Re: [Qemu-devel] softmmu thoughts

[Fabrice Bellard <fabrice@bellard.org>]

Hi,

Here are a few points to look at:

1) MMU_MAP -> SOFT_MMU transition : more work is needed there, but I 
will look at it too, but not in the near future (I think the best 
solution is to recompile the TB directly in the fault handler - the goal 
is to suppress 'HF_SOFTMMU_MASK' which slows down the emulator).

2) Try to make a Windows port. It seems doable because when you create a 
64 KB mapping in Windows you can select which 4 KB subpages are mapped.

3) Test it with the ppc emulation if not already done.

4) Use assembly code in most of the SOFTMMU code to accelerate unaligned 
and I/O accesses (the current code is not optimized).

5) For you and me: reduce the number of ifdefs for SOFTMMU/MMU_MAP and 
mmap().

6) You can go even faster (at least on Linux or *BSD) by using hard mmu 
for pages between 0 and a given address 'L' by using mmap() and by using 
segment limits. You can fall back to MMU_MAP if the address is >= 'L', 
and fall back to soft MMU if I/O accesses are done. The advantage is 
that the 'code copy' mode can be used in that case, so you get closer to 
1:1 performance on most of the user code.

Fabrice.

Piotras wrote:
> Hi!
> 
> I updated the patch to solve problems reported by Jens Arm:
>   * compilation problem on Mandrake 9.1
>   * SEGV problem when working with non-TLS glibc.
> Jens, thanks for your help!
> 
> I also cleaned-up the code and comments in several places (for 
> example page_unprotect).
> 
> The patch should apply cleanly to recent CVS. It should work 
> with 0.6.1 without problems too.
> 
> 
> Regards,
> 
> Piotrek
> 
> On Tue, 14 Dec 2004 18:54:04 +0100, Piotras <piotras@gmail.com> wrote:
> 
>>Hi!
>>
>>I'm attaching a new patch based on your suggestions. On my machine
>>nbench gives:
>>    memory index: 37% up
>>    integer index: 36% up
>>    fp index: 4% up
>>
>>The patch is divided in three files to simplify review. Part 1 contains
>>mostly rearrangement existing code necessary for the patch. Part 2
>>adds the new mmu mode. Part 3 contains heuristic to optimize
>>performance of iomem and self modifying code.
>>
>>To test the patch create a build directory and run:
>> > tar -zxf _PATH_TO_qemu-0.6.1.tar.gz
>> > tar -zxf _PATH_TO_linux-test-0.5.1.tar.gz
>> > cd qemu-0.6.1
>> > ./configure --target-list=i386-softmmu
>> > gunzip < _PATH_TO_mmu-part1.patch.gz | patch -p1
>> > gunzip < _PATH_TO_mmu-part2.patch.gz | patch -p1
>> > gunzip < _PATH_TO_mmu-part3.patch.gz | patch -p1
>> > make
>> > ./i386-softmmu/qemu -m 64 -L pc-bios -hda ../linux-test/linux.img
>>
>>Last but not least. I'd like to acknowledge Magnus contribution -- VM
>>setup code is derived from his work.
>>
>>Regards,
>>
>>Piotrek
>>
>>On Wed, 20 Oct 2004 14:41:42 +0200, Fabrice Bellard <fabrice@bellard.org> wrote:
>>
>>>Hi,
>>>
>>>The idea is interesting. Here are several suggestions:
>>>
>>>- It would be more efficient and simpler to map one 4KB host memory page
>>>every 8 KB. Then you can have a fixed mmap() mapping (no syscall
>>>overhead to change the mappings) and a simple way to handle unaligned
>>>accesses.
>>>- The critical point would be to keep standard soft MMU accesses for
>>>device access. An architectural change is needed to do that, but it
>>>seems easy to add.
>>>- This patch should work with qemu, not qemu-fast. The future of
>>>qemu-fast is to use a kernel module to have near native performances. It
>>>is not worthwhile to invest time in soft MMU or dynamic translation when
>>>you can just execute the code as is !
>>>
>>>Fabrice.