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/error-injection.h>
9#include <linux/types.h>
10#include <linux/kernel.h>
11#include <linux/slab.h>
12#include <linux/device.h>
13
14#include <net/netdev_rx_queue.h>
15#include <net/page_pool/helpers.h>
16#include <net/xdp.h>
17
18#include <linux/dma-direction.h>
19#include <linux/dma-mapping.h>
20#include <linux/page-flags.h>
21#include <linux/mm.h> /* for put_page() */
22#include <linux/poison.h>
23#include <linux/ethtool.h>
24#include <linux/netdevice.h>
25
26#include <trace/events/page_pool.h>
27
28#include "mp_dmabuf_devmem.h"
29#include "netmem_priv.h"
30#include "page_pool_priv.h"
31
32DEFINE_STATIC_KEY_FALSE(page_pool_mem_providers);
33
34#define DEFER_TIME (msecs_to_jiffies(1000))
35#define DEFER_WARN_INTERVAL (60 * HZ)
36
37#define BIAS_MAX (LONG_MAX >> 1)
38
39#ifdef CONFIG_PAGE_POOL_STATS
40static DEFINE_PER_CPU(struct page_pool_recycle_stats, pp_system_recycle_stats);
41
42/* alloc_stat_inc is intended to be used in softirq context */
43#define alloc_stat_inc(pool, __stat) (pool->alloc_stats.__stat++)
44/* recycle_stat_inc is safe to use when preemption is possible. */
45#define recycle_stat_inc(pool, __stat) \
46 do { \
47 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats; \
48 this_cpu_inc(s->__stat); \
49 } while (0)
50
51#define recycle_stat_add(pool, __stat, val) \
52 do { \
53 struct page_pool_recycle_stats __percpu *s = pool->recycle_stats; \
54 this_cpu_add(s->__stat, val); \
55 } while (0)
56
57static const char pp_stats[][ETH_GSTRING_LEN] = {
58 "rx_pp_alloc_fast",
59 "rx_pp_alloc_slow",
60 "rx_pp_alloc_slow_ho",
61 "rx_pp_alloc_empty",
62 "rx_pp_alloc_refill",
63 "rx_pp_alloc_waive",
64 "rx_pp_recycle_cached",
65 "rx_pp_recycle_cache_full",
66 "rx_pp_recycle_ring",
67 "rx_pp_recycle_ring_full",
68 "rx_pp_recycle_released_ref",
69};
70
71/**
72 * page_pool_get_stats() - fetch page pool stats
73 * @pool: pool from which page was allocated
74 * @stats: struct page_pool_stats to fill in
75 *
76 * Retrieve statistics about the page_pool. This API is only available
77 * if the kernel has been configured with ``CONFIG_PAGE_POOL_STATS=y``.
78 * A pointer to a caller allocated struct page_pool_stats structure
79 * is passed to this API which is filled in. The caller can then report
80 * those stats to the user (perhaps via ethtool, debugfs, etc.).
81 */
82bool page_pool_get_stats(const struct page_pool *pool,
83 struct page_pool_stats *stats)
84{
85 int cpu = 0;
86
87 if (!stats)
88 return false;
89
90 /* The caller is responsible to initialize stats. */
91 stats->alloc_stats.fast += pool->alloc_stats.fast;
92 stats->alloc_stats.slow += pool->alloc_stats.slow;
93 stats->alloc_stats.slow_high_order += pool->alloc_stats.slow_high_order;
94 stats->alloc_stats.empty += pool->alloc_stats.empty;
95 stats->alloc_stats.refill += pool->alloc_stats.refill;
96 stats->alloc_stats.waive += pool->alloc_stats.waive;
97
98 for_each_possible_cpu(cpu) {
99 const struct page_pool_recycle_stats *pcpu =
100 per_cpu_ptr(pool->recycle_stats, cpu);
101
102 stats->recycle_stats.cached += pcpu->cached;
103 stats->recycle_stats.cache_full += pcpu->cache_full;
104 stats->recycle_stats.ring += pcpu->ring;
105 stats->recycle_stats.ring_full += pcpu->ring_full;
106 stats->recycle_stats.released_refcnt += pcpu->released_refcnt;
107 }
108
109 return true;
110}
111EXPORT_SYMBOL(page_pool_get_stats);
112
113u8 *page_pool_ethtool_stats_get_strings(u8 *data)
114{
115 int i;
116
117 for (i = 0; i < ARRAY_SIZE(pp_stats); i++) {
118 memcpy(data, pp_stats[i], ETH_GSTRING_LEN);
119 data += ETH_GSTRING_LEN;
120 }
121
122 return data;
123}
124EXPORT_SYMBOL(page_pool_ethtool_stats_get_strings);
125
126int page_pool_ethtool_stats_get_count(void)
127{
128 return ARRAY_SIZE(pp_stats);
129}
130EXPORT_SYMBOL(page_pool_ethtool_stats_get_count);
131
132u64 *page_pool_ethtool_stats_get(u64 *data, const void *stats)
133{
134 const struct page_pool_stats *pool_stats = stats;
135
136 *data++ = pool_stats->alloc_stats.fast;
137 *data++ = pool_stats->alloc_stats.slow;
138 *data++ = pool_stats->alloc_stats.slow_high_order;
139 *data++ = pool_stats->alloc_stats.empty;
140 *data++ = pool_stats->alloc_stats.refill;
141 *data++ = pool_stats->alloc_stats.waive;
142 *data++ = pool_stats->recycle_stats.cached;
143 *data++ = pool_stats->recycle_stats.cache_full;
144 *data++ = pool_stats->recycle_stats.ring;
145 *data++ = pool_stats->recycle_stats.ring_full;
146 *data++ = pool_stats->recycle_stats.released_refcnt;
147
148 return data;
149}
150EXPORT_SYMBOL(page_pool_ethtool_stats_get);
151
152#else
153#define alloc_stat_inc(pool, __stat)
154#define recycle_stat_inc(pool, __stat)
155#define recycle_stat_add(pool, __stat, val)
156#endif
157
158static bool page_pool_producer_lock(struct page_pool *pool)
159 __acquires(&pool->ring.producer_lock)
160{
161 bool in_softirq = in_softirq();
162
163 if (in_softirq)
164 spin_lock(&pool->ring.producer_lock);
165 else
166 spin_lock_bh(&pool->ring.producer_lock);
167
168 return in_softirq;
169}
170
171static void page_pool_producer_unlock(struct page_pool *pool,
172 bool in_softirq)
173 __releases(&pool->ring.producer_lock)
174{
175 if (in_softirq)
176 spin_unlock(&pool->ring.producer_lock);
177 else
178 spin_unlock_bh(&pool->ring.producer_lock);
179}
180
181static void page_pool_struct_check(void)
182{
183 CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_users);
184 CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_page);
185 CACHELINE_ASSERT_GROUP_MEMBER(struct page_pool, frag, frag_offset);
186 CACHELINE_ASSERT_GROUP_SIZE(struct page_pool, frag,
187 PAGE_POOL_FRAG_GROUP_ALIGN);
188}
189
190static int page_pool_init(struct page_pool *pool,
191 const struct page_pool_params *params,
192 int cpuid)
193{
194 unsigned int ring_qsize = 1024; /* Default */
195 struct netdev_rx_queue *rxq;
196 int err;
197
198 page_pool_struct_check();
199
200 memcpy(&pool->p, ¶ms->fast, sizeof(pool->p));
201 memcpy(&pool->slow, ¶ms->slow, sizeof(pool->slow));
202
203 pool->cpuid = cpuid;
204
205 /* Validate only known flags were used */
206 if (pool->slow.flags & ~PP_FLAG_ALL)
207 return -EINVAL;
208
209 if (pool->p.pool_size)
210 ring_qsize = pool->p.pool_size;
211
212 /* Sanity limit mem that can be pinned down */
213 if (ring_qsize > 32768)
214 return -E2BIG;
215
216 /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
217 * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
218 * which is the XDP_TX use-case.
219 */
220 if (pool->slow.flags & PP_FLAG_DMA_MAP) {
221 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
222 (pool->p.dma_dir != DMA_BIDIRECTIONAL))
223 return -EINVAL;
224
225 pool->dma_map = true;
226 }
227
228 if (pool->slow.flags & PP_FLAG_DMA_SYNC_DEV) {
229 /* In order to request DMA-sync-for-device the page
230 * needs to be mapped
231 */
232 if (!(pool->slow.flags & PP_FLAG_DMA_MAP))
233 return -EINVAL;
234
235 if (!pool->p.max_len)
236 return -EINVAL;
237
238 pool->dma_sync = true;
239
240 /* pool->p.offset has to be set according to the address
241 * offset used by the DMA engine to start copying rx data
242 */
243 }
244
245 pool->has_init_callback = !!pool->slow.init_callback;
246
247#ifdef CONFIG_PAGE_POOL_STATS
248 if (!(pool->slow.flags & PP_FLAG_SYSTEM_POOL)) {
249 pool->recycle_stats = alloc_percpu(struct page_pool_recycle_stats);
250 if (!pool->recycle_stats)
251 return -ENOMEM;
252 } else {
253 /* For system page pool instance we use a singular stats object
254 * instead of allocating a separate percpu variable for each
255 * (also percpu) page pool instance.
256 */
257 pool->recycle_stats = &pp_system_recycle_stats;
258 pool->system = true;
259 }
260#endif
261
262 if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0) {
263#ifdef CONFIG_PAGE_POOL_STATS
264 if (!pool->system)
265 free_percpu(pool->recycle_stats);
266#endif
267 return -ENOMEM;
268 }
269
270 atomic_set(&pool->pages_state_release_cnt, 0);
271
272 /* Driver calling page_pool_create() also call page_pool_destroy() */
273 refcount_set(&pool->user_cnt, 1);
274
275 if (pool->dma_map)
276 get_device(pool->p.dev);
277
278 if (pool->slow.flags & PP_FLAG_ALLOW_UNREADABLE_NETMEM) {
279 /* We rely on rtnl_lock()ing to make sure netdev_rx_queue
280 * configuration doesn't change while we're initializing
281 * the page_pool.
282 */
283 ASSERT_RTNL();
284 rxq = __netif_get_rx_queue(pool->slow.netdev,
285 pool->slow.queue_idx);
286 pool->mp_priv = rxq->mp_params.mp_priv;
287 }
288
289 if (pool->mp_priv) {
290 err = mp_dmabuf_devmem_init(pool);
291 if (err) {
292 pr_warn("%s() mem-provider init failed %d\n", __func__,
293 err);
294 goto free_ptr_ring;
295 }
296
297 static_branch_inc(&page_pool_mem_providers);
298 }
299
300 return 0;
301
302free_ptr_ring:
303 ptr_ring_cleanup(&pool->ring, NULL);
304#ifdef CONFIG_PAGE_POOL_STATS
305 if (!pool->system)
306 free_percpu(pool->recycle_stats);
307#endif
308 return err;
309}
310
311static void page_pool_uninit(struct page_pool *pool)
312{
313 ptr_ring_cleanup(&pool->ring, NULL);
314
315 if (pool->dma_map)
316 put_device(pool->p.dev);
317
318#ifdef CONFIG_PAGE_POOL_STATS
319 if (!pool->system)
320 free_percpu(pool->recycle_stats);
321#endif
322}
323
324/**
325 * page_pool_create_percpu() - create a page pool for a given cpu.
326 * @params: parameters, see struct page_pool_params
327 * @cpuid: cpu identifier
328 */
329struct page_pool *
330page_pool_create_percpu(const struct page_pool_params *params, int cpuid)
331{
332 struct page_pool *pool;
333 int err;
334
335 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
336 if (!pool)
337 return ERR_PTR(-ENOMEM);
338
339 err = page_pool_init(pool, params, cpuid);
340 if (err < 0)
341 goto err_free;
342
343 err = page_pool_list(pool);
344 if (err)
345 goto err_uninit;
346
347 return pool;
348
349err_uninit:
350 page_pool_uninit(pool);
351err_free:
352 pr_warn("%s() gave up with errno %d\n", __func__, err);
353 kfree(pool);
354 return ERR_PTR(err);
355}
356EXPORT_SYMBOL(page_pool_create_percpu);
357
358/**
359 * page_pool_create() - create a page pool
360 * @params: parameters, see struct page_pool_params
361 */
362struct page_pool *page_pool_create(const struct page_pool_params *params)
363{
364 return page_pool_create_percpu(params, -1);
365}
366EXPORT_SYMBOL(page_pool_create);
367
368static void page_pool_return_page(struct page_pool *pool, netmem_ref netmem);
369
370static noinline netmem_ref page_pool_refill_alloc_cache(struct page_pool *pool)
371{
372 struct ptr_ring *r = &pool->ring;
373 netmem_ref netmem;
374 int pref_nid; /* preferred NUMA node */
375
376 /* Quicker fallback, avoid locks when ring is empty */
377 if (__ptr_ring_empty(r)) {
378 alloc_stat_inc(pool, empty);
379 return 0;
380 }
381
382 /* Softirq guarantee CPU and thus NUMA node is stable. This,
383 * assumes CPU refilling driver RX-ring will also run RX-NAPI.
384 */
385#ifdef CONFIG_NUMA
386 pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
387#else
388 /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
389 pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
390#endif
391
392 /* Refill alloc array, but only if NUMA match */
393 do {
394 netmem = (__force netmem_ref)__ptr_ring_consume(r);
395 if (unlikely(!netmem))
396 break;
397
398 if (likely(netmem_is_pref_nid(netmem, pref_nid))) {
399 pool->alloc.cache[pool->alloc.count++] = netmem;
400 } else {
401 /* NUMA mismatch;
402 * (1) release 1 page to page-allocator and
403 * (2) break out to fallthrough to alloc_pages_node.
404 * This limit stress on page buddy alloactor.
405 */
406 page_pool_return_page(pool, netmem);
407 alloc_stat_inc(pool, waive);
408 netmem = 0;
409 break;
410 }
411 } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
412
413 /* Return last page */
414 if (likely(pool->alloc.count > 0)) {
415 netmem = pool->alloc.cache[--pool->alloc.count];
416 alloc_stat_inc(pool, refill);
417 }
418
419 return netmem;
420}
421
422/* fast path */
423static netmem_ref __page_pool_get_cached(struct page_pool *pool)
424{
425 netmem_ref netmem;
426
427 /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
428 if (likely(pool->alloc.count)) {
429 /* Fast-path */
430 netmem = pool->alloc.cache[--pool->alloc.count];
431 alloc_stat_inc(pool, fast);
432 } else {
433 netmem = page_pool_refill_alloc_cache(pool);
434 }
435
436 return netmem;
437}
438
439static void __page_pool_dma_sync_for_device(const struct page_pool *pool,
440 netmem_ref netmem,
441 u32 dma_sync_size)
442{
443#if defined(CONFIG_HAS_DMA) && defined(CONFIG_DMA_NEED_SYNC)
444 dma_addr_t dma_addr = page_pool_get_dma_addr_netmem(netmem);
445
446 dma_sync_size = min(dma_sync_size, pool->p.max_len);
447 __dma_sync_single_for_device(pool->p.dev, dma_addr + pool->p.offset,
448 dma_sync_size, pool->p.dma_dir);
449#endif
450}
451
452static __always_inline void
453page_pool_dma_sync_for_device(const struct page_pool *pool,
454 netmem_ref netmem,
455 u32 dma_sync_size)
456{
457 if (pool->dma_sync && dma_dev_need_sync(pool->p.dev))
458 __page_pool_dma_sync_for_device(pool, netmem, dma_sync_size);
459}
460
461static bool page_pool_dma_map(struct page_pool *pool, netmem_ref netmem)
462{
463 dma_addr_t dma;
464
465 /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
466 * since dma_addr_t can be either 32 or 64 bits and does not always fit
467 * into page private data (i.e 32bit cpu with 64bit DMA caps)
468 * This mapping is kept for lifetime of page, until leaving pool.
469 */
470 dma = dma_map_page_attrs(pool->p.dev, netmem_to_page(netmem), 0,
471 (PAGE_SIZE << pool->p.order), pool->p.dma_dir,
472 DMA_ATTR_SKIP_CPU_SYNC |
473 DMA_ATTR_WEAK_ORDERING);
474 if (dma_mapping_error(pool->p.dev, dma))
475 return false;
476
477 if (page_pool_set_dma_addr_netmem(netmem, dma))
478 goto unmap_failed;
479
480 page_pool_dma_sync_for_device(pool, netmem, pool->p.max_len);
481
482 return true;
483
484unmap_failed:
485 WARN_ONCE(1, "unexpected DMA address, please report to netdev@");
486 dma_unmap_page_attrs(pool->p.dev, dma,
487 PAGE_SIZE << pool->p.order, pool->p.dma_dir,
488 DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
489 return false;
490}
491
492static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
493 gfp_t gfp)
494{
495 struct page *page;
496
497 gfp |= __GFP_COMP;
498 page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
499 if (unlikely(!page))
500 return NULL;
501
502 if (pool->dma_map && unlikely(!page_pool_dma_map(pool, page_to_netmem(page)))) {
503 put_page(page);
504 return NULL;
505 }
506
507 alloc_stat_inc(pool, slow_high_order);
508 page_pool_set_pp_info(pool, page_to_netmem(page));
509
510 /* Track how many pages are held 'in-flight' */
511 pool->pages_state_hold_cnt++;
512 trace_page_pool_state_hold(pool, page_to_netmem(page),
513 pool->pages_state_hold_cnt);
514 return page;
515}
516
517/* slow path */
518static noinline netmem_ref __page_pool_alloc_pages_slow(struct page_pool *pool,
519 gfp_t gfp)
520{
521 const int bulk = PP_ALLOC_CACHE_REFILL;
522 unsigned int pp_order = pool->p.order;
523 bool dma_map = pool->dma_map;
524 netmem_ref netmem;
525 int i, nr_pages;
526
527 /* Don't support bulk alloc for high-order pages */
528 if (unlikely(pp_order))
529 return page_to_netmem(__page_pool_alloc_page_order(pool, gfp));
530
531 /* Unnecessary as alloc cache is empty, but guarantees zero count */
532 if (unlikely(pool->alloc.count > 0))
533 return pool->alloc.cache[--pool->alloc.count];
534
535 /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
536 memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
537
538 nr_pages = alloc_pages_bulk_array_node(gfp,
539 pool->p.nid, bulk,
540 (struct page **)pool->alloc.cache);
541 if (unlikely(!nr_pages))
542 return 0;
543
544 /* Pages have been filled into alloc.cache array, but count is zero and
545 * page element have not been (possibly) DMA mapped.
546 */
547 for (i = 0; i < nr_pages; i++) {
548 netmem = pool->alloc.cache[i];
549 if (dma_map && unlikely(!page_pool_dma_map(pool, netmem))) {
550 put_page(netmem_to_page(netmem));
551 continue;
552 }
553
554 page_pool_set_pp_info(pool, netmem);
555 pool->alloc.cache[pool->alloc.count++] = netmem;
556 /* Track how many pages are held 'in-flight' */
557 pool->pages_state_hold_cnt++;
558 trace_page_pool_state_hold(pool, netmem,
559 pool->pages_state_hold_cnt);
560 }
561
562 /* Return last page */
563 if (likely(pool->alloc.count > 0)) {
564 netmem = pool->alloc.cache[--pool->alloc.count];
565 alloc_stat_inc(pool, slow);
566 } else {
567 netmem = 0;
568 }
569
570 /* When page just alloc'ed is should/must have refcnt 1. */
571 return netmem;
572}
573
574/* For using page_pool replace: alloc_pages() API calls, but provide
575 * synchronization guarantee for allocation side.
576 */
577netmem_ref page_pool_alloc_netmem(struct page_pool *pool, gfp_t gfp)
578{
579 netmem_ref netmem;
580
581 /* Fast-path: Get a page from cache */
582 netmem = __page_pool_get_cached(pool);
583 if (netmem)
584 return netmem;
585
586 /* Slow-path: cache empty, do real allocation */
587 if (static_branch_unlikely(&page_pool_mem_providers) && pool->mp_priv)
588 netmem = mp_dmabuf_devmem_alloc_netmems(pool, gfp);
589 else
590 netmem = __page_pool_alloc_pages_slow(pool, gfp);
591 return netmem;
592}
593EXPORT_SYMBOL(page_pool_alloc_netmem);
594
595struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
596{
597 return netmem_to_page(page_pool_alloc_netmem(pool, gfp));
598}
599EXPORT_SYMBOL(page_pool_alloc_pages);
600ALLOW_ERROR_INJECTION(page_pool_alloc_pages, NULL);
601
602/* Calculate distance between two u32 values, valid if distance is below 2^(31)
603 * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
604 */
605#define _distance(a, b) (s32)((a) - (b))
606
607s32 page_pool_inflight(const struct page_pool *pool, bool strict)
608{
609 u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
610 u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
611 s32 inflight;
612
613 inflight = _distance(hold_cnt, release_cnt);
614
615 if (strict) {
616 trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
617 WARN(inflight < 0, "Negative(%d) inflight packet-pages",
618 inflight);
619 } else {
620 inflight = max(0, inflight);
621 }
622
623 return inflight;
624}
625
626void page_pool_set_pp_info(struct page_pool *pool, netmem_ref netmem)
627{
628 netmem_set_pp(netmem, pool);
629 netmem_or_pp_magic(netmem, PP_SIGNATURE);
630
631 /* Ensuring all pages have been split into one fragment initially:
632 * page_pool_set_pp_info() is only called once for every page when it
633 * is allocated from the page allocator and page_pool_fragment_page()
634 * is dirtying the same cache line as the page->pp_magic above, so
635 * the overhead is negligible.
636 */
637 page_pool_fragment_netmem(netmem, 1);
638 if (pool->has_init_callback)
639 pool->slow.init_callback(netmem, pool->slow.init_arg);
640}
641
642void page_pool_clear_pp_info(netmem_ref netmem)
643{
644 netmem_clear_pp_magic(netmem);
645 netmem_set_pp(netmem, NULL);
646}
647
648static __always_inline void __page_pool_release_page_dma(struct page_pool *pool,
649 netmem_ref netmem)
650{
651 dma_addr_t dma;
652
653 if (!pool->dma_map)
654 /* Always account for inflight pages, even if we didn't
655 * map them
656 */
657 return;
658
659 dma = page_pool_get_dma_addr_netmem(netmem);
660
661 /* When page is unmapped, it cannot be returned to our pool */
662 dma_unmap_page_attrs(pool->p.dev, dma,
663 PAGE_SIZE << pool->p.order, pool->p.dma_dir,
664 DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING);
665 page_pool_set_dma_addr_netmem(netmem, 0);
666}
667
668/* Disconnects a page (from a page_pool). API users can have a need
669 * to disconnect a page (from a page_pool), to allow it to be used as
670 * a regular page (that will eventually be returned to the normal
671 * page-allocator via put_page).
672 */
673void page_pool_return_page(struct page_pool *pool, netmem_ref netmem)
674{
675 int count;
676 bool put;
677
678 put = true;
679 if (static_branch_unlikely(&page_pool_mem_providers) && pool->mp_priv)
680 put = mp_dmabuf_devmem_release_page(pool, netmem);
681 else
682 __page_pool_release_page_dma(pool, netmem);
683
684 /* This may be the last page returned, releasing the pool, so
685 * it is not safe to reference pool afterwards.
686 */
687 count = atomic_inc_return_relaxed(&pool->pages_state_release_cnt);
688 trace_page_pool_state_release(pool, netmem, count);
689
690 if (put) {
691 page_pool_clear_pp_info(netmem);
692 put_page(netmem_to_page(netmem));
693 }
694 /* An optimization would be to call __free_pages(page, pool->p.order)
695 * knowing page is not part of page-cache (thus avoiding a
696 * __page_cache_release() call).
697 */
698}
699
700static bool page_pool_recycle_in_ring(struct page_pool *pool, netmem_ref netmem)
701{
702 int ret;
703 /* BH protection not needed if current is softirq */
704 if (in_softirq())
705 ret = ptr_ring_produce(&pool->ring, (__force void *)netmem);
706 else
707 ret = ptr_ring_produce_bh(&pool->ring, (__force void *)netmem);
708
709 if (!ret) {
710 recycle_stat_inc(pool, ring);
711 return true;
712 }
713
714 return false;
715}
716
717/* Only allow direct recycling in special circumstances, into the
718 * alloc side cache. E.g. during RX-NAPI processing for XDP_DROP use-case.
719 *
720 * Caller must provide appropriate safe context.
721 */
722static bool page_pool_recycle_in_cache(netmem_ref netmem,
723 struct page_pool *pool)
724{
725 if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE)) {
726 recycle_stat_inc(pool, cache_full);
727 return false;
728 }
729
730 /* Caller MUST have verified/know (page_ref_count(page) == 1) */
731 pool->alloc.cache[pool->alloc.count++] = netmem;
732 recycle_stat_inc(pool, cached);
733 return true;
734}
735
736static bool __page_pool_page_can_be_recycled(netmem_ref netmem)
737{
738 return netmem_is_net_iov(netmem) ||
739 (page_ref_count(netmem_to_page(netmem)) == 1 &&
740 !page_is_pfmemalloc(netmem_to_page(netmem)));
741}
742
743/* If the page refcnt == 1, this will try to recycle the page.
744 * If pool->dma_sync is set, we'll try to sync the DMA area for
745 * the configured size min(dma_sync_size, pool->max_len).
746 * If the page refcnt != 1, then the page will be returned to memory
747 * subsystem.
748 */
749static __always_inline netmem_ref
750__page_pool_put_page(struct page_pool *pool, netmem_ref netmem,
751 unsigned int dma_sync_size, bool allow_direct)
752{
753 lockdep_assert_no_hardirq();
754
755 /* This allocator is optimized for the XDP mode that uses
756 * one-frame-per-page, but have fallbacks that act like the
757 * regular page allocator APIs.
758 *
759 * refcnt == 1 means page_pool owns page, and can recycle it.
760 *
761 * page is NOT reusable when allocated when system is under
762 * some pressure. (page_is_pfmemalloc)
763 */
764 if (likely(__page_pool_page_can_be_recycled(netmem))) {
765 /* Read barrier done in page_ref_count / READ_ONCE */
766
767 page_pool_dma_sync_for_device(pool, netmem, dma_sync_size);
768
769 if (allow_direct && page_pool_recycle_in_cache(netmem, pool))
770 return 0;
771
772 /* Page found as candidate for recycling */
773 return netmem;
774 }
775
776 /* Fallback/non-XDP mode: API user have elevated refcnt.
777 *
778 * Many drivers split up the page into fragments, and some
779 * want to keep doing this to save memory and do refcnt based
780 * recycling. Support this use case too, to ease drivers
781 * switching between XDP/non-XDP.
782 *
783 * In-case page_pool maintains the DMA mapping, API user must
784 * call page_pool_put_page once. In this elevated refcnt
785 * case, the DMA is unmapped/released, as driver is likely
786 * doing refcnt based recycle tricks, meaning another process
787 * will be invoking put_page.
788 */
789 recycle_stat_inc(pool, released_refcnt);
790 page_pool_return_page(pool, netmem);
791
792 return 0;
793}
794
795static bool page_pool_napi_local(const struct page_pool *pool)
796{
797 const struct napi_struct *napi;
798 u32 cpuid;
799
800 if (unlikely(!in_softirq()))
801 return false;
802
803 /* Allow direct recycle if we have reasons to believe that we are
804 * in the same context as the consumer would run, so there's
805 * no possible race.
806 * __page_pool_put_page() makes sure we're not in hardirq context
807 * and interrupts are enabled prior to accessing the cache.
808 */
809 cpuid = smp_processor_id();
810 if (READ_ONCE(pool->cpuid) == cpuid)
811 return true;
812
813 napi = READ_ONCE(pool->p.napi);
814
815 return napi && READ_ONCE(napi->list_owner) == cpuid;
816}
817
818void page_pool_put_unrefed_netmem(struct page_pool *pool, netmem_ref netmem,
819 unsigned int dma_sync_size, bool allow_direct)
820{
821 if (!allow_direct)
822 allow_direct = page_pool_napi_local(pool);
823
824 netmem =
825 __page_pool_put_page(pool, netmem, dma_sync_size, allow_direct);
826 if (netmem && !page_pool_recycle_in_ring(pool, netmem)) {
827 /* Cache full, fallback to free pages */
828 recycle_stat_inc(pool, ring_full);
829 page_pool_return_page(pool, netmem);
830 }
831}
832EXPORT_SYMBOL(page_pool_put_unrefed_netmem);
833
834void page_pool_put_unrefed_page(struct page_pool *pool, struct page *page,
835 unsigned int dma_sync_size, bool allow_direct)
836{
837 page_pool_put_unrefed_netmem(pool, page_to_netmem(page), dma_sync_size,
838 allow_direct);
839}
840EXPORT_SYMBOL(page_pool_put_unrefed_page);
841
842/**
843 * page_pool_put_page_bulk() - release references on multiple pages
844 * @pool: pool from which pages were allocated
845 * @data: array holding page pointers
846 * @count: number of pages in @data
847 *
848 * Tries to refill a number of pages into the ptr_ring cache holding ptr_ring
849 * producer lock. If the ptr_ring is full, page_pool_put_page_bulk()
850 * will release leftover pages to the page allocator.
851 * page_pool_put_page_bulk() is suitable to be run inside the driver NAPI tx
852 * completion loop for the XDP_REDIRECT use case.
853 *
854 * Please note the caller must not use data area after running
855 * page_pool_put_page_bulk(), as this function overwrites it.
856 */
857void page_pool_put_page_bulk(struct page_pool *pool, void **data,
858 int count)
859{
860 int i, bulk_len = 0;
861 bool allow_direct;
862 bool in_softirq;
863
864 allow_direct = page_pool_napi_local(pool);
865
866 for (i = 0; i < count; i++) {
867 netmem_ref netmem = page_to_netmem(virt_to_head_page(data[i]));
868
869 /* It is not the last user for the page frag case */
870 if (!page_pool_is_last_ref(netmem))
871 continue;
872
873 netmem = __page_pool_put_page(pool, netmem, -1, allow_direct);
874 /* Approved for bulk recycling in ptr_ring cache */
875 if (netmem)
876 data[bulk_len++] = (__force void *)netmem;
877 }
878
879 if (!bulk_len)
880 return;
881
882 /* Bulk producer into ptr_ring page_pool cache */
883 in_softirq = page_pool_producer_lock(pool);
884 for (i = 0; i < bulk_len; i++) {
885 if (__ptr_ring_produce(&pool->ring, data[i])) {
886 /* ring full */
887 recycle_stat_inc(pool, ring_full);
888 break;
889 }
890 }
891 recycle_stat_add(pool, ring, i);
892 page_pool_producer_unlock(pool, in_softirq);
893
894 /* Hopefully all pages was return into ptr_ring */
895 if (likely(i == bulk_len))
896 return;
897
898 /* ptr_ring cache full, free remaining pages outside producer lock
899 * since put_page() with refcnt == 1 can be an expensive operation
900 */
901 for (; i < bulk_len; i++)
902 page_pool_return_page(pool, (__force netmem_ref)data[i]);
903}
904EXPORT_SYMBOL(page_pool_put_page_bulk);
905
906static netmem_ref page_pool_drain_frag(struct page_pool *pool,
907 netmem_ref netmem)
908{
909 long drain_count = BIAS_MAX - pool->frag_users;
910
911 /* Some user is still using the page frag */
912 if (likely(page_pool_unref_netmem(netmem, drain_count)))
913 return 0;
914
915 if (__page_pool_page_can_be_recycled(netmem)) {
916 page_pool_dma_sync_for_device(pool, netmem, -1);
917 return netmem;
918 }
919
920 page_pool_return_page(pool, netmem);
921 return 0;
922}
923
924static void page_pool_free_frag(struct page_pool *pool)
925{
926 long drain_count = BIAS_MAX - pool->frag_users;
927 netmem_ref netmem = pool->frag_page;
928
929 pool->frag_page = 0;
930
931 if (!netmem || page_pool_unref_netmem(netmem, drain_count))
932 return;
933
934 page_pool_return_page(pool, netmem);
935}
936
937netmem_ref page_pool_alloc_frag_netmem(struct page_pool *pool,
938 unsigned int *offset, unsigned int size,
939 gfp_t gfp)
940{
941 unsigned int max_size = PAGE_SIZE << pool->p.order;
942 netmem_ref netmem = pool->frag_page;
943
944 if (WARN_ON(size > max_size))
945 return 0;
946
947 size = ALIGN(size, dma_get_cache_alignment());
948 *offset = pool->frag_offset;
949
950 if (netmem && *offset + size > max_size) {
951 netmem = page_pool_drain_frag(pool, netmem);
952 if (netmem) {
953 recycle_stat_inc(pool, cached);
954 alloc_stat_inc(pool, fast);
955 goto frag_reset;
956 }
957 }
958
959 if (!netmem) {
960 netmem = page_pool_alloc_netmem(pool, gfp);
961 if (unlikely(!netmem)) {
962 pool->frag_page = 0;
963 return 0;
964 }
965
966 pool->frag_page = netmem;
967
968frag_reset:
969 pool->frag_users = 1;
970 *offset = 0;
971 pool->frag_offset = size;
972 page_pool_fragment_netmem(netmem, BIAS_MAX);
973 return netmem;
974 }
975
976 pool->frag_users++;
977 pool->frag_offset = *offset + size;
978 return netmem;
979}
980EXPORT_SYMBOL(page_pool_alloc_frag_netmem);
981
982struct page *page_pool_alloc_frag(struct page_pool *pool, unsigned int *offset,
983 unsigned int size, gfp_t gfp)
984{
985 return netmem_to_page(page_pool_alloc_frag_netmem(pool, offset, size,
986 gfp));
987}
988EXPORT_SYMBOL(page_pool_alloc_frag);
989
990static void page_pool_empty_ring(struct page_pool *pool)
991{
992 netmem_ref netmem;
993
994 /* Empty recycle ring */
995 while ((netmem = (__force netmem_ref)ptr_ring_consume_bh(&pool->ring))) {
996 /* Verify the refcnt invariant of cached pages */
997 if (!(netmem_ref_count(netmem) == 1))
998 pr_crit("%s() page_pool refcnt %d violation\n",
999 __func__, netmem_ref_count(netmem));
1000
1001 page_pool_return_page(pool, netmem);
1002 }
1003}
1004
1005static void __page_pool_destroy(struct page_pool *pool)
1006{
1007 if (pool->disconnect)
1008 pool->disconnect(pool);
1009
1010 page_pool_unlist(pool);
1011 page_pool_uninit(pool);
1012
1013 if (pool->mp_priv) {
1014 mp_dmabuf_devmem_destroy(pool);
1015 static_branch_dec(&page_pool_mem_providers);
1016 }
1017
1018 kfree(pool);
1019}
1020
1021static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
1022{
1023 netmem_ref netmem;
1024
1025 if (pool->destroy_cnt)
1026 return;
1027
1028 /* Empty alloc cache, assume caller made sure this is
1029 * no-longer in use, and page_pool_alloc_pages() cannot be
1030 * call concurrently.
1031 */
1032 while (pool->alloc.count) {
1033 netmem = pool->alloc.cache[--pool->alloc.count];
1034 page_pool_return_page(pool, netmem);
1035 }
1036}
1037
1038static void page_pool_scrub(struct page_pool *pool)
1039{
1040 page_pool_empty_alloc_cache_once(pool);
1041 pool->destroy_cnt++;
1042
1043 /* No more consumers should exist, but producers could still
1044 * be in-flight.
1045 */
1046 page_pool_empty_ring(pool);
1047}
1048
1049static int page_pool_release(struct page_pool *pool)
1050{
1051 int inflight;
1052
1053 page_pool_scrub(pool);
1054 inflight = page_pool_inflight(pool, true);
1055 if (!inflight)
1056 __page_pool_destroy(pool);
1057
1058 return inflight;
1059}
1060
1061static void page_pool_release_retry(struct work_struct *wq)
1062{
1063 struct delayed_work *dwq = to_delayed_work(wq);
1064 struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
1065 void *netdev;
1066 int inflight;
1067
1068 inflight = page_pool_release(pool);
1069 if (!inflight)
1070 return;
1071
1072 /* Periodic warning for page pools the user can't see */
1073 netdev = READ_ONCE(pool->slow.netdev);
1074 if (time_after_eq(jiffies, pool->defer_warn) &&
1075 (!netdev || netdev == NET_PTR_POISON)) {
1076 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
1077
1078 pr_warn("%s() stalled pool shutdown: id %u, %d inflight %d sec\n",
1079 __func__, pool->user.id, inflight, sec);
1080 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
1081 }
1082
1083 /* Still not ready to be disconnected, retry later */
1084 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
1085}
1086
1087void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *),
1088 const struct xdp_mem_info *mem)
1089{
1090 refcount_inc(&pool->user_cnt);
1091 pool->disconnect = disconnect;
1092 pool->xdp_mem_id = mem->id;
1093}
1094
1095void page_pool_disable_direct_recycling(struct page_pool *pool)
1096{
1097 /* Disable direct recycling based on pool->cpuid.
1098 * Paired with READ_ONCE() in page_pool_napi_local().
1099 */
1100 WRITE_ONCE(pool->cpuid, -1);
1101
1102 if (!pool->p.napi)
1103 return;
1104
1105 /* To avoid races with recycling and additional barriers make sure
1106 * pool and NAPI are unlinked when NAPI is disabled.
1107 */
1108 WARN_ON(!test_bit(NAPI_STATE_SCHED, &pool->p.napi->state));
1109 WARN_ON(READ_ONCE(pool->p.napi->list_owner) != -1);
1110
1111 mutex_lock(&page_pools_lock);
1112 WRITE_ONCE(pool->p.napi, NULL);
1113 mutex_unlock(&page_pools_lock);
1114}
1115EXPORT_SYMBOL(page_pool_disable_direct_recycling);
1116
1117void page_pool_destroy(struct page_pool *pool)
1118{
1119 if (!pool)
1120 return;
1121
1122 if (!page_pool_put(pool))
1123 return;
1124
1125 page_pool_disable_direct_recycling(pool);
1126 page_pool_free_frag(pool);
1127
1128 if (!page_pool_release(pool))
1129 return;
1130
1131 page_pool_detached(pool);
1132 pool->defer_start = jiffies;
1133 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
1134
1135 INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
1136 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
1137}
1138EXPORT_SYMBOL(page_pool_destroy);
1139
1140/* Caller must provide appropriate safe context, e.g. NAPI. */
1141void page_pool_update_nid(struct page_pool *pool, int new_nid)
1142{
1143 netmem_ref netmem;
1144
1145 trace_page_pool_update_nid(pool, new_nid);
1146 pool->p.nid = new_nid;
1147
1148 /* Flush pool alloc cache, as refill will check NUMA node */
1149 while (pool->alloc.count) {
1150 netmem = pool->alloc.cache[--pool->alloc.count];
1151 page_pool_return_page(pool, netmem);
1152 }
1153}
1154EXPORT_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);