Re: [syzbot] [mm?] kernel BUG in sg_init_one

[Barry Song <21cnbao@gmail.com>]

On Tue, Mar 19, 2024 at 10:00 AM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> [..]
> > > > I guess that is because on arm32 , we have highmem but
> > > > sg_init_one supports lowmem only. the below should be
> > > > able to fix?
> > > >
> > > > diff --git a/mm/zswap.c b/mm/zswap.c
> > > > index 9dec853647c8..47c0386caba2 100644
> > > > --- a/mm/zswap.c
> > > > +++ b/mm/zswap.c
> > > > @@ -1086,7 +1086,8 @@ static void zswap_decompress(struct zswap_entry
> > > > *entry, struct page *page)
> > > >                 zpool_unmap_handle(zpool, entry->handle);
> > > >         }
> > > >
> > > > -       sg_init_one(&input, src, entry->length);
> > > > +       sg_init_table(&input, 1);
> > > > +       sg_set_page(&input, kmap_to_page(src), entry->length,
> > > > offset_in_page(src));
> > >
> > > Is this working around the debug check in sg_init_one()? IIUC, only
> >
> > I wouldn't characterize it as a workaround; it's more of a solution.
>
> I assumed that the debug check in sg_set_buf() is because
> sg_set_page() cannot handle highmem pages, sorry if that isn't the
> case. Apparently we are hitting a warning with kmap_to_page() though
> as syzbot just reported.
>
> >
> > > lowmem pages are supported. We may be passing in a highmem page to
> > > sg_set_page() now, right?
> >
> > we can pass highmem to sg_set_page(). This is perfectly fine.
>
> So the debug check is only because we are using virt_to_page() in sg_set_buf()?

yes. it is checking if linear_mapping can apply on the buffer.

>
> >
> > >
> > > Also, it seems like if src is a lowmem address kmap_to_page() will be
> > > doing unnecessary checks (assuming it's working correctly)?
> >
> > In practice, we consistently use kmap and kunmap even on systems with
> > low memory.
> > However, it's worth noting that for low memory scenarios, kmap
> > essentially returns
> > page_to_virt(page_address). Thus, the overhead of kmap_to_page shouldn't be
> > significant on low memory systems, especially considering that it simplifies to
> > virt_to_page().
> >
> > Another approach is to consistently employ page_to_virt() for low
> > memory situations
> > and reserve kmap for high memory scenarios. However, since we always
> > utilize kmap
> > regardless of whether the page is low or high memory, we don't need to concern
> > ourselves with this distinction
>
> I see. Thanks for elaborating.
>
> >
> > >
> > > Would it be more robust to just use the temporary buffer if src is a
> > > kmap address?
> >
> > I don't think so because we will need a memcpy then.
>
> I thought that was necessary because sg_set_page() cannot take in
> highmem pages, but you mentioned that this isn't the case.

I think both sg_init_one() and sg_set_page() lack docs. as apparently
sg_init_one() can't take highmem. sg_set_page() can definitely take
highmem as crypto/scompress.c takes care of both high
and low. and scatterwalk_map_and_copy() can handle both.

Thanks
Barry