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