Commit | Line | Data |
---|---|---|
ff7d6b27 JDB |
1 | /* SPDX-License-Identifier: GPL-2.0 |
2 | * | |
3 | * page_pool.c | |
4 | * Author: Jesper Dangaard Brouer <netoptimizer@brouer.com> | |
5 | * Copyright (C) 2016 Red Hat, Inc. | |
6 | */ | |
32c28f7e | 7 | |
ff7d6b27 JDB |
8 | #include <linux/types.h> |
9 | #include <linux/kernel.h> | |
10 | #include <linux/slab.h> | |
f71fec47 | 11 | #include <linux/device.h> |
ff7d6b27 | 12 | |
a9ca9f9c | 13 | #include <net/page_pool/helpers.h> |
78862447 LB |
14 | #include <net/xdp.h> |
15 | ||
ff7d6b27 JDB |
16 | #include <linux/dma-direction.h> |
17 | #include <linux/dma-mapping.h> | |
18 | #include <linux/page-flags.h> | |
8d29c703 | 19 | #include <linux/mm.h> /* for put_page() */ |
c07aea3e | 20 | #include <linux/poison.h> |
f3c5264f | 21 | #include <linux/ethtool.h> |
8c48eea3 | 22 | #include <linux/netdevice.h> |
ff7d6b27 | 23 | |
32c28f7e JDB |
24 | #include <trace/events/page_pool.h> |
25 | ||
f17c6964 JK |
26 | #include "page_pool_priv.h" |
27 | ||
c3f812ce JL |
28 | #define DEFER_TIME (msecs_to_jiffies(1000)) |
29 | #define DEFER_WARN_INTERVAL (60 * HZ) | |
30 | ||
aaf153ae | 31 | #define BIAS_MAX (LONG_MAX >> 1) |
53e0961d | 32 | |
8610037e | 33 | #ifdef CONFIG_PAGE_POOL_STATS |
f853fa5c LB |
34 | static DEFINE_PER_CPU(struct page_pool_recycle_stats, pp_system_recycle_stats); |
35 | ||
8610037e JD |
36 | /* alloc_stat_inc is intended to be used in softirq context */ |
37 | #define alloc_stat_inc(pool, __stat) (pool->alloc_stats.__stat++) | |
ad6fa1e1 JD |
38 | /* recycle_stat_inc is safe to use when preemption is possible. */ |
39 | #define recycle_stat_inc(pool, __stat) \ | |
40 | do { \ | |
41 | struct page_pool_recycle_stats __percpu *s = pool->recycle_stats; \ | |
42 | this_cpu_inc(s->__stat); \ | |
43 | } while (0) | |
6b95e338 | 44 | |
590032a4 LB |
45 | #define recycle_stat_add(pool, __stat, val) \ |
46 | do { \ | |
47 | struct page_pool_recycle_stats __percpu *s = pool->recycle_stats; \ | |
48 | this_cpu_add(s->__stat, val); \ | |
49 | } while (0) | |
50 | ||
f3c5264f LB |
51 | static const char pp_stats[][ETH_GSTRING_LEN] = { |
52 | "rx_pp_alloc_fast", | |
53 | "rx_pp_alloc_slow", | |
54 | "rx_pp_alloc_slow_ho", | |
55 | "rx_pp_alloc_empty", | |
56 | "rx_pp_alloc_refill", | |
57 | "rx_pp_alloc_waive", | |
58 | "rx_pp_recycle_cached", | |
59 | "rx_pp_recycle_cache_full", | |
60 | "rx_pp_recycle_ring", | |
61 | "rx_pp_recycle_ring_full", | |
62 | "rx_pp_recycle_released_ref", | |
63 | }; | |
64 | ||
82e896d9 JK |
65 | /** |
66 | * page_pool_get_stats() - fetch page pool stats | |
67 | * @pool: pool from which page was allocated | |
68 | * @stats: struct page_pool_stats to fill in | |
69 | * | |
70 | * Retrieve statistics about the page_pool. This API is only available | |
71 | * if the kernel has been configured with ``CONFIG_PAGE_POOL_STATS=y``. | |
72 | * A pointer to a caller allocated struct page_pool_stats structure | |
73 | * is passed to this API which is filled in. The caller can then report | |
74 | * those stats to the user (perhaps via ethtool, debugfs, etc.). | |
75 | */ | |
d49010ad | 76 | bool page_pool_get_stats(const struct page_pool *pool, |
6b95e338 JD |
77 | struct page_pool_stats *stats) |
78 | { | |
79 | int cpu = 0; | |
80 | ||
81 | if (!stats) | |
82 | return false; | |
83 | ||
f3c5264f LB |
84 | /* The caller is responsible to initialize stats. */ |
85 | stats->alloc_stats.fast += pool->alloc_stats.fast; | |
86 | stats->alloc_stats.slow += pool->alloc_stats.slow; | |
87 | stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order; | |
88 | stats->alloc_stats.empty += pool->alloc_stats.empty; | |
89 | stats->alloc_stats.refill += pool->alloc_stats.refill; | |
90 | stats->alloc_stats.waive += pool->alloc_stats.waive; | |
6b95e338 JD |
91 | |
92 | for_each_possible_cpu(cpu) { | |
93 | const struct page_pool_recycle_stats *pcpu = | |
94 | per_cpu_ptr(pool->recycle_stats, cpu); | |
95 | ||
96 | stats->recycle_stats.cached += pcpu->cached; | |
97 | stats->recycle_stats.cache_full += pcpu->cache_full; | |
98 | stats->recycle_stats.ring += pcpu->ring; | |
99 | stats->recycle_stats.ring_full += pcpu->ring_full; | |
100 | stats->recycle_stats.released_refcnt += pcpu->released_refcnt; | |
101 | } | |
102 | ||
103 | return true; | |
104 | } | |
105 | EXPORT_SYMBOL(page_pool_get_stats); | |
f3c5264f LB |
106 | |
107 | u8 *page_pool_ethtool_stats_get_strings(u8 *data) | |
108 | { | |
109 | int i; | |
110 | ||
111 | for (i = 0; i < ARRAY_SIZE(pp_stats); i++) { | |
112 | memcpy(data, pp_stats[i], ETH_GSTRING_LEN); | |
113 | data += ETH_GSTRING_LEN; | |
114 | } | |
115 | ||
116 | return data; | |
117 | } | |
118 | EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings); | |
119 | ||
120 | int page_pool_ethtool_stats_get_count(void) | |
121 | { | |
122 | return ARRAY_SIZE(pp_stats); | |
123 | } | |
124 | EXPORT_SYMBOL(page_pool_ethtool_stats_get_count); | |
125 | ||
126 | u64 *page_pool_ethtool_stats_get(u64 *data, void *stats) | |
127 | { | |
128 | struct page_pool_stats *pool_stats = stats; | |
129 | ||
130 | *data++ = pool_stats->alloc_stats.fast; | |
131 | *data++ = pool_stats->alloc_stats.slow; | |
132 | *data++ = pool_stats->alloc_stats.slow_high_order; | |
133 | *data++ = pool_stats->alloc_stats.empty; | |
134 | *data++ = pool_stats->alloc_stats.refill; | |
135 | *data++ = pool_stats->alloc_stats.waive; | |
136 | *data++ = pool_stats->recycle_stats.cached; | |
137 | *data++ = pool_stats->recycle_stats.cache_full; | |
138 | *data++ = pool_stats->recycle_stats.ring; | |
139 | *data++ = pool_stats->recycle_stats.ring_full; | |
140 | *data++ = pool_stats->recycle_stats.released_refcnt; | |
141 | ||
142 | return data; | |
143 | } | |
144 | EXPORT_SYMBOL(page_pool_ethtool_stats_get); | |
145 | ||
8610037e JD |
146 | #else |
147 | #define alloc_stat_inc(pool, __stat) | |
ad6fa1e1 | 148 | #define recycle_stat_inc(pool, __stat) |
590032a4 | 149 | #define recycle_stat_add(pool, __stat, val) |
8610037e JD |
150 | #endif |
151 | ||
368d3cb4 YL |
152 | static bool page_pool_producer_lock(struct page_pool *pool) |
153 | __acquires(&pool->ring.producer_lock) | |
154 | { | |
155 | bool in_softirq = in_softirq(); | |
156 | ||
157 | if (in_softirq) | |
158 | spin_lock(&pool->ring.producer_lock); | |
159 | else | |
160 | spin_lock_bh(&pool->ring.producer_lock); | |
161 | ||
162 | return in_softirq; | |
163 | } | |
164 | ||
165 | static void page_pool_producer_unlock(struct page_pool *pool, | |
166 | bool in_softirq) | |
167 | __releases(&pool->ring.producer_lock) | |
168 | { | |
169 | if (in_softirq) | |
170 | spin_unlock(&pool->ring.producer_lock); | |
171 | else | |
172 | spin_unlock_bh(&pool->ring.producer_lock); | |
173 | } | |
174 | ||
1f20a576 AL |
175 | static void page_pool_struct_check(void) |
176 | { | |
177 | CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_users); | |
178 | CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_page); | |
179 | CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_offset); | |
180 | CACHELINE_ASSERT_GROUP_SIZE(struct page_pool, frag, 4 * sizeof(long)); | |
181 | } | |
182 | ||
ff7d6b27 | 183 | static int page_pool_init(struct page_pool *pool, |
2b0cfa6e LB |
184 | const struct page_pool_params *params, |
185 | int cpuid) | |
ff7d6b27 JDB |
186 | { |
187 | unsigned int ring_qsize = 1024; /* Default */ | |
188 | ||
1f20a576 AL |
189 | page_pool_struct_check(); |
190 | ||
5027ec19 JK |
191 | memcpy(&pool->p, ¶ms->fast, sizeof(pool->p)); |
192 | memcpy(&pool->slow, ¶ms->slow, sizeof(pool->slow)); | |
ff7d6b27 | 193 | |
2b0cfa6e LB |
194 | pool->cpuid = cpuid; |
195 | ||
ff7d6b27 | 196 | /* Validate only known flags were used */ |
403f11ac | 197 | if (pool->slow.flags & ~PP_FLAG_ALL) |
ff7d6b27 JDB |
198 | return -EINVAL; |
199 | ||
200 | if (pool->p.pool_size) | |
201 | ring_qsize = pool->p.pool_size; | |
202 | ||
203 | /* Sanity limit mem that can be pinned down */ | |
204 | if (ring_qsize > 32768) | |
205 | return -E2BIG; | |
206 | ||
207 | /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL. | |
208 | * DMA_BIDIRECTIONAL is for allowing page used for DMA sending, | |
209 | * which is the XDP_TX use-case. | |
210 | */ | |
403f11ac | 211 | if (pool->slow.flags & PP_FLAG_DMA_MAP) { |
798dda81 DK |
212 | if ((pool->p.dma_dir != DMA_FROM_DEVICE) && |
213 | (pool->p.dma_dir != DMA_BIDIRECTIONAL)) | |
214 | return -EINVAL; | |
403f11ac AL |
215 | |
216 | pool->dma_map = true; | |
798dda81 | 217 | } |
ff7d6b27 | 218 | |
403f11ac | 219 | if (pool->slow.flags & PP_FLAG_DMA_SYNC_DEV) { |
e68bc756 LB |
220 | /* In order to request DMA-sync-for-device the page |
221 | * needs to be mapped | |
222 | */ | |
403f11ac | 223 | if (!(pool->slow.flags & PP_FLAG_DMA_MAP)) |
e68bc756 LB |
224 | return -EINVAL; |
225 | ||
226 | if (!pool->p.max_len) | |
227 | return -EINVAL; | |
228 | ||
403f11ac AL |
229 | pool->dma_sync = true; |
230 | ||
e68bc756 LB |
231 | /* pool->p.offset has to be set according to the address |
232 | * offset used by the DMA engine to start copying rx data | |
233 | */ | |
234 | } | |
235 | ||
2da0cac1 JK |
236 | pool->has_init_callback = !!pool->slow.init_callback; |
237 | ||
ad6fa1e1 | 238 | #ifdef CONFIG_PAGE_POOL_STATS |
403f11ac | 239 | if (!(pool->slow.flags & PP_FLAG_SYSTEM_POOL)) { |
f853fa5c LB |
240 | pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats); |
241 | if (!pool->recycle_stats) | |
242 | return -ENOMEM; | |
243 | } else { | |
244 | /* For system page pool instance we use a singular stats object | |
245 | * instead of allocating a separate percpu variable for each | |
246 | * (also percpu) page pool instance. | |
247 | */ | |
248 | pool->recycle_stats = &pp_system_recycle_stats; | |
403f11ac | 249 | pool->system = true; |
f853fa5c | 250 | } |
ad6fa1e1 JD |
251 | #endif |
252 | ||
8ffbd166 JS |
253 | if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0) { |
254 | #ifdef CONFIG_PAGE_POOL_STATS | |
403f11ac | 255 | if (!pool->system) |
f853fa5c | 256 | free_percpu(pool->recycle_stats); |
8ffbd166 | 257 | #endif |
ff7d6b27 | 258 | return -ENOMEM; |
8ffbd166 | 259 | } |
ff7d6b27 | 260 | |
99c07c43 JDB |
261 | atomic_set(&pool->pages_state_release_cnt, 0); |
262 | ||
1da4bbef IK |
263 | /* Driver calling page_pool_create() also call page_pool_destroy() */ |
264 | refcount_set(&pool->user_cnt, 1); | |
265 | ||
403f11ac | 266 | if (pool->dma_map) |
f71fec47 JDB |
267 | get_device(pool->p.dev); |
268 | ||
ff7d6b27 JDB |
269 | return 0; |
270 | } | |
271 | ||
23cfaf67 JK |
272 | static void page_pool_uninit(struct page_pool *pool) |
273 | { | |
274 | ptr_ring_cleanup(&pool->ring, NULL); | |
275 | ||
403f11ac | 276 | if (pool->dma_map) |
23cfaf67 JK |
277 | put_device(pool->p.dev); |
278 | ||
279 | #ifdef CONFIG_PAGE_POOL_STATS | |
403f11ac | 280 | if (!pool->system) |
f853fa5c | 281 | free_percpu(pool->recycle_stats); |
23cfaf67 JK |
282 | #endif |
283 | } | |
284 | ||
82e896d9 | 285 | /** |
2b0cfa6e | 286 | * page_pool_create_percpu() - create a page pool for a given cpu. |
82e896d9 | 287 | * @params: parameters, see struct page_pool_params |
2b0cfa6e | 288 | * @cpuid: cpu identifier |
82e896d9 | 289 | */ |
2b0cfa6e LB |
290 | struct page_pool * |
291 | page_pool_create_percpu(const struct page_pool_params *params, int cpuid) | |
ff7d6b27 JDB |
292 | { |
293 | struct page_pool *pool; | |
873343e7 | 294 | int err; |
ff7d6b27 JDB |
295 | |
296 | pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid); | |
297 | if (!pool) | |
298 | return ERR_PTR(-ENOMEM); | |
299 | ||
2b0cfa6e | 300 | err = page_pool_init(pool, params, cpuid); |
f17c6964 JK |
301 | if (err < 0) |
302 | goto err_free; | |
303 | ||
304 | err = page_pool_list(pool); | |
305 | if (err) | |
306 | goto err_uninit; | |
1da4bbef | 307 | |
ff7d6b27 | 308 | return pool; |
f17c6964 JK |
309 | |
310 | err_uninit: | |
311 | page_pool_uninit(pool); | |
312 | err_free: | |
313 | pr_warn("%s() gave up with errno %d\n", __func__, err); | |
314 | kfree(pool); | |
315 | return ERR_PTR(err); | |
ff7d6b27 | 316 | } |
2b0cfa6e LB |
317 | EXPORT_SYMBOL(page_pool_create_percpu); |
318 | ||
319 | /** | |
320 | * page_pool_create() - create a page pool | |
321 | * @params: parameters, see struct page_pool_params | |
322 | */ | |
323 | struct page_pool *page_pool_create(const struct page_pool_params *params) | |
324 | { | |
325 | return page_pool_create_percpu(params, -1); | |
326 | } | |
ff7d6b27 JDB |
327 | EXPORT_SYMBOL(page_pool_create); |
328 | ||
458de8a9 | 329 | static void page_pool_return_page(struct page_pool *pool, struct page *page); |
44768dec JDB |
330 | |
331 | noinline | |
304db6cb | 332 | static struct page *page_pool_refill_alloc_cache(struct page_pool *pool) |
44768dec JDB |
333 | { |
334 | struct ptr_ring *r = &pool->ring; | |
335 | struct page *page; | |
336 | int pref_nid; /* preferred NUMA node */ | |
337 | ||
338 | /* Quicker fallback, avoid locks when ring is empty */ | |
8610037e JD |
339 | if (__ptr_ring_empty(r)) { |
340 | alloc_stat_inc(pool, empty); | |
44768dec | 341 | return NULL; |
8610037e | 342 | } |
44768dec JDB |
343 | |
344 | /* Softirq guarantee CPU and thus NUMA node is stable. This, | |
345 | * assumes CPU refilling driver RX-ring will also run RX-NAPI. | |
346 | */ | |
f13fc107 | 347 | #ifdef CONFIG_NUMA |
44768dec | 348 | pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid; |
f13fc107 JDB |
349 | #else |
350 | /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */ | |
351 | pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */ | |
352 | #endif | |
44768dec | 353 | |
44768dec JDB |
354 | /* Refill alloc array, but only if NUMA match */ |
355 | do { | |
356 | page = __ptr_ring_consume(r); | |
357 | if (unlikely(!page)) | |
358 | break; | |
359 | ||
360 | if (likely(page_to_nid(page) == pref_nid)) { | |
361 | pool->alloc.cache[pool->alloc.count++] = page; | |
362 | } else { | |
363 | /* NUMA mismatch; | |
364 | * (1) release 1 page to page-allocator and | |
365 | * (2) break out to fallthrough to alloc_pages_node. | |
366 | * This limit stress on page buddy alloactor. | |
367 | */ | |
458de8a9 | 368 | page_pool_return_page(pool, page); |
8610037e | 369 | alloc_stat_inc(pool, waive); |
44768dec JDB |
370 | page = NULL; |
371 | break; | |
372 | } | |
304db6cb | 373 | } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL); |
44768dec JDB |
374 | |
375 | /* Return last page */ | |
8610037e | 376 | if (likely(pool->alloc.count > 0)) { |
44768dec | 377 | page = pool->alloc.cache[--pool->alloc.count]; |
8610037e JD |
378 | alloc_stat_inc(pool, refill); |
379 | } | |
44768dec | 380 | |
44768dec JDB |
381 | return page; |
382 | } | |
383 | ||
ff7d6b27 JDB |
384 | /* fast path */ |
385 | static struct page *__page_pool_get_cached(struct page_pool *pool) | |
386 | { | |
ff7d6b27 JDB |
387 | struct page *page; |
388 | ||
304db6cb LR |
389 | /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */ |
390 | if (likely(pool->alloc.count)) { | |
391 | /* Fast-path */ | |
392 | page = pool->alloc.cache[--pool->alloc.count]; | |
8610037e | 393 | alloc_stat_inc(pool, fast); |
304db6cb LR |
394 | } else { |
395 | page = page_pool_refill_alloc_cache(pool); | |
ff7d6b27 JDB |
396 | } |
397 | ||
ff7d6b27 JDB |
398 | return page; |
399 | } | |
400 | ||
4321de44 AL |
401 | static void __page_pool_dma_sync_for_device(const struct page_pool *pool, |
402 | struct page *page, | |
403 | u32 dma_sync_size) | |
e68bc756 | 404 | { |
4321de44 | 405 | #if defined(CONFIG_HAS_DMA) && defined(CONFIG_DMA_NEED_SYNC) |
9ddb3c14 MWO |
406 | dma_addr_t dma_addr = page_pool_get_dma_addr(page); |
407 | ||
e68bc756 | 408 | dma_sync_size = min(dma_sync_size, pool->p.max_len); |
4321de44 AL |
409 | __dma_sync_single_for_device(pool->p.dev, dma_addr + pool->p.offset, |
410 | dma_sync_size, pool->p.dma_dir); | |
411 | #endif | |
412 | } | |
413 | ||
414 | static __always_inline void | |
415 | page_pool_dma_sync_for_device(const struct page_pool *pool, | |
416 | struct page *page, | |
417 | u32 dma_sync_size) | |
418 | { | |
419 | if (pool->dma_sync && dma_dev_need_sync(pool->p.dev)) | |
420 | __page_pool_dma_sync_for_device(pool, page, dma_sync_size); | |
e68bc756 LB |
421 | } |
422 | ||
dfa59717 JDB |
423 | static bool page_pool_dma_map(struct page_pool *pool, struct page *page) |
424 | { | |
425 | dma_addr_t dma; | |
426 | ||
427 | /* Setup DMA mapping: use 'struct page' area for storing DMA-addr | |
428 | * since dma_addr_t can be either 32 or 64 bits and does not always fit | |
429 | * into page private data (i.e 32bit cpu with 64bit DMA caps) | |
430 | * This mapping is kept for lifetime of page, until leaving pool. | |
431 | */ | |
432 | dma = dma_map_page_attrs(pool->p.dev, page, 0, | |
433 | (PAGE_SIZE << pool->p.order), | |
8e4c62c7 JK |
434 | pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC | |
435 | DMA_ATTR_WEAK_ORDERING); | |
dfa59717 JDB |
436 | if (dma_mapping_error(pool->p.dev, dma)) |
437 | return false; | |
438 | ||
90de47f0 YL |
439 | if (page_pool_set_dma_addr(page, dma)) |
440 | goto unmap_failed; | |
dfa59717 | 441 | |
4321de44 | 442 | page_pool_dma_sync_for_device(pool, page, pool->p.max_len); |
dfa59717 JDB |
443 | |
444 | return true; | |
90de47f0 YL |
445 | |
446 | unmap_failed: | |
447 | WARN_ON_ONCE("unexpected DMA address, please report to netdev@"); | |
448 | dma_unmap_page_attrs(pool->p.dev, dma, | |
449 | PAGE_SIZE << pool->p.order, pool->p.dma_dir, | |
450 | DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING); | |
451 | return false; | |
dfa59717 JDB |
452 | } |
453 | ||
57f05bc2 YL |
454 | static void page_pool_set_pp_info(struct page_pool *pool, |
455 | struct page *page) | |
456 | { | |
457 | page->pp = pool; | |
458 | page->pp_magic |= PP_SIGNATURE; | |
58d53d8f YL |
459 | |
460 | /* Ensuring all pages have been split into one fragment initially: | |
461 | * page_pool_set_pp_info() is only called once for every page when it | |
462 | * is allocated from the page allocator and page_pool_fragment_page() | |
463 | * is dirtying the same cache line as the page->pp_magic above, so | |
464 | * the overhead is negligible. | |
465 | */ | |
466 | page_pool_fragment_page(page, 1); | |
2da0cac1 | 467 | if (pool->has_init_callback) |
5027ec19 | 468 | pool->slow.init_callback(page, pool->slow.init_arg); |
57f05bc2 YL |
469 | } |
470 | ||
471 | static void page_pool_clear_pp_info(struct page *page) | |
472 | { | |
473 | page->pp_magic = 0; | |
474 | page->pp = NULL; | |
475 | } | |
476 | ||
be5dba25 JDB |
477 | static struct page *__page_pool_alloc_page_order(struct page_pool *pool, |
478 | gfp_t gfp) | |
ff7d6b27 JDB |
479 | { |
480 | struct page *page; | |
ff7d6b27 | 481 | |
be5dba25 | 482 | gfp |= __GFP_COMP; |
ff7d6b27 | 483 | page = alloc_pages_node(pool->p.nid, gfp, pool->p.order); |
be5dba25 | 484 | if (unlikely(!page)) |
ff7d6b27 JDB |
485 | return NULL; |
486 | ||
403f11ac | 487 | if (pool->dma_map && unlikely(!page_pool_dma_map(pool, page))) { |
ff7d6b27 JDB |
488 | put_page(page); |
489 | return NULL; | |
490 | } | |
ff7d6b27 | 491 | |
8610037e | 492 | alloc_stat_inc(pool, slow_high_order); |
57f05bc2 | 493 | page_pool_set_pp_info(pool, page); |
c07aea3e | 494 | |
99c07c43 JDB |
495 | /* Track how many pages are held 'in-flight' */ |
496 | pool->pages_state_hold_cnt++; | |
32c28f7e | 497 | trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt); |
be5dba25 JDB |
498 | return page; |
499 | } | |
500 | ||
501 | /* slow path */ | |
502 | noinline | |
503 | static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool, | |
504 | gfp_t gfp) | |
505 | { | |
506 | const int bulk = PP_ALLOC_CACHE_REFILL; | |
be5dba25 | 507 | unsigned int pp_order = pool->p.order; |
403f11ac | 508 | bool dma_map = pool->dma_map; |
be5dba25 JDB |
509 | struct page *page; |
510 | int i, nr_pages; | |
511 | ||
512 | /* Don't support bulk alloc for high-order pages */ | |
513 | if (unlikely(pp_order)) | |
514 | return __page_pool_alloc_page_order(pool, gfp); | |
515 | ||
516 | /* Unnecessary as alloc cache is empty, but guarantees zero count */ | |
517 | if (unlikely(pool->alloc.count > 0)) | |
518 | return pool->alloc.cache[--pool->alloc.count]; | |
519 | ||
520 | /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */ | |
521 | memset(&pool->alloc.cache, 0, sizeof(void *) * bulk); | |
522 | ||
d810d367 JW |
523 | nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk, |
524 | pool->alloc.cache); | |
be5dba25 JDB |
525 | if (unlikely(!nr_pages)) |
526 | return NULL; | |
527 | ||
528 | /* Pages have been filled into alloc.cache array, but count is zero and | |
529 | * page element have not been (possibly) DMA mapped. | |
530 | */ | |
531 | for (i = 0; i < nr_pages; i++) { | |
532 | page = pool->alloc.cache[i]; | |
403f11ac | 533 | if (dma_map && unlikely(!page_pool_dma_map(pool, page))) { |
be5dba25 JDB |
534 | put_page(page); |
535 | continue; | |
536 | } | |
57f05bc2 YL |
537 | |
538 | page_pool_set_pp_info(pool, page); | |
be5dba25 JDB |
539 | pool->alloc.cache[pool->alloc.count++] = page; |
540 | /* Track how many pages are held 'in-flight' */ | |
541 | pool->pages_state_hold_cnt++; | |
542 | trace_page_pool_state_hold(pool, page, | |
543 | pool->pages_state_hold_cnt); | |
544 | } | |
545 | ||
546 | /* Return last page */ | |
8610037e | 547 | if (likely(pool->alloc.count > 0)) { |
be5dba25 | 548 | page = pool->alloc.cache[--pool->alloc.count]; |
8610037e JD |
549 | alloc_stat_inc(pool, slow); |
550 | } else { | |
be5dba25 | 551 | page = NULL; |
8610037e | 552 | } |
32c28f7e | 553 | |
ff7d6b27 JDB |
554 | /* When page just alloc'ed is should/must have refcnt 1. */ |
555 | return page; | |
556 | } | |
557 | ||
558 | /* For using page_pool replace: alloc_pages() API calls, but provide | |
559 | * synchronization guarantee for allocation side. | |
560 | */ | |
561 | struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp) | |
562 | { | |
563 | struct page *page; | |
564 | ||
565 | /* Fast-path: Get a page from cache */ | |
566 | page = __page_pool_get_cached(pool); | |
567 | if (page) | |
568 | return page; | |
569 | ||
570 | /* Slow-path: cache empty, do real allocation */ | |
571 | page = __page_pool_alloc_pages_slow(pool, gfp); | |
572 | return page; | |
573 | } | |
574 | EXPORT_SYMBOL(page_pool_alloc_pages); | |
575 | ||
99c07c43 JDB |
576 | /* Calculate distance between two u32 values, valid if distance is below 2^(31) |
577 | * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution | |
578 | */ | |
579 | #define _distance(a, b) (s32)((a) - (b)) | |
580 | ||
7aee8429 | 581 | s32 page_pool_inflight(const struct page_pool *pool, bool strict) |
99c07c43 JDB |
582 | { |
583 | u32 release_cnt = atomic_read(&pool->pages_state_release_cnt); | |
584 | u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt); | |
c3f812ce | 585 | s32 inflight; |
99c07c43 | 586 | |
c3f812ce | 587 | inflight = _distance(hold_cnt, release_cnt); |
99c07c43 | 588 | |
7aee8429 JK |
589 | if (strict) { |
590 | trace_page_pool_release(pool, inflight, hold_cnt, release_cnt); | |
591 | WARN(inflight < 0, "Negative(%d) inflight packet-pages", | |
592 | inflight); | |
593 | } else { | |
594 | inflight = max(0, inflight); | |
595 | } | |
99c07c43 | 596 | |
c3f812ce | 597 | return inflight; |
99c07c43 JDB |
598 | } |
599 | ||
c3f687d8 JK |
600 | static __always_inline |
601 | void __page_pool_release_page_dma(struct page_pool *pool, struct page *page) | |
ff7d6b27 | 602 | { |
1567b85e IA |
603 | dma_addr_t dma; |
604 | ||
403f11ac | 605 | if (!pool->dma_map) |
458de8a9 IA |
606 | /* Always account for inflight pages, even if we didn't |
607 | * map them | |
608 | */ | |
c3f687d8 | 609 | return; |
ff7d6b27 | 610 | |
9ddb3c14 | 611 | dma = page_pool_get_dma_addr(page); |
458de8a9 | 612 | |
9ddb3c14 | 613 | /* When page is unmapped, it cannot be returned to our pool */ |
13f16d9d JDB |
614 | dma_unmap_page_attrs(pool->p.dev, dma, |
615 | PAGE_SIZE << pool->p.order, pool->p.dma_dir, | |
8e4c62c7 | 616 | DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING); |
9ddb3c14 | 617 | page_pool_set_dma_addr(page, 0); |
c3f687d8 JK |
618 | } |
619 | ||
620 | /* Disconnects a page (from a page_pool). API users can have a need | |
621 | * to disconnect a page (from a page_pool), to allow it to be used as | |
622 | * a regular page (that will eventually be returned to the normal | |
623 | * page-allocator via put_page). | |
624 | */ | |
625 | void page_pool_return_page(struct page_pool *pool, struct page *page) | |
626 | { | |
627 | int count; | |
628 | ||
629 | __page_pool_release_page_dma(pool, page); | |
630 | ||
57f05bc2 | 631 | page_pool_clear_pp_info(page); |
c07aea3e | 632 | |
c3f812ce JL |
633 | /* This may be the last page returned, releasing the pool, so |
634 | * it is not safe to reference pool afterwards. | |
635 | */ | |
7fb9b66d | 636 | count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt); |
c3f812ce | 637 | trace_page_pool_state_release(pool, page, count); |
99c07c43 | 638 | |
ff7d6b27 JDB |
639 | put_page(page); |
640 | /* An optimization would be to call __free_pages(page, pool->p.order) | |
641 | * knowing page is not part of page-cache (thus avoiding a | |
642 | * __page_cache_release() call). | |
643 | */ | |
644 | } | |
645 | ||
458de8a9 | 646 | static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page) |
ff7d6b27 JDB |
647 | { |
648 | int ret; | |
542bcea4 QD |
649 | /* BH protection not needed if current is softirq */ |
650 | if (in_softirq()) | |
ff7d6b27 JDB |
651 | ret = ptr_ring_produce(&pool->ring, page); |
652 | else | |
653 | ret = ptr_ring_produce_bh(&pool->ring, page); | |
654 | ||
ad6fa1e1 JD |
655 | if (!ret) { |
656 | recycle_stat_inc(pool, ring); | |
657 | return true; | |
658 | } | |
659 | ||
660 | return false; | |
ff7d6b27 JDB |
661 | } |
662 | ||
663 | /* Only allow direct recycling in special circumstances, into the | |
664 | * alloc side cache. E.g. during RX-NAPI processing for XDP_DROP use-case. | |
665 | * | |
666 | * Caller must provide appropriate safe context. | |
667 | */ | |
458de8a9 | 668 | static bool page_pool_recycle_in_cache(struct page *page, |
ff7d6b27 JDB |
669 | struct page_pool *pool) |
670 | { | |
ad6fa1e1 JD |
671 | if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) { |
672 | recycle_stat_inc(pool, cache_full); | |
ff7d6b27 | 673 | return false; |
ad6fa1e1 | 674 | } |
ff7d6b27 JDB |
675 | |
676 | /* Caller MUST have verified/know (page_ref_count(page) == 1) */ | |
677 | pool->alloc.cache[pool->alloc.count++] = page; | |
ad6fa1e1 | 678 | recycle_stat_inc(pool, cached); |
ff7d6b27 JDB |
679 | return true; |
680 | } | |
681 | ||
46f40172 MA |
682 | static bool __page_pool_page_can_be_recycled(const struct page *page) |
683 | { | |
684 | return page_ref_count(page) == 1 && !page_is_pfmemalloc(page); | |
685 | } | |
686 | ||
458de8a9 | 687 | /* If the page refcnt == 1, this will try to recycle the page. |
403f11ac | 688 | * If pool->dma_sync is set, we'll try to sync the DMA area for |
458de8a9 IA |
689 | * the configured size min(dma_sync_size, pool->max_len). |
690 | * If the page refcnt != 1, then the page will be returned to memory | |
691 | * subsystem. | |
692 | */ | |
78862447 LB |
693 | static __always_inline struct page * |
694 | __page_pool_put_page(struct page_pool *pool, struct page *page, | |
695 | unsigned int dma_sync_size, bool allow_direct) | |
ff7d6b27 | 696 | { |
ff4e538c JK |
697 | lockdep_assert_no_hardirq(); |
698 | ||
ff7d6b27 JDB |
699 | /* This allocator is optimized for the XDP mode that uses |
700 | * one-frame-per-page, but have fallbacks that act like the | |
701 | * regular page allocator APIs. | |
702 | * | |
703 | * refcnt == 1 means page_pool owns page, and can recycle it. | |
05656132 AL |
704 | * |
705 | * page is NOT reusable when allocated when system is under | |
706 | * some pressure. (page_is_pfmemalloc) | |
ff7d6b27 | 707 | */ |
46f40172 | 708 | if (likely(__page_pool_page_can_be_recycled(page))) { |
ff7d6b27 JDB |
709 | /* Read barrier done in page_ref_count / READ_ONCE */ |
710 | ||
4321de44 | 711 | page_pool_dma_sync_for_device(pool, page, dma_sync_size); |
e68bc756 | 712 | |
542bcea4 | 713 | if (allow_direct && in_softirq() && |
78862447 LB |
714 | page_pool_recycle_in_cache(page, pool)) |
715 | return NULL; | |
ff7d6b27 | 716 | |
78862447 LB |
717 | /* Page found as candidate for recycling */ |
718 | return page; | |
ff7d6b27 JDB |
719 | } |
720 | /* Fallback/non-XDP mode: API user have elevated refcnt. | |
721 | * | |
722 | * Many drivers split up the page into fragments, and some | |
723 | * want to keep doing this to save memory and do refcnt based | |
724 | * recycling. Support this use case too, to ease drivers | |
725 | * switching between XDP/non-XDP. | |
726 | * | |
727 | * In-case page_pool maintains the DMA mapping, API user must | |
728 | * call page_pool_put_page once. In this elevated refcnt | |
729 | * case, the DMA is unmapped/released, as driver is likely | |
730 | * doing refcnt based recycle tricks, meaning another process | |
731 | * will be invoking put_page. | |
732 | */ | |
ad6fa1e1 | 733 | recycle_stat_inc(pool, released_refcnt); |
07e0c7d3 | 734 | page_pool_return_page(pool, page); |
78862447 LB |
735 | |
736 | return NULL; | |
737 | } | |
738 | ||
0a149ab7 LC |
739 | void page_pool_put_unrefed_page(struct page_pool *pool, struct page *page, |
740 | unsigned int dma_sync_size, bool allow_direct) | |
78862447 LB |
741 | { |
742 | page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct); | |
743 | if (page && !page_pool_recycle_in_ring(pool, page)) { | |
744 | /* Cache full, fallback to free pages */ | |
ad6fa1e1 | 745 | recycle_stat_inc(pool, ring_full); |
78862447 LB |
746 | page_pool_return_page(pool, page); |
747 | } | |
ff7d6b27 | 748 | } |
0a149ab7 | 749 | EXPORT_SYMBOL(page_pool_put_unrefed_page); |
ff7d6b27 | 750 | |
82e896d9 JK |
751 | /** |
752 | * page_pool_put_page_bulk() - release references on multiple pages | |
753 | * @pool: pool from which pages were allocated | |
754 | * @data: array holding page pointers | |
755 | * @count: number of pages in @data | |
756 | * | |
757 | * Tries to refill a number of pages into the ptr_ring cache holding ptr_ring | |
758 | * producer lock. If the ptr_ring is full, page_pool_put_page_bulk() | |
759 | * will release leftover pages to the page allocator. | |
760 | * page_pool_put_page_bulk() is suitable to be run inside the driver NAPI tx | |
761 | * completion loop for the XDP_REDIRECT use case. | |
762 | * | |
763 | * Please note the caller must not use data area after running | |
764 | * page_pool_put_page_bulk(), as this function overwrites it. | |
765 | */ | |
78862447 LB |
766 | void page_pool_put_page_bulk(struct page_pool *pool, void **data, |
767 | int count) | |
768 | { | |
769 | int i, bulk_len = 0; | |
368d3cb4 | 770 | bool in_softirq; |
78862447 LB |
771 | |
772 | for (i = 0; i < count; i++) { | |
773 | struct page *page = virt_to_head_page(data[i]); | |
774 | ||
52cc6ffc | 775 | /* It is not the last user for the page frag case */ |
0a149ab7 | 776 | if (!page_pool_is_last_ref(page)) |
52cc6ffc AD |
777 | continue; |
778 | ||
78862447 LB |
779 | page = __page_pool_put_page(pool, page, -1, false); |
780 | /* Approved for bulk recycling in ptr_ring cache */ | |
781 | if (page) | |
782 | data[bulk_len++] = page; | |
783 | } | |
784 | ||
785 | if (unlikely(!bulk_len)) | |
786 | return; | |
787 | ||
788 | /* Bulk producer into ptr_ring page_pool cache */ | |
368d3cb4 | 789 | in_softirq = page_pool_producer_lock(pool); |
78862447 | 790 | for (i = 0; i < bulk_len; i++) { |
590032a4 LB |
791 | if (__ptr_ring_produce(&pool->ring, data[i])) { |
792 | /* ring full */ | |
793 | recycle_stat_inc(pool, ring_full); | |
794 | break; | |
795 | } | |
78862447 | 796 | } |
590032a4 | 797 | recycle_stat_add(pool, ring, i); |
368d3cb4 | 798 | page_pool_producer_unlock(pool, in_softirq); |
78862447 LB |
799 | |
800 | /* Hopefully all pages was return into ptr_ring */ | |
801 | if (likely(i == bulk_len)) | |
802 | return; | |
803 | ||
804 | /* ptr_ring cache full, free remaining pages outside producer lock | |
805 | * since put_page() with refcnt == 1 can be an expensive operation | |
806 | */ | |
807 | for (; i < bulk_len; i++) | |
808 | page_pool_return_page(pool, data[i]); | |
809 | } | |
810 | EXPORT_SYMBOL(page_pool_put_page_bulk); | |
811 | ||
53e0961d YL |
812 | static struct page *page_pool_drain_frag(struct page_pool *pool, |
813 | struct page *page) | |
814 | { | |
815 | long drain_count = BIAS_MAX - pool->frag_users; | |
816 | ||
817 | /* Some user is still using the page frag */ | |
0a149ab7 | 818 | if (likely(page_pool_unref_page(page, drain_count))) |
53e0961d YL |
819 | return NULL; |
820 | ||
46f40172 | 821 | if (__page_pool_page_can_be_recycled(page)) { |
4321de44 | 822 | page_pool_dma_sync_for_device(pool, page, -1); |
53e0961d YL |
823 | return page; |
824 | } | |
825 | ||
826 | page_pool_return_page(pool, page); | |
827 | return NULL; | |
828 | } | |
829 | ||
830 | static void page_pool_free_frag(struct page_pool *pool) | |
831 | { | |
832 | long drain_count = BIAS_MAX - pool->frag_users; | |
833 | struct page *page = pool->frag_page; | |
834 | ||
835 | pool->frag_page = NULL; | |
836 | ||
0a149ab7 | 837 | if (!page || page_pool_unref_page(page, drain_count)) |
53e0961d YL |
838 | return; |
839 | ||
840 | page_pool_return_page(pool, page); | |
841 | } | |
842 | ||
843 | struct page *page_pool_alloc_frag(struct page_pool *pool, | |
844 | unsigned int *offset, | |
845 | unsigned int size, gfp_t gfp) | |
846 | { | |
847 | unsigned int max_size = PAGE_SIZE << pool->p.order; | |
848 | struct page *page = pool->frag_page; | |
849 | ||
09d96ee5 | 850 | if (WARN_ON(size > max_size)) |
53e0961d YL |
851 | return NULL; |
852 | ||
853 | size = ALIGN(size, dma_get_cache_alignment()); | |
854 | *offset = pool->frag_offset; | |
855 | ||
856 | if (page && *offset + size > max_size) { | |
857 | page = page_pool_drain_frag(pool, page); | |
0f6deac3 JW |
858 | if (page) { |
859 | alloc_stat_inc(pool, fast); | |
53e0961d | 860 | goto frag_reset; |
0f6deac3 | 861 | } |
53e0961d YL |
862 | } |
863 | ||
864 | if (!page) { | |
865 | page = page_pool_alloc_pages(pool, gfp); | |
866 | if (unlikely(!page)) { | |
867 | pool->frag_page = NULL; | |
868 | return NULL; | |
869 | } | |
870 | ||
871 | pool->frag_page = page; | |
872 | ||
873 | frag_reset: | |
874 | pool->frag_users = 1; | |
875 | *offset = 0; | |
876 | pool->frag_offset = size; | |
52cc6ffc | 877 | page_pool_fragment_page(page, BIAS_MAX); |
53e0961d YL |
878 | return page; |
879 | } | |
880 | ||
881 | pool->frag_users++; | |
882 | pool->frag_offset = *offset + size; | |
0f6deac3 | 883 | alloc_stat_inc(pool, fast); |
53e0961d YL |
884 | return page; |
885 | } | |
886 | EXPORT_SYMBOL(page_pool_alloc_frag); | |
887 | ||
458de8a9 | 888 | static void page_pool_empty_ring(struct page_pool *pool) |
ff7d6b27 JDB |
889 | { |
890 | struct page *page; | |
891 | ||
892 | /* Empty recycle ring */ | |
4905bd9a | 893 | while ((page = ptr_ring_consume_bh(&pool->ring))) { |
ff7d6b27 JDB |
894 | /* Verify the refcnt invariant of cached pages */ |
895 | if (!(page_ref_count(page) == 1)) | |
896 | pr_crit("%s() page_pool refcnt %d violation\n", | |
897 | __func__, page_ref_count(page)); | |
898 | ||
458de8a9 | 899 | page_pool_return_page(pool, page); |
ff7d6b27 JDB |
900 | } |
901 | } | |
902 | ||
de97502e | 903 | static void __page_pool_destroy(struct page_pool *pool) |
d956a048 | 904 | { |
c3f812ce JL |
905 | if (pool->disconnect) |
906 | pool->disconnect(pool); | |
e54cfd7e | 907 | |
f17c6964 | 908 | page_pool_unlist(pool); |
23cfaf67 | 909 | page_pool_uninit(pool); |
e54cfd7e JDB |
910 | kfree(pool); |
911 | } | |
e54cfd7e | 912 | |
7c9e6942 | 913 | static void page_pool_empty_alloc_cache_once(struct page_pool *pool) |
ff7d6b27 JDB |
914 | { |
915 | struct page *page; | |
916 | ||
7c9e6942 JDB |
917 | if (pool->destroy_cnt) |
918 | return; | |
919 | ||
ff7d6b27 JDB |
920 | /* Empty alloc cache, assume caller made sure this is |
921 | * no-longer in use, and page_pool_alloc_pages() cannot be | |
922 | * call concurrently. | |
923 | */ | |
924 | while (pool->alloc.count) { | |
925 | page = pool->alloc.cache[--pool->alloc.count]; | |
458de8a9 | 926 | page_pool_return_page(pool, page); |
ff7d6b27 | 927 | } |
7c9e6942 JDB |
928 | } |
929 | ||
930 | static void page_pool_scrub(struct page_pool *pool) | |
931 | { | |
932 | page_pool_empty_alloc_cache_once(pool); | |
933 | pool->destroy_cnt++; | |
ff7d6b27 JDB |
934 | |
935 | /* No more consumers should exist, but producers could still | |
936 | * be in-flight. | |
937 | */ | |
458de8a9 | 938 | page_pool_empty_ring(pool); |
c3f812ce JL |
939 | } |
940 | ||
941 | static int page_pool_release(struct page_pool *pool) | |
942 | { | |
943 | int inflight; | |
944 | ||
945 | page_pool_scrub(pool); | |
7aee8429 | 946 | inflight = page_pool_inflight(pool, true); |
c3f812ce | 947 | if (!inflight) |
de97502e | 948 | __page_pool_destroy(pool); |
c3f812ce JL |
949 | |
950 | return inflight; | |
951 | } | |
952 | ||
953 | static void page_pool_release_retry(struct work_struct *wq) | |
954 | { | |
955 | struct delayed_work *dwq = to_delayed_work(wq); | |
956 | struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw); | |
be009667 | 957 | void *netdev; |
c3f812ce JL |
958 | int inflight; |
959 | ||
960 | inflight = page_pool_release(pool); | |
961 | if (!inflight) | |
962 | return; | |
963 | ||
be009667 JK |
964 | /* Periodic warning for page pools the user can't see */ |
965 | netdev = READ_ONCE(pool->slow.netdev); | |
966 | if (time_after_eq(jiffies, pool->defer_warn) && | |
967 | (!netdev || netdev == NET_PTR_POISON)) { | |
c3f812ce JL |
968 | int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ; |
969 | ||
be009667 JK |
970 | pr_warn("%s() stalled pool shutdown: id %u, %d inflight %d sec\n", |
971 | __func__, pool->user.id, inflight, sec); | |
c3f812ce JL |
972 | pool->defer_warn = jiffies + DEFER_WARN_INTERVAL; |
973 | } | |
974 | ||
975 | /* Still not ready to be disconnected, retry later */ | |
976 | schedule_delayed_work(&pool->release_dw, DEFER_TIME); | |
977 | } | |
978 | ||
64693ec7 THJ |
979 | void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *), |
980 | struct xdp_mem_info *mem) | |
c3f812ce JL |
981 | { |
982 | refcount_inc(&pool->user_cnt); | |
983 | pool->disconnect = disconnect; | |
64693ec7 | 984 | pool->xdp_mem_id = mem->id; |
c3f812ce JL |
985 | } |
986 | ||
56ef27e3 | 987 | static void page_pool_disable_direct_recycling(struct page_pool *pool) |
dd64b232 | 988 | { |
56ef27e3 AL |
989 | /* Disable direct recycling based on pool->cpuid. |
990 | * Paired with READ_ONCE() in napi_pp_put_page(). | |
991 | */ | |
992 | WRITE_ONCE(pool->cpuid, -1); | |
993 | ||
dd64b232 JK |
994 | if (!pool->p.napi) |
995 | return; | |
996 | ||
997 | /* To avoid races with recycling and additional barriers make sure | |
998 | * pool and NAPI are unlinked when NAPI is disabled. | |
999 | */ | |
1000 | WARN_ON(!test_bit(NAPI_STATE_SCHED, &pool->p.napi->state) || | |
1001 | READ_ONCE(pool->p.napi->list_owner) != -1); | |
1002 | ||
1003 | WRITE_ONCE(pool->p.napi, NULL); | |
1004 | } | |
dd64b232 | 1005 | |
c3f812ce JL |
1006 | void page_pool_destroy(struct page_pool *pool) |
1007 | { | |
1008 | if (!pool) | |
1009 | return; | |
1010 | ||
1011 | if (!page_pool_put(pool)) | |
1012 | return; | |
1013 | ||
56ef27e3 | 1014 | page_pool_disable_direct_recycling(pool); |
53e0961d YL |
1015 | page_pool_free_frag(pool); |
1016 | ||
c3f812ce JL |
1017 | if (!page_pool_release(pool)) |
1018 | return; | |
1019 | ||
69cb4952 | 1020 | page_pool_detached(pool); |
c3f812ce JL |
1021 | pool->defer_start = jiffies; |
1022 | pool->defer_warn = jiffies + DEFER_WARN_INTERVAL; | |
ff7d6b27 | 1023 | |
c3f812ce JL |
1024 | INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry); |
1025 | schedule_delayed_work(&pool->release_dw, DEFER_TIME); | |
ff7d6b27 | 1026 | } |
c3f812ce | 1027 | EXPORT_SYMBOL(page_pool_destroy); |
bc836748 SM |
1028 | |
1029 | /* Caller must provide appropriate safe context, e.g. NAPI. */ | |
1030 | void page_pool_update_nid(struct page_pool *pool, int new_nid) | |
1031 | { | |
44768dec JDB |
1032 | struct page *page; |
1033 | ||
bc836748 SM |
1034 | trace_page_pool_update_nid(pool, new_nid); |
1035 | pool->p.nid = new_nid; | |
44768dec JDB |
1036 | |
1037 | /* Flush pool alloc cache, as refill will check NUMA node */ | |
1038 | while (pool->alloc.count) { | |
1039 | page = pool->alloc.cache[--pool->alloc.count]; | |
458de8a9 | 1040 | page_pool_return_page(pool, page); |
44768dec | 1041 | } |
bc836748 SM |
1042 | } |
1043 | EXPORT_SYMBOL(page_pool_update_nid); |