Hi Andrew,
On 2023/1/6 am 10:59, Andrew Morton wrote:
On Tue, 3 Jan 2023 14:33:03 +0800 Hongchen Zhang <zhanghongchen@xxxxxxxxxxx> wrote:
Use spinlock in pipe_read/write cost too much time,IMO
pipe->{head,tail} can be protected by __pipe_{lock,unlock}.
On the other hand, we can use __pipe_lock/unlock to protect the
pipe->head/tail in pipe_resize_ring and post_one_notification.
Can you please test this with the test code in Linus's 0ddad21d3e99 and
check that performance is good?
I tested with the test code in Linus's 0ddad21d3e99,and get the
following result:
1) before this patch
13,136.54 msec task-clock # 3.870 CPUs utilized
1,186,779 context-switches # 90.342 K/sec
668,867 cpu-migrations # 50.917 K/sec
895 page-faults # 68.131 /sec
29,875,711,543 cycles # 2.274 GHz
12,372,397,462 instructions # 0.41 insn per cycle
2,480,235,723 branches # 188.804 M/sec
47,191,943 branch-misses # 1.90% of all branches
3.394806886 seconds time elapsed
0.037869000 seconds user
0.189346000 seconds sys
2) after this patch
12,395.63 msec task-clock # 4.138 CPUs utilized
1,193,381 context-switches # 96.274 K/sec
585,543 cpu-migrations # 47.238 K/sec
1,063 page-faults # 85.756 /sec
27,691,587,226 cycles # 2.234 GHz
11,738,307,999 instructions # 0.42 insn per cycle
2,351,299,522 branches # 189.688 M/sec
45,404,526 branch-misses # 1.93% of all branches
2.995280878 seconds time elapsed
0.010615000 seconds user
0.206999000 seconds sys
After adding this patch, the time used on this test program becomes less.
Another thing, because of my carelessness, trinity tool tested a bug, I
will remove the unnecessary __pipe_{lock,unlock} in pipe_resize_ring in
v3, because the lock is owned in pipe_fcntl/pipe_ioctl.
Thanks.
