net/core/page_pool.c

   1 /* SPDX-License-Identifier: GPL-2.0
   2  *
   3  * page_pool.c
   4  *      Author: Jesper Dangaard Brouer <netoptimizer@brouer.com>
   5  *      Copyright (C) 2016 Red Hat, Inc.
   6  */
   7
   8 #include <linux/types.h>
   9 #include <linux/kernel.h>
  10 #include <linux/slab.h>
  11 #include <linux/device.h>
  12
  13 #include <net/page_pool.h>
  14 #include <net/xdp.h>
  15
  16 #include <linux/dma-direction.h>
  17 #include <linux/dma-mapping.h>
  18 #include <linux/page-flags.h>
  19 #include <linux/mm.h> /* for __put_page() */
  20 #include <linux/poison.h>
  21 #include <linux/ethtool.h>
  22
  23 #include <trace/events/page_pool.h>
  24
  25 #define DEFER_TIME (msecs_to_jiffies(1000))
  26 #define DEFER_WARN_INTERVAL (60 * HZ)
  27
  28 #define BIAS_MAX        LONG_MAX
  29
  30 #ifdef CONFIG_PAGE_POOL_STATS
  31 /* alloc_stat_inc is intended to be used in softirq context */
  32 #define alloc_stat_inc(pool, __stat)    (pool->alloc_stats.__stat++)
  33 /* recycle_stat_inc is safe to use when preemption is possible. */
  34 #define recycle_stat_inc(pool, __stat)                                                  \
  35         do {                                                                            \
  36                 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats;       \
  37                 this_cpu_inc(s->__stat);                                                \
  38         } while (0)
  39
  40 #define recycle_stat_add(pool, __stat, val)                                             \
  41         do {                                                                            \
  42                 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats;       \
  43                 this_cpu_add(s->__stat, val);                                           \
  44         } while (0)
  45
  46 static const char pp_stats[][ETH_GSTRING_LEN] = {
  47         "rx_pp_alloc_fast",
  48         "rx_pp_alloc_slow",
  49         "rx_pp_alloc_slow_ho",
  50         "rx_pp_alloc_empty",
  51         "rx_pp_alloc_refill",
  52         "rx_pp_alloc_waive",
  53         "rx_pp_recycle_cached",
  54         "rx_pp_recycle_cache_full",
  55         "rx_pp_recycle_ring",
  56         "rx_pp_recycle_ring_full",
  57         "rx_pp_recycle_released_ref",
  58 };
  59
  60 bool page_pool_get_stats(struct page_pool *pool,
  61                          struct page_pool_stats *stats)
  62 {
  63         int cpu = 0;
  64
  65         if (!stats)
  66                 return false;
  67
  68         /* The caller is responsible to initialize stats. */
  69         stats->alloc_stats.fast += pool->alloc_stats.fast;
  70         stats->alloc_stats.slow += pool->alloc_stats.slow;
  71         stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
  72         stats->alloc_stats.empty += pool->alloc_stats.empty;
  73         stats->alloc_stats.refill += pool->alloc_stats.refill;
  74         stats->alloc_stats.waive += pool->alloc_stats.waive;
  75
  76         for_each_possible_cpu(cpu) {
  77                 const struct page_pool_recycle_stats *pcpu =
  78                         per_cpu_ptr(pool->recycle_stats, cpu);
  79
  80                 stats->recycle_stats.cached += pcpu->cached;
  81                 stats->recycle_stats.cache_full += pcpu->cache_full;
  82                 stats->recycle_stats.ring += pcpu->ring;
  83                 stats->recycle_stats.ring_full += pcpu->ring_full;
  84                 stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
  85         }
  86
  87         return true;
  88 }
  89 EXPORT_SYMBOL(page_pool_get_stats);
  90
  91 u8 *page_pool_ethtool_stats_get_strings(u8 *data)
  92 {
  93         int i;
  94
  95         for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
  96                 memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
  97                 data += ETH_GSTRING_LEN;
  98         }
  99
 100         return data;
 101 }
 102 EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);
 103
 104 int page_pool_ethtool_stats_get_count(void)
 105 {
 106         return ARRAY_SIZE(pp_stats);
 107 }
 108 EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);
 109
 110 u64 *page_pool_ethtool_stats_get(u64 *data, void *stats)
 111 {
 112         struct page_pool_stats *pool_stats = stats;
 113
 114         *data++ = pool_stats->alloc_stats.fast;
 115         *data++ = pool_stats->alloc_stats.slow;
 116         *data++ = pool_stats->alloc_stats.slow_high_order;
 117         *data++ = pool_stats->alloc_stats.empty;
 118         *data++ = pool_stats->alloc_stats.refill;
 119         *data++ = pool_stats->alloc_stats.waive;
 120         *data++ = pool_stats->recycle_stats.cached;
 121         *data++ = pool_stats->recycle_stats.cache_full;
 122         *data++ = pool_stats->recycle_stats.ring;
 123         *data++ = pool_stats->recycle_stats.ring_full;
 124         *data++ = pool_stats->recycle_stats.released_refcnt;
 125
 126         return data;
 127 }
 128 EXPORT_SYMBOL(page_pool_ethtool_stats_get);
 129
 130 #else
 131 #define alloc_stat_inc(pool, __stat)
 132 #define recycle_stat_inc(pool, __stat)
 133 #define recycle_stat_add(pool, __stat, val)
 134 #endif
 135
 136 static int page_pool_init(struct page_pool *pool,
 137                           const struct page_pool_params *params)
 138 {
 139         unsigned int ring_qsize = 1024; /* Default */
 140
 141         memcpy(&pool->p, params, sizeof(pool->p));
 142
 143         /* Validate only known flags were used */
 144         if (pool->p.flags & ~(PP_FLAG_ALL))
 145                 return -EINVAL;
 146
 147         if (pool->p.pool_size)
 148                 ring_qsize = pool->p.pool_size;
 149
 150         /* Sanity limit mem that can be pinned down */
 151         if (ring_qsize > 32768)
 152                 return -E2BIG;
 153
 154         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
 155          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
 156          * which is the XDP_TX use-case.
 157          */
 158         if (pool->p.flags & PP_FLAG_DMA_MAP) {
 159                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
 160                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
 161                         return -EINVAL;
 162         }
 163
 164         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
 165                 /* In order to request DMA-sync-for-device the page
 166                  * needs to be mapped
 167                  */
 168                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 169                         return -EINVAL;
 170
 171                 if (!pool->p.max_len)
 172                         return -EINVAL;
 173
 174                 /* pool->p.offset has to be set according to the address
 175                  * offset used by the DMA engine to start copying rx data
 176                  */
 177         }
 178
 179         if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT &&
 180             pool->p.flags & PP_FLAG_PAGE_FRAG)
 181                 return -EINVAL;
 182
 183 #ifdef CONFIG_PAGE_POOL_STATS
 184         pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
 185         if (!pool->recycle_stats)
 186                 return -ENOMEM;
 187 #endif
 188
 189         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
 190                 return -ENOMEM;
 191
 192         atomic_set(&pool->pages_state_release_cnt, 0);
 193
 194         /* Driver calling page_pool_create() also call page_pool_destroy() */
 195         refcount_set(&pool->user_cnt, 1);
 196
 197         if (pool->p.flags & PP_FLAG_DMA_MAP)
 198                 get_device(pool->p.dev);
 199
 200         return 0;
 201 }
 202
 203 struct page_pool *page_pool_create(const struct page_pool_params *params)
 204 {
 205         struct page_pool *pool;
 206         int err;
 207
 208         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
 209         if (!pool)
 210                 return ERR_PTR(-ENOMEM);
 211
 212         err = page_pool_init(pool, params);
 213         if (err < 0) {
 214                 pr_warn("%s() gave up with errno %d\n", __func__, err);
 215                 kfree(pool);
 216                 return ERR_PTR(err);
 217         }
 218
 219         return pool;
 220 }
 221 EXPORT_SYMBOL(page_pool_create);
 222
 223 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 224
 225 noinline
 226 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 227 {
 228         struct ptr_ring *r = &pool->ring;
 229         struct page *page;
 230         int pref_nid; /* preferred NUMA node */
 231
 232         /* Quicker fallback, avoid locks when ring is empty */
 233         if (__ptr_ring_empty(r)) {
 234                 alloc_stat_inc(pool, empty);
 235                 return NULL;
 236         }
 237
 238         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 239          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 240          */
 241 #ifdef CONFIG_NUMA
 242         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 243 #else
 244         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 245         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 246 #endif
 247
 248         /* Refill alloc array, but only if NUMA match */
 249         do {
 250                 page = __ptr_ring_consume(r);
 251                 if (unlikely(!page))
 252                         break;
 253
 254                 if (likely(page_to_nid(page) == pref_nid)) {
 255                         pool->alloc.cache[pool->alloc.count++] = page;
 256                 } else {
 257                         /* NUMA mismatch;
 258                          * (1) release 1 page to page-allocator and
 259                          * (2) break out to fallthrough to alloc_pages_node.
 260                          * This limit stress on page buddy alloactor.
 261                          */
 262                         page_pool_return_page(pool, page);
 263                         alloc_stat_inc(pool, waive);
 264                         page = NULL;
 265                         break;
 266                 }
 267         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 268
 269         /* Return last page */
 270         if (likely(pool->alloc.count > 0)) {
 271                 page = pool->alloc.cache[--pool->alloc.count];
 272                 alloc_stat_inc(pool, refill);
 273         }
 274
 275         return page;
 276 }
 277
 278 /* fast path */
 279 static struct page *__page_pool_get_cached(struct page_pool *pool)
 280 {
 281         struct page *page;
 282
 283         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 284         if (likely(pool->alloc.count)) {
 285                 /* Fast-path */
 286                 page = pool->alloc.cache[--pool->alloc.count];
 287                 alloc_stat_inc(pool, fast);
 288         } else {
 289                 page = page_pool_refill_alloc_cache(pool);
 290         }
 291
 292         return page;
 293 }
 294
 295 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 296                                           struct page *page,
 297                                           unsigned int dma_sync_size)
 298 {
 299         dma_addr_t dma_addr = page_pool_get_dma_addr(page);
 300
 301         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 302         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 303                                          pool->p.offset, dma_sync_size,
 304                                          pool->p.dma_dir);
 305 }
 306
 307 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 308 {
 309         dma_addr_t dma;
 310
 311         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 312          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 313          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 314          * This mapping is kept for lifetime of page, until leaving pool.
 315          */
 316         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 317                                  (PAGE_SIZE << pool->p.order),
 318                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 319         if (dma_mapping_error(pool->p.dev, dma))
 320                 return false;
 321
 322         page_pool_set_dma_addr(page, dma);
 323
 324         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 325                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 326
 327         return true;
 328 }
 329
 330 static void page_pool_set_pp_info(struct page_pool *pool,
 331                                   struct page *page)
 332 {
 333         page->pp = pool;
 334         page->pp_magic |= PP_SIGNATURE;
 335         if (pool->p.init_callback)
 336                 pool->p.init_callback(page, pool->p.init_arg);
 337 }
 338
 339 static void page_pool_clear_pp_info(struct page *page)
 340 {
 341         page->pp_magic = 0;
 342         page->pp = NULL;
 343 }
 344
 345 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 346                                                  gfp_t gfp)
 347 {
 348         struct page *page;
 349
 350         gfp |= __GFP_COMP;
 351         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 352         if (unlikely(!page))
 353                 return NULL;
 354
 355         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 356             unlikely(!page_pool_dma_map(pool, page))) {
 357                 put_page(page);
 358                 return NULL;
 359         }
 360
 361         alloc_stat_inc(pool, slow_high_order);
 362         page_pool_set_pp_info(pool, page);
 363
 364         /* Track how many pages are held 'in-flight' */
 365         pool->pages_state_hold_cnt++;
 366         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 367         return page;
 368 }
 369
 370 /* slow path */
 371 noinline
 372 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 373                                                  gfp_t gfp)
 374 {
 375         const int bulk = PP_ALLOC_CACHE_REFILL;
 376         unsigned int pp_flags = pool->p.flags;
 377         unsigned int pp_order = pool->p.order;
 378         struct page *page;
 379         int i, nr_pages;
 380
 381         /* Don't support bulk alloc for high-order pages */
 382         if (unlikely(pp_order))
 383                 return __page_pool_alloc_page_order(pool, gfp);
 384
 385         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 386         if (unlikely(pool->alloc.count > 0))
 387                 return pool->alloc.cache[--pool->alloc.count];
 388
 389         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 390         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 391
 392         nr_pages = alloc_pages_bulk_array(gfp, bulk, pool->alloc.cache);
 393         if (unlikely(!nr_pages))
 394                 return NULL;
 395
 396         /* Pages have been filled into alloc.cache array, but count is zero and
 397          * page element have not been (possibly) DMA mapped.
 398          */
 399         for (i = 0; i < nr_pages; i++) {
 400                 page = pool->alloc.cache[i];
 401                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 402                     unlikely(!page_pool_dma_map(pool, page))) {
 403                         put_page(page);
 404                         continue;
 405                 }
 406
 407                 page_pool_set_pp_info(pool, page);
 408                 pool->alloc.cache[pool->alloc.count++] = page;
 409                 /* Track how many pages are held 'in-flight' */
 410                 pool->pages_state_hold_cnt++;
 411                 trace_page_pool_state_hold(pool, page,
 412                                            pool->pages_state_hold_cnt);
 413         }
 414
 415         /* Return last page */
 416         if (likely(pool->alloc.count > 0)) {
 417                 page = pool->alloc.cache[--pool->alloc.count];
 418                 alloc_stat_inc(pool, slow);
 419         } else {
 420                 page = NULL;
 421         }
 422
 423         /* When page just alloc'ed is should/must have refcnt 1. */
 424         return page;
 425 }
 426
 427 /* For using page_pool replace: alloc_pages() API calls, but provide
 428  * synchronization guarantee for allocation side.
 429  */
 430 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 431 {
 432         struct page *page;
 433
 434         /* Fast-path: Get a page from cache */
 435         page = __page_pool_get_cached(pool);
 436         if (page)
 437                 return page;
 438
 439         /* Slow-path: cache empty, do real allocation */
 440         page = __page_pool_alloc_pages_slow(pool, gfp);
 441         return page;
 442 }
 443 EXPORT_SYMBOL(page_pool_alloc_pages);
 444
 445 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 446  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 447  */
 448 #define _distance(a, b) (s32)((a) - (b))
 449
 450 static s32 page_pool_inflight(struct page_pool *pool)
 451 {
 452         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 453         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 454         s32 inflight;
 455
 456         inflight = _distance(hold_cnt, release_cnt);
 457
 458         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 459         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 460
 461         return inflight;
 462 }
 463
 464 /* Disconnects a page (from a page_pool).  API users can have a need
 465  * to disconnect a page (from a page_pool), to allow it to be used as
 466  * a regular page (that will eventually be returned to the normal
 467  * page-allocator via put_page).
 468  */
 469 void page_pool_release_page(struct page_pool *pool, struct page *page)
 470 {
 471         dma_addr_t dma;
 472         int count;
 473
 474         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 475                 /* Always account for inflight pages, even if we didn't
 476                  * map them
 477                  */
 478                 goto skip_dma_unmap;
 479
 480         dma = page_pool_get_dma_addr(page);
 481
 482         /* When page is unmapped, it cannot be returned to our pool */
 483         dma_unmap_page_attrs(pool->p.dev, dma,
 484                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 485                              DMA_ATTR_SKIP_CPU_SYNC);
 486         page_pool_set_dma_addr(page, 0);
 487 skip_dma_unmap:
 488         page_pool_clear_pp_info(page);
 489
 490         /* This may be the last page returned, releasing the pool, so
 491          * it is not safe to reference pool afterwards.
 492          */
 493         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 494         trace_page_pool_state_release(pool, page, count);
 495 }
 496 EXPORT_SYMBOL(page_pool_release_page);
 497
 498 /* Return a page to the page allocator, cleaning up our state */
 499 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 500 {
 501         page_pool_release_page(pool, page);
 502
 503         put_page(page);
 504         /* An optimization would be to call __free_pages(page, pool->p.order)
 505          * knowing page is not part of page-cache (thus avoiding a
 506          * __page_cache_release() call).
 507          */
 508 }
 509
 510 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 511 {
 512         int ret;
 513         /* BH protection not needed if current is serving softirq */
 514         if (in_serving_softirq())
 515                 ret = ptr_ring_produce(&pool->ring, page);
 516         else
 517                 ret = ptr_ring_produce_bh(&pool->ring, page);
 518
 519         if (!ret) {
 520                 recycle_stat_inc(pool, ring);
 521                 return true;
 522         }
 523
 524         return false;
 525 }
 526
 527 /* Only allow direct recycling in special circumstances, into the
 528  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 529  *
 530  * Caller must provide appropriate safe context.
 531  */
 532 static bool page_pool_recycle_in_cache(struct page *page,
 533                                        struct page_pool *pool)
 534 {
 535         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
 536                 recycle_stat_inc(pool, cache_full);
 537                 return false;
 538         }
 539
 540         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 541         pool->alloc.cache[pool->alloc.count++] = page;
 542         recycle_stat_inc(pool, cached);
 543         return true;
 544 }
 545
 546 /* If the page refcnt == 1, this will try to recycle the page.
 547  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 548  * the configured size min(dma_sync_size, pool->max_len).
 549  * If the page refcnt != 1, then the page will be returned to memory
 550  * subsystem.
 551  */
 552 static __always_inline struct page *
 553 __page_pool_put_page(struct page_pool *pool, struct page *page,
 554                      unsigned int dma_sync_size, bool allow_direct)
 555 {
 556         /* This allocator is optimized for the XDP mode that uses
 557          * one-frame-per-page, but have fallbacks that act like the
 558          * regular page allocator APIs.
 559          *
 560          * refcnt == 1 means page_pool owns page, and can recycle it.
 561          *
 562          * page is NOT reusable when allocated when system is under
 563          * some pressure. (page_is_pfmemalloc)
 564          */
 565         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 566                 /* Read barrier done in page_ref_count / READ_ONCE */
 567
 568                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 569                         page_pool_dma_sync_for_device(pool, page,
 570                                                       dma_sync_size);
 571
 572                 if (allow_direct && in_serving_softirq() &&
 573                     page_pool_recycle_in_cache(page, pool))
 574                         return NULL;
 575
 576                 /* Page found as candidate for recycling */
 577                 return page;
 578         }
 579         /* Fallback/non-XDP mode: API user have elevated refcnt.
 580          *
 581          * Many drivers split up the page into fragments, and some
 582          * want to keep doing this to save memory and do refcnt based
 583          * recycling. Support this use case too, to ease drivers
 584          * switching between XDP/non-XDP.
 585          *
 586          * In-case page_pool maintains the DMA mapping, API user must
 587          * call page_pool_put_page once.  In this elevated refcnt
 588          * case, the DMA is unmapped/released, as driver is likely
 589          * doing refcnt based recycle tricks, meaning another process
 590          * will be invoking put_page.
 591          */
 592         recycle_stat_inc(pool, released_refcnt);
 593         /* Do not replace this with page_pool_return_page() */
 594         page_pool_release_page(pool, page);
 595         put_page(page);
 596
 597         return NULL;
 598 }
 599
 600 void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
 601                                   unsigned int dma_sync_size, bool allow_direct)
 602 {
 603         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 604         if (page && !page_pool_recycle_in_ring(pool, page)) {
 605                 /* Cache full, fallback to free pages */
 606                 recycle_stat_inc(pool, ring_full);
 607                 page_pool_return_page(pool, page);
 608         }
 609 }
 610 EXPORT_SYMBOL(page_pool_put_defragged_page);
 611
 612 /* Caller must not use data area after call, as this function overwrites it */
 613 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 614                              int count)
 615 {
 616         int i, bulk_len = 0;
 617
 618         for (i = 0; i < count; i++) {
 619                 struct page *page = virt_to_head_page(data[i]);
 620
 621                 /* It is not the last user for the page frag case */
 622                 if (!page_pool_is_last_frag(pool, page))
 623                         continue;
 624
 625                 page = __page_pool_put_page(pool, page, -1, false);
 626                 /* Approved for bulk recycling in ptr_ring cache */
 627                 if (page)
 628                         data[bulk_len++] = page;
 629         }
 630
 631         if (unlikely(!bulk_len))
 632                 return;
 633
 634         /* Bulk producer into ptr_ring page_pool cache */
 635         page_pool_ring_lock(pool);
 636         for (i = 0; i < bulk_len; i++) {
 637                 if (__ptr_ring_produce(&pool->ring, data[i])) {
 638                         /* ring full */
 639                         recycle_stat_inc(pool, ring_full);
 640                         break;
 641                 }
 642         }
 643         recycle_stat_add(pool, ring, i);
 644         page_pool_ring_unlock(pool);
 645
 646         /* Hopefully all pages was return into ptr_ring */
 647         if (likely(i == bulk_len))
 648                 return;
 649
 650         /* ptr_ring cache full, free remaining pages outside producer lock
 651          * since put_page() with refcnt == 1 can be an expensive operation
 652          */
 653         for (; i < bulk_len; i++)
 654                 page_pool_return_page(pool, data[i]);
 655 }
 656 EXPORT_SYMBOL(page_pool_put_page_bulk);
 657
 658 static struct page *page_pool_drain_frag(struct page_pool *pool,
 659                                          struct page *page)
 660 {
 661         long drain_count = BIAS_MAX - pool->frag_users;
 662
 663         /* Some user is still using the page frag */
 664         if (likely(page_pool_defrag_page(page, drain_count)))
 665                 return NULL;
 666
 667         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 668                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 669                         page_pool_dma_sync_for_device(pool, page, -1);
 670
 671                 return page;
 672         }
 673
 674         page_pool_return_page(pool, page);
 675         return NULL;
 676 }
 677
 678 static void page_pool_free_frag(struct page_pool *pool)
 679 {
 680         long drain_count = BIAS_MAX - pool->frag_users;
 681         struct page *page = pool->frag_page;
 682
 683         pool->frag_page = NULL;
 684
 685         if (!page || page_pool_defrag_page(page, drain_count))
 686                 return;
 687
 688         page_pool_return_page(pool, page);
 689 }
 690
 691 struct page *page_pool_alloc_frag(struct page_pool *pool,
 692                                   unsigned int *offset,
 693                                   unsigned int size, gfp_t gfp)
 694 {
 695         unsigned int max_size = PAGE_SIZE << pool->p.order;
 696         struct page *page = pool->frag_page;
 697
 698         if (WARN_ON(!(pool->p.flags & PP_FLAG_PAGE_FRAG) ||
 699                     size > max_size))
 700                 return NULL;
 701
 702         size = ALIGN(size, dma_get_cache_alignment());
 703         *offset = pool->frag_offset;
 704
 705         if (page && *offset + size > max_size) {
 706                 page = page_pool_drain_frag(pool, page);
 707                 if (page) {
 708                         alloc_stat_inc(pool, fast);
 709                         goto frag_reset;
 710                 }
 711         }
 712
 713         if (!page) {
 714                 page = page_pool_alloc_pages(pool, gfp);
 715                 if (unlikely(!page)) {
 716                         pool->frag_page = NULL;
 717                         return NULL;
 718                 }
 719
 720                 pool->frag_page = page;
 721
 722 frag_reset:
 723                 pool->frag_users = 1;
 724                 *offset = 0;
 725                 pool->frag_offset = size;
 726                 page_pool_fragment_page(page, BIAS_MAX);
 727                 return page;
 728         }
 729
 730         pool->frag_users++;
 731         pool->frag_offset = *offset + size;
 732         alloc_stat_inc(pool, fast);
 733         return page;
 734 }
 735 EXPORT_SYMBOL(page_pool_alloc_frag);
 736
 737 static void page_pool_empty_ring(struct page_pool *pool)
 738 {
 739         struct page *page;
 740
 741         /* Empty recycle ring */
 742         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 743                 /* Verify the refcnt invariant of cached pages */
 744                 if (!(page_ref_count(page) == 1))
 745                         pr_crit("%s() page_pool refcnt %d violation\n",
 746                                 __func__, page_ref_count(page));
 747
 748                 page_pool_return_page(pool, page);
 749         }
 750 }
 751
 752 static void page_pool_free(struct page_pool *pool)
 753 {
 754         if (pool->disconnect)
 755                 pool->disconnect(pool);
 756
 757         ptr_ring_cleanup(&pool->ring, NULL);
 758
 759         if (pool->p.flags & PP_FLAG_DMA_MAP)
 760                 put_device(pool->p.dev);
 761
 762 #ifdef CONFIG_PAGE_POOL_STATS
 763         free_percpu(pool->recycle_stats);
 764 #endif
 765         kfree(pool);
 766 }
 767
 768 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 769 {
 770         struct page *page;
 771
 772         if (pool->destroy_cnt)
 773                 return;
 774
 775         /* Empty alloc cache, assume caller made sure this is
 776          * no-longer in use, and page_pool_alloc_pages() cannot be
 777          * call concurrently.
 778          */
 779         while (pool->alloc.count) {
 780                 page = pool->alloc.cache[--pool->alloc.count];
 781                 page_pool_return_page(pool, page);
 782         }
 783 }
 784
 785 static void page_pool_scrub(struct page_pool *pool)
 786 {
 787         page_pool_empty_alloc_cache_once(pool);
 788         pool->destroy_cnt++;
 789
 790         /* No more consumers should exist, but producers could still
 791          * be in-flight.
 792          */
 793         page_pool_empty_ring(pool);
 794 }
 795
 796 static int page_pool_release(struct page_pool *pool)
 797 {
 798         int inflight;
 799
 800         page_pool_scrub(pool);
 801         inflight = page_pool_inflight(pool);
 802         if (!inflight)
 803                 page_pool_free(pool);
 804
 805         return inflight;
 806 }
 807
 808 static void page_pool_release_retry(struct work_struct *wq)
 809 {
 810         struct delayed_work *dwq = to_delayed_work(wq);
 811         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 812         int inflight;
 813
 814         inflight = page_pool_release(pool);
 815         if (!inflight)
 816                 return;
 817
 818         /* Periodic warning */
 819         if (time_after_eq(jiffies, pool->defer_warn)) {
 820                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 821
 822                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 823                         __func__, inflight, sec);
 824                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 825         }
 826
 827         /* Still not ready to be disconnected, retry later */
 828         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 829 }
 830
 831 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
 832                            struct xdp_mem_info *mem)
 833 {
 834         refcount_inc(&pool->user_cnt);
 835         pool->disconnect = disconnect;
 836         pool->xdp_mem_id = mem->id;
 837 }
 838
 839 void page_pool_destroy(struct page_pool *pool)
 840 {
 841         if (!pool)
 842                 return;
 843
 844         if (!page_pool_put(pool))
 845                 return;
 846
 847         page_pool_free_frag(pool);
 848
 849         if (!page_pool_release(pool))
 850                 return;
 851
 852         pool->defer_start = jiffies;
 853         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 854
 855         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 856         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 857 }
 858 EXPORT_SYMBOL(page_pool_destroy);
 859
 860 /* Caller must provide appropriate safe context, e.g. NAPI. */
 861 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 862 {
 863         struct page *page;
 864
 865         trace_page_pool_update_nid(pool, new_nid);
 866         pool->p.nid = new_nid;
 867
 868         /* Flush pool alloc cache, as refill will check NUMA node */
 869         while (pool->alloc.count) {
 870                 page = pool->alloc.cache[--pool->alloc.count];
 871                 page_pool_return_page(pool, page);
 872         }
 873 }
 874 EXPORT_SYMBOL(page_pool_update_nid);
 875
 876 bool page_pool_return_skb_page(struct page *page)
 877 {
 878         struct page_pool *pp;
 879
 880         page = compound_head(page);
 881
 882         /* page->pp_magic is OR'ed with PP_SIGNATURE after the allocation
 883          * in order to preserve any existing bits, such as bit 0 for the
 884          * head page of compound page and bit 1 for pfmemalloc page, so
 885          * mask those bits for freeing side when doing below checking,
 886          * and page_is_pfmemalloc() is checked in __page_pool_put_page()
 887          * to avoid recycling the pfmemalloc page.
 888          */
 889         if (unlikely((page->pp_magic & ~0x3UL) != PP_SIGNATURE))
 890                 return false;
 891
 892         pp = page->pp;
 893
 894         /* Driver set this to memory recycling info. Reset it on recycle.
 895          * This will *not* work for NIC using a split-page memory model.
 896          * The page will be returned to the pool here regardless of the
 897          * 'flipped' fragment being in use or not.
 898          */
 899         page_pool_put_full_page(pp, page, false);
 900
 901         return true;
 902 }
 903 EXPORT_SYMBOL(page_pool_return_skb_page);