1 The text below describes the locking rules for VFS-related methods. 2It is (believed to be) up-to-date. *Please*, if you change anything in 3prototypes or locking protocols - update this file. And update the relevant 4instances in the tree, don't leave that to maintainers of filesystems/devices/ 5etc. At the very least, put the list of dubious cases in the end of this file. 6Don't turn it into log - maintainers of out-of-the-tree code are supposed to 7be able to use diff(1). 8 Thing currently missing here: socket operations. Alexey? 9 10--------------------------- dentry_operations -------------------------- 11prototypes: 12 int (*d_revalidate)(struct dentry *, unsigned int); 13 int (*d_weak_revalidate)(struct dentry *, unsigned int); 14 int (*d_hash)(const struct dentry *, struct qstr *); 15 int (*d_compare)(const struct dentry *, const struct dentry *, 16 unsigned int, const char *, const struct qstr *); 17 int (*d_delete)(struct dentry *); 18 void (*d_release)(struct dentry *); 19 void (*d_iput)(struct dentry *, struct inode *); 20 char *(*d_dname)((struct dentry *dentry, char *buffer, int buflen); 21 struct vfsmount *(*d_automount)(struct path *path); 22 int (*d_manage)(struct dentry *, bool); 23 24locking rules: 25 rename_lock ->d_lock may block rcu-walk 26d_revalidate: no no yes (ref-walk) maybe 27d_weak_revalidate:no no yes no 28d_hash no no no maybe 29d_compare: yes no no maybe 30d_delete: no yes no no 31d_release: no no yes no 32d_prune: no yes no no 33d_iput: no no yes no 34d_dname: no no no no 35d_automount: no no yes no 36d_manage: no no yes (ref-walk) maybe 37 38--------------------------- inode_operations --------------------------- 39prototypes: 40 int (*create) (struct inode *,struct dentry *,umode_t, bool); 41 struct dentry * (*lookup) (struct inode *,struct dentry *, unsigned int); 42 int (*link) (struct dentry *,struct inode *,struct dentry *); 43 int (*unlink) (struct inode *,struct dentry *); 44 int (*symlink) (struct inode *,struct dentry *,const char *); 45 int (*mkdir) (struct inode *,struct dentry *,umode_t); 46 int (*rmdir) (struct inode *,struct dentry *); 47 int (*mknod) (struct inode *,struct dentry *,umode_t,dev_t); 48 int (*rename) (struct inode *, struct dentry *, 49 struct inode *, struct dentry *); 50 int (*rename2) (struct inode *, struct dentry *, 51 struct inode *, struct dentry *, unsigned int); 52 int (*readlink) (struct dentry *, char __user *,int); 53 void * (*follow_link) (struct dentry *, struct nameidata *); 54 void (*put_link) (struct dentry *, struct nameidata *, void *); 55 void (*truncate) (struct inode *); 56 int (*permission) (struct inode *, int, unsigned int); 57 int (*get_acl)(struct inode *, int); 58 int (*setattr) (struct dentry *, struct iattr *); 59 int (*getattr) (struct vfsmount *, struct dentry *, struct kstat *); 60 int (*setxattr) (struct dentry *, const char *,const void *,size_t,int); 61 ssize_t (*getxattr) (struct dentry *, const char *, void *, size_t); 62 ssize_t (*listxattr) (struct dentry *, char *, size_t); 63 int (*removexattr) (struct dentry *, const char *); 64 int (*fiemap)(struct inode *, struct fiemap_extent_info *, u64 start, u64 len); 65 void (*update_time)(struct inode *, struct timespec *, int); 66 int (*atomic_open)(struct inode *, struct dentry *, 67 struct file *, unsigned open_flag, 68 umode_t create_mode, int *opened); 69 int (*tmpfile) (struct inode *, struct dentry *, umode_t); 70 int (*dentry_open)(struct dentry *, struct file *, const struct cred *); 71 72locking rules: 73 all may block 74 i_mutex(inode) 75lookup: yes 76create: yes 77link: yes (both) 78mknod: yes 79symlink: yes 80mkdir: yes 81unlink: yes (both) 82rmdir: yes (both) (see below) 83rename: yes (all) (see below) 84rename2: yes (all) (see below) 85readlink: no 86follow_link: no 87put_link: no 88setattr: yes 89permission: no (may not block if called in rcu-walk mode) 90get_acl: no 91getattr: no 92setxattr: yes 93getxattr: no 94listxattr: no 95removexattr: yes 96fiemap: no 97update_time: no 98atomic_open: yes 99tmpfile: no 100dentry_open: no 101 102 Additionally, ->rmdir(), ->unlink() and ->rename() have ->i_mutex on 103victim. 104 cross-directory ->rename() and rename2() has (per-superblock) 105->s_vfs_rename_sem. 106 107See Documentation/filesystems/directory-locking for more detailed discussion 108of the locking scheme for directory operations. 109 110--------------------------- super_operations --------------------------- 111prototypes: 112 struct inode *(*alloc_inode)(struct super_block *sb); 113 void (*destroy_inode)(struct inode *); 114 void (*dirty_inode) (struct inode *, int flags); 115 int (*write_inode) (struct inode *, struct writeback_control *wbc); 116 int (*drop_inode) (struct inode *); 117 void (*evict_inode) (struct inode *); 118 void (*put_super) (struct super_block *); 119 int (*sync_fs)(struct super_block *sb, int wait); 120 int (*freeze_fs) (struct super_block *); 121 int (*unfreeze_fs) (struct super_block *); 122 int (*statfs) (struct dentry *, struct kstatfs *); 123 int (*remount_fs) (struct super_block *, int *, char *); 124 void (*umount_begin) (struct super_block *); 125 int (*show_options)(struct seq_file *, struct dentry *); 126 ssize_t (*quota_read)(struct super_block *, int, char *, size_t, loff_t); 127 ssize_t (*quota_write)(struct super_block *, int, const char *, size_t, loff_t); 128 int (*bdev_try_to_free_page)(struct super_block*, struct page*, gfp_t); 129 130locking rules: 131 All may block [not true, see below] 132 s_umount 133alloc_inode: 134destroy_inode: 135dirty_inode: 136write_inode: 137drop_inode: !!!inode->i_lock!!! 138evict_inode: 139put_super: write 140sync_fs: read 141freeze_fs: write 142unfreeze_fs: write 143statfs: maybe(read) (see below) 144remount_fs: write 145umount_begin: no 146show_options: no (namespace_sem) 147quota_read: no (see below) 148quota_write: no (see below) 149bdev_try_to_free_page: no (see below) 150 151->statfs() has s_umount (shared) when called by ustat(2) (native or 152compat), but that's an accident of bad API; s_umount is used to pin 153the superblock down when we only have dev_t given us by userland to 154identify the superblock. Everything else (statfs(), fstatfs(), etc.) 155doesn't hold it when calling ->statfs() - superblock is pinned down 156by resolving the pathname passed to syscall. 157->quota_read() and ->quota_write() functions are both guaranteed to 158be the only ones operating on the quota file by the quota code (via 159dqio_sem) (unless an admin really wants to screw up something and 160writes to quota files with quotas on). For other details about locking 161see also dquot_operations section. 162->bdev_try_to_free_page is called from the ->releasepage handler of 163the block device inode. See there for more details. 164 165--------------------------- file_system_type --------------------------- 166prototypes: 167 struct dentry *(*mount) (struct file_system_type *, int, 168 const char *, void *); 169 void (*kill_sb) (struct super_block *); 170locking rules: 171 may block 172mount yes 173kill_sb yes 174 175->mount() returns ERR_PTR or the root dentry; its superblock should be locked 176on return. 177->kill_sb() takes a write-locked superblock, does all shutdown work on it, 178unlocks and drops the reference. 179 180--------------------------- address_space_operations -------------------------- 181prototypes: 182 int (*writepage)(struct page *page, struct writeback_control *wbc); 183 int (*readpage)(struct file *, struct page *); 184 int (*sync_page)(struct page *); 185 int (*writepages)(struct address_space *, struct writeback_control *); 186 int (*set_page_dirty)(struct page *page); 187 int (*readpages)(struct file *filp, struct address_space *mapping, 188 struct list_head *pages, unsigned nr_pages); 189 int (*write_begin)(struct file *, struct address_space *mapping, 190 loff_t pos, unsigned len, unsigned flags, 191 struct page **pagep, void **fsdata); 192 int (*write_end)(struct file *, struct address_space *mapping, 193 loff_t pos, unsigned len, unsigned copied, 194 struct page *page, void *fsdata); 195 sector_t (*bmap)(struct address_space *, sector_t); 196 void (*invalidatepage) (struct page *, unsigned int, unsigned int); 197 int (*releasepage) (struct page *, int); 198 void (*freepage)(struct page *); 199 int (*direct_IO)(struct kiocb *, struct iov_iter *iter, loff_t offset); 200 int (*migratepage)(struct address_space *, struct page *, struct page *); 201 int (*launder_page)(struct page *); 202 int (*is_partially_uptodate)(struct page *, unsigned long, unsigned long); 203 int (*error_remove_page)(struct address_space *, struct page *); 204 int (*swap_activate)(struct file *); 205 int (*swap_deactivate)(struct file *); 206 207locking rules: 208 All except set_page_dirty and freepage may block 209 210 PageLocked(page) i_mutex 211writepage: yes, unlocks (see below) 212readpage: yes, unlocks 213sync_page: maybe 214writepages: 215set_page_dirty no 216readpages: 217write_begin: locks the page yes 218write_end: yes, unlocks yes 219bmap: 220invalidatepage: yes 221releasepage: yes 222freepage: yes 223direct_IO: 224migratepage: yes (both) 225launder_page: yes 226is_partially_uptodate: yes 227error_remove_page: yes 228swap_activate: no 229swap_deactivate: no 230 231 ->write_begin(), ->write_end(), ->sync_page() and ->readpage() 232may be called from the request handler (/dev/loop). 233 234 ->readpage() unlocks the page, either synchronously or via I/O 235completion. 236 237 ->readpages() populates the pagecache with the passed pages and starts 238I/O against them. They come unlocked upon I/O completion. 239 240 ->writepage() is used for two purposes: for "memory cleansing" and for 241"sync". These are quite different operations and the behaviour may differ 242depending upon the mode. 243 244If writepage is called for sync (wbc->sync_mode != WBC_SYNC_NONE) then 245it *must* start I/O against the page, even if that would involve 246blocking on in-progress I/O. 247 248If writepage is called for memory cleansing (sync_mode == 249WBC_SYNC_NONE) then its role is to get as much writeout underway as 250possible. So writepage should try to avoid blocking against 251currently-in-progress I/O. 252 253If the filesystem is not called for "sync" and it determines that it 254would need to block against in-progress I/O to be able to start new I/O 255against the page the filesystem should redirty the page with 256redirty_page_for_writepage(), then unlock the page and return zero. 257This may also be done to avoid internal deadlocks, but rarely. 258 259If the filesystem is called for sync then it must wait on any 260in-progress I/O and then start new I/O. 261 262The filesystem should unlock the page synchronously, before returning to the 263caller, unless ->writepage() returns special WRITEPAGE_ACTIVATE 264value. WRITEPAGE_ACTIVATE means that page cannot really be written out 265currently, and VM should stop calling ->writepage() on this page for some 266time. VM does this by moving page to the head of the active list, hence the 267name. 268 269Unless the filesystem is going to redirty_page_for_writepage(), unlock the page 270and return zero, writepage *must* run set_page_writeback() against the page, 271followed by unlocking it. Once set_page_writeback() has been run against the 272page, write I/O can be submitted and the write I/O completion handler must run 273end_page_writeback() once the I/O is complete. If no I/O is submitted, the 274filesystem must run end_page_writeback() against the page before returning from 275writepage. 276 277That is: after 2.5.12, pages which are under writeout are *not* locked. Note, 278if the filesystem needs the page to be locked during writeout, that is ok, too, 279the page is allowed to be unlocked at any point in time between the calls to 280set_page_writeback() and end_page_writeback(). 281 282Note, failure to run either redirty_page_for_writepage() or the combination of 283set_page_writeback()/end_page_writeback() on a page submitted to writepage 284will leave the page itself marked clean but it will be tagged as dirty in the 285radix tree. This incoherency can lead to all sorts of hard-to-debug problems 286in the filesystem like having dirty inodes at umount and losing written data. 287 288 ->sync_page() locking rules are not well-defined - usually it is called 289with lock on page, but that is not guaranteed. Considering the currently 290existing instances of this method ->sync_page() itself doesn't look 291well-defined... 292 293 ->writepages() is used for periodic writeback and for syscall-initiated 294sync operations. The address_space should start I/O against at least 295*nr_to_write pages. *nr_to_write must be decremented for each page which is 296written. The address_space implementation may write more (or less) pages 297than *nr_to_write asks for, but it should try to be reasonably close. If 298nr_to_write is NULL, all dirty pages must be written. 299 300writepages should _only_ write pages which are present on 301mapping->io_pages. 302 303 ->set_page_dirty() is called from various places in the kernel 304when the target page is marked as needing writeback. It may be called 305under spinlock (it cannot block) and is sometimes called with the page 306not locked. 307 308 ->bmap() is currently used by legacy ioctl() (FIBMAP) provided by some 309filesystems and by the swapper. The latter will eventually go away. Please, 310keep it that way and don't breed new callers. 311 312 ->invalidatepage() is called when the filesystem must attempt to drop 313some or all of the buffers from the page when it is being truncated. It 314returns zero on success. If ->invalidatepage is zero, the kernel uses 315block_invalidatepage() instead. 316 317 ->releasepage() is called when the kernel is about to try to drop the 318buffers from the page in preparation for freeing it. It returns zero to 319indicate that the buffers are (or may be) freeable. If ->releasepage is zero, 320the kernel assumes that the fs has no private interest in the buffers. 321 322 ->freepage() is called when the kernel is done dropping the page 323from the page cache. 324 325 ->launder_page() may be called prior to releasing a page if 326it is still found to be dirty. It returns zero if the page was successfully 327cleaned, or an error value if not. Note that in order to prevent the page 328getting mapped back in and redirtied, it needs to be kept locked 329across the entire operation. 330 331 ->swap_activate will be called with a non-zero argument on 332files backing (non block device backed) swapfiles. A return value 333of zero indicates success, in which case this file can be used for 334backing swapspace. The swapspace operations will be proxied to the 335address space operations. 336 337 ->swap_deactivate() will be called in the sys_swapoff() 338path after ->swap_activate() returned success. 339 340----------------------- file_lock_operations ------------------------------ 341prototypes: 342 void (*fl_copy_lock)(struct file_lock *, struct file_lock *); 343 void (*fl_release_private)(struct file_lock *); 344 345 346locking rules: 347 inode->i_lock may block 348fl_copy_lock: yes no 349fl_release_private: maybe maybe[1] 350 351[1]: ->fl_release_private for flock or POSIX locks is currently allowed 352to block. Leases however can still be freed while the i_lock is held and 353so fl_release_private called on a lease should not block. 354 355----------------------- lock_manager_operations --------------------------- 356prototypes: 357 int (*lm_compare_owner)(struct file_lock *, struct file_lock *); 358 unsigned long (*lm_owner_key)(struct file_lock *); 359 void (*lm_notify)(struct file_lock *); /* unblock callback */ 360 int (*lm_grant)(struct file_lock *, struct file_lock *, int); 361 void (*lm_break)(struct file_lock *); /* break_lease callback */ 362 int (*lm_change)(struct file_lock **, int); 363 364locking rules: 365 366 inode->i_lock blocked_lock_lock may block 367lm_compare_owner: yes[1] maybe no 368lm_owner_key yes[1] yes no 369lm_notify: yes yes no 370lm_grant: no no no 371lm_break: yes no no 372lm_change yes no no 373 374[1]: ->lm_compare_owner and ->lm_owner_key are generally called with 375*an* inode->i_lock held. It may not be the i_lock of the inode 376associated with either file_lock argument! This is the case with deadlock 377detection, since the code has to chase down the owners of locks that may 378be entirely unrelated to the one on which the lock is being acquired. 379For deadlock detection however, the blocked_lock_lock is also held. The 380fact that these locks are held ensures that the file_locks do not 381disappear out from under you while doing the comparison or generating an 382owner key. 383 384--------------------------- buffer_head ----------------------------------- 385prototypes: 386 void (*b_end_io)(struct buffer_head *bh, int uptodate); 387 388locking rules: 389 called from interrupts. In other words, extreme care is needed here. 390bh is locked, but that's all warranties we have here. Currently only RAID1, 391highmem, fs/buffer.c, and fs/ntfs/aops.c are providing these. Block devices 392call this method upon the IO completion. 393 394--------------------------- block_device_operations ----------------------- 395prototypes: 396 int (*open) (struct block_device *, fmode_t); 397 int (*release) (struct gendisk *, fmode_t); 398 int (*ioctl) (struct block_device *, fmode_t, unsigned, unsigned long); 399 int (*compat_ioctl) (struct block_device *, fmode_t, unsigned, unsigned long); 400 int (*direct_access) (struct block_device *, sector_t, void **, unsigned long *); 401 int (*media_changed) (struct gendisk *); 402 void (*unlock_native_capacity) (struct gendisk *); 403 int (*revalidate_disk) (struct gendisk *); 404 int (*getgeo)(struct block_device *, struct hd_geometry *); 405 void (*swap_slot_free_notify) (struct block_device *, unsigned long); 406 407locking rules: 408 bd_mutex 409open: yes 410release: yes 411ioctl: no 412compat_ioctl: no 413direct_access: no 414media_changed: no 415unlock_native_capacity: no 416revalidate_disk: no 417getgeo: no 418swap_slot_free_notify: no (see below) 419 420media_changed, unlock_native_capacity and revalidate_disk are called only from 421check_disk_change(). 422 423swap_slot_free_notify is called with swap_lock and sometimes the page lock 424held. 425 426 427--------------------------- file_operations ------------------------------- 428prototypes: 429 loff_t (*llseek) (struct file *, loff_t, int); 430 ssize_t (*read) (struct file *, char __user *, size_t, loff_t *); 431 ssize_t (*write) (struct file *, const char __user *, size_t, loff_t *); 432 ssize_t (*read_iter) (struct kiocb *, struct iov_iter *); 433 ssize_t (*write_iter) (struct kiocb *, struct iov_iter *); 434 int (*iterate) (struct file *, struct dir_context *); 435 unsigned int (*poll) (struct file *, struct poll_table_struct *); 436 long (*unlocked_ioctl) (struct file *, unsigned int, unsigned long); 437 long (*compat_ioctl) (struct file *, unsigned int, unsigned long); 438 int (*mmap) (struct file *, struct vm_area_struct *); 439 int (*open) (struct inode *, struct file *); 440 int (*flush) (struct file *); 441 int (*release) (struct inode *, struct file *); 442 int (*fsync) (struct file *, loff_t start, loff_t end, int datasync); 443 int (*aio_fsync) (struct kiocb *, int datasync); 444 int (*fasync) (int, struct file *, int); 445 int (*lock) (struct file *, int, struct file_lock *); 446 ssize_t (*readv) (struct file *, const struct iovec *, unsigned long, 447 loff_t *); 448 ssize_t (*writev) (struct file *, const struct iovec *, unsigned long, 449 loff_t *); 450 ssize_t (*sendfile) (struct file *, loff_t *, size_t, read_actor_t, 451 void __user *); 452 ssize_t (*sendpage) (struct file *, struct page *, int, size_t, 453 loff_t *, int); 454 unsigned long (*get_unmapped_area)(struct file *, unsigned long, 455 unsigned long, unsigned long, unsigned long); 456 int (*check_flags)(int); 457 int (*flock) (struct file *, int, struct file_lock *); 458 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *, loff_t *, 459 size_t, unsigned int); 460 ssize_t (*splice_read)(struct file *, loff_t *, struct pipe_inode_info *, 461 size_t, unsigned int); 462 int (*setlease)(struct file *, long, struct file_lock **, void **); 463 long (*fallocate)(struct file *, int, loff_t, loff_t); 464}; 465 466locking rules: 467 All may block. 468 469->llseek() locking has moved from llseek to the individual llseek 470implementations. If your fs is not using generic_file_llseek, you 471need to acquire and release the appropriate locks in your ->llseek(). 472For many filesystems, it is probably safe to acquire the inode 473mutex or just to use i_size_read() instead. 474Note: this does not protect the file->f_pos against concurrent modifications 475since this is something the userspace has to take care about. 476 477->fasync() is responsible for maintaining the FASYNC bit in filp->f_flags. 478Most instances call fasync_helper(), which does that maintenance, so it's 479not normally something one needs to worry about. Return values > 0 will be 480mapped to zero in the VFS layer. 481 482->readdir() and ->ioctl() on directories must be changed. Ideally we would 483move ->readdir() to inode_operations and use a separate method for directory 484->ioctl() or kill the latter completely. One of the problems is that for 485anything that resembles union-mount we won't have a struct file for all 486components. And there are other reasons why the current interface is a mess... 487 488->read on directories probably must go away - we should just enforce -EISDIR 489in sys_read() and friends. 490 491->setlease operations should call generic_setlease() before or after setting 492the lease within the individual filesystem to record the result of the 493operation 494 495--------------------------- dquot_operations ------------------------------- 496prototypes: 497 int (*write_dquot) (struct dquot *); 498 int (*acquire_dquot) (struct dquot *); 499 int (*release_dquot) (struct dquot *); 500 int (*mark_dirty) (struct dquot *); 501 int (*write_info) (struct super_block *, int); 502 503These operations are intended to be more or less wrapping functions that ensure 504a proper locking wrt the filesystem and call the generic quota operations. 505 506What filesystem should expect from the generic quota functions: 507 508 FS recursion Held locks when called 509write_dquot: yes dqonoff_sem or dqptr_sem 510acquire_dquot: yes dqonoff_sem or dqptr_sem 511release_dquot: yes dqonoff_sem or dqptr_sem 512mark_dirty: no - 513write_info: yes dqonoff_sem 514 515FS recursion means calling ->quota_read() and ->quota_write() from superblock 516operations. 517 518More details about quota locking can be found in fs/dquot.c. 519 520--------------------------- vm_operations_struct ----------------------------- 521prototypes: 522 void (*open)(struct vm_area_struct*); 523 void (*close)(struct vm_area_struct*); 524 int (*fault)(struct vm_area_struct*, struct vm_fault *); 525 int (*page_mkwrite)(struct vm_area_struct *, struct vm_fault *); 526 int (*pfn_mkwrite)(struct vm_area_struct *, struct vm_fault *); 527 int (*access)(struct vm_area_struct *, unsigned long, void*, int, int); 528 529locking rules: 530 mmap_sem PageLocked(page) 531open: yes 532close: yes 533fault: yes can return with page locked 534map_pages: yes 535page_mkwrite: yes can return with page locked 536pfn_mkwrite: yes 537access: yes 538 539 ->fault() is called when a previously not present pte is about 540to be faulted in. The filesystem must find and return the page associated 541with the passed in "pgoff" in the vm_fault structure. If it is possible that 542the page may be truncated and/or invalidated, then the filesystem must lock 543the page, then ensure it is not already truncated (the page lock will block 544subsequent truncate), and then return with VM_FAULT_LOCKED, and the page 545locked. The VM will unlock the page. 546 547 ->map_pages() is called when VM asks to map easy accessible pages. 548Filesystem should find and map pages associated with offsets from "pgoff" 549till "max_pgoff". ->map_pages() is called with page table locked and must 550not block. If it's not possible to reach a page without blocking, 551filesystem should skip it. Filesystem should use do_set_pte() to setup 552page table entry. Pointer to entry associated with offset "pgoff" is 553passed in "pte" field in vm_fault structure. Pointers to entries for other 554offsets should be calculated relative to "pte". 555 556 ->page_mkwrite() is called when a previously read-only pte is 557about to become writeable. The filesystem again must ensure that there are 558no truncate/invalidate races, and then return with the page locked. If 559the page has been truncated, the filesystem should not look up a new page 560like the ->fault() handler, but simply return with VM_FAULT_NOPAGE, which 561will cause the VM to retry the fault. 562 563 ->pfn_mkwrite() is the same as page_mkwrite but when the pte is 564VM_PFNMAP or VM_MIXEDMAP with a page-less entry. Expected return is 565VM_FAULT_NOPAGE. Or one of the VM_FAULT_ERROR types. The default behavior 566after this call is to make the pte read-write, unless pfn_mkwrite returns 567an error. 568 569 ->access() is called when get_user_pages() fails in 570access_process_vm(), typically used to debug a process through 571/proc/pid/mem or ptrace. This function is needed only for 572VM_IO | VM_PFNMAP VMAs. 573 574================================================================================ 575 Dubious stuff 576 577(if you break something or notice that it is broken and do not fix it yourself 578- at least put it here) 579