Re: [PATCH] KVM: arm64: vgic: Use list_del_rcu() when flushing pending LPIs

[Marc Zyngier <maz@kernel.org>]

On Fri, 05 Jun 2026 06:47:17 +0100,
Oliver Upton <oupton@kernel.org> wrote:
> 
> Hi Hyunwoo,
> 
> On Fri, Jun 05, 2026 at 06:16:08AM +0900, Hyunwoo Kim wrote:
> > vgic_v3_fold_lr_state() walks the ap_list from last_lr_irq without holding
> > the ap_list_lock, relying on vgic_irq being freed via kfree_rcu() and on
> > interrupts being disabled. vgic_flush_pending_lpis() removes entries with
> > list_del(), which clobbers a node's next pointer, so when another vCPU
> > disables LPIs via GICR_CTLR the walk can follow the clobbered next pointer
> > from a removed node, or from the node that last_lr_irq points to.
> > 
> > Remove entries with list_del_rcu() so that the next pointer stays valid
> > until the walk completes.
> > 
> > Fixes: 3cfd59f81e0f ("KVM: arm64: GICv3: Handle LR overflow when EOImode==0")
> > Signed-off-by: Hyunwoo Kim <imv4bel@gmail.com>
> 
> Changing only one of the writer paths to use the rculist helpers does
> not make the ap_list an rculist. Insertions are not RCU-safe, nor are
> deleations from vgic_prune_ap_list().
> 
> And TBH, the real bug here is the fact that vgic_v3_fold_lr_state() isn't
> taking the ap_list_lock.

Yup, that'd be more sensible. I need to convince myself that there is
no possible path from vgic_v*_fold_lr() to vgic_irq_queue_unlock(),
because that one does actually acquire that lock.

	M.

-- 
Jazz isn't dead. It just smells funny.