linux_dsm_epyc7002/Documentation/filesystems/directory-locking

128 lines
5.8 KiB
Plaintext
Raw Normal View History

Locking scheme used for directory operations is based on two
kinds of locks - per-inode (->i_mutex) and per-filesystem
(->s_vfs_rename_mutex).
When taking the i_mutex on multiple non-directory objects, we
always acquire the locks in order by increasing address. We'll call
that "inode pointer" order in the following.
For our purposes all operations fall in 5 classes:
1) read access. Locking rules: caller locks directory we are accessing.
2) object creation. Locking rules: same as above.
3) object removal. Locking rules: caller locks parent, finds victim,
locks victim and calls the method.
4) rename() that is _not_ cross-directory. Locking rules: caller locks
the parent and finds source and target. If target already exists, lock
it. If source is a non-directory, lock it. If that means we need to
lock both, lock them in inode pointer order.
5) link creation. Locking rules:
* lock parent
* check that source is not a directory
* lock source
* call the method.
6) cross-directory rename. The trickiest in the whole bunch. Locking
rules:
* lock the filesystem
* lock parents in "ancestors first" order.
* find source and target.
* if old parent is equal to or is a descendent of target
fail with -ENOTEMPTY
* if new parent is equal to or is a descendent of source
fail with -ELOOP
* If target exists, lock it. If source is a non-directory, lock
it. In case that means we need to lock both source and target,
do so in inode pointer order.
* call the method.
The rules above obviously guarantee that all directories that are going to be
read, modified or removed by method will be locked by caller.
If no directory is its own ancestor, the scheme above is deadlock-free.
Proof:
First of all, at any moment we have a partial ordering of the
objects - A < B iff A is an ancestor of B.
That ordering can change. However, the following is true:
(1) if object removal or non-cross-directory rename holds lock on A and
attempts to acquire lock on B, A will remain the parent of B until we
acquire the lock on B. (Proof: only cross-directory rename can change
the parent of object and it would have to lock the parent).
(2) if cross-directory rename holds the lock on filesystem, order will not
change until rename acquires all locks. (Proof: other cross-directory
renames will be blocked on filesystem lock and we don't start changing
the order until we had acquired all locks).
(3) locks on non-directory objects are acquired only after locks on
directory objects, and are acquired in inode pointer order.
(Proof: all operations but renames take lock on at most one
non-directory object, except renames, which take locks on source and
target in inode pointer order in the case they are not directories.)
Now consider the minimal deadlock. Each process is blocked on
attempt to acquire some lock and already holds at least one lock. Let's
consider the set of contended locks. First of all, filesystem lock is
not contended, since any process blocked on it is not holding any locks.
Thus all processes are blocked on ->i_mutex.
By (3), any process holding a non-directory lock can only be
waiting on another non-directory lock with a larger address. Therefore
the process holding the "largest" such lock can always make progress, and
non-directory objects are not included in the set of contended locks.
Thus link creation can't be a part of deadlock - it can't be
blocked on source and it means that it doesn't hold any locks.
Any contended object is either held by cross-directory rename or
has a child that is also contended. Indeed, suppose that it is held by
operation other than cross-directory rename. Then the lock this operation
is blocked on belongs to child of that object due to (1).
It means that one of the operations is cross-directory rename.
Otherwise the set of contended objects would be infinite - each of them
would have a contended child and we had assumed that no object is its
own descendent. Moreover, there is exactly one cross-directory rename
(see above).
Consider the object blocking the cross-directory rename. One
of its descendents is locked by cross-directory rename (otherwise we
would again have an infinite set of contended objects). But that
means that cross-directory rename is taking locks out of order. Due
to (2) the order hadn't changed since we had acquired filesystem lock.
But locking rules for cross-directory rename guarantee that we do not
try to acquire lock on descendent before the lock on ancestor.
Contradiction. I.e. deadlock is impossible. Q.E.D.
These operations are guaranteed to avoid loop creation. Indeed,
the only operation that could introduce loops is cross-directory rename.
Since the only new (parent, child) pair added by rename() is (new parent,
source), such loop would have to contain these objects and the rest of it
would have to exist before rename(). I.e. at the moment of loop creation
rename() responsible for that would be holding filesystem lock and new parent
would have to be equal to or a descendent of source. But that means that
new parent had been equal to or a descendent of source since the moment when
we had acquired filesystem lock and rename() would fail with -ELOOP in that
case.
While this locking scheme works for arbitrary DAGs, it relies on
ability to check that directory is a descendent of another object. Current
implementation assumes that directory graph is a tree. This assumption is
also preserved by all operations (cross-directory rename on a tree that would
not introduce a cycle will leave it a tree and link() fails for directories).
Notice that "directory" in the above == "anything that might have
children", so if we are going to introduce hybrid objects we will need
either to make sure that link(2) doesn't work for them or to make changes
in is_subdir() that would make it work even in presence of such beasts.