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 VM_BUG_ON_PAGE(page_count(page), page);
1077 list_move(&page->lru, &h->hugepage_freelists[nid]);
1078 h->free_huge_pages++;
1079 h->free_huge_pages_node[nid]++;
1080 SetHPageFreed(page);
1083 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1086 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1088 lockdep_assert_held(&hugetlb_lock);
1089 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1090 if (pin && !is_pinnable_page(page))
1093 if (PageHWPoison(page))
1096 list_move(&page->lru, &h->hugepage_activelist);
1097 set_page_refcounted(page);
1098 ClearHPageFreed(page);
1099 h->free_huge_pages--;
1100 h->free_huge_pages_node[nid]--;
1107 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1110 unsigned int cpuset_mems_cookie;
1111 struct zonelist *zonelist;
1114 int node = NUMA_NO_NODE;
1116 zonelist = node_zonelist(nid, gfp_mask);
1119 cpuset_mems_cookie = read_mems_allowed_begin();
1120 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1123 if (!cpuset_zone_allowed(zone, gfp_mask))
1126 * no need to ask again on the same node. Pool is node rather than
1129 if (zone_to_nid(zone) == node)
1131 node = zone_to_nid(zone);
1133 page = dequeue_huge_page_node_exact(h, node);
1137 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1143 static struct page *dequeue_huge_page_vma(struct hstate *h,
1144 struct vm_area_struct *vma,
1145 unsigned long address, int avoid_reserve,
1149 struct mempolicy *mpol;
1151 nodemask_t *nodemask;
1155 * A child process with MAP_PRIVATE mappings created by their parent
1156 * have no page reserves. This check ensures that reservations are
1157 * not "stolen". The child may still get SIGKILLed
1159 if (!vma_has_reserves(vma, chg) &&
1160 h->free_huge_pages - h->resv_huge_pages == 0)
1163 /* If reserves cannot be used, ensure enough pages are in the pool */
1164 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1167 gfp_mask = htlb_alloc_mask(h);
1168 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1169 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1170 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1171 SetHPageRestoreReserve(page);
1172 h->resv_huge_pages--;
1175 mpol_cond_put(mpol);
1183 * common helper functions for hstate_next_node_to_{alloc|free}.
1184 * We may have allocated or freed a huge page based on a different
1185 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1186 * be outside of *nodes_allowed. Ensure that we use an allowed
1187 * node for alloc or free.
1189 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1191 nid = next_node_in(nid, *nodes_allowed);
1192 VM_BUG_ON(nid >= MAX_NUMNODES);
1197 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1199 if (!node_isset(nid, *nodes_allowed))
1200 nid = next_node_allowed(nid, nodes_allowed);
1205 * returns the previously saved node ["this node"] from which to
1206 * allocate a persistent huge page for the pool and advance the
1207 * next node from which to allocate, handling wrap at end of node
1210 static int hstate_next_node_to_alloc(struct hstate *h,
1211 nodemask_t *nodes_allowed)
1215 VM_BUG_ON(!nodes_allowed);
1217 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1218 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1224 * helper for remove_pool_huge_page() - return the previously saved
1225 * node ["this node"] from which to free a huge page. Advance the
1226 * next node id whether or not we find a free huge page to free so
1227 * that the next attempt to free addresses the next node.
1229 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1233 VM_BUG_ON(!nodes_allowed);
1235 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1236 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1241 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1242 for (nr_nodes = nodes_weight(*mask); \
1244 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1247 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1248 for (nr_nodes = nodes_weight(*mask); \
1250 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1253 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1254 static void destroy_compound_gigantic_page(struct page *page,
1258 int nr_pages = 1 << order;
1259 struct page *p = page + 1;
1261 atomic_set(compound_mapcount_ptr(page), 0);
1262 atomic_set(compound_pincount_ptr(page), 0);
1264 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1265 clear_compound_head(p);
1266 set_page_refcounted(p);
1269 set_compound_order(page, 0);
1270 page[1].compound_nr = 0;
1271 __ClearPageHead(page);
1274 static void free_gigantic_page(struct page *page, unsigned int order)
1277 * If the page isn't allocated using the cma allocator,
1278 * cma_release() returns false.
1281 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1285 free_contig_range(page_to_pfn(page), 1 << order);
1288 #ifdef CONFIG_CONTIG_ALLOC
1289 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1290 int nid, nodemask_t *nodemask)
1292 unsigned long nr_pages = pages_per_huge_page(h);
1293 if (nid == NUMA_NO_NODE)
1294 nid = numa_mem_id();
1301 if (hugetlb_cma[nid]) {
1302 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1303 huge_page_order(h), true);
1308 if (!(gfp_mask & __GFP_THISNODE)) {
1309 for_each_node_mask(node, *nodemask) {
1310 if (node == nid || !hugetlb_cma[node])
1313 page = cma_alloc(hugetlb_cma[node], nr_pages,
1314 huge_page_order(h), true);
1322 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
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]--;
1376 * For non-gigantic pages set the destructor to the normal compound
1377 * page dtor. This is needed in case someone takes an additional
1378 * temporary ref to the page, and freeing is delayed until they drop
1381 * For gigantic pages set the destructor to the null dtor. This
1382 * destructor will never be called. Before freeing the gigantic
1383 * page destroy_compound_gigantic_page will turn the compound page
1384 * into a simple group of pages. After this the destructor does not
1387 * This handles the case where more than one ref is held when and
1388 * after update_and_free_page is called.
1390 set_page_refcounted(page);
1391 if (hstate_is_gigantic(h))
1392 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1394 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
1397 h->nr_huge_pages_node[nid]--;
1400 static void add_hugetlb_page(struct hstate *h, struct page *page,
1401 bool adjust_surplus)
1404 int nid = page_to_nid(page);
1406 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1408 lockdep_assert_held(&hugetlb_lock);
1410 INIT_LIST_HEAD(&page->lru);
1412 h->nr_huge_pages_node[nid]++;
1414 if (adjust_surplus) {
1415 h->surplus_huge_pages++;
1416 h->surplus_huge_pages_node[nid]++;
1419 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1420 set_page_private(page, 0);
1421 SetHPageVmemmapOptimized(page);
1424 * This page is about to be managed by the hugetlb allocator and
1425 * should have no users. Drop our reference, and check for others
1428 zeroed = put_page_testzero(page);
1431 * It is VERY unlikely soneone else has taken a ref on
1432 * the page. In this case, we simply return as the
1433 * hugetlb destructor (free_huge_page) will be called
1434 * when this other ref is dropped.
1438 arch_clear_hugepage_flags(page);
1439 enqueue_huge_page(h, page);
1442 static void __update_and_free_page(struct hstate *h, struct page *page)
1445 struct page *subpage = page;
1447 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1450 if (alloc_huge_page_vmemmap(h, page)) {
1451 spin_lock_irq(&hugetlb_lock);
1453 * If we cannot allocate vmemmap pages, just refuse to free the
1454 * page and put the page back on the hugetlb free list and treat
1455 * as a surplus page.
1457 add_hugetlb_page(h, page, true);
1458 spin_unlock_irq(&hugetlb_lock);
1462 for (i = 0; i < pages_per_huge_page(h);
1463 i++, subpage = mem_map_next(subpage, page, i)) {
1464 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1465 1 << PG_referenced | 1 << PG_dirty |
1466 1 << PG_active | 1 << PG_private |
1469 if (hstate_is_gigantic(h)) {
1470 destroy_compound_gigantic_page(page, huge_page_order(h));
1471 free_gigantic_page(page, huge_page_order(h));
1473 __free_pages(page, huge_page_order(h));
1478 * As update_and_free_page() can be called under any context, so we cannot
1479 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1480 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1481 * the vmemmap pages.
1483 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1484 * freed and frees them one-by-one. As the page->mapping pointer is going
1485 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1486 * structure of a lockless linked list of huge pages to be freed.
1488 static LLIST_HEAD(hpage_freelist);
1490 static void free_hpage_workfn(struct work_struct *work)
1492 struct llist_node *node;
1494 node = llist_del_all(&hpage_freelist);
1500 page = container_of((struct address_space **)node,
1501 struct page, mapping);
1503 page->mapping = NULL;
1505 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1506 * is going to trigger because a previous call to
1507 * remove_hugetlb_page() will set_compound_page_dtor(page,
1508 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1510 h = size_to_hstate(page_size(page));
1512 __update_and_free_page(h, page);
1517 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1519 static inline void flush_free_hpage_work(struct hstate *h)
1521 if (free_vmemmap_pages_per_hpage(h))
1522 flush_work(&free_hpage_work);
1525 static void update_and_free_page(struct hstate *h, struct page *page,
1528 if (!HPageVmemmapOptimized(page) || !atomic) {
1529 __update_and_free_page(h, page);
1534 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1536 * Only call schedule_work() if hpage_freelist is previously
1537 * empty. Otherwise, schedule_work() had been called but the workfn
1538 * hasn't retrieved the list yet.
1540 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1541 schedule_work(&free_hpage_work);
1544 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1546 struct page *page, *t_page;
1548 list_for_each_entry_safe(page, t_page, list, lru) {
1549 update_and_free_page(h, page, false);
1554 struct hstate *size_to_hstate(unsigned long size)
1558 for_each_hstate(h) {
1559 if (huge_page_size(h) == size)
1565 void free_huge_page(struct page *page)
1568 * Can't pass hstate in here because it is called from the
1569 * compound page destructor.
1571 struct hstate *h = page_hstate(page);
1572 int nid = page_to_nid(page);
1573 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1574 bool restore_reserve;
1575 unsigned long flags;
1577 VM_BUG_ON_PAGE(page_count(page), page);
1578 VM_BUG_ON_PAGE(page_mapcount(page), page);
1580 hugetlb_set_page_subpool(page, NULL);
1581 page->mapping = NULL;
1582 restore_reserve = HPageRestoreReserve(page);
1583 ClearHPageRestoreReserve(page);
1586 * If HPageRestoreReserve was set on page, page allocation consumed a
1587 * reservation. If the page was associated with a subpool, there
1588 * would have been a page reserved in the subpool before allocation
1589 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1590 * reservation, do not call hugepage_subpool_put_pages() as this will
1591 * remove the reserved page from the subpool.
1593 if (!restore_reserve) {
1595 * A return code of zero implies that the subpool will be
1596 * under its minimum size if the reservation is not restored
1597 * after page is free. Therefore, force restore_reserve
1600 if (hugepage_subpool_put_pages(spool, 1) == 0)
1601 restore_reserve = true;
1604 spin_lock_irqsave(&hugetlb_lock, flags);
1605 ClearHPageMigratable(page);
1606 hugetlb_cgroup_uncharge_page(hstate_index(h),
1607 pages_per_huge_page(h), page);
1608 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1609 pages_per_huge_page(h), page);
1610 if (restore_reserve)
1611 h->resv_huge_pages++;
1613 if (HPageTemporary(page)) {
1614 remove_hugetlb_page(h, page, false);
1615 spin_unlock_irqrestore(&hugetlb_lock, flags);
1616 update_and_free_page(h, page, true);
1617 } else if (h->surplus_huge_pages_node[nid]) {
1618 /* remove the page from active list */
1619 remove_hugetlb_page(h, page, true);
1620 spin_unlock_irqrestore(&hugetlb_lock, flags);
1621 update_and_free_page(h, page, true);
1623 arch_clear_hugepage_flags(page);
1624 enqueue_huge_page(h, page);
1625 spin_unlock_irqrestore(&hugetlb_lock, flags);
1630 * Must be called with the hugetlb lock held
1632 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1634 lockdep_assert_held(&hugetlb_lock);
1636 h->nr_huge_pages_node[nid]++;
1639 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1641 free_huge_page_vmemmap(h, page);
1642 INIT_LIST_HEAD(&page->lru);
1643 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1644 hugetlb_set_page_subpool(page, NULL);
1645 set_hugetlb_cgroup(page, NULL);
1646 set_hugetlb_cgroup_rsvd(page, NULL);
1649 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1651 __prep_new_huge_page(h, page);
1652 spin_lock_irq(&hugetlb_lock);
1653 __prep_account_new_huge_page(h, nid);
1654 spin_unlock_irq(&hugetlb_lock);
1657 static bool prep_compound_gigantic_page(struct page *page, unsigned int order)
1660 int nr_pages = 1 << order;
1661 struct page *p = page + 1;
1663 /* we rely on prep_new_huge_page to set the destructor */
1664 set_compound_order(page, order);
1665 __ClearPageReserved(page);
1666 __SetPageHead(page);
1667 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1669 * For gigantic hugepages allocated through bootmem at
1670 * boot, it's safer to be consistent with the not-gigantic
1671 * hugepages and clear the PG_reserved bit from all tail pages
1672 * too. Otherwise drivers using get_user_pages() to access tail
1673 * pages may get the reference counting wrong if they see
1674 * PG_reserved set on a tail page (despite the head page not
1675 * having PG_reserved set). Enforcing this consistency between
1676 * head and tail pages allows drivers to optimize away a check
1677 * on the head page when they need know if put_page() is needed
1678 * after get_user_pages().
1680 __ClearPageReserved(p);
1682 * Subtle and very unlikely
1684 * Gigantic 'page allocators' such as memblock or cma will
1685 * return a set of pages with each page ref counted. We need
1686 * to turn this set of pages into a compound page with tail
1687 * page ref counts set to zero. Code such as speculative page
1688 * cache adding could take a ref on a 'to be' tail page.
1689 * We need to respect any increased ref count, and only set
1690 * the ref count to zero if count is currently 1. If count
1691 * is not 1, we return an error. An error return indicates
1692 * the set of pages can not be converted to a gigantic page.
1693 * The caller who allocated the pages should then discard the
1694 * pages using the appropriate free interface.
1696 if (!page_ref_freeze(p, 1)) {
1697 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
1700 set_page_count(p, 0);
1701 set_compound_head(p, page);
1703 atomic_set(compound_mapcount_ptr(page), -1);
1704 atomic_set(compound_pincount_ptr(page), 0);
1708 /* undo tail page modifications made above */
1710 for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) {
1711 clear_compound_head(p);
1712 set_page_refcounted(p);
1714 /* need to clear PG_reserved on remaining tail pages */
1715 for (; j < nr_pages; j++, p = mem_map_next(p, page, j))
1716 __ClearPageReserved(p);
1717 set_compound_order(page, 0);
1718 page[1].compound_nr = 0;
1719 __ClearPageHead(page);
1724 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1725 * transparent huge pages. See the PageTransHuge() documentation for more
1728 int PageHuge(struct page *page)
1730 if (!PageCompound(page))
1733 page = compound_head(page);
1734 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1736 EXPORT_SYMBOL_GPL(PageHuge);
1739 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1740 * normal or transparent huge pages.
1742 int PageHeadHuge(struct page *page_head)
1744 if (!PageHead(page_head))
1747 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1751 * Find and lock address space (mapping) in write mode.
1753 * Upon entry, the page is locked which means that page_mapping() is
1754 * stable. Due to locking order, we can only trylock_write. If we can
1755 * not get the lock, simply return NULL to caller.
1757 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1759 struct address_space *mapping = page_mapping(hpage);
1764 if (i_mmap_trylock_write(mapping))
1770 pgoff_t hugetlb_basepage_index(struct page *page)
1772 struct page *page_head = compound_head(page);
1773 pgoff_t index = page_index(page_head);
1774 unsigned long compound_idx;
1776 if (compound_order(page_head) >= MAX_ORDER)
1777 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1779 compound_idx = page - page_head;
1781 return (index << compound_order(page_head)) + compound_idx;
1784 static struct page *alloc_buddy_huge_page(struct hstate *h,
1785 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1786 nodemask_t *node_alloc_noretry)
1788 int order = huge_page_order(h);
1790 bool alloc_try_hard = true;
1793 * By default we always try hard to allocate the page with
1794 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1795 * a loop (to adjust global huge page counts) and previous allocation
1796 * failed, do not continue to try hard on the same node. Use the
1797 * node_alloc_noretry bitmap to manage this state information.
1799 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1800 alloc_try_hard = false;
1801 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1803 gfp_mask |= __GFP_RETRY_MAYFAIL;
1804 if (nid == NUMA_NO_NODE)
1805 nid = numa_mem_id();
1806 page = __alloc_pages(gfp_mask, order, nid, nmask);
1808 __count_vm_event(HTLB_BUDDY_PGALLOC);
1810 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1813 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1814 * indicates an overall state change. Clear bit so that we resume
1815 * normal 'try hard' allocations.
1817 if (node_alloc_noretry && page && !alloc_try_hard)
1818 node_clear(nid, *node_alloc_noretry);
1821 * If we tried hard to get a page but failed, set bit so that
1822 * subsequent attempts will not try as hard until there is an
1823 * overall state change.
1825 if (node_alloc_noretry && !page && alloc_try_hard)
1826 node_set(nid, *node_alloc_noretry);
1832 * Common helper to allocate a fresh hugetlb page. All specific allocators
1833 * should use this function to get new hugetlb pages
1835 static struct page *alloc_fresh_huge_page(struct hstate *h,
1836 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1837 nodemask_t *node_alloc_noretry)
1843 if (hstate_is_gigantic(h))
1844 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1846 page = alloc_buddy_huge_page(h, gfp_mask,
1847 nid, nmask, node_alloc_noretry);
1851 if (hstate_is_gigantic(h)) {
1852 if (!prep_compound_gigantic_page(page, huge_page_order(h))) {
1854 * Rare failure to convert pages to compound page.
1855 * Free pages and try again - ONCE!
1857 free_gigantic_page(page, huge_page_order(h));
1865 prep_new_huge_page(h, page, page_to_nid(page));
1871 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1874 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1875 nodemask_t *node_alloc_noretry)
1879 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1881 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1882 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1883 node_alloc_noretry);
1891 put_page(page); /* free it into the hugepage allocator */
1897 * Remove huge page from pool from next node to free. Attempt to keep
1898 * persistent huge pages more or less balanced over allowed nodes.
1899 * This routine only 'removes' the hugetlb page. The caller must make
1900 * an additional call to free the page to low level allocators.
1901 * Called with hugetlb_lock locked.
1903 static struct page *remove_pool_huge_page(struct hstate *h,
1904 nodemask_t *nodes_allowed,
1908 struct page *page = NULL;
1910 lockdep_assert_held(&hugetlb_lock);
1911 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1913 * If we're returning unused surplus pages, only examine
1914 * nodes with surplus pages.
1916 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1917 !list_empty(&h->hugepage_freelists[node])) {
1918 page = list_entry(h->hugepage_freelists[node].next,
1920 remove_hugetlb_page(h, page, acct_surplus);
1929 * Dissolve a given free hugepage into free buddy pages. This function does
1930 * nothing for in-use hugepages and non-hugepages.
1931 * This function returns values like below:
1933 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
1934 * when the system is under memory pressure and the feature of
1935 * freeing unused vmemmap pages associated with each hugetlb page
1937 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1938 * (allocated or reserved.)
1939 * 0: successfully dissolved free hugepages or the page is not a
1940 * hugepage (considered as already dissolved)
1942 int dissolve_free_huge_page(struct page *page)
1947 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1948 if (!PageHuge(page))
1951 spin_lock_irq(&hugetlb_lock);
1952 if (!PageHuge(page)) {
1957 if (!page_count(page)) {
1958 struct page *head = compound_head(page);
1959 struct hstate *h = page_hstate(head);
1960 if (h->free_huge_pages - h->resv_huge_pages == 0)
1964 * We should make sure that the page is already on the free list
1965 * when it is dissolved.
1967 if (unlikely(!HPageFreed(head))) {
1968 spin_unlock_irq(&hugetlb_lock);
1972 * Theoretically, we should return -EBUSY when we
1973 * encounter this race. In fact, we have a chance
1974 * to successfully dissolve the page if we do a
1975 * retry. Because the race window is quite small.
1976 * If we seize this opportunity, it is an optimization
1977 * for increasing the success rate of dissolving page.
1982 remove_hugetlb_page(h, head, false);
1983 h->max_huge_pages--;
1984 spin_unlock_irq(&hugetlb_lock);
1987 * Normally update_and_free_page will allocate required vmemmmap
1988 * before freeing the page. update_and_free_page will fail to
1989 * free the page if it can not allocate required vmemmap. We
1990 * need to adjust max_huge_pages if the page is not freed.
1991 * Attempt to allocate vmemmmap here so that we can take
1992 * appropriate action on failure.
1994 rc = alloc_huge_page_vmemmap(h, head);
1997 * Move PageHWPoison flag from head page to the raw
1998 * error page, which makes any subpages rather than
1999 * the error page reusable.
2001 if (PageHWPoison(head) && page != head) {
2002 SetPageHWPoison(page);
2003 ClearPageHWPoison(head);
2005 update_and_free_page(h, head, false);
2007 spin_lock_irq(&hugetlb_lock);
2008 add_hugetlb_page(h, head, false);
2009 h->max_huge_pages++;
2010 spin_unlock_irq(&hugetlb_lock);
2016 spin_unlock_irq(&hugetlb_lock);
2021 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2022 * make specified memory blocks removable from the system.
2023 * Note that this will dissolve a free gigantic hugepage completely, if any
2024 * part of it lies within the given range.
2025 * Also note that if dissolve_free_huge_page() returns with an error, all
2026 * free hugepages that were dissolved before that error are lost.
2028 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2034 if (!hugepages_supported())
2037 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
2038 page = pfn_to_page(pfn);
2039 rc = dissolve_free_huge_page(page);
2048 * Allocates a fresh surplus page from the page allocator.
2050 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
2051 int nid, nodemask_t *nmask, bool zero_ref)
2053 struct page *page = NULL;
2056 if (hstate_is_gigantic(h))
2059 spin_lock_irq(&hugetlb_lock);
2060 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2062 spin_unlock_irq(&hugetlb_lock);
2065 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2069 spin_lock_irq(&hugetlb_lock);
2071 * We could have raced with the pool size change.
2072 * Double check that and simply deallocate the new page
2073 * if we would end up overcommiting the surpluses. Abuse
2074 * temporary page to workaround the nasty free_huge_page
2077 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2078 SetHPageTemporary(page);
2079 spin_unlock_irq(&hugetlb_lock);
2086 * Caller requires a page with zero ref count.
2087 * We will drop ref count here. If someone else is holding
2088 * a ref, the page will be freed when they drop it. Abuse
2089 * temporary page flag to accomplish this.
2091 SetHPageTemporary(page);
2092 if (!put_page_testzero(page)) {
2094 * Unexpected inflated ref count on freshly allocated
2097 pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n");
2098 spin_unlock_irq(&hugetlb_lock);
2105 ClearHPageTemporary(page);
2108 h->surplus_huge_pages++;
2109 h->surplus_huge_pages_node[page_to_nid(page)]++;
2112 spin_unlock_irq(&hugetlb_lock);
2117 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2118 int nid, nodemask_t *nmask)
2122 if (hstate_is_gigantic(h))
2125 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2130 * We do not account these pages as surplus because they are only
2131 * temporary and will be released properly on the last reference
2133 SetHPageTemporary(page);
2139 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2142 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2143 struct vm_area_struct *vma, unsigned long addr)
2146 struct mempolicy *mpol;
2147 gfp_t gfp_mask = htlb_alloc_mask(h);
2149 nodemask_t *nodemask;
2151 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2152 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false);
2153 mpol_cond_put(mpol);
2158 /* page migration callback function */
2159 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2160 nodemask_t *nmask, gfp_t gfp_mask)
2162 spin_lock_irq(&hugetlb_lock);
2163 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2166 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2168 spin_unlock_irq(&hugetlb_lock);
2172 spin_unlock_irq(&hugetlb_lock);
2174 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2177 /* mempolicy aware migration callback */
2178 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2179 unsigned long address)
2181 struct mempolicy *mpol;
2182 nodemask_t *nodemask;
2187 gfp_mask = htlb_alloc_mask(h);
2188 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2189 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2190 mpol_cond_put(mpol);
2196 * Increase the hugetlb pool such that it can accommodate a reservation
2199 static int gather_surplus_pages(struct hstate *h, long delta)
2200 __must_hold(&hugetlb_lock)
2202 struct list_head surplus_list;
2203 struct page *page, *tmp;
2206 long needed, allocated;
2207 bool alloc_ok = true;
2209 lockdep_assert_held(&hugetlb_lock);
2210 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2212 h->resv_huge_pages += delta;
2217 INIT_LIST_HEAD(&surplus_list);
2221 spin_unlock_irq(&hugetlb_lock);
2222 for (i = 0; i < needed; i++) {
2223 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2224 NUMA_NO_NODE, NULL, true);
2229 list_add(&page->lru, &surplus_list);
2235 * After retaking hugetlb_lock, we need to recalculate 'needed'
2236 * because either resv_huge_pages or free_huge_pages may have changed.
2238 spin_lock_irq(&hugetlb_lock);
2239 needed = (h->resv_huge_pages + delta) -
2240 (h->free_huge_pages + allocated);
2245 * We were not able to allocate enough pages to
2246 * satisfy the entire reservation so we free what
2247 * we've allocated so far.
2252 * The surplus_list now contains _at_least_ the number of extra pages
2253 * needed to accommodate the reservation. Add the appropriate number
2254 * of pages to the hugetlb pool and free the extras back to the buddy
2255 * allocator. Commit the entire reservation here to prevent another
2256 * process from stealing the pages as they are added to the pool but
2257 * before they are reserved.
2259 needed += allocated;
2260 h->resv_huge_pages += delta;
2263 /* Free the needed pages to the hugetlb pool */
2264 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2267 /* Add the page to the hugetlb allocator */
2268 enqueue_huge_page(h, page);
2271 spin_unlock_irq(&hugetlb_lock);
2274 * Free unnecessary surplus pages to the buddy allocator.
2275 * Pages have no ref count, call free_huge_page directly.
2277 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2278 free_huge_page(page);
2279 spin_lock_irq(&hugetlb_lock);
2285 * This routine has two main purposes:
2286 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2287 * in unused_resv_pages. This corresponds to the prior adjustments made
2288 * to the associated reservation map.
2289 * 2) Free any unused surplus pages that may have been allocated to satisfy
2290 * the reservation. As many as unused_resv_pages may be freed.
2292 static void return_unused_surplus_pages(struct hstate *h,
2293 unsigned long unused_resv_pages)
2295 unsigned long nr_pages;
2297 LIST_HEAD(page_list);
2299 lockdep_assert_held(&hugetlb_lock);
2300 /* Uncommit the reservation */
2301 h->resv_huge_pages -= unused_resv_pages;
2303 /* Cannot return gigantic pages currently */
2304 if (hstate_is_gigantic(h))
2308 * Part (or even all) of the reservation could have been backed
2309 * by pre-allocated pages. Only free surplus pages.
2311 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2314 * We want to release as many surplus pages as possible, spread
2315 * evenly across all nodes with memory. Iterate across these nodes
2316 * until we can no longer free unreserved surplus pages. This occurs
2317 * when the nodes with surplus pages have no free pages.
2318 * remove_pool_huge_page() will balance the freed pages across the
2319 * on-line nodes with memory and will handle the hstate accounting.
2321 while (nr_pages--) {
2322 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2326 list_add(&page->lru, &page_list);
2330 spin_unlock_irq(&hugetlb_lock);
2331 update_and_free_pages_bulk(h, &page_list);
2332 spin_lock_irq(&hugetlb_lock);
2337 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2338 * are used by the huge page allocation routines to manage reservations.
2340 * vma_needs_reservation is called to determine if the huge page at addr
2341 * within the vma has an associated reservation. If a reservation is
2342 * needed, the value 1 is returned. The caller is then responsible for
2343 * managing the global reservation and subpool usage counts. After
2344 * the huge page has been allocated, vma_commit_reservation is called
2345 * to add the page to the reservation map. If the page allocation fails,
2346 * the reservation must be ended instead of committed. vma_end_reservation
2347 * is called in such cases.
2349 * In the normal case, vma_commit_reservation returns the same value
2350 * as the preceding vma_needs_reservation call. The only time this
2351 * is not the case is if a reserve map was changed between calls. It
2352 * is the responsibility of the caller to notice the difference and
2353 * take appropriate action.
2355 * vma_add_reservation is used in error paths where a reservation must
2356 * be restored when a newly allocated huge page must be freed. It is
2357 * to be called after calling vma_needs_reservation to determine if a
2358 * reservation exists.
2360 * vma_del_reservation is used in error paths where an entry in the reserve
2361 * map was created during huge page allocation and must be removed. It is to
2362 * be called after calling vma_needs_reservation to determine if a reservation
2365 enum vma_resv_mode {
2372 static long __vma_reservation_common(struct hstate *h,
2373 struct vm_area_struct *vma, unsigned long addr,
2374 enum vma_resv_mode mode)
2376 struct resv_map *resv;
2379 long dummy_out_regions_needed;
2381 resv = vma_resv_map(vma);
2385 idx = vma_hugecache_offset(h, vma, addr);
2387 case VMA_NEEDS_RESV:
2388 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2389 /* We assume that vma_reservation_* routines always operate on
2390 * 1 page, and that adding to resv map a 1 page entry can only
2391 * ever require 1 region.
2393 VM_BUG_ON(dummy_out_regions_needed != 1);
2395 case VMA_COMMIT_RESV:
2396 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2397 /* region_add calls of range 1 should never fail. */
2401 region_abort(resv, idx, idx + 1, 1);
2405 if (vma->vm_flags & VM_MAYSHARE) {
2406 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2407 /* region_add calls of range 1 should never fail. */
2410 region_abort(resv, idx, idx + 1, 1);
2411 ret = region_del(resv, idx, idx + 1);
2415 if (vma->vm_flags & VM_MAYSHARE) {
2416 region_abort(resv, idx, idx + 1, 1);
2417 ret = region_del(resv, idx, idx + 1);
2419 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2420 /* region_add calls of range 1 should never fail. */
2428 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2431 * We know private mapping must have HPAGE_RESV_OWNER set.
2433 * In most cases, reserves always exist for private mappings.
2434 * However, a file associated with mapping could have been
2435 * hole punched or truncated after reserves were consumed.
2436 * As subsequent fault on such a range will not use reserves.
2437 * Subtle - The reserve map for private mappings has the
2438 * opposite meaning than that of shared mappings. If NO
2439 * entry is in the reserve map, it means a reservation exists.
2440 * If an entry exists in the reserve map, it means the
2441 * reservation has already been consumed. As a result, the
2442 * return value of this routine is the opposite of the
2443 * value returned from reserve map manipulation routines above.
2452 static long vma_needs_reservation(struct hstate *h,
2453 struct vm_area_struct *vma, unsigned long addr)
2455 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2458 static long vma_commit_reservation(struct hstate *h,
2459 struct vm_area_struct *vma, unsigned long addr)
2461 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2464 static void vma_end_reservation(struct hstate *h,
2465 struct vm_area_struct *vma, unsigned long addr)
2467 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2470 static long vma_add_reservation(struct hstate *h,
2471 struct vm_area_struct *vma, unsigned long addr)
2473 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2476 static long vma_del_reservation(struct hstate *h,
2477 struct vm_area_struct *vma, unsigned long addr)
2479 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2483 * This routine is called to restore reservation information on error paths.
2484 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2485 * the hugetlb mutex should remain held when calling this routine.
2487 * It handles two specific cases:
2488 * 1) A reservation was in place and the page consumed the reservation.
2489 * HPageRestoreReserve is set in the page.
2490 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2491 * not set. However, alloc_huge_page always updates the reserve map.
2493 * In case 1, free_huge_page later in the error path will increment the
2494 * global reserve count. But, free_huge_page does not have enough context
2495 * to adjust the reservation map. This case deals primarily with private
2496 * mappings. Adjust the reserve map here to be consistent with global
2497 * reserve count adjustments to be made by free_huge_page. Make sure the
2498 * reserve map indicates there is a reservation present.
2500 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2502 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2503 unsigned long address, struct page *page)
2505 long rc = vma_needs_reservation(h, vma, address);
2507 if (HPageRestoreReserve(page)) {
2508 if (unlikely(rc < 0))
2510 * Rare out of memory condition in reserve map
2511 * manipulation. Clear HPageRestoreReserve so that
2512 * global reserve count will not be incremented
2513 * by free_huge_page. This will make it appear
2514 * as though the reservation for this page was
2515 * consumed. This may prevent the task from
2516 * faulting in the page at a later time. This
2517 * is better than inconsistent global huge page
2518 * accounting of reserve counts.
2520 ClearHPageRestoreReserve(page);
2522 (void)vma_add_reservation(h, vma, address);
2524 vma_end_reservation(h, vma, address);
2528 * This indicates there is an entry in the reserve map
2529 * not added by alloc_huge_page. We know it was added
2530 * before the alloc_huge_page call, otherwise
2531 * HPageRestoreReserve would be set on the page.
2532 * Remove the entry so that a subsequent allocation
2533 * does not consume a reservation.
2535 rc = vma_del_reservation(h, vma, address);
2538 * VERY rare out of memory condition. Since
2539 * we can not delete the entry, set
2540 * HPageRestoreReserve so that the reserve
2541 * count will be incremented when the page
2542 * is freed. This reserve will be consumed
2543 * on a subsequent allocation.
2545 SetHPageRestoreReserve(page);
2546 } else if (rc < 0) {
2548 * Rare out of memory condition from
2549 * vma_needs_reservation call. Memory allocation is
2550 * only attempted if a new entry is needed. Therefore,
2551 * this implies there is not an entry in the
2554 * For shared mappings, no entry in the map indicates
2555 * no reservation. We are done.
2557 if (!(vma->vm_flags & VM_MAYSHARE))
2559 * For private mappings, no entry indicates
2560 * a reservation is present. Since we can
2561 * not add an entry, set SetHPageRestoreReserve
2562 * on the page so reserve count will be
2563 * incremented when freed. This reserve will
2564 * be consumed on a subsequent allocation.
2566 SetHPageRestoreReserve(page);
2569 * No reservation present, do nothing
2571 vma_end_reservation(h, vma, address);
2576 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2577 * @h: struct hstate old page belongs to
2578 * @old_page: Old page to dissolve
2579 * @list: List to isolate the page in case we need to
2580 * Returns 0 on success, otherwise negated error.
2582 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2583 struct list_head *list)
2585 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2586 int nid = page_to_nid(old_page);
2587 bool alloc_retry = false;
2588 struct page *new_page;
2592 * Before dissolving the page, we need to allocate a new one for the
2593 * pool to remain stable. Here, we allocate the page and 'prep' it
2594 * by doing everything but actually updating counters and adding to
2595 * the pool. This simplifies and let us do most of the processing
2599 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2603 * If all goes well, this page will be directly added to the free
2604 * list in the pool. For this the ref count needs to be zero.
2605 * Attempt to drop now, and retry once if needed. It is VERY
2606 * unlikely there is another ref on the page.
2608 * If someone else has a reference to the page, it will be freed
2609 * when they drop their ref. Abuse temporary page flag to accomplish
2610 * this. Retry once if there is an inflated ref count.
2612 SetHPageTemporary(new_page);
2613 if (!put_page_testzero(new_page)) {
2620 ClearHPageTemporary(new_page);
2622 __prep_new_huge_page(h, new_page);
2625 spin_lock_irq(&hugetlb_lock);
2626 if (!PageHuge(old_page)) {
2628 * Freed from under us. Drop new_page too.
2631 } else if (page_count(old_page)) {
2633 * Someone has grabbed the page, try to isolate it here.
2634 * Fail with -EBUSY if not possible.
2636 spin_unlock_irq(&hugetlb_lock);
2637 if (!isolate_huge_page(old_page, list))
2639 spin_lock_irq(&hugetlb_lock);
2641 } else if (!HPageFreed(old_page)) {
2643 * Page's refcount is 0 but it has not been enqueued in the
2644 * freelist yet. Race window is small, so we can succeed here if
2647 spin_unlock_irq(&hugetlb_lock);
2652 * Ok, old_page is still a genuine free hugepage. Remove it from
2653 * the freelist and decrease the counters. These will be
2654 * incremented again when calling __prep_account_new_huge_page()
2655 * and enqueue_huge_page() for new_page. The counters will remain
2656 * stable since this happens under the lock.
2658 remove_hugetlb_page(h, old_page, false);
2661 * Ref count on new page is already zero as it was dropped
2662 * earlier. It can be directly added to the pool free list.
2664 __prep_account_new_huge_page(h, nid);
2665 enqueue_huge_page(h, new_page);
2668 * Pages have been replaced, we can safely free the old one.
2670 spin_unlock_irq(&hugetlb_lock);
2671 update_and_free_page(h, old_page, false);
2677 spin_unlock_irq(&hugetlb_lock);
2678 /* Page has a zero ref count, but needs a ref to be freed */
2679 set_page_refcounted(new_page);
2680 update_and_free_page(h, new_page, false);
2685 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2692 * The page might have been dissolved from under our feet, so make sure
2693 * to carefully check the state under the lock.
2694 * Return success when racing as if we dissolved the page ourselves.
2696 spin_lock_irq(&hugetlb_lock);
2697 if (PageHuge(page)) {
2698 head = compound_head(page);
2699 h = page_hstate(head);
2701 spin_unlock_irq(&hugetlb_lock);
2704 spin_unlock_irq(&hugetlb_lock);
2707 * Fence off gigantic pages as there is a cyclic dependency between
2708 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2709 * of bailing out right away without further retrying.
2711 if (hstate_is_gigantic(h))
2714 if (page_count(head) && isolate_huge_page(head, list))
2716 else if (!page_count(head))
2717 ret = alloc_and_dissolve_huge_page(h, head, list);
2722 struct page *alloc_huge_page(struct vm_area_struct *vma,
2723 unsigned long addr, int avoid_reserve)
2725 struct hugepage_subpool *spool = subpool_vma(vma);
2726 struct hstate *h = hstate_vma(vma);
2728 long map_chg, map_commit;
2731 struct hugetlb_cgroup *h_cg;
2732 bool deferred_reserve;
2734 idx = hstate_index(h);
2736 * Examine the region/reserve map to determine if the process
2737 * has a reservation for the page to be allocated. A return
2738 * code of zero indicates a reservation exists (no change).
2740 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2742 return ERR_PTR(-ENOMEM);
2745 * Processes that did not create the mapping will have no
2746 * reserves as indicated by the region/reserve map. Check
2747 * that the allocation will not exceed the subpool limit.
2748 * Allocations for MAP_NORESERVE mappings also need to be
2749 * checked against any subpool limit.
2751 if (map_chg || avoid_reserve) {
2752 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2754 vma_end_reservation(h, vma, addr);
2755 return ERR_PTR(-ENOSPC);
2759 * Even though there was no reservation in the region/reserve
2760 * map, there could be reservations associated with the
2761 * subpool that can be used. This would be indicated if the
2762 * return value of hugepage_subpool_get_pages() is zero.
2763 * However, if avoid_reserve is specified we still avoid even
2764 * the subpool reservations.
2770 /* If this allocation is not consuming a reservation, charge it now.
2772 deferred_reserve = map_chg || avoid_reserve;
2773 if (deferred_reserve) {
2774 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2775 idx, pages_per_huge_page(h), &h_cg);
2777 goto out_subpool_put;
2780 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2782 goto out_uncharge_cgroup_reservation;
2784 spin_lock_irq(&hugetlb_lock);
2786 * glb_chg is passed to indicate whether or not a page must be taken
2787 * from the global free pool (global change). gbl_chg == 0 indicates
2788 * a reservation exists for the allocation.
2790 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2792 spin_unlock_irq(&hugetlb_lock);
2793 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2795 goto out_uncharge_cgroup;
2796 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2797 SetHPageRestoreReserve(page);
2798 h->resv_huge_pages--;
2800 spin_lock_irq(&hugetlb_lock);
2801 list_add(&page->lru, &h->hugepage_activelist);
2804 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2805 /* If allocation is not consuming a reservation, also store the
2806 * hugetlb_cgroup pointer on the page.
2808 if (deferred_reserve) {
2809 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2813 spin_unlock_irq(&hugetlb_lock);
2815 hugetlb_set_page_subpool(page, spool);
2817 map_commit = vma_commit_reservation(h, vma, addr);
2818 if (unlikely(map_chg > map_commit)) {
2820 * The page was added to the reservation map between
2821 * vma_needs_reservation and vma_commit_reservation.
2822 * This indicates a race with hugetlb_reserve_pages.
2823 * Adjust for the subpool count incremented above AND
2824 * in hugetlb_reserve_pages for the same page. Also,
2825 * the reservation count added in hugetlb_reserve_pages
2826 * no longer applies.
2830 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2831 hugetlb_acct_memory(h, -rsv_adjust);
2832 if (deferred_reserve)
2833 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2834 pages_per_huge_page(h), page);
2838 out_uncharge_cgroup:
2839 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2840 out_uncharge_cgroup_reservation:
2841 if (deferred_reserve)
2842 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2845 if (map_chg || avoid_reserve)
2846 hugepage_subpool_put_pages(spool, 1);
2847 vma_end_reservation(h, vma, addr);
2848 return ERR_PTR(-ENOSPC);
2851 int alloc_bootmem_huge_page(struct hstate *h)
2852 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2853 int __alloc_bootmem_huge_page(struct hstate *h)
2855 struct huge_bootmem_page *m;
2858 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2861 addr = memblock_alloc_try_nid_raw(
2862 huge_page_size(h), huge_page_size(h),
2863 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2866 * Use the beginning of the huge page to store the
2867 * huge_bootmem_page struct (until gather_bootmem
2868 * puts them into the mem_map).
2877 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2878 /* Put them into a private list first because mem_map is not up yet */
2879 INIT_LIST_HEAD(&m->list);
2880 list_add(&m->list, &huge_boot_pages);
2886 * Put bootmem huge pages into the standard lists after mem_map is up.
2887 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
2889 static void __init gather_bootmem_prealloc(void)
2891 struct huge_bootmem_page *m;
2893 list_for_each_entry(m, &huge_boot_pages, list) {
2894 struct page *page = virt_to_page(m);
2895 struct hstate *h = m->hstate;
2897 VM_BUG_ON(!hstate_is_gigantic(h));
2898 WARN_ON(page_count(page) != 1);
2899 if (prep_compound_gigantic_page(page, huge_page_order(h))) {
2900 WARN_ON(PageReserved(page));
2901 prep_new_huge_page(h, page, page_to_nid(page));
2902 put_page(page); /* add to the hugepage allocator */
2904 /* VERY unlikely inflated ref count on a tail page */
2905 free_gigantic_page(page, huge_page_order(h));
2909 * We need to restore the 'stolen' pages to totalram_pages
2910 * in order to fix confusing memory reports from free(1) and
2911 * other side-effects, like CommitLimit going negative.
2913 adjust_managed_page_count(page, pages_per_huge_page(h));
2918 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2921 nodemask_t *node_alloc_noretry;
2923 if (!hstate_is_gigantic(h)) {
2925 * Bit mask controlling how hard we retry per-node allocations.
2926 * Ignore errors as lower level routines can deal with
2927 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2928 * time, we are likely in bigger trouble.
2930 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2933 /* allocations done at boot time */
2934 node_alloc_noretry = NULL;
2937 /* bit mask controlling how hard we retry per-node allocations */
2938 if (node_alloc_noretry)
2939 nodes_clear(*node_alloc_noretry);
2941 for (i = 0; i < h->max_huge_pages; ++i) {
2942 if (hstate_is_gigantic(h)) {
2943 if (hugetlb_cma_size) {
2944 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2947 if (!alloc_bootmem_huge_page(h))
2949 } else if (!alloc_pool_huge_page(h,
2950 &node_states[N_MEMORY],
2951 node_alloc_noretry))
2955 if (i < h->max_huge_pages) {
2958 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2959 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2960 h->max_huge_pages, buf, i);
2961 h->max_huge_pages = i;
2964 kfree(node_alloc_noretry);
2967 static void __init hugetlb_init_hstates(void)
2971 for_each_hstate(h) {
2972 if (minimum_order > huge_page_order(h))
2973 minimum_order = huge_page_order(h);
2975 /* oversize hugepages were init'ed in early boot */
2976 if (!hstate_is_gigantic(h))
2977 hugetlb_hstate_alloc_pages(h);
2979 VM_BUG_ON(minimum_order == UINT_MAX);
2982 static void __init report_hugepages(void)
2986 for_each_hstate(h) {
2989 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2990 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2991 buf, h->free_huge_pages);
2995 #ifdef CONFIG_HIGHMEM
2996 static void try_to_free_low(struct hstate *h, unsigned long count,
2997 nodemask_t *nodes_allowed)
3000 LIST_HEAD(page_list);
3002 lockdep_assert_held(&hugetlb_lock);
3003 if (hstate_is_gigantic(h))
3007 * Collect pages to be freed on a list, and free after dropping lock
3009 for_each_node_mask(i, *nodes_allowed) {
3010 struct page *page, *next;
3011 struct list_head *freel = &h->hugepage_freelists[i];
3012 list_for_each_entry_safe(page, next, freel, lru) {
3013 if (count >= h->nr_huge_pages)
3015 if (PageHighMem(page))
3017 remove_hugetlb_page(h, page, false);
3018 list_add(&page->lru, &page_list);
3023 spin_unlock_irq(&hugetlb_lock);
3024 update_and_free_pages_bulk(h, &page_list);
3025 spin_lock_irq(&hugetlb_lock);
3028 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3029 nodemask_t *nodes_allowed)
3035 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3036 * balanced by operating on them in a round-robin fashion.
3037 * Returns 1 if an adjustment was made.
3039 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3044 lockdep_assert_held(&hugetlb_lock);
3045 VM_BUG_ON(delta != -1 && delta != 1);
3048 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3049 if (h->surplus_huge_pages_node[node])
3053 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3054 if (h->surplus_huge_pages_node[node] <
3055 h->nr_huge_pages_node[node])
3062 h->surplus_huge_pages += delta;
3063 h->surplus_huge_pages_node[node] += delta;
3067 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
3068 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3069 nodemask_t *nodes_allowed)
3071 unsigned long min_count, ret;
3073 LIST_HEAD(page_list);
3074 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3077 * Bit mask controlling how hard we retry per-node allocations.
3078 * If we can not allocate the bit mask, do not attempt to allocate
3079 * the requested huge pages.
3081 if (node_alloc_noretry)
3082 nodes_clear(*node_alloc_noretry);
3087 * resize_lock mutex prevents concurrent adjustments to number of
3088 * pages in hstate via the proc/sysfs interfaces.
3090 mutex_lock(&h->resize_lock);
3091 flush_free_hpage_work(h);
3092 spin_lock_irq(&hugetlb_lock);
3095 * Check for a node specific request.
3096 * Changing node specific huge page count may require a corresponding
3097 * change to the global count. In any case, the passed node mask
3098 * (nodes_allowed) will restrict alloc/free to the specified node.
3100 if (nid != NUMA_NO_NODE) {
3101 unsigned long old_count = count;
3103 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3105 * User may have specified a large count value which caused the
3106 * above calculation to overflow. In this case, they wanted
3107 * to allocate as many huge pages as possible. Set count to
3108 * largest possible value to align with their intention.
3110 if (count < old_count)
3115 * Gigantic pages runtime allocation depend on the capability for large
3116 * page range allocation.
3117 * If the system does not provide this feature, return an error when
3118 * the user tries to allocate gigantic pages but let the user free the
3119 * boottime allocated gigantic pages.
3121 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3122 if (count > persistent_huge_pages(h)) {
3123 spin_unlock_irq(&hugetlb_lock);
3124 mutex_unlock(&h->resize_lock);
3125 NODEMASK_FREE(node_alloc_noretry);
3128 /* Fall through to decrease pool */
3132 * Increase the pool size
3133 * First take pages out of surplus state. Then make up the
3134 * remaining difference by allocating fresh huge pages.
3136 * We might race with alloc_surplus_huge_page() here and be unable
3137 * to convert a surplus huge page to a normal huge page. That is
3138 * not critical, though, it just means the overall size of the
3139 * pool might be one hugepage larger than it needs to be, but
3140 * within all the constraints specified by the sysctls.
3142 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3143 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3147 while (count > persistent_huge_pages(h)) {
3149 * If this allocation races such that we no longer need the
3150 * page, free_huge_page will handle it by freeing the page
3151 * and reducing the surplus.
3153 spin_unlock_irq(&hugetlb_lock);
3155 /* yield cpu to avoid soft lockup */
3158 ret = alloc_pool_huge_page(h, nodes_allowed,
3159 node_alloc_noretry);
3160 spin_lock_irq(&hugetlb_lock);
3164 /* Bail for signals. Probably ctrl-c from user */
3165 if (signal_pending(current))
3170 * Decrease the pool size
3171 * First return free pages to the buddy allocator (being careful
3172 * to keep enough around to satisfy reservations). Then place
3173 * pages into surplus state as needed so the pool will shrink
3174 * to the desired size as pages become free.
3176 * By placing pages into the surplus state independent of the
3177 * overcommit value, we are allowing the surplus pool size to
3178 * exceed overcommit. There are few sane options here. Since
3179 * alloc_surplus_huge_page() is checking the global counter,
3180 * though, we'll note that we're not allowed to exceed surplus
3181 * and won't grow the pool anywhere else. Not until one of the
3182 * sysctls are changed, or the surplus pages go out of use.
3184 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3185 min_count = max(count, min_count);
3186 try_to_free_low(h, min_count, nodes_allowed);
3189 * Collect pages to be removed on list without dropping lock
3191 while (min_count < persistent_huge_pages(h)) {
3192 page = remove_pool_huge_page(h, nodes_allowed, 0);
3196 list_add(&page->lru, &page_list);
3198 /* free the pages after dropping lock */
3199 spin_unlock_irq(&hugetlb_lock);
3200 update_and_free_pages_bulk(h, &page_list);
3201 flush_free_hpage_work(h);
3202 spin_lock_irq(&hugetlb_lock);
3204 while (count < persistent_huge_pages(h)) {
3205 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3209 h->max_huge_pages = persistent_huge_pages(h);
3210 spin_unlock_irq(&hugetlb_lock);
3211 mutex_unlock(&h->resize_lock);
3213 NODEMASK_FREE(node_alloc_noretry);
3218 #define HSTATE_ATTR_RO(_name) \
3219 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3221 #define HSTATE_ATTR(_name) \
3222 static struct kobj_attribute _name##_attr = \
3223 __ATTR(_name, 0644, _name##_show, _name##_store)
3225 static struct kobject *hugepages_kobj;
3226 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3228 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3230 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3234 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3235 if (hstate_kobjs[i] == kobj) {
3237 *nidp = NUMA_NO_NODE;
3241 return kobj_to_node_hstate(kobj, nidp);
3244 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3245 struct kobj_attribute *attr, char *buf)
3248 unsigned long nr_huge_pages;
3251 h = kobj_to_hstate(kobj, &nid);
3252 if (nid == NUMA_NO_NODE)
3253 nr_huge_pages = h->nr_huge_pages;
3255 nr_huge_pages = h->nr_huge_pages_node[nid];
3257 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3260 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3261 struct hstate *h, int nid,
3262 unsigned long count, size_t len)
3265 nodemask_t nodes_allowed, *n_mask;
3267 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3270 if (nid == NUMA_NO_NODE) {
3272 * global hstate attribute
3274 if (!(obey_mempolicy &&
3275 init_nodemask_of_mempolicy(&nodes_allowed)))
3276 n_mask = &node_states[N_MEMORY];
3278 n_mask = &nodes_allowed;
3281 * Node specific request. count adjustment happens in
3282 * set_max_huge_pages() after acquiring hugetlb_lock.
3284 init_nodemask_of_node(&nodes_allowed, nid);
3285 n_mask = &nodes_allowed;
3288 err = set_max_huge_pages(h, count, nid, n_mask);
3290 return err ? err : len;
3293 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3294 struct kobject *kobj, const char *buf,
3298 unsigned long count;
3302 err = kstrtoul(buf, 10, &count);
3306 h = kobj_to_hstate(kobj, &nid);
3307 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3310 static ssize_t nr_hugepages_show(struct kobject *kobj,
3311 struct kobj_attribute *attr, char *buf)
3313 return nr_hugepages_show_common(kobj, attr, buf);
3316 static ssize_t nr_hugepages_store(struct kobject *kobj,
3317 struct kobj_attribute *attr, const char *buf, size_t len)
3319 return nr_hugepages_store_common(false, kobj, buf, len);
3321 HSTATE_ATTR(nr_hugepages);
3326 * hstate attribute for optionally mempolicy-based constraint on persistent
3327 * huge page alloc/free.
3329 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3330 struct kobj_attribute *attr,
3333 return nr_hugepages_show_common(kobj, attr, buf);
3336 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3337 struct kobj_attribute *attr, const char *buf, size_t len)
3339 return nr_hugepages_store_common(true, kobj, buf, len);
3341 HSTATE_ATTR(nr_hugepages_mempolicy);
3345 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3346 struct kobj_attribute *attr, char *buf)
3348 struct hstate *h = kobj_to_hstate(kobj, NULL);
3349 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3352 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3353 struct kobj_attribute *attr, const char *buf, size_t count)
3356 unsigned long input;
3357 struct hstate *h = kobj_to_hstate(kobj, NULL);
3359 if (hstate_is_gigantic(h))
3362 err = kstrtoul(buf, 10, &input);
3366 spin_lock_irq(&hugetlb_lock);
3367 h->nr_overcommit_huge_pages = input;
3368 spin_unlock_irq(&hugetlb_lock);
3372 HSTATE_ATTR(nr_overcommit_hugepages);
3374 static ssize_t free_hugepages_show(struct kobject *kobj,
3375 struct kobj_attribute *attr, char *buf)
3378 unsigned long free_huge_pages;
3381 h = kobj_to_hstate(kobj, &nid);
3382 if (nid == NUMA_NO_NODE)
3383 free_huge_pages = h->free_huge_pages;
3385 free_huge_pages = h->free_huge_pages_node[nid];
3387 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3389 HSTATE_ATTR_RO(free_hugepages);
3391 static ssize_t resv_hugepages_show(struct kobject *kobj,
3392 struct kobj_attribute *attr, char *buf)
3394 struct hstate *h = kobj_to_hstate(kobj, NULL);
3395 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3397 HSTATE_ATTR_RO(resv_hugepages);
3399 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3400 struct kobj_attribute *attr, char *buf)
3403 unsigned long surplus_huge_pages;
3406 h = kobj_to_hstate(kobj, &nid);
3407 if (nid == NUMA_NO_NODE)
3408 surplus_huge_pages = h->surplus_huge_pages;
3410 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3412 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3414 HSTATE_ATTR_RO(surplus_hugepages);
3416 static struct attribute *hstate_attrs[] = {
3417 &nr_hugepages_attr.attr,
3418 &nr_overcommit_hugepages_attr.attr,
3419 &free_hugepages_attr.attr,
3420 &resv_hugepages_attr.attr,
3421 &surplus_hugepages_attr.attr,
3423 &nr_hugepages_mempolicy_attr.attr,
3428 static const struct attribute_group hstate_attr_group = {
3429 .attrs = hstate_attrs,
3432 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3433 struct kobject **hstate_kobjs,
3434 const struct attribute_group *hstate_attr_group)
3437 int hi = hstate_index(h);
3439 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3440 if (!hstate_kobjs[hi])
3443 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3445 kobject_put(hstate_kobjs[hi]);
3446 hstate_kobjs[hi] = NULL;
3452 static void __init hugetlb_sysfs_init(void)
3457 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3458 if (!hugepages_kobj)
3461 for_each_hstate(h) {
3462 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3463 hstate_kobjs, &hstate_attr_group);
3465 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3472 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3473 * with node devices in node_devices[] using a parallel array. The array
3474 * index of a node device or _hstate == node id.
3475 * This is here to avoid any static dependency of the node device driver, in
3476 * the base kernel, on the hugetlb module.
3478 struct node_hstate {
3479 struct kobject *hugepages_kobj;
3480 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3482 static struct node_hstate node_hstates[MAX_NUMNODES];
3485 * A subset of global hstate attributes for node devices
3487 static struct attribute *per_node_hstate_attrs[] = {
3488 &nr_hugepages_attr.attr,
3489 &free_hugepages_attr.attr,
3490 &surplus_hugepages_attr.attr,
3494 static const struct attribute_group per_node_hstate_attr_group = {
3495 .attrs = per_node_hstate_attrs,
3499 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3500 * Returns node id via non-NULL nidp.
3502 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3506 for (nid = 0; nid < nr_node_ids; nid++) {
3507 struct node_hstate *nhs = &node_hstates[nid];
3509 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3510 if (nhs->hstate_kobjs[i] == kobj) {
3522 * Unregister hstate attributes from a single node device.
3523 * No-op if no hstate attributes attached.
3525 static void hugetlb_unregister_node(struct node *node)
3528 struct node_hstate *nhs = &node_hstates[node->dev.id];
3530 if (!nhs->hugepages_kobj)
3531 return; /* no hstate attributes */
3533 for_each_hstate(h) {
3534 int idx = hstate_index(h);
3535 if (nhs->hstate_kobjs[idx]) {
3536 kobject_put(nhs->hstate_kobjs[idx]);
3537 nhs->hstate_kobjs[idx] = NULL;
3541 kobject_put(nhs->hugepages_kobj);
3542 nhs->hugepages_kobj = NULL;
3547 * Register hstate attributes for a single node device.
3548 * No-op if attributes already registered.
3550 static void hugetlb_register_node(struct node *node)
3553 struct node_hstate *nhs = &node_hstates[node->dev.id];
3556 if (nhs->hugepages_kobj)
3557 return; /* already allocated */
3559 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3561 if (!nhs->hugepages_kobj)
3564 for_each_hstate(h) {
3565 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3567 &per_node_hstate_attr_group);
3569 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3570 h->name, node->dev.id);
3571 hugetlb_unregister_node(node);
3578 * hugetlb init time: register hstate attributes for all registered node
3579 * devices of nodes that have memory. All on-line nodes should have
3580 * registered their associated device by this time.
3582 static void __init hugetlb_register_all_nodes(void)
3586 for_each_node_state(nid, N_MEMORY) {
3587 struct node *node = node_devices[nid];
3588 if (node->dev.id == nid)
3589 hugetlb_register_node(node);
3593 * Let the node device driver know we're here so it can
3594 * [un]register hstate attributes on node hotplug.
3596 register_hugetlbfs_with_node(hugetlb_register_node,
3597 hugetlb_unregister_node);
3599 #else /* !CONFIG_NUMA */
3601 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3609 static void hugetlb_register_all_nodes(void) { }
3613 static int __init hugetlb_init(void)
3617 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3620 if (!hugepages_supported()) {
3621 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3622 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3627 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3628 * architectures depend on setup being done here.
3630 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3631 if (!parsed_default_hugepagesz) {
3633 * If we did not parse a default huge page size, set
3634 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3635 * number of huge pages for this default size was implicitly
3636 * specified, set that here as well.
3637 * Note that the implicit setting will overwrite an explicit
3638 * setting. A warning will be printed in this case.
3640 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3641 if (default_hstate_max_huge_pages) {
3642 if (default_hstate.max_huge_pages) {
3645 string_get_size(huge_page_size(&default_hstate),
3646 1, STRING_UNITS_2, buf, 32);
3647 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3648 default_hstate.max_huge_pages, buf);
3649 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3650 default_hstate_max_huge_pages);
3652 default_hstate.max_huge_pages =
3653 default_hstate_max_huge_pages;
3657 hugetlb_cma_check();
3658 hugetlb_init_hstates();
3659 gather_bootmem_prealloc();
3662 hugetlb_sysfs_init();
3663 hugetlb_register_all_nodes();
3664 hugetlb_cgroup_file_init();
3667 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3669 num_fault_mutexes = 1;
3671 hugetlb_fault_mutex_table =
3672 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3674 BUG_ON(!hugetlb_fault_mutex_table);
3676 for (i = 0; i < num_fault_mutexes; i++)
3677 mutex_init(&hugetlb_fault_mutex_table[i]);
3680 subsys_initcall(hugetlb_init);
3682 /* Overwritten by architectures with more huge page sizes */
3683 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3685 return size == HPAGE_SIZE;
3688 void __init hugetlb_add_hstate(unsigned int order)
3693 if (size_to_hstate(PAGE_SIZE << order)) {
3696 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3698 h = &hstates[hugetlb_max_hstate++];
3699 mutex_init(&h->resize_lock);
3701 h->mask = ~(huge_page_size(h) - 1);
3702 for (i = 0; i < MAX_NUMNODES; ++i)
3703 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3704 INIT_LIST_HEAD(&h->hugepage_activelist);
3705 h->next_nid_to_alloc = first_memory_node;
3706 h->next_nid_to_free = first_memory_node;
3707 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3708 huge_page_size(h)/1024);
3709 hugetlb_vmemmap_init(h);
3715 * hugepages command line processing
3716 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3717 * specification. If not, ignore the hugepages value. hugepages can also
3718 * be the first huge page command line option in which case it implicitly
3719 * specifies the number of huge pages for the default size.
3721 static int __init hugepages_setup(char *s)
3724 static unsigned long *last_mhp;
3726 if (!parsed_valid_hugepagesz) {
3727 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3728 parsed_valid_hugepagesz = true;
3733 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3734 * yet, so this hugepages= parameter goes to the "default hstate".
3735 * Otherwise, it goes with the previously parsed hugepagesz or
3736 * default_hugepagesz.
3738 else if (!hugetlb_max_hstate)
3739 mhp = &default_hstate_max_huge_pages;
3741 mhp = &parsed_hstate->max_huge_pages;
3743 if (mhp == last_mhp) {
3744 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3748 if (sscanf(s, "%lu", mhp) <= 0)
3752 * Global state is always initialized later in hugetlb_init.
3753 * But we need to allocate gigantic hstates here early to still
3754 * use the bootmem allocator.
3756 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3757 hugetlb_hstate_alloc_pages(parsed_hstate);
3763 __setup("hugepages=", hugepages_setup);
3766 * hugepagesz command line processing
3767 * A specific huge page size can only be specified once with hugepagesz.
3768 * hugepagesz is followed by hugepages on the command line. The global
3769 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3770 * hugepagesz argument was valid.
3772 static int __init hugepagesz_setup(char *s)
3777 parsed_valid_hugepagesz = false;
3778 size = (unsigned long)memparse(s, NULL);
3780 if (!arch_hugetlb_valid_size(size)) {
3781 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3785 h = size_to_hstate(size);
3788 * hstate for this size already exists. This is normally
3789 * an error, but is allowed if the existing hstate is the
3790 * default hstate. More specifically, it is only allowed if
3791 * the number of huge pages for the default hstate was not
3792 * previously specified.
3794 if (!parsed_default_hugepagesz || h != &default_hstate ||
3795 default_hstate.max_huge_pages) {
3796 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3801 * No need to call hugetlb_add_hstate() as hstate already
3802 * exists. But, do set parsed_hstate so that a following
3803 * hugepages= parameter will be applied to this hstate.
3806 parsed_valid_hugepagesz = true;
3810 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3811 parsed_valid_hugepagesz = true;
3814 __setup("hugepagesz=", hugepagesz_setup);
3817 * default_hugepagesz command line input
3818 * Only one instance of default_hugepagesz allowed on command line.
3820 static int __init default_hugepagesz_setup(char *s)
3824 parsed_valid_hugepagesz = false;
3825 if (parsed_default_hugepagesz) {
3826 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3830 size = (unsigned long)memparse(s, NULL);
3832 if (!arch_hugetlb_valid_size(size)) {
3833 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3837 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3838 parsed_valid_hugepagesz = true;
3839 parsed_default_hugepagesz = true;
3840 default_hstate_idx = hstate_index(size_to_hstate(size));
3843 * The number of default huge pages (for this size) could have been
3844 * specified as the first hugetlb parameter: hugepages=X. If so,
3845 * then default_hstate_max_huge_pages is set. If the default huge
3846 * page size is gigantic (>= MAX_ORDER), then the pages must be
3847 * allocated here from bootmem allocator.
3849 if (default_hstate_max_huge_pages) {
3850 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3851 if (hstate_is_gigantic(&default_hstate))
3852 hugetlb_hstate_alloc_pages(&default_hstate);
3853 default_hstate_max_huge_pages = 0;
3858 __setup("default_hugepagesz=", default_hugepagesz_setup);
3860 static unsigned int allowed_mems_nr(struct hstate *h)
3863 unsigned int nr = 0;
3864 nodemask_t *mpol_allowed;
3865 unsigned int *array = h->free_huge_pages_node;
3866 gfp_t gfp_mask = htlb_alloc_mask(h);
3868 mpol_allowed = policy_nodemask_current(gfp_mask);
3870 for_each_node_mask(node, cpuset_current_mems_allowed) {
3871 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3878 #ifdef CONFIG_SYSCTL
3879 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3880 void *buffer, size_t *length,
3881 loff_t *ppos, unsigned long *out)
3883 struct ctl_table dup_table;
3886 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3887 * can duplicate the @table and alter the duplicate of it.
3890 dup_table.data = out;
3892 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3895 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3896 struct ctl_table *table, int write,
3897 void *buffer, size_t *length, loff_t *ppos)
3899 struct hstate *h = &default_hstate;
3900 unsigned long tmp = h->max_huge_pages;
3903 if (!hugepages_supported())
3906 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3912 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3913 NUMA_NO_NODE, tmp, *length);
3918 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3919 void *buffer, size_t *length, loff_t *ppos)
3922 return hugetlb_sysctl_handler_common(false, table, write,
3923 buffer, length, ppos);
3927 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3928 void *buffer, size_t *length, loff_t *ppos)
3930 return hugetlb_sysctl_handler_common(true, table, write,
3931 buffer, length, ppos);
3933 #endif /* CONFIG_NUMA */
3935 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3936 void *buffer, size_t *length, loff_t *ppos)
3938 struct hstate *h = &default_hstate;
3942 if (!hugepages_supported())
3945 tmp = h->nr_overcommit_huge_pages;
3947 if (write && hstate_is_gigantic(h))
3950 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3956 spin_lock_irq(&hugetlb_lock);
3957 h->nr_overcommit_huge_pages = tmp;
3958 spin_unlock_irq(&hugetlb_lock);
3964 #endif /* CONFIG_SYSCTL */
3966 void hugetlb_report_meminfo(struct seq_file *m)
3969 unsigned long total = 0;
3971 if (!hugepages_supported())
3974 for_each_hstate(h) {
3975 unsigned long count = h->nr_huge_pages;
3977 total += huge_page_size(h) * count;
3979 if (h == &default_hstate)
3981 "HugePages_Total: %5lu\n"
3982 "HugePages_Free: %5lu\n"
3983 "HugePages_Rsvd: %5lu\n"
3984 "HugePages_Surp: %5lu\n"
3985 "Hugepagesize: %8lu kB\n",
3989 h->surplus_huge_pages,
3990 huge_page_size(h) / SZ_1K);
3993 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
3996 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3998 struct hstate *h = &default_hstate;
4000 if (!hugepages_supported())
4003 return sysfs_emit_at(buf, len,
4004 "Node %d HugePages_Total: %5u\n"
4005 "Node %d HugePages_Free: %5u\n"
4006 "Node %d HugePages_Surp: %5u\n",
4007 nid, h->nr_huge_pages_node[nid],
4008 nid, h->free_huge_pages_node[nid],
4009 nid, h->surplus_huge_pages_node[nid]);
4012 void hugetlb_show_meminfo(void)
4017 if (!hugepages_supported())
4020 for_each_node_state(nid, N_MEMORY)
4022 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4024 h->nr_huge_pages_node[nid],
4025 h->free_huge_pages_node[nid],
4026 h->surplus_huge_pages_node[nid],
4027 huge_page_size(h) / SZ_1K);
4030 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4032 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4033 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
4036 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
4037 unsigned long hugetlb_total_pages(void)
4040 unsigned long nr_total_pages = 0;
4043 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4044 return nr_total_pages;
4047 static int hugetlb_acct_memory(struct hstate *h, long delta)
4054 spin_lock_irq(&hugetlb_lock);
4056 * When cpuset is configured, it breaks the strict hugetlb page
4057 * reservation as the accounting is done on a global variable. Such
4058 * reservation is completely rubbish in the presence of cpuset because
4059 * the reservation is not checked against page availability for the
4060 * current cpuset. Application can still potentially OOM'ed by kernel
4061 * with lack of free htlb page in cpuset that the task is in.
4062 * Attempt to enforce strict accounting with cpuset is almost
4063 * impossible (or too ugly) because cpuset is too fluid that
4064 * task or memory node can be dynamically moved between cpusets.
4066 * The change of semantics for shared hugetlb mapping with cpuset is
4067 * undesirable. However, in order to preserve some of the semantics,
4068 * we fall back to check against current free page availability as
4069 * a best attempt and hopefully to minimize the impact of changing
4070 * semantics that cpuset has.
4072 * Apart from cpuset, we also have memory policy mechanism that
4073 * also determines from which node the kernel will allocate memory
4074 * in a NUMA system. So similar to cpuset, we also should consider
4075 * the memory policy of the current task. Similar to the description
4079 if (gather_surplus_pages(h, delta) < 0)
4082 if (delta > allowed_mems_nr(h)) {
4083 return_unused_surplus_pages(h, delta);
4090 return_unused_surplus_pages(h, (unsigned long) -delta);
4093 spin_unlock_irq(&hugetlb_lock);
4097 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4099 struct resv_map *resv = vma_resv_map(vma);
4102 * This new VMA should share its siblings reservation map if present.
4103 * The VMA will only ever have a valid reservation map pointer where
4104 * it is being copied for another still existing VMA. As that VMA
4105 * has a reference to the reservation map it cannot disappear until
4106 * after this open call completes. It is therefore safe to take a
4107 * new reference here without additional locking.
4109 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4110 kref_get(&resv->refs);
4113 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4115 struct hstate *h = hstate_vma(vma);
4116 struct resv_map *resv = vma_resv_map(vma);
4117 struct hugepage_subpool *spool = subpool_vma(vma);
4118 unsigned long reserve, start, end;
4121 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4124 start = vma_hugecache_offset(h, vma, vma->vm_start);
4125 end = vma_hugecache_offset(h, vma, vma->vm_end);
4127 reserve = (end - start) - region_count(resv, start, end);
4128 hugetlb_cgroup_uncharge_counter(resv, start, end);
4131 * Decrement reserve counts. The global reserve count may be
4132 * adjusted if the subpool has a minimum size.
4134 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4135 hugetlb_acct_memory(h, -gbl_reserve);
4138 kref_put(&resv->refs, resv_map_release);
4141 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4143 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4148 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4150 return huge_page_size(hstate_vma(vma));
4154 * We cannot handle pagefaults against hugetlb pages at all. They cause
4155 * handle_mm_fault() to try to instantiate regular-sized pages in the
4156 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4159 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4166 * When a new function is introduced to vm_operations_struct and added
4167 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4168 * This is because under System V memory model, mappings created via
4169 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4170 * their original vm_ops are overwritten with shm_vm_ops.
4172 const struct vm_operations_struct hugetlb_vm_ops = {
4173 .fault = hugetlb_vm_op_fault,
4174 .open = hugetlb_vm_op_open,
4175 .close = hugetlb_vm_op_close,
4176 .may_split = hugetlb_vm_op_split,
4177 .pagesize = hugetlb_vm_op_pagesize,
4180 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4184 unsigned int shift = huge_page_shift(hstate_vma(vma));
4187 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4188 vma->vm_page_prot)));
4190 entry = huge_pte_wrprotect(mk_huge_pte(page,
4191 vma->vm_page_prot));
4193 entry = pte_mkyoung(entry);
4194 entry = pte_mkhuge(entry);
4195 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4200 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4201 unsigned long address, pte_t *ptep)
4205 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4206 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4207 update_mmu_cache(vma, address, ptep);
4210 bool is_hugetlb_entry_migration(pte_t pte)
4214 if (huge_pte_none(pte) || pte_present(pte))
4216 swp = pte_to_swp_entry(pte);
4217 if (is_migration_entry(swp))
4223 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4227 if (huge_pte_none(pte) || pte_present(pte))
4229 swp = pte_to_swp_entry(pte);
4230 if (is_hwpoison_entry(swp))
4237 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4238 struct page *new_page)
4240 __SetPageUptodate(new_page);
4241 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4242 hugepage_add_new_anon_rmap(new_page, vma, addr);
4243 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4244 ClearHPageRestoreReserve(new_page);
4245 SetHPageMigratable(new_page);
4248 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4249 struct vm_area_struct *vma)
4251 pte_t *src_pte, *dst_pte, entry, dst_entry;
4252 struct page *ptepage;
4254 bool cow = is_cow_mapping(vma->vm_flags);
4255 struct hstate *h = hstate_vma(vma);
4256 unsigned long sz = huge_page_size(h);
4257 unsigned long npages = pages_per_huge_page(h);
4258 struct address_space *mapping = vma->vm_file->f_mapping;
4259 struct mmu_notifier_range range;
4263 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
4266 mmu_notifier_invalidate_range_start(&range);
4269 * For shared mappings i_mmap_rwsem must be held to call
4270 * huge_pte_alloc, otherwise the returned ptep could go
4271 * away if part of a shared pmd and another thread calls
4274 i_mmap_lock_read(mapping);
4277 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
4278 spinlock_t *src_ptl, *dst_ptl;
4279 src_pte = huge_pte_offset(src, addr, sz);
4282 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
4289 * If the pagetables are shared don't copy or take references.
4290 * dst_pte == src_pte is the common case of src/dest sharing.
4292 * However, src could have 'unshared' and dst shares with
4293 * another vma. If dst_pte !none, this implies sharing.
4294 * Check here before taking page table lock, and once again
4295 * after taking the lock below.
4297 dst_entry = huge_ptep_get(dst_pte);
4298 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4301 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4302 src_ptl = huge_pte_lockptr(h, src, src_pte);
4303 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4304 entry = huge_ptep_get(src_pte);
4305 dst_entry = huge_ptep_get(dst_pte);
4307 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4309 * Skip if src entry none. Also, skip in the
4310 * unlikely case dst entry !none as this implies
4311 * sharing with another vma.
4314 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4315 is_hugetlb_entry_hwpoisoned(entry))) {
4316 swp_entry_t swp_entry = pte_to_swp_entry(entry);
4318 if (is_writable_migration_entry(swp_entry) && cow) {
4320 * COW mappings require pages in both
4321 * parent and child to be set to read.
4323 swp_entry = make_readable_migration_entry(
4324 swp_offset(swp_entry));
4325 entry = swp_entry_to_pte(swp_entry);
4326 set_huge_swap_pte_at(src, addr, src_pte,
4329 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4331 entry = huge_ptep_get(src_pte);
4332 ptepage = pte_page(entry);
4336 * This is a rare case where we see pinned hugetlb
4337 * pages while they're prone to COW. We need to do the
4338 * COW earlier during fork.
4340 * When pre-allocating the page or copying data, we
4341 * need to be without the pgtable locks since we could
4342 * sleep during the process.
4344 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4345 pte_t src_pte_old = entry;
4348 spin_unlock(src_ptl);
4349 spin_unlock(dst_ptl);
4350 /* Do not use reserve as it's private owned */
4351 new = alloc_huge_page(vma, addr, 1);
4357 copy_user_huge_page(new, ptepage, addr, vma,
4361 /* Install the new huge page if src pte stable */
4362 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4363 src_ptl = huge_pte_lockptr(h, src, src_pte);
4364 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4365 entry = huge_ptep_get(src_pte);
4366 if (!pte_same(src_pte_old, entry)) {
4367 restore_reserve_on_error(h, vma, addr,
4370 /* dst_entry won't change as in child */
4373 hugetlb_install_page(vma, dst_pte, addr, new);
4374 spin_unlock(src_ptl);
4375 spin_unlock(dst_ptl);
4381 * No need to notify as we are downgrading page
4382 * table protection not changing it to point
4385 * See Documentation/vm/mmu_notifier.rst
4387 huge_ptep_set_wrprotect(src, addr, src_pte);
4388 entry = huge_pte_wrprotect(entry);
4391 page_dup_rmap(ptepage, true);
4392 set_huge_pte_at(dst, addr, dst_pte, entry);
4393 hugetlb_count_add(npages, dst);
4395 spin_unlock(src_ptl);
4396 spin_unlock(dst_ptl);
4400 mmu_notifier_invalidate_range_end(&range);
4402 i_mmap_unlock_read(mapping);
4407 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4408 unsigned long start, unsigned long end,
4409 struct page *ref_page)
4411 struct mm_struct *mm = vma->vm_mm;
4412 unsigned long address;
4417 struct hstate *h = hstate_vma(vma);
4418 unsigned long sz = huge_page_size(h);
4419 struct mmu_notifier_range range;
4421 WARN_ON(!is_vm_hugetlb_page(vma));
4422 BUG_ON(start & ~huge_page_mask(h));
4423 BUG_ON(end & ~huge_page_mask(h));
4426 * This is a hugetlb vma, all the pte entries should point
4429 tlb_change_page_size(tlb, sz);
4430 tlb_start_vma(tlb, vma);
4433 * If sharing possible, alert mmu notifiers of worst case.
4435 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4437 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4438 mmu_notifier_invalidate_range_start(&range);
4440 for (; address < end; address += sz) {
4441 ptep = huge_pte_offset(mm, address, sz);
4445 ptl = huge_pte_lock(h, mm, ptep);
4446 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4449 * We just unmapped a page of PMDs by clearing a PUD.
4450 * The caller's TLB flush range should cover this area.
4455 pte = huge_ptep_get(ptep);
4456 if (huge_pte_none(pte)) {
4462 * Migrating hugepage or HWPoisoned hugepage is already
4463 * unmapped and its refcount is dropped, so just clear pte here.
4465 if (unlikely(!pte_present(pte))) {
4466 huge_pte_clear(mm, address, ptep, sz);
4471 page = pte_page(pte);
4473 * If a reference page is supplied, it is because a specific
4474 * page is being unmapped, not a range. Ensure the page we
4475 * are about to unmap is the actual page of interest.
4478 if (page != ref_page) {
4483 * Mark the VMA as having unmapped its page so that
4484 * future faults in this VMA will fail rather than
4485 * looking like data was lost
4487 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4490 pte = huge_ptep_get_and_clear(mm, address, ptep);
4491 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4492 if (huge_pte_dirty(pte))
4493 set_page_dirty(page);
4495 hugetlb_count_sub(pages_per_huge_page(h), mm);
4496 page_remove_rmap(page, true);
4499 tlb_remove_page_size(tlb, page, huge_page_size(h));
4501 * Bail out after unmapping reference page if supplied
4506 mmu_notifier_invalidate_range_end(&range);
4507 tlb_end_vma(tlb, vma);
4510 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4511 struct vm_area_struct *vma, unsigned long start,
4512 unsigned long end, struct page *ref_page)
4514 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4517 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4518 * test will fail on a vma being torn down, and not grab a page table
4519 * on its way out. We're lucky that the flag has such an appropriate
4520 * name, and can in fact be safely cleared here. We could clear it
4521 * before the __unmap_hugepage_range above, but all that's necessary
4522 * is to clear it before releasing the i_mmap_rwsem. This works
4523 * because in the context this is called, the VMA is about to be
4524 * destroyed and the i_mmap_rwsem is held.
4526 vma->vm_flags &= ~VM_MAYSHARE;
4529 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4530 unsigned long end, struct page *ref_page)
4532 struct mmu_gather tlb;
4534 tlb_gather_mmu(&tlb, vma->vm_mm);
4535 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4536 tlb_finish_mmu(&tlb);
4540 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4541 * mapping it owns the reserve page for. The intention is to unmap the page
4542 * from other VMAs and let the children be SIGKILLed if they are faulting the
4545 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4546 struct page *page, unsigned long address)
4548 struct hstate *h = hstate_vma(vma);
4549 struct vm_area_struct *iter_vma;
4550 struct address_space *mapping;
4554 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4555 * from page cache lookup which is in HPAGE_SIZE units.
4557 address = address & huge_page_mask(h);
4558 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4560 mapping = vma->vm_file->f_mapping;
4563 * Take the mapping lock for the duration of the table walk. As
4564 * this mapping should be shared between all the VMAs,
4565 * __unmap_hugepage_range() is called as the lock is already held
4567 i_mmap_lock_write(mapping);
4568 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4569 /* Do not unmap the current VMA */
4570 if (iter_vma == vma)
4574 * Shared VMAs have their own reserves and do not affect
4575 * MAP_PRIVATE accounting but it is possible that a shared
4576 * VMA is using the same page so check and skip such VMAs.
4578 if (iter_vma->vm_flags & VM_MAYSHARE)
4582 * Unmap the page from other VMAs without their own reserves.
4583 * They get marked to be SIGKILLed if they fault in these
4584 * areas. This is because a future no-page fault on this VMA
4585 * could insert a zeroed page instead of the data existing
4586 * from the time of fork. This would look like data corruption
4588 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4589 unmap_hugepage_range(iter_vma, address,
4590 address + huge_page_size(h), page);
4592 i_mmap_unlock_write(mapping);
4596 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4597 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4598 * cannot race with other handlers or page migration.
4599 * Keep the pte_same checks anyway to make transition from the mutex easier.
4601 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4602 unsigned long address, pte_t *ptep,
4603 struct page *pagecache_page, spinlock_t *ptl)
4606 struct hstate *h = hstate_vma(vma);
4607 struct page *old_page, *new_page;
4608 int outside_reserve = 0;
4610 unsigned long haddr = address & huge_page_mask(h);
4611 struct mmu_notifier_range range;
4613 pte = huge_ptep_get(ptep);
4614 old_page = pte_page(pte);
4617 /* If no-one else is actually using this page, avoid the copy
4618 * and just make the page writable */
4619 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4620 page_move_anon_rmap(old_page, vma);
4621 set_huge_ptep_writable(vma, haddr, ptep);
4626 * If the process that created a MAP_PRIVATE mapping is about to
4627 * perform a COW due to a shared page count, attempt to satisfy
4628 * the allocation without using the existing reserves. The pagecache
4629 * page is used to determine if the reserve at this address was
4630 * consumed or not. If reserves were used, a partial faulted mapping
4631 * at the time of fork() could consume its reserves on COW instead
4632 * of the full address range.
4634 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4635 old_page != pagecache_page)
4636 outside_reserve = 1;
4641 * Drop page table lock as buddy allocator may be called. It will
4642 * be acquired again before returning to the caller, as expected.
4645 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4647 if (IS_ERR(new_page)) {
4649 * If a process owning a MAP_PRIVATE mapping fails to COW,
4650 * it is due to references held by a child and an insufficient
4651 * huge page pool. To guarantee the original mappers
4652 * reliability, unmap the page from child processes. The child
4653 * may get SIGKILLed if it later faults.
4655 if (outside_reserve) {
4656 struct address_space *mapping = vma->vm_file->f_mapping;
4661 BUG_ON(huge_pte_none(pte));
4663 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4664 * unmapping. unmapping needs to hold i_mmap_rwsem
4665 * in write mode. Dropping i_mmap_rwsem in read mode
4666 * here is OK as COW mappings do not interact with
4669 * Reacquire both after unmap operation.
4671 idx = vma_hugecache_offset(h, vma, haddr);
4672 hash = hugetlb_fault_mutex_hash(mapping, idx);
4673 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4674 i_mmap_unlock_read(mapping);
4676 unmap_ref_private(mm, vma, old_page, haddr);
4678 i_mmap_lock_read(mapping);
4679 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4681 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4683 pte_same(huge_ptep_get(ptep), pte)))
4684 goto retry_avoidcopy;
4686 * race occurs while re-acquiring page table
4687 * lock, and our job is done.
4692 ret = vmf_error(PTR_ERR(new_page));
4693 goto out_release_old;
4697 * When the original hugepage is shared one, it does not have
4698 * anon_vma prepared.
4700 if (unlikely(anon_vma_prepare(vma))) {
4702 goto out_release_all;
4705 copy_user_huge_page(new_page, old_page, address, vma,
4706 pages_per_huge_page(h));
4707 __SetPageUptodate(new_page);
4709 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4710 haddr + huge_page_size(h));
4711 mmu_notifier_invalidate_range_start(&range);
4714 * Retake the page table lock to check for racing updates
4715 * before the page tables are altered
4718 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4719 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4720 ClearHPageRestoreReserve(new_page);
4723 huge_ptep_clear_flush(vma, haddr, ptep);
4724 mmu_notifier_invalidate_range(mm, range.start, range.end);
4725 set_huge_pte_at(mm, haddr, ptep,
4726 make_huge_pte(vma, new_page, 1));
4727 page_remove_rmap(old_page, true);
4728 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4729 SetHPageMigratable(new_page);
4730 /* Make the old page be freed below */
4731 new_page = old_page;
4734 mmu_notifier_invalidate_range_end(&range);
4736 /* No restore in case of successful pagetable update (Break COW) */
4737 if (new_page != old_page)
4738 restore_reserve_on_error(h, vma, haddr, new_page);
4743 spin_lock(ptl); /* Caller expects lock to be held */
4747 /* Return the pagecache page at a given address within a VMA */
4748 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4749 struct vm_area_struct *vma, unsigned long address)
4751 struct address_space *mapping;
4754 mapping = vma->vm_file->f_mapping;
4755 idx = vma_hugecache_offset(h, vma, address);
4757 return find_lock_page(mapping, idx);
4761 * Return whether there is a pagecache page to back given address within VMA.
4762 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4764 static bool hugetlbfs_pagecache_present(struct hstate *h,
4765 struct vm_area_struct *vma, unsigned long address)
4767 struct address_space *mapping;
4771 mapping = vma->vm_file->f_mapping;
4772 idx = vma_hugecache_offset(h, vma, address);
4774 page = find_get_page(mapping, idx);
4777 return page != NULL;
4780 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4783 struct inode *inode = mapping->host;
4784 struct hstate *h = hstate_inode(inode);
4785 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4789 ClearHPageRestoreReserve(page);
4792 * set page dirty so that it will not be removed from cache/file
4793 * by non-hugetlbfs specific code paths.
4795 set_page_dirty(page);
4797 spin_lock(&inode->i_lock);
4798 inode->i_blocks += blocks_per_huge_page(h);
4799 spin_unlock(&inode->i_lock);
4803 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
4804 struct address_space *mapping,
4807 unsigned long haddr,
4808 unsigned long reason)
4812 struct vm_fault vmf = {
4818 * Hard to debug if it ends up being
4819 * used by a callee that assumes
4820 * something about the other
4821 * uninitialized fields... same as in
4827 * hugetlb_fault_mutex and i_mmap_rwsem must be
4828 * dropped before handling userfault. Reacquire
4829 * after handling fault to make calling code simpler.
4831 hash = hugetlb_fault_mutex_hash(mapping, idx);
4832 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4833 i_mmap_unlock_read(mapping);
4834 ret = handle_userfault(&vmf, reason);
4835 i_mmap_lock_read(mapping);
4836 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4841 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4842 struct vm_area_struct *vma,
4843 struct address_space *mapping, pgoff_t idx,
4844 unsigned long address, pte_t *ptep, unsigned int flags)
4846 struct hstate *h = hstate_vma(vma);
4847 vm_fault_t ret = VM_FAULT_SIGBUS;
4853 unsigned long haddr = address & huge_page_mask(h);
4854 bool new_page, new_pagecache_page = false;
4857 * Currently, we are forced to kill the process in the event the
4858 * original mapper has unmapped pages from the child due to a failed
4859 * COW. Warn that such a situation has occurred as it may not be obvious
4861 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4862 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4868 * We can not race with truncation due to holding i_mmap_rwsem.
4869 * i_size is modified when holding i_mmap_rwsem, so check here
4870 * once for faults beyond end of file.
4872 size = i_size_read(mapping->host) >> huge_page_shift(h);
4878 page = find_lock_page(mapping, idx);
4880 /* Check for page in userfault range */
4881 if (userfaultfd_missing(vma)) {
4882 ret = hugetlb_handle_userfault(vma, mapping, idx,
4888 page = alloc_huge_page(vma, haddr, 0);
4891 * Returning error will result in faulting task being
4892 * sent SIGBUS. The hugetlb fault mutex prevents two
4893 * tasks from racing to fault in the same page which
4894 * could result in false unable to allocate errors.
4895 * Page migration does not take the fault mutex, but
4896 * does a clear then write of pte's under page table
4897 * lock. Page fault code could race with migration,
4898 * notice the clear pte and try to allocate a page
4899 * here. Before returning error, get ptl and make
4900 * sure there really is no pte entry.
4902 ptl = huge_pte_lock(h, mm, ptep);
4904 if (huge_pte_none(huge_ptep_get(ptep)))
4905 ret = vmf_error(PTR_ERR(page));
4909 clear_huge_page(page, address, pages_per_huge_page(h));
4910 __SetPageUptodate(page);
4913 if (vma->vm_flags & VM_MAYSHARE) {
4914 int err = huge_add_to_page_cache(page, mapping, idx);
4921 new_pagecache_page = true;
4924 if (unlikely(anon_vma_prepare(vma))) {
4926 goto backout_unlocked;
4932 * If memory error occurs between mmap() and fault, some process
4933 * don't have hwpoisoned swap entry for errored virtual address.
4934 * So we need to block hugepage fault by PG_hwpoison bit check.
4936 if (unlikely(PageHWPoison(page))) {
4937 ret = VM_FAULT_HWPOISON_LARGE |
4938 VM_FAULT_SET_HINDEX(hstate_index(h));
4939 goto backout_unlocked;
4942 /* Check for page in userfault range. */
4943 if (userfaultfd_minor(vma)) {
4946 ret = hugetlb_handle_userfault(vma, mapping, idx,
4954 * If we are going to COW a private mapping later, we examine the
4955 * pending reservations for this page now. This will ensure that
4956 * any allocations necessary to record that reservation occur outside
4959 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4960 if (vma_needs_reservation(h, vma, haddr) < 0) {
4962 goto backout_unlocked;
4964 /* Just decrements count, does not deallocate */
4965 vma_end_reservation(h, vma, haddr);
4968 ptl = huge_pte_lock(h, mm, ptep);
4970 if (!huge_pte_none(huge_ptep_get(ptep)))
4974 ClearHPageRestoreReserve(page);
4975 hugepage_add_new_anon_rmap(page, vma, haddr);
4977 page_dup_rmap(page, true);
4978 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4979 && (vma->vm_flags & VM_SHARED)));
4980 set_huge_pte_at(mm, haddr, ptep, new_pte);
4982 hugetlb_count_add(pages_per_huge_page(h), mm);
4983 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4984 /* Optimization, do the COW without a second fault */
4985 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4991 * Only set HPageMigratable in newly allocated pages. Existing pages
4992 * found in the pagecache may not have HPageMigratableset if they have
4993 * been isolated for migration.
4996 SetHPageMigratable(page);
5006 /* restore reserve for newly allocated pages not in page cache */
5007 if (new_page && !new_pagecache_page)
5008 restore_reserve_on_error(h, vma, haddr, page);
5014 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5016 unsigned long key[2];
5019 key[0] = (unsigned long) mapping;
5022 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
5024 return hash & (num_fault_mutexes - 1);
5028 * For uniprocessor systems we always use a single mutex, so just
5029 * return 0 and avoid the hashing overhead.
5031 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
5037 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
5038 unsigned long address, unsigned int flags)
5045 struct page *page = NULL;
5046 struct page *pagecache_page = NULL;
5047 struct hstate *h = hstate_vma(vma);
5048 struct address_space *mapping;
5049 int need_wait_lock = 0;
5050 unsigned long haddr = address & huge_page_mask(h);
5052 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
5055 * Since we hold no locks, ptep could be stale. That is
5056 * OK as we are only making decisions based on content and
5057 * not actually modifying content here.
5059 entry = huge_ptep_get(ptep);
5060 if (unlikely(is_hugetlb_entry_migration(entry))) {
5061 migration_entry_wait_huge(vma, mm, ptep);
5063 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
5064 return VM_FAULT_HWPOISON_LARGE |
5065 VM_FAULT_SET_HINDEX(hstate_index(h));
5069 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
5070 * until finished with ptep. This serves two purposes:
5071 * 1) It prevents huge_pmd_unshare from being called elsewhere
5072 * and making the ptep no longer valid.
5073 * 2) It synchronizes us with i_size modifications during truncation.
5075 * ptep could have already be assigned via huge_pte_offset. That
5076 * is OK, as huge_pte_alloc will return the same value unless
5077 * something has changed.
5079 mapping = vma->vm_file->f_mapping;
5080 i_mmap_lock_read(mapping);
5081 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
5083 i_mmap_unlock_read(mapping);
5084 return VM_FAULT_OOM;
5088 * Serialize hugepage allocation and instantiation, so that we don't
5089 * get spurious allocation failures if two CPUs race to instantiate
5090 * the same page in the page cache.
5092 idx = vma_hugecache_offset(h, vma, haddr);
5093 hash = hugetlb_fault_mutex_hash(mapping, idx);
5094 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5096 entry = huge_ptep_get(ptep);
5097 if (huge_pte_none(entry)) {
5098 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
5105 * entry could be a migration/hwpoison entry at this point, so this
5106 * check prevents the kernel from going below assuming that we have
5107 * an active hugepage in pagecache. This goto expects the 2nd page
5108 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
5109 * properly handle it.
5111 if (!pte_present(entry))
5115 * If we are going to COW the mapping later, we examine the pending
5116 * reservations for this page now. This will ensure that any
5117 * allocations necessary to record that reservation occur outside the
5118 * spinlock. For private mappings, we also lookup the pagecache
5119 * page now as it is used to determine if a reservation has been
5122 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
5123 if (vma_needs_reservation(h, vma, haddr) < 0) {
5127 /* Just decrements count, does not deallocate */
5128 vma_end_reservation(h, vma, haddr);
5130 if (!(vma->vm_flags & VM_MAYSHARE))
5131 pagecache_page = hugetlbfs_pagecache_page(h,
5135 ptl = huge_pte_lock(h, mm, ptep);
5137 /* Check for a racing update before calling hugetlb_cow */
5138 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5142 * hugetlb_cow() requires page locks of pte_page(entry) and
5143 * pagecache_page, so here we need take the former one
5144 * when page != pagecache_page or !pagecache_page.
5146 page = pte_page(entry);
5147 if (page != pagecache_page)
5148 if (!trylock_page(page)) {
5155 if (flags & FAULT_FLAG_WRITE) {
5156 if (!huge_pte_write(entry)) {
5157 ret = hugetlb_cow(mm, vma, address, ptep,
5158 pagecache_page, ptl);
5161 entry = huge_pte_mkdirty(entry);
5163 entry = pte_mkyoung(entry);
5164 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5165 flags & FAULT_FLAG_WRITE))
5166 update_mmu_cache(vma, haddr, ptep);
5168 if (page != pagecache_page)
5174 if (pagecache_page) {
5175 unlock_page(pagecache_page);
5176 put_page(pagecache_page);
5179 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5180 i_mmap_unlock_read(mapping);
5182 * Generally it's safe to hold refcount during waiting page lock. But
5183 * here we just wait to defer the next page fault to avoid busy loop and
5184 * the page is not used after unlocked before returning from the current
5185 * page fault. So we are safe from accessing freed page, even if we wait
5186 * here without taking refcount.
5189 wait_on_page_locked(page);
5193 #ifdef CONFIG_USERFAULTFD
5195 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5196 * modifications for huge pages.
5198 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5200 struct vm_area_struct *dst_vma,
5201 unsigned long dst_addr,
5202 unsigned long src_addr,
5203 enum mcopy_atomic_mode mode,
5204 struct page **pagep)
5206 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5207 struct hstate *h = hstate_vma(dst_vma);
5208 struct address_space *mapping = dst_vma->vm_file->f_mapping;
5209 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5211 int vm_shared = dst_vma->vm_flags & VM_SHARED;
5217 bool new_pagecache_page = false;
5221 page = find_lock_page(mapping, idx);
5224 } else if (!*pagep) {
5225 /* If a page already exists, then it's UFFDIO_COPY for
5226 * a non-missing case. Return -EEXIST.
5229 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5234 page = alloc_huge_page(dst_vma, dst_addr, 0);
5240 ret = copy_huge_page_from_user(page,
5241 (const void __user *) src_addr,
5242 pages_per_huge_page(h), false);
5244 /* fallback to copy_from_user outside mmap_lock */
5245 if (unlikely(ret)) {
5247 /* Free the allocated page which may have
5248 * consumed a reservation.
5250 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5253 /* Allocate a temporary page to hold the copied
5256 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5262 /* Set the outparam pagep and return to the caller to
5263 * copy the contents outside the lock. Don't free the
5270 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5277 page = alloc_huge_page(dst_vma, dst_addr, 0);
5283 copy_huge_page(page, *pagep);
5289 * The memory barrier inside __SetPageUptodate makes sure that
5290 * preceding stores to the page contents become visible before
5291 * the set_pte_at() write.
5293 __SetPageUptodate(page);
5295 /* Add shared, newly allocated pages to the page cache. */
5296 if (vm_shared && !is_continue) {
5297 size = i_size_read(mapping->host) >> huge_page_shift(h);
5300 goto out_release_nounlock;
5303 * Serialization between remove_inode_hugepages() and
5304 * huge_add_to_page_cache() below happens through the
5305 * hugetlb_fault_mutex_table that here must be hold by
5308 ret = huge_add_to_page_cache(page, mapping, idx);
5310 goto out_release_nounlock;
5311 new_pagecache_page = true;
5314 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5318 * Recheck the i_size after holding PT lock to make sure not
5319 * to leave any page mapped (as page_mapped()) beyond the end
5320 * of the i_size (remove_inode_hugepages() is strict about
5321 * enforcing that). If we bail out here, we'll also leave a
5322 * page in the radix tree in the vm_shared case beyond the end
5323 * of the i_size, but remove_inode_hugepages() will take care
5324 * of it as soon as we drop the hugetlb_fault_mutex_table.
5326 size = i_size_read(mapping->host) >> huge_page_shift(h);
5329 goto out_release_unlock;
5332 if (!huge_pte_none(huge_ptep_get(dst_pte)))
5333 goto out_release_unlock;
5336 page_dup_rmap(page, true);
5338 ClearHPageRestoreReserve(page);
5339 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5342 /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5343 if (is_continue && !vm_shared)
5346 writable = dst_vma->vm_flags & VM_WRITE;
5348 _dst_pte = make_huge_pte(dst_vma, page, writable);
5350 _dst_pte = huge_pte_mkdirty(_dst_pte);
5351 _dst_pte = pte_mkyoung(_dst_pte);
5353 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5355 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5356 dst_vma->vm_flags & VM_WRITE);
5357 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5359 /* No need to invalidate - it was non-present before */
5360 update_mmu_cache(dst_vma, dst_addr, dst_pte);
5364 SetHPageMigratable(page);
5365 if (vm_shared || is_continue)
5372 if (vm_shared || is_continue)
5374 out_release_nounlock:
5375 if (!new_pagecache_page)
5376 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5380 #endif /* CONFIG_USERFAULTFD */
5382 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5383 int refs, struct page **pages,
5384 struct vm_area_struct **vmas)
5388 for (nr = 0; nr < refs; nr++) {
5390 pages[nr] = mem_map_offset(page, nr);
5396 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5397 struct page **pages, struct vm_area_struct **vmas,
5398 unsigned long *position, unsigned long *nr_pages,
5399 long i, unsigned int flags, int *locked)
5401 unsigned long pfn_offset;
5402 unsigned long vaddr = *position;
5403 unsigned long remainder = *nr_pages;
5404 struct hstate *h = hstate_vma(vma);
5405 int err = -EFAULT, refs;
5407 while (vaddr < vma->vm_end && remainder) {
5409 spinlock_t *ptl = NULL;
5414 * If we have a pending SIGKILL, don't keep faulting pages and
5415 * potentially allocating memory.
5417 if (fatal_signal_pending(current)) {
5423 * Some archs (sparc64, sh*) have multiple pte_ts to
5424 * each hugepage. We have to make sure we get the
5425 * first, for the page indexing below to work.
5427 * Note that page table lock is not held when pte is null.
5429 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5432 ptl = huge_pte_lock(h, mm, pte);
5433 absent = !pte || huge_pte_none(huge_ptep_get(pte));
5436 * When coredumping, it suits get_dump_page if we just return
5437 * an error where there's an empty slot with no huge pagecache
5438 * to back it. This way, we avoid allocating a hugepage, and
5439 * the sparse dumpfile avoids allocating disk blocks, but its
5440 * huge holes still show up with zeroes where they need to be.
5442 if (absent && (flags & FOLL_DUMP) &&
5443 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
5451 * We need call hugetlb_fault for both hugepages under migration
5452 * (in which case hugetlb_fault waits for the migration,) and
5453 * hwpoisoned hugepages (in which case we need to prevent the
5454 * caller from accessing to them.) In order to do this, we use
5455 * here is_swap_pte instead of is_hugetlb_entry_migration and
5456 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5457 * both cases, and because we can't follow correct pages
5458 * directly from any kind of swap entries.
5460 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
5461 ((flags & FOLL_WRITE) &&
5462 !huge_pte_write(huge_ptep_get(pte)))) {
5464 unsigned int fault_flags = 0;
5468 if (flags & FOLL_WRITE)
5469 fault_flags |= FAULT_FLAG_WRITE;
5471 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5472 FAULT_FLAG_KILLABLE;
5473 if (flags & FOLL_NOWAIT)
5474 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5475 FAULT_FLAG_RETRY_NOWAIT;
5476 if (flags & FOLL_TRIED) {
5478 * Note: FAULT_FLAG_ALLOW_RETRY and
5479 * FAULT_FLAG_TRIED can co-exist
5481 fault_flags |= FAULT_FLAG_TRIED;
5483 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
5484 if (ret & VM_FAULT_ERROR) {
5485 err = vm_fault_to_errno(ret, flags);
5489 if (ret & VM_FAULT_RETRY) {
5491 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
5495 * VM_FAULT_RETRY must not return an
5496 * error, it will return zero
5499 * No need to update "position" as the
5500 * caller will not check it after
5501 * *nr_pages is set to 0.
5508 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
5509 page = pte_page(huge_ptep_get(pte));
5512 * If subpage information not requested, update counters
5513 * and skip the same_page loop below.
5515 if (!pages && !vmas && !pfn_offset &&
5516 (vaddr + huge_page_size(h) < vma->vm_end) &&
5517 (remainder >= pages_per_huge_page(h))) {
5518 vaddr += huge_page_size(h);
5519 remainder -= pages_per_huge_page(h);
5520 i += pages_per_huge_page(h);
5525 /* vaddr may not be aligned to PAGE_SIZE */
5526 refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
5527 (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
5530 record_subpages_vmas(mem_map_offset(page, pfn_offset),
5532 likely(pages) ? pages + i : NULL,
5533 vmas ? vmas + i : NULL);
5537 * try_grab_compound_head() should always succeed here,
5538 * because: a) we hold the ptl lock, and b) we've just
5539 * checked that the huge page is present in the page
5540 * tables. If the huge page is present, then the tail
5541 * pages must also be present. The ptl prevents the
5542 * head page and tail pages from being rearranged in
5543 * any way. So this page must be available at this
5544 * point, unless the page refcount overflowed:
5546 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
5556 vaddr += (refs << PAGE_SHIFT);
5562 *nr_pages = remainder;
5564 * setting position is actually required only if remainder is
5565 * not zero but it's faster not to add a "if (remainder)"
5573 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5574 unsigned long address, unsigned long end, pgprot_t newprot)
5576 struct mm_struct *mm = vma->vm_mm;
5577 unsigned long start = address;
5580 struct hstate *h = hstate_vma(vma);
5581 unsigned long pages = 0;
5582 bool shared_pmd = false;
5583 struct mmu_notifier_range range;
5586 * In the case of shared PMDs, the area to flush could be beyond
5587 * start/end. Set range.start/range.end to cover the maximum possible
5588 * range if PMD sharing is possible.
5590 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5591 0, vma, mm, start, end);
5592 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5594 BUG_ON(address >= end);
5595 flush_cache_range(vma, range.start, range.end);
5597 mmu_notifier_invalidate_range_start(&range);
5598 i_mmap_lock_write(vma->vm_file->f_mapping);
5599 for (; address < end; address += huge_page_size(h)) {
5601 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5604 ptl = huge_pte_lock(h, mm, ptep);
5605 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5611 pte = huge_ptep_get(ptep);
5612 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5616 if (unlikely(is_hugetlb_entry_migration(pte))) {
5617 swp_entry_t entry = pte_to_swp_entry(pte);
5619 if (is_writable_migration_entry(entry)) {
5622 entry = make_readable_migration_entry(
5624 newpte = swp_entry_to_pte(entry);
5625 set_huge_swap_pte_at(mm, address, ptep,
5626 newpte, huge_page_size(h));
5632 if (!huge_pte_none(pte)) {
5634 unsigned int shift = huge_page_shift(hstate_vma(vma));
5636 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5637 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5638 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
5639 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5645 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5646 * may have cleared our pud entry and done put_page on the page table:
5647 * once we release i_mmap_rwsem, another task can do the final put_page
5648 * and that page table be reused and filled with junk. If we actually
5649 * did unshare a page of pmds, flush the range corresponding to the pud.
5652 flush_hugetlb_tlb_range(vma, range.start, range.end);
5654 flush_hugetlb_tlb_range(vma, start, end);
5656 * No need to call mmu_notifier_invalidate_range() we are downgrading
5657 * page table protection not changing it to point to a new page.
5659 * See Documentation/vm/mmu_notifier.rst
5661 i_mmap_unlock_write(vma->vm_file->f_mapping);
5662 mmu_notifier_invalidate_range_end(&range);
5664 return pages << h->order;
5667 /* Return true if reservation was successful, false otherwise. */
5668 bool hugetlb_reserve_pages(struct inode *inode,
5670 struct vm_area_struct *vma,
5671 vm_flags_t vm_flags)
5674 struct hstate *h = hstate_inode(inode);
5675 struct hugepage_subpool *spool = subpool_inode(inode);
5676 struct resv_map *resv_map;
5677 struct hugetlb_cgroup *h_cg = NULL;
5678 long gbl_reserve, regions_needed = 0;
5680 /* This should never happen */
5682 VM_WARN(1, "%s called with a negative range\n", __func__);
5687 * Only apply hugepage reservation if asked. At fault time, an
5688 * attempt will be made for VM_NORESERVE to allocate a page
5689 * without using reserves
5691 if (vm_flags & VM_NORESERVE)
5695 * Shared mappings base their reservation on the number of pages that
5696 * are already allocated on behalf of the file. Private mappings need
5697 * to reserve the full area even if read-only as mprotect() may be
5698 * called to make the mapping read-write. Assume !vma is a shm mapping
5700 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5702 * resv_map can not be NULL as hugetlb_reserve_pages is only
5703 * called for inodes for which resv_maps were created (see
5704 * hugetlbfs_get_inode).
5706 resv_map = inode_resv_map(inode);
5708 chg = region_chg(resv_map, from, to, ®ions_needed);
5711 /* Private mapping. */
5712 resv_map = resv_map_alloc();
5718 set_vma_resv_map(vma, resv_map);
5719 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5725 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5726 chg * pages_per_huge_page(h), &h_cg) < 0)
5729 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5730 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5733 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5737 * There must be enough pages in the subpool for the mapping. If
5738 * the subpool has a minimum size, there may be some global
5739 * reservations already in place (gbl_reserve).
5741 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5742 if (gbl_reserve < 0)
5743 goto out_uncharge_cgroup;
5746 * Check enough hugepages are available for the reservation.
5747 * Hand the pages back to the subpool if there are not
5749 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5753 * Account for the reservations made. Shared mappings record regions
5754 * that have reservations as they are shared by multiple VMAs.
5755 * When the last VMA disappears, the region map says how much
5756 * the reservation was and the page cache tells how much of
5757 * the reservation was consumed. Private mappings are per-VMA and
5758 * only the consumed reservations are tracked. When the VMA
5759 * disappears, the original reservation is the VMA size and the
5760 * consumed reservations are stored in the map. Hence, nothing
5761 * else has to be done for private mappings here
5763 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5764 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5766 if (unlikely(add < 0)) {
5767 hugetlb_acct_memory(h, -gbl_reserve);
5769 } else if (unlikely(chg > add)) {
5771 * pages in this range were added to the reserve
5772 * map between region_chg and region_add. This
5773 * indicates a race with alloc_huge_page. Adjust
5774 * the subpool and reserve counts modified above
5775 * based on the difference.
5780 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5781 * reference to h_cg->css. See comment below for detail.
5783 hugetlb_cgroup_uncharge_cgroup_rsvd(
5785 (chg - add) * pages_per_huge_page(h), h_cg);
5787 rsv_adjust = hugepage_subpool_put_pages(spool,
5789 hugetlb_acct_memory(h, -rsv_adjust);
5792 * The file_regions will hold their own reference to
5793 * h_cg->css. So we should release the reference held
5794 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5797 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5803 /* put back original number of pages, chg */
5804 (void)hugepage_subpool_put_pages(spool, chg);
5805 out_uncharge_cgroup:
5806 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5807 chg * pages_per_huge_page(h), h_cg);
5809 if (!vma || vma->vm_flags & VM_MAYSHARE)
5810 /* Only call region_abort if the region_chg succeeded but the
5811 * region_add failed or didn't run.
5813 if (chg >= 0 && add < 0)
5814 region_abort(resv_map, from, to, regions_needed);
5815 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5816 kref_put(&resv_map->refs, resv_map_release);
5820 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5823 struct hstate *h = hstate_inode(inode);
5824 struct resv_map *resv_map = inode_resv_map(inode);
5826 struct hugepage_subpool *spool = subpool_inode(inode);
5830 * Since this routine can be called in the evict inode path for all
5831 * hugetlbfs inodes, resv_map could be NULL.
5834 chg = region_del(resv_map, start, end);
5836 * region_del() can fail in the rare case where a region
5837 * must be split and another region descriptor can not be
5838 * allocated. If end == LONG_MAX, it will not fail.
5844 spin_lock(&inode->i_lock);
5845 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5846 spin_unlock(&inode->i_lock);
5849 * If the subpool has a minimum size, the number of global
5850 * reservations to be released may be adjusted.
5852 * Note that !resv_map implies freed == 0. So (chg - freed)
5853 * won't go negative.
5855 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5856 hugetlb_acct_memory(h, -gbl_reserve);
5861 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5862 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5863 struct vm_area_struct *vma,
5864 unsigned long addr, pgoff_t idx)
5866 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5868 unsigned long sbase = saddr & PUD_MASK;
5869 unsigned long s_end = sbase + PUD_SIZE;
5871 /* Allow segments to share if only one is marked locked */
5872 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5873 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5876 * match the virtual addresses, permission and the alignment of the
5879 if (pmd_index(addr) != pmd_index(saddr) ||
5880 vm_flags != svm_flags ||
5881 !range_in_vma(svma, sbase, s_end))
5887 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5889 unsigned long base = addr & PUD_MASK;
5890 unsigned long end = base + PUD_SIZE;
5893 * check on proper vm_flags and page table alignment
5895 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5900 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5902 #ifdef CONFIG_USERFAULTFD
5903 if (uffd_disable_huge_pmd_share(vma))
5906 return vma_shareable(vma, addr);
5910 * Determine if start,end range within vma could be mapped by shared pmd.
5911 * If yes, adjust start and end to cover range associated with possible
5912 * shared pmd mappings.
5914 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5915 unsigned long *start, unsigned long *end)
5917 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5918 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5921 * vma needs to span at least one aligned PUD size, and the range
5922 * must be at least partially within in.
5924 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5925 (*end <= v_start) || (*start >= v_end))
5928 /* Extend the range to be PUD aligned for a worst case scenario */
5929 if (*start > v_start)
5930 *start = ALIGN_DOWN(*start, PUD_SIZE);
5933 *end = ALIGN(*end, PUD_SIZE);
5937 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5938 * and returns the corresponding pte. While this is not necessary for the
5939 * !shared pmd case because we can allocate the pmd later as well, it makes the
5940 * code much cleaner.
5942 * This routine must be called with i_mmap_rwsem held in at least read mode if
5943 * sharing is possible. For hugetlbfs, this prevents removal of any page
5944 * table entries associated with the address space. This is important as we
5945 * are setting up sharing based on existing page table entries (mappings).
5947 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5948 * huge_pte_alloc know that sharing is not possible and do not take
5949 * i_mmap_rwsem as a performance optimization. This is handled by the
5950 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5951 * only required for subsequent processing.
5953 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5954 unsigned long addr, pud_t *pud)
5956 struct address_space *mapping = vma->vm_file->f_mapping;
5957 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5959 struct vm_area_struct *svma;
5960 unsigned long saddr;
5965 i_mmap_assert_locked(mapping);
5966 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5970 saddr = page_table_shareable(svma, vma, addr, idx);
5972 spte = huge_pte_offset(svma->vm_mm, saddr,
5973 vma_mmu_pagesize(svma));
5975 get_page(virt_to_page(spte));
5984 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5985 if (pud_none(*pud)) {
5986 pud_populate(mm, pud,
5987 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5990 put_page(virt_to_page(spte));
5994 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5999 * unmap huge page backed by shared pte.
6001 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
6002 * indicated by page_count > 1, unmap is achieved by clearing pud and
6003 * decrementing the ref count. If count == 1, the pte page is not shared.
6005 * Called with page table lock held and i_mmap_rwsem held in write mode.
6007 * returns: 1 successfully unmapped a shared pte page
6008 * 0 the underlying pte page is not shared, or it is the last user
6010 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6011 unsigned long *addr, pte_t *ptep)
6013 pgd_t *pgd = pgd_offset(mm, *addr);
6014 p4d_t *p4d = p4d_offset(pgd, *addr);
6015 pud_t *pud = pud_offset(p4d, *addr);
6017 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
6018 BUG_ON(page_count(virt_to_page(ptep)) == 0);
6019 if (page_count(virt_to_page(ptep)) == 1)
6023 put_page(virt_to_page(ptep));
6025 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
6029 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6030 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6031 unsigned long addr, pud_t *pud)
6036 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
6037 unsigned long *addr, pte_t *ptep)
6042 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6043 unsigned long *start, unsigned long *end)
6047 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6051 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
6053 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
6054 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
6055 unsigned long addr, unsigned long sz)
6062 pgd = pgd_offset(mm, addr);
6063 p4d = p4d_alloc(mm, pgd, addr);
6066 pud = pud_alloc(mm, p4d, addr);
6068 if (sz == PUD_SIZE) {
6071 BUG_ON(sz != PMD_SIZE);
6072 if (want_pmd_share(vma, addr) && pud_none(*pud))
6073 pte = huge_pmd_share(mm, vma, addr, pud);
6075 pte = (pte_t *)pmd_alloc(mm, pud, addr);
6078 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
6084 * huge_pte_offset() - Walk the page table to resolve the hugepage
6085 * entry at address @addr
6087 * Return: Pointer to page table entry (PUD or PMD) for
6088 * address @addr, or NULL if a !p*d_present() entry is encountered and the
6089 * size @sz doesn't match the hugepage size at this level of the page
6092 pte_t *huge_pte_offset(struct mm_struct *mm,
6093 unsigned long addr, unsigned long sz)
6100 pgd = pgd_offset(mm, addr);
6101 if (!pgd_present(*pgd))
6103 p4d = p4d_offset(pgd, addr);
6104 if (!p4d_present(*p4d))
6107 pud = pud_offset(p4d, addr);
6109 /* must be pud huge, non-present or none */
6110 return (pte_t *)pud;
6111 if (!pud_present(*pud))
6113 /* must have a valid entry and size to go further */
6115 pmd = pmd_offset(pud, addr);
6116 /* must be pmd huge, non-present or none */
6117 return (pte_t *)pmd;
6120 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
6123 * These functions are overwritable if your architecture needs its own
6126 struct page * __weak
6127 follow_huge_addr(struct mm_struct *mm, unsigned long address,
6130 return ERR_PTR(-EINVAL);
6133 struct page * __weak
6134 follow_huge_pd(struct vm_area_struct *vma,
6135 unsigned long address, hugepd_t hpd, int flags, int pdshift)
6137 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6141 struct page * __weak
6142 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6143 pmd_t *pmd, int flags)
6145 struct page *page = NULL;
6149 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6150 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6151 (FOLL_PIN | FOLL_GET)))
6155 ptl = pmd_lockptr(mm, pmd);
6158 * make sure that the address range covered by this pmd is not
6159 * unmapped from other threads.
6161 if (!pmd_huge(*pmd))
6163 pte = huge_ptep_get((pte_t *)pmd);
6164 if (pte_present(pte)) {
6165 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6167 * try_grab_page() should always succeed here, because: a) we
6168 * hold the pmd (ptl) lock, and b) we've just checked that the
6169 * huge pmd (head) page is present in the page tables. The ptl
6170 * prevents the head page and tail pages from being rearranged
6171 * in any way. So this page must be available at this point,
6172 * unless the page refcount overflowed:
6174 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6179 if (is_hugetlb_entry_migration(pte)) {
6181 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
6185 * hwpoisoned entry is treated as no_page_table in
6186 * follow_page_mask().
6194 struct page * __weak
6195 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6196 pud_t *pud, int flags)
6198 if (flags & (FOLL_GET | FOLL_PIN))
6201 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6204 struct page * __weak
6205 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6207 if (flags & (FOLL_GET | FOLL_PIN))
6210 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6213 bool isolate_huge_page(struct page *page, struct list_head *list)
6217 spin_lock_irq(&hugetlb_lock);
6218 if (!PageHeadHuge(page) ||
6219 !HPageMigratable(page) ||
6220 !get_page_unless_zero(page)) {
6224 ClearHPageMigratable(page);
6225 list_move_tail(&page->lru, list);
6227 spin_unlock_irq(&hugetlb_lock);
6231 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6236 spin_lock_irq(&hugetlb_lock);
6237 if (PageHeadHuge(page)) {
6239 if (HPageFreed(page) || HPageMigratable(page))
6240 ret = get_page_unless_zero(page);
6244 spin_unlock_irq(&hugetlb_lock);
6248 void putback_active_hugepage(struct page *page)
6250 spin_lock_irq(&hugetlb_lock);
6251 SetHPageMigratable(page);
6252 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
6253 spin_unlock_irq(&hugetlb_lock);
6257 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6259 struct hstate *h = page_hstate(oldpage);
6261 hugetlb_cgroup_migrate(oldpage, newpage);
6262 set_page_owner_migrate_reason(newpage, reason);
6265 * transfer temporary state of the new huge page. This is
6266 * reverse to other transitions because the newpage is going to
6267 * be final while the old one will be freed so it takes over
6268 * the temporary status.
6270 * Also note that we have to transfer the per-node surplus state
6271 * here as well otherwise the global surplus count will not match
6274 if (HPageTemporary(newpage)) {
6275 int old_nid = page_to_nid(oldpage);
6276 int new_nid = page_to_nid(newpage);
6278 SetHPageTemporary(oldpage);
6279 ClearHPageTemporary(newpage);
6282 * There is no need to transfer the per-node surplus state
6283 * when we do not cross the node.
6285 if (new_nid == old_nid)
6287 spin_lock_irq(&hugetlb_lock);
6288 if (h->surplus_huge_pages_node[old_nid]) {
6289 h->surplus_huge_pages_node[old_nid]--;
6290 h->surplus_huge_pages_node[new_nid]++;
6292 spin_unlock_irq(&hugetlb_lock);
6297 * This function will unconditionally remove all the shared pmd pgtable entries
6298 * within the specific vma for a hugetlbfs memory range.
6300 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6302 struct hstate *h = hstate_vma(vma);
6303 unsigned long sz = huge_page_size(h);
6304 struct mm_struct *mm = vma->vm_mm;
6305 struct mmu_notifier_range range;
6306 unsigned long address, start, end;
6310 if (!(vma->vm_flags & VM_MAYSHARE))
6313 start = ALIGN(vma->vm_start, PUD_SIZE);
6314 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6320 * No need to call adjust_range_if_pmd_sharing_possible(), because
6321 * we have already done the PUD_SIZE alignment.
6323 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6325 mmu_notifier_invalidate_range_start(&range);
6326 i_mmap_lock_write(vma->vm_file->f_mapping);
6327 for (address = start; address < end; address += PUD_SIZE) {
6328 unsigned long tmp = address;
6330 ptep = huge_pte_offset(mm, address, sz);
6333 ptl = huge_pte_lock(h, mm, ptep);
6334 /* We don't want 'address' to be changed */
6335 huge_pmd_unshare(mm, vma, &tmp, ptep);
6338 flush_hugetlb_tlb_range(vma, start, end);
6339 i_mmap_unlock_write(vma->vm_file->f_mapping);
6341 * No need to call mmu_notifier_invalidate_range(), see
6342 * Documentation/vm/mmu_notifier.rst.
6344 mmu_notifier_invalidate_range_end(&range);
6348 static bool cma_reserve_called __initdata;
6350 static int __init cmdline_parse_hugetlb_cma(char *p)
6352 hugetlb_cma_size = memparse(p, &p);
6356 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6358 void __init hugetlb_cma_reserve(int order)
6360 unsigned long size, reserved, per_node;
6363 cma_reserve_called = true;
6365 if (!hugetlb_cma_size)
6368 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6369 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6370 (PAGE_SIZE << order) / SZ_1M);
6375 * If 3 GB area is requested on a machine with 4 numa nodes,
6376 * let's allocate 1 GB on first three nodes and ignore the last one.
6378 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6379 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6380 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6383 for_each_node_state(nid, N_ONLINE) {
6385 char name[CMA_MAX_NAME];
6387 size = min(per_node, hugetlb_cma_size - reserved);
6388 size = round_up(size, PAGE_SIZE << order);
6390 snprintf(name, sizeof(name), "hugetlb%d", nid);
6391 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
6393 &hugetlb_cma[nid], nid);
6395 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
6401 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
6404 if (reserved >= hugetlb_cma_size)
6409 void __init hugetlb_cma_check(void)
6411 if (!hugetlb_cma_size || cma_reserve_called)
6414 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
6417 #endif /* CONFIG_CMA */