Messages in this thread | | | Date | Tue, 20 Oct 2015 12:11:33 +0200 | From | Peter Zijlstra <> | Subject | Re: [RFC] perf: fix building for ARCv1 |
| |
On Tue, Oct 20, 2015 at 08:00:46AM +0000, Vineet Gupta wrote: > On Monday 19 October 2015 03:22 PM, Peter Zijlstra wrote: > > On Mon, Oct 19, 2015 at 09:46:35AM +0000, Vineet Gupta wrote: > >> On ARC we could use the atomic EXchange to implement a user space only binary > >> semaphore - these atomic ops will be small duration so it is OK to spin wait for a > >> little bit. That's how the old pthread library worked for ARC w/o any atomic support. > > That has the obvious problem of lock-holder-preemption and the horrible > > performance issues that result from that. > > > > I think the syscall at least has deterministic behaviour, whereas that > > userspace spin loop has this abysmal worst case thing. > > I don't have issue with adding the syscall per-se. But that comes with it's own > headaches of ABI change - more importantly it requires several things to match, > libc, kernel... It would be easier if change was confined to say perf.
OTOH fixing all those would get you a 'sane' system :-)
> Can we use existing syscall(s) - again this is what our good old pthread library > code did. > > static void __pthread_acquire(int * spinlock) > { > int cnt = 0; > struct timespec tm; > > READ_MEMORY_BARRIER(); > > while (testandset(spinlock)) { <---- atomic EXchange > if (cnt < 50) { > sched_yield(); > cnt++; > } else { > tm.tv_sec = 0; > tm.tv_nsec = 2000001; > nanosleep(&tm, ((void *)0)); > cnt = 0; > } > }
*shudder* that is quite horrible.
This means all your 'atomics' are broken for anything SCHED_FIFO and the like. You simply _cannot_ run a realtime system.
(also, for ACQUIRE you want the READ_MEMORY_BARRIER() _after_ the test-and-set control dependency.)
But its your arch..
| |