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