Re: [PATCH v7 7/9] iommu/vt-d: Allow qi_submit_sync() to return the QI faults

[Yi Liu]

On 2023/12/26 14:15, Yi Liu wrote:
> On 2023/12/26 12:13, Tian, Kevin wrote:
> > > From: Liu, Yi L <yi.l.liu@xxxxxxxxx>
> > > Sent: Tuesday, December 26, 2023 12:03 PM
> > >
> > > On 2023/12/22 12:23, Tian, Kevin wrote:
> > > > > From: Liu, Yi L <yi.l.liu@xxxxxxxxx>
> > > > > Sent: Thursday, December 21, 2023 11:40 PM
> > > > >
> > > > > +    fault &= DMA_FSTS_IQE | DMA_FSTS_ITE | DMA_FSTS_ICE;
> > > > > +    if (fault) {
> > > > > +        if (fsts)
> > > > > +            *fsts |= fault;
> > > >
> > > > do we expect the fault to be accumulated? otherwise it's clearer to
> > > > just do direct assignment instead of asking for the caller to clear
> > > > the variable before invocation.
> > >
> > > not quite get. do you mean the fault should not be cleared in the caller
> > > side?
> >
> > I meant:
> >
> >     if (fsts)
> >         *fsts = fault;
> >
> > unless there is a reason to *OR* the original value.
>
> I guess no such a reason. :) let me modify it.

hmmm, replied too soon. The qi_check_fault() would be called multiple
times in one invalidation circle as qi_submit_sync() needs to see if any
fault happened before the hw writes back QI_DONE to the wait descriptor.
There can be ICE which may eventually result in ITE. So caller of 
qi_check_fault()
would continue to wait for QI_DONE. So qi_check_fault() returns 0 to let
qi_submit_sync() go on though ICE detected. If we use '*fsts = fault;',
then ICE would be missed since the input fsts pointer is the same in
one qi_submit_sync() call.

--
Regards,
Yi Liu