Re: RFR: 8227269: Slow class loading when running JVM in debug mode

[Chris Plummer]

Hi Roman,

I assume JVMTI maintains separate tagging data for each agent so having 
two agents doing tagging won't result in confusion. I didn't actually 
find this in the spec. Would be nice to confirm that it is the case. 
However, your implementation does seem to conflict with other uses of 
tagging in the debug agent:

 - During the execution of ObjectReference.ReferringObjects, the object 
being checked is tagged. If this happens to be a Class instance, the tag 
you setup will end up being cleared.

- During the execution of VirtualMachine.InstanceCounts, each Class 
instance being counted is tagged. So that means your tag is cleared for 
any Class passed to this API

- SetTag is used in commonRef.c. I believe any Object for which an 
objectID is created and sent to the front end (debugger), a weakref to 
that object is created and tagged with a RefNode*. So you will have many 
Weakref objects with tags. When these are freed, they are passed to 
cbTrackingObjectFree() and these tags are incorrectly added to 
deletedSignatures(). This means you end up treating a RefNode* as a 
char* in synthesizeUnloadEvent(), and a ClassUnload event gets created 
with garbage for the classname. I also think this could cause issues 
when eventually this RefNode* is passed to jvmtiDeallocate(). However, I 
think you have a bug where you never actually free up signatures for 
Classes that get unloaded. Only signatures for loaded classes seem to 
get deleted, and that is done when the agent detaches.

What would cause classTrack_addPreparedClass() to be called for a Class 
you've already seen? I don't understand the need for the "tag != 0l" check.

thanks,

Chris

On 3/20/20 12:52 PM, Chris Plummer wrote:
On 3/20/20 8:30 AM, Roman Kennke wrote:
I believe I came up with a much simpler solution that also solves the
problems of the existing one, and the ones I proposed earlier.

It turns out that we can take advantage of the fact that we can use
*anything* as tags in JVMTI, even pointers to stuff (this is explicitely
mentioned in the JVMTI spec). This means we can simply stick a pointer
to the signature of a class into the tag, and pull it out again when we
get notified that the class gets unloaded.

This means we don't need an extra data-structure to keep track of
classes and signatures, and it also makes the story around locking
*much* simpler. Performance-wise this is O(1), i.e. no scanning of all
classes needed (as in the current implementation) and no searching of
table needed (like in my previous attempts).

Please review this new revision:
http://cr.openjdk.java.net/~rkennke/JDK-8227269/webrev.06/
I'll have a look at this.

(Notice that there still appears to be a performance bottleneck with
class-unloading when an actual debugger is attached. This doesn't seem
to be related to the classTrack.c implementation though, but looks like
a consequence of getting all those class-unload notifications over the
wire. My testcase generates 1000s of them, and it's clogging up the
buffers.)
At least this is only a one-shot hit when the classes are unloaded, 
and the performance hit is based on the number of classes being 
unloaded. The main issue is happening every GC, and is O(n) where n is 
the number of loaded classes.
I am not sure why jdb needs to enable class-unload listener always. A
simple hack disables it, and performance is brilliant, even when jdb is
attached:
http://cr.openjdk.java.net/~rkennke/disable-jdk-class-unload.patch
This is JDI, not jdb. It looks like it needs ClassUnload events so it 
can maintain typesBySignature, which is used by public APIs like 
allClasses(). So we have caching of loaded classes both in the debug 
agent and in JDI.

Chris
But this is not in the scope of this bug.)

Roman


On 3/16/20 8:05 AM, serguei.spit...@oracle.com wrote:
Sorry, forgot to complete my comments at the end (see below).


On 3/15/20 23:57, serguei.spit...@oracle.com wrote:
Hi Roman,

Thank you for the update and sorry for the latency in review.

Some comments are below.

http://cr.openjdk.java.net/~rkennke/JDK-8227269/webrev.05/src/jdk.jdwp.agent/share/native/libjdwp/classTrack.c.frames.html 


87 cbTrackingObjectFree(jvmtiEnv* jvmti_env, jlong tag)
   88 {
89 debugMonitorEnter(deletedSignatureLock);
90 if (currentClassTag == -1) {
91 // Class tracking not initialized, nobody's interested
92 debugMonitorExit(deletedSignatureLock);
93 return;
   94     }
Just a question:
   Q1: Should the ObjectFree events be disabled for the jvmtiEnv 
that does
       the class tracking if class tracking has not been initialized?

70 static jlong currentClassTag; I'm thinking if the name is better to
be something like: lastClassTag or highestClassTag.

99 KlassNode* klass = *klass_ptr;
100 102 while (klass != NULL && klass->klass_tag != tag) { 103
klass_ptr = &klass->next; 104 klass = *klass_ptr;
105 } 106 if (klass != NULL || klass->klass_tag != tag) { // klass not
found - ignore.
107 debugMonitorExit(deletedSignatureLock);
108 return;
  109     }
  It seems to me, something is wrong in the condition at L106 above.
  Should it be? :
     if (klass == NULL || klass->klass_tag != tag)

  Otherwise, how can the second check ever work correctly as the 
return
will always happen when (klass != NULL)?

  There are several places in this file with the the indent:
90 if (currentClassTag == -1) {
91 // Class tracking not initialized, nobody's interested
92 debugMonitorExit(deletedSignatureLock);
93 return;
   94     }
  ...
152 if (currentClassTag == -1) {
153 // Class tracking not initialized yet, nobody's interested
154 debugMonitorExit(deletedSignatureLock);
155 return;
  156     }
  ...
161 if (error != JVMTI_ERROR_NONE) {
162 EXIT_ERROR(error, "Unable to GetTag with class trackingEnv");
  163     }
164 if (tag != 0l) {
165 debugMonitorExit(deletedSignatureLock);
166 return; // Already added
  167     }
  ...
281 cleanDeleted(void *signatureVoid, void *arg)
282 {
283 char* sig = (char*)signatureVoid;
284 jvmtiDeallocate(sig);
285 return JNI_TRUE;
  286 }
  ...
  291 void
  292 classTrack_reset(void)
  293 {
294 int idx;
295 debugMonitorEnter(deletedSignatureLock);
296
297 for (idx = 0; idx < CT_SLOT_COUNT; ++idx) {
298 KlassNode* node = table[idx];
299 while (node != NULL) {
300 KlassNode* next = node->next;
301 jvmtiDeallocate(node->signature);
302 jvmtiDeallocate(node);
303 node = next;
304 }
305 }
306 jvmtiDeallocate(table);
307
308 bagEnumerateOver(deletedSignatureBag, cleanDeleted, NULL);
309 bagDestroyBag(deletedSignatureBag);
310
311 currentClassTag = -1;
312
313 (void)JVMTI_FUNC_PTR(trackingEnv,DisposeEnvironment)(trackingEnv);
314 trackingEnv = NULL;
315
316 debugMonitorExit(deletedSignatureLock);

Could you, please, fix several comments below?
63 * The JVMTI tracking env to keep track of klass tags, for 
class-unloads
  The comma is not needed.
  Would it better to replace: klass tags => klass_tag's ?


73 * Lock to keep table, currentClassTag and deletedSignatureBag
consistent
  Maybe: Lock to guard ... or lock to keep integrity of ...

84 * Callback when classes are freed, Finds the signature and
remembers it in deletedSignatureBag. Would be better to use words like
"store" or "record", "Find" should not start from capital letter:
Invoke the callback when classes are freed, find and record the
signature in deletedSignatureBag.

96 // Find deleted KlassNode 133 // Class tracking not initialized,
nobody's interested 153 // Class tracking not initialized yet,
nobody's interested 158 /* Check this is not a duplicate */ Missed dot
at the end. 106 if (klass != NULL || klass->klass_tag != tag) { //
klass not found - ignore. In opposite, dot is not needed as the
comment does not start from a capital letter. 111 // At this point we
have the KlassNode corresponding to the tag
112 // in klass, and the pointer to it in klass_node.
  The comment above can be better. Maybe, something like:
    " At this point, we found the KlassNode matching the klass 
tag(and it is
linked).

113 // Remember the unloaded signature.
  Better: Record the signature of the unloaded class and unlink it.

Thanks,
Serguei

Thanks,
Serguei

On 3/9/20 05:39, Roman Kennke wrote:
Hello all,

Can I please get reviews of this change? In the meantime, we've done
more testing and also field-/torture-testing by a customer who is 
happy
now. :-)

Thanks,
Roman


Hi Serguei,

Thanks for reviewing!

I updated the patch to reflect your suggestions, very good!
It also includes a fix to allow re-connecting an agent after 
disconnect,
namely move setup of the trackingEnv and deletedSignatureBag to
_activate() to ensure have those structures after re-connect.

http://cr.openjdk.java.net/~rkennke/JDK-8227269/webrev.05/

Let me know what you think!
Roman

Hi Roman,

Thank you for taking care about this scalability issue!

I have a couple of quick comments.

http://cr.openjdk.java.net/~rkennke/JDK-8227269/webrev.04/src/jdk.jdwp.agent/share/native/libjdwp/classTrack.c.frames.html 


72 /*
73 * Lock to protect deletedSignatureBag
74 */
75 static jrawMonitorID deletedSignatureLock; 76 77 /*
78 * A bag containing all the deleted classes' signatures. Must be
accessed under
79 * deletedTagLock,
   80  */
81 struct bag* deletedSignatureBag;

   The comments contradict to each other.
   I guess, the lock name at line 79 has to be deletedSignatureLock
instead of deletedTagLock.
   Also, comma at the end must be replaced with dot.


101 // Tag not found? Ignore.
102 if (klass == NULL) {
103 debugMonitorExit(deletedSignatureLock);
104 return;
105 }
  106
107 // Scan linked-list.
108 jlong found_tag = klass->klass_tag;
109 while (klass != NULL && found_tag != tag) {
110 klass_ptr = &klass->next;
111 klass = *klass_ptr;
112 found_tag = klass->klass_tag;
  113     }
114
115 // Tag not found? Ignore.
116 if (found_tag != tag) {
117 debugMonitorExit(deletedSignatureLock);
118 return;
  119     }


  The code above can be simplified, so that the lines 101-105 
are not
needed anymore.
  It can be something like this:

// Scan linked-list.
while (klass != NULL && klass->klass_tag != tag) {
klass_ptr = &klass->next;
klass = *klass_ptr;
      }
if (klass == NULL || klass->klass_tag != tag) { // klass not 
found - ignore.
debugMonitorExit(deletedSignatureLock);
return;
      }

It will take more time when I get a chance to look at the rest.


Thanks,
Serguei




On 12/21/19 13:24, Roman Kennke wrote:
Here comes an update that resolves some races that happen when
disconnecting an agent. In particular, we need to take the lock on
basically every operation, and also need to check whether or not
class-tracking is active and return an appropriate result (e.g. 
an empty
list) when we're not.

Updated webrev:
http://cr.openjdk.java.net/~rkennke/JDK-8227269/webrev.04/

Thanks,
Roman


So, here comes the O(1) implementation:

- Whenever a class is 'prepared', it is registered with a tag, 
and we
set-up a listener to get notified when it is unloaded.
- Prepared classes are kept in a datastructure that is a 
table, which
each entry being the head of a linked-list of KlassNode*. The 
table is
indexed by tag % slot-count, and then simply prepend the new 
KlassNode*.
This is O(1) operation.
- When we get notified of unloading a class, we look up the 
signature of
the reported tag in that table, and remember it in a bag. The 
KlassNode*
is then unlinked from the table and deallocated. This is ~O(1) 
operation
too, depending on the depth of the table. In my testcase which 
hammered
the code with class-loads and unloads, I usually see depths of 
like 2-3,
but not usually more. It should be ok.
- when processUnloads() gets called, we simply hand out that 
bag, and
allocate a new one.
- I also added cleanup-code in classTrack_reset() to avoid 
leaking the
signatures and KlassNode* etc when debug agent gets detached 
and/or
re-attached (was missing before).
- I also added locks around data-structure-manipulation (was 
missing
before).
- Also, I only activate this whole process when an actual 
listener gets
registered on EI_GC_FINISH. This seems to happen right when 
attaching a
jdb, not sure why jdb does that though. This may be something 
to improve
in the future?

In my tests, the performance of class-tracking itself looks 
really good.
The bottleneck now is clearly actual synthesizing the 
class-unload
events. I don't see how this can be helped when the debug 
agent asks for it?

Updated webrev:
http://cr.openjdk.java.net/~rkennke/JDK-8227269/webrev.03/

Please let me know what you think of it.

Thanks,
Roman


Alright, the perfectionist in me got me. I am implementing 
the even more
efficient ~O(1) class tracking. Please hold off reviewing for 
now.

Thanks,Roman

  Hi Chris,
I'll have a look at this, although it might not be for a 
few days. In
the meantime, maybe you can describe your new 
implementation in
classTrack.c so it's easier to look through the changes.
Sure.

The purpose of this class-tracking is to be able to 
determine the
signatures of unloaded classes when GC/class-unloading 
happened, so that
we can generate the appropriate JDWP event.

The current implementation does so by maintaining a table of 
currently
prepared classes by building that table when classTrack is 
initialized,
and then add new classes whenever a class gets loaded. When 
unloading
occurs, that cache is rebuilt into a new table, and compared 
with the
old table, and whatever is in the old, but not in the new 
table gets
returned. The problem is that when GCs happen frequently 
and/or many
classes get loaded+unloaded, this amounts to 
O(classCount*gcCount)
complexity.

The new implementation keeps a linked-list of prepared 
classes, and also
tracks unloads via the listener cbTrackingObjectFree(). 
Whenever an
unload/GC occurs, the list of prepared classes is scanned, 
and classes
that are also in the deletedTagBag are unlinked (thus 
maintaining the
prepared-classes-list) and its signature put in the list 
that gets returned.

The implementation is not perfect. In order to determine 
whether or not
a class is unloaded, it needs to scan the deletedTagBag. 
That process is
therefore still O(unloadedClassCount). The assumption here 
is that
unloadedClassCount << classCount. In my experiments this 
seems to be
true, and also reasonable to expect.

(I have some ideas how to improve the implementation to 
~O(1) but it
would be considerably more complex: have to maintain a 
(hash)table that
maps tags -> KlassNode*, unlink them directly upon unload, 
and build the
unloaded-signatures list there, but I don't currently see 
that it's
worth the effort).

In addition to all that, this process is only activated when 
there's an
actual listener registered for EI_GC_FINISH.

Thanks,
Roman


Chris

On 12/18/19 5:05 AM, Roman Kennke wrote:
Hello all,

Issue:
https://bugs.openjdk.java.net/browse/JDK-8227269

I am proposing what amounts to a rewrite of classTrack.c. 
It avoids
throwing away the class cache on GC, and instead keeps 
track of
loaded/unloaded classes one-by-one.

In addition to that, it avoids this whole dance until an 
agent
registers interest in EI_GC_FINISH.

Webrev:
http://cr.openjdk.java.net/~rkennke/JDK-8227269/webrev.01/

Testing: manual testing of provided test scenarios and 
timing.

Eg with the testcase provided here:
https://bugzilla.redhat.com/show_bug.cgi?id=1751985

I am getting those numbers:
unpatched: no debug: 84s with debug: 225s
patched:   no debug: 85s with debug: 95s

I also tested successfully through jdk/submit repo

Can I please get a review?

Thanks,
Roman