Re: [RFC PATCH] arm64: remove CONFIG_DEBUG_ALIGN_RODATA feature

[Ard Biesheuvel <ardb@kernel.org>]

On Mon, 30 Mar 2020 at 16:28, Will Deacon <will@kernel.org> wrote:
>
> On Mon, Mar 30, 2020 at 04:22:24PM +0200, Ard Biesheuvel wrote:
> > On Mon, 30 Mar 2020 at 16:04, Will Deacon <will@kernel.org> wrote:
> > >
> > > On Mon, Mar 30, 2020 at 03:53:04PM +0200, Ard Biesheuvel wrote:
> > > > On Mon, 30 Mar 2020 at 15:51, Will Deacon <will@kernel.org> wrote:
> > > > >
> > > > > On Sun, Mar 29, 2020 at 04:12:58PM +0200, Ard Biesheuvel wrote:
> > > > > > When CONFIG_DEBUG_ALIGN_RODATA is enabled, kernel segments mapped with
> > > > > > different permissions (r-x for .text, r-- for .rodata, rw- for .data,
> > > > > > etc) are rounded up to 2 MiB so they can be mapped more efficiently.
> > > > > > In particular, it permits the segments to be mapped using level 2
> > > > > > block entries when using 4k pages, which is expected to result in less
> > > > > > TLB pressure.
> > > > > >
> > > > > > However, the mappings for the bulk of the kernel will use level 2
> > > > > > entries anyway, and the misaligned fringes are organized such that they
> > > > > > can take advantage of the contiguous bit, and use far fewer level 3
> > > > > > entries than would be needed otherwise.
> > > > > >
> > > > > > This makes the value of this feature dubious at best, and since it is not
> > > > > > enabled in defconfig or in the distro configs, it does not appear to be
> > > > > > in wide use either. So let's just remove it.
> > > > > >
> > > > > > Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
> > > > > > ---
> > > > > >  arch/arm64/Kconfig.debug                  | 13 -------------
> > > > > >  arch/arm64/include/asm/memory.h           | 12 +-----------
> > > > > >  drivers/firmware/efi/libstub/arm64-stub.c |  8 +++-----
> > > > > >  3 files changed, 4 insertions(+), 29 deletions(-)
> > > > >
> > > > > Acked-by: Will Deacon <will@kernel.org>
> > > > >
> > > > > But I would really like to go a step further and rip out the block mapping
> > > > > support altogether so that we can fix non-coherent DMA aliases:
> > > > >
> > > > > https://lore.kernel.org/lkml/20200224194446.690816-1-hch@lst.de
> > > > >
> > > >
> > > > I'm not sure I follow - is this about mapping parts of the static
> > > > kernel Image for non-coherent DMA?
> > >
> > > Sorry, it's not directly related to your patch, just that if we're removing
> > > options relating to kernel mappings then I'd be quite keen on effectively
> > > forcing page-granularity on the linear map, as is currently done by default
> > > thanks to RODATA_FULL_DEFAULT_ENABLED, so that we can nobble cacheable
> > > aliases for non-coherent streaming DMA mappings by hooking into Christoph's
> > > series above.
> > >
> >
> > Right. I don't remember seeing any complaints about
> > RODATA_FULL_DEFAULT_ENABLED, but maybe it's too early for that.
> >
> > > This series just reminded me of it because it's another
> > > "off-by-default-behaviour-for-block-mappings-probably-because-of-performance-
> > > but-never-actually-measured" type of thing which really just gets in the
> > > way.
> > >
> >
> > Well, even though I agree that the lack of actual numbers is a bit
> > disturbing here, I'd hate to penalize all systems even more than they
> > already are (due to ARCH_KMALLOC_MINALIGN == ARCH_DMA_MINALIGN) by
> > adding another workaround that is only needed on devices that have
> > non-coherent masters.
>
> Fair enough, but I'd still like to see some numbers. If they're compelling,
> then we could explore something like CONFIG_OF_DMA_DEFAULT_COHERENT, but
> that doesn't really help the kconfig maze :(
>

Could we make this a runtime thing? E.g., remap the entire linear
region down to pages under stop_machine() the first time we probe a
device that uses non-coherent DMA?

(/me ducks)