On Sat, 07 Nov 2009 23:10:13 +0100, Pavel Filipensky <Pavel.Filipensky at sun.com> wrote:
> can I get a code review for: 6897605 Deadlock in nfs4_active_reclaim() > on r_lock if file and its XATTR are hashed in the same r4hashq_t > > Webrev: > > http://cr.opensolaris.org/~pavelf/6897605 > > Background: > > If the system is low on memory, in can decide to reclaim memory from > "rnode4_cache". > This can lead to a deadlock if the rnode for a file uses extended > attributes and both > rnodes (file and its xattr) are hashed in the same r4hashq_t since the > nfsv4 code will ask for the same r_lock twice. > > The fix makes the reclaim less strict but safe - we avoid releasing of > xattr vnode. I'm not much comfortable fix this fix. since this cached xattr dir vnode is accociated with the rnode we're going to deactivate - who's going to release the xattr vnode eventually ? have you considered using VN_RELE_ASYNC(xattr vnode) here instead, which would issue a taskq thread handling the VN_RELE() for the asynchronously ? of course since we're in a low memory situation, a taskq may also have difficulties being created, but at least it will eventually and we have this all consistant. --- frankB
