lkml.org 
[lkml]   [2015]   [Sep]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC PATCH v2 3/7] arm64: alternative: Apply alternatives early in boot process
On Wed, Sep 16, 2015 at 04:51:12PM +0100, Daniel Thompson wrote:
> On 16/09/15 14:05, Will Deacon wrote:
> > On Mon, Sep 14, 2015 at 02:26:17PM +0100, Daniel Thompson wrote:
> >> Currently alternatives are applied very late in the boot process (and
> >> a long time after we enable scheduling). Some alternative sequences,
> >> such as those that alter the way CPU context is stored, must be applied
> >> much earlier in the boot sequence.
> >>
> >> Introduce apply_alternatives_early() to allow some alternatives to be
> >> applied immediately after we detect the CPU features of the boot CPU.
> >>
> >> Currently apply_alternatives_all() is not optimized and will re-patch
> >> code that has already been updated. This is harmless but could be
> >> removed by adding extra flags to the alternatives store.
> >>
> >> Signed-off-by: Daniel Thompson <daniel.thompson@linaro.org>
> >> ---
> > [snip]
> >> /*
> >> + * This is called very early in the boot process (directly after we run
> >> + * a feature detect on the boot CPU). No need to worry about other CPUs
> >> + * here.
> >> + */
> >> +void apply_alternatives_early(void)
> >> +{
> >> + struct alt_region region = {
> >> + .begin = __alt_instructions,
> >> + .end = __alt_instructions_end,
> >> + };
> >> +
> >> + __apply_alternatives(&region);
> >> +}
> >
> > How do you choose which alternatives are applied early and which are
> > applied later? AFAICT, this just applies everything before we've
> > established the capabilities of the CPUs in the system, which could cause
> > problems for big/little SoCs.
>
> They are applied twice. This relies for correctness on the fact that
> cpufeatures can be set but not unset.
>
> In other words the boot CPU does a feature detect and, as a result, a
> subset of the required alternatives will be applied. However after this
> the other CPUs will boot and the the remaining alternatives applied as
> before.
>
> The current implementation is inefficient (because it will redundantly
> patch the same code twice) but I don't think it is broken.

What about a big/little system where we boot on the big cores and only
they support LSE atomics?

> > Also, why do we need this for the NMI?
>
> I was/am concerned that a context saved before the alternatives are
> applied might be restored afterwards. If that happens the bit that
> indicates what value to put into the PMR would read during the restore
> without having been saved first. Applying early ensures that the context
> save/restore code is updated before it is ever used.

Damn, and stop_machine makes use of local_irq_restore immediately after
the patching has completed, so it's a non-starter. Still, special-casing
this feature via an explicit apply_alternatives call would be better
than moving everything earlier, I think.

We also need to think about how an incoming NMI interacts with
concurrent patching of later features. I suspect we want to set the I
bit, like you do for WFI, unless you can guarantee that no patched
sequences run in NMI context.

Will


\
 
 \ /
  Last update: 2015-09-16 18:41    [W:5.304 / U:0.032 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site