Re: new text patching for review

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Zachary Amsden (zach@vmware.com) wrote:
> Mathieu Desnoyers wrote:
> >Yes, kprobes is case 1: atomic update. And we don't even have to bother
> >about Intel's erratum. This one is ok. That's mainly the
> >alternatives/paravirt code I worry about.
> >  
> 
> Paravirt and alternatives should all be ok because they are done before 
> SMP bringup and with NMIs disabled.  NMI watchdog is not setup until 
> smp_prepare_cpus/check_nmi_watchdog, which happens way later, not during 
> parse_args/setup_nmi_watchdog, which just decides which type of watchdog 
> to setup.
> 

I'm not so sure about this. You are right in that it has nothing to do
with parse_args, but I just went in detail through the source, and the
order seems to be:

1 - NMI is activated
2 - MCE is activated
3 - alternatives are applied

In detail:

start_kernel()
  ...
  include/asm-i386/smp.h:smp_prepare_boot_cpu()
    smp_ops.smp_prepare_boot_cpu()
       arch/i386/kernel/smpboot.c:native_smp_prepare_boot_cpu()
        arch/i386/kernel/smpboot.c:native_smp_prepare_cpus()
          arch/i386/kernel/smpboot.c:smp_boot_cpus()
            arch/i386/kernel/apic.c:setup_local_APIC()
              arch/i386/kernel/nmi.c:setup_apic_nmi_watchdog()
                arch/i386/kernel/cpu/perfctr-watchdog.c:lapic_watchdog_init()
                  wd_ops->setup() (e.g. setup_intel_arch_watchdog)
  ...
  arch/i386/kernel/cpu/bugs.c:check_bugs()
    arch/i386/kernel/cpu/common.c:identify_boot_cpu()
      identify_cpu()
        arch/i386/kernel/cpu/mcheck/mce.c:mcheck_init()
          arch/i386/kernel/cpu/mcheck/p4.c:intel_p4_mcheck_init()
    ...
    arch/i386/kernel/alternative_instructions()
  ...

Therefore, applying alternatives instructions seems to be done after NMI
and MCE init at boot time, or am I misunderstanding something ?


> I originally considered the NMI problem for paravirt-ops patching done 
> during module load, and found that I would need to modify 
> stop_machine_run to have an architecture specific callout to mask and 
> unmask NMIs.  I didn't imagine that would be very popular, and VMI was 
> the only paravirt-ops that were considering module load time patching, 
> so I flushed it.
> 
> You get some other nasty issues as well with run-time switching, like 
> missing early init calls (in particular, we would have to go to some 
> heroics to retake the land surrounding the APIC local timer interrupt).
> 
> Zach

-- 
Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68