ftrace: Handle tracing when switching between context

When an interrupt or NMI comes in and switches the context, there's a delay
from when the preempt_count() shows the update. As the preempt_count() is
used to detect recursion having each context have its own bit get set when
tracing starts, and if that bit is already set, it is considered a recursion
and the function exits. But if this happens in that section where context
has changed but preempt_count() has not been updated, this will be
incorrectly flagged as a recursion.

To handle this case, create another bit call TRANSITION and test it if the
current context bit is already set. Flag the call as a recursion if the
TRANSITION bit is already set, and if not, set it and continue. The
TRANSITION bit will be cleared normally on the return of the function that
set it, or if the current context bit is clear, set it and clear the
TRANSITION bit to allow for another transition between the current context
and an even higher one.

Cc: [email protected]
Fixes: edc15ca ("tracing: Avoid unnecessary multiple recursion checks")
Signed-off-by: Steven Rostedt (VMware) <[email protected]>

Author: rostedt

kernel/trace/trace.h

@@ -637,6 +637,12 @@ enum {
 	 * function is called to clear it.
 	 */
 	TRACE_GRAPH_NOTRACE_BIT,
+
+	/*
+	 * When transitioning between context, the preempt_count() may
+	 * not be correct. Allow for a single recursion to cover this case.
+	 */
+	TRACE_TRANSITION_BIT,
 };
 
 #define trace_recursion_set(bit)	do { (current)->trace_recursion |= (1<<(bit)); } while (0)
@@ -691,8 +697,21 @@ static __always_inline int trace_test_and_set_recursion(int start, int max)
 		return 0;
 
 	bit = trace_get_context_bit() + start;
-	if (unlikely(val & (1 << bit)))
-		return -1;
+	if (unlikely(val & (1 << bit))) {
+		/*
+		 * It could be that preempt_count has not been updated during
+		 * a switch between contexts. Allow for a single recursion.
+		 */
+		bit = TRACE_TRANSITION_BIT;
+		if (trace_recursion_test(bit))
+			return -1;
+		trace_recursion_set(bit);
+		barrier();
+		return bit + 1;
+	}
+
+	/* Normal check passed, clear the transition to allow it again */
+	trace_recursion_clear(TRACE_TRANSITION_BIT);
 
 	val |= 1 << bit;
 	current->trace_recursion = val;

kernel/trace/trace_selftest.c

@@ -492,8 +492,13 @@ trace_selftest_function_recursion(void)
 	unregister_ftrace_function(&test_rec_probe);
 
 	ret = -1;
-	if (trace_selftest_recursion_cnt != 1) {
-		pr_cont("*callback not called once (%d)* ",
+	/*
+	 * Recursion allows for transitions between context,
+	 * and may call the callback twice.
+	 */
+	if (trace_selftest_recursion_cnt != 1 &&
+	    trace_selftest_recursion_cnt != 2) {
+		pr_cont("*callback not called once (or twice) (%d)* ",
 			trace_selftest_recursion_cnt);
 		goto out;
 	}