On Fri, 03 Dec 2021 10:08:49 +0000,
Ard Biesheuvel [off-list ref] wrote:

First, enable the use of the TLS register to hold the 'current' pointer
for all configurations that can support it, including non-SMP ones that
target v6k or later CPUs, and multi-platform SMP ones that also support
v6 based UP systems.

The remaining configurations are all strictly UP, which means we can
switch to a global variable to hold the current pointer. By doing this,
we can enable THREAD_INFO_IN_TASK, which moves thread info off the
stack, protecting it from overflows. It also permits us to enable IRQ
stacks and vmap'ed stacks for UP configurations as well.

Supporting v6 cores without SMP extensions in SMP configurations (e.g.,
omap2plus_defconfig or imx_v6_v7_defconfig) makes this a bit tricky, and
this is a feature we may consider dropping entirely in the future. But
for the time being, we can support this mode as well.

The accesses to the global variable holding 'current' are constructed in
a way that ensures that no literal pool accesses (and associated D-cache
misses) are needed unless the access is from a module and module PLTs
are enabled. This means that accessing 'current' is just as costly as
before, as it used to require some arithmetic involving the stack
pointer and a load from the thread_info::task field.

However, accessing thread_info itself now also involves a load, although
it should be noted that all thread_info and current accesses now go via
the same variable, which is therefore expected to be hot in the caches
at all times.

FWIW, I've given this a go on a Netwinder (of the Footbridge family),
and nothing caught fire. I'm not using modules though.

Tested-by: Marc Zyngier <maz@kernel.org>

	M.

-- 
Without deviation from the norm, progress is not possible.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel