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 #include <linux/types.h>
9 #include <linux/xarray.h>
10 #include <net/netmem.h>
11
12 #define PP_FLAG_DMA_MAP BIT(0) /* Should page_pool do the DMA
13 * map/unmap
14 */
15 #define PP_FLAG_DMA_SYNC_DEV BIT(1) /* If set all pages that the driver gets
16 * from page_pool will be
17 * DMA-synced-for-device according to
18 * the length provided by the device
19 * driver.
20 * Please note DMA-sync-for-CPU is still
21 * device driver responsibility
22 */
23 #define PP_FLAG_SYSTEM_POOL BIT(2) /* Global system page_pool */
24
25 /* Allow unreadable (net_iov backed) netmem in this page_pool. Drivers setting
26 * this must be able to support unreadable netmem, where netmem_address() would
27 * return NULL. This flag should not be set for header page_pools.
28 *
29 * If the driver sets PP_FLAG_ALLOW_UNREADABLE_NETMEM, it should also set
30 * page_pool_params.slow.queue_idx.
31 */
32 #define PP_FLAG_ALLOW_UNREADABLE_NETMEM BIT(3)
33
34 #define PP_FLAG_ALL (PP_FLAG_DMA_MAP | PP_FLAG_DMA_SYNC_DEV | \
35 PP_FLAG_SYSTEM_POOL | PP_FLAG_ALLOW_UNREADABLE_NETMEM)
36
37 /* Index limit to stay within PP_DMA_INDEX_BITS for DMA indices */
38 #define PP_DMA_INDEX_LIMIT XA_LIMIT(1, BIT(PP_DMA_INDEX_BITS) - 1)
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 netmem_ref cache[PP_ALLOC_CACHE_SIZE];
59 };
60
61 /**
62 * struct page_pool_params - page pool parameters
63 * @fast: params accessed frequently on hotpath
64 * @order: 2^order pages on allocation
65 * @pool_size: size of the ptr_ring
66 * @nid: NUMA node id to allocate from pages from
67 * @dev: device, for DMA pre-mapping purposes
68 * @napi: NAPI which is the sole consumer of pages, otherwise NULL
69 * @dma_dir: DMA mapping direction
70 * @max_len: max DMA sync memory size for PP_FLAG_DMA_SYNC_DEV
71 * @offset: DMA sync address offset for PP_FLAG_DMA_SYNC_DEV
72 * @slow: params with slowpath access only (initialization and Netlink)
73 * @netdev: netdev this pool will serve (leave as NULL if none or multiple)
74 * @queue_idx: queue idx this page_pool is being created for.
75 * @flags: PP_FLAG_DMA_MAP, PP_FLAG_DMA_SYNC_DEV, PP_FLAG_SYSTEM_POOL,
76 * PP_FLAG_ALLOW_UNREADABLE_NETMEM.
77 */
78 struct page_pool_params {
79 struct_group_tagged(page_pool_params_fast, fast,
80 unsigned int order;
81 unsigned int pool_size;
82 int nid;
83 struct device *dev;
84 struct napi_struct *napi;
85 enum dma_data_direction dma_dir;
86 unsigned int max_len;
87 unsigned int offset;
88 );
89 struct_group_tagged(page_pool_params_slow, slow,
90 struct net_device *netdev;
91 unsigned int queue_idx;
92 unsigned int flags;
93 /* private: used by test code only */
94 void (*init_callback)(netmem_ref netmem, void *arg);
95 void *init_arg;
96 );
97 };
98
99 #ifdef CONFIG_PAGE_POOL_STATS
100 /**
101 * struct page_pool_alloc_stats - allocation statistics
102 * @fast: successful fast path allocations
103 * @slow: slow path order-0 allocations
104 * @slow_high_order: slow path high order allocations
105 * @empty: ptr ring is empty, so a slow path allocation was forced
106 * @refill: an allocation which triggered a refill of the cache
107 * @waive: pages obtained from the ptr ring that cannot be added to
108 * the cache due to a NUMA mismatch
109 */
110 struct page_pool_alloc_stats {
111 u64 fast;
112 u64 slow;
113 u64 slow_high_order;
114 u64 empty;
115 u64 refill;
116 u64 waive;
117 };
118
119 /**
120 * struct page_pool_recycle_stats - recycling (freeing) statistics
121 * @cached: recycling placed page in the page pool cache
122 * @cache_full: page pool cache was full
123 * @ring: page placed into the ptr ring
124 * @ring_full: page released from page pool because the ptr ring was full
125 * @released_refcnt: page released (and not recycled) because refcnt > 1
126 */
127 struct page_pool_recycle_stats {
128 u64 cached;
129 u64 cache_full;
130 u64 ring;
131 u64 ring_full;
132 u64 released_refcnt;
133 };
134
135 /**
136 * struct page_pool_stats - combined page pool use statistics
137 * @alloc_stats: see struct page_pool_alloc_stats
138 * @recycle_stats: see struct page_pool_recycle_stats
139 *
140 * Wrapper struct for combining page pool stats with different storage
141 * requirements.
142 */
143 struct page_pool_stats {
144 struct page_pool_alloc_stats alloc_stats;
145 struct page_pool_recycle_stats recycle_stats;
146 };
147 #endif
148
149 /* The whole frag API block must stay within one cacheline. On 32-bit systems,
150 * sizeof(long) == sizeof(int), so that the block size is ``3 * sizeof(long)``.
151 * On 64-bit systems, the actual size is ``2 * sizeof(long) + sizeof(int)``.
152 * The closest pow-2 to both of them is ``4 * sizeof(long)``, so just use that
153 * one for simplicity.
154 * Having it aligned to a cacheline boundary may be excessive and doesn't bring
155 * any good.
156 */
157 #define PAGE_POOL_FRAG_GROUP_ALIGN (4 * sizeof(long))
158
159 struct memory_provider_ops;
160
161 struct pp_memory_provider_params {
162 void *mp_priv;
163 const struct memory_provider_ops *mp_ops;
164 };
165
166 struct page_pool {
167 struct page_pool_params_fast p;
168
169 int cpuid;
170 u32 pages_state_hold_cnt;
171
172 bool has_init_callback:1; /* slow::init_callback is set */
173 bool dma_map:1; /* Perform DMA mapping */
174 bool dma_sync:1; /* Perform DMA sync for device */
175 bool dma_sync_for_cpu:1; /* Perform DMA sync for cpu */
176 #ifdef CONFIG_PAGE_POOL_STATS
177 bool system:1; /* This is a global percpu pool */
178 #endif
179
180 __cacheline_group_begin_aligned(frag, PAGE_POOL_FRAG_GROUP_ALIGN);
181 long frag_users;
182 netmem_ref frag_page;
183 unsigned int frag_offset;
184 __cacheline_group_end_aligned(frag, PAGE_POOL_FRAG_GROUP_ALIGN);
185
186 struct delayed_work release_dw;
187 void (*disconnect)(void *pool);
188 unsigned long defer_start;
189 unsigned long defer_warn;
190
191 #ifdef CONFIG_PAGE_POOL_STATS
192 /* these stats are incremented while in softirq context */
193 struct page_pool_alloc_stats alloc_stats;
194 #endif
195 u32 xdp_mem_id;
196
197 /*
198 * Data structure for allocation side
199 *
200 * Drivers allocation side usually already perform some kind
201 * of resource protection. Piggyback on this protection, and
202 * require driver to protect allocation side.
203 *
204 * For NIC drivers this means, allocate a page_pool per
205 * RX-queue. As the RX-queue is already protected by
206 * Softirq/BH scheduling and napi_schedule. NAPI schedule
207 * guarantee that a single napi_struct will only be scheduled
208 * on a single CPU (see napi_schedule).
209 */
210 struct pp_alloc_cache alloc ____cacheline_aligned_in_smp;
211
212 /* Data structure for storing recycled pages.
213 *
214 * Returning/freeing pages is more complicated synchronization
215 * wise, because free's can happen on remote CPUs, with no
216 * association with allocation resource.
217 *
218 * Use ptr_ring, as it separates consumer and producer
219 * efficiently, it a way that doesn't bounce cache-lines.
220 *
221 * TODO: Implement bulk return pages into this structure.
222 */
223 struct ptr_ring ring;
224
225 void *mp_priv;
226 const struct memory_provider_ops *mp_ops;
227
228 struct xarray dma_mapped;
229
230 #ifdef CONFIG_PAGE_POOL_STATS
231 /* recycle stats are per-cpu to avoid locking */
232 struct page_pool_recycle_stats __percpu *recycle_stats;
233 #endif
234 atomic_t pages_state_release_cnt;
235
236 /* A page_pool is strictly tied to a single RX-queue being
237 * protected by NAPI, due to above pp_alloc_cache. This
238 * refcnt serves purpose is to simplify drivers error handling.
239 */
240 refcount_t user_cnt;
241
242 u64 destroy_cnt;
243
244 /* Slow/Control-path information follows */
245 struct page_pool_params_slow slow;
246 /* User-facing fields, protected by page_pools_lock */
247 struct {
248 struct hlist_node list;
249 u64 detach_time;
250 u32 id;
251 } user;
252 };
253
254 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp);
255 netmem_ref page_pool_alloc_netmems(struct page_pool *pool, gfp_t gfp);
256 struct page *page_pool_alloc_frag(struct page_pool *pool, unsigned int *offset,
257 unsigned int size, gfp_t gfp);
258 netmem_ref page_pool_alloc_frag_netmem(struct page_pool *pool,
259 unsigned int *offset, unsigned int size,
260 gfp_t gfp);
261 struct page_pool *page_pool_create(const struct page_pool_params *params);
262 struct page_pool *page_pool_create_percpu(const struct page_pool_params *params,
263 int cpuid);
264
265 struct xdp_mem_info;
266
267 #ifdef CONFIG_PAGE_POOL
268 void page_pool_disable_direct_recycling(struct page_pool *pool);
269 void page_pool_destroy(struct page_pool *pool);
270 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
271 const struct xdp_mem_info *mem);
272 void page_pool_put_netmem_bulk(netmem_ref *data, u32 count);
273 #else
page_pool_destroy(struct page_pool * pool)274 static inline void page_pool_destroy(struct page_pool *pool)
275 {
276 }
277
page_pool_use_xdp_mem(struct page_pool * pool,void (* disconnect)(void *),const struct xdp_mem_info * mem)278 static inline void page_pool_use_xdp_mem(struct page_pool *pool,
279 void (*disconnect)(void *),
280 const struct xdp_mem_info *mem)
281 {
282 }
283
page_pool_put_netmem_bulk(netmem_ref * data,u32 count)284 static inline void page_pool_put_netmem_bulk(netmem_ref *data, u32 count)
285 {
286 }
287 #endif
288
289 void page_pool_put_unrefed_netmem(struct page_pool *pool, netmem_ref netmem,
290 unsigned int dma_sync_size,
291 bool allow_direct);
292 void page_pool_put_unrefed_page(struct page_pool *pool, struct page *page,
293 unsigned int dma_sync_size,
294 bool allow_direct);
295
is_page_pool_compiled_in(void)296 static inline bool is_page_pool_compiled_in(void)
297 {
298 #ifdef CONFIG_PAGE_POOL
299 return true;
300 #else
301 return false;
302 #endif
303 }
304
305 /* Caller must provide appropriate safe context, e.g. NAPI. */
306 void page_pool_update_nid(struct page_pool *pool, int new_nid);
307
308 #endif /* _NET_PAGE_POOL_H */
309