summaryrefslogtreecommitdiff
path: root/mm/khugepaged.c (plain)
blob: e0cfc3a54b6a64b66c6d5f2e775f48356da5bf11
1#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
2
3#include <linux/mm.h>
4#include <linux/sched.h>
5#include <linux/mmu_notifier.h>
6#include <linux/rmap.h>
7#include <linux/swap.h>
8#include <linux/mm_inline.h>
9#include <linux/kthread.h>
10#include <linux/khugepaged.h>
11#include <linux/freezer.h>
12#include <linux/mman.h>
13#include <linux/hashtable.h>
14#include <linux/userfaultfd_k.h>
15#include <linux/page_idle.h>
16#include <linux/swapops.h>
17#include <linux/shmem_fs.h>
18
19#include <asm/tlb.h>
20#include <asm/pgalloc.h>
21#include "internal.h"
22
23enum scan_result {
24 SCAN_FAIL,
25 SCAN_SUCCEED,
26 SCAN_PMD_NULL,
27 SCAN_EXCEED_NONE_PTE,
28 SCAN_PTE_NON_PRESENT,
29 SCAN_PAGE_RO,
30 SCAN_LACK_REFERENCED_PAGE,
31 SCAN_PAGE_NULL,
32 SCAN_SCAN_ABORT,
33 SCAN_PAGE_COUNT,
34 SCAN_PAGE_LRU,
35 SCAN_PAGE_LOCK,
36 SCAN_PAGE_ANON,
37 SCAN_PAGE_COMPOUND,
38 SCAN_ANY_PROCESS,
39 SCAN_VMA_NULL,
40 SCAN_VMA_CHECK,
41 SCAN_ADDRESS_RANGE,
42 SCAN_SWAP_CACHE_PAGE,
43 SCAN_DEL_PAGE_LRU,
44 SCAN_ALLOC_HUGE_PAGE_FAIL,
45 SCAN_CGROUP_CHARGE_FAIL,
46 SCAN_EXCEED_SWAP_PTE,
47 SCAN_TRUNCATED,
48};
49
50#define CREATE_TRACE_POINTS
51#include <trace/events/huge_memory.h>
52
53/* default scan 8*512 pte (or vmas) every 30 second */
54static unsigned int khugepaged_pages_to_scan __read_mostly;
55static unsigned int khugepaged_pages_collapsed;
56static unsigned int khugepaged_full_scans;
57static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
58/* during fragmentation poll the hugepage allocator once every minute */
59static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
60static unsigned long khugepaged_sleep_expire;
61static DEFINE_SPINLOCK(khugepaged_mm_lock);
62static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
63/*
64 * default collapse hugepages if there is at least one pte mapped like
65 * it would have happened if the vma was large enough during page
66 * fault.
67 */
68static unsigned int khugepaged_max_ptes_none __read_mostly;
69static unsigned int khugepaged_max_ptes_swap __read_mostly;
70
71#define MM_SLOTS_HASH_BITS 10
72static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
73
74static struct kmem_cache *mm_slot_cache __read_mostly;
75
76/**
77 * struct mm_slot - hash lookup from mm to mm_slot
78 * @hash: hash collision list
79 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
80 * @mm: the mm that this information is valid for
81 */
82struct mm_slot {
83 struct hlist_node hash;
84 struct list_head mm_node;
85 struct mm_struct *mm;
86};
87
88/**
89 * struct khugepaged_scan - cursor for scanning
90 * @mm_head: the head of the mm list to scan
91 * @mm_slot: the current mm_slot we are scanning
92 * @address: the next address inside that to be scanned
93 *
94 * There is only the one khugepaged_scan instance of this cursor structure.
95 */
96struct khugepaged_scan {
97 struct list_head mm_head;
98 struct mm_slot *mm_slot;
99 unsigned long address;
100};
101
102static struct khugepaged_scan khugepaged_scan = {
103 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
104};
105
106#ifdef CONFIG_SYSFS
107static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
108 struct kobj_attribute *attr,
109 char *buf)
110{
111 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
112}
113
114static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
115 struct kobj_attribute *attr,
116 const char *buf, size_t count)
117{
118 unsigned long msecs;
119 int err;
120
121 err = kstrtoul(buf, 10, &msecs);
122 if (err || msecs > UINT_MAX)
123 return -EINVAL;
124
125 khugepaged_scan_sleep_millisecs = msecs;
126 khugepaged_sleep_expire = 0;
127 wake_up_interruptible(&khugepaged_wait);
128
129 return count;
130}
131static struct kobj_attribute scan_sleep_millisecs_attr =
132 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
133 scan_sleep_millisecs_store);
134
135static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
136 struct kobj_attribute *attr,
137 char *buf)
138{
139 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
140}
141
142static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
143 struct kobj_attribute *attr,
144 const char *buf, size_t count)
145{
146 unsigned long msecs;
147 int err;
148
149 err = kstrtoul(buf, 10, &msecs);
150 if (err || msecs > UINT_MAX)
151 return -EINVAL;
152
153 khugepaged_alloc_sleep_millisecs = msecs;
154 khugepaged_sleep_expire = 0;
155 wake_up_interruptible(&khugepaged_wait);
156
157 return count;
158}
159static struct kobj_attribute alloc_sleep_millisecs_attr =
160 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
161 alloc_sleep_millisecs_store);
162
163static ssize_t pages_to_scan_show(struct kobject *kobj,
164 struct kobj_attribute *attr,
165 char *buf)
166{
167 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
168}
169static ssize_t pages_to_scan_store(struct kobject *kobj,
170 struct kobj_attribute *attr,
171 const char *buf, size_t count)
172{
173 int err;
174 unsigned long pages;
175
176 err = kstrtoul(buf, 10, &pages);
177 if (err || !pages || pages > UINT_MAX)
178 return -EINVAL;
179
180 khugepaged_pages_to_scan = pages;
181
182 return count;
183}
184static struct kobj_attribute pages_to_scan_attr =
185 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
186 pages_to_scan_store);
187
188static ssize_t pages_collapsed_show(struct kobject *kobj,
189 struct kobj_attribute *attr,
190 char *buf)
191{
192 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
193}
194static struct kobj_attribute pages_collapsed_attr =
195 __ATTR_RO(pages_collapsed);
196
197static ssize_t full_scans_show(struct kobject *kobj,
198 struct kobj_attribute *attr,
199 char *buf)
200{
201 return sprintf(buf, "%u\n", khugepaged_full_scans);
202}
203static struct kobj_attribute full_scans_attr =
204 __ATTR_RO(full_scans);
205
206static ssize_t khugepaged_defrag_show(struct kobject *kobj,
207 struct kobj_attribute *attr, char *buf)
208{
209 return single_hugepage_flag_show(kobj, attr, buf,
210 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
211}
212static ssize_t khugepaged_defrag_store(struct kobject *kobj,
213 struct kobj_attribute *attr,
214 const char *buf, size_t count)
215{
216 return single_hugepage_flag_store(kobj, attr, buf, count,
217 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
218}
219static struct kobj_attribute khugepaged_defrag_attr =
220 __ATTR(defrag, 0644, khugepaged_defrag_show,
221 khugepaged_defrag_store);
222
223/*
224 * max_ptes_none controls if khugepaged should collapse hugepages over
225 * any unmapped ptes in turn potentially increasing the memory
226 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
227 * reduce the available free memory in the system as it
228 * runs. Increasing max_ptes_none will instead potentially reduce the
229 * free memory in the system during the khugepaged scan.
230 */
231static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
232 struct kobj_attribute *attr,
233 char *buf)
234{
235 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
236}
237static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
238 struct kobj_attribute *attr,
239 const char *buf, size_t count)
240{
241 int err;
242 unsigned long max_ptes_none;
243
244 err = kstrtoul(buf, 10, &max_ptes_none);
245 if (err || max_ptes_none > HPAGE_PMD_NR-1)
246 return -EINVAL;
247
248 khugepaged_max_ptes_none = max_ptes_none;
249
250 return count;
251}
252static struct kobj_attribute khugepaged_max_ptes_none_attr =
253 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
254 khugepaged_max_ptes_none_store);
255
256static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
257 struct kobj_attribute *attr,
258 char *buf)
259{
260 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
261}
262
263static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
264 struct kobj_attribute *attr,
265 const char *buf, size_t count)
266{
267 int err;
268 unsigned long max_ptes_swap;
269
270 err = kstrtoul(buf, 10, &max_ptes_swap);
271 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
272 return -EINVAL;
273
274 khugepaged_max_ptes_swap = max_ptes_swap;
275
276 return count;
277}
278
279static struct kobj_attribute khugepaged_max_ptes_swap_attr =
280 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
281 khugepaged_max_ptes_swap_store);
282
283static struct attribute *khugepaged_attr[] = {
284 &khugepaged_defrag_attr.attr,
285 &khugepaged_max_ptes_none_attr.attr,
286 &pages_to_scan_attr.attr,
287 &pages_collapsed_attr.attr,
288 &full_scans_attr.attr,
289 &scan_sleep_millisecs_attr.attr,
290 &alloc_sleep_millisecs_attr.attr,
291 &khugepaged_max_ptes_swap_attr.attr,
292 NULL,
293};
294
295struct attribute_group khugepaged_attr_group = {
296 .attrs = khugepaged_attr,
297 .name = "khugepaged",
298};
299#endif /* CONFIG_SYSFS */
300
301#define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
302
303int hugepage_madvise(struct vm_area_struct *vma,
304 unsigned long *vm_flags, int advice)
305{
306 switch (advice) {
307 case MADV_HUGEPAGE:
308#ifdef CONFIG_S390
309 /*
310 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
311 * can't handle this properly after s390_enable_sie, so we simply
312 * ignore the madvise to prevent qemu from causing a SIGSEGV.
313 */
314 if (mm_has_pgste(vma->vm_mm))
315 return 0;
316#endif
317 *vm_flags &= ~VM_NOHUGEPAGE;
318 *vm_flags |= VM_HUGEPAGE;
319 /*
320 * If the vma become good for khugepaged to scan,
321 * register it here without waiting a page fault that
322 * may not happen any time soon.
323 */
324 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
325 khugepaged_enter_vma_merge(vma, *vm_flags))
326 return -ENOMEM;
327 break;
328 case MADV_NOHUGEPAGE:
329 *vm_flags &= ~VM_HUGEPAGE;
330 *vm_flags |= VM_NOHUGEPAGE;
331 /*
332 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
333 * this vma even if we leave the mm registered in khugepaged if
334 * it got registered before VM_NOHUGEPAGE was set.
335 */
336 break;
337 }
338
339 return 0;
340}
341
342int __init khugepaged_init(void)
343{
344 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
345 sizeof(struct mm_slot),
346 __alignof__(struct mm_slot), 0, NULL);
347 if (!mm_slot_cache)
348 return -ENOMEM;
349
350 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
351 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
352 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
353
354 return 0;
355}
356
357void __init khugepaged_destroy(void)
358{
359 kmem_cache_destroy(mm_slot_cache);
360}
361
362static inline struct mm_slot *alloc_mm_slot(void)
363{
364 if (!mm_slot_cache) /* initialization failed */
365 return NULL;
366 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
367}
368
369static inline void free_mm_slot(struct mm_slot *mm_slot)
370{
371 kmem_cache_free(mm_slot_cache, mm_slot);
372}
373
374static struct mm_slot *get_mm_slot(struct mm_struct *mm)
375{
376 struct mm_slot *mm_slot;
377
378 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
379 if (mm == mm_slot->mm)
380 return mm_slot;
381
382 return NULL;
383}
384
385static void insert_to_mm_slots_hash(struct mm_struct *mm,
386 struct mm_slot *mm_slot)
387{
388 mm_slot->mm = mm;
389 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
390}
391
392static inline int khugepaged_test_exit(struct mm_struct *mm)
393{
394 return atomic_read(&mm->mm_users) == 0;
395}
396
397int __khugepaged_enter(struct mm_struct *mm)
398{
399 struct mm_slot *mm_slot;
400 int wakeup;
401
402 mm_slot = alloc_mm_slot();
403 if (!mm_slot)
404 return -ENOMEM;
405
406 /* __khugepaged_exit() must not run from under us */
407 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
408 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
409 free_mm_slot(mm_slot);
410 return 0;
411 }
412
413 spin_lock(&khugepaged_mm_lock);
414 insert_to_mm_slots_hash(mm, mm_slot);
415 /*
416 * Insert just behind the scanning cursor, to let the area settle
417 * down a little.
418 */
419 wakeup = list_empty(&khugepaged_scan.mm_head);
420 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
421 spin_unlock(&khugepaged_mm_lock);
422
423 atomic_inc(&mm->mm_count);
424 if (wakeup)
425 wake_up_interruptible(&khugepaged_wait);
426
427 return 0;
428}
429
430int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
431 unsigned long vm_flags)
432{
433 unsigned long hstart, hend;
434 if (!vma->anon_vma)
435 /*
436 * Not yet faulted in so we will register later in the
437 * page fault if needed.
438 */
439 return 0;
440 if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
441 /* khugepaged not yet working on file or special mappings */
442 return 0;
443 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
444 hend = vma->vm_end & HPAGE_PMD_MASK;
445 if (hstart < hend)
446 return khugepaged_enter(vma, vm_flags);
447 return 0;
448}
449
450void __khugepaged_exit(struct mm_struct *mm)
451{
452 struct mm_slot *mm_slot;
453 int free = 0;
454
455 spin_lock(&khugepaged_mm_lock);
456 mm_slot = get_mm_slot(mm);
457 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
458 hash_del(&mm_slot->hash);
459 list_del(&mm_slot->mm_node);
460 free = 1;
461 }
462 spin_unlock(&khugepaged_mm_lock);
463
464 if (free) {
465 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
466 free_mm_slot(mm_slot);
467 mmdrop(mm);
468 } else if (mm_slot) {
469 /*
470 * This is required to serialize against
471 * khugepaged_test_exit() (which is guaranteed to run
472 * under mmap sem read mode). Stop here (after we
473 * return all pagetables will be destroyed) until
474 * khugepaged has finished working on the pagetables
475 * under the mmap_sem.
476 */
477 down_write(&mm->mmap_sem);
478 up_write(&mm->mmap_sem);
479 }
480}
481
482static void release_pte_page(struct page *page)
483{
484 /* 0 stands for page_is_file_cache(page) == false */
485 dec_node_page_state(page, NR_ISOLATED_ANON + 0);
486 unlock_page(page);
487 putback_lru_page(page);
488}
489
490static void release_pte_pages(pte_t *pte, pte_t *_pte)
491{
492 while (--_pte >= pte) {
493 pte_t pteval = *_pte;
494 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
495 release_pte_page(pte_page(pteval));
496 }
497}
498
499static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
500 unsigned long address,
501 pte_t *pte)
502{
503 struct page *page = NULL;
504 pte_t *_pte;
505 int none_or_zero = 0, result = 0, referenced = 0;
506 bool writable = false;
507
508 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
509 _pte++, address += PAGE_SIZE) {
510 pte_t pteval = *_pte;
511 if (pte_none(pteval) || (pte_present(pteval) &&
512 is_zero_pfn(pte_pfn(pteval)))) {
513 if (!userfaultfd_armed(vma) &&
514 ++none_or_zero <= khugepaged_max_ptes_none) {
515 continue;
516 } else {
517 result = SCAN_EXCEED_NONE_PTE;
518 goto out;
519 }
520 }
521 if (!pte_present(pteval)) {
522 result = SCAN_PTE_NON_PRESENT;
523 goto out;
524 }
525 page = vm_normal_page(vma, address, pteval);
526 if (unlikely(!page)) {
527 result = SCAN_PAGE_NULL;
528 goto out;
529 }
530
531 /* TODO: teach khugepaged to collapse THP mapped with pte */
532 if (PageCompound(page)) {
533 result = SCAN_PAGE_COMPOUND;
534 goto out;
535 }
536
537 VM_BUG_ON_PAGE(!PageAnon(page), page);
538 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
539
540 /*
541 * We can do it before isolate_lru_page because the
542 * page can't be freed from under us. NOTE: PG_lock
543 * is needed to serialize against split_huge_page
544 * when invoked from the VM.
545 */
546 if (!trylock_page(page)) {
547 result = SCAN_PAGE_LOCK;
548 goto out;
549 }
550
551 /*
552 * cannot use mapcount: can't collapse if there's a gup pin.
553 * The page must only be referenced by the scanned process
554 * and page swap cache.
555 */
556 if (page_count(page) != 1 + !!PageSwapCache(page)) {
557 unlock_page(page);
558 result = SCAN_PAGE_COUNT;
559 goto out;
560 }
561 if (pte_write(pteval)) {
562 writable = true;
563 } else {
564 if (PageSwapCache(page) &&
565 !reuse_swap_page(page, NULL)) {
566 unlock_page(page);
567 result = SCAN_SWAP_CACHE_PAGE;
568 goto out;
569 }
570 /*
571 * Page is not in the swap cache. It can be collapsed
572 * into a THP.
573 */
574 }
575
576 /*
577 * Isolate the page to avoid collapsing an hugepage
578 * currently in use by the VM.
579 */
580 if (isolate_lru_page(page)) {
581 unlock_page(page);
582 result = SCAN_DEL_PAGE_LRU;
583 goto out;
584 }
585 /* 0 stands for page_is_file_cache(page) == false */
586 inc_node_page_state(page, NR_ISOLATED_ANON + 0);
587 VM_BUG_ON_PAGE(!PageLocked(page), page);
588 VM_BUG_ON_PAGE(PageLRU(page), page);
589
590 /* There should be enough young pte to collapse the page */
591 if (pte_young(pteval) ||
592 page_is_young(page) || PageReferenced(page) ||
593 mmu_notifier_test_young(vma->vm_mm, address))
594 referenced++;
595 }
596 if (likely(writable)) {
597 if (likely(referenced)) {
598 result = SCAN_SUCCEED;
599 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
600 referenced, writable, result);
601 return 1;
602 }
603 } else {
604 result = SCAN_PAGE_RO;
605 }
606
607out:
608 release_pte_pages(pte, _pte);
609 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
610 referenced, writable, result);
611 return 0;
612}
613
614static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
615 struct vm_area_struct *vma,
616 unsigned long address,
617 spinlock_t *ptl)
618{
619 pte_t *_pte;
620 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
621 pte_t pteval = *_pte;
622 struct page *src_page;
623
624 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
625 clear_user_highpage(page, address);
626 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
627 if (is_zero_pfn(pte_pfn(pteval))) {
628 /*
629 * ptl mostly unnecessary.
630 */
631 spin_lock(ptl);
632 /*
633 * paravirt calls inside pte_clear here are
634 * superfluous.
635 */
636 pte_clear(vma->vm_mm, address, _pte);
637 spin_unlock(ptl);
638 }
639 } else {
640 src_page = pte_page(pteval);
641 copy_user_highpage(page, src_page, address, vma);
642 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
643 release_pte_page(src_page);
644 /*
645 * ptl mostly unnecessary, but preempt has to
646 * be disabled to update the per-cpu stats
647 * inside page_remove_rmap().
648 */
649 spin_lock(ptl);
650 /*
651 * paravirt calls inside pte_clear here are
652 * superfluous.
653 */
654 pte_clear(vma->vm_mm, address, _pte);
655 page_remove_rmap(src_page, false);
656 spin_unlock(ptl);
657 free_page_and_swap_cache(src_page);
658 }
659
660 address += PAGE_SIZE;
661 page++;
662 }
663}
664
665static void khugepaged_alloc_sleep(void)
666{
667 DEFINE_WAIT(wait);
668
669 add_wait_queue(&khugepaged_wait, &wait);
670 freezable_schedule_timeout_interruptible(
671 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
672 remove_wait_queue(&khugepaged_wait, &wait);
673}
674
675static int khugepaged_node_load[MAX_NUMNODES];
676
677static bool khugepaged_scan_abort(int nid)
678{
679 int i;
680
681 /*
682 * If node_reclaim_mode is disabled, then no extra effort is made to
683 * allocate memory locally.
684 */
685 if (!node_reclaim_mode)
686 return false;
687
688 /* If there is a count for this node already, it must be acceptable */
689 if (khugepaged_node_load[nid])
690 return false;
691
692 for (i = 0; i < MAX_NUMNODES; i++) {
693 if (!khugepaged_node_load[i])
694 continue;
695 if (node_distance(nid, i) > RECLAIM_DISTANCE)
696 return true;
697 }
698 return false;
699}
700
701/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
702static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
703{
704 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
705}
706
707#ifdef CONFIG_NUMA
708static int khugepaged_find_target_node(void)
709{
710 static int last_khugepaged_target_node = NUMA_NO_NODE;
711 int nid, target_node = 0, max_value = 0;
712
713 /* find first node with max normal pages hit */
714 for (nid = 0; nid < MAX_NUMNODES; nid++)
715 if (khugepaged_node_load[nid] > max_value) {
716 max_value = khugepaged_node_load[nid];
717 target_node = nid;
718 }
719
720 /* do some balance if several nodes have the same hit record */
721 if (target_node <= last_khugepaged_target_node)
722 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
723 nid++)
724 if (max_value == khugepaged_node_load[nid]) {
725 target_node = nid;
726 break;
727 }
728
729 last_khugepaged_target_node = target_node;
730 return target_node;
731}
732
733static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
734{
735 if (IS_ERR(*hpage)) {
736 if (!*wait)
737 return false;
738
739 *wait = false;
740 *hpage = NULL;
741 khugepaged_alloc_sleep();
742 } else if (*hpage) {
743 put_page(*hpage);
744 *hpage = NULL;
745 }
746
747 return true;
748}
749
750static struct page *
751khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
752{
753 VM_BUG_ON_PAGE(*hpage, *hpage);
754
755 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
756 if (unlikely(!*hpage)) {
757 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
758 *hpage = ERR_PTR(-ENOMEM);
759 return NULL;
760 }
761
762 prep_transhuge_page(*hpage);
763 count_vm_event(THP_COLLAPSE_ALLOC);
764 return *hpage;
765}
766#else
767static int khugepaged_find_target_node(void)
768{
769 return 0;
770}
771
772static inline struct page *alloc_khugepaged_hugepage(void)
773{
774 struct page *page;
775
776 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
777 HPAGE_PMD_ORDER);
778 if (page)
779 prep_transhuge_page(page);
780 return page;
781}
782
783static struct page *khugepaged_alloc_hugepage(bool *wait)
784{
785 struct page *hpage;
786
787 do {
788 hpage = alloc_khugepaged_hugepage();
789 if (!hpage) {
790 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
791 if (!*wait)
792 return NULL;
793
794 *wait = false;
795 khugepaged_alloc_sleep();
796 } else
797 count_vm_event(THP_COLLAPSE_ALLOC);
798 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
799
800 return hpage;
801}
802
803static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
804{
805 if (!*hpage)
806 *hpage = khugepaged_alloc_hugepage(wait);
807
808 if (unlikely(!*hpage))
809 return false;
810
811 return true;
812}
813
814static struct page *
815khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
816{
817 VM_BUG_ON(!*hpage);
818
819 return *hpage;
820}
821#endif
822
823static bool hugepage_vma_check(struct vm_area_struct *vma)
824{
825 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
826 (vma->vm_flags & VM_NOHUGEPAGE))
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
848static 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
878static 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 pte_t pteval;
884 int swapped_in = 0, ret = 0;
885 struct fault_env fe = {
886 .vma = vma,
887 .address = address,
888 .flags = FAULT_FLAG_ALLOW_RETRY,
889 .pmd = pmd,
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 fe.pte = pte_offset_map(pmd, address);
898 for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
899 fe.pte++, fe.address += PAGE_SIZE) {
900 pteval = *fe.pte;
901 if (!is_swap_pte(pteval))
902 continue;
903 swapped_in++;
904 ret = do_swap_page(&fe, pteval);
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, &fe.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 return false;
917 }
918 if (ret & VM_FAULT_ERROR) {
919 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
920 return false;
921 }
922 /* pte is unmapped now, we need to map it */
923 fe.pte = pte_offset_map(pmd, fe.address);
924 }
925 fe.pte--;
926 pte_unmap(fe.pte);
927 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
928 return true;
929}
930
931static void collapse_huge_page(struct mm_struct *mm,
932 unsigned long address,
933 struct page **hpage,
934 int node, int referenced)
935{
936 pmd_t *pmd, _pmd;
937 pte_t *pte;
938 pgtable_t pgtable;
939 struct page *new_page;
940 spinlock_t *pmd_ptl, *pte_ptl;
941 int isolated = 0, result = 0;
942 struct mem_cgroup *memcg;
943 struct vm_area_struct *vma;
944 unsigned long mmun_start; /* For mmu_notifiers */
945 unsigned long mmun_end; /* For mmu_notifiers */
946 gfp_t gfp;
947
948 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
949
950 /* Only allocate from the target node */
951 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
952
953 /*
954 * Before allocating the hugepage, release the mmap_sem read lock.
955 * The allocation can take potentially a long time if it involves
956 * sync compaction, and we do not need to hold the mmap_sem during
957 * that. We will recheck the vma after taking it again in write mode.
958 */
959 up_read(&mm->mmap_sem);
960 new_page = khugepaged_alloc_page(hpage, gfp, node);
961 if (!new_page) {
962 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
963 goto out_nolock;
964 }
965
966 /* Do not oom kill for khugepaged charges */
967 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp | __GFP_NORETRY,
968 &memcg, true))) {
969 result = SCAN_CGROUP_CHARGE_FAIL;
970 goto out_nolock;
971 }
972
973 down_read(&mm->mmap_sem);
974 result = hugepage_vma_revalidate(mm, address, &vma);
975 if (result) {
976 mem_cgroup_cancel_charge(new_page, memcg, true);
977 up_read(&mm->mmap_sem);
978 goto out_nolock;
979 }
980
981 pmd = mm_find_pmd(mm, address);
982 if (!pmd) {
983 result = SCAN_PMD_NULL;
984 mem_cgroup_cancel_charge(new_page, memcg, true);
985 up_read(&mm->mmap_sem);
986 goto out_nolock;
987 }
988
989 /*
990 * __collapse_huge_page_swapin always returns with mmap_sem locked.
991 * If it fails, we release mmap_sem and jump out_nolock.
992 * Continuing to collapse causes inconsistency.
993 */
994 if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
995 mem_cgroup_cancel_charge(new_page, memcg, true);
996 up_read(&mm->mmap_sem);
997 goto out_nolock;
998 }
999
1000 up_read(&mm->mmap_sem);
1001 /*
1002 * Prevent all access to pagetables with the exception of
1003 * gup_fast later handled by the ptep_clear_flush and the VM
1004 * handled by the anon_vma lock + PG_lock.
1005 */
1006 down_write(&mm->mmap_sem);
1007 result = hugepage_vma_revalidate(mm, address, &vma);
1008 if (result)
1009 goto out;
1010 /* check if the pmd is still valid */
1011 if (mm_find_pmd(mm, address) != pmd)
1012 goto out;
1013
1014 anon_vma_lock_write(vma->anon_vma);
1015
1016 pte = pte_offset_map(pmd, address);
1017 pte_ptl = pte_lockptr(mm, pmd);
1018
1019 mmun_start = address;
1020 mmun_end = address + HPAGE_PMD_SIZE;
1021 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1022 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1023 /*
1024 * After this gup_fast can't run anymore. This also removes
1025 * any huge TLB entry from the CPU so we won't allow
1026 * huge and small TLB entries for the same virtual address
1027 * to avoid the risk of CPU bugs in that area.
1028 */
1029 _pmd = pmdp_collapse_flush(vma, address, pmd);
1030 spin_unlock(pmd_ptl);
1031 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1032
1033 spin_lock(pte_ptl);
1034 isolated = __collapse_huge_page_isolate(vma, address, pte);
1035 spin_unlock(pte_ptl);
1036
1037 if (unlikely(!isolated)) {
1038 pte_unmap(pte);
1039 spin_lock(pmd_ptl);
1040 BUG_ON(!pmd_none(*pmd));
1041 /*
1042 * We can only use set_pmd_at when establishing
1043 * hugepmds and never for establishing regular pmds that
1044 * points to regular pagetables. Use pmd_populate for that
1045 */
1046 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1047 spin_unlock(pmd_ptl);
1048 anon_vma_unlock_write(vma->anon_vma);
1049 result = SCAN_FAIL;
1050 goto out;
1051 }
1052
1053 /*
1054 * All pages are isolated and locked so anon_vma rmap
1055 * can't run anymore.
1056 */
1057 anon_vma_unlock_write(vma->anon_vma);
1058
1059 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1060 pte_unmap(pte);
1061 __SetPageUptodate(new_page);
1062 pgtable = pmd_pgtable(_pmd);
1063
1064 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1065 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1066
1067 /*
1068 * spin_lock() below is not the equivalent of smp_wmb(), so
1069 * this is needed to avoid the copy_huge_page writes to become
1070 * visible after the set_pmd_at() write.
1071 */
1072 smp_wmb();
1073
1074 spin_lock(pmd_ptl);
1075 BUG_ON(!pmd_none(*pmd));
1076 page_add_new_anon_rmap(new_page, vma, address, true);
1077 mem_cgroup_commit_charge(new_page, memcg, false, true);
1078 lru_cache_add_active_or_unevictable(new_page, vma);
1079 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1080 set_pmd_at(mm, address, pmd, _pmd);
1081 update_mmu_cache_pmd(vma, address, pmd);
1082 spin_unlock(pmd_ptl);
1083
1084 *hpage = NULL;
1085
1086 khugepaged_pages_collapsed++;
1087 result = SCAN_SUCCEED;
1088out_up_write:
1089 up_write(&mm->mmap_sem);
1090out_nolock:
1091 trace_mm_collapse_huge_page(mm, isolated, result);
1092 return;
1093out:
1094 mem_cgroup_cancel_charge(new_page, memcg, true);
1095 goto out_up_write;
1096}
1097
1098static int khugepaged_scan_pmd(struct mm_struct *mm,
1099 struct vm_area_struct *vma,
1100 unsigned long address,
1101 struct page **hpage)
1102{
1103 pmd_t *pmd;
1104 pte_t *pte, *_pte;
1105 int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1106 struct page *page = NULL;
1107 unsigned long _address;
1108 spinlock_t *ptl;
1109 int node = NUMA_NO_NODE, unmapped = 0;
1110 bool writable = false;
1111
1112 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1113
1114 pmd = mm_find_pmd(mm, address);
1115 if (!pmd) {
1116 result = SCAN_PMD_NULL;
1117 goto out;
1118 }
1119
1120 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1121 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1122 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1123 _pte++, _address += PAGE_SIZE) {
1124 pte_t pteval = *_pte;
1125 if (is_swap_pte(pteval)) {
1126 if (++unmapped <= khugepaged_max_ptes_swap) {
1127 continue;
1128 } else {
1129 result = SCAN_EXCEED_SWAP_PTE;
1130 goto out_unmap;
1131 }
1132 }
1133 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1134 if (!userfaultfd_armed(vma) &&
1135 ++none_or_zero <= khugepaged_max_ptes_none) {
1136 continue;
1137 } else {
1138 result = SCAN_EXCEED_NONE_PTE;
1139 goto out_unmap;
1140 }
1141 }
1142 if (!pte_present(pteval)) {
1143 result = SCAN_PTE_NON_PRESENT;
1144 goto out_unmap;
1145 }
1146 if (pte_write(pteval))
1147 writable = true;
1148
1149 page = vm_normal_page(vma, _address, pteval);
1150 if (unlikely(!page)) {
1151 result = SCAN_PAGE_NULL;
1152 goto out_unmap;
1153 }
1154
1155 /* TODO: teach khugepaged to collapse THP mapped with pte */
1156 if (PageCompound(page)) {
1157 result = SCAN_PAGE_COMPOUND;
1158 goto out_unmap;
1159 }
1160
1161 /*
1162 * Record which node the original page is from and save this
1163 * information to khugepaged_node_load[].
1164 * Khupaged will allocate hugepage from the node has the max
1165 * hit record.
1166 */
1167 node = page_to_nid(page);
1168 if (khugepaged_scan_abort(node)) {
1169 result = SCAN_SCAN_ABORT;
1170 goto out_unmap;
1171 }
1172 khugepaged_node_load[node]++;
1173 if (!PageLRU(page)) {
1174 result = SCAN_PAGE_LRU;
1175 goto out_unmap;
1176 }
1177 if (PageLocked(page)) {
1178 result = SCAN_PAGE_LOCK;
1179 goto out_unmap;
1180 }
1181 if (!PageAnon(page)) {
1182 result = SCAN_PAGE_ANON;
1183 goto out_unmap;
1184 }
1185
1186 /*
1187 * cannot use mapcount: can't collapse if there's a gup pin.
1188 * The page must only be referenced by the scanned process
1189 * and page swap cache.
1190 */
1191 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1192 result = SCAN_PAGE_COUNT;
1193 goto out_unmap;
1194 }
1195 if (pte_young(pteval) ||
1196 page_is_young(page) || PageReferenced(page) ||
1197 mmu_notifier_test_young(vma->vm_mm, address))
1198 referenced++;
1199 }
1200 if (writable) {
1201 if (referenced) {
1202 result = SCAN_SUCCEED;
1203 ret = 1;
1204 } else {
1205 result = SCAN_LACK_REFERENCED_PAGE;
1206 }
1207 } else {
1208 result = SCAN_PAGE_RO;
1209 }
1210out_unmap:
1211 pte_unmap_unlock(pte, ptl);
1212 if (ret) {
1213 node = khugepaged_find_target_node();
1214 /* collapse_huge_page will return with the mmap_sem released */
1215 collapse_huge_page(mm, address, hpage, node, referenced);
1216 }
1217out:
1218 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1219 none_or_zero, result, unmapped);
1220 return ret;
1221}
1222
1223static void collect_mm_slot(struct mm_slot *mm_slot)
1224{
1225 struct mm_struct *mm = mm_slot->mm;
1226
1227 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1228
1229 if (khugepaged_test_exit(mm)) {
1230 /* free mm_slot */
1231 hash_del(&mm_slot->hash);
1232 list_del(&mm_slot->mm_node);
1233
1234 /*
1235 * Not strictly needed because the mm exited already.
1236 *
1237 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1238 */
1239
1240 /* khugepaged_mm_lock actually not necessary for the below */
1241 free_mm_slot(mm_slot);
1242 mmdrop(mm);
1243 }
1244}
1245
1246#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1247static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1248{
1249 struct vm_area_struct *vma;
1250 unsigned long addr;
1251 pmd_t *pmd, _pmd;
1252
1253 i_mmap_lock_write(mapping);
1254 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1255 /* probably overkill */
1256 if (vma->anon_vma)
1257 continue;
1258 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1259 if (addr & ~HPAGE_PMD_MASK)
1260 continue;
1261 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1262 continue;
1263 pmd = mm_find_pmd(vma->vm_mm, addr);
1264 if (!pmd)
1265 continue;
1266 /*
1267 * We need exclusive mmap_sem to retract page table.
1268 * If trylock fails we would end up with pte-mapped THP after
1269 * re-fault. Not ideal, but it's more important to not disturb
1270 * the system too much.
1271 */
1272 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1273 spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1274 /* assume page table is clear */
1275 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1276 spin_unlock(ptl);
1277 up_write(&vma->vm_mm->mmap_sem);
1278 atomic_long_dec(&vma->vm_mm->nr_ptes);
1279 pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1280 }
1281 }
1282 i_mmap_unlock_write(mapping);
1283}
1284
1285/**
1286 * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1287 *
1288 * Basic scheme is simple, details are more complex:
1289 * - allocate and lock a new huge page;
1290 * - scan over radix tree replacing old pages the new one
1291 * + swap in pages if necessary;
1292 * + fill in gaps;
1293 * + keep old pages around in case if rollback is required;
1294 * - if replacing succeed:
1295 * + copy data over;
1296 * + free old pages;
1297 * + unlock huge page;
1298 * - if replacing failed;
1299 * + put all pages back and unfreeze them;
1300 * + restore gaps in the radix-tree;
1301 * + unlock and free huge page;
1302 */
1303static void collapse_shmem(struct mm_struct *mm,
1304 struct address_space *mapping, pgoff_t start,
1305 struct page **hpage, int node)
1306{
1307 gfp_t gfp;
1308 struct page *page, *new_page, *tmp;
1309 struct mem_cgroup *memcg;
1310 pgoff_t index, end = start + HPAGE_PMD_NR;
1311 LIST_HEAD(pagelist);
1312 struct radix_tree_iter iter;
1313 void **slot;
1314 int nr_none = 0, result = SCAN_SUCCEED;
1315
1316 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1317
1318 /* Only allocate from the target node */
1319 gfp = alloc_hugepage_khugepaged_gfpmask() |
1320 __GFP_OTHER_NODE | __GFP_THISNODE;
1321
1322 new_page = khugepaged_alloc_page(hpage, gfp, node);
1323 if (!new_page) {
1324 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1325 goto out;
1326 }
1327
1328 /* Do not oom kill for khugepaged charges */
1329 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp | __GFP_NORETRY,
1330 &memcg, true))) {
1331 result = SCAN_CGROUP_CHARGE_FAIL;
1332 goto out;
1333 }
1334
1335 __SetPageLocked(new_page);
1336 __SetPageSwapBacked(new_page);
1337 new_page->index = start;
1338 new_page->mapping = mapping;
1339
1340 /*
1341 * At this point the new_page is locked and not up-to-date.
1342 * It's safe to insert it into the page cache, because nobody would
1343 * be able to map it or use it in another way until we unlock it.
1344 */
1345
1346 index = start;
1347 spin_lock_irq(&mapping->tree_lock);
1348 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1349 int n = min(iter.index, end) - index;
1350
1351 /*
1352 * Stop if extent has been hole-punched, and is now completely
1353 * empty (the more obvious i_size_read() check would take an
1354 * irq-unsafe seqlock on 32-bit).
1355 */
1356 if (n >= HPAGE_PMD_NR) {
1357 result = SCAN_TRUNCATED;
1358 goto tree_locked;
1359 }
1360
1361 /*
1362 * Handle holes in the radix tree: charge it from shmem and
1363 * insert relevant subpage of new_page into the radix-tree.
1364 */
1365 if (n && !shmem_charge(mapping->host, n)) {
1366 result = SCAN_FAIL;
1367 goto tree_locked;
1368 }
1369 for (; index < min(iter.index, end); index++) {
1370 radix_tree_insert(&mapping->page_tree, index,
1371 new_page + (index % HPAGE_PMD_NR));
1372 }
1373 nr_none += n;
1374
1375 /* We are done. */
1376 if (index >= end)
1377 break;
1378
1379 page = radix_tree_deref_slot_protected(slot,
1380 &mapping->tree_lock);
1381 if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1382 spin_unlock_irq(&mapping->tree_lock);
1383 /* swap in or instantiate fallocated page */
1384 if (shmem_getpage(mapping->host, index, &page,
1385 SGP_NOHUGE)) {
1386 result = SCAN_FAIL;
1387 goto tree_unlocked;
1388 }
1389 } else if (trylock_page(page)) {
1390 get_page(page);
1391 spin_unlock_irq(&mapping->tree_lock);
1392 } else {
1393 result = SCAN_PAGE_LOCK;
1394 goto tree_locked;
1395 }
1396
1397 /*
1398 * The page must be locked, so we can drop the tree_lock
1399 * without racing with truncate.
1400 */
1401 VM_BUG_ON_PAGE(!PageLocked(page), page);
1402 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1403
1404 /*
1405 * If file was truncated then extended, or hole-punched, before
1406 * we locked the first page, then a THP might be there already.
1407 */
1408 if (PageTransCompound(page)) {
1409 result = SCAN_PAGE_COMPOUND;
1410 goto out_unlock;
1411 }
1412
1413 if (page_mapping(page) != mapping) {
1414 result = SCAN_TRUNCATED;
1415 goto out_unlock;
1416 }
1417
1418 if (isolate_lru_page(page)) {
1419 result = SCAN_DEL_PAGE_LRU;
1420 goto out_unlock;
1421 }
1422
1423 if (page_mapped(page))
1424 unmap_mapping_range(mapping, index << PAGE_SHIFT,
1425 PAGE_SIZE, 0);
1426
1427 spin_lock_irq(&mapping->tree_lock);
1428
1429 slot = radix_tree_lookup_slot(&mapping->page_tree, index);
1430 VM_BUG_ON_PAGE(page != radix_tree_deref_slot_protected(slot,
1431 &mapping->tree_lock), page);
1432 VM_BUG_ON_PAGE(page_mapped(page), page);
1433
1434 /*
1435 * The page is expected to have page_count() == 3:
1436 * - we hold a pin on it;
1437 * - one reference from radix tree;
1438 * - one from isolate_lru_page;
1439 */
1440 if (!page_ref_freeze(page, 3)) {
1441 result = SCAN_PAGE_COUNT;
1442 spin_unlock_irq(&mapping->tree_lock);
1443 putback_lru_page(page);
1444 goto out_unlock;
1445 }
1446
1447 /*
1448 * Add the page to the list to be able to undo the collapse if
1449 * something go wrong.
1450 */
1451 list_add_tail(&page->lru, &pagelist);
1452
1453 /* Finally, replace with the new page. */
1454 radix_tree_replace_slot(slot,
1455 new_page + (index % HPAGE_PMD_NR));
1456
1457 slot = radix_tree_iter_next(&iter);
1458 index++;
1459 continue;
1460out_unlock:
1461 unlock_page(page);
1462 put_page(page);
1463 goto tree_unlocked;
1464 }
1465
1466 /*
1467 * Handle hole in radix tree at the end of the range.
1468 * This code only triggers if there's nothing in radix tree
1469 * beyond 'end'.
1470 */
1471 if (index < end) {
1472 int n = end - index;
1473
1474 /* Stop if extent has been truncated, and is now empty */
1475 if (n >= HPAGE_PMD_NR) {
1476 result = SCAN_TRUNCATED;
1477 goto tree_locked;
1478 }
1479 if (!shmem_charge(mapping->host, n)) {
1480 result = SCAN_FAIL;
1481 goto tree_locked;
1482 }
1483 for (; index < end; index++) {
1484 radix_tree_insert(&mapping->page_tree, index,
1485 new_page + (index % HPAGE_PMD_NR));
1486 }
1487 nr_none += n;
1488 }
1489
1490 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1491 if (nr_none) {
1492 struct zone *zone = page_zone(new_page);
1493
1494 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1495 __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1496 }
1497
1498tree_locked:
1499 spin_unlock_irq(&mapping->tree_lock);
1500tree_unlocked:
1501
1502 if (result == SCAN_SUCCEED) {
1503 /*
1504 * Replacing old pages with new one has succeed, now we need to
1505 * copy the content and free old pages.
1506 */
1507 index = start;
1508 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1509 while (index < page->index) {
1510 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1511 index++;
1512 }
1513 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1514 page);
1515 list_del(&page->lru);
1516 page->mapping = NULL;
1517 page_ref_unfreeze(page, 1);
1518 ClearPageActive(page);
1519 ClearPageUnevictable(page);
1520 unlock_page(page);
1521 put_page(page);
1522 index++;
1523 }
1524 while (index < end) {
1525 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1526 index++;
1527 }
1528
1529 SetPageUptodate(new_page);
1530 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1531 set_page_dirty(new_page);
1532 mem_cgroup_commit_charge(new_page, memcg, false, true);
1533 lru_cache_add_anon(new_page);
1534
1535 /*
1536 * Remove pte page tables, so we can re-fault the page as huge.
1537 */
1538 retract_page_tables(mapping, start);
1539 *hpage = NULL;
1540 } else {
1541 /* Something went wrong: rollback changes to the radix-tree */
1542 spin_lock_irq(&mapping->tree_lock);
1543 mapping->nrpages -= nr_none;
1544 shmem_uncharge(mapping->host, nr_none);
1545
1546 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
1547 start) {
1548 if (iter.index >= end)
1549 break;
1550 page = list_first_entry_or_null(&pagelist,
1551 struct page, lru);
1552 if (!page || iter.index < page->index) {
1553 if (!nr_none)
1554 break;
1555 nr_none--;
1556 /* Put holes back where they were */
1557 radix_tree_delete(&mapping->page_tree,
1558 iter.index);
1559 slot = radix_tree_iter_next(&iter);
1560 continue;
1561 }
1562
1563 VM_BUG_ON_PAGE(page->index != iter.index, page);
1564
1565 /* Unfreeze the page. */
1566 list_del(&page->lru);
1567 page_ref_unfreeze(page, 2);
1568 radix_tree_replace_slot(slot, page);
1569 spin_unlock_irq(&mapping->tree_lock);
1570 unlock_page(page);
1571 putback_lru_page(page);
1572 spin_lock_irq(&mapping->tree_lock);
1573 slot = radix_tree_iter_next(&iter);
1574 }
1575 VM_BUG_ON(nr_none);
1576 spin_unlock_irq(&mapping->tree_lock);
1577
1578 mem_cgroup_cancel_charge(new_page, memcg, true);
1579 new_page->mapping = NULL;
1580 }
1581
1582 unlock_page(new_page);
1583out:
1584 VM_BUG_ON(!list_empty(&pagelist));
1585 /* TODO: tracepoints */
1586}
1587
1588static void khugepaged_scan_shmem(struct mm_struct *mm,
1589 struct address_space *mapping,
1590 pgoff_t start, struct page **hpage)
1591{
1592 struct page *page = NULL;
1593 struct radix_tree_iter iter;
1594 void **slot;
1595 int present, swap;
1596 int node = NUMA_NO_NODE;
1597 int result = SCAN_SUCCEED;
1598
1599 present = 0;
1600 swap = 0;
1601 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1602 rcu_read_lock();
1603 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1604 if (iter.index >= start + HPAGE_PMD_NR)
1605 break;
1606
1607 page = radix_tree_deref_slot(slot);
1608 if (radix_tree_deref_retry(page)) {
1609 slot = radix_tree_iter_retry(&iter);
1610 continue;
1611 }
1612
1613 if (radix_tree_exception(page)) {
1614 if (++swap > khugepaged_max_ptes_swap) {
1615 result = SCAN_EXCEED_SWAP_PTE;
1616 break;
1617 }
1618 continue;
1619 }
1620
1621 if (PageTransCompound(page)) {
1622 result = SCAN_PAGE_COMPOUND;
1623 break;
1624 }
1625
1626 node = page_to_nid(page);
1627 if (khugepaged_scan_abort(node)) {
1628 result = SCAN_SCAN_ABORT;
1629 break;
1630 }
1631 khugepaged_node_load[node]++;
1632
1633 if (!PageLRU(page)) {
1634 result = SCAN_PAGE_LRU;
1635 break;
1636 }
1637
1638 if (page_count(page) != 1 + page_mapcount(page)) {
1639 result = SCAN_PAGE_COUNT;
1640 break;
1641 }
1642
1643 /*
1644 * We probably should check if the page is referenced here, but
1645 * nobody would transfer pte_young() to PageReferenced() for us.
1646 * And rmap walk here is just too costly...
1647 */
1648
1649 present++;
1650
1651 if (need_resched()) {
1652 cond_resched_rcu();
1653 slot = radix_tree_iter_next(&iter);
1654 }
1655 }
1656 rcu_read_unlock();
1657
1658 if (result == SCAN_SUCCEED) {
1659 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1660 result = SCAN_EXCEED_NONE_PTE;
1661 } else {
1662 node = khugepaged_find_target_node();
1663 collapse_shmem(mm, mapping, start, hpage, node);
1664 }
1665 }
1666
1667 /* TODO: tracepoints */
1668}
1669#else
1670static void khugepaged_scan_shmem(struct mm_struct *mm,
1671 struct address_space *mapping,
1672 pgoff_t start, struct page **hpage)
1673{
1674 BUILD_BUG();
1675}
1676#endif
1677
1678static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1679 struct page **hpage)
1680 __releases(&khugepaged_mm_lock)
1681 __acquires(&khugepaged_mm_lock)
1682{
1683 struct mm_slot *mm_slot;
1684 struct mm_struct *mm;
1685 struct vm_area_struct *vma;
1686 int progress = 0;
1687
1688 VM_BUG_ON(!pages);
1689 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1690
1691 if (khugepaged_scan.mm_slot)
1692 mm_slot = khugepaged_scan.mm_slot;
1693 else {
1694 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1695 struct mm_slot, mm_node);
1696 khugepaged_scan.address = 0;
1697 khugepaged_scan.mm_slot = mm_slot;
1698 }
1699 spin_unlock(&khugepaged_mm_lock);
1700
1701 mm = mm_slot->mm;
1702 /*
1703 * Don't wait for semaphore (to avoid long wait times). Just move to
1704 * the next mm on the list.
1705 */
1706 vma = NULL;
1707 if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1708 goto breakouterloop_mmap_sem;
1709 if (likely(!khugepaged_test_exit(mm)))
1710 vma = find_vma(mm, khugepaged_scan.address);
1711
1712 progress++;
1713 for (; vma; vma = vma->vm_next) {
1714 unsigned long hstart, hend;
1715
1716 cond_resched();
1717 if (unlikely(khugepaged_test_exit(mm))) {
1718 progress++;
1719 break;
1720 }
1721 if (!hugepage_vma_check(vma)) {
1722skip:
1723 progress++;
1724 continue;
1725 }
1726 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1727 hend = vma->vm_end & HPAGE_PMD_MASK;
1728 if (hstart >= hend)
1729 goto skip;
1730 if (khugepaged_scan.address > hend)
1731 goto skip;
1732 if (khugepaged_scan.address < hstart)
1733 khugepaged_scan.address = hstart;
1734 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1735
1736 while (khugepaged_scan.address < hend) {
1737 int ret;
1738 cond_resched();
1739 if (unlikely(khugepaged_test_exit(mm)))
1740 goto breakouterloop;
1741
1742 VM_BUG_ON(khugepaged_scan.address < hstart ||
1743 khugepaged_scan.address + HPAGE_PMD_SIZE >
1744 hend);
1745 if (shmem_file(vma->vm_file)) {
1746 struct file *file;
1747 pgoff_t pgoff = linear_page_index(vma,
1748 khugepaged_scan.address);
1749 if (!shmem_huge_enabled(vma))
1750 goto skip;
1751 file = get_file(vma->vm_file);
1752 up_read(&mm->mmap_sem);
1753 ret = 1;
1754 khugepaged_scan_shmem(mm, file->f_mapping,
1755 pgoff, hpage);
1756 fput(file);
1757 } else {
1758 ret = khugepaged_scan_pmd(mm, vma,
1759 khugepaged_scan.address,
1760 hpage);
1761 }
1762 /* move to next address */
1763 khugepaged_scan.address += HPAGE_PMD_SIZE;
1764 progress += HPAGE_PMD_NR;
1765 if (ret)
1766 /* we released mmap_sem so break loop */
1767 goto breakouterloop_mmap_sem;
1768 if (progress >= pages)
1769 goto breakouterloop;
1770 }
1771 }
1772breakouterloop:
1773 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1774breakouterloop_mmap_sem:
1775
1776 spin_lock(&khugepaged_mm_lock);
1777 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1778 /*
1779 * Release the current mm_slot if this mm is about to die, or
1780 * if we scanned all vmas of this mm.
1781 */
1782 if (khugepaged_test_exit(mm) || !vma) {
1783 /*
1784 * Make sure that if mm_users is reaching zero while
1785 * khugepaged runs here, khugepaged_exit will find
1786 * mm_slot not pointing to the exiting mm.
1787 */
1788 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1789 khugepaged_scan.mm_slot = list_entry(
1790 mm_slot->mm_node.next,
1791 struct mm_slot, mm_node);
1792 khugepaged_scan.address = 0;
1793 } else {
1794 khugepaged_scan.mm_slot = NULL;
1795 khugepaged_full_scans++;
1796 }
1797
1798 collect_mm_slot(mm_slot);
1799 }
1800
1801 return progress;
1802}
1803
1804static int khugepaged_has_work(void)
1805{
1806 return !list_empty(&khugepaged_scan.mm_head) &&
1807 khugepaged_enabled();
1808}
1809
1810static int khugepaged_wait_event(void)
1811{
1812 return !list_empty(&khugepaged_scan.mm_head) ||
1813 kthread_should_stop();
1814}
1815
1816static void khugepaged_do_scan(void)
1817{
1818 struct page *hpage = NULL;
1819 unsigned int progress = 0, pass_through_head = 0;
1820 unsigned int pages = khugepaged_pages_to_scan;
1821 bool wait = true;
1822
1823 barrier(); /* write khugepaged_pages_to_scan to local stack */
1824
1825 while (progress < pages) {
1826 if (!khugepaged_prealloc_page(&hpage, &wait))
1827 break;
1828
1829 cond_resched();
1830
1831 if (unlikely(kthread_should_stop() || try_to_freeze()))
1832 break;
1833
1834 spin_lock(&khugepaged_mm_lock);
1835 if (!khugepaged_scan.mm_slot)
1836 pass_through_head++;
1837 if (khugepaged_has_work() &&
1838 pass_through_head < 2)
1839 progress += khugepaged_scan_mm_slot(pages - progress,
1840 &hpage);
1841 else
1842 progress = pages;
1843 spin_unlock(&khugepaged_mm_lock);
1844 }
1845
1846 if (!IS_ERR_OR_NULL(hpage))
1847 put_page(hpage);
1848}
1849
1850static bool khugepaged_should_wakeup(void)
1851{
1852 return kthread_should_stop() ||
1853 time_after_eq(jiffies, khugepaged_sleep_expire);
1854}
1855
1856static void khugepaged_wait_work(void)
1857{
1858 if (khugepaged_has_work()) {
1859 const unsigned long scan_sleep_jiffies =
1860 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1861
1862 if (!scan_sleep_jiffies)
1863 return;
1864
1865 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1866 wait_event_freezable_timeout(khugepaged_wait,
1867 khugepaged_should_wakeup(),
1868 scan_sleep_jiffies);
1869 return;
1870 }
1871
1872 if (khugepaged_enabled())
1873 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1874}
1875
1876static int khugepaged(void *none)
1877{
1878 struct mm_slot *mm_slot;
1879
1880 set_freezable();
1881 set_user_nice(current, MAX_NICE);
1882
1883 while (!kthread_should_stop()) {
1884 khugepaged_do_scan();
1885 khugepaged_wait_work();
1886 }
1887
1888 spin_lock(&khugepaged_mm_lock);
1889 mm_slot = khugepaged_scan.mm_slot;
1890 khugepaged_scan.mm_slot = NULL;
1891 if (mm_slot)
1892 collect_mm_slot(mm_slot);
1893 spin_unlock(&khugepaged_mm_lock);
1894 return 0;
1895}
1896
1897static void set_recommended_min_free_kbytes(void)
1898{
1899 struct zone *zone;
1900 int nr_zones = 0;
1901 unsigned long recommended_min;
1902
1903 for_each_populated_zone(zone)
1904 nr_zones++;
1905
1906 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1907 recommended_min = pageblock_nr_pages * nr_zones * 2;
1908
1909 /*
1910 * Make sure that on average at least two pageblocks are almost free
1911 * of another type, one for a migratetype to fall back to and a
1912 * second to avoid subsequent fallbacks of other types There are 3
1913 * MIGRATE_TYPES we care about.
1914 */
1915 recommended_min += pageblock_nr_pages * nr_zones *
1916 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1917
1918 /* don't ever allow to reserve more than 5% of the lowmem */
1919 recommended_min = min(recommended_min,
1920 (unsigned long) nr_free_buffer_pages() / 20);
1921 recommended_min <<= (PAGE_SHIFT-10);
1922
1923 if (recommended_min > min_free_kbytes) {
1924 if (user_min_free_kbytes >= 0)
1925 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1926 min_free_kbytes, recommended_min);
1927
1928 min_free_kbytes = recommended_min;
1929 }
1930 setup_per_zone_wmarks();
1931}
1932
1933int start_stop_khugepaged(void)
1934{
1935 static struct task_struct *khugepaged_thread __read_mostly;
1936 static DEFINE_MUTEX(khugepaged_mutex);
1937 int err = 0;
1938
1939 mutex_lock(&khugepaged_mutex);
1940 if (khugepaged_enabled()) {
1941 if (!khugepaged_thread)
1942 khugepaged_thread = kthread_run(khugepaged, NULL,
1943 "khugepaged");
1944 if (IS_ERR(khugepaged_thread)) {
1945 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1946 err = PTR_ERR(khugepaged_thread);
1947 khugepaged_thread = NULL;
1948 goto fail;
1949 }
1950
1951 if (!list_empty(&khugepaged_scan.mm_head))
1952 wake_up_interruptible(&khugepaged_wait);
1953
1954 set_recommended_min_free_kbytes();
1955 } else if (khugepaged_thread) {
1956 kthread_stop(khugepaged_thread);
1957 khugepaged_thread = NULL;
1958 }
1959fail:
1960 mutex_unlock(&khugepaged_mutex);
1961 return err;
1962}
1963
generated by cgit at 2024-05-18 16:50:02 (GMT)