Messages in this thread | | | Date | Mon, 29 Apr 2013 14:53:07 -0400 | From | KOSAKI Motohiro <> | Subject | Re: [PATCH 2/2] posix-cpu-timers: fix wrong timer initialization |
| |
>>> @@ -749,7 +756,13 @@ static int posix_cpu_timer_set(struct k_itimer *timer, int flags, >>> } >>> >>> if (new_expires.sched != 0 && !(flags & TIMER_ABSTIME)) { >>> - cpu_time_add(timer->it_clock, &new_expires, val); >>> + union cpu_time_count now; >>> + >>> + if (CPUCLOCK_PERTHREAD(timer->it_clock)) >>> + cpu_clock_sample(timer->it_clock, p, &now); >>> + else >>> + cpu_clock_sample_group(timer->it_clock, p, &now); >> >> This triggered a pattern match against earlier in this function; but they're >> different now; timer vs clock. So nothing to merge... > > Not different, I think. > Relative timeout need to calculate "now + timeout" by definition. > > But which time is "now"? > > Example, thread1 has 10ms sum_exec_runtime and 4ms delta and call timer_settime(4ms). > Old code calculate an expire is 10+4=14. New one calculate 10+4+4=18. > > Which expire is correct? When using old one, timer will fire just after syscall. This > is posix violation. > > In the other words, > > sighandler(){ > t1 = clock_gettime() > } > > t0 = clock_gettime() > timer_settime(timeout); > ... wait to fire > > assert (t1 - t0 >= timeout) > > This pseudo code must be true. it is snippest what glibc rt/tst-cputimer1 test and failed.
In the other hands, following two calculations need to timer time (aka time without delta).
1) Initialization signal->cputimer for avoiding double delta count. 2) calculating old tiemr because timer firing logic (run_posix_cpu_timers) don't care delta_exec.
| |