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