From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail.innovsys.com (smtp.innovsys.com [66.115.232.196]) by ozlabs.org (Postfix) with ESMTP id EE1B367C06 for ; Mon, 16 Oct 2006 23:52:30 +1000 (EST) MIME-Version: 1.0 Content-Type: multipart/alternative; boundary="----_=_NextPart_001_01C6F12A.517511E3" Subject: RE: preempt crash in 2.6.18 Date: Mon, 16 Oct 2006 08:50:40 -0500 Message-ID: References: <20061014012105.6ca4ff05@localhost.localdomain><1160781278.4792.279.camel@localhost.localdomain> <20061015132058.01c68605@localhost.localdomain> From: "Rune Torgersen" To: "Vitaly Bordug" , "Benjamin Herrenschmidt" Cc: linuxppc-dev@ozlabs.org List-Id: Linux on PowerPC Developers Mail List List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , This is a multi-part message in MIME format. ------_=_NextPart_001_01C6F12A.517511E3 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable -----Original Message----- From: Vitaly Bordug [mailto:vbordug@ru.mvista.com] =20 On Sat, 14 Oct 2006 09:14:38 +1000 Benjamin Herrenschmidt wrote: > On Sat, 2006-10-14 at 01:21 +0400, Vitaly Bordug wrote: > > On Fri, 13 Oct 2006 09:16:52 -0500 > > Rune Torgersen wrote: > >=20 > > >=20 > > > This is from a Freescale 8265 running 2.6.18 > > > Anyone have any idea what this is and how to fix it? > > >=20 > > Yes. Actually, I recall alike thing and used to fix with proper > > spin-locking if the cascade PCI irq, which apparently does do_irq > > that seems to confuse preempt counters. > >=20 > > Just a pure guess. >=20 > How so ? Cascades shouldn't do do_IRQ with the new irq code anyway > unless this is still arch/ppc, they should do either __do_IRQ or > better, generic_handle_irq(). >=20 To clarify - I was about __do_IRQ and arch/ppc.=20 Anyway, cascade for PCI is the first place I'd give a look. In our case it turned out to be in the TIPC network code. post_timeout = was called and didn't unlock a lock. The patch had been in the TIPC tree for abpout 2 days when we saw the = problem. -- Sincerely, -Vitaly ------_=_NextPart_001_01C6F12A.517511E3 Content-Type: text/html; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable RE: preempt crash in 2.6.18

-----Original Message-----
From: Vitaly Bordug [mailto:vbordug@ru.mvista.com]
On Sat, 14 Oct 2006 09:14:38 +1000
Benjamin Herrenschmidt wrote:

> On Sat, 2006-10-14 at 01:21 +0400, Vitaly Bordug wrote:
> > On Fri, 13 Oct 2006 09:16:52 -0500
> > Rune Torgersen wrote:
> >
> > >
> > > This is from a Freescale 8265 running 2.6.18
> > > Anyone have any idea what this is and how to fix it?
> > >
> > Yes. Actually, I recall alike thing and used to fix with = proper
> > spin-locking if the cascade PCI irq, which apparently does = do_irq
> > that seems to confuse preempt counters.
> >
> > Just a pure guess.
>
> How so ? Cascades shouldn't do do_IRQ with the new irq code = anyway
> unless this is still arch/ppc, they should do either __do_IRQ = or
> better, generic_handle_irq().
>
To clarify - I was about __do_IRQ and arch/ppc.
Anyway, cascade for PCI is the first place I'd give a look.

In our case it turned out to be in the TIPC network code. post_timeout = was called and didn't unlock a lock.
The patch had been in the TIPC tree for abpout 2 days when we saw the = problem.


--
Sincerely,

-Vitaly

------_=_NextPart_001_01C6F12A.517511E3--