@@ -1402,6 +1402,12 @@ static target_ulong h_get_cpu_characteristics(PowerPCCPU *cpu,
uint8_t count_cache_flush_assist = spapr_get_cap(spapr,
SPAPR_CAP_CCF_ASSIST);
+ if (kvm_enabled() && spapr->chars.character) {
+ args[0] = spapr->chars.character;
+ args[1] = spapr->chars.behaviour;
+ return H_SUCCESS;
+ }
+
switch (safe_cache) {
case SPAPR_CAP_WORKAROUND:
characteristics |= H_CPU_CHAR_L1D_FLUSH_ORI30;
@@ -280,6 +280,7 @@ struct SpaprMachineState {
Error *fwnmi_migration_blocker;
SpaprWatchdog wds[WDT_MAX_WATCHDOGS];
+ struct kvm_ppc_cpu_char chars;
};
#define H_SUCCESS 0
@@ -2499,7 +2499,8 @@ bool kvmppc_has_cap_xive(void)
static void kvmppc_get_cpu_characteristics(KVMState *s)
{
- struct kvm_ppc_cpu_char c;
+ SpaprMachineState *spapr = SPAPR_MACHINE(qdev_get_machine());
+ struct kvm_ppc_cpu_char c = {0};
int ret;
/* Assume broken */
@@ -2509,18 +2510,24 @@ static void kvmppc_get_cpu_characteristics(KVMState *s)
ret = kvm_vm_check_extension(s, KVM_CAP_PPC_GET_CPU_CHAR);
if (!ret) {
- return;
+ goto err;
}
ret = kvm_vm_ioctl(s, KVM_PPC_GET_CPU_CHAR, &c);
if (ret < 0) {
- return;
+ goto err;
}
+ spapr->chars = c;
cap_ppc_safe_cache = parse_cap_ppc_safe_cache(c);
cap_ppc_safe_bounds_check = parse_cap_ppc_safe_bounds_check(c);
cap_ppc_safe_indirect_branch = parse_cap_ppc_safe_indirect_branch(c);
cap_ppc_count_cache_flush_assist =
parse_cap_ppc_count_cache_flush_assist(c);
+
+ return;
+
+err:
+ memset(&(spapr->chars), 0, sizeof(struct kvm_ppc_cpu_char));
}
int kvmppc_get_cap_safe_cache(void)
Currently, on a P10 KVM guest, the mitigations seen in the output of "lscpu" command are different from the host. The reason for this behaviour is that when the KVM guest makes the "h_get_cpu_characteristics" hcall, QEMU does not consider the data it received from the host via the KVM_PPC_GET_CPU_CHAR ioctl, and just uses the values present in spapr->eff.caps[], which in turn just contain the default values set in spapr_machine_class_init(). Fix this behaviour by making sure that h_get_cpu_characteristics() returns the data received from the KVM ioctl for a KVM guest. Perf impact: With null syscall benchmark[1], ~45% improvement is observed. 1. Vanilla QEMU $ ./null_syscall 132.19 ns 456.54 cycles 2. With this patch $ ./null_syscall 91.18 ns 314.57 cycles [1]: https://ozlabs.org/~anton/junkcode/null_syscall.c Signed-off-by: Gautam Menghani <gautam@linux.ibm.com> --- v1 -> v2: Handle the case where KVM_PPC_GET_CPU_CHAR ioctl fails hw/ppc/spapr_hcall.c | 6 ++++++ include/hw/ppc/spapr.h | 1 + target/ppc/kvm.c | 13 ++++++++++--- 3 files changed, 17 insertions(+), 3 deletions(-)