1 /* SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 2016 Red Hat, Inc.
9 * DOC: page_pool allocator
11 * This page_pool allocator is optimized for the XDP mode that
12 * uses one-frame-per-page, but have fallbacks that act like the
13 * regular page allocator APIs.
15 * Basic use involve replacing alloc_pages() calls with the
16 * page_pool_alloc_pages() call. Drivers should likely use
17 * page_pool_dev_alloc_pages() replacing dev_alloc_pages().
19 * API keeps track of in-flight pages, in-order to let API user know
20 * when it is safe to dealloactor page_pool object. Thus, API users
21 * must make sure to call page_pool_release_page() when a page is
22 * "leaving" the page_pool. Or call page_pool_put_page() where
23 * appropiate. For maintaining correct accounting.
25 * API user must only call page_pool_put_page() once on a page, as it
26 * will either recycle the page, or in case of elevated refcnt, it
27 * will release the DMA mapping and in-flight state accounting. We
28 * hope to lift this requirement in the future.
30 #ifndef _NET_PAGE_POOL_H
31 #define _NET_PAGE_POOL_H
33 #include <linux/mm.h> /* Needed by ptr_ring */
34 #include <linux/ptr_ring.h>
35 #include <linux/dma-direction.h>
37 #define PP_FLAG_DMA_MAP BIT(0) /* Should page_pool do the DMA
40 #define PP_FLAG_DMA_SYNC_DEV BIT(1) /* If set all pages that the driver gets
41 * from page_pool will be
42 * DMA-synced-for-device according to
43 * the length provided by the device
45 * Please note DMA-sync-for-CPU is still
46 * device driver responsibility
48 #define PP_FLAG_PAGE_FRAG BIT(2) /* for page frag feature */
49 #define PP_FLAG_ALL (PP_FLAG_DMA_MAP |\
50 PP_FLAG_DMA_SYNC_DEV |\
54 * Fast allocation side cache array/stack
56 * The cache size and refill watermark is related to the network
57 * use-case. The NAPI budget is 64 packets. After a NAPI poll the RX
58 * ring is usually refilled and the max consumed elements will be 64,
59 * thus a natural max size of objects needed in the cache.
61 * Keeping room for more objects, is due to XDP_DROP use-case. As
62 * XDP_DROP allows the opportunity to recycle objects directly into
63 * this array, as it shares the same softirq/NAPI protection. If
64 * cache is already full (or partly full) then the XDP_DROP recycles
65 * would have to take a slower code path.
67 #define PP_ALLOC_CACHE_SIZE 128
68 #define PP_ALLOC_CACHE_REFILL 64
69 struct pp_alloc_cache {
71 struct page *cache[PP_ALLOC_CACHE_SIZE];
74 struct page_pool_params {
77 unsigned int pool_size;
78 int nid; /* Numa node id to allocate from pages from */
79 struct device *dev; /* device, for DMA pre-mapping purposes */
80 enum dma_data_direction dma_dir; /* DMA mapping direction */
81 unsigned int max_len; /* max DMA sync memory size */
82 unsigned int offset; /* DMA addr offset */
83 void (*init_callback)(struct page *page, void *arg);
88 struct page_pool_params p;
90 struct delayed_work release_dw;
91 void (*disconnect)(void *);
92 unsigned long defer_start;
93 unsigned long defer_warn;
95 u32 pages_state_hold_cnt;
96 unsigned int frag_offset;
97 struct page *frag_page;
102 * Data structure for allocation side
104 * Drivers allocation side usually already perform some kind
105 * of resource protection. Piggyback on this protection, and
106 * require driver to protect allocation side.
108 * For NIC drivers this means, allocate a page_pool per
109 * RX-queue. As the RX-queue is already protected by
110 * Softirq/BH scheduling and napi_schedule. NAPI schedule
111 * guarantee that a single napi_struct will only be scheduled
112 * on a single CPU (see napi_schedule).
114 struct pp_alloc_cache alloc ____cacheline_aligned_in_smp;
116 /* Data structure for storing recycled pages.
118 * Returning/freeing pages is more complicated synchronization
119 * wise, because free's can happen on remote CPUs, with no
120 * association with allocation resource.
122 * Use ptr_ring, as it separates consumer and producer
123 * effeciently, it a way that doesn't bounce cache-lines.
125 * TODO: Implement bulk return pages into this structure.
127 struct ptr_ring ring;
129 atomic_t pages_state_release_cnt;
131 /* A page_pool is strictly tied to a single RX-queue being
132 * protected by NAPI, due to above pp_alloc_cache. This
133 * refcnt serves purpose is to simplify drivers error handling.
140 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp);
142 static inline struct page *page_pool_dev_alloc_pages(struct page_pool *pool)
144 gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN);
146 return page_pool_alloc_pages(pool, gfp);
149 struct page *page_pool_alloc_frag(struct page_pool *pool, unsigned int *offset,
150 unsigned int size, gfp_t gfp);
152 static inline struct page *page_pool_dev_alloc_frag(struct page_pool *pool,
153 unsigned int *offset,
156 gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN);
158 return page_pool_alloc_frag(pool, offset, size, gfp);
161 /* get the stored dma direction. A driver might decide to treat this locally and
162 * avoid the extra cache line from page_pool to determine the direction
165 inline enum dma_data_direction page_pool_get_dma_dir(struct page_pool *pool)
167 return pool->p.dma_dir;
170 bool page_pool_return_skb_page(struct page *page);
172 struct page_pool *page_pool_create(const struct page_pool_params *params);
176 #ifdef CONFIG_PAGE_POOL
177 void page_pool_destroy(struct page_pool *pool);
178 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
179 struct xdp_mem_info *mem);
180 void page_pool_release_page(struct page_pool *pool, struct page *page);
181 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
184 static inline void page_pool_destroy(struct page_pool *pool)
188 static inline void page_pool_use_xdp_mem(struct page_pool *pool,
189 void (*disconnect)(void *),
190 struct xdp_mem_info *mem)
193 static inline void page_pool_release_page(struct page_pool *pool,
198 static inline void page_pool_put_page_bulk(struct page_pool *pool, void **data,
204 void page_pool_put_page(struct page_pool *pool, struct page *page,
205 unsigned int dma_sync_size, bool allow_direct);
207 /* Same as above but will try to sync the entire area pool->max_len */
208 static inline void page_pool_put_full_page(struct page_pool *pool,
209 struct page *page, bool allow_direct)
211 /* When page_pool isn't compiled-in, net/core/xdp.c doesn't
212 * allow registering MEM_TYPE_PAGE_POOL, but shield linker.
214 #ifdef CONFIG_PAGE_POOL
215 page_pool_put_page(pool, page, -1, allow_direct);
219 /* Same as above but the caller must guarantee safe context. e.g NAPI */
220 static inline void page_pool_recycle_direct(struct page_pool *pool,
223 page_pool_put_full_page(pool, page, true);
226 #define PAGE_POOL_DMA_USE_PP_FRAG_COUNT \
227 (sizeof(dma_addr_t) > sizeof(unsigned long))
229 static inline dma_addr_t page_pool_get_dma_addr(struct page *page)
231 dma_addr_t ret = page->dma_addr;
233 if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT)
234 ret |= (dma_addr_t)page->dma_addr_upper << 16 << 16;
239 static inline void page_pool_set_dma_addr(struct page *page, dma_addr_t addr)
241 page->dma_addr = addr;
242 if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT)
243 page->dma_addr_upper = upper_32_bits(addr);
246 static inline void page_pool_set_frag_count(struct page *page, long nr)
248 atomic_long_set(&page->pp_frag_count, nr);
251 static inline long page_pool_atomic_sub_frag_count_return(struct page *page,
256 /* As suggested by Alexander, atomic_long_read() may cover up the
257 * reference count errors, so avoid calling atomic_long_read() in
258 * the cases of freeing or draining the page_frags, where we would
259 * not expect it to match or that are slowpath anyway.
261 if (__builtin_constant_p(nr) &&
262 atomic_long_read(&page->pp_frag_count) == nr)
265 ret = atomic_long_sub_return(nr, &page->pp_frag_count);
270 static inline bool is_page_pool_compiled_in(void)
272 #ifdef CONFIG_PAGE_POOL
279 static inline bool page_pool_put(struct page_pool *pool)
281 return refcount_dec_and_test(&pool->user_cnt);
284 /* Caller must provide appropriate safe context, e.g. NAPI. */
285 void page_pool_update_nid(struct page_pool *pool, int new_nid);
286 static inline void page_pool_nid_changed(struct page_pool *pool, int new_nid)
288 if (unlikely(pool->p.nid != new_nid))
289 page_pool_update_nid(pool, new_nid);
292 static inline void page_pool_ring_lock(struct page_pool *pool)
293 __acquires(&pool->ring.producer_lock)
295 if (in_serving_softirq())
296 spin_lock(&pool->ring.producer_lock);
298 spin_lock_bh(&pool->ring.producer_lock);
301 static inline void page_pool_ring_unlock(struct page_pool *pool)
302 __releases(&pool->ring.producer_lock)
304 if (in_serving_softirq())
305 spin_unlock(&pool->ring.producer_lock);
307 spin_unlock_bh(&pool->ring.producer_lock);
310 #endif /* _NET_PAGE_POOL_H */