Re: [Qemu-devel] [PATCH v3 6/9] target-arm: support QMP dump-guest-memory

[Andrew Jones <drjones@redhat.com>]

On Fri, Dec 18, 2015 at 11:59:39AM +0000, Peter Maydell wrote:
> On 15 December 2015 at 22:51, Andrew Jones <drjones@redhat.com> wrote:
> > Add the support needed for creating prstatus elf notes. This
> > allows us to use QMP dump-guest-memory.
> >
> > Signed-off-by: Andrew Jones <drjones@redhat.com>
> 
> > +int arm_cpu_write_elf64_note(WriteCoreDumpFunction f, CPUState *cs,
> > +                             int cpuid, void *opaque)
> > +{
> > +    struct aarch64_note note;
> > +    CPUARMState *env = &ARM_CPU(cs)->env;
> > +    DumpState *s = opaque;
> > +    uint64_t pstate, sp;
> > +    int ret, i;
> > +
> > +    aarch64_note_init(&note, s, "CORE", 5, NT_PRSTATUS, sizeof(note.prstatus));
> > +
> > +    note.prstatus.pr_pid = cpu_to_dump32(s, cpuid);
> > +
> > +    if (!is_a64(env)) {
> > +        aarch64_sync_32_to_64(env);
> > +        pstate = cpsr_read(env);
> > +        sp = aarch64_compat_sp(env);
> 
> I don't understand why we need to do this. If this is an
> AArch64 dump then we should just treat it as an AArch64
> dump, and presumably the consumer of the dump knows enough
> to know what the "hypervisor view" of a CPU that's currently
> in 32-bit mode is. It has to anyway to be able to figure
> out where all the other registers are, so why can't it
> also figure out what mode the CPU is currently in and thus
> where r13 is in the xregs array?

You're probably right that this shouldn't be necessary. But, in order for
it not to be necessary, I'll need to write another crash patch. Currently,
if you do a dump-guest-memory on a running guest, i.e. one where the kernel
has not called panic(), and thus the cpus are actually in 32-bit usermode,
rather than in the 64-bit cpu-stop IPI handler, then the crash utility
segfaults if sp == xregs[31]. crash does properly decode the registers
it digs out of the stack frame on a panic'ed cpu though, and setting sp
to aarch64_compat_sp here also allows crash to work properly in the non-
panic'd case.

So, I could teach crash to do what I'm doing here in qemu instead, but
there's still one more reason why it may make sense to do it here. That
reason is that I don't know what else to put in prstatus.pr_reg.sp. Does
xregs[31] make the most sense? or just zero? prstatus.pr_reg.pc is the
correct 32-bit userspace pc, prstatus.pr_reg.pstate is the correct cpsr,
so why not set sp to the correct userspace sp?

Thanks,
drew

> 
> > +    } else {
> > +        pstate = pstate_read(env);
> > +        sp = env->xregs[31];
> > +    }
> 
> thanks
> -- PMM
>