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/falloc.h>
16 #include <linux/mount.h>
17 #include <linux/file.h>
18 #include <linux/kernel.h>
19 #include <linux/writeback.h>
20 #include <linux/pagemap.h>
21 #include <linux/highmem.h>
22 #include <linux/init.h>
23 #include <linux/string.h>
24 #include <linux/capability.h>
25 #include <linux/ctype.h>
26 #include <linux/backing-dev.h>
27 #include <linux/hugetlb.h>
28 #include <linux/pagevec.h>
29 #include <linux/fs_parser.h>
30 #include <linux/mman.h>
31 #include <linux/slab.h>
32 #include <linux/dnotify.h>
33 #include <linux/statfs.h>
34 #include <linux/security.h>
35 #include <linux/magic.h>
36 #include <linux/migrate.h>
37 #include <linux/uio.h>
39 #include <linux/uaccess.h>
40 #include <linux/sched/mm.h>
42 static const struct address_space_operations hugetlbfs_aops;
43 const struct file_operations hugetlbfs_file_operations;
44 static const struct inode_operations hugetlbfs_dir_inode_operations;
45 static const struct inode_operations hugetlbfs_inode_operations;
47 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
49 struct hugetlbfs_fs_context {
50 struct hstate *hstate;
51 unsigned long long max_size_opt;
52 unsigned long long min_size_opt;
56 enum hugetlbfs_size_type max_val_type;
57 enum hugetlbfs_size_type min_val_type;
63 int sysctl_hugetlb_shm_group;
75 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
76 fsparam_u32 ("gid", Opt_gid),
77 fsparam_string("min_size", Opt_min_size),
78 fsparam_u32oct("mode", Opt_mode),
79 fsparam_string("nr_inodes", Opt_nr_inodes),
80 fsparam_string("pagesize", Opt_pagesize),
81 fsparam_string("size", Opt_size),
82 fsparam_u32 ("uid", Opt_uid),
87 * Mask used when checking the page offset value passed in via system
88 * calls. This value will be converted to a loff_t which is signed.
89 * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
90 * value. The extra bit (- 1 in the shift value) is to take the sign
93 #define PGOFF_LOFFT_MAX \
94 (((1UL << (PAGE_SHIFT + 1)) - 1) << (BITS_PER_LONG - (PAGE_SHIFT + 1)))
96 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
98 struct inode *inode = file_inode(file);
99 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
102 struct hstate *h = hstate_file(file);
105 * vma address alignment (but not the pgoff alignment) has
106 * already been checked by prepare_hugepage_range. If you add
107 * any error returns here, do so after setting VM_HUGETLB, so
108 * is_vm_hugetlb_page tests below unmap_region go the right
109 * way when do_mmap unwinds (may be important on powerpc
112 vm_flags_set(vma, VM_HUGETLB | VM_DONTEXPAND);
113 vma->vm_ops = &hugetlb_vm_ops;
115 ret = seal_check_write(info->seals, vma);
120 * page based offset in vm_pgoff could be sufficiently large to
121 * overflow a loff_t when converted to byte offset. This can
122 * only happen on architectures where sizeof(loff_t) ==
123 * sizeof(unsigned long). So, only check in those instances.
125 if (sizeof(unsigned long) == sizeof(loff_t)) {
126 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
130 /* must be huge page aligned */
131 if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
134 vma_len = (loff_t)(vma->vm_end - vma->vm_start);
135 len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
136 /* check for overflow */
144 if (!hugetlb_reserve_pages(inode,
145 vma->vm_pgoff >> huge_page_order(h),
146 len >> huge_page_shift(h), vma,
151 if (vma->vm_flags & VM_WRITE && inode->i_size < len)
152 i_size_write(inode, len);
160 * Called under mmap_write_lock(mm).
164 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
165 unsigned long len, unsigned long pgoff, unsigned long flags)
167 struct hstate *h = hstate_file(file);
168 struct vm_unmapped_area_info info;
172 info.low_limit = current->mm->mmap_base;
173 info.high_limit = arch_get_mmap_end(addr, len, flags);
174 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
175 info.align_offset = 0;
176 return vm_unmapped_area(&info);
180 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
181 unsigned long len, unsigned long pgoff, unsigned long flags)
183 struct hstate *h = hstate_file(file);
184 struct vm_unmapped_area_info info;
186 info.flags = VM_UNMAPPED_AREA_TOPDOWN;
188 info.low_limit = PAGE_SIZE;
189 info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
190 info.align_mask = PAGE_MASK & ~huge_page_mask(h);
191 info.align_offset = 0;
192 addr = vm_unmapped_area(&info);
195 * A failed mmap() very likely causes application failure,
196 * so fall back to the bottom-up function here. This scenario
197 * can happen with large stack limits and large mmap()
200 if (unlikely(offset_in_page(addr))) {
201 VM_BUG_ON(addr != -ENOMEM);
203 info.low_limit = current->mm->mmap_base;
204 info.high_limit = arch_get_mmap_end(addr, len, flags);
205 addr = vm_unmapped_area(&info);
212 generic_hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
213 unsigned long len, unsigned long pgoff,
216 struct mm_struct *mm = current->mm;
217 struct vm_area_struct *vma;
218 struct hstate *h = hstate_file(file);
219 const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
221 if (len & ~huge_page_mask(h))
226 if (flags & MAP_FIXED) {
227 if (prepare_hugepage_range(file, addr, len))
233 addr = ALIGN(addr, huge_page_size(h));
234 vma = find_vma(mm, addr);
235 if (mmap_end - len >= addr &&
236 (!vma || addr + len <= vm_start_gap(vma)))
241 * Use mm->get_unmapped_area value as a hint to use topdown routine.
242 * If architectures have special needs, they should define their own
243 * version of hugetlb_get_unmapped_area.
245 if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
246 return hugetlb_get_unmapped_area_topdown(file, addr, len,
248 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
252 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
254 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
255 unsigned long len, unsigned long pgoff,
258 return generic_hugetlb_get_unmapped_area(file, addr, len, pgoff, flags);
263 * Someone wants to read @bytes from a HWPOISON hugetlb @page from @offset.
264 * Returns the maximum number of bytes one can read without touching the 1st raw
267 * The implementation borrows the iteration logic from copy_page_to_iter*.
269 static size_t adjust_range_hwpoison(struct page *page, size_t offset, size_t bytes)
274 /* First subpage to start the loop. */
275 page = nth_page(page, offset / PAGE_SIZE);
278 if (is_raw_hwpoison_page_in_hugepage(page))
281 /* Safe to read n bytes without touching HWPOISON subpage. */
282 n = min(bytes, (size_t)PAGE_SIZE - offset);
288 if (offset == PAGE_SIZE) {
289 page = nth_page(page, 1);
298 * Support for read() - Find the page attached to f_mapping and copy out the
299 * data. This provides functionality similar to filemap_read().
301 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
303 struct file *file = iocb->ki_filp;
304 struct hstate *h = hstate_file(file);
305 struct address_space *mapping = file->f_mapping;
306 struct inode *inode = mapping->host;
307 unsigned long index = iocb->ki_pos >> huge_page_shift(h);
308 unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
309 unsigned long end_index;
313 while (iov_iter_count(to)) {
315 size_t nr, copied, want;
317 /* nr is the maximum number of bytes to copy from this page */
318 nr = huge_page_size(h);
319 isize = i_size_read(inode);
322 end_index = (isize - 1) >> huge_page_shift(h);
323 if (index > end_index)
325 if (index == end_index) {
326 nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
333 folio = filemap_lock_hugetlb_folio(h, mapping, index);
336 * We have a HOLE, zero out the user-buffer for the
337 * length of the hole or request.
339 copied = iov_iter_zero(nr, to);
343 if (!folio_test_has_hwpoisoned(folio))
347 * Adjust how many bytes safe to read without
348 * touching the 1st raw HWPOISON subpage after
351 want = adjust_range_hwpoison(&folio->page, offset, nr);
360 * We have the folio, copy it to user space buffer.
362 copied = copy_folio_to_iter(folio, offset, want, to);
367 if (copied != nr && iov_iter_count(to)) {
372 index += offset >> huge_page_shift(h);
373 offset &= ~huge_page_mask(h);
375 iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
379 static int hugetlbfs_write_begin(struct file *file,
380 struct address_space *mapping,
381 loff_t pos, unsigned len,
382 struct page **pagep, void **fsdata)
387 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
388 loff_t pos, unsigned len, unsigned copied,
389 struct page *page, void *fsdata)
395 static void hugetlb_delete_from_page_cache(struct folio *folio)
397 folio_clear_dirty(folio);
398 folio_clear_uptodate(folio);
399 filemap_remove_folio(folio);
403 * Called with i_mmap_rwsem held for inode based vma maps. This makes
404 * sure vma (and vm_mm) will not go away. We also hold the hugetlb fault
405 * mutex for the page in the mapping. So, we can not race with page being
406 * faulted into the vma.
408 static bool hugetlb_vma_maps_page(struct vm_area_struct *vma,
409 unsigned long addr, struct page *page)
413 ptep = hugetlb_walk(vma, addr, huge_page_size(hstate_vma(vma)));
417 pte = huge_ptep_get(ptep);
418 if (huge_pte_none(pte) || !pte_present(pte))
421 if (pte_page(pte) == page)
428 * Can vma_offset_start/vma_offset_end overflow on 32-bit arches?
429 * No, because the interval tree returns us only those vmas
430 * which overlap the truncated area starting at pgoff,
431 * and no vma on a 32-bit arch can span beyond the 4GB.
433 static unsigned long vma_offset_start(struct vm_area_struct *vma, pgoff_t start)
435 unsigned long offset = 0;
437 if (vma->vm_pgoff < start)
438 offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
440 return vma->vm_start + offset;
443 static unsigned long vma_offset_end(struct vm_area_struct *vma, pgoff_t end)
450 t_end = ((end - vma->vm_pgoff) << PAGE_SHIFT) + vma->vm_start;
451 if (t_end > vma->vm_end)
457 * Called with hugetlb fault mutex held. Therefore, no more mappings to
458 * this folio can be created while executing the routine.
460 static void hugetlb_unmap_file_folio(struct hstate *h,
461 struct address_space *mapping,
462 struct folio *folio, pgoff_t index)
464 struct rb_root_cached *root = &mapping->i_mmap;
465 struct hugetlb_vma_lock *vma_lock;
466 struct page *page = &folio->page;
467 struct vm_area_struct *vma;
468 unsigned long v_start;
472 start = index * pages_per_huge_page(h);
473 end = (index + 1) * pages_per_huge_page(h);
475 i_mmap_lock_write(mapping);
478 vma_interval_tree_foreach(vma, root, start, end - 1) {
479 v_start = vma_offset_start(vma, start);
480 v_end = vma_offset_end(vma, end);
482 if (!hugetlb_vma_maps_page(vma, v_start, page))
485 if (!hugetlb_vma_trylock_write(vma)) {
486 vma_lock = vma->vm_private_data;
488 * If we can not get vma lock, we need to drop
489 * immap_sema and take locks in order. First,
490 * take a ref on the vma_lock structure so that
491 * we can be guaranteed it will not go away when
492 * dropping immap_sema.
494 kref_get(&vma_lock->refs);
498 unmap_hugepage_range(vma, v_start, v_end, NULL,
499 ZAP_FLAG_DROP_MARKER);
500 hugetlb_vma_unlock_write(vma);
503 i_mmap_unlock_write(mapping);
507 * Wait on vma_lock. We know it is still valid as we have
508 * a reference. We must 'open code' vma locking as we do
509 * not know if vma_lock is still attached to vma.
511 down_write(&vma_lock->rw_sema);
512 i_mmap_lock_write(mapping);
517 * If lock is no longer attached to vma, then just
518 * unlock, drop our reference and retry looking for
521 up_write(&vma_lock->rw_sema);
522 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
527 * vma_lock is still attached to vma. Check to see if vma
528 * still maps page and if so, unmap.
530 v_start = vma_offset_start(vma, start);
531 v_end = vma_offset_end(vma, end);
532 if (hugetlb_vma_maps_page(vma, v_start, page))
533 unmap_hugepage_range(vma, v_start, v_end, NULL,
534 ZAP_FLAG_DROP_MARKER);
536 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
537 hugetlb_vma_unlock_write(vma);
544 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
545 zap_flags_t zap_flags)
547 struct vm_area_struct *vma;
550 * end == 0 indicates that the entire range after start should be
551 * unmapped. Note, end is exclusive, whereas the interval tree takes
552 * an inclusive "last".
554 vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
555 unsigned long v_start;
558 if (!hugetlb_vma_trylock_write(vma))
561 v_start = vma_offset_start(vma, start);
562 v_end = vma_offset_end(vma, end);
564 unmap_hugepage_range(vma, v_start, v_end, NULL, zap_flags);
567 * Note that vma lock only exists for shared/non-private
568 * vmas. Therefore, lock is not held when calling
569 * unmap_hugepage_range for private vmas.
571 hugetlb_vma_unlock_write(vma);
576 * Called with hugetlb fault mutex held.
577 * Returns true if page was actually removed, false otherwise.
579 static bool remove_inode_single_folio(struct hstate *h, struct inode *inode,
580 struct address_space *mapping,
581 struct folio *folio, pgoff_t index,
587 * If folio is mapped, it was faulted in after being
588 * unmapped in caller. Unmap (again) while holding
589 * the fault mutex. The mutex will prevent faults
590 * until we finish removing the folio.
592 if (unlikely(folio_mapped(folio)))
593 hugetlb_unmap_file_folio(h, mapping, folio, index);
597 * We must remove the folio from page cache before removing
598 * the region/ reserve map (hugetlb_unreserve_pages). In
599 * rare out of memory conditions, removal of the region/reserve
600 * map could fail. Correspondingly, the subpool and global
601 * reserve usage count can need to be adjusted.
603 VM_BUG_ON_FOLIO(folio_test_hugetlb_restore_reserve(folio), folio);
604 hugetlb_delete_from_page_cache(folio);
607 if (unlikely(hugetlb_unreserve_pages(inode, index,
609 hugetlb_fix_reserve_counts(inode);
617 * remove_inode_hugepages handles two distinct cases: truncation and hole
618 * punch. There are subtle differences in operation for each case.
620 * truncation is indicated by end of range being LLONG_MAX
621 * In this case, we first scan the range and release found pages.
622 * After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
623 * maps and global counts. Page faults can race with truncation.
624 * During faults, hugetlb_no_page() checks i_size before page allocation,
625 * and again after obtaining page table lock. It will 'back out'
626 * allocations in the truncated range.
627 * hole punch is indicated if end is not LLONG_MAX
628 * In the hole punch case we scan the range and release found pages.
629 * Only when releasing a page is the associated region/reserve map
630 * deleted. The region/reserve map for ranges without associated
631 * pages are not modified. Page faults can race with hole punch.
632 * This is indicated if we find a mapped page.
633 * Note: If the passed end of range value is beyond the end of file, but
634 * not LLONG_MAX this routine still performs a hole punch operation.
636 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
639 struct hstate *h = hstate_inode(inode);
640 struct address_space *mapping = &inode->i_data;
641 const pgoff_t end = lend >> PAGE_SHIFT;
642 struct folio_batch fbatch;
645 bool truncate_op = (lend == LLONG_MAX);
647 folio_batch_init(&fbatch);
648 next = lstart >> PAGE_SHIFT;
649 while (filemap_get_folios(mapping, &next, end - 1, &fbatch)) {
650 for (i = 0; i < folio_batch_count(&fbatch); ++i) {
651 struct folio *folio = fbatch.folios[i];
654 index = folio->index >> huge_page_order(h);
655 hash = hugetlb_fault_mutex_hash(mapping, index);
656 mutex_lock(&hugetlb_fault_mutex_table[hash]);
659 * Remove folio that was part of folio_batch.
661 if (remove_inode_single_folio(h, inode, mapping, folio,
665 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
667 folio_batch_release(&fbatch);
672 (void)hugetlb_unreserve_pages(inode,
673 lstart >> huge_page_shift(h),
677 static void hugetlbfs_evict_inode(struct inode *inode)
679 struct resv_map *resv_map;
681 remove_inode_hugepages(inode, 0, LLONG_MAX);
684 * Get the resv_map from the address space embedded in the inode.
685 * This is the address space which points to any resv_map allocated
686 * at inode creation time. If this is a device special inode,
687 * i_mapping may not point to the original address space.
689 resv_map = (struct resv_map *)(&inode->i_data)->i_private_data;
690 /* Only regular and link inodes have associated reserve maps */
692 resv_map_release(&resv_map->refs);
696 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
699 struct address_space *mapping = inode->i_mapping;
700 struct hstate *h = hstate_inode(inode);
702 BUG_ON(offset & ~huge_page_mask(h));
703 pgoff = offset >> PAGE_SHIFT;
705 i_size_write(inode, offset);
706 i_mmap_lock_write(mapping);
707 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
708 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
709 ZAP_FLAG_DROP_MARKER);
710 i_mmap_unlock_write(mapping);
711 remove_inode_hugepages(inode, offset, LLONG_MAX);
714 static void hugetlbfs_zero_partial_page(struct hstate *h,
715 struct address_space *mapping,
719 pgoff_t idx = start >> huge_page_shift(h);
722 folio = filemap_lock_hugetlb_folio(h, mapping, idx);
726 start = start & ~huge_page_mask(h);
727 end = end & ~huge_page_mask(h);
729 end = huge_page_size(h);
731 folio_zero_segment(folio, (size_t)start, (size_t)end);
737 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
739 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
740 struct address_space *mapping = inode->i_mapping;
741 struct hstate *h = hstate_inode(inode);
742 loff_t hpage_size = huge_page_size(h);
743 loff_t hole_start, hole_end;
746 * hole_start and hole_end indicate the full pages within the hole.
748 hole_start = round_up(offset, hpage_size);
749 hole_end = round_down(offset + len, hpage_size);
753 /* protected by i_rwsem */
754 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
759 i_mmap_lock_write(mapping);
761 /* If range starts before first full page, zero partial page. */
762 if (offset < hole_start)
763 hugetlbfs_zero_partial_page(h, mapping,
764 offset, min(offset + len, hole_start));
766 /* Unmap users of full pages in the hole. */
767 if (hole_end > hole_start) {
768 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
769 hugetlb_vmdelete_list(&mapping->i_mmap,
770 hole_start >> PAGE_SHIFT,
771 hole_end >> PAGE_SHIFT, 0);
774 /* If range extends beyond last full page, zero partial page. */
775 if ((offset + len) > hole_end && (offset + len) > hole_start)
776 hugetlbfs_zero_partial_page(h, mapping,
777 hole_end, offset + len);
779 i_mmap_unlock_write(mapping);
781 /* Remove full pages from the file. */
782 if (hole_end > hole_start)
783 remove_inode_hugepages(inode, hole_start, hole_end);
790 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
793 struct inode *inode = file_inode(file);
794 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
795 struct address_space *mapping = inode->i_mapping;
796 struct hstate *h = hstate_inode(inode);
797 struct vm_area_struct pseudo_vma;
798 struct mm_struct *mm = current->mm;
799 loff_t hpage_size = huge_page_size(h);
800 unsigned long hpage_shift = huge_page_shift(h);
801 pgoff_t start, index, end;
805 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
808 if (mode & FALLOC_FL_PUNCH_HOLE)
809 return hugetlbfs_punch_hole(inode, offset, len);
812 * Default preallocate case.
813 * For this range, start is rounded down and end is rounded up
814 * as well as being converted to page offsets.
816 start = offset >> hpage_shift;
817 end = (offset + len + hpage_size - 1) >> hpage_shift;
821 /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
822 error = inode_newsize_ok(inode, offset + len);
826 if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
832 * Initialize a pseudo vma as this is required by the huge page
833 * allocation routines.
835 vma_init(&pseudo_vma, mm);
836 vm_flags_init(&pseudo_vma, VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
837 pseudo_vma.vm_file = file;
839 for (index = start; index < end; index++) {
841 * This is supposed to be the vaddr where the page is being
842 * faulted in, but we have no vaddr here.
850 * fallocate(2) manpage permits EINTR; we may have been
851 * interrupted because we are using up too much memory.
853 if (signal_pending(current)) {
858 /* addr is the offset within the file (zero based) */
859 addr = index * hpage_size;
861 /* mutex taken here, fault path and hole punch */
862 hash = hugetlb_fault_mutex_hash(mapping, index);
863 mutex_lock(&hugetlb_fault_mutex_table[hash]);
865 /* See if already present in mapping to avoid alloc/free */
866 folio = filemap_get_folio(mapping, index << huge_page_order(h));
867 if (!IS_ERR(folio)) {
869 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
874 * Allocate folio without setting the avoid_reserve argument.
875 * There certainly are no reserves associated with the
876 * pseudo_vma. However, there could be shared mappings with
877 * reserves for the file at the inode level. If we fallocate
878 * folios in these areas, we need to consume the reserves
879 * to keep reservation accounting consistent.
881 folio = alloc_hugetlb_folio(&pseudo_vma, addr, 0);
883 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
884 error = PTR_ERR(folio);
887 clear_huge_page(&folio->page, addr, pages_per_huge_page(h));
888 __folio_mark_uptodate(folio);
889 error = hugetlb_add_to_page_cache(folio, mapping, index);
890 if (unlikely(error)) {
891 restore_reserve_on_error(h, &pseudo_vma, addr, folio);
893 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
897 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
899 folio_set_hugetlb_migratable(folio);
901 * folio_unlock because locked by hugetlb_add_to_page_cache()
902 * folio_put() due to reference from alloc_hugetlb_folio()
908 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
909 i_size_write(inode, offset + len);
910 inode_set_ctime_current(inode);
916 static int hugetlbfs_setattr(struct mnt_idmap *idmap,
917 struct dentry *dentry, struct iattr *attr)
919 struct inode *inode = d_inode(dentry);
920 struct hstate *h = hstate_inode(inode);
922 unsigned int ia_valid = attr->ia_valid;
923 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
925 error = setattr_prepare(&nop_mnt_idmap, dentry, attr);
929 if (ia_valid & ATTR_SIZE) {
930 loff_t oldsize = inode->i_size;
931 loff_t newsize = attr->ia_size;
933 if (newsize & ~huge_page_mask(h))
935 /* protected by i_rwsem */
936 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
937 (newsize > oldsize && (info->seals & F_SEAL_GROW)))
939 hugetlb_vmtruncate(inode, newsize);
942 setattr_copy(&nop_mnt_idmap, inode, attr);
943 mark_inode_dirty(inode);
947 static struct inode *hugetlbfs_get_root(struct super_block *sb,
948 struct hugetlbfs_fs_context *ctx)
952 inode = new_inode(sb);
954 inode->i_ino = get_next_ino();
955 inode->i_mode = S_IFDIR | ctx->mode;
956 inode->i_uid = ctx->uid;
957 inode->i_gid = ctx->gid;
958 simple_inode_init_ts(inode);
959 inode->i_op = &hugetlbfs_dir_inode_operations;
960 inode->i_fop = &simple_dir_operations;
961 /* directory inodes start off with i_nlink == 2 (for "." entry) */
963 lockdep_annotate_inode_mutex_key(inode);
969 * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
970 * be taken from reclaim -- unlike regular filesystems. This needs an
971 * annotation because huge_pmd_share() does an allocation under hugetlb's
974 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
976 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
978 umode_t mode, dev_t dev)
981 struct resv_map *resv_map = NULL;
984 * Reserve maps are only needed for inodes that can have associated
987 if (S_ISREG(mode) || S_ISLNK(mode)) {
988 resv_map = resv_map_alloc();
993 inode = new_inode(sb);
995 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
997 inode->i_ino = get_next_ino();
998 inode_init_owner(&nop_mnt_idmap, inode, dir, mode);
999 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
1000 &hugetlbfs_i_mmap_rwsem_key);
1001 inode->i_mapping->a_ops = &hugetlbfs_aops;
1002 simple_inode_init_ts(inode);
1003 inode->i_mapping->i_private_data = resv_map;
1004 info->seals = F_SEAL_SEAL;
1005 switch (mode & S_IFMT) {
1007 init_special_inode(inode, mode, dev);
1010 inode->i_op = &hugetlbfs_inode_operations;
1011 inode->i_fop = &hugetlbfs_file_operations;
1014 inode->i_op = &hugetlbfs_dir_inode_operations;
1015 inode->i_fop = &simple_dir_operations;
1017 /* directory inodes start off with i_nlink == 2 (for "." entry) */
1021 inode->i_op = &page_symlink_inode_operations;
1022 inode_nohighmem(inode);
1025 lockdep_annotate_inode_mutex_key(inode);
1028 kref_put(&resv_map->refs, resv_map_release);
1035 * File creation. Allocate an inode, and we're done..
1037 static int hugetlbfs_mknod(struct mnt_idmap *idmap, struct inode *dir,
1038 struct dentry *dentry, umode_t mode, dev_t dev)
1040 struct inode *inode;
1042 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
1045 inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1046 d_instantiate(dentry, inode);
1047 dget(dentry);/* Extra count - pin the dentry in core */
1051 static int hugetlbfs_mkdir(struct mnt_idmap *idmap, struct inode *dir,
1052 struct dentry *dentry, umode_t mode)
1054 int retval = hugetlbfs_mknod(&nop_mnt_idmap, dir, dentry,
1061 static int hugetlbfs_create(struct mnt_idmap *idmap,
1062 struct inode *dir, struct dentry *dentry,
1063 umode_t mode, bool excl)
1065 return hugetlbfs_mknod(&nop_mnt_idmap, dir, dentry, mode | S_IFREG, 0);
1068 static int hugetlbfs_tmpfile(struct mnt_idmap *idmap,
1069 struct inode *dir, struct file *file,
1072 struct inode *inode;
1074 inode = hugetlbfs_get_inode(dir->i_sb, dir, mode | S_IFREG, 0);
1077 inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1078 d_tmpfile(file, inode);
1079 return finish_open_simple(file, 0);
1082 static int hugetlbfs_symlink(struct mnt_idmap *idmap,
1083 struct inode *dir, struct dentry *dentry,
1084 const char *symname)
1086 struct inode *inode;
1087 int error = -ENOSPC;
1089 inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
1091 int l = strlen(symname)+1;
1092 error = page_symlink(inode, symname, l);
1094 d_instantiate(dentry, inode);
1099 inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
1104 #ifdef CONFIG_MIGRATION
1105 static int hugetlbfs_migrate_folio(struct address_space *mapping,
1106 struct folio *dst, struct folio *src,
1107 enum migrate_mode mode)
1111 rc = migrate_huge_page_move_mapping(mapping, dst, src);
1112 if (rc != MIGRATEPAGE_SUCCESS)
1115 if (hugetlb_folio_subpool(src)) {
1116 hugetlb_set_folio_subpool(dst,
1117 hugetlb_folio_subpool(src));
1118 hugetlb_set_folio_subpool(src, NULL);
1121 if (mode != MIGRATE_SYNC_NO_COPY)
1122 folio_migrate_copy(dst, src);
1124 folio_migrate_flags(dst, src);
1126 return MIGRATEPAGE_SUCCESS;
1129 #define hugetlbfs_migrate_folio NULL
1132 static int hugetlbfs_error_remove_folio(struct address_space *mapping,
1133 struct folio *folio)
1139 * Display the mount options in /proc/mounts.
1141 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1143 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1144 struct hugepage_subpool *spool = sbinfo->spool;
1145 unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1146 unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1149 if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1150 seq_printf(m, ",uid=%u",
1151 from_kuid_munged(&init_user_ns, sbinfo->uid));
1152 if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1153 seq_printf(m, ",gid=%u",
1154 from_kgid_munged(&init_user_ns, sbinfo->gid));
1155 if (sbinfo->mode != 0755)
1156 seq_printf(m, ",mode=%o", sbinfo->mode);
1157 if (sbinfo->max_inodes != -1)
1158 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1162 if (hpage_size >= 1024) {
1166 seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1168 if (spool->max_hpages != -1)
1169 seq_printf(m, ",size=%llu",
1170 (unsigned long long)spool->max_hpages << hpage_shift);
1171 if (spool->min_hpages != -1)
1172 seq_printf(m, ",min_size=%llu",
1173 (unsigned long long)spool->min_hpages << hpage_shift);
1178 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1180 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1181 struct hstate *h = hstate_inode(d_inode(dentry));
1182 u64 id = huge_encode_dev(dentry->d_sb->s_dev);
1184 buf->f_fsid = u64_to_fsid(id);
1185 buf->f_type = HUGETLBFS_MAGIC;
1186 buf->f_bsize = huge_page_size(h);
1188 spin_lock(&sbinfo->stat_lock);
1189 /* If no limits set, just report 0 or -1 for max/free/used
1190 * blocks, like simple_statfs() */
1191 if (sbinfo->spool) {
1194 spin_lock_irq(&sbinfo->spool->lock);
1195 buf->f_blocks = sbinfo->spool->max_hpages;
1196 free_pages = sbinfo->spool->max_hpages
1197 - sbinfo->spool->used_hpages;
1198 buf->f_bavail = buf->f_bfree = free_pages;
1199 spin_unlock_irq(&sbinfo->spool->lock);
1200 buf->f_files = sbinfo->max_inodes;
1201 buf->f_ffree = sbinfo->free_inodes;
1203 spin_unlock(&sbinfo->stat_lock);
1205 buf->f_namelen = NAME_MAX;
1209 static void hugetlbfs_put_super(struct super_block *sb)
1211 struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1214 sb->s_fs_info = NULL;
1217 hugepage_put_subpool(sbi->spool);
1223 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1225 if (sbinfo->free_inodes >= 0) {
1226 spin_lock(&sbinfo->stat_lock);
1227 if (unlikely(!sbinfo->free_inodes)) {
1228 spin_unlock(&sbinfo->stat_lock);
1231 sbinfo->free_inodes--;
1232 spin_unlock(&sbinfo->stat_lock);
1238 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1240 if (sbinfo->free_inodes >= 0) {
1241 spin_lock(&sbinfo->stat_lock);
1242 sbinfo->free_inodes++;
1243 spin_unlock(&sbinfo->stat_lock);
1248 static struct kmem_cache *hugetlbfs_inode_cachep;
1250 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1252 struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1253 struct hugetlbfs_inode_info *p;
1255 if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1257 p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1259 hugetlbfs_inc_free_inodes(sbinfo);
1262 return &p->vfs_inode;
1265 static void hugetlbfs_free_inode(struct inode *inode)
1267 kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1270 static void hugetlbfs_destroy_inode(struct inode *inode)
1272 hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1275 static const struct address_space_operations hugetlbfs_aops = {
1276 .write_begin = hugetlbfs_write_begin,
1277 .write_end = hugetlbfs_write_end,
1278 .dirty_folio = noop_dirty_folio,
1279 .migrate_folio = hugetlbfs_migrate_folio,
1280 .error_remove_folio = hugetlbfs_error_remove_folio,
1284 static void init_once(void *foo)
1286 struct hugetlbfs_inode_info *ei = foo;
1288 inode_init_once(&ei->vfs_inode);
1291 const struct file_operations hugetlbfs_file_operations = {
1292 .read_iter = hugetlbfs_read_iter,
1293 .mmap = hugetlbfs_file_mmap,
1294 .fsync = noop_fsync,
1295 .get_unmapped_area = hugetlb_get_unmapped_area,
1296 .llseek = default_llseek,
1297 .fallocate = hugetlbfs_fallocate,
1300 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1301 .create = hugetlbfs_create,
1302 .lookup = simple_lookup,
1303 .link = simple_link,
1304 .unlink = simple_unlink,
1305 .symlink = hugetlbfs_symlink,
1306 .mkdir = hugetlbfs_mkdir,
1307 .rmdir = simple_rmdir,
1308 .mknod = hugetlbfs_mknod,
1309 .rename = simple_rename,
1310 .setattr = hugetlbfs_setattr,
1311 .tmpfile = hugetlbfs_tmpfile,
1314 static const struct inode_operations hugetlbfs_inode_operations = {
1315 .setattr = hugetlbfs_setattr,
1318 static const struct super_operations hugetlbfs_ops = {
1319 .alloc_inode = hugetlbfs_alloc_inode,
1320 .free_inode = hugetlbfs_free_inode,
1321 .destroy_inode = hugetlbfs_destroy_inode,
1322 .evict_inode = hugetlbfs_evict_inode,
1323 .statfs = hugetlbfs_statfs,
1324 .put_super = hugetlbfs_put_super,
1325 .show_options = hugetlbfs_show_options,
1329 * Convert size option passed from command line to number of huge pages
1330 * in the pool specified by hstate. Size option could be in bytes
1331 * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1334 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1335 enum hugetlbfs_size_type val_type)
1337 if (val_type == NO_SIZE)
1340 if (val_type == SIZE_PERCENT) {
1341 size_opt <<= huge_page_shift(h);
1342 size_opt *= h->max_huge_pages;
1343 do_div(size_opt, 100);
1346 size_opt >>= huge_page_shift(h);
1351 * Parse one mount parameter.
1353 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1355 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1356 struct fs_parse_result result;
1361 opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1367 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1368 if (!uid_valid(ctx->uid))
1373 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1374 if (!gid_valid(ctx->gid))
1379 ctx->mode = result.uint_32 & 01777U;
1383 /* memparse() will accept a K/M/G without a digit */
1384 if (!param->string || !isdigit(param->string[0]))
1386 ctx->max_size_opt = memparse(param->string, &rest);
1387 ctx->max_val_type = SIZE_STD;
1389 ctx->max_val_type = SIZE_PERCENT;
1393 /* memparse() will accept a K/M/G without a digit */
1394 if (!param->string || !isdigit(param->string[0]))
1396 ctx->nr_inodes = memparse(param->string, &rest);
1400 ps = memparse(param->string, &rest);
1401 ctx->hstate = size_to_hstate(ps);
1403 pr_err("Unsupported page size %lu MB\n", ps / SZ_1M);
1409 /* memparse() will accept a K/M/G without a digit */
1410 if (!param->string || !isdigit(param->string[0]))
1412 ctx->min_size_opt = memparse(param->string, &rest);
1413 ctx->min_val_type = SIZE_STD;
1415 ctx->min_val_type = SIZE_PERCENT;
1423 return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1424 param->string, param->key);
1428 * Validate the parsed options.
1430 static int hugetlbfs_validate(struct fs_context *fc)
1432 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1435 * Use huge page pool size (in hstate) to convert the size
1436 * options to number of huge pages. If NO_SIZE, -1 is returned.
1438 ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1441 ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1446 * If max_size was specified, then min_size must be smaller
1448 if (ctx->max_val_type > NO_SIZE &&
1449 ctx->min_hpages > ctx->max_hpages) {
1450 pr_err("Minimum size can not be greater than maximum size\n");
1458 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1460 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1461 struct hugetlbfs_sb_info *sbinfo;
1463 sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1466 sb->s_fs_info = sbinfo;
1467 spin_lock_init(&sbinfo->stat_lock);
1468 sbinfo->hstate = ctx->hstate;
1469 sbinfo->max_inodes = ctx->nr_inodes;
1470 sbinfo->free_inodes = ctx->nr_inodes;
1471 sbinfo->spool = NULL;
1472 sbinfo->uid = ctx->uid;
1473 sbinfo->gid = ctx->gid;
1474 sbinfo->mode = ctx->mode;
1477 * Allocate and initialize subpool if maximum or minimum size is
1478 * specified. Any needed reservations (for minimum size) are taken
1479 * when the subpool is created.
1481 if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1482 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1488 sb->s_maxbytes = MAX_LFS_FILESIZE;
1489 sb->s_blocksize = huge_page_size(ctx->hstate);
1490 sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1491 sb->s_magic = HUGETLBFS_MAGIC;
1492 sb->s_op = &hugetlbfs_ops;
1493 sb->s_time_gran = 1;
1496 * Due to the special and limited functionality of hugetlbfs, it does
1497 * not work well as a stacking filesystem.
1499 sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1500 sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1505 kfree(sbinfo->spool);
1510 static int hugetlbfs_get_tree(struct fs_context *fc)
1512 int err = hugetlbfs_validate(fc);
1515 return get_tree_nodev(fc, hugetlbfs_fill_super);
1518 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1520 kfree(fc->fs_private);
1523 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1524 .free = hugetlbfs_fs_context_free,
1525 .parse_param = hugetlbfs_parse_param,
1526 .get_tree = hugetlbfs_get_tree,
1529 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1531 struct hugetlbfs_fs_context *ctx;
1533 ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1537 ctx->max_hpages = -1; /* No limit on size by default */
1538 ctx->nr_inodes = -1; /* No limit on number of inodes by default */
1539 ctx->uid = current_fsuid();
1540 ctx->gid = current_fsgid();
1542 ctx->hstate = &default_hstate;
1543 ctx->min_hpages = -1; /* No default minimum size */
1544 ctx->max_val_type = NO_SIZE;
1545 ctx->min_val_type = NO_SIZE;
1546 fc->fs_private = ctx;
1547 fc->ops = &hugetlbfs_fs_context_ops;
1551 static struct file_system_type hugetlbfs_fs_type = {
1552 .name = "hugetlbfs",
1553 .init_fs_context = hugetlbfs_init_fs_context,
1554 .parameters = hugetlb_fs_parameters,
1555 .kill_sb = kill_litter_super,
1558 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1560 static int can_do_hugetlb_shm(void)
1563 shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1564 return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1567 static int get_hstate_idx(int page_size_log)
1569 struct hstate *h = hstate_sizelog(page_size_log);
1573 return hstate_index(h);
1577 * Note that size should be aligned to proper hugepage size in caller side,
1578 * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1580 struct file *hugetlb_file_setup(const char *name, size_t size,
1581 vm_flags_t acctflag, int creat_flags,
1584 struct inode *inode;
1585 struct vfsmount *mnt;
1589 hstate_idx = get_hstate_idx(page_size_log);
1591 return ERR_PTR(-ENODEV);
1593 mnt = hugetlbfs_vfsmount[hstate_idx];
1595 return ERR_PTR(-ENOENT);
1597 if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1598 struct ucounts *ucounts = current_ucounts();
1600 if (user_shm_lock(size, ucounts)) {
1601 pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1602 current->comm, current->pid);
1603 user_shm_unlock(size, ucounts);
1605 return ERR_PTR(-EPERM);
1608 file = ERR_PTR(-ENOSPC);
1609 inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1612 if (creat_flags == HUGETLB_SHMFS_INODE)
1613 inode->i_flags |= S_PRIVATE;
1615 inode->i_size = size;
1618 if (!hugetlb_reserve_pages(inode, 0,
1619 size >> huge_page_shift(hstate_inode(inode)), NULL,
1621 file = ERR_PTR(-ENOMEM);
1623 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1624 &hugetlbfs_file_operations);
1633 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1635 struct fs_context *fc;
1636 struct vfsmount *mnt;
1638 fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1642 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1648 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1649 huge_page_size(h) / SZ_1K);
1653 static int __init init_hugetlbfs_fs(void)
1655 struct vfsmount *mnt;
1660 if (!hugepages_supported()) {
1661 pr_info("disabling because there are no supported hugepage sizes\n");
1666 hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1667 sizeof(struct hugetlbfs_inode_info),
1668 0, SLAB_ACCOUNT, init_once);
1669 if (hugetlbfs_inode_cachep == NULL)
1672 error = register_filesystem(&hugetlbfs_fs_type);
1676 /* default hstate mount is required */
1677 mnt = mount_one_hugetlbfs(&default_hstate);
1679 error = PTR_ERR(mnt);
1682 hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1684 /* other hstates are optional */
1686 for_each_hstate(h) {
1687 if (i == default_hstate_idx) {
1692 mnt = mount_one_hugetlbfs(h);
1694 hugetlbfs_vfsmount[i] = NULL;
1696 hugetlbfs_vfsmount[i] = mnt;
1703 (void)unregister_filesystem(&hugetlbfs_fs_type);
1705 kmem_cache_destroy(hugetlbfs_inode_cachep);
1709 fs_initcall(init_hugetlbfs_fs)