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

[Keith Busch]

On Tue, Mar 02, 2021 at 08:18:40AM +0100, Hannes Reinecke wrote:
> On 3/1/21 9:59 PM, Keith Busch wrote:
> > On Mon, Mar 01, 2021 at 05:53:25PM +0100, Hannes Reinecke wrote:
> >> On 3/1/21 5:05 PM, Keith Busch wrote:
> >>> On Mon, Mar 01, 2021 at 02:55:30PM +0100, Hannes Reinecke wrote:
> >>>> On 3/1/21 2:26 PM, Daniel Wagner wrote:
> >>>>> On Sat, Feb 27, 2021 at 02:19:01AM +0900, Keith Busch wrote:
> >>>>>> Crashing is bad, silent data corruption is worse. Is there truly no
> >>>>>> defense against that? If not, why should anyone rely on this?
> >>>>>
> >>>>> If we receive an response for which we don't have a started request, we
> >>>>> know that something is wrong. Couldn't we in just reset the connection
> >>>>> in this case? We don't have to pretend nothing has happened and
> >>>>> continuing normally. This would avoid a host crash and would not create
> >>>>> (more) data corruption. Or I am just too naive?
> >>>>>
> >>>> This is actually a sensible solution.
> >>>> Please send a patch for that.
> >>>
> >>> Is a bad frame a problem that can be resolved with a reset?
> >>>
> >>> Even if so, the reset doesn't indicate to the user if previous commands
> >>> completed with bad data, so it still seems unreliable.
> >>>
> >> We need to distinguish two cases here.
> >> The one is use receiving a frame with an invalid tag, leading to a crash.
> >> This can be easily resolved by issuing a reset, as clearly the command was
> >> garbage and we need to invoke error handling (which is reset).
> >>
> >> The other case is us receiving a frame with a _duplicate_ tag, ie a tag
> >> which is _currently_ valid. This is a case which will fail _even now_, as we
> >> have simply no way of detecting this.
> >>
> >> So what again do we miss by fixing the first case?
> >> Apart from a system which does _not_ crash?
> > 
> > I'm just saying each case is a symptom of the same problem. The only
> > difference from observing one vs the other is a race with the host's
> > dispatch. And since you're proposing this patch, it sounds like this
> > condition does happen on tcp compared to other transports where we don't
> > observe it. I just thought the implication that data corruption happens
> > is a alarming.
> > 
> Oh yes, it is.
> But sadly TCP inherently suffers from this, as literally anyone can
> spoof frames on the network.
> Other transports like RDMA or FC do not suffer to that extend as
> spoofing frames there is far more elaborate, and not really possible
> without dedicated hardware equipment.
> 
> That's why there is header and data digest; that will protect you
> against accidental frame corruption (as this case clearly is; the
> remainder of the frame is filled with zeroes).
> It would not protect you against deliberate frame corruption; that's why
> there is TPAR 8010 (TLS encryption for NVMe-TCP).
> 
> Be it as it may, none of these methods are in use here, and none of
> these methods can be made mandatory. So we need to deal with the case at
> hand.
> 
> And in my opinion crashing is the _worst_ options of all.
> Tear the connection down, reset the thing, whatever.
> 
> But do not crash.
> Customers tend to have a very dim view on crashing machines, and have a
> very limited capacity for being susceptible to our reasoning in these cases.

I was pushing the data corruption angle because fixing that should
address both cases. If there's really nothing you can do about
corruption, your approach here makes sense, and I defer to Sagi and
Christoph for inclusion.

I still wouldn't trust my data to storage behaving this way, though. :)