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 #ifdef CONFIG_PAGE_POOL_STATS
 222                 free_percpu(pool->recycle_stats);
 223 #endif
 224                 return -ENOMEM;
 225         }
 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         if (page_pool_set_dma_addr(page, dma))
 363                 goto unmap_failed;
 364
 365         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 366                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 367
 368         return true;
 369
 370 unmap_failed:
 371         WARN_ON_ONCE("unexpected DMA address, please report to netdev@");
 372         dma_unmap_page_attrs(pool->p.dev, dma,
 373                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 374                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 375         return false;
 376 }
 377
 378 static void page_pool_set_pp_info(struct page_pool *pool,
 379                                   struct page *page)
 380 {
 381         page->pp = pool;
 382         page->pp_magic |= PP_SIGNATURE;
 383
 384         /* Ensuring all pages have been split into one fragment initially:
 385          * page_pool_set_pp_info() is only called once for every page when it
 386          * is allocated from the page allocator and page_pool_fragment_page()
 387          * is dirtying the same cache line as the page->pp_magic above, so
 388          * the overhead is negligible.
 389          */
 390         page_pool_fragment_page(page, 1);
 391         if (pool->p.init_callback)
 392                 pool->p.init_callback(page, pool->p.init_arg);
 393 }
 394
 395 static void page_pool_clear_pp_info(struct page *page)
 396 {
 397         page->pp_magic = 0;
 398         page->pp = NULL;
 399 }
 400
 401 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 402                                                  gfp_t gfp)
 403 {
 404         struct page *page;
 405
 406         gfp |= __GFP_COMP;
 407         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 408         if (unlikely(!page))
 409                 return NULL;
 410
 411         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 412             unlikely(!page_pool_dma_map(pool, page))) {
 413                 put_page(page);
 414                 return NULL;
 415         }
 416
 417         alloc_stat_inc(pool, slow_high_order);
 418         page_pool_set_pp_info(pool, page);
 419
 420         /* Track how many pages are held 'in-flight' */
 421         pool->pages_state_hold_cnt++;
 422         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 423         return page;
 424 }
 425
 426 /* slow path */
 427 noinline
 428 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 429                                                  gfp_t gfp)
 430 {
 431         const int bulk = PP_ALLOC_CACHE_REFILL;
 432         unsigned int pp_flags = pool->p.flags;
 433         unsigned int pp_order = pool->p.order;
 434         struct page *page;
 435         int i, nr_pages;
 436
 437         /* Don't support bulk alloc for high-order pages */
 438         if (unlikely(pp_order))
 439                 return __page_pool_alloc_page_order(pool, gfp);
 440
 441         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 442         if (unlikely(pool->alloc.count > 0))
 443                 return pool->alloc.cache[--pool->alloc.count];
 444
 445         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 446         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 447
 448         nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk,
 449                                                pool->alloc.cache);
 450         if (unlikely(!nr_pages))
 451                 return NULL;
 452
 453         /* Pages have been filled into alloc.cache array, but count is zero and
 454          * page element have not been (possibly) DMA mapped.
 455          */
 456         for (i = 0; i < nr_pages; i++) {
 457                 page = pool->alloc.cache[i];
 458                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 459                     unlikely(!page_pool_dma_map(pool, page))) {
 460                         put_page(page);
 461                         continue;
 462                 }
 463
 464                 page_pool_set_pp_info(pool, page);
 465                 pool->alloc.cache[pool->alloc.count++] = page;
 466                 /* Track how many pages are held 'in-flight' */
 467                 pool->pages_state_hold_cnt++;
 468                 trace_page_pool_state_hold(pool, page,
 469                                            pool->pages_state_hold_cnt);
 470         }
 471
 472         /* Return last page */
 473         if (likely(pool->alloc.count > 0)) {
 474                 page = pool->alloc.cache[--pool->alloc.count];
 475                 alloc_stat_inc(pool, slow);
 476         } else {
 477                 page = NULL;
 478         }
 479
 480         /* When page just alloc'ed is should/must have refcnt 1. */
 481         return page;
 482 }
 483
 484 /* For using page_pool replace: alloc_pages() API calls, but provide
 485  * synchronization guarantee for allocation side.
 486  */
 487 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 488 {
 489         struct page *page;
 490
 491         /* Fast-path: Get a page from cache */
 492         page = __page_pool_get_cached(pool);
 493         if (page)
 494                 return page;
 495
 496         /* Slow-path: cache empty, do real allocation */
 497         page = __page_pool_alloc_pages_slow(pool, gfp);
 498         return page;
 499 }
 500 EXPORT_SYMBOL(page_pool_alloc_pages);
 501
 502 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 503  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 504  */
 505 #define _distance(a, b) (s32)((a) - (b))
 506
 507 static s32 page_pool_inflight(struct page_pool *pool)
 508 {
 509         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 510         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 511         s32 inflight;
 512
 513         inflight = _distance(hold_cnt, release_cnt);
 514
 515         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 516         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 517
 518         return inflight;
 519 }
 520
 521 /* Disconnects a page (from a page_pool).  API users can have a need
 522  * to disconnect a page (from a page_pool), to allow it to be used as
 523  * a regular page (that will eventually be returned to the normal
 524  * page-allocator via put_page).
 525  */
 526 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 527 {
 528         dma_addr_t dma;
 529         int count;
 530
 531         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 532                 /* Always account for inflight pages, even if we didn't
 533                  * map them
 534                  */
 535                 goto skip_dma_unmap;
 536
 537         dma = page_pool_get_dma_addr(page);
 538
 539         /* When page is unmapped, it cannot be returned to our pool */
 540         dma_unmap_page_attrs(pool->p.dev, dma,
 541                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 542                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 543         page_pool_set_dma_addr(page, 0);
 544 skip_dma_unmap:
 545         page_pool_clear_pp_info(page);
 546
 547         /* This may be the last page returned, releasing the pool, so
 548          * it is not safe to reference pool afterwards.
 549          */
 550         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 551         trace_page_pool_state_release(pool, page, count);
 552
 553         put_page(page);
 554         /* An optimization would be to call __free_pages(page, pool->p.order)
 555          * knowing page is not part of page-cache (thus avoiding a
 556          * __page_cache_release() call).
 557          */
 558 }
 559
 560 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 561 {
 562         int ret;
 563         /* BH protection not needed if current is softirq */
 564         if (in_softirq())
 565                 ret = ptr_ring_produce(&pool->ring, page);
 566         else
 567                 ret = ptr_ring_produce_bh(&pool->ring, page);
 568
 569         if (!ret) {
 570                 recycle_stat_inc(pool, ring);
 571                 return true;
 572         }
 573
 574         return false;
 575 }
 576
 577 /* Only allow direct recycling in special circumstances, into the
 578  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 579  *
 580  * Caller must provide appropriate safe context.
 581  */
 582 static bool page_pool_recycle_in_cache(struct page *page,
 583                                        struct page_pool *pool)
 584 {
 585         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
 586                 recycle_stat_inc(pool, cache_full);
 587                 return false;
 588         }
 589
 590         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 591         pool->alloc.cache[pool->alloc.count++] = page;
 592         recycle_stat_inc(pool, cached);
 593         return true;
 594 }
 595
 596 /* If the page refcnt == 1, this will try to recycle the page.
 597  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 598  * the configured size min(dma_sync_size, pool->max_len).
 599  * If the page refcnt != 1, then the page will be returned to memory
 600  * subsystem.
 601  */
 602 static __always_inline struct page *
 603 __page_pool_put_page(struct page_pool *pool, struct page *page,
 604                      unsigned int dma_sync_size, bool allow_direct)
 605 {
 606         lockdep_assert_no_hardirq();
 607
 608         /* This allocator is optimized for the XDP mode that uses
 609          * one-frame-per-page, but have fallbacks that act like the
 610          * regular page allocator APIs.
 611          *
 612          * refcnt == 1 means page_pool owns page, and can recycle it.
 613          *
 614          * page is NOT reusable when allocated when system is under
 615          * some pressure. (page_is_pfmemalloc)
 616          */
 617         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 618                 /* Read barrier done in page_ref_count / READ_ONCE */
 619
 620                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 621                         page_pool_dma_sync_for_device(pool, page,
 622                                                       dma_sync_size);
 623
 624                 if (allow_direct && in_softirq() &&
 625                     page_pool_recycle_in_cache(page, pool))
 626                         return NULL;
 627
 628                 /* Page found as candidate for recycling */
 629                 return page;
 630         }
 631         /* Fallback/non-XDP mode: API user have elevated refcnt.
 632          *
 633          * Many drivers split up the page into fragments, and some
 634          * want to keep doing this to save memory and do refcnt based
 635          * recycling. Support this use case too, to ease drivers
 636          * switching between XDP/non-XDP.
 637          *
 638          * In-case page_pool maintains the DMA mapping, API user must
 639          * call page_pool_put_page once.  In this elevated refcnt
 640          * case, the DMA is unmapped/released, as driver is likely
 641          * doing refcnt based recycle tricks, meaning another process
 642          * will be invoking put_page.
 643          */
 644         recycle_stat_inc(pool, released_refcnt);
 645         page_pool_return_page(pool, page);
 646
 647         return NULL;
 648 }
 649
 650 void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
 651                                   unsigned int dma_sync_size, bool allow_direct)
 652 {
 653         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 654         if (page && !page_pool_recycle_in_ring(pool, page)) {
 655                 /* Cache full, fallback to free pages */
 656                 recycle_stat_inc(pool, ring_full);
 657                 page_pool_return_page(pool, page);
 658         }
 659 }
 660 EXPORT_SYMBOL(page_pool_put_defragged_page);
 661
 662 /**
 663  * page_pool_put_page_bulk() - release references on multiple pages
 664  * @pool:       pool from which pages were allocated
 665  * @data:       array holding page pointers
 666  * @count:      number of pages in @data
 667  *
 668  * Tries to refill a number of pages into the ptr_ring cache holding ptr_ring
 669  * producer lock. If the ptr_ring is full, page_pool_put_page_bulk()
 670  * will release leftover pages to the page allocator.
 671  * page_pool_put_page_bulk() is suitable to be run inside the driver NAPI tx
 672  * completion loop for the XDP_REDIRECT use case.
 673  *
 674  * Please note the caller must not use data area after running
 675  * page_pool_put_page_bulk(), as this function overwrites it.
 676  */
 677 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 678                              int count)
 679 {
 680         int i, bulk_len = 0;
 681         bool in_softirq;
 682
 683         for (i = 0; i < count; i++) {
 684                 struct page *page = virt_to_head_page(data[i]);
 685
 686                 /* It is not the last user for the page frag case */
 687                 if (!page_pool_is_last_frag(page))
 688                         continue;
 689
 690                 page = __page_pool_put_page(pool, page, -1, false);
 691                 /* Approved for bulk recycling in ptr_ring cache */
 692                 if (page)
 693                         data[bulk_len++] = page;
 694         }
 695
 696         if (unlikely(!bulk_len))
 697                 return;
 698
 699         /* Bulk producer into ptr_ring page_pool cache */
 700         in_softirq = page_pool_producer_lock(pool);
 701         for (i = 0; i < bulk_len; i++) {
 702                 if (__ptr_ring_produce(&pool->ring, data[i])) {
 703                         /* ring full */
 704                         recycle_stat_inc(pool, ring_full);
 705                         break;
 706                 }
 707         }
 708         recycle_stat_add(pool, ring, i);
 709         page_pool_producer_unlock(pool, in_softirq);
 710
 711         /* Hopefully all pages was return into ptr_ring */
 712         if (likely(i == bulk_len))
 713                 return;
 714
 715         /* ptr_ring cache full, free remaining pages outside producer lock
 716          * since put_page() with refcnt == 1 can be an expensive operation
 717          */
 718         for (; i < bulk_len; i++)
 719                 page_pool_return_page(pool, data[i]);
 720 }
 721 EXPORT_SYMBOL(page_pool_put_page_bulk);
 722
 723 static struct page *page_pool_drain_frag(struct page_pool *pool,
 724                                          struct page *page)
 725 {
 726         long drain_count = BIAS_MAX - pool->frag_users;
 727
 728         /* Some user is still using the page frag */
 729         if (likely(page_pool_defrag_page(page, drain_count)))
 730                 return NULL;
 731
 732         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 733                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 734                         page_pool_dma_sync_for_device(pool, page, -1);
 735
 736                 return page;
 737         }
 738
 739         page_pool_return_page(pool, page);
 740         return NULL;
 741 }
 742
 743 static void page_pool_free_frag(struct page_pool *pool)
 744 {
 745         long drain_count = BIAS_MAX - pool->frag_users;
 746         struct page *page = pool->frag_page;
 747
 748         pool->frag_page = NULL;
 749
 750         if (!page || page_pool_defrag_page(page, drain_count))
 751                 return;
 752
 753         page_pool_return_page(pool, page);
 754 }
 755
 756 struct page *page_pool_alloc_frag(struct page_pool *pool,
 757                                   unsigned int *offset,
 758                                   unsigned int size, gfp_t gfp)
 759 {
 760         unsigned int max_size = PAGE_SIZE << pool->p.order;
 761         struct page *page = pool->frag_page;
 762
 763         if (WARN_ON(size > max_size))
 764                 return NULL;
 765
 766         size = ALIGN(size, dma_get_cache_alignment());
 767         *offset = pool->frag_offset;
 768
 769         if (page && *offset + size > max_size) {
 770                 page = page_pool_drain_frag(pool, page);
 771                 if (page) {
 772                         alloc_stat_inc(pool, fast);
 773                         goto frag_reset;
 774                 }
 775         }
 776
 777         if (!page) {
 778                 page = page_pool_alloc_pages(pool, gfp);
 779                 if (unlikely(!page)) {
 780                         pool->frag_page = NULL;
 781                         return NULL;
 782                 }
 783
 784                 pool->frag_page = page;
 785
 786 frag_reset:
 787                 pool->frag_users = 1;
 788                 *offset = 0;
 789                 pool->frag_offset = size;
 790                 page_pool_fragment_page(page, BIAS_MAX);
 791                 return page;
 792         }
 793
 794         pool->frag_users++;
 795         pool->frag_offset = *offset + size;
 796         alloc_stat_inc(pool, fast);
 797         return page;
 798 }
 799 EXPORT_SYMBOL(page_pool_alloc_frag);
 800
 801 static void page_pool_empty_ring(struct page_pool *pool)
 802 {
 803         struct page *page;
 804
 805         /* Empty recycle ring */
 806         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 807                 /* Verify the refcnt invariant of cached pages */
 808                 if (!(page_ref_count(page) == 1))
 809                         pr_crit("%s() page_pool refcnt %d violation\n",
 810                                 __func__, page_ref_count(page));
 811
 812                 page_pool_return_page(pool, page);
 813         }
 814 }
 815
 816 static void __page_pool_destroy(struct page_pool *pool)
 817 {
 818         if (pool->disconnect)
 819                 pool->disconnect(pool);
 820
 821         ptr_ring_cleanup(&pool->ring, NULL);
 822
 823         if (pool->p.flags & PP_FLAG_DMA_MAP)
 824                 put_device(pool->p.dev);
 825
 826 #ifdef CONFIG_PAGE_POOL_STATS
 827         free_percpu(pool->recycle_stats);
 828 #endif
 829         kfree(pool);
 830 }
 831
 832 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 833 {
 834         struct page *page;
 835
 836         if (pool->destroy_cnt)
 837                 return;
 838
 839         /* Empty alloc cache, assume caller made sure this is
 840          * no-longer in use, and page_pool_alloc_pages() cannot be
 841          * call concurrently.
 842          */
 843         while (pool->alloc.count) {
 844                 page = pool->alloc.cache[--pool->alloc.count];
 845                 page_pool_return_page(pool, page);
 846         }
 847 }
 848
 849 static void page_pool_scrub(struct page_pool *pool)
 850 {
 851         page_pool_empty_alloc_cache_once(pool);
 852         pool->destroy_cnt++;
 853
 854         /* No more consumers should exist, but producers could still
 855          * be in-flight.
 856          */
 857         page_pool_empty_ring(pool);
 858 }
 859
 860 static int page_pool_release(struct page_pool *pool)
 861 {
 862         int inflight;
 863
 864         page_pool_scrub(pool);
 865         inflight = page_pool_inflight(pool);
 866         if (!inflight)
 867                 __page_pool_destroy(pool);
 868
 869         return inflight;
 870 }
 871
 872 static void page_pool_release_retry(struct work_struct *wq)
 873 {
 874         struct delayed_work *dwq = to_delayed_work(wq);
 875         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 876         int inflight;
 877
 878         inflight = page_pool_release(pool);
 879         if (!inflight)
 880                 return;
 881
 882         /* Periodic warning */
 883         if (time_after_eq(jiffies, pool->defer_warn)) {
 884                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 885
 886                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 887                         __func__, inflight, sec);
 888                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 889         }
 890
 891         /* Still not ready to be disconnected, retry later */
 892         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 893 }
 894
 895 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
 896                            struct xdp_mem_info *mem)
 897 {
 898         refcount_inc(&pool->user_cnt);
 899         pool->disconnect = disconnect;
 900         pool->xdp_mem_id = mem->id;
 901 }
 902
 903 void page_pool_unlink_napi(struct page_pool *pool)
 904 {
 905         if (!pool->p.napi)
 906                 return;
 907
 908         /* To avoid races with recycling and additional barriers make sure
 909          * pool and NAPI are unlinked when NAPI is disabled.
 910          */
 911         WARN_ON(!test_bit(NAPI_STATE_SCHED, &pool->p.napi->state) ||
 912                 READ_ONCE(pool->p.napi->list_owner) != -1);
 913
 914         WRITE_ONCE(pool->p.napi, NULL);
 915 }
 916 EXPORT_SYMBOL(page_pool_unlink_napi);
 917
 918 void page_pool_destroy(struct page_pool *pool)
 919 {
 920         if (!pool)
 921                 return;
 922
 923         if (!page_pool_put(pool))
 924                 return;
 925
 926         page_pool_unlink_napi(pool);
 927         page_pool_free_frag(pool);
 928
 929         if (!page_pool_release(pool))
 930                 return;
 931
 932         pool->defer_start = jiffies;
 933         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 934
 935         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 936         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 937 }
 938 EXPORT_SYMBOL(page_pool_destroy);
 939
 940 /* Caller must provide appropriate safe context, e.g. NAPI. */
 941 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 942 {
 943         struct page *page;
 944
 945         trace_page_pool_update_nid(pool, new_nid);
 946         pool->p.nid = new_nid;
 947
 948         /* Flush pool alloc cache, as refill will check NUMA node */
 949         while (pool->alloc.count) {
 950                 page = pool->alloc.cache[--pool->alloc.count];
 951                 page_pool_return_page(pool, page);
 952         }
 953 }
 954 EXPORT_SYMBOL(page_pool_update_nid);