Re: [PATCH] genirq: use relaxed access by default for irq_reg_{readl,writel}

From: Marc Zyngier
Date: Tue Jan 30 2024 - 07:00:34 EST


On Mon, 29 Jan 2024 14:45:02 +0000,
Lucas Stach <l.stach@xxxxxxxxxxxxxx> wrote:
>
> irqchip access does not require any memory ordering between other
> memory transactions and the IRQ controller peripheral access.
> As all architectures now implement the relaxed MMIO accessors we
> can switch the irq_reg_{readl,writel} helpers to use them, in
> order to avoid potentially costly barriers in the IRQ handling
> hotpath.
>
> Signed-off-by: Lucas Stach <l.stach@xxxxxxxxxxxxxx>
> ---
> include/linux/irq.h | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/include/linux/irq.h b/include/linux/irq.h
> index 90081afa10ce..fa1597db7887 100644
> --- a/include/linux/irq.h
> +++ b/include/linux/irq.h
> @@ -1218,7 +1218,7 @@ static inline void irq_reg_writel(struct irq_chip_generic *gc,
> if (gc->reg_writel)
> gc->reg_writel(val, gc->reg_base + reg_offset);
> else
> - writel(val, gc->reg_base + reg_offset);
> + writel_relaxed(val, gc->reg_base + reg_offset);
> }
>
> static inline u32 irq_reg_readl(struct irq_chip_generic *gc,
> @@ -1227,7 +1227,7 @@ static inline u32 irq_reg_readl(struct irq_chip_generic *gc,
> if (gc->reg_readl)
> return gc->reg_readl(gc->reg_base + reg_offset);
> else
> - return readl(gc->reg_base + reg_offset);
> + return readl_relaxed(gc->reg_base + reg_offset);
> }
>

If this relaxation is introduced, it really should be documented and
require a buy-in, because unsuspecting drivers may implicitly depend
on the stronger ordering.

I'm a strong advocate of the relaxed ordering, but changing this
wholesale is potentially dangerous.

M.

--
Without deviation from the norm, progress is not possible.