lkml.org 
[lkml]   [2021]   [Dec]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH v2 1/3] sched/fair: avoid burst statistic if it's not set
From


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)

\
 
 \ /
  Last update: 2021-12-09 16:25    [W:0.052 / U:0.116 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site