2 * hugetlbpage-backed filesystem. Based on ramfs.
4 * Nadia Yvette Chambers, 2002
6 * Copyright (C) 2002 Linus Torvalds.
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/module.h>
12 #include <linux/thread_info.h>
13 #include <asm/current.h>
14 #include <linux/sched.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/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 <asm/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 struct hugetlbfs_config {
54 struct hstate *hstate;
58 struct hugetlbfs_inode_info {
59 struct shared_policy policy;
60 struct inode vfs_inode;
63 static inline struct hugetlbfs_inode_info *HUGETLBFS_I(struct inode *inode)
65 return container_of(inode, struct hugetlbfs_inode_info, vfs_inode);
68 int sysctl_hugetlb_shm_group;
71 Opt_size, Opt_nr_inodes,
72 Opt_mode, Opt_uid, Opt_gid,
73 Opt_pagesize, Opt_min_size,
77 static const match_table_t tokens = {
78 {Opt_size, "size=%s"},
79 {Opt_nr_inodes, "nr_inodes=%s"},
80 {Opt_mode, "mode=%o"},
83 {Opt_pagesize, "pagesize=%s"},
84 {Opt_min_size, "min_size=%s"},
89 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
90 struct inode *inode, pgoff_t index)
92 vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
96 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
98 mpol_cond_put(vma->vm_policy);
101 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
102 struct inode *inode, pgoff_t index)
106 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
111 static void huge_pagevec_release(struct pagevec *pvec)
115 for (i = 0; i < pagevec_count(pvec); ++i)
116 put_page(pvec->pages[i]);
118 pagevec_reinit(pvec);
122 * Mask used when checking the page offset value passed in via system
123 * calls. This value will be converted to a loff_t which is signed.
124 * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
125 * value. The extra bit (- 1 in the shift value) is to take the sign
128 #define PGOFF_LOFFT_MAX \
129 (((1UL << (PAGE_SHIFT + 1)) - 1) << (BITS_PER_LONG - (PAGE_SHIFT + 1)))
131 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
133 struct inode *inode = file_inode(file);
136 struct hstate *h = hstate_file(file);
139 * vma address alignment (but not the pgoff alignment) has
140 * already been checked by prepare_hugepage_range. If you add
141 * any error returns here, do so after setting VM_HUGETLB, so
142 * is_vm_hugetlb_page tests below unmap_region go the right
143 * way when do_mmap_pgoff unwinds (may be important on powerpc
146 vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
147 vma->vm_ops = &hugetlb_vm_ops;
150 * page based offset in vm_pgoff could be sufficiently large to
151 * overflow a loff_t when converted to byte offset. This can
152 * only happen on architectures where sizeof(loff_t) ==
153 * sizeof(unsigned long). So, only check in those instances.
155 if (sizeof(unsigned long) == sizeof(loff_t)) {
156 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
160 /* must be huge page aligned */
161 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
164 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
165 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
166 /* check for overflow */
170 mutex_lock(&inode->i_mutex);
174 if (hugetlb_reserve_pages(inode,
175 vma->vm_pgoff >> huge_page_order(h),
176 len >> huge_page_shift(h), vma,
181 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
182 i_size_write(inode, len);
184 mutex_unlock(&inode->i_mutex);
190 * Called under down_write(mmap_sem).
193 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
195 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
196 unsigned long len, unsigned long pgoff, unsigned long flags)
198 struct mm_struct *mm = current->mm;
199 struct vm_area_struct *vma;
200 struct hstate *h = hstate_file(file);
201 struct vm_unmapped_area_info info;
203 if (len & ~huge_page_mask(h))
208 if (flags & MAP_FIXED) {
209 if (prepare_hugepage_range(file, addr, len))
215 addr = ALIGN(addr, huge_page_size(h));
216 vma = find_vma(mm, addr);
217 if (TASK_SIZE - len >= addr &&
218 (!vma || addr + len <= vm_start_gap(vma)))
224 info.low_limit = TASK_UNMAPPED_BASE;
225 info.high_limit = TASK_SIZE;
226 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
227 info.align_offset = 0;
228 return vm_unmapped_area(&info);
233 hugetlbfs_read_actor(struct page *page, unsigned long offset,
234 struct iov_iter *to, unsigned long size)
239 /* Find which 4k chunk and offset with in that chunk */
240 i = offset >> PAGE_CACHE_SHIFT;
241 offset = offset & ~PAGE_CACHE_MASK;
245 chunksize = PAGE_CACHE_SIZE;
248 if (chunksize > size)
250 n = copy_page_to_iter(&page[i], offset, chunksize, to);
262 * Support for read() - Find the page attached to f_mapping and copy out the
263 * data. Its *very* similar to do_generic_mapping_read(), we can't use that
264 * since it has PAGE_CACHE_SIZE assumptions.
266 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
268 struct file *file = iocb->ki_filp;
269 struct hstate *h = hstate_file(file);
270 struct address_space *mapping = file->f_mapping;
271 struct inode *inode = mapping->host;
272 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
273 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
274 unsigned long end_index;
278 while (iov_iter_count(to)) {
282 /* nr is the maximum number of bytes to copy from this page */
283 nr = huge_page_size(h);
284 isize = i_size_read(inode);
287 end_index = (isize - 1) >> huge_page_shift(h);
288 if (index > end_index)
290 if (index == end_index) {
291 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
298 page = find_lock_page(mapping, index);
299 if (unlikely(page == NULL)) {
301 * We have a HOLE, zero out the user-buffer for the
302 * length of the hole or request.
304 copied = iov_iter_zero(nr, to);
309 * We have the page, copy it to user space buffer.
311 copied = hugetlbfs_read_actor(page, offset, to, nr);
312 page_cache_release(page);
316 if (copied != nr && iov_iter_count(to)) {
321 index += offset >> huge_page_shift(h);
322 offset &= ~huge_page_mask(h);
324 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
328 static int hugetlbfs_write_begin(struct file *file,
329 struct address_space *mapping,
330 loff_t pos, unsigned len, unsigned flags,
331 struct page **pagep, void **fsdata)
336 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
337 loff_t pos, unsigned len, unsigned copied,
338 struct page *page, void *fsdata)
344 static void remove_huge_page(struct page *page)
346 ClearPageDirty(page);
347 ClearPageUptodate(page);
348 delete_from_page_cache(page);
353 * remove_inode_hugepages handles two distinct cases: truncation and hole
354 * punch. There are subtle differences in operation for each case.
356 * truncation is indicated by end of range being LLONG_MAX
357 * In this case, we first scan the range and release found pages.
358 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
359 * maps and global counts. Page faults can not race with truncation
360 * in this routine. hugetlb_no_page() prevents page faults in the
361 * truncated range. It checks i_size before allocation, and again after
362 * with the page table lock for the page held. The same lock must be
363 * acquired to unmap a page.
364 * hole punch is indicated if end is not LLONG_MAX
365 * In the hole punch case we scan the range and release found pages.
366 * Only when releasing a page is the associated region/reserv map
367 * deleted. The region/reserv map for ranges without associated
368 * pages are not modified. Page faults can race with hole punch.
369 * This is indicated if we find a mapped page.
370 * Note: If the passed end of range value is beyond the end of file, but
371 * not LLONG_MAX this routine still performs a hole punch operation.
373 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
376 struct hstate *h = hstate_inode(inode);
377 struct address_space *mapping = &inode->i_data;
378 const pgoff_t start = lstart >> huge_page_shift(h);
379 const pgoff_t end = lend >> huge_page_shift(h);
380 struct vm_area_struct pseudo_vma;
384 long lookup_nr = PAGEVEC_SIZE;
385 bool truncate_op = (lend == LLONG_MAX);
387 memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
388 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
389 pagevec_init(&pvec, 0);
393 * Don't grab more pages than the number left in the range.
395 if (end - next < lookup_nr)
396 lookup_nr = end - next;
399 * When no more pages are found, we are done.
401 if (!pagevec_lookup(&pvec, mapping, next, lookup_nr))
404 for (i = 0; i < pagevec_count(&pvec); ++i) {
405 struct page *page = pvec.pages[i];
409 * The page (index) could be beyond end. This is
410 * only possible in the punch hole case as end is
411 * max page offset in the truncate case.
417 hash = hugetlb_fault_mutex_hash(h, mapping, next);
418 mutex_lock(&hugetlb_fault_mutex_table[hash]);
421 if (likely(!page_mapped(page))) {
422 bool rsv_on_error = !PagePrivate(page);
424 * We must free the huge page and remove
425 * from page cache (remove_huge_page) BEFORE
426 * removing the region/reserve map
427 * (hugetlb_unreserve_pages). In rare out
428 * of memory conditions, removal of the
429 * region/reserve map could fail. Before
430 * free'ing the page, note PagePrivate which
431 * is used in case of error.
433 remove_huge_page(page);
436 if (unlikely(hugetlb_unreserve_pages(
439 hugetlb_fix_reserve_counts(
440 inode, rsv_on_error);
444 * If page is mapped, it was faulted in after
445 * being unmapped. It indicates a race between
446 * hole punch and page fault. Do nothing in
447 * this case. Getting here in a truncate
448 * operation is a bug.
454 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
457 huge_pagevec_release(&pvec);
462 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
465 static void hugetlbfs_evict_inode(struct inode *inode)
467 struct resv_map *resv_map;
469 remove_inode_hugepages(inode, 0, LLONG_MAX);
470 resv_map = (struct resv_map *)inode->i_mapping->private_data;
471 /* root inode doesn't have the resv_map, so we should check it */
473 resv_map_release(&resv_map->refs);
478 hugetlb_vmdelete_list(struct rb_root *root, pgoff_t start, pgoff_t end)
480 struct vm_area_struct *vma;
483 * end == 0 indicates that the entire range after
484 * start should be unmapped.
486 vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
487 unsigned long v_offset;
491 * Can the expression below overflow on 32-bit arches?
492 * No, because the interval tree returns us only those vmas
493 * which overlap the truncated area starting at pgoff,
494 * and no vma on a 32-bit arch can span beyond the 4GB.
496 if (vma->vm_pgoff < start)
497 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
504 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
506 if (v_end > vma->vm_end)
510 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
515 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
518 struct address_space *mapping = inode->i_mapping;
519 struct hstate *h = hstate_inode(inode);
521 BUG_ON(offset & ~huge_page_mask(h));
522 pgoff = offset >> PAGE_SHIFT;
524 i_size_write(inode, offset);
525 i_mmap_lock_write(mapping);
526 if (!RB_EMPTY_ROOT(&mapping->i_mmap))
527 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
528 i_mmap_unlock_write(mapping);
529 remove_inode_hugepages(inode, offset, LLONG_MAX);
533 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
535 struct hstate *h = hstate_inode(inode);
536 loff_t hpage_size = huge_page_size(h);
537 loff_t hole_start, hole_end;
540 * For hole punch round up the beginning offset of the hole and
541 * round down the end.
543 hole_start = round_up(offset, hpage_size);
544 hole_end = round_down(offset + len, hpage_size);
546 if (hole_end > hole_start) {
547 struct address_space *mapping = inode->i_mapping;
549 mutex_lock(&inode->i_mutex);
550 i_mmap_lock_write(mapping);
551 if (!RB_EMPTY_ROOT(&mapping->i_mmap))
552 hugetlb_vmdelete_list(&mapping->i_mmap,
553 hole_start >> PAGE_SHIFT,
554 hole_end >> PAGE_SHIFT);
555 i_mmap_unlock_write(mapping);
556 remove_inode_hugepages(inode, hole_start, hole_end);
557 mutex_unlock(&inode->i_mutex);
563 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
566 struct inode *inode = file_inode(file);
567 struct address_space *mapping = inode->i_mapping;
568 struct hstate *h = hstate_inode(inode);
569 struct vm_area_struct pseudo_vma;
570 loff_t hpage_size = huge_page_size(h);
571 unsigned long hpage_shift = huge_page_shift(h);
572 pgoff_t start, index, end;
576 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
579 if (mode & FALLOC_FL_PUNCH_HOLE)
580 return hugetlbfs_punch_hole(inode, offset, len);
583 * Default preallocate case.
584 * For this range, start is rounded down and end is rounded up
585 * as well as being converted to page offsets.
587 start = offset >> hpage_shift;
588 end = (offset + len + hpage_size - 1) >> hpage_shift;
590 mutex_lock(&inode->i_mutex);
592 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
593 error = inode_newsize_ok(inode, offset + len);
598 * Initialize a pseudo vma as this is required by the huge page
599 * allocation routines. If NUMA is configured, use page index
600 * as input to create an allocation policy.
602 memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
603 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
604 pseudo_vma.vm_file = file;
606 for (index = start; index < end; index++) {
608 * This is supposed to be the vaddr where the page is being
609 * faulted in, but we have no vaddr here.
613 int avoid_reserve = 0;
618 * fallocate(2) manpage permits EINTR; we may have been
619 * interrupted because we are using up too much memory.
621 if (signal_pending(current)) {
626 /* Set numa allocation policy based on index */
627 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
629 /* addr is the offset within the file (zero based) */
630 addr = index * hpage_size;
632 /* mutex taken here, fault path and hole punch */
633 hash = hugetlb_fault_mutex_hash(h, mapping, index);
634 mutex_lock(&hugetlb_fault_mutex_table[hash]);
636 /* See if already present in mapping to avoid alloc/free */
637 page = find_get_page(mapping, index);
640 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
641 hugetlb_drop_vma_policy(&pseudo_vma);
645 /* Allocate page and add to page cache */
646 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
647 hugetlb_drop_vma_policy(&pseudo_vma);
649 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
650 error = PTR_ERR(page);
653 clear_huge_page(page, addr, pages_per_huge_page(h));
654 __SetPageUptodate(page);
655 error = huge_add_to_page_cache(page, mapping, index);
656 if (unlikely(error)) {
658 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
662 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
664 set_page_huge_active(page);
666 * put_page() due to reference from alloc_huge_page()
667 * unlock_page because locked by add_to_page_cache()
673 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
674 i_size_write(inode, offset + len);
675 inode->i_ctime = CURRENT_TIME;
677 mutex_unlock(&inode->i_mutex);
681 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
683 struct inode *inode = d_inode(dentry);
684 struct hstate *h = hstate_inode(inode);
686 unsigned int ia_valid = attr->ia_valid;
690 error = inode_change_ok(inode, attr);
694 if (ia_valid & ATTR_SIZE) {
696 if (attr->ia_size & ~huge_page_mask(h))
698 error = hugetlb_vmtruncate(inode, attr->ia_size);
703 setattr_copy(inode, attr);
704 mark_inode_dirty(inode);
708 static struct inode *hugetlbfs_get_root(struct super_block *sb,
709 struct hugetlbfs_config *config)
713 inode = new_inode(sb);
715 struct hugetlbfs_inode_info *info;
716 inode->i_ino = get_next_ino();
717 inode->i_mode = S_IFDIR | config->mode;
718 inode->i_uid = config->uid;
719 inode->i_gid = config->gid;
720 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
721 info = HUGETLBFS_I(inode);
722 mpol_shared_policy_init(&info->policy, NULL);
723 inode->i_op = &hugetlbfs_dir_inode_operations;
724 inode->i_fop = &simple_dir_operations;
725 /* directory inodes start off with i_nlink == 2 (for "." entry) */
727 lockdep_annotate_inode_mutex_key(inode);
733 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
734 * be taken from reclaim -- unlike regular filesystems. This needs an
735 * annotation because huge_pmd_share() does an allocation under
738 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
740 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
742 umode_t mode, dev_t dev)
745 struct resv_map *resv_map = NULL;
748 * Reserve maps are only needed for inodes that can have associated
751 if (S_ISREG(mode) || S_ISLNK(mode)) {
752 resv_map = resv_map_alloc();
757 inode = new_inode(sb);
759 struct hugetlbfs_inode_info *info;
760 inode->i_ino = get_next_ino();
761 inode_init_owner(inode, dir, mode);
762 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
763 &hugetlbfs_i_mmap_rwsem_key);
764 inode->i_mapping->a_ops = &hugetlbfs_aops;
765 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
766 inode->i_mapping->private_data = resv_map;
767 info = HUGETLBFS_I(inode);
769 * The policy is initialized here even if we are creating a
770 * private inode because initialization simply creates an
771 * an empty rb tree and calls spin_lock_init(), later when we
772 * call mpol_free_shared_policy() it will just return because
773 * the rb tree will still be empty.
775 mpol_shared_policy_init(&info->policy, NULL);
776 switch (mode & S_IFMT) {
778 init_special_inode(inode, mode, dev);
781 inode->i_op = &hugetlbfs_inode_operations;
782 inode->i_fop = &hugetlbfs_file_operations;
785 inode->i_op = &hugetlbfs_dir_inode_operations;
786 inode->i_fop = &simple_dir_operations;
788 /* directory inodes start off with i_nlink == 2 (for "." entry) */
792 inode->i_op = &page_symlink_inode_operations;
795 lockdep_annotate_inode_mutex_key(inode);
798 kref_put(&resv_map->refs, resv_map_release);
805 * File creation. Allocate an inode, and we're done..
807 static int hugetlbfs_mknod(struct inode *dir,
808 struct dentry *dentry, umode_t mode, dev_t dev)
813 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
815 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
816 d_instantiate(dentry, inode);
817 dget(dentry); /* Extra count - pin the dentry in core */
823 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
825 int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
831 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
833 return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
836 static int hugetlbfs_symlink(struct inode *dir,
837 struct dentry *dentry, const char *symname)
842 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
844 int l = strlen(symname)+1;
845 error = page_symlink(inode, symname, l);
847 d_instantiate(dentry, inode);
852 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
858 * mark the head page dirty
860 static int hugetlbfs_set_page_dirty(struct page *page)
862 struct page *head = compound_head(page);
868 static int hugetlbfs_migrate_page(struct address_space *mapping,
869 struct page *newpage, struct page *page,
870 enum migrate_mode mode)
874 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
875 if (rc != MIGRATEPAGE_SUCCESS)
879 * page_private is subpool pointer in hugetlb pages. Transfer to
880 * new page. PagePrivate is not associated with page_private for
881 * hugetlb pages and can not be set here as only page_huge_active
882 * pages can be migrated.
884 if (page_private(page)) {
885 set_page_private(newpage, page_private(page));
886 set_page_private(page, 0);
889 migrate_page_copy(newpage, page);
891 return MIGRATEPAGE_SUCCESS;
894 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
896 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
897 struct hstate *h = hstate_inode(d_inode(dentry));
899 buf->f_type = HUGETLBFS_MAGIC;
900 buf->f_bsize = huge_page_size(h);
902 spin_lock(&sbinfo->stat_lock);
903 /* If no limits set, just report 0 for max/free/used
904 * blocks, like simple_statfs() */
908 spin_lock(&sbinfo->spool->lock);
909 buf->f_blocks = sbinfo->spool->max_hpages;
910 free_pages = sbinfo->spool->max_hpages
911 - sbinfo->spool->used_hpages;
912 buf->f_bavail = buf->f_bfree = free_pages;
913 spin_unlock(&sbinfo->spool->lock);
914 buf->f_files = sbinfo->max_inodes;
915 buf->f_ffree = sbinfo->free_inodes;
917 spin_unlock(&sbinfo->stat_lock);
919 buf->f_namelen = NAME_MAX;
923 static void hugetlbfs_put_super(struct super_block *sb)
925 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
928 sb->s_fs_info = NULL;
931 hugepage_put_subpool(sbi->spool);
937 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
939 if (sbinfo->free_inodes >= 0) {
940 spin_lock(&sbinfo->stat_lock);
941 if (unlikely(!sbinfo->free_inodes)) {
942 spin_unlock(&sbinfo->stat_lock);
945 sbinfo->free_inodes--;
946 spin_unlock(&sbinfo->stat_lock);
952 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
954 if (sbinfo->free_inodes >= 0) {
955 spin_lock(&sbinfo->stat_lock);
956 sbinfo->free_inodes++;
957 spin_unlock(&sbinfo->stat_lock);
962 static struct kmem_cache *hugetlbfs_inode_cachep;
964 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
966 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
967 struct hugetlbfs_inode_info *p;
969 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
971 p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
973 hugetlbfs_inc_free_inodes(sbinfo);
976 return &p->vfs_inode;
979 static void hugetlbfs_i_callback(struct rcu_head *head)
981 struct inode *inode = container_of(head, struct inode, i_rcu);
982 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
985 static void hugetlbfs_destroy_inode(struct inode *inode)
987 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
988 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
989 call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
992 static const struct address_space_operations hugetlbfs_aops = {
993 .write_begin = hugetlbfs_write_begin,
994 .write_end = hugetlbfs_write_end,
995 .set_page_dirty = hugetlbfs_set_page_dirty,
996 .migratepage = hugetlbfs_migrate_page,
1000 static void init_once(void *foo)
1002 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1004 inode_init_once(&ei->vfs_inode);
1007 const struct file_operations hugetlbfs_file_operations = {
1008 .read_iter = hugetlbfs_read_iter,
1009 .mmap = hugetlbfs_file_mmap,
1010 .fsync = noop_fsync,
1011 .get_unmapped_area = hugetlb_get_unmapped_area,
1012 .llseek = default_llseek,
1013 .fallocate = hugetlbfs_fallocate,
1016 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1017 .create = hugetlbfs_create,
1018 .lookup = simple_lookup,
1019 .link = simple_link,
1020 .unlink = simple_unlink,
1021 .symlink = hugetlbfs_symlink,
1022 .mkdir = hugetlbfs_mkdir,
1023 .rmdir = simple_rmdir,
1024 .mknod = hugetlbfs_mknod,
1025 .rename = simple_rename,
1026 .setattr = hugetlbfs_setattr,
1029 static const struct inode_operations hugetlbfs_inode_operations = {
1030 .setattr = hugetlbfs_setattr,
1033 static const struct super_operations hugetlbfs_ops = {
1034 .alloc_inode = hugetlbfs_alloc_inode,
1035 .destroy_inode = hugetlbfs_destroy_inode,
1036 .evict_inode = hugetlbfs_evict_inode,
1037 .statfs = hugetlbfs_statfs,
1038 .put_super = hugetlbfs_put_super,
1039 .show_options = generic_show_options,
1042 enum { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1045 * Convert size option passed from command line to number of huge pages
1046 * in the pool specified by hstate. Size option could be in bytes
1047 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1050 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1053 if (val_type == NO_SIZE)
1056 if (val_type == SIZE_PERCENT) {
1057 size_opt <<= huge_page_shift(h);
1058 size_opt *= h->max_huge_pages;
1059 do_div(size_opt, 100);
1062 size_opt >>= huge_page_shift(h);
1067 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1070 substring_t args[MAX_OPT_ARGS];
1072 unsigned long long max_size_opt = 0, min_size_opt = 0;
1073 int max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1078 while ((p = strsep(&options, ",")) != NULL) {
1083 token = match_token(p, tokens, args);
1086 if (match_int(&args[0], &option))
1088 pconfig->uid = make_kuid(current_user_ns(), option);
1089 if (!uid_valid(pconfig->uid))
1094 if (match_int(&args[0], &option))
1096 pconfig->gid = make_kgid(current_user_ns(), option);
1097 if (!gid_valid(pconfig->gid))
1102 if (match_octal(&args[0], &option))
1104 pconfig->mode = option & 01777U;
1108 /* memparse() will accept a K/M/G without a digit */
1109 if (!isdigit(*args[0].from))
1111 max_size_opt = memparse(args[0].from, &rest);
1112 max_val_type = SIZE_STD;
1114 max_val_type = SIZE_PERCENT;
1119 /* memparse() will accept a K/M/G without a digit */
1120 if (!isdigit(*args[0].from))
1122 pconfig->nr_inodes = memparse(args[0].from, &rest);
1125 case Opt_pagesize: {
1127 ps = memparse(args[0].from, &rest);
1128 pconfig->hstate = size_to_hstate(ps);
1129 if (!pconfig->hstate) {
1130 pr_err("Unsupported page size %lu MB\n",
1137 case Opt_min_size: {
1138 /* memparse() will accept a K/M/G without a digit */
1139 if (!isdigit(*args[0].from))
1141 min_size_opt = memparse(args[0].from, &rest);
1142 min_val_type = SIZE_STD;
1144 min_val_type = SIZE_PERCENT;
1149 pr_err("Bad mount option: \"%s\"\n", p);
1156 * Use huge page pool size (in hstate) to convert the size
1157 * options to number of huge pages. If NO_SIZE, -1 is returned.
1159 pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1160 max_size_opt, max_val_type);
1161 pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1162 min_size_opt, min_val_type);
1165 * If max_size was specified, then min_size must be smaller
1167 if (max_val_type > NO_SIZE &&
1168 pconfig->min_hpages > pconfig->max_hpages) {
1169 pr_err("minimum size can not be greater than maximum size\n");
1176 pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1181 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1184 struct hugetlbfs_config config;
1185 struct hugetlbfs_sb_info *sbinfo;
1187 save_mount_options(sb, data);
1189 config.max_hpages = -1; /* No limit on size by default */
1190 config.nr_inodes = -1; /* No limit on number of inodes by default */
1191 config.uid = current_fsuid();
1192 config.gid = current_fsgid();
1194 config.hstate = &default_hstate;
1195 config.min_hpages = -1; /* No default minimum size */
1196 ret = hugetlbfs_parse_options(data, &config);
1200 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1203 sb->s_fs_info = sbinfo;
1204 sbinfo->hstate = config.hstate;
1205 spin_lock_init(&sbinfo->stat_lock);
1206 sbinfo->max_inodes = config.nr_inodes;
1207 sbinfo->free_inodes = config.nr_inodes;
1208 sbinfo->spool = NULL;
1210 * Allocate and initialize subpool if maximum or minimum size is
1211 * specified. Any needed reservations (for minimim size) are taken
1212 * taken when the subpool is created.
1214 if (config.max_hpages != -1 || config.min_hpages != -1) {
1215 sbinfo->spool = hugepage_new_subpool(config.hstate,
1221 sb->s_maxbytes = MAX_LFS_FILESIZE;
1222 sb->s_blocksize = huge_page_size(config.hstate);
1223 sb->s_blocksize_bits = huge_page_shift(config.hstate);
1224 sb->s_magic = HUGETLBFS_MAGIC;
1225 sb->s_op = &hugetlbfs_ops;
1226 sb->s_time_gran = 1;
1227 sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1232 kfree(sbinfo->spool);
1237 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1238 int flags, const char *dev_name, void *data)
1240 return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1243 static struct file_system_type hugetlbfs_fs_type = {
1244 .name = "hugetlbfs",
1245 .mount = hugetlbfs_mount,
1246 .kill_sb = kill_litter_super,
1248 MODULE_ALIAS_FS("hugetlbfs");
1250 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1252 static int can_do_hugetlb_shm(void)
1255 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1256 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1259 static int get_hstate_idx(int page_size_log)
1261 struct hstate *h = hstate_sizelog(page_size_log);
1268 static const struct dentry_operations anon_ops = {
1269 .d_dname = simple_dname
1273 * Note that size should be aligned to proper hugepage size in caller side,
1274 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1276 struct file *hugetlb_file_setup(const char *name, size_t size,
1277 vm_flags_t acctflag, struct user_struct **user,
1278 int creat_flags, int page_size_log)
1280 struct file *file = ERR_PTR(-ENOMEM);
1281 struct inode *inode;
1283 struct super_block *sb;
1284 struct qstr quick_string;
1287 hstate_idx = get_hstate_idx(page_size_log);
1289 return ERR_PTR(-ENODEV);
1292 if (!hugetlbfs_vfsmount[hstate_idx])
1293 return ERR_PTR(-ENOENT);
1295 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1296 *user = current_user();
1297 if (user_shm_lock(size, *user)) {
1299 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1300 current->comm, current->pid);
1301 task_unlock(current);
1304 return ERR_PTR(-EPERM);
1308 sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1309 quick_string.name = name;
1310 quick_string.len = strlen(quick_string.name);
1311 quick_string.hash = 0;
1312 path.dentry = d_alloc_pseudo(sb, &quick_string);
1314 goto out_shm_unlock;
1316 d_set_d_op(path.dentry, &anon_ops);
1317 path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1318 file = ERR_PTR(-ENOSPC);
1319 inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1322 if (creat_flags == HUGETLB_SHMFS_INODE)
1323 inode->i_flags |= S_PRIVATE;
1325 file = ERR_PTR(-ENOMEM);
1326 if (hugetlb_reserve_pages(inode, 0,
1327 size >> huge_page_shift(hstate_inode(inode)), NULL,
1331 d_instantiate(path.dentry, inode);
1332 inode->i_size = size;
1335 file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1336 &hugetlbfs_file_operations);
1338 goto out_dentry; /* inode is already attached */
1348 user_shm_unlock(size, *user);
1354 static int __init init_hugetlbfs_fs(void)
1360 if (!hugepages_supported()) {
1361 pr_info("disabling because there are no supported hugepage sizes\n");
1366 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1367 sizeof(struct hugetlbfs_inode_info),
1369 if (hugetlbfs_inode_cachep == NULL)
1372 error = register_filesystem(&hugetlbfs_fs_type);
1377 for_each_hstate(h) {
1379 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1381 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1382 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1385 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1386 pr_err("Cannot mount internal hugetlbfs for "
1387 "page size %uK", ps_kb);
1388 error = PTR_ERR(hugetlbfs_vfsmount[i]);
1389 hugetlbfs_vfsmount[i] = NULL;
1393 /* Non default hstates are optional */
1394 if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1398 kmem_cache_destroy(hugetlbfs_inode_cachep);
1403 static void __exit exit_hugetlbfs_fs(void)
1410 * Make sure all delayed rcu free inodes are flushed before we
1414 kmem_cache_destroy(hugetlbfs_inode_cachep);
1417 kern_unmount(hugetlbfs_vfsmount[i++]);
1418 unregister_filesystem(&hugetlbfs_fs_type);
1421 module_init(init_hugetlbfs_fs)
1422 module_exit(exit_hugetlbfs_fs)
1424 MODULE_LICENSE("GPL");