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 super_block *, 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         struct super_block *(*get_sb) (struct file_system_type *, int,
 146                         const char *, void *);
 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 a locked superblock (exclusive on ->s_umount).
 154 ->kill_sb() takes a write-locked superblock, does all shutdown work on it,
 155 unlocks and drops the reference.
 156
 157 --------------------------- address_space_operations --------------------------
 158 prototypes:
 159         int (*writepage)(struct page *page, struct writeback_control *wbc);
 160         int (*readpage)(struct file *, struct page *);
 161         int (*sync_page)(struct page *);
 162         int (*writepages)(struct address_space *, struct writeback_control *);
 163         int (*set_page_dirty)(struct page *page);
 164         int (*readpages)(struct file *filp, struct address_space *mapping,
 165                         struct list_head *pages, unsigned nr_pages);
 166         int (*prepare_write)(struct file *, struct page *, unsigned, unsigned);
 167         int (*commit_write)(struct file *, struct page *, unsigned, unsigned);
 168         sector_t (*bmap)(struct address_space *, sector_t);
 169         int (*invalidatepage) (struct page *, unsigned long);
 170         int (*releasepage) (struct page *, int);
 171         int (*direct_IO)(int, struct kiocb *, const struct iovec *iov,
 172                         loff_t offset, unsigned long nr_segs);
 173
 174 locking rules:
 175         All except set_page_dirty may block
 176
 177                         BKL     PageLocked(page)
 178 writepage:              no      yes, unlocks (see below)
 179 readpage:               no      yes, unlocks
 180 sync_page:              no      maybe
 181 writepages:             no
 182 set_page_dirty          no      no
 183 readpages:              no
 184 prepare_write:          no      yes
 185 commit_write:           no      yes
 186 bmap:                   yes
 187 invalidatepage:         no      yes
 188 releasepage:            no      yes
 189 direct_IO:              no
 190
 191         ->prepare_write(), ->commit_write(), ->sync_page() and ->readpage()
 192 may be called from the request handler (/dev/loop).
 193
 194         ->readpage() unlocks the page, either synchronously or via I/O
 195 completion.
 196
 197         ->readpages() populates the pagecache with the passed pages and starts
 198 I/O against them.  They come unlocked upon I/O completion.
 199
 200         ->writepage() is used for two purposes: for "memory cleansing" and for
 201 "sync".  These are quite different operations and the behaviour may differ
 202 depending upon the mode.
 203
 204 If writepage is called for sync (wbc->sync_mode != WBC_SYNC_NONE) then
 205 it *must* start I/O against the page, even if that would involve
 206 blocking on in-progress I/O.
 207
 208 If writepage is called for memory cleansing (sync_mode ==
 209 WBC_SYNC_NONE) then its role is to get as much writeout underway as
 210 possible.  So writepage should try to avoid blocking against
 211 currently-in-progress I/O.
 212
 213 If the filesystem is not called for "sync" and it determines that it
 214 would need to block against in-progress I/O to be able to start new I/O
 215 against the page the filesystem should redirty the page with
 216 redirty_page_for_writepage(), then unlock the page and return zero.
 217 This may also be done to avoid internal deadlocks, but rarely.
 218
 219 If the filesytem is called for sync then it must wait on any
 220 in-progress I/O and then start new I/O.
 221
 222 The filesystem should unlock the page synchronously, before returning
 223 to the caller.
 224
 225 Unless the filesystem is going to redirty_page_for_writepage(), unlock the page
 226 and return zero, writepage *must* run set_page_writeback() against the page,
 227 followed by unlocking it.  Once set_page_writeback() has been run against the
 228 page, write I/O can be submitted and the write I/O completion handler must run
 229 end_page_writeback() once the I/O is complete.  If no I/O is submitted, the
 230 filesystem must run end_page_writeback() against the page before returning from
 231 writepage.
 232
 233 That is: after 2.5.12, pages which are under writeout are *not* locked.  Note,
 234 if the filesystem needs the page to be locked during writeout, that is ok, too,
 235 the page is allowed to be unlocked at any point in time between the calls to
 236 set_page_writeback() and end_page_writeback().
 237
 238 Note, failure to run either redirty_page_for_writepage() or the combination of
 239 set_page_writeback()/end_page_writeback() on a page submitted to writepage
 240 will leave the page itself marked clean but it will be tagged as dirty in the
 241 radix tree.  This incoherency can lead to all sorts of hard-to-debug problems
 242 in the filesystem like having dirty inodes at umount and losing written data.
 243
 244         ->sync_page() locking rules are not well-defined - usually it is called
 245 with lock on page, but that is not guaranteed. Considering the currently
 246 existing instances of this method ->sync_page() itself doesn't look
 247 well-defined...
 248
 249         ->writepages() is used for periodic writeback and for syscall-initiated
 250 sync operations.  The address_space should start I/O against at least
 251 *nr_to_write pages.  *nr_to_write must be decremented for each page which is
 252 written.  The address_space implementation may write more (or less) pages
 253 than *nr_to_write asks for, but it should try to be reasonably close.  If
 254 nr_to_write is NULL, all dirty pages must be written.
 255
 256 writepages should _only_ write pages which are present on
 257 mapping->io_pages.
 258
 259         ->set_page_dirty() is called from various places in the kernel
 260 when the target page is marked as needing writeback.  It may be called
 261 under spinlock (it cannot block) and is sometimes called with the page
 262 not locked.
 263
 264         ->bmap() is currently used by legacy ioctl() (FIBMAP) provided by some
 265 filesystems and by the swapper. The latter will eventually go away. All
 266 instances do not actually need the BKL. Please, keep it that way and don't
 267 breed new callers.
 268
 269         ->invalidatepage() is called when the filesystem must attempt to drop
 270 some or all of the buffers from the page when it is being truncated.  It
 271 returns zero on success.  If ->invalidatepage is zero, the kernel uses
 272 block_invalidatepage() instead.
 273
 274         ->releasepage() is called when the kernel is about to try to drop the
 275 buffers from the page in preparation for freeing it.  It returns zero to
 276 indicate that the buffers are (or may be) freeable.  If ->releasepage is zero,
 277 the kernel assumes that the fs has no private interest in the buffers.
 278
 279         Note: currently almost all instances of address_space methods are
 280 using BKL for internal serialization and that's one of the worst sources
 281 of contention. Normally they are calling library functions (in fs/buffer.c)
 282 and pass foo_get_block() as a callback (on local block-based filesystems,
 283 indeed). BKL is not needed for library stuff and is usually taken by
 284 foo_get_block(). It's an overkill, since block bitmaps can be protected by
 285 internal fs locking and real critical areas are much smaller than the areas
 286 filesystems protect now.
 287
 288 ----------------------- file_lock_operations ------------------------------
 289 prototypes:
 290         void (*fl_insert)(struct file_lock *);  /* lock insertion callback */
 291         void (*fl_remove)(struct file_lock *);  /* lock removal callback */
 292         void (*fl_copy_lock)(struct file_lock *, struct file_lock *);
 293         void (*fl_release_private)(struct file_lock *);
 294
 295
 296 locking rules:
 297                         BKL     may block
 298 fl_insert:              yes     no
 299 fl_remove:              yes     no
 300 fl_copy_lock:           yes     no
 301 fl_release_private:     yes     yes
 302
 303 ----------------------- lock_manager_operations ---------------------------
 304 prototypes:
 305         int (*fl_compare_owner)(struct file_lock *, struct file_lock *);
 306         void (*fl_notify)(struct file_lock *);  /* unblock callback */
 307         void (*fl_copy_lock)(struct file_lock *, struct file_lock *);
 308         void (*fl_release_private)(struct file_lock *);
 309         void (*fl_break)(struct file_lock *); /* break_lease callback */
 310
 311 locking rules:
 312                         BKL     may block
 313 fl_compare_owner:       yes     no
 314 fl_notify:              yes     no
 315 fl_copy_lock:           yes     no
 316 fl_release_private:     yes     yes
 317 fl_break:               yes     no
 318
 319         Currently only NFSD and NLM provide instances of this class. None of the
 320 them block. If you have out-of-tree instances - please, show up. Locking
 321 in that area will change.
 322 --------------------------- buffer_head -----------------------------------
 323 prototypes:
 324         void (*b_end_io)(struct buffer_head *bh, int uptodate);
 325
 326 locking rules:
 327         called from interrupts. In other words, extreme care is needed here.
 328 bh is locked, but that's all warranties we have here. Currently only RAID1,
 329 highmem, fs/buffer.c, and fs/ntfs/aops.c are providing these. Block devices
 330 call this method upon the IO completion.
 331
 332 --------------------------- block_device_operations -----------------------
 333 prototypes:
 334         int (*open) (struct inode *, struct file *);
 335         int (*release) (struct inode *, struct file *);
 336         int (*ioctl) (struct inode *, struct file *, unsigned, unsigned long);
 337         int (*media_changed) (struct gendisk *);
 338         int (*revalidate_disk) (struct gendisk *);
 339
 340 locking rules:
 341                         BKL     bd_sem
 342 open:                   yes     yes
 343 release:                yes     yes
 344 ioctl:                  yes     no
 345 media_changed:          no      no
 346 revalidate_disk:        no      no
 347
 348 The last two are called only from check_disk_change().
 349
 350 --------------------------- file_operations -------------------------------
 351 prototypes:
 352         loff_t (*llseek) (struct file *, loff_t, int);
 353         ssize_t (*read) (struct file *, char __user *, size_t, loff_t *);
 354         ssize_t (*aio_read) (struct kiocb *, char __user *, size_t, loff_t);
 355         ssize_t (*write) (struct file *, const char __user *, size_t, loff_t *);
 356         ssize_t (*aio_write) (struct kiocb *, const char __user *, size_t,
 357                         loff_t);
 358         int (*readdir) (struct file *, void *, filldir_t);
 359         unsigned int (*poll) (struct file *, struct poll_table_struct *);
 360         int (*ioctl) (struct inode *, struct file *, unsigned int,
 361                         unsigned long);
 362         long (*unlocked_ioctl) (struct file *, unsigned int, unsigned long);
 363         long (*compat_ioctl) (struct file *, unsigned int, unsigned long);
 364         int (*mmap) (struct file *, struct vm_area_struct *);
 365         int (*open) (struct inode *, struct file *);
 366         int (*flush) (struct file *);
 367         int (*release) (struct inode *, struct file *);
 368         int (*fsync) (struct file *, struct dentry *, int datasync);
 369         int (*aio_fsync) (struct kiocb *, int datasync);
 370         int (*fasync) (int, struct file *, int);
 371         int (*lock) (struct file *, int, struct file_lock *);
 372         ssize_t (*readv) (struct file *, const struct iovec *, unsigned long,
 373                         loff_t *);
 374         ssize_t (*writev) (struct file *, const struct iovec *, unsigned long,
 375                         loff_t *);
 376         ssize_t (*sendfile) (struct file *, loff_t *, size_t, read_actor_t,
 377                         void __user *);
 378         ssize_t (*sendpage) (struct file *, struct page *, int, size_t,
 379                         loff_t *, int);
 380         unsigned long (*get_unmapped_area)(struct file *, unsigned long,
 381                         unsigned long, unsigned long, unsigned long);
 382         int (*check_flags)(int);
 383         int (*dir_notify)(struct file *, unsigned long);
 384 };
 385
 386 locking rules:
 387         All except ->poll() may block.
 388                         BKL
 389 llseek:                 no      (see below)
 390 read:                   no
 391 aio_read:               no
 392 write:                  no
 393 aio_write:              no
 394 readdir:                no
 395 poll:                   no
 396 ioctl:                  yes     (see below)
 397 unlocked_ioctl:         no      (see below)
 398 compat_ioctl:           no
 399 mmap:                   no
 400 open:                   maybe   (see below)
 401 flush:                  no
 402 release:                no
 403 fsync:                  no      (see below)
 404 aio_fsync:              no
 405 fasync:                 yes     (see below)
 406 lock:                   yes
 407 readv:                  no
 408 writev:                 no
 409 sendfile:               no
 410 sendpage:               no
 411 get_unmapped_area:      no
 412 check_flags:            no
 413 dir_notify:             no
 414
 415 ->llseek() locking has moved from llseek to the individual llseek
 416 implementations.  If your fs is not using generic_file_llseek, you
 417 need to acquire and release the appropriate locks in your ->llseek().
 418 For many filesystems, it is probably safe to acquire the inode
 419 semaphore.  Note some filesystems (i.e. remote ones) provide no
 420 protection for i_size so you will need to use the BKL.
 421
 422 ->open() locking is in-transit: big lock partially moved into the methods.
 423 The only exception is ->open() in the instances of file_operations that never
 424 end up in ->i_fop/->proc_fops, i.e. ones that belong to character devices
 425 (chrdev_open() takes lock before replacing ->f_op and calling the secondary
 426 method. As soon as we fix the handling of module reference counters all
 427 instances of ->open() will be called without the BKL.
 428
 429 Note: ext2_release() was *the* source of contention on fs-intensive
 430 loads and dropping BKL on ->release() helps to get rid of that (we still
 431 grab BKL for cases when we close a file that had been opened r/w, but that
 432 can and should be done using the internal locking with smaller critical areas).
 433 Current worst offender is ext2_get_block()...
 434
 435 ->fasync() is a mess. This area needs a big cleanup and that will probably
 436 affect locking.
 437
 438 ->readdir() and ->ioctl() on directories must be changed. Ideally we would
 439 move ->readdir() to inode_operations and use a separate method for directory
 440 ->ioctl() or kill the latter completely. One of the problems is that for
 441 anything that resembles union-mount we won't have a struct file for all
 442 components. And there are other reasons why the current interface is a mess...
 443
 444 ->ioctl() on regular files is superceded by the ->unlocked_ioctl() that
 445 doesn't take the BKL.
 446
 447 ->read on directories probably must go away - we should just enforce -EISDIR
 448 in sys_read() and friends.
 449
 450 ->fsync() has i_sem on inode.
 451
 452 --------------------------- dquot_operations -------------------------------
 453 prototypes:
 454         int (*initialize) (struct inode *, int);
 455         int (*drop) (struct inode *);
 456         int (*alloc_space) (struct inode *, qsize_t, int);
 457         int (*alloc_inode) (const struct inode *, unsigned long);
 458         int (*free_space) (struct inode *, qsize_t);
 459         int (*free_inode) (const struct inode *, unsigned long);
 460         int (*transfer) (struct inode *, struct iattr *);
 461         int (*write_dquot) (struct dquot *);
 462         int (*acquire_dquot) (struct dquot *);
 463         int (*release_dquot) (struct dquot *);
 464         int (*mark_dirty) (struct dquot *);
 465         int (*write_info) (struct super_block *, int);
 466
 467 These operations are intended to be more or less wrapping functions that ensure
 468 a proper locking wrt the filesystem and call the generic quota operations.
 469
 470 What filesystem should expect from the generic quota functions:
 471
 472                 FS recursion    Held locks when called
 473 initialize:     yes             maybe dqonoff_sem
 474 drop:           yes             -
 475 alloc_space:    ->mark_dirty()  -
 476 alloc_inode:    ->mark_dirty()  -
 477 free_space:     ->mark_dirty()  -
 478 free_inode:     ->mark_dirty()  -
 479 transfer:       yes             -
 480 write_dquot:    yes             dqonoff_sem or dqptr_sem
 481 acquire_dquot:  yes             dqonoff_sem or dqptr_sem
 482 release_dquot:  yes             dqonoff_sem or dqptr_sem
 483 mark_dirty:     no              -
 484 write_info:     yes             dqonoff_sem
 485
 486 FS recursion means calling ->quota_read() and ->quota_write() from superblock
 487 operations.
 488
 489 ->alloc_space(), ->alloc_inode(), ->free_space(), ->free_inode() are called
 490 only directly by the filesystem and do not call any fs functions only
 491 the ->mark_dirty() operation.
 492
 493 More details about quota locking can be found in fs/dquot.c.
 494
 495 --------------------------- vm_operations_struct -----------------------------
 496 prototypes:
 497         void (*open)(struct vm_area_struct*);
 498         void (*close)(struct vm_area_struct*);
 499         struct page *(*nopage)(struct vm_area_struct*, unsigned long, int *);
 500
 501 locking rules:
 502                 BKL     mmap_sem
 503 open:           no      yes
 504 close:          no      yes
 505 nopage:         no      yes
 506
 507 ================================================================================
 508                         Dubious stuff
 509
 510 (if you break something or notice that it is broken and do not fix it yourself
 511 - at least put it here)
 512
 513 ipc/shm.c::shm_delete() - may need BKL.
 514 ->read() and ->write() in many drivers are (probably) missing BKL.
 515 drivers/sgi/char/graphics.c::sgi_graphics_nopage() - may need BKL.