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() */
21 #include <trace/events/page_pool.h>
23 #define DEFER_TIME (msecs_to_jiffies(1000))
24 #define DEFER_WARN_INTERVAL (60 * HZ)
26 static int page_pool_init(struct page_pool *pool,
27 const struct page_pool_params *params)
29 unsigned int ring_qsize = 1024; /* Default */
31 memcpy(&pool->p, params, sizeof(pool->p));
33 /* Validate only known flags were used */
34 if (pool->p.flags & ~(PP_FLAG_ALL))
37 if (pool->p.pool_size)
38 ring_qsize = pool->p.pool_size;
40 /* Sanity limit mem that can be pinned down */
41 if (ring_qsize > 32768)
44 /* DMA direction is either DMA_FROM_DEVICE or DMA_BIDIRECTIONAL.
45 * DMA_BIDIRECTIONAL is for allowing page used for DMA sending,
46 * which is the XDP_TX use-case.
48 if (pool->p.flags & PP_FLAG_DMA_MAP) {
49 if ((pool->p.dma_dir != DMA_FROM_DEVICE) &&
50 (pool->p.dma_dir != DMA_BIDIRECTIONAL))
54 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
55 /* In order to request DMA-sync-for-device the page
58 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
64 /* pool->p.offset has to be set according to the address
65 * offset used by the DMA engine to start copying rx data
69 if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
72 atomic_set(&pool->pages_state_release_cnt, 0);
74 /* Driver calling page_pool_create() also call page_pool_destroy() */
75 refcount_set(&pool->user_cnt, 1);
77 if (pool->p.flags & PP_FLAG_DMA_MAP)
78 get_device(pool->p.dev);
83 struct page_pool *page_pool_create(const struct page_pool_params *params)
85 struct page_pool *pool;
88 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
90 return ERR_PTR(-ENOMEM);
92 err = page_pool_init(pool, params);
94 pr_warn("%s() gave up with errno %d\n", __func__, err);
101 EXPORT_SYMBOL(page_pool_create);
103 static void page_pool_return_page(struct page_pool *pool, struct page *page);
106 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
108 struct ptr_ring *r = &pool->ring;
110 int pref_nid; /* preferred NUMA node */
112 /* Quicker fallback, avoid locks when ring is empty */
113 if (__ptr_ring_empty(r))
116 /* Softirq guarantee CPU and thus NUMA node is stable. This,
117 * assumes CPU refilling driver RX-ring will also run RX-NAPI.
120 pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
122 /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
123 pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
126 /* Slower-path: Get pages from locked ring queue */
127 spin_lock(&r->consumer_lock);
129 /* Refill alloc array, but only if NUMA match */
131 page = __ptr_ring_consume(r);
135 if (likely(page_to_nid(page) == pref_nid)) {
136 pool->alloc.cache[pool->alloc.count++] = page;
139 * (1) release 1 page to page-allocator and
140 * (2) break out to fallthrough to alloc_pages_node.
141 * This limit stress on page buddy alloactor.
143 page_pool_return_page(pool, page);
147 } while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
149 /* Return last page */
150 if (likely(pool->alloc.count > 0))
151 page = pool->alloc.cache[--pool->alloc.count];
153 spin_unlock(&r->consumer_lock);
158 static struct page *__page_pool_get_cached(struct page_pool *pool)
162 /* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
163 if (likely(pool->alloc.count)) {
165 page = pool->alloc.cache[--pool->alloc.count];
167 page = page_pool_refill_alloc_cache(pool);
173 static void page_pool_dma_sync_for_device(struct page_pool *pool,
175 unsigned int dma_sync_size)
177 dma_addr_t dma_addr = page_pool_get_dma_addr(page);
179 dma_sync_size = min(dma_sync_size, pool->p.max_len);
180 dma_sync_single_range_for_device(pool->p.dev, dma_addr,
181 pool->p.offset, dma_sync_size,
185 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
189 /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
190 * since dma_addr_t can be either 32 or 64 bits and does not always fit
191 * into page private data (i.e 32bit cpu with 64bit DMA caps)
192 * This mapping is kept for lifetime of page, until leaving pool.
194 dma = dma_map_page_attrs(pool->p.dev, page, 0,
195 (PAGE_SIZE << pool->p.order),
196 pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
197 if (dma_mapping_error(pool->p.dev, dma))
200 page_pool_set_dma_addr(page, dma);
202 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
203 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
208 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
214 page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
218 if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
219 unlikely(!page_pool_dma_map(pool, page))) {
224 /* Track how many pages are held 'in-flight' */
225 pool->pages_state_hold_cnt++;
226 trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
232 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
235 const int bulk = PP_ALLOC_CACHE_REFILL;
236 unsigned int pp_flags = pool->p.flags;
237 unsigned int pp_order = pool->p.order;
241 /* Don't support bulk alloc for high-order pages */
242 if (unlikely(pp_order))
243 return __page_pool_alloc_page_order(pool, gfp);
245 /* Unnecessary as alloc cache is empty, but guarantees zero count */
246 if (unlikely(pool->alloc.count > 0))
247 return pool->alloc.cache[--pool->alloc.count];
249 /* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
250 memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
252 nr_pages = alloc_pages_bulk_array(gfp, bulk, pool->alloc.cache);
253 if (unlikely(!nr_pages))
256 /* Pages have been filled into alloc.cache array, but count is zero and
257 * page element have not been (possibly) DMA mapped.
259 for (i = 0; i < nr_pages; i++) {
260 page = pool->alloc.cache[i];
261 if ((pp_flags & PP_FLAG_DMA_MAP) &&
262 unlikely(!page_pool_dma_map(pool, page))) {
266 pool->alloc.cache[pool->alloc.count++] = page;
267 /* Track how many pages are held 'in-flight' */
268 pool->pages_state_hold_cnt++;
269 trace_page_pool_state_hold(pool, page,
270 pool->pages_state_hold_cnt);
273 /* Return last page */
274 if (likely(pool->alloc.count > 0))
275 page = pool->alloc.cache[--pool->alloc.count];
279 /* When page just alloc'ed is should/must have refcnt 1. */
283 /* For using page_pool replace: alloc_pages() API calls, but provide
284 * synchronization guarantee for allocation side.
286 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
290 /* Fast-path: Get a page from cache */
291 page = __page_pool_get_cached(pool);
295 /* Slow-path: cache empty, do real allocation */
296 page = __page_pool_alloc_pages_slow(pool, gfp);
299 EXPORT_SYMBOL(page_pool_alloc_pages);
301 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
302 * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
304 #define _distance(a, b) (s32)((a) - (b))
306 static s32 page_pool_inflight(struct page_pool *pool)
308 u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
309 u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
312 inflight = _distance(hold_cnt, release_cnt);
314 trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
315 WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
320 /* Disconnects a page (from a page_pool). API users can have a need
321 * to disconnect a page (from a page_pool), to allow it to be used as
322 * a regular page (that will eventually be returned to the normal
323 * page-allocator via put_page).
325 void page_pool_release_page(struct page_pool *pool, struct page *page)
330 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
331 /* Always account for inflight pages, even if we didn't
336 dma = page_pool_get_dma_addr(page);
338 /* When page is unmapped, it cannot be returned to our pool */
339 dma_unmap_page_attrs(pool->p.dev, dma,
340 PAGE_SIZE << pool->p.order, pool->p.dma_dir,
341 DMA_ATTR_SKIP_CPU_SYNC);
342 page_pool_set_dma_addr(page, 0);
344 /* This may be the last page returned, releasing the pool, so
345 * it is not safe to reference pool afterwards.
347 count = atomic_inc_return(&pool->pages_state_release_cnt);
348 trace_page_pool_state_release(pool, page, count);
350 EXPORT_SYMBOL(page_pool_release_page);
352 /* Return a page to the page allocator, cleaning up our state */
353 static void page_pool_return_page(struct page_pool *pool, struct page *page)
355 page_pool_release_page(pool, page);
358 /* An optimization would be to call __free_pages(page, pool->p.order)
359 * knowing page is not part of page-cache (thus avoiding a
360 * __page_cache_release() call).
364 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
367 /* BH protection not needed if current is serving softirq */
368 if (in_serving_softirq())
369 ret = ptr_ring_produce(&pool->ring, page);
371 ret = ptr_ring_produce_bh(&pool->ring, page);
373 return (ret == 0) ? true : false;
376 /* Only allow direct recycling in special circumstances, into the
377 * alloc side cache. E.g. during RX-NAPI processing for XDP_DROP use-case.
379 * Caller must provide appropriate safe context.
381 static bool page_pool_recycle_in_cache(struct page *page,
382 struct page_pool *pool)
384 if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
387 /* Caller MUST have verified/know (page_ref_count(page) == 1) */
388 pool->alloc.cache[pool->alloc.count++] = page;
392 /* If the page refcnt == 1, this will try to recycle the page.
393 * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
394 * the configured size min(dma_sync_size, pool->max_len).
395 * If the page refcnt != 1, then the page will be returned to memory
398 static __always_inline struct page *
399 __page_pool_put_page(struct page_pool *pool, struct page *page,
400 unsigned int dma_sync_size, bool allow_direct)
402 /* This allocator is optimized for the XDP mode that uses
403 * one-frame-per-page, but have fallbacks that act like the
404 * regular page allocator APIs.
406 * refcnt == 1 means page_pool owns page, and can recycle it.
408 * page is NOT reusable when allocated when system is under
409 * some pressure. (page_is_pfmemalloc)
411 if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
412 /* Read barrier done in page_ref_count / READ_ONCE */
414 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
415 page_pool_dma_sync_for_device(pool, page,
418 if (allow_direct && in_serving_softirq() &&
419 page_pool_recycle_in_cache(page, pool))
422 /* Page found as candidate for recycling */
425 /* Fallback/non-XDP mode: API user have elevated refcnt.
427 * Many drivers split up the page into fragments, and some
428 * want to keep doing this to save memory and do refcnt based
429 * recycling. Support this use case too, to ease drivers
430 * switching between XDP/non-XDP.
432 * In-case page_pool maintains the DMA mapping, API user must
433 * call page_pool_put_page once. In this elevated refcnt
434 * case, the DMA is unmapped/released, as driver is likely
435 * doing refcnt based recycle tricks, meaning another process
436 * will be invoking put_page.
438 /* Do not replace this with page_pool_return_page() */
439 page_pool_release_page(pool, page);
445 void page_pool_put_page(struct page_pool *pool, struct page *page,
446 unsigned int dma_sync_size, bool allow_direct)
448 page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
449 if (page && !page_pool_recycle_in_ring(pool, page)) {
450 /* Cache full, fallback to free pages */
451 page_pool_return_page(pool, page);
454 EXPORT_SYMBOL(page_pool_put_page);
456 /* Caller must not use data area after call, as this function overwrites it */
457 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
462 for (i = 0; i < count; i++) {
463 struct page *page = virt_to_head_page(data[i]);
465 page = __page_pool_put_page(pool, page, -1, false);
466 /* Approved for bulk recycling in ptr_ring cache */
468 data[bulk_len++] = page;
471 if (unlikely(!bulk_len))
474 /* Bulk producer into ptr_ring page_pool cache */
475 page_pool_ring_lock(pool);
476 for (i = 0; i < bulk_len; i++) {
477 if (__ptr_ring_produce(&pool->ring, data[i]))
478 break; /* ring full */
480 page_pool_ring_unlock(pool);
482 /* Hopefully all pages was return into ptr_ring */
483 if (likely(i == bulk_len))
486 /* ptr_ring cache full, free remaining pages outside producer lock
487 * since put_page() with refcnt == 1 can be an expensive operation
489 for (; i < bulk_len; i++)
490 page_pool_return_page(pool, data[i]);
492 EXPORT_SYMBOL(page_pool_put_page_bulk);
494 static void page_pool_empty_ring(struct page_pool *pool)
498 /* Empty recycle ring */
499 while ((page = ptr_ring_consume_bh(&pool->ring))) {
500 /* Verify the refcnt invariant of cached pages */
501 if (!(page_ref_count(page) == 1))
502 pr_crit("%s() page_pool refcnt %d violation\n",
503 __func__, page_ref_count(page));
505 page_pool_return_page(pool, page);
509 static void page_pool_free(struct page_pool *pool)
511 if (pool->disconnect)
512 pool->disconnect(pool);
514 ptr_ring_cleanup(&pool->ring, NULL);
516 if (pool->p.flags & PP_FLAG_DMA_MAP)
517 put_device(pool->p.dev);
522 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
526 if (pool->destroy_cnt)
529 /* Empty alloc cache, assume caller made sure this is
530 * no-longer in use, and page_pool_alloc_pages() cannot be
533 while (pool->alloc.count) {
534 page = pool->alloc.cache[--pool->alloc.count];
535 page_pool_return_page(pool, page);
539 static void page_pool_scrub(struct page_pool *pool)
541 page_pool_empty_alloc_cache_once(pool);
544 /* No more consumers should exist, but producers could still
547 page_pool_empty_ring(pool);
550 static int page_pool_release(struct page_pool *pool)
554 page_pool_scrub(pool);
555 inflight = page_pool_inflight(pool);
557 page_pool_free(pool);
562 static void page_pool_release_retry(struct work_struct *wq)
564 struct delayed_work *dwq = to_delayed_work(wq);
565 struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
568 inflight = page_pool_release(pool);
572 /* Periodic warning */
573 if (time_after_eq(jiffies, pool->defer_warn)) {
574 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
576 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
577 __func__, inflight, sec);
578 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
581 /* Still not ready to be disconnected, retry later */
582 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
585 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *))
587 refcount_inc(&pool->user_cnt);
588 pool->disconnect = disconnect;
591 void page_pool_destroy(struct page_pool *pool)
596 if (!page_pool_put(pool))
599 if (!page_pool_release(pool))
602 pool->defer_start = jiffies;
603 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
605 INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
606 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
608 EXPORT_SYMBOL(page_pool_destroy);
610 /* Caller must provide appropriate safe context, e.g. NAPI. */
611 void page_pool_update_nid(struct page_pool *pool, int new_nid)
615 trace_page_pool_update_nid(pool, new_nid);
616 pool->p.nid = new_nid;
618 /* Flush pool alloc cache, as refill will check NUMA node */
619 while (pool->alloc.count) {
620 page = pool->alloc.cache[--pool->alloc.count];
621 page_pool_return_page(pool, page);
624 EXPORT_SYMBOL(page_pool_update_nid);