1 The text below describes the locking rules for VFS-related methods.
2 It is (believed to be) up-to-date. *Please*, if you change anything in
3 prototypes or locking protocols - update this file. And update the relevant
4 instances in the tree, don't leave that to maintainers of filesystems/devices/
5 etc. At the very least, put the list of dubious cases in the end of this file.
6 Don't turn it into log - maintainers of out-of-the-tree code are supposed to
7 be able to use diff(1).
8 Thing currently missing here: socket operations. Alexey?
10 --------------------------- dentry_operations --------------------------
12 int (*d_revalidate)(struct dentry *, int);
13 int (*d_hash) (struct dentry *, struct qstr *);
14 int (*d_compare) (struct dentry *, struct qstr *, struct qstr *);
15 int (*d_delete)(struct dentry *);
16 void (*d_release)(struct dentry *);
17 void (*d_iput)(struct dentry *, struct inode *);
18 char *(*d_dname)((struct dentry *dentry, char *buffer, int buflen);
22 dcache_lock rename_lock ->d_lock may block
23 d_revalidate: no no no yes
25 d_compare: no yes no no
26 d_delete: yes no yes no
27 d_release: no no no yes
31 --------------------------- inode_operations ---------------------------
33 int (*create) (struct inode *,struct dentry *,int, struct nameidata *);
34 struct dentry * (*lookup) (struct inode *,struct dentry *, struct nameid
36 int (*link) (struct dentry *,struct inode *,struct dentry *);
37 int (*unlink) (struct inode *,struct dentry *);
38 int (*symlink) (struct inode *,struct dentry *,const char *);
39 int (*mkdir) (struct inode *,struct dentry *,int);
40 int (*rmdir) (struct inode *,struct dentry *);
41 int (*mknod) (struct inode *,struct dentry *,int,dev_t);
42 int (*rename) (struct inode *, struct dentry *,
43 struct inode *, struct dentry *);
44 int (*readlink) (struct dentry *, char __user *,int);
45 int (*follow_link) (struct dentry *, struct nameidata *);
46 void (*truncate) (struct inode *);
47 int (*permission) (struct inode *, int, struct nameidata *);
48 int (*setattr) (struct dentry *, struct iattr *);
49 int (*getattr) (struct vfsmount *, struct dentry *, struct kstat *);
50 int (*setxattr) (struct dentry *, const char *,const void *,size_t,int);
51 ssize_t (*getxattr) (struct dentry *, const char *, void *, size_t);
52 ssize_t (*listxattr) (struct dentry *, char *, size_t);
53 int (*removexattr) (struct dentry *, const char *);
56 all may block, none have BKL
65 rmdir: yes (both) (see below)
66 rename: yes (all) (see below)
69 truncate: yes (see below)
77 Additionally, ->rmdir(), ->unlink() and ->rename() have ->i_mutex on
79 cross-directory ->rename() has (per-superblock) ->s_vfs_rename_sem.
80 ->truncate() is never called directly - it's a callback, not a
81 method. It's called by vmtruncate() - library function normally used by
82 ->setattr(). Locking information above applies to that call (i.e. is
83 inherited from ->setattr() - vmtruncate() is used when ATTR_SIZE had been
86 See Documentation/filesystems/directory-locking for more detailed discussion
87 of the locking scheme for directory operations.
89 --------------------------- super_operations ---------------------------
91 struct inode *(*alloc_inode)(struct super_block *sb);
92 void (*destroy_inode)(struct inode *);
93 void (*dirty_inode) (struct inode *);
94 int (*write_inode) (struct inode *, int);
95 int (*drop_inode) (struct inode *);
96 void (*evict_inode) (struct inode *);
97 void (*put_super) (struct super_block *);
98 void (*write_super) (struct super_block *);
99 int (*sync_fs)(struct super_block *sb, int wait);
100 int (*freeze_fs) (struct super_block *);
101 int (*unfreeze_fs) (struct super_block *);
102 int (*statfs) (struct dentry *, struct kstatfs *);
103 int (*remount_fs) (struct super_block *, int *, char *);
104 void (*umount_begin) (struct super_block *);
105 int (*show_options)(struct seq_file *, struct vfsmount *);
106 ssize_t (*quota_read)(struct super_block *, int, char *, size_t, loff_t);
107 ssize_t (*quota_write)(struct super_block *, int, const char *, size_t, loff_t);
110 All may block [not true, see below]
115 dirty_inode: (must not sleep)
117 drop_inode: !!!inode_lock!!!
124 statfs: maybe(read) (see below)
127 show_options: no (namespace_sem)
128 quota_read: no (see below)
129 quota_write: no (see below)
131 ->statfs() has s_umount (shared) when called by ustat(2) (native or
132 compat), but that's an accident of bad API; s_umount is used to pin
133 the superblock down when we only have dev_t given us by userland to
134 identify the superblock. Everything else (statfs(), fstatfs(), etc.)
135 doesn't hold it when calling ->statfs() - superblock is pinned down
136 by resolving the pathname passed to syscall.
137 ->quota_read() and ->quota_write() functions are both guaranteed to
138 be the only ones operating on the quota file by the quota code (via
139 dqio_sem) (unless an admin really wants to screw up something and
140 writes to quota files with quotas on). For other details about locking
141 see also dquot_operations section.
143 --------------------------- file_system_type ---------------------------
145 int (*get_sb) (struct file_system_type *, int,
146 const char *, void *, struct vfsmount *);
147 void (*kill_sb) (struct super_block *);
153 ->get_sb() returns error or 0 with locked superblock attached to the vfsmount
154 (exclusive on ->s_umount).
155 ->kill_sb() takes a write-locked superblock, does all shutdown work on it,
156 unlocks and drops the reference.
158 --------------------------- address_space_operations --------------------------
160 int (*writepage)(struct page *page, struct writeback_control *wbc);
161 int (*readpage)(struct file *, struct page *);
162 int (*sync_page)(struct page *);
163 int (*writepages)(struct address_space *, struct writeback_control *);
164 int (*set_page_dirty)(struct page *page);
165 int (*readpages)(struct file *filp, struct address_space *mapping,
166 struct list_head *pages, unsigned nr_pages);
167 int (*write_begin)(struct file *, struct address_space *mapping,
168 loff_t pos, unsigned len, unsigned flags,
169 struct page **pagep, void **fsdata);
170 int (*write_end)(struct file *, struct address_space *mapping,
171 loff_t pos, unsigned len, unsigned copied,
172 struct page *page, void *fsdata);
173 sector_t (*bmap)(struct address_space *, sector_t);
174 int (*invalidatepage) (struct page *, unsigned long);
175 int (*releasepage) (struct page *, int);
176 int (*direct_IO)(int, struct kiocb *, const struct iovec *iov,
177 loff_t offset, unsigned long nr_segs);
178 int (*launder_page) (struct page *);
181 All except set_page_dirty may block
183 BKL PageLocked(page) i_mutex
184 writepage: no yes, unlocks (see below)
185 readpage: no yes, unlocks
190 write_begin: no locks the page yes
191 write_end: no yes, unlocks yes
192 perform_write: no n/a yes
194 invalidatepage: no yes
199 ->write_begin(), ->write_end(), ->sync_page() and ->readpage()
200 may be called from the request handler (/dev/loop).
202 ->readpage() unlocks the page, either synchronously or via I/O
205 ->readpages() populates the pagecache with the passed pages and starts
206 I/O against them. They come unlocked upon I/O completion.
208 ->writepage() is used for two purposes: for "memory cleansing" and for
209 "sync". These are quite different operations and the behaviour may differ
210 depending upon the mode.
212 If writepage is called for sync (wbc->sync_mode != WBC_SYNC_NONE) then
213 it *must* start I/O against the page, even if that would involve
214 blocking on in-progress I/O.
216 If writepage is called for memory cleansing (sync_mode ==
217 WBC_SYNC_NONE) then its role is to get as much writeout underway as
218 possible. So writepage should try to avoid blocking against
219 currently-in-progress I/O.
221 If the filesystem is not called for "sync" and it determines that it
222 would need to block against in-progress I/O to be able to start new I/O
223 against the page the filesystem should redirty the page with
224 redirty_page_for_writepage(), then unlock the page and return zero.
225 This may also be done to avoid internal deadlocks, but rarely.
227 If the filesystem is called for sync then it must wait on any
228 in-progress I/O and then start new I/O.
230 The filesystem should unlock the page synchronously, before returning to the
231 caller, unless ->writepage() returns special WRITEPAGE_ACTIVATE
232 value. WRITEPAGE_ACTIVATE means that page cannot really be written out
233 currently, and VM should stop calling ->writepage() on this page for some
234 time. VM does this by moving page to the head of the active list, hence the
237 Unless the filesystem is going to redirty_page_for_writepage(), unlock the page
238 and return zero, writepage *must* run set_page_writeback() against the page,
239 followed by unlocking it. Once set_page_writeback() has been run against the
240 page, write I/O can be submitted and the write I/O completion handler must run
241 end_page_writeback() once the I/O is complete. If no I/O is submitted, the
242 filesystem must run end_page_writeback() against the page before returning from
245 That is: after 2.5.12, pages which are under writeout are *not* locked. Note,
246 if the filesystem needs the page to be locked during writeout, that is ok, too,
247 the page is allowed to be unlocked at any point in time between the calls to
248 set_page_writeback() and end_page_writeback().
250 Note, failure to run either redirty_page_for_writepage() or the combination of
251 set_page_writeback()/end_page_writeback() on a page submitted to writepage
252 will leave the page itself marked clean but it will be tagged as dirty in the
253 radix tree. This incoherency can lead to all sorts of hard-to-debug problems
254 in the filesystem like having dirty inodes at umount and losing written data.
256 ->sync_page() locking rules are not well-defined - usually it is called
257 with lock on page, but that is not guaranteed. Considering the currently
258 existing instances of this method ->sync_page() itself doesn't look
261 ->writepages() is used for periodic writeback and for syscall-initiated
262 sync operations. The address_space should start I/O against at least
263 *nr_to_write pages. *nr_to_write must be decremented for each page which is
264 written. The address_space implementation may write more (or less) pages
265 than *nr_to_write asks for, but it should try to be reasonably close. If
266 nr_to_write is NULL, all dirty pages must be written.
268 writepages should _only_ write pages which are present on
271 ->set_page_dirty() is called from various places in the kernel
272 when the target page is marked as needing writeback. It may be called
273 under spinlock (it cannot block) and is sometimes called with the page
276 ->bmap() is currently used by legacy ioctl() (FIBMAP) provided by some
277 filesystems and by the swapper. The latter will eventually go away. All
278 instances do not actually need the BKL. Please, keep it that way and don't
281 ->invalidatepage() is called when the filesystem must attempt to drop
282 some or all of the buffers from the page when it is being truncated. It
283 returns zero on success. If ->invalidatepage is zero, the kernel uses
284 block_invalidatepage() instead.
286 ->releasepage() is called when the kernel is about to try to drop the
287 buffers from the page in preparation for freeing it. It returns zero to
288 indicate that the buffers are (or may be) freeable. If ->releasepage is zero,
289 the kernel assumes that the fs has no private interest in the buffers.
291 ->launder_page() may be called prior to releasing a page if
292 it is still found to be dirty. It returns zero if the page was successfully
293 cleaned, or an error value if not. Note that in order to prevent the page
294 getting mapped back in and redirtied, it needs to be kept locked
295 across the entire operation.
297 Note: currently almost all instances of address_space methods are
298 using BKL for internal serialization and that's one of the worst sources
299 of contention. Normally they are calling library functions (in fs/buffer.c)
300 and pass foo_get_block() as a callback (on local block-based filesystems,
301 indeed). BKL is not needed for library stuff and is usually taken by
302 foo_get_block(). It's an overkill, since block bitmaps can be protected by
303 internal fs locking and real critical areas are much smaller than the areas
304 filesystems protect now.
306 ----------------------- file_lock_operations ------------------------------
308 void (*fl_insert)(struct file_lock *); /* lock insertion callback */
309 void (*fl_remove)(struct file_lock *); /* lock removal callback */
310 void (*fl_copy_lock)(struct file_lock *, struct file_lock *);
311 void (*fl_release_private)(struct file_lock *);
319 fl_release_private: yes yes
321 ----------------------- lock_manager_operations ---------------------------
323 int (*fl_compare_owner)(struct file_lock *, struct file_lock *);
324 void (*fl_notify)(struct file_lock *); /* unblock callback */
325 void (*fl_copy_lock)(struct file_lock *, struct file_lock *);
326 void (*fl_release_private)(struct file_lock *);
327 void (*fl_break)(struct file_lock *); /* break_lease callback */
331 fl_compare_owner: yes no
334 fl_release_private: yes yes
337 Currently only NFSD and NLM provide instances of this class. None of the
338 them block. If you have out-of-tree instances - please, show up. Locking
339 in that area will change.
340 --------------------------- buffer_head -----------------------------------
342 void (*b_end_io)(struct buffer_head *bh, int uptodate);
345 called from interrupts. In other words, extreme care is needed here.
346 bh is locked, but that's all warranties we have here. Currently only RAID1,
347 highmem, fs/buffer.c, and fs/ntfs/aops.c are providing these. Block devices
348 call this method upon the IO completion.
350 --------------------------- block_device_operations -----------------------
352 int (*open) (struct inode *, struct file *);
353 int (*release) (struct inode *, struct file *);
354 int (*ioctl) (struct inode *, struct file *, unsigned, unsigned long);
355 int (*media_changed) (struct gendisk *);
356 int (*revalidate_disk) (struct gendisk *);
364 revalidate_disk: no no
366 The last two are called only from check_disk_change().
368 --------------------------- file_operations -------------------------------
370 loff_t (*llseek) (struct file *, loff_t, int);
371 ssize_t (*read) (struct file *, char __user *, size_t, loff_t *);
372 ssize_t (*write) (struct file *, const char __user *, size_t, loff_t *);
373 ssize_t (*aio_read) (struct kiocb *, const struct iovec *, unsigned long, loff_t);
374 ssize_t (*aio_write) (struct kiocb *, const struct iovec *, unsigned long, loff_t);
375 int (*readdir) (struct file *, void *, filldir_t);
376 unsigned int (*poll) (struct file *, struct poll_table_struct *);
377 long (*unlocked_ioctl) (struct file *, unsigned int, unsigned long);
378 long (*compat_ioctl) (struct file *, unsigned int, unsigned long);
379 int (*mmap) (struct file *, struct vm_area_struct *);
380 int (*open) (struct inode *, struct file *);
381 int (*flush) (struct file *);
382 int (*release) (struct inode *, struct file *);
383 int (*fsync) (struct file *, int datasync);
384 int (*aio_fsync) (struct kiocb *, int datasync);
385 int (*fasync) (int, struct file *, int);
386 int (*lock) (struct file *, int, struct file_lock *);
387 ssize_t (*readv) (struct file *, const struct iovec *, unsigned long,
389 ssize_t (*writev) (struct file *, const struct iovec *, unsigned long,
391 ssize_t (*sendfile) (struct file *, loff_t *, size_t, read_actor_t,
393 ssize_t (*sendpage) (struct file *, struct page *, int, size_t,
395 unsigned long (*get_unmapped_area)(struct file *, unsigned long,
396 unsigned long, unsigned long, unsigned long);
397 int (*check_flags)(int);
403 llseek: no (see below)
416 fsync: no (see below)
424 get_unmapped_area: no
427 ->llseek() locking has moved from llseek to the individual llseek
428 implementations. If your fs is not using generic_file_llseek, you
429 need to acquire and release the appropriate locks in your ->llseek().
430 For many filesystems, it is probably safe to acquire the inode
431 mutex or just to use i_size_read() instead.
432 Note: this does not protect the file->f_pos against concurrent modifications
433 since this is something the userspace has to take care about.
435 Note: ext2_release() was *the* source of contention on fs-intensive
436 loads and dropping BKL on ->release() helps to get rid of that (we still
437 grab BKL for cases when we close a file that had been opened r/w, but that
438 can and should be done using the internal locking with smaller critical areas).
439 Current worst offender is ext2_get_block()...
441 ->fasync() is called without BKL protection, and is responsible for
442 maintaining the FASYNC bit in filp->f_flags. Most instances call
443 fasync_helper(), which does that maintenance, so it's not normally
444 something one needs to worry about. Return values > 0 will be mapped to
445 zero in the VFS layer.
447 ->readdir() and ->ioctl() on directories must be changed. Ideally we would
448 move ->readdir() to inode_operations and use a separate method for directory
449 ->ioctl() or kill the latter completely. One of the problems is that for
450 anything that resembles union-mount we won't have a struct file for all
451 components. And there are other reasons why the current interface is a mess...
453 ->read on directories probably must go away - we should just enforce -EISDIR
454 in sys_read() and friends.
456 ->fsync() has i_mutex on inode.
458 --------------------------- dquot_operations -------------------------------
460 int (*write_dquot) (struct dquot *);
461 int (*acquire_dquot) (struct dquot *);
462 int (*release_dquot) (struct dquot *);
463 int (*mark_dirty) (struct dquot *);
464 int (*write_info) (struct super_block *, int);
466 These operations are intended to be more or less wrapping functions that ensure
467 a proper locking wrt the filesystem and call the generic quota operations.
469 What filesystem should expect from the generic quota functions:
471 FS recursion Held locks when called
472 write_dquot: yes dqonoff_sem or dqptr_sem
473 acquire_dquot: yes dqonoff_sem or dqptr_sem
474 release_dquot: yes dqonoff_sem or dqptr_sem
476 write_info: yes dqonoff_sem
478 FS recursion means calling ->quota_read() and ->quota_write() from superblock
481 More details about quota locking can be found in fs/dquot.c.
483 --------------------------- vm_operations_struct -----------------------------
485 void (*open)(struct vm_area_struct*);
486 void (*close)(struct vm_area_struct*);
487 int (*fault)(struct vm_area_struct*, struct vm_fault *);
488 int (*page_mkwrite)(struct vm_area_struct *, struct vm_fault *);
489 int (*access)(struct vm_area_struct *, unsigned long, void*, int, int);
492 BKL mmap_sem PageLocked(page)
495 fault: no yes can return with page locked
496 page_mkwrite: no yes can return with page locked
499 ->fault() is called when a previously not present pte is about
500 to be faulted in. The filesystem must find and return the page associated
501 with the passed in "pgoff" in the vm_fault structure. If it is possible that
502 the page may be truncated and/or invalidated, then the filesystem must lock
503 the page, then ensure it is not already truncated (the page lock will block
504 subsequent truncate), and then return with VM_FAULT_LOCKED, and the page
505 locked. The VM will unlock the page.
507 ->page_mkwrite() is called when a previously read-only pte is
508 about to become writeable. The filesystem again must ensure that there are
509 no truncate/invalidate races, and then return with the page locked. If
510 the page has been truncated, the filesystem should not look up a new page
511 like the ->fault() handler, but simply return with VM_FAULT_NOPAGE, which
512 will cause the VM to retry the fault.
514 ->access() is called when get_user_pages() fails in
515 acces_process_vm(), typically used to debug a process through
516 /proc/pid/mem or ptrace. This function is needed only for
517 VM_IO | VM_PFNMAP VMAs.
519 ================================================================================
522 (if you break something or notice that it is broken and do not fix it yourself
523 - at least put it here)
525 ipc/shm.c::shm_delete() - may need BKL.
526 ->read() and ->write() in many drivers are (probably) missing BKL.