Re: [block:for-3.14/core] kernel BUG at fs/bio.c:1748

[Kent Overstreet <kmo@daterainc.com>]

On Wed, Jan 08, 2014 at 01:18:46PM -0800, Muthu Kumar wrote:
> On Wed, Jan 8, 2014 at 1:14 PM, Kent Overstreet <kmo@daterainc.com> wrote:
> > On Wed, Jan 08, 2014 at 09:11:49PM +0000, Chris Mason wrote:
> >> On Wed, 2014-01-08 at 13:01 -0800, Muthu Kumar wrote:
> >> > On Wed, Jan 8, 2014 at 12:51 PM, Chris Mason <clm@fb.com> wrote:
> >> > > On Wed, 2014-01-08 at 12:40 -0800, Muthu Kumar wrote:
> >> > >> On Wed, Jan 8, 2014 at 12:16 PM, Chris Mason <clm@fb.com> wrote:
> >> > >> > On Wed, 2014-01-08 at 11:54 -0800, Muthu Kumar wrote:
> >> > >> >> Chris,
> >> > >> >>
> >> > >> >> [    8.336061] WARNING: CPU: 0 PID: 0 at fs/bio.c:1778 bio_endio+0xbe/0x100()
> >> > >> >> [    8.336062] bio_endio: bio for (unknown) without endio
> >> > >> >>
> >> > >> >> This is my recent change to avoid memory leak in bio_endio. But I
> >> > >> >> think the problem is higher up, most likely bio_endio is called twice
> >> > >> >> on the same bio (which was freed before).
> >> > >> >>
> >> > >> >
> >> > >> > I think these are just two separate problems.  Lets ignore the WARN_ON
> >> > >> > for now.
> >> > >> >
> >> > >>
> >> > >> Not really... the BUG that is triggered:
> >> > >>
> >> > >> kernel BUG at fs/bio.c:523!
> >> > >>
> >> > >> It is in bio_put() (added to bio_endio() as part of recent change)
> >> > >> which gets an already freed bio.
> >> > >>
> >> > >
> >> > > Oh! I see.  Let me try with that one reverted.  Thanks!
> >> > >
> >> > > -chris
> >> > >
> >> >
> >> > But, like I said, problem is in different place. I am running a "dd"
> >> > on ext4 fs for a while now, but didn't hit the problem. Any idea to
> >> > repro locally? I would also suggest running just the for-3.1/core to
> >> > isolate the issue.
> >>
> >> Just reverting that change fixes it for me.  Jens mentioned it was
> >> broken for on-stack bios.
> >
> > On-stack bios? I don't recall ever coming across such a thing, who what
> > where why?
> >
> > i would expect on stack bios to work though, i'm really curious how it
> > was broken
> 
> New change added a bio_put() which might not work if the bio is on stack.
> 
> I don't remember seeing a on-stack-bio either, any help to jog my memory?

That's code that logically belongs in bio_chain_endio(), it's just a
hack to avoid blowing the stack since the kernel is compiled with
-fno-sibling-call-optimization when you enable frame pointers (otherwise
would optimize those tail calls to jumps and we'd have no stack blowing
issues).