Re: [PATCH v2 3/3] hw/386: Fix uninitialized memory with -device and CPU hotplug

[Igor Mammedov]

On Wed, 01 Jul 2020 12:31:15 -0500
Babu Moger <babu.moger@amd.com> wrote:

> Noticed the following command failure while testing CPU hotplug.
> 
> $ qemu-system-x86_64 -machine q35,accel=kvm -smp 1,maxcpus=2,
>   cores=1, threads=1,sockets=2 -cpu EPYC -device EPYC-x86_64-
>   cpu,core-id=0,socket-id=1,thread-id=0
> 
>   qemu-system-x86_64: -device EPYC-x86_64-cpu,core-id=0,socket-id=1,
>   thread-id=0: Invalid CPU [socket: 21855, die: 0, core: 0, thread: 0]
>   with APIC ID 21855, valid index range 0:1
> 
> This happens because APIC ID is calculated using uninitialized memory.
> This is happening after the addition of new field node_id in X86CPUTopoIDs
> structure. The node_id field is uninitialized while calling
> apicid_from_topo_ids. The problem is discussed in the thread below.
> https://lore.kernel.org/qemu-devel/20200602171838.GG577771@habkost.net/
> 
> Fix the problem by initializing the node_id from the device being added.
> 
> Fixes:
> Link: https://bugzilla.redhat.com/show_bug.cgi?id=1828750
> 
> Signed-off-by: Babu Moger <babu.moger@amd.com>
> ---
>  hw/i386/pc.c |   10 ++++++++++
>  1 file changed, 10 insertions(+)
> 
> diff --git a/hw/i386/pc.c b/hw/i386/pc.c
> index e613b2299f..aa9fb48834 100644
> --- a/hw/i386/pc.c
> +++ b/hw/i386/pc.c
> @@ -1553,6 +1553,15 @@ static void pc_cpu_pre_plug(HotplugHandler 
> *hotplug_dev,
>              cpu->die_id = 0;
>          }
>  
so this is from 
 'if (cpu->apic_id == UNASSIGNED_APIC_ID) {'
branch, meaning cpu comes from -device/device_add

> +        /*
> +         * If node_id is not set, initialize it to zero for now. If the user
> +         * does not pass the correct node in case of numa configuration, it
> +         * will be rejected eventually.
> +         */
> +        if (cpu->node_id < 0) {
which means that user hasn't provided 'node-id',
in which case we should error out asking for specifying NUMA node-id along with 
other options

(1)
However that's not enough since by contract query-hotpluggbale-cpus shall 
provide all attributes
necessary to hotplug CPU, which makes node-id is not an optional in case of 
EPYC cpu.
So we need to initialize ms->possible_cpus->cpus[i].props.[has_]node_id by the 
time
we start creating CPUs.

here are 2 variants:
(2)
  * single node:
      nodes_per_pkg 1 and ms->smp.sockets == 1

    since it's the only node and mapping of RAM/CPU is unambigiuos,
    we can deal with it by moving auto_enable_numa into MachineState
    and enabling it in case EPYC CPU is used

  * multiple nodes:
      - ms->smp.sockets > 1
      - nodes_per_pkg > 1
    we can't make up NUMA nodes automatically, and have to ask user to use 
-numa options
    to provide nodes along with CPU/RAM mapping. So in case NUMA wasn't 
configured
    explicitly, we can only error out. (that also applies to CPU created 
implicitly by board '-smp X')

(3) Once user supplied mapping we need to checks that it matches EPYC topology,


(4) As for CPUID, current code in CPUID_Fn8000001E_ECX
      if (nodes <= 4) { /* here nodes is nodes_per_pkg */
         / goes by stricly spec /
         *ecx = ((nodes - 1) << 8) | (topo_ids.pkg_id << 2) | topo_ids.node_id;
         /* makes up system wide NUMA node IDs which happen to match system wide
            NUMA node IDs created by -numa, when -smp + -numa produces 
nodes_per_pkg is in that range,
            basically user has no idea when this happens
          */
      } else {
         /* makeup new algorithm system wide NUMA node IDs generation for out 
of spec behaviour */
      }
     
    problem with both branches is that might lead to inconsistentcy between 
system wide
    NUMA node id in CPUID_Fn8000001E_ECX and the one configured with -numa 
which goes to
    SRAT ACPI table and should go to CPU::node-id property.

    Considering that out of spec behaviour is allowed we probably schould 
replace both branches
    with
       *ecx = ((nodes - 1) << 8) | cpu->node_id;
    which ensures consistency of system wide NUMA node ids and add checks for 
max nodes/max node id.

checks could be done early in cpu's realize() function.



> +            cpu->node_id = 0;
> +        }

>          if (cpu->socket_id < 0) {
>              error_setg(errp, "CPU socket-id is not set");
>              return;
> @@ -1587,6 +1596,7 @@ static void pc_cpu_pre_plug(HotplugHandler *hotplug_dev,
>          }
>  
>          topo_ids.pkg_id = cpu->socket_id;
> +        topo_ids.node_id = cpu->node_id;
>          topo_ids.die_id = cpu->die_id;
>          topo_ids.core_id = cpu->core_id;
>          topo_ids.smt_id = cpu->thread_id;
>