Re: [RFC PATCH v6] mm, pmem, xfs: Introduce MF_MEM_REMOVE for unbind

[Dan Williams]

Darrick J. Wong wrote:
> On Thu, Jul 14, 2022 at 11:21:44AM -0700, Dan Williams wrote:
> > ruansy.fnst@xxxxxxxxxxx wrote:
> > > This patch is inspired by Dan's "mm, dax, pmem: Introduce
> > > dev_pagemap_failure()"[1].  With the help of dax_holder and
> > > ->notify_failure() mechanism, the pmem driver is able to ask filesystem
> > > (or mapped device) on it to unmap all files in use and notify processes
> > > who are using those files.
> > > 
> > > Call trace:
> > > trigger unbind
> > >  -> unbind_store()
> > >   -> ... (skip)
> > >    -> devres_release_all()   # was pmem driver ->remove() in v1
> > >     -> kill_dax()
> > >      -> dax_holder_notify_failure(dax_dev, 0, U64_MAX, MF_MEM_PRE_REMOVE)
> > >       -> xfs_dax_notify_failure()
> > > 
> > > Introduce MF_MEM_PRE_REMOVE to let filesystem know this is a remove
> > > event.  So do not shutdown filesystem directly if something not
> > > supported, or if failure range includes metadata area.  Make sure all
> > > files and processes are handled correctly.
> > > 
> > > ==
> > > Changes since v5:
> > >   1. Renamed MF_MEM_REMOVE to MF_MEM_PRE_REMOVE
> > >   2. hold s_umount before sync_filesystem()
> > >   3. move sync_filesystem() after SB_BORN check
> > >   4. Rebased on next-20220714
> > > 
> > > Changes since v4:
> > >   1. sync_filesystem() at the beginning when MF_MEM_REMOVE
> > >   2. Rebased on next-20220706
> > > 
> > > [1]: https://lore.kernel.org/linux-mm/161604050314.1463742.14151665140035795571.stgit@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
> > > 
> > > Signed-off-by: Shiyang Ruan <ruansy.fnst@xxxxxxxxxxx>
> > > ---
> > >  drivers/dax/super.c         |  3 ++-
> > >  fs/xfs/xfs_notify_failure.c | 15 +++++++++++++++
> > >  include/linux/mm.h          |  1 +
> > >  3 files changed, 18 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/dax/super.c b/drivers/dax/super.c
> > > index 9b5e2a5eb0ae..cf9a64563fbe 100644
> > > --- a/drivers/dax/super.c
> > > +++ b/drivers/dax/super.c
> > > @@ -323,7 +323,8 @@ void kill_dax(struct dax_device *dax_dev)
> > >  		return;
> > >  
> > >  	if (dax_dev->holder_data != NULL)
> > > -		dax_holder_notify_failure(dax_dev, 0, U64_MAX, 0);
> > > +		dax_holder_notify_failure(dax_dev, 0, U64_MAX,
> > > +				MF_MEM_PRE_REMOVE);
> > >  
> > >  	clear_bit(DAXDEV_ALIVE, &dax_dev->flags);
> > >  	synchronize_srcu(&dax_srcu);
> > > diff --git a/fs/xfs/xfs_notify_failure.c b/fs/xfs/xfs_notify_failure.c
> > > index 69d9c83ea4b2..6da6747435eb 100644
> > > --- a/fs/xfs/xfs_notify_failure.c
> > > +++ b/fs/xfs/xfs_notify_failure.c
> > > @@ -76,6 +76,9 @@ xfs_dax_failure_fn(
> > >  
> > >  	if (XFS_RMAP_NON_INODE_OWNER(rec->rm_owner) ||
> > >  	    (rec->rm_flags & (XFS_RMAP_ATTR_FORK | XFS_RMAP_BMBT_BLOCK))) {
> > > +		/* Do not shutdown so early when device is to be removed */
> > > +		if (notify->mf_flags & MF_MEM_PRE_REMOVE)
> > > +			return 0;
> > >  		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_ONDISK);
> > >  		return -EFSCORRUPTED;
> > >  	}
> > > @@ -174,12 +177,22 @@ xfs_dax_notify_failure(
> > >  	struct xfs_mount	*mp = dax_holder(dax_dev);
> > >  	u64			ddev_start;
> > >  	u64			ddev_end;
> > > +	int			error;
> > >  
> > >  	if (!(mp->m_sb.sb_flags & SB_BORN)) {
> > >  		xfs_warn(mp, "filesystem is not ready for notify_failure()!");
> > >  		return -EIO;
> > >  	}
> > >  
> > > +	if (mf_flags & MF_MEM_PRE_REMOVE) {
> > > +		xfs_info(mp, "device is about to be removed!");
> > > +		down_write(&mp->m_super->s_umount);
> > > +		error = sync_filesystem(mp->m_super);
> > > +		up_write(&mp->m_super->s_umount);
> > 
> > Are all mappings invalidated after this point?
> 
> No; all this step does is pushes dirty filesystem [meta]data to pmem
> before we lose DAXDEV_ALIVE...
> 
> > The goal of the removal notification is to invalidate all DAX mappings
> > that are no pointing to pfns that do not exist anymore, so just syncing
> > does not seem like enough, and the shutdown is skipped above. What am I
> > missing?
> 
> ...however, the shutdown above only applies to filesystem metadata.  In
> effect, we avoid the fs shutdown in MF_MEM_PRE_REMOVE mode, which
> enables the mf_dax_kill_procs calls to proceed against mapped file data.
> I have a nagging suspicion that in non-PREREMOVE mode, we can end up
> shutting down the filesytem on an xattr block and the 'return
> -EFSCORRUPTED' actually prevents us from reaching all the remaining file
> data mappings.
> 
> IOWs, I think that clause above really ought to have returned zero so
> that we keep the filesystem up while we're tearing down mappings, and
> only call xfs_force_shutdown() after we've had a chance to let
> xfs_dax_notify_ddev_failure() tear down all the mappings.
> 
> I missed that subtlety in the initial ~30 rounds of review, but I figure
> at this point let's just land it in 5.20 and clean up that quirk for
> -rc1.

Sure, this is a good baseline to incrementally improve.

> 
> > Notice that kill_dev_dax() does unmap_mapping_range() after invalidating
> > the dax device and that ensures that all existing mappings are gone and
> > cannot be re-established. As far as I can see a process with an existing
> > dax mapping will still be able to use it after this runs, no?
> 
> I'm not sure where in akpm's tree I find kill_dev_dax()?  I'm cribbing
> off of:
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm.git/tree/fs/xfs/xfs_notify_failure.c?h=mm-stable

https://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm.git/tree/drivers/dax/bus.c?h=mm-stable#n381

Where the observation is that when device-dax is told that the device is
going away it invalidates all the active mappings to that single
character-device-inode. The hope being that in the fsdax case all the
dax-mapped filesystem inodes would experience the same irreversible
invalidation as the device is exiting.