netfilter: nf_tables: nft_set_rbtree: fix spurious insertion failure

[ Upstream commit 08738827 ]

nft_rbtree_gc_elem() walks back and removes the end interval element that
comes before the expired element.

There is a small chance that we've cached this element as 'rbe_ge'.
If this happens, we hold and test a pointer that has been queued for
freeing.

It also causes spurious insertion failures:

$ cat test-testcases-sets-0044interval_overlap_0.1/testout.log
Error: Could not process rule: File exists
add element t s {  0 -  2 }
                   ^^^^^^
Failed to insert  0 -  2 given:
table ip t {
        set s {
                type inet_service
                flags interval,timeout
                timeout 2s
                gc-interval 2s
        }
}

The set (rbtree) is empty. The 'failure' doesn't happen on next attempt.

Reason is that when we try to insert, the tree may hold an expired
element that collides with the range we're adding.
While we do evict/erase this element, we can trip over this check:

if (rbe_ge && nft_rbtree_interval_end(rbe_ge) && nft_rbtree_interval_end(new))
      return -ENOTEMPTY;

rbe_ge was erased by the synchronous gc, we should not have done this
check.  Next attempt won't find it, so retry results in successful
insertion.

Restart in-kernel to avoid such spurious errors.

Such restart are rare, unless userspace intentionally adds very large
numbers of elements with very short timeouts while setting a huge
gc interval.

Even in this case, this cannot loop forever, on each retry an existing
element has been removed.

As the caller is holding the transaction mutex, its impossible
for a second entity to add more expiring elements to the tree.

After this it also becomes feasible to remove the async gc worker
and perform all garbage collection from the commit path.

Fixes: c9e6978e

 ("netfilter: nft_set_rbtree: Switch to node list walk for overlap detection")
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Sasha Levin <sashal@kernel.org>