Re: [Qemu-devel] [RFC PATCH 00/11] speedup vring processing with MemoryRegionCaches

[Stefan Hajnoczi]

On Mon, Dec 12, 2016 at 12:18:46PM +0100, Paolo Bonzini wrote:
> It is known that virtio's usage of ld_*_phys and st_*_phys functions
> wastes time in address_space_translate, visiting the
> AddressSpaceDispatch's radix tree.
> 
> This series introduces a small cache that changes these functions
> to a simple range check and memory access.  The effect is a bit
> underwhelming, because the improvement is only 1-2Kiops/second.
> Nevertheless I'm throwing out the patches so that for example they
> can be tested on s390.
> 
> Things to fix: handle address_space_cache_init failures
> in virtio_init_region_cache.  Also, once virtio breaks free of
> address_space_memoory, we'll need to handle invalidation in IOMMU regions.
> For the latter, maybe it's worth introducing a new abstraction that
> is higher-level than MemoryListener and covers both regular and IOMMU
> memory regions.

I think this new API will be important for getting the best
performance in emulated devices.

Stefan

signature.asc
Description: PGP signature