Re: [Starpu-devel] Conversion tasks

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Nathalie Furmento, le Mon 14 Oct 2013 11:17:09 +0200, a écrit :
> Witth the current version of the trunk, submission of conversion tasks are
> broken.
> 
> Here some backtraces.
> 
> Looks like the mutex object has been destroyed too early.

Well, the pointer itself is completely bogus:

> (gdb) bt
> #1  0x00007fffef2c0612 in starpu_pthread_mutex_lock (mutex=0x40) at
> /home/nfurmento/softs/starpu/soft/starpu.git/src/common/thread.c:269
> #2  0x00007fffef2be5e1 in _starpu_barrier_counter_increment (barrier_c=0x30)
> at
> /home/nfurmento/softs/starpu/soft/starpu.git/src/common/barrier_counter.c:94

And indeed the ctx_id is STARPU_NMAX_SCHED_CTXS (expected since this
test doesn't use contexts):

> #3  0x00007fffef2f1974 in _starpu_increment_nsubmitted_tasks_of_sched_ctx
> (sched_ctx_id=10) at

that makes _starpu_get_sched_ctx_struct return NULL, and thus bogus
mutex pointer.

It has however been so since ages, I wonder why that triggers a segfault
only now.

AIUI, _starpu_increment_nsubmitted_tasks_of_sched_ctx() should simply
return when _starpu_get_sched_ctx_struct returns NULL?

Samuel