Re: [patch 4/8] fs, exportfs: Add export_encode_inode_fh helper

[Pavel Emelyanov <xemul@parallels.com>]

On 08/16/2012 06:03 PM, James Bottomley wrote:
> On Thu, 2012-08-16 at 17:54 +0400, Cyrill Gorcunov wrote:
>> On Thu, Aug 16, 2012 at 02:50:19PM +0100, Al Viro wrote:
>>> On Thu, Aug 16, 2012 at 05:47:06PM +0400, Pavel Emelyanov wrote:
>>>> On 08/16/2012 05:43 PM, Al Viro wrote:
>>>>> On Thu, Aug 16, 2012 at 04:38:14PM +0400, Cyrill Gorcunov wrote:
>>>>>
>>>>>> Hi Bruce, thinking a bit more I guess using general encode_fh is not that
>>>>>> convenient since it operates with dentries while our fdinfo output deals
>>>>>> with inodes. Thus I should either provide some new encode_fh variant
>>>>>> which would deal with inodes directly without "parents". Which doesn't
>>>>>> look for me anyhow better than the new export_encode_inode_fh helper.
>>>>>
>>>>> Huh?  You do have dentries, for crying out loud...
>>>>
>>>> Sometimes we don't -- the inotify thing gets an inode only.
>>>> Unlike other notifies that have dentries at hands...
>>>
>>> What's wrong with saying "we don't support idiotify"?
>>
>> Al, we need some way to restore inotifies after checkpoint.
>> At the very early versions of these patches I simply added
>> dentry to the inotify mark thus once inotify created we always
>> have a dentry to refer on in encode_fh, but I'm not sure if
>> this will be good design.
> 
> Actually, I was about to suggest this.  This can be done internally
> within fs/notify without actually modifying the syscall interface, can't
> it, since they take a path which is used to obtain the inode?  It looks
> like the whole of the inotify interface could be internally recast to
> use dentries instead of inodes.  Unless I've missed something obvious?

This will change the observable by userspace behavior. Various apps inotify a
file, then rename/unlink/link it or do tricks with mounts container the file.
And it works one way if the dentry+mount reference is 0 (now) and some other
way if it's not (after the proposed change).

The dentries-related behavior is especially bad on NFS with its silly-renames
decisions based on dentry reference counters. The mount-related one is bad in
general.

I'm saying this, because we were facing such problems at approx. once-a-week
rate when we did this in OpenVZ :(

> James
> 

Thanks,
Pavel