Re: [PATCH] Revert "x86: optimize page faults like all other achitectures and kill notifier cruft"

From: Ingo Molnar <mingo@elte.hu>
To: David Miller <davem@davemloft.net>
Cc: hch@lst.de, airlied@linux.ie, akpm@linux-foundation.org,
	torvalds@linux-foundation.org, ak@suse.de, pq@iki.fi,
	jbeulich@novell.com, linux-kernel@vger.kernel.org,
	Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>,
	Ananth N Mavinakayanahalli <ananth@in.ibm.com>,
	Masami Hiramatsu <mhiramat@redhat.com>
Subject: Re: [PATCH] Revert "x86: optimize page faults like all other achitectures and kill notifier cruft"
Date: Wed, 9 Jan 2008 11:41:49 +0100	[thread overview]
Message-ID: <20080109104149.GB8041@elte.hu> (raw)
In-Reply-To: <20080108.232311.79648776.davem@davemloft.net>

(kprobes folks Cc:-ed)

* David Miller <davem@davemloft.net> wrote:

> From: Christoph Hellwig <hch@lst.de>
> Date: Wed, 9 Jan 2008 08:19:45 +0100
> 
> > On Wed, Jan 09, 2008 at 03:55:20AM +0000, Dave Airlie wrote:
> > > now because Linus said send him a patch to revert regressions rather than 
> > > just complain,
> > 
> > this is not a regression by any definition.  You were abusing 
> > exported symbols for out of tree junk, so you'll lose.
> 
> And furthermore, they don't even need it, use a kprobe.

i agree. There a few practical complication on x86: the do_page_fault() 
function is currently excluded from kprobe probing, for recursion 
reasons. handle_mm_fault() can be probed OTOH - but that does not catch 
vmalloc()-ed faults. The middle of do_page_fault() [line 348] should 
work better [the point after notify_page_fault()] - but it's usually 
more fragile to insert probes to such middle-of-the-function places.

So probing pagefaults is not as easy as it should/could be. We should 
put a practical NOP marker to around line 348, to make it easier (and 
faster) for systemtap to probe there.

(__kprobes is a highly confusing newspeak name btw - it should be 
__noprobe instead.)

	Ingo