Re: Question on fscrypt_d_revalidate() and fstest generic/429

[Eric Biggers]

On Mon, May 15, 2017 at 09:51:03PM +0200, Richard Weinberger wrote:
> > 
> > The test is repeatedly creating and removing a directory "dir" while lookups are
> > being done in it.  It seems the problem is that many dentries are being created
> > for "dir", and they pin many different inodes, all at the same time.  This
> > actually happens for ext4 too; it just doesn't cause an observable error.
> > 
> > I doubt it's the right solution to make fscrypt_d_revalidate() look at
> > ->i_nlink, since ->d_revalidate() is meant to validate the filename, not the
> > inode.  I think there is probably a VFS bug that is causing the dentries to not
> > be freed.
> 
> Not sure. Al? :-)
> 

I can reproduce this on an unencrypted directory after updating path_init() in
fs/namei.c to always clear LOOKUP_RCU, so that all path lookups are done in
ref-walk mode.  So I think fscrypt_d_revalidate() was only relevant because it
causes all path lookups to drop out of rcu-walk mode.

It seems that what's happening is the "dir" dentries are not being freed because
each one has a child dentry "file" that is a negative dentry.  The "file" dentry
would normally be freed by shrink_dcache_parent() called from vfs_rmdir(), but
due to a race with stat("dir/file") the "file" dentry sometimes has nonzero
reference count at that time, causing it to remain in the subdirs list.  So we
end up with a negative dentry "file" with 0 refcount and on the dentry LRU list,
and its parent the positive dentry "dir" with 1 refcount.  And the test program
generates thousands of copies of that, with each "dir" referring to a different
inode, and they get freed only when the shrinker runs.

I'm not sure how to fix it...

Eric