Re: [jlayton:mgtime 5/13] inode.c:undefined reference to `__invalid_cmpxchg_size'

[Jeff Layton <jlayton@kernel.org>]

On Tue, 2024-07-09 at 17:07 +0200, Arnd Bergmann wrote:
> On Tue, Jul 9, 2024, at 16:23, Jeff Layton wrote:
> > On Tue, 2024-07-09 at 16:16 +0200, Arnd Bergmann wrote:
> > > On Tue, Jul 9, 2024, at 15:45, Geert Uytterhoeven wrote:
> > 
> > I think the simplest solution is to make the floor value I'm
> > tracking
> > be an atomic64_t. That looks like it should smooth over the
> > differences
> > between arches. I'm testing a patch to do that now.
> 
> Yes, atomic64_t should work, but be careful about using this
> in a fast path since it can turn into a global spinlock
> in lib/atomic64.c on architectures that don't support it
> natively.
> 
> I'm still reading through the rest of your series, but
> it appears that you pass the time value into 
> ktime_to_timespec64() directly afterwards, so I guess
> that is already a fairly large overhead on 32-bit
> architectures and an extra spinlock doesn't hurt too
> much.
> 

Thanks Arnd.

The context for this is generally a write or other change to an inode,
so I too am hoping the overhead won't be too bad. It does take great
pains to avoid changing the ctime_floor value whenever possible.


> Two more things I noticed in your patch:
> 
> - smp_load_acquire() on a 64-bit variable seems problematic
>   as well, maybe this needs a spinlock on 32-bit
>   architectures?
> 

That should go away with the conversion of ctime_floor to atomic64_t.
AFAICT, arches that don't have native a 64-bit cmpxchg op usually
emulate that with hashed spinlocks. 

> - for the coarse_ctime function, I think you should be
>   able to avoid the conversion to timespec by just calling
>   ktime_get_coarse_real_ts64() again instead of converting
>   monotonic to real and then to timespec.
> 

Note that we might get different values for the coarse timestamps, but
if we do then the second fetch will just be a little later (which is
OK). I'll plan to make this change.


> - inode_set_ctime_current() seems to now store a fine-grained
>   timespec in the inode even for the !is_mgtime case, skipping
>   the timestamp_truncate() step. This appears to potentially
>   leak a non-truncated value to userspace, which would be
>   inconsistent with the value read back from disk.

Oof, you're right. I'll fix that up for the next version.

Thanks for the review!
-- 
Jeff Layton <jlayton@kernel.org>