Re: [PATCH] kvm: defer huge page recovery vhost task to later

[Sean Christopherson]

On Fri, Jan 24, 2025, Keith Busch wrote:
> On Fri, Jan 24, 2025 at 12:07:24PM -0800, Sean Christopherson wrote:
> > This is broken.  If the module param is toggled before the first KVM_RUN, KVM
> > will hit a NULL pointer deref due to trying to start a non-existent vhost task:
> > 
> >   BUG: kernel NULL pointer dereference, address: 0000000000000040
> >   #PF: supervisor read access in kernel mode
> >   #PF: error_code(0x0000) - not-present page
> >   PGD 0 P4D 0 
> >   Oops: Oops: 0000 [#1] SMP
> >   CPU: 16 UID: 0 PID: 1190 Comm: bash Not tainted 6.13.0-rc3-9bb02e874121-x86/xen_msr_fixes-vm #2382
> >   Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
> >   RIP: 0010:vhost_task_wake+0x5/0x10
> >   Call Trace:
> >    <TASK>
> >    set_nx_huge_pages+0xcc/0x1e0 [kvm]
> 
> Thanks for pointing out this gap. It looks like we'd have to hold the
> kvm_lock in kvm_mmu_post_init_vm(), and add NULL checks in
> set_nx_huge_pages() and set_nx_huge_pages_recovery_param() to prevent
> the NULL deref. Is that okay?

I don't _think_ we need to take kvm_lock.  And I don't want to take kvm_lock,
because we're also trying to eliminate a (very theoretical) deadlock[1] due to
taking kvm_lock in the params helpers.

There is a race that can happen with my proposed fix[2], but I'm not sure we care
enough to address it.  If kvm_nx_huge_page_recovery_worker() runs before the params
are set, and the param setter processes the VM before nx_huge_page_recovery_thread
is set, then the worker could sleep for too long, relative to what userspace expects.

I suppose if we care then we could fix that by taking kvm->arch.nx_once.mutex
when waking the task?

[1] https://lore.kernel.org/all/20250124191109.205955-2-pbonzini@xxxxxxxxxx
[2] https://lore.kernel.org/all/20250124234623.3609069-1-seanjc@xxxxxxxxxx