net/core/page_pool.c

   1 /* SPDX-License-Identifier: GPL-2.0
   2  *
   3  * page_pool.c
   4  *      Author: Jesper Dangaard Brouer <netoptimizer@brouer.com>
   5  *      Copyright (C) 2016 Red Hat, Inc.
   6  */
   7
   8 #include <linux/types.h>
   9 #include <linux/kernel.h>
  10 #include <linux/slab.h>
  11 #include <linux/device.h>
  12
  13 #include <net/page_pool.h>
  14 #include <net/xdp.h>
  15
  16 #include <linux/dma-direction.h>
  17 #include <linux/dma-mapping.h>
  18 #include <linux/page-flags.h>
  19 #include <linux/mm.h> /* for put_page() */
  20 #include <linux/poison.h>
  21 #include <linux/ethtool.h>
  22 #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 bool page_pool_get_stats(struct page_pool *pool,
  62                          struct page_pool_stats *stats)
  63 {
  64         int cpu = 0;
  65
  66         if (!stats)
  67                 return false;
  68
  69         /* The caller is responsible to initialize stats. */
  70         stats->alloc_stats.fast += pool->alloc_stats.fast;
  71         stats->alloc_stats.slow += pool->alloc_stats.slow;
  72         stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
  73         stats->alloc_stats.empty += pool->alloc_stats.empty;
  74         stats->alloc_stats.refill += pool->alloc_stats.refill;
  75         stats->alloc_stats.waive += pool->alloc_stats.waive;
  76
  77         for_each_possible_cpu(cpu) {
  78                 const struct page_pool_recycle_stats *pcpu =
  79                         per_cpu_ptr(pool->recycle_stats, cpu);
  80
  81                 stats->recycle_stats.cached += pcpu->cached;
  82                 stats->recycle_stats.cache_full += pcpu->cache_full;
  83                 stats->recycle_stats.ring += pcpu->ring;
  84                 stats->recycle_stats.ring_full += pcpu->ring_full;
  85                 stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
  86         }
  87
  88         return true;
  89 }
  90 EXPORT_SYMBOL(page_pool_get_stats);
  91
  92 u8 *page_pool_ethtool_stats_get_strings(u8 *data)
  93 {
  94         int i;
  95
  96         for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
  97                 memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
  98                 data += ETH_GSTRING_LEN;
  99         }
 100
 101         return data;
 102 }
 103 EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);
 104
 105 int page_pool_ethtool_stats_get_count(void)
 106 {
 107         return ARRAY_SIZE(pp_stats);
 108 }
 109 EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);
 110
 111 u64 *page_pool_ethtool_stats_get(u64 *data, void *stats)
 112 {
 113         struct page_pool_stats *pool_stats = stats;
 114
 115         *data++ = pool_stats->alloc_stats.fast;
 116         *data++ = pool_stats->alloc_stats.slow;
 117         *data++ = pool_stats->alloc_stats.slow_high_order;
 118         *data++ = pool_stats->alloc_stats.empty;
 119         *data++ = pool_stats->alloc_stats.refill;
 120         *data++ = pool_stats->alloc_stats.waive;
 121         *data++ = pool_stats->recycle_stats.cached;
 122         *data++ = pool_stats->recycle_stats.cache_full;
 123         *data++ = pool_stats->recycle_stats.ring;
 124         *data++ = pool_stats->recycle_stats.ring_full;
 125         *data++ = pool_stats->recycle_stats.released_refcnt;
 126
 127         return data;
 128 }
 129 EXPORT_SYMBOL(page_pool_ethtool_stats_get);
 130
 131 #else
 132 #define alloc_stat_inc(pool, __stat)
 133 #define recycle_stat_inc(pool, __stat)
 134 #define recycle_stat_add(pool, __stat, val)
 135 #endif
 136
 137 static bool page_pool_producer_lock(struct page_pool *pool)
 138         __acquires(&pool->ring.producer_lock)
 139 {
 140         bool in_softirq = in_softirq();
 141
 142         if (in_softirq)
 143                 spin_lock(&pool->ring.producer_lock);
 144         else
 145                 spin_lock_bh(&pool->ring.producer_lock);
 146
 147         return in_softirq;
 148 }
 149
 150 static void page_pool_producer_unlock(struct page_pool *pool,
 151                                       bool in_softirq)
 152         __releases(&pool->ring.producer_lock)
 153 {
 154         if (in_softirq)
 155                 spin_unlock(&pool->ring.producer_lock);
 156         else
 157                 spin_unlock_bh(&pool->ring.producer_lock);
 158 }
 159
 160 static int page_pool_init(struct page_pool *pool,
 161                           const struct page_pool_params *params)
 162 {
 163         unsigned int ring_qsize = 1024; /* Default */
 164
 165         memcpy(&pool->p, params, sizeof(pool->p));
 166
 167         /* Validate only known flags were used */
 168         if (pool->p.flags & ~(PP_FLAG_ALL))
 169                 return -EINVAL;
 170
 171         if (pool->p.pool_size)
 172                 ring_qsize = pool->p.pool_size;
 173
 174         /* Sanity limit mem that can be pinned down */
 175         if (ring_qsize > 32768)
 176                 return -E2BIG;
 177
 178         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
 179          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
 180          * which is the XDP_TX use-case.
 181          */
 182         if (pool->p.flags & PP_FLAG_DMA_MAP) {
 183                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
 184                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
 185                         return -EINVAL;
 186         }
 187
 188         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
 189                 /* In order to request DMA-sync-for-device the page
 190                  * needs to be mapped
 191                  */
 192                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 193                         return -EINVAL;
 194
 195                 if (!pool->p.max_len)
 196                         return -EINVAL;
 197
 198                 /* pool->p.offset has to be set according to the address
 199                  * offset used by the DMA engine to start copying rx data
 200                  */
 201         }
 202
 203         if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT &&
 204             pool->p.flags & PP_FLAG_PAGE_FRAG)
 205                 return -EINVAL;
 206
 207 #ifdef CONFIG_PAGE_POOL_STATS
 208         pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
 209         if (!pool->recycle_stats)
 210                 return -ENOMEM;
 211 #endif
 212
 213         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
 214                 return -ENOMEM;
 215
 216         atomic_set(&pool->pages_state_release_cnt, 0);
 217
 218         /* Driver calling page_pool_create() also call page_pool_destroy() */
 219         refcount_set(&pool->user_cnt, 1);
 220
 221         if (pool->p.flags & PP_FLAG_DMA_MAP)
 222                 get_device(pool->p.dev);
 223
 224         return 0;
 225 }
 226
 227 struct page_pool *page_pool_create(const struct page_pool_params *params)
 228 {
 229         struct page_pool *pool;
 230         int err;
 231
 232         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
 233         if (!pool)
 234                 return ERR_PTR(-ENOMEM);
 235
 236         err = page_pool_init(pool, params);
 237         if (err < 0) {
 238                 pr_warn("%s() gave up with errno %d\n", __func__, err);
 239                 kfree(pool);
 240                 return ERR_PTR(err);
 241         }
 242
 243         return pool;
 244 }
 245 EXPORT_SYMBOL(page_pool_create);
 246
 247 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 248
 249 noinline
 250 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 251 {
 252         struct ptr_ring *r = &pool->ring;
 253         struct page *page;
 254         int pref_nid; /* preferred NUMA node */
 255
 256         /* Quicker fallback, avoid locks when ring is empty */
 257         if (__ptr_ring_empty(r)) {
 258                 alloc_stat_inc(pool, empty);
 259                 return NULL;
 260         }
 261
 262         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 263          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 264          */
 265 #ifdef CONFIG_NUMA
 266         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 267 #else
 268         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 269         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 270 #endif
 271
 272         /* Refill alloc array, but only if NUMA match */
 273         do {
 274                 page = __ptr_ring_consume(r);
 275                 if (unlikely(!page))
 276                         break;
 277
 278                 if (likely(page_to_nid(page) == pref_nid)) {
 279                         pool->alloc.cache[pool->alloc.count++] = page;
 280                 } else {
 281                         /* NUMA mismatch;
 282                          * (1) release 1 page to page-allocator and
 283                          * (2) break out to fallthrough to alloc_pages_node.
 284                          * This limit stress on page buddy alloactor.
 285                          */
 286                         page_pool_return_page(pool, page);
 287                         alloc_stat_inc(pool, waive);
 288                         page = NULL;
 289                         break;
 290                 }
 291         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 292
 293         /* Return last page */
 294         if (likely(pool->alloc.count > 0)) {
 295                 page = pool->alloc.cache[--pool->alloc.count];
 296                 alloc_stat_inc(pool, refill);
 297         }
 298
 299         return page;
 300 }
 301
 302 /* fast path */
 303 static struct page *__page_pool_get_cached(struct page_pool *pool)
 304 {
 305         struct page *page;
 306
 307         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 308         if (likely(pool->alloc.count)) {
 309                 /* Fast-path */
 310                 page = pool->alloc.cache[--pool->alloc.count];
 311                 alloc_stat_inc(pool, fast);
 312         } else {
 313                 page = page_pool_refill_alloc_cache(pool);
 314         }
 315
 316         return page;
 317 }
 318
 319 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 320                                           struct page *page,
 321                                           unsigned int dma_sync_size)
 322 {
 323         dma_addr_t dma_addr = page_pool_get_dma_addr(page);
 324
 325         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 326         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 327                                          pool->p.offset, dma_sync_size,
 328                                          pool->p.dma_dir);
 329 }
 330
 331 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 332 {
 333         dma_addr_t dma;
 334
 335         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 336          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 337          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 338          * This mapping is kept for lifetime of page, until leaving pool.
 339          */
 340         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 341                                  (PAGE_SIZE << pool->p.order),
 342                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC |
 343                                                   DMA_ATTR_WEAK_ORDERING);
 344         if (dma_mapping_error(pool->p.dev, dma))
 345                 return false;
 346
 347         page_pool_set_dma_addr(page, dma);
 348
 349         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 350                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 351
 352         return true;
 353 }
 354
 355 static void page_pool_set_pp_info(struct page_pool *pool,
 356                                   struct page *page)
 357 {
 358         page->pp = pool;
 359         page->pp_magic |= PP_SIGNATURE;
 360         if (pool->p.init_callback)
 361                 pool->p.init_callback(page, pool->p.init_arg);
 362 }
 363
 364 static void page_pool_clear_pp_info(struct page *page)
 365 {
 366         page->pp_magic = 0;
 367         page->pp = NULL;
 368 }
 369
 370 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 371                                                  gfp_t gfp)
 372 {
 373         struct page *page;
 374
 375         gfp |= __GFP_COMP;
 376         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 377         if (unlikely(!page))
 378                 return NULL;
 379
 380         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 381             unlikely(!page_pool_dma_map(pool, page))) {
 382                 put_page(page);
 383                 return NULL;
 384         }
 385
 386         alloc_stat_inc(pool, slow_high_order);
 387         page_pool_set_pp_info(pool, page);
 388
 389         /* Track how many pages are held 'in-flight' */
 390         pool->pages_state_hold_cnt++;
 391         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 392         return page;
 393 }
 394
 395 /* slow path */
 396 noinline
 397 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 398                                                  gfp_t gfp)
 399 {
 400         const int bulk = PP_ALLOC_CACHE_REFILL;
 401         unsigned int pp_flags = pool->p.flags;
 402         unsigned int pp_order = pool->p.order;
 403         struct page *page;
 404         int i, nr_pages;
 405
 406         /* Don't support bulk alloc for high-order pages */
 407         if (unlikely(pp_order))
 408                 return __page_pool_alloc_page_order(pool, gfp);
 409
 410         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 411         if (unlikely(pool->alloc.count > 0))
 412                 return pool->alloc.cache[--pool->alloc.count];
 413
 414         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 415         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 416
 417         nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk,
 418                                                pool->alloc.cache);
 419         if (unlikely(!nr_pages))
 420                 return NULL;
 421
 422         /* Pages have been filled into alloc.cache array, but count is zero and
 423          * page element have not been (possibly) DMA mapped.
 424          */
 425         for (i = 0; i < nr_pages; i++) {
 426                 page = pool->alloc.cache[i];
 427                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 428                     unlikely(!page_pool_dma_map(pool, page))) {
 429                         put_page(page);
 430                         continue;
 431                 }
 432
 433                 page_pool_set_pp_info(pool, page);
 434                 pool->alloc.cache[pool->alloc.count++] = page;
 435                 /* Track how many pages are held 'in-flight' */
 436                 pool->pages_state_hold_cnt++;
 437                 trace_page_pool_state_hold(pool, page,
 438                                            pool->pages_state_hold_cnt);
 439         }
 440
 441         /* Return last page */
 442         if (likely(pool->alloc.count > 0)) {
 443                 page = pool->alloc.cache[--pool->alloc.count];
 444                 alloc_stat_inc(pool, slow);
 445         } else {
 446                 page = NULL;
 447         }
 448
 449         /* When page just alloc'ed is should/must have refcnt 1. */
 450         return page;
 451 }
 452
 453 /* For using page_pool replace: alloc_pages() API calls, but provide
 454  * synchronization guarantee for allocation side.
 455  */
 456 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 457 {
 458         struct page *page;
 459
 460         /* Fast-path: Get a page from cache */
 461         page = __page_pool_get_cached(pool);
 462         if (page)
 463                 return page;
 464
 465         /* Slow-path: cache empty, do real allocation */
 466         page = __page_pool_alloc_pages_slow(pool, gfp);
 467         return page;
 468 }
 469 EXPORT_SYMBOL(page_pool_alloc_pages);
 470
 471 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 472  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 473  */
 474 #define _distance(a, b) (s32)((a) - (b))
 475
 476 static s32 page_pool_inflight(struct page_pool *pool)
 477 {
 478         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 479         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 480         s32 inflight;
 481
 482         inflight = _distance(hold_cnt, release_cnt);
 483
 484         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 485         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 486
 487         return inflight;
 488 }
 489
 490 /* Disconnects a page (from a page_pool).  API users can have a need
 491  * to disconnect a page (from a page_pool), to allow it to be used as
 492  * a regular page (that will eventually be returned to the normal
 493  * page-allocator via put_page).
 494  */
 495 void page_pool_release_page(struct page_pool *pool, struct page *page)
 496 {
 497         dma_addr_t dma;
 498         int count;
 499
 500         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 501                 /* Always account for inflight pages, even if we didn't
 502                  * map them
 503                  */
 504                 goto skip_dma_unmap;
 505
 506         dma = page_pool_get_dma_addr(page);
 507
 508         /* When page is unmapped, it cannot be returned to our pool */
 509         dma_unmap_page_attrs(pool->p.dev, dma,
 510                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 511                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 512         page_pool_set_dma_addr(page, 0);
 513 skip_dma_unmap:
 514         page_pool_clear_pp_info(page);
 515
 516         /* This may be the last page returned, releasing the pool, so
 517          * it is not safe to reference pool afterwards.
 518          */
 519         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 520         trace_page_pool_state_release(pool, page, count);
 521 }
 522 EXPORT_SYMBOL(page_pool_release_page);
 523
 524 /* Return a page to the page allocator, cleaning up our state */
 525 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 526 {
 527         page_pool_release_page(pool, page);
 528
 529         put_page(page);
 530         /* An optimization would be to call __free_pages(page, pool->p.order)
 531          * knowing page is not part of page-cache (thus avoiding a
 532          * __page_cache_release() call).
 533          */
 534 }
 535
 536 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 537 {
 538         int ret;
 539         /* BH protection not needed if current is softirq */
 540         if (in_softirq())
 541                 ret = ptr_ring_produce(&pool->ring, page);
 542         else
 543                 ret = ptr_ring_produce_bh(&pool->ring, page);
 544
 545         if (!ret) {
 546                 recycle_stat_inc(pool, ring);
 547                 return true;
 548         }
 549
 550         return false;
 551 }
 552
 553 /* Only allow direct recycling in special circumstances, into the
 554  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 555  *
 556  * Caller must provide appropriate safe context.
 557  */
 558 static bool page_pool_recycle_in_cache(struct page *page,
 559                                        struct page_pool *pool)
 560 {
 561         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
 562                 recycle_stat_inc(pool, cache_full);
 563                 return false;
 564         }
 565
 566         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 567         pool->alloc.cache[pool->alloc.count++] = page;
 568         recycle_stat_inc(pool, cached);
 569         return true;
 570 }
 571
 572 /* If the page refcnt == 1, this will try to recycle the page.
 573  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 574  * the configured size min(dma_sync_size, pool->max_len).
 575  * If the page refcnt != 1, then the page will be returned to memory
 576  * subsystem.
 577  */
 578 static __always_inline struct page *
 579 __page_pool_put_page(struct page_pool *pool, struct page *page,
 580                      unsigned int dma_sync_size, bool allow_direct)
 581 {
 582         /* This allocator is optimized for the XDP mode that uses
 583          * one-frame-per-page, but have fallbacks that act like the
 584          * regular page allocator APIs.
 585          *
 586          * refcnt == 1 means page_pool owns page, and can recycle it.
 587          *
 588          * page is NOT reusable when allocated when system is under
 589          * some pressure. (page_is_pfmemalloc)
 590          */
 591         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 592                 /* Read barrier done in page_ref_count / READ_ONCE */
 593
 594                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 595                         page_pool_dma_sync_for_device(pool, page,
 596                                                       dma_sync_size);
 597
 598                 if (allow_direct && in_softirq() &&
 599                     page_pool_recycle_in_cache(page, pool))
 600                         return NULL;
 601
 602                 /* Page found as candidate for recycling */
 603                 return page;
 604         }
 605         /* Fallback/non-XDP mode: API user have elevated refcnt.
 606          *
 607          * Many drivers split up the page into fragments, and some
 608          * want to keep doing this to save memory and do refcnt based
 609          * recycling. Support this use case too, to ease drivers
 610          * switching between XDP/non-XDP.
 611          *
 612          * In-case page_pool maintains the DMA mapping, API user must
 613          * call page_pool_put_page once.  In this elevated refcnt
 614          * case, the DMA is unmapped/released, as driver is likely
 615          * doing refcnt based recycle tricks, meaning another process
 616          * will be invoking put_page.
 617          */
 618         recycle_stat_inc(pool, released_refcnt);
 619         /* Do not replace this with page_pool_return_page() */
 620         page_pool_release_page(pool, page);
 621         put_page(page);
 622
 623         return NULL;
 624 }
 625
 626 void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
 627                                   unsigned int dma_sync_size, bool allow_direct)
 628 {
 629         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 630         if (page && !page_pool_recycle_in_ring(pool, page)) {
 631                 /* Cache full, fallback to free pages */
 632                 recycle_stat_inc(pool, ring_full);
 633                 page_pool_return_page(pool, page);
 634         }
 635 }
 636 EXPORT_SYMBOL(page_pool_put_defragged_page);
 637
 638 /* Caller must not use data area after call, as this function overwrites it */
 639 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 640                              int count)
 641 {
 642         int i, bulk_len = 0;
 643         bool in_softirq;
 644
 645         for (i = 0; i < count; i++) {
 646                 struct page *page = virt_to_head_page(data[i]);
 647
 648                 /* It is not the last user for the page frag case */
 649                 if (!page_pool_is_last_frag(pool, page))
 650                         continue;
 651
 652                 page = __page_pool_put_page(pool, page, -1, false);
 653                 /* Approved for bulk recycling in ptr_ring cache */
 654                 if (page)
 655                         data[bulk_len++] = page;
 656         }
 657
 658         if (unlikely(!bulk_len))
 659                 return;
 660
 661         /* Bulk producer into ptr_ring page_pool cache */
 662         in_softirq = page_pool_producer_lock(pool);
 663         for (i = 0; i < bulk_len; i++) {
 664                 if (__ptr_ring_produce(&pool->ring, data[i])) {
 665                         /* ring full */
 666                         recycle_stat_inc(pool, ring_full);
 667                         break;
 668                 }
 669         }
 670         recycle_stat_add(pool, ring, i);
 671         page_pool_producer_unlock(pool, in_softirq);
 672
 673         /* Hopefully all pages was return into ptr_ring */
 674         if (likely(i == bulk_len))
 675                 return;
 676
 677         /* ptr_ring cache full, free remaining pages outside producer lock
 678          * since put_page() with refcnt == 1 can be an expensive operation
 679          */
 680         for (; i < bulk_len; i++)
 681                 page_pool_return_page(pool, data[i]);
 682 }
 683 EXPORT_SYMBOL(page_pool_put_page_bulk);
 684
 685 static struct page *page_pool_drain_frag(struct page_pool *pool,
 686                                          struct page *page)
 687 {
 688         long drain_count = BIAS_MAX - pool->frag_users;
 689
 690         /* Some user is still using the page frag */
 691         if (likely(page_pool_defrag_page(page, drain_count)))
 692                 return NULL;
 693
 694         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 695                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 696                         page_pool_dma_sync_for_device(pool, page, -1);
 697
 698                 return page;
 699         }
 700
 701         page_pool_return_page(pool, page);
 702         return NULL;
 703 }
 704
 705 static void page_pool_free_frag(struct page_pool *pool)
 706 {
 707         long drain_count = BIAS_MAX - pool->frag_users;
 708         struct page *page = pool->frag_page;
 709
 710         pool->frag_page = NULL;
 711
 712         if (!page || page_pool_defrag_page(page, drain_count))
 713                 return;
 714
 715         page_pool_return_page(pool, page);
 716 }
 717
 718 struct page *page_pool_alloc_frag(struct page_pool *pool,
 719                                   unsigned int *offset,
 720                                   unsigned int size, gfp_t gfp)
 721 {
 722         unsigned int max_size = PAGE_SIZE << pool->p.order;
 723         struct page *page = pool->frag_page;
 724
 725         if (WARN_ON(!(pool->p.flags & PP_FLAG_PAGE_FRAG) ||
 726                     size > max_size))
 727                 return NULL;
 728
 729         size = ALIGN(size, dma_get_cache_alignment());
 730         *offset = pool->frag_offset;
 731
 732         if (page && *offset + size > max_size) {
 733                 page = page_pool_drain_frag(pool, page);
 734                 if (page) {
 735                         alloc_stat_inc(pool, fast);
 736                         goto frag_reset;
 737                 }
 738         }
 739
 740         if (!page) {
 741                 page = page_pool_alloc_pages(pool, gfp);
 742                 if (unlikely(!page)) {
 743                         pool->frag_page = NULL;
 744                         return NULL;
 745                 }
 746
 747                 pool->frag_page = page;
 748
 749 frag_reset:
 750                 pool->frag_users = 1;
 751                 *offset = 0;
 752                 pool->frag_offset = size;
 753                 page_pool_fragment_page(page, BIAS_MAX);
 754                 return page;
 755         }
 756
 757         pool->frag_users++;
 758         pool->frag_offset = *offset + size;
 759         alloc_stat_inc(pool, fast);
 760         return page;
 761 }
 762 EXPORT_SYMBOL(page_pool_alloc_frag);
 763
 764 static void page_pool_empty_ring(struct page_pool *pool)
 765 {
 766         struct page *page;
 767
 768         /* Empty recycle ring */
 769         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 770                 /* Verify the refcnt invariant of cached pages */
 771                 if (!(page_ref_count(page) == 1))
 772                         pr_crit("%s() page_pool refcnt %d violation\n",
 773                                 __func__, page_ref_count(page));
 774
 775                 page_pool_return_page(pool, page);
 776         }
 777 }
 778
 779 static void page_pool_free(struct page_pool *pool)
 780 {
 781         if (pool->disconnect)
 782                 pool->disconnect(pool);
 783
 784         ptr_ring_cleanup(&pool->ring, NULL);
 785
 786         if (pool->p.flags & PP_FLAG_DMA_MAP)
 787                 put_device(pool->p.dev);
 788
 789 #ifdef CONFIG_PAGE_POOL_STATS
 790         free_percpu(pool->recycle_stats);
 791 #endif
 792         kfree(pool);
 793 }
 794
 795 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 796 {
 797         struct page *page;
 798
 799         if (pool->destroy_cnt)
 800                 return;
 801
 802         /* Empty alloc cache, assume caller made sure this is
 803          * no-longer in use, and page_pool_alloc_pages() cannot be
 804          * call concurrently.
 805          */
 806         while (pool->alloc.count) {
 807                 page = pool->alloc.cache[--pool->alloc.count];
 808                 page_pool_return_page(pool, page);
 809         }
 810 }
 811
 812 static void page_pool_scrub(struct page_pool *pool)
 813 {
 814         page_pool_empty_alloc_cache_once(pool);
 815         pool->destroy_cnt++;
 816
 817         /* No more consumers should exist, but producers could still
 818          * be in-flight.
 819          */
 820         page_pool_empty_ring(pool);
 821 }
 822
 823 static int page_pool_release(struct page_pool *pool)
 824 {
 825         int inflight;
 826
 827         page_pool_scrub(pool);
 828         inflight = page_pool_inflight(pool);
 829         if (!inflight)
 830                 page_pool_free(pool);
 831
 832         return inflight;
 833 }
 834
 835 static void page_pool_release_retry(struct work_struct *wq)
 836 {
 837         struct delayed_work *dwq = to_delayed_work(wq);
 838         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 839         int inflight;
 840
 841         inflight = page_pool_release(pool);
 842         if (!inflight)
 843                 return;
 844
 845         /* Periodic warning */
 846         if (time_after_eq(jiffies, pool->defer_warn)) {
 847                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 848
 849                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 850                         __func__, inflight, sec);
 851                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 852         }
 853
 854         /* Still not ready to be disconnected, retry later */
 855         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 856 }
 857
 858 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
 859                            struct xdp_mem_info *mem)
 860 {
 861         refcount_inc(&pool->user_cnt);
 862         pool->disconnect = disconnect;
 863         pool->xdp_mem_id = mem->id;
 864 }
 865
 866 void page_pool_unlink_napi(struct page_pool *pool)
 867 {
 868         if (!pool->p.napi)
 869                 return;
 870
 871         /* To avoid races with recycling and additional barriers make sure
 872          * pool and NAPI are unlinked when NAPI is disabled.
 873          */
 874         WARN_ON(!test_bit(NAPI_STATE_SCHED, &pool->p.napi->state) ||
 875                 READ_ONCE(pool->p.napi->list_owner) != -1);
 876
 877         WRITE_ONCE(pool->p.napi, NULL);
 878 }
 879 EXPORT_SYMBOL(page_pool_unlink_napi);
 880
 881 void page_pool_destroy(struct page_pool *pool)
 882 {
 883         if (!pool)
 884                 return;
 885
 886         if (!page_pool_put(pool))
 887                 return;
 888
 889         page_pool_unlink_napi(pool);
 890         page_pool_free_frag(pool);
 891
 892         if (!page_pool_release(pool))
 893                 return;
 894
 895         pool->defer_start = jiffies;
 896         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 897
 898         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 899         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 900 }
 901 EXPORT_SYMBOL(page_pool_destroy);
 902
 903 /* Caller must provide appropriate safe context, e.g. NAPI. */
 904 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 905 {
 906         struct page *page;
 907
 908         trace_page_pool_update_nid(pool, new_nid);
 909         pool->p.nid = new_nid;
 910
 911         /* Flush pool alloc cache, as refill will check NUMA node */
 912         while (pool->alloc.count) {
 913                 page = pool->alloc.cache[--pool->alloc.count];
 914                 page_pool_return_page(pool, page);
 915         }
 916 }
 917 EXPORT_SYMBOL(page_pool_update_nid);
 918
 919 bool page_pool_return_skb_page(struct page *page, bool napi_safe)
 920 {
 921         struct napi_struct *napi;
 922         struct page_pool *pp;
 923         bool allow_direct;
 924
 925         page = compound_head(page);
 926
 927         /* page->pp_magic is OR'ed with PP_SIGNATURE after the allocation
 928          * in order to preserve any existing bits, such as bit 0 for the
 929          * head page of compound page and bit 1 for pfmemalloc page, so
 930          * mask those bits for freeing side when doing below checking,
 931          * and page_is_pfmemalloc() is checked in __page_pool_put_page()
 932          * to avoid recycling the pfmemalloc page.
 933          */
 934         if (unlikely((page->pp_magic & ~0x3UL) != PP_SIGNATURE))
 935                 return false;
 936
 937         pp = page->pp;
 938
 939         /* Allow direct recycle if we have reasons to believe that we are
 940          * in the same context as the consumer would run, so there's
 941          * no possible race.
 942          */
 943         napi = READ_ONCE(pp->p.napi);
 944         allow_direct = napi_safe && napi &&
 945                 READ_ONCE(napi->list_owner) == smp_processor_id();
 946
 947         /* Driver set this to memory recycling info. Reset it on recycle.
 948          * This will *not* work for NIC using a split-page memory model.
 949          * The page will be returned to the pool here regardless of the
 950          * 'flipped' fragment being in use or not.
 951          */
 952         page_pool_put_full_page(pp, page, allow_direct);
 953
 954         return true;
 955 }
 956 EXPORT_SYMBOL(page_pool_return_skb_page);