Hi,
we encountered problems with the initial distribution of the
migration_tasks across the CPUs. Machines with 16 and more CPUs
sometimes won't boot. Here is a fix which works reliably:
- all migration_tasks are started on CPU#0,
- migration_task #0 is used to migrate the others to their destination,
- synchronization with migration_mask isn't needed any more.
Thanks,
best regards,
Erich
diff -urN 2.5.6-pre1-fix/kernel/sched.c 2.5.6-pre1-fix2/kernel/sched.c
--- 2.5.6-pre1-fix/kernel/sched.c Thu Feb 28 19:14:29 2002
+++ 2.5.6-pre1-fix2/kernel/sched.c Wed Mar 6 11:28:59 2002
@@ -1557,8 +1557,9 @@
static volatile unsigned long migration_mask;
-static int migration_thread(void * unused)
+static int migration_thread(void * bind_cpu)
{
+ int cpu = cpu_logical_map((int) (long) bind_cpu);
struct sched_param param = { sched_priority: 99 };
runqueue_t *rq;
int ret;
@@ -1566,33 +1567,19 @@
daemonize();
sigfillset(¤t->blocked);
set_fs(KERNEL_DS);
- ret = setscheduler(0, SCHED_FIFO, ¶m);
/*
- * We have to migrate manually - there is no migration thread
- * to do this for us yet :-)
- *
- * We use the following property of the Linux scheduler. At
- * this point no other task is running, so by keeping all
- * migration threads running, the load-balancer will distribute
- * them between all CPUs equally. At that point every migration
- * task binds itself to the current CPU.
+ * The first migration task is started on CPU #0. This one can migrate
+ * the tasks to their destination CPUs.
*/
-
- /* wait for all migration threads to start up. */
- while (!migration_mask)
- yield();
-
- for (;;) {
- preempt_disable();
- if (test_and_clear_bit(smp_processor_id(), &migration_mask))
- current->cpus_allowed = 1 << smp_processor_id();
- if (test_thread_flag(TIF_NEED_RESCHED))
- schedule();
- if (!migration_mask)
- break;
- preempt_enable();
+ if (cpu != 0) {
+ while (!cpu_rq(cpu_logical_map(0))->migration_thread)
+ yield();
+ set_cpus_allowed(current, 1UL << cpu);
}
+ printk("migration_task %d on cpu=%d\n",cpu,smp_processor_id());
+ ret = setscheduler(0, SCHED_FIFO, ¶m);
+
rq = this_rq();
rq->migration_thread = current;
preempt_enable();
@@ -1651,17 +1638,15 @@
{
int cpu;
+ current->cpus_allowed = 1UL << cpu_logical_map(0);
for (cpu = 0; cpu < smp_num_cpus; cpu++)
- if (kernel_thread(migration_thread, NULL,
+ if (kernel_thread(migration_thread, (void *) (long) cpu,
CLONE_FS | CLONE_FILES | CLONE_SIGNAL) < 0)
BUG();
-
- migration_mask = (1 << smp_num_cpus) - 1;
+ current->cpus_allowed = -1L;
for (cpu = 0; cpu < smp_num_cpus; cpu++)
while (!cpu_rq(cpu)->migration_thread)
schedule_timeout(2);
- if (migration_mask)
- BUG();
}
#endif
> we encountered problems with the initial distribution of the
> migration_tasks across the CPUs. Machines with 16 and more CPUs
> sometimes won't boot.
We found this on a 31 way and have sent a fix to Ingo already, we needed
to do a set_current_state(TASK_UNINTERRUPTIBLE) before schedule_timeout()
or we just busy loop.
Anton
On Thu, 7 Mar 2002, Anton Blanchard wrote:
> > we encountered problems with the initial distribution of the
> > migration_tasks across the CPUs. Machines with 16 and more CPUs
> > sometimes won't boot.
>
> We found this on a 31 way and have sent a fix to Ingo already, we needed
> to do a set_current_state(TASK_UNINTERRUPTIBLE) before schedule_timeout()
> or we just busy loop.
OK, good to know that it wasn't a tipical IA64 problem.
As far as I understand, your solution still relies on the specific
behavior of the scheduler for distributing the tasks uniformly and on
synchronizing among migration_tasks. I'd really prefer replacing this
shaky method by the usage of migration_task on CPU#0. Thich would always
work, no matter how many tasks are around. This saves some debugging work
to those who play around with the scheduler or want to start some kernel
threads earlier...
Regards,
Erich