[RFC PATCH] fork: report pid reservation failure properly

[Michal Hocko]

Hi,
while debugging an unexpected ENOMEM from fork (there was no memory
pressure and OVERCOMMIT_ALWAYS) I have found out that fork returns
ENOMEM even when not short on memory.

In this particular case it was due to depleted pid space which is
documented to return EAGAIN in man pages.

Here is a quick fix up. I am sending that as a RFC because I am not
entirely sure about interaction with pid namespace which might cause
fork to fail. pid_ns_prepare_proc might return few error codes which are
not documented by man page but I am not sure whether that is actually
going to happen on a properly configured system.

There doesn't seem to be a good error code for an already "closed"
namespace (PIDNS_HASH_ADDING disabled) as well.  I've chosen EBUSY but
that might be completely wrong.  I am also wondering how would this
error case happen in the first place because parent should still exist
while fork is going on so the pid namespace shouldn't go away but I
smell this might have something to do with setns or something similar.

Any feedback would be appreciated.
---