Re: BUG: unable to handle kernel NULL pointer dereference in xlog_cil_commit

[Darrick J. Wong]

On Thu, Oct 07, 2021 at 09:27:49AM +1100, Dave Chinner wrote:
> On Wed, Oct 06, 2021 at 08:43:27AM -0700, Darrick J. Wong wrote:
> > On Wed, Oct 06, 2021 at 04:14:43PM +0800, Hao Sun wrote:
> > > Hello,
> > > 
> > > When using Healer to fuzz the latest Linux kernel, the following crash
> > > was triggered.
> > > 
> > > HEAD commit: 0513e464f900 Merge tag 'perf-tools-fixes-for-v5.15-2021-09-27'
> > > git tree: upstream
> > > console output:
> > > https://drive.google.com/file/d/1vm5fDM220kkghoiGa3Aw_Prl4O_pqAXF/view?usp=sharing
> > > kernel config: https://drive.google.com/file/d/1Jqhc4DpCVE8X7d-XBdQnrMoQzifTG5ho/view?usp=sharing
> > > 
> > > Sorry, I don't have a reproducer for this crash, hope the symbolized
> > > report can help.
> > > If you fix this issue, please add the following tag to the commit:
> > > Reported-by: Hao Sun <sunhao.th@xxxxxxxxx>
> > 
> > So figure out how to fix the problem and send a patch.  You don't get to
> > hand out fixit tasks like you're some kind of manager for people you
> > don't employ.
> 
> I fully agree with this Darrick but, OTOH, the cynical, jaded
> engineer in me says "I don't think people that run bots and
> copy/paste their output to mailing lists have the capability to fix
> the problems the bots find."

They gotta learn to level up (or leave us alone) somehow...

> Quite frankly, it's even more of a waste of our time trying to
> review crap patches and make suggestions to fix it and then going
> around the review loop 15 times getting nowhere like we have in teh
> past.
> 
> So, kvmalloc() sucks dogs balls, as I pointed out in this recent
> patch in the intent whiteouts series:
> 
> https://lore.kernel.org/linux-xfs/20210902095927.911100-8-david@xxxxxxxxxxxxx/
> 
> Because of the crap implementation of kvmalloc(), we can't just pass
> __GFP_NOFAIL because that will cause it to try to run
> kmalloc_node(__GFP_NORETRY | __GFP_NOFAIL) and that will cause heads
> to go all explodey. Not to mention that kvmalloc won't even allow
> GFP_NOFS to be passed and still actually do the vmalloc() fallback.
> 
> So, basically, we've got to go back to doing an open coded kvmalloc
> loop here that cannot fail. Because kvmalloc can fail and we can't
> tell it that it must succeed or die trying.
> 
> That's what the above patch does - gets rid of the garbage kvmalloc
> direct reclaim -> memory compaction behaviour, and wraps it in a
> loop so that the fail-fast memory allocation semantics it uses does
> not end up in a shadow buffer allocation failure.
> 
> So, yeah, I've already fixed this in my trees....

Er, do you want me to queue that patch for 5.17?  I had assumed it was
only really needed for allison's patches, but I don't mind pushing it
in sooner.

--D

> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@xxxxxxxxxxxxx