Re: [PATCH v5.5 26/30] KVM: Keep memslots in tree-based structures instead of array-based ones

[Maciej S. Szmigiero]

On 12.11.2021 01:51, Sean Christopherson wrote:
> On Fri, Nov 12, 2021, Maciej S. Szmigiero wrote:
> > On 04.11.2021 01:25, Sean Christopherson wrote:
> > > -	/*
> > > -	 * Remove the old memslot from the hash list and interval tree, copying
> > > -	 * the node data would corrupt the structures.
> > > -	 */
> > > +	int as_id = kvm_memslots_get_as_id(old, new);
> > > +	struct kvm_memslots *slots = kvm_get_inactive_memslots(kvm, as_id);
> > > +	int idx = slots->node_idx;
> > > +
> > >    	if (old) {
> > > -		hash_del(&old->id_node);
> > > -		interval_tree_remove(&old->hva_node, &slots->hva_tree);
> > > +		hash_del(&old->id_node[idx]);
> > > +		interval_tree_remove(&old->hva_node[idx], &slots->hva_tree);
> > > -		if (!new)
> > > +		if ((long)old == atomic_long_read(&slots->last_used_slot))
> > > +			atomic_long_set(&slots->last_used_slot, (long)new);
> >
> > Open-coding cmpxchg() is way less readable than a direct call.
>
> Doh, I meant to call this out and/or add a comment.
>
> My objection to cmpxchg() is that it implies atomicity is required (the kernel's
> version adds the lock), which is very much not the case.  So this isn't strictly
> an open-coded version of cmpxchg().
>
> > The open-coded version also compiles on x86 to multiple instructions with
> > a branch, instead of just a single instruction.
>
> Yeah.  The lock can't be contended, so that part of cmpxchg is a non-issue.  But
> that's also why I don't love using cmpxchg.
>
> I don't have a strong preference, I just got briefly confused by the atomicity part.

We can simply add a comment there to explain that the atomicity isn't actually
strictly required here - will do that.

> > > +static void kvm_invalidate_memslot(struct kvm *kvm,
> > > +				   struct kvm_memory_slot *old,
> > > +				   struct kvm_memory_slot *working_slot)
> > > +{
> > > +	/*
> > > +	 * Mark the current slot INVALID.  As with all memslot modifications,
> > > +	 * this must be done on an unreachable slot to avoid modifying the
> > > +	 * current slot in the active tree.
> > > +	 */
> > > +	kvm_copy_memslot(working_slot, old);
> > > +	working_slot->flags |= KVM_MEMSLOT_INVALID;
> > > +	kvm_replace_memslot(kvm, old, working_slot);
> > > +
> > > +	/*
> > > +	 * Activate the slot that is now marked INVALID, but don't propagate
> > > +	 * the slot to the now inactive slots. The slot is either going to be
> > > +	 * deleted or recreated as a new slot.
> > > +	 */
> > > +	kvm_swap_active_memslots(kvm, old->as_id);
> > > +
> > > +	/*
> > > +	 * From this point no new shadow pages pointing to a deleted, or moved,
> > > +	 * memslot will be created.  Validation of sp->gfn happens in:
> > > +	 *	- gfn_to_hva (kvm_read_guest, gfn_to_pfn)
> > > +	 *	- kvm_is_visible_gfn (mmu_check_root)
> > > +	 */
> > > +	kvm_arch_flush_shadow_memslot(kvm, old);
> >
> > This should flush the currently active slot (that is, "working_slot",
> > not "old") to not introduce a behavior change with respect to the existing
> > code.
> >
> > That's also what the previous version of this patch set did.
>
> Eww.  I would much prefer to "fix" the existing code in a prep patch.  It shouldn't
> matter, but arch code really should not get passed an INVALID slot.

I will add a separate patch that switches that kvm_arch_flush_shadow_memslot()
call to use a valid (old) memslot instead.

It is actually simpler to do it *after* the main patch series to not add
more dead code that next patches remove anyway.

Thanks,
Maciej