Re: linux-stable-5.10-y CVE-2022-1508 of io_uring module

[Pavel Begunkov]

On 5/7/22 15:18, Jens Axboe wrote:
> On 5/7/22 3:16 AM, Pavel Begunkov wrote:
> > On 5/6/22 19:22, Jens Axboe wrote:
> > > On 5/6/22 10:15 AM, Jens Axboe wrote:
> > > > On 5/6/22 9:57 AM, Pavel Begunkov wrote:
> > > > > On 5/6/22 03:16, Jens Axboe wrote:
> > > > > > On 5/5/22 8:11 AM, Guo Xuenan wrote:
> > > > > > > Hi, Pavel & Jens
> > > > > > >
> > > > > > > CVE-2022-1508[1] contains an patch[2] of io_uring. As Jones reported,
> > > > > > > it is not enough only apply [2] to stable-5.10.
> > > > > > > Io_uring is very valuable and active module of linux kernel.
> > > > > > > I've tried to apply these two patches[3] [4] to my local 5.10 code, I
> > > > > > > found my understanding of io_uring is not enough to resolve all conflicts.
> > > > > > >
> > > > > > > Since 5.10 is an important stable branch of linux, we would appreciate
> > > > > > > your help in solving this problem.
> > > > > >
> > > > > > Yes, this really needs to get buttoned up for 5.10. I seem to recall
> > > > > > there was a reproducer for this that was somewhat saner than the
> > > > > > syzbot one (which doesn't do anything for me). Pavel, do you have one?
> > > > >
> > > > > No, it was the only repro and was triggering the problem
> > > > > just fine back then
> > > >
> > > > I modified it a bit and I can now trigger it.
> > >
> > > Pavel, why don't we just keep it really simple and just always save the
> > > iter state in read/write, and use the restore instead of the revert?
> >
> > The problem here is where we're doing revert. If it's done deep in
> > the stack and then while unwinding someone decides to revert it again,
> > e.g. blkdev_read_iter(), we're screwed.
> >
> > The last attempt was backporting 20+ patches that would move revert
> > into io_read/io_write, i.e. REQ_F_REISSUE, back that failed some of
> > your tests back then. (was it read retry tests iirc?)
>
> Do you still have that series? Yes, if I recall correctly, the series

Yep, still in the repo:

https://github.com/isilence/linux/tree/5.10_revert

> had an issue with the resubmit. Which might just be minor, I don't
> believe we really took a closer look at that.
>
> Let's resurrect that series and see if we can pull it to completion,
> would be nice to finally close the chapter on this issue for 5.10...

We can try, but I'm not too comfortable with those backports, I had
to considerably rewrite last three patches or so. Another option
is to disable retries from the rw callback if the iter has been
truncated.

--
Pavel Begunkov