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

[Hannes Reinecke]

On 2/15/21 10:23 PM, Sagi Grimberg wrote:
> > > > > > blk_mq_tag_to_rq() will always return a request if the command_id is
> > > > > > in the valid range. Check if the request has been started. If we
> > > > > > blindly process the request we might double complete a request which
> > > > > > can be fatal.
> > > > >
> > > > > How did you get to this one? did the controller send a completion for
> > > > > a completed/bogus request?
> > > >
> > > > If that is the case, then that must mean it's possible the driver could
> > > > have started the command id just before the bogus completion check. 
> > > > Data
> > > > iorruption, right?
> > >
> > > Yes, which is why I don't think this check is very useful..
> >
> > I actually view that as a valid protection against spoofed frames.
> > Without it it's easy to crash the machine by injecting fake 
> > completions with random command ids.
>
> And this doesn't help because the command can have been easily reused
> and started... What is this protecting against? Note that none of the
> other transports checks that, why should tcp?

Because it's particularly easy to spoof packets on tcp.
All other nvme-of transports are layered on top of other transports 
which do some sanity checks already, so it becomes really hard to inject 
invalid NVMe-oF frames for those.
NVMe-TCP has none of these protections, making it really easy to inject 
faulty frames (or, heaven forbid, running a packet fuzzer).

And crashing the machine on invalid frames is always a bad idea; I would 
have expected NVMe-TCP to drop them.

Cheers,

Hannes
--
Dr. Hannes Reinecke                Kernel Storage Architect
hare@xxxxxxx                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer