Re: [RFC PATCH v2 2/6] PM / Runtime: introduce pm_runtime_set_memalloc_noio()

[Alan Cox <alan@lxorguk.ukuu.org.uk>]

On Wed, 24 Oct 2012 10:06:36 +0100
"David Laight" <David.Laight@ACULAB.COM> wrote:

> > Looks the problem is worse than above, not only bitfields are affected, the
> > adjacent fields might be involved too, see:
> > 
> >            http://lwn.net/Articles/478657/
> 
> Not mentioned in there is that even with x86/amd64 given
> a struct with the following adjacent fields:
> 	char a;
> 	char b;
> 	char c;
> then foo->b |= 0x80; might do a 32bit RMW cycle.

There are processors that will do this for the char case at least as they
do byte ops by a mix of 32bit ops and rotate.

> This will (well might - but probably does) happen
> if compiled to a 'BTS' instruction.
> The x86 instruction set docs are actually unclear
> as to whether the 32bit cycle might even be misaligned!
> amd64 might do a 64bit cycle (not checked the docs).

Even with a suitably aligned field the compiler is at liberty to generate
things like

	reg = 0x80
	reg |= foo->b
	foo->b = reg;

One reason it's a good idea to use set_bit/test_bit and friends.

Alan

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>