1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/debugfs.h>
4 #include <linux/slab.h>
5 #include <linux/uaccess.h>
6 #include <linux/memblock.h>
7 #include <linux/stacktrace.h>
8 #include <linux/page_owner.h>
9 #include <linux/jump_label.h>
10 #include <linux/migrate.h>
11 #include <linux/stackdepot.h>
12 #include <linux/seq_file.h>
13 #include <linux/memcontrol.h>
14 #include <linux/sched/clock.h>
19 * TODO: teach PAGE_OWNER_STACK_DEPTH (__dump_page_owner and save_stack)
20 * to use off stack temporal storage
22 #define PAGE_OWNER_STACK_DEPTH (16)
26 short last_migrate_reason;
28 depot_stack_handle_t handle;
29 depot_stack_handle_t free_handle;
32 char comm[TASK_COMM_LEN];
37 static bool page_owner_enabled = false;
38 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
40 static depot_stack_handle_t dummy_handle;
41 static depot_stack_handle_t failure_handle;
42 static depot_stack_handle_t early_handle;
44 static void init_early_allocated_pages(void);
46 static int __init early_page_owner_param(char *buf)
48 return kstrtobool(buf, &page_owner_enabled);
50 early_param("page_owner", early_page_owner_param);
52 static __init bool need_page_owner(void)
54 return page_owner_enabled;
57 static __always_inline depot_stack_handle_t create_dummy_stack(void)
59 unsigned long entries[4];
60 unsigned int nr_entries;
62 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
63 return stack_depot_save(entries, nr_entries, GFP_KERNEL);
66 static noinline void register_dummy_stack(void)
68 dummy_handle = create_dummy_stack();
71 static noinline void register_failure_stack(void)
73 failure_handle = create_dummy_stack();
76 static noinline void register_early_stack(void)
78 early_handle = create_dummy_stack();
81 static __init void init_page_owner(void)
83 if (!page_owner_enabled)
88 register_dummy_stack();
89 register_failure_stack();
90 register_early_stack();
91 static_branch_enable(&page_owner_inited);
92 init_early_allocated_pages();
95 struct page_ext_operations page_owner_ops = {
96 .size = sizeof(struct page_owner),
97 .need = need_page_owner,
98 .init = init_page_owner,
101 static inline struct page_owner *get_page_owner(struct page_ext *page_ext)
103 return (void *)page_ext + page_owner_ops.offset;
106 static noinline depot_stack_handle_t save_stack(gfp_t flags)
108 unsigned long entries[PAGE_OWNER_STACK_DEPTH];
109 depot_stack_handle_t handle;
110 unsigned int nr_entries;
115 * Sometimes page metadata allocation tracking requires more
116 * memory to be allocated:
117 * - when new stack trace is saved to stack depot
118 * - when backtrace itself is calculated (ia64)
120 if (current->in_page_owner)
122 current->in_page_owner = 1;
124 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2);
125 handle = stack_depot_save(entries, nr_entries, flags);
127 handle = failure_handle;
129 current->in_page_owner = 0;
133 void __reset_page_owner(struct page *page, unsigned short order)
136 struct page_ext *page_ext;
137 depot_stack_handle_t handle;
138 struct page_owner *page_owner;
139 u64 free_ts_nsec = local_clock();
141 page_ext = lookup_page_ext(page);
142 if (unlikely(!page_ext))
145 handle = save_stack(GFP_NOWAIT | __GFP_NOWARN);
146 for (i = 0; i < (1 << order); i++) {
147 __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
148 page_owner = get_page_owner(page_ext);
149 page_owner->free_handle = handle;
150 page_owner->free_ts_nsec = free_ts_nsec;
151 page_ext = page_ext_next(page_ext);
155 static inline void __set_page_owner_handle(struct page_ext *page_ext,
156 depot_stack_handle_t handle,
157 unsigned short order, gfp_t gfp_mask)
159 struct page_owner *page_owner;
162 for (i = 0; i < (1 << order); i++) {
163 page_owner = get_page_owner(page_ext);
164 page_owner->handle = handle;
165 page_owner->order = order;
166 page_owner->gfp_mask = gfp_mask;
167 page_owner->last_migrate_reason = -1;
168 page_owner->pid = current->pid;
169 page_owner->tgid = current->tgid;
170 page_owner->ts_nsec = local_clock();
171 strlcpy(page_owner->comm, current->comm,
172 sizeof(page_owner->comm));
173 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
174 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
176 page_ext = page_ext_next(page_ext);
180 noinline void __set_page_owner(struct page *page, unsigned short order,
183 struct page_ext *page_ext = lookup_page_ext(page);
184 depot_stack_handle_t handle;
186 if (unlikely(!page_ext))
189 handle = save_stack(gfp_mask);
190 __set_page_owner_handle(page_ext, handle, order, gfp_mask);
193 void __set_page_owner_migrate_reason(struct page *page, int reason)
195 struct page_ext *page_ext = lookup_page_ext(page);
196 struct page_owner *page_owner;
198 if (unlikely(!page_ext))
201 page_owner = get_page_owner(page_ext);
202 page_owner->last_migrate_reason = reason;
205 void __split_page_owner(struct page *page, unsigned int nr)
208 struct page_ext *page_ext = lookup_page_ext(page);
209 struct page_owner *page_owner;
211 if (unlikely(!page_ext))
214 for (i = 0; i < nr; i++) {
215 page_owner = get_page_owner(page_ext);
216 page_owner->order = 0;
217 page_ext = page_ext_next(page_ext);
221 void __folio_copy_owner(struct folio *newfolio, struct folio *old)
223 struct page_ext *old_ext = lookup_page_ext(&old->page);
224 struct page_ext *new_ext = lookup_page_ext(&newfolio->page);
225 struct page_owner *old_page_owner, *new_page_owner;
227 if (unlikely(!old_ext || !new_ext))
230 old_page_owner = get_page_owner(old_ext);
231 new_page_owner = get_page_owner(new_ext);
232 new_page_owner->order = old_page_owner->order;
233 new_page_owner->gfp_mask = old_page_owner->gfp_mask;
234 new_page_owner->last_migrate_reason =
235 old_page_owner->last_migrate_reason;
236 new_page_owner->handle = old_page_owner->handle;
237 new_page_owner->pid = old_page_owner->pid;
238 new_page_owner->tgid = old_page_owner->tgid;
239 new_page_owner->ts_nsec = old_page_owner->ts_nsec;
240 new_page_owner->free_ts_nsec = old_page_owner->ts_nsec;
241 strcpy(new_page_owner->comm, old_page_owner->comm);
244 * We don't clear the bit on the old folio 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 overall 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 old folio 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.
252 __set_bit(PAGE_EXT_OWNER, &new_ext->flags);
253 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &new_ext->flags);
256 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
257 pg_data_t *pgdat, struct zone *zone)
260 struct page_ext *page_ext;
261 struct page_owner *page_owner;
262 unsigned long pfn, block_end_pfn;
263 unsigned long end_pfn = zone_end_pfn(zone);
264 unsigned long count[MIGRATE_TYPES] = { 0, };
265 int pageblock_mt, page_mt;
268 /* Scan block by block. First and last block may be incomplete */
269 pfn = zone->zone_start_pfn;
272 * Walk the zone in pageblock_nr_pages steps. If a page block spans
273 * a zone boundary, it will be double counted between zones. This does
274 * not matter as the mixed block count will still be correct
276 for (; pfn < end_pfn; ) {
277 page = pfn_to_online_page(pfn);
279 pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
283 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
284 block_end_pfn = min(block_end_pfn, end_pfn);
286 pageblock_mt = get_pageblock_migratetype(page);
288 for (; pfn < block_end_pfn; pfn++) {
289 /* The pageblock is online, no need to recheck. */
290 page = pfn_to_page(pfn);
292 if (page_zone(page) != zone)
295 if (PageBuddy(page)) {
296 unsigned long freepage_order;
298 freepage_order = buddy_order_unsafe(page);
299 if (freepage_order < MAX_ORDER)
300 pfn += (1UL << freepage_order) - 1;
304 if (PageReserved(page))
307 page_ext = lookup_page_ext(page);
308 if (unlikely(!page_ext))
311 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
314 page_owner = get_page_owner(page_ext);
315 page_mt = gfp_migratetype(page_owner->gfp_mask);
316 if (pageblock_mt != page_mt) {
317 if (is_migrate_cma(pageblock_mt))
318 count[MIGRATE_MOVABLE]++;
320 count[pageblock_mt]++;
325 pfn += (1UL << page_owner->order) - 1;
330 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
331 for (i = 0; i < MIGRATE_TYPES; i++)
332 seq_printf(m, "%12lu ", count[i]);
337 * Looking for memcg information and print it out
339 static inline int print_page_owner_memcg(char *kbuf, size_t count, int ret,
343 unsigned long memcg_data;
344 struct mem_cgroup *memcg;
349 memcg_data = READ_ONCE(page->memcg_data);
353 if (memcg_data & MEMCG_DATA_OBJCGS)
354 ret += scnprintf(kbuf + ret, count - ret,
355 "Slab cache page\n");
357 memcg = page_memcg_check(page);
361 online = (memcg->css.flags & CSS_ONLINE);
362 cgroup_name(memcg->css.cgroup, name, sizeof(name));
363 ret += scnprintf(kbuf + ret, count - ret,
364 "Charged %sto %smemcg %s\n",
365 PageMemcgKmem(page) ? "(via objcg) " : "",
366 online ? "" : "offline ",
370 #endif /* CONFIG_MEMCG */
376 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
377 struct page *page, struct page_owner *page_owner,
378 depot_stack_handle_t handle)
380 int ret, pageblock_mt, page_mt;
383 count = min_t(size_t, count, PAGE_SIZE);
384 kbuf = kmalloc(count, GFP_KERNEL);
388 ret = scnprintf(kbuf, count,
389 "Page allocated via order %u, mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu ns, free_ts %llu ns\n",
390 page_owner->order, page_owner->gfp_mask,
391 &page_owner->gfp_mask, page_owner->pid,
392 page_owner->tgid, page_owner->comm,
393 page_owner->ts_nsec, page_owner->free_ts_nsec);
395 /* Print information relevant to grouping pages by mobility */
396 pageblock_mt = get_pageblock_migratetype(page);
397 page_mt = gfp_migratetype(page_owner->gfp_mask);
398 ret += scnprintf(kbuf + ret, count - ret,
399 "PFN %lu type %s Block %lu type %s Flags %pGp\n",
401 migratetype_names[page_mt],
402 pfn >> pageblock_order,
403 migratetype_names[pageblock_mt],
406 ret += stack_depot_snprint(handle, kbuf + ret, count - ret, 0);
410 if (page_owner->last_migrate_reason != -1) {
411 ret += scnprintf(kbuf + ret, count - ret,
412 "Page has been migrated, last migrate reason: %s\n",
413 migrate_reason_names[page_owner->last_migrate_reason]);
416 ret = print_page_owner_memcg(kbuf, count, ret, page);
418 ret += snprintf(kbuf + ret, count - ret, "\n");
422 if (copy_to_user(buf, kbuf, ret))
433 void __dump_page_owner(const struct page *page)
435 struct page_ext *page_ext = lookup_page_ext(page);
436 struct page_owner *page_owner;
437 depot_stack_handle_t handle;
441 if (unlikely(!page_ext)) {
442 pr_alert("There is not page extension available.\n");
446 page_owner = get_page_owner(page_ext);
447 gfp_mask = page_owner->gfp_mask;
448 mt = gfp_migratetype(gfp_mask);
450 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
451 pr_alert("page_owner info is not present (never set?)\n");
455 if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
456 pr_alert("page_owner tracks the page as allocated\n");
458 pr_alert("page_owner tracks the page as freed\n");
460 pr_alert("page last allocated via order %u, migratetype %s, gfp_mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu, free_ts %llu\n",
461 page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
462 page_owner->pid, page_owner->tgid, page_owner->comm,
463 page_owner->ts_nsec, page_owner->free_ts_nsec);
465 handle = READ_ONCE(page_owner->handle);
467 pr_alert("page_owner allocation stack trace missing\n");
469 stack_depot_print(handle);
471 handle = READ_ONCE(page_owner->free_handle);
473 pr_alert("page_owner free stack trace missing\n");
475 pr_alert("page last free stack trace:\n");
476 stack_depot_print(handle);
479 if (page_owner->last_migrate_reason != -1)
480 pr_alert("page has been migrated, last migrate reason: %s\n",
481 migrate_reason_names[page_owner->last_migrate_reason]);
485 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
489 struct page_ext *page_ext;
490 struct page_owner *page_owner;
491 depot_stack_handle_t handle;
493 if (!static_branch_unlikely(&page_owner_inited))
497 pfn = min_low_pfn + *ppos;
499 /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
500 while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
503 drain_all_pages(NULL);
505 /* Find an allocated page */
506 for (; pfn < max_pfn; pfn++) {
508 * If the new page is in a new MAX_ORDER_NR_PAGES area,
509 * validate the area as existing, skip it if not
511 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
512 pfn += MAX_ORDER_NR_PAGES - 1;
516 page = pfn_to_page(pfn);
517 if (PageBuddy(page)) {
518 unsigned long freepage_order = buddy_order_unsafe(page);
520 if (freepage_order < MAX_ORDER)
521 pfn += (1UL << freepage_order) - 1;
525 page_ext = lookup_page_ext(page);
526 if (unlikely(!page_ext))
530 * Some pages could be missed by concurrent allocation or free,
531 * because we don't hold the zone lock.
533 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
537 * Although we do have the info about past allocation of free
538 * pages, it's not relevant for current memory usage.
540 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
543 page_owner = get_page_owner(page_ext);
546 * Don't print "tail" pages of high-order allocations as that
547 * would inflate the stats.
549 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
553 * Access to page_ext->handle isn't synchronous so we should
554 * be careful to access it.
556 handle = READ_ONCE(page_owner->handle);
560 /* Record the next PFN to read in the file offset */
561 *ppos = (pfn - min_low_pfn) + 1;
563 return print_page_owner(buf, count, pfn, page,
570 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
572 unsigned long pfn = zone->zone_start_pfn;
573 unsigned long end_pfn = zone_end_pfn(zone);
574 unsigned long count = 0;
577 * Walk the zone in pageblock_nr_pages steps. If a page block spans
578 * a zone boundary, it will be double counted between zones. This does
579 * not matter as the mixed block count will still be correct
581 for (; pfn < end_pfn; ) {
582 unsigned long block_end_pfn;
584 if (!pfn_valid(pfn)) {
585 pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
589 block_end_pfn = ALIGN(pfn + 1, pageblock_nr_pages);
590 block_end_pfn = min(block_end_pfn, end_pfn);
592 for (; pfn < block_end_pfn; pfn++) {
593 struct page *page = pfn_to_page(pfn);
594 struct page_ext *page_ext;
596 if (page_zone(page) != zone)
600 * To avoid having to grab zone->lock, be a little
601 * careful when reading buddy page order. The only
602 * danger is that we skip too much and potentially miss
603 * some early allocated pages, which is better than
604 * heavy lock contention.
606 if (PageBuddy(page)) {
607 unsigned long order = buddy_order_unsafe(page);
609 if (order > 0 && order < MAX_ORDER)
610 pfn += (1UL << order) - 1;
614 if (PageReserved(page))
617 page_ext = lookup_page_ext(page);
618 if (unlikely(!page_ext))
621 /* Maybe overlapping zone */
622 if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
625 /* Found early allocated page */
626 __set_page_owner_handle(page_ext, early_handle,
633 pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
634 pgdat->node_id, zone->name, count);
637 static void init_zones_in_node(pg_data_t *pgdat)
640 struct zone *node_zones = pgdat->node_zones;
642 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
643 if (!populated_zone(zone))
646 init_pages_in_zone(pgdat, zone);
650 static void init_early_allocated_pages(void)
654 for_each_online_pgdat(pgdat)
655 init_zones_in_node(pgdat);
658 static const struct file_operations proc_page_owner_operations = {
659 .read = read_page_owner,
662 static int __init pageowner_init(void)
664 if (!static_branch_unlikely(&page_owner_inited)) {
665 pr_info("page_owner is disabled\n");
669 debugfs_create_file("page_owner", 0400, NULL, NULL,
670 &proc_page_owner_operations);
674 late_initcall(pageowner_init)