Re: [PATCH] tpm: of: avoid __va() translation for event log address

[Jarkko Sakkinen]

On Mon, Sep 28, 2020 at 08:20:18AM +0200, Ard Biesheuvel wrote:
> On Mon, 28 Sep 2020 at 07:56, Christophe Leroy
> <christophe.leroy@xxxxxxxxxx> wrote:
> >
> >
> >
> > Le 28/09/2020 à 01:44, Jarkko Sakkinen a écrit :
> > > On Fri, Sep 25, 2020 at 09:00:18AM -0300, Jason Gunthorpe wrote:
> > >> On Fri, Sep 25, 2020 at 01:29:20PM +0300, Jarkko Sakkinen wrote:
> > >>> On Fri, Sep 25, 2020 at 09:00:56AM +0200, Ard Biesheuvel wrote:
> > >>>> On Fri, 25 Sep 2020 at 07:56, Jarkko Sakkinen
> > >>>> <jarkko.sakkinen@xxxxxxxxxxxxxxx> wrote:
> > >>>>>
> > >>>>> On Tue, Sep 22, 2020 at 11:41:28AM +0200, Ard Biesheuvel wrote:
> > >>>>>> The TPM event log is provided to the OS by the firmware, by loading
> > >>>>>> it into an area in memory and passing the physical address via a node
> > >>>>>> in the device tree.
> > >>>>>>
> > >>>>>> Currently, we use __va() to access the memory via the kernel's linear
> > >>>>>> map: however, it is not guaranteed that the linear map covers this
> > >>>>>> particular address, as we may be running under HIGHMEM on a 32-bit
> > >>>>>> architecture, or running firmware that uses a memory type for the
> > >>>>>> event log that is omitted from the linear map (such as EfiReserved).
> > >>>>>
> > >>>>> Makes perfect sense to the level that I wonder if this should have a
> > >>>>> fixes tag and/or needs to be backported to the stable kernels?
> > >>>>>
> > >>>>
> > >>>> AIUI, the code was written specifically for ppc64, which is a
> > >>>> non-highmem, non-EFI architecture. However, when we start reusing this
> > >>>> driver for ARM, this issue could pop up.
> > >>>>
> > >>>> The code itself has been refactored a couple of times, so I think it
> > >>>> will require different versions of the patch for different generations
> > >>>> of stable kernels.
> > >>>>
> > >>>> So perhaps just add Cc: <stable@xxxxxxxxxxxxxxx>, and wait and see how
> > >>>> far back it applies cleanly?
> > >>>
> > >>> Yeah, I think I'll cc it with some note before the diffstat.
> > >>>
> > >>> I'm thinking to cap it to only 5.x kernels (at least first) unless it is
> > >>> dead easy to backport below that.
> > >>
> > >> I have this vauge recollection of pointing at this before and being
> > >> told that it had to be __va for some PPC reason?
> > >>
> > >> Do check with the PPC people first, I see none on the CC list.
> > >>
> > >> Jason
> > >
> > > Thanks, added arch/powerpc maintainers.
> > >
> >
> > As far as I can see, memremap() won't work on PPC32 at least:
> >
> > IIUC, memremap() calls arch_memremap_wb()
> > arch_memremap_wb() calls ioremap_cache()
> > In case of failure, then ioremap_wt() and ioremap_wc() are tried.
> >
> > All ioremap calls end up in __ioremap_caller() which will return NULL in case you try to ioremap RAM.
> >
> > So the statement "So instead, use memremap(), which will reuse the linear mapping if
> > it is valid, or create another mapping otherwise." seems to be wrong, at least for PPC32.
> >
> > Even for PPC64 which doesn't seem to have the RAM check, I can't see that it will "reuse the linear
> > mapping".
> >
> 
> It is there, please look again. Before any of the above happens,
> memremap() will call try_ram_remap() for regions that are covered by a
> IORESOURCE_SYSTEM_RAM, and map it using __va() if its PFN is valid and
> it is not highmem.
> 
> So as far as I can tell, this change has no effect on PPC at all
> unless its RAM is not described as IORESOURCE_SYSTEM_RAM.

Any chance for someone to test this on PPC32?

/Jarkko