Re: [PATCH 3/5] asm-generic, kcsan: Add KCSAN instrumentation for bitops

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, 21 Jan 2020 at 15:47, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Tue, Jan 21, 2020 at 06:21:09AM -0800, Paul E. McKenney wrote:
> > On Tue, Jan 21, 2020 at 10:15:01AM +0100, Peter Zijlstra wrote:
> > > On Mon, Jan 20, 2020 at 12:23:59PM -0800, Paul E. McKenney wrote:
> > > > We also don't have __atomic_read() and __atomic_set(), yet atomic_read()
> > > > and atomic_set() are considered to be non-racy, right?
> > >
> > > What is racy? :-) You can make data races with atomic_{read,set}() just
> > > fine.
> >
> > Like "fairness", lots of definitions of "racy".  ;-)
> >
> > > Anyway, traditionally we call the read-modify-write stuff atomic, not
> > > the trivial load-store stuff. The only reason we care about the
> > > load-store stuff in the first place is because C compilers are shit.
> > >
> > > atomic_read() / test_bit() are just a load, all we need is the C
> > > compiler not to be an ass and split it. Yes, we've invented the term
> > > single-copy atomicity for that, but that doesn't make it more or less of
> > > a load.
> > >
> > > And exactly because it is just a load, there is no __test_bit(), which
> > > would be the exact same load.
> >
> > Very good!  Shouldn't KCSAN then define test_bit() as non-racy just as
> > for atomic_read()?
>
> Sure it does; but my comment was aimed at the gripe that test_bit()
> lives in the non-atomic bitops header. That is arguably entirely
> correct.

I will also point out that test_bit() is listed in
Documentation/atomic_bitops.txt.  What I gather from
atomic_bitops.txt, is that the semantics of test_bit() is simply an
unordered atomic operation: the interface promises that the load will
be executed as one indivisible step, i.e. (single-copy) atomically
(after compiler optimizations etc.).

Although at this point probably not too important, I checked Alpha's
implementation of test_bit(), and there is no
smp_read_barrier_depends(). Is it safe to say that test_bit() should
then be weaker in terms of ordering than READ_ONCE()?

My assumption was that test_bit() is as strong as READ_ONCE().

Thanks,
-- Marco



[Index of Archives]     [Linux Kernel]     [Kernel Newbies]     [x86 Platform Driver]     [Netdev]     [Linux Wireless]     [Netfilter]     [Bugtraq]     [Linux Filesystems]     [Yosemite Discussion]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Device Mapper]

  Powered by Linux