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 return kstrtobool(buf, &page_owner_enabled);
46 }
47 early_param("page_owner", early_page_owner_param);
48
need_page_owner(void)49 static bool need_page_owner(void)
50 {
51 return page_owner_enabled;
52 }
53
create_dummy_stack(void)54 static __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
register_dummy_stack(void)63 static noinline void register_dummy_stack(void)
64 {
65 dummy_handle = create_dummy_stack();
66 }
67
register_failure_stack(void)68 static noinline void register_failure_stack(void)
69 {
70 failure_handle = create_dummy_stack();
71 }
72
register_early_stack(void)73 static noinline void register_early_stack(void)
74 {
75 early_handle = create_dummy_stack();
76 }
77
init_page_owner(void)78 static 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
90 struct page_ext_operations page_owner_ops = {
91 .size = sizeof(struct page_owner),
92 .need = need_page_owner,
93 .init = init_page_owner,
94 };
95
get_page_owner(struct page_ext * page_ext)96 static inline struct page_owner *get_page_owner(struct page_ext *page_ext)
97 {
98 return (void *)page_ext + page_owner_ops.offset;
99 }
100
save_stack(gfp_t flags)101 static 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
__reset_page_owner(struct page * page,unsigned int order)128 void __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
__set_page_owner_handle(struct page_ext * page_ext,depot_stack_handle_t handle,unsigned int order,gfp_t gfp_mask)150 static 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
__set_page_owner(struct page * page,unsigned int order,gfp_t gfp_mask)172 noinline 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
__set_page_owner_migrate_reason(struct page * page,int reason)185 void __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
__split_page_owner(struct page * page,unsigned int nr)197 void __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
__copy_page_owner(struct page * oldpage,struct page * newpage)213 void __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
pagetypeinfo_showmixedcount_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)246 void 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 /* The pageblock is online, no need to recheck. */
280 page = pfn_to_page(pfn);
281
282 if (page_zone(page) != zone)
283 continue;
284
285 if (PageBuddy(page)) {
286 unsigned long freepage_order;
287
288 freepage_order = buddy_order_unsafe(page);
289 if (freepage_order < MAX_ORDER)
290 pfn += (1UL << freepage_order) - 1;
291 continue;
292 }
293
294 if (PageReserved(page))
295 continue;
296
297 page_ext = lookup_page_ext(page);
298 if (unlikely(!page_ext))
299 continue;
300
301 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
302 continue;
303
304 page_owner = get_page_owner(page_ext);
305 page_mt = gfp_migratetype(page_owner->gfp_mask);
306 if (pageblock_mt != page_mt) {
307 if (is_migrate_cma(pageblock_mt))
308 count[MIGRATE_MOVABLE]++;
309 else
310 count[pageblock_mt]++;
311
312 pfn = block_end_pfn;
313 break;
314 }
315 pfn += (1UL << page_owner->order) - 1;
316 }
317 }
318
319 /* Print counts */
320 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
321 for (i = 0; i < MIGRATE_TYPES; i++)
322 seq_printf(m, "%12lu ", count[i]);
323 seq_putc(m, '\n');
324 }
325
326 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)327 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
328 struct page *page, struct page_owner *page_owner,
329 depot_stack_handle_t handle)
330 {
331 int ret, pageblock_mt, page_mt;
332 unsigned long *entries;
333 unsigned int nr_entries;
334 char *kbuf;
335
336 count = min_t(size_t, count, PAGE_SIZE);
337 kbuf = kmalloc(count, GFP_KERNEL);
338 if (!kbuf)
339 return -ENOMEM;
340
341 ret = snprintf(kbuf, count,
342 "Page allocated via order %u, mask %#x(%pGg), pid %d, ts %llu ns, free_ts %llu ns\n",
343 page_owner->order, page_owner->gfp_mask,
344 &page_owner->gfp_mask, page_owner->pid,
345 page_owner->ts_nsec, page_owner->free_ts_nsec);
346
347 if (ret >= count)
348 goto err;
349
350 /* Print information relevant to grouping pages by mobility */
351 pageblock_mt = get_pageblock_migratetype(page);
352 page_mt = gfp_migratetype(page_owner->gfp_mask);
353 ret += snprintf(kbuf + ret, count - ret,
354 "PFN %lu type %s Block %lu type %s Flags %#lx(%pGp)\n",
355 pfn,
356 migratetype_names[page_mt],
357 pfn >> pageblock_order,
358 migratetype_names[pageblock_mt],
359 page->flags, &page->flags);
360
361 if (ret >= count)
362 goto err;
363
364 nr_entries = stack_depot_fetch(handle, &entries);
365 ret += stack_trace_snprint(kbuf + ret, count - ret, entries, nr_entries, 0);
366 if (ret >= count)
367 goto err;
368
369 if (page_owner->last_migrate_reason != -1) {
370 ret += snprintf(kbuf + ret, count - ret,
371 "Page has been migrated, last migrate reason: %s\n",
372 migrate_reason_names[page_owner->last_migrate_reason]);
373 if (ret >= count)
374 goto err;
375 }
376
377 ret += snprintf(kbuf + ret, count - ret, "\n");
378 if (ret >= count)
379 goto err;
380
381 if (copy_to_user(buf, kbuf, ret))
382 ret = -EFAULT;
383
384 kfree(kbuf);
385 return ret;
386
387 err:
388 kfree(kbuf);
389 return -ENOMEM;
390 }
391
__dump_page_owner(const struct page * page)392 void __dump_page_owner(const struct page *page)
393 {
394 struct page_ext *page_ext = lookup_page_ext(page);
395 struct page_owner *page_owner;
396 depot_stack_handle_t handle;
397 unsigned long *entries;
398 unsigned int nr_entries;
399 gfp_t gfp_mask;
400 int mt;
401
402 if (unlikely(!page_ext)) {
403 pr_alert("There is not page extension available.\n");
404 return;
405 }
406
407 page_owner = get_page_owner(page_ext);
408 gfp_mask = page_owner->gfp_mask;
409 mt = gfp_migratetype(gfp_mask);
410
411 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
412 pr_alert("page_owner info is not present (never set?)\n");
413 return;
414 }
415
416 if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
417 pr_alert("page_owner tracks the page as allocated\n");
418 else
419 pr_alert("page_owner tracks the page as freed\n");
420
421 pr_alert("page last allocated via order %u, migratetype %s, gfp_mask %#x(%pGg), pid %d, ts %llu, free_ts %llu\n",
422 page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
423 page_owner->pid, page_owner->ts_nsec, page_owner->free_ts_nsec);
424
425 handle = READ_ONCE(page_owner->handle);
426 if (!handle) {
427 pr_alert("page_owner allocation stack trace missing\n");
428 } else {
429 nr_entries = stack_depot_fetch(handle, &entries);
430 stack_trace_print(entries, nr_entries, 0);
431 }
432
433 handle = READ_ONCE(page_owner->free_handle);
434 if (!handle) {
435 pr_alert("page_owner free stack trace missing\n");
436 } else {
437 nr_entries = stack_depot_fetch(handle, &entries);
438 pr_alert("page last free stack trace:\n");
439 stack_trace_print(entries, nr_entries, 0);
440 }
441
442 if (page_owner->last_migrate_reason != -1)
443 pr_alert("page has been migrated, last migrate reason: %s\n",
444 migrate_reason_names[page_owner->last_migrate_reason]);
445 }
446
447 static ssize_t
read_page_owner(struct file * file,char __user * buf,size_t count,loff_t * ppos)448 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
449 {
450 unsigned long pfn;
451 struct page *page;
452 struct page_ext *page_ext;
453 struct page_owner *page_owner;
454 depot_stack_handle_t handle;
455
456 if (!static_branch_unlikely(&page_owner_inited))
457 return -EINVAL;
458
459 page = NULL;
460 pfn = min_low_pfn + *ppos;
461
462 /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
463 while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
464 pfn++;
465
466 drain_all_pages(NULL);
467
468 /* Find an allocated page */
469 for (; pfn < max_pfn; pfn++) {
470 /*
471 * If the new page is in a new MAX_ORDER_NR_PAGES area,
472 * validate the area as existing, skip it if not
473 */
474 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
475 pfn += MAX_ORDER_NR_PAGES - 1;
476 continue;
477 }
478
479 page = pfn_to_page(pfn);
480 if (PageBuddy(page)) {
481 unsigned long freepage_order = buddy_order_unsafe(page);
482
483 if (freepage_order < MAX_ORDER)
484 pfn += (1UL << freepage_order) - 1;
485 continue;
486 }
487
488 page_ext = lookup_page_ext(page);
489 if (unlikely(!page_ext))
490 continue;
491
492 /*
493 * Some pages could be missed by concurrent allocation or free,
494 * because we don't hold the zone lock.
495 */
496 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
497 continue;
498
499 /*
500 * Although we do have the info about past allocation of free
501 * pages, it's not relevant for current memory usage.
502 */
503 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
504 continue;
505
506 page_owner = get_page_owner(page_ext);
507
508 /*
509 * Don't print "tail" pages of high-order allocations as that
510 * would inflate the stats.
511 */
512 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
513 continue;
514
515 /*
516 * Access to page_ext->handle isn't synchronous so we should
517 * be careful to access it.
518 */
519 handle = READ_ONCE(page_owner->handle);
520 if (!handle)
521 continue;
522
523 /* Record the next PFN to read in the file offset */
524 *ppos = (pfn - min_low_pfn) + 1;
525
526 return print_page_owner(buf, count, pfn, page,
527 page_owner, handle);
528 }
529
530 return 0;
531 }
532
init_pages_in_zone(pg_data_t * pgdat,struct zone * zone)533 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
534 {
535 unsigned long pfn = zone->zone_start_pfn;
536 unsigned long end_pfn = zone_end_pfn(zone);
537 unsigned long count = 0;
538
539 /*
540 * Walk the zone in pageblock_nr_pages steps. If a page block spans
541 * a zone boundary, it will be double counted between zones. This does
542 * not matter as the mixed block count will still be correct
543 */
544 for (; pfn < end_pfn; ) {
545 unsigned long block_end_pfn;
546
547 if (!pfn_valid(pfn)) {
548 pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
549 continue;
550 }
551
552 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
553 block_end_pfn = min(block_end_pfn, end_pfn);
554
555 for (; pfn < block_end_pfn; pfn++) {
556 struct page *page = pfn_to_page(pfn);
557 struct page_ext *page_ext;
558
559 if (page_zone(page) != zone)
560 continue;
561
562 /*
563 * To avoid having to grab zone->lock, be a little
564 * careful when reading buddy page order. The only
565 * danger is that we skip too much and potentially miss
566 * some early allocated pages, which is better than
567 * heavy lock contention.
568 */
569 if (PageBuddy(page)) {
570 unsigned long order = buddy_order_unsafe(page);
571
572 if (order > 0 && order < MAX_ORDER)
573 pfn += (1UL << order) - 1;
574 continue;
575 }
576
577 if (PageReserved(page))
578 continue;
579
580 page_ext = lookup_page_ext(page);
581 if (unlikely(!page_ext))
582 continue;
583
584 /* Maybe overlapping zone */
585 if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
586 continue;
587
588 /* Found early allocated page */
589 __set_page_owner_handle(page_ext, early_handle,
590 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
init_zones_in_node(pg_data_t * pgdat)600 static 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
init_early_allocated_pages(void)613 static 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
621 static const struct file_operations proc_page_owner_operations = {
622 .read = read_page_owner,
623 };
624
pageowner_init(void)625 static int __init pageowner_init(void)
626 {
627 if (!static_branch_unlikely(&page_owner_inited)) {
628 pr_info("page_owner is disabled\n");
629 return 0;
630 }
631
632 debugfs_create_file("page_owner", 0400, NULL, NULL,
633 &proc_page_owner_operations);
634
635 return 0;
636 }
637 late_initcall(pageowner_init)
638
