On Mon, Nov 04, 2013 at 12:22:35PM +0100, Paolo Bonzini wrote: > Il 04/11/2013 12:14, Michael S. Tsirkin ha scritto: > >> > > >> > This patch looks good; however, on top of it can you test > >> > kvm-unit-tests with TARGET_PHYS_ADDR_SPACE_BITS=64 and see whether > >> > there is a measurable slowdown (in the inl_from_qemu tests)? If not, > >> > we can just get rid of TARGET_PHYS_ADDR_SPACE_BITS in exec.c. > > I'd rather we fixed a bug first - we need to fix it on stable too - any > > cleanups can come on top. > > This is not necessarily a cleanup. Getting rid of > TARGET_PHYS_ADDR_SPACE_BITS in exec.c means fixing device-to-device DMA > bugs for example. > > Of course a smaller patch can be done that avoids the renaming of L2_* > constants. > > > Also, I'm not sure what will this test tell > > us: inl reads io space, not memory, right? > > The number of levels in the dispatch radix tree is independent of the > size of the AddressSpace; it is P_L2_LEVELS for both the 64K io space > and the 2^TARGET_PHYS_ADDRESS_SPACE_BITS memory space. > > Paolo
Hmm I think it's *at most* that deep but can be more shallow, no? -- MST