From: Christian Weisgerber <naddy@mips.inka.de>
Subject: Loss of original errors
To: gameoftrees@openbsd.org
Date: Tue, 11 Jul 2023 20:22:20 +0200

There are a few places in got(1) where, if an error occurs, a
clean-up action is executed, whose potential error is ignored, and
the original error is returned.

Or at least that's the intent.  In practice, the original error can
be lost because the custom_errors[] ring buffer is too small.

Specifically, there are three instances in cmd_histedit():

        error = histedit_...(...);
        if (error) {
                got_worktree_histedit_abort(...);
                print_merge_progress_stats(...);
                goto done;
        }

Here got_worktree_histedit_abort() creates a number of ignored
errors that overrun the ring buffer.  I first tried increasing the
buffer size, but a mkdir(2) is run on every directory in the worktree,
filling the buffer with ignored EEXIST errors, so that doesn't
scale.

There is similar-ish code in cmd_integrate() with calls to
got_worktree_integrate_abort(), but that routine probably doesn't
do enough work to cause problems.

The obvious solution would be to save *error into a variable before
calling the clean-up action, but I haven't found a precedent in the
code, and it can't be a local variable...

Suggestions?

-- 
Christian "naddy" Weisgerber                          naddy@mips.inka.de