Re: [PATCH 2/3] nbd: notify userland even if nbd has already disconnected

[Yao Liu]

On Tue, May 28, 2019 at 11:36:21AM -0500, Mike Christie wrote:
> On 05/27/2019 01:23 PM, Yao Liu wrote:
> > On Fri, May 24, 2019 at 09:08:58AM -0400, Josef Bacik wrote:
> >> On Fri, May 24, 2019 at 05:43:55PM +0800, Yao Liu wrote:
> >>> Some nbd client implementations have a userland's daemon, so we should
> >>> inform client daemon to clean up and exit.
> >>>
> >>> Signed-off-by: Yao Liu <yotta.liu@xxxxxxxxx>
> >>
> >> Except the nbd_disconnected() check is for the case that the client told us
> >> specifically to disconnect, so we don't want to send the notification to
> >> re-connect because we've already been told we want to tear everything down.
> >> Nack to this as well.  Thanks,
> >>
> >> Josef
> >>
> > 
> > But in userland, client daemon process and process which send disconnect
> > command are not same process, so they are not clear to each other, so
> > client daemon expect driver inform it to exit.
> > In addition, client daemon will get nbd status with nbd_genl_status interface
> > after it get notified and it should not re-connect if status connected == 0
> > 
> 
> When using the netlink interface you get the NBD_CMD_LINK_DEAD first
> then the configs_refs goes to zero right?
> 
> nbd_disconnect_and_put -> sock_shutdown -> nbd_mark_nsock_dead
> 
> then later we do the final nbd_config_put?
> 
> Maybe it would be best to add a new netlink event to signal what has
> happened, because the above nl and stat algorithm seems like a pain. The
> NBD_CMD_LINK_DEAD will be sent, then userspace has to possibly poll the
> status to check if this was caused due to nbd_genl_disconnect instead of
> a downed link due to something like a command timeout, because the
> refcount may not be down when userspace gets the NL event.
> 
> Or, I guess the admin/tool process could just send a msg to the daemon
> process to tell it to do the netlink disconnect request.
> 

Adding a new netlink event sames good.