• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <net/netmem.h>
10 #include <linux/android_kabi.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 /*
38  * Fast allocation side cache array/stack
39  *
40  * The cache size and refill watermark is related to the network
41  * use-case.  The NAPI budget is 64 packets.  After a NAPI poll the RX
42  * ring is usually refilled and the max consumed elements will be 64,
43  * thus a natural max size of objects needed in the cache.
44  *
45  * Keeping room for more objects, is due to XDP_DROP use-case.  As
46  * XDP_DROP allows the opportunity to recycle objects directly into
47  * this array, as it shares the same softirq/NAPI protection.  If
48  * cache is already full (or partly full) then the XDP_DROP recycles
49  * would have to take a slower code path.
50  */
51 #define PP_ALLOC_CACHE_SIZE	128
52 #define PP_ALLOC_CACHE_REFILL	64
53 struct pp_alloc_cache {
54 	u32 count;
55 	netmem_ref cache[PP_ALLOC_CACHE_SIZE];
56 };
57 
58 /**
59  * struct page_pool_params - page pool parameters
60  * @fast:	params accessed frequently on hotpath
61  * @order:	2^order pages on allocation
62  * @pool_size:	size of the ptr_ring
63  * @nid:	NUMA node id to allocate from pages from
64  * @dev:	device, for DMA pre-mapping purposes
65  * @napi:	NAPI which is the sole consumer of pages, otherwise NULL
66  * @dma_dir:	DMA mapping direction
67  * @max_len:	max DMA sync memory size for PP_FLAG_DMA_SYNC_DEV
68  * @offset:	DMA sync address offset for PP_FLAG_DMA_SYNC_DEV
69  * @slow:	params with slowpath access only (initialization and Netlink)
70  * @netdev:	netdev this pool will serve (leave as NULL if none or multiple)
71  * @queue_idx:	queue idx this page_pool is being created for.
72  * @flags:	PP_FLAG_DMA_MAP, PP_FLAG_DMA_SYNC_DEV, PP_FLAG_SYSTEM_POOL,
73  *		PP_FLAG_ALLOW_UNREADABLE_NETMEM.
74  */
75 struct page_pool_params {
76 	struct_group_tagged(page_pool_params_fast, fast,
77 		unsigned int	order;
78 		unsigned int	pool_size;
79 		int		nid;
80 		struct device	*dev;
81 		struct napi_struct *napi;
82 		enum dma_data_direction dma_dir;
83 		unsigned int	max_len;
84 		unsigned int	offset;
85 	);
86 	struct_group_tagged(page_pool_params_slow, slow,
87 		struct net_device *netdev;
88 		unsigned int queue_idx;
89 		unsigned int	flags;
90 /* private: used by test code only */
91 		void (*init_callback)(netmem_ref netmem, void *arg);
92 		void *init_arg;
93 	);
94 };
95 
96 #ifdef CONFIG_PAGE_POOL_STATS
97 /**
98  * struct page_pool_alloc_stats - allocation statistics
99  * @fast:	successful fast path allocations
100  * @slow:	slow path order-0 allocations
101  * @slow_high_order: slow path high order allocations
102  * @empty:	ptr ring is empty, so a slow path allocation was forced
103  * @refill:	an allocation which triggered a refill of the cache
104  * @waive:	pages obtained from the ptr ring that cannot be added to
105  *		the cache due to a NUMA mismatch
106  */
107 struct page_pool_alloc_stats {
108 	u64 fast;
109 	u64 slow;
110 	u64 slow_high_order;
111 	u64 empty;
112 	u64 refill;
113 	u64 waive;
114 };
115 
116 /**
117  * struct page_pool_recycle_stats - recycling (freeing) statistics
118  * @cached:	recycling placed page in the page pool cache
119  * @cache_full:	page pool cache was full
120  * @ring:	page placed into the ptr ring
121  * @ring_full:	page released from page pool because the ptr ring was full
122  * @released_refcnt:	page released (and not recycled) because refcnt > 1
123  */
124 struct page_pool_recycle_stats {
125 	u64 cached;
126 	u64 cache_full;
127 	u64 ring;
128 	u64 ring_full;
129 	u64 released_refcnt;
130 };
131 
132 /**
133  * struct page_pool_stats - combined page pool use statistics
134  * @alloc_stats:	see struct page_pool_alloc_stats
135  * @recycle_stats:	see struct page_pool_recycle_stats
136  *
137  * Wrapper struct for combining page pool stats with different storage
138  * requirements.
139  */
140 struct page_pool_stats {
141 	struct page_pool_alloc_stats alloc_stats;
142 	struct page_pool_recycle_stats recycle_stats;
143 };
144 #endif
145 
146 /* The whole frag API block must stay within one cacheline. On 32-bit systems,
147  * sizeof(long) == sizeof(int), so that the block size is ``3 * sizeof(long)``.
148  * On 64-bit systems, the actual size is ``2 * sizeof(long) + sizeof(int)``.
149  * The closest pow-2 to both of them is ``4 * sizeof(long)``, so just use that
150  * one for simplicity.
151  * Having it aligned to a cacheline boundary may be excessive and doesn't bring
152  * any good.
153  */
154 #define PAGE_POOL_FRAG_GROUP_ALIGN	(4 * sizeof(long))
155 
156 struct pp_memory_provider_params {
157 	void *mp_priv;
158 };
159 
160 struct page_pool {
161 	struct page_pool_params_fast p;
162 
163 	int cpuid;
164 	u32 pages_state_hold_cnt;
165 
166 	bool has_init_callback:1;	/* slow::init_callback is set */
167 	bool dma_map:1;			/* Perform DMA mapping */
168 	bool dma_sync:1;		/* Perform DMA sync */
169 #ifdef CONFIG_PAGE_POOL_STATS
170 	bool system:1;			/* This is a global percpu pool */
171 #endif
172 
173 	__cacheline_group_begin_aligned(frag, PAGE_POOL_FRAG_GROUP_ALIGN);
174 	long frag_users;
175 	netmem_ref frag_page;
176 	unsigned int frag_offset;
177 	__cacheline_group_end_aligned(frag, PAGE_POOL_FRAG_GROUP_ALIGN);
178 
179 	struct delayed_work release_dw;
180 	void (*disconnect)(void *pool);
181 	unsigned long defer_start;
182 	unsigned long defer_warn;
183 
184 #ifdef CONFIG_PAGE_POOL_STATS
185 	/* these stats are incremented while in softirq context */
186 	struct page_pool_alloc_stats alloc_stats;
187 #endif
188 	u32 xdp_mem_id;
189 
190 	/*
191 	 * Data structure for allocation side
192 	 *
193 	 * Drivers allocation side usually already perform some kind
194 	 * of resource protection.  Piggyback on this protection, and
195 	 * require driver to protect allocation side.
196 	 *
197 	 * For NIC drivers this means, allocate a page_pool per
198 	 * RX-queue. As the RX-queue is already protected by
199 	 * Softirq/BH scheduling and napi_schedule. NAPI schedule
200 	 * guarantee that a single napi_struct will only be scheduled
201 	 * on a single CPU (see napi_schedule).
202 	 */
203 	struct pp_alloc_cache alloc ____cacheline_aligned_in_smp;
204 
205 	/* Data structure for storing recycled pages.
206 	 *
207 	 * Returning/freeing pages is more complicated synchronization
208 	 * wise, because free's can happen on remote CPUs, with no
209 	 * association with allocation resource.
210 	 *
211 	 * Use ptr_ring, as it separates consumer and producer
212 	 * efficiently, it a way that doesn't bounce cache-lines.
213 	 *
214 	 * TODO: Implement bulk return pages into this structure.
215 	 */
216 	struct ptr_ring ring;
217 
218 	void *mp_priv;
219 
220 #ifdef CONFIG_PAGE_POOL_STATS
221 	/* recycle stats are per-cpu to avoid locking */
222 	struct page_pool_recycle_stats __percpu *recycle_stats;
223 #endif
224 	atomic_t pages_state_release_cnt;
225 
226 	/* A page_pool is strictly tied to a single RX-queue being
227 	 * protected by NAPI, due to above pp_alloc_cache. This
228 	 * refcnt serves purpose is to simplify drivers error handling.
229 	 */
230 	refcount_t user_cnt;
231 
232 	u64 destroy_cnt;
233 
234 	/* Slow/Control-path information follows */
235 	struct page_pool_params_slow slow;
236 	/* User-facing fields, protected by page_pools_lock */
237 	struct {
238 		struct hlist_node list;
239 		u64 detach_time;
240 		u32 napi_id;
241 		u32 id;
242 	} user;
243 	ANDROID_KABI_RESERVE(1);
244 };
245 
246 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp);
247 netmem_ref page_pool_alloc_netmem(struct page_pool *pool, gfp_t gfp);
248 struct page *page_pool_alloc_frag(struct page_pool *pool, unsigned int *offset,
249 				  unsigned int size, gfp_t gfp);
250 netmem_ref page_pool_alloc_frag_netmem(struct page_pool *pool,
251 				       unsigned int *offset, unsigned int size,
252 				       gfp_t gfp);
253 struct page_pool *page_pool_create(const struct page_pool_params *params);
254 struct page_pool *page_pool_create_percpu(const struct page_pool_params *params,
255 					  int cpuid);
256 
257 struct xdp_mem_info;
258 
259 #ifdef CONFIG_PAGE_POOL
260 void page_pool_disable_direct_recycling(struct page_pool *pool);
261 void page_pool_destroy(struct page_pool *pool);
262 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
263 			   const struct xdp_mem_info *mem);
264 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
265 			     int count);
266 #else
page_pool_destroy(struct page_pool * pool)267 static inline void page_pool_destroy(struct page_pool *pool)
268 {
269 }
270 
page_pool_use_xdp_mem(struct page_pool * pool,void (* disconnect)(void *),const struct xdp_mem_info * mem)271 static inline void page_pool_use_xdp_mem(struct page_pool *pool,
272 					 void (*disconnect)(void *),
273 					 const struct xdp_mem_info *mem)
274 {
275 }
276 
page_pool_put_page_bulk(struct page_pool * pool,void ** data,int count)277 static inline void page_pool_put_page_bulk(struct page_pool *pool, void **data,
278 					   int count)
279 {
280 }
281 #endif
282 
283 void page_pool_put_unrefed_netmem(struct page_pool *pool, netmem_ref netmem,
284 				  unsigned int dma_sync_size,
285 				  bool allow_direct);
286 void page_pool_put_unrefed_page(struct page_pool *pool, struct page *page,
287 				unsigned int dma_sync_size,
288 				bool allow_direct);
289 
is_page_pool_compiled_in(void)290 static inline bool is_page_pool_compiled_in(void)
291 {
292 #ifdef CONFIG_PAGE_POOL
293 	return true;
294 #else
295 	return false;
296 #endif
297 }
298 
299 /* Caller must provide appropriate safe context, e.g. NAPI. */
300 void page_pool_update_nid(struct page_pool *pool, int new_nid);
301 
302 #endif /* _NET_PAGE_POOL_H */
303