On Tue, 2018-08-14 at 13:42 +0000, Vineet Gupta wrote: > On 08/11/2018 09:09 AM, Eugeniy Paltsev wrote: > > Current implementation of 'atomic64_add_unless' function > > (and hence 'atomic64_inc_not_zero') return incorrect value > > if lover 32 bits of compared 64-bit number are equal and > > higher 32 bits aren't. > > > > For in following example atomic64_add_unless must return '1' > > but it actually returns '0': > > --------->8--------- > > atomic64_t val = ATOMIC64_INIT(0x4444000000000000LL); > > int ret = atomic64_add_unless(&val, 1LL, 0LL) > > --------->8--------- > > > > This happens because we write '0' to returned variable regardless > > of higher 32 bits comparison result. > > > > So fix it. > > > > NOTE: > > this change was tested with atomic64_test. > > > > Signed-off-by: Eugeniy Paltsev <Eugeniy.Paltsev at synopsys.com> > > LGTM. Curious, was this from code review or did u actually run into this ? I've accidentally run into this when I played with atomic64_* functions trying to implement hack to automatically align LL64/SC64 data for atomic 64-bit operations on ARC to avoid problems like: https://www.mail-archive.com/linux-snps-arc at lists.infradead.org/msg03791.html > Thx, > -Vineet > > > --- > > arch/arc/include/asm/atomic.h | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > diff --git a/arch/arc/include/asm/atomic.h b/arch/arc/include/asm/atomic.h > > index 11859287c52a..e840cb1763b2 100644 > > --- a/arch/arc/include/asm/atomic.h > > +++ b/arch/arc/include/asm/atomic.h > > @@ -578,11 +578,11 @@ static inline int atomic64_add_unless(atomic64_t *v, long long a, long long u) > > > > __asm__ __volatile__( > > "1: llockd %0, [%2] \n" > > - " mov %1, 1 \n" > > " brne %L0, %L4, 2f # continue to add since v != u \n" > > " breq.d %H0, %H4, 3f # return since v == u \n" > > " mov %1, 0 \n" > > "2: \n" > > + " mov %1, 1 \n" > > " add.f %L0, %L0, %L3 \n" > > " adc %H0, %H0, %H3 \n" > > " scondd %0, [%2] \n" > > -- Eugeniy Paltsev
