UMA MD Small Allocator Runtime Switching
nwhitehorn at freebsd.org
Wed Aug 13 14:45:26 UTC 2008
John Baldwin wrote:
> On Tuesday 05 August 2008 05:23:37 am Nathan Whitehorn wrote:
>> I'm working on the PowerPC G5 port right now, and have run into a
>> problem with the way the UMA small allocator works. On G3/G4 systems,
>> there is a direct physical->virtual mapping, and on G5s there isn't. All
>> of the infrastructure is in place to support both types of system with a
>> single kernel image, except that UMA_MD_SMALL_ALLOC must be switched
>> on/off at runtime.
>> One solution is to put if (direct_map) use_nonsmall_case() into the MD
>> small_alloc/free() routines and define UMA_MD_SMALL_ALLOC everywhere.
>> This works well, except that the MI UMA code then sets booted = 1 too
>> early in the boot process, before the kmem_alloc*() routines are available.
>> Basically, I need to find a way have an MD UMA allocator without the MI
>> UMA code assuming anything about how it works internally. Maybe adding a
>> UMA_MD_ALLOC_LATE define to prevent setting booted=1 early on?
> Have you considered creating an artificial direct map region in the address
> space on the G5? Some of the other 64-bit ports (amd64 and sparc64) do this
> to gain the benefits of the direct map even though it isn't a mandated part
> of the architecture like it is on some other platforms (alpha and mips).
I thought about it, but we can only use 4K pages on the G5 so this would
put a large amount of pressure on the page table. IBM removed the block
translation mechanism from the G5 and the CPU's superpage support is not
available in the 32-bit compatibility mode under which we currently run.
More information about the freebsd-arch