Documentation/filesystems/Locking

   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?
   9
  10 --------------------------- dentry_operations --------------------------
  11 prototypes:
  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
  19 locking rules:
  20         none have BKL
  21                 dcache_lock     rename_lock     ->d_lock        may block
  22 d_revalidate:   no              no              no              yes
  23 d_hash          no              no              no              yes
  24 d_compare:      no              yes             no              no
  25 d_delete:       yes             no              yes             no
  26 d_release:      no              no              no              yes
  27 d_iput:         no              no              no              yes
  28
  29 --------------------------- inode_operations ---------------------------
  30 prototypes:
  31         int (*create) (struct inode *,struct dentry *,int, struct nameidata *);
  32         struct dentry * (*lookup) (struct inode *,struct dentry *, struct nameid
  33 ata *);
  34         int (*link) (struct dentry *,struct inode *,struct dentry *);
  35         int (*unlink) (struct inode *,struct dentry *);
  36         int (*symlink) (struct inode *,struct dentry *,const char *);
  37         int (*mkdir) (struct inode *,struct dentry *,int);
  38         int (*rmdir) (struct inode *,struct dentry *);
  39         int (*mknod) (struct inode *,struct dentry *,int,dev_t);
  40         int (*rename) (struct inode *, struct dentry *,
  41                         struct inode *, struct dentry *);
  42         int (*readlink) (struct dentry *, char __user *,int);
  43         int (*follow_link) (struct dentry *, struct nameidata *);
  44         void (*truncate) (struct inode *);
  45         int (*permission) (struct inode *, int, struct nameidata *);
  46         int (*setattr) (struct dentry *, struct iattr *);
  47         int (*getattr) (struct vfsmount *, struct dentry *, struct kstat *);
  48         int (*setxattr) (struct dentry *, const char *,const void *,size_t,int);
  49         ssize_t (*getxattr) (struct dentry *, const char *, void *, size_t);
  50         ssize_t (*listxattr) (struct dentry *, char *, size_t);
  51         int (*removexattr) (struct dentry *, const char *);
  52
  53 locking rules:
  54         all may block, none have BKL
  55                 i_sem(inode)
  56 lookup:         yes
  57 create:         yes
  58 link:           yes (both)
  59 mknod:          yes
  60 symlink:        yes
  61 mkdir:          yes
  62 unlink:         yes (both)
  63 rmdir:          yes (both)      (see below)
  64 rename:         yes (all)       (see below)
  65 readlink:       no
  66 follow_link:    no
  67 truncate:       yes             (see below)
  68 setattr:        yes
  69 permission:     no
  70 getattr:        no
  71 setxattr:       yes
  72 getxattr:       no
  73 listxattr:      no
  74 removexattr:    yes
  75         Additionally, ->rmdir(), ->unlink() and ->rename() have ->i_sem on
  76 victim.
  77         cross-directory ->rename() has (per-superblock) ->s_vfs_rename_sem.
  78         ->truncate() is never called directly - it's a callback, not a
  79 method. It's called by vmtruncate() - library function normally used by
  80 ->setattr(). Locking information above applies to that call (i.e. is
  81 inherited from ->setattr() - vmtruncate() is used when ATTR_SIZE had been
  82 passed).
  83
  84 See Documentation/filesystems/directory-locking for more detailed discussion
  85 of the locking scheme for directory operations.
  86
  87 --------------------------- super_operations ---------------------------
  88 prototypes:
  89         struct inode *(*alloc_inode)(struct super_block *sb);
  90         void (*destroy_inode)(struct inode *);
  91         void (*read_inode) (struct inode *);
  92         void (*dirty_inode) (struct inode *);
  93         int (*write_inode) (struct inode *, int);
  94         void (*put_inode) (struct inode *);
  95         void (*drop_inode) (struct inode *);
  96         void (*delete_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         void (*write_super_lockfs) (struct super_block *);
 101         void (*unlockfs) (struct super_block *);
 102         int (*statfs) (struct dentry *, struct kstatfs *);
 103         int (*remount_fs) (struct super_block *, int *, char *);
 104         void (*clear_inode) (struct inode *);
 105         void (*umount_begin) (struct super_block *);
 106         int (*show_options)(struct seq_file *, struct vfsmount *);
 107         ssize_t (*quota_read)(struct super_block *, int, char *, size_t, loff_t);
 108         ssize_t (*quota_write)(struct super_block *, int, const char *, size_t, loff_t);
 109
 110 locking rules:
 111         All may block.
 112                         BKL     s_lock  s_umount
 113 alloc_inode:            no      no      no
 114 destroy_inode:          no
 115 read_inode:             no                              (see below)
 116 dirty_inode:            no                              (must not sleep)
 117 write_inode:            no
 118 put_inode:              no
 119 drop_inode:             no                              !!!inode_lock!!!
 120 delete_inode:           no
 121 put_super:              yes     yes     no
 122 write_super:            no      yes     read
 123 sync_fs:                no      no      read
 124 write_super_lockfs:     ?
 125 unlockfs:               ?
 126 statfs:                 no      no      no
 127 remount_fs:             no      yes     maybe           (see below)
 128 clear_inode:            no
 129 umount_begin:           yes     no      no
 130 show_options:           no                              (vfsmount->sem)
 131 quota_read:             no      no      no              (see below)
 132 quota_write:            no      no      no              (see below)
 133
 134 ->read_inode() is not a method - it's a callback used in iget().
 135 ->remount_fs() will have the s_umount lock if it's already mounted.
 136 When called from get_sb_single, it does NOT have the s_umount lock.
 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.
 142
 143 --------------------------- file_system_type ---------------------------
 144 prototypes:
 145         int (*get_sb) (struct file_system_type *, int,
 146                        const char *, void *, struct vfsmount *);
 147         void (*kill_sb) (struct super_block *);
 148 locking rules:
 149                 may block       BKL
 150 get_sb          yes             yes
 151 kill_sb         yes             yes
 152
 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.
 157
 158 --------------------------- address_space_operations --------------------------
 159 prototypes:
 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 (*prepare_write)(struct file *, struct page *, unsigned, unsigned);
 168         int (*commit_write)(struct file *, struct page *, unsigned, unsigned);
 169         sector_t (*bmap)(struct address_space *, sector_t);
 170         int (*invalidatepage) (struct page *, unsigned long);
 171         int (*releasepage) (struct page *, int);
 172         int (*direct_IO)(int, struct kiocb *, const struct iovec *iov,
 173                         loff_t offset, unsigned long nr_segs);
 174
 175 locking rules:
 176         All except set_page_dirty may block
 177
 178                         BKL     PageLocked(page)
 179 writepage:              no      yes, unlocks (see below)
 180 readpage:               no      yes, unlocks
 181 sync_page:              no      maybe
 182 writepages:             no
 183 set_page_dirty          no      no
 184 readpages:              no
 185 prepare_write:          no      yes
 186 commit_write:           no      yes
 187 bmap:                   yes
 188 invalidatepage:         no      yes
 189 releasepage:            no      yes
 190 direct_IO:              no
 191
 192         ->prepare_write(), ->commit_write(), ->sync_page() and ->readpage()
 193 may be called from the request handler (/dev/loop).
 194
 195         ->readpage() unlocks the page, either synchronously or via I/O
 196 completion.
 197
 198         ->readpages() populates the pagecache with the passed pages and starts
 199 I/O against them.  They come unlocked upon I/O completion.
 200
 201         ->writepage() is used for two purposes: for "memory cleansing" and for
 202 "sync".  These are quite different operations and the behaviour may differ
 203 depending upon the mode.
 204
 205 If writepage is called for sync (wbc->sync_mode != WBC_SYNC_NONE) then
 206 it *must* start I/O against the page, even if that would involve
 207 blocking on in-progress I/O.
 208
 209 If writepage is called for memory cleansing (sync_mode ==
 210 WBC_SYNC_NONE) then its role is to get as much writeout underway as
 211 possible.  So writepage should try to avoid blocking against
 212 currently-in-progress I/O.
 213
 214 If the filesystem is not called for "sync" and it determines that it
 215 would need to block against in-progress I/O to be able to start new I/O
 216 against the page the filesystem should redirty the page with
 217 redirty_page_for_writepage(), then unlock the page and return zero.
 218 This may also be done to avoid internal deadlocks, but rarely.
 219
 220 If the filesytem is called for sync then it must wait on any
 221 in-progress I/O and then start new I/O.
 222
 223 The filesystem should unlock the page synchronously, before returning to the
 224 caller, unless ->writepage() returns special WRITEPAGE_ACTIVATE
 225 value. WRITEPAGE_ACTIVATE means that page cannot really be written out
 226 currently, and VM should stop calling ->writepage() on this page for some
 227 time. VM does this by moving page to the head of the active list, hence the
 228 name.
 229
 230 Unless the filesystem is going to redirty_page_for_writepage(), unlock the page
 231 and return zero, writepage *must* run set_page_writeback() against the page,
 232 followed by unlocking it.  Once set_page_writeback() has been run against the
 233 page, write I/O can be submitted and the write I/O completion handler must run
 234 end_page_writeback() once the I/O is complete.  If no I/O is submitted, the
 235 filesystem must run end_page_writeback() against the page before returning from
 236 writepage.
 237
 238 That is: after 2.5.12, pages which are under writeout are *not* locked.  Note,
 239 if the filesystem needs the page to be locked during writeout, that is ok, too,
 240 the page is allowed to be unlocked at any point in time between the calls to
 241 set_page_writeback() and end_page_writeback().
 242
 243 Note, failure to run either redirty_page_for_writepage() or the combination of
 244 set_page_writeback()/end_page_writeback() on a page submitted to writepage
 245 will leave the page itself marked clean but it will be tagged as dirty in the
 246 radix tree.  This incoherency can lead to all sorts of hard-to-debug problems
 247 in the filesystem like having dirty inodes at umount and losing written data.
 248
 249         ->sync_page() locking rules are not well-defined - usually it is called
 250 with lock on page, but that is not guaranteed. Considering the currently
 251 existing instances of this method ->sync_page() itself doesn't look
 252 well-defined...
 253
 254         ->writepages() is used for periodic writeback and for syscall-initiated
 255 sync operations.  The address_space should start I/O against at least
 256 *nr_to_write pages.  *nr_to_write must be decremented for each page which is
 257 written.  The address_space implementation may write more (or less) pages
 258 than *nr_to_write asks for, but it should try to be reasonably close.  If
 259 nr_to_write is NULL, all dirty pages must be written.
 260
 261 writepages should _only_ write pages which are present on
 262 mapping->io_pages.
 263
 264         ->set_page_dirty() is called from various places in the kernel
 265 when the target page is marked as needing writeback.  It may be called
 266 under spinlock (it cannot block) and is sometimes called with the page
 267 not locked.
 268
 269         ->bmap() is currently used by legacy ioctl() (FIBMAP) provided by some
 270 filesystems and by the swapper. The latter will eventually go away. All
 271 instances do not actually need the BKL. Please, keep it that way and don't
 272 breed new callers.
 273
 274         ->invalidatepage() is called when the filesystem must attempt to drop
 275 some or all of the buffers from the page when it is being truncated.  It
 276 returns zero on success.  If ->invalidatepage is zero, the kernel uses
 277 block_invalidatepage() instead.
 278
 279         ->releasepage() is called when the kernel is about to try to drop the
 280 buffers from the page in preparation for freeing it.  It returns zero to
 281 indicate that the buffers are (or may be) freeable.  If ->releasepage is zero,
 282 the kernel assumes that the fs has no private interest in the buffers.
 283
 284         Note: currently almost all instances of address_space methods are
 285 using BKL for internal serialization and that's one of the worst sources
 286 of contention. Normally they are calling library functions (in fs/buffer.c)
 287 and pass foo_get_block() as a callback (on local block-based filesystems,
 288 indeed). BKL is not needed for library stuff and is usually taken by
 289 foo_get_block(). It's an overkill, since block bitmaps can be protected by
 290 internal fs locking and real critical areas are much smaller than the areas
 291 filesystems protect now.
 292
 293 ----------------------- file_lock_operations ------------------------------
 294 prototypes:
 295         void (*fl_insert)(struct file_lock *);  /* lock insertion callback */
 296         void (*fl_remove)(struct file_lock *);  /* lock removal callback */
 297         void (*fl_copy_lock)(struct file_lock *, struct file_lock *);
 298         void (*fl_release_private)(struct file_lock *);
 299
 300
 301 locking rules:
 302                         BKL     may block
 303 fl_insert:              yes     no
 304 fl_remove:              yes     no
 305 fl_copy_lock:           yes     no
 306 fl_release_private:     yes     yes
 307
 308 ----------------------- lock_manager_operations ---------------------------
 309 prototypes:
 310         int (*fl_compare_owner)(struct file_lock *, struct file_lock *);
 311         void (*fl_notify)(struct file_lock *);  /* unblock callback */
 312         void (*fl_copy_lock)(struct file_lock *, struct file_lock *);
 313         void (*fl_release_private)(struct file_lock *);
 314         void (*fl_break)(struct file_lock *); /* break_lease callback */
 315
 316 locking rules:
 317                         BKL     may block
 318 fl_compare_owner:       yes     no
 319 fl_notify:              yes     no
 320 fl_copy_lock:           yes     no
 321 fl_release_private:     yes     yes
 322 fl_break:               yes     no
 323
 324         Currently only NFSD and NLM provide instances of this class. None of the
 325 them block. If you have out-of-tree instances - please, show up. Locking
 326 in that area will change.
 327 --------------------------- buffer_head -----------------------------------
 328 prototypes:
 329         void (*b_end_io)(struct buffer_head *bh, int uptodate);
 330
 331 locking rules:
 332         called from interrupts. In other words, extreme care is needed here.
 333 bh is locked, but that's all warranties we have here. Currently only RAID1,
 334 highmem, fs/buffer.c, and fs/ntfs/aops.c are providing these. Block devices
 335 call this method upon the IO completion.
 336
 337 --------------------------- block_device_operations -----------------------
 338 prototypes:
 339         int (*open) (struct inode *, struct file *);
 340         int (*release) (struct inode *, struct file *);
 341         int (*ioctl) (struct inode *, struct file *, unsigned, unsigned long);
 342         int (*media_changed) (struct gendisk *);
 343         int (*revalidate_disk) (struct gendisk *);
 344
 345 locking rules:
 346                         BKL     bd_sem
 347 open:                   yes     yes
 348 release:                yes     yes
 349 ioctl:                  yes     no
 350 media_changed:          no      no
 351 revalidate_disk:        no      no
 352
 353 The last two are called only from check_disk_change().
 354
 355 --------------------------- file_operations -------------------------------
 356 prototypes:
 357         loff_t (*llseek) (struct file *, loff_t, int);
 358         ssize_t (*read) (struct file *, char __user *, size_t, loff_t *);
 359         ssize_t (*write) (struct file *, const char __user *, size_t, loff_t *);
 360         ssize_t (*aio_read) (struct kiocb *, const struct iovec *, unsigned long, loff_t);
 361         ssize_t (*aio_write) (struct kiocb *, const struct iovec *, unsigned long, loff_t);
 362         int (*readdir) (struct file *, void *, filldir_t);
 363         unsigned int (*poll) (struct file *, struct poll_table_struct *);
 364         int (*ioctl) (struct inode *, struct file *, unsigned int,
 365                         unsigned long);
 366         long (*unlocked_ioctl) (struct file *, unsigned int, unsigned long);
 367         long (*compat_ioctl) (struct file *, unsigned int, unsigned long);
 368         int (*mmap) (struct file *, struct vm_area_struct *);
 369         int (*open) (struct inode *, struct file *);
 370         int (*flush) (struct file *);
 371         int (*release) (struct inode *, struct file *);
 372         int (*fsync) (struct file *, struct dentry *, int datasync);
 373         int (*aio_fsync) (struct kiocb *, int datasync);
 374         int (*fasync) (int, struct file *, int);
 375         int (*lock) (struct file *, int, struct file_lock *);
 376         ssize_t (*readv) (struct file *, const struct iovec *, unsigned long,
 377                         loff_t *);
 378         ssize_t (*writev) (struct file *, const struct iovec *, unsigned long,
 379                         loff_t *);
 380         ssize_t (*sendfile) (struct file *, loff_t *, size_t, read_actor_t,
 381                         void __user *);
 382         ssize_t (*sendpage) (struct file *, struct page *, int, size_t,
 383                         loff_t *, int);
 384         unsigned long (*get_unmapped_area)(struct file *, unsigned long,
 385                         unsigned long, unsigned long, unsigned long);
 386         int (*check_flags)(int);
 387         int (*dir_notify)(struct file *, unsigned long);
 388 };
 389
 390 locking rules:
 391         All except ->poll() may block.
 392                         BKL
 393 llseek:                 no      (see below)
 394 read:                   no
 395 aio_read:               no
 396 write:                  no
 397 aio_write:              no
 398 readdir:                no
 399 poll:                   no
 400 ioctl:                  yes     (see below)
 401 unlocked_ioctl:         no      (see below)
 402 compat_ioctl:           no
 403 mmap:                   no
 404 open:                   maybe   (see below)
 405 flush:                  no
 406 release:                no
 407 fsync:                  no      (see below)
 408 aio_fsync:              no
 409 fasync:                 yes     (see below)
 410 lock:                   yes
 411 readv:                  no
 412 writev:                 no
 413 sendfile:               no
 414 sendpage:               no
 415 get_unmapped_area:      no
 416 check_flags:            no
 417 dir_notify:             no
 418
 419 ->llseek() locking has moved from llseek to the individual llseek
 420 implementations.  If your fs is not using generic_file_llseek, you
 421 need to acquire and release the appropriate locks in your ->llseek().
 422 For many filesystems, it is probably safe to acquire the inode
 423 semaphore.  Note some filesystems (i.e. remote ones) provide no
 424 protection for i_size so you will need to use the BKL.
 425
 426 ->open() locking is in-transit: big lock partially moved into the methods.
 427 The only exception is ->open() in the instances of file_operations that never
 428 end up in ->i_fop/->proc_fops, i.e. ones that belong to character devices
 429 (chrdev_open() takes lock before replacing ->f_op and calling the secondary
 430 method. As soon as we fix the handling of module reference counters all
 431 instances of ->open() will be called without the BKL.
 432
 433 Note: ext2_release() was *the* source of contention on fs-intensive
 434 loads and dropping BKL on ->release() helps to get rid of that (we still
 435 grab BKL for cases when we close a file that had been opened r/w, but that
 436 can and should be done using the internal locking with smaller critical areas).
 437 Current worst offender is ext2_get_block()...
 438
 439 ->fasync() is a mess. This area needs a big cleanup and that will probably
 440 affect locking.
 441
 442 ->readdir() and ->ioctl() on directories must be changed. Ideally we would
 443 move ->readdir() to inode_operations and use a separate method for directory
 444 ->ioctl() or kill the latter completely. One of the problems is that for
 445 anything that resembles union-mount we won't have a struct file for all
 446 components. And there are other reasons why the current interface is a mess...
 447
 448 ->ioctl() on regular files is superceded by the ->unlocked_ioctl() that
 449 doesn't take the BKL.
 450
 451 ->read on directories probably must go away - we should just enforce -EISDIR
 452 in sys_read() and friends.
 453
 454 ->fsync() has i_sem on inode.
 455
 456 --------------------------- dquot_operations -------------------------------
 457 prototypes:
 458         int (*initialize) (struct inode *, int);
 459         int (*drop) (struct inode *);
 460         int (*alloc_space) (struct inode *, qsize_t, int);
 461         int (*alloc_inode) (const struct inode *, unsigned long);
 462         int (*free_space) (struct inode *, qsize_t);
 463         int (*free_inode) (const struct inode *, unsigned long);
 464         int (*transfer) (struct inode *, struct iattr *);
 465         int (*write_dquot) (struct dquot *);
 466         int (*acquire_dquot) (struct dquot *);
 467         int (*release_dquot) (struct dquot *);
 468         int (*mark_dirty) (struct dquot *);
 469         int (*write_info) (struct super_block *, int);
 470
 471 These operations are intended to be more or less wrapping functions that ensure
 472 a proper locking wrt the filesystem and call the generic quota operations.
 473
 474 What filesystem should expect from the generic quota functions:
 475
 476                 FS recursion    Held locks when called
 477 initialize:     yes             maybe dqonoff_sem
 478 drop:           yes             -
 479 alloc_space:    ->mark_dirty()  -
 480 alloc_inode:    ->mark_dirty()  -
 481 free_space:     ->mark_dirty()  -
 482 free_inode:     ->mark_dirty()  -
 483 transfer:       yes             -
 484 write_dquot:    yes             dqonoff_sem or dqptr_sem
 485 acquire_dquot:  yes             dqonoff_sem or dqptr_sem
 486 release_dquot:  yes             dqonoff_sem or dqptr_sem
 487 mark_dirty:     no              -
 488 write_info:     yes             dqonoff_sem
 489
 490 FS recursion means calling ->quota_read() and ->quota_write() from superblock
 491 operations.
 492
 493 ->alloc_space(), ->alloc_inode(), ->free_space(), ->free_inode() are called
 494 only directly by the filesystem and do not call any fs functions only
 495 the ->mark_dirty() operation.
 496
 497 More details about quota locking can be found in fs/dquot.c.
 498
 499 --------------------------- vm_operations_struct -----------------------------
 500 prototypes:
 501         void (*open)(struct vm_area_struct*);
 502         void (*close)(struct vm_area_struct*);
 503         struct page *(*nopage)(struct vm_area_struct*, unsigned long, int *);
 504
 505 locking rules:
 506                 BKL     mmap_sem
 507 open:           no      yes
 508 close:          no      yes
 509 nopage:         no      yes
 510
 511 ================================================================================
 512                         Dubious stuff
 513
 514 (if you break something or notice that it is broken and do not fix it yourself
 515 - at least put it here)
 516
 517 ipc/shm.c::shm_delete() - may need BKL.
 518 ->read() and ->write() in many drivers are (probably) missing BKL.
 519 drivers/sgi/char/graphics.c::sgi_graphics_nopage() - may need BKL.