Segher Boessenkool <seg...@kernel.crashing.org> writes:
> Hi!
>
> On Fri, Dec 13, 2019 at 11:07:55PM +1100, Michael Ellerman wrote:
>> I tried this:
>>
>> > @@ -295,6 +296,23 @@ void __write_once_size(volatile void *p, void *res,
>> > int size)
>> > */
>> > #define READ_ONCE_NOCHECK(x) __READ_ONCE(x, 0)
>> >
>> > +#else /* GCC_VERSION < 40800 */
>> > +
>> > +#define READ_ONCE_NOCHECK(x)
>> > \
>> > +({
>> > \
>> > + typeof(x) __x = *(volatile typeof(x))&(x); \
>>
>> Didn't compile, needed:
>>
>> typeof(x) __x = *(volatile typeof(&x))&(x); \
>>
>>
>> > + smp_read_barrier_depends(); \
>> > + __x;
>> > +})
>>
>>
>> And that works for me. No extra stack check stuff.
>>
>> I guess the question is does that version of READ_ONCE() implement the
>> read once semantics. Do we have a good way to test that?
>>
>> The only differences are because of the early return in the generic
>> test_and_set_bit_lock():
>
> No, there is another difference:
>
>> 30 ld r10,560(r9)
>> 31 std r10,104(r1)
>> 32 ld r10,104(r1)
>> 33 andi. r10,r10,1
>> 34 bne <ext4_resize_begin_generic+0xd0> 29 bne
>> <ext4_resize_begin_ppc+0xd0>
>
> The stack var is volatile, so it is read back immediately after writing
> it, here. This is a bad idea for performance, in general.
Argh, yuck. Thanks, I shouldn't try to read asm listings at 11pm.
So that just confirms what Will was saying further up the thread about
the volatile pointer, rather than READ_ONCE() per se.
cheers
