On Mon, 2022-01-31 at 14:59 +0100, Borislav Petkov wrote:
> On Sat, Jan 29, 2022 at 12:04:19PM +0000, David Woodhouse wrote:
> > I've rebased and pushed to
> > https://git.infradead.org/users/dwmw2/linux.git/shortlog/refs/heads/parallel-5.17
> > 
> > 
> > I'll do some more testing and repost the series during next week. The
> > win is slightly more modest than the original patch sets because it now
> > only parallelises x86/cpu:kick. I'm going to do more careful review and
> > testing before doing the same for x86/cpu:wait-init in a later series.
> > You can see that coming together in the git tree but I'm only going to
> > post up to the 'Serialise topology updates' patch again for now.
> 
> Btw, Mr. Cooper points out a very important aspect and I don't know
> whether you've verified this already or whether this is not affected
> by your series ... yet. In any case it should be checked: microcode
> loading.
> 
> See __reload_late() and all that dance we do to keep SMT siblings do
> nothing at the same time while updating microcode.
> 
> With the current boot order, the APs should all do nothing so they won't
> need that sync for early loading - load_ucode_{ap,bsp} - but I don't
> know if you're changing that order with the parallel startup.
> 
> If you do, you'll probably need such syncing for the early loading
> too...

Thanks. It looks like that is only invoked after boot, with a write to
/sys/devices/system/cpu/microcode/reload.

My series is only parallelising the initial bringup at boot time, so it
shouldn't make any difference.

However... it does look like there's nothing preventing a sibling being
brought online *while* the dance you mention above is occurring.

Shouldn't __reload_late() take the device_hotplug_lock to prevent that?