[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH V2] intel_iommu: refine iotlb hash calculation
From: |
Alex Bennée |
Subject: |
Re: [PATCH V2] intel_iommu: refine iotlb hash calculation |
Date: |
Wed, 12 Apr 2023 09:41:56 +0100 |
User-agent: |
mu4e 1.10.0; emacs 29.0.90 |
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.
--
Alex Bennée
Virtualisation Tech Lead @ Linaro