Re: [v5.4 stable] arm: stm32: Regression observed on "no-map" reserved memory region

[Ard Biesheuvel]

On Tue, 20 Apr 2021 at 17:54, Rob Herring <robh+dt@xxxxxxxxxx> wrote:
>
> On Tue, Apr 20, 2021 at 10:12 AM Alexandre TORGUE
> <alexandre.torgue@xxxxxxxxxxx> wrote:
> >
> >
> >
> > On 4/20/21 4:45 PM, Rob Herring wrote:
> > > On Tue, Apr 20, 2021 at 9:03 AM Alexandre TORGUE
> > > <alexandre.torgue@xxxxxxxxxxx> wrote:
> > >>
> > >> Hi,
> > >
> > > Greg or Sasha won't know what to do with this. Not sure who follows
> > > the stable list either. Quentin sent the patch, but is not the author.
> > > Given the patch in question is about consistency between EFI memory
> > > map boot and DT memory map boot, copying EFI knowledgeable folks would
> > > help (Ard B for starters).
> >
> > Ok thanks for the tips. I add Ard in the loop.
>
> Sigh. If it was only Ard I was suggesting I would have done that
> myself. Now everyone on the patch in question and relevant lists are
> Cc'ed.
>

Thanks for the cc.

> >
> > Ard, let me know if other people have to be directly added or if I have
> > to resend to another mailing list.
> >
> > thanks
> > alex
> >
> > >
> > >>
> > >> Since v5.4.102 I observe a regression on stm32mp1 platform: "no-map"
> > >> reserved-memory regions are no more "reserved" and make part of the
> > >> kernel System RAM. This causes allocation failure for devices which try
> > >> to take a reserved-memory region.
> > >>
> > >> It has been introduced by the following path:
> > >>
> > >> "fdt: Properly handle "no-map" field in the memory region
> > >> [ Upstream commit 86588296acbfb1591e92ba60221e95677ecadb43 ]"
> > >> which replace memblock_remove by memblock_mark_nomap in no-map case.
> > >>

Why was this backported? It doesn't look like a bugfix to me.

> > >> Reverting this patch it's fine.
> > >>
> > >> I add part of my DT (something is maybe wrong inside):
> > >>
> > >> memory@c0000000 {
> > >>          reg = <0xc0000000 0x20000000>;
> > >> };
> > >>
> > >> reserved-memory {
> > >>          #address-cells = <1>;
> > >>          #size-cells = <1>;
> > >>          ranges;
> > >>
> > >>          gpu_reserved: gpu@d4000000 {
> > >>                  reg = <0xd4000000 0x4000000>;
> > >>                  no-map;
> > >>          };
> > >> };
> > >>
> > >> Sorry if this issue has already been raised and discussed.
> > >>

Could you explain why it fails? The region is clearly part of system
memory, and tagged as no-map, so the patch in itself is not
unreasonable. However, we obviously have code that relies on how the
region is represented in /proc/iomem, so it would be helpful to get
some insight into why this is the case.

In any case, the mere fact that this causes a regression should be
sufficient justification to revert/withdraw it from v5.4, as I don't
see a reason why it was merged there in the first place. (It has no
fixes tag or cc:stable)