Re: [PATCH] dma/debug: Fix dma vs cow-page collision detection

[Dan Williams <dan.j.williams@intel.com>]

On Tue, Nov 19, 2019 at 4:02 PM Alexander Duyck
<alexander.duyck@gmail.com> wrote:
>
> On Tue, Nov 19, 2019 at 9:49 AM Dan Williams <dan.j.williams@intel.com> wrote:
> >
> > The debug_dma_assert_idle() infrastructure was put in place to catch a
> > data corruption scenario first identified by the now defunct NET_DMA
> > receive offload feature. It caught cases where dma was in flight to a
> > stale page because the dma raced the cpu writing the page, and the cpu
> > write triggered cow_user_page().
> >
> > However, the dma-debug tracking is overeager and also triggers in cases
> > where the dma device is reading from a page that is also undergoing
> > cow_user_page().
> >
> > The fix proposed was originally posted in 2016, and Russell reported
> > "Yes, that seems to avoid the warning for me from an initial test", and
> > now Don is also reporting that this fix is addressing a similar false
> > positive report that he is seeing.
> >
> > Link: https://lore.kernel.org/r/CAPcyv4j8fWqwAaX5oCdg5atc+vmp57HoAGT6AfBFwaCiv0RbAQ@mail.gmail.com
> > Reported-by: Russell King <linux@armlinux.org.uk>
> > Reported-by: Don Dutile <ddutile@redhat.com>
> > Fixes: 0abdd7a81b7e ("dma-debug: introduce debug_dma_assert_idle()")
> > Cc: <stable@vger.kernel.org>
> > Cc: Christoph Hellwig <hch@lst.de>
> > Cc: Marek Szyprowski <m.szyprowski@samsung.com>
> > Cc: Robin Murphy <robin.murphy@arm.com>
> > Signed-off-by: Dan Williams <dan.j.williams@intel.com>
> > ---
> >  kernel/dma/debug.c |    2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/kernel/dma/debug.c b/kernel/dma/debug.c
> > index 099002d84f46..11a6db53d193 100644
> > --- a/kernel/dma/debug.c
> > +++ b/kernel/dma/debug.c
> > @@ -587,7 +587,7 @@ void debug_dma_assert_idle(struct page *page)
> >         }
> >         spin_unlock_irqrestore(&radix_lock, flags);
> >
> > -       if (!entry)
> > +       if (!entry || entry->direction != DMA_FROM_DEVICE)
> >                 return;
> >
> >         cln = to_cacheline_number(entry);
>
> If I am understanding right DMA_TO_DEVICE is fine, but won't  you also
> need to cover the DMA_BIDIRECTIONAL case since it is possible for a
> device to also write the memory in that case?

True, DMA_BIDIRECTIONAL and DMA_TO_DEVICE are being treated equally in
this case. Given this is the second time this facility needed to be
taught to be less eager [1], I'd be inclined to let the tie-break /
BIDIR case be treated like TO. This facility was always meant as a
"there might be a problem here", but not a definitive checker, and it
certainly loses value if the reports are ambiguous.

[1]: 3b7a6418c749 dma debug: account for cachelines and read-only
mappings in overlap tracking