Re: [PATCH 1/2] sched/wait: Break up long wake list walk

[Tim Chen]

On 08/14/2017 08:28 PM, Linus Torvalds wrote:
> On Mon, Aug 14, 2017 at 8:15 PM, Andi Kleen <ak@xxxxxxxxxxxxxxx> wrote:
>> But what should we do when some other (non page) wait queue runs into the
>> same problem?
> 
> Hopefully the same: root-cause it.
> 
> Once you have a test-case, it should generally be fairly simple to do
> with profiles, just seeing who the caller is when ttwu() (or whatever
> it is that ends up being the most noticeable part of the wakeup chain)
> shows up very heavily.

We have a test case but it is a customer workload.  We'll try to get
a bit more info.

> 
> And I think that ends up being true whether the "break up long chains"
> patch goes in or not. Even if we end up allowing interrupts in the
> middle, a long wait-queue is a problem.
> 
> I think the "break up long chains" thing may be the right thing
> against actual malicious attacks, but not for any actual real
> benchmark or load.

This is a concern from our customer as we could trigger the watchdog timer
by running user space workloads.  

> 
> I don't think we normally have cases of long wait-queues, though. At
> least not the kinds that cause problems. The real (and valid)
> thundering herd cases should already be using exclusive waiters that
> only wake up one process at a time.
> 
> The page bit-waiting is hopefully special. As mentioned, we used to
> have some _really_ special code for it for other reasons, and I
> suspect you see this problem with them because we over-simplified it
> from being a per-zone dynamically sized one (where the per-zone thing
> caused both performance problems and actual bugs) to being that
> "static small array".
> 
> So I think/hope that just re-introducing some dynamic sizing will help
> sufficiently, and that this really is an odd and unusual case.

I agree that dynamic sizing makes a lot of sense.  We'll check to
see if additional size to the hash table helps, assuming that the
waiters are distributed among different pages for our test case.  

Thanks.

Tim