Hi Peter, Thanks for your comments.
On 2019/4/24 2:44, Peter Zijlstra wrote: > On Sat, Apr 20, 2019 at 04:34:16PM +0800, Xie XiuQi wrote: >> We meet a divide error on 3.10.0 kernel, the error message is bellow: > > That is a _realllllllyyyy_ old kernel. I would urge you to upgrade. I will. > >> [499992.287996] divide error: 0000 [#1] SMP > >> sched_clock_cpu may not be consistent bwtwen cpus. If a task migrate >> to another cpu, the se.exec_start was set to that cpu's rq_clock_task >> by update_stats_curr_start(). Which may not be monotonic. >> >> update_stats_curr_start >> <- set_next_entity >> <- set_curr_task_fair >> <- sched_move_task > > That is not in fact a cross-cpu migration path. But I see the point. > Also many migration paths do in fact preserve monotonicity, even when > the clock is busted, but you're right, not all of them. > >> So, if now - p->last_task_numa_placement is -1, then (*period + 1) is >> 0, and divide error was triggerred at the div operation: >> task_numa_placement: >> runtime = numa_get_avg_runtime(p, &period); >> f_weight = div64_u64(runtime << 16, period + 1); // divide error here >> >> In this patch, we make sure period is not less than 0 to avoid this >> divide error. >> >> Signed-off-by: Xie XiuQi <[email protected]> >> Cc: [email protected] >> --- >> kernel/sched/fair.c | 4 ++++ >> 1 file changed, 4 insertions(+) >> >> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c >> index 40bd1e27b1b7..f2abb258fc85 100644 >> --- a/kernel/sched/fair.c >> +++ b/kernel/sched/fair.c >> @@ -2007,6 +2007,10 @@ static u64 numa_get_avg_runtime(struct task_struct >> *p, u64 *period) >> if (p->last_task_numa_placement) { >> delta = runtime - p->last_sum_exec_runtime; >> *period = now - p->last_task_numa_placement; >> + >> + /* Avoid backward, and prevent potential divide error */ >> + if ((s64)*period < 0) >> + *period = 0; >> } else { >> delta = p->se.avg.load_sum; >> *period = LOAD_AVG_MAX; > > Yeah, I suppose that is indeed correct. > > I'll try and come up with a better Changelog tomorrow. Thanks. -- Thanks, Xie XiuQi
