git.itanic.dy.fi Git - linux-stable/blob - 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
  23 #include <trace/events/page_pool.h>
  24
  25 #define DEFER_TIME (msecs_to_jiffies(1000))
  26 #define DEFER_WARN_INTERVAL (60 * HZ)
  27
  28 #define BIAS_MAX        LONG_MAX
  29
  30 #ifdef CONFIG_PAGE_POOL_STATS
  31 /* alloc_stat_inc is intended to be used in softirq context */
  32 #define alloc_stat_inc(pool, __stat)    (pool->alloc_stats.__stat++)
  33 /* recycle_stat_inc is safe to use when preemption is possible. */
  34 #define recycle_stat_inc(pool, __stat)                                                  \
  35         do {                                                                            \
  36                 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats;       \
  37                 this_cpu_inc(s->__stat);                                                \
  38         } while (0)
  39
  40 #define recycle_stat_add(pool, __stat, val)                                             \
  41         do {                                                                            \
  42                 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats;       \
  43                 this_cpu_add(s->__stat, val);                                           \
  44         } while (0)
  45
  46 static const char pp_stats[][ETH_GSTRING_LEN] = {
  47         "rx_pp_alloc_fast",
  48         "rx_pp_alloc_slow",
  49         "rx_pp_alloc_slow_ho",
  50         "rx_pp_alloc_empty",
  51         "rx_pp_alloc_refill",
  52         "rx_pp_alloc_waive",
  53         "rx_pp_recycle_cached",
  54         "rx_pp_recycle_cache_full",
  55         "rx_pp_recycle_ring",
  56         "rx_pp_recycle_ring_full",
  57         "rx_pp_recycle_released_ref",
  58 };
  59
  60 bool page_pool_get_stats(struct page_pool *pool,
  61                          struct page_pool_stats *stats)
  62 {
  63         int cpu = 0;
  64
  65         if (!stats)
  66                 return false;
  67
  68         /* The caller is responsible to initialize stats. */
  69         stats->alloc_stats.fast += pool->alloc_stats.fast;
  70         stats->alloc_stats.slow += pool->alloc_stats.slow;
  71         stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
  72         stats->alloc_stats.empty += pool->alloc_stats.empty;
  73         stats->alloc_stats.refill += pool->alloc_stats.refill;
  74         stats->alloc_stats.waive += pool->alloc_stats.waive;
  75
  76         for_each_possible_cpu(cpu) {
  77                 const struct page_pool_recycle_stats *pcpu =
  78                         per_cpu_ptr(pool->recycle_stats, cpu);
  79
  80                 stats->recycle_stats.cached += pcpu->cached;
  81                 stats->recycle_stats.cache_full += pcpu->cache_full;
  82                 stats->recycle_stats.ring += pcpu->ring;
  83                 stats->recycle_stats.ring_full += pcpu->ring_full;
  84                 stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
  85         }
  86
  87         return true;
  88 }
  89 EXPORT_SYMBOL(page_pool_get_stats);
  90
  91 u8 *page_pool_ethtool_stats_get_strings(u8 *data)
  92 {
  93         int i;
  94
  95         for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
  96                 memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
  97                 data += ETH_GSTRING_LEN;
  98         }
  99
 100         return data;
 101 }
 102 EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);
 103
 104 int page_pool_ethtool_stats_get_count(void)
 105 {
 106         return ARRAY_SIZE(pp_stats);
 107 }
 108 EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);
 109
 110 u64 *page_pool_ethtool_stats_get(u64 *data, void *stats)
 111 {
 112         struct page_pool_stats *pool_stats = stats;
 113
 114         *data++ = pool_stats->alloc_stats.fast;
 115         *data++ = pool_stats->alloc_stats.slow;
 116         *data++ = pool_stats->alloc_stats.slow_high_order;
 117         *data++ = pool_stats->alloc_stats.empty;
 118         *data++ = pool_stats->alloc_stats.refill;
 119         *data++ = pool_stats->alloc_stats.waive;
 120         *data++ = pool_stats->recycle_stats.cached;
 121         *data++ = pool_stats->recycle_stats.cache_full;
 122         *data++ = pool_stats->recycle_stats.ring;
 123         *data++ = pool_stats->recycle_stats.ring_full;
 124         *data++ = pool_stats->recycle_stats.released_refcnt;
 125
 126         return data;
 127 }
 128 EXPORT_SYMBOL(page_pool_ethtool_stats_get);
 129
 130 #else
 131 #define alloc_stat_inc(pool, __stat)
 132 #define recycle_stat_inc(pool, __stat)
 133 #define recycle_stat_add(pool, __stat, val)
 134 #endif
 135
 136 static bool page_pool_producer_lock(struct page_pool *pool)
 137         __acquires(&pool->ring.producer_lock)
 138 {
 139         bool in_softirq = in_softirq();
 140
 141         if (in_softirq)
 142                 spin_lock(&pool->ring.producer_lock);
 143         else
 144                 spin_lock_bh(&pool->ring.producer_lock);
 145
 146         return in_softirq;
 147 }
 148
 149 static void page_pool_producer_unlock(struct page_pool *pool,
 150                                       bool in_softirq)
 151         __releases(&pool->ring.producer_lock)
 152 {
 153         if (in_softirq)
 154                 spin_unlock(&pool->ring.producer_lock);
 155         else
 156                 spin_unlock_bh(&pool->ring.producer_lock);
 157 }
 158
 159 static int page_pool_init(struct page_pool *pool,
 160                           const struct page_pool_params *params)
 161 {
 162         unsigned int ring_qsize = 1024; /* Default */
 163
 164         memcpy(&pool->p, params, sizeof(pool->p));
 165
 166         /* Validate only known flags were used */
 167         if (pool->p.flags & ~(PP_FLAG_ALL))
 168                 return -EINVAL;
 169
 170         if (pool->p.pool_size)
 171                 ring_qsize = pool->p.pool_size;
 172
 173         /* Sanity limit mem that can be pinned down */
 174         if (ring_qsize > 32768)
 175                 return -E2BIG;
 176
 177         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
 178          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
 179          * which is the XDP_TX use-case.
 180          */
 181         if (pool->p.flags & PP_FLAG_DMA_MAP) {
 182                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
 183                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
 184                         return -EINVAL;
 185         }
 186
 187         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
 188                 /* In order to request DMA-sync-for-device the page
 189                  * needs to be mapped
 190                  */
 191                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 192                         return -EINVAL;
 193
 194                 if (!pool->p.max_len)
 195                         return -EINVAL;
 196
 197                 /* pool->p.offset has to be set according to the address
 198                  * offset used by the DMA engine to start copying rx data
 199                  */
 200         }
 201
 202         if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT &&
 203             pool->p.flags & PP_FLAG_PAGE_FRAG)
 204                 return -EINVAL;
 205
 206 #ifdef CONFIG_PAGE_POOL_STATS
 207         pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
 208         if (!pool->recycle_stats)
 209                 return -ENOMEM;
 210 #endif
 211
 212         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
 213                 return -ENOMEM;
 214
 215         atomic_set(&pool->pages_state_release_cnt, 0);
 216
 217         /* Driver calling page_pool_create() also call page_pool_destroy() */
 218         refcount_set(&pool->user_cnt, 1);
 219
 220         if (pool->p.flags & PP_FLAG_DMA_MAP)
 221                 get_device(pool->p.dev);
 222
 223         return 0;
 224 }
 225
 226 struct page_pool *page_pool_create(const struct page_pool_params *params)
 227 {
 228         struct page_pool *pool;
 229         int err;
 230
 231         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
 232         if (!pool)
 233                 return ERR_PTR(-ENOMEM);
 234
 235         err = page_pool_init(pool, params);
 236         if (err < 0) {
 237                 pr_warn("%s() gave up with errno %d\n", __func__, err);
 238                 kfree(pool);
 239                 return ERR_PTR(err);
 240         }
 241
 242         return pool;
 243 }
 244 EXPORT_SYMBOL(page_pool_create);
 245
 246 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 247
 248 noinline
 249 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 250 {
 251         struct ptr_ring *r = &pool->ring;
 252         struct page *page;
 253         int pref_nid; /* preferred NUMA node */
 254
 255         /* Quicker fallback, avoid locks when ring is empty */
 256         if (__ptr_ring_empty(r)) {
 257                 alloc_stat_inc(pool, empty);
 258                 return NULL;
 259         }
 260
 261         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 262          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 263          */
 264 #ifdef CONFIG_NUMA
 265         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 266 #else
 267         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 268         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 269 #endif
 270
 271         /* Refill alloc array, but only if NUMA match */
 272         do {
 273                 page = __ptr_ring_consume(r);
 274                 if (unlikely(!page))
 275                         break;
 276
 277                 if (likely(page_to_nid(page) == pref_nid)) {
 278                         pool->alloc.cache[pool->alloc.count++] = page;
 279                 } else {
 280                         /* NUMA mismatch;
 281                          * (1) release 1 page to page-allocator and
 282                          * (2) break out to fallthrough to alloc_pages_node.
 283                          * This limit stress on page buddy alloactor.
 284                          */
 285                         page_pool_return_page(pool, page);
 286                         alloc_stat_inc(pool, waive);
 287                         page = NULL;
 288                         break;
 289                 }
 290         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 291
 292         /* Return last page */
 293         if (likely(pool->alloc.count > 0)) {
 294                 page = pool->alloc.cache[--pool->alloc.count];
 295                 alloc_stat_inc(pool, refill);
 296         }
 297
 298         return page;
 299 }
 300
 301 /* fast path */
 302 static struct page *__page_pool_get_cached(struct page_pool *pool)
 303 {
 304         struct page *page;
 305
 306         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 307         if (likely(pool->alloc.count)) {
 308                 /* Fast-path */
 309                 page = pool->alloc.cache[--pool->alloc.count];
 310                 alloc_stat_inc(pool, fast);
 311         } else {
 312                 page = page_pool_refill_alloc_cache(pool);
 313         }
 314
 315         return page;
 316 }
 317
 318 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 319                                           struct page *page,
 320                                           unsigned int dma_sync_size)
 321 {
 322         dma_addr_t dma_addr = page_pool_get_dma_addr(page);
 323
 324         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 325         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 326                                          pool->p.offset, dma_sync_size,
 327                                          pool->p.dma_dir);
 328 }
 329
 330 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 331 {
 332         dma_addr_t dma;
 333
 334         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 335          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 336          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 337          * This mapping is kept for lifetime of page, until leaving pool.
 338          */
 339         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 340                                  (PAGE_SIZE << pool->p.order),
 341                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 342         if (dma_mapping_error(pool->p.dev, dma))
 343                 return false;
 344
 345         page_pool_set_dma_addr(page, dma);
 346
 347         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 348                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 349
 350         return true;
 351 }
 352
 353 static void page_pool_set_pp_info(struct page_pool *pool,
 354                                   struct page *page)
 355 {
 356         page->pp = pool;
 357         page->pp_magic |= PP_SIGNATURE;
 358         if (pool->p.init_callback)
 359                 pool->p.init_callback(page, pool->p.init_arg);
 360 }
 361
 362 static void page_pool_clear_pp_info(struct page *page)
 363 {
 364         page->pp_magic = 0;
 365         page->pp = NULL;
 366 }
 367
 368 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 369                                                  gfp_t gfp)
 370 {
 371         struct page *page;
 372
 373         gfp |= __GFP_COMP;
 374         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 375         if (unlikely(!page))
 376                 return NULL;
 377
 378         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 379             unlikely(!page_pool_dma_map(pool, page))) {
 380                 put_page(page);
 381                 return NULL;
 382         }
 383
 384         alloc_stat_inc(pool, slow_high_order);
 385         page_pool_set_pp_info(pool, page);
 386
 387         /* Track how many pages are held 'in-flight' */
 388         pool->pages_state_hold_cnt++;
 389         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 390         return page;
 391 }
 392
 393 /* slow path */
 394 noinline
 395 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 396                                                  gfp_t gfp)
 397 {
 398         const int bulk = PP_ALLOC_CACHE_REFILL;
 399         unsigned int pp_flags = pool->p.flags;
 400         unsigned int pp_order = pool->p.order;
 401         struct page *page;
 402         int i, nr_pages;
 403
 404         /* Don't support bulk alloc for high-order pages */
 405         if (unlikely(pp_order))
 406                 return __page_pool_alloc_page_order(pool, gfp);
 407
 408         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 409         if (unlikely(pool->alloc.count > 0))
 410                 return pool->alloc.cache[--pool->alloc.count];
 411
 412         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 413         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 414
 415         nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk,
 416                                                pool->alloc.cache);
 417         if (unlikely(!nr_pages))
 418                 return NULL;
 419
 420         /* Pages have been filled into alloc.cache array, but count is zero and
 421          * page element have not been (possibly) DMA mapped.
 422          */
 423         for (i = 0; i < nr_pages; i++) {
 424                 page = pool->alloc.cache[i];
 425                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 426                     unlikely(!page_pool_dma_map(pool, page))) {
 427                         put_page(page);
 428                         continue;
 429                 }
 430
 431                 page_pool_set_pp_info(pool, page);
 432                 pool->alloc.cache[pool->alloc.count++] = page;
 433                 /* Track how many pages are held 'in-flight' */
 434                 pool->pages_state_hold_cnt++;
 435                 trace_page_pool_state_hold(pool, page,
 436                                            pool->pages_state_hold_cnt);
 437         }
 438
 439         /* Return last page */
 440         if (likely(pool->alloc.count > 0)) {
 441                 page = pool->alloc.cache[--pool->alloc.count];
 442                 alloc_stat_inc(pool, slow);
 443         } else {
 444                 page = NULL;
 445         }
 446
 447         /* When page just alloc'ed is should/must have refcnt 1. */
 448         return page;
 449 }
 450
 451 /* For using page_pool replace: alloc_pages() API calls, but provide
 452  * synchronization guarantee for allocation side.
 453  */
 454 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 455 {
 456         struct page *page;
 457
 458         /* Fast-path: Get a page from cache */
 459         page = __page_pool_get_cached(pool);
 460         if (page)
 461                 return page;
 462
 463         /* Slow-path: cache empty, do real allocation */
 464         page = __page_pool_alloc_pages_slow(pool, gfp);
 465         return page;
 466 }
 467 EXPORT_SYMBOL(page_pool_alloc_pages);
 468
 469 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 470  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 471  */
 472 #define _distance(a, b) (s32)((a) - (b))
 473
 474 static s32 page_pool_inflight(struct page_pool *pool)
 475 {
 476         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 477         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 478         s32 inflight;
 479
 480         inflight = _distance(hold_cnt, release_cnt);
 481
 482         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 483         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 484
 485         return inflight;
 486 }
 487
 488 /* Disconnects a page (from a page_pool).  API users can have a need
 489  * to disconnect a page (from a page_pool), to allow it to be used as
 490  * a regular page (that will eventually be returned to the normal
 491  * page-allocator via put_page).
 492  */
 493 void page_pool_release_page(struct page_pool *pool, struct page *page)
 494 {
 495         dma_addr_t dma;
 496         int count;
 497
 498         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 499                 /* Always account for inflight pages, even if we didn't
 500                  * map them
 501                  */
 502                 goto skip_dma_unmap;
 503
 504         dma = page_pool_get_dma_addr(page);
 505
 506         /* When page is unmapped, it cannot be returned to our pool */
 507         dma_unmap_page_attrs(pool->p.dev, dma,
 508                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 509                              DMA_ATTR_SKIP_CPU_SYNC);
 510         page_pool_set_dma_addr(page, 0);
 511 skip_dma_unmap:
 512         page_pool_clear_pp_info(page);
 513
 514         /* This may be the last page returned, releasing the pool, so
 515          * it is not safe to reference pool afterwards.
 516          */
 517         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 518         trace_page_pool_state_release(pool, page, count);
 519 }
 520 EXPORT_SYMBOL(page_pool_release_page);
 521
 522 /* Return a page to the page allocator, cleaning up our state */
 523 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 524 {
 525         page_pool_release_page(pool, page);
 526
 527         put_page(page);
 528         /* An optimization would be to call __free_pages(page, pool->p.order)
 529          * knowing page is not part of page-cache (thus avoiding a
 530          * __page_cache_release() call).
 531          */
 532 }
 533
 534 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 535 {
 536         int ret;
 537         /* BH protection not needed if current is softirq */
 538         if (in_softirq())
 539                 ret = ptr_ring_produce(&pool->ring, page);
 540         else
 541                 ret = ptr_ring_produce_bh(&pool->ring, page);
 542
 543         if (!ret) {
 544                 recycle_stat_inc(pool, ring);
 545                 return true;
 546         }
 547
 548         return false;
 549 }
 550
 551 /* Only allow direct recycling in special circumstances, into the
 552  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 553  *
 554  * Caller must provide appropriate safe context.
 555  */
 556 static bool page_pool_recycle_in_cache(struct page *page,
 557                                        struct page_pool *pool)
 558 {
 559         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
 560                 recycle_stat_inc(pool, cache_full);
 561                 return false;
 562         }
 563
 564         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 565         pool->alloc.cache[pool->alloc.count++] = page;
 566         recycle_stat_inc(pool, cached);
 567         return true;
 568 }
 569
 570 /* If the page refcnt == 1, this will try to recycle the page.
 571  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 572  * the configured size min(dma_sync_size, pool->max_len).
 573  * If the page refcnt != 1, then the page will be returned to memory
 574  * subsystem.
 575  */
 576 static __always_inline struct page *
 577 __page_pool_put_page(struct page_pool *pool, struct page *page,
 578                      unsigned int dma_sync_size, bool allow_direct)
 579 {
 580         /* This allocator is optimized for the XDP mode that uses
 581          * one-frame-per-page, but have fallbacks that act like the
 582          * regular page allocator APIs.
 583          *
 584          * refcnt == 1 means page_pool owns page, and can recycle it.
 585          *
 586          * page is NOT reusable when allocated when system is under
 587          * some pressure. (page_is_pfmemalloc)
 588          */
 589         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 590                 /* Read barrier done in page_ref_count / READ_ONCE */
 591
 592                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 593                         page_pool_dma_sync_for_device(pool, page,
 594                                                       dma_sync_size);
 595
 596                 if (allow_direct && in_softirq() &&
 597                     page_pool_recycle_in_cache(page, pool))
 598                         return NULL;
 599
 600                 /* Page found as candidate for recycling */
 601                 return page;
 602         }
 603         /* Fallback/non-XDP mode: API user have elevated refcnt.
 604          *
 605          * Many drivers split up the page into fragments, and some
 606          * want to keep doing this to save memory and do refcnt based
 607          * recycling. Support this use case too, to ease drivers
 608          * switching between XDP/non-XDP.
 609          *
 610          * In-case page_pool maintains the DMA mapping, API user must
 611          * call page_pool_put_page once.  In this elevated refcnt
 612          * case, the DMA is unmapped/released, as driver is likely
 613          * doing refcnt based recycle tricks, meaning another process
 614          * will be invoking put_page.
 615          */
 616         recycle_stat_inc(pool, released_refcnt);
 617         /* Do not replace this with page_pool_return_page() */
 618         page_pool_release_page(pool, page);
 619         put_page(page);
 620
 621         return NULL;
 622 }
 623
 624 void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
 625                                   unsigned int dma_sync_size, bool allow_direct)
 626 {
 627         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 628         if (page && !page_pool_recycle_in_ring(pool, page)) {
 629                 /* Cache full, fallback to free pages */
 630                 recycle_stat_inc(pool, ring_full);
 631                 page_pool_return_page(pool, page);
 632         }
 633 }
 634 EXPORT_SYMBOL(page_pool_put_defragged_page);
 635
 636 /* Caller must not use data area after call, as this function overwrites it */
 637 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 638                              int count)
 639 {
 640         int i, bulk_len = 0;
 641         bool in_softirq;
 642
 643         for (i = 0; i < count; i++) {
 644                 struct page *page = virt_to_head_page(data[i]);
 645
 646                 /* It is not the last user for the page frag case */
 647                 if (!page_pool_is_last_frag(pool, page))
 648                         continue;
 649
 650                 page = __page_pool_put_page(pool, page, -1, false);
 651                 /* Approved for bulk recycling in ptr_ring cache */
 652                 if (page)
 653                         data[bulk_len++] = page;
 654         }
 655
 656         if (unlikely(!bulk_len))
 657                 return;
 658
 659         /* Bulk producer into ptr_ring page_pool cache */
 660         in_softirq = page_pool_producer_lock(pool);
 661         for (i = 0; i < bulk_len; i++) {
 662                 if (__ptr_ring_produce(&pool->ring, data[i])) {
 663                         /* ring full */
 664                         recycle_stat_inc(pool, ring_full);
 665                         break;
 666                 }
 667         }
 668         recycle_stat_add(pool, ring, i);
 669         page_pool_producer_unlock(pool, in_softirq);
 670
 671         /* Hopefully all pages was return into ptr_ring */
 672         if (likely(i == bulk_len))
 673                 return;
 674
 675         /* ptr_ring cache full, free remaining pages outside producer lock
 676          * since put_page() with refcnt == 1 can be an expensive operation
 677          */
 678         for (; i < bulk_len; i++)
 679                 page_pool_return_page(pool, data[i]);
 680 }
 681 EXPORT_SYMBOL(page_pool_put_page_bulk);
 682
 683 static struct page *page_pool_drain_frag(struct page_pool *pool,
 684                                          struct page *page)
 685 {
 686         long drain_count = BIAS_MAX - pool->frag_users;
 687
 688         /* Some user is still using the page frag */
 689         if (likely(page_pool_defrag_page(page, drain_count)))
 690                 return NULL;
 691
 692         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 693                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 694                         page_pool_dma_sync_for_device(pool, page, -1);
 695
 696                 return page;
 697         }
 698
 699         page_pool_return_page(pool, page);
 700         return NULL;
 701 }
 702
 703 static void page_pool_free_frag(struct page_pool *pool)
 704 {
 705         long drain_count = BIAS_MAX - pool->frag_users;
 706         struct page *page = pool->frag_page;
 707
 708         pool->frag_page = NULL;
 709
 710         if (!page || page_pool_defrag_page(page, drain_count))
 711                 return;
 712
 713         page_pool_return_page(pool, page);
 714 }
 715
 716 struct page *page_pool_alloc_frag(struct page_pool *pool,
 717                                   unsigned int *offset,
 718                                   unsigned int size, gfp_t gfp)
 719 {
 720         unsigned int max_size = PAGE_SIZE << pool->p.order;
 721         struct page *page = pool->frag_page;
 722
 723         if (WARN_ON(!(pool->p.flags & PP_FLAG_PAGE_FRAG) ||
 724                     size > max_size))
 725                 return NULL;
 726
 727         size = ALIGN(size, dma_get_cache_alignment());
 728         *offset = pool->frag_offset;
 729
 730         if (page && *offset + size > max_size) {
 731                 page = page_pool_drain_frag(pool, page);
 732                 if (page) {
 733                         alloc_stat_inc(pool, fast);
 734                         goto frag_reset;
 735                 }
 736         }
 737
 738         if (!page) {
 739                 page = page_pool_alloc_pages(pool, gfp);
 740                 if (unlikely(!page)) {
 741                         pool->frag_page = NULL;
 742                         return NULL;
 743                 }
 744
 745                 pool->frag_page = page;
 746
 747 frag_reset:
 748                 pool->frag_users = 1;
 749                 *offset = 0;
 750                 pool->frag_offset = size;
 751                 page_pool_fragment_page(page, BIAS_MAX);
 752                 return page;
 753         }
 754
 755         pool->frag_users++;
 756         pool->frag_offset = *offset + size;
 757         alloc_stat_inc(pool, fast);
 758         return page;
 759 }
 760 EXPORT_SYMBOL(page_pool_alloc_frag);
 761
 762 static void page_pool_empty_ring(struct page_pool *pool)
 763 {
 764         struct page *page;
 765
 766         /* Empty recycle ring */
 767         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 768                 /* Verify the refcnt invariant of cached pages */
 769                 if (!(page_ref_count(page) == 1))
 770                         pr_crit("%s() page_pool refcnt %d violation\n",
 771                                 __func__, page_ref_count(page));
 772
 773                 page_pool_return_page(pool, page);
 774         }
 775 }
 776
 777 static void page_pool_free(struct page_pool *pool)
 778 {
 779         if (pool->disconnect)
 780                 pool->disconnect(pool);
 781
 782         ptr_ring_cleanup(&pool->ring, NULL);
 783
 784         if (pool->p.flags & PP_FLAG_DMA_MAP)
 785                 put_device(pool->p.dev);
 786
 787 #ifdef CONFIG_PAGE_POOL_STATS
 788         free_percpu(pool->recycle_stats);
 789 #endif
 790         kfree(pool);
 791 }
 792
 793 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 794 {
 795         struct page *page;
 796
 797         if (pool->destroy_cnt)
 798                 return;
 799
 800         /* Empty alloc cache, assume caller made sure this is
 801          * no-longer in use, and page_pool_alloc_pages() cannot be
 802          * call concurrently.
 803          */
 804         while (pool->alloc.count) {
 805                 page = pool->alloc.cache[--pool->alloc.count];
 806                 page_pool_return_page(pool, page);
 807         }
 808 }
 809
 810 static void page_pool_scrub(struct page_pool *pool)
 811 {
 812         page_pool_empty_alloc_cache_once(pool);
 813         pool->destroy_cnt++;
 814
 815         /* No more consumers should exist, but producers could still
 816          * be in-flight.
 817          */
 818         page_pool_empty_ring(pool);
 819 }
 820
 821 static int page_pool_release(struct page_pool *pool)
 822 {
 823         int inflight;
 824
 825         page_pool_scrub(pool);
 826         inflight = page_pool_inflight(pool);
 827         if (!inflight)
 828                 page_pool_free(pool);
 829
 830         return inflight;
 831 }
 832
 833 static void page_pool_release_retry(struct work_struct *wq)
 834 {
 835         struct delayed_work *dwq = to_delayed_work(wq);
 836         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 837         int inflight;
 838
 839         inflight = page_pool_release(pool);
 840         if (!inflight)
 841                 return;
 842
 843         /* Periodic warning */
 844         if (time_after_eq(jiffies, pool->defer_warn)) {
 845                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 846
 847                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 848                         __func__, inflight, sec);
 849                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 850         }
 851
 852         /* Still not ready to be disconnected, retry later */
 853         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 854 }
 855
 856 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
 857                            struct xdp_mem_info *mem)
 858 {
 859         refcount_inc(&pool->user_cnt);
 860         pool->disconnect = disconnect;
 861         pool->xdp_mem_id = mem->id;
 862 }
 863
 864 void page_pool_destroy(struct page_pool *pool)
 865 {
 866         if (!pool)
 867                 return;
 868
 869         if (!page_pool_put(pool))
 870                 return;
 871
 872         page_pool_free_frag(pool);
 873
 874         if (!page_pool_release(pool))
 875                 return;
 876
 877         pool->defer_start = jiffies;
 878         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 879
 880         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 881         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 882 }
 883 EXPORT_SYMBOL(page_pool_destroy);
 884
 885 /* Caller must provide appropriate safe context, e.g. NAPI. */
 886 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 887 {
 888         struct page *page;
 889
 890         trace_page_pool_update_nid(pool, new_nid);
 891         pool->p.nid = new_nid;
 892
 893         /* Flush pool alloc cache, as refill will check NUMA node */
 894         while (pool->alloc.count) {
 895                 page = pool->alloc.cache[--pool->alloc.count];
 896                 page_pool_return_page(pool, page);
 897         }
 898 }
 899 EXPORT_SYMBOL(page_pool_update_nid);
 900
 901 bool page_pool_return_skb_page(struct page *page)
 902 {
 903         struct page_pool *pp;
 904
 905         page = compound_head(page);
 906
 907         /* page->pp_magic is OR'ed with PP_SIGNATURE after the allocation
 908          * in order to preserve any existing bits, such as bit 0 for the
 909          * head page of compound page and bit 1 for pfmemalloc page, so
 910          * mask those bits for freeing side when doing below checking,
 911          * and page_is_pfmemalloc() is checked in __page_pool_put_page()
 912          * to avoid recycling the pfmemalloc page.
 913          */
 914         if (unlikely((page->pp_magic & ~0x3UL) != PP_SIGNATURE))
 915                 return false;
 916
 917         pp = page->pp;
 918
 919         /* Driver set this to memory recycling info. Reset it on recycle.
 920          * This will *not* work for NIC using a split-page memory model.
 921          * The page will be returned to the pool here regardless of the
 922          * 'flipped' fragment being in use or not.
 923          */
 924         page_pool_put_full_page(pp, page, false);
 925
 926         return true;
 927 }
 928 EXPORT_SYMBOL(page_pool_return_skb_page);