On 05/25/2013 08:32 PM, Davidlohr Bueso wrote: > Yep, could you please explain what benefits you see in keeping FIFO order? a) It's user space visible. b) It's a well-defined behavior that might even make sense for some applications. Right now, a 2 semop operation with "+1, then -2" is priorized over a semop with "-1". And: It doesn't cost much: - no impact for users that use only single-op operations. - no impact for users that use only multi-op operations - for users that use both types: In the worst case some linked list splicing. Actually, the code is probably faster because wait-for-zero ops are only scanned when the semaphore values are 0. >> Acked-by: Rik van Riel >> >>> - simpler check_restart logic. >>> - Efficient handling of wait-for-zero semops, both simple and complex. >>> - Fewer restarts in update_queue(), because pending wait-for-zero do not >>> force a restart anymore. >>> >>> Other changes: >>> - try_atomic_semop() also performs the semop. Thus rename the function. >>> >>> It passes tests with qemu, but not boot-tested due to EFI problems. > I think this still needs a *lot* of testing - I don't have my Oracle > workload available right now, but I will definitely see how this patch > behaves on it. That said, I believe Oracle is are already quite happy > with the sem improvements. Ah, ok. The change is good for one application and the risk of breaking other apps is considered as negligible. > > Furthermore, this patch is way too invasive for considering it for 3.10 > - I like Rik's patch better because it simply addresses the issue and > nothing more. I would disagree: My patch is testable - with it applied, linux-3.0.10 should behave exactly as linux-3.0.9. Except the scalability - the new sem_lock from Rik is great. My problem with Rik's patch is that it is untestable: It changes the behavior and we must hope that nothing breaks. Actually, the latest patch makes it a bit worse: > @@ -720,16 +718,11 @@ static int update_queue(struct sem_array *sma, int semnum, struct list_head *pt) > > unlink_queue(sma, q); > > - if (error) { > - restart = 0; > - } else { > - semop_completed = 1; > - restart = check_restart(sma, q); > - } > + semop_completed = 1; > + if (check_restart(sma, q)) > + *restart = 1; > > wake_up_sem_queue_prepare(pt, q, error); > - if (restart) > - goto again; If check_restart returns "1", then the current (3.0.10-rc1) code restarts immediately ("goto a again"). Now the rest of the queue is processed completely and only afterwards it is scanned again. This means that wait-for-zero now succeeds only if a semaphore value stays zero. For 3.0.9, it was sufficient if the value was temporarily zero. Before the change, complex wait-for-zero would work, only simple wait-for-zero would be starved. Now all operations are starved. I've attached a test case: ./test5.sh linux-3.0.9 completes all operations With Rik's patch, the wait-for-zero remains running. -- Manfred P.S.: Btw, I found some code that uses a semop with 2 ops: http://publib.boulder.ibm.com/infocenter/iseries/v5r3/index.jsp?topic=%2Fapis%2Fapiexusmem.htm