Re: [Qemu-devel] KVM/QEMU: Support executing from flash device

[Avi Kivity]

On 03/04/2012 08:56 AM, Jordan Justen wrote:
> On Tue, Oct 25, 2011 at 00:47, Avi Kivity <address@hidden> wrote:
> > The core issue that kvm (the kernel part) supports two styles of memory:
> > read/write RAM, and read/write MMIO.  ROM wants writes to be ignored,
> > and rom/device wants reads serviced from memory and writes serviced by
> > userspace (as MMIO).
> >
> > It should not be too hard to patch kvm to support these additional two
> > styles.  The entry point is the KVM_SET_USER_MEMORY_REGION ioctl to
> > define the new attributes for the region, and kvm_mmu_page_fault() to
> > map these pages as read-only and emulate writes (for ROM/device regions).
>
> Additional context for CC'd kvm-devel: I would like to support
> executing from an emulated flash memory on kvm.  The flash memory
> would be a CFI (Firmware Hub-like) device which is programmed via MMIO
> at the same addresses in memory where the execution would occur.
>
> What is needed is:
> 1. The ability to set a region to allow read/exec,
>    but trap to qemu on writes.
> 2. qemu should then be able to set the region
>    to trap on reads/exec/writes.
> 3. qemu should be able to then restore the region
>    to the initial state.
>
> Here is my plan:
> 1. Add KVM_CAP_REGION_WRITE_TRAP
> 2. Add kvm_memory_region::flags
>    KVM_MEMSLOT_TRAP_WRITES
> 3. Update page table creation to trap on writes when
>    KVM_MEMSLOT_TRAP_WRITES is set
>
> Is this plan heading in the right direction?

Sure.  I'd name things differently, call this type of slot readonly
instead of write trapping (that is, say what works in kvm (reads)
instead of what doesn't (writes).

Note you need to update kvm_userspace_memory_region, not
kvm_memory_region, which has been deprecated since 2.6.24 or so.

> Regarding emulating ROM (as you mentioned above): I think that it can
> trap to qemu and QEMU can ignore it.  This would be a performance hit,
> but I don't think it is expected that lots of writes to a ROM will
> occur.  Do you agree?

I do, I don't foresee any problems in this area.

Only potential issue is vapic writing to a ROM region, but Jan's recent
work adds a hackaround for that (vapic_map_rom_writable()).

-- 
error compiling committee.c: too many arguments to function