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 /* slow path */
186 noinline
187 static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
188 gfp_t _gfp)
189 {
190 struct page *page;
191 gfp_t gfp = _gfp;
192 dma_addr_t dma;
193
194 /* We could always set __GFP_COMP, and avoid this branch, as
195 * prep_new_page() can handle order-0 with __GFP_COMP.
196 */
197 if (pool->p.order)
198 gfp |= __GFP_COMP;
199
200 /* FUTURE development:
201 *
202 * Current slow-path essentially falls back to single page
203 * allocations, which doesn't improve performance. This code
204 * need bulk allocation support from the page allocator code.
205 */
206
207 /* Cache was empty, do real allocation */
208 #ifdef CONFIG_NUMA
209 page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
210 #else
211 page = alloc_pages(gfp, pool->p.order);
212 #endif
213 if (!page)
214 return NULL;
215
216 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
217 goto skip_dma_map;
218
219 /* Setup DMA mapping: use 'struct page' area for storing DMA-addr
220 * since dma_addr_t can be either 32 or 64 bits and does not always fit
221 * into page private data (i.e 32bit cpu with 64bit DMA caps)
222 * This mapping is kept for lifetime of page, until leaving pool.
223 */
224 dma = dma_map_page_attrs(pool->p.dev, page, 0,
225 (PAGE_SIZE << pool->p.order),
226 pool->p.dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
227 if (dma_mapping_error(pool->p.dev, dma)) {
228 put_page(page);
229 return NULL;
230 }
231 page_pool_set_dma_addr(page, dma);
232
233 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
234 page_pool_dma_sync_for_device(pool, page, pool->p.max_len);
235
236 skip_dma_map:
237 /* Track how many pages are held 'in-flight' */
238 pool->pages_state_hold_cnt++;
239
240 trace_page_pool_state_hold(pool, page, pool->pages_state_hold_cnt);
241
242 /* When page just alloc'ed is should/must have refcnt 1. */
243 return page;
244 }
245
246 /* For using page_pool replace: alloc_pages() API calls, but provide
247 * synchronization guarantee for allocation side.
248 */
249 struct page *page_pool_alloc_pages(struct page_pool *pool, gfp_t gfp)
250 {
251 struct page *page;
252
253 /* Fast-path: Get a page from cache */
254 page = __page_pool_get_cached(pool);
255 if (page)
256 return page;
257
258 /* Slow-path: cache empty, do real allocation */
259 page = __page_pool_alloc_pages_slow(pool, gfp);
260 return page;
261 }
262 EXPORT_SYMBOL(page_pool_alloc_pages);
263
264 /* Calculate distance between two u32 values, valid if distance is below 2^(31)
265 * https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution
266 */
267 #define _distance(a, b) (s32)((a) - (b))
268
269 static s32 page_pool_inflight(struct page_pool *pool)
270 {
271 u32 release_cnt = atomic_read(&pool->pages_state_release_cnt);
272 u32 hold_cnt = READ_ONCE(pool->pages_state_hold_cnt);
273 s32 inflight;
274
275 inflight = _distance(hold_cnt, release_cnt);
276
277 trace_page_pool_release(pool, inflight, hold_cnt, release_cnt);
278 WARN(inflight < 0, "Negative(%d) inflight packet-pages", inflight);
279
280 return inflight;
281 }
282
283 /* Disconnects a page (from a page_pool). API users can have a need
284 * to disconnect a page (from a page_pool), to allow it to be used as
285 * a regular page (that will eventually be returned to the normal
286 * page-allocator via put_page).
287 */
288 void page_pool_release_page(struct page_pool *pool, struct page *page)
289 {
290 dma_addr_t dma;
291 int count;
292
293 if (!(pool->p.flags & PP_FLAG_DMA_MAP))
294 /* Always account for inflight pages, even if we didn't
295 * map them
296 */
297 goto skip_dma_unmap;
298
299 dma = page_pool_get_dma_addr(page);
300
301 /* When page is unmapped, it cannot be returned to our pool */
302 dma_unmap_page_attrs(pool->p.dev, dma,
303 PAGE_SIZE << pool->p.order, pool->p.dma_dir,
304 DMA_ATTR_SKIP_CPU_SYNC);
305 page_pool_set_dma_addr(page, 0);
306 skip_dma_unmap:
307 /* This may be the last page returned, releasing the pool, so
308 * it is not safe to reference pool afterwards.
309 */
310 count = atomic_inc_return(&pool->pages_state_release_cnt);
311 trace_page_pool_state_release(pool, page, count);
312 }
313 EXPORT_SYMBOL(page_pool_release_page);
314
315 /* Return a page to the page allocator, cleaning up our state */
316 static void page_pool_return_page(struct page_pool *pool, struct page *page)
317 {
318 page_pool_release_page(pool, page);
319
320 put_page(page);
321 /* An optimization would be to call __free_pages(page, pool->p.order)
322 * knowing page is not part of page-cache (thus avoiding a
323 * __page_cache_release() call).
324 */
325 }
326
327 static bool page_pool_recycle_in_ring(struct page_pool *pool, struct page *page)
328 {
329 int ret;
330 /* BH protection not needed if current is serving softirq */
331 if (in_serving_softirq())
332 ret = ptr_ring_produce(&pool->ring, page);
333 else
334 ret = ptr_ring_produce_bh(&pool->ring, page);
335
336 return (ret == 0) ? true : false;
337 }
338
339 /* Only allow direct recycling in special circumstances, into the
340 * alloc side cache. E.g. during RX-NAPI processing for XDP_DROP use-case.
341 *
342 * Caller must provide appropriate safe context.
343 */
344 static bool page_pool_recycle_in_cache(struct page *page,
345 struct page_pool *pool)
346 {
347 if (unlikely(pool->alloc.count == PP_ALLOC_CACHE_SIZE))
348 return false;
349
350 /* Caller MUST have verified/know (page_ref_count(page) == 1) */
351 pool->alloc.cache[pool->alloc.count++] = page;
352 return true;
353 }
354
355 /* If the page refcnt == 1, this will try to recycle the page.
356 * if PP_FLAG_DMA_SYNC_DEV is set, we'll try to sync the DMA area for
357 * the configured size min(dma_sync_size, pool->max_len).
358 * If the page refcnt != 1, then the page will be returned to memory
359 * subsystem.
360 */
361 static __always_inline struct page *
362 __page_pool_put_page(struct page_pool *pool, struct page *page,
363 unsigned int dma_sync_size, bool allow_direct)
364 {
365 /* This allocator is optimized for the XDP mode that uses
366 * one-frame-per-page, but have fallbacks that act like the
367 * regular page allocator APIs.
368 *
369 * refcnt == 1 means page_pool owns page, and can recycle it.
370 *
371 * page is NOT reusable when allocated when system is under
372 * some pressure. (page_is_pfmemalloc)
373 */
374 if (likely(page_ref_count(page) == 1 && !page_is_pfmemalloc(page))) {
375 /* Read barrier done in page_ref_count / READ_ONCE */
376
377 if (pool->p.flags & PP_FLAG_DMA_SYNC_DEV)
378 page_pool_dma_sync_for_device(pool, page,
379 dma_sync_size);
380
381 if (allow_direct && in_serving_softirq() &&
382 page_pool_recycle_in_cache(page, pool))
383 return NULL;
384
385 /* Page found as candidate for recycling */
386 return page;
387 }
388 /* Fallback/non-XDP mode: API user have elevated refcnt.
389 *
390 * Many drivers split up the page into fragments, and some
391 * want to keep doing this to save memory and do refcnt based
392 * recycling. Support this use case too, to ease drivers
393 * switching between XDP/non-XDP.
394 *
395 * In-case page_pool maintains the DMA mapping, API user must
396 * call page_pool_put_page once. In this elevated refcnt
397 * case, the DMA is unmapped/released, as driver is likely
398 * doing refcnt based recycle tricks, meaning another process
399 * will be invoking put_page.
400 */
401 /* Do not replace this with page_pool_return_page() */
402 page_pool_release_page(pool, page);
403 put_page(page);
404
405 return NULL;
406 }
407
408 void page_pool_put_page(struct page_pool *pool, struct page *page,
409 unsigned int dma_sync_size, bool allow_direct)
410 {
411 page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
412 if (page && !page_pool_recycle_in_ring(pool, page)) {
413 /* Cache full, fallback to free pages */
414 page_pool_return_page(pool, page);
415 }
416 }
417 EXPORT_SYMBOL(page_pool_put_page);
418
419 /* Caller must not use data area after call, as this function overwrites it */
420 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
421 int count)
422 {
423 int i, bulk_len = 0;
424
425 for (i = 0; i < count; i++) {
426 struct page *page = virt_to_head_page(data[i]);
427
428 page = __page_pool_put_page(pool, page, -1, false);
429 /* Approved for bulk recycling in ptr_ring cache */
430 if (page)
431 data[bulk_len++] = page;
432 }
433
434 if (unlikely(!bulk_len))
435 return;
436
437 /* Bulk producer into ptr_ring page_pool cache */
438 page_pool_ring_lock(pool);
439 for (i = 0; i < bulk_len; i++) {
440 if (__ptr_ring_produce(&pool->ring, data[i]))
441 break; /* ring full */
442 }
443 page_pool_ring_unlock(pool);
444
445 /* Hopefully all pages was return into ptr_ring */
446 if (likely(i == bulk_len))
447 return;
448
449 /* ptr_ring cache full, free remaining pages outside producer lock
450 * since put_page() with refcnt == 1 can be an expensive operation
451 */
452 for (; i < bulk_len; i++)
453 page_pool_return_page(pool, data[i]);
454 }
455 EXPORT_SYMBOL(page_pool_put_page_bulk);
456
457 static void page_pool_empty_ring(struct page_pool *pool)
458 {
459 struct page *page;
460
461 /* Empty recycle ring */
462 while ((page = ptr_ring_consume_bh(&pool->ring))) {
463 /* Verify the refcnt invariant of cached pages */
464 if (!(page_ref_count(page) == 1))
465 pr_crit("%s() page_pool refcnt %d violation\n",
466 __func__, page_ref_count(page));
467
468 page_pool_return_page(pool, page);
469 }
470 }
471
472 static void page_pool_free(struct page_pool *pool)
473 {
474 if (pool->disconnect)
475 pool->disconnect(pool);
476
477 ptr_ring_cleanup(&pool->ring, NULL);
478
479 if (pool->p.flags & PP_FLAG_DMA_MAP)
480 put_device(pool->p.dev);
481
482 kfree(pool);
483 }
484
485 static void page_pool_empty_alloc_cache_once(struct page_pool *pool)
486 {
487 struct page *page;
488
489 if (pool->destroy_cnt)
490 return;
491
492 /* Empty alloc cache, assume caller made sure this is
493 * no-longer in use, and page_pool_alloc_pages() cannot be
494 * call concurrently.
495 */
496 while (pool->alloc.count) {
497 page = pool->alloc.cache[--pool->alloc.count];
498 page_pool_return_page(pool, page);
499 }
500 }
501
502 static void page_pool_scrub(struct page_pool *pool)
503 {
504 page_pool_empty_alloc_cache_once(pool);
505 pool->destroy_cnt++;
506
507 /* No more consumers should exist, but producers could still
508 * be in-flight.
509 */
510 page_pool_empty_ring(pool);
511 }
512
513 static int page_pool_release(struct page_pool *pool)
514 {
515 int inflight;
516
517 page_pool_scrub(pool);
518 inflight = page_pool_inflight(pool);
519 if (!inflight)
520 page_pool_free(pool);
521
522 return inflight;
523 }
524
525 static void page_pool_release_retry(struct work_struct *wq)
526 {
527 struct delayed_work *dwq = to_delayed_work(wq);
528 struct page_pool *pool = container_of(dwq, typeof(*pool), release_dw);
529 int inflight;
530
531 inflight = page_pool_release(pool);
532 if (!inflight)
533 return;
534
535 /* Periodic warning */
536 if (time_after_eq(jiffies, pool->defer_warn)) {
537 int sec = (s32)((u32)jiffies - (u32)pool->defer_start) / HZ;
538
539 pr_warn("%s() stalled pool shutdown %d inflight %d sec\n",
540 __func__, inflight, sec);
541 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
542 }
543
544 /* Still not ready to be disconnected, retry later */
545 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
546 }
547
548 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *))
549 {
550 refcount_inc(&pool->user_cnt);
551 pool->disconnect = disconnect;
552 }
553
554 void page_pool_destroy(struct page_pool *pool)
555 {
556 if (!pool)
557 return;
558
559 if (!page_pool_put(pool))
560 return;
561
562 if (!page_pool_release(pool))
563 return;
564
565 pool->defer_start = jiffies;
566 pool->defer_warn = jiffies + DEFER_WARN_INTERVAL;
567
568 INIT_DELAYED_WORK(&pool->release_dw, page_pool_release_retry);
569 schedule_delayed_work(&pool->release_dw, DEFER_TIME);
570 }
571 EXPORT_SYMBOL(page_pool_destroy);
572
573 /* Caller must provide appropriate safe context, e.g. NAPI. */
574 void page_pool_update_nid(struct page_pool *pool, int new_nid)
575 {
576 struct page *page;
577
578 trace_page_pool_update_nid(pool, new_nid);
579 pool->p.nid = new_nid;
580
581 /* Flush pool alloc cache, as refill will check NUMA node */
582 while (pool->alloc.count) {
583 page = pool->alloc.cache[--pool->alloc.count];
584 page_pool_return_page(pool, page);
585 }
586 }
587 EXPORT_SYMBOL(page_pool_update_nid);
588
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.