At 15:04 +0100 on 12 Nov (1605193496), Jan Beulich wrote: > On 12.11.2020 14:07, Roger Pau Monné wrote: > > On Thu, Nov 12, 2020 at 01:29:33PM +0100, Jan Beulich wrote: > >> I agree with all this. If only it was merely about TLB flushes. In > >> the shadow case, shadow_blow_all_tables() gets invoked, and that > >> one - looking at the other call sites - wants the paging lock held. [...] > > The post hook for shadow could pick the lock again, as I don't think > > the removal of the tables needs to be strictly done inside of the same > > locked region? > > I think it does, or else a use of the now stale tables may occur > before they got blown away. Tim?
Is this the call to shadow_blow_tables() in the write_p2m_entry path? I think it would be safe to drop and re-take the paging lock there as long as the call happens before the write is considered to have finished. But it would not be a useful performance improvement - any update that takes this path is going to be very slow regardless. So unless you have another pressing reason to split it up, I would be inclined to leave it as it is. That way it's easier to see that the locking is correct. Cheers, Tim.
