autofs4: avoid taking fs_lock during rcu-walk
authorNeilBrown <neilb@suse.de>
Mon, 13 Oct 2014 22:52:20 +0000 (15:52 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 14 Oct 2014 00:18:16 +0000 (02:18 +0200)
commit4d885f90e3784df77945101d8ec32dc798a49862
treea034e761a43d14bb24f27e98a0cd3f568bbf7732
parent6ece08e6187a62dd5bdf7c49539980571922c551
autofs4: avoid taking fs_lock during rcu-walk

->fs_lock protects AUTOFS_INF_EXPIRING.  We need to be sure that once
the flag is set, no new references beneath the dentry are taken.  So
rcu-walk currently needs to take fs_lock before checking the flag.  This
hurts performance.

Change the expiry to a two-stage process.  First set AUTOFS_INF_NO_RCU
which forces any path walk into ref-walk mode, then drop the lock and
call synchronize_rcu().  Once that returns we can be sure no rcu-walk is
active beneath the dentry and we can check reference counts again.

Now during an RCU-walk we can test AUTOFS_INF_EXPIRING without taking
the lock as along as we test AUTOFS_INF_NO_RCU too.  If either are set,
we must abort the RCU-walk If neither are set, we know that refcounts
will be tested again after we finish the RCU-walk so we are safe to
continue.

->fs_lock is still taken in d_manage() to check for a non-trap
directory.  That will be resolved in the next patch.

Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Ian Kent <raven@themaw.net>
Tested-by: Ian Kent <raven@themaw.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/autofs4/autofs_i.h
fs/autofs4/expire.c