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_addr_t dma_addr = page_pool_get_dma_addr(page);
 176
 177         dma_sync_size = min(dma_sync_size, pool->p.max_len);
 178         dma_sync_single_range_for_device(pool->p.dev, dma_addr,
 179                                          pool->p.offset, dma_sync_size,
 180                                          pool->p.dma_dir);
 181 }
 182
 183 /* slow path */
 184 noinline
 185 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 186                                                  gfp_t _gfp)
 187 {
 188         struct page *page;
 189         gfp_t gfp = _gfp;
 190         dma_addr_t dma;
 191
 192         /* We could always set __GFP_COMP, and avoid this branch, as
 193          * prep_new_page() can handle order-0 with __GFP_COMP.
 194          */
 195         if (pool->p.order)
 196                 gfp |= __GFP_COMP;
 197
 198         /* FUTURE development:
 199          *
 200          * Current slow-path essentially falls back to single page
 201          * allocations, which doesn't improve performance.  This code
 202          * need bulk allocation support from the page allocator code.
 203          */
 204
 205         /* Cache was empty, do real allocation */
 206 #ifdef CONFIG_NUMA
 207         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 208 #else
 209         page = alloc_pages(gfp, pool->p.order);
 210 #endif
 211         if (!page)
 212                 return NULL;
 213
 214         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 215                 goto skip_dma_map;
 216
 217         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 218          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 219          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 220          * This mapping is kept for lifetime of page, until leaving pool.
 221          */
 222         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 223                                  (PAGE_SIZE << pool->p.order),
 224                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 225         if (dma_mapping_error(pool->p.dev, dma)) {
 226                 put_page(page);
 227                 return NULL;
 228         }
 229         page_pool_set_dma_addr(page, dma);
 230
 231         if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 232                 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
 233
 234 skip_dma_map:
 235         /* Track how many pages are held 'in-flight' */
 236         pool->pages_state_hold_cnt++;
 237
 238         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 239
 240         /* When page just alloc'ed is should/must have refcnt 1. */
 241         return page;
 242 }
 243
 244 /* For using page_pool replace: alloc_pages() API calls, but provide
 245  * synchronization guarantee for allocation side.
 246  */
 247 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 248 {
 249         struct page *page;
 250
 251         /* Fast-path: Get a page from cache */
 252         page = __page_pool_get_cached(pool);
 253         if (page)
 254                 return page;
 255
 256         /* Slow-path: cache empty, do real allocation */
 257         page = __page_pool_alloc_pages_slow(pool, gfp);
 258         return page;
 259 }
 260 EXPORT_SYMBOL(page_pool_alloc_pages);
 261
 262 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 263  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 264  */
 265 #define _distance(a, b) (s32)((a) - (b))
 266
 267 static s32 page_pool_inflight(struct page_pool *pool)
 268 {
 269         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 270         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 271         s32 inflight;
 272
 273         inflight = _distance(hold_cnt, release_cnt);
 274
 275         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 276         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 277
 278         return inflight;
 279 }
 280
 281 /* Disconnects a page (from a page_pool).  API users can have a need
 282  * to disconnect a page (from a page_pool), to allow it to be used as
 283  * a regular page (that will eventually be returned to the normal
 284  * page-allocator via put_page).
 285  */
 286 void page_pool_release_page(struct page_pool *pool, struct page *page)
 287 {
 288         dma_addr_t dma;
 289         int count;
 290
 291         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 292                 /* Always account for inflight pages, even if we didn't
 293                  * map them
 294                  */
 295                 goto skip_dma_unmap;
 296
 297         dma = page_pool_get_dma_addr(page);
 298
 299         /* When page is unmapped, it cannot be returned to our pool */
 300         dma_unmap_page_attrs(pool->p.dev, dma,
 301                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 302                              DMA_ATTR_SKIP_CPU_SYNC);
 303         page_pool_set_dma_addr(page, 0);
 304 skip_dma_unmap:
 305         /* This may be the last page returned, releasing the pool, so
 306          * it is not safe to reference pool afterwards.
 307          */
 308         count = atomic_inc_return(&pool->pages_state_release_cnt);
 309         trace_page_pool_state_release(pool, page, count);
 310 }
 311 EXPORT_SYMBOL(page_pool_release_page);
 312
 313 /* Return a page to the page allocator, cleaning up our state */
 314 static void page_pool_return_page(struct page_pool *pool, struct page *page)
 315 {
 316         page_pool_release_page(pool, page);
 317
 318         put_page(page);
 319         /* An optimization would be to call __free_pages(page, pool->p.order)
 320          * knowing page is not part of page-cache (thus avoiding a
 321          * __page_cache_release() call).
 322          */
 323 }
 324
 325 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
 326 {
 327         int ret;
 328         /* BH protection not needed if current is serving softirq */
 329         if (in_serving_softirq())
 330                 ret = ptr_ring_produce(&pool->ring, page);
 331         else
 332                 ret = ptr_ring_produce_bh(&pool->ring, page);
 333
 334         return (ret == 0) ? true : false;
 335 }
 336
 337 /* Only allow direct recycling in special circumstances, into the
 338  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 339  *
 340  * Caller must provide appropriate safe context.
 341  */
 342 static bool page_pool_recycle_in_cache(struct page *page,
 343                                        struct page_pool *pool)
 344 {
 345         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
 346                 return false;
 347
 348         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 349         pool->alloc.cache[pool->alloc.count++] = page;
 350         return true;
 351 }
 352
 353 /* page is NOT reusable when:
 354  * 1) allocated when system is under some pressure. (page_is_pfmemalloc)
 355  */
 356 static bool pool_page_reusable(struct page_pool *pool, struct page *page)
 357 {
 358         return !page_is_pfmemalloc(page);
 359 }
 360
 361 /* If the page refcnt == 1, this will try to recycle the page.
 362  * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
 363  * the configured size min(dma_sync_size, pool->max_len).
 364  * If the page refcnt != 1, then the page will be returned to memory
 365  * subsystem.
 366  */
 367 void page_pool_put_page(struct page_pool *pool, struct page *page,
 368                         unsigned int dma_sync_size, bool allow_direct)
 369 {
 370         /* This allocator is optimized for the XDP mode that uses
 371          * one-frame-per-page, but have fallbacks that act like the
 372          * regular page allocator APIs.
 373          *
 374          * refcnt == 1 means page_pool owns page, and can recycle it.
 375          */
 376         if (likely(page_ref_count(page) == 1 &&
 377                    pool_page_reusable(pool, page))) {
 378                 /* Read barrier done in page_ref_count / READ_ONCE */
 379
 380                 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
 381                         page_pool_dma_sync_for_device(pool, page,
 382                                                       dma_sync_size);
 383
 384                 if (allow_direct && in_serving_softirq())
 385                         if (page_pool_recycle_in_cache(page, pool))
 386                                 return;
 387
 388                 if (!page_pool_recycle_in_ring(pool, page)) {
 389                         /* Cache full, fallback to free pages */
 390                         page_pool_return_page(pool, page);
 391                 }
 392                 return;
 393         }
 394         /* Fallback/non-XDP mode: API user have elevated refcnt.
 395          *
 396          * Many drivers split up the page into fragments, and some
 397          * want to keep doing this to save memory and do refcnt based
 398          * recycling. Support this use case too, to ease drivers
 399          * switching between XDP/non-XDP.
 400          *
 401          * In-case page_pool maintains the DMA mapping, API user must
 402          * call page_pool_put_page once.  In this elevated refcnt
 403          * case, the DMA is unmapped/released, as driver is likely
 404          * doing refcnt based recycle tricks, meaning another process
 405          * will be invoking put_page.
 406          */
 407         /* Do not replace this with page_pool_return_page() */
 408         page_pool_release_page(pool, page);
 409         put_page(page);
 410 }
 411 EXPORT_SYMBOL(page_pool_put_page);
 412
 413 static void page_pool_empty_ring(struct page_pool *pool)
 414 {
 415         struct page *page;
 416
 417         /* Empty recycle ring */
 418         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 419                 /* Verify the refcnt invariant of cached pages */
 420                 if (!(page_ref_count(page) == 1))
 421                         pr_crit("%s() page_pool refcnt %d violation\n",
 422                                 __func__, page_ref_count(page));
 423
 424                 page_pool_return_page(pool, page);
 425         }
 426 }
 427
 428 static void page_pool_free(struct page_pool *pool)
 429 {
 430         if (pool->disconnect)
 431                 pool->disconnect(pool);
 432
 433         ptr_ring_cleanup(&pool->ring, NULL);
 434
 435         if (pool->p.flags & PP_FLAG_DMA_MAP)
 436                 put_device(pool->p.dev);
 437
 438         kfree(pool);
 439 }
 440
 441 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 442 {
 443         struct page *page;
 444
 445         if (pool->destroy_cnt)
 446                 return;
 447
 448         /* Empty alloc cache, assume caller made sure this is
 449          * no-longer in use, and page_pool_alloc_pages() cannot be
 450          * call concurrently.
 451          */
 452         while (pool->alloc.count) {
 453                 page = pool->alloc.cache[--pool->alloc.count];
 454                 page_pool_return_page(pool, page);
 455         }
 456 }
 457
 458 static void page_pool_scrub(struct page_pool *pool)
 459 {
 460         page_pool_empty_alloc_cache_once(pool);
 461         pool->destroy_cnt++;
 462
 463         /* No more consumers should exist, but producers could still
 464          * be in-flight.
 465          */
 466         page_pool_empty_ring(pool);
 467 }
 468
 469 static int page_pool_release(struct page_pool *pool)
 470 {
 471         int inflight;
 472
 473         page_pool_scrub(pool);
 474         inflight = page_pool_inflight(pool);
 475         if (!inflight)
 476                 page_pool_free(pool);
 477
 478         return inflight;
 479 }
 480
 481 static void page_pool_release_retry(struct work_struct *wq)
 482 {
 483         struct delayed_work *dwq = to_delayed_work(wq);
 484         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 485         int inflight;
 486
 487         inflight = page_pool_release(pool);
 488         if (!inflight)
 489                 return;
 490
 491         /* Periodic warning */
 492         if (time_after_eq(jiffies, pool->defer_warn)) {
 493                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 494
 495                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 496                         __func__, inflight, sec);
 497                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 498         }
 499
 500         /* Still not ready to be disconnected, retry later */
 501         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 502 }
 503
 504 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *))
 505 {
 506         refcount_inc(&pool->user_cnt);
 507         pool->disconnect = disconnect;
 508 }
 509
 510 void page_pool_destroy(struct page_pool *pool)
 511 {
 512         if (!pool)
 513                 return;
 514
 515         if (!page_pool_put(pool))
 516                 return;
 517
 518         if (!page_pool_release(pool))
 519                 return;
 520
 521         pool->defer_start = jiffies;
 522         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 523
 524         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 525         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 526 }
 527 EXPORT_SYMBOL(page_pool_destroy);
 528
 529 /* Caller must provide appropriate safe context, e.g. NAPI. */
 530 void page_pool_update_nid(struct page_pool *pool, int new_nid)
 531 {
 532         struct page *page;
 533
 534         trace_page_pool_update_nid(pool, new_nid);
 535         pool->p.nid = new_nid;
 536
 537         /* Flush pool alloc cache, as refill will check NUMA node */
 538         while (pool->alloc.count) {
 539                 page = pool->alloc.cache[--pool->alloc.count];
 540                 page_pool_return_page(pool, page);
 541         }
 542 }
 543 EXPORT_SYMBOL(page_pool_update_nid);