1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
36 #include <asm/pgalloc.h>
40 #include <linux/hugetlb.h>
41 #include <linux/hugetlb_cgroup.h>
42 #include <linux/node.h>
43 #include <linux/page_owner.h>
45 #include "hugetlb_vmemmap.h"
47 int hugetlb_max_hstate __read_mostly;
48 unsigned int default_hstate_idx;
49 struct hstate hstates[HUGE_MAX_HSTATE];
52 static struct cma *hugetlb_cma[MAX_NUMNODES];
54 static unsigned long hugetlb_cma_size __initdata;
57 * Minimum page order among possible hugepage sizes, set to a proper value
60 static unsigned int minimum_order __read_mostly = UINT_MAX;
62 __initdata LIST_HEAD(huge_boot_pages);
64 /* for command line parsing */
65 static struct hstate * __initdata parsed_hstate;
66 static unsigned long __initdata default_hstate_max_huge_pages;
67 static bool __initdata parsed_valid_hugepagesz = true;
68 static bool __initdata parsed_default_hugepagesz;
71 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
72 * free_huge_pages, and surplus_huge_pages.
74 DEFINE_SPINLOCK(hugetlb_lock);
77 * Serializes faults on the same logical page. This is used to
78 * prevent spurious OOMs when the hugepage pool is fully utilized.
80 static int num_fault_mutexes;
81 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
83 /* Forward declaration */
84 static int hugetlb_acct_memory(struct hstate *h, long delta);
86 static inline bool subpool_is_free(struct hugepage_subpool *spool)
90 if (spool->max_hpages != -1)
91 return spool->used_hpages == 0;
92 if (spool->min_hpages != -1)
93 return spool->rsv_hpages == spool->min_hpages;
98 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
99 unsigned long irq_flags)
101 spin_unlock_irqrestore(&spool->lock, irq_flags);
103 /* If no pages are used, and no other handles to the subpool
104 * remain, give up any reservations based on minimum size and
105 * free the subpool */
106 if (subpool_is_free(spool)) {
107 if (spool->min_hpages != -1)
108 hugetlb_acct_memory(spool->hstate,
114 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
117 struct hugepage_subpool *spool;
119 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
123 spin_lock_init(&spool->lock);
125 spool->max_hpages = max_hpages;
127 spool->min_hpages = min_hpages;
129 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
133 spool->rsv_hpages = min_hpages;
138 void hugepage_put_subpool(struct hugepage_subpool *spool)
142 spin_lock_irqsave(&spool->lock, flags);
143 BUG_ON(!spool->count);
145 unlock_or_release_subpool(spool, flags);
149 * Subpool accounting for allocating and reserving pages.
150 * Return -ENOMEM if there are not enough resources to satisfy the
151 * request. Otherwise, return the number of pages by which the
152 * global pools must be adjusted (upward). The returned value may
153 * only be different than the passed value (delta) in the case where
154 * a subpool minimum size must be maintained.
156 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
164 spin_lock_irq(&spool->lock);
166 if (spool->max_hpages != -1) { /* maximum size accounting */
167 if ((spool->used_hpages + delta) <= spool->max_hpages)
168 spool->used_hpages += delta;
175 /* minimum size accounting */
176 if (spool->min_hpages != -1 && spool->rsv_hpages) {
177 if (delta > spool->rsv_hpages) {
179 * Asking for more reserves than those already taken on
180 * behalf of subpool. Return difference.
182 ret = delta - spool->rsv_hpages;
183 spool->rsv_hpages = 0;
185 ret = 0; /* reserves already accounted for */
186 spool->rsv_hpages -= delta;
191 spin_unlock_irq(&spool->lock);
196 * Subpool accounting for freeing and unreserving pages.
197 * Return the number of global page reservations that must be dropped.
198 * The return value may only be different than the passed value (delta)
199 * in the case where a subpool minimum size must be maintained.
201 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
210 spin_lock_irqsave(&spool->lock, flags);
212 if (spool->max_hpages != -1) /* maximum size accounting */
213 spool->used_hpages -= delta;
215 /* minimum size accounting */
216 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
217 if (spool->rsv_hpages + delta <= spool->min_hpages)
220 ret = spool->rsv_hpages + delta - spool->min_hpages;
222 spool->rsv_hpages += delta;
223 if (spool->rsv_hpages > spool->min_hpages)
224 spool->rsv_hpages = spool->min_hpages;
228 * If hugetlbfs_put_super couldn't free spool due to an outstanding
229 * quota reference, free it now.
231 unlock_or_release_subpool(spool, flags);
236 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
238 return HUGETLBFS_SB(inode->i_sb)->spool;
241 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
243 return subpool_inode(file_inode(vma->vm_file));
246 /* Helper that removes a struct file_region from the resv_map cache and returns
249 static struct file_region *
250 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
252 struct file_region *nrg = NULL;
254 VM_BUG_ON(resv->region_cache_count <= 0);
256 resv->region_cache_count--;
257 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
258 list_del(&nrg->link);
266 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
267 struct file_region *rg)
269 #ifdef CONFIG_CGROUP_HUGETLB
270 nrg->reservation_counter = rg->reservation_counter;
277 /* Helper that records hugetlb_cgroup uncharge info. */
278 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
280 struct resv_map *resv,
281 struct file_region *nrg)
283 #ifdef CONFIG_CGROUP_HUGETLB
285 nrg->reservation_counter =
286 &h_cg->rsvd_hugepage[hstate_index(h)];
287 nrg->css = &h_cg->css;
289 * The caller will hold exactly one h_cg->css reference for the
290 * whole contiguous reservation region. But this area might be
291 * scattered when there are already some file_regions reside in
292 * it. As a result, many file_regions may share only one css
293 * reference. In order to ensure that one file_region must hold
294 * exactly one h_cg->css reference, we should do css_get for
295 * each file_region and leave the reference held by caller
299 if (!resv->pages_per_hpage)
300 resv->pages_per_hpage = pages_per_huge_page(h);
301 /* pages_per_hpage should be the same for all entries in
304 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
306 nrg->reservation_counter = NULL;
312 static void put_uncharge_info(struct file_region *rg)
314 #ifdef CONFIG_CGROUP_HUGETLB
320 static bool has_same_uncharge_info(struct file_region *rg,
321 struct file_region *org)
323 #ifdef CONFIG_CGROUP_HUGETLB
325 rg->reservation_counter == org->reservation_counter &&
333 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
335 struct file_region *nrg = NULL, *prg = NULL;
337 prg = list_prev_entry(rg, link);
338 if (&prg->link != &resv->regions && prg->to == rg->from &&
339 has_same_uncharge_info(prg, rg)) {
343 put_uncharge_info(rg);
349 nrg = list_next_entry(rg, link);
350 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
351 has_same_uncharge_info(nrg, rg)) {
352 nrg->from = rg->from;
355 put_uncharge_info(rg);
361 hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
362 long to, struct hstate *h, struct hugetlb_cgroup *cg,
363 long *regions_needed)
365 struct file_region *nrg;
367 if (!regions_needed) {
368 nrg = get_file_region_entry_from_cache(map, from, to);
369 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
370 list_add(&nrg->link, rg->link.prev);
371 coalesce_file_region(map, nrg);
373 *regions_needed += 1;
379 * Must be called with resv->lock held.
381 * Calling this with regions_needed != NULL will count the number of pages
382 * to be added but will not modify the linked list. And regions_needed will
383 * indicate the number of file_regions needed in the cache to carry out to add
384 * the regions for this range.
386 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
387 struct hugetlb_cgroup *h_cg,
388 struct hstate *h, long *regions_needed)
391 struct list_head *head = &resv->regions;
392 long last_accounted_offset = f;
393 struct file_region *rg = NULL, *trg = NULL;
398 /* In this loop, we essentially handle an entry for the range
399 * [last_accounted_offset, rg->from), at every iteration, with some
402 list_for_each_entry_safe(rg, trg, head, link) {
403 /* Skip irrelevant regions that start before our range. */
405 /* If this region ends after the last accounted offset,
406 * then we need to update last_accounted_offset.
408 if (rg->to > last_accounted_offset)
409 last_accounted_offset = rg->to;
413 /* When we find a region that starts beyond our range, we've
419 /* Add an entry for last_accounted_offset -> rg->from, and
420 * update last_accounted_offset.
422 if (rg->from > last_accounted_offset)
423 add += hugetlb_resv_map_add(resv, rg,
424 last_accounted_offset,
428 last_accounted_offset = rg->to;
431 /* Handle the case where our range extends beyond
432 * last_accounted_offset.
434 if (last_accounted_offset < t)
435 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
436 t, h, h_cg, regions_needed);
442 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
444 static int allocate_file_region_entries(struct resv_map *resv,
446 __must_hold(&resv->lock)
448 struct list_head allocated_regions;
449 int to_allocate = 0, i = 0;
450 struct file_region *trg = NULL, *rg = NULL;
452 VM_BUG_ON(regions_needed < 0);
454 INIT_LIST_HEAD(&allocated_regions);
457 * Check for sufficient descriptors in the cache to accommodate
458 * the number of in progress add operations plus regions_needed.
460 * This is a while loop because when we drop the lock, some other call
461 * to region_add or region_del may have consumed some region_entries,
462 * so we keep looping here until we finally have enough entries for
463 * (adds_in_progress + regions_needed).
465 while (resv->region_cache_count <
466 (resv->adds_in_progress + regions_needed)) {
467 to_allocate = resv->adds_in_progress + regions_needed -
468 resv->region_cache_count;
470 /* At this point, we should have enough entries in the cache
471 * for all the existing adds_in_progress. We should only be
472 * needing to allocate for regions_needed.
474 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
476 spin_unlock(&resv->lock);
477 for (i = 0; i < to_allocate; i++) {
478 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
481 list_add(&trg->link, &allocated_regions);
484 spin_lock(&resv->lock);
486 list_splice(&allocated_regions, &resv->region_cache);
487 resv->region_cache_count += to_allocate;
493 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
501 * Add the huge page range represented by [f, t) to the reserve
502 * map. Regions will be taken from the cache to fill in this range.
503 * Sufficient regions should exist in the cache due to the previous
504 * call to region_chg with the same range, but in some cases the cache will not
505 * have sufficient entries due to races with other code doing region_add or
506 * region_del. The extra needed entries will be allocated.
508 * regions_needed is the out value provided by a previous call to region_chg.
510 * Return the number of new huge pages added to the map. This number is greater
511 * than or equal to zero. If file_region entries needed to be allocated for
512 * this operation and we were not able to allocate, it returns -ENOMEM.
513 * region_add of regions of length 1 never allocate file_regions and cannot
514 * fail; region_chg will always allocate at least 1 entry and a region_add for
515 * 1 page will only require at most 1 entry.
517 static long region_add(struct resv_map *resv, long f, long t,
518 long in_regions_needed, struct hstate *h,
519 struct hugetlb_cgroup *h_cg)
521 long add = 0, actual_regions_needed = 0;
523 spin_lock(&resv->lock);
526 /* Count how many regions are actually needed to execute this add. */
527 add_reservation_in_range(resv, f, t, NULL, NULL,
528 &actual_regions_needed);
531 * Check for sufficient descriptors in the cache to accommodate
532 * this add operation. Note that actual_regions_needed may be greater
533 * than in_regions_needed, as the resv_map may have been modified since
534 * the region_chg call. In this case, we need to make sure that we
535 * allocate extra entries, such that we have enough for all the
536 * existing adds_in_progress, plus the excess needed for this
539 if (actual_regions_needed > in_regions_needed &&
540 resv->region_cache_count <
541 resv->adds_in_progress +
542 (actual_regions_needed - in_regions_needed)) {
543 /* region_add operation of range 1 should never need to
544 * allocate file_region entries.
546 VM_BUG_ON(t - f <= 1);
548 if (allocate_file_region_entries(
549 resv, actual_regions_needed - in_regions_needed)) {
556 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
558 resv->adds_in_progress -= in_regions_needed;
560 spin_unlock(&resv->lock);
565 * Examine the existing reserve map and determine how many
566 * huge pages in the specified range [f, t) are NOT currently
567 * represented. This routine is called before a subsequent
568 * call to region_add that will actually modify the reserve
569 * map to add the specified range [f, t). region_chg does
570 * not change the number of huge pages represented by the
571 * map. A number of new file_region structures is added to the cache as a
572 * placeholder, for the subsequent region_add call to use. At least 1
573 * file_region structure is added.
575 * out_regions_needed is the number of regions added to the
576 * resv->adds_in_progress. This value needs to be provided to a follow up call
577 * to region_add or region_abort for proper accounting.
579 * Returns the number of huge pages that need to be added to the existing
580 * reservation map for the range [f, t). This number is greater or equal to
581 * zero. -ENOMEM is returned if a new file_region structure or cache entry
582 * is needed and can not be allocated.
584 static long region_chg(struct resv_map *resv, long f, long t,
585 long *out_regions_needed)
589 spin_lock(&resv->lock);
591 /* Count how many hugepages in this range are NOT represented. */
592 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
595 if (*out_regions_needed == 0)
596 *out_regions_needed = 1;
598 if (allocate_file_region_entries(resv, *out_regions_needed))
601 resv->adds_in_progress += *out_regions_needed;
603 spin_unlock(&resv->lock);
608 * Abort the in progress add operation. The adds_in_progress field
609 * of the resv_map keeps track of the operations in progress between
610 * calls to region_chg and region_add. Operations are sometimes
611 * aborted after the call to region_chg. In such cases, region_abort
612 * is called to decrement the adds_in_progress counter. regions_needed
613 * is the value returned by the region_chg call, it is used to decrement
614 * the adds_in_progress counter.
616 * NOTE: The range arguments [f, t) are not needed or used in this
617 * routine. They are kept to make reading the calling code easier as
618 * arguments will match the associated region_chg call.
620 static void region_abort(struct resv_map *resv, long f, long t,
623 spin_lock(&resv->lock);
624 VM_BUG_ON(!resv->region_cache_count);
625 resv->adds_in_progress -= regions_needed;
626 spin_unlock(&resv->lock);
630 * Delete the specified range [f, t) from the reserve map. If the
631 * t parameter is LONG_MAX, this indicates that ALL regions after f
632 * should be deleted. Locate the regions which intersect [f, t)
633 * and either trim, delete or split the existing regions.
635 * Returns the number of huge pages deleted from the reserve map.
636 * In the normal case, the return value is zero or more. In the
637 * case where a region must be split, a new region descriptor must
638 * be allocated. If the allocation fails, -ENOMEM will be returned.
639 * NOTE: If the parameter t == LONG_MAX, then we will never split
640 * a region and possibly return -ENOMEM. Callers specifying
641 * t == LONG_MAX do not need to check for -ENOMEM error.
643 static long region_del(struct resv_map *resv, long f, long t)
645 struct list_head *head = &resv->regions;
646 struct file_region *rg, *trg;
647 struct file_region *nrg = NULL;
651 spin_lock(&resv->lock);
652 list_for_each_entry_safe(rg, trg, head, link) {
654 * Skip regions before the range to be deleted. file_region
655 * ranges are normally of the form [from, to). However, there
656 * may be a "placeholder" entry in the map which is of the form
657 * (from, to) with from == to. Check for placeholder entries
658 * at the beginning of the range to be deleted.
660 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
666 if (f > rg->from && t < rg->to) { /* Must split region */
668 * Check for an entry in the cache before dropping
669 * lock and attempting allocation.
672 resv->region_cache_count > resv->adds_in_progress) {
673 nrg = list_first_entry(&resv->region_cache,
676 list_del(&nrg->link);
677 resv->region_cache_count--;
681 spin_unlock(&resv->lock);
682 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
689 hugetlb_cgroup_uncharge_file_region(
690 resv, rg, t - f, false);
692 /* New entry for end of split region */
696 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
698 INIT_LIST_HEAD(&nrg->link);
700 /* Original entry is trimmed */
703 list_add(&nrg->link, &rg->link);
708 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
709 del += rg->to - rg->from;
710 hugetlb_cgroup_uncharge_file_region(resv, rg,
711 rg->to - rg->from, true);
717 if (f <= rg->from) { /* Trim beginning of region */
718 hugetlb_cgroup_uncharge_file_region(resv, rg,
719 t - rg->from, false);
723 } else { /* Trim end of region */
724 hugetlb_cgroup_uncharge_file_region(resv, rg,
732 spin_unlock(&resv->lock);
738 * A rare out of memory error was encountered which prevented removal of
739 * the reserve map region for a page. The huge page itself was free'ed
740 * and removed from the page cache. This routine will adjust the subpool
741 * usage count, and the global reserve count if needed. By incrementing
742 * these counts, the reserve map entry which could not be deleted will
743 * appear as a "reserved" entry instead of simply dangling with incorrect
746 void hugetlb_fix_reserve_counts(struct inode *inode)
748 struct hugepage_subpool *spool = subpool_inode(inode);
750 bool reserved = false;
752 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
753 if (rsv_adjust > 0) {
754 struct hstate *h = hstate_inode(inode);
756 if (!hugetlb_acct_memory(h, 1))
758 } else if (!rsv_adjust) {
763 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
767 * Count and return the number of huge pages in the reserve map
768 * that intersect with the range [f, t).
770 static long region_count(struct resv_map *resv, long f, long t)
772 struct list_head *head = &resv->regions;
773 struct file_region *rg;
776 spin_lock(&resv->lock);
777 /* Locate each segment we overlap with, and count that overlap. */
778 list_for_each_entry(rg, head, link) {
787 seg_from = max(rg->from, f);
788 seg_to = min(rg->to, t);
790 chg += seg_to - seg_from;
792 spin_unlock(&resv->lock);
798 * Convert the address within this vma to the page offset within
799 * the mapping, in pagecache page units; huge pages here.
801 static pgoff_t vma_hugecache_offset(struct hstate *h,
802 struct vm_area_struct *vma, unsigned long address)
804 return ((address - vma->vm_start) >> huge_page_shift(h)) +
805 (vma->vm_pgoff >> huge_page_order(h));
808 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
809 unsigned long address)
811 return vma_hugecache_offset(hstate_vma(vma), vma, address);
813 EXPORT_SYMBOL_GPL(linear_hugepage_index);
816 * Return the size of the pages allocated when backing a VMA. In the majority
817 * cases this will be same size as used by the page table entries.
819 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
821 if (vma->vm_ops && vma->vm_ops->pagesize)
822 return vma->vm_ops->pagesize(vma);
825 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
828 * Return the page size being used by the MMU to back a VMA. In the majority
829 * of cases, the page size used by the kernel matches the MMU size. On
830 * architectures where it differs, an architecture-specific 'strong'
831 * version of this symbol is required.
833 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
835 return vma_kernel_pagesize(vma);
839 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
840 * bits of the reservation map pointer, which are always clear due to
843 #define HPAGE_RESV_OWNER (1UL << 0)
844 #define HPAGE_RESV_UNMAPPED (1UL << 1)
845 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
848 * These helpers are used to track how many pages are reserved for
849 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
850 * is guaranteed to have their future faults succeed.
852 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
853 * the reserve counters are updated with the hugetlb_lock held. It is safe
854 * to reset the VMA at fork() time as it is not in use yet and there is no
855 * chance of the global counters getting corrupted as a result of the values.
857 * The private mapping reservation is represented in a subtly different
858 * manner to a shared mapping. A shared mapping has a region map associated
859 * with the underlying file, this region map represents the backing file
860 * pages which have ever had a reservation assigned which this persists even
861 * after the page is instantiated. A private mapping has a region map
862 * associated with the original mmap which is attached to all VMAs which
863 * reference it, this region map represents those offsets which have consumed
864 * reservation ie. where pages have been instantiated.
866 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
868 return (unsigned long)vma->vm_private_data;
871 static void set_vma_private_data(struct vm_area_struct *vma,
874 vma->vm_private_data = (void *)value;
878 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
879 struct hugetlb_cgroup *h_cg,
882 #ifdef CONFIG_CGROUP_HUGETLB
884 resv_map->reservation_counter = NULL;
885 resv_map->pages_per_hpage = 0;
886 resv_map->css = NULL;
888 resv_map->reservation_counter =
889 &h_cg->rsvd_hugepage[hstate_index(h)];
890 resv_map->pages_per_hpage = pages_per_huge_page(h);
891 resv_map->css = &h_cg->css;
896 struct resv_map *resv_map_alloc(void)
898 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
899 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
901 if (!resv_map || !rg) {
907 kref_init(&resv_map->refs);
908 spin_lock_init(&resv_map->lock);
909 INIT_LIST_HEAD(&resv_map->regions);
911 resv_map->adds_in_progress = 0;
913 * Initialize these to 0. On shared mappings, 0's here indicate these
914 * fields don't do cgroup accounting. On private mappings, these will be
915 * re-initialized to the proper values, to indicate that hugetlb cgroup
916 * reservations are to be un-charged from here.
918 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
920 INIT_LIST_HEAD(&resv_map->region_cache);
921 list_add(&rg->link, &resv_map->region_cache);
922 resv_map->region_cache_count = 1;
927 void resv_map_release(struct kref *ref)
929 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
930 struct list_head *head = &resv_map->region_cache;
931 struct file_region *rg, *trg;
933 /* Clear out any active regions before we release the map. */
934 region_del(resv_map, 0, LONG_MAX);
936 /* ... and any entries left in the cache */
937 list_for_each_entry_safe(rg, trg, head, link) {
942 VM_BUG_ON(resv_map->adds_in_progress);
947 static inline struct resv_map *inode_resv_map(struct inode *inode)
950 * At inode evict time, i_mapping may not point to the original
951 * address space within the inode. This original address space
952 * contains the pointer to the resv_map. So, always use the
953 * address space embedded within the inode.
954 * The VERY common case is inode->mapping == &inode->i_data but,
955 * this may not be true for device special inodes.
957 return (struct resv_map *)(&inode->i_data)->private_data;
960 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
962 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
963 if (vma->vm_flags & VM_MAYSHARE) {
964 struct address_space *mapping = vma->vm_file->f_mapping;
965 struct inode *inode = mapping->host;
967 return inode_resv_map(inode);
970 return (struct resv_map *)(get_vma_private_data(vma) &
975 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
977 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
978 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
980 set_vma_private_data(vma, (get_vma_private_data(vma) &
981 HPAGE_RESV_MASK) | (unsigned long)map);
984 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
986 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
987 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
989 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
992 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
994 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
996 return (get_vma_private_data(vma) & flag) != 0;
999 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
1000 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1002 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1003 if (!(vma->vm_flags & VM_MAYSHARE))
1004 vma->vm_private_data = (void *)0;
1007 /* Returns true if the VMA has associated reserve pages */
1008 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1010 if (vma->vm_flags & VM_NORESERVE) {
1012 * This address is already reserved by other process(chg == 0),
1013 * so, we should decrement reserved count. Without decrementing,
1014 * reserve count remains after releasing inode, because this
1015 * allocated page will go into page cache and is regarded as
1016 * coming from reserved pool in releasing step. Currently, we
1017 * don't have any other solution to deal with this situation
1018 * properly, so add work-around here.
1020 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1026 /* Shared mappings always use reserves */
1027 if (vma->vm_flags & VM_MAYSHARE) {
1029 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1030 * be a region map for all pages. The only situation where
1031 * there is no region map is if a hole was punched via
1032 * fallocate. In this case, there really are no reserves to
1033 * use. This situation is indicated if chg != 0.
1042 * Only the process that called mmap() has reserves for
1045 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1047 * Like the shared case above, a hole punch or truncate
1048 * could have been performed on the private mapping.
1049 * Examine the value of chg to determine if reserves
1050 * actually exist or were previously consumed.
1051 * Very Subtle - The value of chg comes from a previous
1052 * call to vma_needs_reserves(). The reserve map for
1053 * private mappings has different (opposite) semantics
1054 * than that of shared mappings. vma_needs_reserves()
1055 * has already taken this difference in semantics into
1056 * account. Therefore, the meaning of chg is the same
1057 * as in the shared case above. Code could easily be
1058 * combined, but keeping it separate draws attention to
1059 * subtle differences.
1070 static void enqueue_huge_page(struct hstate *h, struct page *page)
1072 int nid = page_to_nid(page);
1074 lockdep_assert_held(&hugetlb_lock);
1075 list_move(&page->lru, &h->hugepage_freelists[nid]);
1076 h->free_huge_pages++;
1077 h->free_huge_pages_node[nid]++;
1078 SetHPageFreed(page);
1081 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1084 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1086 lockdep_assert_held(&hugetlb_lock);
1087 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1088 if (pin && !is_pinnable_page(page))
1091 if (PageHWPoison(page))
1094 list_move(&page->lru, &h->hugepage_activelist);
1095 set_page_refcounted(page);
1096 ClearHPageFreed(page);
1097 h->free_huge_pages--;
1098 h->free_huge_pages_node[nid]--;
1105 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1108 unsigned int cpuset_mems_cookie;
1109 struct zonelist *zonelist;
1112 int node = NUMA_NO_NODE;
1114 zonelist = node_zonelist(nid, gfp_mask);
1117 cpuset_mems_cookie = read_mems_allowed_begin();
1118 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1121 if (!cpuset_zone_allowed(zone, gfp_mask))
1124 * no need to ask again on the same node. Pool is node rather than
1127 if (zone_to_nid(zone) == node)
1129 node = zone_to_nid(zone);
1131 page = dequeue_huge_page_node_exact(h, node);
1135 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1141 static struct page *dequeue_huge_page_vma(struct hstate *h,
1142 struct vm_area_struct *vma,
1143 unsigned long address, int avoid_reserve,
1147 struct mempolicy *mpol;
1149 nodemask_t *nodemask;
1153 * A child process with MAP_PRIVATE mappings created by their parent
1154 * have no page reserves. This check ensures that reservations are
1155 * not "stolen". The child may still get SIGKILLed
1157 if (!vma_has_reserves(vma, chg) &&
1158 h->free_huge_pages - h->resv_huge_pages == 0)
1161 /* If reserves cannot be used, ensure enough pages are in the pool */
1162 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1165 gfp_mask = htlb_alloc_mask(h);
1166 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1167 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1168 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1169 SetHPageRestoreReserve(page);
1170 h->resv_huge_pages--;
1173 mpol_cond_put(mpol);
1181 * common helper functions for hstate_next_node_to_{alloc|free}.
1182 * We may have allocated or freed a huge page based on a different
1183 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1184 * be outside of *nodes_allowed. Ensure that we use an allowed
1185 * node for alloc or free.
1187 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1189 nid = next_node_in(nid, *nodes_allowed);
1190 VM_BUG_ON(nid >= MAX_NUMNODES);
1195 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1197 if (!node_isset(nid, *nodes_allowed))
1198 nid = next_node_allowed(nid, nodes_allowed);
1203 * returns the previously saved node ["this node"] from which to
1204 * allocate a persistent huge page for the pool and advance the
1205 * next node from which to allocate, handling wrap at end of node
1208 static int hstate_next_node_to_alloc(struct hstate *h,
1209 nodemask_t *nodes_allowed)
1213 VM_BUG_ON(!nodes_allowed);
1215 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1216 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1222 * helper for remove_pool_huge_page() - return the previously saved
1223 * node ["this node"] from which to free a huge page. Advance the
1224 * next node id whether or not we find a free huge page to free so
1225 * that the next attempt to free addresses the next node.
1227 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1231 VM_BUG_ON(!nodes_allowed);
1233 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1234 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1239 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1240 for (nr_nodes = nodes_weight(*mask); \
1242 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1245 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1246 for (nr_nodes = nodes_weight(*mask); \
1248 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1251 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1252 static void destroy_compound_gigantic_page(struct page *page,
1256 int nr_pages = 1 << order;
1257 struct page *p = page + 1;
1259 atomic_set(compound_mapcount_ptr(page), 0);
1260 atomic_set(compound_pincount_ptr(page), 0);
1262 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1263 clear_compound_head(p);
1264 set_page_refcounted(p);
1267 set_compound_order(page, 0);
1268 page[1].compound_nr = 0;
1269 __ClearPageHead(page);
1272 static void free_gigantic_page(struct page *page, unsigned int order)
1275 * If the page isn't allocated using the cma allocator,
1276 * cma_release() returns false.
1279 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1283 free_contig_range(page_to_pfn(page), 1 << order);
1286 #ifdef CONFIG_CONTIG_ALLOC
1287 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1288 int nid, nodemask_t *nodemask)
1290 unsigned long nr_pages = pages_per_huge_page(h);
1291 if (nid == NUMA_NO_NODE)
1292 nid = numa_mem_id();
1299 if (hugetlb_cma[nid]) {
1300 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1301 huge_page_order(h), true);
1306 if (!(gfp_mask & __GFP_THISNODE)) {
1307 for_each_node_mask(node, *nodemask) {
1308 if (node == nid || !hugetlb_cma[node])
1311 page = cma_alloc(hugetlb_cma[node], nr_pages,
1312 huge_page_order(h), true);
1320 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1323 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1324 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1325 #else /* !CONFIG_CONTIG_ALLOC */
1326 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1327 int nid, nodemask_t *nodemask)
1331 #endif /* CONFIG_CONTIG_ALLOC */
1333 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1334 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1335 int nid, nodemask_t *nodemask)
1339 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1340 static inline void destroy_compound_gigantic_page(struct page *page,
1341 unsigned int order) { }
1345 * Remove hugetlb page from lists, and update dtor so that page appears
1346 * as just a compound page. A reference is held on the page.
1348 * Must be called with hugetlb lock held.
1350 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1351 bool adjust_surplus)
1353 int nid = page_to_nid(page);
1355 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1356 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1358 lockdep_assert_held(&hugetlb_lock);
1359 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1362 list_del(&page->lru);
1364 if (HPageFreed(page)) {
1365 h->free_huge_pages--;
1366 h->free_huge_pages_node[nid]--;
1368 if (adjust_surplus) {
1369 h->surplus_huge_pages--;
1370 h->surplus_huge_pages_node[nid]--;
1373 set_page_refcounted(page);
1374 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1377 h->nr_huge_pages_node[nid]--;
1380 static void add_hugetlb_page(struct hstate *h, struct page *page,
1381 bool adjust_surplus)
1384 int nid = page_to_nid(page);
1386 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1388 lockdep_assert_held(&hugetlb_lock);
1390 INIT_LIST_HEAD(&page->lru);
1392 h->nr_huge_pages_node[nid]++;
1394 if (adjust_surplus) {
1395 h->surplus_huge_pages++;
1396 h->surplus_huge_pages_node[nid]++;
1399 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1400 set_page_private(page, 0);
1401 SetHPageVmemmapOptimized(page);
1404 * This page is now managed by the hugetlb allocator and has
1405 * no users -- drop the last reference.
1407 zeroed = put_page_testzero(page);
1408 VM_BUG_ON_PAGE(!zeroed, page);
1409 arch_clear_hugepage_flags(page);
1410 enqueue_huge_page(h, page);
1413 static void __update_and_free_page(struct hstate *h, struct page *page)
1416 struct page *subpage = page;
1418 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1421 if (alloc_huge_page_vmemmap(h, page)) {
1422 spin_lock_irq(&hugetlb_lock);
1424 * If we cannot allocate vmemmap pages, just refuse to free the
1425 * page and put the page back on the hugetlb free list and treat
1426 * as a surplus page.
1428 add_hugetlb_page(h, page, true);
1429 spin_unlock_irq(&hugetlb_lock);
1433 for (i = 0; i < pages_per_huge_page(h);
1434 i++, subpage = mem_map_next(subpage, page, i)) {
1435 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1436 1 << PG_referenced | 1 << PG_dirty |
1437 1 << PG_active | 1 << PG_private |
1440 if (hstate_is_gigantic(h)) {
1441 destroy_compound_gigantic_page(page, huge_page_order(h));
1442 free_gigantic_page(page, huge_page_order(h));
1444 __free_pages(page, huge_page_order(h));
1449 * As update_and_free_page() can be called under any context, so we cannot
1450 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1451 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1452 * the vmemmap pages.
1454 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1455 * freed and frees them one-by-one. As the page->mapping pointer is going
1456 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1457 * structure of a lockless linked list of huge pages to be freed.
1459 static LLIST_HEAD(hpage_freelist);
1461 static void free_hpage_workfn(struct work_struct *work)
1463 struct llist_node *node;
1465 node = llist_del_all(&hpage_freelist);
1471 page = container_of((struct address_space **)node,
1472 struct page, mapping);
1474 page->mapping = NULL;
1476 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1477 * is going to trigger because a previous call to
1478 * remove_hugetlb_page() will set_compound_page_dtor(page,
1479 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1481 h = size_to_hstate(page_size(page));
1483 __update_and_free_page(h, page);
1488 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1490 static inline void flush_free_hpage_work(struct hstate *h)
1492 if (free_vmemmap_pages_per_hpage(h))
1493 flush_work(&free_hpage_work);
1496 static void update_and_free_page(struct hstate *h, struct page *page,
1499 if (!HPageVmemmapOptimized(page) || !atomic) {
1500 __update_and_free_page(h, page);
1505 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1507 * Only call schedule_work() if hpage_freelist is previously
1508 * empty. Otherwise, schedule_work() had been called but the workfn
1509 * hasn't retrieved the list yet.
1511 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1512 schedule_work(&free_hpage_work);
1515 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1517 struct page *page, *t_page;
1519 list_for_each_entry_safe(page, t_page, list, lru) {
1520 update_and_free_page(h, page, false);
1525 struct hstate *size_to_hstate(unsigned long size)
1529 for_each_hstate(h) {
1530 if (huge_page_size(h) == size)
1536 void free_huge_page(struct page *page)
1539 * Can't pass hstate in here because it is called from the
1540 * compound page destructor.
1542 struct hstate *h = page_hstate(page);
1543 int nid = page_to_nid(page);
1544 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1545 bool restore_reserve;
1546 unsigned long flags;
1548 VM_BUG_ON_PAGE(page_count(page), page);
1549 VM_BUG_ON_PAGE(page_mapcount(page), page);
1551 hugetlb_set_page_subpool(page, NULL);
1552 page->mapping = NULL;
1553 restore_reserve = HPageRestoreReserve(page);
1554 ClearHPageRestoreReserve(page);
1557 * If HPageRestoreReserve was set on page, page allocation consumed a
1558 * reservation. If the page was associated with a subpool, there
1559 * would have been a page reserved in the subpool before allocation
1560 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1561 * reservation, do not call hugepage_subpool_put_pages() as this will
1562 * remove the reserved page from the subpool.
1564 if (!restore_reserve) {
1566 * A return code of zero implies that the subpool will be
1567 * under its minimum size if the reservation is not restored
1568 * after page is free. Therefore, force restore_reserve
1571 if (hugepage_subpool_put_pages(spool, 1) == 0)
1572 restore_reserve = true;
1575 spin_lock_irqsave(&hugetlb_lock, flags);
1576 ClearHPageMigratable(page);
1577 hugetlb_cgroup_uncharge_page(hstate_index(h),
1578 pages_per_huge_page(h), page);
1579 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1580 pages_per_huge_page(h), page);
1581 if (restore_reserve)
1582 h->resv_huge_pages++;
1584 if (HPageTemporary(page)) {
1585 remove_hugetlb_page(h, page, false);
1586 spin_unlock_irqrestore(&hugetlb_lock, flags);
1587 update_and_free_page(h, page, true);
1588 } else if (h->surplus_huge_pages_node[nid]) {
1589 /* remove the page from active list */
1590 remove_hugetlb_page(h, page, true);
1591 spin_unlock_irqrestore(&hugetlb_lock, flags);
1592 update_and_free_page(h, page, true);
1594 arch_clear_hugepage_flags(page);
1595 enqueue_huge_page(h, page);
1596 spin_unlock_irqrestore(&hugetlb_lock, flags);
1601 * Must be called with the hugetlb lock held
1603 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1605 lockdep_assert_held(&hugetlb_lock);
1607 h->nr_huge_pages_node[nid]++;
1610 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1612 free_huge_page_vmemmap(h, page);
1613 INIT_LIST_HEAD(&page->lru);
1614 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1615 hugetlb_set_page_subpool(page, NULL);
1616 set_hugetlb_cgroup(page, NULL);
1617 set_hugetlb_cgroup_rsvd(page, NULL);
1620 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1622 __prep_new_huge_page(h, page);
1623 spin_lock_irq(&hugetlb_lock);
1624 __prep_account_new_huge_page(h, nid);
1625 spin_unlock_irq(&hugetlb_lock);
1628 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1631 int nr_pages = 1 << order;
1632 struct page *p = page + 1;
1634 /* we rely on prep_new_huge_page to set the destructor */
1635 set_compound_order(page, order);
1636 __ClearPageReserved(page);
1637 __SetPageHead(page);
1638 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1640 * For gigantic hugepages allocated through bootmem at
1641 * boot, it's safer to be consistent with the not-gigantic
1642 * hugepages and clear the PG_reserved bit from all tail pages
1643 * too. Otherwise drivers using get_user_pages() to access tail
1644 * pages may get the reference counting wrong if they see
1645 * PG_reserved set on a tail page (despite the head page not
1646 * having PG_reserved set). Enforcing this consistency between
1647 * head and tail pages allows drivers to optimize away a check
1648 * on the head page when they need know if put_page() is needed
1649 * after get_user_pages().
1651 __ClearPageReserved(p);
1652 set_page_count(p, 0);
1653 set_compound_head(p, page);
1655 atomic_set(compound_mapcount_ptr(page), -1);
1656 atomic_set(compound_pincount_ptr(page), 0);
1660 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1661 * transparent huge pages. See the PageTransHuge() documentation for more
1664 int PageHuge(struct page *page)
1666 if (!PageCompound(page))
1669 page = compound_head(page);
1670 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1672 EXPORT_SYMBOL_GPL(PageHuge);
1675 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1676 * normal or transparent huge pages.
1678 int PageHeadHuge(struct page *page_head)
1680 if (!PageHead(page_head))
1683 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1687 * Find and lock address space (mapping) in write mode.
1689 * Upon entry, the page is locked which means that page_mapping() is
1690 * stable. Due to locking order, we can only trylock_write. If we can
1691 * not get the lock, simply return NULL to caller.
1693 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1695 struct address_space *mapping = page_mapping(hpage);
1700 if (i_mmap_trylock_write(mapping))
1706 pgoff_t hugetlb_basepage_index(struct page *page)
1708 struct page *page_head = compound_head(page);
1709 pgoff_t index = page_index(page_head);
1710 unsigned long compound_idx;
1712 if (compound_order(page_head) >= MAX_ORDER)
1713 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1715 compound_idx = page - page_head;
1717 return (index << compound_order(page_head)) + compound_idx;
1720 static struct page *alloc_buddy_huge_page(struct hstate *h,
1721 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1722 nodemask_t *node_alloc_noretry)
1724 int order = huge_page_order(h);
1726 bool alloc_try_hard = true;
1729 * By default we always try hard to allocate the page with
1730 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1731 * a loop (to adjust global huge page counts) and previous allocation
1732 * failed, do not continue to try hard on the same node. Use the
1733 * node_alloc_noretry bitmap to manage this state information.
1735 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1736 alloc_try_hard = false;
1737 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1739 gfp_mask |= __GFP_RETRY_MAYFAIL;
1740 if (nid == NUMA_NO_NODE)
1741 nid = numa_mem_id();
1742 page = __alloc_pages(gfp_mask, order, nid, nmask);
1744 __count_vm_event(HTLB_BUDDY_PGALLOC);
1746 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1749 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1750 * indicates an overall state change. Clear bit so that we resume
1751 * normal 'try hard' allocations.
1753 if (node_alloc_noretry && page && !alloc_try_hard)
1754 node_clear(nid, *node_alloc_noretry);
1757 * If we tried hard to get a page but failed, set bit so that
1758 * subsequent attempts will not try as hard until there is an
1759 * overall state change.
1761 if (node_alloc_noretry && !page && alloc_try_hard)
1762 node_set(nid, *node_alloc_noretry);
1768 * Common helper to allocate a fresh hugetlb page. All specific allocators
1769 * should use this function to get new hugetlb pages
1771 static struct page *alloc_fresh_huge_page(struct hstate *h,
1772 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1773 nodemask_t *node_alloc_noretry)
1777 if (hstate_is_gigantic(h))
1778 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1780 page = alloc_buddy_huge_page(h, gfp_mask,
1781 nid, nmask, node_alloc_noretry);
1785 if (hstate_is_gigantic(h))
1786 prep_compound_gigantic_page(page, huge_page_order(h));
1787 prep_new_huge_page(h, page, page_to_nid(page));
1793 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1796 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1797 nodemask_t *node_alloc_noretry)
1801 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1803 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1804 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1805 node_alloc_noretry);
1813 put_page(page); /* free it into the hugepage allocator */
1819 * Remove huge page from pool from next node to free. Attempt to keep
1820 * persistent huge pages more or less balanced over allowed nodes.
1821 * This routine only 'removes' the hugetlb page. The caller must make
1822 * an additional call to free the page to low level allocators.
1823 * Called with hugetlb_lock locked.
1825 static struct page *remove_pool_huge_page(struct hstate *h,
1826 nodemask_t *nodes_allowed,
1830 struct page *page = NULL;
1832 lockdep_assert_held(&hugetlb_lock);
1833 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1835 * If we're returning unused surplus pages, only examine
1836 * nodes with surplus pages.
1838 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1839 !list_empty(&h->hugepage_freelists[node])) {
1840 page = list_entry(h->hugepage_freelists[node].next,
1842 remove_hugetlb_page(h, page, acct_surplus);
1851 * Dissolve a given free hugepage into free buddy pages. This function does
1852 * nothing for in-use hugepages and non-hugepages.
1853 * This function returns values like below:
1855 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
1856 * when the system is under memory pressure and the feature of
1857 * freeing unused vmemmap pages associated with each hugetlb page
1859 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1860 * (allocated or reserved.)
1861 * 0: successfully dissolved free hugepages or the page is not a
1862 * hugepage (considered as already dissolved)
1864 int dissolve_free_huge_page(struct page *page)
1869 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1870 if (!PageHuge(page))
1873 spin_lock_irq(&hugetlb_lock);
1874 if (!PageHuge(page)) {
1879 if (!page_count(page)) {
1880 struct page *head = compound_head(page);
1881 struct hstate *h = page_hstate(head);
1882 if (h->free_huge_pages - h->resv_huge_pages == 0)
1886 * We should make sure that the page is already on the free list
1887 * when it is dissolved.
1889 if (unlikely(!HPageFreed(head))) {
1890 spin_unlock_irq(&hugetlb_lock);
1894 * Theoretically, we should return -EBUSY when we
1895 * encounter this race. In fact, we have a chance
1896 * to successfully dissolve the page if we do a
1897 * retry. Because the race window is quite small.
1898 * If we seize this opportunity, it is an optimization
1899 * for increasing the success rate of dissolving page.
1904 remove_hugetlb_page(h, head, false);
1905 h->max_huge_pages--;
1906 spin_unlock_irq(&hugetlb_lock);
1909 * Normally update_and_free_page will allocate required vmemmmap
1910 * before freeing the page. update_and_free_page will fail to
1911 * free the page if it can not allocate required vmemmap. We
1912 * need to adjust max_huge_pages if the page is not freed.
1913 * Attempt to allocate vmemmmap here so that we can take
1914 * appropriate action on failure.
1916 rc = alloc_huge_page_vmemmap(h, head);
1919 * Move PageHWPoison flag from head page to the raw
1920 * error page, which makes any subpages rather than
1921 * the error page reusable.
1923 if (PageHWPoison(head) && page != head) {
1924 SetPageHWPoison(page);
1925 ClearPageHWPoison(head);
1927 update_and_free_page(h, head, false);
1929 spin_lock_irq(&hugetlb_lock);
1930 add_hugetlb_page(h, head, false);
1931 h->max_huge_pages++;
1932 spin_unlock_irq(&hugetlb_lock);
1938 spin_unlock_irq(&hugetlb_lock);
1943 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1944 * make specified memory blocks removable from the system.
1945 * Note that this will dissolve a free gigantic hugepage completely, if any
1946 * part of it lies within the given range.
1947 * Also note that if dissolve_free_huge_page() returns with an error, all
1948 * free hugepages that were dissolved before that error are lost.
1950 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1956 if (!hugepages_supported())
1959 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1960 page = pfn_to_page(pfn);
1961 rc = dissolve_free_huge_page(page);
1970 * Allocates a fresh surplus page from the page allocator.
1972 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1973 int nid, nodemask_t *nmask)
1975 struct page *page = NULL;
1977 if (hstate_is_gigantic(h))
1980 spin_lock_irq(&hugetlb_lock);
1981 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1983 spin_unlock_irq(&hugetlb_lock);
1985 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1989 spin_lock_irq(&hugetlb_lock);
1991 * We could have raced with the pool size change.
1992 * Double check that and simply deallocate the new page
1993 * if we would end up overcommiting the surpluses. Abuse
1994 * temporary page to workaround the nasty free_huge_page
1997 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1998 SetHPageTemporary(page);
1999 spin_unlock_irq(&hugetlb_lock);
2003 h->surplus_huge_pages++;
2004 h->surplus_huge_pages_node[page_to_nid(page)]++;
2008 spin_unlock_irq(&hugetlb_lock);
2013 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2014 int nid, nodemask_t *nmask)
2018 if (hstate_is_gigantic(h))
2021 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2026 * We do not account these pages as surplus because they are only
2027 * temporary and will be released properly on the last reference
2029 SetHPageTemporary(page);
2035 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2038 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2039 struct vm_area_struct *vma, unsigned long addr)
2042 struct mempolicy *mpol;
2043 gfp_t gfp_mask = htlb_alloc_mask(h);
2045 nodemask_t *nodemask;
2047 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2048 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
2049 mpol_cond_put(mpol);
2054 /* page migration callback function */
2055 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2056 nodemask_t *nmask, gfp_t gfp_mask)
2058 spin_lock_irq(&hugetlb_lock);
2059 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2062 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2064 spin_unlock_irq(&hugetlb_lock);
2068 spin_unlock_irq(&hugetlb_lock);
2070 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2073 /* mempolicy aware migration callback */
2074 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2075 unsigned long address)
2077 struct mempolicy *mpol;
2078 nodemask_t *nodemask;
2083 gfp_mask = htlb_alloc_mask(h);
2084 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2085 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2086 mpol_cond_put(mpol);
2092 * Increase the hugetlb pool such that it can accommodate a reservation
2095 static int gather_surplus_pages(struct hstate *h, long delta)
2096 __must_hold(&hugetlb_lock)
2098 struct list_head surplus_list;
2099 struct page *page, *tmp;
2102 long needed, allocated;
2103 bool alloc_ok = true;
2105 lockdep_assert_held(&hugetlb_lock);
2106 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2108 h->resv_huge_pages += delta;
2113 INIT_LIST_HEAD(&surplus_list);
2117 spin_unlock_irq(&hugetlb_lock);
2118 for (i = 0; i < needed; i++) {
2119 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2120 NUMA_NO_NODE, NULL);
2125 list_add(&page->lru, &surplus_list);
2131 * After retaking hugetlb_lock, we need to recalculate 'needed'
2132 * because either resv_huge_pages or free_huge_pages may have changed.
2134 spin_lock_irq(&hugetlb_lock);
2135 needed = (h->resv_huge_pages + delta) -
2136 (h->free_huge_pages + allocated);
2141 * We were not able to allocate enough pages to
2142 * satisfy the entire reservation so we free what
2143 * we've allocated so far.
2148 * The surplus_list now contains _at_least_ the number of extra pages
2149 * needed to accommodate the reservation. Add the appropriate number
2150 * of pages to the hugetlb pool and free the extras back to the buddy
2151 * allocator. Commit the entire reservation here to prevent another
2152 * process from stealing the pages as they are added to the pool but
2153 * before they are reserved.
2155 needed += allocated;
2156 h->resv_huge_pages += delta;
2159 /* Free the needed pages to the hugetlb pool */
2160 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2166 * This page is now managed by the hugetlb allocator and has
2167 * no users -- drop the buddy allocator's reference.
2169 zeroed = put_page_testzero(page);
2170 VM_BUG_ON_PAGE(!zeroed, page);
2171 enqueue_huge_page(h, page);
2174 spin_unlock_irq(&hugetlb_lock);
2176 /* Free unnecessary surplus pages to the buddy allocator */
2177 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2179 spin_lock_irq(&hugetlb_lock);
2185 * This routine has two main purposes:
2186 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2187 * in unused_resv_pages. This corresponds to the prior adjustments made
2188 * to the associated reservation map.
2189 * 2) Free any unused surplus pages that may have been allocated to satisfy
2190 * the reservation. As many as unused_resv_pages may be freed.
2192 static void return_unused_surplus_pages(struct hstate *h,
2193 unsigned long unused_resv_pages)
2195 unsigned long nr_pages;
2197 LIST_HEAD(page_list);
2199 lockdep_assert_held(&hugetlb_lock);
2200 /* Uncommit the reservation */
2201 h->resv_huge_pages -= unused_resv_pages;
2203 /* Cannot return gigantic pages currently */
2204 if (hstate_is_gigantic(h))
2208 * Part (or even all) of the reservation could have been backed
2209 * by pre-allocated pages. Only free surplus pages.
2211 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2214 * We want to release as many surplus pages as possible, spread
2215 * evenly across all nodes with memory. Iterate across these nodes
2216 * until we can no longer free unreserved surplus pages. This occurs
2217 * when the nodes with surplus pages have no free pages.
2218 * remove_pool_huge_page() will balance the freed pages across the
2219 * on-line nodes with memory and will handle the hstate accounting.
2221 while (nr_pages--) {
2222 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2226 list_add(&page->lru, &page_list);
2230 spin_unlock_irq(&hugetlb_lock);
2231 update_and_free_pages_bulk(h, &page_list);
2232 spin_lock_irq(&hugetlb_lock);
2237 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2238 * are used by the huge page allocation routines to manage reservations.
2240 * vma_needs_reservation is called to determine if the huge page at addr
2241 * within the vma has an associated reservation. If a reservation is
2242 * needed, the value 1 is returned. The caller is then responsible for
2243 * managing the global reservation and subpool usage counts. After
2244 * the huge page has been allocated, vma_commit_reservation is called
2245 * to add the page to the reservation map. If the page allocation fails,
2246 * the reservation must be ended instead of committed. vma_end_reservation
2247 * is called in such cases.
2249 * In the normal case, vma_commit_reservation returns the same value
2250 * as the preceding vma_needs_reservation call. The only time this
2251 * is not the case is if a reserve map was changed between calls. It
2252 * is the responsibility of the caller to notice the difference and
2253 * take appropriate action.
2255 * vma_add_reservation is used in error paths where a reservation must
2256 * be restored when a newly allocated huge page must be freed. It is
2257 * to be called after calling vma_needs_reservation to determine if a
2258 * reservation exists.
2260 * vma_del_reservation is used in error paths where an entry in the reserve
2261 * map was created during huge page allocation and must be removed. It is to
2262 * be called after calling vma_needs_reservation to determine if a reservation
2265 enum vma_resv_mode {
2272 static long __vma_reservation_common(struct hstate *h,
2273 struct vm_area_struct *vma, unsigned long addr,
2274 enum vma_resv_mode mode)
2276 struct resv_map *resv;
2279 long dummy_out_regions_needed;
2281 resv = vma_resv_map(vma);
2285 idx = vma_hugecache_offset(h, vma, addr);
2287 case VMA_NEEDS_RESV:
2288 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2289 /* We assume that vma_reservation_* routines always operate on
2290 * 1 page, and that adding to resv map a 1 page entry can only
2291 * ever require 1 region.
2293 VM_BUG_ON(dummy_out_regions_needed != 1);
2295 case VMA_COMMIT_RESV:
2296 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2297 /* region_add calls of range 1 should never fail. */
2301 region_abort(resv, idx, idx + 1, 1);
2305 if (vma->vm_flags & VM_MAYSHARE) {
2306 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2307 /* region_add calls of range 1 should never fail. */
2310 region_abort(resv, idx, idx + 1, 1);
2311 ret = region_del(resv, idx, idx + 1);
2315 if (vma->vm_flags & VM_MAYSHARE) {
2316 region_abort(resv, idx, idx + 1, 1);
2317 ret = region_del(resv, idx, idx + 1);
2319 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2320 /* region_add calls of range 1 should never fail. */
2328 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2331 * We know private mapping must have HPAGE_RESV_OWNER set.
2333 * In most cases, reserves always exist for private mappings.
2334 * However, a file associated with mapping could have been
2335 * hole punched or truncated after reserves were consumed.
2336 * As subsequent fault on such a range will not use reserves.
2337 * Subtle - The reserve map for private mappings has the
2338 * opposite meaning than that of shared mappings. If NO
2339 * entry is in the reserve map, it means a reservation exists.
2340 * If an entry exists in the reserve map, it means the
2341 * reservation has already been consumed. As a result, the
2342 * return value of this routine is the opposite of the
2343 * value returned from reserve map manipulation routines above.
2352 static long vma_needs_reservation(struct hstate *h,
2353 struct vm_area_struct *vma, unsigned long addr)
2355 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2358 static long vma_commit_reservation(struct hstate *h,
2359 struct vm_area_struct *vma, unsigned long addr)
2361 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2364 static void vma_end_reservation(struct hstate *h,
2365 struct vm_area_struct *vma, unsigned long addr)
2367 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2370 static long vma_add_reservation(struct hstate *h,
2371 struct vm_area_struct *vma, unsigned long addr)
2373 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2376 static long vma_del_reservation(struct hstate *h,
2377 struct vm_area_struct *vma, unsigned long addr)
2379 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2383 * This routine is called to restore reservation information on error paths.
2384 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2385 * the hugetlb mutex should remain held when calling this routine.
2387 * It handles two specific cases:
2388 * 1) A reservation was in place and the page consumed the reservation.
2389 * HPageRestoreReserve is set in the page.
2390 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2391 * not set. However, alloc_huge_page always updates the reserve map.
2393 * In case 1, free_huge_page later in the error path will increment the
2394 * global reserve count. But, free_huge_page does not have enough context
2395 * to adjust the reservation map. This case deals primarily with private
2396 * mappings. Adjust the reserve map here to be consistent with global
2397 * reserve count adjustments to be made by free_huge_page. Make sure the
2398 * reserve map indicates there is a reservation present.
2400 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2402 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2403 unsigned long address, struct page *page)
2405 long rc = vma_needs_reservation(h, vma, address);
2407 if (HPageRestoreReserve(page)) {
2408 if (unlikely(rc < 0))
2410 * Rare out of memory condition in reserve map
2411 * manipulation. Clear HPageRestoreReserve so that
2412 * global reserve count will not be incremented
2413 * by free_huge_page. This will make it appear
2414 * as though the reservation for this page was
2415 * consumed. This may prevent the task from
2416 * faulting in the page at a later time. This
2417 * is better than inconsistent global huge page
2418 * accounting of reserve counts.
2420 ClearHPageRestoreReserve(page);
2422 (void)vma_add_reservation(h, vma, address);
2424 vma_end_reservation(h, vma, address);
2428 * This indicates there is an entry in the reserve map
2429 * added by alloc_huge_page. We know it was added
2430 * before the alloc_huge_page call, otherwise
2431 * HPageRestoreReserve would be set on the page.
2432 * Remove the entry so that a subsequent allocation
2433 * does not consume a reservation.
2435 rc = vma_del_reservation(h, vma, address);
2438 * VERY rare out of memory condition. Since
2439 * we can not delete the entry, set
2440 * HPageRestoreReserve so that the reserve
2441 * count will be incremented when the page
2442 * is freed. This reserve will be consumed
2443 * on a subsequent allocation.
2445 SetHPageRestoreReserve(page);
2446 } else if (rc < 0) {
2448 * Rare out of memory condition from
2449 * vma_needs_reservation call. Memory allocation is
2450 * only attempted if a new entry is needed. Therefore,
2451 * this implies there is not an entry in the
2454 * For shared mappings, no entry in the map indicates
2455 * no reservation. We are done.
2457 if (!(vma->vm_flags & VM_MAYSHARE))
2459 * For private mappings, no entry indicates
2460 * a reservation is present. Since we can
2461 * not add an entry, set SetHPageRestoreReserve
2462 * on the page so reserve count will be
2463 * incremented when freed. This reserve will
2464 * be consumed on a subsequent allocation.
2466 SetHPageRestoreReserve(page);
2469 * No reservation present, do nothing
2471 vma_end_reservation(h, vma, address);
2476 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2477 * @h: struct hstate old page belongs to
2478 * @old_page: Old page to dissolve
2479 * @list: List to isolate the page in case we need to
2480 * Returns 0 on success, otherwise negated error.
2482 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2483 struct list_head *list)
2485 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2486 int nid = page_to_nid(old_page);
2487 struct page *new_page;
2491 * Before dissolving the page, we need to allocate a new one for the
2492 * pool to remain stable. Here, we allocate the page and 'prep' it
2493 * by doing everything but actually updating counters and adding to
2494 * the pool. This simplifies and let us do most of the processing
2497 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2500 __prep_new_huge_page(h, new_page);
2503 spin_lock_irq(&hugetlb_lock);
2504 if (!PageHuge(old_page)) {
2506 * Freed from under us. Drop new_page too.
2509 } else if (page_count(old_page)) {
2511 * Someone has grabbed the page, try to isolate it here.
2512 * Fail with -EBUSY if not possible.
2514 spin_unlock_irq(&hugetlb_lock);
2515 if (!isolate_huge_page(old_page, list))
2517 spin_lock_irq(&hugetlb_lock);
2519 } else if (!HPageFreed(old_page)) {
2521 * Page's refcount is 0 but it has not been enqueued in the
2522 * freelist yet. Race window is small, so we can succeed here if
2525 spin_unlock_irq(&hugetlb_lock);
2530 * Ok, old_page is still a genuine free hugepage. Remove it from
2531 * the freelist and decrease the counters. These will be
2532 * incremented again when calling __prep_account_new_huge_page()
2533 * and enqueue_huge_page() for new_page. The counters will remain
2534 * stable since this happens under the lock.
2536 remove_hugetlb_page(h, old_page, false);
2539 * Reference count trick is needed because allocator gives us
2540 * referenced page but the pool requires pages with 0 refcount.
2542 __prep_account_new_huge_page(h, nid);
2543 page_ref_dec(new_page);
2544 enqueue_huge_page(h, new_page);
2547 * Pages have been replaced, we can safely free the old one.
2549 spin_unlock_irq(&hugetlb_lock);
2550 update_and_free_page(h, old_page, false);
2556 spin_unlock_irq(&hugetlb_lock);
2557 update_and_free_page(h, new_page, false);
2562 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2569 * The page might have been dissolved from under our feet, so make sure
2570 * to carefully check the state under the lock.
2571 * Return success when racing as if we dissolved the page ourselves.
2573 spin_lock_irq(&hugetlb_lock);
2574 if (PageHuge(page)) {
2575 head = compound_head(page);
2576 h = page_hstate(head);
2578 spin_unlock_irq(&hugetlb_lock);
2581 spin_unlock_irq(&hugetlb_lock);
2584 * Fence off gigantic pages as there is a cyclic dependency between
2585 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2586 * of bailing out right away without further retrying.
2588 if (hstate_is_gigantic(h))
2591 if (page_count(head) && isolate_huge_page(head, list))
2593 else if (!page_count(head))
2594 ret = alloc_and_dissolve_huge_page(h, head, list);
2599 struct page *alloc_huge_page(struct vm_area_struct *vma,
2600 unsigned long addr, int avoid_reserve)
2602 struct hugepage_subpool *spool = subpool_vma(vma);
2603 struct hstate *h = hstate_vma(vma);
2605 long map_chg, map_commit;
2608 struct hugetlb_cgroup *h_cg;
2609 bool deferred_reserve;
2611 idx = hstate_index(h);
2613 * Examine the region/reserve map to determine if the process
2614 * has a reservation for the page to be allocated. A return
2615 * code of zero indicates a reservation exists (no change).
2617 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2619 return ERR_PTR(-ENOMEM);
2622 * Processes that did not create the mapping will have no
2623 * reserves as indicated by the region/reserve map. Check
2624 * that the allocation will not exceed the subpool limit.
2625 * Allocations for MAP_NORESERVE mappings also need to be
2626 * checked against any subpool limit.
2628 if (map_chg || avoid_reserve) {
2629 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2631 vma_end_reservation(h, vma, addr);
2632 return ERR_PTR(-ENOSPC);
2636 * Even though there was no reservation in the region/reserve
2637 * map, there could be reservations associated with the
2638 * subpool that can be used. This would be indicated if the
2639 * return value of hugepage_subpool_get_pages() is zero.
2640 * However, if avoid_reserve is specified we still avoid even
2641 * the subpool reservations.
2647 /* If this allocation is not consuming a reservation, charge it now.
2649 deferred_reserve = map_chg || avoid_reserve;
2650 if (deferred_reserve) {
2651 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2652 idx, pages_per_huge_page(h), &h_cg);
2654 goto out_subpool_put;
2657 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2659 goto out_uncharge_cgroup_reservation;
2661 spin_lock_irq(&hugetlb_lock);
2663 * glb_chg is passed to indicate whether or not a page must be taken
2664 * from the global free pool (global change). gbl_chg == 0 indicates
2665 * a reservation exists for the allocation.
2667 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2669 spin_unlock_irq(&hugetlb_lock);
2670 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2672 goto out_uncharge_cgroup;
2673 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2674 SetHPageRestoreReserve(page);
2675 h->resv_huge_pages--;
2677 spin_lock_irq(&hugetlb_lock);
2678 list_add(&page->lru, &h->hugepage_activelist);
2681 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2682 /* If allocation is not consuming a reservation, also store the
2683 * hugetlb_cgroup pointer on the page.
2685 if (deferred_reserve) {
2686 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2690 spin_unlock_irq(&hugetlb_lock);
2692 hugetlb_set_page_subpool(page, spool);
2694 map_commit = vma_commit_reservation(h, vma, addr);
2695 if (unlikely(map_chg > map_commit)) {
2697 * The page was added to the reservation map between
2698 * vma_needs_reservation and vma_commit_reservation.
2699 * This indicates a race with hugetlb_reserve_pages.
2700 * Adjust for the subpool count incremented above AND
2701 * in hugetlb_reserve_pages for the same page. Also,
2702 * the reservation count added in hugetlb_reserve_pages
2703 * no longer applies.
2707 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2708 hugetlb_acct_memory(h, -rsv_adjust);
2709 if (deferred_reserve)
2710 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2711 pages_per_huge_page(h), page);
2715 out_uncharge_cgroup:
2716 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2717 out_uncharge_cgroup_reservation:
2718 if (deferred_reserve)
2719 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2722 if (map_chg || avoid_reserve)
2723 hugepage_subpool_put_pages(spool, 1);
2724 vma_end_reservation(h, vma, addr);
2725 return ERR_PTR(-ENOSPC);
2728 int alloc_bootmem_huge_page(struct hstate *h)
2729 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2730 int __alloc_bootmem_huge_page(struct hstate *h)
2732 struct huge_bootmem_page *m;
2735 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2738 addr = memblock_alloc_try_nid_raw(
2739 huge_page_size(h), huge_page_size(h),
2740 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2743 * Use the beginning of the huge page to store the
2744 * huge_bootmem_page struct (until gather_bootmem
2745 * puts them into the mem_map).
2754 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2755 /* Put them into a private list first because mem_map is not up yet */
2756 INIT_LIST_HEAD(&m->list);
2757 list_add(&m->list, &huge_boot_pages);
2762 static void __init prep_compound_huge_page(struct page *page,
2765 if (unlikely(order > (MAX_ORDER - 1)))
2766 prep_compound_gigantic_page(page, order);
2768 prep_compound_page(page, order);
2771 /* Put bootmem huge pages into the standard lists after mem_map is up */
2772 static void __init gather_bootmem_prealloc(void)
2774 struct huge_bootmem_page *m;
2776 list_for_each_entry(m, &huge_boot_pages, list) {
2777 struct page *page = virt_to_page(m);
2778 struct hstate *h = m->hstate;
2780 WARN_ON(page_count(page) != 1);
2781 prep_compound_huge_page(page, huge_page_order(h));
2782 WARN_ON(PageReserved(page));
2783 prep_new_huge_page(h, page, page_to_nid(page));
2784 put_page(page); /* free it into the hugepage allocator */
2787 * If we had gigantic hugepages allocated at boot time, we need
2788 * to restore the 'stolen' pages to totalram_pages in order to
2789 * fix confusing memory reports from free(1) and another
2790 * side-effects, like CommitLimit going negative.
2792 if (hstate_is_gigantic(h))
2793 adjust_managed_page_count(page, pages_per_huge_page(h));
2798 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2801 nodemask_t *node_alloc_noretry;
2803 if (!hstate_is_gigantic(h)) {
2805 * Bit mask controlling how hard we retry per-node allocations.
2806 * Ignore errors as lower level routines can deal with
2807 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2808 * time, we are likely in bigger trouble.
2810 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2813 /* allocations done at boot time */
2814 node_alloc_noretry = NULL;
2817 /* bit mask controlling how hard we retry per-node allocations */
2818 if (node_alloc_noretry)
2819 nodes_clear(*node_alloc_noretry);
2821 for (i = 0; i < h->max_huge_pages; ++i) {
2822 if (hstate_is_gigantic(h)) {
2823 if (hugetlb_cma_size) {
2824 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2827 if (!alloc_bootmem_huge_page(h))
2829 } else if (!alloc_pool_huge_page(h,
2830 &node_states[N_MEMORY],
2831 node_alloc_noretry))
2835 if (i < h->max_huge_pages) {
2838 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2839 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2840 h->max_huge_pages, buf, i);
2841 h->max_huge_pages = i;
2844 kfree(node_alloc_noretry);
2847 static void __init hugetlb_init_hstates(void)
2851 for_each_hstate(h) {
2852 if (minimum_order > huge_page_order(h))
2853 minimum_order = huge_page_order(h);
2855 /* oversize hugepages were init'ed in early boot */
2856 if (!hstate_is_gigantic(h))
2857 hugetlb_hstate_alloc_pages(h);
2859 VM_BUG_ON(minimum_order == UINT_MAX);
2862 static void __init report_hugepages(void)
2866 for_each_hstate(h) {
2869 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2870 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2871 buf, h->free_huge_pages);
2875 #ifdef CONFIG_HIGHMEM
2876 static void try_to_free_low(struct hstate *h, unsigned long count,
2877 nodemask_t *nodes_allowed)
2880 LIST_HEAD(page_list);
2882 lockdep_assert_held(&hugetlb_lock);
2883 if (hstate_is_gigantic(h))
2887 * Collect pages to be freed on a list, and free after dropping lock
2889 for_each_node_mask(i, *nodes_allowed) {
2890 struct page *page, *next;
2891 struct list_head *freel = &h->hugepage_freelists[i];
2892 list_for_each_entry_safe(page, next, freel, lru) {
2893 if (count >= h->nr_huge_pages)
2895 if (PageHighMem(page))
2897 remove_hugetlb_page(h, page, false);
2898 list_add(&page->lru, &page_list);
2903 spin_unlock_irq(&hugetlb_lock);
2904 update_and_free_pages_bulk(h, &page_list);
2905 spin_lock_irq(&hugetlb_lock);
2908 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2909 nodemask_t *nodes_allowed)
2915 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2916 * balanced by operating on them in a round-robin fashion.
2917 * Returns 1 if an adjustment was made.
2919 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2924 lockdep_assert_held(&hugetlb_lock);
2925 VM_BUG_ON(delta != -1 && delta != 1);
2928 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2929 if (h->surplus_huge_pages_node[node])
2933 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2934 if (h->surplus_huge_pages_node[node] <
2935 h->nr_huge_pages_node[node])
2942 h->surplus_huge_pages += delta;
2943 h->surplus_huge_pages_node[node] += delta;
2947 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2948 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2949 nodemask_t *nodes_allowed)
2951 unsigned long min_count, ret;
2953 LIST_HEAD(page_list);
2954 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2957 * Bit mask controlling how hard we retry per-node allocations.
2958 * If we can not allocate the bit mask, do not attempt to allocate
2959 * the requested huge pages.
2961 if (node_alloc_noretry)
2962 nodes_clear(*node_alloc_noretry);
2967 * resize_lock mutex prevents concurrent adjustments to number of
2968 * pages in hstate via the proc/sysfs interfaces.
2970 mutex_lock(&h->resize_lock);
2971 flush_free_hpage_work(h);
2972 spin_lock_irq(&hugetlb_lock);
2975 * Check for a node specific request.
2976 * Changing node specific huge page count may require a corresponding
2977 * change to the global count. In any case, the passed node mask
2978 * (nodes_allowed) will restrict alloc/free to the specified node.
2980 if (nid != NUMA_NO_NODE) {
2981 unsigned long old_count = count;
2983 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2985 * User may have specified a large count value which caused the
2986 * above calculation to overflow. In this case, they wanted
2987 * to allocate as many huge pages as possible. Set count to
2988 * largest possible value to align with their intention.
2990 if (count < old_count)
2995 * Gigantic pages runtime allocation depend on the capability for large
2996 * page range allocation.
2997 * If the system does not provide this feature, return an error when
2998 * the user tries to allocate gigantic pages but let the user free the
2999 * boottime allocated gigantic pages.
3001 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3002 if (count > persistent_huge_pages(h)) {
3003 spin_unlock_irq(&hugetlb_lock);
3004 mutex_unlock(&h->resize_lock);
3005 NODEMASK_FREE(node_alloc_noretry);
3008 /* Fall through to decrease pool */
3012 * Increase the pool size
3013 * First take pages out of surplus state. Then make up the
3014 * remaining difference by allocating fresh huge pages.
3016 * We might race with alloc_surplus_huge_page() here and be unable
3017 * to convert a surplus huge page to a normal huge page. That is
3018 * not critical, though, it just means the overall size of the
3019 * pool might be one hugepage larger than it needs to be, but
3020 * within all the constraints specified by the sysctls.
3022 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3023 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3027 while (count > persistent_huge_pages(h)) {
3029 * If this allocation races such that we no longer need the
3030 * page, free_huge_page will handle it by freeing the page
3031 * and reducing the surplus.
3033 spin_unlock_irq(&hugetlb_lock);
3035 /* yield cpu to avoid soft lockup */
3038 ret = alloc_pool_huge_page(h, nodes_allowed,
3039 node_alloc_noretry);
3040 spin_lock_irq(&hugetlb_lock);
3044 /* Bail for signals. Probably ctrl-c from user */
3045 if (signal_pending(current))
3050 * Decrease the pool size
3051 * First return free pages to the buddy allocator (being careful
3052 * to keep enough around to satisfy reservations). Then place
3053 * pages into surplus state as needed so the pool will shrink
3054 * to the desired size as pages become free.
3056 * By placing pages into the surplus state independent of the
3057 * overcommit value, we are allowing the surplus pool size to
3058 * exceed overcommit. There are few sane options here. Since
3059 * alloc_surplus_huge_page() is checking the global counter,
3060 * though, we'll note that we're not allowed to exceed surplus
3061 * and won't grow the pool anywhere else. Not until one of the
3062 * sysctls are changed, or the surplus pages go out of use.
3064 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3065 min_count = max(count, min_count);
3066 try_to_free_low(h, min_count, nodes_allowed);
3069 * Collect pages to be removed on list without dropping lock
3071 while (min_count < persistent_huge_pages(h)) {
3072 page = remove_pool_huge_page(h, nodes_allowed, 0);
3076 list_add(&page->lru, &page_list);
3078 /* free the pages after dropping lock */
3079 spin_unlock_irq(&hugetlb_lock);
3080 update_and_free_pages_bulk(h, &page_list);
3081 flush_free_hpage_work(h);
3082 spin_lock_irq(&hugetlb_lock);
3084 while (count < persistent_huge_pages(h)) {
3085 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3089 h->max_huge_pages = persistent_huge_pages(h);
3090 spin_unlock_irq(&hugetlb_lock);
3091 mutex_unlock(&h->resize_lock);
3093 NODEMASK_FREE(node_alloc_noretry);
3098 #define HSTATE_ATTR_RO(_name) \
3099 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3101 #define HSTATE_ATTR(_name) \
3102 static struct kobj_attribute _name##_attr = \
3103 __ATTR(_name, 0644, _name##_show, _name##_store)
3105 static struct kobject *hugepages_kobj;
3106 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3108 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3110 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3114 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3115 if (hstate_kobjs[i] == kobj) {
3117 *nidp = NUMA_NO_NODE;
3121 return kobj_to_node_hstate(kobj, nidp);
3124 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3125 struct kobj_attribute *attr, char *buf)
3128 unsigned long nr_huge_pages;
3131 h = kobj_to_hstate(kobj, &nid);
3132 if (nid == NUMA_NO_NODE)
3133 nr_huge_pages = h->nr_huge_pages;
3135 nr_huge_pages = h->nr_huge_pages_node[nid];
3137 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3140 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3141 struct hstate *h, int nid,
3142 unsigned long count, size_t len)
3145 nodemask_t nodes_allowed, *n_mask;
3147 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3150 if (nid == NUMA_NO_NODE) {
3152 * global hstate attribute
3154 if (!(obey_mempolicy &&
3155 init_nodemask_of_mempolicy(&nodes_allowed)))
3156 n_mask = &node_states[N_MEMORY];
3158 n_mask = &nodes_allowed;
3161 * Node specific request. count adjustment happens in
3162 * set_max_huge_pages() after acquiring hugetlb_lock.
3164 init_nodemask_of_node(&nodes_allowed, nid);
3165 n_mask = &nodes_allowed;
3168 err = set_max_huge_pages(h, count, nid, n_mask);
3170 return err ? err : len;
3173 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3174 struct kobject *kobj, const char *buf,
3178 unsigned long count;
3182 err = kstrtoul(buf, 10, &count);
3186 h = kobj_to_hstate(kobj, &nid);
3187 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3190 static ssize_t nr_hugepages_show(struct kobject *kobj,
3191 struct kobj_attribute *attr, char *buf)
3193 return nr_hugepages_show_common(kobj, attr, buf);
3196 static ssize_t nr_hugepages_store(struct kobject *kobj,
3197 struct kobj_attribute *attr, const char *buf, size_t len)
3199 return nr_hugepages_store_common(false, kobj, buf, len);
3201 HSTATE_ATTR(nr_hugepages);
3206 * hstate attribute for optionally mempolicy-based constraint on persistent
3207 * huge page alloc/free.
3209 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3210 struct kobj_attribute *attr,
3213 return nr_hugepages_show_common(kobj, attr, buf);
3216 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3217 struct kobj_attribute *attr, const char *buf, size_t len)
3219 return nr_hugepages_store_common(true, kobj, buf, len);
3221 HSTATE_ATTR(nr_hugepages_mempolicy);
3225 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3226 struct kobj_attribute *attr, char *buf)
3228 struct hstate *h = kobj_to_hstate(kobj, NULL);
3229 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3232 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3233 struct kobj_attribute *attr, const char *buf, size_t count)
3236 unsigned long input;
3237 struct hstate *h = kobj_to_hstate(kobj, NULL);
3239 if (hstate_is_gigantic(h))
3242 err = kstrtoul(buf, 10, &input);
3246 spin_lock_irq(&hugetlb_lock);
3247 h->nr_overcommit_huge_pages = input;
3248 spin_unlock_irq(&hugetlb_lock);
3252 HSTATE_ATTR(nr_overcommit_hugepages);
3254 static ssize_t free_hugepages_show(struct kobject *kobj,
3255 struct kobj_attribute *attr, char *buf)
3258 unsigned long free_huge_pages;
3261 h = kobj_to_hstate(kobj, &nid);
3262 if (nid == NUMA_NO_NODE)
3263 free_huge_pages = h->free_huge_pages;
3265 free_huge_pages = h->free_huge_pages_node[nid];
3267 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3269 HSTATE_ATTR_RO(free_hugepages);
3271 static ssize_t resv_hugepages_show(struct kobject *kobj,
3272 struct kobj_attribute *attr, char *buf)
3274 struct hstate *h = kobj_to_hstate(kobj, NULL);
3275 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3277 HSTATE_ATTR_RO(resv_hugepages);
3279 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3280 struct kobj_attribute *attr, char *buf)
3283 unsigned long surplus_huge_pages;
3286 h = kobj_to_hstate(kobj, &nid);
3287 if (nid == NUMA_NO_NODE)
3288 surplus_huge_pages = h->surplus_huge_pages;
3290 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3292 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3294 HSTATE_ATTR_RO(surplus_hugepages);
3296 static struct attribute *hstate_attrs[] = {
3297 &nr_hugepages_attr.attr,
3298 &nr_overcommit_hugepages_attr.attr,
3299 &free_hugepages_attr.attr,
3300 &resv_hugepages_attr.attr,
3301 &surplus_hugepages_attr.attr,
3303 &nr_hugepages_mempolicy_attr.attr,
3308 static const struct attribute_group hstate_attr_group = {
3309 .attrs = hstate_attrs,
3312 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3313 struct kobject **hstate_kobjs,
3314 const struct attribute_group *hstate_attr_group)
3317 int hi = hstate_index(h);
3319 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3320 if (!hstate_kobjs[hi])
3323 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3325 kobject_put(hstate_kobjs[hi]);
3326 hstate_kobjs[hi] = NULL;
3332 static void __init hugetlb_sysfs_init(void)
3337 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3338 if (!hugepages_kobj)
3341 for_each_hstate(h) {
3342 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3343 hstate_kobjs, &hstate_attr_group);
3345 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3352 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3353 * with node devices in node_devices[] using a parallel array. The array
3354 * index of a node device or _hstate == node id.
3355 * This is here to avoid any static dependency of the node device driver, in
3356 * the base kernel, on the hugetlb module.
3358 struct node_hstate {
3359 struct kobject *hugepages_kobj;
3360 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3362 static struct node_hstate node_hstates[MAX_NUMNODES];
3365 * A subset of global hstate attributes for node devices
3367 static struct attribute *per_node_hstate_attrs[] = {
3368 &nr_hugepages_attr.attr,
3369 &free_hugepages_attr.attr,
3370 &surplus_hugepages_attr.attr,
3374 static const struct attribute_group per_node_hstate_attr_group = {
3375 .attrs = per_node_hstate_attrs,
3379 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3380 * Returns node id via non-NULL nidp.
3382 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3386 for (nid = 0; nid < nr_node_ids; nid++) {
3387 struct node_hstate *nhs = &node_hstates[nid];
3389 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3390 if (nhs->hstate_kobjs[i] == kobj) {
3402 * Unregister hstate attributes from a single node device.
3403 * No-op if no hstate attributes attached.
3405 static void hugetlb_unregister_node(struct node *node)
3408 struct node_hstate *nhs = &node_hstates[node->dev.id];
3410 if (!nhs->hugepages_kobj)
3411 return; /* no hstate attributes */
3413 for_each_hstate(h) {
3414 int idx = hstate_index(h);
3415 if (nhs->hstate_kobjs[idx]) {
3416 kobject_put(nhs->hstate_kobjs[idx]);
3417 nhs->hstate_kobjs[idx] = NULL;
3421 kobject_put(nhs->hugepages_kobj);
3422 nhs->hugepages_kobj = NULL;
3427 * Register hstate attributes for a single node device.
3428 * No-op if attributes already registered.
3430 static void hugetlb_register_node(struct node *node)
3433 struct node_hstate *nhs = &node_hstates[node->dev.id];
3436 if (nhs->hugepages_kobj)
3437 return; /* already allocated */
3439 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3441 if (!nhs->hugepages_kobj)
3444 for_each_hstate(h) {
3445 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3447 &per_node_hstate_attr_group);
3449 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3450 h->name, node->dev.id);
3451 hugetlb_unregister_node(node);
3458 * hugetlb init time: register hstate attributes for all registered node
3459 * devices of nodes that have memory. All on-line nodes should have
3460 * registered their associated device by this time.
3462 static void __init hugetlb_register_all_nodes(void)
3466 for_each_node_state(nid, N_MEMORY) {
3467 struct node *node = node_devices[nid];
3468 if (node->dev.id == nid)
3469 hugetlb_register_node(node);
3473 * Let the node device driver know we're here so it can
3474 * [un]register hstate attributes on node hotplug.
3476 register_hugetlbfs_with_node(hugetlb_register_node,
3477 hugetlb_unregister_node);
3479 #else /* !CONFIG_NUMA */
3481 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3489 static void hugetlb_register_all_nodes(void) { }
3493 static int __init hugetlb_init(void)
3497 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3500 if (!hugepages_supported()) {
3501 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3502 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3507 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3508 * architectures depend on setup being done here.
3510 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3511 if (!parsed_default_hugepagesz) {
3513 * If we did not parse a default huge page size, set
3514 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3515 * number of huge pages for this default size was implicitly
3516 * specified, set that here as well.
3517 * Note that the implicit setting will overwrite an explicit
3518 * setting. A warning will be printed in this case.
3520 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3521 if (default_hstate_max_huge_pages) {
3522 if (default_hstate.max_huge_pages) {
3525 string_get_size(huge_page_size(&default_hstate),
3526 1, STRING_UNITS_2, buf, 32);
3527 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3528 default_hstate.max_huge_pages, buf);
3529 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3530 default_hstate_max_huge_pages);
3532 default_hstate.max_huge_pages =
3533 default_hstate_max_huge_pages;
3537 hugetlb_cma_check();
3538 hugetlb_init_hstates();
3539 gather_bootmem_prealloc();
3542 hugetlb_sysfs_init();
3543 hugetlb_register_all_nodes();
3544 hugetlb_cgroup_file_init();
3547 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3549 num_fault_mutexes = 1;
3551 hugetlb_fault_mutex_table =
3552 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3554 BUG_ON(!hugetlb_fault_mutex_table);
3556 for (i = 0; i < num_fault_mutexes; i++)
3557 mutex_init(&hugetlb_fault_mutex_table[i]);
3560 subsys_initcall(hugetlb_init);
3562 /* Overwritten by architectures with more huge page sizes */
3563 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3565 return size == HPAGE_SIZE;
3568 void __init hugetlb_add_hstate(unsigned int order)
3573 if (size_to_hstate(PAGE_SIZE << order)) {
3576 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3578 h = &hstates[hugetlb_max_hstate++];
3579 mutex_init(&h->resize_lock);
3581 h->mask = ~(huge_page_size(h) - 1);
3582 for (i = 0; i < MAX_NUMNODES; ++i)
3583 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3584 INIT_LIST_HEAD(&h->hugepage_activelist);
3585 h->next_nid_to_alloc = first_memory_node;
3586 h->next_nid_to_free = first_memory_node;
3587 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3588 huge_page_size(h)/1024);
3589 hugetlb_vmemmap_init(h);
3595 * hugepages command line processing
3596 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3597 * specification. If not, ignore the hugepages value. hugepages can also
3598 * be the first huge page command line option in which case it implicitly
3599 * specifies the number of huge pages for the default size.
3601 static int __init hugepages_setup(char *s)
3604 static unsigned long *last_mhp;
3606 if (!parsed_valid_hugepagesz) {
3607 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3608 parsed_valid_hugepagesz = true;
3613 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3614 * yet, so this hugepages= parameter goes to the "default hstate".
3615 * Otherwise, it goes with the previously parsed hugepagesz or
3616 * default_hugepagesz.
3618 else if (!hugetlb_max_hstate)
3619 mhp = &default_hstate_max_huge_pages;
3621 mhp = &parsed_hstate->max_huge_pages;
3623 if (mhp == last_mhp) {
3624 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3628 if (sscanf(s, "%lu", mhp) <= 0)
3632 * Global state is always initialized later in hugetlb_init.
3633 * But we need to allocate gigantic hstates here early to still
3634 * use the bootmem allocator.
3636 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3637 hugetlb_hstate_alloc_pages(parsed_hstate);
3643 __setup("hugepages=", hugepages_setup);
3646 * hugepagesz command line processing
3647 * A specific huge page size can only be specified once with hugepagesz.
3648 * hugepagesz is followed by hugepages on the command line. The global
3649 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3650 * hugepagesz argument was valid.
3652 static int __init hugepagesz_setup(char *s)
3657 parsed_valid_hugepagesz = false;
3658 size = (unsigned long)memparse(s, NULL);
3660 if (!arch_hugetlb_valid_size(size)) {
3661 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3665 h = size_to_hstate(size);
3668 * hstate for this size already exists. This is normally
3669 * an error, but is allowed if the existing hstate is the
3670 * default hstate. More specifically, it is only allowed if
3671 * the number of huge pages for the default hstate was not
3672 * previously specified.
3674 if (!parsed_default_hugepagesz || h != &default_hstate ||
3675 default_hstate.max_huge_pages) {
3676 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3681 * No need to call hugetlb_add_hstate() as hstate already
3682 * exists. But, do set parsed_hstate so that a following
3683 * hugepages= parameter will be applied to this hstate.
3686 parsed_valid_hugepagesz = true;
3690 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3691 parsed_valid_hugepagesz = true;
3694 __setup("hugepagesz=", hugepagesz_setup);
3697 * default_hugepagesz command line input
3698 * Only one instance of default_hugepagesz allowed on command line.
3700 static int __init default_hugepagesz_setup(char *s)
3704 parsed_valid_hugepagesz = false;
3705 if (parsed_default_hugepagesz) {
3706 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3710 size = (unsigned long)memparse(s, NULL);
3712 if (!arch_hugetlb_valid_size(size)) {
3713 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3717 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3718 parsed_valid_hugepagesz = true;
3719 parsed_default_hugepagesz = true;
3720 default_hstate_idx = hstate_index(size_to_hstate(size));
3723 * The number of default huge pages (for this size) could have been
3724 * specified as the first hugetlb parameter: hugepages=X. If so,
3725 * then default_hstate_max_huge_pages is set. If the default huge
3726 * page size is gigantic (>= MAX_ORDER), then the pages must be
3727 * allocated here from bootmem allocator.
3729 if (default_hstate_max_huge_pages) {
3730 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3731 if (hstate_is_gigantic(&default_hstate))
3732 hugetlb_hstate_alloc_pages(&default_hstate);
3733 default_hstate_max_huge_pages = 0;
3738 __setup("default_hugepagesz=", default_hugepagesz_setup);
3740 static unsigned int allowed_mems_nr(struct hstate *h)
3743 unsigned int nr = 0;
3744 nodemask_t *mpol_allowed;
3745 unsigned int *array = h->free_huge_pages_node;
3746 gfp_t gfp_mask = htlb_alloc_mask(h);
3748 mpol_allowed = policy_nodemask_current(gfp_mask);
3750 for_each_node_mask(node, cpuset_current_mems_allowed) {
3751 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3758 #ifdef CONFIG_SYSCTL
3759 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3760 void *buffer, size_t *length,
3761 loff_t *ppos, unsigned long *out)
3763 struct ctl_table dup_table;
3766 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3767 * can duplicate the @table and alter the duplicate of it.
3770 dup_table.data = out;
3772 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3775 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3776 struct ctl_table *table, int write,
3777 void *buffer, size_t *length, loff_t *ppos)
3779 struct hstate *h = &default_hstate;
3780 unsigned long tmp = h->max_huge_pages;
3783 if (!hugepages_supported())
3786 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3792 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3793 NUMA_NO_NODE, tmp, *length);
3798 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3799 void *buffer, size_t *length, loff_t *ppos)
3802 return hugetlb_sysctl_handler_common(false, table, write,
3803 buffer, length, ppos);
3807 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3808 void *buffer, size_t *length, loff_t *ppos)
3810 return hugetlb_sysctl_handler_common(true, table, write,
3811 buffer, length, ppos);
3813 #endif /* CONFIG_NUMA */
3815 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3816 void *buffer, size_t *length, loff_t *ppos)
3818 struct hstate *h = &default_hstate;
3822 if (!hugepages_supported())
3825 tmp = h->nr_overcommit_huge_pages;
3827 if (write && hstate_is_gigantic(h))
3830 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3836 spin_lock_irq(&hugetlb_lock);
3837 h->nr_overcommit_huge_pages = tmp;
3838 spin_unlock_irq(&hugetlb_lock);
3844 #endif /* CONFIG_SYSCTL */
3846 void hugetlb_report_meminfo(struct seq_file *m)
3849 unsigned long total = 0;
3851 if (!hugepages_supported())
3854 for_each_hstate(h) {
3855 unsigned long count = h->nr_huge_pages;
3857 total += huge_page_size(h) * count;
3859 if (h == &default_hstate)
3861 "HugePages_Total: %5lu\n"
3862 "HugePages_Free: %5lu\n"
3863 "HugePages_Rsvd: %5lu\n"
3864 "HugePages_Surp: %5lu\n"
3865 "Hugepagesize: %8lu kB\n",
3869 h->surplus_huge_pages,
3870 huge_page_size(h) / SZ_1K);
3873 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
3876 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3878 struct hstate *h = &default_hstate;
3880 if (!hugepages_supported())
3883 return sysfs_emit_at(buf, len,
3884 "Node %d HugePages_Total: %5u\n"
3885 "Node %d HugePages_Free: %5u\n"
3886 "Node %d HugePages_Surp: %5u\n",
3887 nid, h->nr_huge_pages_node[nid],
3888 nid, h->free_huge_pages_node[nid],
3889 nid, h->surplus_huge_pages_node[nid]);
3892 void hugetlb_show_meminfo(void)
3897 if (!hugepages_supported())
3900 for_each_node_state(nid, N_MEMORY)
3902 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3904 h->nr_huge_pages_node[nid],
3905 h->free_huge_pages_node[nid],
3906 h->surplus_huge_pages_node[nid],
3907 huge_page_size(h) / SZ_1K);
3910 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3912 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3913 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3916 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3917 unsigned long hugetlb_total_pages(void)
3920 unsigned long nr_total_pages = 0;
3923 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3924 return nr_total_pages;
3927 static int hugetlb_acct_memory(struct hstate *h, long delta)
3934 spin_lock_irq(&hugetlb_lock);
3936 * When cpuset is configured, it breaks the strict hugetlb page
3937 * reservation as the accounting is done on a global variable. Such
3938 * reservation is completely rubbish in the presence of cpuset because
3939 * the reservation is not checked against page availability for the
3940 * current cpuset. Application can still potentially OOM'ed by kernel
3941 * with lack of free htlb page in cpuset that the task is in.
3942 * Attempt to enforce strict accounting with cpuset is almost
3943 * impossible (or too ugly) because cpuset is too fluid that
3944 * task or memory node can be dynamically moved between cpusets.
3946 * The change of semantics for shared hugetlb mapping with cpuset is
3947 * undesirable. However, in order to preserve some of the semantics,
3948 * we fall back to check against current free page availability as
3949 * a best attempt and hopefully to minimize the impact of changing
3950 * semantics that cpuset has.
3952 * Apart from cpuset, we also have memory policy mechanism that
3953 * also determines from which node the kernel will allocate memory
3954 * in a NUMA system. So similar to cpuset, we also should consider
3955 * the memory policy of the current task. Similar to the description
3959 if (gather_surplus_pages(h, delta) < 0)
3962 if (delta > allowed_mems_nr(h)) {
3963 return_unused_surplus_pages(h, delta);
3970 return_unused_surplus_pages(h, (unsigned long) -delta);
3973 spin_unlock_irq(&hugetlb_lock);
3977 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3979 struct resv_map *resv = vma_resv_map(vma);
3982 * This new VMA should share its siblings reservation map if present.
3983 * The VMA will only ever have a valid reservation map pointer where
3984 * it is being copied for another still existing VMA. As that VMA
3985 * has a reference to the reservation map it cannot disappear until
3986 * after this open call completes. It is therefore safe to take a
3987 * new reference here without additional locking.
3989 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3990 kref_get(&resv->refs);
3993 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3995 struct hstate *h = hstate_vma(vma);
3996 struct resv_map *resv = vma_resv_map(vma);
3997 struct hugepage_subpool *spool = subpool_vma(vma);
3998 unsigned long reserve, start, end;
4001 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4004 start = vma_hugecache_offset(h, vma, vma->vm_start);
4005 end = vma_hugecache_offset(h, vma, vma->vm_end);
4007 reserve = (end - start) - region_count(resv, start, end);
4008 hugetlb_cgroup_uncharge_counter(resv, start, end);
4011 * Decrement reserve counts. The global reserve count may be
4012 * adjusted if the subpool has a minimum size.
4014 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4015 hugetlb_acct_memory(h, -gbl_reserve);
4018 kref_put(&resv->refs, resv_map_release);
4021 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4023 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4028 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4030 return huge_page_size(hstate_vma(vma));
4034 * We cannot handle pagefaults against hugetlb pages at all. They cause
4035 * handle_mm_fault() to try to instantiate regular-sized pages in the
4036 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4039 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4046 * When a new function is introduced to vm_operations_struct and added
4047 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4048 * This is because under System V memory model, mappings created via
4049 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4050 * their original vm_ops are overwritten with shm_vm_ops.
4052 const struct vm_operations_struct hugetlb_vm_ops = {
4053 .fault = hugetlb_vm_op_fault,
4054 .open = hugetlb_vm_op_open,
4055 .close = hugetlb_vm_op_close,
4056 .may_split = hugetlb_vm_op_split,
4057 .pagesize = hugetlb_vm_op_pagesize,
4060 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4064 unsigned int shift = huge_page_shift(hstate_vma(vma));
4067 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4068 vma->vm_page_prot)));
4070 entry = huge_pte_wrprotect(mk_huge_pte(page,
4071 vma->vm_page_prot));
4073 entry = pte_mkyoung(entry);
4074 entry = pte_mkhuge(entry);
4075 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4080 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4081 unsigned long address, pte_t *ptep)
4085 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4086 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4087 update_mmu_cache(vma, address, ptep);
4090 bool is_hugetlb_entry_migration(pte_t pte)
4094 if (huge_pte_none(pte) || pte_present(pte))
4096 swp = pte_to_swp_entry(pte);
4097 if (is_migration_entry(swp))
4103 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4107 if (huge_pte_none(pte) || pte_present(pte))
4109 swp = pte_to_swp_entry(pte);
4110 if (is_hwpoison_entry(swp))
4117 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4118 struct page *new_page)
4120 __SetPageUptodate(new_page);
4121 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4122 hugepage_add_new_anon_rmap(new_page, vma, addr);
4123 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4124 ClearHPageRestoreReserve(new_page);
4125 SetHPageMigratable(new_page);
4128 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4129 struct vm_area_struct *vma)
4131 pte_t *src_pte, *dst_pte, entry, dst_entry;
4132 struct page *ptepage;
4134 bool cow = is_cow_mapping(vma->vm_flags);
4135 struct hstate *h = hstate_vma(vma);
4136 unsigned long sz = huge_page_size(h);
4137 unsigned long npages = pages_per_huge_page(h);
4138 struct address_space *mapping = vma->vm_file->f_mapping;
4139 struct mmu_notifier_range range;
4143 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
4146 mmu_notifier_invalidate_range_start(&range);
4149 * For shared mappings i_mmap_rwsem must be held to call
4150 * huge_pte_alloc, otherwise the returned ptep could go
4151 * away if part of a shared pmd and another thread calls
4154 i_mmap_lock_read(mapping);
4157 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
4158 spinlock_t *src_ptl, *dst_ptl;
4159 src_pte = huge_pte_offset(src, addr, sz);
4162 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
4169 * If the pagetables are shared don't copy or take references.
4170 * dst_pte == src_pte is the common case of src/dest sharing.
4172 * However, src could have 'unshared' and dst shares with
4173 * another vma. If dst_pte !none, this implies sharing.
4174 * Check here before taking page table lock, and once again
4175 * after taking the lock below.
4177 dst_entry = huge_ptep_get(dst_pte);
4178 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4181 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4182 src_ptl = huge_pte_lockptr(h, src, src_pte);
4183 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4184 entry = huge_ptep_get(src_pte);
4185 dst_entry = huge_ptep_get(dst_pte);
4187 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4189 * Skip if src entry none. Also, skip in the
4190 * unlikely case dst entry !none as this implies
4191 * sharing with another vma.
4194 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4195 is_hugetlb_entry_hwpoisoned(entry))) {
4196 swp_entry_t swp_entry = pte_to_swp_entry(entry);
4198 if (is_write_migration_entry(swp_entry) && cow) {
4200 * COW mappings require pages in both
4201 * parent and child to be set to read.
4203 make_migration_entry_read(&swp_entry);
4204 entry = swp_entry_to_pte(swp_entry);
4205 set_huge_swap_pte_at(src, addr, src_pte,
4208 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4210 entry = huge_ptep_get(src_pte);
4211 ptepage = pte_page(entry);
4215 * This is a rare case where we see pinned hugetlb
4216 * pages while they're prone to COW. We need to do the
4217 * COW earlier during fork.
4219 * When pre-allocating the page or copying data, we
4220 * need to be without the pgtable locks since we could
4221 * sleep during the process.
4223 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4224 pte_t src_pte_old = entry;
4227 spin_unlock(src_ptl);
4228 spin_unlock(dst_ptl);
4229 /* Do not use reserve as it's private owned */
4230 new = alloc_huge_page(vma, addr, 1);
4236 copy_user_huge_page(new, ptepage, addr, vma,
4240 /* Install the new huge page if src pte stable */
4241 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4242 src_ptl = huge_pte_lockptr(h, src, src_pte);
4243 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4244 entry = huge_ptep_get(src_pte);
4245 if (!pte_same(src_pte_old, entry)) {
4246 restore_reserve_on_error(h, vma, addr,
4249 /* dst_entry won't change as in child */
4252 hugetlb_install_page(vma, dst_pte, addr, new);
4253 spin_unlock(src_ptl);
4254 spin_unlock(dst_ptl);
4260 * No need to notify as we are downgrading page
4261 * table protection not changing it to point
4264 * See Documentation/vm/mmu_notifier.rst
4266 huge_ptep_set_wrprotect(src, addr, src_pte);
4267 entry = huge_pte_wrprotect(entry);
4270 page_dup_rmap(ptepage, true);
4271 set_huge_pte_at(dst, addr, dst_pte, entry);
4272 hugetlb_count_add(npages, dst);
4274 spin_unlock(src_ptl);
4275 spin_unlock(dst_ptl);
4279 mmu_notifier_invalidate_range_end(&range);
4281 i_mmap_unlock_read(mapping);
4286 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4287 unsigned long start, unsigned long end,
4288 struct page *ref_page)
4290 struct mm_struct *mm = vma->vm_mm;
4291 unsigned long address;
4296 struct hstate *h = hstate_vma(vma);
4297 unsigned long sz = huge_page_size(h);
4298 struct mmu_notifier_range range;
4300 WARN_ON(!is_vm_hugetlb_page(vma));
4301 BUG_ON(start & ~huge_page_mask(h));
4302 BUG_ON(end & ~huge_page_mask(h));
4305 * This is a hugetlb vma, all the pte entries should point
4308 tlb_change_page_size(tlb, sz);
4309 tlb_start_vma(tlb, vma);
4312 * If sharing possible, alert mmu notifiers of worst case.
4314 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4316 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4317 mmu_notifier_invalidate_range_start(&range);
4319 for (; address < end; address += sz) {
4320 ptep = huge_pte_offset(mm, address, sz);
4324 ptl = huge_pte_lock(h, mm, ptep);
4325 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4328 * We just unmapped a page of PMDs by clearing a PUD.
4329 * The caller's TLB flush range should cover this area.
4334 pte = huge_ptep_get(ptep);
4335 if (huge_pte_none(pte)) {
4341 * Migrating hugepage or HWPoisoned hugepage is already
4342 * unmapped and its refcount is dropped, so just clear pte here.
4344 if (unlikely(!pte_present(pte))) {
4345 huge_pte_clear(mm, address, ptep, sz);
4350 page = pte_page(pte);
4352 * If a reference page is supplied, it is because a specific
4353 * page is being unmapped, not a range. Ensure the page we
4354 * are about to unmap is the actual page of interest.
4357 if (page != ref_page) {
4362 * Mark the VMA as having unmapped its page so that
4363 * future faults in this VMA will fail rather than
4364 * looking like data was lost
4366 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4369 pte = huge_ptep_get_and_clear(mm, address, ptep);
4370 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4371 if (huge_pte_dirty(pte))
4372 set_page_dirty(page);
4374 hugetlb_count_sub(pages_per_huge_page(h), mm);
4375 page_remove_rmap(page, true);
4378 tlb_remove_page_size(tlb, page, huge_page_size(h));
4380 * Bail out after unmapping reference page if supplied
4385 mmu_notifier_invalidate_range_end(&range);
4386 tlb_end_vma(tlb, vma);
4389 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4390 struct vm_area_struct *vma, unsigned long start,
4391 unsigned long end, struct page *ref_page)
4393 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4396 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4397 * test will fail on a vma being torn down, and not grab a page table
4398 * on its way out. We're lucky that the flag has such an appropriate
4399 * name, and can in fact be safely cleared here. We could clear it
4400 * before the __unmap_hugepage_range above, but all that's necessary
4401 * is to clear it before releasing the i_mmap_rwsem. This works
4402 * because in the context this is called, the VMA is about to be
4403 * destroyed and the i_mmap_rwsem is held.
4405 vma->vm_flags &= ~VM_MAYSHARE;
4408 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4409 unsigned long end, struct page *ref_page)
4411 struct mmu_gather tlb;
4413 tlb_gather_mmu(&tlb, vma->vm_mm);
4414 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4415 tlb_finish_mmu(&tlb);
4419 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4420 * mapping it owns the reserve page for. The intention is to unmap the page
4421 * from other VMAs and let the children be SIGKILLed if they are faulting the
4424 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4425 struct page *page, unsigned long address)
4427 struct hstate *h = hstate_vma(vma);
4428 struct vm_area_struct *iter_vma;
4429 struct address_space *mapping;
4433 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4434 * from page cache lookup which is in HPAGE_SIZE units.
4436 address = address & huge_page_mask(h);
4437 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4439 mapping = vma->vm_file->f_mapping;
4442 * Take the mapping lock for the duration of the table walk. As
4443 * this mapping should be shared between all the VMAs,
4444 * __unmap_hugepage_range() is called as the lock is already held
4446 i_mmap_lock_write(mapping);
4447 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4448 /* Do not unmap the current VMA */
4449 if (iter_vma == vma)
4453 * Shared VMAs have their own reserves and do not affect
4454 * MAP_PRIVATE accounting but it is possible that a shared
4455 * VMA is using the same page so check and skip such VMAs.
4457 if (iter_vma->vm_flags & VM_MAYSHARE)
4461 * Unmap the page from other VMAs without their own reserves.
4462 * They get marked to be SIGKILLed if they fault in these
4463 * areas. This is because a future no-page fault on this VMA
4464 * could insert a zeroed page instead of the data existing
4465 * from the time of fork. This would look like data corruption
4467 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4468 unmap_hugepage_range(iter_vma, address,
4469 address + huge_page_size(h), page);
4471 i_mmap_unlock_write(mapping);
4475 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4476 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4477 * cannot race with other handlers or page migration.
4478 * Keep the pte_same checks anyway to make transition from the mutex easier.
4480 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4481 unsigned long address, pte_t *ptep,
4482 struct page *pagecache_page, spinlock_t *ptl)
4485 struct hstate *h = hstate_vma(vma);
4486 struct page *old_page, *new_page;
4487 int outside_reserve = 0;
4489 unsigned long haddr = address & huge_page_mask(h);
4490 struct mmu_notifier_range range;
4492 pte = huge_ptep_get(ptep);
4493 old_page = pte_page(pte);
4496 /* If no-one else is actually using this page, avoid the copy
4497 * and just make the page writable */
4498 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4499 page_move_anon_rmap(old_page, vma);
4500 set_huge_ptep_writable(vma, haddr, ptep);
4505 * If the process that created a MAP_PRIVATE mapping is about to
4506 * perform a COW due to a shared page count, attempt to satisfy
4507 * the allocation without using the existing reserves. The pagecache
4508 * page is used to determine if the reserve at this address was
4509 * consumed or not. If reserves were used, a partial faulted mapping
4510 * at the time of fork() could consume its reserves on COW instead
4511 * of the full address range.
4513 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4514 old_page != pagecache_page)
4515 outside_reserve = 1;
4520 * Drop page table lock as buddy allocator may be called. It will
4521 * be acquired again before returning to the caller, as expected.
4524 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4526 if (IS_ERR(new_page)) {
4528 * If a process owning a MAP_PRIVATE mapping fails to COW,
4529 * it is due to references held by a child and an insufficient
4530 * huge page pool. To guarantee the original mappers
4531 * reliability, unmap the page from child processes. The child
4532 * may get SIGKILLed if it later faults.
4534 if (outside_reserve) {
4535 struct address_space *mapping = vma->vm_file->f_mapping;
4540 BUG_ON(huge_pte_none(pte));
4542 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4543 * unmapping. unmapping needs to hold i_mmap_rwsem
4544 * in write mode. Dropping i_mmap_rwsem in read mode
4545 * here is OK as COW mappings do not interact with
4548 * Reacquire both after unmap operation.
4550 idx = vma_hugecache_offset(h, vma, haddr);
4551 hash = hugetlb_fault_mutex_hash(mapping, idx);
4552 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4553 i_mmap_unlock_read(mapping);
4555 unmap_ref_private(mm, vma, old_page, haddr);
4557 i_mmap_lock_read(mapping);
4558 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4560 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4562 pte_same(huge_ptep_get(ptep), pte)))
4563 goto retry_avoidcopy;
4565 * race occurs while re-acquiring page table
4566 * lock, and our job is done.
4571 ret = vmf_error(PTR_ERR(new_page));
4572 goto out_release_old;
4576 * When the original hugepage is shared one, it does not have
4577 * anon_vma prepared.
4579 if (unlikely(anon_vma_prepare(vma))) {
4581 goto out_release_all;
4584 copy_user_huge_page(new_page, old_page, address, vma,
4585 pages_per_huge_page(h));
4586 __SetPageUptodate(new_page);
4588 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4589 haddr + huge_page_size(h));
4590 mmu_notifier_invalidate_range_start(&range);
4593 * Retake the page table lock to check for racing updates
4594 * before the page tables are altered
4597 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4598 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4599 ClearHPageRestoreReserve(new_page);
4602 huge_ptep_clear_flush(vma, haddr, ptep);
4603 mmu_notifier_invalidate_range(mm, range.start, range.end);
4604 set_huge_pte_at(mm, haddr, ptep,
4605 make_huge_pte(vma, new_page, 1));
4606 page_remove_rmap(old_page, true);
4607 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4608 SetHPageMigratable(new_page);
4609 /* Make the old page be freed below */
4610 new_page = old_page;
4613 mmu_notifier_invalidate_range_end(&range);
4615 restore_reserve_on_error(h, vma, haddr, new_page);
4620 spin_lock(ptl); /* Caller expects lock to be held */
4624 /* Return the pagecache page at a given address within a VMA */
4625 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4626 struct vm_area_struct *vma, unsigned long address)
4628 struct address_space *mapping;
4631 mapping = vma->vm_file->f_mapping;
4632 idx = vma_hugecache_offset(h, vma, address);
4634 return find_lock_page(mapping, idx);
4638 * Return whether there is a pagecache page to back given address within VMA.
4639 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4641 static bool hugetlbfs_pagecache_present(struct hstate *h,
4642 struct vm_area_struct *vma, unsigned long address)
4644 struct address_space *mapping;
4648 mapping = vma->vm_file->f_mapping;
4649 idx = vma_hugecache_offset(h, vma, address);
4651 page = find_get_page(mapping, idx);
4654 return page != NULL;
4657 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4660 struct inode *inode = mapping->host;
4661 struct hstate *h = hstate_inode(inode);
4662 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4666 ClearHPageRestoreReserve(page);
4669 * set page dirty so that it will not be removed from cache/file
4670 * by non-hugetlbfs specific code paths.
4672 set_page_dirty(page);
4674 spin_lock(&inode->i_lock);
4675 inode->i_blocks += blocks_per_huge_page(h);
4676 spin_unlock(&inode->i_lock);
4680 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
4681 struct address_space *mapping,
4684 unsigned long haddr,
4685 unsigned long reason)
4689 struct vm_fault vmf = {
4695 * Hard to debug if it ends up being
4696 * used by a callee that assumes
4697 * something about the other
4698 * uninitialized fields... same as in
4704 * hugetlb_fault_mutex and i_mmap_rwsem must be
4705 * dropped before handling userfault. Reacquire
4706 * after handling fault to make calling code simpler.
4708 hash = hugetlb_fault_mutex_hash(mapping, idx);
4709 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4710 i_mmap_unlock_read(mapping);
4711 ret = handle_userfault(&vmf, reason);
4712 i_mmap_lock_read(mapping);
4713 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4718 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4719 struct vm_area_struct *vma,
4720 struct address_space *mapping, pgoff_t idx,
4721 unsigned long address, pte_t *ptep, unsigned int flags)
4723 struct hstate *h = hstate_vma(vma);
4724 vm_fault_t ret = VM_FAULT_SIGBUS;
4730 unsigned long haddr = address & huge_page_mask(h);
4731 bool new_page = false;
4734 * Currently, we are forced to kill the process in the event the
4735 * original mapper has unmapped pages from the child due to a failed
4736 * COW. Warn that such a situation has occurred as it may not be obvious
4738 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4739 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4745 * We can not race with truncation due to holding i_mmap_rwsem.
4746 * i_size is modified when holding i_mmap_rwsem, so check here
4747 * once for faults beyond end of file.
4749 size = i_size_read(mapping->host) >> huge_page_shift(h);
4754 page = find_lock_page(mapping, idx);
4756 /* Check for page in userfault range */
4757 if (userfaultfd_missing(vma)) {
4758 ret = hugetlb_handle_userfault(vma, mapping, idx,
4764 page = alloc_huge_page(vma, haddr, 0);
4767 * Returning error will result in faulting task being
4768 * sent SIGBUS. The hugetlb fault mutex prevents two
4769 * tasks from racing to fault in the same page which
4770 * could result in false unable to allocate errors.
4771 * Page migration does not take the fault mutex, but
4772 * does a clear then write of pte's under page table
4773 * lock. Page fault code could race with migration,
4774 * notice the clear pte and try to allocate a page
4775 * here. Before returning error, get ptl and make
4776 * sure there really is no pte entry.
4778 ptl = huge_pte_lock(h, mm, ptep);
4780 if (huge_pte_none(huge_ptep_get(ptep)))
4781 ret = vmf_error(PTR_ERR(page));
4785 clear_huge_page(page, address, pages_per_huge_page(h));
4786 __SetPageUptodate(page);
4789 if (vma->vm_flags & VM_MAYSHARE) {
4790 int err = huge_add_to_page_cache(page, mapping, idx);
4799 if (unlikely(anon_vma_prepare(vma))) {
4801 goto backout_unlocked;
4807 * If memory error occurs between mmap() and fault, some process
4808 * don't have hwpoisoned swap entry for errored virtual address.
4809 * So we need to block hugepage fault by PG_hwpoison bit check.
4811 if (unlikely(PageHWPoison(page))) {
4812 ret = VM_FAULT_HWPOISON_LARGE |
4813 VM_FAULT_SET_HINDEX(hstate_index(h));
4814 goto backout_unlocked;
4817 /* Check for page in userfault range. */
4818 if (userfaultfd_minor(vma)) {
4821 ret = hugetlb_handle_userfault(vma, mapping, idx,
4829 * If we are going to COW a private mapping later, we examine the
4830 * pending reservations for this page now. This will ensure that
4831 * any allocations necessary to record that reservation occur outside
4834 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4835 if (vma_needs_reservation(h, vma, haddr) < 0) {
4837 goto backout_unlocked;
4839 /* Just decrements count, does not deallocate */
4840 vma_end_reservation(h, vma, haddr);
4843 ptl = huge_pte_lock(h, mm, ptep);
4845 if (!huge_pte_none(huge_ptep_get(ptep)))
4849 ClearHPageRestoreReserve(page);
4850 hugepage_add_new_anon_rmap(page, vma, haddr);
4852 page_dup_rmap(page, true);
4853 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4854 && (vma->vm_flags & VM_SHARED)));
4855 set_huge_pte_at(mm, haddr, ptep, new_pte);
4857 hugetlb_count_add(pages_per_huge_page(h), mm);
4858 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4859 /* Optimization, do the COW without a second fault */
4860 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4866 * Only set HPageMigratable in newly allocated pages. Existing pages
4867 * found in the pagecache may not have HPageMigratableset if they have
4868 * been isolated for migration.
4871 SetHPageMigratable(page);
4881 restore_reserve_on_error(h, vma, haddr, page);
4887 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4889 unsigned long key[2];
4892 key[0] = (unsigned long) mapping;
4895 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4897 return hash & (num_fault_mutexes - 1);
4901 * For uniprocessor systems we always use a single mutex, so just
4902 * return 0 and avoid the hashing overhead.
4904 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4910 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4911 unsigned long address, unsigned int flags)
4918 struct page *page = NULL;
4919 struct page *pagecache_page = NULL;
4920 struct hstate *h = hstate_vma(vma);
4921 struct address_space *mapping;
4922 int need_wait_lock = 0;
4923 unsigned long haddr = address & huge_page_mask(h);
4925 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4928 * Since we hold no locks, ptep could be stale. That is
4929 * OK as we are only making decisions based on content and
4930 * not actually modifying content here.
4932 entry = huge_ptep_get(ptep);
4933 if (unlikely(is_hugetlb_entry_migration(entry))) {
4934 migration_entry_wait_huge(vma, mm, ptep);
4936 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4937 return VM_FAULT_HWPOISON_LARGE |
4938 VM_FAULT_SET_HINDEX(hstate_index(h));
4942 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4943 * until finished with ptep. This serves two purposes:
4944 * 1) It prevents huge_pmd_unshare from being called elsewhere
4945 * and making the ptep no longer valid.
4946 * 2) It synchronizes us with i_size modifications during truncation.
4948 * ptep could have already be assigned via huge_pte_offset. That
4949 * is OK, as huge_pte_alloc will return the same value unless
4950 * something has changed.
4952 mapping = vma->vm_file->f_mapping;
4953 i_mmap_lock_read(mapping);
4954 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
4956 i_mmap_unlock_read(mapping);
4957 return VM_FAULT_OOM;
4961 * Serialize hugepage allocation and instantiation, so that we don't
4962 * get spurious allocation failures if two CPUs race to instantiate
4963 * the same page in the page cache.
4965 idx = vma_hugecache_offset(h, vma, haddr);
4966 hash = hugetlb_fault_mutex_hash(mapping, idx);
4967 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4969 entry = huge_ptep_get(ptep);
4970 if (huge_pte_none(entry)) {
4971 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4978 * entry could be a migration/hwpoison entry at this point, so this
4979 * check prevents the kernel from going below assuming that we have
4980 * an active hugepage in pagecache. This goto expects the 2nd page
4981 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4982 * properly handle it.
4984 if (!pte_present(entry))
4988 * If we are going to COW the mapping later, we examine the pending
4989 * reservations for this page now. This will ensure that any
4990 * allocations necessary to record that reservation occur outside the
4991 * spinlock. For private mappings, we also lookup the pagecache
4992 * page now as it is used to determine if a reservation has been
4995 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4996 if (vma_needs_reservation(h, vma, haddr) < 0) {
5000 /* Just decrements count, does not deallocate */
5001 vma_end_reservation(h, vma, haddr);
5003 if (!(vma->vm_flags & VM_MAYSHARE))
5004 pagecache_page = hugetlbfs_pagecache_page(h,
5008 ptl = huge_pte_lock(h, mm, ptep);
5010 /* Check for a racing update before calling hugetlb_cow */
5011 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5015 * hugetlb_cow() requires page locks of pte_page(entry) and
5016 * pagecache_page, so here we need take the former one
5017 * when page != pagecache_page or !pagecache_page.
5019 page = pte_page(entry);
5020 if (page != pagecache_page)
5021 if (!trylock_page(page)) {
5028 if (flags & FAULT_FLAG_WRITE) {
5029 if (!huge_pte_write(entry)) {
5030 ret = hugetlb_cow(mm, vma, address, ptep,
5031 pagecache_page, ptl);
5034 entry = huge_pte_mkdirty(entry);
5036 entry = pte_mkyoung(entry);
5037 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5038 flags & FAULT_FLAG_WRITE))
5039 update_mmu_cache(vma, haddr, ptep);
5041 if (page != pagecache_page)
5047 if (pagecache_page) {
5048 unlock_page(pagecache_page);
5049 put_page(pagecache_page);
5052 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5053 i_mmap_unlock_read(mapping);
5055 * Generally it's safe to hold refcount during waiting page lock. But
5056 * here we just wait to defer the next page fault to avoid busy loop and
5057 * the page is not used after unlocked before returning from the current
5058 * page fault. So we are safe from accessing freed page, even if we wait
5059 * here without taking refcount.
5062 wait_on_page_locked(page);
5066 #ifdef CONFIG_USERFAULTFD
5068 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5069 * modifications for huge pages.
5071 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5073 struct vm_area_struct *dst_vma,
5074 unsigned long dst_addr,
5075 unsigned long src_addr,
5076 enum mcopy_atomic_mode mode,
5077 struct page **pagep)
5079 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5080 struct hstate *h = hstate_vma(dst_vma);
5081 struct address_space *mapping = dst_vma->vm_file->f_mapping;
5082 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5084 int vm_shared = dst_vma->vm_flags & VM_SHARED;
5093 page = find_lock_page(mapping, idx);
5096 } else if (!*pagep) {
5097 /* If a page already exists, then it's UFFDIO_COPY for
5098 * a non-missing case. Return -EEXIST.
5101 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5106 page = alloc_huge_page(dst_vma, dst_addr, 0);
5112 ret = copy_huge_page_from_user(page,
5113 (const void __user *) src_addr,
5114 pages_per_huge_page(h), false);
5116 /* fallback to copy_from_user outside mmap_lock */
5117 if (unlikely(ret)) {
5119 /* Free the allocated page which may have
5120 * consumed a reservation.
5122 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5125 /* Allocate a temporary page to hold the copied
5128 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5134 /* Set the outparam pagep and return to the caller to
5135 * copy the contents outside the lock. Don't free the
5142 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5149 page = alloc_huge_page(dst_vma, dst_addr, 0);
5155 copy_huge_page(page, *pagep);
5161 * The memory barrier inside __SetPageUptodate makes sure that
5162 * preceding stores to the page contents become visible before
5163 * the set_pte_at() write.
5165 __SetPageUptodate(page);
5167 /* Add shared, newly allocated pages to the page cache. */
5168 if (vm_shared && !is_continue) {
5169 size = i_size_read(mapping->host) >> huge_page_shift(h);
5172 goto out_release_nounlock;
5175 * Serialization between remove_inode_hugepages() and
5176 * huge_add_to_page_cache() below happens through the
5177 * hugetlb_fault_mutex_table that here must be hold by
5180 ret = huge_add_to_page_cache(page, mapping, idx);
5182 goto out_release_nounlock;
5185 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5189 * Recheck the i_size after holding PT lock to make sure not
5190 * to leave any page mapped (as page_mapped()) beyond the end
5191 * of the i_size (remove_inode_hugepages() is strict about
5192 * enforcing that). If we bail out here, we'll also leave a
5193 * page in the radix tree in the vm_shared case beyond the end
5194 * of the i_size, but remove_inode_hugepages() will take care
5195 * of it as soon as we drop the hugetlb_fault_mutex_table.
5197 size = i_size_read(mapping->host) >> huge_page_shift(h);
5200 goto out_release_unlock;
5203 if (!huge_pte_none(huge_ptep_get(dst_pte)))
5204 goto out_release_unlock;
5207 page_dup_rmap(page, true);
5209 ClearHPageRestoreReserve(page);
5210 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5213 /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5214 if (is_continue && !vm_shared)
5217 writable = dst_vma->vm_flags & VM_WRITE;
5219 _dst_pte = make_huge_pte(dst_vma, page, writable);
5221 _dst_pte = huge_pte_mkdirty(_dst_pte);
5222 _dst_pte = pte_mkyoung(_dst_pte);
5224 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5226 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5227 dst_vma->vm_flags & VM_WRITE);
5228 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5230 /* No need to invalidate - it was non-present before */
5231 update_mmu_cache(dst_vma, dst_addr, dst_pte);
5235 SetHPageMigratable(page);
5236 if (vm_shared || is_continue)
5243 if (vm_shared || is_continue)
5245 out_release_nounlock:
5246 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5250 #endif /* CONFIG_USERFAULTFD */
5252 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5253 int refs, struct page **pages,
5254 struct vm_area_struct **vmas)
5258 for (nr = 0; nr < refs; nr++) {
5260 pages[nr] = mem_map_offset(page, nr);
5266 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5267 struct page **pages, struct vm_area_struct **vmas,
5268 unsigned long *position, unsigned long *nr_pages,
5269 long i, unsigned int flags, int *locked)
5271 unsigned long pfn_offset;
5272 unsigned long vaddr = *position;
5273 unsigned long remainder = *nr_pages;
5274 struct hstate *h = hstate_vma(vma);
5275 int err = -EFAULT, refs;
5277 while (vaddr < vma->vm_end && remainder) {
5279 spinlock_t *ptl = NULL;
5284 * If we have a pending SIGKILL, don't keep faulting pages and
5285 * potentially allocating memory.
5287 if (fatal_signal_pending(current)) {
5293 * Some archs (sparc64, sh*) have multiple pte_ts to
5294 * each hugepage. We have to make sure we get the
5295 * first, for the page indexing below to work.
5297 * Note that page table lock is not held when pte is null.
5299 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5302 ptl = huge_pte_lock(h, mm, pte);
5303 absent = !pte || huge_pte_none(huge_ptep_get(pte));
5306 * When coredumping, it suits get_dump_page if we just return
5307 * an error where there's an empty slot with no huge pagecache
5308 * to back it. This way, we avoid allocating a hugepage, and
5309 * the sparse dumpfile avoids allocating disk blocks, but its
5310 * huge holes still show up with zeroes where they need to be.
5312 if (absent && (flags & FOLL_DUMP) &&
5313 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
5321 * We need call hugetlb_fault for both hugepages under migration
5322 * (in which case hugetlb_fault waits for the migration,) and
5323 * hwpoisoned hugepages (in which case we need to prevent the
5324 * caller from accessing to them.) In order to do this, we use
5325 * here is_swap_pte instead of is_hugetlb_entry_migration and
5326 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5327 * both cases, and because we can't follow correct pages
5328 * directly from any kind of swap entries.
5330 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
5331 ((flags & FOLL_WRITE) &&
5332 !huge_pte_write(huge_ptep_get(pte)))) {
5334 unsigned int fault_flags = 0;
5338 if (flags & FOLL_WRITE)
5339 fault_flags |= FAULT_FLAG_WRITE;
5341 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5342 FAULT_FLAG_KILLABLE;
5343 if (flags & FOLL_NOWAIT)
5344 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5345 FAULT_FLAG_RETRY_NOWAIT;
5346 if (flags & FOLL_TRIED) {
5348 * Note: FAULT_FLAG_ALLOW_RETRY and
5349 * FAULT_FLAG_TRIED can co-exist
5351 fault_flags |= FAULT_FLAG_TRIED;
5353 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
5354 if (ret & VM_FAULT_ERROR) {
5355 err = vm_fault_to_errno(ret, flags);
5359 if (ret & VM_FAULT_RETRY) {
5361 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
5365 * VM_FAULT_RETRY must not return an
5366 * error, it will return zero
5369 * No need to update "position" as the
5370 * caller will not check it after
5371 * *nr_pages is set to 0.
5378 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
5379 page = pte_page(huge_ptep_get(pte));
5382 * If subpage information not requested, update counters
5383 * and skip the same_page loop below.
5385 if (!pages && !vmas && !pfn_offset &&
5386 (vaddr + huge_page_size(h) < vma->vm_end) &&
5387 (remainder >= pages_per_huge_page(h))) {
5388 vaddr += huge_page_size(h);
5389 remainder -= pages_per_huge_page(h);
5390 i += pages_per_huge_page(h);
5395 refs = min3(pages_per_huge_page(h) - pfn_offset,
5396 (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
5399 record_subpages_vmas(mem_map_offset(page, pfn_offset),
5401 likely(pages) ? pages + i : NULL,
5402 vmas ? vmas + i : NULL);
5406 * try_grab_compound_head() should always succeed here,
5407 * because: a) we hold the ptl lock, and b) we've just
5408 * checked that the huge page is present in the page
5409 * tables. If the huge page is present, then the tail
5410 * pages must also be present. The ptl prevents the
5411 * head page and tail pages from being rearranged in
5412 * any way. So this page must be available at this
5413 * point, unless the page refcount overflowed:
5415 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
5425 vaddr += (refs << PAGE_SHIFT);
5431 *nr_pages = remainder;
5433 * setting position is actually required only if remainder is
5434 * not zero but it's faster not to add a "if (remainder)"
5442 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5443 unsigned long address, unsigned long end, pgprot_t newprot)
5445 struct mm_struct *mm = vma->vm_mm;
5446 unsigned long start = address;
5449 struct hstate *h = hstate_vma(vma);
5450 unsigned long pages = 0;
5451 bool shared_pmd = false;
5452 struct mmu_notifier_range range;
5455 * In the case of shared PMDs, the area to flush could be beyond
5456 * start/end. Set range.start/range.end to cover the maximum possible
5457 * range if PMD sharing is possible.
5459 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5460 0, vma, mm, start, end);
5461 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5463 BUG_ON(address >= end);
5464 flush_cache_range(vma, range.start, range.end);
5466 mmu_notifier_invalidate_range_start(&range);
5467 i_mmap_lock_write(vma->vm_file->f_mapping);
5468 for (; address < end; address += huge_page_size(h)) {
5470 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5473 ptl = huge_pte_lock(h, mm, ptep);
5474 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5480 pte = huge_ptep_get(ptep);
5481 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5485 if (unlikely(is_hugetlb_entry_migration(pte))) {
5486 swp_entry_t entry = pte_to_swp_entry(pte);
5488 if (is_write_migration_entry(entry)) {
5491 make_migration_entry_read(&entry);
5492 newpte = swp_entry_to_pte(entry);
5493 set_huge_swap_pte_at(mm, address, ptep,
5494 newpte, huge_page_size(h));
5500 if (!huge_pte_none(pte)) {
5502 unsigned int shift = huge_page_shift(hstate_vma(vma));
5504 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5505 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5506 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
5507 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5513 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5514 * may have cleared our pud entry and done put_page on the page table:
5515 * once we release i_mmap_rwsem, another task can do the final put_page
5516 * and that page table be reused and filled with junk. If we actually
5517 * did unshare a page of pmds, flush the range corresponding to the pud.
5520 flush_hugetlb_tlb_range(vma, range.start, range.end);
5522 flush_hugetlb_tlb_range(vma, start, end);
5524 * No need to call mmu_notifier_invalidate_range() we are downgrading
5525 * page table protection not changing it to point to a new page.
5527 * See Documentation/vm/mmu_notifier.rst
5529 i_mmap_unlock_write(vma->vm_file->f_mapping);
5530 mmu_notifier_invalidate_range_end(&range);
5532 return pages << h->order;
5535 /* Return true if reservation was successful, false otherwise. */
5536 bool hugetlb_reserve_pages(struct inode *inode,
5538 struct vm_area_struct *vma,
5539 vm_flags_t vm_flags)
5542 struct hstate *h = hstate_inode(inode);
5543 struct hugepage_subpool *spool = subpool_inode(inode);
5544 struct resv_map *resv_map;
5545 struct hugetlb_cgroup *h_cg = NULL;
5546 long gbl_reserve, regions_needed = 0;
5548 /* This should never happen */
5550 VM_WARN(1, "%s called with a negative range\n", __func__);
5555 * Only apply hugepage reservation if asked. At fault time, an
5556 * attempt will be made for VM_NORESERVE to allocate a page
5557 * without using reserves
5559 if (vm_flags & VM_NORESERVE)
5563 * Shared mappings base their reservation on the number of pages that
5564 * are already allocated on behalf of the file. Private mappings need
5565 * to reserve the full area even if read-only as mprotect() may be
5566 * called to make the mapping read-write. Assume !vma is a shm mapping
5568 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5570 * resv_map can not be NULL as hugetlb_reserve_pages is only
5571 * called for inodes for which resv_maps were created (see
5572 * hugetlbfs_get_inode).
5574 resv_map = inode_resv_map(inode);
5576 chg = region_chg(resv_map, from, to, ®ions_needed);
5579 /* Private mapping. */
5580 resv_map = resv_map_alloc();
5586 set_vma_resv_map(vma, resv_map);
5587 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5593 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5594 chg * pages_per_huge_page(h), &h_cg) < 0)
5597 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5598 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5601 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5605 * There must be enough pages in the subpool for the mapping. If
5606 * the subpool has a minimum size, there may be some global
5607 * reservations already in place (gbl_reserve).
5609 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5610 if (gbl_reserve < 0)
5611 goto out_uncharge_cgroup;
5614 * Check enough hugepages are available for the reservation.
5615 * Hand the pages back to the subpool if there are not
5617 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5621 * Account for the reservations made. Shared mappings record regions
5622 * that have reservations as they are shared by multiple VMAs.
5623 * When the last VMA disappears, the region map says how much
5624 * the reservation was and the page cache tells how much of
5625 * the reservation was consumed. Private mappings are per-VMA and
5626 * only the consumed reservations are tracked. When the VMA
5627 * disappears, the original reservation is the VMA size and the
5628 * consumed reservations are stored in the map. Hence, nothing
5629 * else has to be done for private mappings here
5631 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5632 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5634 if (unlikely(add < 0)) {
5635 hugetlb_acct_memory(h, -gbl_reserve);
5637 } else if (unlikely(chg > add)) {
5639 * pages in this range were added to the reserve
5640 * map between region_chg and region_add. This
5641 * indicates a race with alloc_huge_page. Adjust
5642 * the subpool and reserve counts modified above
5643 * based on the difference.
5648 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5649 * reference to h_cg->css. See comment below for detail.
5651 hugetlb_cgroup_uncharge_cgroup_rsvd(
5653 (chg - add) * pages_per_huge_page(h), h_cg);
5655 rsv_adjust = hugepage_subpool_put_pages(spool,
5657 hugetlb_acct_memory(h, -rsv_adjust);
5660 * The file_regions will hold their own reference to
5661 * h_cg->css. So we should release the reference held
5662 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5665 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5671 /* put back original number of pages, chg */
5672 (void)hugepage_subpool_put_pages(spool, chg);
5673 out_uncharge_cgroup:
5674 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5675 chg * pages_per_huge_page(h), h_cg);
5677 if (!vma || vma->vm_flags & VM_MAYSHARE)
5678 /* Only call region_abort if the region_chg succeeded but the
5679 * region_add failed or didn't run.
5681 if (chg >= 0 && add < 0)
5682 region_abort(resv_map, from, to, regions_needed);
5683 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5684 kref_put(&resv_map->refs, resv_map_release);
5688 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5691 struct hstate *h = hstate_inode(inode);
5692 struct resv_map *resv_map = inode_resv_map(inode);
5694 struct hugepage_subpool *spool = subpool_inode(inode);
5698 * Since this routine can be called in the evict inode path for all
5699 * hugetlbfs inodes, resv_map could be NULL.
5702 chg = region_del(resv_map, start, end);
5704 * region_del() can fail in the rare case where a region
5705 * must be split and another region descriptor can not be
5706 * allocated. If end == LONG_MAX, it will not fail.
5712 spin_lock(&inode->i_lock);
5713 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5714 spin_unlock(&inode->i_lock);
5717 * If the subpool has a minimum size, the number of global
5718 * reservations to be released may be adjusted.
5720 * Note that !resv_map implies freed == 0. So (chg - freed)
5721 * won't go negative.
5723 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5724 hugetlb_acct_memory(h, -gbl_reserve);
5729 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5730 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5731 struct vm_area_struct *vma,
5732 unsigned long addr, pgoff_t idx)
5734 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5736 unsigned long sbase = saddr & PUD_MASK;
5737 unsigned long s_end = sbase + PUD_SIZE;
5739 /* Allow segments to share if only one is marked locked */
5740 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5741 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5744 * match the virtual addresses, permission and the alignment of the
5747 if (pmd_index(addr) != pmd_index(saddr) ||
5748 vm_flags != svm_flags ||
5749 !range_in_vma(svma, sbase, s_end))
5755 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5757 unsigned long base = addr & PUD_MASK;
5758 unsigned long end = base + PUD_SIZE;
5761 * check on proper vm_flags and page table alignment
5763 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5768 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5770 #ifdef CONFIG_USERFAULTFD
5771 if (uffd_disable_huge_pmd_share(vma))
5774 return vma_shareable(vma, addr);
5778 * Determine if start,end range within vma could be mapped by shared pmd.
5779 * If yes, adjust start and end to cover range associated with possible
5780 * shared pmd mappings.
5782 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5783 unsigned long *start, unsigned long *end)
5785 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5786 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5789 * vma needs to span at least one aligned PUD size, and the range
5790 * must be at least partially within in.
5792 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5793 (*end <= v_start) || (*start >= v_end))
5796 /* Extend the range to be PUD aligned for a worst case scenario */
5797 if (*start > v_start)
5798 *start = ALIGN_DOWN(*start, PUD_SIZE);
5801 *end = ALIGN(*end, PUD_SIZE);
5805 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5806 * and returns the corresponding pte. While this is not necessary for the
5807 * !shared pmd case because we can allocate the pmd later as well, it makes the
5808 * code much cleaner.
5810 * This routine must be called with i_mmap_rwsem held in at least read mode if
5811 * sharing is possible. For hugetlbfs, this prevents removal of any page
5812 * table entries associated with the address space. This is important as we
5813 * are setting up sharing based on existing page table entries (mappings).
5815 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5816 * huge_pte_alloc know that sharing is not possible and do not take
5817 * i_mmap_rwsem as a performance optimization. This is handled by the
5818 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5819 * only required for subsequent processing.
5821 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5822 unsigned long addr, pud_t *pud)
5824 struct address_space *mapping = vma->vm_file->f_mapping;
5825 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5827 struct vm_area_struct *svma;
5828 unsigned long saddr;
5833 i_mmap_assert_locked(mapping);
5834 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5838 saddr = page_table_shareable(svma, vma, addr, idx);
5840 spte = huge_pte_offset(svma->vm_mm, saddr,
5841 vma_mmu_pagesize(svma));
5843 get_page(virt_to_page(spte));
5852 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5853 if (pud_none(*pud)) {
5854 pud_populate(mm, pud,
5855 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5858 put_page(virt_to_page(spte));
5862 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5867 * unmap huge page backed by shared pte.
5869 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5870 * indicated by page_count > 1, unmap is achieved by clearing pud and
5871 * decrementing the ref count. If count == 1, the pte page is not shared.
5873 * Called with page table lock held and i_mmap_rwsem held in write mode.
5875 * returns: 1 successfully unmapped a shared pte page
5876 * 0 the underlying pte page is not shared, or it is the last user
5878 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5879 unsigned long *addr, pte_t *ptep)
5881 pgd_t *pgd = pgd_offset(mm, *addr);
5882 p4d_t *p4d = p4d_offset(pgd, *addr);
5883 pud_t *pud = pud_offset(p4d, *addr);
5885 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5886 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5887 if (page_count(virt_to_page(ptep)) == 1)
5891 put_page(virt_to_page(ptep));
5893 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5897 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5898 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5899 unsigned long addr, pud_t *pud)
5904 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5905 unsigned long *addr, pte_t *ptep)
5910 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5911 unsigned long *start, unsigned long *end)
5915 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5919 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5921 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5922 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
5923 unsigned long addr, unsigned long sz)
5930 pgd = pgd_offset(mm, addr);
5931 p4d = p4d_alloc(mm, pgd, addr);
5934 pud = pud_alloc(mm, p4d, addr);
5936 if (sz == PUD_SIZE) {
5939 BUG_ON(sz != PMD_SIZE);
5940 if (want_pmd_share(vma, addr) && pud_none(*pud))
5941 pte = huge_pmd_share(mm, vma, addr, pud);
5943 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5946 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5952 * huge_pte_offset() - Walk the page table to resolve the hugepage
5953 * entry at address @addr
5955 * Return: Pointer to page table entry (PUD or PMD) for
5956 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5957 * size @sz doesn't match the hugepage size at this level of the page
5960 pte_t *huge_pte_offset(struct mm_struct *mm,
5961 unsigned long addr, unsigned long sz)
5968 pgd = pgd_offset(mm, addr);
5969 if (!pgd_present(*pgd))
5971 p4d = p4d_offset(pgd, addr);
5972 if (!p4d_present(*p4d))
5975 pud = pud_offset(p4d, addr);
5977 /* must be pud huge, non-present or none */
5978 return (pte_t *)pud;
5979 if (!pud_present(*pud))
5981 /* must have a valid entry and size to go further */
5983 pmd = pmd_offset(pud, addr);
5984 /* must be pmd huge, non-present or none */
5985 return (pte_t *)pmd;
5988 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5991 * These functions are overwritable if your architecture needs its own
5994 struct page * __weak
5995 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5998 return ERR_PTR(-EINVAL);
6001 struct page * __weak
6002 follow_huge_pd(struct vm_area_struct *vma,
6003 unsigned long address, hugepd_t hpd, int flags, int pdshift)
6005 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6009 struct page * __weak
6010 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6011 pmd_t *pmd, int flags)
6013 struct page *page = NULL;
6017 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6018 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6019 (FOLL_PIN | FOLL_GET)))
6023 ptl = pmd_lockptr(mm, pmd);
6026 * make sure that the address range covered by this pmd is not
6027 * unmapped from other threads.
6029 if (!pmd_huge(*pmd))
6031 pte = huge_ptep_get((pte_t *)pmd);
6032 if (pte_present(pte)) {
6033 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6035 * try_grab_page() should always succeed here, because: a) we
6036 * hold the pmd (ptl) lock, and b) we've just checked that the
6037 * huge pmd (head) page is present in the page tables. The ptl
6038 * prevents the head page and tail pages from being rearranged
6039 * in any way. So this page must be available at this point,
6040 * unless the page refcount overflowed:
6042 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6047 if (is_hugetlb_entry_migration(pte)) {
6049 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
6053 * hwpoisoned entry is treated as no_page_table in
6054 * follow_page_mask().
6062 struct page * __weak
6063 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6064 pud_t *pud, int flags)
6066 if (flags & (FOLL_GET | FOLL_PIN))
6069 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6072 struct page * __weak
6073 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6075 if (flags & (FOLL_GET | FOLL_PIN))
6078 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6081 bool isolate_huge_page(struct page *page, struct list_head *list)
6085 spin_lock_irq(&hugetlb_lock);
6086 if (!PageHeadHuge(page) ||
6087 !HPageMigratable(page) ||
6088 !get_page_unless_zero(page)) {
6092 ClearHPageMigratable(page);
6093 list_move_tail(&page->lru, list);
6095 spin_unlock_irq(&hugetlb_lock);
6099 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6104 spin_lock_irq(&hugetlb_lock);
6105 if (PageHeadHuge(page)) {
6107 if (HPageFreed(page) || HPageMigratable(page))
6108 ret = get_page_unless_zero(page);
6112 spin_unlock_irq(&hugetlb_lock);
6116 void putback_active_hugepage(struct page *page)
6118 spin_lock_irq(&hugetlb_lock);
6119 SetHPageMigratable(page);
6120 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
6121 spin_unlock_irq(&hugetlb_lock);
6125 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6127 struct hstate *h = page_hstate(oldpage);
6129 hugetlb_cgroup_migrate(oldpage, newpage);
6130 set_page_owner_migrate_reason(newpage, reason);
6133 * transfer temporary state of the new huge page. This is
6134 * reverse to other transitions because the newpage is going to
6135 * be final while the old one will be freed so it takes over
6136 * the temporary status.
6138 * Also note that we have to transfer the per-node surplus state
6139 * here as well otherwise the global surplus count will not match
6142 if (HPageTemporary(newpage)) {
6143 int old_nid = page_to_nid(oldpage);
6144 int new_nid = page_to_nid(newpage);
6146 SetHPageTemporary(oldpage);
6147 ClearHPageTemporary(newpage);
6150 * There is no need to transfer the per-node surplus state
6151 * when we do not cross the node.
6153 if (new_nid == old_nid)
6155 spin_lock_irq(&hugetlb_lock);
6156 if (h->surplus_huge_pages_node[old_nid]) {
6157 h->surplus_huge_pages_node[old_nid]--;
6158 h->surplus_huge_pages_node[new_nid]++;
6160 spin_unlock_irq(&hugetlb_lock);
6165 * This function will unconditionally remove all the shared pmd pgtable entries
6166 * within the specific vma for a hugetlbfs memory range.
6168 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6170 struct hstate *h = hstate_vma(vma);
6171 unsigned long sz = huge_page_size(h);
6172 struct mm_struct *mm = vma->vm_mm;
6173 struct mmu_notifier_range range;
6174 unsigned long address, start, end;
6178 if (!(vma->vm_flags & VM_MAYSHARE))
6181 start = ALIGN(vma->vm_start, PUD_SIZE);
6182 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6188 * No need to call adjust_range_if_pmd_sharing_possible(), because
6189 * we have already done the PUD_SIZE alignment.
6191 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6193 mmu_notifier_invalidate_range_start(&range);
6194 i_mmap_lock_write(vma->vm_file->f_mapping);
6195 for (address = start; address < end; address += PUD_SIZE) {
6196 unsigned long tmp = address;
6198 ptep = huge_pte_offset(mm, address, sz);
6201 ptl = huge_pte_lock(h, mm, ptep);
6202 /* We don't want 'address' to be changed */
6203 huge_pmd_unshare(mm, vma, &tmp, ptep);
6206 flush_hugetlb_tlb_range(vma, start, end);
6207 i_mmap_unlock_write(vma->vm_file->f_mapping);
6209 * No need to call mmu_notifier_invalidate_range(), see
6210 * Documentation/vm/mmu_notifier.rst.
6212 mmu_notifier_invalidate_range_end(&range);
6216 static bool cma_reserve_called __initdata;
6218 static int __init cmdline_parse_hugetlb_cma(char *p)
6220 hugetlb_cma_size = memparse(p, &p);
6224 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6226 void __init hugetlb_cma_reserve(int order)
6228 unsigned long size, reserved, per_node;
6231 cma_reserve_called = true;
6233 if (!hugetlb_cma_size)
6236 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6237 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6238 (PAGE_SIZE << order) / SZ_1M);
6243 * If 3 GB area is requested on a machine with 4 numa nodes,
6244 * let's allocate 1 GB on first three nodes and ignore the last one.
6246 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6247 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6248 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6251 for_each_node_state(nid, N_ONLINE) {
6253 char name[CMA_MAX_NAME];
6255 size = min(per_node, hugetlb_cma_size - reserved);
6256 size = round_up(size, PAGE_SIZE << order);
6258 snprintf(name, sizeof(name), "hugetlb%d", nid);
6259 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
6261 &hugetlb_cma[nid], nid);
6263 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
6269 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
6272 if (reserved >= hugetlb_cma_size)
6277 void __init hugetlb_cma_check(void)
6279 if (!hugetlb_cma_size || cma_reserve_called)
6282 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
6285 #endif /* CONFIG_CMA */