1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
36 #include <asm/pgalloc.h>
40 #include <linux/hugetlb.h>
41 #include <linux/hugetlb_cgroup.h>
42 #include <linux/node.h>
43 #include <linux/page_owner.h>
45 #include "hugetlb_vmemmap.h"
47 int hugetlb_max_hstate __read_mostly;
48 unsigned int default_hstate_idx;
49 struct hstate hstates[HUGE_MAX_HSTATE];
52 static struct cma *hugetlb_cma[MAX_NUMNODES];
54 static unsigned long hugetlb_cma_size __initdata;
57 * Minimum page order among possible hugepage sizes, set to a proper value
60 static unsigned int minimum_order __read_mostly = UINT_MAX;
62 __initdata LIST_HEAD(huge_boot_pages);
64 /* for command line parsing */
65 static struct hstate * __initdata parsed_hstate;
66 static unsigned long __initdata default_hstate_max_huge_pages;
67 static bool __initdata parsed_valid_hugepagesz = true;
68 static bool __initdata parsed_default_hugepagesz;
71 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
72 * free_huge_pages, and surplus_huge_pages.
74 DEFINE_SPINLOCK(hugetlb_lock);
77 * Serializes faults on the same logical page. This is used to
78 * prevent spurious OOMs when the hugepage pool is fully utilized.
80 static int num_fault_mutexes;
81 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
83 /* Forward declaration */
84 static int hugetlb_acct_memory(struct hstate *h, long delta);
86 static inline bool subpool_is_free(struct hugepage_subpool *spool)
90 if (spool->max_hpages != -1)
91 return spool->used_hpages == 0;
92 if (spool->min_hpages != -1)
93 return spool->rsv_hpages == spool->min_hpages;
98 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
99 unsigned long irq_flags)
101 spin_unlock_irqrestore(&spool->lock, irq_flags);
103 /* If no pages are used, and no other handles to the subpool
104 * remain, give up any reservations based on minimum size and
105 * free the subpool */
106 if (subpool_is_free(spool)) {
107 if (spool->min_hpages != -1)
108 hugetlb_acct_memory(spool->hstate,
114 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
117 struct hugepage_subpool *spool;
119 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
123 spin_lock_init(&spool->lock);
125 spool->max_hpages = max_hpages;
127 spool->min_hpages = min_hpages;
129 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
133 spool->rsv_hpages = min_hpages;
138 void hugepage_put_subpool(struct hugepage_subpool *spool)
142 spin_lock_irqsave(&spool->lock, flags);
143 BUG_ON(!spool->count);
145 unlock_or_release_subpool(spool, flags);
149 * Subpool accounting for allocating and reserving pages.
150 * Return -ENOMEM if there are not enough resources to satisfy the
151 * request. Otherwise, return the number of pages by which the
152 * global pools must be adjusted (upward). The returned value may
153 * only be different than the passed value (delta) in the case where
154 * a subpool minimum size must be maintained.
156 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
164 spin_lock_irq(&spool->lock);
166 if (spool->max_hpages != -1) { /* maximum size accounting */
167 if ((spool->used_hpages + delta) <= spool->max_hpages)
168 spool->used_hpages += delta;
175 /* minimum size accounting */
176 if (spool->min_hpages != -1 && spool->rsv_hpages) {
177 if (delta > spool->rsv_hpages) {
179 * Asking for more reserves than those already taken on
180 * behalf of subpool. Return difference.
182 ret = delta - spool->rsv_hpages;
183 spool->rsv_hpages = 0;
185 ret = 0; /* reserves already accounted for */
186 spool->rsv_hpages -= delta;
191 spin_unlock_irq(&spool->lock);
196 * Subpool accounting for freeing and unreserving pages.
197 * Return the number of global page reservations that must be dropped.
198 * The return value may only be different than the passed value (delta)
199 * in the case where a subpool minimum size must be maintained.
201 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
210 spin_lock_irqsave(&spool->lock, flags);
212 if (spool->max_hpages != -1) /* maximum size accounting */
213 spool->used_hpages -= delta;
215 /* minimum size accounting */
216 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
217 if (spool->rsv_hpages + delta <= spool->min_hpages)
220 ret = spool->rsv_hpages + delta - spool->min_hpages;
222 spool->rsv_hpages += delta;
223 if (spool->rsv_hpages > spool->min_hpages)
224 spool->rsv_hpages = spool->min_hpages;
228 * If hugetlbfs_put_super couldn't free spool due to an outstanding
229 * quota reference, free it now.
231 unlock_or_release_subpool(spool, flags);
236 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
238 return HUGETLBFS_SB(inode->i_sb)->spool;
241 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
243 return subpool_inode(file_inode(vma->vm_file));
246 /* Helper that removes a struct file_region from the resv_map cache and returns
249 static struct file_region *
250 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
252 struct file_region *nrg = NULL;
254 VM_BUG_ON(resv->region_cache_count <= 0);
256 resv->region_cache_count--;
257 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
258 list_del(&nrg->link);
266 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
267 struct file_region *rg)
269 #ifdef CONFIG_CGROUP_HUGETLB
270 nrg->reservation_counter = rg->reservation_counter;
277 /* Helper that records hugetlb_cgroup uncharge info. */
278 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
280 struct resv_map *resv,
281 struct file_region *nrg)
283 #ifdef CONFIG_CGROUP_HUGETLB
285 nrg->reservation_counter =
286 &h_cg->rsvd_hugepage[hstate_index(h)];
287 nrg->css = &h_cg->css;
289 * The caller will hold exactly one h_cg->css reference for the
290 * whole contiguous reservation region. But this area might be
291 * scattered when there are already some file_regions reside in
292 * it. As a result, many file_regions may share only one css
293 * reference. In order to ensure that one file_region must hold
294 * exactly one h_cg->css reference, we should do css_get for
295 * each file_region and leave the reference held by caller
299 if (!resv->pages_per_hpage)
300 resv->pages_per_hpage = pages_per_huge_page(h);
301 /* pages_per_hpage should be the same for all entries in
304 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
306 nrg->reservation_counter = NULL;
312 static void put_uncharge_info(struct file_region *rg)
314 #ifdef CONFIG_CGROUP_HUGETLB
320 static bool has_same_uncharge_info(struct file_region *rg,
321 struct file_region *org)
323 #ifdef CONFIG_CGROUP_HUGETLB
325 rg->reservation_counter == org->reservation_counter &&
333 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
335 struct file_region *nrg = NULL, *prg = NULL;
337 prg = list_prev_entry(rg, link);
338 if (&prg->link != &resv->regions && prg->to == rg->from &&
339 has_same_uncharge_info(prg, rg)) {
343 put_uncharge_info(rg);
349 nrg = list_next_entry(rg, link);
350 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
351 has_same_uncharge_info(nrg, rg)) {
352 nrg->from = rg->from;
355 put_uncharge_info(rg);
361 hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
362 long to, struct hstate *h, struct hugetlb_cgroup *cg,
363 long *regions_needed)
365 struct file_region *nrg;
367 if (!regions_needed) {
368 nrg = get_file_region_entry_from_cache(map, from, to);
369 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
370 list_add(&nrg->link, rg->link.prev);
371 coalesce_file_region(map, nrg);
373 *regions_needed += 1;
379 * Must be called with resv->lock held.
381 * Calling this with regions_needed != NULL will count the number of pages
382 * to be added but will not modify the linked list. And regions_needed will
383 * indicate the number of file_regions needed in the cache to carry out to add
384 * the regions for this range.
386 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
387 struct hugetlb_cgroup *h_cg,
388 struct hstate *h, long *regions_needed)
391 struct list_head *head = &resv->regions;
392 long last_accounted_offset = f;
393 struct file_region *rg = NULL, *trg = NULL;
398 /* In this loop, we essentially handle an entry for the range
399 * [last_accounted_offset, rg->from), at every iteration, with some
402 list_for_each_entry_safe(rg, trg, head, link) {
403 /* Skip irrelevant regions that start before our range. */
405 /* If this region ends after the last accounted offset,
406 * then we need to update last_accounted_offset.
408 if (rg->to > last_accounted_offset)
409 last_accounted_offset = rg->to;
413 /* When we find a region that starts beyond our range, we've
419 /* Add an entry for last_accounted_offset -> rg->from, and
420 * update last_accounted_offset.
422 if (rg->from > last_accounted_offset)
423 add += hugetlb_resv_map_add(resv, rg,
424 last_accounted_offset,
428 last_accounted_offset = rg->to;
431 /* Handle the case where our range extends beyond
432 * last_accounted_offset.
434 if (last_accounted_offset < t)
435 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
436 t, h, h_cg, regions_needed);
442 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
444 static int allocate_file_region_entries(struct resv_map *resv,
446 __must_hold(&resv->lock)
448 struct list_head allocated_regions;
449 int to_allocate = 0, i = 0;
450 struct file_region *trg = NULL, *rg = NULL;
452 VM_BUG_ON(regions_needed < 0);
454 INIT_LIST_HEAD(&allocated_regions);
457 * Check for sufficient descriptors in the cache to accommodate
458 * the number of in progress add operations plus regions_needed.
460 * This is a while loop because when we drop the lock, some other call
461 * to region_add or region_del may have consumed some region_entries,
462 * so we keep looping here until we finally have enough entries for
463 * (adds_in_progress + regions_needed).
465 while (resv->region_cache_count <
466 (resv->adds_in_progress + regions_needed)) {
467 to_allocate = resv->adds_in_progress + regions_needed -
468 resv->region_cache_count;
470 /* At this point, we should have enough entries in the cache
471 * for all the existing adds_in_progress. We should only be
472 * needing to allocate for regions_needed.
474 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
476 spin_unlock(&resv->lock);
477 for (i = 0; i < to_allocate; i++) {
478 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
481 list_add(&trg->link, &allocated_regions);
484 spin_lock(&resv->lock);
486 list_splice(&allocated_regions, &resv->region_cache);
487 resv->region_cache_count += to_allocate;
493 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
501 * Add the huge page range represented by [f, t) to the reserve
502 * map. Regions will be taken from the cache to fill in this range.
503 * Sufficient regions should exist in the cache due to the previous
504 * call to region_chg with the same range, but in some cases the cache will not
505 * have sufficient entries due to races with other code doing region_add or
506 * region_del. The extra needed entries will be allocated.
508 * regions_needed is the out value provided by a previous call to region_chg.
510 * Return the number of new huge pages added to the map. This number is greater
511 * than or equal to zero. If file_region entries needed to be allocated for
512 * this operation and we were not able to allocate, it returns -ENOMEM.
513 * region_add of regions of length 1 never allocate file_regions and cannot
514 * fail; region_chg will always allocate at least 1 entry and a region_add for
515 * 1 page will only require at most 1 entry.
517 static long region_add(struct resv_map *resv, long f, long t,
518 long in_regions_needed, struct hstate *h,
519 struct hugetlb_cgroup *h_cg)
521 long add = 0, actual_regions_needed = 0;
523 spin_lock(&resv->lock);
526 /* Count how many regions are actually needed to execute this add. */
527 add_reservation_in_range(resv, f, t, NULL, NULL,
528 &actual_regions_needed);
531 * Check for sufficient descriptors in the cache to accommodate
532 * this add operation. Note that actual_regions_needed may be greater
533 * than in_regions_needed, as the resv_map may have been modified since
534 * the region_chg call. In this case, we need to make sure that we
535 * allocate extra entries, such that we have enough for all the
536 * existing adds_in_progress, plus the excess needed for this
539 if (actual_regions_needed > in_regions_needed &&
540 resv->region_cache_count <
541 resv->adds_in_progress +
542 (actual_regions_needed - in_regions_needed)) {
543 /* region_add operation of range 1 should never need to
544 * allocate file_region entries.
546 VM_BUG_ON(t - f <= 1);
548 if (allocate_file_region_entries(
549 resv, actual_regions_needed - in_regions_needed)) {
556 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
558 resv->adds_in_progress -= in_regions_needed;
560 spin_unlock(&resv->lock);
565 * Examine the existing reserve map and determine how many
566 * huge pages in the specified range [f, t) are NOT currently
567 * represented. This routine is called before a subsequent
568 * call to region_add that will actually modify the reserve
569 * map to add the specified range [f, t). region_chg does
570 * not change the number of huge pages represented by the
571 * map. A number of new file_region structures is added to the cache as a
572 * placeholder, for the subsequent region_add call to use. At least 1
573 * file_region structure is added.
575 * out_regions_needed is the number of regions added to the
576 * resv->adds_in_progress. This value needs to be provided to a follow up call
577 * to region_add or region_abort for proper accounting.
579 * Returns the number of huge pages that need to be added to the existing
580 * reservation map for the range [f, t). This number is greater or equal to
581 * zero. -ENOMEM is returned if a new file_region structure or cache entry
582 * is needed and can not be allocated.
584 static long region_chg(struct resv_map *resv, long f, long t,
585 long *out_regions_needed)
589 spin_lock(&resv->lock);
591 /* Count how many hugepages in this range are NOT represented. */
592 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
595 if (*out_regions_needed == 0)
596 *out_regions_needed = 1;
598 if (allocate_file_region_entries(resv, *out_regions_needed))
601 resv->adds_in_progress += *out_regions_needed;
603 spin_unlock(&resv->lock);
608 * Abort the in progress add operation. The adds_in_progress field
609 * of the resv_map keeps track of the operations in progress between
610 * calls to region_chg and region_add. Operations are sometimes
611 * aborted after the call to region_chg. In such cases, region_abort
612 * is called to decrement the adds_in_progress counter. regions_needed
613 * is the value returned by the region_chg call, it is used to decrement
614 * the adds_in_progress counter.
616 * NOTE: The range arguments [f, t) are not needed or used in this
617 * routine. They are kept to make reading the calling code easier as
618 * arguments will match the associated region_chg call.
620 static void region_abort(struct resv_map *resv, long f, long t,
623 spin_lock(&resv->lock);
624 VM_BUG_ON(!resv->region_cache_count);
625 resv->adds_in_progress -= regions_needed;
626 spin_unlock(&resv->lock);
630 * Delete the specified range [f, t) from the reserve map. If the
631 * t parameter is LONG_MAX, this indicates that ALL regions after f
632 * should be deleted. Locate the regions which intersect [f, t)
633 * and either trim, delete or split the existing regions.
635 * Returns the number of huge pages deleted from the reserve map.
636 * In the normal case, the return value is zero or more. In the
637 * case where a region must be split, a new region descriptor must
638 * be allocated. If the allocation fails, -ENOMEM will be returned.
639 * NOTE: If the parameter t == LONG_MAX, then we will never split
640 * a region and possibly return -ENOMEM. Callers specifying
641 * t == LONG_MAX do not need to check for -ENOMEM error.
643 static long region_del(struct resv_map *resv, long f, long t)
645 struct list_head *head = &resv->regions;
646 struct file_region *rg, *trg;
647 struct file_region *nrg = NULL;
651 spin_lock(&resv->lock);
652 list_for_each_entry_safe(rg, trg, head, link) {
654 * Skip regions before the range to be deleted. file_region
655 * ranges are normally of the form [from, to). However, there
656 * may be a "placeholder" entry in the map which is of the form
657 * (from, to) with from == to. Check for placeholder entries
658 * at the beginning of the range to be deleted.
660 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
666 if (f > rg->from && t < rg->to) { /* Must split region */
668 * Check for an entry in the cache before dropping
669 * lock and attempting allocation.
672 resv->region_cache_count > resv->adds_in_progress) {
673 nrg = list_first_entry(&resv->region_cache,
676 list_del(&nrg->link);
677 resv->region_cache_count--;
681 spin_unlock(&resv->lock);
682 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
689 hugetlb_cgroup_uncharge_file_region(
690 resv, rg, t - f, false);
692 /* New entry for end of split region */
696 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
698 INIT_LIST_HEAD(&nrg->link);
700 /* Original entry is trimmed */
703 list_add(&nrg->link, &rg->link);
708 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
709 del += rg->to - rg->from;
710 hugetlb_cgroup_uncharge_file_region(resv, rg,
711 rg->to - rg->from, true);
717 if (f <= rg->from) { /* Trim beginning of region */
718 hugetlb_cgroup_uncharge_file_region(resv, rg,
719 t - rg->from, false);
723 } else { /* Trim end of region */
724 hugetlb_cgroup_uncharge_file_region(resv, rg,
732 spin_unlock(&resv->lock);
738 * A rare out of memory error was encountered which prevented removal of
739 * the reserve map region for a page. The huge page itself was free'ed
740 * and removed from the page cache. This routine will adjust the subpool
741 * usage count, and the global reserve count if needed. By incrementing
742 * these counts, the reserve map entry which could not be deleted will
743 * appear as a "reserved" entry instead of simply dangling with incorrect
746 void hugetlb_fix_reserve_counts(struct inode *inode)
748 struct hugepage_subpool *spool = subpool_inode(inode);
750 bool reserved = false;
752 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
753 if (rsv_adjust > 0) {
754 struct hstate *h = hstate_inode(inode);
756 if (!hugetlb_acct_memory(h, 1))
758 } else if (!rsv_adjust) {
763 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
767 * Count and return the number of huge pages in the reserve map
768 * that intersect with the range [f, t).
770 static long region_count(struct resv_map *resv, long f, long t)
772 struct list_head *head = &resv->regions;
773 struct file_region *rg;
776 spin_lock(&resv->lock);
777 /* Locate each segment we overlap with, and count that overlap. */
778 list_for_each_entry(rg, head, link) {
787 seg_from = max(rg->from, f);
788 seg_to = min(rg->to, t);
790 chg += seg_to - seg_from;
792 spin_unlock(&resv->lock);
798 * Convert the address within this vma to the page offset within
799 * the mapping, in pagecache page units; huge pages here.
801 static pgoff_t vma_hugecache_offset(struct hstate *h,
802 struct vm_area_struct *vma, unsigned long address)
804 return ((address - vma->vm_start) >> huge_page_shift(h)) +
805 (vma->vm_pgoff >> huge_page_order(h));
808 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
809 unsigned long address)
811 return vma_hugecache_offset(hstate_vma(vma), vma, address);
813 EXPORT_SYMBOL_GPL(linear_hugepage_index);
816 * Return the size of the pages allocated when backing a VMA. In the majority
817 * cases this will be same size as used by the page table entries.
819 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
821 if (vma->vm_ops && vma->vm_ops->pagesize)
822 return vma->vm_ops->pagesize(vma);
825 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
828 * Return the page size being used by the MMU to back a VMA. In the majority
829 * of cases, the page size used by the kernel matches the MMU size. On
830 * architectures where it differs, an architecture-specific 'strong'
831 * version of this symbol is required.
833 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
835 return vma_kernel_pagesize(vma);
839 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
840 * bits of the reservation map pointer, which are always clear due to
843 #define HPAGE_RESV_OWNER (1UL << 0)
844 #define HPAGE_RESV_UNMAPPED (1UL << 1)
845 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
848 * These helpers are used to track how many pages are reserved for
849 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
850 * is guaranteed to have their future faults succeed.
852 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
853 * the reserve counters are updated with the hugetlb_lock held. It is safe
854 * to reset the VMA at fork() time as it is not in use yet and there is no
855 * chance of the global counters getting corrupted as a result of the values.
857 * The private mapping reservation is represented in a subtly different
858 * manner to a shared mapping. A shared mapping has a region map associated
859 * with the underlying file, this region map represents the backing file
860 * pages which have ever had a reservation assigned which this persists even
861 * after the page is instantiated. A private mapping has a region map
862 * associated with the original mmap which is attached to all VMAs which
863 * reference it, this region map represents those offsets which have consumed
864 * reservation ie. where pages have been instantiated.
866 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
868 return (unsigned long)vma->vm_private_data;
871 static void set_vma_private_data(struct vm_area_struct *vma,
874 vma->vm_private_data = (void *)value;
878 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
879 struct hugetlb_cgroup *h_cg,
882 #ifdef CONFIG_CGROUP_HUGETLB
884 resv_map->reservation_counter = NULL;
885 resv_map->pages_per_hpage = 0;
886 resv_map->css = NULL;
888 resv_map->reservation_counter =
889 &h_cg->rsvd_hugepage[hstate_index(h)];
890 resv_map->pages_per_hpage = pages_per_huge_page(h);
891 resv_map->css = &h_cg->css;
896 struct resv_map *resv_map_alloc(void)
898 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
899 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
901 if (!resv_map || !rg) {
907 kref_init(&resv_map->refs);
908 spin_lock_init(&resv_map->lock);
909 INIT_LIST_HEAD(&resv_map->regions);
911 resv_map->adds_in_progress = 0;
913 * Initialize these to 0. On shared mappings, 0's here indicate these
914 * fields don't do cgroup accounting. On private mappings, these will be
915 * re-initialized to the proper values, to indicate that hugetlb cgroup
916 * reservations are to be un-charged from here.
918 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
920 INIT_LIST_HEAD(&resv_map->region_cache);
921 list_add(&rg->link, &resv_map->region_cache);
922 resv_map->region_cache_count = 1;
927 void resv_map_release(struct kref *ref)
929 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
930 struct list_head *head = &resv_map->region_cache;
931 struct file_region *rg, *trg;
933 /* Clear out any active regions before we release the map. */
934 region_del(resv_map, 0, LONG_MAX);
936 /* ... and any entries left in the cache */
937 list_for_each_entry_safe(rg, trg, head, link) {
942 VM_BUG_ON(resv_map->adds_in_progress);
947 static inline struct resv_map *inode_resv_map(struct inode *inode)
950 * At inode evict time, i_mapping may not point to the original
951 * address space within the inode. This original address space
952 * contains the pointer to the resv_map. So, always use the
953 * address space embedded within the inode.
954 * The VERY common case is inode->mapping == &inode->i_data but,
955 * this may not be true for device special inodes.
957 return (struct resv_map *)(&inode->i_data)->private_data;
960 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
962 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
963 if (vma->vm_flags & VM_MAYSHARE) {
964 struct address_space *mapping = vma->vm_file->f_mapping;
965 struct inode *inode = mapping->host;
967 return inode_resv_map(inode);
970 return (struct resv_map *)(get_vma_private_data(vma) &
975 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
977 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
978 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
980 set_vma_private_data(vma, (get_vma_private_data(vma) &
981 HPAGE_RESV_MASK) | (unsigned long)map);
984 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
986 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
987 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
989 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
992 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
994 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
996 return (get_vma_private_data(vma) & flag) != 0;
999 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
1000 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1002 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1003 if (!(vma->vm_flags & VM_MAYSHARE))
1004 vma->vm_private_data = (void *)0;
1007 /* Returns true if the VMA has associated reserve pages */
1008 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1010 if (vma->vm_flags & VM_NORESERVE) {
1012 * This address is already reserved by other process(chg == 0),
1013 * so, we should decrement reserved count. Without decrementing,
1014 * reserve count remains after releasing inode, because this
1015 * allocated page will go into page cache and is regarded as
1016 * coming from reserved pool in releasing step. Currently, we
1017 * don't have any other solution to deal with this situation
1018 * properly, so add work-around here.
1020 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1026 /* Shared mappings always use reserves */
1027 if (vma->vm_flags & VM_MAYSHARE) {
1029 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1030 * be a region map for all pages. The only situation where
1031 * there is no region map is if a hole was punched via
1032 * fallocate. In this case, there really are no reserves to
1033 * use. This situation is indicated if chg != 0.
1042 * Only the process that called mmap() has reserves for
1045 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1047 * Like the shared case above, a hole punch or truncate
1048 * could have been performed on the private mapping.
1049 * Examine the value of chg to determine if reserves
1050 * actually exist or were previously consumed.
1051 * Very Subtle - The value of chg comes from a previous
1052 * call to vma_needs_reserves(). The reserve map for
1053 * private mappings has different (opposite) semantics
1054 * than that of shared mappings. vma_needs_reserves()
1055 * has already taken this difference in semantics into
1056 * account. Therefore, the meaning of chg is the same
1057 * as in the shared case above. Code could easily be
1058 * combined, but keeping it separate draws attention to
1059 * subtle differences.
1070 static void enqueue_huge_page(struct hstate *h, struct page *page)
1072 int nid = page_to_nid(page);
1074 lockdep_assert_held(&hugetlb_lock);
1075 list_move(&page->lru, &h->hugepage_freelists[nid]);
1076 h->free_huge_pages++;
1077 h->free_huge_pages_node[nid]++;
1078 SetHPageFreed(page);
1081 static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
1084 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1086 lockdep_assert_held(&hugetlb_lock);
1087 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1088 if (pin && !is_pinnable_page(page))
1091 if (PageHWPoison(page))
1094 list_move(&page->lru, &h->hugepage_activelist);
1095 set_page_refcounted(page);
1096 ClearHPageFreed(page);
1097 h->free_huge_pages--;
1098 h->free_huge_pages_node[nid]--;
1105 static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1108 unsigned int cpuset_mems_cookie;
1109 struct zonelist *zonelist;
1112 int node = NUMA_NO_NODE;
1114 zonelist = node_zonelist(nid, gfp_mask);
1117 cpuset_mems_cookie = read_mems_allowed_begin();
1118 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1121 if (!cpuset_zone_allowed(zone, gfp_mask))
1124 * no need to ask again on the same node. Pool is node rather than
1127 if (zone_to_nid(zone) == node)
1129 node = zone_to_nid(zone);
1131 page = dequeue_huge_page_node_exact(h, node);
1135 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1141 static struct page *dequeue_huge_page_vma(struct hstate *h,
1142 struct vm_area_struct *vma,
1143 unsigned long address, int avoid_reserve,
1147 struct mempolicy *mpol;
1149 nodemask_t *nodemask;
1153 * A child process with MAP_PRIVATE mappings created by their parent
1154 * have no page reserves. This check ensures that reservations are
1155 * not "stolen". The child may still get SIGKILLed
1157 if (!vma_has_reserves(vma, chg) &&
1158 h->free_huge_pages - h->resv_huge_pages == 0)
1161 /* If reserves cannot be used, ensure enough pages are in the pool */
1162 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
1165 gfp_mask = htlb_alloc_mask(h);
1166 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1167 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1168 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1169 SetHPageRestoreReserve(page);
1170 h->resv_huge_pages--;
1173 mpol_cond_put(mpol);
1181 * common helper functions for hstate_next_node_to_{alloc|free}.
1182 * We may have allocated or freed a huge page based on a different
1183 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1184 * be outside of *nodes_allowed. Ensure that we use an allowed
1185 * node for alloc or free.
1187 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1189 nid = next_node_in(nid, *nodes_allowed);
1190 VM_BUG_ON(nid >= MAX_NUMNODES);
1195 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1197 if (!node_isset(nid, *nodes_allowed))
1198 nid = next_node_allowed(nid, nodes_allowed);
1203 * returns the previously saved node ["this node"] from which to
1204 * allocate a persistent huge page for the pool and advance the
1205 * next node from which to allocate, handling wrap at end of node
1208 static int hstate_next_node_to_alloc(struct hstate *h,
1209 nodemask_t *nodes_allowed)
1213 VM_BUG_ON(!nodes_allowed);
1215 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1216 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1222 * helper for remove_pool_huge_page() - return the previously saved
1223 * node ["this node"] from which to free a huge page. Advance the
1224 * next node id whether or not we find a free huge page to free so
1225 * that the next attempt to free addresses the next node.
1227 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1231 VM_BUG_ON(!nodes_allowed);
1233 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1234 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1239 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1240 for (nr_nodes = nodes_weight(*mask); \
1242 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1245 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1246 for (nr_nodes = nodes_weight(*mask); \
1248 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1251 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1252 static void destroy_compound_gigantic_page(struct page *page,
1256 int nr_pages = 1 << order;
1257 struct page *p = page + 1;
1259 atomic_set(compound_mapcount_ptr(page), 0);
1260 atomic_set(compound_pincount_ptr(page), 0);
1262 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1263 clear_compound_head(p);
1264 set_page_refcounted(p);
1267 set_compound_order(page, 0);
1268 page[1].compound_nr = 0;
1269 __ClearPageHead(page);
1272 static void free_gigantic_page(struct page *page, unsigned int order)
1275 * If the page isn't allocated using the cma allocator,
1276 * cma_release() returns false.
1279 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
1283 free_contig_range(page_to_pfn(page), 1 << order);
1286 #ifdef CONFIG_CONTIG_ALLOC
1287 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1288 int nid, nodemask_t *nodemask)
1290 unsigned long nr_pages = pages_per_huge_page(h);
1291 if (nid == NUMA_NO_NODE)
1292 nid = numa_mem_id();
1299 if (hugetlb_cma[nid]) {
1300 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1301 huge_page_order(h), true);
1306 if (!(gfp_mask & __GFP_THISNODE)) {
1307 for_each_node_mask(node, *nodemask) {
1308 if (node == nid || !hugetlb_cma[node])
1311 page = cma_alloc(hugetlb_cma[node], nr_pages,
1312 huge_page_order(h), true);
1320 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1323 #else /* !CONFIG_CONTIG_ALLOC */
1324 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1325 int nid, nodemask_t *nodemask)
1329 #endif /* CONFIG_CONTIG_ALLOC */
1331 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1332 static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1333 int nid, nodemask_t *nodemask)
1337 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1338 static inline void destroy_compound_gigantic_page(struct page *page,
1339 unsigned int order) { }
1343 * Remove hugetlb page from lists, and update dtor so that page appears
1344 * as just a compound page. A reference is held on the page.
1346 * Must be called with hugetlb lock held.
1348 static void remove_hugetlb_page(struct hstate *h, struct page *page,
1349 bool adjust_surplus)
1351 int nid = page_to_nid(page);
1353 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1354 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1356 lockdep_assert_held(&hugetlb_lock);
1357 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1360 list_del(&page->lru);
1362 if (HPageFreed(page)) {
1363 h->free_huge_pages--;
1364 h->free_huge_pages_node[nid]--;
1366 if (adjust_surplus) {
1367 h->surplus_huge_pages--;
1368 h->surplus_huge_pages_node[nid]--;
1371 set_page_refcounted(page);
1372 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1375 h->nr_huge_pages_node[nid]--;
1378 static void add_hugetlb_page(struct hstate *h, struct page *page,
1379 bool adjust_surplus)
1382 int nid = page_to_nid(page);
1384 VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page);
1386 lockdep_assert_held(&hugetlb_lock);
1388 INIT_LIST_HEAD(&page->lru);
1390 h->nr_huge_pages_node[nid]++;
1392 if (adjust_surplus) {
1393 h->surplus_huge_pages++;
1394 h->surplus_huge_pages_node[nid]++;
1397 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1398 set_page_private(page, 0);
1399 SetHPageVmemmapOptimized(page);
1402 * This page is now managed by the hugetlb allocator and has
1403 * no users -- drop the last reference.
1405 zeroed = put_page_testzero(page);
1406 VM_BUG_ON_PAGE(!zeroed, page);
1407 arch_clear_hugepage_flags(page);
1408 enqueue_huge_page(h, page);
1411 static void __update_and_free_page(struct hstate *h, struct page *page)
1414 struct page *subpage = page;
1416 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1419 if (alloc_huge_page_vmemmap(h, page)) {
1420 spin_lock_irq(&hugetlb_lock);
1422 * If we cannot allocate vmemmap pages, just refuse to free the
1423 * page and put the page back on the hugetlb free list and treat
1424 * as a surplus page.
1426 add_hugetlb_page(h, page, true);
1427 spin_unlock_irq(&hugetlb_lock);
1431 for (i = 0; i < pages_per_huge_page(h);
1432 i++, subpage = mem_map_next(subpage, page, i)) {
1433 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
1434 1 << PG_referenced | 1 << PG_dirty |
1435 1 << PG_active | 1 << PG_private |
1438 if (hstate_is_gigantic(h)) {
1439 destroy_compound_gigantic_page(page, huge_page_order(h));
1440 free_gigantic_page(page, huge_page_order(h));
1442 __free_pages(page, huge_page_order(h));
1447 * As update_and_free_page() can be called under any context, so we cannot
1448 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1449 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1450 * the vmemmap pages.
1452 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1453 * freed and frees them one-by-one. As the page->mapping pointer is going
1454 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1455 * structure of a lockless linked list of huge pages to be freed.
1457 static LLIST_HEAD(hpage_freelist);
1459 static void free_hpage_workfn(struct work_struct *work)
1461 struct llist_node *node;
1463 node = llist_del_all(&hpage_freelist);
1469 page = container_of((struct address_space **)node,
1470 struct page, mapping);
1472 page->mapping = NULL;
1474 * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate()
1475 * is going to trigger because a previous call to
1476 * remove_hugetlb_page() will set_compound_page_dtor(page,
1477 * NULL_COMPOUND_DTOR), so do not use page_hstate() directly.
1479 h = size_to_hstate(page_size(page));
1481 __update_and_free_page(h, page);
1486 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1488 static inline void flush_free_hpage_work(struct hstate *h)
1490 if (free_vmemmap_pages_per_hpage(h))
1491 flush_work(&free_hpage_work);
1494 static void update_and_free_page(struct hstate *h, struct page *page,
1497 if (!HPageVmemmapOptimized(page) || !atomic) {
1498 __update_and_free_page(h, page);
1503 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1505 * Only call schedule_work() if hpage_freelist is previously
1506 * empty. Otherwise, schedule_work() had been called but the workfn
1507 * hasn't retrieved the list yet.
1509 if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist))
1510 schedule_work(&free_hpage_work);
1513 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1515 struct page *page, *t_page;
1517 list_for_each_entry_safe(page, t_page, list, lru) {
1518 update_and_free_page(h, page, false);
1523 struct hstate *size_to_hstate(unsigned long size)
1527 for_each_hstate(h) {
1528 if (huge_page_size(h) == size)
1534 void free_huge_page(struct page *page)
1537 * Can't pass hstate in here because it is called from the
1538 * compound page destructor.
1540 struct hstate *h = page_hstate(page);
1541 int nid = page_to_nid(page);
1542 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
1543 bool restore_reserve;
1544 unsigned long flags;
1546 VM_BUG_ON_PAGE(page_count(page), page);
1547 VM_BUG_ON_PAGE(page_mapcount(page), page);
1549 hugetlb_set_page_subpool(page, NULL);
1550 page->mapping = NULL;
1551 restore_reserve = HPageRestoreReserve(page);
1552 ClearHPageRestoreReserve(page);
1555 * If HPageRestoreReserve was set on page, page allocation consumed a
1556 * reservation. If the page was associated with a subpool, there
1557 * would have been a page reserved in the subpool before allocation
1558 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1559 * reservation, do not call hugepage_subpool_put_pages() as this will
1560 * remove the reserved page from the subpool.
1562 if (!restore_reserve) {
1564 * A return code of zero implies that the subpool will be
1565 * under its minimum size if the reservation is not restored
1566 * after page is free. Therefore, force restore_reserve
1569 if (hugepage_subpool_put_pages(spool, 1) == 0)
1570 restore_reserve = true;
1573 spin_lock_irqsave(&hugetlb_lock, flags);
1574 ClearHPageMigratable(page);
1575 hugetlb_cgroup_uncharge_page(hstate_index(h),
1576 pages_per_huge_page(h), page);
1577 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1578 pages_per_huge_page(h), page);
1579 if (restore_reserve)
1580 h->resv_huge_pages++;
1582 if (HPageTemporary(page)) {
1583 remove_hugetlb_page(h, page, false);
1584 spin_unlock_irqrestore(&hugetlb_lock, flags);
1585 update_and_free_page(h, page, true);
1586 } else if (h->surplus_huge_pages_node[nid]) {
1587 /* remove the page from active list */
1588 remove_hugetlb_page(h, page, true);
1589 spin_unlock_irqrestore(&hugetlb_lock, flags);
1590 update_and_free_page(h, page, true);
1592 arch_clear_hugepage_flags(page);
1593 enqueue_huge_page(h, page);
1594 spin_unlock_irqrestore(&hugetlb_lock, flags);
1599 * Must be called with the hugetlb lock held
1601 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1603 lockdep_assert_held(&hugetlb_lock);
1605 h->nr_huge_pages_node[nid]++;
1608 static void __prep_new_huge_page(struct hstate *h, struct page *page)
1610 free_huge_page_vmemmap(h, page);
1611 INIT_LIST_HEAD(&page->lru);
1612 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1613 hugetlb_set_page_subpool(page, NULL);
1614 set_hugetlb_cgroup(page, NULL);
1615 set_hugetlb_cgroup_rsvd(page, NULL);
1618 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1620 __prep_new_huge_page(h, page);
1621 spin_lock_irq(&hugetlb_lock);
1622 __prep_account_new_huge_page(h, nid);
1623 spin_unlock_irq(&hugetlb_lock);
1626 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1629 int nr_pages = 1 << order;
1630 struct page *p = page + 1;
1632 /* we rely on prep_new_huge_page to set the destructor */
1633 set_compound_order(page, order);
1634 __ClearPageReserved(page);
1635 __SetPageHead(page);
1636 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1638 * For gigantic hugepages allocated through bootmem at
1639 * boot, it's safer to be consistent with the not-gigantic
1640 * hugepages and clear the PG_reserved bit from all tail pages
1641 * too. Otherwise drivers using get_user_pages() to access tail
1642 * pages may get the reference counting wrong if they see
1643 * PG_reserved set on a tail page (despite the head page not
1644 * having PG_reserved set). Enforcing this consistency between
1645 * head and tail pages allows drivers to optimize away a check
1646 * on the head page when they need know if put_page() is needed
1647 * after get_user_pages().
1649 __ClearPageReserved(p);
1650 set_page_count(p, 0);
1651 set_compound_head(p, page);
1653 atomic_set(compound_mapcount_ptr(page), -1);
1654 atomic_set(compound_pincount_ptr(page), 0);
1658 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1659 * transparent huge pages. See the PageTransHuge() documentation for more
1662 int PageHuge(struct page *page)
1664 if (!PageCompound(page))
1667 page = compound_head(page);
1668 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1670 EXPORT_SYMBOL_GPL(PageHuge);
1673 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1674 * normal or transparent huge pages.
1676 int PageHeadHuge(struct page *page_head)
1678 if (!PageHead(page_head))
1681 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
1685 * Find and lock address space (mapping) in write mode.
1687 * Upon entry, the page is locked which means that page_mapping() is
1688 * stable. Due to locking order, we can only trylock_write. If we can
1689 * not get the lock, simply return NULL to caller.
1691 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1693 struct address_space *mapping = page_mapping(hpage);
1698 if (i_mmap_trylock_write(mapping))
1704 pgoff_t hugetlb_basepage_index(struct page *page)
1706 struct page *page_head = compound_head(page);
1707 pgoff_t index = page_index(page_head);
1708 unsigned long compound_idx;
1710 if (compound_order(page_head) >= MAX_ORDER)
1711 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1713 compound_idx = page - page_head;
1715 return (index << compound_order(page_head)) + compound_idx;
1718 static struct page *alloc_buddy_huge_page(struct hstate *h,
1719 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1720 nodemask_t *node_alloc_noretry)
1722 int order = huge_page_order(h);
1724 bool alloc_try_hard = true;
1727 * By default we always try hard to allocate the page with
1728 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1729 * a loop (to adjust global huge page counts) and previous allocation
1730 * failed, do not continue to try hard on the same node. Use the
1731 * node_alloc_noretry bitmap to manage this state information.
1733 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1734 alloc_try_hard = false;
1735 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1737 gfp_mask |= __GFP_RETRY_MAYFAIL;
1738 if (nid == NUMA_NO_NODE)
1739 nid = numa_mem_id();
1740 page = __alloc_pages(gfp_mask, order, nid, nmask);
1742 __count_vm_event(HTLB_BUDDY_PGALLOC);
1744 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1747 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1748 * indicates an overall state change. Clear bit so that we resume
1749 * normal 'try hard' allocations.
1751 if (node_alloc_noretry && page && !alloc_try_hard)
1752 node_clear(nid, *node_alloc_noretry);
1755 * If we tried hard to get a page but failed, set bit so that
1756 * subsequent attempts will not try as hard until there is an
1757 * overall state change.
1759 if (node_alloc_noretry && !page && alloc_try_hard)
1760 node_set(nid, *node_alloc_noretry);
1766 * Common helper to allocate a fresh hugetlb page. All specific allocators
1767 * should use this function to get new hugetlb pages
1769 static struct page *alloc_fresh_huge_page(struct hstate *h,
1770 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1771 nodemask_t *node_alloc_noretry)
1775 if (hstate_is_gigantic(h))
1776 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1778 page = alloc_buddy_huge_page(h, gfp_mask,
1779 nid, nmask, node_alloc_noretry);
1783 if (hstate_is_gigantic(h))
1784 prep_compound_gigantic_page(page, huge_page_order(h));
1785 prep_new_huge_page(h, page, page_to_nid(page));
1791 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1794 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1795 nodemask_t *node_alloc_noretry)
1799 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1801 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1802 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1803 node_alloc_noretry);
1811 put_page(page); /* free it into the hugepage allocator */
1817 * Remove huge page from pool from next node to free. Attempt to keep
1818 * persistent huge pages more or less balanced over allowed nodes.
1819 * This routine only 'removes' the hugetlb page. The caller must make
1820 * an additional call to free the page to low level allocators.
1821 * Called with hugetlb_lock locked.
1823 static struct page *remove_pool_huge_page(struct hstate *h,
1824 nodemask_t *nodes_allowed,
1828 struct page *page = NULL;
1830 lockdep_assert_held(&hugetlb_lock);
1831 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1833 * If we're returning unused surplus pages, only examine
1834 * nodes with surplus pages.
1836 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1837 !list_empty(&h->hugepage_freelists[node])) {
1838 page = list_entry(h->hugepage_freelists[node].next,
1840 remove_hugetlb_page(h, page, acct_surplus);
1849 * Dissolve a given free hugepage into free buddy pages. This function does
1850 * nothing for in-use hugepages and non-hugepages.
1851 * This function returns values like below:
1853 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
1854 * when the system is under memory pressure and the feature of
1855 * freeing unused vmemmap pages associated with each hugetlb page
1857 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1858 * (allocated or reserved.)
1859 * 0: successfully dissolved free hugepages or the page is not a
1860 * hugepage (considered as already dissolved)
1862 int dissolve_free_huge_page(struct page *page)
1867 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1868 if (!PageHuge(page))
1871 spin_lock_irq(&hugetlb_lock);
1872 if (!PageHuge(page)) {
1877 if (!page_count(page)) {
1878 struct page *head = compound_head(page);
1879 struct hstate *h = page_hstate(head);
1880 if (h->free_huge_pages - h->resv_huge_pages == 0)
1884 * We should make sure that the page is already on the free list
1885 * when it is dissolved.
1887 if (unlikely(!HPageFreed(head))) {
1888 spin_unlock_irq(&hugetlb_lock);
1892 * Theoretically, we should return -EBUSY when we
1893 * encounter this race. In fact, we have a chance
1894 * to successfully dissolve the page if we do a
1895 * retry. Because the race window is quite small.
1896 * If we seize this opportunity, it is an optimization
1897 * for increasing the success rate of dissolving page.
1902 remove_hugetlb_page(h, head, false);
1903 h->max_huge_pages--;
1904 spin_unlock_irq(&hugetlb_lock);
1907 * Normally update_and_free_page will allocate required vmemmmap
1908 * before freeing the page. update_and_free_page will fail to
1909 * free the page if it can not allocate required vmemmap. We
1910 * need to adjust max_huge_pages if the page is not freed.
1911 * Attempt to allocate vmemmmap here so that we can take
1912 * appropriate action on failure.
1914 rc = alloc_huge_page_vmemmap(h, head);
1917 * Move PageHWPoison flag from head page to the raw
1918 * error page, which makes any subpages rather than
1919 * the error page reusable.
1921 if (PageHWPoison(head) && page != head) {
1922 SetPageHWPoison(page);
1923 ClearPageHWPoison(head);
1925 update_and_free_page(h, head, false);
1927 spin_lock_irq(&hugetlb_lock);
1928 add_hugetlb_page(h, head, false);
1929 h->max_huge_pages++;
1930 spin_unlock_irq(&hugetlb_lock);
1936 spin_unlock_irq(&hugetlb_lock);
1941 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1942 * make specified memory blocks removable from the system.
1943 * Note that this will dissolve a free gigantic hugepage completely, if any
1944 * part of it lies within the given range.
1945 * Also note that if dissolve_free_huge_page() returns with an error, all
1946 * free hugepages that were dissolved before that error are lost.
1948 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1954 if (!hugepages_supported())
1957 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1958 page = pfn_to_page(pfn);
1959 rc = dissolve_free_huge_page(page);
1968 * Allocates a fresh surplus page from the page allocator.
1970 static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1971 int nid, nodemask_t *nmask)
1973 struct page *page = NULL;
1975 if (hstate_is_gigantic(h))
1978 spin_lock_irq(&hugetlb_lock);
1979 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1981 spin_unlock_irq(&hugetlb_lock);
1983 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1987 spin_lock_irq(&hugetlb_lock);
1989 * We could have raced with the pool size change.
1990 * Double check that and simply deallocate the new page
1991 * if we would end up overcommiting the surpluses. Abuse
1992 * temporary page to workaround the nasty free_huge_page
1995 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1996 SetHPageTemporary(page);
1997 spin_unlock_irq(&hugetlb_lock);
2001 h->surplus_huge_pages++;
2002 h->surplus_huge_pages_node[page_to_nid(page)]++;
2006 spin_unlock_irq(&hugetlb_lock);
2011 static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
2012 int nid, nodemask_t *nmask)
2016 if (hstate_is_gigantic(h))
2019 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
2024 * We do not account these pages as surplus because they are only
2025 * temporary and will be released properly on the last reference
2027 SetHPageTemporary(page);
2033 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2036 struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
2037 struct vm_area_struct *vma, unsigned long addr)
2040 struct mempolicy *mpol;
2041 gfp_t gfp_mask = htlb_alloc_mask(h);
2043 nodemask_t *nodemask;
2045 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2046 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
2047 mpol_cond_put(mpol);
2052 /* page migration callback function */
2053 struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
2054 nodemask_t *nmask, gfp_t gfp_mask)
2056 spin_lock_irq(&hugetlb_lock);
2057 if (h->free_huge_pages - h->resv_huge_pages > 0) {
2060 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
2062 spin_unlock_irq(&hugetlb_lock);
2066 spin_unlock_irq(&hugetlb_lock);
2068 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
2071 /* mempolicy aware migration callback */
2072 struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2073 unsigned long address)
2075 struct mempolicy *mpol;
2076 nodemask_t *nodemask;
2081 gfp_mask = htlb_alloc_mask(h);
2082 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2083 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
2084 mpol_cond_put(mpol);
2090 * Increase the hugetlb pool such that it can accommodate a reservation
2093 static int gather_surplus_pages(struct hstate *h, long delta)
2094 __must_hold(&hugetlb_lock)
2096 struct list_head surplus_list;
2097 struct page *page, *tmp;
2100 long needed, allocated;
2101 bool alloc_ok = true;
2103 lockdep_assert_held(&hugetlb_lock);
2104 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2106 h->resv_huge_pages += delta;
2111 INIT_LIST_HEAD(&surplus_list);
2115 spin_unlock_irq(&hugetlb_lock);
2116 for (i = 0; i < needed; i++) {
2117 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
2118 NUMA_NO_NODE, NULL);
2123 list_add(&page->lru, &surplus_list);
2129 * After retaking hugetlb_lock, we need to recalculate 'needed'
2130 * because either resv_huge_pages or free_huge_pages may have changed.
2132 spin_lock_irq(&hugetlb_lock);
2133 needed = (h->resv_huge_pages + delta) -
2134 (h->free_huge_pages + allocated);
2139 * We were not able to allocate enough pages to
2140 * satisfy the entire reservation so we free what
2141 * we've allocated so far.
2146 * The surplus_list now contains _at_least_ the number of extra pages
2147 * needed to accommodate the reservation. Add the appropriate number
2148 * of pages to the hugetlb pool and free the extras back to the buddy
2149 * allocator. Commit the entire reservation here to prevent another
2150 * process from stealing the pages as they are added to the pool but
2151 * before they are reserved.
2153 needed += allocated;
2154 h->resv_huge_pages += delta;
2157 /* Free the needed pages to the hugetlb pool */
2158 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
2164 * This page is now managed by the hugetlb allocator and has
2165 * no users -- drop the buddy allocator's reference.
2167 zeroed = put_page_testzero(page);
2168 VM_BUG_ON_PAGE(!zeroed, page);
2169 enqueue_huge_page(h, page);
2172 spin_unlock_irq(&hugetlb_lock);
2174 /* Free unnecessary surplus pages to the buddy allocator */
2175 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2177 spin_lock_irq(&hugetlb_lock);
2183 * This routine has two main purposes:
2184 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2185 * in unused_resv_pages. This corresponds to the prior adjustments made
2186 * to the associated reservation map.
2187 * 2) Free any unused surplus pages that may have been allocated to satisfy
2188 * the reservation. As many as unused_resv_pages may be freed.
2190 static void return_unused_surplus_pages(struct hstate *h,
2191 unsigned long unused_resv_pages)
2193 unsigned long nr_pages;
2195 LIST_HEAD(page_list);
2197 lockdep_assert_held(&hugetlb_lock);
2198 /* Uncommit the reservation */
2199 h->resv_huge_pages -= unused_resv_pages;
2201 /* Cannot return gigantic pages currently */
2202 if (hstate_is_gigantic(h))
2206 * Part (or even all) of the reservation could have been backed
2207 * by pre-allocated pages. Only free surplus pages.
2209 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2212 * We want to release as many surplus pages as possible, spread
2213 * evenly across all nodes with memory. Iterate across these nodes
2214 * until we can no longer free unreserved surplus pages. This occurs
2215 * when the nodes with surplus pages have no free pages.
2216 * remove_pool_huge_page() will balance the freed pages across the
2217 * on-line nodes with memory and will handle the hstate accounting.
2219 while (nr_pages--) {
2220 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2224 list_add(&page->lru, &page_list);
2228 spin_unlock_irq(&hugetlb_lock);
2229 update_and_free_pages_bulk(h, &page_list);
2230 spin_lock_irq(&hugetlb_lock);
2235 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2236 * are used by the huge page allocation routines to manage reservations.
2238 * vma_needs_reservation is called to determine if the huge page at addr
2239 * within the vma has an associated reservation. If a reservation is
2240 * needed, the value 1 is returned. The caller is then responsible for
2241 * managing the global reservation and subpool usage counts. After
2242 * the huge page has been allocated, vma_commit_reservation is called
2243 * to add the page to the reservation map. If the page allocation fails,
2244 * the reservation must be ended instead of committed. vma_end_reservation
2245 * is called in such cases.
2247 * In the normal case, vma_commit_reservation returns the same value
2248 * as the preceding vma_needs_reservation call. The only time this
2249 * is not the case is if a reserve map was changed between calls. It
2250 * is the responsibility of the caller to notice the difference and
2251 * take appropriate action.
2253 * vma_add_reservation is used in error paths where a reservation must
2254 * be restored when a newly allocated huge page must be freed. It is
2255 * to be called after calling vma_needs_reservation to determine if a
2256 * reservation exists.
2258 * vma_del_reservation is used in error paths where an entry in the reserve
2259 * map was created during huge page allocation and must be removed. It is to
2260 * be called after calling vma_needs_reservation to determine if a reservation
2263 enum vma_resv_mode {
2270 static long __vma_reservation_common(struct hstate *h,
2271 struct vm_area_struct *vma, unsigned long addr,
2272 enum vma_resv_mode mode)
2274 struct resv_map *resv;
2277 long dummy_out_regions_needed;
2279 resv = vma_resv_map(vma);
2283 idx = vma_hugecache_offset(h, vma, addr);
2285 case VMA_NEEDS_RESV:
2286 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2287 /* We assume that vma_reservation_* routines always operate on
2288 * 1 page, and that adding to resv map a 1 page entry can only
2289 * ever require 1 region.
2291 VM_BUG_ON(dummy_out_regions_needed != 1);
2293 case VMA_COMMIT_RESV:
2294 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2295 /* region_add calls of range 1 should never fail. */
2299 region_abort(resv, idx, idx + 1, 1);
2303 if (vma->vm_flags & VM_MAYSHARE) {
2304 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2305 /* region_add calls of range 1 should never fail. */
2308 region_abort(resv, idx, idx + 1, 1);
2309 ret = region_del(resv, idx, idx + 1);
2313 if (vma->vm_flags & VM_MAYSHARE) {
2314 region_abort(resv, idx, idx + 1, 1);
2315 ret = region_del(resv, idx, idx + 1);
2317 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2318 /* region_add calls of range 1 should never fail. */
2326 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2329 * We know private mapping must have HPAGE_RESV_OWNER set.
2331 * In most cases, reserves always exist for private mappings.
2332 * However, a file associated with mapping could have been
2333 * hole punched or truncated after reserves were consumed.
2334 * As subsequent fault on such a range will not use reserves.
2335 * Subtle - The reserve map for private mappings has the
2336 * opposite meaning than that of shared mappings. If NO
2337 * entry is in the reserve map, it means a reservation exists.
2338 * If an entry exists in the reserve map, it means the
2339 * reservation has already been consumed. As a result, the
2340 * return value of this routine is the opposite of the
2341 * value returned from reserve map manipulation routines above.
2350 static long vma_needs_reservation(struct hstate *h,
2351 struct vm_area_struct *vma, unsigned long addr)
2353 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2356 static long vma_commit_reservation(struct hstate *h,
2357 struct vm_area_struct *vma, unsigned long addr)
2359 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2362 static void vma_end_reservation(struct hstate *h,
2363 struct vm_area_struct *vma, unsigned long addr)
2365 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2368 static long vma_add_reservation(struct hstate *h,
2369 struct vm_area_struct *vma, unsigned long addr)
2371 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2374 static long vma_del_reservation(struct hstate *h,
2375 struct vm_area_struct *vma, unsigned long addr)
2377 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2381 * This routine is called to restore reservation information on error paths.
2382 * It should ONLY be called for pages allocated via alloc_huge_page(), and
2383 * the hugetlb mutex should remain held when calling this routine.
2385 * It handles two specific cases:
2386 * 1) A reservation was in place and the page consumed the reservation.
2387 * HPageRestoreReserve is set in the page.
2388 * 2) No reservation was in place for the page, so HPageRestoreReserve is
2389 * not set. However, alloc_huge_page always updates the reserve map.
2391 * In case 1, free_huge_page later in the error path will increment the
2392 * global reserve count. But, free_huge_page does not have enough context
2393 * to adjust the reservation map. This case deals primarily with private
2394 * mappings. Adjust the reserve map here to be consistent with global
2395 * reserve count adjustments to be made by free_huge_page. Make sure the
2396 * reserve map indicates there is a reservation present.
2398 * In case 2, simply undo reserve map modifications done by alloc_huge_page.
2400 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2401 unsigned long address, struct page *page)
2403 long rc = vma_needs_reservation(h, vma, address);
2405 if (HPageRestoreReserve(page)) {
2406 if (unlikely(rc < 0))
2408 * Rare out of memory condition in reserve map
2409 * manipulation. Clear HPageRestoreReserve so that
2410 * global reserve count will not be incremented
2411 * by free_huge_page. This will make it appear
2412 * as though the reservation for this page was
2413 * consumed. This may prevent the task from
2414 * faulting in the page at a later time. This
2415 * is better than inconsistent global huge page
2416 * accounting of reserve counts.
2418 ClearHPageRestoreReserve(page);
2420 (void)vma_add_reservation(h, vma, address);
2422 vma_end_reservation(h, vma, address);
2426 * This indicates there is an entry in the reserve map
2427 * added by alloc_huge_page. We know it was added
2428 * before the alloc_huge_page call, otherwise
2429 * HPageRestoreReserve would be set on the page.
2430 * Remove the entry so that a subsequent allocation
2431 * does not consume a reservation.
2433 rc = vma_del_reservation(h, vma, address);
2436 * VERY rare out of memory condition. Since
2437 * we can not delete the entry, set
2438 * HPageRestoreReserve so that the reserve
2439 * count will be incremented when the page
2440 * is freed. This reserve will be consumed
2441 * on a subsequent allocation.
2443 SetHPageRestoreReserve(page);
2444 } else if (rc < 0) {
2446 * Rare out of memory condition from
2447 * vma_needs_reservation call. Memory allocation is
2448 * only attempted if a new entry is needed. Therefore,
2449 * this implies there is not an entry in the
2452 * For shared mappings, no entry in the map indicates
2453 * no reservation. We are done.
2455 if (!(vma->vm_flags & VM_MAYSHARE))
2457 * For private mappings, no entry indicates
2458 * a reservation is present. Since we can
2459 * not add an entry, set SetHPageRestoreReserve
2460 * on the page so reserve count will be
2461 * incremented when freed. This reserve will
2462 * be consumed on a subsequent allocation.
2464 SetHPageRestoreReserve(page);
2467 * No reservation present, do nothing
2469 vma_end_reservation(h, vma, address);
2474 * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one
2475 * @h: struct hstate old page belongs to
2476 * @old_page: Old page to dissolve
2477 * @list: List to isolate the page in case we need to
2478 * Returns 0 on success, otherwise negated error.
2480 static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page,
2481 struct list_head *list)
2483 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2484 int nid = page_to_nid(old_page);
2485 struct page *new_page;
2489 * Before dissolving the page, we need to allocate a new one for the
2490 * pool to remain stable. Here, we allocate the page and 'prep' it
2491 * by doing everything but actually updating counters and adding to
2492 * the pool. This simplifies and let us do most of the processing
2495 new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL);
2498 __prep_new_huge_page(h, new_page);
2501 spin_lock_irq(&hugetlb_lock);
2502 if (!PageHuge(old_page)) {
2504 * Freed from under us. Drop new_page too.
2507 } else if (page_count(old_page)) {
2509 * Someone has grabbed the page, try to isolate it here.
2510 * Fail with -EBUSY if not possible.
2512 spin_unlock_irq(&hugetlb_lock);
2513 if (!isolate_huge_page(old_page, list))
2515 spin_lock_irq(&hugetlb_lock);
2517 } else if (!HPageFreed(old_page)) {
2519 * Page's refcount is 0 but it has not been enqueued in the
2520 * freelist yet. Race window is small, so we can succeed here if
2523 spin_unlock_irq(&hugetlb_lock);
2528 * Ok, old_page is still a genuine free hugepage. Remove it from
2529 * the freelist and decrease the counters. These will be
2530 * incremented again when calling __prep_account_new_huge_page()
2531 * and enqueue_huge_page() for new_page. The counters will remain
2532 * stable since this happens under the lock.
2534 remove_hugetlb_page(h, old_page, false);
2537 * Reference count trick is needed because allocator gives us
2538 * referenced page but the pool requires pages with 0 refcount.
2540 __prep_account_new_huge_page(h, nid);
2541 page_ref_dec(new_page);
2542 enqueue_huge_page(h, new_page);
2545 * Pages have been replaced, we can safely free the old one.
2547 spin_unlock_irq(&hugetlb_lock);
2548 update_and_free_page(h, old_page, false);
2554 spin_unlock_irq(&hugetlb_lock);
2555 update_and_free_page(h, new_page, false);
2560 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2567 * The page might have been dissolved from under our feet, so make sure
2568 * to carefully check the state under the lock.
2569 * Return success when racing as if we dissolved the page ourselves.
2571 spin_lock_irq(&hugetlb_lock);
2572 if (PageHuge(page)) {
2573 head = compound_head(page);
2574 h = page_hstate(head);
2576 spin_unlock_irq(&hugetlb_lock);
2579 spin_unlock_irq(&hugetlb_lock);
2582 * Fence off gigantic pages as there is a cyclic dependency between
2583 * alloc_contig_range and them. Return -ENOMEM as this has the effect
2584 * of bailing out right away without further retrying.
2586 if (hstate_is_gigantic(h))
2589 if (page_count(head) && isolate_huge_page(head, list))
2591 else if (!page_count(head))
2592 ret = alloc_and_dissolve_huge_page(h, head, list);
2597 struct page *alloc_huge_page(struct vm_area_struct *vma,
2598 unsigned long addr, int avoid_reserve)
2600 struct hugepage_subpool *spool = subpool_vma(vma);
2601 struct hstate *h = hstate_vma(vma);
2603 long map_chg, map_commit;
2606 struct hugetlb_cgroup *h_cg;
2607 bool deferred_reserve;
2609 idx = hstate_index(h);
2611 * Examine the region/reserve map to determine if the process
2612 * has a reservation for the page to be allocated. A return
2613 * code of zero indicates a reservation exists (no change).
2615 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2617 return ERR_PTR(-ENOMEM);
2620 * Processes that did not create the mapping will have no
2621 * reserves as indicated by the region/reserve map. Check
2622 * that the allocation will not exceed the subpool limit.
2623 * Allocations for MAP_NORESERVE mappings also need to be
2624 * checked against any subpool limit.
2626 if (map_chg || avoid_reserve) {
2627 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2629 vma_end_reservation(h, vma, addr);
2630 return ERR_PTR(-ENOSPC);
2634 * Even though there was no reservation in the region/reserve
2635 * map, there could be reservations associated with the
2636 * subpool that can be used. This would be indicated if the
2637 * return value of hugepage_subpool_get_pages() is zero.
2638 * However, if avoid_reserve is specified we still avoid even
2639 * the subpool reservations.
2645 /* If this allocation is not consuming a reservation, charge it now.
2647 deferred_reserve = map_chg || avoid_reserve;
2648 if (deferred_reserve) {
2649 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2650 idx, pages_per_huge_page(h), &h_cg);
2652 goto out_subpool_put;
2655 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2657 goto out_uncharge_cgroup_reservation;
2659 spin_lock_irq(&hugetlb_lock);
2661 * glb_chg is passed to indicate whether or not a page must be taken
2662 * from the global free pool (global change). gbl_chg == 0 indicates
2663 * a reservation exists for the allocation.
2665 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2667 spin_unlock_irq(&hugetlb_lock);
2668 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2670 goto out_uncharge_cgroup;
2671 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2672 SetHPageRestoreReserve(page);
2673 h->resv_huge_pages--;
2675 spin_lock_irq(&hugetlb_lock);
2676 list_add(&page->lru, &h->hugepage_activelist);
2679 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2680 /* If allocation is not consuming a reservation, also store the
2681 * hugetlb_cgroup pointer on the page.
2683 if (deferred_reserve) {
2684 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2688 spin_unlock_irq(&hugetlb_lock);
2690 hugetlb_set_page_subpool(page, spool);
2692 map_commit = vma_commit_reservation(h, vma, addr);
2693 if (unlikely(map_chg > map_commit)) {
2695 * The page was added to the reservation map between
2696 * vma_needs_reservation and vma_commit_reservation.
2697 * This indicates a race with hugetlb_reserve_pages.
2698 * Adjust for the subpool count incremented above AND
2699 * in hugetlb_reserve_pages for the same page. Also,
2700 * the reservation count added in hugetlb_reserve_pages
2701 * no longer applies.
2705 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2706 hugetlb_acct_memory(h, -rsv_adjust);
2707 if (deferred_reserve)
2708 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2709 pages_per_huge_page(h), page);
2713 out_uncharge_cgroup:
2714 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2715 out_uncharge_cgroup_reservation:
2716 if (deferred_reserve)
2717 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2720 if (map_chg || avoid_reserve)
2721 hugepage_subpool_put_pages(spool, 1);
2722 vma_end_reservation(h, vma, addr);
2723 return ERR_PTR(-ENOSPC);
2726 int alloc_bootmem_huge_page(struct hstate *h)
2727 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2728 int __alloc_bootmem_huge_page(struct hstate *h)
2730 struct huge_bootmem_page *m;
2733 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2736 addr = memblock_alloc_try_nid_raw(
2737 huge_page_size(h), huge_page_size(h),
2738 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2741 * Use the beginning of the huge page to store the
2742 * huge_bootmem_page struct (until gather_bootmem
2743 * puts them into the mem_map).
2752 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2753 /* Put them into a private list first because mem_map is not up yet */
2754 INIT_LIST_HEAD(&m->list);
2755 list_add(&m->list, &huge_boot_pages);
2761 * Put bootmem huge pages into the standard lists after mem_map is up.
2762 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
2764 static void __init gather_bootmem_prealloc(void)
2766 struct huge_bootmem_page *m;
2768 list_for_each_entry(m, &huge_boot_pages, list) {
2769 struct page *page = virt_to_page(m);
2770 struct hstate *h = m->hstate;
2772 VM_BUG_ON(!hstate_is_gigantic(h));
2773 WARN_ON(page_count(page) != 1);
2774 prep_compound_gigantic_page(page, huge_page_order(h));
2775 WARN_ON(PageReserved(page));
2776 prep_new_huge_page(h, page, page_to_nid(page));
2777 put_page(page); /* free it into the hugepage allocator */
2780 * We need to restore the 'stolen' pages to totalram_pages
2781 * in order to fix confusing memory reports from free(1) and
2782 * other side-effects, like CommitLimit going negative.
2784 adjust_managed_page_count(page, pages_per_huge_page(h));
2789 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2792 nodemask_t *node_alloc_noretry;
2794 if (!hstate_is_gigantic(h)) {
2796 * Bit mask controlling how hard we retry per-node allocations.
2797 * Ignore errors as lower level routines can deal with
2798 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2799 * time, we are likely in bigger trouble.
2801 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2804 /* allocations done at boot time */
2805 node_alloc_noretry = NULL;
2808 /* bit mask controlling how hard we retry per-node allocations */
2809 if (node_alloc_noretry)
2810 nodes_clear(*node_alloc_noretry);
2812 for (i = 0; i < h->max_huge_pages; ++i) {
2813 if (hstate_is_gigantic(h)) {
2814 if (hugetlb_cma_size) {
2815 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2818 if (!alloc_bootmem_huge_page(h))
2820 } else if (!alloc_pool_huge_page(h,
2821 &node_states[N_MEMORY],
2822 node_alloc_noretry))
2826 if (i < h->max_huge_pages) {
2829 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2830 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2831 h->max_huge_pages, buf, i);
2832 h->max_huge_pages = i;
2835 kfree(node_alloc_noretry);
2838 static void __init hugetlb_init_hstates(void)
2842 for_each_hstate(h) {
2843 if (minimum_order > huge_page_order(h))
2844 minimum_order = huge_page_order(h);
2846 /* oversize hugepages were init'ed in early boot */
2847 if (!hstate_is_gigantic(h))
2848 hugetlb_hstate_alloc_pages(h);
2850 VM_BUG_ON(minimum_order == UINT_MAX);
2853 static void __init report_hugepages(void)
2857 for_each_hstate(h) {
2860 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2861 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2862 buf, h->free_huge_pages);
2866 #ifdef CONFIG_HIGHMEM
2867 static void try_to_free_low(struct hstate *h, unsigned long count,
2868 nodemask_t *nodes_allowed)
2871 LIST_HEAD(page_list);
2873 lockdep_assert_held(&hugetlb_lock);
2874 if (hstate_is_gigantic(h))
2878 * Collect pages to be freed on a list, and free after dropping lock
2880 for_each_node_mask(i, *nodes_allowed) {
2881 struct page *page, *next;
2882 struct list_head *freel = &h->hugepage_freelists[i];
2883 list_for_each_entry_safe(page, next, freel, lru) {
2884 if (count >= h->nr_huge_pages)
2886 if (PageHighMem(page))
2888 remove_hugetlb_page(h, page, false);
2889 list_add(&page->lru, &page_list);
2894 spin_unlock_irq(&hugetlb_lock);
2895 update_and_free_pages_bulk(h, &page_list);
2896 spin_lock_irq(&hugetlb_lock);
2899 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2900 nodemask_t *nodes_allowed)
2906 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2907 * balanced by operating on them in a round-robin fashion.
2908 * Returns 1 if an adjustment was made.
2910 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2915 lockdep_assert_held(&hugetlb_lock);
2916 VM_BUG_ON(delta != -1 && delta != 1);
2919 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2920 if (h->surplus_huge_pages_node[node])
2924 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2925 if (h->surplus_huge_pages_node[node] <
2926 h->nr_huge_pages_node[node])
2933 h->surplus_huge_pages += delta;
2934 h->surplus_huge_pages_node[node] += delta;
2938 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2939 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2940 nodemask_t *nodes_allowed)
2942 unsigned long min_count, ret;
2944 LIST_HEAD(page_list);
2945 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2948 * Bit mask controlling how hard we retry per-node allocations.
2949 * If we can not allocate the bit mask, do not attempt to allocate
2950 * the requested huge pages.
2952 if (node_alloc_noretry)
2953 nodes_clear(*node_alloc_noretry);
2958 * resize_lock mutex prevents concurrent adjustments to number of
2959 * pages in hstate via the proc/sysfs interfaces.
2961 mutex_lock(&h->resize_lock);
2962 flush_free_hpage_work(h);
2963 spin_lock_irq(&hugetlb_lock);
2966 * Check for a node specific request.
2967 * Changing node specific huge page count may require a corresponding
2968 * change to the global count. In any case, the passed node mask
2969 * (nodes_allowed) will restrict alloc/free to the specified node.
2971 if (nid != NUMA_NO_NODE) {
2972 unsigned long old_count = count;
2974 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2976 * User may have specified a large count value which caused the
2977 * above calculation to overflow. In this case, they wanted
2978 * to allocate as many huge pages as possible. Set count to
2979 * largest possible value to align with their intention.
2981 if (count < old_count)
2986 * Gigantic pages runtime allocation depend on the capability for large
2987 * page range allocation.
2988 * If the system does not provide this feature, return an error when
2989 * the user tries to allocate gigantic pages but let the user free the
2990 * boottime allocated gigantic pages.
2992 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2993 if (count > persistent_huge_pages(h)) {
2994 spin_unlock_irq(&hugetlb_lock);
2995 mutex_unlock(&h->resize_lock);
2996 NODEMASK_FREE(node_alloc_noretry);
2999 /* Fall through to decrease pool */
3003 * Increase the pool size
3004 * First take pages out of surplus state. Then make up the
3005 * remaining difference by allocating fresh huge pages.
3007 * We might race with alloc_surplus_huge_page() here and be unable
3008 * to convert a surplus huge page to a normal huge page. That is
3009 * not critical, though, it just means the overall size of the
3010 * pool might be one hugepage larger than it needs to be, but
3011 * within all the constraints specified by the sysctls.
3013 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3014 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3018 while (count > persistent_huge_pages(h)) {
3020 * If this allocation races such that we no longer need the
3021 * page, free_huge_page will handle it by freeing the page
3022 * and reducing the surplus.
3024 spin_unlock_irq(&hugetlb_lock);
3026 /* yield cpu to avoid soft lockup */
3029 ret = alloc_pool_huge_page(h, nodes_allowed,
3030 node_alloc_noretry);
3031 spin_lock_irq(&hugetlb_lock);
3035 /* Bail for signals. Probably ctrl-c from user */
3036 if (signal_pending(current))
3041 * Decrease the pool size
3042 * First return free pages to the buddy allocator (being careful
3043 * to keep enough around to satisfy reservations). Then place
3044 * pages into surplus state as needed so the pool will shrink
3045 * to the desired size as pages become free.
3047 * By placing pages into the surplus state independent of the
3048 * overcommit value, we are allowing the surplus pool size to
3049 * exceed overcommit. There are few sane options here. Since
3050 * alloc_surplus_huge_page() is checking the global counter,
3051 * though, we'll note that we're not allowed to exceed surplus
3052 * and won't grow the pool anywhere else. Not until one of the
3053 * sysctls are changed, or the surplus pages go out of use.
3055 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3056 min_count = max(count, min_count);
3057 try_to_free_low(h, min_count, nodes_allowed);
3060 * Collect pages to be removed on list without dropping lock
3062 while (min_count < persistent_huge_pages(h)) {
3063 page = remove_pool_huge_page(h, nodes_allowed, 0);
3067 list_add(&page->lru, &page_list);
3069 /* free the pages after dropping lock */
3070 spin_unlock_irq(&hugetlb_lock);
3071 update_and_free_pages_bulk(h, &page_list);
3072 flush_free_hpage_work(h);
3073 spin_lock_irq(&hugetlb_lock);
3075 while (count < persistent_huge_pages(h)) {
3076 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3080 h->max_huge_pages = persistent_huge_pages(h);
3081 spin_unlock_irq(&hugetlb_lock);
3082 mutex_unlock(&h->resize_lock);
3084 NODEMASK_FREE(node_alloc_noretry);
3089 #define HSTATE_ATTR_RO(_name) \
3090 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3092 #define HSTATE_ATTR(_name) \
3093 static struct kobj_attribute _name##_attr = \
3094 __ATTR(_name, 0644, _name##_show, _name##_store)
3096 static struct kobject *hugepages_kobj;
3097 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3099 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3101 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3105 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3106 if (hstate_kobjs[i] == kobj) {
3108 *nidp = NUMA_NO_NODE;
3112 return kobj_to_node_hstate(kobj, nidp);
3115 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3116 struct kobj_attribute *attr, char *buf)
3119 unsigned long nr_huge_pages;
3122 h = kobj_to_hstate(kobj, &nid);
3123 if (nid == NUMA_NO_NODE)
3124 nr_huge_pages = h->nr_huge_pages;
3126 nr_huge_pages = h->nr_huge_pages_node[nid];
3128 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3131 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3132 struct hstate *h, int nid,
3133 unsigned long count, size_t len)
3136 nodemask_t nodes_allowed, *n_mask;
3138 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3141 if (nid == NUMA_NO_NODE) {
3143 * global hstate attribute
3145 if (!(obey_mempolicy &&
3146 init_nodemask_of_mempolicy(&nodes_allowed)))
3147 n_mask = &node_states[N_MEMORY];
3149 n_mask = &nodes_allowed;
3152 * Node specific request. count adjustment happens in
3153 * set_max_huge_pages() after acquiring hugetlb_lock.
3155 init_nodemask_of_node(&nodes_allowed, nid);
3156 n_mask = &nodes_allowed;
3159 err = set_max_huge_pages(h, count, nid, n_mask);
3161 return err ? err : len;
3164 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3165 struct kobject *kobj, const char *buf,
3169 unsigned long count;
3173 err = kstrtoul(buf, 10, &count);
3177 h = kobj_to_hstate(kobj, &nid);
3178 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3181 static ssize_t nr_hugepages_show(struct kobject *kobj,
3182 struct kobj_attribute *attr, char *buf)
3184 return nr_hugepages_show_common(kobj, attr, buf);
3187 static ssize_t nr_hugepages_store(struct kobject *kobj,
3188 struct kobj_attribute *attr, const char *buf, size_t len)
3190 return nr_hugepages_store_common(false, kobj, buf, len);
3192 HSTATE_ATTR(nr_hugepages);
3197 * hstate attribute for optionally mempolicy-based constraint on persistent
3198 * huge page alloc/free.
3200 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3201 struct kobj_attribute *attr,
3204 return nr_hugepages_show_common(kobj, attr, buf);
3207 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3208 struct kobj_attribute *attr, const char *buf, size_t len)
3210 return nr_hugepages_store_common(true, kobj, buf, len);
3212 HSTATE_ATTR(nr_hugepages_mempolicy);
3216 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3217 struct kobj_attribute *attr, char *buf)
3219 struct hstate *h = kobj_to_hstate(kobj, NULL);
3220 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3223 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3224 struct kobj_attribute *attr, const char *buf, size_t count)
3227 unsigned long input;
3228 struct hstate *h = kobj_to_hstate(kobj, NULL);
3230 if (hstate_is_gigantic(h))
3233 err = kstrtoul(buf, 10, &input);
3237 spin_lock_irq(&hugetlb_lock);
3238 h->nr_overcommit_huge_pages = input;
3239 spin_unlock_irq(&hugetlb_lock);
3243 HSTATE_ATTR(nr_overcommit_hugepages);
3245 static ssize_t free_hugepages_show(struct kobject *kobj,
3246 struct kobj_attribute *attr, char *buf)
3249 unsigned long free_huge_pages;
3252 h = kobj_to_hstate(kobj, &nid);
3253 if (nid == NUMA_NO_NODE)
3254 free_huge_pages = h->free_huge_pages;
3256 free_huge_pages = h->free_huge_pages_node[nid];
3258 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3260 HSTATE_ATTR_RO(free_hugepages);
3262 static ssize_t resv_hugepages_show(struct kobject *kobj,
3263 struct kobj_attribute *attr, char *buf)
3265 struct hstate *h = kobj_to_hstate(kobj, NULL);
3266 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3268 HSTATE_ATTR_RO(resv_hugepages);
3270 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3271 struct kobj_attribute *attr, char *buf)
3274 unsigned long surplus_huge_pages;
3277 h = kobj_to_hstate(kobj, &nid);
3278 if (nid == NUMA_NO_NODE)
3279 surplus_huge_pages = h->surplus_huge_pages;
3281 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3283 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3285 HSTATE_ATTR_RO(surplus_hugepages);
3287 static struct attribute *hstate_attrs[] = {
3288 &nr_hugepages_attr.attr,
3289 &nr_overcommit_hugepages_attr.attr,
3290 &free_hugepages_attr.attr,
3291 &resv_hugepages_attr.attr,
3292 &surplus_hugepages_attr.attr,
3294 &nr_hugepages_mempolicy_attr.attr,
3299 static const struct attribute_group hstate_attr_group = {
3300 .attrs = hstate_attrs,
3303 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3304 struct kobject **hstate_kobjs,
3305 const struct attribute_group *hstate_attr_group)
3308 int hi = hstate_index(h);
3310 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3311 if (!hstate_kobjs[hi])
3314 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
3316 kobject_put(hstate_kobjs[hi]);
3317 hstate_kobjs[hi] = NULL;
3323 static void __init hugetlb_sysfs_init(void)
3328 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3329 if (!hugepages_kobj)
3332 for_each_hstate(h) {
3333 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3334 hstate_kobjs, &hstate_attr_group);
3336 pr_err("HugeTLB: Unable to add hstate %s", h->name);
3343 * node_hstate/s - associate per node hstate attributes, via their kobjects,
3344 * with node devices in node_devices[] using a parallel array. The array
3345 * index of a node device or _hstate == node id.
3346 * This is here to avoid any static dependency of the node device driver, in
3347 * the base kernel, on the hugetlb module.
3349 struct node_hstate {
3350 struct kobject *hugepages_kobj;
3351 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3353 static struct node_hstate node_hstates[MAX_NUMNODES];
3356 * A subset of global hstate attributes for node devices
3358 static struct attribute *per_node_hstate_attrs[] = {
3359 &nr_hugepages_attr.attr,
3360 &free_hugepages_attr.attr,
3361 &surplus_hugepages_attr.attr,
3365 static const struct attribute_group per_node_hstate_attr_group = {
3366 .attrs = per_node_hstate_attrs,
3370 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
3371 * Returns node id via non-NULL nidp.
3373 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3377 for (nid = 0; nid < nr_node_ids; nid++) {
3378 struct node_hstate *nhs = &node_hstates[nid];
3380 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3381 if (nhs->hstate_kobjs[i] == kobj) {
3393 * Unregister hstate attributes from a single node device.
3394 * No-op if no hstate attributes attached.
3396 static void hugetlb_unregister_node(struct node *node)
3399 struct node_hstate *nhs = &node_hstates[node->dev.id];
3401 if (!nhs->hugepages_kobj)
3402 return; /* no hstate attributes */
3404 for_each_hstate(h) {
3405 int idx = hstate_index(h);
3406 if (nhs->hstate_kobjs[idx]) {
3407 kobject_put(nhs->hstate_kobjs[idx]);
3408 nhs->hstate_kobjs[idx] = NULL;
3412 kobject_put(nhs->hugepages_kobj);
3413 nhs->hugepages_kobj = NULL;
3418 * Register hstate attributes for a single node device.
3419 * No-op if attributes already registered.
3421 static void hugetlb_register_node(struct node *node)
3424 struct node_hstate *nhs = &node_hstates[node->dev.id];
3427 if (nhs->hugepages_kobj)
3428 return; /* already allocated */
3430 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
3432 if (!nhs->hugepages_kobj)
3435 for_each_hstate(h) {
3436 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3438 &per_node_hstate_attr_group);
3440 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
3441 h->name, node->dev.id);
3442 hugetlb_unregister_node(node);
3449 * hugetlb init time: register hstate attributes for all registered node
3450 * devices of nodes that have memory. All on-line nodes should have
3451 * registered their associated device by this time.
3453 static void __init hugetlb_register_all_nodes(void)
3457 for_each_node_state(nid, N_MEMORY) {
3458 struct node *node = node_devices[nid];
3459 if (node->dev.id == nid)
3460 hugetlb_register_node(node);
3464 * Let the node device driver know we're here so it can
3465 * [un]register hstate attributes on node hotplug.
3467 register_hugetlbfs_with_node(hugetlb_register_node,
3468 hugetlb_unregister_node);
3470 #else /* !CONFIG_NUMA */
3472 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3480 static void hugetlb_register_all_nodes(void) { }
3484 static int __init hugetlb_init(void)
3488 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3491 if (!hugepages_supported()) {
3492 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3493 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
3498 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3499 * architectures depend on setup being done here.
3501 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3502 if (!parsed_default_hugepagesz) {
3504 * If we did not parse a default huge page size, set
3505 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3506 * number of huge pages for this default size was implicitly
3507 * specified, set that here as well.
3508 * Note that the implicit setting will overwrite an explicit
3509 * setting. A warning will be printed in this case.
3511 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3512 if (default_hstate_max_huge_pages) {
3513 if (default_hstate.max_huge_pages) {
3516 string_get_size(huge_page_size(&default_hstate),
3517 1, STRING_UNITS_2, buf, 32);
3518 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3519 default_hstate.max_huge_pages, buf);
3520 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3521 default_hstate_max_huge_pages);
3523 default_hstate.max_huge_pages =
3524 default_hstate_max_huge_pages;
3528 hugetlb_cma_check();
3529 hugetlb_init_hstates();
3530 gather_bootmem_prealloc();
3533 hugetlb_sysfs_init();
3534 hugetlb_register_all_nodes();
3535 hugetlb_cgroup_file_init();
3538 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3540 num_fault_mutexes = 1;
3542 hugetlb_fault_mutex_table =
3543 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3545 BUG_ON(!hugetlb_fault_mutex_table);
3547 for (i = 0; i < num_fault_mutexes; i++)
3548 mutex_init(&hugetlb_fault_mutex_table[i]);
3551 subsys_initcall(hugetlb_init);
3553 /* Overwritten by architectures with more huge page sizes */
3554 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
3556 return size == HPAGE_SIZE;
3559 void __init hugetlb_add_hstate(unsigned int order)
3564 if (size_to_hstate(PAGE_SIZE << order)) {
3567 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
3569 h = &hstates[hugetlb_max_hstate++];
3570 mutex_init(&h->resize_lock);
3572 h->mask = ~(huge_page_size(h) - 1);
3573 for (i = 0; i < MAX_NUMNODES; ++i)
3574 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
3575 INIT_LIST_HEAD(&h->hugepage_activelist);
3576 h->next_nid_to_alloc = first_memory_node;
3577 h->next_nid_to_free = first_memory_node;
3578 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3579 huge_page_size(h)/1024);
3580 hugetlb_vmemmap_init(h);
3586 * hugepages command line processing
3587 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3588 * specification. If not, ignore the hugepages value. hugepages can also
3589 * be the first huge page command line option in which case it implicitly
3590 * specifies the number of huge pages for the default size.
3592 static int __init hugepages_setup(char *s)
3595 static unsigned long *last_mhp;
3597 if (!parsed_valid_hugepagesz) {
3598 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
3599 parsed_valid_hugepagesz = true;
3604 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3605 * yet, so this hugepages= parameter goes to the "default hstate".
3606 * Otherwise, it goes with the previously parsed hugepagesz or
3607 * default_hugepagesz.
3609 else if (!hugetlb_max_hstate)
3610 mhp = &default_hstate_max_huge_pages;
3612 mhp = &parsed_hstate->max_huge_pages;
3614 if (mhp == last_mhp) {
3615 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3619 if (sscanf(s, "%lu", mhp) <= 0)
3623 * Global state is always initialized later in hugetlb_init.
3624 * But we need to allocate gigantic hstates here early to still
3625 * use the bootmem allocator.
3627 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
3628 hugetlb_hstate_alloc_pages(parsed_hstate);
3634 __setup("hugepages=", hugepages_setup);
3637 * hugepagesz command line processing
3638 * A specific huge page size can only be specified once with hugepagesz.
3639 * hugepagesz is followed by hugepages on the command line. The global
3640 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3641 * hugepagesz argument was valid.
3643 static int __init hugepagesz_setup(char *s)
3648 parsed_valid_hugepagesz = false;
3649 size = (unsigned long)memparse(s, NULL);
3651 if (!arch_hugetlb_valid_size(size)) {
3652 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
3656 h = size_to_hstate(size);
3659 * hstate for this size already exists. This is normally
3660 * an error, but is allowed if the existing hstate is the
3661 * default hstate. More specifically, it is only allowed if
3662 * the number of huge pages for the default hstate was not
3663 * previously specified.
3665 if (!parsed_default_hugepagesz || h != &default_hstate ||
3666 default_hstate.max_huge_pages) {
3667 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3672 * No need to call hugetlb_add_hstate() as hstate already
3673 * exists. But, do set parsed_hstate so that a following
3674 * hugepages= parameter will be applied to this hstate.
3677 parsed_valid_hugepagesz = true;
3681 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3682 parsed_valid_hugepagesz = true;
3685 __setup("hugepagesz=", hugepagesz_setup);
3688 * default_hugepagesz command line input
3689 * Only one instance of default_hugepagesz allowed on command line.
3691 static int __init default_hugepagesz_setup(char *s)
3695 parsed_valid_hugepagesz = false;
3696 if (parsed_default_hugepagesz) {
3697 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3701 size = (unsigned long)memparse(s, NULL);
3703 if (!arch_hugetlb_valid_size(size)) {
3704 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3708 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3709 parsed_valid_hugepagesz = true;
3710 parsed_default_hugepagesz = true;
3711 default_hstate_idx = hstate_index(size_to_hstate(size));
3714 * The number of default huge pages (for this size) could have been
3715 * specified as the first hugetlb parameter: hugepages=X. If so,
3716 * then default_hstate_max_huge_pages is set. If the default huge
3717 * page size is gigantic (>= MAX_ORDER), then the pages must be
3718 * allocated here from bootmem allocator.
3720 if (default_hstate_max_huge_pages) {
3721 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3722 if (hstate_is_gigantic(&default_hstate))
3723 hugetlb_hstate_alloc_pages(&default_hstate);
3724 default_hstate_max_huge_pages = 0;
3729 __setup("default_hugepagesz=", default_hugepagesz_setup);
3731 static unsigned int allowed_mems_nr(struct hstate *h)
3734 unsigned int nr = 0;
3735 nodemask_t *mpol_allowed;
3736 unsigned int *array = h->free_huge_pages_node;
3737 gfp_t gfp_mask = htlb_alloc_mask(h);
3739 mpol_allowed = policy_nodemask_current(gfp_mask);
3741 for_each_node_mask(node, cpuset_current_mems_allowed) {
3742 if (!mpol_allowed || node_isset(node, *mpol_allowed))
3749 #ifdef CONFIG_SYSCTL
3750 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3751 void *buffer, size_t *length,
3752 loff_t *ppos, unsigned long *out)
3754 struct ctl_table dup_table;
3757 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3758 * can duplicate the @table and alter the duplicate of it.
3761 dup_table.data = out;
3763 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3766 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3767 struct ctl_table *table, int write,
3768 void *buffer, size_t *length, loff_t *ppos)
3770 struct hstate *h = &default_hstate;
3771 unsigned long tmp = h->max_huge_pages;
3774 if (!hugepages_supported())
3777 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3783 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3784 NUMA_NO_NODE, tmp, *length);
3789 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
3790 void *buffer, size_t *length, loff_t *ppos)
3793 return hugetlb_sysctl_handler_common(false, table, write,
3794 buffer, length, ppos);
3798 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3799 void *buffer, size_t *length, loff_t *ppos)
3801 return hugetlb_sysctl_handler_common(true, table, write,
3802 buffer, length, ppos);
3804 #endif /* CONFIG_NUMA */
3806 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3807 void *buffer, size_t *length, loff_t *ppos)
3809 struct hstate *h = &default_hstate;
3813 if (!hugepages_supported())
3816 tmp = h->nr_overcommit_huge_pages;
3818 if (write && hstate_is_gigantic(h))
3821 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3827 spin_lock_irq(&hugetlb_lock);
3828 h->nr_overcommit_huge_pages = tmp;
3829 spin_unlock_irq(&hugetlb_lock);
3835 #endif /* CONFIG_SYSCTL */
3837 void hugetlb_report_meminfo(struct seq_file *m)
3840 unsigned long total = 0;
3842 if (!hugepages_supported())
3845 for_each_hstate(h) {
3846 unsigned long count = h->nr_huge_pages;
3848 total += huge_page_size(h) * count;
3850 if (h == &default_hstate)
3852 "HugePages_Total: %5lu\n"
3853 "HugePages_Free: %5lu\n"
3854 "HugePages_Rsvd: %5lu\n"
3855 "HugePages_Surp: %5lu\n"
3856 "Hugepagesize: %8lu kB\n",
3860 h->surplus_huge_pages,
3861 huge_page_size(h) / SZ_1K);
3864 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
3867 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
3869 struct hstate *h = &default_hstate;
3871 if (!hugepages_supported())
3874 return sysfs_emit_at(buf, len,
3875 "Node %d HugePages_Total: %5u\n"
3876 "Node %d HugePages_Free: %5u\n"
3877 "Node %d HugePages_Surp: %5u\n",
3878 nid, h->nr_huge_pages_node[nid],
3879 nid, h->free_huge_pages_node[nid],
3880 nid, h->surplus_huge_pages_node[nid]);
3883 void hugetlb_show_meminfo(void)
3888 if (!hugepages_supported())
3891 for_each_node_state(nid, N_MEMORY)
3893 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3895 h->nr_huge_pages_node[nid],
3896 h->free_huge_pages_node[nid],
3897 h->surplus_huge_pages_node[nid],
3898 huge_page_size(h) / SZ_1K);
3901 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3903 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3904 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3907 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3908 unsigned long hugetlb_total_pages(void)
3911 unsigned long nr_total_pages = 0;
3914 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3915 return nr_total_pages;
3918 static int hugetlb_acct_memory(struct hstate *h, long delta)
3925 spin_lock_irq(&hugetlb_lock);
3927 * When cpuset is configured, it breaks the strict hugetlb page
3928 * reservation as the accounting is done on a global variable. Such
3929 * reservation is completely rubbish in the presence of cpuset because
3930 * the reservation is not checked against page availability for the
3931 * current cpuset. Application can still potentially OOM'ed by kernel
3932 * with lack of free htlb page in cpuset that the task is in.
3933 * Attempt to enforce strict accounting with cpuset is almost
3934 * impossible (or too ugly) because cpuset is too fluid that
3935 * task or memory node can be dynamically moved between cpusets.
3937 * The change of semantics for shared hugetlb mapping with cpuset is
3938 * undesirable. However, in order to preserve some of the semantics,
3939 * we fall back to check against current free page availability as
3940 * a best attempt and hopefully to minimize the impact of changing
3941 * semantics that cpuset has.
3943 * Apart from cpuset, we also have memory policy mechanism that
3944 * also determines from which node the kernel will allocate memory
3945 * in a NUMA system. So similar to cpuset, we also should consider
3946 * the memory policy of the current task. Similar to the description
3950 if (gather_surplus_pages(h, delta) < 0)
3953 if (delta > allowed_mems_nr(h)) {
3954 return_unused_surplus_pages(h, delta);
3961 return_unused_surplus_pages(h, (unsigned long) -delta);
3964 spin_unlock_irq(&hugetlb_lock);
3968 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3970 struct resv_map *resv = vma_resv_map(vma);
3973 * This new VMA should share its siblings reservation map if present.
3974 * The VMA will only ever have a valid reservation map pointer where
3975 * it is being copied for another still existing VMA. As that VMA
3976 * has a reference to the reservation map it cannot disappear until
3977 * after this open call completes. It is therefore safe to take a
3978 * new reference here without additional locking.
3980 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3981 kref_get(&resv->refs);
3984 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3986 struct hstate *h = hstate_vma(vma);
3987 struct resv_map *resv = vma_resv_map(vma);
3988 struct hugepage_subpool *spool = subpool_vma(vma);
3989 unsigned long reserve, start, end;
3992 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3995 start = vma_hugecache_offset(h, vma, vma->vm_start);
3996 end = vma_hugecache_offset(h, vma, vma->vm_end);
3998 reserve = (end - start) - region_count(resv, start, end);
3999 hugetlb_cgroup_uncharge_counter(resv, start, end);
4002 * Decrement reserve counts. The global reserve count may be
4003 * adjusted if the subpool has a minimum size.
4005 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4006 hugetlb_acct_memory(h, -gbl_reserve);
4009 kref_put(&resv->refs, resv_map_release);
4012 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4014 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4019 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4021 return huge_page_size(hstate_vma(vma));
4025 * We cannot handle pagefaults against hugetlb pages at all. They cause
4026 * handle_mm_fault() to try to instantiate regular-sized pages in the
4027 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4030 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4037 * When a new function is introduced to vm_operations_struct and added
4038 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4039 * This is because under System V memory model, mappings created via
4040 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4041 * their original vm_ops are overwritten with shm_vm_ops.
4043 const struct vm_operations_struct hugetlb_vm_ops = {
4044 .fault = hugetlb_vm_op_fault,
4045 .open = hugetlb_vm_op_open,
4046 .close = hugetlb_vm_op_close,
4047 .may_split = hugetlb_vm_op_split,
4048 .pagesize = hugetlb_vm_op_pagesize,
4051 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4055 unsigned int shift = huge_page_shift(hstate_vma(vma));
4058 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4059 vma->vm_page_prot)));
4061 entry = huge_pte_wrprotect(mk_huge_pte(page,
4062 vma->vm_page_prot));
4064 entry = pte_mkyoung(entry);
4065 entry = pte_mkhuge(entry);
4066 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4071 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4072 unsigned long address, pte_t *ptep)
4076 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4077 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4078 update_mmu_cache(vma, address, ptep);
4081 bool is_hugetlb_entry_migration(pte_t pte)
4085 if (huge_pte_none(pte) || pte_present(pte))
4087 swp = pte_to_swp_entry(pte);
4088 if (is_migration_entry(swp))
4094 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4098 if (huge_pte_none(pte) || pte_present(pte))
4100 swp = pte_to_swp_entry(pte);
4101 if (is_hwpoison_entry(swp))
4108 hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
4109 struct page *new_page)
4111 __SetPageUptodate(new_page);
4112 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
4113 hugepage_add_new_anon_rmap(new_page, vma, addr);
4114 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
4115 ClearHPageRestoreReserve(new_page);
4116 SetHPageMigratable(new_page);
4119 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
4120 struct vm_area_struct *vma)
4122 pte_t *src_pte, *dst_pte, entry, dst_entry;
4123 struct page *ptepage;
4125 bool cow = is_cow_mapping(vma->vm_flags);
4126 struct hstate *h = hstate_vma(vma);
4127 unsigned long sz = huge_page_size(h);
4128 unsigned long npages = pages_per_huge_page(h);
4129 struct address_space *mapping = vma->vm_file->f_mapping;
4130 struct mmu_notifier_range range;
4134 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
4137 mmu_notifier_invalidate_range_start(&range);
4140 * For shared mappings i_mmap_rwsem must be held to call
4141 * huge_pte_alloc, otherwise the returned ptep could go
4142 * away if part of a shared pmd and another thread calls
4145 i_mmap_lock_read(mapping);
4148 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
4149 spinlock_t *src_ptl, *dst_ptl;
4150 src_pte = huge_pte_offset(src, addr, sz);
4153 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
4160 * If the pagetables are shared don't copy or take references.
4161 * dst_pte == src_pte is the common case of src/dest sharing.
4163 * However, src could have 'unshared' and dst shares with
4164 * another vma. If dst_pte !none, this implies sharing.
4165 * Check here before taking page table lock, and once again
4166 * after taking the lock below.
4168 dst_entry = huge_ptep_get(dst_pte);
4169 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
4172 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4173 src_ptl = huge_pte_lockptr(h, src, src_pte);
4174 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4175 entry = huge_ptep_get(src_pte);
4176 dst_entry = huge_ptep_get(dst_pte);
4178 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
4180 * Skip if src entry none. Also, skip in the
4181 * unlikely case dst entry !none as this implies
4182 * sharing with another vma.
4185 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
4186 is_hugetlb_entry_hwpoisoned(entry))) {
4187 swp_entry_t swp_entry = pte_to_swp_entry(entry);
4189 if (is_write_migration_entry(swp_entry) && cow) {
4191 * COW mappings require pages in both
4192 * parent and child to be set to read.
4194 make_migration_entry_read(&swp_entry);
4195 entry = swp_entry_to_pte(swp_entry);
4196 set_huge_swap_pte_at(src, addr, src_pte,
4199 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
4201 entry = huge_ptep_get(src_pte);
4202 ptepage = pte_page(entry);
4206 * This is a rare case where we see pinned hugetlb
4207 * pages while they're prone to COW. We need to do the
4208 * COW earlier during fork.
4210 * When pre-allocating the page or copying data, we
4211 * need to be without the pgtable locks since we could
4212 * sleep during the process.
4214 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
4215 pte_t src_pte_old = entry;
4218 spin_unlock(src_ptl);
4219 spin_unlock(dst_ptl);
4220 /* Do not use reserve as it's private owned */
4221 new = alloc_huge_page(vma, addr, 1);
4227 copy_user_huge_page(new, ptepage, addr, vma,
4231 /* Install the new huge page if src pte stable */
4232 dst_ptl = huge_pte_lock(h, dst, dst_pte);
4233 src_ptl = huge_pte_lockptr(h, src, src_pte);
4234 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
4235 entry = huge_ptep_get(src_pte);
4236 if (!pte_same(src_pte_old, entry)) {
4237 restore_reserve_on_error(h, vma, addr,
4240 /* dst_entry won't change as in child */
4243 hugetlb_install_page(vma, dst_pte, addr, new);
4244 spin_unlock(src_ptl);
4245 spin_unlock(dst_ptl);
4251 * No need to notify as we are downgrading page
4252 * table protection not changing it to point
4255 * See Documentation/vm/mmu_notifier.rst
4257 huge_ptep_set_wrprotect(src, addr, src_pte);
4258 entry = huge_pte_wrprotect(entry);
4261 page_dup_rmap(ptepage, true);
4262 set_huge_pte_at(dst, addr, dst_pte, entry);
4263 hugetlb_count_add(npages, dst);
4265 spin_unlock(src_ptl);
4266 spin_unlock(dst_ptl);
4270 mmu_notifier_invalidate_range_end(&range);
4272 i_mmap_unlock_read(mapping);
4277 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
4278 unsigned long start, unsigned long end,
4279 struct page *ref_page)
4281 struct mm_struct *mm = vma->vm_mm;
4282 unsigned long address;
4287 struct hstate *h = hstate_vma(vma);
4288 unsigned long sz = huge_page_size(h);
4289 struct mmu_notifier_range range;
4291 WARN_ON(!is_vm_hugetlb_page(vma));
4292 BUG_ON(start & ~huge_page_mask(h));
4293 BUG_ON(end & ~huge_page_mask(h));
4296 * This is a hugetlb vma, all the pte entries should point
4299 tlb_change_page_size(tlb, sz);
4300 tlb_start_vma(tlb, vma);
4303 * If sharing possible, alert mmu notifiers of worst case.
4305 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
4307 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4308 mmu_notifier_invalidate_range_start(&range);
4310 for (; address < end; address += sz) {
4311 ptep = huge_pte_offset(mm, address, sz);
4315 ptl = huge_pte_lock(h, mm, ptep);
4316 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
4319 * We just unmapped a page of PMDs by clearing a PUD.
4320 * The caller's TLB flush range should cover this area.
4325 pte = huge_ptep_get(ptep);
4326 if (huge_pte_none(pte)) {
4332 * Migrating hugepage or HWPoisoned hugepage is already
4333 * unmapped and its refcount is dropped, so just clear pte here.
4335 if (unlikely(!pte_present(pte))) {
4336 huge_pte_clear(mm, address, ptep, sz);
4341 page = pte_page(pte);
4343 * If a reference page is supplied, it is because a specific
4344 * page is being unmapped, not a range. Ensure the page we
4345 * are about to unmap is the actual page of interest.
4348 if (page != ref_page) {
4353 * Mark the VMA as having unmapped its page so that
4354 * future faults in this VMA will fail rather than
4355 * looking like data was lost
4357 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4360 pte = huge_ptep_get_and_clear(mm, address, ptep);
4361 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
4362 if (huge_pte_dirty(pte))
4363 set_page_dirty(page);
4365 hugetlb_count_sub(pages_per_huge_page(h), mm);
4366 page_remove_rmap(page, true);
4369 tlb_remove_page_size(tlb, page, huge_page_size(h));
4371 * Bail out after unmapping reference page if supplied
4376 mmu_notifier_invalidate_range_end(&range);
4377 tlb_end_vma(tlb, vma);
4380 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4381 struct vm_area_struct *vma, unsigned long start,
4382 unsigned long end, struct page *ref_page)
4384 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4387 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4388 * test will fail on a vma being torn down, and not grab a page table
4389 * on its way out. We're lucky that the flag has such an appropriate
4390 * name, and can in fact be safely cleared here. We could clear it
4391 * before the __unmap_hugepage_range above, but all that's necessary
4392 * is to clear it before releasing the i_mmap_rwsem. This works
4393 * because in the context this is called, the VMA is about to be
4394 * destroyed and the i_mmap_rwsem is held.
4396 vma->vm_flags &= ~VM_MAYSHARE;
4399 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
4400 unsigned long end, struct page *ref_page)
4402 struct mmu_gather tlb;
4404 tlb_gather_mmu(&tlb, vma->vm_mm);
4405 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
4406 tlb_finish_mmu(&tlb);
4410 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4411 * mapping it owns the reserve page for. The intention is to unmap the page
4412 * from other VMAs and let the children be SIGKILLed if they are faulting the
4415 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4416 struct page *page, unsigned long address)
4418 struct hstate *h = hstate_vma(vma);
4419 struct vm_area_struct *iter_vma;
4420 struct address_space *mapping;
4424 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4425 * from page cache lookup which is in HPAGE_SIZE units.
4427 address = address & huge_page_mask(h);
4428 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4430 mapping = vma->vm_file->f_mapping;
4433 * Take the mapping lock for the duration of the table walk. As
4434 * this mapping should be shared between all the VMAs,
4435 * __unmap_hugepage_range() is called as the lock is already held
4437 i_mmap_lock_write(mapping);
4438 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
4439 /* Do not unmap the current VMA */
4440 if (iter_vma == vma)
4444 * Shared VMAs have their own reserves and do not affect
4445 * MAP_PRIVATE accounting but it is possible that a shared
4446 * VMA is using the same page so check and skip such VMAs.
4448 if (iter_vma->vm_flags & VM_MAYSHARE)
4452 * Unmap the page from other VMAs without their own reserves.
4453 * They get marked to be SIGKILLed if they fault in these
4454 * areas. This is because a future no-page fault on this VMA
4455 * could insert a zeroed page instead of the data existing
4456 * from the time of fork. This would look like data corruption
4458 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
4459 unmap_hugepage_range(iter_vma, address,
4460 address + huge_page_size(h), page);
4462 i_mmap_unlock_write(mapping);
4466 * Hugetlb_cow() should be called with page lock of the original hugepage held.
4467 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4468 * cannot race with other handlers or page migration.
4469 * Keep the pte_same checks anyway to make transition from the mutex easier.
4471 static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
4472 unsigned long address, pte_t *ptep,
4473 struct page *pagecache_page, spinlock_t *ptl)
4476 struct hstate *h = hstate_vma(vma);
4477 struct page *old_page, *new_page;
4478 int outside_reserve = 0;
4480 unsigned long haddr = address & huge_page_mask(h);
4481 struct mmu_notifier_range range;
4483 pte = huge_ptep_get(ptep);
4484 old_page = pte_page(pte);
4487 /* If no-one else is actually using this page, avoid the copy
4488 * and just make the page writable */
4489 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
4490 page_move_anon_rmap(old_page, vma);
4491 set_huge_ptep_writable(vma, haddr, ptep);
4496 * If the process that created a MAP_PRIVATE mapping is about to
4497 * perform a COW due to a shared page count, attempt to satisfy
4498 * the allocation without using the existing reserves. The pagecache
4499 * page is used to determine if the reserve at this address was
4500 * consumed or not. If reserves were used, a partial faulted mapping
4501 * at the time of fork() could consume its reserves on COW instead
4502 * of the full address range.
4504 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
4505 old_page != pagecache_page)
4506 outside_reserve = 1;
4511 * Drop page table lock as buddy allocator may be called. It will
4512 * be acquired again before returning to the caller, as expected.
4515 new_page = alloc_huge_page(vma, haddr, outside_reserve);
4517 if (IS_ERR(new_page)) {
4519 * If a process owning a MAP_PRIVATE mapping fails to COW,
4520 * it is due to references held by a child and an insufficient
4521 * huge page pool. To guarantee the original mappers
4522 * reliability, unmap the page from child processes. The child
4523 * may get SIGKILLed if it later faults.
4525 if (outside_reserve) {
4526 struct address_space *mapping = vma->vm_file->f_mapping;
4531 BUG_ON(huge_pte_none(pte));
4533 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4534 * unmapping. unmapping needs to hold i_mmap_rwsem
4535 * in write mode. Dropping i_mmap_rwsem in read mode
4536 * here is OK as COW mappings do not interact with
4539 * Reacquire both after unmap operation.
4541 idx = vma_hugecache_offset(h, vma, haddr);
4542 hash = hugetlb_fault_mutex_hash(mapping, idx);
4543 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4544 i_mmap_unlock_read(mapping);
4546 unmap_ref_private(mm, vma, old_page, haddr);
4548 i_mmap_lock_read(mapping);
4549 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4551 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4553 pte_same(huge_ptep_get(ptep), pte)))
4554 goto retry_avoidcopy;
4556 * race occurs while re-acquiring page table
4557 * lock, and our job is done.
4562 ret = vmf_error(PTR_ERR(new_page));
4563 goto out_release_old;
4567 * When the original hugepage is shared one, it does not have
4568 * anon_vma prepared.
4570 if (unlikely(anon_vma_prepare(vma))) {
4572 goto out_release_all;
4575 copy_user_huge_page(new_page, old_page, address, vma,
4576 pages_per_huge_page(h));
4577 __SetPageUptodate(new_page);
4579 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
4580 haddr + huge_page_size(h));
4581 mmu_notifier_invalidate_range_start(&range);
4584 * Retake the page table lock to check for racing updates
4585 * before the page tables are altered
4588 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4589 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
4590 ClearHPageRestoreReserve(new_page);
4593 huge_ptep_clear_flush(vma, haddr, ptep);
4594 mmu_notifier_invalidate_range(mm, range.start, range.end);
4595 set_huge_pte_at(mm, haddr, ptep,
4596 make_huge_pte(vma, new_page, 1));
4597 page_remove_rmap(old_page, true);
4598 hugepage_add_new_anon_rmap(new_page, vma, haddr);
4599 SetHPageMigratable(new_page);
4600 /* Make the old page be freed below */
4601 new_page = old_page;
4604 mmu_notifier_invalidate_range_end(&range);
4606 restore_reserve_on_error(h, vma, haddr, new_page);
4611 spin_lock(ptl); /* Caller expects lock to be held */
4615 /* Return the pagecache page at a given address within a VMA */
4616 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4617 struct vm_area_struct *vma, unsigned long address)
4619 struct address_space *mapping;
4622 mapping = vma->vm_file->f_mapping;
4623 idx = vma_hugecache_offset(h, vma, address);
4625 return find_lock_page(mapping, idx);
4629 * Return whether there is a pagecache page to back given address within VMA.
4630 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4632 static bool hugetlbfs_pagecache_present(struct hstate *h,
4633 struct vm_area_struct *vma, unsigned long address)
4635 struct address_space *mapping;
4639 mapping = vma->vm_file->f_mapping;
4640 idx = vma_hugecache_offset(h, vma, address);
4642 page = find_get_page(mapping, idx);
4645 return page != NULL;
4648 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4651 struct inode *inode = mapping->host;
4652 struct hstate *h = hstate_inode(inode);
4653 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4657 ClearHPageRestoreReserve(page);
4660 * set page dirty so that it will not be removed from cache/file
4661 * by non-hugetlbfs specific code paths.
4663 set_page_dirty(page);
4665 spin_lock(&inode->i_lock);
4666 inode->i_blocks += blocks_per_huge_page(h);
4667 spin_unlock(&inode->i_lock);
4671 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
4672 struct address_space *mapping,
4675 unsigned long haddr,
4676 unsigned long reason)
4680 struct vm_fault vmf = {
4686 * Hard to debug if it ends up being
4687 * used by a callee that assumes
4688 * something about the other
4689 * uninitialized fields... same as in
4695 * hugetlb_fault_mutex and i_mmap_rwsem must be
4696 * dropped before handling userfault. Reacquire
4697 * after handling fault to make calling code simpler.
4699 hash = hugetlb_fault_mutex_hash(mapping, idx);
4700 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4701 i_mmap_unlock_read(mapping);
4702 ret = handle_userfault(&vmf, reason);
4703 i_mmap_lock_read(mapping);
4704 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4709 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4710 struct vm_area_struct *vma,
4711 struct address_space *mapping, pgoff_t idx,
4712 unsigned long address, pte_t *ptep, unsigned int flags)
4714 struct hstate *h = hstate_vma(vma);
4715 vm_fault_t ret = VM_FAULT_SIGBUS;
4721 unsigned long haddr = address & huge_page_mask(h);
4722 bool new_page = false;
4725 * Currently, we are forced to kill the process in the event the
4726 * original mapper has unmapped pages from the child due to a failed
4727 * COW. Warn that such a situation has occurred as it may not be obvious
4729 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
4730 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
4736 * We can not race with truncation due to holding i_mmap_rwsem.
4737 * i_size is modified when holding i_mmap_rwsem, so check here
4738 * once for faults beyond end of file.
4740 size = i_size_read(mapping->host) >> huge_page_shift(h);
4745 page = find_lock_page(mapping, idx);
4747 /* Check for page in userfault range */
4748 if (userfaultfd_missing(vma)) {
4749 ret = hugetlb_handle_userfault(vma, mapping, idx,
4755 page = alloc_huge_page(vma, haddr, 0);
4758 * Returning error will result in faulting task being
4759 * sent SIGBUS. The hugetlb fault mutex prevents two
4760 * tasks from racing to fault in the same page which
4761 * could result in false unable to allocate errors.
4762 * Page migration does not take the fault mutex, but
4763 * does a clear then write of pte's under page table
4764 * lock. Page fault code could race with migration,
4765 * notice the clear pte and try to allocate a page
4766 * here. Before returning error, get ptl and make
4767 * sure there really is no pte entry.
4769 ptl = huge_pte_lock(h, mm, ptep);
4771 if (huge_pte_none(huge_ptep_get(ptep)))
4772 ret = vmf_error(PTR_ERR(page));
4776 clear_huge_page(page, address, pages_per_huge_page(h));
4777 __SetPageUptodate(page);
4780 if (vma->vm_flags & VM_MAYSHARE) {
4781 int err = huge_add_to_page_cache(page, mapping, idx);
4790 if (unlikely(anon_vma_prepare(vma))) {
4792 goto backout_unlocked;
4798 * If memory error occurs between mmap() and fault, some process
4799 * don't have hwpoisoned swap entry for errored virtual address.
4800 * So we need to block hugepage fault by PG_hwpoison bit check.
4802 if (unlikely(PageHWPoison(page))) {
4803 ret = VM_FAULT_HWPOISON_LARGE |
4804 VM_FAULT_SET_HINDEX(hstate_index(h));
4805 goto backout_unlocked;
4808 /* Check for page in userfault range. */
4809 if (userfaultfd_minor(vma)) {
4812 ret = hugetlb_handle_userfault(vma, mapping, idx,
4820 * If we are going to COW a private mapping later, we examine the
4821 * pending reservations for this page now. This will ensure that
4822 * any allocations necessary to record that reservation occur outside
4825 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4826 if (vma_needs_reservation(h, vma, haddr) < 0) {
4828 goto backout_unlocked;
4830 /* Just decrements count, does not deallocate */
4831 vma_end_reservation(h, vma, haddr);
4834 ptl = huge_pte_lock(h, mm, ptep);
4836 if (!huge_pte_none(huge_ptep_get(ptep)))
4840 ClearHPageRestoreReserve(page);
4841 hugepage_add_new_anon_rmap(page, vma, haddr);
4843 page_dup_rmap(page, true);
4844 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4845 && (vma->vm_flags & VM_SHARED)));
4846 set_huge_pte_at(mm, haddr, ptep, new_pte);
4848 hugetlb_count_add(pages_per_huge_page(h), mm);
4849 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
4850 /* Optimization, do the COW without a second fault */
4851 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
4857 * Only set HPageMigratable in newly allocated pages. Existing pages
4858 * found in the pagecache may not have HPageMigratableset if they have
4859 * been isolated for migration.
4862 SetHPageMigratable(page);
4872 restore_reserve_on_error(h, vma, haddr, page);
4878 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4880 unsigned long key[2];
4883 key[0] = (unsigned long) mapping;
4886 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
4888 return hash & (num_fault_mutexes - 1);
4892 * For uniprocessor systems we always use a single mutex, so just
4893 * return 0 and avoid the hashing overhead.
4895 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
4901 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
4902 unsigned long address, unsigned int flags)
4909 struct page *page = NULL;
4910 struct page *pagecache_page = NULL;
4911 struct hstate *h = hstate_vma(vma);
4912 struct address_space *mapping;
4913 int need_wait_lock = 0;
4914 unsigned long haddr = address & huge_page_mask(h);
4916 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4919 * Since we hold no locks, ptep could be stale. That is
4920 * OK as we are only making decisions based on content and
4921 * not actually modifying content here.
4923 entry = huge_ptep_get(ptep);
4924 if (unlikely(is_hugetlb_entry_migration(entry))) {
4925 migration_entry_wait_huge(vma, mm, ptep);
4927 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4928 return VM_FAULT_HWPOISON_LARGE |
4929 VM_FAULT_SET_HINDEX(hstate_index(h));
4933 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
4934 * until finished with ptep. This serves two purposes:
4935 * 1) It prevents huge_pmd_unshare from being called elsewhere
4936 * and making the ptep no longer valid.
4937 * 2) It synchronizes us with i_size modifications during truncation.
4939 * ptep could have already be assigned via huge_pte_offset. That
4940 * is OK, as huge_pte_alloc will return the same value unless
4941 * something has changed.
4943 mapping = vma->vm_file->f_mapping;
4944 i_mmap_lock_read(mapping);
4945 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
4947 i_mmap_unlock_read(mapping);
4948 return VM_FAULT_OOM;
4952 * Serialize hugepage allocation and instantiation, so that we don't
4953 * get spurious allocation failures if two CPUs race to instantiate
4954 * the same page in the page cache.
4956 idx = vma_hugecache_offset(h, vma, haddr);
4957 hash = hugetlb_fault_mutex_hash(mapping, idx);
4958 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4960 entry = huge_ptep_get(ptep);
4961 if (huge_pte_none(entry)) {
4962 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4969 * entry could be a migration/hwpoison entry at this point, so this
4970 * check prevents the kernel from going below assuming that we have
4971 * an active hugepage in pagecache. This goto expects the 2nd page
4972 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4973 * properly handle it.
4975 if (!pte_present(entry))
4979 * If we are going to COW the mapping later, we examine the pending
4980 * reservations for this page now. This will ensure that any
4981 * allocations necessary to record that reservation occur outside the
4982 * spinlock. For private mappings, we also lookup the pagecache
4983 * page now as it is used to determine if a reservation has been
4986 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4987 if (vma_needs_reservation(h, vma, haddr) < 0) {
4991 /* Just decrements count, does not deallocate */
4992 vma_end_reservation(h, vma, haddr);
4994 if (!(vma->vm_flags & VM_MAYSHARE))
4995 pagecache_page = hugetlbfs_pagecache_page(h,
4999 ptl = huge_pte_lock(h, mm, ptep);
5001 /* Check for a racing update before calling hugetlb_cow */
5002 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
5006 * hugetlb_cow() requires page locks of pte_page(entry) and
5007 * pagecache_page, so here we need take the former one
5008 * when page != pagecache_page or !pagecache_page.
5010 page = pte_page(entry);
5011 if (page != pagecache_page)
5012 if (!trylock_page(page)) {
5019 if (flags & FAULT_FLAG_WRITE) {
5020 if (!huge_pte_write(entry)) {
5021 ret = hugetlb_cow(mm, vma, address, ptep,
5022 pagecache_page, ptl);
5025 entry = huge_pte_mkdirty(entry);
5027 entry = pte_mkyoung(entry);
5028 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
5029 flags & FAULT_FLAG_WRITE))
5030 update_mmu_cache(vma, haddr, ptep);
5032 if (page != pagecache_page)
5038 if (pagecache_page) {
5039 unlock_page(pagecache_page);
5040 put_page(pagecache_page);
5043 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5044 i_mmap_unlock_read(mapping);
5046 * Generally it's safe to hold refcount during waiting page lock. But
5047 * here we just wait to defer the next page fault to avoid busy loop and
5048 * the page is not used after unlocked before returning from the current
5049 * page fault. So we are safe from accessing freed page, even if we wait
5050 * here without taking refcount.
5053 wait_on_page_locked(page);
5057 #ifdef CONFIG_USERFAULTFD
5059 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
5060 * modifications for huge pages.
5062 int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
5064 struct vm_area_struct *dst_vma,
5065 unsigned long dst_addr,
5066 unsigned long src_addr,
5067 enum mcopy_atomic_mode mode,
5068 struct page **pagep)
5070 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
5071 struct hstate *h = hstate_vma(dst_vma);
5072 struct address_space *mapping = dst_vma->vm_file->f_mapping;
5073 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
5075 int vm_shared = dst_vma->vm_flags & VM_SHARED;
5084 page = find_lock_page(mapping, idx);
5087 } else if (!*pagep) {
5088 /* If a page already exists, then it's UFFDIO_COPY for
5089 * a non-missing case. Return -EEXIST.
5092 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5097 page = alloc_huge_page(dst_vma, dst_addr, 0);
5103 ret = copy_huge_page_from_user(page,
5104 (const void __user *) src_addr,
5105 pages_per_huge_page(h), false);
5107 /* fallback to copy_from_user outside mmap_lock */
5108 if (unlikely(ret)) {
5110 /* Free the allocated page which may have
5111 * consumed a reservation.
5113 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5116 /* Allocate a temporary page to hold the copied
5119 page = alloc_huge_page_vma(h, dst_vma, dst_addr);
5125 /* Set the outparam pagep and return to the caller to
5126 * copy the contents outside the lock. Don't free the
5133 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
5140 page = alloc_huge_page(dst_vma, dst_addr, 0);
5146 copy_huge_page(page, *pagep);
5152 * The memory barrier inside __SetPageUptodate makes sure that
5153 * preceding stores to the page contents become visible before
5154 * the set_pte_at() write.
5156 __SetPageUptodate(page);
5158 /* Add shared, newly allocated pages to the page cache. */
5159 if (vm_shared && !is_continue) {
5160 size = i_size_read(mapping->host) >> huge_page_shift(h);
5163 goto out_release_nounlock;
5166 * Serialization between remove_inode_hugepages() and
5167 * huge_add_to_page_cache() below happens through the
5168 * hugetlb_fault_mutex_table that here must be hold by
5171 ret = huge_add_to_page_cache(page, mapping, idx);
5173 goto out_release_nounlock;
5176 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
5180 * Recheck the i_size after holding PT lock to make sure not
5181 * to leave any page mapped (as page_mapped()) beyond the end
5182 * of the i_size (remove_inode_hugepages() is strict about
5183 * enforcing that). If we bail out here, we'll also leave a
5184 * page in the radix tree in the vm_shared case beyond the end
5185 * of the i_size, but remove_inode_hugepages() will take care
5186 * of it as soon as we drop the hugetlb_fault_mutex_table.
5188 size = i_size_read(mapping->host) >> huge_page_shift(h);
5191 goto out_release_unlock;
5194 if (!huge_pte_none(huge_ptep_get(dst_pte)))
5195 goto out_release_unlock;
5198 page_dup_rmap(page, true);
5200 ClearHPageRestoreReserve(page);
5201 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
5204 /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
5205 if (is_continue && !vm_shared)
5208 writable = dst_vma->vm_flags & VM_WRITE;
5210 _dst_pte = make_huge_pte(dst_vma, page, writable);
5212 _dst_pte = huge_pte_mkdirty(_dst_pte);
5213 _dst_pte = pte_mkyoung(_dst_pte);
5215 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
5217 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
5218 dst_vma->vm_flags & VM_WRITE);
5219 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
5221 /* No need to invalidate - it was non-present before */
5222 update_mmu_cache(dst_vma, dst_addr, dst_pte);
5226 SetHPageMigratable(page);
5227 if (vm_shared || is_continue)
5234 if (vm_shared || is_continue)
5236 out_release_nounlock:
5237 restore_reserve_on_error(h, dst_vma, dst_addr, page);
5241 #endif /* CONFIG_USERFAULTFD */
5243 static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
5244 int refs, struct page **pages,
5245 struct vm_area_struct **vmas)
5249 for (nr = 0; nr < refs; nr++) {
5251 pages[nr] = mem_map_offset(page, nr);
5257 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
5258 struct page **pages, struct vm_area_struct **vmas,
5259 unsigned long *position, unsigned long *nr_pages,
5260 long i, unsigned int flags, int *locked)
5262 unsigned long pfn_offset;
5263 unsigned long vaddr = *position;
5264 unsigned long remainder = *nr_pages;
5265 struct hstate *h = hstate_vma(vma);
5266 int err = -EFAULT, refs;
5268 while (vaddr < vma->vm_end && remainder) {
5270 spinlock_t *ptl = NULL;
5275 * If we have a pending SIGKILL, don't keep faulting pages and
5276 * potentially allocating memory.
5278 if (fatal_signal_pending(current)) {
5284 * Some archs (sparc64, sh*) have multiple pte_ts to
5285 * each hugepage. We have to make sure we get the
5286 * first, for the page indexing below to work.
5288 * Note that page table lock is not held when pte is null.
5290 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
5293 ptl = huge_pte_lock(h, mm, pte);
5294 absent = !pte || huge_pte_none(huge_ptep_get(pte));
5297 * When coredumping, it suits get_dump_page if we just return
5298 * an error where there's an empty slot with no huge pagecache
5299 * to back it. This way, we avoid allocating a hugepage, and
5300 * the sparse dumpfile avoids allocating disk blocks, but its
5301 * huge holes still show up with zeroes where they need to be.
5303 if (absent && (flags & FOLL_DUMP) &&
5304 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
5312 * We need call hugetlb_fault for both hugepages under migration
5313 * (in which case hugetlb_fault waits for the migration,) and
5314 * hwpoisoned hugepages (in which case we need to prevent the
5315 * caller from accessing to them.) In order to do this, we use
5316 * here is_swap_pte instead of is_hugetlb_entry_migration and
5317 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
5318 * both cases, and because we can't follow correct pages
5319 * directly from any kind of swap entries.
5321 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
5322 ((flags & FOLL_WRITE) &&
5323 !huge_pte_write(huge_ptep_get(pte)))) {
5325 unsigned int fault_flags = 0;
5329 if (flags & FOLL_WRITE)
5330 fault_flags |= FAULT_FLAG_WRITE;
5332 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5333 FAULT_FLAG_KILLABLE;
5334 if (flags & FOLL_NOWAIT)
5335 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
5336 FAULT_FLAG_RETRY_NOWAIT;
5337 if (flags & FOLL_TRIED) {
5339 * Note: FAULT_FLAG_ALLOW_RETRY and
5340 * FAULT_FLAG_TRIED can co-exist
5342 fault_flags |= FAULT_FLAG_TRIED;
5344 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
5345 if (ret & VM_FAULT_ERROR) {
5346 err = vm_fault_to_errno(ret, flags);
5350 if (ret & VM_FAULT_RETRY) {
5352 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
5356 * VM_FAULT_RETRY must not return an
5357 * error, it will return zero
5360 * No need to update "position" as the
5361 * caller will not check it after
5362 * *nr_pages is set to 0.
5369 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
5370 page = pte_page(huge_ptep_get(pte));
5373 * If subpage information not requested, update counters
5374 * and skip the same_page loop below.
5376 if (!pages && !vmas && !pfn_offset &&
5377 (vaddr + huge_page_size(h) < vma->vm_end) &&
5378 (remainder >= pages_per_huge_page(h))) {
5379 vaddr += huge_page_size(h);
5380 remainder -= pages_per_huge_page(h);
5381 i += pages_per_huge_page(h);
5386 refs = min3(pages_per_huge_page(h) - pfn_offset,
5387 (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
5390 record_subpages_vmas(mem_map_offset(page, pfn_offset),
5392 likely(pages) ? pages + i : NULL,
5393 vmas ? vmas + i : NULL);
5397 * try_grab_compound_head() should always succeed here,
5398 * because: a) we hold the ptl lock, and b) we've just
5399 * checked that the huge page is present in the page
5400 * tables. If the huge page is present, then the tail
5401 * pages must also be present. The ptl prevents the
5402 * head page and tail pages from being rearranged in
5403 * any way. So this page must be available at this
5404 * point, unless the page refcount overflowed:
5406 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
5416 vaddr += (refs << PAGE_SHIFT);
5422 *nr_pages = remainder;
5424 * setting position is actually required only if remainder is
5425 * not zero but it's faster not to add a "if (remainder)"
5433 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
5434 unsigned long address, unsigned long end, pgprot_t newprot)
5436 struct mm_struct *mm = vma->vm_mm;
5437 unsigned long start = address;
5440 struct hstate *h = hstate_vma(vma);
5441 unsigned long pages = 0;
5442 bool shared_pmd = false;
5443 struct mmu_notifier_range range;
5446 * In the case of shared PMDs, the area to flush could be beyond
5447 * start/end. Set range.start/range.end to cover the maximum possible
5448 * range if PMD sharing is possible.
5450 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5451 0, vma, mm, start, end);
5452 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5454 BUG_ON(address >= end);
5455 flush_cache_range(vma, range.start, range.end);
5457 mmu_notifier_invalidate_range_start(&range);
5458 i_mmap_lock_write(vma->vm_file->f_mapping);
5459 for (; address < end; address += huge_page_size(h)) {
5461 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5464 ptl = huge_pte_lock(h, mm, ptep);
5465 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
5471 pte = huge_ptep_get(ptep);
5472 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5476 if (unlikely(is_hugetlb_entry_migration(pte))) {
5477 swp_entry_t entry = pte_to_swp_entry(pte);
5479 if (is_write_migration_entry(entry)) {
5482 make_migration_entry_read(&entry);
5483 newpte = swp_entry_to_pte(entry);
5484 set_huge_swap_pte_at(mm, address, ptep,
5485 newpte, huge_page_size(h));
5491 if (!huge_pte_none(pte)) {
5493 unsigned int shift = huge_page_shift(hstate_vma(vma));
5495 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5496 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
5497 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
5498 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
5504 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
5505 * may have cleared our pud entry and done put_page on the page table:
5506 * once we release i_mmap_rwsem, another task can do the final put_page
5507 * and that page table be reused and filled with junk. If we actually
5508 * did unshare a page of pmds, flush the range corresponding to the pud.
5511 flush_hugetlb_tlb_range(vma, range.start, range.end);
5513 flush_hugetlb_tlb_range(vma, start, end);
5515 * No need to call mmu_notifier_invalidate_range() we are downgrading
5516 * page table protection not changing it to point to a new page.
5518 * See Documentation/vm/mmu_notifier.rst
5520 i_mmap_unlock_write(vma->vm_file->f_mapping);
5521 mmu_notifier_invalidate_range_end(&range);
5523 return pages << h->order;
5526 /* Return true if reservation was successful, false otherwise. */
5527 bool hugetlb_reserve_pages(struct inode *inode,
5529 struct vm_area_struct *vma,
5530 vm_flags_t vm_flags)
5533 struct hstate *h = hstate_inode(inode);
5534 struct hugepage_subpool *spool = subpool_inode(inode);
5535 struct resv_map *resv_map;
5536 struct hugetlb_cgroup *h_cg = NULL;
5537 long gbl_reserve, regions_needed = 0;
5539 /* This should never happen */
5541 VM_WARN(1, "%s called with a negative range\n", __func__);
5546 * Only apply hugepage reservation if asked. At fault time, an
5547 * attempt will be made for VM_NORESERVE to allocate a page
5548 * without using reserves
5550 if (vm_flags & VM_NORESERVE)
5554 * Shared mappings base their reservation on the number of pages that
5555 * are already allocated on behalf of the file. Private mappings need
5556 * to reserve the full area even if read-only as mprotect() may be
5557 * called to make the mapping read-write. Assume !vma is a shm mapping
5559 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5561 * resv_map can not be NULL as hugetlb_reserve_pages is only
5562 * called for inodes for which resv_maps were created (see
5563 * hugetlbfs_get_inode).
5565 resv_map = inode_resv_map(inode);
5567 chg = region_chg(resv_map, from, to, ®ions_needed);
5570 /* Private mapping. */
5571 resv_map = resv_map_alloc();
5577 set_vma_resv_map(vma, resv_map);
5578 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5584 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5585 chg * pages_per_huge_page(h), &h_cg) < 0)
5588 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5589 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5592 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5596 * There must be enough pages in the subpool for the mapping. If
5597 * the subpool has a minimum size, there may be some global
5598 * reservations already in place (gbl_reserve).
5600 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5601 if (gbl_reserve < 0)
5602 goto out_uncharge_cgroup;
5605 * Check enough hugepages are available for the reservation.
5606 * Hand the pages back to the subpool if there are not
5608 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
5612 * Account for the reservations made. Shared mappings record regions
5613 * that have reservations as they are shared by multiple VMAs.
5614 * When the last VMA disappears, the region map says how much
5615 * the reservation was and the page cache tells how much of
5616 * the reservation was consumed. Private mappings are per-VMA and
5617 * only the consumed reservations are tracked. When the VMA
5618 * disappears, the original reservation is the VMA size and the
5619 * consumed reservations are stored in the map. Hence, nothing
5620 * else has to be done for private mappings here
5622 if (!vma || vma->vm_flags & VM_MAYSHARE) {
5623 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
5625 if (unlikely(add < 0)) {
5626 hugetlb_acct_memory(h, -gbl_reserve);
5628 } else if (unlikely(chg > add)) {
5630 * pages in this range were added to the reserve
5631 * map between region_chg and region_add. This
5632 * indicates a race with alloc_huge_page. Adjust
5633 * the subpool and reserve counts modified above
5634 * based on the difference.
5639 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5640 * reference to h_cg->css. See comment below for detail.
5642 hugetlb_cgroup_uncharge_cgroup_rsvd(
5644 (chg - add) * pages_per_huge_page(h), h_cg);
5646 rsv_adjust = hugepage_subpool_put_pages(spool,
5648 hugetlb_acct_memory(h, -rsv_adjust);
5651 * The file_regions will hold their own reference to
5652 * h_cg->css. So we should release the reference held
5653 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5656 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
5662 /* put back original number of pages, chg */
5663 (void)hugepage_subpool_put_pages(spool, chg);
5664 out_uncharge_cgroup:
5665 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5666 chg * pages_per_huge_page(h), h_cg);
5668 if (!vma || vma->vm_flags & VM_MAYSHARE)
5669 /* Only call region_abort if the region_chg succeeded but the
5670 * region_add failed or didn't run.
5672 if (chg >= 0 && add < 0)
5673 region_abort(resv_map, from, to, regions_needed);
5674 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5675 kref_put(&resv_map->refs, resv_map_release);
5679 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5682 struct hstate *h = hstate_inode(inode);
5683 struct resv_map *resv_map = inode_resv_map(inode);
5685 struct hugepage_subpool *spool = subpool_inode(inode);
5689 * Since this routine can be called in the evict inode path for all
5690 * hugetlbfs inodes, resv_map could be NULL.
5693 chg = region_del(resv_map, start, end);
5695 * region_del() can fail in the rare case where a region
5696 * must be split and another region descriptor can not be
5697 * allocated. If end == LONG_MAX, it will not fail.
5703 spin_lock(&inode->i_lock);
5704 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
5705 spin_unlock(&inode->i_lock);
5708 * If the subpool has a minimum size, the number of global
5709 * reservations to be released may be adjusted.
5711 * Note that !resv_map implies freed == 0. So (chg - freed)
5712 * won't go negative.
5714 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5715 hugetlb_acct_memory(h, -gbl_reserve);
5720 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5721 static unsigned long page_table_shareable(struct vm_area_struct *svma,
5722 struct vm_area_struct *vma,
5723 unsigned long addr, pgoff_t idx)
5725 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5727 unsigned long sbase = saddr & PUD_MASK;
5728 unsigned long s_end = sbase + PUD_SIZE;
5730 /* Allow segments to share if only one is marked locked */
5731 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5732 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
5735 * match the virtual addresses, permission and the alignment of the
5738 if (pmd_index(addr) != pmd_index(saddr) ||
5739 vm_flags != svm_flags ||
5740 !range_in_vma(svma, sbase, s_end))
5746 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
5748 unsigned long base = addr & PUD_MASK;
5749 unsigned long end = base + PUD_SIZE;
5752 * check on proper vm_flags and page table alignment
5754 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
5759 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5761 #ifdef CONFIG_USERFAULTFD
5762 if (uffd_disable_huge_pmd_share(vma))
5765 return vma_shareable(vma, addr);
5769 * Determine if start,end range within vma could be mapped by shared pmd.
5770 * If yes, adjust start and end to cover range associated with possible
5771 * shared pmd mappings.
5773 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5774 unsigned long *start, unsigned long *end)
5776 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5777 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5780 * vma needs to span at least one aligned PUD size, and the range
5781 * must be at least partially within in.
5783 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5784 (*end <= v_start) || (*start >= v_end))
5787 /* Extend the range to be PUD aligned for a worst case scenario */
5788 if (*start > v_start)
5789 *start = ALIGN_DOWN(*start, PUD_SIZE);
5792 *end = ALIGN(*end, PUD_SIZE);
5796 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5797 * and returns the corresponding pte. While this is not necessary for the
5798 * !shared pmd case because we can allocate the pmd later as well, it makes the
5799 * code much cleaner.
5801 * This routine must be called with i_mmap_rwsem held in at least read mode if
5802 * sharing is possible. For hugetlbfs, this prevents removal of any page
5803 * table entries associated with the address space. This is important as we
5804 * are setting up sharing based on existing page table entries (mappings).
5806 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5807 * huge_pte_alloc know that sharing is not possible and do not take
5808 * i_mmap_rwsem as a performance optimization. This is handled by the
5809 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5810 * only required for subsequent processing.
5812 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5813 unsigned long addr, pud_t *pud)
5815 struct address_space *mapping = vma->vm_file->f_mapping;
5816 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5818 struct vm_area_struct *svma;
5819 unsigned long saddr;
5824 i_mmap_assert_locked(mapping);
5825 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5829 saddr = page_table_shareable(svma, vma, addr, idx);
5831 spte = huge_pte_offset(svma->vm_mm, saddr,
5832 vma_mmu_pagesize(svma));
5834 get_page(virt_to_page(spte));
5843 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
5844 if (pud_none(*pud)) {
5845 pud_populate(mm, pud,
5846 (pmd_t *)((unsigned long)spte & PAGE_MASK));
5849 put_page(virt_to_page(spte));
5853 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5858 * unmap huge page backed by shared pte.
5860 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5861 * indicated by page_count > 1, unmap is achieved by clearing pud and
5862 * decrementing the ref count. If count == 1, the pte page is not shared.
5864 * Called with page table lock held and i_mmap_rwsem held in write mode.
5866 * returns: 1 successfully unmapped a shared pte page
5867 * 0 the underlying pte page is not shared, or it is the last user
5869 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5870 unsigned long *addr, pte_t *ptep)
5872 pgd_t *pgd = pgd_offset(mm, *addr);
5873 p4d_t *p4d = p4d_offset(pgd, *addr);
5874 pud_t *pud = pud_offset(p4d, *addr);
5876 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
5877 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5878 if (page_count(virt_to_page(ptep)) == 1)
5882 put_page(virt_to_page(ptep));
5884 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5888 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5889 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5890 unsigned long addr, pud_t *pud)
5895 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5896 unsigned long *addr, pte_t *ptep)
5901 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5902 unsigned long *start, unsigned long *end)
5906 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5910 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5912 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5913 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
5914 unsigned long addr, unsigned long sz)
5921 pgd = pgd_offset(mm, addr);
5922 p4d = p4d_alloc(mm, pgd, addr);
5925 pud = pud_alloc(mm, p4d, addr);
5927 if (sz == PUD_SIZE) {
5930 BUG_ON(sz != PMD_SIZE);
5931 if (want_pmd_share(vma, addr) && pud_none(*pud))
5932 pte = huge_pmd_share(mm, vma, addr, pud);
5934 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5937 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
5943 * huge_pte_offset() - Walk the page table to resolve the hugepage
5944 * entry at address @addr
5946 * Return: Pointer to page table entry (PUD or PMD) for
5947 * address @addr, or NULL if a !p*d_present() entry is encountered and the
5948 * size @sz doesn't match the hugepage size at this level of the page
5951 pte_t *huge_pte_offset(struct mm_struct *mm,
5952 unsigned long addr, unsigned long sz)
5959 pgd = pgd_offset(mm, addr);
5960 if (!pgd_present(*pgd))
5962 p4d = p4d_offset(pgd, addr);
5963 if (!p4d_present(*p4d))
5966 pud = pud_offset(p4d, addr);
5968 /* must be pud huge, non-present or none */
5969 return (pte_t *)pud;
5970 if (!pud_present(*pud))
5972 /* must have a valid entry and size to go further */
5974 pmd = pmd_offset(pud, addr);
5975 /* must be pmd huge, non-present or none */
5976 return (pte_t *)pmd;
5979 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5982 * These functions are overwritable if your architecture needs its own
5985 struct page * __weak
5986 follow_huge_addr(struct mm_struct *mm, unsigned long address,
5989 return ERR_PTR(-EINVAL);
5992 struct page * __weak
5993 follow_huge_pd(struct vm_area_struct *vma,
5994 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5996 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
6000 struct page * __weak
6001 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
6002 pmd_t *pmd, int flags)
6004 struct page *page = NULL;
6008 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
6009 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
6010 (FOLL_PIN | FOLL_GET)))
6014 ptl = pmd_lockptr(mm, pmd);
6017 * make sure that the address range covered by this pmd is not
6018 * unmapped from other threads.
6020 if (!pmd_huge(*pmd))
6022 pte = huge_ptep_get((pte_t *)pmd);
6023 if (pte_present(pte)) {
6024 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
6026 * try_grab_page() should always succeed here, because: a) we
6027 * hold the pmd (ptl) lock, and b) we've just checked that the
6028 * huge pmd (head) page is present in the page tables. The ptl
6029 * prevents the head page and tail pages from being rearranged
6030 * in any way. So this page must be available at this point,
6031 * unless the page refcount overflowed:
6033 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
6038 if (is_hugetlb_entry_migration(pte)) {
6040 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
6044 * hwpoisoned entry is treated as no_page_table in
6045 * follow_page_mask().
6053 struct page * __weak
6054 follow_huge_pud(struct mm_struct *mm, unsigned long address,
6055 pud_t *pud, int flags)
6057 if (flags & (FOLL_GET | FOLL_PIN))
6060 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
6063 struct page * __weak
6064 follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
6066 if (flags & (FOLL_GET | FOLL_PIN))
6069 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
6072 bool isolate_huge_page(struct page *page, struct list_head *list)
6076 spin_lock_irq(&hugetlb_lock);
6077 if (!PageHeadHuge(page) ||
6078 !HPageMigratable(page) ||
6079 !get_page_unless_zero(page)) {
6083 ClearHPageMigratable(page);
6084 list_move_tail(&page->lru, list);
6086 spin_unlock_irq(&hugetlb_lock);
6090 int get_hwpoison_huge_page(struct page *page, bool *hugetlb)
6095 spin_lock_irq(&hugetlb_lock);
6096 if (PageHeadHuge(page)) {
6098 if (HPageFreed(page) || HPageMigratable(page))
6099 ret = get_page_unless_zero(page);
6103 spin_unlock_irq(&hugetlb_lock);
6107 void putback_active_hugepage(struct page *page)
6109 spin_lock_irq(&hugetlb_lock);
6110 SetHPageMigratable(page);
6111 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
6112 spin_unlock_irq(&hugetlb_lock);
6116 void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
6118 struct hstate *h = page_hstate(oldpage);
6120 hugetlb_cgroup_migrate(oldpage, newpage);
6121 set_page_owner_migrate_reason(newpage, reason);
6124 * transfer temporary state of the new huge page. This is
6125 * reverse to other transitions because the newpage is going to
6126 * be final while the old one will be freed so it takes over
6127 * the temporary status.
6129 * Also note that we have to transfer the per-node surplus state
6130 * here as well otherwise the global surplus count will not match
6133 if (HPageTemporary(newpage)) {
6134 int old_nid = page_to_nid(oldpage);
6135 int new_nid = page_to_nid(newpage);
6137 SetHPageTemporary(oldpage);
6138 ClearHPageTemporary(newpage);
6141 * There is no need to transfer the per-node surplus state
6142 * when we do not cross the node.
6144 if (new_nid == old_nid)
6146 spin_lock_irq(&hugetlb_lock);
6147 if (h->surplus_huge_pages_node[old_nid]) {
6148 h->surplus_huge_pages_node[old_nid]--;
6149 h->surplus_huge_pages_node[new_nid]++;
6151 spin_unlock_irq(&hugetlb_lock);
6156 * This function will unconditionally remove all the shared pmd pgtable entries
6157 * within the specific vma for a hugetlbfs memory range.
6159 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
6161 struct hstate *h = hstate_vma(vma);
6162 unsigned long sz = huge_page_size(h);
6163 struct mm_struct *mm = vma->vm_mm;
6164 struct mmu_notifier_range range;
6165 unsigned long address, start, end;
6169 if (!(vma->vm_flags & VM_MAYSHARE))
6172 start = ALIGN(vma->vm_start, PUD_SIZE);
6173 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6179 * No need to call adjust_range_if_pmd_sharing_possible(), because
6180 * we have already done the PUD_SIZE alignment.
6182 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
6184 mmu_notifier_invalidate_range_start(&range);
6185 i_mmap_lock_write(vma->vm_file->f_mapping);
6186 for (address = start; address < end; address += PUD_SIZE) {
6187 unsigned long tmp = address;
6189 ptep = huge_pte_offset(mm, address, sz);
6192 ptl = huge_pte_lock(h, mm, ptep);
6193 /* We don't want 'address' to be changed */
6194 huge_pmd_unshare(mm, vma, &tmp, ptep);
6197 flush_hugetlb_tlb_range(vma, start, end);
6198 i_mmap_unlock_write(vma->vm_file->f_mapping);
6200 * No need to call mmu_notifier_invalidate_range(), see
6201 * Documentation/vm/mmu_notifier.rst.
6203 mmu_notifier_invalidate_range_end(&range);
6207 static bool cma_reserve_called __initdata;
6209 static int __init cmdline_parse_hugetlb_cma(char *p)
6211 hugetlb_cma_size = memparse(p, &p);
6215 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
6217 void __init hugetlb_cma_reserve(int order)
6219 unsigned long size, reserved, per_node;
6222 cma_reserve_called = true;
6224 if (!hugetlb_cma_size)
6227 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
6228 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
6229 (PAGE_SIZE << order) / SZ_1M);
6234 * If 3 GB area is requested on a machine with 4 numa nodes,
6235 * let's allocate 1 GB on first three nodes and ignore the last one.
6237 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
6238 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
6239 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
6242 for_each_node_state(nid, N_ONLINE) {
6244 char name[CMA_MAX_NAME];
6246 size = min(per_node, hugetlb_cma_size - reserved);
6247 size = round_up(size, PAGE_SIZE << order);
6249 snprintf(name, sizeof(name), "hugetlb%d", nid);
6250 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
6252 &hugetlb_cma[nid], nid);
6254 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
6260 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
6263 if (reserved >= hugetlb_cma_size)
6268 void __init hugetlb_cma_check(void)
6270 if (!hugetlb_cma_size || cma_reserve_called)
6273 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
6276 #endif /* CONFIG_CMA */