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