Re: [RFC PATCH v2 01/18] vfs: export new_inode_pseudo

[Eric Biggers <ebiggers@kernel.org>]

On Wed, Sep 09, 2020 at 06:47:28AM -0400, Jeff Layton wrote:
> On Tue, 2020-09-08 at 15:31 -0700, Eric Biggers wrote:
> > On Tue, Sep 08, 2020 at 07:27:58AM -0400, Jeff Layton wrote:
> > > On Mon, 2020-09-07 at 20:38 -0700, Eric Biggers wrote:
> > > > On Fri, Sep 04, 2020 at 12:05:20PM -0400, Jeff Layton wrote:
> > > > > Ceph needs to be able to allocate inodes ahead of a create that might
> > > > > involve a fscrypt-encrypted inode. new_inode() almost fits the bill,
> > > > > but it puts the inode on the sb->s_inodes list, and we don't want to
> > > > > do that until we're ready to insert it into the hash.
> > > > > 
> > > > > Signed-off-by: Jeff Layton <jlayton@kernel.org>
> > > > > ---
> > > > >  fs/inode.c | 1 +
> > > > >  1 file changed, 1 insertion(+)
> > > > > 
> > > > > diff --git a/fs/inode.c b/fs/inode.c
> > > > > index 72c4c347afb7..61554c2477ab 100644
> > > > > --- a/fs/inode.c
> > > > > +++ b/fs/inode.c
> > > > > @@ -935,6 +935,7 @@ struct inode *new_inode_pseudo(struct super_block *sb)
> > > > >  	}
> > > > >  	return inode;
> > > > >  }
> > > > > +EXPORT_SYMBOL(new_inode_pseudo);
> > > > >  
> > > > 
> > > > What's the problem with putting the new inode on sb->s_inodes already?
> > > > That's what all the other filesystems do.
> > > > 
> > > 
> > > The existing ones are all local filesystems that use
> > > insert_inode_locked() and similar paths. Ceph needs to use the '5'
> > > variants of those functions (inode_insert5(), iget5_locked(), etc.).
> > > 
> > > When we go to insert it into the hash in inode_insert5(), we'd need to
> > > set I_CREATING if allocated from new_inode(). But, if you do _that_,
> > > then you'll get back ESTALE from find_inode() if (eg.) someone calls
> > > iget5_locked() before you can clear I_CREATING.
> > > 
> > > Hitting that race is easy with an asynchronous create. The simplest
> > > scheme to avoid that is to just export new_inode_pseudo and keep it off
> > > of s_inodes until we're ready to do the insert. The only real issue here
> > > is that this inode won't be findable by evict_inodes during umount, but
> > > that shouldn't be happening during an active syscall anyway.
> > 
> > Is your concern the following scenario?
> > 
> > 1. ceph successfully created a new file on the server
> > 2. inode_insert5() is called for the new file's inode
> > 3. error occurs in ceph_fill_inode()
> > 4. discard_new_inode() is called
> > 5. another thread looks up the inode and gets ESTALE
> > 6. iput() is finally called
> > 
> > And the claim is that the ESTALE in (5) is unexpected?  I'm not sure that it's
> > unexpected, given that the file allegedly failed to be created...  Also, it
> > seems that maybe (3) isn't something that should actually happen, since after
> > (1) it's too late to fail the file creation.
> > 
> 
> No, more like:
> 
> Syscall					Workqueue
> ------------------------------------------------------------------------------
> 1. allocate an inode
> 2. determine we can do an async create
>    and allocate an inode number for it
> 3. hash the inode (must set I_CREATING
>    if we allocated with new_inode()) 
> 4. issue the call to the MDS
> 5. finish filling out the inode()
> 6.					MDS reply comes in, and workqueue thread
> 					looks up new inode (-ESTALE)
> 7. unlock_new_inode()
> 
> 
> Because 6 happens before 7 in this case, we get an ESTALE on that
> lookup.

How is ESTALE at (6) possible?  (3) will set I_NEW on the inode when inserting
it into the inode hash table.  Then (6) will wait for I_NEW to be cleared before
returning the inode.  (7) will clear I_NEW and I_CREATING together.

- Eric