During an IPI procedure, we need to ensure all previous write operations
are visible to other CPUs before sending a real IPI. We use wmb() barrier
to ensure this as CLINT issues IPI via mmio writes.

Signed-off-by: Xu Lu <luxu.kernel@bytedance.com>
---
 drivers/clocksource/timer-clint.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/drivers/clocksource/timer-clint.c b/drivers/clocksource/timer-clint.c
index 0bdd9d7ec545..8d73b45f9966 100644
--- a/drivers/clocksource/timer-clint.c
+++ b/drivers/clocksource/timer-clint.c
@@ -48,6 +48,12 @@ EXPORT_SYMBOL(clint_time_val);
 #ifdef CONFIG_SMP
 static void clint_send_ipi(unsigned int cpu)
 {
+	/*
+	 * Ensure that stores to normal memory are visible to the other CPUs
+	 * before issuing IPI.
+	 */
+	wmb();
+
 	writel(1, clint_ipi_base + cpuid_to_hartid_map(cpu));
 }
 
-- 
2.20.1

On Mon, Jan 13, 2025 at 8:39 PM Xu Lu <luxu.kernel@bytedance.com> wrote:
>
> During an IPI procedure, we need to ensure all previous write operations
> are visible to other CPUs before sending a real IPI. We use wmb() barrier
> to ensure this as CLINT issues IPI via mmio writes.
>
> Signed-off-by: Xu Lu <luxu.kernel@bytedance.com>
> ---
>  drivers/clocksource/timer-clint.c | 6 ++++++
>  1 file changed, 6 insertions(+)
>
> diff --git a/drivers/clocksource/timer-clint.c b/drivers/clocksource/timer-clint.c
> index 0bdd9d7ec545..8d73b45f9966 100644
> --- a/drivers/clocksource/timer-clint.c
> +++ b/drivers/clocksource/timer-clint.c
> @@ -48,6 +48,12 @@ EXPORT_SYMBOL(clint_time_val);
>  #ifdef CONFIG_SMP
>  static void clint_send_ipi(unsigned int cpu)
>  {
> +       /*
> +        * Ensure that stores to normal memory are visible to the other CPUs
> +        * before issuing IPI.
> +        */
> +       wmb();
> +

Same comment as PATCH3.

The clint_send_ipi() is called through ipi_mux_send_mask()
which does smp_mb__after_atomic() before calling so no need
for any barrier here. Also, barriers need to be in-pair so adding
a single barrier at random location is inappropriate.
(Refer, https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/irq/ipi-mux.c?h=v6.13-rc7#n78)

Based on the above, this patch is not needed.

Regards,
Anup