[Qemu-devel] Are "info pic" and "info irq" still of any use?

[Qemu-devel] Are "info pic" and "info irq" still of any use?

[Markus Armbruster]

These commands look like bit-rotted development aids to me.

They're limited to just a few interrupt controllers.  For the most
common machine types and accelerators, they do nothing.

They complicate David Gibson's work on disentangling dependencies on
ISA.

I'm cc'ing the maintainers of all machines that can be configured in a
way that makes these commands do something.  Please speak up if you
think they provide value.



The gory details:

Only some targets provide "info pic" and "info irq", namely:

#if defined(TARGET_I386) || defined(TARGET_PPC) || defined(TARGET_MIPS) || \
    defined(TARGET_LM32) || (defined(TARGET_SPARC) && !defined(TARGET_SPARC64))

They're HMP-only.  Their documentation in full:

    info pic  -- show i8259 (PIC) state
    info irq  -- show the interrupts statistics (if available)

The first line is a lie: as we'll see, it's not just i8259, but it isn't
all PICs, either.  Not even all i8259.

Three separate implementations:

* sparc-softmmu

  This is sun4m_hmp_info_pic(), sun4m_hmp_info_irq().  Both do nothing
  when !slavio_intctl.  Set by sun4m_hw_init().  Thus, two cases:

  - The sun4m machine types

    These are SS-5, SS-10, SS-600MP, SS-20, Voyager, LX, SS-4,
    SPARCclassic, SPARCbook.  Their board initialization runs
    sun4m_hw_init().

    "info pic" prints selected registers of the "slavio_intctl" device
    model.  This isn't a i8259.

    "info irq" prints "irq statistic code not compiled" unless you
    manually enable DEBUG_IRQ_COUNT to make slavio_set_irq() count IRQs.
    Then it prints IRQ counts.

  - The other machine types

    These are leon3_generic and none.

    Both commands do nothing.

* lm32-softmmu

    This is lm32_hmp_info_pic(), lm32_hmp_info_irq().  Both do nothing
    when !pic.  Set by init() method of "lm32-pic".  Thus, two cases:

    - Machines sporting an "lm32-pic"

      These are lm32-evr, lm32-uclinux, milkymist.

      "info pic" prints selected registers of the "lm32-pic" device
      model realized last.  Fortunately, all machines have just one.
      This isn't a i8259.

      "info irq" prints its IRQ counts.

    - Other machine types

      This is just none right now.

      Both commands do nothing.

* {i386,mips,mips64,mips64el,mipsel,ppc,ppc64,ppcemb,x86_64}-softmmu

  This is hmp_info_pic(), hmp_info_irq().

  hmp_info_pic() does nothing when !isa_pic.  Set by the i8259_init()
  helper that creates a master/slave pair of the "isa-i8259" device
  model.  Cases:

  - Machines using the helper

    These are
        alpha clipper
        i386 & x86_64 pc*, isapc when accel=tcg or kernel_irqchip=off
        mips* magnum, pica64, malta, mips
        mips64el fulong2e
        ppc & ppc64 prep

    "info pic" prints selected registers of the last "isa-i8259"
    master/slave pair created with i8259_init().  I can't see a machine
    creating more than one.

    For the pc machine types, the pair of i8259s is just a piece of
    legacy crap hanging off LAPIC or I/O APIC.  The commands are thus of
    dubious utility there.

    "info irq" prints "irq statistic code not compiled" unless you
    manually enable DEBUG_IRQ_COUNT to make pic_set_irq() count IRQs.
    Then it prints IRQ counts for *all* "isa-i8259", whether they're
    created with i8259_init() or not.  Looks like all are.

  - Machines not using the helper

    These include i386 & x86_64 pc*, isapc when accel=kvm and
    kernel_irqchip=on (this is the default).

    Both commands do nothing.  Even when the machine has in fact a pair
    of i8259 implemented by device model "kvm-i8259".

Re: [Qemu-devel] Are "info pic" and "info irq" still of any use?

[Paolo Bonzini]

On 31/03/2015 11:56, Markus Armbruster wrote:
> These commands look like bit-rotted development aids to me.
> 
> They're limited to just a few interrupt controllers.  For the most
> common machine types and accelerators, they do nothing.
> 
> They complicate David Gibson's work on disentangling dependencies on
> ISA.
> 
> I'm cc'ing the maintainers of all machines that can be configured in a
> way that makes these commands do something.  Please speak up if you
> think they provide value.

For x86, they can be removed.  However, it would be nice to provide
access to the same information via QOM, especially for "info pic".  This
is the old QIDL project.

Paolo

Re: [Qemu-devel] Are "info pic" and "info irq" still of any use?

[Michael Walle]

Am 2015-03-31 11:56, schrieb Markus Armbruster:
[..]
> * lm32-softmmu
> 
>     This is lm32_hmp_info_pic(), lm32_hmp_info_irq().  Both do nothing
>     when !pic.  Set by init() method of "lm32-pic".  Thus, two cases:
> 
>     - Machines sporting an "lm32-pic"
> 
>       These are lm32-evr, lm32-uclinux, milkymist.
> 
>       "info pic" prints selected registers of the "lm32-pic" device
>       model realized last.  Fortunately, all machines have just one.
>       This isn't a i8259.
> 
>       "info irq" prints its IRQ counts.
> 
>     - Other machine types
> 
>       This is just none right now.
> 
>       Both commands do nothing.

Can be removed. Is there another way to make these information visible? 
Esp. the irq counters.

-michael

Re: [Qemu-devel] Are "info pic" and "info irq" still of any use?

[Markus Armbruster]

Paolo Bonzini <pbonzini@redhat.com> writes:

> On 31/03/2015 11:56, Markus Armbruster wrote:
>> These commands look like bit-rotted development aids to me.
>> 
>> They're limited to just a few interrupt controllers.  For the most
>> common machine types and accelerators, they do nothing.
>> 
>> They complicate David Gibson's work on disentangling dependencies on
>> ISA.
>> 
>> I'm cc'ing the maintainers of all machines that can be configured in a
>> way that makes these commands do something.  Please speak up if you
>> think they provide value.
>
> For x86, they can be removed.  However, it would be nice to provide
> access to the same information via QOM, especially for "info pic".  This
> is the old QIDL project.

I guess you mean introspecting device state via QOM, thus QMP.  Would be
nice indeed, and much more useful than on-off hacks like "info pic".

Should be no harder than serializing device state, which we already do
for migration.  At least in theory.  In practice, ...

Re: [Qemu-devel] Are "info pic" and "info irq" still of any use?

[Kevin Wolf]

Am 31.03.2015 um 11:56 hat Markus Armbruster geschrieben:
> These commands look like bit-rotted development aids to me.
> 
> They're limited to just a few interrupt controllers.  For the most
> common machine types and accelerators, they do nothing.
> 
> They complicate David Gibson's work on disentangling dependencies on
> ISA.
> 
> I'm cc'ing the maintainers of all machines that can be configured in a
> way that makes these commands do something.  Please speak up if you
> think they provide value.

I've used 'info pic' before (on i386) and I think it has helped me a bit
in debugging a guest. Definitely not mainstream usage, but there is some
value.

It's not specifically this feature, but in my experience TCG guests are
much friendlier for debugging (or I just don't know how to do it right
for KVM guests - which would mean that for TCG the features are more
discoverable and therefore friendlier again), so I actively avoid KVM
when doing that.

I would appreciate if the disparity would be solved by KVM guests
getting friendlier instead of TCG guests becoming harder to debug.

That said, if there are good replacements, there is nothing that should
stop us from removing old HMP commands when they are worse than newer
commands.

Kevin