On Wed, Nov 22, 2023 at 04:47:51AM -0800, Christoph Hellwig wrote:
> On Wed, Nov 22, 2023 at 04:19:40PM +0800, Ming Lei wrote:
> > On Tue, Nov 21, 2023 at 11:53:17PM -0800, Christoph Hellwig wrote:
> > > On Wed, Nov 22, 2023 at 03:45:24PM +0800, Ming Lei wrote:
> > > > All the existed 'bool' flags are not atomic RW, so I think it isn't
> > > > necessary to define 'bd_flags' as 'unsigned long' for replacing them.
> > >
> > > So because the old code wasn't correct we'll never bother? The new
> > > flag and the new placement certainly make this more critical as well.
> >
> > Can you explain why the old code was wrong?
> >
> > 1) ->bd_read_only and ->bd_make_it_fail
> >
> > - set from userspace interface(ioctl or sysfs)
> > - check in IO code path
> >
> > so changing it into atomic bit doesn't make difference from user
> > viewpoint.
>
> >
> > 2) ->bd_write_holder
> >
> > disk->open_mutex is held for read & write this flag
> >
> > 3) ->bd_has_submit_bio
> >
> > This flag is setup as oneshot before adding disk, and check in FS io code
> > path.
>
> On architectures that can't do byte-level atomics all three can corrupt
> each other
Yeah, C/C++ doesn't provide such guarantee, but many modern ARCHs [1]
guarantees that RW on naturally aligned type is atomic.
I verified the point on x86/arm64/ppc64le by the following code, and
all three STOREs are done in single instruction.
struct data {
int b;
char a;
char a2;
char a3;
char a4;
} __attribute__((aligned(8)));
void atomic_test()
{
struct data d;
d.b = 1;
d.a = 2;
d.a3 = 3;
printf("%d %d %d\n", d.b, d.a, d.a3);
}
[1] https://preshing.com/20130618/atomic-vs-non-atomic-operations/
> and even worse bd_partno. Granted that is only alpha these
> days IIRC, but it's still buggy.
bd_has_submit_bio and bd_partno can be thought as read only, and the
two can be corrupted?
bd_dev may have similar trouble with bd_partno for ARCHs which don't
provide atomic RW on naturally aligned int.
Thanks,
Ming
