]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commitdiff
bcachefs: Fix transaction restarts due to upgrading of cloned iterators
authorKent Overstreet <kent.overstreet@gmail.com>
Fri, 16 Apr 2021 18:29:26 +0000 (14:29 -0400)
committerKent Overstreet <kent.overstreet@linux.dev>
Sun, 22 Oct 2023 21:09:00 +0000 (17:09 -0400)
This fixes a regression from
  52d86202fd bcachefs: Improve bch2_btree_iter_traverse_all()

We want to avoid mucking with other iterators in the btree transaction
in operations that are only supposed to be touching individual iterators
- that patch was a cleanup to move lock ordering handling to
bch2_btree_iter_traverse_all(). But it broke upgrading of cloned
iterators.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
fs/bcachefs/btree_iter.c

index f82976aab7d93b50c624ceedeb5236dae5b45bf0..11f7b47e3e7f4eaaa042fa5271225ac7ca30a217 100644 (file)
@@ -387,11 +387,44 @@ bool bch2_btree_iter_relock(struct btree_iter *iter, bool trace)
 bool __bch2_btree_iter_upgrade(struct btree_iter *iter,
                               unsigned new_locks_want)
 {
+       struct btree_iter *linked;
+
        EBUG_ON(iter->locks_want >= new_locks_want);
 
        iter->locks_want = new_locks_want;
 
-       return btree_iter_get_locks(iter, true, true);
+       if (btree_iter_get_locks(iter, true, true))
+               return true;
+
+       /*
+        * XXX: this is ugly - we'd prefer to not be mucking with other
+        * iterators in the btree_trans here.
+        *
+        * On failure to upgrade the iterator, setting iter->locks_want and
+        * calling get_locks() is sufficient to make bch2_btree_iter_traverse()
+        * get the locks we want on transaction restart.
+        *
+        * But if this iterator was a clone, on transaction restart what we did
+        * to this iterator isn't going to be preserved.
+        *
+        * Possibly we could add an iterator field for the parent iterator when
+        * an iterator is a copy - for now, we'll just upgrade any other
+        * iterators with the same btree id.
+        *
+        * The code below used to be needed to ensure ancestor nodes get locked
+        * before interior nodes - now that's handled by
+        * bch2_btree_iter_traverse_all().
+        */
+       trans_for_each_iter(iter->trans, linked)
+               if (linked != iter &&
+                   btree_iter_type(linked) == btree_iter_type(iter) &&
+                   linked->btree_id == iter->btree_id &&
+                   linked->locks_want < new_locks_want) {
+                       linked->locks_want = new_locks_want;
+                       btree_iter_get_locks(linked, true, false);
+               }
+
+       return false;
 }
 
 void __bch2_btree_iter_downgrade(struct btree_iter *iter,