On POWER9 and newer, rather than the complex HMI synchronisation and
subcore state, have each thread un-apply the guest TB offset before
calling into the early HMI handler.
This allows the subcore state to be avoided, including subcore enter
/ exit guest, which includes an expensive divide that shows up
slightly in profiles.
Signed-off-by: Nicholas Piggin <npiggin at gmail.com>
---
arch/powerpc/include/asm/kvm_ppc.h | 1 +
arch/powerpc/kvm/book3s_hv.c | 12 +++---
arch/powerpc/kvm/book3s_hv_hmi.c | 7 +++-
arch/powerpc/kvm/book3s_hv_p9_entry.c | 2 +-
arch/powerpc/kvm/book3s_hv_ras.c | 54 +++++++++++++++++++++++++++
5 files changed, 67 insertions(+), 9 deletions(-)
diff --git a/arch/powerpc/include/asm/kvm_ppc.h b/arch/powerpc/include/asm/kvm_ppc.h
index 671fbd1a765e..70ffcb3c91bf 100644
--- a/arch/powerpc/include/asm/kvm_ppc.h
+++ b/arch/powerpc/include/asm/kvm_ppc.h
@@ -760,6 +760,7 @@ void kvmppc_realmode_machine_check(struct kvm_vcpu *vcpu);
void kvmppc_subcore_enter_guest(void);
void kvmppc_subcore_exit_guest(void);
long kvmppc_realmode_hmi_handler(void);
+long kvmppc_p9_realmode_hmi_handler(struct kvm_vcpu *vcpu);
long kvmppc_h_enter(struct kvm_vcpu *vcpu, unsigned long flags,
long pte_index, unsigned long pteh, unsigned long ptel);
long kvmppc_h_remove(struct kvm_vcpu *vcpu, unsigned long flags,
diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c
index 351018f617fb..449ac0a19ceb 100644
--- a/arch/powerpc/kvm/book3s_hv.c
+++ b/arch/powerpc/kvm/book3s_hv.c
@@ -4014,8 +4014,6 @@ static int kvmhv_p9_guest_entry(struct kvm_vcpu *vcpu, u64 time_limit,
vcpu->arch.ceded = 0;
- kvmppc_subcore_enter_guest();
-
vcpu_vpa_increment_dispatch(vcpu);
if (kvmhv_on_pseries()) {
@@ -4068,8 +4066,6 @@ static int kvmhv_p9_guest_entry(struct kvm_vcpu *vcpu, u64 time_limit,
vcpu_vpa_increment_dispatch(vcpu);
- kvmppc_subcore_exit_guest();
-
return trap;
}
@@ -6069,9 +6065,11 @@ static int kvmppc_book3s_init_hv(void)
if (r)
return r;
- r = kvm_init_subcore_bitmap();
- if (r)
- return r;
+ if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
+ r = kvm_init_subcore_bitmap();
+ if (r)
+ return r;
+ }
/*
* We need a way of accessing the XICS interrupt controller,
diff --git a/arch/powerpc/kvm/book3s_hv_hmi.c b/arch/powerpc/kvm/book3s_hv_hmi.c
index 9af660476314..1ec50c69678b 100644
--- a/arch/powerpc/kvm/book3s_hv_hmi.c
+++ b/arch/powerpc/kvm/book3s_hv_hmi.c
@@ -20,10 +20,15 @@ void wait_for_subcore_guest_exit(void)
/*
* NULL bitmap pointer indicates that KVM module hasn't
- * been loaded yet and hence no guests are running.
+ * been loaded yet and hence no guests are running, or running
+ * on POWER9 or newer CPU.
+ *
* If no KVM is in use, no need to co-ordinate among threads
* as all of them will always be in host and no one is going
* to modify TB other than the opal hmi handler.
+ *
+ * POWER9 and newer don't need this synchronisation.
+ *
* Hence, just return from here.
*/
if (!local_paca->sibling_subcore_state)
diff --git a/arch/powerpc/kvm/book3s_hv_p9_entry.c b/arch/powerpc/kvm/book3s_hv_p9_entry.c
index fbecbdc42c26..86a222f97e8e 100644
--- a/arch/powerpc/kvm/book3s_hv_p9_entry.c
+++ b/arch/powerpc/kvm/book3s_hv_p9_entry.c
@@ -938,7 +938,7 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
kvmppc_realmode_machine_check(vcpu);
} else if (unlikely(trap == BOOK3S_INTERRUPT_HMI)) {
- kvmppc_realmode_hmi_handler();
+ kvmppc_p9_realmode_hmi_handler(vcpu);
} else if (trap == BOOK3S_INTERRUPT_H_EMUL_ASSIST) {
vcpu->arch.emul_inst = mfspr(SPRN_HEIR);
diff --git a/arch/powerpc/kvm/book3s_hv_ras.c b/arch/powerpc/kvm/book3s_hv_ras.c
index d4bca93b79f6..ccfd96965630 100644
--- a/arch/powerpc/kvm/book3s_hv_ras.c
+++ b/arch/powerpc/kvm/book3s_hv_ras.c
@@ -136,6 +136,60 @@ void kvmppc_realmode_machine_check(struct kvm_vcpu *vcpu)
vcpu->arch.mce_evt = mce_evt;
}
+
+long kvmppc_p9_realmode_hmi_handler(struct kvm_vcpu *vcpu)
+{
+ struct kvmppc_vcore *vc = vcpu->arch.vcore;
+ long ret = 0;
+
+ /*
+ * Unapply and clear the offset first. That way, if the TB was not
+ * resynced then it will remain in host-offset, and if it was resynced
+ * then it is brought into host-offset. Then the tb offset is
+ * re-applied before continuing with the KVM exit.
+ *
+ * This way, we don't need to actually know whether not OPAL resynced
+ * the timebase or do any of the complicated dance that the P7/8
+ * path requires.
+ */
+ if (vc->tb_offset_applied) {
+ u64 new_tb = mftb() - vc->tb_offset_applied;
+ mtspr(SPRN_TBU40, new_tb);
+ if ((mftb() & 0xffffff) < (new_tb & 0xffffff)) {
+ new_tb += 0x1000000;
+ mtspr(SPRN_TBU40, new_tb);
+ }
+ vc->tb_offset_applied = 0;
+ }
+
+ local_paca->hmi_irqs++;
+
+ if (hmi_handle_debugtrig(NULL) >= 0) {
+ ret = 1;
+ goto out;
+ }
+
+ if (ppc_md.hmi_exception_early)
+ ppc_md.hmi_exception_early(NULL);
+
+out:
+ if (vc->tb_offset) {
+ u64 new_tb = mftb() + vc->tb_offset;
+ mtspr(SPRN_TBU40, new_tb);
+ if ((mftb() & 0xffffff) < (new_tb & 0xffffff)) {
+ new_tb += 0x1000000;
+ mtspr(SPRN_TBU40, new_tb);
+ }
+ vc->tb_offset_applied = vc->tb_offset;
+ }
+
+ return ret;
+}
+
+/*
+ * The following subcore HMI handling is all only for pre-POWER9 CPUs.
+ */
+
/* Check if dynamic split is in force and return subcore size accordingly. */
static inline int kvmppc_cur_subcore_size(void)
{
--
2.23.0