Re: Circular lock dep in btrfs triggered by shrinker

[David Sterba]

On Mon, Oct 22, 2018 at 03:07:26AM -0700, Olof Johansson wrote:
> I hit the below circular locking dependency. Seems like the assumption made in
> 712e36c5f2a7fa56 ("btrfs: use GFP_KERNEL in btrfs_alloc_inode") either isn't
> true, or has since changed?

I think it must have been there from the beginning. There were reports
of this lockdep warning like the below and IIRC a few more
(https://lkml.kernel.org/r/20180627120523.GA2287@xxxxxxx), but without a
resolution.

Incidentally, there was a fix that's now in the 4.20 pull and only after
I had seen your report I realized that it was the fix for the warning:

https://git.kernel.org/pub/scm/linux/kernel/git/kdave/linux.git/commit/?h=for-4.20-part1&id=84de76a2fb217dc1b6bc2965cc397d1648aa1404

It disables the filesystem allocations using the memalloc_nofs mechanism
around free space inode allocation, while my original patch expected
only regular inodes created by VFS.

> ======================================================
> WARNING: possible circular locking dependency detected
> 4.19.0 #25 Not tainted
> ------------------------------------------------------
> kswapd0/414 is trying to acquire lock:
> 000000008b8f1971 (&delayed_node->mutex){+.+.}, at: __btrfs_release_delayed_node+0x35/0x1e0
> but task is already holding lock:
> 00000000032f657e (fs_reclaim){+.+.}, at: __fs_reclaim_acquire+0x0/0x30
> which lock already depends on the new lock.
> 
> the existing dependency chain (in reverse order) is:
> -> #3 (fs_reclaim){+.+.}:
>        kmem_cache_alloc+0x24/0x270
>        btrfs_alloc_inode+0x1f/0x260
>        alloc_inode+0x13/0x80
>        iget5_locked+0x3f/0x80
>        btrfs_iget+0x52/0x680
>        __lookup_free_space_inode+0xd9/0x110

The patch adds memalloc_nofs into ^^^^.  It's scheduled for stable so
the warning will disappear eventually, thanks for the report.