1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
4 #include <linux/sched.h>
5 #include <linux/mmu_notifier.h>
6 #include <linux/rmap.h>
7 #include <linux/swap.h>
8 #include <linux/mm_inline.h>
9 #include <linux/kthread.h>
10 #include <linux/khugepaged.h>
11 #include <linux/freezer.h>
12 #include <linux/mman.h>
13 #include <linux/hashtable.h>
14 #include <linux/userfaultfd_k.h>
15 #include <linux/page_idle.h>
16 #include <linux/swapops.h>
17 #include <linux/shmem_fs.h>
20 #include <asm/pgalloc.h>
30 SCAN_LACK_REFERENCED_PAGE,
44 SCAN_ALLOC_HUGE_PAGE_FAIL,
45 SCAN_CGROUP_CHARGE_FAIL,
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/huge_memory.h>
53 /* default scan 8*512 pte (or vmas) every 30 second */
54 static unsigned int khugepaged_pages_to_scan __read_mostly;
55 static unsigned int khugepaged_pages_collapsed;
56 static unsigned int khugepaged_full_scans;
57 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
58 /* during fragmentation poll the hugepage allocator once every minute */
59 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
60 static unsigned long khugepaged_sleep_expire;
61 static DEFINE_SPINLOCK(khugepaged_mm_lock);
62 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
64 * default collapse hugepages if there is at least one pte mapped like
65 * it would have happened if the vma was large enough during page
68 static unsigned int khugepaged_max_ptes_none __read_mostly;
69 static unsigned int khugepaged_max_ptes_swap __read_mostly;
71 #define MM_SLOTS_HASH_BITS 10
72 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
74 static struct kmem_cache *mm_slot_cache __read_mostly;
77 * struct mm_slot - hash lookup from mm to mm_slot
78 * @hash: hash collision list
79 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
80 * @mm: the mm that this information is valid for
83 struct hlist_node hash;
84 struct list_head mm_node;
89 * struct khugepaged_scan - cursor for scanning
90 * @mm_head: the head of the mm list to scan
91 * @mm_slot: the current mm_slot we are scanning
92 * @address: the next address inside that to be scanned
94 * There is only the one khugepaged_scan instance of this cursor structure.
96 struct khugepaged_scan {
97 struct list_head mm_head;
98 struct mm_slot *mm_slot;
99 unsigned long address;
102 static struct khugepaged_scan khugepaged_scan = {
103 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
107 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
108 struct kobj_attribute *attr,
111 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
114 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
115 struct kobj_attribute *attr,
116 const char *buf, size_t count)
121 err = kstrtoul(buf, 10, &msecs);
122 if (err || msecs > UINT_MAX)
125 khugepaged_scan_sleep_millisecs = msecs;
126 khugepaged_sleep_expire = 0;
127 wake_up_interruptible(&khugepaged_wait);
131 static struct kobj_attribute scan_sleep_millisecs_attr =
132 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
133 scan_sleep_millisecs_store);
135 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
136 struct kobj_attribute *attr,
139 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
142 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
143 struct kobj_attribute *attr,
144 const char *buf, size_t count)
149 err = kstrtoul(buf, 10, &msecs);
150 if (err || msecs > UINT_MAX)
153 khugepaged_alloc_sleep_millisecs = msecs;
154 khugepaged_sleep_expire = 0;
155 wake_up_interruptible(&khugepaged_wait);
159 static struct kobj_attribute alloc_sleep_millisecs_attr =
160 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
161 alloc_sleep_millisecs_store);
163 static ssize_t pages_to_scan_show(struct kobject *kobj,
164 struct kobj_attribute *attr,
167 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
169 static ssize_t pages_to_scan_store(struct kobject *kobj,
170 struct kobj_attribute *attr,
171 const char *buf, size_t count)
176 err = kstrtoul(buf, 10, &pages);
177 if (err || !pages || pages > UINT_MAX)
180 khugepaged_pages_to_scan = pages;
184 static struct kobj_attribute pages_to_scan_attr =
185 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
186 pages_to_scan_store);
188 static ssize_t pages_collapsed_show(struct kobject *kobj,
189 struct kobj_attribute *attr,
192 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
194 static struct kobj_attribute pages_collapsed_attr =
195 __ATTR_RO(pages_collapsed);
197 static ssize_t full_scans_show(struct kobject *kobj,
198 struct kobj_attribute *attr,
201 return sprintf(buf, "%u\n", khugepaged_full_scans);
203 static struct kobj_attribute full_scans_attr =
204 __ATTR_RO(full_scans);
206 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
207 struct kobj_attribute *attr, char *buf)
209 return single_hugepage_flag_show(kobj, attr, buf,
210 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
212 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
213 struct kobj_attribute *attr,
214 const char *buf, size_t count)
216 return single_hugepage_flag_store(kobj, attr, buf, count,
217 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
219 static struct kobj_attribute khugepaged_defrag_attr =
220 __ATTR(defrag, 0644, khugepaged_defrag_show,
221 khugepaged_defrag_store);
224 * max_ptes_none controls if khugepaged should collapse hugepages over
225 * any unmapped ptes in turn potentially increasing the memory
226 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
227 * reduce the available free memory in the system as it
228 * runs. Increasing max_ptes_none will instead potentially reduce the
229 * free memory in the system during the khugepaged scan.
231 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
232 struct kobj_attribute *attr,
235 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
237 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
238 struct kobj_attribute *attr,
239 const char *buf, size_t count)
242 unsigned long max_ptes_none;
244 err = kstrtoul(buf, 10, &max_ptes_none);
245 if (err || max_ptes_none > HPAGE_PMD_NR-1)
248 khugepaged_max_ptes_none = max_ptes_none;
252 static struct kobj_attribute khugepaged_max_ptes_none_attr =
253 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
254 khugepaged_max_ptes_none_store);
256 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
257 struct kobj_attribute *attr,
260 return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
263 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
264 struct kobj_attribute *attr,
265 const char *buf, size_t count)
268 unsigned long max_ptes_swap;
270 err = kstrtoul(buf, 10, &max_ptes_swap);
271 if (err || max_ptes_swap > HPAGE_PMD_NR-1)
274 khugepaged_max_ptes_swap = max_ptes_swap;
279 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
280 __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
281 khugepaged_max_ptes_swap_store);
283 static struct attribute *khugepaged_attr[] = {
284 &khugepaged_defrag_attr.attr,
285 &khugepaged_max_ptes_none_attr.attr,
286 &pages_to_scan_attr.attr,
287 &pages_collapsed_attr.attr,
288 &full_scans_attr.attr,
289 &scan_sleep_millisecs_attr.attr,
290 &alloc_sleep_millisecs_attr.attr,
291 &khugepaged_max_ptes_swap_attr.attr,
295 struct attribute_group khugepaged_attr_group = {
296 .attrs = khugepaged_attr,
297 .name = "khugepaged",
299 #endif /* CONFIG_SYSFS */
301 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
303 int hugepage_madvise(struct vm_area_struct *vma,
304 unsigned long *vm_flags, int advice)
310 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
311 * can't handle this properly after s390_enable_sie, so we simply
312 * ignore the madvise to prevent qemu from causing a SIGSEGV.
314 if (mm_has_pgste(vma->vm_mm))
317 *vm_flags &= ~VM_NOHUGEPAGE;
318 *vm_flags |= VM_HUGEPAGE;
320 * If the vma become good for khugepaged to scan,
321 * register it here without waiting a page fault that
322 * may not happen any time soon.
324 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
325 khugepaged_enter_vma_merge(vma, *vm_flags))
328 case MADV_NOHUGEPAGE:
329 *vm_flags &= ~VM_HUGEPAGE;
330 *vm_flags |= VM_NOHUGEPAGE;
332 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
333 * this vma even if we leave the mm registered in khugepaged if
334 * it got registered before VM_NOHUGEPAGE was set.
342 int __init khugepaged_init(void)
344 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
345 sizeof(struct mm_slot),
346 __alignof__(struct mm_slot), 0, NULL);
350 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
351 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
352 khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
357 void __init khugepaged_destroy(void)
359 kmem_cache_destroy(mm_slot_cache);
362 static inline struct mm_slot *alloc_mm_slot(void)
364 if (!mm_slot_cache) /* initialization failed */
366 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
369 static inline void free_mm_slot(struct mm_slot *mm_slot)
371 kmem_cache_free(mm_slot_cache, mm_slot);
374 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
376 struct mm_slot *mm_slot;
378 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
379 if (mm == mm_slot->mm)
385 static void insert_to_mm_slots_hash(struct mm_struct *mm,
386 struct mm_slot *mm_slot)
389 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
392 static inline int khugepaged_test_exit(struct mm_struct *mm)
394 return atomic_read(&mm->mm_users) == 0;
397 int __khugepaged_enter(struct mm_struct *mm)
399 struct mm_slot *mm_slot;
402 mm_slot = alloc_mm_slot();
406 /* __khugepaged_exit() must not run from under us */
407 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
408 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
409 free_mm_slot(mm_slot);
413 spin_lock(&khugepaged_mm_lock);
414 insert_to_mm_slots_hash(mm, mm_slot);
416 * Insert just behind the scanning cursor, to let the area settle
419 wakeup = list_empty(&khugepaged_scan.mm_head);
420 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
421 spin_unlock(&khugepaged_mm_lock);
423 atomic_inc(&mm->mm_count);
425 wake_up_interruptible(&khugepaged_wait);
430 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
431 unsigned long vm_flags)
433 unsigned long hstart, hend;
436 * Not yet faulted in so we will register later in the
437 * page fault if needed.
440 if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
441 /* khugepaged not yet working on file or special mappings */
443 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
444 hend = vma->vm_end & HPAGE_PMD_MASK;
446 return khugepaged_enter(vma, vm_flags);
450 void __khugepaged_exit(struct mm_struct *mm)
452 struct mm_slot *mm_slot;
455 spin_lock(&khugepaged_mm_lock);
456 mm_slot = get_mm_slot(mm);
457 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
458 hash_del(&mm_slot->hash);
459 list_del(&mm_slot->mm_node);
462 spin_unlock(&khugepaged_mm_lock);
465 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
466 free_mm_slot(mm_slot);
468 } else if (mm_slot) {
470 * This is required to serialize against
471 * khugepaged_test_exit() (which is guaranteed to run
472 * under mmap sem read mode). Stop here (after we
473 * return all pagetables will be destroyed) until
474 * khugepaged has finished working on the pagetables
475 * under the mmap_sem.
477 down_write(&mm->mmap_sem);
478 up_write(&mm->mmap_sem);
482 static void release_pte_page(struct page *page)
484 /* 0 stands for page_is_file_cache(page) == false */
485 dec_node_page_state(page, NR_ISOLATED_ANON + 0);
487 putback_lru_page(page);
490 static void release_pte_pages(pte_t *pte, pte_t *_pte)
492 while (--_pte >= pte) {
493 pte_t pteval = *_pte;
494 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
495 release_pte_page(pte_page(pteval));
499 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
500 unsigned long address,
503 struct page *page = NULL;
505 int none_or_zero = 0, result = 0, referenced = 0;
506 bool writable = false;
508 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
509 _pte++, address += PAGE_SIZE) {
510 pte_t pteval = *_pte;
511 if (pte_none(pteval) || (pte_present(pteval) &&
512 is_zero_pfn(pte_pfn(pteval)))) {
513 if (!userfaultfd_armed(vma) &&
514 ++none_or_zero <= khugepaged_max_ptes_none) {
517 result = SCAN_EXCEED_NONE_PTE;
521 if (!pte_present(pteval)) {
522 result = SCAN_PTE_NON_PRESENT;
525 page = vm_normal_page(vma, address, pteval);
526 if (unlikely(!page)) {
527 result = SCAN_PAGE_NULL;
531 VM_BUG_ON_PAGE(PageCompound(page), page);
532 VM_BUG_ON_PAGE(!PageAnon(page), page);
533 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
536 * We can do it before isolate_lru_page because the
537 * page can't be freed from under us. NOTE: PG_lock
538 * is needed to serialize against split_huge_page
539 * when invoked from the VM.
541 if (!trylock_page(page)) {
542 result = SCAN_PAGE_LOCK;
547 * cannot use mapcount: can't collapse if there's a gup pin.
548 * The page must only be referenced by the scanned process
549 * and page swap cache.
551 if (page_count(page) != 1 + !!PageSwapCache(page)) {
553 result = SCAN_PAGE_COUNT;
556 if (pte_write(pteval)) {
559 if (PageSwapCache(page) &&
560 !reuse_swap_page(page, NULL)) {
562 result = SCAN_SWAP_CACHE_PAGE;
566 * Page is not in the swap cache. It can be collapsed
572 * Isolate the page to avoid collapsing an hugepage
573 * currently in use by the VM.
575 if (isolate_lru_page(page)) {
577 result = SCAN_DEL_PAGE_LRU;
580 /* 0 stands for page_is_file_cache(page) == false */
581 inc_node_page_state(page, NR_ISOLATED_ANON + 0);
582 VM_BUG_ON_PAGE(!PageLocked(page), page);
583 VM_BUG_ON_PAGE(PageLRU(page), page);
585 /* There should be enough young pte to collapse the page */
586 if (pte_young(pteval) ||
587 page_is_young(page) || PageReferenced(page) ||
588 mmu_notifier_test_young(vma->vm_mm, address))
591 if (likely(writable)) {
592 if (likely(referenced)) {
593 result = SCAN_SUCCEED;
594 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
595 referenced, writable, result);
599 result = SCAN_PAGE_RO;
603 release_pte_pages(pte, _pte);
604 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
605 referenced, writable, result);
609 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
610 struct vm_area_struct *vma,
611 unsigned long address,
615 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
616 pte_t pteval = *_pte;
617 struct page *src_page;
619 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
620 clear_user_highpage(page, address);
621 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
622 if (is_zero_pfn(pte_pfn(pteval))) {
624 * ptl mostly unnecessary.
628 * paravirt calls inside pte_clear here are
631 pte_clear(vma->vm_mm, address, _pte);
635 src_page = pte_page(pteval);
636 copy_user_highpage(page, src_page, address, vma);
637 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
638 release_pte_page(src_page);
640 * ptl mostly unnecessary, but preempt has to
641 * be disabled to update the per-cpu stats
642 * inside page_remove_rmap().
646 * paravirt calls inside pte_clear here are
649 pte_clear(vma->vm_mm, address, _pte);
650 page_remove_rmap(src_page, false);
652 free_page_and_swap_cache(src_page);
655 address += PAGE_SIZE;
660 static void khugepaged_alloc_sleep(void)
664 add_wait_queue(&khugepaged_wait, &wait);
665 freezable_schedule_timeout_interruptible(
666 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
667 remove_wait_queue(&khugepaged_wait, &wait);
670 static int khugepaged_node_load[MAX_NUMNODES];
672 static bool khugepaged_scan_abort(int nid)
677 * If node_reclaim_mode is disabled, then no extra effort is made to
678 * allocate memory locally.
680 if (!node_reclaim_mode)
683 /* If there is a count for this node already, it must be acceptable */
684 if (khugepaged_node_load[nid])
687 for (i = 0; i < MAX_NUMNODES; i++) {
688 if (!khugepaged_node_load[i])
690 if (node_distance(nid, i) > RECLAIM_DISTANCE)
696 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
697 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
699 return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
703 static int khugepaged_find_target_node(void)
705 static int last_khugepaged_target_node = NUMA_NO_NODE;
706 int nid, target_node = 0, max_value = 0;
708 /* find first node with max normal pages hit */
709 for (nid = 0; nid < MAX_NUMNODES; nid++)
710 if (khugepaged_node_load[nid] > max_value) {
711 max_value = khugepaged_node_load[nid];
715 /* do some balance if several nodes have the same hit record */
716 if (target_node <= last_khugepaged_target_node)
717 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
719 if (max_value == khugepaged_node_load[nid]) {
724 last_khugepaged_target_node = target_node;
728 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
730 if (IS_ERR(*hpage)) {
736 khugepaged_alloc_sleep();
746 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
748 VM_BUG_ON_PAGE(*hpage, *hpage);
750 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
751 if (unlikely(!*hpage)) {
752 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
753 *hpage = ERR_PTR(-ENOMEM);
757 prep_transhuge_page(*hpage);
758 count_vm_event(THP_COLLAPSE_ALLOC);
762 static int khugepaged_find_target_node(void)
767 static inline struct page *alloc_khugepaged_hugepage(void)
771 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
774 prep_transhuge_page(page);
778 static struct page *khugepaged_alloc_hugepage(bool *wait)
783 hpage = alloc_khugepaged_hugepage();
785 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
790 khugepaged_alloc_sleep();
792 count_vm_event(THP_COLLAPSE_ALLOC);
793 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
798 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
801 *hpage = khugepaged_alloc_hugepage(wait);
803 if (unlikely(!*hpage))
810 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
818 static bool hugepage_vma_check(struct vm_area_struct *vma)
820 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
821 (vma->vm_flags & VM_NOHUGEPAGE))
823 if (shmem_file(vma->vm_file)) {
824 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
826 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
829 if (!vma->anon_vma || vma->vm_ops)
831 if (is_vma_temporary_stack(vma))
833 return !(vma->vm_flags & VM_NO_KHUGEPAGED);
837 * If mmap_sem temporarily dropped, revalidate vma
838 * before taking mmap_sem.
839 * Return 0 if succeeds, otherwise return none-zero
843 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
844 struct vm_area_struct **vmap)
846 struct vm_area_struct *vma;
847 unsigned long hstart, hend;
849 if (unlikely(khugepaged_test_exit(mm)))
850 return SCAN_ANY_PROCESS;
852 *vmap = vma = find_vma(mm, address);
854 return SCAN_VMA_NULL;
856 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
857 hend = vma->vm_end & HPAGE_PMD_MASK;
858 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
859 return SCAN_ADDRESS_RANGE;
860 if (!hugepage_vma_check(vma))
861 return SCAN_VMA_CHECK;
866 * Bring missing pages in from swap, to complete THP collapse.
867 * Only done if khugepaged_scan_pmd believes it is worthwhile.
869 * Called and returns without pte mapped or spinlocks held,
870 * but with mmap_sem held to protect against vma changes.
873 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
874 struct vm_area_struct *vma,
875 unsigned long address, pmd_t *pmd,
879 int swapped_in = 0, ret = 0;
880 struct fault_env fe = {
883 .flags = FAULT_FLAG_ALLOW_RETRY,
887 /* we only decide to swapin, if there is enough young ptes */
888 if (referenced < HPAGE_PMD_NR/2) {
889 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
892 fe.pte = pte_offset_map(pmd, address);
893 for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
894 fe.pte++, fe.address += PAGE_SIZE) {
896 if (!is_swap_pte(pteval))
899 ret = do_swap_page(&fe, pteval);
901 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
902 if (ret & VM_FAULT_RETRY) {
903 down_read(&mm->mmap_sem);
904 if (hugepage_vma_revalidate(mm, address, &fe.vma)) {
905 /* vma is no longer available, don't continue to swapin */
906 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
909 /* check if the pmd is still valid */
910 if (mm_find_pmd(mm, address) != pmd)
913 if (ret & VM_FAULT_ERROR) {
914 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
917 /* pte is unmapped now, we need to map it */
918 fe.pte = pte_offset_map(pmd, fe.address);
922 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
926 static void collapse_huge_page(struct mm_struct *mm,
927 unsigned long address,
929 int node, int referenced)
934 struct page *new_page;
935 spinlock_t *pmd_ptl, *pte_ptl;
936 int isolated = 0, result = 0;
937 struct mem_cgroup *memcg;
938 struct vm_area_struct *vma;
939 unsigned long mmun_start; /* For mmu_notifiers */
940 unsigned long mmun_end; /* For mmu_notifiers */
943 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
945 /* Only allocate from the target node */
946 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
949 * Before allocating the hugepage, release the mmap_sem read lock.
950 * The allocation can take potentially a long time if it involves
951 * sync compaction, and we do not need to hold the mmap_sem during
952 * that. We will recheck the vma after taking it again in write mode.
954 up_read(&mm->mmap_sem);
955 new_page = khugepaged_alloc_page(hpage, gfp, node);
957 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
961 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
962 result = SCAN_CGROUP_CHARGE_FAIL;
966 down_read(&mm->mmap_sem);
967 result = hugepage_vma_revalidate(mm, address, &vma);
969 mem_cgroup_cancel_charge(new_page, memcg, true);
970 up_read(&mm->mmap_sem);
974 pmd = mm_find_pmd(mm, address);
976 result = SCAN_PMD_NULL;
977 mem_cgroup_cancel_charge(new_page, memcg, true);
978 up_read(&mm->mmap_sem);
983 * __collapse_huge_page_swapin always returns with mmap_sem locked.
984 * If it fails, we release mmap_sem and jump out_nolock.
985 * Continuing to collapse causes inconsistency.
987 if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
988 mem_cgroup_cancel_charge(new_page, memcg, true);
989 up_read(&mm->mmap_sem);
993 up_read(&mm->mmap_sem);
995 * Prevent all access to pagetables with the exception of
996 * gup_fast later handled by the ptep_clear_flush and the VM
997 * handled by the anon_vma lock + PG_lock.
999 down_write(&mm->mmap_sem);
1000 result = hugepage_vma_revalidate(mm, address, &vma);
1003 /* check if the pmd is still valid */
1004 if (mm_find_pmd(mm, address) != pmd)
1007 anon_vma_lock_write(vma->anon_vma);
1009 pte = pte_offset_map(pmd, address);
1010 pte_ptl = pte_lockptr(mm, pmd);
1012 mmun_start = address;
1013 mmun_end = address + HPAGE_PMD_SIZE;
1014 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1015 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1017 * After this gup_fast can't run anymore. This also removes
1018 * any huge TLB entry from the CPU so we won't allow
1019 * huge and small TLB entries for the same virtual address
1020 * to avoid the risk of CPU bugs in that area.
1022 _pmd = pmdp_collapse_flush(vma, address, pmd);
1023 spin_unlock(pmd_ptl);
1024 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1027 isolated = __collapse_huge_page_isolate(vma, address, pte);
1028 spin_unlock(pte_ptl);
1030 if (unlikely(!isolated)) {
1033 BUG_ON(!pmd_none(*pmd));
1035 * We can only use set_pmd_at when establishing
1036 * hugepmds and never for establishing regular pmds that
1037 * points to regular pagetables. Use pmd_populate for that
1039 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1040 spin_unlock(pmd_ptl);
1041 anon_vma_unlock_write(vma->anon_vma);
1047 * All pages are isolated and locked so anon_vma rmap
1048 * can't run anymore.
1050 anon_vma_unlock_write(vma->anon_vma);
1052 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1054 __SetPageUptodate(new_page);
1055 pgtable = pmd_pgtable(_pmd);
1057 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1058 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1061 * spin_lock() below is not the equivalent of smp_wmb(), so
1062 * this is needed to avoid the copy_huge_page writes to become
1063 * visible after the set_pmd_at() write.
1068 BUG_ON(!pmd_none(*pmd));
1069 page_add_new_anon_rmap(new_page, vma, address, true);
1070 mem_cgroup_commit_charge(new_page, memcg, false, true);
1071 lru_cache_add_active_or_unevictable(new_page, vma);
1072 pgtable_trans_huge_deposit(mm, pmd, pgtable);
1073 set_pmd_at(mm, address, pmd, _pmd);
1074 update_mmu_cache_pmd(vma, address, pmd);
1075 spin_unlock(pmd_ptl);
1079 khugepaged_pages_collapsed++;
1080 result = SCAN_SUCCEED;
1082 up_write(&mm->mmap_sem);
1084 trace_mm_collapse_huge_page(mm, isolated, result);
1087 mem_cgroup_cancel_charge(new_page, memcg, true);
1091 static int khugepaged_scan_pmd(struct mm_struct *mm,
1092 struct vm_area_struct *vma,
1093 unsigned long address,
1094 struct page **hpage)
1098 int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1099 struct page *page = NULL;
1100 unsigned long _address;
1102 int node = NUMA_NO_NODE, unmapped = 0;
1103 bool writable = false;
1105 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1107 pmd = mm_find_pmd(mm, address);
1109 result = SCAN_PMD_NULL;
1113 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1114 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1115 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1116 _pte++, _address += PAGE_SIZE) {
1117 pte_t pteval = *_pte;
1118 if (is_swap_pte(pteval)) {
1119 if (++unmapped <= khugepaged_max_ptes_swap) {
1122 result = SCAN_EXCEED_SWAP_PTE;
1126 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1127 if (!userfaultfd_armed(vma) &&
1128 ++none_or_zero <= khugepaged_max_ptes_none) {
1131 result = SCAN_EXCEED_NONE_PTE;
1135 if (!pte_present(pteval)) {
1136 result = SCAN_PTE_NON_PRESENT;
1139 if (pte_write(pteval))
1142 page = vm_normal_page(vma, _address, pteval);
1143 if (unlikely(!page)) {
1144 result = SCAN_PAGE_NULL;
1148 /* TODO: teach khugepaged to collapse THP mapped with pte */
1149 if (PageCompound(page)) {
1150 result = SCAN_PAGE_COMPOUND;
1155 * Record which node the original page is from and save this
1156 * information to khugepaged_node_load[].
1157 * Khupaged will allocate hugepage from the node has the max
1160 node = page_to_nid(page);
1161 if (khugepaged_scan_abort(node)) {
1162 result = SCAN_SCAN_ABORT;
1165 khugepaged_node_load[node]++;
1166 if (!PageLRU(page)) {
1167 result = SCAN_PAGE_LRU;
1170 if (PageLocked(page)) {
1171 result = SCAN_PAGE_LOCK;
1174 if (!PageAnon(page)) {
1175 result = SCAN_PAGE_ANON;
1180 * cannot use mapcount: can't collapse if there's a gup pin.
1181 * The page must only be referenced by the scanned process
1182 * and page swap cache.
1184 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1185 result = SCAN_PAGE_COUNT;
1188 if (pte_young(pteval) ||
1189 page_is_young(page) || PageReferenced(page) ||
1190 mmu_notifier_test_young(vma->vm_mm, address))
1195 result = SCAN_SUCCEED;
1198 result = SCAN_LACK_REFERENCED_PAGE;
1201 result = SCAN_PAGE_RO;
1204 pte_unmap_unlock(pte, ptl);
1206 node = khugepaged_find_target_node();
1207 /* collapse_huge_page will return with the mmap_sem released */
1208 collapse_huge_page(mm, address, hpage, node, referenced);
1211 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1212 none_or_zero, result, unmapped);
1216 static void collect_mm_slot(struct mm_slot *mm_slot)
1218 struct mm_struct *mm = mm_slot->mm;
1220 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1222 if (khugepaged_test_exit(mm)) {
1224 hash_del(&mm_slot->hash);
1225 list_del(&mm_slot->mm_node);
1228 * Not strictly needed because the mm exited already.
1230 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1233 /* khugepaged_mm_lock actually not necessary for the below */
1234 free_mm_slot(mm_slot);
1239 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1240 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1242 struct vm_area_struct *vma;
1246 i_mmap_lock_write(mapping);
1247 vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1248 /* probably overkill */
1251 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1252 if (addr & ~HPAGE_PMD_MASK)
1254 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1256 pmd = mm_find_pmd(vma->vm_mm, addr);
1260 * We need exclusive mmap_sem to retract page table.
1261 * If trylock fails we would end up with pte-mapped THP after
1262 * re-fault. Not ideal, but it's more important to not disturb
1263 * the system too much.
1265 if (down_write_trylock(&vma->vm_mm->mmap_sem)) {
1266 spinlock_t *ptl = pmd_lock(vma->vm_mm, pmd);
1267 /* assume page table is clear */
1268 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1270 up_write(&vma->vm_mm->mmap_sem);
1271 atomic_long_dec(&vma->vm_mm->nr_ptes);
1272 pte_free(vma->vm_mm, pmd_pgtable(_pmd));
1275 i_mmap_unlock_write(mapping);
1279 * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1281 * Basic scheme is simple, details are more complex:
1282 * - allocate and freeze a new huge page;
1283 * - scan over radix tree replacing old pages the new one
1284 * + swap in pages if necessary;
1286 * + keep old pages around in case if rollback is required;
1287 * - if replacing succeed:
1290 * + unfreeze huge page;
1291 * - if replacing failed;
1292 * + put all pages back and unfreeze them;
1293 * + restore gaps in the radix-tree;
1296 static void collapse_shmem(struct mm_struct *mm,
1297 struct address_space *mapping, pgoff_t start,
1298 struct page **hpage, int node)
1301 struct page *page, *new_page, *tmp;
1302 struct mem_cgroup *memcg;
1303 pgoff_t index, end = start + HPAGE_PMD_NR;
1304 LIST_HEAD(pagelist);
1305 struct radix_tree_iter iter;
1307 int nr_none = 0, result = SCAN_SUCCEED;
1309 VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1311 /* Only allocate from the target node */
1312 gfp = alloc_hugepage_khugepaged_gfpmask() |
1313 __GFP_OTHER_NODE | __GFP_THISNODE;
1315 new_page = khugepaged_alloc_page(hpage, gfp, node);
1317 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1321 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
1322 result = SCAN_CGROUP_CHARGE_FAIL;
1326 new_page->index = start;
1327 new_page->mapping = mapping;
1328 __SetPageSwapBacked(new_page);
1329 __SetPageLocked(new_page);
1330 BUG_ON(!page_ref_freeze(new_page, 1));
1334 * At this point the new_page is 'frozen' (page_count() is zero), locked
1335 * and not up-to-date. It's safe to insert it into radix tree, because
1336 * nobody would be able to map it or use it in other way until we
1341 spin_lock_irq(&mapping->tree_lock);
1342 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1343 int n = min(iter.index, end) - index;
1346 * Handle holes in the radix tree: charge it from shmem and
1347 * insert relevant subpage of new_page into the radix-tree.
1349 if (n && !shmem_charge(mapping->host, n)) {
1354 for (; index < min(iter.index, end); index++) {
1355 radix_tree_insert(&mapping->page_tree, index,
1356 new_page + (index % HPAGE_PMD_NR));
1363 page = radix_tree_deref_slot_protected(slot,
1364 &mapping->tree_lock);
1365 if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1366 spin_unlock_irq(&mapping->tree_lock);
1367 /* swap in or instantiate fallocated page */
1368 if (shmem_getpage(mapping->host, index, &page,
1373 spin_lock_irq(&mapping->tree_lock);
1374 } else if (trylock_page(page)) {
1377 result = SCAN_PAGE_LOCK;
1382 * The page must be locked, so we can drop the tree_lock
1383 * without racing with truncate.
1385 VM_BUG_ON_PAGE(!PageLocked(page), page);
1386 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1387 VM_BUG_ON_PAGE(PageTransCompound(page), page);
1389 if (page_mapping(page) != mapping) {
1390 result = SCAN_TRUNCATED;
1393 spin_unlock_irq(&mapping->tree_lock);
1395 if (isolate_lru_page(page)) {
1396 result = SCAN_DEL_PAGE_LRU;
1397 goto out_isolate_failed;
1400 if (page_mapped(page))
1401 unmap_mapping_range(mapping, index << PAGE_SHIFT,
1404 spin_lock_irq(&mapping->tree_lock);
1406 VM_BUG_ON_PAGE(page_mapped(page), page);
1409 * The page is expected to have page_count() == 3:
1410 * - we hold a pin on it;
1411 * - one reference from radix tree;
1412 * - one from isolate_lru_page;
1414 if (!page_ref_freeze(page, 3)) {
1415 result = SCAN_PAGE_COUNT;
1420 * Add the page to the list to be able to undo the collapse if
1421 * something go wrong.
1423 list_add_tail(&page->lru, &pagelist);
1425 /* Finally, replace with the new page. */
1426 radix_tree_replace_slot(slot,
1427 new_page + (index % HPAGE_PMD_NR));
1432 spin_unlock_irq(&mapping->tree_lock);
1433 putback_lru_page(page);
1445 * Handle hole in radix tree at the end of the range.
1446 * This code only triggers if there's nothing in radix tree
1449 if (result == SCAN_SUCCEED && index < end) {
1450 int n = end - index;
1452 if (!shmem_charge(mapping->host, n)) {
1457 for (; index < end; index++) {
1458 radix_tree_insert(&mapping->page_tree, index,
1459 new_page + (index % HPAGE_PMD_NR));
1465 spin_unlock_irq(&mapping->tree_lock);
1468 if (result == SCAN_SUCCEED) {
1469 unsigned long flags;
1470 struct zone *zone = page_zone(new_page);
1473 * Replacing old pages with new one has succeed, now we need to
1474 * copy the content and free old pages.
1476 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1477 copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1479 list_del(&page->lru);
1481 page_ref_unfreeze(page, 1);
1482 page->mapping = NULL;
1483 ClearPageActive(page);
1484 ClearPageUnevictable(page);
1488 local_irq_save(flags);
1489 __inc_node_page_state(new_page, NR_SHMEM_THPS);
1491 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1492 __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1494 local_irq_restore(flags);
1497 * Remove pte page tables, so we can re-faulti
1500 retract_page_tables(mapping, start);
1502 /* Everything is ready, let's unfreeze the new_page */
1503 set_page_dirty(new_page);
1504 SetPageUptodate(new_page);
1505 page_ref_unfreeze(new_page, HPAGE_PMD_NR);
1506 mem_cgroup_commit_charge(new_page, memcg, false, true);
1507 lru_cache_add_anon(new_page);
1508 unlock_page(new_page);
1512 /* Something went wrong: rollback changes to the radix-tree */
1513 shmem_uncharge(mapping->host, nr_none);
1514 spin_lock_irq(&mapping->tree_lock);
1515 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
1517 if (iter.index >= end)
1519 page = list_first_entry_or_null(&pagelist,
1521 if (!page || iter.index < page->index) {
1524 /* Put holes back where they were */
1525 radix_tree_replace_slot(slot, NULL);
1530 VM_BUG_ON_PAGE(page->index != iter.index, page);
1532 /* Unfreeze the page. */
1533 list_del(&page->lru);
1534 page_ref_unfreeze(page, 2);
1535 radix_tree_replace_slot(slot, page);
1536 spin_unlock_irq(&mapping->tree_lock);
1537 putback_lru_page(page);
1539 spin_lock_irq(&mapping->tree_lock);
1542 spin_unlock_irq(&mapping->tree_lock);
1544 /* Unfreeze new_page, caller would take care about freeing it */
1545 page_ref_unfreeze(new_page, 1);
1546 mem_cgroup_cancel_charge(new_page, memcg, true);
1547 unlock_page(new_page);
1548 new_page->mapping = NULL;
1551 VM_BUG_ON(!list_empty(&pagelist));
1552 /* TODO: tracepoints */
1555 static void khugepaged_scan_shmem(struct mm_struct *mm,
1556 struct address_space *mapping,
1557 pgoff_t start, struct page **hpage)
1559 struct page *page = NULL;
1560 struct radix_tree_iter iter;
1563 int node = NUMA_NO_NODE;
1564 int result = SCAN_SUCCEED;
1568 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1570 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1571 if (iter.index >= start + HPAGE_PMD_NR)
1574 page = radix_tree_deref_slot(slot);
1575 if (radix_tree_deref_retry(page)) {
1576 slot = radix_tree_iter_retry(&iter);
1580 if (radix_tree_exception(page)) {
1581 if (++swap > khugepaged_max_ptes_swap) {
1582 result = SCAN_EXCEED_SWAP_PTE;
1588 if (PageTransCompound(page)) {
1589 result = SCAN_PAGE_COMPOUND;
1593 node = page_to_nid(page);
1594 if (khugepaged_scan_abort(node)) {
1595 result = SCAN_SCAN_ABORT;
1598 khugepaged_node_load[node]++;
1600 if (!PageLRU(page)) {
1601 result = SCAN_PAGE_LRU;
1605 if (page_count(page) != 1 + page_mapcount(page)) {
1606 result = SCAN_PAGE_COUNT;
1611 * We probably should check if the page is referenced here, but
1612 * nobody would transfer pte_young() to PageReferenced() for us.
1613 * And rmap walk here is just too costly...
1618 if (need_resched()) {
1620 slot = radix_tree_iter_next(&iter);
1625 if (result == SCAN_SUCCEED) {
1626 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1627 result = SCAN_EXCEED_NONE_PTE;
1629 node = khugepaged_find_target_node();
1630 collapse_shmem(mm, mapping, start, hpage, node);
1634 /* TODO: tracepoints */
1637 static void khugepaged_scan_shmem(struct mm_struct *mm,
1638 struct address_space *mapping,
1639 pgoff_t start, struct page **hpage)
1645 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1646 struct page **hpage)
1647 __releases(&khugepaged_mm_lock)
1648 __acquires(&khugepaged_mm_lock)
1650 struct mm_slot *mm_slot;
1651 struct mm_struct *mm;
1652 struct vm_area_struct *vma;
1656 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1658 if (khugepaged_scan.mm_slot)
1659 mm_slot = khugepaged_scan.mm_slot;
1661 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1662 struct mm_slot, mm_node);
1663 khugepaged_scan.address = 0;
1664 khugepaged_scan.mm_slot = mm_slot;
1666 spin_unlock(&khugepaged_mm_lock);
1669 down_read(&mm->mmap_sem);
1670 if (unlikely(khugepaged_test_exit(mm)))
1673 vma = find_vma(mm, khugepaged_scan.address);
1676 for (; vma; vma = vma->vm_next) {
1677 unsigned long hstart, hend;
1680 if (unlikely(khugepaged_test_exit(mm))) {
1684 if (!hugepage_vma_check(vma)) {
1689 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1690 hend = vma->vm_end & HPAGE_PMD_MASK;
1693 if (khugepaged_scan.address > hend)
1695 if (khugepaged_scan.address < hstart)
1696 khugepaged_scan.address = hstart;
1697 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1699 while (khugepaged_scan.address < hend) {
1702 if (unlikely(khugepaged_test_exit(mm)))
1703 goto breakouterloop;
1705 VM_BUG_ON(khugepaged_scan.address < hstart ||
1706 khugepaged_scan.address + HPAGE_PMD_SIZE >
1708 if (shmem_file(vma->vm_file)) {
1710 pgoff_t pgoff = linear_page_index(vma,
1711 khugepaged_scan.address);
1712 if (!shmem_huge_enabled(vma))
1714 file = get_file(vma->vm_file);
1715 up_read(&mm->mmap_sem);
1717 khugepaged_scan_shmem(mm, file->f_mapping,
1721 ret = khugepaged_scan_pmd(mm, vma,
1722 khugepaged_scan.address,
1725 /* move to next address */
1726 khugepaged_scan.address += HPAGE_PMD_SIZE;
1727 progress += HPAGE_PMD_NR;
1729 /* we released mmap_sem so break loop */
1730 goto breakouterloop_mmap_sem;
1731 if (progress >= pages)
1732 goto breakouterloop;
1736 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1737 breakouterloop_mmap_sem:
1739 spin_lock(&khugepaged_mm_lock);
1740 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1742 * Release the current mm_slot if this mm is about to die, or
1743 * if we scanned all vmas of this mm.
1745 if (khugepaged_test_exit(mm) || !vma) {
1747 * Make sure that if mm_users is reaching zero while
1748 * khugepaged runs here, khugepaged_exit will find
1749 * mm_slot not pointing to the exiting mm.
1751 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1752 khugepaged_scan.mm_slot = list_entry(
1753 mm_slot->mm_node.next,
1754 struct mm_slot, mm_node);
1755 khugepaged_scan.address = 0;
1757 khugepaged_scan.mm_slot = NULL;
1758 khugepaged_full_scans++;
1761 collect_mm_slot(mm_slot);
1767 static int khugepaged_has_work(void)
1769 return !list_empty(&khugepaged_scan.mm_head) &&
1770 khugepaged_enabled();
1773 static int khugepaged_wait_event(void)
1775 return !list_empty(&khugepaged_scan.mm_head) ||
1776 kthread_should_stop();
1779 static void khugepaged_do_scan(void)
1781 struct page *hpage = NULL;
1782 unsigned int progress = 0, pass_through_head = 0;
1783 unsigned int pages = khugepaged_pages_to_scan;
1786 barrier(); /* write khugepaged_pages_to_scan to local stack */
1788 while (progress < pages) {
1789 if (!khugepaged_prealloc_page(&hpage, &wait))
1794 if (unlikely(kthread_should_stop() || try_to_freeze()))
1797 spin_lock(&khugepaged_mm_lock);
1798 if (!khugepaged_scan.mm_slot)
1799 pass_through_head++;
1800 if (khugepaged_has_work() &&
1801 pass_through_head < 2)
1802 progress += khugepaged_scan_mm_slot(pages - progress,
1806 spin_unlock(&khugepaged_mm_lock);
1809 if (!IS_ERR_OR_NULL(hpage))
1813 static bool khugepaged_should_wakeup(void)
1815 return kthread_should_stop() ||
1816 time_after_eq(jiffies, khugepaged_sleep_expire);
1819 static void khugepaged_wait_work(void)
1821 if (khugepaged_has_work()) {
1822 const unsigned long scan_sleep_jiffies =
1823 msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1825 if (!scan_sleep_jiffies)
1828 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1829 wait_event_freezable_timeout(khugepaged_wait,
1830 khugepaged_should_wakeup(),
1831 scan_sleep_jiffies);
1835 if (khugepaged_enabled())
1836 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1839 static int khugepaged(void *none)
1841 struct mm_slot *mm_slot;
1844 set_user_nice(current, MAX_NICE);
1846 while (!kthread_should_stop()) {
1847 khugepaged_do_scan();
1848 khugepaged_wait_work();
1851 spin_lock(&khugepaged_mm_lock);
1852 mm_slot = khugepaged_scan.mm_slot;
1853 khugepaged_scan.mm_slot = NULL;
1855 collect_mm_slot(mm_slot);
1856 spin_unlock(&khugepaged_mm_lock);
1860 static void set_recommended_min_free_kbytes(void)
1864 unsigned long recommended_min;
1866 for_each_populated_zone(zone)
1869 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1870 recommended_min = pageblock_nr_pages * nr_zones * 2;
1873 * Make sure that on average at least two pageblocks are almost free
1874 * of another type, one for a migratetype to fall back to and a
1875 * second to avoid subsequent fallbacks of other types There are 3
1876 * MIGRATE_TYPES we care about.
1878 recommended_min += pageblock_nr_pages * nr_zones *
1879 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1881 /* don't ever allow to reserve more than 5% of the lowmem */
1882 recommended_min = min(recommended_min,
1883 (unsigned long) nr_free_buffer_pages() / 20);
1884 recommended_min <<= (PAGE_SHIFT-10);
1886 if (recommended_min > min_free_kbytes) {
1887 if (user_min_free_kbytes >= 0)
1888 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1889 min_free_kbytes, recommended_min);
1891 min_free_kbytes = recommended_min;
1893 setup_per_zone_wmarks();
1896 int start_stop_khugepaged(void)
1898 static struct task_struct *khugepaged_thread __read_mostly;
1899 static DEFINE_MUTEX(khugepaged_mutex);
1902 mutex_lock(&khugepaged_mutex);
1903 if (khugepaged_enabled()) {
1904 if (!khugepaged_thread)
1905 khugepaged_thread = kthread_run(khugepaged, NULL,
1907 if (IS_ERR(khugepaged_thread)) {
1908 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1909 err = PTR_ERR(khugepaged_thread);
1910 khugepaged_thread = NULL;
1914 if (!list_empty(&khugepaged_scan.mm_head))
1915 wake_up_interruptible(&khugepaged_wait);
1917 set_recommended_min_free_kbytes();
1918 } else if (khugepaged_thread) {
1919 kthread_stop(khugepaged_thread);
1920 khugepaged_thread = NULL;
1923 mutex_unlock(&khugepaged_mutex);