obsstore: break the repo → obstore → repo loop This should help the garbage collector to do its job. On repository with many markers, the memory pressure from the obsstore can get quite serious.

relnotes: mention chg behavior change when given --cwd

dispatch: change cwd when loading local config Previously, the `_getlocal` function would not correctly load the repo config when given a relative `rpath` and an alternate cwd via the `wd` parameter. Normally when `--cwd` is specified, hg changes to the given directory before attempting to load the local config (and therefore does not specify a `wd`). The only time the function is called with `wd` set is when hg is running as a command server (e.g., with chg), in which case each forked worker process will attempt to configure itself via `_getlocal` before responding to the client. When given a relative repo path, the worker fails to load the repo config, detects a config mismatch with the client, and enters a redirect/respawn loop. To fix this, we can simply change to the desired working directory during config loading. (Note that simply concatenating `wd` and `rpath` won't work in all cases. The repo path could be something more complicated than a simple relative path, such as a `union:` repo.)

contrib: add support for rhel9

packagelib: use python3 by default

tests: work around libmagic bug in svn subrepo tests libmagic 5.40 introduced a bug [1] wherein ASCII text files with fewer than 3 distinct character values would be reported as binary data rather than as text. This bug was later fixed in version 5.41 [2]. SVN uses libmagic to determine the MIME type of added files with missing or unknown extensions [3]. This results in test failures on systems with libmagic 5.40 installed: $ echo a > a $ svn add a - A a + A (bin) a A simple workaround is to change the test file's content to include 3 distinct ASCII values (including the terminating newline). [1] https://bugs.astron.com/view.php?id=180 [2] https://bugs.astron.com/view.php?id=261 [3] https://svnbook.red-bean.com/en/1.8/svn.advanced.props.html#idm2649

revlog: make _partialmatch fail fast on almost-hex inputs Before this change, resolving a revision like [0123456789^] on a large repo can take multiple seconds because: - hg does not realize this is a revset, so it tries various things, including _partialmatch(b"0123456789^") - after the rust lookup fails, it falls back to pure hg - pure hg takes all-but-last chars and converts them to binary, which *succeeds*, so it does the expensive part.

perf-unbundle: add a perf command to time the unbundle operation Check documentation for details.

perf-bundle: accept --type argument

perf-bundle: accept --rev arguments This is fairly standard nowaday.

perf-bundle: add a new command to benchmark bundle creation time

bundle: introduce a --exact option I have been wanting this options for a long time.

bundlespec: add documentation about existing option We have some documentation, lets make it complete.

setup: use the full executable manifest from `python.exe` The manifest embedded by the build process (before the string here is added) already accounts for the `<requestedExecutionLevel level="asInvoker" ...>` setting. (Note that the PyOxidizer build is missing this, so it will likely trigger the UAC escalation prompt on each run.) However, using `mt.exe` to merge the fragment with what is already in the manifest seems to strip all whitespace, making it unreadable. Since Mercurial can be run via `python.exe`, it makes sense that we would have the same manifest settings (like the supported OS list), though I'm unaware of any functionality this enables. It also has the nice effect of making the content readable from a resource editor. The manifest comes from python 3.9.12. Note that this seems to strip the `<?xml ... ?>` declaration when viewed with ResourceHacker 5.1.7, but this was also the state of things with the previous commit, and `mt.exe "-inputresource:hg.exe;#1" -out:extracted` does contain the declaration and the BOM in both cases. No idea why this differs from other executables.

setup: unconditionally enable the `long-paths-support` option on Windows I don't see anything talking about why this was experimental in the first place, but maybe it was concern about the level of python2 support for it. But now, both `python.exe` and the PyOxidizer build of `hg.exe` have a manifest that enables it, so leaving it off would mean some Mercurial installations could operate on a repo with long paths, and others couldn't. Note that only the wide character functions (XxxW) will have the length restriction lifted. Sadly, distutils applies `/MANIFEST:EMBED` to the linker in a way that can't easily be turned off, so we can't use `/MANIFESTFILE` with `extra_preargs` on `link_executable`. Fortunately, the compiler object provides a path to the `mt.exe` it found during initialization, because the previous incarnation seems to have assumed it is being run within an activated Visual Studio environment. That causes MSYS builds to fail, and probably would have broke the CI environment.

setup: stop shadowing the builtin `dir` symbol I hit this when debugging what's available on the compiler.

subrepo: avoid opening console window for non-native subrepos on Windows Prevent annoying command prompt windows popping up when using TortoiseHG with Git and SVN subrepos by passing creationflags=subprocess.CREATE_NO_WINDOW to subprocess.Popen.

ci: bump pytype to 2022.03.29 This is as far as we can go without running into issues with the vendored `attr` package. I tried updating that to the latest, and not only did it not fix the issue, but test-util.py failed due to some poking at `attr` internals that apparently is no longer valid. The `libcst` package is now pinned to what I have locally because trying to install the latest (0.4.7) complains that it can't find the Rust compiler. We should probably use a requirements file instead (and/or figure out why it can't find the Rust compiler), but I don't feel like dealing with another side quest.

typing: suppress a few attribute errors in url.py These are newly detected by pytype 2022.03.21. Not sure what is going on here- `realhostport` and `headers` are added outside of the constructor, so that makes sense. But PyCharm also thinks the private methods don't exist, though when clicking through the class hierarchy, it shows in the py3.9 source code.

typing: suppress a few pyi-errors with more recent pytype Not sure what's going on here, but these were flagged with pytype 2022.03.21. We can't update to something much more recent, because newer versions complain about various `attr` uses.

sslutil: another use proper attribute to select python 3.7+ The previous attribute was python 3.6+, but guarded a python 3.7+ block Using the correct attribute avoids: + File "/tmp/hgtests.bc0_uk2d/install/lib/python/mercurial/sslutil.py", line 577, in wrapserversocket + sslcontext.minimum_version = ssl.TLSVersion.TLSv1_1 + AttributeError: module 'ssl' has no attribute 'TLSVersion'

sslutil: use proper attribute to select python 3.7+ The previous attribute was python 3.6+, but guarded a python 3.7+ block. Using the correct attribute avoids: File "/usr/lib64/python3.6/site-packages/mercurial/sslutil.py", line 334, in wrapsocket sslcontext.minimum_version = ssl.TLSVersion.TLSv1_1 AttributeError: module 'ssl' has no attribute 'TLSVersion'

branching: merge stable into default

branching: merge stable into default

lfs: fix interpolation of int and %s in an exception case Seen in the wild in a server log when MS antivirus was quarantining a file on the client side.

tests: catch "Can't assign requested address" in test-https.t (issue6726)

tests: add another variation of EADDRNOTAVAIL message (e.g. from NetBSD)

configitems: change the `verify.skipflags` default value to avoid a py3 crash The revlog and LFS modules use various `&` and `&=` operations with this value, which no longer treats `None` as 0. Since nothing cares if it was actually set in the config or not, just default to 0 for simplicity.