[PATCH v5 0/7] arm64: efi: leave MMU and caches on at boot

[Ard Biesheuvel]

On Tue, 15 Nov 2022 at 12:31, Will Deacon <will at kernel.org> wrote:
>
> On Tue, Nov 15, 2022 at 12:21:55PM +0100, Ard Biesheuvel wrote:
> > On Tue, 15 Nov 2022 at 12:17, Will Deacon <will at kernel.org> wrote:
> > >
> > > On Fri, Nov 11, 2022 at 05:36:19PM +0000, Mark Rutland wrote:
> > > > On Tue, Nov 08, 2022 at 07:21:57PM +0100, Ard Biesheuvel wrote:
> > > > > The purpose of this series is to remove any explicit cache maintenance
> > > > > for coherency during early boot that becomes unnecessary if we simply
> > > > > retain the cacheable 1:1 mapping of all of system RAM provided by EFI,
> > > > > and use it to populate the ID map page tables. After setting up this
> > > > > preliminary ID map, we disable the MMU, drop to EL1, reprogram the MAIR,
> > > > > TCR and SCTLR registers as before, and proceed as usual, avoiding the
> > > > > need for any manipulations of memory while the MMU and caches are off.
> > > > >
> > > > > The only properties of the firmware provided 1:1 map we rely on is that
> > > > > it does not require any explicit cache maintenance for coherency, and
> > > > > that it covers the entire memory footprint of the image, including the
> > > > > BSS and padding at the end - all else is under control of the kernel
> > > > > itself, as before.
> > > >
> > > > As a high-level thing, I'm still very much not keen on entering the kernel with
> > > > the MMU on. Given that we have to support booting with the MMU off for !EFI
> > > > boot (including kexec when EFI is in use), I think this makes it harder to
> > > > reason about the boot code overall (e.g. due to the conditional maintenance
> > > > added to head.S), and adds more scope for error, even if it simplifies the EFI
> > > > stub itself.
> > >
> > > As discussed offline, two things that would help the current series are:
> > >
> > >   (1) Some performance numbers comparing MMU off vs MMU on boot
> > >

Finally got around to measuring this - I lost access to my TX2 machine
for a couple of days during the past week,

With the patch below applied to mainline, I measure ~6 ms spent
cleaning the entire image to the PoC (which is the bulk of it) and
subsequently populating the initial ID map and activating it.

This drops to about 0.6 ms with my changes applied. This is unlikely
to ever matter in practice, perhaps, but I will note that booting a VM
in EFI mode using Tianocore/EDK2 from the point where KVM clears the
counter to the point where we start user space can be done (on the
same machine) in 500-700 ms so it is not entirely insignificant
either.

I could try and measure it on bare metal as well, but I suppose that
launch times are even less relevant there so I didn't bother.