On 29/9/23 17:10, Miklos Szeredi wrote:

On Fri, 29 Sept 2023 at 02:42, Ian Kent [off-list ref] wrote:

quoted

On 28/9/23 21:01, Miklos Szeredi wrote:

quoted

+static struct vfsmount *lookup_mnt_in_ns(u64 id, struct mnt_namespace *ns)
+{
+     struct mount *mnt;
+     struct vfsmount *res = NULL;
+
+     lock_ns_list(ns);
+     list_for_each_entry(mnt, &ns->list, mnt_list) {
+             if (!mnt_is_cursor(mnt) && id == mnt->mnt_id_unique) {
+                     res = &mnt->mnt;
+                     break;
+             }
+     }
+     unlock_ns_list(ns);
+     return res;
+}

Seems like we might need to consider making (struct mnt_namespace)->list

a hashed list.

Yes, linear search needs to go.  A hash table is probably the easiest solution.

But I'd also consider replacing ns->list with an rbtree.  Not as
trivial as adding a system hash table and probably also slightly
slower, but it would have some advantages:

  - most space efficient (no overhead of hash buckets)

  - cursor can go away (f_pos can just contain last ID)

I guess that would be ok.

Avoiding the cursor is a big plus.


An rbtree is used in kernfs and its readdir function is rather painful so

I wonder what the implications might be for other enumeration needs.


Ian