Re: [Qemu-devel] [PATCH V3 10/11] vcpu: introduce lockmap

[Igor Mitsyanko]

On 09/19/2012 04:12 PM, Avi Kivity wrote:
> On 09/19/2012 02:46 PM, Edgar E. Iglesias wrote:
> > On Wed, Sep 19, 2012 at 10:55:30AM +0300, Avi Kivity wrote:
> > > On 09/19/2012 07:40 AM, Peter Crosthwaite wrote:
> > > > On Wed, Sep 19, 2012 at 2:32 PM, Edgar E. Iglesias
> > > > <address@hidden> wrote:
> > > > > On Wed, Sep 19, 2012 at 02:25:48PM +1000, Peter Crosthwaite wrote:
> > > > > > Ping for PMM,
> > > > > >
> > > > > > This is the root case of your block on the SDHCI series - this is a
> > > > > > discussion on resolution to bogus infinite looping DMA. For current
> > > > > > participants in this discussion, heres our thread on the same topic
> > > > > > over in SD land:
> > > > > >
> > > > > > http://lists.gnu.org/archive/html/qemu-devel/2012-08/msg01017.html
> > > > > >
> > > > > > With the findings here and acknowledgement that this is a general
> > > > > > problem, we either need to declare this issue of scope for SDHCI, or
> > > > > > work with these guys (in the immediate future) on the DMA infinite
> > > > > > loop problem flagged here. I dont mind if SDHCI+ADMA is the guinea pig
> > > > > > here, but can we get a decisive plan for going forward with this issue
> > > > > > if it is going to continue to block SDHCI.
> > > > > >
> > > > > > Thanks to Igor for identifying the overlap between these discussions.
> > > > > Hi,
> > > > >
> > > > > A couple of dumb questions.
> > > > >
> > > > > What is the reason for the blocker? that possible guest dos is worse
> > > > > than no functionality at all?
> > > > >
> > > > > Can't you put the DMA/IO processing into the background?
> > > > I dont know a way do do asynchronous DMA unless I am missing something
> > > > here.
> > > You could schedule a bottom half and do the accesses from there.  Solves
> > > nothing though.
> > >
> > > > So what happens is we have a device that walks a SG list
> > > > synchronously while holding all the locks and stuff being discussed
> > > > here. If that SG list infinite loops then crash.
> > > Did you mean loop, or recursion into the memory region that initiates DMA?
> > I think we were discussing the loop issue (I hadn't thought of recursion
> > issues before) :)
> >
> > Jan's point of resetability was interesting.
> >
> > Processing a finite amount of dma descriptors and scheduling bh's
> > to continously get back and continue processing should be OK no?
> Yes.

Bottom half idea sounds good.
Also, for this particular device, rather then limiting amount of 
processed descriptor entries in bottom half, it seems reasonable to 
interrupt bh handler if we encounter a LINK descriptor since only 
recursive LINKs can generate infinite loops. In that case, a very long 
descriptor table without a single LINK entry could potentially hung QEMU 
for a long time, but that time will be finite anyway. Long term, this 
problem will disappear when/if someone converts QEMU SD card model to 
use async io.
> > That would avoid the lockups and allow the device to be reset at
> > any time. Or am I missing something?
> Not that I can see.  If real hardware can be looped, so can qemu.  I'm
> only worried about recursion and deadlocks (while real hardware can
> deadlock, we'd prefer to avoid that).

So, I think the idea here is that if real hardware can be locked we 
should lock too, but provide a guest CPU a possibility to abort locked 
operation. For this particular example, SD card controller can deadlock 
on recursive descriptor LINK entries, but CPU can abort ongoing 
transaction at any time by issuing ABORT command. And if guest CPU 
busy-waits in infinite while() loop for a TRANSFER OVER flag to be set, 
then it had it coming.