RFC: ksmbd ABI for ksmbd-tools...

Hi Namjae,

I looked through the interfaces used between userspace (ksmbd.mountd and ksmbd.control)
and the kernel module.

After loading the ksmbd.ko module and calling 'ksmbd.mountd', I see
the following related proceses/kernel-threads:

  12200 ?        I      0:00 [kworker/0:0-ksmbd-io]
  12247 ?        Ss     0:00 ksmbd.mountd
  12248 ?        S      0:00 ksmbd.mountd
  12249 ?        S      0:00 [ksmbd-lo]
  12250 ?        S      0:00 [ksmbd-enp0s3]
  12251 ?        S      0:00 [ksmbd-enp0s8]
  12252 ?        S      0:00 [ksmbd-enp0s9]
  12253 ?        S      0:00 [ksmbd-enp0s10]
  12254 ?        I<     0:00 [ksmbd-smb_direc]
  12255 ?        S      0:00 [ksmbd:38794]
  12257 ?        S      0:00 [ksmbd:51579]

I haven't found the exact place, but ksmbd.mountd starts the kernel-part.

ksmbd.mountd also acts as some kind of upcall, for the server part,
that takes care of authentication and some basic DCERPC calls.

I'm wondering why there are two separate ways to kill the running server,
'killall ksmbd.mountd' for the userspace part and
'ksmbd.control -s' (which is just a wrapper for
'echo -n "hard" > /sys/class/ksmbd-control/kill_server') to shutdown the server part.

As it's not useful to run any of these two components on its own,
so I'm wondering why there's no stronger relationship.

As naive admin I'd assume that the kernel part would detect the exit of ksmbd.mountd
and shutdown itself.

It would also be great to bind to specific ip addresses instead of devices
and allow to run more than one instance of ksmbd.mountd (with different config files
and or within containers). That's why I think single global hardcoded path like
'/sys/class/ksmbd-control/kill_server' should be avoided, something like:

26 more lines

Hi,

On (21/02/12 15:38), Stefan Metzmacher via Linux-cifsd-devel wrote:

I haven't found the exact place, but ksmbd.mountd starts the kernel-part.

ksmbd.mountd also acts as some kind of upcall, for the server part,
that takes care of authentication and some basic DCERPC calls.

I'm wondering why there are two separate ways to kill the running server,
'killall ksmbd.mountd' for the userspace part and
'ksmbd.control -s' (which is just a wrapper for
'echo -n "hard" > /sys/class/ksmbd-control/kill_server') to shutdown the server part.

As it's not useful to run any of these two components on its own,
so I'm wondering why there's no stronger relationship.

As naive admin I'd assume that the kernel part would detect the exit of ksmbd.mountd
and shutdown itself.

User-space daemon is just some sort of a database engine, kernel module
queries it when it needs something; otherwise kernel module works without
user-space part just fine and doesn't need it. The goal is that when
user-space crashes or gets restarted after update (critical fix) the server
doesn't panic and doesn't kill itself.

E.g. - when you restart mysql you don't expect httpd to kill itself and
to terminate all existing TCP connections.

[..]

Can we require that the userspace tool matches the kernel version for a while?

I think such a check exists (or at least it used to). Note that the only
time when version mismatch matters is when sizeof() or layout of the
structures that are used for RPC get updated, or new request commands or
status codes added. IOW, RPC version mismatch is the critical thing.
Otherwise there is no real reason (case by case) to forbid version
mismatch. If user-space gets a memory leak or a NULL pointer dereference
fix then there is no real reasons to force server restart or to force
server module version update, because the fix is completely internal to
the user-space daemon.

	-ss