RFR: 8239563 - Reduce public exports in dynamic libraries built from static JDK libraries

Thu Feb 27 14:52:05 UTC 2020

> On Feb 27, 2020, at 7:15 AM, Magnus Ihse Bursie <magnus.ihse.bursie at oracle.com> wrote:
> 
> On 2020-02-26 22:01, Bob Vandette wrote:
>> Here’s an updated webrev that implementes the suggestion that allows JNIEXPORT in jni.h to be overridden
>> and the build limits visibility for static libraries.
>> 
>> If this webrev is accepted, I’ll update the CSR solution to match this implementation.
>> 
>> http://cr.openjdk.java.net/~bobv/8239563/webrev.01
> This looks basically ok, but some remarks:
> 
> You have forgotten to update the copyright year in the header files. 

Thanks, I’ll update them.

> 
> Also, the quoting looks suspicious. I would have guessed, thinking more carefully about this than the post yesterday, that the proper syntax would be -DJNIEXPORT='__attribute__((visibility("hidden")))' since otherwise the ' will make the \ literal. But, I usually end up being wrong about 50% of the time regarding this kind of stuff. :-) Have you verified that you get the proper define?

I did verify that the quoting works on Mac and Linux.  I needed to add \” before hidden or the quotes would be eliminated causing 
the compiler to complain that visibility was expecting a string but didn’t see one.

Bob.

> 
> /Magnus
> 
>> 
>> Bob.
>> 
>> 
>>> On Feb 26, 2020, at 10:35 AM, Magnus Ihse Bursie <magnus.ihse.bursie at oracle.com> wrote:
>>> 
>>> On 2020-02-26 15:56, Bob Vandette wrote:
>>>> 
>>>>> On Feb 26, 2020, at 9:17 AM, Magnus Ihse Bursie <magnus.ihse.bursie at oracle.com> wrote:
>>>>> 
>>>>> On 2020-02-26 14:31, Bob Vandette wrote:
>>>>>>> On Feb 26, 2020, at 7:31 AM, Magnus Ihse Bursie <magnus.ihse.bursie at oracle.com> wrote:
>>>>>>> 
>>>>>>> On 2020-02-26 08:41, David Holmes wrote:
>>>>>>>> Hi Bob,
>>>>>>>> 
>>>>>>>> Adding build-dev.
>>>>>>> Thanks for noticing that we were missing, David!
>>>>>> Sorry, I should have included you folks.
>>>>>> 
>>>>>>>> On 26/02/2020 6:37 am, Bob Vandette wrote:
>>>>>>>>> Please review this RFE that alters the visibility of JNI entrypoints to hidden when a shared library
>>>>>>>>> is created using static JDK libraries.
>>>>>>>>> 
>>>>>>>>> RFE:
>>>>>>>>> 
>>>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8239563
>>>>>>>>> 
>>>>>>>>> WEBREV:
>>>>>>>>> 
>>>>>>>>> http://cr.openjdk.java.net/~bobv/8239563/webrev.00/
>>>>>>>>> 
>>>>>>>>> CSR:
>>>>>>>>> 
>>>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8239791
>>>>>>>>> 
>>>>>>>>> All JNI entrypoints that exist in JDK static libraries are declared as exported or visible.
>>>>>>>>> If a dynamic library is built from these static libraries, we end up with many exported
>>>>>>>>> functions that we don't want to provide access to,
>>>>>>>>> 
>>>>>>>>> This RFE will change the definition of JNIEXPORT for libraries built when JNI_STATIC_BUILD
>>>>>>>>> is defined.  When defined, functions declared with JNIEXPORT will not be exported when
>>>>>>>>> linked into shared or dynamic libraries.  This will still allow linking of these functions into
>>>>>>>>> dynamic libraries but will not export the JDK symbols outside of the shared library.
>>>>>>>>> 
>>>>>>>>> A CSR has been filed (https://bugs.openjdk.java.net/browse/JDK-8239791) to add the JNI_STATIC_BUILD
>>>>>>>>> define support in jni.h.
>>>>>>>> I have reservations about turning this into something we have to expose and support, rather than being something totally handled within the OpenJDK build system.
>>>>>>> I fully agree. The JNI headers are an exported interface. Our internal build mechanisms have nothing to do there.
>>>>>>> 
>>>>>>>> I'm tempted to suggest the header files be adjusted to have:
>>>>>>>> 
>>>>>>>>     #ifndef JNIEXPORT
>>>>>>>>     <JNIEXPORT basic definitions as they are now >
>>>>>>>>     #endif
>>>>>>>> 
>>>>>>>> and then we define the modified JNIEXPORT via the build system when doing a static build.
>>>>>>>> 
>>>>>>>> Is that feasible?
>>>>>>> It's definitely doable, and a far better solution.
>>>>>> Yes, I like this solution.
>>>>>> 
>>>>>>> A patch something akin to this would be needed:
>>>>>>> 
>>>>>>> diff --git a/make/autoconf/flags-cflags.m4 b/make/autoconf/flags-cflags.m4
>>>>>>> --- a/make/autoconf/flags-cflags.m4
>>>>>>> +++ b/make/autoconf/flags-cflags.m4
>>>>>>> @@ -709,7 +709,10 @@
>>>>>>>    # JDK libraries.
>>>>>>>    STATIC_LIBS_CFLAGS="-DSTATIC_BUILD=1"
>>>>>>>    if test "x$TOOLCHAIN_TYPE" = xgcc || test "x$TOOLCHAIN_TYPE" = xclang; then
>>>>>>> -    STATIC_LIBS_CFLAGS="$STATIC_LIBS_CFLAGS -ffunction-sections -fdata-sections"
>>>>>>> +    STATIC_LIBS_CFLAGS="$STATIC_LIBS_CFLAGS -ffunction-sections \
>>>>>>> +        -fdata-sections -DJNIEXPORT=__attribute__((visibility(\"hidden\")))"
>>>>>>> +  else
>>>>>>> +    STATIC_LIBS_CFLAGS="$STATIC_LIBS_CFLAGS -DJNIEXPORT="
>>>>>>>    fi
>>>>>>>    if test "x$TOOLCHAIN_TYPE" = xgcc; then
>>>>>>>      # Disable relax-relocation to enable compatibility with older linkers
>>>>>>> 
>>>>>>> (With the reservation that I just wrote this on the fly and have not tested it -- things like quoting might be off. Also, I'm not sure if the match of
>>>>>>> compilers is correct -- it might be the case that all compilers except Microsoft defines __GNUC__, so maybe the addition of this -D flag might need
>>>>>>> a separate if statement to cover all our compilers correctly.)
>>>>>> Most of the STATIC_BUILD support is done in jni_util.h.  We could define JNIEXPORT in that header file after allowing it to be overridden in jni.h.
>>>>> I'm not sure I understand you correctly here. Do you mean that you'd like to re-define JNIEXPORT inside jni_util.h instead of using compiler command line flags? I don't think that'd work -- all libraries using JNIEXPORT that does not include jni_util.h first would then export their symbols just the same. Even if you fixed those, the system would be very fragile.
>>>> I was just trying to keep all static library building options in one place.  The static libraries that we produce need to include jni_util.h
>>>> or the JNI_OnLoad_xxx functions will not be declared properly.  Why not expand that dependency to the JNIEXPORT?
>>> Unless *all* libraries that include jni.h also include jni_util.h, then the current definition of JNIEXPORT in jni.h will be used -- meaning that the so decorated functions will be exported -- which was exactly what you wanted to prevent. So I fail to see how this can be a solution.
>>>> 
>>>> Do we really have access to all of these compiler defines from within our Makefiles?
>>>> 
>>>> #if (defined(__GNUC__) && ((__GNUC__ > 4) || (__GNUC__ == 4) && (__GNUC_MINOR__ > 2))) || __has_attribute(visibility)
>>> Well, yes and no. I'm not certain which compilers define __GNUC__ just to show compatibility with gcc, but otoh that does not really matter. All that matters is that we know how we want JNIEXPORT to be defined when creating a static build -- and that we know, since we can check which toolchain we're using. (This is btw a far better check than to look for __GNUC__).
>>> 
>>> /Magnus
>>>> 
>>>> 
>>>> Bob.
>>>> 
>>>> 
>>>>> /Magnus
>>>>>> Bob.
>>>>>> 
>>>>>>> /Magnus
>>>>>>>> Thanks,
>>>>>>>> David
>>>>>>>> 
>>>>>>>>> BACKGROUND:
>>>>>>>>> 
>>>>>>>>> In JDK8 the JNI specification and JDK implementation was enhanced to support static JNI libraries
>>>>>>>>> but we didn’t consider the issue of exportibility of JNI entrypoint symbols.
>>>>>>>>> 
>>>>>>>>>    https://bugs.openjdk.java.net/browse/JDK-8005716
>>>>>>>>> 
>>>>>>>>> If developers use these static JDK libraries in order to produce a custom shared library, all of the
>>>>>>>>> JNIEXPORTS will be exposed by this library even if the developer did not choose to export these.
>>>>>>>>> This is a security issue and a potential problem if this library is mixed with other libraries containing
>>>>>>>>> these symbols.
>>>>>>>>> 
>>>>>>>>> Bob.
>>>>>>>>> 
>>>>>>>>> 
>>>>>>>>> 
>>> 
>> 
> 