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 static int page_pool_init(struct page_pool *pool,
  22                           const struct page_pool_params *params)
  23 {
  24         unsigned int ring_qsize = 1024; /* Default */
  25
  26         memcpy(&pool->p, params, sizeof(pool->p));
  27
  28         /* Validate only known flags were used */
  29         if (pool->p.flags & ~(PP_FLAG_ALL))
  30                 return -EINVAL;
  31
  32         if (pool->p.pool_size)
  33                 ring_qsize = pool->p.pool_size;
  34
  35         /* Sanity limit mem that can be pinned down */
  36         if (ring_qsize > 32768)
  37                 return -E2BIG;
  38
  39         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
  40          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
  41          * which is the XDP_TX use-case.
  42          */
  43         if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
  44             (pool->p.dma_dir != DMA_BIDIRECTIONAL))
  45                 return -EINVAL;
  46
  47         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
  48                 return -ENOMEM;
  49
  50         atomic_set(&pool->pages_state_release_cnt, 0);
  51
  52         /* Driver calling page_pool_create() also call page_pool_destroy() */
  53         refcount_set(&pool->user_cnt, 1);
  54
  55         if (pool->p.flags & PP_FLAG_DMA_MAP)
  56                 get_device(pool->p.dev);
  57
  58         return 0;
  59 }
  60
  61 struct page_pool *page_pool_create(const struct page_pool_params *params)
  62 {
  63         struct page_pool *pool;
  64         int err = 0;
  65
  66         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
  67         if (!pool)
  68                 return ERR_PTR(-ENOMEM);
  69
  70         err = page_pool_init(pool, params);
  71         if (err < 0) {
  72                 pr_warn("%s() gave up with errno %d\n", __func__, err);
  73                 kfree(pool);
  74                 return ERR_PTR(err);
  75         }
  76
  77         return pool;
  78 }
  79 EXPORT_SYMBOL(page_pool_create);
  80
  81 /* fast path */
  82 static struct page *__page_pool_get_cached(struct page_pool *pool)
  83 {
  84         struct ptr_ring *r = &pool->ring;
  85         struct page *page;
  86
  87         /* Quicker fallback, avoid locks when ring is empty */
  88         if (__ptr_ring_empty(r))
  89                 return NULL;
  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                 /* Slower-path: Alloc array empty, time to refill
  99                  *
 100                  * Open-coded bulk ptr_ring consumer.
 101                  *
 102                  * Discussion: the ring consumer lock is not really
 103                  * needed due to the softirq/NAPI protection, but
 104                  * later need the ability to reclaim pages on the
 105                  * ring. Thus, keeping the locks.
 106                  */
 107                 spin_lock(&r->consumer_lock);
 108                 while ((page = __ptr_ring_consume(r))) {
 109                         if (pool->alloc.count == PP_ALLOC_CACHE_REFILL)
 110                                 break;
 111                         pool->alloc.cache[pool->alloc.count++] = page;
 112                 }
 113                 spin_unlock(&r->consumer_lock);
 114                 return page;
 115         }
 116
 117         /* Slow-path: Get page from locked ring queue */
 118         page = ptr_ring_consume(&pool->ring);
 119         return page;
 120 }
 121
 122 /* slow path */
 123 noinline
 124 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 125                                                  gfp_t _gfp)
 126 {
 127         struct page *page;
 128         gfp_t gfp = _gfp;
 129         dma_addr_t dma;
 130
 131         /* We could always set __GFP_COMP, and avoid this branch, as
 132          * prep_new_page() can handle order-0 with __GFP_COMP.
 133          */
 134         if (pool->p.order)
 135                 gfp |= __GFP_COMP;
 136
 137         /* FUTURE development:
 138          *
 139          * Current slow-path essentially falls back to single page
 140          * allocations, which doesn't improve performance.  This code
 141          * need bulk allocation support from the page allocator code.
 142          */
 143
 144         /* Cache was empty, do real allocation */
 145         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 146         if (!page)
 147                 return NULL;
 148
 149         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 150                 goto skip_dma_map;
 151
 152         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 153          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 154          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 155          * This mapping is kept for lifetime of page, until leaving pool.
 156          */
 157         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 158                                  (PAGE_SIZE << pool->p.order),
 159                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 160         if (dma_mapping_error(pool->p.dev, dma)) {
 161                 put_page(page);
 162                 return NULL;
 163         }
 164         page->dma_addr = dma;
 165
 166 skip_dma_map:
 167         /* Track how many pages are held 'in-flight' */
 168         pool->pages_state_hold_cnt++;
 169
 170         trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
 171
 172         /* When page just alloc'ed is should/must have refcnt 1. */
 173         return page;
 174 }
 175
 176 /* For using page_pool replace: alloc_pages() API calls, but provide
 177  * synchronization guarantee for allocation side.
 178  */
 179 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 180 {
 181         struct page *page;
 182
 183         /* Fast-path: Get a page from cache */
 184         page = __page_pool_get_cached(pool);
 185         if (page)
 186                 return page;
 187
 188         /* Slow-path: cache empty, do real allocation */
 189         page = __page_pool_alloc_pages_slow(pool, gfp);
 190         return page;
 191 }
 192 EXPORT_SYMBOL(page_pool_alloc_pages);
 193
 194 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
 195  *  https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
 196  */
 197 #define _distance(a, b) (s32)((a) - (b))
 198
 199 static s32 page_pool_inflight(struct page_pool *pool)
 200 {
 201         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 202         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 203         s32 distance;
 204
 205         distance = _distance(hold_cnt, release_cnt);
 206
 207         trace_page_pool_inflight(pool, distance, hold_cnt, release_cnt);
 208         return distance;
 209 }
 210
 211 static bool __page_pool_safe_to_destroy(struct page_pool *pool)
 212 {
 213         s32 inflight = page_pool_inflight(pool);
 214
 215         /* The distance should not be able to become negative */
 216         WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
 217
 218         return (inflight == 0);
 219 }
 220
 221 /* Cleanup page_pool state from page */
 222 static void __page_pool_clean_page(struct page_pool *pool,
 223                                    struct page *page)
 224 {
 225         dma_addr_t dma;
 226
 227         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 228                 goto skip_dma_unmap;
 229
 230         dma = page->dma_addr;
 231         /* DMA unmap */
 232         dma_unmap_page_attrs(pool->p.dev, dma,
 233                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 234                              DMA_ATTR_SKIP_CPU_SYNC);
 235         page->dma_addr = 0;
 236 skip_dma_unmap:
 237         atomic_inc(&pool->pages_state_release_cnt);
 238         trace_page_pool_state_release(pool, page,
 239                               atomic_read(&pool->pages_state_release_cnt));
 240 }
 241
 242 /* unmap the page and clean our state */
 243 void page_pool_unmap_page(struct page_pool *pool, struct page *page)
 244 {
 245         /* When page is unmapped, this implies page will not be
 246          * returned to page_pool.
 247          */
 248         __page_pool_clean_page(pool, page);
 249 }
 250 EXPORT_SYMBOL(page_pool_unmap_page);
 251
 252 /* Return a page to the page allocator, cleaning up our state */
 253 static void __page_pool_return_page(struct page_pool *pool, struct page *page)
 254 {
 255         __page_pool_clean_page(pool, page);
 256
 257         put_page(page);
 258         /* An optimization would be to call __free_pages(page, pool->p.order)
 259          * knowing page is not part of page-cache (thus avoiding a
 260          * __page_cache_release() call).
 261          */
 262 }
 263
 264 static bool __page_pool_recycle_into_ring(struct page_pool *pool,
 265                                    struct page *page)
 266 {
 267         int ret;
 268         /* BH protection not needed if current is serving softirq */
 269         if (in_serving_softirq())
 270                 ret = ptr_ring_produce(&pool->ring, page);
 271         else
 272                 ret = ptr_ring_produce_bh(&pool->ring, page);
 273
 274         return (ret == 0) ? true : false;
 275 }
 276
 277 /* Only allow direct recycling in special circumstances, into the
 278  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 279  *
 280  * Caller must provide appropriate safe context.
 281  */
 282 static bool __page_pool_recycle_direct(struct page *page,
 283                                        struct page_pool *pool)
 284 {
 285         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
 286                 return false;
 287
 288         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 289         pool->alloc.cache[pool->alloc.count++] = page;
 290         return true;
 291 }
 292
 293 void __page_pool_put_page(struct page_pool *pool,
 294                           struct page *page, bool allow_direct)
 295 {
 296         /* This allocator is optimized for the XDP mode that uses
 297          * one-frame-per-page, but have fallbacks that act like the
 298          * regular page allocator APIs.
 299          *
 300          * refcnt == 1 means page_pool owns page, and can recycle it.
 301          */
 302         if (likely(page_ref_count(page) == 1)) {
 303                 /* Read barrier done in page_ref_count / READ_ONCE */
 304
 305                 if (allow_direct && in_serving_softirq())
 306                         if (__page_pool_recycle_direct(page, pool))
 307                                 return;
 308
 309                 if (!__page_pool_recycle_into_ring(pool, page)) {
 310                         /* Cache full, fallback to free pages */
 311                         __page_pool_return_page(pool, page);
 312                 }
 313                 return;
 314         }
 315         /* Fallback/non-XDP mode: API user have elevated refcnt.
 316          *
 317          * Many drivers split up the page into fragments, and some
 318          * want to keep doing this to save memory and do refcnt based
 319          * recycling. Support this use case too, to ease drivers
 320          * switching between XDP/non-XDP.
 321          *
 322          * In-case page_pool maintains the DMA mapping, API user must
 323          * call page_pool_put_page once.  In this elevated refcnt
 324          * case, the DMA is unmapped/released, as driver is likely
 325          * doing refcnt based recycle tricks, meaning another process
 326          * will be invoking put_page.
 327          */
 328         __page_pool_clean_page(pool, page);
 329         put_page(page);
 330 }
 331 EXPORT_SYMBOL(__page_pool_put_page);
 332
 333 static void __page_pool_empty_ring(struct page_pool *pool)
 334 {
 335         struct page *page;
 336
 337         /* Empty recycle ring */
 338         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 339                 /* Verify the refcnt invariant of cached pages */
 340                 if (!(page_ref_count(page) == 1))
 341                         pr_crit("%s() page_pool refcnt %d violation\n",
 342                                 __func__, page_ref_count(page));
 343
 344                 __page_pool_return_page(pool, page);
 345         }
 346 }
 347
 348 static void __warn_in_flight(struct page_pool *pool)
 349 {
 350         u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
 351         u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
 352         s32 distance;
 353
 354         distance = _distance(hold_cnt, release_cnt);
 355
 356         /* Drivers should fix this, but only problematic when DMA is used */
 357         WARN(1, "Still in-flight pages:%d hold:%u released:%u",
 358              distance, hold_cnt, release_cnt);
 359 }
 360
 361 void __page_pool_free(struct page_pool *pool)
 362 {
 363         /* Only last user actually free/release resources */
 364         if (!page_pool_put(pool))
 365                 return;
 366
 367         WARN(pool->alloc.count, "API usage violation");
 368         WARN(!ptr_ring_empty(&pool->ring), "ptr_ring is not empty");
 369
 370         /* Can happen due to forced shutdown */
 371         if (!__page_pool_safe_to_destroy(pool))
 372                 __warn_in_flight(pool);
 373
 374         ptr_ring_cleanup(&pool->ring, NULL);
 375
 376         if (pool->p.flags & PP_FLAG_DMA_MAP)
 377                 put_device(pool->p.dev);
 378
 379         kfree(pool);
 380 }
 381 EXPORT_SYMBOL(__page_pool_free);
 382
 383 /* Request to shutdown: release pages cached by page_pool, and check
 384  * for in-flight pages
 385  */
 386 bool __page_pool_request_shutdown(struct page_pool *pool)
 387 {
 388         struct page *page;
 389
 390         /* Empty alloc cache, assume caller made sure this is
 391          * no-longer in use, and page_pool_alloc_pages() cannot be
 392          * call concurrently.
 393          */
 394         while (pool->alloc.count) {
 395                 page = pool->alloc.cache[--pool->alloc.count];
 396                 __page_pool_return_page(pool, page);
 397         }
 398
 399         /* No more consumers should exist, but producers could still
 400          * be in-flight.
 401          */
 402         __page_pool_empty_ring(pool);
 403
 404         return __page_pool_safe_to_destroy(pool);
 405 }
 406 EXPORT_SYMBOL(__page_pool_request_shutdown);