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