Re: [PATCH] xfs: clear PF_MEMALLOC before exiting xfsaild thread

[Eric Biggers]

On Mon, Mar 09, 2020 at 10:03:07AM +1100, Dave Chinner wrote:
> On Sat, Mar 07, 2020 at 08:35:40PM -0800, Eric Biggers wrote:
> > From: Eric Biggers <ebiggers@xxxxxxxxxx>
> > 
> > Leaving PF_MEMALLOC set when exiting a kthread causes it to remain set
> > during do_exit().  That can confuse things.  For example, if BSD process
> > accounting is enabled, then it's possible for do_exit() to end up
> > calling ext4_write_inode().  That triggers the
> > WARN_ON_ONCE(current->flags & PF_MEMALLOC) there, as it assumes
> > (appropriately) that inodes aren't written when allocating memory.
> 
> And just how the hell does and XFS kernel thread end up calling
> ext4_write_inode()? That's kinda a key factor in all this, and
> it's not explained here.
> 
> > This case was reported by syzbot at
> > https://lkml.kernel.org/r/0000000000000e7156059f751d7b@xxxxxxxxxxx
> 
> Which doesn't really explain it, either.
> 
> What is the configuration conditions under which this triggers? It
> looks like some weird combination of a no-journal ext4 root
> filesystem and the audit subsystem being configured with O_SYNC
> files?
> 
> People trying to decide if this is something that needs to be
> backported to stable kernels need to be able to unerstand how this
> bug is actually triggered so they can make sane decisions about
> it...

My guess is that syzbot enabled BSD process accounting to a file with FS_SYNC_FL
on an ext4 nojournal fs.  It didn't provide a reproducer itself.  Sure, I'll try
to write a reproducer and include it in the commit message.  I felt it wasn't
quite as important for this one compared to most of the other syzbot bugs, since
BSD process accounting can only be enabled by root, and once we're talking about
do_exit() being able to write an arbitrary file, it's not hard to see why
*something* could get tripped up by PF_MEMALLOC.  There are probably other ways
it could cause problems besides this specific one.  But sure, I'll try.

> 
> I also note that cifs_demultiplex_thread() has the same problem -
> can you please do a complete audit of all the users of PF_MEMALLOC
> and fix all of them?

I already did, that's why at the same time I sent out this patch, I also sent
out one to fix cifs_demultiplex_thread()
(https://lkml.kernel.org/linux-cifs/20200308043645.1034870-1-ebiggers@xxxxxxxxxx/T/#t).
These were the only two; I didn't find any others that needed to be fixed.

- Eric