Re: msync(2) bug(?), returns AOP_WRITEPAGE_ACTIVATE to userland

[Ryan Finnie]

On 10/11/07, Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> shit.  That's a nasty bug.  Really userspace should be testing for -1, but
> the msync() library function should only ever return 0 or -1.
>
> Does this fix it?
>
> --- a/mm/page-writeback.c~a
> +++ a/mm/page-writeback.c
> @@ -850,8 +850,10 @@ retry:
>
>                         ret = (*writepage)(page, wbc, data);
>
> -                       if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE))
> +                       if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
>                                 unlock_page(page);
> +                               ret = 0;
> +                       }
>                         if (ret || (--(wbc->nr_to_write) <= 0))
>                                 done = 1;
>                         if (wbc->nonblocking && bdi_write_congested(bdi)) {
> _
>

Pekka Enberg replied with an identical patch a few days ago, but for
some reason the same condition flows up to msync as -1 EIO instead of
AOP_WRITEPAGE_ACTIVATE with that patch applied.  The last part of the
thread is below.  Thanks.

Ryan

On 10/7/07, Ryan Finnie <ryan@xxxxxxxxxx> wrote:
> On 10/7/07, Pekka J Enberg <penberg@xxxxxxxxxxxxxx> wrote:
> > On 10/7/07, Erez Zadok <ezk@xxxxxxxxxxxxx> wrote:
> > > Anyway, some Ubuntu users of Unionfs reported that msync(2) sometimes
> > > returns AOP_WRITEPAGE_ACTIVATE (decimal 524288) back to userland.
> > > Therefore, some user programs fail, esp. if they're written such as
> > > this:
> >
> ...
> > It's a kernel bug. AOP_WRITEPAGE_ACTIVATE is a hint to the VM to avoid
> > writeback of the page in the near future. I wonder if it's enough that we
> > change the return value to zero from
> > mm/page-writeback.c:write_cache_pages() in case we hit AOP_WRITEPAGE_ACTIVE...
>
> Doesn't appear to be enough.  I can't figure out why (since it appears
> write_cache_pages bubbles up directly to sys_msync), but with that
> patch applied, in my test case[1], msync returns -1 EIO.  However,
> with the exact same kernel without that patch applied, msync returns
> 524288 (AOP_WRITEPAGE_ACTIVATE).  But as your patch specifically flips
> 524288 to 0, I can't figure out how it eventually returns  -1 EIO.
>
> Ryan
>
> [1] "apt-get check" on a unionfs2 mount backed by tmpfs over cdrom,
> standard livecd setup
>
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/