[Qemu-devel] qemu ARM host support likely to be broken in 1.0

[Qemu-devel] qemu ARM host support likely to be broken in 1.0

[Peter Maydell]

This is just a quick email to summarise a discussion on IRC.

QEMU on ARM hosts (not ARM guests!) is currently broken in
at least the following ways:
 * configure will select the ucontext coroutine implementation
but on ARM makecontext returns ENOSYS and we will abort() on
startup
 * the fixed address we mmap the code gen buffer to in system
mode tends to overlap the libc heap (at least on Ubuntu oneiric);
when this happens you typically get a sysmalloc abort
 * there is a bug I have not investigated which causes an
i386 guest image to loop infinitely resetting before it brings
up the grub menu
 * the TCG_AREG0 is r7, which clashes with the Thumb frame
pointer, with the effect that --enable-debug fails to compile

ARM host support for TCG is not a very high priority for me and
I don't expect to fix any of these before 1.0 is released. I'm
happy to test patches if anybody else wishes to write them :-)

-- PMM

Re: [Qemu-devel] qemu ARM host support likely to be broken in 1.0

[Paolo Bonzini]

On 11/04/2011 02:52 AM, Peter Maydell wrote:
>   * the fixed address we mmap the code gen buffer to in system
> mode tends to overlap the libc heap (at least on Ubuntu oneiric);
> when this happens you typically get a sysmalloc abort

Let's change this to use the .bss code gen buffer.  We can make it 
depend on __arm__ for 1.0, and toggle it for all architectures for 1.1 
after testing on Windows (Linux and Mac OS X should have no problem).

Paolo

Re: [Qemu-devel] qemu ARM host support likely to be broken in 1.0

[Peter Maydell]

On 4 November 2011 07:53, Paolo Bonzini <pbonzini@redhat.com> wrote:
> On 11/04/2011 02:52 AM, Peter Maydell wrote:
>>
>>  * the fixed address we mmap the code gen buffer to in system
>> mode tends to overlap the libc heap (at least on Ubuntu oneiric);
>> when this happens you typically get a sysmalloc abort
>
> Let's change this to use the .bss code gen buffer

Does anybody know why we don't just use the .bss codegen
buffer all the time for everything anyway?

-- PMM