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 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
313 int hugepage_madvise(struct vm_area_struct *vma,
314 unsigned long *vm_flags, int advice)
320 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
321 * can't handle this properly after s390_enable_sie, so we simply
322 * ignore the madvise to prevent qemu from causing a SIGSEGV.
324 if (mm_has_pgste(vma->vm_mm))
327 *vm_flags &= ~VM_NOHUGEPAGE;
328 *vm_flags |= VM_HUGEPAGE;
330 * If the vma become good for khugepaged to scan,
331 * register it here without waiting a page fault that
332 * may not happen any time soon.
334 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
335 khugepaged_enter_vma_merge(vma, *vm_flags))
338 case MADV_NOHUGEPAGE:
339 *vm_flags &= ~VM_HUGEPAGE;
340 *vm_flags |= VM_NOHUGEPAGE;
342 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
343 * this vma even if we leave the mm registered in khugepaged if
344 * it got registered before VM_NOHUGEPAGE was set.
352 int __init khugepaged_init(void)
354 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
355 sizeof(struct mm_slot),
356 __alignof__(struct mm_slot), 0, NULL);
360 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
361 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
362 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
367 void __init khugepaged_destroy(void)
369 kmem_cache_destroy(mm_slot_cache);
372 static inline struct mm_slot *alloc_mm_slot(void)
374 if (!mm_slot_cache) /* initialization failed */
376 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
379 static inline void free_mm_slot(struct mm_slot *mm_slot)
381 kmem_cache_free(mm_slot_cache, mm_slot);
384 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
386 struct mm_slot *mm_slot;
388 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
389 if (mm == mm_slot->mm)
395 static void insert_to_mm_slots_hash(struct mm_struct *mm,
396 struct mm_slot *mm_slot)
399 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
402 static inline int khugepaged_test_exit(struct mm_struct *mm)
404 return atomic_read(&mm->mm_users) == 0;
407 static bool hugepage_vma_check(struct vm_area_struct *vma,
408 unsigned long vm_flags)
410 if ((!(vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
411 (vm_flags & VM_NOHUGEPAGE) ||
412 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
415 if (shmem_file(vma->vm_file) ||
416 (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) &&
418 (vm_flags & VM_DENYWRITE))) {
419 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
421 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
424 if (!vma->anon_vma || vma->vm_ops)
426 if (is_vma_temporary_stack(vma))
428 return !(vm_flags & VM_NO_KHUGEPAGED);
431 int __khugepaged_enter(struct mm_struct *mm)
433 struct mm_slot *mm_slot;
436 mm_slot = alloc_mm_slot();
440 /* __khugepaged_exit() must not run from under us */
441 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
442 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
443 free_mm_slot(mm_slot);
447 spin_lock(&khugepaged_mm_lock);
448 insert_to_mm_slots_hash(mm, mm_slot);
450 * Insert just behind the scanning cursor, to let the area settle
453 wakeup = list_empty(&khugepaged_scan.mm_head);
454 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
455 spin_unlock(&khugepaged_mm_lock);
459 wake_up_interruptible(&khugepaged_wait);
464 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
465 unsigned long vm_flags)
467 unsigned long hstart, hend;
470 * khugepaged only supports read-only files for non-shmem files.
471 * khugepaged does not yet work on special mappings. And
472 * file-private shmem THP is not supported.
474 if (!hugepage_vma_check(vma, vm_flags))
477 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
478 hend = vma->vm_end & HPAGE_PMD_MASK;
480 return khugepaged_enter(vma, vm_flags);
484 void __khugepaged_exit(struct mm_struct *mm)
486 struct mm_slot *mm_slot;
489 spin_lock(&khugepaged_mm_lock);
490 mm_slot = get_mm_slot(mm);
491 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
492 hash_del(&mm_slot->hash);
493 list_del(&mm_slot->mm_node);
496 spin_unlock(&khugepaged_mm_lock);
499 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
500 free_mm_slot(mm_slot);
502 } else if (mm_slot) {
504 * This is required to serialize against
505 * khugepaged_test_exit() (which is guaranteed to run
506 * under mmap sem read mode). Stop here (after we
507 * return all pagetables will be destroyed) until
508 * khugepaged has finished working on the pagetables
509 * under the mmap_sem.
511 down_write(&mm->mmap_sem);
512 up_write(&mm->mmap_sem);
516 static void release_pte_page(struct page *page)
518 dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
520 putback_lru_page(page);
523 static void release_pte_pages(pte_t *pte, pte_t *_pte)
525 while (--_pte >= pte) {
526 pte_t pteval = *_pte;
527 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
528 release_pte_page(pte_page(pteval));
532 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
533 unsigned long address,
536 struct page *page = NULL;
538 int none_or_zero = 0, result = 0, referenced = 0;
539 bool writable = false;
541 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
542 _pte++, address += PAGE_SIZE) {
543 pte_t pteval = *_pte;
544 if (pte_none(pteval) || (pte_present(pteval) &&
545 is_zero_pfn(pte_pfn(pteval)))) {
546 if (!userfaultfd_armed(vma) &&
547 ++none_or_zero <= khugepaged_max_ptes_none) {
550 result = SCAN_EXCEED_NONE_PTE;
554 if (!pte_present(pteval)) {
555 result = SCAN_PTE_NON_PRESENT;
558 page = vm_normal_page(vma, address, pteval);
559 if (unlikely(!page)) {
560 result = SCAN_PAGE_NULL;
564 /* TODO: teach khugepaged to collapse THP mapped with pte */
565 if (PageCompound(page)) {
566 result = SCAN_PAGE_COMPOUND;
570 VM_BUG_ON_PAGE(!PageAnon(page), page);
573 * We can do it before isolate_lru_page because the
574 * page can't be freed from under us. NOTE: PG_lock
575 * is needed to serialize against split_huge_page
576 * when invoked from the VM.
578 if (!trylock_page(page)) {
579 result = SCAN_PAGE_LOCK;
584 * cannot use mapcount: can't collapse if there's a gup pin.
585 * The page must only be referenced by the scanned process
586 * and page swap cache.
588 if (page_count(page) != 1 + PageSwapCache(page)) {
590 result = SCAN_PAGE_COUNT;
593 if (pte_write(pteval)) {
596 if (PageSwapCache(page) &&
597 !reuse_swap_page(page, NULL)) {
599 result = SCAN_SWAP_CACHE_PAGE;
603 * Page is not in the swap cache. It can be collapsed
609 * Isolate the page to avoid collapsing an hugepage
610 * currently in use by the VM.
612 if (isolate_lru_page(page)) {
614 result = SCAN_DEL_PAGE_LRU;
617 inc_node_page_state(page,
618 NR_ISOLATED_ANON + page_is_file_cache(page));
619 VM_BUG_ON_PAGE(!PageLocked(page), page);
620 VM_BUG_ON_PAGE(PageLRU(page), page);
622 /* There should be enough young pte to collapse the page */
623 if (pte_young(pteval) ||
624 page_is_young(page) || PageReferenced(page) ||
625 mmu_notifier_test_young(vma->vm_mm, address))
628 if (likely(writable)) {
629 if (likely(referenced)) {
630 result = SCAN_SUCCEED;
631 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
632 referenced, writable, result);
636 result = SCAN_PAGE_RO;
640 release_pte_pages(pte, _pte);
641 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
642 referenced, writable, result);
646 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
647 struct vm_area_struct *vma,
648 unsigned long address,
652 for (_pte = pte; _pte < pte + HPAGE_PMD_NR;
653 _pte++, page++, address += PAGE_SIZE) {
654 pte_t pteval = *_pte;
655 struct page *src_page;
657 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
658 clear_user_highpage(page, address);
659 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
660 if (is_zero_pfn(pte_pfn(pteval))) {
662 * ptl mostly unnecessary.
666 * paravirt calls inside pte_clear here are
669 pte_clear(vma->vm_mm, address, _pte);
673 src_page = pte_page(pteval);
674 copy_user_highpage(page, src_page, address, vma);
675 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
676 release_pte_page(src_page);
678 * ptl mostly unnecessary, but preempt has to
679 * be disabled to update the per-cpu stats
680 * inside page_remove_rmap().
684 * paravirt calls inside pte_clear here are
687 pte_clear(vma->vm_mm, address, _pte);
688 page_remove_rmap(src_page, false);
690 free_page_and_swap_cache(src_page);
695 static void khugepaged_alloc_sleep(void)
699 add_wait_queue(&khugepaged_wait, &wait);
700 freezable_schedule_timeout_interruptible(
701 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
702 remove_wait_queue(&khugepaged_wait, &wait);
705 static int khugepaged_node_load[MAX_NUMNODES];
707 static bool khugepaged_scan_abort(int nid)
712 * If node_reclaim_mode is disabled, then no extra effort is made to
713 * allocate memory locally.
715 if (!node_reclaim_mode)
718 /* If there is a count for this node already, it must be acceptable */
719 if (khugepaged_node_load[nid])
722 for (i = 0; i < MAX_NUMNODES; i++) {
723 if (!khugepaged_node_load[i])
725 if (node_distance(nid, i) > node_reclaim_distance)
731 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
732 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
734 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
738 static int khugepaged_find_target_node(void)
740 static int last_khugepaged_target_node = NUMA_NO_NODE;
741 int nid, target_node = 0, max_value = 0;
743 /* find first node with max normal pages hit */
744 for (nid = 0; nid < MAX_NUMNODES; nid++)
745 if (khugepaged_node_load[nid] > max_value) {
746 max_value = khugepaged_node_load[nid];
750 /* do some balance if several nodes have the same hit record */
751 if (target_node <= last_khugepaged_target_node)
752 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
754 if (max_value == khugepaged_node_load[nid]) {
759 last_khugepaged_target_node = target_node;
763 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
765 if (IS_ERR(*hpage)) {
771 khugepaged_alloc_sleep();
781 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
783 VM_BUG_ON_PAGE(*hpage, *hpage);
785 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
786 if (unlikely(!*hpage)) {
787 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
788 *hpage = ERR_PTR(-ENOMEM);
792 prep_transhuge_page(*hpage);
793 count_vm_event(THP_COLLAPSE_ALLOC);
797 static int khugepaged_find_target_node(void)
802 static inline struct page *alloc_khugepaged_hugepage(void)
806 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
809 prep_transhuge_page(page);
813 static struct page *khugepaged_alloc_hugepage(bool *wait)
818 hpage = alloc_khugepaged_hugepage();
820 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
825 khugepaged_alloc_sleep();
827 count_vm_event(THP_COLLAPSE_ALLOC);
828 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
833 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
836 *hpage = khugepaged_alloc_hugepage(wait);
838 if (unlikely(!*hpage))
845 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
854 * If mmap_sem temporarily dropped, revalidate vma
855 * before taking mmap_sem.
856 * Return 0 if succeeds, otherwise return none-zero
860 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
861 struct vm_area_struct **vmap)
863 struct vm_area_struct *vma;
864 unsigned long hstart, hend;
866 if (unlikely(khugepaged_test_exit(mm)))
867 return SCAN_ANY_PROCESS;
869 *vmap = vma = find_vma(mm, address);
871 return SCAN_VMA_NULL;
873 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
874 hend = vma->vm_end & HPAGE_PMD_MASK;
875 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
876 return SCAN_ADDRESS_RANGE;
877 if (!hugepage_vma_check(vma, vma->vm_flags))
878 return SCAN_VMA_CHECK;
883 * Bring missing pages in from swap, to complete THP collapse.
884 * Only done if khugepaged_scan_pmd believes it is worthwhile.
886 * Called and returns without pte mapped or spinlocks held,
887 * but with mmap_sem held to protect against vma changes.
890 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
891 struct vm_area_struct *vma,
892 unsigned long address, pmd_t *pmd,
897 struct vm_fault vmf = {
900 .flags = FAULT_FLAG_ALLOW_RETRY,
902 .pgoff = linear_page_index(vma, address),
905 /* we only decide to swapin, if there is enough young ptes */
906 if (referenced < HPAGE_PMD_NR/2) {
907 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
910 vmf.pte = pte_offset_map(pmd, address);
911 for (; vmf.address < address + HPAGE_PMD_NR*PAGE_SIZE;
912 vmf.pte++, vmf.address += PAGE_SIZE) {
913 vmf.orig_pte = *vmf.pte;
914 if (!is_swap_pte(vmf.orig_pte))
917 ret = do_swap_page(&vmf);
919 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
920 if (ret & VM_FAULT_RETRY) {
921 down_read(&mm->mmap_sem);
922 if (hugepage_vma_revalidate(mm, address, &vmf.vma)) {
923 /* vma is no longer available, don't continue to swapin */
924 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
927 /* check if the pmd is still valid */
928 if (mm_find_pmd(mm, address) != pmd) {
929 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
933 if (ret & VM_FAULT_ERROR) {
934 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
937 /* pte is unmapped now, we need to map it */
938 vmf.pte = pte_offset_map(pmd, vmf.address);
942 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
946 static void collapse_huge_page(struct mm_struct *mm,
947 unsigned long address,
949 int node, int referenced)
954 struct page *new_page;
955 spinlock_t *pmd_ptl, *pte_ptl;
956 int isolated = 0, result = 0;
957 struct mem_cgroup *memcg;
958 struct vm_area_struct *vma;
959 struct mmu_notifier_range range;
962 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
964 /* Only allocate from the target node */
965 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
968 * Before allocating the hugepage, release the mmap_sem read lock.
969 * The allocation can take potentially a long time if it involves
970 * sync compaction, and we do not need to hold the mmap_sem during
971 * that. We will recheck the vma after taking it again in write mode.
973 up_read(&mm->mmap_sem);
974 new_page = khugepaged_alloc_page(hpage, gfp, node);
976 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
980 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
981 result = SCAN_CGROUP_CHARGE_FAIL;
985 down_read(&mm->mmap_sem);
986 result = hugepage_vma_revalidate(mm, address, &vma);
988 mem_cgroup_cancel_charge(new_page, memcg, true);
989 up_read(&mm->mmap_sem);
993 pmd = mm_find_pmd(mm, address);
995 result = SCAN_PMD_NULL;
996 mem_cgroup_cancel_charge(new_page, memcg, true);
997 up_read(&mm->mmap_sem);
1002 * __collapse_huge_page_swapin always returns with mmap_sem locked.
1003 * If it fails, we release mmap_sem and jump out_nolock.
1004 * Continuing to collapse causes inconsistency.
1006 if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1007 mem_cgroup_cancel_charge(new_page, memcg, true);
1008 up_read(&mm->mmap_sem);
1012 up_read(&mm->mmap_sem);
1014 * Prevent all access to pagetables with the exception of
1015 * gup_fast later handled by the ptep_clear_flush and the VM
1016 * handled by the anon_vma lock + PG_lock.
1018 down_write(&mm->mmap_sem);
1019 result = SCAN_ANY_PROCESS;
1020 if (!mmget_still_valid(mm))
1022 result = hugepage_vma_revalidate(mm, address, &vma);
1025 /* check if the pmd is still valid */
1026 if (mm_find_pmd(mm, address) != pmd)
1029 anon_vma_lock_write(vma->anon_vma);
1031 pte = pte_offset_map(pmd, address);
1032 pte_ptl = pte_lockptr(mm, pmd);
1034 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
1035 address, address + HPAGE_PMD_SIZE);
1036 mmu_notifier_invalidate_range_start(&range);
1037 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1039 * After this gup_fast can't run anymore. This also removes
1040 * any huge TLB entry from the CPU so we won't allow
1041 * huge and small TLB entries for the same virtual address
1042 * to avoid the risk of CPU bugs in that area.
1044 _pmd = pmdp_collapse_flush(vma, address, pmd);
1045 spin_unlock(pmd_ptl);
1046 mmu_notifier_invalidate_range_end(&range);
1049 isolated = __collapse_huge_page_isolate(vma, address, pte);
1050 spin_unlock(pte_ptl);
1052 if (unlikely(!isolated)) {
1055 BUG_ON(!pmd_none(*pmd));
1057 * We can only use set_pmd_at when establishing
1058 * hugepmds and never for establishing regular pmds that
1059 * points to regular pagetables. Use pmd_populate for that
1061 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1062 spin_unlock(pmd_ptl);
1063 anon_vma_unlock_write(vma->anon_vma);
1069 * All pages are isolated and locked so anon_vma rmap
1070 * can't run anymore.
1072 anon_vma_unlock_write(vma->anon_vma);
1074 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1076 __SetPageUptodate(new_page);
1077 pgtable = pmd_pgtable(_pmd);
1079 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1080 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1083 * spin_lock() below is not the equivalent of smp_wmb(), so
1084 * this is needed to avoid the copy_huge_page writes to become
1085 * visible after the set_pmd_at() write.
1090 BUG_ON(!pmd_none(*pmd));
1091 page_add_new_anon_rmap(new_page, vma, address, true);
1092 mem_cgroup_commit_charge(new_page, memcg, false, true);
1093 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1094 lru_cache_add_active_or_unevictable(new_page, vma);
1095 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1096 set_pmd_at(mm, address, pmd, _pmd);
1097 update_mmu_cache_pmd(vma, address, pmd);
1098 spin_unlock(pmd_ptl);
1102 khugepaged_pages_collapsed++;
1103 result = SCAN_SUCCEED;
1105 up_write(&mm->mmap_sem);
1107 trace_mm_collapse_huge_page(mm, isolated, result);
1110 mem_cgroup_cancel_charge(new_page, memcg, true);
1114 static int khugepaged_scan_pmd(struct mm_struct *mm,
1115 struct vm_area_struct *vma,
1116 unsigned long address,
1117 struct page **hpage)
1121 int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1122 struct page *page = NULL;
1123 unsigned long _address;
1125 int node = NUMA_NO_NODE, unmapped = 0;
1126 bool writable = false;
1128 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1130 pmd = mm_find_pmd(mm, address);
1132 result = SCAN_PMD_NULL;
1136 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1137 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1138 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1139 _pte++, _address += PAGE_SIZE) {
1140 pte_t pteval = *_pte;
1141 if (is_swap_pte(pteval)) {
1142 if (++unmapped <= khugepaged_max_ptes_swap) {
1145 result = SCAN_EXCEED_SWAP_PTE;
1149 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1150 if (!userfaultfd_armed(vma) &&
1151 ++none_or_zero <= khugepaged_max_ptes_none) {
1154 result = SCAN_EXCEED_NONE_PTE;
1158 if (!pte_present(pteval)) {
1159 result = SCAN_PTE_NON_PRESENT;
1162 if (pte_write(pteval))
1165 page = vm_normal_page(vma, _address, pteval);
1166 if (unlikely(!page)) {
1167 result = SCAN_PAGE_NULL;
1171 /* TODO: teach khugepaged to collapse THP mapped with pte */
1172 if (PageCompound(page)) {
1173 result = SCAN_PAGE_COMPOUND;
1178 * Record which node the original page is from and save this
1179 * information to khugepaged_node_load[].
1180 * Khupaged will allocate hugepage from the node has the max
1183 node = page_to_nid(page);
1184 if (khugepaged_scan_abort(node)) {
1185 result = SCAN_SCAN_ABORT;
1188 khugepaged_node_load[node]++;
1189 if (!PageLRU(page)) {
1190 result = SCAN_PAGE_LRU;
1193 if (PageLocked(page)) {
1194 result = SCAN_PAGE_LOCK;
1197 if (!PageAnon(page)) {
1198 result = SCAN_PAGE_ANON;
1203 * cannot use mapcount: can't collapse if there's a gup pin.
1204 * The page must only be referenced by the scanned process
1205 * and page swap cache.
1207 if (page_count(page) != 1 + PageSwapCache(page)) {
1208 result = SCAN_PAGE_COUNT;
1211 if (pte_young(pteval) ||
1212 page_is_young(page) || PageReferenced(page) ||
1213 mmu_notifier_test_young(vma->vm_mm, address))
1218 result = SCAN_SUCCEED;
1221 result = SCAN_LACK_REFERENCED_PAGE;
1224 result = SCAN_PAGE_RO;
1227 pte_unmap_unlock(pte, ptl);
1229 node = khugepaged_find_target_node();
1230 /* collapse_huge_page will return with the mmap_sem released */
1231 collapse_huge_page(mm, address, hpage, node, referenced);
1234 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1235 none_or_zero, result, unmapped);
1239 static void collect_mm_slot(struct mm_slot *mm_slot)
1241 struct mm_struct *mm = mm_slot->mm;
1243 lockdep_assert_held(&khugepaged_mm_lock);
1245 if (khugepaged_test_exit(mm)) {
1247 hash_del(&mm_slot->hash);
1248 list_del(&mm_slot->mm_node);
1251 * Not strictly needed because the mm exited already.
1253 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1256 /* khugepaged_mm_lock actually not necessary for the below */
1257 free_mm_slot(mm_slot);
1262 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1264 * Notify khugepaged that given addr of the mm is pte-mapped THP. Then
1265 * khugepaged should try to collapse the page table.
1267 static int khugepaged_add_pte_mapped_thp(struct mm_struct *mm,
1270 struct mm_slot *mm_slot;
1272 VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
1274 spin_lock(&khugepaged_mm_lock);
1275 mm_slot = get_mm_slot(mm);
1276 if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP))
1277 mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr;
1278 spin_unlock(&khugepaged_mm_lock);
1283 * Try to collapse a pte-mapped THP for mm at address haddr.
1285 * This function checks whether all the PTEs in the PMD are pointing to the
1286 * right THP. If so, retract the page table so the THP can refault in with
1289 void collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr)
1291 unsigned long haddr = addr & HPAGE_PMD_MASK;
1292 struct vm_area_struct *vma = find_vma(mm, haddr);
1293 struct page *hpage = NULL;
1294 pte_t *start_pte, *pte;
1300 if (!vma || !vma->vm_file ||
1301 vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE)
1305 * This vm_flags may not have VM_HUGEPAGE if the page was not
1306 * collapsed by this mm. But we can still collapse if the page is
1307 * the valid THP. Add extra VM_HUGEPAGE so hugepage_vma_check()
1308 * will not fail the vma for missing VM_HUGEPAGE
1310 if (!hugepage_vma_check(vma, vma->vm_flags | VM_HUGEPAGE))
1313 pmd = mm_find_pmd(mm, haddr);
1317 start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl);
1319 /* step 1: check all mapped PTEs are to the right huge page */
1320 for (i = 0, addr = haddr, pte = start_pte;
1321 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1324 /* empty pte, skip */
1328 /* page swapped out, abort */
1329 if (!pte_present(*pte))
1332 page = vm_normal_page(vma, addr, *pte);
1334 if (!page || !PageCompound(page))
1338 hpage = compound_head(page);
1340 * The mapping of the THP should not change.
1342 * Note that uprobe, debugger, or MAP_PRIVATE may
1343 * change the page table, but the new page will
1344 * not pass PageCompound() check.
1346 if (WARN_ON(hpage->mapping != vma->vm_file->f_mapping))
1351 * Confirm the page maps to the correct subpage.
1353 * Note that uprobe, debugger, or MAP_PRIVATE may change
1354 * the page table, but the new page will not pass
1355 * PageCompound() check.
1357 if (WARN_ON(hpage + i != page))
1362 /* step 2: adjust rmap */
1363 for (i = 0, addr = haddr, pte = start_pte;
1364 i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
1369 page = vm_normal_page(vma, addr, *pte);
1370 page_remove_rmap(page, false);
1373 pte_unmap_unlock(start_pte, ptl);
1375 /* step 3: set proper refcount and mm_counters. */
1377 page_ref_sub(hpage, count);
1378 add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count);
1381 /* step 4: collapse pmd */
1382 ptl = pmd_lock(vma->vm_mm, pmd);
1383 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1386 pte_free(mm, pmd_pgtable(_pmd));
1390 pte_unmap_unlock(start_pte, ptl);
1393 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1395 struct mm_struct *mm = mm_slot->mm;
1398 if (likely(mm_slot->nr_pte_mapped_thp == 0))
1401 if (!down_write_trylock(&mm->mmap_sem))
1404 if (unlikely(khugepaged_test_exit(mm)))
1407 for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++)
1408 collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i]);
1411 mm_slot->nr_pte_mapped_thp = 0;
1412 up_write(&mm->mmap_sem);
1416 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1418 struct vm_area_struct *vma;
1422 i_mmap_lock_write(mapping);
1423 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1425 * Check vma->anon_vma to exclude MAP_PRIVATE mappings that
1426 * got written to. These VMAs are likely not worth investing
1427 * down_write(mmap_sem) as PMD-mapping is likely to be split
1430 * Not that vma->anon_vma check is racy: it can be set up after
1431 * the check but before we took mmap_sem by the fault path.
1432 * But page lock would prevent establishing any new ptes of the
1433 * page, so we are safe.
1435 * An alternative would be drop the check, but check that page
1436 * table is clear before calling pmdp_collapse_flush() under
1437 * ptl. It has higher chance to recover THP for the VMA, but
1438 * has higher cost too.
1442 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1443 if (addr & ~HPAGE_PMD_MASK)
1445 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1447 pmd = mm_find_pmd(vma->vm_mm, addr);
1451 * We need exclusive mmap_sem to retract page table.
1453 * We use trylock due to lock inversion: we need to acquire
1454 * mmap_sem while holding page lock. Fault path does it in
1455 * reverse order. Trylock is a way to avoid deadlock.
1457 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1458 spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1459 /* assume page table is clear */
1460 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1462 up_write(&vma->vm_mm->mmap_sem);
1463 mm_dec_nr_ptes(vma->vm_mm);
1464 pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1466 /* Try again later */
1467 khugepaged_add_pte_mapped_thp(vma->vm_mm, addr);
1470 i_mmap_unlock_write(mapping);
1474 * collapse_file - collapse filemap/tmpfs/shmem pages into huge one.
1476 * Basic scheme is simple, details are more complex:
1477 * - allocate and lock a new huge page;
1478 * - scan page cache replacing old pages with the new one
1479 * + swap/gup in pages if necessary;
1481 * + keep old pages around in case rollback is required;
1482 * - if replacing succeeds:
1485 * + unlock huge page;
1486 * - if replacing failed;
1487 * + put all pages back and unfreeze them;
1488 * + restore gaps in the page cache;
1489 * + unlock and free huge page;
1491 static void collapse_file(struct mm_struct *mm,
1492 struct file *file, pgoff_t start,
1493 struct page **hpage, int node)
1495 struct address_space *mapping = file->f_mapping;
1497 struct page *new_page;
1498 struct mem_cgroup *memcg;
1499 pgoff_t index, end = start + HPAGE_PMD_NR;
1500 LIST_HEAD(pagelist);
1501 XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER);
1502 int nr_none = 0, result = SCAN_SUCCEED;
1503 bool is_shmem = shmem_file(file);
1505 VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);
1506 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1508 /* Only allocate from the target node */
1509 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_THISNODE;
1511 new_page = khugepaged_alloc_page(hpage, gfp, node);
1513 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1517 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1518 result = SCAN_CGROUP_CHARGE_FAIL;
1522 /* This will be less messy when we use multi-index entries */
1525 xas_create_range(&xas);
1526 if (!xas_error(&xas))
1528 xas_unlock_irq(&xas);
1529 if (!xas_nomem(&xas, GFP_KERNEL)) {
1530 mem_cgroup_cancel_charge(new_page, memcg, true);
1536 __SetPageLocked(new_page);
1538 __SetPageSwapBacked(new_page);
1539 new_page->index = start;
1540 new_page->mapping = mapping;
1543 * At this point the new_page is locked and not up-to-date.
1544 * It's safe to insert it into the page cache, because nobody would
1545 * be able to map it or use it in another way until we unlock it.
1548 xas_set(&xas, start);
1549 for (index = start; index < end; index++) {
1550 struct page *page = xas_next(&xas);
1552 VM_BUG_ON(index != xas.xa_index);
1556 * Stop if extent has been truncated or
1557 * hole-punched, and is now completely
1560 if (index == start) {
1561 if (!xas_next_entry(&xas, end - 1)) {
1562 result = SCAN_TRUNCATED;
1565 xas_set(&xas, index);
1567 if (!shmem_charge(mapping->host, 1)) {
1571 xas_store(&xas, new_page);
1576 if (xa_is_value(page) || !PageUptodate(page)) {
1577 xas_unlock_irq(&xas);
1578 /* swap in or instantiate fallocated page */
1579 if (shmem_getpage(mapping->host, index, &page,
1584 } else if (trylock_page(page)) {
1586 xas_unlock_irq(&xas);
1588 result = SCAN_PAGE_LOCK;
1591 } else { /* !is_shmem */
1592 if (!page || xa_is_value(page)) {
1593 xas_unlock_irq(&xas);
1594 page_cache_sync_readahead(mapping, &file->f_ra,
1597 /* drain pagevecs to help isolate_lru_page() */
1599 page = find_lock_page(mapping, index);
1600 if (unlikely(page == NULL)) {
1604 } else if (!PageUptodate(page)) {
1605 xas_unlock_irq(&xas);
1606 wait_on_page_locked(page);
1607 if (!trylock_page(page)) {
1608 result = SCAN_PAGE_LOCK;
1612 } else if (PageDirty(page)) {
1615 } else if (trylock_page(page)) {
1617 xas_unlock_irq(&xas);
1619 result = SCAN_PAGE_LOCK;
1625 * The page must be locked, so we can drop the i_pages lock
1626 * without racing with truncate.
1628 VM_BUG_ON_PAGE(!PageLocked(page), page);
1629 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1632 * If file was truncated then extended, or hole-punched, before
1633 * we locked the first page, then a THP might be there already.
1635 if (PageTransCompound(page)) {
1636 result = SCAN_PAGE_COMPOUND;
1640 if (page_mapping(page) != mapping) {
1641 result = SCAN_TRUNCATED;
1645 if (isolate_lru_page(page)) {
1646 result = SCAN_DEL_PAGE_LRU;
1650 if (page_has_private(page) &&
1651 !try_to_release_page(page, GFP_KERNEL)) {
1652 result = SCAN_PAGE_HAS_PRIVATE;
1656 if (page_mapped(page))
1657 unmap_mapping_pages(mapping, index, 1, false);
1660 xas_set(&xas, index);
1662 VM_BUG_ON_PAGE(page != xas_load(&xas), page);
1663 VM_BUG_ON_PAGE(page_mapped(page), page);
1666 * The page is expected to have page_count() == 3:
1667 * - we hold a pin on it;
1668 * - one reference from page cache;
1669 * - one from isolate_lru_page;
1671 if (!page_ref_freeze(page, 3)) {
1672 result = SCAN_PAGE_COUNT;
1673 xas_unlock_irq(&xas);
1674 putback_lru_page(page);
1679 * Add the page to the list to be able to undo the collapse if
1680 * something go wrong.
1682 list_add_tail(&page->lru, &pagelist);
1684 /* Finally, replace with the new page. */
1685 xas_store(&xas, new_page);
1694 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1696 __inc_node_page_state(new_page, NR_FILE_THPS);
1697 filemap_nr_thps_inc(mapping);
1701 struct zone *zone = page_zone(new_page);
1703 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1705 __mod_node_page_state(zone->zone_pgdat,
1710 xas_unlock_irq(&xas);
1713 if (result == SCAN_SUCCEED) {
1714 struct page *page, *tmp;
1717 * Replacing old pages with new one has succeeded, now we
1718 * need to copy the content and free the old pages.
1721 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1722 while (index < page->index) {
1723 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1726 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1728 list_del(&page->lru);
1729 page->mapping = NULL;
1730 page_ref_unfreeze(page, 1);
1731 ClearPageActive(page);
1732 ClearPageUnevictable(page);
1737 while (index < end) {
1738 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1742 SetPageUptodate(new_page);
1743 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1744 mem_cgroup_commit_charge(new_page, memcg, false, true);
1747 set_page_dirty(new_page);
1748 lru_cache_add_anon(new_page);
1750 lru_cache_add_file(new_page);
1752 count_memcg_events(memcg, THP_COLLAPSE_ALLOC, 1);
1755 * Remove pte page tables, so we can re-fault the page as huge.
1757 retract_page_tables(mapping, start);
1760 khugepaged_pages_collapsed++;
1764 /* Something went wrong: roll back page cache changes */
1766 mapping->nrpages -= nr_none;
1769 shmem_uncharge(mapping->host, nr_none);
1771 xas_set(&xas, start);
1772 xas_for_each(&xas, page, end - 1) {
1773 page = list_first_entry_or_null(&pagelist,
1775 if (!page || xas.xa_index < page->index) {
1779 /* Put holes back where they were */
1780 xas_store(&xas, NULL);
1784 VM_BUG_ON_PAGE(page->index != xas.xa_index, page);
1786 /* Unfreeze the page. */
1787 list_del(&page->lru);
1788 page_ref_unfreeze(page, 2);
1789 xas_store(&xas, page);
1791 xas_unlock_irq(&xas);
1793 putback_lru_page(page);
1797 xas_unlock_irq(&xas);
1799 mem_cgroup_cancel_charge(new_page, memcg, true);
1800 new_page->mapping = NULL;
1803 unlock_page(new_page);
1805 VM_BUG_ON(!list_empty(&pagelist));
1806 /* TODO: tracepoints */
1809 static void khugepaged_scan_file(struct mm_struct *mm,
1810 struct file *file, pgoff_t start, struct page **hpage)
1812 struct page *page = NULL;
1813 struct address_space *mapping = file->f_mapping;
1814 XA_STATE(xas, &mapping->i_pages, start);
1816 int node = NUMA_NO_NODE;
1817 int result = SCAN_SUCCEED;
1821 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1823 xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) {
1824 if (xas_retry(&xas, page))
1827 if (xa_is_value(page)) {
1828 if (++swap > khugepaged_max_ptes_swap) {
1829 result = SCAN_EXCEED_SWAP_PTE;
1835 if (PageTransCompound(page)) {
1836 result = SCAN_PAGE_COMPOUND;
1840 node = page_to_nid(page);
1841 if (khugepaged_scan_abort(node)) {
1842 result = SCAN_SCAN_ABORT;
1845 khugepaged_node_load[node]++;
1847 if (!PageLRU(page)) {
1848 result = SCAN_PAGE_LRU;
1852 if (page_count(page) !=
1853 1 + page_mapcount(page) + page_has_private(page)) {
1854 result = SCAN_PAGE_COUNT;
1859 * We probably should check if the page is referenced here, but
1860 * nobody would transfer pte_young() to PageReferenced() for us.
1861 * And rmap walk here is just too costly...
1866 if (need_resched()) {
1873 if (result == SCAN_SUCCEED) {
1874 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1875 result = SCAN_EXCEED_NONE_PTE;
1877 node = khugepaged_find_target_node();
1878 collapse_file(mm, file, start, hpage, node);
1882 /* TODO: tracepoints */
1885 static void khugepaged_scan_file(struct mm_struct *mm,
1886 struct file *file, pgoff_t start, struct page **hpage)
1891 static int khugepaged_collapse_pte_mapped_thps(struct mm_slot *mm_slot)
1897 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1898 struct page **hpage)
1899 __releases(&khugepaged_mm_lock)
1900 __acquires(&khugepaged_mm_lock)
1902 struct mm_slot *mm_slot;
1903 struct mm_struct *mm;
1904 struct vm_area_struct *vma;
1908 lockdep_assert_held(&khugepaged_mm_lock);
1910 if (khugepaged_scan.mm_slot)
1911 mm_slot = khugepaged_scan.mm_slot;
1913 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1914 struct mm_slot, mm_node);
1915 khugepaged_scan.address = 0;
1916 khugepaged_scan.mm_slot = mm_slot;
1918 spin_unlock(&khugepaged_mm_lock);
1919 khugepaged_collapse_pte_mapped_thps(mm_slot);
1923 * Don't wait for semaphore (to avoid long wait times). Just move to
1924 * the next mm on the list.
1927 if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1928 goto breakouterloop_mmap_sem;
1929 if (likely(!khugepaged_test_exit(mm)))
1930 vma = find_vma(mm, khugepaged_scan.address);
1933 for (; vma; vma = vma->vm_next) {
1934 unsigned long hstart, hend;
1937 if (unlikely(khugepaged_test_exit(mm))) {
1941 if (!hugepage_vma_check(vma, vma->vm_flags)) {
1946 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1947 hend = vma->vm_end & HPAGE_PMD_MASK;
1950 if (khugepaged_scan.address > hend)
1952 if (khugepaged_scan.address < hstart)
1953 khugepaged_scan.address = hstart;
1954 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1956 while (khugepaged_scan.address < hend) {
1959 if (unlikely(khugepaged_test_exit(mm)))
1960 goto breakouterloop;
1962 VM_BUG_ON(khugepaged_scan.address < hstart ||
1963 khugepaged_scan.address + HPAGE_PMD_SIZE >
1965 if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {
1967 pgoff_t pgoff = linear_page_index(vma,
1968 khugepaged_scan.address);
1970 if (shmem_file(vma->vm_file)
1971 && !shmem_huge_enabled(vma))
1973 file = get_file(vma->vm_file);
1974 up_read(&mm->mmap_sem);
1976 khugepaged_scan_file(mm, file, pgoff, hpage);
1979 ret = khugepaged_scan_pmd(mm, vma,
1980 khugepaged_scan.address,
1983 /* move to next address */
1984 khugepaged_scan.address += HPAGE_PMD_SIZE;
1985 progress += HPAGE_PMD_NR;
1987 /* we released mmap_sem so break loop */
1988 goto breakouterloop_mmap_sem;
1989 if (progress >= pages)
1990 goto breakouterloop;
1994 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1995 breakouterloop_mmap_sem:
1997 spin_lock(&khugepaged_mm_lock);
1998 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2000 * Release the current mm_slot if this mm is about to die, or
2001 * if we scanned all vmas of this mm.
2003 if (khugepaged_test_exit(mm) || !vma) {
2005 * Make sure that if mm_users is reaching zero while
2006 * khugepaged runs here, khugepaged_exit will find
2007 * mm_slot not pointing to the exiting mm.
2009 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2010 khugepaged_scan.mm_slot = list_entry(
2011 mm_slot->mm_node.next,
2012 struct mm_slot, mm_node);
2013 khugepaged_scan.address = 0;
2015 khugepaged_scan.mm_slot = NULL;
2016 khugepaged_full_scans++;
2019 collect_mm_slot(mm_slot);
2025 static int khugepaged_has_work(void)
2027 return !list_empty(&khugepaged_scan.mm_head) &&
2028 khugepaged_enabled();
2031 static int khugepaged_wait_event(void)
2033 return !list_empty(&khugepaged_scan.mm_head) ||
2034 kthread_should_stop();
2037 static void khugepaged_do_scan(void)
2039 struct page *hpage = NULL;
2040 unsigned int progress = 0, pass_through_head = 0;
2041 unsigned int pages = khugepaged_pages_to_scan;
2044 barrier(); /* write khugepaged_pages_to_scan to local stack */
2046 while (progress < pages) {
2047 if (!khugepaged_prealloc_page(&hpage, &wait))
2052 if (unlikely(kthread_should_stop() || try_to_freeze()))
2055 spin_lock(&khugepaged_mm_lock);
2056 if (!khugepaged_scan.mm_slot)
2057 pass_through_head++;
2058 if (khugepaged_has_work() &&
2059 pass_through_head < 2)
2060 progress += khugepaged_scan_mm_slot(pages - progress,
2064 spin_unlock(&khugepaged_mm_lock);
2067 if (!IS_ERR_OR_NULL(hpage))
2071 static bool khugepaged_should_wakeup(void)
2073 return kthread_should_stop() ||
2074 time_after_eq(jiffies, khugepaged_sleep_expire);
2077 static void khugepaged_wait_work(void)
2079 if (khugepaged_has_work()) {
2080 const unsigned long scan_sleep_jiffies =
2081 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
2083 if (!scan_sleep_jiffies)
2086 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
2087 wait_event_freezable_timeout(khugepaged_wait,
2088 khugepaged_should_wakeup(),
2089 scan_sleep_jiffies);
2093 if (khugepaged_enabled())
2094 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2097 static int khugepaged(void *none)
2099 struct mm_slot *mm_slot;
2102 set_user_nice(current, MAX_NICE);
2104 while (!kthread_should_stop()) {
2105 khugepaged_do_scan();
2106 khugepaged_wait_work();
2109 spin_lock(&khugepaged_mm_lock);
2110 mm_slot = khugepaged_scan.mm_slot;
2111 khugepaged_scan.mm_slot = NULL;
2113 collect_mm_slot(mm_slot);
2114 spin_unlock(&khugepaged_mm_lock);
2118 static void set_recommended_min_free_kbytes(void)
2122 unsigned long recommended_min;
2124 for_each_populated_zone(zone) {
2126 * We don't need to worry about fragmentation of
2127 * ZONE_MOVABLE since it only has movable pages.
2129 if (zone_idx(zone) > gfp_zone(GFP_USER))
2135 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
2136 recommended_min = pageblock_nr_pages * nr_zones * 2;
2139 * Make sure that on average at least two pageblocks are almost free
2140 * of another type, one for a migratetype to fall back to and a
2141 * second to avoid subsequent fallbacks of other types There are 3
2142 * MIGRATE_TYPES we care about.
2144 recommended_min += pageblock_nr_pages * nr_zones *
2145 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
2147 /* don't ever allow to reserve more than 5% of the lowmem */
2148 recommended_min = min(recommended_min,
2149 (unsigned long) nr_free_buffer_pages() / 20);
2150 recommended_min <<= (PAGE_SHIFT-10);
2152 if (recommended_min > min_free_kbytes) {
2153 if (user_min_free_kbytes >= 0)
2154 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
2155 min_free_kbytes, recommended_min);
2157 min_free_kbytes = recommended_min;
2159 setup_per_zone_wmarks();
2162 int start_stop_khugepaged(void)
2164 static struct task_struct *khugepaged_thread __read_mostly;
2165 static DEFINE_MUTEX(khugepaged_mutex);
2168 mutex_lock(&khugepaged_mutex);
2169 if (khugepaged_enabled()) {
2170 if (!khugepaged_thread)
2171 khugepaged_thread = kthread_run(khugepaged, NULL,
2173 if (IS_ERR(khugepaged_thread)) {
2174 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
2175 err = PTR_ERR(khugepaged_thread);
2176 khugepaged_thread = NULL;
2180 if (!list_empty(&khugepaged_scan.mm_head))
2181 wake_up_interruptible(&khugepaged_wait);
2183 set_recommended_min_free_kbytes();
2184 } else if (khugepaged_thread) {
2185 kthread_stop(khugepaged_thread);
2186 khugepaged_thread = NULL;
2189 mutex_unlock(&khugepaged_mutex);