On Sun, Feb 1, 2026 at 9:05 AM Leo Yan [off-list ref] wrote:

On Sat, Jan 31, 2026 at 12:02:20PM -0800, Ian Rogers wrote:

[...]

quoted

@@ -1666,7 +1661,7 @@ kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)
              return ret;
      }

-     for (events_tp = kvm_events_tp; *events_tp; events_tp++)
+     for (events_tp = kvm_events_tp(); *events_tp; events_tp++)
              events_tp_size++;

      rec_argc = ARRAY_SIZE(record_args) + argc + 2 +

@@ -1681,7 +1676,7 @@ kvm_events_record(struct perf_kvm_stat *kvm, int argc, const char **argv)

      for (j = 0; j < events_tp_size; j++) {
              rec_argv[i++] = STRDUP_FAIL_EXIT("-e");
-             rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp[j]);
+             rec_argv[i++] = STRDUP_FAIL_EXIT(kvm_events_tp()[j]);
      }

Nitpick: we can assign reuse events_tp throughout the
kvm_events_record().  Something like:

  events_tp = kvm_events_tp();
  for (j = 0; events_tp[j]; j++)
       events_tp_size++;

  ...

  for (j = 0; j < events_tp_size; j++) {
       rec_argv[i++] = STRDUP_FAIL_EXIT("-e");
       rec_argv[i++] = STRDUP_FAIL_EXIT(events_tp[j]);
  }

[...]

quoted

+int setup_kvm_events_tp(struct perf_kvm_stat *kvm)
+{
+     switch (EM_HOST) {
+     case EM_PPC:
+     case EM_PPC64:
+             return __setup_kvm_events_tp_powerpc(kvm);
+     default:
+             return 0;
+     }
+}
+
+int cpu_isa_init(struct perf_kvm_stat *kvm, const char *cpuid)
+{
+     switch (EM_HOST) {
+     case EM_AARCH64:
+             return __cpu_isa_init_arm64(kvm);
+     case EM_LOONGARCH:
+             return __cpu_isa_init_loongarch(kvm);
+     case EM_PPC:
+     case EM_PPC64:
+             return __cpu_isa_init_powerpc(kvm);
+     case EM_RISCV:
+             return __cpu_isa_init_riscv(kvm);
+     case EM_S390:
+             return __cpu_isa_init_s390(kvm, cpuid);
+     case EM_X86_64:
+     case EM_386:
+             return __cpu_isa_init_x86(kvm, cpuid);
+     default:
+             pr_err("Unsupported kvm-stat host %d\n", EM_HOST);
+             return -1;
+     }
+}

For a general solution, I'd prefer to use "install" + callback methods
rather than the opened code.  E.g., each arch installs a structure
with callbacks:

  struct kvm_stat_arch kvm_stat_aarch64 {
      .setup_events_tp = NULL;
      .cpu_isa_init    = __cpu_isa_init_arm64;
      ...
  };

Then at the init phase, we can install arch's structure:

  switch (EM_HOST) {
  case EM_AARCH64:
      kvm_stat_arch_init(&kvm_stat_aarch64);
      break;
  ...
  }

Afterwards, it is no need to check EM_HOST anymore, it can simply
invoke the arch's callback:

  int cpu_isa_init(struct perf_kvm_stat *kvm, const char *cpuid)
  {
      if (!arch_kvm_stat || !arch_kvm_stat->cpu_isa_init)
          return -1;

      return arch_kvm_stat->cpu_isa_init(kvm, cpuid);
  }

As a result, we can avoid spreading "switch (EM_HOST)" everywhere.

So the use of switch(EM_HOST) is to just show the equivalence between
the before and after code. The EM_HOST has become an e_machine read
from the session/env by the end of the series. I agree there is naming
clean up, etc. that can be done to the kvm_stat code but I just wanted
to focus the series on the removal of the arch directory by way of
switching to using the e_machine value.

Thanks,
Ian

Thanks,
Leo