net/core/page_pool.c

   1 /* SPDX-License-Identifier: GPL-2.0
   2  *
   3  * page_pool.c
   4  *      Author: Jesper Dangaard Brouer <[email protected]>
   5  *      Copyright (C) 2016 Red Hat, Inc.
   6  */
   7
   8 #include <linux/types.h>
   9 #include <linux/kernel.h>
  10 #include <linux/slab.h>
  11 #include <linux/device.h>
  12
  13 #include <net/page_pool.h>
  14 #include <net/xdp.h>
  15
  16 #include <linux/dma-direction.h>
  17 #include <linux/dma-mapping.h>
  18 #include <linux/page-flags.h>
  19 #include <linux/mm.h> /* for __put_page() */
  20 #include <linux/poison.h>
  21
  22 #include <trace/events/page_pool.h>
  23
  24 #define DEFER_TIME (msecs_to_jiffies(1000))
  25 #define DEFER_WARN_INTERVAL (60 * HZ)
  26
  27 #define BIAS_MAX        LONG_MAX
  28
  29 static int page_pool_init(struct page_pool *pool,
  30                           const struct page_pool_params *params)
  31 {
  32         unsigned int ring_qsize = 1024; /* Default */
  33
  34         memcpy(&pool->p, params, sizeof(pool->p));
  35
  36         /* Validate only known flags were used */
  37         if (pool->p.flags & ~(PP_FLAG_ALL))
  38                 return -EINVAL;
  39
  40         if (pool->p.pool_size)
  41                 ring_qsize = pool->p.pool_size;
  42
  43         /* Sanity limit mem that can be pinned down */
  44         if (ring_qsize > 32768)
  45                 return -E2BIG;
  46
  47         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
  48          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
  49          * which is the XDP_TX use-case.
  50          */
  51         if (pool->p.flags & PP_FLAG_DMA_MAP) {
  52                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
  53                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
  54                         return -EINVAL;
  55         }
  56
  57         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
  58                 /* In order to request DMA-sync-for-device the page
  59                  * needs to be mapped
  60                  */
  61                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
  62                         return -EINVAL;
  63
  64                 if (!pool->p.max_len)
  65                         return -EINVAL;
  66
  67                 /* pool->p.offset has to be set according to the address
  68                  * offset used by the DMA engine to start copying rx data
  69                  */
  70         }
  71
  72         if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT &&
  73             pool->p.flags & PP_FLAG_PAGE_FRAG)
  74                 return -EINVAL;
  75
  76         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
  77                 return -ENOMEM;
  78
  79         atomic_set(&pool->pages_state_release_cnt, 0);
  80
  81         /* Driver calling page_pool_create() also call page_pool_destroy() */
  82         refcount_set(&pool->user_cnt, 1);
  83
  84         if (pool->p.flags & PP_FLAG_DMA_MAP)
  85                 get_device(pool->p.dev);
  86
  87         return 0;
  88 }
  89
  90 struct page_pool *page_pool_create(const struct page_pool_params *params)
  91 {
  92         struct page_pool *pool;
  93         int err;
  94
  95         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
  96         if (!pool)
  97                 return ERR_PTR(-ENOMEM);
  98
  99         err = page_pool_init(pool, params);
 100         if (err < 0) {
 101                 pr_warn("%s() gave up with errno %d\n", __func__, err);
 102                 kfree(pool);
 103                 return ERR_PTR(err);
 104         }
 105
 106         return pool;
 107 }
 108 EXPORT_SYMBOL(page_pool_create);
 109
 110 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 111
 112 noinline
 113 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 114 {
 115         struct ptr_ring *r = &pool->ring;
 116         struct page *page;
 117         int pref_nid; /* preferred NUMA node */
 118
 119         /* Quicker fallback, avoid locks when ring is empty */
 120         if (__ptr_ring_empty(r))
 121                 return NULL;
 122
 123         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 124          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 125          */
 126 #ifdef CONFIG_NUMA
 127         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 128 #else
 129         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 130         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 131 #endif
 132
 133         /* Refill alloc array, but only if NUMA match */
 134         do {
 135                 page = __ptr_ring_consume(r);
 136                 if (unlikely(!page))
 137                         break;
 138
 139                 if (likely(page_to_nid(page) == pref_nid)) {
 140                         pool->alloc.cache[pool->alloc.count++] = page;
 141                 } else {
 142                         /* NUMA mismatch;
 143                          * (1) release 1 page to page-allocator and
 144                          * (2) break out to fallthrough to alloc_pages_node.
 145                          * This limit stress on page buddy alloactor.
 146                          */
 147                         page_pool_return_page(pool, page);
 148                         page = NULL;
 149                         break;
 150                 }
 151         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 152
 153         /* Return last page */
 154         if (likely(pool->alloc.count > 0))
 155                 page = pool->alloc.cache[--pool->alloc.count];
 156
 157         return page;
 158 }
 159
 160 /* fast path */
 161 static struct page *__page_pool_get_cached(struct page_pool *pool)
 162 {
 163         struct page *page;
 164
 165         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 166         if (likely(pool->alloc.count)) {
 167                 /* Fast-path */
 168                 page = pool->alloc.cache[--pool->alloc.count];
 169         } else {
 170                 page = page_pool_refill_alloc_cache(pool);
 171         }
 172
 173         return page;
 174 }
 175
 176 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 177                                           struct page *page,
 178                                           unsigned int dma_sync_size)
 179 {
 180         dma_addr_t dma_addr = page_pool_get_dma_addr(page);
 181
 182         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 183         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 184                                          pool->p.offset, dma_sync_size,
 185                                          pool->p.dma_dir);
 186 }
 187
 188 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 189 {
 190         dma_addr_t dma;
 191
 192         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 193          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 194          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 195          * This mapping is kept for lifetime of page, until leaving pool.
 196          */
 197         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 198                                  (PAGE_SIZE << pool->p.order),
 199                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 200         if (dma_mapping_error(pool->p.dev, dma))
 201                 return false;
 202
 203         page_pool_set_dma_addr(page, dma);
 204
 205         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 206                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 207
 208         return true;
 209 }
 210
 211 static void page_pool_set_pp_info(struct page_pool *pool,
 212                                   struct page *page)
 213 {
 214         page->pp = pool;
 215         page->pp_magic |= PP_SIGNATURE;
 216         if (pool->p.init_callback)
 217                 pool->p.init_callback(page, pool->p.init_arg);
 218 }
 219
 220 static void page_pool_clear_pp_info(struct page *page)
 221 {
 222         page->pp_magic = 0;
 223         page->pp = NULL;
 224 }
 225
 226 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 227                                                  gfp_t gfp)
 228 {
 229         struct page *page;
 230
 231         gfp |= __GFP_COMP;
 232         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 233         if (unlikely(!page))
 234                 return NULL;
 235
 236         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 237             unlikely(!page_pool_dma_map(pool, page))) {
 238                 put_page(page);
 239                 return NULL;
 240         }
 241
 242         page_pool_set_pp_info(pool, page);
 243
 244         /* Track how many pages are held 'in-flight' */
 245         pool->pages_state_hold_cnt++;
 246         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 247         return page;
 248 }
 249
 250 /* slow path */
 251 noinline
 252 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 253                                                  gfp_t gfp)
 254 {
 255         const int bulk = PP_ALLOC_CACHE_REFILL;
 256         unsigned int pp_flags = pool->p.flags;
 257         unsigned int pp_order = pool->p.order;
 258         struct page *page;
 259         int i, nr_pages;
 260
 261         /* Don't support bulk alloc for high-order pages */
 262         if (unlikely(pp_order))
 263                 return __page_pool_alloc_page_order(pool, gfp);
 264
 265         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 266         if (unlikely(pool->alloc.count > 0))
 267                 return pool->alloc.cache[--pool->alloc.count];
 268
 269         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 270         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 271
 272         nr_pages = alloc_pages_bulk_array(gfp, bulk, pool->alloc.cache);
 273         if (unlikely(!nr_pages))
 274                 return NULL;
 275
 276         /* Pages have been filled into alloc.cache array, but count is zero and
 277          * page element have not been (possibly) DMA mapped.
 278          */
 279         for (i = 0; i < nr_pages; i++) {
 280                 page = pool->alloc.cache[i];
 281                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 282                     unlikely(!page_pool_dma_map(pool, page))) {
 283                         put_page(page);
 284                         continue;
 285                 }
 286
 287                 page_pool_set_pp_info(pool, page);
 288                 pool->alloc.cache[pool->alloc.count++] = page;
 289                 /* Track how many pages are held 'in-flight' */
 290                 pool->pages_state_hold_cnt++;
 291                 trace_page_pool_state_hold(pool, page,
 292                                            pool->pages_state_hold_cnt);
 293         }
 294
 295         /* Return last page */
 296         if (likely(pool->alloc.count > 0))
 297                 page = pool->alloc.cache[--pool->alloc.count];
 298         else
 299                 page = NULL;
 300
 301         /* When page just alloc'ed is should/must have refcnt 1. */
 302         return page;
 303 }
 304
 305 /* For using page_pool replace: alloc_pages() API calls, but provide
 306  * synchronization guarantee for allocation side.
 307  */
 308 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 309 {
 310         struct page *page;
 311
 312         /* Fast-path: Get a page from cache */
 313         page = __page_pool_get_cached(pool);
 314         if (page)
 315                 return page;
 316
 317         /* Slow-path: cache empty, do real allocation */
 318         page = __page_pool_alloc_pages_slow(pool, gfp);
 319         return page;
 320 }
 321 EXPORT_SYMBOL(page_pool_alloc_pages);
 322
 323 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 324  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 325  */
 326 #define _distance(a, b) (s32)((a) - (b))
 327
 328 static s32 page_pool_inflight(struct page_pool *pool)
 329 {
 330         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 331         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 332         s32 inflight;
 333
 334         inflight = _distance(hold_cnt, release_cnt);
 335
 336         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 337         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 338
 339         return inflight;
 340 }
 341
 342 /* Disconnects a page (from a page_pool).  API users can have a need
 343  * to disconnect a page (from a page_pool), to allow it to be used as
 344  * a regular page (that will eventually be returned to the normal
 345  * page-allocator via put_page).
 346  */
 347 void page_pool_release_page(struct page_pool *pool, struct page *page)
 348 {
 349         dma_addr_t dma;
 350         int count;
 351
 352         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 353                 /* Always account for inflight pages, even if we didn't
 354                  * map them
 355                  */
 356                 goto skip_dma_unmap;
 357
 358         dma = page_pool_get_dma_addr(page);
 359
 360         /* When page is unmapped, it cannot be returned to our pool */
 361         dma_unmap_page_attrs(pool->p.dev, dma,
 362                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 363                              DMA_ATTR_SKIP_CPU_SYNC);
 364         page_pool_set_dma_addr(page, 0);
 365 skip_dma_unmap:
 366         page_pool_clear_pp_info(page);
 367
 368         /* This may be the last page returned, releasing the pool, so
 369          * it is not safe to reference pool afterwards.
 370          */
 371         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 372         trace_page_pool_state_release(pool, page, count);
 373 }
 374 EXPORT_SYMBOL(page_pool_release_page);
 375
 376 /* Return a page to the page allocator, cleaning up our state */
 377 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 378 {
 379         page_pool_release_page(pool, page);
 380
 381         put_page(page);
 382         /* An optimization would be to call __free_pages(page, pool->p.order)
 383          * knowing page is not part of page-cache (thus avoiding a
 384          * __page_cache_release() call).
 385          */
 386 }
 387
 388 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 389 {
 390         int ret;
 391         /* BH protection not needed if current is serving softirq */
 392         if (in_serving_softirq())
 393                 ret = ptr_ring_produce(&pool->ring, page);
 394         else
 395                 ret = ptr_ring_produce_bh(&pool->ring, page);
 396
 397         return (ret == 0) ? true : false;
 398 }
 399
 400 /* Only allow direct recycling in special circumstances, into the
 401  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 402  *
 403  * Caller must provide appropriate safe context.
 404  */
 405 static bool page_pool_recycle_in_cache(struct page *page,
 406                                        struct page_pool *pool)
 407 {
 408         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
 409                 return false;
 410
 411         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 412         pool->alloc.cache[pool->alloc.count++] = page;
 413         return true;
 414 }
 415
 416 /* If the page refcnt == 1, this will try to recycle the page.
 417  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 418  * the configured size min(dma_sync_size, pool->max_len).
 419  * If the page refcnt != 1, then the page will be returned to memory
 420  * subsystem.
 421  */
 422 static __always_inline struct page *
 423 __page_pool_put_page(struct page_pool *pool, struct page *page,
 424                      unsigned int dma_sync_size, bool allow_direct)
 425 {
 426         /* It is not the last user for the page frag case */
 427         if (pool->p.flags & PP_FLAG_PAGE_FRAG &&
 428             page_pool_atomic_sub_frag_count_return(page, 1))
 429                 return NULL;
 430
 431         /* This allocator is optimized for the XDP mode that uses
 432          * one-frame-per-page, but have fallbacks that act like the
 433          * regular page allocator APIs.
 434          *
 435          * refcnt == 1 means page_pool owns page, and can recycle it.
 436          *
 437          * page is NOT reusable when allocated when system is under
 438          * some pressure. (page_is_pfmemalloc)
 439          */
 440         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 441                 /* Read barrier done in page_ref_count / READ_ONCE */
 442
 443                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 444                         page_pool_dma_sync_for_device(pool, page,
 445                                                       dma_sync_size);
 446
 447                 if (allow_direct && in_serving_softirq() &&
 448                     page_pool_recycle_in_cache(page, pool))
 449                         return NULL;
 450
 451                 /* Page found as candidate for recycling */
 452                 return page;
 453         }
 454         /* Fallback/non-XDP mode: API user have elevated refcnt.
 455          *
 456          * Many drivers split up the page into fragments, and some
 457          * want to keep doing this to save memory and do refcnt based
 458          * recycling. Support this use case too, to ease drivers
 459          * switching between XDP/non-XDP.
 460          *
 461          * In-case page_pool maintains the DMA mapping, API user must
 462          * call page_pool_put_page once.  In this elevated refcnt
 463          * case, the DMA is unmapped/released, as driver is likely
 464          * doing refcnt based recycle tricks, meaning another process
 465          * will be invoking put_page.
 466          */
 467         /* Do not replace this with page_pool_return_page() */
 468         page_pool_release_page(pool, page);
 469         put_page(page);
 470
 471         return NULL;
 472 }
 473
 474 void page_pool_put_page(struct page_pool *pool, struct page *page,
 475                         unsigned int dma_sync_size, bool allow_direct)
 476 {
 477         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 478         if (page && !page_pool_recycle_in_ring(pool, page)) {
 479                 /* Cache full, fallback to free pages */
 480                 page_pool_return_page(pool, page);
 481         }
 482 }
 483 EXPORT_SYMBOL(page_pool_put_page);
 484
 485 /* Caller must not use data area after call, as this function overwrites it */
 486 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 487                              int count)
 488 {
 489         int i, bulk_len = 0;
 490
 491         for (i = 0; i < count; i++) {
 492                 struct page *page = virt_to_head_page(data[i]);
 493
 494                 page = __page_pool_put_page(pool, page, -1, false);
 495                 /* Approved for bulk recycling in ptr_ring cache */
 496                 if (page)
 497                         data[bulk_len++] = page;
 498         }
 499
 500         if (unlikely(!bulk_len))
 501                 return;
 502
 503         /* Bulk producer into ptr_ring page_pool cache */
 504         page_pool_ring_lock(pool);
 505         for (i = 0; i < bulk_len; i++) {
 506                 if (__ptr_ring_produce(&pool->ring, data[i]))
 507                         break; /* ring full */
 508         }
 509         page_pool_ring_unlock(pool);
 510
 511         /* Hopefully all pages was return into ptr_ring */
 512         if (likely(i == bulk_len))
 513                 return;
 514
 515         /* ptr_ring cache full, free remaining pages outside producer lock
 516          * since put_page() with refcnt == 1 can be an expensive operation
 517          */
 518         for (; i < bulk_len; i++)
 519                 page_pool_return_page(pool, data[i]);
 520 }
 521 EXPORT_SYMBOL(page_pool_put_page_bulk);
 522
 523 static struct page *page_pool_drain_frag(struct page_pool *pool,
 524                                          struct page *page)
 525 {
 526         long drain_count = BIAS_MAX - pool->frag_users;
 527
 528         /* Some user is still using the page frag */
 529         if (likely(page_pool_atomic_sub_frag_count_return(page,
 530                                                           drain_count)))
 531                 return NULL;
 532
 533         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 534                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 535                         page_pool_dma_sync_for_device(pool, page, -1);
 536
 537                 return page;
 538         }
 539
 540         page_pool_return_page(pool, page);
 541         return NULL;
 542 }
 543
 544 static void page_pool_free_frag(struct page_pool *pool)
 545 {
 546         long drain_count = BIAS_MAX - pool->frag_users;
 547         struct page *page = pool->frag_page;
 548
 549         pool->frag_page = NULL;
 550
 551         if (!page ||
 552             page_pool_atomic_sub_frag_count_return(page, drain_count))
 553                 return;
 554
 555         page_pool_return_page(pool, page);
 556 }
 557
 558 struct page *page_pool_alloc_frag(struct page_pool *pool,
 559                                   unsigned int *offset,
 560                                   unsigned int size, gfp_t gfp)
 561 {
 562         unsigned int max_size = PAGE_SIZE << pool->p.order;
 563         struct page *page = pool->frag_page;
 564
 565         if (WARN_ON(!(pool->p.flags & PP_FLAG_PAGE_FRAG) ||
 566                     size > max_size))
 567                 return NULL;
 568
 569         size = ALIGN(size, dma_get_cache_alignment());
 570         *offset = pool->frag_offset;
 571
 572         if (page && *offset + size > max_size) {
 573                 page = page_pool_drain_frag(pool, page);
 574                 if (page)
 575                         goto frag_reset;
 576         }
 577
 578         if (!page) {
 579                 page = page_pool_alloc_pages(pool, gfp);
 580                 if (unlikely(!page)) {
 581                         pool->frag_page = NULL;
 582                         return NULL;
 583                 }
 584
 585                 pool->frag_page = page;
 586
 587 frag_reset:
 588                 pool->frag_users = 1;
 589                 *offset = 0;
 590                 pool->frag_offset = size;
 591                 page_pool_set_frag_count(page, BIAS_MAX);
 592                 return page;
 593         }
 594
 595         pool->frag_users++;
 596         pool->frag_offset = *offset + size;
 597         return page;
 598 }
 599 EXPORT_SYMBOL(page_pool_alloc_frag);
 600
 601 static void page_pool_empty_ring(struct page_pool *pool)
 602 {
 603         struct page *page;
 604
 605         /* Empty recycle ring */
 606         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 607                 /* Verify the refcnt invariant of cached pages */
 608                 if (!(page_ref_count(page) == 1))
 609                         pr_crit("%s() page_pool refcnt %d violation\n",
 610                                 __func__, page_ref_count(page));
 611
 612                 page_pool_return_page(pool, page);
 613         }
 614 }
 615
 616 static void page_pool_free(struct page_pool *pool)
 617 {
 618         if (pool->disconnect)
 619                 pool->disconnect(pool);
 620
 621         ptr_ring_cleanup(&pool->ring, NULL);
 622
 623         if (pool->p.flags & PP_FLAG_DMA_MAP)
 624                 put_device(pool->p.dev);
 625
 626         kfree(pool);
 627 }
 628
 629 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 630 {
 631         struct page *page;
 632
 633         if (pool->destroy_cnt)
 634                 return;
 635
 636         /* Empty alloc cache, assume caller made sure this is
 637          * no-longer in use, and page_pool_alloc_pages() cannot be
 638          * call concurrently.
 639          */
 640         while (pool->alloc.count) {
 641                 page = pool->alloc.cache[--pool->alloc.count];
 642                 page_pool_return_page(pool, page);
 643         }
 644 }
 645
 646 static void page_pool_scrub(struct page_pool *pool)
 647 {
 648         page_pool_empty_alloc_cache_once(pool);
 649         pool->destroy_cnt++;
 650
 651         /* No more consumers should exist, but producers could still
 652          * be in-flight.
 653          */
 654         page_pool_empty_ring(pool);
 655 }
 656
 657 static int page_pool_release(struct page_pool *pool)
 658 {
 659         int inflight;
 660
 661         page_pool_scrub(pool);
 662         inflight = page_pool_inflight(pool);
 663         if (!inflight)
 664                 page_pool_free(pool);
 665
 666         return inflight;
 667 }
 668
 669 static void page_pool_release_retry(struct work_struct *wq)
 670 {
 671         struct delayed_work *dwq = to_delayed_work(wq);
 672         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 673         int inflight;
 674
 675         inflight = page_pool_release(pool);
 676         if (!inflight)
 677                 return;
 678
 679         /* Periodic warning */
 680         if (time_after_eq(jiffies, pool->defer_warn)) {
 681                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 682
 683                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 684                         __func__, inflight, sec);
 685                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 686         }
 687
 688         /* Still not ready to be disconnected, retry later */
 689         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 690 }
 691
 692 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
 693                            struct xdp_mem_info *mem)
 694 {
 695         refcount_inc(&pool->user_cnt);
 696         pool->disconnect = disconnect;
 697         pool->xdp_mem_id = mem->id;
 698 }
 699
 700 void page_pool_destroy(struct page_pool *pool)
 701 {
 702         if (!pool)
 703                 return;
 704
 705         if (!page_pool_put(pool))
 706                 return;
 707
 708         page_pool_free_frag(pool);
 709
 710         if (!page_pool_release(pool))
 711                 return;
 712
 713         pool->defer_start = jiffies;
 714         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 715
 716         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 717         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 718 }
 719 EXPORT_SYMBOL(page_pool_destroy);
 720
 721 /* Caller must provide appropriate safe context, e.g. NAPI. */
 722 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 723 {
 724         struct page *page;
 725
 726         trace_page_pool_update_nid(pool, new_nid);
 727         pool->p.nid = new_nid;
 728
 729         /* Flush pool alloc cache, as refill will check NUMA node */
 730         while (pool->alloc.count) {
 731                 page = pool->alloc.cache[--pool->alloc.count];
 732                 page_pool_return_page(pool, page);
 733         }
 734 }
 735 EXPORT_SYMBOL(page_pool_update_nid);
 736
 737 bool page_pool_return_skb_page(struct page *page)
 738 {
 739         struct page_pool *pp;
 740
 741         page = compound_head(page);
 742
 743         /* page->pp_magic is OR'ed with PP_SIGNATURE after the allocation
 744          * in order to preserve any existing bits, such as bit 0 for the
 745          * head page of compound page and bit 1 for pfmemalloc page, so
 746          * mask those bits for freeing side when doing below checking,
 747          * and page_is_pfmemalloc() is checked in __page_pool_put_page()
 748          * to avoid recycling the pfmemalloc page.
 749          */
 750         if (unlikely((page->pp_magic & ~0x3UL) != PP_SIGNATURE))
 751                 return false;
 752
 753         pp = page->pp;
 754
 755         /* Driver set this to memory recycling info. Reset it on recycle.
 756          * This will *not* work for NIC using a split-page memory model.
 757          * The page will be returned to the pool here regardless of the
 758          * 'flipped' fragment being in use or not.
 759          */
 760         page_pool_put_full_page(pp, page, false);
 761
 762         return true;
 763 }
 764 EXPORT_SYMBOL(page_pool_return_skb_page);