Re: [PATCH STABLE 5.10 5.11] swap: fix swapfile page to sector mapping

[Anthony Iliopoulos <ailiop@suse.com>]

On Thu, Mar 04, 2021 at 05:58:49PM +0100, Greg Kroah-Hartman wrote:
> On Thu, Mar 04, 2021 at 05:30:00PM +0100, Anthony Iliopoulos wrote:
> > On Thu, Mar 04, 2021 at 04:16:26PM +0100, Greg Kroah-Hartman wrote:
> > > On Thu, Mar 04, 2021 at 04:08:24PM +0100, Anthony Iliopoulos wrote:
> > > > commit caf6912f3f4af7232340d500a4a2008f81b93f14 upstream.
> > > 
> > > No, this does not look like that commit.
> > > 
> > > Why can I not just take caf6912f3f4a ("swap: fix swapfile read/write
> > > offset") directly for 5.10 and 5.11?  WHat has changed to prevent that?
> > 
> > You're right of course, the upstream fix applies even on v5.4 so you
> > could just take it directly for those branches if this is preferable.
> 
> But, that commit says it fixes 48d15436fde6 ("mm: remove get_swap_bio"),
> which is NOT what you are saying here in these patches.

It is admittedly a bit confusing as the upstream commit fixes two issues
in one swoop:

- the bug which was introduced in v5.12-rc1 via 48d15436fde6 ("mm:
  remove get_swap_bio"), which affected swapfiles running on regular
  block devices, in addition to:

- an identical bug which up until 48d15436fde6 was only applicable to
  swapfiles on top of blockdevs that can do page io without the block
  layer, which was introduced with dd6bd0d9c7db ("swap: use
  bdev_read_page() / bdev_write_page()")

> So which is it?  Is there a problem in 5.11 and older kernels
> (48d15436fde6 ("mm: remove get_swap_bio") showed up in 5.12-rc1), that
> requires this fix, or is there nothing needed to be backported?

The second point/bug mentioned above is present on 5.11 and all older
kernels, so some form of this fix is required.

> As a note, I've been running swapfiles on 5.11 and earlier just fine for
> a very long time now, so is this really an issue?

Yes there is an issue on all kernels since v3.16-rc1 when dd6bd0d9c7db
was introduced, but it is applicable only to setups with swapfiles on
filesystems sitting on top of brd, zram, btt or pmem.

I can trivially reproduce this e.g. on v5.11 by creating a swapfile on
top of a zram or pmem blockdev and pushing the system to swap out pages,
at which point it corrupts filesystem blocks that don't belong to the
swapfile.

Regards,
Anthony