Re: [PATCH v2] RISC-V: KVM: Convert IMSIC vsfile_lock to a raw spinlock

[Andrew Jones <andrew.jones@linux.dev>]

On Mon, Oct 19, 2026 at 03:47:09PM +0200, Andrew Jones wrote:
> kvm_riscv_vcpu_aia_imsic_put() runs from the scheduler preemption
> notifier with the raw runqueue lock held, so it must not take
> vsfile_lock, an rwlock_t, which sleeps on PREEMPT_RT. All vsfile_lock
> critical sections are short, so a raw spinlock is fine.
> 
> Fixes: 4cec89db80ba ("RISC-V: KVM: Move HGEI[E|P] CSR access to IMSIC virtualization")
> Signed-off-by: Andrew Jones <andrew.jones@linux.dev>
> ---
> v2:
>  - Convert vsfile_lock to raw_spinlock_t instead of adding a separate
>    hgei_lock [Anup]

Hi Anup,

Sashiko found an ABBA deadlock with this v2[1]. Making vsfile_lock
exclusive serializes kvm_riscv_vcpu_aia_imsic_put(), which runs under
rq->__lock, against the SW-file inject paths, which kick the vCPU
(pi_lock/rq->__lock) while holding vsfile_lock. Lockdep confirms it with
an EMUL mode guest, and the host then hangs (I should have tested with
EMUL before...). v1 doesn't have the problem, since its hgei_lock never
has anything that wakes tasks nested under it. I tested v1 in both EMUL
mode and hwaccel mode [again] with lockdep and it's good.

Should we take v1 instead?

[1] https://sashiko.dev/#/patchset/20261019134709.2610129-1-andrew.jones@linux.dev?part=1

Thanks,
drew