On Wed, Dec 09, 2020 at 08:14:15PM +0000, Matthew Wilcox wrote:
> On Wed, Dec 09, 2020 at 11:47:56AM -0800, Dan Williams wrote:
> > On Tue, Dec 8, 2020 at 8:03 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > > On Tue, Dec 08, 2020 at 06:22:50PM -0800, Ira Weiny wrote:
> > > > Therefore, I tend to agree with Dan that if anything is to be done it should be
> > > > a WARN_ON() which is only going to throw an error that something has probably
> > > > been wrong all along and should be fixed but continue running as before.
> > >
> > > Silent data corruption is for ever. Are you absolutely sure nobody has
> > > done:
> > >
> > > page = alloc_pages(GFP_HIGHUSER_MOVABLE, 3);
> > > memcpy_to_page(page, PAGE_SIZE * 2, p, PAGE_SIZE * 2);
> > >
> > > because that will work fine if the pages come from ZONE_NORMAL and fail
> > > miserably if they came from ZONE_HIGHMEM.
> >
> > ...and violently regress with the BUG_ON.
>
> ... which is what we want, no?
>
> > The question to me is: which is more likely that any bad usages have
> > been covered up by being limited to ZONE_NORMAL / 64-bit only, or that
> > silent data corruption has been occurring with no ill effects?
>
> I wouldn't be at all surprised to learn that there is silent data
> corruption on 32-bit systems with HIGHMEM. Would you? How much testing
> do you do on 32-bit HIGHMEM systems?
>
> Actually, I wouldn't be at all surprised if we can hit this problem today.
> Look at this:
>
> size_t _copy_from_iter(void *addr, size_t bytes, struct iov_iter *i)
> {
> char *to = addr;
> if (unlikely(iov_iter_is_pipe(i))) {
> WARN_ON(1);
> return 0;
> }
> if (iter_is_iovec(i))
> might_fault();
> iterate_and_advance(i, bytes, v,
> copyin((to += v.iov_len) - v.iov_len, v.iov_base, v.iov_len),
> memcpy_from_page((to += v.bv_len) - v.bv_len, v.bv_page,
> v.bv_offset, v.bv_len),
> memcpy((to += v.iov_len) - v.iov_len, v.iov_base, v.iov_len)
> )
>
> return bytes;
> }
> EXPORT_SYMBOL(_copy_from_iter);
>
> There's a lot of macrology in there, so for those following along who
> aren't familiar with the iov_iter code, if the iter is operating on a
> bvec, then iterate_and_advance() will call memcpy_from_page(), passing
> it the bv_page, bv_offset and bv_len stored in the bvec. Since 2019,
> Linux has supported multipage bvecs (commit 07173c3ec276). So bv_len
> absolutely *can* be > PAGE_SIZE.
>
> Does this ever happen in practice? I have no idea; I don't know whether
> any multipage BIOs are currently handed to copy_from_iter(). But I
> have no confidence in your audit if you didn't catch this one.
Ah... This call site has been there since 2014 and is not a new caller I have
been 'auditing'.[1]
>
> > > > FWIW I think this is a 'bad BUG_ON' use because we are "checking something that
> > > > we know we might be getting wrong".[1] And because, "BUG() is only good for
> > > > something that never happens and that we really have no other option for".[2]
> > >
> > > BUG() is our only option here. Both limiting how much we copy or
> > > copying the requested amount result in data corruption or leaking
> > > information to a process that isn't supposed to see it.
> >
> > At a minimum I think this should be debated in a follow on patch to
> > add assertion checking where there was none before. There is no
> > evidence of a page being overrun in the audit Ira performed.
>
> If we put in into a separate patch, someone will suggest backing out the
> patch which tells us that there's a problem. You know, like this guy ...
> https://lore.kernel.org/linux-mm/CAPcyv4jNVroYmirzKw_=CsEixOEACdL3M1Wc4xjd_TFv3h+o8Q@xxxxxxxxxxxxxx/
I'm not following this. Regardless I've already added the BUG_ON's.
Ira
[1]
commit 0dbca9a4b5d69a7e4b8c1d55b98312fcd9aafcf7
Author: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
Date: Thu Nov 27 14:26:43 2014 -0500
iov_iter.c: convert copy_from_iter() to iterate_and_advance
Signed-off-by: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
