CODEWITHSUNDEEP
javajvm
Alexander Arendar
Alexander Arendar

Reputation: 3435

Is there a way to know maximally reached JVM call stack depth for a particular program run?

I've been coding one recursive function today and the recursion depth depends on the input length.

I was wondering from the pure interest point of view, is there some way to monitor, probably in some JVM logs or elsewhere, what has been the maximal call stack depth during a particular program execution?

After some thinking I can imagine an analytical approach to calculate this approximately but that would be very time-intensive and would require quite good knowledge of JVM internals and bytecode.

JVM allows to configure the limit of a stack size memory, but I've never seen anything about how to get the actually reached limit and not in memory size units but the number of allocated stack frames.

Upvotes: 5

Views: 488

Answers (1)

apangin
apangin

Reputation: 98630

One can easily make JVMTI agent that will trace MethodEntry / MethodExit events and correspondingly increase or decrease stack depth counter. Here is an example of such agent. When the program ends, it will print the maximum recorded Java stack depth.

#include <jvmti.h>
#include <stdint.h>
#include <stdio.h>

static volatile int max_depth = 0;

static int adjust_stack_depth(jvmtiEnv *jvmti, int delta) {
    intptr_t depth = 0;
    (*jvmti)->GetThreadLocalStorage(jvmti, NULL, (void**)&depth);
    (*jvmti)->SetThreadLocalStorage(jvmti, NULL, (const void*)(depth + delta));
    return (int)depth;
}

void JNICALL MethodEntry(jvmtiEnv *jvmti, JNIEnv* jni, jthread thread, jmethodID method) {
    adjust_stack_depth(jvmti, +1);
}

void JNICALL MethodExit(jvmtiEnv *jvmti, JNIEnv* jni, jthread thread, jmethodID method,
                        jboolean was_popped_by_exception, jvalue return_value) {
    int depth = adjust_stack_depth(jvmti, -1);
    if (depth > max_depth) {
        max_depth = depth;  // TODO: replace with atomic CAS to avoid race condition
    }
}

JNIEXPORT jint JNICALL Agent_OnLoad(JavaVM *vm, char *options, void *reserved) {
    jvmtiEnv* jvmti;
    (*vm)->GetEnv(vm, (void**)&jvmti, JVMTI_VERSION_1_0);

    jvmtiCapabilities capabilities = {0};
    capabilities.can_generate_method_entry_events = 1;
    capabilities.can_generate_method_exit_events = 1;
    (*jvmti)->AddCapabilities(jvmti, &capabilities);

    jvmtiEventCallbacks callbacks = {0};
    callbacks.MethodEntry = MethodEntry;
    callbacks.MethodExit = MethodExit;
    (*jvmti)->SetEventCallbacks(jvmti, &callbacks, sizeof(callbacks));

    (*jvmti)->SetEventNotificationMode(jvmti, JVMTI_ENABLE, JVMTI_EVENT_METHOD_ENTRY, NULL);
    (*jvmti)->SetEventNotificationMode(jvmti, JVMTI_ENABLE, JVMTI_EVENT_METHOD_EXIT, NULL);

    return 0;
}

JNIEXPORT void JNICALL Agent_OnUnload(JavaVM *vm) {
    printf("Max stack depth = %d\n", max_depth);
}

Compile:

gcc -fPIC -shared -I $JAVA_HOME/include -I $JAVA_HOME/include/linux -o libmaxdepth.so maxdepth.c

Run:

java -agentpath:/path/to/libmaxdepth.so MyProgram

However, tracing each method entry and exit is very expensive. A less accurate, but much more efficient alternative would be a sampling profiler which periodically records a stack trace of a running thread, e.g. async-profiler or Java Flight Recorder.

Upvotes: 4

Related Questions