Re: [PATCH] loop: don't call vfs_flush() with queue frozen

[Christoph Hellwig <hch@infradead.org>]

On Wed, Jan 15, 2025 at 04:21:40PM +0800, Ming Lei wrote:
> On Tue, Jan 14, 2025 at 11:05:00PM -0800, Christoph Hellwig wrote:
> > On Wed, Jan 15, 2025 at 10:37:08AM +0800, Ming Lei wrote:
> > > Actually some FSs may call kmalloc(GFP_KERNEL) with i_rwsem grabbed,
> > > which could call into real deadlock if IO on the loop disk is caused by
> > > the kmalloc(GFP_KERNEL).
> > 
> > Well, loop can always deadlock when the lower fs is doing allocations,
> > even without the freeze.  I'm actually kinda surprised loop doesn't
> > force a context noio as we'd really need that.
> 
> Loop does call mapping_set_gfp_mask(~(__GFP_IO|__GFP_FS)).

But that only helps with page cache allocations, not other allocations
don't in the I/O path, of which there are plenty in modern file systems.

> > And all of these are caused by not breaking the dependency loop..
> 
> Can you share how to break the dependency? Such as fs_reclaim &
> mm->mmap_lock.

fs_reclaim should not be broken.  It's a fake lockdep key annotation
the reclaim path.  Everything under it needs to be NOFS, or in case
of loop NOIO (see above).  I think mmap_lock is just an indirect chain
here, but I need to take a closer look.  A little busty this week and
only sporadically online, sorry.