Re: [PATCH] dma-mapping: Lift address space checks out of debug code

[Kees Cook]

On Thu, Oct 03, 2019 at 10:42:45AM +0100, Robin Murphy wrote:
> On 03/10/2019 00:58, Kees Cook wrote:
> > On Wed, Oct 02, 2019 at 10:15:43PM +0100, Robin Murphy wrote:
> > > Hi Kees,
> > > 
> > > On 2019-10-02 9:46 pm, Kees Cook wrote:
> > > > As we've seen from USB and other areas, we need to always do runtime
> > > > checks for DMA operating on memory regions that might be remapped. This
> > > > consolidates the (existing!) checks and makes them on by default. A
> > > > warning will be triggered for any drivers still using DMA on the stack
> > > > (as has been seen in a few recent reports).
> > > > 
> > > > Suggested-by: Laura Abbott <labbott@xxxxxxxxxx>
> > > > Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
> > > > ---
> > > >    include/linux/dma-debug.h   |  8 --------
> > > >    include/linux/dma-mapping.h |  8 +++++++-
> > > >    kernel/dma/debug.c          | 16 ----------------
> > > >    3 files changed, 7 insertions(+), 25 deletions(-)
> > > > 
> > > > diff --git a/include/linux/dma-debug.h b/include/linux/dma-debug.h
> > > > index 4208f94d93f7..2af9765d9af7 100644
> > > > --- a/include/linux/dma-debug.h
> > > > +++ b/include/linux/dma-debug.h
> > > > @@ -18,9 +18,6 @@ struct bus_type;
> > > >    extern void dma_debug_add_bus(struct bus_type *bus);
> > > > -extern void debug_dma_map_single(struct device *dev, const void *addr,
> > > > -				 unsigned long len);
> > > > -
> > > >    extern void debug_dma_map_page(struct device *dev, struct page *page,
> > > >    			       size_t offset, size_t size,
> > > >    			       int direction, dma_addr_t dma_addr);
> > > > @@ -75,11 +72,6 @@ static inline void dma_debug_add_bus(struct bus_type *bus)
> > > >    {
> > > >    }
> > > > -static inline void debug_dma_map_single(struct device *dev, const void *addr,
> > > > -					unsigned long len)
> > > > -{
> > > > -}
> > > > -
> > > >    static inline void debug_dma_map_page(struct device *dev, struct page *page,
> > > >    				      size_t offset, size_t size,
> > > >    				      int direction, dma_addr_t dma_addr)
> > > > diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
> > > > index 4a1c4fca475a..2d6b8382eab1 100644
> > > > --- a/include/linux/dma-mapping.h
> > > > +++ b/include/linux/dma-mapping.h
> > > > @@ -583,7 +583,13 @@ static inline unsigned long dma_get_merge_boundary(struct device *dev)
> > > >    static inline dma_addr_t dma_map_single_attrs(struct device *dev, void *ptr,
> > > >    		size_t size, enum dma_data_direction dir, unsigned long attrs)
> > > >    {
> > > > -	debug_dma_map_single(dev, ptr, size);
> > > > +	/* DMA must never operate on stack or other remappable places. */
> > > > +	WARN_ONCE(is_vmalloc_addr(ptr) || !virt_addr_valid(ptr),
> > > 
> > > This stands to absolutely cripple I/O performance on arm64, because every
> > > valid call will end up going off and scanning the memblock list, which is
> > > not something we want on a fastpath in non-debug configurations. We'd need a
> > > much better solution to the "pfn_valid() vs. EFI no-map" problem before this
> > > might be viable.
> > 
> > Ah! Interesting. I didn't realize this was fast-path (I don't know the
> > DMA code at all). I thought it was more of a "one time setup" before
> > actual DMA activity started.
> 
> That's strictly true, it's just that many workloads can involve tens of
> thousands of "one time"s per second ;)
> 
> Overhead on the dma_map_* paths has shown to have a direct impact on
> throughput in such situations, hence various optimisation effort in IOVA
> allocation for IOMMU-based DMA ops, and the recent work to remove indirect
> calls entirely for the common dma-direct/SWIOTLB cases.
> 
> > Regardless, is_vmalloc_addr() is extremely light (a bounds check), and is the
> > most important part of this as far as catching stack-based DMA attempts.
> > I thought virt_addr_valid() was cheap too, but I see it's much heavier on
> > arm64.
> > 
> > I just went to compare what the existing USB check does, and it happens
> > immediately before its call to dma_map_single(). Both checks are simple
> > bounds checks, so it shouldn't be an issue:
> > 
> > 			if (is_vmalloc_addr(urb->setup_packet)) {
> > 				WARN_ONCE(1, "setup packet is not dma capable\n");
> > 				return -EAGAIN;
> > 			} else if (object_is_on_stack(urb->setup_packet)) {
> > 				WARN_ONCE(1, "setup packet is on stack\n");
> > 				return -EAGAIN;
> > 			}
> > 
> > 			urb->setup_dma = dma_map_single(
> > 					hcd->self.sysdev,
> > 					urb->setup_packet,
> > 					sizeof(struct usb_ctrlrequest),
> > 
> > 
> > In the USB case, it'll actually refuse to do the operation. Should
> > dma_map_single() similarly fail? I could push these checks down into
> > dma_map_single(), which would be a no-change on behavior for USB and
> > gain the checks on all other callers...
> 
> I think it would be reasonable to pull the is_vmalloc_addr() check inline,
> as that probably covers 90+% of badness (especially given vmapped stacks),
> and as you say should be reliably cheap everywhere. Callers are certainly
> expected to use dma_mapping_error() and handle failure, so refusing to do a
> bogus mapping operation should be OK API-wise - ultimately if a driver goes
> ahead and uses DMA_MAPPING_ERROR as an address anyway, that's not likely to
> be any *more* catastrophic than if it did the same with whatever nonsense
> virt_to_phys() of a vmalloc address had returned.

What do you think about the object_is_on_stack() check? That does a
dereference through "current" to find the stack bounds...

-- 
Kees Cook