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/helpers.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 /**
  62  * page_pool_get_stats() - fetch page pool stats
  63  * @pool:       pool from which page was allocated
  64  * @stats:      struct page_pool_stats to fill in
  65  *
  66  * Retrieve statistics about the page_pool. This API is only available
  67  * if the kernel has been configured with ``CONFIG_PAGE_POOL_STATS=y``.
  68  * A pointer to a caller allocated struct page_pool_stats structure
  69  * is passed to this API which is filled in. The caller can then report
  70  * those stats to the user (perhaps via ethtool, debugfs, etc.).
  71  */
  72 bool page_pool_get_stats(struct page_pool *pool,
  73                          struct page_pool_stats *stats)
  74 {
  75         int cpu = 0;
  76
  77         if (!stats)
  78                 return false;
  79
  80         /* The caller is responsible to initialize stats. */
  81         stats->alloc_stats.fast += pool->alloc_stats.fast;
  82         stats->alloc_stats.slow += pool->alloc_stats.slow;
  83         stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
  84         stats->alloc_stats.empty += pool->alloc_stats.empty;
  85         stats->alloc_stats.refill += pool->alloc_stats.refill;
  86         stats->alloc_stats.waive += pool->alloc_stats.waive;
  87
  88         for_each_possible_cpu(cpu) {
  89                 const struct page_pool_recycle_stats *pcpu =
  90                         per_cpu_ptr(pool->recycle_stats, cpu);
  91
  92                 stats->recycle_stats.cached += pcpu->cached;
  93                 stats->recycle_stats.cache_full += pcpu->cache_full;
  94                 stats->recycle_stats.ring += pcpu->ring;
  95                 stats->recycle_stats.ring_full += pcpu->ring_full;
  96                 stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
  97         }
  98
  99         return true;
 100 }
 101 EXPORT_SYMBOL(page_pool_get_stats);
 102
 103 u8 *page_pool_ethtool_stats_get_strings(u8 *data)
 104 {
 105         int i;
 106
 107         for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
 108                 memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
 109                 data += ETH_GSTRING_LEN;
 110         }
 111
 112         return data;
 113 }
 114 EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);
 115
 116 int page_pool_ethtool_stats_get_count(void)
 117 {
 118         return ARRAY_SIZE(pp_stats);
 119 }
 120 EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);
 121
 122 u64 *page_pool_ethtool_stats_get(u64 *data, void *stats)
 123 {
 124         struct page_pool_stats *pool_stats = stats;
 125
 126         *data++ = pool_stats->alloc_stats.fast;
 127         *data++ = pool_stats->alloc_stats.slow;
 128         *data++ = pool_stats->alloc_stats.slow_high_order;
 129         *data++ = pool_stats->alloc_stats.empty;
 130         *data++ = pool_stats->alloc_stats.refill;
 131         *data++ = pool_stats->alloc_stats.waive;
 132         *data++ = pool_stats->recycle_stats.cached;
 133         *data++ = pool_stats->recycle_stats.cache_full;
 134         *data++ = pool_stats->recycle_stats.ring;
 135         *data++ = pool_stats->recycle_stats.ring_full;
 136         *data++ = pool_stats->recycle_stats.released_refcnt;
 137
 138         return data;
 139 }
 140 EXPORT_SYMBOL(page_pool_ethtool_stats_get);
 141
 142 #else
 143 #define alloc_stat_inc(pool, __stat)
 144 #define recycle_stat_inc(pool, __stat)
 145 #define recycle_stat_add(pool, __stat, val)
 146 #endif
 147
 148 static bool page_pool_producer_lock(struct page_pool *pool)
 149         __acquires(&pool->ring.producer_lock)
 150 {
 151         bool in_softirq = in_softirq();
 152
 153         if (in_softirq)
 154                 spin_lock(&pool->ring.producer_lock);
 155         else
 156                 spin_lock_bh(&pool->ring.producer_lock);
 157
 158         return in_softirq;
 159 }
 160
 161 static void page_pool_producer_unlock(struct page_pool *pool,
 162                                       bool in_softirq)
 163         __releases(&pool->ring.producer_lock)
 164 {
 165         if (in_softirq)
 166                 spin_unlock(&pool->ring.producer_lock);
 167         else
 168                 spin_unlock_bh(&pool->ring.producer_lock);
 169 }
 170
 171 static int page_pool_init(struct page_pool *pool,
 172                           const struct page_pool_params *params)
 173 {
 174         unsigned int ring_qsize = 1024; /* Default */
 175
 176         memcpy(&pool->p, &params->fast, sizeof(pool->p));
 177         memcpy(&pool->slow, &params->slow, sizeof(pool->slow));
 178
 179         /* Validate only known flags were used */
 180         if (pool->p.flags & ~(PP_FLAG_ALL))
 181                 return -EINVAL;
 182
 183         if (pool->p.pool_size)
 184                 ring_qsize = pool->p.pool_size;
 185
 186         /* Sanity limit mem that can be pinned down */
 187         if (ring_qsize > 32768)
 188                 return -E2BIG;
 189
 190         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
 191          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
 192          * which is the XDP_TX use-case.
 193          */
 194         if (pool->p.flags & PP_FLAG_DMA_MAP) {
 195                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
 196                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
 197                         return -EINVAL;
 198         }
 199
 200         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
 201                 /* In order to request DMA-sync-for-device the page
 202                  * needs to be mapped
 203                  */
 204                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 205                         return -EINVAL;
 206
 207                 if (!pool->p.max_len)
 208                         return -EINVAL;
 209
 210                 /* pool->p.offset has to be set according to the address
 211                  * offset used by the DMA engine to start copying rx data
 212                  */
 213         }
 214
 215         pool->has_init_callback = !!pool->slow.init_callback;
 216
 217 #ifdef CONFIG_PAGE_POOL_STATS
 218         pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
 219         if (!pool->recycle_stats)
 220                 return -ENOMEM;
 221 #endif
 222
 223         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0) {
 224 #ifdef CONFIG_PAGE_POOL_STATS
 225                 free_percpu(pool->recycle_stats);
 226 #endif
 227                 return -ENOMEM;
 228         }
 229
 230         atomic_set(&pool->pages_state_release_cnt, 0);
 231
 232         /* Driver calling page_pool_create() also call page_pool_destroy() */
 233         refcount_set(&pool->user_cnt, 1);
 234
 235         if (pool->p.flags & PP_FLAG_DMA_MAP)
 236                 get_device(pool->p.dev);
 237
 238         return 0;
 239 }
 240
 241 /**
 242  * page_pool_create() - create a page pool.
 243  * @params: parameters, see struct page_pool_params
 244  */
 245 struct page_pool *page_pool_create(const struct page_pool_params *params)
 246 {
 247         struct page_pool *pool;
 248         int err;
 249
 250         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
 251         if (!pool)
 252                 return ERR_PTR(-ENOMEM);
 253
 254         err = page_pool_init(pool, params);
 255         if (err < 0) {
 256                 pr_warn("%s() gave up with errno %d\n", __func__, err);
 257                 kfree(pool);
 258                 return ERR_PTR(err);
 259         }
 260
 261         return pool;
 262 }
 263 EXPORT_SYMBOL(page_pool_create);
 264
 265 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 266
 267 noinline
 268 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 269 {
 270         struct ptr_ring *r = &pool->ring;
 271         struct page *page;
 272         int pref_nid; /* preferred NUMA node */
 273
 274         /* Quicker fallback, avoid locks when ring is empty */
 275         if (__ptr_ring_empty(r)) {
 276                 alloc_stat_inc(pool, empty);
 277                 return NULL;
 278         }
 279
 280         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 281          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 282          */
 283 #ifdef CONFIG_NUMA
 284         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 285 #else
 286         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 287         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 288 #endif
 289
 290         /* Refill alloc array, but only if NUMA match */
 291         do {
 292                 page = __ptr_ring_consume(r);
 293                 if (unlikely(!page))
 294                         break;
 295
 296                 if (likely(page_to_nid(page) == pref_nid)) {
 297                         pool->alloc.cache[pool->alloc.count++] = page;
 298                 } else {
 299                         /* NUMA mismatch;
 300                          * (1) release 1 page to page-allocator and
 301                          * (2) break out to fallthrough to alloc_pages_node.
 302                          * This limit stress on page buddy alloactor.
 303                          */
 304                         page_pool_return_page(pool, page);
 305                         alloc_stat_inc(pool, waive);
 306                         page = NULL;
 307                         break;
 308                 }
 309         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 310
 311         /* Return last page */
 312         if (likely(pool->alloc.count > 0)) {
 313                 page = pool->alloc.cache[--pool->alloc.count];
 314                 alloc_stat_inc(pool, refill);
 315         }
 316
 317         return page;
 318 }
 319
 320 /* fast path */
 321 static struct page *__page_pool_get_cached(struct page_pool *pool)
 322 {
 323         struct page *page;
 324
 325         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 326         if (likely(pool->alloc.count)) {
 327                 /* Fast-path */
 328                 page = pool->alloc.cache[--pool->alloc.count];
 329                 alloc_stat_inc(pool, fast);
 330         } else {
 331                 page = page_pool_refill_alloc_cache(pool);
 332         }
 333
 334         return page;
 335 }
 336
 337 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 338                                           struct page *page,
 339                                           unsigned int dma_sync_size)
 340 {
 341         dma_addr_t dma_addr = page_pool_get_dma_addr(page);
 342
 343         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 344         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 345                                          pool->p.offset, dma_sync_size,
 346                                          pool->p.dma_dir);
 347 }
 348
 349 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 350 {
 351         dma_addr_t dma;
 352
 353         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 354          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 355          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 356          * This mapping is kept for lifetime of page, until leaving pool.
 357          */
 358         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 359                                  (PAGE_SIZE << pool->p.order),
 360                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC |
 361                                                   DMA_ATTR_WEAK_ORDERING);
 362         if (dma_mapping_error(pool->p.dev, dma))
 363                 return false;
 364
 365         if (page_pool_set_dma_addr(page, dma))
 366                 goto unmap_failed;
 367
 368         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 369                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 370
 371         return true;
 372
 373 unmap_failed:
 374         WARN_ON_ONCE("unexpected DMA address, please report to netdev@");
 375         dma_unmap_page_attrs(pool->p.dev, dma,
 376                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 377                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 378         return false;
 379 }
 380
 381 static void page_pool_set_pp_info(struct page_pool *pool,
 382                                   struct page *page)
 383 {
 384         page->pp = pool;
 385         page->pp_magic |= PP_SIGNATURE;
 386
 387         /* Ensuring all pages have been split into one fragment initially:
 388          * page_pool_set_pp_info() is only called once for every page when it
 389          * is allocated from the page allocator and page_pool_fragment_page()
 390          * is dirtying the same cache line as the page->pp_magic above, so
 391          * the overhead is negligible.
 392          */
 393         page_pool_fragment_page(page, 1);
 394         if (pool->has_init_callback)
 395                 pool->slow.init_callback(page, pool->slow.init_arg);
 396 }
 397
 398 static void page_pool_clear_pp_info(struct page *page)
 399 {
 400         page->pp_magic = 0;
 401         page->pp = NULL;
 402 }
 403
 404 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 405                                                  gfp_t gfp)
 406 {
 407         struct page *page;
 408
 409         gfp |= __GFP_COMP;
 410         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 411         if (unlikely(!page))
 412                 return NULL;
 413
 414         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 415             unlikely(!page_pool_dma_map(pool, page))) {
 416                 put_page(page);
 417                 return NULL;
 418         }
 419
 420         alloc_stat_inc(pool, slow_high_order);
 421         page_pool_set_pp_info(pool, page);
 422
 423         /* Track how many pages are held 'in-flight' */
 424         pool->pages_state_hold_cnt++;
 425         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 426         return page;
 427 }
 428
 429 /* slow path */
 430 noinline
 431 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 432                                                  gfp_t gfp)
 433 {
 434         const int bulk = PP_ALLOC_CACHE_REFILL;
 435         unsigned int pp_flags = pool->p.flags;
 436         unsigned int pp_order = pool->p.order;
 437         struct page *page;
 438         int i, nr_pages;
 439
 440         /* Don't support bulk alloc for high-order pages */
 441         if (unlikely(pp_order))
 442                 return __page_pool_alloc_page_order(pool, gfp);
 443
 444         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 445         if (unlikely(pool->alloc.count > 0))
 446                 return pool->alloc.cache[--pool->alloc.count];
 447
 448         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 449         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 450
 451         nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk,
 452                                                pool->alloc.cache);
 453         if (unlikely(!nr_pages))
 454                 return NULL;
 455
 456         /* Pages have been filled into alloc.cache array, but count is zero and
 457          * page element have not been (possibly) DMA mapped.
 458          */
 459         for (i = 0; i < nr_pages; i++) {
 460                 page = pool->alloc.cache[i];
 461                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 462                     unlikely(!page_pool_dma_map(pool, page))) {
 463                         put_page(page);
 464                         continue;
 465                 }
 466
 467                 page_pool_set_pp_info(pool, page);
 468                 pool->alloc.cache[pool->alloc.count++] = page;
 469                 /* Track how many pages are held 'in-flight' */
 470                 pool->pages_state_hold_cnt++;
 471                 trace_page_pool_state_hold(pool, page,
 472                                            pool->pages_state_hold_cnt);
 473         }
 474
 475         /* Return last page */
 476         if (likely(pool->alloc.count > 0)) {
 477                 page = pool->alloc.cache[--pool->alloc.count];
 478                 alloc_stat_inc(pool, slow);
 479         } else {
 480                 page = NULL;
 481         }
 482
 483         /* When page just alloc'ed is should/must have refcnt 1. */
 484         return page;
 485 }
 486
 487 /* For using page_pool replace: alloc_pages() API calls, but provide
 488  * synchronization guarantee for allocation side.
 489  */
 490 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 491 {
 492         struct page *page;
 493
 494         /* Fast-path: Get a page from cache */
 495         page = __page_pool_get_cached(pool);
 496         if (page)
 497                 return page;
 498
 499         /* Slow-path: cache empty, do real allocation */
 500         page = __page_pool_alloc_pages_slow(pool, gfp);
 501         return page;
 502 }
 503 EXPORT_SYMBOL(page_pool_alloc_pages);
 504
 505 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 506  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 507  */
 508 #define _distance(a, b) (s32)((a) - (b))
 509
 510 static s32 page_pool_inflight(struct page_pool *pool)
 511 {
 512         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 513         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 514         s32 inflight;
 515
 516         inflight = _distance(hold_cnt, release_cnt);
 517
 518         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 519         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 520
 521         return inflight;
 522 }
 523
 524 /* Disconnects a page (from a page_pool).  API users can have a need
 525  * to disconnect a page (from a page_pool), to allow it to be used as
 526  * a regular page (that will eventually be returned to the normal
 527  * page-allocator via put_page).
 528  */
 529 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 530 {
 531         dma_addr_t dma;
 532         int count;
 533
 534         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 535                 /* Always account for inflight pages, even if we didn't
 536                  * map them
 537                  */
 538                 goto skip_dma_unmap;
 539
 540         dma = page_pool_get_dma_addr(page);
 541
 542         /* When page is unmapped, it cannot be returned to our pool */
 543         dma_unmap_page_attrs(pool->p.dev, dma,
 544                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 545                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 546         page_pool_set_dma_addr(page, 0);
 547 skip_dma_unmap:
 548         page_pool_clear_pp_info(page);
 549
 550         /* This may be the last page returned, releasing the pool, so
 551          * it is not safe to reference pool afterwards.
 552          */
 553         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 554         trace_page_pool_state_release(pool, page, count);
 555
 556         put_page(page);
 557         /* An optimization would be to call __free_pages(page, pool->p.order)
 558          * knowing page is not part of page-cache (thus avoiding a
 559          * __page_cache_release() call).
 560          */
 561 }
 562
 563 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 564 {
 565         int ret;
 566         /* BH protection not needed if current is softirq */
 567         if (in_softirq())
 568                 ret = ptr_ring_produce(&pool->ring, page);
 569         else
 570                 ret = ptr_ring_produce_bh(&pool->ring, page);
 571
 572         if (!ret) {
 573                 recycle_stat_inc(pool, ring);
 574                 return true;
 575         }
 576
 577         return false;
 578 }
 579
 580 /* Only allow direct recycling in special circumstances, into the
 581  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 582  *
 583  * Caller must provide appropriate safe context.
 584  */
 585 static bool page_pool_recycle_in_cache(struct page *page,
 586                                        struct page_pool *pool)
 587 {
 588         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
 589                 recycle_stat_inc(pool, cache_full);
 590                 return false;
 591         }
 592
 593         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 594         pool->alloc.cache[pool->alloc.count++] = page;
 595         recycle_stat_inc(pool, cached);
 596         return true;
 597 }
 598
 599 /* If the page refcnt == 1, this will try to recycle the page.
 600  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 601  * the configured size min(dma_sync_size, pool->max_len).
 602  * If the page refcnt != 1, then the page will be returned to memory
 603  * subsystem.
 604  */
 605 static __always_inline struct page *
 606 __page_pool_put_page(struct page_pool *pool, struct page *page,
 607                      unsigned int dma_sync_size, bool allow_direct)
 608 {
 609         lockdep_assert_no_hardirq();
 610
 611         /* This allocator is optimized for the XDP mode that uses
 612          * one-frame-per-page, but have fallbacks that act like the
 613          * regular page allocator APIs.
 614          *
 615          * refcnt == 1 means page_pool owns page, and can recycle it.
 616          *
 617          * page is NOT reusable when allocated when system is under
 618          * some pressure. (page_is_pfmemalloc)
 619          */
 620         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 621                 /* Read barrier done in page_ref_count / READ_ONCE */
 622
 623                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 624                         page_pool_dma_sync_for_device(pool, page,
 625                                                       dma_sync_size);
 626
 627                 if (allow_direct && in_softirq() &&
 628                     page_pool_recycle_in_cache(page, pool))
 629                         return NULL;
 630
 631                 /* Page found as candidate for recycling */
 632                 return page;
 633         }
 634         /* Fallback/non-XDP mode: API user have elevated refcnt.
 635          *
 636          * Many drivers split up the page into fragments, and some
 637          * want to keep doing this to save memory and do refcnt based
 638          * recycling. Support this use case too, to ease drivers
 639          * switching between XDP/non-XDP.
 640          *
 641          * In-case page_pool maintains the DMA mapping, API user must
 642          * call page_pool_put_page once.  In this elevated refcnt
 643          * case, the DMA is unmapped/released, as driver is likely
 644          * doing refcnt based recycle tricks, meaning another process
 645          * will be invoking put_page.
 646          */
 647         recycle_stat_inc(pool, released_refcnt);
 648         page_pool_return_page(pool, page);
 649
 650         return NULL;
 651 }
 652
 653 void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
 654                                   unsigned int dma_sync_size, bool allow_direct)
 655 {
 656         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 657         if (page && !page_pool_recycle_in_ring(pool, page)) {
 658                 /* Cache full, fallback to free pages */
 659                 recycle_stat_inc(pool, ring_full);
 660                 page_pool_return_page(pool, page);
 661         }
 662 }
 663 EXPORT_SYMBOL(page_pool_put_defragged_page);
 664
 665 /**
 666  * page_pool_put_page_bulk() - release references on multiple pages
 667  * @pool:       pool from which pages were allocated
 668  * @data:       array holding page pointers
 669  * @count:      number of pages in @data
 670  *
 671  * Tries to refill a number of pages into the ptr_ring cache holding ptr_ring
 672  * producer lock. If the ptr_ring is full, page_pool_put_page_bulk()
 673  * will release leftover pages to the page allocator.
 674  * page_pool_put_page_bulk() is suitable to be run inside the driver NAPI tx
 675  * completion loop for the XDP_REDIRECT use case.
 676  *
 677  * Please note the caller must not use data area after running
 678  * page_pool_put_page_bulk(), as this function overwrites it.
 679  */
 680 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 681                              int count)
 682 {
 683         int i, bulk_len = 0;
 684         bool in_softirq;
 685
 686         for (i = 0; i < count; i++) {
 687                 struct page *page = virt_to_head_page(data[i]);
 688
 689                 /* It is not the last user for the page frag case */
 690                 if (!page_pool_is_last_frag(page))
 691                         continue;
 692
 693                 page = __page_pool_put_page(pool, page, -1, false);
 694                 /* Approved for bulk recycling in ptr_ring cache */
 695                 if (page)
 696                         data[bulk_len++] = page;
 697         }
 698
 699         if (unlikely(!bulk_len))
 700                 return;
 701
 702         /* Bulk producer into ptr_ring page_pool cache */
 703         in_softirq = page_pool_producer_lock(pool);
 704         for (i = 0; i < bulk_len; i++) {
 705                 if (__ptr_ring_produce(&pool->ring, data[i])) {
 706                         /* ring full */
 707                         recycle_stat_inc(pool, ring_full);
 708                         break;
 709                 }
 710         }
 711         recycle_stat_add(pool, ring, i);
 712         page_pool_producer_unlock(pool, in_softirq);
 713
 714         /* Hopefully all pages was return into ptr_ring */
 715         if (likely(i == bulk_len))
 716                 return;
 717
 718         /* ptr_ring cache full, free remaining pages outside producer lock
 719          * since put_page() with refcnt == 1 can be an expensive operation
 720          */
 721         for (; i < bulk_len; i++)
 722                 page_pool_return_page(pool, data[i]);
 723 }
 724 EXPORT_SYMBOL(page_pool_put_page_bulk);
 725
 726 static struct page *page_pool_drain_frag(struct page_pool *pool,
 727                                          struct page *page)
 728 {
 729         long drain_count = BIAS_MAX - pool->frag_users;
 730
 731         /* Some user is still using the page frag */
 732         if (likely(page_pool_defrag_page(page, drain_count)))
 733                 return NULL;
 734
 735         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 736                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 737                         page_pool_dma_sync_for_device(pool, page, -1);
 738
 739                 return page;
 740         }
 741
 742         page_pool_return_page(pool, page);
 743         return NULL;
 744 }
 745
 746 static void page_pool_free_frag(struct page_pool *pool)
 747 {
 748         long drain_count = BIAS_MAX - pool->frag_users;
 749         struct page *page = pool->frag_page;
 750
 751         pool->frag_page = NULL;
 752
 753         if (!page || page_pool_defrag_page(page, drain_count))
 754                 return;
 755
 756         page_pool_return_page(pool, page);
 757 }
 758
 759 struct page *page_pool_alloc_frag(struct page_pool *pool,
 760                                   unsigned int *offset,
 761                                   unsigned int size, gfp_t gfp)
 762 {
 763         unsigned int max_size = PAGE_SIZE << pool->p.order;
 764         struct page *page = pool->frag_page;
 765
 766         if (WARN_ON(size > max_size))
 767                 return NULL;
 768
 769         size = ALIGN(size, dma_get_cache_alignment());
 770         *offset = pool->frag_offset;
 771
 772         if (page && *offset + size > max_size) {
 773                 page = page_pool_drain_frag(pool, page);
 774                 if (page) {
 775                         alloc_stat_inc(pool, fast);
 776                         goto frag_reset;
 777                 }
 778         }
 779
 780         if (!page) {
 781                 page = page_pool_alloc_pages(pool, gfp);
 782                 if (unlikely(!page)) {
 783                         pool->frag_page = NULL;
 784                         return NULL;
 785                 }
 786
 787                 pool->frag_page = page;
 788
 789 frag_reset:
 790                 pool->frag_users = 1;
 791                 *offset = 0;
 792                 pool->frag_offset = size;
 793                 page_pool_fragment_page(page, BIAS_MAX);
 794                 return page;
 795         }
 796
 797         pool->frag_users++;
 798         pool->frag_offset = *offset + size;
 799         alloc_stat_inc(pool, fast);
 800         return page;
 801 }
 802 EXPORT_SYMBOL(page_pool_alloc_frag);
 803
 804 static void page_pool_empty_ring(struct page_pool *pool)
 805 {
 806         struct page *page;
 807
 808         /* Empty recycle ring */
 809         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 810                 /* Verify the refcnt invariant of cached pages */
 811                 if (!(page_ref_count(page) == 1))
 812                         pr_crit("%s() page_pool refcnt %d violation\n",
 813                                 __func__, page_ref_count(page));
 814
 815                 page_pool_return_page(pool, page);
 816         }
 817 }
 818
 819 static void __page_pool_destroy(struct page_pool *pool)
 820 {
 821         if (pool->disconnect)
 822                 pool->disconnect(pool);
 823
 824         ptr_ring_cleanup(&pool->ring, NULL);
 825
 826         if (pool->p.flags & PP_FLAG_DMA_MAP)
 827                 put_device(pool->p.dev);
 828
 829 #ifdef CONFIG_PAGE_POOL_STATS
 830         free_percpu(pool->recycle_stats);
 831 #endif
 832         kfree(pool);
 833 }
 834
 835 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 836 {
 837         struct page *page;
 838
 839         if (pool->destroy_cnt)
 840                 return;
 841
 842         /* Empty alloc cache, assume caller made sure this is
 843          * no-longer in use, and page_pool_alloc_pages() cannot be
 844          * call concurrently.
 845          */
 846         while (pool->alloc.count) {
 847                 page = pool->alloc.cache[--pool->alloc.count];
 848                 page_pool_return_page(pool, page);
 849         }
 850 }
 851
 852 static void page_pool_scrub(struct page_pool *pool)
 853 {
 854         page_pool_empty_alloc_cache_once(pool);
 855         pool->destroy_cnt++;
 856
 857         /* No more consumers should exist, but producers could still
 858          * be in-flight.
 859          */
 860         page_pool_empty_ring(pool);
 861 }
 862
 863 static int page_pool_release(struct page_pool *pool)
 864 {
 865         int inflight;
 866
 867         page_pool_scrub(pool);
 868         inflight = page_pool_inflight(pool);
 869         if (!inflight)
 870                 __page_pool_destroy(pool);
 871
 872         return inflight;
 873 }
 874
 875 static void page_pool_release_retry(struct work_struct *wq)
 876 {
 877         struct delayed_work *dwq = to_delayed_work(wq);
 878         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 879         int inflight;
 880
 881         inflight = page_pool_release(pool);
 882         if (!inflight)
 883                 return;
 884
 885         /* Periodic warning */
 886         if (time_after_eq(jiffies, pool->defer_warn)) {
 887                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 888
 889                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 890                         __func__, inflight, sec);
 891                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 892         }
 893
 894         /* Still not ready to be disconnected, retry later */
 895         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 896 }
 897
 898 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
 899                            struct xdp_mem_info *mem)
 900 {
 901         refcount_inc(&pool->user_cnt);
 902         pool->disconnect = disconnect;
 903         pool->xdp_mem_id = mem->id;
 904 }
 905
 906 void page_pool_unlink_napi(struct page_pool *pool)
 907 {
 908         if (!pool->p.napi)
 909                 return;
 910
 911         /* To avoid races with recycling and additional barriers make sure
 912          * pool and NAPI are unlinked when NAPI is disabled.
 913          */
 914         WARN_ON(!test_bit(NAPI_STATE_SCHED, &pool->p.napi->state) ||
 915                 READ_ONCE(pool->p.napi->list_owner) != -1);
 916
 917         WRITE_ONCE(pool->p.napi, NULL);
 918 }
 919 EXPORT_SYMBOL(page_pool_unlink_napi);
 920
 921 void page_pool_destroy(struct page_pool *pool)
 922 {
 923         if (!pool)
 924                 return;
 925
 926         if (!page_pool_put(pool))
 927                 return;
 928
 929         page_pool_unlink_napi(pool);
 930         page_pool_free_frag(pool);
 931
 932         if (!page_pool_release(pool))
 933                 return;
 934
 935         pool->defer_start = jiffies;
 936         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 937
 938         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 939         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 940 }
 941 EXPORT_SYMBOL(page_pool_destroy);
 942
 943 /* Caller must provide appropriate safe context, e.g. NAPI. */
 944 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 945 {
 946         struct page *page;
 947
 948         trace_page_pool_update_nid(pool, new_nid);
 949         pool->p.nid = new_nid;
 950
 951         /* Flush pool alloc cache, as refill will check NUMA node */
 952         while (pool->alloc.count) {
 953                 page = pool->alloc.cache[--pool->alloc.count];
 954                 page_pool_return_page(pool, page);
 955         }
 956 }
 957 EXPORT_SYMBOL(page_pool_update_nid);