Re: [PATCH RT 10/30] hrtimer: Prevent using hrtimer_grab_expiry_lock() on migration_base

[Rasmus Villemoes]

On 23/01/2020 21.39, Steven Rostedt wrote:
> 4.19.94-rt39-rc2 stable review patch.
> If anyone has any objections, please let me know.
>
> ------------------
> 
> From: Julien Grall <julien.grall@xxxxxxx>
> 
> [ Upstream commit cef1b87f98823af923a386f3f69149acb212d4a1 ]
> 
> As tglx puts it:
> |If base == migration_base then there is no point to lock soft_expiry_lock
> |simply because the timer is not executing the callback in soft irq context
> |and the whole lock/unlock dance can be avoided.

Hold on a second. This patch (hrtimer: Prevent using
hrtimer_grab_expiry_lock() on migration_base) indeed seems to implement
the optimization implied by the above, namely avoid the lock/unlock in
case base == migration_base:

> -	if (timer->is_soft && base && base->cpu_base) {
> +	if (timer->is_soft && base != &migration_base) {

But the followup patch (hrtimer: Add a missing bracket and hide
`migration_base on !SMP) to fix the build on !SMP [the missing bracket
part seems to have been fixed when backporting the above to 4.19-rt]
replaces that logic by

+static inline bool is_migration_base(struct hrtimer_clock_base *base)
+{
+	return base == &migration_base;
+}
+
...
-	if (timer->is_soft && base != &migration_base) {
+	if (timer->is_soft && is_migration_base(base)) {

in the SMP case, i.e. the exact opposite condition. One of these can't
be correct.

Assuming the followup patch was wrong and the condition should have read

  timer->is_soft && !is_migration_base(base)

while keeping is_migration_base() false on !SMP might explain the
problem I see. But I'd like someone who knows this code to chime in.

Thanks,
Rasmus