#464: drakma / cl+ssl failure ----------------------------+--------------------------------- Reporter: charmon | Owner: Evenson Not Org Type: defect | Status: assigned Priority: blocker | Milestone: 1.6.0 Component: libraries | Version: 1.6.0-dev Resolution: | Keywords: cl+ssl, ssl, drakma Parent Tickets: | ----------------------------+---------------------------------
Comment (by Evenson Not Org):
I'm still suffering from the feeling that we are dealing with some sort of Heisenbug around the memory allocation/initialization strategy of openssl 1.1.
Running the {{{cl+ssl.test}}} suite against released abcl-1.5.0 binaries which uses jna-4.2.2 segfaults on macOS, has the two test failures with {{{bio_write_intern:uninitialized}}} on Linux, and segfaults on freebsd.
The space of potential variables here is large:
* Platform (macOS fails "sometimes", Linux "sometimes", has always failed on FreeBSD, Windows untested) * OpenSSL version (abcl-1.5.0 was released when openssl 1.0.x/1.1.y were in use) * jna version (abcl-1.5.0 used jna-4.2.2; abcl-1.6.0-dev has jna-5.5.0 but "downgrading" to jna-4.2.2 doesn't show much of a difference * CFFI (presumably better with the varargs patch)
My current strategy towards working on this:
1. Get CI artifacts into source tree so we can run the tests in a much more controlled manner. I will work on using Github/Travis as the first target, as that seems easiest for me at the moment
2. Figure out how to run the JVM under GDB so that we can inspect the smashed stack, and hopefully set a breakpoint before it happens. Not sure if this is entirely possible in 2019 without a whole lot of elbow grease.
-- Ticket URL: http://abcl.org/trac/ticket/464#comment:23 armedbear http://abcl.org armedbear