The process is already gone, so there's not going to be any new data
showing up on its stderr; we only want to make sure that we haven't
missed something that was already written. So polling once is enough.
This change is motivated by an ssh oddity: when ControlPersist is
enabled, the first (i.e. master) connection goes into the background; we
see EOF on its stdout and the process exits, but we never see EOF on its
stderr. So if we ran a command like this:
ANSIBLE_SSH_PIPELINING=1 ansible -T 30 -vvv somehost -u someuser -m command -a whoami
We would first do select([stdout,stderr], timeout) and read the command
module output, then select([stdout,stderr], timeout) again and read EOF
on stdout, then select([stderr], timeout) AGAIN (though the process has
exited), and select() would wait for the full timeout before returning
rfd=[], and then we would exit. The use of a very short timeout in the
code masked the underlying problem (that we don't see EOF on stderr).
It's always preferable to call select() with a long timeout so that the
process doesn't use any CPU until one of the events it's interested in
happens (and then select will return independent of elapsed time).
(A long timeout value means "if nothing happens, sleep for up to <x>";
omitting the timeout value means "if nothing happens, sleep forever";
specifying a zero timeout means "don't sleep at all", i.e. poll for
events and return immediately.)
This commit uses a long timeout, but explicitly detects the condition
where we've seen EOF on stdout and the process has exited, but we have
not seen EOF on stderr. If and only if that happens, it reruns select()
with a short timeout (in practice it could just exit at that point, but
I chose to be extra cautious). As a result, we end up calling select()
far less often, and use less CPU while waiting, but don't sleep for a
long time waiting for something that will never happen.
Note that we don't omit the timeout to select() altogether because if
we're waiting for an escalation prompt, we DO want to give up with an
error after some time. We also don't set exceptfds, because we're not
actually acting on any notifications of exceptional conditions.
On Python 2, shlex.split() raises if you pass it a unicode object with
non-ASCII characters in it. The Ansible codebase copes by explicitly
converting the string using to_bytes() before passing it to
shlex.split().
On Python 3, shlex.split() raises ('bytes' object has no attribute 'read')
if you pass a bytes object. Oops.
This commit introduces a new wrapper function, shlex_split, that
transparently performs the to_bytes/to_unicode conversions only on
Python 2.
Currently I've only converted one call site (the one that was causing a
unit test to fail on Python 3). If this approach is deemed suitable,
I'll convert them all.
Without this, we could execute «ssh -q ...» and call select(), which
would timeout after the default 10s, and only then send initial data.
(This is a relic of the earlier change where we always ran ssh with
-vvv, so the situation where it would sit quietly never happened in
practice; but this would have been the right thing to do even then.)
Make the code compatible with Pythons 2.4 through 3.5 by using
sys.exc_info()[1] instead.
This is necessary but not sufficient for Python 3 compatibility.
The event loop (even after it was brought into one place in _run in the
previous commit) was hard to follow. The states and transitions weren't
clear or documented, and the privilege escalation code was non-blocking
while the rest was blocking.
Now we have a state machine with four states: awaiting_prompt,
awaiting_escalation, ready_to_send (initial data), and awaiting_exit.
The actions in each state and the transitions between then are clearly
documented.
The check_incorrect_password() method no longer checks for empty strings
(since they will always match), and check_become_success() uses equality
rather than a substring match to avoid thinking an echoed command is an
indication of successful escalation. Also adds a check_missing_password
connection method to detect the error from sudo -n/doas -n.
The main exec_command/put_file/fetch_file methods now _build_command and
call _run to handle input from/output to the ssh process. The purpose is
to bring connection handling together in one place so that the locking
doesn't have to be split across functions.
Note that this doesn't change the privilege escalation and connection IO
code at all—just puts it all into one function.
Most of the changes are just moving code from one place to another (e.g.
from _connect to _build_command, from _exec_command and _communicate to
_run), but there are some other notable changes:
1. We test for the existence of sshpass the first time we need to use
password authentication, and remember the result.
2. We set _persistent in _build_command if we're using ControlPersist,
for later use in close(). (The detection could be smarter.)
3. Some apparently inadvertent inconsistencies between put_file and
fetch_file (e.g. argument quoting, sftp -b use) have been removed.
Also reorders functions into a logical sequence, removes unused imports
and functions, etc.
Aside: the high-level EXEC/PUT/FETCH description should really be logged
from ConnectionBase, while individual subclasses log transport-specific
details.
* Make LookupBase an abc with required methods (run()) marked as an
abstractmethod
* Mark methods that don't use self as @staticmethod
* Document how to implement the run method of a lookup plugin.
Toshio Kuratomi