1 /* SPDX-License-Identifier: GPL-2.0
4 * Author: Jesper Dangaard Brouer <netoptimizer@brouer.com>
5 * Copyright (C) 2016 Red Hat, Inc.
8 #include <linux/types.h>
9 #include <linux/kernel.h>
10 #include <linux/slab.h>
11 #include <linux/device.h>
13 #include <net/page_pool.h>
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>
22 #include <trace/events/page_pool.h>
24 #define DEFER_TIME (msecs_to_jiffies(1000))
25 #define DEFER_WARN_INTERVAL (60 * HZ)
27 static int page_pool_init(struct page_pool *pool,
28 const struct page_pool_params *params)
30 unsigned int ring_qsize = 1024; /* Default */
32 memcpy(&pool->p, params, sizeof(pool->p));
34 /* Validate only known flags were used */
35 if (pool->p.flags & ~(PP_FLAG_ALL))
38 if (pool->p.pool_size)
39 ring_qsize = pool->p.pool_size;
41 /* Sanity limit mem that can be pinned down */
42 if (ring_qsize > 32768)
45 /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
46 * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
47 * which is the XDP_TX use-case.
49 if (pool->p.flags & PP_FLAG_DMA_MAP) {
50 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
51 (pool->p.dma_dir != DMA_BIDIRECTIONAL))
55 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
56 /* In order to request DMA-sync-for-device the page
59 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
65 /* pool->p.offset has to be set according to the address
66 * offset used by the DMA engine to start copying rx data
70 if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
73 atomic_set(&pool->pages_state_release_cnt, 0);
75 /* Driver calling page_pool_create() also call page_pool_destroy() */
76 refcount_set(&pool->user_cnt, 1);
78 if (pool->p.flags & PP_FLAG_DMA_MAP)
79 get_device(pool->p.dev);
84 struct page_pool *page_pool_create(const struct page_pool_params *params)
86 struct page_pool *pool;
89 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
91 return ERR_PTR(-ENOMEM);
93 err = page_pool_init(pool, params);
95 pr_warn("%s() gave up with errno %d\n", __func__, err);
102 EXPORT_SYMBOL(page_pool_create);
104 static void page_pool_return_page(struct page_pool *pool, struct page *page);
107 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
109 struct ptr_ring *r = &pool->ring;
111 int pref_nid; /* preferred NUMA node */
113 /* Quicker fallback, avoid locks when ring is empty */
114 if (__ptr_ring_empty(r))
117 /* Softirq guarantee CPU and thus NUMA node is stable. This,
118 * assumes CPU refilling driver RX-ring will also run RX-NAPI.
121 pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
123 /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
124 pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
127 /* Slower-path: Get pages from locked ring queue */
128 spin_lock(&r->consumer_lock);
130 /* Refill alloc array, but only if NUMA match */
132 page = __ptr_ring_consume(r);
136 if (likely(page_to_nid(page) == pref_nid)) {
137 pool->alloc.cache[pool->alloc.count++] = page;
140 * (1) release 1 page to page-allocator and
141 * (2) break out to fallthrough to alloc_pages_node.
142 * This limit stress on page buddy alloactor.
144 page_pool_return_page(pool, page);
148 } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
150 /* Return last page */
151 if (likely(pool->alloc.count > 0))
152 page = pool->alloc.cache[--pool->alloc.count];
154 spin_unlock(&r->consumer_lock);
159 static struct page *__page_pool_get_cached(struct page_pool *pool)
163 /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
164 if (likely(pool->alloc.count)) {
166 page = pool->alloc.cache[--pool->alloc.count];
168 page = page_pool_refill_alloc_cache(pool);
174 static void page_pool_dma_sync_for_device(struct page_pool *pool,
176 unsigned int dma_sync_size)
178 dma_addr_t dma_addr = page_pool_get_dma_addr(page);
180 dma_sync_size = min(dma_sync_size, pool->p.max_len);
181 dma_sync_single_range_for_device(pool->p.dev, dma_addr,
182 pool->p.offset, dma_sync_size,
186 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
190 /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
191 * since dma_addr_t can be either 32 or 64 bits and does not always fit
192 * into page private data (i.e 32bit cpu with 64bit DMA caps)
193 * This mapping is kept for lifetime of page, until leaving pool.
195 dma = dma_map_page_attrs(pool->p.dev, page, 0,
196 (PAGE_SIZE << pool->p.order),
197 pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
198 if (dma_mapping_error(pool->p.dev, dma))
201 page_pool_set_dma_addr(page, dma);
203 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
204 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
209 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
215 page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
219 if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
220 unlikely(!page_pool_dma_map(pool, page))) {
225 page->pp_magic |= PP_SIGNATURE;
227 /* Track how many pages are held 'in-flight' */
228 pool->pages_state_hold_cnt++;
229 trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
235 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
238 const int bulk = PP_ALLOC_CACHE_REFILL;
239 unsigned int pp_flags = pool->p.flags;
240 unsigned int pp_order = pool->p.order;
244 /* Don't support bulk alloc for high-order pages */
245 if (unlikely(pp_order))
246 return __page_pool_alloc_page_order(pool, gfp);
248 /* Unnecessary as alloc cache is empty, but guarantees zero count */
249 if (unlikely(pool->alloc.count > 0))
250 return pool->alloc.cache[--pool->alloc.count];
252 /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
253 memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
255 nr_pages = alloc_pages_bulk_array(gfp, bulk, pool->alloc.cache);
256 if (unlikely(!nr_pages))
259 /* Pages have been filled into alloc.cache array, but count is zero and
260 * page element have not been (possibly) DMA mapped.
262 for (i = 0; i < nr_pages; i++) {
263 page = pool->alloc.cache[i];
264 if ((pp_flags & PP_FLAG_DMA_MAP) &&
265 unlikely(!page_pool_dma_map(pool, page))) {
269 page->pp_magic |= PP_SIGNATURE;
270 pool->alloc.cache[pool->alloc.count++] = page;
271 /* Track how many pages are held 'in-flight' */
272 pool->pages_state_hold_cnt++;
273 trace_page_pool_state_hold(pool, page,
274 pool->pages_state_hold_cnt);
277 /* Return last page */
278 if (likely(pool->alloc.count > 0))
279 page = pool->alloc.cache[--pool->alloc.count];
283 /* When page just alloc'ed is should/must have refcnt 1. */
287 /* For using page_pool replace: alloc_pages() API calls, but provide
288 * synchronization guarantee for allocation side.
290 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
294 /* Fast-path: Get a page from cache */
295 page = __page_pool_get_cached(pool);
299 /* Slow-path: cache empty, do real allocation */
300 page = __page_pool_alloc_pages_slow(pool, gfp);
303 EXPORT_SYMBOL(page_pool_alloc_pages);
305 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
306 * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
308 #define _distance(a, b) (s32)((a) - (b))
310 static s32 page_pool_inflight(struct page_pool *pool)
312 u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
313 u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
316 inflight = _distance(hold_cnt, release_cnt);
318 trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
319 WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
324 /* Disconnects a page (from a page_pool). API users can have a need
325 * to disconnect a page (from a page_pool), to allow it to be used as
326 * a regular page (that will eventually be returned to the normal
327 * page-allocator via put_page).
329 void page_pool_release_page(struct page_pool *pool, struct page *page)
334 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
335 /* Always account for inflight pages, even if we didn't
340 dma = page_pool_get_dma_addr(page);
342 /* When page is unmapped, it cannot be returned to our pool */
343 dma_unmap_page_attrs(pool->p.dev, dma,
344 PAGE_SIZE << pool->p.order, pool->p.dma_dir,
345 DMA_ATTR_SKIP_CPU_SYNC);
346 page_pool_set_dma_addr(page, 0);
350 /* This may be the last page returned, releasing the pool, so
351 * it is not safe to reference pool afterwards.
353 count = atomic_inc_return(&pool->pages_state_release_cnt);
354 trace_page_pool_state_release(pool, page, count);
356 EXPORT_SYMBOL(page_pool_release_page);
358 /* Return a page to the page allocator, cleaning up our state */
359 static void page_pool_return_page(struct page_pool *pool, struct page *page)
361 page_pool_release_page(pool, page);
364 /* An optimization would be to call __free_pages(page, pool->p.order)
365 * knowing page is not part of page-cache (thus avoiding a
366 * __page_cache_release() call).
370 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
373 /* BH protection not needed if current is serving softirq */
374 if (in_serving_softirq())
375 ret = ptr_ring_produce(&pool->ring, page);
377 ret = ptr_ring_produce_bh(&pool->ring, page);
379 return (ret == 0) ? true : false;
382 /* Only allow direct recycling in special circumstances, into the
383 * alloc side cache. E.g. during RX-NAPI processing for XDP_DROP use-case.
385 * Caller must provide appropriate safe context.
387 static bool page_pool_recycle_in_cache(struct page *page,
388 struct page_pool *pool)
390 if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
393 /* Caller MUST have verified/know (page_ref_count(page) == 1) */
394 pool->alloc.cache[pool->alloc.count++] = page;
398 /* If the page refcnt == 1, this will try to recycle the page.
399 * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
400 * the configured size min(dma_sync_size, pool->max_len).
401 * If the page refcnt != 1, then the page will be returned to memory
404 static __always_inline struct page *
405 __page_pool_put_page(struct page_pool *pool, struct page *page,
406 unsigned int dma_sync_size, bool allow_direct)
408 /* This allocator is optimized for the XDP mode that uses
409 * one-frame-per-page, but have fallbacks that act like the
410 * regular page allocator APIs.
412 * refcnt == 1 means page_pool owns page, and can recycle it.
414 * page is NOT reusable when allocated when system is under
415 * some pressure. (page_is_pfmemalloc)
417 if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
418 /* Read barrier done in page_ref_count / READ_ONCE */
420 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
421 page_pool_dma_sync_for_device(pool, page,
424 if (allow_direct && in_serving_softirq() &&
425 page_pool_recycle_in_cache(page, pool))
428 /* Page found as candidate for recycling */
431 /* Fallback/non-XDP mode: API user have elevated refcnt.
433 * Many drivers split up the page into fragments, and some
434 * want to keep doing this to save memory and do refcnt based
435 * recycling. Support this use case too, to ease drivers
436 * switching between XDP/non-XDP.
438 * In-case page_pool maintains the DMA mapping, API user must
439 * call page_pool_put_page once. In this elevated refcnt
440 * case, the DMA is unmapped/released, as driver is likely
441 * doing refcnt based recycle tricks, meaning another process
442 * will be invoking put_page.
444 /* Do not replace this with page_pool_return_page() */
445 page_pool_release_page(pool, page);
451 void page_pool_put_page(struct page_pool *pool, struct page *page,
452 unsigned int dma_sync_size, bool allow_direct)
454 page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
455 if (page && !page_pool_recycle_in_ring(pool, page)) {
456 /* Cache full, fallback to free pages */
457 page_pool_return_page(pool, page);
460 EXPORT_SYMBOL(page_pool_put_page);
462 /* Caller must not use data area after call, as this function overwrites it */
463 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
468 for (i = 0; i < count; i++) {
469 struct page *page = virt_to_head_page(data[i]);
471 page = __page_pool_put_page(pool, page, -1, false);
472 /* Approved for bulk recycling in ptr_ring cache */
474 data[bulk_len++] = page;
477 if (unlikely(!bulk_len))
480 /* Bulk producer into ptr_ring page_pool cache */
481 page_pool_ring_lock(pool);
482 for (i = 0; i < bulk_len; i++) {
483 if (__ptr_ring_produce(&pool->ring, data[i]))
484 break; /* ring full */
486 page_pool_ring_unlock(pool);
488 /* Hopefully all pages was return into ptr_ring */
489 if (likely(i == bulk_len))
492 /* ptr_ring cache full, free remaining pages outside producer lock
493 * since put_page() with refcnt == 1 can be an expensive operation
495 for (; i < bulk_len; i++)
496 page_pool_return_page(pool, data[i]);
498 EXPORT_SYMBOL(page_pool_put_page_bulk);
500 static void page_pool_empty_ring(struct page_pool *pool)
504 /* Empty recycle ring */
505 while ((page = ptr_ring_consume_bh(&pool->ring))) {
506 /* Verify the refcnt invariant of cached pages */
507 if (!(page_ref_count(page) == 1))
508 pr_crit("%s() page_pool refcnt %d violation\n",
509 __func__, page_ref_count(page));
511 page_pool_return_page(pool, page);
515 static void page_pool_free(struct page_pool *pool)
517 if (pool->disconnect)
518 pool->disconnect(pool);
520 ptr_ring_cleanup(&pool->ring, NULL);
522 if (pool->p.flags & PP_FLAG_DMA_MAP)
523 put_device(pool->p.dev);
528 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
532 if (pool->destroy_cnt)
535 /* Empty alloc cache, assume caller made sure this is
536 * no-longer in use, and page_pool_alloc_pages() cannot be
539 while (pool->alloc.count) {
540 page = pool->alloc.cache[--pool->alloc.count];
541 page_pool_return_page(pool, page);
545 static void page_pool_scrub(struct page_pool *pool)
547 page_pool_empty_alloc_cache_once(pool);
550 /* No more consumers should exist, but producers could still
553 page_pool_empty_ring(pool);
556 static int page_pool_release(struct page_pool *pool)
560 page_pool_scrub(pool);
561 inflight = page_pool_inflight(pool);
563 page_pool_free(pool);
568 static void page_pool_release_retry(struct work_struct *wq)
570 struct delayed_work *dwq = to_delayed_work(wq);
571 struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
574 inflight = page_pool_release(pool);
578 /* Periodic warning */
579 if (time_after_eq(jiffies, pool->defer_warn)) {
580 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
582 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
583 __func__, inflight, sec);
584 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
587 /* Still not ready to be disconnected, retry later */
588 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
591 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *))
593 refcount_inc(&pool->user_cnt);
594 pool->disconnect = disconnect;
597 void page_pool_destroy(struct page_pool *pool)
602 if (!page_pool_put(pool))
605 if (!page_pool_release(pool))
608 pool->defer_start = jiffies;
609 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
611 INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
612 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
614 EXPORT_SYMBOL(page_pool_destroy);
616 /* Caller must provide appropriate safe context, e.g. NAPI. */
617 void page_pool_update_nid(struct page_pool *pool, int new_nid)
621 trace_page_pool_update_nid(pool, new_nid);
622 pool->p.nid = new_nid;
624 /* Flush pool alloc cache, as refill will check NUMA node */
625 while (pool->alloc.count) {
626 page = pool->alloc.cache[--pool->alloc.count];
627 page_pool_return_page(pool, page);
630 EXPORT_SYMBOL(page_pool_update_nid);
632 bool page_pool_return_skb_page(struct page *page)
634 struct page_pool *pp;
636 page = compound_head(page);
638 /* page->pp_magic is OR'ed with PP_SIGNATURE after the allocation
639 * in order to preserve any existing bits, such as bit 0 for the
640 * head page of compound page and bit 1 for pfmemalloc page, so
641 * mask those bits for freeing side when doing below checking,
642 * and page_is_pfmemalloc() is checked in __page_pool_put_page()
643 * to avoid recycling the pfmemalloc page.
645 if (unlikely((page->pp_magic & ~0x3UL) != PP_SIGNATURE))
650 /* Driver set this to memory recycling info. Reset it on recycle.
651 * This will *not* work for NIC using a split-page memory model.
652 * The page will be returned to the pool here regardless of the
653 * 'flipped' fragment being in use or not.
656 page_pool_put_full_page(pp, page, false);
660 EXPORT_SYMBOL(page_pool_return_skb_page);