1================= 2Directory Locking 3================= 4 5 6Locking scheme used for directory operations is based on two 7kinds of locks - per-inode (->i_rwsem) and per-filesystem 8(->s_vfs_rename_mutex). 9 10When taking the i_rwsem on multiple non-directory objects, we 11always acquire the locks in order by increasing address. We'll call 12that "inode pointer" order in the following. 13 14For our purposes all operations fall in 5 classes: 15 161) read access. Locking rules: caller locks directory we are accessing. 17The lock is taken shared. 18 192) object creation. Locking rules: same as above, but the lock is taken 20exclusive. 21 223) object removal. Locking rules: caller locks parent, finds victim, 23locks victim and calls the method. Locks are exclusive. 24 254) rename() that is _not_ cross-directory. Locking rules: caller locks 26the parent and finds source and target. Then we decide which of the 27source and target need to be locked. Source needs to be locked if it's a 28non-directory; target - if it's a non-directory or about to be removed. 29Take the locks that need to be taken, in inode pointer order if need 30to take both (that can happen only when both source and target are 31non-directories - the source because it wouldn't be locked otherwise 32and the target because mixing directory and non-directory is allowed 33only with RENAME_EXCHANGE, and that won't be removing the target). 34After the locks had been taken, call the method. All locks are exclusive. 35 365) link creation. Locking rules: 37 38 * lock parent 39 * check that source is not a directory 40 * lock source 41 * call the method. 42 43All locks are exclusive. 44 456) cross-directory rename. The trickiest in the whole bunch. Locking 46rules: 47 48 * lock the filesystem 49 * lock parents in "ancestors first" order. If one is not ancestor of 50 the other, lock the parent of source first. 51 * find source and target. 52 * if old parent is equal to or is a descendent of target 53 fail with -ENOTEMPTY 54 * if new parent is equal to or is a descendent of source 55 fail with -ELOOP 56 * Lock both the source and the target provided they exist. If we 57 need to lock two inodes of different type (dir vs non-dir), we lock 58 the directory first. If we need to lock two inodes of the same type, 59 lock them in inode pointer order. 60 * Lock subdirectories involved (source before target). 61 * Lock non-directories involved, in inode pointer order. 62 * call the method. 63 64All ->i_rwsem are taken exclusive. 65 66The rules above obviously guarantee that all directories that are going to be 67read, modified or removed by method will be locked by caller. 68 69 70If no directory is its own ancestor, the scheme above is deadlock-free. 71 72Proof: 73 74[XXX: will be updated once we are done massaging the lock_rename()] 75 First of all, at any moment we have a linear ordering of the 76 objects - A < B iff (A is an ancestor of B) or (B is not an ancestor 77 of A and ptr(A) < ptr(B)). 78 79 That ordering can change. However, the following is true: 80 81(1) if object removal or non-cross-directory rename holds lock on A and 82 attempts to acquire lock on B, A will remain the parent of B until we 83 acquire the lock on B. (Proof: only cross-directory rename can change 84 the parent of object and it would have to lock the parent). 85 86(2) if cross-directory rename holds the lock on filesystem, order will not 87 change until rename acquires all locks. (Proof: other cross-directory 88 renames will be blocked on filesystem lock and we don't start changing 89 the order until we had acquired all locks). 90 91(3) locks on non-directory objects are acquired only after locks on 92 directory objects, and are acquired in inode pointer order. 93 (Proof: all operations but renames take lock on at most one 94 non-directory object, except renames, which take locks on source and 95 target in inode pointer order in the case they are not directories.) 96 97Now consider the minimal deadlock. Each process is blocked on 98attempt to acquire some lock and already holds at least one lock. Let's 99consider the set of contended locks. First of all, filesystem lock is 100not contended, since any process blocked on it is not holding any locks. 101Thus all processes are blocked on ->i_rwsem. 102 103By (3), any process holding a non-directory lock can only be 104waiting on another non-directory lock with a larger address. Therefore 105the process holding the "largest" such lock can always make progress, and 106non-directory objects are not included in the set of contended locks. 107 108Thus link creation can't be a part of deadlock - it can't be 109blocked on source and it means that it doesn't hold any locks. 110 111Any contended object is either held by cross-directory rename or 112has a child that is also contended. Indeed, suppose that it is held by 113operation other than cross-directory rename. Then the lock this operation 114is blocked on belongs to child of that object due to (1). 115 116It means that one of the operations is cross-directory rename. 117Otherwise the set of contended objects would be infinite - each of them 118would have a contended child and we had assumed that no object is its 119own descendent. Moreover, there is exactly one cross-directory rename 120(see above). 121 122Consider the object blocking the cross-directory rename. One 123of its descendents is locked by cross-directory rename (otherwise we 124would again have an infinite set of contended objects). But that 125means that cross-directory rename is taking locks out of order. Due 126to (2) the order hadn't changed since we had acquired filesystem lock. 127But locking rules for cross-directory rename guarantee that we do not 128try to acquire lock on descendent before the lock on ancestor. 129Contradiction. I.e. deadlock is impossible. Q.E.D. 130 131 132These operations are guaranteed to avoid loop creation. Indeed, 133the only operation that could introduce loops is cross-directory rename. 134Since the only new (parent, child) pair added by rename() is (new parent, 135source), such loop would have to contain these objects and the rest of it 136would have to exist before rename(). I.e. at the moment of loop creation 137rename() responsible for that would be holding filesystem lock and new parent 138would have to be equal to or a descendent of source. But that means that 139new parent had been equal to or a descendent of source since the moment when 140we had acquired filesystem lock and rename() would fail with -ELOOP in that 141case. 142 143While this locking scheme works for arbitrary DAGs, it relies on 144ability to check that directory is a descendent of another object. Current 145implementation assumes that directory graph is a tree. This assumption is 146also preserved by all operations (cross-directory rename on a tree that would 147not introduce a cycle will leave it a tree and link() fails for directories). 148 149Notice that "directory" in the above == "anything that might have 150children", so if we are going to introduce hybrid objects we will need 151either to make sure that link(2) doesn't work for them or to make changes 152in is_subdir() that would make it work even in presence of such beasts. 153