On Thu, 28 Aug 2025, Amir Goldstein wrote:
>
> Neil,
>
> FYI, if your future work for vfs assumes that fs will alway have the
> dentry hashed after create, you may want to look at:
>
> static int ovl_instantiate(struct dentry *dentry, struct inode *inode,
> ...
> /* Force lookup of new upper hardlink to find its lower */
> if (hardlink)
> d_drop(dentry);
>
> return 0;
> }
>
> If your assumption is not true for overlayfs, it may not be true for other fs
> as well. How could you verify that it is correct?
I don't need the dentry to be hashed after the create has completed (or
failed).
I only need it to be hashed when the create starts, and ideally for the
duration of the creation process.
Several filesystems d_drop() a newly created dentry so as to trigger a
lookup - overlayfs is not unique.
>
> I really hope that you have some opt-in strategy in mind, so those new
> dirops assumptions would not have to include all possible filesystems.
Filesystems will need to opt-in to not having the parent locked. If
a fs still has the parent locked across operations it doesn't really
matter when the d_drop() happens. However I want to move all the
d_drop()s to the end (which is where ovl has it) to ensure there are no
structural issues that mean an early d_drop() is needed. e.g. Some
filesystems d_drop() and then d_splice_alias() and I want to add a new
d_splice_alias() variant that doesn't require the d_drop().
So it is only at the start of an operation (create, remove, rename) that
I need the dentry to be hashed. That raises questions about ext4_lookup
not hashing a negative dentry as a lookup-create pair in do_mknodat or
lookup_open could call vfs_create with a non-hashed dentry.
That isn't *actually* a problem (I think - I should double-check) as the
dentry is still d_in_lookup() so it is hashed in the separate
in_lookup_hashtable(). So a d_lookup() will find it even though it
isn't hashed.
That suggests an alternate fix for ovl_parent_lock(). Rather than
insisting that the child is hashed, we can insist that either
d_in_lookup(child) || !d_unhashed(child)
Such a dentry really is hashed: it might be hashed in one table, it
might be hashed in the other.
However that wouldn't protect against filesystems which deliberately
d_drop() during create, so I think ovl still needs to perform a lookup
after a create and before a rename - if the create succeeds but the
dentry is negative.
Thanks,
NeilBrown
