From 015d08b66af5ebd10665aa90f5426930e56c540d Mon Sep 17 00:00:00 2001
From: Pierre-Marc Fournier <pierre-marc.fournier@polymtl.ca>
Date: Mon, 1 Mar 2010 16:19:38 -0500
Subject: [PATCH] convert ltt_nesting to a per_thread variable, and access it
 with volatile access

This fixes an ltt_nesting corruption bug.
---
 libust/buffers.h    |  2 +-
 libust/serialize.c  | 14 ++++++++------
 libust/tracercore.c |  2 +-
 libust/tracercore.h |  3 +--
 4 files changed, 11 insertions(+), 10 deletions(-)

diff --git a/libust/buffers.h b/libust/buffers.h
index 28faad7..db4b708 100644
--- a/libust/buffers.h
+++ b/libust/buffers.h
@@ -356,7 +356,7 @@ static __inline__ int ltt_reserve_slot(struct ust_trace *trace,
 	 * Perform retryable operations.
 	 */
 	/* FIXME: make this rellay per cpu? */
-	if (unlikely(__get_cpu_var(ltt_nesting) > 4)) {
+	if (unlikely(LOAD_SHARED(ltt_nesting) > 4)) {
 		local_inc(&buf->events_lost);
 		return -EPERM;
 	}
diff --git a/libust/serialize.c b/libust/serialize.c
index 011206a..4c23e8d 100644
--- a/libust/serialize.c
+++ b/libust/serialize.c
@@ -623,11 +623,11 @@ notrace void ltt_vtrace(const struct marker *mdata, void *probe_data,
 		return;
 
 	rcu_read_lock(); //ust// rcu_read_lock_sched_notrace();
-//ust//	cpu = smp_processor_id();
 	cpu = ust_get_cpu();
-//ust//	__get_cpu_var(ltt_nesting)++;
-	/* FIXME: should nesting be per-cpu? */
-	ltt_nesting++;
+
+	/* Force volatile access. */
+	STORE_SHARED(ltt_nesting, LOAD_SHARED(ltt_nesting) + 1);
+	barrier();
 
 	pdata = (struct ltt_active_marker *)probe_data;
 	eID = mdata->event_id;
@@ -712,8 +712,10 @@ notrace void ltt_vtrace(const struct marker *mdata, void *probe_data,
 		ltt_commit_slot(channel, buf, buf_offset, data_size, slot_size);
 		DBG("just commited event (%s/%s) at offset %ld and size %zd", mdata->channel, mdata->name, buf_offset, slot_size);
 	}
-//ust//	__get_cpu_var(ltt_nesting)--;
-	ltt_nesting--;
+
+	barrier();
+	STORE_SHARED(ltt_nesting, LOAD_SHARED(ltt_nesting) - 1);
+
 	rcu_read_unlock(); //ust// rcu_read_unlock_sched_notrace();
 }
 
diff --git a/libust/tracercore.c b/libust/tracercore.c
index 51f9a70..72c2887 100644
--- a/libust/tracercore.c
+++ b/libust/tracercore.c
@@ -42,7 +42,7 @@ void ltt_unlock_traces(void)
 
 //ust// DEFINE_PER_CPU(unsigned int, ltt_nesting);
 //ust// EXPORT_PER_CPU_SYMBOL(ltt_nesting);
-unsigned int ltt_nesting;
+__thread int ltt_nesting;
 
 int ltt_run_filter_default(void *trace, uint16_t eID)
 {
diff --git a/libust/tracercore.h b/libust/tracercore.h
index 3113383..2f3d7c4 100644
--- a/libust/tracercore.h
+++ b/libust/tracercore.h
@@ -42,8 +42,7 @@ struct ltt_traces {
 extern struct ltt_traces ltt_traces;
 
 /* Keep track of trap nesting inside LTT */
-//ust// DECLARE_PER_CPU(unsigned int, ltt_nesting);
-extern unsigned int ltt_nesting;
+extern __thread int ltt_nesting;
 
 typedef int (*ltt_run_filter_functor)(void *trace, uint16_t eID);
 //typedef int (*ltt_run_filter_functor)(void *, __u16);
-- 
2.34.1