lkml.org 
[lkml]   [2019]   [Dec]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH 0/5] iommu/vt-d: Consolidate various cache flush ops
From
Date
On Fri, 2019-11-22 at 11:04 +0800, Lu Baolu wrote:
> Intel VT-d 3.0 introduces more caches and interfaces for software to
> flush when it runs in the scalable mode. Currently various cache flush
> helpers are scattered around. This consolidates them by putting them in
> the existing iommu_flush structure.
>
> /* struct iommu_flush - Intel IOMMU cache invalidation ops
> *
> * @cc_inv: invalidate context cache
> * @iotlb_inv: Invalidate IOTLB and paging structure caches when software
> * has changed second-level tables.
> * @p_iotlb_inv: Invalidate IOTLB and paging structure caches when software
> * has changed first-level tables.
> * @pc_inv: invalidate pasid cache
> * @dev_tlb_inv: invalidate cached mappings used by requests-without-PASID
> * from the Device-TLB on a endpoint device.
> * @p_dev_tlb_inv: invalidate cached mappings used by requests-with-PASID
> * from the Device-TLB on an endpoint device
> */
> struct iommu_flush {
> void (*cc_inv)(struct intel_iommu *iommu, u16 did,
> u16 sid, u8 fm, u64 type);
> void (*iotlb_inv)(struct intel_iommu *iommu, u16 did, u64 addr,
> unsigned int size_order, u64 type);
> void (*p_iotlb_inv)(struct intel_iommu *iommu, u16 did, u32 pasid,
> u64 addr, unsigned long npages, bool ih);
> void (*pc_inv)(struct intel_iommu *iommu, u16 did, u32 pasid,
> u64 granu);
> void (*dev_tlb_inv)(struct intel_iommu *iommu, u16 sid, u16 pfsid,
> u16 qdep, u64 addr, unsigned int mask);
> void (*p_dev_tlb_inv)(struct intel_iommu *iommu, u16 sid, u16 pfsid,
> u32 pasid, u16 qdep, u64 addr,
> unsigned long npages);
> };
>
> The name of each cache flush ops is defined according to the spec section 6.5
> so that people are easy to look up them in the spec.

Hm, indirect function calls are quite expensive these days.

I would have preferred to go in the opposite direction, since surely
aren't going to have *many* of these implementations. Currently there's
only one for register-based and one for queued invalidation, right?
Even if VT-d 3.0 throws an extra version in, I think I'd prefer to take
out the indirection completely and have an if/then helper.

Would love to see a microbenchmark of unmap operations before and after
this patch series with retpoline enabled, to see the effect.



[unhandled content-type:application/x-pkcs7-signature]
\
 
 \ /
  Last update: 2019-12-03 09:50    [W:0.065 / U:3.104 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site