samba - Jan 2023 - oplocks, kernel oplocks, kernel share modes, .. - how it all works?

25.01.2023 00:19, Ralph Boehme via samba ?????:
..
>> so setting "kernel oplocks = yes" should prevent
SMB2_CAP_LEASING
>> being returned to the client. A wireshark trace when you remove
>> the "smb2 leases = no" line would be good to look at.
> 
> oh! That is the only place where we treat kernel oplocks as a global
option. The docs and the rest of the code allow this as a per share option.
Looks
> like we have a problem here.
> 
> @Michael: move the kernel oplocks setting to the global section, then
things may start to work as expected.
With kernel oplocks = yes in [global] but without mentioning
smb2 leases, this thing behaves in a way similar to when setting
smb2 leases to no (also globally, as it is a global parameter), --
I see the original NONE/BATCH/LEVEL_II oplocks in there as in the
old good times.

So looks like Ralph is correct here. Effectively, kernel oplocks
turn smb2 leases off, and work on the global level, not on a share
level.

Hmm..

It looks like this needs some more experiments on my side though,
I'm not sure if it actually works and the cache is being invalidated
properly.  Also from my own description of whole thing, it looks
like there's no way to do anything when the file is actually locked.

Why inotify can't be used instead of F_SETLEASE here?

Thanks!

/mjt

On Wed, Jan 25, 2023 at 12:33:57AM +0300, Michael Tokarev
wrote:
>>@Michael: move the kernel oplocks setting to the global section, then
things may start to work as expected.
>
>With kernel oplocks = yes in [global] but without mentioning
>smb2 leases, this thing behaves in a way similar to when setting
>smb2 leases to no (also globally, as it is a global parameter), --
>I see the original NONE/BATCH/LEVEL_II oplocks in there as in the
>old good times.
>
>So looks like Ralph is correct here. Effectively, kernel oplocks
>turn smb2 leases off, and work on the global level, not on a share
>level.
Easiest fix is to just change the docs here :-).
>It looks like this needs some more experiments on my side though,
>I'm not sure if it actually works and the cache is being invalidated
>properly.  Also from my own description of whole thing, it looks
>like there's no way to do anything when the file is actually locked.
Locking and share modes have bugger-all to do with leases and oplocks.

Please separate the two in your mind, things will make
much more sense when you do :-).
>Why inotify can't be used instead of F_SETLEASE here?
Because the local unix process *must* be blocked from
doing *anything* to the file until the SMB client has
flushed all changes back to the file.

inotify just notifies, it doesn't stop any local
client doing anything.