Re: [RFC PATCH v4 19/28] swiotlb: Add warnings for use of bounce buffers with SME

[Tom Lendacky]

On 2/27/2017 11:52 AM, Borislav Petkov wrote:
> On Thu, Feb 16, 2017 at 09:46:19AM -0600, Tom Lendacky wrote:
> > Add warnings to let the user know when bounce buffers are being used for
> > DMA when SME is active.  Since the bounce buffers are not in encrypted
> > memory, these notifications are to allow the user to determine some
> > appropriate action - if necessary.
> >
> > Signed-off-by: Tom Lendacky <thomas.lendacky@xxxxxxx>
> > ---
> >  arch/x86/include/asm/mem_encrypt.h |   11 +++++++++++
> >  include/linux/dma-mapping.h        |   11 +++++++++++
> >  include/linux/mem_encrypt.h        |    6 ++++++
> >  lib/swiotlb.c                      |    3 +++
> >  4 files changed, 31 insertions(+)
> >
> > diff --git a/arch/x86/include/asm/mem_encrypt.h b/arch/x86/include/asm/mem_encrypt.h
> > index 87e816f..5a17f1b 100644
> > --- a/arch/x86/include/asm/mem_encrypt.h
> > +++ b/arch/x86/include/asm/mem_encrypt.h
> > @@ -26,6 +26,11 @@ static inline bool sme_active(void)
> >  	return (sme_me_mask) ? true : false;
> >  }
> >
> > +static inline u64 sme_dma_mask(void)
> > +{
> > +	return ((u64)sme_me_mask << 1) - 1;
> > +}
> > +
> >  void __init sme_early_encrypt(resource_size_t paddr,
> >  			      unsigned long size);
> >  void __init sme_early_decrypt(resource_size_t paddr,
> > @@ -53,6 +58,12 @@ static inline bool sme_active(void)
> >  {
> >  	return false;
> >  }
> > +
> > +static inline u64 sme_dma_mask(void)
> > +{
> > +	return 0ULL;
> > +}
> > +
> >  #endif
> >
> >  static inline void __init sme_early_encrypt(resource_size_t paddr,
> > diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h
> > index 10c5a17..130bef7 100644
> > --- a/include/linux/dma-mapping.h
> > +++ b/include/linux/dma-mapping.h
> > @@ -10,6 +10,7 @@
> >  #include <linux/scatterlist.h>
> >  #include <linux/kmemcheck.h>
> >  #include <linux/bug.h>
> > +#include <linux/mem_encrypt.h>
> >
> >  /**
> >   * List of possible attributes associated with a DMA mapping. The semantics
> > @@ -557,6 +558,11 @@ static inline int dma_set_mask(struct device *dev, u64 mask)
> >
> >  	if (!dev->dma_mask || !dma_supported(dev, mask))
> >  		return -EIO;
> > +
> > +	if (sme_active() && (mask < sme_dma_mask()))
> > +		dev_warn(dev,
> > +			 "SME is active, device will require DMA bounce buffers\n");
> > +
>
> Yes, definitely _once() here.

Setting the mask is a probe/init type event, so I think not having the
_once() would be better so that all devices that set a mask to something
less than the SME encryption mask would be identified.  This isn't done
for every DMA, etc.

> It could be extended later to be per-device if the need arises.
>
> Also, a bit above in this function, we test if (ops->set_dma_mask) so
> device drivers which supply even an empty ->set_dma_mask will circumvent
> this check.
>
> It probably doesn't matter all that much right now because the
> only driver I see right now defining this method, though, is
> ethernet/intel/fm10k/fm10k_pf.c and some other arches' functionality
> which is unrelated here.

Device drivers don't supply set_dma_mask() since that is part of the
dma_map_ops structure. The fm10k_pf.c file function is unrelated to this
(it's part of an internal driver structure). The dma_map_ops structure
is setup by the arch or an iommu.

Thanks,
Tom

> But still...