efx_filter_is_mc_recipient(const struct efx_filter_spec *spec)
 635 {
 636         if (!(spec->flags & EFX_FILTER_FLAG_RX) ||
 637             spec->dmaq_id == EFX_FILTER_RX_DMAQ_ID_DROP)
 638                 return false;
 639
 640         if (spec->match_flags &
 641             (EFX_FILTER_MATCH_LOC_MAC | EFX_FILTER_MATCH_LOC_MAC_IG) &&
 642             is_multicast_ether_addr(spec->loc_mac))
 643                 return true;
 644
 645         if ((spec->match_flags &
 646              (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) ==
 647             (EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_LOC_HOST)) {
 648                 if (spec->ether_type == htons(ETH_P_IP) &&
 649                     ipv4_is_multicast(spec->loc_host[0]))
 650                         return true;
 651                 if (spec->ether_type == htons(ETH_P_IPV6) &&
 652                     ((const u8 *)spec->loc_host)[0] == 0xff)
 653                         return true;
 654         }
 655
 656         return false;
 657 }
 658
 659 bool efx_filter_spec_equal(const struct efx_filter_spec *left,
 660                            const struct efx_filter_spec *right)
 661 {
 662         if ((left->match_flags ^ right->match_flags) |
 663             ((left->flags ^ right->flags) &
 664              (EFX_FILTER_FLAG_RX | EFX_FILTER_FLAG_TX)))
 665                 return false;
 666
 667         return memcmp(&left->outer_vid, &right->outer_vid,
 668                       sizeof(struct efx_filter_spec) -
 669                       offsetof(struct efx_filter_spec, outer_vid)) == 0;
 670 }
 671
 672 u32 efx_filter_spec_hash(const struct efx_filter_spec *spec)
 673 {
 674         BUILD_BUG_ON(offsetof(struct efx_filter_spec, outer_vid) & 3);
 675         return jhash2((const u32 *)&spec->outer_vid,
 676                       (sizeof(struct efx_filter_spec) -
 677                        offsetof(struct efx_filter_spec, outer_vid)) / 4,
 678                       0);
 679 }
 680
 681 #ifdef CONFIG_RFS_ACCEL
 682 bool efx_rps_check_rule(struct efx_arfs_rule *rule, unsigned int filter_idx,
 683                         bool *force)
 684 {
 685         if (rule->filter_id == EFX_ARFS_FILTER_ID_PENDING) {
 686                 /* ARFS is currently updating this entry, leave it */
 687                 return false;
 688         }
 689         if (rule->filter_id == EFX_ARFS_FILTER_ID_ERROR) {
 690                 /* ARFS tried and failed to update this, so it's probably out
 691                  * of date.  Remove the filter and the ARFS rule entry.
 692                  */
 693                 rule->filter_id = EFX_ARFS_FILTER_ID_REMOVING;
 694                 *force = true;
 695                 return true;
 696         } else if (WARN_ON(rule->filter_id != filter_idx)) { /* can't happen */
 697                 /* ARFS has moved on, so old filter is not needed.  Since we did
 698                  * not mark the rule with EFX_ARFS_FILTER_ID_REMOVING, it will
 699                  * not be removed by efx_rps_hash_del() subsequently.
 700                  */
 701                 *force = true;
 702                 return true;
 703         }
 704         /* Remove it iff ARFS wants to. */
 705         return true;
 706 }
 707
 708 static
 709 struct hlist_head *efx_rps_hash_bucket(struct efx_nic *efx,
 710                                        const struct efx_filter_spec *spec)
 711 {
 712         u32 hash = efx_filter_spec_hash(spec);
 713
 714         lockdep_assert_held(&efx->rps_hash_lock);
 715         if (!efx->rps_hash_table)
 716                 return NULL;
 717         return &efx->rps_hash_table[hash % EFX_ARFS_HASH_TABLE_SIZE];
 718 }
 719
 720 struct efx_arfs_rule *efx_rps_hash_find(struct efx_nic *efx,
 721                                         const struct efx_filter_spec *spec)
 722 {
 723         struct efx_arfs_rule *rule;
 724         struct hlist_head *head;
 725         struct hlist_node *node;
 726
 727         head = efx_rps_hash_bucket(efx, spec);
 728         if (!head)
 729                 return NULL;
 730         hlist_for_each(node, head) {
 731                 rule = container_of(node, struct efx_arfs_rule, node);
 732                 if (efx_filter_spec_equal(spec, &rule->spec))
 733                         return rule;
 734         }
 735         return NULL;
 736 }
 737
 738 struct efx_arfs_rule *efx_rps_hash_add(struct efx_nic *efx,
 739                                        const struct efx_filter_spec *spec,
 740                                        bool *new)
 741 {
 742         struct efx_arfs_rule *rule;
 743         struct hlist_head *head;
 744         struct hlist_node *node;
 745
 746         head = efx_rps_hash_bucket(efx, spec);
 747         if (!head)
 748                 return NULL;
 749         hlist_for_each(node, head) {
 750                 rule = container_of(node, struct efx_arfs_rule, node);
 751                 if (efx_filter_spec_equal(spec, &rule->spec)) {
 752                         *new = false;
 753                         return rule;
 754                 }
 755         }
 756         rule = kmalloc(sizeof(*rule), GFP_ATOMIC);
 757         *new = true;
 758         if (rule) {
 759                 memcpy(&rule->spec, spec, sizeof(rule->spec));
 760                 hlist_add_head(&rule->node, head);
 761         }
 762         return rule;
 763 }
 764
 765 void efx_rps_hash_del(struct efx_nic *efx, const struct efx_filter_spec *spec)
 766 {
 767         struct efx_arfs_rule *rule;
 768         struct hlist_head *head;
 769         struct hlist_node *node;
 770
 771         head = efx_rps_hash_bucket(efx, spec);
 772         if (WARN_ON(!head))
 773                 return;
 774         hlist_for_each(node, head) {
 775                 rule = container_of(node, struct efx_arfs_rule, node);
 776                 if (efx_filter_spec_equal(spec, &rule->spec)) {
 777                         /* Someone already reused the entry.  We know that if
 778                          * this check doesn't fire (i.e. filter_id == REMOVING)
 779                          * then the REMOVING mark was put there by our caller,
 780                          * because caller is holding a lock on filter table and
 781                          * only holders of that lock set REMOVING.
 782                          */
 783                         if (rule->filter_id != EFX_ARFS_FILTER_ID_REMOVING)
 784                                 return;
 785                         hlist_del(node);
 786                         kfree(rule);
 787                         return;
 788                 }
 789         }
 790         /* We didn't find it. */
 791         WARN_ON(1);
 792 }
 793 #endif
 794
 795 int efx_probe_filters(struct efx_nic *efx)
 796 {
 797         int rc;
 798
 799         init_rwsem(&efx->filter_sem);
 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         }
 852 #endif
 853         down_write(&efx->filter_sem);
 854         efx->type->filter_table_remove(efx);
 855         up_write(&efx->filter_sem);
 856 }
 857
 858 #ifdef CONFIG_RFS_ACCEL
 859
 860 static void efx_filter_rfs_work(struct work_struct *data)
 861 {
 862         struct efx_async_filter_insertion *req = container_of(data, struct efx_async_filter_insertion,
 863                                                               work);
 864         struct efx_nic *efx = netdev_priv(req->net_dev);
 865         struct efx_channel *channel = efx_get_channel(efx, req->rxq_index);
 866         int slot_idx = req - efx->rps_slot;
 867         struct efx_arfs_rule *rule;
 868         u16 arfs_id = 0;
 869         int rc;
 870
 871         rc = efx->type->filter_insert(efx, &req->spec, true);
 872         if (rc >= 0)
 873                 /* Discard 'priority' part of EF10+ filter ID (mcdi_filters) */
 874                 rc %= efx->type->max_rx_ip_filters;
 875         if (efx->rps_hash_table) {
 876                 spin_lock_bh(&efx->rps_hash_lock);
 877                 rule = efx_rps_hash_find(efx, &req->spec);
 878                 /* The rule might have already gone, if someone else's request
 879                  * for the same spec was already worked and then expired before
 880                  * we got around to our work.  In that case we have nothing
 881                  * tying us to an arfs_id, meaning that as soon as the filter
 882                  * is considered for expiry it will be removed.
 883                  */
 884                 if (rule) {
 885                         if (rc < 0)
 886                                 rule->filter_id = EFX_ARFS_FILTER_ID_ERROR;
 887                         else
 888                                 rule->filter_id = rc;
 889                         arfs_id = rule->arfs_id;
 890                 }
 891                 spin_unlock_bh(&efx->rps_hash_lock);
 892         }
 893         if (rc >= 0) {
 894                 /* Remember this so we can check whether to expire the filter
 895                  * later.
 896                  */
 897                 mutex_lock(&efx->rps_mutex);
 898                 if (channel->rps_flow_id[rc] == RPS_FLOW_ID_INVALID)
 899                         channel->rfs_filter_count++;
 900                 channel->rps_flow_id[rc] = req->flow_id;
 901                 mutex_unlock(&efx->rps_mutex);
 902
 903                 if (req->spec.ether_type == htons(ETH_P_IP))
 904                         netif_info(efx, rx_status, efx->net_dev,
 905                                    "steering %s %pI4:%u:%pI4:%u to queue %u [flow %u filter %d id %u]\n",
 906                                    (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 907                                    req->spec.rem_host, ntohs(req->spec.rem_port),
 908                                    req->spec.loc_host, ntohs(req->spec.loc_port),
 909                                    req->rxq_index, req->flow_id, rc, arfs_id);
 910                 else
 911                         netif_info(efx, rx_status, efx->net_dev,
 912                                    "steering %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u filter %d id %u]\n",
 913                                    (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 914                                    req->spec.rem_host, ntohs(req->spec.rem_port),
 915                                    req->spec.loc_host, ntohs(req->spec.loc_port),
 916                                    req->rxq_index, req->flow_id, rc, arfs_id);
 917                 channel->n_rfs_succeeded++;
 918         } else {
 919                 if (req->spec.ether_type == htons(ETH_P_IP))
 920                         netif_dbg(efx, rx_status, efx->net_dev,
 921                                   "failed to steer %s %pI4:%u:%pI4:%u to queue %u [flow %u rc %d id %u]\n",
 922                                   (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 923                                   req->spec.rem_host, ntohs(req->spec.rem_port),
 924                                   req->spec.loc_host, ntohs(req->spec.loc_port),
 925                                   req->rxq_index, req->flow_id, rc, arfs_id);
 926                 else
 927                         netif_dbg(efx, rx_status, efx->net_dev,
 928                                   "failed to steer %s [%pI6]:%u:[%pI6]:%u to queue %u [flow %u rc %d id %u]\n",
 929                                   (req->spec.ip_proto == IPPROTO_TCP) ? "TCP" : "UDP",
 930                                   req->spec.rem_host, ntohs(req->spec.rem_port),
 931                                   req->spec.loc_host, ntohs(req->spec.loc_port),
 932                                   req->rxq_index, req->flow_id, rc, arfs_id);
 933                 channel->n_rfs_failed++;
 934                 /* We're overloading the NIC's filter tables, so let's do a
 935                  * chunk of extra expiry work.
 936                  */
 937                 __efx_filter_rfs_expire(channel, min(channel->rfs_filter_count,
 938                                                      100u));
 939         }
 940
 941         /* Release references */
 942         clear_bit(slot_idx, &efx->rps_slot_map);
 943         dev_put(req->net_dev);
 944 }
 945
 946 int efx_filter_rfs(struct net_device *net_dev, const struct sk_buff *skb,
 947                    u16 rxq_index, u32 flow_id)
 948 {
 949         struct efx_nic *efx = netdev_priv(net_dev);
 950         struct efx_async_filter_insertion *req;
 951         struct efx_arfs_rule *rule;
 952         struct flow_keys fk;
 953         int slot_idx;
 954         bool new;
 955         int rc;
 956
 957         /* find a free slot */
 958         for (slot_idx = 0; slot_idx < EFX_RPS_MAX_IN_FLIGHT; slot_idx++)
 959                 if (!test_and_set_bit(slot_idx, &efx->rps_slot_map))
 960                         break;
 961         if (slot_idx >= EFX_RPS_MAX_IN_FLIGHT)
 962                 return -EBUSY;
 963
 964         if (flow_id == RPS_FLOW_ID_INVALID) {
 965                 rc = -EINVAL;
 966                 goto out_clear;
 967         }
 968
 969         if (!skb_flow_dissect_flow_keys(skb, &fk, 0)) {
 970                 rc = -EPROTONOSUPPORT;
 971                 goto out_clear;
 972         }
 973
 974         if (fk.basic.n_proto != htons(ETH_P_IP) && fk.basic.n_proto != htons(ETH_P_IPV6)) {
 975                 rc = -EPROTONOSUPPORT;
 976                 goto out_clear;
 977         }
 978         if (fk.control.flags & FLOW_DIS_IS_FRAGMENT) {
 979                 rc = -EPROTONOSUPPORT;
 980                 goto out_clear;
 981         }
 982
 983         req = efx->rps_slot + slot_idx;
 984         efx_filter_init_rx(&req->spec, EFX_FILTER_PRI_HINT,
 985                            efx->rx_scatter ? EFX_FILTER_FLAG_RX_SCATTER : 0,
 986                            rxq_index);
 987         req->spec.match_flags =
 988                 EFX_FILTER_MATCH_ETHER_TYPE | EFX_FILTER_MATCH_IP_PROTO |
 989                 EFX_FILTER_MATCH_LOC_HOST | EFX_FILTER_MATCH_LOC_PORT |
 990                 EFX_FILTER_MATCH_REM_HOST | EFX_FILTER_MATCH_REM_PORT;
 991         req->spec.ether_type = fk.basic.n_proto;
 992         req->spec.ip_proto = fk.basic.ip_proto;
 993
 994         if (fk.basic.n_proto == htons(ETH_P_IP)) {
 995                 req->spec.rem_host[0] = fk.addrs.v4addrs.src;
 996                 req->spec.loc_host[0] = fk.addrs.v4addrs.dst;
 997         } else {
 998                 memcpy(req->spec.rem_host, &fk.addrs.v6addrs.src,
 999                        sizeof(struct in6_addr));
1000                 memcpy(req->spec.loc_host, &fk.addrs.v6addrs.dst,
1001                        sizeof(struct in6_addr));
1002         }
1003
1004         req->spec.rem_port = fk.ports.src;
1005         req->spec.loc_port = fk.ports.dst;
1006
1007         if (efx->rps_hash_table) {
1008                 /* Add it to ARFS hash table */
1009                 spin_lock(&efx->rps_hash_lock);
1010                 rule = efx_rps_hash_add(efx, &req->spec, &new);
1011                 if (!rule) {
1012                         rc = -ENOMEM;
1013                         goto out_unlock;
1014                 }
1015                 if (new)
1016                         rule->arfs_id = efx->rps_next_id++ % RPS_NO_FILTER;
1017                 rc = rule->arfs_id;
1018                 /* Skip if existing or pending filter already does the right thing */
1019                 if (!new && rule->rxq_index == rxq_index &&
1020                     rule->filter_id >= EFX_ARFS_FILTER_ID_PENDING)
1021                         goto out_unlock;
1022                 rule->rxq_index = rxq_index;
1023                 rule->filter_id = EFX_ARFS_FILTER_ID_PENDING;
1024                 spin_unlock(&efx->rps_hash_lock);
1025         } else {
1026                 /* Without an ARFS hash table, we just use arfs_id 0 for all
1027                  * filters.  This means if multiple flows hash to the same
1028                  * flow_id, all but the most recently touched will be eligible
1029                  * for expiry.
1030                  */
1031                 rc = 0;
1032         }
1033
1034         /* Queue the request */
1035         dev_hold(req->net_dev = net_dev);
1036         INIT_WORK(&req->work, efx_filter_rfs_work);
1037         req->rxq_index = rxq_index;
1038         req->flow_id = flow_id;
1039         schedule_work(&req->work);
1040         return rc;
1041 out_unlock:
1042         spin_unlock(&efx->rps_hash_lock);
1043 out_clear:
1044         clear_bit(slot_idx, &efx->rps_slot_map);
1045         return rc;
1046 }
1047
1048 bool __efx_filter_rfs_expire(struct efx_channel *channel, unsigned int quota)
1049 {
1050         bool (*expire_one)(struct efx_nic *efx, u32 flow_id, unsigned int index);
1051         struct efx_nic *efx = channel->efx;
1052         unsigned int index, size, start;
1053         u32 flow_id;
1054
1055         if (!mutex_trylock(&efx->rps_mutex))
1056                 return false;
1057         expire_one = efx->type->filter_rfs_expire_one;
1058         index = channel->rfs_expire_index;
1059         start = index;
1060         size = efx->type->max_rx_ip_filters;
1061         while (quota) {
1062                 flow_id = channel->rps_flow_id[index];
1063
1064                 if (flow_id != RPS_FLOW_ID_INVALID) {
1065                         quota--;
1066                         if (expire_one(efx, flow_id, index)) {
1067                                 netif_info(efx, rx_status, efx->net_dev,
1068                                            "expired filter %d [channel %u flow %u]\n",
1069                                            index, channel->channel, flow_id);
1070                                 channel->rps_flow_id[index] = RPS_FLOW_ID_INVALID;
1071                                 channel->rfs_filter_count--;
1072                         }
1073                 }
1074                 if (++index == size)
1075                         index = 0;
1076                 /* If we were called with a quota that exceeds the total number
1077                  * of filters in the table (which shouldn't happen, but could
1078                  * if two callers race), ensure that we don't loop forever -
1079                  * stop when we've examined every row of the table.
1080                  */
1081                 if (index == start)
1082                         break;
1083         }
1084
1085         channel->rfs_expire_index = index;
1086         mutex_unlock(&efx->rps_mutex);
1087         return true;
1088 }
1089
1090 #endif /* CONFIG_RFS_ACCEL */