cifs.upcall: try to use container ipc/uts/net/pid/mnt/user namespaces
In certain scenarios (e.g. kerberos multimount), when a process does
syscalls, the kernel sometimes has to query information or trigger
some actions in userspace. To do so it calls the cifs.upcall binary
with information on the process that triggered the syscall in the
first place.
ls(pid=10) ====> open("foo") ====> kernel
that user doesn't have an SMB
session, lets create one using his
kerberos credential cache
call cifs.upcall and ask for krb info
for whoever owns pid=10
|
cifs.upcall --pid 10 <=================+
...gather info...
return binary blob used
when establishing SMB session
===================> kernel
open SMB session, handle
open() syscall
ls <=================================== return open() result to ls
On a system using containers, the kernel is still calling the host
cifs.upcall and using the host configuration (for network, pid, etc).
This patch changes the behaviour of cifs.upcall so that it uses the
calling process namespaces (ls in the example) when doing its
job.
Note that the kernel still calls the binary in the host, but the
binary will place itself the contexts of the calling process
namespaces.
This code makes use of (but shouldn't require) the following kernel
config options and syscall flags:
approx. year |
introduced | config/flags
---------------+----------------
2008 | CONFIG_NAMESPACES=y
2007 | CONFIG_UTS_NS=y
2020 | CONFIG_TIME_NS=y
2006 | CONFIG_IPC_NS=y
2007 | CONFIG_USER_NS
2008 | CONFIG_PID_NS=y
2007 | CONFIG_NET_NS=y
2007 | CONFIG_CGROUPS
2016 | CLONE_NEWCGROUP setns() flag
Signed-off-by: Aurelien Aptel <aaptel@suse.com>
Signed-off-by: Alastair Houghton <alastair@alastairs-place.net>