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