Re: [PATCH v2 1/4] mm: Fix multiple evaluvations of totalram_pages and managed_pages

[Michal Hocko]

On Wed 07-11-18 09:44:00, Vlastimil Babka wrote:
> On 11/7/18 9:20 AM, Michal Hocko wrote:
> > On Tue 06-11-18 21:51:47, Arun KS wrote:
> 
> Hi,
> 
> there's typo in subject: evaluvations -> evaluations.
> 
> However, "fix" is also misleading (more below), so I'd suggest something
> like:
> 
> mm: reference totalram_pages and managed_pages once per function
> 
> >> This patch is in preparation to a later patch which converts totalram_pages
> >> and zone->managed_pages to atomic variables. This patch does not introduce
> >> any functional changes.
> > 
> > I forgot to comment on this one. The patch makes a lot of sense. But I
> > would be little bit more conservative and won't claim "no functional
> > changes". As things stand now multiple reads in the same function are
> > racy (without holding the lock). I do not see any example of an
> > obviously harmful case but claiming the above is too strong of a
> > statement. I would simply go with something like "Please note that
> > re-reading the value might lead to a different value and as such it
> > could lead to unexpected behavior. There are no known bugs as a result
> > of the current code but it is better to prevent from them in principle."
> 
> However, the new code doesn't use READ_ONCE(), so the compiler is free
> to read the value multiple times, and before the patch it was free to
> read it just once, as the variables are not volatile. So strictly
> speaking this is indeed not a functional change (if compiler decides
> differently based on the patch, it's an implementation detail).

Yes, compiler is allowed to optimize this either way without READ_ONCE
but it is allowed to do two reads so claiming no functional change is a
bit problematic. Not that this would be a reason to discuss this in
length...
-- 
Michal Hocko
SUSE Labs