dispatch: handle late KeyboardInterrupt occurred in run() User can press Ctrl+C while flushing streams in dispatch.run(). In such case, I think exiting with 255 is better than printing Python traceback and exiting with 1.

dispatch: indent run() function I'll add KeyboardInterrupt handling there.

debuginstall: don't translate encoding messages While Microsoft trained us to read mojibake text, it doesn't make sense to print mojibake messages when debugging the exact issue.

merge with stable

patch: refactor content diffing part in separate fn so extensions can wrap Right now extdiff uses it's own logic using archival to diff two versions of file using external diff tools. This makes the extdiff functionality non-extensible. This series is an attempt to refactor core patch and diff functionality so that extdiff can wrap and reuse it. This will help us in using external diffing tools at more places and not just extdiff command only then. Differential Revision: https://phab.mercurial-scm.org/D8685

tests: check that procutil.std{out,err}.write() returns correct result On Windows, we currently don’t fully test the case when the stream is connected to a TTY, but we test the child process side by connecting them to NUL, which is recognized as a TTY by Python. To make the large write test a bit more useful besides checking that it doesn’t crash, we can check that the write() method returns the correct result.

tests: add tests for when stdout or stderr is connected to `os.devnull` The original motivation was that creating PTYs on Windows is not possible, but `NUL` is recognized as a TTY, so we can have at least some test coverage for the TTY case. I think it doesn’t hurt to run the test cases on all systems.

procutil: ensure that procutil.std{out,err}.write() writes all bytes Python 3 offers different kind of streams and it’s not guaranteed for all of them that calling write() writes all bytes. When Python is started in unbuffered mode, sys.std{out,err}.buffer are instances of io.FileIO, whose write() can write less bytes for platform-specific reasons (e.g. Linux has a 0x7ffff000 bytes maximum and could write less if interrupted by a signal; when writing to Windows consoles, it’s limited to 32767 bytes to avoid the "not enough space" error). This can lead to silent loss of data, both when using sys.std{out,err}.buffer (which may in fact not be a buffered stream) and when using the text streams sys.std{out,err} (I’ve created a CPython bug report for that: https://bugs.python.org/issue41221). Python may fix the problem at some point. For now, we implement our own wrapper for procutil.std{out,err} that calls the raw stream’s write() method until all bytes have been written. We don’t use sys.std{out,err} for larger writes, so I think it’s not worth the effort to patch them.

procutil: move assignments This should probably be part of the previous patch, but folding it results in a less useful word diff, so I decided to keep it separate for review.

procutil: distribute code for stdout It makes sense to have the distinction between Python 2 and 3 at the top level, as we have to fight a different kind of battle on each: On Python 3, we get consistent behavior on all platforms, but need to create correctly-behaving binary streams. On Python 2, we have to account for platform differences.

convert: use bytes for value in extra dict The keys and values of the extra dict are expected to be bytes. Before this fix, there was a crash in mercurial.changelog.encodeextra().

overlayworkingctx: remove unused `nofilechanges()` and `_compact()` methods Differential Revision: https://phab.mercurial-scm.org/D8733

rebase: correctly check for empty commit in in-memory mode The new code has a small overhead in the empty commit case, as a `memctx` object is always created, but I don’t think it’s justified here to duplicate code to optimize a relatively unlikely code path. Differential Revision: https://phab.mercurial-scm.org/D8732

memctx: make `parents()` return list of one element if it’s not a merge It is part of the contract of `parents()` that non-merges return a list of one element. `self._parents` is always a list of two elements in `memctx`. Differential Revision: https://phab.mercurial-scm.org/D8731

commit: factor out empty commit check to `basectx.isempty()` This enables reuse in other places, e.g. those dealing with `memctx`. Differential Revision: https://phab.mercurial-scm.org/D8729

overlayworkingctx: rename misleadingly named `isempty()` method This method is only about whether there are file changes, not about whether the commit will be empty or not. One user of the method was incorrectly assuming the latter meaning, leading to the bug for which a test case was added in D8727. I’ve added a FIXME to the code. The original motivation for the rename was that I want to add `committablectx.isempty()`, that properly checks if a commit will be empty, using the exact same logic as in `repo.commit()`, and I wanted to avoid a name clash. Differential Revision: https://phab.mercurial-scm.org/D8728

rebase: add test case that shows that inmemory rebase does not preserve merges Differential Revision: https://phab.mercurial-scm.org/D8727

rebase: remove now unnecessary logic to allow empty commit when branch changes This was a workaround for a bug in the empty commit check in repo.commit(), where the parent branch name was incorrectly compared with the wdir branch name instead of the branch name passed via `extra`. The bug was fixed in D8724. The workaround was introduced in b2415e94b2f5. Differential Revision: https://phab.mercurial-scm.org/D8725

commit: check branch change on to-be-committed changeset This makes a difference when a branch name is passed via the `extra` argument, which will be used as the branch name for the created changeset. The empty commit check should use that branch name instead of whatever was set on the working directory. Besides fixing the bug for which a test case was extended, it enables to remove a workaround in the rebase extension, which will be done in a follow-up patch (D8725). Differential Revision: https://phab.mercurial-scm.org/D8724

merge with stable

procutil: move comment closer to relevant code

procutil: use mercurial.windows.winstdout only on Python 2 and TTYs Python 3 already works around the bug. The workaround is only needed when writing to consoles. If stdout is a console, sys.stdout.isatty() is true.

procutil: split if condition This prepares the code for subsequent changes where we need to differentiate between the two cases.

procutil: explain better why line buffering is not possible The sentence “On Python 3, buffered binary streams can't be set line-buffered.” was imprecise, as all streams are just Python classes and we can implement our own (which we did).

tests: make subprocess handling reusable for different tests in test-stdio.py

procutil: back out 8403cc54bc83 (make ....procutil.stderr unbuffered) Changeset 8403cc54bc83 introduced code that opens a second file object referring to the stderr file descriptor. This broke tests on Windows. The reason is that on Windows, sys.stderr is buffered and procutil.stderr closed the file descriptor when it got garbage collected before sys.stderr had the chance to flush buffered data. `procutil.stdout` had the same problem for a long time, but we didn’t realize, as in CI test runs, stdout is not a TTY and in this case no second file object is opened.

test-clone-uncompressed: use config to silence the lock warning The ui.timeout.warn option can solve the issue6237 in a cleaner way than what was done in 752da6863e39. That way, possible other unexpected error output will not be silenced. Differential Revision: https://phab.mercurial-scm.org/D8726

hooklib: fix detection of successors for changeset_obsoleted Provide a hook for obsutil.getobsolete to be used with either a transaction or the changes item of the transaction, since hooks only have access to the latter. Use that to find the correct list of revisions with obsmarkers, even new ones, and then filter out revisions with known successors. Move the processing from pretxnclose to txnclose as the transaction access itself is no longer necessary. This is more in line with notify and ensures that sanity checks can abort the transaction first. Differential Revision: https://phab.mercurial-scm.org/D8575

fix: prefetch file contents This prevents the worker subprocesses from contacting the server individually, which is either inefficient, or leads to problems if the connection is shared among them. Differential Revision: https://phab.mercurial-scm.org/D8723

fix: obtain base paths before starting workers This moves calculation of base paths to before work is dispatched. While this does mean that copy tracing will be serialized instead of parallel, it is necessary to be able to prefetch the base contents in a batch, which will likely be more efficient. Differential Revision: https://phab.mercurial-scm.org/D8722