Message ID | 1271212509.13059.135.camel@pasglop (mailing list archive) |
---|---|
State | Not Applicable |
Headers | show |
On Wed, 2010-04-14 at 12:35 +1000, Benjamin Herrenschmidt wrote: > Hi Peter ! > > I -may- have found a bug with mutex adaptative spinning. We hit it when > torture testing CPU unplug. .../... In fact, I wonder if there's another potential problem: If the owner is actually running, it may do so for a very long time. It looks to me that everybody trying to take the mutex will thus spin and never get out of the spin loop until the owner stops running. That sounds very wrong to me :-) Shouldn't you at least remove the test for !owner when testing need_resched() in __mutex_lock_common() ? Cheers, Ben.
On Wed, 2010-04-14 at 12:56 +1000, Benjamin Herrenschmidt wrote: > On Wed, 2010-04-14 at 12:35 +1000, Benjamin Herrenschmidt wrote: > > Hi Peter ! > > > > I -may- have found a bug with mutex adaptative spinning. We hit it when > > torture testing CPU unplug. > > .../... > > In fact, I wonder if there's another potential problem: > > If the owner is actually running, it may do so for a very long time. It > looks to me that everybody trying to take the mutex will thus spin and > never get out of the spin loop until the owner stops running. The inner-most spin loop breaks out on need_resched(): if (task_thread_info(rq->curr) != owner || need_resched()) return 0;
On Fri, 2010-04-16 at 23:27 +0200, Peter Zijlstra wrote: > > If the owner is actually running, it may do so for a very long time. It > > looks to me that everybody trying to take the mutex will thus spin and > > never get out of the spin loop until the owner stops running. > > The inner-most spin loop breaks out on need_resched(): > > if (task_thread_info(rq->curr) != owner || need_resched()) > return 0; You are right, this was only a problem in conjunction with the other bug returning 1 all the time, causing us to ignore need_resched(). With that fixed, it looks fine now. Thanks ! Cheers, Ben.
diff --git a/kernel/sched.c b/kernel/sched.c index a3dff1f..11b7f4a 100644 --- a/kernel/sched.c +++ b/kernel/sched.c @@ -3780,7 +3780,7 @@ int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner) * the mutex owner just released it and exited. */ if (probe_kernel_address(&owner->cpu, cpu)) - goto out; + return 0; #else cpu = owner->cpu;
Hi Peter ! I -may- have found a bug with mutex adaptative spinning. We hit it when torture testing CPU unplug. Basically, what happens is mutex_spin_on_owner() returns 1 if the owner CPU is offline. That means that the caller (__mutex_lock_common()) will spin until the mutex is released since there's a valid owner (so the need_resched() test doesn't trigger). This will deadlock if this is the only remaining CPU online. In fact, it even deadlocks with more than 1 CPU online, for example, I have a case where the 2 remaining online CPUs got into __mutex_lock_common() at the same time, and so got into that spin loop before any of them added itself to the wait_list, thus never hitting the exist condition there. I believe your test against nr_cpumask_bits is also a bit fragile for similar reasons. IE. You have what looks like a valid owner but it seems to be on an invalid CPU. It could be a freed thread_info, in which case returning 1 is fine, but if it's anything else, kaboom. I think it's better to be safe than sorry here and just go to sleep (ie return 0). Same comment with the DEBUG_PAGE_ALLOC case. In fact, this (untested) patch makes the function simpler and I don't think will have any negative effect on performances. Let me know what you think: ---- cut here ---- [PATCH] mutex: Don't spin when the owner CPU is offline or other weird cases The current code might spin forever if the CPU owning the mutex has been offlined, and the last CPU in the system is trying to acquire it, since mutex_spin_on_owner() will always return 1, telling the caller to spin until the mutex has been released. This patch changes mutex_spin_on_owner() to return 0 (don't spin) in any case where we aren't sure about the owner struct validity or CPU number, and if the said CPU is offline. There is no point going back & re-evaluate spinning in corner cases like that, let's just go to sleep. Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> --- #endif @@ -3788,16 +3788,21 @@ int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner) /* * Even if the access succeeded (likely case), * the cpu field may no longer be valid. + * + * We stop spinning in this case since the owner may never */ if (cpu >= nr_cpumask_bits) - goto out; + return 0; /* * We need to validate that we can do a * get_cpu() and that we have the percpu area. + * + * If the CPU is offline, the owner will never release it so + * we must not spin */ if (!cpu_online(cpu)) - goto out; + return 0; rq = cpu_rq(cpu); @@ -3816,7 +3821,6 @@ int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner) cpu_relax(); } -out: return 1; } #endif