On Sun, Jun 20, 2021 at 12:22:53PM +1000, Nicholas Piggin wrote:
> Excerpts from David Gibson's message of June 19, 2021 7:26 pm:
> > On Tue, Jun 15, 2021 at 02:41:07PM +1000, Nicholas Piggin wrote:
> >> There are several new L1D cache flush bits added to the hcall which reflect
> >> hardware security features for speculative cache access issues.
> >> 
> >> These behaviours are now being specified as negative in order to simplify
> >> patched kernel compatibility with older firmware (a new problem found in
> >> existing systems would automatically be vulnerable).
> > 
> > I don't really understand all the consequences of that.  What I need
> > to know here, is if it's safe to unconditionally enable these bits,
> > even for older machine types.
> 
> Unconditionally on the condition that the cap is set to fixed?

Sorry, poor phrashing.  I meant it's not conditional on the machine
version, which means we are technically changing behaviour for
existing machine versions.  That's a big red flag, though there can be
circumstances in which it's justified.

> It
> should be fine AFAIKS. If the older machine types are running on
> older hardware that acutally does require the flush, then the fixed
> cap would cause the existing flush bit to be clear and the kernel to 
> skip the exit flush, so that would be broken already. Does that sound
> right?

Urgh.  The fact that some bits have different sense to others is doing
my head in.

> One thing I'm not entirely clear on is:
> 
> All these (entry/exit/uaccess) flush requirements stem from basically 
> the same underlying mechanism, so that gets resolved in hardware and
> software can stop doing all of them. That's fine, but it was decided to
> add different bits to the hcall basically to have flexibility let's say
> in case a new issue is discovered one day that requires just the uaccess 
> flush, for example.
> 
> In that case we can just set the right combination of bits in firmware,
> and kernels in the field will just do the right thing, and we don't
> need to do all the other flushes that would be worse for performance.

Right, but that doesn't work for qemu guests.  Qemu needs to advertise
things so that guests will do all the things that are necessary not
just on this host, but on any host we might be migrated to.

> How would that work with qemu? I assume we don't have a cap per bit in 
> the hcall, but rather a cap per vulnerability class, so you would set 
> that new class as vulnerable, and this code will have to translate those
> and work out the correct combination of bits to set in these fields.

Something like that, yes.  Actually doing so can be pretty complicated
because of the need to have a consistent migration domain.

> If I'm way off base or there's a better way to do it, that could mean
> this patch needs to be done a different way.

I think we need to find a time to discuss this (Slack or call or
whatever), so you can step me through this bit by bit.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson