Re: 2.4.20pre5aa2

[Stephen Lord <lord@sgi.com>]

On Thu, 2002-09-12 at 19:23, Andrea Arcangeli wrote:
> 
> that seems a bug in xfs, it BUG() if vmap fails, it must not BUG(), it
> must return -ENOMEM to userspace instead, or it can try to recollect and
> release some of the other vmalloced entries. Most probably you run into
> an address space shortage, not a real ram shortage, so to workaround it
> you can recompile with CONFIG_2G and it'll probably work, also dropping
> the gap page in vmalloc may help workaround it (there's no config option
> for it though). It could be also a vmap leak, maybe a missing vfree,
> just some idea.
> 

We hold vmalloced space for very short periods of time, in fact
filesystem recovery and large extended attributes are the only
cases. In this case we should be attempting to remap 2 pages
together. The only way out of this would be to fail the whole
mount at this point. I suspect a leak elsewhere.

Samuel, when you mounted xfs and it oopsed, was it shortly after bootup?
Also, how far did your dbench run get before it hung? I tried the
kernel, but I paniced during startup - then I realized I did not 
apply the patch to fix the xfs/scheduler interactions first.

How much memory is in the machine by the way? And Andrea, is the
vmalloc space size reduced in the 3G user space configuration?

Steve

-- 

Steve Lord                                      voice: +1-651-683-3511
Principal Engineer, Filesystem Software         email: lord@sgi.com