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/memcontrol.h>
14 #include <linux/sched/clock.h>
15
16 #include "internal.h"
17
18 /*
19 * TODO: teach PAGE_OWNER_STACK_DEPTH (__dump_page_owner and save_stack)
20 * to use off stack temporal storage
21 */
22 #define PAGE_OWNER_STACK_DEPTH (16)
23
24 struct page_owner {
25 unsigned short order;
26 short last_migrate_reason;
27 gfp_t gfp_mask;
28 depot_stack_handle_t handle;
29 depot_stack_handle_t free_handle;
30 u64 ts_nsec;
31 u64 free_ts_nsec;
32 char comm[TASK_COMM_LEN];
33 pid_t pid;
34 pid_t tgid;
35 pid_t free_pid;
36 pid_t free_tgid;
37 };
38
39 struct stack {
40 struct stack_record *stack_record;
41 struct stack *next;
42 };
43 static struct stack dummy_stack;
44 static struct stack failure_stack;
45 static struct stack *stack_list;
46 static DEFINE_SPINLOCK(stack_list_lock);
47
48 static bool page_owner_enabled __initdata;
49 DEFINE_STATIC_KEY_FALSE(page_owner_inited);
50
51 static depot_stack_handle_t dummy_handle;
52 static depot_stack_handle_t failure_handle;
53 static depot_stack_handle_t early_handle;
54
55 static void init_early_allocated_pages(void);
56
set_current_in_page_owner(void)57 static inline void set_current_in_page_owner(void)
58 {
59 /*
60 * Avoid recursion.
61 *
62 * We might need to allocate more memory from page_owner code, so make
63 * sure to signal it in order to avoid recursion.
64 */
65 current->in_page_owner = 1;
66 }
67
unset_current_in_page_owner(void)68 static inline void unset_current_in_page_owner(void)
69 {
70 current->in_page_owner = 0;
71 }
72
early_page_owner_param(char * buf)73 static int __init early_page_owner_param(char *buf)
74 {
75 int ret = kstrtobool(buf, &page_owner_enabled);
76
77 if (page_owner_enabled)
78 stack_depot_request_early_init();
79
80 return ret;
81 }
82 early_param("page_owner", early_page_owner_param);
83
need_page_owner(void)84 static __init bool need_page_owner(void)
85 {
86 return page_owner_enabled;
87 }
88
create_dummy_stack(void)89 static __always_inline depot_stack_handle_t create_dummy_stack(void)
90 {
91 unsigned long entries[4];
92 unsigned int nr_entries;
93
94 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 0);
95 return stack_depot_save(entries, nr_entries, GFP_KERNEL);
96 }
97
register_dummy_stack(void)98 static noinline void register_dummy_stack(void)
99 {
100 dummy_handle = create_dummy_stack();
101 }
102
register_failure_stack(void)103 static noinline void register_failure_stack(void)
104 {
105 failure_handle = create_dummy_stack();
106 }
107
register_early_stack(void)108 static noinline void register_early_stack(void)
109 {
110 early_handle = create_dummy_stack();
111 }
112
init_page_owner(void)113 static __init void init_page_owner(void)
114 {
115 if (!page_owner_enabled)
116 return;
117
118 register_dummy_stack();
119 register_failure_stack();
120 register_early_stack();
121 init_early_allocated_pages();
122 /* Initialize dummy and failure stacks and link them to stack_list */
123 dummy_stack.stack_record = __stack_depot_get_stack_record(dummy_handle);
124 failure_stack.stack_record = __stack_depot_get_stack_record(failure_handle);
125 if (dummy_stack.stack_record)
126 refcount_set(&dummy_stack.stack_record->count, 1);
127 if (failure_stack.stack_record)
128 refcount_set(&failure_stack.stack_record->count, 1);
129 dummy_stack.next = &failure_stack;
130 stack_list = &dummy_stack;
131 static_branch_enable(&page_owner_inited);
132 }
133
134 struct page_ext_operations page_owner_ops = {
135 .size = sizeof(struct page_owner),
136 .need = need_page_owner,
137 .init = init_page_owner,
138 .need_shared_flags = true,
139 };
140
get_page_owner(struct page_ext * page_ext)141 static inline struct page_owner *get_page_owner(struct page_ext *page_ext)
142 {
143 return page_ext_data(page_ext, &page_owner_ops);
144 }
145
save_stack(gfp_t flags)146 static noinline depot_stack_handle_t save_stack(gfp_t flags)
147 {
148 unsigned long entries[PAGE_OWNER_STACK_DEPTH];
149 depot_stack_handle_t handle;
150 unsigned int nr_entries;
151
152 if (current->in_page_owner)
153 return dummy_handle;
154
155 set_current_in_page_owner();
156 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 2);
157 handle = stack_depot_save(entries, nr_entries, flags);
158 if (!handle)
159 handle = failure_handle;
160 unset_current_in_page_owner();
161
162 return handle;
163 }
164
add_stack_record_to_list(struct stack_record * stack_record,gfp_t gfp_mask)165 static void add_stack_record_to_list(struct stack_record *stack_record,
166 gfp_t gfp_mask)
167 {
168 unsigned long flags;
169 struct stack *stack;
170
171 if (!gfpflags_allow_spinning(gfp_mask))
172 return;
173
174 set_current_in_page_owner();
175 stack = kmalloc(sizeof(*stack), gfp_nested_mask(gfp_mask));
176 if (!stack) {
177 unset_current_in_page_owner();
178 return;
179 }
180 unset_current_in_page_owner();
181
182 stack->stack_record = stack_record;
183 stack->next = NULL;
184
185 spin_lock_irqsave(&stack_list_lock, flags);
186 stack->next = stack_list;
187 /*
188 * This pairs with smp_load_acquire() from function
189 * stack_start(). This guarantees that stack_start()
190 * will see an updated stack_list before starting to
191 * traverse the list.
192 */
193 smp_store_release(&stack_list, stack);
194 spin_unlock_irqrestore(&stack_list_lock, flags);
195 }
196
inc_stack_record_count(depot_stack_handle_t handle,gfp_t gfp_mask,int nr_base_pages)197 static void inc_stack_record_count(depot_stack_handle_t handle, gfp_t gfp_mask,
198 int nr_base_pages)
199 {
200 struct stack_record *stack_record = __stack_depot_get_stack_record(handle);
201
202 if (!stack_record)
203 return;
204
205 /*
206 * New stack_record's that do not use STACK_DEPOT_FLAG_GET start
207 * with REFCOUNT_SATURATED to catch spurious increments of their
208 * refcount.
209 * Since we do not use STACK_DEPOT_FLAG_GET API, let us
210 * set a refcount of 1 ourselves.
211 */
212 if (refcount_read(&stack_record->count) == REFCOUNT_SATURATED) {
213 int old = REFCOUNT_SATURATED;
214
215 if (atomic_try_cmpxchg_relaxed(&stack_record->count.refs, &old, 1))
216 /* Add the new stack_record to our list */
217 add_stack_record_to_list(stack_record, gfp_mask);
218 }
219 refcount_add(nr_base_pages, &stack_record->count);
220 }
221
dec_stack_record_count(depot_stack_handle_t handle,int nr_base_pages)222 static void dec_stack_record_count(depot_stack_handle_t handle,
223 int nr_base_pages)
224 {
225 struct stack_record *stack_record = __stack_depot_get_stack_record(handle);
226
227 if (!stack_record)
228 return;
229
230 if (refcount_sub_and_test(nr_base_pages, &stack_record->count))
231 pr_warn("%s: refcount went to 0 for %u handle\n", __func__,
232 handle);
233 }
234
__update_page_owner_handle(struct page * page,depot_stack_handle_t handle,unsigned short order,gfp_t gfp_mask,short last_migrate_reason,u64 ts_nsec,pid_t pid,pid_t tgid,char * comm)235 static inline void __update_page_owner_handle(struct page *page,
236 depot_stack_handle_t handle,
237 unsigned short order,
238 gfp_t gfp_mask,
239 short last_migrate_reason, u64 ts_nsec,
240 pid_t pid, pid_t tgid, char *comm)
241 {
242 struct page_ext_iter iter;
243 struct page_ext *page_ext;
244 struct page_owner *page_owner;
245
246 rcu_read_lock();
247 for_each_page_ext(page, 1 << order, page_ext, iter) {
248 page_owner = get_page_owner(page_ext);
249 page_owner->handle = handle;
250 page_owner->order = order;
251 page_owner->gfp_mask = gfp_mask;
252 page_owner->last_migrate_reason = last_migrate_reason;
253 page_owner->pid = pid;
254 page_owner->tgid = tgid;
255 page_owner->ts_nsec = ts_nsec;
256 strscpy(page_owner->comm, comm,
257 sizeof(page_owner->comm));
258 __set_bit(PAGE_EXT_OWNER, &page_ext->flags);
259 __set_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
260 }
261 rcu_read_unlock();
262 }
263
__update_page_owner_free_handle(struct page * page,depot_stack_handle_t handle,unsigned short order,pid_t pid,pid_t tgid,u64 free_ts_nsec)264 static inline void __update_page_owner_free_handle(struct page *page,
265 depot_stack_handle_t handle,
266 unsigned short order,
267 pid_t pid, pid_t tgid,
268 u64 free_ts_nsec)
269 {
270 struct page_ext_iter iter;
271 struct page_ext *page_ext;
272 struct page_owner *page_owner;
273
274 rcu_read_lock();
275 for_each_page_ext(page, 1 << order, page_ext, iter) {
276 page_owner = get_page_owner(page_ext);
277 /* Only __reset_page_owner() wants to clear the bit */
278 if (handle) {
279 __clear_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
280 page_owner->free_handle = handle;
281 }
282 page_owner->free_ts_nsec = free_ts_nsec;
283 page_owner->free_pid = current->pid;
284 page_owner->free_tgid = current->tgid;
285 }
286 rcu_read_unlock();
287 }
288
__reset_page_owner(struct page * page,unsigned short order)289 void __reset_page_owner(struct page *page, unsigned short order)
290 {
291 struct page_ext *page_ext;
292 depot_stack_handle_t handle;
293 depot_stack_handle_t alloc_handle;
294 struct page_owner *page_owner;
295 u64 free_ts_nsec = local_clock();
296
297 page_ext = page_ext_get(page);
298 if (unlikely(!page_ext))
299 return;
300
301 page_owner = get_page_owner(page_ext);
302 alloc_handle = page_owner->handle;
303 page_ext_put(page_ext);
304
305 /*
306 * Do not specify GFP_NOWAIT to make gfpflags_allow_spinning() == false
307 * to prevent issues in stack_depot_save().
308 * This is similar to alloc_pages_nolock() gfp flags, but only used
309 * to signal stack_depot to avoid spin_locks.
310 */
311 handle = save_stack(__GFP_NOWARN);
312 __update_page_owner_free_handle(page, handle, order, current->pid,
313 current->tgid, free_ts_nsec);
314
315 if (alloc_handle != early_handle)
316 /*
317 * early_handle is being set as a handle for all those
318 * early allocated pages. See init_pages_in_zone().
319 * Since their refcount is not being incremented because
320 * the machinery is not ready yet, we cannot decrement
321 * their refcount either.
322 */
323 dec_stack_record_count(alloc_handle, 1 << order);
324 }
325
__set_page_owner(struct page * page,unsigned short order,gfp_t gfp_mask)326 noinline void __set_page_owner(struct page *page, unsigned short order,
327 gfp_t gfp_mask)
328 {
329 u64 ts_nsec = local_clock();
330 depot_stack_handle_t handle;
331
332 handle = save_stack(gfp_mask);
333 __update_page_owner_handle(page, handle, order, gfp_mask, -1,
334 ts_nsec, current->pid, current->tgid,
335 current->comm);
336 inc_stack_record_count(handle, gfp_mask, 1 << order);
337 }
338
__folio_set_owner_migrate_reason(struct folio * folio,int reason)339 void __folio_set_owner_migrate_reason(struct folio *folio, int reason)
340 {
341 struct page_ext *page_ext = page_ext_get(&folio->page);
342 struct page_owner *page_owner;
343
344 if (unlikely(!page_ext))
345 return;
346
347 page_owner = get_page_owner(page_ext);
348 page_owner->last_migrate_reason = reason;
349 page_ext_put(page_ext);
350 }
351
__split_page_owner(struct page * page,int old_order,int new_order)352 void __split_page_owner(struct page *page, int old_order, int new_order)
353 {
354 struct page_ext_iter iter;
355 struct page_ext *page_ext;
356 struct page_owner *page_owner;
357
358 rcu_read_lock();
359 for_each_page_ext(page, 1 << old_order, page_ext, iter) {
360 page_owner = get_page_owner(page_ext);
361 page_owner->order = new_order;
362 }
363 rcu_read_unlock();
364 }
365
__folio_copy_owner(struct folio * newfolio,struct folio * old)366 void __folio_copy_owner(struct folio *newfolio, struct folio *old)
367 {
368 struct page_ext *page_ext;
369 struct page_ext_iter iter;
370 struct page_owner *old_page_owner;
371 struct page_owner *new_page_owner;
372 depot_stack_handle_t migrate_handle;
373
374 page_ext = page_ext_get(&old->page);
375 if (unlikely(!page_ext))
376 return;
377
378 old_page_owner = get_page_owner(page_ext);
379 page_ext_put(page_ext);
380
381 page_ext = page_ext_get(&newfolio->page);
382 if (unlikely(!page_ext))
383 return;
384
385 new_page_owner = get_page_owner(page_ext);
386 page_ext_put(page_ext);
387
388 migrate_handle = new_page_owner->handle;
389 __update_page_owner_handle(&newfolio->page, old_page_owner->handle,
390 old_page_owner->order, old_page_owner->gfp_mask,
391 old_page_owner->last_migrate_reason,
392 old_page_owner->ts_nsec, old_page_owner->pid,
393 old_page_owner->tgid, old_page_owner->comm);
394 /*
395 * Do not proactively clear PAGE_EXT_OWNER{_ALLOCATED} bits as the folio
396 * will be freed after migration. Keep them until then as they may be
397 * useful.
398 */
399 __update_page_owner_free_handle(&newfolio->page, 0, old_page_owner->order,
400 old_page_owner->free_pid,
401 old_page_owner->free_tgid,
402 old_page_owner->free_ts_nsec);
403 /*
404 * We linked the original stack to the new folio, we need to do the same
405 * for the new one and the old folio otherwise there will be an imbalance
406 * when subtracting those pages from the stack.
407 */
408 rcu_read_lock();
409 for_each_page_ext(&old->page, 1 << new_page_owner->order, page_ext, iter) {
410 old_page_owner = get_page_owner(page_ext);
411 old_page_owner->handle = migrate_handle;
412 }
413 rcu_read_unlock();
414 }
415
pagetypeinfo_showmixedcount_print(struct seq_file * m,pg_data_t * pgdat,struct zone * zone)416 void pagetypeinfo_showmixedcount_print(struct seq_file *m,
417 pg_data_t *pgdat, struct zone *zone)
418 {
419 struct page *page;
420 struct page_ext *page_ext;
421 struct page_owner *page_owner;
422 unsigned long pfn, block_end_pfn;
423 unsigned long end_pfn = zone_end_pfn(zone);
424 unsigned long count[MIGRATE_TYPES] = { 0, };
425 int pageblock_mt, page_mt;
426 int i;
427
428 /* Scan block by block. First and last block may be incomplete */
429 pfn = zone->zone_start_pfn;
430
431 /*
432 * Walk the zone in pageblock_nr_pages steps. If a page block spans
433 * a zone boundary, it will be double counted between zones. This does
434 * not matter as the mixed block count will still be correct
435 */
436 for (; pfn < end_pfn; ) {
437 page = pfn_to_online_page(pfn);
438 if (!page) {
439 pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
440 continue;
441 }
442
443 block_end_pfn = pageblock_end_pfn(pfn);
444 block_end_pfn = min(block_end_pfn, end_pfn);
445
446 pageblock_mt = get_pageblock_migratetype(page);
447
448 for (; pfn < block_end_pfn; pfn++) {
449 /* The pageblock is online, no need to recheck. */
450 page = pfn_to_page(pfn);
451
452 if (page_zone(page) != zone)
453 continue;
454
455 if (PageBuddy(page)) {
456 unsigned long freepage_order;
457
458 freepage_order = buddy_order_unsafe(page);
459 if (freepage_order <= MAX_PAGE_ORDER)
460 pfn += (1UL << freepage_order) - 1;
461 continue;
462 }
463
464 if (PageReserved(page))
465 continue;
466
467 page_ext = page_ext_get(page);
468 if (unlikely(!page_ext))
469 continue;
470
471 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
472 goto ext_put_continue;
473
474 page_owner = get_page_owner(page_ext);
475 page_mt = gfp_migratetype(page_owner->gfp_mask);
476 if (pageblock_mt != page_mt) {
477 if (is_migrate_cma(pageblock_mt))
478 count[MIGRATE_MOVABLE]++;
479 else
480 count[pageblock_mt]++;
481
482 pfn = block_end_pfn;
483 page_ext_put(page_ext);
484 break;
485 }
486 pfn += (1UL << page_owner->order) - 1;
487 ext_put_continue:
488 page_ext_put(page_ext);
489 }
490 }
491
492 /* Print counts */
493 seq_printf(m, "Node %d, zone %8s ", pgdat->node_id, zone->name);
494 for (i = 0; i < MIGRATE_TYPES; i++)
495 seq_printf(m, "%12lu ", count[i]);
496 seq_putc(m, '\n');
497 }
498
499 /*
500 * Looking for memcg information and print it out
501 */
print_page_owner_memcg(char * kbuf,size_t count,int ret,struct page * page)502 static inline int print_page_owner_memcg(char *kbuf, size_t count, int ret,
503 struct page *page)
504 {
505 #ifdef CONFIG_MEMCG
506 unsigned long memcg_data;
507 struct mem_cgroup *memcg;
508 bool online;
509 char name[80];
510
511 rcu_read_lock();
512 memcg_data = READ_ONCE(page->memcg_data);
513 if (!memcg_data || PageTail(page))
514 goto out_unlock;
515
516 if (memcg_data & MEMCG_DATA_OBJEXTS)
517 ret += scnprintf(kbuf + ret, count - ret,
518 "Slab cache page\n");
519
520 memcg = page_memcg_check(page);
521 if (!memcg)
522 goto out_unlock;
523
524 online = (memcg->css.flags & CSS_ONLINE);
525 cgroup_name(memcg->css.cgroup, name, sizeof(name));
526 ret += scnprintf(kbuf + ret, count - ret,
527 "Charged %sto %smemcg %s\n",
528 PageMemcgKmem(page) ? "(via objcg) " : "",
529 online ? "" : "offline ",
530 name);
531 out_unlock:
532 rcu_read_unlock();
533 #endif /* CONFIG_MEMCG */
534
535 return ret;
536 }
537
538 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)539 print_page_owner(char __user *buf, size_t count, unsigned long pfn,
540 struct page *page, struct page_owner *page_owner,
541 depot_stack_handle_t handle)
542 {
543 int ret, pageblock_mt, page_mt;
544 char *kbuf;
545
546 count = min_t(size_t, count, PAGE_SIZE);
547 kbuf = kmalloc(count, GFP_KERNEL);
548 if (!kbuf)
549 return -ENOMEM;
550
551 ret = scnprintf(kbuf, count,
552 "Page allocated via order %u, mask %#x(%pGg), pid %d, tgid %d (%s), ts %llu ns\n",
553 page_owner->order, page_owner->gfp_mask,
554 &page_owner->gfp_mask, page_owner->pid,
555 page_owner->tgid, page_owner->comm,
556 page_owner->ts_nsec);
557
558 /* Print information relevant to grouping pages by mobility */
559 pageblock_mt = get_pageblock_migratetype(page);
560 page_mt = gfp_migratetype(page_owner->gfp_mask);
561 ret += scnprintf(kbuf + ret, count - ret,
562 "PFN 0x%lx type %s Block %lu type %s Flags %pGp\n",
563 pfn,
564 migratetype_names[page_mt],
565 pfn >> pageblock_order,
566 migratetype_names[pageblock_mt],
567 &page->flags);
568
569 ret += stack_depot_snprint(handle, kbuf + ret, count - ret, 0);
570 if (ret >= count)
571 goto err;
572
573 if (page_owner->last_migrate_reason != -1) {
574 ret += scnprintf(kbuf + ret, count - ret,
575 "Page has been migrated, last migrate reason: %s\n",
576 migrate_reason_names[page_owner->last_migrate_reason]);
577 }
578
579 ret = print_page_owner_memcg(kbuf, count, ret, page);
580
581 ret += snprintf(kbuf + ret, count - ret, "\n");
582 if (ret >= count)
583 goto err;
584
585 if (copy_to_user(buf, kbuf, ret))
586 ret = -EFAULT;
587
588 kfree(kbuf);
589 return ret;
590
591 err:
592 kfree(kbuf);
593 return -ENOMEM;
594 }
595
__dump_page_owner(const struct page * page)596 void __dump_page_owner(const struct page *page)
597 {
598 struct page_ext *page_ext = page_ext_get((void *)page);
599 struct page_owner *page_owner;
600 depot_stack_handle_t handle;
601 gfp_t gfp_mask;
602 int mt;
603
604 if (unlikely(!page_ext)) {
605 pr_alert("There is not page extension available.\n");
606 return;
607 }
608
609 page_owner = get_page_owner(page_ext);
610 gfp_mask = page_owner->gfp_mask;
611 mt = gfp_migratetype(gfp_mask);
612
613 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags)) {
614 pr_alert("page_owner info is not present (never set?)\n");
615 page_ext_put(page_ext);
616 return;
617 }
618
619 if (test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
620 pr_alert("page_owner tracks the page as allocated\n");
621 else
622 pr_alert("page_owner tracks the page as freed\n");
623
624 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",
625 page_owner->order, migratetype_names[mt], gfp_mask, &gfp_mask,
626 page_owner->pid, page_owner->tgid, page_owner->comm,
627 page_owner->ts_nsec, page_owner->free_ts_nsec);
628
629 handle = READ_ONCE(page_owner->handle);
630 if (!handle)
631 pr_alert("page_owner allocation stack trace missing\n");
632 else
633 stack_depot_print(handle);
634
635 handle = READ_ONCE(page_owner->free_handle);
636 if (!handle) {
637 pr_alert("page_owner free stack trace missing\n");
638 } else {
639 pr_alert("page last free pid %d tgid %d stack trace:\n",
640 page_owner->free_pid, page_owner->free_tgid);
641 stack_depot_print(handle);
642 }
643
644 if (page_owner->last_migrate_reason != -1)
645 pr_alert("page has been migrated, last migrate reason: %s\n",
646 migrate_reason_names[page_owner->last_migrate_reason]);
647 page_ext_put(page_ext);
648 }
649
650 static ssize_t
read_page_owner(struct file * file,char __user * buf,size_t count,loff_t * ppos)651 read_page_owner(struct file *file, char __user *buf, size_t count, loff_t *ppos)
652 {
653 unsigned long pfn;
654 struct page *page;
655 struct page_ext *page_ext;
656 struct page_owner *page_owner;
657 depot_stack_handle_t handle;
658
659 if (!static_branch_unlikely(&page_owner_inited))
660 return -EINVAL;
661
662 page = NULL;
663 if (*ppos == 0)
664 pfn = min_low_pfn;
665 else
666 pfn = *ppos;
667 /* Find a valid PFN or the start of a MAX_ORDER_NR_PAGES area */
668 while (!pfn_valid(pfn) && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0)
669 pfn++;
670
671 /* Find an allocated page */
672 for (; pfn < max_pfn; pfn++) {
673 /*
674 * This temporary page_owner is required so
675 * that we can avoid the context switches while holding
676 * the rcu lock and copying the page owner information to
677 * user through copy_to_user() or GFP_KERNEL allocations.
678 */
679 struct page_owner page_owner_tmp;
680
681 /*
682 * If the new page is in a new MAX_ORDER_NR_PAGES area,
683 * validate the area as existing, skip it if not
684 */
685 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0 && !pfn_valid(pfn)) {
686 pfn += MAX_ORDER_NR_PAGES - 1;
687 continue;
688 }
689
690 page = pfn_to_page(pfn);
691 if (PageBuddy(page)) {
692 unsigned long freepage_order = buddy_order_unsafe(page);
693
694 if (freepage_order <= MAX_PAGE_ORDER)
695 pfn += (1UL << freepage_order) - 1;
696 continue;
697 }
698
699 page_ext = page_ext_get(page);
700 if (unlikely(!page_ext))
701 continue;
702
703 /*
704 * Some pages could be missed by concurrent allocation or free,
705 * because we don't hold the zone lock.
706 */
707 if (!test_bit(PAGE_EXT_OWNER, &page_ext->flags))
708 goto ext_put_continue;
709
710 /*
711 * Although we do have the info about past allocation of free
712 * pages, it's not relevant for current memory usage.
713 */
714 if (!test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags))
715 goto ext_put_continue;
716
717 page_owner = get_page_owner(page_ext);
718
719 /*
720 * Don't print "tail" pages of high-order allocations as that
721 * would inflate the stats.
722 */
723 if (!IS_ALIGNED(pfn, 1 << page_owner->order))
724 goto ext_put_continue;
725
726 /*
727 * Access to page_ext->handle isn't synchronous so we should
728 * be careful to access it.
729 */
730 handle = READ_ONCE(page_owner->handle);
731 if (!handle)
732 goto ext_put_continue;
733
734 /* Record the next PFN to read in the file offset */
735 *ppos = pfn + 1;
736
737 page_owner_tmp = *page_owner;
738 page_ext_put(page_ext);
739 return print_page_owner(buf, count, pfn, page,
740 &page_owner_tmp, handle);
741 ext_put_continue:
742 page_ext_put(page_ext);
743 }
744
745 return 0;
746 }
747
lseek_page_owner(struct file * file,loff_t offset,int orig)748 static loff_t lseek_page_owner(struct file *file, loff_t offset, int orig)
749 {
750 switch (orig) {
751 case SEEK_SET:
752 file->f_pos = offset;
753 break;
754 case SEEK_CUR:
755 file->f_pos += offset;
756 break;
757 default:
758 return -EINVAL;
759 }
760 return file->f_pos;
761 }
762
init_pages_in_zone(pg_data_t * pgdat,struct zone * zone)763 static void init_pages_in_zone(pg_data_t *pgdat, struct zone *zone)
764 {
765 unsigned long pfn = zone->zone_start_pfn;
766 unsigned long end_pfn = zone_end_pfn(zone);
767 unsigned long count = 0;
768
769 /*
770 * Walk the zone in pageblock_nr_pages steps. If a page block spans
771 * a zone boundary, it will be double counted between zones. This does
772 * not matter as the mixed block count will still be correct
773 */
774 for (; pfn < end_pfn; ) {
775 unsigned long block_end_pfn;
776
777 if (!pfn_valid(pfn)) {
778 pfn = ALIGN(pfn + 1, MAX_ORDER_NR_PAGES);
779 continue;
780 }
781
782 block_end_pfn = pageblock_end_pfn(pfn);
783 block_end_pfn = min(block_end_pfn, end_pfn);
784
785 for (; pfn < block_end_pfn; pfn++) {
786 struct page *page = pfn_to_page(pfn);
787 struct page_ext *page_ext;
788
789 if (page_zone(page) != zone)
790 continue;
791
792 /*
793 * To avoid having to grab zone->lock, be a little
794 * careful when reading buddy page order. The only
795 * danger is that we skip too much and potentially miss
796 * some early allocated pages, which is better than
797 * heavy lock contention.
798 */
799 if (PageBuddy(page)) {
800 unsigned long order = buddy_order_unsafe(page);
801
802 if (order > 0 && order <= MAX_PAGE_ORDER)
803 pfn += (1UL << order) - 1;
804 continue;
805 }
806
807 if (PageReserved(page))
808 continue;
809
810 page_ext = page_ext_get(page);
811 if (unlikely(!page_ext))
812 continue;
813
814 /* Maybe overlapping zone */
815 if (test_bit(PAGE_EXT_OWNER, &page_ext->flags))
816 goto ext_put_continue;
817
818 /* Found early allocated page */
819 __update_page_owner_handle(page, early_handle, 0, 0,
820 -1, local_clock(), current->pid,
821 current->tgid, current->comm);
822 count++;
823 ext_put_continue:
824 page_ext_put(page_ext);
825 }
826 cond_resched();
827 }
828
829 pr_info("Node %d, zone %8s: page owner found early allocated %lu pages\n",
830 pgdat->node_id, zone->name, count);
831 }
832
init_zones_in_node(pg_data_t * pgdat)833 static void init_zones_in_node(pg_data_t *pgdat)
834 {
835 struct zone *zone;
836 struct zone *node_zones = pgdat->node_zones;
837
838 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
839 if (!populated_zone(zone))
840 continue;
841
842 init_pages_in_zone(pgdat, zone);
843 }
844 }
845
init_early_allocated_pages(void)846 static void init_early_allocated_pages(void)
847 {
848 pg_data_t *pgdat;
849
850 for_each_online_pgdat(pgdat)
851 init_zones_in_node(pgdat);
852 }
853
854 static const struct file_operations proc_page_owner_operations = {
855 .read = read_page_owner,
856 .llseek = lseek_page_owner,
857 };
858
stack_start(struct seq_file * m,loff_t * ppos)859 static void *stack_start(struct seq_file *m, loff_t *ppos)
860 {
861 struct stack *stack;
862
863 if (*ppos == -1UL)
864 return NULL;
865
866 if (!*ppos) {
867 /*
868 * This pairs with smp_store_release() from function
869 * add_stack_record_to_list(), so we get a consistent
870 * value of stack_list.
871 */
872 stack = smp_load_acquire(&stack_list);
873 m->private = stack;
874 } else {
875 stack = m->private;
876 }
877
878 return stack;
879 }
880
stack_next(struct seq_file * m,void * v,loff_t * ppos)881 static void *stack_next(struct seq_file *m, void *v, loff_t *ppos)
882 {
883 struct stack *stack = v;
884
885 stack = stack->next;
886 *ppos = stack ? *ppos + 1 : -1UL;
887 m->private = stack;
888
889 return stack;
890 }
891
892 static unsigned long page_owner_pages_threshold;
893
stack_print(struct seq_file * m,void * v)894 static int stack_print(struct seq_file *m, void *v)
895 {
896 int i, nr_base_pages;
897 struct stack *stack = v;
898 unsigned long *entries;
899 unsigned long nr_entries;
900 struct stack_record *stack_record = stack->stack_record;
901
902 if (!stack->stack_record)
903 return 0;
904
905 nr_entries = stack_record->size;
906 entries = stack_record->entries;
907 nr_base_pages = refcount_read(&stack_record->count) - 1;
908
909 if (nr_base_pages < 1 || nr_base_pages < page_owner_pages_threshold)
910 return 0;
911
912 for (i = 0; i < nr_entries; i++)
913 seq_printf(m, " %pS\n", (void *)entries[i]);
914 seq_printf(m, "nr_base_pages: %d\n\n", nr_base_pages);
915
916 return 0;
917 }
918
stack_stop(struct seq_file * m,void * v)919 static void stack_stop(struct seq_file *m, void *v)
920 {
921 }
922
923 static const struct seq_operations page_owner_stack_op = {
924 .start = stack_start,
925 .next = stack_next,
926 .stop = stack_stop,
927 .show = stack_print
928 };
929
page_owner_stack_open(struct inode * inode,struct file * file)930 static int page_owner_stack_open(struct inode *inode, struct file *file)
931 {
932 return seq_open_private(file, &page_owner_stack_op, 0);
933 }
934
935 static const struct file_operations page_owner_stack_operations = {
936 .open = page_owner_stack_open,
937 .read = seq_read,
938 .llseek = seq_lseek,
939 .release = seq_release,
940 };
941
page_owner_threshold_get(void * data,u64 * val)942 static int page_owner_threshold_get(void *data, u64 *val)
943 {
944 *val = READ_ONCE(page_owner_pages_threshold);
945 return 0;
946 }
947
page_owner_threshold_set(void * data,u64 val)948 static int page_owner_threshold_set(void *data, u64 val)
949 {
950 WRITE_ONCE(page_owner_pages_threshold, val);
951 return 0;
952 }
953
954 DEFINE_SIMPLE_ATTRIBUTE(proc_page_owner_threshold, &page_owner_threshold_get,
955 &page_owner_threshold_set, "%llu");
956
957
pageowner_init(void)958 static int __init pageowner_init(void)
959 {
960 struct dentry *dir;
961
962 if (!static_branch_unlikely(&page_owner_inited)) {
963 pr_info("page_owner is disabled\n");
964 return 0;
965 }
966
967 debugfs_create_file("page_owner", 0400, NULL, NULL,
968 &proc_page_owner_operations);
969 dir = debugfs_create_dir("page_owner_stacks", NULL);
970 debugfs_create_file("show_stacks", 0400, dir, NULL,
971 &page_owner_stack_operations);
972 debugfs_create_file("count_threshold", 0600, dir, NULL,
973 &proc_page_owner_threshold);
974
975 return 0;
976 }
977 late_initcall(pageowner_init)
978