xref: /linux/net/core/page_pool.c (revision 87c9c16317882dd6dbbc07e349bc3223e14f3244)
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 
21 #include <trace/events/page_pool.h>
22 
23 #define DEFER_TIME (msecs_to_jiffies(1000))
24 #define DEFER_WARN_INTERVAL (60 * HZ)
25 
26 static int page_pool_init(struct page_pool *pool,
27 			  const struct page_pool_params *params)
28 {
29 	unsigned int ring_qsize = 1024; /* Default */
30 
31 	memcpy(&pool->p, params, sizeof(pool->p));
32 
33 	/* Validate only known flags were used */
34 	if (pool->p.flags & ~(PP_FLAG_ALL))
35 		return -EINVAL;
36 
37 	if (pool->p.pool_size)
38 		ring_qsize = pool->p.pool_size;
39 
40 	/* Sanity limit mem that can be pinned down */
41 	if (ring_qsize > 32768)
42 		return -E2BIG;
43 
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.
47 	 */
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))
51 			return -EINVAL;
52 	}
53 
54 	if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV) {
55 		/* In order to request DMA-sync-for-device the page
56 		 * needs to be mapped
57 		 */
58 		if (!(pool->p.flags & PP_FLAG_DMA_MAP))
59 			return -EINVAL;
60 
61 		if (!pool->p.max_len)
62 			return -EINVAL;
63 
64 		/* pool->p.offset has to be set according to the address
65 		 * offset used by the DMA engine to start copying rx data
66 		 */
67 	}
68 
69 	if (ptr_ring_init(&pool->ring, ring_qsize, GFP_KERNEL) < 0)
70 		return -ENOMEM;
71 
72 	atomic_set(&pool->pages_state_release_cnt, 0);
73 
74 	/* Driver calling page_pool_create() also call page_pool_destroy() */
75 	refcount_set(&pool->user_cnt, 1);
76 
77 	if (pool->p.flags & PP_FLAG_DMA_MAP)
78 		get_device(pool->p.dev);
79 
80 	return 0;
81 }
82 
83 struct page_pool *page_pool_create(const struct page_pool_params *params)
84 {
85 	struct page_pool *pool;
86 	int err;
87 
88 	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, params->nid);
89 	if (!pool)
90 		return ERR_PTR(-ENOMEM);
91 
92 	err = page_pool_init(pool, params);
93 	if (err < 0) {
94 		pr_warn("%s() gave up with errno %d\n", __func__, err);
95 		kfree(pool);
96 		return ERR_PTR(err);
97 	}
98 
99 	return pool;
100 }
101 EXPORT_SYMBOL(page_pool_create);
102 
103 static void page_pool_return_page(struct page_pool *pool, struct page *page);
104 
105 noinline
106 static struct page *page_pool_refill_alloc_cache(struct page_pool *pool)
107 {
108 	struct ptr_ring *r = &pool->ring;
109 	struct page *page;
110 	int pref_nid; /* preferred NUMA node */
111 
112 	/* Quicker fallback, avoid locks when ring is empty */
113 	if (__ptr_ring_empty(r))
114 		return NULL;
115 
116 	/* Softirq guarantee CPU and thus NUMA node is stable. This,
117 	 * assumes CPU refilling driver RX-ring will also run RX-NAPI.
118 	 */
119 #ifdef CONFIG_NUMA
120 	pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
121 #else
122 	/* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
123 	pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
124 #endif
125 
126 	/* Slower-path: Get pages from locked ring queue */
127 	spin_lock(&r->consumer_lock);
128 
129 	/* Refill alloc array, but only if NUMA match */
130 	do {
131 		page = __ptr_ring_consume(r);
132 		if (unlikely(!page))
133 			break;
134 
135 		if (likely(page_to_nid(page) == pref_nid)) {
136 			pool->alloc.cache[pool->alloc.count++] = page;
137 		} else {
138 			/* NUMA mismatch;
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.
142 			 */
143 			page_pool_return_page(pool, page);
144 			page = NULL;
145 			break;
146 		}
147 	} while (pool->alloc.count < PP_ALLOC_CACHE_REFILL);
148 
149 	/* Return last page */
150 	if (likely(pool->alloc.count > 0))
151 		page = pool->alloc.cache[--pool->alloc.count];
152 
153 	spin_unlock(&r->consumer_lock);
154 	return page;
155 }
156 
157 /* fast path */
158 static struct page *__page_pool_get_cached(struct page_pool *pool)
159 {
160 	struct page *page;
161 
162 	/* Caller MUST guarantee safe non-concurrent access, e.g. softirq */
163 	if (likely(pool->alloc.count)) {
164 		/* Fast-path */
165 		page = pool->alloc.cache[--pool->alloc.count];
166 	} else {
167 		page = page_pool_refill_alloc_cache(pool);
168 	}
169 
170 	return page;
171 }
172 
173 static void page_pool_dma_sync_for_device(struct page_pool *pool,
174 					  struct page *page,
175 					  unsigned int dma_sync_size)
176 {
177 	dma_addr_t dma_addr = page_pool_get_dma_addr(page);
178 
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,
182 					 pool->p.dma_dir);
183 }
184 
185 static bool page_pool_dma_map(struct page_pool *pool, struct page *page)
186 {
187 	dma_addr_t dma;
188 
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.
193 	 */
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))
198 		return false;
199 
200 	page_pool_set_dma_addr(page, dma);
201 
202 	if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
203 		page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
204 
205 	return true;
206 }
207 
208 static struct page *__page_pool_alloc_page_order(struct page_pool *pool,
209 						 gfp_t gfp)
210 {
211 	struct page *page;
212 
213 	gfp |= __GFP_COMP;
214 	page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
215 	if (unlikely(!page))
216 		return NULL;
217 
218 	if ((pool->p.flags & PP_FLAG_DMA_MAP) &&
219 	    unlikely(!page_pool_dma_map(pool, page))) {
220 		put_page(page);
221 		return NULL;
222 	}
223 
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);
227 	return page;
228 }
229 
230 /* slow path */
231 noinline
232 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
233 						 gfp_t gfp)
234 {
235 	const int bulk = PP_ALLOC_CACHE_REFILL;
236 	unsigned int pp_flags = pool->p.flags;
237 	unsigned int pp_order = pool->p.order;
238 	struct page *page;
239 	int i, nr_pages;
240 
241 	/* Don't support bulk alloc for high-order pages */
242 	if (unlikely(pp_order))
243 		return __page_pool_alloc_page_order(pool, gfp);
244 
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];
248 
249 	/* Mark empty alloc.cache slots "empty" for alloc_pages_bulk_array */
250 	memset(&pool->alloc.cache, 0, sizeof(void *) * bulk);
251 
252 	nr_pages = alloc_pages_bulk_array(gfp, bulk, pool->alloc.cache);
253 	if (unlikely(!nr_pages))
254 		return NULL;
255 
256 	/* Pages have been filled into alloc.cache array, but count is zero and
257 	 * page element have not been (possibly) DMA mapped.
258 	 */
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))) {
263 			put_page(page);
264 			continue;
265 		}
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);
271 	}
272 
273 	/* Return last page */
274 	if (likely(pool->alloc.count > 0))
275 		page = pool->alloc.cache[--pool->alloc.count];
276 	else
277 		page = NULL;
278 
279 	/* When page just alloc'ed is should/must have refcnt 1. */
280 	return page;
281 }
282 
283 /* For using page_pool replace: alloc_pages() API calls, but provide
284  * synchronization guarantee for allocation side.
285  */
286 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
287 {
288 	struct page *page;
289 
290 	/* Fast-path: Get a page from cache */
291 	page = __page_pool_get_cached(pool);
292 	if (page)
293 		return page;
294 
295 	/* Slow-path: cache empty, do real allocation */
296 	page = __page_pool_alloc_pages_slow(pool, gfp);
297 	return page;
298 }
299 EXPORT_SYMBOL(page_pool_alloc_pages);
300 
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
303  */
304 #define _distance(a, b)	(s32)((a) - (b))
305 
306 static s32 page_pool_inflight(struct page_pool *pool)
307 {
308 	u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
309 	u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
310 	s32 inflight;
311 
312 	inflight = _distance(hold_cnt, release_cnt);
313 
314 	trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
315 	WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
316 
317 	return inflight;
318 }
319 
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).
324  */
325 void page_pool_release_page(struct page_pool *pool, struct page *page)
326 {
327 	dma_addr_t dma;
328 	int count;
329 
330 	if (!(pool->p.flags & PP_FLAG_DMA_MAP))
331 		/* Always account for inflight pages, even if we didn't
332 		 * map them
333 		 */
334 		goto skip_dma_unmap;
335 
336 	dma = page_pool_get_dma_addr(page);
337 
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);
343 skip_dma_unmap:
344 	/* This may be the last page returned, releasing the pool, so
345 	 * it is not safe to reference pool afterwards.
346 	 */
347 	count = atomic_inc_return(&pool->pages_state_release_cnt);
348 	trace_page_pool_state_release(pool, page, count);
349 }
350 EXPORT_SYMBOL(page_pool_release_page);
351 
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)
354 {
355 	page_pool_release_page(pool, page);
356 
357 	put_page(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).
361 	 */
362 }
363 
364 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
365 {
366 	int ret;
367 	/* BH protection not needed if current is serving softirq */
368 	if (in_serving_softirq())
369 		ret = ptr_ring_produce(&pool->ring, page);
370 	else
371 		ret = ptr_ring_produce_bh(&pool->ring, page);
372 
373 	return (ret == 0) ? true : false;
374 }
375 
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.
378  *
379  * Caller must provide appropriate safe context.
380  */
381 static bool page_pool_recycle_in_cache(struct page *page,
382 				       struct page_pool *pool)
383 {
384 	if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
385 		return false;
386 
387 	/* Caller MUST have verified/know (page_ref_count(page) == 1) */
388 	pool->alloc.cache[pool->alloc.count++] = page;
389 	return true;
390 }
391 
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
396  * subsystem.
397  */
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)
401 {
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.
405 	 *
406 	 * refcnt == 1 means page_pool owns page, and can recycle it.
407 	 *
408 	 * page is NOT reusable when allocated when system is under
409 	 * some pressure. (page_is_pfmemalloc)
410 	 */
411 	if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
412 		/* Read barrier done in page_ref_count / READ_ONCE */
413 
414 		if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
415 			page_pool_dma_sync_for_device(pool, page,
416 						      dma_sync_size);
417 
418 		if (allow_direct && in_serving_softirq() &&
419 		    page_pool_recycle_in_cache(page, pool))
420 			return NULL;
421 
422 		/* Page found as candidate for recycling */
423 		return page;
424 	}
425 	/* Fallback/non-XDP mode: API user have elevated refcnt.
426 	 *
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.
431 	 *
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.
437 	 */
438 	/* Do not replace this with page_pool_return_page() */
439 	page_pool_release_page(pool, page);
440 	put_page(page);
441 
442 	return NULL;
443 }
444 
445 void page_pool_put_page(struct page_pool *pool, struct page *page,
446 			unsigned int dma_sync_size, bool allow_direct)
447 {
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);
452 	}
453 }
454 EXPORT_SYMBOL(page_pool_put_page);
455 
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,
458 			     int count)
459 {
460 	int i, bulk_len = 0;
461 
462 	for (i = 0; i < count; i++) {
463 		struct page *page = virt_to_head_page(data[i]);
464 
465 		page = __page_pool_put_page(pool, page, -1, false);
466 		/* Approved for bulk recycling in ptr_ring cache */
467 		if (page)
468 			data[bulk_len++] = page;
469 	}
470 
471 	if (unlikely(!bulk_len))
472 		return;
473 
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 */
479 	}
480 	page_pool_ring_unlock(pool);
481 
482 	/* Hopefully all pages was return into ptr_ring */
483 	if (likely(i == bulk_len))
484 		return;
485 
486 	/* ptr_ring cache full, free remaining pages outside producer lock
487 	 * since put_page() with refcnt == 1 can be an expensive operation
488 	 */
489 	for (; i < bulk_len; i++)
490 		page_pool_return_page(pool, data[i]);
491 }
492 EXPORT_SYMBOL(page_pool_put_page_bulk);
493 
494 static void page_pool_empty_ring(struct page_pool *pool)
495 {
496 	struct page *page;
497 
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));
504 
505 		page_pool_return_page(pool, page);
506 	}
507 }
508 
509 static void page_pool_free(struct page_pool *pool)
510 {
511 	if (pool->disconnect)
512 		pool->disconnect(pool);
513 
514 	ptr_ring_cleanup(&pool->ring, NULL);
515 
516 	if (pool->p.flags & PP_FLAG_DMA_MAP)
517 		put_device(pool->p.dev);
518 
519 	kfree(pool);
520 }
521 
522 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
523 {
524 	struct page *page;
525 
526 	if (pool->destroy_cnt)
527 		return;
528 
529 	/* Empty alloc cache, assume caller made sure this is
530 	 * no-longer in use, and page_pool_alloc_pages() cannot be
531 	 * call concurrently.
532 	 */
533 	while (pool->alloc.count) {
534 		page = pool->alloc.cache[--pool->alloc.count];
535 		page_pool_return_page(pool, page);
536 	}
537 }
538 
539 static void page_pool_scrub(struct page_pool *pool)
540 {
541 	page_pool_empty_alloc_cache_once(pool);
542 	pool->destroy_cnt++;
543 
544 	/* No more consumers should exist, but producers could still
545 	 * be in-flight.
546 	 */
547 	page_pool_empty_ring(pool);
548 }
549 
550 static int page_pool_release(struct page_pool *pool)
551 {
552 	int inflight;
553 
554 	page_pool_scrub(pool);
555 	inflight = page_pool_inflight(pool);
556 	if (!inflight)
557 		page_pool_free(pool);
558 
559 	return inflight;
560 }
561 
562 static void page_pool_release_retry(struct work_struct *wq)
563 {
564 	struct delayed_work *dwq = to_delayed_work(wq);
565 	struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
566 	int inflight;
567 
568 	inflight = page_pool_release(pool);
569 	if (!inflight)
570 		return;
571 
572 	/* Periodic warning */
573 	if (time_after_eq(jiffies, pool->defer_warn)) {
574 		int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
575 
576 		pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
577 			__func__, inflight, sec);
578 		pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
579 	}
580 
581 	/* Still not ready to be disconnected, retry later */
582 	schedule_delayed_work(&pool->release_dw, DEFER_TIME);
583 }
584 
585 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *))
586 {
587 	refcount_inc(&pool->user_cnt);
588 	pool->disconnect = disconnect;
589 }
590 
591 void page_pool_destroy(struct page_pool *pool)
592 {
593 	if (!pool)
594 		return;
595 
596 	if (!page_pool_put(pool))
597 		return;
598 
599 	if (!page_pool_release(pool))
600 		return;
601 
602 	pool->defer_start = jiffies;
603 	pool->defer_warn  = jiffies + DEFER_WARN_INTERVAL;
604 
605 	INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
606 	schedule_delayed_work(&pool->release_dw, DEFER_TIME);
607 }
608 EXPORT_SYMBOL(page_pool_destroy);
609 
610 /* Caller must provide appropriate safe context, e.g. NAPI. */
611 void page_pool_update_nid(struct page_pool *pool, int new_nid)
612 {
613 	struct page *page;
614 
615 	trace_page_pool_update_nid(pool, new_nid);
616 	pool->p.nid = new_nid;
617 
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);
622 	}
623 }
624 EXPORT_SYMBOL(page_pool_update_nid);
625