Hi Greg,
On Mon, 2018-07-09 at 07:48 +0200, Greg KH wrote:
> On Mon, Jul 09, 2018 at 07:44:44AM +0300, Alexey Brodkin wrote:
> > Depending on ABI "long long" type of a particular 32-bit CPU
> > might be aligned by either word (32-bits) or double word (64-bits).
> > Make sure "data" is really 64-bit aligned for any 32-bit CPU.
> >
> > At least for 32-bit ARC cores ABI requires "long long" types
> > to be aligned by normal 32-bit word. This makes "data" field aligned to
> > 12 bytes. Which is still OK as long as we use 32-bit data only.
> >
> > But once we want to use native atomic64_t type (i.e. when we use special
> > instructions LLOCKD/SCONDD for accessing 64-bit data) we easily hit
> > misaligned access exception.
>
> So is this something you hit today? If not, why is this needed for
> stable kernels?
Indeed we hit that problem recently when Etnaviv driver was switched to
DRM GPU scheduler, see
commit e93b6deeb45a ("drm/etnaviv: hook up DRM GPU scheduler").
The most important part of DRM GPU scheduler is "job_id_count" member of
"drm_gpu_scheduler" structure of type "atomic64_t". This structure is put
in a buffer allocated by devm_kzalloc() and if "job_id_count" is not 64-bit
aligned atomic instruction fails with an exception.
As for stable requirements - mentioned commit was a part of 4.17 kernel
which broke GPU driver for one of our HSDK board so I guess back-porting
to 4.17 is a no-brainer.
Now given a nature of the problem at hand I may expect other breakages
whenever another driver/fs etc that use "atomic64_t" gets enabled.
> > That's because even on CPUs capable of non-aligned data access LL/SC
> > instructions require strict alignment.
>
> Are you going to hit this code with all types of structures?
If there're other cases which lead to 4-byte aligned "atomic64_t" variables
there will be a problem as well but it's quite hard to predict those cases.
That said if we manage to reproduce more similar issues there will be more
patches with fixes :)
> What happens when you do have an unaligned access?
Atomic instructions are a bit special as compared to normal loads and stores.
Even if normal loads and stores may deal with unaligned data atomic instructions
still require data to be aligned because it's hard to manage atomic value that
spans through multiple cache lines or even MMU pages. And hardware just
raises an alignment fault exception.
And that's not something special for ARC, I guess all CPUs are the same in
that regard, see here's an extract from ARM(r) Architecture Reference
Manual ARMv7-A and ARMv7-R edition: https://lkml.org/lkml/2017/12/5/440
