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 * Support for read() - Find the page attached to f_mapping and copy out the
287 * data. This provides functionality similar to filemap_read().
289 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
291 struct file *file = iocb->ki_filp;
292 struct hstate *h = hstate_file(file);
293 struct address_space *mapping = file->f_mapping;
294 struct inode *inode = mapping->host;
295 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
296 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
297 unsigned long end_index;
301 while (iov_iter_count(to)) {
305 /* nr is the maximum number of bytes to copy from this page */
306 nr = huge_page_size(h);
307 isize = i_size_read(inode);
310 end_index = (isize - 1) >> huge_page_shift(h);
311 if (index > end_index)
313 if (index == end_index) {
314 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
321 page = find_lock_page(mapping, index);
322 if (unlikely(page == NULL)) {
324 * We have a HOLE, zero out the user-buffer for the
325 * length of the hole or request.
327 copied = iov_iter_zero(nr, to);
332 * We have the page, copy it to user space buffer.
334 copied = copy_page_to_iter(page, offset, nr, to);
339 if (copied != nr && iov_iter_count(to)) {
344 index += offset >> huge_page_shift(h);
345 offset &= ~huge_page_mask(h);
347 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
351 static int hugetlbfs_write_begin(struct file *file,
352 struct address_space *mapping,
353 loff_t pos, unsigned len,
354 struct page **pagep, void **fsdata)
359 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
360 loff_t pos, unsigned len, unsigned copied,
361 struct page *page, void *fsdata)
367 static void remove_huge_page(struct page *page)
369 ClearPageDirty(page);
370 ClearPageUptodate(page);
371 delete_from_page_cache(page);
375 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
376 zap_flags_t zap_flags)
378 struct vm_area_struct *vma;
381 * end == 0 indicates that the entire range after start should be
382 * unmapped. Note, end is exclusive, whereas the interval tree takes
383 * an inclusive "last".
385 vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
386 unsigned long v_offset;
390 * Can the expression below overflow on 32-bit arches?
391 * No, because the interval tree returns us only those vmas
392 * which overlap the truncated area starting at pgoff,
393 * and no vma on a 32-bit arch can span beyond the 4GB.
395 if (vma->vm_pgoff < start)
396 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
403 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
405 if (v_end > vma->vm_end)
409 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
415 * remove_inode_hugepages handles two distinct cases: truncation and hole
416 * punch. There are subtle differences in operation for each case.
418 * truncation is indicated by end of range being LLONG_MAX
419 * In this case, we first scan the range and release found pages.
420 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
421 * maps and global counts. Page faults can not race with truncation
422 * in this routine. hugetlb_no_page() holds i_mmap_rwsem and prevents
423 * page faults in the truncated range by checking i_size. i_size is
424 * modified while holding i_mmap_rwsem.
425 * hole punch is indicated if end is not LLONG_MAX
426 * In the hole punch case we scan the range and release found pages.
427 * Only when releasing a page is the associated region/reserve map
428 * deleted. The region/reserve map for ranges without associated
429 * pages are not modified. Page faults can race with hole punch.
430 * This is indicated if we find a mapped page.
431 * Note: If the passed end of range value is beyond the end of file, but
432 * not LLONG_MAX this routine still performs a hole punch operation.
434 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
437 struct hstate *h = hstate_inode(inode);
438 struct address_space *mapping = &inode->i_data;
439 const pgoff_t start = lstart >> huge_page_shift(h);
440 const pgoff_t end = lend >> huge_page_shift(h);
441 struct folio_batch fbatch;
444 bool truncate_op = (lend == LLONG_MAX);
446 folio_batch_init(&fbatch);
448 while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
449 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
450 struct folio *folio = fbatch.folios[i];
453 index = folio->index;
456 * Only need to hold the fault mutex in the
457 * hole punch case. This prevents races with
458 * page faults. Races are not possible in the
459 * case of truncation.
461 hash = hugetlb_fault_mutex_hash(mapping, index);
462 mutex_lock(&hugetlb_fault_mutex_table[hash]);
466 * If folio is mapped, it was faulted in after being
467 * unmapped in caller. Unmap (again) now after taking
468 * the fault mutex. The mutex will prevent faults
469 * until we finish removing the folio.
471 * This race can only happen in the hole punch case.
472 * Getting here in a truncate operation is a bug.
474 if (unlikely(folio_mapped(folio))) {
477 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
478 i_mmap_lock_write(mapping);
479 mutex_lock(&hugetlb_fault_mutex_table[hash]);
480 hugetlb_vmdelete_list(&mapping->i_mmap,
481 index * pages_per_huge_page(h),
482 (index + 1) * pages_per_huge_page(h),
483 ZAP_FLAG_DROP_MARKER);
484 i_mmap_unlock_write(mapping);
489 * We must free the huge page and remove from page
490 * cache (remove_huge_page) BEFORE removing the
491 * region/reserve map (hugetlb_unreserve_pages). In
492 * rare out of memory conditions, removal of the
493 * region/reserve map could fail. Correspondingly,
494 * the subpool and global reserve usage count can need
497 VM_BUG_ON(HPageRestoreReserve(&folio->page));
498 remove_huge_page(&folio->page);
501 if (unlikely(hugetlb_unreserve_pages(inode,
502 index, index + 1, 1)))
503 hugetlb_fix_reserve_counts(inode);
508 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
510 folio_batch_release(&fbatch);
515 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
518 static void hugetlbfs_evict_inode(struct inode *inode)
520 struct resv_map *resv_map;
522 remove_inode_hugepages(inode, 0, LLONG_MAX);
525 * Get the resv_map from the address space embedded in the inode.
526 * This is the address space which points to any resv_map allocated
527 * at inode creation time. If this is a device special inode,
528 * i_mapping may not point to the original address space.
530 resv_map = (struct resv_map *)(&inode->i_data)->private_data;
531 /* Only regular and link inodes have associated reserve maps */
533 resv_map_release(&resv_map->refs);
537 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
540 struct address_space *mapping = inode->i_mapping;
541 struct hstate *h = hstate_inode(inode);
543 BUG_ON(offset & ~huge_page_mask(h));
544 pgoff = offset >> PAGE_SHIFT;
546 i_mmap_lock_write(mapping);
547 i_size_write(inode, offset);
548 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
549 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
550 ZAP_FLAG_DROP_MARKER);
551 i_mmap_unlock_write(mapping);
552 remove_inode_hugepages(inode, offset, LLONG_MAX);
555 static void hugetlbfs_zero_partial_page(struct hstate *h,
556 struct address_space *mapping,
560 pgoff_t idx = start >> huge_page_shift(h);
563 folio = filemap_lock_folio(mapping, idx);
567 start = start & ~huge_page_mask(h);
568 end = end & ~huge_page_mask(h);
570 end = huge_page_size(h);
572 folio_zero_segment(folio, (size_t)start, (size_t)end);
578 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
580 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
581 struct address_space *mapping = inode->i_mapping;
582 struct hstate *h = hstate_inode(inode);
583 loff_t hpage_size = huge_page_size(h);
584 loff_t hole_start, hole_end;
587 * hole_start and hole_end indicate the full pages within the hole.
589 hole_start = round_up(offset, hpage_size);
590 hole_end = round_down(offset + len, hpage_size);
594 /* protected by i_rwsem */
595 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
600 i_mmap_lock_write(mapping);
602 /* If range starts before first full page, zero partial page. */
603 if (offset < hole_start)
604 hugetlbfs_zero_partial_page(h, mapping,
605 offset, min(offset + len, hole_start));
607 /* Unmap users of full pages in the hole. */
608 if (hole_end > hole_start) {
609 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
610 hugetlb_vmdelete_list(&mapping->i_mmap,
611 hole_start >> PAGE_SHIFT,
612 hole_end >> PAGE_SHIFT, 0);
615 /* If range extends beyond last full page, zero partial page. */
616 if ((offset + len) > hole_end && (offset + len) > hole_start)
617 hugetlbfs_zero_partial_page(h, mapping,
618 hole_end, offset + len);
620 i_mmap_unlock_write(mapping);
622 /* Remove full pages from the file. */
623 if (hole_end > hole_start)
624 remove_inode_hugepages(inode, hole_start, hole_end);
631 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
634 struct inode *inode = file_inode(file);
635 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
636 struct address_space *mapping = inode->i_mapping;
637 struct hstate *h = hstate_inode(inode);
638 struct vm_area_struct pseudo_vma;
639 struct mm_struct *mm = current->mm;
640 loff_t hpage_size = huge_page_size(h);
641 unsigned long hpage_shift = huge_page_shift(h);
642 pgoff_t start, index, end;
646 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
649 if (mode & FALLOC_FL_PUNCH_HOLE)
650 return hugetlbfs_punch_hole(inode, offset, len);
653 * Default preallocate case.
654 * For this range, start is rounded down and end is rounded up
655 * as well as being converted to page offsets.
657 start = offset >> hpage_shift;
658 end = (offset + len + hpage_size - 1) >> hpage_shift;
662 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
663 error = inode_newsize_ok(inode, offset + len);
667 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
673 * Initialize a pseudo vma as this is required by the huge page
674 * allocation routines. If NUMA is configured, use page index
675 * as input to create an allocation policy.
677 vma_init(&pseudo_vma, mm);
678 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
679 pseudo_vma.vm_file = file;
681 for (index = start; index < end; index++) {
683 * This is supposed to be the vaddr where the page is being
684 * faulted in, but we have no vaddr here.
692 * fallocate(2) manpage permits EINTR; we may have been
693 * interrupted because we are using up too much memory.
695 if (signal_pending(current)) {
700 /* Set numa allocation policy based on index */
701 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
703 /* addr is the offset within the file (zero based) */
704 addr = index * hpage_size;
707 * fault mutex taken here, protects against fault path
708 * and hole punch. inode_lock previously taken protects
709 * against truncation.
711 hash = hugetlb_fault_mutex_hash(mapping, index);
712 mutex_lock(&hugetlb_fault_mutex_table[hash]);
714 /* See if already present in mapping to avoid alloc/free */
715 page = find_get_page(mapping, index);
718 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
719 hugetlb_drop_vma_policy(&pseudo_vma);
724 * Allocate page without setting the avoid_reserve argument.
725 * There certainly are no reserves associated with the
726 * pseudo_vma. However, there could be shared mappings with
727 * reserves for the file at the inode level. If we fallocate
728 * pages in these areas, we need to consume the reserves
729 * to keep reservation accounting consistent.
731 page = alloc_huge_page(&pseudo_vma, addr, 0);
732 hugetlb_drop_vma_policy(&pseudo_vma);
734 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
735 error = PTR_ERR(page);
738 clear_huge_page(page, addr, pages_per_huge_page(h));
739 __SetPageUptodate(page);
740 error = huge_add_to_page_cache(page, mapping, index);
741 if (unlikely(error)) {
742 restore_reserve_on_error(h, &pseudo_vma, addr, page);
744 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
748 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
750 SetHPageMigratable(page);
752 * unlock_page because locked by huge_add_to_page_cache()
753 * put_page() due to reference from alloc_huge_page()
759 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
760 i_size_write(inode, offset + len);
761 inode->i_ctime = current_time(inode);
767 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
768 struct dentry *dentry, struct iattr *attr)
770 struct inode *inode = d_inode(dentry);
771 struct hstate *h = hstate_inode(inode);
773 unsigned int ia_valid = attr->ia_valid;
774 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
776 error = setattr_prepare(&init_user_ns, dentry, attr);
780 if (ia_valid & ATTR_SIZE) {
781 loff_t oldsize = inode->i_size;
782 loff_t newsize = attr->ia_size;
784 if (newsize & ~huge_page_mask(h))
786 /* protected by i_rwsem */
787 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
788 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
790 hugetlb_vmtruncate(inode, newsize);
793 setattr_copy(&init_user_ns, inode, attr);
794 mark_inode_dirty(inode);
798 static struct inode *hugetlbfs_get_root(struct super_block *sb,
799 struct hugetlbfs_fs_context *ctx)
803 inode = new_inode(sb);
805 inode->i_ino = get_next_ino();
806 inode->i_mode = S_IFDIR | ctx->mode;
807 inode->i_uid = ctx->uid;
808 inode->i_gid = ctx->gid;
809 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
810 inode->i_op = &hugetlbfs_dir_inode_operations;
811 inode->i_fop = &simple_dir_operations;
812 /* directory inodes start off with i_nlink == 2 (for "." entry) */
814 lockdep_annotate_inode_mutex_key(inode);
820 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
821 * be taken from reclaim -- unlike regular filesystems. This needs an
822 * annotation because huge_pmd_share() does an allocation under hugetlb's
825 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
827 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
829 umode_t mode, dev_t dev)
832 struct resv_map *resv_map = NULL;
835 * Reserve maps are only needed for inodes that can have associated
838 if (S_ISREG(mode) || S_ISLNK(mode)) {
839 resv_map = resv_map_alloc();
844 inode = new_inode(sb);
846 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
848 inode->i_ino = get_next_ino();
849 inode_init_owner(&init_user_ns, inode, dir, mode);
850 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
851 &hugetlbfs_i_mmap_rwsem_key);
852 inode->i_mapping->a_ops = &hugetlbfs_aops;
853 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
854 inode->i_mapping->private_data = resv_map;
855 info->seals = F_SEAL_SEAL;
856 switch (mode & S_IFMT) {
858 init_special_inode(inode, mode, dev);
861 inode->i_op = &hugetlbfs_inode_operations;
862 inode->i_fop = &hugetlbfs_file_operations;
865 inode->i_op = &hugetlbfs_dir_inode_operations;
866 inode->i_fop = &simple_dir_operations;
868 /* directory inodes start off with i_nlink == 2 (for "." entry) */
872 inode->i_op = &page_symlink_inode_operations;
873 inode_nohighmem(inode);
876 lockdep_annotate_inode_mutex_key(inode);
879 kref_put(&resv_map->refs, resv_map_release);
886 * File creation. Allocate an inode, and we're done..
888 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
889 struct dentry *dentry, umode_t mode, dev_t dev)
893 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
896 dir->i_ctime = dir->i_mtime = current_time(dir);
897 d_instantiate(dentry, inode);
898 dget(dentry);/* Extra count - pin the dentry in core */
902 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
903 struct dentry *dentry, umode_t mode)
905 int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
912 static int hugetlbfs_create(struct user_namespace *mnt_userns,
913 struct inode *dir, struct dentry *dentry,
914 umode_t mode, bool excl)
916 return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
919 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
920 struct inode *dir, struct dentry *dentry,
925 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode | S_IFREG, 0);
928 dir->i_ctime = dir->i_mtime = current_time(dir);
929 d_tmpfile(dentry, inode);
933 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
934 struct inode *dir, struct dentry *dentry,
940 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
942 int l = strlen(symname)+1;
943 error = page_symlink(inode, symname, l);
945 d_instantiate(dentry, inode);
950 dir->i_ctime = dir->i_mtime = current_time(dir);
955 #ifdef CONFIG_MIGRATION
956 static int hugetlbfs_migrate_folio(struct address_space *mapping,
957 struct folio *dst, struct folio *src,
958 enum migrate_mode mode)
962 rc = migrate_huge_page_move_mapping(mapping, dst, src);
963 if (rc != MIGRATEPAGE_SUCCESS)
966 if (hugetlb_page_subpool(&src->page)) {
967 hugetlb_set_page_subpool(&dst->page,
968 hugetlb_page_subpool(&src->page));
969 hugetlb_set_page_subpool(&src->page, NULL);
972 if (mode != MIGRATE_SYNC_NO_COPY)
973 folio_migrate_copy(dst, src);
975 folio_migrate_flags(dst, src);
977 return MIGRATEPAGE_SUCCESS;
980 #define hugetlbfs_migrate_folio NULL
983 static int hugetlbfs_error_remove_page(struct address_space *mapping,
986 struct inode *inode = mapping->host;
987 pgoff_t index = page->index;
989 remove_huge_page(page);
990 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
991 hugetlb_fix_reserve_counts(inode);
997 * Display the mount options in /proc/mounts.
999 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1001 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1002 struct hugepage_subpool *spool = sbinfo->spool;
1003 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1004 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1007 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1008 seq_printf(m, ",uid=%u",
1009 from_kuid_munged(&init_user_ns, sbinfo->uid));
1010 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1011 seq_printf(m, ",gid=%u",
1012 from_kgid_munged(&init_user_ns, sbinfo->gid));
1013 if (sbinfo->mode != 0755)
1014 seq_printf(m, ",mode=%o", sbinfo->mode);
1015 if (sbinfo->max_inodes != -1)
1016 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1020 if (hpage_size >= 1024) {
1024 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1026 if (spool->max_hpages != -1)
1027 seq_printf(m, ",size=%llu",
1028 (unsigned long long)spool->max_hpages << hpage_shift);
1029 if (spool->min_hpages != -1)
1030 seq_printf(m, ",min_size=%llu",
1031 (unsigned long long)spool->min_hpages << hpage_shift);
1036 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1038 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1039 struct hstate *h = hstate_inode(d_inode(dentry));
1041 buf->f_type = HUGETLBFS_MAGIC;
1042 buf->f_bsize = huge_page_size(h);
1044 spin_lock(&sbinfo->stat_lock);
1045 /* If no limits set, just report 0 or -1 for max/free/used
1046 * blocks, like simple_statfs() */
1047 if (sbinfo->spool) {
1050 spin_lock_irq(&sbinfo->spool->lock);
1051 buf->f_blocks = sbinfo->spool->max_hpages;
1052 free_pages = sbinfo->spool->max_hpages
1053 - sbinfo->spool->used_hpages;
1054 buf->f_bavail = buf->f_bfree = free_pages;
1055 spin_unlock_irq(&sbinfo->spool->lock);
1056 buf->f_files = sbinfo->max_inodes;
1057 buf->f_ffree = sbinfo->free_inodes;
1059 spin_unlock(&sbinfo->stat_lock);
1061 buf->f_namelen = NAME_MAX;
1065 static void hugetlbfs_put_super(struct super_block *sb)
1067 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1070 sb->s_fs_info = NULL;
1073 hugepage_put_subpool(sbi->spool);
1079 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1081 if (sbinfo->free_inodes >= 0) {
1082 spin_lock(&sbinfo->stat_lock);
1083 if (unlikely(!sbinfo->free_inodes)) {
1084 spin_unlock(&sbinfo->stat_lock);
1087 sbinfo->free_inodes--;
1088 spin_unlock(&sbinfo->stat_lock);
1094 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1096 if (sbinfo->free_inodes >= 0) {
1097 spin_lock(&sbinfo->stat_lock);
1098 sbinfo->free_inodes++;
1099 spin_unlock(&sbinfo->stat_lock);
1104 static struct kmem_cache *hugetlbfs_inode_cachep;
1106 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1108 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1109 struct hugetlbfs_inode_info *p;
1111 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1113 p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1115 hugetlbfs_inc_free_inodes(sbinfo);
1120 * Any time after allocation, hugetlbfs_destroy_inode can be called
1121 * for the inode. mpol_free_shared_policy is unconditionally called
1122 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1123 * in case of a quick call to destroy.
1125 * Note that the policy is initialized even if we are creating a
1126 * private inode. This simplifies hugetlbfs_destroy_inode.
1128 mpol_shared_policy_init(&p->policy, NULL);
1130 return &p->vfs_inode;
1133 static void hugetlbfs_free_inode(struct inode *inode)
1135 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1138 static void hugetlbfs_destroy_inode(struct inode *inode)
1140 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1141 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1144 static const struct address_space_operations hugetlbfs_aops = {
1145 .write_begin = hugetlbfs_write_begin,
1146 .write_end = hugetlbfs_write_end,
1147 .dirty_folio = noop_dirty_folio,
1148 .migrate_folio = hugetlbfs_migrate_folio,
1149 .error_remove_page = hugetlbfs_error_remove_page,
1153 static void init_once(void *foo)
1155 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1157 inode_init_once(&ei->vfs_inode);
1160 const struct file_operations hugetlbfs_file_operations = {
1161 .read_iter = hugetlbfs_read_iter,
1162 .mmap = hugetlbfs_file_mmap,
1163 .fsync = noop_fsync,
1164 .get_unmapped_area = hugetlb_get_unmapped_area,
1165 .llseek = default_llseek,
1166 .fallocate = hugetlbfs_fallocate,
1169 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1170 .create = hugetlbfs_create,
1171 .lookup = simple_lookup,
1172 .link = simple_link,
1173 .unlink = simple_unlink,
1174 .symlink = hugetlbfs_symlink,
1175 .mkdir = hugetlbfs_mkdir,
1176 .rmdir = simple_rmdir,
1177 .mknod = hugetlbfs_mknod,
1178 .rename = simple_rename,
1179 .setattr = hugetlbfs_setattr,
1180 .tmpfile = hugetlbfs_tmpfile,
1183 static const struct inode_operations hugetlbfs_inode_operations = {
1184 .setattr = hugetlbfs_setattr,
1187 static const struct super_operations hugetlbfs_ops = {
1188 .alloc_inode = hugetlbfs_alloc_inode,
1189 .free_inode = hugetlbfs_free_inode,
1190 .destroy_inode = hugetlbfs_destroy_inode,
1191 .evict_inode = hugetlbfs_evict_inode,
1192 .statfs = hugetlbfs_statfs,
1193 .put_super = hugetlbfs_put_super,
1194 .show_options = hugetlbfs_show_options,
1198 * Convert size option passed from command line to number of huge pages
1199 * in the pool specified by hstate. Size option could be in bytes
1200 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1203 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1204 enum hugetlbfs_size_type val_type)
1206 if (val_type == NO_SIZE)
1209 if (val_type == SIZE_PERCENT) {
1210 size_opt <<= huge_page_shift(h);
1211 size_opt *= h->max_huge_pages;
1212 do_div(size_opt, 100);
1215 size_opt >>= huge_page_shift(h);
1220 * Parse one mount parameter.
1222 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1224 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1225 struct fs_parse_result result;
1230 opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1236 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1237 if (!uid_valid(ctx->uid))
1242 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1243 if (!gid_valid(ctx->gid))
1248 ctx->mode = result.uint_32 & 01777U;
1252 /* memparse() will accept a K/M/G without a digit */
1253 if (!isdigit(param->string[0]))
1255 ctx->max_size_opt = memparse(param->string, &rest);
1256 ctx->max_val_type = SIZE_STD;
1258 ctx->max_val_type = SIZE_PERCENT;
1262 /* memparse() will accept a K/M/G without a digit */
1263 if (!isdigit(param->string[0]))
1265 ctx->nr_inodes = memparse(param->string, &rest);
1269 ps = memparse(param->string, &rest);
1270 ctx->hstate = size_to_hstate(ps);
1272 pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1278 /* memparse() will accept a K/M/G without a digit */
1279 if (!isdigit(param->string[0]))
1281 ctx->min_size_opt = memparse(param->string, &rest);
1282 ctx->min_val_type = SIZE_STD;
1284 ctx->min_val_type = SIZE_PERCENT;
1292 return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1293 param->string, param->key);
1297 * Validate the parsed options.
1299 static int hugetlbfs_validate(struct fs_context *fc)
1301 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1304 * Use huge page pool size (in hstate) to convert the size
1305 * options to number of huge pages. If NO_SIZE, -1 is returned.
1307 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1310 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1315 * If max_size was specified, then min_size must be smaller
1317 if (ctx->max_val_type > NO_SIZE &&
1318 ctx->min_hpages > ctx->max_hpages) {
1319 pr_err("Minimum size can not be greater than maximum size\n");
1327 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1329 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1330 struct hugetlbfs_sb_info *sbinfo;
1332 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1335 sb->s_fs_info = sbinfo;
1336 spin_lock_init(&sbinfo->stat_lock);
1337 sbinfo->hstate = ctx->hstate;
1338 sbinfo->max_inodes = ctx->nr_inodes;
1339 sbinfo->free_inodes = ctx->nr_inodes;
1340 sbinfo->spool = NULL;
1341 sbinfo->uid = ctx->uid;
1342 sbinfo->gid = ctx->gid;
1343 sbinfo->mode = ctx->mode;
1346 * Allocate and initialize subpool if maximum or minimum size is
1347 * specified. Any needed reservations (for minimum size) are taken
1348 * when the subpool is created.
1350 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1351 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1357 sb->s_maxbytes = MAX_LFS_FILESIZE;
1358 sb->s_blocksize = huge_page_size(ctx->hstate);
1359 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1360 sb->s_magic = HUGETLBFS_MAGIC;
1361 sb->s_op = &hugetlbfs_ops;
1362 sb->s_time_gran = 1;
1365 * Due to the special and limited functionality of hugetlbfs, it does
1366 * not work well as a stacking filesystem.
1368 sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1369 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1374 kfree(sbinfo->spool);
1379 static int hugetlbfs_get_tree(struct fs_context *fc)
1381 int err = hugetlbfs_validate(fc);
1384 return get_tree_nodev(fc, hugetlbfs_fill_super);
1387 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1389 kfree(fc->fs_private);
1392 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1393 .free = hugetlbfs_fs_context_free,
1394 .parse_param = hugetlbfs_parse_param,
1395 .get_tree = hugetlbfs_get_tree,
1398 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1400 struct hugetlbfs_fs_context *ctx;
1402 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1406 ctx->max_hpages = -1; /* No limit on size by default */
1407 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1408 ctx->uid = current_fsuid();
1409 ctx->gid = current_fsgid();
1411 ctx->hstate = &default_hstate;
1412 ctx->min_hpages = -1; /* No default minimum size */
1413 ctx->max_val_type = NO_SIZE;
1414 ctx->min_val_type = NO_SIZE;
1415 fc->fs_private = ctx;
1416 fc->ops = &hugetlbfs_fs_context_ops;
1420 static struct file_system_type hugetlbfs_fs_type = {
1421 .name = "hugetlbfs",
1422 .init_fs_context = hugetlbfs_init_fs_context,
1423 .parameters = hugetlb_fs_parameters,
1424 .kill_sb = kill_litter_super,
1427 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1429 static int can_do_hugetlb_shm(void)
1432 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1433 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1436 static int get_hstate_idx(int page_size_log)
1438 struct hstate *h = hstate_sizelog(page_size_log);
1442 return hstate_index(h);
1446 * Note that size should be aligned to proper hugepage size in caller side,
1447 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1449 struct file *hugetlb_file_setup(const char *name, size_t size,
1450 vm_flags_t acctflag, int creat_flags,
1453 struct inode *inode;
1454 struct vfsmount *mnt;
1458 hstate_idx = get_hstate_idx(page_size_log);
1460 return ERR_PTR(-ENODEV);
1462 mnt = hugetlbfs_vfsmount[hstate_idx];
1464 return ERR_PTR(-ENOENT);
1466 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1467 struct ucounts *ucounts = current_ucounts();
1469 if (user_shm_lock(size, ucounts)) {
1470 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1471 current->comm, current->pid);
1472 user_shm_unlock(size, ucounts);
1474 return ERR_PTR(-EPERM);
1477 file = ERR_PTR(-ENOSPC);
1478 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1481 if (creat_flags == HUGETLB_SHMFS_INODE)
1482 inode->i_flags |= S_PRIVATE;
1484 inode->i_size = size;
1487 if (!hugetlb_reserve_pages(inode, 0,
1488 size >> huge_page_shift(hstate_inode(inode)), NULL,
1490 file = ERR_PTR(-ENOMEM);
1492 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1493 &hugetlbfs_file_operations);
1502 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1504 struct fs_context *fc;
1505 struct vfsmount *mnt;
1507 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1511 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1517 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1518 huge_page_size(h) / SZ_1K);
1522 static int __init init_hugetlbfs_fs(void)
1524 struct vfsmount *mnt;
1529 if (!hugepages_supported()) {
1530 pr_info("disabling because there are no supported hugepage sizes\n");
1535 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1536 sizeof(struct hugetlbfs_inode_info),
1537 0, SLAB_ACCOUNT, init_once);
1538 if (hugetlbfs_inode_cachep == NULL)
1541 error = register_filesystem(&hugetlbfs_fs_type);
1545 /* default hstate mount is required */
1546 mnt = mount_one_hugetlbfs(&default_hstate);
1548 error = PTR_ERR(mnt);
1551 hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1553 /* other hstates are optional */
1555 for_each_hstate(h) {
1556 if (i == default_hstate_idx) {
1561 mnt = mount_one_hugetlbfs(h);
1563 hugetlbfs_vfsmount[i] = NULL;
1565 hugetlbfs_vfsmount[i] = mnt;
1572 (void)unregister_filesystem(&hugetlbfs_fs_type);
1574 kmem_cache_destroy(hugetlbfs_inode_cachep);
1578 fs_initcall(init_hugetlbfs_fs)