On Tue, Oct 05, 2021, Paolo Bonzini wrote:
> On 05/10/21 17:41, Sean Christopherson wrote:
> > > if (*gfn_track == NULL) {
> > > mutex_unlock(&kvm->slots_arch_lock);
> > Hrm, this fails to free the gfn_track allocations for previous memslots. The
> > on-demand rmaps code has the exact same bug (it frees rmaps for previous lpages
> > in the_current_ slot, but does not free previous slots).
>
> That's not a huge deal because the syscall is failing. So as long as it's
> not leaked forever, it's okay. The problem is the
> WARN_ON(slot->arch.rmap[i]), or the missing check in
> kvm_page_track_enable_mmu_write_tracking, but that's easily fixed. I'd even
> remove the call to memslot_rmaps_free.
It can be leaked forever though, e.g. if userspace invokes KVM_RUN over and over
on -ENOMEM. That would trigger the WARN_ON(slot->arch.rmap[i]) and leak the
previous allocation. I think it would be safe to change that WARN_ON to a
check-and-continue, i.e. to preserve the previous allocation
> > And having two separate flows (and flags) for rmaps vs. gfn_track is pointless,
> > and means we have to maintain two near-identical copies of non-obvious code.
>
> I was thinking the separate flow (not so much the flag) is needed because,
> if KVMGT is enabled, gfn_track is allocated unconditionally. rmaps are added
> on top of that if shadow paging is enabled; but
> kvm_page_track_create_memslot will have already created the counter,
> including the one for KVM_PAGE_TRACK_WRITE.
>
> But looking at the code again, I guess you could call
> kvm_page_track_enable_mmu_write_tracking inside alloc_all_memslots_rmaps
> (with a little bit of renaming), and with that the flag would go away.
Yes, and reuse the control flow, which is what I really care about since that's
the part that both features get wrong.
> I'll take a look tomorrow, but I'd rather avoid reverting the patch.
I can poke at it too if you don't have time. I wasn't suggesting a full revert,
rather a "drop and pretend it never got applied", with a plan to apply a new
version instead of fixing up the current code.
