From mboxrd@z Thu Jan  1 00:00:00 1970
From: swise@opengridcomputing.com (Steve Wise)
Date: Wed, 21 Sep 2016 16:20:37 -0500
Subject: nvmf/rdma host crash during heavy load and keep alive recovery
In-Reply-To: <02db01d2128b$e9244c70$bb6ce550$@opengridcomputing.com>
References: <021401d20a16$ed60d470$c8227d50$@opengridcomputing.com>
 <021501d20a19$327ba5b0$9772f110$@opengridcomputing.com>
 <00ab01d20ab1$ed212ff0$c7638fd0$@opengridcomputing.com>
 <022701d20d31$a9645850$fc2d08f0$@opengridcomputing.com>
 <da2e918b-0f18-e032-272d-368c6ec49c62@grimberg.me>
 <011501d20f5f$b94e6c80$2beb4580$@opengridcomputing.com>
 <012001d20f63$5c8f7490$15ae5db0$@opengridcomputing.com>
 <01d201d20f69$449abce0$cdd036a0$@opengridcomputing.com>
 <020001d20f70$9998fde0$cccaf9a0$@opengridcomputing.com>
 <02c001d20f93$e6a88a60$b3f99f20$@opengridcomputing.com>
 <20160916110412.GC5476@lst.de>
 <8fc2cefe-76b6-b0a3-12af-701833c286f7@grimberg.me>
 <02db01d2128b$e9244c70$bb6ce550$@opengridcomputing.com>
Message-ID: <02c601d2144d$ff453a50$fdcfaef0$@opengridcomputing.com>

> > >
> > > Oh.  Actually we'll probably need to take care of the connect_q just
> > > about anywhere we do anything to the other queues..
> >
> > Why should we?
> >
> > We control the IOs on the connect_q (we only submit connect to it) and
> > we only submit to it if our queue is established.
> >
> > I still don't see how this explains why Steves is seeing bogus
> > queue/hctx mappings...
> 
> I don't think I'm seeing bogus mappings necessarily.  I think my debug
code
> uncovered (to me at least) that connect_q hctx's use the same
> nvme_rdma_queues
> as the ioq hctxs.   And I thought that was not a valid configuration, but
> apparently its normal.  So I still don't know how/why a pending request
gets
> run
> on an nvme_rdma_queue that has blown away its rdma qp and cm_id.  It
> _could_ be
> due to queue/hctx bogus mappings, but I haven't proven it.  I'm not sure
> how to
> prove it (or how to further debug this issue)...

I added debug code to save off the 2 blk_mq_hw_ctx pointers that get
associated with each nvme_rdma_queue.  This allows me to assert that the
hctx passed into nvme_rdma_queue_rq() is not bogus.  And indeed the hctx
passed in during the crash is the correct hctx.  So we know there isn't a
problem with a bogus hctx being used.

The hctx.state has BLK_MQ_S_TAG_ACTIVE set and _not_ BLK_MQ_S_STOPPED.  The
ns->queue->queue_flags has QUEUE_FLAG_STOPPED bit set.  So the blk_mq queue
is active and the nvme queue is STOPPED.  I don't know how it gets in this
state...

Steve.