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);
121 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
123 struct inode *inode = file_inode(file);
126 struct hstate *h = hstate_file(file);
129 * vma address alignment (but not the pgoff alignment) has
130 * already been checked by prepare_hugepage_range. If you add
131 * any error returns here, do so after setting VM_HUGETLB, so
132 * is_vm_hugetlb_page tests below unmap_region go the right
133 * way when do_mmap_pgoff unwinds (may be important on powerpc
136 vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
137 vma->vm_ops = &hugetlb_vm_ops;
140 * Offset passed to mmap (before page shift) could have been
141 * negative when represented as a (l)off_t.
143 if (((loff_t)vma->vm_pgoff << PAGE_SHIFT) < 0)
146 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
149 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
150 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
151 /* check for overflow */
159 if (hugetlb_reserve_pages(inode,
160 vma->vm_pgoff >> huge_page_order(h),
161 len >> huge_page_shift(h), vma,
166 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
167 i_size_write(inode, len);
175 * Called under down_write(mmap_sem).
178 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
180 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
181 unsigned long len, unsigned long pgoff, unsigned long flags)
183 struct mm_struct *mm = current->mm;
184 struct vm_area_struct *vma;
185 struct hstate *h = hstate_file(file);
186 struct vm_unmapped_area_info info;
188 if (len & ~huge_page_mask(h))
193 if (flags & MAP_FIXED) {
194 if (prepare_hugepage_range(file, addr, len))
200 addr = ALIGN(addr, huge_page_size(h));
201 vma = find_vma(mm, addr);
202 if (TASK_SIZE - len >= addr &&
203 (!vma || addr + len <= vm_start_gap(vma)))
209 info.low_limit = TASK_UNMAPPED_BASE;
210 info.high_limit = TASK_SIZE;
211 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
212 info.align_offset = 0;
213 return vm_unmapped_area(&info);
218 hugetlbfs_read_actor(struct page *page, unsigned long offset,
219 struct iov_iter *to, unsigned long size)
224 /* Find which 4k chunk and offset with in that chunk */
225 i = offset >> PAGE_SHIFT;
226 offset = offset & ~PAGE_MASK;
230 chunksize = PAGE_SIZE;
233 if (chunksize > size)
235 n = copy_page_to_iter(&page[i], offset, chunksize, to);
247 * Support for read() - Find the page attached to f_mapping and copy out the
248 * data. Its *very* similar to do_generic_mapping_read(), we can't use that
249 * since it has PAGE_SIZE assumptions.
251 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
253 struct file *file = iocb->ki_filp;
254 struct hstate *h = hstate_file(file);
255 struct address_space *mapping = file->f_mapping;
256 struct inode *inode = mapping->host;
257 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
258 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
259 unsigned long end_index;
263 while (iov_iter_count(to)) {
267 /* nr is the maximum number of bytes to copy from this page */
268 nr = huge_page_size(h);
269 isize = i_size_read(inode);
272 end_index = (isize - 1) >> huge_page_shift(h);
273 if (index > end_index)
275 if (index == end_index) {
276 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
283 page = find_lock_page(mapping, index);
284 if (unlikely(page == NULL)) {
286 * We have a HOLE, zero out the user-buffer for the
287 * length of the hole or request.
289 copied = iov_iter_zero(nr, to);
294 * We have the page, copy it to user space buffer.
296 copied = hugetlbfs_read_actor(page, offset, to, nr);
301 if (copied != nr && iov_iter_count(to)) {
306 index += offset >> huge_page_shift(h);
307 offset &= ~huge_page_mask(h);
309 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
313 static int hugetlbfs_write_begin(struct file *file,
314 struct address_space *mapping,
315 loff_t pos, unsigned len, unsigned flags,
316 struct page **pagep, void **fsdata)
321 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
322 loff_t pos, unsigned len, unsigned copied,
323 struct page *page, void *fsdata)
329 static void remove_huge_page(struct page *page)
331 ClearPageDirty(page);
332 ClearPageUptodate(page);
333 delete_from_page_cache(page);
337 hugetlb_vmdelete_list(struct rb_root *root, pgoff_t start, pgoff_t end)
339 struct vm_area_struct *vma;
342 * end == 0 indicates that the entire range after
343 * start should be unmapped.
345 vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
346 unsigned long v_offset;
350 * Can the expression below overflow on 32-bit arches?
351 * No, because the interval tree returns us only those vmas
352 * which overlap the truncated area starting at pgoff,
353 * and no vma on a 32-bit arch can span beyond the 4GB.
355 if (vma->vm_pgoff < start)
356 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
363 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
365 if (v_end > vma->vm_end)
369 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
375 * remove_inode_hugepages handles two distinct cases: truncation and hole
376 * punch. There are subtle differences in operation for each case.
378 * truncation is indicated by end of range being LLONG_MAX
379 * In this case, we first scan the range and release found pages.
380 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
381 * maps and global counts. Page faults can not race with truncation
382 * in this routine. hugetlb_no_page() prevents page faults in the
383 * truncated range. It checks i_size before allocation, and again after
384 * with the page table lock for the page held. The same lock must be
385 * acquired to unmap a page.
386 * hole punch is indicated if end is not LLONG_MAX
387 * In the hole punch case we scan the range and release found pages.
388 * Only when releasing a page is the associated region/reserv map
389 * deleted. The region/reserv map for ranges without associated
390 * pages are not modified. Page faults can race with hole punch.
391 * This is indicated if we find a mapped page.
392 * Note: If the passed end of range value is beyond the end of file, but
393 * not LLONG_MAX this routine still performs a hole punch operation.
395 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
398 struct hstate *h = hstate_inode(inode);
399 struct address_space *mapping = &inode->i_data;
400 const pgoff_t start = lstart >> huge_page_shift(h);
401 const pgoff_t end = lend >> huge_page_shift(h);
402 struct vm_area_struct pseudo_vma;
406 long lookup_nr = PAGEVEC_SIZE;
407 bool truncate_op = (lend == LLONG_MAX);
409 memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
410 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
411 pagevec_init(&pvec, 0);
415 * Don't grab more pages than the number left in the range.
417 if (end - next < lookup_nr)
418 lookup_nr = end - next;
421 * When no more pages are found, we are done.
423 if (!pagevec_lookup(&pvec, mapping, next, lookup_nr))
426 for (i = 0; i < pagevec_count(&pvec); ++i) {
427 struct page *page = pvec.pages[i];
431 * The page (index) could be beyond end. This is
432 * only possible in the punch hole case as end is
433 * max page offset in the truncate case.
439 hash = hugetlb_fault_mutex_hash(h, current->mm,
442 mutex_lock(&hugetlb_fault_mutex_table[hash]);
445 * If page is mapped, it was faulted in after being
446 * unmapped in caller. Unmap (again) now after taking
447 * the fault mutex. The mutex will prevent faults
448 * until we finish removing the page.
450 * This race can only happen in the hole punch case.
451 * Getting here in a truncate operation is a bug.
453 if (unlikely(page_mapped(page))) {
456 i_mmap_lock_write(mapping);
457 hugetlb_vmdelete_list(&mapping->i_mmap,
458 next * pages_per_huge_page(h),
459 (next + 1) * pages_per_huge_page(h));
460 i_mmap_unlock_write(mapping);
465 * We must free the huge page and remove from page
466 * cache (remove_huge_page) BEFORE removing the
467 * region/reserve map (hugetlb_unreserve_pages). In
468 * rare out of memory conditions, removal of the
469 * region/reserve map could fail. Correspondingly,
470 * the subpool and global reserve usage count can need
473 VM_BUG_ON(PagePrivate(page));
474 remove_huge_page(page);
477 if (unlikely(hugetlb_unreserve_pages(inode,
479 hugetlb_fix_reserve_counts(inode);
483 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
486 huge_pagevec_release(&pvec);
491 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
494 static void hugetlbfs_evict_inode(struct inode *inode)
496 struct resv_map *resv_map;
498 remove_inode_hugepages(inode, 0, LLONG_MAX);
499 resv_map = (struct resv_map *)inode->i_mapping->private_data;
500 /* root inode doesn't have the resv_map, so we should check it */
502 resv_map_release(&resv_map->refs);
506 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
509 struct address_space *mapping = inode->i_mapping;
510 struct hstate *h = hstate_inode(inode);
512 BUG_ON(offset & ~huge_page_mask(h));
513 pgoff = offset >> PAGE_SHIFT;
515 i_size_write(inode, offset);
516 i_mmap_lock_write(mapping);
517 if (!RB_EMPTY_ROOT(&mapping->i_mmap))
518 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
519 i_mmap_unlock_write(mapping);
520 remove_inode_hugepages(inode, offset, LLONG_MAX);
524 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
526 struct hstate *h = hstate_inode(inode);
527 loff_t hpage_size = huge_page_size(h);
528 loff_t hole_start, hole_end;
531 * For hole punch round up the beginning offset of the hole and
532 * round down the end.
534 hole_start = round_up(offset, hpage_size);
535 hole_end = round_down(offset + len, hpage_size);
537 if (hole_end > hole_start) {
538 struct address_space *mapping = inode->i_mapping;
541 i_mmap_lock_write(mapping);
542 if (!RB_EMPTY_ROOT(&mapping->i_mmap))
543 hugetlb_vmdelete_list(&mapping->i_mmap,
544 hole_start >> PAGE_SHIFT,
545 hole_end >> PAGE_SHIFT);
546 i_mmap_unlock_write(mapping);
547 remove_inode_hugepages(inode, hole_start, hole_end);
554 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
557 struct inode *inode = file_inode(file);
558 struct address_space *mapping = inode->i_mapping;
559 struct hstate *h = hstate_inode(inode);
560 struct vm_area_struct pseudo_vma;
561 struct mm_struct *mm = current->mm;
562 loff_t hpage_size = huge_page_size(h);
563 unsigned long hpage_shift = huge_page_shift(h);
564 pgoff_t start, index, end;
568 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
571 if (mode & FALLOC_FL_PUNCH_HOLE)
572 return hugetlbfs_punch_hole(inode, offset, len);
575 * Default preallocate case.
576 * For this range, start is rounded down and end is rounded up
577 * as well as being converted to page offsets.
579 start = offset >> hpage_shift;
580 end = (offset + len + hpage_size - 1) >> hpage_shift;
584 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
585 error = inode_newsize_ok(inode, offset + len);
590 * Initialize a pseudo vma as this is required by the huge page
591 * allocation routines. If NUMA is configured, use page index
592 * as input to create an allocation policy.
594 memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
595 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
596 pseudo_vma.vm_file = file;
598 for (index = start; index < end; index++) {
600 * This is supposed to be the vaddr where the page is being
601 * faulted in, but we have no vaddr here.
605 int avoid_reserve = 0;
610 * fallocate(2) manpage permits EINTR; we may have been
611 * interrupted because we are using up too much memory.
613 if (signal_pending(current)) {
618 /* Set numa allocation policy based on index */
619 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
621 /* addr is the offset within the file (zero based) */
622 addr = index * hpage_size;
624 /* mutex taken here, fault path and hole punch */
625 hash = hugetlb_fault_mutex_hash(h, mm, &pseudo_vma, mapping,
627 mutex_lock(&hugetlb_fault_mutex_table[hash]);
629 /* See if already present in mapping to avoid alloc/free */
630 page = find_get_page(mapping, index);
633 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
634 hugetlb_drop_vma_policy(&pseudo_vma);
638 /* Allocate page and add to page cache */
639 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
640 hugetlb_drop_vma_policy(&pseudo_vma);
642 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
643 error = PTR_ERR(page);
646 clear_huge_page(page, addr, pages_per_huge_page(h));
647 __SetPageUptodate(page);
648 error = huge_add_to_page_cache(page, mapping, index);
649 if (unlikely(error)) {
651 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
655 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
658 * page_put due to reference from alloc_huge_page()
659 * unlock_page because locked by add_to_page_cache()
665 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
666 i_size_write(inode, offset + len);
667 inode->i_ctime = current_time(inode);
673 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
675 struct inode *inode = d_inode(dentry);
676 struct hstate *h = hstate_inode(inode);
678 unsigned int ia_valid = attr->ia_valid;
682 error = setattr_prepare(dentry, attr);
686 if (ia_valid & ATTR_SIZE) {
688 if (attr->ia_size & ~huge_page_mask(h))
690 error = hugetlb_vmtruncate(inode, attr->ia_size);
695 setattr_copy(inode, attr);
696 mark_inode_dirty(inode);
700 static struct inode *hugetlbfs_get_root(struct super_block *sb,
701 struct hugetlbfs_config *config)
705 inode = new_inode(sb);
707 inode->i_ino = get_next_ino();
708 inode->i_mode = S_IFDIR | config->mode;
709 inode->i_uid = config->uid;
710 inode->i_gid = config->gid;
711 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
712 inode->i_op = &hugetlbfs_dir_inode_operations;
713 inode->i_fop = &simple_dir_operations;
714 /* directory inodes start off with i_nlink == 2 (for "." entry) */
716 lockdep_annotate_inode_mutex_key(inode);
722 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
723 * be taken from reclaim -- unlike regular filesystems. This needs an
724 * annotation because huge_pmd_share() does an allocation under hugetlb's
727 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
729 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
731 umode_t mode, dev_t dev)
734 struct resv_map *resv_map;
736 resv_map = resv_map_alloc();
740 inode = new_inode(sb);
742 inode->i_ino = get_next_ino();
743 inode_init_owner(inode, dir, mode);
744 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
745 &hugetlbfs_i_mmap_rwsem_key);
746 inode->i_mapping->a_ops = &hugetlbfs_aops;
747 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
748 inode->i_mapping->private_data = resv_map;
749 switch (mode & S_IFMT) {
751 init_special_inode(inode, mode, dev);
754 inode->i_op = &hugetlbfs_inode_operations;
755 inode->i_fop = &hugetlbfs_file_operations;
758 inode->i_op = &hugetlbfs_dir_inode_operations;
759 inode->i_fop = &simple_dir_operations;
761 /* directory inodes start off with i_nlink == 2 (for "." entry) */
765 inode->i_op = &page_symlink_inode_operations;
766 inode_nohighmem(inode);
769 lockdep_annotate_inode_mutex_key(inode);
771 kref_put(&resv_map->refs, resv_map_release);
777 * File creation. Allocate an inode, and we're done..
779 static int hugetlbfs_mknod(struct inode *dir,
780 struct dentry *dentry, umode_t mode, dev_t dev)
785 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
787 dir->i_ctime = dir->i_mtime = current_time(dir);
788 d_instantiate(dentry, inode);
789 dget(dentry); /* Extra count - pin the dentry in core */
795 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
797 int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
803 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
805 return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
808 static int hugetlbfs_symlink(struct inode *dir,
809 struct dentry *dentry, const char *symname)
814 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
816 int l = strlen(symname)+1;
817 error = page_symlink(inode, symname, l);
819 d_instantiate(dentry, inode);
824 dir->i_ctime = dir->i_mtime = current_time(dir);
830 * mark the head page dirty
832 static int hugetlbfs_set_page_dirty(struct page *page)
834 struct page *head = compound_head(page);
840 static int hugetlbfs_migrate_page(struct address_space *mapping,
841 struct page *newpage, struct page *page,
842 enum migrate_mode mode)
846 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
847 if (rc != MIGRATEPAGE_SUCCESS)
849 migrate_page_copy(newpage, page);
851 return MIGRATEPAGE_SUCCESS;
854 static int hugetlbfs_error_remove_page(struct address_space *mapping,
857 struct inode *inode = mapping->host;
859 remove_huge_page(page);
860 hugetlb_fix_reserve_counts(inode);
865 * Display the mount options in /proc/mounts.
867 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
869 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
870 struct hugepage_subpool *spool = sbinfo->spool;
871 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
872 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
875 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
876 seq_printf(m, ",uid=%u",
877 from_kuid_munged(&init_user_ns, sbinfo->uid));
878 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
879 seq_printf(m, ",gid=%u",
880 from_kgid_munged(&init_user_ns, sbinfo->gid));
881 if (sbinfo->mode != 0755)
882 seq_printf(m, ",mode=%o", sbinfo->mode);
883 if (sbinfo->max_inodes != -1)
884 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
888 if (hpage_size >= 1024) {
892 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
894 if (spool->max_hpages != -1)
895 seq_printf(m, ",size=%llu",
896 (unsigned long long)spool->max_hpages << hpage_shift);
897 if (spool->min_hpages != -1)
898 seq_printf(m, ",min_size=%llu",
899 (unsigned long long)spool->min_hpages << hpage_shift);
904 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
906 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
907 struct hstate *h = hstate_inode(d_inode(dentry));
909 buf->f_type = HUGETLBFS_MAGIC;
910 buf->f_bsize = huge_page_size(h);
912 spin_lock(&sbinfo->stat_lock);
913 /* If no limits set, just report 0 for max/free/used
914 * blocks, like simple_statfs() */
918 spin_lock(&sbinfo->spool->lock);
919 buf->f_blocks = sbinfo->spool->max_hpages;
920 free_pages = sbinfo->spool->max_hpages
921 - sbinfo->spool->used_hpages;
922 buf->f_bavail = buf->f_bfree = free_pages;
923 spin_unlock(&sbinfo->spool->lock);
924 buf->f_files = sbinfo->max_inodes;
925 buf->f_ffree = sbinfo->free_inodes;
927 spin_unlock(&sbinfo->stat_lock);
929 buf->f_namelen = NAME_MAX;
933 static void hugetlbfs_put_super(struct super_block *sb)
935 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
938 sb->s_fs_info = NULL;
941 hugepage_put_subpool(sbi->spool);
947 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
949 if (sbinfo->free_inodes >= 0) {
950 spin_lock(&sbinfo->stat_lock);
951 if (unlikely(!sbinfo->free_inodes)) {
952 spin_unlock(&sbinfo->stat_lock);
955 sbinfo->free_inodes--;
956 spin_unlock(&sbinfo->stat_lock);
962 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
964 if (sbinfo->free_inodes >= 0) {
965 spin_lock(&sbinfo->stat_lock);
966 sbinfo->free_inodes++;
967 spin_unlock(&sbinfo->stat_lock);
972 static struct kmem_cache *hugetlbfs_inode_cachep;
974 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
976 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
977 struct hugetlbfs_inode_info *p;
979 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
981 p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
983 hugetlbfs_inc_free_inodes(sbinfo);
988 * Any time after allocation, hugetlbfs_destroy_inode can be called
989 * for the inode. mpol_free_shared_policy is unconditionally called
990 * as part of hugetlbfs_destroy_inode. So, initialize policy here
991 * in case of a quick call to destroy.
993 * Note that the policy is initialized even if we are creating a
994 * private inode. This simplifies hugetlbfs_destroy_inode.
996 mpol_shared_policy_init(&p->policy, NULL);
998 return &p->vfs_inode;
1001 static void hugetlbfs_i_callback(struct rcu_head *head)
1003 struct inode *inode = container_of(head, struct inode, i_rcu);
1004 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1007 static void hugetlbfs_destroy_inode(struct inode *inode)
1009 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1010 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1011 call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
1014 static const struct address_space_operations hugetlbfs_aops = {
1015 .write_begin = hugetlbfs_write_begin,
1016 .write_end = hugetlbfs_write_end,
1017 .set_page_dirty = hugetlbfs_set_page_dirty,
1018 .migratepage = hugetlbfs_migrate_page,
1019 .error_remove_page = hugetlbfs_error_remove_page,
1023 static void init_once(void *foo)
1025 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1027 inode_init_once(&ei->vfs_inode);
1030 const struct file_operations hugetlbfs_file_operations = {
1031 .read_iter = hugetlbfs_read_iter,
1032 .mmap = hugetlbfs_file_mmap,
1033 .fsync = noop_fsync,
1034 .get_unmapped_area = hugetlb_get_unmapped_area,
1035 .llseek = default_llseek,
1036 .fallocate = hugetlbfs_fallocate,
1039 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1040 .create = hugetlbfs_create,
1041 .lookup = simple_lookup,
1042 .link = simple_link,
1043 .unlink = simple_unlink,
1044 .symlink = hugetlbfs_symlink,
1045 .mkdir = hugetlbfs_mkdir,
1046 .rmdir = simple_rmdir,
1047 .mknod = hugetlbfs_mknod,
1048 .rename = simple_rename,
1049 .setattr = hugetlbfs_setattr,
1052 static const struct inode_operations hugetlbfs_inode_operations = {
1053 .setattr = hugetlbfs_setattr,
1056 static const struct super_operations hugetlbfs_ops = {
1057 .alloc_inode = hugetlbfs_alloc_inode,
1058 .destroy_inode = hugetlbfs_destroy_inode,
1059 .evict_inode = hugetlbfs_evict_inode,
1060 .statfs = hugetlbfs_statfs,
1061 .put_super = hugetlbfs_put_super,
1062 .show_options = hugetlbfs_show_options,
1065 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1068 * Convert size option passed from command line to number of huge pages
1069 * in the pool specified by hstate. Size option could be in bytes
1070 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1073 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1074 enum hugetlbfs_size_type val_type)
1076 if (val_type == NO_SIZE)
1079 if (val_type == SIZE_PERCENT) {
1080 size_opt <<= huge_page_shift(h);
1081 size_opt *= h->max_huge_pages;
1082 do_div(size_opt, 100);
1085 size_opt >>= huge_page_shift(h);
1090 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1093 substring_t args[MAX_OPT_ARGS];
1095 unsigned long long max_size_opt = 0, min_size_opt = 0;
1096 enum hugetlbfs_size_type max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1101 while ((p = strsep(&options, ",")) != NULL) {
1106 token = match_token(p, tokens, args);
1109 if (match_int(&args[0], &option))
1111 pconfig->uid = make_kuid(current_user_ns(), option);
1112 if (!uid_valid(pconfig->uid))
1117 if (match_int(&args[0], &option))
1119 pconfig->gid = make_kgid(current_user_ns(), option);
1120 if (!gid_valid(pconfig->gid))
1125 if (match_octal(&args[0], &option))
1127 pconfig->mode = option & 01777U;
1131 /* memparse() will accept a K/M/G without a digit */
1132 if (!isdigit(*args[0].from))
1134 max_size_opt = memparse(args[0].from, &rest);
1135 max_val_type = SIZE_STD;
1137 max_val_type = SIZE_PERCENT;
1142 /* memparse() will accept a K/M/G without a digit */
1143 if (!isdigit(*args[0].from))
1145 pconfig->nr_inodes = memparse(args[0].from, &rest);
1148 case Opt_pagesize: {
1150 ps = memparse(args[0].from, &rest);
1151 pconfig->hstate = size_to_hstate(ps);
1152 if (!pconfig->hstate) {
1153 pr_err("Unsupported page size %lu MB\n",
1160 case Opt_min_size: {
1161 /* memparse() will accept a K/M/G without a digit */
1162 if (!isdigit(*args[0].from))
1164 min_size_opt = memparse(args[0].from, &rest);
1165 min_val_type = SIZE_STD;
1167 min_val_type = SIZE_PERCENT;
1172 pr_err("Bad mount option: \"%s\"\n", p);
1179 * Use huge page pool size (in hstate) to convert the size
1180 * options to number of huge pages. If NO_SIZE, -1 is returned.
1182 pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1183 max_size_opt, max_val_type);
1184 pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1185 min_size_opt, min_val_type);
1188 * If max_size was specified, then min_size must be smaller
1190 if (max_val_type > NO_SIZE &&
1191 pconfig->min_hpages > pconfig->max_hpages) {
1192 pr_err("minimum size can not be greater than maximum size\n");
1199 pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1204 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1207 struct hugetlbfs_config config;
1208 struct hugetlbfs_sb_info *sbinfo;
1210 config.max_hpages = -1; /* No limit on size by default */
1211 config.nr_inodes = -1; /* No limit on number of inodes by default */
1212 config.uid = current_fsuid();
1213 config.gid = current_fsgid();
1215 config.hstate = &default_hstate;
1216 config.min_hpages = -1; /* No default minimum size */
1217 ret = hugetlbfs_parse_options(data, &config);
1221 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1224 sb->s_fs_info = sbinfo;
1225 sbinfo->hstate = config.hstate;
1226 spin_lock_init(&sbinfo->stat_lock);
1227 sbinfo->max_inodes = config.nr_inodes;
1228 sbinfo->free_inodes = config.nr_inodes;
1229 sbinfo->spool = NULL;
1230 sbinfo->uid = config.uid;
1231 sbinfo->gid = config.gid;
1232 sbinfo->mode = config.mode;
1235 * Allocate and initialize subpool if maximum or minimum size is
1236 * specified. Any needed reservations (for minimim size) are taken
1237 * taken when the subpool is created.
1239 if (config.max_hpages != -1 || config.min_hpages != -1) {
1240 sbinfo->spool = hugepage_new_subpool(config.hstate,
1246 sb->s_maxbytes = MAX_LFS_FILESIZE;
1247 sb->s_blocksize = huge_page_size(config.hstate);
1248 sb->s_blocksize_bits = huge_page_shift(config.hstate);
1249 sb->s_magic = HUGETLBFS_MAGIC;
1250 sb->s_op = &hugetlbfs_ops;
1251 sb->s_time_gran = 1;
1252 sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1257 kfree(sbinfo->spool);
1262 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1263 int flags, const char *dev_name, void *data)
1265 return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1268 static struct file_system_type hugetlbfs_fs_type = {
1269 .name = "hugetlbfs",
1270 .mount = hugetlbfs_mount,
1271 .kill_sb = kill_litter_super,
1274 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1276 static int can_do_hugetlb_shm(void)
1279 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1280 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1283 static int get_hstate_idx(int page_size_log)
1285 struct hstate *h = hstate_sizelog(page_size_log);
1292 static const struct dentry_operations anon_ops = {
1293 .d_dname = simple_dname
1297 * Note that size should be aligned to proper hugepage size in caller side,
1298 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1300 struct file *hugetlb_file_setup(const char *name, size_t size,
1301 vm_flags_t acctflag, struct user_struct **user,
1302 int creat_flags, int page_size_log)
1304 struct file *file = ERR_PTR(-ENOMEM);
1305 struct inode *inode;
1307 struct super_block *sb;
1308 struct qstr quick_string;
1311 hstate_idx = get_hstate_idx(page_size_log);
1313 return ERR_PTR(-ENODEV);
1316 if (!hugetlbfs_vfsmount[hstate_idx])
1317 return ERR_PTR(-ENOENT);
1319 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1320 *user = current_user();
1321 if (user_shm_lock(size, *user)) {
1323 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1324 current->comm, current->pid);
1325 task_unlock(current);
1328 return ERR_PTR(-EPERM);
1332 sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1333 quick_string.name = name;
1334 quick_string.len = strlen(quick_string.name);
1335 quick_string.hash = 0;
1336 path.dentry = d_alloc_pseudo(sb, &quick_string);
1338 goto out_shm_unlock;
1340 d_set_d_op(path.dentry, &anon_ops);
1341 path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1342 file = ERR_PTR(-ENOSPC);
1343 inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1346 if (creat_flags == HUGETLB_SHMFS_INODE)
1347 inode->i_flags |= S_PRIVATE;
1349 file = ERR_PTR(-ENOMEM);
1350 if (hugetlb_reserve_pages(inode, 0,
1351 size >> huge_page_shift(hstate_inode(inode)), NULL,
1355 d_instantiate(path.dentry, inode);
1356 inode->i_size = size;
1359 file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1360 &hugetlbfs_file_operations);
1362 goto out_dentry; /* inode is already attached */
1372 user_shm_unlock(size, *user);
1378 static int __init init_hugetlbfs_fs(void)
1384 if (!hugepages_supported()) {
1385 pr_info("disabling because there are no supported hugepage sizes\n");
1390 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1391 sizeof(struct hugetlbfs_inode_info),
1392 0, SLAB_ACCOUNT, init_once);
1393 if (hugetlbfs_inode_cachep == NULL)
1396 error = register_filesystem(&hugetlbfs_fs_type);
1401 for_each_hstate(h) {
1403 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1405 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1406 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1409 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1410 pr_err("Cannot mount internal hugetlbfs for "
1411 "page size %uK", ps_kb);
1412 error = PTR_ERR(hugetlbfs_vfsmount[i]);
1413 hugetlbfs_vfsmount[i] = NULL;
1417 /* Non default hstates are optional */
1418 if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1422 kmem_cache_destroy(hugetlbfs_inode_cachep);
1426 fs_initcall(init_hugetlbfs_fs)