Btrfs: stop caching thread if extent_commit_sem is contended
authorJosef Bacik <jbacik@fusionio.com>
Thu, 19 Sep 2013 14:02:11 +0000 (10:02 -0400)
committerChris Mason <clm@fb.com>
Tue, 28 Jan 2014 21:20:10 +0000 (13:20 -0800)
We can starve out the transaction commit with a bunch of caching threads all
running at the same time.  This is because we will only drop the
extent_commit_sem if we need_resched(), which isn't likely to happen since we
will be reading a lot from the disk so have already schedule()'ed plenty.  Alex
observed that he could starve out a transaction commit for up to a minute with
32 caching threads all running at once.  This will allow us to drop the
extent_commit_sem to allow the transaction commit to swap the commit_root out
and then all the cachers will start back up. Here is an explanation provided by
Igno

So, just to fill in what happens in this loop:

                                mutex_unlock(&caching_ctl->mutex);
                                cond_resched();
                                goto again;

where 'again:' takes caching_ctl->mutex and fs_info->extent_commit_sem
again:

        again:
                mutex_lock(&caching_ctl->mutex);
                /* need to make sure the commit_root doesn't disappear */
                down_read(&fs_info->extent_commit_sem);

So, if I'm reading the code correct, there can be a fair amount of
concurrency here: there may be multiple 'caching kthreads' per filesystem
active, while there's one fs_info->extent_commit_sem per filesystem
AFAICS.

So, what happens if there are a lot of CPUs all busy holding the
->extent_commit_sem rwsem read-locked and a writer arrives? They'd all
rush to try to release the fs_info->extent_commit_sem, and they'd block in
the down_read() because there's a writer waiting.

So there's a guarantee of forward progress. This should answer akpm's
concern I think.

Thanks,

Acked-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Josef Bacik <jbacik@fusionio.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/extent-tree.c

index 1c82bead2c088596269b823ae6b7ae0f84c29b7b..3d19dcc553aae23cd1fa47b59de493e896504f9e 100644 (file)
@@ -442,7 +442,8 @@ next:
                        if (ret)
                                break;
 
-                       if (need_resched()) {
+                       if (need_resched() ||
+                           rwsem_is_contended(&fs_info->extent_commit_sem)) {
                                caching_ctl->progress = last;
                                btrfs_release_path(path);
                                up_read(&fs_info->extent_commit_sem);