Re: [patch 2.6.28-rc2] at91_mci: workaround lockdep

[Peter Zijlstra <peterz@infradead.org>]

On Mon, 2008-11-17 at 10:28 +0100, Nicolas Ferre wrote:
> David Brownell :
> > From: David Brownell <dbrownell@users.sourceforge.net>
> > 
> > Lockdep reported a problem in the at91_mci driver ... in this case, the
> > issue is with lockdep, not with the driver.  A trimmed stack dump, from
> > trying to boot with root on MMC, shows:
> > 
> >   WARNING: at kernel/lockdep.c:2195 trace_hardirqs_on_caller+0xf4/0x170()
> >   Modules linked in:
> >   [<c005bc98>] (trace_hardirqs_on+0x0/0x18) from [<c0213bf4>] (_spin_unlock_irq+0x2c/0x3c)
> >   [<c0213bc8>] (_spin_unlock_irq+0x0/0x3c) from [<c0029a88>] (flush_dcache_page+0x114/0x144)
> >   [<c0029974>] (flush_dcache_page+0x0/0x144) from [<c019b034>] (at91_mci_irq+0x150/0x414)
> >   [<c019aee4>] (at91_mci_irq+0x0/0x414) from [<c0066c5c>] (handle_IRQ_event+0x2c/0x6c)
> >   [<c0066c30>] (handle_IRQ_event+0x0/0x6c) from [<c0068a60>] (handle_level_irq+0x108/0x124)
> >   [<c0068958>] (handle_level_irq+0x0/0x124) from [<c0022064>] (__exception_text_start+0x64/0x90)
> > 
> > When __flush_dcache_aliases() returns -- inlined into flush_dcache_page(),
> > above -- it re-enables IRQs ... since that evidently may only be called with
> > IRQs enabled.  That's OK since the (unshared) IRQ handler doesn't ask for IRQs
> > to be disabled.   Except ... that lockdep went and disabled them, then went on
> > to complains about the breakage *it* caused!
> > 
> > Workaround: depend on LOCKDEP=n ... and for paranoia, disable IRQF_SHARED
> > for this interrupt.  (At the hardware level, this is dedicated to MCI, so
> > there's never a need for multiple handlers.)
> > 
> > Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
> 
> Acked-by: Nicolas Ferre <nicolas.ferre@atmel.com>

Nacked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>

> 
> > ---
> > An alternate fix might be to have ARM dcache mmap locking save and restore
> > the IRQ flags ... but there are several such "can't run with IRQs disabled"
> > restrictions scattered through that code, with no rationales commented.  Or
> > stop lockdep from mucking with IRQ flags; probably not any time soon!
> > 
> > So for now, this seems to be the best "solution" available...
> > 
> >  drivers/mmc/host/Kconfig    |    1 +
> >  drivers/mmc/host/at91_mci.c |    3 +--
> >  2 files changed, 2 insertions(+), 2 deletions(-)
> > 
> > --- a/drivers/mmc/host/Kconfig
> > +++ b/drivers/mmc/host/Kconfig
> > @@ -99,6 +99,7 @@ config MMC_AU1X
> >  config MMC_AT91
> >  	tristate "AT91 SD/MMC Card Interface support"
> >  	depends on ARCH_AT91
> > +	depends on LOCKDEP=n
> >  	help
> >  	  This selects the AT91 MCI controller.
> >  
> > --- a/drivers/mmc/host/at91_mci.c
> > +++ b/drivers/mmc/host/at91_mci.c
> > @@ -1081,8 +1081,7 @@ static int __init at91_mci_probe(struct 
> >  	 * Allocate the MCI interrupt
> >  	 */
> >  	host->irq = platform_get_irq(pdev, 0);
> > -	ret = request_irq(host->irq, at91_mci_irq, IRQF_SHARED,
> > -			mmc_hostname(mmc), host);
> > +	ret = request_irq(host->irq, at91_mci_irq, 0, mmc_hostname(mmc), host);
> >  	if (ret) {
> >  		dev_dbg(&pdev->dev, "request MCI interrupt failed\n");
> >  		goto fail0;
> > 
> 
> Thanks, regards,