On Thu 29-08-24 13:26:17, Amir Goldstein wrote: > On Thu, Aug 29, 2024 at 1:15 PM Jan Kara <jack@xxxxxxx> wrote: > > > > On Wed 14-08-24 17:25:33, Josef Bacik wrote: > > > gfs2 takes the glock before calling into filemap fault, so add the > > > fsnotify hook for ->fault before we take the glock in order to avoid any > > > possible deadlock with the HSM. > > > > > > Signed-off-by: Josef Bacik <josef@xxxxxxxxxxxxxx> > > > > The idea of interactions between GFS2 cluster locking and HSM gives me > > creeps. But yes, this patch looks good to me. Would be nice to get ack from > > GFS2 guys. Andreas? > > If we are being honest, I think that the fact that HSM events require careful > handling in ->fault() and not to mention no documentation of this fact, > perhaps we should let HSM events be an opt-in file_system_type feature? > > Additionally, we had to introduce FS_DISALLOW_NOTIFY_PERM > and restrict sb marks on SB_NOUSER, all because these fanotify > features did not require fs opt-in to begin with. > > I think we would be repeating this mistake if we do not add > FS_ALLOW_HSM from the start. > > After all, I cannot imagine HSM being used on anything but > the major disk filesystems. > > Hmm? Yeah, I was considering this already when thinking about btrfs quirks with readahead and various special filesystem ioctls and I agree that a need to be careful with page faults is another good reason to make this a per-filesystem opt in. Will you send a patch? Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
