?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
scsi: ufs: Fix RX_TERMINATION_FORCE_ENABLE define value
[linux-stable.git] / mm / khugepaged.c
blobacf66c5dc9bd969321759bb0d5ddf499b06b8e63
1 // SPDX-License-Identifier: GPL-2.0
2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
3
4 #include <linux/mm.h>
5 #include <linux/sched.h>
6 #include <linux/sched/mm.h>
7 #include <linux/sched/coredump.h>
8 #include <linux/mmu_notifier.h>
9 #include <linux/rmap.h>
10 #include <linux/swap.h>
11 #include <linux/mm_inline.h>
12 #include <linux/kthread.h>
13 #include <linux/khugepaged.h>
14 #include <linux/freezer.h>
15 #include <linux/mman.h>
16 #include <linux/hashtable.h>
17 #include <linux/userfaultfd_k.h>
18 #include <linux/page_idle.h>
19 #include <linux/swapops.h>
20 #include <linux/shmem_fs.h>
21
22 #include <asm/tlb.h>
23 #include <asm/pgalloc.h>
24 #include "internal.h"
25
26 enum scan_result {
27 SCAN_FAIL,
28 SCAN_SUCCEED,
29 SCAN_PMD_NULL,
30 SCAN_EXCEED_NONE_PTE,
31 SCAN_PTE_NON_PRESENT,
32 SCAN_PAGE_RO,
33 SCAN_LACK_REFERENCED_PAGE,
34 SCAN_PAGE_NULL,
35 SCAN_SCAN_ABORT,
36 SCAN_PAGE_COUNT,
37 SCAN_PAGE_LRU,
38 SCAN_PAGE_LOCK,
39 SCAN_PAGE_ANON,
40 SCAN_PAGE_COMPOUND,
41 SCAN_ANY_PROCESS,
42 SCAN_VMA_NULL,
43 SCAN_VMA_CHECK,
44 SCAN_ADDRESS_RANGE,
45 SCAN_SWAP_CACHE_PAGE,
46 SCAN_DEL_PAGE_LRU,
47 SCAN_ALLOC_HUGE_PAGE_FAIL,
48 SCAN_CGROUP_CHARGE_FAIL,
49 SCAN_EXCEED_SWAP_PTE,
50 SCAN_TRUNCATED,
51 };
52
53 #define CREATE_TRACE_POINTS
54 #include <trace/events/huge_memory.h>
55
56 /* default scan 8*512 pte (or vmas) every 30 second */
57 static unsigned int khugepaged_pages_to_scan __read_mostly;
58 static unsigned int khugepaged_pages_collapsed;
59 static unsigned int khugepaged_full_scans;
60 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
61 /* during fragmentation poll the hugepage allocator once every minute */
62 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
63 static unsigned long khugepaged_sleep_expire;
64 static DEFINE_SPINLOCK(khugepaged_mm_lock);
65 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
66 /*
67 * default collapse hugepages if there is at least one pte mapped like
68 * it would have happened if the vma was large enough during page
69 * fault.
70 */
71 static unsigned int khugepaged_max_ptes_none __read_mostly;
72 static unsigned int khugepaged_max_ptes_swap __read_mostly;
73
74 #define MM_SLOTS_HASH_BITS 10
75 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
76
77 static struct kmem_cache *mm_slot_cache __read_mostly;
78
79 /**
80 * struct mm_slot - hash lookup from mm to mm_slot
81 * @hash: hash collision list
82 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
83 * @mm: the mm that this information is valid for
84 */
85 struct mm_slot {
86 struct hlist_node hash;
87 struct list_head mm_node;
88 struct mm_struct *mm;
89 };
90
91 /**
92 * struct khugepaged_scan - cursor for scanning
93 * @mm_head: the head of the mm list to scan
94 * @mm_slot: the current mm_slot we are scanning
95 * @address: the next address inside that to be scanned
96 *
97 * There is only the one khugepaged_scan instance of this cursor structure.
98 */
99 struct khugepaged_scan {
100 struct list_head mm_head;
101 struct mm_slot *mm_slot;
102 unsigned long address;
103 };
104
105 static struct khugepaged_scan khugepaged_scan = {
106 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
107 };
108
109 #ifdef CONFIG_SYSFS
110 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
111 struct kobj_attribute *attr,
112 char *buf)
113 {
114 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
115 }
116
117 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
118 struct kobj_attribute *attr,
119 const char *buf, size_t count)
120 {
121 unsigned long msecs;
122 int err;
123
124 err = kstrtoul(buf, 10, &msecs);
125 if (err || msecs > UINT_MAX)
126 return -EINVAL;
127
128 khugepaged_scan_sleep_millisecs = msecs;
129 khugepaged_sleep_expire = 0;
130 wake_up_interruptible(&khugepaged_wait);
131
132 return count;
133 }
134 static struct kobj_attribute scan_sleep_millisecs_attr =
135 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
136 scan_sleep_millisecs_store);
137
138 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
139 struct kobj_attribute *attr,
140 char *buf)
141 {
142 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
143 }
144
145 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
146 struct kobj_attribute *attr,
147 const char *buf, size_t count)
148 {
149 unsigned long msecs;
150 int err;
151
152 err = kstrtoul(buf, 10, &msecs);
153 if (err || msecs > UINT_MAX)
154 return -EINVAL;
155
156 khugepaged_alloc_sleep_millisecs = msecs;
157 khugepaged_sleep_expire = 0;
158 wake_up_interruptible(&khugepaged_wait);
159
160 return count;
161 }
162 static struct kobj_attribute alloc_sleep_millisecs_attr =
163 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
164 alloc_sleep_millisecs_store);
165
166 static ssize_t pages_to_scan_show(struct kobject *kobj,
167 struct kobj_attribute *attr,
168 char *buf)
169 {
170 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
171 }
172 static ssize_t pages_to_scan_store(struct kobject *kobj,
173 struct kobj_attribute *attr,
174 const char *buf, size_t count)
175 {
176 int err;
177 unsigned long pages;
178
179 err = kstrtoul(buf, 10, &pages);
180 if (err || !pages || pages > UINT_MAX)
181 return -EINVAL;
182
183 khugepaged_pages_to_scan = pages;
184
185 return count;
186 }
187 static struct kobj_attribute pages_to_scan_attr =
188 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
189 pages_to_scan_store);
190
191 static ssize_t pages_collapsed_show(struct kobject *kobj,
192 struct kobj_attribute *attr,
193 char *buf)
194 {
195 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
196 }
197 static struct kobj_attribute pages_collapsed_attr =
198 __ATTR_RO(pages_collapsed);
199
200 static ssize_t full_scans_show(struct kobject *kobj,
201 struct kobj_attribute *attr,
202 char *buf)
203 {
204 return sprintf(buf, "%u\n", khugepaged_full_scans);
205 }
206 static struct kobj_attribute full_scans_attr =
207 __ATTR_RO(full_scans);
208
209 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
210 struct kobj_attribute *attr, char *buf)
211 {
212 return single_hugepage_flag_show(kobj, attr, buf,
213 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
214 }
215 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
216 struct kobj_attribute *attr,
217 const char *buf, size_t count)
218 {
219 return single_hugepage_flag_store(kobj, attr, buf, count,
220 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
221 }
222 static struct kobj_attribute khugepaged_defrag_attr =
223 __ATTR(defrag, 0644, khugepaged_defrag_show,
224 khugepaged_defrag_store);
225
226 /*
227 * max_ptes_none controls if khugepaged should collapse hugepages over
228 * any unmapped ptes in turn potentially increasing the memory
229 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
230 * reduce the available free memory in the system as it
231 * runs. Increasing max_ptes_none will instead potentially reduce the
232 * free memory in the system during the khugepaged scan.
233 */
234 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
235 struct kobj_attribute *attr,
236 char *buf)
237 {
238 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
239 }
240 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
241 struct kobj_attribute *attr,
242 const char *buf, size_t count)
243 {
244 int err;
245 unsigned long max_ptes_none;
246
247 err = kstrtoul(buf, 10, &max_ptes_none);
248 if (err || max_ptes_none > HPAGE_PMD_NR-1)
249 return -EINVAL;
250
251 khugepaged_max_ptes_none = max_ptes_none;
252
253 return count;
254 }
255 static struct kobj_attribute khugepaged_max_ptes_none_attr =
256 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
257 khugepaged_max_ptes_none_store);
258
259 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
260 struct kobj_attribute *attr,
261 char *buf)
262 {
263 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
264 }
265
266 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
267 struct kobj_attribute *attr,
268 const char *buf, size_t count)
269 {
270 int err;
271 unsigned long max_ptes_swap;
272
273 err = kstrtoul(buf, 10, &max_ptes_swap);
274 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
275 return -EINVAL;
276
277 khugepaged_max_ptes_swap = max_ptes_swap;
278
279 return count;
280 }
281
282 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
283 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
284 khugepaged_max_ptes_swap_store);
285
286 static struct attribute *khugepaged_attr[] = {
287 &khugepaged_defrag_attr.attr,
288 &khugepaged_max_ptes_none_attr.attr,
289 &pages_to_scan_attr.attr,
290 &pages_collapsed_attr.attr,
291 &full_scans_attr.attr,
292 &scan_sleep_millisecs_attr.attr,
293 &alloc_sleep_millisecs_attr.attr,
294 &khugepaged_max_ptes_swap_attr.attr,
295 NULL,
296 };
297
298 struct attribute_group khugepaged_attr_group = {
299 .attrs = khugepaged_attr,
300 .name = "khugepaged",
301 };
302 #endif /* CONFIG_SYSFS */
303
304 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
305
306 int hugepage_madvise(struct vm_area_struct *vma,
307 unsigned long *vm_flags, int advice)
308 {
309 switch (advice) {
310 case MADV_HUGEPAGE:
311 #ifdef CONFIG_S390
312 /*
313 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
314 * can't handle this properly after s390_enable_sie, so we simply
315 * ignore the madvise to prevent qemu from causing a SIGSEGV.
316 */
317 if (mm_has_pgste(vma->vm_mm))
318 return 0;
319 #endif
320 *vm_flags &= ~VM_NOHUGEPAGE;
321 *vm_flags |= VM_HUGEPAGE;
322 /*
323 * If the vma become good for khugepaged to scan,
324 * register it here without waiting a page fault that
325 * may not happen any time soon.
326 */
327 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
328 khugepaged_enter_vma_merge(vma, *vm_flags))
329 return -ENOMEM;
330 break;
331 case MADV_NOHUGEPAGE:
332 *vm_flags &= ~VM_HUGEPAGE;
333 *vm_flags |= VM_NOHUGEPAGE;
334 /*
335 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
336 * this vma even if we leave the mm registered in khugepaged if
337 * it got registered before VM_NOHUGEPAGE was set.
338 */
339 break;
340 }
341
342 return 0;
343 }
344
345 int __init khugepaged_init(void)
346 {
347 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
348 sizeof(struct mm_slot),
349 __alignof__(struct mm_slot), 0, NULL);
350 if (!mm_slot_cache)
351 return -ENOMEM;
352
353 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
354 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
355 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
356
357 return 0;
358 }
359
360 void __init khugepaged_destroy(void)
361 {
362 kmem_cache_destroy(mm_slot_cache);
363 }
364
365 static inline struct mm_slot *alloc_mm_slot(void)
366 {
367 if (!mm_slot_cache) /* initialization failed */
368 return NULL;
369 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
370 }
371
372 static inline void free_mm_slot(struct mm_slot *mm_slot)
373 {
374 kmem_cache_free(mm_slot_cache, mm_slot);
375 }
376
377 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
378 {
379 struct mm_slot *mm_slot;
380
381 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
382 if (mm == mm_slot->mm)
383 return mm_slot;
384
385 return NULL;
386 }
387
388 static void insert_to_mm_slots_hash(struct mm_struct *mm,
389 struct mm_slot *mm_slot)
390 {
391 mm_slot->mm = mm;
392 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
393 }
394
395 static inline int khugepaged_test_exit(struct mm_struct *mm)
396 {
397 return atomic_read(&mm->mm_users) == 0;
398 }
399
400 int __khugepaged_enter(struct mm_struct *mm)
401 {
402 struct mm_slot *mm_slot;
403 int wakeup;
404
405 mm_slot = alloc_mm_slot();
406 if (!mm_slot)
407 return -ENOMEM;
408
409 /* __khugepaged_exit() must not run from under us */
410 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
411 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
412 free_mm_slot(mm_slot);
413 return 0;
414 }
415
416 spin_lock(&khugepaged_mm_lock);
417 insert_to_mm_slots_hash(mm, mm_slot);
418 /*
419 * Insert just behind the scanning cursor, to let the area settle
420 * down a little.
421 */
422 wakeup = list_empty(&khugepaged_scan.mm_head);
423 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
424 spin_unlock(&khugepaged_mm_lock);
425
426 mmgrab(mm);
427 if (wakeup)
428 wake_up_interruptible(&khugepaged_wait);
429
430 return 0;
431 }
432
433 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
434 unsigned long vm_flags)
435 {
436 unsigned long hstart, hend;
437 if (!vma->anon_vma)
438 /*
439 * Not yet faulted in so we will register later in the
440 * page fault if needed.
441 */
442 return 0;
443 if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
444 /* khugepaged not yet working on file or special mappings */
445 return 0;
446 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
447 hend = vma->vm_end & HPAGE_PMD_MASK;
448 if (hstart < hend)
449 return khugepaged_enter(vma, vm_flags);
450 return 0;
451 }
452
453 void __khugepaged_exit(struct mm_struct *mm)
454 {
455 struct mm_slot *mm_slot;
456 int free = 0;
457
458 spin_lock(&khugepaged_mm_lock);
459 mm_slot = get_mm_slot(mm);
460 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
461 hash_del(&mm_slot->hash);
462 list_del(&mm_slot->mm_node);
463 free = 1;
464 }
465 spin_unlock(&khugepaged_mm_lock);
466
467 if (free) {
468 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
469 free_mm_slot(mm_slot);
470 mmdrop(mm);
471 } else if (mm_slot) {
472 /*
473 * This is required to serialize against
474 * khugepaged_test_exit() (which is guaranteed to run
475 * under mmap sem read mode). Stop here (after we
476 * return all pagetables will be destroyed) until
477 * khugepaged has finished working on the pagetables
478 * under the mmap_sem.
479 */
480 down_write(&mm->mmap_sem);
481 up_write(&mm->mmap_sem);
482 }
483 }
484
485 static void release_pte_page(struct page *page)
486 {
487 dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
488 unlock_page(page);
489 putback_lru_page(page);
490 }
491
492 static void release_pte_pages(pte_t *pte, pte_t *_pte)
493 {
494 while (--_pte >= pte) {
495 pte_t pteval = *_pte;
496 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
497 release_pte_page(pte_page(pteval));
498 }
499 }
500
501 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
502 unsigned long address,
503 pte_t *pte)
504 {
505 struct page *page = NULL;
506 pte_t *_pte;
507 int none_or_zero = 0, result = 0, referenced = 0;
508 bool writable = false;
509
510 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
511 _pte++, address += PAGE_SIZE) {
512 pte_t pteval = *_pte;
513 if (pte_none(pteval) || (pte_present(pteval) &&
514 is_zero_pfn(pte_pfn(pteval)))) {
515 if (!userfaultfd_armed(vma) &&
516 ++none_or_zero <= khugepaged_max_ptes_none) {
517 continue;
518 } else {
519 result = SCAN_EXCEED_NONE_PTE;
520 goto out;
521 }
522 }
523 if (!pte_present(pteval)) {
524 result = SCAN_PTE_NON_PRESENT;
525 goto out;
526 }
527 page = vm_normal_page(vma, address, pteval);
528 if (unlikely(!page)) {
529 result = SCAN_PAGE_NULL;
530 goto out;
531 }
532
533 /* TODO: teach khugepaged to collapse THP mapped with pte */
534 if (PageCompound(page)) {
535 result = SCAN_PAGE_COMPOUND;
536 goto out;
537 }
538
539 VM_BUG_ON_PAGE(!PageAnon(page), page);
540
541 /*
542 * We can do it before isolate_lru_page because the
543 * page can't be freed from under us. NOTE: PG_lock
544 * is needed to serialize against split_huge_page
545 * when invoked from the VM.
546 */
547 if (!trylock_page(page)) {
548 result = SCAN_PAGE_LOCK;
549 goto out;
550 }
551
552 /*
553 * cannot use mapcount: can't collapse if there's a gup pin.
554 * The page must only be referenced by the scanned process
555 * and page swap cache.
556 */
557 if (page_count(page) != 1 + PageSwapCache(page)) {
558 unlock_page(page);
559 result = SCAN_PAGE_COUNT;
560 goto out;
561 }
562 if (pte_write(pteval)) {
563 writable = true;
564 } else {
565 if (PageSwapCache(page) &&
566 !reuse_swap_page(page, NULL)) {
567 unlock_page(page);
568 result = SCAN_SWAP_CACHE_PAGE;
569 goto out;
570 }
571 /*
572 * Page is not in the swap cache. It can be collapsed
573 * into a THP.
574 */
575 }
576
577 /*
578 * Isolate the page to avoid collapsing an hugepage
579 * currently in use by the VM.
580 */
581 if (isolate_lru_page(page)) {
582 unlock_page(page);
583 result = SCAN_DEL_PAGE_LRU;
584 goto out;
585 }
586 inc_node_page_state(page,
587 NR_ISOLATED_ANON + page_is_file_cache(page));
588 VM_BUG_ON_PAGE(!PageLocked(page), page);
589 VM_BUG_ON_PAGE(PageLRU(page), page);
590
591 /* There should be enough young pte to collapse the page */
592 if (pte_young(pteval) ||
593 page_is_young(page) || PageReferenced(page) ||
594 mmu_notifier_test_young(vma->vm_mm, address))
595 referenced++;
596 }
597 if (likely(writable)) {
598 if (likely(referenced)) {
599 result = SCAN_SUCCEED;
600 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
601 referenced, writable, result);
602 return 1;
603 }
604 } else {
605 result = SCAN_PAGE_RO;
606 }
607
608 out:
609 release_pte_pages(pte, _pte);
610 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
611 referenced, writable, result);
612 return 0;
613 }
614
615 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
616 struct vm_area_struct *vma,
617 unsigned long address,
618 spinlock_t *ptl)
619 {
620 pte_t *_pte;
621 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
622 _pte++, page++, address += PAGE_SIZE) {
623 pte_t pteval = *_pte;
624 struct page *src_page;
625
626 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
627 clear_user_highpage(page, address);
628 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
629 if (is_zero_pfn(pte_pfn(pteval))) {
630 /*
631 * ptl mostly unnecessary.
632 */
633 spin_lock(ptl);
634 /*
635 * paravirt calls inside pte_clear here are
636 * superfluous.
637 */
638 pte_clear(vma->vm_mm, address, _pte);
639 spin_unlock(ptl);
640 }
641 } else {
642 src_page = pte_page(pteval);
643 copy_user_highpage(page, src_page, address, vma);
644 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
645 release_pte_page(src_page);
646 /*
647 * ptl mostly unnecessary, but preempt has to
648 * be disabled to update the per-cpu stats
649 * inside page_remove_rmap().
650 */
651 spin_lock(ptl);
652 /*
653 * paravirt calls inside pte_clear here are
654 * superfluous.
655 */
656 pte_clear(vma->vm_mm, address, _pte);
657 page_remove_rmap(src_page, false);
658 spin_unlock(ptl);
659 free_page_and_swap_cache(src_page);
660 }
661 }
662 }
663
664 static void khugepaged_alloc_sleep(void)
665 {
666 DEFINE_WAIT(wait);
667
668 add_wait_queue(&khugepaged_wait, &wait);
669 freezable_schedule_timeout_interruptible(
670 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
671 remove_wait_queue(&khugepaged_wait, &wait);
672 }
673
674 static int khugepaged_node_load[MAX_NUMNODES];
675
676 static bool khugepaged_scan_abort(int nid)
677 {
678 int i;
679
680 /*
681 * If node_reclaim_mode is disabled, then no extra effort is made to
682 * allocate memory locally.
683 */
684 if (!node_reclaim_mode)
685 return false;
686
687 /* If there is a count for this node already, it must be acceptable */
688 if (khugepaged_node_load[nid])
689 return false;
690
691 for (i = 0; i < MAX_NUMNODES; i++) {
692 if (!khugepaged_node_load[i])
693 continue;
694 if (node_distance(nid, i) > RECLAIM_DISTANCE)
695 return true;
696 }
697 return false;
698 }
699
700 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
701 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
702 {
703 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
704 }
705
706 #ifdef CONFIG_NUMA
707 static int khugepaged_find_target_node(void)
708 {
709 static int last_khugepaged_target_node = NUMA_NO_NODE;
710 int nid, target_node = 0, max_value = 0;
711
712 /* find first node with max normal pages hit */
713 for (nid = 0; nid < MAX_NUMNODES; nid++)
714 if (khugepaged_node_load[nid] > max_value) {
715 max_value = khugepaged_node_load[nid];
716 target_node = nid;
717 }
718
719 /* do some balance if several nodes have the same hit record */
720 if (target_node <= last_khugepaged_target_node)
721 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
722 nid++)
723 if (max_value == khugepaged_node_load[nid]) {
724 target_node = nid;
725 break;
726 }
727
728 last_khugepaged_target_node = target_node;
729 return target_node;
730 }
731
732 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
733 {
734 if (IS_ERR(*hpage)) {
735 if (!*wait)
736 return false;
737
738 *wait = false;
739 *hpage = NULL;
740 khugepaged_alloc_sleep();
741 } else if (*hpage) {
742 put_page(*hpage);
743 *hpage = NULL;
744 }
745
746 return true;
747 }
748
749 static struct page *
750 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
751 {
752 VM_BUG_ON_PAGE(*hpage, *hpage);
753
754 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
755 if (unlikely(!*hpage)) {
756 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
757 *hpage = ERR_PTR(-ENOMEM);
758 return NULL;
759 }
760
761 prep_transhuge_page(*hpage);
762 count_vm_event(THP_COLLAPSE_ALLOC);
763 return *hpage;
764 }
765 #else
766 static int khugepaged_find_target_node(void)
767 {
768 return 0;
769 }
770
771 static inline struct page *alloc_khugepaged_hugepage(void)
772 {
773 struct page *page;
774
775 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
776 HPAGE_PMD_ORDER);
777 if (page)
778 prep_transhuge_page(page);
779 return page;
780 }
781
782 static struct page *khugepaged_alloc_hugepage(bool *wait)
783 {
784 struct page *hpage;
785
786 do {
787 hpage = alloc_khugepaged_hugepage();
788 if (!hpage) {
789 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
790 if (!*wait)
791 return NULL;
792
793 *wait = false;
794 khugepaged_alloc_sleep();
795 } else
796 count_vm_event(THP_COLLAPSE_ALLOC);
797 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
798
799 return hpage;
800 }
801
802 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
803 {
804 if (!*hpage)
805 *hpage = khugepaged_alloc_hugepage(wait);
806
807 if (unlikely(!*hpage))
808 return false;
809
810 return true;
811 }
812
813 static struct page *
814 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
815 {
816 VM_BUG_ON(!*hpage);
817
818 return *hpage;
819 }
820 #endif
821
822 static bool hugepage_vma_check(struct vm_area_struct *vma)
823 {
824 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
825 (vma->vm_flags & VM_NOHUGEPAGE) ||
826 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
827 return false;
828 if (shmem_file(vma->vm_file)) {
829 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
830 return false;
831 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
832 HPAGE_PMD_NR);
833 }
834 if (!vma->anon_vma || vma->vm_ops)
835 return false;
836 if (is_vma_temporary_stack(vma))
837 return false;
838 return !(vma->vm_flags & VM_NO_KHUGEPAGED);
839 }
840
841 /*
842 * If mmap_sem temporarily dropped, revalidate vma
843 * before taking mmap_sem.
844 * Return 0 if succeeds, otherwise return none-zero
845 * value (scan code).
846 */
847
848 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
849 struct vm_area_struct **vmap)
850 {
851 struct vm_area_struct *vma;
852 unsigned long hstart, hend;
853
854 if (unlikely(khugepaged_test_exit(mm)))
855 return SCAN_ANY_PROCESS;
856
857 *vmap = vma = find_vma(mm, address);
858 if (!vma)
859 return SCAN_VMA_NULL;
860
861 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
862 hend = vma->vm_end & HPAGE_PMD_MASK;
863 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
864 return SCAN_ADDRESS_RANGE;
865 if (!hugepage_vma_check(vma))
866 return SCAN_VMA_CHECK;
867 return 0;
868 }
869
870 /*
871 * Bring missing pages in from swap, to complete THP collapse.
872 * Only done if khugepaged_scan_pmd believes it is worthwhile.
873 *
874 * Called and returns without pte mapped or spinlocks held,
875 * but with mmap_sem held to protect against vma changes.
876 */
877
878 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
879 struct vm_area_struct *vma,
880 unsigned long address, pmd_t *pmd,
881 int referenced)
882 {
883 int swapped_in = 0, ret = 0;
884 struct vm_fault vmf = {
885 .vma = vma,
886 .address = address,
887 .flags = FAULT_FLAG_ALLOW_RETRY,
888 .pmd = pmd,
889 .pgoff = linear_page_index(vma, address),
890 };
891
892 /* we only decide to swapin, if there is enough young ptes */
893 if (referenced < HPAGE_PMD_NR/2) {
894 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
895 return false;
896 }
897 vmf.pte = pte_offset_map(pmd, address);
898 for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
899 vmf.pte++, vmf.address += PAGE_SIZE) {
900 vmf.orig_pte = *vmf.pte;
901 if (!is_swap_pte(vmf.orig_pte))
902 continue;
903 swapped_in++;
904 ret = do_swap_page(&vmf);
905
906 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
907 if (ret & VM_FAULT_RETRY) {
908 down_read(&mm->mmap_sem);
909 if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
910 /* vma is no longer available, don't continue to swapin */
911 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
912 return false;
913 }
914 /* check if the pmd is still valid */
915 if (mm_find_pmd(mm, address) != pmd) {
916 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
917 return false;
918 }
919 }
920 if (ret & VM_FAULT_ERROR) {
921 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
922 return false;
923 }
924 /* pte is unmapped now, we need to map it */
925 vmf.pte = pte_offset_map(pmd, vmf.address);
926 }
927 vmf.pte--;
928 pte_unmap(vmf.pte);
929 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
930 return true;
931 }
932
933 static void collapse_huge_page(struct mm_struct *mm,
934 unsigned long address,
935 struct page **hpage,
936 int node, int referenced)
937 {
938 pmd_t *pmd, _pmd;
939 pte_t *pte;
940 pgtable_t pgtable;
941 struct page *new_page;
942 spinlock_t *pmd_ptl, *pte_ptl;
943 int isolated = 0, result = 0;
944 struct mem_cgroup *memcg;
945 struct vm_area_struct *vma;
946 unsigned long mmun_start; /* For mmu_notifiers */
947 unsigned long mmun_end; /* For mmu_notifiers */
948 gfp_t gfp;
949
950 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
951
952 /* Only allocate from the target node */
953 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
954
955 /*
956 * Before allocating the hugepage, release the mmap_sem read lock.
957 * The allocation can take potentially a long time if it involves
958 * sync compaction, and we do not need to hold the mmap_sem during
959 * that. We will recheck the vma after taking it again in write mode.
960 */
961 up_read(&mm->mmap_sem);
962 new_page = khugepaged_alloc_page(hpage, gfp, node);
963 if (!new_page) {
964 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
965 goto out_nolock;
966 }
967
968 /* Do not oom kill for khugepaged charges */
969 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp | __GFP_NORETRY,
970 &memcg, true))) {
971 result = SCAN_CGROUP_CHARGE_FAIL;
972 goto out_nolock;
973 }
974
975 down_read(&mm->mmap_sem);
976 result = hugepage_vma_revalidate(mm, address, &vma);
977 if (result) {
978 mem_cgroup_cancel_charge(new_page, memcg, true);
979 up_read(&mm->mmap_sem);
980 goto out_nolock;
981 }
982
983 pmd = mm_find_pmd(mm, address);
984 if (!pmd) {
985 result = SCAN_PMD_NULL;
986 mem_cgroup_cancel_charge(new_page, memcg, true);
987 up_read(&mm->mmap_sem);
988 goto out_nolock;
989 }
990
991 /*
992 * __collapse_huge_page_swapin always returns with mmap_sem locked.
993 * If it fails, we release mmap_sem and jump out_nolock.
994 * Continuing to collapse causes inconsistency.
995 */
996 if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
997 mem_cgroup_cancel_charge(new_page, memcg, true);
998 up_read(&mm->mmap_sem);
999 goto out_nolock;
1000 }
1001
1002 up_read(&mm->mmap_sem);
1003 /*
1004 * Prevent all access to pagetables with the exception of
1005 * gup_fast later handled by the ptep_clear_flush and the VM
1006 * handled by the anon_vma lock + PG_lock.
1007 */
1008 down_write(&mm->mmap_sem);
1009 result = SCAN_ANY_PROCESS;
1010 if (!mmget_still_valid(mm))
1011 goto out;
1012 result = hugepage_vma_revalidate(mm, address, &vma);
1013 if (result)
1014 goto out;
1015 /* check if the pmd is still valid */
1016 if (mm_find_pmd(mm, address) != pmd)
1017 goto out;
1018
1019 anon_vma_lock_write(vma->anon_vma);
1020
1021 pte = pte_offset_map(pmd, address);
1022 pte_ptl = pte_lockptr(mm, pmd);
1023
1024 mmun_start = address;
1025 mmun_end = address + HPAGE_PMD_SIZE;
1026 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1027 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1028 /*
1029 * After this gup_fast can't run anymore. This also removes
1030 * any huge TLB entry from the CPU so we won't allow
1031 * huge and small TLB entries for the same virtual address
1032 * to avoid the risk of CPU bugs in that area.
1033 */
1034 _pmd = pmdp_collapse_flush(vma, address, pmd);
1035 spin_unlock(pmd_ptl);
1036 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1037
1038 spin_lock(pte_ptl);
1039 isolated = __collapse_huge_page_isolate(vma, address, pte);
1040 spin_unlock(pte_ptl);
1041
1042 if (unlikely(!isolated)) {
1043 pte_unmap(pte);
1044 spin_lock(pmd_ptl);
1045 BUG_ON(!pmd_none(*pmd));
1046 /*
1047 * We can only use set_pmd_at when establishing
1048 * hugepmds and never for establishing regular pmds that
1049 * points to regular pagetables. Use pmd_populate for that
1050 */
1051 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1052 spin_unlock(pmd_ptl);
1053 anon_vma_unlock_write(vma->anon_vma);
1054 result = SCAN_FAIL;
1055 goto out;
1056 }
1057
1058 /*
1059 * All pages are isolated and locked so anon_vma rmap
1060 * can't run anymore.
1061 */
1062 anon_vma_unlock_write(vma->anon_vma);
1063
1064 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1065 pte_unmap(pte);
1066 __SetPageUptodate(new_page);
1067 pgtable = pmd_pgtable(_pmd);
1068
1069 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1070 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1071
1072 /*
1073 * spin_lock() below is not the equivalent of smp_wmb(), so
1074 * this is needed to avoid the copy_huge_page writes to become
1075 * visible after the set_pmd_at() write.
1076 */
1077 smp_wmb();
1078
1079 spin_lock(pmd_ptl);
1080 BUG_ON(!pmd_none(*pmd));
1081 page_add_new_anon_rmap(new_page, vma, address, true);
1082 mem_cgroup_commit_charge(new_page, memcg, false, true);
1083 lru_cache_add_active_or_unevictable(new_page, vma);
1084 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1085 set_pmd_at(mm, address, pmd, _pmd);
1086 update_mmu_cache_pmd(vma, address, pmd);
1087 spin_unlock(pmd_ptl);
1088
1089 *hpage = NULL;
1090
1091 khugepaged_pages_collapsed++;
1092 result = SCAN_SUCCEED;
1093 out_up_write:
1094 up_write(&mm->mmap_sem);
1095 out_nolock:
1096 trace_mm_collapse_huge_page(mm, isolated, result);
1097 return;
1098 out:
1099 mem_cgroup_cancel_charge(new_page, memcg, true);
1100 goto out_up_write;
1101 }
1102
1103 static int khugepaged_scan_pmd(struct mm_struct *mm,
1104 struct vm_area_struct *vma,
1105 unsigned long address,
1106 struct page **hpage)
1107 {
1108 pmd_t *pmd;
1109 pte_t *pte, *_pte;
1110 int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1111 struct page *page = NULL;
1112 unsigned long _address;
1113 spinlock_t *ptl;
1114 int node = NUMA_NO_NODE, unmapped = 0;
1115 bool writable = false;
1116
1117 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1118
1119 pmd = mm_find_pmd(mm, address);
1120 if (!pmd) {
1121 result = SCAN_PMD_NULL;
1122 goto out;
1123 }
1124
1125 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1126 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1127 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1128 _pte++, _address += PAGE_SIZE) {
1129 pte_t pteval = *_pte;
1130 if (is_swap_pte(pteval)) {
1131 if (++unmapped <= khugepaged_max_ptes_swap) {
1132 continue;
1133 } else {
1134 result = SCAN_EXCEED_SWAP_PTE;
1135 goto out_unmap;
1136 }
1137 }
1138 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1139 if (!userfaultfd_armed(vma) &&
1140 ++none_or_zero <= khugepaged_max_ptes_none) {
1141 continue;
1142 } else {
1143 result = SCAN_EXCEED_NONE_PTE;
1144 goto out_unmap;
1145 }
1146 }
1147 if (!pte_present(pteval)) {
1148 result = SCAN_PTE_NON_PRESENT;
1149 goto out_unmap;
1150 }
1151 if (pte_write(pteval))
1152 writable = true;
1153
1154 page = vm_normal_page(vma, _address, pteval);
1155 if (unlikely(!page)) {
1156 result = SCAN_PAGE_NULL;
1157 goto out_unmap;
1158 }
1159
1160 /* TODO: teach khugepaged to collapse THP mapped with pte */
1161 if (PageCompound(page)) {
1162 result = SCAN_PAGE_COMPOUND;
1163 goto out_unmap;
1164 }
1165
1166 /*
1167 * Record which node the original page is from and save this
1168 * information to khugepaged_node_load[].
1169 * Khupaged will allocate hugepage from the node has the max
1170 * hit record.
1171 */
1172 node = page_to_nid(page);
1173 if (khugepaged_scan_abort(node)) {
1174 result = SCAN_SCAN_ABORT;
1175 goto out_unmap;
1176 }
1177 khugepaged_node_load[node]++;
1178 if (!PageLRU(page)) {
1179 result = SCAN_PAGE_LRU;
1180 goto out_unmap;
1181 }
1182 if (PageLocked(page)) {
1183 result = SCAN_PAGE_LOCK;
1184 goto out_unmap;
1185 }
1186 if (!PageAnon(page)) {
1187 result = SCAN_PAGE_ANON;
1188 goto out_unmap;
1189 }
1190
1191 /*
1192 * cannot use mapcount: can't collapse if there's a gup pin.
1193 * The page must only be referenced by the scanned process
1194 * and page swap cache.
1195 */
1196 if (page_count(page) != 1 + PageSwapCache(page)) {
1197 result = SCAN_PAGE_COUNT;
1198 goto out_unmap;
1199 }
1200 if (pte_young(pteval) ||
1201 page_is_young(page) || PageReferenced(page) ||
1202 mmu_notifier_test_young(vma->vm_mm, address))
1203 referenced++;
1204 }
1205 if (writable) {
1206 if (referenced) {
1207 result = SCAN_SUCCEED;
1208 ret = 1;
1209 } else {
1210 result = SCAN_LACK_REFERENCED_PAGE;
1211 }
1212 } else {
1213 result = SCAN_PAGE_RO;
1214 }
1215 out_unmap:
1216 pte_unmap_unlock(pte, ptl);
1217 if (ret) {
1218 node = khugepaged_find_target_node();
1219 /* collapse_huge_page will return with the mmap_sem released */
1220 collapse_huge_page(mm, address, hpage, node, referenced);
1221 }
1222 out:
1223 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1224 none_or_zero, result, unmapped);
1225 return ret;
1226 }
1227
1228 static void collect_mm_slot(struct mm_slot *mm_slot)
1229 {
1230 struct mm_struct *mm = mm_slot->mm;
1231
1232 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1233
1234 if (khugepaged_test_exit(mm)) {
1235 /* free mm_slot */
1236 hash_del(&mm_slot->hash);
1237 list_del(&mm_slot->mm_node);
1238
1239 /*
1240 * Not strictly needed because the mm exited already.
1241 *
1242 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1243 */
1244
1245 /* khugepaged_mm_lock actually not necessary for the below */
1246 free_mm_slot(mm_slot);
1247 mmdrop(mm);
1248 }
1249 }
1250
1251 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1252 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1253 {
1254 struct vm_area_struct *vma;
1255 unsigned long addr;
1256 pmd_t *pmd, _pmd;
1257
1258 i_mmap_lock_write(mapping);
1259 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1260 /* probably overkill */
1261 if (vma->anon_vma)
1262 continue;
1263 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1264 if (addr & ~HPAGE_PMD_MASK)
1265 continue;
1266 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1267 continue;
1268 pmd = mm_find_pmd(vma->vm_mm, addr);
1269 if (!pmd)
1270 continue;
1271 /*
1272 * We need exclusive mmap_sem to retract page table.
1273 * If trylock fails we would end up with pte-mapped THP after
1274 * re-fault. Not ideal, but it's more important to not disturb
1275 * the system too much.
1276 */
1277 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1278 spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1279 /* assume page table is clear */
1280 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1281 spin_unlock(ptl);
1282 up_write(&vma->vm_mm->mmap_sem);
1283 atomic_long_dec(&vma->vm_mm->nr_ptes);
1284 pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1285 }
1286 }
1287 i_mmap_unlock_write(mapping);
1288 }
1289
1290 /**
1291 * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1292 *
1293 * Basic scheme is simple, details are more complex:
1294 * - allocate and lock a new huge page;
1295 * - scan over radix tree replacing old pages the new one
1296 * + swap in pages if necessary;
1297 * + fill in gaps;
1298 * + keep old pages around in case if rollback is required;
1299 * - if replacing succeed:
1300 * + copy data over;
1301 * + free old pages;
1302 * + unlock huge page;
1303 * - if replacing failed;
1304 * + put all pages back and unfreeze them;
1305 * + restore gaps in the radix-tree;
1306 * + unlock and free huge page;
1307 */
1308 static void collapse_shmem(struct mm_struct *mm,
1309 struct address_space *mapping, pgoff_t start,
1310 struct page **hpage, int node)
1311 {
1312 gfp_t gfp;
1313 struct page *page, *new_page, *tmp;
1314 struct mem_cgroup *memcg;
1315 pgoff_t index, end = start + HPAGE_PMD_NR;
1316 LIST_HEAD(pagelist);
1317 struct radix_tree_iter iter;
1318 void **slot;
1319 int nr_none = 0, result = SCAN_SUCCEED;
1320
1321 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1322
1323 /* Only allocate from the target node */
1324 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1325
1326 new_page = khugepaged_alloc_page(hpage, gfp, node);
1327 if (!new_page) {
1328 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1329 goto out;
1330 }
1331
1332 /* Do not oom kill for khugepaged charges */
1333 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp | __GFP_NORETRY,
1334 &memcg, true))) {
1335 result = SCAN_CGROUP_CHARGE_FAIL;
1336 goto out;
1337 }
1338
1339 __SetPageLocked(new_page);
1340 __SetPageSwapBacked(new_page);
1341 new_page->index = start;
1342 new_page->mapping = mapping;
1343
1344 /*
1345 * At this point the new_page is locked and not up-to-date.
1346 * It's safe to insert it into the page cache, because nobody would
1347 * be able to map it or use it in another way until we unlock it.
1348 */
1349
1350 index = start;
1351 spin_lock_irq(&mapping->tree_lock);
1352 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1353 int n = min(iter.index, end) - index;
1354
1355 /*
1356 * Stop if extent has been hole-punched, and is now completely
1357 * empty (the more obvious i_size_read() check would take an
1358 * irq-unsafe seqlock on 32-bit).
1359 */
1360 if (n >= HPAGE_PMD_NR) {
1361 result = SCAN_TRUNCATED;
1362 goto tree_locked;
1363 }
1364
1365 /*
1366 * Handle holes in the radix tree: charge it from shmem and
1367 * insert relevant subpage of new_page into the radix-tree.
1368 */
1369 if (n && !shmem_charge(mapping->host, n)) {
1370 result = SCAN_FAIL;
1371 goto tree_locked;
1372 }
1373 for (; index < min(iter.index, end); index++) {
1374 radix_tree_insert(&mapping->page_tree, index,
1375 new_page + (index % HPAGE_PMD_NR));
1376 }
1377 nr_none += n;
1378
1379 /* We are done. */
1380 if (index >= end)
1381 break;
1382
1383 page = radix_tree_deref_slot_protected(slot,
1384 &mapping->tree_lock);
1385 if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1386 spin_unlock_irq(&mapping->tree_lock);
1387 /* swap in or instantiate fallocated page */
1388 if (shmem_getpage(mapping->host, index, &page,
1389 SGP_NOHUGE)) {
1390 result = SCAN_FAIL;
1391 goto tree_unlocked;
1392 }
1393 } else if (trylock_page(page)) {
1394 get_page(page);
1395 spin_unlock_irq(&mapping->tree_lock);
1396 } else {
1397 result = SCAN_PAGE_LOCK;
1398 goto tree_locked;
1399 }
1400
1401 /*
1402 * The page must be locked, so we can drop the tree_lock
1403 * without racing with truncate.
1404 */
1405 VM_BUG_ON_PAGE(!PageLocked(page), page);
1406 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1407
1408 /*
1409 * If file was truncated then extended, or hole-punched, before
1410 * we locked the first page, then a THP might be there already.
1411 */
1412 if (PageTransCompound(page)) {
1413 result = SCAN_PAGE_COMPOUND;
1414 goto out_unlock;
1415 }
1416
1417 if (page_mapping(page) != mapping) {
1418 result = SCAN_TRUNCATED;
1419 goto out_unlock;
1420 }
1421
1422 if (isolate_lru_page(page)) {
1423 result = SCAN_DEL_PAGE_LRU;
1424 goto out_unlock;
1425 }
1426
1427 if (page_mapped(page))
1428 unmap_mapping_range(mapping, index << PAGE_SHIFT,
1429 PAGE_SIZE, 0);
1430
1431 spin_lock_irq(&mapping->tree_lock);
1432
1433 slot = radix_tree_lookup_slot(&mapping->page_tree, index);
1434 VM_BUG_ON_PAGE(page != radix_tree_deref_slot_protected(slot,
1435 &mapping->tree_lock), page);
1436 VM_BUG_ON_PAGE(page_mapped(page), page);
1437
1438 /*
1439 * The page is expected to have page_count() == 3:
1440 * - we hold a pin on it;
1441 * - one reference from radix tree;
1442 * - one from isolate_lru_page;
1443 */
1444 if (!page_ref_freeze(page, 3)) {
1445 result = SCAN_PAGE_COUNT;
1446 spin_unlock_irq(&mapping->tree_lock);
1447 putback_lru_page(page);
1448 goto out_unlock;
1449 }
1450
1451 /*
1452 * Add the page to the list to be able to undo the collapse if
1453 * something go wrong.
1454 */
1455 list_add_tail(&page->lru, &pagelist);
1456
1457 /* Finally, replace with the new page. */
1458 radix_tree_replace_slot(&mapping->page_tree, slot,
1459 new_page + (index % HPAGE_PMD_NR));
1460
1461 slot = radix_tree_iter_resume(slot, &iter);
1462 index++;
1463 continue;
1464 out_unlock:
1465 unlock_page(page);
1466 put_page(page);
1467 goto tree_unlocked;
1468 }
1469
1470 /*
1471 * Handle hole in radix tree at the end of the range.
1472 * This code only triggers if there's nothing in radix tree
1473 * beyond 'end'.
1474 */
1475 if (index < end) {
1476 int n = end - index;
1477
1478 /* Stop if extent has been truncated, and is now empty */
1479 if (n >= HPAGE_PMD_NR) {
1480 result = SCAN_TRUNCATED;
1481 goto tree_locked;
1482 }
1483 if (!shmem_charge(mapping->host, n)) {
1484 result = SCAN_FAIL;
1485 goto tree_locked;
1486 }
1487 for (; index < end; index++) {
1488 radix_tree_insert(&mapping->page_tree, index,
1489 new_page + (index % HPAGE_PMD_NR));
1490 }
1491 nr_none += n;
1492 }
1493
1494 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1495 if (nr_none) {
1496 struct zone *zone = page_zone(new_page);
1497
1498 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1499 __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1500 }
1501
1502 tree_locked:
1503 spin_unlock_irq(&mapping->tree_lock);
1504 tree_unlocked:
1505
1506 if (result == SCAN_SUCCEED) {
1507 /*
1508 * Replacing old pages with new one has succeed, now we need to
1509 * copy the content and free old pages.
1510 */
1511 index = start;
1512 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1513 while (index < page->index) {
1514 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1515 index++;
1516 }
1517 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1518 page);
1519 list_del(&page->lru);
1520 page->mapping = NULL;
1521 page_ref_unfreeze(page, 1);
1522 ClearPageActive(page);
1523 ClearPageUnevictable(page);
1524 unlock_page(page);
1525 put_page(page);
1526 index++;
1527 }
1528 while (index < end) {
1529 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1530 index++;
1531 }
1532
1533 SetPageUptodate(new_page);
1534 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1535 set_page_dirty(new_page);
1536 mem_cgroup_commit_charge(new_page, memcg, false, true);
1537 lru_cache_add_anon(new_page);
1538
1539 /*
1540 * Remove pte page tables, so we can re-fault the page as huge.
1541 */
1542 retract_page_tables(mapping, start);
1543 *hpage = NULL;
1544 } else {
1545 /* Something went wrong: rollback changes to the radix-tree */
1546 spin_lock_irq(&mapping->tree_lock);
1547 mapping->nrpages -= nr_none;
1548 shmem_uncharge(mapping->host, nr_none);
1549
1550 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
1551 start) {
1552 if (iter.index >= end)
1553 break;
1554 page = list_first_entry_or_null(&pagelist,
1555 struct page, lru);
1556 if (!page || iter.index < page->index) {
1557 if (!nr_none)
1558 break;
1559 nr_none--;
1560 /* Put holes back where they were */
1561 radix_tree_delete(&mapping->page_tree,
1562 iter.index);
1563 continue;
1564 }
1565
1566 VM_BUG_ON_PAGE(page->index != iter.index, page);
1567
1568 /* Unfreeze the page. */
1569 list_del(&page->lru);
1570 page_ref_unfreeze(page, 2);
1571 radix_tree_replace_slot(&mapping->page_tree,
1572 slot, page);
1573 slot = radix_tree_iter_resume(slot, &iter);
1574 spin_unlock_irq(&mapping->tree_lock);
1575 unlock_page(page);
1576 putback_lru_page(page);
1577 spin_lock_irq(&mapping->tree_lock);
1578 }
1579 VM_BUG_ON(nr_none);
1580 spin_unlock_irq(&mapping->tree_lock);
1581
1582 mem_cgroup_cancel_charge(new_page, memcg, true);
1583 new_page->mapping = NULL;
1584 }
1585
1586 unlock_page(new_page);
1587 out:
1588 VM_BUG_ON(!list_empty(&pagelist));
1589 /* TODO: tracepoints */
1590 }
1591
1592 static void khugepaged_scan_shmem(struct mm_struct *mm,
1593 struct address_space *mapping,
1594 pgoff_t start, struct page **hpage)
1595 {
1596 struct page *page = NULL;
1597 struct radix_tree_iter iter;
1598 void **slot;
1599 int present, swap;
1600 int node = NUMA_NO_NODE;
1601 int result = SCAN_SUCCEED;
1602
1603 present = 0;
1604 swap = 0;
1605 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1606 rcu_read_lock();
1607 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1608 if (iter.index >= start + HPAGE_PMD_NR)
1609 break;
1610
1611 page = radix_tree_deref_slot(slot);
1612 if (radix_tree_deref_retry(page)) {
1613 slot = radix_tree_iter_retry(&iter);
1614 continue;
1615 }
1616
1617 if (radix_tree_exception(page)) {
1618 if (++swap > khugepaged_max_ptes_swap) {
1619 result = SCAN_EXCEED_SWAP_PTE;
1620 break;
1621 }
1622 continue;
1623 }
1624
1625 if (PageTransCompound(page)) {
1626 result = SCAN_PAGE_COMPOUND;
1627 break;
1628 }
1629
1630 node = page_to_nid(page);
1631 if (khugepaged_scan_abort(node)) {
1632 result = SCAN_SCAN_ABORT;
1633 break;
1634 }
1635 khugepaged_node_load[node]++;
1636
1637 if (!PageLRU(page)) {
1638 result = SCAN_PAGE_LRU;
1639 break;
1640 }
1641
1642 if (page_count(page) != 1 + page_mapcount(page)) {
1643 result = SCAN_PAGE_COUNT;
1644 break;
1645 }
1646
1647 /*
1648 * We probably should check if the page is referenced here, but
1649 * nobody would transfer pte_young() to PageReferenced() for us.
1650 * And rmap walk here is just too costly...
1651 */
1652
1653 present++;
1654
1655 if (need_resched()) {
1656 slot = radix_tree_iter_resume(slot, &iter);
1657 cond_resched_rcu();
1658 }
1659 }
1660 rcu_read_unlock();
1661
1662 if (result == SCAN_SUCCEED) {
1663 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1664 result = SCAN_EXCEED_NONE_PTE;
1665 } else {
1666 node = khugepaged_find_target_node();
1667 collapse_shmem(mm, mapping, start, hpage, node);
1668 }
1669 }
1670
1671 /* TODO: tracepoints */
1672 }
1673 #else
1674 static void khugepaged_scan_shmem(struct mm_struct *mm,
1675 struct address_space *mapping,
1676 pgoff_t start, struct page **hpage)
1677 {
1678 BUILD_BUG();
1679 }
1680 #endif
1681
1682 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1683 struct page **hpage)
1684 __releases(&khugepaged_mm_lock)
1685 __acquires(&khugepaged_mm_lock)
1686 {
1687 struct mm_slot *mm_slot;
1688 struct mm_struct *mm;
1689 struct vm_area_struct *vma;
1690 int progress = 0;
1691
1692 VM_BUG_ON(!pages);
1693 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1694
1695 if (khugepaged_scan.mm_slot)
1696 mm_slot = khugepaged_scan.mm_slot;
1697 else {
1698 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1699 struct mm_slot, mm_node);
1700 khugepaged_scan.address = 0;
1701 khugepaged_scan.mm_slot = mm_slot;
1702 }
1703 spin_unlock(&khugepaged_mm_lock);
1704
1705 mm = mm_slot->mm;
1706 /*
1707 * Don't wait for semaphore (to avoid long wait times). Just move to
1708 * the next mm on the list.
1709 */
1710 vma = NULL;
1711 if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1712 goto breakouterloop_mmap_sem;
1713 if (likely(!khugepaged_test_exit(mm)))
1714 vma = find_vma(mm, khugepaged_scan.address);
1715
1716 progress++;
1717 for (; vma; vma = vma->vm_next) {
1718 unsigned long hstart, hend;
1719
1720 cond_resched();
1721 if (unlikely(khugepaged_test_exit(mm))) {
1722 progress++;
1723 break;
1724 }
1725 if (!hugepage_vma_check(vma)) {
1726 skip:
1727 progress++;
1728 continue;
1729 }
1730 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1731 hend = vma->vm_end & HPAGE_PMD_MASK;
1732 if (hstart >= hend)
1733 goto skip;
1734 if (khugepaged_scan.address > hend)
1735 goto skip;
1736 if (khugepaged_scan.address < hstart)
1737 khugepaged_scan.address = hstart;
1738 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1739
1740 while (khugepaged_scan.address < hend) {
1741 int ret;
1742 cond_resched();
1743 if (unlikely(khugepaged_test_exit(mm)))
1744 goto breakouterloop;
1745
1746 VM_BUG_ON(khugepaged_scan.address < hstart ||
1747 khugepaged_scan.address + HPAGE_PMD_SIZE >
1748 hend);
1749 if (shmem_file(vma->vm_file)) {
1750 struct file *file;
1751 pgoff_t pgoff = linear_page_index(vma,
1752 khugepaged_scan.address);
1753 if (!shmem_huge_enabled(vma))
1754 goto skip;
1755 file = get_file(vma->vm_file);
1756 up_read(&mm->mmap_sem);
1757 ret = 1;
1758 khugepaged_scan_shmem(mm, file->f_mapping,
1759 pgoff, hpage);
1760 fput(file);
1761 } else {
1762 ret = khugepaged_scan_pmd(mm, vma,
1763 khugepaged_scan.address,
1764 hpage);
1765 }
1766 /* move to next address */
1767 khugepaged_scan.address += HPAGE_PMD_SIZE;
1768 progress += HPAGE_PMD_NR;
1769 if (ret)
1770 /* we released mmap_sem so break loop */
1771 goto breakouterloop_mmap_sem;
1772 if (progress >= pages)
1773 goto breakouterloop;
1774 }
1775 }
1776 breakouterloop:
1777 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1778 breakouterloop_mmap_sem:
1779
1780 spin_lock(&khugepaged_mm_lock);
1781 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1782 /*
1783 * Release the current mm_slot if this mm is about to die, or
1784 * if we scanned all vmas of this mm.
1785 */
1786 if (khugepaged_test_exit(mm) || !vma) {
1787 /*
1788 * Make sure that if mm_users is reaching zero while
1789 * khugepaged runs here, khugepaged_exit will find
1790 * mm_slot not pointing to the exiting mm.
1791 */
1792 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1793 khugepaged_scan.mm_slot = list_entry(
1794 mm_slot->mm_node.next,
1795 struct mm_slot, mm_node);
1796 khugepaged_scan.address = 0;
1797 } else {
1798 khugepaged_scan.mm_slot = NULL;
1799 khugepaged_full_scans++;
1800 }
1801
1802 collect_mm_slot(mm_slot);
1803 }
1804
1805 return progress;
1806 }
1807
1808 static int khugepaged_has_work(void)
1809 {
1810 return !list_empty(&khugepaged_scan.mm_head) &&
1811 khugepaged_enabled();
1812 }
1813
1814 static int khugepaged_wait_event(void)
1815 {
1816 return !list_empty(&khugepaged_scan.mm_head) ||
1817 kthread_should_stop();
1818 }
1819
1820 static void khugepaged_do_scan(void)
1821 {
1822 struct page *hpage = NULL;
1823 unsigned int progress = 0, pass_through_head = 0;
1824 unsigned int pages = khugepaged_pages_to_scan;
1825 bool wait = true;
1826
1827 barrier(); /* write khugepaged_pages_to_scan to local stack */
1828
1829 while (progress < pages) {
1830 if (!khugepaged_prealloc_page(&hpage, &wait))
1831 break;
1832
1833 cond_resched();
1834
1835 if (unlikely(kthread_should_stop() || try_to_freeze()))
1836 break;
1837
1838 spin_lock(&khugepaged_mm_lock);
1839 if (!khugepaged_scan.mm_slot)
1840 pass_through_head++;
1841 if (khugepaged_has_work() &&
1842 pass_through_head < 2)
1843 progress += khugepaged_scan_mm_slot(pages - progress,
1844 &hpage);
1845 else
1846 progress = pages;
1847 spin_unlock(&khugepaged_mm_lock);
1848 }
1849
1850 if (!IS_ERR_OR_NULL(hpage))
1851 put_page(hpage);
1852 }
1853
1854 static bool khugepaged_should_wakeup(void)
1855 {
1856 return kthread_should_stop() ||
1857 time_after_eq(jiffies, khugepaged_sleep_expire);
1858 }
1859
1860 static void khugepaged_wait_work(void)
1861 {
1862 if (khugepaged_has_work()) {
1863 const unsigned long scan_sleep_jiffies =
1864 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1865
1866 if (!scan_sleep_jiffies)
1867 return;
1868
1869 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1870 wait_event_freezable_timeout(khugepaged_wait,
1871 khugepaged_should_wakeup(),
1872 scan_sleep_jiffies);
1873 return;
1874 }
1875
1876 if (khugepaged_enabled())
1877 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1878 }
1879
1880 static int khugepaged(void *none)
1881 {
1882 struct mm_slot *mm_slot;
1883
1884 set_freezable();
1885 set_user_nice(current, MAX_NICE);
1886
1887 while (!kthread_should_stop()) {
1888 khugepaged_do_scan();
1889 khugepaged_wait_work();
1890 }
1891
1892 spin_lock(&khugepaged_mm_lock);
1893 mm_slot = khugepaged_scan.mm_slot;
1894 khugepaged_scan.mm_slot = NULL;
1895 if (mm_slot)
1896 collect_mm_slot(mm_slot);
1897 spin_unlock(&khugepaged_mm_lock);
1898 return 0;
1899 }
1900
1901 static void set_recommended_min_free_kbytes(void)
1902 {
1903 struct zone *zone;
1904 int nr_zones = 0;
1905 unsigned long recommended_min;
1906
1907 for_each_populated_zone(zone)
1908 nr_zones++;
1909
1910 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1911 recommended_min = pageblock_nr_pages * nr_zones * 2;
1912
1913 /*
1914 * Make sure that on average at least two pageblocks are almost free
1915 * of another type, one for a migratetype to fall back to and a
1916 * second to avoid subsequent fallbacks of other types There are 3
1917 * MIGRATE_TYPES we care about.
1918 */
1919 recommended_min += pageblock_nr_pages * nr_zones *
1920 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1921
1922 /* don't ever allow to reserve more than 5% of the lowmem */
1923 recommended_min = min(recommended_min,
1924 (unsigned long) nr_free_buffer_pages() / 20);
1925 recommended_min <<= (PAGE_SHIFT-10);
1926
1927 if (recommended_min > min_free_kbytes) {
1928 if (user_min_free_kbytes >= 0)
1929 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1930 min_free_kbytes, recommended_min);
1931
1932 min_free_kbytes = recommended_min;
1933 }
1934 setup_per_zone_wmarks();
1935 }
1936
1937 int start_stop_khugepaged(void)
1938 {
1939 static struct task_struct *khugepaged_thread __read_mostly;
1940 static DEFINE_MUTEX(khugepaged_mutex);
1941 int err = 0;
1942
1943 mutex_lock(&khugepaged_mutex);
1944 if (khugepaged_enabled()) {
1945 if (!khugepaged_thread)
1946 khugepaged_thread = kthread_run(khugepaged, NULL,
1947 "khugepaged");
1948 if (IS_ERR(khugepaged_thread)) {
1949 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1950 err = PTR_ERR(khugepaged_thread);
1951 khugepaged_thread = NULL;
1952 goto fail;
1953 }
1954
1955 if (!list_empty(&khugepaged_scan.mm_head))
1956 wake_up_interruptible(&khugepaged_wait);
1957
1958 set_recommended_min_free_kbytes();
1959 } else if (khugepaged_thread) {
1960 kthread_stop(khugepaged_thread);
1961 khugepaged_thread = NULL;
1962 }
1963 fail:
1964 mutex_unlock(&khugepaged_mutex);
1965 return err;
1966 }
Linux kernel stable tree