[Skiboot] [PATCH skiboot] npu2: Clear fence on all bricks
alistair at popple.id.au
Tue Dec 10 15:29:38 AEDT 2019
On Tuesday, 10 December 2019 9:21:18 AM AEDT Reza Arbab wrote:
> On 02/12/2019 13:05, Alexey Kardashevskiy wrote:
> > On 30/11/2019 03:47, Reza Arbab wrote:
> >> On Fri, Nov 22, 2019 at 11:04:22AM +1100, Alexey Kardashevskiy wrote:
> >> Each GPU has a 640kb protected region which will result in a
> >> "unsupported request" (UR) response. The root bug is that the driver
> >> maps and accidentally accesses that area.
> > Oh. Is this address range described anywhere? We could disable mapping
> > these as a precaution measure.
> It's only been communicated to us ad hoc during bug investigation, as
> far as I know. I'm going to try requesting documentation of all
> cpu-access-limited regions so we have something to refer to.
I'm not a fan of this kind of whack-a-mole at all. If the driver is requesting
a mapping of something it shouldn't be mapping then it's a bug and the driver
needs to be fixed.
Fixing the recovery paths makes sense, but adding arbitrary validation checks
around the place will simply create more hard to check co-dependencies and
strange bugs when those locations inevitably change and it still won't prevent
bugs causing UR responses or other bad state.
More information about the Skiboot