On 2023/9/8 Tim Chen wrote:
On Thu, 2023-09-07 at 16:59 +0800, Hao Jia wrote:
On 2023/9/5 Peter Zijlstra wrote:
On Thu, Aug 31, 2023 at 04:48:29PM +0800, Hao Jia wrote:
If I understand correctly, rq->clock_update_flags may be set to RQCF_ACT_SKIP after __schedule() holds the rq lock, and sometimes the rq lock may be released briefly in __schedule(), such as newidle_balance(). At this time Other CPUs hold this rq lock, and then calling rq_clock_start_loop_update() may trigger this warning.
This warning check might be wrong. We need to add assert_clock_updated() to check that the rq clock has been updated before calling rq_clock_start_loop_update().
Maybe some things can be like this?
Urgh, aside from it being white space mangled, I think this is entirely going in the wrong direction.
Leaking ACT_SKIP is dodgy as heck.. it's entirely too late to think clearly though, I'll have to try again tomorrow.
I am trying to understand why this is an ACT_SKIP leak. Before call to __cfsb_csd_unthrottle(), is it possible someone else lock the runqueue, set ACT_SKIP and release rq_lock? And then that someone never update the rq_clock?
Yes, we want to set rq->clock_update_flags to RQCF_ACT_SKIP to avoid updating the rq clock multiple times in __cfsb_csd_unthrottle().
But now we find ACT_SKIP leak, so we cannot unconditionally set rq->clock_update_flags to RQCF_ACT_SKIP in rq_clock_start_loop_update().
Hi Peter,
Do you think this fix method is correct? Or should we go back to the beginning and move update_rq_clock() from unthrottle_cfs_rq()?
If anyone who locked the runqueue set ACT_SKIP also will update rq_clock, I think your change is okay. Otherwise rq_clock could be missing update.
Thanks.
Tim