Emilio Pozuelo Monfort
GSoC 2010: Hurd: Fix Compatibility Problems Exposed by Testsuites
Mentor: Carl Fredrik Hammar
This project will consist of identifiying some projects' test suite failures when executed on GNU/Hurd, debugging them, and if they are truly GNU/Hurd issues (and not problems in the projects themselves), fixing them.
- July 18th: GLib finished.
- July 22nd: coreutils finished.
- July 25th: All Perl failures investigated.
- August 5th: Perl finished.
- August 8th: All Python failures investigated.
- August 16th: Python finished.
- August 16th: Firm 'pencils down' date
- Investigate why coreutils' nice test fails.
- Analyze Perl's testsuite failures.
- Towards a New Strategy of OS Design, an architectural overview by Thomas Bushnell, BSG
- The Hurd, a presentation by Marcus Brinkmann
- The Hurd Hacking Guide
- MIG - The MACH Interface Generator
July 26th - August 1st
- Tested /dev/fd/N patches and sent them for review.
- Finished SCM_RIGHTS patch. Created a minimal testcase without using glibc to demonstrate the socket_send/recv failure with non-socket fds. Sent the testcase and the patch for review.
- Investigated cp issues with O_NOFOLLOW & O_NOTRANS. Sent a mail to bug-hurd explaining both issues and possible solutions.
July 19th - July 25th
- Initial SCM_RIGHTS implementation. Seems to work when sending pipes, but fails miserably when sending fds from an open syscall. No idea why yet.
- Fixed memleaks in sendmsg() while implementing SCM_RIGHTS. Patch accepted upstream.
- Had to build glibc thrice because the system crashed and the fs was totally corrupted. I'll build stuff in a separate partition as suggested from now on. Doesn't help. It turns out the issue seems to be with kvm, or at least it's only reproducible for me there. I've switched to VirtualBox and there are no filesystem issues there.
- Addressed comments in the /dev/fd/N patches. Need to test them (when I can build glibc and Hurd without the system crashing).
July 14th - July 18th
- Catched up with email.
- Prepared a patch to implement getsockopt(fd, SOL_SOCKET, SO_TYPE, ...). Patch committed to Hurd.
- Addressed comments in the /dev/fd/N patches and resent them.
- Investigated another glib's unix-fd failure: passing fds over a socket using sendmsg() doesn't dup them. Created a minimal testcase. Prepared a preliminary patch, needs testing and fixing.
May 26 - July 13th
- Copyright assignment on file.
- Studied a lot to finish my BSc.
- Got the linkat patch (Savannah #29655) committed upstream.
May 19 - May 26
- Read MIG - The MACH Interface Generator
- Worked on bug 28934. Send patches for review.
- Requested GNU libc copyright papers to the FSF again since they didn't arrive the first time.
May 12 - May 19
- Read http://www.gnu.org/software/hurd/hurd-talk.html
- Half read http://www.gnu.org/software/hurd/hacking-guide/hhg.html
- Read many Hurd interfaces (fs.defs, auth.defs, fsys.defs, io.defs, password.defs).
May 5 - May 12
- Read http://www.gnu.org/software/hurd/hurd-paper.html
- Improved the linkat() patch
- Fixed the issues mentioned in bug 28934, but after doing so I realized that wouldn't work well. The only good solution is to pass file_name from execve() to the exec server, so we need new RPCs.
April 28 - May 5
- Submitted a talk proposal on Hurd with Michael Banck for DebConf
- Sent request to assign copyright to the FSF for Hurd/Mach/libc
Before April 28
- Investigated the glib's gtester problem and tracked it down: bug 28934. Prepared a patch but it's not good.
- glib's garray test timeouts on Hurd. The tests passes if the timeout is increased. The current upstream timeout (10s) is quite small (it fails on many Debian builds for some Linux arches). However on Hurd it needs a very big time it seems (like more than 100s). Maybe do an allocation benchmarch?
- Investigated glib's unix-fd test failure: getsockopt() isn't implemented on Hurd. Need to implement it in hurd/pflocal/socket.c.
- Investigated coreutils' ln EIEIO, with Samuel's help. linkat() is buggy. Reported as bug 29655. Prepared a patch for it.
- Investigated coreutils' cp EACCES. Test case: 'mkfifo a && cp -R --copy-contents a b'. Problem is that O_NOFOLLOW adds O_NOTRANS.
- Assigned copyright to the FSF.
- Read many documentation and source code.
- /dev/fd/N bug fixed
- Prepared a patch for getsockopt()
- Fixed linkat() problems.
- Investigated bug with O_NOFOLLOW & O_NOTRANS (needs more work).
Investigated a glib test failure (garray). Not a Hurd issue.
Studied a lot to finish my BSc. Didn't work on Hurd for a month because of that, so that's why I couldn't make a lot of progress (this was known in advance, although in the end the downtime was a bit larger than expected).
- There's no expected downtime from now on.