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 #include <linux/android_kabi.h>
37
38 #define PP_FLAG_DMA_MAP BIT(0) /* Should page_pool do the DMA
39 * map/unmap
40 */
41 #define PP_FLAG_DMA_SYNC_DEV BIT(1) /* If set all pages that the driver gets
42 * from page_pool will be
43 * DMA-synced-for-device according to
44 * the length provided by the device
45 * driver.
46 * Please note DMA-sync-for-CPU is still
47 * device driver responsibility
48 */
49 #define PP_FLAG_PAGE_FRAG BIT(2) /* for page frag feature */
50 #define PP_FLAG_ALL (PP_FLAG_DMA_MAP |\
51 PP_FLAG_DMA_SYNC_DEV |\
52 PP_FLAG_PAGE_FRAG)
53
54 /*
55 * Fast allocation side cache array/stack
56 *
57 * The cache size and refill watermark is related to the network
58 * use-case. The NAPI budget is 64 packets. After a NAPI poll the RX
59 * ring is usually refilled and the max consumed elements will be 64,
60 * thus a natural max size of objects needed in the cache.
61 *
62 * Keeping room for more objects, is due to XDP_DROP use-case. As
63 * XDP_DROP allows the opportunity to recycle objects directly into
64 * this array, as it shares the same softirq/NAPI protection. If
65 * cache is already full (or partly full) then the XDP_DROP recycles
66 * would have to take a slower code path.
67 */
68 #define PP_ALLOC_CACHE_SIZE 128
69 #define PP_ALLOC_CACHE_REFILL 64
70 struct pp_alloc_cache {
71 u32 count;
72 struct page *cache[PP_ALLOC_CACHE_SIZE];
73 };
74
75 struct page_pool_params {
76 unsigned int flags;
77 unsigned int order;
78 unsigned int pool_size;
79 int nid; /* Numa node id to allocate from pages from */
80 struct device *dev; /* device, for DMA pre-mapping purposes */
81 enum dma_data_direction dma_dir; /* DMA mapping direction */
82 unsigned int max_len; /* max DMA sync memory size */
83 unsigned int offset; /* DMA addr offset */
84 };
85
86 struct page_pool {
87 struct page_pool_params p;
88
89 struct delayed_work release_dw;
90 void (*disconnect)(void *);
91 unsigned long defer_start;
92 unsigned long defer_warn;
93
94 u32 pages_state_hold_cnt;
95 unsigned int frag_offset;
96 struct page *frag_page;
97 long frag_users;
98
99 /*
100 * Data structure for allocation side
101 *
102 * Drivers allocation side usually already perform some kind
103 * of resource protection. Piggyback on this protection, and
104 * require driver to protect allocation side.
105 *
106 * For NIC drivers this means, allocate a page_pool per
107 * RX-queue. As the RX-queue is already protected by
108 * Softirq/BH scheduling and napi_schedule. NAPI schedule
109 * guarantee that a single napi_struct will only be scheduled
110 * on a single CPU (see napi_schedule).
111 */
112 struct pp_alloc_cache alloc ____cacheline_aligned_in_smp;
113
114 /* Data structure for storing recycled pages.
115 *
116 * Returning/freeing pages is more complicated synchronization
117 * wise, because free's can happen on remote CPUs, with no
118 * association with allocation resource.
119 *
120 * Use ptr_ring, as it separates consumer and producer
121 * effeciently, it a way that doesn't bounce cache-lines.
122 *
123 * TODO: Implement bulk return pages into this structure.
124 */
125 struct ptr_ring ring;
126
127 atomic_t pages_state_release_cnt;
128
129 /* A page_pool is strictly tied to a single RX-queue being
130 * protected by NAPI, due to above pp_alloc_cache. This
131 * refcnt serves purpose is to simplify drivers error handling.
132 */
133 refcount_t user_cnt;
134
135 u64 destroy_cnt;
136
137 ANDROID_KABI_RESERVE(1);
138 };
139
140 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp);
141
page_pool_dev_alloc_pages(struct page_pool * pool)142 static inline struct page *page_pool_dev_alloc_pages(struct page_pool *pool)
143 {
144 gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN);
145
146 return page_pool_alloc_pages(pool, gfp);
147 }
148
149 struct page *page_pool_alloc_frag(struct page_pool *pool, unsigned int *offset,
150 unsigned int size, gfp_t gfp);
151
page_pool_dev_alloc_frag(struct page_pool * pool,unsigned int * offset,unsigned int size)152 static inline struct page *page_pool_dev_alloc_frag(struct page_pool *pool,
153 unsigned int *offset,
154 unsigned int size)
155 {
156 gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN);
157
158 return page_pool_alloc_frag(pool, offset, size, gfp);
159 }
160
161 /* get the stored dma direction. A driver might decide to treat this locally and
162 * avoid the extra cache line from page_pool to determine the direction
163 */
164 static
page_pool_get_dma_dir(struct page_pool * pool)165 inline enum dma_data_direction page_pool_get_dma_dir(struct page_pool *pool)
166 {
167 return pool->p.dma_dir;
168 }
169
170 bool page_pool_return_skb_page(struct page *page);
171
172 struct page_pool *page_pool_create(const struct page_pool_params *params);
173
174 #ifdef CONFIG_PAGE_POOL
175 void page_pool_destroy(struct page_pool *pool);
176 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *));
177 void page_pool_release_page(struct page_pool *pool, struct page *page);
178 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
179 int count);
180 #else
page_pool_destroy(struct page_pool * pool)181 static inline void page_pool_destroy(struct page_pool *pool)
182 {
183 }
184
page_pool_use_xdp_mem(struct page_pool * pool,void (* disconnect)(void *))185 static inline void page_pool_use_xdp_mem(struct page_pool *pool,
186 void (*disconnect)(void *))
187 {
188 }
page_pool_release_page(struct page_pool * pool,struct page * page)189 static inline void page_pool_release_page(struct page_pool *pool,
190 struct page *page)
191 {
192 }
193
page_pool_put_page_bulk(struct page_pool * pool,void ** data,int count)194 static inline void page_pool_put_page_bulk(struct page_pool *pool, void **data,
195 int count)
196 {
197 }
198 #endif
199
200 void page_pool_put_page(struct page_pool *pool, struct page *page,
201 unsigned int dma_sync_size, bool allow_direct);
202
203 /* Same as above but will try to sync the entire area pool->max_len */
page_pool_put_full_page(struct page_pool * pool,struct page * page,bool allow_direct)204 static inline void page_pool_put_full_page(struct page_pool *pool,
205 struct page *page, bool allow_direct)
206 {
207 /* When page_pool isn't compiled-in, net/core/xdp.c doesn't
208 * allow registering MEM_TYPE_PAGE_POOL, but shield linker.
209 */
210 #ifdef CONFIG_PAGE_POOL
211 page_pool_put_page(pool, page, -1, allow_direct);
212 #endif
213 }
214
215 /* Same as above but the caller must guarantee safe context. e.g NAPI */
page_pool_recycle_direct(struct page_pool * pool,struct page * page)216 static inline void page_pool_recycle_direct(struct page_pool *pool,
217 struct page *page)
218 {
219 page_pool_put_full_page(pool, page, true);
220 }
221
222 #define PAGE_POOL_DMA_USE_PP_FRAG_COUNT \
223 (sizeof(dma_addr_t) > sizeof(unsigned long))
224
page_pool_get_dma_addr(struct page * page)225 static inline dma_addr_t page_pool_get_dma_addr(struct page *page)
226 {
227 dma_addr_t ret = page->dma_addr;
228
229 if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT)
230 ret |= (dma_addr_t)page->dma_addr_upper << 16 << 16;
231
232 return ret;
233 }
234
page_pool_set_dma_addr(struct page * page,dma_addr_t addr)235 static inline void page_pool_set_dma_addr(struct page *page, dma_addr_t addr)
236 {
237 page->dma_addr = addr;
238 if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT)
239 page->dma_addr_upper = upper_32_bits(addr);
240 }
241
page_pool_set_frag_count(struct page * page,long nr)242 static inline void page_pool_set_frag_count(struct page *page, long nr)
243 {
244 atomic_long_set(&page->pp_frag_count, nr);
245 }
246
page_pool_atomic_sub_frag_count_return(struct page * page,long nr)247 static inline long page_pool_atomic_sub_frag_count_return(struct page *page,
248 long nr)
249 {
250 long ret;
251
252 /* As suggested by Alexander, atomic_long_read() may cover up the
253 * reference count errors, so avoid calling atomic_long_read() in
254 * the cases of freeing or draining the page_frags, where we would
255 * not expect it to match or that are slowpath anyway.
256 */
257 if (__builtin_constant_p(nr) &&
258 atomic_long_read(&page->pp_frag_count) == nr)
259 return 0;
260
261 ret = atomic_long_sub_return(nr, &page->pp_frag_count);
262 WARN_ON(ret < 0);
263 return ret;
264 }
265
is_page_pool_compiled_in(void)266 static inline bool is_page_pool_compiled_in(void)
267 {
268 #ifdef CONFIG_PAGE_POOL
269 return true;
270 #else
271 return false;
272 #endif
273 }
274
page_pool_put(struct page_pool * pool)275 static inline bool page_pool_put(struct page_pool *pool)
276 {
277 return refcount_dec_and_test(&pool->user_cnt);
278 }
279
280 /* Caller must provide appropriate safe context, e.g. NAPI. */
281 void page_pool_update_nid(struct page_pool *pool, int new_nid);
page_pool_nid_changed(struct page_pool * pool,int new_nid)282 static inline void page_pool_nid_changed(struct page_pool *pool, int new_nid)
283 {
284 if (unlikely(pool->p.nid != new_nid))
285 page_pool_update_nid(pool, new_nid);
286 }
287
288 #endif /* _NET_PAGE_POOL_H */
289