Zygo reported the following panic when testing my error handling patches
for relocation:

  kernel BUG at fs/btrfs/backref.c:2545!
  invalid opcode: 0000 [#1] SMP KASAN PTI CPU: 3 PID: 8472 Comm: btrfs Tainted: G        W 14
  Hardware name: QEMU Standard PC (i440FX + PIIX,

  Call Trace:
   btrfs_backref_error_cleanup+0x4df/0x530
   build_backref_tree+0x1a5/0x700
   ? _raw_spin_unlock+0x22/0x30
   ? release_extent_buffer+0x225/0x280
   ? free_extent_buffer.part.52+0xd7/0x140
   relocate_tree_blocks+0x2a6/0xb60
   ? kasan_unpoison_shadow+0x35/0x50
   ? do_relocation+0xc10/0xc10
   ? kasan_kmalloc+0x9/0x10
   ? kmem_cache_alloc_trace+0x6a3/0xcb0
   ? free_extent_buffer.part.52+0xd7/0x140
   ? rb_insert_color+0x342/0x360
   ? add_tree_block.isra.36+0x236/0x2b0
   relocate_block_group+0x2eb/0x780
   ? merge_reloc_roots+0x470/0x470
   btrfs_relocate_block_group+0x26e/0x4c0
   btrfs_relocate_chunk+0x52/0x120
   btrfs_balance+0xe2e/0x18f0
   ? pvclock_clocksource_read+0xeb/0x190
   ? btrfs_relocate_chunk+0x120/0x120
   ? lock_contended+0x620/0x6e0
   ? do_raw_spin_lock+0x1e0/0x1e0
   ? do_raw_spin_unlock+0xa8/0x140
   btrfs_ioctl_balance+0x1f9/0x460
   btrfs_ioctl+0x24c8/0x4380
   ? __kasan_check_read+0x11/0x20
   ? check_chain_key+0x1f4/0x2f0
   ? __asan_loadN+0xf/0x20
   ? btrfs_ioctl_get_supported_features+0x30/0x30
   ? kvm_sched_clock_read+0x18/0x30
   ? check_chain_key+0x1f4/0x2f0
   ? lock_downgrade+0x3f0/0x3f0
   ? handle_mm_fault+0xad6/0x2150
   ? do_vfs_ioctl+0xfc/0x9d0
   ? ioctl_file_clone+0xe0/0xe0
   ? check_flags.part.50+0x6c/0x1e0
   ? check_flags.part.50+0x6c/0x1e0
   ? check_flags+0x26/0x30
   ? lock_is_held_type+0xc3/0xf0
   ? syscall_enter_from_user_mode+0x1b/0x60
   ? do_syscall_64+0x13/0x80
   ? rcu_read_lock_sched_held+0xa1/0xd0
   ? __kasan_check_read+0x11/0x20
   ? __fget_light+0xae/0x110
   __x64_sys_ioctl+0xc3/0x100
   do_syscall_64+0x37/0x80
   entry_SYSCALL_64_after_hwframe+0x44/0xa9

This occurs because of this check

  if (RB_EMPTY_NODE(&upper->rb_node))
	  BUG_ON(!list_empty(&node->upper));

As we are dropping the backref node, if we discover that our upper node
in the edge we just cleaned up isn't linked into the cache that we are
now done with this node, thus the BUG_ON().

However this is an erroneous assumption, as we will look up all the
references for a node first, and then process the pending edges.  All of
the 'upper' nodes in our pending edges won't be in the cache's rb_tree
yet, because they haven't been processed.  We could very well have many
edges still left to cleanup on this node.

The fact is we simply do not need this check, we can just process all of
the edges only for this node, because below this check we do the
following

  if (list_empty(&upper->lower)) {
	  list_add_tail(&upper->lower, &cache->leaves);
	  upper->lowest = 1;
  }

If the upper node truly isn't used yet, then we add it to the
cache->leaves list to be cleaned up later.  If it is still used then the
last child node that has it linked into its node will add it to the
leaves list and then it will be cleaned up.

Fix this problem by dropping this logic altogether.  With this fix I no
longer see the panic when testing with error injection in the backref
code.

CC: stable@vger.kernel.org # 4.4+
Reviewed-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: David Sterba <dsterba@suse.com>