1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2009 Red Hat, Inc.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 #include <linux/sched.h>
10 #include <linux/sched/mm.h>
11 #include <linux/sched/coredump.h>
12 #include <linux/sched/numa_balancing.h>
13 #include <linux/highmem.h>
14 #include <linux/hugetlb.h>
15 #include <linux/mmu_notifier.h>
16 #include <linux/rmap.h>
17 #include <linux/swap.h>
18 #include <linux/shrinker.h>
19 #include <linux/mm_inline.h>
20 #include <linux/swapops.h>
21 #include <linux/dax.h>
22 #include <linux/khugepaged.h>
23 #include <linux/freezer.h>
24 #include <linux/pfn_t.h>
25 #include <linux/mman.h>
26 #include <linux/memremap.h>
27 #include <linux/pagemap.h>
28 #include <linux/debugfs.h>
29 #include <linux/migrate.h>
30 #include <linux/hashtable.h>
31 #include <linux/userfaultfd_k.h>
32 #include <linux/page_idle.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/oom.h>
35 #include <linux/numa.h>
36 #include <linux/page_owner.h>
39 #include <asm/pgalloc.h>
43 * By default, transparent hugepage support is disabled in order to avoid
44 * risking an increased memory footprint for applications that are not
45 * guaranteed to benefit from it. When transparent hugepage support is
46 * enabled, it is for all mappings, and khugepaged scans all mappings.
47 * Defrag is invoked by khugepaged hugepage allocations and by page faults
48 * for all hugepage allocations.
50 unsigned long transparent_hugepage_flags __read_mostly =
51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
52 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
54 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
55 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
58 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
59 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
61 static struct shrinker deferred_split_shrinker;
63 static atomic_t huge_zero_refcount;
64 struct page *huge_zero_page __read_mostly;
65 unsigned long huge_zero_pfn __read_mostly = ~0UL;
67 static inline bool file_thp_enabled(struct vm_area_struct *vma)
69 return transhuge_vma_enabled(vma, vma->vm_flags) && vma->vm_file &&
70 !inode_is_open_for_write(vma->vm_file->f_inode) &&
71 (vma->vm_flags & VM_EXEC);
74 bool transparent_hugepage_active(struct vm_area_struct *vma)
76 /* The addr is used to check if the vma size fits */
77 unsigned long addr = (vma->vm_end & HPAGE_PMD_MASK) - HPAGE_PMD_SIZE;
79 if (!transhuge_vma_suitable(vma, addr))
81 if (vma_is_anonymous(vma))
82 return __transparent_hugepage_enabled(vma);
83 if (vma_is_shmem(vma))
84 return shmem_huge_enabled(vma);
85 if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS))
86 return file_thp_enabled(vma);
91 static bool get_huge_zero_page(void)
93 struct page *zero_page;
95 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
98 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
101 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
104 count_vm_event(THP_ZERO_PAGE_ALLOC);
106 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
108 __free_pages(zero_page, compound_order(zero_page));
111 WRITE_ONCE(huge_zero_pfn, page_to_pfn(zero_page));
113 /* We take additional reference here. It will be put back by shrinker */
114 atomic_set(&huge_zero_refcount, 2);
119 static void put_huge_zero_page(void)
122 * Counter should never go to zero here. Only shrinker can put
125 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
128 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
130 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
131 return READ_ONCE(huge_zero_page);
133 if (!get_huge_zero_page())
136 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
137 put_huge_zero_page();
139 return READ_ONCE(huge_zero_page);
142 void mm_put_huge_zero_page(struct mm_struct *mm)
144 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
145 put_huge_zero_page();
148 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
149 struct shrink_control *sc)
151 /* we can free zero page only if last reference remains */
152 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
155 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
156 struct shrink_control *sc)
158 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
159 struct page *zero_page = xchg(&huge_zero_page, NULL);
160 BUG_ON(zero_page == NULL);
161 WRITE_ONCE(huge_zero_pfn, ~0UL);
162 __free_pages(zero_page, compound_order(zero_page));
169 static struct shrinker huge_zero_page_shrinker = {
170 .count_objects = shrink_huge_zero_page_count,
171 .scan_objects = shrink_huge_zero_page_scan,
172 .seeks = DEFAULT_SEEKS,
176 static ssize_t enabled_show(struct kobject *kobj,
177 struct kobj_attribute *attr, char *buf)
181 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
182 output = "[always] madvise never";
183 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
184 &transparent_hugepage_flags))
185 output = "always [madvise] never";
187 output = "always madvise [never]";
189 return sysfs_emit(buf, "%s\n", output);
192 static ssize_t enabled_store(struct kobject *kobj,
193 struct kobj_attribute *attr,
194 const char *buf, size_t count)
198 if (sysfs_streq(buf, "always")) {
199 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
200 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
201 } else if (sysfs_streq(buf, "madvise")) {
202 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
203 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
204 } else if (sysfs_streq(buf, "never")) {
205 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
206 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
211 int err = start_stop_khugepaged();
217 static struct kobj_attribute enabled_attr =
218 __ATTR(enabled, 0644, enabled_show, enabled_store);
220 ssize_t single_hugepage_flag_show(struct kobject *kobj,
221 struct kobj_attribute *attr, char *buf,
222 enum transparent_hugepage_flag flag)
224 return sysfs_emit(buf, "%d\n",
225 !!test_bit(flag, &transparent_hugepage_flags));
228 ssize_t single_hugepage_flag_store(struct kobject *kobj,
229 struct kobj_attribute *attr,
230 const char *buf, size_t count,
231 enum transparent_hugepage_flag flag)
236 ret = kstrtoul(buf, 10, &value);
243 set_bit(flag, &transparent_hugepage_flags);
245 clear_bit(flag, &transparent_hugepage_flags);
250 static ssize_t defrag_show(struct kobject *kobj,
251 struct kobj_attribute *attr, char *buf)
255 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
256 &transparent_hugepage_flags))
257 output = "[always] defer defer+madvise madvise never";
258 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
259 &transparent_hugepage_flags))
260 output = "always [defer] defer+madvise madvise never";
261 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
262 &transparent_hugepage_flags))
263 output = "always defer [defer+madvise] madvise never";
264 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
265 &transparent_hugepage_flags))
266 output = "always defer defer+madvise [madvise] never";
268 output = "always defer defer+madvise madvise [never]";
270 return sysfs_emit(buf, "%s\n", output);
273 static ssize_t defrag_store(struct kobject *kobj,
274 struct kobj_attribute *attr,
275 const char *buf, size_t count)
277 if (sysfs_streq(buf, "always")) {
278 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
279 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
280 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
281 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
282 } else if (sysfs_streq(buf, "defer+madvise")) {
283 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
284 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
285 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
286 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
287 } else if (sysfs_streq(buf, "defer")) {
288 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
289 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
290 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
291 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
292 } else if (sysfs_streq(buf, "madvise")) {
293 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
294 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
295 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
296 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
297 } else if (sysfs_streq(buf, "never")) {
298 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
299 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
300 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
301 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
307 static struct kobj_attribute defrag_attr =
308 __ATTR(defrag, 0644, defrag_show, defrag_store);
310 static ssize_t use_zero_page_show(struct kobject *kobj,
311 struct kobj_attribute *attr, char *buf)
313 return single_hugepage_flag_show(kobj, attr, buf,
314 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
316 static ssize_t use_zero_page_store(struct kobject *kobj,
317 struct kobj_attribute *attr, const char *buf, size_t count)
319 return single_hugepage_flag_store(kobj, attr, buf, count,
320 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
322 static struct kobj_attribute use_zero_page_attr =
323 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
325 static ssize_t hpage_pmd_size_show(struct kobject *kobj,
326 struct kobj_attribute *attr, char *buf)
328 return sysfs_emit(buf, "%lu\n", HPAGE_PMD_SIZE);
330 static struct kobj_attribute hpage_pmd_size_attr =
331 __ATTR_RO(hpage_pmd_size);
333 static struct attribute *hugepage_attr[] = {
336 &use_zero_page_attr.attr,
337 &hpage_pmd_size_attr.attr,
339 &shmem_enabled_attr.attr,
344 static const struct attribute_group hugepage_attr_group = {
345 .attrs = hugepage_attr,
348 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
352 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
353 if (unlikely(!*hugepage_kobj)) {
354 pr_err("failed to create transparent hugepage kobject\n");
358 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
360 pr_err("failed to register transparent hugepage group\n");
364 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
366 pr_err("failed to register transparent hugepage group\n");
367 goto remove_hp_group;
373 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
375 kobject_put(*hugepage_kobj);
379 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
381 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
382 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
383 kobject_put(hugepage_kobj);
386 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
391 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
394 #endif /* CONFIG_SYSFS */
396 static int __init hugepage_init(void)
399 struct kobject *hugepage_kobj;
401 if (!has_transparent_hugepage()) {
403 * Hardware doesn't support hugepages, hence disable
406 transparent_hugepage_flags = 1 << TRANSPARENT_HUGEPAGE_NEVER_DAX;
411 * hugepages can't be allocated by the buddy allocator
413 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
415 * we use page->mapping and page->index in second tail page
416 * as list_head: assuming THP order >= 2
418 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
420 err = hugepage_init_sysfs(&hugepage_kobj);
424 err = khugepaged_init();
428 err = register_shrinker(&huge_zero_page_shrinker);
430 goto err_hzp_shrinker;
431 err = register_shrinker(&deferred_split_shrinker);
433 goto err_split_shrinker;
436 * By default disable transparent hugepages on smaller systems,
437 * where the extra memory used could hurt more than TLB overhead
438 * is likely to save. The admin can still enable it through /sys.
440 if (totalram_pages() < (512 << (20 - PAGE_SHIFT))) {
441 transparent_hugepage_flags = 0;
445 err = start_stop_khugepaged();
451 unregister_shrinker(&deferred_split_shrinker);
453 unregister_shrinker(&huge_zero_page_shrinker);
455 khugepaged_destroy();
457 hugepage_exit_sysfs(hugepage_kobj);
461 subsys_initcall(hugepage_init);
463 static int __init setup_transparent_hugepage(char *str)
468 if (!strcmp(str, "always")) {
469 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
470 &transparent_hugepage_flags);
471 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
472 &transparent_hugepage_flags);
474 } else if (!strcmp(str, "madvise")) {
475 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
476 &transparent_hugepage_flags);
477 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
478 &transparent_hugepage_flags);
480 } else if (!strcmp(str, "never")) {
481 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
482 &transparent_hugepage_flags);
483 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
484 &transparent_hugepage_flags);
489 pr_warn("transparent_hugepage= cannot parse, ignored\n");
492 __setup("transparent_hugepage=", setup_transparent_hugepage);
494 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
496 if (likely(vma->vm_flags & VM_WRITE))
497 pmd = pmd_mkwrite(pmd);
502 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
504 struct mem_cgroup *memcg = page_memcg(compound_head(page));
505 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
508 return &memcg->deferred_split_queue;
510 return &pgdat->deferred_split_queue;
513 static inline struct deferred_split *get_deferred_split_queue(struct page *page)
515 struct pglist_data *pgdat = NODE_DATA(page_to_nid(page));
517 return &pgdat->deferred_split_queue;
521 void prep_transhuge_page(struct page *page)
524 * we use page->mapping and page->indexlru in second tail page
525 * as list_head: assuming THP order >= 2
528 INIT_LIST_HEAD(page_deferred_list(page));
529 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
532 bool is_transparent_hugepage(struct page *page)
534 if (!PageCompound(page))
537 page = compound_head(page);
538 return is_huge_zero_page(page) ||
539 page[1].compound_dtor == TRANSHUGE_PAGE_DTOR;
541 EXPORT_SYMBOL_GPL(is_transparent_hugepage);
543 static unsigned long __thp_get_unmapped_area(struct file *filp,
544 unsigned long addr, unsigned long len,
545 loff_t off, unsigned long flags, unsigned long size)
547 loff_t off_end = off + len;
548 loff_t off_align = round_up(off, size);
549 unsigned long len_pad, ret;
551 if (off_end <= off_align || (off_end - off_align) < size)
554 len_pad = len + size;
555 if (len_pad < len || (off + len_pad) < off)
558 ret = current->mm->get_unmapped_area(filp, addr, len_pad,
559 off >> PAGE_SHIFT, flags);
562 * The failure might be due to length padding. The caller will retry
563 * without the padding.
565 if (IS_ERR_VALUE(ret))
569 * Do not try to align to THP boundary if allocation at the address
575 ret += (off - ret) & (size - 1);
579 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
580 unsigned long len, unsigned long pgoff, unsigned long flags)
583 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
585 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
588 ret = __thp_get_unmapped_area(filp, addr, len, off, flags, PMD_SIZE);
592 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
594 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
596 static vm_fault_t __do_huge_pmd_anonymous_page(struct vm_fault *vmf,
597 struct page *page, gfp_t gfp)
599 struct vm_area_struct *vma = vmf->vma;
601 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
604 VM_BUG_ON_PAGE(!PageCompound(page), page);
606 if (mem_cgroup_charge(page, vma->vm_mm, gfp)) {
608 count_vm_event(THP_FAULT_FALLBACK);
609 count_vm_event(THP_FAULT_FALLBACK_CHARGE);
610 return VM_FAULT_FALLBACK;
612 cgroup_throttle_swaprate(page, gfp);
614 pgtable = pte_alloc_one(vma->vm_mm);
615 if (unlikely(!pgtable)) {
620 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
622 * The memory barrier inside __SetPageUptodate makes sure that
623 * clear_huge_page writes become visible before the set_pmd_at()
626 __SetPageUptodate(page);
628 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
629 if (unlikely(!pmd_none(*vmf->pmd))) {
634 ret = check_stable_address_space(vma->vm_mm);
638 /* Deliver the page fault to userland */
639 if (userfaultfd_missing(vma)) {
640 spin_unlock(vmf->ptl);
642 pte_free(vma->vm_mm, pgtable);
643 ret = handle_userfault(vmf, VM_UFFD_MISSING);
644 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
648 entry = mk_huge_pmd(page, vma->vm_page_prot);
649 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
650 page_add_new_anon_rmap(page, vma, haddr, true);
651 lru_cache_add_inactive_or_unevictable(page, vma);
652 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
653 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
654 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
655 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
656 mm_inc_nr_ptes(vma->vm_mm);
657 spin_unlock(vmf->ptl);
658 count_vm_event(THP_FAULT_ALLOC);
659 count_memcg_event_mm(vma->vm_mm, THP_FAULT_ALLOC);
664 spin_unlock(vmf->ptl);
667 pte_free(vma->vm_mm, pgtable);
674 * always: directly stall for all thp allocations
675 * defer: wake kswapd and fail if not immediately available
676 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
677 * fail if not immediately available
678 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
680 * never: never stall for any thp allocation
682 gfp_t vma_thp_gfp_mask(struct vm_area_struct *vma)
684 const bool vma_madvised = vma && (vma->vm_flags & VM_HUGEPAGE);
686 /* Always do synchronous compaction */
687 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
688 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
690 /* Kick kcompactd and fail quickly */
691 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
692 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
694 /* Synchronous compaction if madvised, otherwise kick kcompactd */
695 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
696 return GFP_TRANSHUGE_LIGHT |
697 (vma_madvised ? __GFP_DIRECT_RECLAIM :
698 __GFP_KSWAPD_RECLAIM);
700 /* Only do synchronous compaction if madvised */
701 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
702 return GFP_TRANSHUGE_LIGHT |
703 (vma_madvised ? __GFP_DIRECT_RECLAIM : 0);
705 return GFP_TRANSHUGE_LIGHT;
708 /* Caller must hold page table lock. */
709 static void set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
710 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
711 struct page *zero_page)
716 entry = mk_pmd(zero_page, vma->vm_page_prot);
717 entry = pmd_mkhuge(entry);
719 pgtable_trans_huge_deposit(mm, pmd, pgtable);
720 set_pmd_at(mm, haddr, pmd, entry);
724 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf)
726 struct vm_area_struct *vma = vmf->vma;
729 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
731 if (!transhuge_vma_suitable(vma, haddr))
732 return VM_FAULT_FALLBACK;
733 if (unlikely(anon_vma_prepare(vma)))
735 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
737 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
738 !mm_forbids_zeropage(vma->vm_mm) &&
739 transparent_hugepage_use_zero_page()) {
741 struct page *zero_page;
743 pgtable = pte_alloc_one(vma->vm_mm);
744 if (unlikely(!pgtable))
746 zero_page = mm_get_huge_zero_page(vma->vm_mm);
747 if (unlikely(!zero_page)) {
748 pte_free(vma->vm_mm, pgtable);
749 count_vm_event(THP_FAULT_FALLBACK);
750 return VM_FAULT_FALLBACK;
752 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
754 if (pmd_none(*vmf->pmd)) {
755 ret = check_stable_address_space(vma->vm_mm);
757 spin_unlock(vmf->ptl);
758 pte_free(vma->vm_mm, pgtable);
759 } else if (userfaultfd_missing(vma)) {
760 spin_unlock(vmf->ptl);
761 pte_free(vma->vm_mm, pgtable);
762 ret = handle_userfault(vmf, VM_UFFD_MISSING);
763 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
765 set_huge_zero_page(pgtable, vma->vm_mm, vma,
766 haddr, vmf->pmd, zero_page);
767 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
768 spin_unlock(vmf->ptl);
771 spin_unlock(vmf->ptl);
772 pte_free(vma->vm_mm, pgtable);
776 gfp = vma_thp_gfp_mask(vma);
777 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
778 if (unlikely(!page)) {
779 count_vm_event(THP_FAULT_FALLBACK);
780 return VM_FAULT_FALLBACK;
782 prep_transhuge_page(page);
783 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
786 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
787 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
790 struct mm_struct *mm = vma->vm_mm;
794 ptl = pmd_lock(mm, pmd);
795 if (!pmd_none(*pmd)) {
797 if (pmd_pfn(*pmd) != pfn_t_to_pfn(pfn)) {
798 WARN_ON_ONCE(!is_huge_zero_pmd(*pmd));
801 entry = pmd_mkyoung(*pmd);
802 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
803 if (pmdp_set_access_flags(vma, addr, pmd, entry, 1))
804 update_mmu_cache_pmd(vma, addr, pmd);
810 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
811 if (pfn_t_devmap(pfn))
812 entry = pmd_mkdevmap(entry);
814 entry = pmd_mkyoung(pmd_mkdirty(entry));
815 entry = maybe_pmd_mkwrite(entry, vma);
819 pgtable_trans_huge_deposit(mm, pmd, pgtable);
824 set_pmd_at(mm, addr, pmd, entry);
825 update_mmu_cache_pmd(vma, addr, pmd);
830 pte_free(mm, pgtable);
834 * vmf_insert_pfn_pmd_prot - insert a pmd size pfn
835 * @vmf: Structure describing the fault
836 * @pfn: pfn to insert
837 * @pgprot: page protection to use
838 * @write: whether it's a write fault
840 * Insert a pmd size pfn. See vmf_insert_pfn() for additional info and
841 * also consult the vmf_insert_mixed_prot() documentation when
842 * @pgprot != @vmf->vma->vm_page_prot.
844 * Return: vm_fault_t value.
846 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
847 pgprot_t pgprot, bool write)
849 unsigned long addr = vmf->address & PMD_MASK;
850 struct vm_area_struct *vma = vmf->vma;
851 pgtable_t pgtable = NULL;
854 * If we had pmd_special, we could avoid all these restrictions,
855 * but we need to be consistent with PTEs and architectures that
856 * can't support a 'special' bit.
858 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
860 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
861 (VM_PFNMAP|VM_MIXEDMAP));
862 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
864 if (addr < vma->vm_start || addr >= vma->vm_end)
865 return VM_FAULT_SIGBUS;
867 if (arch_needs_pgtable_deposit()) {
868 pgtable = pte_alloc_one(vma->vm_mm);
873 track_pfn_insert(vma, &pgprot, pfn);
875 insert_pfn_pmd(vma, addr, vmf->pmd, pfn, pgprot, write, pgtable);
876 return VM_FAULT_NOPAGE;
878 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd_prot);
880 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
881 static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
883 if (likely(vma->vm_flags & VM_WRITE))
884 pud = pud_mkwrite(pud);
888 static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
889 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
891 struct mm_struct *mm = vma->vm_mm;
895 ptl = pud_lock(mm, pud);
896 if (!pud_none(*pud)) {
898 if (pud_pfn(*pud) != pfn_t_to_pfn(pfn)) {
899 WARN_ON_ONCE(!is_huge_zero_pud(*pud));
902 entry = pud_mkyoung(*pud);
903 entry = maybe_pud_mkwrite(pud_mkdirty(entry), vma);
904 if (pudp_set_access_flags(vma, addr, pud, entry, 1))
905 update_mmu_cache_pud(vma, addr, pud);
910 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
911 if (pfn_t_devmap(pfn))
912 entry = pud_mkdevmap(entry);
914 entry = pud_mkyoung(pud_mkdirty(entry));
915 entry = maybe_pud_mkwrite(entry, vma);
917 set_pud_at(mm, addr, pud, entry);
918 update_mmu_cache_pud(vma, addr, pud);
925 * vmf_insert_pfn_pud_prot - insert a pud size pfn
926 * @vmf: Structure describing the fault
927 * @pfn: pfn to insert
928 * @pgprot: page protection to use
929 * @write: whether it's a write fault
931 * Insert a pud size pfn. See vmf_insert_pfn() for additional info and
932 * also consult the vmf_insert_mixed_prot() documentation when
933 * @pgprot != @vmf->vma->vm_page_prot.
935 * Return: vm_fault_t value.
937 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
938 pgprot_t pgprot, bool write)
940 unsigned long addr = vmf->address & PUD_MASK;
941 struct vm_area_struct *vma = vmf->vma;
944 * If we had pud_special, we could avoid all these restrictions,
945 * but we need to be consistent with PTEs and architectures that
946 * can't support a 'special' bit.
948 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) &&
950 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
951 (VM_PFNMAP|VM_MIXEDMAP));
952 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
954 if (addr < vma->vm_start || addr >= vma->vm_end)
955 return VM_FAULT_SIGBUS;
957 track_pfn_insert(vma, &pgprot, pfn);
959 insert_pfn_pud(vma, addr, vmf->pud, pfn, pgprot, write);
960 return VM_FAULT_NOPAGE;
962 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud_prot);
963 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
965 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
966 pmd_t *pmd, int flags)
970 _pmd = pmd_mkyoung(*pmd);
971 if (flags & FOLL_WRITE)
972 _pmd = pmd_mkdirty(_pmd);
973 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
974 pmd, _pmd, flags & FOLL_WRITE))
975 update_mmu_cache_pmd(vma, addr, pmd);
978 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
979 pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
981 unsigned long pfn = pmd_pfn(*pmd);
982 struct mm_struct *mm = vma->vm_mm;
985 assert_spin_locked(pmd_lockptr(mm, pmd));
988 * When we COW a devmap PMD entry, we split it into PTEs, so we should
989 * not be in this function with `flags & FOLL_COW` set.
991 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
993 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
994 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
995 (FOLL_PIN | FOLL_GET)))
998 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1001 if (pmd_present(*pmd) && pmd_devmap(*pmd))
1006 if (flags & FOLL_TOUCH)
1007 touch_pmd(vma, addr, pmd, flags);
1010 * device mapped pages can only be returned if the
1011 * caller will manage the page reference count.
1013 if (!(flags & (FOLL_GET | FOLL_PIN)))
1014 return ERR_PTR(-EEXIST);
1016 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1017 *pgmap = get_dev_pagemap(pfn, *pgmap);
1019 return ERR_PTR(-EFAULT);
1020 page = pfn_to_page(pfn);
1021 if (!try_grab_page(page, flags))
1022 page = ERR_PTR(-ENOMEM);
1027 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1028 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1029 struct vm_area_struct *vma)
1031 spinlock_t *dst_ptl, *src_ptl;
1032 struct page *src_page;
1034 pgtable_t pgtable = NULL;
1037 /* Skip if can be re-fill on fault */
1038 if (!vma_is_anonymous(vma))
1041 pgtable = pte_alloc_one(dst_mm);
1042 if (unlikely(!pgtable))
1045 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1046 src_ptl = pmd_lockptr(src_mm, src_pmd);
1047 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1053 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
1054 * does not have the VM_UFFD_WP, which means that the uffd
1055 * fork event is not enabled.
1057 if (!(vma->vm_flags & VM_UFFD_WP))
1058 pmd = pmd_clear_uffd_wp(pmd);
1060 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1061 if (unlikely(is_swap_pmd(pmd))) {
1062 swp_entry_t entry = pmd_to_swp_entry(pmd);
1064 VM_BUG_ON(!is_pmd_migration_entry(pmd));
1065 if (is_write_migration_entry(entry)) {
1066 make_migration_entry_read(&entry);
1067 pmd = swp_entry_to_pmd(entry);
1068 if (pmd_swp_soft_dirty(*src_pmd))
1069 pmd = pmd_swp_mksoft_dirty(pmd);
1070 set_pmd_at(src_mm, addr, src_pmd, pmd);
1072 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1073 mm_inc_nr_ptes(dst_mm);
1074 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1075 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1081 if (unlikely(!pmd_trans_huge(pmd))) {
1082 pte_free(dst_mm, pgtable);
1086 * When page table lock is held, the huge zero pmd should not be
1087 * under splitting since we don't split the page itself, only pmd to
1090 if (is_huge_zero_pmd(pmd)) {
1091 struct page *zero_page;
1093 * get_huge_zero_page() will never allocate a new page here,
1094 * since we already have a zero page to copy. It just takes a
1097 zero_page = mm_get_huge_zero_page(dst_mm);
1098 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1104 src_page = pmd_page(pmd);
1105 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1108 * If this page is a potentially pinned page, split and retry the fault
1109 * with smaller page size. Normally this should not happen because the
1110 * userspace should use MADV_DONTFORK upon pinned regions. This is a
1111 * best effort that the pinned pages won't be replaced by another
1112 * random page during the coming copy-on-write.
1114 if (unlikely(page_needs_cow_for_dma(vma, src_page))) {
1115 pte_free(dst_mm, pgtable);
1116 spin_unlock(src_ptl);
1117 spin_unlock(dst_ptl);
1118 __split_huge_pmd(vma, src_pmd, addr, false, NULL);
1123 page_dup_rmap(src_page, true);
1124 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1125 mm_inc_nr_ptes(dst_mm);
1126 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1128 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1129 pmd = pmd_mkold(pmd_wrprotect(pmd));
1130 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1134 spin_unlock(src_ptl);
1135 spin_unlock(dst_ptl);
1140 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1141 static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
1142 pud_t *pud, int flags)
1146 _pud = pud_mkyoung(*pud);
1147 if (flags & FOLL_WRITE)
1148 _pud = pud_mkdirty(_pud);
1149 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1150 pud, _pud, flags & FOLL_WRITE))
1151 update_mmu_cache_pud(vma, addr, pud);
1154 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1155 pud_t *pud, int flags, struct dev_pagemap **pgmap)
1157 unsigned long pfn = pud_pfn(*pud);
1158 struct mm_struct *mm = vma->vm_mm;
1161 assert_spin_locked(pud_lockptr(mm, pud));
1163 if (flags & FOLL_WRITE && !pud_write(*pud))
1166 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
1167 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
1168 (FOLL_PIN | FOLL_GET)))
1171 if (pud_present(*pud) && pud_devmap(*pud))
1176 if (flags & FOLL_TOUCH)
1177 touch_pud(vma, addr, pud, flags);
1180 * device mapped pages can only be returned if the
1181 * caller will manage the page reference count.
1183 * At least one of FOLL_GET | FOLL_PIN must be set, so assert that here:
1185 if (!(flags & (FOLL_GET | FOLL_PIN)))
1186 return ERR_PTR(-EEXIST);
1188 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1189 *pgmap = get_dev_pagemap(pfn, *pgmap);
1191 return ERR_PTR(-EFAULT);
1192 page = pfn_to_page(pfn);
1193 if (!try_grab_page(page, flags))
1194 page = ERR_PTR(-ENOMEM);
1199 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1200 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1201 struct vm_area_struct *vma)
1203 spinlock_t *dst_ptl, *src_ptl;
1207 dst_ptl = pud_lock(dst_mm, dst_pud);
1208 src_ptl = pud_lockptr(src_mm, src_pud);
1209 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1213 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1217 * When page table lock is held, the huge zero pud should not be
1218 * under splitting since we don't split the page itself, only pud to
1221 if (is_huge_zero_pud(pud)) {
1222 /* No huge zero pud yet */
1225 /* Please refer to comments in copy_huge_pmd() */
1226 if (unlikely(page_needs_cow_for_dma(vma, pud_page(pud)))) {
1227 spin_unlock(src_ptl);
1228 spin_unlock(dst_ptl);
1229 __split_huge_pud(vma, src_pud, addr);
1233 pudp_set_wrprotect(src_mm, addr, src_pud);
1234 pud = pud_mkold(pud_wrprotect(pud));
1235 set_pud_at(dst_mm, addr, dst_pud, pud);
1239 spin_unlock(src_ptl);
1240 spin_unlock(dst_ptl);
1244 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1247 unsigned long haddr;
1248 bool write = vmf->flags & FAULT_FLAG_WRITE;
1250 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1251 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1254 entry = pud_mkyoung(orig_pud);
1256 entry = pud_mkdirty(entry);
1257 haddr = vmf->address & HPAGE_PUD_MASK;
1258 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1259 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1262 spin_unlock(vmf->ptl);
1264 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1266 void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1269 unsigned long haddr;
1270 bool write = vmf->flags & FAULT_FLAG_WRITE;
1272 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1273 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1276 entry = pmd_mkyoung(orig_pmd);
1278 entry = pmd_mkdirty(entry);
1279 haddr = vmf->address & HPAGE_PMD_MASK;
1280 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1281 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1284 spin_unlock(vmf->ptl);
1287 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1289 struct vm_area_struct *vma = vmf->vma;
1291 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1293 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1294 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1296 if (is_huge_zero_pmd(orig_pmd))
1299 spin_lock(vmf->ptl);
1301 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1302 spin_unlock(vmf->ptl);
1306 page = pmd_page(orig_pmd);
1307 VM_BUG_ON_PAGE(!PageHead(page), page);
1309 /* Lock page for reuse_swap_page() */
1310 if (!trylock_page(page)) {
1312 spin_unlock(vmf->ptl);
1314 spin_lock(vmf->ptl);
1315 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1316 spin_unlock(vmf->ptl);
1325 * We can only reuse the page if nobody else maps the huge page or it's
1328 if (reuse_swap_page(page, NULL)) {
1330 entry = pmd_mkyoung(orig_pmd);
1331 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1332 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1333 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1335 spin_unlock(vmf->ptl);
1336 return VM_FAULT_WRITE;
1340 spin_unlock(vmf->ptl);
1342 __split_huge_pmd(vma, vmf->pmd, vmf->address, false, NULL);
1343 return VM_FAULT_FALLBACK;
1347 * FOLL_FORCE can write to even unwritable pmd's, but only
1348 * after we've gone through a COW cycle and they are dirty.
1350 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1352 return pmd_write(pmd) ||
1353 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1356 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1361 struct mm_struct *mm = vma->vm_mm;
1362 struct page *page = NULL;
1364 assert_spin_locked(pmd_lockptr(mm, pmd));
1366 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1369 /* Avoid dumping huge zero page */
1370 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1371 return ERR_PTR(-EFAULT);
1373 /* Full NUMA hinting faults to serialise migration in fault paths */
1374 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1377 page = pmd_page(*pmd);
1378 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1380 if (!try_grab_page(page, flags))
1381 return ERR_PTR(-ENOMEM);
1383 if (flags & FOLL_TOUCH)
1384 touch_pmd(vma, addr, pmd, flags);
1386 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1388 * We don't mlock() pte-mapped THPs. This way we can avoid
1389 * leaking mlocked pages into non-VM_LOCKED VMAs.
1393 * In most cases the pmd is the only mapping of the page as we
1394 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1395 * writable private mappings in populate_vma_page_range().
1397 * The only scenario when we have the page shared here is if we
1398 * mlocking read-only mapping shared over fork(). We skip
1399 * mlocking such pages.
1403 * We can expect PageDoubleMap() to be stable under page lock:
1404 * for file pages we set it in page_add_file_rmap(), which
1405 * requires page to be locked.
1408 if (PageAnon(page) && compound_mapcount(page) != 1)
1410 if (PageDoubleMap(page) || !page->mapping)
1412 if (!trylock_page(page))
1414 if (page->mapping && !PageDoubleMap(page))
1415 mlock_vma_page(page);
1419 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1420 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1426 /* NUMA hinting page fault entry point for trans huge pmds */
1427 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1429 struct vm_area_struct *vma = vmf->vma;
1430 struct anon_vma *anon_vma = NULL;
1432 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1433 int page_nid = NUMA_NO_NODE, this_nid = numa_node_id();
1434 int target_nid, last_cpupid = -1;
1436 bool migrated = false;
1440 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1441 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1445 * If there are potential migrations, wait for completion and retry
1446 * without disrupting NUMA hinting information. Do not relock and
1447 * check_same as the page may no longer be mapped.
1449 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1450 page = pmd_page(*vmf->pmd);
1451 if (!get_page_unless_zero(page))
1453 spin_unlock(vmf->ptl);
1454 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
1458 page = pmd_page(pmd);
1459 BUG_ON(is_huge_zero_page(page));
1460 page_nid = page_to_nid(page);
1461 last_cpupid = page_cpupid_last(page);
1462 count_vm_numa_event(NUMA_HINT_FAULTS);
1463 if (page_nid == this_nid) {
1464 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1465 flags |= TNF_FAULT_LOCAL;
1468 /* See similar comment in do_numa_page for explanation */
1469 if (!pmd_savedwrite(pmd))
1470 flags |= TNF_NO_GROUP;
1473 * Acquire the page lock to serialise THP migrations but avoid dropping
1474 * page_table_lock if at all possible
1476 page_locked = trylock_page(page);
1477 target_nid = mpol_misplaced(page, vma, haddr);
1478 /* Migration could have started since the pmd_trans_migrating check */
1480 page_nid = NUMA_NO_NODE;
1481 if (!get_page_unless_zero(page))
1483 spin_unlock(vmf->ptl);
1484 put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE);
1486 } else if (target_nid == NUMA_NO_NODE) {
1487 /* There are no parallel migrations and page is in the right
1488 * node. Clear the numa hinting info in this pmd.
1494 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1495 * to serialises splits
1498 spin_unlock(vmf->ptl);
1499 anon_vma = page_lock_anon_vma_read(page);
1501 /* Confirm the PMD did not change while page_table_lock was released */
1502 spin_lock(vmf->ptl);
1503 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1506 page_nid = NUMA_NO_NODE;
1510 /* Bail if we fail to protect against THP splits for any reason */
1511 if (unlikely(!anon_vma)) {
1513 page_nid = NUMA_NO_NODE;
1518 * Since we took the NUMA fault, we must have observed the !accessible
1519 * bit. Make sure all other CPUs agree with that, to avoid them
1520 * modifying the page we're about to migrate.
1522 * Must be done under PTL such that we'll observe the relevant
1523 * inc_tlb_flush_pending().
1525 * We are not sure a pending tlb flush here is for a huge page
1526 * mapping or not. Hence use the tlb range variant
1528 if (mm_tlb_flush_pending(vma->vm_mm)) {
1529 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1531 * change_huge_pmd() released the pmd lock before
1532 * invalidating the secondary MMUs sharing the primary
1533 * MMU pagetables (with ->invalidate_range()). The
1534 * mmu_notifier_invalidate_range_end() (which
1535 * internally calls ->invalidate_range()) in
1536 * change_pmd_range() will run after us, so we can't
1537 * rely on it here and we need an explicit invalidate.
1539 mmu_notifier_invalidate_range(vma->vm_mm, haddr,
1540 haddr + HPAGE_PMD_SIZE);
1544 * Migrate the THP to the requested node, returns with page unlocked
1545 * and access rights restored.
1547 spin_unlock(vmf->ptl);
1549 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1550 vmf->pmd, pmd, vmf->address, page, target_nid);
1552 flags |= TNF_MIGRATED;
1553 page_nid = target_nid;
1555 flags |= TNF_MIGRATE_FAIL;
1559 BUG_ON(!PageLocked(page));
1560 was_writable = pmd_savedwrite(pmd);
1561 pmd = pmd_modify(pmd, vma->vm_page_prot);
1562 pmd = pmd_mkyoung(pmd);
1564 pmd = pmd_mkwrite(pmd);
1565 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1566 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1569 spin_unlock(vmf->ptl);
1573 page_unlock_anon_vma_read(anon_vma);
1575 if (page_nid != NUMA_NO_NODE)
1576 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1583 * Return true if we do MADV_FREE successfully on entire pmd page.
1584 * Otherwise, return false.
1586 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1587 pmd_t *pmd, unsigned long addr, unsigned long next)
1592 struct mm_struct *mm = tlb->mm;
1595 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1597 ptl = pmd_trans_huge_lock(pmd, vma);
1602 if (is_huge_zero_pmd(orig_pmd))
1605 if (unlikely(!pmd_present(orig_pmd))) {
1606 VM_BUG_ON(thp_migration_supported() &&
1607 !is_pmd_migration_entry(orig_pmd));
1611 page = pmd_page(orig_pmd);
1613 * If other processes are mapping this page, we couldn't discard
1614 * the page unless they all do MADV_FREE so let's skip the page.
1616 if (total_mapcount(page) != 1)
1619 if (!trylock_page(page))
1623 * If user want to discard part-pages of THP, split it so MADV_FREE
1624 * will deactivate only them.
1626 if (next - addr != HPAGE_PMD_SIZE) {
1629 split_huge_page(page);
1635 if (PageDirty(page))
1636 ClearPageDirty(page);
1639 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1640 pmdp_invalidate(vma, addr, pmd);
1641 orig_pmd = pmd_mkold(orig_pmd);
1642 orig_pmd = pmd_mkclean(orig_pmd);
1644 set_pmd_at(mm, addr, pmd, orig_pmd);
1645 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1648 mark_page_lazyfree(page);
1656 static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1660 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1661 pte_free(mm, pgtable);
1665 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1666 pmd_t *pmd, unsigned long addr)
1671 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
1673 ptl = __pmd_trans_huge_lock(pmd, vma);
1677 * For architectures like ppc64 we look at deposited pgtable
1678 * when calling pmdp_huge_get_and_clear. So do the
1679 * pgtable_trans_huge_withdraw after finishing pmdp related
1682 orig_pmd = pmdp_huge_get_and_clear_full(vma, addr, pmd,
1684 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1685 if (vma_is_special_huge(vma)) {
1686 if (arch_needs_pgtable_deposit())
1687 zap_deposited_table(tlb->mm, pmd);
1689 } else if (is_huge_zero_pmd(orig_pmd)) {
1690 zap_deposited_table(tlb->mm, pmd);
1693 struct page *page = NULL;
1694 int flush_needed = 1;
1696 if (pmd_present(orig_pmd)) {
1697 page = pmd_page(orig_pmd);
1698 page_remove_rmap(page, true);
1699 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1700 VM_BUG_ON_PAGE(!PageHead(page), page);
1701 } else if (thp_migration_supported()) {
1704 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1705 entry = pmd_to_swp_entry(orig_pmd);
1706 page = migration_entry_to_page(entry);
1709 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1711 if (PageAnon(page)) {
1712 zap_deposited_table(tlb->mm, pmd);
1713 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1715 if (arch_needs_pgtable_deposit())
1716 zap_deposited_table(tlb->mm, pmd);
1717 add_mm_counter(tlb->mm, mm_counter_file(page), -HPAGE_PMD_NR);
1722 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1727 #ifndef pmd_move_must_withdraw
1728 static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1729 spinlock_t *old_pmd_ptl,
1730 struct vm_area_struct *vma)
1733 * With split pmd lock we also need to move preallocated
1734 * PTE page table if new_pmd is on different PMD page table.
1736 * We also don't deposit and withdraw tables for file pages.
1738 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1742 static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1744 #ifdef CONFIG_MEM_SOFT_DIRTY
1745 if (unlikely(is_pmd_migration_entry(pmd)))
1746 pmd = pmd_swp_mksoft_dirty(pmd);
1747 else if (pmd_present(pmd))
1748 pmd = pmd_mksoft_dirty(pmd);
1753 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1754 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd)
1756 spinlock_t *old_ptl, *new_ptl;
1758 struct mm_struct *mm = vma->vm_mm;
1759 bool force_flush = false;
1762 * The destination pmd shouldn't be established, free_pgtables()
1763 * should have release it.
1765 if (WARN_ON(!pmd_none(*new_pmd))) {
1766 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1771 * We don't have to worry about the ordering of src and dst
1772 * ptlocks because exclusive mmap_lock prevents deadlock.
1774 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1776 new_ptl = pmd_lockptr(mm, new_pmd);
1777 if (new_ptl != old_ptl)
1778 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1779 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1780 if (pmd_present(pmd))
1782 VM_BUG_ON(!pmd_none(*new_pmd));
1784 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1786 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1787 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1789 pmd = move_soft_dirty_pmd(pmd);
1790 set_pmd_at(mm, new_addr, new_pmd, pmd);
1792 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1793 if (new_ptl != old_ptl)
1794 spin_unlock(new_ptl);
1795 spin_unlock(old_ptl);
1803 * - 0 if PMD could not be locked
1804 * - 1 if PMD was locked but protections unchanged and TLB flush unnecessary
1805 * - HPAGE_PMD_NR if protections changed and TLB flush necessary
1807 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1808 unsigned long addr, pgprot_t newprot, unsigned long cp_flags)
1810 struct mm_struct *mm = vma->vm_mm;
1813 bool preserve_write;
1815 bool prot_numa = cp_flags & MM_CP_PROT_NUMA;
1816 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
1817 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
1819 ptl = __pmd_trans_huge_lock(pmd, vma);
1823 preserve_write = prot_numa && pmd_write(*pmd);
1826 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1827 if (is_swap_pmd(*pmd)) {
1828 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1830 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1831 if (is_write_migration_entry(entry)) {
1834 * A protection check is difficult so
1835 * just be safe and disable write
1837 make_migration_entry_read(&entry);
1838 newpmd = swp_entry_to_pmd(entry);
1839 if (pmd_swp_soft_dirty(*pmd))
1840 newpmd = pmd_swp_mksoft_dirty(newpmd);
1841 set_pmd_at(mm, addr, pmd, newpmd);
1848 * Avoid trapping faults against the zero page. The read-only
1849 * data is likely to be read-cached on the local CPU and
1850 * local/remote hits to the zero page are not interesting.
1852 if (prot_numa && is_huge_zero_pmd(*pmd))
1855 if (prot_numa && pmd_protnone(*pmd))
1859 * In case prot_numa, we are under mmap_read_lock(mm). It's critical
1860 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1861 * which is also under mmap_read_lock(mm):
1864 * change_huge_pmd(prot_numa=1)
1865 * pmdp_huge_get_and_clear_notify()
1866 * madvise_dontneed()
1868 * pmd_trans_huge(*pmd) == 0 (without ptl)
1871 * // pmd is re-established
1873 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1874 * which may break userspace.
1876 * pmdp_invalidate() is required to make sure we don't miss
1877 * dirty/young flags set by hardware.
1879 entry = pmdp_invalidate(vma, addr, pmd);
1881 entry = pmd_modify(entry, newprot);
1883 entry = pmd_mk_savedwrite(entry);
1885 entry = pmd_wrprotect(entry);
1886 entry = pmd_mkuffd_wp(entry);
1887 } else if (uffd_wp_resolve) {
1889 * Leave the write bit to be handled by PF interrupt
1890 * handler, then things like COW could be properly
1893 entry = pmd_clear_uffd_wp(entry);
1896 set_pmd_at(mm, addr, pmd, entry);
1897 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1904 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1906 * Note that if it returns page table lock pointer, this routine returns without
1907 * unlocking page table lock. So callers must unlock it.
1909 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1912 ptl = pmd_lock(vma->vm_mm, pmd);
1913 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1921 * Returns true if a given pud maps a thp, false otherwise.
1923 * Note that if it returns true, this routine returns without unlocking page
1924 * table lock. So callers must unlock it.
1926 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1930 ptl = pud_lock(vma->vm_mm, pud);
1931 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1937 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1938 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1939 pud_t *pud, unsigned long addr)
1943 ptl = __pud_trans_huge_lock(pud, vma);
1947 * For architectures like ppc64 we look at deposited pgtable
1948 * when calling pudp_huge_get_and_clear. So do the
1949 * pgtable_trans_huge_withdraw after finishing pudp related
1952 pudp_huge_get_and_clear_full(tlb->mm, addr, pud, tlb->fullmm);
1953 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1954 if (vma_is_special_huge(vma)) {
1956 /* No zero page support yet */
1958 /* No support for anonymous PUD pages yet */
1964 static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1965 unsigned long haddr)
1967 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1968 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1969 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1970 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1972 count_vm_event(THP_SPLIT_PUD);
1974 pudp_huge_clear_flush_notify(vma, haddr, pud);
1977 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
1978 unsigned long address)
1981 struct mmu_notifier_range range;
1983 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1984 address & HPAGE_PUD_MASK,
1985 (address & HPAGE_PUD_MASK) + HPAGE_PUD_SIZE);
1986 mmu_notifier_invalidate_range_start(&range);
1987 ptl = pud_lock(vma->vm_mm, pud);
1988 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
1990 __split_huge_pud_locked(vma, pud, range.start);
1995 * No need to double call mmu_notifier->invalidate_range() callback as
1996 * the above pudp_huge_clear_flush_notify() did already call it.
1998 mmu_notifier_invalidate_range_only_end(&range);
2000 #endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2002 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2003 unsigned long haddr, pmd_t *pmd)
2005 struct mm_struct *mm = vma->vm_mm;
2011 * Leave pmd empty until pte is filled note that it is fine to delay
2012 * notification until mmu_notifier_invalidate_range_end() as we are
2013 * replacing a zero pmd write protected page with a zero pte write
2016 * See Documentation/vm/mmu_notifier.rst
2018 pmdp_huge_clear_flush(vma, haddr, pmd);
2020 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2021 pmd_populate(mm, &_pmd, pgtable);
2023 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2025 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2026 entry = pte_mkspecial(entry);
2027 pte = pte_offset_map(&_pmd, haddr);
2028 VM_BUG_ON(!pte_none(*pte));
2029 set_pte_at(mm, haddr, pte, entry);
2032 smp_wmb(); /* make pte visible before pmd */
2033 pmd_populate(mm, pmd, pgtable);
2036 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2037 unsigned long haddr, bool freeze)
2039 struct mm_struct *mm = vma->vm_mm;
2042 pmd_t old_pmd, _pmd;
2043 bool young, write, soft_dirty, pmd_migration = false, uffd_wp = false;
2047 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2048 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2049 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2050 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2051 && !pmd_devmap(*pmd));
2053 count_vm_event(THP_SPLIT_PMD);
2055 if (!vma_is_anonymous(vma)) {
2056 old_pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2058 * We are going to unmap this huge page. So
2059 * just go ahead and zap it
2061 if (arch_needs_pgtable_deposit())
2062 zap_deposited_table(mm, pmd);
2063 if (vma_is_special_huge(vma))
2065 if (unlikely(is_pmd_migration_entry(old_pmd))) {
2068 entry = pmd_to_swp_entry(old_pmd);
2069 page = migration_entry_to_page(entry);
2071 page = pmd_page(old_pmd);
2072 if (!PageDirty(page) && pmd_dirty(old_pmd))
2073 set_page_dirty(page);
2074 if (!PageReferenced(page) && pmd_young(old_pmd))
2075 SetPageReferenced(page);
2076 page_remove_rmap(page, true);
2079 add_mm_counter(mm, mm_counter_file(page), -HPAGE_PMD_NR);
2083 if (is_huge_zero_pmd(*pmd)) {
2085 * FIXME: Do we want to invalidate secondary mmu by calling
2086 * mmu_notifier_invalidate_range() see comments below inside
2087 * __split_huge_pmd() ?
2089 * We are going from a zero huge page write protected to zero
2090 * small page also write protected so it does not seems useful
2091 * to invalidate secondary mmu at this time.
2093 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2097 * Up to this point the pmd is present and huge and userland has the
2098 * whole access to the hugepage during the split (which happens in
2099 * place). If we overwrite the pmd with the not-huge version pointing
2100 * to the pte here (which of course we could if all CPUs were bug
2101 * free), userland could trigger a small page size TLB miss on the
2102 * small sized TLB while the hugepage TLB entry is still established in
2103 * the huge TLB. Some CPU doesn't like that.
2104 * See http://support.amd.com/TechDocs/41322_10h_Rev_Gd.pdf, Erratum
2105 * 383 on page 105. Intel should be safe but is also warns that it's
2106 * only safe if the permission and cache attributes of the two entries
2107 * loaded in the two TLB is identical (which should be the case here).
2108 * But it is generally safer to never allow small and huge TLB entries
2109 * for the same virtual address to be loaded simultaneously. So instead
2110 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2111 * current pmd notpresent (atomically because here the pmd_trans_huge
2112 * must remain set at all times on the pmd until the split is complete
2113 * for this pmd), then we flush the SMP TLB and finally we write the
2114 * non-huge version of the pmd entry with pmd_populate.
2116 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2118 pmd_migration = is_pmd_migration_entry(old_pmd);
2119 if (unlikely(pmd_migration)) {
2122 entry = pmd_to_swp_entry(old_pmd);
2123 page = migration_entry_to_page(entry);
2124 write = is_write_migration_entry(entry);
2126 soft_dirty = pmd_swp_soft_dirty(old_pmd);
2127 uffd_wp = pmd_swp_uffd_wp(old_pmd);
2129 page = pmd_page(old_pmd);
2130 if (pmd_dirty(old_pmd))
2132 write = pmd_write(old_pmd);
2133 young = pmd_young(old_pmd);
2134 soft_dirty = pmd_soft_dirty(old_pmd);
2135 uffd_wp = pmd_uffd_wp(old_pmd);
2137 VM_BUG_ON_PAGE(!page_count(page), page);
2138 page_ref_add(page, HPAGE_PMD_NR - 1);
2141 * Withdraw the table only after we mark the pmd entry invalid.
2142 * This's critical for some architectures (Power).
2144 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2145 pmd_populate(mm, &_pmd, pgtable);
2147 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2150 * Note that NUMA hinting access restrictions are not
2151 * transferred to avoid any possibility of altering
2152 * permissions across VMAs.
2154 if (freeze || pmd_migration) {
2155 swp_entry_t swp_entry;
2156 swp_entry = make_migration_entry(page + i, write);
2157 entry = swp_entry_to_pte(swp_entry);
2159 entry = pte_swp_mksoft_dirty(entry);
2161 entry = pte_swp_mkuffd_wp(entry);
2163 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2164 entry = maybe_mkwrite(entry, vma);
2166 entry = pte_wrprotect(entry);
2168 entry = pte_mkold(entry);
2170 entry = pte_mksoft_dirty(entry);
2172 entry = pte_mkuffd_wp(entry);
2174 pte = pte_offset_map(&_pmd, addr);
2175 BUG_ON(!pte_none(*pte));
2176 set_pte_at(mm, addr, pte, entry);
2178 atomic_inc(&page[i]._mapcount);
2182 if (!pmd_migration) {
2184 * Set PG_double_map before dropping compound_mapcount to avoid
2185 * false-negative page_mapped().
2187 if (compound_mapcount(page) > 1 &&
2188 !TestSetPageDoubleMap(page)) {
2189 for (i = 0; i < HPAGE_PMD_NR; i++)
2190 atomic_inc(&page[i]._mapcount);
2193 lock_page_memcg(page);
2194 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2195 /* Last compound_mapcount is gone. */
2196 __mod_lruvec_page_state(page, NR_ANON_THPS,
2198 if (TestClearPageDoubleMap(page)) {
2199 /* No need in mapcount reference anymore */
2200 for (i = 0; i < HPAGE_PMD_NR; i++)
2201 atomic_dec(&page[i]._mapcount);
2204 unlock_page_memcg(page);
2207 smp_wmb(); /* make pte visible before pmd */
2208 pmd_populate(mm, pmd, pgtable);
2211 for (i = 0; i < HPAGE_PMD_NR; i++) {
2212 page_remove_rmap(page + i, false);
2218 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2219 unsigned long address, bool freeze, struct page *page)
2222 struct mmu_notifier_range range;
2223 bool do_unlock_page = false;
2226 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
2227 address & HPAGE_PMD_MASK,
2228 (address & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE);
2229 mmu_notifier_invalidate_range_start(&range);
2230 ptl = pmd_lock(vma->vm_mm, pmd);
2233 * If caller asks to setup a migration entries, we need a page to check
2234 * pmd against. Otherwise we can end up replacing wrong page.
2236 VM_BUG_ON(freeze && !page);
2238 VM_WARN_ON_ONCE(!PageLocked(page));
2239 if (page != pmd_page(*pmd))
2244 if (pmd_trans_huge(*pmd)) {
2246 page = pmd_page(*pmd);
2248 * An anonymous page must be locked, to ensure that a
2249 * concurrent reuse_swap_page() sees stable mapcount;
2250 * but reuse_swap_page() is not used on shmem or file,
2251 * and page lock must not be taken when zap_pmd_range()
2252 * calls __split_huge_pmd() while i_mmap_lock is held.
2254 if (PageAnon(page)) {
2255 if (unlikely(!trylock_page(page))) {
2261 if (unlikely(!pmd_same(*pmd, _pmd))) {
2269 do_unlock_page = true;
2272 if (PageMlocked(page))
2273 clear_page_mlock(page);
2274 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2276 __split_huge_pmd_locked(vma, pmd, range.start, freeze);
2282 * No need to double call mmu_notifier->invalidate_range() callback.
2283 * They are 3 cases to consider inside __split_huge_pmd_locked():
2284 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2285 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2286 * fault will trigger a flush_notify before pointing to a new page
2287 * (it is fine if the secondary mmu keeps pointing to the old zero
2288 * page in the meantime)
2289 * 3) Split a huge pmd into pte pointing to the same page. No need
2290 * to invalidate secondary tlb entry they are all still valid.
2291 * any further changes to individual pte will notify. So no need
2292 * to call mmu_notifier->invalidate_range()
2294 mmu_notifier_invalidate_range_only_end(&range);
2297 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2298 bool freeze, struct page *page)
2305 pgd = pgd_offset(vma->vm_mm, address);
2306 if (!pgd_present(*pgd))
2309 p4d = p4d_offset(pgd, address);
2310 if (!p4d_present(*p4d))
2313 pud = pud_offset(p4d, address);
2314 if (!pud_present(*pud))
2317 pmd = pmd_offset(pud, address);
2319 __split_huge_pmd(vma, pmd, address, freeze, page);
2322 static inline void split_huge_pmd_if_needed(struct vm_area_struct *vma, unsigned long address)
2325 * If the new address isn't hpage aligned and it could previously
2326 * contain an hugepage: check if we need to split an huge pmd.
2328 if (!IS_ALIGNED(address, HPAGE_PMD_SIZE) &&
2329 range_in_vma(vma, ALIGN_DOWN(address, HPAGE_PMD_SIZE),
2330 ALIGN(address, HPAGE_PMD_SIZE)))
2331 split_huge_pmd_address(vma, address, false, NULL);
2334 void vma_adjust_trans_huge(struct vm_area_struct *vma,
2335 unsigned long start,
2339 /* Check if we need to split start first. */
2340 split_huge_pmd_if_needed(vma, start);
2342 /* Check if we need to split end next. */
2343 split_huge_pmd_if_needed(vma, end);
2346 * If we're also updating the vma->vm_next->vm_start,
2347 * check if we need to split it.
2349 if (adjust_next > 0) {
2350 struct vm_area_struct *next = vma->vm_next;
2351 unsigned long nstart = next->vm_start;
2352 nstart += adjust_next;
2353 split_huge_pmd_if_needed(next, nstart);
2357 static void unmap_page(struct page *page)
2359 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_SYNC |
2360 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2362 VM_BUG_ON_PAGE(!PageHead(page), page);
2365 ttu_flags |= TTU_SPLIT_FREEZE;
2367 try_to_unmap(page, ttu_flags);
2369 VM_WARN_ON_ONCE_PAGE(page_mapped(page), page);
2372 static void remap_page(struct page *page, unsigned int nr)
2375 if (PageTransHuge(page)) {
2376 remove_migration_ptes(page, page, true);
2378 for (i = 0; i < nr; i++)
2379 remove_migration_ptes(page + i, page + i, true);
2383 static void lru_add_page_tail(struct page *head, struct page *tail,
2384 struct lruvec *lruvec, struct list_head *list)
2386 VM_BUG_ON_PAGE(!PageHead(head), head);
2387 VM_BUG_ON_PAGE(PageCompound(tail), head);
2388 VM_BUG_ON_PAGE(PageLRU(tail), head);
2389 lockdep_assert_held(&lruvec->lru_lock);
2392 /* page reclaim is reclaiming a huge page */
2393 VM_WARN_ON(PageLRU(head));
2395 list_add_tail(&tail->lru, list);
2397 /* head is still on lru (and we have it frozen) */
2398 VM_WARN_ON(!PageLRU(head));
2400 list_add_tail(&tail->lru, &head->lru);
2404 static void __split_huge_page_tail(struct page *head, int tail,
2405 struct lruvec *lruvec, struct list_head *list)
2407 struct page *page_tail = head + tail;
2409 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2412 * Clone page flags before unfreezing refcount.
2414 * After successful get_page_unless_zero() might follow flags change,
2415 * for example lock_page() which set PG_waiters.
2417 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2418 page_tail->flags |= (head->flags &
2419 ((1L << PG_referenced) |
2420 (1L << PG_swapbacked) |
2421 (1L << PG_swapcache) |
2422 (1L << PG_mlocked) |
2423 (1L << PG_uptodate) |
2425 (1L << PG_workingset) |
2427 (1L << PG_unevictable) |
2433 /* ->mapping in first tail page is compound_mapcount */
2434 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2436 page_tail->mapping = head->mapping;
2437 page_tail->index = head->index + tail;
2439 /* Page flags must be visible before we make the page non-compound. */
2443 * Clear PageTail before unfreezing page refcount.
2445 * After successful get_page_unless_zero() might follow put_page()
2446 * which needs correct compound_head().
2448 clear_compound_head(page_tail);
2450 /* Finally unfreeze refcount. Additional reference from page cache. */
2451 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2452 PageSwapCache(head)));
2454 if (page_is_young(head))
2455 set_page_young(page_tail);
2456 if (page_is_idle(head))
2457 set_page_idle(page_tail);
2459 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2462 * always add to the tail because some iterators expect new
2463 * pages to show after the currently processed elements - e.g.
2466 lru_add_page_tail(head, page_tail, lruvec, list);
2469 static void __split_huge_page(struct page *page, struct list_head *list,
2472 struct page *head = compound_head(page);
2473 struct lruvec *lruvec;
2474 struct address_space *swap_cache = NULL;
2475 unsigned long offset = 0;
2476 unsigned int nr = thp_nr_pages(head);
2479 /* complete memcg works before add pages to LRU */
2480 split_page_memcg(head, nr);
2482 if (PageAnon(head) && PageSwapCache(head)) {
2483 swp_entry_t entry = { .val = page_private(head) };
2485 offset = swp_offset(entry);
2486 swap_cache = swap_address_space(entry);
2487 xa_lock(&swap_cache->i_pages);
2490 /* lock lru list/PageCompound, ref frozen by page_ref_freeze */
2491 lruvec = lock_page_lruvec(head);
2493 for (i = nr - 1; i >= 1; i--) {
2494 __split_huge_page_tail(head, i, lruvec, list);
2495 /* Some pages can be beyond i_size: drop them from page cache */
2496 if (head[i].index >= end) {
2497 ClearPageDirty(head + i);
2498 __delete_from_page_cache(head + i, NULL);
2499 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2500 shmem_uncharge(head->mapping->host, 1);
2502 } else if (!PageAnon(page)) {
2503 __xa_store(&head->mapping->i_pages, head[i].index,
2505 } else if (swap_cache) {
2506 __xa_store(&swap_cache->i_pages, offset + i,
2511 ClearPageCompound(head);
2512 unlock_page_lruvec(lruvec);
2513 /* Caller disabled irqs, so they are still disabled here */
2515 split_page_owner(head, nr);
2517 /* See comment in __split_huge_page_tail() */
2518 if (PageAnon(head)) {
2519 /* Additional pin to swap cache */
2520 if (PageSwapCache(head)) {
2521 page_ref_add(head, 2);
2522 xa_unlock(&swap_cache->i_pages);
2527 /* Additional pin to page cache */
2528 page_ref_add(head, 2);
2529 xa_unlock(&head->mapping->i_pages);
2533 remap_page(head, nr);
2535 if (PageSwapCache(head)) {
2536 swp_entry_t entry = { .val = page_private(head) };
2538 split_swap_cluster(entry);
2541 for (i = 0; i < nr; i++) {
2542 struct page *subpage = head + i;
2543 if (subpage == page)
2545 unlock_page(subpage);
2548 * Subpages may be freed if there wasn't any mapping
2549 * like if add_to_swap() is running on a lru page that
2550 * had its mapping zapped. And freeing these pages
2551 * requires taking the lru_lock so we do the put_page
2552 * of the tail pages after the split is complete.
2558 int total_mapcount(struct page *page)
2560 int i, compound, nr, ret;
2562 VM_BUG_ON_PAGE(PageTail(page), page);
2564 if (likely(!PageCompound(page)))
2565 return atomic_read(&page->_mapcount) + 1;
2567 compound = compound_mapcount(page);
2568 nr = compound_nr(page);
2572 for (i = 0; i < nr; i++)
2573 ret += atomic_read(&page[i]._mapcount) + 1;
2574 /* File pages has compound_mapcount included in _mapcount */
2575 if (!PageAnon(page))
2576 return ret - compound * nr;
2577 if (PageDoubleMap(page))
2583 * This calculates accurately how many mappings a transparent hugepage
2584 * has (unlike page_mapcount() which isn't fully accurate). This full
2585 * accuracy is primarily needed to know if copy-on-write faults can
2586 * reuse the page and change the mapping to read-write instead of
2587 * copying them. At the same time this returns the total_mapcount too.
2589 * The function returns the highest mapcount any one of the subpages
2590 * has. If the return value is one, even if different processes are
2591 * mapping different subpages of the transparent hugepage, they can
2592 * all reuse it, because each process is reusing a different subpage.
2594 * The total_mapcount is instead counting all virtual mappings of the
2595 * subpages. If the total_mapcount is equal to "one", it tells the
2596 * caller all mappings belong to the same "mm" and in turn the
2597 * anon_vma of the transparent hugepage can become the vma->anon_vma
2598 * local one as no other process may be mapping any of the subpages.
2600 * It would be more accurate to replace page_mapcount() with
2601 * page_trans_huge_mapcount(), however we only use
2602 * page_trans_huge_mapcount() in the copy-on-write faults where we
2603 * need full accuracy to avoid breaking page pinning, because
2604 * page_trans_huge_mapcount() is slower than page_mapcount().
2606 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2608 int i, ret, _total_mapcount, mapcount;
2610 /* hugetlbfs shouldn't call it */
2611 VM_BUG_ON_PAGE(PageHuge(page), page);
2613 if (likely(!PageTransCompound(page))) {
2614 mapcount = atomic_read(&page->_mapcount) + 1;
2616 *total_mapcount = mapcount;
2620 page = compound_head(page);
2622 _total_mapcount = ret = 0;
2623 for (i = 0; i < thp_nr_pages(page); i++) {
2624 mapcount = atomic_read(&page[i]._mapcount) + 1;
2625 ret = max(ret, mapcount);
2626 _total_mapcount += mapcount;
2628 if (PageDoubleMap(page)) {
2630 _total_mapcount -= thp_nr_pages(page);
2632 mapcount = compound_mapcount(page);
2634 _total_mapcount += mapcount;
2636 *total_mapcount = _total_mapcount;
2640 /* Racy check whether the huge page can be split */
2641 bool can_split_huge_page(struct page *page, int *pextra_pins)
2645 /* Additional pins from page cache */
2647 extra_pins = PageSwapCache(page) ? thp_nr_pages(page) : 0;
2649 extra_pins = thp_nr_pages(page);
2651 *pextra_pins = extra_pins;
2652 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2656 * This function splits huge page into normal pages. @page can point to any
2657 * subpage of huge page to split. Split doesn't change the position of @page.
2659 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2660 * The huge page must be locked.
2662 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2664 * Both head page and tail pages will inherit mapping, flags, and so on from
2667 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2668 * they are not mapped.
2670 * Returns 0 if the hugepage is split successfully.
2671 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2674 int split_huge_page_to_list(struct page *page, struct list_head *list)
2676 struct page *head = compound_head(page);
2677 struct deferred_split *ds_queue = get_deferred_split_queue(head);
2678 struct anon_vma *anon_vma = NULL;
2679 struct address_space *mapping = NULL;
2680 int extra_pins, ret;
2683 VM_BUG_ON_PAGE(is_huge_zero_page(head), head);
2684 VM_BUG_ON_PAGE(!PageLocked(head), head);
2685 VM_BUG_ON_PAGE(!PageCompound(head), head);
2687 if (PageWriteback(head))
2690 if (PageAnon(head)) {
2692 * The caller does not necessarily hold an mmap_lock that would
2693 * prevent the anon_vma disappearing so we first we take a
2694 * reference to it and then lock the anon_vma for write. This
2695 * is similar to page_lock_anon_vma_read except the write lock
2696 * is taken to serialise against parallel split or collapse
2699 anon_vma = page_get_anon_vma(head);
2706 anon_vma_lock_write(anon_vma);
2708 mapping = head->mapping;
2717 i_mmap_lock_read(mapping);
2720 *__split_huge_page() may need to trim off pages beyond EOF:
2721 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2722 * which cannot be nested inside the page tree lock. So note
2723 * end now: i_size itself may be changed at any moment, but
2724 * head page lock is good enough to serialize the trimming.
2726 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2730 * Racy check if we can split the page, before unmap_page() will
2733 if (!can_split_huge_page(head, &extra_pins)) {
2740 /* block interrupt reentry in xa_lock and spinlock */
2741 local_irq_disable();
2743 XA_STATE(xas, &mapping->i_pages, page_index(head));
2746 * Check if the head page is present in page cache.
2747 * We assume all tail are present too, if head is there.
2749 xa_lock(&mapping->i_pages);
2750 if (xas_load(&xas) != head)
2754 /* Prevent deferred_split_scan() touching ->_refcount */
2755 spin_lock(&ds_queue->split_queue_lock);
2756 if (page_ref_freeze(head, 1 + extra_pins)) {
2757 if (!list_empty(page_deferred_list(head))) {
2758 ds_queue->split_queue_len--;
2759 list_del(page_deferred_list(head));
2761 spin_unlock(&ds_queue->split_queue_lock);
2763 int nr = thp_nr_pages(head);
2765 if (PageSwapBacked(head))
2766 __mod_lruvec_page_state(head, NR_SHMEM_THPS,
2769 __mod_lruvec_page_state(head, NR_FILE_THPS,
2773 __split_huge_page(page, list, end);
2776 spin_unlock(&ds_queue->split_queue_lock);
2779 xa_unlock(&mapping->i_pages);
2781 remap_page(head, thp_nr_pages(head));
2787 anon_vma_unlock_write(anon_vma);
2788 put_anon_vma(anon_vma);
2791 i_mmap_unlock_read(mapping);
2793 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2797 void free_transhuge_page(struct page *page)
2799 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2800 unsigned long flags;
2802 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2803 if (!list_empty(page_deferred_list(page))) {
2804 ds_queue->split_queue_len--;
2805 list_del(page_deferred_list(page));
2807 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2808 free_compound_page(page);
2811 void deferred_split_huge_page(struct page *page)
2813 struct deferred_split *ds_queue = get_deferred_split_queue(page);
2815 struct mem_cgroup *memcg = page_memcg(compound_head(page));
2817 unsigned long flags;
2819 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2822 * The try_to_unmap() in page reclaim path might reach here too,
2823 * this may cause a race condition to corrupt deferred split queue.
2824 * And, if page reclaim is already handling the same page, it is
2825 * unnecessary to handle it again in shrinker.
2827 * Check PageSwapCache to determine if the page is being
2828 * handled by page reclaim since THP swap would add the page into
2829 * swap cache before calling try_to_unmap().
2831 if (PageSwapCache(page))
2834 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2835 if (list_empty(page_deferred_list(page))) {
2836 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2837 list_add_tail(page_deferred_list(page), &ds_queue->split_queue);
2838 ds_queue->split_queue_len++;
2841 set_shrinker_bit(memcg, page_to_nid(page),
2842 deferred_split_shrinker.id);
2845 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2848 static unsigned long deferred_split_count(struct shrinker *shrink,
2849 struct shrink_control *sc)
2851 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2852 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2856 ds_queue = &sc->memcg->deferred_split_queue;
2858 return READ_ONCE(ds_queue->split_queue_len);
2861 static unsigned long deferred_split_scan(struct shrinker *shrink,
2862 struct shrink_control *sc)
2864 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2865 struct deferred_split *ds_queue = &pgdata->deferred_split_queue;
2866 unsigned long flags;
2867 LIST_HEAD(list), *pos, *next;
2873 ds_queue = &sc->memcg->deferred_split_queue;
2876 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2877 /* Take pin on all head pages to avoid freeing them under us */
2878 list_for_each_safe(pos, next, &ds_queue->split_queue) {
2879 page = list_entry((void *)pos, struct page, deferred_list);
2880 page = compound_head(page);
2881 if (get_page_unless_zero(page)) {
2882 list_move(page_deferred_list(page), &list);
2884 /* We lost race with put_compound_page() */
2885 list_del_init(page_deferred_list(page));
2886 ds_queue->split_queue_len--;
2888 if (!--sc->nr_to_scan)
2891 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2893 list_for_each_safe(pos, next, &list) {
2894 page = list_entry((void *)pos, struct page, deferred_list);
2895 if (!trylock_page(page))
2897 /* split_huge_page() removes page from list on success */
2898 if (!split_huge_page(page))
2905 spin_lock_irqsave(&ds_queue->split_queue_lock, flags);
2906 list_splice_tail(&list, &ds_queue->split_queue);
2907 spin_unlock_irqrestore(&ds_queue->split_queue_lock, flags);
2910 * Stop shrinker if we didn't split any page, but the queue is empty.
2911 * This can happen if pages were freed under us.
2913 if (!split && list_empty(&ds_queue->split_queue))
2918 static struct shrinker deferred_split_shrinker = {
2919 .count_objects = deferred_split_count,
2920 .scan_objects = deferred_split_scan,
2921 .seeks = DEFAULT_SEEKS,
2922 .flags = SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE |
2926 #ifdef CONFIG_DEBUG_FS
2927 static void split_huge_pages_all(void)
2931 unsigned long pfn, max_zone_pfn;
2932 unsigned long total = 0, split = 0;
2934 pr_debug("Split all THPs\n");
2935 for_each_populated_zone(zone) {
2936 max_zone_pfn = zone_end_pfn(zone);
2937 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2938 if (!pfn_valid(pfn))
2941 page = pfn_to_page(pfn);
2942 if (!get_page_unless_zero(page))
2945 if (zone != page_zone(page))
2948 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2953 if (!split_huge_page(page))
2962 pr_debug("%lu of %lu THP split\n", split, total);
2965 static inline bool vma_not_suitable_for_thp_split(struct vm_area_struct *vma)
2967 return vma_is_special_huge(vma) || (vma->vm_flags & VM_IO) ||
2968 is_vm_hugetlb_page(vma);
2971 static int split_huge_pages_pid(int pid, unsigned long vaddr_start,
2972 unsigned long vaddr_end)
2975 struct task_struct *task;
2976 struct mm_struct *mm;
2977 unsigned long total = 0, split = 0;
2980 vaddr_start &= PAGE_MASK;
2981 vaddr_end &= PAGE_MASK;
2983 /* Find the task_struct from pid */
2985 task = find_task_by_vpid(pid);
2991 get_task_struct(task);
2994 /* Find the mm_struct */
2995 mm = get_task_mm(task);
2996 put_task_struct(task);
3003 pr_debug("Split huge pages in pid: %d, vaddr: [0x%lx - 0x%lx]\n",
3004 pid, vaddr_start, vaddr_end);
3008 * always increase addr by PAGE_SIZE, since we could have a PTE page
3009 * table filled with PTE-mapped THPs, each of which is distinct.
3011 for (addr = vaddr_start; addr < vaddr_end; addr += PAGE_SIZE) {
3012 struct vm_area_struct *vma = find_vma(mm, addr);
3013 unsigned int follflags;
3016 if (!vma || addr < vma->vm_start)
3019 /* skip special VMA and hugetlb VMA */
3020 if (vma_not_suitable_for_thp_split(vma)) {
3025 /* FOLL_DUMP to ignore special (like zero) pages */
3026 follflags = FOLL_GET | FOLL_DUMP;
3027 page = follow_page(vma, addr, follflags);
3034 if (!is_transparent_hugepage(page))
3038 if (!can_split_huge_page(compound_head(page), NULL))
3041 if (!trylock_page(page))
3044 if (!split_huge_page(page))
3052 mmap_read_unlock(mm);
3055 pr_debug("%lu of %lu THP split\n", split, total);
3061 static int split_huge_pages_in_file(const char *file_path, pgoff_t off_start,
3064 struct filename *file;
3065 struct file *candidate;
3066 struct address_space *mapping;
3070 unsigned long total = 0, split = 0;
3072 file = getname_kernel(file_path);
3076 candidate = file_open_name(file, O_RDONLY, 0);
3077 if (IS_ERR(candidate))
3080 pr_debug("split file-backed THPs in file: %s, page offset: [0x%lx - 0x%lx]\n",
3081 file_path, off_start, off_end);
3083 mapping = candidate->f_mapping;
3085 for (index = off_start; index < off_end; index += nr_pages) {
3086 struct page *fpage = pagecache_get_page(mapping, index,
3087 FGP_ENTRY | FGP_HEAD, 0);
3090 if (xa_is_value(fpage) || !fpage)
3093 if (!is_transparent_hugepage(fpage))
3097 nr_pages = thp_nr_pages(fpage);
3099 if (!trylock_page(fpage))
3102 if (!split_huge_page(fpage))
3111 filp_close(candidate, NULL);
3114 pr_debug("%lu of %lu file-backed THP split\n", split, total);
3120 #define MAX_INPUT_BUF_SZ 255
3122 static ssize_t split_huge_pages_write(struct file *file, const char __user *buf,
3123 size_t count, loff_t *ppops)
3125 static DEFINE_MUTEX(split_debug_mutex);
3127 /* hold pid, start_vaddr, end_vaddr or file_path, off_start, off_end */
3128 char input_buf[MAX_INPUT_BUF_SZ];
3130 unsigned long vaddr_start, vaddr_end;
3132 ret = mutex_lock_interruptible(&split_debug_mutex);
3138 memset(input_buf, 0, MAX_INPUT_BUF_SZ);
3139 if (copy_from_user(input_buf, buf, min_t(size_t, count, MAX_INPUT_BUF_SZ)))
3142 input_buf[MAX_INPUT_BUF_SZ - 1] = '\0';
3144 if (input_buf[0] == '/') {
3146 char *buf = input_buf;
3147 char file_path[MAX_INPUT_BUF_SZ];
3148 pgoff_t off_start = 0, off_end = 0;
3149 size_t input_len = strlen(input_buf);
3151 tok = strsep(&buf, ",");
3153 strncpy(file_path, tok, MAX_INPUT_BUF_SZ);
3159 ret = sscanf(buf, "0x%lx,0x%lx", &off_start, &off_end);
3164 ret = split_huge_pages_in_file(file_path, off_start, off_end);
3171 ret = sscanf(input_buf, "%d,0x%lx,0x%lx", &pid, &vaddr_start, &vaddr_end);
3172 if (ret == 1 && pid == 1) {
3173 split_huge_pages_all();
3174 ret = strlen(input_buf);
3176 } else if (ret != 3) {
3181 ret = split_huge_pages_pid(pid, vaddr_start, vaddr_end);
3183 ret = strlen(input_buf);
3185 mutex_unlock(&split_debug_mutex);
3190 static const struct file_operations split_huge_pages_fops = {
3191 .owner = THIS_MODULE,
3192 .write = split_huge_pages_write,
3193 .llseek = no_llseek,
3196 static int __init split_huge_pages_debugfs(void)
3198 debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3199 &split_huge_pages_fops);
3202 late_initcall(split_huge_pages_debugfs);
3205 #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
3206 void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
3209 struct vm_area_struct *vma = pvmw->vma;
3210 struct mm_struct *mm = vma->vm_mm;
3211 unsigned long address = pvmw->address;
3216 if (!(pvmw->pmd && !pvmw->pte))
3219 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
3220 pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
3221 if (pmd_dirty(pmdval))
3222 set_page_dirty(page);
3223 entry = make_migration_entry(page, pmd_write(pmdval));
3224 pmdswp = swp_entry_to_pmd(entry);
3225 if (pmd_soft_dirty(pmdval))
3226 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
3227 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
3228 page_remove_rmap(page, true);
3232 void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
3234 struct vm_area_struct *vma = pvmw->vma;
3235 struct mm_struct *mm = vma->vm_mm;
3236 unsigned long address = pvmw->address;
3237 unsigned long mmun_start = address & HPAGE_PMD_MASK;
3241 if (!(pvmw->pmd && !pvmw->pte))
3244 entry = pmd_to_swp_entry(*pvmw->pmd);
3246 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
3247 if (pmd_swp_soft_dirty(*pvmw->pmd))
3248 pmde = pmd_mksoft_dirty(pmde);
3249 if (is_write_migration_entry(entry))
3250 pmde = maybe_pmd_mkwrite(pmde, vma);
3252 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
3254 page_add_anon_rmap(new, vma, mmun_start, true);
3256 page_add_file_rmap(new, true);
3257 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
3258 if ((vma->vm_flags & VM_LOCKED) && !PageDoubleMap(new))
3259 mlock_vma_page(new);
3260 update_mmu_cache_pmd(vma, address, pvmw->pmd);