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 /* default scan 8*512 pte (or vmas) every 30 second */
58 static unsigned int khugepaged_pages_to_scan __read_mostly;
59 static unsigned int khugepaged_pages_collapsed;
60 static unsigned int khugepaged_full_scans;
61 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
62 /* during fragmentation poll the hugepage allocator once every minute */
63 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
64 static unsigned long khugepaged_sleep_expire;
65 static DEFINE_SPINLOCK(khugepaged_mm_lock);
66 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
68 * default collapse hugepages if there is at least one pte mapped like
69 * it would have happened if the vma was large enough during page
72 static unsigned int khugepaged_max_ptes_none __read_mostly;
73 static unsigned int khugepaged_max_ptes_swap __read_mostly;
75 #define MM_SLOTS_HASH_BITS 10
76 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
78 static struct kmem_cache *mm_slot_cache __read_mostly;
80 #define MAX_PTE_MAPPED_THP 8
83 * struct mm_slot - hash lookup from mm to mm_slot
84 * @hash: hash collision list
85 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
86 * @mm: the mm that this information is valid for
89 struct hlist_node hash;
90 struct list_head mm_node;
93 /* pte-mapped THP in this mm */
94 int nr_pte_mapped_thp;
95 unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP];
99 * struct khugepaged_scan - cursor for scanning
100 * @mm_head: the head of the mm list to scan
101 * @mm_slot: the current mm_slot we are scanning
102 * @address: the next address inside that to be scanned
104 * There is only the one khugepaged_scan instance of this cursor structure.
106 struct khugepaged_scan {
107 struct list_head mm_head;
108 struct mm_slot *mm_slot;
109 unsigned long address;
112 static struct khugepaged_scan khugepaged_scan = {
113 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
117 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
118 struct kobj_attribute *attr,
121 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
124 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
125 struct kobj_attribute *attr,
126 const char *buf, size_t count)
131 err = kstrtoul(buf, 10, &msecs);
132 if (err || msecs > UINT_MAX)
135 khugepaged_scan_sleep_millisecs = msecs;
136 khugepaged_sleep_expire = 0;
137 wake_up_interruptible(&khugepaged_wait);
141 static struct kobj_attribute scan_sleep_millisecs_attr =
142 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
143 scan_sleep_millisecs_store);
145 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
146 struct kobj_attribute *attr,
149 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
152 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
153 struct kobj_attribute *attr,
154 const char *buf, size_t count)
159 err = kstrtoul(buf, 10, &msecs);
160 if (err || msecs > UINT_MAX)
163 khugepaged_alloc_sleep_millisecs = msecs;
164 khugepaged_sleep_expire = 0;
165 wake_up_interruptible(&khugepaged_wait);
169 static struct kobj_attribute alloc_sleep_millisecs_attr =
170 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
171 alloc_sleep_millisecs_store);
173 static ssize_t pages_to_scan_show(struct kobject *kobj,
174 struct kobj_attribute *attr,
177 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
179 static ssize_t pages_to_scan_store(struct kobject *kobj,
180 struct kobj_attribute *attr,
181 const char *buf, size_t count)
186 err = kstrtoul(buf, 10, &pages);
187 if (err || !pages || pages > UINT_MAX)
190 khugepaged_pages_to_scan = pages;
194 static struct kobj_attribute pages_to_scan_attr =
195 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
196 pages_to_scan_store);
198 static ssize_t pages_collapsed_show(struct kobject *kobj,
199 struct kobj_attribute *attr,
202 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
204 static struct kobj_attribute pages_collapsed_attr =
205 __ATTR_RO(pages_collapsed);
207 static ssize_t full_scans_show(struct kobject *kobj,
208 struct kobj_attribute *attr,
211 return sprintf(buf, "%u\n", khugepaged_full_scans);
213 static struct kobj_attribute full_scans_attr =
214 __ATTR_RO(full_scans);
216 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
217 struct kobj_attribute *attr, char *buf)
219 return single_hugepage_flag_show(kobj, attr, buf,
220 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
222 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
223 struct kobj_attribute *attr,
224 const char *buf, size_t count)
226 return single_hugepage_flag_store(kobj, attr, buf, count,
227 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
229 static struct kobj_attribute khugepaged_defrag_attr =
230 __ATTR(defrag, 0644, khugepaged_defrag_show,
231 khugepaged_defrag_store);
234 * max_ptes_none controls if khugepaged should collapse hugepages over
235 * any unmapped ptes in turn potentially increasing the memory
236 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
237 * reduce the available free memory in the system as it
238 * runs. Increasing max_ptes_none will instead potentially reduce the
239 * free memory in the system during the khugepaged scan.
241 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
242 struct kobj_attribute *attr,
245 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
247 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
248 struct kobj_attribute *attr,
249 const char *buf, size_t count)
252 unsigned long max_ptes_none;
254 err = kstrtoul(buf, 10, &max_ptes_none);
255 if (err || max_ptes_none > HPAGE_PMD_NR-1)
258 khugepaged_max_ptes_none = max_ptes_none;
262 static struct kobj_attribute khugepaged_max_ptes_none_attr =
263 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
264 khugepaged_max_ptes_none_store);
266 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
267 struct kobj_attribute *attr,
270 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
273 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
274 struct kobj_attribute *attr,
275 const char *buf, size_t count)
278 unsigned long max_ptes_swap;
280 err = kstrtoul(buf, 10, &max_ptes_swap);
281 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
284 khugepaged_max_ptes_swap = max_ptes_swap;
289 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
290 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
291 khugepaged_max_ptes_swap_store);
293 static struct attribute *khugepaged_attr[] = {
294 &khugepaged_defrag_attr.attr,
295 &khugepaged_max_ptes_none_attr.attr,
296 &pages_to_scan_attr.attr,
297 &pages_collapsed_attr.attr,
298 &full_scans_attr.attr,
299 &scan_sleep_millisecs_attr.attr,
300 &alloc_sleep_millisecs_attr.attr,
301 &khugepaged_max_ptes_swap_attr.attr,
305 struct attribute_group khugepaged_attr_group = {
306 .attrs = khugepaged_attr,
307 .name = "khugepaged",
309 #endif /* CONFIG_SYSFS */
311 int hugepage_madvise(struct vm_area_struct *vma,
312 unsigned long *vm_flags, int advice)
318 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
319 * can't handle this properly after s390_enable_sie, so we simply
320 * ignore the madvise to prevent qemu from causing a SIGSEGV.
322 if (mm_has_pgste(vma->vm_mm))
325 *vm_flags &= ~VM_NOHUGEPAGE;
326 *vm_flags |= VM_HUGEPAGE;
328 * If the vma become good for khugepaged to scan,
329 * register it here without waiting a page fault that
330 * may not happen any time soon.
332 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
333 khugepaged_enter_vma_merge(vma, *vm_flags))
336 case MADV_NOHUGEPAGE:
337 *vm_flags &= ~VM_HUGEPAGE;
338 *vm_flags |= VM_NOHUGEPAGE;
340 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
341 * this vma even if we leave the mm registered in khugepaged if
342 * it got registered before VM_NOHUGEPAGE was set.
350 int __init khugepaged_init(void)
352 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
353 sizeof(struct mm_slot),
354 __alignof__(struct mm_slot), 0, NULL);
358 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
359 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
360 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
365 void __init khugepaged_destroy(void)
367 kmem_cache_destroy(mm_slot_cache);
370 static inline struct mm_slot *alloc_mm_slot(void)
372 if (!mm_slot_cache) /* initialization failed */
374 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
377 static inline void free_mm_slot(struct mm_slot *mm_slot)
379 kmem_cache_free(mm_slot_cache, mm_slot);
382 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
384 struct mm_slot *mm_slot;
386 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
387 if (mm == mm_slot->mm)
393 static void insert_to_mm_slots_hash(struct mm_struct *mm,
394 struct mm_slot *mm_slot)
397 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
400 static inline int khugepaged_test_exit(struct mm_struct *mm)
402 return atomic_read(&mm->mm_users) == 0;
405 static bool hugepage_vma_check(struct vm_area_struct *vma,
406 unsigned long vm_flags)
408 if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
409 (vm_flags & VM_NOHUGEPAGE) ||
410 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
413 if (shmem_file(vma->vm_file) ||
414 (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
416 (vm_flags & VM_DENYWRITE))) {
417 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
419 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
422 if (!vma->anon_vma || vma->vm_ops)
424 if (vma_is_temporary_stack(vma))
426 return !(vm_flags & VM_NO_KHUGEPAGED);
429 int __khugepaged_enter(struct mm_struct *mm)
431 struct mm_slot *mm_slot;
434 mm_slot = alloc_mm_slot();
438 /* __khugepaged_exit() must not run from under us */
439 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
440 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
441 free_mm_slot(mm_slot);
445 spin_lock(&khugepaged_mm_lock);
446 insert_to_mm_slots_hash(mm, mm_slot);
448 * Insert just behind the scanning cursor, to let the area settle
451 wakeup = list_empty(&khugepaged_scan.mm_head);
452 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
453 spin_unlock(&khugepaged_mm_lock);
457 wake_up_interruptible(&khugepaged_wait);
462 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
463 unsigned long vm_flags)
465 unsigned long hstart, hend;
468 * khugepaged only supports read-only files for non-shmem files.
469 * khugepaged does not yet work on special mappings. And
470 * file-private shmem THP is not supported.
472 if (!hugepage_vma_check(vma, vm_flags))
475 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
476 hend = vma->vm_end & HPAGE_PMD_MASK;
478 return khugepaged_enter(vma, vm_flags);
482 void __khugepaged_exit(struct mm_struct *mm)
484 struct mm_slot *mm_slot;
487 spin_lock(&khugepaged_mm_lock);
488 mm_slot = get_mm_slot(mm);
489 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
490 hash_del(&mm_slot->hash);
491 list_del(&mm_slot->mm_node);
494 spin_unlock(&khugepaged_mm_lock);
497 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
498 free_mm_slot(mm_slot);
500 } else if (mm_slot) {
502 * This is required to serialize against
503 * khugepaged_test_exit() (which is guaranteed to run
504 * under mmap sem read mode). Stop here (after we
505 * return all pagetables will be destroyed) until
506 * khugepaged has finished working on the pagetables
507 * under the mmap_sem.
509 down_write(&mm->mmap_sem);
510 up_write(&mm->mmap_sem);
514 static void release_pte_page(struct page *page)
516 dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
518 putback_lru_page(page);
521 static void release_pte_pages(pte_t *pte, pte_t *_pte)
523 while (--_pte >= pte) {
524 pte_t pteval = *_pte;
525 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
526 release_pte_page(pte_page(pteval));
530 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
531 unsigned long address,
534 struct page *page = NULL;
536 int none_or_zero = 0, result = 0, referenced = 0;
537 bool writable = false;
539 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
540 _pte++, address += PAGE_SIZE) {
541 pte_t pteval = *_pte;
542 if (pte_none(pteval) || (pte_present(pteval) &&
543 is_zero_pfn(pte_pfn(pteval)))) {
544 if (!userfaultfd_armed(vma) &&
545 ++none_or_zero <= khugepaged_max_ptes_none) {
548 result = SCAN_EXCEED_NONE_PTE;
552 if (!pte_present(pteval)) {
553 result = SCAN_PTE_NON_PRESENT;
556 page = vm_normal_page(vma, address, pteval);
557 if (unlikely(!page)) {
558 result = SCAN_PAGE_NULL;
562 /* TODO: teach khugepaged to collapse THP mapped with pte */
563 if (PageCompound(page)) {
564 result = SCAN_PAGE_COMPOUND;
568 VM_BUG_ON_PAGE(!PageAnon(page), page);
571 * We can do it before isolate_lru_page because the
572 * page can't be freed from under us. NOTE: PG_lock
573 * is needed to serialize against split_huge_page
574 * when invoked from the VM.
576 if (!trylock_page(page)) {
577 result = SCAN_PAGE_LOCK;
582 * cannot use mapcount: can't collapse if there's a gup pin.
583 * The page must only be referenced by the scanned process
584 * and page swap cache.
586 if (page_count(page) != 1 + PageSwapCache(page)) {
588 result = SCAN_PAGE_COUNT;
591 if (pte_write(pteval)) {
594 if (PageSwapCache(page) &&
595 !reuse_swap_page(page, NULL)) {
597 result = SCAN_SWAP_CACHE_PAGE;
601 * Page is not in the swap cache. It can be collapsed
607 * Isolate the page to avoid collapsing an hugepage
608 * currently in use by the VM.
610 if (isolate_lru_page(page)) {
612 result = SCAN_DEL_PAGE_LRU;
615 inc_node_page_state(page,
616 NR_ISOLATED_ANON + page_is_file_cache(page));
617 VM_BUG_ON_PAGE(!PageLocked(page), page);
618 VM_BUG_ON_PAGE(PageLRU(page), page);
620 /* There should be enough young pte to collapse the page */
621 if (pte_young(pteval) ||
622 page_is_young(page) || PageReferenced(page) ||
623 mmu_notifier_test_young(vma->vm_mm, address))
626 if (likely(writable)) {
627 if (likely(referenced)) {
628 result = SCAN_SUCCEED;
629 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
630 referenced, writable, result);
634 result = SCAN_PAGE_RO;
638 release_pte_pages(pte, _pte);
639 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
640 referenced, writable, result);
644 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
645 struct vm_area_struct *vma,
646 unsigned long address,
650 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
651 _pte++, page++, address += PAGE_SIZE) {
652 pte_t pteval = *_pte;
653 struct page *src_page;
655 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
656 clear_user_highpage(page, address);
657 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
658 if (is_zero_pfn(pte_pfn(pteval))) {
660 * ptl mostly unnecessary.
664 * paravirt calls inside pte_clear here are
667 pte_clear(vma->vm_mm, address, _pte);
671 src_page = pte_page(pteval);
672 copy_user_highpage(page, src_page, address, vma);
673 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
674 release_pte_page(src_page);
676 * ptl mostly unnecessary, but preempt has to
677 * be disabled to update the per-cpu stats
678 * inside page_remove_rmap().
682 * paravirt calls inside pte_clear here are
685 pte_clear(vma->vm_mm, address, _pte);
686 page_remove_rmap(src_page, false);
688 free_page_and_swap_cache(src_page);
693 static void khugepaged_alloc_sleep(void)
697 add_wait_queue(&khugepaged_wait, &wait);
698 freezable_schedule_timeout_interruptible(
699 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
700 remove_wait_queue(&khugepaged_wait, &wait);
703 static int khugepaged_node_load[MAX_NUMNODES];
705 static bool khugepaged_scan_abort(int nid)
710 * If node_reclaim_mode is disabled, then no extra effort is made to
711 * allocate memory locally.
713 if (!node_reclaim_mode)
716 /* If there is a count for this node already, it must be acceptable */
717 if (khugepaged_node_load[nid])
720 for (i = 0; i < MAX_NUMNODES; i++) {
721 if (!khugepaged_node_load[i])
723 if (node_distance(nid, i) > node_reclaim_distance)
729 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
730 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
732 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
736 static int khugepaged_find_target_node(void)
738 static int last_khugepaged_target_node = NUMA_NO_NODE;
739 int nid, target_node = 0, max_value = 0;
741 /* find first node with max normal pages hit */
742 for (nid = 0; nid < MAX_NUMNODES; nid++)
743 if (khugepaged_node_load[nid] > max_value) {
744 max_value = khugepaged_node_load[nid];
748 /* do some balance if several nodes have the same hit record */
749 if (target_node <= last_khugepaged_target_node)
750 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
752 if (max_value == khugepaged_node_load[nid]) {
757 last_khugepaged_target_node = target_node;
761 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
763 if (IS_ERR(*hpage)) {
769 khugepaged_alloc_sleep();
779 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
781 VM_BUG_ON_PAGE(*hpage, *hpage);
783 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
784 if (unlikely(!*hpage)) {
785 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
786 *hpage = ERR_PTR(-ENOMEM);
790 prep_transhuge_page(*hpage);
791 count_vm_event(THP_COLLAPSE_ALLOC);
795 static int khugepaged_find_target_node(void)
800 static inline struct page *alloc_khugepaged_hugepage(void)
804 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
807 prep_transhuge_page(page);
811 static struct page *khugepaged_alloc_hugepage(bool *wait)
816 hpage = alloc_khugepaged_hugepage();
818 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
823 khugepaged_alloc_sleep();
825 count_vm_event(THP_COLLAPSE_ALLOC);
826 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
831 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
834 *hpage = khugepaged_alloc_hugepage(wait);
836 if (unlikely(!*hpage))
843 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
852 * If mmap_sem temporarily dropped, revalidate vma
853 * before taking mmap_sem.
854 * Return 0 if succeeds, otherwise return none-zero
858 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
859 struct vm_area_struct **vmap)
861 struct vm_area_struct *vma;
862 unsigned long hstart, hend;
864 if (unlikely(khugepaged_test_exit(mm)))
865 return SCAN_ANY_PROCESS;
867 *vmap = vma = find_vma(mm, address);
869 return SCAN_VMA_NULL;
871 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
872 hend = vma->vm_end & HPAGE_PMD_MASK;
873 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
874 return SCAN_ADDRESS_RANGE;
875 if (!hugepage_vma_check(vma, vma->vm_flags))
876 return SCAN_VMA_CHECK;
881 * Bring missing pages in from swap, to complete THP collapse.
882 * Only done if khugepaged_scan_pmd believes it is worthwhile.
884 * Called and returns without pte mapped or spinlocks held,
885 * but with mmap_sem held to protect against vma changes.
888 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
889 struct vm_area_struct *vma,
890 unsigned long address, pmd_t *pmd,
895 struct vm_fault vmf = {
898 .flags = FAULT_FLAG_ALLOW_RETRY,
900 .pgoff = linear_page_index(vma, address),
903 /* we only decide to swapin, if there is enough young ptes */
904 if (referenced < HPAGE_PMD_NR/2) {
905 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
908 vmf.pte = pte_offset_map(pmd, address);
909 for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
910 vmf.pte++, vmf.address += PAGE_SIZE) {
911 vmf.orig_pte = *vmf.pte;
912 if (!is_swap_pte(vmf.orig_pte))
915 ret = do_swap_page(&vmf);
917 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
918 if (ret & VM_FAULT_RETRY) {
919 down_read(&mm->mmap_sem);
920 if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
921 /* vma is no longer available, don't continue to swapin */
922 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
925 /* check if the pmd is still valid */
926 if (mm_find_pmd(mm, address) != pmd) {
927 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
931 if (ret & VM_FAULT_ERROR) {
932 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
935 /* pte is unmapped now, we need to map it */
936 vmf.pte = pte_offset_map(pmd, vmf.address);
940 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
944 static void collapse_huge_page(struct mm_struct *mm,
945 unsigned long address,
947 int node, int referenced)
952 struct page *new_page;
953 spinlock_t *pmd_ptl, *pte_ptl;
954 int isolated = 0, result = 0;
955 struct mem_cgroup *memcg;
956 struct vm_area_struct *vma;
957 struct mmu_notifier_range range;
960 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
962 /* Only allocate from the target node */
963 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
966 * Before allocating the hugepage, release the mmap_sem read lock.
967 * The allocation can take potentially a long time if it involves
968 * sync compaction, and we do not need to hold the mmap_sem during
969 * that. We will recheck the vma after taking it again in write mode.
971 up_read(&mm->mmap_sem);
972 new_page = khugepaged_alloc_page(hpage, gfp, node);
974 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
978 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
979 result = SCAN_CGROUP_CHARGE_FAIL;
983 down_read(&mm->mmap_sem);
984 result = hugepage_vma_revalidate(mm, address, &vma);
986 mem_cgroup_cancel_charge(new_page, memcg, true);
987 up_read(&mm->mmap_sem);
991 pmd = mm_find_pmd(mm, address);
993 result = SCAN_PMD_NULL;
994 mem_cgroup_cancel_charge(new_page, memcg, true);
995 up_read(&mm->mmap_sem);
1000 * __collapse_huge_page_swapin always returns with mmap_sem locked.
1001 * If it fails, we release mmap_sem and jump out_nolock.
1002 * Continuing to collapse causes inconsistency.
1004 if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1005 mem_cgroup_cancel_charge(new_page, memcg, true);
1006 up_read(&mm->mmap_sem);
1010 up_read(&mm->mmap_sem);
1012 * Prevent all access to pagetables with the exception of
1013 * gup_fast later handled by the ptep_clear_flush and the VM
1014 * handled by the anon_vma lock + PG_lock.
1016 down_write(&mm->mmap_sem);
1017 result = SCAN_ANY_PROCESS;
1018 if (!mmget_still_valid(mm))
1020 result = hugepage_vma_revalidate(mm, address, &vma);
1023 /* check if the pmd is still valid */
1024 if (mm_find_pmd(mm, address) != pmd)
1027 anon_vma_lock_write(vma->anon_vma);
1029 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1030 address, address + HPAGE_PMD_SIZE);
1031 mmu_notifier_invalidate_range_start(&range);
1033 pte = pte_offset_map(pmd, address);
1034 pte_ptl = pte_lockptr(mm, pmd);
1036 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1038 * After this gup_fast can't run anymore. This also removes
1039 * any huge TLB entry from the CPU so we won't allow
1040 * huge and small TLB entries for the same virtual address
1041 * to avoid the risk of CPU bugs in that area.
1043 _pmd = pmdp_collapse_flush(vma, address, pmd);
1044 spin_unlock(pmd_ptl);
1045 mmu_notifier_invalidate_range_end(&range);
1048 isolated = __collapse_huge_page_isolate(vma, address, pte);
1049 spin_unlock(pte_ptl);
1051 if (unlikely(!isolated)) {
1054 BUG_ON(!pmd_none(*pmd));
1056 * We can only use set_pmd_at when establishing
1057 * hugepmds and never for establishing regular pmds that
1058 * points to regular pagetables. Use pmd_populate for that
1060 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1061 spin_unlock(pmd_ptl);
1062 anon_vma_unlock_write(vma->anon_vma);
1068 * All pages are isolated and locked so anon_vma rmap
1069 * can't run anymore.
1071 anon_vma_unlock_write(vma->anon_vma);
1073 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1075 __SetPageUptodate(new_page);
1076 pgtable = pmd_pgtable(_pmd);
1078 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1079 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1082 * spin_lock() below is not the equivalent of smp_wmb(), so
1083 * this is needed to avoid the copy_huge_page writes to become
1084 * visible after the set_pmd_at() write.
1089 BUG_ON(!pmd_none(*pmd));
1090 page_add_new_anon_rmap(new_page, vma, address, true);
1091 mem_cgroup_commit_charge(new_page, memcg, false, true);
1092 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1093 lru_cache_add_active_or_unevictable(new_page, vma);
1094 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1095 set_pmd_at(mm, address, pmd, _pmd);
1096 update_mmu_cache_pmd(vma, address, pmd);
1097 spin_unlock(pmd_ptl);
1101 khugepaged_pages_collapsed++;
1102 result = SCAN_SUCCEED;
1104 up_write(&mm->mmap_sem);
1106 trace_mm_collapse_huge_page(mm, isolated, result);
1109 mem_cgroup_cancel_charge(new_page, memcg, true);
1113 static int khugepaged_scan_pmd(struct mm_struct *mm,
1114 struct vm_area_struct *vma,
1115 unsigned long address,
1116 struct page **hpage)
1120 int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1121 struct page *page = NULL;
1122 unsigned long _address;
1124 int node = NUMA_NO_NODE, unmapped = 0;
1125 bool writable = false;
1127 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1129 pmd = mm_find_pmd(mm, address);
1131 result = SCAN_PMD_NULL;
1135 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1136 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1137 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1138 _pte++, _address += PAGE_SIZE) {
1139 pte_t pteval = *_pte;
1140 if (is_swap_pte(pteval)) {
1141 if (++unmapped <= khugepaged_max_ptes_swap) {
1144 result = SCAN_EXCEED_SWAP_PTE;
1148 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1149 if (!userfaultfd_armed(vma) &&
1150 ++none_or_zero <= khugepaged_max_ptes_none) {
1153 result = SCAN_EXCEED_NONE_PTE;
1157 if (!pte_present(pteval)) {
1158 result = SCAN_PTE_NON_PRESENT;
1161 if (pte_write(pteval))
1164 page = vm_normal_page(vma, _address, pteval);
1165 if (unlikely(!page)) {
1166 result = SCAN_PAGE_NULL;
1170 /* TODO: teach khugepaged to collapse THP mapped with pte */
1171 if (PageCompound(page)) {
1172 result = SCAN_PAGE_COMPOUND;
1177 * Record which node the original page is from and save this
1178 * information to khugepaged_node_load[].
1179 * Khupaged will allocate hugepage from the node has the max
1182 node = page_to_nid(page);
1183 if (khugepaged_scan_abort(node)) {
1184 result = SCAN_SCAN_ABORT;
1187 khugepaged_node_load[node]++;
1188 if (!PageLRU(page)) {
1189 result = SCAN_PAGE_LRU;
1192 if (PageLocked(page)) {
1193 result = SCAN_PAGE_LOCK;
1196 if (!PageAnon(page)) {
1197 result = SCAN_PAGE_ANON;
1202 * cannot use mapcount: can't collapse if there's a gup pin.
1203 * The page must only be referenced by the scanned process
1204 * and page swap cache.
1206 if (page_count(page) != 1 + PageSwapCache(page)) {
1207 result = SCAN_PAGE_COUNT;
1210 if (pte_young(pteval) ||
1211 page_is_young(page) || PageReferenced(page) ||
1212 mmu_notifier_test_young(vma->vm_mm, address))
1217 result = SCAN_SUCCEED;
1220 result = SCAN_LACK_REFERENCED_PAGE;
1223 result = SCAN_PAGE_RO;
1226 pte_unmap_unlock(pte, ptl);
1228 node = khugepaged_find_target_node();
1229 /* collapse_huge_page will return with the mmap_sem released */
1230 collapse_huge_page(mm, address, hpage, node, referenced);
1233 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1234 none_or_zero, result, unmapped);
1238 static void collect_mm_slot(struct mm_slot *mm_slot)
1240 struct mm_struct *mm = mm_slot->mm;
1242 lockdep_assert_held(&khugepaged_mm_lock);
1244 if (khugepaged_test_exit(mm)) {
1246 hash_del(&mm_slot->hash);
1247 list_del(&mm_slot->mm_node);
1250 * Not strictly needed because the mm exited already.
1252 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1255 /* khugepaged_mm_lock actually not necessary for the below */
1256 free_mm_slot(mm_slot);
1261 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1263 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1264 * khugepaged should try to collapse the page table.
1266 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1269 struct mm_slot *mm_slot;
1271 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1273 spin_lock(&khugepaged_mm_lock);
1274 mm_slot = get_mm_slot(mm);
1275 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1276 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1277 spin_unlock(&khugepaged_mm_lock);
1282 * Try to collapse a pte-mapped THP for mm at address haddr.
1284 * This function checks whether all the PTEs in the PMD are pointing to the
1285 * right THP. If so, retract the page table so the THP can refault in with
1288 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1290 unsigned long haddr = addr & HPAGE_PMD_MASK;
1291 struct vm_area_struct *vma = find_vma(mm, haddr);
1292 struct page *hpage = NULL;
1293 pte_t *start_pte, *pte;
1299 if (!vma || !vma->vm_file ||
1300 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1304 * This vm_flags may not have VM_HUGEPAGE if the page was not
1305 * collapsed by this mm. But we can still collapse if the page is
1306 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1307 * will not fail the vma for missing VM_HUGEPAGE
1309 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1312 pmd = mm_find_pmd(mm, haddr);
1316 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1318 /* step 1: check all mapped PTEs are to the right huge page */
1319 for (i = 0, addr = haddr, pte = start_pte;
1320 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1323 /* empty pte, skip */
1327 /* page swapped out, abort */
1328 if (!pte_present(*pte))
1331 page = vm_normal_page(vma, addr, *pte);
1333 if (!page || !PageCompound(page))
1337 hpage = compound_head(page);
1339 * The mapping of the THP should not change.
1341 * Note that uprobe, debugger, or MAP_PRIVATE may
1342 * change the page table, but the new page will
1343 * not pass PageCompound() check.
1345 if (WARN_ON(hpage->mapping != vma->vm_file->f_mapping))
1350 * Confirm the page maps to the correct subpage.
1352 * Note that uprobe, debugger, or MAP_PRIVATE may change
1353 * the page table, but the new page will not pass
1354 * PageCompound() check.
1356 if (WARN_ON(hpage + i != page))
1361 /* step 2: adjust rmap */
1362 for (i = 0, addr = haddr, pte = start_pte;
1363 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1368 page = vm_normal_page(vma, addr, *pte);
1369 page_remove_rmap(page, false);
1372 pte_unmap_unlock(start_pte, ptl);
1374 /* step 3: set proper refcount and mm_counters. */
1376 page_ref_sub(hpage, count);
1377 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1380 /* step 4: collapse pmd */
1381 ptl = pmd_lock(vma->vm_mm, pmd);
1382 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1385 pte_free(mm, pmd_pgtable(_pmd));
1389 pte_unmap_unlock(start_pte, ptl);
1392 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1394 struct mm_struct *mm = mm_slot->mm;
1397 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1400 if (!down_write_trylock(&mm->mmap_sem))
1403 if (unlikely(khugepaged_test_exit(mm)))
1406 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1407 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1410 mm_slot->nr_pte_mapped_thp = 0;
1411 up_write(&mm->mmap_sem);
1415 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1417 struct vm_area_struct *vma;
1421 i_mmap_lock_write(mapping);
1422 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1424 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1425 * got written to. These VMAs are likely not worth investing
1426 * down_write(mmap_sem) as PMD-mapping is likely to be split
1429 * Not that vma->anon_vma check is racy: it can be set up after
1430 * the check but before we took mmap_sem by the fault path.
1431 * But page lock would prevent establishing any new ptes of the
1432 * page, so we are safe.
1434 * An alternative would be drop the check, but check that page
1435 * table is clear before calling pmdp_collapse_flush() under
1436 * ptl. It has higher chance to recover THP for the VMA, but
1437 * has higher cost too.
1441 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1442 if (addr & ~HPAGE_PMD_MASK)
1444 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1446 pmd = mm_find_pmd(vma->vm_mm, addr);
1450 * We need exclusive mmap_sem to retract page table.
1452 * We use trylock due to lock inversion: we need to acquire
1453 * mmap_sem while holding page lock. Fault path does it in
1454 * reverse order. Trylock is a way to avoid deadlock.
1456 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1457 spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1458 /* assume page table is clear */
1459 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1461 up_write(&vma->vm_mm->mmap_sem);
1462 mm_dec_nr_ptes(vma->vm_mm);
1463 pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1465 /* Try again later */
1466 khugepaged_add_pte_mapped_thp(vma->vm_mm, addr);
1469 i_mmap_unlock_write(mapping);
1473 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1475 * Basic scheme is simple, details are more complex:
1476 * - allocate and lock a new huge page;
1477 * - scan page cache replacing old pages with the new one
1478 * + swap/gup in pages if necessary;
1480 * + keep old pages around in case rollback is required;
1481 * - if replacing succeeds:
1484 * + unlock huge page;
1485 * - if replacing failed;
1486 * + put all pages back and unfreeze them;
1487 * + restore gaps in the page cache;
1488 * + unlock and free huge page;
1490 static void collapse_file(struct mm_struct *mm,
1491 struct file *file, pgoff_t start,
1492 struct page **hpage, int node)
1494 struct address_space *mapping = file->f_mapping;
1496 struct page *new_page;
1497 struct mem_cgroup *memcg;
1498 pgoff_t index, end = start + HPAGE_PMD_NR;
1499 LIST_HEAD(pagelist);
1500 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1501 int nr_none = 0, result = SCAN_SUCCEED;
1502 bool is_shmem = shmem_file(file);
1504 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1505 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1507 /* Only allocate from the target node */
1508 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1510 new_page = khugepaged_alloc_page(hpage, gfp, node);
1512 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1516 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1517 result = SCAN_CGROUP_CHARGE_FAIL;
1521 /* This will be less messy when we use multi-index entries */
1524 xas_create_range(&xas);
1525 if (!xas_error(&xas))
1527 xas_unlock_irq(&xas);
1528 if (!xas_nomem(&xas, GFP_KERNEL)) {
1529 mem_cgroup_cancel_charge(new_page, memcg, true);
1535 __SetPageLocked(new_page);
1537 __SetPageSwapBacked(new_page);
1538 new_page->index = start;
1539 new_page->mapping = mapping;
1542 * At this point the new_page is locked and not up-to-date.
1543 * It's safe to insert it into the page cache, because nobody would
1544 * be able to map it or use it in another way until we unlock it.
1547 xas_set(&xas, start);
1548 for (index = start; index < end; index++) {
1549 struct page *page = xas_next(&xas);
1551 VM_BUG_ON(index != xas.xa_index);
1555 * Stop if extent has been truncated or
1556 * hole-punched, and is now completely
1559 if (index == start) {
1560 if (!xas_next_entry(&xas, end - 1)) {
1561 result = SCAN_TRUNCATED;
1564 xas_set(&xas, index);
1566 if (!shmem_charge(mapping->host, 1)) {
1570 xas_store(&xas, new_page);
1575 if (xa_is_value(page) || !PageUptodate(page)) {
1576 xas_unlock_irq(&xas);
1577 /* swap in or instantiate fallocated page */
1578 if (shmem_getpage(mapping->host, index, &page,
1583 } else if (trylock_page(page)) {
1585 xas_unlock_irq(&xas);
1587 result = SCAN_PAGE_LOCK;
1590 } else { /* !is_shmem */
1591 if (!page || xa_is_value(page)) {
1592 xas_unlock_irq(&xas);
1593 page_cache_sync_readahead(mapping, &file->f_ra,
1596 /* drain pagevecs to help isolate_lru_page() */
1598 page = find_lock_page(mapping, index);
1599 if (unlikely(page == NULL)) {
1603 } else if (PageDirty(page)) {
1605 * khugepaged only works on read-only fd,
1606 * so this page is dirty because it hasn't
1607 * been flushed since first write. There
1608 * won't be new dirty pages.
1610 * Trigger async flush here and hope the
1611 * writeback is done when khugepaged
1612 * revisits this page.
1614 * This is a one-off situation. We are not
1615 * forcing writeback in loop.
1617 xas_unlock_irq(&xas);
1618 filemap_flush(mapping);
1621 } else if (trylock_page(page)) {
1623 xas_unlock_irq(&xas);
1625 result = SCAN_PAGE_LOCK;
1631 * The page must be locked, so we can drop the i_pages lock
1632 * without racing with truncate.
1634 VM_BUG_ON_PAGE(!PageLocked(page), page);
1636 /* make sure the page is up to date */
1637 if (unlikely(!PageUptodate(page))) {
1643 * If file was truncated then extended, or hole-punched, before
1644 * we locked the first page, then a THP might be there already.
1646 if (PageTransCompound(page)) {
1647 result = SCAN_PAGE_COMPOUND;
1651 if (page_mapping(page) != mapping) {
1652 result = SCAN_TRUNCATED;
1656 if (!is_shmem && PageDirty(page)) {
1658 * khugepaged only works on read-only fd, so this
1659 * page is dirty because it hasn't been flushed
1660 * since first write.
1666 if (isolate_lru_page(page)) {
1667 result = SCAN_DEL_PAGE_LRU;
1671 if (page_has_private(page) &&
1672 !try_to_release_page(page, GFP_KERNEL)) {
1673 result = SCAN_PAGE_HAS_PRIVATE;
1677 if (page_mapped(page))
1678 unmap_mapping_pages(mapping, index, 1, false);
1681 xas_set(&xas, index);
1683 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1684 VM_BUG_ON_PAGE(page_mapped(page), page);
1687 * The page is expected to have page_count() == 3:
1688 * - we hold a pin on it;
1689 * - one reference from page cache;
1690 * - one from isolate_lru_page;
1692 if (!page_ref_freeze(page, 3)) {
1693 result = SCAN_PAGE_COUNT;
1694 xas_unlock_irq(&xas);
1695 putback_lru_page(page);
1700 * Add the page to the list to be able to undo the collapse if
1701 * something go wrong.
1703 list_add_tail(&page->lru, &pagelist);
1705 /* Finally, replace with the new page. */
1706 xas_store(&xas, new_page);
1715 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1717 __inc_node_page_state(new_page, NR_FILE_THPS);
1718 filemap_nr_thps_inc(mapping);
1722 struct zone *zone = page_zone(new_page);
1724 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1726 __mod_node_page_state(zone->zone_pgdat,
1731 xas_unlock_irq(&xas);
1734 if (result == SCAN_SUCCEED) {
1735 struct page *page, *tmp;
1738 * Replacing old pages with new one has succeeded, now we
1739 * need to copy the content and free the old pages.
1742 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1743 while (index < page->index) {
1744 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1747 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1749 list_del(&page->lru);
1750 page->mapping = NULL;
1751 page_ref_unfreeze(page, 1);
1752 ClearPageActive(page);
1753 ClearPageUnevictable(page);
1758 while (index < end) {
1759 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1763 SetPageUptodate(new_page);
1764 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1765 mem_cgroup_commit_charge(new_page, memcg, false, true);
1768 set_page_dirty(new_page);
1769 lru_cache_add_anon(new_page);
1771 lru_cache_add_file(new_page);
1773 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1776 * Remove pte page tables, so we can re-fault the page as huge.
1778 retract_page_tables(mapping, start);
1781 khugepaged_pages_collapsed++;
1785 /* Something went wrong: roll back page cache changes */
1787 mapping->nrpages -= nr_none;
1790 shmem_uncharge(mapping->host, nr_none);
1792 xas_set(&xas, start);
1793 xas_for_each(&xas, page, end - 1) {
1794 page = list_first_entry_or_null(&pagelist,
1796 if (!page || xas.xa_index < page->index) {
1800 /* Put holes back where they were */
1801 xas_store(&xas, NULL);
1805 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1807 /* Unfreeze the page. */
1808 list_del(&page->lru);
1809 page_ref_unfreeze(page, 2);
1810 xas_store(&xas, page);
1812 xas_unlock_irq(&xas);
1814 putback_lru_page(page);
1818 xas_unlock_irq(&xas);
1820 mem_cgroup_cancel_charge(new_page, memcg, true);
1821 new_page->mapping = NULL;
1824 unlock_page(new_page);
1826 VM_BUG_ON(!list_empty(&pagelist));
1827 /* TODO: tracepoints */
1830 static void khugepaged_scan_file(struct mm_struct *mm,
1831 struct file *file, pgoff_t start, struct page **hpage)
1833 struct page *page = NULL;
1834 struct address_space *mapping = file->f_mapping;
1835 XA_STATE(xas, &mapping->i_pages, start);
1837 int node = NUMA_NO_NODE;
1838 int result = SCAN_SUCCEED;
1842 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1844 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1845 if (xas_retry(&xas, page))
1848 if (xa_is_value(page)) {
1849 if (++swap > khugepaged_max_ptes_swap) {
1850 result = SCAN_EXCEED_SWAP_PTE;
1856 if (PageTransCompound(page)) {
1857 result = SCAN_PAGE_COMPOUND;
1861 node = page_to_nid(page);
1862 if (khugepaged_scan_abort(node)) {
1863 result = SCAN_SCAN_ABORT;
1866 khugepaged_node_load[node]++;
1868 if (!PageLRU(page)) {
1869 result = SCAN_PAGE_LRU;
1873 if (page_count(page) !=
1874 1 + page_mapcount(page) + page_has_private(page)) {
1875 result = SCAN_PAGE_COUNT;
1880 * We probably should check if the page is referenced here, but
1881 * nobody would transfer pte_young() to PageReferenced() for us.
1882 * And rmap walk here is just too costly...
1887 if (need_resched()) {
1894 if (result == SCAN_SUCCEED) {
1895 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1896 result = SCAN_EXCEED_NONE_PTE;
1898 node = khugepaged_find_target_node();
1899 collapse_file(mm, file, start, hpage, node);
1903 /* TODO: tracepoints */
1906 static void khugepaged_scan_file(struct mm_struct *mm,
1907 struct file *file, pgoff_t start, struct page **hpage)
1912 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1918 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1919 struct page **hpage)
1920 __releases(&khugepaged_mm_lock)
1921 __acquires(&khugepaged_mm_lock)
1923 struct mm_slot *mm_slot;
1924 struct mm_struct *mm;
1925 struct vm_area_struct *vma;
1929 lockdep_assert_held(&khugepaged_mm_lock);
1931 if (khugepaged_scan.mm_slot)
1932 mm_slot = khugepaged_scan.mm_slot;
1934 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1935 struct mm_slot, mm_node);
1936 khugepaged_scan.address = 0;
1937 khugepaged_scan.mm_slot = mm_slot;
1939 spin_unlock(&khugepaged_mm_lock);
1940 khugepaged_collapse_pte_mapped_thps(mm_slot);
1944 * Don't wait for semaphore (to avoid long wait times). Just move to
1945 * the next mm on the list.
1948 if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1949 goto breakouterloop_mmap_sem;
1950 if (likely(!khugepaged_test_exit(mm)))
1951 vma = find_vma(mm, khugepaged_scan.address);
1954 for (; vma; vma = vma->vm_next) {
1955 unsigned long hstart, hend;
1958 if (unlikely(khugepaged_test_exit(mm))) {
1962 if (!hugepage_vma_check(vma, vma->vm_flags)) {
1967 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1968 hend = vma->vm_end & HPAGE_PMD_MASK;
1971 if (khugepaged_scan.address > hend)
1973 if (khugepaged_scan.address < hstart)
1974 khugepaged_scan.address = hstart;
1975 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1977 while (khugepaged_scan.address < hend) {
1980 if (unlikely(khugepaged_test_exit(mm)))
1981 goto breakouterloop;
1983 VM_BUG_ON(khugepaged_scan.address < hstart ||
1984 khugepaged_scan.address + HPAGE_PMD_SIZE >
1986 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
1988 pgoff_t pgoff = linear_page_index(vma,
1989 khugepaged_scan.address);
1991 if (shmem_file(vma->vm_file)
1992 && !shmem_huge_enabled(vma))
1994 file = get_file(vma->vm_file);
1995 up_read(&mm->mmap_sem);
1997 khugepaged_scan_file(mm, file, pgoff, hpage);
2000 ret = khugepaged_scan_pmd(mm, vma,
2001 khugepaged_scan.address,
2004 /* move to next address */
2005 khugepaged_scan.address += HPAGE_PMD_SIZE;
2006 progress += HPAGE_PMD_NR;
2008 /* we released mmap_sem so break loop */
2009 goto breakouterloop_mmap_sem;
2010 if (progress >= pages)
2011 goto breakouterloop;
2015 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2016 breakouterloop_mmap_sem:
2018 spin_lock(&khugepaged_mm_lock);
2019 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2021 * Release the current mm_slot if this mm is about to die, or
2022 * if we scanned all vmas of this mm.
2024 if (khugepaged_test_exit(mm) || !vma) {
2026 * Make sure that if mm_users is reaching zero while
2027 * khugepaged runs here, khugepaged_exit will find
2028 * mm_slot not pointing to the exiting mm.
2030 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2031 khugepaged_scan.mm_slot = list_entry(
2032 mm_slot->mm_node.next,
2033 struct mm_slot, mm_node);
2034 khugepaged_scan.address = 0;
2036 khugepaged_scan.mm_slot = NULL;
2037 khugepaged_full_scans++;
2040 collect_mm_slot(mm_slot);
2046 static int khugepaged_has_work(void)
2048 return !list_empty(&khugepaged_scan.mm_head) &&
2049 khugepaged_enabled();
2052 static int khugepaged_wait_event(void)
2054 return !list_empty(&khugepaged_scan.mm_head) ||
2055 kthread_should_stop();
2058 static void khugepaged_do_scan(void)
2060 struct page *hpage = NULL;
2061 unsigned int progress = 0, pass_through_head = 0;
2062 unsigned int pages = khugepaged_pages_to_scan;
2065 barrier(); /* write khugepaged_pages_to_scan to local stack */
2067 while (progress < pages) {
2068 if (!khugepaged_prealloc_page(&hpage, &wait))
2073 if (unlikely(kthread_should_stop() || try_to_freeze()))
2076 spin_lock(&khugepaged_mm_lock);
2077 if (!khugepaged_scan.mm_slot)
2078 pass_through_head++;
2079 if (khugepaged_has_work() &&
2080 pass_through_head < 2)
2081 progress += khugepaged_scan_mm_slot(pages - progress,
2085 spin_unlock(&khugepaged_mm_lock);
2088 if (!IS_ERR_OR_NULL(hpage))
2092 static bool khugepaged_should_wakeup(void)
2094 return kthread_should_stop() ||
2095 time_after_eq(jiffies, khugepaged_sleep_expire);
2098 static void khugepaged_wait_work(void)
2100 if (khugepaged_has_work()) {
2101 const unsigned long scan_sleep_jiffies =
2102 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2104 if (!scan_sleep_jiffies)
2107 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2108 wait_event_freezable_timeout(khugepaged_wait,
2109 khugepaged_should_wakeup(),
2110 scan_sleep_jiffies);
2114 if (khugepaged_enabled())
2115 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2118 static int khugepaged(void *none)
2120 struct mm_slot *mm_slot;
2123 set_user_nice(current, MAX_NICE);
2125 while (!kthread_should_stop()) {
2126 khugepaged_do_scan();
2127 khugepaged_wait_work();
2130 spin_lock(&khugepaged_mm_lock);
2131 mm_slot = khugepaged_scan.mm_slot;
2132 khugepaged_scan.mm_slot = NULL;
2134 collect_mm_slot(mm_slot);
2135 spin_unlock(&khugepaged_mm_lock);
2139 static void set_recommended_min_free_kbytes(void)
2143 unsigned long recommended_min;
2145 for_each_populated_zone(zone) {
2147 * We don't need to worry about fragmentation of
2148 * ZONE_MOVABLE since it only has movable pages.
2150 if (zone_idx(zone) > gfp_zone(GFP_USER))
2156 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2157 recommended_min = pageblock_nr_pages * nr_zones * 2;
2160 * Make sure that on average at least two pageblocks are almost free
2161 * of another type, one for a migratetype to fall back to and a
2162 * second to avoid subsequent fallbacks of other types There are 3
2163 * MIGRATE_TYPES we care about.
2165 recommended_min += pageblock_nr_pages * nr_zones *
2166 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2168 /* don't ever allow to reserve more than 5% of the lowmem */
2169 recommended_min = min(recommended_min,
2170 (unsigned long) nr_free_buffer_pages() / 20);
2171 recommended_min <<= (PAGE_SHIFT-10);
2173 if (recommended_min > min_free_kbytes) {
2174 if (user_min_free_kbytes >= 0)
2175 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2176 min_free_kbytes, recommended_min);
2178 min_free_kbytes = recommended_min;
2180 setup_per_zone_wmarks();
2183 int start_stop_khugepaged(void)
2185 static struct task_struct *khugepaged_thread __read_mostly;
2186 static DEFINE_MUTEX(khugepaged_mutex);
2189 mutex_lock(&khugepaged_mutex);
2190 if (khugepaged_enabled()) {
2191 if (!khugepaged_thread)
2192 khugepaged_thread = kthread_run(khugepaged, NULL,
2194 if (IS_ERR(khugepaged_thread)) {
2195 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2196 err = PTR_ERR(khugepaged_thread);
2197 khugepaged_thread = NULL;
2201 if (!list_empty(&khugepaged_scan.mm_head))
2202 wake_up_interruptible(&khugepaged_wait);
2204 set_recommended_min_free_kbytes();
2205 } else if (khugepaged_thread) {
2206 kthread_stop(khugepaged_thread);
2207 khugepaged_thread = NULL;
2210 mutex_unlock(&khugepaged_mutex);