directory-locking.rst (6484B)
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. In case of exchange (with 27RENAME_EXCHANGE in flags argument) lock both. In any case, 28if the target already exists, lock it. If the source is a non-directory, 29lock it. If we need to lock both, lock them in inode pointer order. 30Then call the method. All locks are exclusive. 31NB: we might get away with locking the source (and target in exchange 32case) shared. 33 345) link creation. Locking rules: 35 36 * lock parent 37 * check that source is not a directory 38 * lock source 39 * call the method. 40 41All locks are exclusive. 42 436) cross-directory rename. The trickiest in the whole bunch. Locking 44rules: 45 46 * lock the filesystem 47 * lock parents in "ancestors first" order. 48 * find source and target. 49 * if old parent is equal to or is a descendent of target 50 fail with -ENOTEMPTY 51 * if new parent is equal to or is a descendent of source 52 fail with -ELOOP 53 * If it's an exchange, lock both the source and the target. 54 * If the target exists, lock it. If the source is a non-directory, 55 lock it. If we need to lock both, do so in inode pointer order. 56 * call the method. 57 58All ->i_rwsem are taken exclusive. Again, we might get away with locking 59the source (and target in exchange case) shared. 60 61The rules above obviously guarantee that all directories that are going to be 62read, modified or removed by method will be locked by caller. 63 64 65If no directory is its own ancestor, the scheme above is deadlock-free. 66 67Proof: 68 69 First of all, at any moment we have a partial ordering of the 70 objects - A < B iff A is an ancestor of B. 71 72 That ordering can change. However, the following is true: 73 74(1) if object removal or non-cross-directory rename holds lock on A and 75 attempts to acquire lock on B, A will remain the parent of B until we 76 acquire the lock on B. (Proof: only cross-directory rename can change 77 the parent of object and it would have to lock the parent). 78 79(2) if cross-directory rename holds the lock on filesystem, order will not 80 change until rename acquires all locks. (Proof: other cross-directory 81 renames will be blocked on filesystem lock and we don't start changing 82 the order until we had acquired all locks). 83 84(3) locks on non-directory objects are acquired only after locks on 85 directory objects, and are acquired in inode pointer order. 86 (Proof: all operations but renames take lock on at most one 87 non-directory object, except renames, which take locks on source and 88 target in inode pointer order in the case they are not directories.) 89 90Now consider the minimal deadlock. Each process is blocked on 91attempt to acquire some lock and already holds at least one lock. Let's 92consider the set of contended locks. First of all, filesystem lock is 93not contended, since any process blocked on it is not holding any locks. 94Thus all processes are blocked on ->i_rwsem. 95 96By (3), any process holding a non-directory lock can only be 97waiting on another non-directory lock with a larger address. Therefore 98the process holding the "largest" such lock can always make progress, and 99non-directory objects are not included in the set of contended locks. 100 101Thus link creation can't be a part of deadlock - it can't be 102blocked on source and it means that it doesn't hold any locks. 103 104Any contended object is either held by cross-directory rename or 105has a child that is also contended. Indeed, suppose that it is held by 106operation other than cross-directory rename. Then the lock this operation 107is blocked on belongs to child of that object due to (1). 108 109It means that one of the operations is cross-directory rename. 110Otherwise the set of contended objects would be infinite - each of them 111would have a contended child and we had assumed that no object is its 112own descendent. Moreover, there is exactly one cross-directory rename 113(see above). 114 115Consider the object blocking the cross-directory rename. One 116of its descendents is locked by cross-directory rename (otherwise we 117would again have an infinite set of contended objects). But that 118means that cross-directory rename is taking locks out of order. Due 119to (2) the order hadn't changed since we had acquired filesystem lock. 120But locking rules for cross-directory rename guarantee that we do not 121try to acquire lock on descendent before the lock on ancestor. 122Contradiction. I.e. deadlock is impossible. Q.E.D. 123 124 125These operations are guaranteed to avoid loop creation. Indeed, 126the only operation that could introduce loops is cross-directory rename. 127Since the only new (parent, child) pair added by rename() is (new parent, 128source), such loop would have to contain these objects and the rest of it 129would have to exist before rename(). I.e. at the moment of loop creation 130rename() responsible for that would be holding filesystem lock and new parent 131would have to be equal to or a descendent of source. But that means that 132new parent had been equal to or a descendent of source since the moment when 133we had acquired filesystem lock and rename() would fail with -ELOOP in that 134case. 135 136While this locking scheme works for arbitrary DAGs, it relies on 137ability to check that directory is a descendent of another object. Current 138implementation assumes that directory graph is a tree. This assumption is 139also preserved by all operations (cross-directory rename on a tree that would 140not introduce a cycle will leave it a tree and link() fails for directories). 141 142Notice that "directory" in the above == "anything that might have 143children", so if we are going to introduce hybrid objects we will need 144either to make sure that link(2) doesn't work for them or to make changes 145in is_subdir() that would make it work even in presence of such beasts.