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