net/core/page_pool.c

   1 /* SPDX-License-Identifier: GPL-2.0
   2  *
   3  * page_pool.c
   4  *      Author: Jesper Dangaard Brouer <[email protected]>
   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 {
 104         struct ptr_ring *r = &pool->ring;
 105         struct page *page;
 106         int pref_nid; /* preferred NUMA node */
 107
 108         /* Quicker fallback, avoid locks when ring is empty */
 109         if (__ptr_ring_empty(r))
 110                 return NULL;
 111
 112         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 113          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 114          */
 115 #ifdef CONFIG_NUMA
 116         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 117 #else
 118         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 119         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 120 #endif
 121
 122         /* Slower-path: Get pages from locked ring queue */
 123         spin_lock(&r->consumer_lock);
 124
 125         /* Refill alloc array, but only if NUMA match */
 126         do {
 127                 page = __ptr_ring_consume(r);
 128                 if (unlikely(!page))
 129                         break;
 130
 131                 if (likely(page_to_nid(page) == pref_nid)) {
 132                         pool->alloc.cache[pool->alloc.count++] = page;
 133                 } else {
 134                         /* NUMA mismatch;
 135                          * (1) release 1 page to page-allocator and
 136                          * (2) break out to fallthrough to alloc_pages_node.
 137                          * This limit stress on page buddy alloactor.
 138                          */
 139                         page_pool_return_page(pool, page);
 140                         page = NULL;
 141                         break;
 142                 }
 143         } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
 144
 145         /* Return last page */
 146         if (likely(pool->alloc.count > 0))
 147                 page = pool->alloc.cache[--pool->alloc.count];
 148
 149         spin_unlock(&r->consumer_lock);
 150         return page;
 151 }
 152
 153 /* fast path */
 154 static struct page *__page_pool_get_cached(struct page_pool *pool)
 155 {
 156         struct page *page;
 157
 158         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 159         if (likely(pool->alloc.count)) {
 160                 /* Fast-path */
 161                 page = pool->alloc.cache[--pool->alloc.count];
 162         } else {
 163                 page = page_pool_refill_alloc_cache(pool);
 164         }
 165
 166         return page;
 167 }
 168
 169 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 170                                           struct page *page,
 171                                           unsigned int dma_sync_size)
 172 {
 173         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 174         dma_sync_single_range_for_device(pool->p.dev, page->dma_addr,
 175                                          pool->p.offset, dma_sync_size,
 176                                          pool->p.dma_dir);
 177 }
 178
 179 /* slow path */
 180 noinline
 181 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 182                                                  gfp_t _gfp)
 183 {
 184         struct page *page;
 185         gfp_t gfp = _gfp;
 186         dma_addr_t dma;
 187
 188         /* We could always set __GFP_COMP, and avoid this branch, as
 189          * prep_new_page() can handle order-0 with __GFP_COMP.
 190          */
 191         if (pool->p.order)
 192                 gfp |= __GFP_COMP;
 193
 194         /* FUTURE development:
 195          *
 196          * Current slow-path essentially falls back to single page
 197          * allocations, which doesn't improve performance.  This code
 198          * need bulk allocation support from the page allocator code.
 199          */
 200
 201         /* Cache was empty, do real allocation */
 202 #ifdef CONFIG_NUMA
 203         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 204 #else
 205         page = alloc_pages(gfp, pool->p.order);
 206 #endif
 207         if (!page)
 208                 return NULL;
 209
 210         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 211                 goto skip_dma_map;
 212
 213         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 214          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 215          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 216          * This mapping is kept for lifetime of page, until leaving pool.
 217          */
 218         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 219                                  (PAGE_SIZE << pool->p.order),
 220                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 221         if (dma_mapping_error(pool->p.dev, dma)) {
 222                 put_page(page);
 223                 return NULL;
 224         }
 225         page->dma_addr = dma;
 226
 227         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 228                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 229
 230 skip_dma_map:
 231         /* Track how many pages are held 'in-flight' */
 232         pool->pages_state_hold_cnt++;
 233
 234         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 235
 236         /* When page just alloc'ed is should/must have refcnt 1. */
 237         return page;
 238 }
 239
 240 /* For using page_pool replace: alloc_pages() API calls, but provide
 241  * synchronization guarantee for allocation side.
 242  */
 243 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 244 {
 245         struct page *page;
 246
 247         /* Fast-path: Get a page from cache */
 248         page = __page_pool_get_cached(pool);
 249         if (page)
 250                 return page;
 251
 252         /* Slow-path: cache empty, do real allocation */
 253         page = __page_pool_alloc_pages_slow(pool, gfp);
 254         return page;
 255 }
 256 EXPORT_SYMBOL(page_pool_alloc_pages);
 257
 258 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 259  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 260  */
 261 #define _distance(a, b) (s32)((a) - (b))
 262
 263 static s32 page_pool_inflight(struct page_pool *pool)
 264 {
 265         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 266         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 267         s32 inflight;
 268
 269         inflight = _distance(hold_cnt, release_cnt);
 270
 271         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 272         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 273
 274         return inflight;
 275 }
 276
 277 /* Disconnects a page (from a page_pool).  API users can have a need
 278  * to disconnect a page (from a page_pool), to allow it to be used as
 279  * a regular page (that will eventually be returned to the normal
 280  * page-allocator via put_page).
 281  */
 282 void page_pool_release_page(struct page_pool *pool, struct page *page)
 283 {
 284         dma_addr_t dma;
 285         int count;
 286
 287         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 288                 /* Always account for inflight pages, even if we didn't
 289                  * map them
 290                  */
 291                 goto skip_dma_unmap;
 292
 293         dma = page->dma_addr;
 294
 295         /* When page is unmapped, it cannot be returned our pool */
 296         dma_unmap_page_attrs(pool->p.dev, dma,
 297                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 298                              DMA_ATTR_SKIP_CPU_SYNC);
 299         page->dma_addr = 0;
 300 skip_dma_unmap:
 301         /* This may be the last page returned, releasing the pool, so
 302          * it is not safe to reference pool afterwards.
 303          */
 304         count = atomic_inc_return(&pool->pages_state_release_cnt);
 305         trace_page_pool_state_release(pool, page, count);
 306 }
 307 EXPORT_SYMBOL(page_pool_release_page);
 308
 309 /* Return a page to the page allocator, cleaning up our state */
 310 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 311 {
 312         page_pool_release_page(pool, page);
 313
 314         put_page(page);
 315         /* An optimization would be to call __free_pages(page, pool->p.order)
 316          * knowing page is not part of page-cache (thus avoiding a
 317          * __page_cache_release() call).
 318          */
 319 }
 320
 321 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 322 {
 323         int ret;
 324         /* BH protection not needed if current is serving softirq */
 325         if (in_serving_softirq())
 326                 ret = ptr_ring_produce(&pool->ring, page);
 327         else
 328                 ret = ptr_ring_produce_bh(&pool->ring, page);
 329
 330         return (ret == 0) ? true : false;
 331 }
 332
 333 /* Only allow direct recycling in special circumstances, into the
 334  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 335  *
 336  * Caller must provide appropriate safe context.
 337  */
 338 static bool page_pool_recycle_in_cache(struct page *page,
 339                                        struct page_pool *pool)
 340 {
 341         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
 342                 return false;
 343
 344         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 345         pool->alloc.cache[pool->alloc.count++] = page;
 346         return true;
 347 }
 348
 349 /* page is NOT reusable when:
 350  * 1) allocated when system is under some pressure. (page_is_pfmemalloc)
 351  */
 352 static bool pool_page_reusable(struct page_pool *pool, struct page *page)
 353 {
 354         return !page_is_pfmemalloc(page);
 355 }
 356
 357 /* If the page refcnt == 1, this will try to recycle the page.
 358  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 359  * the configured size min(dma_sync_size, pool->max_len).
 360  * If the page refcnt != 1, then the page will be returned to memory
 361  * subsystem.
 362  */
 363 void page_pool_put_page(struct page_pool *pool, struct page *page,
 364                         unsigned int dma_sync_size, bool allow_direct)
 365 {
 366         /* This allocator is optimized for the XDP mode that uses
 367          * one-frame-per-page, but have fallbacks that act like the
 368          * regular page allocator APIs.
 369          *
 370          * refcnt == 1 means page_pool owns page, and can recycle it.
 371          */
 372         if (likely(page_ref_count(page) == 1 &&
 373                    pool_page_reusable(pool, page))) {
 374                 /* Read barrier done in page_ref_count / READ_ONCE */
 375
 376                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 377                         page_pool_dma_sync_for_device(pool, page,
 378                                                       dma_sync_size);
 379
 380                 if (allow_direct && in_serving_softirq())
 381                         if (page_pool_recycle_in_cache(page, pool))
 382                                 return;
 383
 384                 if (!page_pool_recycle_in_ring(pool, page)) {
 385                         /* Cache full, fallback to free pages */
 386                         page_pool_return_page(pool, page);
 387                 }
 388                 return;
 389         }
 390         /* Fallback/non-XDP mode: API user have elevated refcnt.
 391          *
 392          * Many drivers split up the page into fragments, and some
 393          * want to keep doing this to save memory and do refcnt based
 394          * recycling. Support this use case too, to ease drivers
 395          * switching between XDP/non-XDP.
 396          *
 397          * In-case page_pool maintains the DMA mapping, API user must
 398          * call page_pool_put_page once.  In this elevated refcnt
 399          * case, the DMA is unmapped/released, as driver is likely
 400          * doing refcnt based recycle tricks, meaning another process
 401          * will be invoking put_page.
 402          */
 403         /* Do not replace this with page_pool_return_page() */
 404         page_pool_release_page(pool, page);
 405         put_page(page);
 406 }
 407 EXPORT_SYMBOL(page_pool_put_page);
 408
 409 static void page_pool_empty_ring(struct page_pool *pool)
 410 {
 411         struct page *page;
 412
 413         /* Empty recycle ring */
 414         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 415                 /* Verify the refcnt invariant of cached pages */
 416                 if (!(page_ref_count(page) == 1))
 417                         pr_crit("%s() page_pool refcnt %d violation\n",
 418                                 __func__, page_ref_count(page));
 419
 420                 page_pool_return_page(pool, page);
 421         }
 422 }
 423
 424 static void page_pool_free(struct page_pool *pool)
 425 {
 426         if (pool->disconnect)
 427                 pool->disconnect(pool);
 428
 429         ptr_ring_cleanup(&pool->ring, NULL);
 430
 431         if (pool->p.flags & PP_FLAG_DMA_MAP)
 432                 put_device(pool->p.dev);
 433
 434         kfree(pool);
 435 }
 436
 437 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 438 {
 439         struct page *page;
 440
 441         if (pool->destroy_cnt)
 442                 return;
 443
 444         /* Empty alloc cache, assume caller made sure this is
 445          * no-longer in use, and page_pool_alloc_pages() cannot be
 446          * call concurrently.
 447          */
 448         while (pool->alloc.count) {
 449                 page = pool->alloc.cache[--pool->alloc.count];
 450                 page_pool_return_page(pool, page);
 451         }
 452 }
 453
 454 static void page_pool_scrub(struct page_pool *pool)
 455 {
 456         page_pool_empty_alloc_cache_once(pool);
 457         pool->destroy_cnt++;
 458
 459         /* No more consumers should exist, but producers could still
 460          * be in-flight.
 461          */
 462         page_pool_empty_ring(pool);
 463 }
 464
 465 static int page_pool_release(struct page_pool *pool)
 466 {
 467         int inflight;
 468
 469         page_pool_scrub(pool);
 470         inflight = page_pool_inflight(pool);
 471         if (!inflight)
 472                 page_pool_free(pool);
 473
 474         return inflight;
 475 }
 476
 477 static void page_pool_release_retry(struct work_struct *wq)
 478 {
 479         struct delayed_work *dwq = to_delayed_work(wq);
 480         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 481         int inflight;
 482
 483         inflight = page_pool_release(pool);
 484         if (!inflight)
 485                 return;
 486
 487         /* Periodic warning */
 488         if (time_after_eq(jiffies, pool->defer_warn)) {
 489                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 490
 491                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 492                         __func__, inflight, sec);
 493                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 494         }
 495
 496         /* Still not ready to be disconnected, retry later */
 497         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 498 }
 499
 500 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *))
 501 {
 502         refcount_inc(&pool->user_cnt);
 503         pool->disconnect = disconnect;
 504 }
 505
 506 void page_pool_destroy(struct page_pool *pool)
 507 {
 508         if (!pool)
 509                 return;
 510
 511         if (!page_pool_put(pool))
 512                 return;
 513
 514         if (!page_pool_release(pool))
 515                 return;
 516
 517         pool->defer_start = jiffies;
 518         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 519
 520         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 521         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 522 }
 523 EXPORT_SYMBOL(page_pool_destroy);
 524
 525 /* Caller must provide appropriate safe context, e.g. NAPI. */
 526 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 527 {
 528         struct page *page;
 529
 530         trace_page_pool_update_nid(pool, new_nid);
 531         pool->p.nid = new_nid;
 532
 533         /* Flush pool alloc cache, as refill will check NUMA node */
 534         while (pool->alloc.count) {
 535                 page = pool->alloc.cache[--pool->alloc.count];
 536                 page_pool_return_page(pool, page);
 537         }
 538 }
 539 EXPORT_SYMBOL(page_pool_update_nid);