[ofw][patch][ibal] make IBBUS start up synchronous.

[Sean Hefty]

>What about the following implementation ?
>(The patch adds an event field to CI_CA object, and post the event at
>the end of the callback (__pnp_process_add_ca).

This looks better to me.  See comments:

>Index: core/al/kernel/al_ci_ca.c
>===================================================================
>--- core/al/kernel/al_ci_ca.c	(revision 4194)
>+++ core/al/kernel/al_ci_ca.c	(working copy)
>@@ -112,6 +112,8 @@
> 	cl_qlist_init( &p_ci_ca->shmid_list );
> 	cl_qpool_construct( &p_ci_ca->event_pool );
> 	p_ci_ca->verbs = *p_ci;
>+	cl_event_construct( &p_ci_ca->event );
>+	cl_event_init( &p_ci_ca->event, FALSE );

Is there a cl_event_destroy() that needs to be called?

>
> 	cl_status = cl_spinlock_init( &p_ci_ca->attr_lock );
> 	if( cl_status != CL_SUCCESS )
>Index: core/al/kernel/al_pnp.c
>===================================================================
>--- core/al/kernel/al_pnp.c	(revision 4193)
>+++ core/al/kernel/al_pnp.c	(working copy)
>@@ -994,6 +994,7 @@
>
> 	/* Cleanup the event record. */
> 	deref_al_obj( &gp_pnp->obj );
>+	cl_event_signal( &p_event_rec->p_ci_ca->event );
> 	cl_free( p_event_rec );
>
> 	AL_EXIT( AL_DBG_PNP );
>@@ -1152,7 +1153,6 @@
> 	cl_status_t			cl_status;
> 	al_pnp_ca_event_t	*p_event_rec;
> 	ib_ca_attr_t		*p_old_ca_attr;
>-	uint32_t			ref_cnt, prev_cnt, i;
>
> 	AL_ENTER( AL_DBG_PNP );
>
>@@ -1233,17 +1233,15 @@
> 	}
>
> 	/* Queue the event to the async processing manager. */
>-	ref_cnt = ref_al_obj( &gp_pnp->obj );
>+	ref_al_obj( &gp_pnp->obj );
> 	cl_async_proc_queue( gp_async_pnp_mgr, &p_event_rec->async_item
>);
>
> 	/* wait up to 3 seconds for the end of event propagation
> 	It is needed for enabling quick HCA disable/enable scenarios. */
>-	prev_cnt = ref_cnt - 1;
>-	for ( i = 0; i < 30; ++i ) {
>-		if (ref_cnt <= prev_cnt)
>-			break;
>-		cl_thread_suspend(100);
>-	}
>+	cl_status = cl_event_wait_on(
>+		&p_ci_ca->event, 3000, AL_WAIT_ALERTABLE );
>+	if (cl_status != CL_SUCCESS)
>+		return IB_ERROR;

Any reason not to wait indefinitely?  What would prevent the event from not
being signaled or can we ensure that it will be?

- Sean