Re: [PATCH] ARM: 8936/1: decompressor: avoid CP15 barrier instructions in v7 cache setup code

[Ard Biesheuvel]

On Sat, 25 Jan 2020 at 19:06, Russell King - ARM Linux admin
<linux@xxxxxxxxxxxxxxx> wrote:
>
> On Sat, Jan 25, 2020 at 09:39:50AM -0800, Guenter Roeck wrote:
> > Hi,
> >
> > On Fri, Nov 08, 2019 at 01:44:32PM +0100, Ard Biesheuvel wrote:
> > > Commit e17b1af96b2afc38e684aa2f1033387e2ed10029
> > >
> > >   "ARM: 8857/1: efi: enable CP15 DMB instructions before cleaning the cache"
> > >
> > > added some explicit handling of the CP15BEN bit in the SCTLR system
> > > register, to ensure that CP15 barrier instructions are enabled, even
> > > if we enter the decompressor via the EFI stub.
> > >
> > > However, as it turns out, there are other ways in which we may end up
> > > using CP15 barrier instructions without them being enabled. I.e., when
> > > the decompressor startup code skips the cache_on() initially, we end
> > > up calling cache_clean_flush() with the caches and MMU off, in which
> > > case the CP15BEN bit in SCTLR may not be programmed either. And in
> > > fact, cache_on() itself issues CP15 barrier instructions before actually
> > > enabling them by programming the new SCTLR value (and issuing an ISB)
> > >
> > > Since all these routines are specific to v7, let's clean this up by
> > > using the ordinary v7 barrier instructions in the v7 specific cache
> > > handling routines, so that we never rely on the CP15 ones. This also
> > > avoids the issue where a barrier is required between programming SCTLR
> > > and using the CP15 barrier instructions, which would result in two
> > > different kinds of barriers being used in the same function.
> > >
> > > Acked-by: Marc Zyngier <maz@xxxxxxxxxx>
> > > Signed-off-by: Ard Biesheuvel <ardb@xxxxxxxxxx>
> > > Signed-off-by: Russell King <rmk+kernel@xxxxxxxxxxxxxxx>
> >
> > This patch causes all qemu emulations for ARM1176 to fail hard (stall with
> > no console output even with earlycon enabled). This affects witherspoon-bmc,
> > ast2500-evb, romulus-bmc, and swift-bmc. It does not affect emulations
> > for other CPU types, even with the same kernel configuration (such as
> > ast2600-evb).
>
> Hmm, looks like we're going to have to drop 8936/1, 8941/1 and 8942/1
> in that case.
>

8941 was intended as an alternative approach to 8936, as the latter is
flawed, given that the v7 cache maintenance routines are shared with
CPUID capable non-v7 CPUs such as the ARM1176. So it was never the
intention for both to be applied.

It should be sufficient to revert 8936. Apologies for the confusion.