[PATCH v4 47/53] KVM: PPC: Book3S HV P9: Avoid changing MSR[RI] in entry and exit

[Nicholas Piggin]

kvm_hstate.in_guest provides the equivalent of MSR[RI]=0 protection,
and it covers the existing MSR[RI]=0 section in late entry and early
exit, so clearing and setting MSR[RI] in those cases does not
actually do anything useful.

Remove the RI manipulation and replace it with comments. Make the
in_guest memory accesses a bit closer to a proper critical section
pattern. This speeds up guest entry/exit performance.

This also removes the MSR[RI] warnings which aren't very interesting
and would cause crashes if they hit due to causing an interrupt in
non-recoverable code.

Signed-off-by: Nicholas Piggin <npiggin at gmail.com>
---
 arch/powerpc/kvm/book3s_hv_p9_entry.c | 50 ++++++++++++---------------
 1 file changed, 23 insertions(+), 27 deletions(-)

diff --git a/arch/powerpc/kvm/book3s_hv_p9_entry.c b/arch/powerpc/kvm/book3s_hv_p9_entry.c
index 8fa48ba01f79..6120cdf281b9 100644
--- a/arch/powerpc/kvm/book3s_hv_p9_entry.c
+++ b/arch/powerpc/kvm/book3s_hv_p9_entry.c
@@ -904,7 +904,15 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
 	 * But TM could be split out if this would be a significant benefit.
 	 */
 
-	local_paca->kvm_hstate.in_guest = KVM_GUEST_MODE_HV_P9;
+	/*
+	 * MSR[RI] does not need to be cleared (and is not, for radix guests
+	 * with no prefetch bug), because in_guest is set. If we take a SRESET
+	 * or MCE with in_guest set but still in HV mode, then
+	 * kvmppc_p9_bad_interrupt handles the interrupt, which effectively
+	 * clears MSR[RI] and doesn't return.
+	 */
+	WRITE_ONCE(local_paca->kvm_hstate.in_guest, KVM_GUEST_MODE_HV_P9);
+	barrier(); /* Open in_guest critical section */
 
 	/*
 	 * Hash host, hash guest, or radix guest with prefetch bug, all have
@@ -916,14 +924,10 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
 
 	save_clear_host_mmu(kvm);
 
-	if (kvm_is_radix(kvm)) {
+	if (kvm_is_radix(kvm))
 		switch_mmu_to_guest_radix(kvm, vcpu, lpcr);
-		if (!cpu_has_feature(CPU_FTR_P9_RADIX_PREFETCH_BUG))
-			__mtmsrd(0, 1); /* clear RI */
-
-	} else {
+	else
 		switch_mmu_to_guest_hpt(kvm, vcpu, lpcr);
-	}
 
 	/* TLBIEL uses LPID=LPIDR, so run this after setting guest LPID */
 	check_need_tlb_flush(kvm, vc->pcpu, nested);
@@ -978,19 +982,16 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
 	vcpu->arch.regs.gpr[3] = local_paca->kvm_hstate.scratch2;
 
 	/*
-	 * Only set RI after reading machine check regs (DAR, DSISR, SRR0/1)
-	 * and hstate scratch (which we need to move into exsave to make
-	 * re-entrant vs SRESET/MCE)
+	 * After reading machine check regs (DAR, DSISR, SRR0/1) and hstate
+	 * scratch (which we need to move into exsave to make re-entrant vs
+	 * SRESET/MCE), register state is protected from reentrancy. However
+	 * timebase, MMU, among other state is still set to guest, so don't
+	 * enable MSR[RI] here. It gets enabled at the end, after in_guest
+	 * is cleared.
+	 *
+	 * It is possible an NMI could come in here, which is why it is
+	 * important to save the above state early so it can be debugged.
 	 */
-	if (ri_set) {
-		if (unlikely(!(mfmsr() & MSR_RI))) {
-			__mtmsrd(MSR_RI, 1);
-			WARN_ON_ONCE(1);
-		}
-	} else {
-		WARN_ON_ONCE(mfmsr() & MSR_RI);
-		__mtmsrd(MSR_RI, 1);
-	}
 
 	vcpu->arch.regs.gpr[9] = exsave[EX_R9/sizeof(u64)];
 	vcpu->arch.regs.gpr[10] = exsave[EX_R10/sizeof(u64)];
@@ -1048,13 +1049,6 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
 				 */
 				mtspr(SPRN_HSRR0, vcpu->arch.regs.nip);
 				mtspr(SPRN_HSRR1, vcpu->arch.shregs.msr);
-
-				/*
-				 * tm_return_to_guest re-loads SRR0/1, DAR,
-				 * DSISR after RI is cleared, in case they had
-				 * been clobbered by a MCE.
-				 */
-				__mtmsrd(0, 1); /* clear RI */
 				goto tm_return_to_guest;
 			}
 		}
@@ -1154,7 +1148,9 @@ int kvmhv_vcpu_entry_p9(struct kvm_vcpu *vcpu, u64 time_limit, unsigned long lpc
 
 	restore_p9_host_os_sprs(vcpu, &host_os_sprs);
 
-	local_paca->kvm_hstate.in_guest = KVM_GUEST_MODE_NONE;
+	barrier(); /* Close in_guest critical section */
+	WRITE_ONCE(local_paca->kvm_hstate.in_guest, KVM_GUEST_MODE_NONE);
+	/* Interrupts are recoverable at this point */
 
 	/*
 	 * cp_abort is required if the processor supports local copy-paste
-- 
2.23.0