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