Re: [PATCH v4 2/2] mm: make lru_add_drain_all() selective

[Chris Metcalf <cmetcalf@tilera.com>]

On 8/13/2013 6:26 PM, Andrew Morton wrote:
> On Tue, 13 Aug 2013 18:13:48 -0400 Chris Metcalf <cmetcalf@tilera.com> wrote:
>
>> On 8/13/2013 5:13 PM, Andrew Morton wrote:
>>> On Tue, 13 Aug 2013 16:59:54 -0400 Chris Metcalf <cmetcalf@tilera.com> wrote:
>>>
>>>>> Then again, why does this patchset exist?  It's a performance
>>>>> optimisation so presumably someone cares.  But not enough to perform
>>>>> actual measurements :(
>>>> The patchset exists because of the difference between zero overhead on
>>>> cpus that don't have drainable lrus, and non-zero overhead.  This turns
>>>> out to be important on workloads where nohz cores are handling 10 Gb
>>>> traffic in userspace and really, really don't want to be interrupted,
>>>> or they drop packets on the floor.
>>> But what is the effect of the patchset?  Has it been tested against the
>>> problematic workload(s)?
>> Yes.  The result is that syscalls such as mlockall(), which otherwise interrupt
>> every core, don't interrupt the cores that are running purely in userspace.
>> Since they are purely in userspace they don't have any drainable pagevecs,
>> so the patchset means they don't get interrupted and don't drop packets.
>>
>> I implemented this against Linux 2.6.38 and our home-grown version of nohz
>> cpusets back in July 2012, and we have been shipping it to customers since then.
> argh.
>
> Those per-cpu LRU pagevecs were a nasty but very effective locking
> amortization hack back in, umm, 2002.  They have caused quite a lot of
> weird corner-case behaviour, resulting in all the lru_add_drain_all()
> calls sprinkled around the place.  I'd like to nuke the whole thing,
> but that would require a fundamental rethnik/rework of all the LRU list
> locking.
>
> According to the 8891d6da17db0f changelog, the lru_add_drain_all() in
> sys_mlock() isn't really required: "it isn't must.  but it reduce the
> failure of moving to unevictable list.  its failure can rescue in
> vmscan later.  but reducing is better."
>
> I suspect we could just kill it.

That's probably true, but I suspect this change is still worthwhile for
nohz environments.  There are other calls of lru_add_drain_all(), and
you just don't want anything in the kernel that interrupts every core
when only a subset could be interrupted.  If the kernel can avoid
generating unnecessary interrupts to uninvolved cores, you can make
guarantees about jitter on cores that are running dedicated userspace code.

-- 
Chris Metcalf, Tilera Corp.
http://www.tilera.com