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.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 */
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 *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 long lookup_nr = PAGEVEC_SIZE;
422 bool truncate_op = (lend == LLONG_MAX);
424 memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
425 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
426 pagevec_init(&pvec, 0);
430 * Don't grab more pages than the number left in the range.
432 if (end - next < lookup_nr)
433 lookup_nr = end - next;
436 * When no more pages are found, we are done.
438 if (!pagevec_lookup(&pvec, mapping, next, lookup_nr))
441 for (i = 0; i < pagevec_count(&pvec); ++i) {
442 struct page *page = pvec.pages[i];
446 * The page (index) could be beyond end. This is
447 * only possible in the punch hole case as end is
448 * max page offset in the truncate case.
454 hash = hugetlb_fault_mutex_hash(h, mapping, next);
455 mutex_lock(&hugetlb_fault_mutex_table[hash]);
458 * If page is mapped, it was faulted in after being
459 * unmapped in caller. Unmap (again) now after taking
460 * the fault mutex. The mutex will prevent faults
461 * until we finish removing the page.
463 * This race can only happen in the hole punch case.
464 * Getting here in a truncate operation is a bug.
466 if (unlikely(page_mapped(page))) {
469 i_mmap_lock_write(mapping);
470 hugetlb_vmdelete_list(&mapping->i_mmap,
471 next * pages_per_huge_page(h),
472 (next + 1) * pages_per_huge_page(h));
473 i_mmap_unlock_write(mapping);
478 * We must free the huge page and remove from page
479 * cache (remove_huge_page) BEFORE removing the
480 * region/reserve map (hugetlb_unreserve_pages). In
481 * rare out of memory conditions, removal of the
482 * region/reserve map could fail. Correspondingly,
483 * the subpool and global reserve usage count can need
486 VM_BUG_ON(PagePrivate(page));
487 remove_huge_page(page);
490 if (unlikely(hugetlb_unreserve_pages(inode,
492 hugetlb_fix_reserve_counts(inode);
496 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
499 huge_pagevec_release(&pvec);
504 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
507 static void hugetlbfs_evict_inode(struct inode *inode)
509 struct resv_map *resv_map;
511 remove_inode_hugepages(inode, 0, LLONG_MAX);
512 resv_map = (struct resv_map *)inode->i_mapping->private_data;
513 /* root inode doesn't have the resv_map, so we should check it */
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_size_write(inode, offset);
529 i_mmap_lock_write(mapping);
530 if (!RB_EMPTY_ROOT(&mapping->i_mmap))
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;
554 i_mmap_lock_write(mapping);
555 if (!RB_EMPTY_ROOT(&mapping->i_mmap))
556 hugetlb_vmdelete_list(&mapping->i_mmap,
557 hole_start >> PAGE_SHIFT,
558 hole_end >> PAGE_SHIFT);
559 i_mmap_unlock_write(mapping);
560 remove_inode_hugepages(inode, hole_start, hole_end);
567 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
570 struct inode *inode = file_inode(file);
571 struct address_space *mapping = inode->i_mapping;
572 struct hstate *h = hstate_inode(inode);
573 struct vm_area_struct pseudo_vma;
574 loff_t hpage_size = huge_page_size(h);
575 unsigned long hpage_shift = huge_page_shift(h);
576 pgoff_t start, index, end;
580 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
583 if (mode & FALLOC_FL_PUNCH_HOLE)
584 return hugetlbfs_punch_hole(inode, offset, len);
587 * Default preallocate case.
588 * For this range, start is rounded down and end is rounded up
589 * as well as being converted to page offsets.
591 start = offset >> hpage_shift;
592 end = (offset + len + hpage_size - 1) >> hpage_shift;
596 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
597 error = inode_newsize_ok(inode, offset + len);
602 * Initialize a pseudo vma as this is required by the huge page
603 * allocation routines. If NUMA is configured, use page index
604 * as input to create an allocation policy.
606 memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
607 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
608 pseudo_vma.vm_file = file;
610 for (index = start; index < end; index++) {
612 * This is supposed to be the vaddr where the page is being
613 * faulted in, but we have no vaddr here.
617 int avoid_reserve = 0;
622 * fallocate(2) manpage permits EINTR; we may have been
623 * interrupted because we are using up too much memory.
625 if (signal_pending(current)) {
630 /* Set numa allocation policy based on index */
631 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
633 /* addr is the offset within the file (zero based) */
634 addr = index * hpage_size;
636 /* mutex taken here, fault path and hole punch */
637 hash = hugetlb_fault_mutex_hash(h, mapping, index);
638 mutex_lock(&hugetlb_fault_mutex_table[hash]);
640 /* See if already present in mapping to avoid alloc/free */
641 page = find_get_page(mapping, index);
644 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
645 hugetlb_drop_vma_policy(&pseudo_vma);
649 /* Allocate page and add to page cache */
650 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
651 hugetlb_drop_vma_policy(&pseudo_vma);
653 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
654 error = PTR_ERR(page);
657 clear_huge_page(page, addr, pages_per_huge_page(h));
658 __SetPageUptodate(page);
659 error = huge_add_to_page_cache(page, mapping, index);
660 if (unlikely(error)) {
662 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
666 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
668 set_page_huge_active(page);
670 * put_page() due to reference from alloc_huge_page()
671 * unlock_page because locked by add_to_page_cache()
677 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
678 i_size_write(inode, offset + len);
679 inode->i_ctime = current_time(inode);
685 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
687 struct inode *inode = d_inode(dentry);
688 struct hstate *h = hstate_inode(inode);
690 unsigned int ia_valid = attr->ia_valid;
694 error = setattr_prepare(dentry, attr);
698 if (ia_valid & ATTR_SIZE) {
700 if (attr->ia_size & ~huge_page_mask(h))
702 error = hugetlb_vmtruncate(inode, attr->ia_size);
707 setattr_copy(inode, attr);
708 mark_inode_dirty(inode);
712 static struct inode *hugetlbfs_get_root(struct super_block *sb,
713 struct hugetlbfs_config *config)
717 inode = new_inode(sb);
719 inode->i_ino = get_next_ino();
720 inode->i_mode = S_IFDIR | config->mode;
721 inode->i_uid = config->uid;
722 inode->i_gid = config->gid;
723 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
724 inode->i_op = &hugetlbfs_dir_inode_operations;
725 inode->i_fop = &simple_dir_operations;
726 /* directory inodes start off with i_nlink == 2 (for "." entry) */
728 lockdep_annotate_inode_mutex_key(inode);
734 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
735 * be taken from reclaim -- unlike regular filesystems. This needs an
736 * annotation because huge_pmd_share() does an allocation under hugetlb's
739 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
741 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
743 umode_t mode, dev_t dev)
746 struct resv_map *resv_map = NULL;
749 * Reserve maps are only needed for inodes that can have associated
752 if (S_ISREG(mode) || S_ISLNK(mode)) {
753 resv_map = resv_map_alloc();
758 inode = new_inode(sb);
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(inode);
766 inode->i_mapping->private_data = resv_map;
767 switch (mode & S_IFMT) {
769 init_special_inode(inode, mode, dev);
772 inode->i_op = &hugetlbfs_inode_operations;
773 inode->i_fop = &hugetlbfs_file_operations;
776 inode->i_op = &hugetlbfs_dir_inode_operations;
777 inode->i_fop = &simple_dir_operations;
779 /* directory inodes start off with i_nlink == 2 (for "." entry) */
783 inode->i_op = &page_symlink_inode_operations;
784 inode_nohighmem(inode);
787 lockdep_annotate_inode_mutex_key(inode);
790 kref_put(&resv_map->refs, resv_map_release);
797 * File creation. Allocate an inode, and we're done..
799 static int hugetlbfs_mknod(struct inode *dir,
800 struct dentry *dentry, umode_t mode, dev_t dev)
805 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
807 dir->i_ctime = dir->i_mtime = current_time(dir);
808 d_instantiate(dentry, inode);
809 dget(dentry); /* Extra count - pin the dentry in core */
815 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
817 int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
823 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
825 return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
828 static int hugetlbfs_symlink(struct inode *dir,
829 struct dentry *dentry, const char *symname)
834 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
836 int l = strlen(symname)+1;
837 error = page_symlink(inode, symname, l);
839 d_instantiate(dentry, inode);
844 dir->i_ctime = dir->i_mtime = current_time(dir);
850 * mark the head page dirty
852 static int hugetlbfs_set_page_dirty(struct page *page)
854 struct page *head = compound_head(page);
860 static int hugetlbfs_migrate_page(struct address_space *mapping,
861 struct page *newpage, struct page *page,
862 enum migrate_mode mode)
866 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
867 if (rc != MIGRATEPAGE_SUCCESS)
871 * page_private is subpool pointer in hugetlb pages. Transfer to
872 * new page. PagePrivate is not associated with page_private for
873 * hugetlb pages and can not be set here as only page_huge_active
874 * pages can be migrated.
876 if (page_private(page)) {
877 set_page_private(newpage, page_private(page));
878 set_page_private(page, 0);
881 migrate_page_copy(newpage, page);
883 return MIGRATEPAGE_SUCCESS;
886 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
888 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
889 struct hstate *h = hstate_inode(d_inode(dentry));
891 buf->f_type = HUGETLBFS_MAGIC;
892 buf->f_bsize = huge_page_size(h);
894 spin_lock(&sbinfo->stat_lock);
895 /* If no limits set, just report 0 for max/free/used
896 * blocks, like simple_statfs() */
900 spin_lock(&sbinfo->spool->lock);
901 buf->f_blocks = sbinfo->spool->max_hpages;
902 free_pages = sbinfo->spool->max_hpages
903 - sbinfo->spool->used_hpages;
904 buf->f_bavail = buf->f_bfree = free_pages;
905 spin_unlock(&sbinfo->spool->lock);
906 buf->f_files = sbinfo->max_inodes;
907 buf->f_ffree = sbinfo->free_inodes;
909 spin_unlock(&sbinfo->stat_lock);
911 buf->f_namelen = NAME_MAX;
915 static void hugetlbfs_put_super(struct super_block *sb)
917 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
920 sb->s_fs_info = NULL;
923 hugepage_put_subpool(sbi->spool);
929 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
931 if (sbinfo->free_inodes >= 0) {
932 spin_lock(&sbinfo->stat_lock);
933 if (unlikely(!sbinfo->free_inodes)) {
934 spin_unlock(&sbinfo->stat_lock);
937 sbinfo->free_inodes--;
938 spin_unlock(&sbinfo->stat_lock);
944 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
946 if (sbinfo->free_inodes >= 0) {
947 spin_lock(&sbinfo->stat_lock);
948 sbinfo->free_inodes++;
949 spin_unlock(&sbinfo->stat_lock);
954 static struct kmem_cache *hugetlbfs_inode_cachep;
956 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
958 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
959 struct hugetlbfs_inode_info *p;
961 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
963 p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
965 hugetlbfs_inc_free_inodes(sbinfo);
970 * Any time after allocation, hugetlbfs_destroy_inode can be called
971 * for the inode. mpol_free_shared_policy is unconditionally called
972 * as part of hugetlbfs_destroy_inode. So, initialize policy here
973 * in case of a quick call to destroy.
975 * Note that the policy is initialized even if we are creating a
976 * private inode. This simplifies hugetlbfs_destroy_inode.
978 mpol_shared_policy_init(&p->policy, NULL);
980 return &p->vfs_inode;
983 static void hugetlbfs_i_callback(struct rcu_head *head)
985 struct inode *inode = container_of(head, struct inode, i_rcu);
986 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
989 static void hugetlbfs_destroy_inode(struct inode *inode)
991 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
992 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
993 call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
996 static const struct address_space_operations hugetlbfs_aops = {
997 .write_begin = hugetlbfs_write_begin,
998 .write_end = hugetlbfs_write_end,
999 .set_page_dirty = hugetlbfs_set_page_dirty,
1000 .migratepage = hugetlbfs_migrate_page,
1004 static void init_once(void *foo)
1006 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1008 inode_init_once(&ei->vfs_inode);
1011 const struct file_operations hugetlbfs_file_operations = {
1012 .read_iter = hugetlbfs_read_iter,
1013 .mmap = hugetlbfs_file_mmap,
1014 .fsync = noop_fsync,
1015 .get_unmapped_area = hugetlb_get_unmapped_area,
1016 .llseek = default_llseek,
1017 .fallocate = hugetlbfs_fallocate,
1020 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1021 .create = hugetlbfs_create,
1022 .lookup = simple_lookup,
1023 .link = simple_link,
1024 .unlink = simple_unlink,
1025 .symlink = hugetlbfs_symlink,
1026 .mkdir = hugetlbfs_mkdir,
1027 .rmdir = simple_rmdir,
1028 .mknod = hugetlbfs_mknod,
1029 .rename = simple_rename,
1030 .setattr = hugetlbfs_setattr,
1033 static const struct inode_operations hugetlbfs_inode_operations = {
1034 .setattr = hugetlbfs_setattr,
1037 static const struct super_operations hugetlbfs_ops = {
1038 .alloc_inode = hugetlbfs_alloc_inode,
1039 .destroy_inode = hugetlbfs_destroy_inode,
1040 .evict_inode = hugetlbfs_evict_inode,
1041 .statfs = hugetlbfs_statfs,
1042 .put_super = hugetlbfs_put_super,
1043 .show_options = generic_show_options,
1046 enum { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1049 * Convert size option passed from command line to number of huge pages
1050 * in the pool specified by hstate. Size option could be in bytes
1051 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1054 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1057 if (val_type == NO_SIZE)
1060 if (val_type == SIZE_PERCENT) {
1061 size_opt <<= huge_page_shift(h);
1062 size_opt *= h->max_huge_pages;
1063 do_div(size_opt, 100);
1066 size_opt >>= huge_page_shift(h);
1071 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1074 substring_t args[MAX_OPT_ARGS];
1076 unsigned long long max_size_opt = 0, min_size_opt = 0;
1077 int max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1082 while ((p = strsep(&options, ",")) != NULL) {
1087 token = match_token(p, tokens, args);
1090 if (match_int(&args[0], &option))
1092 pconfig->uid = make_kuid(current_user_ns(), option);
1093 if (!uid_valid(pconfig->uid))
1098 if (match_int(&args[0], &option))
1100 pconfig->gid = make_kgid(current_user_ns(), option);
1101 if (!gid_valid(pconfig->gid))
1106 if (match_octal(&args[0], &option))
1108 pconfig->mode = option & 01777U;
1112 /* memparse() will accept a K/M/G without a digit */
1113 if (!isdigit(*args[0].from))
1115 max_size_opt = memparse(args[0].from, &rest);
1116 max_val_type = SIZE_STD;
1118 max_val_type = SIZE_PERCENT;
1123 /* memparse() will accept a K/M/G without a digit */
1124 if (!isdigit(*args[0].from))
1126 pconfig->nr_inodes = memparse(args[0].from, &rest);
1129 case Opt_pagesize: {
1131 ps = memparse(args[0].from, &rest);
1132 pconfig->hstate = size_to_hstate(ps);
1133 if (!pconfig->hstate) {
1134 pr_err("Unsupported page size %lu MB\n",
1141 case Opt_min_size: {
1142 /* memparse() will accept a K/M/G without a digit */
1143 if (!isdigit(*args[0].from))
1145 min_size_opt = memparse(args[0].from, &rest);
1146 min_val_type = SIZE_STD;
1148 min_val_type = SIZE_PERCENT;
1153 pr_err("Bad mount option: \"%s\"\n", p);
1160 * Use huge page pool size (in hstate) to convert the size
1161 * options to number of huge pages. If NO_SIZE, -1 is returned.
1163 pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1164 max_size_opt, max_val_type);
1165 pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1166 min_size_opt, min_val_type);
1169 * If max_size was specified, then min_size must be smaller
1171 if (max_val_type > NO_SIZE &&
1172 pconfig->min_hpages > pconfig->max_hpages) {
1173 pr_err("minimum size can not be greater than maximum size\n");
1180 pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1185 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1188 struct hugetlbfs_config config;
1189 struct hugetlbfs_sb_info *sbinfo;
1191 save_mount_options(sb, data);
1193 config.max_hpages = -1; /* No limit on size by default */
1194 config.nr_inodes = -1; /* No limit on number of inodes by default */
1195 config.uid = current_fsuid();
1196 config.gid = current_fsgid();
1198 config.hstate = &default_hstate;
1199 config.min_hpages = -1; /* No default minimum size */
1200 ret = hugetlbfs_parse_options(data, &config);
1204 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1207 sb->s_fs_info = sbinfo;
1208 sbinfo->hstate = config.hstate;
1209 spin_lock_init(&sbinfo->stat_lock);
1210 sbinfo->max_inodes = config.nr_inodes;
1211 sbinfo->free_inodes = config.nr_inodes;
1212 sbinfo->spool = NULL;
1214 * Allocate and initialize subpool if maximum or minimum size is
1215 * specified. Any needed reservations (for minimim size) are taken
1216 * taken when the subpool is created.
1218 if (config.max_hpages != -1 || config.min_hpages != -1) {
1219 sbinfo->spool = hugepage_new_subpool(config.hstate,
1225 sb->s_maxbytes = MAX_LFS_FILESIZE;
1226 sb->s_blocksize = huge_page_size(config.hstate);
1227 sb->s_blocksize_bits = huge_page_shift(config.hstate);
1228 sb->s_magic = HUGETLBFS_MAGIC;
1229 sb->s_op = &hugetlbfs_ops;
1230 sb->s_time_gran = 1;
1231 sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1236 kfree(sbinfo->spool);
1241 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1242 int flags, const char *dev_name, void *data)
1244 return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1247 static struct file_system_type hugetlbfs_fs_type = {
1248 .name = "hugetlbfs",
1249 .mount = hugetlbfs_mount,
1250 .kill_sb = kill_litter_super,
1253 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1255 static int can_do_hugetlb_shm(void)
1258 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1259 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1262 static int get_hstate_idx(int page_size_log)
1264 struct hstate *h = hstate_sizelog(page_size_log);
1271 static const struct dentry_operations anon_ops = {
1272 .d_dname = simple_dname
1276 * Note that size should be aligned to proper hugepage size in caller side,
1277 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1279 struct file *hugetlb_file_setup(const char *name, size_t size,
1280 vm_flags_t acctflag, struct user_struct **user,
1281 int creat_flags, int page_size_log)
1283 struct file *file = ERR_PTR(-ENOMEM);
1284 struct inode *inode;
1286 struct super_block *sb;
1287 struct qstr quick_string;
1290 hstate_idx = get_hstate_idx(page_size_log);
1292 return ERR_PTR(-ENODEV);
1295 if (!hugetlbfs_vfsmount[hstate_idx])
1296 return ERR_PTR(-ENOENT);
1298 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1299 *user = current_user();
1300 if (user_shm_lock(size, *user)) {
1302 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1303 current->comm, current->pid);
1304 task_unlock(current);
1307 return ERR_PTR(-EPERM);
1311 sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1312 quick_string.name = name;
1313 quick_string.len = strlen(quick_string.name);
1314 quick_string.hash = 0;
1315 path.dentry = d_alloc_pseudo(sb, &quick_string);
1317 goto out_shm_unlock;
1319 d_set_d_op(path.dentry, &anon_ops);
1320 path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1321 file = ERR_PTR(-ENOSPC);
1322 inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1325 if (creat_flags == HUGETLB_SHMFS_INODE)
1326 inode->i_flags |= S_PRIVATE;
1328 file = ERR_PTR(-ENOMEM);
1329 if (hugetlb_reserve_pages(inode, 0,
1330 size >> huge_page_shift(hstate_inode(inode)), NULL,
1334 d_instantiate(path.dentry, inode);
1335 inode->i_size = size;
1338 file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1339 &hugetlbfs_file_operations);
1341 goto out_dentry; /* inode is already attached */
1351 user_shm_unlock(size, *user);
1357 static int __init init_hugetlbfs_fs(void)
1363 if (!hugepages_supported()) {
1364 pr_info("disabling because there are no supported hugepage sizes\n");
1369 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1370 sizeof(struct hugetlbfs_inode_info),
1371 0, SLAB_ACCOUNT, init_once);
1372 if (hugetlbfs_inode_cachep == NULL)
1375 error = register_filesystem(&hugetlbfs_fs_type);
1380 for_each_hstate(h) {
1382 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1384 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1385 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1388 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1389 pr_err("Cannot mount internal hugetlbfs for "
1390 "page size %uK", ps_kb);
1391 error = PTR_ERR(hugetlbfs_vfsmount[i]);
1392 hugetlbfs_vfsmount[i] = NULL;
1396 /* Non default hstates are optional */
1397 if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1401 kmem_cache_destroy(hugetlbfs_inode_cachep);
1405 fs_initcall(init_hugetlbfs_fs)