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 #include <linux/ethtool.h>
  22 #include <linux/netdevice.h>
  23
  24 #include <trace/events/page_pool.h>
  25
  26 #define DEFER_TIME (msecs_to_jiffies(1000))
  27 #define DEFER_WARN_INTERVAL (60 * HZ)
  28
  29 #define BIAS_MAX        LONG_MAX
  30
  31 #ifdef CONFIG_PAGE_POOL_STATS
  32 /* alloc_stat_inc is intended to be used in softirq context */
  33 #define alloc_stat_inc(pool, __stat)    (pool->alloc_stats.__stat++)
  34 /* recycle_stat_inc is safe to use when preemption is possible. */
  35 #define recycle_stat_inc(pool, __stat)                                                  \
  36         do {                                                                            \
  37                 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats;       \
  38                 this_cpu_inc(s->__stat);                                                \
  39         } while (0)
  40
  41 #define recycle_stat_add(pool, __stat, val)                                             \
  42         do {                                                                            \
  43                 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats;       \
  44                 this_cpu_add(s->__stat, val);                                           \
  45         } while (0)
  46
  47 static const char pp_stats[][ETH_GSTRING_LEN] = {
  48         "rx_pp_alloc_fast",
  49         "rx_pp_alloc_slow",
  50         "rx_pp_alloc_slow_ho",
  51         "rx_pp_alloc_empty",
  52         "rx_pp_alloc_refill",
  53         "rx_pp_alloc_waive",
  54         "rx_pp_recycle_cached",
  55         "rx_pp_recycle_cache_full",
  56         "rx_pp_recycle_ring",
  57         "rx_pp_recycle_ring_full",
  58         "rx_pp_recycle_released_ref",
  59 };
  60
  61 bool page_pool_get_stats(struct page_pool *pool,
  62                          struct page_pool_stats *stats)
  63 {
  64         int cpu = 0;
  65
  66         if (!stats)
  67                 return false;
  68
  69         /* The caller is responsible to initialize stats. */
  70         stats->alloc_stats.fast += pool->alloc_stats.fast;
  71         stats->alloc_stats.slow += pool->alloc_stats.slow;
  72         stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
  73         stats->alloc_stats.empty += pool->alloc_stats.empty;
  74         stats->alloc_stats.refill += pool->alloc_stats.refill;
  75         stats->alloc_stats.waive += pool->alloc_stats.waive;
  76
  77         for_each_possible_cpu(cpu) {
  78                 const struct page_pool_recycle_stats *pcpu =
  79                         per_cpu_ptr(pool->recycle_stats, cpu);
  80
  81                 stats->recycle_stats.cached += pcpu->cached;
  82                 stats->recycle_stats.cache_full += pcpu->cache_full;
  83                 stats->recycle_stats.ring += pcpu->ring;
  84                 stats->recycle_stats.ring_full += pcpu->ring_full;
  85                 stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
  86         }
  87
  88         return true;
  89 }
  90 EXPORT_SYMBOL(page_pool_get_stats);
  91
  92 u8 *page_pool_ethtool_stats_get_strings(u8 *data)
  93 {
  94         int i;
  95
  96         for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
  97                 memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
  98                 data += ETH_GSTRING_LEN;
  99         }
 100
 101         return data;
 102 }
 103 EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);
 104
 105 int page_pool_ethtool_stats_get_count(void)
 106 {
 107         return ARRAY_SIZE(pp_stats);
 108 }
 109 EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);
 110
 111 u64 *page_pool_ethtool_stats_get(u64 *data, void *stats)
 112 {
 113         struct page_pool_stats *pool_stats = stats;
 114
 115         *data++ = pool_stats->alloc_stats.fast;
 116         *data++ = pool_stats->alloc_stats.slow;
 117         *data++ = pool_stats->alloc_stats.slow_high_order;
 118         *data++ = pool_stats->alloc_stats.empty;
 119         *data++ = pool_stats->alloc_stats.refill;
 120         *data++ = pool_stats->alloc_stats.waive;
 121         *data++ = pool_stats->recycle_stats.cached;
 122         *data++ = pool_stats->recycle_stats.cache_full;
 123         *data++ = pool_stats->recycle_stats.ring;
 124         *data++ = pool_stats->recycle_stats.ring_full;
 125         *data++ = pool_stats->recycle_stats.released_refcnt;
 126
 127         return data;
 128 }
 129 EXPORT_SYMBOL(page_pool_ethtool_stats_get);
 130
 131 #else
 132 #define alloc_stat_inc(pool, __stat)
 133 #define recycle_stat_inc(pool, __stat)
 134 #define recycle_stat_add(pool, __stat, val)
 135 #endif
 136
 137 static int page_pool_init(struct page_pool *pool,
 138                           const struct page_pool_params *params)
 139 {
 140         unsigned int ring_qsize = 1024; /* Default */
 141
 142         memcpy(&pool->p, params, sizeof(pool->p));
 143
 144         /* Validate only known flags were used */
 145         if (pool->p.flags & ~(PP_FLAG_ALL))
 146                 return -EINVAL;
 147
 148         if (pool->p.pool_size)
 149                 ring_qsize = pool->p.pool_size;
 150
 151         /* Sanity limit mem that can be pinned down */
 152         if (ring_qsize > 32768)
 153                 return -E2BIG;
 154
 155         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
 156          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
 157          * which is the XDP_TX use-case.
 158          */
 159         if (pool->p.flags & PP_FLAG_DMA_MAP) {
 160                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
 161                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
 162                         return -EINVAL;
 163         }
 164
 165         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
 166                 /* In order to request DMA-sync-for-device the page
 167                  * needs to be mapped
 168                  */
 169                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 170                         return -EINVAL;
 171
 172                 if (!pool->p.max_len)
 173                         return -EINVAL;
 174
 175                 /* pool->p.offset has to be set according to the address
 176                  * offset used by the DMA engine to start copying rx data
 177                  */
 178         }
 179
 180         if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT &&
 181             pool->p.flags & PP_FLAG_PAGE_FRAG)
 182                 return -EINVAL;
 183
 184 #ifdef CONFIG_PAGE_POOL_STATS
 185         pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
 186         if (!pool->recycle_stats)
 187                 return -ENOMEM;
 188 #endif
 189
 190         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
 191                 return -ENOMEM;
 192
 193         atomic_set(&pool->pages_state_release_cnt, 0);
 194
 195         /* Driver calling page_pool_create() also call page_pool_destroy() */
 196         refcount_set(&pool->user_cnt, 1);
 197
 198         if (pool->p.flags & PP_FLAG_DMA_MAP)
 199                 get_device(pool->p.dev);
 200
 201         return 0;
 202 }
 203
 204 struct page_pool *page_pool_create(const struct page_pool_params *params)
 205 {
 206         struct page_pool *pool;
 207         int err;
 208
 209         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
 210         if (!pool)
 211                 return ERR_PTR(-ENOMEM);
 212
 213         err = page_pool_init(pool, params);
 214         if (err < 0) {
 215                 pr_warn("%s() gave up with errno %d\n", __func__, err);
 216                 kfree(pool);
 217                 return ERR_PTR(err);
 218         }
 219
 220         return pool;
 221 }
 222 EXPORT_SYMBOL(page_pool_create);
 223
 224 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 225
 226 noinline
 227 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 228 {
 229         struct ptr_ring *r = &pool->ring;
 230         struct page *page;
 231         int pref_nid; /* preferred NUMA node */
 232
 233         /* Quicker fallback, avoid locks when ring is empty */
 234         if (__ptr_ring_empty(r)) {
 235                 alloc_stat_inc(pool, empty);
 236                 return NULL;
 237         }
 238
 239         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 240          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 241          */
 242 #ifdef CONFIG_NUMA
 243         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 244 #else
 245         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 246         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 247 #endif
 248
 249         /* Refill alloc array, but only if NUMA match */
 250         do {
 251                 page = __ptr_ring_consume(r);
 252                 if (unlikely(!page))
 253                         break;
 254
 255                 if (likely(page_to_nid(page) == pref_nid)) {
 256                         pool->alloc.cache[pool->alloc.count++] = page;
 257                 } else {
 258                         /* NUMA mismatch;
 259                          * (1) release 1 page to page-allocator and
 260                          * (2) break out to fallthrough to alloc_pages_node.
 261                          * This limit stress on page buddy alloactor.
 262                          */
 263                         page_pool_return_page(pool, page);
 264                         alloc_stat_inc(pool, waive);
 265                         page = NULL;
 266                         break;
 267                 }
 268         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 269
 270         /* Return last page */
 271         if (likely(pool->alloc.count > 0)) {
 272                 page = pool->alloc.cache[--pool->alloc.count];
 273                 alloc_stat_inc(pool, refill);
 274         }
 275
 276         return page;
 277 }
 278
 279 /* fast path */
 280 static struct page *__page_pool_get_cached(struct page_pool *pool)
 281 {
 282         struct page *page;
 283
 284         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 285         if (likely(pool->alloc.count)) {
 286                 /* Fast-path */
 287                 page = pool->alloc.cache[--pool->alloc.count];
 288                 alloc_stat_inc(pool, fast);
 289         } else {
 290                 page = page_pool_refill_alloc_cache(pool);
 291         }
 292
 293         return page;
 294 }
 295
 296 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 297                                           struct page *page,
 298                                           unsigned int dma_sync_size)
 299 {
 300         dma_addr_t dma_addr = page_pool_get_dma_addr(page);
 301
 302         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 303         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 304                                          pool->p.offset, dma_sync_size,
 305                                          pool->p.dma_dir);
 306 }
 307
 308 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 309 {
 310         dma_addr_t dma;
 311
 312         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 313          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 314          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 315          * This mapping is kept for lifetime of page, until leaving pool.
 316          */
 317         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 318                                  (PAGE_SIZE << pool->p.order),
 319                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC |
 320                                                   DMA_ATTR_WEAK_ORDERING);
 321         if (dma_mapping_error(pool->p.dev, dma))
 322                 return false;
 323
 324         page_pool_set_dma_addr(page, dma);
 325
 326         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 327                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 328
 329         return true;
 330 }
 331
 332 static void page_pool_set_pp_info(struct page_pool *pool,
 333                                   struct page *page)
 334 {
 335         page->pp = pool;
 336         page->pp_magic |= PP_SIGNATURE;
 337         if (pool->p.init_callback)
 338                 pool->p.init_callback(page, pool->p.init_arg);
 339 }
 340
 341 static void page_pool_clear_pp_info(struct page *page)
 342 {
 343         page->pp_magic = 0;
 344         page->pp = NULL;
 345 }
 346
 347 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 348                                                  gfp_t gfp)
 349 {
 350         struct page *page;
 351
 352         gfp |= __GFP_COMP;
 353         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 354         if (unlikely(!page))
 355                 return NULL;
 356
 357         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 358             unlikely(!page_pool_dma_map(pool, page))) {
 359                 put_page(page);
 360                 return NULL;
 361         }
 362
 363         alloc_stat_inc(pool, slow_high_order);
 364         page_pool_set_pp_info(pool, page);
 365
 366         /* Track how many pages are held 'in-flight' */
 367         pool->pages_state_hold_cnt++;
 368         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 369         return page;
 370 }
 371
 372 /* slow path */
 373 noinline
 374 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 375                                                  gfp_t gfp)
 376 {
 377         const int bulk = PP_ALLOC_CACHE_REFILL;
 378         unsigned int pp_flags = pool->p.flags;
 379         unsigned int pp_order = pool->p.order;
 380         struct page *page;
 381         int i, nr_pages;
 382
 383         /* Don't support bulk alloc for high-order pages */
 384         if (unlikely(pp_order))
 385                 return __page_pool_alloc_page_order(pool, gfp);
 386
 387         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 388         if (unlikely(pool->alloc.count > 0))
 389                 return pool->alloc.cache[--pool->alloc.count];
 390
 391         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 392         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 393
 394         nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk,
 395                                                pool->alloc.cache);
 396         if (unlikely(!nr_pages))
 397                 return NULL;
 398
 399         /* Pages have been filled into alloc.cache array, but count is zero and
 400          * page element have not been (possibly) DMA mapped.
 401          */
 402         for (i = 0; i < nr_pages; i++) {
 403                 page = pool->alloc.cache[i];
 404                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 405                     unlikely(!page_pool_dma_map(pool, page))) {
 406                         put_page(page);
 407                         continue;
 408                 }
 409
 410                 page_pool_set_pp_info(pool, page);
 411                 pool->alloc.cache[pool->alloc.count++] = page;
 412                 /* Track how many pages are held 'in-flight' */
 413                 pool->pages_state_hold_cnt++;
 414                 trace_page_pool_state_hold(pool, page,
 415                                            pool->pages_state_hold_cnt);
 416         }
 417
 418         /* Return last page */
 419         if (likely(pool->alloc.count > 0)) {
 420                 page = pool->alloc.cache[--pool->alloc.count];
 421                 alloc_stat_inc(pool, slow);
 422         } else {
 423                 page = NULL;
 424         }
 425
 426         /* When page just alloc'ed is should/must have refcnt 1. */
 427         return page;
 428 }
 429
 430 /* For using page_pool replace: alloc_pages() API calls, but provide
 431  * synchronization guarantee for allocation side.
 432  */
 433 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 434 {
 435         struct page *page;
 436
 437         /* Fast-path: Get a page from cache */
 438         page = __page_pool_get_cached(pool);
 439         if (page)
 440                 return page;
 441
 442         /* Slow-path: cache empty, do real allocation */
 443         page = __page_pool_alloc_pages_slow(pool, gfp);
 444         return page;
 445 }
 446 EXPORT_SYMBOL(page_pool_alloc_pages);
 447
 448 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 449  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 450  */
 451 #define _distance(a, b) (s32)((a) - (b))
 452
 453 static s32 page_pool_inflight(struct page_pool *pool)
 454 {
 455         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 456         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 457         s32 inflight;
 458
 459         inflight = _distance(hold_cnt, release_cnt);
 460
 461         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 462         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 463
 464         return inflight;
 465 }
 466
 467 /* Disconnects a page (from a page_pool).  API users can have a need
 468  * to disconnect a page (from a page_pool), to allow it to be used as
 469  * a regular page (that will eventually be returned to the normal
 470  * page-allocator via put_page).
 471  */
 472 void page_pool_release_page(struct page_pool *pool, struct page *page)
 473 {
 474         dma_addr_t dma;
 475         int count;
 476
 477         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 478                 /* Always account for inflight pages, even if we didn't
 479                  * map them
 480                  */
 481                 goto skip_dma_unmap;
 482
 483         dma = page_pool_get_dma_addr(page);
 484
 485         /* When page is unmapped, it cannot be returned to our pool */
 486         dma_unmap_page_attrs(pool->p.dev, dma,
 487                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 488                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 489         page_pool_set_dma_addr(page, 0);
 490 skip_dma_unmap:
 491         page_pool_clear_pp_info(page);
 492
 493         /* This may be the last page returned, releasing the pool, so
 494          * it is not safe to reference pool afterwards.
 495          */
 496         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 497         trace_page_pool_state_release(pool, page, count);
 498 }
 499 EXPORT_SYMBOL(page_pool_release_page);
 500
 501 /* Return a page to the page allocator, cleaning up our state */
 502 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 503 {
 504         page_pool_release_page(pool, page);
 505
 506         put_page(page);
 507         /* An optimization would be to call __free_pages(page, pool->p.order)
 508          * knowing page is not part of page-cache (thus avoiding a
 509          * __page_cache_release() call).
 510          */
 511 }
 512
 513 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 514 {
 515         int ret;
 516         /* BH protection not needed if current is softirq */
 517         if (in_softirq())
 518                 ret = ptr_ring_produce(&pool->ring, page);
 519         else
 520                 ret = ptr_ring_produce_bh(&pool->ring, page);
 521
 522         if (!ret) {
 523                 recycle_stat_inc(pool, ring);
 524                 return true;
 525         }
 526
 527         return false;
 528 }
 529
 530 /* Only allow direct recycling in special circumstances, into the
 531  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 532  *
 533  * Caller must provide appropriate safe context.
 534  */
 535 static bool page_pool_recycle_in_cache(struct page *page,
 536                                        struct page_pool *pool)
 537 {
 538         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
 539                 recycle_stat_inc(pool, cache_full);
 540                 return false;
 541         }
 542
 543         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 544         pool->alloc.cache[pool->alloc.count++] = page;
 545         recycle_stat_inc(pool, cached);
 546         return true;
 547 }
 548
 549 /* If the page refcnt == 1, this will try to recycle the page.
 550  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 551  * the configured size min(dma_sync_size, pool->max_len).
 552  * If the page refcnt != 1, then the page will be returned to memory
 553  * subsystem.
 554  */
 555 static __always_inline struct page *
 556 __page_pool_put_page(struct page_pool *pool, struct page *page,
 557                      unsigned int dma_sync_size, bool allow_direct)
 558 {
 559         /* This allocator is optimized for the XDP mode that uses
 560          * one-frame-per-page, but have fallbacks that act like the
 561          * regular page allocator APIs.
 562          *
 563          * refcnt == 1 means page_pool owns page, and can recycle it.
 564          *
 565          * page is NOT reusable when allocated when system is under
 566          * some pressure. (page_is_pfmemalloc)
 567          */
 568         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 569                 /* Read barrier done in page_ref_count / READ_ONCE */
 570
 571                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 572                         page_pool_dma_sync_for_device(pool, page,
 573                                                       dma_sync_size);
 574
 575                 if (allow_direct && in_softirq() &&
 576                     page_pool_recycle_in_cache(page, pool))
 577                         return NULL;
 578
 579                 /* Page found as candidate for recycling */
 580                 return page;
 581         }
 582         /* Fallback/non-XDP mode: API user have elevated refcnt.
 583          *
 584          * Many drivers split up the page into fragments, and some
 585          * want to keep doing this to save memory and do refcnt based
 586          * recycling. Support this use case too, to ease drivers
 587          * switching between XDP/non-XDP.
 588          *
 589          * In-case page_pool maintains the DMA mapping, API user must
 590          * call page_pool_put_page once.  In this elevated refcnt
 591          * case, the DMA is unmapped/released, as driver is likely
 592          * doing refcnt based recycle tricks, meaning another process
 593          * will be invoking put_page.
 594          */
 595         recycle_stat_inc(pool, released_refcnt);
 596         /* Do not replace this with page_pool_return_page() */
 597         page_pool_release_page(pool, page);
 598         put_page(page);
 599
 600         return NULL;
 601 }
 602
 603 void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
 604                                   unsigned int dma_sync_size, bool allow_direct)
 605 {
 606         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 607         if (page && !page_pool_recycle_in_ring(pool, page)) {
 608                 /* Cache full, fallback to free pages */
 609                 recycle_stat_inc(pool, ring_full);
 610                 page_pool_return_page(pool, page);
 611         }
 612 }
 613 EXPORT_SYMBOL(page_pool_put_defragged_page);
 614
 615 /* Caller must not use data area after call, as this function overwrites it */
 616 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 617                              int count)
 618 {
 619         int i, bulk_len = 0;
 620
 621         for (i = 0; i < count; i++) {
 622                 struct page *page = virt_to_head_page(data[i]);
 623
 624                 /* It is not the last user for the page frag case */
 625                 if (!page_pool_is_last_frag(pool, page))
 626                         continue;
 627
 628                 page = __page_pool_put_page(pool, page, -1, false);
 629                 /* Approved for bulk recycling in ptr_ring cache */
 630                 if (page)
 631                         data[bulk_len++] = page;
 632         }
 633
 634         if (unlikely(!bulk_len))
 635                 return;
 636
 637         /* Bulk producer into ptr_ring page_pool cache */
 638         page_pool_ring_lock(pool);
 639         for (i = 0; i < bulk_len; i++) {
 640                 if (__ptr_ring_produce(&pool->ring, data[i])) {
 641                         /* ring full */
 642                         recycle_stat_inc(pool, ring_full);
 643                         break;
 644                 }
 645         }
 646         recycle_stat_add(pool, ring, i);
 647         page_pool_ring_unlock(pool);
 648
 649         /* Hopefully all pages was return into ptr_ring */
 650         if (likely(i == bulk_len))
 651                 return;
 652
 653         /* ptr_ring cache full, free remaining pages outside producer lock
 654          * since put_page() with refcnt == 1 can be an expensive operation
 655          */
 656         for (; i < bulk_len; i++)
 657                 page_pool_return_page(pool, data[i]);
 658 }
 659 EXPORT_SYMBOL(page_pool_put_page_bulk);
 660
 661 static struct page *page_pool_drain_frag(struct page_pool *pool,
 662                                          struct page *page)
 663 {
 664         long drain_count = BIAS_MAX - pool->frag_users;
 665
 666         /* Some user is still using the page frag */
 667         if (likely(page_pool_defrag_page(page, drain_count)))
 668                 return NULL;
 669
 670         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 671                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 672                         page_pool_dma_sync_for_device(pool, page, -1);
 673
 674                 return page;
 675         }
 676
 677         page_pool_return_page(pool, page);
 678         return NULL;
 679 }
 680
 681 static void page_pool_free_frag(struct page_pool *pool)
 682 {
 683         long drain_count = BIAS_MAX - pool->frag_users;
 684         struct page *page = pool->frag_page;
 685
 686         pool->frag_page = NULL;
 687
 688         if (!page || page_pool_defrag_page(page, drain_count))
 689                 return;
 690
 691         page_pool_return_page(pool, page);
 692 }
 693
 694 struct page *page_pool_alloc_frag(struct page_pool *pool,
 695                                   unsigned int *offset,
 696                                   unsigned int size, gfp_t gfp)
 697 {
 698         unsigned int max_size = PAGE_SIZE << pool->p.order;
 699         struct page *page = pool->frag_page;
 700
 701         if (WARN_ON(!(pool->p.flags & PP_FLAG_PAGE_FRAG) ||
 702                     size > max_size))
 703                 return NULL;
 704
 705         size = ALIGN(size, dma_get_cache_alignment());
 706         *offset = pool->frag_offset;
 707
 708         if (page && *offset + size > max_size) {
 709                 page = page_pool_drain_frag(pool, page);
 710                 if (page) {
 711                         alloc_stat_inc(pool, fast);
 712                         goto frag_reset;
 713                 }
 714         }
 715
 716         if (!page) {
 717                 page = page_pool_alloc_pages(pool, gfp);
 718                 if (unlikely(!page)) {
 719                         pool->frag_page = NULL;
 720                         return NULL;
 721                 }
 722
 723                 pool->frag_page = page;
 724
 725 frag_reset:
 726                 pool->frag_users = 1;
 727                 *offset = 0;
 728                 pool->frag_offset = size;
 729                 page_pool_fragment_page(page, BIAS_MAX);
 730                 return page;
 731         }
 732
 733         pool->frag_users++;
 734         pool->frag_offset = *offset + size;
 735         alloc_stat_inc(pool, fast);
 736         return page;
 737 }
 738 EXPORT_SYMBOL(page_pool_alloc_frag);
 739
 740 static void page_pool_empty_ring(struct page_pool *pool)
 741 {
 742         struct page *page;
 743
 744         /* Empty recycle ring */
 745         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 746                 /* Verify the refcnt invariant of cached pages */
 747                 if (!(page_ref_count(page) == 1))
 748                         pr_crit("%s() page_pool refcnt %d violation\n",
 749                                 __func__, page_ref_count(page));
 750
 751                 page_pool_return_page(pool, page);
 752         }
 753 }
 754
 755 static void page_pool_free(struct page_pool *pool)
 756 {
 757         if (pool->disconnect)
 758                 pool->disconnect(pool);
 759
 760         ptr_ring_cleanup(&pool->ring, NULL);
 761
 762         if (pool->p.flags & PP_FLAG_DMA_MAP)
 763                 put_device(pool->p.dev);
 764
 765 #ifdef CONFIG_PAGE_POOL_STATS
 766         free_percpu(pool->recycle_stats);
 767 #endif
 768         kfree(pool);
 769 }
 770
 771 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 772 {
 773         struct page *page;
 774
 775         if (pool->destroy_cnt)
 776                 return;
 777
 778         /* Empty alloc cache, assume caller made sure this is
 779          * no-longer in use, and page_pool_alloc_pages() cannot be
 780          * call concurrently.
 781          */
 782         while (pool->alloc.count) {
 783                 page = pool->alloc.cache[--pool->alloc.count];
 784                 page_pool_return_page(pool, page);
 785         }
 786 }
 787
 788 static void page_pool_scrub(struct page_pool *pool)
 789 {
 790         page_pool_empty_alloc_cache_once(pool);
 791         pool->destroy_cnt++;
 792
 793         /* No more consumers should exist, but producers could still
 794          * be in-flight.
 795          */
 796         page_pool_empty_ring(pool);
 797 }
 798
 799 static int page_pool_release(struct page_pool *pool)
 800 {
 801         int inflight;
 802
 803         page_pool_scrub(pool);
 804         inflight = page_pool_inflight(pool);
 805         if (!inflight)
 806                 page_pool_free(pool);
 807
 808         return inflight;
 809 }
 810
 811 static void page_pool_release_retry(struct work_struct *wq)
 812 {
 813         struct delayed_work *dwq = to_delayed_work(wq);
 814         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 815         int inflight;
 816
 817         inflight = page_pool_release(pool);
 818         if (!inflight)
 819                 return;
 820
 821         /* Periodic warning */
 822         if (time_after_eq(jiffies, pool->defer_warn)) {
 823                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 824
 825                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 826                         __func__, inflight, sec);
 827                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 828         }
 829
 830         /* Still not ready to be disconnected, retry later */
 831         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 832 }
 833
 834 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
 835                            struct xdp_mem_info *mem)
 836 {
 837         refcount_inc(&pool->user_cnt);
 838         pool->disconnect = disconnect;
 839         pool->xdp_mem_id = mem->id;
 840 }
 841
 842 void page_pool_unlink_napi(struct page_pool *pool)
 843 {
 844         if (!pool->p.napi)
 845                 return;
 846
 847         /* To avoid races with recycling and additional barriers make sure
 848          * pool and NAPI are unlinked when NAPI is disabled.
 849          */
 850         WARN_ON(!test_bit(NAPI_STATE_SCHED, &pool->p.napi->state) ||
 851                 READ_ONCE(pool->p.napi->list_owner) != -1);
 852
 853         WRITE_ONCE(pool->p.napi, NULL);
 854 }
 855 EXPORT_SYMBOL(page_pool_unlink_napi);
 856
 857 void page_pool_destroy(struct page_pool *pool)
 858 {
 859         if (!pool)
 860                 return;
 861
 862         if (!page_pool_put(pool))
 863                 return;
 864
 865         page_pool_unlink_napi(pool);
 866         page_pool_free_frag(pool);
 867
 868         if (!page_pool_release(pool))
 869                 return;
 870
 871         pool->defer_start = jiffies;
 872         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 873
 874         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 875         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 876 }
 877 EXPORT_SYMBOL(page_pool_destroy);
 878
 879 /* Caller must provide appropriate safe context, e.g. NAPI. */
 880 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 881 {
 882         struct page *page;
 883
 884         trace_page_pool_update_nid(pool, new_nid);
 885         pool->p.nid = new_nid;
 886
 887         /* Flush pool alloc cache, as refill will check NUMA node */
 888         while (pool->alloc.count) {
 889                 page = pool->alloc.cache[--pool->alloc.count];
 890                 page_pool_return_page(pool, page);
 891         }
 892 }
 893 EXPORT_SYMBOL(page_pool_update_nid);
 894
 895 bool page_pool_return_skb_page(struct page *page, bool napi_safe)
 896 {
 897         struct napi_struct *napi;
 898         struct page_pool *pp;
 899         bool allow_direct;
 900
 901         page = compound_head(page);
 902
 903         /* page->pp_magic is OR'ed with PP_SIGNATURE after the allocation
 904          * in order to preserve any existing bits, such as bit 0 for the
 905          * head page of compound page and bit 1 for pfmemalloc page, so
 906          * mask those bits for freeing side when doing below checking,
 907          * and page_is_pfmemalloc() is checked in __page_pool_put_page()
 908          * to avoid recycling the pfmemalloc page.
 909          */
 910         if (unlikely((page->pp_magic & ~0x3UL) != PP_SIGNATURE))
 911                 return false;
 912
 913         pp = page->pp;
 914
 915         /* Allow direct recycle if we have reasons to believe that we are
 916          * in the same context as the consumer would run, so there's
 917          * no possible race.
 918          */
 919         napi = READ_ONCE(pp->p.napi);
 920         allow_direct = napi_safe && napi &&
 921                 READ_ONCE(napi->list_owner) == smp_processor_id();
 922
 923         /* Driver set this to memory recycling info. Reset it on recycle.
 924          * This will *not* work for NIC using a split-page memory model.
 925          * The page will be returned to the pool here regardless of the
 926          * 'flipped' fragment being in use or not.
 927          */
 928         page_pool_put_full_page(pp, page, allow_direct);
 929
 930         return true;
 931 }
 932 EXPORT_SYMBOL(page_pool_return_skb_page);