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

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

On Tue, 26 Feb 2008 11:32:24 -0500 Jeff Garzik <jgarzik@pobox.com> wrote:

> Ingo Molnar wrote:
> > * Jeff Garzik <jgarzik@pobox.com> wrote:
> > 
> >> +	unsigned long flags;
> >> +
> >> +	local_irq_save(flags);
> > 
> > hm, couldnt we attach the irq disabling to some spinlock, in a natural 
> > way? Explicit flags fiddling is a PITA once we do things like threaded 
> > irq handlers, -rt, etc.
> 
> Attaching the irq disabling to some spinlock is what would be 
> artificial...  See the ahci.c patch earlier in this thread.  It is taken 
> without spin_lock_irqsave() in the interrupt handler, and there is no 
> reason to disable interrupts for the entirety of the interrupt handler 
> run -- only the part where we call kmap.
> 
> This is only being done to satisfy kmap_atomic's requirements, not libata's.
> 
> I could add a "kmap lock" but that just seems silly.
> 

It's a bit sad to disable interupts across a memset (how big is it?) just
for the small proportion of cases which are accessing a highmem page.

What you could do is to add an `unsigned long *flags' arg to
ata_scsi_rbuf_get() and ata_scsi_rbuf_put(), and then, in
ata_scsi_rbuf_get() do

	if (PageHighmem(page))
		local_irq_disable(*flags);