[PATCH] sched/membarrier: Fix redundant load of membarrier_state

[Nysal Jan K.A.]

On Fri, Oct 25, 2024 at 11:29:38AM +1100, Michael Ellerman wrote:
> [To += Mathieu]
> 
> "Nysal Jan K.A." <nysal at linux.ibm.com> writes:
> > From: "Nysal Jan K.A" <nysal at linux.ibm.com>
> >
> > On architectures where ARCH_HAS_SYNC_CORE_BEFORE_USERMODE
> > is not selected, sync_core_before_usermode() is a no-op.
> > In membarrier_mm_sync_core_before_usermode() the compiler does not
> > eliminate redundant branches and the load of mm->membarrier_state
> > for this case as the atomic_read() cannot be optimized away.
> 
> I was wondering if this was caused by powerpc's arch_atomic_read() which
> uses asm volatile.
> 

Yes, that's my understanding as well

> But replacing arch_atomic_read() with READ_ONCE() makes no difference,
> presumably because the compiler still can't see that the READ_ONCE() is
> unnecessary (which is kind of by design).
> 

In READ_ONCE() we cast to a volatile pointer, I think the compiler cannot eliminate
the code in that case.

> > Here's a snippet of the code generated for finish_task_switch() on powerpc:
> >
> > 1b786c:   ld      r26,2624(r30)   # mm = rq->prev_mm;
> > .......
> > 1b78c8:   cmpdi   cr7,r26,0
> > 1b78cc:   beq     cr7,1b78e4 <finish_task_switch+0xd0>
> > 1b78d0:   ld      r9,2312(r13)    # current
> > 1b78d4:   ld      r9,1888(r9)     # current->mm
> > 1b78d8:   cmpd    cr7,r26,r9
> > 1b78dc:   beq     cr7,1b7a70 <finish_task_switch+0x25c>
> > 1b78e0:   hwsync
> > 1b78e4:   cmplwi  cr7,r27,128
> > .......
> > 1b7a70:   lwz     r9,176(r26)     # atomic_read(&mm->membarrier_state)
> > 1b7a74:   b       1b78e0 <finish_task_switch+0xcc>
> >
> > This was found while analyzing "perf c2c" reports on kernels prior
> > to commit c1753fd02a00 ("mm: move mm_count into its own cache line")
> > where mm_count was false sharing with membarrier_state.
> 
> So it was causing a noticable performance blip? But isn't anymore?
> 

It was noticeable in that it showed up amongst the top entries in perf c2c reports.
There was similar false sharing with other fields that share the cache line with
mm_count, so the gains were minimal with just this patch. c1753fd02a00 addresses
these cases too.

> > There is a minor improvement in the size of finish_task_switch().
> > The following are results from bloat-o-meter:
> >
> > GCC 7.5.0:
> > ----------
> > add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-32 (-32)
> > Function                                     old     new   delta
> > finish_task_switch                           884     852     -32
> >
> > GCC 12.2.1:
> > -----------
> > add/remove: 0/0 grow/shrink: 0/1 up/down: 0/-32 (-32)
> > Function                                     old     new   delta
> > finish_task_switch.isra                      852     820     -32
> 
> GCC 12 is a couple of years old, I assume GCC 14 behaves similarly?
> 

I cross compiled for aarch64 with gcc 14.1.1 and see similar results:

add/remove: 0/2 grow/shrink: 1/1 up/down: 4/-60 (-56)
Function                                     old     new   delta
get_nohz_timer_target                        352     356      +4
e843419 at 0b02_0000d7e7_408                      8       -      -8
e843419 at 01bb_000021d2_868                      8       -      -8
finish_task_switch.isra                      592     548     -44
Total: Before=31013792, After=31013736, chg -0.00%

> > LLVM 17.0.6:
> > ------------
> > add/remove: 0/0 grow/shrink: 0/2 up/down: 0/-36 (-36)
> > Function                                     old     new   delta
> > rt_mutex_schedule                            120     104     -16
> > finish_task_switch                           792     772     -20
> >
> > Signed-off-by: Nysal Jan K.A <nysal at linux.ibm.com>
> > ---
> >  include/linux/sched/mm.h | 2 ++
> >  1 file changed, 2 insertions(+)
> >
> > diff --git a/include/linux/sched/mm.h b/include/linux/sched/mm.h
> > index 07bb8d4181d7..042e60ab853a 100644
> > --- a/include/linux/sched/mm.h
> > +++ b/include/linux/sched/mm.h
> > @@ -540,6 +540,8 @@ enum {
> >  
> >  static inline void membarrier_mm_sync_core_before_usermode(struct mm_struct *mm)
> >  {
> > +	if (!IS_ENABLED(CONFIG_ARCH_HAS_SYNC_CORE_BEFORE_USERMODE))
> > +		return;
> >  	if (current->mm != mm)
> >  		return;
> >  	if (likely(!(atomic_read(&mm->membarrier_state) &
> 
> The other option would be to have a completely separate stub, eg:
> 
>   #ifdef CONFIG_ARCH_HAS_SYNC_CORE_BEFORE_USERMODE
>   static inline void membarrier_mm_sync_core_before_usermode(struct mm_struct *mm)
>   {
>           if (current->mm != mm)
>                   return;
>           if (likely(!(atomic_read(&mm->membarrier_state) &
>                        MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE)))
>                   return;
>           sync_core_before_usermode();
>   }
>   #else
>   static inline void membarrier_mm_sync_core_before_usermode(struct mm_struct *mm) { }
>   #endif
> 
> Not sure what folks prefer.
> 
> In either case I think it's probably worth a short comment explaining
> why it's worth the trouble (ie. that the atomic_read() prevents the
> compiler from doing DCE).
>

I'll send a v2 with a comment added in there. Thanks for the review.

--Nysal