1 | /* SPDX-License-Identifier: GPL-2.0 |
2 | * |
3 | * page_pool/helpers.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 | * The page_pool allocator is optimized for recycling page or page fragment used |
12 | * by skb packet and xdp frame. |
13 | * |
14 | * Basic use involves replacing and alloc_pages() calls with page_pool_alloc(), |
15 | * which allocate memory with or without page splitting depending on the |
16 | * requested memory size. |
17 | * |
18 | * If the driver knows that it always requires full pages or its allocations are |
19 | * always smaller than half a page, it can use one of the more specific API |
20 | * calls: |
21 | * |
22 | * 1. page_pool_alloc_pages(): allocate memory without page splitting when |
23 | * driver knows that the memory it need is always bigger than half of the page |
24 | * allocated from page pool. There is no cache line dirtying for 'struct page' |
25 | * when a page is recycled back to the page pool. |
26 | * |
27 | * 2. page_pool_alloc_frag(): allocate memory with page splitting when driver |
28 | * knows that the memory it need is always smaller than or equal to half of the |
29 | * page allocated from page pool. Page splitting enables memory saving and thus |
30 | * avoids TLB/cache miss for data access, but there also is some cost to |
31 | * implement page splitting, mainly some cache line dirtying/bouncing for |
32 | * 'struct page' and atomic operation for page->pp_frag_count. |
33 | * |
34 | * The API keeps track of in-flight pages, in order to let API users know when |
35 | * it is safe to free a page_pool object, the API users must call |
36 | * page_pool_put_page() or page_pool_free_va() to free the page_pool object, or |
37 | * attach the page_pool object to a page_pool-aware object like skbs marked with |
38 | * skb_mark_for_recycle(). |
39 | * |
40 | * page_pool_put_page() may be called multi times on the same page if a page is |
41 | * split into multi fragments. For the last fragment, it will either recycle the |
42 | * page, or in case of page->_refcount > 1, it will release the DMA mapping and |
43 | * in-flight state accounting. |
44 | * |
45 | * dma_sync_single_range_for_device() is only called for the last fragment when |
46 | * page_pool is created with PP_FLAG_DMA_SYNC_DEV flag, so it depends on the |
47 | * last freed fragment to do the sync_for_device operation for all fragments in |
48 | * the same page when a page is split, the API user must setup pool->p.max_len |
49 | * and pool->p.offset correctly and ensure that page_pool_put_page() is called |
50 | * with dma_sync_size being -1 for fragment API. |
51 | */ |
52 | #ifndef _NET_PAGE_POOL_HELPERS_H |
53 | #define _NET_PAGE_POOL_HELPERS_H |
54 | |
55 | #include <net/page_pool/types.h> |
56 | |
57 | #ifdef CONFIG_PAGE_POOL_STATS |
58 | int page_pool_ethtool_stats_get_count(void); |
59 | u8 *page_pool_ethtool_stats_get_strings(u8 *data); |
60 | u64 *page_pool_ethtool_stats_get(u64 *data, void *stats); |
61 | |
62 | /* |
63 | * Drivers that wish to harvest page pool stats and report them to users |
64 | * (perhaps via ethtool, debugfs, or another mechanism) can allocate a |
65 | * struct page_pool_stats call page_pool_get_stats to get stats for the specified pool. |
66 | */ |
67 | bool page_pool_get_stats(struct page_pool *pool, |
68 | struct page_pool_stats *stats); |
69 | #else |
70 | static inline int page_pool_ethtool_stats_get_count(void) |
71 | { |
72 | return 0; |
73 | } |
74 | |
75 | static inline u8 *page_pool_ethtool_stats_get_strings(u8 *data) |
76 | { |
77 | return data; |
78 | } |
79 | |
80 | static inline u64 *page_pool_ethtool_stats_get(u64 *data, void *stats) |
81 | { |
82 | return data; |
83 | } |
84 | #endif |
85 | |
86 | /** |
87 | * page_pool_dev_alloc_pages() - allocate a page. |
88 | * @pool: pool from which to allocate |
89 | * |
90 | * Get a page from the page allocator or page_pool caches. |
91 | */ |
92 | static inline struct page *page_pool_dev_alloc_pages(struct page_pool *pool) |
93 | { |
94 | gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN); |
95 | |
96 | return page_pool_alloc_pages(pool, gfp); |
97 | } |
98 | |
99 | /** |
100 | * page_pool_dev_alloc_frag() - allocate a page fragment. |
101 | * @pool: pool from which to allocate |
102 | * @offset: offset to the allocated page |
103 | * @size: requested size |
104 | * |
105 | * Get a page fragment from the page allocator or page_pool caches. |
106 | * |
107 | * Return: |
108 | * Return allocated page fragment, otherwise return NULL. |
109 | */ |
110 | static inline struct page *page_pool_dev_alloc_frag(struct page_pool *pool, |
111 | unsigned int *offset, |
112 | unsigned int size) |
113 | { |
114 | gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN); |
115 | |
116 | return page_pool_alloc_frag(pool, offset, size, gfp); |
117 | } |
118 | |
119 | static inline struct page *page_pool_alloc(struct page_pool *pool, |
120 | unsigned int *offset, |
121 | unsigned int *size, gfp_t gfp) |
122 | { |
123 | unsigned int max_size = PAGE_SIZE << pool->p.order; |
124 | struct page *page; |
125 | |
126 | if ((*size << 1) > max_size) { |
127 | *size = max_size; |
128 | *offset = 0; |
129 | return page_pool_alloc_pages(pool, gfp); |
130 | } |
131 | |
132 | page = page_pool_alloc_frag(pool, offset, size: *size, gfp); |
133 | if (unlikely(!page)) |
134 | return NULL; |
135 | |
136 | /* There is very likely not enough space for another fragment, so append |
137 | * the remaining size to the current fragment to avoid truesize |
138 | * underestimate problem. |
139 | */ |
140 | if (pool->frag_offset + *size > max_size) { |
141 | *size = max_size - *offset; |
142 | pool->frag_offset = max_size; |
143 | } |
144 | |
145 | return page; |
146 | } |
147 | |
148 | /** |
149 | * page_pool_dev_alloc() - allocate a page or a page fragment. |
150 | * @pool: pool from which to allocate |
151 | * @offset: offset to the allocated page |
152 | * @size: in as the requested size, out as the allocated size |
153 | * |
154 | * Get a page or a page fragment from the page allocator or page_pool caches |
155 | * depending on the requested size in order to allocate memory with least memory |
156 | * utilization and performance penalty. |
157 | * |
158 | * Return: |
159 | * Return allocated page or page fragment, otherwise return NULL. |
160 | */ |
161 | static inline struct page *page_pool_dev_alloc(struct page_pool *pool, |
162 | unsigned int *offset, |
163 | unsigned int *size) |
164 | { |
165 | gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN); |
166 | |
167 | return page_pool_alloc(pool, offset, size, gfp); |
168 | } |
169 | |
170 | static inline void *page_pool_alloc_va(struct page_pool *pool, |
171 | unsigned int *size, gfp_t gfp) |
172 | { |
173 | unsigned int offset; |
174 | struct page *page; |
175 | |
176 | /* Mask off __GFP_HIGHMEM to ensure we can use page_address() */ |
177 | page = page_pool_alloc(pool, offset: &offset, size, gfp: gfp & ~__GFP_HIGHMEM); |
178 | if (unlikely(!page)) |
179 | return NULL; |
180 | |
181 | return page_address(page) + offset; |
182 | } |
183 | |
184 | /** |
185 | * page_pool_dev_alloc_va() - allocate a page or a page fragment and return its |
186 | * va. |
187 | * @pool: pool from which to allocate |
188 | * @size: in as the requested size, out as the allocated size |
189 | * |
190 | * This is just a thin wrapper around the page_pool_alloc() API, and |
191 | * it returns va of the allocated page or page fragment. |
192 | * |
193 | * Return: |
194 | * Return the va for the allocated page or page fragment, otherwise return NULL. |
195 | */ |
196 | static inline void *page_pool_dev_alloc_va(struct page_pool *pool, |
197 | unsigned int *size) |
198 | { |
199 | gfp_t gfp = (GFP_ATOMIC | __GFP_NOWARN); |
200 | |
201 | return page_pool_alloc_va(pool, size, gfp); |
202 | } |
203 | |
204 | /** |
205 | * page_pool_get_dma_dir() - Retrieve the stored DMA direction. |
206 | * @pool: pool from which page was allocated |
207 | * |
208 | * Get the stored dma direction. A driver might decide to store this locally |
209 | * and avoid the extra cache line from page_pool to determine the direction. |
210 | */ |
211 | static |
212 | inline enum dma_data_direction page_pool_get_dma_dir(struct page_pool *pool) |
213 | { |
214 | return pool->p.dma_dir; |
215 | } |
216 | |
217 | /* pp_frag_count represents the number of writers who can update the page |
218 | * either by updating skb->data or via DMA mappings for the device. |
219 | * We can't rely on the page refcnt for that as we don't know who might be |
220 | * holding page references and we can't reliably destroy or sync DMA mappings |
221 | * of the fragments. |
222 | * |
223 | * When pp_frag_count reaches 0 we can either recycle the page if the page |
224 | * refcnt is 1 or return it back to the memory allocator and destroy any |
225 | * mappings we have. |
226 | */ |
227 | static inline void page_pool_fragment_page(struct page *page, long nr) |
228 | { |
229 | atomic_long_set(v: &page->pp_frag_count, i: nr); |
230 | } |
231 | |
232 | static inline long page_pool_defrag_page(struct page *page, long nr) |
233 | { |
234 | long ret; |
235 | |
236 | /* If nr == pp_frag_count then we have cleared all remaining |
237 | * references to the page: |
238 | * 1. 'n == 1': no need to actually overwrite it. |
239 | * 2. 'n != 1': overwrite it with one, which is the rare case |
240 | * for pp_frag_count draining. |
241 | * |
242 | * The main advantage to doing this is that not only we avoid a atomic |
243 | * update, as an atomic_read is generally a much cheaper operation than |
244 | * an atomic update, especially when dealing with a page that may be |
245 | * partitioned into only 2 or 3 pieces; but also unify the pp_frag_count |
246 | * handling by ensuring all pages have partitioned into only 1 piece |
247 | * initially, and only overwrite it when the page is partitioned into |
248 | * more than one piece. |
249 | */ |
250 | if (atomic_long_read(v: &page->pp_frag_count) == nr) { |
251 | /* As we have ensured nr is always one for constant case using |
252 | * the BUILD_BUG_ON(), only need to handle the non-constant case |
253 | * here for pp_frag_count draining, which is a rare case. |
254 | */ |
255 | BUILD_BUG_ON(__builtin_constant_p(nr) && nr != 1); |
256 | if (!__builtin_constant_p(nr)) |
257 | atomic_long_set(v: &page->pp_frag_count, i: 1); |
258 | |
259 | return 0; |
260 | } |
261 | |
262 | ret = atomic_long_sub_return(i: nr, v: &page->pp_frag_count); |
263 | WARN_ON(ret < 0); |
264 | |
265 | /* We are the last user here too, reset pp_frag_count back to 1 to |
266 | * ensure all pages have been partitioned into 1 piece initially, |
267 | * this should be the rare case when the last two fragment users call |
268 | * page_pool_defrag_page() currently. |
269 | */ |
270 | if (unlikely(!ret)) |
271 | atomic_long_set(v: &page->pp_frag_count, i: 1); |
272 | |
273 | return ret; |
274 | } |
275 | |
276 | static inline bool page_pool_is_last_frag(struct page *page) |
277 | { |
278 | /* If page_pool_defrag_page() returns 0, we were the last user */ |
279 | return page_pool_defrag_page(page, nr: 1) == 0; |
280 | } |
281 | |
282 | /** |
283 | * page_pool_put_page() - release a reference to a page pool page |
284 | * @pool: pool from which page was allocated |
285 | * @page: page to release a reference on |
286 | * @dma_sync_size: how much of the page may have been touched by the device |
287 | * @allow_direct: released by the consumer, allow lockless caching |
288 | * |
289 | * The outcome of this depends on the page refcnt. If the driver bumps |
290 | * the refcnt > 1 this will unmap the page. If the page refcnt is 1 |
291 | * the allocator owns the page and will try to recycle it in one of the pool |
292 | * caches. If PP_FLAG_DMA_SYNC_DEV is set, the page will be synced for_device |
293 | * using dma_sync_single_range_for_device(). |
294 | */ |
295 | static inline void page_pool_put_page(struct page_pool *pool, |
296 | struct page *page, |
297 | unsigned int dma_sync_size, |
298 | bool allow_direct) |
299 | { |
300 | /* When page_pool isn't compiled-in, net/core/xdp.c doesn't |
301 | * allow registering MEM_TYPE_PAGE_POOL, but shield linker. |
302 | */ |
303 | #ifdef CONFIG_PAGE_POOL |
304 | if (!page_pool_is_last_frag(page)) |
305 | return; |
306 | |
307 | page_pool_put_defragged_page(pool, page, dma_sync_size, allow_direct); |
308 | #endif |
309 | } |
310 | |
311 | /** |
312 | * page_pool_put_full_page() - release a reference on a page pool page |
313 | * @pool: pool from which page was allocated |
314 | * @page: page to release a reference on |
315 | * @allow_direct: released by the consumer, allow lockless caching |
316 | * |
317 | * Similar to page_pool_put_page(), but will DMA sync the entire memory area |
318 | * as configured in &page_pool_params.max_len. |
319 | */ |
320 | static inline void page_pool_put_full_page(struct page_pool *pool, |
321 | struct page *page, bool allow_direct) |
322 | { |
323 | page_pool_put_page(pool, page, dma_sync_size: -1, allow_direct); |
324 | } |
325 | |
326 | /** |
327 | * page_pool_recycle_direct() - release a reference on a page pool page |
328 | * @pool: pool from which page was allocated |
329 | * @page: page to release a reference on |
330 | * |
331 | * Similar to page_pool_put_full_page() but caller must guarantee safe context |
332 | * (e.g NAPI), since it will recycle the page directly into the pool fast cache. |
333 | */ |
334 | static inline void page_pool_recycle_direct(struct page_pool *pool, |
335 | struct page *page) |
336 | { |
337 | page_pool_put_full_page(pool, page, allow_direct: true); |
338 | } |
339 | |
340 | #define PAGE_POOL_32BIT_ARCH_WITH_64BIT_DMA \ |
341 | (sizeof(dma_addr_t) > sizeof(unsigned long)) |
342 | |
343 | /** |
344 | * page_pool_free_va() - free a va into the page_pool |
345 | * @pool: pool from which va was allocated |
346 | * @va: va to be freed |
347 | * @allow_direct: freed by the consumer, allow lockless caching |
348 | * |
349 | * Free a va allocated from page_pool_allo_va(). |
350 | */ |
351 | static inline void page_pool_free_va(struct page_pool *pool, void *va, |
352 | bool allow_direct) |
353 | { |
354 | page_pool_put_page(pool, page: virt_to_head_page(x: va), dma_sync_size: -1, allow_direct); |
355 | } |
356 | |
357 | /** |
358 | * page_pool_get_dma_addr() - Retrieve the stored DMA address. |
359 | * @page: page allocated from a page pool |
360 | * |
361 | * Fetch the DMA address of the page. The page pool to which the page belongs |
362 | * must had been created with PP_FLAG_DMA_MAP. |
363 | */ |
364 | static inline dma_addr_t page_pool_get_dma_addr(struct page *page) |
365 | { |
366 | dma_addr_t ret = page->dma_addr; |
367 | |
368 | if (PAGE_POOL_32BIT_ARCH_WITH_64BIT_DMA) |
369 | ret <<= PAGE_SHIFT; |
370 | |
371 | return ret; |
372 | } |
373 | |
374 | static inline bool page_pool_set_dma_addr(struct page *page, dma_addr_t addr) |
375 | { |
376 | if (PAGE_POOL_32BIT_ARCH_WITH_64BIT_DMA) { |
377 | page->dma_addr = addr >> PAGE_SHIFT; |
378 | |
379 | /* We assume page alignment to shave off bottom bits, |
380 | * if this "compression" doesn't work we need to drop. |
381 | */ |
382 | return addr != (dma_addr_t)page->dma_addr << PAGE_SHIFT; |
383 | } |
384 | |
385 | page->dma_addr = addr; |
386 | return false; |
387 | } |
388 | |
389 | static inline bool page_pool_put(struct page_pool *pool) |
390 | { |
391 | return refcount_dec_and_test(r: &pool->user_cnt); |
392 | } |
393 | |
394 | static inline void page_pool_nid_changed(struct page_pool *pool, int new_nid) |
395 | { |
396 | if (unlikely(pool->p.nid != new_nid)) |
397 | page_pool_update_nid(pool, new_nid); |
398 | } |
399 | |
400 | #endif /* _NET_PAGE_POOL_HELPERS_H */ |
401 | |