Re: bcachefs: Kernel panic - not syncing: trans path oveflow

[Kent Overstreet <kent.overstreet@gmail.com>]

On Wed, Mar 09, 2022 at 01:14:58PM -0800, Eric Wheeler wrote:
> Hi Kent,
> 
> We just started testing bcachefs snapshots this week: we have a bunch of 
> mysql replicas, each in its own subvolume.  Every 4 hours we stop mysql, 
> run a subvolume snapshot and restart mysql, so it gets lots of snapshot 
> and sync IO from the many database instances.

Cool! Would love to hear any comments you've got so far.

> We hit the following bcachefs panic while testing commit# 
> 5490c9c529770aa18b2571bd98f5416ed9ae24c6 from March 3rd. Can you tell what 
> the issue might be?
> 
> It is easily reproducable, the same problem hits shortly after we reboot 
> and remount so happy to test patches or git-pull's to rebuild with:
> 
> Here is the stack trace (more logs below):

So it looks like there's some code that iterates over btree keys and goes
further than it's supposed to - we have paths that point to different inode
numbers and that's not supposed to happen in the write path, we're only updating
a single inode.

I've had a report of a similar bug in the data move path, which may or may not
be the same as this bug - but I haven't worked up a repro for it yet so I
haven't figured out yet which code path is allocating these btree paths. Could
you enable CONFIG_BCACHEFS_DEBUG, then run your log through
scripts/decode_stacktrace.sh from the kernel source tree?