[PATCH v1 0/4] powerpc/64s: interrupt speedups

@@ -148,9 +148,21 @@ static inline void interrupt_enter_prepare(struct pt_regs *regs, struct interrup
 #endif
 
 #ifdef CONFIG_PPC64
-	if (irq_soft_mask_set_return(IRQS_ALL_DISABLED) == IRQS_ENABLED)
+	bool trace_enable = false;
+
+	if (IS_ENABLED(CONFIG_TRACE_IRQFLAGS)) {
+		if (irq_soft_mask_set_return(IRQS_DISABLED) == IRQS_ENABLED)
+			trace_enable = true;
+	} else {
+		irq_soft_mask_set(IRQS_DISABLED);
+	}
+	/* If the interrupt was taken with HARD_DIS set, don't enable MSR[EE] */
+	if (local_paca->irq_happened & PACA_IRQ_HARD_DIS)
+		__hard_RI_enable();
+	else
+		__hard_irq_enable();
+	if (trace_enable)
 		trace_hardirqs_off();
-	local_paca->irq_happened |= PACA_IRQ_HARD_DIS;
 
 	if (user_mode(regs)) {
 		CT_WARN_ON(ct_state() != CONTEXT_USER);

@@ -200,13 +212,20 @@ static inline void interrupt_exit_prepare(struct pt_regs *regs, struct interrupt
 
 static inline void interrupt_async_enter_prepare(struct pt_regs *regs, struct interrupt_state *state)
 {
+#ifdef CONFIG_PPC64
+	/* Ensure interrupt_enter_prepare does not enable MSR[EE] */
+	local_paca->irq_happened |= PACA_IRQ_HARD_DIS;
+#endif
+	interrupt_enter_prepare(regs, state);
 #ifdef CONFIG_PPC_BOOK3S_64
+	/*
+	 * MSR[RI] is only enabled after interrupt_enter_prepare, so this
+	 * thread flags access has to come afterward.
+	 */
 	if (cpu_has_feature(CPU_FTR_CTRL) &&
 	    !test_thread_local_flags(_TLF_RUNLATCH))
 		__ppc64_runlatch_on();
 #endif
-
-	interrupt_enter_prepare(regs, state);
 	irq_enter();
 }
 

@@ -273,6 +292,8 @@ static inline void interrupt_nmi_enter_prepare(struct pt_regs *regs, struct inte
 	if (IS_ENABLED(CONFIG_PPC_IRQ_SOFT_MASK_DEBUG))
 		BUG_ON(!arch_irq_disabled_regs(regs) && !(regs->msr & MSR_EE));
 
+	__hard_RI_enable();
+
 	/* Don't do any per-CPU operations until interrupt state is fixed */
 
 	if (nmi_disables_ftrace(regs)) {

@@ -370,6 +391,8 @@ interrupt_handler long func(struct pt_regs *regs)			\
 {									\
 	long ret;							\
 									\
+	__hard_RI_enable();						\
+									\
 	ret = ____##func (regs);					\
 									\
 	return ret;							\

@@ -113,7 +113,6 @@ name:
 #define IISIDE		.L_IISIDE_\name\()	/* Uses SRR0/1 not DAR/DSISR */
 #define IDAR		.L_IDAR_\name\()	/* Uses DAR (or SRR0) */
 #define IDSISR		.L_IDSISR_\name\()	/* Uses DSISR (or SRR1) */
-#define ISET_RI		.L_ISET_RI_\name\()	/* Run common code w/ MSR[RI]=1 */
 #define IBRANCH_TO_COMMON	.L_IBRANCH_TO_COMMON_\name\() /* ENTRY branch to common */
 #define IREALMODE_COMMON	.L_IREALMODE_COMMON_\name\() /* Common runs in realmode */
 #define IMASK		.L_IMASK_\name\()	/* IRQ soft-mask bit */

@@ -157,9 +156,6 @@ do_define_int n
 	.ifndef IDSISR
 		IDSISR=0
 	.endif
-	.ifndef ISET_RI
-		ISET_RI=1
-	.endif
 	.ifndef IBRANCH_TO_COMMON
 		IBRANCH_TO_COMMON=1
 	.endif

@@ -512,11 +508,6 @@ DEFINE_FIXED_SYMBOL(\name\()_common_real)
 	stb	r10,PACASRR_VALID(r13)
 	.endif
 
-	.if ISET_RI
-	li	r10,MSR_RI
-	mtmsrd	r10,1			/* Set MSR_RI */
-	.endif
-
 	.if ISTACK
 	.if IKUAP
 	kuap_save_amr_and_lock r9, r10, cr1, cr0

@@ -901,11 +892,6 @@ INT_DEFINE_BEGIN(system_reset)
 	IVEC=0x100
 	IAREA=PACA_EXNMI
 	IVIRT=0 /* no virt entry point */
-	/*
-	 * MSR_RI is not enabled, because PACA_EXNMI and nmi stack is
-	 * being used, so a nested NMI exception would corrupt it.
-	 */
-	ISET_RI=0
 	ISTACK=0
 	IKVM_REAL=1
 INT_DEFINE_END(system_reset)

@@ -986,8 +972,6 @@ EXC_COMMON_BEGIN(system_reset_common)
 	lhz	r10,PACA_IN_NMI(r13)
 	addi	r10,r10,1
 	sth	r10,PACA_IN_NMI(r13)
-	li	r10,MSR_RI
-	mtmsrd 	r10,1
 
 	mr	r10,r1
 	ld	r1,PACA_NMI_EMERG_SP(r13)

@@ -1061,12 +1045,6 @@ INT_DEFINE_BEGIN(machine_check_early)
 	IAREA=PACA_EXMC
 	IVIRT=0 /* no virt entry point */
 	IREALMODE_COMMON=1
-	/*
-	 * MSR_RI is not enabled, because PACA_EXMC is being used, so a
-	 * nested machine check corrupts it. machine_check_common enables
-	 * MSR_RI.
-	 */
-	ISET_RI=0
 	ISTACK=0
 	IDAR=1
 	IDSISR=1

@@ -1077,7 +1055,6 @@ INT_DEFINE_BEGIN(machine_check)
 	IVEC=0x200
 	IAREA=PACA_EXMC
 	IVIRT=0 /* no virt entry point */
-	ISET_RI=0
 	IDAR=1
 	IDSISR=1
 	IKVM_REAL=1

@@ -1147,9 +1124,6 @@ EXC_COMMON_BEGIN(machine_check_early_common)
 BEGIN_FTR_SECTION
 	bl	enable_machine_check
 END_FTR_SECTION_IFSET(CPU_FTR_HVMODE)
-	li	r10,MSR_RI
-	mtmsrd	r10,1
-
 	addi	r3,r1,STACK_FRAME_OVERHEAD
 	bl	machine_check_early
 	std	r3,RESULT(r1)	/* Save result */

@@ -1237,10 +1211,6 @@ EXC_COMMON_BEGIN(machine_check_common)
 	 * save area: PACA_EXMC instead of PACA_EXGEN.
 	 */
 	GEN_COMMON machine_check
-
-	/* Enable MSR_RI when finished with PACA_EXMC */
-	li	r10,MSR_RI
-	mtmsrd 	r10,1
 	addi	r3,r1,STACK_FRAME_OVERHEAD
 	bl	machine_check_exception
 	b	interrupt_return_srr

@@ -81,7 +81,12 @@ EXPORT_SYMBOL(store_fp_state)
  */
 _GLOBAL(load_up_fpu)
 	mfmsr	r5
+#ifdef CONFIG_PPC_BOOK3S_64
+	/* interrupt doesn't set MSR[RI] and HPT can fault on current access */
+	ori	r5,r5,MSR_FP|MSR_RI
+#else
 	ori	r5,r5,MSR_FP
+#endif
 #ifdef CONFIG_VSX
 BEGIN_FTR_SECTION
 	oris	r5,r5,MSR_VSX@h

@@ -47,6 +47,10 @@ EXPORT_SYMBOL(store_vr_state)
  */
 _GLOBAL(load_up_altivec)
 	mfmsr	r5			/* grab the current MSR */
+#ifdef CONFIG_PPC_BOOK3S_64
+	/* interrupt doesn't set MSR[RI] and HPT can fault on current access */
+	ori	r5,r5,MSR_RI
+#endif
 	oris	r5,r5,MSR_VEC@h
 	MTMSRD(r5)			/* enable use of AltiVec now */
 	isync

@@ -128,6 +132,10 @@ _GLOBAL(load_up_vsx)
 	andis.	r5,r12,MSR_VEC@h
 	beql+	load_up_altivec		/* skip if already loaded */
 
+	/* interrupt doesn't set MSR[RI] and HPT can fault on current access */
+	li	r5,MSR_RI
+	mtmsrd	r5,1
+
 	ld	r4,PACACURRENT(r13)
 	addi	r4,r4,THREAD		/* Get THREAD */
 	li	r6,1

@@ -306,6 +306,8 @@ static inline bool lazy_irq_pending_nocheck(void)
 	return __lazy_irq_pending(local_paca->irq_happened);
 }
 
+bool power_pmu_running(void);
+
 /*
  * This is called by asynchronous interrupts to conditionally
  * re-enable hard interrupts after having cleared the source

@@ -2380,6 +2380,19 @@ static void perf_event_interrupt(struct pt_regs *regs)
 	perf_sample_event_took(sched_clock() - start_clock);
 }
 
+bool power_pmu_running(void)
+{
+	struct cpu_hw_events *cpuhw;
+
+	/* Could this simply test local_paca->pmcregs_in_use? */
+
+	if (!ppmu)
+		return false;
+
+	cpuhw = this_cpu_ptr(&cpu_hw_events);
+	return cpuhw->n_events;
+}
+
 static int power_pmu_prepare_cpu(unsigned int cpu)
 {
 	struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);

@@ -309,17 +309,46 @@ static inline bool lazy_irq_pending_nocheck(void)
 bool power_pmu_running(void);
 
 /*
- * This is called by asynchronous interrupts to conditionally
- * re-enable hard interrupts after having cleared the source
- * of the interrupt. They are kept disabled if there is a different
- * soft-masked interrupt pending that requires hard masking.
+ * This is called by asynchronous interrupts to check whether to
+ * conditionally re-enable hard interrupts after having cleared
+ * the source of the interrupt. They are kept disabled if there
+ * is a different soft-masked interrupt pending that requires hard
+ * masking.
  */
-static inline void may_hard_irq_enable(void)
+static inline bool may_hard_irq_enable(void)
 {
-	if (!(get_paca()->irq_happened & PACA_IRQ_MUST_HARD_MASK)) {
-		get_paca()->irq_happened &= ~PACA_IRQ_HARD_DIS;
-		__hard_irq_enable();
-	}
+#ifdef CONFIG_PPC_IRQ_SOFT_MASK_DEBUG
+	BUG_ON(mfmsr() & MSR_EE);
+#endif
+#ifdef CONFIG_PERF_EVENTS
+	if (!power_pmu_running())
+		return false;
+
+	if (get_paca()->irq_happened & PACA_IRQ_MUST_HARD_MASK)
+		return false;
+
+	return true;
+#else
+	return false;
+#endif
+}
+
+/*
+ * Do the hard enabling, only call this if may_hard_irq_enable is true.
+ */
+static inline void do_hard_irq_enable(void)
+{
+#ifdef CONFIG_PPC_IRQ_SOFT_MASK_DEBUG
+	WARN_ON(irq_soft_mask_return() != IRQS_ALL_DISABLED);
+	WARN_ON(get_paca()->irq_happened & PACA_IRQ_MUST_HARD_MASK);
+	BUG_ON(mfmsr() & MSR_EE);
+#endif
+	/*
+	 * This allows PMI interrupts (and watchdog soft-NMIs) through.
+	 * There is no other reason to enable this way.
+	 */
+	get_paca()->irq_happened &= ~PACA_IRQ_HARD_DIS;
+	__hard_irq_enable();
 }
 
 static inline bool arch_irq_disabled_regs(struct pt_regs *regs)

@@ -27,7 +27,8 @@ DEFINE_INTERRUPT_HANDLER_ASYNC(doorbell_exception)
 
 	ppc_msgsync();
 
-	may_hard_irq_enable();
+	if (may_hard_irq_enable())
+		do_hard_irq_enable();
 
 	kvmppc_clear_host_ipi(smp_processor_id());
 	__this_cpu_inc(irq_stat.doorbell_irqs);

@@ -739,7 +739,8 @@ void __do_irq(struct pt_regs *regs)
 	irq = ppc_md.get_irq();
 
 	/* We can hard enable interrupts now to allow perf interrupts */
-	may_hard_irq_enable();
+	if (may_hard_irq_enable())
+		do_hard_irq_enable();
 
 	/* And finally process it */
 	if (unlikely(!irq))

@@ -567,22 +567,22 @@ DEFINE_INTERRUPT_HANDLER_ASYNC(timer_interrupt)
 		return;
 	}
 
-	/* Ensure a positive value is written to the decrementer, or else
-	 * some CPUs will continue to take decrementer exceptions. When the
-	 * PPC_WATCHDOG (decrementer based) is configured, keep this at most
-	 * 31 bits, which is about 4 seconds on most systems, which gives
-	 * the watchdog a chance of catching timer interrupt hard lockups.
-	 */
-	if (IS_ENABLED(CONFIG_PPC_WATCHDOG))
-		set_dec(0x7fffffff);
-	else
-		set_dec(decrementer_max);
-
-	/* Conditionally hard-enable interrupts now that the DEC has been
-	 * bumped to its maximum value
-	 */
-	may_hard_irq_enable();
+	/* Conditionally hard-enable interrupts. */
+	if (may_hard_irq_enable()) {
+		/* Ensure a positive value is written to the decrementer, or
+		 * else some CPUs will continue to take decrementer exceptions.
+		 * When the PPC_WATCHDOG (decrementer based) is configured,
+		 * keep this at most 31 bits, which is about 4 seconds on most
+		 * systems, which gives the watchdog a chance of catching timer
+		 * interrupt hard lockups.
+		 */
+		if (IS_ENABLED(CONFIG_PPC_WATCHDOG))
+			set_dec(0x7fffffff);
+		else
+			set_dec(decrementer_max);
 
+		do_hard_irq_enable();
+	}
 
 #if defined(CONFIG_PPC32) && defined(CONFIG_PPC_PMAC)
 	if (atomic_read(&ppc_n_lost_interrupts) != 0)

@@ -111,6 +111,8 @@ name:
 #define IAREA		.L_IAREA_\name\()	/* PACA save area */
 #define IVIRT		.L_IVIRT_\name\()	/* Has virt mode entry point */
 #define IISIDE		.L_IISIDE_\name\()	/* Uses SRR0/1 not DAR/DSISR */
+#define ICFAR		.L_ICFAR_\name\()	/* Uses CFAR */
+#define ICFAR_IF_HVMODE	.L_ICFAR_IF_HVMODE_\name\() /* Uses CFAR if HV */
 #define IDAR		.L_IDAR_\name\()	/* Uses DAR (or SRR0) */
 #define IDSISR		.L_IDSISR_\name\()	/* Uses DSISR (or SRR1) */
 #define IBRANCH_TO_COMMON	.L_IBRANCH_TO_COMMON_\name\() /* ENTRY branch to common */

@@ -150,6 +152,12 @@ do_define_int n
 	.ifndef IISIDE
 		IISIDE=0
 	.endif
+	.ifndef ICFAR
+		ICFAR=1
+	.endif
+	.ifndef ICFAR_IF_HVMODE
+		ICFAR_IF_HVMODE=0
+	.endif
 	.ifndef IDAR
 		IDAR=0
 	.endif

@@ -287,9 +295,21 @@ BEGIN_FTR_SECTION
 END_FTR_SECTION_IFSET(CPU_FTR_HAS_PPR)
 	HMT_MEDIUM
 	std	r10,IAREA+EX_R10(r13)		/* save r10 - r12 */
+	.if ICFAR
 BEGIN_FTR_SECTION
 	mfspr	r10,SPRN_CFAR
 END_FTR_SECTION_IFSET(CPU_FTR_CFAR)
+	.elseif ICFAR_IF_HVMODE
+BEGIN_FTR_SECTION
+  BEGIN_FTR_SECTION_NESTED(69)
+	mfspr	r10,SPRN_CFAR
+  END_FTR_SECTION_NESTED(CPU_FTR_CFAR, CPU_FTR_CFAR, 69)
+FTR_SECTION_ELSE
+  BEGIN_FTR_SECTION_NESTED(69)
+	li	r10,0
+  END_FTR_SECTION_NESTED(CPU_FTR_CFAR, CPU_FTR_CFAR, 69)
+ALT_FTR_SECTION_END_IFSET(CPU_FTR_HVMODE | CPU_FTR_ARCH_206)
+	.endif
 	.if \ool
 	.if !\virt
 	b	tramp_real_\name

@@ -305,9 +325,11 @@ END_FTR_SECTION_IFSET(CPU_FTR_CFAR)
 BEGIN_FTR_SECTION
 	std	r9,IAREA+EX_PPR(r13)
 END_FTR_SECTION_IFSET(CPU_FTR_HAS_PPR)
+	.if ICFAR || ICFAR_IF_HVMODE
 BEGIN_FTR_SECTION
 	std	r10,IAREA+EX_CFAR(r13)
 END_FTR_SECTION_IFSET(CPU_FTR_CFAR)
+	.endif
 	INTERRUPT_TO_KERNEL
 	mfctr	r10
 	std	r10,IAREA+EX_CTR(r13)

@@ -559,7 +581,11 @@ END_FTR_SECTION_IFSET(CPU_FTR_HAS_PPR)
 	.endif
 
 BEGIN_FTR_SECTION
+	.if ICFAR || ICFAR_IF_HVMODE
 	ld	r10,IAREA+EX_CFAR(r13)
+	.else
+	li	r10,0
+	.endif
 	std	r10,ORIG_GPR3(r1)
 END_FTR_SECTION_IFSET(CPU_FTR_CFAR)
 	ld	r10,IAREA+EX_CTR(r13)

@@ -1501,6 +1527,12 @@ ALT_MMU_FTR_SECTION_END_IFCLR(MMU_FTR_TYPE_RADIX)
  *
  * If soft masked, the masked handler will note the pending interrupt for
  * replay, and clear MSR[EE] in the interrupted context.
+ *
+ * CFAR is not required because this is an asynchronous interrupt that in
+ * general won't have much bearing on the state of the CPU, with the possible
+ * exception of crash/debug IPIs, but those are generally moving to use SRESET
+ * IPIs. Unless this is an HV interrupt and KVM HV is possible, in which case
+ * it may be exiting the guest and need CFAR to be saved.
  */
 INT_DEFINE_BEGIN(hardware_interrupt)
 	IVEC=0x500

@@ -1508,6 +1540,10 @@ INT_DEFINE_BEGIN(hardware_interrupt)
 	IMASK=IRQS_DISABLED
 	IKVM_REAL=1
 	IKVM_VIRT=1
+	ICFAR=0
+#ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
+	ICFAR_IF_HVMODE=1
+#endif
 INT_DEFINE_END(hardware_interrupt)
 
 EXC_REAL_BEGIN(hardware_interrupt, 0x500, 0x100)

@@ -1726,6 +1762,10 @@ END_FTR_SECTION_IFSET(CPU_FTR_TM)
  * If PPC_WATCHDOG is configured, the soft masked handler will actually set
  * things back up to run soft_nmi_interrupt as a regular interrupt handler
  * on the emergency stack.
+ *
+ * CFAR is not required because this is asynchronous (see hardware_interrupt).
+ * A watchdog interrupt may like to have CFAR, but usually the interesting
+ * branch is long gone by that point (e.g., infinite loop).
  */
 INT_DEFINE_BEGIN(decrementer)
 	IVEC=0x900

@@ -1733,6 +1773,7 @@ INT_DEFINE_BEGIN(decrementer)
 #ifdef CONFIG_KVM_BOOK3S_PR_POSSIBLE
 	IKVM_REAL=1
 #endif
+	ICFAR=0
 INT_DEFINE_END(decrementer)
 
 EXC_REAL_BEGIN(decrementer, 0x900, 0x80)

@@ -1808,6 +1849,8 @@ EXC_COMMON_BEGIN(hdecrementer_common)
  * If soft masked, the masked handler will note the pending interrupt for
  * replay, leaving MSR[EE] enabled in the interrupted context because the
  * doorbells are edge triggered.
+ *
+ * CFAR is not required, similarly to hardware_interrupt.
  */
 INT_DEFINE_BEGIN(doorbell_super)
 	IVEC=0xa00

@@ -1815,6 +1858,7 @@ INT_DEFINE_BEGIN(doorbell_super)
 #ifdef CONFIG_KVM_BOOK3S_PR_POSSIBLE
 	IKVM_REAL=1
 #endif
+	ICFAR=0
 INT_DEFINE_END(doorbell_super)
 
 EXC_REAL_BEGIN(doorbell_super, 0xa00, 0x100)

@@ -1866,6 +1910,7 @@ INT_DEFINE_BEGIN(system_call)
 	IVEC=0xc00
 	IKVM_REAL=1
 	IKVM_VIRT=1
+	ICFAR=0
 INT_DEFINE_END(system_call)
 
 .macro SYSTEM_CALL virt

@@ -2164,6 +2209,11 @@ EXC_COMMON_BEGIN(hmi_exception_common)
  * Interrupt 0xe80 - Directed Hypervisor Doorbell Interrupt.
  * This is an asynchronous interrupt in response to a msgsnd doorbell.
  * Similar to the 0xa00 doorbell but for host rather than guest.
+ *
+ * CFAR is not required (similar to doorbell_interrupt), unless KVM HV
+ * is enabled, in which case it may be a guest exit. Most PowerNV kernels
+ * include KVM support so it would be nice if this could be dynamically
+ * patched out if KVM was not currently running any guests.
  */
 INT_DEFINE_BEGIN(h_doorbell)
 	IVEC=0xe80

@@ -2171,6 +2221,9 @@ INT_DEFINE_BEGIN(h_doorbell)
 	IMASK=IRQS_DISABLED
 	IKVM_REAL=1
 	IKVM_VIRT=1
+#ifndef CONFIG_KVM_BOOK3S_HV_POSSIBLE
+	ICFAR=0
+#endif
 INT_DEFINE_END(h_doorbell)
 
 EXC_REAL_BEGIN(h_doorbell, 0xe80, 0x20)

@@ -2194,6 +2247,9 @@ EXC_COMMON_BEGIN(h_doorbell_common)
  * Interrupt 0xea0 - Hypervisor Virtualization Interrupt.
  * This is an asynchronous interrupt in response to an "external exception".
  * Similar to 0x500 but for host only.
+ *
+ * Like h_doorbell, CFAR is only required for KVM HV because this can be
+ * a guest exit.
  */
 INT_DEFINE_BEGIN(h_virt_irq)
 	IVEC=0xea0

@@ -2201,6 +2257,9 @@ INT_DEFINE_BEGIN(h_virt_irq)
 	IMASK=IRQS_DISABLED
 	IKVM_REAL=1
 	IKVM_VIRT=1
+#ifndef CONFIG_KVM_BOOK3S_HV_POSSIBLE
+	ICFAR=0
+#endif
 INT_DEFINE_END(h_virt_irq)
 
 EXC_REAL_BEGIN(h_virt_irq, 0xea0, 0x20)

@@ -2237,6 +2296,8 @@ EXC_VIRT_NONE(0x4ee0, 0x20)
  *
  * If soft masked, the masked handler will note the pending interrupt for
  * replay, and clear MSR[EE] in the interrupted context.
+ *
+ * CFAR is not used by perf interrupts so not required.
  */
 INT_DEFINE_BEGIN(performance_monitor)
 	IVEC=0xf00

@@ -2244,6 +2305,7 @@ INT_DEFINE_BEGIN(performance_monitor)
 #ifdef CONFIG_KVM_BOOK3S_PR_POSSIBLE
 	IKVM_REAL=1
 #endif
+	ICFAR=0
 INT_DEFINE_END(performance_monitor)
 
 EXC_REAL_BEGIN(performance_monitor, 0xf00, 0x20)

@@ -2668,6 +2730,7 @@ EXC_VIRT_NONE(0x5800, 0x100)
 INT_DEFINE_BEGIN(soft_nmi)
 	IVEC=0x900
 	ISTACK=0
+	ICFAR=0
 INT_DEFINE_END(soft_nmi)
 
 /*