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