Gilles Chanteperdrix wrote:
Jan Kiszka wrote:
Am 05.11.2010 00:24, Gilles Chanteperdrix wrote:
Jan Kiszka wrote:
Am 04.11.2010 23:06, Gilles Chanteperdrix wrote:
Jan Kiszka wrote:
At first sight, here you are more breaking things than cleaning them.
Still, it has the SMP record for my test program, still runs with ftrace on (after 2 hours, where it previously failed after maximum 23 minutes).
My version was indeed still buggy, I'm reworking it ATM.

If I get the gist of Jan's changes, they are (using the IPI to transfer one bit of information: your cpu needs to reschedule):

xnsched_set_resched:
-      setbits((__sched__)->status, XNRESCHED);

xnpod_schedule_handler:
+       xnsched_set_resched(sched);
        
If you (we?) decide to keep the debug checks, under what circumstances would the current check trigger (in laymans language, that I'll be able to understand)?
That's actually what /me is wondering as well. I do not see yet how you
can reliably detect a missed reschedule reliably (that was the purpose
of the debug check) given the racy nature between signaling resched and
processing the resched hints.
The purpose of the debugging change is to detect a change of the
scheduler state which was not followed by setting the XNRESCHED bit.
But that is nucleus business, nothing skins can screw up (as long as
they do not misuse APIs).
Yes, but it happens that we modify the nucleus from time to time.

Getting it to work is relatively simple: we add a "scheduler change set
remotely" bit to the sched structure which is NOT in the status bit, set
this bit when changing a remote sched (under nklock). In the debug check
code, if the scheduler state changed, and the XNRESCHED bit is not set,
only consider this a but if this new bit is not set. All this is
compiled out if the debug is not enabled.
I still see no benefit in this check. Where to you want to place the bit
set? Aren't that just the same locations where
xnsched_set_[self_]resched already is today?
Well no, that would be another bit in the sched structure which would
allow us to manipulate the status bits from the local cpu. That
supplementary bit would only be changed from a distant CPU, and serve to
detect the race which causes the false positive. The resched bits are
set on the local cpu to get xnpod_schedule to trigger a rescheduling on
the distance cpu. That bit would be set on the remote cpu's sched. Only
when debugging is enabled.

But maybe you can provide some motivating bug scenarios, real ones of
the past or realistic ones of the future.
Of course. The bug is anything which changes the scheduler state but
does not set the XNRESCHED bit. This happened when we started the SMP
port. New scheduling policies would be good candidates for a revival of
this bug.

You don't gain any worthwhile check if you cannot make the
instrumentation required for a stable detection simpler than the proper
problem solution itself. And this is what I'm still skeptical of.

The solution is simple, but finding the problem without the instrumentation is way harder than with the instrumentation, so the instrumentation is worth something.

Reproducing the false positive is surprisingly easy with a simple
dual-cpu semaphore ping-pong test. So, here is the (tested) patch, using a ridiculous long variable name to illustrate what I was thinking about:

diff --git a/include/nucleus/sched.h b/include/nucleus/sched.h
index 8888cf4..454b8e8 100644
--- a/include/nucleus/sched.h
+++ b/include/nucleus/sched.h
@@ -108,6 +108,9 @@ typedef struct xnsched {
        struct xnthread *gktarget;
 #endif

+#ifdef CONFIG_XENO_OPT_DEBUG_NUCLEUS
+       int debug_resched_from_remote;
+#endif
 } xnsched_t;

 union xnsched_policy_param;
@@ -185,6 +188,8 @@ static inline int xnsched_resched_p(struct xnsched *sched)
   xnsched_t *current_sched = xnpod_current_sched();                    \
   __setbits(current_sched->status, XNRESCHED);                         \
   if (current_sched != (__sched__))    {                               \
+         if (XENO_DEBUG(NUCLEUS))                                      \
+                 __sched__->debug_resched_from_remote = 1;             \
       xnarch_cpu_set(xnsched_cpu(__sched__), current_sched->resched);  \
   }                                                                    \
 } while (0)
diff --git a/ksrc/nucleus/pod.c b/ksrc/nucleus/pod.c
index 4cb707a..50b0f49 100644
--- a/ksrc/nucleus/pod.c
+++ b/ksrc/nucleus/pod.c
@@ -2177,6 +2177,10 @@ static inline int __xnpod_test_resched(struct xnsched 
*sched)
                xnarch_cpus_clear(sched->resched);
        }
 #endif
+       if (XENO_DEBUG(NUCLEUS) && sched->debug_resched_from_remote) {
+               sched->debug_resched_from_remote = 0;
+               resched = 1;
+       }
        clrbits(sched->status, XNRESCHED);
        return resched;
 }


I am still uncertain.
Will only work if all is done under nklock, otherwise two almost simultaneous xnsched_resched_p from different cpus, might lead to one of the ipi wakeups sees the 0 written due to handling the first ipi interrupt.

/Anders



_______________________________________________
Xenomai-core mailing list
Xenomai-core@gna.org
https://mail.gna.org/listinfo/xenomai-core

Reply via email to