On Mon Mar 30, 2026 at 2:58 AM PDT, Troy Mitchell wrote: > From: Troy Mitchell > > Currently, the RISC-V implementation of machine_restart(), machine_halt(), > and machine_power_off() invokes the kernel teardown chains (e.g., > do_kernel_restart()) with local interrupts enabled and other CPUs still > running. > > This implementation fails to provide a deterministic execution environment > for registered handlers in the restart or power-off notifier chains. These > chains are intended to be executed in a strict atomic and single-threaded > context. > > Specifically, under CONFIG_PREEMPT_RCU, rcu_read_lock() does not increment > the preempt_count. If local interrupts remain enabled, the environment > is not guaranteed to be atomic. This can lead to a context misidentification > within generic kernel teardown code, causing it to incorrectly enter > non-atomic paths (such as attempting to acquire sleeping locks), which > results in fatal "scheduling while atomic" splats or system hangs. > > Additionally, stopping other CPUs ensures the primary CPU has exclusive > access to the hardware state during the final teardown phase, preventing > unpredictable interference from other active cores. > > Align RISC-V with other major architectures by disabling local interrupts > and stopping other CPUs at the beginning of the shutdown sequences. This > guarantees the architectural expectations of the kernel's restart and > power-off handlers are met. > > Signed-off-by: Troy Mitchell > Tested-by: Aurelien Jarno Hi Paul, A gentle ping on this patch. When you have a chance, could you please take a look and let me know whether this approach looks reasonable? - Troy