1 /* SPDX-License-Identifier: GPL-2.0
2 *
3 * page_pool.h
4 * Author: Jesper Dangaard Brouer <netoptimizer@brouer.com>
5 * Copyright (C) 2016 Red Hat, Inc.
6 */
7
8 /**
9 * DOC: page_pool allocator
10 *
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.
14 *
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().
18 *
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.
24 *
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.
29 */
30 #ifndef _NET_PAGE_POOL_H
31 #define _NET_PAGE_POOL_H
32
33 #include <linux/mm.h> /* Needed by ptr_ring */
34 #include <linux/ptr_ring.h>
35 #include <linux/dma-direction.h>
36
37 #define PP_FLAG_DMA_MAP 1 /* Should page_pool do the DMA map/unmap */
38 #define PP_FLAG_ALL PP_FLAG_DMA_MAP
39
40 /*
41 * Fast allocation side cache array/stack
42 *
43 * The cache size and refill watermark is related to the network
44 * use-case. The NAPI budget is 64 packets. After a NAPI poll the RX
45 * ring is usually refilled and the max consumed elements will be 64,
46 * thus a natural max size of objects needed in the cache.
47 *
48 * Keeping room for more objects, is due to XDP_DROP use-case. As
49 * XDP_DROP allows the opportunity to recycle objects directly into
50 * this array, as it shares the same softirq/NAPI protection. If
51 * cache is already full (or partly full) then the XDP_DROP recycles
52 * would have to take a slower code path.
53 */
54 #define PP_ALLOC_CACHE_SIZE 128
55 #define PP_ALLOC_CACHE_REFILL 64
56 struct pp_alloc_cache {
57 u32 count;
58 void *cache[PP_ALLOC_CACHE_SIZE];
59 };
60
61 struct page_pool_params {
62 unsigned int flags;
63 unsigned int order;
64 unsigned int pool_size;
65 int nid; /* Numa node id to allocate from pages from */
66 struct device *dev; /* device, for DMA pre-mapping purposes */
67 enum dma_data_direction dma_dir; /* DMA mapping direction */
68 };
69
70 struct page_pool {
71 struct page_pool_params p;
72
73 struct delayed_work release_dw;
74 void (*disconnect)(void *);
75 unsigned long defer_start;
76 unsigned long defer_warn;
77
78 u32 pages_state_hold_cnt;
79
80 /*
81 * Data structure for allocation side
82 *
83 * Drivers allocation side usually already perform some kind
84 * of resource protection. Piggyback on this protection, and
85 * require driver to protect allocation side.
86 *
87 * For NIC drivers this means, allocate a page_pool per
88 * RX-queue. As the RX-queue is already protected by
89 * Softirq/BH scheduling and napi_schedule. NAPI schedule
90 * guarantee that a single napi_struct will only be scheduled
91 * on a single CPU (see napi_schedule).
92 */
93 struct pp_alloc_cache alloc ____cacheline_aligned_in_smp;
94
95 /* Data structure for storing recycled pages.
96 *
97 * Returning/freeing pages is more complicated synchronization
98 * wise, because free's can happen on remote CPUs, with no
99 * association with allocation resource.
100 *
101 * Use ptr_ring, as it separates consumer and producer
102 * effeciently, it a way that doesn't bounce cache-lines.
103 *
104 * TODO: Implement bulk return pages into this structure.
105 */
106 struct ptr_ring ring;
107
108 atomic_t pages_state_release_cnt;
109
110 /* A page_pool is strictly tied to a single RX-queue being
111 * protected by NAPI, due to above pp_alloc_cache. This
112 * refcnt serves purpose is to simplify drivers error handling.
113 */
114 refcount_t user_cnt;
115 };
116
117 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp);
118
119 static inline struct page *page_pool_dev_alloc_pages(struct page_pool *pool)
120 {
121 gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN);
122
123 return page_pool_alloc_pages(pool, gfp);
124 }
125
126 /* get the stored dma direction. A driver might decide to treat this locally and
127 * avoid the extra cache line from page_pool to determine the direction
128 */
129 static
130 inline enum dma_data_direction page_pool_get_dma_dir(struct page_pool *pool)
131 {
132 return pool->p.dma_dir;
133 }
134
135 struct page_pool *page_pool_create(const struct page_pool_params *params);
136
137 #ifdef CONFIG_PAGE_POOL
138 void page_pool_destroy(struct page_pool *pool);
139 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *));
140 #else
141 static inline void page_pool_destroy(struct page_pool *pool)
142 {
143 }
144
145 static inline void page_pool_use_xdp_mem(struct page_pool *pool,
146 void (*disconnect)(void *))
147 {
148 }
149 #endif
150
151 /* Never call this directly, use helpers below */
152 void __page_pool_put_page(struct page_pool *pool,
153 struct page *page, bool allow_direct);
154
155 static inline void page_pool_put_page(struct page_pool *pool,
156 struct page *page, bool allow_direct)
157 {
158 /* When page_pool isn't compiled-in, net/core/xdp.c doesn't
159 * allow registering MEM_TYPE_PAGE_POOL, but shield linker.
160 */
161 #ifdef CONFIG_PAGE_POOL
162 __page_pool_put_page(pool, page, allow_direct);
163 #endif
164 }
165 /* Very limited use-cases allow recycle direct */
166 static inline void page_pool_recycle_direct(struct page_pool *pool,
167 struct page *page)
168 {
169 __page_pool_put_page(pool, page, true);
170 }
171
172 /* Disconnects a page (from a page_pool). API users can have a need
173 * to disconnect a page (from a page_pool), to allow it to be used as
174 * a regular page (that will eventually be returned to the normal
175 * page-allocator via put_page).
176 */
177 void page_pool_unmap_page(struct page_pool *pool, struct page *page);
178 static inline void page_pool_release_page(struct page_pool *pool,
179 struct page *page)
180 {
181 #ifdef CONFIG_PAGE_POOL
182 page_pool_unmap_page(pool, page);
183 #endif
184 }
185
186 static inline dma_addr_t page_pool_get_dma_addr(struct page *page)
187 {
188 return page->dma_addr;
189 }
190
191 static inline bool is_page_pool_compiled_in(void)
192 {
193 #ifdef CONFIG_PAGE_POOL
194 return true;
195 #else
196 return false;
197 #endif
198 }
199
200 static inline bool page_pool_put(struct page_pool *pool)
201 {
202 return refcount_dec_and_test(&pool->user_cnt);
203 }
204
205 #endif /* _NET_PAGE_POOL_H */
206
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.