From: Honglei Wang <wanghonglei@didichuxing.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@redhat.com>,
Juri Lelli <juri.lelli@redhat.com>,
Vincent Guittot <vincent.guittot@linaro.org>,
Dietmar Eggemann <dietmar.eggemann@arm.com>,
Steven Rostedt <rostedt@goodmis.org>,
Ben Segall <bsegall@google.com>, Mel Gorman <mgorman@suse.de>,
"Daniel Bristot de Oliveira" <bristot@redhat.com>,
<linux-kernel@vger.kernel.org>,
Huaixin Chang <changhuaixin@linux.alibaba.com>,
Honglei Wang <jameshongleiwang@126.com>
Subject: Re: [PATCH v2 1/3] sched/fair: avoid burst statistic if it's not set
Date: Thu, 9 Dec 2021 23:24:27 +0800 [thread overview]
Message-ID: <24777493-e182-d641-9e21-41102b0962c8@didichuxing.com> (raw)
In-Reply-To: <YbH08Nw3WpyxElE9@hirez.programming.kicks-ass.net>
On 2021/12/9 20:22, Peter Zijlstra wrote:
> On Wed, Dec 08, 2021 at 10:50:18PM +0800, Honglei Wang wrote:
>> It's not necessary to do burst associated statistic and calculation for
>> runtime if the burst feature is not set at all. Just return at the very
>> start point if so.
>>
>> Signed-off-by: Honglei Wang <wanghonglei@didichuxing.com>
>> ---
>> kernel/sched/fair.c | 5 +++++
>> 1 file changed, 5 insertions(+)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 6e476f6d9435..2cd626c22912 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -4640,6 +4640,11 @@ void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
>> if (unlikely(cfs_b->quota == RUNTIME_INF))
>> return;
>>
>> + if (cfs_b->burst == 0 && cfs_b->runtime_snap == cfs_b->quota) {
>> + cfs_b->runtime = cfs_b->quota;
>> + return;
>> + }
>> +
>> cfs_b->runtime += cfs_b->quota;
>> runtime = cfs_b->runtime_snap - cfs_b->runtime;
>> if (runtime > 0) {
>
> What actual purpose does this patch serve, other than to increase the
> line-count?
>
Yep, there is no functional improvement with this tiny section. The
original thought is why do we do the judgement, calculation and
comparing for one feature if it is not even set? Could we go a bit
faster if skip the stuff?
I disassembled the function when finished the code. If the burst feature
is not open, we just need 8 instructions (it's 6 at most cases) with the
change VS 12 ones without it. So seems there is a bit help? And it can
be cut to 5 instructions if we don't mind loss 1 potential nr_burst
statistic.
000000000000b530 <__refill_cfs_bandwidth_runtime>:
{
b530: e8 00 00 00 00 callq b535
<__refill_cfs_bandwidth_runtime+0x5>
if (unlikely(cfs_b->quota == RUNTIME_INF))
b535: 48 8b 47 10 mov 0x10(%rdi),%rax
b539: 48 83 f8 ff cmp $0xffffffffffffffff,%rax
b53d: 74 45 je b584
<__refill_cfs_bandwidth_runtime+0x54>
if (cfs_b->burst == 0 && cfs_b->runtime_snap == cfs_b->quota) {
b53f: 48 8b 77 20 mov 0x20(%rdi),%rsi---->1,3
b543: 48 8b 57 28 mov 0x28(%rdi),%rdx---->1,3
b547: 48 85 f6 test %rsi,%rsi---------->1,3
b54a: 75 05 jne b551 -------------->1,3
<__refill_cfs_bandwidth_runtime+0x21>
b54c: 48 39 d0 cmp %rdx,%rax---------->1
b54f: 74 35 je b586
<__refill_cfs_bandwidth_runtime+0x56> ---------------------------->1
cfs_b->runtime += cfs_b->quota;
b551: 48 8b 4f 18 mov 0x18(%rdi),%rcx---->2
b555: 48 01 c1 add %rax,%rcx---------->2
runtime = cfs_b->runtime_snap - cfs_b->runtime;
b558: 48 29 ca sub %rcx,%rdx---------->2
cfs_b->runtime += cfs_b->quota;
b55b: 48 89 4f 18 mov %rcx,0x18(%rdi)---->2
if (runtime > 0) {
b55f: 48 85 d2 test %rdx,%rdx---------->2
b562: 7e 0e jle b572
<__refill_cfs_bandwidth_runtime+0x42>----------------------------->2
cfs_b->burst_time += runtime;
b564: 48 01 97 e8 00 00 00 add %rdx,0xe8(%rdi)
cfs_b->nr_burst++;
b56b: 83 87 d8 00 00 00 01 addl $0x1,0xd8(%rdi)
cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
b572: 48 01 f0 add %rsi,%rax---------->2
b575: 48 39 c8 cmp %rcx,%rax---------->2
b578: 48 0f 47 c1 cmova %rcx,%rax---------->2
b57c: 48 89 47 18 mov %rax,0x18(%rdi)---->2
cfs_b->runtime_snap = cfs_b->runtime;
b580: 48 89 47 28 mov %rax,0x28(%rdi)---->2
}
b584: f3 c3 repz retq----------------->2
cfs_b->runtime = cfs_b->quota;
b586: 48 89 47 18 mov %rax,0x18(%rdi)---->1
return;
b58a: c3 retq---------------------->1
b58b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
prev parent reply other threads:[~2021-12-09 15:24 UTC|newest]
Thread overview: 3+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-12-08 14:50 Honglei Wang
2021-12-09 12:22 ` Peter Zijlstra
2021-12-09 15:24 ` Honglei Wang [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=24777493-e182-d641-9e21-41102b0962c8@didichuxing.com \
--to=wanghonglei@didichuxing.com \
--cc=bristot@redhat.com \
--cc=bsegall@google.com \
--cc=changhuaixin@linux.alibaba.com \
--cc=dietmar.eggemann@arm.com \
--cc=jameshongleiwang@126.com \
--cc=juri.lelli@redhat.com \
--cc=linux-kernel@vger.kernel.org \
--cc=mgorman@suse.de \
--cc=mingo@redhat.com \
--cc=peterz@infradead.org \
--cc=rostedt@goodmis.org \
--cc=vincent.guittot@linaro.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®