Re: [PATCH v3 1/3] numa: Validate cluster and NUMA node boundary if required

[Gavin Shan <gshan@redhat.com>]

On 3/13/23 7:40 PM, Philippe Mathieu-Daudé wrote:
> On 25/2/23 07:35, Gavin Shan wrote:
>> For some architectures like ARM64, multiple CPUs in one cluster can be
>> associated with different NUMA nodes, which is irregular configuration
>> because we shouldn't have this in baremetal environment. The irregular
>> configuration causes Linux guest to misbehave, as the following warning
>> messages indicate.
>>
>>    -smp 6,maxcpus=6,sockets=2,clusters=1,cores=3,threads=1 \
>>    -numa node,nodeid=0,cpus=0-1,memdev=ram0                \
>>    -numa node,nodeid=1,cpus=2-3,memdev=ram1                \
>>    -numa node,nodeid=2,cpus=4-5,memdev=ram2                \
>>
>>    ------------[ cut here ]------------
>>    WARNING: CPU: 0 PID: 1 at kernel/sched/topology.c:2271 build_sched_domains+0x284/0x910
>>    Modules linked in:
>>    CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.14.0-268.el9.aarch64 #1
>>    pstate: 00400005 (nzcv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>>    pc : build_sched_domains+0x284/0x910
>>    lr : build_sched_domains+0x184/0x910
>>    sp : ffff80000804bd50
>>    x29: ffff80000804bd50 x28: 0000000000000002 x27: 0000000000000000
>>    x26: ffff800009cf9a80 x25: 0000000000000000 x24: ffff800009cbf840
>>    x23: ffff000080325000 x22: ffff0000005df800 x21: ffff80000a4ce508
>>    x20: 0000000000000000 x19: ffff000080324440 x18: 0000000000000014
>>    x17: 00000000388925c0 x16: 000000005386a066 x15: 000000009c10cc2e
>>    x14: 00000000000001c0 x13: 0000000000000001 x12: ffff00007fffb1a0
>>    x11: ffff00007fffb180 x10: ffff80000a4ce508 x9 : 0000000000000041
>>    x8 : ffff80000a4ce500 x7 : ffff80000a4cf920 x6 : 0000000000000001
>>    x5 : 0000000000000001 x4 : 0000000000000007 x3 : 0000000000000002
>>    x2 : 0000000000001000 x1 : ffff80000a4cf928 x0 : 0000000000000001
>>    Call trace:
>>     build_sched_domains+0x284/0x910
>>     sched_init_domains+0xac/0xe0
>>     sched_init_smp+0x48/0xc8
>>     kernel_init_freeable+0x140/0x1ac
>>     kernel_init+0x28/0x140
>>     ret_from_fork+0x10/0x20
>>
>> Improve the situation to warn when multiple CPUs in one cluster have
>> been associated with different NUMA nodes. However, one NUMA node is
>> allowed to be associated with different clusters.
>>
>> Signed-off-by: Gavin Shan <gshan@redhat.com>
>> ---
>>   hw/core/machine.c   | 42 ++++++++++++++++++++++++++++++++++++++++++
>>   include/hw/boards.h |  1 +
>>   2 files changed, 43 insertions(+)
>>
>> diff --git a/hw/core/machine.c b/hw/core/machine.c
>> index f29e700ee4..3513df5a86 100644
>> --- a/hw/core/machine.c
>> +++ b/hw/core/machine.c
>> @@ -1252,6 +1252,45 @@ static void machine_numa_finish_cpu_init(MachineState *machine)
>>       g_string_free(s, true);
>>   }
>> +static void validate_cpu_cluster_to_numa_boundary(MachineState *ms)
>> +{
>> +    MachineClass *mc = MACHINE_GET_CLASS(ms);
>> +    NumaState *state = ms->numa_state;
>> +    const CPUArchIdList *possible_cpus = mc->possible_cpu_arch_ids(ms);
>> +    const CPUArchId *cpus = possible_cpus->cpus;
>> +    int len = possible_cpus->len, i, j;
> 
> (Nitpicking, 'len' variable is not very useful).
> 

Yes, Lets drop it if I need to post a new revision :)

>> +
>> +    if (state->num_nodes <= 1 || len <= 1) {
>> +        return;
>> +    }
>> +
>> +    /*
>> +     * The Linux scheduling domain can't be parsed when the multiple CPUs
>> +     * in one cluster have been associated with different NUMA nodes. However,
>> +     * it's fine to associate one NUMA node with CPUs in different clusters.
>> +     */
>> +    for (i = 0; i < len; i++) {
>> +        for (j = i + 1; j < len; j++) {
>> +            if (cpus[i].props.has_socket_id &&
>> +                cpus[i].props.has_cluster_id &&
>> +                cpus[i].props.has_node_id &&
>> +                cpus[j].props.has_socket_id &&
>> +                cpus[j].props.has_cluster_id &&
>> +                cpus[j].props.has_node_id &&
>> +                cpus[i].props.socket_id == cpus[j].props.socket_id &&
>> +                cpus[i].props.cluster_id == cpus[j].props.cluster_id &&
>> +                cpus[i].props.node_id != cpus[j].props.node_id) {
>> +                warn_report("CPU-%d and CPU-%d in socket-%ld-cluster-%ld "
>> +                             "have been associated with node-%ld and node-%ld "
>> +                             "respectively. It can cause OSes like Linux to"
>> +                             "misbehave", i, j, cpus[i].props.socket_id,
>> +                             cpus[i].props.cluster_id, cpus[i].props.node_id,
>> +                             cpus[j].props.node_id);
> 
> machine_run_board_init() takes an Error* argument, but is only called
> once by qemu_init_board() with errp=&error_fatal. I suppose using
> warn_report() here is OK.
> 
> Acked-by: Philippe Mathieu-Daudé <philmd@linaro.org>
> 

warn_report() here is correct because it's inappropriate to propogate the
warning message to @error_fatal through error_setg(). When the messages
included in @error_fatal is handled and printed in util/error.c::error_handle(),
the QEMU process will be terminated unexpectedly.

>> +            }
>> +        }
>> +    }
>> +}
>> +
>>   MemoryRegion *machine_consume_memdev(MachineState *machine,
>>                                        HostMemoryBackend *backend)
>>   {
>> @@ -1337,6 +1376,9 @@ void machine_run_board_init(MachineState *machine, const char *mem_path, Error *
>>           numa_complete_configuration(machine);
>>           if (machine->numa_state->num_nodes) {
>>               machine_numa_finish_cpu_init(machine);
>> +            if (machine_class->cpu_cluster_has_numa_boundary) {
>> +                validate_cpu_cluster_to_numa_boundary(machine);
>> +            }
>>           }
>>       }

Thanks,
Gavin