nodemap: drop the 'exp-' prefix for internal opener option The feature is now in a descent shape and we can consider having it "less" experimental. We won't be able to make it "totally" non-experimental, because its benefit rely on rust, which is totally experimental. Differential Revision: https://phab.mercurial-scm.org/D8418

nodemap: gate the feature behind a new requirement Now that the feature is working smoothly, a question was still open, should we gate the feature behind a new requirement or just treat it as a cache to be warmed by those who can and ignored by other. The advantage of using the cache approach is a transparent upgrade/downgrade story, making the feature easier to move to. However having out of date cache can come with a significant performance hit for process who expect an up to date cache but found none. In this case the file needs to be stored under `.hg/cache`. The "requirement" approach guarantee that the persistent nodemap is up to date. However, it comes with a less flexible activation story since an explicite upgrade is required. In this case the file can be stored in `.hg/store`. This wiki page is relevant to this questions: https://www.mercurial-scm.org/wiki/ComputedIndexPlan So which one should we take? Another element came into plan, the persistent nodemap use the `add` method of the transaction, it is used to keep track of a file content before a transaction in case we need to rollback it back. It turns out that the `transaction.add` API does not support file stored anywhere than `.hg/store`. Making it support file stored elsewhere is possible, require a change in on disk transaction format. Updating on disk file requires… introducing a new requirements. As a result, we pick the second option "gating the persistent nodemap behind a new requirements". Differential Revision: https://phab.mercurial-scm.org/D8417

nodemap: move on disk file to version 1 The current format contains the information we need, lets freeze it before the release. Differential Revision: https://phab.mercurial-scm.org/D8416

nodemap: add a new mode value, "strict" When "strict" is set, operation will refuse to use the slow path and abort if they would. This is useful for testing, benchmarking and server operation. Differential Revision: https://phab.mercurial-scm.org/D8415

nodemap: add a new mode option, with an optional "warn" value When "warn" is set, user will get notified when the slow code, used for compatibility is used. This can help people to detect situation were using that feature will give them a slowdown instead of a speedup. Differential Revision: https://phab.mercurial-scm.org/D8414

nodemap: also warm manifest nodemap with other caches The `hg debugupdatecache` command now also warm the persistent nodemap for the manifest (when applicable). Differential Revision: https://phab.mercurial-scm.org/D8411

nodemap: also use persistent nodemap for manifest The manifest as a different usage pattern than the changelog. First, while the lookup in changelog are not garanteed to match, the lookup in the manifest nodemap come from changelog and will exist in the manifest. In addition, looking up a manifest almost always result in unpacking a manifest an operation that rarely come cheap. Nevertheless, using a persistent nodemap provide a significant gain for some operations. For our measurementw, we use `hg cat --rev REV FILE` on the our reference mozilla-try. On this repository the persistent nodemap cache is about 29 MB in side for a total store side of 11,988 MB File with large history (file: b2g/config/gaia.json, revision: 195a1146daa0) no optimisation: 0.358s using mmap for index: 0.297s (-0.061s) persistent nodemap for changelog only: 0.275s (-0.024s) persistent nodemap for manifest too: 0.258s (-0.017s) File with small history (file: .hgignore, revision: 195a1146daa0) no optimisation: 0.377s using mmap for index: 0.296s (-0.061s) persistent nodemap for changelog only: 0.274s (-0.022s) persistent nodemap for manifest too: 0.257s (-0.017s) Same file but using a revision (8ba995b74e18) with a smaller manifest (3944829 bytes vs 10 bytes) no optimisation: 0.192s (-0.185s) using mmap for index: 0.131s (-0.061s) persistent nodemap for changelog only: 0.106s (-0.025s) persistent nodemap for manifest too: 0.087s (-0.019s) Differential Revision: https://phab.mercurial-scm.org/D8410

nodemap: create files in the repository used in the test We need a manifest with more content to test persistent nodemap for manifest. This change the repository content and affect all the hashes. Differential Revision: https://phab.mercurial-scm.org/D8409

rust-matchers: add timing tracing to regex compilation This might be useful to diagnose later performance issues or just to show the difference between engines. Differential Revision: https://phab.mercurial-scm.org/D8498

url: fix a bytes vs str crash in processing proxy headers (issue6249) I have no idea how to make a test for this, so if somebody knows, feel free to add one or follow up on this. The bug reporter reported that it worked for them, so there may not be other hidden issues here. Differential Revision: https://phab.mercurial-scm.org/D8485

pullbundles: use unfiltered repo for head/base matching The unfiltered view works even when changeset transistion from draft to hidden phase. The normal visibility is already ensured by discovery as invisible heads would have been filtered out before. Skipping the filtering has a positive impact on performance, too. Differential Revision: https://phab.mercurial-scm.org/D8481

procutil: always waiting on child processes to prevent zombies with 'hg serve' When runbgcommand is invoked by an extension with ensurestart=False, we never called waitpid - which is fine in most cases, except if that's happening on a command server (e.g. chg), in which case the child defunct process will just sit there for as long as the server is running. The actual semantics of SIGCHLD signal handling is a lot more complex than it seems, and the POSIX standard *seems* to read that it's ignored by default and everything would just work without the waitpid if we're not listening for it, but the truth is that it's only ignored if we *explicitly* set it to SIG_IGN. We further cannot set it to SIG_IGN or to a catch-all handler across all of 'hg serve', because Python's suprocess.Popen relies on that signal, and a few specific parts of hg also set custom handlers, so instead we wait for specific PIDs in dedicated threads. I did a poor-man's benchmark of the thread creation and it seems to take about 1ms, which is way better than the 20+ms from ensurestart=True. Differential Revision: https://phab.mercurial-scm.org/D8497