2 * hugetlbpage-backed filesystem. Based on ramfs.
4 * Nadia Yvette Chambers, 2002
6 * Copyright (C) 2002 Linus Torvalds.
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 #include <linux/thread_info.h>
13 #include <asm/current.h>
14 #include <linux/sched/signal.h> /* remove ASAP */
15 #include <linux/falloc.h>
17 #include <linux/mount.h>
18 #include <linux/file.h>
19 #include <linux/kernel.h>
20 #include <linux/writeback.h>
21 #include <linux/pagemap.h>
22 #include <linux/highmem.h>
23 #include <linux/init.h>
24 #include <linux/string.h>
25 #include <linux/capability.h>
26 #include <linux/ctype.h>
27 #include <linux/backing-dev.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagevec.h>
30 #include <linux/fs_parser.h>
31 #include <linux/mman.h>
32 #include <linux/slab.h>
33 #include <linux/dnotify.h>
34 #include <linux/statfs.h>
35 #include <linux/security.h>
36 #include <linux/magic.h>
37 #include <linux/migrate.h>
38 #include <linux/uio.h>
40 #include <linux/uaccess.h>
41 #include <linux/sched/mm.h>
43 static const struct super_operations hugetlbfs_ops;
44 static const struct address_space_operations hugetlbfs_aops;
45 const struct file_operations hugetlbfs_file_operations;
46 static const struct inode_operations hugetlbfs_dir_inode_operations;
47 static const struct inode_operations hugetlbfs_inode_operations;
49 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
51 struct hugetlbfs_fs_context {
52 struct hstate *hstate;
53 unsigned long long max_size_opt;
54 unsigned long long min_size_opt;
58 enum hugetlbfs_size_type max_val_type;
59 enum hugetlbfs_size_type min_val_type;
65 int sysctl_hugetlb_shm_group;
77 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
78 fsparam_u32 ("gid", Opt_gid),
79 fsparam_string("min_size", Opt_min_size),
80 fsparam_u32oct("mode", Opt_mode),
81 fsparam_string("nr_inodes", Opt_nr_inodes),
82 fsparam_string("pagesize", Opt_pagesize),
83 fsparam_string("size", Opt_size),
84 fsparam_u32 ("uid", Opt_uid),
89 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
90 struct inode *inode, pgoff_t index)
92 vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
96 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
98 mpol_cond_put(vma->vm_policy);
101 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
102 struct inode *inode, pgoff_t index)
106 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
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);
134 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
137 struct hstate *h = hstate_file(file);
141 * vma address alignment (but not the pgoff alignment) has
142 * already been checked by prepare_hugepage_range. If you add
143 * any error returns here, do so after setting VM_HUGETLB, so
144 * is_vm_hugetlb_page tests below unmap_region go the right
145 * way when do_mmap unwinds (may be important on powerpc
148 vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
149 vma->vm_ops = &hugetlb_vm_ops;
151 ret = seal_check_future_write(info->seals, vma);
156 * page based offset in vm_pgoff could be sufficiently large to
157 * overflow a loff_t when converted to byte offset. This can
158 * only happen on architectures where sizeof(loff_t) ==
159 * sizeof(unsigned long). So, only check in those instances.
161 if (sizeof(unsigned long) == sizeof(loff_t)) {
162 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
166 /* must be huge page aligned */
167 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
170 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
171 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
172 /* check for overflow */
181 vm_flags = vma->vm_flags;
183 * for SHM_HUGETLB, the pages are reserved in the shmget() call so skip
184 * reserving here. Note: only for SHM hugetlbfs file, the inode
185 * flag S_PRIVATE is set.
187 if (inode->i_flags & S_PRIVATE)
188 vm_flags |= VM_NORESERVE;
190 if (!hugetlb_reserve_pages(inode,
191 vma->vm_pgoff >> huge_page_order(h),
192 len >> huge_page_shift(h), vma,
197 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
198 i_size_write(inode, len);
206 * Called under mmap_write_lock(mm).
209 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
211 hugetlb_get_unmapped_area_bottomup(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;
219 info.low_limit = current->mm->mmap_base;
220 info.high_limit = arch_get_mmap_end(addr);
221 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
222 info.align_offset = 0;
223 return vm_unmapped_area(&info);
227 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
228 unsigned long len, unsigned long pgoff, unsigned long flags)
230 struct hstate *h = hstate_file(file);
231 struct vm_unmapped_area_info info;
233 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
235 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
236 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
237 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
238 info.align_offset = 0;
239 addr = vm_unmapped_area(&info);
242 * A failed mmap() very likely causes application failure,
243 * so fall back to the bottom-up function here. This scenario
244 * can happen with large stack limits and large mmap()
247 if (unlikely(offset_in_page(addr))) {
248 VM_BUG_ON(addr != -ENOMEM);
250 info.low_limit = current->mm->mmap_base;
251 info.high_limit = arch_get_mmap_end(addr);
252 addr = vm_unmapped_area(&info);
259 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
260 unsigned long len, unsigned long pgoff, unsigned long flags)
262 struct mm_struct *mm = current->mm;
263 struct vm_area_struct *vma;
264 struct hstate *h = hstate_file(file);
265 const unsigned long mmap_end = arch_get_mmap_end(addr);
267 if (len & ~huge_page_mask(h))
272 if (flags & MAP_FIXED) {
273 if (prepare_hugepage_range(file, addr, len))
279 addr = ALIGN(addr, huge_page_size(h));
280 vma = find_vma(mm, addr);
281 if (mmap_end - len >= addr &&
282 (!vma || addr + len <= vm_start_gap(vma)))
287 * Use mm->get_unmapped_area value as a hint to use topdown routine.
288 * If architectures have special needs, they should define their own
289 * version of hugetlb_get_unmapped_area.
291 if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
292 return hugetlb_get_unmapped_area_topdown(file, addr, len,
294 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
300 hugetlbfs_read_actor(struct page *page, unsigned long offset,
301 struct iov_iter *to, unsigned long size)
306 /* Find which 4k chunk and offset with in that chunk */
307 i = offset >> PAGE_SHIFT;
308 offset = offset & ~PAGE_MASK;
312 chunksize = PAGE_SIZE;
315 if (chunksize > size)
317 n = copy_page_to_iter(&page[i], offset, chunksize, to);
329 * Support for read() - Find the page attached to f_mapping and copy out the
330 * data. Its *very* similar to do_generic_mapping_read(), we can't use that
331 * since it has PAGE_SIZE assumptions.
333 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
335 struct file *file = iocb->ki_filp;
336 struct hstate *h = hstate_file(file);
337 struct address_space *mapping = file->f_mapping;
338 struct inode *inode = mapping->host;
339 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
340 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
341 unsigned long end_index;
345 while (iov_iter_count(to)) {
349 /* nr is the maximum number of bytes to copy from this page */
350 nr = huge_page_size(h);
351 isize = i_size_read(inode);
354 end_index = (isize - 1) >> huge_page_shift(h);
355 if (index > end_index)
357 if (index == end_index) {
358 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
365 page = find_lock_page(mapping, index);
366 if (unlikely(page == NULL)) {
368 * We have a HOLE, zero out the user-buffer for the
369 * length of the hole or request.
371 copied = iov_iter_zero(nr, to);
376 * We have the page, copy it to user space buffer.
378 copied = hugetlbfs_read_actor(page, offset, to, nr);
383 if (copied != nr && iov_iter_count(to)) {
388 index += offset >> huge_page_shift(h);
389 offset &= ~huge_page_mask(h);
391 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
395 static int hugetlbfs_write_begin(struct file *file,
396 struct address_space *mapping,
397 loff_t pos, unsigned len, unsigned flags,
398 struct page **pagep, void **fsdata)
403 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
404 loff_t pos, unsigned len, unsigned copied,
405 struct page *page, void *fsdata)
411 static void remove_huge_page(struct page *page)
413 ClearPageDirty(page);
414 ClearPageUptodate(page);
415 delete_from_page_cache(page);
419 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
421 struct vm_area_struct *vma;
424 * end == 0 indicates that the entire range after
425 * start should be unmapped.
427 vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
428 unsigned long v_offset;
432 * Can the expression below overflow on 32-bit arches?
433 * No, because the interval tree returns us only those vmas
434 * which overlap the truncated area starting at pgoff,
435 * and no vma on a 32-bit arch can span beyond the 4GB.
437 if (vma->vm_pgoff < start)
438 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
445 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
447 if (v_end > vma->vm_end)
451 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
457 * remove_inode_hugepages handles two distinct cases: truncation and hole
458 * punch. There are subtle differences in operation for each case.
460 * truncation is indicated by end of range being LLONG_MAX
461 * In this case, we first scan the range and release found pages.
462 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
463 * maps and global counts. Page faults can not race with truncation
464 * in this routine. hugetlb_no_page() holds i_mmap_rwsem and prevents
465 * page faults in the truncated range by checking i_size. i_size is
466 * modified while holding i_mmap_rwsem.
467 * hole punch is indicated if end is not LLONG_MAX
468 * In the hole punch case we scan the range and release found pages.
469 * Only when releasing a page is the associated region/reserv map
470 * deleted. The region/reserv map for ranges without associated
471 * pages are not modified. Page faults can race with hole punch.
472 * This is indicated if we find a mapped page.
473 * Note: If the passed end of range value is beyond the end of file, but
474 * not LLONG_MAX this routine still performs a hole punch operation.
476 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
479 struct hstate *h = hstate_inode(inode);
480 struct address_space *mapping = &inode->i_data;
481 const pgoff_t start = lstart >> huge_page_shift(h);
482 const pgoff_t end = lend >> huge_page_shift(h);
483 struct vm_area_struct pseudo_vma;
487 bool truncate_op = (lend == LLONG_MAX);
489 vma_init(&pseudo_vma, current->mm);
490 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
495 * When no more pages are found, we are done.
497 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
500 for (i = 0; i < pagevec_count(&pvec); ++i) {
501 struct page *page = pvec.pages[i];
505 hash = hugetlb_fault_mutex_hash(mapping, index);
508 * Only need to hold the fault mutex in the
509 * hole punch case. This prevents races with
510 * page faults. Races are not possible in the
511 * case of truncation.
513 mutex_lock(&hugetlb_fault_mutex_table[hash]);
517 * If page is mapped, it was faulted in after being
518 * unmapped in caller. Unmap (again) now after taking
519 * the fault mutex. The mutex will prevent faults
520 * until we finish removing the page.
522 * This race can only happen in the hole punch case.
523 * Getting here in a truncate operation is a bug.
525 if (unlikely(page_mapped(page))) {
528 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
529 i_mmap_lock_write(mapping);
530 mutex_lock(&hugetlb_fault_mutex_table[hash]);
531 hugetlb_vmdelete_list(&mapping->i_mmap,
532 index * pages_per_huge_page(h),
533 (index + 1) * pages_per_huge_page(h));
534 i_mmap_unlock_write(mapping);
539 * We must free the huge page and remove from page
540 * cache (remove_huge_page) BEFORE removing the
541 * region/reserve map (hugetlb_unreserve_pages). In
542 * rare out of memory conditions, removal of the
543 * region/reserve map could fail. Correspondingly,
544 * the subpool and global reserve usage count can need
547 VM_BUG_ON(PagePrivate(page));
548 remove_huge_page(page);
551 if (unlikely(hugetlb_unreserve_pages(inode,
552 index, index + 1, 1)))
553 hugetlb_fix_reserve_counts(inode);
558 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
560 huge_pagevec_release(&pvec);
565 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
568 static void hugetlbfs_evict_inode(struct inode *inode)
570 struct resv_map *resv_map;
572 remove_inode_hugepages(inode, 0, LLONG_MAX);
575 * Get the resv_map from the address space embedded in the inode.
576 * This is the address space which points to any resv_map allocated
577 * at inode creation time. If this is a device special inode,
578 * i_mapping may not point to the original address space.
580 resv_map = (struct resv_map *)(&inode->i_data)->private_data;
581 /* Only regular and link inodes have associated reserve maps */
583 resv_map_release(&resv_map->refs);
587 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
590 struct address_space *mapping = inode->i_mapping;
591 struct hstate *h = hstate_inode(inode);
593 BUG_ON(offset & ~huge_page_mask(h));
594 pgoff = offset >> PAGE_SHIFT;
596 i_mmap_lock_write(mapping);
597 i_size_write(inode, offset);
598 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
599 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
600 i_mmap_unlock_write(mapping);
601 remove_inode_hugepages(inode, offset, LLONG_MAX);
605 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
607 struct hstate *h = hstate_inode(inode);
608 loff_t hpage_size = huge_page_size(h);
609 loff_t hole_start, hole_end;
612 * For hole punch round up the beginning offset of the hole and
613 * round down the end.
615 hole_start = round_up(offset, hpage_size);
616 hole_end = round_down(offset + len, hpage_size);
618 if (hole_end > hole_start) {
619 struct address_space *mapping = inode->i_mapping;
620 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
624 /* protected by i_mutex */
625 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
630 i_mmap_lock_write(mapping);
631 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
632 hugetlb_vmdelete_list(&mapping->i_mmap,
633 hole_start >> PAGE_SHIFT,
634 hole_end >> PAGE_SHIFT);
635 i_mmap_unlock_write(mapping);
636 remove_inode_hugepages(inode, hole_start, hole_end);
643 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
646 struct inode *inode = file_inode(file);
647 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
648 struct address_space *mapping = inode->i_mapping;
649 struct hstate *h = hstate_inode(inode);
650 struct vm_area_struct pseudo_vma;
651 struct mm_struct *mm = current->mm;
652 loff_t hpage_size = huge_page_size(h);
653 unsigned long hpage_shift = huge_page_shift(h);
654 pgoff_t start, index, end;
658 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
661 if (mode & FALLOC_FL_PUNCH_HOLE)
662 return hugetlbfs_punch_hole(inode, offset, len);
665 * Default preallocate case.
666 * For this range, start is rounded down and end is rounded up
667 * as well as being converted to page offsets.
669 start = offset >> hpage_shift;
670 end = (offset + len + hpage_size - 1) >> hpage_shift;
674 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
675 error = inode_newsize_ok(inode, offset + len);
679 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
685 * Initialize a pseudo vma as this is required by the huge page
686 * allocation routines. If NUMA is configured, use page index
687 * as input to create an allocation policy.
689 vma_init(&pseudo_vma, mm);
690 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
691 pseudo_vma.vm_file = file;
693 for (index = start; index < end; index++) {
695 * This is supposed to be the vaddr where the page is being
696 * faulted in, but we have no vaddr here.
700 int avoid_reserve = 0;
705 * fallocate(2) manpage permits EINTR; we may have been
706 * interrupted because we are using up too much memory.
708 if (signal_pending(current)) {
713 /* Set numa allocation policy based on index */
714 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
716 /* addr is the offset within the file (zero based) */
717 addr = index * hpage_size;
720 * fault mutex taken here, protects against fault path
721 * and hole punch. inode_lock previously taken protects
722 * against truncation.
724 hash = hugetlb_fault_mutex_hash(mapping, index);
725 mutex_lock(&hugetlb_fault_mutex_table[hash]);
727 /* See if already present in mapping to avoid alloc/free */
728 page = find_get_page(mapping, index);
731 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
732 hugetlb_drop_vma_policy(&pseudo_vma);
736 /* Allocate page and add to page cache */
737 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
738 hugetlb_drop_vma_policy(&pseudo_vma);
740 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
741 error = PTR_ERR(page);
744 clear_huge_page(page, addr, pages_per_huge_page(h));
745 __SetPageUptodate(page);
746 error = huge_add_to_page_cache(page, mapping, index);
747 if (unlikely(error)) {
749 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
753 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
755 set_page_huge_active(page);
757 * unlock_page because locked by add_to_page_cache()
758 * put_page() due to reference from alloc_huge_page()
764 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
765 i_size_write(inode, offset + len);
766 inode->i_ctime = current_time(inode);
772 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
774 struct inode *inode = d_inode(dentry);
775 struct hstate *h = hstate_inode(inode);
777 unsigned int ia_valid = attr->ia_valid;
778 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
782 error = setattr_prepare(dentry, attr);
786 if (ia_valid & ATTR_SIZE) {
787 loff_t oldsize = inode->i_size;
788 loff_t newsize = attr->ia_size;
790 if (newsize & ~huge_page_mask(h))
792 /* protected by i_mutex */
793 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
794 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
796 error = hugetlb_vmtruncate(inode, newsize);
801 setattr_copy(inode, attr);
802 mark_inode_dirty(inode);
806 static struct inode *hugetlbfs_get_root(struct super_block *sb,
807 struct hugetlbfs_fs_context *ctx)
811 inode = new_inode(sb);
813 inode->i_ino = get_next_ino();
814 inode->i_mode = S_IFDIR | ctx->mode;
815 inode->i_uid = ctx->uid;
816 inode->i_gid = ctx->gid;
817 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
818 inode->i_op = &hugetlbfs_dir_inode_operations;
819 inode->i_fop = &simple_dir_operations;
820 /* directory inodes start off with i_nlink == 2 (for "." entry) */
822 lockdep_annotate_inode_mutex_key(inode);
828 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
829 * be taken from reclaim -- unlike regular filesystems. This needs an
830 * annotation because huge_pmd_share() does an allocation under hugetlb's
833 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
835 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
837 umode_t mode, dev_t dev)
840 struct resv_map *resv_map = NULL;
843 * Reserve maps are only needed for inodes that can have associated
846 if (S_ISREG(mode) || S_ISLNK(mode)) {
847 resv_map = resv_map_alloc();
852 inode = new_inode(sb);
854 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
856 inode->i_ino = get_next_ino();
857 inode_init_owner(inode, dir, mode);
858 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
859 &hugetlbfs_i_mmap_rwsem_key);
860 inode->i_mapping->a_ops = &hugetlbfs_aops;
861 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
862 inode->i_mapping->private_data = resv_map;
863 info->seals = F_SEAL_SEAL;
864 switch (mode & S_IFMT) {
866 init_special_inode(inode, mode, dev);
869 inode->i_op = &hugetlbfs_inode_operations;
870 inode->i_fop = &hugetlbfs_file_operations;
873 inode->i_op = &hugetlbfs_dir_inode_operations;
874 inode->i_fop = &simple_dir_operations;
876 /* directory inodes start off with i_nlink == 2 (for "." entry) */
880 inode->i_op = &page_symlink_inode_operations;
881 inode_nohighmem(inode);
884 lockdep_annotate_inode_mutex_key(inode);
887 kref_put(&resv_map->refs, resv_map_release);
894 * File creation. Allocate an inode, and we're done..
896 static int do_hugetlbfs_mknod(struct inode *dir,
897 struct dentry *dentry,
905 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
907 dir->i_ctime = dir->i_mtime = current_time(dir);
909 d_tmpfile(dentry, inode);
911 d_instantiate(dentry, inode);
912 dget(dentry);/* Extra count - pin the dentry in core */
919 static int hugetlbfs_mknod(struct inode *dir,
920 struct dentry *dentry, umode_t mode, dev_t dev)
922 return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
925 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
927 int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
933 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
935 return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
938 static int hugetlbfs_tmpfile(struct inode *dir,
939 struct dentry *dentry, umode_t mode)
941 return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
944 static int hugetlbfs_symlink(struct inode *dir,
945 struct dentry *dentry, const char *symname)
950 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
952 int l = strlen(symname)+1;
953 error = page_symlink(inode, symname, l);
955 d_instantiate(dentry, inode);
960 dir->i_ctime = dir->i_mtime = current_time(dir);
966 * mark the head page dirty
968 static int hugetlbfs_set_page_dirty(struct page *page)
970 struct page *head = compound_head(page);
976 static int hugetlbfs_migrate_page(struct address_space *mapping,
977 struct page *newpage, struct page *page,
978 enum migrate_mode mode)
982 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
983 if (rc != MIGRATEPAGE_SUCCESS)
987 * page_private is subpool pointer in hugetlb pages. Transfer to
988 * new page. PagePrivate is not associated with page_private for
989 * hugetlb pages and can not be set here as only page_huge_active
990 * pages can be migrated.
992 if (page_private(page)) {
993 set_page_private(newpage, page_private(page));
994 set_page_private(page, 0);
997 if (mode != MIGRATE_SYNC_NO_COPY)
998 migrate_page_copy(newpage, page);
1000 migrate_page_states(newpage, page);
1002 return MIGRATEPAGE_SUCCESS;
1005 static int hugetlbfs_error_remove_page(struct address_space *mapping,
1008 struct inode *inode = mapping->host;
1009 pgoff_t index = page->index;
1011 remove_huge_page(page);
1012 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
1013 hugetlb_fix_reserve_counts(inode);
1019 * Display the mount options in /proc/mounts.
1021 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1023 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1024 struct hugepage_subpool *spool = sbinfo->spool;
1025 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1026 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1029 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1030 seq_printf(m, ",uid=%u",
1031 from_kuid_munged(&init_user_ns, sbinfo->uid));
1032 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1033 seq_printf(m, ",gid=%u",
1034 from_kgid_munged(&init_user_ns, sbinfo->gid));
1035 if (sbinfo->mode != 0755)
1036 seq_printf(m, ",mode=%o", sbinfo->mode);
1037 if (sbinfo->max_inodes != -1)
1038 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1042 if (hpage_size >= 1024) {
1046 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1048 if (spool->max_hpages != -1)
1049 seq_printf(m, ",size=%llu",
1050 (unsigned long long)spool->max_hpages << hpage_shift);
1051 if (spool->min_hpages != -1)
1052 seq_printf(m, ",min_size=%llu",
1053 (unsigned long long)spool->min_hpages << hpage_shift);
1058 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1060 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1061 struct hstate *h = hstate_inode(d_inode(dentry));
1063 buf->f_type = HUGETLBFS_MAGIC;
1064 buf->f_bsize = huge_page_size(h);
1066 spin_lock(&sbinfo->stat_lock);
1067 /* If no limits set, just report 0 for max/free/used
1068 * blocks, like simple_statfs() */
1069 if (sbinfo->spool) {
1072 spin_lock(&sbinfo->spool->lock);
1073 buf->f_blocks = sbinfo->spool->max_hpages;
1074 free_pages = sbinfo->spool->max_hpages
1075 - sbinfo->spool->used_hpages;
1076 buf->f_bavail = buf->f_bfree = free_pages;
1077 spin_unlock(&sbinfo->spool->lock);
1078 buf->f_files = sbinfo->max_inodes;
1079 buf->f_ffree = sbinfo->free_inodes;
1081 spin_unlock(&sbinfo->stat_lock);
1083 buf->f_namelen = NAME_MAX;
1087 static void hugetlbfs_put_super(struct super_block *sb)
1089 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1092 sb->s_fs_info = NULL;
1095 hugepage_put_subpool(sbi->spool);
1101 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1103 if (sbinfo->free_inodes >= 0) {
1104 spin_lock(&sbinfo->stat_lock);
1105 if (unlikely(!sbinfo->free_inodes)) {
1106 spin_unlock(&sbinfo->stat_lock);
1109 sbinfo->free_inodes--;
1110 spin_unlock(&sbinfo->stat_lock);
1116 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1118 if (sbinfo->free_inodes >= 0) {
1119 spin_lock(&sbinfo->stat_lock);
1120 sbinfo->free_inodes++;
1121 spin_unlock(&sbinfo->stat_lock);
1126 static struct kmem_cache *hugetlbfs_inode_cachep;
1128 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1130 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1131 struct hugetlbfs_inode_info *p;
1133 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1135 p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1137 hugetlbfs_inc_free_inodes(sbinfo);
1142 * Any time after allocation, hugetlbfs_destroy_inode can be called
1143 * for the inode. mpol_free_shared_policy is unconditionally called
1144 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1145 * in case of a quick call to destroy.
1147 * Note that the policy is initialized even if we are creating a
1148 * private inode. This simplifies hugetlbfs_destroy_inode.
1150 mpol_shared_policy_init(&p->policy, NULL);
1152 return &p->vfs_inode;
1155 static void hugetlbfs_free_inode(struct inode *inode)
1157 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1160 static void hugetlbfs_destroy_inode(struct inode *inode)
1162 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1163 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1166 static const struct address_space_operations hugetlbfs_aops = {
1167 .write_begin = hugetlbfs_write_begin,
1168 .write_end = hugetlbfs_write_end,
1169 .set_page_dirty = hugetlbfs_set_page_dirty,
1170 .migratepage = hugetlbfs_migrate_page,
1171 .error_remove_page = hugetlbfs_error_remove_page,
1175 static void init_once(void *foo)
1177 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1179 inode_init_once(&ei->vfs_inode);
1182 const struct file_operations hugetlbfs_file_operations = {
1183 .read_iter = hugetlbfs_read_iter,
1184 .mmap = hugetlbfs_file_mmap,
1185 .fsync = noop_fsync,
1186 .get_unmapped_area = hugetlb_get_unmapped_area,
1187 .llseek = default_llseek,
1188 .fallocate = hugetlbfs_fallocate,
1191 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1192 .create = hugetlbfs_create,
1193 .lookup = simple_lookup,
1194 .link = simple_link,
1195 .unlink = simple_unlink,
1196 .symlink = hugetlbfs_symlink,
1197 .mkdir = hugetlbfs_mkdir,
1198 .rmdir = simple_rmdir,
1199 .mknod = hugetlbfs_mknod,
1200 .rename = simple_rename,
1201 .setattr = hugetlbfs_setattr,
1202 .tmpfile = hugetlbfs_tmpfile,
1205 static const struct inode_operations hugetlbfs_inode_operations = {
1206 .setattr = hugetlbfs_setattr,
1209 static const struct super_operations hugetlbfs_ops = {
1210 .alloc_inode = hugetlbfs_alloc_inode,
1211 .free_inode = hugetlbfs_free_inode,
1212 .destroy_inode = hugetlbfs_destroy_inode,
1213 .evict_inode = hugetlbfs_evict_inode,
1214 .statfs = hugetlbfs_statfs,
1215 .put_super = hugetlbfs_put_super,
1216 .show_options = hugetlbfs_show_options,
1220 * Convert size option passed from command line to number of huge pages
1221 * in the pool specified by hstate. Size option could be in bytes
1222 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1225 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1226 enum hugetlbfs_size_type val_type)
1228 if (val_type == NO_SIZE)
1231 if (val_type == SIZE_PERCENT) {
1232 size_opt <<= huge_page_shift(h);
1233 size_opt *= h->max_huge_pages;
1234 do_div(size_opt, 100);
1237 size_opt >>= huge_page_shift(h);
1242 * Parse one mount parameter.
1244 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1246 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1247 struct fs_parse_result result;
1253 opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1259 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1260 if (!uid_valid(ctx->uid))
1265 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1266 if (!gid_valid(ctx->gid))
1271 ctx->mode = result.uint_32 & 01777U;
1275 /* memparse() will accept a K/M/G without a digit */
1276 if (!param->string || !isdigit(param->string[0]))
1278 ctx->max_size_opt = memparse(param->string, &rest);
1279 ctx->max_val_type = SIZE_STD;
1281 ctx->max_val_type = SIZE_PERCENT;
1285 /* memparse() will accept a K/M/G without a digit */
1286 if (!param->string || !isdigit(param->string[0]))
1288 ctx->nr_inodes = memparse(param->string, &rest);
1292 ps = memparse(param->string, &rest);
1293 h = size_to_hstate(ps);
1295 pr_err("Unsupported page size %lu MB\n", ps >> 20);
1302 /* memparse() will accept a K/M/G without a digit */
1303 if (!param->string || !isdigit(param->string[0]))
1305 ctx->min_size_opt = memparse(param->string, &rest);
1306 ctx->min_val_type = SIZE_STD;
1308 ctx->min_val_type = SIZE_PERCENT;
1316 return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1317 param->string, param->key);
1321 * Validate the parsed options.
1323 static int hugetlbfs_validate(struct fs_context *fc)
1325 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1328 * Use huge page pool size (in hstate) to convert the size
1329 * options to number of huge pages. If NO_SIZE, -1 is returned.
1331 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1334 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1339 * If max_size was specified, then min_size must be smaller
1341 if (ctx->max_val_type > NO_SIZE &&
1342 ctx->min_hpages > ctx->max_hpages) {
1343 pr_err("Minimum size can not be greater than maximum size\n");
1351 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1353 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1354 struct hugetlbfs_sb_info *sbinfo;
1356 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1359 sb->s_fs_info = sbinfo;
1360 spin_lock_init(&sbinfo->stat_lock);
1361 sbinfo->hstate = ctx->hstate;
1362 sbinfo->max_inodes = ctx->nr_inodes;
1363 sbinfo->free_inodes = ctx->nr_inodes;
1364 sbinfo->spool = NULL;
1365 sbinfo->uid = ctx->uid;
1366 sbinfo->gid = ctx->gid;
1367 sbinfo->mode = ctx->mode;
1370 * Allocate and initialize subpool if maximum or minimum size is
1371 * specified. Any needed reservations (for minimim size) are taken
1372 * taken when the subpool is created.
1374 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1375 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1381 sb->s_maxbytes = MAX_LFS_FILESIZE;
1382 sb->s_blocksize = huge_page_size(ctx->hstate);
1383 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1384 sb->s_magic = HUGETLBFS_MAGIC;
1385 sb->s_op = &hugetlbfs_ops;
1386 sb->s_time_gran = 1;
1389 * Due to the special and limited functionality of hugetlbfs, it does
1390 * not work well as a stacking filesystem.
1392 sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1393 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1398 kfree(sbinfo->spool);
1403 static int hugetlbfs_get_tree(struct fs_context *fc)
1405 int err = hugetlbfs_validate(fc);
1408 return get_tree_nodev(fc, hugetlbfs_fill_super);
1411 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1413 kfree(fc->fs_private);
1416 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1417 .free = hugetlbfs_fs_context_free,
1418 .parse_param = hugetlbfs_parse_param,
1419 .get_tree = hugetlbfs_get_tree,
1422 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1424 struct hugetlbfs_fs_context *ctx;
1426 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1430 ctx->max_hpages = -1; /* No limit on size by default */
1431 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1432 ctx->uid = current_fsuid();
1433 ctx->gid = current_fsgid();
1435 ctx->hstate = &default_hstate;
1436 ctx->min_hpages = -1; /* No default minimum size */
1437 ctx->max_val_type = NO_SIZE;
1438 ctx->min_val_type = NO_SIZE;
1439 fc->fs_private = ctx;
1440 fc->ops = &hugetlbfs_fs_context_ops;
1444 static struct file_system_type hugetlbfs_fs_type = {
1445 .name = "hugetlbfs",
1446 .init_fs_context = hugetlbfs_init_fs_context,
1447 .parameters = hugetlb_fs_parameters,
1448 .kill_sb = kill_litter_super,
1451 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1453 static int can_do_hugetlb_shm(void)
1456 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1457 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1460 static int get_hstate_idx(int page_size_log)
1462 struct hstate *h = hstate_sizelog(page_size_log);
1470 * Note that size should be aligned to proper hugepage size in caller side,
1471 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1473 struct file *hugetlb_file_setup(const char *name, size_t size,
1474 vm_flags_t acctflag, struct user_struct **user,
1475 int creat_flags, int page_size_log)
1477 struct inode *inode;
1478 struct vfsmount *mnt;
1482 hstate_idx = get_hstate_idx(page_size_log);
1484 return ERR_PTR(-ENODEV);
1487 mnt = hugetlbfs_vfsmount[hstate_idx];
1489 return ERR_PTR(-ENOENT);
1491 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1492 *user = current_user();
1493 if (user_shm_lock(size, *user)) {
1495 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1496 current->comm, current->pid);
1497 task_unlock(current);
1500 return ERR_PTR(-EPERM);
1504 file = ERR_PTR(-ENOSPC);
1505 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1508 if (creat_flags == HUGETLB_SHMFS_INODE)
1509 inode->i_flags |= S_PRIVATE;
1511 inode->i_size = size;
1514 if (!hugetlb_reserve_pages(inode, 0,
1515 size >> huge_page_shift(hstate_inode(inode)), NULL,
1517 file = ERR_PTR(-ENOMEM);
1519 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1520 &hugetlbfs_file_operations);
1527 user_shm_unlock(size, *user);
1533 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1535 struct fs_context *fc;
1536 struct vfsmount *mnt;
1538 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1542 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1548 pr_err("Cannot mount internal hugetlbfs for page size %uK",
1549 1U << (h->order + PAGE_SHIFT - 10));
1553 static int __init init_hugetlbfs_fs(void)
1555 struct vfsmount *mnt;
1560 if (!hugepages_supported()) {
1561 pr_info("disabling because there are no supported hugepage sizes\n");
1566 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1567 sizeof(struct hugetlbfs_inode_info),
1568 0, SLAB_ACCOUNT, init_once);
1569 if (hugetlbfs_inode_cachep == NULL)
1572 error = register_filesystem(&hugetlbfs_fs_type);
1576 /* default hstate mount is required */
1577 mnt = mount_one_hugetlbfs(&hstates[default_hstate_idx]);
1579 error = PTR_ERR(mnt);
1582 hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1584 /* other hstates are optional */
1586 for_each_hstate(h) {
1587 if (i == default_hstate_idx) {
1592 mnt = mount_one_hugetlbfs(h);
1594 hugetlbfs_vfsmount[i] = NULL;
1596 hugetlbfs_vfsmount[i] = mnt;
1603 (void)unregister_filesystem(&hugetlbfs_fs_type);
1605 kmem_cache_destroy(hugetlbfs_inode_cachep);
1609 fs_initcall(init_hugetlbfs_fs)