Re: [PATCH] Only print kernel debug information for OOMs caused by kernel allocations

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 28 Jan 2008 07:10:07 +0100 Andi Kleen <ak@suse.de> wrote:

> On Monday 28 January 2008 06:52, Andrew Morton wrote:
> > On Wed, 16 Jan 2008 23:24:21 +0100 Andi Kleen <ak@suse.de> wrote:
> > > I recently suffered an 20+ minutes oom thrash disk to death and computer
> > > completely unresponsive situation on my desktop when some user program
> > > decided to grab all memory. It eventually recovered, but left lots
> > > of ugly and imho misleading messages in the kernel log. here's a minor
> > > improvement
> 
> As a followup this was with swap over dm crypt. I've recently heard
> about other people having trouble with this too so this setup seems to trigger
> something bad in the VM.

Where's the backtrace and show_mem() output? :)

> > That information is useful for working out why a userspace allocation
> > attempt failed.  If we don't print it, and the application gets killed and
> > thus frees a lot of memory, we will just never know why the allocation
> > failed.
> 
> But it's basically only either page fault (direct or indirect) or write et.al.
> who do these page cache allocations. Do you really think it is that important
> to distingush these cases individually? In 95+% of all cases it should
> be a standard user page fault which always has the same backtrace.

Sure, the backtrace isn't very important.  The show_mem() output is vital.

> To figure out why the application really oom'ed for those you would
> need a user level backtrace, but the message doesn't supply that one anyways. 
> 
> All other cases will still print the full backtrace so if some kernel 
> subsystem runs amok it should be still possible to diagnose it.
> 

We need the show_mem() output to see where all the memory went, and to see
what state page reclaim is in.

> 
> >
> > >  struct page *__page_cache_alloc(gfp_t gfp)
> > >  {
> > > +	struct task_struct *me = current;
> > > +	unsigned old = (~me->flags) & PF_USER_ALLOC;
> > > +	struct page *p;
> > > +
> > > +	me->flags |= PF_USER_ALLOC;
> > >  	if (cpuset_do_page_mem_spread()) {
> > >  		int n = cpuset_mem_spread_node();
> > > -		return alloc_pages_node(n, gfp, 0);
> > > -	}
> > > -	return alloc_pages(gfp, 0);
> > > +		p = alloc_pages_node(n, gfp, 0);
> > > +	} else
> > > +		p = alloc_pages(gfp, 0);
> > > +	/* Clear USER_ALLOC if it wasn't set originally */
> > > +	me->flags ^= old;
> > > +	return p;
> > >  }
> >
> > That's appreciable amount of new overhead for at best a fairly marginal
> > benefit.  Perhaps __GFP_USER could be [re|ab]used.
> 
> It's a few non atomic bit operations. You really think that is considerable
> overhead? Also all should be cache hot already. My guess is that even with the 
> additional function call it's < 10 cycles more.

Plus an additional function call.  On the already-deep page allocation
path, I might add.

> > Alternatively: if we've printed the diagnostic on behalf of this process
> > and then decided to kill it, set some flag to prevent us from printing it
> > again.
> 
> Do you really think that would help?  I thought these messages came usually
> from different processes.

Dunno.