Re: [PATCH 1/1] Fix: trace sched switch start/stop racy updates
From: Paul E. McKenney
Date: Tue Aug 20 2019 - 16:30:27 EST
On Tue, Aug 20, 2019 at 03:56:12PM +0200, Peter Zijlstra wrote:
> On Sat, Aug 17, 2019 at 01:08:02AM -0700, Linus Torvalds wrote:
>
> > The data tearing issue is almost a non-issue. We're not going to add
> > WRITE_ONCE() to these kinds of places for no good reason.
>
> Paulmck actually has an example of that somewhere; ISTR that particular
> case actually got fixed by GCC, but I'd really _love_ for some compiler
> people (both GCC and LLVM) to state that their respective compilers will
> not do load/store tearing for machine word sized load/stores.
I do very much recall such an example, but I am now unable to either
find it or reproduce it. :-/
If I cannot turn it up in a few days, I will ask the LWN editors to
make appropriate changes to the "Who is afraid" article.
> Without this written guarantee (which supposedly was in older GCC
> manuals but has since gone missing), I'm loathe to rely on it.
>
> Yes, it is very rare, but it is a massive royal pain to debug if/when it
> does do happen.
But from what I can see, Linus is OK with use of WRITE_ONCE() for data
races on any variable for which there is at least one READ_ONCE().
So we can still use WRITE_ONCE() as we would like in our own code.
Yes, you or I might be hit by someone else's omission of WRITE_ONCE(),
it is better than the proverbial kick in the teeth.
Of course, if anyone knows of a compiler/architecture combination that
really does tear stores of 32-bit constants, please do not keep it
a secret! After all, it would be good to get that addressed easily
starting now rather than after a difficult and painful series of
debugging sessions.
Thanx, Paul