Move register saving and loading from kvmhv_p9_guest_entry() into the HV
and nested entry handlers.
Accesses are scheduled to reduce mtSPR / mfSPR interleaving which
reduces SPR scoreboard stalls.
XXX +212 cycles here somewhere (7566), investigate POWER9 virt-mode NULL hcall
Signed-off-by: Nicholas Piggin <npiggin@xxxxxxxxx>
---
arch/powerpc/kvm/book3s_hv.c | 79 ++++++++++------------
arch/powerpc/kvm/book3s_hv_p9_entry.c | 96 ++++++++++++++++++++-------
2 files changed, 109 insertions(+), 66 deletions(-)
diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c
index cb66c9534dbf..8c1c93ebd669 100644
--- a/arch/powerpc/kvm/book3s_hv.c
+++ b/arch/powerpc/kvm/book3s_hv.c
@@ -3794,9 +3794,15 @@ static int kvmhv_vcpu_entry_p9_nested(struct kvm_vcpu *vcpu, u64 time_limit, uns
{
struct kvmppc_vcore *vc = vcpu->arch.vcore;
unsigned long host_psscr;
+ unsigned long msr;
struct hv_guest_state hvregs;
- int trap;
+ struct p9_host_os_sprs host_os_sprs;
s64 dec;
+ int trap;
+
+ switch_pmu_to_guest(vcpu, &host_os_sprs);
+
+ save_p9_host_os_sprs(&host_os_sprs);
/*
* We need to save and restore the guest visible part of the
@@ -3805,6 +3811,27 @@ static int kvmhv_vcpu_entry_p9_nested(struct kvm_vcpu *vcpu, u64 time_limit, uns
* this is done in kvmhv_vcpu_entry_p9() below otherwise.
*/
host_psscr = mfspr(SPRN_PSSCR_PR);
+
+ hard_irq_disable();
+ if (lazy_irq_pending())
+ return 0;
+
+ /* MSR bits may have been cleared by context switch */
+ msr = 0;
+ if (IS_ENABLED(CONFIG_PPC_FPU))
+ msr |= MSR_FP;
+ if (cpu_has_feature(CPU_FTR_ALTIVEC))
+ msr |= MSR_VEC;
+ if (cpu_has_feature(CPU_FTR_VSX))
+ msr |= MSR_VSX;
+ if (cpu_has_feature(CPU_FTR_TM) ||
+ cpu_has_feature(CPU_FTR_P9_TM_HV_ASSIST))
+ msr |= MSR_TM;
+ msr = msr_check_and_set(msr);
+
+ if (unlikely(load_vcpu_state(vcpu, &host_os_sprs)))
+ msr = mfmsr(); /* TM restore can update msr */
+
mtspr(SPRN_PSSCR_PR, vcpu->arch.psscr);
kvmhv_save_hv_regs(vcpu, &hvregs);
hvregs.lpcr = lpcr;
@@ -3846,12 +3873,20 @@ static int kvmhv_vcpu_entry_p9_nested(struct kvm_vcpu *vcpu, u64 time_limit, uns
vcpu->arch.psscr = mfspr(SPRN_PSSCR_PR);
mtspr(SPRN_PSSCR_PR, host_psscr);
+ store_vcpu_state(vcpu);
+
dec = mfspr(SPRN_DEC);
if (!(lpcr & LPCR_LD)) /* Sign extend if not using large decrementer */
dec = (s32) dec;
*tb = mftb();
vcpu->arch.dec_expires = dec + (*tb + vc->tb_offset);
+ timer_rearm_host_dec(*tb);
+
+ restore_p9_host_os_sprs(vcpu, &host_os_sprs);
+
+ switch_pmu_to_host(vcpu, &host_os_sprs);
+
return trap;
}
@@ -3862,9 +3897,7 @@ static int kvmhv_p9_guest_entry(struct kvm_vcpu *vcpu, u64 time_limit,
unsigned long lpcr, u64 *tb)
{
struct kvmppc_vcore *vc = vcpu->arch.vcore;
- struct p9_host_os_sprs host_os_sprs;
u64 next_timer;
- unsigned long msr;
int trap;
next_timer = timer_get_next_tb();
@@ -3875,33 +3908,6 @@ static int kvmhv_p9_guest_entry(struct kvm_vcpu *vcpu, u64 time_limit,
vcpu->arch.ceded = 0;
- save_p9_host_os_sprs(&host_os_sprs);
-
- /*
- * This could be combined with MSR[RI] clearing, but that expands
- * the unrecoverable window. It would be better to cover unrecoverable
- * with KVM bad interrupt handling rather than use MSR[RI] at all.
- *
- * Much more difficult and less worthwhile to combine with IR/DR
- * disable.
- */
- hard_irq_disable();
- if (lazy_irq_pending())
- return 0;
-
- /* MSR bits may have been cleared by context switch */
- msr = 0;
- if (IS_ENABLED(CONFIG_PPC_FPU))
- msr |= MSR_FP;
- if (cpu_has_feature(CPU_FTR_ALTIVEC))
- msr |= MSR_VEC;
- if (cpu_has_feature(CPU_FTR_VSX))
- msr |= MSR_VSX;
- if (cpu_has_feature(CPU_FTR_TM) ||
- cpu_has_feature(CPU_FTR_P9_TM_HV_ASSIST))
- msr |= MSR_TM;
- msr = msr_check_and_set(msr);
-
kvmppc_subcore_enter_guest();
vc->entry_exit_map = 1;
@@ -3909,11 +3915,6 @@ static int kvmhv_p9_guest_entry(struct kvm_vcpu *vcpu, u64 time_limit,
vcpu_vpa_increment_dispatch(vcpu);
- if (unlikely(load_vcpu_state(vcpu, &host_os_sprs)))
- msr = mfmsr(); /* MSR may have been updated */
-
- switch_pmu_to_guest(vcpu, &host_os_sprs);
-
if (kvmhv_on_pseries()) {
trap = kvmhv_vcpu_entry_p9_nested(vcpu, time_limit, lpcr, tb);
@@ -3956,16 +3957,8 @@ static int kvmhv_p9_guest_entry(struct kvm_vcpu *vcpu, u64 time_limit,
vcpu->arch.slb_max = 0;
}
- switch_pmu_to_host(vcpu, &host_os_sprs);
-
- store_vcpu_state(vcpu);
-
vcpu_vpa_increment_dispatch(vcpu);
- timer_rearm_host_dec(*tb);
-
- restore_p9_host_os_sprs(vcpu, &host_os_sprs);
-
vc->entry_exit_map = 0x101;
vc->in_guest = 0;
diff --git a/arch/powerpc/kvm/book3s_hv_p9_entry.c b/arch/powerpc/kvm/book3s_hv_p9_entry.c
index 5a34f0199bfe..ea531f76f116 100644
--- a/arch/powerpc/kvm/book3s_hv_p9_entry.c
+++ b/arch/powerpc/kvm/book3s_hv_p9_entry.c
@@ -530,6 +530,7 @@ static void save_clear_guest_mmu(struct kvm *kvm, struct kvm_vcpu *vcpu)
int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpcr, u64 *tb)
{
+ struct p9_host_os_sprs host_os_sprs;
struct kvm *kvm = vcpu->kvm;
struct kvm_nested_guest *nested = vcpu->arch.nested;
struct kvmppc_vcore *vc = vcpu->arch.vcore;
@@ -559,9 +560,6 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
vcpu->arch.ceded = 0;
- /* Could avoid mfmsr by passing around, but probably no big deal */
- msr = mfmsr();
-
host_hfscr = mfspr(SPRN_HFSCR);
host_ciabr = mfspr(SPRN_CIABR);
host_dawr0 = mfspr(SPRN_DAWR0);
@@ -576,6 +574,41 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
local_paca->kvm_hstate.host_purr = mfspr(SPRN_PURR);
local_paca->kvm_hstate.host_spurr = mfspr(SPRN_SPURR);
+ switch_pmu_to_guest(vcpu, &host_os_sprs);
+
+ save_p9_host_os_sprs(&host_os_sprs);
+
+ /*
+ * This could be combined with MSR[RI] clearing, but that expands
+ * the unrecoverable window. It would be better to cover unrecoverable
+ * with KVM bad interrupt handling rather than use MSR[RI] at all.
+ *
+ * Much more difficult and less worthwhile to combine with IR/DR
+ * disable.
+ */
+ hard_irq_disable();
+ if (lazy_irq_pending()) {
+ trap = 0;
+ goto out;
+ }
+
+ /* MSR bits may have been cleared by context switch */
+ msr = 0;
+ if (IS_ENABLED(CONFIG_PPC_FPU))
+ msr |= MSR_FP;
+ if (cpu_has_feature(CPU_FTR_ALTIVEC))
+ msr |= MSR_VEC;
+ if (cpu_has_feature(CPU_FTR_VSX))
+ msr |= MSR_VSX;
+ if (cpu_has_feature(CPU_FTR_TM) ||
+ cpu_has_feature(CPU_FTR_P9_TM_HV_ASSIST))
+ msr |= MSR_TM;
+ msr = msr_check_and_set(msr);
+ /* Save MSR for restore. This is after hard disable, so EE is clear. */
+
+ if (unlikely(load_vcpu_state(vcpu, &host_os_sprs)))
+ msr = mfmsr(); /* MSR may have been updated */
+
if (vc->tb_offset) {
u64 new_tb = *tb + vc->tb_offset;
mtspr(SPRN_TBU40, new_tb);
@@ -634,6 +667,14 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
mtspr(SPRN_SPRG2, vcpu->arch.shregs.sprg2);
mtspr(SPRN_SPRG3, vcpu->arch.shregs.sprg3);
+ /*
+ * It might be preferable to load_vcpu_state here, in order to get the
+ * GPR/FP register loads executing in parallel with the previous mtSPR
+ * instructions, but for now that can't be done because the TM handling
+ * in load_vcpu_state can change some SPRs and vcpu state (nip, msr).
+ * But TM could be split out if this would be a significant benefit.
+ */
+
local_paca->kvm_hstate.in_guest = KVM_GUEST_MODE_HV_P9;
/*
@@ -811,6 +852,20 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
vc->dpdes = mfspr(SPRN_DPDES);
vc->vtb = mfspr(SPRN_VTB);
+ save_clear_guest_mmu(kvm, vcpu);
+ switch_mmu_to_host(kvm, host_pidr);
+
+ /*
+ * If we are in real mode, only switch MMU on after the MMU is
+ * switched to host, to avoid the P9_RADIX_PREFETCH_BUG.
+ */
+ if (IS_ENABLED(CONFIG_PPC_TRANSACTIONAL_MEM) &&
+ vcpu->arch.shregs.msr & MSR_TS_MASK)
+ msr |= MSR_TS_S;
+ __mtmsrd(msr, 0);
+
+ store_vcpu_state(vcpu);
+
dec = mfspr(SPRN_DEC);
if (!(lpcr & LPCR_LD)) /* Sign extend if not using large decrementer */
dec = (s32) dec;
@@ -843,6 +898,19 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
mtspr(SPRN_DAWRX1, host_dawrx1);
}
+ mtspr(SPRN_DPDES, 0);
+ if (vc->pcr)
+ mtspr(SPRN_PCR, PCR_MASK);
+
+ /* HDEC must be at least as large as DEC, so decrementer_max fits */
+ mtspr(SPRN_HDEC, decrementer_max);
+
+ timer_rearm_host_dec(*tb);
+
+ restore_p9_host_os_sprs(vcpu, &host_os_sprs);
+
+ local_paca->kvm_hstate.in_guest = KVM_GUEST_MODE_NONE;
+
if (kvm_is_radix(kvm)) {
/*
* Since this is radix, do a eieio; tlbsync; ptesync sequence
@@ -859,26 +927,8 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
if (cpu_has_feature(CPU_FTR_ARCH_31))
asm volatile(PPC_CP_ABORT);
- mtspr(SPRN_DPDES, 0);
- if (vc->pcr)
- mtspr(SPRN_PCR, PCR_MASK);
-
- /* HDEC must be at least as large as DEC, so decrementer_max fits */
- mtspr(SPRN_HDEC, decrementer_max);
-
- save_clear_guest_mmu(kvm, vcpu);
- switch_mmu_to_host(kvm, host_pidr);
- local_paca->kvm_hstate.in_guest = KVM_GUEST_MODE_NONE;
-
- /*
- * If we are in real mode, only switch MMU on after the MMU is
- * switched to host, to avoid the P9_RADIX_PREFETCH_BUG.
- */
- if (IS_ENABLED(CONFIG_PPC_TRANSACTIONAL_MEM) &&
- vcpu->arch.shregs.msr & MSR_TS_MASK)
- msr |= MSR_TS_S;
-
- __mtmsrd(msr, 0);
+out:
+ switch_pmu_to_host(vcpu, &host_os_sprs);
end_timing(vcpu);
--
2.23.0
