Re: [PATCH 2/2] drm: Revert syncobj timeline changes.

[Chris Wilson]

Quoting Christian KÃnig (2018-11-12 10:16:01)
> Am 09.11.18 um 23:26 schrieb Eric Anholt:
> 
>     Eric Anholt <eric@xxxxxxxxxx> writes:
> 
> 
>         [ Unknown signature status ]
>         zhoucm1 <zhoucm1@xxxxxxx> writes:
> 
> 
>             On 2018å11æ09æ 00:52, Christian KÃnig wrote:
> 
>                 Am 08.11.18 um 17:07 schrieb Koenig, Christian:
> 
>                     Am 08.11.18 um 17:04 schrieb Eric Anholt:
> 
>                         Daniel suggested I submit this, since we're still seeing regressions
>                         from it. This is a revert to before 48197bc564c7 ("drm: add syncobj
>                         timeline support v9") and its followon fixes.
> 
>                     This is a harmless false positive from lockdep, Chouming and I are
>                     already working on a fix.
> 
>                 On the other hand we had enough trouble with that patch, so if it
>                 really bothers you feel free to add my Acked-by: Christian KÃnig
>                 <christian.koenig@xxxxxxx> and push it.
> 
>             NAK, please no, I don't think this needed, the Warning totally isn't
>             related to syncobj timeline, but fence-array implementation flaw, just
>             exposed by syncobj.
>             In addition, Christian already has a fix for this Warning, I've tested.
>             Please Christian send to public review.
> 
>         I backed out my revert of #2 (#1 still necessary) after adding the
>         lockdep regression fix, and now my CTS run got oomkilled after just a
>         few hours, with these notable lines in the unreclaimable slab info list:
> 
>         [ 6314.373099] drm_sched_fence        69095KB      69095KB
>         [ 6314.373653] kmemleak_object       428249KB     428384KB
>         [ 6314.373736] kmalloc-262144           256KB        256KB
>         [ 6314.373743] kmalloc-131072           128KB        128KB
>         [ 6314.373750] kmalloc-65536             64KB         64KB
>         [ 6314.373756] kmalloc-32768           1472KB       1728KB
>         [ 6314.373763] kmalloc-16384             64KB         64KB
>         [ 6314.373770] kmalloc-8192             208KB        208KB
>         [ 6314.373778] kmalloc-4096            2408KB       2408KB
>         [ 6314.373784] kmalloc-2048             288KB        336KB
>         [ 6314.373792] kmalloc-1024            1457KB       1512KB
>         [ 6314.373800] kmalloc-512              854KB       1048KB
>         [ 6314.373808] kmalloc-256              188KB        268KB
>         [ 6314.373817] kmalloc-192            69141KB      69142KB
>         [ 6314.373824] kmalloc-64             47703KB      47704KB
>         [ 6314.373886] kmalloc-128            46396KB      46396KB
>         [ 6314.373894] kmem_cache                31KB         35KB
> 
>         No results from kmemleak, though.
> 
>     OK, it looks like the #2 revert probably isn't related to the OOM issue.
>     Running a single job on otherwise unused DRM, watching /proc/slabinfo
>     every second for drm_sched_fence, I get:
> 
>     drm_sched_fence        0      0    192   21    1 : tunables   32   16    8 : slabdata      0      0      0 : globalstat       0      0     0    0    0    0    0    0    0 : cpustat      0      0      0      0
>     drm_sched_fence       16     21    192   21    1 : tunables   32   16    8 : slabdata      1      1      0 : globalstat      16     16     1    0    0    0    0    0    0 : cpustat      5      1      6      0
>     drm_sched_fence       13     21    192   21    1 : tunables   32   16    8 : slabdata      1      1      0 : globalstat      16     16     1    0    0    0    0    0    0 : cpustat      5      1      6      0
>     drm_sched_fence        6     21    192   21    1 : tunables   32   16    8 : slabdata      1      1      0 : globalstat      16     16     1    0    0    0    0    0    0 : cpustat      5      1      6      0
>     drm_sched_fence        4     21    192   21    1 : tunables   32   16    8 : slabdata      1      1      0 : globalstat      16     16     1    0    0    0    0    0    0 : cpustat      5      1      6      0
>     drm_sched_fence        2     21    192   21    1 : tunables   32   16    8 : slabdata      1      1      0 : globalstat      16     16     1    0    0    0    0    0    0 : cpustat      5      1      6      0
>     drm_sched_fence        0     21    192   21    1 : tunables   32   16    8 : slabdata      0      1      0 : globalstat      16     16     1    0    0    0    0    0    0 : cpustat      5      1      6      0
> 
>     So we generate a ton of fences, and I guess free them slowly because of
>     RCU?  And presumably kmemleak was sucking up lots of memory because of
>     how many of these objects were laying around.
> 
> 
> That is certainly possible. Another possibility is that we don't drop the
> reference in dma-fence-array early enough.
> 
> E.g. the dma-fence-array will keep the reference to its fences until it is
> destroyed, which is a bit late when you chain multiple dma-fence-array objects
> together.
> 
> David can you take a look at this and propose a fix? That would probably be
> good to have fixed in dma-fence-array separately to the timeline work.

Note that drm_syncobj_replace_fence() leaks any existing fence for
!timeline syncobjs. Which coupled with the linear search ends up with
a severe regression in both time and memory.
-Chris