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