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.flags & PP_FLAG_DMA_MAP) {
  47                 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
  48                     (pool->p.dma_dir != DMA_BIDIRECTIONAL))
  49                         return -EINVAL;
  50         }
  51
  52         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
  53                 /* In order to request DMA-sync-for-device the page
  54                  * needs to be mapped
  55                  */
  56                 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
  57                         return -EINVAL;
  58
  59                 if (!pool->p.max_len)
  60                         return -EINVAL;
  61
  62                 /* pool->p.offset has to be set according to the address
  63                  * offset used by the DMA engine to start copying rx data
  64                  */
  65         }
  66
  67         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
  68                 return -ENOMEM;
  69
  70         atomic_set(&pool->pages_state_release_cnt, 0);
  71
  72         /* Driver calling page_pool_create() also call page_pool_destroy() */
  73         refcount_set(&pool->user_cnt, 1);
  74
  75         if (pool->p.flags & PP_FLAG_DMA_MAP)
  76                 get_device(pool->p.dev);
  77
  78         return 0;
  79 }
  80
  81 struct page_pool *page_pool_create(const struct page_pool_params *params)
  82 {
  83         struct page_pool *pool;
  84         int err;
  85
  86         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
  87         if (!pool)
  88                 return ERR_PTR(-ENOMEM);
  89
  90         err = page_pool_init(pool, params);
  91         if (err < 0) {
  92                 pr_warn("%s() gave up with errno %d\n", __func__, err);
  93                 kfree(pool);
  94                 return ERR_PTR(err);
  95         }
  96
  97         return pool;
  98 }
  99 EXPORT_SYMBOL(page_pool_create);
 100
 101 static void page_pool_return_page(struct page_pool *pool, struct page *page);
 102
 103 noinline
 104 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
 105 {
 106         struct ptr_ring *r = &pool->ring;
 107         struct page *page;
 108         int pref_nid; /* preferred NUMA node */
 109
 110         /* Quicker fallback, avoid locks when ring is empty */
 111         if (__ptr_ring_empty(r))
 112                 return NULL;
 113
 114         /* Softirq guarantee CPU and thus NUMA node is stable. This,
 115          * assumes CPU refilling driver RX-ring will also run RX-NAPI.
 116          */
 117 #ifdef CONFIG_NUMA
 118         pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
 119 #else
 120         /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
 121         pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
 122 #endif
 123
 124         /* Slower-path: Get pages from locked ring queue */
 125         spin_lock(&r->consumer_lock);
 126
 127         /* Refill alloc array, but only if NUMA match */
 128         do {
 129                 page = __ptr_ring_consume(r);
 130                 if (unlikely(!page))
 131                         break;
 132
 133                 if (likely(page_to_nid(page) == pref_nid)) {
 134                         pool->alloc.cache[pool->alloc.count++] = page;
 135                 } else {
 136                         /* NUMA mismatch;
 137                          * (1) release 1 page to page-allocator and
 138                          * (2) break out to fallthrough to alloc_pages_node.
 139                          * This limit stress on page buddy alloactor.
 140                          */
 141                         page_pool_return_page(pool, page);
 142                         page = NULL;
 143                         break;
 144                 }
 145         } while (pool->alloc.count < PP_ALLOC_CACHE_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         struct page *page;
 159
 160         /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
 161         if (likely(pool->alloc.count)) {
 162                 /* Fast-path */
 163                 page = pool->alloc.cache[--pool->alloc.count];
 164         } else {
 165                 page = page_pool_refill_alloc_cache(pool);
 166         }
 167
 168         return page;
 169 }
 170
 171 static void page_pool_dma_sync_for_device(struct page_pool *pool,
 172                                           struct page *page,
 173                                           unsigned int dma_sync_size)
 174 {
 175         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 176         dma_sync_single_range_for_device(pool->p.dev, page->dma_addr,
 177                                          pool->p.offset, dma_sync_size,
 178                                          pool->p.dma_dir);
 179 }
 180
 181 /* slow path */
 182 noinline
 183 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 184                                                  gfp_t _gfp)
 185 {
 186         struct page *page;
 187         gfp_t gfp = _gfp;
 188         dma_addr_t dma;
 189
 190         /* We could always set __GFP_COMP, and avoid this branch, as
 191          * prep_new_page() can handle order-0 with __GFP_COMP.
 192          */
 193         if (pool->p.order)
 194                 gfp |= __GFP_COMP;
 195
 196         /* FUTURE development:
 197          *
 198          * Current slow-path essentially falls back to single page
 199          * allocations, which doesn't improve performance.  This code
 200          * need bulk allocation support from the page allocator code.
 201          */
 202
 203         /* Cache was empty, do real allocation */
 204 #ifdef CONFIG_NUMA
 205         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 206 #else
 207         page = alloc_pages(gfp, pool->p.order);
 208 #endif
 209         if (!page)
 210                 return NULL;
 211
 212         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 213                 goto skip_dma_map;
 214
 215         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 216          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 217          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 218          * This mapping is kept for lifetime of page, until leaving pool.
 219          */
 220         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 221                                  (PAGE_SIZE << pool->p.order),
 222                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 223         if (dma_mapping_error(pool->p.dev, dma)) {
 224                 put_page(page);
 225                 return NULL;
 226         }
 227         page->dma_addr = dma;
 228
 229         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 230                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 231
 232 skip_dma_map:
 233         /* Track how many pages are held 'in-flight' */
 234         pool->pages_state_hold_cnt++;
 235
 236         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 237
 238         /* When page just alloc'ed is should/must have refcnt 1. */
 239         return page;
 240 }
 241
 242 /* For using page_pool replace: alloc_pages() API calls, but provide
 243  * synchronization guarantee for allocation side.
 244  */
 245 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 246 {
 247         struct page *page;
 248
 249         /* Fast-path: Get a page from cache */
 250         page = __page_pool_get_cached(pool);
 251         if (page)
 252                 return page;
 253
 254         /* Slow-path: cache empty, do real allocation */
 255         page = __page_pool_alloc_pages_slow(pool, gfp);
 256         return page;
 257 }
 258 EXPORT_SYMBOL(page_pool_alloc_pages);
 259
 260 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 261  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 262  */
 263 #define _distance(a, b) (s32)((a) - (b))
 264
 265 static s32 page_pool_inflight(struct page_pool *pool)
 266 {
 267         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 268         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 269         s32 inflight;
 270
 271         inflight = _distance(hold_cnt, release_cnt);
 272
 273         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 274         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 275
 276         return inflight;
 277 }
 278
 279 /* Disconnects a page (from a page_pool).  API users can have a need
 280  * to disconnect a page (from a page_pool), to allow it to be used as
 281  * a regular page (that will eventually be returned to the normal
 282  * page-allocator via put_page).
 283  */
 284 void page_pool_release_page(struct page_pool *pool, struct page *page)
 285 {
 286         dma_addr_t dma;
 287         int count;
 288
 289         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 290                 /* Always account for inflight pages, even if we didn't
 291                  * map them
 292                  */
 293                 goto skip_dma_unmap;
 294
 295         dma = page->dma_addr;
 296
 297         /* When page is unmapped, it cannot be returned our pool */
 298         dma_unmap_page_attrs(pool->p.dev, dma,
 299                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 300                              DMA_ATTR_SKIP_CPU_SYNC);
 301         page->dma_addr = 0;
 302 skip_dma_unmap:
 303         /* This may be the last page returned, releasing the pool, so
 304          * it is not safe to reference pool afterwards.
 305          */
 306         count = atomic_inc_return(&pool->pages_state_release_cnt);
 307         trace_page_pool_state_release(pool, page, count);
 308 }
 309 EXPORT_SYMBOL(page_pool_release_page);
 310
 311 /* Return a page to the page allocator, cleaning up our state */
 312 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 313 {
 314         page_pool_release_page(pool, page);
 315
 316         put_page(page);
 317         /* An optimization would be to call __free_pages(page, pool->p.order)
 318          * knowing page is not part of page-cache (thus avoiding a
 319          * __page_cache_release() call).
 320          */
 321 }
 322
 323 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 324 {
 325         int ret;
 326         /* BH protection not needed if current is serving softirq */
 327         if (in_serving_softirq())
 328                 ret = ptr_ring_produce(&pool->ring, page);
 329         else
 330                 ret = ptr_ring_produce_bh(&pool->ring, page);
 331
 332         return (ret == 0) ? true : false;
 333 }
 334
 335 /* Only allow direct recycling in special circumstances, into the
 336  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 337  *
 338  * Caller must provide appropriate safe context.
 339  */
 340 static bool page_pool_recycle_in_cache(struct page *page,
 341                                        struct page_pool *pool)
 342 {
 343         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
 344                 return false;
 345
 346         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 347         pool->alloc.cache[pool->alloc.count++] = page;
 348         return true;
 349 }
 350
 351 /* page is NOT reusable when:
 352  * 1) allocated when system is under some pressure. (page_is_pfmemalloc)
 353  */
 354 static bool pool_page_reusable(struct page_pool *pool, struct page *page)
 355 {
 356         return !page_is_pfmemalloc(page);
 357 }
 358
 359 /* If the page refcnt == 1, this will try to recycle the page.
 360  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 361  * the configured size min(dma_sync_size, pool->max_len).
 362  * If the page refcnt != 1, then the page will be returned to memory
 363  * subsystem.
 364  */
 365 void page_pool_put_page(struct page_pool *pool, struct page *page,
 366                         unsigned int dma_sync_size, bool allow_direct)
 367 {
 368         /* This allocator is optimized for the XDP mode that uses
 369          * one-frame-per-page, but have fallbacks that act like the
 370          * regular page allocator APIs.
 371          *
 372          * refcnt == 1 means page_pool owns page, and can recycle it.
 373          */
 374         if (likely(page_ref_count(page) == 1 &&
 375                    pool_page_reusable(pool, page))) {
 376                 /* Read barrier done in page_ref_count / READ_ONCE */
 377
 378                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 379                         page_pool_dma_sync_for_device(pool, page,
 380                                                       dma_sync_size);
 381
 382                 if (allow_direct && in_serving_softirq())
 383                         if (page_pool_recycle_in_cache(page, pool))
 384                                 return;
 385
 386                 if (!page_pool_recycle_in_ring(pool, page)) {
 387                         /* Cache full, fallback to free pages */
 388                         page_pool_return_page(pool, page);
 389                 }
 390                 return;
 391         }
 392         /* Fallback/non-XDP mode: API user have elevated refcnt.
 393          *
 394          * Many drivers split up the page into fragments, and some
 395          * want to keep doing this to save memory and do refcnt based
 396          * recycling. Support this use case too, to ease drivers
 397          * switching between XDP/non-XDP.
 398          *
 399          * In-case page_pool maintains the DMA mapping, API user must
 400          * call page_pool_put_page once.  In this elevated refcnt
 401          * case, the DMA is unmapped/released, as driver is likely
 402          * doing refcnt based recycle tricks, meaning another process
 403          * will be invoking put_page.
 404          */
 405         /* Do not replace this with page_pool_return_page() */
 406         page_pool_release_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);