cvs commit: src/sys/sys proc.h umtx.h src/sys/kernkern_thread.c kern_umtx.c

[Scott Long]

Alexey Dokuchaev wrote:
> On Sat, Mar 05, 2005 at 09:15:03AM +0000, David Xu wrote:
> 
>>davidxu     2005-03-05 09:15:03 UTC
>>
>>  FreeBSD src repository
>>
>>  Modified files:
>>    sys/sys              proc.h umtx.h 
>>    sys/kern             kern_thread.c kern_umtx.c 
>>  Log:
>>  Allocate umtx_q from heap instead of stack, this avoids
>>  page fault panic in kernel under heavy swapping.
> 
> 
> So..  Slow malloc/free path at last.  As a side note, could someone (not
> necessarily David) comment on my impression that, for example, recently
> reported not-so-optimal performance of our threading model(s) is largely due
> to heavy use of malloc/free, as opposed to other operating systems out
> there?  Am I right thinking that this is main bottleneck?  If malloc'ing is
> so costly, why we're taking this path?  Can kernel malloc() be optimized?
> 
> Thanks.
> 
> ./danfe

All David's change does is allocate the umtx_q object at thread
creation, it doesn't allocate it on every contested lock operation
like was first suggested.  So yes, it likely slows down thread creation
by a small amount, but there is already a lot of other overhead there.
Thread allocation is already optimized by UMA, and it might be possible
to optmize the sub-allocations by putting them under UMA too.

Scott