LoongArch: KVM: Add PMU support for guest

On LoongArch, the host and guest have their own PMU CSRs registers and
they share PMU hardware resources. A set of PMU CSRs consists of a CTRL
register and a CNTR register. We can set which PMU CSRs are used by the
guest by writing to the GCFG register [24:26] bits.

On KVM side:
- Save the host PMU CSRs into structure kvm_context.
- If the host supports the PMU feature.
  - When entering guest mode, save the host PMU CSRs and restore the guest PMU CSRs.
  - When exiting guest mode, save the guest PMU CSRs and restore the host PMU CSRs.
Reviewed-by: Bibo Mao <maobibo@loongson.cn>
Signed-off-by: Song Gao <gaosong@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>

arch/loongarch/include/asm/kvm_csr.h

@@ -30,6 +30,7 @@
 		: [val] "+r" (__v)				\
 		: [reg] "i" (csr)				\
 		: "memory");					\
+	__v;							\
 })
 
 #define gcsr_xchg(v, m, csr)					\
@@ -181,6 +182,8 @@ __BUILD_GCSR_OP(tlbidx)
 #define kvm_save_hw_gcsr(csr, gid)	(csr->csrs[gid] = gcsr_read(gid))
 #define kvm_restore_hw_gcsr(csr, gid)	(gcsr_write(csr->csrs[gid], gid))
 
+#define kvm_read_clear_hw_gcsr(csr, gid)	(csr->csrs[gid] = gcsr_write(0, gid))
+
 int kvm_emu_iocsr(larch_inst inst, struct kvm_run *run, struct kvm_vcpu *vcpu);
 
 static __always_inline unsigned long kvm_read_sw_gcsr(struct loongarch_csrs *csr, int gid)
@@ -208,4 +211,7 @@ static __always_inline void kvm_change_sw_gcsr(struct loongarch_csrs *csr,
 	csr->csrs[gid] |= val & _mask;
 }
 
+#define KVM_PMU_EVENT_ENABLED	(CSR_PERFCTRL_PLV0 | CSR_PERFCTRL_PLV1 | \
+					CSR_PERFCTRL_PLV2 | CSR_PERFCTRL_PLV3)
+
 #endif	/* __ASM_LOONGARCH_KVM_CSR_H__ */

arch/loongarch/include/asm/kvm_host.h

@@ -30,6 +30,7 @@
 #define KVM_HALT_POLL_NS_DEFAULT	500000
 #define KVM_REQ_TLB_FLUSH_GPA		KVM_ARCH_REQ(0)
 #define KVM_REQ_STEAL_UPDATE		KVM_ARCH_REQ(1)
+#define KVM_REQ_PMU			KVM_ARCH_REQ(2)
 
 #define KVM_GUESTDBG_SW_BP_MASK		\
 	(KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP)
@@ -60,9 +61,13 @@ struct kvm_arch_memory_slot {
 	unsigned long flags;
 };
 
+#define HOST_MAX_PMNUM			16
 struct kvm_context {
 	unsigned long vpid_cache;
 	struct kvm_vcpu *last_vcpu;
+	/* Host PMU CSR */
+	u64 perf_ctrl[HOST_MAX_PMNUM];
+	u64 perf_cntr[HOST_MAX_PMNUM];
 };
 
 struct kvm_world_switch {
@@ -134,8 +139,9 @@ enum emulation_result {
 #define KVM_LARCH_LSX		(0x1 << 1)
 #define KVM_LARCH_LASX		(0x1 << 2)
 #define KVM_LARCH_LBT		(0x1 << 3)
-#define KVM_LARCH_SWCSR_LATEST	(0x1 << 4)
-#define KVM_LARCH_HWCSR_USABLE	(0x1 << 5)
+#define KVM_LARCH_PMU		(0x1 << 4)
+#define KVM_LARCH_SWCSR_LATEST	(0x1 << 5)
+#define KVM_LARCH_HWCSR_USABLE	(0x1 << 6)
 
 struct kvm_vcpu_arch {
 	/*
@@ -174,6 +180,9 @@ struct kvm_vcpu_arch {
 	/* CSR state */
 	struct loongarch_csrs *csr;
 
+	/* Guest max PMU CSR id */
+	int max_pmu_csrid;
+
 	/* GPR used as IO source/target */
 	u32 io_gpr;
 
@@ -246,6 +255,16 @@ static inline bool kvm_guest_has_lbt(struct kvm_vcpu_arch *arch)
 	return arch->cpucfg[2] & (CPUCFG2_X86BT | CPUCFG2_ARMBT | CPUCFG2_MIPSBT);
 }
 
+static inline bool kvm_guest_has_pmu(struct kvm_vcpu_arch *arch)
+{
+	return arch->cpucfg[6] & CPUCFG6_PMP;
+}
+
+static inline int kvm_get_pmu_num(struct kvm_vcpu_arch *arch)
+{
+	return (arch->cpucfg[6] & CPUCFG6_PMNUM) >> CPUCFG6_PMNUM_SHIFT;
+}
+
 /* Debug: dump vcpu state */
 int kvm_arch_vcpu_dump_regs(struct kvm_vcpu *vcpu);
 

arch/loongarch/include/asm/loongarch.h

@@ -119,6 +119,7 @@
 #define  CPUCFG6_PMP			BIT(0)
 #define  CPUCFG6_PAMVER			GENMASK(3, 1)
 #define  CPUCFG6_PMNUM			GENMASK(7, 4)
+#define  CPUCFG6_PMNUM_SHIFT		4
 #define  CPUCFG6_PMBITS			GENMASK(13, 8)
 #define  CPUCFG6_UPM			BIT(14)
 

arch/loongarch/include/uapi/asm/kvm.h

@@ -98,6 +98,7 @@ struct kvm_fpu {
 #define  KVM_LOONGARCH_VM_FEAT_X86BT	2
 #define  KVM_LOONGARCH_VM_FEAT_ARMBT	3
 #define  KVM_LOONGARCH_VM_FEAT_MIPSBT	4
+#define  KVM_LOONGARCH_VM_FEAT_PMU	5
 
 /* Device Control API on vcpu fd */
 #define KVM_LOONGARCH_VCPU_CPUCFG	0

arch/loongarch/kvm/exit.c

@@ -127,6 +127,14 @@ static int kvm_handle_csr(struct kvm_vcpu *vcpu, larch_inst inst)
 	rj = inst.reg2csr_format.rj;
 	csrid = inst.reg2csr_format.csr;
 
+	if (csrid >= LOONGARCH_CSR_PERFCTRL0 && csrid <= vcpu->arch.max_pmu_csrid) {
+		if (kvm_guest_has_pmu(&vcpu->arch)) {
+			vcpu->arch.pc -= 4;
+			kvm_make_request(KVM_REQ_PMU, vcpu);
+			return EMULATE_DONE;
+		}
+	}
+
 	/* Process CSR ops */
 	switch (rj) {
 	case 0: /* process csrrd */

arch/loongarch/kvm/vcpu.c

@@ -32,6 +32,126 @@ const struct kvm_stats_header kvm_vcpu_stats_header = {
 		       sizeof(kvm_vcpu_stats_desc),
 };
 
+static inline void kvm_save_host_pmu(struct kvm_vcpu *vcpu)
+{
+	struct kvm_context *context;
+
+	context = this_cpu_ptr(vcpu->kvm->arch.vmcs);
+	context->perf_cntr[0] = read_csr_perfcntr0();
+	context->perf_cntr[1] = read_csr_perfcntr1();
+	context->perf_cntr[2] = read_csr_perfcntr2();
+	context->perf_cntr[3] = read_csr_perfcntr3();
+	context->perf_ctrl[0] = write_csr_perfctrl0(0);
+	context->perf_ctrl[1] = write_csr_perfctrl1(0);
+	context->perf_ctrl[2] = write_csr_perfctrl2(0);
+	context->perf_ctrl[3] = write_csr_perfctrl3(0);
+}
+
+static inline void kvm_restore_host_pmu(struct kvm_vcpu *vcpu)
+{
+	struct kvm_context *context;
+
+	context = this_cpu_ptr(vcpu->kvm->arch.vmcs);
+	write_csr_perfcntr0(context->perf_cntr[0]);
+	write_csr_perfcntr1(context->perf_cntr[1]);
+	write_csr_perfcntr2(context->perf_cntr[2]);
+	write_csr_perfcntr3(context->perf_cntr[3]);
+	write_csr_perfctrl0(context->perf_ctrl[0]);
+	write_csr_perfctrl1(context->perf_ctrl[1]);
+	write_csr_perfctrl2(context->perf_ctrl[2]);
+	write_csr_perfctrl3(context->perf_ctrl[3]);
+}
+
+
+static inline void kvm_save_guest_pmu(struct kvm_vcpu *vcpu)
+{
+	struct loongarch_csrs *csr = vcpu->arch.csr;
+
+	kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR0);
+	kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR1);
+	kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR2);
+	kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR3);
+	kvm_read_clear_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0);
+	kvm_read_clear_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1);
+	kvm_read_clear_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2);
+	kvm_read_clear_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3);
+}
+
+static inline void kvm_restore_guest_pmu(struct kvm_vcpu *vcpu)
+{
+	struct loongarch_csrs *csr = vcpu->arch.csr;
+
+	kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR0);
+	kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR1);
+	kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR2);
+	kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR3);
+	kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0);
+	kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1);
+	kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2);
+	kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3);
+}
+
+static int kvm_own_pmu(struct kvm_vcpu *vcpu)
+{
+	unsigned long val;
+
+	if (!kvm_guest_has_pmu(&vcpu->arch))
+		return -EINVAL;
+
+	kvm_save_host_pmu(vcpu);
+
+	/* Set PM0-PM(num) to guest */
+	val = read_csr_gcfg() & ~CSR_GCFG_GPERF;
+	val |= (kvm_get_pmu_num(&vcpu->arch) + 1) << CSR_GCFG_GPERF_SHIFT;
+	write_csr_gcfg(val);
+
+	kvm_restore_guest_pmu(vcpu);
+
+	return 0;
+}
+
+static void kvm_lose_pmu(struct kvm_vcpu *vcpu)
+{
+	unsigned long val;
+	struct loongarch_csrs *csr = vcpu->arch.csr;
+
+	if (!(vcpu->arch.aux_inuse & KVM_LARCH_PMU))
+		return;
+
+	kvm_save_guest_pmu(vcpu);
+
+	/* Disable pmu access from guest */
+	write_csr_gcfg(read_csr_gcfg() & ~CSR_GCFG_GPERF);
+
+	/*
+	 * Clear KVM_LARCH_PMU if the guest is not using PMU CSRs when
+	 * exiting the guest, so that the next time trap into the guest.
+	 * We don't need to deal with PMU CSRs contexts.
+	 */
+	val = kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0);
+	val |= kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1);
+	val |= kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2);
+	val |= kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3);
+	if (!(val & KVM_PMU_EVENT_ENABLED))
+		vcpu->arch.aux_inuse &= ~KVM_LARCH_PMU;
+
+	kvm_restore_host_pmu(vcpu);
+}
+
+static void kvm_restore_pmu(struct kvm_vcpu *vcpu)
+{
+	if ((vcpu->arch.aux_inuse & KVM_LARCH_PMU))
+		kvm_make_request(KVM_REQ_PMU, vcpu);
+}
+
+static void kvm_check_pmu(struct kvm_vcpu *vcpu)
+{
+	if (kvm_check_request(KVM_REQ_PMU, vcpu)) {
+		kvm_own_pmu(vcpu);
+		vcpu->arch.aux_inuse |= KVM_LARCH_PMU;
+	}
+}
+
 static void kvm_update_stolen_time(struct kvm_vcpu *vcpu)
 {
 	u32 version;
@@ -159,6 +279,7 @@ static int kvm_pre_enter_guest(struct kvm_vcpu *vcpu)
 		/* Make sure the vcpu mode has been written */
 		smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 		kvm_check_vpid(vcpu);
+		kvm_check_pmu(vcpu);
 
 		/*
 		 * Called after function kvm_check_vpid()
@@ -196,6 +317,8 @@ static int kvm_handle_exit(struct kvm_run *run, struct kvm_vcpu *vcpu)
 	/* Set a default exit reason */
 	run->exit_reason = KVM_EXIT_UNKNOWN;
 
+	kvm_lose_pmu(vcpu);
+
 	guest_timing_exit_irqoff();
 	guest_state_exit_irqoff();
 	local_irq_enable();
@@ -469,6 +592,22 @@ static int _kvm_setcsr(struct kvm_vcpu *vcpu, unsigned int id, u64 val)
 
 	kvm_write_sw_gcsr(csr, id, val);
 
+	/*
+	 * After modifying the PMU CSR register value of the vcpu.
+	 * If the PMU CSRs are used, we need to set KVM_REQ_PMU.
+	 */
+	if (id >= LOONGARCH_CSR_PERFCTRL0 && id <= LOONGARCH_CSR_PERFCNTR3) {
+		unsigned long val;
+
+		val = kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0) |
+		      kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1) |
+		      kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2) |
+		      kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3);
+
+		if (val & KVM_PMU_EVENT_ENABLED)
+			kvm_make_request(KVM_REQ_PMU, vcpu);
+	}
+
 	return ret;
 }
 
@@ -513,6 +652,12 @@ static int _kvm_get_cpucfg_mask(int id, u64 *v)
 	case LOONGARCH_CPUCFG5:
 		*v = GENMASK(31, 0);
 		return 0;
+	case LOONGARCH_CPUCFG6:
+		if (cpu_has_pmp)
+			*v = GENMASK(14, 0);
+		else
+			*v = 0;
+		return 0;
 	case LOONGARCH_CPUCFG16:
 		*v = GENMASK(16, 0);
 		return 0;
@@ -557,6 +702,17 @@ static int kvm_check_cpucfg(int id, u64 val)
 			/* LASX architecturally implies LSX and FP but val does not satisfy that */
 			return -EINVAL;
 		return 0;
+	case LOONGARCH_CPUCFG6:
+		if (val & CPUCFG6_PMP) {
+			u32 host = read_cpucfg(LOONGARCH_CPUCFG6);
+			if ((val & CPUCFG6_PMBITS) != (host & CPUCFG6_PMBITS))
+				return -EINVAL;
+			if ((val & CPUCFG6_PMNUM) > (host & CPUCFG6_PMNUM))
+				return -EINVAL;
+			if ((val & CPUCFG6_UPM) && !(host & CPUCFG6_UPM))
+				return -EINVAL;
+		}
+		return 0;
 	default:
 		/*
 		 * Values for the other CPUCFG IDs are not being further validated
@@ -670,6 +826,9 @@ static int kvm_set_one_reg(struct kvm_vcpu *vcpu,
 		if (ret)
 			break;
 		vcpu->arch.cpucfg[id] = (u32)v;
+		if (id == LOONGARCH_CPUCFG6)
+			vcpu->arch.max_pmu_csrid =
+				LOONGARCH_CSR_PERFCTRL0 + 2 * kvm_get_pmu_num(&vcpu->arch) + 1;
 		break;
 	case KVM_REG_LOONGARCH_LBT:
 		if (!kvm_guest_has_lbt(&vcpu->arch))
@@ -791,7 +950,8 @@ static int kvm_loongarch_cpucfg_has_attr(struct kvm_vcpu *vcpu,
 					 struct kvm_device_attr *attr)
 {
 	switch (attr->attr) {
-	case 2:
+	case LOONGARCH_CPUCFG2:
+	case LOONGARCH_CPUCFG6:
 		return 0;
 	default:
 		return -ENXIO;
@@ -1356,6 +1516,9 @@ static int _kvm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	change_csr_gcfg(CSR_GCFG_MATC_MASK, CSR_GCFG_MATC_ROOT);
 	kvm_make_request(KVM_REQ_STEAL_UPDATE, vcpu);
 
+	/* Restore hardware PMU CSRs */
+	kvm_restore_pmu(vcpu);
+
 	/* Don't bother restoring registers multiple times unless necessary */
 	if (vcpu->arch.aux_inuse & KVM_LARCH_HWCSR_USABLE)
 		return 0;

arch/loongarch/kvm/vm.c

@@ -122,6 +122,10 @@ static int kvm_vm_feature_has_attr(struct kvm *kvm, struct kvm_device_attr *attr
 		if (cpu_has_lbt_mips)
 			return 0;
 		return -ENXIO;
+	case KVM_LOONGARCH_VM_FEAT_PMU:
+		if (cpu_has_pmp)
+			return 0;
+		return -ENXIO;
 	default:
 		return -ENXIO;
 	}