Re: [PATCH v1 5/8] cpus-common: ensure auto-assigned cpu_indexes don't clash

[Igor Mammedov]

On Thu, 14 May 2020 17:27:53 +0100
Alex Bennée <address@hidden> wrote:

> a
> Alex Bennée <address@hidden> writes:
> 
> > Basing the cpu_index on the number of currently allocated vCPUs fails
> > when vCPUs aren't removed in a LIFO manner. This is especially true
> > when we are allocating a cpu_index for each guest thread in
> > linux-user where there is no ordering constraint on their allocation
> > and de-allocation.
> >
> > [I've dropped the assert which is there to guard against out-of-order
> > removal as this should probably be caught higher up the stack. Maybe
> > we could just ifdef CONFIG_SOFTTMU it?]

for machines where we care about cross version migration 
(arm/virt,s390,x86,spapr),
we do manual cpu_index assignment on keep control on its stability
So orderining probably shouldn't matter for other softmmu boards,
but what I'd watch for is arrays within devices where cpu_index is used as index
(ex: would be apic emulation (but its not affected by this patch since x86 
control
cpu_index assignment))


> >
> > Signed-off-by: Alex Bennée <address@hidden>
> > Cc: Nikolay Igotti <address@hidden>
> > Cc: Paolo Bonzini <address@hidden>
> > Cc: Igor Mammedov <address@hidden>
> > Cc: Eduardo Habkost <address@hidden>
> > ---
> >  cpus-common.c | 9 ++++-----
> >  1 file changed, 4 insertions(+), 5 deletions(-)
> >
> > diff --git a/cpus-common.c b/cpus-common.c
> > index 55d5df89237..5a7d2f6132b 100644
> > --- a/cpus-common.c
> > +++ b/cpus-common.c
> > @@ -61,13 +61,14 @@ static bool cpu_index_auto_assigned;
> >  static int cpu_get_free_index(void)
> >  {
> >      CPUState *some_cpu;
> > -    int cpu_index = 0;
> > +    int max_cpu_index = 0;
> >  
> >      cpu_index_auto_assigned = true;
> >      CPU_FOREACH(some_cpu) {
> > -        cpu_index++;
> > +        max_cpu_index = MAX(some_cpu->cpu_index, max_cpu_index);
> >      }
> > -    return cpu_index;
> > +    max_cpu_index++;
> > +    return max_cpu_index;
> >  }  
> 
> OK some ending up with cpu_index = 1 threw off devices that would do
> qemu_get_cpu(0) so I've tweaked the algorithm to:
> 
>   static int cpu_get_free_index(void)
>   {
>       CPUState *some_cpu;
>       int max_cpu_index = 0;
> 
>       cpu_index_auto_assigned = true;
>       CPU_FOREACH(some_cpu) {
>           if (some_cpu->cpu_index >= max_cpu_index) {
>               max_cpu_index = some_cpu->cpu_index + 1;
>           }
>       }
>       return max_cpu_index;
>   }
> 
> >  
> >  void cpu_list_add(CPUState *cpu)
> > @@ -90,8 +91,6 @@ void cpu_list_remove(CPUState *cpu)
> >          return;
> >      }
> >  
> > -    assert(!(cpu_index_auto_assigned && cpu != QTAILQ_LAST(&cpus)));
> > -
> >      QTAILQ_REMOVE_RCU(&cpus, cpu, node);
> >      cpu->cpu_index = UNASSIGNED_CPU_INDEX;
> >  }  
> 
>