1 // SPDX-License-Identifier: GPL-2.0
2 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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>
23 #include <asm/pgalloc.h>
33 SCAN_LACK_REFERENCED_PAGE,
47 SCAN_ALLOC_HUGE_PAGE_FAIL,
48 SCAN_CGROUP_CHARGE_FAIL,
51 SCAN_PAGE_HAS_PRIVATE,
54 #define CREATE_TRACE_POINTS
55 #include <trace/events/huge_memory.h>
57 static struct task_struct *khugepaged_thread __read_mostly;
58 static DEFINE_MUTEX(khugepaged_mutex);
60 /* default scan 8*512 pte (or vmas) every 30 second */
61 static unsigned int khugepaged_pages_to_scan __read_mostly;
62 static unsigned int khugepaged_pages_collapsed;
63 static unsigned int khugepaged_full_scans;
64 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
65 /* during fragmentation poll the hugepage allocator once every minute */
66 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
67 static unsigned long khugepaged_sleep_expire;
68 static DEFINE_SPINLOCK(khugepaged_mm_lock);
69 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
71 * default collapse hugepages if there is at least one pte mapped like
72 * it would have happened if the vma was large enough during page
75 static unsigned int khugepaged_max_ptes_none __read_mostly;
76 static unsigned int khugepaged_max_ptes_swap __read_mostly;
78 #define MM_SLOTS_HASH_BITS 10
79 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
81 static struct kmem_cache *mm_slot_cache __read_mostly;
83 #define MAX_PTE_MAPPED_THP 8
86 * struct mm_slot - hash lookup from mm to mm_slot
87 * @hash: hash collision list
88 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
89 * @mm: the mm that this information is valid for
92 struct hlist_node hash;
93 struct list_head mm_node;
96 /* pte-mapped THP in this mm */
97 int nr_pte_mapped_thp;
98 unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
102 * struct khugepaged_scan - cursor for scanning
103 * @mm_head: the head of the mm list to scan
104 * @mm_slot: the current mm_slot we are scanning
105 * @address: the next address inside that to be scanned
107 * There is only the one khugepaged_scan instance of this cursor structure.
109 struct khugepaged_scan {
110 struct list_head mm_head;
111 struct mm_slot *mm_slot;
112 unsigned long address;
115 static struct khugepaged_scan khugepaged_scan = {
116 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
120 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
121 struct kobj_attribute *attr,
124 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
127 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
128 struct kobj_attribute *attr,
129 const char *buf, size_t count)
134 err = kstrtoul(buf, 10, &msecs);
135 if (err || msecs > UINT_MAX)
138 khugepaged_scan_sleep_millisecs = msecs;
139 khugepaged_sleep_expire = 0;
140 wake_up_interruptible(&khugepaged_wait);
144 static struct kobj_attribute scan_sleep_millisecs_attr =
145 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
146 scan_sleep_millisecs_store);
148 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
149 struct kobj_attribute *attr,
152 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
155 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
156 struct kobj_attribute *attr,
157 const char *buf, size_t count)
162 err = kstrtoul(buf, 10, &msecs);
163 if (err || msecs > UINT_MAX)
166 khugepaged_alloc_sleep_millisecs = msecs;
167 khugepaged_sleep_expire = 0;
168 wake_up_interruptible(&khugepaged_wait);
172 static struct kobj_attribute alloc_sleep_millisecs_attr =
173 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
174 alloc_sleep_millisecs_store);
176 static ssize_t pages_to_scan_show(struct kobject *kobj,
177 struct kobj_attribute *attr,
180 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
182 static ssize_t pages_to_scan_store(struct kobject *kobj,
183 struct kobj_attribute *attr,
184 const char *buf, size_t count)
189 err = kstrtoul(buf, 10, &pages);
190 if (err || !pages || pages > UINT_MAX)
193 khugepaged_pages_to_scan = pages;
197 static struct kobj_attribute pages_to_scan_attr =
198 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
199 pages_to_scan_store);
201 static ssize_t pages_collapsed_show(struct kobject *kobj,
202 struct kobj_attribute *attr,
205 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
207 static struct kobj_attribute pages_collapsed_attr =
208 __ATTR_RO(pages_collapsed);
210 static ssize_t full_scans_show(struct kobject *kobj,
211 struct kobj_attribute *attr,
214 return sprintf(buf, "%u\n", khugepaged_full_scans);
216 static struct kobj_attribute full_scans_attr =
217 __ATTR_RO(full_scans);
219 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
220 struct kobj_attribute *attr, char *buf)
222 return single_hugepage_flag_show(kobj, attr, buf,
223 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
225 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
226 struct kobj_attribute *attr,
227 const char *buf, size_t count)
229 return single_hugepage_flag_store(kobj, attr, buf, count,
230 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
232 static struct kobj_attribute khugepaged_defrag_attr =
233 __ATTR(defrag, 0644, khugepaged_defrag_show,
234 khugepaged_defrag_store);
237 * max_ptes_none controls if khugepaged should collapse hugepages over
238 * any unmapped ptes in turn potentially increasing the memory
239 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
240 * reduce the available free memory in the system as it
241 * runs. Increasing max_ptes_none will instead potentially reduce the
242 * free memory in the system during the khugepaged scan.
244 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
245 struct kobj_attribute *attr,
248 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
250 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
251 struct kobj_attribute *attr,
252 const char *buf, size_t count)
255 unsigned long max_ptes_none;
257 err = kstrtoul(buf, 10, &max_ptes_none);
258 if (err || max_ptes_none > HPAGE_PMD_NR-1)
261 khugepaged_max_ptes_none = max_ptes_none;
265 static struct kobj_attribute khugepaged_max_ptes_none_attr =
266 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
267 khugepaged_max_ptes_none_store);
269 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
270 struct kobj_attribute *attr,
273 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
276 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
277 struct kobj_attribute *attr,
278 const char *buf, size_t count)
281 unsigned long max_ptes_swap;
283 err = kstrtoul(buf, 10, &max_ptes_swap);
284 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
287 khugepaged_max_ptes_swap = max_ptes_swap;
292 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
293 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
294 khugepaged_max_ptes_swap_store);
296 static struct attribute *khugepaged_attr[] = {
297 &khugepaged_defrag_attr.attr,
298 &khugepaged_max_ptes_none_attr.attr,
299 &pages_to_scan_attr.attr,
300 &pages_collapsed_attr.attr,
301 &full_scans_attr.attr,
302 &scan_sleep_millisecs_attr.attr,
303 &alloc_sleep_millisecs_attr.attr,
304 &khugepaged_max_ptes_swap_attr.attr,
308 struct attribute_group khugepaged_attr_group = {
309 .attrs = khugepaged_attr,
310 .name = "khugepaged",
312 #endif /* CONFIG_SYSFS */
314 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
316 int hugepage_madvise(struct vm_area_struct *vma,
317 unsigned long *vm_flags, int advice)
323 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
324 * can't handle this properly after s390_enable_sie, so we simply
325 * ignore the madvise to prevent qemu from causing a SIGSEGV.
327 if (mm_has_pgste(vma->vm_mm))
330 *vm_flags &= ~VM_NOHUGEPAGE;
331 *vm_flags |= VM_HUGEPAGE;
333 * If the vma become good for khugepaged to scan,
334 * register it here without waiting a page fault that
335 * may not happen any time soon.
337 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
338 khugepaged_enter_vma_merge(vma, *vm_flags))
341 case MADV_NOHUGEPAGE:
342 *vm_flags &= ~VM_HUGEPAGE;
343 *vm_flags |= VM_NOHUGEPAGE;
345 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
346 * this vma even if we leave the mm registered in khugepaged if
347 * it got registered before VM_NOHUGEPAGE was set.
355 int __init khugepaged_init(void)
357 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
358 sizeof(struct mm_slot),
359 __alignof__(struct mm_slot), 0, NULL);
363 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
364 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
365 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
370 void __init khugepaged_destroy(void)
372 kmem_cache_destroy(mm_slot_cache);
375 static inline struct mm_slot *alloc_mm_slot(void)
377 if (!mm_slot_cache) /* initialization failed */
379 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
382 static inline void free_mm_slot(struct mm_slot *mm_slot)
384 kmem_cache_free(mm_slot_cache, mm_slot);
387 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
389 struct mm_slot *mm_slot;
391 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
392 if (mm == mm_slot->mm)
398 static void insert_to_mm_slots_hash(struct mm_struct *mm,
399 struct mm_slot *mm_slot)
402 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
405 static inline int khugepaged_test_exit(struct mm_struct *mm)
407 return atomic_read(&mm->mm_users) == 0 || !mmget_still_valid(mm);
410 static bool hugepage_vma_check(struct vm_area_struct *vma,
411 unsigned long vm_flags)
413 if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
414 (vm_flags & VM_NOHUGEPAGE) ||
415 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
418 if (shmem_file(vma->vm_file) ||
419 (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
421 (vm_flags & VM_DENYWRITE))) {
422 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
424 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
427 if (!vma->anon_vma || vma->vm_ops)
429 if (is_vma_temporary_stack(vma))
431 return !(vm_flags & VM_NO_KHUGEPAGED);
434 int __khugepaged_enter(struct mm_struct *mm)
436 struct mm_slot *mm_slot;
439 mm_slot = alloc_mm_slot();
443 /* __khugepaged_exit() must not run from under us */
444 VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
445 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
446 free_mm_slot(mm_slot);
450 spin_lock(&khugepaged_mm_lock);
451 insert_to_mm_slots_hash(mm, mm_slot);
453 * Insert just behind the scanning cursor, to let the area settle
456 wakeup = list_empty(&khugepaged_scan.mm_head);
457 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
458 spin_unlock(&khugepaged_mm_lock);
462 wake_up_interruptible(&khugepaged_wait);
467 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
468 unsigned long vm_flags)
470 unsigned long hstart, hend;
473 * khugepaged only supports read-only files for non-shmem files.
474 * khugepaged does not yet work on special mappings. And
475 * file-private shmem THP is not supported.
477 if (!hugepage_vma_check(vma, vm_flags))
480 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
481 hend = vma->vm_end & HPAGE_PMD_MASK;
483 return khugepaged_enter(vma, vm_flags);
487 void __khugepaged_exit(struct mm_struct *mm)
489 struct mm_slot *mm_slot;
492 spin_lock(&khugepaged_mm_lock);
493 mm_slot = get_mm_slot(mm);
494 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
495 hash_del(&mm_slot->hash);
496 list_del(&mm_slot->mm_node);
499 spin_unlock(&khugepaged_mm_lock);
502 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
503 free_mm_slot(mm_slot);
505 } else if (mm_slot) {
507 * This is required to serialize against
508 * khugepaged_test_exit() (which is guaranteed to run
509 * under mmap sem read mode). Stop here (after we
510 * return all pagetables will be destroyed) until
511 * khugepaged has finished working on the pagetables
512 * under the mmap_sem.
514 down_write(&mm->mmap_sem);
515 up_write(&mm->mmap_sem);
519 static void release_pte_page(struct page *page)
521 dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
523 putback_lru_page(page);
526 static void release_pte_pages(pte_t *pte, pte_t *_pte)
528 while (--_pte >= pte) {
529 pte_t pteval = *_pte;
530 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
531 release_pte_page(pte_page(pteval));
535 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
536 unsigned long address,
539 struct page *page = NULL;
541 int none_or_zero = 0, result = 0, referenced = 0;
542 bool writable = false;
544 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
545 _pte++, address += PAGE_SIZE) {
546 pte_t pteval = *_pte;
547 if (pte_none(pteval) || (pte_present(pteval) &&
548 is_zero_pfn(pte_pfn(pteval)))) {
549 if (!userfaultfd_armed(vma) &&
550 ++none_or_zero <= khugepaged_max_ptes_none) {
553 result = SCAN_EXCEED_NONE_PTE;
557 if (!pte_present(pteval)) {
558 result = SCAN_PTE_NON_PRESENT;
561 page = vm_normal_page(vma, address, pteval);
562 if (unlikely(!page)) {
563 result = SCAN_PAGE_NULL;
567 /* TODO: teach khugepaged to collapse THP mapped with pte */
568 if (PageCompound(page)) {
569 result = SCAN_PAGE_COMPOUND;
573 VM_BUG_ON_PAGE(!PageAnon(page), page);
576 * We can do it before isolate_lru_page because the
577 * page can't be freed from under us. NOTE: PG_lock
578 * is needed to serialize against split_huge_page
579 * when invoked from the VM.
581 if (!trylock_page(page)) {
582 result = SCAN_PAGE_LOCK;
587 * cannot use mapcount: can't collapse if there's a gup pin.
588 * The page must only be referenced by the scanned process
589 * and page swap cache.
591 if (page_count(page) != 1 + PageSwapCache(page)) {
593 result = SCAN_PAGE_COUNT;
596 if (pte_write(pteval)) {
599 if (PageSwapCache(page) &&
600 !reuse_swap_page(page, NULL)) {
602 result = SCAN_SWAP_CACHE_PAGE;
606 * Page is not in the swap cache. It can be collapsed
612 * Isolate the page to avoid collapsing an hugepage
613 * currently in use by the VM.
615 if (isolate_lru_page(page)) {
617 result = SCAN_DEL_PAGE_LRU;
620 inc_node_page_state(page,
621 NR_ISOLATED_ANON + page_is_file_cache(page));
622 VM_BUG_ON_PAGE(!PageLocked(page), page);
623 VM_BUG_ON_PAGE(PageLRU(page), page);
625 /* There should be enough young pte to collapse the page */
626 if (pte_young(pteval) ||
627 page_is_young(page) || PageReferenced(page) ||
628 mmu_notifier_test_young(vma->vm_mm, address))
632 if (unlikely(!writable)) {
633 result = SCAN_PAGE_RO;
634 } else if (unlikely(!referenced)) {
635 result = SCAN_LACK_REFERENCED_PAGE;
637 result = SCAN_SUCCEED;
638 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
639 referenced, writable, result);
643 release_pte_pages(pte, _pte);
644 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
645 referenced, writable, result);
649 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
650 struct vm_area_struct *vma,
651 unsigned long address,
655 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
656 _pte++, page++, address += PAGE_SIZE) {
657 pte_t pteval = *_pte;
658 struct page *src_page;
660 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
661 clear_user_highpage(page, address);
662 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
663 if (is_zero_pfn(pte_pfn(pteval))) {
665 * ptl mostly unnecessary.
669 * paravirt calls inside pte_clear here are
672 pte_clear(vma->vm_mm, address, _pte);
676 src_page = pte_page(pteval);
677 copy_user_highpage(page, src_page, address, vma);
678 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
679 release_pte_page(src_page);
681 * ptl mostly unnecessary, but preempt has to
682 * be disabled to update the per-cpu stats
683 * inside page_remove_rmap().
687 * paravirt calls inside pte_clear here are
690 pte_clear(vma->vm_mm, address, _pte);
691 page_remove_rmap(src_page, false);
693 free_page_and_swap_cache(src_page);
698 static void khugepaged_alloc_sleep(void)
702 add_wait_queue(&khugepaged_wait, &wait);
703 freezable_schedule_timeout_interruptible(
704 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
705 remove_wait_queue(&khugepaged_wait, &wait);
708 static int khugepaged_node_load[MAX_NUMNODES];
710 static bool khugepaged_scan_abort(int nid)
715 * If node_reclaim_mode is disabled, then no extra effort is made to
716 * allocate memory locally.
718 if (!node_reclaim_mode)
721 /* If there is a count for this node already, it must be acceptable */
722 if (khugepaged_node_load[nid])
725 for (i = 0; i < MAX_NUMNODES; i++) {
726 if (!khugepaged_node_load[i])
728 if (node_distance(nid, i) > node_reclaim_distance)
734 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
735 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
737 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
741 static int khugepaged_find_target_node(void)
743 static int last_khugepaged_target_node = NUMA_NO_NODE;
744 int nid, target_node = 0, max_value = 0;
746 /* find first node with max normal pages hit */
747 for (nid = 0; nid < MAX_NUMNODES; nid++)
748 if (khugepaged_node_load[nid] > max_value) {
749 max_value = khugepaged_node_load[nid];
753 /* do some balance if several nodes have the same hit record */
754 if (target_node <= last_khugepaged_target_node)
755 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
757 if (max_value == khugepaged_node_load[nid]) {
762 last_khugepaged_target_node = target_node;
766 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
768 if (IS_ERR(*hpage)) {
774 khugepaged_alloc_sleep();
784 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
786 VM_BUG_ON_PAGE(*hpage, *hpage);
788 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
789 if (unlikely(!*hpage)) {
790 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
791 *hpage = ERR_PTR(-ENOMEM);
795 prep_transhuge_page(*hpage);
796 count_vm_event(THP_COLLAPSE_ALLOC);
800 static int khugepaged_find_target_node(void)
805 static inline struct page *alloc_khugepaged_hugepage(void)
809 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
812 prep_transhuge_page(page);
816 static struct page *khugepaged_alloc_hugepage(bool *wait)
821 hpage = alloc_khugepaged_hugepage();
823 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
828 khugepaged_alloc_sleep();
830 count_vm_event(THP_COLLAPSE_ALLOC);
831 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
836 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
839 * If the hpage allocated earlier was briefly exposed in page cache
840 * before collapse_file() failed, it is possible that racing lookups
841 * have not yet completed, and would then be unpleasantly surprised by
842 * finding the hpage reused for the same mapping at a different offset.
843 * Just release the previous allocation if there is any danger of that.
845 if (*hpage && page_count(*hpage) > 1) {
851 *hpage = khugepaged_alloc_hugepage(wait);
853 if (unlikely(!*hpage))
860 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
869 * If mmap_sem temporarily dropped, revalidate vma
870 * before taking mmap_sem.
871 * Return 0 if succeeds, otherwise return none-zero
875 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
876 struct vm_area_struct **vmap)
878 struct vm_area_struct *vma;
879 unsigned long hstart, hend;
881 if (unlikely(khugepaged_test_exit(mm)))
882 return SCAN_ANY_PROCESS;
884 *vmap = vma = find_vma(mm, address);
886 return SCAN_VMA_NULL;
888 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
889 hend = vma->vm_end & HPAGE_PMD_MASK;
890 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
891 return SCAN_ADDRESS_RANGE;
892 if (!hugepage_vma_check(vma, vma->vm_flags))
893 return SCAN_VMA_CHECK;
894 /* Anon VMA expected */
895 if (!vma->anon_vma || vma->vm_ops)
896 return SCAN_VMA_CHECK;
901 * Bring missing pages in from swap, to complete THP collapse.
902 * Only done if khugepaged_scan_pmd believes it is worthwhile.
904 * Called and returns without pte mapped or spinlocks held,
905 * but with mmap_sem held to protect against vma changes.
908 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
909 struct vm_area_struct *vma,
910 unsigned long address, pmd_t *pmd,
915 struct vm_fault vmf = {
918 .flags = FAULT_FLAG_ALLOW_RETRY,
920 .pgoff = linear_page_index(vma, address),
923 /* we only decide to swapin, if there is enough young ptes */
924 if (referenced < HPAGE_PMD_NR/2) {
925 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
928 vmf.pte = pte_offset_map(pmd, address);
929 for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
930 vmf.pte++, vmf.address += PAGE_SIZE) {
931 vmf.orig_pte = *vmf.pte;
932 if (!is_swap_pte(vmf.orig_pte))
935 ret = do_swap_page(&vmf);
937 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
938 if (ret & VM_FAULT_RETRY) {
939 down_read(&mm->mmap_sem);
940 if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
941 /* vma is no longer available, don't continue to swapin */
942 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
945 /* check if the pmd is still valid */
946 if (mm_find_pmd(mm, address) != pmd) {
947 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
951 if (ret & VM_FAULT_ERROR) {
952 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
955 /* pte is unmapped now, we need to map it */
956 vmf.pte = pte_offset_map(pmd, vmf.address);
960 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
964 static void collapse_huge_page(struct mm_struct *mm,
965 unsigned long address,
967 int node, int referenced)
972 struct page *new_page;
973 spinlock_t *pmd_ptl, *pte_ptl;
974 int isolated = 0, result = 0;
975 struct mem_cgroup *memcg;
976 struct vm_area_struct *vma;
977 struct mmu_notifier_range range;
980 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
982 /* Only allocate from the target node */
983 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
986 * Before allocating the hugepage, release the mmap_sem read lock.
987 * The allocation can take potentially a long time if it involves
988 * sync compaction, and we do not need to hold the mmap_sem during
989 * that. We will recheck the vma after taking it again in write mode.
991 up_read(&mm->mmap_sem);
992 new_page = khugepaged_alloc_page(hpage, gfp, node);
994 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
998 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
999 result = SCAN_CGROUP_CHARGE_FAIL;
1003 down_read(&mm->mmap_sem);
1004 result = hugepage_vma_revalidate(mm, address, &vma);
1006 mem_cgroup_cancel_charge(new_page, memcg, true);
1007 up_read(&mm->mmap_sem);
1011 pmd = mm_find_pmd(mm, address);
1013 result = SCAN_PMD_NULL;
1014 mem_cgroup_cancel_charge(new_page, memcg, true);
1015 up_read(&mm->mmap_sem);
1020 * __collapse_huge_page_swapin always returns with mmap_sem locked.
1021 * If it fails, we release mmap_sem and jump out_nolock.
1022 * Continuing to collapse causes inconsistency.
1024 if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1025 mem_cgroup_cancel_charge(new_page, memcg, true);
1026 up_read(&mm->mmap_sem);
1030 up_read(&mm->mmap_sem);
1032 * Prevent all access to pagetables with the exception of
1033 * gup_fast later handled by the ptep_clear_flush and the VM
1034 * handled by the anon_vma lock + PG_lock.
1036 down_write(&mm->mmap_sem);
1037 result = hugepage_vma_revalidate(mm, address, &vma);
1040 /* check if the pmd is still valid */
1041 if (mm_find_pmd(mm, address) != pmd)
1044 anon_vma_lock_write(vma->anon_vma);
1046 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1047 address, address + HPAGE_PMD_SIZE);
1048 mmu_notifier_invalidate_range_start(&range);
1050 pte = pte_offset_map(pmd, address);
1051 pte_ptl = pte_lockptr(mm, pmd);
1053 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1055 * After this gup_fast can't run anymore. This also removes
1056 * any huge TLB entry from the CPU so we won't allow
1057 * huge and small TLB entries for the same virtual address
1058 * to avoid the risk of CPU bugs in that area.
1060 _pmd = pmdp_collapse_flush(vma, address, pmd);
1061 spin_unlock(pmd_ptl);
1062 mmu_notifier_invalidate_range_end(&range);
1063 tlb_remove_table_sync_one();
1066 isolated = __collapse_huge_page_isolate(vma, address, pte);
1067 spin_unlock(pte_ptl);
1069 if (unlikely(!isolated)) {
1072 BUG_ON(!pmd_none(*pmd));
1074 * We can only use set_pmd_at when establishing
1075 * hugepmds and never for establishing regular pmds that
1076 * points to regular pagetables. Use pmd_populate for that
1078 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1079 spin_unlock(pmd_ptl);
1080 anon_vma_unlock_write(vma->anon_vma);
1086 * All pages are isolated and locked so anon_vma rmap
1087 * can't run anymore.
1089 anon_vma_unlock_write(vma->anon_vma);
1091 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1093 __SetPageUptodate(new_page);
1094 pgtable = pmd_pgtable(_pmd);
1096 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1097 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1100 * spin_lock() below is not the equivalent of smp_wmb(), so
1101 * this is needed to avoid the copy_huge_page writes to become
1102 * visible after the set_pmd_at() write.
1107 BUG_ON(!pmd_none(*pmd));
1108 page_add_new_anon_rmap(new_page, vma, address, true);
1109 mem_cgroup_commit_charge(new_page, memcg, false, true);
1110 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1111 lru_cache_add_active_or_unevictable(new_page, vma);
1112 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1113 set_pmd_at(mm, address, pmd, _pmd);
1114 update_mmu_cache_pmd(vma, address, pmd);
1115 spin_unlock(pmd_ptl);
1119 khugepaged_pages_collapsed++;
1120 result = SCAN_SUCCEED;
1122 up_write(&mm->mmap_sem);
1124 trace_mm_collapse_huge_page(mm, isolated, result);
1127 mem_cgroup_cancel_charge(new_page, memcg, true);
1131 static int khugepaged_scan_pmd(struct mm_struct *mm,
1132 struct vm_area_struct *vma,
1133 unsigned long address,
1134 struct page **hpage)
1138 int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1139 struct page *page = NULL;
1140 unsigned long _address;
1142 int node = NUMA_NO_NODE, unmapped = 0;
1143 bool writable = false;
1145 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1147 pmd = mm_find_pmd(mm, address);
1149 result = SCAN_PMD_NULL;
1153 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1154 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1155 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1156 _pte++, _address += PAGE_SIZE) {
1157 pte_t pteval = *_pte;
1158 if (is_swap_pte(pteval)) {
1159 if (++unmapped <= khugepaged_max_ptes_swap) {
1162 result = SCAN_EXCEED_SWAP_PTE;
1166 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1167 if (!userfaultfd_armed(vma) &&
1168 ++none_or_zero <= khugepaged_max_ptes_none) {
1171 result = SCAN_EXCEED_NONE_PTE;
1175 if (!pte_present(pteval)) {
1176 result = SCAN_PTE_NON_PRESENT;
1179 if (pte_write(pteval))
1182 page = vm_normal_page(vma, _address, pteval);
1183 if (unlikely(!page)) {
1184 result = SCAN_PAGE_NULL;
1188 /* TODO: teach khugepaged to collapse THP mapped with pte */
1189 if (PageCompound(page)) {
1190 result = SCAN_PAGE_COMPOUND;
1195 * Record which node the original page is from and save this
1196 * information to khugepaged_node_load[].
1197 * Khupaged will allocate hugepage from the node has the max
1200 node = page_to_nid(page);
1201 if (khugepaged_scan_abort(node)) {
1202 result = SCAN_SCAN_ABORT;
1205 khugepaged_node_load[node]++;
1206 if (!PageLRU(page)) {
1207 result = SCAN_PAGE_LRU;
1210 if (PageLocked(page)) {
1211 result = SCAN_PAGE_LOCK;
1214 if (!PageAnon(page)) {
1215 result = SCAN_PAGE_ANON;
1220 * cannot use mapcount: can't collapse if there's a gup pin.
1221 * The page must only be referenced by the scanned process
1222 * and page swap cache.
1224 if (page_count(page) != 1 + PageSwapCache(page)) {
1225 result = SCAN_PAGE_COUNT;
1228 if (pte_young(pteval) ||
1229 page_is_young(page) || PageReferenced(page) ||
1230 mmu_notifier_test_young(vma->vm_mm, address))
1235 result = SCAN_SUCCEED;
1238 result = SCAN_LACK_REFERENCED_PAGE;
1241 result = SCAN_PAGE_RO;
1244 pte_unmap_unlock(pte, ptl);
1246 node = khugepaged_find_target_node();
1247 /* collapse_huge_page will return with the mmap_sem released */
1248 collapse_huge_page(mm, address, hpage, node, referenced);
1251 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1252 none_or_zero, result, unmapped);
1256 static void collect_mm_slot(struct mm_slot *mm_slot)
1258 struct mm_struct *mm = mm_slot->mm;
1260 lockdep_assert_held(&khugepaged_mm_lock);
1262 if (khugepaged_test_exit(mm)) {
1264 hash_del(&mm_slot->hash);
1265 list_del(&mm_slot->mm_node);
1268 * Not strictly needed because the mm exited already.
1270 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1273 /* khugepaged_mm_lock actually not necessary for the below */
1274 free_mm_slot(mm_slot);
1279 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1281 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1282 * khugepaged should try to collapse the page table.
1284 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1287 struct mm_slot *mm_slot;
1289 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1291 spin_lock(&khugepaged_mm_lock);
1292 mm_slot = get_mm_slot(mm);
1293 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1294 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1295 spin_unlock(&khugepaged_mm_lock);
1300 * Try to collapse a pte-mapped THP for mm at address haddr.
1302 * This function checks whether all the PTEs in the PMD are pointing to the
1303 * right THP. If so, retract the page table so the THP can refault in with
1306 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1308 unsigned long haddr = addr & HPAGE_PMD_MASK;
1309 struct vm_area_struct *vma = find_vma(mm, haddr);
1311 pte_t *start_pte, *pte;
1316 struct mmu_notifier_range range;
1318 if (!vma || !vma->vm_file ||
1319 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1323 * This vm_flags may not have VM_HUGEPAGE if the page was not
1324 * collapsed by this mm. But we can still collapse if the page is
1325 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1326 * will not fail the vma for missing VM_HUGEPAGE
1328 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1331 hpage = find_lock_page(vma->vm_file->f_mapping,
1332 linear_page_index(vma, haddr));
1336 if (!PageHead(hpage))
1339 pmd = mm_find_pmd(mm, haddr);
1344 * We need to lock the mapping so that from here on, only GUP-fast and
1345 * hardware page walks can access the parts of the page tables that
1346 * we're operating on.
1348 i_mmap_lock_write(vma->vm_file->f_mapping);
1351 * This spinlock should be unnecessary: Nobody else should be accessing
1352 * the page tables under spinlock protection here, only
1353 * lockless_pages_from_mm() and the hardware page walker can access page
1354 * tables while all the high-level locks are held in write mode.
1356 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1358 /* step 1: check all mapped PTEs are to the right huge page */
1359 for (i = 0, addr = haddr, pte = start_pte;
1360 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1363 /* empty pte, skip */
1367 /* page swapped out, abort */
1368 if (!pte_present(*pte))
1371 page = vm_normal_page(vma, addr, *pte);
1374 * Note that uprobe, debugger, or MAP_PRIVATE may change the
1375 * page table, but the new page will not be a subpage of hpage.
1377 if (hpage + i != page)
1382 /* step 2: adjust rmap */
1383 for (i = 0, addr = haddr, pte = start_pte;
1384 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1389 page = vm_normal_page(vma, addr, *pte);
1390 page_remove_rmap(page, false);
1393 pte_unmap_unlock(start_pte, ptl);
1395 /* step 3: set proper refcount and mm_counters. */
1397 page_ref_sub(hpage, count);
1398 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1401 /* step 4: collapse pmd */
1402 /* we make no change to anon, but protect concurrent anon page lookup */
1404 anon_vma_lock_write(vma->anon_vma);
1406 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, haddr,
1407 haddr + HPAGE_PMD_SIZE);
1408 mmu_notifier_invalidate_range_start(&range);
1409 _pmd = pmdp_collapse_flush(vma, haddr, pmd);
1411 tlb_remove_table_sync_one();
1412 mmu_notifier_invalidate_range_end(&range);
1413 pte_free(mm, pmd_pgtable(_pmd));
1416 anon_vma_unlock_write(vma->anon_vma);
1417 i_mmap_unlock_write(vma->vm_file->f_mapping);
1425 pte_unmap_unlock(start_pte, ptl);
1426 i_mmap_unlock_write(vma->vm_file->f_mapping);
1430 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1432 struct mm_struct *mm = mm_slot->mm;
1435 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1438 if (!down_write_trylock(&mm->mmap_sem))
1441 if (unlikely(khugepaged_test_exit(mm)))
1444 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1445 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1448 mm_slot->nr_pte_mapped_thp = 0;
1449 up_write(&mm->mmap_sem);
1453 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1455 struct vm_area_struct *vma;
1456 struct mm_struct *mm;
1460 i_mmap_lock_write(mapping);
1461 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1463 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1464 * got written to. These VMAs are likely not worth investing
1465 * down_write(mmap_sem) as PMD-mapping is likely to be split
1468 * Not that vma->anon_vma check is racy: it can be set up after
1469 * the check but before we took mmap_sem by the fault path.
1470 * But page lock would prevent establishing any new ptes of the
1471 * page, so we are safe.
1473 * An alternative would be drop the check, but check that page
1474 * table is clear before calling pmdp_collapse_flush() under
1475 * ptl. It has higher chance to recover THP for the VMA, but
1476 * has higher cost too. It would also probably require locking
1481 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1482 if (addr & ~HPAGE_PMD_MASK)
1484 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1487 pmd = mm_find_pmd(mm, addr);
1491 * We need exclusive mmap_sem to retract page table.
1493 * We use trylock due to lock inversion: we need to acquire
1494 * mmap_sem while holding page lock. Fault path does it in
1495 * reverse order. Trylock is a way to avoid deadlock.
1497 if (down_write_trylock(&mm->mmap_sem)) {
1498 if (!khugepaged_test_exit(mm)) {
1499 struct mmu_notifier_range range;
1501 mmu_notifier_range_init(&range,
1502 MMU_NOTIFY_CLEAR, 0,
1504 addr + HPAGE_PMD_SIZE);
1505 mmu_notifier_invalidate_range_start(&range);
1506 /* assume page table is clear */
1507 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1509 tlb_remove_table_sync_one();
1510 pte_free(mm, pmd_pgtable(_pmd));
1511 mmu_notifier_invalidate_range_end(&range);
1513 up_write(&mm->mmap_sem);
1515 /* Try again later */
1516 khugepaged_add_pte_mapped_thp(mm, addr);
1519 i_mmap_unlock_write(mapping);
1523 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1525 * Basic scheme is simple, details are more complex:
1526 * - allocate and lock a new huge page;
1527 * - scan page cache replacing old pages with the new one
1528 * + swap/gup in pages if necessary;
1530 * + keep old pages around in case rollback is required;
1531 * - if replacing succeeds:
1534 * + unlock huge page;
1535 * - if replacing failed;
1536 * + put all pages back and unfreeze them;
1537 * + restore gaps in the page cache;
1538 * + unlock and free huge page;
1540 static void collapse_file(struct mm_struct *mm,
1541 struct file *file, pgoff_t start,
1542 struct page **hpage, int node)
1544 struct address_space *mapping = file->f_mapping;
1546 struct page *new_page;
1547 struct mem_cgroup *memcg;
1548 pgoff_t index, end = start + HPAGE_PMD_NR;
1549 LIST_HEAD(pagelist);
1550 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1551 int nr_none = 0, result = SCAN_SUCCEED;
1552 bool is_shmem = shmem_file(file);
1554 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1555 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1557 /* Only allocate from the target node */
1558 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1560 new_page = khugepaged_alloc_page(hpage, gfp, node);
1562 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1566 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1567 result = SCAN_CGROUP_CHARGE_FAIL;
1571 /* This will be less messy when we use multi-index entries */
1574 xas_create_range(&xas);
1575 if (!xas_error(&xas))
1577 xas_unlock_irq(&xas);
1578 if (!xas_nomem(&xas, GFP_KERNEL)) {
1579 mem_cgroup_cancel_charge(new_page, memcg, true);
1585 __SetPageLocked(new_page);
1587 __SetPageSwapBacked(new_page);
1588 new_page->index = start;
1589 new_page->mapping = mapping;
1592 * At this point the new_page is locked and not up-to-date.
1593 * It's safe to insert it into the page cache, because nobody would
1594 * be able to map it or use it in another way until we unlock it.
1597 xas_set(&xas, start);
1598 for (index = start; index < end; index++) {
1599 struct page *page = xas_next(&xas);
1601 VM_BUG_ON(index != xas.xa_index);
1605 * Stop if extent has been truncated or
1606 * hole-punched, and is now completely
1609 if (index == start) {
1610 if (!xas_next_entry(&xas, end - 1)) {
1611 result = SCAN_TRUNCATED;
1614 xas_set(&xas, index);
1616 if (!shmem_charge(mapping->host, 1)) {
1620 xas_store(&xas, new_page);
1625 if (xa_is_value(page) || !PageUptodate(page)) {
1626 xas_unlock_irq(&xas);
1627 /* swap in or instantiate fallocated page */
1628 if (shmem_getpage(mapping->host, index, &page,
1633 } else if (trylock_page(page)) {
1635 xas_unlock_irq(&xas);
1637 result = SCAN_PAGE_LOCK;
1640 } else { /* !is_shmem */
1641 if (!page || xa_is_value(page)) {
1642 xas_unlock_irq(&xas);
1643 page_cache_sync_readahead(mapping, &file->f_ra,
1646 /* drain pagevecs to help isolate_lru_page() */
1648 page = find_lock_page(mapping, index);
1649 if (unlikely(page == NULL)) {
1653 } else if (trylock_page(page)) {
1655 xas_unlock_irq(&xas);
1657 result = SCAN_PAGE_LOCK;
1663 * The page must be locked, so we can drop the i_pages lock
1664 * without racing with truncate.
1666 VM_BUG_ON_PAGE(!PageLocked(page), page);
1668 /* make sure the page is up to date */
1669 if (unlikely(!PageUptodate(page))) {
1675 * If file was truncated then extended, or hole-punched, before
1676 * we locked the first page, then a THP might be there already.
1678 if (PageTransCompound(page)) {
1679 result = SCAN_PAGE_COMPOUND;
1683 if (page_mapping(page) != mapping) {
1684 result = SCAN_TRUNCATED;
1688 if (!is_shmem && PageDirty(page)) {
1690 * khugepaged only works on read-only fd, so this
1691 * page is dirty because it hasn't been flushed
1692 * since first write.
1698 if (isolate_lru_page(page)) {
1699 result = SCAN_DEL_PAGE_LRU;
1703 if (page_has_private(page) &&
1704 !try_to_release_page(page, GFP_KERNEL)) {
1705 result = SCAN_PAGE_HAS_PRIVATE;
1706 putback_lru_page(page);
1710 if (page_mapped(page))
1711 unmap_mapping_pages(mapping, index, 1, false);
1714 xas_set(&xas, index);
1716 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1717 VM_BUG_ON_PAGE(page_mapped(page), page);
1720 * The page is expected to have page_count() == 3:
1721 * - we hold a pin on it;
1722 * - one reference from page cache;
1723 * - one from isolate_lru_page;
1725 if (!page_ref_freeze(page, 3)) {
1726 result = SCAN_PAGE_COUNT;
1727 xas_unlock_irq(&xas);
1728 putback_lru_page(page);
1733 * Add the page to the list to be able to undo the collapse if
1734 * something go wrong.
1736 list_add_tail(&page->lru, &pagelist);
1738 /* Finally, replace with the new page. */
1739 xas_store(&xas, new_page);
1748 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1750 __inc_node_page_state(new_page, NR_FILE_THPS);
1751 filemap_nr_thps_inc(mapping);
1755 struct zone *zone = page_zone(new_page);
1757 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1759 __mod_node_page_state(zone->zone_pgdat,
1764 xas_unlock_irq(&xas);
1767 if (result == SCAN_SUCCEED) {
1768 struct page *page, *tmp;
1771 * Replacing old pages with new one has succeeded, now we
1772 * need to copy the content and free the old pages.
1775 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1776 while (index < page->index) {
1777 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1780 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1782 list_del(&page->lru);
1783 page->mapping = NULL;
1784 page_ref_unfreeze(page, 1);
1785 ClearPageActive(page);
1786 ClearPageUnevictable(page);
1791 while (index < end) {
1792 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1796 SetPageUptodate(new_page);
1797 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1798 mem_cgroup_commit_charge(new_page, memcg, false, true);
1801 set_page_dirty(new_page);
1802 lru_cache_add_anon(new_page);
1804 lru_cache_add_file(new_page);
1806 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1809 * Remove pte page tables, so we can re-fault the page as huge.
1811 retract_page_tables(mapping, start);
1814 khugepaged_pages_collapsed++;
1818 /* Something went wrong: roll back page cache changes */
1820 mapping->nrpages -= nr_none;
1823 shmem_uncharge(mapping->host, nr_none);
1825 xas_set(&xas, start);
1826 xas_for_each(&xas, page, end - 1) {
1827 page = list_first_entry_or_null(&pagelist,
1829 if (!page || xas.xa_index < page->index) {
1833 /* Put holes back where they were */
1834 xas_store(&xas, NULL);
1838 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1840 /* Unfreeze the page. */
1841 list_del(&page->lru);
1842 page_ref_unfreeze(page, 2);
1843 xas_store(&xas, page);
1845 xas_unlock_irq(&xas);
1847 putback_lru_page(page);
1851 xas_unlock_irq(&xas);
1853 mem_cgroup_cancel_charge(new_page, memcg, true);
1854 new_page->mapping = NULL;
1857 unlock_page(new_page);
1859 VM_BUG_ON(!list_empty(&pagelist));
1860 /* TODO: tracepoints */
1863 static void khugepaged_scan_file(struct mm_struct *mm,
1864 struct file *file, pgoff_t start, struct page **hpage)
1866 struct page *page = NULL;
1867 struct address_space *mapping = file->f_mapping;
1868 XA_STATE(xas, &mapping->i_pages, start);
1870 int node = NUMA_NO_NODE;
1871 int result = SCAN_SUCCEED;
1875 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1877 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1878 if (xas_retry(&xas, page))
1881 if (xa_is_value(page)) {
1882 if (++swap > khugepaged_max_ptes_swap) {
1883 result = SCAN_EXCEED_SWAP_PTE;
1889 if (PageTransCompound(page)) {
1890 result = SCAN_PAGE_COMPOUND;
1894 node = page_to_nid(page);
1895 if (khugepaged_scan_abort(node)) {
1896 result = SCAN_SCAN_ABORT;
1899 khugepaged_node_load[node]++;
1901 if (!PageLRU(page)) {
1902 result = SCAN_PAGE_LRU;
1906 if (page_count(page) !=
1907 1 + page_mapcount(page) + page_has_private(page)) {
1908 result = SCAN_PAGE_COUNT;
1913 * We probably should check if the page is referenced here, but
1914 * nobody would transfer pte_young() to PageReferenced() for us.
1915 * And rmap walk here is just too costly...
1920 if (need_resched()) {
1927 if (result == SCAN_SUCCEED) {
1928 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1929 result = SCAN_EXCEED_NONE_PTE;
1931 node = khugepaged_find_target_node();
1932 collapse_file(mm, file, start, hpage, node);
1936 /* TODO: tracepoints */
1939 static void khugepaged_scan_file(struct mm_struct *mm,
1940 struct file *file, pgoff_t start, struct page **hpage)
1945 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1951 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1952 struct page **hpage)
1953 __releases(&khugepaged_mm_lock)
1954 __acquires(&khugepaged_mm_lock)
1956 struct mm_slot *mm_slot;
1957 struct mm_struct *mm;
1958 struct vm_area_struct *vma;
1962 lockdep_assert_held(&khugepaged_mm_lock);
1964 if (khugepaged_scan.mm_slot)
1965 mm_slot = khugepaged_scan.mm_slot;
1967 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1968 struct mm_slot, mm_node);
1969 khugepaged_scan.address = 0;
1970 khugepaged_scan.mm_slot = mm_slot;
1972 spin_unlock(&khugepaged_mm_lock);
1973 khugepaged_collapse_pte_mapped_thps(mm_slot);
1977 * Don't wait for semaphore (to avoid long wait times). Just move to
1978 * the next mm on the list.
1981 if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1982 goto breakouterloop_mmap_sem;
1983 if (likely(!khugepaged_test_exit(mm)))
1984 vma = find_vma(mm, khugepaged_scan.address);
1987 for (; vma; vma = vma->vm_next) {
1988 unsigned long hstart, hend;
1991 if (unlikely(khugepaged_test_exit(mm))) {
1995 if (!hugepage_vma_check(vma, vma->vm_flags)) {
2000 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2001 hend = vma->vm_end & HPAGE_PMD_MASK;
2004 if (khugepaged_scan.address > hend)
2006 if (khugepaged_scan.address < hstart)
2007 khugepaged_scan.address = hstart;
2008 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2010 while (khugepaged_scan.address < hend) {
2013 if (unlikely(khugepaged_test_exit(mm)))
2014 goto breakouterloop;
2016 VM_BUG_ON(khugepaged_scan.address < hstart ||
2017 khugepaged_scan.address + HPAGE_PMD_SIZE >
2019 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
2021 pgoff_t pgoff = linear_page_index(vma,
2022 khugepaged_scan.address);
2024 if (shmem_file(vma->vm_file)
2025 && !shmem_huge_enabled(vma))
2027 file = get_file(vma->vm_file);
2028 up_read(&mm->mmap_sem);
2030 khugepaged_scan_file(mm, file, pgoff, hpage);
2033 ret = khugepaged_scan_pmd(mm, vma,
2034 khugepaged_scan.address,
2037 /* move to next address */
2038 khugepaged_scan.address += HPAGE_PMD_SIZE;
2039 progress += HPAGE_PMD_NR;
2041 /* we released mmap_sem so break loop */
2042 goto breakouterloop_mmap_sem;
2043 if (progress >= pages)
2044 goto breakouterloop;
2048 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2049 breakouterloop_mmap_sem:
2051 spin_lock(&khugepaged_mm_lock);
2052 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2054 * Release the current mm_slot if this mm is about to die, or
2055 * if we scanned all vmas of this mm.
2057 if (khugepaged_test_exit(mm) || !vma) {
2059 * Make sure that if mm_users is reaching zero while
2060 * khugepaged runs here, khugepaged_exit will find
2061 * mm_slot not pointing to the exiting mm.
2063 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2064 khugepaged_scan.mm_slot = list_entry(
2065 mm_slot->mm_node.next,
2066 struct mm_slot, mm_node);
2067 khugepaged_scan.address = 0;
2069 khugepaged_scan.mm_slot = NULL;
2070 khugepaged_full_scans++;
2073 collect_mm_slot(mm_slot);
2079 static int khugepaged_has_work(void)
2081 return !list_empty(&khugepaged_scan.mm_head) &&
2082 khugepaged_enabled();
2085 static int khugepaged_wait_event(void)
2087 return !list_empty(&khugepaged_scan.mm_head) ||
2088 kthread_should_stop();
2091 static void khugepaged_do_scan(void)
2093 struct page *hpage = NULL;
2094 unsigned int progress = 0, pass_through_head = 0;
2095 unsigned int pages = khugepaged_pages_to_scan;
2098 barrier(); /* write khugepaged_pages_to_scan to local stack */
2100 while (progress < pages) {
2101 if (!khugepaged_prealloc_page(&hpage, &wait))
2106 if (unlikely(kthread_should_stop() || try_to_freeze()))
2109 spin_lock(&khugepaged_mm_lock);
2110 if (!khugepaged_scan.mm_slot)
2111 pass_through_head++;
2112 if (khugepaged_has_work() &&
2113 pass_through_head < 2)
2114 progress += khugepaged_scan_mm_slot(pages - progress,
2118 spin_unlock(&khugepaged_mm_lock);
2121 if (!IS_ERR_OR_NULL(hpage))
2125 static bool khugepaged_should_wakeup(void)
2127 return kthread_should_stop() ||
2128 time_after_eq(jiffies, khugepaged_sleep_expire);
2131 static void khugepaged_wait_work(void)
2133 if (khugepaged_has_work()) {
2134 const unsigned long scan_sleep_jiffies =
2135 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2137 if (!scan_sleep_jiffies)
2140 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2141 wait_event_freezable_timeout(khugepaged_wait,
2142 khugepaged_should_wakeup(),
2143 scan_sleep_jiffies);
2147 if (khugepaged_enabled())
2148 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2151 static int khugepaged(void *none)
2153 struct mm_slot *mm_slot;
2156 set_user_nice(current, MAX_NICE);
2158 while (!kthread_should_stop()) {
2159 khugepaged_do_scan();
2160 khugepaged_wait_work();
2163 spin_lock(&khugepaged_mm_lock);
2164 mm_slot = khugepaged_scan.mm_slot;
2165 khugepaged_scan.mm_slot = NULL;
2167 collect_mm_slot(mm_slot);
2168 spin_unlock(&khugepaged_mm_lock);
2172 static void set_recommended_min_free_kbytes(void)
2176 unsigned long recommended_min;
2178 for_each_populated_zone(zone) {
2180 * We don't need to worry about fragmentation of
2181 * ZONE_MOVABLE since it only has movable pages.
2183 if (zone_idx(zone) > gfp_zone(GFP_USER))
2189 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2190 recommended_min = pageblock_nr_pages * nr_zones * 2;
2193 * Make sure that on average at least two pageblocks are almost free
2194 * of another type, one for a migratetype to fall back to and a
2195 * second to avoid subsequent fallbacks of other types There are 3
2196 * MIGRATE_TYPES we care about.
2198 recommended_min += pageblock_nr_pages * nr_zones *
2199 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2201 /* don't ever allow to reserve more than 5% of the lowmem */
2202 recommended_min = min(recommended_min,
2203 (unsigned long) nr_free_buffer_pages() / 20);
2204 recommended_min <<= (PAGE_SHIFT-10);
2206 if (recommended_min > min_free_kbytes) {
2207 if (user_min_free_kbytes >= 0)
2208 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2209 min_free_kbytes, recommended_min);
2211 min_free_kbytes = recommended_min;
2213 setup_per_zone_wmarks();
2216 int start_stop_khugepaged(void)
2220 mutex_lock(&khugepaged_mutex);
2221 if (khugepaged_enabled()) {
2222 if (!khugepaged_thread)
2223 khugepaged_thread = kthread_run(khugepaged, NULL,
2225 if (IS_ERR(khugepaged_thread)) {
2226 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2227 err = PTR_ERR(khugepaged_thread);
2228 khugepaged_thread = NULL;
2232 if (!list_empty(&khugepaged_scan.mm_head))
2233 wake_up_interruptible(&khugepaged_wait);
2235 set_recommended_min_free_kbytes();
2236 } else if (khugepaged_thread) {
2237 kthread_stop(khugepaged_thread);
2238 khugepaged_thread = NULL;
2241 mutex_unlock(&khugepaged_mutex);
2245 void khugepaged_min_free_kbytes_update(void)
2247 mutex_lock(&khugepaged_mutex);
2248 if (khugepaged_enabled() && khugepaged_thread)
2249 set_recommended_min_free_kbytes();
2250 mutex_unlock(&khugepaged_mutex);