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 int sysctl_hugetlb_shm_group;
61 Opt_size, Opt_nr_inodes,
62 Opt_mode, Opt_uid, Opt_gid,
63 Opt_pagesize, Opt_min_size,
67 static const match_table_t tokens = {
68 {Opt_size, "size=%s"},
69 {Opt_nr_inodes, "nr_inodes=%s"},
70 {Opt_mode, "mode=%o"},
73 {Opt_pagesize, "pagesize=%s"},
74 {Opt_min_size, "min_size=%s"},
79 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
80 struct inode *inode, pgoff_t index)
82 vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
86 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
88 mpol_cond_put(vma->vm_policy);
91 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
92 struct inode *inode, pgoff_t index)
96 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
101 static void huge_pagevec_release(struct pagevec *pvec)
105 for (i = 0; i < pagevec_count(pvec); ++i)
106 put_page(pvec->pages[i]);
108 pagevec_reinit(pvec);
112 * Mask used when checking the page offset value passed in via system
113 * calls. This value will be converted to a loff_t which is signed.
114 * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
115 * value. The extra bit (- 1 in the shift value) is to take the sign
118 #define PGOFF_LOFFT_MAX \
119 (((1UL << (PAGE_SHIFT + 1)) - 1) << (BITS_PER_LONG - (PAGE_SHIFT + 1)))
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 * page based offset in vm_pgoff could be sufficiently large to
141 * overflow a loff_t when converted to byte offset. This can
142 * only happen on architectures where sizeof(loff_t) ==
143 * sizeof(unsigned long). So, only check in those instances.
145 if (sizeof(unsigned long) == sizeof(loff_t)) {
146 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
150 /* must be huge page aligned */
151 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
154 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
155 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
156 /* check for overflow */
164 if (hugetlb_reserve_pages(inode,
165 vma->vm_pgoff >> huge_page_order(h),
166 len >> huge_page_shift(h), vma,
171 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
172 i_size_write(inode, len);
180 * Called under down_write(mmap_sem).
183 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
185 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
186 unsigned long len, unsigned long pgoff, unsigned long flags)
188 struct mm_struct *mm = current->mm;
189 struct vm_area_struct *vma;
190 struct hstate *h = hstate_file(file);
191 struct vm_unmapped_area_info info;
193 if (len & ~huge_page_mask(h))
198 if (flags & MAP_FIXED) {
199 if (prepare_hugepage_range(file, addr, len))
205 addr = ALIGN(addr, huge_page_size(h));
206 vma = find_vma(mm, addr);
207 if (TASK_SIZE - len >= addr &&
208 (!vma || addr + len <= vm_start_gap(vma)))
214 info.low_limit = TASK_UNMAPPED_BASE;
215 info.high_limit = TASK_SIZE;
216 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
217 info.align_offset = 0;
218 return vm_unmapped_area(&info);
223 hugetlbfs_read_actor(struct page *page, unsigned long offset,
224 struct iov_iter *to, unsigned long size)
229 /* Find which 4k chunk and offset with in that chunk */
230 i = offset >> PAGE_SHIFT;
231 offset = offset & ~PAGE_MASK;
235 chunksize = PAGE_SIZE;
238 if (chunksize > size)
240 n = copy_page_to_iter(&page[i], offset, chunksize, to);
252 * Support for read() - Find the page attached to f_mapping and copy out the
253 * data. Its *very* similar to do_generic_mapping_read(), we can't use that
254 * since it has PAGE_SIZE assumptions.
256 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
258 struct file *file = iocb->ki_filp;
259 struct hstate *h = hstate_file(file);
260 struct address_space *mapping = file->f_mapping;
261 struct inode *inode = mapping->host;
262 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
263 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
264 unsigned long end_index;
268 while (iov_iter_count(to)) {
272 /* nr is the maximum number of bytes to copy from this page */
273 nr = huge_page_size(h);
274 isize = i_size_read(inode);
277 end_index = (isize - 1) >> huge_page_shift(h);
278 if (index > end_index)
280 if (index == end_index) {
281 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
288 page = find_lock_page(mapping, index);
289 if (unlikely(page == NULL)) {
291 * We have a HOLE, zero out the user-buffer for the
292 * length of the hole or request.
294 copied = iov_iter_zero(nr, to);
299 * We have the page, copy it to user space buffer.
301 copied = hugetlbfs_read_actor(page, offset, to, nr);
306 if (copied != nr && iov_iter_count(to)) {
311 index += offset >> huge_page_shift(h);
312 offset &= ~huge_page_mask(h);
314 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
318 static int hugetlbfs_write_begin(struct file *file,
319 struct address_space *mapping,
320 loff_t pos, unsigned len, unsigned flags,
321 struct page **pagep, void **fsdata)
326 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
327 loff_t pos, unsigned len, unsigned copied,
328 struct page *page, void *fsdata)
334 static void remove_huge_page(struct page *page)
336 ClearPageDirty(page);
337 ClearPageUptodate(page);
338 delete_from_page_cache(page);
342 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
344 struct vm_area_struct *vma;
347 * end == 0 indicates that the entire range after
348 * start should be unmapped.
350 vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
351 unsigned long v_offset;
355 * Can the expression below overflow on 32-bit arches?
356 * No, because the interval tree returns us only those vmas
357 * which overlap the truncated area starting at pgoff,
358 * and no vma on a 32-bit arch can span beyond the 4GB.
360 if (vma->vm_pgoff < start)
361 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
368 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
370 if (v_end > vma->vm_end)
374 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
380 * remove_inode_hugepages handles two distinct cases: truncation and hole
381 * punch. There are subtle differences in operation for each case.
383 * truncation is indicated by end of range being LLONG_MAX
384 * In this case, we first scan the range and release found pages.
385 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
386 * maps and global counts.
387 * hole punch is indicated if end is not LLONG_MAX
388 * In the hole punch case we scan the range and release found pages.
389 * Only when releasing a page is the associated region/reserv map
390 * deleted. The region/reserv map for ranges without associated
391 * pages are not modified.
393 * Callers of this routine must hold the i_mmap_rwsem in write mode to prevent
394 * races with page faults.
396 * Note: If the passed end of range value is beyond the end of file, but
397 * not LLONG_MAX this routine still performs a hole punch operation.
399 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
402 struct hstate *h = hstate_inode(inode);
403 struct address_space *mapping = &inode->i_data;
404 const pgoff_t start = lstart >> huge_page_shift(h);
405 const pgoff_t end = lend >> huge_page_shift(h);
406 struct vm_area_struct pseudo_vma;
410 bool truncate_op = (lend == LLONG_MAX);
412 vma_init(&pseudo_vma, current->mm);
413 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
418 * When no more pages are found, we are done.
420 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
423 for (i = 0; i < pagevec_count(&pvec); ++i) {
424 struct page *page = pvec.pages[i];
428 * A mapped page is impossible as callers should unmap
429 * all references before calling. And, i_mmap_rwsem
430 * prevents the creation of additional mappings.
432 VM_BUG_ON(page_mapped(page));
436 * We must free the huge page and remove from page
437 * cache (remove_huge_page) BEFORE removing the
438 * region/reserve map (hugetlb_unreserve_pages). In
439 * rare out of memory conditions, removal of the
440 * region/reserve map could fail. Correspondingly,
441 * the subpool and global reserve usage count can need
444 VM_BUG_ON(PagePrivate(page));
445 remove_huge_page(page);
448 if (unlikely(hugetlb_unreserve_pages(inode,
449 index, index + 1, 1)))
450 hugetlb_fix_reserve_counts(inode);
455 huge_pagevec_release(&pvec);
460 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
463 static void hugetlbfs_evict_inode(struct inode *inode)
465 struct address_space *mapping = inode->i_mapping;
466 struct resv_map *resv_map;
469 * The vfs layer guarantees that there are no other users of this
470 * inode. Therefore, it would be safe to call remove_inode_hugepages
471 * without holding i_mmap_rwsem. We acquire and hold here to be
472 * consistent with other callers. Since there will be no contention
473 * on the semaphore, overhead is negligible.
475 i_mmap_lock_write(mapping);
476 remove_inode_hugepages(inode, 0, LLONG_MAX);
477 i_mmap_unlock_write(mapping);
479 resv_map = (struct resv_map *)inode->i_mapping->private_data;
480 /* root inode doesn't have the resv_map, so we should check it */
482 resv_map_release(&resv_map->refs);
486 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
489 struct address_space *mapping = inode->i_mapping;
490 struct hstate *h = hstate_inode(inode);
492 BUG_ON(offset & ~huge_page_mask(h));
493 pgoff = offset >> PAGE_SHIFT;
495 i_size_write(inode, offset);
496 i_mmap_lock_write(mapping);
497 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
498 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
499 remove_inode_hugepages(inode, offset, LLONG_MAX);
500 i_mmap_unlock_write(mapping);
504 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
506 struct hstate *h = hstate_inode(inode);
507 loff_t hpage_size = huge_page_size(h);
508 loff_t hole_start, hole_end;
511 * For hole punch round up the beginning offset of the hole and
512 * round down the end.
514 hole_start = round_up(offset, hpage_size);
515 hole_end = round_down(offset + len, hpage_size);
517 if (hole_end > hole_start) {
518 struct address_space *mapping = inode->i_mapping;
519 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
523 /* protected by i_mutex */
524 if (info->seals & F_SEAL_WRITE) {
529 i_mmap_lock_write(mapping);
530 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
531 hugetlb_vmdelete_list(&mapping->i_mmap,
532 hole_start >> PAGE_SHIFT,
533 hole_end >> PAGE_SHIFT);
534 remove_inode_hugepages(inode, hole_start, hole_end);
535 i_mmap_unlock_write(mapping);
542 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
545 struct inode *inode = file_inode(file);
546 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
547 struct address_space *mapping = inode->i_mapping;
548 struct hstate *h = hstate_inode(inode);
549 struct vm_area_struct pseudo_vma;
550 struct mm_struct *mm = current->mm;
551 loff_t hpage_size = huge_page_size(h);
552 unsigned long hpage_shift = huge_page_shift(h);
553 pgoff_t start, index, end;
557 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
560 if (mode & FALLOC_FL_PUNCH_HOLE)
561 return hugetlbfs_punch_hole(inode, offset, len);
564 * Default preallocate case.
565 * For this range, start is rounded down and end is rounded up
566 * as well as being converted to page offsets.
568 start = offset >> hpage_shift;
569 end = (offset + len + hpage_size - 1) >> hpage_shift;
573 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
574 error = inode_newsize_ok(inode, offset + len);
578 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
584 * Initialize a pseudo vma as this is required by the huge page
585 * allocation routines. If NUMA is configured, use page index
586 * as input to create an allocation policy.
588 vma_init(&pseudo_vma, mm);
589 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
590 pseudo_vma.vm_file = file;
592 for (index = start; index < end; index++) {
594 * This is supposed to be the vaddr where the page is being
595 * faulted in, but we have no vaddr here.
599 int avoid_reserve = 0;
604 * fallocate(2) manpage permits EINTR; we may have been
605 * interrupted because we are using up too much memory.
607 if (signal_pending(current)) {
612 /* Set numa allocation policy based on index */
613 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
615 /* addr is the offset within the file (zero based) */
616 addr = index * hpage_size;
619 * fault mutex taken here, protects against fault path
620 * and hole punch. inode_lock previously taken protects
621 * against truncation.
623 hash = hugetlb_fault_mutex_hash(h, mm, &pseudo_vma, mapping,
625 mutex_lock(&hugetlb_fault_mutex_table[hash]);
627 /* See if already present in mapping to avoid alloc/free */
628 page = find_get_page(mapping, index);
631 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
632 hugetlb_drop_vma_policy(&pseudo_vma);
636 /* Allocate page and add to page cache */
637 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
638 hugetlb_drop_vma_policy(&pseudo_vma);
640 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
641 error = PTR_ERR(page);
644 clear_huge_page(page, addr, pages_per_huge_page(h));
645 __SetPageUptodate(page);
646 error = huge_add_to_page_cache(page, mapping, index);
647 if (unlikely(error)) {
649 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
653 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
656 * unlock_page because locked by add_to_page_cache()
657 * page_put due to reference from alloc_huge_page()
663 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
664 i_size_write(inode, offset + len);
665 inode->i_ctime = current_time(inode);
671 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
673 struct inode *inode = d_inode(dentry);
674 struct hstate *h = hstate_inode(inode);
676 unsigned int ia_valid = attr->ia_valid;
677 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
681 error = setattr_prepare(dentry, attr);
685 if (ia_valid & ATTR_SIZE) {
686 loff_t oldsize = inode->i_size;
687 loff_t newsize = attr->ia_size;
689 if (newsize & ~huge_page_mask(h))
691 /* protected by i_mutex */
692 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
693 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
695 error = hugetlb_vmtruncate(inode, newsize);
700 setattr_copy(inode, attr);
701 mark_inode_dirty(inode);
705 static struct inode *hugetlbfs_get_root(struct super_block *sb,
706 struct hugetlbfs_config *config)
710 inode = new_inode(sb);
712 inode->i_ino = get_next_ino();
713 inode->i_mode = S_IFDIR | config->mode;
714 inode->i_uid = config->uid;
715 inode->i_gid = config->gid;
716 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
717 inode->i_op = &hugetlbfs_dir_inode_operations;
718 inode->i_fop = &simple_dir_operations;
719 /* directory inodes start off with i_nlink == 2 (for "." entry) */
721 lockdep_annotate_inode_mutex_key(inode);
727 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
728 * be taken from reclaim -- unlike regular filesystems. This needs an
729 * annotation because huge_pmd_share() does an allocation under hugetlb's
732 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
734 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
736 umode_t mode, dev_t dev)
739 struct resv_map *resv_map;
741 resv_map = resv_map_alloc();
745 inode = new_inode(sb);
747 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
749 inode->i_ino = get_next_ino();
750 inode_init_owner(inode, dir, mode);
751 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
752 &hugetlbfs_i_mmap_rwsem_key);
753 inode->i_mapping->a_ops = &hugetlbfs_aops;
754 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
755 inode->i_mapping->private_data = resv_map;
756 info->seals = F_SEAL_SEAL;
757 switch (mode & S_IFMT) {
759 init_special_inode(inode, mode, dev);
762 inode->i_op = &hugetlbfs_inode_operations;
763 inode->i_fop = &hugetlbfs_file_operations;
766 inode->i_op = &hugetlbfs_dir_inode_operations;
767 inode->i_fop = &simple_dir_operations;
769 /* directory inodes start off with i_nlink == 2 (for "." entry) */
773 inode->i_op = &page_symlink_inode_operations;
774 inode_nohighmem(inode);
777 lockdep_annotate_inode_mutex_key(inode);
779 kref_put(&resv_map->refs, resv_map_release);
785 * File creation. Allocate an inode, and we're done..
787 static int hugetlbfs_mknod(struct inode *dir,
788 struct dentry *dentry, umode_t mode, dev_t dev)
793 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
795 dir->i_ctime = dir->i_mtime = current_time(dir);
796 d_instantiate(dentry, inode);
797 dget(dentry); /* Extra count - pin the dentry in core */
803 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
805 int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
811 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
813 return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
816 static int hugetlbfs_symlink(struct inode *dir,
817 struct dentry *dentry, const char *symname)
822 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
824 int l = strlen(symname)+1;
825 error = page_symlink(inode, symname, l);
827 d_instantiate(dentry, inode);
832 dir->i_ctime = dir->i_mtime = current_time(dir);
838 * mark the head page dirty
840 static int hugetlbfs_set_page_dirty(struct page *page)
842 struct page *head = compound_head(page);
848 static int hugetlbfs_migrate_page(struct address_space *mapping,
849 struct page *newpage, struct page *page,
850 enum migrate_mode mode)
854 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
855 if (rc != MIGRATEPAGE_SUCCESS)
857 if (mode != MIGRATE_SYNC_NO_COPY)
858 migrate_page_copy(newpage, page);
860 migrate_page_states(newpage, page);
862 return MIGRATEPAGE_SUCCESS;
865 static int hugetlbfs_error_remove_page(struct address_space *mapping,
868 struct inode *inode = mapping->host;
869 pgoff_t index = page->index;
871 remove_huge_page(page);
872 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
873 hugetlb_fix_reserve_counts(inode);
879 * Display the mount options in /proc/mounts.
881 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
883 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
884 struct hugepage_subpool *spool = sbinfo->spool;
885 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
886 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
889 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
890 seq_printf(m, ",uid=%u",
891 from_kuid_munged(&init_user_ns, sbinfo->uid));
892 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
893 seq_printf(m, ",gid=%u",
894 from_kgid_munged(&init_user_ns, sbinfo->gid));
895 if (sbinfo->mode != 0755)
896 seq_printf(m, ",mode=%o", sbinfo->mode);
897 if (sbinfo->max_inodes != -1)
898 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
902 if (hpage_size >= 1024) {
906 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
908 if (spool->max_hpages != -1)
909 seq_printf(m, ",size=%llu",
910 (unsigned long long)spool->max_hpages << hpage_shift);
911 if (spool->min_hpages != -1)
912 seq_printf(m, ",min_size=%llu",
913 (unsigned long long)spool->min_hpages << hpage_shift);
918 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
920 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
921 struct hstate *h = hstate_inode(d_inode(dentry));
923 buf->f_type = HUGETLBFS_MAGIC;
924 buf->f_bsize = huge_page_size(h);
926 spin_lock(&sbinfo->stat_lock);
927 /* If no limits set, just report 0 for max/free/used
928 * blocks, like simple_statfs() */
932 spin_lock(&sbinfo->spool->lock);
933 buf->f_blocks = sbinfo->spool->max_hpages;
934 free_pages = sbinfo->spool->max_hpages
935 - sbinfo->spool->used_hpages;
936 buf->f_bavail = buf->f_bfree = free_pages;
937 spin_unlock(&sbinfo->spool->lock);
938 buf->f_files = sbinfo->max_inodes;
939 buf->f_ffree = sbinfo->free_inodes;
941 spin_unlock(&sbinfo->stat_lock);
943 buf->f_namelen = NAME_MAX;
947 static void hugetlbfs_put_super(struct super_block *sb)
949 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
952 sb->s_fs_info = NULL;
955 hugepage_put_subpool(sbi->spool);
961 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
963 if (sbinfo->free_inodes >= 0) {
964 spin_lock(&sbinfo->stat_lock);
965 if (unlikely(!sbinfo->free_inodes)) {
966 spin_unlock(&sbinfo->stat_lock);
969 sbinfo->free_inodes--;
970 spin_unlock(&sbinfo->stat_lock);
976 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
978 if (sbinfo->free_inodes >= 0) {
979 spin_lock(&sbinfo->stat_lock);
980 sbinfo->free_inodes++;
981 spin_unlock(&sbinfo->stat_lock);
986 static struct kmem_cache *hugetlbfs_inode_cachep;
988 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
990 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
991 struct hugetlbfs_inode_info *p;
993 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
995 p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
997 hugetlbfs_inc_free_inodes(sbinfo);
1002 * Any time after allocation, hugetlbfs_destroy_inode can be called
1003 * for the inode. mpol_free_shared_policy is unconditionally called
1004 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1005 * in case of a quick call to destroy.
1007 * Note that the policy is initialized even if we are creating a
1008 * private inode. This simplifies hugetlbfs_destroy_inode.
1010 mpol_shared_policy_init(&p->policy, NULL);
1012 return &p->vfs_inode;
1015 static void hugetlbfs_i_callback(struct rcu_head *head)
1017 struct inode *inode = container_of(head, struct inode, i_rcu);
1018 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1021 static void hugetlbfs_destroy_inode(struct inode *inode)
1023 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1024 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1025 call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
1028 static const struct address_space_operations hugetlbfs_aops = {
1029 .write_begin = hugetlbfs_write_begin,
1030 .write_end = hugetlbfs_write_end,
1031 .set_page_dirty = hugetlbfs_set_page_dirty,
1032 .migratepage = hugetlbfs_migrate_page,
1033 .error_remove_page = hugetlbfs_error_remove_page,
1037 static void init_once(void *foo)
1039 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1041 inode_init_once(&ei->vfs_inode);
1044 const struct file_operations hugetlbfs_file_operations = {
1045 .read_iter = hugetlbfs_read_iter,
1046 .mmap = hugetlbfs_file_mmap,
1047 .fsync = noop_fsync,
1048 .get_unmapped_area = hugetlb_get_unmapped_area,
1049 .llseek = default_llseek,
1050 .fallocate = hugetlbfs_fallocate,
1053 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1054 .create = hugetlbfs_create,
1055 .lookup = simple_lookup,
1056 .link = simple_link,
1057 .unlink = simple_unlink,
1058 .symlink = hugetlbfs_symlink,
1059 .mkdir = hugetlbfs_mkdir,
1060 .rmdir = simple_rmdir,
1061 .mknod = hugetlbfs_mknod,
1062 .rename = simple_rename,
1063 .setattr = hugetlbfs_setattr,
1066 static const struct inode_operations hugetlbfs_inode_operations = {
1067 .setattr = hugetlbfs_setattr,
1070 static const struct super_operations hugetlbfs_ops = {
1071 .alloc_inode = hugetlbfs_alloc_inode,
1072 .destroy_inode = hugetlbfs_destroy_inode,
1073 .evict_inode = hugetlbfs_evict_inode,
1074 .statfs = hugetlbfs_statfs,
1075 .put_super = hugetlbfs_put_super,
1076 .show_options = hugetlbfs_show_options,
1079 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1082 * Convert size option passed from command line to number of huge pages
1083 * in the pool specified by hstate. Size option could be in bytes
1084 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1087 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1088 enum hugetlbfs_size_type val_type)
1090 if (val_type == NO_SIZE)
1093 if (val_type == SIZE_PERCENT) {
1094 size_opt <<= huge_page_shift(h);
1095 size_opt *= h->max_huge_pages;
1096 do_div(size_opt, 100);
1099 size_opt >>= huge_page_shift(h);
1104 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1107 substring_t args[MAX_OPT_ARGS];
1109 unsigned long long max_size_opt = 0, min_size_opt = 0;
1110 enum hugetlbfs_size_type max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1115 while ((p = strsep(&options, ",")) != NULL) {
1120 token = match_token(p, tokens, args);
1123 if (match_int(&args[0], &option))
1125 pconfig->uid = make_kuid(current_user_ns(), option);
1126 if (!uid_valid(pconfig->uid))
1131 if (match_int(&args[0], &option))
1133 pconfig->gid = make_kgid(current_user_ns(), option);
1134 if (!gid_valid(pconfig->gid))
1139 if (match_octal(&args[0], &option))
1141 pconfig->mode = option & 01777U;
1145 /* memparse() will accept a K/M/G without a digit */
1146 if (!isdigit(*args[0].from))
1148 max_size_opt = memparse(args[0].from, &rest);
1149 max_val_type = SIZE_STD;
1151 max_val_type = SIZE_PERCENT;
1156 /* memparse() will accept a K/M/G without a digit */
1157 if (!isdigit(*args[0].from))
1159 pconfig->nr_inodes = memparse(args[0].from, &rest);
1162 case Opt_pagesize: {
1164 ps = memparse(args[0].from, &rest);
1165 pconfig->hstate = size_to_hstate(ps);
1166 if (!pconfig->hstate) {
1167 pr_err("Unsupported page size %lu MB\n",
1174 case Opt_min_size: {
1175 /* memparse() will accept a K/M/G without a digit */
1176 if (!isdigit(*args[0].from))
1178 min_size_opt = memparse(args[0].from, &rest);
1179 min_val_type = SIZE_STD;
1181 min_val_type = SIZE_PERCENT;
1186 pr_err("Bad mount option: \"%s\"\n", p);
1193 * Use huge page pool size (in hstate) to convert the size
1194 * options to number of huge pages. If NO_SIZE, -1 is returned.
1196 pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1197 max_size_opt, max_val_type);
1198 pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1199 min_size_opt, min_val_type);
1202 * If max_size was specified, then min_size must be smaller
1204 if (max_val_type > NO_SIZE &&
1205 pconfig->min_hpages > pconfig->max_hpages) {
1206 pr_err("minimum size can not be greater than maximum size\n");
1213 pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1218 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1221 struct hugetlbfs_config config;
1222 struct hugetlbfs_sb_info *sbinfo;
1224 config.max_hpages = -1; /* No limit on size by default */
1225 config.nr_inodes = -1; /* No limit on number of inodes by default */
1226 config.uid = current_fsuid();
1227 config.gid = current_fsgid();
1229 config.hstate = &default_hstate;
1230 config.min_hpages = -1; /* No default minimum size */
1231 ret = hugetlbfs_parse_options(data, &config);
1235 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1238 sb->s_fs_info = sbinfo;
1239 sbinfo->hstate = config.hstate;
1240 spin_lock_init(&sbinfo->stat_lock);
1241 sbinfo->max_inodes = config.nr_inodes;
1242 sbinfo->free_inodes = config.nr_inodes;
1243 sbinfo->spool = NULL;
1244 sbinfo->uid = config.uid;
1245 sbinfo->gid = config.gid;
1246 sbinfo->mode = config.mode;
1249 * Allocate and initialize subpool if maximum or minimum size is
1250 * specified. Any needed reservations (for minimim size) are taken
1251 * taken when the subpool is created.
1253 if (config.max_hpages != -1 || config.min_hpages != -1) {
1254 sbinfo->spool = hugepage_new_subpool(config.hstate,
1260 sb->s_maxbytes = MAX_LFS_FILESIZE;
1261 sb->s_blocksize = huge_page_size(config.hstate);
1262 sb->s_blocksize_bits = huge_page_shift(config.hstate);
1263 sb->s_magic = HUGETLBFS_MAGIC;
1264 sb->s_op = &hugetlbfs_ops;
1265 sb->s_time_gran = 1;
1266 sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1271 kfree(sbinfo->spool);
1276 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1277 int flags, const char *dev_name, void *data)
1279 return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1282 static struct file_system_type hugetlbfs_fs_type = {
1283 .name = "hugetlbfs",
1284 .mount = hugetlbfs_mount,
1285 .kill_sb = kill_litter_super,
1288 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1290 static int can_do_hugetlb_shm(void)
1293 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1294 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1297 static int get_hstate_idx(int page_size_log)
1299 struct hstate *h = hstate_sizelog(page_size_log);
1307 * Note that size should be aligned to proper hugepage size in caller side,
1308 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1310 struct file *hugetlb_file_setup(const char *name, size_t size,
1311 vm_flags_t acctflag, struct user_struct **user,
1312 int creat_flags, int page_size_log)
1314 struct inode *inode;
1315 struct vfsmount *mnt;
1319 hstate_idx = get_hstate_idx(page_size_log);
1321 return ERR_PTR(-ENODEV);
1324 mnt = hugetlbfs_vfsmount[hstate_idx];
1326 return ERR_PTR(-ENOENT);
1328 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1329 *user = current_user();
1330 if (user_shm_lock(size, *user)) {
1332 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1333 current->comm, current->pid);
1334 task_unlock(current);
1337 return ERR_PTR(-EPERM);
1341 file = ERR_PTR(-ENOSPC);
1342 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1345 if (creat_flags == HUGETLB_SHMFS_INODE)
1346 inode->i_flags |= S_PRIVATE;
1348 inode->i_size = size;
1351 if (hugetlb_reserve_pages(inode, 0,
1352 size >> huge_page_shift(hstate_inode(inode)), NULL,
1354 file = ERR_PTR(-ENOMEM);
1356 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1357 &hugetlbfs_file_operations);
1364 user_shm_unlock(size, *user);
1370 static int __init init_hugetlbfs_fs(void)
1376 if (!hugepages_supported()) {
1377 pr_info("disabling because there are no supported hugepage sizes\n");
1382 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1383 sizeof(struct hugetlbfs_inode_info),
1384 0, SLAB_ACCOUNT, init_once);
1385 if (hugetlbfs_inode_cachep == NULL)
1388 error = register_filesystem(&hugetlbfs_fs_type);
1393 for_each_hstate(h) {
1395 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1397 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1398 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1401 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1402 pr_err("Cannot mount internal hugetlbfs for "
1403 "page size %uK", ps_kb);
1404 error = PTR_ERR(hugetlbfs_vfsmount[i]);
1405 hugetlbfs_vfsmount[i] = NULL;
1409 /* Non default hstates are optional */
1410 if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1414 kmem_cache_destroy(hugetlbfs_inode_cachep);
1418 fs_initcall(init_hugetlbfs_fs)