Merge tag 'leds-5.19-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/pavel...
[linux-2.6-microblaze.git] / mm / page_owner.c
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/debugfs.h>
3 #include <linux/mm.h>
4 #include <linux/slab.h>
5 #include <linux/uaccess.h>
6 #include <linux/memblock.h>
7 #include <linux/stacktrace.h>
8 #include <linux/page_owner.h>
9 #include <linux/jump_label.h>
10 #include <linux/migrate.h>
11 #include <linux/stackdepot.h>
12 #include <linux/seq_file.h>
13 #include <linux/memcontrol.h>
14 #include <linux/sched/clock.h>
15
16 #include "internal.h"
17
18 /*
19  * TODO: teach PAGE_OWNER_STACK_DEPTH (__dump_page_owner and save_stack)
20  * to use off stack temporal storage
21  */
22 #define PAGE_OWNER_STACK_DEPTH (16)
23
24 struct page_owner {
25         unsigned short order;
26         short last_migrate_reason;
27         gfp_t gfp_mask;
28         depot_stack_handle_t handle;
29         depot_stack_handle_t free_handle;
30         u64 ts_nsec;
31         u64 free_ts_nsec;
32         char comm[TASK_COMM_LEN];
33         pid_t pid;
34         pid_t tgid;
35 };
36
37 static bool page_owner_enabled __initdata;
38 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
39
40 static depot_stack_handle_t dummy_handle;
41 static depot_stack_handle_t failure_handle;
42 static depot_stack_handle_t early_handle;
43
44 static void init_early_allocated_pages(void);
45
46 static int __init early_page_owner_param(char *buf)
47 {
48         int ret = kstrtobool(buf, &page_owner_enabled);
49
50         if (page_owner_enabled)
51                 stack_depot_want_early_init();
52
53         return ret;
54 }
55 early_param("page_owner", early_page_owner_param);
56
57 static __init bool need_page_owner(void)
58 {
59         return page_owner_enabled;
60 }
61
62 static __always_inline depot_stack_handle_t create_dummy_stack(void)
63 {
64         unsigned long entries[4];
65         unsigned int nr_entries;
66
67         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
68         return stack_depot_save(entries, nr_entries, GFP_KERNEL);
69 }
70
71 static noinline void register_dummy_stack(void)
72 {
73         dummy_handle = create_dummy_stack();
74 }
75
76 static noinline void register_failure_stack(void)
77 {
78         failure_handle = create_dummy_stack();
79 }
80
81 static noinline void register_early_stack(void)
82 {
83         early_handle = create_dummy_stack();
84 }
85
86 static __init void init_page_owner(void)
87 {
88         if (!page_owner_enabled)
89                 return;
90
91         register_dummy_stack();
92         register_failure_stack();
93         register_early_stack();
94         static_branch_enable(&page_owner_inited);
95         init_early_allocated_pages();
96 }
97
98 struct page_ext_operations page_owner_ops = {
99         .size = sizeof(struct page_owner),
100         .need = need_page_owner,
101         .init = init_page_owner,
102 };
103
104 static inline struct page_owner *get_page_owner(struct page_ext *page_ext)
105 {
106         return (void *)page_ext + page_owner_ops.offset;
107 }
108
109 static noinline depot_stack_handle_t save_stack(gfp_t flags)
110 {
111         unsigned long entries[PAGE_OWNER_STACK_DEPTH];
112         depot_stack_handle_t handle;
113         unsigned int nr_entries;
114
115         /*
116          * Avoid recursion.
117          *
118          * Sometimes page metadata allocation tracking requires more
119          * memory to be allocated:
120          * - when new stack trace is saved to stack depot
121          * - when backtrace itself is calculated (ia64)
122          */
123         if (current->in_page_owner)
124                 return dummy_handle;
125         current->in_page_owner = 1;
126
127         nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2);
128         handle = stack_depot_save(entries, nr_entries, flags);
129         if (!handle)
130                 handle = failure_handle;
131
132         current->in_page_owner = 0;
133         return handle;
134 }
135
136 void __reset_page_owner(struct page *page, unsigned short order)
137 {
138         int i;
139         struct page_ext *page_ext;
140         depot_stack_handle_t handle;
141         struct page_owner *page_owner;
142         u64 free_ts_nsec = local_clock();
143
144         page_ext = lookup_page_ext(page);
145         if (unlikely(!page_ext))
146                 return;
147
148         handle = save_stack(GFP_NOWAIT | __GFP_NOWARN);
149         for (i = 0; i < (1 << order); i++) {
150                 __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
151                 page_owner = get_page_owner(page_ext);
152                 page_owner->free_handle = handle;
153                 page_owner->free_ts_nsec = free_ts_nsec;
154                 page_ext = page_ext_next(page_ext);
155         }
156 }
157
158 static inline void __set_page_owner_handle(struct page_ext *page_ext,
159                                         depot_stack_handle_t handle,
160                                         unsigned short order, gfp_t gfp_mask)
161 {
162         struct page_owner *page_owner;
163         int i;
164
165         for (i = 0; i < (1 << order); i++) {
166                 page_owner = get_page_owner(page_ext);
167                 page_owner->handle = handle;
168                 page_owner->order = order;
169                 page_owner->gfp_mask = gfp_mask;
170                 page_owner->last_migrate_reason = -1;
171                 page_owner->pid = current->pid;
172                 page_owner->tgid = current->tgid;
173                 page_owner->ts_nsec = local_clock();
174                 strscpy(page_owner->comm, current->comm,
175                         sizeof(page_owner->comm));
176                 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
177                 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
178
179                 page_ext = page_ext_next(page_ext);
180         }
181 }
182
183 noinline void __set_page_owner(struct page *page, unsigned short order,
184                                         gfp_t gfp_mask)
185 {
186         struct page_ext *page_ext = lookup_page_ext(page);
187         depot_stack_handle_t handle;
188
189         if (unlikely(!page_ext))
190                 return;
191
192         handle = save_stack(gfp_mask);
193         __set_page_owner_handle(page_ext, handle, order, gfp_mask);
194 }
195
196 void __set_page_owner_migrate_reason(struct page *page, int reason)
197 {
198         struct page_ext *page_ext = lookup_page_ext(page);
199         struct page_owner *page_owner;
200
201         if (unlikely(!page_ext))
202                 return;
203
204         page_owner = get_page_owner(page_ext);
205         page_owner->last_migrate_reason = reason;
206 }
207
208 void __split_page_owner(struct page *page, unsigned int nr)
209 {
210         int i;
211         struct page_ext *page_ext = lookup_page_ext(page);
212         struct page_owner *page_owner;
213
214         if (unlikely(!page_ext))
215                 return;
216
217         for (i = 0; i < nr; i++) {
218                 page_owner = get_page_owner(page_ext);
219                 page_owner->order = 0;
220                 page_ext = page_ext_next(page_ext);
221         }
222 }
223
224 void __folio_copy_owner(struct folio *newfolio, struct folio *old)
225 {
226         struct page_ext *old_ext = lookup_page_ext(&old->page);
227         struct page_ext *new_ext = lookup_page_ext(&newfolio->page);
228         struct page_owner *old_page_owner, *new_page_owner;
229
230         if (unlikely(!old_ext || !new_ext))
231                 return;
232
233         old_page_owner = get_page_owner(old_ext);
234         new_page_owner = get_page_owner(new_ext);
235         new_page_owner->order = old_page_owner->order;
236         new_page_owner->gfp_mask = old_page_owner->gfp_mask;
237         new_page_owner->last_migrate_reason =
238                 old_page_owner->last_migrate_reason;
239         new_page_owner->handle = old_page_owner->handle;
240         new_page_owner->pid = old_page_owner->pid;
241         new_page_owner->tgid = old_page_owner->tgid;
242         new_page_owner->ts_nsec = old_page_owner->ts_nsec;
243         new_page_owner->free_ts_nsec = old_page_owner->ts_nsec;
244         strcpy(new_page_owner->comm, old_page_owner->comm);
245
246         /*
247          * We don't clear the bit on the old folio as it's going to be freed
248          * after migration. Until then, the info can be useful in case of
249          * a bug, and the overall stats will be off a bit only temporarily.
250          * Also, migrate_misplaced_transhuge_page() can still fail the
251          * migration and then we want the old folio to retain the info. But
252          * in that case we also don't need to explicitly clear the info from
253          * the new page, which will be freed.
254          */
255         __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
256         __set_bit(PAGE_EXT_OWNER_ALLOCATED, &new_ext->flags);
257 }
258
259 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
260                                        pg_data_t *pgdat, struct zone *zone)
261 {
262         struct page *page;
263         struct page_ext *page_ext;
264         struct page_owner *page_owner;
265         unsigned long pfn, block_end_pfn;
266         unsigned long end_pfn = zone_end_pfn(zone);
267         unsigned long count[MIGRATE_TYPES] = { 0, };
268         int pageblock_mt, page_mt;
269         int i;
270
271         /* Scan block by block. First and last block may be incomplete */
272         pfn = zone->zone_start_pfn;
273
274         /*
275          * Walk the zone in pageblock_nr_pages steps. If a page block spans
276          * a zone boundary, it will be double counted between zones. This does
277          * not matter as the mixed block count will still be correct
278          */
279         for (; pfn < end_pfn; ) {
280                 page = pfn_to_online_page(pfn);
281                 if (!page) {
282                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
283                         continue;
284                 }
285
286                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
287                 block_end_pfn = min(block_end_pfn, end_pfn);
288
289                 pageblock_mt = get_pageblock_migratetype(page);
290
291                 for (; pfn < block_end_pfn; pfn++) {
292                         /* The pageblock is online, no need to recheck. */
293                         page = pfn_to_page(pfn);
294
295                         if (page_zone(page) != zone)
296                                 continue;
297
298                         if (PageBuddy(page)) {
299                                 unsigned long freepage_order;
300
301                                 freepage_order = buddy_order_unsafe(page);
302                                 if (freepage_order < MAX_ORDER)
303                                         pfn += (1UL << freepage_order) - 1;
304                                 continue;
305                         }
306
307                         if (PageReserved(page))
308                                 continue;
309
310                         page_ext = lookup_page_ext(page);
311                         if (unlikely(!page_ext))
312                                 continue;
313
314                         if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
315                                 continue;
316
317                         page_owner = get_page_owner(page_ext);
318                         page_mt = gfp_migratetype(page_owner->gfp_mask);
319                         if (pageblock_mt != page_mt) {
320                                 if (is_migrate_cma(pageblock_mt))
321                                         count[MIGRATE_MOVABLE]++;
322                                 else
323                                         count[pageblock_mt]++;
324
325                                 pfn = block_end_pfn;
326                                 break;
327                         }
328                         pfn += (1UL << page_owner->order) - 1;
329                 }
330         }
331
332         /* Print counts */
333         seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
334         for (i = 0; i < MIGRATE_TYPES; i++)
335                 seq_printf(m, "%12lu ", count[i]);
336         seq_putc(m, '\n');
337 }
338
339 /*
340  * Looking for memcg information and print it out
341  */
342 static inline int print_page_owner_memcg(char *kbuf, size_t count, int ret,
343                                          struct page *page)
344 {
345 #ifdef CONFIG_MEMCG
346         unsigned long memcg_data;
347         struct mem_cgroup *memcg;
348         bool online;
349         char name[80];
350
351         rcu_read_lock();
352         memcg_data = READ_ONCE(page->memcg_data);
353         if (!memcg_data)
354                 goto out_unlock;
355
356         if (memcg_data & MEMCG_DATA_OBJCGS)
357                 ret += scnprintf(kbuf + ret, count - ret,
358                                 "Slab cache page\n");
359
360         memcg = page_memcg_check(page);
361         if (!memcg)
362                 goto out_unlock;
363
364         online = (memcg->css.flags & CSS_ONLINE);
365         cgroup_name(memcg->css.cgroup, name, sizeof(name));
366         ret += scnprintf(kbuf + ret, count - ret,
367                         "Charged %sto %smemcg %s\n",
368                         PageMemcgKmem(page) ? "(via objcg) " : "",
369                         online ? "" : "offline ",
370                         name);
371 out_unlock:
372         rcu_read_unlock();
373 #endif /* CONFIG_MEMCG */
374
375         return ret;
376 }
377
378 static ssize_t
379 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
380                 struct page *page, struct page_owner *page_owner,
381                 depot_stack_handle_t handle)
382 {
383         int ret, pageblock_mt, page_mt;
384         char *kbuf;
385
386         count = min_t(size_t, count, PAGE_SIZE);
387         kbuf = kmalloc(count, GFP_KERNEL);
388         if (!kbuf)
389                 return -ENOMEM;
390
391         ret = scnprintf(kbuf, count,
392                         "Page allocated via order %u, mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu ns, free_ts %llu ns\n",
393                         page_owner->order, page_owner->gfp_mask,
394                         &page_owner->gfp_mask, page_owner->pid,
395                         page_owner->tgid, page_owner->comm,
396                         page_owner->ts_nsec, page_owner->free_ts_nsec);
397
398         /* Print information relevant to grouping pages by mobility */
399         pageblock_mt = get_pageblock_migratetype(page);
400         page_mt  = gfp_migratetype(page_owner->gfp_mask);
401         ret += scnprintf(kbuf + ret, count - ret,
402                         "PFN %lu type %s Block %lu type %s Flags %pGp\n",
403                         pfn,
404                         migratetype_names[page_mt],
405                         pfn >> pageblock_order,
406                         migratetype_names[pageblock_mt],
407                         &page->flags);
408
409         ret += stack_depot_snprint(handle, kbuf + ret, count - ret, 0);
410         if (ret >= count)
411                 goto err;
412
413         if (page_owner->last_migrate_reason != -1) {
414                 ret += scnprintf(kbuf + ret, count - ret,
415                         "Page has been migrated, last migrate reason: %s\n",
416                         migrate_reason_names[page_owner->last_migrate_reason]);
417         }
418
419         ret = print_page_owner_memcg(kbuf, count, ret, page);
420
421         ret += snprintf(kbuf + ret, count - ret, "\n");
422         if (ret >= count)
423                 goto err;
424
425         if (copy_to_user(buf, kbuf, ret))
426                 ret = -EFAULT;
427
428         kfree(kbuf);
429         return ret;
430
431 err:
432         kfree(kbuf);
433         return -ENOMEM;
434 }
435
436 void __dump_page_owner(const struct page *page)
437 {
438         struct page_ext *page_ext = lookup_page_ext(page);
439         struct page_owner *page_owner;
440         depot_stack_handle_t handle;
441         gfp_t gfp_mask;
442         int mt;
443
444         if (unlikely(!page_ext)) {
445                 pr_alert("There is not page extension available.\n");
446                 return;
447         }
448
449         page_owner = get_page_owner(page_ext);
450         gfp_mask = page_owner->gfp_mask;
451         mt = gfp_migratetype(gfp_mask);
452
453         if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
454                 pr_alert("page_owner info is not present (never set?)\n");
455                 return;
456         }
457
458         if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
459                 pr_alert("page_owner tracks the page as allocated\n");
460         else
461                 pr_alert("page_owner tracks the page as freed\n");
462
463         pr_alert("page last allocated via order %u, migratetype %s, gfp_mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu, free_ts %llu\n",
464                  page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
465                  page_owner->pid, page_owner->tgid, page_owner->comm,
466                  page_owner->ts_nsec, page_owner->free_ts_nsec);
467
468         handle = READ_ONCE(page_owner->handle);
469         if (!handle)
470                 pr_alert("page_owner allocation stack trace missing\n");
471         else
472                 stack_depot_print(handle);
473
474         handle = READ_ONCE(page_owner->free_handle);
475         if (!handle) {
476                 pr_alert("page_owner free stack trace missing\n");
477         } else {
478                 pr_alert("page last free stack trace:\n");
479                 stack_depot_print(handle);
480         }
481
482         if (page_owner->last_migrate_reason != -1)
483                 pr_alert("page has been migrated, last migrate reason: %s\n",
484                         migrate_reason_names[page_owner->last_migrate_reason]);
485 }
486
487 static ssize_t
488 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
489 {
490         unsigned long pfn;
491         struct page *page;
492         struct page_ext *page_ext;
493         struct page_owner *page_owner;
494         depot_stack_handle_t handle;
495
496         if (!static_branch_unlikely(&page_owner_inited))
497                 return -EINVAL;
498
499         page = NULL;
500         pfn = min_low_pfn + *ppos;
501
502         /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
503         while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
504                 pfn++;
505
506         drain_all_pages(NULL);
507
508         /* Find an allocated page */
509         for (; pfn < max_pfn; pfn++) {
510                 /*
511                  * If the new page is in a new MAX_ORDER_NR_PAGES area,
512                  * validate the area as existing, skip it if not
513                  */
514                 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
515                         pfn += MAX_ORDER_NR_PAGES - 1;
516                         continue;
517                 }
518
519                 page = pfn_to_page(pfn);
520                 if (PageBuddy(page)) {
521                         unsigned long freepage_order = buddy_order_unsafe(page);
522
523                         if (freepage_order < MAX_ORDER)
524                                 pfn += (1UL << freepage_order) - 1;
525                         continue;
526                 }
527
528                 page_ext = lookup_page_ext(page);
529                 if (unlikely(!page_ext))
530                         continue;
531
532                 /*
533                  * Some pages could be missed by concurrent allocation or free,
534                  * because we don't hold the zone lock.
535                  */
536                 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
537                         continue;
538
539                 /*
540                  * Although we do have the info about past allocation of free
541                  * pages, it's not relevant for current memory usage.
542                  */
543                 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
544                         continue;
545
546                 page_owner = get_page_owner(page_ext);
547
548                 /*
549                  * Don't print "tail" pages of high-order allocations as that
550                  * would inflate the stats.
551                  */
552                 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
553                         continue;
554
555                 /*
556                  * Access to page_ext->handle isn't synchronous so we should
557                  * be careful to access it.
558                  */
559                 handle = READ_ONCE(page_owner->handle);
560                 if (!handle)
561                         continue;
562
563                 /* Record the next PFN to read in the file offset */
564                 *ppos = (pfn - min_low_pfn) + 1;
565
566                 return print_page_owner(buf, count, pfn, page,
567                                 page_owner, handle);
568         }
569
570         return 0;
571 }
572
573 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
574 {
575         unsigned long pfn = zone->zone_start_pfn;
576         unsigned long end_pfn = zone_end_pfn(zone);
577         unsigned long count = 0;
578
579         /*
580          * Walk the zone in pageblock_nr_pages steps. If a page block spans
581          * a zone boundary, it will be double counted between zones. This does
582          * not matter as the mixed block count will still be correct
583          */
584         for (; pfn < end_pfn; ) {
585                 unsigned long block_end_pfn;
586
587                 if (!pfn_valid(pfn)) {
588                         pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
589                         continue;
590                 }
591
592                 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
593                 block_end_pfn = min(block_end_pfn, end_pfn);
594
595                 for (; pfn < block_end_pfn; pfn++) {
596                         struct page *page = pfn_to_page(pfn);
597                         struct page_ext *page_ext;
598
599                         if (page_zone(page) != zone)
600                                 continue;
601
602                         /*
603                          * To avoid having to grab zone->lock, be a little
604                          * careful when reading buddy page order. The only
605                          * danger is that we skip too much and potentially miss
606                          * some early allocated pages, which is better than
607                          * heavy lock contention.
608                          */
609                         if (PageBuddy(page)) {
610                                 unsigned long order = buddy_order_unsafe(page);
611
612                                 if (order > 0 && order < MAX_ORDER)
613                                         pfn += (1UL << order) - 1;
614                                 continue;
615                         }
616
617                         if (PageReserved(page))
618                                 continue;
619
620                         page_ext = lookup_page_ext(page);
621                         if (unlikely(!page_ext))
622                                 continue;
623
624                         /* Maybe overlapping zone */
625                         if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
626                                 continue;
627
628                         /* Found early allocated page */
629                         __set_page_owner_handle(page_ext, early_handle,
630                                                 0, 0);
631                         count++;
632                 }
633                 cond_resched();
634         }
635
636         pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
637                 pgdat->node_id, zone->name, count);
638 }
639
640 static void init_zones_in_node(pg_data_t *pgdat)
641 {
642         struct zone *zone;
643         struct zone *node_zones = pgdat->node_zones;
644
645         for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
646                 if (!populated_zone(zone))
647                         continue;
648
649                 init_pages_in_zone(pgdat, zone);
650         }
651 }
652
653 static void init_early_allocated_pages(void)
654 {
655         pg_data_t *pgdat;
656
657         for_each_online_pgdat(pgdat)
658                 init_zones_in_node(pgdat);
659 }
660
661 static const struct file_operations proc_page_owner_operations = {
662         .read           = read_page_owner,
663 };
664
665 static int __init pageowner_init(void)
666 {
667         if (!static_branch_unlikely(&page_owner_inited)) {
668                 pr_info("page_owner is disabled\n");
669                 return 0;
670         }
671
672         debugfs_create_file("page_owner", 0400, NULL, NULL,
673                             &proc_page_owner_operations);
674
675         return 0;
676 }
677 late_initcall(pageowner_init)