lkml.org 
[lkml]   [2018]   [Apr]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v5] ANDROID: binder: change down_write to down_read
Hi Ganesh,

On Mon, Apr 02, 2018 at 06:01:59PM +0800, Ganesh Mahendran wrote:
> 2018-04-02 15:11 GMT+08:00 Minchan Kim <minchan@kernel.org>:
> > On Mon, Apr 02, 2018 at 02:46:14PM +0800, Ganesh Mahendran wrote:
> >> 2018-04-02 14:34 GMT+08:00 Minchan Kim <minchan@kernel.org>:
> >> > On Fri, Mar 30, 2018 at 12:04:07PM +0200, Greg Kroah-Hartman wrote:
> >> >> On Fri, Mar 30, 2018 at 10:29:21AM +0900, Minchan Kim wrote:
> >> >> > Hi Ganesh,
> >> >> >
> >> >> > On Fri, Mar 30, 2018 at 09:21:55AM +0800, Ganesh Mahendran wrote:
> >> >> > > 2018-03-29 14:54 GMT+08:00 Minchan Kim <minchan@kernel.org>:
> >> >> > > > binder_update_page_range needs down_write of mmap_sem because
> >> >> > > > vm_insert_page need to change vma->vm_flags to VM_MIXEDMAP unless
> >> >> > > > it is set. However, when I profile binder working, it seems
> >> >> > > > every binder buffers should be mapped in advance by binder_mmap.
> >> >> > > > It means we could set VM_MIXEDMAP in binder_mmap time which is
> >> >> > > > already hold a mmap_sem as down_write so binder_update_page_range
> >> >> > > > doesn't need to hold a mmap_sem as down_write.
> >> >> > > >
> >> >> > > > Android suffers from mmap_sem contention so let's reduce mmap_sem
> >> >> > > > down_write.
> >> >> > >
> >> >> > > Hi, Minchan:
> >> >> > >
> >> >> > > It seems there is performance regression of this patch.
> >> >> >
> >> >> > You mean "This patch aims for solving performance regression" not "This patch
> >> >> > makes performance regression"?
> >> >> >
> >> >> > >
> >> >> > > Do you have some test result of android app launch time or binderThroughput?
> >> >> >
> >> >> > Unfortunately, I don't have any number. The goal is to reduce the number of
> >> >> > call mmap_sem as write-side lock because it makes priority inversion of threads
> >> >> > easily and that's one of clear part I spot that we don't need write-side lock.
> >> >>
> >> >> Please always run the binderThroughput tests when making binder changes
> >> >> (there is a binder test suite in the CTS Android tests), as that ensures
> >> >> that you are not causing performance regressions as well as just normal
> >> >> bug regressions :)
> >> >
> >> > Thanks for the information. I didn't notice that such kinds of tests for
> >> > binder. I will keep it in mind.
> >> >
> >> > Today, I have setup the testing for my phone and found testing was very
> >> > fluctuating even without my patch. It might be not good with my test
> >> > skill. I emulated user's behavior with various touch event. With it, I open
> >> > various apps and play with them several times. Before starting the test,
> >> > I did "adb shell stop && adb shell start && echo 3 > /proc/sys/vm/drop_caches"
> >> >
> >> > Such 15% noise was very easy to make it.
> >> >
> >> > Ganesh, How did you measure? What's the stddev?
> >>
> >> Hi, Minchan:
> >>
> >> Sorry for the late response, a little busy these days. :)
> >>
> >> We have our own test tools to measure app launch time, or you can use
> >> android systrace to get the app launch time. We tested your V1 patch:
> >> https://patchwork.kernel.org/patch/10312057/
> >> and found app lunch time regression.
> >
> > V1 had a bug with VM_MAYWRITE. Could you confirm it with v5?
>
> I have finished binder Throughput test. The test result is stable,
> there is no performance
> regression found both in v1 and v5.

Thanks for the test! Now I'm struggling with setting up BinderThrough test.
Binder matainers:
If it's really one every binder contributors should do before the
sending their patch, couldn't we have them in kernel directory like kselftest?
Like me who understand just a part of code, it's hard to download android
userspace full code and build/test.


>
> base patch_v1 patch_v5
> -----------------------------------------------------------
> 91223.4 90560.2 89644.5
> 90520.3 89583.1 89048.2
> 89833.2 90247.6 90091.3
> 90740.2 90276.7 90994.2
> 89703.5 90112.4 89994.6
> 89945.1 89122.8 88937.7
> 89872.8 90357.3 89307.4
> 89913.2 90355.4 89563.8
> 88979 90393.4 90182.8
> 89577.3 90946.8 90441.4
> AVG 90030.8 90195.57 89820.59

Yes, no regression.

>
> Before the test, I stop the android framework by:
> adb shell stop
>
> >
> > Please tell me more detail. What apps are slower compared to old?
> > Every apps are slowed with avg 15%? Then, what's the stddev?
>
> Not all of the apps slowed 15%, The app *avg* launch time slowed 15%.
> And We will re-launch the test tomorrow: base, v1,v5. We will get the
> test result in two days later. Then I will post all the app launch time details.

I'm also trying to make stable result in my side but it's really hard to
get. Please post stddev of each app as well as avg when you finished testing.
I really appreicate you.

>
> >
> > The reason I'm asking is as I mentioned, it would be caused by rw_semaphore
> > implementation and priority of threads which calls binder operation so I
> > guess it would be not deterministic.
> >
> > When I had an simple experiment, it was very fluctuating as I expected.
> > (the testing enviroment might be not good in my side).
> > If it's real problem on real practice, better fix is not using write_lock
> > of mmap_sem(it's abusing the write-side lock) but should adjust priority,
> > I think. What do you think?
>
> If you want to narrow the range of the problem. We can disable binder priority
> inherit, and do not set the priority(currently it is nice -10 or fifo)
> of top app in Android AMS.
> I think we need to wait for the test result to see whether it really
> has performance
> regression.

Look at up_write.

(Let's assume process B is head of wait list of rw_semaphore, and then C, D, E)
If the process B is trying to down_write and previous lock holder A is
called up_write, the only B could be waked up so there is no contention
to get CPU slice. It's the current as-is but if we changes B to try to
down_read instead of down_write, B should be competed with other down_read
C,D,E in so the chance would be rare to be scheduled.

It's really (timing|priority of binder and other threads) problem so I don't
understand what you said how we could narrow down the problem with disabling
binder priority.

\
 
 \ /
  Last update: 2018-04-02 12:32    [W:0.087 / U:4.260 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site