Messages in this thread | | | Date | Thu, 8 May 2008 22:19:56 +0200 | From | Ingo Molnar <> | Subject | Re: [patch] speed up / fix the new generic semaphore code (fix AIM7 40% regression with 2.6.26-rc1) |
| |
* Linus Torvalds <torvalds@linux-foundation.org> wrote:
> On Thu, 8 May 2008, Linus Torvalds wrote: > > > > Why don't we just make it do the same thing that the x86 semaphores used > > to do: make it signed, and decrement unconditionally. And callt eh > > slow-path if it became negative. > > ... > > and now we have an existing known-good implementation to look at? > > Ok, after having thought that over, and looked at the code, I think I > like your version after all. The old implementation was pretty complex > due to the need to be so extra careful about the count that could > change outside of the lock, so everything considered, a new > implementation that is simpler is probably the better choice.
yeah. I thought about that too, the problem i found is this thing in the old lib/semaphore-sleepers.c code's __down() path:
remove_wait_queue_locked(&sem->wait, &wait); wake_up_locked(&sem->wait); spin_unlock_irqrestore(&sem->wait.lock, flags); tsk->state = TASK_RUNNING;
that mystery wakeup once i understood to be necessary for some weird ordering reason, but it would probably be hard to justify in the new code, because it's done unconditionally, regardless of whether there are sleepers around.
And once we deviate from the old code, we might as well go for the simplest approach - which also happens to be rather close to the mutex code's current slowpath - just with counting property added, legacy semantics and no lockdep coverage.
> Ergo, I will just pull your scheduler tree.
great! Meanwhile a 100 randconfigs booted fine with that tree so i'd say the implementation is robust.
i also did a quick re-test of AIM7 because the wakeup logic changed a bit from what i tested initially (from round-robin to strict FIFO), and as expected not much changed in the AIM7 results on the quad:
Tasks Jobs/Min JTI Real CPU Jobs/sec/task 2000 55019.9 96 211.6 806.5 0.4585 2000 55116.2 90 211.2 804.7 0.4593 2000 55082.3 82 211.3 805.5 0.4590
this is slightly lower but the test was not fully apples to apples because this also had some tracing active and other small details. It's still very close to the v2.6.25 numbers. I suspect some more performance could be won in this particular workload by getting rid of the BKL dependency altogether.
Ingo
| |