As reported by teor, stem.descriptor's parse_file() function cannot accept
stdin...

  https://trac.torproject.org/projects/tor/ticket/23859

The trouble is that not all file objects in python are seekable. I'd *like*
to handle address this transparently by buffering the content...

  try:
    descriptor_file.tell()
  except IOError:
    # file's not seekable, wrapping in a buffer that is

    descriptor_file = io.BytesIO(descriptor_file.read())

This works great if our stream has content...

  % cat my_descriptors | python demo.py

*But* hangs indefinitely if no EOF is present in the stream.

  % python demo.py   <= hangs

Turns out non-blocking, platform independent reading of streams like stdin is
pretty tricky...

  http://eyalarubas.com/python-subproc-nonblock.html

As such simply providing callers with a more descriptive exception. If they
know their stream won't block *they* can add the above wrapper to provide
us with a seekable file object.

Irc discussion Dave partook in indicated that we should flush our cache when we
receive a DNS_USELESS event, not just EXTERNAL_ADDRESS.

  https://trac.torproject.org/projects/tor/ticket/26129

When not configured to be a tor relay tor's 'GETINFO exit-policy/full' query
won't function. Rather than attempt to guess the exit policy from other
configuration options providing None for 552 responses...

  https://trac.torproject.org/projects/tor/ticket/25853
  https://gitweb.torproject.org/torspec.git/commit/?id=c5453a0

The above spec addition makes it ambiguous if 552 means 'not a relay' or
'we had an error' so reached out for clarificaition on how these should
be handled...

  https://trac.torproject.org/projects/tor/ticket/27034

In the process also had to fix get_info to raise OperationFailed rather than
a ProtocolError so callers can get the underlying response codes.

Couple regressions that mostly slipped in with the recent hashing changes.

  https://trac.torproject.org/projects/tor/ticket/27030

Thanks to rl1987 tor now has a GETINFO method to retrieve microdescriptors
(similar to 'GETINFO desc/all-recent', 'GETINFO ns/all', etc).

  https://gitweb.torproject.org/torspec.git/commit/?id=b5396d5
  https://trac.torproject.org/projects/tor/ticket/26210

Good point from irl that our tests fail if the gitignore file isn't present...

  https://trac.torproject.org/projects/tor/ticket/26984

  % ./run_tests.py --all
  Traceback (most recent call last):
    File "./run_tests.py", line 36, in <module>
      import test
    File "/home/atagar/Desktop/stem/test/__init__.py", line 71, in <module>
      with open(os.path.join(STEM_BASE, '.gitignore')) as ignore_file:
  IOError: [Errno 2] No such file or directory: '/home/atagar/Desktop/stem/.gitignore'

Link protocols are ints. I made a LinkProtocol class so we can bundle
protocol characteristics (field size and such) but at the end of the
day it's nice for LinkProtocol to still behave like an int.

Dave made the good point that LinkProtocols should be comparable with
ints...

  https://trac.torproject.org/projects/tor/ticket/26432

Taking that one step further and making them in int subclass instead.
Experimenting, the __init__ and __new__ methods seem to work in practice
for our purposes but the python docs advise __new__ so using that...

  https://jfine-python-classes.readthedocs.io/en/latest/subclass-int.html

Python3 evidently no longer bundles distutils. We only used it in our utils for
simplicity, prior to commit 943289db we manually crawled the path. Effectively
reverting commit 943289db.

  https://trac.torproject.org/projects/tor/ticket/26967

Running `tox` (or e.g. `tox -- -u`) will now be substantially more
useful, by default.

Fixes #26916

  https://trac.torproject.org/projects/tor/ticket/26916

In practice, it appears to redundantly reference whatever py3 version is
installed.

Note that this applies only to the summary and exit status of tox.
During operation, tox will still display an error for each missing
interpreter.
(This may be confusing.)

The initial dir-spec definition of this field allowed blank values. Then it got
revised to disallow them. Now they're allowed again. ;P

  https://gitweb.torproject.org/torspec.git/commit/?id=2cce07d

No big whoop. Until we cut a release this is easy to twiddle.

Skimmed over our history to fill in changelog gaps. We only include user facing
adjustments to existing features so this didn't include most of our recent work
(stem.client wasn't part of Stem 1.6, and we don't note testing or development
changes).

Matt's narrowed sbws issues he's been having to a buggy implementation
of lru_cache in python 3.5. Using our bundled version instead for that
particular series.

Immutable classes have immutable cache values, and as frequntly used
equality/dictionary values it's useful to cache these.

Hashing classes (ie. "hash(type(my_obj))") sometimes provides inconsistant
values between interpreter invocations. Common builtin types (int, str, etc)
are consistant, but custom classes appear to vary.

Generally speaking this is fine. Hash value consistancy is only highly
important within an interpreter invocation. But full consistancy is nice and
easy to achieve.

Hashing the string representation of our module name instead.

We parse quite a number of 'key=value' descriptor fields. Clear spot where we
can deduplicate things a bit.

This flag, denoted as deprecated in pip 9, was causing pip 10 to error,
preventing tox from being used for new virtualenvs.

Fixes #26811

  https://trac.torproject.org/projects/tor/ticket/26811

Router status entry 'a' lines previously only appeared in the standard
consensus but now they appear in the microdescriptor consensus as well...

  https://trac.torproject.org/projects/tor/ticket/26405
  https://trac.torproject.org/projects/tor/ticket/23826

Newish change in tor's dirauth configuration.

Resyncing our manual information from tor to pick up new torrc options.

Previously, unpacking would popular the 'unused' attribute with any
remainder content, but it would never be used by anything.

This also has various testing improvements to make the Cell tests more
data-driven.

  https://trac.torproject.org/projects/tor/ticket/26766

Especially with the test case added just below the line this commit
changes, it becomes easier to identify the bytes that specify the
payload length and see it change line-to-line.

All implemented cell types that may have 'unused' content now have a
test case that checks it.

This could've also been done inline in the earlier commit, but wasn't so
that it would be easier to identify the moved content therein.

This could've been directly fixed in the previous commit, but it was
such a minor change that doing so might've hidden it, so it's been
separated out for clarity.

By refactoring this check into test data, and with the logic set up in
previous commits, the commit also ensures more test coverage for this
case.

Also clean up the comments for context.

This is the other logical thing to do (the complement of the previous
commit) with the test cases.

Note that just because the logic is in the tests now doesn't necessarily
mean that there's test cases that use it yet.

Don't try to test cell packing with non-blank,non-ZERO 'unused' content without specifying it in constructor

The test for DestroyCell already did this, so apply this structure to
tests of all other cell types that may have 'unused' content.

And since this is now such a common pattern in the test file, having a
comment for it is less necessary, so also removing that.

This was the only implemented cell type without any tests of 'unused'
content, so now all have tests. However they don't necessarily test
non-blank or non-ZERO data yet.

This is in preparation to make it much easier to manage data-driven test
cases for 'unused' content.

This removes most - but not all - use of magic numbers for link_protocol
values.

This fixes #26766, as well as makes the aforementioned failing
DestroyCell test pass.

Also change a few tests that used _pack() directly, since the parameter
order changed with keyword arguments.