Re: bo problem, file-max limit reached within two days

[Chris Wilson <chris@chris-wilson.co.uk>]

On Thu, 12 Apr 2012 09:55:11 +0200, Knut Petersen <Knut_Petersen@t-online.de> wrote:
> Am 02.04.2012 10:44, schrieb Chris Wilson:
> > On Mon, 02 Apr 2012 10:36:44 +0200, Knut Petersen<Knut_Petersen@t-online.de>  wrote:
> >> After a few unattended hours Xorg was still running,
> >> but only a terminal window had survived the night :-(
> >>
> >> Xorg: git, a few days old, kernel: 3.3.
> >>
> >> Xorg.0.log: Nothing unusual
> >>
> >> dmesg:
> >>
> >> [156859.078080] [drm:drm_gem_create_mmap_offset] *ERROR* failed to allocate offset for bo 0
> >> [179417.374016] VFS: file-max limit 204863 reached
> >>
> >> I built a new X server and tried kernel 3.2.12 - that does not seem to help as
> >> both the number of open files  and the number of gem objects still grow without
> >> obvious reasons.
> > It's the EFILE that is truly worrying. We have a patch in the queue to
> > help ease the ENOSPC issue, but the EFILE implies a bo reference leak.
> > And that I have not found yet.
> >
> > Happy hunting,
> > -Chris
> >
> 
> Well, it takes more than two days to trigger the EFILE limit here, but yesterday
> the ENOSPC bo problem appeared 6.5 hours after booting the system with current
> Xorg git and kernel 3.3.1. I wonder if this is one problem or if we face two independent bugs.

Where there is one, there is likely to be a cluster. However both ENOSPC
and EFILE can be explained by a bo leak, just differing by how they were
used prior to the leak. Eventually we run out of fd handles on the
system (EFILE). If we mapped those into our process space before the
leak then we hit the much smaller address space limit (ENOSPC).
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre