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_param_specs[] = {
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),
88 static const struct fs_parameter_description hugetlb_fs_parameters = {
90 .specs = hugetlb_param_specs,
94 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
95 struct inode *inode, pgoff_t index)
97 vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
101 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
103 mpol_cond_put(vma->vm_policy);
106 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
107 struct inode *inode, pgoff_t index)
111 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
116 static void huge_pagevec_release(struct pagevec *pvec)
120 for (i = 0; i < pagevec_count(pvec); ++i)
121 put_page(pvec->pages[i]);
123 pagevec_reinit(pvec);
127 * Mask used when checking the page offset value passed in via system
128 * calls. This value will be converted to a loff_t which is signed.
129 * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
130 * value. The extra bit (- 1 in the shift value) is to take the sign
133 #define PGOFF_LOFFT_MAX \
134 (((1UL << (PAGE_SHIFT + 1)) - 1) << (BITS_PER_LONG - (PAGE_SHIFT + 1)))
136 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
138 struct inode *inode = file_inode(file);
139 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
142 struct hstate *h = hstate_file(file);
145 * vma address alignment (but not the pgoff alignment) has
146 * already been checked by prepare_hugepage_range. If you add
147 * any error returns here, do so after setting VM_HUGETLB, so
148 * is_vm_hugetlb_page tests below unmap_region go the right
149 * way when do_mmap_pgoff unwinds (may be important on powerpc
152 vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
153 vma->vm_ops = &hugetlb_vm_ops;
155 ret = seal_check_future_write(info->seals, vma);
160 * page based offset in vm_pgoff could be sufficiently large to
161 * overflow a loff_t when converted to byte offset. This can
162 * only happen on architectures where sizeof(loff_t) ==
163 * sizeof(unsigned long). So, only check in those instances.
165 if (sizeof(unsigned long) == sizeof(loff_t)) {
166 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
170 /* must be huge page aligned */
171 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
174 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
175 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
176 /* check for overflow */
184 if (hugetlb_reserve_pages(inode,
185 vma->vm_pgoff >> huge_page_order(h),
186 len >> huge_page_shift(h), vma,
191 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
192 i_size_write(inode, len);
200 * Called under down_write(mmap_sem).
203 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
205 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
206 unsigned long len, unsigned long pgoff, unsigned long flags)
208 struct hstate *h = hstate_file(file);
209 struct vm_unmapped_area_info info;
213 info.low_limit = current->mm->mmap_base;
214 info.high_limit = arch_get_mmap_end(addr);
215 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
216 info.align_offset = 0;
217 return vm_unmapped_area(&info);
221 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
222 unsigned long len, unsigned long pgoff, unsigned long flags)
224 struct hstate *h = hstate_file(file);
225 struct vm_unmapped_area_info info;
227 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
229 info.low_limit = max(PAGE_SIZE, mmap_min_addr);
230 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
231 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
232 info.align_offset = 0;
233 addr = vm_unmapped_area(&info);
236 * A failed mmap() very likely causes application failure,
237 * so fall back to the bottom-up function here. This scenario
238 * can happen with large stack limits and large mmap()
241 if (unlikely(offset_in_page(addr))) {
242 VM_BUG_ON(addr != -ENOMEM);
244 info.low_limit = current->mm->mmap_base;
245 info.high_limit = arch_get_mmap_end(addr);
246 addr = vm_unmapped_area(&info);
253 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
254 unsigned long len, unsigned long pgoff, unsigned long flags)
256 struct mm_struct *mm = current->mm;
257 struct vm_area_struct *vma;
258 struct hstate *h = hstate_file(file);
259 const unsigned long mmap_end = arch_get_mmap_end(addr);
261 if (len & ~huge_page_mask(h))
266 if (flags & MAP_FIXED) {
267 if (prepare_hugepage_range(file, addr, len))
273 addr = ALIGN(addr, huge_page_size(h));
274 vma = find_vma(mm, addr);
275 if (mmap_end - len >= addr &&
276 (!vma || addr + len <= vm_start_gap(vma)))
281 * Use mm->get_unmapped_area value as a hint to use topdown routine.
282 * If architectures have special needs, they should define their own
283 * version of hugetlb_get_unmapped_area.
285 if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
286 return hugetlb_get_unmapped_area_topdown(file, addr, len,
288 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
294 hugetlbfs_read_actor(struct page *page, unsigned long offset,
295 struct iov_iter *to, unsigned long size)
300 /* Find which 4k chunk and offset with in that chunk */
301 i = offset >> PAGE_SHIFT;
302 offset = offset & ~PAGE_MASK;
306 chunksize = PAGE_SIZE;
309 if (chunksize > size)
311 n = copy_page_to_iter(&page[i], offset, chunksize, to);
323 * Support for read() - Find the page attached to f_mapping and copy out the
324 * data. Its *very* similar to do_generic_mapping_read(), we can't use that
325 * since it has PAGE_SIZE assumptions.
327 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
329 struct file *file = iocb->ki_filp;
330 struct hstate *h = hstate_file(file);
331 struct address_space *mapping = file->f_mapping;
332 struct inode *inode = mapping->host;
333 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
334 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
335 unsigned long end_index;
339 while (iov_iter_count(to)) {
343 /* nr is the maximum number of bytes to copy from this page */
344 nr = huge_page_size(h);
345 isize = i_size_read(inode);
348 end_index = (isize - 1) >> huge_page_shift(h);
349 if (index > end_index)
351 if (index == end_index) {
352 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
359 page = find_lock_page(mapping, index);
360 if (unlikely(page == NULL)) {
362 * We have a HOLE, zero out the user-buffer for the
363 * length of the hole or request.
365 copied = iov_iter_zero(nr, to);
370 * We have the page, copy it to user space buffer.
372 copied = hugetlbfs_read_actor(page, offset, to, nr);
377 if (copied != nr && iov_iter_count(to)) {
382 index += offset >> huge_page_shift(h);
383 offset &= ~huge_page_mask(h);
385 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
389 static int hugetlbfs_write_begin(struct file *file,
390 struct address_space *mapping,
391 loff_t pos, unsigned len, unsigned flags,
392 struct page **pagep, void **fsdata)
397 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
398 loff_t pos, unsigned len, unsigned copied,
399 struct page *page, void *fsdata)
405 static void remove_huge_page(struct page *page)
407 ClearPageDirty(page);
408 ClearPageUptodate(page);
409 delete_from_page_cache(page);
413 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
415 struct vm_area_struct *vma;
418 * end == 0 indicates that the entire range after
419 * start should be unmapped.
421 vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
422 unsigned long v_offset;
426 * Can the expression below overflow on 32-bit arches?
427 * No, because the interval tree returns us only those vmas
428 * which overlap the truncated area starting at pgoff,
429 * and no vma on a 32-bit arch can span beyond the 4GB.
431 if (vma->vm_pgoff < start)
432 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
439 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
441 if (v_end > vma->vm_end)
445 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
451 * remove_inode_hugepages handles two distinct cases: truncation and hole
452 * punch. There are subtle differences in operation for each case.
454 * truncation is indicated by end of range being LLONG_MAX
455 * In this case, we first scan the range and release found pages.
456 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
457 * maps and global counts. Page faults can not race with truncation
458 * in this routine. hugetlb_no_page() prevents page faults in the
459 * truncated range. It checks i_size before allocation, and again after
460 * with the page table lock for the page held. The same lock must be
461 * acquired to unmap a page.
462 * hole punch is indicated if end is not LLONG_MAX
463 * In the hole punch case we scan the range and release found pages.
464 * Only when releasing a page is the associated region/reserv map
465 * deleted. The region/reserv map for ranges without associated
466 * pages are not modified. Page faults can race with hole punch.
467 * This is indicated if we find a mapped page.
468 * Note: If the passed end of range value is beyond the end of file, but
469 * not LLONG_MAX this routine still performs a hole punch operation.
471 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
474 struct hstate *h = hstate_inode(inode);
475 struct address_space *mapping = &inode->i_data;
476 const pgoff_t start = lstart >> huge_page_shift(h);
477 const pgoff_t end = lend >> huge_page_shift(h);
478 struct vm_area_struct pseudo_vma;
482 bool truncate_op = (lend == LLONG_MAX);
484 vma_init(&pseudo_vma, current->mm);
485 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
490 * When no more pages are found, we are done.
492 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
495 for (i = 0; i < pagevec_count(&pvec); ++i) {
496 struct page *page = pvec.pages[i];
500 hash = hugetlb_fault_mutex_hash(h, mapping, index);
501 mutex_lock(&hugetlb_fault_mutex_table[hash]);
504 * If page is mapped, it was faulted in after being
505 * unmapped in caller. Unmap (again) now after taking
506 * the fault mutex. The mutex will prevent faults
507 * until we finish removing the page.
509 * This race can only happen in the hole punch case.
510 * Getting here in a truncate operation is a bug.
512 if (unlikely(page_mapped(page))) {
515 i_mmap_lock_write(mapping);
516 hugetlb_vmdelete_list(&mapping->i_mmap,
517 index * pages_per_huge_page(h),
518 (index + 1) * pages_per_huge_page(h));
519 i_mmap_unlock_write(mapping);
524 * We must free the huge page and remove from page
525 * cache (remove_huge_page) BEFORE removing the
526 * region/reserve map (hugetlb_unreserve_pages). In
527 * rare out of memory conditions, removal of the
528 * region/reserve map could fail. Correspondingly,
529 * the subpool and global reserve usage count can need
532 VM_BUG_ON(PagePrivate(page));
533 remove_huge_page(page);
536 if (unlikely(hugetlb_unreserve_pages(inode,
537 index, index + 1, 1)))
538 hugetlb_fix_reserve_counts(inode);
542 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
544 huge_pagevec_release(&pvec);
549 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
552 static void hugetlbfs_evict_inode(struct inode *inode)
554 struct resv_map *resv_map;
556 remove_inode_hugepages(inode, 0, LLONG_MAX);
559 * Get the resv_map from the address space embedded in the inode.
560 * This is the address space which points to any resv_map allocated
561 * at inode creation time. If this is a device special inode,
562 * i_mapping may not point to the original address space.
564 resv_map = (struct resv_map *)(&inode->i_data)->private_data;
565 /* Only regular and link inodes have associated reserve maps */
567 resv_map_release(&resv_map->refs);
571 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
574 struct address_space *mapping = inode->i_mapping;
575 struct hstate *h = hstate_inode(inode);
577 BUG_ON(offset & ~huge_page_mask(h));
578 pgoff = offset >> PAGE_SHIFT;
580 i_size_write(inode, offset);
581 i_mmap_lock_write(mapping);
582 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
583 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
584 i_mmap_unlock_write(mapping);
585 remove_inode_hugepages(inode, offset, LLONG_MAX);
589 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
591 struct hstate *h = hstate_inode(inode);
592 loff_t hpage_size = huge_page_size(h);
593 loff_t hole_start, hole_end;
596 * For hole punch round up the beginning offset of the hole and
597 * round down the end.
599 hole_start = round_up(offset, hpage_size);
600 hole_end = round_down(offset + len, hpage_size);
602 if (hole_end > hole_start) {
603 struct address_space *mapping = inode->i_mapping;
604 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
608 /* protected by i_mutex */
609 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
614 i_mmap_lock_write(mapping);
615 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
616 hugetlb_vmdelete_list(&mapping->i_mmap,
617 hole_start >> PAGE_SHIFT,
618 hole_end >> PAGE_SHIFT);
619 i_mmap_unlock_write(mapping);
620 remove_inode_hugepages(inode, hole_start, hole_end);
627 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
630 struct inode *inode = file_inode(file);
631 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
632 struct address_space *mapping = inode->i_mapping;
633 struct hstate *h = hstate_inode(inode);
634 struct vm_area_struct pseudo_vma;
635 struct mm_struct *mm = current->mm;
636 loff_t hpage_size = huge_page_size(h);
637 unsigned long hpage_shift = huge_page_shift(h);
638 pgoff_t start, index, end;
642 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
645 if (mode & FALLOC_FL_PUNCH_HOLE)
646 return hugetlbfs_punch_hole(inode, offset, len);
649 * Default preallocate case.
650 * For this range, start is rounded down and end is rounded up
651 * as well as being converted to page offsets.
653 start = offset >> hpage_shift;
654 end = (offset + len + hpage_size - 1) >> hpage_shift;
658 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
659 error = inode_newsize_ok(inode, offset + len);
663 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
669 * Initialize a pseudo vma as this is required by the huge page
670 * allocation routines. If NUMA is configured, use page index
671 * as input to create an allocation policy.
673 vma_init(&pseudo_vma, mm);
674 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
675 pseudo_vma.vm_file = file;
677 for (index = start; index < end; index++) {
679 * This is supposed to be the vaddr where the page is being
680 * faulted in, but we have no vaddr here.
684 int avoid_reserve = 0;
689 * fallocate(2) manpage permits EINTR; we may have been
690 * interrupted because we are using up too much memory.
692 if (signal_pending(current)) {
697 /* Set numa allocation policy based on index */
698 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
700 /* addr is the offset within the file (zero based) */
701 addr = index * hpage_size;
703 /* mutex taken here, fault path and hole punch */
704 hash = hugetlb_fault_mutex_hash(h, mapping, index);
705 mutex_lock(&hugetlb_fault_mutex_table[hash]);
707 /* See if already present in mapping to avoid alloc/free */
708 page = find_get_page(mapping, index);
711 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
712 hugetlb_drop_vma_policy(&pseudo_vma);
716 /* Allocate page and add to page cache */
717 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
718 hugetlb_drop_vma_policy(&pseudo_vma);
720 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
721 error = PTR_ERR(page);
724 clear_huge_page(page, addr, pages_per_huge_page(h));
725 __SetPageUptodate(page);
726 error = huge_add_to_page_cache(page, mapping, index);
727 if (unlikely(error)) {
729 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
733 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
735 set_page_huge_active(page);
737 * unlock_page because locked by add_to_page_cache()
738 * put_page() due to reference from alloc_huge_page()
744 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
745 i_size_write(inode, offset + len);
746 inode->i_ctime = current_time(inode);
752 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
754 struct inode *inode = d_inode(dentry);
755 struct hstate *h = hstate_inode(inode);
757 unsigned int ia_valid = attr->ia_valid;
758 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
762 error = setattr_prepare(dentry, attr);
766 if (ia_valid & ATTR_SIZE) {
767 loff_t oldsize = inode->i_size;
768 loff_t newsize = attr->ia_size;
770 if (newsize & ~huge_page_mask(h))
772 /* protected by i_mutex */
773 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
774 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
776 error = hugetlb_vmtruncate(inode, newsize);
781 setattr_copy(inode, attr);
782 mark_inode_dirty(inode);
786 static struct inode *hugetlbfs_get_root(struct super_block *sb,
787 struct hugetlbfs_fs_context *ctx)
791 inode = new_inode(sb);
793 inode->i_ino = get_next_ino();
794 inode->i_mode = S_IFDIR | ctx->mode;
795 inode->i_uid = ctx->uid;
796 inode->i_gid = ctx->gid;
797 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
798 inode->i_op = &hugetlbfs_dir_inode_operations;
799 inode->i_fop = &simple_dir_operations;
800 /* directory inodes start off with i_nlink == 2 (for "." entry) */
802 lockdep_annotate_inode_mutex_key(inode);
808 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
809 * be taken from reclaim -- unlike regular filesystems. This needs an
810 * annotation because huge_pmd_share() does an allocation under hugetlb's
813 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
815 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
817 umode_t mode, dev_t dev)
820 struct resv_map *resv_map = NULL;
823 * Reserve maps are only needed for inodes that can have associated
826 if (S_ISREG(mode) || S_ISLNK(mode)) {
827 resv_map = resv_map_alloc();
832 inode = new_inode(sb);
834 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
836 inode->i_ino = get_next_ino();
837 inode_init_owner(inode, dir, mode);
838 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
839 &hugetlbfs_i_mmap_rwsem_key);
840 inode->i_mapping->a_ops = &hugetlbfs_aops;
841 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
842 inode->i_mapping->private_data = resv_map;
843 info->seals = F_SEAL_SEAL;
844 switch (mode & S_IFMT) {
846 init_special_inode(inode, mode, dev);
849 inode->i_op = &hugetlbfs_inode_operations;
850 inode->i_fop = &hugetlbfs_file_operations;
853 inode->i_op = &hugetlbfs_dir_inode_operations;
854 inode->i_fop = &simple_dir_operations;
856 /* directory inodes start off with i_nlink == 2 (for "." entry) */
860 inode->i_op = &page_symlink_inode_operations;
861 inode_nohighmem(inode);
864 lockdep_annotate_inode_mutex_key(inode);
867 kref_put(&resv_map->refs, resv_map_release);
874 * File creation. Allocate an inode, and we're done..
876 static int hugetlbfs_mknod(struct inode *dir,
877 struct dentry *dentry, umode_t mode, dev_t dev)
882 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
884 dir->i_ctime = dir->i_mtime = current_time(dir);
885 d_instantiate(dentry, inode);
886 dget(dentry); /* Extra count - pin the dentry in core */
892 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
894 int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
900 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
902 return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
905 static int hugetlbfs_symlink(struct inode *dir,
906 struct dentry *dentry, const char *symname)
911 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
913 int l = strlen(symname)+1;
914 error = page_symlink(inode, symname, l);
916 d_instantiate(dentry, inode);
921 dir->i_ctime = dir->i_mtime = current_time(dir);
927 * mark the head page dirty
929 static int hugetlbfs_set_page_dirty(struct page *page)
931 struct page *head = compound_head(page);
937 static int hugetlbfs_migrate_page(struct address_space *mapping,
938 struct page *newpage, struct page *page,
939 enum migrate_mode mode)
943 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
944 if (rc != MIGRATEPAGE_SUCCESS)
948 * page_private is subpool pointer in hugetlb pages. Transfer to
949 * new page. PagePrivate is not associated with page_private for
950 * hugetlb pages and can not be set here as only page_huge_active
951 * pages can be migrated.
953 if (page_private(page)) {
954 set_page_private(newpage, page_private(page));
955 set_page_private(page, 0);
958 if (mode != MIGRATE_SYNC_NO_COPY)
959 migrate_page_copy(newpage, page);
961 migrate_page_states(newpage, page);
963 return MIGRATEPAGE_SUCCESS;
966 static int hugetlbfs_error_remove_page(struct address_space *mapping,
969 struct inode *inode = mapping->host;
970 pgoff_t index = page->index;
972 remove_huge_page(page);
973 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
974 hugetlb_fix_reserve_counts(inode);
980 * Display the mount options in /proc/mounts.
982 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
984 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
985 struct hugepage_subpool *spool = sbinfo->spool;
986 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
987 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
990 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
991 seq_printf(m, ",uid=%u",
992 from_kuid_munged(&init_user_ns, sbinfo->uid));
993 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
994 seq_printf(m, ",gid=%u",
995 from_kgid_munged(&init_user_ns, sbinfo->gid));
996 if (sbinfo->mode != 0755)
997 seq_printf(m, ",mode=%o", sbinfo->mode);
998 if (sbinfo->max_inodes != -1)
999 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1003 if (hpage_size >= 1024) {
1007 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1009 if (spool->max_hpages != -1)
1010 seq_printf(m, ",size=%llu",
1011 (unsigned long long)spool->max_hpages << hpage_shift);
1012 if (spool->min_hpages != -1)
1013 seq_printf(m, ",min_size=%llu",
1014 (unsigned long long)spool->min_hpages << hpage_shift);
1019 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1021 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1022 struct hstate *h = hstate_inode(d_inode(dentry));
1024 buf->f_type = HUGETLBFS_MAGIC;
1025 buf->f_bsize = huge_page_size(h);
1027 spin_lock(&sbinfo->stat_lock);
1028 /* If no limits set, just report 0 for max/free/used
1029 * blocks, like simple_statfs() */
1030 if (sbinfo->spool) {
1033 spin_lock(&sbinfo->spool->lock);
1034 buf->f_blocks = sbinfo->spool->max_hpages;
1035 free_pages = sbinfo->spool->max_hpages
1036 - sbinfo->spool->used_hpages;
1037 buf->f_bavail = buf->f_bfree = free_pages;
1038 spin_unlock(&sbinfo->spool->lock);
1039 buf->f_files = sbinfo->max_inodes;
1040 buf->f_ffree = sbinfo->free_inodes;
1042 spin_unlock(&sbinfo->stat_lock);
1044 buf->f_namelen = NAME_MAX;
1048 static void hugetlbfs_put_super(struct super_block *sb)
1050 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1053 sb->s_fs_info = NULL;
1056 hugepage_put_subpool(sbi->spool);
1062 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1064 if (sbinfo->free_inodes >= 0) {
1065 spin_lock(&sbinfo->stat_lock);
1066 if (unlikely(!sbinfo->free_inodes)) {
1067 spin_unlock(&sbinfo->stat_lock);
1070 sbinfo->free_inodes--;
1071 spin_unlock(&sbinfo->stat_lock);
1077 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1079 if (sbinfo->free_inodes >= 0) {
1080 spin_lock(&sbinfo->stat_lock);
1081 sbinfo->free_inodes++;
1082 spin_unlock(&sbinfo->stat_lock);
1087 static struct kmem_cache *hugetlbfs_inode_cachep;
1089 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1091 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1092 struct hugetlbfs_inode_info *p;
1094 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1096 p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1098 hugetlbfs_inc_free_inodes(sbinfo);
1103 * Any time after allocation, hugetlbfs_destroy_inode can be called
1104 * for the inode. mpol_free_shared_policy is unconditionally called
1105 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1106 * in case of a quick call to destroy.
1108 * Note that the policy is initialized even if we are creating a
1109 * private inode. This simplifies hugetlbfs_destroy_inode.
1111 mpol_shared_policy_init(&p->policy, NULL);
1113 return &p->vfs_inode;
1116 static void hugetlbfs_free_inode(struct inode *inode)
1118 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1121 static void hugetlbfs_destroy_inode(struct inode *inode)
1123 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1124 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1127 static const struct address_space_operations hugetlbfs_aops = {
1128 .write_begin = hugetlbfs_write_begin,
1129 .write_end = hugetlbfs_write_end,
1130 .set_page_dirty = hugetlbfs_set_page_dirty,
1131 .migratepage = hugetlbfs_migrate_page,
1132 .error_remove_page = hugetlbfs_error_remove_page,
1136 static void init_once(void *foo)
1138 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1140 inode_init_once(&ei->vfs_inode);
1143 const struct file_operations hugetlbfs_file_operations = {
1144 .read_iter = hugetlbfs_read_iter,
1145 .mmap = hugetlbfs_file_mmap,
1146 .fsync = noop_fsync,
1147 .get_unmapped_area = hugetlb_get_unmapped_area,
1148 .llseek = default_llseek,
1149 .fallocate = hugetlbfs_fallocate,
1152 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1153 .create = hugetlbfs_create,
1154 .lookup = simple_lookup,
1155 .link = simple_link,
1156 .unlink = simple_unlink,
1157 .symlink = hugetlbfs_symlink,
1158 .mkdir = hugetlbfs_mkdir,
1159 .rmdir = simple_rmdir,
1160 .mknod = hugetlbfs_mknod,
1161 .rename = simple_rename,
1162 .setattr = hugetlbfs_setattr,
1165 static const struct inode_operations hugetlbfs_inode_operations = {
1166 .setattr = hugetlbfs_setattr,
1169 static const struct super_operations hugetlbfs_ops = {
1170 .alloc_inode = hugetlbfs_alloc_inode,
1171 .free_inode = hugetlbfs_free_inode,
1172 .destroy_inode = hugetlbfs_destroy_inode,
1173 .evict_inode = hugetlbfs_evict_inode,
1174 .statfs = hugetlbfs_statfs,
1175 .put_super = hugetlbfs_put_super,
1176 .show_options = hugetlbfs_show_options,
1180 * Convert size option passed from command line to number of huge pages
1181 * in the pool specified by hstate. Size option could be in bytes
1182 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1185 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1186 enum hugetlbfs_size_type val_type)
1188 if (val_type == NO_SIZE)
1191 if (val_type == SIZE_PERCENT) {
1192 size_opt <<= huge_page_shift(h);
1193 size_opt *= h->max_huge_pages;
1194 do_div(size_opt, 100);
1197 size_opt >>= huge_page_shift(h);
1202 * Parse one mount parameter.
1204 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1206 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1207 struct fs_parse_result result;
1212 opt = fs_parse(fc, &hugetlb_fs_parameters, param, &result);
1218 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1219 if (!uid_valid(ctx->uid))
1224 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1225 if (!gid_valid(ctx->gid))
1230 ctx->mode = result.uint_32 & 01777U;
1234 /* memparse() will accept a K/M/G without a digit */
1235 if (!isdigit(param->string[0]))
1237 ctx->max_size_opt = memparse(param->string, &rest);
1238 ctx->max_val_type = SIZE_STD;
1240 ctx->max_val_type = SIZE_PERCENT;
1244 /* memparse() will accept a K/M/G without a digit */
1245 if (!isdigit(param->string[0]))
1247 ctx->nr_inodes = memparse(param->string, &rest);
1251 ps = memparse(param->string, &rest);
1252 ctx->hstate = size_to_hstate(ps);
1254 pr_err("Unsupported page size %lu MB\n", ps >> 20);
1260 /* memparse() will accept a K/M/G without a digit */
1261 if (!isdigit(param->string[0]))
1263 ctx->min_size_opt = memparse(param->string, &rest);
1264 ctx->min_val_type = SIZE_STD;
1266 ctx->min_val_type = SIZE_PERCENT;
1274 return invalf(fc, "hugetlbfs: Bad value '%s' for mount option '%s'\n",
1275 param->string, param->key);
1279 * Validate the parsed options.
1281 static int hugetlbfs_validate(struct fs_context *fc)
1283 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1286 * Use huge page pool size (in hstate) to convert the size
1287 * options to number of huge pages. If NO_SIZE, -1 is returned.
1289 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1292 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1297 * If max_size was specified, then min_size must be smaller
1299 if (ctx->max_val_type > NO_SIZE &&
1300 ctx->min_hpages > ctx->max_hpages) {
1301 pr_err("Minimum size can not be greater than maximum size\n");
1309 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1311 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1312 struct hugetlbfs_sb_info *sbinfo;
1314 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1317 sb->s_fs_info = sbinfo;
1318 spin_lock_init(&sbinfo->stat_lock);
1319 sbinfo->hstate = ctx->hstate;
1320 sbinfo->max_inodes = ctx->nr_inodes;
1321 sbinfo->free_inodes = ctx->nr_inodes;
1322 sbinfo->spool = NULL;
1323 sbinfo->uid = ctx->uid;
1324 sbinfo->gid = ctx->gid;
1325 sbinfo->mode = ctx->mode;
1328 * Allocate and initialize subpool if maximum or minimum size is
1329 * specified. Any needed reservations (for minimim size) are taken
1330 * taken when the subpool is created.
1332 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1333 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1339 sb->s_maxbytes = MAX_LFS_FILESIZE;
1340 sb->s_blocksize = huge_page_size(ctx->hstate);
1341 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1342 sb->s_magic = HUGETLBFS_MAGIC;
1343 sb->s_op = &hugetlbfs_ops;
1344 sb->s_time_gran = 1;
1347 * Due to the special and limited functionality of hugetlbfs, it does
1348 * not work well as a stacking filesystem.
1350 sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1351 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1356 kfree(sbinfo->spool);
1361 static int hugetlbfs_get_tree(struct fs_context *fc)
1363 int err = hugetlbfs_validate(fc);
1366 return get_tree_nodev(fc, hugetlbfs_fill_super);
1369 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1371 kfree(fc->fs_private);
1374 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1375 .free = hugetlbfs_fs_context_free,
1376 .parse_param = hugetlbfs_parse_param,
1377 .get_tree = hugetlbfs_get_tree,
1380 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1382 struct hugetlbfs_fs_context *ctx;
1384 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1388 ctx->max_hpages = -1; /* No limit on size by default */
1389 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1390 ctx->uid = current_fsuid();
1391 ctx->gid = current_fsgid();
1393 ctx->hstate = &default_hstate;
1394 ctx->min_hpages = -1; /* No default minimum size */
1395 ctx->max_val_type = NO_SIZE;
1396 ctx->min_val_type = NO_SIZE;
1397 fc->fs_private = ctx;
1398 fc->ops = &hugetlbfs_fs_context_ops;
1402 static struct file_system_type hugetlbfs_fs_type = {
1403 .name = "hugetlbfs",
1404 .init_fs_context = hugetlbfs_init_fs_context,
1405 .parameters = &hugetlb_fs_parameters,
1406 .kill_sb = kill_litter_super,
1409 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1411 static int can_do_hugetlb_shm(void)
1414 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1415 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1418 static int get_hstate_idx(int page_size_log)
1420 struct hstate *h = hstate_sizelog(page_size_log);
1428 * Note that size should be aligned to proper hugepage size in caller side,
1429 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1431 struct file *hugetlb_file_setup(const char *name, size_t size,
1432 vm_flags_t acctflag, struct user_struct **user,
1433 int creat_flags, int page_size_log)
1435 struct inode *inode;
1436 struct vfsmount *mnt;
1440 hstate_idx = get_hstate_idx(page_size_log);
1442 return ERR_PTR(-ENODEV);
1445 mnt = hugetlbfs_vfsmount[hstate_idx];
1447 return ERR_PTR(-ENOENT);
1449 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1450 *user = current_user();
1451 if (user_shm_lock(size, *user)) {
1453 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1454 current->comm, current->pid);
1455 task_unlock(current);
1458 return ERR_PTR(-EPERM);
1462 file = ERR_PTR(-ENOSPC);
1463 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1466 if (creat_flags == HUGETLB_SHMFS_INODE)
1467 inode->i_flags |= S_PRIVATE;
1469 inode->i_size = size;
1472 if (hugetlb_reserve_pages(inode, 0,
1473 size >> huge_page_shift(hstate_inode(inode)), NULL,
1475 file = ERR_PTR(-ENOMEM);
1477 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1478 &hugetlbfs_file_operations);
1485 user_shm_unlock(size, *user);
1491 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1493 struct fs_context *fc;
1494 struct vfsmount *mnt;
1496 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1500 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1506 pr_err("Cannot mount internal hugetlbfs for page size %uK",
1507 1U << (h->order + PAGE_SHIFT - 10));
1511 static int __init init_hugetlbfs_fs(void)
1513 struct vfsmount *mnt;
1518 if (!hugepages_supported()) {
1519 pr_info("disabling because there are no supported hugepage sizes\n");
1524 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1525 sizeof(struct hugetlbfs_inode_info),
1526 0, SLAB_ACCOUNT, init_once);
1527 if (hugetlbfs_inode_cachep == NULL)
1530 error = register_filesystem(&hugetlbfs_fs_type);
1534 /* default hstate mount is required */
1535 mnt = mount_one_hugetlbfs(&hstates[default_hstate_idx]);
1537 error = PTR_ERR(mnt);
1540 hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1542 /* other hstates are optional */
1544 for_each_hstate(h) {
1545 if (i == default_hstate_idx) {
1550 mnt = mount_one_hugetlbfs(h);
1552 hugetlbfs_vfsmount[i] = NULL;
1554 hugetlbfs_vfsmount[i] = mnt;
1561 (void)unregister_filesystem(&hugetlbfs_fs_type);
1563 kmem_cache_destroy(hugetlbfs_inode_cachep);
1567 fs_initcall(init_hugetlbfs_fs)