-----Original Message----- From: Mike Galbraith [mailto:efault@gmx.de] Sent: Tuesday, April 27, 2021 4:22 PM To: Song Bao Hua (Barry Song) song.bao.hua@hisilicon.com; vincent.guittot@linaro.org; mingo@redhat.com; peterz@infradead.org; dietmar.eggemann@arm.com; rostedt@goodmis.org; bsegall@google.com; mgorman@suse.de Cc: valentin.schneider@arm.com; juri.lelli@redhat.com; bristot@redhat.com; linux-arm-kernel@lists.infradead.org; linux-kernel@vger.kernel.org; xuwei (O) xuwei5@huawei.com; Zengtao (B) prime.zeng@hisilicon.com; guodong.xu@linaro.org; yangyicong yangyicong@huawei.com; Liguozhu (Kenneth) liguozhu@hisilicon.com; linuxarm@openeuler.org; wanghuiqiang wanghuiqiang@huawei.com; xieyongjia (A) xieyongjia1@huawei.com Subject: Re: [PATCH] sched/fair: don't use waker's cpu if the waker of sync wake-up is interrupt
On Tue, 2021-04-27 at 14:37 +1200, Barry Song wrote:
To fix this issue, this patch checks the waker of sync wake-up is a task but not an interrupt. In this case, the waker will schedule out and give CPU to wakee.
That rash "the waker will schedule out" assumption, ie this really really is a synchronous wakeup, may be true in your particular case,
Hi Mike,
For my particular case, sync hint is used by interrupt but not task. so "the waker will schedule out" is false because the existing task is relevant at all.
Here the description "the waker will schedule out" is just copying the general idea of sync wake-up though the real code might overuse this hint.
but the sync hint is so overused as to be fairly meaningless. We've squabbled with it repeatedly over the years because of that. It really should either become more of a communication of intent than it currently is, or just go away.
I agree sync hint might have been overused by other kernel subsystem. But this patch will at least fix a case: sync waker is interrupt, in this case, the existing task has nothing to do with waker and wakee, so this case should be excluded from wake_affine_idle().
I'd argue for go away, simply because there is no way for the kernel to know that there isn't more work directly behind any particular wakeup.
To some extent, "go way" might be a good choice. But this is a more aggressive action. For those cases waker will really scheduler out, wakee won't be able to take the advantage of getting the idle cpu of waker.
Take a pipe, does shoving some bits through a pipe mean you have no further use of your CPU? IFF you're doing nothing but playing ping- pong, sure it does, but how many real proggies have zero overlap among its threads of execution? The mere notion of threaded apps having no overlap *to be converted to throughput* is dainbramaged, which should be the death knell of the sync wakeup hint. Threaded apps can't do stuff like, oh, networking, which uses the sync hint heavily, without at least to some extent defeating the purpose of threading if we were to take the hint seriously. Heck, just look at the beauty (gak) of wake_wide(). It was born specifically to combat the pure-evil side of the sync wakeup hint.
As above, removing the code of migrating wakee to the cpu of sync waker could be an option, but needs more investigations.
Bah, 'nuff "Danger Will Robinson, that thing will *eat you*!!" ;-)
-Mike
Thanks Barry