Re: 2.6.24-git: kmap_atomic() WARN_ON()

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

On Mon, 25 Feb 2008 23:01:59 +0100 (CET) Thomas Gleixner <tglx@linutronix.de> wrote:

> On Mon, 25 Feb 2008, Andrew Morton wrote:
> > On Mon, 25 Feb 2008 15:08:35 -0500 Jeff Garzik <jgarzik@pobox.com> wrote:
> > > There are plenty of drivers that do the same thing that ahci does, in 
> > > terms of interrupt handler locking...  and I will definitely push back 
> > > on efforts to convert otherwise-100%-safe spin_lock() into 
> > > spin_lock_irqsave() just to quiet lockdep.
> > > 
> > > Very interesting email, thanks...
> > > 
> > 
> > I suspect this is a bug in my old kmap_atomic debugging patch.  It doesn't
> > know about the implicit irq-disablememnt which interrupt handlers enjoy.  I
> > don't think...
> 
> I suspect here is confusion. The implicit irq-disablement of lockdep
> is actually hiding the warning.
> 
> The code which emits the warning is:
> 
>         if (type == KM_IRQ0 || type == KM_IRQ1 || type == KM_BOUNCE_READ ||
>                         type == KM_BIO_SRC_IRQ || type == KM_BIO_DST_IRQ) {
>                 if (!irqs_disabled()) {
>                         WARN_ON(1);
>                         warn_count--;
>                 }
> 
> It checks for _NOT_ irqs_disabled. The calling code is
> ata_scsi_rbuf_get() which calls with:
> 
>      buf = kmap_atomic(sg_page(sg), KM_IRQ0) + sg->offset;
> 
> This happens with interrupts enabled. So the warning is according to
> the well documented km_type enum and the equally well documented
> highmem debug code correct.
> 
> Bjoern decoded it very well, just Jeff jumped to very interesting
> conclusions.
> 

Ah, OK, yes.  ata is wrong.  It must disable interrupts here.  Otherwise
this CPU could get interrupted by some other device whose handler also uses
KM_IRQ0, resulting in data corruption.