Re: Questions about nfs_sb_active

[Trond Myklebust <trondmy@hammerspace.com>]

On Thu, 2021-09-16 at 11:48 +0800, zhangxiaoxu (A) wrote:
> 
> 
> 在 2021/9/15 21:05, Trond Myklebust 写道:
> > 
> > 
> > > On Sep 15, 2021, at 04:03, zhangxiaoxu (A)
> > > <zhangxiaoxu5@huawei.com> wrote:
> > > 
> > > Hi Trond,
> > > 
> > > I have some confuse about 'nfs_sb_active'.
> > > 
> > > The following commit increase the 'sb->s_active' to prevent
> > > concurrent with umount process when handle the callback rpc
> > > message.
> > > 
> > >   e39d8a186ed0 ("NFSv4: Fix an Oops during delegation callbacks")
> > >   113aac6d567b ("NFS: nfs_delegation_find_inode_server must first
> > > reference the superblock")
> > > 
> > > But it also delay the process in function
> > > 'generic_shutdown_super', such as 'sync_filesystem' and
> > > 'fsnotify_sb_delete'.
> > > 
> > > For the common file system, when umount success, the data should
> > > be stable to the disk, but in nfs, it maybe delay?
> > > 
> > > I want know :
> > >   1. whether we _must_ stable the data to the server?
> > >   2. how to ensure the data not lost when umount success but
> > > client crash?
> > >   3. the delayed fsnotify umount event is reasonable or not?
> > >   4. the 'nfs_sb_active' should be used under what scenario?
> > > 
> > > Thanks.
> > 
> > That has nothing to do with I/O. Delegations are state.
> Since the callbacks hold the 'sb->s_active',
> the umount maybe return success without shutdown the superblock.
> 
> In general, the superblock should be shutdown before umount success,
> but in the concurrent scenario, the superblock is shutdown after the
> callbacks finish.
> 
> If the system is crashed in this period, we may lost
> 'sync_filesystem',
> then the page caches (which not flush to server since hold the write
> delegation when close the file)
> and metadata caches maybe lost?

No. We still flush writes on close. Even if that were the case, then it
is no different from the behaviour of block devices.

> 
> And the 'fsnotify_sb_delete' is also called after the callbacks
> finish.
> IOW, the umount already return with success, but the FS_UNMOUNT event
> maybe delay?
> 
> I have no idea about it is reasonable or not.
> > 

I do.

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com