[PATCH v2] nvme-tcp: Check if request has started before processing it

[Ewan D. Milne]

On Wed, 2021-03-31 at 09:11 +0200, Hannes Reinecke wrote:
> On 3/31/21 1:28 AM, Keith Busch wrote:
> > On Tue, Mar 30, 2021 at 10:34:25AM -0700, Sagi Grimberg wrote:
> > > 
> > > > > It is, but in this situation, the controller is sending a
> > > > > second
> > > > > completion that results in a use-after-free, which makes the
> > > > > transport irrelevant. Unless there is some other flow (which
> > > > > is
> > > > > unclear
> > > > > to me) that causes this which is a bug that needs to be fixed
> > > > > rather
> > > > > than hidden with a safeguard.
> > > > > 
> > > > 
> > > > The kernel should not crash regardless of any network traffic
> > > > that is
> > > > sent to the system.  It should not be possible to either
> > > > intentionally
> > > > of mistakenly contruct packets that will deny service in this
> > > > way.
> > > 
> > > This is not specific to nvme-tcp. I can build an rdma or pci
> > > controller
> > > that can trigger the same crash... I saw a similar patch from
> > > Hannes
> > > implemented in the scsi level, and not the individual scsi
> > > transports..
> > 
> > If scsi wants this too, this could be made generic at the blk-mq
> > level.
> > We just need to make something like blk_mq_tag_to_rq(), but return
> > NULL
> > if the request isn't started.
> >  
> > > I would also mention, that a crash is not even the scariest issue
> > > that
> > > we can see here, because if the request happened to be reused we
> > > are
> > > in the silent data corruption realm...
> > 
> > If this does happen, I think we have to come up with some way to
> > mitigate it. We're not utilizing the full 16 bits of the
> > command_id, so
> > maybe we can append something like a generation sequence number
> > that can
> > be checked for validity.
> > 
> 
> ... which will be near impossible.
> We can protect against crashing on invalid frames.
> We can _not_ protect against maliciously crafted packets referencing
> any
> random _existing_ tag; that's what TLS is for.
> 
> What we can do, though, is checking the 'state' field in the tcp
> request, and only allow completions for commands which are in a state
> allowing for completions.
> 
> Let's see if I can whip up a patch.

That would be great.  BTW in the crash dump I am looking at now, it
looks like pdu->command_id was zero in nvme_tcp_recv_data(), and
blk_mq_tag_to_rq() returned a request struct that had not been used.
So I think we do need to check that the tag was actually allocated.

-Ewan

> 
> Cheers,
> 
> Hannes