Loading...
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 */
7
8#include <linux/types.h>
9#include <linux/kernel.h>
10#include <linux/slab.h>
11#include <linux/device.h>
12
13#include <net/page_pool/helpers.h>
14#include <net/xdp.h>
15
16#include <linux/dma-direction.h>
17#include <linux/dma-mapping.h>
18#include <linux/page-flags.h>
19#include <linux/mm.h> /* for put_page() */
20#include <linux/poison.h>
21#include <linux/ethtool.h>
22#include <linux/netdevice.h>
23
24#include <trace/events/page_pool.h>
25
26#include "page_pool_priv.h"
27
28#define DEFER_TIME (msecs_to_jiffies(1000))
29#define DEFER_WARN_INTERVAL (60 * HZ)
30
31#define BIAS_MAX (LONG_MAX >> 1)
32
33#ifdef CONFIG_PAGE_POOL_STATS
34/* alloc_stat_inc is intended to be used in softirq context */
35#define alloc_stat_inc(pool, __stat) (pool->alloc_stats.__stat++)
36/* recycle_stat_inc is safe to use when preemption is possible. */
37#define recycle_stat_inc(pool, __stat) \
38 do { \
39 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats; \
40 this_cpu_inc(s->__stat); \
41 } while (0)
42
43#define recycle_stat_add(pool, __stat, val) \
44 do { \
45 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats; \
46 this_cpu_add(s->__stat, val); \
47 } while (0)
48
49static const char pp_stats[][ETH_GSTRING_LEN] = {
50 "rx_pp_alloc_fast",
51 "rx_pp_alloc_slow",
52 "rx_pp_alloc_slow_ho",
53 "rx_pp_alloc_empty",
54 "rx_pp_alloc_refill",
55 "rx_pp_alloc_waive",
56 "rx_pp_recycle_cached",
57 "rx_pp_recycle_cache_full",
58 "rx_pp_recycle_ring",
59 "rx_pp_recycle_ring_full",
60 "rx_pp_recycle_released_ref",
61};
62
63/**
64 * page_pool_get_stats() - fetch page pool stats
65 * @pool: pool from which page was allocated
66 * @stats: struct page_pool_stats to fill in
67 *
68 * Retrieve statistics about the page_pool. This API is only available
69 * if the kernel has been configured with ``CONFIG_PAGE_POOL_STATS=y``.
70 * A pointer to a caller allocated struct page_pool_stats structure
71 * is passed to this API which is filled in. The caller can then report
72 * those stats to the user (perhaps via ethtool, debugfs, etc.).
73 */
74bool page_pool_get_stats(const struct page_pool *pool,
75 struct page_pool_stats *stats)
76{
77 int cpu = 0;
78
79 if (!stats)
80 return false;
81
82 /* The caller is responsible to initialize stats. */
83 stats->alloc_stats.fast += pool->alloc_stats.fast;
84 stats->alloc_stats.slow += pool->alloc_stats.slow;
85 stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
86 stats->alloc_stats.empty += pool->alloc_stats.empty;
87 stats->alloc_stats.refill += pool->alloc_stats.refill;
88 stats->alloc_stats.waive += pool->alloc_stats.waive;
89
90 for_each_possible_cpu(cpu) {
91 const struct page_pool_recycle_stats *pcpu =
92 per_cpu_ptr(pool->recycle_stats, cpu);
93
94 stats->recycle_stats.cached += pcpu->cached;
95 stats->recycle_stats.cache_full += pcpu->cache_full;
96 stats->recycle_stats.ring += pcpu->ring;
97 stats->recycle_stats.ring_full += pcpu->ring_full;
98 stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
99 }
100
101 return true;
102}
103EXPORT_SYMBOL(page_pool_get_stats);
104
105u8 *page_pool_ethtool_stats_get_strings(u8 *data)
106{
107 int i;
108
109 for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
110 memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
111 data += ETH_GSTRING_LEN;
112 }
113
114 return data;
115}
116EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);
117
118int page_pool_ethtool_stats_get_count(void)
119{
120 return ARRAY_SIZE(pp_stats);
121}
122EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);
123
124u64 *page_pool_ethtool_stats_get(u64 *data, void *stats)
125{
126 struct page_pool_stats *pool_stats = stats;
127
128 *data++ = pool_stats->alloc_stats.fast;
129 *data++ = pool_stats->alloc_stats.slow;
130 *data++ = pool_stats->alloc_stats.slow_high_order;
131 *data++ = pool_stats->alloc_stats.empty;
132 *data++ = pool_stats->alloc_stats.refill;
133 *data++ = pool_stats->alloc_stats.waive;
134 *data++ = pool_stats->recycle_stats.cached;
135 *data++ = pool_stats->recycle_stats.cache_full;
136 *data++ = pool_stats->recycle_stats.ring;
137 *data++ = pool_stats->recycle_stats.ring_full;
138 *data++ = pool_stats->recycle_stats.released_refcnt;
139
140 return data;
141}
142EXPORT_SYMBOL(page_pool_ethtool_stats_get);
143
144#else
145#define alloc_stat_inc(pool, __stat)
146#define recycle_stat_inc(pool, __stat)
147#define recycle_stat_add(pool, __stat, val)
148#endif
149
150static bool page_pool_producer_lock(struct page_pool *pool)
151 __acquires(&pool->ring.producer_lock)
152{
153 bool in_softirq = in_softirq();
154
155 if (in_softirq)
156 spin_lock(&pool->ring.producer_lock);
157 else
158 spin_lock_bh(&pool->ring.producer_lock);
159
160 return in_softirq;
161}
162
163static void page_pool_producer_unlock(struct page_pool *pool,
164 bool in_softirq)
165 __releases(&pool->ring.producer_lock)
166{
167 if (in_softirq)
168 spin_unlock(&pool->ring.producer_lock);
169 else
170 spin_unlock_bh(&pool->ring.producer_lock);
171}
172
173static int page_pool_init(struct page_pool *pool,
174 const struct page_pool_params *params)
175{
176 unsigned int ring_qsize = 1024; /* Default */
177
178 memcpy(&pool->p, ¶ms->fast, sizeof(pool->p));
179 memcpy(&pool->slow, ¶ms->slow, sizeof(pool->slow));
180
181 /* Validate only known flags were used */
182 if (pool->p.flags & ~(PP_FLAG_ALL))
183 return -EINVAL;
184
185 if (pool->p.pool_size)
186 ring_qsize = pool->p.pool_size;
187
188 /* Sanity limit mem that can be pinned down */
189 if (ring_qsize > 32768)
190 return -E2BIG;
191
192 /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
193 * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
194 * which is the XDP_TX use-case.
195 */
196 if (pool->p.flags & PP_FLAG_DMA_MAP) {
197 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
198 (pool->p.dma_dir != DMA_BIDIRECTIONAL))
199 return -EINVAL;
200 }
201
202 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
203 /* In order to request DMA-sync-for-device the page
204 * needs to be mapped
205 */
206 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
207 return -EINVAL;
208
209 if (!pool->p.max_len)
210 return -EINVAL;
211
212 /* pool->p.offset has to be set according to the address
213 * offset used by the DMA engine to start copying rx data
214 */
215 }
216
217 pool->has_init_callback = !!pool->slow.init_callback;
218
219#ifdef CONFIG_PAGE_POOL_STATS
220 pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
221 if (!pool->recycle_stats)
222 return -ENOMEM;
223#endif
224
225 if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0) {
226#ifdef CONFIG_PAGE_POOL_STATS
227 free_percpu(pool->recycle_stats);
228#endif
229 return -ENOMEM;
230 }
231
232 atomic_set(&pool->pages_state_release_cnt, 0);
233
234 /* Driver calling page_pool_create() also call page_pool_destroy() */
235 refcount_set(&pool->user_cnt, 1);
236
237 if (pool->p.flags & PP_FLAG_DMA_MAP)
238 get_device(pool->p.dev);
239
240 return 0;
241}
242
243static void page_pool_uninit(struct page_pool *pool)
244{
245 ptr_ring_cleanup(&pool->ring, NULL);
246
247 if (pool->p.flags & PP_FLAG_DMA_MAP)
248 put_device(pool->p.dev);
249
250#ifdef CONFIG_PAGE_POOL_STATS
251 free_percpu(pool->recycle_stats);
252#endif
253}
254
255/**
256 * page_pool_create() - create a page pool.
257 * @params: parameters, see struct page_pool_params
258 */
259struct page_pool *page_pool_create(const struct page_pool_params *params)
260{
261 struct page_pool *pool;
262 int err;
263
264 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
265 if (!pool)
266 return ERR_PTR(-ENOMEM);
267
268 err = page_pool_init(pool, params);
269 if (err < 0)
270 goto err_free;
271
272 err = page_pool_list(pool);
273 if (err)
274 goto err_uninit;
275
276 return pool;
277
278err_uninit:
279 page_pool_uninit(pool);
280err_free:
281 pr_warn("%s() gave up with errno %d\n", __func__, err);
282 kfree(pool);
283 return ERR_PTR(err);
284}
285EXPORT_SYMBOL(page_pool_create);
286
287static void page_pool_return_page(struct page_pool *pool, struct page *page);
288
289noinline
290static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
291{
292 struct ptr_ring *r = &pool->ring;
293 struct page *page;
294 int pref_nid; /* preferred NUMA node */
295
296 /* Quicker fallback, avoid locks when ring is empty */
297 if (__ptr_ring_empty(r)) {
298 alloc_stat_inc(pool, empty);
299 return NULL;
300 }
301
302 /* Softirq guarantee CPU and thus NUMA node is stable. This,
303 * assumes CPU refilling driver RX-ring will also run RX-NAPI.
304 */
305#ifdef CONFIG_NUMA
306 pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
307#else
308 /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
309 pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
310#endif
311
312 /* Refill alloc array, but only if NUMA match */
313 do {
314 page = __ptr_ring_consume(r);
315 if (unlikely(!page))
316 break;
317
318 if (likely(page_to_nid(page) == pref_nid)) {
319 pool->alloc.cache[pool->alloc.count++] = page;
320 } else {
321 /* NUMA mismatch;
322 * (1) release 1 page to page-allocator and
323 * (2) break out to fallthrough to alloc_pages_node.
324 * This limit stress on page buddy alloactor.
325 */
326 page_pool_return_page(pool, page);
327 alloc_stat_inc(pool, waive);
328 page = NULL;
329 break;
330 }
331 } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
332
333 /* Return last page */
334 if (likely(pool->alloc.count > 0)) {
335 page = pool->alloc.cache[--pool->alloc.count];
336 alloc_stat_inc(pool, refill);
337 }
338
339 return page;
340}
341
342/* fast path */
343static struct page *__page_pool_get_cached(struct page_pool *pool)
344{
345 struct page *page;
346
347 /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
348 if (likely(pool->alloc.count)) {
349 /* Fast-path */
350 page = pool->alloc.cache[--pool->alloc.count];
351 alloc_stat_inc(pool, fast);
352 } else {
353 page = page_pool_refill_alloc_cache(pool);
354 }
355
356 return page;
357}
358
359static void page_pool_dma_sync_for_device(struct page_pool *pool,
360 struct page *page,
361 unsigned int dma_sync_size)
362{
363 dma_addr_t dma_addr = page_pool_get_dma_addr(page);
364
365 dma_sync_size = min(dma_sync_size, pool->p.max_len);
366 dma_sync_single_range_for_device(pool->p.dev, dma_addr,
367 pool->p.offset, dma_sync_size,
368 pool->p.dma_dir);
369}
370
371static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
372{
373 dma_addr_t dma;
374
375 /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
376 * since dma_addr_t can be either 32 or 64 bits and does not always fit
377 * into page private data (i.e 32bit cpu with 64bit DMA caps)
378 * This mapping is kept for lifetime of page, until leaving pool.
379 */
380 dma = dma_map_page_attrs(pool->p.dev, page, 0,
381 (PAGE_SIZE << pool->p.order),
382 pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC |
383 DMA_ATTR_WEAK_ORDERING);
384 if (dma_mapping_error(pool->p.dev, dma))
385 return false;
386
387 if (page_pool_set_dma_addr(page, dma))
388 goto unmap_failed;
389
390 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
391 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
392
393 return true;
394
395unmap_failed:
396 WARN_ON_ONCE("unexpected DMA address, please report to netdev@");
397 dma_unmap_page_attrs(pool->p.dev, dma,
398 PAGE_SIZE << pool->p.order, pool->p.dma_dir,
399 DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
400 return false;
401}
402
403static void page_pool_set_pp_info(struct page_pool *pool,
404 struct page *page)
405{
406 page->pp = pool;
407 page->pp_magic |= PP_SIGNATURE;
408
409 /* Ensuring all pages have been split into one fragment initially:
410 * page_pool_set_pp_info() is only called once for every page when it
411 * is allocated from the page allocator and page_pool_fragment_page()
412 * is dirtying the same cache line as the page->pp_magic above, so
413 * the overhead is negligible.
414 */
415 page_pool_fragment_page(page, 1);
416 if (pool->has_init_callback)
417 pool->slow.init_callback(page, pool->slow.init_arg);
418}
419
420static void page_pool_clear_pp_info(struct page *page)
421{
422 page->pp_magic = 0;
423 page->pp = NULL;
424}
425
426static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
427 gfp_t gfp)
428{
429 struct page *page;
430
431 gfp |= __GFP_COMP;
432 page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
433 if (unlikely(!page))
434 return NULL;
435
436 if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
437 unlikely(!page_pool_dma_map(pool, page))) {
438 put_page(page);
439 return NULL;
440 }
441
442 alloc_stat_inc(pool, slow_high_order);
443 page_pool_set_pp_info(pool, page);
444
445 /* Track how many pages are held 'in-flight' */
446 pool->pages_state_hold_cnt++;
447 trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
448 return page;
449}
450
451/* slow path */
452noinline
453static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
454 gfp_t gfp)
455{
456 const int bulk = PP_ALLOC_CACHE_REFILL;
457 unsigned int pp_flags = pool->p.flags;
458 unsigned int pp_order = pool->p.order;
459 struct page *page;
460 int i, nr_pages;
461
462 /* Don't support bulk alloc for high-order pages */
463 if (unlikely(pp_order))
464 return __page_pool_alloc_page_order(pool, gfp);
465
466 /* Unnecessary as alloc cache is empty, but guarantees zero count */
467 if (unlikely(pool->alloc.count > 0))
468 return pool->alloc.cache[--pool->alloc.count];
469
470 /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
471 memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
472
473 nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk,
474 pool->alloc.cache);
475 if (unlikely(!nr_pages))
476 return NULL;
477
478 /* Pages have been filled into alloc.cache array, but count is zero and
479 * page element have not been (possibly) DMA mapped.
480 */
481 for (i = 0; i < nr_pages; i++) {
482 page = pool->alloc.cache[i];
483 if ((pp_flags & PP_FLAG_DMA_MAP) &&
484 unlikely(!page_pool_dma_map(pool, page))) {
485 put_page(page);
486 continue;
487 }
488
489 page_pool_set_pp_info(pool, page);
490 pool->alloc.cache[pool->alloc.count++] = page;
491 /* Track how many pages are held 'in-flight' */
492 pool->pages_state_hold_cnt++;
493 trace_page_pool_state_hold(pool, page,
494 pool->pages_state_hold_cnt);
495 }
496
497 /* Return last page */
498 if (likely(pool->alloc.count > 0)) {
499 page = pool->alloc.cache[--pool->alloc.count];
500 alloc_stat_inc(pool, slow);
501 } else {
502 page = NULL;
503 }
504
505 /* When page just alloc'ed is should/must have refcnt 1. */
506 return page;
507}
508
509/* For using page_pool replace: alloc_pages() API calls, but provide
510 * synchronization guarantee for allocation side.
511 */
512struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
513{
514 struct page *page;
515
516 /* Fast-path: Get a page from cache */
517 page = __page_pool_get_cached(pool);
518 if (page)
519 return page;
520
521 /* Slow-path: cache empty, do real allocation */
522 page = __page_pool_alloc_pages_slow(pool, gfp);
523 return page;
524}
525EXPORT_SYMBOL(page_pool_alloc_pages);
526
527/* Calculate distance between two u32 values, valid if distance is below 2^(31)
528 * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
529 */
530#define _distance(a, b) (s32)((a) - (b))
531
532s32 page_pool_inflight(const struct page_pool *pool, bool strict)
533{
534 u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
535 u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
536 s32 inflight;
537
538 inflight = _distance(hold_cnt, release_cnt);
539
540 if (strict) {
541 trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
542 WARN(inflight < 0, "Negative(%d) inflight packet-pages",
543 inflight);
544 } else {
545 inflight = max(0, inflight);
546 }
547
548 return inflight;
549}
550
551static __always_inline
552void __page_pool_release_page_dma(struct page_pool *pool, struct page *page)
553{
554 dma_addr_t dma;
555
556 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
557 /* Always account for inflight pages, even if we didn't
558 * map them
559 */
560 return;
561
562 dma = page_pool_get_dma_addr(page);
563
564 /* When page is unmapped, it cannot be returned to our pool */
565 dma_unmap_page_attrs(pool->p.dev, dma,
566 PAGE_SIZE << pool->p.order, pool->p.dma_dir,
567 DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
568 page_pool_set_dma_addr(page, 0);
569}
570
571/* Disconnects a page (from a page_pool). API users can have a need
572 * to disconnect a page (from a page_pool), to allow it to be used as
573 * a regular page (that will eventually be returned to the normal
574 * page-allocator via put_page).
575 */
576void page_pool_return_page(struct page_pool *pool, struct page *page)
577{
578 int count;
579
580 __page_pool_release_page_dma(pool, page);
581
582 page_pool_clear_pp_info(page);
583
584 /* This may be the last page returned, releasing the pool, so
585 * it is not safe to reference pool afterwards.
586 */
587 count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
588 trace_page_pool_state_release(pool, page, count);
589
590 put_page(page);
591 /* An optimization would be to call __free_pages(page, pool->p.order)
592 * knowing page is not part of page-cache (thus avoiding a
593 * __page_cache_release() call).
594 */
595}
596
597static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
598{
599 int ret;
600 /* BH protection not needed if current is softirq */
601 if (in_softirq())
602 ret = ptr_ring_produce(&pool->ring, page);
603 else
604 ret = ptr_ring_produce_bh(&pool->ring, page);
605
606 if (!ret) {
607 recycle_stat_inc(pool, ring);
608 return true;
609 }
610
611 return false;
612}
613
614/* Only allow direct recycling in special circumstances, into the
615 * alloc side cache. E.g. during RX-NAPI processing for XDP_DROP use-case.
616 *
617 * Caller must provide appropriate safe context.
618 */
619static bool page_pool_recycle_in_cache(struct page *page,
620 struct page_pool *pool)
621{
622 if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
623 recycle_stat_inc(pool, cache_full);
624 return false;
625 }
626
627 /* Caller MUST have verified/know (page_ref_count(page) == 1) */
628 pool->alloc.cache[pool->alloc.count++] = page;
629 recycle_stat_inc(pool, cached);
630 return true;
631}
632
633/* If the page refcnt == 1, this will try to recycle the page.
634 * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
635 * the configured size min(dma_sync_size, pool->max_len).
636 * If the page refcnt != 1, then the page will be returned to memory
637 * subsystem.
638 */
639static __always_inline struct page *
640__page_pool_put_page(struct page_pool *pool, struct page *page,
641 unsigned int dma_sync_size, bool allow_direct)
642{
643 lockdep_assert_no_hardirq();
644
645 /* This allocator is optimized for the XDP mode that uses
646 * one-frame-per-page, but have fallbacks that act like the
647 * regular page allocator APIs.
648 *
649 * refcnt == 1 means page_pool owns page, and can recycle it.
650 *
651 * page is NOT reusable when allocated when system is under
652 * some pressure. (page_is_pfmemalloc)
653 */
654 if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
655 /* Read barrier done in page_ref_count / READ_ONCE */
656
657 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
658 page_pool_dma_sync_for_device(pool, page,
659 dma_sync_size);
660
661 if (allow_direct && in_softirq() &&
662 page_pool_recycle_in_cache(page, pool))
663 return NULL;
664
665 /* Page found as candidate for recycling */
666 return page;
667 }
668 /* Fallback/non-XDP mode: API user have elevated refcnt.
669 *
670 * Many drivers split up the page into fragments, and some
671 * want to keep doing this to save memory and do refcnt based
672 * recycling. Support this use case too, to ease drivers
673 * switching between XDP/non-XDP.
674 *
675 * In-case page_pool maintains the DMA mapping, API user must
676 * call page_pool_put_page once. In this elevated refcnt
677 * case, the DMA is unmapped/released, as driver is likely
678 * doing refcnt based recycle tricks, meaning another process
679 * will be invoking put_page.
680 */
681 recycle_stat_inc(pool, released_refcnt);
682 page_pool_return_page(pool, page);
683
684 return NULL;
685}
686
687void page_pool_put_unrefed_page(struct page_pool *pool, struct page *page,
688 unsigned int dma_sync_size, bool allow_direct)
689{
690 page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
691 if (page && !page_pool_recycle_in_ring(pool, page)) {
692 /* Cache full, fallback to free pages */
693 recycle_stat_inc(pool, ring_full);
694 page_pool_return_page(pool, page);
695 }
696}
697EXPORT_SYMBOL(page_pool_put_unrefed_page);
698
699/**
700 * page_pool_put_page_bulk() - release references on multiple pages
701 * @pool: pool from which pages were allocated
702 * @data: array holding page pointers
703 * @count: number of pages in @data
704 *
705 * Tries to refill a number of pages into the ptr_ring cache holding ptr_ring
706 * producer lock. If the ptr_ring is full, page_pool_put_page_bulk()
707 * will release leftover pages to the page allocator.
708 * page_pool_put_page_bulk() is suitable to be run inside the driver NAPI tx
709 * completion loop for the XDP_REDIRECT use case.
710 *
711 * Please note the caller must not use data area after running
712 * page_pool_put_page_bulk(), as this function overwrites it.
713 */
714void page_pool_put_page_bulk(struct page_pool *pool, void **data,
715 int count)
716{
717 int i, bulk_len = 0;
718 bool in_softirq;
719
720 for (i = 0; i < count; i++) {
721 struct page *page = virt_to_head_page(data[i]);
722
723 /* It is not the last user for the page frag case */
724 if (!page_pool_is_last_ref(page))
725 continue;
726
727 page = __page_pool_put_page(pool, page, -1, false);
728 /* Approved for bulk recycling in ptr_ring cache */
729 if (page)
730 data[bulk_len++] = page;
731 }
732
733 if (unlikely(!bulk_len))
734 return;
735
736 /* Bulk producer into ptr_ring page_pool cache */
737 in_softirq = page_pool_producer_lock(pool);
738 for (i = 0; i < bulk_len; i++) {
739 if (__ptr_ring_produce(&pool->ring, data[i])) {
740 /* ring full */
741 recycle_stat_inc(pool, ring_full);
742 break;
743 }
744 }
745 recycle_stat_add(pool, ring, i);
746 page_pool_producer_unlock(pool, in_softirq);
747
748 /* Hopefully all pages was return into ptr_ring */
749 if (likely(i == bulk_len))
750 return;
751
752 /* ptr_ring cache full, free remaining pages outside producer lock
753 * since put_page() with refcnt == 1 can be an expensive operation
754 */
755 for (; i < bulk_len; i++)
756 page_pool_return_page(pool, data[i]);
757}
758EXPORT_SYMBOL(page_pool_put_page_bulk);
759
760static struct page *page_pool_drain_frag(struct page_pool *pool,
761 struct page *page)
762{
763 long drain_count = BIAS_MAX - pool->frag_users;
764
765 /* Some user is still using the page frag */
766 if (likely(page_pool_unref_page(page, drain_count)))
767 return NULL;
768
769 if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
770 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
771 page_pool_dma_sync_for_device(pool, page, -1);
772
773 return page;
774 }
775
776 page_pool_return_page(pool, page);
777 return NULL;
778}
779
780static void page_pool_free_frag(struct page_pool *pool)
781{
782 long drain_count = BIAS_MAX - pool->frag_users;
783 struct page *page = pool->frag_page;
784
785 pool->frag_page = NULL;
786
787 if (!page || page_pool_unref_page(page, drain_count))
788 return;
789
790 page_pool_return_page(pool, page);
791}
792
793struct page *page_pool_alloc_frag(struct page_pool *pool,
794 unsigned int *offset,
795 unsigned int size, gfp_t gfp)
796{
797 unsigned int max_size = PAGE_SIZE << pool->p.order;
798 struct page *page = pool->frag_page;
799
800 if (WARN_ON(size > max_size))
801 return NULL;
802
803 size = ALIGN(size, dma_get_cache_alignment());
804 *offset = pool->frag_offset;
805
806 if (page && *offset + size > max_size) {
807 page = page_pool_drain_frag(pool, page);
808 if (page) {
809 alloc_stat_inc(pool, fast);
810 goto frag_reset;
811 }
812 }
813
814 if (!page) {
815 page = page_pool_alloc_pages(pool, gfp);
816 if (unlikely(!page)) {
817 pool->frag_page = NULL;
818 return NULL;
819 }
820
821 pool->frag_page = page;
822
823frag_reset:
824 pool->frag_users = 1;
825 *offset = 0;
826 pool->frag_offset = size;
827 page_pool_fragment_page(page, BIAS_MAX);
828 return page;
829 }
830
831 pool->frag_users++;
832 pool->frag_offset = *offset + size;
833 alloc_stat_inc(pool, fast);
834 return page;
835}
836EXPORT_SYMBOL(page_pool_alloc_frag);
837
838static void page_pool_empty_ring(struct page_pool *pool)
839{
840 struct page *page;
841
842 /* Empty recycle ring */
843 while ((page = ptr_ring_consume_bh(&pool->ring))) {
844 /* Verify the refcnt invariant of cached pages */
845 if (!(page_ref_count(page) == 1))
846 pr_crit("%s() page_pool refcnt %d violation\n",
847 __func__, page_ref_count(page));
848
849 page_pool_return_page(pool, page);
850 }
851}
852
853static void __page_pool_destroy(struct page_pool *pool)
854{
855 if (pool->disconnect)
856 pool->disconnect(pool);
857
858 page_pool_unlist(pool);
859 page_pool_uninit(pool);
860 kfree(pool);
861}
862
863static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
864{
865 struct page *page;
866
867 if (pool->destroy_cnt)
868 return;
869
870 /* Empty alloc cache, assume caller made sure this is
871 * no-longer in use, and page_pool_alloc_pages() cannot be
872 * call concurrently.
873 */
874 while (pool->alloc.count) {
875 page = pool->alloc.cache[--pool->alloc.count];
876 page_pool_return_page(pool, page);
877 }
878}
879
880static void page_pool_scrub(struct page_pool *pool)
881{
882 page_pool_empty_alloc_cache_once(pool);
883 pool->destroy_cnt++;
884
885 /* No more consumers should exist, but producers could still
886 * be in-flight.
887 */
888 page_pool_empty_ring(pool);
889}
890
891static int page_pool_release(struct page_pool *pool)
892{
893 int inflight;
894
895 page_pool_scrub(pool);
896 inflight = page_pool_inflight(pool, true);
897 if (!inflight)
898 __page_pool_destroy(pool);
899
900 return inflight;
901}
902
903static void page_pool_release_retry(struct work_struct *wq)
904{
905 struct delayed_work *dwq = to_delayed_work(wq);
906 struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
907 void *netdev;
908 int inflight;
909
910 inflight = page_pool_release(pool);
911 if (!inflight)
912 return;
913
914 /* Periodic warning for page pools the user can't see */
915 netdev = READ_ONCE(pool->slow.netdev);
916 if (time_after_eq(jiffies, pool->defer_warn) &&
917 (!netdev || netdev == NET_PTR_POISON)) {
918 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
919
920 pr_warn("%s() stalled pool shutdown: id %u, %d inflight %d sec\n",
921 __func__, pool->user.id, inflight, sec);
922 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
923 }
924
925 /* Still not ready to be disconnected, retry later */
926 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
927}
928
929void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
930 struct xdp_mem_info *mem)
931{
932 refcount_inc(&pool->user_cnt);
933 pool->disconnect = disconnect;
934 pool->xdp_mem_id = mem->id;
935}
936
937void page_pool_unlink_napi(struct page_pool *pool)
938{
939 if (!pool->p.napi)
940 return;
941
942 /* To avoid races with recycling and additional barriers make sure
943 * pool and NAPI are unlinked when NAPI is disabled.
944 */
945 WARN_ON(!test_bit(NAPI_STATE_SCHED, &pool->p.napi->state) ||
946 READ_ONCE(pool->p.napi->list_owner) != -1);
947
948 WRITE_ONCE(pool->p.napi, NULL);
949}
950EXPORT_SYMBOL(page_pool_unlink_napi);
951
952void page_pool_destroy(struct page_pool *pool)
953{
954 if (!pool)
955 return;
956
957 if (!page_pool_put(pool))
958 return;
959
960 page_pool_unlink_napi(pool);
961 page_pool_free_frag(pool);
962
963 if (!page_pool_release(pool))
964 return;
965
966 page_pool_detached(pool);
967 pool->defer_start = jiffies;
968 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
969
970 INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
971 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
972}
973EXPORT_SYMBOL(page_pool_destroy);
974
975/* Caller must provide appropriate safe context, e.g. NAPI. */
976void page_pool_update_nid(struct page_pool *pool, int new_nid)
977{
978 struct page *page;
979
980 trace_page_pool_update_nid(pool, new_nid);
981 pool->p.nid = new_nid;
982
983 /* Flush pool alloc cache, as refill will check NUMA node */
984 while (pool->alloc.count) {
985 page = pool->alloc.cache[--pool->alloc.count];
986 page_pool_return_page(pool, page);
987 }
988}
989EXPORT_SYMBOL(page_pool_update_nid);
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 */
7
8#include <linux/types.h>
9#include <linux/kernel.h>
10#include <linux/slab.h>
11#include <linux/device.h>
12
13#include <net/page_pool.h>
14#include <net/xdp.h>
15
16#include <linux/dma-direction.h>
17#include <linux/dma-mapping.h>
18#include <linux/page-flags.h>
19#include <linux/mm.h> /* for put_page() */
20#include <linux/poison.h>
21#include <linux/ethtool.h>
22
23#include <trace/events/page_pool.h>
24
25#define DEFER_TIME (msecs_to_jiffies(1000))
26#define DEFER_WARN_INTERVAL (60 * HZ)
27
28#define BIAS_MAX LONG_MAX
29
30#ifdef CONFIG_PAGE_POOL_STATS
31/* alloc_stat_inc is intended to be used in softirq context */
32#define alloc_stat_inc(pool, __stat) (pool->alloc_stats.__stat++)
33/* recycle_stat_inc is safe to use when preemption is possible. */
34#define recycle_stat_inc(pool, __stat) \
35 do { \
36 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats; \
37 this_cpu_inc(s->__stat); \
38 } while (0)
39
40#define recycle_stat_add(pool, __stat, val) \
41 do { \
42 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats; \
43 this_cpu_add(s->__stat, val); \
44 } while (0)
45
46static const char pp_stats[][ETH_GSTRING_LEN] = {
47 "rx_pp_alloc_fast",
48 "rx_pp_alloc_slow",
49 "rx_pp_alloc_slow_ho",
50 "rx_pp_alloc_empty",
51 "rx_pp_alloc_refill",
52 "rx_pp_alloc_waive",
53 "rx_pp_recycle_cached",
54 "rx_pp_recycle_cache_full",
55 "rx_pp_recycle_ring",
56 "rx_pp_recycle_ring_full",
57 "rx_pp_recycle_released_ref",
58};
59
60bool page_pool_get_stats(struct page_pool *pool,
61 struct page_pool_stats *stats)
62{
63 int cpu = 0;
64
65 if (!stats)
66 return false;
67
68 /* The caller is responsible to initialize stats. */
69 stats->alloc_stats.fast += pool->alloc_stats.fast;
70 stats->alloc_stats.slow += pool->alloc_stats.slow;
71 stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
72 stats->alloc_stats.empty += pool->alloc_stats.empty;
73 stats->alloc_stats.refill += pool->alloc_stats.refill;
74 stats->alloc_stats.waive += pool->alloc_stats.waive;
75
76 for_each_possible_cpu(cpu) {
77 const struct page_pool_recycle_stats *pcpu =
78 per_cpu_ptr(pool->recycle_stats, cpu);
79
80 stats->recycle_stats.cached += pcpu->cached;
81 stats->recycle_stats.cache_full += pcpu->cache_full;
82 stats->recycle_stats.ring += pcpu->ring;
83 stats->recycle_stats.ring_full += pcpu->ring_full;
84 stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
85 }
86
87 return true;
88}
89EXPORT_SYMBOL(page_pool_get_stats);
90
91u8 *page_pool_ethtool_stats_get_strings(u8 *data)
92{
93 int i;
94
95 for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
96 memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
97 data += ETH_GSTRING_LEN;
98 }
99
100 return data;
101}
102EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);
103
104int page_pool_ethtool_stats_get_count(void)
105{
106 return ARRAY_SIZE(pp_stats);
107}
108EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);
109
110u64 *page_pool_ethtool_stats_get(u64 *data, void *stats)
111{
112 struct page_pool_stats *pool_stats = stats;
113
114 *data++ = pool_stats->alloc_stats.fast;
115 *data++ = pool_stats->alloc_stats.slow;
116 *data++ = pool_stats->alloc_stats.slow_high_order;
117 *data++ = pool_stats->alloc_stats.empty;
118 *data++ = pool_stats->alloc_stats.refill;
119 *data++ = pool_stats->alloc_stats.waive;
120 *data++ = pool_stats->recycle_stats.cached;
121 *data++ = pool_stats->recycle_stats.cache_full;
122 *data++ = pool_stats->recycle_stats.ring;
123 *data++ = pool_stats->recycle_stats.ring_full;
124 *data++ = pool_stats->recycle_stats.released_refcnt;
125
126 return data;
127}
128EXPORT_SYMBOL(page_pool_ethtool_stats_get);
129
130#else
131#define alloc_stat_inc(pool, __stat)
132#define recycle_stat_inc(pool, __stat)
133#define recycle_stat_add(pool, __stat, val)
134#endif
135
136static int page_pool_init(struct page_pool *pool,
137 const struct page_pool_params *params)
138{
139 unsigned int ring_qsize = 1024; /* Default */
140
141 memcpy(&pool->p, params, sizeof(pool->p));
142
143 /* Validate only known flags were used */
144 if (pool->p.flags & ~(PP_FLAG_ALL))
145 return -EINVAL;
146
147 if (pool->p.pool_size)
148 ring_qsize = pool->p.pool_size;
149
150 /* Sanity limit mem that can be pinned down */
151 if (ring_qsize > 32768)
152 return -E2BIG;
153
154 /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
155 * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
156 * which is the XDP_TX use-case.
157 */
158 if (pool->p.flags & PP_FLAG_DMA_MAP) {
159 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
160 (pool->p.dma_dir != DMA_BIDIRECTIONAL))
161 return -EINVAL;
162 }
163
164 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
165 /* In order to request DMA-sync-for-device the page
166 * needs to be mapped
167 */
168 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
169 return -EINVAL;
170
171 if (!pool->p.max_len)
172 return -EINVAL;
173
174 /* pool->p.offset has to be set according to the address
175 * offset used by the DMA engine to start copying rx data
176 */
177 }
178
179 if (PAGE_POOL_DMA_USE_PP_FRAG_COUNT &&
180 pool->p.flags & PP_FLAG_PAGE_FRAG)
181 return -EINVAL;
182
183#ifdef CONFIG_PAGE_POOL_STATS
184 pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
185 if (!pool->recycle_stats)
186 return -ENOMEM;
187#endif
188
189 if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
190 return -ENOMEM;
191
192 atomic_set(&pool->pages_state_release_cnt, 0);
193
194 /* Driver calling page_pool_create() also call page_pool_destroy() */
195 refcount_set(&pool->user_cnt, 1);
196
197 if (pool->p.flags & PP_FLAG_DMA_MAP)
198 get_device(pool->p.dev);
199
200 return 0;
201}
202
203struct page_pool *page_pool_create(const struct page_pool_params *params)
204{
205 struct page_pool *pool;
206 int err;
207
208 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
209 if (!pool)
210 return ERR_PTR(-ENOMEM);
211
212 err = page_pool_init(pool, params);
213 if (err < 0) {
214 pr_warn("%s() gave up with errno %d\n", __func__, err);
215 kfree(pool);
216 return ERR_PTR(err);
217 }
218
219 return pool;
220}
221EXPORT_SYMBOL(page_pool_create);
222
223static void page_pool_return_page(struct page_pool *pool, struct page *page);
224
225noinline
226static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
227{
228 struct ptr_ring *r = &pool->ring;
229 struct page *page;
230 int pref_nid; /* preferred NUMA node */
231
232 /* Quicker fallback, avoid locks when ring is empty */
233 if (__ptr_ring_empty(r)) {
234 alloc_stat_inc(pool, empty);
235 return NULL;
236 }
237
238 /* Softirq guarantee CPU and thus NUMA node is stable. This,
239 * assumes CPU refilling driver RX-ring will also run RX-NAPI.
240 */
241#ifdef CONFIG_NUMA
242 pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
243#else
244 /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
245 pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
246#endif
247
248 /* Refill alloc array, but only if NUMA match */
249 do {
250 page = __ptr_ring_consume(r);
251 if (unlikely(!page))
252 break;
253
254 if (likely(page_to_nid(page) == pref_nid)) {
255 pool->alloc.cache[pool->alloc.count++] = page;
256 } else {
257 /* NUMA mismatch;
258 * (1) release 1 page to page-allocator and
259 * (2) break out to fallthrough to alloc_pages_node.
260 * This limit stress on page buddy alloactor.
261 */
262 page_pool_return_page(pool, page);
263 alloc_stat_inc(pool, waive);
264 page = NULL;
265 break;
266 }
267 } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
268
269 /* Return last page */
270 if (likely(pool->alloc.count > 0)) {
271 page = pool->alloc.cache[--pool->alloc.count];
272 alloc_stat_inc(pool, refill);
273 }
274
275 return page;
276}
277
278/* fast path */
279static struct page *__page_pool_get_cached(struct page_pool *pool)
280{
281 struct page *page;
282
283 /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
284 if (likely(pool->alloc.count)) {
285 /* Fast-path */
286 page = pool->alloc.cache[--pool->alloc.count];
287 alloc_stat_inc(pool, fast);
288 } else {
289 page = page_pool_refill_alloc_cache(pool);
290 }
291
292 return page;
293}
294
295static void page_pool_dma_sync_for_device(struct page_pool *pool,
296 struct page *page,
297 unsigned int dma_sync_size)
298{
299 dma_addr_t dma_addr = page_pool_get_dma_addr(page);
300
301 dma_sync_size = min(dma_sync_size, pool->p.max_len);
302 dma_sync_single_range_for_device(pool->p.dev, dma_addr,
303 pool->p.offset, dma_sync_size,
304 pool->p.dma_dir);
305}
306
307static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
308{
309 dma_addr_t dma;
310
311 /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
312 * since dma_addr_t can be either 32 or 64 bits and does not always fit
313 * into page private data (i.e 32bit cpu with 64bit DMA caps)
314 * This mapping is kept for lifetime of page, until leaving pool.
315 */
316 dma = dma_map_page_attrs(pool->p.dev, page, 0,
317 (PAGE_SIZE << pool->p.order),
318 pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
319 if (dma_mapping_error(pool->p.dev, dma))
320 return false;
321
322 page_pool_set_dma_addr(page, dma);
323
324 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
325 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
326
327 return true;
328}
329
330static void page_pool_set_pp_info(struct page_pool *pool,
331 struct page *page)
332{
333 page->pp = pool;
334 page->pp_magic |= PP_SIGNATURE;
335 if (pool->p.init_callback)
336 pool->p.init_callback(page, pool->p.init_arg);
337}
338
339static void page_pool_clear_pp_info(struct page *page)
340{
341 page->pp_magic = 0;
342 page->pp = NULL;
343}
344
345static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
346 gfp_t gfp)
347{
348 struct page *page;
349
350 gfp |= __GFP_COMP;
351 page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
352 if (unlikely(!page))
353 return NULL;
354
355 if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
356 unlikely(!page_pool_dma_map(pool, page))) {
357 put_page(page);
358 return NULL;
359 }
360
361 alloc_stat_inc(pool, slow_high_order);
362 page_pool_set_pp_info(pool, page);
363
364 /* Track how many pages are held 'in-flight' */
365 pool->pages_state_hold_cnt++;
366 trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
367 return page;
368}
369
370/* slow path */
371noinline
372static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
373 gfp_t gfp)
374{
375 const int bulk = PP_ALLOC_CACHE_REFILL;
376 unsigned int pp_flags = pool->p.flags;
377 unsigned int pp_order = pool->p.order;
378 struct page *page;
379 int i, nr_pages;
380
381 /* Don't support bulk alloc for high-order pages */
382 if (unlikely(pp_order))
383 return __page_pool_alloc_page_order(pool, gfp);
384
385 /* Unnecessary as alloc cache is empty, but guarantees zero count */
386 if (unlikely(pool->alloc.count > 0))
387 return pool->alloc.cache[--pool->alloc.count];
388
389 /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
390 memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
391
392 nr_pages = alloc_pages_bulk_array_node(gfp, pool->p.nid, bulk,
393 pool->alloc.cache);
394 if (unlikely(!nr_pages))
395 return NULL;
396
397 /* Pages have been filled into alloc.cache array, but count is zero and
398 * page element have not been (possibly) DMA mapped.
399 */
400 for (i = 0; i < nr_pages; i++) {
401 page = pool->alloc.cache[i];
402 if ((pp_flags & PP_FLAG_DMA_MAP) &&
403 unlikely(!page_pool_dma_map(pool, page))) {
404 put_page(page);
405 continue;
406 }
407
408 page_pool_set_pp_info(pool, page);
409 pool->alloc.cache[pool->alloc.count++] = page;
410 /* Track how many pages are held 'in-flight' */
411 pool->pages_state_hold_cnt++;
412 trace_page_pool_state_hold(pool, page,
413 pool->pages_state_hold_cnt);
414 }
415
416 /* Return last page */
417 if (likely(pool->alloc.count > 0)) {
418 page = pool->alloc.cache[--pool->alloc.count];
419 alloc_stat_inc(pool, slow);
420 } else {
421 page = NULL;
422 }
423
424 /* When page just alloc'ed is should/must have refcnt 1. */
425 return page;
426}
427
428/* For using page_pool replace: alloc_pages() API calls, but provide
429 * synchronization guarantee for allocation side.
430 */
431struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
432{
433 struct page *page;
434
435 /* Fast-path: Get a page from cache */
436 page = __page_pool_get_cached(pool);
437 if (page)
438 return page;
439
440 /* Slow-path: cache empty, do real allocation */
441 page = __page_pool_alloc_pages_slow(pool, gfp);
442 return page;
443}
444EXPORT_SYMBOL(page_pool_alloc_pages);
445
446/* Calculate distance between two u32 values, valid if distance is below 2^(31)
447 * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
448 */
449#define _distance(a, b) (s32)((a) - (b))
450
451static s32 page_pool_inflight(struct page_pool *pool)
452{
453 u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
454 u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
455 s32 inflight;
456
457 inflight = _distance(hold_cnt, release_cnt);
458
459 trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
460 WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
461
462 return inflight;
463}
464
465/* Disconnects a page (from a page_pool). API users can have a need
466 * to disconnect a page (from a page_pool), to allow it to be used as
467 * a regular page (that will eventually be returned to the normal
468 * page-allocator via put_page).
469 */
470void page_pool_release_page(struct page_pool *pool, struct page *page)
471{
472 dma_addr_t dma;
473 int count;
474
475 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
476 /* Always account for inflight pages, even if we didn't
477 * map them
478 */
479 goto skip_dma_unmap;
480
481 dma = page_pool_get_dma_addr(page);
482
483 /* When page is unmapped, it cannot be returned to our pool */
484 dma_unmap_page_attrs(pool->p.dev, dma,
485 PAGE_SIZE << pool->p.order, pool->p.dma_dir,
486 DMA_ATTR_SKIP_CPU_SYNC);
487 page_pool_set_dma_addr(page, 0);
488skip_dma_unmap:
489 page_pool_clear_pp_info(page);
490
491 /* This may be the last page returned, releasing the pool, so
492 * it is not safe to reference pool afterwards.
493 */
494 count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
495 trace_page_pool_state_release(pool, page, count);
496}
497EXPORT_SYMBOL(page_pool_release_page);
498
499/* Return a page to the page allocator, cleaning up our state */
500static void page_pool_return_page(struct page_pool *pool, struct page *page)
501{
502 page_pool_release_page(pool, page);
503
504 put_page(page);
505 /* An optimization would be to call __free_pages(page, pool->p.order)
506 * knowing page is not part of page-cache (thus avoiding a
507 * __page_cache_release() call).
508 */
509}
510
511static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
512{
513 int ret;
514 /* BH protection not needed if current is serving softirq */
515 if (in_serving_softirq())
516 ret = ptr_ring_produce(&pool->ring, page);
517 else
518 ret = ptr_ring_produce_bh(&pool->ring, page);
519
520 if (!ret) {
521 recycle_stat_inc(pool, ring);
522 return true;
523 }
524
525 return false;
526}
527
528/* Only allow direct recycling in special circumstances, into the
529 * alloc side cache. E.g. during RX-NAPI processing for XDP_DROP use-case.
530 *
531 * Caller must provide appropriate safe context.
532 */
533static bool page_pool_recycle_in_cache(struct page *page,
534 struct page_pool *pool)
535{
536 if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
537 recycle_stat_inc(pool, cache_full);
538 return false;
539 }
540
541 /* Caller MUST have verified/know (page_ref_count(page) == 1) */
542 pool->alloc.cache[pool->alloc.count++] = page;
543 recycle_stat_inc(pool, cached);
544 return true;
545}
546
547/* If the page refcnt == 1, this will try to recycle the page.
548 * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
549 * the configured size min(dma_sync_size, pool->max_len).
550 * If the page refcnt != 1, then the page will be returned to memory
551 * subsystem.
552 */
553static __always_inline struct page *
554__page_pool_put_page(struct page_pool *pool, struct page *page,
555 unsigned int dma_sync_size, bool allow_direct)
556{
557 /* This allocator is optimized for the XDP mode that uses
558 * one-frame-per-page, but have fallbacks that act like the
559 * regular page allocator APIs.
560 *
561 * refcnt == 1 means page_pool owns page, and can recycle it.
562 *
563 * page is NOT reusable when allocated when system is under
564 * some pressure. (page_is_pfmemalloc)
565 */
566 if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
567 /* Read barrier done in page_ref_count / READ_ONCE */
568
569 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
570 page_pool_dma_sync_for_device(pool, page,
571 dma_sync_size);
572
573 if (allow_direct && in_serving_softirq() &&
574 page_pool_recycle_in_cache(page, pool))
575 return NULL;
576
577 /* Page found as candidate for recycling */
578 return page;
579 }
580 /* Fallback/non-XDP mode: API user have elevated refcnt.
581 *
582 * Many drivers split up the page into fragments, and some
583 * want to keep doing this to save memory and do refcnt based
584 * recycling. Support this use case too, to ease drivers
585 * switching between XDP/non-XDP.
586 *
587 * In-case page_pool maintains the DMA mapping, API user must
588 * call page_pool_put_page once. In this elevated refcnt
589 * case, the DMA is unmapped/released, as driver is likely
590 * doing refcnt based recycle tricks, meaning another process
591 * will be invoking put_page.
592 */
593 recycle_stat_inc(pool, released_refcnt);
594 /* Do not replace this with page_pool_return_page() */
595 page_pool_release_page(pool, page);
596 put_page(page);
597
598 return NULL;
599}
600
601void page_pool_put_defragged_page(struct page_pool *pool, struct page *page,
602 unsigned int dma_sync_size, bool allow_direct)
603{
604 page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
605 if (page && !page_pool_recycle_in_ring(pool, page)) {
606 /* Cache full, fallback to free pages */
607 recycle_stat_inc(pool, ring_full);
608 page_pool_return_page(pool, page);
609 }
610}
611EXPORT_SYMBOL(page_pool_put_defragged_page);
612
613/* Caller must not use data area after call, as this function overwrites it */
614void page_pool_put_page_bulk(struct page_pool *pool, void **data,
615 int count)
616{
617 int i, bulk_len = 0;
618
619 for (i = 0; i < count; i++) {
620 struct page *page = virt_to_head_page(data[i]);
621
622 /* It is not the last user for the page frag case */
623 if (!page_pool_is_last_frag(pool, page))
624 continue;
625
626 page = __page_pool_put_page(pool, page, -1, false);
627 /* Approved for bulk recycling in ptr_ring cache */
628 if (page)
629 data[bulk_len++] = page;
630 }
631
632 if (unlikely(!bulk_len))
633 return;
634
635 /* Bulk producer into ptr_ring page_pool cache */
636 page_pool_ring_lock(pool);
637 for (i = 0; i < bulk_len; i++) {
638 if (__ptr_ring_produce(&pool->ring, data[i])) {
639 /* ring full */
640 recycle_stat_inc(pool, ring_full);
641 break;
642 }
643 }
644 recycle_stat_add(pool, ring, i);
645 page_pool_ring_unlock(pool);
646
647 /* Hopefully all pages was return into ptr_ring */
648 if (likely(i == bulk_len))
649 return;
650
651 /* ptr_ring cache full, free remaining pages outside producer lock
652 * since put_page() with refcnt == 1 can be an expensive operation
653 */
654 for (; i < bulk_len; i++)
655 page_pool_return_page(pool, data[i]);
656}
657EXPORT_SYMBOL(page_pool_put_page_bulk);
658
659static struct page *page_pool_drain_frag(struct page_pool *pool,
660 struct page *page)
661{
662 long drain_count = BIAS_MAX - pool->frag_users;
663
664 /* Some user is still using the page frag */
665 if (likely(page_pool_defrag_page(page, drain_count)))
666 return NULL;
667
668 if (page_ref_count(page) == 1 && !page_is_pfmemalloc(page)) {
669 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
670 page_pool_dma_sync_for_device(pool, page, -1);
671
672 return page;
673 }
674
675 page_pool_return_page(pool, page);
676 return NULL;
677}
678
679static void page_pool_free_frag(struct page_pool *pool)
680{
681 long drain_count = BIAS_MAX - pool->frag_users;
682 struct page *page = pool->frag_page;
683
684 pool->frag_page = NULL;
685
686 if (!page || page_pool_defrag_page(page, drain_count))
687 return;
688
689 page_pool_return_page(pool, page);
690}
691
692struct page *page_pool_alloc_frag(struct page_pool *pool,
693 unsigned int *offset,
694 unsigned int size, gfp_t gfp)
695{
696 unsigned int max_size = PAGE_SIZE << pool->p.order;
697 struct page *page = pool->frag_page;
698
699 if (WARN_ON(!(pool->p.flags & PP_FLAG_PAGE_FRAG) ||
700 size > max_size))
701 return NULL;
702
703 size = ALIGN(size, dma_get_cache_alignment());
704 *offset = pool->frag_offset;
705
706 if (page && *offset + size > max_size) {
707 page = page_pool_drain_frag(pool, page);
708 if (page) {
709 alloc_stat_inc(pool, fast);
710 goto frag_reset;
711 }
712 }
713
714 if (!page) {
715 page = page_pool_alloc_pages(pool, gfp);
716 if (unlikely(!page)) {
717 pool->frag_page = NULL;
718 return NULL;
719 }
720
721 pool->frag_page = page;
722
723frag_reset:
724 pool->frag_users = 1;
725 *offset = 0;
726 pool->frag_offset = size;
727 page_pool_fragment_page(page, BIAS_MAX);
728 return page;
729 }
730
731 pool->frag_users++;
732 pool->frag_offset = *offset + size;
733 alloc_stat_inc(pool, fast);
734 return page;
735}
736EXPORT_SYMBOL(page_pool_alloc_frag);
737
738static void page_pool_empty_ring(struct page_pool *pool)
739{
740 struct page *page;
741
742 /* Empty recycle ring */
743 while ((page = ptr_ring_consume_bh(&pool->ring))) {
744 /* Verify the refcnt invariant of cached pages */
745 if (!(page_ref_count(page) == 1))
746 pr_crit("%s() page_pool refcnt %d violation\n",
747 __func__, page_ref_count(page));
748
749 page_pool_return_page(pool, page);
750 }
751}
752
753static void page_pool_free(struct page_pool *pool)
754{
755 if (pool->disconnect)
756 pool->disconnect(pool);
757
758 ptr_ring_cleanup(&pool->ring, NULL);
759
760 if (pool->p.flags & PP_FLAG_DMA_MAP)
761 put_device(pool->p.dev);
762
763#ifdef CONFIG_PAGE_POOL_STATS
764 free_percpu(pool->recycle_stats);
765#endif
766 kfree(pool);
767}
768
769static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
770{
771 struct page *page;
772
773 if (pool->destroy_cnt)
774 return;
775
776 /* Empty alloc cache, assume caller made sure this is
777 * no-longer in use, and page_pool_alloc_pages() cannot be
778 * call concurrently.
779 */
780 while (pool->alloc.count) {
781 page = pool->alloc.cache[--pool->alloc.count];
782 page_pool_return_page(pool, page);
783 }
784}
785
786static void page_pool_scrub(struct page_pool *pool)
787{
788 page_pool_empty_alloc_cache_once(pool);
789 pool->destroy_cnt++;
790
791 /* No more consumers should exist, but producers could still
792 * be in-flight.
793 */
794 page_pool_empty_ring(pool);
795}
796
797static int page_pool_release(struct page_pool *pool)
798{
799 int inflight;
800
801 page_pool_scrub(pool);
802 inflight = page_pool_inflight(pool);
803 if (!inflight)
804 page_pool_free(pool);
805
806 return inflight;
807}
808
809static void page_pool_release_retry(struct work_struct *wq)
810{
811 struct delayed_work *dwq = to_delayed_work(wq);
812 struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
813 int inflight;
814
815 inflight = page_pool_release(pool);
816 if (!inflight)
817 return;
818
819 /* Periodic warning */
820 if (time_after_eq(jiffies, pool->defer_warn)) {
821 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
822
823 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
824 __func__, inflight, sec);
825 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
826 }
827
828 /* Still not ready to be disconnected, retry later */
829 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
830}
831
832void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
833 struct xdp_mem_info *mem)
834{
835 refcount_inc(&pool->user_cnt);
836 pool->disconnect = disconnect;
837 pool->xdp_mem_id = mem->id;
838}
839
840void page_pool_destroy(struct page_pool *pool)
841{
842 if (!pool)
843 return;
844
845 if (!page_pool_put(pool))
846 return;
847
848 page_pool_free_frag(pool);
849
850 if (!page_pool_release(pool))
851 return;
852
853 pool->defer_start = jiffies;
854 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
855
856 INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
857 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
858}
859EXPORT_SYMBOL(page_pool_destroy);
860
861/* Caller must provide appropriate safe context, e.g. NAPI. */
862void page_pool_update_nid(struct page_pool *pool, int new_nid)
863{
864 struct page *page;
865
866 trace_page_pool_update_nid(pool, new_nid);
867 pool->p.nid = new_nid;
868
869 /* Flush pool alloc cache, as refill will check NUMA node */
870 while (pool->alloc.count) {
871 page = pool->alloc.cache[--pool->alloc.count];
872 page_pool_return_page(pool, page);
873 }
874}
875EXPORT_SYMBOL(page_pool_update_nid);
876
877bool page_pool_return_skb_page(struct page *page)
878{
879 struct page_pool *pp;
880
881 page = compound_head(page);
882
883 /* page->pp_magic is OR'ed with PP_SIGNATURE after the allocation
884 * in order to preserve any existing bits, such as bit 0 for the
885 * head page of compound page and bit 1 for pfmemalloc page, so
886 * mask those bits for freeing side when doing below checking,
887 * and page_is_pfmemalloc() is checked in __page_pool_put_page()
888 * to avoid recycling the pfmemalloc page.
889 */
890 if (unlikely((page->pp_magic & ~0x3UL) != PP_SIGNATURE))
891 return false;
892
893 pp = page->pp;
894
895 /* Driver set this to memory recycling info. Reset it on recycle.
896 * This will *not* work for NIC using a split-page memory model.
897 * The page will be returned to the pool here regardless of the
898 * 'flipped' fragment being in use or not.
899 */
900 page_pool_put_full_page(pp, page, false);
901
902 return true;
903}
904EXPORT_SYMBOL(page_pool_return_skb_page);