Re: CpuSet on PAGG (Re: PAGG in Open Source projects?)

Hello, Paul.
Thanks for your comments.

 >>I wondered why every advanced feature duplicates its own hooks
 >>in fork()/exit() for those event handling.
 >
 >
 > Sometimes the simple mechanism is the best.  If dozens of features
 > require a special subroutine to be called from the same place, such as
 > fork, exec or init (kernel boot), then perhaps dozens of subroutine
 > calls, all in a row, is best.
 >
 > Just because a way of doing things requires a minimum of compute cycles,
 > and is so simple that even a beginning programmer can understand it
 > almost immediately, doesn't make it inferior.

Hmm..., 'Simple is the best' is truth, I also think.


 >>__alloc_pages() has possibility to be called from a section which
 >>cannot block. Thus, I replaced PAGG semaphore by RCU.
 >
 > Are you saying that pagg_sem can't block?  I thought semaphores
 > could block.

No, semaphores always have possibility to block, and pagg_sem is also same.
(Who dose need a never-blocking-semaphore?)
When I modified the CpuSet feature for using PAGG, I noticed we need call
pagg_get() from some critical sections.
Thus, I modified PAGG to replace task->pagg_sem by RCU.
pagg_get() does not block on the PAGG implemented by RCU.

 > I am missing something here.  What is the sequence of events that
 > would lead to trying to hold a semaphore from interrupt context
 > (before your lockless changes)?

For example, pid_array_load() in kernel/cpuset.c is relevant.
(This is the Paul Jackson's original implementation.)

int pid_array_load()
{
   read_lock(&tasklist_lock);              ---------
   do_each_thread(g, p) {                      ^
     if (p->cpuset == cs) {                    |
              :                             Critical
     }                                      Section
   } while_each_thread(g, p);                  |
array_full:                                   v
   read_unlock(&tasklist_lock);            ---------
}

In this case, if CpuSet would be a PAGG's custmer, we call pagg_call()
to refer the PAGG object related to each task under the read_lock(&tasklist_lock).
Thus, pagg_get() must not block.

 >>Indeed, I think CpuSet and PAGG(+Job/CSA) are so worthwhile solution.
 >>I'll try to move the discussion into LKML.
 >>But we might start from the improvement of PAGG before it.
 >
 > I think that there is a good chance that later this month, February 2005,
 > the subject of the cpuset patch will again become active on lkml.
 > I will be pleased and grateful for any support you might provide to
 > encourage the acceptance of cpusets.

Today, my colleague acclaim CpuSet also. :)
It's good, if we can use CpuSet on stock-kernel.

Thank you.
-- 
Linux Promotion Center, NEC
KaiGai Kohei <[email protected]>