Re: [NFS] 2.6.23-rc1-mm2

[Oleg Nesterov <oleg@tv-sign.ru>]

On 08/07, Trond Myklebust wrote:
>
> On Wed, 2007-08-08 at 02:20 +0400, Oleg Nesterov wrote:
> 
> > But. nfs4_renew_state() checks list_empty(&clp->cl_superblocks) under
> > clp->cl_sem? So, if it is possible that clp->cl_renewd was scheduled
> > at the time when nfs4_kill_renewd(), we can deadlock, no? Because
> > nfs4_renew_state() needs clp->cl_sem to complete, but nfs4_kill_renewd()
> > holds this sem, and waits for nfs4_renew_state() completion.
> 
> They both take read locks,

Aaaaaaaaaah. Please ignore me, thanks!

> which means that they can take them
> simultaneously. AFAICS, the deadlock can only occur if something manages
> to insert a request for a write lock after nfs4_kill_renewd() takes its
> read lock, but before nfs4_renew_state() takes its read lock:
> 
> 1) nfs4_kill_renewd()		2) nfs4_renew_state()		3) somebody else
> -------------------             ------------------		-------------
> read lock
> wait on (2) to complete
> 								write lock <waits on (1)>
> 				read lock <waits on (3),
> 					because rw_semaphores
> 					don't allow a read lock
> 					request to jump a write
> 					lock request>
> 
> however as I explained earlier, the only process that can take a write
> lock is the reclaimer daemon, but we _know_ that cannot be running (for
> one thing, the reference count on nfs_client is zero, for the other,
> there are no superblocks).

Oleg.