Re: [Patch v3 Part2 4/9] x86/microcode: Do not call apply_microcode() on sibling threads

[Ashok Raj]

On Wed, Feb 01, 2023 at 11:40:58PM +0100, Borislav Petkov wrote:
> On Wed, Feb 01, 2023 at 02:21:18PM -0800, Dave Hansen wrote:
> > That works great, unless T0 experiences an error.  In that case, T0 will
> > jump out of __reload_late() after failing to do the update.  T1 will
> > come bumbling along after it and will enter ->apply_microcode(),
> > blissfully unaware of T0's failure.  T1 will assume that it is supposed
> > to do T0's job, noting "rev < mc->hdr.rev".  T1 will write the MSR while
> > T0 is off doing god knows what.
> > 
> > T1 should not even be attempting to do ->apply_microcode() because T0 is
> > not quiescent.
> 
> Yah, thanks for explaining properly.
> 
> So, if T0 fails, then we will say that it failed. The ->apply_microcode()
> call on T1 was never meant to apply any microcode - just to update the
> cached data.

The commit log "attempted" to convey that, replace primary with T0, and
secondary with T1.

> 
> Now, if T0 fails, then it doesn't matter what T1 does - you have a
> bigger problem:
> 
> A subset of the cores is running with new microcode while other subset
> with the old one. Now this is a shit situation I don't want to be in.
> 
> And I don't have a good way out of it.

T1 shouldn't be sent down the apply_microcode() path, but instead
just gather and update the per-cpu info reflecting the current revision.

Patch 3 and 4 attempted to that. 

In addition....to ensure cores being out of sync within themselves

At wait_for_siblings: Each thread can check their rev against the BSP (yes
BSP can be removed, but we can elect a core leader) and if they are
different we can either warn/taint or pull the plug and panic.


> 
> Revert to the old patch? Maybe...
> 
> Retry to application on all again with the hope that it works this time?
> 
> What if some core touches a MSR being added with the new microcode
> patch?
> 
> Late loading is a big PITA. As we've been preaching for a while now.
> 

Cheers,
Ashok