Andrew Schwartzmeyer created MESOS-8563:
-------------------------------------------
Summary: Windows executors cannot re-register
Key: MESOS-8563
URL: https://issues.apache.org/jira/browse/MESOS-8563
Project: Mesos
Issue Type: Bug
Components: agent, executor, libprocess
Environment: Windows 10
Reporter: Andrew Schwartzmeyer
Assignee: Andrew Schwartzmeyer
This issue captures an important (but already resolved) bug due to incorrect
inheritance of sockets.
When enabling agent recovery, it was discovered that the executors could not
re-register to the new agent. They would send the re-register message, and then
fail silently. The agent never received the re-register message.
This turned out to be due to incorrect inheritance semantics of sockets. On
POSIX systems, {{os::cloexec}} was used to prevent file descriptors (or socket
handles, on Windows) from being inherited by child processes. On Windows, we
were creating {{SOCKET}} handles using the CRT API {{::socket}}, which by
default created _inheritable_ socket handles. The subsequent call to
{{os::cloexec}} to prevent this was a no-op, leaving us leaking socket handles
to all child processes, causing the described bug.
The solution was to split {{net::socket}} into a POSIX and Windows
implementation, where on Windows we use the WinSock 2 API {{WSASocket}}, which
allows us to create the socket upfront with {{WSA_FLAG_NO_HANDLE_INHERIT}},
preventing the leaks. This is somewhat like using {{O_CLOEXEC}} on Linux.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
