[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <nszn2c226bf6xslp7r5axmiov4fuhmwqqyzap7b2lysuch7fnl@uhgtmsaxttf3>
Date: Sat, 26 Oct 2024 00:00:37 +0530
From: "Nysal Jan K.A." <nysal@...ux.ibm.com>
To: Michael Ellerman <mpe@...erman.id.au>,
Andrew Morton <akpm@...ux-foundation.org>,
Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
Cc: linuxppc-dev@...ts.ozlabs.org, Nathan Chancellor <nathan@...nel.org>,
Nick Desaulniers <ndesaulniers@...gle.com>,
Bill Wendling <morbo@...gle.com>,
Justin Stitt <justinstitt@...gle.com>,
Vlastimil Babka <vbabka@...e.cz>,
Kent Overstreet <kent.overstreet@...ux.dev>,
Rick Edgecombe <rick.p.edgecombe@...el.com>,
Roman Gushchin <roman.gushchin@...ux.dev>,
linux-kernel@...r.kernel.org, llvm@...ts.linux.dev
Subject: Re: [PATCH] sched/membarrier: Fix redundant load of membarrier_state
On Fri, Oct 25, 2024 at 11:29:38AM +1100, Michael Ellerman wrote:
> [To += Mathieu]
>
> "Nysal Jan K.A." <nysal@...ux.ibm.com> writes:
> > From: "Nysal Jan K.A" <nysal@...ux.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@...2_0000d7e7_408 8 - -8
e843419@...b_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@...ux.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
Powered by blists - more mailing lists