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 (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
  51                 return -ENOMEM;
  52
  53         atomic_set(&pool->pages_state_release_cnt, 0);
  54
  55         /* Driver calling page_pool_create() also call page_pool_destroy() */
  56         refcount_set(&pool->user_cnt, 1);
  57
  58         if (pool->p.flags & PP_FLAG_DMA_MAP)
  59                 get_device(pool->p.dev);
  60
  61         return 0;
  62 }
  63
  64 struct page_pool *page_pool_create(const struct page_pool_params *params)
  65 {
  66         struct page_pool *pool;
  67         int err;
  68
  69         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
  70         if (!pool)
  71                 return ERR_PTR(-ENOMEM);
  72
  73         err = page_pool_init(pool, params);
  74         if (err < 0) {
  75                 pr_warn("%s() gave up with errno %d\n", __func__, err);
  76                 kfree(pool);
  77                 return ERR_PTR(err);
  78         }
  79
  80         return pool;
  81 }
  82 EXPORT_SYMBOL(page_pool_create);
  83
  84 /* fast path */
  85 static struct page *__page_pool_get_cached(struct page_pool *pool)
  86 {
  87         struct ptr_ring *r = &pool->ring;
  88         bool refill = false;
  89         struct page *page;
  90
  91         /* Test for safe-context, caller should provide this guarantee */
  92         if (likely(in_serving_softirq())) {
  93                 if (likely(pool->alloc.count)) {
  94                         /* Fast-path */
  95                         page = pool->alloc.cache[--pool->alloc.count];
  96                         return page;
  97                 }
  98                 refill = true;
  99         }
 100
 101         /* Quicker fallback, avoid locks when ring is empty */
 102         if (__ptr_ring_empty(r))
 103                 return NULL;
 104
 105         /* Slow-path: Get page from locked ring queue,
 106          * refill alloc array if requested.
 107          */
 108         spin_lock(&r->consumer_lock);
 109         page = __ptr_ring_consume(r);
 110         if (refill)
 111                 pool->alloc.count = __ptr_ring_consume_batched(r,
 112                                                         pool->alloc.cache,
 113                                                         PP_ALLOC_CACHE_REFILL);
 114         spin_unlock(&r->consumer_lock);
 115         return page;
 116 }
 117
 118 /* slow path */
 119 noinline
 120 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 121                                                  gfp_t _gfp)
 122 {
 123         struct page *page;
 124         gfp_t gfp = _gfp;
 125         dma_addr_t dma;
 126
 127         /* We could always set __GFP_COMP, and avoid this branch, as
 128          * prep_new_page() can handle order-0 with __GFP_COMP.
 129          */
 130         if (pool->p.order)
 131                 gfp |= __GFP_COMP;
 132
 133         /* FUTURE development:
 134          *
 135          * Current slow-path essentially falls back to single page
 136          * allocations, which doesn't improve performance.  This code
 137          * need bulk allocation support from the page allocator code.
 138          */
 139
 140         /* Cache was empty, do real allocation */
 141         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 142         if (!page)
 143                 return NULL;
 144
 145         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 146                 goto skip_dma_map;
 147
 148         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 149          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 150          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 151          * This mapping is kept for lifetime of page, until leaving pool.
 152          */
 153         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 154                                  (PAGE_SIZE << pool->p.order),
 155                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 156         if (dma_mapping_error(pool->p.dev, dma)) {
 157                 put_page(page);
 158                 return NULL;
 159         }
 160         page->dma_addr = dma;
 161
 162 skip_dma_map:
 163         /* Track how many pages are held 'in-flight' */
 164         pool->pages_state_hold_cnt++;
 165
 166         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 167
 168         /* When page just alloc'ed is should/must have refcnt 1. */
 169         return page;
 170 }
 171
 172 /* For using page_pool replace: alloc_pages() API calls, but provide
 173  * synchronization guarantee for allocation side.
 174  */
 175 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 176 {
 177         struct page *page;
 178
 179         /* Fast-path: Get a page from cache */
 180         page = __page_pool_get_cached(pool);
 181         if (page)
 182                 return page;
 183
 184         /* Slow-path: cache empty, do real allocation */
 185         page = __page_pool_alloc_pages_slow(pool, gfp);
 186         return page;
 187 }
 188 EXPORT_SYMBOL(page_pool_alloc_pages);
 189
 190 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 191  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 192  */
 193 #define _distance(a, b) (s32)((a) - (b))
 194
 195 static s32 page_pool_inflight(struct page_pool *pool)
 196 {
 197         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 198         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 199         s32 inflight;
 200
 201         inflight = _distance(hold_cnt, release_cnt);
 202
 203         trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
 204         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 205
 206         return inflight;
 207 }
 208
 209 /* Cleanup page_pool state from page */
 210 static void __page_pool_clean_page(struct page_pool *pool,
 211                                    struct page *page)
 212 {
 213         dma_addr_t dma;
 214         int count;
 215
 216         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 217                 goto skip_dma_unmap;
 218
 219         dma = page->dma_addr;
 220         /* DMA unmap */
 221         dma_unmap_page_attrs(pool->p.dev, dma,
 222                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 223                              DMA_ATTR_SKIP_CPU_SYNC);
 224         page->dma_addr = 0;
 225 skip_dma_unmap:
 226         /* This may be the last page returned, releasing the pool, so
 227          * it is not safe to reference pool afterwards.
 228          */
 229         count = atomic_inc_return(&pool->pages_state_release_cnt);
 230         trace_page_pool_state_release(pool, page, count);
 231 }
 232
 233 /* unmap the page and clean our state */
 234 void page_pool_unmap_page(struct page_pool *pool, struct page *page)
 235 {
 236         /* When page is unmapped, this implies page will not be
 237          * returned to page_pool.
 238          */
 239         __page_pool_clean_page(pool, page);
 240 }
 241 EXPORT_SYMBOL(page_pool_unmap_page);
 242
 243 /* Return a page to the page allocator, cleaning up our state */
 244 static void __page_pool_return_page(struct page_pool *pool, struct page *page)
 245 {
 246         __page_pool_clean_page(pool, page);
 247
 248         put_page(page);
 249         /* An optimization would be to call __free_pages(page, pool->p.order)
 250          * knowing page is not part of page-cache (thus avoiding a
 251          * __page_cache_release() call).
 252          */
 253 }
 254
 255 static bool __page_pool_recycle_into_ring(struct page_pool *pool,
 256                                    struct page *page)
 257 {
 258         int ret;
 259         /* BH protection not needed if current is serving softirq */
 260         if (in_serving_softirq())
 261                 ret = ptr_ring_produce(&pool->ring, page);
 262         else
 263                 ret = ptr_ring_produce_bh(&pool->ring, page);
 264
 265         return (ret == 0) ? true : false;
 266 }
 267
 268 /* Only allow direct recycling in special circumstances, into the
 269  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 270  *
 271  * Caller must provide appropriate safe context.
 272  */
 273 static bool __page_pool_recycle_direct(struct page *page,
 274                                        struct page_pool *pool)
 275 {
 276         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
 277                 return false;
 278
 279         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 280         pool->alloc.cache[pool->alloc.count++] = page;
 281         return true;
 282 }
 283
 284 void __page_pool_put_page(struct page_pool *pool,
 285                           struct page *page, bool allow_direct)
 286 {
 287         /* This allocator is optimized for the XDP mode that uses
 288          * one-frame-per-page, but have fallbacks that act like the
 289          * regular page allocator APIs.
 290          *
 291          * refcnt == 1 means page_pool owns page, and can recycle it.
 292          */
 293         if (likely(page_ref_count(page) == 1)) {
 294                 /* Read barrier done in page_ref_count / READ_ONCE */
 295
 296                 if (allow_direct && in_serving_softirq())
 297                         if (__page_pool_recycle_direct(page, pool))
 298                                 return;
 299
 300                 if (!__page_pool_recycle_into_ring(pool, page)) {
 301                         /* Cache full, fallback to free pages */
 302                         __page_pool_return_page(pool, page);
 303                 }
 304                 return;
 305         }
 306         /* Fallback/non-XDP mode: API user have elevated refcnt.
 307          *
 308          * Many drivers split up the page into fragments, and some
 309          * want to keep doing this to save memory and do refcnt based
 310          * recycling. Support this use case too, to ease drivers
 311          * switching between XDP/non-XDP.
 312          *
 313          * In-case page_pool maintains the DMA mapping, API user must
 314          * call page_pool_put_page once.  In this elevated refcnt
 315          * case, the DMA is unmapped/released, as driver is likely
 316          * doing refcnt based recycle tricks, meaning another process
 317          * will be invoking put_page.
 318          */
 319         __page_pool_clean_page(pool, page);
 320         put_page(page);
 321 }
 322 EXPORT_SYMBOL(__page_pool_put_page);
 323
 324 static void __page_pool_empty_ring(struct page_pool *pool)
 325 {
 326         struct page *page;
 327
 328         /* Empty recycle ring */
 329         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 330                 /* Verify the refcnt invariant of cached pages */
 331                 if (!(page_ref_count(page) == 1))
 332                         pr_crit("%s() page_pool refcnt %d violation\n",
 333                                 __func__, page_ref_count(page));
 334
 335                 __page_pool_return_page(pool, page);
 336         }
 337 }
 338
 339 static void page_pool_free(struct page_pool *pool)
 340 {
 341         if (pool->disconnect)
 342                 pool->disconnect(pool);
 343
 344         ptr_ring_cleanup(&pool->ring, NULL);
 345
 346         if (pool->p.flags & PP_FLAG_DMA_MAP)
 347                 put_device(pool->p.dev);
 348
 349         kfree(pool);
 350 }
 351
 352 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
 353 {
 354         struct page *page;
 355
 356         if (pool->destroy_cnt)
 357                 return;
 358
 359         /* Empty alloc cache, assume caller made sure this is
 360          * no-longer in use, and page_pool_alloc_pages() cannot be
 361          * call concurrently.
 362          */
 363         while (pool->alloc.count) {
 364                 page = pool->alloc.cache[--pool->alloc.count];
 365                 __page_pool_return_page(pool, page);
 366         }
 367 }
 368
 369 static void page_pool_scrub(struct page_pool *pool)
 370 {
 371         page_pool_empty_alloc_cache_once(pool);
 372         pool->destroy_cnt++;
 373
 374         /* No more consumers should exist, but producers could still
 375          * be in-flight.
 376          */
 377         __page_pool_empty_ring(pool);
 378 }
 379
 380 static int page_pool_release(struct page_pool *pool)
 381 {
 382         int inflight;
 383
 384         page_pool_scrub(pool);
 385         inflight = page_pool_inflight(pool);
 386         if (!inflight)
 387                 page_pool_free(pool);
 388
 389         return inflight;
 390 }
 391
 392 static void page_pool_release_retry(struct work_struct *wq)
 393 {
 394         struct delayed_work *dwq = to_delayed_work(wq);
 395         struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
 396         int inflight;
 397
 398         inflight = page_pool_release(pool);
 399         if (!inflight)
 400                 return;
 401
 402         /* Periodic warning */
 403         if (time_after_eq(jiffies, pool->defer_warn)) {
 404                 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
 405
 406                 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
 407                         __func__, inflight, sec);
 408                 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
 409         }
 410
 411         /* Still not ready to be disconnected, retry later */
 412         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 413 }
 414
 415 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *))
 416 {
 417         refcount_inc(&pool->user_cnt);
 418         pool->disconnect = disconnect;
 419 }
 420
 421 void page_pool_destroy(struct page_pool *pool)
 422 {
 423         if (!pool)
 424                 return;
 425
 426         if (!page_pool_put(pool))
 427                 return;
 428
 429         if (!page_pool_release(pool))
 430                 return;
 431
 432         pool->defer_start = jiffies;
 433         pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
 434
 435         INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
 436         schedule_delayed_work(&pool->release_dw, DEFER_TIME);
 437 }
 438 EXPORT_SYMBOL(page_pool_destroy);