It seems flushing our final message can fail if we forcefully exit the event loop here. This would leave gotsysd's sysconf process stuck in a state where it was waiting for a message which never arrived.

The gotsysd_imsg_flush() wrapper enforces a sync write by polling the file descriptor and writing data in a loop. Since libevent already does polling for us we can simply use imsgbuf_flush() here instead.

THe main process will always send these messages so don't skip parsing them even if we are going to ignore the current user.

The -q option suppresses too many important errors, including the warning printed when a host key has changed. Requiring users to pass -v in order to see this message is too cumbersome.

Make the semantics of the author_time parameter self-contained by treating author_time == 0 as the fallback case, instead of treating committer == NULL as the fallback case. Having the semantics of one parameter depend on the semantics of another might lead to subtle bugs introduced during potential future refactoring of this code.

No functional change, but keeps the code in sync with got. Patch by Sven M. Hallberg

Adds test_rebase_preserves_author_data to regress/cmdline/rebase.sh. Patch by Sven M. Hallberg

Since gotwebd can now be given read access to gotd repositories by making the directory readable for the _gotwebd group, these checks are too strict. Stop requiring a specific GID, and allow group rx bits to be set for group read access. As before, keep rejecting world/other rwx bits for safety. Read access to repositories is supposed to be managed by gotd.

Add test coverage for this issue.

Dequeue notifications from the list only once the notify process has confirmed that the notification has been sent. This should be more robust for the following reason: In some spots were are checking the list head against NULL to see if any notifications are still pending. By removing notifications from this list too early, this check did not cover notifications which were still in flight. We could thus end up deciding to shut down the session process too early, and the last notification of several would fail. Issue observed when sending a new commit on a branch and a new tag which tags this commit to gotd in a single 'got send' operation. Only the tag notification would make it, while the other notification never arrived after an 'unexpected EOF' error in the notify process while trying to send a confirmation to the session process. ok op@

Ensure that gotsh receives its end of the pack file data pipe before repo_write starts sending pack file status messages. Messages of type GOTD_IMSG_PACKFILE_StATUS would end up being received in gotsh's serve_write() function too early. This race is similar to the one fixed for repo_read back in commit c2274a511a7415078e2628f969b8459f69432411 Reported by martijn@, who pin-pointed the problematic case in the code, thanks!

spotted by Denis Bodor, thanks!

Instead of storing all referenced object IDs in a set, use the pack index of the large pack file we have created to check whether an object exists. Apply the modification time check only to packs which contain unreferenced objects. Otherwise we can leave many duplicate pack files behind. I observed this when all refs were packed and the packed-refs file was days old. ok op@

ok op@