[systemd-devel] kdbus performance regression by ~70% on 3.15 kernels ?

[Djalal Harouni]

On Tue, Jul 01, 2014 at 09:26:01AM +0200, Daniel Mack wrote:
> On 06/28/2014 01:55 AM, Steven Noonan wrote:
> > Was going to try to repro this perf regression as well, but instead got
> > kdbus to oops (via test-kdbus-benchmark):
> > 
> > $ test/test-kdbus-benchmark 
> > -- opening /dev/kdbus/control
> > -- creating bus '1000-testbus'
> > -- opening bus connection /dev/kdbus/1000-testbus/bus
> > -- Our peer ID for /dev/kdbus/1000-testbus/bus: 1 -- bus uuid:
> > 'b65bfdd23d3e4696aae2992a0857aa33'
> > -- opening bus connection /dev/kdbus/1000-testbus/bus
> > -- Our peer ID for /dev/kdbus/1000-testbus/bus: 2 -- bus uuid:
> > 'b65bfdd23d3e4696aae2992a0857aa33'
> > name_acquire(): flags after call: 0x0                                                                                                
> > Killed
> > $
> 
> ...
> 
> 
> > [   33.558642] Stack:
> > [   33.558650]  ffffffffa11f45d6 0000000000000000 ffff8800c82cff50 0000000000000010
> > [   33.558681]  00007fff80d23800 ffff8800c82cff50 0000000000000000 ffff8800c82cfef8
> > [   33.558712]  ffffffff811d386a 00007fff80d23800 0000000000000010 ffff8803f9d5d400
> > [   33.558743] Call Trace:
> > [   33.558757]  [<ffffffffa11f45d6>] ? kdbus_memfd_writev+0x66/0xa0 [kdbus]
> 
> This bug is fixed upstream with 7da2745e ("memfd: switch to
> f_op->{read,write}_iter"). This fix was neccessary due to vfs internal
> refactoring that came in for 3.16-rc1.
Yes,

> Bisecting between v3.15 and 3.16-rc1 can hence be a little tricky. You
> might have to revert this particular patch for some bisect iterations in
> case you get build breakage.
Indeed, revert and switch to the old aio_{read,write} to compare, since
my suspicions points to this API change, and it's not easy to get it
right, the old generic_file_aio_{read,write} was removed... so I'm
really stuck... but I'll try to continue and try to report.

I don't know if memfd needs to do something special, or just convert!
I really don't know...

> It might also be a good idea to investigate whether one specific meta
> data attachments causes the regression.
Ok will do, thanks!


> 
> Thanks,
> Daniel

-- 
Djalal Harouni
http://opendz.org