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 * Support for read() - Find the page attached to f_mapping and copy out the
289 * data. Its *very* similar to generic_file_buffered_read(), we can't use that
290 * since it has PAGE_SIZE assumptions.
292 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
294 struct file *file = iocb->ki_filp;
295 struct hstate *h = hstate_file(file);
296 struct address_space *mapping = file->f_mapping;
297 struct inode *inode = mapping->host;
298 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
299 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
300 unsigned long end_index;
304 while (iov_iter_count(to)) {
308 /* nr is the maximum number of bytes to copy from this page */
309 nr = huge_page_size(h);
310 isize = i_size_read(inode);
313 end_index = (isize - 1) >> huge_page_shift(h);
314 if (index > end_index)
316 if (index == end_index) {
317 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
324 page = find_lock_page(mapping, index);
325 if (unlikely(page == NULL)) {
327 * We have a HOLE, zero out the user-buffer for the
328 * length of the hole or request.
330 copied = iov_iter_zero(nr, to);
335 * We have the page, copy it to user space buffer.
337 copied = copy_page_to_iter(page, offset, nr, to);
342 if (copied != nr && iov_iter_count(to)) {
347 index += offset >> huge_page_shift(h);
348 offset &= ~huge_page_mask(h);
350 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
354 static int hugetlbfs_write_begin(struct file *file,
355 struct address_space *mapping,
356 loff_t pos, unsigned len,
357 struct page **pagep, void **fsdata)
362 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
363 loff_t pos, unsigned len, unsigned copied,
364 struct page *page, void *fsdata)
370 static void remove_huge_page(struct page *page)
372 ClearPageDirty(page);
373 ClearPageUptodate(page);
374 delete_from_page_cache(page);
378 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
379 zap_flags_t zap_flags)
381 struct vm_area_struct *vma;
384 * end == 0 indicates that the entire range after start should be
385 * unmapped. Note, end is exclusive, whereas the interval tree takes
386 * an inclusive "last".
388 vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
389 unsigned long v_offset;
393 * Can the expression below overflow on 32-bit arches?
394 * No, because the interval tree returns us only those vmas
395 * which overlap the truncated area starting at pgoff,
396 * and no vma on a 32-bit arch can span beyond the 4GB.
398 if (vma->vm_pgoff < start)
399 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
406 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
408 if (v_end > vma->vm_end)
412 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
418 * remove_inode_hugepages handles two distinct cases: truncation and hole
419 * punch. There are subtle differences in operation for each case.
421 * truncation is indicated by end of range being LLONG_MAX
422 * In this case, we first scan the range and release found pages.
423 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
424 * maps and global counts. Page faults can not race with truncation
425 * in this routine. hugetlb_no_page() holds i_mmap_rwsem and prevents
426 * page faults in the truncated range by checking i_size. i_size is
427 * modified while holding i_mmap_rwsem.
428 * hole punch is indicated if end is not LLONG_MAX
429 * In the hole punch case we scan the range and release found pages.
430 * Only when releasing a page is the associated region/reserve map
431 * deleted. The region/reserve map for ranges without associated
432 * pages are not modified. Page faults can race with hole punch.
433 * This is indicated if we find a mapped page.
434 * Note: If the passed end of range value is beyond the end of file, but
435 * not LLONG_MAX this routine still performs a hole punch operation.
437 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
440 struct hstate *h = hstate_inode(inode);
441 struct address_space *mapping = &inode->i_data;
442 const pgoff_t start = lstart >> huge_page_shift(h);
443 const pgoff_t end = lend >> huge_page_shift(h);
444 struct folio_batch fbatch;
447 bool truncate_op = (lend == LLONG_MAX);
449 folio_batch_init(&fbatch);
451 while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
452 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
453 struct folio *folio = fbatch.folios[i];
456 index = folio->index;
459 * Only need to hold the fault mutex in the
460 * hole punch case. This prevents races with
461 * page faults. Races are not possible in the
462 * case of truncation.
464 hash = hugetlb_fault_mutex_hash(mapping, index);
465 mutex_lock(&hugetlb_fault_mutex_table[hash]);
469 * If folio is mapped, it was faulted in after being
470 * unmapped in caller. Unmap (again) now after taking
471 * the fault mutex. The mutex will prevent faults
472 * until we finish removing the folio.
474 * This race can only happen in the hole punch case.
475 * Getting here in a truncate operation is a bug.
477 if (unlikely(folio_mapped(folio))) {
480 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
481 i_mmap_lock_write(mapping);
482 mutex_lock(&hugetlb_fault_mutex_table[hash]);
483 hugetlb_vmdelete_list(&mapping->i_mmap,
484 index * pages_per_huge_page(h),
485 (index + 1) * pages_per_huge_page(h),
486 ZAP_FLAG_DROP_MARKER);
487 i_mmap_unlock_write(mapping);
492 * We must free the huge page and remove from page
493 * cache (remove_huge_page) BEFORE removing the
494 * region/reserve map (hugetlb_unreserve_pages). In
495 * rare out of memory conditions, removal of the
496 * region/reserve map could fail. Correspondingly,
497 * the subpool and global reserve usage count can need
500 VM_BUG_ON(HPageRestoreReserve(&folio->page));
501 remove_huge_page(&folio->page);
504 if (unlikely(hugetlb_unreserve_pages(inode,
505 index, index + 1, 1)))
506 hugetlb_fix_reserve_counts(inode);
511 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
513 folio_batch_release(&fbatch);
518 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
521 static void hugetlbfs_evict_inode(struct inode *inode)
523 struct resv_map *resv_map;
525 remove_inode_hugepages(inode, 0, LLONG_MAX);
528 * Get the resv_map from the address space embedded in the inode.
529 * This is the address space which points to any resv_map allocated
530 * at inode creation time. If this is a device special inode,
531 * i_mapping may not point to the original address space.
533 resv_map = (struct resv_map *)(&inode->i_data)->private_data;
534 /* Only regular and link inodes have associated reserve maps */
536 resv_map_release(&resv_map->refs);
540 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
543 struct address_space *mapping = inode->i_mapping;
544 struct hstate *h = hstate_inode(inode);
546 BUG_ON(offset & ~huge_page_mask(h));
547 pgoff = offset >> PAGE_SHIFT;
549 i_mmap_lock_write(mapping);
550 i_size_write(inode, offset);
551 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
552 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
553 ZAP_FLAG_DROP_MARKER);
554 i_mmap_unlock_write(mapping);
555 remove_inode_hugepages(inode, offset, LLONG_MAX);
558 static void hugetlbfs_zero_partial_page(struct hstate *h,
559 struct address_space *mapping,
563 pgoff_t idx = start >> huge_page_shift(h);
566 folio = filemap_lock_folio(mapping, idx);
570 start = start & ~huge_page_mask(h);
571 end = end & ~huge_page_mask(h);
573 end = huge_page_size(h);
575 folio_zero_segment(folio, (size_t)start, (size_t)end);
581 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
583 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
584 struct address_space *mapping = inode->i_mapping;
585 struct hstate *h = hstate_inode(inode);
586 loff_t hpage_size = huge_page_size(h);
587 loff_t hole_start, hole_end;
590 * hole_start and hole_end indicate the full pages within the hole.
592 hole_start = round_up(offset, hpage_size);
593 hole_end = round_down(offset + len, hpage_size);
597 /* protected by i_rwsem */
598 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
603 i_mmap_lock_write(mapping);
605 /* If range starts before first full page, zero partial page. */
606 if (offset < hole_start)
607 hugetlbfs_zero_partial_page(h, mapping,
608 offset, min(offset + len, hole_start));
610 /* Unmap users of full pages in the hole. */
611 if (hole_end > hole_start) {
612 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
613 hugetlb_vmdelete_list(&mapping->i_mmap,
614 hole_start >> PAGE_SHIFT,
615 hole_end >> PAGE_SHIFT, 0);
618 /* If range extends beyond last full page, zero partial page. */
619 if ((offset + len) > hole_end && (offset + len) > hole_start)
620 hugetlbfs_zero_partial_page(h, mapping,
621 hole_end, offset + len);
623 i_mmap_unlock_write(mapping);
625 /* Remove full pages from the file. */
626 if (hole_end > hole_start)
627 remove_inode_hugepages(inode, hole_start, hole_end);
634 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
637 struct inode *inode = file_inode(file);
638 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
639 struct address_space *mapping = inode->i_mapping;
640 struct hstate *h = hstate_inode(inode);
641 struct vm_area_struct pseudo_vma;
642 struct mm_struct *mm = current->mm;
643 loff_t hpage_size = huge_page_size(h);
644 unsigned long hpage_shift = huge_page_shift(h);
645 pgoff_t start, index, end;
649 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
652 if (mode & FALLOC_FL_PUNCH_HOLE)
653 return hugetlbfs_punch_hole(inode, offset, len);
656 * Default preallocate case.
657 * For this range, start is rounded down and end is rounded up
658 * as well as being converted to page offsets.
660 start = offset >> hpage_shift;
661 end = (offset + len + hpage_size - 1) >> hpage_shift;
665 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
666 error = inode_newsize_ok(inode, offset + len);
670 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
676 * Initialize a pseudo vma as this is required by the huge page
677 * allocation routines. If NUMA is configured, use page index
678 * as input to create an allocation policy.
680 vma_init(&pseudo_vma, mm);
681 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
682 pseudo_vma.vm_file = file;
684 for (index = start; index < end; index++) {
686 * This is supposed to be the vaddr where the page is being
687 * faulted in, but we have no vaddr here.
695 * fallocate(2) manpage permits EINTR; we may have been
696 * interrupted because we are using up too much memory.
698 if (signal_pending(current)) {
703 /* Set numa allocation policy based on index */
704 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
706 /* addr is the offset within the file (zero based) */
707 addr = index * hpage_size;
710 * fault mutex taken here, protects against fault path
711 * and hole punch. inode_lock previously taken protects
712 * against truncation.
714 hash = hugetlb_fault_mutex_hash(mapping, index);
715 mutex_lock(&hugetlb_fault_mutex_table[hash]);
717 /* See if already present in mapping to avoid alloc/free */
718 page = find_get_page(mapping, index);
721 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
722 hugetlb_drop_vma_policy(&pseudo_vma);
727 * Allocate page without setting the avoid_reserve argument.
728 * There certainly are no reserves associated with the
729 * pseudo_vma. However, there could be shared mappings with
730 * reserves for the file at the inode level. If we fallocate
731 * pages in these areas, we need to consume the reserves
732 * to keep reservation accounting consistent.
734 page = alloc_huge_page(&pseudo_vma, addr, 0);
735 hugetlb_drop_vma_policy(&pseudo_vma);
737 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
738 error = PTR_ERR(page);
741 clear_huge_page(page, addr, pages_per_huge_page(h));
742 __SetPageUptodate(page);
743 error = huge_add_to_page_cache(page, mapping, index);
744 if (unlikely(error)) {
745 restore_reserve_on_error(h, &pseudo_vma, addr, page);
747 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
751 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
753 SetHPageMigratable(page);
755 * unlock_page because locked by huge_add_to_page_cache()
756 * put_page() due to reference from alloc_huge_page()
762 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
763 i_size_write(inode, offset + len);
764 inode->i_ctime = current_time(inode);
770 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
771 struct dentry *dentry, struct iattr *attr)
773 struct inode *inode = d_inode(dentry);
774 struct hstate *h = hstate_inode(inode);
776 unsigned int ia_valid = attr->ia_valid;
777 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
779 error = setattr_prepare(&init_user_ns, dentry, attr);
783 if (ia_valid & ATTR_SIZE) {
784 loff_t oldsize = inode->i_size;
785 loff_t newsize = attr->ia_size;
787 if (newsize & ~huge_page_mask(h))
789 /* protected by i_rwsem */
790 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
791 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
793 hugetlb_vmtruncate(inode, newsize);
796 setattr_copy(&init_user_ns, inode, attr);
797 mark_inode_dirty(inode);
801 static struct inode *hugetlbfs_get_root(struct super_block *sb,
802 struct hugetlbfs_fs_context *ctx)
806 inode = new_inode(sb);
808 inode->i_ino = get_next_ino();
809 inode->i_mode = S_IFDIR | ctx->mode;
810 inode->i_uid = ctx->uid;
811 inode->i_gid = ctx->gid;
812 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
813 inode->i_op = &hugetlbfs_dir_inode_operations;
814 inode->i_fop = &simple_dir_operations;
815 /* directory inodes start off with i_nlink == 2 (for "." entry) */
817 lockdep_annotate_inode_mutex_key(inode);
823 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
824 * be taken from reclaim -- unlike regular filesystems. This needs an
825 * annotation because huge_pmd_share() does an allocation under hugetlb's
828 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
830 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
832 umode_t mode, dev_t dev)
835 struct resv_map *resv_map = NULL;
838 * Reserve maps are only needed for inodes that can have associated
841 if (S_ISREG(mode) || S_ISLNK(mode)) {
842 resv_map = resv_map_alloc();
847 inode = new_inode(sb);
849 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
851 inode->i_ino = get_next_ino();
852 inode_init_owner(&init_user_ns, inode, dir, mode);
853 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
854 &hugetlbfs_i_mmap_rwsem_key);
855 inode->i_mapping->a_ops = &hugetlbfs_aops;
856 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
857 inode->i_mapping->private_data = resv_map;
858 info->seals = F_SEAL_SEAL;
859 switch (mode & S_IFMT) {
861 init_special_inode(inode, mode, dev);
864 inode->i_op = &hugetlbfs_inode_operations;
865 inode->i_fop = &hugetlbfs_file_operations;
868 inode->i_op = &hugetlbfs_dir_inode_operations;
869 inode->i_fop = &simple_dir_operations;
871 /* directory inodes start off with i_nlink == 2 (for "." entry) */
875 inode->i_op = &page_symlink_inode_operations;
876 inode_nohighmem(inode);
879 lockdep_annotate_inode_mutex_key(inode);
882 kref_put(&resv_map->refs, resv_map_release);
889 * File creation. Allocate an inode, and we're done..
891 static int do_hugetlbfs_mknod(struct inode *dir,
892 struct dentry *dentry,
900 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
902 dir->i_ctime = dir->i_mtime = current_time(dir);
904 d_tmpfile(dentry, inode);
906 d_instantiate(dentry, inode);
907 dget(dentry);/* Extra count - pin the dentry in core */
914 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
915 struct dentry *dentry, umode_t mode, dev_t dev)
917 return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
920 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
921 struct dentry *dentry, umode_t mode)
923 int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
930 static int hugetlbfs_create(struct user_namespace *mnt_userns,
931 struct inode *dir, struct dentry *dentry,
932 umode_t mode, bool excl)
934 return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
937 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
938 struct inode *dir, struct dentry *dentry,
941 return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
944 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
945 struct inode *dir, struct dentry *dentry,
951 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
953 int l = strlen(symname)+1;
954 error = page_symlink(inode, symname, l);
956 d_instantiate(dentry, inode);
961 dir->i_ctime = dir->i_mtime = current_time(dir);
966 #ifdef CONFIG_MIGRATION
967 static int hugetlbfs_migrate_folio(struct address_space *mapping,
968 struct folio *dst, struct folio *src,
969 enum migrate_mode mode)
973 rc = migrate_huge_page_move_mapping(mapping, dst, src);
974 if (rc != MIGRATEPAGE_SUCCESS)
977 if (hugetlb_page_subpool(&src->page)) {
978 hugetlb_set_page_subpool(&dst->page,
979 hugetlb_page_subpool(&src->page));
980 hugetlb_set_page_subpool(&src->page, NULL);
983 if (mode != MIGRATE_SYNC_NO_COPY)
984 folio_migrate_copy(dst, src);
986 folio_migrate_flags(dst, src);
988 return MIGRATEPAGE_SUCCESS;
991 #define hugetlbfs_migrate_folio NULL
994 static int hugetlbfs_error_remove_page(struct address_space *mapping,
997 struct inode *inode = mapping->host;
998 pgoff_t index = page->index;
1000 remove_huge_page(page);
1001 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
1002 hugetlb_fix_reserve_counts(inode);
1008 * Display the mount options in /proc/mounts.
1010 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1012 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1013 struct hugepage_subpool *spool = sbinfo->spool;
1014 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1015 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1018 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1019 seq_printf(m, ",uid=%u",
1020 from_kuid_munged(&init_user_ns, sbinfo->uid));
1021 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1022 seq_printf(m, ",gid=%u",
1023 from_kgid_munged(&init_user_ns, sbinfo->gid));
1024 if (sbinfo->mode != 0755)
1025 seq_printf(m, ",mode=%o", sbinfo->mode);
1026 if (sbinfo->max_inodes != -1)
1027 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1031 if (hpage_size >= 1024) {
1035 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1037 if (spool->max_hpages != -1)
1038 seq_printf(m, ",size=%llu",
1039 (unsigned long long)spool->max_hpages << hpage_shift);
1040 if (spool->min_hpages != -1)
1041 seq_printf(m, ",min_size=%llu",
1042 (unsigned long long)spool->min_hpages << hpage_shift);
1047 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1049 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1050 struct hstate *h = hstate_inode(d_inode(dentry));
1052 buf->f_type = HUGETLBFS_MAGIC;
1053 buf->f_bsize = huge_page_size(h);
1055 spin_lock(&sbinfo->stat_lock);
1056 /* If no limits set, just report 0 for max/free/used
1057 * blocks, like simple_statfs() */
1058 if (sbinfo->spool) {
1061 spin_lock_irq(&sbinfo->spool->lock);
1062 buf->f_blocks = sbinfo->spool->max_hpages;
1063 free_pages = sbinfo->spool->max_hpages
1064 - sbinfo->spool->used_hpages;
1065 buf->f_bavail = buf->f_bfree = free_pages;
1066 spin_unlock_irq(&sbinfo->spool->lock);
1067 buf->f_files = sbinfo->max_inodes;
1068 buf->f_ffree = sbinfo->free_inodes;
1070 spin_unlock(&sbinfo->stat_lock);
1072 buf->f_namelen = NAME_MAX;
1076 static void hugetlbfs_put_super(struct super_block *sb)
1078 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1081 sb->s_fs_info = NULL;
1084 hugepage_put_subpool(sbi->spool);
1090 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1092 if (sbinfo->free_inodes >= 0) {
1093 spin_lock(&sbinfo->stat_lock);
1094 if (unlikely(!sbinfo->free_inodes)) {
1095 spin_unlock(&sbinfo->stat_lock);
1098 sbinfo->free_inodes--;
1099 spin_unlock(&sbinfo->stat_lock);
1105 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1107 if (sbinfo->free_inodes >= 0) {
1108 spin_lock(&sbinfo->stat_lock);
1109 sbinfo->free_inodes++;
1110 spin_unlock(&sbinfo->stat_lock);
1115 static struct kmem_cache *hugetlbfs_inode_cachep;
1117 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1119 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1120 struct hugetlbfs_inode_info *p;
1122 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1124 p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1126 hugetlbfs_inc_free_inodes(sbinfo);
1131 * Any time after allocation, hugetlbfs_destroy_inode can be called
1132 * for the inode. mpol_free_shared_policy is unconditionally called
1133 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1134 * in case of a quick call to destroy.
1136 * Note that the policy is initialized even if we are creating a
1137 * private inode. This simplifies hugetlbfs_destroy_inode.
1139 mpol_shared_policy_init(&p->policy, NULL);
1141 return &p->vfs_inode;
1144 static void hugetlbfs_free_inode(struct inode *inode)
1146 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1149 static void hugetlbfs_destroy_inode(struct inode *inode)
1151 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1152 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1155 static const struct address_space_operations hugetlbfs_aops = {
1156 .write_begin = hugetlbfs_write_begin,
1157 .write_end = hugetlbfs_write_end,
1158 .dirty_folio = noop_dirty_folio,
1159 .migrate_folio = hugetlbfs_migrate_folio,
1160 .error_remove_page = hugetlbfs_error_remove_page,
1164 static void init_once(void *foo)
1166 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1168 inode_init_once(&ei->vfs_inode);
1171 const struct file_operations hugetlbfs_file_operations = {
1172 .read_iter = hugetlbfs_read_iter,
1173 .mmap = hugetlbfs_file_mmap,
1174 .fsync = noop_fsync,
1175 .get_unmapped_area = hugetlb_get_unmapped_area,
1176 .llseek = default_llseek,
1177 .fallocate = hugetlbfs_fallocate,
1180 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1181 .create = hugetlbfs_create,
1182 .lookup = simple_lookup,
1183 .link = simple_link,
1184 .unlink = simple_unlink,
1185 .symlink = hugetlbfs_symlink,
1186 .mkdir = hugetlbfs_mkdir,
1187 .rmdir = simple_rmdir,
1188 .mknod = hugetlbfs_mknod,
1189 .rename = simple_rename,
1190 .setattr = hugetlbfs_setattr,
1191 .tmpfile = hugetlbfs_tmpfile,
1194 static const struct inode_operations hugetlbfs_inode_operations = {
1195 .setattr = hugetlbfs_setattr,
1198 static const struct super_operations hugetlbfs_ops = {
1199 .alloc_inode = hugetlbfs_alloc_inode,
1200 .free_inode = hugetlbfs_free_inode,
1201 .destroy_inode = hugetlbfs_destroy_inode,
1202 .evict_inode = hugetlbfs_evict_inode,
1203 .statfs = hugetlbfs_statfs,
1204 .put_super = hugetlbfs_put_super,
1205 .show_options = hugetlbfs_show_options,
1209 * Convert size option passed from command line to number of huge pages
1210 * in the pool specified by hstate. Size option could be in bytes
1211 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1214 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1215 enum hugetlbfs_size_type val_type)
1217 if (val_type == NO_SIZE)
1220 if (val_type == SIZE_PERCENT) {
1221 size_opt <<= huge_page_shift(h);
1222 size_opt *= h->max_huge_pages;
1223 do_div(size_opt, 100);
1226 size_opt >>= huge_page_shift(h);
1231 * Parse one mount parameter.
1233 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1235 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1236 struct fs_parse_result result;
1241 opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1247 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1248 if (!uid_valid(ctx->uid))
1253 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1254 if (!gid_valid(ctx->gid))
1259 ctx->mode = result.uint_32 & 01777U;
1263 /* memparse() will accept a K/M/G without a digit */
1264 if (!isdigit(param->string[0]))
1266 ctx->max_size_opt = memparse(param->string, &rest);
1267 ctx->max_val_type = SIZE_STD;
1269 ctx->max_val_type = SIZE_PERCENT;
1273 /* memparse() will accept a K/M/G without a digit */
1274 if (!isdigit(param->string[0]))
1276 ctx->nr_inodes = memparse(param->string, &rest);
1280 ps = memparse(param->string, &rest);
1281 ctx->hstate = size_to_hstate(ps);
1283 pr_err("Unsupported page size %lu MB\n", ps >> 20);
1289 /* memparse() will accept a K/M/G without a digit */
1290 if (!isdigit(param->string[0]))
1292 ctx->min_size_opt = memparse(param->string, &rest);
1293 ctx->min_val_type = SIZE_STD;
1295 ctx->min_val_type = SIZE_PERCENT;
1303 return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1304 param->string, param->key);
1308 * Validate the parsed options.
1310 static int hugetlbfs_validate(struct fs_context *fc)
1312 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1315 * Use huge page pool size (in hstate) to convert the size
1316 * options to number of huge pages. If NO_SIZE, -1 is returned.
1318 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1321 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1326 * If max_size was specified, then min_size must be smaller
1328 if (ctx->max_val_type > NO_SIZE &&
1329 ctx->min_hpages > ctx->max_hpages) {
1330 pr_err("Minimum size can not be greater than maximum size\n");
1338 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1340 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1341 struct hugetlbfs_sb_info *sbinfo;
1343 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1346 sb->s_fs_info = sbinfo;
1347 spin_lock_init(&sbinfo->stat_lock);
1348 sbinfo->hstate = ctx->hstate;
1349 sbinfo->max_inodes = ctx->nr_inodes;
1350 sbinfo->free_inodes = ctx->nr_inodes;
1351 sbinfo->spool = NULL;
1352 sbinfo->uid = ctx->uid;
1353 sbinfo->gid = ctx->gid;
1354 sbinfo->mode = ctx->mode;
1357 * Allocate and initialize subpool if maximum or minimum size is
1358 * specified. Any needed reservations (for minimum size) are taken
1359 * taken when the subpool is created.
1361 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1362 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1368 sb->s_maxbytes = MAX_LFS_FILESIZE;
1369 sb->s_blocksize = huge_page_size(ctx->hstate);
1370 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1371 sb->s_magic = HUGETLBFS_MAGIC;
1372 sb->s_op = &hugetlbfs_ops;
1373 sb->s_time_gran = 1;
1376 * Due to the special and limited functionality of hugetlbfs, it does
1377 * not work well as a stacking filesystem.
1379 sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1380 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1385 kfree(sbinfo->spool);
1390 static int hugetlbfs_get_tree(struct fs_context *fc)
1392 int err = hugetlbfs_validate(fc);
1395 return get_tree_nodev(fc, hugetlbfs_fill_super);
1398 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1400 kfree(fc->fs_private);
1403 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1404 .free = hugetlbfs_fs_context_free,
1405 .parse_param = hugetlbfs_parse_param,
1406 .get_tree = hugetlbfs_get_tree,
1409 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1411 struct hugetlbfs_fs_context *ctx;
1413 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1417 ctx->max_hpages = -1; /* No limit on size by default */
1418 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1419 ctx->uid = current_fsuid();
1420 ctx->gid = current_fsgid();
1422 ctx->hstate = &default_hstate;
1423 ctx->min_hpages = -1; /* No default minimum size */
1424 ctx->max_val_type = NO_SIZE;
1425 ctx->min_val_type = NO_SIZE;
1426 fc->fs_private = ctx;
1427 fc->ops = &hugetlbfs_fs_context_ops;
1431 static struct file_system_type hugetlbfs_fs_type = {
1432 .name = "hugetlbfs",
1433 .init_fs_context = hugetlbfs_init_fs_context,
1434 .parameters = hugetlb_fs_parameters,
1435 .kill_sb = kill_litter_super,
1438 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1440 static int can_do_hugetlb_shm(void)
1443 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1444 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1447 static int get_hstate_idx(int page_size_log)
1449 struct hstate *h = hstate_sizelog(page_size_log);
1453 return hstate_index(h);
1457 * Note that size should be aligned to proper hugepage size in caller side,
1458 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1460 struct file *hugetlb_file_setup(const char *name, size_t size,
1461 vm_flags_t acctflag, int creat_flags,
1464 struct inode *inode;
1465 struct vfsmount *mnt;
1469 hstate_idx = get_hstate_idx(page_size_log);
1471 return ERR_PTR(-ENODEV);
1473 mnt = hugetlbfs_vfsmount[hstate_idx];
1475 return ERR_PTR(-ENOENT);
1477 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1478 struct ucounts *ucounts = current_ucounts();
1480 if (user_shm_lock(size, ucounts)) {
1481 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1482 current->comm, current->pid);
1483 user_shm_unlock(size, ucounts);
1485 return ERR_PTR(-EPERM);
1488 file = ERR_PTR(-ENOSPC);
1489 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1492 if (creat_flags == HUGETLB_SHMFS_INODE)
1493 inode->i_flags |= S_PRIVATE;
1495 inode->i_size = size;
1498 if (!hugetlb_reserve_pages(inode, 0,
1499 size >> huge_page_shift(hstate_inode(inode)), NULL,
1501 file = ERR_PTR(-ENOMEM);
1503 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1504 &hugetlbfs_file_operations);
1513 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1515 struct fs_context *fc;
1516 struct vfsmount *mnt;
1518 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1522 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1528 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1529 huge_page_size(h) >> 10);
1533 static int __init init_hugetlbfs_fs(void)
1535 struct vfsmount *mnt;
1540 if (!hugepages_supported()) {
1541 pr_info("disabling because there are no supported hugepage sizes\n");
1546 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1547 sizeof(struct hugetlbfs_inode_info),
1548 0, SLAB_ACCOUNT, init_once);
1549 if (hugetlbfs_inode_cachep == NULL)
1552 error = register_filesystem(&hugetlbfs_fs_type);
1556 /* default hstate mount is required */
1557 mnt = mount_one_hugetlbfs(&default_hstate);
1559 error = PTR_ERR(mnt);
1562 hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1564 /* other hstates are optional */
1566 for_each_hstate(h) {
1567 if (i == default_hstate_idx) {
1572 mnt = mount_one_hugetlbfs(h);
1574 hugetlbfs_vfsmount[i] = NULL;
1576 hugetlbfs_vfsmount[i] = mnt;
1583 (void)unregister_filesystem(&hugetlbfs_fs_type);
1585 kmem_cache_destroy(hugetlbfs_inode_cachep);
1589 fs_initcall(init_hugetlbfs_fs)