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/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 struct hugetlbfs_config {
49 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 */
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);
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_SHIFT;
241 offset = offset & ~PAGE_MASK;
245 chunksize = PAGE_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_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);
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);
352 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
354 struct vm_area_struct *vma;
357 * end == 0 indicates that the entire range after
358 * start should be unmapped.
360 vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
361 unsigned long v_offset;
365 * Can the expression below overflow on 32-bit arches?
366 * No, because the interval tree returns us only those vmas
367 * which overlap the truncated area starting at pgoff,
368 * and no vma on a 32-bit arch can span beyond the 4GB.
370 if (vma->vm_pgoff < start)
371 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
378 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
380 if (v_end > vma->vm_end)
384 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
390 * remove_inode_hugepages handles two distinct cases: truncation and hole
391 * punch. There are subtle differences in operation for each case.
393 * truncation is indicated by end of range being LLONG_MAX
394 * In this case, we first scan the range and release found pages.
395 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
396 * maps and global counts. Page faults can not race with truncation
397 * in this routine. hugetlb_no_page() prevents page faults in the
398 * truncated range. It checks i_size before allocation, and again after
399 * with the page table lock for the page held. The same lock must be
400 * acquired to unmap a page.
401 * hole punch is indicated if end is not LLONG_MAX
402 * In the hole punch case we scan the range and release found pages.
403 * Only when releasing a page is the associated region/reserv map
404 * deleted. The region/reserv map for ranges without associated
405 * pages are not modified. Page faults can race with hole punch.
406 * This is indicated if we find a mapped page.
407 * Note: If the passed end of range value is beyond the end of file, but
408 * not LLONG_MAX this routine still performs a hole punch operation.
410 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
413 struct hstate *h = hstate_inode(inode);
414 struct address_space *mapping = &inode->i_data;
415 const pgoff_t start = lstart >> huge_page_shift(h);
416 const pgoff_t end = lend >> huge_page_shift(h);
417 struct vm_area_struct pseudo_vma;
421 bool truncate_op = (lend == LLONG_MAX);
423 memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
424 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
425 pagevec_init(&pvec, 0);
429 * When no more pages are found, we are done.
431 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
434 for (i = 0; i < pagevec_count(&pvec); ++i) {
435 struct page *page = pvec.pages[i];
439 hash = hugetlb_fault_mutex_hash(h, mapping, index);
440 mutex_lock(&hugetlb_fault_mutex_table[hash]);
443 * If page is mapped, it was faulted in after being
444 * unmapped in caller. Unmap (again) now after taking
445 * the fault mutex. The mutex will prevent faults
446 * until we finish removing the page.
448 * This race can only happen in the hole punch case.
449 * Getting here in a truncate operation is a bug.
451 if (unlikely(page_mapped(page))) {
454 i_mmap_lock_write(mapping);
455 hugetlb_vmdelete_list(&mapping->i_mmap,
456 index * pages_per_huge_page(h),
457 (index + 1) * pages_per_huge_page(h));
458 i_mmap_unlock_write(mapping);
463 * We must free the huge page and remove from page
464 * cache (remove_huge_page) BEFORE removing the
465 * region/reserve map (hugetlb_unreserve_pages). In
466 * rare out of memory conditions, removal of the
467 * region/reserve map could fail. Correspondingly,
468 * the subpool and global reserve usage count can need
471 VM_BUG_ON(PagePrivate(page));
472 remove_huge_page(page);
475 if (unlikely(hugetlb_unreserve_pages(inode,
476 index, index + 1, 1)))
477 hugetlb_fix_reserve_counts(inode);
481 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
483 huge_pagevec_release(&pvec);
488 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
491 static void hugetlbfs_evict_inode(struct inode *inode)
493 struct resv_map *resv_map;
495 remove_inode_hugepages(inode, 0, LLONG_MAX);
496 resv_map = (struct resv_map *)inode->i_mapping->private_data;
497 /* root inode doesn't have the resv_map, so we should check it */
499 resv_map_release(&resv_map->refs);
503 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
506 struct address_space *mapping = inode->i_mapping;
507 struct hstate *h = hstate_inode(inode);
509 BUG_ON(offset & ~huge_page_mask(h));
510 pgoff = offset >> PAGE_SHIFT;
512 i_size_write(inode, offset);
513 i_mmap_lock_write(mapping);
514 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
515 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
516 i_mmap_unlock_write(mapping);
517 remove_inode_hugepages(inode, offset, LLONG_MAX);
521 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
523 struct hstate *h = hstate_inode(inode);
524 loff_t hpage_size = huge_page_size(h);
525 loff_t hole_start, hole_end;
528 * For hole punch round up the beginning offset of the hole and
529 * round down the end.
531 hole_start = round_up(offset, hpage_size);
532 hole_end = round_down(offset + len, hpage_size);
534 if (hole_end > hole_start) {
535 struct address_space *mapping = inode->i_mapping;
538 i_mmap_lock_write(mapping);
539 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
540 hugetlb_vmdelete_list(&mapping->i_mmap,
541 hole_start >> PAGE_SHIFT,
542 hole_end >> PAGE_SHIFT);
543 i_mmap_unlock_write(mapping);
544 remove_inode_hugepages(inode, hole_start, hole_end);
551 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
554 struct inode *inode = file_inode(file);
555 struct address_space *mapping = inode->i_mapping;
556 struct hstate *h = hstate_inode(inode);
557 struct vm_area_struct pseudo_vma;
558 loff_t hpage_size = huge_page_size(h);
559 unsigned long hpage_shift = huge_page_shift(h);
560 pgoff_t start, index, end;
564 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
567 if (mode & FALLOC_FL_PUNCH_HOLE)
568 return hugetlbfs_punch_hole(inode, offset, len);
571 * Default preallocate case.
572 * For this range, start is rounded down and end is rounded up
573 * as well as being converted to page offsets.
575 start = offset >> hpage_shift;
576 end = (offset + len + hpage_size - 1) >> hpage_shift;
580 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
581 error = inode_newsize_ok(inode, offset + len);
586 * Initialize a pseudo vma as this is required by the huge page
587 * allocation routines. If NUMA is configured, use page index
588 * as input to create an allocation policy.
590 memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
591 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
592 pseudo_vma.vm_file = file;
594 for (index = start; index < end; index++) {
596 * This is supposed to be the vaddr where the page is being
597 * faulted in, but we have no vaddr here.
601 int avoid_reserve = 0;
606 * fallocate(2) manpage permits EINTR; we may have been
607 * interrupted because we are using up too much memory.
609 if (signal_pending(current)) {
614 /* Set numa allocation policy based on index */
615 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
617 /* addr is the offset within the file (zero based) */
618 addr = index * hpage_size;
620 /* mutex taken here, fault path and hole punch */
621 hash = hugetlb_fault_mutex_hash(h, mapping, index);
622 mutex_lock(&hugetlb_fault_mutex_table[hash]);
624 /* See if already present in mapping to avoid alloc/free */
625 page = find_get_page(mapping, index);
628 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
629 hugetlb_drop_vma_policy(&pseudo_vma);
633 /* Allocate page and add to page cache */
634 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
635 hugetlb_drop_vma_policy(&pseudo_vma);
637 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
638 error = PTR_ERR(page);
641 clear_huge_page(page, addr, pages_per_huge_page(h));
642 __SetPageUptodate(page);
643 error = huge_add_to_page_cache(page, mapping, index);
644 if (unlikely(error)) {
646 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
650 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
652 set_page_huge_active(page);
654 * put_page() due to reference from alloc_huge_page()
655 * unlock_page because locked by add_to_page_cache()
661 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
662 i_size_write(inode, offset + len);
663 inode->i_ctime = current_time(inode);
669 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
671 struct inode *inode = d_inode(dentry);
672 struct hstate *h = hstate_inode(inode);
674 unsigned int ia_valid = attr->ia_valid;
678 error = setattr_prepare(dentry, attr);
682 if (ia_valid & ATTR_SIZE) {
684 if (attr->ia_size & ~huge_page_mask(h))
686 error = hugetlb_vmtruncate(inode, attr->ia_size);
691 setattr_copy(inode, attr);
692 mark_inode_dirty(inode);
696 static struct inode *hugetlbfs_get_root(struct super_block *sb,
697 struct hugetlbfs_config *config)
701 inode = new_inode(sb);
703 inode->i_ino = get_next_ino();
704 inode->i_mode = S_IFDIR | config->mode;
705 inode->i_uid = config->uid;
706 inode->i_gid = config->gid;
707 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
708 inode->i_op = &hugetlbfs_dir_inode_operations;
709 inode->i_fop = &simple_dir_operations;
710 /* directory inodes start off with i_nlink == 2 (for "." entry) */
712 lockdep_annotate_inode_mutex_key(inode);
718 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
719 * be taken from reclaim -- unlike regular filesystems. This needs an
720 * annotation because huge_pmd_share() does an allocation under hugetlb's
723 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
725 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
727 umode_t mode, dev_t dev)
730 struct resv_map *resv_map = NULL;
733 * Reserve maps are only needed for inodes that can have associated
736 if (S_ISREG(mode) || S_ISLNK(mode)) {
737 resv_map = resv_map_alloc();
742 inode = new_inode(sb);
744 inode->i_ino = get_next_ino();
745 inode_init_owner(inode, dir, mode);
746 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
747 &hugetlbfs_i_mmap_rwsem_key);
748 inode->i_mapping->a_ops = &hugetlbfs_aops;
749 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
750 inode->i_mapping->private_data = resv_map;
751 switch (mode & S_IFMT) {
753 init_special_inode(inode, mode, dev);
756 inode->i_op = &hugetlbfs_inode_operations;
757 inode->i_fop = &hugetlbfs_file_operations;
760 inode->i_op = &hugetlbfs_dir_inode_operations;
761 inode->i_fop = &simple_dir_operations;
763 /* directory inodes start off with i_nlink == 2 (for "." entry) */
767 inode->i_op = &page_symlink_inode_operations;
768 inode_nohighmem(inode);
771 lockdep_annotate_inode_mutex_key(inode);
774 kref_put(&resv_map->refs, resv_map_release);
781 * File creation. Allocate an inode, and we're done..
783 static int hugetlbfs_mknod(struct inode *dir,
784 struct dentry *dentry, umode_t mode, dev_t dev)
789 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
791 dir->i_ctime = dir->i_mtime = current_time(dir);
792 d_instantiate(dentry, inode);
793 dget(dentry); /* Extra count - pin the dentry in core */
799 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
801 int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
807 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
809 return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
812 static int hugetlbfs_symlink(struct inode *dir,
813 struct dentry *dentry, const char *symname)
818 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
820 int l = strlen(symname)+1;
821 error = page_symlink(inode, symname, l);
823 d_instantiate(dentry, inode);
828 dir->i_ctime = dir->i_mtime = current_time(dir);
834 * mark the head page dirty
836 static int hugetlbfs_set_page_dirty(struct page *page)
838 struct page *head = compound_head(page);
844 static int hugetlbfs_migrate_page(struct address_space *mapping,
845 struct page *newpage, struct page *page,
846 enum migrate_mode mode)
850 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
851 if (rc != MIGRATEPAGE_SUCCESS)
855 * page_private is subpool pointer in hugetlb pages. Transfer to
856 * new page. PagePrivate is not associated with page_private for
857 * hugetlb pages and can not be set here as only page_huge_active
858 * pages can be migrated.
860 if (page_private(page)) {
861 set_page_private(newpage, page_private(page));
862 set_page_private(page, 0);
865 if (mode != MIGRATE_SYNC_NO_COPY)
866 migrate_page_copy(newpage, page);
868 migrate_page_states(newpage, page);
870 return MIGRATEPAGE_SUCCESS;
873 static int hugetlbfs_error_remove_page(struct address_space *mapping,
876 struct inode *inode = mapping->host;
877 pgoff_t index = page->index;
879 remove_huge_page(page);
880 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
881 hugetlb_fix_reserve_counts(inode);
887 * Display the mount options in /proc/mounts.
889 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
891 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
892 struct hugepage_subpool *spool = sbinfo->spool;
893 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
894 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
897 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
898 seq_printf(m, ",uid=%u",
899 from_kuid_munged(&init_user_ns, sbinfo->uid));
900 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
901 seq_printf(m, ",gid=%u",
902 from_kgid_munged(&init_user_ns, sbinfo->gid));
903 if (sbinfo->mode != 0755)
904 seq_printf(m, ",mode=%o", sbinfo->mode);
905 if (sbinfo->max_inodes != -1)
906 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
910 if (hpage_size >= 1024) {
914 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
916 if (spool->max_hpages != -1)
917 seq_printf(m, ",size=%llu",
918 (unsigned long long)spool->max_hpages << hpage_shift);
919 if (spool->min_hpages != -1)
920 seq_printf(m, ",min_size=%llu",
921 (unsigned long long)spool->min_hpages << hpage_shift);
926 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
928 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
929 struct hstate *h = hstate_inode(d_inode(dentry));
931 buf->f_type = HUGETLBFS_MAGIC;
932 buf->f_bsize = huge_page_size(h);
934 spin_lock(&sbinfo->stat_lock);
935 /* If no limits set, just report 0 for max/free/used
936 * blocks, like simple_statfs() */
940 spin_lock(&sbinfo->spool->lock);
941 buf->f_blocks = sbinfo->spool->max_hpages;
942 free_pages = sbinfo->spool->max_hpages
943 - sbinfo->spool->used_hpages;
944 buf->f_bavail = buf->f_bfree = free_pages;
945 spin_unlock(&sbinfo->spool->lock);
946 buf->f_files = sbinfo->max_inodes;
947 buf->f_ffree = sbinfo->free_inodes;
949 spin_unlock(&sbinfo->stat_lock);
951 buf->f_namelen = NAME_MAX;
955 static void hugetlbfs_put_super(struct super_block *sb)
957 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
960 sb->s_fs_info = NULL;
963 hugepage_put_subpool(sbi->spool);
969 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
971 if (sbinfo->free_inodes >= 0) {
972 spin_lock(&sbinfo->stat_lock);
973 if (unlikely(!sbinfo->free_inodes)) {
974 spin_unlock(&sbinfo->stat_lock);
977 sbinfo->free_inodes--;
978 spin_unlock(&sbinfo->stat_lock);
984 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
986 if (sbinfo->free_inodes >= 0) {
987 spin_lock(&sbinfo->stat_lock);
988 sbinfo->free_inodes++;
989 spin_unlock(&sbinfo->stat_lock);
994 static struct kmem_cache *hugetlbfs_inode_cachep;
996 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
998 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
999 struct hugetlbfs_inode_info *p;
1001 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1003 p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1005 hugetlbfs_inc_free_inodes(sbinfo);
1010 * Any time after allocation, hugetlbfs_destroy_inode can be called
1011 * for the inode. mpol_free_shared_policy is unconditionally called
1012 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1013 * in case of a quick call to destroy.
1015 * Note that the policy is initialized even if we are creating a
1016 * private inode. This simplifies hugetlbfs_destroy_inode.
1018 mpol_shared_policy_init(&p->policy, NULL);
1020 return &p->vfs_inode;
1023 static void hugetlbfs_i_callback(struct rcu_head *head)
1025 struct inode *inode = container_of(head, struct inode, i_rcu);
1026 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1029 static void hugetlbfs_destroy_inode(struct inode *inode)
1031 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1032 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1033 call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
1036 static const struct address_space_operations hugetlbfs_aops = {
1037 .write_begin = hugetlbfs_write_begin,
1038 .write_end = hugetlbfs_write_end,
1039 .set_page_dirty = hugetlbfs_set_page_dirty,
1040 .migratepage = hugetlbfs_migrate_page,
1041 .error_remove_page = hugetlbfs_error_remove_page,
1045 static void init_once(void *foo)
1047 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1049 inode_init_once(&ei->vfs_inode);
1052 const struct file_operations hugetlbfs_file_operations = {
1053 .read_iter = hugetlbfs_read_iter,
1054 .mmap = hugetlbfs_file_mmap,
1055 .fsync = noop_fsync,
1056 .get_unmapped_area = hugetlb_get_unmapped_area,
1057 .llseek = default_llseek,
1058 .fallocate = hugetlbfs_fallocate,
1061 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1062 .create = hugetlbfs_create,
1063 .lookup = simple_lookup,
1064 .link = simple_link,
1065 .unlink = simple_unlink,
1066 .symlink = hugetlbfs_symlink,
1067 .mkdir = hugetlbfs_mkdir,
1068 .rmdir = simple_rmdir,
1069 .mknod = hugetlbfs_mknod,
1070 .rename = simple_rename,
1071 .setattr = hugetlbfs_setattr,
1074 static const struct inode_operations hugetlbfs_inode_operations = {
1075 .setattr = hugetlbfs_setattr,
1078 static const struct super_operations hugetlbfs_ops = {
1079 .alloc_inode = hugetlbfs_alloc_inode,
1080 .destroy_inode = hugetlbfs_destroy_inode,
1081 .evict_inode = hugetlbfs_evict_inode,
1082 .statfs = hugetlbfs_statfs,
1083 .put_super = hugetlbfs_put_super,
1084 .show_options = hugetlbfs_show_options,
1087 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1090 * Convert size option passed from command line to number of huge pages
1091 * in the pool specified by hstate. Size option could be in bytes
1092 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1095 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1096 enum hugetlbfs_size_type val_type)
1098 if (val_type == NO_SIZE)
1101 if (val_type == SIZE_PERCENT) {
1102 size_opt <<= huge_page_shift(h);
1103 size_opt *= h->max_huge_pages;
1104 do_div(size_opt, 100);
1107 size_opt >>= huge_page_shift(h);
1112 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1115 substring_t args[MAX_OPT_ARGS];
1117 unsigned long long max_size_opt = 0, min_size_opt = 0;
1118 enum hugetlbfs_size_type max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1123 while ((p = strsep(&options, ",")) != NULL) {
1128 token = match_token(p, tokens, args);
1131 if (match_int(&args[0], &option))
1133 pconfig->uid = make_kuid(current_user_ns(), option);
1134 if (!uid_valid(pconfig->uid))
1139 if (match_int(&args[0], &option))
1141 pconfig->gid = make_kgid(current_user_ns(), option);
1142 if (!gid_valid(pconfig->gid))
1147 if (match_octal(&args[0], &option))
1149 pconfig->mode = option & 01777U;
1153 /* memparse() will accept a K/M/G without a digit */
1154 if (!isdigit(*args[0].from))
1156 max_size_opt = memparse(args[0].from, &rest);
1157 max_val_type = SIZE_STD;
1159 max_val_type = SIZE_PERCENT;
1164 /* memparse() will accept a K/M/G without a digit */
1165 if (!isdigit(*args[0].from))
1167 pconfig->nr_inodes = memparse(args[0].from, &rest);
1170 case Opt_pagesize: {
1172 ps = memparse(args[0].from, &rest);
1173 pconfig->hstate = size_to_hstate(ps);
1174 if (!pconfig->hstate) {
1175 pr_err("Unsupported page size %lu MB\n",
1182 case Opt_min_size: {
1183 /* memparse() will accept a K/M/G without a digit */
1184 if (!isdigit(*args[0].from))
1186 min_size_opt = memparse(args[0].from, &rest);
1187 min_val_type = SIZE_STD;
1189 min_val_type = SIZE_PERCENT;
1194 pr_err("Bad mount option: \"%s\"\n", p);
1201 * Use huge page pool size (in hstate) to convert the size
1202 * options to number of huge pages. If NO_SIZE, -1 is returned.
1204 pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1205 max_size_opt, max_val_type);
1206 pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1207 min_size_opt, min_val_type);
1210 * If max_size was specified, then min_size must be smaller
1212 if (max_val_type > NO_SIZE &&
1213 pconfig->min_hpages > pconfig->max_hpages) {
1214 pr_err("minimum size can not be greater than maximum size\n");
1221 pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1226 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1229 struct hugetlbfs_config config;
1230 struct hugetlbfs_sb_info *sbinfo;
1232 config.max_hpages = -1; /* No limit on size by default */
1233 config.nr_inodes = -1; /* No limit on number of inodes by default */
1234 config.uid = current_fsuid();
1235 config.gid = current_fsgid();
1237 config.hstate = &default_hstate;
1238 config.min_hpages = -1; /* No default minimum size */
1239 ret = hugetlbfs_parse_options(data, &config);
1243 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1246 sb->s_fs_info = sbinfo;
1247 sbinfo->hstate = config.hstate;
1248 spin_lock_init(&sbinfo->stat_lock);
1249 sbinfo->max_inodes = config.nr_inodes;
1250 sbinfo->free_inodes = config.nr_inodes;
1251 sbinfo->spool = NULL;
1252 sbinfo->uid = config.uid;
1253 sbinfo->gid = config.gid;
1254 sbinfo->mode = config.mode;
1257 * Allocate and initialize subpool if maximum or minimum size is
1258 * specified. Any needed reservations (for minimim size) are taken
1259 * taken when the subpool is created.
1261 if (config.max_hpages != -1 || config.min_hpages != -1) {
1262 sbinfo->spool = hugepage_new_subpool(config.hstate,
1268 sb->s_maxbytes = MAX_LFS_FILESIZE;
1269 sb->s_blocksize = huge_page_size(config.hstate);
1270 sb->s_blocksize_bits = huge_page_shift(config.hstate);
1271 sb->s_magic = HUGETLBFS_MAGIC;
1272 sb->s_op = &hugetlbfs_ops;
1273 sb->s_time_gran = 1;
1274 sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1279 kfree(sbinfo->spool);
1284 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1285 int flags, const char *dev_name, void *data)
1287 return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1290 static struct file_system_type hugetlbfs_fs_type = {
1291 .name = "hugetlbfs",
1292 .mount = hugetlbfs_mount,
1293 .kill_sb = kill_litter_super,
1296 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1298 static int can_do_hugetlb_shm(void)
1301 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1302 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1305 static int get_hstate_idx(int page_size_log)
1307 struct hstate *h = hstate_sizelog(page_size_log);
1314 static const struct dentry_operations anon_ops = {
1315 .d_dname = simple_dname
1319 * Note that size should be aligned to proper hugepage size in caller side,
1320 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1322 struct file *hugetlb_file_setup(const char *name, size_t size,
1323 vm_flags_t acctflag, struct user_struct **user,
1324 int creat_flags, int page_size_log)
1326 struct file *file = ERR_PTR(-ENOMEM);
1327 struct inode *inode;
1329 struct super_block *sb;
1330 struct qstr quick_string;
1333 hstate_idx = get_hstate_idx(page_size_log);
1335 return ERR_PTR(-ENODEV);
1338 if (!hugetlbfs_vfsmount[hstate_idx])
1339 return ERR_PTR(-ENOENT);
1341 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1342 *user = current_user();
1343 if (user_shm_lock(size, *user)) {
1345 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1346 current->comm, current->pid);
1347 task_unlock(current);
1350 return ERR_PTR(-EPERM);
1354 sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1355 quick_string.name = name;
1356 quick_string.len = strlen(quick_string.name);
1357 quick_string.hash = 0;
1358 path.dentry = d_alloc_pseudo(sb, &quick_string);
1360 goto out_shm_unlock;
1362 d_set_d_op(path.dentry, &anon_ops);
1363 path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1364 file = ERR_PTR(-ENOSPC);
1365 inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1368 if (creat_flags == HUGETLB_SHMFS_INODE)
1369 inode->i_flags |= S_PRIVATE;
1371 file = ERR_PTR(-ENOMEM);
1372 if (hugetlb_reserve_pages(inode, 0,
1373 size >> huge_page_shift(hstate_inode(inode)), NULL,
1377 d_instantiate(path.dentry, inode);
1378 inode->i_size = size;
1381 file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1382 &hugetlbfs_file_operations);
1384 goto out_dentry; /* inode is already attached */
1394 user_shm_unlock(size, *user);
1400 static int __init init_hugetlbfs_fs(void)
1406 if (!hugepages_supported()) {
1407 pr_info("disabling because there are no supported hugepage sizes\n");
1412 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1413 sizeof(struct hugetlbfs_inode_info),
1414 0, SLAB_ACCOUNT, init_once);
1415 if (hugetlbfs_inode_cachep == NULL)
1418 error = register_filesystem(&hugetlbfs_fs_type);
1423 for_each_hstate(h) {
1425 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1427 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1428 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1431 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1432 pr_err("Cannot mount internal hugetlbfs for "
1433 "page size %uK", ps_kb);
1434 error = PTR_ERR(hugetlbfs_vfsmount[i]);
1435 hugetlbfs_vfsmount[i] = NULL;
1439 /* Non default hstates are optional */
1440 if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1444 kmem_cache_destroy(hugetlbfs_inode_cachep);
1448 fs_initcall(init_hugetlbfs_fs)