On Mon, Apr 15, 2024 at 11:11:50PM +0800, Ming Lei wrote:
> On Mon, Apr 15, 2024 at 04:53:42PM +0200, Christian Brauner wrote:
> > On Mon, Apr 15, 2024 at 10:35:53PM +0800, Ming Lei wrote:
> > > On Mon, Apr 15, 2024 at 02:35:17PM +0200, Christian Brauner wrote:
> > > > On Mon, Apr 15, 2024 at 05:26:19PM +0800, Ming Lei wrote:
> > > > > Hello,
> > > > >
> > > > > On Tue, Jan 23, 2024 at 02:26:21PM +0100, Christian Brauner wrote:
> > > > > > Signed-off-by: Christian Brauner <brauner@xxxxxxxxxx>
> > > > > > ---
> > > > > > drivers/md/dm.c | 23 +++++++++++++----------
> > > > > > drivers/md/md.c | 12 ++++++------
> > > > > > drivers/md/md.h | 2 +-
> > > > > > include/linux/device-mapper.h | 2 +-
> > > > > > 4 files changed, 21 insertions(+), 18 deletions(-)
> > > > > >
> > > > > > diff --git a/drivers/md/dm.c b/drivers/md/dm.c
> > > > > > index 8dcabf84d866..87de5b5682ad 100644
> > > > > > --- a/drivers/md/dm.c
> > > > > > +++ b/drivers/md/dm.c
> > > > >
> > > > > ...
> > > > >
> > > > > > @@ -775,7 +778,7 @@ static void close_table_device(struct table_device *td, struct mapped_device *md
> > > > > > {
> > > > > > if (md->disk->slave_dir)
> > > > > > bd_unlink_disk_holder(td->dm_dev.bdev, md->disk);
> > > > > > - bdev_release(td->dm_dev.bdev_handle);
> > > > > > + fput(td->dm_dev.bdev_file);
> > > > >
> > > > > The above change caused regression on 'dmsetup remove_all'.
> > > > >
> > > > > blkdev_release() is delayed because of fput(), so dm_lock_for_deletion
> > > > > returns -EBUSY, then this dm disk is skipped in remove_all().
> > > > >
> > > > > Force to mark DMF_DEFERRED_REMOVE might solve it, but need our device
> > > > > mapper guys to check if it is safe.
> > > > >
> > > > > Or other better solution?
> > > >
> > > > Yeah, I think there is. You can just switch all fput() instances in
> > > > device mapper to bdev_fput() which is mainline now. This will yield the
> > > > device and make it able to be reclaimed. Should be as simple as the
> > > > patch below. Could you test this and send a patch based on this (I'm on
> > > > a prolonged vacation so I don't have time right now.):
> > >
> > > Unfortunately it doesn't work.
> > >
> > > Here the problem is that blkdev_release() is delayed, which changes
> > > 'dmsetup remove_all' behavior, and causes that some of dm disks aren't
> > > removed.
> > >
> > > Please see dm_lock_for_deletion() and dm_blk_open()/dm_blk_close().
> >
> > So you really need blkdev_release() itself to be synchronous? Groan, in
>
> At least the current dm implementation relies on this way sort of, and
> it could be addressed by forcing to mark DMF_DEFERRED_REMOVE in
> remove_all().
You floated that earlier in this thread, etc: no, that would change
the interface. DMF_DEFERRED_REMOVE gives people options to allow for
async device closes, etc. But I don't want to impose it as some faux
equivalent to the sync model remove_all has always provided.
And what about simple 'dmsetup remove'? remove_all just loops doing
remove... so isn't 'dmsetup remove' also being forced to be async as
of commit a28d893eb3270 ("md: port block device access to file")?
dm.c:dm_put_device -> dm_put_table_device -> close_table_device
> > that case use __fput_sync() instead of fput() which ensures that this
> > file is closed synchronously.
>
> I tried __fput_sync(), but the following panic is caused:
Ok, so more work needed. But we need to preserve the existing sync
interface for DM device removal.
Mike
