Hi, Sundar,
I agree with you in general, but I want to extend SA in some case.
I just want to add gateway to SA.
So I proposed to add some classes in jdk.hotspot.agent for exporting.
It does not mean to maintain all SA classes for it.
I think that the user who want to use it should use it by own risk.
Thus my proposal needs to add --add-exports to build SA plugin.
As another approach, we can might custom CLHSDB command with
@FunctionalInterface.
For example, if CLHSDB provides "add_function [jar] [FQCN] [command name]" for
it,
the user add new command(s) through it by own risk.
It does not affect SA code. I believe it is not to expose "Unsupported Java
API".
Yasumasa
On 2019/12/20 12:25, sundararajan.athijegannat...@oracle.com wrote:
Hi,
I am going to reiterate. This will lead to maintenance nightmare! Any "Unsupported Java
API" is still an API! Remember Unsafe? Once blessed in any form, it is very difficult to
remove. Exposing hotspot VM internals as a Java API is very bad idea. No, not even as
"unsupported API".
Exposing SA to scripts was a different beast. Scripting languages are dynamically typed. It is quite common
in scripting world to check the existence of an attribute & then use it (if (!obj["foo"]) /
if (typeof obj["foo"] == 'function') kind of code). That'd be extremely painful in a statically
typed language like Java (reflection or method/var handles?). Platform debuggers like dbx support a
scripting language along with access to the data structures from the target process. SA scripting was
inspired from that model. Also scripts were never meant to be written & maintained for long time! Most
scripts were expected to be written for a specific debugging exercise/session (thrown away).
-Sundar
On 20/12/19 6:38 am, Yasumasa Suenaga wrote:
Hi Chris,
Can we treat (part of) jdk.hotspot.agent like jdk.unsupported module?
jdk.unsupported exports unspec'd API like Unsafe.
If we do so, we might need to separate SA API into exported class and internal
class.
I've proposed to export all SA packages in JDK-8157947, but it was rejected.
Thanks,
Yasumasa
On 2019/12/20 1:28, Chris Plummer wrote:
Hi Yasumasa,
I've had similar thoughts about how to extend clhsdb. Why not export everything
since what we would export is not part of a spec, and the javascript support
had the same issue of potentially breaking when the SA API changed. But maybe
this type of unspec'd API exporting is considered bad policy. I'm not sure.
I'll let the API spec guru's comment on that aspect of it.
thanks,
Chris
On 12/19/19 12:17 AM, Yasumasa Suenaga wrote:
Hi,
I think we can provide API for SA as following:
Patch: http://cr.openjdk.java.net/~ysuenaga/sa-api/webrev/
Plugin examples:
browse: http://cr.openjdk.java.net/~ysuenaga/sa-api/plugin-examples/
download: http://cr.openjdk.java.net/~ysuenaga/sa-api/plugin-examples.tar.gz
I think JS plugin (loading via `jsload` CLHSDB command) was supported "AS IS".
If HotSpot and/or SA code is changed, the user should follow it if need.
SA is not part of Java SE. We need not to maintain SA API when it happen IMHO.
The user who want to expand SA features (includes me!) should have responsible
for it.
So I did not expose jdk.hotspot.agent module - the user need to build with
--add-exports.
My proposal can write SA plugins with pure Java. So we don't need to depend on
script engine.
Comments are welcome.
Thanks,
Yasumasa
On 2019/12/11 21:47, sundararajan.athijegannat...@oracle.com wrote:
Effectively you're asking for SA as API. I don't think that is a good idea.
That implies supporting hotspot data structures as Java *API*. That will be
maintainability nightmare - we've to keep tracking hotspot data structures in
SA code. That itself is problematic. API would be next level nightmare.
-Sundar
On 11/12/19 11:57 am, Yasumasa Suenaga wrote:
Hi,
IMHO we need to export all packages in SA if we do not provide new API for SA.
sa.js in jdk.hotspot.agent could access all SA classes until JDK 8 (before
Jigsaw), so we could make various functions if we need.
OTOH we cannot know what classes are needed by the SA users. All packages in
jdk.hotspot.agent module provides features, and they require other packages.
For example, sun.jvm.hotspot.oops.Oop requires sun.jvm.hotspot.types, and it
requires sun.jvm.hotspot.debugger .
It is difficult to track and to export minimally.
(I worked for it in JDK-8157947, but I gave up...)
Thus I guess it is a big challenge to export SA classes without refactoring.
If we provide new API for SA plugin, I guess we need to work some refactoring.
Yasumasa
On 2019/12/11 15:00, Chris Plummer wrote:
On 12/10/19 9:56 PM, Yasumasa Suenaga wrote:
On 2019/12/11 14:39, Krystal Mok wrote:
Hi Yasumasa,
That's a very nice idea. Basically what you're asking for is exposing the
Command interface [1] so that plugins can implement it and get dynamically
loaded / registered into CLHSDB / HSDB, right?
Yes, but we also need proxy API to access internal SA objects e.g. CodeCache,
JavaThread, TypeDataBase, etc...
Yes, or export them. I should have read this email before posting my previous
one.
Chris
Yasumasa
[1]:
http://hg.openjdk.java.net/jdk/jdk/file/c71ec1f09f21/src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/CommandProcessor.java#l246
- Kris
On Tue, Dec 10, 2019 at 9:33 PM Yasumasa Suenaga <suen...@oss.nttdata.com
<mailto:suen...@oss.nttdata.com>> wrote:
Hi Chris,
It's a sad proposal, but I agree with you. To maintain SA in JS is
difficult since Jigsaw.
However I want SA to implement pluggable feature.
I use custom script to list compiled codes in CodeCache.
I guess other troubleshooters also want similar feature (via jsload) in
future if they encounter JVM crash.
Thanks,
Yasumasa
On 2019/12/11 11:52, Chris Plummer wrote:
> Hi,
>
> I like to propose the removal of SA javascript support. Few people even
realize this support exists, and hopefully even fewer are using it since I'd like
to remove it. Since I'm new to this myself, let me first explain what I know about
it's existence, and then explain why I want to remove it.
>
> If you run "jhsdb clhsdb", there are jsload and jseval commands. Don't
look for them in anything post JDK 8. I'll explain why later. jsload is used to load a
javascript file. In that file you can register new clhsdb commands that are written in
javascript. You can also evaluate javascript using the jseval command. Some of this is
explained in [1], which is the only place I can find any reference to this support. It does
not appear to be officially supported, nor is there any oracle provided documentation.
>
> There also appear to be a few clhsdb commands that are written in javascript.
Doing a grep for "registerCommand" in sa.js shows the following:
>
> registerCommand("class", "class name", "jclass");
> registerCommand("classes", "classes", "jclasses");
> registerCommand("dumpclass", "dumpclass { address | name } [ directory ]",
"dclass");
> registerCommand("dumpheap", "dumpheap [ file ]", "dumpHeap");
> registerCommand("mem", "mem address [ length ]", "printMem");
> registerCommand("sysprops", "sysprops", "sysProps");
> registerCommand("whatis", "whatis address", "printWhatis");
>
> Once again, don't go looking for these in anything newer than JDK8. You
won't find them. Again the only documentation I can fine is [1].
>
> The other use of Javascript is the SOQL command (Simple Object Query
Language), a tool used to query the heap, and also the JSDB command. The only SOQL
documentation I could find is the blog reference [2]. I could not find HSDB
documentation, but I believe is is a javascript support for looking at hotspot. So
once again, neither of these seem to be officially supported or documented.
>
> The real purpose of the email is to propose removal of this support.
Here are the reasons:
>
> (1) It's broken, and has been since 9. See [3]. This is why you don't
see the javascript related commands in clhsdb. Javascript fails to initialize, so
none of the javascript related commands are registered.
> (2) Nashorn is deprecated and will be removed eventually.
> (3) We have very little understanding of the javascript support.
> (4) No resources to work on it (unless there is a community volunteer).
> (5) Very questionable value (lack of users). The fact this support has
been broken since JDK 9 and no bug was filed until I did so this week is a good
indication of that. Another is that there are no other SA Javascript related bugs
filed. Lastly, the lack of any official documentation and only minimal mention of
it on the web is another good indication of it's (lack of) value.
>
> Also, regarding the 7 commands listed above that would be lost (but
currently don't work now anyway), if they are really wanted, they could be
implemented in java instead of javascript.
>
> I'd like to remove javascript support in two steps. The first is simply
disable the clhsdb code that tries to initialize the javascript support. I'd like
to do this in 14 (actually as soon as possible). I'd like to actually do this now
even if we decide to keep javascript support and eventually fix it because it will
get rid of the warning you see whenever you attach from clhsdb:
>
> Warning! JS Engine can't start, some commands will not be
available.
>
> This warning will become more of an issue for the clhsdb tests after I
push [4] because then you will also see the full stacktrace for the underlying
exception that caused the Javascript to fail to start. Besides being unnecessary
noise in passing test cases, it can also be misleading in any test that fails
because the exception will be unrelated to the failure. This is actually what got
me going down this path of what the javascript support is all about.
>
> The next step would be to strip out all Javascript related code,
including the SOQL and JSDB tools. This would be done in 15.
>
> Please let me know what you think.
>
> thanks,
>
> Chris
>
> [1]
https://cr.openjdk.java.net/~minqi/6830717/raw_files/new/agent/doc/clhsdb.html
> [2]
http://javatroubleshooting.blogspot.com/2015/12/serviceability-agent-part-3.html
> [3] https://bugs.openjdk.java.net/browse/JDK-8235594
> [4] https://bugs.openjdk.java.net/browse/JDK-8234277
>