Created attachment 591717 [details]
calls semget 1 millions times

Description of problem:
It seemed that the system call sys_semget costs more time than that in rhel 5.


Version-Release number of selected component (if applicable):
Rhel-6.2
kernel-2.6.32-220

How reproducible:
compile the and execute the two programs a(calls semget for 1 million times ) and b(calls semget in a while(1) statement) on rhel5 and rhel6 .

Steps to Reproduce:
1.On rhel5, execute a singlely , then execute b and re-execute a , compare the results(time cost) outputed by a.
2.On rhel6, execute a singlely , then execute b and re-execute a , compare the results(time cost) outputed by a.
3.
  
Actual results:

DELL OPTIPLEX 380 (Pentium Dual-Core E5800 ,4G RAM)
On rhel5 , a costs about 0.2 seconds when it is executed singley , and costs about 0.5 seconds when b is running.
On rhel6 , a costs about 0.2 seconds when it is executed singley , but costs about 7 seconds when b is running.
 

Expected results:
At least ，we should get the similar results . 

Additional Info:
I searched it on google , and found that the lock in sys_semget had been changed from mutex to rw_semaphore .
It seemed that rw_semaphore down_write() was not as efficient as mutex's mutex_lock(), it was slower.