Bug#848652: [Pkg-openssl-devel] Bug#848652: build fails on ppc64el, double free or corruption, stack suggests libssl

2016-12-20 Thread Kurt Roeckx 
On Tue, Dec 20, 2016 at 05:33:12PM +0100, Daniel Pocock wrote:
> 
> 
> On 19/12/16 23:05, Kurt Roeckx wrote:
> 
> > 
> > You should use SSL_COMP_free_compression_methods() so that we can
> > put the internal pointer to NULL.
> >
> 
> Thanks for suggesting that, I notice that method is only available with
> OpenSSL 1.0.2 and I'm also trying to build backports for jessie (OpenSSL
> 1.0.1)
> 
> Is there anything I can do that will be 1.0.1 compatible?

Doesn't look like there is.

> Maybe I could also do this:
> 
> #if OPENSSL_VERSION_NUMBER < 0x01000200f
>   ErrLog(<<"Unable to free compression methods on OpenSSL < 1.0.2");
> #else
>   SSL_COMP_free_compression_methods()
> #endif
> 
> 
> so that the code will compile with either OpenSSL version, but people
> stuck with an older OpenSSL will potentially have the leak.

Just leak it, even for 1.0.2. (In 1.1.0 the function doesn't do
anything anymore, OpenSSL will clean it up itself.)


Kurt

Bug#848652: [Pkg-openssl-devel] Bug#848652: build fails on ppc64el, double free or corruption, stack suggests libssl

2016-12-20 Thread Daniel Pocock 


On 19/12/16 23:05, Kurt Roeckx wrote:

> 
> You should use SSL_COMP_free_compression_methods() so that we can
> put the internal pointer to NULL.
>

Thanks for suggesting that, I notice that method is only available with
OpenSSL 1.0.2 and I'm also trying to build backports for jessie (OpenSSL
1.0.1)

Is there anything I can do that will be 1.0.1 compatible?

Maybe I could also do this:

#if OPENSSL_VERSION_NUMBER < 0x01000200f
  ErrLog(<<"Unable to free compression methods on OpenSSL < 1.0.2");
#else
  SSL_COMP_free_compression_methods()
#endif


so that the code will compile with either OpenSSL version, but people
stuck with an older OpenSSL will potentially have the leak.

I saw some other discussions of this API method in various Github issues:
https://github.com/curl/curl/issues/905
https://github.com/curl/curl/issues/817

Regards,

Daniel

Bug#848652: [Pkg-openssl-devel] Bug#848652: build fails on ppc64el, double free or corruption, stack suggests libssl

2016-12-19 Thread Kurt Roeckx 
On Mon, Dec 19, 2016 at 09:57:42PM +0100, Daniel Pocock wrote:
> 
> 
> On 19/12/16 21:26, Sebastian Andrzej Siewior wrote:
> > On 2016-12-19 09:40:38 [+0100], Daniel Pocock wrote:
> >> Could this be a known issue on ppc64el or with libssl?
> > 
> > Nothing comes to mind. But it explodes the same way (but without the
> > fancy backtrace) on hurd, too.
> > 
> >> FAIL: stunTestVectors
> >> =
> >>
> >> INFO | 20161218-094209.416 |  | RESIP:TEST | 70366852992352 | 
> >> stunTestVectors.cxx:168 | All tests passed!
> >> *** Error in `/«PKGBUILDDIR»/reTurn/test/.libs/stunTestVectors': double 
> >> free or corruption (out): 0x01001be00e20 ***
> >> === Backtrace: =
> >> /lib/powerpc64le-linux-gnu/libc.so.6(+0x829e8)[0x3fff8ea529e8]
> >> /lib/powerpc64le-linux-gnu/libc.so.6(+0x8c098)[0x3fff8ea5c098]
> >> /lib/powerpc64le-linux-gnu/libc.so.6(+0x8cbec)[0x3fff8ea5cbec]
> >> /usr/lib/powerpc64le-linux-gnu/libcrypto.so.1.0.2(CRYPTO_free+0x60)[0x3fff8ef79430]
> >> /usr/lib/powerpc64le-linux-gnu/libssl.so.1.0.2(+0x4ded8)[0x3fff8f19ded8]
> >> /usr/lib/powerpc64le-linux-gnu/libcrypto.so.1.0.2(sk_pop_free+0x74)[0x3fff8f0134c4]
> >> /usr/lib/powerpc64le-linux-gnu/libssl.so.1.0.2(SSL_COMP_free_compression_methods+0x3c)[0x3fff8f1a0b6c]
> >> /«PKGBUILDDIR»/reTurn/test/.libs/stunTestVectors(_ZNSt15_Sp_counted_ptrIPN4asio3ssl6detail17openssl_init_base7do_initELN9__gnu_cxx12_Lock_policyE2EE10_M_disposeEv+0x64)[0x5e41a044]
> >> /«PKGBUILDDIR»/reTurn/test/.libs/stunTestVectors(_ZNSt10shared_ptrIN4asio3ssl6detail17openssl_init_base7do_initEED1Ev+0xdc)[0x5e41a41c]
> > 
> > So a deconstructor in asio invokes SSL_COMP_free_compression_methods()
> > on the end and we explode while the list is cleaned up. Now lets look a
> > little… Oh what is this rutil/ssl/OpenSSLInit.cxx:
> > 
> > | OpenSSLInit::~OpenSSLInit()
> > | {
> > |mInitialized = false;
> > |ERR_remove_state(0);// free thread error queue
> > |EVP_cleanup();// Clean up data allocated during 
> > OpenSSL_add_all_algorithms
> > |CRYPTO_cleanup_all_ex_data();
> > |ERR_free_strings();// Clean up data allocated during 
> > SSL_load_error_strings
> > |sk_SSL_COMP_free (SSL_COMP_get_compression_methods()); 
> > | 
> > | //  CRYPTO_mem_leaks_fp(stderr);
> > |cerr << "HELLO" << endl;
> > | delete [] mMutexes;
> > | }
> >  
> > |(sid)bigeasy@debbuildd:/build/resiprocate-1.11.0~alpha8/reTurn/test$ 
> > ./stunTestVectors 
> > |INFO | 20161219-202446.416 |  | RESIP:TEST | 140378940325248 | 
> > stunTestVectors.cxx:168 | All tests passed!
> > |HELLO
> > 
> > Hmmm. It gives me the impression that exploding here is what should be
> > expected and libssl is the victim here.
> > 
> 
> Thanks for taking the time to look at this.
> 
> Is there an example of how this cleanup should be done, or is there a
> specific line of code there that you believe is out of place?
> 
> If it is definitely not an openssl issue then the bug can be moved back
> to src:resiprocate

You should use SSL_COMP_free_compression_methods() so that we can
put the internal pointer to NULL.


Kurt