Re: RFR(XXS): 8162945: HotspotDiagnosticMXBean getFlags erroneously reports OutOfMemory

[Daniel D. Daugherty]

On 8/2/16 6:58 AM, Markus Gronlund wrote:

Greetings,

Please review this small fix to address some new issues seen in testing where OOM is erroneously being reported:

Bug: https://bugs.openjdk.java.net/browse/JDK-8162945

Changeset:

# HG changeset patch

# User mgronlun

# Date 1470141649 -7200

#      Tue Aug 02 14:40:49 2016 +0200

# Node ID e06e6474ff5f5798f9249aeaa6b33ec6aa021563

# Parent 9672159305d72f5dd430a3edd4b97c4e5bc816e0

[mq]: 8162945

diff --git a/src/jdk.management/share/native/libmanagement_ext/Flag.c b/src/jdk.management/share/native/libmanagement_ext/Flag.c

--- a/src/jdk.management/share/native/libmanagement_ext/Flag.c

+++ b/src/jdk.management/share/native/libmanagement_ext/Flag.c

@@ -142,7 +142,7 @@

             continue;

         }

-        if (valueObj == NULL) {

+ if (valueObj == NULL && !(globals[i].type == JMM_VMGLOBAL_TYPE_JSTRING)) {

Might be too late, but I would prefer:

if (valueObj == NULL && globals[i].type != JMM_VMGLOBAL_TYPE_JSTRING) {

which I think is more clear.

Dan

free(globals);

JNU_ThrowOutOfMemoryError(env, 0);

             return 0;

Summary:

OOM’s have manifested after the following change was integrated:

8162524: src/jdk.management/share/native/libmanagement_ext/Flag.c doesn't handle JNI exceptions

The following check was then added to jdk\src\jdk.management\share\native\libmanagement_ext\flags.c:

        if (valueObj == NULL) {
            free(globals);
            JNU_ThrowOutOfMemoryError(env, 0);
            return 0;
        }

However, valueObj is not always the direct result of a failed JNI allocation, for example:

        case JMM_VMGLOBAL_TYPE_JSTRING:
            valueObj = globals[i].value.l;

The valueObj here (a JSTRING) is coming a VM allocation, more specifically from :

services/management.cpp jmm_GetVMGlobals()
...
add_global_entry()

} else if (flag->is_ccstr()) {

Handle str = java_lang_String::create_from_str(flag->get_ccstr(), CHECK_false); <<--

    global->value.l = (jobject)JNIHandles::make_local(env, str());
    global->type = JMM_VMGLOBAL_TYPE_JSTRING;

For certain ccstr flags, such as the "HeapDumpFlag" for example, the ccstr() is NULL (i.e. not set).

So returning a NULL is fine here, the JNI NULL check validation needs to take this into account.

Thanks

Markus