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/falloc.h>
16 #include <linux/mount.h>
17 #include <linux/file.h>
18 #include <linux/kernel.h>
19 #include <linux/writeback.h>
20 #include <linux/pagemap.h>
21 #include <linux/highmem.h>
22 #include <linux/init.h>
23 #include <linux/string.h>
24 #include <linux/capability.h>
25 #include <linux/ctype.h>
26 #include <linux/backing-dev.h>
27 #include <linux/hugetlb.h>
28 #include <linux/pagevec.h>
29 #include <linux/fs_parser.h>
30 #include <linux/mman.h>
31 #include <linux/slab.h>
32 #include <linux/dnotify.h>
33 #include <linux/statfs.h>
34 #include <linux/security.h>
35 #include <linux/magic.h>
36 #include <linux/migrate.h>
37 #include <linux/uio.h>
39 #include <linux/uaccess.h>
40 #include <linux/sched/mm.h>
42 static const struct address_space_operations hugetlbfs_aops;
43 const struct file_operations hugetlbfs_file_operations;
44 static const struct inode_operations hugetlbfs_dir_inode_operations;
45 static const struct inode_operations hugetlbfs_inode_operations;
47 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
49 struct hugetlbfs_fs_context {
50 struct hstate *hstate;
51 unsigned long long max_size_opt;
52 unsigned long long min_size_opt;
56 enum hugetlbfs_size_type max_val_type;
57 enum hugetlbfs_size_type min_val_type;
63 int sysctl_hugetlb_shm_group;
75 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
76 fsparam_u32 ("gid", Opt_gid),
77 fsparam_string("min_size", Opt_min_size),
78 fsparam_u32oct("mode", Opt_mode),
79 fsparam_string("nr_inodes", Opt_nr_inodes),
80 fsparam_string("pagesize", Opt_pagesize),
81 fsparam_string("size", Opt_size),
82 fsparam_u32 ("uid", Opt_uid),
87 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
88 struct inode *inode, pgoff_t index)
90 vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
94 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
96 mpol_cond_put(vma->vm_policy);
99 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
100 struct inode *inode, pgoff_t index)
104 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
110 * Mask used when checking the page offset value passed in via system
111 * calls. This value will be converted to a loff_t which is signed.
112 * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
113 * value. The extra bit (- 1 in the shift value) is to take the sign
116 #define PGOFF_LOFFT_MAX \
117 (((1UL << (PAGE_SHIFT + 1)) - 1) << (BITS_PER_LONG - (PAGE_SHIFT + 1)))
119 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
121 struct inode *inode = file_inode(file);
122 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
125 struct hstate *h = hstate_file(file);
128 * vma address alignment (but not the pgoff alignment) has
129 * already been checked by prepare_hugepage_range. If you add
130 * any error returns here, do so after setting VM_HUGETLB, so
131 * is_vm_hugetlb_page tests below unmap_region go the right
132 * way when do_mmap unwinds (may be important on powerpc
135 vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
136 vma->vm_ops = &hugetlb_vm_ops;
138 ret = seal_check_future_write(info->seals, vma);
143 * page based offset in vm_pgoff could be sufficiently large to
144 * overflow a loff_t when converted to byte offset. This can
145 * only happen on architectures where sizeof(loff_t) ==
146 * sizeof(unsigned long). So, only check in those instances.
148 if (sizeof(unsigned long) == sizeof(loff_t)) {
149 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
153 /* must be huge page aligned */
154 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
157 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
158 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
159 /* check for overflow */
167 if (!hugetlb_reserve_pages(inode,
168 vma->vm_pgoff >> huge_page_order(h),
169 len >> huge_page_shift(h), vma,
174 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
175 i_size_write(inode, len);
183 * Called under mmap_write_lock(mm).
187 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
188 unsigned long len, unsigned long pgoff, unsigned long flags)
190 struct hstate *h = hstate_file(file);
191 struct vm_unmapped_area_info info;
195 info.low_limit = current->mm->mmap_base;
196 info.high_limit = arch_get_mmap_end(addr, len, flags);
197 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
198 info.align_offset = 0;
199 return vm_unmapped_area(&info);
203 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
204 unsigned long len, unsigned long pgoff, unsigned long flags)
206 struct hstate *h = hstate_file(file);
207 struct vm_unmapped_area_info info;
209 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
211 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
212 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
213 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
214 info.align_offset = 0;
215 addr = vm_unmapped_area(&info);
218 * A failed mmap() very likely causes application failure,
219 * so fall back to the bottom-up function here. This scenario
220 * can happen with large stack limits and large mmap()
223 if (unlikely(offset_in_page(addr))) {
224 VM_BUG_ON(addr != -ENOMEM);
226 info.low_limit = current->mm->mmap_base;
227 info.high_limit = arch_get_mmap_end(addr, len, flags);
228 addr = vm_unmapped_area(&info);
235 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
236 unsigned long len, unsigned long pgoff,
239 struct mm_struct *mm = current->mm;
240 struct vm_area_struct *vma;
241 struct hstate *h = hstate_file(file);
242 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
244 if (len & ~huge_page_mask(h))
249 if (flags & MAP_FIXED) {
250 if (prepare_hugepage_range(file, addr, len))
256 addr = ALIGN(addr, huge_page_size(h));
257 vma = find_vma(mm, addr);
258 if (mmap_end - len >= addr &&
259 (!vma || addr + len <= vm_start_gap(vma)))
264 * Use mm->get_unmapped_area value as a hint to use topdown routine.
265 * If architectures have special needs, they should define their own
266 * version of hugetlb_get_unmapped_area.
268 if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
269 return hugetlb_get_unmapped_area_topdown(file, addr, len,
271 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
275 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
277 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
278 unsigned long len, unsigned long pgoff,
281 return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
286 hugetlbfs_read_actor(struct page *page, unsigned long offset,
287 struct iov_iter *to, unsigned long size)
292 /* Find which 4k chunk and offset with in that chunk */
293 i = offset >> PAGE_SHIFT;
294 offset = offset & ~PAGE_MASK;
298 chunksize = PAGE_SIZE;
301 if (chunksize > size)
303 n = copy_page_to_iter(&page[i], offset, chunksize, to);
315 * Support for read() - Find the page attached to f_mapping and copy out the
316 * data. This provides functionality similar to filemap_read().
318 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
320 struct file *file = iocb->ki_filp;
321 struct hstate *h = hstate_file(file);
322 struct address_space *mapping = file->f_mapping;
323 struct inode *inode = mapping->host;
324 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
325 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
326 unsigned long end_index;
330 while (iov_iter_count(to)) {
334 /* nr is the maximum number of bytes to copy from this page */
335 nr = huge_page_size(h);
336 isize = i_size_read(inode);
339 end_index = (isize - 1) >> huge_page_shift(h);
340 if (index > end_index)
342 if (index == end_index) {
343 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
350 page = find_lock_page(mapping, index);
351 if (unlikely(page == NULL)) {
353 * We have a HOLE, zero out the user-buffer for the
354 * length of the hole or request.
356 copied = iov_iter_zero(nr, to);
361 * We have the page, copy it to user space buffer.
363 copied = hugetlbfs_read_actor(page, offset, to, nr);
368 if (copied != nr && iov_iter_count(to)) {
373 index += offset >> huge_page_shift(h);
374 offset &= ~huge_page_mask(h);
376 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
380 static int hugetlbfs_write_begin(struct file *file,
381 struct address_space *mapping,
382 loff_t pos, unsigned len,
383 struct page **pagep, void **fsdata)
388 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
389 loff_t pos, unsigned len, unsigned copied,
390 struct page *page, void *fsdata)
396 static void remove_huge_page(struct page *page)
398 ClearPageDirty(page);
399 ClearPageUptodate(page);
400 delete_from_page_cache(page);
404 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
405 zap_flags_t zap_flags)
407 struct vm_area_struct *vma;
410 * end == 0 indicates that the entire range after start should be
411 * unmapped. Note, end is exclusive, whereas the interval tree takes
412 * an inclusive "last".
414 vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
415 unsigned long v_offset;
419 * Can the expression below overflow on 32-bit arches?
420 * No, because the interval tree returns us only those vmas
421 * which overlap the truncated area starting at pgoff,
422 * and no vma on a 32-bit arch can span beyond the 4GB.
424 if (vma->vm_pgoff < start)
425 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
432 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
434 if (v_end > vma->vm_end)
438 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
444 * remove_inode_hugepages handles two distinct cases: truncation and hole
445 * punch. There are subtle differences in operation for each case.
447 * truncation is indicated by end of range being LLONG_MAX
448 * In this case, we first scan the range and release found pages.
449 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
450 * maps and global counts. Page faults can not race with truncation
451 * in this routine. hugetlb_no_page() holds i_mmap_rwsem and prevents
452 * page faults in the truncated range by checking i_size. i_size is
453 * modified while holding i_mmap_rwsem.
454 * hole punch is indicated if end is not LLONG_MAX
455 * In the hole punch case we scan the range and release found pages.
456 * Only when releasing a page is the associated region/reserve map
457 * deleted. The region/reserve map for ranges without associated
458 * pages are not modified. Page faults can race with hole punch.
459 * This is indicated if we find a mapped page.
460 * Note: If the passed end of range value is beyond the end of file, but
461 * not LLONG_MAX this routine still performs a hole punch operation.
463 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
466 struct hstate *h = hstate_inode(inode);
467 struct address_space *mapping = &inode->i_data;
468 const pgoff_t start = lstart >> huge_page_shift(h);
469 const pgoff_t end = lend >> huge_page_shift(h);
470 struct folio_batch fbatch;
473 bool truncate_op = (lend == LLONG_MAX);
475 folio_batch_init(&fbatch);
477 while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
478 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
479 struct folio *folio = fbatch.folios[i];
482 index = folio->index;
485 * Only need to hold the fault mutex in the
486 * hole punch case. This prevents races with
487 * page faults. Races are not possible in the
488 * case of truncation.
490 hash = hugetlb_fault_mutex_hash(mapping, index);
491 mutex_lock(&hugetlb_fault_mutex_table[hash]);
495 * If folio is mapped, it was faulted in after being
496 * unmapped in caller. Unmap (again) now after taking
497 * the fault mutex. The mutex will prevent faults
498 * until we finish removing the folio.
500 * This race can only happen in the hole punch case.
501 * Getting here in a truncate operation is a bug.
503 if (unlikely(folio_mapped(folio))) {
506 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
507 i_mmap_lock_write(mapping);
508 mutex_lock(&hugetlb_fault_mutex_table[hash]);
509 hugetlb_vmdelete_list(&mapping->i_mmap,
510 index * pages_per_huge_page(h),
511 (index + 1) * pages_per_huge_page(h),
512 ZAP_FLAG_DROP_MARKER);
513 i_mmap_unlock_write(mapping);
518 * We must free the huge page and remove from page
519 * cache (remove_huge_page) BEFORE removing the
520 * region/reserve map (hugetlb_unreserve_pages). In
521 * rare out of memory conditions, removal of the
522 * region/reserve map could fail. Correspondingly,
523 * the subpool and global reserve usage count can need
526 VM_BUG_ON(HPageRestoreReserve(&folio->page));
527 remove_huge_page(&folio->page);
530 if (unlikely(hugetlb_unreserve_pages(inode,
531 index, index + 1, 1)))
532 hugetlb_fix_reserve_counts(inode);
537 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
539 folio_batch_release(&fbatch);
544 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
547 static void hugetlbfs_evict_inode(struct inode *inode)
549 struct resv_map *resv_map;
551 remove_inode_hugepages(inode, 0, LLONG_MAX);
554 * Get the resv_map from the address space embedded in the inode.
555 * This is the address space which points to any resv_map allocated
556 * at inode creation time. If this is a device special inode,
557 * i_mapping may not point to the original address space.
559 resv_map = (struct resv_map *)(&inode->i_data)->private_data;
560 /* Only regular and link inodes have associated reserve maps */
562 resv_map_release(&resv_map->refs);
566 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
569 struct address_space *mapping = inode->i_mapping;
570 struct hstate *h = hstate_inode(inode);
572 BUG_ON(offset & ~huge_page_mask(h));
573 pgoff = offset >> PAGE_SHIFT;
575 i_mmap_lock_write(mapping);
576 i_size_write(inode, offset);
577 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
578 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
579 ZAP_FLAG_DROP_MARKER);
580 i_mmap_unlock_write(mapping);
581 remove_inode_hugepages(inode, offset, LLONG_MAX);
584 static void hugetlbfs_zero_partial_page(struct hstate *h,
585 struct address_space *mapping,
589 pgoff_t idx = start >> huge_page_shift(h);
592 folio = filemap_lock_folio(mapping, idx);
596 start = start & ~huge_page_mask(h);
597 end = end & ~huge_page_mask(h);
599 end = huge_page_size(h);
601 folio_zero_segment(folio, (size_t)start, (size_t)end);
607 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
609 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
610 struct address_space *mapping = inode->i_mapping;
611 struct hstate *h = hstate_inode(inode);
612 loff_t hpage_size = huge_page_size(h);
613 loff_t hole_start, hole_end;
616 * hole_start and hole_end indicate the full pages within the hole.
618 hole_start = round_up(offset, hpage_size);
619 hole_end = round_down(offset + len, hpage_size);
623 /* protected by i_rwsem */
624 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
629 i_mmap_lock_write(mapping);
631 /* If range starts before first full page, zero partial page. */
632 if (offset < hole_start)
633 hugetlbfs_zero_partial_page(h, mapping,
634 offset, min(offset + len, hole_start));
636 /* Unmap users of full pages in the hole. */
637 if (hole_end > hole_start) {
638 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
639 hugetlb_vmdelete_list(&mapping->i_mmap,
640 hole_start >> PAGE_SHIFT,
641 hole_end >> PAGE_SHIFT, 0);
644 /* If range extends beyond last full page, zero partial page. */
645 if ((offset + len) > hole_end && (offset + len) > hole_start)
646 hugetlbfs_zero_partial_page(h, mapping,
647 hole_end, offset + len);
649 i_mmap_unlock_write(mapping);
651 /* Remove full pages from the file. */
652 if (hole_end > hole_start)
653 remove_inode_hugepages(inode, hole_start, hole_end);
660 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
663 struct inode *inode = file_inode(file);
664 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
665 struct address_space *mapping = inode->i_mapping;
666 struct hstate *h = hstate_inode(inode);
667 struct vm_area_struct pseudo_vma;
668 struct mm_struct *mm = current->mm;
669 loff_t hpage_size = huge_page_size(h);
670 unsigned long hpage_shift = huge_page_shift(h);
671 pgoff_t start, index, end;
675 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
678 if (mode & FALLOC_FL_PUNCH_HOLE)
679 return hugetlbfs_punch_hole(inode, offset, len);
682 * Default preallocate case.
683 * For this range, start is rounded down and end is rounded up
684 * as well as being converted to page offsets.
686 start = offset >> hpage_shift;
687 end = (offset + len + hpage_size - 1) >> hpage_shift;
691 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
692 error = inode_newsize_ok(inode, offset + len);
696 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
702 * Initialize a pseudo vma as this is required by the huge page
703 * allocation routines. If NUMA is configured, use page index
704 * as input to create an allocation policy.
706 vma_init(&pseudo_vma, mm);
707 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
708 pseudo_vma.vm_file = file;
710 for (index = start; index < end; index++) {
712 * This is supposed to be the vaddr where the page is being
713 * faulted in, but we have no vaddr here.
721 * fallocate(2) manpage permits EINTR; we may have been
722 * interrupted because we are using up too much memory.
724 if (signal_pending(current)) {
729 /* Set numa allocation policy based on index */
730 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
732 /* addr is the offset within the file (zero based) */
733 addr = index * hpage_size;
736 * fault mutex taken here, protects against fault path
737 * and hole punch. inode_lock previously taken protects
738 * against truncation.
740 hash = hugetlb_fault_mutex_hash(mapping, index);
741 mutex_lock(&hugetlb_fault_mutex_table[hash]);
743 /* See if already present in mapping to avoid alloc/free */
744 page = find_get_page(mapping, index);
747 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
748 hugetlb_drop_vma_policy(&pseudo_vma);
753 * Allocate page without setting the avoid_reserve argument.
754 * There certainly are no reserves associated with the
755 * pseudo_vma. However, there could be shared mappings with
756 * reserves for the file at the inode level. If we fallocate
757 * pages in these areas, we need to consume the reserves
758 * to keep reservation accounting consistent.
760 page = alloc_huge_page(&pseudo_vma, addr, 0);
761 hugetlb_drop_vma_policy(&pseudo_vma);
763 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
764 error = PTR_ERR(page);
767 clear_huge_page(page, addr, pages_per_huge_page(h));
768 __SetPageUptodate(page);
769 error = huge_add_to_page_cache(page, mapping, index);
770 if (unlikely(error)) {
771 restore_reserve_on_error(h, &pseudo_vma, addr, page);
773 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
777 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
779 SetHPageMigratable(page);
781 * unlock_page because locked by huge_add_to_page_cache()
782 * put_page() due to reference from alloc_huge_page()
788 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
789 i_size_write(inode, offset + len);
790 inode->i_ctime = current_time(inode);
796 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
797 struct dentry *dentry, struct iattr *attr)
799 struct inode *inode = d_inode(dentry);
800 struct hstate *h = hstate_inode(inode);
802 unsigned int ia_valid = attr->ia_valid;
803 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
805 error = setattr_prepare(&init_user_ns, dentry, attr);
809 if (ia_valid & ATTR_SIZE) {
810 loff_t oldsize = inode->i_size;
811 loff_t newsize = attr->ia_size;
813 if (newsize & ~huge_page_mask(h))
815 /* protected by i_rwsem */
816 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
817 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
819 hugetlb_vmtruncate(inode, newsize);
822 setattr_copy(&init_user_ns, inode, attr);
823 mark_inode_dirty(inode);
827 static struct inode *hugetlbfs_get_root(struct super_block *sb,
828 struct hugetlbfs_fs_context *ctx)
832 inode = new_inode(sb);
834 inode->i_ino = get_next_ino();
835 inode->i_mode = S_IFDIR | ctx->mode;
836 inode->i_uid = ctx->uid;
837 inode->i_gid = ctx->gid;
838 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
839 inode->i_op = &hugetlbfs_dir_inode_operations;
840 inode->i_fop = &simple_dir_operations;
841 /* directory inodes start off with i_nlink == 2 (for "." entry) */
843 lockdep_annotate_inode_mutex_key(inode);
849 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
850 * be taken from reclaim -- unlike regular filesystems. This needs an
851 * annotation because huge_pmd_share() does an allocation under hugetlb's
854 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
856 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
858 umode_t mode, dev_t dev)
861 struct resv_map *resv_map = NULL;
864 * Reserve maps are only needed for inodes that can have associated
867 if (S_ISREG(mode) || S_ISLNK(mode)) {
868 resv_map = resv_map_alloc();
873 inode = new_inode(sb);
875 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
877 inode->i_ino = get_next_ino();
878 inode_init_owner(&init_user_ns, inode, dir, mode);
879 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
880 &hugetlbfs_i_mmap_rwsem_key);
881 inode->i_mapping->a_ops = &hugetlbfs_aops;
882 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
883 inode->i_mapping->private_data = resv_map;
884 info->seals = F_SEAL_SEAL;
885 switch (mode & S_IFMT) {
887 init_special_inode(inode, mode, dev);
890 inode->i_op = &hugetlbfs_inode_operations;
891 inode->i_fop = &hugetlbfs_file_operations;
894 inode->i_op = &hugetlbfs_dir_inode_operations;
895 inode->i_fop = &simple_dir_operations;
897 /* directory inodes start off with i_nlink == 2 (for "." entry) */
901 inode->i_op = &page_symlink_inode_operations;
902 inode_nohighmem(inode);
905 lockdep_annotate_inode_mutex_key(inode);
908 kref_put(&resv_map->refs, resv_map_release);
915 * File creation. Allocate an inode, and we're done..
917 static int do_hugetlbfs_mknod(struct inode *dir,
918 struct dentry *dentry,
926 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
928 dir->i_ctime = dir->i_mtime = current_time(dir);
930 d_tmpfile(dentry, inode);
932 d_instantiate(dentry, inode);
933 dget(dentry);/* Extra count - pin the dentry in core */
940 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
941 struct dentry *dentry, umode_t mode, dev_t dev)
943 return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
946 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
947 struct dentry *dentry, umode_t mode)
949 int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
956 static int hugetlbfs_create(struct user_namespace *mnt_userns,
957 struct inode *dir, struct dentry *dentry,
958 umode_t mode, bool excl)
960 return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
963 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
964 struct inode *dir, struct dentry *dentry,
967 return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
970 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
971 struct inode *dir, struct dentry *dentry,
977 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
979 int l = strlen(symname)+1;
980 error = page_symlink(inode, symname, l);
982 d_instantiate(dentry, inode);
987 dir->i_ctime = dir->i_mtime = current_time(dir);
992 #ifdef CONFIG_MIGRATION
993 static int hugetlbfs_migrate_folio(struct address_space *mapping,
994 struct folio *dst, struct folio *src,
995 enum migrate_mode mode)
999 rc = migrate_huge_page_move_mapping(mapping, dst, src);
1000 if (rc != MIGRATEPAGE_SUCCESS)
1003 if (hugetlb_page_subpool(&src->page)) {
1004 hugetlb_set_page_subpool(&dst->page,
1005 hugetlb_page_subpool(&src->page));
1006 hugetlb_set_page_subpool(&src->page, NULL);
1009 if (mode != MIGRATE_SYNC_NO_COPY)
1010 folio_migrate_copy(dst, src);
1012 folio_migrate_flags(dst, src);
1014 return MIGRATEPAGE_SUCCESS;
1017 #define hugetlbfs_migrate_folio NULL
1020 static int hugetlbfs_error_remove_page(struct address_space *mapping,
1023 struct inode *inode = mapping->host;
1024 pgoff_t index = page->index;
1026 remove_huge_page(page);
1027 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
1028 hugetlb_fix_reserve_counts(inode);
1034 * Display the mount options in /proc/mounts.
1036 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1038 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1039 struct hugepage_subpool *spool = sbinfo->spool;
1040 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1041 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1044 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1045 seq_printf(m, ",uid=%u",
1046 from_kuid_munged(&init_user_ns, sbinfo->uid));
1047 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1048 seq_printf(m, ",gid=%u",
1049 from_kgid_munged(&init_user_ns, sbinfo->gid));
1050 if (sbinfo->mode != 0755)
1051 seq_printf(m, ",mode=%o", sbinfo->mode);
1052 if (sbinfo->max_inodes != -1)
1053 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1057 if (hpage_size >= 1024) {
1061 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1063 if (spool->max_hpages != -1)
1064 seq_printf(m, ",size=%llu",
1065 (unsigned long long)spool->max_hpages << hpage_shift);
1066 if (spool->min_hpages != -1)
1067 seq_printf(m, ",min_size=%llu",
1068 (unsigned long long)spool->min_hpages << hpage_shift);
1073 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1075 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1076 struct hstate *h = hstate_inode(d_inode(dentry));
1078 buf->f_type = HUGETLBFS_MAGIC;
1079 buf->f_bsize = huge_page_size(h);
1081 spin_lock(&sbinfo->stat_lock);
1082 /* If no limits set, just report 0 or -1 for max/free/used
1083 * blocks, like simple_statfs() */
1084 if (sbinfo->spool) {
1087 spin_lock_irq(&sbinfo->spool->lock);
1088 buf->f_blocks = sbinfo->spool->max_hpages;
1089 free_pages = sbinfo->spool->max_hpages
1090 - sbinfo->spool->used_hpages;
1091 buf->f_bavail = buf->f_bfree = free_pages;
1092 spin_unlock_irq(&sbinfo->spool->lock);
1093 buf->f_files = sbinfo->max_inodes;
1094 buf->f_ffree = sbinfo->free_inodes;
1096 spin_unlock(&sbinfo->stat_lock);
1098 buf->f_namelen = NAME_MAX;
1102 static void hugetlbfs_put_super(struct super_block *sb)
1104 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1107 sb->s_fs_info = NULL;
1110 hugepage_put_subpool(sbi->spool);
1116 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1118 if (sbinfo->free_inodes >= 0) {
1119 spin_lock(&sbinfo->stat_lock);
1120 if (unlikely(!sbinfo->free_inodes)) {
1121 spin_unlock(&sbinfo->stat_lock);
1124 sbinfo->free_inodes--;
1125 spin_unlock(&sbinfo->stat_lock);
1131 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1133 if (sbinfo->free_inodes >= 0) {
1134 spin_lock(&sbinfo->stat_lock);
1135 sbinfo->free_inodes++;
1136 spin_unlock(&sbinfo->stat_lock);
1141 static struct kmem_cache *hugetlbfs_inode_cachep;
1143 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1145 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1146 struct hugetlbfs_inode_info *p;
1148 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1150 p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1152 hugetlbfs_inc_free_inodes(sbinfo);
1157 * Any time after allocation, hugetlbfs_destroy_inode can be called
1158 * for the inode. mpol_free_shared_policy is unconditionally called
1159 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1160 * in case of a quick call to destroy.
1162 * Note that the policy is initialized even if we are creating a
1163 * private inode. This simplifies hugetlbfs_destroy_inode.
1165 mpol_shared_policy_init(&p->policy, NULL);
1167 return &p->vfs_inode;
1170 static void hugetlbfs_free_inode(struct inode *inode)
1172 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1175 static void hugetlbfs_destroy_inode(struct inode *inode)
1177 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1178 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1181 static const struct address_space_operations hugetlbfs_aops = {
1182 .write_begin = hugetlbfs_write_begin,
1183 .write_end = hugetlbfs_write_end,
1184 .dirty_folio = noop_dirty_folio,
1185 .migrate_folio = hugetlbfs_migrate_folio,
1186 .error_remove_page = hugetlbfs_error_remove_page,
1190 static void init_once(void *foo)
1192 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1194 inode_init_once(&ei->vfs_inode);
1197 const struct file_operations hugetlbfs_file_operations = {
1198 .read_iter = hugetlbfs_read_iter,
1199 .mmap = hugetlbfs_file_mmap,
1200 .fsync = noop_fsync,
1201 .get_unmapped_area = hugetlb_get_unmapped_area,
1202 .llseek = default_llseek,
1203 .fallocate = hugetlbfs_fallocate,
1206 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1207 .create = hugetlbfs_create,
1208 .lookup = simple_lookup,
1209 .link = simple_link,
1210 .unlink = simple_unlink,
1211 .symlink = hugetlbfs_symlink,
1212 .mkdir = hugetlbfs_mkdir,
1213 .rmdir = simple_rmdir,
1214 .mknod = hugetlbfs_mknod,
1215 .rename = simple_rename,
1216 .setattr = hugetlbfs_setattr,
1217 .tmpfile = hugetlbfs_tmpfile,
1220 static const struct inode_operations hugetlbfs_inode_operations = {
1221 .setattr = hugetlbfs_setattr,
1224 static const struct super_operations hugetlbfs_ops = {
1225 .alloc_inode = hugetlbfs_alloc_inode,
1226 .free_inode = hugetlbfs_free_inode,
1227 .destroy_inode = hugetlbfs_destroy_inode,
1228 .evict_inode = hugetlbfs_evict_inode,
1229 .statfs = hugetlbfs_statfs,
1230 .put_super = hugetlbfs_put_super,
1231 .show_options = hugetlbfs_show_options,
1235 * Convert size option passed from command line to number of huge pages
1236 * in the pool specified by hstate. Size option could be in bytes
1237 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1240 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1241 enum hugetlbfs_size_type val_type)
1243 if (val_type == NO_SIZE)
1246 if (val_type == SIZE_PERCENT) {
1247 size_opt <<= huge_page_shift(h);
1248 size_opt *= h->max_huge_pages;
1249 do_div(size_opt, 100);
1252 size_opt >>= huge_page_shift(h);
1257 * Parse one mount parameter.
1259 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1261 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1262 struct fs_parse_result result;
1267 opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1273 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1274 if (!uid_valid(ctx->uid))
1279 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1280 if (!gid_valid(ctx->gid))
1285 ctx->mode = result.uint_32 & 01777U;
1289 /* memparse() will accept a K/M/G without a digit */
1290 if (!isdigit(param->string[0]))
1292 ctx->max_size_opt = memparse(param->string, &rest);
1293 ctx->max_val_type = SIZE_STD;
1295 ctx->max_val_type = SIZE_PERCENT;
1299 /* memparse() will accept a K/M/G without a digit */
1300 if (!isdigit(param->string[0]))
1302 ctx->nr_inodes = memparse(param->string, &rest);
1306 ps = memparse(param->string, &rest);
1307 ctx->hstate = size_to_hstate(ps);
1309 pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1315 /* memparse() will accept a K/M/G without a digit */
1316 if (!isdigit(param->string[0]))
1318 ctx->min_size_opt = memparse(param->string, &rest);
1319 ctx->min_val_type = SIZE_STD;
1321 ctx->min_val_type = SIZE_PERCENT;
1329 return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1330 param->string, param->key);
1334 * Validate the parsed options.
1336 static int hugetlbfs_validate(struct fs_context *fc)
1338 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1341 * Use huge page pool size (in hstate) to convert the size
1342 * options to number of huge pages. If NO_SIZE, -1 is returned.
1344 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1347 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1352 * If max_size was specified, then min_size must be smaller
1354 if (ctx->max_val_type > NO_SIZE &&
1355 ctx->min_hpages > ctx->max_hpages) {
1356 pr_err("Minimum size can not be greater than maximum size\n");
1364 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1366 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1367 struct hugetlbfs_sb_info *sbinfo;
1369 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1372 sb->s_fs_info = sbinfo;
1373 spin_lock_init(&sbinfo->stat_lock);
1374 sbinfo->hstate = ctx->hstate;
1375 sbinfo->max_inodes = ctx->nr_inodes;
1376 sbinfo->free_inodes = ctx->nr_inodes;
1377 sbinfo->spool = NULL;
1378 sbinfo->uid = ctx->uid;
1379 sbinfo->gid = ctx->gid;
1380 sbinfo->mode = ctx->mode;
1383 * Allocate and initialize subpool if maximum or minimum size is
1384 * specified. Any needed reservations (for minimum size) are taken
1385 * when the subpool is created.
1387 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1388 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1394 sb->s_maxbytes = MAX_LFS_FILESIZE;
1395 sb->s_blocksize = huge_page_size(ctx->hstate);
1396 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1397 sb->s_magic = HUGETLBFS_MAGIC;
1398 sb->s_op = &hugetlbfs_ops;
1399 sb->s_time_gran = 1;
1402 * Due to the special and limited functionality of hugetlbfs, it does
1403 * not work well as a stacking filesystem.
1405 sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1406 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1411 kfree(sbinfo->spool);
1416 static int hugetlbfs_get_tree(struct fs_context *fc)
1418 int err = hugetlbfs_validate(fc);
1421 return get_tree_nodev(fc, hugetlbfs_fill_super);
1424 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1426 kfree(fc->fs_private);
1429 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1430 .free = hugetlbfs_fs_context_free,
1431 .parse_param = hugetlbfs_parse_param,
1432 .get_tree = hugetlbfs_get_tree,
1435 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1437 struct hugetlbfs_fs_context *ctx;
1439 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1443 ctx->max_hpages = -1; /* No limit on size by default */
1444 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1445 ctx->uid = current_fsuid();
1446 ctx->gid = current_fsgid();
1448 ctx->hstate = &default_hstate;
1449 ctx->min_hpages = -1; /* No default minimum size */
1450 ctx->max_val_type = NO_SIZE;
1451 ctx->min_val_type = NO_SIZE;
1452 fc->fs_private = ctx;
1453 fc->ops = &hugetlbfs_fs_context_ops;
1457 static struct file_system_type hugetlbfs_fs_type = {
1458 .name = "hugetlbfs",
1459 .init_fs_context = hugetlbfs_init_fs_context,
1460 .parameters = hugetlb_fs_parameters,
1461 .kill_sb = kill_litter_super,
1464 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1466 static int can_do_hugetlb_shm(void)
1469 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1470 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1473 static int get_hstate_idx(int page_size_log)
1475 struct hstate *h = hstate_sizelog(page_size_log);
1479 return hstate_index(h);
1483 * Note that size should be aligned to proper hugepage size in caller side,
1484 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1486 struct file *hugetlb_file_setup(const char *name, size_t size,
1487 vm_flags_t acctflag, int creat_flags,
1490 struct inode *inode;
1491 struct vfsmount *mnt;
1495 hstate_idx = get_hstate_idx(page_size_log);
1497 return ERR_PTR(-ENODEV);
1499 mnt = hugetlbfs_vfsmount[hstate_idx];
1501 return ERR_PTR(-ENOENT);
1503 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1504 struct ucounts *ucounts = current_ucounts();
1506 if (user_shm_lock(size, ucounts)) {
1507 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1508 current->comm, current->pid);
1509 user_shm_unlock(size, ucounts);
1511 return ERR_PTR(-EPERM);
1514 file = ERR_PTR(-ENOSPC);
1515 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1518 if (creat_flags == HUGETLB_SHMFS_INODE)
1519 inode->i_flags |= S_PRIVATE;
1521 inode->i_size = size;
1524 if (!hugetlb_reserve_pages(inode, 0,
1525 size >> huge_page_shift(hstate_inode(inode)), NULL,
1527 file = ERR_PTR(-ENOMEM);
1529 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1530 &hugetlbfs_file_operations);
1539 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1541 struct fs_context *fc;
1542 struct vfsmount *mnt;
1544 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1548 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1554 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1555 huge_page_size(h) / SZ_1K);
1559 static int __init init_hugetlbfs_fs(void)
1561 struct vfsmount *mnt;
1566 if (!hugepages_supported()) {
1567 pr_info("disabling because there are no supported hugepage sizes\n");
1572 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1573 sizeof(struct hugetlbfs_inode_info),
1574 0, SLAB_ACCOUNT, init_once);
1575 if (hugetlbfs_inode_cachep == NULL)
1578 error = register_filesystem(&hugetlbfs_fs_type);
1582 /* default hstate mount is required */
1583 mnt = mount_one_hugetlbfs(&default_hstate);
1585 error = PTR_ERR(mnt);
1588 hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1590 /* other hstates are optional */
1592 for_each_hstate(h) {
1593 if (i == default_hstate_idx) {
1598 mnt = mount_one_hugetlbfs(h);
1600 hugetlbfs_vfsmount[i] = NULL;
1602 hugetlbfs_vfsmount[i] = mnt;
1609 (void)unregister_filesystem(&hugetlbfs_fs_type);
1611 kmem_cache_destroy(hugetlbfs_inode_cachep);
1615 fs_initcall(init_hugetlbfs_fs)