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