Re: [PATCH v3 09/10] virtiofsd: Optionally fill lo_inode.fhandle

[Vivek Goyal]

On Tue, Aug 10, 2021 at 10:32:55AM +0200, Hanna Reitz wrote:
> On 09.08.21 20:41, Vivek Goyal wrote:
> > On Fri, Jul 30, 2021 at 05:01:33PM +0200, Max Reitz wrote:
> > > When the inode_file_handles option is set, try to generate a file handle
> > > for new inodes instead of opening an O_PATH FD.
> > > 
> > > Being able to open these again will require CAP_DAC_READ_SEARCH, so the
> > > description text tells the user they will also need to specify
> > > -o modcaps=+dac_read_search.
> > > 
> > > Generating a file handle returns the mount ID it is valid for.  Opening
> > > it will require an FD instead.  We have mount_fds to map an ID to an FD.
> > > get_file_handle() fills the hash map by opening the file we have
> > > generated a handle for.  To verify that the resulting FD indeed
> > > represents the handle's mount ID, we use statx().  Therefore, using file
> > > handles requires statx() support.
> > So opening the file and storing that fd in mount_fds table might be
> > a potential problem with inotify work Ioannis is doing.
> > 
> > So say a file foo.txt was opened O_RDONLY and fd stored in mount_fs. Now
> > say user unlinks foo.txt. If notifications are enabled, final notification
> > will not be generated till this mount_fds fd is closed.
> > 
> > Now question is when will this fd be closed? If it closed at some
> > later point and then notification is generated, that will break
> > notificaitons.
> 
> Currently, it is never closed.
> 
> > In fact even O_PATH fd is delaying notifications due to same reason.
> > But its not too bad as we close O_PATH fd pretty quickly after
> > unlinking. And we were hoping that file handle support will get rid
> > of this problem because we will not keep O_PATH fd open.
> > 
> > But, IIUC, mount_fds stuff will make it even worse. I did not see
> > the code which removes this fd from mount_fds. So I am not sure what's
> > the life time of this fd.
> 
> The lifetime is forever.  If we wanted to remove it at some point, we’d need
> to track how many file handles we have open for the given mount fd and then
> remove it from the table once the count reaches 0, so it would still be
> delayed.
> 
> I think in practice the first thing that is looked up from some mount will
> probably be the root directory, which cannot be deleted before everything
> else on the mount is gone, so that would work.  We track how many handles
> are there, if the whole mount were to be deleted, I hope all lo_inodes are
> evicted, the count goes to 0, and we can drop the mount fd.

Keeping a reference count on mount_fd object make sense. So we probably
maintain this hash table and lookup using mount_id (as you are already
doing). All subsequent inodes from same filesystem will use same
object. Once all inodes have been flushed out, then mount_fd object
should go away as well (allowing for unmount on host).

> 
> I think we can make the assumption that the mount fd is the root directory
> certain by, well, looking into mountinfo...  That would result in us always
> opening the root node of the filesystem, so that first the whole filesystem
> needs to disappear before it can be deleted (and our mount fd closed) –
> which should work, I guess?

This seems more reasonable. And I think that's what man page seems to 
suggest.

       The  mount_id  argument  returns an identifier for the filesystem mount
       that corresponds to pathname.  This corresponds to the first  field  in
       one  of  the  records in /proc/self/mountinfo.  Opening the pathname in
       the fifth field of that record yields a file descriptor for  the  mount
       point;  that  file  descriptor  can  be  used  in  a subsequent call to
       open_by_handle_at().

Fifth field seems to be the mount point. man proc says.

              (5)  mount  point:  the  pathname of the mount point relative to
                   the process's root directory.

So opening mount point and saving as mount_fd (if it is not already
in hash table) and then take a per inode reference count on mount_fd
object looks like will solve the life time issue of mount_fd as
well as the issue of temporary failures arising because we can't
open a device special file.

> 
> It’s a bit tricky because our sandboxing prevents easy access to mountinfo,
> but if that’s the only way...

yes. We already have lo->proc_self_fd. Maybe we need to keep
/proc/self/mountinfo open in lo->proc_self_mountinfo. I am assuming
that any mount table changes will still be visible despite the fact
I have fd open (and don't have to open new fd to notice new mount/unmount
changes).

Vivek