Re: RFR [9]: 8177036: Class.checkMemberAccess throws NPE when calling Class methods via JNI

[Mandy Chung]

I saw the thread on JDK-8177136.  I share your concern that throwing ICE does 
not seem a good choice.  I’ll reply on that thread.  It seems to me that it 
should default to the unnamed module if no caller frame rather than throwing an 
exception.

For 8177036, no change in behavior.  This actually fixes a regression that 
throws NPE that I didn’t catch during the review. 

Mandy

> On Mar 20, 2017, at 1:29 PM, David Holmes  wrote:
> 
> As I'm discussing in the RFR for
> 
> "8177136: Caller sensitive methods Logger.getLogger, 
> Logger.getAnonymousLogger, and System.getLogger should throw 
> IllegalCallerException if there is no caller on the stack."
> 
> I am quite concerned by this, seemingly sudden, problem that we have a whole 
> bunch of methods that now rely on there being a Java caller on the stack and 
> as a result if called from JNI they fail! It's fine for the internal methods 
> that depend on a caller to throw exceptions, but the public APIs should have 
> defined semantics about how they depend on any kind of "caller context" and 
> not just suddenly make the call illegal by throwing an IllegalCallerException 
> as in 8177136!
> 
> Cheers,
> David
> 
> On 21/03/2017 4:06 AM, Mandy Chung wrote:
>> 
>>> On Mar 20, 2017, at 5:15 AM, Daniel Fuchs  wrote:
>>> 
>>> Hi Peter,
>>> 
>>> On 20/03/2017 12:01, Peter Levart wrote:
 Perhaps the best way to rectify those problems in one place would be for
 Reflection.getCallerClass() to return a special internal class in its
 own package, such as:
 
 jdk.internal.solitary.NoCaller
 
 ...when there is no caller. This would work correctly for class loader
 checks and would only allow invoking public exported members by core
 reflection if invoked with no caller...
>>> 
>>> I believe this might be dangerous as it would probably hide bugs
>>> in places where 'null' results in NPE being thrown in today's
>>> implementation.
>>> 
>>> Allowing the code to succeed is not always the right thing to
>>> do, and I don't believe it can be fixed in one place.
>>> It's probably better to let the caller of Reflection.getCallerClass()
>>> decide what to do when null is returned, even if this means
>>> we might have to analyze all places where @CallerSensitive is
>>> used.
>>> 
>> 
>> Exactly.  We ought to examine all @CS methods and determine if it handles 
>> the no caller case properly and as intended.  I file a JBS issue to track 
>> this:
>> 
>> https://bugs.openjdk.java.net/browse/JDK-8177155
>> 
>> FYI.  Several options were discussed what StackWalker::getCallerClass should 
>> return and captured in [1].
>> 
>> Mandy
>> [1] 
>> https://bugs.openjdk.java.net/browse/JDK-8140450?focusedCommentId=13867764=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13867764
>> 

Re: RFR [9]: 8177036: Class.checkMemberAccess throws NPE when calling Class methods via JNI

[David Holmes]

As I'm discussing in the RFR for

"8177136: Caller sensitive methods Logger.getLogger, 
Logger.getAnonymousLogger, and System.getLogger should throw 
IllegalCallerException if there is no caller on the stack."


I am quite concerned by this, seemingly sudden, problem that we have a 
whole bunch of methods that now rely on there being a Java caller on the 
stack and as a result if called from JNI they fail! It's fine for the 
internal methods that depend on a caller to throw exceptions, but the 
public APIs should have defined semantics about how they depend on any 
kind of "caller context" and not just suddenly make the call illegal by 
throwing an IllegalCallerException as in 8177136!


Cheers,
David

On 21/03/2017 4:06 AM, Mandy Chung wrote:



On Mar 20, 2017, at 5:15 AM, Daniel Fuchs  wrote:

Hi Peter,

On 20/03/2017 12:01, Peter Levart wrote:

Perhaps the best way to rectify those problems in one place would be for
Reflection.getCallerClass() to return a special internal class in its
own package, such as:

jdk.internal.solitary.NoCaller

...when there is no caller. This would work correctly for class loader
checks and would only allow invoking public exported members by core
reflection if invoked with no caller...


I believe this might be dangerous as it would probably hide bugs
in places where 'null' results in NPE being thrown in today's
implementation.

Allowing the code to succeed is not always the right thing to
do, and I don't believe it can be fixed in one place.
It's probably better to let the caller of Reflection.getCallerClass()
decide what to do when null is returned, even if this means
we might have to analyze all places where @CallerSensitive is
used.



Exactly.  We ought to examine all @CS methods and determine if it handles the 
no caller case properly and as intended.  I file a JBS issue to track this:

https://bugs.openjdk.java.net/browse/JDK-8177155

FYI.  Several options were discussed what StackWalker::getCallerClass should 
return and captured in [1].

Mandy
[1] 
https://bugs.openjdk.java.net/browse/JDK-8140450?focusedCommentId=13867764=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13867764

Re: RFR [9]: 8177036: Class.checkMemberAccess throws NPE when calling Class methods via JNI

[Mandy Chung]

> On Mar 20, 2017, at 2:19 AM, Claes Redestad  wrote:
> 
> Hi,
> 
> please review this fix to avoid NPEs when calling certain Class methods via 
> JNI:
> 
> Bug: https://bugs.openjdk.java.net/browse/JDK-8177036
> Webrev: http://cr.openjdk.java.net/~redestad/8177036/jdk.01/


Looks okay.  I agree that at some point we should look at @CS methods and if 
the no caller case is properly handled.  I filed a JBS issue for it:
  https://bugs.openjdk.java.net/browse/JDK-8177155

Mandy

Re: RFR [9]: 8177036: Class.checkMemberAccess throws NPE when calling Class methods via JNI

[Mandy Chung]

> On Mar 20, 2017, at 5:15 AM, Daniel Fuchs  wrote:
> 
> Hi Peter,
> 
> On 20/03/2017 12:01, Peter Levart wrote:
>> Perhaps the best way to rectify those problems in one place would be for
>> Reflection.getCallerClass() to return a special internal class in its
>> own package, such as:
>> 
>> jdk.internal.solitary.NoCaller
>> 
>> ...when there is no caller. This would work correctly for class loader
>> checks and would only allow invoking public exported members by core
>> reflection if invoked with no caller...
> 
> I believe this might be dangerous as it would probably hide bugs
> in places where 'null' results in NPE being thrown in today's
> implementation.
> 
> Allowing the code to succeed is not always the right thing to
> do, and I don't believe it can be fixed in one place.
> It's probably better to let the caller of Reflection.getCallerClass()
> decide what to do when null is returned, even if this means
> we might have to analyze all places where @CallerSensitive is
> used.
> 

Exactly.  We ought to examine all @CS methods and determine if it handles the 
no caller case properly and as intended.  I file a JBS issue to track this:

https://bugs.openjdk.java.net/browse/JDK-8177155

FYI.  Several options were discussed what StackWalker::getCallerClass should 
return and captured in [1].  

Mandy
[1] 
https://bugs.openjdk.java.net/browse/JDK-8140450?focusedCommentId=13867764=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13867764

Re: RFR [9]: 8177036: Class.checkMemberAccess throws NPE when calling Class methods via JNI

[Daniel Fuchs]

Hi Peter,

On 20/03/2017 12:01, Peter Levart wrote:

Perhaps the best way to rectify those problems in one place would be for
Reflection.getCallerClass() to return a special internal class in its
own package, such as:

jdk.internal.solitary.NoCaller

...when there is no caller. This would work correctly for class loader
checks and would only allow invoking public exported members by core
reflection if invoked with no caller...


I believe this might be dangerous as it would probably hide bugs
in places where 'null' results in NPE being thrown in today's
implementation.

Allowing the code to succeed is not always the right thing to
do, and I don't believe it can be fixed in one place.
It's probably better to let the caller of Reflection.getCallerClass()
decide what to do when null is returned, even if this means
we might have to analyze all places where @CallerSensitive is
used.

best regards,

-- daniel




What do you think?


Regards, Peter

Re: RFR [9]: 8177036: Class.checkMemberAccess throws NPE when calling Class methods via JNI

[Peter Levart]

On 03/20/2017 12:42 PM, Peter Levart wrote:
Currently core reflection throws InternalError (because of the check 
in Reflection::ensureMemberAccess). But jake repo already contains 
changes that remove this check and NPE is thrown later on in 
verifyModuleAccess...


Regards, Peter 


...it is actually even more erratic. If the invocation of core 
reflection is performed from JNI with no caller on a freshly constructed 
Member object, then the call often always succeeds (unless the member is 
protected instance member accessed from subclass), because 
AccessibleObject cache with no cached entry is mistakenly treated as 
'null' caller.  If the Member object is 1st used from a non-null caller, 
only then InternalError is thrown when such Member is later invoked from 
JNI with no caller.


Regards, Peter

Re: RFR [9]: 8177036: Class.checkMemberAccess throws NPE when calling Class methods via JNI

[Alan Bateman]

On 20/03/2017 11:42, Peter Levart wrote:



I guess that when only the ClassLoader of the caller is checked in 
logic of @CS method, such calls should behave as though the caller was 
some class loaded by bootstrap ClassLoader. But what about @CS methods 
that inspect the caller class more deeply (such as reflection API) 
where the identity of the caller, its package and module play role in 
decisions? What behavior is expected in such cases? Should core 
reflection always allow access when invoked from "null" caller?


Currently core reflection throws InternalError (because of the check 
in Reflection::ensureMemberAccess). But jake repo already contains 
changes that remove this check and NPE is thrown later on in 
verifyModuleAccess...
I think we'll need to create some tests to catch all the cases where @CS 
methods are involved from thread attached via JNI and without any caller 
frames. In some cases then IllegalCallerException might be best, in 
other cases then it may be okay to assume the caller is java.base.


-Alan.

Re: RFR [9]: 8177036: Class.checkMemberAccess throws NPE when calling Class methods via JNI

[Peter Levart]

On 03/20/2017 11:41 AM, Alan Bateman wrote:

On 20/03/2017 09:19, Claes Redestad wrote:


Hi,

please review this fix to avoid NPEs when calling certain Class 
methods via JNI:


Bug: https://bugs.openjdk.java.net/browse/JDK-8177036
Webrev: http://cr.openjdk.java.net/~redestad/8177036/jdk.01/
This looks okay. At some point then we need to see how we can test 
these methods called from attached thread (with no caller).


-Alan


I guess that when only the ClassLoader of the caller is checked in logic 
of @CS method, such calls should behave as though the caller was some 
class loaded by bootstrap ClassLoader. But what about @CS methods that 
inspect the caller class more deeply (such as reflection API) where the 
identity of the caller, its package and module play role in decisions? 
What behavior is expected in such cases? Should core reflection always 
allow access when invoked from "null" caller?


Currently core reflection throws InternalError (because of the check in 
Reflection::ensureMemberAccess). But jake repo already contains changes 
that remove this check and NPE is thrown later on in verifyModuleAccess...


Regards, Peter

Re: RFR [9]: 8177036: Class.checkMemberAccess throws NPE when calling Class methods via JNI

[Alan Bateman]

On 20/03/2017 09:19, Claes Redestad wrote:


Hi,

please review this fix to avoid NPEs when calling certain Class 
methods via JNI:


Bug: https://bugs.openjdk.java.net/browse/JDK-8177036
Webrev: http://cr.openjdk.java.net/~redestad/8177036/jdk.01/
This looks okay. At some point then we need to see how we can test these 
methods called from attached thread (with no caller).


-Alan