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 #include "page_pool_priv.h"
  27
  28 #define DEFER_TIME (msecs_to_jiffies(1000))
  29 #define DEFER_WARN_INTERVAL (60 * HZ)
  30
  31 #define BIAS_MAX        LONG_MAX
  32
  33 #ifdef CONFIG_PAGE_POOL_STATS
  34 /* alloc_stat_inc is intended to be used in softirq context */
  35 #define alloc_stat_inc(pool, __stat)    (pool->alloc_stats.__stat++)
  36 /* recycle_stat_inc is safe to use when preemption is possible. */
  37 #define recycle_stat_inc(pool, __stat)                                                  \
  38         do {                                                                            \
  39                 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats;       \
  40                 this_cpu_inc(s->__stat);                                                \
  41         } while (0)
  42
  43 #define recycle_stat_add(pool, __stat, val)                                             \
  44         do {                                                                            \
  45                 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats;       \
  46                 this_cpu_add(s->__stat, val);                                           \
  47         } while (0)
  48
  49 static const char pp_stats[][ETH_GSTRING_LEN] = {
  50         "rx_pp_alloc_fast",
  51         "rx_pp_alloc_slow",
  52         "rx_pp_alloc_slow_ho",
  53         "rx_pp_alloc_empty",
  54         "rx_pp_alloc_refill",
  55         "rx_pp_alloc_waive",
  56         "rx_pp_recycle_cached",
  57         "rx_pp_recycle_cache_full",
  58         "rx_pp_recycle_ring",
  59         "rx_pp_recycle_ring_full",
  60         "rx_pp_recycle_released_ref",
  61 };
  62
  63 /**
  64  * page_pool_get_stats() - fetch page pool stats
  65  * @pool:       pool from which page was allocated
  66  * @stats:      struct page_pool_stats to fill in
  67  *
  68  * Retrieve statistics about the page_pool. This API is only available
  69  * if the kernel has been configured with ``CONFIG_PAGE_POOL_STATS=y``.
  70  * A pointer to a caller allocated struct page_pool_stats structure
  71  * is passed to this API which is filled in. The caller can then report
  72  * those stats to the user (perhaps via ethtool, debugfs, etc.).
  73  */
  74 bool page_pool_get_stats(struct page_pool *pool,
  75                          struct page_pool_stats *stats)
  76 {
  77         int cpu = 0;
  78
  79         if (!stats)
  80                 return false;
  81
  82         /* The caller is responsible to initialize stats. */
  83         stats->alloc_stats.fast += pool->alloc_stats.fast;
  84         stats->alloc_stats.slow += pool->alloc_stats.slow;
  85         stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
  86         stats->alloc_stats.empty += pool->alloc_stats.empty;
  87         stats->alloc_stats.refill += pool->alloc_stats.refill;
  88         stats->alloc_stats.waive += pool->alloc_stats.waive;
  89
  90         for_each_possible_cpu(cpu) {
  91                 const struct page_pool_recycle_stats *pcpu =
  92                         per_cpu_ptr(pool->recycle_stats, cpu);
  93
  94                 stats->recycle_stats.cached += pcpu->cached;
  95                 stats->recycle_stats.cache_full += pcpu->cache_full;
  96                 stats->recycle_stats.ring += pcpu->ring;
  97                 stats->recycle_stats.ring_full += pcpu->ring_full;
  98                 stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
  99         }
 100
 101         return true;
 102 }
 103 EXPORT_SYMBOL(page_pool_get_stats);
 104
 105 u8 *page_pool_ethtool_stats_get_strings(u8 *data)
 106 {
 107         int i;
 108
 109         for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
 110                 memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
 111                 data += ETH_GSTRING_LEN;
 112         }
 113
 114         return data;
 115 }
 116 EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);
 117
 118 int page_pool_ethtool_stats_get_count(void)
 119 {
 120         return ARRAY_SIZE(pp_stats);
 121 }
 122 EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);
 123
 124 u64 *page_pool_ethtool_stats_get(u64 *data, void *stats)
 125 {
 126         struct page_pool_stats *pool_stats = stats;
 127
 128         *data++ = pool_stats->alloc_stats.fast;
 129         *data++ = pool_stats->alloc_stats.slow;
 130         *data++ = pool_stats->alloc_stats.slow_high_order;
 131         *data++ = pool_stats->alloc_stats.empty;
 132         *data++ = pool_stats->alloc_stats.refill;
 133         *data++ = pool_stats->alloc_stats.waive;
 134         *data++ = pool_stats->recycle_stats.cached;
 135         *data++ = pool_stats->recycle_stats.cache_full;
 136         *data++ = pool_stats->recycle_stats.ring;
 137         *data++ = pool_stats->recycle_stats.ring_full;
 138         *data++ = pool_stats->recycle_stats.released_refcnt;
 139
 140         return data;
 141 }
 142 EXPORT_SYMBOL(page_pool_ethtool_stats_get);
 143
 144 #else
 145 #define alloc_stat_inc(pool, __stat)
 146 #define recycle_stat_inc(pool, __stat)
 147 #define recycle_stat_add(pool, __stat, val)
 148 #endif
 149
 150 static bool page_pool_producer_lock(struct page_pool *pool)
 151         __acquires(&pool->ring.producer_lock)
 152 {
 153         bool in_softirq = in_softirq();
 154
 155         if (in_softirq)
 156                 spin_lock(&pool->ring.producer_lock);
 157         else
 158                 spin_lock_bh(&pool->ring.producer_lock);
 159
 160         return in_softirq;
 161 }
 162
 163 static void page_pool_producer_unlock(struct page_pool *pool,
 164                                       bool in_softirq)
 165         __releases(&pool->ring.producer_lock)
 166 {
 167         if (in_softirq)
 168                 spin_unlock(&pool->ring.producer_lock);
 169         else
 170                 spin_unlock_bh(&pool->ring.producer_lock);
 171 }
 172
 173 static int page_pool_init(struct page_pool *pool,
 174                           const struct page_pool_params *params)
 175 {
 176         unsigned int ring_qsize = 1024; /* Default */
 177
 178         memcpy(&pool->p, &params->fast, sizeof(pool->p));
 179         memcpy(&pool->slow, &params->slow, sizeof(pool->slow));
 180
 181         /* Validate only known flags were used */
 182         if (pool->p.flags & ~(PP_FLAG_ALL))
 183                 return -EINVAL;
 184
 185         if (pool->p.pool_size)
 186                 ring_qsize = pool->p.pool_size;
 187
 188         /* Sanity limit mem that can be pinned down */
 189         if (ring_qsize > 32768)
 190                 return -E2BIG;
 191
 192         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
 193          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
 194          * which is the XDP_TX use-case.
 195          */
 196         if (pool->p.flags & PP_FLAG_DMA_MAP) {
 197                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
 198                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
 199                         return -EINVAL;
 200         }
 201
 202         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
 203                 /* In order to request DMA-sync-for-device the page
 204                  * needs to be mapped
 205                  */
 206                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 207                         return -EINVAL;
 208
 209                 if (!pool->p.max_len)
 210                         return -EINVAL;
 211
 212                 /* pool->p.offset has to be set according to the address
 213                  * offset used by the DMA engine to start copying rx data
 214                  */
 215         }
 216
 217         pool->has_init_callback = !!pool->slow.init_callback;
 218
 219 #ifdef CONFIG_PAGE_POOL_STATS
 220         pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
 221         if (!pool->recycle_stats)
 222                 return -ENOMEM;
 223 #endif
 224
 225         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0) {
 226 #ifdef CONFIG_PAGE_POOL_STATS
 227                 free_percpu(pool->recycle_stats);
 228 #endif
 229                 return -ENOMEM;
 230         }
 231
 232         atomic_set(&pool->pages_state_release_cnt, 0);
 233
 234         /* Driver calling page_pool_create() also call page_pool_destroy() */
 235         refcount_set(&pool->user_cnt, 1);
 236
 237         if (pool->p.flags & PP_FLAG_DMA_MAP)
 238                 get_device(pool->p.dev);
 239
 240         return 0;
 241 }
 242
 243 static void page_pool_uninit(struct page_pool *pool)
 244 {
 245         ptr_ring_cleanup(&pool->ring, NULL);
 246
 247         if (pool->p.flags & PP_FLAG_DMA_MAP)
 248                 put_device(pool->p.dev);
 249
 250 #ifdef CONFIG_PAGE_POOL_STATS
 251         free_percpu(pool->recycle_stats);
 252 #endif
 253 }
 254
 255 /**
 256  * page_pool_create() - create a page pool.
 257  * @params: parameters, see struct page_pool_params
 258  */
 259 struct page_pool *page_pool_create(const struct page_pool_params *params)
 260 {
 261         struct page_pool *pool;
 262         int err;
 263
 264         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
 265         if (!pool)
 266                 return ERR_PTR(-ENOMEM);
 267
 268         err = page_pool_init(pool, params);
 269         if (err < 0)
 270                 goto err_free;
 271
 272         err = page_pool_list(pool);
 273         if (err)
 274                 goto err_uninit;
 275
 276         return pool;
 277
 278 err_uninit:
 279         page_pool_uninit(pool);
 280 err_free:
 281         pr_warn("%s() gave up with errno %d\n", __func__, err);
 282         kfree(pool);
 283         return ERR_PTR(err);
 284 }
 285 EXPORT_SYMBOL(page_pool_create);
 286
 287 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 288
 289 noinline
 290 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 291 {
 292         struct ptr_ring *r = &pool->ring;
 293         struct page *page;
 294         int pref_nid; /* preferred NUMA node */
 295
 296         /* Quicker fallback, avoid locks when ring is empty */
 297         if (__ptr_ring_empty(r)) {
 298                 alloc_stat_inc(pool, empty);
 299                 return NULL;
 300         }
 301
 302         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 303          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 304          */
 305 #ifdef CONFIG_NUMA
 306         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 307 #else
 308         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 309         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 310 #endif
 311
 312         /* Refill alloc array, but only if NUMA match */
 313         do {
 314                 page = __ptr_ring_consume(r);
 315                 if (unlikely(!page))
 316                         break;
 317
 318                 if (likely(page_to_nid(page) == pref_nid)) {
 319                         pool->alloc.cache[pool->alloc.count++] = page;
 320                 } else {
 321                         /* NUMA mismatch;
 322                          * (1) release 1 page to page-allocator and
 323                          * (2) break out to fallthrough to alloc_pages_node.
 324                          * This limit stress on page buddy alloactor.
 325                          */
 326                         page_pool_return_page(pool, page);
 327                         alloc_stat_inc(pool, waive);
 328                         page = NULL;
 329                         break;
 330                 }
 331         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 332
 333         /* Return last page */
 334         if (likely(pool->alloc.count > 0)) {
 335                 page = pool->alloc.cache[--pool->alloc.count];
 336                 alloc_stat_inc(pool, refill);
 337         }
 338
 339         return page;
 340 }
 341
 342 /* fast path */
 343 static struct page *__page_pool_get_cached(struct page_pool *pool)
 344 {
 345         struct page *page;
 346
 347         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 348         if (likely(pool->alloc.count)) {
 349                 /* Fast-path */
 350                 page = pool->alloc.cache[--pool->alloc.count];
 351                 alloc_stat_inc(pool, fast);
 352         } else {
 353                 page = page_pool_refill_alloc_cache(pool);
 354         }
 355
 356         return page;
 357 }
 358
 359 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 360                                           struct page *page,
 361                                           unsigned int dma_sync_size)
 362 {
 363         dma_addr_t dma_addr = page_pool_get_dma_addr(page);
 364
 365         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 366         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 367                                          pool->p.offset, dma_sync_size,
 368                                          pool->p.dma_dir);
 369 }
 370
 371 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 372 {
 373         dma_addr_t dma;
 374
 375         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 376          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 377          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 378          * This mapping is kept for lifetime of page, until leaving pool.
 379          */
 380         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 381                                  (PAGE_SIZE << pool->p.order),
 382                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC |
 383                                                   DMA_ATTR_WEAK_ORDERING);
 384         if (dma_mapping_error(pool->p.dev, dma))
 385                 return false;
 386
 387         if (page_pool_set_dma_addr(page, dma))
 388                 goto unmap_failed;
 389
 390         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 391                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 392
 393         return true;
 394
 395 unmap_failed:
 396         WARN_ON_ONCE("unexpected DMA address, please report to netdev@");
 397         dma_unmap_page_attrs(pool->p.dev, dma,
 398                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 399                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 400         return false;
 401 }
 402
 403 static void page_pool_set_pp_info(struct page_pool *pool,
 404                                   struct page *page)
 405 {
 406         page->pp = pool;
 407         page->pp_magic |= PP_SIGNATURE;
 408
 409         /* Ensuring all pages have been split into one fragment initially:
 410          * page_pool_set_pp_info() is only called once for every page when it
 411          * is allocated from the page allocator and page_pool_fragment_page()
 412          * is dirtying the same cache line as the page->pp_magic above, so
 413          * the overhead is negligible.
 414          */
 415         page_pool_fragment_page(page, 1);
 416         if (pool->has_init_callback)
 417                 pool->slow.init_callback(page, pool->slow.init_arg);
 418 }
 419
 420 static void page_pool_clear_pp_info(struct page *page)
 421 {
 422         page->pp_magic = 0;
 423         page->pp = NULL;
 424 }
 425
 426 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 427                                                  gfp_t gfp)
 428 {
 429         struct page *page;
 430
 431         gfp |= __GFP_COMP;
 432         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 433         if (unlikely(!page))
 434                 return NULL;
 435
 436         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 437             unlikely(!page_pool_dma_map(pool, page))) {
 438                 put_page(page);
 439                 return NULL;
 440         }
 441
 442         alloc_stat_inc(pool, slow_high_order);
 443         page_pool_set_pp_info(pool, page);
 444
 445         /* Track how many pages are held 'in-flight' */
 446         pool->pages_state_hold_cnt++;
 447         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 448         return page;
 449 }
 450
 451 /* slow path */
 452 noinline
 453 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 454                                                  gfp_t gfp)
 455 {
 456         const int bulk = PP_ALLOC_CACHE_REFILL;
 457         unsigned int pp_flags = pool->p.flags;
 458         unsigned int pp_order = pool->p.order;
 459         struct page *page;
 460         int i, nr_pages;
 461
 462         /* Don't support bulk alloc for high-order pages */
 463         if (unlikely(pp_order))
 464                 return __page_pool_alloc_page_order(pool, gfp);
 465
 466         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 467         if (unlikely(pool->alloc.count > 0))
 468                 return pool->alloc.cache[--pool->alloc.count];
 469
 470         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 471         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 472
 473         nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk,
 474                                                pool->alloc.cache);
 475         if (unlikely(!nr_pages))
 476                 return NULL;
 477
 478         /* Pages have been filled into alloc.cache array, but count is zero and
 479          * page element have not been (possibly) DMA mapped.
 480          */
 481         for (i = 0; i < nr_pages; i++) {
 482                 page = pool->alloc.cache[i];
 483                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 484                     unlikely(!page_pool_dma_map(pool, page))) {
 485                         put_page(page);
 486                         continue;
 487                 }
 488
 489                 page_pool_set_pp_info(pool, page);
 490                 pool->alloc.cache[pool->alloc.count++] = page;
 491                 /* Track how many pages are held 'in-flight' */
 492                 pool->pages_state_hold_cnt++;
 493                 trace_page_pool_state_hold(pool, page,
 494                                            pool->pages_state_hold_cnt);
 495         }
 496
 497         /* Return last page */
 498         if (likely(pool->alloc.count > 0)) {
 499                 page = pool->alloc.cache[--pool->alloc.count];
 500                 alloc_stat_inc(pool, slow);
 501         } else {
 502                 page = NULL;
 503         }
 504
 505         /* When page just alloc'ed is should/must have refcnt 1. */
 506         return page;
 507 }
 508
 509 /* For using page_pool replace: alloc_pages() API calls, but provide
 510  * synchronization guarantee for allocation side.
 511  */
 512 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 513 {
 514         struct page *page;
 515
 516         /* Fast-path: Get a page from cache */
 517         page = __page_pool_get_cached(pool);
 518         if (page)
 519                 return page;
 520
 521         /* Slow-path: cache empty, do real allocation */
 522         page = __page_pool_alloc_pages_slow(pool, gfp);
 523         return page;
 524 }
 525 EXPORT_SYMBOL(page_pool_alloc_pages);
 526
 527 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 528  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 529  */
 530 #define _distance(a, b) (s32)((a) - (b))
 531
 532 s32 page_pool_inflight(const struct page_pool *pool, bool strict)
 533 {
 534         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 535         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 536         s32 inflight;
 537
 538         inflight = _distance(hold_cnt, release_cnt);
 539
 540         if (strict) {
 541                 trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 542                 WARN(inflight < 0, "Negative(%d) inflight packet-pages",
 543                      inflight);
 544         } else {
 545                 inflight = max(0, inflight);
 546         }
 547
 548         return inflight;
 549 }
 550
 551 /* Disconnects a page (from a page_pool).  API users can have a need
 552  * to disconnect a page (from a page_pool), to allow it to be used as
 553  * a regular page (that will eventually be returned to the normal
 554  * page-allocator via put_page).
 555  */
 556 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 557 {
 558         dma_addr_t dma;
 559         int count;
 560
 561         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 562                 /* Always account for inflight pages, even if we didn't
 563                  * map them
 564                  */
 565                 goto skip_dma_unmap;
 566
 567         dma = page_pool_get_dma_addr(page);
 568
 569         /* When page is unmapped, it cannot be returned to our pool */
 570         dma_unmap_page_attrs(pool->p.dev, dma,
 571                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 572                              DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
 573         page_pool_set_dma_addr(page, 0);
 574 skip_dma_unmap:
 575         page_pool_clear_pp_info(page);
 576
 577         /* This may be the last page returned, releasing the pool, so
 578          * it is not safe to reference pool afterwards.
 579          */
 580         count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
 581         trace_page_pool_state_release(pool, page, count);
 582
 583         put_page(page);
 584         /* An optimization would be to call __free_pages(page, pool->p.order)
 585          * knowing page is not part of page-cache (thus avoiding a
 586          * __page_cache_release() call).
 587          */
 588 }
 589
 590 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 591 {
 592         int ret;
 593         /* BH protection not needed if current is softirq */
 594         if (in_softirq())
 595                 ret = ptr_ring_produce(&pool->ring, page);
 596         else
 597                 ret = ptr_ring_produce_bh(&pool->ring, page);
 598
 599         if (!ret) {
 600                 recycle_stat_inc(pool, ring);
 601                 return true;
 602         }
 603
 604         return false;
 605 }
 606
 607 /* Only allow direct recycling in special circumstances, into the
 608  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 609  *
 610  * Caller must provide appropriate safe context.
 611  */
 612 static bool page_pool_recycle_in_cache(struct page *page,
 613                                        struct page_pool *pool)
 614 {
 615         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
 616                 recycle_stat_inc(pool, cache_full);
 617                 return false;
 618         }
 619
 620         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 621         pool->alloc.cache[pool->alloc.count++] = page;
 622         recycle_stat_inc(pool, cached);
 623         return true;
 624 }
 625
 626 /* If the page refcnt == 1, this will try to recycle the page.
 627  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 628  * the configured size min(dma_sync_size, pool->max_len).
 629  * If the page refcnt != 1, then the page will be returned to memory
 630  * subsystem.
 631  */
 632 static __always_inline struct page *
 633 __page_pool_put_page(struct page_pool *pool, struct page *page,
 634                      unsigned int dma_sync_size, bool allow_direct)
 635 {
 636         lockdep_assert_no_hardirq();
 637
 638         /* This allocator is optimized for the XDP mode that uses
 639          * one-frame-per-page, but have fallbacks that act like the
 640          * regular page allocator APIs.
 641          *
 642          * refcnt == 1 means page_pool owns page, and can recycle it.
 643          *
 644          * page is NOT reusable when allocated when system is under
 645          * some pressure. (page_is_pfmemalloc)
 646          */
 647         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 648                 /* Read barrier done in page_ref_count / READ_ONCE */
 649
 650                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 651                         page_pool_dma_sync_for_device(pool, page,
 652                                                       dma_sync_size);
 653
 654                 if (allow_direct && in_softirq() &&
 655                     page_pool_recycle_in_cache(page, pool))
 656                         return NULL;
 657
 658                 /* Page found as candidate for recycling */
 659                 return page;
 660         }
 661         /* Fallback/non-XDP mode: API user have elevated refcnt.
 662          *
 663          * Many drivers split up the page into fragments, and some
 664          * want to keep doing this to save memory and do refcnt based
 665          * recycling. Support this use case too, to ease drivers
 666          * switching between XDP/non-XDP.
 667          *
 668          * In-case page_pool maintains the DMA mapping, API user must
 669          * call page_pool_put_page once.  In this elevated refcnt
 670          * case, the DMA is unmapped/released, as driver is likely
 671          * doing refcnt based recycle tricks, meaning another process
 672          * will be invoking put_page.
 673          */
 674         recycle_stat_inc(pool, released_refcnt);
 675         page_pool_return_page(pool, page);
 676
 677         return NULL;
 678 }
 679
 680 void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
 681                                   unsigned int dma_sync_size, bool allow_direct)
 682 {
 683         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 684         if (page && !page_pool_recycle_in_ring(pool, page)) {
 685                 /* Cache full, fallback to free pages */
 686                 recycle_stat_inc(pool, ring_full);
 687                 page_pool_return_page(pool, page);
 688         }
 689 }
 690 EXPORT_SYMBOL(page_pool_put_defragged_page);
 691
 692 /**
 693  * page_pool_put_page_bulk() - release references on multiple pages
 694  * @pool:       pool from which pages were allocated
 695  * @data:       array holding page pointers
 696  * @count:      number of pages in @data
 697  *
 698  * Tries to refill a number of pages into the ptr_ring cache holding ptr_ring
 699  * producer lock. If the ptr_ring is full, page_pool_put_page_bulk()
 700  * will release leftover pages to the page allocator.
 701  * page_pool_put_page_bulk() is suitable to be run inside the driver NAPI tx
 702  * completion loop for the XDP_REDIRECT use case.
 703  *
 704  * Please note the caller must not use data area after running
 705  * page_pool_put_page_bulk(), as this function overwrites it.
 706  */
 707 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 708                              int count)
 709 {
 710         int i, bulk_len = 0;
 711         bool in_softirq;
 712
 713         for (i = 0; i < count; i++) {
 714                 struct page *page = virt_to_head_page(data[i]);
 715
 716                 /* It is not the last user for the page frag case */
 717                 if (!page_pool_is_last_frag(page))
 718                         continue;
 719
 720                 page = __page_pool_put_page(pool, page, -1, false);
 721                 /* Approved for bulk recycling in ptr_ring cache */
 722                 if (page)
 723                         data[bulk_len++] = page;
 724         }
 725
 726         if (unlikely(!bulk_len))
 727                 return;
 728
 729         /* Bulk producer into ptr_ring page_pool cache */
 730         in_softirq = page_pool_producer_lock(pool);
 731         for (i = 0; i < bulk_len; i++) {
 732                 if (__ptr_ring_produce(&pool->ring, data[i])) {
 733                         /* ring full */
 734                         recycle_stat_inc(pool, ring_full);
 735                         break;
 736                 }
 737         }
 738         recycle_stat_add(pool, ring, i);
 739         page_pool_producer_unlock(pool, in_softirq);
 740
 741         /* Hopefully all pages was return into ptr_ring */
 742         if (likely(i == bulk_len))
 743                 return;
 744
 745         /* ptr_ring cache full, free remaining pages outside producer lock
 746          * since put_page() with refcnt == 1 can be an expensive operation
 747          */
 748         for (; i < bulk_len; i++)
 749                 page_pool_return_page(pool, data[i]);
 750 }
 751 EXPORT_SYMBOL(page_pool_put_page_bulk);
 752
 753 static struct page *page_pool_drain_frag(struct page_pool *pool,
 754                                          struct page *page)
 755 {
 756         long drain_count = BIAS_MAX - pool->frag_users;
 757
 758         /* Some user is still using the page frag */
 759         if (likely(page_pool_defrag_page(page, drain_count)))
 760                 return NULL;
 761
 762         if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
 763                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 764                         page_pool_dma_sync_for_device(pool, page, -1);
 765
 766                 return page;
 767         }
 768
 769         page_pool_return_page(pool, page);
 770         return NULL;
 771 }
 772
 773 static void page_pool_free_frag(struct page_pool *pool)
 774 {
 775         long drain_count = BIAS_MAX - pool->frag_users;
 776         struct page *page = pool->frag_page;
 777
 778         pool->frag_page = NULL;
 779
 780         if (!page || page_pool_defrag_page(page, drain_count))
 781                 return;
 782
 783         page_pool_return_page(pool, page);
 784 }
 785
 786 struct page *page_pool_alloc_frag(struct page_pool *pool,
 787                                   unsigned int *offset,
 788                                   unsigned int size, gfp_t gfp)
 789 {
 790         unsigned int max_size = PAGE_SIZE << pool->p.order;
 791         struct page *page = pool->frag_page;
 792
 793         if (WARN_ON(size > max_size))
 794                 return NULL;
 795
 796         size = ALIGN(size, dma_get_cache_alignment());
 797         *offset = pool->frag_offset;
 798
 799         if (page && *offset + size > max_size) {
 800                 page = page_pool_drain_frag(pool, page);
 801                 if (page) {
 802                         alloc_stat_inc(pool, fast);
 803                         goto frag_reset;
 804                 }
 805         }
 806
 807         if (!page) {
 808                 page = page_pool_alloc_pages(pool, gfp);
 809                 if (unlikely(!page)) {
 810                         pool->frag_page = NULL;
 811                         return NULL;
 812                 }
 813
 814                 pool->frag_page = page;
 815
 816 frag_reset:
 817                 pool->frag_users = 1;
 818                 *offset = 0;
 819                 pool->frag_offset = size;
 820                 page_pool_fragment_page(page, BIAS_MAX);
 821                 return page;
 822         }
 823
 824         pool->frag_users++;
 825         pool->frag_offset = *offset + size;
 826         alloc_stat_inc(pool, fast);
 827         return page;
 828 }
 829 EXPORT_SYMBOL(page_pool_alloc_frag);
 830
 831 static void page_pool_empty_ring(struct page_pool *pool)
 832 {
 833         struct page *page;
 834
 835         /* Empty recycle ring */
 836         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 837                 /* Verify the refcnt invariant of cached pages */
 838                 if (!(page_ref_count(page) == 1))
 839                         pr_crit("%s() page_pool refcnt %d violation\n",
 840                                 __func__, page_ref_count(page));
 841
 842                 page_pool_return_page(pool, page);
 843         }
 844 }
 845
 846 static void __page_pool_destroy(struct page_pool *pool)
 847 {
 848         if (pool->disconnect)
 849                 pool->disconnect(pool);
 850
 851         page_pool_unlist(pool);
 852         page_pool_uninit(pool);
 853         kfree(pool);
 854 }
 855
 856 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 857 {
 858         struct page *page;
 859
 860         if (pool->destroy_cnt)
 861                 return;
 862
 863         /* Empty alloc cache, assume caller made sure this is
 864          * no-longer in use, and page_pool_alloc_pages() cannot be
 865          * call concurrently.
 866          */
 867         while (pool->alloc.count) {
 868                 page = pool->alloc.cache[--pool->alloc.count];
 869                 page_pool_return_page(pool, page);
 870         }
 871 }
 872
 873 static void page_pool_scrub(struct page_pool *pool)
 874 {
 875         page_pool_empty_alloc_cache_once(pool);
 876         pool->destroy_cnt++;
 877
 878         /* No more consumers should exist, but producers could still
 879          * be in-flight.
 880          */
 881         page_pool_empty_ring(pool);
 882 }
 883
 884 static int page_pool_release(struct page_pool *pool)
 885 {
 886         int inflight;
 887
 888         page_pool_scrub(pool);
 889         inflight = page_pool_inflight(pool, true);
 890         if (!inflight)
 891                 __page_pool_destroy(pool);
 892
 893         return inflight;
 894 }
 895
 896 static void page_pool_release_retry(struct work_struct *wq)
 897 {
 898         struct delayed_work *dwq = to_delayed_work(wq);
 899         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 900         int inflight;
 901
 902         inflight = page_pool_release(pool);
 903         if (!inflight)
 904                 return;
 905
 906         /* Periodic warning */
 907         if (time_after_eq(jiffies, pool->defer_warn)) {
 908                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 909
 910                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 911                         __func__, inflight, sec);
 912                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 913         }
 914
 915         /* Still not ready to be disconnected, retry later */
 916         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 917 }
 918
 919 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
 920                            struct xdp_mem_info *mem)
 921 {
 922         refcount_inc(&pool->user_cnt);
 923         pool->disconnect = disconnect;
 924         pool->xdp_mem_id = mem->id;
 925 }
 926
 927 void page_pool_unlink_napi(struct page_pool *pool)
 928 {
 929         if (!pool->p.napi)
 930                 return;
 931
 932         /* To avoid races with recycling and additional barriers make sure
 933          * pool and NAPI are unlinked when NAPI is disabled.
 934          */
 935         WARN_ON(!test_bit(NAPI_STATE_SCHED, &pool->p.napi->state) ||
 936                 READ_ONCE(pool->p.napi->list_owner) != -1);
 937
 938         WRITE_ONCE(pool->p.napi, NULL);
 939 }
 940 EXPORT_SYMBOL(page_pool_unlink_napi);
 941
 942 void page_pool_destroy(struct page_pool *pool)
 943 {
 944         if (!pool)
 945                 return;
 946
 947         if (!page_pool_put(pool))
 948                 return;
 949
 950         page_pool_unlink_napi(pool);
 951         page_pool_free_frag(pool);
 952
 953         if (!page_pool_release(pool))
 954                 return;
 955
 956         pool->defer_start = jiffies;
 957         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 958
 959         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 960         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 961 }
 962 EXPORT_SYMBOL(page_pool_destroy);
 963
 964 /* Caller must provide appropriate safe context, e.g. NAPI. */
 965 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 966 {
 967         struct page *page;
 968
 969         trace_page_pool_update_nid(pool, new_nid);
 970         pool->p.nid = new_nid;
 971
 972         /* Flush pool alloc cache, as refill will check NUMA node */
 973         while (pool->alloc.count) {
 974                 page = pool->alloc.cache[--pool->alloc.count];
 975                 page_pool_return_page(pool, page);
 976         }
 977 }
 978 EXPORT_SYMBOL(page_pool_update_nid);