Re: mmap2 is probably still broken in linux-user & other things

[Philippe Mathieu-Daudé]

Cc'ing user-mode team.

On 11/8/20 4:53 PM, Catherine A. Frederick wrote:
> Hi, I submitted a patch a while ago and then dropped off the face of the
> planet like most people do. In my journey to fix DRM/radeonsi in
> user-mode emulation I discovered a few bugs. I don't really have the
> time to write the infrastructure to make wrapping DRM IOCTLs remotely
> not-ugly, but I do want to get patches in for the last bug I encountered
> as it's kinda sneaky.
> 
> Currently(as of 6 months ago, and I doubt anyone noticed) mmap2 is
> completely broken for it's intended use(mapping pages outside the
> address limit). The value of the address input to mmap is clipped to an
> abi_ulong(4 bytes of 32 bit platforms afaik?) and because of this the
> value in pages when converted to addresses and passed to mmap on the
> host is invalid and this breaks radeonsi(and probably any other GPU
> driver in linux-user.
> 
> I don't really know how to fix this in a sane way so I'm looking for ideas.
> 
> Another thing I'm looking for ideas wrt is I'm thinking about is some
> form of automagic(opt-in, through a new type) pointer/union translation
> in IOCTLs over the guest barrier. Currently it's really ugly to wrap
> most of the IOCTLs in DRM and amdgpu because the DRM ioctls frequently
> pass pointers to the kernel and because of this I don't feel comfortable
> merging my patches(manually translating the struct to the types of the
> host sucks and feels like redundant work). On top of that AMDGPU uses a
> lot of structs where currently having one type for the call would
> corrupt the data in the return union.
> 
> Last thing: I never really got any performance benefit out of my
> scheduler, but I think that's mostly from bad implementation. Is there
> any information on TCG's performance characteristics? Thanks.
> 
>