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_u32 ("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)
111 static void huge_pagevec_release(struct pagevec *pvec)
115 for (i = 0; i < pagevec_count(pvec); ++i)
116 put_page(pvec->pages[i]);
118 pagevec_reinit(pvec);
122 * Mask used when checking the page offset value passed in via system
123 * calls. This value will be converted to a loff_t which is signed.
124 * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
125 * value. The extra bit (- 1 in the shift value) is to take the sign
128 #define PGOFF_LOFFT_MAX \
129 (((1UL << (PAGE_SHIFT + 1)) - 1) << (BITS_PER_LONG - (PAGE_SHIFT + 1)))
131 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
133 struct inode *inode = file_inode(file);
136 struct hstate *h = hstate_file(file);
139 * vma address alignment (but not the pgoff alignment) has
140 * already been checked by prepare_hugepage_range. If you add
141 * any error returns here, do so after setting VM_HUGETLB, so
142 * is_vm_hugetlb_page tests below unmap_region go the right
143 * way when do_mmap unwinds (may be important on powerpc
146 vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
147 vma->vm_ops = &hugetlb_vm_ops;
150 * page based offset in vm_pgoff could be sufficiently large to
151 * overflow a loff_t when converted to byte offset. This can
152 * only happen on architectures where sizeof(loff_t) ==
153 * sizeof(unsigned long). So, only check in those instances.
155 if (sizeof(unsigned long) == sizeof(loff_t)) {
156 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
160 /* must be huge page aligned */
161 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
164 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
165 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
166 /* check for overflow */
174 if (!hugetlb_reserve_pages(inode,
175 vma->vm_pgoff >> huge_page_order(h),
176 len >> huge_page_shift(h), vma,
181 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
182 i_size_write(inode, len);
190 * Called under mmap_write_lock(mm).
193 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
195 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
196 unsigned long len, unsigned long pgoff, unsigned long flags)
198 struct hstate *h = hstate_file(file);
199 struct vm_unmapped_area_info info;
203 info.low_limit = current->mm->mmap_base;
204 info.high_limit = TASK_SIZE;
205 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
206 info.align_offset = 0;
207 return vm_unmapped_area(&info);
211 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
212 unsigned long len, unsigned long pgoff, unsigned long flags)
214 struct hstate *h = hstate_file(file);
215 struct vm_unmapped_area_info info;
217 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
219 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
220 info.high_limit = current->mm->mmap_base;
221 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
222 info.align_offset = 0;
223 addr = vm_unmapped_area(&info);
226 * A failed mmap() very likely causes application failure,
227 * so fall back to the bottom-up function here. This scenario
228 * can happen with large stack limits and large mmap()
231 if (unlikely(offset_in_page(addr))) {
232 VM_BUG_ON(addr != -ENOMEM);
234 info.low_limit = current->mm->mmap_base;
235 info.high_limit = TASK_SIZE;
236 addr = vm_unmapped_area(&info);
243 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
244 unsigned long len, unsigned long pgoff, unsigned long flags)
246 struct mm_struct *mm = current->mm;
247 struct vm_area_struct *vma;
248 struct hstate *h = hstate_file(file);
250 if (len & ~huge_page_mask(h))
255 if (flags & MAP_FIXED) {
256 if (prepare_hugepage_range(file, addr, len))
262 addr = ALIGN(addr, huge_page_size(h));
263 vma = find_vma(mm, addr);
264 if (TASK_SIZE - len >= addr &&
265 (!vma || addr + len <= vm_start_gap(vma)))
270 * Use mm->get_unmapped_area value as a hint to use topdown routine.
271 * If architectures have special needs, they should define their own
272 * version of hugetlb_get_unmapped_area.
274 if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
275 return hugetlb_get_unmapped_area_topdown(file, addr, len,
277 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
283 hugetlbfs_read_actor(struct page *page, unsigned long offset,
284 struct iov_iter *to, unsigned long size)
289 /* Find which 4k chunk and offset with in that chunk */
290 i = offset >> PAGE_SHIFT;
291 offset = offset & ~PAGE_MASK;
295 chunksize = PAGE_SIZE;
298 if (chunksize > size)
300 n = copy_page_to_iter(&page[i], offset, chunksize, to);
312 * Support for read() - Find the page attached to f_mapping and copy out the
313 * data. Its *very* similar to generic_file_buffered_read(), we can't use that
314 * since it has PAGE_SIZE assumptions.
316 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
318 struct file *file = iocb->ki_filp;
319 struct hstate *h = hstate_file(file);
320 struct address_space *mapping = file->f_mapping;
321 struct inode *inode = mapping->host;
322 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
323 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
324 unsigned long end_index;
328 while (iov_iter_count(to)) {
332 /* nr is the maximum number of bytes to copy from this page */
333 nr = huge_page_size(h);
334 isize = i_size_read(inode);
337 end_index = (isize - 1) >> huge_page_shift(h);
338 if (index > end_index)
340 if (index == end_index) {
341 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
348 page = find_lock_page(mapping, index);
349 if (unlikely(page == NULL)) {
351 * We have a HOLE, zero out the user-buffer for the
352 * length of the hole or request.
354 copied = iov_iter_zero(nr, to);
359 * We have the page, copy it to user space buffer.
361 copied = hugetlbfs_read_actor(page, offset, to, nr);
366 if (copied != nr && iov_iter_count(to)) {
371 index += offset >> huge_page_shift(h);
372 offset &= ~huge_page_mask(h);
374 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
378 static int hugetlbfs_write_begin(struct file *file,
379 struct address_space *mapping,
380 loff_t pos, unsigned len, unsigned flags,
381 struct page **pagep, void **fsdata)
386 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
387 loff_t pos, unsigned len, unsigned copied,
388 struct page *page, void *fsdata)
394 static void remove_huge_page(struct page *page)
396 ClearPageDirty(page);
397 ClearPageUptodate(page);
398 delete_from_page_cache(page);
402 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
404 struct vm_area_struct *vma;
407 * end == 0 indicates that the entire range after
408 * start should be unmapped.
410 vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
411 unsigned long v_offset;
415 * Can the expression below overflow on 32-bit arches?
416 * No, because the interval tree returns us only those vmas
417 * which overlap the truncated area starting at pgoff,
418 * and no vma on a 32-bit arch can span beyond the 4GB.
420 if (vma->vm_pgoff < start)
421 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
428 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
430 if (v_end > vma->vm_end)
434 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
440 * remove_inode_hugepages handles two distinct cases: truncation and hole
441 * punch. There are subtle differences in operation for each case.
443 * truncation is indicated by end of range being LLONG_MAX
444 * In this case, we first scan the range and release found pages.
445 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
446 * maps and global counts. Page faults can not race with truncation
447 * in this routine. hugetlb_no_page() holds i_mmap_rwsem and prevents
448 * page faults in the truncated range by checking i_size. i_size is
449 * modified while holding i_mmap_rwsem.
450 * hole punch is indicated if end is not LLONG_MAX
451 * In the hole punch case we scan the range and release found pages.
452 * Only when releasing a page is the associated region/reserve map
453 * deleted. The region/reserve map for ranges without associated
454 * pages are not modified. Page faults can race with hole punch.
455 * This is indicated if we find a mapped page.
456 * Note: If the passed end of range value is beyond the end of file, but
457 * not LLONG_MAX this routine still performs a hole punch operation.
459 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
462 struct hstate *h = hstate_inode(inode);
463 struct address_space *mapping = &inode->i_data;
464 const pgoff_t start = lstart >> huge_page_shift(h);
465 const pgoff_t end = lend >> huge_page_shift(h);
469 bool truncate_op = (lend == LLONG_MAX);
475 * When no more pages are found, we are done.
477 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
480 for (i = 0; i < pagevec_count(&pvec); ++i) {
481 struct page *page = pvec.pages[i];
487 * Only need to hold the fault mutex in the
488 * hole punch case. This prevents races with
489 * page faults. Races are not possible in the
490 * case of truncation.
492 hash = hugetlb_fault_mutex_hash(mapping, index);
493 mutex_lock(&hugetlb_fault_mutex_table[hash]);
497 * If page is mapped, it was faulted in after being
498 * unmapped in caller. Unmap (again) now after taking
499 * the fault mutex. The mutex will prevent faults
500 * until we finish removing the page.
502 * This race can only happen in the hole punch case.
503 * Getting here in a truncate operation is a bug.
505 if (unlikely(page_mapped(page))) {
508 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
509 i_mmap_lock_write(mapping);
510 mutex_lock(&hugetlb_fault_mutex_table[hash]);
511 hugetlb_vmdelete_list(&mapping->i_mmap,
512 index * pages_per_huge_page(h),
513 (index + 1) * pages_per_huge_page(h));
514 i_mmap_unlock_write(mapping);
519 * We must free the huge page and remove from page
520 * cache (remove_huge_page) BEFORE removing the
521 * region/reserve map (hugetlb_unreserve_pages). In
522 * rare out of memory conditions, removal of the
523 * region/reserve map could fail. Correspondingly,
524 * the subpool and global reserve usage count can need
527 VM_BUG_ON(PagePrivate(page));
528 remove_huge_page(page);
531 if (unlikely(hugetlb_unreserve_pages(inode,
532 index, index + 1, 1)))
533 hugetlb_fix_reserve_counts(inode);
538 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
540 huge_pagevec_release(&pvec);
545 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
548 static void hugetlbfs_evict_inode(struct inode *inode)
550 struct resv_map *resv_map;
552 remove_inode_hugepages(inode, 0, LLONG_MAX);
555 * Get the resv_map from the address space embedded in the inode.
556 * This is the address space which points to any resv_map allocated
557 * at inode creation time. If this is a device special inode,
558 * i_mapping may not point to the original address space.
560 resv_map = (struct resv_map *)(&inode->i_data)->private_data;
561 /* Only regular and link inodes have associated reserve maps */
563 resv_map_release(&resv_map->refs);
567 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
570 struct address_space *mapping = inode->i_mapping;
571 struct hstate *h = hstate_inode(inode);
573 BUG_ON(offset & ~huge_page_mask(h));
574 pgoff = offset >> PAGE_SHIFT;
576 i_mmap_lock_write(mapping);
577 i_size_write(inode, offset);
578 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
579 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
580 i_mmap_unlock_write(mapping);
581 remove_inode_hugepages(inode, offset, LLONG_MAX);
584 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
586 struct hstate *h = hstate_inode(inode);
587 loff_t hpage_size = huge_page_size(h);
588 loff_t hole_start, hole_end;
591 * For hole punch round up the beginning offset of the hole and
592 * round down the end.
594 hole_start = round_up(offset, hpage_size);
595 hole_end = round_down(offset + len, hpage_size);
597 if (hole_end > hole_start) {
598 struct address_space *mapping = inode->i_mapping;
599 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
603 /* protected by i_rwsem */
604 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
609 i_mmap_lock_write(mapping);
610 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
611 hugetlb_vmdelete_list(&mapping->i_mmap,
612 hole_start >> PAGE_SHIFT,
613 hole_end >> PAGE_SHIFT);
614 i_mmap_unlock_write(mapping);
615 remove_inode_hugepages(inode, hole_start, hole_end);
622 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
625 struct inode *inode = file_inode(file);
626 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
627 struct address_space *mapping = inode->i_mapping;
628 struct hstate *h = hstate_inode(inode);
629 struct vm_area_struct pseudo_vma;
630 struct mm_struct *mm = current->mm;
631 loff_t hpage_size = huge_page_size(h);
632 unsigned long hpage_shift = huge_page_shift(h);
633 pgoff_t start, index, end;
637 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
640 if (mode & FALLOC_FL_PUNCH_HOLE)
641 return hugetlbfs_punch_hole(inode, offset, len);
644 * Default preallocate case.
645 * For this range, start is rounded down and end is rounded up
646 * as well as being converted to page offsets.
648 start = offset >> hpage_shift;
649 end = (offset + len + hpage_size - 1) >> hpage_shift;
653 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
654 error = inode_newsize_ok(inode, offset + len);
658 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
664 * Initialize a pseudo vma as this is required by the huge page
665 * allocation routines. If NUMA is configured, use page index
666 * as input to create an allocation policy.
668 vma_init(&pseudo_vma, mm);
669 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
670 pseudo_vma.vm_file = file;
672 for (index = start; index < end; index++) {
674 * This is supposed to be the vaddr where the page is being
675 * faulted in, but we have no vaddr here.
683 * fallocate(2) manpage permits EINTR; we may have been
684 * interrupted because we are using up too much memory.
686 if (signal_pending(current)) {
691 /* Set numa allocation policy based on index */
692 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
694 /* addr is the offset within the file (zero based) */
695 addr = index * hpage_size;
698 * fault mutex taken here, protects against fault path
699 * and hole punch. inode_lock previously taken protects
700 * against truncation.
702 hash = hugetlb_fault_mutex_hash(mapping, index);
703 mutex_lock(&hugetlb_fault_mutex_table[hash]);
705 /* See if already present in mapping to avoid alloc/free */
706 page = find_get_page(mapping, index);
709 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
710 hugetlb_drop_vma_policy(&pseudo_vma);
715 * Allocate page without setting the avoid_reserve argument.
716 * There certainly are no reserves associated with the
717 * pseudo_vma. However, there could be shared mappings with
718 * reserves for the file at the inode level. If we fallocate
719 * pages in these areas, we need to consume the reserves
720 * to keep reservation accounting consistent.
722 page = alloc_huge_page(&pseudo_vma, addr, 0);
723 hugetlb_drop_vma_policy(&pseudo_vma);
725 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
726 error = PTR_ERR(page);
729 clear_huge_page(page, addr, pages_per_huge_page(h));
730 __SetPageUptodate(page);
731 error = huge_add_to_page_cache(page, mapping, index);
732 if (unlikely(error)) {
734 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
738 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
740 SetHPageMigratable(page);
742 * unlock_page because locked by add_to_page_cache()
743 * put_page() due to reference from alloc_huge_page()
749 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
750 i_size_write(inode, offset + len);
751 inode->i_ctime = current_time(inode);
757 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
758 struct dentry *dentry, struct iattr *attr)
760 struct inode *inode = d_inode(dentry);
761 struct hstate *h = hstate_inode(inode);
763 unsigned int ia_valid = attr->ia_valid;
764 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
766 error = setattr_prepare(&init_user_ns, dentry, attr);
770 if (ia_valid & ATTR_SIZE) {
771 loff_t oldsize = inode->i_size;
772 loff_t newsize = attr->ia_size;
774 if (newsize & ~huge_page_mask(h))
776 /* protected by i_rwsem */
777 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
778 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
780 hugetlb_vmtruncate(inode, newsize);
783 setattr_copy(&init_user_ns, inode, attr);
784 mark_inode_dirty(inode);
788 static struct inode *hugetlbfs_get_root(struct super_block *sb,
789 struct hugetlbfs_fs_context *ctx)
793 inode = new_inode(sb);
795 inode->i_ino = get_next_ino();
796 inode->i_mode = S_IFDIR | ctx->mode;
797 inode->i_uid = ctx->uid;
798 inode->i_gid = ctx->gid;
799 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
800 inode->i_op = &hugetlbfs_dir_inode_operations;
801 inode->i_fop = &simple_dir_operations;
802 /* directory inodes start off with i_nlink == 2 (for "." entry) */
804 lockdep_annotate_inode_mutex_key(inode);
810 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
811 * be taken from reclaim -- unlike regular filesystems. This needs an
812 * annotation because huge_pmd_share() does an allocation under hugetlb's
815 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
817 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
819 umode_t mode, dev_t dev)
822 struct resv_map *resv_map = NULL;
825 * Reserve maps are only needed for inodes that can have associated
828 if (S_ISREG(mode) || S_ISLNK(mode)) {
829 resv_map = resv_map_alloc();
834 inode = new_inode(sb);
836 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
838 inode->i_ino = get_next_ino();
839 inode_init_owner(&init_user_ns, inode, dir, mode);
840 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
841 &hugetlbfs_i_mmap_rwsem_key);
842 inode->i_mapping->a_ops = &hugetlbfs_aops;
843 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
844 inode->i_mapping->private_data = resv_map;
845 info->seals = F_SEAL_SEAL;
846 switch (mode & S_IFMT) {
848 init_special_inode(inode, mode, dev);
851 inode->i_op = &hugetlbfs_inode_operations;
852 inode->i_fop = &hugetlbfs_file_operations;
855 inode->i_op = &hugetlbfs_dir_inode_operations;
856 inode->i_fop = &simple_dir_operations;
858 /* directory inodes start off with i_nlink == 2 (for "." entry) */
862 inode->i_op = &page_symlink_inode_operations;
863 inode_nohighmem(inode);
866 lockdep_annotate_inode_mutex_key(inode);
869 kref_put(&resv_map->refs, resv_map_release);
876 * File creation. Allocate an inode, and we're done..
878 static int do_hugetlbfs_mknod(struct inode *dir,
879 struct dentry *dentry,
887 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
889 dir->i_ctime = dir->i_mtime = current_time(dir);
891 d_tmpfile(dentry, inode);
893 d_instantiate(dentry, inode);
894 dget(dentry);/* Extra count - pin the dentry in core */
901 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
902 struct dentry *dentry, umode_t mode, dev_t dev)
904 return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
907 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
908 struct dentry *dentry, umode_t mode)
910 int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
917 static int hugetlbfs_create(struct user_namespace *mnt_userns,
918 struct inode *dir, struct dentry *dentry,
919 umode_t mode, bool excl)
921 return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
924 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
925 struct inode *dir, struct dentry *dentry,
928 return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
931 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
932 struct inode *dir, struct dentry *dentry,
938 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
940 int l = strlen(symname)+1;
941 error = page_symlink(inode, symname, l);
943 d_instantiate(dentry, inode);
948 dir->i_ctime = dir->i_mtime = current_time(dir);
953 static int hugetlbfs_migrate_page(struct address_space *mapping,
954 struct page *newpage, struct page *page,
955 enum migrate_mode mode)
959 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
960 if (rc != MIGRATEPAGE_SUCCESS)
963 if (hugetlb_page_subpool(page)) {
964 hugetlb_set_page_subpool(newpage, hugetlb_page_subpool(page));
965 hugetlb_set_page_subpool(page, NULL);
968 if (mode != MIGRATE_SYNC_NO_COPY)
969 migrate_page_copy(newpage, page);
971 migrate_page_states(newpage, page);
973 return MIGRATEPAGE_SUCCESS;
976 static int hugetlbfs_error_remove_page(struct address_space *mapping,
979 struct inode *inode = mapping->host;
980 pgoff_t index = page->index;
982 remove_huge_page(page);
983 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
984 hugetlb_fix_reserve_counts(inode);
990 * Display the mount options in /proc/mounts.
992 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
994 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
995 struct hugepage_subpool *spool = sbinfo->spool;
996 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
997 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1000 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1001 seq_printf(m, ",uid=%u",
1002 from_kuid_munged(&init_user_ns, sbinfo->uid));
1003 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1004 seq_printf(m, ",gid=%u",
1005 from_kgid_munged(&init_user_ns, sbinfo->gid));
1006 if (sbinfo->mode != 0755)
1007 seq_printf(m, ",mode=%o", sbinfo->mode);
1008 if (sbinfo->max_inodes != -1)
1009 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1013 if (hpage_size >= 1024) {
1017 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1019 if (spool->max_hpages != -1)
1020 seq_printf(m, ",size=%llu",
1021 (unsigned long long)spool->max_hpages << hpage_shift);
1022 if (spool->min_hpages != -1)
1023 seq_printf(m, ",min_size=%llu",
1024 (unsigned long long)spool->min_hpages << hpage_shift);
1029 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1031 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1032 struct hstate *h = hstate_inode(d_inode(dentry));
1034 buf->f_type = HUGETLBFS_MAGIC;
1035 buf->f_bsize = huge_page_size(h);
1037 spin_lock(&sbinfo->stat_lock);
1038 /* If no limits set, just report 0 for max/free/used
1039 * blocks, like simple_statfs() */
1040 if (sbinfo->spool) {
1043 spin_lock(&sbinfo->spool->lock);
1044 buf->f_blocks = sbinfo->spool->max_hpages;
1045 free_pages = sbinfo->spool->max_hpages
1046 - sbinfo->spool->used_hpages;
1047 buf->f_bavail = buf->f_bfree = free_pages;
1048 spin_unlock(&sbinfo->spool->lock);
1049 buf->f_files = sbinfo->max_inodes;
1050 buf->f_ffree = sbinfo->free_inodes;
1052 spin_unlock(&sbinfo->stat_lock);
1054 buf->f_namelen = NAME_MAX;
1058 static void hugetlbfs_put_super(struct super_block *sb)
1060 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1063 sb->s_fs_info = NULL;
1066 hugepage_put_subpool(sbi->spool);
1072 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1074 if (sbinfo->free_inodes >= 0) {
1075 spin_lock(&sbinfo->stat_lock);
1076 if (unlikely(!sbinfo->free_inodes)) {
1077 spin_unlock(&sbinfo->stat_lock);
1080 sbinfo->free_inodes--;
1081 spin_unlock(&sbinfo->stat_lock);
1087 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1089 if (sbinfo->free_inodes >= 0) {
1090 spin_lock(&sbinfo->stat_lock);
1091 sbinfo->free_inodes++;
1092 spin_unlock(&sbinfo->stat_lock);
1097 static struct kmem_cache *hugetlbfs_inode_cachep;
1099 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1101 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1102 struct hugetlbfs_inode_info *p;
1104 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1106 p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1108 hugetlbfs_inc_free_inodes(sbinfo);
1113 * Any time after allocation, hugetlbfs_destroy_inode can be called
1114 * for the inode. mpol_free_shared_policy is unconditionally called
1115 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1116 * in case of a quick call to destroy.
1118 * Note that the policy is initialized even if we are creating a
1119 * private inode. This simplifies hugetlbfs_destroy_inode.
1121 mpol_shared_policy_init(&p->policy, NULL);
1123 return &p->vfs_inode;
1126 static void hugetlbfs_free_inode(struct inode *inode)
1128 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1131 static void hugetlbfs_destroy_inode(struct inode *inode)
1133 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1134 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1137 static const struct address_space_operations hugetlbfs_aops = {
1138 .write_begin = hugetlbfs_write_begin,
1139 .write_end = hugetlbfs_write_end,
1140 .set_page_dirty = __set_page_dirty_no_writeback,
1141 .migratepage = hugetlbfs_migrate_page,
1142 .error_remove_page = hugetlbfs_error_remove_page,
1146 static void init_once(void *foo)
1148 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1150 inode_init_once(&ei->vfs_inode);
1153 const struct file_operations hugetlbfs_file_operations = {
1154 .read_iter = hugetlbfs_read_iter,
1155 .mmap = hugetlbfs_file_mmap,
1156 .fsync = noop_fsync,
1157 .get_unmapped_area = hugetlb_get_unmapped_area,
1158 .llseek = default_llseek,
1159 .fallocate = hugetlbfs_fallocate,
1162 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1163 .create = hugetlbfs_create,
1164 .lookup = simple_lookup,
1165 .link = simple_link,
1166 .unlink = simple_unlink,
1167 .symlink = hugetlbfs_symlink,
1168 .mkdir = hugetlbfs_mkdir,
1169 .rmdir = simple_rmdir,
1170 .mknod = hugetlbfs_mknod,
1171 .rename = simple_rename,
1172 .setattr = hugetlbfs_setattr,
1173 .tmpfile = hugetlbfs_tmpfile,
1176 static const struct inode_operations hugetlbfs_inode_operations = {
1177 .setattr = hugetlbfs_setattr,
1180 static const struct super_operations hugetlbfs_ops = {
1181 .alloc_inode = hugetlbfs_alloc_inode,
1182 .free_inode = hugetlbfs_free_inode,
1183 .destroy_inode = hugetlbfs_destroy_inode,
1184 .evict_inode = hugetlbfs_evict_inode,
1185 .statfs = hugetlbfs_statfs,
1186 .put_super = hugetlbfs_put_super,
1187 .show_options = hugetlbfs_show_options,
1191 * Convert size option passed from command line to number of huge pages
1192 * in the pool specified by hstate. Size option could be in bytes
1193 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1196 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1197 enum hugetlbfs_size_type val_type)
1199 if (val_type == NO_SIZE)
1202 if (val_type == SIZE_PERCENT) {
1203 size_opt <<= huge_page_shift(h);
1204 size_opt *= h->max_huge_pages;
1205 do_div(size_opt, 100);
1208 size_opt >>= huge_page_shift(h);
1213 * Parse one mount parameter.
1215 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1217 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1218 struct fs_parse_result result;
1223 opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1229 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1230 if (!uid_valid(ctx->uid))
1235 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1236 if (!gid_valid(ctx->gid))
1241 ctx->mode = result.uint_32 & 01777U;
1245 /* memparse() will accept a K/M/G without a digit */
1246 if (!isdigit(param->string[0]))
1248 ctx->max_size_opt = memparse(param->string, &rest);
1249 ctx->max_val_type = SIZE_STD;
1251 ctx->max_val_type = SIZE_PERCENT;
1255 /* memparse() will accept a K/M/G without a digit */
1256 if (!isdigit(param->string[0]))
1258 ctx->nr_inodes = memparse(param->string, &rest);
1262 ps = memparse(param->string, &rest);
1263 ctx->hstate = size_to_hstate(ps);
1265 pr_err("Unsupported page size %lu MB\n", ps >> 20);
1271 /* memparse() will accept a K/M/G without a digit */
1272 if (!isdigit(param->string[0]))
1274 ctx->min_size_opt = memparse(param->string, &rest);
1275 ctx->min_val_type = SIZE_STD;
1277 ctx->min_val_type = SIZE_PERCENT;
1285 return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1286 param->string, param->key);
1290 * Validate the parsed options.
1292 static int hugetlbfs_validate(struct fs_context *fc)
1294 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1297 * Use huge page pool size (in hstate) to convert the size
1298 * options to number of huge pages. If NO_SIZE, -1 is returned.
1300 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1303 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1308 * If max_size was specified, then min_size must be smaller
1310 if (ctx->max_val_type > NO_SIZE &&
1311 ctx->min_hpages > ctx->max_hpages) {
1312 pr_err("Minimum size can not be greater than maximum size\n");
1320 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1322 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1323 struct hugetlbfs_sb_info *sbinfo;
1325 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1328 sb->s_fs_info = sbinfo;
1329 spin_lock_init(&sbinfo->stat_lock);
1330 sbinfo->hstate = ctx->hstate;
1331 sbinfo->max_inodes = ctx->nr_inodes;
1332 sbinfo->free_inodes = ctx->nr_inodes;
1333 sbinfo->spool = NULL;
1334 sbinfo->uid = ctx->uid;
1335 sbinfo->gid = ctx->gid;
1336 sbinfo->mode = ctx->mode;
1339 * Allocate and initialize subpool if maximum or minimum size is
1340 * specified. Any needed reservations (for minimum size) are taken
1341 * taken when the subpool is created.
1343 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1344 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1350 sb->s_maxbytes = MAX_LFS_FILESIZE;
1351 sb->s_blocksize = huge_page_size(ctx->hstate);
1352 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1353 sb->s_magic = HUGETLBFS_MAGIC;
1354 sb->s_op = &hugetlbfs_ops;
1355 sb->s_time_gran = 1;
1358 * Due to the special and limited functionality of hugetlbfs, it does
1359 * not work well as a stacking filesystem.
1361 sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1362 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1367 kfree(sbinfo->spool);
1372 static int hugetlbfs_get_tree(struct fs_context *fc)
1374 int err = hugetlbfs_validate(fc);
1377 return get_tree_nodev(fc, hugetlbfs_fill_super);
1380 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1382 kfree(fc->fs_private);
1385 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1386 .free = hugetlbfs_fs_context_free,
1387 .parse_param = hugetlbfs_parse_param,
1388 .get_tree = hugetlbfs_get_tree,
1391 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1393 struct hugetlbfs_fs_context *ctx;
1395 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1399 ctx->max_hpages = -1; /* No limit on size by default */
1400 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1401 ctx->uid = current_fsuid();
1402 ctx->gid = current_fsgid();
1404 ctx->hstate = &default_hstate;
1405 ctx->min_hpages = -1; /* No default minimum size */
1406 ctx->max_val_type = NO_SIZE;
1407 ctx->min_val_type = NO_SIZE;
1408 fc->fs_private = ctx;
1409 fc->ops = &hugetlbfs_fs_context_ops;
1413 static struct file_system_type hugetlbfs_fs_type = {
1414 .name = "hugetlbfs",
1415 .init_fs_context = hugetlbfs_init_fs_context,
1416 .parameters = hugetlb_fs_parameters,
1417 .kill_sb = kill_litter_super,
1420 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1422 static int can_do_hugetlb_shm(void)
1425 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1426 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1429 static int get_hstate_idx(int page_size_log)
1431 struct hstate *h = hstate_sizelog(page_size_log);
1435 return hstate_index(h);
1439 * Note that size should be aligned to proper hugepage size in caller side,
1440 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1442 struct file *hugetlb_file_setup(const char *name, size_t size,
1443 vm_flags_t acctflag, struct user_struct **user,
1444 int creat_flags, int page_size_log)
1446 struct inode *inode;
1447 struct vfsmount *mnt;
1451 hstate_idx = get_hstate_idx(page_size_log);
1453 return ERR_PTR(-ENODEV);
1456 mnt = hugetlbfs_vfsmount[hstate_idx];
1458 return ERR_PTR(-ENOENT);
1460 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1461 *user = current_user();
1462 if (user_shm_lock(size, *user)) {
1464 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1465 current->comm, current->pid);
1466 task_unlock(current);
1469 return ERR_PTR(-EPERM);
1473 file = ERR_PTR(-ENOSPC);
1474 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1477 if (creat_flags == HUGETLB_SHMFS_INODE)
1478 inode->i_flags |= S_PRIVATE;
1480 inode->i_size = size;
1483 if (!hugetlb_reserve_pages(inode, 0,
1484 size >> huge_page_shift(hstate_inode(inode)), NULL,
1486 file = ERR_PTR(-ENOMEM);
1488 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1489 &hugetlbfs_file_operations);
1496 user_shm_unlock(size, *user);
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) >> 10);
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)