Re: [PATCH v10 07/36] function_graph: Allow multiple users to attach to function graph

On Fri, 24 May 2024 21:32:08 -0400
Steven Rostedt [off-list ref] wrote:

On Tue,  7 May 2024 23:09:22 +0900
"Masami Hiramatsu (Google)" [off-list ref] wrote:

quoted

@@ -109,6 +244,21 @@ ftrace_push_return_trace(unsigned long ret, unsigned long func,
 	if (!current->ret_stack)
 		return -EBUSY;
 
+	/*
+	 * At first, check whether the previous fgraph callback is pushed by
+	 * the fgraph on the same function entry.
+	 * But if @func is the self tail-call function, we also need to ensure
+	 * the ret_stack is not for the previous call by checking whether the
+	 * bit of @fgraph_idx is set or not.
+	 */
+	ret_stack = get_ret_stack(current, current->curr_ret_stack, &offset);
+	if (ret_stack && ret_stack->func == func &&
+	    get_fgraph_type(current, offset + FGRAPH_FRAME_OFFSET) == FGRAPH_TYPE_BITMAP &&
+	    !is_fgraph_index_set(current, offset + FGRAPH_FRAME_OFFSET, fgraph_idx))
+		return offset + FGRAPH_FRAME_OFFSET;
+
+	val = (FGRAPH_TYPE_RESERVED << FGRAPH_TYPE_SHIFT) | FGRAPH_FRAME_OFFSET;
+
 	BUILD_BUG_ON(SHADOW_STACK_SIZE % sizeof(long));

I'm trying to figure out what the above is trying to do. This gets called
once in function_graph_enter() (or function_graph_enter_ops()). What
exactly are you trying to catch here?

Aah, good catch! This was originally for catching the self tail-call case with
multiple fgraph callback on the same function, but it was my misread.
In later patch ([12/36]), we introduced function_graph_enter_ops() so that
we can skip checking hash table and directly pass the fgraph_ops to user
callback. I thought this function_graph_enter_ops() is used even if multiple
fgraph is set on the same function. In this case, we always need to check the
stack can be reused(pushed by other fgraph_ops on the same function) or not.
But as we discussed, the function_graph_enter_ops() is used only when only
one fgraph is set on the function (if there are multiple fgraphs are set on
the same function, use function_graph_enter() ), we are sure that 
ftrace_push_return_trace() is called only once on hooking the function entry.
Thus we don't need to reuse it.

Is it from this email:

  https://lore.kernel.org/all/20231110105154.df937bf9f200a0c16806c522@kernel.org/ (local)

As that's the last version before you added the above code.

But you also noticed it may not be needed, but triggered a crash without it
in v3:

  https://lore.kernel.org/all/20231205234511.3839128259dfec153ea7da81@kernel.org/ (local)

I removed this code in my version and it runs just fine. Perhaps there was
another bug that this was hiding that you fixed in later versions?

No problem. I think we can remove this block safely.

Thank you,

-- Steve

-- 
Masami Hiramatsu (Google) [off-list ref]