Re: [Qemu-devel] [PATCH 1/2] virtio-9p: print error message and exit instead of BUG_ON()

[Greg Kurz <groug@kaod.org>]

On Fri, 09 Sep 2016 08:38:13 +0200
Markus Armbruster <armbru@redhat.com> wrote:

> Greg Kurz <groug@kaod.org> writes:
> 
> > On Thu, 8 Sep 2016 18:19:27 +0300
> > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> >  
> >> On Thu, Sep 08, 2016 at 05:04:47PM +0200, Cornelia Huck wrote:  
> >> > On Thu, 8 Sep 2016 18:00:28 +0300
> >> > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> >> >     
> >> > > On Thu, Sep 08, 2016 at 11:12:16AM +0200, Greg Kurz wrote:    
> >> > > > On Thu, 8 Sep 2016 10:59:26 +0200
> >> > > > Cornelia Huck <cornelia.huck@de.ibm.com> wrote:
> >> > > >     
> >> > > > > On Wed, 07 Sep 2016 19:19:24 +0200
> >> > > > > Greg Kurz <groug@kaod.org> wrote:
> >> > > > >     
> >> > > > > > Calling assert() really makes sense when hitting a genuine bug, which calls
> >> > > > > > for a fix in QEMU. However, when something goes wrong because the guest
> >> > > > > > sends a malformed message, it is better to write down a more meaningul
> >> > > > > > error message and exit.
> >> > > > > > 
> >> > > > > > Signed-off-by: Greg Kurz <groug@kaod.org>
> >> > > > > > ---
> >> > > > > >  hw/9pfs/virtio-9p-device.c |   20 ++++++++++++++++++--
> >> > > > > >  1 file changed, 18 insertions(+), 2 deletions(-)      
> >> > > > > 
> >> > > > > While this is an improvement over the current state, I don't think the
> >> > > > > guest should be able to kill qemu just by doing something stupid.
> >> > > > >     
> >> > > > 
> >> > > > Hi Connie,
> >> > > > 
> >> > > > I'm glad you're pointing this out... this was also my impression, but
> >> > > > since there are a bunch of sanity checks in the virtio code that cause
> >> > > > QEMU to exit (even recently added like 1e7aed70144b), I did not dare
> >> > > > stand up :)    
> >> > > 
> >> > > It's true that it's broken in many places but we should just
> >> > > fix them all.
> >> > > 
> >> > > 
> >> > > A separate question is how to log such hardware/guest bugs generally.
> >> > > People already complained about disk filling up because of us printing
> >> > > errors on each such bug.  Maybe print each message only N times, and
> >> > > then set a flag to skip the log until management tells us to restart
> >> > > logging again.    
> >> > 
> >> > I'd expect to get the message just once per device if we set the device
> >> > to broken (unless the guess continuously resets it again...)    
> >> 
> >> Which it can do, so we should limit that anyway.
> >>   
> >> > Do we have
> >> > a generic print/log ratelimit infrastructure in qemu?    
> >> 
> >> There are actually two kinds of errors
> >> host side ones and ones triggered by guests.
> >> 
> >> We should distinguish between them API-wise, then
> >> we will be able to limit the logging of those
> >> that guest can trigger.
> >>   
> >
> > FWIW it makes sense to use error_report() if QEMU exits.  
> 
> exit(STATUS) with STATUS != 0 without printing a message is always
> wrong.
> 

I fully agree.

> >                                                          If it continues
> > execution, this means we're expecting the guest or the host to do something
> > to fix the error condition. This requires QEMU to emit an event of some
> > sort, but not necessarily to log an error message in a file. I guess this
> > depends if QEMU is run by some tooling, or by a human.  
> 
> error_report() normally goes to stderr.  Tooling or humans can of course
> make it go to a file instead.
> 
> error_report() is indeed a sub-par way to send an "attention" signal to
> the host, because recognizing such a signal reliably is unnecessary hard
> for management applications.  QMP events are much easier.
> 

My wording was poor but yes, that was my point. :)

> Both are useless when the signal needs to go to the guest.  Signalling
> the guest is a device model job.
> 

I also agree with that. In the case of virtio, this is explained in section
2.1.2 of the spec.

> error_report() without exit() has its uses.  Error conditions in need of
> fixing aren't the only reason to call error_report().  But when you add
> a call, ask yourself whether management application or guest would like
> to respond to it.

In the case of the present patch, we currently have BUG_ON() which generates
a cryptic and unusable message.

It turns out that the first one (elem->out_num == 0 || elem->in_num == 0) is
correct since it is now [1] impossible to hit this according to the code (see
virtqueue_pop() and virtqueue_map_desc()).

The second one (len != sizeof out) though matches a potential guest originated
error. If I do as suggested by Connie, then the error_report() isn't needed
anymore.

Cheers.

--
Greg

[1] sending an empty buffer was sufficient before commit 1e7aed70144b4 as said
    in my previous answer