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/fs_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>
41 #include <linux/sched/mm.h>
43 static const struct super_operations hugetlbfs_ops;
44 static const struct address_space_operations hugetlbfs_aops;
45 const struct file_operations hugetlbfs_file_operations;
46 static const struct inode_operations hugetlbfs_dir_inode_operations;
47 static const struct inode_operations hugetlbfs_inode_operations;
49 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
51 struct hugetlbfs_fs_context {
52 struct hstate *hstate;
53 unsigned long long max_size_opt;
54 unsigned long long min_size_opt;
58 enum hugetlbfs_size_type max_val_type;
59 enum hugetlbfs_size_type min_val_type;
65 int sysctl_hugetlb_shm_group;
77 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
78 fsparam_u32 ("gid", Opt_gid),
79 fsparam_string("min_size", Opt_min_size),
80 fsparam_u32oct("mode", Opt_mode),
81 fsparam_string("nr_inodes", Opt_nr_inodes),
82 fsparam_string("pagesize", Opt_pagesize),
83 fsparam_string("size", Opt_size),
84 fsparam_u32 ("uid", Opt_uid),
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)
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);
124 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
127 struct hstate *h = hstate_file(file);
130 * vma address alignment (but not the pgoff alignment) has
131 * already been checked by prepare_hugepage_range. If you add
132 * any error returns here, do so after setting VM_HUGETLB, so
133 * is_vm_hugetlb_page tests below unmap_region go the right
134 * way when do_mmap unwinds (may be important on powerpc
137 vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
138 vma->vm_ops = &hugetlb_vm_ops;
140 ret = seal_check_future_write(info->seals, vma);
145 * page based offset in vm_pgoff could be sufficiently large to
146 * overflow a loff_t when converted to byte offset. This can
147 * only happen on architectures where sizeof(loff_t) ==
148 * sizeof(unsigned long). So, only check in those instances.
150 if (sizeof(unsigned long) == sizeof(loff_t)) {
151 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
155 /* must be huge page aligned */
156 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
159 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
160 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
161 /* check for overflow */
169 if (!hugetlb_reserve_pages(inode,
170 vma->vm_pgoff >> huge_page_order(h),
171 len >> huge_page_shift(h), vma,
176 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
177 i_size_write(inode, len);
185 * Called under mmap_write_lock(mm).
189 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
190 unsigned long len, unsigned long pgoff, unsigned long flags)
192 struct hstate *h = hstate_file(file);
193 struct vm_unmapped_area_info info;
197 info.low_limit = current->mm->mmap_base;
198 info.high_limit = arch_get_mmap_end(addr, len, flags);
199 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
200 info.align_offset = 0;
201 return vm_unmapped_area(&info);
205 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
206 unsigned long len, unsigned long pgoff, unsigned long flags)
208 struct hstate *h = hstate_file(file);
209 struct vm_unmapped_area_info info;
211 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
213 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
214 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
215 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
216 info.align_offset = 0;
217 addr = vm_unmapped_area(&info);
220 * A failed mmap() very likely causes application failure,
221 * so fall back to the bottom-up function here. This scenario
222 * can happen with large stack limits and large mmap()
225 if (unlikely(offset_in_page(addr))) {
226 VM_BUG_ON(addr != -ENOMEM);
228 info.low_limit = current->mm->mmap_base;
229 info.high_limit = arch_get_mmap_end(addr, len, flags);
230 addr = vm_unmapped_area(&info);
237 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
238 unsigned long len, unsigned long pgoff,
241 struct mm_struct *mm = current->mm;
242 struct vm_area_struct *vma;
243 struct hstate *h = hstate_file(file);
244 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
246 if (len & ~huge_page_mask(h))
251 if (flags & MAP_FIXED) {
252 if (prepare_hugepage_range(file, addr, len))
258 addr = ALIGN(addr, huge_page_size(h));
259 vma = find_vma(mm, addr);
260 if (mmap_end - len >= addr &&
261 (!vma || addr + len <= vm_start_gap(vma)))
266 * Use mm->get_unmapped_area value as a hint to use topdown routine.
267 * If architectures have special needs, they should define their own
268 * version of hugetlb_get_unmapped_area.
270 if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
271 return hugetlb_get_unmapped_area_topdown(file, addr, len,
273 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
277 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
279 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
280 unsigned long len, unsigned long pgoff,
283 return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
288 hugetlbfs_read_actor(struct page *page, unsigned long offset,
289 struct iov_iter *to, unsigned long size)
294 /* Find which 4k chunk and offset with in that chunk */
295 i = offset >> PAGE_SHIFT;
296 offset = offset & ~PAGE_MASK;
300 chunksize = PAGE_SIZE;
303 if (chunksize > size)
305 n = copy_page_to_iter(&page[i], offset, chunksize, to);
317 * Support for read() - Find the page attached to f_mapping and copy out the
318 * data. Its *very* similar to generic_file_buffered_read(), we can't use that
319 * since it has PAGE_SIZE assumptions.
321 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
323 struct file *file = iocb->ki_filp;
324 struct hstate *h = hstate_file(file);
325 struct address_space *mapping = file->f_mapping;
326 struct inode *inode = mapping->host;
327 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
328 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
329 unsigned long end_index;
333 while (iov_iter_count(to)) {
337 /* nr is the maximum number of bytes to copy from this page */
338 nr = huge_page_size(h);
339 isize = i_size_read(inode);
342 end_index = (isize - 1) >> huge_page_shift(h);
343 if (index > end_index)
345 if (index == end_index) {
346 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
353 page = find_lock_page(mapping, index);
354 if (unlikely(page == NULL)) {
356 * We have a HOLE, zero out the user-buffer for the
357 * length of the hole or request.
359 copied = iov_iter_zero(nr, to);
364 * We have the page, copy it to user space buffer.
366 copied = hugetlbfs_read_actor(page, offset, to, nr);
371 if (copied != nr && iov_iter_count(to)) {
376 index += offset >> huge_page_shift(h);
377 offset &= ~huge_page_mask(h);
379 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
383 static int hugetlbfs_write_begin(struct file *file,
384 struct address_space *mapping,
385 loff_t pos, unsigned len,
386 struct page **pagep, void **fsdata)
391 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
392 loff_t pos, unsigned len, unsigned copied,
393 struct page *page, void *fsdata)
399 static void remove_huge_page(struct page *page)
401 ClearPageDirty(page);
402 ClearPageUptodate(page);
403 delete_from_page_cache(page);
407 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
408 zap_flags_t zap_flags)
410 struct vm_area_struct *vma;
413 * end == 0 indicates that the entire range after start should be
414 * unmapped. Note, end is exclusive, whereas the interval tree takes
415 * an inclusive "last".
417 vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
418 unsigned long v_offset;
422 * Can the expression below overflow on 32-bit arches?
423 * No, because the interval tree returns us only those vmas
424 * which overlap the truncated area starting at pgoff,
425 * and no vma on a 32-bit arch can span beyond the 4GB.
427 if (vma->vm_pgoff < start)
428 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
435 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
437 if (v_end > vma->vm_end)
441 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
447 * remove_inode_hugepages handles two distinct cases: truncation and hole
448 * punch. There are subtle differences in operation for each case.
450 * truncation is indicated by end of range being LLONG_MAX
451 * In this case, we first scan the range and release found pages.
452 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
453 * maps and global counts. Page faults can not race with truncation
454 * in this routine. hugetlb_no_page() holds i_mmap_rwsem and prevents
455 * page faults in the truncated range by checking i_size. i_size is
456 * modified while holding i_mmap_rwsem.
457 * hole punch is indicated if end is not LLONG_MAX
458 * In the hole punch case we scan the range and release found pages.
459 * Only when releasing a page is the associated region/reserve map
460 * deleted. The region/reserve map for ranges without associated
461 * pages are not modified. Page faults can race with hole punch.
462 * This is indicated if we find a mapped page.
463 * Note: If the passed end of range value is beyond the end of file, but
464 * not LLONG_MAX this routine still performs a hole punch operation.
466 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
469 struct hstate *h = hstate_inode(inode);
470 struct address_space *mapping = &inode->i_data;
471 const pgoff_t start = lstart >> huge_page_shift(h);
472 const pgoff_t end = lend >> huge_page_shift(h);
473 struct folio_batch fbatch;
476 bool truncate_op = (lend == LLONG_MAX);
478 folio_batch_init(&fbatch);
480 while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
481 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
482 struct folio *folio = fbatch.folios[i];
485 index = folio->index;
488 * Only need to hold the fault mutex in the
489 * hole punch case. This prevents races with
490 * page faults. Races are not possible in the
491 * case of truncation.
493 hash = hugetlb_fault_mutex_hash(mapping, index);
494 mutex_lock(&hugetlb_fault_mutex_table[hash]);
498 * If folio is mapped, it was faulted in after being
499 * unmapped in caller. Unmap (again) now after taking
500 * the fault mutex. The mutex will prevent faults
501 * until we finish removing the folio.
503 * This race can only happen in the hole punch case.
504 * Getting here in a truncate operation is a bug.
506 if (unlikely(folio_mapped(folio))) {
509 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
510 i_mmap_lock_write(mapping);
511 mutex_lock(&hugetlb_fault_mutex_table[hash]);
512 hugetlb_vmdelete_list(&mapping->i_mmap,
513 index * pages_per_huge_page(h),
514 (index + 1) * pages_per_huge_page(h),
515 ZAP_FLAG_DROP_MARKER);
516 i_mmap_unlock_write(mapping);
521 * We must free the huge page and remove from page
522 * cache (remove_huge_page) BEFORE removing the
523 * region/reserve map (hugetlb_unreserve_pages). In
524 * rare out of memory conditions, removal of the
525 * region/reserve map could fail. Correspondingly,
526 * the subpool and global reserve usage count can need
529 VM_BUG_ON(HPageRestoreReserve(&folio->page));
530 remove_huge_page(&folio->page);
533 if (unlikely(hugetlb_unreserve_pages(inode,
534 index, index + 1, 1)))
535 hugetlb_fix_reserve_counts(inode);
540 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
542 folio_batch_release(&fbatch);
547 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
550 static void hugetlbfs_evict_inode(struct inode *inode)
552 struct resv_map *resv_map;
554 remove_inode_hugepages(inode, 0, LLONG_MAX);
557 * Get the resv_map from the address space embedded in the inode.
558 * This is the address space which points to any resv_map allocated
559 * at inode creation time. If this is a device special inode,
560 * i_mapping may not point to the original address space.
562 resv_map = (struct resv_map *)(&inode->i_data)->private_data;
563 /* Only regular and link inodes have associated reserve maps */
565 resv_map_release(&resv_map->refs);
569 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
572 struct address_space *mapping = inode->i_mapping;
573 struct hstate *h = hstate_inode(inode);
575 BUG_ON(offset & ~huge_page_mask(h));
576 pgoff = offset >> PAGE_SHIFT;
578 i_mmap_lock_write(mapping);
579 i_size_write(inode, offset);
580 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
581 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
582 ZAP_FLAG_DROP_MARKER);
583 i_mmap_unlock_write(mapping);
584 remove_inode_hugepages(inode, offset, LLONG_MAX);
587 static void hugetlbfs_zero_partial_page(struct hstate *h,
588 struct address_space *mapping,
592 pgoff_t idx = start >> huge_page_shift(h);
595 folio = filemap_lock_folio(mapping, idx);
599 start = start & ~huge_page_mask(h);
600 end = end & ~huge_page_mask(h);
602 end = huge_page_size(h);
604 folio_zero_segment(folio, (size_t)start, (size_t)end);
610 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
612 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
613 struct address_space *mapping = inode->i_mapping;
614 struct hstate *h = hstate_inode(inode);
615 loff_t hpage_size = huge_page_size(h);
616 loff_t hole_start, hole_end;
619 * hole_start and hole_end indicate the full pages within the hole.
621 hole_start = round_up(offset, hpage_size);
622 hole_end = round_down(offset + len, hpage_size);
626 /* protected by i_rwsem */
627 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
632 i_mmap_lock_write(mapping);
634 /* If range starts before first full page, zero partial page. */
635 if (offset < hole_start)
636 hugetlbfs_zero_partial_page(h, mapping,
637 offset, min(offset + len, hole_start));
639 /* Unmap users of full pages in the hole. */
640 if (hole_end > hole_start) {
641 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
642 hugetlb_vmdelete_list(&mapping->i_mmap,
643 hole_start >> PAGE_SHIFT,
644 hole_end >> PAGE_SHIFT, 0);
647 /* If range extends beyond last full page, zero partial page. */
648 if ((offset + len) > hole_end && (offset + len) > hole_start)
649 hugetlbfs_zero_partial_page(h, mapping,
650 hole_end, offset + len);
652 i_mmap_unlock_write(mapping);
654 /* Remove full pages from the file. */
655 if (hole_end > hole_start)
656 remove_inode_hugepages(inode, hole_start, hole_end);
663 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
666 struct inode *inode = file_inode(file);
667 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
668 struct address_space *mapping = inode->i_mapping;
669 struct hstate *h = hstate_inode(inode);
670 struct vm_area_struct pseudo_vma;
671 struct mm_struct *mm = current->mm;
672 loff_t hpage_size = huge_page_size(h);
673 unsigned long hpage_shift = huge_page_shift(h);
674 pgoff_t start, index, end;
678 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
681 if (mode & FALLOC_FL_PUNCH_HOLE)
682 return hugetlbfs_punch_hole(inode, offset, len);
685 * Default preallocate case.
686 * For this range, start is rounded down and end is rounded up
687 * as well as being converted to page offsets.
689 start = offset >> hpage_shift;
690 end = (offset + len + hpage_size - 1) >> hpage_shift;
694 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
695 error = inode_newsize_ok(inode, offset + len);
699 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
705 * Initialize a pseudo vma as this is required by the huge page
706 * allocation routines. If NUMA is configured, use page index
707 * as input to create an allocation policy.
709 vma_init(&pseudo_vma, mm);
710 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
711 pseudo_vma.vm_file = file;
713 for (index = start; index < end; index++) {
715 * This is supposed to be the vaddr where the page is being
716 * faulted in, but we have no vaddr here.
724 * fallocate(2) manpage permits EINTR; we may have been
725 * interrupted because we are using up too much memory.
727 if (signal_pending(current)) {
732 /* Set numa allocation policy based on index */
733 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
735 /* addr is the offset within the file (zero based) */
736 addr = index * hpage_size;
739 * fault mutex taken here, protects against fault path
740 * and hole punch. inode_lock previously taken protects
741 * against truncation.
743 hash = hugetlb_fault_mutex_hash(mapping, index);
744 mutex_lock(&hugetlb_fault_mutex_table[hash]);
746 /* See if already present in mapping to avoid alloc/free */
747 page = find_get_page(mapping, index);
750 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
751 hugetlb_drop_vma_policy(&pseudo_vma);
756 * Allocate page without setting the avoid_reserve argument.
757 * There certainly are no reserves associated with the
758 * pseudo_vma. However, there could be shared mappings with
759 * reserves for the file at the inode level. If we fallocate
760 * pages in these areas, we need to consume the reserves
761 * to keep reservation accounting consistent.
763 page = alloc_huge_page(&pseudo_vma, addr, 0);
764 hugetlb_drop_vma_policy(&pseudo_vma);
766 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
767 error = PTR_ERR(page);
770 clear_huge_page(page, addr, pages_per_huge_page(h));
771 __SetPageUptodate(page);
772 error = huge_add_to_page_cache(page, mapping, index);
773 if (unlikely(error)) {
774 restore_reserve_on_error(h, &pseudo_vma, addr, page);
776 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
780 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
782 SetHPageMigratable(page);
784 * unlock_page because locked by huge_add_to_page_cache()
785 * put_page() due to reference from alloc_huge_page()
791 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
792 i_size_write(inode, offset + len);
793 inode->i_ctime = current_time(inode);
799 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
800 struct dentry *dentry, struct iattr *attr)
802 struct inode *inode = d_inode(dentry);
803 struct hstate *h = hstate_inode(inode);
805 unsigned int ia_valid = attr->ia_valid;
806 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
808 error = setattr_prepare(&init_user_ns, dentry, attr);
812 if (ia_valid & ATTR_SIZE) {
813 loff_t oldsize = inode->i_size;
814 loff_t newsize = attr->ia_size;
816 if (newsize & ~huge_page_mask(h))
818 /* protected by i_rwsem */
819 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
820 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
822 hugetlb_vmtruncate(inode, newsize);
825 setattr_copy(&init_user_ns, inode, attr);
826 mark_inode_dirty(inode);
830 static struct inode *hugetlbfs_get_root(struct super_block *sb,
831 struct hugetlbfs_fs_context *ctx)
835 inode = new_inode(sb);
837 inode->i_ino = get_next_ino();
838 inode->i_mode = S_IFDIR | ctx->mode;
839 inode->i_uid = ctx->uid;
840 inode->i_gid = ctx->gid;
841 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
842 inode->i_op = &hugetlbfs_dir_inode_operations;
843 inode->i_fop = &simple_dir_operations;
844 /* directory inodes start off with i_nlink == 2 (for "." entry) */
846 lockdep_annotate_inode_mutex_key(inode);
852 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
853 * be taken from reclaim -- unlike regular filesystems. This needs an
854 * annotation because huge_pmd_share() does an allocation under hugetlb's
857 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
859 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
861 umode_t mode, dev_t dev)
864 struct resv_map *resv_map = NULL;
867 * Reserve maps are only needed for inodes that can have associated
870 if (S_ISREG(mode) || S_ISLNK(mode)) {
871 resv_map = resv_map_alloc();
876 inode = new_inode(sb);
878 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
880 inode->i_ino = get_next_ino();
881 inode_init_owner(&init_user_ns, inode, dir, mode);
882 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
883 &hugetlbfs_i_mmap_rwsem_key);
884 inode->i_mapping->a_ops = &hugetlbfs_aops;
885 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
886 inode->i_mapping->private_data = resv_map;
887 info->seals = F_SEAL_SEAL;
888 switch (mode & S_IFMT) {
890 init_special_inode(inode, mode, dev);
893 inode->i_op = &hugetlbfs_inode_operations;
894 inode->i_fop = &hugetlbfs_file_operations;
897 inode->i_op = &hugetlbfs_dir_inode_operations;
898 inode->i_fop = &simple_dir_operations;
900 /* directory inodes start off with i_nlink == 2 (for "." entry) */
904 inode->i_op = &page_symlink_inode_operations;
905 inode_nohighmem(inode);
908 lockdep_annotate_inode_mutex_key(inode);
911 kref_put(&resv_map->refs, resv_map_release);
918 * File creation. Allocate an inode, and we're done..
920 static int do_hugetlbfs_mknod(struct inode *dir,
921 struct dentry *dentry,
929 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
931 dir->i_ctime = dir->i_mtime = current_time(dir);
933 d_tmpfile(dentry, inode);
935 d_instantiate(dentry, inode);
936 dget(dentry);/* Extra count - pin the dentry in core */
943 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
944 struct dentry *dentry, umode_t mode, dev_t dev)
946 return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
949 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
950 struct dentry *dentry, umode_t mode)
952 int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
959 static int hugetlbfs_create(struct user_namespace *mnt_userns,
960 struct inode *dir, struct dentry *dentry,
961 umode_t mode, bool excl)
963 return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
966 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
967 struct inode *dir, struct dentry *dentry,
970 return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
973 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
974 struct inode *dir, struct dentry *dentry,
980 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
982 int l = strlen(symname)+1;
983 error = page_symlink(inode, symname, l);
985 d_instantiate(dentry, inode);
990 dir->i_ctime = dir->i_mtime = current_time(dir);
995 #ifdef CONFIG_MIGRATION
996 static int hugetlbfs_migrate_folio(struct address_space *mapping,
997 struct folio *dst, struct folio *src,
998 enum migrate_mode mode)
1002 rc = migrate_huge_page_move_mapping(mapping, dst, src);
1003 if (rc != MIGRATEPAGE_SUCCESS)
1006 if (hugetlb_page_subpool(&src->page)) {
1007 hugetlb_set_page_subpool(&dst->page,
1008 hugetlb_page_subpool(&src->page));
1009 hugetlb_set_page_subpool(&src->page, NULL);
1012 if (mode != MIGRATE_SYNC_NO_COPY)
1013 folio_migrate_copy(dst, src);
1015 folio_migrate_flags(dst, src);
1017 return MIGRATEPAGE_SUCCESS;
1020 #define hugetlbfs_migrate_folio NULL
1023 static int hugetlbfs_error_remove_page(struct address_space *mapping,
1026 struct inode *inode = mapping->host;
1027 pgoff_t index = page->index;
1029 remove_huge_page(page);
1030 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
1031 hugetlb_fix_reserve_counts(inode);
1037 * Display the mount options in /proc/mounts.
1039 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1041 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1042 struct hugepage_subpool *spool = sbinfo->spool;
1043 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1044 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1047 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1048 seq_printf(m, ",uid=%u",
1049 from_kuid_munged(&init_user_ns, sbinfo->uid));
1050 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1051 seq_printf(m, ",gid=%u",
1052 from_kgid_munged(&init_user_ns, sbinfo->gid));
1053 if (sbinfo->mode != 0755)
1054 seq_printf(m, ",mode=%o", sbinfo->mode);
1055 if (sbinfo->max_inodes != -1)
1056 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1060 if (hpage_size >= 1024) {
1064 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1066 if (spool->max_hpages != -1)
1067 seq_printf(m, ",size=%llu",
1068 (unsigned long long)spool->max_hpages << hpage_shift);
1069 if (spool->min_hpages != -1)
1070 seq_printf(m, ",min_size=%llu",
1071 (unsigned long long)spool->min_hpages << hpage_shift);
1076 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1078 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1079 struct hstate *h = hstate_inode(d_inode(dentry));
1081 buf->f_type = HUGETLBFS_MAGIC;
1082 buf->f_bsize = huge_page_size(h);
1084 spin_lock(&sbinfo->stat_lock);
1085 /* If no limits set, just report 0 for max/free/used
1086 * blocks, like simple_statfs() */
1087 if (sbinfo->spool) {
1090 spin_lock_irq(&sbinfo->spool->lock);
1091 buf->f_blocks = sbinfo->spool->max_hpages;
1092 free_pages = sbinfo->spool->max_hpages
1093 - sbinfo->spool->used_hpages;
1094 buf->f_bavail = buf->f_bfree = free_pages;
1095 spin_unlock_irq(&sbinfo->spool->lock);
1096 buf->f_files = sbinfo->max_inodes;
1097 buf->f_ffree = sbinfo->free_inodes;
1099 spin_unlock(&sbinfo->stat_lock);
1101 buf->f_namelen = NAME_MAX;
1105 static void hugetlbfs_put_super(struct super_block *sb)
1107 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1110 sb->s_fs_info = NULL;
1113 hugepage_put_subpool(sbi->spool);
1119 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1121 if (sbinfo->free_inodes >= 0) {
1122 spin_lock(&sbinfo->stat_lock);
1123 if (unlikely(!sbinfo->free_inodes)) {
1124 spin_unlock(&sbinfo->stat_lock);
1127 sbinfo->free_inodes--;
1128 spin_unlock(&sbinfo->stat_lock);
1134 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1136 if (sbinfo->free_inodes >= 0) {
1137 spin_lock(&sbinfo->stat_lock);
1138 sbinfo->free_inodes++;
1139 spin_unlock(&sbinfo->stat_lock);
1144 static struct kmem_cache *hugetlbfs_inode_cachep;
1146 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1148 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1149 struct hugetlbfs_inode_info *p;
1151 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1153 p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1155 hugetlbfs_inc_free_inodes(sbinfo);
1160 * Any time after allocation, hugetlbfs_destroy_inode can be called
1161 * for the inode. mpol_free_shared_policy is unconditionally called
1162 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1163 * in case of a quick call to destroy.
1165 * Note that the policy is initialized even if we are creating a
1166 * private inode. This simplifies hugetlbfs_destroy_inode.
1168 mpol_shared_policy_init(&p->policy, NULL);
1170 return &p->vfs_inode;
1173 static void hugetlbfs_free_inode(struct inode *inode)
1175 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1178 static void hugetlbfs_destroy_inode(struct inode *inode)
1180 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1181 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1184 static const struct address_space_operations hugetlbfs_aops = {
1185 .write_begin = hugetlbfs_write_begin,
1186 .write_end = hugetlbfs_write_end,
1187 .dirty_folio = noop_dirty_folio,
1188 .migrate_folio = hugetlbfs_migrate_folio,
1189 .error_remove_page = hugetlbfs_error_remove_page,
1193 static void init_once(void *foo)
1195 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1197 inode_init_once(&ei->vfs_inode);
1200 const struct file_operations hugetlbfs_file_operations = {
1201 .read_iter = hugetlbfs_read_iter,
1202 .mmap = hugetlbfs_file_mmap,
1203 .fsync = noop_fsync,
1204 .get_unmapped_area = hugetlb_get_unmapped_area,
1205 .llseek = default_llseek,
1206 .fallocate = hugetlbfs_fallocate,
1209 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1210 .create = hugetlbfs_create,
1211 .lookup = simple_lookup,
1212 .link = simple_link,
1213 .unlink = simple_unlink,
1214 .symlink = hugetlbfs_symlink,
1215 .mkdir = hugetlbfs_mkdir,
1216 .rmdir = simple_rmdir,
1217 .mknod = hugetlbfs_mknod,
1218 .rename = simple_rename,
1219 .setattr = hugetlbfs_setattr,
1220 .tmpfile = hugetlbfs_tmpfile,
1223 static const struct inode_operations hugetlbfs_inode_operations = {
1224 .setattr = hugetlbfs_setattr,
1227 static const struct super_operations hugetlbfs_ops = {
1228 .alloc_inode = hugetlbfs_alloc_inode,
1229 .free_inode = hugetlbfs_free_inode,
1230 .destroy_inode = hugetlbfs_destroy_inode,
1231 .evict_inode = hugetlbfs_evict_inode,
1232 .statfs = hugetlbfs_statfs,
1233 .put_super = hugetlbfs_put_super,
1234 .show_options = hugetlbfs_show_options,
1238 * Convert size option passed from command line to number of huge pages
1239 * in the pool specified by hstate. Size option could be in bytes
1240 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1243 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1244 enum hugetlbfs_size_type val_type)
1246 if (val_type == NO_SIZE)
1249 if (val_type == SIZE_PERCENT) {
1250 size_opt <<= huge_page_shift(h);
1251 size_opt *= h->max_huge_pages;
1252 do_div(size_opt, 100);
1255 size_opt >>= huge_page_shift(h);
1260 * Parse one mount parameter.
1262 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1264 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1265 struct fs_parse_result result;
1270 opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1276 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1277 if (!uid_valid(ctx->uid))
1282 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1283 if (!gid_valid(ctx->gid))
1288 ctx->mode = result.uint_32 & 01777U;
1292 /* memparse() will accept a K/M/G without a digit */
1293 if (!isdigit(param->string[0]))
1295 ctx->max_size_opt = memparse(param->string, &rest);
1296 ctx->max_val_type = SIZE_STD;
1298 ctx->max_val_type = SIZE_PERCENT;
1302 /* memparse() will accept a K/M/G without a digit */
1303 if (!isdigit(param->string[0]))
1305 ctx->nr_inodes = memparse(param->string, &rest);
1309 ps = memparse(param->string, &rest);
1310 ctx->hstate = size_to_hstate(ps);
1312 pr_err("Unsupported page size %lu MB\n", ps >> 20);
1318 /* memparse() will accept a K/M/G without a digit */
1319 if (!isdigit(param->string[0]))
1321 ctx->min_size_opt = memparse(param->string, &rest);
1322 ctx->min_val_type = SIZE_STD;
1324 ctx->min_val_type = SIZE_PERCENT;
1332 return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1333 param->string, param->key);
1337 * Validate the parsed options.
1339 static int hugetlbfs_validate(struct fs_context *fc)
1341 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1344 * Use huge page pool size (in hstate) to convert the size
1345 * options to number of huge pages. If NO_SIZE, -1 is returned.
1347 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1350 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1355 * If max_size was specified, then min_size must be smaller
1357 if (ctx->max_val_type > NO_SIZE &&
1358 ctx->min_hpages > ctx->max_hpages) {
1359 pr_err("Minimum size can not be greater than maximum size\n");
1367 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1369 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1370 struct hugetlbfs_sb_info *sbinfo;
1372 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1375 sb->s_fs_info = sbinfo;
1376 spin_lock_init(&sbinfo->stat_lock);
1377 sbinfo->hstate = ctx->hstate;
1378 sbinfo->max_inodes = ctx->nr_inodes;
1379 sbinfo->free_inodes = ctx->nr_inodes;
1380 sbinfo->spool = NULL;
1381 sbinfo->uid = ctx->uid;
1382 sbinfo->gid = ctx->gid;
1383 sbinfo->mode = ctx->mode;
1386 * Allocate and initialize subpool if maximum or minimum size is
1387 * specified. Any needed reservations (for minimum size) are taken
1388 * taken when the subpool is created.
1390 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1391 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1397 sb->s_maxbytes = MAX_LFS_FILESIZE;
1398 sb->s_blocksize = huge_page_size(ctx->hstate);
1399 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1400 sb->s_magic = HUGETLBFS_MAGIC;
1401 sb->s_op = &hugetlbfs_ops;
1402 sb->s_time_gran = 1;
1405 * Due to the special and limited functionality of hugetlbfs, it does
1406 * not work well as a stacking filesystem.
1408 sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1409 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1414 kfree(sbinfo->spool);
1419 static int hugetlbfs_get_tree(struct fs_context *fc)
1421 int err = hugetlbfs_validate(fc);
1424 return get_tree_nodev(fc, hugetlbfs_fill_super);
1427 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1429 kfree(fc->fs_private);
1432 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1433 .free = hugetlbfs_fs_context_free,
1434 .parse_param = hugetlbfs_parse_param,
1435 .get_tree = hugetlbfs_get_tree,
1438 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1440 struct hugetlbfs_fs_context *ctx;
1442 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1446 ctx->max_hpages = -1; /* No limit on size by default */
1447 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1448 ctx->uid = current_fsuid();
1449 ctx->gid = current_fsgid();
1451 ctx->hstate = &default_hstate;
1452 ctx->min_hpages = -1; /* No default minimum size */
1453 ctx->max_val_type = NO_SIZE;
1454 ctx->min_val_type = NO_SIZE;
1455 fc->fs_private = ctx;
1456 fc->ops = &hugetlbfs_fs_context_ops;
1460 static struct file_system_type hugetlbfs_fs_type = {
1461 .name = "hugetlbfs",
1462 .init_fs_context = hugetlbfs_init_fs_context,
1463 .parameters = hugetlb_fs_parameters,
1464 .kill_sb = kill_litter_super,
1467 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1469 static int can_do_hugetlb_shm(void)
1472 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1473 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1476 static int get_hstate_idx(int page_size_log)
1478 struct hstate *h = hstate_sizelog(page_size_log);
1482 return hstate_index(h);
1486 * Note that size should be aligned to proper hugepage size in caller side,
1487 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1489 struct file *hugetlb_file_setup(const char *name, size_t size,
1490 vm_flags_t acctflag, int creat_flags,
1493 struct inode *inode;
1494 struct vfsmount *mnt;
1498 hstate_idx = get_hstate_idx(page_size_log);
1500 return ERR_PTR(-ENODEV);
1502 mnt = hugetlbfs_vfsmount[hstate_idx];
1504 return ERR_PTR(-ENOENT);
1506 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1507 struct ucounts *ucounts = current_ucounts();
1509 if (user_shm_lock(size, ucounts)) {
1510 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1511 current->comm, current->pid);
1512 user_shm_unlock(size, ucounts);
1514 return ERR_PTR(-EPERM);
1517 file = ERR_PTR(-ENOSPC);
1518 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1521 if (creat_flags == HUGETLB_SHMFS_INODE)
1522 inode->i_flags |= S_PRIVATE;
1524 inode->i_size = size;
1527 if (!hugetlb_reserve_pages(inode, 0,
1528 size >> huge_page_shift(hstate_inode(inode)), NULL,
1530 file = ERR_PTR(-ENOMEM);
1532 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1533 &hugetlbfs_file_operations);
1542 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1544 struct fs_context *fc;
1545 struct vfsmount *mnt;
1547 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1551 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1557 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1558 huge_page_size(h) >> 10);
1562 static int __init init_hugetlbfs_fs(void)
1564 struct vfsmount *mnt;
1569 if (!hugepages_supported()) {
1570 pr_info("disabling because there are no supported hugepage sizes\n");
1575 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1576 sizeof(struct hugetlbfs_inode_info),
1577 0, SLAB_ACCOUNT, init_once);
1578 if (hugetlbfs_inode_cachep == NULL)
1581 error = register_filesystem(&hugetlbfs_fs_type);
1585 /* default hstate mount is required */
1586 mnt = mount_one_hugetlbfs(&default_hstate);
1588 error = PTR_ERR(mnt);
1591 hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1593 /* other hstates are optional */
1595 for_each_hstate(h) {
1596 if (i == default_hstate_idx) {
1601 mnt = mount_one_hugetlbfs(h);
1603 hugetlbfs_vfsmount[i] = NULL;
1605 hugetlbfs_vfsmount[i] = mnt;
1612 (void)unregister_filesystem(&hugetlbfs_fs_type);
1614 kmem_cache_destroy(hugetlbfs_inode_cachep);
1618 fs_initcall(init_hugetlbfs_fs)