Re: [BUG] cpu hotplug vs scheduler

[Dmitry Adamushko]

2008/5/14 Avi Kivity <avi@xxxxxxxxxxxx>:

> [ ... ]
>
>  [4302727.900184] Call Trace:
>  [4302727.900184]  [<ffffffff803249de>] spin_bug+0x9e/0xe9
>  [4302727.900184]  [<ffffffff80324af4>] _raw_spin_lock+0x41/0x123
>  [4302727.900184]  [<ffffffff80439638>] _spin_lock_irqsave+0x2f/0x37
>  [4302727.900184]  [<ffffffff8022ef7c>] print_cfs_rq+0xca/0x46a
>  [4302727.900184]  [<ffffffff80231f97>] sched_debug_show+0x7a3/0xb8c
>  [4302727.900184]  [<ffffffff8023238d>] sysrq_sched_debug_show+0xd/0xf
>  [4302727.900184]  [<ffffffff802323ee>] pick_next_task_fair+0x5f/0x86


Err... sorry for the broken patch. The patch below on top of the
previous one should address this issue (ugly, but should be ok for
debugging).
'tasklist_lock' shouldn't cause a double lock, I guess.

Sorry for rather 'blind' attempts. If no, then I'll
prepare/test/take-a-closer-look at it later today when I'm at home.

TIA,


------ kernel/sched_debug-prev.c   2008-05-14 14:53:28.000000000 +0200
+++ kernel/sched_debug.c        2008-05-14 14:58:12.000000000 +0200
@@ -125,6 +125,7 @@ void print_cfs_rq(struct seq_file *m, in
        char path[128] = "";
        struct cgroup *cgroup = NULL;
        struct task_group *tg = cfs_rq->tg;
+       int was_locked;

        if (tg)
                cgroup = tg->css.cgroup;
@@ -138,7 +139,11 @@ void print_cfs_rq(struct seq_file *m, in
        SEQ_printf(m, "  .%-30s: %Ld.%06ld\n", "exec_clock",
                        SPLIT_NS(cfs_rq->exec_clock));

-       spin_lock_irqsave(&rq->lock, flags);
+       was_locked = spin_is_locked(&rq->lock);
+
+       if (!was_locked)
+               spin_lock_irqsave(&rq->lock, flags);
+
        if (cfs_rq->rb_leftmost)
                MIN_vruntime = (__pick_next_entity(cfs_rq))->vruntime;
        last = __pick_last_entity(cfs_rq);
@@ -146,7 +151,10 @@ void print_cfs_rq(struct seq_file *m, in
                max_vruntime = last->vruntime;
        min_vruntime = rq->cfs.min_vruntime;
        rq0_min_vruntime = per_cpu(runqueues, 0).cfs.min_vruntime;
-       spin_unlock_irqrestore(&rq->lock, flags);
+
+       if (!was_locked)
+               spin_unlock_irqrestore(&rq->lock, flags);
+
        SEQ_printf(m, "  .%-30s: %Ld.%06ld\n", "MIN_vruntime",
                        SPLIT_NS(MIN_vruntime));
        SEQ_printf(m, "  .%-30s: %Ld.%06ld\n", "min_vruntime",

---


-- 
Best regards,
Dmitry Adamushko
--- kernel/sched_debug-prev.c	2008-05-14 14:53:28.000000000 +0200
+++ kernel/sched_debug.c	2008-05-14 14:58:12.000000000 +0200
@@ -125,6 +125,7 @@ void print_cfs_rq(struct seq_file *m, in
 	char path[128] = "";
 	struct cgroup *cgroup = NULL;
 	struct task_group *tg = cfs_rq->tg;
+	int was_locked;
 
 	if (tg)
 		cgroup = tg->css.cgroup;
@@ -138,7 +139,11 @@ void print_cfs_rq(struct seq_file *m, in
 	SEQ_printf(m, "  .%-30s: %Ld.%06ld\n", "exec_clock",
 			SPLIT_NS(cfs_rq->exec_clock));
 
-	spin_lock_irqsave(&rq->lock, flags);
+	was_locked = spin_is_locked(&rq->lock);
+
+	if (!was_locked)
+		spin_lock_irqsave(&rq->lock, flags);
+	
 	if (cfs_rq->rb_leftmost)
 		MIN_vruntime = (__pick_next_entity(cfs_rq))->vruntime;
 	last = __pick_last_entity(cfs_rq);
@@ -146,7 +151,10 @@ void print_cfs_rq(struct seq_file *m, in
 		max_vruntime = last->vruntime;
 	min_vruntime = rq->cfs.min_vruntime;
 	rq0_min_vruntime = per_cpu(runqueues, 0).cfs.min_vruntime;
-	spin_unlock_irqrestore(&rq->lock, flags);
+
+	if (!was_locked)
+		spin_unlock_irqrestore(&rq->lock, flags);
+
 	SEQ_printf(m, "  .%-30s: %Ld.%06ld\n", "MIN_vruntime",
 			SPLIT_NS(MIN_vruntime));
 	SEQ_printf(m, "  .%-30s: %Ld.%06ld\n", "min_vruntime",