Note: there is more. With this applied, my desktop box will no longer reproduce when booted to init 3 with nowatchdog on the command line. My 8 socket DL980 OTOH still will, though it takes longer, and is seemingly no longer interested in following up with a permanent RCU stall after the tsc clocksource is killed, as it does in virgin source.
--- timers_update_migration() is called by tick_nohz_activate() before the late initcall tmigr_init() sets tmigr_enabled to true, resulting in it updating neither timer_base.nohz_active nor .migration_enabled, meaning we'll not kick an idling cpu in add_timer_on(). Remove redundant loop avoidance such that tick_nohz_activate() updates timer_bases[].nohz_active as intended, and call it in tmigr_init() to update timer_bases[].migration_enabled. Signed-off-by: Mike Galbraith <efa...@gmx.de> Fixes: ec2206b91d43 timer: Implement the hierarchical pull model --- kernel/time/timer.c | 4 ---- kernel/time/timer_migration.c | 1 + 2 files changed, 1 insertion(+), 4 deletions(-) --- a/kernel/time/timer.c +++ b/kernel/time/timer.c @@ -224,10 +224,6 @@ void timers_update_migration(bool update bool on = sysctl_timer_migration && tick_nohz_active && tmigr_enabled; unsigned int cpu; - /* Avoid the loop, if nothing to update */ - if (this_cpu_read(timer_bases[BASE_GLOBAL].migration_enabled) == on) - return; - for_each_possible_cpu(cpu) { per_cpu(timer_bases[BASE_LOCAL].migration_enabled, cpu) = on; per_cpu(timer_bases[BASE_GLOBAL].migration_enabled, cpu) = on; --- a/kernel/time/timer_migration.c +++ b/kernel/time/timer_migration.c @@ -649,6 +649,7 @@ static int __init tmigr_init(void) goto hp_err; tmigr_enabled = true; + timers_update_migration(false); pr_info("Timer migration: %d hierarchy levels\n", tmigr_hierarchy_levels); return 0;