Re: [PATCH 10/9] x86/fpu: Fix 'struct fpu' misalignment on 32-bit kernels
From: Oleg Nesterov
Date: Sat Jun 15 2024 - 06:25:48 EST
On 06/14, Oleg Nesterov wrote:
>
> On 06/13, Ingo Molnar wrote:
> >
> > > --- a/include/linux/sched.h
> > > +++ b/include/linux/sched.h
> > > @@ -1562,7 +1562,7 @@ struct task_struct {
> > > * they are included in the randomized portion of task_struct.
> > > */
> > > randomized_struct_fields_end
> > > -};
> > > +} __attribute__ ((aligned (64)));
>
> I guess __aligned(64) will look a bit better, but this is minor.
...
> But Ingo, it was a shot in the dark ;) I still don't really understand
> what exactly should be aligned. Is it the fpstate->regs member? Or what?
> If yes, perhaps this member needs __aligned(64) too to be safe?
Ah, I didn't notice that fpregs_state->xregs_state has
__attribute__ ((packed, aligned (64))), so everything is clear.
>From your previous email:
- The extra alignment attribute in <linux/sched.h> will affect other
architecture as well, although in practice the alignment of init_task is
not critical, and is very likely at least 32 bytes, probably more.
Still, it's a bit ugly in its current form.
Agreed, but afaics we need to align task_struct only to ensure that
(void *)task + sizeof(*task);
doesn't break the alignment.
So perhaps we can (later) change x86_task_fpu(), fpu_clone(), and
fpu__init_task_struct_size() to use
ALIGN(sizeof(struct task_struct), 64)
and remove the alignment attribute in sched.h?
Or use ARCH_MIN_TASKALIGN == __alignof__(union fpregs_state) which is
also used in fork_init()->kmem_cache_create().
Oleg.