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);
140 * vma address alignment (but not the pgoff alignment) has
141 * already been checked by prepare_hugepage_range. If you add
142 * any error returns here, do so after setting VM_HUGETLB, so
143 * is_vm_hugetlb_page tests below unmap_region go the right
144 * way when do_mmap unwinds (may be important on powerpc
147 vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
148 vma->vm_ops = &hugetlb_vm_ops;
150 ret = seal_check_future_write(info->seals, vma);
155 * page based offset in vm_pgoff could be sufficiently large to
156 * overflow a loff_t when converted to byte offset. This can
157 * only happen on architectures where sizeof(loff_t) ==
158 * sizeof(unsigned long). So, only check in those instances.
160 if (sizeof(unsigned long) == sizeof(loff_t)) {
161 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
165 /* must be huge page aligned */
166 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
169 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
170 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
171 /* check for overflow */
179 if (!hugetlb_reserve_pages(inode,
180 vma->vm_pgoff >> huge_page_order(h),
181 len >> huge_page_shift(h), vma,
186 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
187 i_size_write(inode, len);
195 * Called under mmap_write_lock(mm).
199 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
200 unsigned long len, unsigned long pgoff, unsigned long flags)
202 struct hstate *h = hstate_file(file);
203 struct vm_unmapped_area_info info;
207 info.low_limit = current->mm->mmap_base;
208 info.high_limit = arch_get_mmap_end(addr, len, flags);
209 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
210 info.align_offset = 0;
211 return vm_unmapped_area(&info);
215 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
216 unsigned long len, unsigned long pgoff, unsigned long flags)
218 struct hstate *h = hstate_file(file);
219 struct vm_unmapped_area_info info;
221 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
223 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
224 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
225 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
226 info.align_offset = 0;
227 addr = vm_unmapped_area(&info);
230 * A failed mmap() very likely causes application failure,
231 * so fall back to the bottom-up function here. This scenario
232 * can happen with large stack limits and large mmap()
235 if (unlikely(offset_in_page(addr))) {
236 VM_BUG_ON(addr != -ENOMEM);
238 info.low_limit = current->mm->mmap_base;
239 info.high_limit = arch_get_mmap_end(addr, len, flags);
240 addr = vm_unmapped_area(&info);
247 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
248 unsigned long len, unsigned long pgoff,
251 struct mm_struct *mm = current->mm;
252 struct vm_area_struct *vma;
253 struct hstate *h = hstate_file(file);
254 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
256 if (len & ~huge_page_mask(h))
261 if (flags & MAP_FIXED) {
262 if (prepare_hugepage_range(file, addr, len))
268 addr = ALIGN(addr, huge_page_size(h));
269 vma = find_vma(mm, addr);
270 if (mmap_end - len >= addr &&
271 (!vma || addr + len <= vm_start_gap(vma)))
276 * Use mm->get_unmapped_area value as a hint to use topdown routine.
277 * If architectures have special needs, they should define their own
278 * version of hugetlb_get_unmapped_area.
280 if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
281 return hugetlb_get_unmapped_area_topdown(file, addr, len,
283 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
287 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
289 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
290 unsigned long len, unsigned long pgoff,
293 return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
298 hugetlbfs_read_actor(struct page *page, unsigned long offset,
299 struct iov_iter *to, unsigned long size)
304 /* Find which 4k chunk and offset with in that chunk */
305 i = offset >> PAGE_SHIFT;
306 offset = offset & ~PAGE_MASK;
310 chunksize = PAGE_SIZE;
313 if (chunksize > size)
315 n = copy_page_to_iter(&page[i], offset, chunksize, to);
327 * Support for read() - Find the page attached to f_mapping and copy out the
328 * data. Its *very* similar to generic_file_buffered_read(), we can't use that
329 * since it has PAGE_SIZE assumptions.
331 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
333 struct file *file = iocb->ki_filp;
334 struct hstate *h = hstate_file(file);
335 struct address_space *mapping = file->f_mapping;
336 struct inode *inode = mapping->host;
337 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
338 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
339 unsigned long end_index;
343 while (iov_iter_count(to)) {
347 /* nr is the maximum number of bytes to copy from this page */
348 nr = huge_page_size(h);
349 isize = i_size_read(inode);
352 end_index = (isize - 1) >> huge_page_shift(h);
353 if (index > end_index)
355 if (index == end_index) {
356 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
363 page = find_lock_page(mapping, index);
364 if (unlikely(page == NULL)) {
366 * We have a HOLE, zero out the user-buffer for the
367 * length of the hole or request.
369 copied = iov_iter_zero(nr, to);
374 * We have the page, copy it to user space buffer.
376 copied = hugetlbfs_read_actor(page, offset, to, nr);
381 if (copied != nr && iov_iter_count(to)) {
386 index += offset >> huge_page_shift(h);
387 offset &= ~huge_page_mask(h);
389 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
393 static int hugetlbfs_write_begin(struct file *file,
394 struct address_space *mapping,
395 loff_t pos, unsigned len,
396 struct page **pagep, void **fsdata)
401 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
402 loff_t pos, unsigned len, unsigned copied,
403 struct page *page, void *fsdata)
409 static void remove_huge_page(struct page *page)
411 ClearPageDirty(page);
412 ClearPageUptodate(page);
413 delete_from_page_cache(page);
417 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
418 zap_flags_t zap_flags)
420 struct vm_area_struct *vma;
423 * end == 0 indicates that the entire range after start should be
424 * unmapped. Note, end is exclusive, whereas the interval tree takes
425 * an inclusive "last".
427 vma_interval_tree_foreach(vma, root, start, end ? end - 1 : 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/reserve
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/reserve map
470 * deleted. The region/reserve 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);
486 bool truncate_op = (lend == LLONG_MAX);
492 * When no more pages are found, we are done.
494 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
497 for (i = 0; i < pagevec_count(&pvec); ++i) {
498 struct page *page = pvec.pages[i];
504 * Only need to hold the fault mutex in the
505 * hole punch case. This prevents races with
506 * page faults. Races are not possible in the
507 * case of truncation.
509 hash = hugetlb_fault_mutex_hash(mapping, index);
510 mutex_lock(&hugetlb_fault_mutex_table[hash]);
514 * If page is mapped, it was faulted in after being
515 * unmapped in caller. Unmap (again) now after taking
516 * the fault mutex. The mutex will prevent faults
517 * until we finish removing the page.
519 * This race can only happen in the hole punch case.
520 * Getting here in a truncate operation is a bug.
522 if (unlikely(page_mapped(page))) {
525 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
526 i_mmap_lock_write(mapping);
527 mutex_lock(&hugetlb_fault_mutex_table[hash]);
528 hugetlb_vmdelete_list(&mapping->i_mmap,
529 index * pages_per_huge_page(h),
530 (index + 1) * pages_per_huge_page(h),
531 ZAP_FLAG_DROP_MARKER);
532 i_mmap_unlock_write(mapping);
537 * We must free the huge page and remove from page
538 * cache (remove_huge_page) BEFORE removing the
539 * region/reserve map (hugetlb_unreserve_pages). In
540 * rare out of memory conditions, removal of the
541 * region/reserve map could fail. Correspondingly,
542 * the subpool and global reserve usage count can need
545 VM_BUG_ON(HPageRestoreReserve(page));
546 remove_huge_page(page);
549 if (unlikely(hugetlb_unreserve_pages(inode,
550 index, index + 1, 1)))
551 hugetlb_fix_reserve_counts(inode);
556 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
558 huge_pagevec_release(&pvec);
563 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
566 static void hugetlbfs_evict_inode(struct inode *inode)
568 struct resv_map *resv_map;
570 remove_inode_hugepages(inode, 0, LLONG_MAX);
573 * Get the resv_map from the address space embedded in the inode.
574 * This is the address space which points to any resv_map allocated
575 * at inode creation time. If this is a device special inode,
576 * i_mapping may not point to the original address space.
578 resv_map = (struct resv_map *)(&inode->i_data)->private_data;
579 /* Only regular and link inodes have associated reserve maps */
581 resv_map_release(&resv_map->refs);
585 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
588 struct address_space *mapping = inode->i_mapping;
589 struct hstate *h = hstate_inode(inode);
591 BUG_ON(offset & ~huge_page_mask(h));
592 pgoff = offset >> PAGE_SHIFT;
594 i_mmap_lock_write(mapping);
595 i_size_write(inode, offset);
596 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
597 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
598 ZAP_FLAG_DROP_MARKER);
599 i_mmap_unlock_write(mapping);
600 remove_inode_hugepages(inode, offset, LLONG_MAX);
603 static void hugetlbfs_zero_partial_page(struct hstate *h,
604 struct address_space *mapping,
608 pgoff_t idx = start >> huge_page_shift(h);
611 folio = filemap_lock_folio(mapping, idx);
615 start = start & ~huge_page_mask(h);
616 end = end & ~huge_page_mask(h);
618 end = huge_page_size(h);
620 folio_zero_segment(folio, (size_t)start, (size_t)end);
626 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
628 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
629 struct address_space *mapping = inode->i_mapping;
630 struct hstate *h = hstate_inode(inode);
631 loff_t hpage_size = huge_page_size(h);
632 loff_t hole_start, hole_end;
635 * hole_start and hole_end indicate the full pages within the hole.
637 hole_start = round_up(offset, hpage_size);
638 hole_end = round_down(offset + len, hpage_size);
642 /* protected by i_rwsem */
643 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
648 i_mmap_lock_write(mapping);
650 /* If range starts before first full page, zero partial page. */
651 if (offset < hole_start)
652 hugetlbfs_zero_partial_page(h, mapping,
653 offset, min(offset + len, hole_start));
655 /* Unmap users of full pages in the hole. */
656 if (hole_end > hole_start) {
657 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
658 hugetlb_vmdelete_list(&mapping->i_mmap,
659 hole_start >> PAGE_SHIFT,
660 hole_end >> PAGE_SHIFT, 0);
663 /* If range extends beyond last full page, zero partial page. */
664 if ((offset + len) > hole_end && (offset + len) > hole_start)
665 hugetlbfs_zero_partial_page(h, mapping,
666 hole_end, offset + len);
668 i_mmap_unlock_write(mapping);
670 /* Remove full pages from the file. */
671 if (hole_end > hole_start)
672 remove_inode_hugepages(inode, hole_start, hole_end);
679 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
682 struct inode *inode = file_inode(file);
683 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
684 struct address_space *mapping = inode->i_mapping;
685 struct hstate *h = hstate_inode(inode);
686 struct vm_area_struct pseudo_vma;
687 struct mm_struct *mm = current->mm;
688 loff_t hpage_size = huge_page_size(h);
689 unsigned long hpage_shift = huge_page_shift(h);
690 pgoff_t start, index, end;
694 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
697 if (mode & FALLOC_FL_PUNCH_HOLE)
698 return hugetlbfs_punch_hole(inode, offset, len);
701 * Default preallocate case.
702 * For this range, start is rounded down and end is rounded up
703 * as well as being converted to page offsets.
705 start = offset >> hpage_shift;
706 end = (offset + len + hpage_size - 1) >> hpage_shift;
710 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
711 error = inode_newsize_ok(inode, offset + len);
715 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
721 * Initialize a pseudo vma as this is required by the huge page
722 * allocation routines. If NUMA is configured, use page index
723 * as input to create an allocation policy.
725 vma_init(&pseudo_vma, mm);
726 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
727 pseudo_vma.vm_file = file;
729 for (index = start; index < end; index++) {
731 * This is supposed to be the vaddr where the page is being
732 * faulted in, but we have no vaddr here.
740 * fallocate(2) manpage permits EINTR; we may have been
741 * interrupted because we are using up too much memory.
743 if (signal_pending(current)) {
748 /* Set numa allocation policy based on index */
749 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
751 /* addr is the offset within the file (zero based) */
752 addr = index * hpage_size;
755 * fault mutex taken here, protects against fault path
756 * and hole punch. inode_lock previously taken protects
757 * against truncation.
759 hash = hugetlb_fault_mutex_hash(mapping, index);
760 mutex_lock(&hugetlb_fault_mutex_table[hash]);
762 /* See if already present in mapping to avoid alloc/free */
763 page = find_get_page(mapping, index);
766 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
767 hugetlb_drop_vma_policy(&pseudo_vma);
772 * Allocate page without setting the avoid_reserve argument.
773 * There certainly are no reserves associated with the
774 * pseudo_vma. However, there could be shared mappings with
775 * reserves for the file at the inode level. If we fallocate
776 * pages in these areas, we need to consume the reserves
777 * to keep reservation accounting consistent.
779 page = alloc_huge_page(&pseudo_vma, addr, 0);
780 hugetlb_drop_vma_policy(&pseudo_vma);
782 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
783 error = PTR_ERR(page);
786 clear_huge_page(page, addr, pages_per_huge_page(h));
787 __SetPageUptodate(page);
788 error = huge_add_to_page_cache(page, mapping, index);
789 if (unlikely(error)) {
790 restore_reserve_on_error(h, &pseudo_vma, addr, page);
792 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
796 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
798 SetHPageMigratable(page);
800 * unlock_page because locked by add_to_page_cache()
801 * put_page() due to reference from alloc_huge_page()
807 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
808 i_size_write(inode, offset + len);
809 inode->i_ctime = current_time(inode);
815 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
816 struct dentry *dentry, struct iattr *attr)
818 struct inode *inode = d_inode(dentry);
819 struct hstate *h = hstate_inode(inode);
821 unsigned int ia_valid = attr->ia_valid;
822 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
824 error = setattr_prepare(&init_user_ns, dentry, attr);
828 if (ia_valid & ATTR_SIZE) {
829 loff_t oldsize = inode->i_size;
830 loff_t newsize = attr->ia_size;
832 if (newsize & ~huge_page_mask(h))
834 /* protected by i_rwsem */
835 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
836 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
838 hugetlb_vmtruncate(inode, newsize);
841 setattr_copy(&init_user_ns, inode, attr);
842 mark_inode_dirty(inode);
846 static struct inode *hugetlbfs_get_root(struct super_block *sb,
847 struct hugetlbfs_fs_context *ctx)
851 inode = new_inode(sb);
853 inode->i_ino = get_next_ino();
854 inode->i_mode = S_IFDIR | ctx->mode;
855 inode->i_uid = ctx->uid;
856 inode->i_gid = ctx->gid;
857 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
858 inode->i_op = &hugetlbfs_dir_inode_operations;
859 inode->i_fop = &simple_dir_operations;
860 /* directory inodes start off with i_nlink == 2 (for "." entry) */
862 lockdep_annotate_inode_mutex_key(inode);
868 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
869 * be taken from reclaim -- unlike regular filesystems. This needs an
870 * annotation because huge_pmd_share() does an allocation under hugetlb's
873 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
875 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
877 umode_t mode, dev_t dev)
880 struct resv_map *resv_map = NULL;
883 * Reserve maps are only needed for inodes that can have associated
886 if (S_ISREG(mode) || S_ISLNK(mode)) {
887 resv_map = resv_map_alloc();
892 inode = new_inode(sb);
894 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
896 inode->i_ino = get_next_ino();
897 inode_init_owner(&init_user_ns, inode, dir, mode);
898 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
899 &hugetlbfs_i_mmap_rwsem_key);
900 inode->i_mapping->a_ops = &hugetlbfs_aops;
901 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
902 inode->i_mapping->private_data = resv_map;
903 info->seals = F_SEAL_SEAL;
904 switch (mode & S_IFMT) {
906 init_special_inode(inode, mode, dev);
909 inode->i_op = &hugetlbfs_inode_operations;
910 inode->i_fop = &hugetlbfs_file_operations;
913 inode->i_op = &hugetlbfs_dir_inode_operations;
914 inode->i_fop = &simple_dir_operations;
916 /* directory inodes start off with i_nlink == 2 (for "." entry) */
920 inode->i_op = &page_symlink_inode_operations;
921 inode_nohighmem(inode);
924 lockdep_annotate_inode_mutex_key(inode);
927 kref_put(&resv_map->refs, resv_map_release);
934 * File creation. Allocate an inode, and we're done..
936 static int do_hugetlbfs_mknod(struct inode *dir,
937 struct dentry *dentry,
945 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
947 dir->i_ctime = dir->i_mtime = current_time(dir);
949 d_tmpfile(dentry, inode);
951 d_instantiate(dentry, inode);
952 dget(dentry);/* Extra count - pin the dentry in core */
959 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
960 struct dentry *dentry, umode_t mode, dev_t dev)
962 return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
965 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
966 struct dentry *dentry, umode_t mode)
968 int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
975 static int hugetlbfs_create(struct user_namespace *mnt_userns,
976 struct inode *dir, struct dentry *dentry,
977 umode_t mode, bool excl)
979 return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
982 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
983 struct inode *dir, struct dentry *dentry,
986 return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
989 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
990 struct inode *dir, struct dentry *dentry,
996 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
998 int l = strlen(symname)+1;
999 error = page_symlink(inode, symname, l);
1001 d_instantiate(dentry, inode);
1006 dir->i_ctime = dir->i_mtime = current_time(dir);
1011 static int hugetlbfs_migrate_page(struct address_space *mapping,
1012 struct page *newpage, struct page *page,
1013 enum migrate_mode mode)
1017 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
1018 if (rc != MIGRATEPAGE_SUCCESS)
1021 if (hugetlb_page_subpool(page)) {
1022 hugetlb_set_page_subpool(newpage, hugetlb_page_subpool(page));
1023 hugetlb_set_page_subpool(page, NULL);
1026 if (mode != MIGRATE_SYNC_NO_COPY)
1027 migrate_page_copy(newpage, page);
1029 migrate_page_states(newpage, page);
1031 return MIGRATEPAGE_SUCCESS;
1034 static int hugetlbfs_error_remove_page(struct address_space *mapping,
1037 struct inode *inode = mapping->host;
1038 pgoff_t index = page->index;
1040 remove_huge_page(page);
1041 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
1042 hugetlb_fix_reserve_counts(inode);
1048 * Display the mount options in /proc/mounts.
1050 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1052 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1053 struct hugepage_subpool *spool = sbinfo->spool;
1054 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1055 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1058 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1059 seq_printf(m, ",uid=%u",
1060 from_kuid_munged(&init_user_ns, sbinfo->uid));
1061 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1062 seq_printf(m, ",gid=%u",
1063 from_kgid_munged(&init_user_ns, sbinfo->gid));
1064 if (sbinfo->mode != 0755)
1065 seq_printf(m, ",mode=%o", sbinfo->mode);
1066 if (sbinfo->max_inodes != -1)
1067 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1071 if (hpage_size >= 1024) {
1075 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1077 if (spool->max_hpages != -1)
1078 seq_printf(m, ",size=%llu",
1079 (unsigned long long)spool->max_hpages << hpage_shift);
1080 if (spool->min_hpages != -1)
1081 seq_printf(m, ",min_size=%llu",
1082 (unsigned long long)spool->min_hpages << hpage_shift);
1087 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1089 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1090 struct hstate *h = hstate_inode(d_inode(dentry));
1092 buf->f_type = HUGETLBFS_MAGIC;
1093 buf->f_bsize = huge_page_size(h);
1095 spin_lock(&sbinfo->stat_lock);
1096 /* If no limits set, just report 0 for max/free/used
1097 * blocks, like simple_statfs() */
1098 if (sbinfo->spool) {
1101 spin_lock_irq(&sbinfo->spool->lock);
1102 buf->f_blocks = sbinfo->spool->max_hpages;
1103 free_pages = sbinfo->spool->max_hpages
1104 - sbinfo->spool->used_hpages;
1105 buf->f_bavail = buf->f_bfree = free_pages;
1106 spin_unlock_irq(&sbinfo->spool->lock);
1107 buf->f_files = sbinfo->max_inodes;
1108 buf->f_ffree = sbinfo->free_inodes;
1110 spin_unlock(&sbinfo->stat_lock);
1112 buf->f_namelen = NAME_MAX;
1116 static void hugetlbfs_put_super(struct super_block *sb)
1118 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1121 sb->s_fs_info = NULL;
1124 hugepage_put_subpool(sbi->spool);
1130 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1132 if (sbinfo->free_inodes >= 0) {
1133 spin_lock(&sbinfo->stat_lock);
1134 if (unlikely(!sbinfo->free_inodes)) {
1135 spin_unlock(&sbinfo->stat_lock);
1138 sbinfo->free_inodes--;
1139 spin_unlock(&sbinfo->stat_lock);
1145 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1147 if (sbinfo->free_inodes >= 0) {
1148 spin_lock(&sbinfo->stat_lock);
1149 sbinfo->free_inodes++;
1150 spin_unlock(&sbinfo->stat_lock);
1155 static struct kmem_cache *hugetlbfs_inode_cachep;
1157 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1159 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1160 struct hugetlbfs_inode_info *p;
1162 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1164 p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1166 hugetlbfs_inc_free_inodes(sbinfo);
1171 * Any time after allocation, hugetlbfs_destroy_inode can be called
1172 * for the inode. mpol_free_shared_policy is unconditionally called
1173 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1174 * in case of a quick call to destroy.
1176 * Note that the policy is initialized even if we are creating a
1177 * private inode. This simplifies hugetlbfs_destroy_inode.
1179 mpol_shared_policy_init(&p->policy, NULL);
1181 return &p->vfs_inode;
1184 static void hugetlbfs_free_inode(struct inode *inode)
1186 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1189 static void hugetlbfs_destroy_inode(struct inode *inode)
1191 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1192 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1195 static const struct address_space_operations hugetlbfs_aops = {
1196 .write_begin = hugetlbfs_write_begin,
1197 .write_end = hugetlbfs_write_end,
1198 .dirty_folio = noop_dirty_folio,
1199 .migratepage = hugetlbfs_migrate_page,
1200 .error_remove_page = hugetlbfs_error_remove_page,
1204 static void init_once(void *foo)
1206 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1208 inode_init_once(&ei->vfs_inode);
1211 const struct file_operations hugetlbfs_file_operations = {
1212 .read_iter = hugetlbfs_read_iter,
1213 .mmap = hugetlbfs_file_mmap,
1214 .fsync = noop_fsync,
1215 .get_unmapped_area = hugetlb_get_unmapped_area,
1216 .llseek = default_llseek,
1217 .fallocate = hugetlbfs_fallocate,
1220 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1221 .create = hugetlbfs_create,
1222 .lookup = simple_lookup,
1223 .link = simple_link,
1224 .unlink = simple_unlink,
1225 .symlink = hugetlbfs_symlink,
1226 .mkdir = hugetlbfs_mkdir,
1227 .rmdir = simple_rmdir,
1228 .mknod = hugetlbfs_mknod,
1229 .rename = simple_rename,
1230 .setattr = hugetlbfs_setattr,
1231 .tmpfile = hugetlbfs_tmpfile,
1234 static const struct inode_operations hugetlbfs_inode_operations = {
1235 .setattr = hugetlbfs_setattr,
1238 static const struct super_operations hugetlbfs_ops = {
1239 .alloc_inode = hugetlbfs_alloc_inode,
1240 .free_inode = hugetlbfs_free_inode,
1241 .destroy_inode = hugetlbfs_destroy_inode,
1242 .evict_inode = hugetlbfs_evict_inode,
1243 .statfs = hugetlbfs_statfs,
1244 .put_super = hugetlbfs_put_super,
1245 .show_options = hugetlbfs_show_options,
1249 * Convert size option passed from command line to number of huge pages
1250 * in the pool specified by hstate. Size option could be in bytes
1251 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1254 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1255 enum hugetlbfs_size_type val_type)
1257 if (val_type == NO_SIZE)
1260 if (val_type == SIZE_PERCENT) {
1261 size_opt <<= huge_page_shift(h);
1262 size_opt *= h->max_huge_pages;
1263 do_div(size_opt, 100);
1266 size_opt >>= huge_page_shift(h);
1271 * Parse one mount parameter.
1273 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1275 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1276 struct fs_parse_result result;
1281 opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1287 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1288 if (!uid_valid(ctx->uid))
1293 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1294 if (!gid_valid(ctx->gid))
1299 ctx->mode = result.uint_32 & 01777U;
1303 /* memparse() will accept a K/M/G without a digit */
1304 if (!isdigit(param->string[0]))
1306 ctx->max_size_opt = memparse(param->string, &rest);
1307 ctx->max_val_type = SIZE_STD;
1309 ctx->max_val_type = SIZE_PERCENT;
1313 /* memparse() will accept a K/M/G without a digit */
1314 if (!isdigit(param->string[0]))
1316 ctx->nr_inodes = memparse(param->string, &rest);
1320 ps = memparse(param->string, &rest);
1321 ctx->hstate = size_to_hstate(ps);
1323 pr_err("Unsupported page size %lu MB\n", ps >> 20);
1329 /* memparse() will accept a K/M/G without a digit */
1330 if (!isdigit(param->string[0]))
1332 ctx->min_size_opt = memparse(param->string, &rest);
1333 ctx->min_val_type = SIZE_STD;
1335 ctx->min_val_type = SIZE_PERCENT;
1343 return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1344 param->string, param->key);
1348 * Validate the parsed options.
1350 static int hugetlbfs_validate(struct fs_context *fc)
1352 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1355 * Use huge page pool size (in hstate) to convert the size
1356 * options to number of huge pages. If NO_SIZE, -1 is returned.
1358 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1361 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1366 * If max_size was specified, then min_size must be smaller
1368 if (ctx->max_val_type > NO_SIZE &&
1369 ctx->min_hpages > ctx->max_hpages) {
1370 pr_err("Minimum size can not be greater than maximum size\n");
1378 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1380 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1381 struct hugetlbfs_sb_info *sbinfo;
1383 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1386 sb->s_fs_info = sbinfo;
1387 spin_lock_init(&sbinfo->stat_lock);
1388 sbinfo->hstate = ctx->hstate;
1389 sbinfo->max_inodes = ctx->nr_inodes;
1390 sbinfo->free_inodes = ctx->nr_inodes;
1391 sbinfo->spool = NULL;
1392 sbinfo->uid = ctx->uid;
1393 sbinfo->gid = ctx->gid;
1394 sbinfo->mode = ctx->mode;
1397 * Allocate and initialize subpool if maximum or minimum size is
1398 * specified. Any needed reservations (for minimum size) are taken
1399 * taken when the subpool is created.
1401 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1402 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1408 sb->s_maxbytes = MAX_LFS_FILESIZE;
1409 sb->s_blocksize = huge_page_size(ctx->hstate);
1410 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1411 sb->s_magic = HUGETLBFS_MAGIC;
1412 sb->s_op = &hugetlbfs_ops;
1413 sb->s_time_gran = 1;
1416 * Due to the special and limited functionality of hugetlbfs, it does
1417 * not work well as a stacking filesystem.
1419 sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1420 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1425 kfree(sbinfo->spool);
1430 static int hugetlbfs_get_tree(struct fs_context *fc)
1432 int err = hugetlbfs_validate(fc);
1435 return get_tree_nodev(fc, hugetlbfs_fill_super);
1438 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1440 kfree(fc->fs_private);
1443 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1444 .free = hugetlbfs_fs_context_free,
1445 .parse_param = hugetlbfs_parse_param,
1446 .get_tree = hugetlbfs_get_tree,
1449 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1451 struct hugetlbfs_fs_context *ctx;
1453 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1457 ctx->max_hpages = -1; /* No limit on size by default */
1458 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1459 ctx->uid = current_fsuid();
1460 ctx->gid = current_fsgid();
1462 ctx->hstate = &default_hstate;
1463 ctx->min_hpages = -1; /* No default minimum size */
1464 ctx->max_val_type = NO_SIZE;
1465 ctx->min_val_type = NO_SIZE;
1466 fc->fs_private = ctx;
1467 fc->ops = &hugetlbfs_fs_context_ops;
1471 static struct file_system_type hugetlbfs_fs_type = {
1472 .name = "hugetlbfs",
1473 .init_fs_context = hugetlbfs_init_fs_context,
1474 .parameters = hugetlb_fs_parameters,
1475 .kill_sb = kill_litter_super,
1478 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1480 static int can_do_hugetlb_shm(void)
1483 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1484 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1487 static int get_hstate_idx(int page_size_log)
1489 struct hstate *h = hstate_sizelog(page_size_log);
1493 return hstate_index(h);
1497 * Note that size should be aligned to proper hugepage size in caller side,
1498 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1500 struct file *hugetlb_file_setup(const char *name, size_t size,
1501 vm_flags_t acctflag, int creat_flags,
1504 struct inode *inode;
1505 struct vfsmount *mnt;
1509 hstate_idx = get_hstate_idx(page_size_log);
1511 return ERR_PTR(-ENODEV);
1513 mnt = hugetlbfs_vfsmount[hstate_idx];
1515 return ERR_PTR(-ENOENT);
1517 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1518 struct ucounts *ucounts = current_ucounts();
1520 if (user_shm_lock(size, ucounts)) {
1521 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1522 current->comm, current->pid);
1523 user_shm_unlock(size, ucounts);
1525 return ERR_PTR(-EPERM);
1528 file = ERR_PTR(-ENOSPC);
1529 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1532 if (creat_flags == HUGETLB_SHMFS_INODE)
1533 inode->i_flags |= S_PRIVATE;
1535 inode->i_size = size;
1538 if (!hugetlb_reserve_pages(inode, 0,
1539 size >> huge_page_shift(hstate_inode(inode)), NULL,
1541 file = ERR_PTR(-ENOMEM);
1543 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1544 &hugetlbfs_file_operations);
1553 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1555 struct fs_context *fc;
1556 struct vfsmount *mnt;
1558 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1562 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1568 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1569 huge_page_size(h) >> 10);
1573 static int __init init_hugetlbfs_fs(void)
1575 struct vfsmount *mnt;
1580 if (!hugepages_supported()) {
1581 pr_info("disabling because there are no supported hugepage sizes\n");
1586 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1587 sizeof(struct hugetlbfs_inode_info),
1588 0, SLAB_ACCOUNT, init_once);
1589 if (hugetlbfs_inode_cachep == NULL)
1592 error = register_filesystem(&hugetlbfs_fs_type);
1596 /* default hstate mount is required */
1597 mnt = mount_one_hugetlbfs(&default_hstate);
1599 error = PTR_ERR(mnt);
1602 hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1604 /* other hstates are optional */
1606 for_each_hstate(h) {
1607 if (i == default_hstate_idx) {
1612 mnt = mount_one_hugetlbfs(h);
1614 hugetlbfs_vfsmount[i] = NULL;
1616 hugetlbfs_vfsmount[i] = mnt;
1623 (void)unregister_filesystem(&hugetlbfs_fs_type);
1625 kmem_cache_destroy(hugetlbfs_inode_cachep);
1629 fs_initcall(init_hugetlbfs_fs)