diff --git a/include/arch/aarch64/arch.h b/include/arch/aarch64/arch.h
index fa857fb..98a8590 100644
--- a/include/arch/aarch64/arch.h
+++ b/include/arch/aarch64/arch.h
@@ -132,12 +132,13 @@
 #define ID_AA64PFR0_EL2_SHIFT	U(8)
 #define ID_AA64PFR0_EL3_SHIFT	U(12)
 #define ID_AA64PFR0_AMU_SHIFT	U(44)
-#define ID_AA64PFR0_AMU_LENGTH	U(4)
 #define ID_AA64PFR0_AMU_MASK	ULL(0xf)
 #define ID_AA64PFR0_ELX_MASK	ULL(0xf)
+#define ID_AA64PFR0_GIC_SHIFT	U(24)
+#define ID_AA64PFR0_GIC_WIDTH	U(4)
+#define ID_AA64PFR0_GIC_MASK	ULL(0xf)
 #define ID_AA64PFR0_SVE_SHIFT	U(32)
 #define ID_AA64PFR0_SVE_MASK	ULL(0xf)
-#define ID_AA64PFR0_SVE_LENGTH	U(4)
 #define ID_AA64PFR0_MPAM_SHIFT	U(40)
 #define ID_AA64PFR0_MPAM_MASK	ULL(0xf)
 #define ID_AA64PFR0_DIT_SHIFT	U(48)
@@ -148,18 +149,14 @@
 #define ID_AA64PFR0_CSV2_MASK	ULL(0xf)
 #define ID_AA64PFR0_CSV2_LENGTH	U(4)
 
-/* ID_AA64DFR0_EL1.PMS definitions (for ARMv8.2+) */
-#define ID_AA64DFR0_PMS_SHIFT	U(32)
-#define ID_AA64DFR0_PMS_LENGTH	U(4)
-#define ID_AA64DFR0_PMS_MASK	ULL(0xf)
-
+/* Exception level handling */
 #define EL_IMPL_NONE		ULL(0)
 #define EL_IMPL_A64ONLY		ULL(1)
 #define EL_IMPL_A64_A32		ULL(2)
 
-#define ID_AA64PFR0_GIC_SHIFT	U(24)
-#define ID_AA64PFR0_GIC_WIDTH	U(4)
-#define ID_AA64PFR0_GIC_MASK	ULL(0xf)
+/* ID_AA64DFR0_EL1.PMS definitions (for ARMv8.2+) */
+#define ID_AA64DFR0_PMS_SHIFT	U(32)
+#define ID_AA64DFR0_PMS_MASK	ULL(0xf)
 
 /* ID_AA64ISAR1_EL1 definitions */
 #define ID_AA64ISAR1_EL1	S3_0_C0_C6_1
@@ -304,20 +301,25 @@
 #define SCR_RESET_VAL		SCR_RES1_BITS
 
 /* MDCR_EL3 definitions */
+#define MDCR_SCCD_BIT		(ULL(1) << 23)
+#define MDCR_SPME_BIT		(ULL(1) << 17)
+#define MDCR_SDD_BIT		(ULL(1) << 16)
 #define MDCR_SPD32(x)		((x) << 14)
 #define MDCR_SPD32_LEGACY	ULL(0x0)
 #define MDCR_SPD32_DISABLE	ULL(0x2)
 #define MDCR_SPD32_ENABLE	ULL(0x3)
-#define MDCR_SDD_BIT		(ULL(1) << 16)
 #define MDCR_NSPB(x)		((x) << 12)
 #define MDCR_NSPB_EL1		ULL(0x3)
 #define MDCR_TDOSA_BIT		(ULL(1) << 10)
 #define MDCR_TDA_BIT		(ULL(1) << 9)
 #define MDCR_TPM_BIT		(ULL(1) << 6)
-#define MDCR_SCCD_BIT		(ULL(1) << 23)
 #define MDCR_EL3_RESET_VAL	ULL(0x0)
 
 /* MDCR_EL2 definitions */
+#define MDCR_EL2_HLP		(U(1) << 26)
+#define MDCR_EL2_HCCD		(U(1) << 23)
+#define MDCR_EL2_TTRF		(U(1) << 19)
+#define MDCR_EL2_HPMD		(U(1) << 17)
 #define MDCR_EL2_TPMS		(U(1) << 14)
 #define MDCR_EL2_E2PB(x)	((x) << 12)
 #define MDCR_EL2_E2PB_EL1	U(0x3)
@@ -677,10 +679,14 @@
 #define PMCR_EL0_N_SHIFT	U(11)
 #define PMCR_EL0_N_MASK		U(0x1f)
 #define PMCR_EL0_N_BITS		(PMCR_EL0_N_MASK << PMCR_EL0_N_SHIFT)
+#define PMCR_EL0_LP_BIT		(U(1) << 7)
 #define PMCR_EL0_LC_BIT		(U(1) << 6)
 #define PMCR_EL0_DP_BIT		(U(1) << 5)
 #define PMCR_EL0_X_BIT		(U(1) << 4)
 #define PMCR_EL0_D_BIT		(U(1) << 3)
+#define PMCR_EL0_C_BIT		(U(1) << 2)
+#define PMCR_EL0_P_BIT		(U(1) << 1)
+#define PMCR_EL0_E_BIT		(U(1) << 0)
 
 /*******************************************************************************
  * Definitions for system register interface to SVE
diff --git a/include/arch/aarch64/el3_common_macros.S b/include/arch/aarch64/el3_common_macros.S
index 22b32b4..a36b7da 100644
--- a/include/arch/aarch64/el3_common_macros.S
+++ b/include/arch/aarch64/el3_common_macros.S
@@ -116,12 +116,42 @@
 	 * ---------------------------------------------------------------------
 	 */
 	mov_imm	x0, ((MDCR_EL3_RESET_VAL | MDCR_SDD_BIT | \
-		      MDCR_SPD32(MDCR_SPD32_DISABLE) | MDCR_SCCD_BIT) \
-		    & ~(MDCR_TDOSA_BIT | MDCR_TDA_BIT | MDCR_TPM_BIT))
+		      MDCR_SPD32(MDCR_SPD32_DISABLE) | MDCR_SCCD_BIT) & \
+		    ~(MDCR_TDOSA_BIT | MDCR_TDA_BIT | MDCR_TPM_BIT))
 
 	msr	mdcr_el3, x0
 
 	/* ---------------------------------------------------------------------
+	 * Initialise PMCR_EL0 setting all fields rather than relying
+	 * on hw. Some fields are architecturally UNKNOWN on reset.
+	 *
+	 * PMCR_EL0.LP: Set to one so that event counter overflow, that
+	 *  is recorded in PMOVSCLR_EL0[0-30], occurs on the increment
+	 *  that changes PMEVCNTR<n>_EL0[63] from 1 to 0, when ARMv8.5-PMU
+	 *  is implemented. This bit is RES0 in versions of the architecture
+	 *  earlier than ARMv8.5, setting it to 1 doesn't have any effect
+	 *  on them.
+	 *
+	 * PMCR_EL0.LC: Set to one so that cycle counter overflow, that
+	 *  is recorded in PMOVSCLR_EL0[31], occurs on the increment
+	 *  that changes PMCCNTR_EL0[63] from 1 to 0.
+	 *
+	 * PMCR_EL0.DP: Set to one so that the cycle counter,
+	 *  PMCCNTR_EL0 does not count when event counting is prohibited.
+	 *
+	 * PMCR_EL0.X: Set to zero to disable export of events.
+	 *
+	 * PMCR_EL0.D: Set to zero so that, when enabled, PMCCNTR_EL0
+	 *  counts on every clock cycle.
+	 * ---------------------------------------------------------------------
+	 */
+	mov_imm	x0, ((PMCR_EL0_RESET_VAL | PMCR_EL0_LP_BIT | \
+		      PMCR_EL0_LC_BIT | PMCR_EL0_DP_BIT) & \
+		    ~(PMCR_EL0_X_BIT | PMCR_EL0_D_BIT))
+
+	msr	pmcr_el0, x0
+
+	/* ---------------------------------------------------------------------
 	 * Enable External Aborts and SError Interrupts now that the exception
 	 * vectors have been setup.
 	 * ---------------------------------------------------------------------
diff --git a/include/lib/el3_runtime/aarch64/context.h b/include/lib/el3_runtime/aarch64/context.h
index a76a59b..64fa8a9 100644
--- a/include/lib/el3_runtime/aarch64/context.h
+++ b/include/lib/el3_runtime/aarch64/context.h
@@ -59,7 +59,7 @@
 #define CTX_RUNTIME_SP		U(0x10)
 #define CTX_SPSR_EL3		U(0x18)
 #define CTX_ELR_EL3		U(0x20)
-#define CTX_UNUSED		U(0x28)
+#define CTX_PMCR_EL0		U(0x28)
 #define CTX_EL3STATE_END	U(0x30)
 
 /*******************************************************************************
@@ -91,22 +91,21 @@
 #define CTX_AFSR1_EL1		U(0x98)
 #define CTX_CONTEXTIDR_EL1	U(0xa0)
 #define CTX_VBAR_EL1		U(0xa8)
-#define CTX_PMCR_EL0		U(0xb0)
 
 /*
  * If the platform is AArch64-only, there is no need to save and restore these
  * AArch32 registers.
  */
 #if CTX_INCLUDE_AARCH32_REGS
-#define CTX_SPSR_ABT		U(0xc0)  /* Align to the next 16 byte boundary */
-#define CTX_SPSR_UND		U(0xc8)
-#define CTX_SPSR_IRQ		U(0xd0)
-#define CTX_SPSR_FIQ		U(0xd8)
-#define CTX_DACR32_EL2		U(0xe0)
-#define CTX_IFSR32_EL2		U(0xe8)
-#define CTX_AARCH32_END		U(0xf0) /* Align to the next 16 byte boundary */
+#define CTX_SPSR_ABT		U(0xb0)	/* Align to the next 16 byte boundary */
+#define CTX_SPSR_UND		U(0xb8)
+#define CTX_SPSR_IRQ		U(0xc0)
+#define CTX_SPSR_FIQ		U(0xc8)
+#define CTX_DACR32_EL2		U(0xd0)
+#define CTX_IFSR32_EL2		U(0xd8)
+#define CTX_AARCH32_END		U(0xe0) /* Align to the next 16 byte boundary */
 #else
-#define CTX_AARCH32_END		U(0xc0)  /* Align to the next 16 byte boundary */
+#define CTX_AARCH32_END		U(0xb0)	/* Align to the next 16 byte boundary */
 #endif /* CTX_INCLUDE_AARCH32_REGS */
 
 /*