net/core/page_pool.c

   1 /* SPDX-License-Identifier: GPL-2.0
   2  *
   3  * page_pool.c
   4  *      Author: Jesper Dangaard Brouer <netoptimizer@brouer.com>
   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                 /* DMA-mapping is not supported on 32-bit systems with
  53                  * 64-bit DMA mapping.
  54                  */
  55                 if (sizeof(dma_addr_t) > sizeof(unsigned long))
  56                         return -EOPNOTSUPP;
  57
  58                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
  59                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
  60                         return -EINVAL;
  61         }
  62
  63         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
  64                 /* In order to request DMA-sync-for-device the page
  65                  * needs to be mapped
  66                  */
  67                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
  68                         return -EINVAL;
  69
  70                 if (!pool->p.max_len)
  71                         return -EINVAL;
  72
  73                 /* pool->p.offset has to be set according to the address
  74                  * offset used by the DMA engine to start copying rx data
  75                  */
  76         }
  77
  78         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
  79                 return -ENOMEM;
  80
  81         atomic_set(&pool->pages_state_release_cnt, 0);
  82
  83         /* Driver calling page_pool_create() also call page_pool_destroy() */
  84         refcount_set(&pool->user_cnt, 1);
  85
  86         if (pool->p.flags & PP_FLAG_DMA_MAP)
  87                 get_device(pool->p.dev);
  88
  89         return 0;
  90 }
  91
  92 struct page_pool *page_pool_create(const struct page_pool_params *params)
  93 {
  94         struct page_pool *pool;
  95         int err;
  96
  97         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
  98         if (!pool)
  99                 return ERR_PTR(-ENOMEM);
 100
 101         err = page_pool_init(pool, params);
 102         if (err < 0) {
 103                 pr_warn("%s() gave up with errno %d\n", __func__, err);
 104                 kfree(pool);
 105                 return ERR_PTR(err);
 106         }
 107
 108         return pool;
 109 }
 110 EXPORT_SYMBOL(page_pool_create);
 111
 112 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 113
 114 noinline
 115 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 116 {
 117         struct ptr_ring *r = &pool->ring;
 118         struct page *page;
 119         int pref_nid; /* preferred NUMA node */
 120
 121         /* Quicker fallback, avoid locks when ring is empty */
 122         if (__ptr_ring_empty(r))
 123                 return NULL;
 124
 125         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 126          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 127          */
 128 #ifdef CONFIG_NUMA
 129         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 130 #else
 131         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 132         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 133 #endif
 134
 135         /* Slower-path: Get pages from locked ring queue */
 136         spin_lock(&r->consumer_lock);
 137
 138         /* Refill alloc array, but only if NUMA match */
 139         do {
 140                 page = __ptr_ring_consume(r);
 141                 if (unlikely(!page))
 142                         break;
 143
 144                 if (likely(page_to_nid(page) == pref_nid)) {
 145                         pool->alloc.cache[pool->alloc.count++] = page;
 146                 } else {
 147                         /* NUMA mismatch;
 148                          * (1) release 1 page to page-allocator and
 149                          * (2) break out to fallthrough to alloc_pages_node.
 150                          * This limit stress on page buddy alloactor.
 151                          */
 152                         page_pool_return_page(pool, page);
 153                         page = NULL;
 154                         break;
 155                 }
 156         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 157
 158         /* Return last page */
 159         if (likely(pool->alloc.count > 0))
 160                 page = pool->alloc.cache[--pool->alloc.count];
 161
 162         spin_unlock(&r->consumer_lock);
 163         return page;
 164 }
 165
 166 /* fast path */
 167 static struct page *__page_pool_get_cached(struct page_pool *pool)
 168 {
 169         struct page *page;
 170
 171         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 172         if (likely(pool->alloc.count)) {
 173                 /* Fast-path */
 174                 page = pool->alloc.cache[--pool->alloc.count];
 175         } else {
 176                 page = page_pool_refill_alloc_cache(pool);
 177         }
 178
 179         return page;
 180 }
 181
 182 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 183                                           struct page *page,
 184                                           unsigned int dma_sync_size)
 185 {
 186         dma_addr_t dma_addr = page_pool_get_dma_addr(page);
 187
 188         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 189         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 190                                          pool->p.offset, dma_sync_size,
 191                                          pool->p.dma_dir);
 192 }
 193
 194 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 195 {
 196         dma_addr_t dma;
 197
 198         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 199          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 200          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 201          * This mapping is kept for lifetime of page, until leaving pool.
 202          */
 203         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 204                                  (PAGE_SIZE << pool->p.order),
 205                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 206         if (dma_mapping_error(pool->p.dev, dma))
 207                 return false;
 208
 209         page_pool_set_dma_addr(page, dma);
 210
 211         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 212                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 213
 214         return true;
 215 }
 216
 217 static void page_pool_set_pp_info(struct page_pool *pool,
 218                                   struct page *page)
 219 {
 220         page->pp = pool;
 221         page->pp_magic |= PP_SIGNATURE;
 222 }
 223
 224 static void page_pool_clear_pp_info(struct page *page)
 225 {
 226         page->pp_magic = 0;
 227         page->pp = NULL;
 228 }
 229
 230 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 231                                                  gfp_t gfp)
 232 {
 233         struct page *page;
 234
 235         gfp |= __GFP_COMP;
 236         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 237         if (unlikely(!page))
 238                 return NULL;
 239
 240         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 241             unlikely(!page_pool_dma_map(pool, page))) {
 242                 put_page(page);
 243                 return NULL;
 244         }
 245
 246         page_pool_set_pp_info(pool, page);
 247
 248         /* Track how many pages are held 'in-flight' */
 249         pool->pages_state_hold_cnt++;
 250         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 251         return page;
 252 }
 253
 254 /* slow path */
 255 noinline
 256 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 257                                                  gfp_t gfp)
 258 {
 259         const int bulk = PP_ALLOC_CACHE_REFILL;
 260         unsigned int pp_flags = pool->p.flags;
 261         unsigned int pp_order = pool->p.order;
 262         struct page *page;
 263         int i, nr_pages;
 264
 265         /* Don't support bulk alloc for high-order pages */
 266         if (unlikely(pp_order))
 267                 return __page_pool_alloc_page_order(pool, gfp);
 268
 269         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 270         if (unlikely(pool->alloc.count > 0))
 271                 return pool->alloc.cache[--pool->alloc.count];
 272
 273         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 274         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 275
 276         nr_pages = alloc_pages_bulk_array(gfp, bulk, pool->alloc.cache);
 277         if (unlikely(!nr_pages))
 278                 return NULL;
 279
 280         /* Pages have been filled into alloc.cache array, but count is zero and
 281          * page element have not been (possibly) DMA mapped.
 282          */
 283         for (i = 0; i < nr_pages; i++) {
 284                 page = pool->alloc.cache[i];
 285                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 286                     unlikely(!page_pool_dma_map(pool, page))) {
 287                         put_page(page);
 288                         continue;
 289                 }
 290
 291                 page_pool_set_pp_info(pool, page);
 292                 pool->alloc.cache[pool->alloc.count++] = page;
 293                 /* Track how many pages are held 'in-flight' */
 294                 pool->pages_state_hold_cnt++;
 295                 trace_page_pool_state_hold(pool, page,
 296                                            pool->pages_state_hold_cnt);
 297         }
 298
 299         /* Return last page */
 300         if (likely(pool->alloc.count > 0))
 301                 page = pool->alloc.cache[--pool->alloc.count];
 302         else
 303                 page = NULL;
 304
 305         /* When page just alloc'ed is should/must have refcnt 1. */
 306         return page;
 307 }
 308
 309 /* For using page_pool replace: alloc_pages() API calls, but provide
 310  * synchronization guarantee for allocation side.
 311  */
 312 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 313 {
 314         struct page *page;
 315
 316         /* Fast-path: Get a page from cache */
 317         page = __page_pool_get_cached(pool);
 318         if (page)
 319                 return page;
 320
 321         /* Slow-path: cache empty, do real allocation */
 322         page = __page_pool_alloc_pages_slow(pool, gfp);
 323         return page;
 324 }
 325 EXPORT_SYMBOL(page_pool_alloc_pages);
 326
 327 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 328  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 329  */
 330 #define _distance(a, b) (s32)((a) - (b))
 331
 332 static s32 page_pool_inflight(struct page_pool *pool)
 333 {
 334         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 335         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 336         s32 inflight;
 337
 338         inflight = _distance(hold_cnt, release_cnt);
 339
 340         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 341         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 342
 343         return inflight;
 344 }
 345
 346 /* Disconnects a page (from a page_pool).  API users can have a need
 347  * to disconnect a page (from a page_pool), to allow it to be used as
 348  * a regular page (that will eventually be returned to the normal
 349  * page-allocator via put_page).
 350  */
 351 void page_pool_release_page(struct page_pool *pool, struct page *page)
 352 {
 353         dma_addr_t dma;
 354         int count;
 355
 356         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 357                 /* Always account for inflight pages, even if we didn't
 358                  * map them
 359                  */
 360                 goto skip_dma_unmap;
 361
 362         dma = page_pool_get_dma_addr(page);
 363
 364         /* When page is unmapped, it cannot be returned to our pool */
 365         dma_unmap_page_attrs(pool->p.dev, dma,
 366                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 367                              DMA_ATTR_SKIP_CPU_SYNC);
 368         page_pool_set_dma_addr(page, 0);
 369 skip_dma_unmap:
 370         page_pool_clear_pp_info(page);
 371
 372         /* This may be the last page returned, releasing the pool, so
 373          * it is not safe to reference pool afterwards.
 374          */
 375         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 376         trace_page_pool_state_release(pool, page, count);
 377 }
 378 EXPORT_SYMBOL(page_pool_release_page);
 379
 380 /* Return a page to the page allocator, cleaning up our state */
 381 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 382 {
 383         page_pool_release_page(pool, page);
 384
 385         put_page(page);
 386         /* An optimization would be to call __free_pages(page, pool->p.order)
 387          * knowing page is not part of page-cache (thus avoiding a
 388          * __page_cache_release() call).
 389          */
 390 }
 391
 392 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 393 {
 394         int ret;
 395         /* BH protection not needed if current is serving softirq */
 396         if (in_serving_softirq())
 397                 ret = ptr_ring_produce(&pool->ring, page);
 398         else
 399                 ret = ptr_ring_produce_bh(&pool->ring, page);
 400
 401         return (ret == 0) ? true : false;
 402 }
 403
 404 /* Only allow direct recycling in special circumstances, into the
 405  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 406  *
 407  * Caller must provide appropriate safe context.
 408  */
 409 static bool page_pool_recycle_in_cache(struct page *page,
 410                                        struct page_pool *pool)
 411 {
 412         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
 413                 return false;
 414
 415         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 416         pool->alloc.cache[pool->alloc.count++] = page;
 417         return true;
 418 }
 419
 420 /* If the page refcnt == 1, this will try to recycle the page.
 421  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 422  * the configured size min(dma_sync_size, pool->max_len).
 423  * If the page refcnt != 1, then the page will be returned to memory
 424  * subsystem.
 425  */
 426 static __always_inline struct page *
 427 __page_pool_put_page(struct page_pool *pool, struct page *page,
 428                      unsigned int dma_sync_size, bool allow_direct)
 429 {
 430         /* It is not the last user for the page frag case */
 431         if (pool->p.flags & PP_FLAG_PAGE_FRAG &&
 432             page_pool_atomic_sub_frag_count_return(page, 1))
 433                 return NULL;
 434
 435         /* This allocator is optimized for the XDP mode that uses
 436          * one-frame-per-page, but have fallbacks that act like the
 437          * regular page allocator APIs.
 438          *
 439          * refcnt == 1 means page_pool owns page, and can recycle it.
 440          *
 441          * page is NOT reusable when allocated when system is under
 442          * some pressure. (page_is_pfmemalloc)
 443          */
 444         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 445                 /* Read barrier done in page_ref_count / READ_ONCE */
 446
 447                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 448                         page_pool_dma_sync_for_device(pool, page,
 449                                                       dma_sync_size);
 450
 451                 if (allow_direct && in_serving_softirq() &&
 452                     page_pool_recycle_in_cache(page, pool))
 453                         return NULL;
 454
 455                 /* Page found as candidate for recycling */
 456                 return page;
 457         }
 458         /* Fallback/non-XDP mode: API user have elevated refcnt.
 459          *
 460          * Many drivers split up the page into fragments, and some
 461          * want to keep doing this to save memory and do refcnt based
 462          * recycling. Support this use case too, to ease drivers
 463          * switching between XDP/non-XDP.
 464          *
 465          * In-case page_pool maintains the DMA mapping, API user must
 466          * call page_pool_put_page once.  In this elevated refcnt
 467          * case, the DMA is unmapped/released, as driver is likely
 468          * doing refcnt based recycle tricks, meaning another process
 469          * will be invoking put_page.
 470          */
 471         /* Do not replace this with page_pool_return_page() */
 472         page_pool_release_page(pool, page);
 473         put_page(page);
 474
 475         return NULL;
 476 }
 477
 478 void page_pool_put_page(struct page_pool *pool, struct page *page,
 479                         unsigned int dma_sync_size, bool allow_direct)
 480 {
 481         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 482         if (page && !page_pool_recycle_in_ring(pool, page)) {
 483                 /* Cache full, fallback to free pages */
 484                 page_pool_return_page(pool, page);
 485         }
 486 }
 487 EXPORT_SYMBOL(page_pool_put_page);
 488
 489 /* Caller must not use data area after call, as this function overwrites it */
 490 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 491                              int count)
 492 {
 493         int i, bulk_len = 0;
 494
 495         for (i = 0; i < count; i++) {
 496                 struct page *page = virt_to_head_page(data[i]);
 497
 498                 page = __page_pool_put_page(pool, page, -1, false);
 499                 /* Approved for bulk recycling in ptr_ring cache */
 500                 if (page)
 501                         data[bulk_len++] = page;
 502         }
 503
 504         if (unlikely(!bulk_len))
 505                 return;
 506
 507         /* Bulk producer into ptr_ring page_pool cache */
 508         page_pool_ring_lock(pool);
 509         for (i = 0; i < bulk_len; i++) {
 510                 if (__ptr_ring_produce(&pool->ring, data[i]))
 511                         break; /* ring full */
 512         }
 513         page_pool_ring_unlock(pool);
 514
 515         /* Hopefully all pages was return into ptr_ring */
 516         if (likely(i == bulk_len))
 517                 return;
 518
 519         /* ptr_ring cache full, free remaining pages outside producer lock
 520          * since put_page() with refcnt == 1 can be an expensive operation
 521          */
 522         for (; i < bulk_len; i++)
 523                 page_pool_return_page(pool, data[i]);
 524 }
 525 EXPORT_SYMBOL(page_pool_put_page_bulk);
 526
 527 static struct page *page_pool_drain_frag(struct page_pool *pool,
 528                                          struct page *page)
 529 {
 530         long drain_count = BIAS_MAX - pool->frag_users;
 531
 532         /* Some user is still using the page frag */
 533         if (likely(page_pool_atomic_sub_frag_count_return(page,
 534                                                           drain_count)))
 535                 return NULL;
 536
 537         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 538                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 539                         page_pool_dma_sync_for_device(pool, page, -1);
 540
 541                 return page;
 542         }
 543
 544         page_pool_return_page(pool, page);
 545         return NULL;
 546 }
 547
 548 static void page_pool_free_frag(struct page_pool *pool)
 549 {
 550         long drain_count = BIAS_MAX - pool->frag_users;
 551         struct page *page = pool->frag_page;
 552
 553         pool->frag_page = NULL;
 554
 555         if (!page ||
 556             page_pool_atomic_sub_frag_count_return(page, drain_count))
 557                 return;
 558
 559         page_pool_return_page(pool, page);
 560 }
 561
 562 struct page *page_pool_alloc_frag(struct page_pool *pool,
 563                                   unsigned int *offset,
 564                                   unsigned int size, gfp_t gfp)
 565 {
 566         unsigned int max_size = PAGE_SIZE << pool->p.order;
 567         struct page *page = pool->frag_page;
 568
 569         if (WARN_ON(!(pool->p.flags & PP_FLAG_PAGE_FRAG) ||
 570                     size > max_size))
 571                 return NULL;
 572
 573         size = ALIGN(size, dma_get_cache_alignment());
 574         *offset = pool->frag_offset;
 575
 576         if (page && *offset + size > max_size) {
 577                 page = page_pool_drain_frag(pool, page);
 578                 if (page)
 579                         goto frag_reset;
 580         }
 581
 582         if (!page) {
 583                 page = page_pool_alloc_pages(pool, gfp);
 584                 if (unlikely(!page)) {
 585                         pool->frag_page = NULL;
 586                         return NULL;
 587                 }
 588
 589                 pool->frag_page = page;
 590
 591 frag_reset:
 592                 pool->frag_users = 1;
 593                 *offset = 0;
 594                 pool->frag_offset = size;
 595                 page_pool_set_frag_count(page, BIAS_MAX);
 596                 return page;
 597         }
 598
 599         pool->frag_users++;
 600         pool->frag_offset = *offset + size;
 601         return page;
 602 }
 603 EXPORT_SYMBOL(page_pool_alloc_frag);
 604
 605 static void page_pool_empty_ring(struct page_pool *pool)
 606 {
 607         struct page *page;
 608
 609         /* Empty recycle ring */
 610         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 611                 /* Verify the refcnt invariant of cached pages */
 612                 if (!(page_ref_count(page) == 1))
 613                         pr_crit("%s() page_pool refcnt %d violation\n",
 614                                 __func__, page_ref_count(page));
 615
 616                 page_pool_return_page(pool, page);
 617         }
 618 }
 619
 620 static void page_pool_free(struct page_pool *pool)
 621 {
 622         if (pool->disconnect)
 623                 pool->disconnect(pool);
 624
 625         ptr_ring_cleanup(&pool->ring, NULL);
 626
 627         if (pool->p.flags & PP_FLAG_DMA_MAP)
 628                 put_device(pool->p.dev);
 629
 630         kfree(pool);
 631 }
 632
 633 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 634 {
 635         struct page *page;
 636
 637         if (pool->destroy_cnt)
 638                 return;
 639
 640         /* Empty alloc cache, assume caller made sure this is
 641          * no-longer in use, and page_pool_alloc_pages() cannot be
 642          * call concurrently.
 643          */
 644         while (pool->alloc.count) {
 645                 page = pool->alloc.cache[--pool->alloc.count];
 646                 page_pool_return_page(pool, page);
 647         }
 648 }
 649
 650 static void page_pool_scrub(struct page_pool *pool)
 651 {
 652         page_pool_empty_alloc_cache_once(pool);
 653         pool->destroy_cnt++;
 654
 655         /* No more consumers should exist, but producers could still
 656          * be in-flight.
 657          */
 658         page_pool_empty_ring(pool);
 659 }
 660
 661 static int page_pool_release(struct page_pool *pool)
 662 {
 663         int inflight;
 664
 665         page_pool_scrub(pool);
 666         inflight = page_pool_inflight(pool);
 667         if (!inflight)
 668                 page_pool_free(pool);
 669
 670         return inflight;
 671 }
 672
 673 static void page_pool_release_retry(struct work_struct *wq)
 674 {
 675         struct delayed_work *dwq = to_delayed_work(wq);
 676         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 677         int inflight;
 678
 679         inflight = page_pool_release(pool);
 680         if (!inflight)
 681                 return;
 682
 683         /* Periodic warning */
 684         if (time_after_eq(jiffies, pool->defer_warn)) {
 685                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 686
 687                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 688                         __func__, inflight, sec);
 689                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 690         }
 691
 692         /* Still not ready to be disconnected, retry later */
 693         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 694 }
 695
 696 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *))
 697 {
 698         refcount_inc(&pool->user_cnt);
 699         pool->disconnect = disconnect;
 700 }
 701
 702 void page_pool_destroy(struct page_pool *pool)
 703 {
 704         if (!pool)
 705                 return;
 706
 707         if (!page_pool_put(pool))
 708                 return;
 709
 710         page_pool_free_frag(pool);
 711
 712         if (!page_pool_release(pool))
 713                 return;
 714
 715         pool->defer_start = jiffies;
 716         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 717
 718         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 719         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 720 }
 721 EXPORT_SYMBOL(page_pool_destroy);
 722
 723 /* Caller must provide appropriate safe context, e.g. NAPI. */
 724 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 725 {
 726         struct page *page;
 727
 728         trace_page_pool_update_nid(pool, new_nid);
 729         pool->p.nid = new_nid;
 730
 731         /* Flush pool alloc cache, as refill will check NUMA node */
 732         while (pool->alloc.count) {
 733                 page = pool->alloc.cache[--pool->alloc.count];
 734                 page_pool_return_page(pool, page);
 735         }
 736 }
 737 EXPORT_SYMBOL(page_pool_update_nid);
 738
 739 bool page_pool_return_skb_page(struct page *page)
 740 {
 741         struct page_pool *pp;
 742
 743         page = compound_head(page);
 744
 745         /* page->pp_magic is OR'ed with PP_SIGNATURE after the allocation
 746          * in order to preserve any existing bits, such as bit 0 for the
 747          * head page of compound page and bit 1 for pfmemalloc page, so
 748          * mask those bits for freeing side when doing below checking,
 749          * and page_is_pfmemalloc() is checked in __page_pool_put_page()
 750          * to avoid recycling the pfmemalloc page.
 751          */
 752         if (unlikely((page->pp_magic & ~0x3UL) != PP_SIGNATURE))
 753                 return false;
 754
 755         pp = page->pp;
 756
 757         /* Driver set this to memory recycling info. Reset it on recycle.
 758          * This will *not* work for NIC using a split-page memory model.
 759          * The page will be returned to the pool here regardless of the
 760          * 'flipped' fragment being in use or not.
 761          */
 762         page_pool_put_full_page(pp, page, false);
 763
 764         return true;
 765 }
 766 EXPORT_SYMBOL(page_pool_return_skb_page);