Re: [PATCH V2] intel_iommu: refine iotlb hash calculation

[Jason Wang]

On Wed, Apr 12, 2023 at 4:43 PM Alex Bennée <alex.bennee@linaro.org> wrote:
>
>
> Jason Wang <jasowang@redhat.com> writes:
>
> > Commit 1b2b12376c8 ("intel-iommu: PASID support") takes PASID into
> > account when calculating iotlb hash like:
> >
> > static guint vtd_iotlb_hash(gconstpointer v)
> > {
> >     const struct vtd_iotlb_key *key = v;
> >
> >     return key->gfn | ((key->sid) << VTD_IOTLB_SID_SHIFT) |
> >            (key->level) << VTD_IOTLB_LVL_SHIFT |
> >            (key->pasid) << VTD_IOTLB_PASID_SHIFT;
> > }
> >
> > This turns out to be problematic since:
> >
> > - the shift will lose bits if not converting to uint64_t
> > - level should be off by one in order to fit into 2 bits
> > - VTD_IOTLB_PASID_SHIFT is 30 but PASID is 20 bits which will waste
> >   some bits
> > - the hash result is uint64_t so we will lose bits when converting to
> >   guint
> >
> > So this patch fixes them by
> >
> > - converting the keys into uint64_t before doing the shift
> > - off level by one to make it fit into two bits
> > - change the sid, lvl and pasid shift to 26, 42 and 44 in order to
> >   take the full width of uint64_t
> > - perform an XOR to the top 32bit with the bottom 32bit for the final
> >   result to fit guint
> >
> > Fixes: Coverity CID 1508100
> > Fixes: 1b2b12376c8 ("intel-iommu: PASID support")
> > Signed-off-by: Jason Wang <jasowang@redhat.com>
> > ---
> > Changes since V1:
> > - perform XOR to avoid losing bits when converting to gint
> > ---
> >  hw/i386/intel_iommu.c          | 9 +++++----
> >  hw/i386/intel_iommu_internal.h | 6 +++---
> >  2 files changed, 8 insertions(+), 7 deletions(-)
> >
> > diff --git a/hw/i386/intel_iommu.c b/hw/i386/intel_iommu.c
> > index a62896759c..94d52f4205 100644
> > --- a/hw/i386/intel_iommu.c
> > +++ b/hw/i386/intel_iommu.c
> > @@ -64,8 +64,8 @@ struct vtd_as_key {
> >  struct vtd_iotlb_key {
> >      uint64_t gfn;
> >      uint32_t pasid;
> > -    uint32_t level;
> >      uint16_t sid;
> > +    uint8_t level;
> >  };
> >
> >  static void vtd_address_space_refresh_all(IntelIOMMUState *s);
> > @@ -221,10 +221,11 @@ static gboolean vtd_iotlb_equal(gconstpointer v1, 
> > gconstpointer v2)
> >  static guint vtd_iotlb_hash(gconstpointer v)
> >  {
> >      const struct vtd_iotlb_key *key = v;
> > +    uint64_t hash64 = key->gfn | ((uint64_t)(key->sid) << 
> > VTD_IOTLB_SID_SHIFT) |
> > +        (uint64_t)(key->level - 1) << VTD_IOTLB_LVL_SHIFT |
> > +        (uint64_t)(key->pasid) << VTD_IOTLB_PASID_SHIFT;
> >
> > -    return key->gfn | ((key->sid) << VTD_IOTLB_SID_SHIFT) |
> > -           (key->level) << VTD_IOTLB_LVL_SHIFT |
> > -           (key->pasid) << VTD_IOTLB_PASID_SHIFT;
> > +    return (guint)((hash64 >> 32) ^ (hash64 & 0xffffffffU));
>
> Have you measured the distribution this hash gives you? Otherwise
> consider using the qemu_xxhash() functions to return a well distributed
> 32 bit hash value.

It depends on a lot of factors and so it won't be even because the
individuals keys are not evenly distributed:

- gfn depends on guest DMA subsystems
- level depends on when huge pages are used
- pasid depends on whether PASID is being used

I'm ok to switch to use qemu_xxhash() if everyone agree. Or if as
Peter said, if it has been dealt with glibc, maybe it's not worth to
bother.

Thanks

>
> --
> Alex Bennée
> Virtualisation Tech Lead @ Linaro
>