2 * hugetlbpage-backed filesystem. Based on ramfs.
4 * Nadia Yvette Chambers, 2002
6 * Copyright (C) 2002 Linus Torvalds.
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 #include <linux/thread_info.h>
13 #include <asm/current.h>
14 #include <linux/sched/signal.h> /* remove ASAP */
15 #include <linux/falloc.h>
17 #include <linux/mount.h>
18 #include <linux/file.h>
19 #include <linux/kernel.h>
20 #include <linux/writeback.h>
21 #include <linux/pagemap.h>
22 #include <linux/highmem.h>
23 #include <linux/init.h>
24 #include <linux/string.h>
25 #include <linux/capability.h>
26 #include <linux/ctype.h>
27 #include <linux/backing-dev.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagevec.h>
30 #include <linux/fs_parser.h>
31 #include <linux/mman.h>
32 #include <linux/slab.h>
33 #include <linux/dnotify.h>
34 #include <linux/statfs.h>
35 #include <linux/security.h>
36 #include <linux/magic.h>
37 #include <linux/migrate.h>
38 #include <linux/uio.h>
40 #include <linux/uaccess.h>
42 static const struct super_operations hugetlbfs_ops;
43 static const struct address_space_operations hugetlbfs_aops;
44 const struct file_operations hugetlbfs_file_operations;
45 static const struct inode_operations hugetlbfs_dir_inode_operations;
46 static const struct inode_operations hugetlbfs_inode_operations;
48 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
50 struct hugetlbfs_fs_context {
51 struct hstate *hstate;
52 unsigned long long max_size_opt;
53 unsigned long long min_size_opt;
57 enum hugetlbfs_size_type max_val_type;
58 enum hugetlbfs_size_type min_val_type;
64 int sysctl_hugetlb_shm_group;
76 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
77 fsparam_u32 ("gid", Opt_gid),
78 fsparam_string("min_size", Opt_min_size),
79 fsparam_u32 ("mode", Opt_mode),
80 fsparam_string("nr_inodes", Opt_nr_inodes),
81 fsparam_string("pagesize", Opt_pagesize),
82 fsparam_string("size", Opt_size),
83 fsparam_u32 ("uid", Opt_uid),
88 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
89 struct inode *inode, pgoff_t index)
91 vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
95 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
97 mpol_cond_put(vma->vm_policy);
100 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
101 struct inode *inode, pgoff_t index)
105 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
110 static void huge_pagevec_release(struct pagevec *pvec)
114 for (i = 0; i < pagevec_count(pvec); ++i)
115 put_page(pvec->pages[i]);
117 pagevec_reinit(pvec);
121 * Mask used when checking the page offset value passed in via system
122 * calls. This value will be converted to a loff_t which is signed.
123 * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
124 * value. The extra bit (- 1 in the shift value) is to take the sign
127 #define PGOFF_LOFFT_MAX \
128 (((1UL << (PAGE_SHIFT + 1)) - 1) << (BITS_PER_LONG - (PAGE_SHIFT + 1)))
130 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
132 struct inode *inode = file_inode(file);
135 struct hstate *h = hstate_file(file);
138 * vma address alignment (but not the pgoff alignment) has
139 * already been checked by prepare_hugepage_range. If you add
140 * any error returns here, do so after setting VM_HUGETLB, so
141 * is_vm_hugetlb_page tests below unmap_region go the right
142 * way when do_mmap_pgoff unwinds (may be important on powerpc
145 vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
146 vma->vm_ops = &hugetlb_vm_ops;
149 * page based offset in vm_pgoff could be sufficiently large to
150 * overflow a loff_t when converted to byte offset. This can
151 * only happen on architectures where sizeof(loff_t) ==
152 * sizeof(unsigned long). So, only check in those instances.
154 if (sizeof(unsigned long) == sizeof(loff_t)) {
155 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
159 /* must be huge page aligned */
160 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
163 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
164 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
165 /* check for overflow */
173 if (hugetlb_reserve_pages(inode,
174 vma->vm_pgoff >> huge_page_order(h),
175 len >> huge_page_shift(h), vma,
180 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
181 i_size_write(inode, len);
189 * Called under down_write(mmap_sem).
192 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
194 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
195 unsigned long len, unsigned long pgoff, unsigned long flags)
197 struct mm_struct *mm = current->mm;
198 struct vm_area_struct *vma;
199 struct hstate *h = hstate_file(file);
200 struct vm_unmapped_area_info info;
202 if (len & ~huge_page_mask(h))
207 if (flags & MAP_FIXED) {
208 if (prepare_hugepage_range(file, addr, len))
214 addr = ALIGN(addr, huge_page_size(h));
215 vma = find_vma(mm, addr);
216 if (TASK_SIZE - len >= addr &&
217 (!vma || addr + len <= vm_start_gap(vma)))
223 info.low_limit = TASK_UNMAPPED_BASE;
224 info.high_limit = TASK_SIZE;
225 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
226 info.align_offset = 0;
227 return vm_unmapped_area(&info);
232 hugetlbfs_read_actor(struct page *page, unsigned long offset,
233 struct iov_iter *to, unsigned long size)
238 /* Find which 4k chunk and offset with in that chunk */
239 i = offset >> PAGE_SHIFT;
240 offset = offset & ~PAGE_MASK;
244 chunksize = PAGE_SIZE;
247 if (chunksize > size)
249 n = copy_page_to_iter(&page[i], offset, chunksize, to);
261 * Support for read() - Find the page attached to f_mapping and copy out the
262 * data. Its *very* similar to do_generic_mapping_read(), we can't use that
263 * since it has PAGE_SIZE assumptions.
265 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
267 struct file *file = iocb->ki_filp;
268 struct hstate *h = hstate_file(file);
269 struct address_space *mapping = file->f_mapping;
270 struct inode *inode = mapping->host;
271 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
272 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
273 unsigned long end_index;
277 while (iov_iter_count(to)) {
281 /* nr is the maximum number of bytes to copy from this page */
282 nr = huge_page_size(h);
283 isize = i_size_read(inode);
286 end_index = (isize - 1) >> huge_page_shift(h);
287 if (index > end_index)
289 if (index == end_index) {
290 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
297 page = find_lock_page(mapping, index);
298 if (unlikely(page == NULL)) {
300 * We have a HOLE, zero out the user-buffer for the
301 * length of the hole or request.
303 copied = iov_iter_zero(nr, to);
308 * We have the page, copy it to user space buffer.
310 copied = hugetlbfs_read_actor(page, offset, to, nr);
315 if (copied != nr && iov_iter_count(to)) {
320 index += offset >> huge_page_shift(h);
321 offset &= ~huge_page_mask(h);
323 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
327 static int hugetlbfs_write_begin(struct file *file,
328 struct address_space *mapping,
329 loff_t pos, unsigned len, unsigned flags,
330 struct page **pagep, void **fsdata)
335 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
336 loff_t pos, unsigned len, unsigned copied,
337 struct page *page, void *fsdata)
343 static void remove_huge_page(struct page *page)
345 ClearPageDirty(page);
346 ClearPageUptodate(page);
347 delete_from_page_cache(page);
351 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
353 struct vm_area_struct *vma;
356 * end == 0 indicates that the entire range after
357 * start should be unmapped.
359 vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
360 unsigned long v_offset;
364 * Can the expression below overflow on 32-bit arches?
365 * No, because the interval tree returns us only those vmas
366 * which overlap the truncated area starting at pgoff,
367 * and no vma on a 32-bit arch can span beyond the 4GB.
369 if (vma->vm_pgoff < start)
370 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
377 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
379 if (v_end > vma->vm_end)
383 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
389 * remove_inode_hugepages handles two distinct cases: truncation and hole
390 * punch. There are subtle differences in operation for each case.
392 * truncation is indicated by end of range being LLONG_MAX
393 * In this case, we first scan the range and release found pages.
394 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
395 * maps and global counts. Page faults can not race with truncation
396 * in this routine. hugetlb_no_page() holds i_mmap_rwsem and prevents
397 * page faults in the truncated range by checking i_size. i_size is
398 * modified while holding i_mmap_rwsem.
399 * hole punch is indicated if end is not LLONG_MAX
400 * In the hole punch case we scan the range and release found pages.
401 * Only when releasing a page is the associated region/reserv map
402 * deleted. The region/reserv map for ranges without associated
403 * pages are not modified. Page faults can race with hole punch.
404 * This is indicated if we find a mapped page.
405 * Note: If the passed end of range value is beyond the end of file, but
406 * not LLONG_MAX this routine still performs a hole punch operation.
408 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
411 struct hstate *h = hstate_inode(inode);
412 struct address_space *mapping = &inode->i_data;
413 const pgoff_t start = lstart >> huge_page_shift(h);
414 const pgoff_t end = lend >> huge_page_shift(h);
415 struct vm_area_struct pseudo_vma;
419 bool truncate_op = (lend == LLONG_MAX);
421 vma_init(&pseudo_vma, current->mm);
422 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
427 * When no more pages are found, we are done.
429 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
432 for (i = 0; i < pagevec_count(&pvec); ++i) {
433 struct page *page = pvec.pages[i];
437 hash = hugetlb_fault_mutex_hash(mapping, index);
440 * Only need to hold the fault mutex in the
441 * hole punch case. This prevents races with
442 * page faults. Races are not possible in the
443 * case of truncation.
445 mutex_lock(&hugetlb_fault_mutex_table[hash]);
449 * If page is mapped, it was faulted in after being
450 * unmapped in caller. Unmap (again) now after taking
451 * the fault mutex. The mutex will prevent faults
452 * until we finish removing the page.
454 * This race can only happen in the hole punch case.
455 * Getting here in a truncate operation is a bug.
457 if (unlikely(page_mapped(page))) {
460 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
461 i_mmap_lock_write(mapping);
462 mutex_lock(&hugetlb_fault_mutex_table[hash]);
463 hugetlb_vmdelete_list(&mapping->i_mmap,
464 index * pages_per_huge_page(h),
465 (index + 1) * pages_per_huge_page(h));
466 i_mmap_unlock_write(mapping);
471 * We must free the huge page and remove from page
472 * cache (remove_huge_page) BEFORE removing the
473 * region/reserve map (hugetlb_unreserve_pages). In
474 * rare out of memory conditions, removal of the
475 * region/reserve map could fail. Correspondingly,
476 * the subpool and global reserve usage count can need
479 VM_BUG_ON(PagePrivate(page));
480 remove_huge_page(page);
483 if (unlikely(hugetlb_unreserve_pages(inode,
484 index, index + 1, 1)))
485 hugetlb_fix_reserve_counts(inode);
490 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
492 huge_pagevec_release(&pvec);
497 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
500 static void hugetlbfs_evict_inode(struct inode *inode)
502 struct resv_map *resv_map;
504 remove_inode_hugepages(inode, 0, LLONG_MAX);
507 * Get the resv_map from the address space embedded in the inode.
508 * This is the address space which points to any resv_map allocated
509 * at inode creation time. If this is a device special inode,
510 * i_mapping may not point to the original address space.
512 resv_map = (struct resv_map *)(&inode->i_data)->private_data;
513 /* Only regular and link inodes have associated reserve maps */
515 resv_map_release(&resv_map->refs);
519 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
522 struct address_space *mapping = inode->i_mapping;
523 struct hstate *h = hstate_inode(inode);
525 BUG_ON(offset & ~huge_page_mask(h));
526 pgoff = offset >> PAGE_SHIFT;
528 i_mmap_lock_write(mapping);
529 i_size_write(inode, offset);
530 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
531 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
532 i_mmap_unlock_write(mapping);
533 remove_inode_hugepages(inode, offset, LLONG_MAX);
537 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
539 struct hstate *h = hstate_inode(inode);
540 loff_t hpage_size = huge_page_size(h);
541 loff_t hole_start, hole_end;
544 * For hole punch round up the beginning offset of the hole and
545 * round down the end.
547 hole_start = round_up(offset, hpage_size);
548 hole_end = round_down(offset + len, hpage_size);
550 if (hole_end > hole_start) {
551 struct address_space *mapping = inode->i_mapping;
552 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
556 /* protected by i_mutex */
557 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
562 i_mmap_lock_write(mapping);
563 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
564 hugetlb_vmdelete_list(&mapping->i_mmap,
565 hole_start >> PAGE_SHIFT,
566 hole_end >> PAGE_SHIFT);
567 i_mmap_unlock_write(mapping);
568 remove_inode_hugepages(inode, hole_start, hole_end);
575 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
578 struct inode *inode = file_inode(file);
579 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
580 struct address_space *mapping = inode->i_mapping;
581 struct hstate *h = hstate_inode(inode);
582 struct vm_area_struct pseudo_vma;
583 struct mm_struct *mm = current->mm;
584 loff_t hpage_size = huge_page_size(h);
585 unsigned long hpage_shift = huge_page_shift(h);
586 pgoff_t start, index, end;
590 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
593 if (mode & FALLOC_FL_PUNCH_HOLE)
594 return hugetlbfs_punch_hole(inode, offset, len);
597 * Default preallocate case.
598 * For this range, start is rounded down and end is rounded up
599 * as well as being converted to page offsets.
601 start = offset >> hpage_shift;
602 end = (offset + len + hpage_size - 1) >> hpage_shift;
606 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
607 error = inode_newsize_ok(inode, offset + len);
611 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
617 * Initialize a pseudo vma as this is required by the huge page
618 * allocation routines. If NUMA is configured, use page index
619 * as input to create an allocation policy.
621 vma_init(&pseudo_vma, mm);
622 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
623 pseudo_vma.vm_file = file;
625 for (index = start; index < end; index++) {
627 * This is supposed to be the vaddr where the page is being
628 * faulted in, but we have no vaddr here.
632 int avoid_reserve = 0;
637 * fallocate(2) manpage permits EINTR; we may have been
638 * interrupted because we are using up too much memory.
640 if (signal_pending(current)) {
645 /* Set numa allocation policy based on index */
646 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
648 /* addr is the offset within the file (zero based) */
649 addr = index * hpage_size;
652 * fault mutex taken here, protects against fault path
653 * and hole punch. inode_lock previously taken protects
654 * against truncation.
656 hash = hugetlb_fault_mutex_hash(mapping, index);
657 mutex_lock(&hugetlb_fault_mutex_table[hash]);
659 /* See if already present in mapping to avoid alloc/free */
660 page = find_get_page(mapping, index);
663 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
664 hugetlb_drop_vma_policy(&pseudo_vma);
668 /* Allocate page and add to page cache */
669 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
670 hugetlb_drop_vma_policy(&pseudo_vma);
672 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
673 error = PTR_ERR(page);
676 clear_huge_page(page, addr, pages_per_huge_page(h));
677 __SetPageUptodate(page);
678 error = huge_add_to_page_cache(page, mapping, index);
679 if (unlikely(error)) {
681 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
685 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
688 * unlock_page because locked by add_to_page_cache()
689 * page_put due to reference from alloc_huge_page()
695 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
696 i_size_write(inode, offset + len);
697 inode->i_ctime = current_time(inode);
703 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
705 struct inode *inode = d_inode(dentry);
706 struct hstate *h = hstate_inode(inode);
708 unsigned int ia_valid = attr->ia_valid;
709 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
713 error = setattr_prepare(dentry, attr);
717 if (ia_valid & ATTR_SIZE) {
718 loff_t oldsize = inode->i_size;
719 loff_t newsize = attr->ia_size;
721 if (newsize & ~huge_page_mask(h))
723 /* protected by i_mutex */
724 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
725 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
727 error = hugetlb_vmtruncate(inode, newsize);
732 setattr_copy(inode, attr);
733 mark_inode_dirty(inode);
737 static struct inode *hugetlbfs_get_root(struct super_block *sb,
738 struct hugetlbfs_fs_context *ctx)
742 inode = new_inode(sb);
744 inode->i_ino = get_next_ino();
745 inode->i_mode = S_IFDIR | ctx->mode;
746 inode->i_uid = ctx->uid;
747 inode->i_gid = ctx->gid;
748 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
749 inode->i_op = &hugetlbfs_dir_inode_operations;
750 inode->i_fop = &simple_dir_operations;
751 /* directory inodes start off with i_nlink == 2 (for "." entry) */
753 lockdep_annotate_inode_mutex_key(inode);
759 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
760 * be taken from reclaim -- unlike regular filesystems. This needs an
761 * annotation because huge_pmd_share() does an allocation under hugetlb's
764 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
766 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
768 umode_t mode, dev_t dev)
771 struct resv_map *resv_map = NULL;
774 * Reserve maps are only needed for inodes that can have associated
777 if (S_ISREG(mode) || S_ISLNK(mode)) {
778 resv_map = resv_map_alloc();
783 inode = new_inode(sb);
785 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
787 inode->i_ino = get_next_ino();
788 inode_init_owner(inode, dir, mode);
789 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
790 &hugetlbfs_i_mmap_rwsem_key);
791 inode->i_mapping->a_ops = &hugetlbfs_aops;
792 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
793 inode->i_mapping->private_data = resv_map;
794 info->seals = F_SEAL_SEAL;
795 switch (mode & S_IFMT) {
797 init_special_inode(inode, mode, dev);
800 inode->i_op = &hugetlbfs_inode_operations;
801 inode->i_fop = &hugetlbfs_file_operations;
804 inode->i_op = &hugetlbfs_dir_inode_operations;
805 inode->i_fop = &simple_dir_operations;
807 /* directory inodes start off with i_nlink == 2 (for "." entry) */
811 inode->i_op = &page_symlink_inode_operations;
812 inode_nohighmem(inode);
815 lockdep_annotate_inode_mutex_key(inode);
818 kref_put(&resv_map->refs, resv_map_release);
825 * File creation. Allocate an inode, and we're done..
827 static int do_hugetlbfs_mknod(struct inode *dir,
828 struct dentry *dentry,
836 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
838 dir->i_ctime = dir->i_mtime = current_time(dir);
840 d_tmpfile(dentry, inode);
842 d_instantiate(dentry, inode);
843 dget(dentry);/* Extra count - pin the dentry in core */
850 static int hugetlbfs_mknod(struct inode *dir,
851 struct dentry *dentry, umode_t mode, dev_t dev)
853 return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
856 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
858 int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
864 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
866 return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
869 static int hugetlbfs_tmpfile(struct inode *dir,
870 struct dentry *dentry, umode_t mode)
872 return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
875 static int hugetlbfs_symlink(struct inode *dir,
876 struct dentry *dentry, const char *symname)
881 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
883 int l = strlen(symname)+1;
884 error = page_symlink(inode, symname, l);
886 d_instantiate(dentry, inode);
891 dir->i_ctime = dir->i_mtime = current_time(dir);
897 * mark the head page dirty
899 static int hugetlbfs_set_page_dirty(struct page *page)
901 struct page *head = compound_head(page);
907 static int hugetlbfs_migrate_page(struct address_space *mapping,
908 struct page *newpage, struct page *page,
909 enum migrate_mode mode)
913 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
914 if (rc != MIGRATEPAGE_SUCCESS)
918 * page_private is subpool pointer in hugetlb pages. Transfer to
919 * new page. PagePrivate is not associated with page_private for
920 * hugetlb pages and can not be set here as only page_huge_active
921 * pages can be migrated.
923 if (page_private(page)) {
924 set_page_private(newpage, page_private(page));
925 set_page_private(page, 0);
928 if (mode != MIGRATE_SYNC_NO_COPY)
929 migrate_page_copy(newpage, page);
931 migrate_page_states(newpage, page);
933 return MIGRATEPAGE_SUCCESS;
936 static int hugetlbfs_error_remove_page(struct address_space *mapping,
939 struct inode *inode = mapping->host;
940 pgoff_t index = page->index;
942 remove_huge_page(page);
943 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
944 hugetlb_fix_reserve_counts(inode);
950 * Display the mount options in /proc/mounts.
952 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
954 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
955 struct hugepage_subpool *spool = sbinfo->spool;
956 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
957 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
960 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
961 seq_printf(m, ",uid=%u",
962 from_kuid_munged(&init_user_ns, sbinfo->uid));
963 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
964 seq_printf(m, ",gid=%u",
965 from_kgid_munged(&init_user_ns, sbinfo->gid));
966 if (sbinfo->mode != 0755)
967 seq_printf(m, ",mode=%o", sbinfo->mode);
968 if (sbinfo->max_inodes != -1)
969 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
973 if (hpage_size >= 1024) {
977 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
979 if (spool->max_hpages != -1)
980 seq_printf(m, ",size=%llu",
981 (unsigned long long)spool->max_hpages << hpage_shift);
982 if (spool->min_hpages != -1)
983 seq_printf(m, ",min_size=%llu",
984 (unsigned long long)spool->min_hpages << hpage_shift);
989 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
991 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
992 struct hstate *h = hstate_inode(d_inode(dentry));
994 buf->f_type = HUGETLBFS_MAGIC;
995 buf->f_bsize = huge_page_size(h);
997 spin_lock(&sbinfo->stat_lock);
998 /* If no limits set, just report 0 for max/free/used
999 * blocks, like simple_statfs() */
1000 if (sbinfo->spool) {
1003 spin_lock(&sbinfo->spool->lock);
1004 buf->f_blocks = sbinfo->spool->max_hpages;
1005 free_pages = sbinfo->spool->max_hpages
1006 - sbinfo->spool->used_hpages;
1007 buf->f_bavail = buf->f_bfree = free_pages;
1008 spin_unlock(&sbinfo->spool->lock);
1009 buf->f_files = sbinfo->max_inodes;
1010 buf->f_ffree = sbinfo->free_inodes;
1012 spin_unlock(&sbinfo->stat_lock);
1014 buf->f_namelen = NAME_MAX;
1018 static void hugetlbfs_put_super(struct super_block *sb)
1020 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1023 sb->s_fs_info = NULL;
1026 hugepage_put_subpool(sbi->spool);
1032 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1034 if (sbinfo->free_inodes >= 0) {
1035 spin_lock(&sbinfo->stat_lock);
1036 if (unlikely(!sbinfo->free_inodes)) {
1037 spin_unlock(&sbinfo->stat_lock);
1040 sbinfo->free_inodes--;
1041 spin_unlock(&sbinfo->stat_lock);
1047 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1049 if (sbinfo->free_inodes >= 0) {
1050 spin_lock(&sbinfo->stat_lock);
1051 sbinfo->free_inodes++;
1052 spin_unlock(&sbinfo->stat_lock);
1057 static struct kmem_cache *hugetlbfs_inode_cachep;
1059 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1061 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1062 struct hugetlbfs_inode_info *p;
1064 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1066 p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1068 hugetlbfs_inc_free_inodes(sbinfo);
1073 * Any time after allocation, hugetlbfs_destroy_inode can be called
1074 * for the inode. mpol_free_shared_policy is unconditionally called
1075 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1076 * in case of a quick call to destroy.
1078 * Note that the policy is initialized even if we are creating a
1079 * private inode. This simplifies hugetlbfs_destroy_inode.
1081 mpol_shared_policy_init(&p->policy, NULL);
1083 return &p->vfs_inode;
1086 static void hugetlbfs_free_inode(struct inode *inode)
1088 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1091 static void hugetlbfs_destroy_inode(struct inode *inode)
1093 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1094 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1097 static const struct address_space_operations hugetlbfs_aops = {
1098 .write_begin = hugetlbfs_write_begin,
1099 .write_end = hugetlbfs_write_end,
1100 .set_page_dirty = hugetlbfs_set_page_dirty,
1101 .migratepage = hugetlbfs_migrate_page,
1102 .error_remove_page = hugetlbfs_error_remove_page,
1106 static void init_once(void *foo)
1108 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1110 inode_init_once(&ei->vfs_inode);
1113 const struct file_operations hugetlbfs_file_operations = {
1114 .read_iter = hugetlbfs_read_iter,
1115 .mmap = hugetlbfs_file_mmap,
1116 .fsync = noop_fsync,
1117 .get_unmapped_area = hugetlb_get_unmapped_area,
1118 .llseek = default_llseek,
1119 .fallocate = hugetlbfs_fallocate,
1122 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1123 .create = hugetlbfs_create,
1124 .lookup = simple_lookup,
1125 .link = simple_link,
1126 .unlink = simple_unlink,
1127 .symlink = hugetlbfs_symlink,
1128 .mkdir = hugetlbfs_mkdir,
1129 .rmdir = simple_rmdir,
1130 .mknod = hugetlbfs_mknod,
1131 .rename = simple_rename,
1132 .setattr = hugetlbfs_setattr,
1133 .tmpfile = hugetlbfs_tmpfile,
1136 static const struct inode_operations hugetlbfs_inode_operations = {
1137 .setattr = hugetlbfs_setattr,
1140 static const struct super_operations hugetlbfs_ops = {
1141 .alloc_inode = hugetlbfs_alloc_inode,
1142 .free_inode = hugetlbfs_free_inode,
1143 .destroy_inode = hugetlbfs_destroy_inode,
1144 .evict_inode = hugetlbfs_evict_inode,
1145 .statfs = hugetlbfs_statfs,
1146 .put_super = hugetlbfs_put_super,
1147 .show_options = hugetlbfs_show_options,
1151 * Convert size option passed from command line to number of huge pages
1152 * in the pool specified by hstate. Size option could be in bytes
1153 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1156 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1157 enum hugetlbfs_size_type val_type)
1159 if (val_type == NO_SIZE)
1162 if (val_type == SIZE_PERCENT) {
1163 size_opt <<= huge_page_shift(h);
1164 size_opt *= h->max_huge_pages;
1165 do_div(size_opt, 100);
1168 size_opt >>= huge_page_shift(h);
1173 * Parse one mount parameter.
1175 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1177 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1178 struct fs_parse_result result;
1183 opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1189 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1190 if (!uid_valid(ctx->uid))
1195 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1196 if (!gid_valid(ctx->gid))
1201 ctx->mode = result.uint_32 & 01777U;
1205 /* memparse() will accept a K/M/G without a digit */
1206 if (!isdigit(param->string[0]))
1208 ctx->max_size_opt = memparse(param->string, &rest);
1209 ctx->max_val_type = SIZE_STD;
1211 ctx->max_val_type = SIZE_PERCENT;
1215 /* memparse() will accept a K/M/G without a digit */
1216 if (!isdigit(param->string[0]))
1218 ctx->nr_inodes = memparse(param->string, &rest);
1222 ps = memparse(param->string, &rest);
1223 ctx->hstate = size_to_hstate(ps);
1225 pr_err("Unsupported page size %lu MB\n", ps >> 20);
1231 /* memparse() will accept a K/M/G without a digit */
1232 if (!isdigit(param->string[0]))
1234 ctx->min_size_opt = memparse(param->string, &rest);
1235 ctx->min_val_type = SIZE_STD;
1237 ctx->min_val_type = SIZE_PERCENT;
1245 return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1246 param->string, param->key);
1250 * Validate the parsed options.
1252 static int hugetlbfs_validate(struct fs_context *fc)
1254 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1257 * Use huge page pool size (in hstate) to convert the size
1258 * options to number of huge pages. If NO_SIZE, -1 is returned.
1260 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1263 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1268 * If max_size was specified, then min_size must be smaller
1270 if (ctx->max_val_type > NO_SIZE &&
1271 ctx->min_hpages > ctx->max_hpages) {
1272 pr_err("Minimum size can not be greater than maximum size\n");
1280 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1282 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1283 struct hugetlbfs_sb_info *sbinfo;
1285 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1288 sb->s_fs_info = sbinfo;
1289 spin_lock_init(&sbinfo->stat_lock);
1290 sbinfo->hstate = ctx->hstate;
1291 sbinfo->max_inodes = ctx->nr_inodes;
1292 sbinfo->free_inodes = ctx->nr_inodes;
1293 sbinfo->spool = NULL;
1294 sbinfo->uid = ctx->uid;
1295 sbinfo->gid = ctx->gid;
1296 sbinfo->mode = ctx->mode;
1299 * Allocate and initialize subpool if maximum or minimum size is
1300 * specified. Any needed reservations (for minimim size) are taken
1301 * taken when the subpool is created.
1303 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1304 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1310 sb->s_maxbytes = MAX_LFS_FILESIZE;
1311 sb->s_blocksize = huge_page_size(ctx->hstate);
1312 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1313 sb->s_magic = HUGETLBFS_MAGIC;
1314 sb->s_op = &hugetlbfs_ops;
1315 sb->s_time_gran = 1;
1316 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1321 kfree(sbinfo->spool);
1326 static int hugetlbfs_get_tree(struct fs_context *fc)
1328 int err = hugetlbfs_validate(fc);
1331 return get_tree_nodev(fc, hugetlbfs_fill_super);
1334 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1336 kfree(fc->fs_private);
1339 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1340 .free = hugetlbfs_fs_context_free,
1341 .parse_param = hugetlbfs_parse_param,
1342 .get_tree = hugetlbfs_get_tree,
1345 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1347 struct hugetlbfs_fs_context *ctx;
1349 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1353 ctx->max_hpages = -1; /* No limit on size by default */
1354 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1355 ctx->uid = current_fsuid();
1356 ctx->gid = current_fsgid();
1358 ctx->hstate = &default_hstate;
1359 ctx->min_hpages = -1; /* No default minimum size */
1360 ctx->max_val_type = NO_SIZE;
1361 ctx->min_val_type = NO_SIZE;
1362 fc->fs_private = ctx;
1363 fc->ops = &hugetlbfs_fs_context_ops;
1367 static struct file_system_type hugetlbfs_fs_type = {
1368 .name = "hugetlbfs",
1369 .init_fs_context = hugetlbfs_init_fs_context,
1370 .parameters = hugetlb_fs_parameters,
1371 .kill_sb = kill_litter_super,
1374 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1376 static int can_do_hugetlb_shm(void)
1379 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1380 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1383 static int get_hstate_idx(int page_size_log)
1385 struct hstate *h = hstate_sizelog(page_size_log);
1393 * Note that size should be aligned to proper hugepage size in caller side,
1394 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1396 struct file *hugetlb_file_setup(const char *name, size_t size,
1397 vm_flags_t acctflag, struct user_struct **user,
1398 int creat_flags, int page_size_log)
1400 struct inode *inode;
1401 struct vfsmount *mnt;
1405 hstate_idx = get_hstate_idx(page_size_log);
1407 return ERR_PTR(-ENODEV);
1410 mnt = hugetlbfs_vfsmount[hstate_idx];
1412 return ERR_PTR(-ENOENT);
1414 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1415 *user = current_user();
1416 if (user_shm_lock(size, *user)) {
1418 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1419 current->comm, current->pid);
1420 task_unlock(current);
1423 return ERR_PTR(-EPERM);
1427 file = ERR_PTR(-ENOSPC);
1428 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1431 if (creat_flags == HUGETLB_SHMFS_INODE)
1432 inode->i_flags |= S_PRIVATE;
1434 inode->i_size = size;
1437 if (hugetlb_reserve_pages(inode, 0,
1438 size >> huge_page_shift(hstate_inode(inode)), NULL,
1440 file = ERR_PTR(-ENOMEM);
1442 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1443 &hugetlbfs_file_operations);
1450 user_shm_unlock(size, *user);
1456 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1458 struct fs_context *fc;
1459 struct vfsmount *mnt;
1461 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1465 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1471 pr_err("Cannot mount internal hugetlbfs for page size %uK",
1472 1U << (h->order + PAGE_SHIFT - 10));
1476 static int __init init_hugetlbfs_fs(void)
1478 struct vfsmount *mnt;
1483 if (!hugepages_supported()) {
1484 pr_info("disabling because there are no supported hugepage sizes\n");
1489 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1490 sizeof(struct hugetlbfs_inode_info),
1491 0, SLAB_ACCOUNT, init_once);
1492 if (hugetlbfs_inode_cachep == NULL)
1495 error = register_filesystem(&hugetlbfs_fs_type);
1499 /* default hstate mount is required */
1500 mnt = mount_one_hugetlbfs(&hstates[default_hstate_idx]);
1502 error = PTR_ERR(mnt);
1505 hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1507 /* other hstates are optional */
1509 for_each_hstate(h) {
1510 if (i == default_hstate_idx) {
1515 mnt = mount_one_hugetlbfs(h);
1517 hugetlbfs_vfsmount[i] = NULL;
1519 hugetlbfs_vfsmount[i] = mnt;
1526 (void)unregister_filesystem(&hugetlbfs_fs_type);
1528 kmem_cache_destroy(hugetlbfs_inode_cachep);
1532 fs_initcall(init_hugetlbfs_fs)