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->fast, sizeof(pool->p));
 177         memcpy(&pool->slow, &params->slow, sizeof(pool->slow));
 178
 179         /* Validate only known flags were used */
 180         if (pool->p.flags & ~(PP_FLAG_ALL))
 181                 return -EINVAL;
 182
 183         if (pool->p.pool_size)
 184                 ring_qsize = pool->p.pool_size;
 185
 186         /* Sanity limit mem that can be pinned down */
 187         if (ring_qsize > 32768)
 188                 return -E2BIG;
 189
 190         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
 191          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
 192          * which is the XDP_TX use-case.
 193          */
 194         if (pool->p.flags & PP_FLAG_DMA_MAP) {
 195                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
 196                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
 197                         return -EINVAL;
 198         }
 199
 200         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
 201                 /* In order to request DMA-sync-for-device the page
 202                  * needs to be mapped
 203                  */
 204                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 205                         return -EINVAL;
 206
 207                 if (!pool->p.max_len)
 208                         return -EINVAL;
 209
 210                 /* pool->p.offset has to be set according to the address
 211                  * offset used by the DMA engine to start copying rx data
 212                  */
 213         }
 214
 215         pool->has_init_callback = !!pool->slow.init_callback;
 216
 217 #ifdef CONFIG_PAGE_POOL_STATS
 218         pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
 219         if (!pool->recycle_stats)
 220                 return -ENOMEM;
 221 #endif
 222
 223         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0) {
 224 #ifdef CONFIG_PAGE_POOL_STATS
 225                 free_percpu(pool->recycle_stats);
 226 #endif
 227                 return -ENOMEM;
 228         }
 229
 230         atomic_set(&pool->pages_state_release_cnt, 0);
 231
 232         /* Driver calling page_pool_create() also call page_pool_destroy() */
 233         refcount_set(&pool->user_cnt, 1);
 234
 235         if (pool->p.flags & PP_FLAG_DMA_MAP)
 236                 get_device(pool->p.dev);
 237
 238         return 0;
 239 }
 240
 241 static void page_pool_uninit(struct page_pool *pool)
 242 {
 243         ptr_ring_cleanup(&pool->ring, NULL);
 244
 245         if (pool->p.flags & PP_FLAG_DMA_MAP)
 246                 put_device(pool->p.dev);
 247
 248 #ifdef CONFIG_PAGE_POOL_STATS
 249         free_percpu(pool->recycle_stats);
 250 #endif
 251 }
 252
 253 /**
 254  * page_pool_create() - create a page pool.
 255  * @params: parameters, see struct page_pool_params
 256  */
 257 struct page_pool *page_pool_create(const struct page_pool_params *params)
 258 {
 259         struct page_pool *pool;
 260         int err;
 261
 262         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
 263         if (!pool)
 264                 return ERR_PTR(-ENOMEM);
 265
 266         err = page_pool_init(pool, params);
 267         if (err < 0) {
 268                 pr_warn("%s() gave up with errno %d\n", __func__, err);
 269                 kfree(pool);
 270                 return ERR_PTR(err);
 271         }
 272
 273         return pool;
 274 }
 275 EXPORT_SYMBOL(page_pool_create);
 276
 277 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 278
 279 noinline
 280 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 281 {
 282         struct ptr_ring *r = &pool->ring;
 283         struct page *page;
 284         int pref_nid; /* preferred NUMA node */
 285
 286         /* Quicker fallback, avoid locks when ring is empty */
 287         if (__ptr_ring_empty(r)) {
 288                 alloc_stat_inc(pool, empty);
 289                 return NULL;
 290         }
 291
 292         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 293          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 294          */
 295 #ifdef CONFIG_NUMA
 296         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 297 #else
 298         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 299         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 300 #endif
 301
 302         /* Refill alloc array, but only if NUMA match */
 303         do {
 304                 page = __ptr_ring_consume(r);
 305                 if (unlikely(!page))
 306                         break;
 307
 308                 if (likely(page_to_nid(page) == pref_nid)) {
 309                         pool->alloc.cache[pool->alloc.count++] = page;
 310                 } else {
 311                         /* NUMA mismatch;
 312                          * (1) release 1 page to page-allocator and
 313                          * (2) break out to fallthrough to alloc_pages_node.
 314                          * This limit stress on page buddy alloactor.
 315                          */
 316                         page_pool_return_page(pool, page);
 317                         alloc_stat_inc(pool, waive);
 318                         page = NULL;
 319                         break;
 320                 }
 321         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 322
 323         /* Return last page */
 324         if (likely(pool->alloc.count > 0)) {
 325                 page = pool->alloc.cache[--pool->alloc.count];
 326                 alloc_stat_inc(pool, refill);
 327         }
 328
 329         return page;
 330 }
 331
 332 /* fast path */
 333 static struct page *__page_pool_get_cached(struct page_pool *pool)
 334 {
 335         struct page *page;
 336
 337         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 338         if (likely(pool->alloc.count)) {
 339                 /* Fast-path */
 340                 page = pool->alloc.cache[--pool->alloc.count];
 341                 alloc_stat_inc(pool, fast);
 342         } else {
 343                 page = page_pool_refill_alloc_cache(pool);
 344         }
 345
 346         return page;
 347 }
 348
 349 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 350                                           struct page *page,
 351                                           unsigned int dma_sync_size)
 352 {
 353         dma_addr_t dma_addr = page_pool_get_dma_addr(page);
 354
 355         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 356         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 357                                          pool->p.offset, dma_sync_size,
 358                                          pool->p.dma_dir);
 359 }
 360
 361 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 362 {
 363         dma_addr_t dma;
 364
 365         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 366          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 367          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 368          * This mapping is kept for lifetime of page, until leaving pool.
 369          */
 370         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 371                                  (PAGE_SIZE << pool->p.order),
 372                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC |
 373                                                   DMA_ATTR_WEAK_ORDERING);
 374         if (dma_mapping_error(pool->p.dev, dma))
 375                 return false;
 376
 377         if (page_pool_set_dma_addr(page, dma))
 378                 goto unmap_failed;
 379
 380         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 381                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 382
 383         return true;
 384
 385 unmap_failed:
 386         WARN_ON_ONCE("unexpected DMA address, please report to netdev@");
 387         dma_unmap_page_attrs(pool->p.dev, dma,
 388                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 389                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 390         return false;
 391 }
 392
 393 static void page_pool_set_pp_info(struct page_pool *pool,
 394                                   struct page *page)
 395 {
 396         page->pp = pool;
 397         page->pp_magic |= PP_SIGNATURE;
 398
 399         /* Ensuring all pages have been split into one fragment initially:
 400          * page_pool_set_pp_info() is only called once for every page when it
 401          * is allocated from the page allocator and page_pool_fragment_page()
 402          * is dirtying the same cache line as the page->pp_magic above, so
 403          * the overhead is negligible.
 404          */
 405         page_pool_fragment_page(page, 1);
 406         if (pool->has_init_callback)
 407                 pool->slow.init_callback(page, pool->slow.init_arg);
 408 }
 409
 410 static void page_pool_clear_pp_info(struct page *page)
 411 {
 412         page->pp_magic = 0;
 413         page->pp = NULL;
 414 }
 415
 416 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 417                                                  gfp_t gfp)
 418 {
 419         struct page *page;
 420
 421         gfp |= __GFP_COMP;
 422         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 423         if (unlikely(!page))
 424                 return NULL;
 425
 426         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 427             unlikely(!page_pool_dma_map(pool, page))) {
 428                 put_page(page);
 429                 return NULL;
 430         }
 431
 432         alloc_stat_inc(pool, slow_high_order);
 433         page_pool_set_pp_info(pool, page);
 434
 435         /* Track how many pages are held 'in-flight' */
 436         pool->pages_state_hold_cnt++;
 437         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 438         return page;
 439 }
 440
 441 /* slow path */
 442 noinline
 443 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 444                                                  gfp_t gfp)
 445 {
 446         const int bulk = PP_ALLOC_CACHE_REFILL;
 447         unsigned int pp_flags = pool->p.flags;
 448         unsigned int pp_order = pool->p.order;
 449         struct page *page;
 450         int i, nr_pages;
 451
 452         /* Don't support bulk alloc for high-order pages */
 453         if (unlikely(pp_order))
 454                 return __page_pool_alloc_page_order(pool, gfp);
 455
 456         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 457         if (unlikely(pool->alloc.count > 0))
 458                 return pool->alloc.cache[--pool->alloc.count];
 459
 460         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 461         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 462
 463         nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk,
 464                                                pool->alloc.cache);
 465         if (unlikely(!nr_pages))
 466                 return NULL;
 467
 468         /* Pages have been filled into alloc.cache array, but count is zero and
 469          * page element have not been (possibly) DMA mapped.
 470          */
 471         for (i = 0; i < nr_pages; i++) {
 472                 page = pool->alloc.cache[i];
 473                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 474                     unlikely(!page_pool_dma_map(pool, page))) {
 475                         put_page(page);
 476                         continue;
 477                 }
 478
 479                 page_pool_set_pp_info(pool, page);
 480                 pool->alloc.cache[pool->alloc.count++] = page;
 481                 /* Track how many pages are held 'in-flight' */
 482                 pool->pages_state_hold_cnt++;
 483                 trace_page_pool_state_hold(pool, page,
 484                                            pool->pages_state_hold_cnt);
 485         }
 486
 487         /* Return last page */
 488         if (likely(pool->alloc.count > 0)) {
 489                 page = pool->alloc.cache[--pool->alloc.count];
 490                 alloc_stat_inc(pool, slow);
 491         } else {
 492                 page = NULL;
 493         }
 494
 495         /* When page just alloc'ed is should/must have refcnt 1. */
 496         return page;
 497 }
 498
 499 /* For using page_pool replace: alloc_pages() API calls, but provide
 500  * synchronization guarantee for allocation side.
 501  */
 502 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 503 {
 504         struct page *page;
 505
 506         /* Fast-path: Get a page from cache */
 507         page = __page_pool_get_cached(pool);
 508         if (page)
 509                 return page;
 510
 511         /* Slow-path: cache empty, do real allocation */
 512         page = __page_pool_alloc_pages_slow(pool, gfp);
 513         return page;
 514 }
 515 EXPORT_SYMBOL(page_pool_alloc_pages);
 516
 517 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 518  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 519  */
 520 #define _distance(a, b) (s32)((a) - (b))
 521
 522 static s32 page_pool_inflight(struct page_pool *pool)
 523 {
 524         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 525         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 526         s32 inflight;
 527
 528         inflight = _distance(hold_cnt, release_cnt);
 529
 530         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 531         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 532
 533         return inflight;
 534 }
 535
 536 /* Disconnects a page (from a page_pool).  API users can have a need
 537  * to disconnect a page (from a page_pool), to allow it to be used as
 538  * a regular page (that will eventually be returned to the normal
 539  * page-allocator via put_page).
 540  */
 541 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 542 {
 543         dma_addr_t dma;
 544         int count;
 545
 546         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 547                 /* Always account for inflight pages, even if we didn't
 548                  * map them
 549                  */
 550                 goto skip_dma_unmap;
 551
 552         dma = page_pool_get_dma_addr(page);
 553
 554         /* When page is unmapped, it cannot be returned to our pool */
 555         dma_unmap_page_attrs(pool->p.dev, dma,
 556                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 557                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 558         page_pool_set_dma_addr(page, 0);
 559 skip_dma_unmap:
 560         page_pool_clear_pp_info(page);
 561
 562         /* This may be the last page returned, releasing the pool, so
 563          * it is not safe to reference pool afterwards.
 564          */
 565         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 566         trace_page_pool_state_release(pool, page, count);
 567
 568         put_page(page);
 569         /* An optimization would be to call __free_pages(page, pool->p.order)
 570          * knowing page is not part of page-cache (thus avoiding a
 571          * __page_cache_release() call).
 572          */
 573 }
 574
 575 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 576 {
 577         int ret;
 578         /* BH protection not needed if current is softirq */
 579         if (in_softirq())
 580                 ret = ptr_ring_produce(&pool->ring, page);
 581         else
 582                 ret = ptr_ring_produce_bh(&pool->ring, page);
 583
 584         if (!ret) {
 585                 recycle_stat_inc(pool, ring);
 586                 return true;
 587         }
 588
 589         return false;
 590 }
 591
 592 /* Only allow direct recycling in special circumstances, into the
 593  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 594  *
 595  * Caller must provide appropriate safe context.
 596  */
 597 static bool page_pool_recycle_in_cache(struct page *page,
 598                                        struct page_pool *pool)
 599 {
 600         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
 601                 recycle_stat_inc(pool, cache_full);
 602                 return false;
 603         }
 604
 605         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 606         pool->alloc.cache[pool->alloc.count++] = page;
 607         recycle_stat_inc(pool, cached);
 608         return true;
 609 }
 610
 611 /* If the page refcnt == 1, this will try to recycle the page.
 612  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 613  * the configured size min(dma_sync_size, pool->max_len).
 614  * If the page refcnt != 1, then the page will be returned to memory
 615  * subsystem.
 616  */
 617 static __always_inline struct page *
 618 __page_pool_put_page(struct page_pool *pool, struct page *page,
 619                      unsigned int dma_sync_size, bool allow_direct)
 620 {
 621         lockdep_assert_no_hardirq();
 622
 623         /* This allocator is optimized for the XDP mode that uses
 624          * one-frame-per-page, but have fallbacks that act like the
 625          * regular page allocator APIs.
 626          *
 627          * refcnt == 1 means page_pool owns page, and can recycle it.
 628          *
 629          * page is NOT reusable when allocated when system is under
 630          * some pressure. (page_is_pfmemalloc)
 631          */
 632         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 633                 /* Read barrier done in page_ref_count / READ_ONCE */
 634
 635                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 636                         page_pool_dma_sync_for_device(pool, page,
 637                                                       dma_sync_size);
 638
 639                 if (allow_direct && in_softirq() &&
 640                     page_pool_recycle_in_cache(page, pool))
 641                         return NULL;
 642
 643                 /* Page found as candidate for recycling */
 644                 return page;
 645         }
 646         /* Fallback/non-XDP mode: API user have elevated refcnt.
 647          *
 648          * Many drivers split up the page into fragments, and some
 649          * want to keep doing this to save memory and do refcnt based
 650          * recycling. Support this use case too, to ease drivers
 651          * switching between XDP/non-XDP.
 652          *
 653          * In-case page_pool maintains the DMA mapping, API user must
 654          * call page_pool_put_page once.  In this elevated refcnt
 655          * case, the DMA is unmapped/released, as driver is likely
 656          * doing refcnt based recycle tricks, meaning another process
 657          * will be invoking put_page.
 658          */
 659         recycle_stat_inc(pool, released_refcnt);
 660         page_pool_return_page(pool, page);
 661
 662         return NULL;
 663 }
 664
 665 void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
 666                                   unsigned int dma_sync_size, bool allow_direct)
 667 {
 668         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 669         if (page && !page_pool_recycle_in_ring(pool, page)) {
 670                 /* Cache full, fallback to free pages */
 671                 recycle_stat_inc(pool, ring_full);
 672                 page_pool_return_page(pool, page);
 673         }
 674 }
 675 EXPORT_SYMBOL(page_pool_put_defragged_page);
 676
 677 /**
 678  * page_pool_put_page_bulk() - release references on multiple pages
 679  * @pool:       pool from which pages were allocated
 680  * @data:       array holding page pointers
 681  * @count:      number of pages in @data
 682  *
 683  * Tries to refill a number of pages into the ptr_ring cache holding ptr_ring
 684  * producer lock. If the ptr_ring is full, page_pool_put_page_bulk()
 685  * will release leftover pages to the page allocator.
 686  * page_pool_put_page_bulk() is suitable to be run inside the driver NAPI tx
 687  * completion loop for the XDP_REDIRECT use case.
 688  *
 689  * Please note the caller must not use data area after running
 690  * page_pool_put_page_bulk(), as this function overwrites it.
 691  */
 692 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 693                              int count)
 694 {
 695         int i, bulk_len = 0;
 696         bool in_softirq;
 697
 698         for (i = 0; i < count; i++) {
 699                 struct page *page = virt_to_head_page(data[i]);
 700
 701                 /* It is not the last user for the page frag case */
 702                 if (!page_pool_is_last_frag(page))
 703                         continue;
 704
 705                 page = __page_pool_put_page(pool, page, -1, false);
 706                 /* Approved for bulk recycling in ptr_ring cache */
 707                 if (page)
 708                         data[bulk_len++] = page;
 709         }
 710
 711         if (unlikely(!bulk_len))
 712                 return;
 713
 714         /* Bulk producer into ptr_ring page_pool cache */
 715         in_softirq = page_pool_producer_lock(pool);
 716         for (i = 0; i < bulk_len; i++) {
 717                 if (__ptr_ring_produce(&pool->ring, data[i])) {
 718                         /* ring full */
 719                         recycle_stat_inc(pool, ring_full);
 720                         break;
 721                 }
 722         }
 723         recycle_stat_add(pool, ring, i);
 724         page_pool_producer_unlock(pool, in_softirq);
 725
 726         /* Hopefully all pages was return into ptr_ring */
 727         if (likely(i == bulk_len))
 728                 return;
 729
 730         /* ptr_ring cache full, free remaining pages outside producer lock
 731          * since put_page() with refcnt == 1 can be an expensive operation
 732          */
 733         for (; i < bulk_len; i++)
 734                 page_pool_return_page(pool, data[i]);
 735 }
 736 EXPORT_SYMBOL(page_pool_put_page_bulk);
 737
 738 static struct page *page_pool_drain_frag(struct page_pool *pool,
 739                                          struct page *page)
 740 {
 741         long drain_count = BIAS_MAX - pool->frag_users;
 742
 743         /* Some user is still using the page frag */
 744         if (likely(page_pool_defrag_page(page, drain_count)))
 745                 return NULL;
 746
 747         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 748                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 749                         page_pool_dma_sync_for_device(pool, page, -1);
 750
 751                 return page;
 752         }
 753
 754         page_pool_return_page(pool, page);
 755         return NULL;
 756 }
 757
 758 static void page_pool_free_frag(struct page_pool *pool)
 759 {
 760         long drain_count = BIAS_MAX - pool->frag_users;
 761         struct page *page = pool->frag_page;
 762
 763         pool->frag_page = NULL;
 764
 765         if (!page || page_pool_defrag_page(page, drain_count))
 766                 return;
 767
 768         page_pool_return_page(pool, page);
 769 }
 770
 771 struct page *page_pool_alloc_frag(struct page_pool *pool,
 772                                   unsigned int *offset,
 773                                   unsigned int size, gfp_t gfp)
 774 {
 775         unsigned int max_size = PAGE_SIZE << pool->p.order;
 776         struct page *page = pool->frag_page;
 777
 778         if (WARN_ON(size > max_size))
 779                 return NULL;
 780
 781         size = ALIGN(size, dma_get_cache_alignment());
 782         *offset = pool->frag_offset;
 783
 784         if (page && *offset + size > max_size) {
 785                 page = page_pool_drain_frag(pool, page);
 786                 if (page) {
 787                         alloc_stat_inc(pool, fast);
 788                         goto frag_reset;
 789                 }
 790         }
 791
 792         if (!page) {
 793                 page = page_pool_alloc_pages(pool, gfp);
 794                 if (unlikely(!page)) {
 795                         pool->frag_page = NULL;
 796                         return NULL;
 797                 }
 798
 799                 pool->frag_page = page;
 800
 801 frag_reset:
 802                 pool->frag_users = 1;
 803                 *offset = 0;
 804                 pool->frag_offset = size;
 805                 page_pool_fragment_page(page, BIAS_MAX);
 806                 return page;
 807         }
 808
 809         pool->frag_users++;
 810         pool->frag_offset = *offset + size;
 811         alloc_stat_inc(pool, fast);
 812         return page;
 813 }
 814 EXPORT_SYMBOL(page_pool_alloc_frag);
 815
 816 static void page_pool_empty_ring(struct page_pool *pool)
 817 {
 818         struct page *page;
 819
 820         /* Empty recycle ring */
 821         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 822                 /* Verify the refcnt invariant of cached pages */
 823                 if (!(page_ref_count(page) == 1))
 824                         pr_crit("%s() page_pool refcnt %d violation\n",
 825                                 __func__, page_ref_count(page));
 826
 827                 page_pool_return_page(pool, page);
 828         }
 829 }
 830
 831 static void __page_pool_destroy(struct page_pool *pool)
 832 {
 833         if (pool->disconnect)
 834                 pool->disconnect(pool);
 835
 836         page_pool_uninit(pool);
 837         kfree(pool);
 838 }
 839
 840 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 841 {
 842         struct page *page;
 843
 844         if (pool->destroy_cnt)
 845                 return;
 846
 847         /* Empty alloc cache, assume caller made sure this is
 848          * no-longer in use, and page_pool_alloc_pages() cannot be
 849          * call concurrently.
 850          */
 851         while (pool->alloc.count) {
 852                 page = pool->alloc.cache[--pool->alloc.count];
 853                 page_pool_return_page(pool, page);
 854         }
 855 }
 856
 857 static void page_pool_scrub(struct page_pool *pool)
 858 {
 859         page_pool_empty_alloc_cache_once(pool);
 860         pool->destroy_cnt++;
 861
 862         /* No more consumers should exist, but producers could still
 863          * be in-flight.
 864          */
 865         page_pool_empty_ring(pool);
 866 }
 867
 868 static int page_pool_release(struct page_pool *pool)
 869 {
 870         int inflight;
 871
 872         page_pool_scrub(pool);
 873         inflight = page_pool_inflight(pool);
 874         if (!inflight)
 875                 __page_pool_destroy(pool);
 876
 877         return inflight;
 878 }
 879
 880 static void page_pool_release_retry(struct work_struct *wq)
 881 {
 882         struct delayed_work *dwq = to_delayed_work(wq);
 883         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 884         int inflight;
 885
 886         inflight = page_pool_release(pool);
 887         if (!inflight)
 888                 return;
 889
 890         /* Periodic warning */
 891         if (time_after_eq(jiffies, pool->defer_warn)) {
 892                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 893
 894                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 895                         __func__, inflight, sec);
 896                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 897         }
 898
 899         /* Still not ready to be disconnected, retry later */
 900         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 901 }
 902
 903 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
 904                            struct xdp_mem_info *mem)
 905 {
 906         refcount_inc(&pool->user_cnt);
 907         pool->disconnect = disconnect;
 908         pool->xdp_mem_id = mem->id;
 909 }
 910
 911 void page_pool_unlink_napi(struct page_pool *pool)
 912 {
 913         if (!pool->p.napi)
 914                 return;
 915
 916         /* To avoid races with recycling and additional barriers make sure
 917          * pool and NAPI are unlinked when NAPI is disabled.
 918          */
 919         WARN_ON(!test_bit(NAPI_STATE_SCHED, &pool->p.napi->state) ||
 920                 READ_ONCE(pool->p.napi->list_owner) != -1);
 921
 922         WRITE_ONCE(pool->p.napi, NULL);
 923 }
 924 EXPORT_SYMBOL(page_pool_unlink_napi);
 925
 926 void page_pool_destroy(struct page_pool *pool)
 927 {
 928         if (!pool)
 929                 return;
 930
 931         if (!page_pool_put(pool))
 932                 return;
 933
 934         page_pool_unlink_napi(pool);
 935         page_pool_free_frag(pool);
 936
 937         if (!page_pool_release(pool))
 938                 return;
 939
 940         pool->defer_start = jiffies;
 941         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 942
 943         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 944         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 945 }
 946 EXPORT_SYMBOL(page_pool_destroy);
 947
 948 /* Caller must provide appropriate safe context, e.g. NAPI. */
 949 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 950 {
 951         struct page *page;
 952
 953         trace_page_pool_update_nid(pool, new_nid);
 954         pool->p.nid = new_nid;
 955
 956         /* Flush pool alloc cache, as refill will check NUMA node */
 957         while (pool->alloc.count) {
 958                 page = pool->alloc.cache[--pool->alloc.count];
 959                 page_pool_return_page(pool, page);
 960         }
 961 }
 962 EXPORT_SYMBOL(page_pool_update_nid);