[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Mon, 06 Jun 2016 12:48:16 +0800
From: xinhui <xinhui.pan@...ux.vnet.ibm.com>
To: Peter Zijlstra <peterz@...radead.org>,
Waiman Long <waiman.long@....com>
CC: linux-kernel@...r.kernel.org, mingo@...hat.com
Subject: Re: [PATCH] locking/qspinlock: Use this_cpu_ptr instead of this_cpu_dec
On 2016年06月04日 05:35, Peter Zijlstra wrote:
> On Fri, Jun 03, 2016 at 05:20:10PM -0400, Waiman Long wrote:
>> On 06/03/2016 05:48 AM, Pan Xinhui wrote:
>>> queued_spin_lock_slowpath should not worry about interrupt change
>>> node->count by accident because ->count is inc and dec when we
>>> enter/leave queued_spin_lock_slowpath.
>>>
>>> So this_cpu_dec() does some no point things here, lets use this_cpu_ptr
>>> for a small optimization.
>>>
>>> Signed-off-by: Pan Xinhui<xinhui.pan@...ux.vnet.ibm.com>
>>> ---
>>> kernel/locking/qspinlock.c | 2 +-
>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c
>>> index 99f31e4..2b4daac 100644
>>> --- a/kernel/locking/qspinlock.c
>>> +++ b/kernel/locking/qspinlock.c
>>> @@ -492,7 +492,7 @@ release:
>>> /*
>>> * release the node
>>> */
>>> - this_cpu_dec(mcs_nodes[0].count);
>>> + this_cpu_ptr(&mcs_nodes[0])->count--;
>>> }
>>> EXPORT_SYMBOL(queued_spin_lock_slowpath);
>>>
>>
>> Is this going to generate better code for PPC? For x86, I think it will
>> cause more instruction to be issued.
>
yes, ppc will do some check when restore irq flags. it's really heavy.
and yes, such change cause more instructions to be issued.
there is a RELOC_HIDE macro in this_cpu_ptr and gcc can't do optimization.
How about ICC. I once used ICC when I was in intel but i forgot the result.
> It does; I think he wants __this_cpu_dec() instead, but the Changelog
> needs improvement to explain why that is ok.
>
oh, great. __this_cpu_dec() is fine. thanks
no any irq flags save/restore again. :)
thanks
xinhui
Powered by blists - more mailing lists