mercurial: Changelog

procutil: redirect ui.fout to stderr while stdio is protected The new behavior seems slightly nicer as we can at least read the output. And this is similar to what the sshserver is doing, so we can probably reuse protectstdio() instead of the weird hook.redirect(True) hack.

procutil: unroll uin/uout loop in protectstdio() I'll change uout to be redirected to stderr.

context: drop support for changeid of type long (API?) I don't see a reason to support type long. It's pretty much the same type as int. There was some discussion about it on the mailing list around the time of ff2f90503d64 (context: work around `long` not existing on Python 3, 2017-03-11), but I couldn't find a good reason to keep it. There was some mention of hgtk doing "repo[long(rev)]", but that was in 2012. Differential Revision: https://phab.mercurial-scm.org/D2989

lfs: drop a duplicate blob verification method

server: minor code cleanup Suggested by Yuya after 77f9e95fe3c4, this is mostly using named values for stdio descriptors. The lockfile is also opened in binary mode when reading back content from the child.

server: refactor 'daemon_postexec' instructions into a dictionary

cbor: import CBORDecoder and CBOREncoder And format the imports so it is cleaner. Differential Revision: https://phab.mercurial-scm.org/D2978

py3: fix fix doctests to be bytes-safe

server: add an error feedback mechanism for when the daemon fails to launch There's a recurring problem on Windows where `hg serve -d` will randomly fail to spawn a detached process. The reason for the failure is completely hidden, and it takes hours to get a single failure on my laptop. All this does is redirect stdout/stderr of the child to a file until the lock file is freed, and then the parent dumps it out if it fails to spawn. I chose to put the output into the lock file because that is always cleaned up. There's no way to report errors after that anyway. On Windows, killdaemons.py is roughly `kill -9`, so this ensures that junk won't pile up. This may end up being a case of EADDRINUSE. At least that's what I saw spit out a few times (among other odd errors and missing output on Windows). But I also managed to get the same thing on Fedora 26 by running test-hgwebdir.t with --loop -j10 for several hours. Running `netstat` immediately after killing that run printed a wall of sockets in the TIME_WAIT state, which were gone a couple seconds later. I couldn't match up ports that failed, because --loop doesn't print out the message about the port that was used. So maybe the fix is to rotate the use of HGPORT[12] in the tests. But, let's collect some more data first.