Messages in this thread | | | From | Binoy Jayan <> | Date | Wed, 21 Sep 2016 11:36:29 +0530 | Subject | Re: [RFC PATCH v7 4/5] tracing: Measure delayed hrtimer offset latency |
| |
On 20 September 2016 at 19:49, Thomas Gleixner <tglx@linutronix.de> wrote: > On Tue, 20 Sep 2016, Binoy Jayan wrote: >> +#ifdef CONFIG_TRACE_DELAYED_TIMER_OFFSETS >> +static inline void latency_hrtimer_timing_start(struct hrtimer *timer, >> + struct hrtimer_clock_base *new_base, >> + ktime_t tim) >> +{ >> + ktime_t now = new_base->get_time(); >> + >> + if (ktime_to_ns(tim) < ktime_to_ns(now)) >> + timer->tim_expiry = now; >> + else >> + timer->tim_expiry = ktime_set(0, 0); > > You still fail to explain why this get_time() magic is required. > > This is executed unconditionally when the config switch is enabled and does > not depend on whether the trace functionality is enabled or not. So you are > imposing the extra get_time() call, which can be expensive depending on the > underlying hardware, on every hrtimer start invocation. > > Tracing is supposed to have ZERO impact when it is not used and even when > it's in use then the impact should be kept as low as possible. The above > does none of that. > > Neither did you provide a proper rationale for this infrastructure in the > changelog. > > You can repost that over and over and it will not go anywhere if you don't > start to address the review comments I give you.
Hi Thomas,
Sorry, I missed to address this comment from you. From what I understand why the get_time() is needed is to get the more accurate current time when the hrtimer base is changed (from the cpu in which it was fired on, to the current cpu on which it is currently made to run or restarted) wherein the hrtimer base needs to be switched to the new cpu provided that it is not running in pinned mode.
Carsten,
Could you please comment on that?
Thanks, Binoy
| |