On Mon, Jan 27, 2014 at 10:38:45AM -0800, David Daney wrote:
> ....
> On 01/27/2014 07:23 AM, Paul Burton wrote:
> >No current systems implementing MSA include support for vector register
> >partitioning which makes it somewhat difficult to implement support for
> >it in the kernel. Thus for the moment the kernel includes no such
> >support. However if the kernel were to be run on a system which
> >implemented register partitioning then it would not function correctly,
> >mishandling MSA disabled exceptions. Calling panic when run on a system
> >with vector register partitioning implemented ensures that we're not
> >caught out by this later but instead reminded to implement support once
> >such a system is available.
> >
> >Signed-off-by: Paul Burton <paul.burton@xxxxxxxxxx>
> >---
> > arch/mips/kernel/cpu-probe.c | 6 +++++-
> > 1 file changed, 5 insertions(+), 1 deletion(-)
> >
> >diff --git a/arch/mips/kernel/cpu-probe.c b/arch/mips/kernel/cpu-probe.c
> >index 852e085..003ba3c 100644
> >--- a/arch/mips/kernel/cpu-probe.c
> >+++ b/arch/mips/kernel/cpu-probe.c
> >@@ -1193,9 +1193,13 @@ void cpu_probe(void)
> > else
> > c->srsets = 1;
> >
> >- if (cpu_has_msa)
> >+ if (cpu_has_msa) {
> > c->msa_id = cpu_get_msa_id();
> >
> >+ if (c->msa_id & MSA_IR_WRPF)
> >+ panic("Vector register partitioning unimplemented!");
>
> You should probably use a WARN_ON() instead. There is no reason to crash
> the kernel for this condition is there?
>
Well mapping vector registers reuses the MSA disabled exception, so if
the kernel were to continue with my current code & userland were to
execute an MSA instruction I believe it would appear to hang. There
would be an initial MSA disabled exception which would lead the kernel
to enable MSA & return to userland to re-execute the MSA instruction.
Then another MSA disabled exception would occur because we'd need to
map a vector register, but the kernel doesn't understand so would
attempt to enable MSA & return again. Then another MSA disabled
exception, etc etc. So if the kernel were to continue then it would
probably want to disable MSA support entirely to avoid userland
appearing to just hang. Additionally since vector registers & FP
registers are aliased the same would apply to scalar FP (cop1)
instructions too, so we'd also need to disable the FPU. To me that all
seems like a lot of hassle to allow a crippled kernel to run on a system
that doesn't exist yet, so personally I'd rather the kernel just panics
as a safeguard. Then once some target system implements vector register
partitioning support for it can be implemented in the kernel.
Thanks,
Paul
> >+ }
> >+
> > cpu_probe_vmbits(c);
> >
> > #ifdef CONFIG_64BIT
> >
>
>
> To report this email as SPAM, please forward it to spam@xxxxxxxxxxxx
