On Sat, Oct 07, 2017 at 03:02:17AM +0100, Al Viro wrote:
> > I do wonder if we shouldn't just use something like
> >
> > "skip leading zeroes, copy to size-limited stack location instead"
> >
> > because the input length really *is* limited once you skip leading
> > zeroes (and whatever base marker we have). We might have at most a
> > 64-bit value in octal, so 22 bytes max.
> >
> > But I guess just changing the two GFP_KERNEL's to GFP_ATOMIC is much
> > simpler.
>
> There's match_strdup() as well...
>
> FWIW, ext2 side also looks fishy; it might be cleaner if we
> collected new state into some object and applied it only after the last
> possible failure exit. The entire "restore the original state" logics
> would go away...
I'm not saying that the bug had been introduced by conversion to
spinlock, BTW - it was racy back when ext2_remount() relied upon BKL.
I hadn't considered the atomicity issues back then - mea culpa...
