linux-nfs.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Trond Myklebust <trondmy@gmail.com>
To: NeilBrown <neilb@suse.de>,
	"bfields@fieldses.org" <bfields@fieldses.org>,
	"anna.schumaker@netapp.com" <anna.schumaker@netapp.com>
Cc: "linux-nfs@vger.kernel.org" <linux-nfs@vger.kernel.org>
Subject: Re: [PATCH] SUNRPC: backchannel RPC request must reference XPRT
Date: Tue, 15 Oct 2019 23:04:53 -0400	[thread overview]
Message-ID: <8ac0b07bd523e4a040b4dc5e53639e1ea3b68db6.camel@hammerspace.com> (raw)
In-Reply-To: <87a7a1r3t2.fsf@notabene.neil.brown.name>

On Wed, 2019-10-16 at 10:23 +1100, NeilBrown wrote:
> On Tue, Oct 15 2019, Trond Myklebust wrote:
> 
> > Hi Neil,
> > 
> > On Tue, 2019-10-15 at 10:36 +1100, NeilBrown wrote:
> > > The backchannel RPC requests - that are queued waiting
> > > for the reply to be sent by the "NFSv4 callback" thread -
> > > have a pointer to the xprt, but it is not reference counted.
> > > It is possible for the xprt to be freed while there are
> > > still queued requests.
> > > 
> > > I think this has been a problem since
> > > Commit fb7a0b9addbd ("nfs41: New backchannel helper routines")
> > > when the code was introduced, but I suspect it became more of
> > > a problem after
> > > Commit 80b14d5e61ca ("SUNRPC: Add a structure to track multiple
> > > transports")
> > > (or there abouts).
> > > Before this second patch, the nfs client would hold a reference
> > > to
> > > the xprt to keep it alive.  After multipath was introduced,
> > > a client holds a reference to a swtich, and the switch can have
> > > multiple
> > > xprts which can be added and removed.
> > > 
> > > I'm not sure of all the causal issues, but this patch has
> > > fixed a customer problem were an NFSv4.1 client would run out
> > > of memory with tens of thousands of backchannel rpc requests
> > > queued for an xprt that had been freed.  This was a 64K-page
> > > machine so each rpc_rqst consumed more than 128K of memory.
> > > 
> > > Fixes: 80b14d5e61ca ("SUNRPC: Add a structure to track multiple
> > > transports")
> > > cc: stable@vger.kernel.org (v4.6)
> > > Signed-off-by: NeilBrown <neilb@suse.de>
> > > ---
> > >  net/sunrpc/backchannel_rqst.c | 3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/net/sunrpc/backchannel_rqst.c
> > > b/net/sunrpc/backchannel_rqst.c
> > > index 339e8c077c2d..c95ca39688b6 100644
> > > --- a/net/sunrpc/backchannel_rqst.c
> > > +++ b/net/sunrpc/backchannel_rqst.c
> > > @@ -61,6 +61,7 @@ static void xprt_free_allocation(struct
> > > rpc_rqst
> > > *req)
> > >  	free_page((unsigned long)xbufp->head[0].iov_base);
> > >  	xbufp = &req->rq_snd_buf;
> > >  	free_page((unsigned long)xbufp->head[0].iov_base);
> > > +	xprt_put(req->rq_xprt);
> > >  	kfree(req);
> > >  }
> > 
> > Would it perhaps make better sense to move the xprt_get() to
> > xprt_lookup_bc_request() and to release it in xprt_free_bc_rqst()? 
> 
> ... maybe ...
> 
> > Otherwise as far as I can tell, we will have freed slots on the
> > xprt-
> > > bc_pa_list that hold a reference to the transport itself, meaning
> > > that
> > the latter never gets released.
> 
> Apart from cleanup-on-error paths, xprt_free_allocation() is called:
>  - in xprt_destroy_bc() - at most 'max_reqs' times.
>  - in xprt_free_bc_rqst(), if the bc_alloc_count >= bc_alloc_max
> 
> So when xprt_destroy_bc() is called (from nfs4_destroy_session, via
> xprt_destroy_backchannel()), bc_alloc_max() is reduced, and possibly
> the requests are freed and bc_alloc_count is reduced accordingly.
> If the requests were busy, they won't have been freed then, but will
> then be freed when xprt_free_bc_reqst is called - because
> bc_alloc_max
> has been reduced.
> 
> Once nfs4_destroy_session() has been called on all session, and
> xprt_free_bc_rqst() has been called on all active requests, I think
> they should be no requests left to be holding a reference to
> the xprt.
> 
> And if there were requests left, and if we change the refcount code
> (like you suggest) so that they weren't holding a reference, then
> they
> would never be freed. - freeing an xprt doesn't clean out the
> bc_pa_list.
> 
> Though ... the connection from a session to an xprt isn't permanent
> (I
> guess, based on the rcu_read_lock in nfs4_destroy_session... which
> itself seems a bit odd as it doesn't inc a refcount while holding the
> lock).
> 
> So maybe the counts can get out of balance.
> 
> Conclusion: I'm not sure the ref counts are entirely correct, but I
>    cannot see a benefit from moving the xprt_get/put requests like
> you
>    suggest.
> 

I don't buy that conclusion.

Nothing stops me from changing the value of NFS41_BC_MIN_CALLBACKS to
zero. Why should that affect the existence or not of the transport by
changing the number of references held? That parameter is supposed to
determine the number of pre-allocated requests and nothing else.

I do agree with your assessment that destroying the xprt does not
currrently destroy the contents of xprt->bc_pa_list if they are non-
zero, but that would be a bug, and an easy one to fix.


-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com




  reply	other threads:[~2019-10-16  3:05 UTC|newest]

Thread overview: 8+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-10-09  5:15 NFSv4.1 backchannel xprt problems NeilBrown
2019-10-11 16:56 ` J. Bruce Fields
2019-10-14 23:36   ` [PATCH] SUNRPC: backchannel RPC request must reference XPRT NeilBrown
2019-10-15 21:16     ` Trond Myklebust
2019-10-15 21:47       ` Chuck Lever
2019-10-15 23:23       ` NeilBrown
2019-10-16  3:04         ` Trond Myklebust [this message]
2019-10-16  4:38           ` NeilBrown

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=8ac0b07bd523e4a040b4dc5e53639e1ea3b68db6.camel@hammerspace.com \
    --to=trondmy@gmail.com \
    --cc=anna.schumaker@netapp.com \
    --cc=bfields@fieldses.org \
    --cc=linux-nfs@vger.kernel.org \
    --cc=neilb@suse.de \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).