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>
42 static const struct super_operations hugetlbfs_ops;
43 static const struct address_space_operations hugetlbfs_aops;
44 const struct file_operations hugetlbfs_file_operations;
45 static const struct inode_operations hugetlbfs_dir_inode_operations;
46 static const struct inode_operations hugetlbfs_inode_operations;
48 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
50 struct hugetlbfs_fs_context {
51 struct hstate *hstate;
52 unsigned long long max_size_opt;
53 unsigned long long min_size_opt;
57 enum hugetlbfs_size_type max_val_type;
58 enum hugetlbfs_size_type min_val_type;
64 int sysctl_hugetlb_shm_group;
76 static const struct fs_parameter_spec hugetlb_param_specs[] = {
77 fsparam_u32 ("gid", Opt_gid),
78 fsparam_string("min_size", Opt_min_size),
79 fsparam_u32 ("mode", Opt_mode),
80 fsparam_string("nr_inodes", Opt_nr_inodes),
81 fsparam_string("pagesize", Opt_pagesize),
82 fsparam_string("size", Opt_size),
83 fsparam_u32 ("uid", Opt_uid),
87 static const struct fs_parameter_description hugetlb_fs_parameters = {
89 .specs = hugetlb_param_specs,
93 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
94 struct inode *inode, pgoff_t index)
96 vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
100 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
102 mpol_cond_put(vma->vm_policy);
105 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
106 struct inode *inode, pgoff_t index)
110 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
115 static void huge_pagevec_release(struct pagevec *pvec)
119 for (i = 0; i < pagevec_count(pvec); ++i)
120 put_page(pvec->pages[i]);
122 pagevec_reinit(pvec);
126 * Mask used when checking the page offset value passed in via system
127 * calls. This value will be converted to a loff_t which is signed.
128 * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
129 * value. The extra bit (- 1 in the shift value) is to take the sign
132 #define PGOFF_LOFFT_MAX \
133 (((1UL << (PAGE_SHIFT + 1)) - 1) << (BITS_PER_LONG - (PAGE_SHIFT + 1)))
135 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
137 struct inode *inode = file_inode(file);
140 struct hstate *h = hstate_file(file);
143 * vma address alignment (but not the pgoff alignment) has
144 * already been checked by prepare_hugepage_range. If you add
145 * any error returns here, do so after setting VM_HUGETLB, so
146 * is_vm_hugetlb_page tests below unmap_region go the right
147 * way when do_mmap_pgoff unwinds (may be important on powerpc
150 vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
151 vma->vm_ops = &hugetlb_vm_ops;
154 * page based offset in vm_pgoff could be sufficiently large to
155 * overflow a loff_t when converted to byte offset. This can
156 * only happen on architectures where sizeof(loff_t) ==
157 * sizeof(unsigned long). So, only check in those instances.
159 if (sizeof(unsigned long) == sizeof(loff_t)) {
160 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
164 /* must be huge page aligned */
165 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
168 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
169 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
170 /* check for overflow */
178 if (hugetlb_reserve_pages(inode,
179 vma->vm_pgoff >> huge_page_order(h),
180 len >> huge_page_shift(h), vma,
185 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
186 i_size_write(inode, len);
194 * Called under down_write(mmap_sem).
197 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
199 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
200 unsigned long len, unsigned long pgoff, unsigned long flags)
202 struct mm_struct *mm = current->mm;
203 struct vm_area_struct *vma;
204 struct hstate *h = hstate_file(file);
205 struct vm_unmapped_area_info info;
207 if (len & ~huge_page_mask(h))
212 if (flags & MAP_FIXED) {
213 if (prepare_hugepage_range(file, addr, len))
219 addr = ALIGN(addr, huge_page_size(h));
220 vma = find_vma(mm, addr);
221 if (TASK_SIZE - len >= addr &&
222 (!vma || addr + len <= vm_start_gap(vma)))
228 info.low_limit = TASK_UNMAPPED_BASE;
229 info.high_limit = TASK_SIZE;
230 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
231 info.align_offset = 0;
232 return vm_unmapped_area(&info);
237 hugetlbfs_read_actor(struct page *page, unsigned long offset,
238 struct iov_iter *to, unsigned long size)
243 /* Find which 4k chunk and offset with in that chunk */
244 i = offset >> PAGE_SHIFT;
245 offset = offset & ~PAGE_MASK;
249 chunksize = PAGE_SIZE;
252 if (chunksize > size)
254 n = copy_page_to_iter(&page[i], offset, chunksize, to);
266 * Support for read() - Find the page attached to f_mapping and copy out the
267 * data. Its *very* similar to do_generic_mapping_read(), we can't use that
268 * since it has PAGE_SIZE assumptions.
270 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
272 struct file *file = iocb->ki_filp;
273 struct hstate *h = hstate_file(file);
274 struct address_space *mapping = file->f_mapping;
275 struct inode *inode = mapping->host;
276 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
277 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
278 unsigned long end_index;
282 while (iov_iter_count(to)) {
286 /* nr is the maximum number of bytes to copy from this page */
287 nr = huge_page_size(h);
288 isize = i_size_read(inode);
291 end_index = (isize - 1) >> huge_page_shift(h);
292 if (index > end_index)
294 if (index == end_index) {
295 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
302 page = find_lock_page(mapping, index);
303 if (unlikely(page == NULL)) {
305 * We have a HOLE, zero out the user-buffer for the
306 * length of the hole or request.
308 copied = iov_iter_zero(nr, to);
313 * We have the page, copy it to user space buffer.
315 copied = hugetlbfs_read_actor(page, offset, to, nr);
320 if (copied != nr && iov_iter_count(to)) {
325 index += offset >> huge_page_shift(h);
326 offset &= ~huge_page_mask(h);
328 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
332 static int hugetlbfs_write_begin(struct file *file,
333 struct address_space *mapping,
334 loff_t pos, unsigned len, unsigned flags,
335 struct page **pagep, void **fsdata)
340 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
341 loff_t pos, unsigned len, unsigned copied,
342 struct page *page, void *fsdata)
348 static void remove_huge_page(struct page *page)
350 ClearPageDirty(page);
351 ClearPageUptodate(page);
352 delete_from_page_cache(page);
356 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
358 struct vm_area_struct *vma;
361 * end == 0 indicates that the entire range after
362 * start should be unmapped.
364 vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
365 unsigned long v_offset;
369 * Can the expression below overflow on 32-bit arches?
370 * No, because the interval tree returns us only those vmas
371 * which overlap the truncated area starting at pgoff,
372 * and no vma on a 32-bit arch can span beyond the 4GB.
374 if (vma->vm_pgoff < start)
375 v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
382 v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
384 if (v_end > vma->vm_end)
388 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
394 * remove_inode_hugepages handles two distinct cases: truncation and hole
395 * punch. There are subtle differences in operation for each case.
397 * truncation is indicated by end of range being LLONG_MAX
398 * In this case, we first scan the range and release found pages.
399 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
400 * maps and global counts. Page faults can not race with truncation
401 * in this routine. hugetlb_no_page() prevents page faults in the
402 * truncated range. It checks i_size before allocation, and again after
403 * with the page table lock for the page held. The same lock must be
404 * acquired to unmap a page.
405 * hole punch is indicated if end is not LLONG_MAX
406 * In the hole punch case we scan the range and release found pages.
407 * Only when releasing a page is the associated region/reserv map
408 * deleted. The region/reserv map for ranges without associated
409 * pages are not modified. Page faults can race with hole punch.
410 * This is indicated if we find a mapped page.
411 * Note: If the passed end of range value is beyond the end of file, but
412 * not LLONG_MAX this routine still performs a hole punch operation.
414 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
417 struct hstate *h = hstate_inode(inode);
418 struct address_space *mapping = &inode->i_data;
419 const pgoff_t start = lstart >> huge_page_shift(h);
420 const pgoff_t end = lend >> huge_page_shift(h);
421 struct vm_area_struct pseudo_vma;
425 bool truncate_op = (lend == LLONG_MAX);
427 vma_init(&pseudo_vma, current->mm);
428 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
433 * When no more pages are found, we are done.
435 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
438 for (i = 0; i < pagevec_count(&pvec); ++i) {
439 struct page *page = pvec.pages[i];
443 hash = hugetlb_fault_mutex_hash(h, mapping, index);
444 mutex_lock(&hugetlb_fault_mutex_table[hash]);
447 * If page is mapped, it was faulted in after being
448 * unmapped in caller. Unmap (again) now after taking
449 * the fault mutex. The mutex will prevent faults
450 * until we finish removing the page.
452 * This race can only happen in the hole punch case.
453 * Getting here in a truncate operation is a bug.
455 if (unlikely(page_mapped(page))) {
458 i_mmap_lock_write(mapping);
459 hugetlb_vmdelete_list(&mapping->i_mmap,
460 index * pages_per_huge_page(h),
461 (index + 1) * pages_per_huge_page(h));
462 i_mmap_unlock_write(mapping);
467 * We must free the huge page and remove from page
468 * cache (remove_huge_page) BEFORE removing the
469 * region/reserve map (hugetlb_unreserve_pages). In
470 * rare out of memory conditions, removal of the
471 * region/reserve map could fail. Correspondingly,
472 * the subpool and global reserve usage count can need
475 VM_BUG_ON(PagePrivate(page));
476 remove_huge_page(page);
479 if (unlikely(hugetlb_unreserve_pages(inode,
480 index, index + 1, 1)))
481 hugetlb_fix_reserve_counts(inode);
485 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
487 huge_pagevec_release(&pvec);
492 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
495 static void hugetlbfs_evict_inode(struct inode *inode)
497 struct resv_map *resv_map;
499 remove_inode_hugepages(inode, 0, LLONG_MAX);
502 * Get the resv_map from the address space embedded in the inode.
503 * This is the address space which points to any resv_map allocated
504 * at inode creation time. If this is a device special inode,
505 * i_mapping may not point to the original address space.
507 resv_map = (struct resv_map *)(&inode->i_data)->private_data;
508 /* Only regular and link inodes have associated reserve maps */
510 resv_map_release(&resv_map->refs);
514 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
517 struct address_space *mapping = inode->i_mapping;
518 struct hstate *h = hstate_inode(inode);
520 BUG_ON(offset & ~huge_page_mask(h));
521 pgoff = offset >> PAGE_SHIFT;
523 i_size_write(inode, offset);
524 i_mmap_lock_write(mapping);
525 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
526 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
527 i_mmap_unlock_write(mapping);
528 remove_inode_hugepages(inode, offset, LLONG_MAX);
532 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
534 struct hstate *h = hstate_inode(inode);
535 loff_t hpage_size = huge_page_size(h);
536 loff_t hole_start, hole_end;
539 * For hole punch round up the beginning offset of the hole and
540 * round down the end.
542 hole_start = round_up(offset, hpage_size);
543 hole_end = round_down(offset + len, hpage_size);
545 if (hole_end > hole_start) {
546 struct address_space *mapping = inode->i_mapping;
547 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
551 /* protected by i_mutex */
552 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
557 i_mmap_lock_write(mapping);
558 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
559 hugetlb_vmdelete_list(&mapping->i_mmap,
560 hole_start >> PAGE_SHIFT,
561 hole_end >> PAGE_SHIFT);
562 i_mmap_unlock_write(mapping);
563 remove_inode_hugepages(inode, hole_start, hole_end);
570 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
573 struct inode *inode = file_inode(file);
574 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
575 struct address_space *mapping = inode->i_mapping;
576 struct hstate *h = hstate_inode(inode);
577 struct vm_area_struct pseudo_vma;
578 struct mm_struct *mm = current->mm;
579 loff_t hpage_size = huge_page_size(h);
580 unsigned long hpage_shift = huge_page_shift(h);
581 pgoff_t start, index, end;
585 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
588 if (mode & FALLOC_FL_PUNCH_HOLE)
589 return hugetlbfs_punch_hole(inode, offset, len);
592 * Default preallocate case.
593 * For this range, start is rounded down and end is rounded up
594 * as well as being converted to page offsets.
596 start = offset >> hpage_shift;
597 end = (offset + len + hpage_size - 1) >> hpage_shift;
601 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
602 error = inode_newsize_ok(inode, offset + len);
606 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
612 * Initialize a pseudo vma as this is required by the huge page
613 * allocation routines. If NUMA is configured, use page index
614 * as input to create an allocation policy.
616 vma_init(&pseudo_vma, mm);
617 pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
618 pseudo_vma.vm_file = file;
620 for (index = start; index < end; index++) {
622 * This is supposed to be the vaddr where the page is being
623 * faulted in, but we have no vaddr here.
627 int avoid_reserve = 0;
632 * fallocate(2) manpage permits EINTR; we may have been
633 * interrupted because we are using up too much memory.
635 if (signal_pending(current)) {
640 /* Set numa allocation policy based on index */
641 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
643 /* addr is the offset within the file (zero based) */
644 addr = index * hpage_size;
646 /* mutex taken here, fault path and hole punch */
647 hash = hugetlb_fault_mutex_hash(h, mapping, index);
648 mutex_lock(&hugetlb_fault_mutex_table[hash]);
650 /* See if already present in mapping to avoid alloc/free */
651 page = find_get_page(mapping, index);
654 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
655 hugetlb_drop_vma_policy(&pseudo_vma);
659 /* Allocate page and add to page cache */
660 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
661 hugetlb_drop_vma_policy(&pseudo_vma);
663 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
664 error = PTR_ERR(page);
667 clear_huge_page(page, addr, pages_per_huge_page(h));
668 __SetPageUptodate(page);
669 error = huge_add_to_page_cache(page, mapping, index);
670 if (unlikely(error)) {
672 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
676 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
679 * unlock_page because locked by add_to_page_cache()
680 * page_put due to reference from alloc_huge_page()
686 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
687 i_size_write(inode, offset + len);
688 inode->i_ctime = current_time(inode);
694 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
696 struct inode *inode = d_inode(dentry);
697 struct hstate *h = hstate_inode(inode);
699 unsigned int ia_valid = attr->ia_valid;
700 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
704 error = setattr_prepare(dentry, attr);
708 if (ia_valid & ATTR_SIZE) {
709 loff_t oldsize = inode->i_size;
710 loff_t newsize = attr->ia_size;
712 if (newsize & ~huge_page_mask(h))
714 /* protected by i_mutex */
715 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
716 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
718 error = hugetlb_vmtruncate(inode, newsize);
723 setattr_copy(inode, attr);
724 mark_inode_dirty(inode);
728 static struct inode *hugetlbfs_get_root(struct super_block *sb,
729 struct hugetlbfs_fs_context *ctx)
733 inode = new_inode(sb);
735 inode->i_ino = get_next_ino();
736 inode->i_mode = S_IFDIR | ctx->mode;
737 inode->i_uid = ctx->uid;
738 inode->i_gid = ctx->gid;
739 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
740 inode->i_op = &hugetlbfs_dir_inode_operations;
741 inode->i_fop = &simple_dir_operations;
742 /* directory inodes start off with i_nlink == 2 (for "." entry) */
744 lockdep_annotate_inode_mutex_key(inode);
750 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
751 * be taken from reclaim -- unlike regular filesystems. This needs an
752 * annotation because huge_pmd_share() does an allocation under hugetlb's
755 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
757 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
759 umode_t mode, dev_t dev)
762 struct resv_map *resv_map = NULL;
765 * Reserve maps are only needed for inodes that can have associated
768 if (S_ISREG(mode) || S_ISLNK(mode)) {
769 resv_map = resv_map_alloc();
774 inode = new_inode(sb);
776 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
778 inode->i_ino = get_next_ino();
779 inode_init_owner(inode, dir, mode);
780 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
781 &hugetlbfs_i_mmap_rwsem_key);
782 inode->i_mapping->a_ops = &hugetlbfs_aops;
783 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
784 inode->i_mapping->private_data = resv_map;
785 info->seals = F_SEAL_SEAL;
786 switch (mode & S_IFMT) {
788 init_special_inode(inode, mode, dev);
791 inode->i_op = &hugetlbfs_inode_operations;
792 inode->i_fop = &hugetlbfs_file_operations;
795 inode->i_op = &hugetlbfs_dir_inode_operations;
796 inode->i_fop = &simple_dir_operations;
798 /* directory inodes start off with i_nlink == 2 (for "." entry) */
802 inode->i_op = &page_symlink_inode_operations;
803 inode_nohighmem(inode);
806 lockdep_annotate_inode_mutex_key(inode);
809 kref_put(&resv_map->refs, resv_map_release);
816 * File creation. Allocate an inode, and we're done..
818 static int hugetlbfs_mknod(struct inode *dir,
819 struct dentry *dentry, umode_t mode, dev_t dev)
824 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
826 dir->i_ctime = dir->i_mtime = current_time(dir);
827 d_instantiate(dentry, inode);
828 dget(dentry); /* Extra count - pin the dentry in core */
834 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
836 int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
842 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
844 return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
847 static int hugetlbfs_symlink(struct inode *dir,
848 struct dentry *dentry, const char *symname)
853 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
855 int l = strlen(symname)+1;
856 error = page_symlink(inode, symname, l);
858 d_instantiate(dentry, inode);
863 dir->i_ctime = dir->i_mtime = current_time(dir);
869 * mark the head page dirty
871 static int hugetlbfs_set_page_dirty(struct page *page)
873 struct page *head = compound_head(page);
879 static int hugetlbfs_migrate_page(struct address_space *mapping,
880 struct page *newpage, struct page *page,
881 enum migrate_mode mode)
885 rc = migrate_huge_page_move_mapping(mapping, newpage, page);
886 if (rc != MIGRATEPAGE_SUCCESS)
890 * page_private is subpool pointer in hugetlb pages. Transfer to
891 * new page. PagePrivate is not associated with page_private for
892 * hugetlb pages and can not be set here as only page_huge_active
893 * pages can be migrated.
895 if (page_private(page)) {
896 set_page_private(newpage, page_private(page));
897 set_page_private(page, 0);
900 if (mode != MIGRATE_SYNC_NO_COPY)
901 migrate_page_copy(newpage, page);
903 migrate_page_states(newpage, page);
905 return MIGRATEPAGE_SUCCESS;
908 static int hugetlbfs_error_remove_page(struct address_space *mapping,
911 struct inode *inode = mapping->host;
912 pgoff_t index = page->index;
914 remove_huge_page(page);
915 if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
916 hugetlb_fix_reserve_counts(inode);
922 * Display the mount options in /proc/mounts.
924 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
926 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
927 struct hugepage_subpool *spool = sbinfo->spool;
928 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
929 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
932 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
933 seq_printf(m, ",uid=%u",
934 from_kuid_munged(&init_user_ns, sbinfo->uid));
935 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
936 seq_printf(m, ",gid=%u",
937 from_kgid_munged(&init_user_ns, sbinfo->gid));
938 if (sbinfo->mode != 0755)
939 seq_printf(m, ",mode=%o", sbinfo->mode);
940 if (sbinfo->max_inodes != -1)
941 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
945 if (hpage_size >= 1024) {
949 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
951 if (spool->max_hpages != -1)
952 seq_printf(m, ",size=%llu",
953 (unsigned long long)spool->max_hpages << hpage_shift);
954 if (spool->min_hpages != -1)
955 seq_printf(m, ",min_size=%llu",
956 (unsigned long long)spool->min_hpages << hpage_shift);
961 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
963 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
964 struct hstate *h = hstate_inode(d_inode(dentry));
966 buf->f_type = HUGETLBFS_MAGIC;
967 buf->f_bsize = huge_page_size(h);
969 spin_lock(&sbinfo->stat_lock);
970 /* If no limits set, just report 0 for max/free/used
971 * blocks, like simple_statfs() */
975 spin_lock(&sbinfo->spool->lock);
976 buf->f_blocks = sbinfo->spool->max_hpages;
977 free_pages = sbinfo->spool->max_hpages
978 - sbinfo->spool->used_hpages;
979 buf->f_bavail = buf->f_bfree = free_pages;
980 spin_unlock(&sbinfo->spool->lock);
981 buf->f_files = sbinfo->max_inodes;
982 buf->f_ffree = sbinfo->free_inodes;
984 spin_unlock(&sbinfo->stat_lock);
986 buf->f_namelen = NAME_MAX;
990 static void hugetlbfs_put_super(struct super_block *sb)
992 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
995 sb->s_fs_info = NULL;
998 hugepage_put_subpool(sbi->spool);
1004 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1006 if (sbinfo->free_inodes >= 0) {
1007 spin_lock(&sbinfo->stat_lock);
1008 if (unlikely(!sbinfo->free_inodes)) {
1009 spin_unlock(&sbinfo->stat_lock);
1012 sbinfo->free_inodes--;
1013 spin_unlock(&sbinfo->stat_lock);
1019 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1021 if (sbinfo->free_inodes >= 0) {
1022 spin_lock(&sbinfo->stat_lock);
1023 sbinfo->free_inodes++;
1024 spin_unlock(&sbinfo->stat_lock);
1029 static struct kmem_cache *hugetlbfs_inode_cachep;
1031 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1033 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1034 struct hugetlbfs_inode_info *p;
1036 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1038 p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1040 hugetlbfs_inc_free_inodes(sbinfo);
1045 * Any time after allocation, hugetlbfs_destroy_inode can be called
1046 * for the inode. mpol_free_shared_policy is unconditionally called
1047 * as part of hugetlbfs_destroy_inode. So, initialize policy here
1048 * in case of a quick call to destroy.
1050 * Note that the policy is initialized even if we are creating a
1051 * private inode. This simplifies hugetlbfs_destroy_inode.
1053 mpol_shared_policy_init(&p->policy, NULL);
1055 return &p->vfs_inode;
1058 static void hugetlbfs_free_inode(struct inode *inode)
1060 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1063 static void hugetlbfs_destroy_inode(struct inode *inode)
1065 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1066 mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1069 static const struct address_space_operations hugetlbfs_aops = {
1070 .write_begin = hugetlbfs_write_begin,
1071 .write_end = hugetlbfs_write_end,
1072 .set_page_dirty = hugetlbfs_set_page_dirty,
1073 .migratepage = hugetlbfs_migrate_page,
1074 .error_remove_page = hugetlbfs_error_remove_page,
1078 static void init_once(void *foo)
1080 struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1082 inode_init_once(&ei->vfs_inode);
1085 const struct file_operations hugetlbfs_file_operations = {
1086 .read_iter = hugetlbfs_read_iter,
1087 .mmap = hugetlbfs_file_mmap,
1088 .fsync = noop_fsync,
1089 .get_unmapped_area = hugetlb_get_unmapped_area,
1090 .llseek = default_llseek,
1091 .fallocate = hugetlbfs_fallocate,
1094 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1095 .create = hugetlbfs_create,
1096 .lookup = simple_lookup,
1097 .link = simple_link,
1098 .unlink = simple_unlink,
1099 .symlink = hugetlbfs_symlink,
1100 .mkdir = hugetlbfs_mkdir,
1101 .rmdir = simple_rmdir,
1102 .mknod = hugetlbfs_mknod,
1103 .rename = simple_rename,
1104 .setattr = hugetlbfs_setattr,
1107 static const struct inode_operations hugetlbfs_inode_operations = {
1108 .setattr = hugetlbfs_setattr,
1111 static const struct super_operations hugetlbfs_ops = {
1112 .alloc_inode = hugetlbfs_alloc_inode,
1113 .free_inode = hugetlbfs_free_inode,
1114 .destroy_inode = hugetlbfs_destroy_inode,
1115 .evict_inode = hugetlbfs_evict_inode,
1116 .statfs = hugetlbfs_statfs,
1117 .put_super = hugetlbfs_put_super,
1118 .show_options = hugetlbfs_show_options,
1122 * Convert size option passed from command line to number of huge pages
1123 * in the pool specified by hstate. Size option could be in bytes
1124 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1127 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1128 enum hugetlbfs_size_type val_type)
1130 if (val_type == NO_SIZE)
1133 if (val_type == SIZE_PERCENT) {
1134 size_opt <<= huge_page_shift(h);
1135 size_opt *= h->max_huge_pages;
1136 do_div(size_opt, 100);
1139 size_opt >>= huge_page_shift(h);
1144 * Parse one mount parameter.
1146 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1148 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1149 struct fs_parse_result result;
1154 opt = fs_parse(fc, &hugetlb_fs_parameters, param, &result);
1160 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1161 if (!uid_valid(ctx->uid))
1166 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1167 if (!gid_valid(ctx->gid))
1172 ctx->mode = result.uint_32 & 01777U;
1176 /* memparse() will accept a K/M/G without a digit */
1177 if (!isdigit(param->string[0]))
1179 ctx->max_size_opt = memparse(param->string, &rest);
1180 ctx->max_val_type = SIZE_STD;
1182 ctx->max_val_type = SIZE_PERCENT;
1186 /* memparse() will accept a K/M/G without a digit */
1187 if (!isdigit(param->string[0]))
1189 ctx->nr_inodes = memparse(param->string, &rest);
1193 ps = memparse(param->string, &rest);
1194 ctx->hstate = size_to_hstate(ps);
1196 pr_err("Unsupported page size %lu MB\n", ps >> 20);
1202 /* memparse() will accept a K/M/G without a digit */
1203 if (!isdigit(param->string[0]))
1205 ctx->min_size_opt = memparse(param->string, &rest);
1206 ctx->min_val_type = SIZE_STD;
1208 ctx->min_val_type = SIZE_PERCENT;
1216 return invalf(fc, "hugetlbfs: Bad value '%s' for mount option '%s'\n",
1217 param->string, param->key);
1221 * Validate the parsed options.
1223 static int hugetlbfs_validate(struct fs_context *fc)
1225 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1228 * Use huge page pool size (in hstate) to convert the size
1229 * options to number of huge pages. If NO_SIZE, -1 is returned.
1231 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1234 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1239 * If max_size was specified, then min_size must be smaller
1241 if (ctx->max_val_type > NO_SIZE &&
1242 ctx->min_hpages > ctx->max_hpages) {
1243 pr_err("Minimum size can not be greater than maximum size\n");
1251 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1253 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1254 struct hugetlbfs_sb_info *sbinfo;
1256 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1259 sb->s_fs_info = sbinfo;
1260 spin_lock_init(&sbinfo->stat_lock);
1261 sbinfo->hstate = ctx->hstate;
1262 sbinfo->max_inodes = ctx->nr_inodes;
1263 sbinfo->free_inodes = ctx->nr_inodes;
1264 sbinfo->spool = NULL;
1265 sbinfo->uid = ctx->uid;
1266 sbinfo->gid = ctx->gid;
1267 sbinfo->mode = ctx->mode;
1270 * Allocate and initialize subpool if maximum or minimum size is
1271 * specified. Any needed reservations (for minimim size) are taken
1272 * taken when the subpool is created.
1274 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1275 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1281 sb->s_maxbytes = MAX_LFS_FILESIZE;
1282 sb->s_blocksize = huge_page_size(ctx->hstate);
1283 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1284 sb->s_magic = HUGETLBFS_MAGIC;
1285 sb->s_op = &hugetlbfs_ops;
1286 sb->s_time_gran = 1;
1287 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1292 kfree(sbinfo->spool);
1297 static int hugetlbfs_get_tree(struct fs_context *fc)
1299 int err = hugetlbfs_validate(fc);
1302 return get_tree_nodev(fc, hugetlbfs_fill_super);
1305 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1307 kfree(fc->fs_private);
1310 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1311 .free = hugetlbfs_fs_context_free,
1312 .parse_param = hugetlbfs_parse_param,
1313 .get_tree = hugetlbfs_get_tree,
1316 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1318 struct hugetlbfs_fs_context *ctx;
1320 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1324 ctx->max_hpages = -1; /* No limit on size by default */
1325 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1326 ctx->uid = current_fsuid();
1327 ctx->gid = current_fsgid();
1329 ctx->hstate = &default_hstate;
1330 ctx->min_hpages = -1; /* No default minimum size */
1331 ctx->max_val_type = NO_SIZE;
1332 ctx->min_val_type = NO_SIZE;
1333 fc->fs_private = ctx;
1334 fc->ops = &hugetlbfs_fs_context_ops;
1338 static struct file_system_type hugetlbfs_fs_type = {
1339 .name = "hugetlbfs",
1340 .init_fs_context = hugetlbfs_init_fs_context,
1341 .parameters = &hugetlb_fs_parameters,
1342 .kill_sb = kill_litter_super,
1345 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1347 static int can_do_hugetlb_shm(void)
1350 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1351 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1354 static int get_hstate_idx(int page_size_log)
1356 struct hstate *h = hstate_sizelog(page_size_log);
1364 * Note that size should be aligned to proper hugepage size in caller side,
1365 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1367 struct file *hugetlb_file_setup(const char *name, size_t size,
1368 vm_flags_t acctflag, struct user_struct **user,
1369 int creat_flags, int page_size_log)
1371 struct inode *inode;
1372 struct vfsmount *mnt;
1376 hstate_idx = get_hstate_idx(page_size_log);
1378 return ERR_PTR(-ENODEV);
1381 mnt = hugetlbfs_vfsmount[hstate_idx];
1383 return ERR_PTR(-ENOENT);
1385 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1386 *user = current_user();
1387 if (user_shm_lock(size, *user)) {
1389 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1390 current->comm, current->pid);
1391 task_unlock(current);
1394 return ERR_PTR(-EPERM);
1398 file = ERR_PTR(-ENOSPC);
1399 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1402 if (creat_flags == HUGETLB_SHMFS_INODE)
1403 inode->i_flags |= S_PRIVATE;
1405 inode->i_size = size;
1408 if (hugetlb_reserve_pages(inode, 0,
1409 size >> huge_page_shift(hstate_inode(inode)), NULL,
1411 file = ERR_PTR(-ENOMEM);
1413 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1414 &hugetlbfs_file_operations);
1421 user_shm_unlock(size, *user);
1427 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1429 struct fs_context *fc;
1430 struct vfsmount *mnt;
1432 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1436 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1442 pr_err("Cannot mount internal hugetlbfs for page size %uK",
1443 1U << (h->order + PAGE_SHIFT - 10));
1447 static int __init init_hugetlbfs_fs(void)
1449 struct vfsmount *mnt;
1454 if (!hugepages_supported()) {
1455 pr_info("disabling because there are no supported hugepage sizes\n");
1460 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1461 sizeof(struct hugetlbfs_inode_info),
1462 0, SLAB_ACCOUNT, init_once);
1463 if (hugetlbfs_inode_cachep == NULL)
1466 error = register_filesystem(&hugetlbfs_fs_type);
1471 for_each_hstate(h) {
1472 mnt = mount_one_hugetlbfs(h);
1473 if (IS_ERR(mnt) && i == 0) {
1474 error = PTR_ERR(mnt);
1477 hugetlbfs_vfsmount[i] = mnt;
1484 kmem_cache_destroy(hugetlbfs_inode_cachep);
1488 fs_initcall(init_hugetlbfs_fs)