This is still not thread-safe but at least avoids errors overwriting each other in single-threaded contexts.

ssizeof() macro idea courtesy of millert@ ok millert@

with input from and ok naddy

This problem could be triggered with the u-boot repository: $ git clone --bare https://gitlab.denx.de/u-boot/u-boot.git $ got log -r u-boot.git got-read-pack: bad object data got: bad object data $

A symptom of this was a segfault in 'got blame distrib/miniroot/install.sub' with the OpenBSD src.git from github, reported by jrick. The problem was that the commit graph traversed one commit too far. This could be more easily reproduced with 'got log': got log -c 05f568 -P distrib/miniroot/install.sub Which listed two commits instead of just the first one: 05f568ecc6aadefa1aff9064a29e798874a71409 <-- install.sub first created here 7c0d87f00e480cdf004324dad6f3e6f4418f8f42 "distrib/miniroot" exists in 7c0d87f00e480cdf004324... but not in the parent of this commit. Ensure that we traverse the full path in tree1 even if an intermediate tree2 cannot be opened. Instead of reporting success and a file change if we can't traverse further through tree2, we now hit this failing attempt to open the file 'install.sub' in tree object 180aa33df8d1 (tree1): te1 = find_entry_by_name(tree1, seg, seglen); if (te1 == NULL) { err = got_error(GOT_ERR_NO_OBJ); goto done; } fix tested and ok jrick

Same change as 16aeacf7088d, for subdirectories other than lib/

These changes are intended to make got more compatible with FreeBSD's Capsicum. ok stsp

This should always be correct, since rebase and histedit start out with a clean and single-base-commit worktree, and end up committing all changes across the entire work tree when they are successful. tested by jrick and myself

This makes it possible to remove just comment lines from log messages during a histedit operation, leaving the rest of the log message as-is. The behaviour of treating changes to comment lines as no-ops was coded to prevent mistakes during 'got commit/import/tag', where the command will error out if the log message template is not modified. This is not appropriate for histedit because histedit is used to fix such mistakes. ok millert@

got_error_fmt() could eventually replace got_error_path() which has already been used to construct errors with strings that are not actually paths... ok millert@

suggested by and ok millert

Otherwise, ncurses will replace them with some printable representation whose width we can't predict, and wunctrl() fails to return the replacement for 0x80..0x9f. ok stsp

ok stsp

If the histedit log message editor exits without saving its buffer, Got threw away log messages of all commits which were folded. Only the last commit message is preserved, which could be something meaningless like "fixup". Instead, preserve the initial editor buffer content as-is. That is not going to be an ideal log message, but doesn't throw away information and stands out visually because the newly created log message will start with a comment like '# log message of folded commit a0ff...' Problem reported by jrick

Also fixed as part of a different patch from jrick ok jrick