On Fri 11-08-23 01:39:10, Ryusuke Konishi wrote:
> On Thu, Aug 10, 2023 at 8:05 PM Jan Kara wrote:
> >
> > On Fri 04-08-23 11:01:39, Ryusuke Konishi wrote:
> > > On Thu, Aug 3, 2023 at 8:46 PM Jan Kara wrote:
> > > >
> > > > On Wed 02-08-23 17:41:21, Christoph Hellwig wrote:
> > > > > Use the generic setup_bdev_super helper to open the main block device
> > > > > and do various bits of superblock setup instead of duplicating the
> > > > > logic. This includes moving to the new scheme implemented in common
> > > > > code that only opens the block device after the superblock has allocated.
> > > > >
> > > > > It does not yet convert nilfs2 to the new mount API, but doing so will
> > > > > become a bit simpler after this first step.
> > > > >
> > > > > Signed-off-by: Christoph Hellwig <hch@xxxxxx>
> > > >
> > > > AFAICS nilfs2 could *almost* use mount_bdev() directly and then just do its
> > >
> > > > snapshot thing after mount_bdev() returns. But it has this weird logic
> > > > that: "if the superblock is already mounted but we can shrink the whole
> > > > dcache, then do remount instead of ignoring mount options". Firstly, this
> > > > looks racy - what prevents someone from say opening a file on the sb just
> > > > after nilfs_tree_is_busy() shrinks dcache? Secondly, it is inconsistent
> > > > with any other filesystem so it's going to surprise sysadmins not
> > > > intimately knowing nilfs2. Thirdly, from userspace you cannot tell what
> > > > your mount call is going to do. Last but not least, what is it really good
> > > > for? Ryusuke, can you explain please?
> > > >
> > > > Honza
> > >
> > > I think you are referring to the following part:
> > >
> > > > if (!s->s_root) {
> > > ...
> > > > } else if (!sd.cno) {
> > > > if (nilfs_tree_is_busy(s->s_root)) {
> > > > if ((flags ^ s->s_flags) & SB_RDONLY) {
> > > > nilfs_err(s,
> > > > "the device already has a %s mount.",
> > > > sb_rdonly(s) ? "read-only" : "read/write");
> > > > err = -EBUSY;
> > > > goto failed_super;
> > > > }
> > > > } else {
> > > > /*
> > > > * Try remount to setup mount states if the current
> > > > * tree is not mounted and only snapshots use this sb.
> > > > */
> > > > err = nilfs_remount(s, &flags, data);
> > > > if (err)
> > > > goto failed_super;
> > > > }
> > > > }
> > >
> > > What this logic is trying to do is, if there is already a nilfs2 mount
> > > instance for the device, and are trying to mounting the current tree
> > > (sd.cno is 0, so this is not a snapshot mount), then will switch
> > > depending on whether the current tree has a mount:
> > >
> > > - If the current tree is mounted, it's just like a normal filesystem.
> > > (A read-only mount and a read/write mount can't coexist, so check
> > > that, and reuse the instance if possible)
> > > - Otherwise, i.e. for snapshot mounts only, do whatever is necessary
> > > to add a new current mount, such as starting a log writer.
> > > Since it does the same thing that nilfs_remount does, so
> > > nilfs_remount() is used there.
> > >
> > > Whether or not there is a current tree mount can be determined by
> > > d_count(s->s_root) > 1 as nilfs_tree_is_busy() does.
> > > Where s->s_root is always the root dentry of the current tree, not
> > > that of the mounted snapshot.
> >
> > I see now, thanks for explanation! But one thing still is not clear to me.
> > If you say have a snapshot mounted read-write and then you mount the
> > current snapshot (cno == 0) read-only, you'll switch the whole superblock
> > to read-only state. So also the mounted snapshot is suddently read-only
> > which is unexpected and actually supposedly breaks things because you can
> > still have file handles open for writing on the snapshot etc.. So how do
> > you solve that?
> >
> > Honza
>
> One thing I have to tell you as a premise is that nilfs2's snapshot
> mounts (cno != 0) are read-only.
>
> The read-only option is mandatory for nilfs2 snapshot mounts, so
> remounting to read/write mode will result in an error.
> This constraint is checked in nilfs_parse_snapshot_option() which is
> called from nilfs_identify().
>
> In fact, any write mode file/inode operations on a snapshot mount will
> result in an EROFS error, regardless of whether the coexisting current
> tree mount is read-only or read/write (i.e. regardless of the
> read-only flag of the superblock instance).
>
> This is mostly (and possibly entirely) accomplished at the vfs layer
> by checking the MNT_READONLY flag in mnt_flags of the vfsmount
> structure, and even on the nilfs2 side, iops->permission
> (=nilfs_permission) rejects write operations on snapshot mounts.
>
> Therefore, the problem you pointed out shouldn't occur in the first
> place since the situation where a snapshot with a handle in write mode
> suddenly becomes read-only doesn't happen. Unless I'm missing
> something..
No, I think you are correct. This particular case should be safe because
MNT_READONLY flags on the mounts used by snapshots will still keep them
read-only even if you remount the superblock to read-write mode for the
current snapshot. So I see why this is useful and I agree this isn't easy
to implement using mount_bdev() so no special code reduction here ;).
Thanks for patient explanation!
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
