At the moment we allocate the entire TCE table, twice (hardware part and
userspace translation cache). This normally works as we normally have
contigous memory and the guest will map entire RAM for 64bit DMA.

However if we have sparse RAM (one example is a memory device), then
we will allocate TCEs which will never be used as the guest only maps
actual memory for DMA. If it is a single level TCE table, there is nothing
we can really do but if it a multilevel table, we can skip allocating
TCEs we know we won't need.

This adds ability to allocate only first level, saving memory.

This changes iommu_table::free() to avoid allocating of an extra level;
iommu_table::set() will do this when needed.

This adds @alloc parameter to iommu_table::exchange() to tell the callback
if it can allocate an extra level; the flag is set to "false" for
the realmode KVM handlers of H_PUT_TCE hcalls and the callback returns
H_TOO_HARD.

This still requires the entire table to be counted in mm::locked_vm.

To be conservative, this only does on-demand allocation when
the usespace cache table is requested which is the case of VFIO.

The example math for a system replicating a powernv setup with NVLink2
in a guest:
16GB RAM mapped at 0x0
128GB GPU RAM window (16GB of actual RAM) mapped at 0x244000000000

the table to cover that all with 64K pages takes:
(((0x244000000000 + 0x2000000000) >> 16)*8)>>20 = 4556MB

If we allocate only necessary TCE levels, we will only need:
(((0x400000000 + 0x400000000) >> 16)*8)>>20 = 4MB (plus some for indirect
levels).
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>