Re: [PATCH 0/2] ipc/sem.c: sem_lock fixes

[Manfred Spraul]

Hi Andrew,

On 07/14/2016 12:05 AM, Andrew Morton wrote:
> On Wed, 13 Jul 2016 07:06:50 +0200 Manfred Spraul <manfred@xxxxxxxxxxxxxxxx> wrote:
>
> > Hi Andrew, Hi Peter,
> >
> > next version of the sem_lock() fixes:
> > The patches are again vs. tip.
> >
> > Patch 1 is ready for merging, Patch 2 is for review.
> >
> > - Patch 1 is the patch as in -next since January
> >    It fixes the race that was found by Felix.
> > - Patch 2 removes the memory barriers that are part of the qspinlock
> >    code.
> > - (The hysteresis patch would be patch 3. The risk of regressions
> >    can't be ruled out, thus it must wait for benchmarks from real
> >    workload tests)
> I think you're saying that if these two patches cause performance
> regressions, we will need ipc-sem-sem_lock-with-hysteresis.patch?
No, the two patches will not cause any regressions.

Commit 6062a8dc051 ("ipc,sem: fine grained locking for semtimedop") was 
a huge scalability improvement for 99% of the users, but introduced a 
regression for one workload.
Noone complained about it, so the workload must be rare.
Patch 3 now fixes the regression.
But we live with the introduced regression for 3 years, so give me (and 
Davidlohr, and anyone else who wants to support) some more time.
Also on my TODO list: The queue merge/unmerge logic, i.e. commit 
f269f40ad5ae ("ipc/sem.c: always use only one queue for alter 
operations") might also win from some hysteresis.

My proposal:
- patch 1: merge towards main tree.
- patch 2: needs update. The barrier in sem_lock() is required. I'm not 
yet sure about complexmode_enter(), perhaps with #ifndef 
CONFIG_QUEUED_SPINLOCKS
- patch 3 should wait.

--

    Manfred