fix task execution/lockstep issues with clear_host_errors

[s-hertel]

SUMMARY

Fixes #35086

Updating the failed host state's also helps fix a clear_host_errors issue where hosts are not brought back into play in lockstep.

Also fixes clearing host errors across plays.

ISSUE TYPE

• Bugfix Pull Request

[mkrizek]

As a note I looked at fixing 31543 in the past and had this in git stash:

diff --git a/lib/ansible/plugins/strategy/linear.py b/lib/ansible/plugins/strategy/linear.py
index d90d347d3e3..01c2bbfeb01 100644
--- a/lib/ansible/plugins/strategy/linear.py
+++ b/lib/ansible/plugins/strategy/linear.py
@@ -411,23 +411,20 @@ class StrategyModule(StrategyBase):
                 for res in results:
                     # execute_meta() does not set 'failed' in the TaskResult
                     # so we skip checking it with the meta tasks and look just at the iterator
-                    if (res.is_failed() or res._task.action in C._ACTION_META) and iterator.is_failed(res._host):
+                    #if (res.is_failed() or res._task.action in C._ACTION_META) and iterator.is_failed(res._host):
+                    if res.is_failed():
                         failed_hosts.append(res._host.name)
                     elif res.is_unreachable():
                         unreachable_hosts.append(res._host.name)

                 # if any_errors_fatal and we had an error, mark all hosts as failed
                 if any_errors_fatal and (len(failed_hosts) > 0 or len(unreachable_hosts) > 0):
-                    dont_fail_states = frozenset([IteratingStates.RESCUE, IteratingStates.ALWAYS])
+                    # would probably need a new flag to abort the play (and for PlayIterator to revert the flag when rescue)
+                    iterator.end_play = True
+                    # FIXME self._tqm.RUN_FAILED_BREAK_PLAY
                     for host in hosts_left:
-                        (s, _) = iterator.get_next_task_for_host(host, peek=True)
-                        # the state may actually be in a child state, use the get_active_state()
-                        # method in the iterator to figure out the true active state
-                        s = iterator.get_active_state(s)
-                        if s.run_state not in dont_fail_states or \
-                           s.run_state == IteratingStates.RESCUE and s.fail_state & FailedStates.RESCUE != 0:
-                            self._tqm._failed_hosts[host.name] = True
-                            result |= self._tqm.RUN_FAILED_BREAK_PLAY
+                        if host.name not in failed_hosts:
+                            #self._tqm._failed_hosts[host.name] = True
+                            iterator.mark_host_failed(host)
                 display.debug("done checking for any_errors_fatal")

                 display.debug("checking for max_fail_percentage")

Basically the idea was to utilize iterator.mark_host_errors() to fail other hosts and for PlayIterator to handle the case where rescue section is present to undo the failure(s). You can see that I commented out some lines, it probably broke something but I am not sure what it was. One thing I know is that with the diff the return code of ansible-playbook changes since strategy does not return self._tqm.RUN_FAILED_BREAK_PLAY anymore.

It would also fix #73246 it seems.

I feel like there are too many fixes in the current (devel) implemetation (dont_fail_states, using iterator.get_active_state) at this point.

There is probably a reason why this wasn't done like this when it was originally written though? 🧐

Anyway just thought I'd share this in case you wanted to experiment with that idea as well.