5 The text below describes the locking rules for VFS-related methods.
6 It is (believed to be) up-to-date. *Please*, if you change anything in
7 prototypes or locking protocols - update this file. And update the relevant
8 instances in the tree, don't leave that to maintainers of filesystems/devices/
9 etc. At the very least, put the list of dubious cases in the end of this file.
10 Don't turn it into log - maintainers of out-of-the-tree code are supposed to
11 be able to use diff(1).
13 Thing currently missing here: socket operations. Alexey?
20 int (*d_revalidate)(struct dentry *, unsigned int);
21 int (*d_weak_revalidate)(struct dentry *, unsigned int);
22 int (*d_hash)(const struct dentry *, struct qstr *);
23 int (*d_compare)(const struct dentry *,
24 unsigned int, const char *, const struct qstr *);
25 int (*d_delete)(struct dentry *);
26 int (*d_init)(struct dentry *);
27 void (*d_release)(struct dentry *);
28 void (*d_iput)(struct dentry *, struct inode *);
29 char *(*d_dname)((struct dentry *dentry, char *buffer, int buflen);
30 struct vfsmount *(*d_automount)(struct path *path);
31 int (*d_manage)(const struct path *, bool);
32 struct dentry *(*d_real)(struct dentry *, const struct inode *);
36 ================== =========== ======== ============== ========
37 ops rename_lock ->d_lock may block rcu-walk
38 ================== =========== ======== ============== ========
39 d_revalidate: no no yes (ref-walk) maybe
40 d_weak_revalidate: no no yes no
42 d_compare: yes no no maybe
43 d_delete: no yes no no
45 d_release: no no yes no
49 d_automount: no no yes no
50 d_manage: no no yes (ref-walk) maybe
52 ================== =========== ======== ============== ========
59 int (*create) (struct inode *,struct dentry *,umode_t, bool);
60 struct dentry * (*lookup) (struct inode *,struct dentry *, unsigned int);
61 int (*link) (struct dentry *,struct inode *,struct dentry *);
62 int (*unlink) (struct inode *,struct dentry *);
63 int (*symlink) (struct inode *,struct dentry *,const char *);
64 int (*mkdir) (struct inode *,struct dentry *,umode_t);
65 int (*rmdir) (struct inode *,struct dentry *);
66 int (*mknod) (struct inode *,struct dentry *,umode_t,dev_t);
67 int (*rename) (struct inode *, struct dentry *,
68 struct inode *, struct dentry *, unsigned int);
69 int (*readlink) (struct dentry *, char __user *,int);
70 const char *(*get_link) (struct dentry *, struct inode *, struct delayed_call *);
71 void (*truncate) (struct inode *);
72 int (*permission) (struct inode *, int, unsigned int);
73 struct posix_acl * (*get_acl)(struct inode *, int, bool);
74 int (*setattr) (struct dentry *, struct iattr *);
75 int (*getattr) (const struct path *, struct kstat *, u32, unsigned int);
76 ssize_t (*listxattr) (struct dentry *, char *, size_t);
77 int (*fiemap)(struct inode *, struct fiemap_extent_info *, u64 start, u64 len);
78 void (*update_time)(struct inode *, struct timespec *, int);
79 int (*atomic_open)(struct inode *, struct dentry *,
80 struct file *, unsigned open_flag,
82 int (*tmpfile) (struct user_namespace *, struct inode *,
83 struct file *, umode_t);
84 int (*fileattr_set)(struct user_namespace *mnt_userns,
85 struct dentry *dentry, struct fileattr *fa);
86 int (*fileattr_get)(struct dentry *dentry, struct fileattr *fa);
91 ============= =============================================
93 ============= =============================================
96 link: exclusive (both)
100 unlink: exclusive (both)
101 rmdir: exclusive (both)(see below)
102 rename: exclusive (all) (see below)
106 permission: no (may not block if called in rcu-walk mode)
112 atomic_open: shared (exclusive if O_CREAT is set in open flags)
114 fileattr_get: no or exclusive
115 fileattr_set: exclusive
116 ============= =============================================
119 Additionally, ->rmdir(), ->unlink() and ->rename() have ->i_rwsem
121 cross-directory ->rename() has (per-superblock) ->s_vfs_rename_sem.
123 See Documentation/filesystems/directory-locking.rst for more detailed discussion
124 of the locking scheme for directory operations.
126 xattr_handler operations
127 ========================
131 bool (*list)(struct dentry *dentry);
132 int (*get)(const struct xattr_handler *handler, struct dentry *dentry,
133 struct inode *inode, const char *name, void *buffer,
135 int (*set)(const struct xattr_handler *handler,
136 struct user_namespace *mnt_userns,
137 struct dentry *dentry, struct inode *inode, const char *name,
138 const void *buffer, size_t size, int flags);
156 struct inode *(*alloc_inode)(struct super_block *sb);
157 void (*free_inode)(struct inode *);
158 void (*destroy_inode)(struct inode *);
159 void (*dirty_inode) (struct inode *, int flags);
160 int (*write_inode) (struct inode *, struct writeback_control *wbc);
161 int (*drop_inode) (struct inode *);
162 void (*evict_inode) (struct inode *);
163 void (*put_super) (struct super_block *);
164 int (*sync_fs)(struct super_block *sb, int wait);
165 int (*freeze_fs) (struct super_block *);
166 int (*unfreeze_fs) (struct super_block *);
167 int (*statfs) (struct dentry *, struct kstatfs *);
168 int (*remount_fs) (struct super_block *, int *, char *);
169 void (*umount_begin) (struct super_block *);
170 int (*show_options)(struct seq_file *, struct dentry *);
171 ssize_t (*quota_read)(struct super_block *, int, char *, size_t, loff_t);
172 ssize_t (*quota_write)(struct super_block *, int, const char *, size_t, loff_t);
175 All may block [not true, see below]
177 ====================== ============ ========================
179 ====================== ============ ========================
181 free_inode: called from RCU callback
185 drop_inode: !!!inode->i_lock!!!
191 statfs: maybe(read) (see below)
194 show_options: no (namespace_sem)
195 quota_read: no (see below)
196 quota_write: no (see below)
197 ====================== ============ ========================
199 ->statfs() has s_umount (shared) when called by ustat(2) (native or
200 compat), but that's an accident of bad API; s_umount is used to pin
201 the superblock down when we only have dev_t given us by userland to
202 identify the superblock. Everything else (statfs(), fstatfs(), etc.)
203 doesn't hold it when calling ->statfs() - superblock is pinned down
204 by resolving the pathname passed to syscall.
206 ->quota_read() and ->quota_write() functions are both guaranteed to
207 be the only ones operating on the quota file by the quota code (via
208 dqio_sem) (unless an admin really wants to screw up something and
209 writes to quota files with quotas on). For other details about locking
210 see also dquot_operations section.
217 struct dentry *(*mount) (struct file_system_type *, int,
218 const char *, void *);
219 void (*kill_sb) (struct super_block *);
230 ->mount() returns ERR_PTR or the root dentry; its superblock should be locked
233 ->kill_sb() takes a write-locked superblock, does all shutdown work on it,
234 unlocks and drops the reference.
236 address_space_operations
237 ========================
240 int (*writepage)(struct page *page, struct writeback_control *wbc);
241 int (*read_folio)(struct file *, struct folio *);
242 int (*writepages)(struct address_space *, struct writeback_control *);
243 bool (*dirty_folio)(struct address_space *, struct folio *folio);
244 void (*readahead)(struct readahead_control *);
245 int (*write_begin)(struct file *, struct address_space *mapping,
246 loff_t pos, unsigned len,
247 struct page **pagep, void **fsdata);
248 int (*write_end)(struct file *, struct address_space *mapping,
249 loff_t pos, unsigned len, unsigned copied,
250 struct page *page, void *fsdata);
251 sector_t (*bmap)(struct address_space *, sector_t);
252 void (*invalidate_folio) (struct folio *, size_t start, size_t len);
253 bool (*release_folio)(struct folio *, gfp_t);
254 void (*free_folio)(struct folio *);
255 int (*direct_IO)(struct kiocb *, struct iov_iter *iter);
256 int (*migrate_folio)(struct address_space *, struct folio *dst,
257 struct folio *src, enum migrate_mode);
258 int (*launder_folio)(struct folio *);
259 bool (*is_partially_uptodate)(struct folio *, size_t from, size_t count);
260 int (*error_remove_page)(struct address_space *, struct page *);
261 int (*swap_activate)(struct swap_info_struct *sis, struct file *f, sector_t *span)
262 int (*swap_deactivate)(struct file *);
263 int (*swap_rw)(struct kiocb *iocb, struct iov_iter *iter);
266 All except dirty_folio and free_folio may block
268 ====================== ======================== ========= ===============
269 ops folio locked i_rwsem invalidate_lock
270 ====================== ======================== ========= ===============
271 writepage: yes, unlocks (see below)
272 read_folio: yes, unlocks shared
275 readahead: yes, unlocks shared
276 write_begin: locks the page exclusive
277 write_end: yes, unlocks exclusive
279 invalidate_folio: yes exclusive
283 migrate_folio: yes (both)
285 is_partially_uptodate: yes
286 error_remove_page: yes
289 swap_rw: yes, unlocks
290 ====================== ======================== ========= ===============
292 ->write_begin(), ->write_end() and ->read_folio() may be called from
293 the request handler (/dev/loop).
295 ->read_folio() unlocks the folio, either synchronously or via I/O
298 ->readahead() unlocks the folios that I/O is attempted on like ->read_folio().
300 ->writepage() is used for two purposes: for "memory cleansing" and for
301 "sync". These are quite different operations and the behaviour may differ
302 depending upon the mode.
304 If writepage is called for sync (wbc->sync_mode != WBC_SYNC_NONE) then
305 it *must* start I/O against the page, even if that would involve
306 blocking on in-progress I/O.
308 If writepage is called for memory cleansing (sync_mode ==
309 WBC_SYNC_NONE) then its role is to get as much writeout underway as
310 possible. So writepage should try to avoid blocking against
311 currently-in-progress I/O.
313 If the filesystem is not called for "sync" and it determines that it
314 would need to block against in-progress I/O to be able to start new I/O
315 against the page the filesystem should redirty the page with
316 redirty_page_for_writepage(), then unlock the page and return zero.
317 This may also be done to avoid internal deadlocks, but rarely.
319 If the filesystem is called for sync then it must wait on any
320 in-progress I/O and then start new I/O.
322 The filesystem should unlock the page synchronously, before returning to the
323 caller, unless ->writepage() returns special WRITEPAGE_ACTIVATE
324 value. WRITEPAGE_ACTIVATE means that page cannot really be written out
325 currently, and VM should stop calling ->writepage() on this page for some
326 time. VM does this by moving page to the head of the active list, hence the
329 Unless the filesystem is going to redirty_page_for_writepage(), unlock the page
330 and return zero, writepage *must* run set_page_writeback() against the page,
331 followed by unlocking it. Once set_page_writeback() has been run against the
332 page, write I/O can be submitted and the write I/O completion handler must run
333 end_page_writeback() once the I/O is complete. If no I/O is submitted, the
334 filesystem must run end_page_writeback() against the page before returning from
337 That is: after 2.5.12, pages which are under writeout are *not* locked. Note,
338 if the filesystem needs the page to be locked during writeout, that is ok, too,
339 the page is allowed to be unlocked at any point in time between the calls to
340 set_page_writeback() and end_page_writeback().
342 Note, failure to run either redirty_page_for_writepage() or the combination of
343 set_page_writeback()/end_page_writeback() on a page submitted to writepage
344 will leave the page itself marked clean but it will be tagged as dirty in the
345 radix tree. This incoherency can lead to all sorts of hard-to-debug problems
346 in the filesystem like having dirty inodes at umount and losing written data.
348 ->writepages() is used for periodic writeback and for syscall-initiated
349 sync operations. The address_space should start I/O against at least
350 ``*nr_to_write`` pages. ``*nr_to_write`` must be decremented for each page
351 which is written. The address_space implementation may write more (or less)
352 pages than ``*nr_to_write`` asks for, but it should try to be reasonably close.
353 If nr_to_write is NULL, all dirty pages must be written.
355 writepages should _only_ write pages which are present on
358 ->dirty_folio() is called from various places in the kernel when
359 the target folio is marked as needing writeback. The folio cannot be
360 truncated because either the caller holds the folio lock, or the caller
361 has found the folio while holding the page table lock which will block
364 ->bmap() is currently used by legacy ioctl() (FIBMAP) provided by some
365 filesystems and by the swapper. The latter will eventually go away. Please,
366 keep it that way and don't breed new callers.
368 ->invalidate_folio() is called when the filesystem must attempt to drop
369 some or all of the buffers from the page when it is being truncated. It
370 returns zero on success. The filesystem must exclusively acquire
371 invalidate_lock before invalidating page cache in truncate / hole punch
372 path (and thus calling into ->invalidate_folio) to block races between page
373 cache invalidation and page cache filling functions (fault, read, ...).
375 ->release_folio() is called when the kernel is about to try to drop the
376 buffers from the folio in preparation for freeing it. It returns false to
377 indicate that the buffers are (or may be) freeable. If ->release_folio is
378 NULL, the kernel assumes that the fs has no private interest in the buffers.
380 ->free_folio() is called when the kernel has dropped the folio
383 ->launder_folio() may be called prior to releasing a folio if
384 it is still found to be dirty. It returns zero if the folio was successfully
385 cleaned, or an error value if not. Note that in order to prevent the folio
386 getting mapped back in and redirtied, it needs to be kept locked
387 across the entire operation.
389 ->swap_activate() will be called to prepare the given file for swap. It
390 should perform any validation and preparation necessary to ensure that
391 writes can be performed with minimal memory allocation. It should call
392 add_swap_extent(), or the helper iomap_swapfile_activate(), and return
393 the number of extents added. If IO should be submitted through
394 ->swap_rw(), it should set SWP_FS_OPS, otherwise IO will be submitted
395 directly to the block device ``sis->bdev``.
397 ->swap_deactivate() will be called in the sys_swapoff()
398 path after ->swap_activate() returned success.
400 ->swap_rw will be called for swap IO if SWP_FS_OPS was set by ->swap_activate().
407 void (*fl_copy_lock)(struct file_lock *, struct file_lock *);
408 void (*fl_release_private)(struct file_lock *);
413 =================== ============= =========
414 ops inode->i_lock may block
415 =================== ============= =========
417 fl_release_private: maybe maybe[1]_
418 =================== ============= =========
421 ->fl_release_private for flock or POSIX locks is currently allowed
422 to block. Leases however can still be freed while the i_lock is held and
423 so fl_release_private called on a lease should not block.
425 lock_manager_operations
426 =======================
430 void (*lm_notify)(struct file_lock *); /* unblock callback */
431 int (*lm_grant)(struct file_lock *, struct file_lock *, int);
432 void (*lm_break)(struct file_lock *); /* break_lease callback */
433 int (*lm_change)(struct file_lock **, int);
434 bool (*lm_breaker_owns_lease)(struct file_lock *);
435 bool (*lm_lock_expirable)(struct file_lock *);
436 void (*lm_expire_lock)(void);
440 ====================== ============= ================= =========
441 ops flc_lock blocked_lock_lock may block
442 ====================== ============= ================= =========
447 lm_breaker_owns_lease: yes no no
448 lm_lock_expirable yes no no
449 lm_expire_lock no no yes
450 ====================== ============= ================= =========
457 void (*b_end_io)(struct buffer_head *bh, int uptodate);
461 called from interrupts. In other words, extreme care is needed here.
462 bh is locked, but that's all warranties we have here. Currently only RAID1,
463 highmem, fs/buffer.c, and fs/ntfs/aops.c are providing these. Block devices
464 call this method upon the IO completion.
466 block_device_operations
467 =======================
470 int (*open) (struct block_device *, fmode_t);
471 int (*release) (struct gendisk *, fmode_t);
472 int (*ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
473 int (*compat_ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
474 int (*direct_access) (struct block_device *, sector_t, void **,
476 void (*unlock_native_capacity) (struct gendisk *);
477 int (*getgeo)(struct block_device *, struct hd_geometry *);
478 void (*swap_slot_free_notify) (struct block_device *, unsigned long);
482 ======================= ===================
484 ======================= ===================
490 unlock_native_capacity: no
492 swap_slot_free_notify: no (see below)
493 ======================= ===================
495 swap_slot_free_notify is called with swap_lock and sometimes the page lock
504 loff_t (*llseek) (struct file *, loff_t, int);
505 ssize_t (*read) (struct file *, char __user *, size_t, loff_t *);
506 ssize_t (*write) (struct file *, const char __user *, size_t, loff_t *);
507 ssize_t (*read_iter) (struct kiocb *, struct iov_iter *);
508 ssize_t (*write_iter) (struct kiocb *, struct iov_iter *);
509 int (*iopoll) (struct kiocb *kiocb, bool spin);
510 int (*iterate) (struct file *, struct dir_context *);
511 int (*iterate_shared) (struct file *, struct dir_context *);
512 __poll_t (*poll) (struct file *, struct poll_table_struct *);
513 long (*unlocked_ioctl) (struct file *, unsigned int, unsigned long);
514 long (*compat_ioctl) (struct file *, unsigned int, unsigned long);
515 int (*mmap) (struct file *, struct vm_area_struct *);
516 int (*open) (struct inode *, struct file *);
517 int (*flush) (struct file *);
518 int (*release) (struct inode *, struct file *);
519 int (*fsync) (struct file *, loff_t start, loff_t end, int datasync);
520 int (*fasync) (int, struct file *, int);
521 int (*lock) (struct file *, int, struct file_lock *);
522 ssize_t (*sendpage) (struct file *, struct page *, int, size_t,
524 unsigned long (*get_unmapped_area)(struct file *, unsigned long,
525 unsigned long, unsigned long, unsigned long);
526 int (*check_flags)(int);
527 int (*flock) (struct file *, int, struct file_lock *);
528 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *, loff_t *,
529 size_t, unsigned int);
530 ssize_t (*splice_read)(struct file *, loff_t *, struct pipe_inode_info *,
531 size_t, unsigned int);
532 int (*setlease)(struct file *, long, struct file_lock **, void **);
533 long (*fallocate)(struct file *, int, loff_t, loff_t);
534 void (*show_fdinfo)(struct seq_file *m, struct file *f);
535 unsigned (*mmap_capabilities)(struct file *);
536 ssize_t (*copy_file_range)(struct file *, loff_t, struct file *,
537 loff_t, size_t, unsigned int);
538 loff_t (*remap_file_range)(struct file *file_in, loff_t pos_in,
539 struct file *file_out, loff_t pos_out,
540 loff_t len, unsigned int remap_flags);
541 int (*fadvise)(struct file *, loff_t, loff_t, int);
546 ->llseek() locking has moved from llseek to the individual llseek
547 implementations. If your fs is not using generic_file_llseek, you
548 need to acquire and release the appropriate locks in your ->llseek().
549 For many filesystems, it is probably safe to acquire the inode
550 mutex or just to use i_size_read() instead.
551 Note: this does not protect the file->f_pos against concurrent modifications
552 since this is something the userspace has to take care about.
554 ->iterate() is called with i_rwsem exclusive.
556 ->iterate_shared() is called with i_rwsem at least shared.
558 ->fasync() is responsible for maintaining the FASYNC bit in filp->f_flags.
559 Most instances call fasync_helper(), which does that maintenance, so it's
560 not normally something one needs to worry about. Return values > 0 will be
561 mapped to zero in the VFS layer.
563 ->readdir() and ->ioctl() on directories must be changed. Ideally we would
564 move ->readdir() to inode_operations and use a separate method for directory
565 ->ioctl() or kill the latter completely. One of the problems is that for
566 anything that resembles union-mount we won't have a struct file for all
567 components. And there are other reasons why the current interface is a mess...
569 ->read on directories probably must go away - we should just enforce -EISDIR
570 in sys_read() and friends.
572 ->setlease operations should call generic_setlease() before or after setting
573 the lease within the individual filesystem to record the result of the
576 ->fallocate implementation must be really careful to maintain page cache
577 consistency when punching holes or performing other operations that invalidate
578 page cache contents. Usually the filesystem needs to call
579 truncate_inode_pages_range() to invalidate relevant range of the page cache.
580 However the filesystem usually also needs to update its internal (and on disk)
581 view of file offset -> disk block mapping. Until this update is finished, the
582 filesystem needs to block page faults and reads from reloading now-stale page
583 cache contents from the disk. Since VFS acquires mapping->invalidate_lock in
584 shared mode when loading pages from disk (filemap_fault(), filemap_read(),
585 readahead paths), the fallocate implementation must take the invalidate_lock to
588 ->copy_file_range and ->remap_file_range implementations need to serialize
589 against modifications of file data while the operation is running. For
590 blocking changes through write(2) and similar operations inode->i_rwsem can be
591 used. To block changes to file contents via a memory mapping during the
592 operation, the filesystem must take mapping->invalidate_lock to coordinate
600 int (*write_dquot) (struct dquot *);
601 int (*acquire_dquot) (struct dquot *);
602 int (*release_dquot) (struct dquot *);
603 int (*mark_dirty) (struct dquot *);
604 int (*write_info) (struct super_block *, int);
606 These operations are intended to be more or less wrapping functions that ensure
607 a proper locking wrt the filesystem and call the generic quota operations.
609 What filesystem should expect from the generic quota functions:
611 ============== ============ =========================
612 ops FS recursion Held locks when called
613 ============== ============ =========================
614 write_dquot: yes dqonoff_sem or dqptr_sem
615 acquire_dquot: yes dqonoff_sem or dqptr_sem
616 release_dquot: yes dqonoff_sem or dqptr_sem
618 write_info: yes dqonoff_sem
619 ============== ============ =========================
621 FS recursion means calling ->quota_read() and ->quota_write() from superblock
624 More details about quota locking can be found in fs/dquot.c.
631 void (*open)(struct vm_area_struct*);
632 void (*close)(struct vm_area_struct*);
633 vm_fault_t (*fault)(struct vm_area_struct*, struct vm_fault *);
634 vm_fault_t (*page_mkwrite)(struct vm_area_struct *, struct vm_fault *);
635 vm_fault_t (*pfn_mkwrite)(struct vm_area_struct *, struct vm_fault *);
636 int (*access)(struct vm_area_struct *, unsigned long, void*, int, int);
640 ============= ========= ===========================
641 ops mmap_lock PageLocked(page)
642 ============= ========= ===========================
645 fault: yes can return with page locked
647 page_mkwrite: yes can return with page locked
650 ============= ========= ===========================
652 ->fault() is called when a previously not present pte is about to be faulted
653 in. The filesystem must find and return the page associated with the passed in
654 "pgoff" in the vm_fault structure. If it is possible that the page may be
655 truncated and/or invalidated, then the filesystem must lock invalidate_lock,
656 then ensure the page is not already truncated (invalidate_lock will block
657 subsequent truncate), and then return with VM_FAULT_LOCKED, and the page
658 locked. The VM will unlock the page.
660 ->map_pages() is called when VM asks to map easy accessible pages.
661 Filesystem should find and map pages associated with offsets from "start_pgoff"
662 till "end_pgoff". ->map_pages() is called with page table locked and must
663 not block. If it's not possible to reach a page without blocking,
664 filesystem should skip it. Filesystem should use do_set_pte() to setup
665 page table entry. Pointer to entry associated with the page is passed in
666 "pte" field in vm_fault structure. Pointers to entries for other offsets
667 should be calculated relative to "pte".
669 ->page_mkwrite() is called when a previously read-only pte is about to become
670 writeable. The filesystem again must ensure that there are no
671 truncate/invalidate races or races with operations such as ->remap_file_range
672 or ->copy_file_range, and then return with the page locked. Usually
673 mapping->invalidate_lock is suitable for proper serialization. If the page has
674 been truncated, the filesystem should not look up a new page like the ->fault()
675 handler, but simply return with VM_FAULT_NOPAGE, which will cause the VM to
678 ->pfn_mkwrite() is the same as page_mkwrite but when the pte is
679 VM_PFNMAP or VM_MIXEDMAP with a page-less entry. Expected return is
680 VM_FAULT_NOPAGE. Or one of the VM_FAULT_ERROR types. The default behavior
681 after this call is to make the pte read-write, unless pfn_mkwrite returns
684 ->access() is called when get_user_pages() fails in
685 access_process_vm(), typically used to debug a process through
686 /proc/pid/mem or ptrace. This function is needed only for
687 VM_IO | VM_PFNMAP VMAs.
689 --------------------------------------------------------------------------------
693 (if you break something or notice that it is broken and do not fix it yourself
694 - at least put it here)