Hi Geert,
On Mon, 2018-07-09 at 09:52 +0200, Geert Uytterhoeven wrote:
> Hi Alexey,
>
> On Mon, Jul 9, 2018 at 9:22 AM Alexey Brodkin
> <Alexey.Brodkin@xxxxxxxxxxxx> wrote:
> > On Mon, 2018-07-09 at 09:07 +0200, Geert Uytterhoeven wrote:
> > > On Mon, Jul 9, 2018 at 7:49 AM Greg KH <greg@xxxxxxxxx> 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).
> > >
> > > Or even 16-bit (on e.g. m68k).
> >
> > Indeed, thanks for the note!
> > Will add this in my v3.
>
> Note that in this particular case, the field will probably always be
> aligned to 4 bytes,
> as the struct will be allocated using *alloc().
Well I think maybe it worth mentioning only 32-bit and 64-bit alignments in
this particular case because as it was mentioned initially in the comment there're
3 pointers before "data" field and for 32-bit machines I guess we may safely
assume that a pointer size is 32-bit.
This leaves us only one problematic situation 32-bit instead of 64-bit alignment.
> > For ARC I'd like this fix to be back-ported starting
> > from v4.8 where we added support of "native" atomic64_t, see commit
> > ce6365270ecd (" ARCv2: Implement atomic64 based on LLOCKD/SCONDD instructions").
> >
> > What about m68k, do you have any preference of earliest kernel version
> > where this fix might be useful?
>
> Given m68k is 32-bit, it will access atomic64_t variables while
> holding a spinlock,
> so it should still be safe without this change.
Well ARC and ARMv7 are 32-bit machines as well still both have
a special instructions to read/write 64-bit data.
> Not to mention no one will try etnaviv on m68k ;-)
See it was just a trigger case but other GPUs that use or will use
DRM GPU scheduler are good candidates to it that problem and not to
mention some other users of atomic64_t.
But from you above comment I may deduce that m68k doesn't have
instructions for 64-bit data; in that case there's no reason to worry
at least about struct devres_node :)
-Alexey
