Re: Query on lock protection in random number driver

[Andi Kleen <andi@firstfloor.org>]

On Fri, Jan 11, 2008 at 04:58:28PM +0530, Nikanth Karthikesan wrote:
> On Fri, 2008-01-11 at 12:12 +0100, Andi Kleen wrote:
> > Nikanth Karthikesan <knikanth@suse.de> writes:
> > >
> > > Also the globals random_read_wakeup_thresh and
> > > random_write_wakeup_thresh are not at all protected by any locks! Why
> > > locks are not needed for these?
> > 
> > Reading variables sizeof <= native word size (32bit or 64bit depending
> > on architecture) is atomic by itself. This is not necessarily
> > guaranteed in ISO-C or POSIX threads, but Linux can assume that.
> 
> Yes, I found that by checking the implementation of atomic_read.
> 
> But I didnt check the implementation of atomic_set before sending the
> mail and assumed assigning to a variable may not be atomic on all arch,
> and because of that, we may be reading a half-written, variable! But
> assigning to an int is also atomic on all arch.

For completeness this both only applies to naturally aligned variables
for portable code.

-Andi