Mercurial: Changelog

errors: use StateError more in merge module Differential Revision: https://phab.mercurial-scm.org/D9912

errors: use exit code 40 for when a hook fails A hook can be used for checking inputs, state, configuration, security, etc., so it's unclear which of the existing exit codes to use. Let's instead add one specifically for failed hooks. I picked 40. Differential Revision: https://phab.mercurial-scm.org/D9910

context: add missing manifest invalidation after write in overlayworkingctx This was breaking my merge-diff logic that will be in the next patch. Differential Revision: https://phab.mercurial-scm.org/D9908

wix: tell ComponentSearch that it is finding a directory (not a file) This is to fix an issue we've noticed where fresh installations start at `C:\Program Files\Mercurial`, and then upgrades "walk up" the tree and end up in `C:\Program Files` and finally `C:\` (where they stay). ComponentSearch defaults to finding files, which I think means "it produces a string like `C:\Program Files\Mercurial`", whereas with the type being explicitly a directory, it would return `C:\Program Files\Mercurial\` (note the final trailing backslash). Presumably, a latter step then tries to turn that file name into a proper directory, by removing everything after the last `\`. This could likely also be fixed by actually searching for the component for hg.exe itself. That seemed a lot more complicated, as the GUID for hg.exe isn't known in this file (it's one of the "auto-derived" ones). We could also consider adding a Condition that I think could check the Property and ensure it's either empty or ends in a trailing slash, but that would be an installer runtime check and I'm not convinced it'd actually be useful. This will *not* cause existing installations that are in one of the bad directories to fix themselves. Doing that would require a fair amount more understanding of wix and windows installer than I have, and it *probably* wouldn't be possible to be 100% correct about it either (there's nothing preventing a user from intentionally installing it in C:\, though I don't know why they would do so). If someone wants to tackle fixing existing installations, I think that the first installation is actually the only one that shows up in "Add or Remove Programs", and that its registry keys still exist. You might be able to find something under HKEY_USERS that lists both the "good" and the "bad" InstallDirs. Mine was under `HKEY_USERS\S-1-5-18\Software\Mercurial\InstallDir` (C:\), and `HKEY_USERS\S-1-5-21-..numbers..\Software\Mercurial\InstallDir` (C:\Program Files\Mercurial). If you find exactly two, with one being the default path, and the other being a prefix of it, the user almost certainly hit this bug :D We had originally thought that this bug might be due to unattended installations/upgrades, but I no longer think that's the case. We were able to reproduce the issue by uninstalling all copies of Mercurial I could find, installing one version (it chose the correct location), and then starting the installer for a different version (higher or lower didn't matter). I did not need to deal with an unattended or headless installation/upgrade to trigger the issue, but it's possible that my system was "primed" for this bug to happen because of a previous unattended installation/upgrade. Differential Revision: https://phab.mercurial-scm.org/D9891

persistent-nodemap: Fix Rust declarations for Revlog_CAPI signatures Use Rust’s `libc::ssize_t` as the closest match to C’s `Py_ssize_t`. See details in test comment. Going forward we should find a way to have such Rust declarations auto-generated from C headers at build time, or auto-checked against them in a test. Differential Revision: https://phab.mercurial-scm.org/D9901

persistent-nodemap: add test case reproducing a Rust panic Thanks to Mitchell Plamann for reporting this bug. The next commit fixes it. Differential Revision: https://phab.mercurial-scm.org/D9900

copies: fix an incorrect comment in graftcopies() from recent D9802 I don't think the bit about creating a merge in the filelog was correct. Or at least I couldn't find a case where it happened. Differential Revision: https://phab.mercurial-scm.org/D9899

narrow: delete a stale TODO about not sending groups the client already has 2c5835b4246b changed the changegroup generation to not send treemanifests for directories the client had before widening. As that commit mentions, we had already stopped before that commit to send the changelog and filelogs for files the client already had. Differential Revision: https://phab.mercurial-scm.org/D9898