Based on kernel version 4.8. Page generated on 2016-10-06 23:15 EST.
1 Locking scheme used for directory operations is based on two 2 kinds of locks - per-inode (->i_rwsem) and per-filesystem 3 (->s_vfs_rename_mutex). 4 5 When taking the i_rwsem on multiple non-directory objects, we 6 always acquire the locks in order by increasing address. We'll call 7 that "inode pointer" order in the following. 8 9 For our purposes all operations fall in 5 classes: 10 11 1) read access. Locking rules: caller locks directory we are accessing. 12 The lock is taken shared. 13 14 2) object creation. Locking rules: same as above, but the lock is taken 15 exclusive. 16 17 3) object removal. Locking rules: caller locks parent, finds victim, 18 locks victim and calls the method. Locks are exclusive. 19 20 4) rename() that is _not_ cross-directory. Locking rules: caller locks 21 the parent and finds source and target. In case of exchange (with 22 RENAME_EXCHANGE in rename2() flags argument) lock both. In any case, 23 if the target already exists, lock it. If the source is a non-directory, 24 lock it. If we need to lock both, lock them in inode pointer order. 25 Then call the method. All locks are exclusive. 26 NB: we might get away with locking the the source (and target in exchange 27 case) shared. 28 29 5) link creation. Locking rules: 30 * lock parent 31 * check that source is not a directory 32 * lock source 33 * call the method. 34 All locks are exclusive. 35 36 6) cross-directory rename. The trickiest in the whole bunch. Locking 37 rules: 38 * lock the filesystem 39 * lock parents in "ancestors first" order. 40 * find source and target. 41 * if old parent is equal to or is a descendent of target 42 fail with -ENOTEMPTY 43 * if new parent is equal to or is a descendent of source 44 fail with -ELOOP 45 * If it's an exchange, lock both the source and the target. 46 * If the target exists, lock it. If the source is a non-directory, 47 lock it. If we need to lock both, do so in inode pointer order. 48 * call the method. 49 All ->i_rwsem are taken exclusive. Again, we might get away with locking 50 the the source (and target in exchange case) shared. 51 52 The rules above obviously guarantee that all directories that are going to be 53 read, modified or removed by method will be locked by caller. 54 55 56 If no directory is its own ancestor, the scheme above is deadlock-free. 57 Proof: 58 59 First of all, at any moment we have a partial ordering of the 60 objects - A < B iff A is an ancestor of B. 61 62 That ordering can change. However, the following is true: 63 64 (1) if object removal or non-cross-directory rename holds lock on A and 65 attempts to acquire lock on B, A will remain the parent of B until we 66 acquire the lock on B. (Proof: only cross-directory rename can change 67 the parent of object and it would have to lock the parent). 68 69 (2) if cross-directory rename holds the lock on filesystem, order will not 70 change until rename acquires all locks. (Proof: other cross-directory 71 renames will be blocked on filesystem lock and we don't start changing 72 the order until we had acquired all locks). 73 74 (3) locks on non-directory objects are acquired only after locks on 75 directory objects, and are acquired in inode pointer order. 76 (Proof: all operations but renames take lock on at most one 77 non-directory object, except renames, which take locks on source and 78 target in inode pointer order in the case they are not directories.) 79 80 Now consider the minimal deadlock. Each process is blocked on 81 attempt to acquire some lock and already holds at least one lock. Let's 82 consider the set of contended locks. First of all, filesystem lock is 83 not contended, since any process blocked on it is not holding any locks. 84 Thus all processes are blocked on ->i_rwsem. 85 86 By (3), any process holding a non-directory lock can only be 87 waiting on another non-directory lock with a larger address. Therefore 88 the process holding the "largest" such lock can always make progress, and 89 non-directory objects are not included in the set of contended locks. 90 91 Thus link creation can't be a part of deadlock - it can't be 92 blocked on source and it means that it doesn't hold any locks. 93 94 Any contended object is either held by cross-directory rename or 95 has a child that is also contended. Indeed, suppose that it is held by 96 operation other than cross-directory rename. Then the lock this operation 97 is blocked on belongs to child of that object due to (1). 98 99 It means that one of the operations is cross-directory rename. 100 Otherwise the set of contended objects would be infinite - each of them 101 would have a contended child and we had assumed that no object is its 102 own descendent. Moreover, there is exactly one cross-directory rename 103 (see above). 104 105 Consider the object blocking the cross-directory rename. One 106 of its descendents is locked by cross-directory rename (otherwise we 107 would again have an infinite set of contended objects). But that 108 means that cross-directory rename is taking locks out of order. Due 109 to (2) the order hadn't changed since we had acquired filesystem lock. 110 But locking rules for cross-directory rename guarantee that we do not 111 try to acquire lock on descendent before the lock on ancestor. 112 Contradiction. I.e. deadlock is impossible. Q.E.D. 113 114 115 These operations are guaranteed to avoid loop creation. Indeed, 116 the only operation that could introduce loops is cross-directory rename. 117 Since the only new (parent, child) pair added by rename() is (new parent, 118 source), such loop would have to contain these objects and the rest of it 119 would have to exist before rename(). I.e. at the moment of loop creation 120 rename() responsible for that would be holding filesystem lock and new parent 121 would have to be equal to or a descendent of source. But that means that 122 new parent had been equal to or a descendent of source since the moment when 123 we had acquired filesystem lock and rename() would fail with -ELOOP in that 124 case. 125 126 While this locking scheme works for arbitrary DAGs, it relies on 127 ability to check that directory is a descendent of another object. Current 128 implementation assumes that directory graph is a tree. This assumption is 129 also preserved by all operations (cross-directory rename on a tree that would 130 not introduce a cycle will leave it a tree and link() fails for directories). 131 132 Notice that "directory" in the above == "anything that might have 133 children", so if we are going to introduce hybrid objects we will need 134 either to make sure that link(2) doesn't work for them or to make changes 135 in is_subdir() that would make it work even in presence of such beasts.