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

[Jeff Garzik <jgarzik@pobox.com>]

Andrew Morton wrote:
> On Tue, 26 Feb 2008 21:49:43 +0100 Ingo Molnar <mingo@elte.hu> wrote:
>> i.e. to introduce a kmap_atomic_irqsave(...,flags) and 
>> kunmap_atomic_irqrestore() API variant. _That_ then could be mapped by 
>> -rt to a non-irq disabling thing.

> Sure.  But iirc we haven't had a need for this before.  Which is a bit odd.

We have such a need -- we just always paper over it with 
spin_lock_irqsave() or local_irq_save() because those are "the rules." 
grep around :)

See ata_pio_sector() in libata-core.c, where we do:

>         if (PageHighMem(page)) {
>                 unsigned long flags;
> 
>                 /* FIXME: use a bounce buffer */
>                 local_irq_save(flags);
>                 buf = kmap_atomic(page, KM_IRQ0);
> 
>                 /* do the actual data transfer */
>                 ap->ops->data_xfer(qc->dev, buf + offset, qc->sect_size, do_write);
> 
>                 kunmap_atomic(buf, KM_IRQ0);
>                 local_irq_restore(flags);
>         } else {
>                 buf = page_address(page);
>                 ap->ops->data_xfer(qc->dev, buf + offset, qc->sect_size, do_write);
>         }

This is a core non-DMA data transfer routine.  Given the high cost of 
disabling interrupts during the relatively-slow PIO data transfer, we 
added extra logic to only disable interrupts for the highmem case.

The code in libata-scsi generally only deals with a single buffer, on 
average less than 128 bytes in length.  So the memcpy() isn't really 
that expensive in the case being discussed -- unlike the 
ata_pio_sector() case, where the cost is very, very high.

Aside: The "FIXME: use bounce buffer" comment above indicates the more 
optimal PIO data xfer approach of

	local_irq_save()
	kmap_atomic()
	memcpy into bounce buffer
	kunmap_atomic()
	local_irq_restore()

	/* do slow PIO bitbanging data transfer */
	ap->ops->data_xfer(...)

Regards,

	Jeff