repeatable btrfs deadlock in unlink, kernel versions v3.15..v3.18.3

Processes keep getting stuck in btrfs_evict_inode during unlink.

I've seen this dozens of times, usually when two subvols on a btrfs filesystem are active at the same time (i.e. it never happens on single-subvol filesystems). It happens on kernel versions v3.15..v3.18.3.

This used to happen with SyS_renameat2 on the stack, but I haven't seen a stack trace including that function since late November (v3.18-rc6).

	[<ffffffff81387dc0>] lock_extent_bits+0x1b0/0x200
	[<ffffffff81375e99>] btrfs_evict_inode+0x139/0x550
	[<ffffffff8120d708>] evict+0xb8/0x190
	[<ffffffff8120dec5>] iput+0x105/0x1a0
	[<ffffffff812001d9>] do_unlinkat+0x189/0x2d0
	[<ffffffff81202e26>] SyS_unlink+0x16/0x20
	[<ffffffff8185e96d>] system_call_fastpath+0x1a/0x1f
	[<ffffffffffffffff>] 0xffffffffffffffff