[Intel-gfx] [Bug #12210] 2.6.28-rc8 big regression in VM

[Wu Fengguang]

Hi Lukas,

On Thu, Feb 26, 2009 at 08:06:36PM +0200, Lukas Hejtmanek wrote:
> Hello,
> 
> you may remember that I reported regression in VM. I used you filecache module
> and discovered that the leaks are caused by these items:
> # filecache 1.0
> #      ino       size   cached cached% refcnt state dev          file
>     167302         16       16     100      1    d- 00:08(tmpfs) /drm\040mm\040object\040(deleted)
>     167301         16       16     100      1    d- 00:08(tmpfs) /drm\040mm\040object\040(deleted)
>     167300         16       16     100      1    d- 00:08(tmpfs) /drm\040mm\040object\040(deleted)
>     167299          4        4     100      1    -- 00:08(tmpfs) /drm\040mm\040object\040(deleted)
>     167290          8        8     100      1    -- 00:08(tmpfs) /drm\040mm\040object\040(deleted)
>     167289          8        8     100      1    -- 00:08(tmpfs) /drm\040mm\040object\040(deleted)
>     167288         16       16     100      1    -- 00:08(tmpfs) /drm\040mm\040object\040(deleted)
>     167287         16       16     100      1    d- 00:08(tmpfs) /drm\040mm\040object\040(deleted)
> 
> there are tons of those items:
> cat /proc/filecache  | grep drm | wc -l
> 14224
> 
> Do you have any clues what happens here? Is it a bug in kernel VM system or it
> is more likely a bug in Intel's GEM drm driver? Right now it results in 500MB
> of undropable cache.

Could you provide the full filecache listing(the 'size' field may
provide some info), and the contents in /proc/dri/0/*?

> Also lsof reports many leaked file descriptors:
> lsof | grep drm | wc -l
> 7326

Simply being "deleted" does not mean that they are leaked files.
shmem files are special.

Thanks,
Fengguang