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 <npig...@gmail.com>
---
arch/powerpc/kvm/book3s_hv.c | 12 +++++-------
arch/powerpc/kvm/book3s_hv_hmi.c | 7 ++++++-
arch/powerpc/kvm/book3s_hv_p9_entry.c | 21 ++++++++++++++++++++-
arch/powerpc/kvm/book3s_hv_ras.c | 4 ++++
4 files changed, 35 insertions(+), 9 deletions(-)
diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c
index 957efde59014..7d5c31537a79 100644
--- a/arch/powerpc/kvm/book3s_hv.c
+++ b/arch/powerpc/kvm/book3s_hv.c
@@ -3985,8 +3985,6 @@ 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()) {
@@ -4039,8 +4037,6 @@ static int kvmhv_p9_guest_entry(struct kvm_vcpu *vcpu,
u64 time_limit,
vcpu_vpa_increment_dispatch(vcpu);
- kvmppc_subcore_exit_guest();
-
return trap;
}
@@ -6022,9 +6018,11 @@ 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 44150a29f6e0..3c9e7a500264 100644
--- a/arch/powerpc/kvm/book3s_hv_p9_entry.c
+++ b/arch/powerpc/kvm/book3s_hv_p9_entry.c
@@ -3,6 +3,7 @@
#include <linux/kvm_host.h>
#include <asm/asm-prototypes.h>
#include <asm/dbell.h>
+#include <asm/interrupt.h>
#include <asm/kvm_ppc.h>
#include <asm/pmc.h>
#include <asm/ppc-opcode.h>
@@ -900,7 +901,25 @@ 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();
+ /*
+ * Unapply and clear the offset first. That way, if the TB
+ * was fine then no harm done, if it is corrupted then the
+ * HMI resync will bring it back to host mode. This way, we
+ * don't need to actualy know whether not OPAL resynced the
+ * timebase. Although it would be cleaner if we could rely
+ * on that, early POWER9 OPAL did not support the
+ * OPAL_HANDLE_HMI2 call.
+ */
+ 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;
+ }
+ hmi_exception_realmode(NULL);
} 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..a49ee9bdab67 100644
--- a/arch/powerpc/kvm/book3s_hv_ras.c
+++ b/arch/powerpc/kvm/book3s_hv_ras.c
@@ -136,6 +136,10 @@ void kvmppc_realmode_machine_check(struct kvm_vcpu *vcpu)
vcpu->arch.mce_evt = mce_evt;
}
+/*
+ * This 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
