Re: [PATCH 1/4] perf tools: Fix struct comm_str removal crash

[Jiri Olsa]

On Tue, Jul 17, 2018 at 10:49:40AM +0900, Namhyung Kim wrote:
> Hi Jiri,
> 
> On Mon, Jul 16, 2018 at 12:29:34PM +0200, Jiri Olsa wrote:
> > On Sun, Jul 15, 2018 at 10:08:27PM +0900, Namhyung Kim wrote:
> > 
> > SNIP
> > 
> > > > Because thread 2 first decrements the refcnt and only after then it
> > > > removes the struct comm_str from the list, the thread 1 can find this
> > > > object on the list with refcnt equls to 0 and hit the assert.
> > > > 
> > > > This patch fixes the thread 2 path, by removing the struct comm_str
> > > > FIRST from the list and only AFTER calling comm_str__put on it. This
> > > > way the thread 1 finds only valid objects on the list.
> > > 
> > > I'm not sure we can unconditionally remove the comm_str from the tree.
> > > It should be removed only if refcount is going to zero IMHO.
> > > Otherwise it could end up having multiple comm_str entry for a same
> > > name.
> > 
> > right, but it wouldn't crash ;-)
> > 
> > how about attached change, that actualy deals with the refcnt
> > race I'm running the tests now, seems ok so far
> 
> I think we can keep if the refcount is back to non-zero.  What about this?
> (not tested..)
> 
> 
> static struct comm_str *comm_str__get(cs)
> {
>     if (cs)
>         refcount_inc_no_warn(&cs->refcnt);  // should be added
>     return cs;
> }
> 
> static void comm_str__put(cs)
> {
>     if (cs && refcount_dec_and_test(&cs->refcnt)) {
>         down_write(&comm_str_lock);
>         /* might race with comm_str__findnew() */
>         if (!refcount_read(&cs->refcnt)) {
>             rb_erase(&cs->rb_node, &comm_str_root);
>             zfree(&cs->str);
>             free(cs);
>         }
>         up_write(&comm_str_lock);
>     }
> }

yea, it's more possitive than my patch
I'm testing attached patch, looks good so far

thanks,
jirka


---
diff --git a/tools/include/linux/refcount.h b/tools/include/linux/refcount.h
index 36cb29bc57c2..11e2be6f68a0 100644
--- a/tools/include/linux/refcount.h
+++ b/tools/include/linux/refcount.h
@@ -109,6 +109,14 @@ static inline void refcount_inc(refcount_t *r)
 	REFCOUNT_WARN(!refcount_inc_not_zero(r), "refcount_t: increment on 0; use-after-free.\n");
 }
 
+/*
+ * Pure refs increase without any chec/warn.
+ */
+static inline void refcount_inc_no_warn(refcount_t *r)
+{
+	atomic_inc(&r->refs);
+}
+
 /*
  * Similar to atomic_dec_and_test(), it will WARN on underflow and fail to
  * decrement when saturated at UINT_MAX.
diff --git a/tools/perf/util/comm.c b/tools/perf/util/comm.c
index 7798a2cc8a86..a2e338cf29d7 100644
--- a/tools/perf/util/comm.c
+++ b/tools/perf/util/comm.c
@@ -21,7 +21,7 @@ static struct rw_semaphore comm_str_lock = {.lock = PTHREAD_RWLOCK_INITIALIZER,}
 static struct comm_str *comm_str__get(struct comm_str *cs)
 {
 	if (cs)
-		refcount_inc(&cs->refcnt);
+		refcount_inc_no_warn(&cs->refcnt);
 	return cs;
 }
 
@@ -29,10 +29,12 @@ static void comm_str__put(struct comm_str *cs)
 {
 	if (cs && refcount_dec_and_test(&cs->refcnt)) {
 		down_write(&comm_str_lock);
-		rb_erase(&cs->rb_node, &comm_str_root);
+		if (refcount_read(&cs->refcnt) == 0) {
+			rb_erase(&cs->rb_node, &comm_str_root);
+			zfree(&cs->str);
+			free(cs);
+		}
 		up_write(&comm_str_lock);
-		zfree(&cs->str);
-		free(cs);
 	}
 }