Re: [PATCH dovetail v8 02/11] riscv: irq_pipeline: add IRQ pipelining core

On Sun, 2026-05-31 at 07:52 +0200, Tobias Schaffner wrote:
> This patchset integrates IRQ pipelining into the RISC-V architecture,
> bringing it in line with the Dovetail/IRQ pipeline model used on other
> architectures. It adds the core pipelining infrastructure and adapts
> low-level primitives to cleanly separate in-band and out-of-band
> interrupt handling.
> 
> Signed-off-by: Tobias Schaffner <[email protected]>
> Co-authored-by: shannmu <[email protected]>
> Co-authored-by: Philippe Gerum <[email protected]>
> ---

> [snip]
>  
> +static __always_inline
> +bool mark_trap_entry(struct pt_regs *regs)
> +{
> +	if (running_oob()) {
> +		return false;
> +	}
> +
> +	if (user_mode(regs))
> +		hard_cond_local_irq_enable();

I had to debug x86 today and noticed a difference here:

hard_cond_local_irq_enable() and the _disable() part below are not
protected by the user_mode() check in x86 and I think that is correct.
We would like to allow hard IRQs to arrive, keeping the pipeline active
while handling a trap.

In addition I would vote for using the same checks/ordering as other
architectures do. I haven't checked arm{64}, but x86 is using a 
running_inband() check instead of running_oob(). That should help to
avoid stupid mistakes in the future.

> +
> +	return true;
> +}
> +
> +static __always_inline
> +void mark_trap_exit(struct pt_regs *regs)
> +{
> +	if (user_mode(regs))
> +		hard_cond_local_irq_disable();
> +}
> +
> 

[snip]