On 02/18, Ingo Molnar wrote:

* Oleg Nesterov [off-list ref] wrote:

quoted

Unlike __mod_timer(..., bool pending_only), it preserves the CPU on
which the timer is pending.

Or, perhaps, we can modify __mod_timer() further,

if then i'd put it into a separate commit.

I think the auto-migration of all the mod_timer() variants is a
scalability feature: if for example a networking socket's main
user migrates to another CPU, then the timer 'follows' it - even
if the timer never actually expires (which is quite common for
high-speed high-reliability networking transports).

OK.

But sometimes it is better (or necessary) to prevent the migration.
Since you already are changed __mod_timer() it would be ugly to
add yet another helper. Perhaps we should turn "bool pending_only"
into "int flags" right now?

This is minor, and perhaps we will never need the TIMER_DONT_MIGRATE
flag. But if ever need, then we have to audit all callers.

Oleg.