heptapod-ci: use shell script in pytype step

tests: migrate the pytype test to a shell script for easier CI processing There have been recent hangs and timeout, but it's hard to debug because the *.t test redirects output to a file and only prints it if `pytype` actually exits. This shell script can be run directly by CI, and will allow more flexibility to try to cache and restore type stubs for further speed increases.

rhg: fix bugs around [use-dirstate-tracked-hint] and repo auto-upgrade This makes two changes: - make rhg support the [dirstate-tracked-key-v1] requirement. I believe rhg never changes the tracked file set, so it's OK that it doesn't have any logic for writing this file. - fix the name of [format.use-dirstate-v2.automatic-upgrade-of-mismatching-repositories] config option in rhg, which makes rhg actually honor the auto-upgrade. These two issues cancelled each other out in tests (auto-upgrade was happening because [dirstate-tracked-key-v1] forced the fallback, not because of the config), which is I think why they went unnoticed earlier.

tests: fix the flaky test test-logtoprocess.t The main change is that we're waiting for the [touched] file to appear for 5 seconds instead of 0.1 seconds. Also, instead of implementing wait-on-file from scratch, we use the existing one from testlib/ that works well.

packaging: update dulwich to drop the certifi dependency on Windows The presence of `certifi` causes the system certificate store to be ignored, which was reported as a bug against TortoiseHg[1]. It was only pulled in on Windows because of `dulwich`, which was copied from the old TortoiseHg install scripts, in order to support `hg-git`. This version of `dulwich` raises the minimum `urllib3` to a version (1.25) that does certificate verification by default, without the help of `certifi`[2]. We already bundle a newer version of `urllib3`. Note that `certifi` can still be imported from the user site directory, if installed there. But the installer no longer disables the system certificates by default. [1] https://foss.heptapod.net/mercurial/tortoisehg/thg/-/issues/5825 [2] https://github.com/jelmer/dulwich/issues/1025

tests: fix tar invocation, to address issue 6740

relnotes: add 6.2.1 and 6.2.2

Added signature for changeset b5c8524827d2

Added tag 6.2.2 for changeset b5c8524827d2

dirstate-v2: no longer register the data-file during transaction If the data file change during the transaction, we cannot truncate it. The content of the file itself is fine as it will get backed up at the same time as the docket. Leaving the trailing data at the end of failed transaction is fine. The dirstate-v2 format supports it. The dead data will simply we written over if necessary.

fsmonitor: use new dirstate APIs (issue6728) On top of fixing fsmonitor, it moves one more "old API" use to the new one. This needs very verbose code to save a few function calls that are very expensive in Python.

dirstate-v2: backup the data file during the transaction (issue6730) If we backup of the docket, without doing a backup of the data file, we highly risk restoring a docket pointing to a missing file in the future. So we now backup the data-file alongside the docket.

dirstate-v2: display a possible issue with transaction If we backup the dirstate without the data file, we can end up in an inconsistent state of the transaction is rolled back.

test-dirstate: do not get out of the test directory It looks like we were doing one `cd ..` too many.

rust-matchers: fix behavior of `IncludeMatcher` with multiple includes This change adds a test for this fix as well as an additional test case that was useful in debugging this behavior.

tests: improve portability by no longer using the flag [tar --force-local] A different way of making sure the path is not interpreted as a URL is to never give this path to [tar], instead making tar read from stdin.

tests: remove flakiness in a time-dependent test

tests: remove flakiness in test-nointerrupt.t The problem was that the reaction to the signal was racing against the completion of the command. Since reaction to the signal is to print a line of warning, we can fix this by waiting for that warning to appear before allowing the command to complete.

automation: set PATH when building on Windows Sometime in the 6.2 release cycle the Windows building automation broke. Building the wheel and even PyOxidizer based installers now fails with: ``` Exception: PowerShell execution failed: error: subprocess-exited-with-error Getting requirements to build wheel did not run successfully. exit code: 1 [1 lines of output] Unable to find a working hg binary to extract the version from the repository tags [end of output] note: This error originates from a subprocess, and is likely not a problem with pip. ``` I have a hunch this is a regression from upgrading pip in 1c00777702da, but I haven't verified this. It may not be, as PyOxidizer has its own bundled Python/pip. So maybe it is something in `setup.py`.

contrib: update Mercurial install in bootstrap environment

automation: transition to Windows Server 2022 Let's keep our Windows build environment modern by upgrading to the latest OS. As part of the upgrade, we pick up a migration to EC2Launch Version 2. This has a different config mechanism. So we need to port how we manage the administrator password. As part of migrating to the new YAML/JSON config file mechanism, we move the code to the powershell script that is run when the instance first launches. This ensures that the config is retained during the reboot we perform as part of building the Windows AMI. The motivation for this is I'm currently unable to build the Windows 2019 AMI due to an issue installing OpenSSH. This _just works_ on Windows Server 2022. I have no clue what the root cause is. I think it might have something to do with Microsoft not publishing the files in the right location. Differential Revision: https://phab.mercurial-scm.org/D12630

automation: refresh requirements I'm hitting errors installing the old version of cffi due to an apparent issue where older versions of cffi aren't compatible with the modern Clang I'm using. So let's upgrade packages to unbreak things and to keep things modern. Differential Revision: https://phab.mercurial-scm.org/D12629

rust-status: ignored directories are now correctly only listed if opted into This fixes the behavior of `hg purge` removing empty ignored directory even without `--all` or `--ignored`.

test: show how purge removes ignored directories Apparently the rust code path removes ignored directories even though the pure code path doesn't. Show this in tests.

rust: bump to memmap2 0.5.3, micro-timer 0.4.0, and crossbeam-channel 0.5.0 The merge in 12adf8c695ed had conflicts in rust/Cargo.lock and rust/hg-core/Cargo.toml . Let's ignore rust/Cargo.lock - it is regenerated. For rust/hg-core/Cargo.toml, stable had dd6b67d5c256 "rust: fix unsound `OwningDirstateMap`" which introduced ouroboros (and dropped stable_deref_trait). Default had ec8d9b5a5e7c "rust-hg-core: upgrade dependencies" which had a lot of churn bumping minimum versions - also patch versions. It is indeed a good idea to bump to *allow* use of latest package. That means that major versions should be bumped for packages after 1.0, and for packages below 1.0 minor versions should be bumped too. But it doesn't work to try enforce a policy of using latest patch by bumping versions at arbitrary times. For good or bad, the merge doesn't seem to have resolved the conflicts correctly, and many of the minor "upgrade dependencies" were lost again. Unfortunately, it also lost the bump of memmap2 to 0.5.3, which is needed for Fedora packaging where 0.4 isn't available. Same with micro-timer bump to 0.4 (which already is used in rhg). crossbeam-channel bump was also lost. This change fixes that regression by redoing these "important" lines of the merge "correctly". I propose this for stable, even though dependency changes on stable branches are annoying.

Added signature for changeset f69bffd00abe

Added tag 6.2.1 for changeset f69bffd00abe

debug-discovery: apply spelling fixes from Raphaël

tree-discovery: fix the request debug output and progress location This is now associated with each request.

debug-discovery: deal with case where common is empty This code was previously confused by case where: `heads_common == {nullid}`

debug-discovery: do not abort on unrelated repositories This is a useful case to consider, so we should not abort in this case. A warning is still issued.

debug-discovery: gather the right number of roundtrips for tree discovery We where not counting the right amount of request before.

debug-discovery: also gather details on tree-discovery queries type This is useful to understand the algorithm.

debug-discovery: properly apply remote filtering in "old" mode Before this change using `--remote-as-revs` with `--old` had no effect and everything was considered as "common", which is really not what we intended.

debug-discovery: fix a typo in the doc The second option should be `--remote-…` as we just talked about `--local-…` already.

packaging: bump dulwich to 0.20.45 I'm told the new dulwich avoids hg-git test failures.

packaging: update keyring on Windows to avoid spurious stacktraces When challenged for a network password, this would spew on Windows before it actually used the stored password: ``` Error initializing plugin EntryPoint(name='libsecret', value='keyring.backends.libsecret', group='keyring.backends'). Traceback (most recent call last): File "keyring.backend", line 198, in _load_plugins init_func = ep.load() File "importlib.metadata", line 77, in load module = import_module(match.group('module')) File "importlib", line 127, in import_module return _bootstrap._gcd_import(name[level:], package, level) File "<frozen importlib._bootstrap>", line 1030, in _gcd_import File "<frozen importlib._bootstrap>", line 1007, in _find_and_load File "<frozen importlib._bootstrap>", line 984, in _find_and_load_unlocked ModuleNotFoundError: No module named 'keyring.backends.libsecret' Error initializing plugin EntryPoint(name='macOS', value='keyring.backends.macOS', group='keyring.backends'). Traceback (most recent call last): File "keyring.backend", line 198, in _load_plugins init_func = ep.load() File "importlib.metadata", line 77, in load module = import_module(match.group('module')) File "importlib", line 127, in import_module return _bootstrap._gcd_import(name[level:], package, level) File "<frozen importlib._bootstrap>", line 1030, in _gcd_import File "<frozen importlib._bootstrap>", line 1007, in _find_and_load File "<frozen importlib._bootstrap>", line 984, in _find_and_load_unlocked ModuleNotFoundError: No module named 'keyring.backends.macOS' ``` We're kinda threading a needle here because the next version of `keyring` (currently at 23.7.0) requires `importlib-metadata` 3.6+, which PyOxidizer 0.22 doesn't support[1]. [1] https://github.com/indygreg/PyOxidizer/issues/609

mergestate: action name was str Apparently the standard for them is still to use byte strings. Found while looking at something else

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'

git: copy findmissingrevs() from revlog.py to gitlog.py (issue6472)

git: add a missing reset_copy keyword argument to dirstate.set_tracked() Since nothing else in hgext/git supports copies yet, the best I can do is avoid TypeError: set_tracked() got an unexpected keyword argument 'reset_copy'.

git: make sure to fsdecode bookmark names everywhere (issue6723)

Added signature for changeset 094a5fa3cf52

Added tag 6.2 for changeset 094a5fa3cf52

procutil: make stream detection in make_line_buffered more correct and strict In make_line_buffered(), we don’t want to wrap the stream if we know that lines get flushed to the underlying raw stream already. Previously, the heuristic was too optimistic. It assumed that any stream which is not an instance of io.BufferedIOBase doesn’t need wrapping. However, there are buffered streams that aren’t instances of io.BufferedIOBase, like Mercurial’s own winstdout. The new logic is different in two ways: First, only for the check, if unwraps any combination of WriteAllWrapper and winstdout. Second, it skips wrapping the stream only if it is an instance of io.RawIOBase (or already wrapped). If it is an instance of io.BufferedIOBase, it gets wrapped. In any other case, the function raises an exception. This ensures that, if an unknown stream is passed or we add another wrapper in the future, we don’t wrap the stream if it’s already line buffered or not wrap the stream if it’s not line buffered. In fact, this was already helpful during development of this change. Without it, I possibly would have forgot that WriteAllWrapper needs to be ignored for the check, leading to unnecessary wrapping if stdout is unbuffered. The alternative would have been to always wrap unknown streams. However, I don’t think that anyone would benefit from being less strict. We can expect streams from the standard library to be subclassing either io.RawIOBase or io.BufferedIOBase, so running Mercurial in the standard way should not regress by this change. Py2exe might replace sys.stdout and sys.stderr, but that currently breaks Mercurial anyway and also these streams don’t claim to be interactive, so this function is not called for them.

repo-upgrade: avoid a crash when multiple optimisation are specified In Python 3, the type are no longer comparable and this expose the error.

rust: remove excessive calls to `#[timed]` This makes trace output *really* noisy and is only useful in case you want to take a look at a single revlog. This is easy to add back on a case-by-case basis and does not need to stay with the more permanent timers.

rhg: add error message for paths outside the repository when cwd != root This mirrors the Python implementation. The relative path handling should probably be refactored into a util, but it it out of scope for this change.

rust: don't swallow valuable error information This helps when diagnosing corruption, and is in general good practice. The information is here, valuable and can be used easily.

rust: add message to `DirstateV2ParseError` to give some context This is useful when debugging.

py3: stop using deprecated Element.getchildren() method in convert/darcs This has been deprecated since py3.2, and removed entirely in py3.9

py3: fix bytes/unicode issues in convert/darcs - don't check for a binary symbol in globals(), which meant it always thought the module wasn't available - don't pass bytes to stdlib methods - return bytes in getchanges where Mercurial expects to see them

convert: remove old ElementTree import cruft from darcs All the `import elementtree` attempts seem to pre-date py2.5, when it was brought into the standard library, and the manual `cElementTree` fast implementation import has been unnecessary and deprecated since py3.3.

relnotes: add 6.2rc0

Added signature for changeset 288de6f5d724

Added tag 6.2rc0 for changeset 288de6f5d724

branching: merge default into stable

sparse: use the rust code even when sparse is present With rust supporting more matcher types, we can now take this route in the sparse case too.

sparse: directly inline the `set_tracked` and `copy` wrapping core is already aware of sparse, so lets move the handful of line of code that deal with this for the sake of simplicity and explicitness.

spares: clarify some test about merging copies This new output clarify the important part out this merging : we want to be able to record the copy source even it is outside of the sparse profile.

sparse: drop the useless wrapping of `dirstate.set_untracked` The code of the wrapper only act in the case where the file is not tracked, so this has not effect for `untracked`. In addition the message explicitly mention `add` of a file and no test are breaking if we drop this. So we drop this for simplicity and cleanup

sparse: directly inline the `rebuild` wrapping Core is already aware of sparse, so lets move the handful of line of code that deal with it in `dirstate.rebuild` for the sake of simplicity.

sparse: directly inline the `walk` wrapping core is already aware of sparse, so lets move the handful of line of code that deal with it in `dirstate.walk` for the sake of simplicity.

sparse: use None as the sparse matcher value when disabled This create a clear signal for when the feature is unused. We could also create an `alwaysmatcher`, but using None is more explicit, so I went for it.

sparse: start moving away from the global variable for detection of usage Code is now checking if the repository using the sparse feature and that's it. Some code in `debugsparse` still rely on "global" state, as it apply sparse logic before updating the requirement. Cleaning that up is more work that we signed up for, but we could narrow the hack to that specific command.

test: fix test-sparse-revlog output after debugdeltachain change We need to get this test covered by the CI, in the meantime, here is a quick fix.

rust-dirstate: add support for nevermatcher This is in case this ever comes up, it's very easy to support, so might as well do it.

rust-dirstate: add `intersectionmatcher` to the allowed matchers `IntersectionMatcher` is now implemented in Rust.

rust: add IntersectionMatcher This will be used in the upcoming support for sparse checkouts in Rust-augmented status and later in rhg support for sparse checkouts.

rust-dirstate: add `unionmatcher` to the allowed matchers `UnionMatcher` is now implemented in Rust.

rust: add UnionMatcher This will be used in the upcoming support for sparse checkouts in Rust-augmented status and later in rhg support for sparse checkouts.

hg-cpython: refactor matcher transformation logic This reduces duplication and will allow for recursive transformation in UnionMatcher.

rust: use owned types in `Matcher` This simplifies the code a lot, allows for some refactoring to come. The original code tried to prevent allocations that were already happening anyway beforehand.

hg-cpython: fallback when encountering an unknown matcher At this point in the process, nothing user-visible has happened, it is still safe to fallback. This can happen now that we're going to be using "container matchers" like unionmatcher and intersectionmatcher. This is easier and less error-prone than recursive checking beforehand since only the presence of a transformation case will allow the process to continue.

auto-upgrade: add an option to silence the safe-mismatch message For well tested case, the message can get in the way, so we add a way to disable it.

auto-upgrade: add an option to silence the tracked-hint message For well tested case, the message can get in the way, so we add a way to disable it.

auto-upgrade: add an option to silence the dirstate-v2 message For well tested case, the message can get in the way, so we add a way to disable it.

auto-upgrade: rename a variable to match the actual content This was the result of a copy paste.

auto-upgrade: add an option to silence the share-safe message For well tested case, the message can get in the way, so we add a way to disable it.

auto-upgrade: add a test that will host "quiet" testing We will add options to suppress the message in the coming changeset. The changes will be clearer if the full test is already in place.

bundlespec: check the `obsolescence` value before adding the caps This does not really matters as the logic to decide wether or not adding the part is correct and elsewhere. However this seems like a good idea to align this logic witht he semantic of the option.

bundlespec: handle the presence of obsmarker part This make `hg debugbundle --spec` more informative about extra part in the bundle.

bundlespec: fix the generation of bundlespec for `cg.version` If the value is non-default, we display it.

bundlespec: allow the bundle spec to control the bundle version This makes it possible to create bundle using changelog-v3. The `hg debugbundle --spec` output is borked, but this will be fixed in coming changesets.

bundlespec: do not check for `-` in the params portion of the bundlespec Otherwise bundle parameter with `-` in their names result in a crash.

bundlespec: add processing of some parameter value The boolean option needs to be turned into boolean.

bundlespec: extract the parseparams closure It has no value being a closure. We extract it before modifying it further.

bundlespec: test `no` value for the `obsolescence` parameter This is currently broken, but let us test for it first.

bundlespec: test that parameter overwrite the local config This is currently working, but havint it explicitly tested seems useful.

bundlespec: do not overwrite bundlespec value with the config one This is finally making the `obsolete` bundlespec paramater work.

bundlespec: add a `overwrite` parameter to set_param This will open the way for the configuration value to yield in front of the explicit bundle type.

bundlespec: merge the contentopts and params dictionnary They are content using the same keys. Using differents object for access open the gates for confusion in the code using them (this is already the case). So we start fusing their usages to make the parameters more useful. More work will be needed to make them really useful, but the first step is here: not throwing the value away. However this is still not making the previously introduced test useful because currently, the default config value overwrite the one from the bundlespec. We will fix this in the coming changesets.

bundlespec: test if the "obsolete=" parameter is read Narrator voice: Actually it is not. We will fix it in the next changeset.

bundlespec: fix lack of title in a evolve tests This is a full new test case and should be "flagged" as such.

bundlespec: make the `stream` case less special The handling of the stream case seems fragile (does not account for newer parts and options that will arise) and has special code dedicated to it. To simplify and strengthen things, we make it use the same mechanism as the other options. So we make it less special by making it a special value in the common case.

bundlespec: phase out the `_bundlespeccgversions` mapping The `_bundlespeccgversions` mapping is redundant with the `cg.version` parameter. We move all users to the `cg.version` version and phase out the `_bundlespeccgversions` mapping.

typing: add a missing suppression directive for `msvcrt`

windows: drop some py2 compatibility code The comment was wrong- the exception handler was the py3 case.

windows: prevent bytes from being passed to registry APIs There was a TortoiseHg bug report in this area[1], and from inspection, it looks like passing `b""` as `valname` would fail to convert to unicode. The underlying API allows both `""` and `NULL` to return the default value for the key. [1] https://foss.heptapod.net/mercurial/tortoisehg/thg/-/issues/5803

relnotes: add 6.1.3

cleanup: return directly instead of assigning variable

commit: remove special handling of IOError (actually dead code) In the past, IOError was used to mark a file as removed. The differentiation between OSError and IOError in this place was introduced in e553a425751d, to avoid that “normal” OSErrors / IOErrors accidentally mark files as removed. This weird internal API was removed in 650b5b6e75ed. It seems like that changeset should have removed the differentiation, at least I don’t see any reason for keeping it. On Python 3, OSError and IOError are aliased. Therefore the removed code was actually dead.

py3: catch specific OSError subclasses instead of checking errno On Python 3, the "not a directory" error is mapped to ENOTDIR instead of EINVAL. Therefore, catching the NotADirectoryError subclass is sufficient.

py3: catch specific OSError subclasses instead of checking errno Contrary to the previous changesets in this series, this covers cases where errno was checked for multiple values. EACCES -> PermissionError ENOENT -> FileNotFoundError ENOTDIR -> NotADirectoryError EISDIR -> IsADirectoryError

py3: catch ProcessLookupError instead of checking errno == ESRCH

py3: catch PermissionError instead of checking errno == EPERM

py3: catch PermissionError instead of checking errno == EACCES

py3: catch FileNotFoundError instead of checking errno == ENOENT

py3: catch FileExistsError instead of checking errno == EEXIST

py3: catch BrokenPipeError instead of checking errno == EPIPE

py3: catch ChildProcessError instead of checking errno == ECHILD

py3: remove retry on EINTR errno Since the implementation of PEP 475 (Python 3.5), Python retries system calls failing with EINTR. Therefore we don’t need the logic that retries it in Python code.

py3: stop catching TypeError that was raised on Python 2

py3: use `x.hex()` instead of `pycompat.sysstr(node.hex(x))`

tests: assume that `raw` attribute is present on original socket file object It seems like the original socket file object is always an io.BufferedIO instance. If not, the code will fail and we should try harder to get the socket object (e.g. if the original socket file object is unbuffered, we can get the `_sock` attribute directly from it).

tests: constant-fold a `pycompat.ispy3` in testlib/badserverext.py

tests: remove Python 2 special cases in test-stdio.py

py3: remove dead code to make file descriptors non-inheritable On Python 3, file descriptors are already non-inheritable by default.

py3: remove hack that removed flush argument from print() calls on Python 2

py3: remove long() compatibility code

zeroconf: constant-fold a `pycompat.ispy3` I’ve checked that both bytes and str gets passed as the `name` parameter, so the rest of the condition is still required. Because there aren’t really any tests for the extensions, I didn’t want to refactor it to pass a single type.

py3: remove conditional to import collections.abc.MutableMapping

py3: remove dead code to open file with O_CLOEXEC on Python 2 The O_CLOEXEC flag is passed by default on Python 3.

py3: don’t encode node.bin() argument It accepts str and bytes.

convert: inline Python 3 variant of url2pathname_like_subversion()

py3: constant-fold some `pycompat.ispy3`

py3: use `zip()` instead of trying to use `itertools.izip()`

py3: replace mention of “xrange” in docstring by “range”

py3: remove xrange() compatibility code Some code used its own xrange() compatibility code instead of pycompat.xrange().

py3: replace `pycompat.xrange` by `range`

hgweb: remove dead code handling UnicodeDecodeError I’m quite confident that the error can’t happen on Python 3, as the main motivation for separating bytes and str in Python 3 was to avoid this class of errors.

cleanup: remove import of already imported module

cleanup: rename some functions to avoid redefinitions

thirdparty: remove Python 2-specific selectors2 copy The selectors module was added in Python 3.4. Because we require Python 3.6, it will always be available. Therefore the selectors2 module is not imported. I’ve verified that the selectors2-specific workaround in commandserver.py is not necessary with the selectors module from the standard library. It returns an empty list if timeout was exceeded. The pytype directive was needed to silence the following error: File "/tmp/mercurial-ci/mercurial/worker.py", line 299, in _posixworker: No attribute 'close' on int [attribute-error] In Union[_typeshed.HasFileno, int] File "/tmp/mercurial-ci/mercurial/worker.py", line 299, in _posixworker: No attribute 'close' on _typeshed.HasFileno [attribute-error] In Union[_typeshed.HasFileno, int]

branching: merge stable into default

chg: replace mercurial.util.recvfds() by simpler pure Python implementation On Python 3, we have socket.socket.recvmsg(). This makes it possible to receive FDs in pure Python code. The new code behaves like the previous implementations, except that it’s more strict about the format of the ancillary data. This works because we know in which format the FDs are passed. Because the code is (and always has been) specific to chg (payload is 1 byte, number of passed FDs is limited) and we now have only one implementation and the code is very short, I decided to stop exposing a function in mercurial.util. Note on terminology: The SCM_RIGHTS mechanism is used to share open file descriptions to another process over a socket. The sending side passes an array of file descriptors and the receiving side receives an array of file descriptors. The file descriptors are different in general on both sides but refer to the same open file descriptions. The two terms are often conflated, even in the official documentation. That’s why I used “FD” above, which could mean both “file descriptor” and “file description”.

parsers: drop one extra argument to PyErr_Format GCC gave the following warning during `make local`: mercurial/cext/parsers.c: In function 'dirstate_item_from_v1_data': mercurial/cext/parsers.c:413:30: warning: too many arguments for format [-Wformat-extra-args] 413 | "unknown state: `%c` (%d, %d, %d)", state, mode, | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ To reproduce, you might need to add the -Wformat-extra-args flag, because it isn't present for me when building for the default python3. But I can see this warning while simply building 6.1 with `make PYTHON=python2 clean local`. I don't think this NULL was useful, because other instances of PyErr_Format() don't have any NULLs as the final argument, but keep in mind that I don't know python's C API.

search-discovery-case: update documentation of a function We return data, it is simpler when we know what these data means.

worker: avoid potential partial write of pickled data Previously, the code wrote the pickled data using os.write(). However, os.write() can write less bytes than passed to it. To trigger the problem, the pickled data had to be larger than 2147479552 bytes on my system. Instead, open a file object and pass it to pickle.dump(). This also has the advantage that it doesn’t buffer the whole pickled data in memory. Note that the opened file must be buffered because pickle doesn’t support unbuffered streams because unbuffered streams’ write() method might write less bytes than passed to it (like os.write()) but pickle.dump() relies on that all bytes are written (see https://github.com/python/cpython/issues/93050). The side effect of using a file object and a with statement is that wfd is explicitly closed now while it seems like before it was implicitly closed by process exit.

mr-template: drop the "title and description" I though I would affect the template in the menu itself, however it just a normal piece of text that should not be here.

debugindex: add a `rank` column

debugindex: add a `sd-chunk-size` column

debugindex: add a `sidedata-offset` column

debugindex: add a `sd-comp-mode` column

debugindex: add a `chunk-size` column

debugindex: add a `data-offset` column

debugindex: add a `comp-mode` column

debugindex: add a `flags` column

debugindex: add a `delta-base` column

debugindex: add a `full-size` column

debugindex: add a `p2-rev` column This will be useful in case of corrupted index.

debugindex: add a `p1-rev` column This will be useful in case of corrupted index.

debugindex: introduce a concept of "verbose-only" column We are about to add a bunch of new column and most of them are probably only relevant to --verbose. We add some more testing of the `--verbose` mode in a sidedata context.

debugindex: move to a flexible column Each column is now declared as a decorated function. This will make it much simpler to add more new column in the future.

debugindex: rename the parent column to mention nodeid We will add new columns with the "revnum" version of the parent. It will be useful in case we need to inspect a corrupted revlog index.

debugindex: align column name on the right It will be simpler to align everything on the same side, and the right seems a better side as it match the value alignment.

debugindex: move the logic into its own module Adding more information will significantly increase the amount of code. So we move the code into its own module before making it more complex.

debugindex: rename to debugindex debug-revlog-index The command dump some content of the revlog index and omit a lot of information. I am going to make it display the missing information. For clarity, we rename the command to explicitly mention revlog.

node: stop converting binascii.Error to TypeError in bin() Changeset f574cc00831a introduced the wrapper, to make bin() behave like on Python 2, where it raised TypeError in many cases. Another previous approach, changing callers to catch binascii.Error in addition to TypeError, was backed out after negative review feedback [1]. However, I think it’s worth reconsidering the approach. Now that we’re on Python 3 only, callers have to catch only binascii.Error instead of both. Catching binascii.Error instead of TypeError has the advantage that it’s less likely to cover a programming error (e.g. passing an int to bin() raises TypeError). Also, raising TypeError never made sense semantically when bin() got an argument of valid type. As a side-effect, this fixed an exception in test-http-bad-server.t. The TODO was outdated: it was not an uncaught ValueError in batch.results() but uncaught TypeError from the now removed wrapper. Now that bin() raises binascii.Error instead of TypeError, it gets converted to a proper error in wirepeer.heads.<locals>.decode() that catches ValueError (superclass of binascii.Error). This is a good example of why this changeset is a good idea. Catching TypeError instead of ValueError there would not make much sense. [1] https://phab.mercurial-scm.org/D2244

revlog: make try block smaller Making try blocks as small as possible is generally a good idea, especially when catching very general errors like TypeError.

revlog: make round-down pattern clearer

branching: merge stable into default

test-revlog: adds a new root revision with a delta against nullrev We add a revision with null parent but that is not the first revision of the revlog. It make it a different a case that is worthy of testing.

test-revlog: adds a non-root revision with a delta against nullrev Using a revision with non-null parents makes it a different case that is worthy of testing.

test-revlog: update the patch used to test delta againts nullrev We need to be careful to not create invalid delta for changelog. Changelog is special because it does not use any delta in practice and don't use general delta in its format. Using the new patch will help use to test more cases.

tests: remove hg10 requirement from test-check-pylint.t Since pylint does its own directory traversal to find files to check and doesn't use hg locate command, this requirement is not valid. See also a29f071751df.

tests: add hg10 requirement to test-check-module-imports.t This requirement exists to make sure for hg has locate command, which is used in this test file for getting a list of files to check. Since in theory current versions of tests could be run on older hg, it doesn't hurt to mark even this version requirement (as old as it is). See also: a29f071751df.

worker: implement _blockingreader.readinto() (issue6444) The core logic for readinto() was already implemented in read(), so this is mostly extracting that code into its own method. Another fix for issue6444 was committed to the stable branch: 2fe4efaa59af. That is a minimal fix that implements readinto() only on Python versions that require readinto() (3.8.0 and 3.8.1), which is the right approach for the stable branch. However, I think that this changeset has its value. It improves performance in cases when pickle can use readinto(), it reduces code duplication compared to the other patch, and by defining readinto() on all Python versions, it makes behavior more consistent across all Python versions. This changesets reverts the other change.

worker: stop relying on garbage collection to release memoryview On CPython, before resizing the bytearray, all memoryviews referencing it must be released. Before this change, we ensured that all references to them were deleted. On CPython, this was enough to set the reference count to zero, which results in garbage collecting and releasing them. On PyPy, releasing the memoryviews is not necessary because they are implemented differently. If it would be necessary however, ensuring that all references are deleted would not be suffient because PyPy doesn’t use reference counting. By using with statements that take care of releasing the memoryviews, we ensure that the bytearray is resizable without relying on implementation details. So while this doesn’t fix any observable bug, it increases compatiblity with other and future Python implementations.

worker: add docstring to _blockingreader

worker: explain why pickle reading stream has to be unbuffered

bundle: quick fix to ludicrous performance penalty We tried a `hg bundle --base ':(tip^)' --rev 'all()'` on a large repository and it spent 3 minutes on this 2 list comprehensions. This change remove this cost. There are still a lot of low hanging fruits as the command still take 30 seconds. However this is a trivial patch with a massive speedup so I'll just sent it.

deltas: add a debug-delta-find command to analyse delta search See command documentation for details. For some reason, pytype is confused by our usage of None/deltainfo variable, so I had to quiet it.

deltas: add a `debug.revlog.debug-delta` config option enable output This provide a way to enable the code introduced in the previous changeset. This will provide a large amount of output when applying a bundle with details about each delta "computation".

deltas: add code to display information about the result of `finddeltainfo` I have been looking into performance issue around pull and getting more information about the computation and time involved into applying each revision is very useful. There will be various way to use this new output, so I am introducing the code first.

test: update `test-sparse-revlog` output This got changed at some point.

branching: fix wrong merge conflict resolution from 13dfad0f9f7a 13dfad0f9f7a merged stable into default, but accidentally added the _blockingreader class from stable (but deindented) instead of merging the changes from stable (2fe4efaa59af) into the existing _blockingreader class. This resulted in the _blockingreader being there two times.

debugdeltachain: detect a special case where parents are "skipped" See inline comment for details, this is a case where the delta is neither against p1 or p2, Yet it is still a simple delta part of a simple chain. We now display them as `skip1/skip2` instead of `other`.

debugdeltachain: document the possible value for deltatype So that one can understand what is displayed by the command.

debugdeltachain: also display p1/p2 Looking at the parents is a common need when trying to understanding why a delta was chosen, having it readily available helps a lot.

debugdeltachain: glob variance of "test-generaldelta" We mostly care about generaldelta happening, the exact details of storage size variation is not really important so we can glob it instead of having multiple lines for each variances. This will make updating the output of the command simpler.

debugdeltachain: use the symbolic constant to access entry information This is more robust and easier to read

debugdeltachain: distinct between snapshot and "other" diffs Snapshot are expected to be healthy behavior, while "other" is a bit more suspicious. So we distinct between the two to make it easier to inspect repositories.

branching: merge stable into default

revlog: use appropriate format char for int ("i" instead of "I") From https://docs.python.org/3.8/c-api/arg.html#numbers : i (int) [int] Convert a Python integer to a plain C int. I (int) [unsigned int] Convert a Python integer to a C unsigned int, without overflow checking.

revlog: use %d to format int instead of %lu (issue6565) The issue says gcc warns that the data types don't match. I couldn't reproduce the warning locally for some reason, but this patch shouldn't break things. Maybe %lu was simply a copy-paste error from 6b1eae313b2f (https://phab.mercurial-scm.org/D10625).

filelog: show the passed argument on error The error now do more than stating what it need. It also state what it got.

manifest: improve error message in case for tree manifest In the case where tree manifest is not enabled but we still receive an sub directory information for the manifest. The error now inform which sub-directory was passed.

branchmap: use a context manager when writing the branchmap This is cleaner and safer. The previous code date from long before we had context manager available.

auto-upgrade: skip the operation if the repository cannot be locked This seems like a fine default behavior for now. If some users wants something more aggressive we can make the behavior configurable in the future. Differential Revision: https://phab.mercurial-scm.org/D12619

auto-upgrade: add a test case where the repository is already locked This show the current behavior when the repository to auto-upgrade is already locked. The current behavior is to abort, which is probably not great. Now that we have a proper test, we can think about the behavior we wants in a later tests. Differential Revision: https://phab.mercurial-scm.org/D12618