CONFIG_THUMB2_KERNEL=y boot failure after Spectre BHB fixes

[Ard Biesheuvel]

On Wed, 30 Mar 2022 at 18:12, Russell King (Oracle)
<linux at armlinux.org.uk> wrote:
>
> On Tue, Mar 22, 2022 at 06:49:17PM +0100, Christian Eggers wrote:
> > Hi Nathan, hi Russel,
> >
> > I stumbled today over the same problem (no output on serial console
> > with v5.15.28-rt36). During `git bisect`, I had also some commits
> > where a few lines of output were visible.
> >
> > At commit 8d9d651ff227 ("ARM: use LOADADDR() to get load address of
> > sections"), the system boots up to here:
> >
> > start_kernel()
> > +--setup_arch()
> >    +--paging_init()
> >       +--devicemaps_init()
> >          +--eary_trap_init(vectors_base = 0xC7FFE000)
> >             +--copy_from_lma(vectors_base = 0xC7FFE000, __vectors_start=0x0, __vectors_end=0x20)
> >                +--__memcpy()
> >
> >                copy_template.S:113
> >                ldr8w   r1, r3, r4, r5, r6, r7, r8, ip, lr, abort=20f
> >                r1 = 0
> >
> >
> > With the final v5.15.28-rt36 I found out that the system boots fine
> > after disabling CONFIG_HARDEN_BRANCH_HISTORY.
> >
> > Is there anything else I could analyze? My SoC system is a NXP i.MX6LL
> > with a ARMv7 core. I have access to a JTAG debugger.
>
> I think this is already fixed in mainline. Commit:
>
> 6c7cb60bff7a ("ARM: fix Thumb2 regression")
>

It's still broken - I sent a couple of patches on Monday, among which
one to fix the boot issue with loop8 on Thumb2. The problem is 'b . +
4', which produces a narrow encoding, and so it skips the subsequent
subs instruction and loops forever.