1 /* SPDX-License-Identifier: GPL-2.0 */
2
3 #ifndef _NET_PAGE_POOL_TYPES_H
4 #define _NET_PAGE_POOL_TYPES_H
5
6 #include <linux/dma-direction.h>
7 #include <linux/ptr_ring.h>
8
9 #define PP_FLAG_DMA_MAP BIT(0) /* Should page_pool do the DMA
10 * map/unmap
11 */
12 #define PP_FLAG_DMA_SYNC_DEV BIT(1) /* If set all pages that the driver gets
13 * from page_pool will be
14 * DMA-synced-for-device according to
15 * the length provided by the device
16 * driver.
17 * Please note DMA-sync-for-CPU is still
18 * device driver responsibility
19 */
20 #define PP_FLAG_PAGE_FRAG BIT(2) /* for page frag feature */
21 #define PP_FLAG_ALL (PP_FLAG_DMA_MAP |\
22 PP_FLAG_DMA_SYNC_DEV |\
23 PP_FLAG_PAGE_FRAG)
24
25 /*
26 * Fast allocation side cache array/stack
27 *
28 * The cache size and refill watermark is related to the network
29 * use-case. The NAPI budget is 64 packets. After a NAPI poll the RX
30 * ring is usually refilled and the max consumed elements will be 64,
31 * thus a natural max size of objects needed in the cache.
32 *
33 * Keeping room for more objects, is due to XDP_DROP use-case. As
34 * XDP_DROP allows the opportunity to recycle objects directly into
35 * this array, as it shares the same softirq/NAPI protection. If
36 * cache is already full (or partly full) then the XDP_DROP recycles
37 * would have to take a slower code path.
38 */
39 #define PP_ALLOC_CACHE_SIZE 128
40 #define PP_ALLOC_CACHE_REFILL 64
41 struct pp_alloc_cache {
42 u32 count;
43 struct page *cache[PP_ALLOC_CACHE_SIZE];
44 };
45
46 /**
47 * struct page_pool_params - page pool parameters
48 * @flags: PP_FLAG_DMA_MAP, PP_FLAG_DMA_SYNC_DEV, PP_FLAG_PAGE_FRAG
49 * @order: 2^order pages on allocation
50 * @pool_size: size of the ptr_ring
51 * @nid: NUMA node id to allocate from pages from
52 * @dev: device, for DMA pre-mapping purposes
53 * @napi: NAPI which is the sole consumer of pages, otherwise NULL
54 * @dma_dir: DMA mapping direction
55 * @max_len: max DMA sync memory size for PP_FLAG_DMA_SYNC_DEV
56 * @offset: DMA sync address offset for PP_FLAG_DMA_SYNC_DEV
57 */
58 struct page_pool_params {
59 unsigned int flags;
60 unsigned int order;
61 unsigned int pool_size;
62 int nid;
63 struct device *dev;
64 struct napi_struct *napi;
65 enum dma_data_direction dma_dir;
66 unsigned int max_len;
67 unsigned int offset;
68 /* private: used by test code only */
69 void (*init_callback)(struct page *page, void *arg);
70 void *init_arg;
71 };
72
73 #ifdef CONFIG_PAGE_POOL_STATS
74 /**
75 * struct page_pool_alloc_stats - allocation statistics
76 * @fast: successful fast path allocations
77 * @slow: slow path order-0 allocations
78 * @slow_high_order: slow path high order allocations
79 * @empty: ptr ring is empty, so a slow path allocation was forced
80 * @refill: an allocation which triggered a refill of the cache
81 * @waive: pages obtained from the ptr ring that cannot be added to
82 * the cache due to a NUMA mismatch
83 */
84 struct page_pool_alloc_stats {
85 u64 fast;
86 u64 slow;
87 u64 slow_high_order;
88 u64 empty;
89 u64 refill;
90 u64 waive;
91 };
92
93 /**
94 * struct page_pool_recycle_stats - recycling (freeing) statistics
95 * @cached: recycling placed page in the page pool cache
96 * @cache_full: page pool cache was full
97 * @ring: page placed into the ptr ring
98 * @ring_full: page released from page pool because the ptr ring was full
99 * @released_refcnt: page released (and not recycled) because refcnt > 1
100 */
101 struct page_pool_recycle_stats {
102 u64 cached;
103 u64 cache_full;
104 u64 ring;
105 u64 ring_full;
106 u64 released_refcnt;
107 };
108
109 /**
110 * struct page_pool_stats - combined page pool use statistics
111 * @alloc_stats: see struct page_pool_alloc_stats
112 * @recycle_stats: see struct page_pool_recycle_stats
113 *
114 * Wrapper struct for combining page pool stats with different storage
115 * requirements.
116 */
117 struct page_pool_stats {
118 struct page_pool_alloc_stats alloc_stats;
119 struct page_pool_recycle_stats recycle_stats;
120 };
121 #endif
122
123 struct page_pool {
124 struct page_pool_params p;
125
126 long frag_users;
127 struct page *frag_page;
128 unsigned int frag_offset;
129 u32 pages_state_hold_cnt;
130
131 struct delayed_work release_dw;
132 void (*disconnect)(void *pool);
133 unsigned long defer_start;
134 unsigned long defer_warn;
135
136 #ifdef CONFIG_PAGE_POOL_STATS
137 /* these stats are incremented while in softirq context */
138 struct page_pool_alloc_stats alloc_stats;
139 #endif
140 u32 xdp_mem_id;
141
142 /*
143 * Data structure for allocation side
144 *
145 * Drivers allocation side usually already perform some kind
146 * of resource protection. Piggyback on this protection, and
147 * require driver to protect allocation side.
148 *
149 * For NIC drivers this means, allocate a page_pool per
150 * RX-queue. As the RX-queue is already protected by
151 * Softirq/BH scheduling and napi_schedule. NAPI schedule
152 * guarantee that a single napi_struct will only be scheduled
153 * on a single CPU (see napi_schedule).
154 */
155 struct pp_alloc_cache alloc ____cacheline_aligned_in_smp;
156
157 /* Data structure for storing recycled pages.
158 *
159 * Returning/freeing pages is more complicated synchronization
160 * wise, because free's can happen on remote CPUs, with no
161 * association with allocation resource.
162 *
163 * Use ptr_ring, as it separates consumer and producer
164 * efficiently, it a way that doesn't bounce cache-lines.
165 *
166 * TODO: Implement bulk return pages into this structure.
167 */
168 struct ptr_ring ring;
169
170 #ifdef CONFIG_PAGE_POOL_STATS
171 /* recycle stats are per-cpu to avoid locking */
172 struct page_pool_recycle_stats __percpu *recycle_stats;
173 #endif
174 atomic_t pages_state_release_cnt;
175
176 /* A page_pool is strictly tied to a single RX-queue being
177 * protected by NAPI, due to above pp_alloc_cache. This
178 * refcnt serves purpose is to simplify drivers error handling.
179 */
180 refcount_t user_cnt;
181
182 u64 destroy_cnt;
183 };
184
185 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp);
186 struct page *page_pool_alloc_frag(struct page_pool *pool, unsigned int *offset,
187 unsigned int size, gfp_t gfp);
188 struct page_pool *page_pool_create(const struct page_pool_params *params);
189
190 struct xdp_mem_info;
191
192 #ifdef CONFIG_PAGE_POOL
193 void page_pool_unlink_napi(struct page_pool *pool);
194 void page_pool_destroy(struct page_pool *pool);
195 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
196 struct xdp_mem_info *mem);
197 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
198 int count);
199 #else
page_pool_unlink_napi(struct page_pool * pool)200 static inline void page_pool_unlink_napi(struct page_pool *pool)
201 {
202 }
203
page_pool_destroy(struct page_pool * pool)204 static inline void page_pool_destroy(struct page_pool *pool)
205 {
206 }
207
page_pool_use_xdp_mem(struct page_pool * pool,void (* disconnect)(void *),struct xdp_mem_info * mem)208 static inline void page_pool_use_xdp_mem(struct page_pool *pool,
209 void (*disconnect)(void *),
210 struct xdp_mem_info *mem)
211 {
212 }
213
page_pool_put_page_bulk(struct page_pool * pool,void ** data,int count)214 static inline void page_pool_put_page_bulk(struct page_pool *pool, void **data,
215 int count)
216 {
217 }
218 #endif
219
220 void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
221 unsigned int dma_sync_size,
222 bool allow_direct);
223
is_page_pool_compiled_in(void)224 static inline bool is_page_pool_compiled_in(void)
225 {
226 #ifdef CONFIG_PAGE_POOL
227 return true;
228 #else
229 return false;
230 #endif
231 }
232
233 /* Caller must provide appropriate safe context, e.g. NAPI. */
234 void page_pool_update_nid(struct page_pool *pool, int new_nid);
235
236 #endif /* _NET_PAGE_POOL_H */
237