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 #include <linux/types.h>
   8 #include <linux/kernel.h>
   9 #include <linux/slab.h>
  10
  11 #include <net/page_pool.h>
  12 #include <linux/dma-direction.h>
  13 #include <linux/dma-mapping.h>
  14 #include <linux/page-flags.h>
  15 #include <linux/mm.h> /* for __put_page() */
  16
  17 static int page_pool_init(struct page_pool *pool,
  18                           const struct page_pool_params *params)
  19 {
  20         unsigned int ring_qsize = 1024; /* Default */
  21
  22         memcpy(&pool->p, params, sizeof(pool->p));
  23
  24         /* Validate only known flags were used */
  25         if (pool->p.flags & ~(PP_FLAG_ALL))
  26                 return -EINVAL;
  27
  28         if (pool->p.pool_size)
  29                 ring_qsize = pool->p.pool_size;
  30
  31         /* Sanity limit mem that can be pinned down */
  32         if (ring_qsize > 32768)
  33                 return -E2BIG;
  34
  35         /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
  36          * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
  37          * which is the XDP_TX use-case.
  38          */
  39         if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
  40             (pool->p.dma_dir != DMA_BIDIRECTIONAL))
  41                 return -EINVAL;
  42
  43         if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
  44                 return -ENOMEM;
  45
  46         return 0;
  47 }
  48
  49 struct page_pool *page_pool_create(const struct page_pool_params *params)
  50 {
  51         struct page_pool *pool;
  52         int err = 0;
  53
  54         pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
  55         if (!pool)
  56                 return ERR_PTR(-ENOMEM);
  57
  58         err = page_pool_init(pool, params);
  59         if (err < 0) {
  60                 pr_warn("%s() gave up with errno %d\n", __func__, err);
  61                 kfree(pool);
  62                 return ERR_PTR(err);
  63         }
  64         return pool;
  65 }
  66 EXPORT_SYMBOL(page_pool_create);
  67
  68 /* fast path */
  69 static struct page *__page_pool_get_cached(struct page_pool *pool)
  70 {
  71         struct ptr_ring *r = &pool->ring;
  72         struct page *page;
  73
  74         /* Quicker fallback, avoid locks when ring is empty */
  75         if (__ptr_ring_empty(r))
  76                 return NULL;
  77
  78         /* Test for safe-context, caller should provide this guarantee */
  79         if (likely(in_serving_softirq())) {
  80                 if (likely(pool->alloc.count)) {
  81                         /* Fast-path */
  82                         page = pool->alloc.cache[--pool->alloc.count];
  83                         return page;
  84                 }
  85                 /* Slower-path: Alloc array empty, time to refill
  86                  *
  87                  * Open-coded bulk ptr_ring consumer.
  88                  *
  89                  * Discussion: the ring consumer lock is not really
  90                  * needed due to the softirq/NAPI protection, but
  91                  * later need the ability to reclaim pages on the
  92                  * ring. Thus, keeping the locks.
  93                  */
  94                 spin_lock(&r->consumer_lock);
  95                 while ((page = __ptr_ring_consume(r))) {
  96                         if (pool->alloc.count == PP_ALLOC_CACHE_REFILL)
  97                                 break;
  98                         pool->alloc.cache[pool->alloc.count++] = page;
  99                 }
 100                 spin_unlock(&r->consumer_lock);
 101                 return page;
 102         }
 103
 104         /* Slow-path: Get page from locked ring queue */
 105         page = ptr_ring_consume(&pool->ring);
 106         return page;
 107 }
 108
 109 /* slow path */
 110 noinline
 111 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
 112                                                  gfp_t _gfp)
 113 {
 114         struct page *page;
 115         gfp_t gfp = _gfp;
 116         dma_addr_t dma;
 117
 118         /* We could always set __GFP_COMP, and avoid this branch, as
 119          * prep_new_page() can handle order-0 with __GFP_COMP.
 120          */
 121         if (pool->p.order)
 122                 gfp |= __GFP_COMP;
 123
 124         /* FUTURE development:
 125          *
 126          * Current slow-path essentially falls back to single page
 127          * allocations, which doesn't improve performance.  This code
 128          * need bulk allocation support from the page allocator code.
 129          */
 130
 131         /* Cache was empty, do real allocation */
 132         page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
 133         if (!page)
 134                 return NULL;
 135
 136         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 137                 goto skip_dma_map;
 138
 139         /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
 140          * since dma_addr_t can be either 32 or 64 bits and does not always fit
 141          * into page private data (i.e 32bit cpu with 64bit DMA caps)
 142          * This mapping is kept for lifetime of page, until leaving pool.
 143          */
 144         dma = dma_map_page_attrs(pool->p.dev, page, 0,
 145                                  (PAGE_SIZE << pool->p.order),
 146                                  pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
 147         if (dma_mapping_error(pool->p.dev, dma)) {
 148                 put_page(page);
 149                 return NULL;
 150         }
 151         page->dma_addr = dma;
 152
 153 skip_dma_map:
 154         /* When page just alloc'ed is should/must have refcnt 1. */
 155         return page;
 156 }
 157
 158 /* For using page_pool replace: alloc_pages() API calls, but provide
 159  * synchronization guarantee for allocation side.
 160  */
 161 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
 162 {
 163         struct page *page;
 164
 165         /* Fast-path: Get a page from cache */
 166         page = __page_pool_get_cached(pool);
 167         if (page)
 168                 return page;
 169
 170         /* Slow-path: cache empty, do real allocation */
 171         page = __page_pool_alloc_pages_slow(pool, gfp);
 172         return page;
 173 }
 174 EXPORT_SYMBOL(page_pool_alloc_pages);
 175
 176 /* Cleanup page_pool state from page */
 177 static void __page_pool_clean_page(struct page_pool *pool,
 178                                    struct page *page)
 179 {
 180         dma_addr_t dma;
 181
 182         if (!(pool->p.flags & PP_FLAG_DMA_MAP))
 183                 return;
 184
 185         dma = page->dma_addr;
 186         /* DMA unmap */
 187         dma_unmap_page_attrs(pool->p.dev, dma,
 188                              PAGE_SIZE << pool->p.order, pool->p.dma_dir,
 189                              DMA_ATTR_SKIP_CPU_SYNC);
 190         page->dma_addr = 0;
 191 }
 192
 193 /* Return a page to the page allocator, cleaning up our state */
 194 static void __page_pool_return_page(struct page_pool *pool, struct page *page)
 195 {
 196         __page_pool_clean_page(pool, page);
 197         put_page(page);
 198         /* An optimization would be to call __free_pages(page, pool->p.order)
 199          * knowing page is not part of page-cache (thus avoiding a
 200          * __page_cache_release() call).
 201          */
 202 }
 203
 204 static bool __page_pool_recycle_into_ring(struct page_pool *pool,
 205                                    struct page *page)
 206 {
 207         int ret;
 208         /* BH protection not needed if current is serving softirq */
 209         if (in_serving_softirq())
 210                 ret = ptr_ring_produce(&pool->ring, page);
 211         else
 212                 ret = ptr_ring_produce_bh(&pool->ring, page);
 213
 214         return (ret == 0) ? true : false;
 215 }
 216
 217 /* Only allow direct recycling in special circumstances, into the
 218  * alloc side cache.  E.g. during RX-NAPI processing for XDP_DROP use-case.
 219  *
 220  * Caller must provide appropriate safe context.
 221  */
 222 static bool __page_pool_recycle_direct(struct page *page,
 223                                        struct page_pool *pool)
 224 {
 225         if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
 226                 return false;
 227
 228         /* Caller MUST have verified/know (page_ref_count(page) == 1) */
 229         pool->alloc.cache[pool->alloc.count++] = page;
 230         return true;
 231 }
 232
 233 void __page_pool_put_page(struct page_pool *pool,
 234                           struct page *page, bool allow_direct)
 235 {
 236         /* This allocator is optimized for the XDP mode that uses
 237          * one-frame-per-page, but have fallbacks that act like the
 238          * regular page allocator APIs.
 239          *
 240          * refcnt == 1 means page_pool owns page, and can recycle it.
 241          */
 242         if (likely(page_ref_count(page) == 1)) {
 243                 /* Read barrier done in page_ref_count / READ_ONCE */
 244
 245                 if (allow_direct && in_serving_softirq())
 246                         if (__page_pool_recycle_direct(page, pool))
 247                                 return;
 248
 249                 if (!__page_pool_recycle_into_ring(pool, page)) {
 250                         /* Cache full, fallback to free pages */
 251                         __page_pool_return_page(pool, page);
 252                 }
 253                 return;
 254         }
 255         /* Fallback/non-XDP mode: API user have elevated refcnt.
 256          *
 257          * Many drivers split up the page into fragments, and some
 258          * want to keep doing this to save memory and do refcnt based
 259          * recycling. Support this use case too, to ease drivers
 260          * switching between XDP/non-XDP.
 261          *
 262          * In-case page_pool maintains the DMA mapping, API user must
 263          * call page_pool_put_page once.  In this elevated refcnt
 264          * case, the DMA is unmapped/released, as driver is likely
 265          * doing refcnt based recycle tricks, meaning another process
 266          * will be invoking put_page.
 267          */
 268         __page_pool_clean_page(pool, page);
 269         put_page(page);
 270 }
 271 EXPORT_SYMBOL(__page_pool_put_page);
 272
 273 static void __page_pool_empty_ring(struct page_pool *pool)
 274 {
 275         struct page *page;
 276
 277         /* Empty recycle ring */
 278         while ((page = ptr_ring_consume_bh(&pool->ring))) {
 279                 /* Verify the refcnt invariant of cached pages */
 280                 if (!(page_ref_count(page) == 1))
 281                         pr_crit("%s() page_pool refcnt %d violation\n",
 282                                 __func__, page_ref_count(page));
 283
 284                 __page_pool_return_page(pool, page);
 285         }
 286 }
 287
 288 static void __page_pool_destroy_rcu(struct rcu_head *rcu)
 289 {
 290         struct page_pool *pool;
 291
 292         pool = container_of(rcu, struct page_pool, rcu);
 293
 294         WARN(pool->alloc.count, "API usage violation");
 295
 296         __page_pool_empty_ring(pool);
 297         ptr_ring_cleanup(&pool->ring, NULL);
 298         kfree(pool);
 299 }
 300
 301 /* Cleanup and release resources */
 302 void page_pool_destroy(struct page_pool *pool)
 303 {
 304         struct page *page;
 305
 306         /* Empty alloc cache, assume caller made sure this is
 307          * no-longer in use, and page_pool_alloc_pages() cannot be
 308          * call concurrently.
 309          */
 310         while (pool->alloc.count) {
 311                 page = pool->alloc.cache[--pool->alloc.count];
 312                 __page_pool_return_page(pool, page);
 313         }
 314
 315         /* No more consumers should exist, but producers could still
 316          * be in-flight.
 317          */
 318         __page_pool_empty_ring(pool);
 319
 320         /* An xdp_mem_allocator can still ref page_pool pointer */
 321         call_rcu(&pool->rcu, __page_pool_destroy_rcu);
 322 }
 323 EXPORT_SYMBOL(page_pool_destroy);