If creating the /tmp display diff for a commitable item failed we would free the commitable item while it was already on the path list. Later on when the path list was freed in got_worktree_commit() a double-free would be detected and the program would be aborted. Found by gonzalo@ ok op@

This prevents gotd from closing the client connection before ref-update status reports have been sent. Seen while testing gotd on Linux, though I see no obvious reason why this race would not trigger on OpenBSD.

The limit commit queue is invariably filled with different commits to the queue containing all commits, so the index of the work tree's base commit in the real queue corresponds to a different commit in the limit queue. As such, the marker is drawn on an incorrect commit. Rather than fix this to draw the marker on the correct commit if it happens to be in the limit queue, don't draw the marker at all in limit view. As pointed out by Mikhail on the list, this information is not wanted in the limit view. Patch and report by Mikhail.

Prevent an endless polling loop on Linux where POLLOUT and POLLHUP can both be reported via pollfd revents.

"thou shalt not include sys/cdefs.h because it is not standardized", and in this case is not really needed either.

We did consider making -d and -P options conflict but doing so would not be very helpful. Instead, document -d as a superset of -P. ok op, jamsek

If show_repo_age is set to off, got_get_repo_age() returns without setting the timestamp. got_get_repo_age() is also used by the branch listing to show how old they are however, and we end up rendering the stack garbage. Instead, respect `show_repo_age off' one layer above, and always return a timestamp in got_get_repo_age(). While here, initialize the timestamp to zero so it's set also in case of failures. Issue reported on IRC by xs, thanks!

When the current matched and selected entry is the last loaded commit, we keep looping the same code path because search_entry is always NULL. Before poking the log thread for more commits, set search_entry to the currently selected commit, which is where the search resumes. Patch by Mikhail

Plus multiple copypasta fixes where I was cmp(1)ing stdout instead of stderr! Fortunately, we were doing the right thing despite not checking it properly. The log -dPp test is from op@. ok + tweaks op@

Only collect changed paths once if both -d and -P are specified; we already checked for -d and -p. Reported by Lucas on IRC. Regress for this and the previous (got log -x keyword) commit still due. ok stsp@ and op@

Suggested by Lucas on IRC. ok stsp@ and op@

By incrementing the resize step from 64 to 256 deltifying takes less time on modestly sized files; the resize is still a small number instead of a fraction of the current table size (which would be more usual for a hash table) since this code is also used in gotd. ok stsp

it's a different way to spell the same number, but it's more evocative since the on-disk format is 32 bits wide. ok stsp

Use a local buffer instead of calling reallocarray() every 8 bytes; the speed up is measurable. The downside is that we're now limited to paths long at most PATH_MAX bytes, but since this is the fileindex it's not an issue in practice. ok jamsek stsp

This makes histedit folding work when a file is modified by commit A and then deleted in commit B, with A being folded into B. Problem reported by naddy@

Patch by jamsek with a request to commit. Thanks!

Aborting tog regress with Ctrl-C would trigger a "tty" pledge violation via the built-in ncurses handler for SIGINT, installed by newterm(). Install tog's own signal handlers earlier to fix this.

leftover from 79c49d8454.

This callback could return any error code. Do not mask all such errors as cancellation.