RFR (M) 8080406: VM_GetOrSetLocal doesn't check local slot type against requested type
JC Beyler
jcbeyler at google.com
Wed Nov 7 17:30:12 UTC 2018
Hi Serguei,
I like the change you made in the VM, it is more clean and easier to see
what is going on :)
A few nits:
http://cr.openjdk.java.net/~sspitsyn/webrevs/2018/8080406-jvmti-get-local.2/test/hotspot/jtreg/vmTestbase/nsk/jvmti/unit/GetLocalVariable/getlocal003/getlocal003.cpp.udiff.html
- There is a type: "Succes:" line 342.
- There is this check:
272 if (jvmti == NULL) {
273 return;
274 }
but the same check line 109 fails the test and prints a message.
http://cr.openjdk.java.net/~sspitsyn/webrevs/2018/8080406-jvmti-get-local.2/test/hotspot/jtreg/vmTestbase/nsk/jvmti/unit/GetLocalVariable/getlocal004/getlocal004.cpp.html
- mid is only used in one method
(Java_nsk_jvmti_unit_GetLocalVariable_getlocal004_getMeth) and
technically it doesn't get anything, it checks if the method is there.
- not sure you wanted the capabilities to be static in the method
The macros look good in
http://cr.openjdk.java.net/~sspitsyn/webrevs/2018/8080406-jvmti-get-local.2/test/hotspot/jtreg/serviceability/jvmti/GetLocalVariable/libGetLocalVars.cpp.html.
For reference,
I think templates are fine, I've used them in the ExceptionJniWrapper
and that has worked there. Templates get initialized and generate a
separate method for each type needed. So it is doing the
same as what you've done here but it does it by hand (except that you
don't have to pass the method in so things are a bit easier at the
call-sites so that is a win in this case as you said :))
- Note that the slot variables ByteSlot could be in the method
Java_GetLocalVars_testLocals
- You gain nothing of really calling the variables in the testLocals
method static, the method is called once...
Thanks,
Jc
On Wed, Nov 7, 2018 at 12:04 AM serguei.spitsyn at oracle.com <
serguei.spitsyn at oracle.com> wrote:
> Hi Jc,
>
> The updated version of webrev:
>
> http://cr.openjdk.java.net/~sspitsyn/webrevs/2018/8080406-jvmti-get-local.2/
>
> I've resolved most of your comments.
> I used macro definitions instead of templates you suggested.
> Two reasons for it:
> - not sure how templates depends on the compiler versions over various
> platforms
> - macro definitions allow to make definitions more complex but not the
> calls
>
>
> Applied the same cleanups to both old tests:
> getlocal003/getlocal003.cpp and getlocal004/getlocal004.cpp
>
> Also, this update includes some change in the VM_GetOrSetLocal
> implementation.
> It is to move the call to check_slot_type_no_lvt() from the doit() to
> prologue().
> So, now the logic is more consistent and clear.
>
> Please, let me know what do you think.
> I hope that Vladimir I. will have a chance to look at the VM changes.
> Also, one more review is needed on the tests side.
>
> Thanks,
> Serguei
>
>
> On 11/6/18 17:13, serguei.spitsyn at oracle.com wrote:
>
> Hi Jc,
>
> Thank you a lot for the code review!
>
> On 11/6/18 9:22 AM, JC Beyler wrote:
>
> Hi Serguei,
>
> I saw this code:
> + BasicType next_slot_type = locals->at(_index + 1)->type();
>
> I think we are not worried about going out of bounds due to the work done
> in the check_slot_type, which is done in doit_prologue:
> 643 if (_index < 0 || _index + extra_slot >=
> method_oop->max_locals()) {
>
> Should we put an assert though in case?
>
>
> It is a good suggestion.
> But I'm thinking about moving the check_slot_type_no_lvt call into the
> check_slot_type().
> Then most likely this assert is not going to be needed.
>
>
> -> For the test
> http://cr.openjdk.java.net/~sspitsyn/webrevs/2018/8080406-jvmti-get-local.1/test/hotspot/jtreg/serviceability/jvmti/GetLocalVariable/libGetLocalVars.cpp.html
> :
> - why not use the TranslateError from
> test/hotspot/jtreg/vmTestbase/nsk/share/jvmti/JVMTITools.cpp
>
>
> We have several other serviceability/jvmti tests that use the same.
> It is not good to use the TranslateError from the the vmTestbase library.
> The TranslateError would better to be copied to the global test library.
> Then the TranslateError macro definition would be removed for all of these
> cases as one action.
>
> - You do this in the test:
> 371 if (!caps.can_access_local_variables) {
> 372 return;
> 373 }
>
> But if you cannot access local variables, on the load of the agent you
> would return JNI_ERR which I believe fails the JVM loading, no? Hence is
> this even needed?
>
>
> Agreed - removed it.
>
>
> - We could get rid of the caps global variable
> - Talking about global variables: I think you can get rid of all of
> them: jvmti is always passed as an argument, mid is not used except to see
> if the method can be found, the slots are used only locally in one method
>
>
> - Why is it PASSED but STATUS_FAILED?
>
>
> Nice catch, fixed.
>
>
> - With templates, you could simplify a bit the repetitive tests it seems:
>
> template<typename T>
> jint testGetter(jvmtiEnv *jvmti, jthread thr, jint depth, jint slot, const
> char* exp_type,
> jvmtiError (jvmtiEnv::*getter)(jthread, jint, jint, T*),
> const char* getter_name) {
> T val = 0;
> jvmtiError err = (jvmti->*getter)(thr, depth, slot, &val);
>
> printf(" %s: %s (%d)\n", getter_name, TranslateError(err), err);
> if (err != JVMTI_ERROR_NONE) {
> printf(" FAIL: %s failed to get value from a local %s\n", getter_name,
> exp_type);
> result = STATUS_FAILED;
> } else {
> printf(" %s got value from a local %s as expected\n", getter_name,
> exp_type);
> }
> }
>
> and then your code:
>
> 259 test_int(jvmti, thr, depth, slot, "byte");
> 260 test_long_inv_slot(jvmti, thr, depth, slot, "byte");
> 261 test_float(jvmti, thr, depth, slot, "byte");
>
> Could become:
> testGetter(jvmti, thr, depth, slot, "byte", &jvmtiEnv::GetLocalInt,
> "GetLocalInt");
> testGetter(jvmti, thr, depth, slot, "byte", &jvmtiEnv::GetLocalLong,
> "GetLocalLong");
> testGetter(jvmti, thr, depth, slot, "byte", &jvmtiEnv::GetLocalFloat,
> "GetLocalFloat");
>
> and by analogy, you could use templates for the invalid and the mismatch
> types.
>
> That way, there really are three methods written with templates and we are
> just calling them with different types. I checked that this seems to work
> with gnu++98 so it should work for OpenJDK.
>
>
> Thank you for the suggestion.
> However, I wouldn't want to go this path.
> I'll check if a macro can be used here in a simple way.
>
> For
> http://cr.openjdk.java.net/~sspitsyn/webrevs/2018/8080406-jvmti-get-local.1/test/hotspot/jtreg/vmTestbase/nsk/jvmti/unit/GetLocalVariable/getlocal003/getlocal003.cpp.html
> :
> - I have the same remarks for the global variables but it is trickier
> because it's a more massive rewrite of the test there it seems
>
>
> I've fixed both getlocal003.cpp and getlocal004.cpp.
>
> - The code you added seems to also be able to be templatized via
> something like:
>
> template<typename T>
> jint testGetter(jvmtiEnv *jvmti, jthread thr, jint slot, jint depth, T*
> value,
> jvmtiError (jvmtiEnv::*getter)(jthread, jint, jint, T*),
> const char* getter_name,
> char sig,
> char expected_sig) {
> jvmtiError err = (jvmti->*getter)(thr, slot, depth, value);
> printf(" %s: %s (%d)\n", getter_name, TranslateError(err), err);
> if (err != JVMTI_ERROR_NONE && sig == expected_sig) {
> printf("FAIL: %s failed to get value of long\n", getter_name);
> result = STATUS_FAILED;
> } else if (err != JVMTI_ERROR_TYPE_MISMATCH && sig != expected_sig) {
> printf("FAIL: %s did not return JVMTI_ERROR_TYPE_MISMATCH for
> non-long\n", getter_name);
> result = STATUS_FAILED;
> }
> }
>
>
> Thanks.
> Please, see my reply above.
>
> I'll send an updated webrev in a separate email.
>
> Thanks!
> Serguei
>
> Apart from that, it looks good to me, these are mostly style choices I
> suppose and trying to reduce code repetitiveness :)
> Jc
>
> On Mon, Nov 5, 2018 at 9:36 PM serguei.spitsyn at oracle.com <
> serguei.spitsyn at oracle.com> wrote:
>
>> Please, review a fix for:
>> https://bugs.openjdk.java.net/browse/JDK-8080406
>>
>> Webrev:
>>
>> http://cr.openjdk.java.net/~sspitsyn/webrevs/2018/8080406-jvmti-get-local.1/
>>
>>
>> Summary:
>> The JVMTI GetLocal<Type>/SetLocal<Type> implementation type checking is
>> based
>> on LVT (Local Variable Table) content. But there is almost no type
>> check if LVT
>> is not present in class file. This fix is an attempt to fill in the gap.
>> When LVT is absent, one issue is that just 3 types are available in the
>> StackValueCollectionfor locals at runtime:
>> - T_OBJECT: if local is an object
>> - T_INT: if local is a primitive type
>> - T_CONFLICT: if local is not valid at current location
>> So there is no way to distinguish primitive types unless the requested
>> type
>> occupies two slots and actual second slot is not T_INT or is out of
>> locals area.
>>
>> Testing:
>> Tested locally on Linux-x64 with:
>> - 1 new jtreg test:
>> hotspot/jtreg/serviceability/jvmti/GetLocalVariable
>> - 2 nsk jtreg tests:
>> hotspot/jtreg/vmTestbase/nsk/jvmti/unit/GetLocalVariable
>> - 2 nsk jtreg tests:
>> hotspot/jtreg/vmTestbase/nsk/jvmti/GetLocalVariable
>> - 4 nsk jtreg tests:
>> hotspot/jtreg/vmTestbase/nsk/jvmti/SetLocalVariable
>>
>> In progress:
>> The same as above but with mach5 in different configs.
>>
>> Thanks,
>> Serguei
>>
>
>
> --
>
> Thanks,
> Jc
>
>
>
>
--
Thanks,
Jc
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openjdk.java.net/pipermail/serviceability-dev/attachments/20181107/8acf71be/attachment-0001.html>
More information about the serviceability-dev
mailing list