release v0.21.18

Caused by 2bc127bb

This reverts commit 4475b8ca.  Further
testing revealed that the threaded resolver still uses a timeout of
0ms.  This revert however lowers the bound to a minimum of 1ms instead
of 10ms.

Like fe598e7d

The packages from the Travis image are good enough, and this speeds up
the Travis build.

Don't bother sending analytics data for a CI runner.  This adds
traffic but doesn't bring anybody any benefit.

This prepares the migration away from strptime() for Windows
portability.

But the real reason I'm doing this is that strptime() on Apple is
buggy: strptime("14", "%H%M%S") (without separating colons) succeeds
even though only the hour has been parsed.  This fixes recent Travis
failures in the ParseISO8601() unit test.

Older clang versions don't support the GCC __attribute__ syntax.  For
those, don't use anything at all, and new clang versions shall use the
standard syntax.

Expand $PATH at evaluation and not at assignment, which fixes the
problem that /usr/lib/ccache was added to $PATH between the
MATRIX_EVAL assignment and its evaluation.

Works around build failures with ccache which may feed processed code
to GCC, which doesn't have the "fall through" code comments.

curl_multi_remove_handle() calls our socket function, and there's no
need to call curl_multi_socket_action().

This was a problem 9 years ago, and apparently, it has been fixed long
ago.

For simplicity, this commit removes a workaround for an old CURL bug.

Can fail if somebody calls Break().

This can cause request completion in the I/O thread before this
constructor returns, leaving the object in an abstract state, causing
a crash due to pure virtual method call.  We should not start the
request until this object is fully constructed.

Closes https://github.com/MusicPlayerDaemon/MPD/issues/665

Works around assertion failure in the boost::intrusive::set class
because we modified the container between insert_check() and
insert_commit().

Closes https://github.com/MusicPlayerDaemon/MPD/issues/691

The ALSA "null" driver opens /dev/null and returns the file handle
from snd_pcm_poll_descriptors(), but /dev/null cannot be used with
epoll, the epoll_ctl() system call returns -EPERM.  This means that
the ALSA output hangs, eventually freezing the whole MPD process.

This commit adds a workaround to the MultiSocketMonitor class which is
used by the ALSA output plugin.

Closes https://github.com/MusicPlayerDaemon/MPD/issues/695

SetReady() does this already.