Re: [PATCH 4/8] i386: bitops: Kill volatile-casting of memory addresses

[Jeff Garzik <jeff@garzik.org>]

Linus Torvalds wrote:
> And before we remove that "volatile", we'd better make damn sure that 
> there isn't any driver that does
> 
> 	/* Wait for the command queue to be cleared by DMA */
> 	while (test_bit(...))
> 		;
> 
> or similar.
> 
> Yes, it's annoying, but this is a scary and subtle area. And we sadly 
> _have_ had code that does things like that.


I certainly cannot speak for all the grotty drivers out there, but I've 
never ever seen anything like the above.  I would consider anyone using 
kernel bit operations on DMA memory to be more than a little crazy.  But 
that's just me :)

Usually you will see

	while (1) {
		rmb();
		if (software_index == hardware_index_in_DMA_memory)
			break;
		
		... handle a unit of work ...
	}

Though ISTR being told that even rmb() was not sufficient in all cases 
[nonetheless, that's what I use in net and SATA drivers and email 
recommendations, and people seem happy]

	Jeff