1 /* SPDX-License-Identifier: GPL-2.0
4 * Author: Jesper Dangaard Brouer <netoptimizer@brouer.com>
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_ALL (PP_FLAG_DMA_MAP | PP_FLAG_DMA_SYNC_DEV)
51 * Fast allocation side cache array/stack
53 * The cache size and refill watermark is related to the network
54 * use-case. The NAPI budget is 64 packets. After a NAPI poll the RX
55 * ring is usually refilled and the max consumed elements will be 64,
56 * thus a natural max size of objects needed in the cache.
58 * Keeping room for more objects, is due to XDP_DROP use-case. As
59 * XDP_DROP allows the opportunity to recycle objects directly into
60 * this array, as it shares the same softirq/NAPI protection. If
61 * cache is already full (or partly full) then the XDP_DROP recycles
62 * would have to take a slower code path.
64 #define PP_ALLOC_CACHE_SIZE 128
65 #define PP_ALLOC_CACHE_REFILL 64
66 struct pp_alloc_cache {
68 struct page *cache[PP_ALLOC_CACHE_SIZE];
71 struct page_pool_params {
74 unsigned int pool_size;
75 int nid; /* Numa node id to allocate from pages from */
76 struct device *dev; /* device, for DMA pre-mapping purposes */
77 enum dma_data_direction dma_dir; /* DMA mapping direction */
78 unsigned int max_len; /* max DMA sync memory size */
79 unsigned int offset; /* DMA addr offset */
83 struct page_pool_params p;
85 struct delayed_work release_dw;
86 void (*disconnect)(void *);
87 unsigned long defer_start;
88 unsigned long defer_warn;
90 u32 pages_state_hold_cnt;
93 * Data structure for allocation side
95 * Drivers allocation side usually already perform some kind
96 * of resource protection. Piggyback on this protection, and
97 * require driver to protect allocation side.
99 * For NIC drivers this means, allocate a page_pool per
100 * RX-queue. As the RX-queue is already protected by
101 * Softirq/BH scheduling and napi_schedule. NAPI schedule
102 * guarantee that a single napi_struct will only be scheduled
103 * on a single CPU (see napi_schedule).
105 struct pp_alloc_cache alloc ____cacheline_aligned_in_smp;
107 /* Data structure for storing recycled pages.
109 * Returning/freeing pages is more complicated synchronization
110 * wise, because free's can happen on remote CPUs, with no
111 * association with allocation resource.
113 * Use ptr_ring, as it separates consumer and producer
114 * effeciently, it a way that doesn't bounce cache-lines.
116 * TODO: Implement bulk return pages into this structure.
118 struct ptr_ring ring;
120 atomic_t pages_state_release_cnt;
122 /* A page_pool is strictly tied to a single RX-queue being
123 * protected by NAPI, due to above pp_alloc_cache. This
124 * refcnt serves purpose is to simplify drivers error handling.
131 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp);
133 static inline struct page *page_pool_dev_alloc_pages(struct page_pool *pool)
135 gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN);
137 return page_pool_alloc_pages(pool, gfp);
140 /* get the stored dma direction. A driver might decide to treat this locally and
141 * avoid the extra cache line from page_pool to determine the direction
144 inline enum dma_data_direction page_pool_get_dma_dir(struct page_pool *pool)
146 return pool->p.dma_dir;
149 bool page_pool_return_skb_page(struct page *page);
151 struct page_pool *page_pool_create(const struct page_pool_params *params);
153 #ifdef CONFIG_PAGE_POOL
154 void page_pool_destroy(struct page_pool *pool);
155 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *));
156 void page_pool_release_page(struct page_pool *pool, struct page *page);
157 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
160 static inline void page_pool_destroy(struct page_pool *pool)
164 static inline void page_pool_use_xdp_mem(struct page_pool *pool,
165 void (*disconnect)(void *))
168 static inline void page_pool_release_page(struct page_pool *pool,
173 static inline void page_pool_put_page_bulk(struct page_pool *pool, void **data,
179 void page_pool_put_page(struct page_pool *pool, struct page *page,
180 unsigned int dma_sync_size, bool allow_direct);
182 /* Same as above but will try to sync the entire area pool->max_len */
183 static inline void page_pool_put_full_page(struct page_pool *pool,
184 struct page *page, bool allow_direct)
186 /* When page_pool isn't compiled-in, net/core/xdp.c doesn't
187 * allow registering MEM_TYPE_PAGE_POOL, but shield linker.
189 #ifdef CONFIG_PAGE_POOL
190 page_pool_put_page(pool, page, -1, allow_direct);
194 /* Same as above but the caller must guarantee safe context. e.g NAPI */
195 static inline void page_pool_recycle_direct(struct page_pool *pool,
198 page_pool_put_full_page(pool, page, true);
201 static inline dma_addr_t page_pool_get_dma_addr(struct page *page)
203 dma_addr_t ret = page->dma_addr[0];
204 if (sizeof(dma_addr_t) > sizeof(unsigned long))
205 ret |= (dma_addr_t)page->dma_addr[1] << 16 << 16;
209 static inline void page_pool_set_dma_addr(struct page *page, dma_addr_t addr)
211 page->dma_addr[0] = addr;
212 if (sizeof(dma_addr_t) > sizeof(unsigned long))
213 page->dma_addr[1] = upper_32_bits(addr);
216 static inline bool is_page_pool_compiled_in(void)
218 #ifdef CONFIG_PAGE_POOL
225 static inline bool page_pool_put(struct page_pool *pool)
227 return refcount_dec_and_test(&pool->user_cnt);
230 /* Caller must provide appropriate safe context, e.g. NAPI. */
231 void page_pool_update_nid(struct page_pool *pool, int new_nid);
232 static inline void page_pool_nid_changed(struct page_pool *pool, int new_nid)
234 if (unlikely(pool->p.nid != new_nid))
235 page_pool_update_nid(pool, new_nid);
238 static inline void page_pool_ring_lock(struct page_pool *pool)
239 __acquires(&pool->ring.producer_lock)
241 if (in_serving_softirq())
242 spin_lock(&pool->ring.producer_lock);
244 spin_lock_bh(&pool->ring.producer_lock);
247 static inline void page_pool_ring_unlock(struct page_pool *pool)
248 __releases(&pool->ring.producer_lock)
250 if (in_serving_softirq())
251 spin_unlock(&pool->ring.producer_lock);
253 spin_unlock_bh(&pool->ring.producer_lock);
256 /* Store mem_info on struct page and use it while recycling skb frags */
258 void page_pool_store_mem_info(struct page *page, struct page_pool *pp)
263 #endif /* _NET_PAGE_POOL_H */