cover:export/1 never returns in Erlang 27

[Michael Uvarov]

Describe the bug
This call never returns in Erlang 27 (even after 3 hours):

cover:export("/tmp/mongoose_combined.coverdata").

To Reproduce
Tricky setup: we get it on MongooseIM test CI. So, we have 5 nodes we have to collect coverage from. We have a lot of tests running before, so there could be a lot of coverage data.

Expected behavior
Same as in Erlang 26: returns after a couple of second.

Affected versions
27

Additional context
I would expect something is wrong with async/remote_calls in the cover.erl. It has a lot of places where it spawns a new process, which would send a reply. If something wrong happens in that new process, caller would never be notified.
It is a bit tricky to figure out what is wrong because of a lot of pmap and remote calls (and the fact that it is hard to get current_stacktraces of these spawned processes). Hope to add more info after debugging.

[John Högberg]

Thanks for your report! Unfortunately most of us are on vacation right now, we'll revisit this once we're back. :-)

[Michael Uvarov]

We meck mongoose_tcp_listener:

(mongooseim@localhost)4> code:get_coverage(cover_id_line, mongoose_acc).
[{1,0},
 {3,0},
...
 {28,...},
 {...}|...]
 
(mongooseim@localhost)5> code:get_coverage(cover_id_line, mongoose_tcp_listener).
** exception error: bad argument
     in function  code:get_coverage/2
        called as code:get_coverage(cover_id_line,mongoose_tcp_listener)
        *** argument 1: must be one of: function or line

so, cover crashes when trying to collect data from the remote node.

btw, calling mongoose_tcp_listener coverage on the node that was not ever mecked works fine:

rpc:call('reg1@localhost', code, get_coverage, [cover_id_line, mongoose_tcp_listener]).
[{1,0},
 {2,0},
 {3,0}
 ...

code:get_coverage/2 is called from cover.erl when exporting coverage report.

[Michael Uvarov]

I've added a bug report to Meck eproxus/meck#248

[Michael Uvarov]

sadly, cover error handling is bad.
For example, that crash just crashes a new spawned process, so the caller waits forever for response.
We should add try..catch at least, or cover_server process should monitor their spawned processes.

Using spawn_link instead of spawn in pmap and other places could prevent unreported errors (though would crash cover_server).

[Michael Uvarov]

Basically, this happens if cover is started in the distributed mode, i.e. cover:start([Node1, Node2...]).
Sadly, meck was broken for this mode before, it was not correctly restoring the modules in meck:unload/1.

Starting cover in local_only mode and importing from all nodes manually (i.e. not using cover distributed mode), "solves" the issue.

[Björn Gustavsson]

There does not seem to be a good way to report problem via the current cover API functions, so #8742 avoids the hanging by catching the call, logging that the call failed using logger, and otherwise ignore the failure.

[Michael Uvarov]

Yeah, it looks like the best way for Meck to support multinode mocking and cover, is to make an explicit PR to OTP with missing functionality. Meck is doing mocking for cover in a bit of a hacky way to begin with. Remote cover nodes not exposing any API to load a mocked version just make stuff worse in this particular case.

We managed to start cover in distributed mode on first node-under-test. Good that we only need to mock stuff on that particular node.