Messages in this thread | | | Subject | Re: [RFC PATCH] timer: Fix bucket_expiry calculation | From | Xiongfeng Wang <> | Date | Thu, 13 May 2021 14:51:15 +0800 |
| |
Hi Thomas,
On 2021/5/12 22:42, Thomas Gleixner wrote: > Xiongfeng, > > On Wed, May 12 2021 at 20:15, Xiongfeng Wang wrote: >> When I use schedule_timeout(5) to put a process into sleep on my machine >> with HZ = 100. It always sleep about 60ms. I enable the timer trace and >> find out, when the timer_list expires, 'now' is always equal to >> 'expires + 1'. I print 'base->next_expiry' in '__run_timers' and find out >> 'next_expiry' is always equal to 'expires + 1'; >> >> It is because we use the following equation to calculate bucket_expiry. >> >> bucket_expiry = ((expires + LVL_GRAN(lvl)) >> LVL_SHIFT(lvl)) << LVL_SHIFT(lvl) >> >> 'bucket_expiry' is equal to 'expires + 1' when lvl = 0. So modify the >> equation as follows to fix the issue. >> >> bucket_expiry = ((expires + LVL_GRAN(lvl) - 1) >> LVL_SHIFT(lvl)) << LVL_SHIFT(lvl) > > That's wrong because you move the expiry of each timer one jiffie ahead, > which violates the guarantee that a timer sleeps at least for one jiffie > for real and not measured in jiffies. > > jiffies = 0 > schedule_timeout(1) > > local_irq_disable() > -> timer interrupt is raised in HW > timer->expires = jiffies + 1 <- 1 > add_timer(timer) > local_irq_enable() > timer interrupt > jiffies++; > softirq() > expire(timer); -> timer is expired immediately > > So the off by one has a reason and is required to prevent too short > timeouts. There is nothing you can do about that because that's a > property of low granularity tick based timer wheels. > > That's even documented in the comment above the code you modified: > > /* > * The timer wheel has to guarantee that a timer does not fire > * early. Early expiry can happen due to: > * - Timer is armed at the edge of a tick > * - Truncation of the expiry time in the outer wheel levels > * > * Round up with level granularity to prevent this. > */
Thanks for your explanation. I got it !
Thanks, Xiongfeng
> > Thanks, > > tglx > . >
| |