net/core/page_pool.c

   1 /* SPDX-License-Identifier: GPL-2.0
   2  *
   3  * page_pool.c
   4  *      Author: Jesper Dangaard Brouer <netoptimizer@brouer.com>
   5  *      Copyright (C) 2016 Red Hat, Inc.
   6  */
   7
   8 #include <linux/types.h>
   9 #include <linux/kernel.h>
  10 #include <linux/slab.h>
  11 #include <linux/device.h>
  12
  13 #include <net/page_pool.h>
  14 #include <net/xdp.h>
  15
  16 #include <linux/dma-direction.h>
  17 #include <linux/dma-mapping.h>
  18 #include <linux/page-flags.h>
  19 #include <linux/mm.h> /* for __put_page() */
  20
  21 #include <trace/events/page_pool.h>
  22
  23 #define DEFER_TIME (msecs_to_jiffies(1000))
  24 #define DEFER_WARN_INTERVAL (60 * HZ)
  25
  26 static int page_pool_init(struct page_pool *pool,
  27                           const struct page_pool_params *params)
  28 {
  29         unsigned int ring_qsize = 1024; /* Default */
  30
  31         memcpy(&pool->p, params, sizeof(pool->p));
  32
  33         /* Validate only known flags were used */
  34         if (pool->p.flags & ~(PP_FLAG_ALL))
  35                 return -EINVAL;
  36
  37         if (pool->p.pool_size)
  38                 ring_qsize = pool->p.pool_size;
  39
  40         /* Sanity limit mem that can be pinned down */
  41         if (ring_qsize > 32768)
  42                 return -E2BIG;
  43
  44         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
  45          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
  46          * which is the XDP_TX use-case.
  47          */
  48         if (pool->p.flags & PP_FLAG_DMA_MAP) {
  49                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
  50                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
  51                         return -EINVAL;
  52         }
  53
  54         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
  55                 /* In order to request DMA-sync-for-device the page
  56                  * needs to be mapped
  57                  */
  58                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
  59                         return -EINVAL;
  60
  61                 if (!pool->p.max_len)
  62                         return -EINVAL;
  63
  64                 /* pool->p.offset has to be set according to the address
  65                  * offset used by the DMA engine to start copying rx data
  66                  */
  67         }
  68
  69         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
  70                 return -ENOMEM;
  71
  72         atomic_set(&pool->pages_state_release_cnt, 0);
  73
  74         /* Driver calling page_pool_create() also call page_pool_destroy() */
  75         refcount_set(&pool->user_cnt, 1);
  76
  77         if (pool->p.flags & PP_FLAG_DMA_MAP)
  78                 get_device(pool->p.dev);
  79
  80         return 0;
  81 }
  82
  83 struct page_pool *page_pool_create(const struct page_pool_params *params)
  84 {
  85         struct page_pool *pool;
  86         int err;
  87
  88         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
  89         if (!pool)
  90                 return ERR_PTR(-ENOMEM);
  91
  92         err = page_pool_init(pool, params);
  93         if (err < 0) {
  94                 pr_warn("%s() gave up with errno %d\n", __func__, err);
  95                 kfree(pool);
  96                 return ERR_PTR(err);
  97         }
  98
  99         return pool;
 100 }
 101 EXPORT_SYMBOL(page_pool_create);
 102
 103 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 104
 105 noinline
 106 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 107 {
 108         struct ptr_ring *r = &pool->ring;
 109         struct page *page;
 110         int pref_nid; /* preferred NUMA node */
 111
 112         /* Quicker fallback, avoid locks when ring is empty */
 113         if (__ptr_ring_empty(r))
 114                 return NULL;
 115
 116         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 117          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 118          */
 119 #ifdef CONFIG_NUMA
 120         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 121 #else
 122         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 123         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 124 #endif
 125
 126         /* Slower-path: Get pages from locked ring queue */
 127         spin_lock(&r->consumer_lock);
 128
 129         /* Refill alloc array, but only if NUMA match */
 130         do {
 131                 page = __ptr_ring_consume(r);
 132                 if (unlikely(!page))
 133                         break;
 134
 135                 if (likely(page_to_nid(page) == pref_nid)) {
 136                         pool->alloc.cache[pool->alloc.count++] = page;
 137                 } else {
 138                         /* NUMA mismatch;
 139                          * (1) release 1 page to page-allocator and
 140                          * (2) break out to fallthrough to alloc_pages_node.
 141                          * This limit stress on page buddy alloactor.
 142                          */
 143                         page_pool_return_page(pool, page);
 144                         page = NULL;
 145                         break;
 146                 }
 147         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 148
 149         /* Return last page */
 150         if (likely(pool->alloc.count > 0))
 151                 page = pool->alloc.cache[--pool->alloc.count];
 152
 153         spin_unlock(&r->consumer_lock);
 154         return page;
 155 }
 156
 157 /* fast path */
 158 static struct page *__page_pool_get_cached(struct page_pool *pool)
 159 {
 160         struct page *page;
 161
 162         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 163         if (likely(pool->alloc.count)) {
 164                 /* Fast-path */
 165                 page = pool->alloc.cache[--pool->alloc.count];
 166         } else {
 167                 page = page_pool_refill_alloc_cache(pool);
 168         }
 169
 170         return page;
 171 }
 172
 173 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 174                                           struct page *page,
 175                                           unsigned int dma_sync_size)
 176 {
 177         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 178         dma_sync_single_range_for_device(pool->p.dev, page->dma_addr,
 179                                          pool->p.offset, dma_sync_size,
 180                                          pool->p.dma_dir);
 181 }
 182
 183 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
 184 {
 185         dma_addr_t dma;
 186
 187         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 188          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 189          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 190          * This mapping is kept for lifetime of page, until leaving pool.
 191          */
 192         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 193                                  (PAGE_SIZE << pool->p.order),
 194                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 195         if (dma_mapping_error(pool->p.dev, dma))
 196                 return false;
 197
 198         page->dma_addr = dma;
 199
 200         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 201                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 202
 203         return true;
 204 }
 205
 206 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
 207                                                  gfp_t gfp)
 208 {
 209         struct page *page;
 210
 211         gfp |= __GFP_COMP;
 212         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 213         if (unlikely(!page))
 214                 return NULL;
 215
 216         if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
 217             unlikely(!page_pool_dma_map(pool, page))) {
 218                 put_page(page);
 219                 return NULL;
 220         }
 221
 222         /* Track how many pages are held 'in-flight' */
 223         pool->pages_state_hold_cnt++;
 224         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 225         return page;
 226 }
 227
 228 /* slow path */
 229 noinline
 230 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 231                                                  gfp_t gfp)
 232 {
 233         const int bulk = PP_ALLOC_CACHE_REFILL;
 234         unsigned int pp_flags = pool->p.flags;
 235         unsigned int pp_order = pool->p.order;
 236         struct page *page;
 237         int i, nr_pages;
 238
 239         /* Don't support bulk alloc for high-order pages */
 240         if (unlikely(pp_order))
 241                 return __page_pool_alloc_page_order(pool, gfp);
 242
 243         /* Unnecessary as alloc cache is empty, but guarantees zero count */
 244         if (unlikely(pool->alloc.count > 0))
 245                 return pool->alloc.cache[--pool->alloc.count];
 246
 247         /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
 248         memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
 249
 250         nr_pages = alloc_pages_bulk_array(gfp, bulk, pool->alloc.cache);
 251         if (unlikely(!nr_pages))
 252                 return NULL;
 253
 254         /* Pages have been filled into alloc.cache array, but count is zero and
 255          * page element have not been (possibly) DMA mapped.
 256          */
 257         for (i = 0; i < nr_pages; i++) {
 258                 page = pool->alloc.cache[i];
 259                 if ((pp_flags & PP_FLAG_DMA_MAP) &&
 260                     unlikely(!page_pool_dma_map(pool, page))) {
 261                         put_page(page);
 262                         continue;
 263                 }
 264                 pool->alloc.cache[pool->alloc.count++] = page;
 265                 /* Track how many pages are held 'in-flight' */
 266                 pool->pages_state_hold_cnt++;
 267                 trace_page_pool_state_hold(pool, page,
 268                                            pool->pages_state_hold_cnt);
 269         }
 270
 271         /* Return last page */
 272         if (likely(pool->alloc.count > 0))
 273                 page = pool->alloc.cache[--pool->alloc.count];
 274         else
 275                 page = NULL;
 276
 277         /* When page just alloc'ed is should/must have refcnt 1. */
 278         return page;
 279 }
 280
 281 /* For using page_pool replace: alloc_pages() API calls, but provide
 282  * synchronization guarantee for allocation side.
 283  */
 284 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 285 {
 286         struct page *page;
 287
 288         /* Fast-path: Get a page from cache */
 289         page = __page_pool_get_cached(pool);
 290         if (page)
 291                 return page;
 292
 293         /* Slow-path: cache empty, do real allocation */
 294         page = __page_pool_alloc_pages_slow(pool, gfp);
 295         return page;
 296 }
 297 EXPORT_SYMBOL(page_pool_alloc_pages);
 298
 299 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 300  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 301  */
 302 #define _distance(a, b) (s32)((a) - (b))
 303
 304 static s32 page_pool_inflight(struct page_pool *pool)
 305 {
 306         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 307         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 308         s32 inflight;
 309
 310         inflight = _distance(hold_cnt, release_cnt);
 311
 312         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 313         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 314
 315         return inflight;
 316 }
 317
 318 /* Disconnects a page (from a page_pool).  API users can have a need
 319  * to disconnect a page (from a page_pool), to allow it to be used as
 320  * a regular page (that will eventually be returned to the normal
 321  * page-allocator via put_page).
 322  */
 323 void page_pool_release_page(struct page_pool *pool, struct page *page)
 324 {
 325         dma_addr_t dma;
 326         int count;
 327
 328         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 329                 /* Always account for inflight pages, even if we didn't
 330                  * map them
 331                  */
 332                 goto skip_dma_unmap;
 333
 334         dma = page->dma_addr;
 335
 336         /* When page is unmapped, it cannot be returned our pool */
 337         dma_unmap_page_attrs(pool->p.dev, dma,
 338                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 339                              DMA_ATTR_SKIP_CPU_SYNC);
 340         page->dma_addr = 0;
 341 skip_dma_unmap:
 342         /* This may be the last page returned, releasing the pool, so
 343          * it is not safe to reference pool afterwards.
 344          */
 345         count = atomic_inc_return(&pool->pages_state_release_cnt);
 346         trace_page_pool_state_release(pool, page, count);
 347 }
 348 EXPORT_SYMBOL(page_pool_release_page);
 349
 350 /* Return a page to the page allocator, cleaning up our state */
 351 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 352 {
 353         page_pool_release_page(pool, page);
 354
 355         put_page(page);
 356         /* An optimization would be to call __free_pages(page, pool->p.order)
 357          * knowing page is not part of page-cache (thus avoiding a
 358          * __page_cache_release() call).
 359          */
 360 }
 361
 362 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 363 {
 364         int ret;
 365         /* BH protection not needed if current is serving softirq */
 366         if (in_serving_softirq())
 367                 ret = ptr_ring_produce(&pool->ring, page);
 368         else
 369                 ret = ptr_ring_produce_bh(&pool->ring, page);
 370
 371         return (ret == 0) ? true : false;
 372 }
 373
 374 /* Only allow direct recycling in special circumstances, into the
 375  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 376  *
 377  * Caller must provide appropriate safe context.
 378  */
 379 static bool page_pool_recycle_in_cache(struct page *page,
 380                                        struct page_pool *pool)
 381 {
 382         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
 383                 return false;
 384
 385         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 386         pool->alloc.cache[pool->alloc.count++] = page;
 387         return true;
 388 }
 389
 390 /* If the page refcnt == 1, this will try to recycle the page.
 391  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 392  * the configured size min(dma_sync_size, pool->max_len).
 393  * If the page refcnt != 1, then the page will be returned to memory
 394  * subsystem.
 395  */
 396 static __always_inline struct page *
 397 __page_pool_put_page(struct page_pool *pool, struct page *page,
 398                      unsigned int dma_sync_size, bool allow_direct)
 399 {
 400         /* This allocator is optimized for the XDP mode that uses
 401          * one-frame-per-page, but have fallbacks that act like the
 402          * regular page allocator APIs.
 403          *
 404          * refcnt == 1 means page_pool owns page, and can recycle it.
 405          *
 406          * page is NOT reusable when allocated when system is under
 407          * some pressure. (page_is_pfmemalloc)
 408          */
 409         if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
 410                 /* Read barrier done in page_ref_count / READ_ONCE */
 411
 412                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 413                         page_pool_dma_sync_for_device(pool, page,
 414                                                       dma_sync_size);
 415
 416                 if (allow_direct && in_serving_softirq() &&
 417                     page_pool_recycle_in_cache(page, pool))
 418                         return NULL;
 419
 420                 /* Page found as candidate for recycling */
 421                 return page;
 422         }
 423         /* Fallback/non-XDP mode: API user have elevated refcnt.
 424          *
 425          * Many drivers split up the page into fragments, and some
 426          * want to keep doing this to save memory and do refcnt based
 427          * recycling. Support this use case too, to ease drivers
 428          * switching between XDP/non-XDP.
 429          *
 430          * In-case page_pool maintains the DMA mapping, API user must
 431          * call page_pool_put_page once.  In this elevated refcnt
 432          * case, the DMA is unmapped/released, as driver is likely
 433          * doing refcnt based recycle tricks, meaning another process
 434          * will be invoking put_page.
 435          */
 436         /* Do not replace this with page_pool_return_page() */
 437         page_pool_release_page(pool, page);
 438         put_page(page);
 439
 440         return NULL;
 441 }
 442
 443 void page_pool_put_page(struct page_pool *pool, struct page *page,
 444                         unsigned int dma_sync_size, bool allow_direct)
 445 {
 446         page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
 447         if (page && !page_pool_recycle_in_ring(pool, page)) {
 448                 /* Cache full, fallback to free pages */
 449                 page_pool_return_page(pool, page);
 450         }
 451 }
 452 EXPORT_SYMBOL(page_pool_put_page);
 453
 454 /* Caller must not use data area after call, as this function overwrites it */
 455 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
 456                              int count)
 457 {
 458         int i, bulk_len = 0;
 459
 460         for (i = 0; i < count; i++) {
 461                 struct page *page = virt_to_head_page(data[i]);
 462
 463                 page = __page_pool_put_page(pool, page, -1, false);
 464                 /* Approved for bulk recycling in ptr_ring cache */
 465                 if (page)
 466                         data[bulk_len++] = page;
 467         }
 468
 469         if (unlikely(!bulk_len))
 470                 return;
 471
 472         /* Bulk producer into ptr_ring page_pool cache */
 473         page_pool_ring_lock(pool);
 474         for (i = 0; i < bulk_len; i++) {
 475                 if (__ptr_ring_produce(&pool->ring, data[i]))
 476                         break; /* ring full */
 477         }
 478         page_pool_ring_unlock(pool);
 479
 480         /* Hopefully all pages was return into ptr_ring */
 481         if (likely(i == bulk_len))
 482                 return;
 483
 484         /* ptr_ring cache full, free remaining pages outside producer lock
 485          * since put_page() with refcnt == 1 can be an expensive operation
 486          */
 487         for (; i < bulk_len; i++)
 488                 page_pool_return_page(pool, data[i]);
 489 }
 490 EXPORT_SYMBOL(page_pool_put_page_bulk);
 491
 492 static void page_pool_empty_ring(struct page_pool *pool)
 493 {
 494         struct page *page;
 495
 496         /* Empty recycle ring */
 497         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 498                 /* Verify the refcnt invariant of cached pages */
 499                 if (!(page_ref_count(page) == 1))
 500                         pr_crit("%s() page_pool refcnt %d violation\n",
 501                                 __func__, page_ref_count(page));
 502
 503                 page_pool_return_page(pool, page);
 504         }
 505 }
 506
 507 static void page_pool_free(struct page_pool *pool)
 508 {
 509         if (pool->disconnect)
 510                 pool->disconnect(pool);
 511
 512         ptr_ring_cleanup(&pool->ring, NULL);
 513
 514         if (pool->p.flags & PP_FLAG_DMA_MAP)
 515                 put_device(pool->p.dev);
 516
 517         kfree(pool);
 518 }
 519
 520 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 521 {
 522         struct page *page;
 523
 524         if (pool->destroy_cnt)
 525                 return;
 526
 527         /* Empty alloc cache, assume caller made sure this is
 528          * no-longer in use, and page_pool_alloc_pages() cannot be
 529          * call concurrently.
 530          */
 531         while (pool->alloc.count) {
 532                 page = pool->alloc.cache[--pool->alloc.count];
 533                 page_pool_return_page(pool, page);
 534         }
 535 }
 536
 537 static void page_pool_scrub(struct page_pool *pool)
 538 {
 539         page_pool_empty_alloc_cache_once(pool);
 540         pool->destroy_cnt++;
 541
 542         /* No more consumers should exist, but producers could still
 543          * be in-flight.
 544          */
 545         page_pool_empty_ring(pool);
 546 }
 547
 548 static int page_pool_release(struct page_pool *pool)
 549 {
 550         int inflight;
 551
 552         page_pool_scrub(pool);
 553         inflight = page_pool_inflight(pool);
 554         if (!inflight)
 555                 page_pool_free(pool);
 556
 557         return inflight;
 558 }
 559
 560 static void page_pool_release_retry(struct work_struct *wq)
 561 {
 562         struct delayed_work *dwq = to_delayed_work(wq);
 563         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 564         int inflight;
 565
 566         inflight = page_pool_release(pool);
 567         if (!inflight)
 568                 return;
 569
 570         /* Periodic warning */
 571         if (time_after_eq(jiffies, pool->defer_warn)) {
 572                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 573
 574                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 575                         __func__, inflight, sec);
 576                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 577         }
 578
 579         /* Still not ready to be disconnected, retry later */
 580         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 581 }
 582
 583 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *))
 584 {
 585         refcount_inc(&pool->user_cnt);
 586         pool->disconnect = disconnect;
 587 }
 588
 589 void page_pool_destroy(struct page_pool *pool)
 590 {
 591         if (!pool)
 592                 return;
 593
 594         if (!page_pool_put(pool))
 595                 return;
 596
 597         if (!page_pool_release(pool))
 598                 return;
 599
 600         pool->defer_start = jiffies;
 601         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 602
 603         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 604         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 605 }
 606 EXPORT_SYMBOL(page_pool_destroy);
 607
 608 /* Caller must provide appropriate safe context, e.g. NAPI. */
 609 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 610 {
 611         struct page *page;
 612
 613         trace_page_pool_update_nid(pool, new_nid);
 614         pool->p.nid = new_nid;
 615
 616         /* Flush pool alloc cache, as refill will check NUMA node */
 617         while (pool->alloc.count) {
 618                 page = pool->alloc.cache[--pool->alloc.count];
 619                 page_pool_return_page(pool, page);
 620         }
 621 }
 622 EXPORT_SYMBOL(page_pool_update_nid);