net/core/page_pool.c

   1 /* SPDX-License-Identifier: GPL-2.0
   2  *
   3  * page_pool.c
   4  *      Author: Jesper Dangaard Brouer <netoptimizer@brouer.com>
   5  *      Copyright (C) 2016 Red Hat, Inc.
   6  */
   7
   8 #include <linux/types.h>
   9 #include <linux/kernel.h>
  10 #include <linux/slab.h>
  11 #include <linux/device.h>
  12
  13 #include <net/page_pool/helpers.h>
  14 #include <net/xdp.h>
  15
  16 #include <linux/dma-direction.h>
  17 #include <linux/dma-mapping.h>
  18 #include <linux/page-flags.h>
  19 #include <linux/mm.h> /* for put_page() */
  20 #include <linux/poison.h>
  21 #include <linux/ethtool.h>
  22 #include <linux/netdevice.h>
  23
  24 #include <trace/events/page_pool.h>
  25
  26 #define DEFER_TIME (msecs_to_jiffies(1000))
  27 #define DEFER_WARN_INTERVAL (60 * HZ)
  28
  29 #define BIAS_MAX        LONG_MAX
  30
  31 #ifdef CONFIG_PAGE_POOL_STATS
  32 /* alloc_stat_inc is intended to be used in softirq context */
  33 #define alloc_stat_inc(pool, __stat)    (pool->alloc_stats.__stat++)
  34 /* recycle_stat_inc is safe to use when preemption is possible. */
  35 #define recycle_stat_inc(pool, __stat)                                                  \
  36         do {                                                                            \
  37                 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats;       \
  38                 this_cpu_inc(s->__stat);                                                \
  39         } while (0)
  40
  41 #define recycle_stat_add(pool, __stat, val)                                             \
  42         do {                                                                            \
  43                 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats;       \
  44                 this_cpu_add(s->__stat, val);                                           \
  45         } while (0)
  46
  47 static const char pp_stats[][ETH_GSTRING_LEN] = {
  48         "rx_pp_alloc_fast",
  49         "rx_pp_alloc_slow",
  50         "rx_pp_alloc_slow_ho",
  51         "rx_pp_alloc_empty",
  52         "rx_pp_alloc_refill",
  53         "rx_pp_alloc_waive",
  54         "rx_pp_recycle_cached",
  55         "rx_pp_recycle_cache_full",
  56         "rx_pp_recycle_ring",
  57         "rx_pp_recycle_ring_full",
  58         "rx_pp_recycle_released_ref",
  59 };
  60
  61 /**
  62  * page_pool_get_stats() - fetch page pool stats
  63  * @pool:       pool from which page was allocated
  64  * @stats:      struct page_pool_stats to fill in
  65  *
  66  * Retrieve statistics about the page_pool. This API is only available
  67  * if the kernel has been configured with ``CONFIG_PAGE_POOL_STATS=y``.
  68  * A pointer to a caller allocated struct page_pool_stats structure
  69  * is passed to this API which is filled in. The caller can then report
  70  * those stats to the user (perhaps via ethtool, debugfs, etc.).
  71  */
  72 bool page_pool_get_stats(struct page_pool *pool,
  73                          struct page_pool_stats *stats)
  74 {
  75         int cpu = 0;
  76
  77         if (!stats)
  78                 return false;
  79
  80         /* The caller is responsible to initialize stats. */
  81         stats->alloc_stats.fast += pool->alloc_stats.fast;
  82         stats->alloc_stats.slow += pool->alloc_stats.slow;
  83         stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
  84         stats->alloc_stats.empty += pool->alloc_stats.empty;
  85         stats->alloc_stats.refill += pool->alloc_stats.refill;
  86         stats->alloc_stats.waive += pool->alloc_stats.waive;
  87
  88         for_each_possible_cpu(cpu) {
  89                 const struct page_pool_recycle_stats *pcpu =
  90                         per_cpu_ptr(pool->recycle_stats, cpu);
  91
  92                 stats->recycle_stats.cached += pcpu->cached;
  93                 stats->recycle_stats.cache_full += pcpu->cache_full;
  94                 stats->recycle_stats.ring += pcpu->ring;
  95                 stats->recycle_stats.ring_full += pcpu->ring_full;
  96                 stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
  97         }
  98
  99         return true;
 100 }
 101 EXPORT_SYMBOL(page_pool_get_stats);
 102
 103 u8 *page_pool_ethtool_stats_get_strings(u8 *data)
 104 {
 105         int i;
 106
 107         for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
 108                 memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
 109                 data += ETH_GSTRING_LEN;
 110         }
 111
 112         return data;
 113 }
 114 EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);
 115
 116 int page_pool_ethtool_stats_get_count(void)
 117 {
 118         return ARRAY_SIZE(pp_stats);
 119 }
 120 EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);
 121
 122 u64 *page_pool_ethtool_stats_get(u64 *data, void *stats)
 123 {
 124         struct page_pool_stats *pool_stats = stats;
 125
 126         *data++ = pool_stats->alloc_stats.fast;
 127         *data++ = pool_stats->alloc_stats.slow;
 128         *data++ = pool_stats->alloc_stats.slow_high_order;
 129         *data++ = pool_stats->alloc_stats.empty;
 130         *data++ = pool_stats->alloc_stats.refill;
 131         *data++ = pool_stats->alloc_stats.waive;
 132         *data++ = pool_stats->recycle_stats.cached;
 133         *data++ = pool_stats->recycle_stats.cache_full;
 134         *data++ = pool_stats->recycle_stats.ring;
 135         *data++ = pool_stats->recycle_stats.ring_full;
 136         *data++ = pool_stats->recycle_stats.released_refcnt;
 137
 138         return data;
 139 }
 140 EXPORT_SYMBOL(page_pool_ethtool_stats_get);
 141
 142 #else
 143 #define alloc_stat_inc(pool, __stat)
 144 #define recycle_stat_inc(pool, __stat)
 145 #define recycle_stat_add(pool, __stat, val)
 146 #endif
 147
 148 static bool page_pool_producer_lock(struct page_pool *pool)
 149         __acquires(&pool->ring.producer_lock)
 150 {
 151         bool in_softirq = in_softirq();
 152
 153         if (in_softirq)
 154                 spin_lock(&pool->ring.producer_lock);
 155         else
 156                 spin_lock_bh(&pool->ring.producer_lock);
 157
 158         return in_softirq;
 159 }
 160
 161 static void page_pool_producer_unlock(struct page_pool *pool,
 162                                       bool in_softirq)
 163         __releases(&pool->ring.producer_lock)
 164 {
 165         if (in_softirq)
 166                 spin_unlock(&pool->ring.producer_lock);
 167         else
 168                 spin_unlock_bh(&pool->ring.producer_lock);
 169 }
 170
 171 static int page_pool_init(struct page_pool *pool,
 172                           const struct page_pool_params *params)
 173 {
 174         unsigned int ring_qsize = 1024; /* Default */
 175
 176         memcpy(&pool->p, params, sizeof(pool->p));
 177
 178         /* Validate only known flags were used */
 179         if (pool->p.flags & ~(PP_FLAG_ALL))
 180                 return -EINVAL;
 181
 182         if (pool->p.pool_size)
 183                 ring_qsize = pool->p.pool_size;
 184
 185         /* Sanity limit mem that can be pinned down */
 186         if (ring_qsize > 32768)
 187                 return -E2BIG;
 188
 189         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
 190          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
 191          * which is the XDP_TX use-case.
 192          */
 193         if (pool->p.flags & PP_FLAG_DMA_MAP) {
 194                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
 195                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
 196                         return -EINVAL;
 197         }
 198
 199         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
 200                 /* In order to request DMA-sync-for-device the page
 201                  * needs to be mapped
 202                  */
 203                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 204                         return -EINVAL;
 205
 206                 if (!pool->p.max_len)
 207                         return -EINVAL;
 208
 209                 /* pool->p.offset has to be set according to the address
 210                  * offset used by the DMA engine to start copying rx data
 211                  */
 212         }
 213
 214         if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT &&
 215             pool->p.flags & PP_FLAG_PAGE_FRAG)
 216                 return -EINVAL;
 217
 218 #ifdef CONFIG_PAGE_POOL_STATS
 219         pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
 220         if (!pool->recycle_stats)
 221                 return -ENOMEM;
 222 #endif
 223
 224         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0) {
 225 #ifdef CONFIG_PAGE_POOL_STATS
 226                 free_percpu(pool->recycle_stats);
 227 #endif
 228                 return -ENOMEM;
 229         }
 230
 231         atomic_set(&pool->pages_state_release_cnt, 0);
 232
 233         /* Driver calling page_pool_create() also call page_pool_destroy() */
 234         refcount_set(&pool->user_cnt, 1);
 235
 236         if (pool->p.flags & PP_FLAG_DMA_MAP)
 237                 get_device(pool->p.dev);
 238
 239         return 0;
 240 }
 241
 242 /**
 243  * page_pool_create() - create a page pool.
 244  * @params: parameters, see struct page_pool_params
 245  */
 246 struct page_pool *page_pool_create(const struct page_pool_params *params)
 247 {
 248         struct page_pool *pool;
 249         int err;
 250
 251         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
 252         if (!pool)
 253                 return ERR_PTR(-ENOMEM);
 254
 255         err = page_pool_init(pool, params);
 256         if (err < 0) {
 257                 pr_warn("%s() gave up with errno %d\n", __func__, err);
 258                 kfree(pool);
 259                 return ERR_PTR(err);
 260         }
 261
 262         return pool;
 263 }
 264 EXPORT_SYMBOL(page_pool_create);
 265
 266 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 267
 268 noinline
 269 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 270 {
 271         struct ptr_ring *r = &pool->ring;
 272         struct page *page;
 273         int pref_nid; /* preferred NUMA node */
 274
 275         /* Quicker fallback, avoid locks when ring is empty */
 276         if (__ptr_ring_empty(r)) {
 277                 alloc_stat_inc(pool, empty);
 278                 return NULL;
 279         }
 280
 281         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 282          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 283          */
 284 #ifdef CONFIG_NUMA
 285         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 286 #else
 287         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 288         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 289 #endif
 290
 291         /* Refill alloc array, but only if NUMA match */
 292         do {
 293                 page = __ptr_ring_consume(r);
 294                 if (unlikely(!page))
 295                         break;
 296
 297                 if (likely(page_to_nid(page) == pref_nid)) {
 298                         pool->alloc.cache[pool->alloc.count++] = page;
 299                 } else {
 300                         /* NUMA mismatch;
 301                          * (1) release 1 page to page-allocator and
 302                          * (2) break out to fallthrough to alloc_pages_node.
 303                          * This limit stress on page buddy alloactor.
 304                          */
 305                         page_pool_return_page(pool, page);
 306                         alloc_stat_inc(pool, waive);
 307                         page = NULL;
 308                         break;
 309                 }
 310         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 311
 312         /* Return last page */
 313         if (likely(pool->alloc.count > 0)) {
 314                 page = pool->alloc.cache[--pool->alloc.count];
 315                 alloc_stat_inc(pool, refill);
 316         }
 317
 318         return page;
 319 }
 320
 321 /* fast path */
 322 static struct page *__page_pool_get_cached(struct page_pool *pool)
 323 {
 324         struct page *page;
 325
 326         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 327         if (likely(pool->alloc.count)) {
 328                 /* Fast-path */
 329                 page = pool->alloc.cache[--pool->alloc.count];
 330                 alloc_stat_inc(pool, fast);
 331         } else {
 332                 page = page_pool_refill_alloc_cache(pool);
 333         }
 334
 335         return page;
 336 }
 337
 338 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 339                                           struct page *page,
 340                                           unsigned int dma_sync_size)
 341 {
 342         dma_addr_t dma_addr = page_pool_get_dma_addr(page);
 343
 344         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 345         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 346                                          pool->p.offset, dma_sync_size,
 347                                          pool->p.dma_dir);
 348 }
 349
 350 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 351 {
 352         dma_addr_t dma;
 353
 354         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 355          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 356          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 357          * This mapping is kept for lifetime of page, until leaving pool.
 358          */
 359         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 360                                  (PAGE_SIZE << pool->p.order),
 361                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC |
 362                                                   DMA_ATTR_WEAK_ORDERING);
 363         if (dma_mapping_error(pool->p.dev, dma))
 364                 return false;
 365
 366         page_pool_set_dma_addr(page, dma);
 367
 368         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 369                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 370
 371         return true;
 372 }
 373
 374 static void page_pool_set_pp_info(struct page_pool *pool,
 375                                   struct page *page)
 376 {
 377         page->pp = pool;
 378         page->pp_magic |= PP_SIGNATURE;
 379         if (pool->p.init_callback)
 380                 pool->p.init_callback(page, pool->p.init_arg);
 381 }
 382
 383 static void page_pool_clear_pp_info(struct page *page)
 384 {
 385         page->pp_magic = 0;
 386         page->pp = NULL;
 387 }
 388
 389 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 390                                                  gfp_t gfp)
 391 {
 392         struct page *page;
 393
 394         gfp |= __GFP_COMP;
 395         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 396         if (unlikely(!page))
 397                 return NULL;
 398
 399         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 400             unlikely(!page_pool_dma_map(pool, page))) {
 401                 put_page(page);
 402                 return NULL;
 403         }
 404
 405         alloc_stat_inc(pool, slow_high_order);
 406         page_pool_set_pp_info(pool, page);
 407
 408         /* Track how many pages are held 'in-flight' */
 409         pool->pages_state_hold_cnt++;
 410         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 411         return page;
 412 }
 413
 414 /* slow path */
 415 noinline
 416 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 417                                                  gfp_t gfp)
 418 {
 419         const int bulk = PP_ALLOC_CACHE_REFILL;
 420         unsigned int pp_flags = pool->p.flags;
 421         unsigned int pp_order = pool->p.order;
 422         struct page *page;
 423         int i, nr_pages;
 424
 425         /* Don't support bulk alloc for high-order pages */
 426         if (unlikely(pp_order))
 427                 return __page_pool_alloc_page_order(pool, gfp);
 428
 429         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 430         if (unlikely(pool->alloc.count > 0))
 431                 return pool->alloc.cache[--pool->alloc.count];
 432
 433         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 434         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 435
 436         nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk,
 437                                                pool->alloc.cache);
 438         if (unlikely(!nr_pages))
 439                 return NULL;
 440
 441         /* Pages have been filled into alloc.cache array, but count is zero and
 442          * page element have not been (possibly) DMA mapped.
 443          */
 444         for (i = 0; i < nr_pages; i++) {
 445                 page = pool->alloc.cache[i];
 446                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 447                     unlikely(!page_pool_dma_map(pool, page))) {
 448                         put_page(page);
 449                         continue;
 450                 }
 451
 452                 page_pool_set_pp_info(pool, page);
 453                 pool->alloc.cache[pool->alloc.count++] = page;
 454                 /* Track how many pages are held 'in-flight' */
 455                 pool->pages_state_hold_cnt++;
 456                 trace_page_pool_state_hold(pool, page,
 457                                            pool->pages_state_hold_cnt);
 458         }
 459
 460         /* Return last page */
 461         if (likely(pool->alloc.count > 0)) {
 462                 page = pool->alloc.cache[--pool->alloc.count];
 463                 alloc_stat_inc(pool, slow);
 464         } else {
 465                 page = NULL;
 466         }
 467
 468         /* When page just alloc'ed is should/must have refcnt 1. */
 469         return page;
 470 }
 471
 472 /* For using page_pool replace: alloc_pages() API calls, but provide
 473  * synchronization guarantee for allocation side.
 474  */
 475 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 476 {
 477         struct page *page;
 478
 479         /* Fast-path: Get a page from cache */
 480         page = __page_pool_get_cached(pool);
 481         if (page)
 482                 return page;
 483
 484         /* Slow-path: cache empty, do real allocation */
 485         page = __page_pool_alloc_pages_slow(pool, gfp);
 486         return page;
 487 }
 488 EXPORT_SYMBOL(page_pool_alloc_pages);
 489
 490 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 491  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 492  */
 493 #define _distance(a, b) (s32)((a) - (b))
 494
 495 static s32 page_pool_inflight(struct page_pool *pool)
 496 {
 497         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 498         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 499         s32 inflight;
 500
 501         inflight = _distance(hold_cnt, release_cnt);
 502
 503         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 504         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 505
 506         return inflight;
 507 }
 508
 509 /* Disconnects a page (from a page_pool).  API users can have a need
 510  * to disconnect a page (from a page_pool), to allow it to be used as
 511  * a regular page (that will eventually be returned to the normal
 512  * page-allocator via put_page).
 513  */
 514 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 515 {
 516         dma_addr_t dma;
 517         int count;
 518
 519         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 520                 /* Always account for inflight pages, even if we didn't
 521                  * map them
 522                  */
 523                 goto skip_dma_unmap;
 524
 525         dma = page_pool_get_dma_addr(page);
 526
 527         /* When page is unmapped, it cannot be returned to our pool */
 528         dma_unmap_page_attrs(pool->p.dev, dma,
 529                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 530                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 531         page_pool_set_dma_addr(page, 0);
 532 skip_dma_unmap:
 533         page_pool_clear_pp_info(page);
 534
 535         /* This may be the last page returned, releasing the pool, so
 536          * it is not safe to reference pool afterwards.
 537          */
 538         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 539         trace_page_pool_state_release(pool, page, count);
 540
 541         put_page(page);
 542         /* An optimization would be to call __free_pages(page, pool->p.order)
 543          * knowing page is not part of page-cache (thus avoiding a
 544          * __page_cache_release() call).
 545          */
 546 }
 547
 548 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 549 {
 550         int ret;
 551         /* BH protection not needed if current is softirq */
 552         if (in_softirq())
 553                 ret = ptr_ring_produce(&pool->ring, page);
 554         else
 555                 ret = ptr_ring_produce_bh(&pool->ring, page);
 556
 557         if (!ret) {
 558                 recycle_stat_inc(pool, ring);
 559                 return true;
 560         }
 561
 562         return false;
 563 }
 564
 565 /* Only allow direct recycling in special circumstances, into the
 566  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 567  *
 568  * Caller must provide appropriate safe context.
 569  */
 570 static bool page_pool_recycle_in_cache(struct page *page,
 571                                        struct page_pool *pool)
 572 {
 573         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
 574                 recycle_stat_inc(pool, cache_full);
 575                 return false;
 576         }
 577
 578         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 579         pool->alloc.cache[pool->alloc.count++] = page;
 580         recycle_stat_inc(pool, cached);
 581         return true;
 582 }
 583
 584 /* If the page refcnt == 1, this will try to recycle the page.
 585  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 586  * the configured size min(dma_sync_size, pool->max_len).
 587  * If the page refcnt != 1, then the page will be returned to memory
 588  * subsystem.
 589  */
 590 static __always_inline struct page *
 591 __page_pool_put_page(struct page_pool *pool, struct page *page,
 592                      unsigned int dma_sync_size, bool allow_direct)
 593 {
 594         lockdep_assert_no_hardirq();
 595
 596         /* This allocator is optimized for the XDP mode that uses
 597          * one-frame-per-page, but have fallbacks that act like the
 598          * regular page allocator APIs.
 599          *
 600          * refcnt == 1 means page_pool owns page, and can recycle it.
 601          *
 602          * page is NOT reusable when allocated when system is under
 603          * some pressure. (page_is_pfmemalloc)
 604          */
 605         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 606                 /* Read barrier done in page_ref_count / READ_ONCE */
 607
 608                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 609                         page_pool_dma_sync_for_device(pool, page,
 610                                                       dma_sync_size);
 611
 612                 if (allow_direct && in_softirq() &&
 613                     page_pool_recycle_in_cache(page, pool))
 614                         return NULL;
 615
 616                 /* Page found as candidate for recycling */
 617                 return page;
 618         }
 619         /* Fallback/non-XDP mode: API user have elevated refcnt.
 620          *
 621          * Many drivers split up the page into fragments, and some
 622          * want to keep doing this to save memory and do refcnt based
 623          * recycling. Support this use case too, to ease drivers
 624          * switching between XDP/non-XDP.
 625          *
 626          * In-case page_pool maintains the DMA mapping, API user must
 627          * call page_pool_put_page once.  In this elevated refcnt
 628          * case, the DMA is unmapped/released, as driver is likely
 629          * doing refcnt based recycle tricks, meaning another process
 630          * will be invoking put_page.
 631          */
 632         recycle_stat_inc(pool, released_refcnt);
 633         page_pool_return_page(pool, page);
 634
 635         return NULL;
 636 }
 637
 638 void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
 639                                   unsigned int dma_sync_size, bool allow_direct)
 640 {
 641         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 642         if (page && !page_pool_recycle_in_ring(pool, page)) {
 643                 /* Cache full, fallback to free pages */
 644                 recycle_stat_inc(pool, ring_full);
 645                 page_pool_return_page(pool, page);
 646         }
 647 }
 648 EXPORT_SYMBOL(page_pool_put_defragged_page);
 649
 650 /**
 651  * page_pool_put_page_bulk() - release references on multiple pages
 652  * @pool:       pool from which pages were allocated
 653  * @data:       array holding page pointers
 654  * @count:      number of pages in @data
 655  *
 656  * Tries to refill a number of pages into the ptr_ring cache holding ptr_ring
 657  * producer lock. If the ptr_ring is full, page_pool_put_page_bulk()
 658  * will release leftover pages to the page allocator.
 659  * page_pool_put_page_bulk() is suitable to be run inside the driver NAPI tx
 660  * completion loop for the XDP_REDIRECT use case.
 661  *
 662  * Please note the caller must not use data area after running
 663  * page_pool_put_page_bulk(), as this function overwrites it.
 664  */
 665 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 666                              int count)
 667 {
 668         int i, bulk_len = 0;
 669         bool in_softirq;
 670
 671         for (i = 0; i < count; i++) {
 672                 struct page *page = virt_to_head_page(data[i]);
 673
 674                 /* It is not the last user for the page frag case */
 675                 if (!page_pool_is_last_frag(pool, page))
 676                         continue;
 677
 678                 page = __page_pool_put_page(pool, page, -1, false);
 679                 /* Approved for bulk recycling in ptr_ring cache */
 680                 if (page)
 681                         data[bulk_len++] = page;
 682         }
 683
 684         if (unlikely(!bulk_len))
 685                 return;
 686
 687         /* Bulk producer into ptr_ring page_pool cache */
 688         in_softirq = page_pool_producer_lock(pool);
 689         for (i = 0; i < bulk_len; i++) {
 690                 if (__ptr_ring_produce(&pool->ring, data[i])) {
 691                         /* ring full */
 692                         recycle_stat_inc(pool, ring_full);
 693                         break;
 694                 }
 695         }
 696         recycle_stat_add(pool, ring, i);
 697         page_pool_producer_unlock(pool, in_softirq);
 698
 699         /* Hopefully all pages was return into ptr_ring */
 700         if (likely(i == bulk_len))
 701                 return;
 702
 703         /* ptr_ring cache full, free remaining pages outside producer lock
 704          * since put_page() with refcnt == 1 can be an expensive operation
 705          */
 706         for (; i < bulk_len; i++)
 707                 page_pool_return_page(pool, data[i]);
 708 }
 709 EXPORT_SYMBOL(page_pool_put_page_bulk);
 710
 711 static struct page *page_pool_drain_frag(struct page_pool *pool,
 712                                          struct page *page)
 713 {
 714         long drain_count = BIAS_MAX - pool->frag_users;
 715
 716         /* Some user is still using the page frag */
 717         if (likely(page_pool_defrag_page(page, drain_count)))
 718                 return NULL;
 719
 720         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 721                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 722                         page_pool_dma_sync_for_device(pool, page, -1);
 723
 724                 return page;
 725         }
 726
 727         page_pool_return_page(pool, page);
 728         return NULL;
 729 }
 730
 731 static void page_pool_free_frag(struct page_pool *pool)
 732 {
 733         long drain_count = BIAS_MAX - pool->frag_users;
 734         struct page *page = pool->frag_page;
 735
 736         pool->frag_page = NULL;
 737
 738         if (!page || page_pool_defrag_page(page, drain_count))
 739                 return;
 740
 741         page_pool_return_page(pool, page);
 742 }
 743
 744 struct page *page_pool_alloc_frag(struct page_pool *pool,
 745                                   unsigned int *offset,
 746                                   unsigned int size, gfp_t gfp)
 747 {
 748         unsigned int max_size = PAGE_SIZE << pool->p.order;
 749         struct page *page = pool->frag_page;
 750
 751         if (WARN_ON(!(pool->p.flags & PP_FLAG_PAGE_FRAG) ||
 752                     size > max_size))
 753                 return NULL;
 754
 755         size = ALIGN(size, dma_get_cache_alignment());
 756         *offset = pool->frag_offset;
 757
 758         if (page && *offset + size > max_size) {
 759                 page = page_pool_drain_frag(pool, page);
 760                 if (page) {
 761                         alloc_stat_inc(pool, fast);
 762                         goto frag_reset;
 763                 }
 764         }
 765
 766         if (!page) {
 767                 page = page_pool_alloc_pages(pool, gfp);
 768                 if (unlikely(!page)) {
 769                         pool->frag_page = NULL;
 770                         return NULL;
 771                 }
 772
 773                 pool->frag_page = page;
 774
 775 frag_reset:
 776                 pool->frag_users = 1;
 777                 *offset = 0;
 778                 pool->frag_offset = size;
 779                 page_pool_fragment_page(page, BIAS_MAX);
 780                 return page;
 781         }
 782
 783         pool->frag_users++;
 784         pool->frag_offset = *offset + size;
 785         alloc_stat_inc(pool, fast);
 786         return page;
 787 }
 788 EXPORT_SYMBOL(page_pool_alloc_frag);
 789
 790 static void page_pool_empty_ring(struct page_pool *pool)
 791 {
 792         struct page *page;
 793
 794         /* Empty recycle ring */
 795         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 796                 /* Verify the refcnt invariant of cached pages */
 797                 if (!(page_ref_count(page) == 1))
 798                         pr_crit("%s() page_pool refcnt %d violation\n",
 799                                 __func__, page_ref_count(page));
 800
 801                 page_pool_return_page(pool, page);
 802         }
 803 }
 804
 805 static void page_pool_free(struct page_pool *pool)
 806 {
 807         if (pool->disconnect)
 808                 pool->disconnect(pool);
 809
 810         ptr_ring_cleanup(&pool->ring, NULL);
 811
 812         if (pool->p.flags & PP_FLAG_DMA_MAP)
 813                 put_device(pool->p.dev);
 814
 815 #ifdef CONFIG_PAGE_POOL_STATS
 816         free_percpu(pool->recycle_stats);
 817 #endif
 818         kfree(pool);
 819 }
 820
 821 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 822 {
 823         struct page *page;
 824
 825         if (pool->destroy_cnt)
 826                 return;
 827
 828         /* Empty alloc cache, assume caller made sure this is
 829          * no-longer in use, and page_pool_alloc_pages() cannot be
 830          * call concurrently.
 831          */
 832         while (pool->alloc.count) {
 833                 page = pool->alloc.cache[--pool->alloc.count];
 834                 page_pool_return_page(pool, page);
 835         }
 836 }
 837
 838 static void page_pool_scrub(struct page_pool *pool)
 839 {
 840         page_pool_empty_alloc_cache_once(pool);
 841         pool->destroy_cnt++;
 842
 843         /* No more consumers should exist, but producers could still
 844          * be in-flight.
 845          */
 846         page_pool_empty_ring(pool);
 847 }
 848
 849 static int page_pool_release(struct page_pool *pool)
 850 {
 851         int inflight;
 852
 853         page_pool_scrub(pool);
 854         inflight = page_pool_inflight(pool);
 855         if (!inflight)
 856                 page_pool_free(pool);
 857
 858         return inflight;
 859 }
 860
 861 static void page_pool_release_retry(struct work_struct *wq)
 862 {
 863         struct delayed_work *dwq = to_delayed_work(wq);
 864         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 865         int inflight;
 866
 867         inflight = page_pool_release(pool);
 868         if (!inflight)
 869                 return;
 870
 871         /* Periodic warning */
 872         if (time_after_eq(jiffies, pool->defer_warn)) {
 873                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 874
 875                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 876                         __func__, inflight, sec);
 877                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 878         }
 879
 880         /* Still not ready to be disconnected, retry later */
 881         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 882 }
 883
 884 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
 885                            struct xdp_mem_info *mem)
 886 {
 887         refcount_inc(&pool->user_cnt);
 888         pool->disconnect = disconnect;
 889         pool->xdp_mem_id = mem->id;
 890 }
 891
 892 void page_pool_unlink_napi(struct page_pool *pool)
 893 {
 894         if (!pool->p.napi)
 895                 return;
 896
 897         /* To avoid races with recycling and additional barriers make sure
 898          * pool and NAPI are unlinked when NAPI is disabled.
 899          */
 900         WARN_ON(!test_bit(NAPI_STATE_SCHED, &pool->p.napi->state) ||
 901                 READ_ONCE(pool->p.napi->list_owner) != -1);
 902
 903         WRITE_ONCE(pool->p.napi, NULL);
 904 }
 905 EXPORT_SYMBOL(page_pool_unlink_napi);
 906
 907 void page_pool_destroy(struct page_pool *pool)
 908 {
 909         if (!pool)
 910                 return;
 911
 912         if (!page_pool_put(pool))
 913                 return;
 914
 915         page_pool_unlink_napi(pool);
 916         page_pool_free_frag(pool);
 917
 918         if (!page_pool_release(pool))
 919                 return;
 920
 921         pool->defer_start = jiffies;
 922         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 923
 924         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 925         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 926 }
 927 EXPORT_SYMBOL(page_pool_destroy);
 928
 929 /* Caller must provide appropriate safe context, e.g. NAPI. */
 930 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 931 {
 932         struct page *page;
 933
 934         trace_page_pool_update_nid(pool, new_nid);
 935         pool->p.nid = new_nid;
 936
 937         /* Flush pool alloc cache, as refill will check NUMA node */
 938         while (pool->alloc.count) {
 939                 page = pool->alloc.cache[--pool->alloc.count];
 940                 page_pool_return_page(pool, page);
 941         }
 942 }
 943 EXPORT_SYMBOL(page_pool_update_nid);