With epoll() there is only one kernel-side object per file descriptor,
which is why _control() is such a pain. Since we merge receive/transmit
watching into that single object, we must always test the mask for both
conditions when reading results.
Kqueue isn't/doesn't appear to be like this. The identity of a Kqueue
event is keyed on (fd, filter), and we register a separate event for
both transmit and receive, so the 'elif' in KqueuePoller.poll() does not
appear to need to change.
Previously, a FD marked for read+write would not indicate writeability
until it was no longer readable.
To support detach, we must be able to preload the target with every
module it will need prior to detachment. This implements the
intermediary part of the process (i.e. the Ansible fork parent) --
receiving LOAD_MODULE/FORWARD_MODULE pairs and ensuring they reach the
child.
This may come back to bite later, but in the meantime it avoids shipping
up to 12KiB of junk metadata for every single task invocation.
For detachment (aka. async), we must ensure the target has two types of
preloads completed (modules and module_utils files) before detaching.
Bootstrap would hang if (as of writing) a pipe sufficient to hold 42,006
bytes was not handed out by the kernel to the first stage. It was luck
that this didn't manifest before, as first stage could write the full
source and exit completely before reading begun.
It is not clear under which circumstances this could previously occur,
but at least since Linux 4.5, it can be triggered if
/proc/sys/fs/pipe-max-size is reduced from the default of 1MiB, which
can have the effect of capping the default pipe buffer size of 64KiB to
something lower.
Suspicion is that buffer pipe size could also be reduced under memory
pressure, as reference to busy machines appeared a few times in the bug
report.
David Wilson