Re: Re: [PATCH] ARM: OMAP2+: Add support for thumb mode on DT bootedN900

[Sebastian Reichel]

Hi,

On Wed, Feb 05, 2014 at 10:30:52PM +0200, Ivaylo Dimitrov wrote:
> In theory having that workaround disabled might be a security
> problem, but honestly, knowing its nature I don't think it is easily
> exploitable, if at all. The final result when bitten by it is a
> SIGILL, but in userspace, not in the kernel(assuming the kernel is
> ARM), and userspace runs in totally different mode (nonsecure,
> nonprivileged) compared to the kernel(nonsecure, privileged) and
> IIRC every mode has its own set of stack, registers etc. BTW I don't
> think the kernel itself can be thumb2-compiled for cores with that
> errata, but I might wrong. Also, as Pali noted, the problem appears
> if and only if there is an userspace binary containing thumb2 code.
> If all of the userspace is pure ARM, there is no problem.

Ok. I assumed, that a single thumb2-compiled binary can potentially
trigger problems for the whole system. Basically because the ERRATA
helptext in KConfig is:

[...], Cortex-A8 does not recover from the stale interworking branch
prediction.

> And as the errata workaround has its drawbacks (BTB is cleared on
> every context switch which affects performance), one might want to
> not have it enabled.

I guess the workaround would be persistently enabled if there were
no drawbacks.

> Maybe that warning should be spit only if CONFIG_THUMB2_KERNEL (or
> whatever the option was) is enabled. Though if that option is
> enabled I'd rather #error during compile time if errata workaround
> is not enabled, instead of printing a warning while booting a
> system that will crash in a matter of seconds.

THUMB2 userland code is supported by the kernel by default, so that
does not work. (Enabling CONFIG_THUMB2_KERNEL will result in a
thumb2-compiled kernel.)

-- Sebastian

Attachment: signature.asc
Description: Digital signature