Re: kernel oops on bcachefs umount, 6.7 kernel

[Brian Foster <bfoster@redhat.com>]

On Fri, Feb 16, 2024 at 10:57:24AM -0500, Kent Overstreet wrote:
> On Fri, Feb 16, 2024 at 07:40:34AM -0500, Brian Foster wrote:
> > On Fri, Feb 16, 2024 at 09:00:17AM +0100, Christoph Hellwig wrote:
> > > On Thu, Feb 15, 2024 at 07:24:23PM -0500, Kent Overstreet wrote:
> > > > > It looks like the warning could be avoided in bcachefs by checking for
> > > > > whether the parent dir/node still exists at cleanup time, but I'm not
> > > > > familiar enough with kobj management to say whether that's the
> > > > > right/best solution. It also looks a little odd to me to see a
> > > > > /sys/block/<dev>/bcachefs dir when I've not seen any other fs or driver
> > > > > do such a thing in the block sysfs dir(s).
> > > > > 
> > > > > Any thoughts on this from the block subsystem folks? Is it reasonable to
> > > > > leave this link around and just fix the removal check, or is another
> > > > > behavior preferred? Thanks.
> > > 
> > > This is the general problem with random cross-subsystem sysfs reference,
> > > and why they are best avoided.  The block layer tears down all the sysfs
> > > objects at del_gendisk time as no one should start using the sysfs files
> > > at that point, but a mounted file system or other opener will of course
> > > keep the bdev itself alive.
> > > 
> > 
> > Yeah, makes sense. The fact that the dir goes away despite having the
> > bdev open is partly what made this seem a little odd to me.
> > 
> > > I'm not sure why bcachefs is doing this, but no one really should be
> > > using the block layer sysfs structures and pointers except for the block
> > > layer itself.  
> > > 
> > 
> > From Kent's comments it sounds like it was just some loose carryover
> > from old bcache stuff. I had poked around a bit for anything similar and
> > it looked to me that current bcache doesn't do this either, but I could
> > have missed something.
> > 
> > > > so there's an existing bd_holder mechanism that e.g. device mapper uses
> > > > for links between block devices. I think the "this block device is going
> > > > away" code knows how to clean those up.
> > > > 
> > > > We're not using that mechanism - and perhaps we should have been, I'd
> > > > need a time machine to ask myself why I did it that way 15 years back.
> > > 
> > > Well, at least Tejun had a very strong opinion that no one should be
> > > abusing sysfs symlinks for linking up subsystems at all, see commit
> > > 49731baa41df404c2c3f44555869ab387363af43, which is also why this code
> > > is marked deprecated and we've not added additional users.
> > > 
> > 
> > Thanks. I'll send a patch to remove this once I'm back from vacation.
> 
> No, we can't remove it - userspace needs to know this topology. When
> we've got one sysfs node with a direct relationship to another sysfs
> node, that needs to be reflected in syfs.
> 

Do you mean that some related userspace tool relies on this to function,
or generally disagreeing with the statement(s) above around links from
/sys/block/<dev>/?

Brian

> What I'm hoping for is that we can get filesystems in general to do this
> right, not just bcachefs, so that they finally start showing up in the
> tree lsblk reports.
>