Re: [PATCH] iommu/vt-d: Fix scatterlist offset handling

[Jacob Pan <jacob.jun.pan-VuQAYsv1563Yd54FQh9/CA@public.gmane.org>]

On Fri, 6 Oct 2017 16:43:09 +0200
Joerg Roedel <joro-zLv9SwRftAIdnm+yROfE0A@public.gmane.org> wrote:

> On Tue, Oct 03, 2017 at 07:05:17PM +0100, Robin Murphy wrote:
> > Now, there are indeed plenty of drivers and subsystems which do
> > work on lists of explicitly single pages - anything doing some
> > variant of "addr = kmap_atomic(sg_page(sg)) + sg->offset;" is easy
> > to spot - but I don't think DMA API implementations are in a
> > position to make any kind of assumption; nearly all of them just
> > shut up and handle sg->length bytes from sg_phys(sg) without
> > questioning the caller, and I reckon that's exactly what they
> > should be doing.  
> 
> I agree with that, it is not explicitly forbidden to have an
> sg->offset > PAGE_SIZE and most IOMMU drivers handle this case.
> 
> So this is a problem I'd like to see resolved in the VT-d driver too.
> If nobody comes up with a correct fix soon I'll apply this one and
> rip out the large-page support from __domain_mapping() to make it
> work.
> 
Hi All,

Just to give an update on the offline debugging of this issue. With
Robin's patch applied, I was able to reproduce the failure with
similar configuration that Jain helped to set up.

I added trace prints just to see the map/unmap activities leading to
the DMAR fault. When fault occurs, the trace shows there is an unmap to
the offending iova pfn. So I think this is a separate problem than
Robin's patch is fixing. I think we should move forward to merge this
patch upstream and stable. The remaining problem is likely a race
condition between unmap and DMA activities.

Here a brief extracted log, ee3d7 is the iova pfn in question.
#1. map sg pfn ee3d7
          <idle>-0     [076] 74124.154254: bprint:               __domain_mapping: vpfn:ee3d7, pgoff=2126, np:1, da:ee3d784e, len:1464
, ppfn:1849c9c                                                                                                                        

#2. unmap ee3d7000
         <idle>-0     [054] 74124.154301: bprint:               intel_unmap: Device 0000:18:00.4 unmapping: pfn ee3d7-ee3d7          
         <idle>-0     [076] 74124.154301: bprint:               __domain_mapping: lvlpg:1, nrpg 0, vpfn:ec2ff, ppfn:183221a, sg_res:0
         <idle>-0     [059] 74124.154302: bprint:               __domain_mapping: lvlpg:1, nrpg 0, vpfn:ee719, ppfn:c3e4dd, sg_res:0 
         <idle>-0     [076] 74124.154302: bprint:               __domain_mapping: vpfn:f183b, pgoff=78, np:1, da:f183b04e, len:1464, 

#3. DMA to unmapped address ee3d7000, DMAR fault raised.
  +2.952861] dmar_fault: 6 callbacks suppressed                                                                                   
  +0.000002] DMAR: DRHD: handling fault status reg 2                                                                              
  +0.005588] turning tracing off                                                                                                  
  +0.003592] DMAR: [DMA Write] Request device [18:00.4] fault addr ee3d7000 [fault reason 05] PTE Write access is not set         
                                                                                                                                  
                                                                                                                                                                                                                                                               
         <idle>-0     [000] 74124.156906: bputs:
         0xffffffffb259916bs: turning tracing off     


Thanks,

Jacob

> Speaking of __domain_mapping(), this function is a big unmaintainable
> mess which should be split and rewritten. A clean and maintainable
> rewrite can alse re-add the large-page support.
> 
> 
> Regards,
> 
> 	Joerg
> 
> _______________________________________________
> iommu mailing list
> iommu-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
> https://lists.linuxfoundation.org/mailman/listinfo/iommu

[Jacob Pan]