On Fri, Oct 22, 2021 at 07:01:35PM +0200, Peter Zijlstra wrote:

quoted

On Fri, Oct 22, 2021 at 05:54:31PM +0100, Mark Rutland wrote:

quoted

quoted

Pardon my thin understanding of the scheduler, but I assume this change
doesn't mean stack_trace_save_tsk() stops working for "current", right?
In trying to answer this for myself, I couldn't convince myself what value
current->__state have here. Is it one of TASK_(UN)INTERRUPTIBLE ?

Regardless of that, current->on_rq will be non-zero, so you're right that this
causes stack_trace_save_tsk() to not work for current, e.g.

| # cat /proc/self/stack 
| # wc  /proc/self/stack 
|         0         0         0 /proc/self/stack

TBH, I think that (taking a step back from this issue in particular)
stack_trace_save_tsk() *shouldn't* work for current, and callers *should* be
forced to explicitly handle current separately from blocked tasks.

That..

So I think I'd prefer the following approach to that (and i'm not
currently volunteering for it):

 - convert all archs to ARCH_STACKWALK; this gets the semantics out of
   arch code and into the single kernel/stacktrace.c file.

 - bike-shed a new/improved stack_trace_save*() API and implement it
   *once* in generic code based on arch_stack_walk().

 - convert users; delete old etc..

For now, current users of stack_trace_save_tsk() very much expect
tsk==current to work.

quoted

quoted

So we could fix this in the stacktrace code with:

| diff --git a/kernel/stacktrace.c b/kernel/stacktrace.c
| index a1cdbf8c3ef8..327af9ff2c55 100644
| --- a/kernel/stacktrace.c
| +++ b/kernel/stacktrace.c
| @@ -149,7 +149,10 @@ unsigned int stack_trace_save_tsk(struct task_struct *tsk, unsigned long *store,
|                 .skip   = skipnr + (current == tsk),
|         };
|  
| -       task_try_func(tsk, try_arch_stack_walk_tsk, &c);
| +       if (tsk == current)
| +               try_arch_stack_walk_tsk(tsk, &c);
| +       else
| +               task_try_func(tsk, try_arch_stack_walk_tsk, &c);
|  
|         return c.len;
|  }

... and we could rename task_try_func() to blocked_task_try_func(), and
later push the distinction into higher-level callers.

I think I favour this fix if we have to. But that's for next week :-)

I ended up with the below delta to this patch.

--- a/kernel/stacktrace.c
+++ b/kernel/stacktrace.c

@@ -101,7 +101,7 @@ static bool stack_trace_consume_entry_no
 }
 
 /**
- * stack_trace_save - Save a stack trace into a storage array
+ * stack_trace_save - Save a stack trace (of current) into a storage array
  * @store:	Pointer to storage array
  * @size:	Size of the storage array
  * @skipnr:	Number of entries to skip at the start of the stack trace

@@ -132,7 +132,7 @@ static int try_arch_stack_walk_tsk(struc
 
 /**
  * stack_trace_save_tsk - Save a task stack trace into a storage array
- * @task:	The task to examine
+ * @task:	The task to examine (current allowed)
  * @store:	Pointer to storage array
  * @size:	Size of the storage array
  * @skipnr:	Number of entries to skip at the start of the stack trace

@@ -149,13 +149,25 @@ unsigned int stack_trace_save_tsk(struct
 		.skip	= skipnr + (current == tsk),
 	};
 
-	task_try_func(tsk, try_arch_stack_walk_tsk, &c);
+	/*
+	 * If the task doesn't have a stack (e.g., a zombie), the stack is
+	 * empty.
+	 */
+	if (!try_get_task_stack(tsk))
+		return 0;
+
+	if (tsk == current)
+		try_arch_stack_walk_tsk(tsk, &c);
+	else
+		task_try_func(tsk, try_arch_stack_walk_tsk, &c);
+
+	put_task_stack(tsk);
 
 	return c.len;
 }
 
 /**
- * stack_trace_save_regs - Save a stack trace based on pt_regs into a storage array
+ * stack_trace_save_regs - Save a stack trace (of current) based on pt_regs into a storage array
  * @regs:	Pointer to pt_regs to examine
  * @store:	Pointer to storage array
  * @size:	Size of the storage array