Merge pull request #1197 from dp-arm/dp/amu

AMUv1 support

16 files changed, 858 insertions, 38 deletions

diff --git a/bl31/bl31.mk b/bl31/bl31.mk
index 8ff8f89b..2db48564 100644
--- a/bl31/bl31.mk
+++ b/bl31/bl31.mk
@@ -51,7 +51,8 @@ BL31_SOURCES		+=	lib/extensions/spe/spe.c
 endif
 
 ifeq (${ENABLE_AMU},1)
-BL31_SOURCES		+=	lib/extensions/amu/aarch64/amu.c
+BL31_SOURCES		+=	lib/extensions/amu/aarch64/amu.c		\
+				lib/extensions/amu/aarch64/amu_helpers.S
 endif
 
 ifeq (${ENABLE_SVE_FOR_NS},1)
diff --git a/docs/porting-guide.rst b/docs/porting-guide.rst
index f020ec97..51d2e64a 100644
--- a/docs/porting-guide.rst
+++ b/docs/porting-guide.rst
@@ -549,6 +549,22 @@ behaviour of the ``assert()`` function (for example, to save memory).
    doesn't print anything to the console. If ``PLAT_LOG_LEVEL_ASSERT`` isn't
    defined, it defaults to ``LOG_LEVEL``.
 
+If the platform port uses the Activity Monitor Unit, the following constants
+may be defined:
+
+-  **PLAT\_AMU\_GROUP1\_COUNTERS\_MASK**
+   This mask reflects the set of group counters that should be enabled.  The
+   maximum number of group 1 counters supported by AMUv1 is 16 so the mask
+   can be at most 0xffff. If the platform does not define this mask, no group 1
+   counters are enabled. If the platform defines this mask, the following
+   constant needs to also be defined.
+
+-  **PLAT\_AMU\_GROUP1\_NR\_COUNTERS**
+   This value is used to allocate an array to save and restore the counters
+   specified by ``PLAT_AMU_GROUP1_COUNTERS_MASK`` on CPU suspend.
+   This value should be equal to the highest bit position set in the
+   mask, plus 1.  The maximum number of group 1 counters in AMUv1 is 16.
+
 File : plat\_macros.S [mandatory]
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
diff --git a/include/lib/aarch32/arch_helpers.h b/include/lib/aarch32/arch_helpers.h
index 0230195a..beae5d06 100644
--- a/include/lib/aarch32/arch_helpers.h
+++ b/include/lib/aarch32/arch_helpers.h
@@ -287,6 +287,11 @@ DEFINE_COPROCR_RW_FUNCS(amcntenset1, AMCNTENSET1)
 DEFINE_COPROCR_RW_FUNCS(amcntenclr0, AMCNTENCLR0)
 DEFINE_COPROCR_RW_FUNCS(amcntenclr1, AMCNTENCLR1)
 
+DEFINE_COPROCR_RW_FUNCS_64(amevcntr00, AMEVCNTR00)
+DEFINE_COPROCR_RW_FUNCS_64(amevcntr01, AMEVCNTR01)
+DEFINE_COPROCR_RW_FUNCS_64(amevcntr02, AMEVCNTR02)
+DEFINE_COPROCR_RW_FUNCS_64(amevcntr03, AMEVCNTR03)
+
 /*
  * TLBI operation prototypes
  */
diff --git a/include/lib/aarch64/arch.h b/include/lib/aarch64/arch.h
index 7f3e9faa..91aa484f 100644
--- a/include/lib/aarch64/arch.h
+++ b/include/lib/aarch64/arch.h
@@ -664,4 +664,45 @@
 #define AMEVTYPER02_EL0		S3_3_C13_C6_2
 #define AMEVTYPER03_EL0		S3_3_C13_C6_3
 
+/* Activity Monitor Group 1 Event Counter Registers */
+#define AMEVCNTR10_EL0		S3_3_C13_C12_0
+#define AMEVCNTR11_EL0		S3_3_C13_C12_1
+#define AMEVCNTR12_EL0		S3_3_C13_C12_2
+#define AMEVCNTR13_EL0		S3_3_C13_C12_3
+#define AMEVCNTR14_EL0		S3_3_C13_C12_4
+#define AMEVCNTR15_EL0		S3_3_C13_C12_5
+#define AMEVCNTR16_EL0		S3_3_C13_C12_6
+#define AMEVCNTR17_EL0		S3_3_C13_C12_7
+#define AMEVCNTR18_EL0		S3_3_C13_C13_0
+#define AMEVCNTR19_EL0		S3_3_C13_C13_1
+#define AMEVCNTR1A_EL0		S3_3_C13_C13_2
+#define AMEVCNTR1B_EL0		S3_3_C13_C13_3
+#define AMEVCNTR1C_EL0		S3_3_C13_C13_4
+#define AMEVCNTR1D_EL0		S3_3_C13_C13_5
+#define AMEVCNTR1E_EL0		S3_3_C13_C13_6
+#define AMEVCNTR1F_EL0		S3_3_C13_C13_7
+
+/* Activity Monitor Group 1 Event Type Registers */
+#define AMEVTYPER10_EL0		S3_3_C13_C14_0
+#define AMEVTYPER11_EL0		S3_3_C13_C14_1
+#define AMEVTYPER12_EL0		S3_3_C13_C14_2
+#define AMEVTYPER13_EL0		S3_3_C13_C14_3
+#define AMEVTYPER14_EL0		S3_3_C13_C14_4
+#define AMEVTYPER15_EL0		S3_3_C13_C14_5
+#define AMEVTYPER16_EL0		S3_3_C13_C14_6
+#define AMEVTYPER17_EL0		S3_3_C13_C14_7
+#define AMEVTYPER18_EL0		S3_3_C13_C15_0
+#define AMEVTYPER19_EL0		S3_3_C13_C15_1
+#define AMEVTYPER1A_EL0		S3_3_C13_C15_2
+#define AMEVTYPER1B_EL0		S3_3_C13_C15_3
+#define AMEVTYPER1C_EL0		S3_3_C13_C15_4
+#define AMEVTYPER1D_EL0		S3_3_C13_C15_5
+#define AMEVTYPER1E_EL0		S3_3_C13_C15_6
+#define AMEVTYPER1F_EL0		S3_3_C13_C15_7
+
+/* AMCGCR_EL0 definitions */
+#define AMCGCR_EL0_CG1NC_SHIFT	U(8)
+#define AMCGCR_EL0_CG1NC_LENGTH	U(8)
+#define AMCGCR_EL0_CG1NC_MASK	U(0xff)
+
 #endif /* __ARCH_H__ */
diff --git a/include/lib/aarch64/arch_helpers.h b/include/lib/aarch64/arch_helpers.h
index 831dfb06..485ed432 100644
--- a/include/lib/aarch64/arch_helpers.h
+++ b/include/lib/aarch64/arch_helpers.h
@@ -322,6 +322,7 @@ DEFINE_RENAME_SYSREG_WRITE_FUNC(icc_eoir0_el1, ICC_EOIR0_EL1)
 DEFINE_RENAME_SYSREG_WRITE_FUNC(icc_eoir1_el1, ICC_EOIR1_EL1)
 DEFINE_RENAME_SYSREG_WRITE_FUNC(icc_sgi0r_el1, ICC_SGI0R_EL1)
 
+DEFINE_RENAME_SYSREG_RW_FUNCS(amcgcr_el0, AMCGCR_EL0)
 DEFINE_RENAME_SYSREG_RW_FUNCS(amcntenclr0_el0, AMCNTENCLR0_EL0)
 DEFINE_RENAME_SYSREG_RW_FUNCS(amcntenset0_el0, AMCNTENSET0_EL0)
 DEFINE_RENAME_SYSREG_RW_FUNCS(amcntenclr1_el0, AMCNTENCLR1_EL0)
diff --git a/include/lib/cpus/aarch64/cortex_a75.h b/include/lib/cpus/aarch64/cortex_a75.h
index d68c9572..940125da 100644
--- a/include/lib/cpus/aarch64/cortex_a75.h
+++ b/include/lib/cpus/aarch64/cortex_a75.h
@@ -50,7 +50,19 @@
  * CPUAMEVTYPER<n> register and are disabled by default. Platforms may
  * enable this with suitable programming.
  */
+#define CORTEX_A75_AMU_NR_COUNTERS	5
 #define CORTEX_A75_AMU_GROUP0_MASK	0x7
 #define CORTEX_A75_AMU_GROUP1_MASK	(0 << 3)
 
+#ifndef __ASSEMBLY__
+#include <stdint.h>
+
+uint64_t cortex_a75_amu_cnt_read(int idx);
+void cortex_a75_amu_cnt_write(int idx, uint64_t val);
+unsigned int cortex_a75_amu_read_cpuamcntenset_el0(void);
+unsigned int cortex_a75_amu_read_cpuamcntenclr_el0(void);
+void cortex_a75_amu_write_cpuamcntenset_el0(unsigned int mask);
+void cortex_a75_amu_write_cpuamcntenclr_el0(unsigned int mask);
+#endif /* __ASSEMBLY__ */
+
 #endif /* __CORTEX_A75_H__ */
diff --git a/include/lib/el3_runtime/pubsub_events.h b/include/lib/el3_runtime/pubsub_events.h
index 9cfedb4d..64b3f630 100644
--- a/include/lib/el3_runtime/pubsub_events.h
+++ b/include/lib/el3_runtime/pubsub_events.h
@@ -17,6 +17,13 @@
  */
 REGISTER_PUBSUB_EVENT(psci_cpu_on_finish);
 
+/*
+ * These events are published before/after a CPU has been powered down/up
+ * via the PSCI CPU SUSPEND API.
+ */
+REGISTER_PUBSUB_EVENT(psci_suspend_pwrdown_start);
+REGISTER_PUBSUB_EVENT(psci_suspend_pwrdown_finish);
+
 #ifdef AARCH64
 /*
  * These events are published by the AArch64 context management framework
diff --git a/include/lib/extensions/amu.h b/include/lib/extensions/amu.h
index bbefe8ff..faa0ee12 100644
--- a/include/lib/extensions/amu.h
+++ b/include/lib/extensions/amu.h
@@ -7,9 +7,39 @@
 #ifndef __AMU_H__
 #define __AMU_H__
 
-/* Enable all group 0 counters */
+#include <sys/cdefs.h> /* for CASSERT() */
+#include <cassert.h>
+#include <platform_def.h>
+#include <stdint.h>
+
+/* All group 0 counters */
 #define AMU_GROUP0_COUNTERS_MASK	0xf
 
+#ifdef PLAT_AMU_GROUP1_COUNTERS_MASK
+#define AMU_GROUP1_COUNTERS_MASK	PLAT_AMU_GROUP1_COUNTERS_MASK
+#else
+#define AMU_GROUP1_COUNTERS_MASK	0
+#endif
+
+#ifdef PLAT_AMU_GROUP1_NR_COUNTERS
+#define AMU_GROUP1_NR_COUNTERS		PLAT_AMU_GROUP1_NR_COUNTERS
+#else
+#define AMU_GROUP1_NR_COUNTERS		0
+#endif
+
+CASSERT(AMU_GROUP1_COUNTERS_MASK <= 0xffff, invalid_amu_group1_counters_mask);
+CASSERT(AMU_GROUP1_NR_COUNTERS <= 16, invalid_amu_group1_nr_counters);
+
+int amu_supported(void);
 void amu_enable(int el2_unused);
 
+/* Group 0 configuration helpers */
+uint64_t amu_group0_cnt_read(int idx);
+void amu_group0_cnt_write(int idx, uint64_t val);
+
+/* Group 1 configuration helpers */
+uint64_t amu_group1_cnt_read(int idx);
+void amu_group1_cnt_write(int idx, uint64_t val);
+void amu_group1_set_evtype(int idx, unsigned int val);
+
 #endif /* __AMU_H__ */
diff --git a/include/lib/extensions/amu_private.h b/include/lib/extensions/amu_private.h
new file mode 100644
index 00000000..0c660bb8
--- /dev/null
+++ b/include/lib/extensions/amu_private.h
@@ -0,0 +1,19 @@
+/*
+ * Copyright (c) 2017, ARM Limited and Contributors. All rights reserved.
+ *
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+
+#ifndef __AMU_PRIVATE_H__
+#define __AMU_PRIVATE_H__
+
+#include <stdint.h>
+
+uint64_t amu_group0_cnt_read_internal(int idx);
+void amu_group0_cnt_write_internal(int idx, uint64_t);
+
+uint64_t amu_group1_cnt_read_internal(int idx);
+void amu_group1_cnt_write_internal(int idx, uint64_t);
+void amu_group1_set_evtype_internal(int idx, unsigned int val);
+
+#endif /* __AMU_PRIVATE_H__ */
diff --git a/lib/cpus/aarch64/cortex_a75.S b/lib/cpus/aarch64/cortex_a75.S
index 9b54b48f..e66ad066 100644
--- a/lib/cpus/aarch64/cortex_a75.S
+++ b/lib/cpus/aarch64/cortex_a75.S
@@ -11,6 +11,104 @@
 #include <plat_macros.S>
 #include <cortex_a75.h>
 
+	.globl	cortex_a75_amu_cnt_read
+	.globl	cortex_a75_amu_cnt_write
+	.globl	cortex_a75_amu_read_cpuamcntenset_el0
+	.globl	cortex_a75_amu_read_cpuamcntenclr_el0
+	.globl	cortex_a75_amu_write_cpuamcntenset_el0
+	.globl	cortex_a75_amu_write_cpuamcntenclr_el0
+
+/*
+ * uint64_t cortex_a75_amu_cnt_read(int idx);
+ *
+ * Given `idx`, read the corresponding AMU counter
+ * and return it in `x0`.
+ */
+func cortex_a75_amu_cnt_read
+	adr	x1, 1f
+	lsl	x0, x0, #3
+	add	x1, x1, x0
+	br	x1
+
+1:
+	mrs	x0, CPUAMEVCNTR0_EL0
+	ret
+	mrs	x0, CPUAMEVCNTR1_EL0
+	ret
+	mrs	x0, CPUAMEVCNTR2_EL0
+	ret
+	mrs	x0, CPUAMEVCNTR3_EL0
+	ret
+	mrs	x0, CPUAMEVCNTR4_EL0
+	ret
+endfunc cortex_a75_amu_cnt_read
+
+/*
+ * void cortex_a75_amu_cnt_write(int idx, uint64_t val);
+ *
+ * Given `idx`, write `val` to the corresponding AMU counter.
+ */
+func cortex_a75_amu_cnt_write
+	adr	x2, 1f
+	lsl	x0, x0, #3
+	add	x2, x2, x0
+	br	x2
+
+1:
+	msr	CPUAMEVCNTR0_EL0, x0
+	ret
+	msr	CPUAMEVCNTR1_EL0, x0
+	ret
+	msr	CPUAMEVCNTR2_EL0, x0
+	ret
+	msr	CPUAMEVCNTR3_EL0, x0
+	ret
+	msr	CPUAMEVCNTR4_EL0, x0
+	ret
+endfunc cortex_a75_amu_cnt_write
+
+/*
+ * unsigned int cortex_a75_amu_read_cpuamcntenset_el0(void);
+ *
+ * Read the `CPUAMCNTENSET_EL0` CPU register and return
+ * it in `x0`.
+ */
+func cortex_a75_amu_read_cpuamcntenset_el0
+	mrs	x0, CPUAMCNTENSET_EL0
+	ret
+endfunc cortex_a75_amu_read_cpuamcntenset_el0
+
+/*
+ * unsigned int cortex_a75_amu_read_cpuamcntenclr_el0(void);
+ *
+ * Read the `CPUAMCNTENCLR_EL0` CPU register and return
+ * it in `x0`.
+ */
+func cortex_a75_amu_read_cpuamcntenclr_el0
+	mrs	x0, CPUAMCNTENCLR_EL0
+	ret
+endfunc cortex_a75_amu_read_cpuamcntenclr_el0
+
+/*
+ * void cortex_a75_amu_write_cpuamcntenset_el0(unsigned int mask);
+ *
+ * Write `mask` to the `CPUAMCNTENSET_EL0` CPU register.
+ */
+func cortex_a75_amu_write_cpuamcntenset_el0
+	msr	CPUAMCNTENSET_EL0, x0
+	ret
+endfunc cortex_a75_amu_write_cpuamcntenset_el0
+
+/*
+ * void cortex_a75_amu_write_cpuamcntenclr_el0(unsigned int mask);
+ *
+ * Write `mask` to the `CPUAMCNTENCLR_EL0` CPU register.
+ */
+func cortex_a75_amu_write_cpuamcntenclr_el0
+	mrs	x0, CPUAMCNTENCLR_EL0
+	ret
+endfunc cortex_a75_amu_write_cpuamcntenclr_el0
+
 func cortex_a75_reset_func
 #if IMAGE_BL31 && WORKAROUND_CVE_2017_5715
 	mrs	x0, id_aa64pfr0_el1
diff --git a/lib/cpus/aarch64/cortex_a75_pubsub.c b/lib/cpus/aarch64/cortex_a75_pubsub.c
new file mode 100644
index 00000000..c1089a60
--- /dev/null
+++ b/lib/cpus/aarch64/cortex_a75_pubsub.c
@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) 2017, ARM Limited and Contributors. All rights reserved.
+ *
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+
+#include <cortex_a75.h>
+#include <pubsub_events.h>
+#include <platform.h>
+
+struct amu_ctx {
+	uint64_t cnts[CORTEX_A75_AMU_NR_COUNTERS];
+	uint16_t mask;
+};
+
+static struct amu_ctx amu_ctxs[PLATFORM_CORE_COUNT];
+
+static void *cortex_a75_context_save(const void *arg)
+{
+	struct amu_ctx *ctx = &amu_ctxs[plat_my_core_pos()];
+	unsigned int midr;
+	unsigned int midr_mask;
+	int i;
+
+	midr = read_midr();
+	midr_mask = (MIDR_IMPL_MASK << MIDR_IMPL_SHIFT) |
+		(MIDR_PN_MASK << MIDR_PN_SHIFT);
+	if ((midr & midr_mask) != (CORTEX_A75_MIDR & midr_mask))
+		return 0;
+
+	/* Save counter configuration */
+	ctx->mask = cortex_a75_amu_read_cpuamcntenset_el0();
+
+	/* Ensure counters are disabled */
+	cortex_a75_amu_write_cpuamcntenclr_el0(ctx->mask);
+	isb();
+
+	/* Save counters */
+	for (i = 0; i < CORTEX_A75_AMU_NR_COUNTERS; i++)
+		ctx->cnts[i] = cortex_a75_amu_cnt_read(i);
+
+	return 0;
+}
+
+static void *cortex_a75_context_restore(const void *arg)
+{
+	struct amu_ctx *ctx = &amu_ctxs[plat_my_core_pos()];
+	unsigned int midr;
+	unsigned int midr_mask;
+	int i;
+
+	midr = read_midr();
+	midr_mask = (MIDR_IMPL_MASK << MIDR_IMPL_SHIFT) |
+		(MIDR_PN_MASK << MIDR_PN_SHIFT);
+	if ((midr & midr_mask) != (CORTEX_A75_MIDR & midr_mask))
+		return 0;
+
+	ctx = &amu_ctxs[plat_my_core_pos()];
+
+	/* Counters were disabled in `cortex_a75_context_save()` */
+	assert(cortex_a75_amu_read_cpuamcntenset_el0() == 0);
+
+	/* Restore counters */
+	for (i = 0; i < CORTEX_A75_AMU_NR_COUNTERS; i++)
+		cortex_a75_amu_cnt_write(i, ctx->cnts[i]);
+	isb();
+
+	/* Restore counter configuration */
+	cortex_a75_amu_write_cpuamcntenset_el0(ctx->mask);
+
+	return 0;
+}
+
+SUBSCRIBE_TO_EVENT(psci_suspend_pwrdown_start, cortex_a75_context_save);
+SUBSCRIBE_TO_EVENT(psci_suspend_pwrdown_finish, cortex_a75_context_restore);
diff --git a/lib/extensions/amu/aarch32/amu.c b/lib/extensions/amu/aarch32/amu.c
index d450bd69..55462cbf 100644
--- a/lib/extensions/amu/aarch32/amu.c
+++ b/lib/extensions/amu/aarch32/amu.c
@@ -7,26 +7,103 @@
 #include <amu.h>
 #include <arch.h>
 #include <arch_helpers.h>
+#include <debug.h>
+#include <platform.h>
+#include <pubsub_events.h>
+
+#define AMU_GROUP0_NR_COUNTERS	4
+
+struct amu_ctx {
+	uint64_t group0_cnts[AMU_GROUP0_NR_COUNTERS];
+};
+
+static struct amu_ctx amu_ctxs[PLATFORM_CORE_COUNT];
 
 void amu_enable(int el2_unused)
 {
 	uint64_t features;
 
 	features = read_id_pfr0() >> ID_PFR0_AMU_SHIFT;
-	if ((features & ID_PFR0_AMU_MASK) == 1) {
-		if (el2_unused) {
-			uint64_t v;
-
-			/*
-			 * Non-secure access from EL0 or EL1 to the Activity Monitor
-			 * registers do not trap to EL2.
-			 */
-			v = read_hcptr();
-			v &= ~TAM_BIT;
-			write_hcptr(v);
-		}
-
-		/* Enable group 0 counters */
-		write_amcntenset0(AMU_GROUP0_COUNTERS_MASK);
+	if ((features & ID_PFR0_AMU_MASK) != 1) {
+		WARN("Cannot enable AMU - not supported\n");
+		return;
+	}
+
+	if (el2_unused) {
+		uint64_t v;
+
+		/*
+		 * Non-secure access from EL0 or EL1 to the Activity Monitor
+		 * registers do not trap to EL2.
+		 */
+		v = read_hcptr();
+		v &= ~TAM_BIT;
+		write_hcptr(v);
 	}
+
+	/* Enable group 0 counters */
+	write_amcntenset0(AMU_GROUP0_COUNTERS_MASK);
+}
+
+static void *amu_context_save(const void *arg)
+{
+	struct amu_ctx *ctx;
+	uint64_t features;
+
+	features = read_id_pfr0() >> ID_PFR0_AMU_SHIFT;
+	if ((features & ID_PFR0_AMU_MASK) != 1)
+		return (void *)-1;
+
+	ctx = &amu_ctxs[plat_my_core_pos()];
+
+	/* Assert that group 0 counter configuration is what we expect */
+	assert(read_amcntenset0() == AMU_GROUP0_COUNTERS_MASK);
+
+	/*
+	 * Disable group 0 counters to avoid other observers like SCP sampling
+	 * counter values from the future via the memory mapped view.
+	 */
+	write_amcntenclr0(AMU_GROUP0_COUNTERS_MASK);
+	isb();
+
+	ctx->group0_cnts[0] = read64_amevcntr00();
+	ctx->group0_cnts[1] = read64_amevcntr01();
+	ctx->group0_cnts[2] = read64_amevcntr02();
+	ctx->group0_cnts[3] = read64_amevcntr03();
+
+	return 0;
 }
+
+static void *amu_context_restore(const void *arg)
+{
+	struct amu_ctx *ctx;
+	uint64_t features;
+
+	features = read_id_pfr0() >> ID_PFR0_AMU_SHIFT;
+	if ((features & ID_PFR0_AMU_MASK) != 1)
+		return (void *)-1;
+
+	ctx = &amu_ctxs[plat_my_core_pos()];
+
+	/* Counters were disabled in `amu_context_save()` */
+	assert(read_amcntenset0() == 0);
+
+	/* Restore group 0 counters */
+	if (AMU_GROUP0_COUNTERS_MASK & (1U << 0))
+		write64_amevcntr00(ctx->group0_cnts[0]);
+	if (AMU_GROUP0_COUNTERS_MASK & (1U << 1))
+		write64_amevcntr01(ctx->group0_cnts[1]);
+	if (AMU_GROUP0_COUNTERS_MASK & (1U << 2))
+		write64_amevcntr02(ctx->group0_cnts[2]);
+	if (AMU_GROUP0_COUNTERS_MASK & (1U << 3))
+		write64_amevcntr03(ctx->group0_cnts[3]);
+	isb();
+
+	/* Enable group 0 counters */
+	write_amcntenset0(AMU_GROUP0_COUNTERS_MASK);
+
+	return 0;
+}
+
+SUBSCRIBE_TO_EVENT(psci_suspend_pwrdown_start, amu_context_save);
+SUBSCRIBE_TO_EVENT(psci_suspend_pwrdown_finish, amu_context_restore);
diff --git a/lib/extensions/amu/aarch64/amu.c b/lib/extensions/amu/aarch64/amu.c
index 007b3494..f743e209 100644
--- a/lib/extensions/amu/aarch64/amu.c
+++ b/lib/extensions/amu/aarch64/amu.c
@@ -5,36 +5,184 @@
  */
 
 #include <amu.h>
+#include <amu_private.h>
 #include <arch.h>
 #include <arch_helpers.h>
+#include <assert.h>
+#include <debug.h>
+#include <platform.h>
+#include <pubsub_events.h>
 
-void amu_enable(int el2_unused)
+#define AMU_GROUP0_NR_COUNTERS	4
+
+struct amu_ctx {
+	uint64_t group0_cnts[AMU_GROUP0_NR_COUNTERS];
+	uint64_t group1_cnts[AMU_GROUP1_NR_COUNTERS];
+};
+
+static struct amu_ctx amu_ctxs[PLATFORM_CORE_COUNT];
+
+int amu_supported(void)
 {
 	uint64_t features;
 
 	features = read_id_aa64pfr0_el1() >> ID_AA64PFR0_AMU_SHIFT;
-	if ((features & ID_AA64PFR0_AMU_MASK) == 1) {
-		uint64_t v;
-
-		if (el2_unused) {
-			/*
-			 * CPTR_EL2.TAM: Set to zero so any accesses to
-			 * the Activity Monitor registers do not trap to EL2.
-			 */
-			v = read_cptr_el2();
-			v &= ~CPTR_EL2_TAM_BIT;
-			write_cptr_el2(v);
-		}
+	return (features & ID_AA64PFR0_AMU_MASK) == 1;
+}
 
+/*
+ * Enable counters.  This function is meant to be invoked
+ * by the context management library before exiting from EL3.
+ */
+void amu_enable(int el2_unused)
+{
+	uint64_t v;
+
+	if (!amu_supported()) {
+		WARN("Cannot enable AMU - not supported\n");
+		return;
+	}
+
+	if (el2_unused) {
 		/*
-		 * CPTR_EL3.TAM: Set to zero so that any accesses to
-		 * the Activity Monitor registers do not trap to EL3.
+		 * CPTR_EL2.TAM: Set to zero so any accesses to
+		 * the Activity Monitor registers do not trap to EL2.
 		 */
-		v = read_cptr_el3();
-		v &= ~TAM_BIT;
-		write_cptr_el3(v);
-
-		/* Enable group 0 counters */
-		write_amcntenset0_el0(AMU_GROUP0_COUNTERS_MASK);
+		v = read_cptr_el2();
+		v &= ~CPTR_EL2_TAM_BIT;
+		write_cptr_el2(v);
 	}
+
+	/*
+	 * CPTR_EL3.TAM: Set to zero so that any accesses to
+	 * the Activity Monitor registers do not trap to EL3.
+	 */
+	v = read_cptr_el3();
+	v &= ~TAM_BIT;
+	write_cptr_el3(v);
+
+	/* Enable group 0 counters */
+	write_amcntenset0_el0(AMU_GROUP0_COUNTERS_MASK);
+	/* Enable group 1 counters */
+	write_amcntenset1_el0(AMU_GROUP1_COUNTERS_MASK);
+}
+
+/* Read the group 0 counter identified by the given `idx`. */
+uint64_t amu_group0_cnt_read(int idx)
+{
+	assert(amu_supported());
+	assert(idx >= 0 && idx < AMU_GROUP0_NR_COUNTERS);
+
+	return amu_group0_cnt_read_internal(idx);
+}
+
+/* Write the group 0 counter identified by the given `idx` with `val`. */
+void amu_group0_cnt_write(int idx, uint64_t val)
+{
+	assert(amu_supported());
+	assert(idx >= 0 && idx < AMU_GROUP0_NR_COUNTERS);
+
+	amu_group0_cnt_write_internal(idx, val);
+	isb();
+}
+
+/* Read the group 1 counter identified by the given `idx`. */
+uint64_t amu_group1_cnt_read(int idx)
+{
+	assert(amu_supported());
+	assert(idx >= 0 && idx < AMU_GROUP1_NR_COUNTERS);
+
+	return amu_group1_cnt_read_internal(idx);
+}
+
+/* Write the group 1 counter identified by the given `idx` with `val`. */
+void amu_group1_cnt_write(int idx, uint64_t val)
+{
+	assert(amu_supported());
+	assert(idx >= 0 && idx < AMU_GROUP1_NR_COUNTERS);
+
+	amu_group1_cnt_write_internal(idx, val);
+	isb();
+}
+
+/*
+ * Program the event type register for the given `idx` with
+ * the event number `val`.
+ */
+void amu_group1_set_evtype(int idx, unsigned int val)
+{
+	assert(amu_supported());
+	assert (idx >= 0 && idx < AMU_GROUP1_NR_COUNTERS);
+
+	amu_group1_set_evtype_internal(idx, val);
+	isb();
 }
+
+static void *amu_context_save(const void *arg)
+{
+	struct amu_ctx *ctx = &amu_ctxs[plat_my_core_pos()];
+	int i;
+
+	if (!amu_supported())
+		return (void *)-1;
+
+	/* Assert that group 0/1 counter configuration is what we expect */
+	assert(read_amcntenset0_el0() == AMU_GROUP0_COUNTERS_MASK &&
+	       read_amcntenset1_el0() == AMU_GROUP1_COUNTERS_MASK);
+
+	assert((sizeof(int) * 8) - __builtin_clz(AMU_GROUP1_COUNTERS_MASK)
+		<= AMU_GROUP1_NR_COUNTERS);
+
+	/*
+	 * Disable group 0/1 counters to avoid other observers like SCP sampling
+	 * counter values from the future via the memory mapped view.
+	 */
+	write_amcntenclr0_el0(AMU_GROUP0_COUNTERS_MASK);
+	write_amcntenclr1_el0(AMU_GROUP1_COUNTERS_MASK);
+	isb();
+
+	/* Save group 0 counters */
+	for (i = 0; i < AMU_GROUP0_NR_COUNTERS; i++)
+		ctx->group0_cnts[i] = amu_group0_cnt_read(i);
+
+	/* Save group 1 counters */
+	for (i = 0; i < AMU_GROUP1_NR_COUNTERS; i++)
+		ctx->group1_cnts[i] = amu_group1_cnt_read(i);
+
+	return 0;
+}
+
+static void *amu_context_restore(const void *arg)
+{
+	struct amu_ctx *ctx = &amu_ctxs[plat_my_core_pos()];
+	int i;
+
+	if (!amu_supported())
+		return (void *)-1;
+
+	/* Counters were disabled in `amu_context_save()` */
+	assert(read_amcntenset0_el0() == 0 && read_amcntenset1_el0() == 0);
+
+	assert((sizeof(int) * 8) - __builtin_clz(AMU_GROUP1_COUNTERS_MASK)
+		<= AMU_GROUP1_NR_COUNTERS);
+
+	/* Restore group 0 counters */
+	for (i = 0; i < AMU_GROUP0_NR_COUNTERS; i++)
+		if (AMU_GROUP0_COUNTERS_MASK & (1U << i))
+			amu_group0_cnt_write(i, ctx->group0_cnts[i]);
+
+	/* Restore group 1 counters */
+	for (i = 0; i < AMU_GROUP1_NR_COUNTERS; i++)
+		if (AMU_GROUP1_COUNTERS_MASK & (1U << i))
+			amu_group1_cnt_write(i, ctx->group1_cnts[i]);
+	isb();
+
+	/* Restore group 0/1 counter configuration */
+	write_amcntenset0_el0(AMU_GROUP0_COUNTERS_MASK);
+	write_amcntenset1_el0(AMU_GROUP1_COUNTERS_MASK);
+
+	return 0;
+}
+
+SUBSCRIBE_TO_EVENT(psci_suspend_pwrdown_start, amu_context_save);
+SUBSCRIBE_TO_EVENT(psci_suspend_pwrdown_finish, amu_context_restore);
diff --git a/lib/extensions/amu/aarch64/amu_helpers.S b/lib/extensions/amu/aarch64/amu_helpers.S
new file mode 100644
index 00000000..e0b1f564
--- /dev/null
+++ b/lib/extensions/amu/aarch64/amu_helpers.S
@@ -0,0 +1,281 @@
+/*
+ * Copyright (c) 2017, ARM Limited and Contributors. All rights reserved.
+ *
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+
+#include <arch.h>
+#include <assert_macros.S>
+#include <asm_macros.S>
+
+	.globl	amu_group0_cnt_read_internal
+	.globl	amu_group0_cnt_write_internal
+	.globl	amu_group1_cnt_read_internal
+	.globl	amu_group1_cnt_write_internal
+	.globl	amu_group1_set_evtype_internal
+
+/*
+ * uint64_t amu_group0_cnt_read_internal(int idx);
+ *
+ * Given `idx`, read the corresponding AMU counter
+ * and return it in `x0`.
+ */
+func amu_group0_cnt_read_internal
+#if ENABLE_ASSERTIONS
+	/*
+	 * It can be dangerous to call this function with an
+	 * out of bounds index.  Ensure `idx` is valid.
+	 */
+	mov	x1, x0
+	lsr	x1, x1, #2
+	cmp	x1, #0
+	ASM_ASSERT(eq)
+#endif
+
+	/*
+	 * Given `idx` calculate address of mrs/ret instruction pair
+	 * in the table below.
+	 */
+	adr	x1, 1f
+	lsl	x0, x0, #3		/* each mrs/ret sequence is 8 bytes */
+	add	x1, x1, x0
+	br	x1
+
+1:
+	mrs	x0, AMEVCNTR00_EL0	/* index 0 */
+	ret
+	mrs	x0, AMEVCNTR01_EL0	/* index 1 */
+	ret
+	mrs	x0, AMEVCNTR02_EL0	/* index 2 */
+	ret
+	mrs	x0, AMEVCNTR03_EL0	/* index 3 */
+	ret
+endfunc amu_group0_cnt_read_internal
+
+/*
+ * void amu_group0_cnt_write_internal(int idx, uint64_t val);
+ *
+ * Given `idx`, write `val` to the corresponding AMU counter.
+ */
+func amu_group0_cnt_write_internal
+#if ENABLE_ASSERTIONS
+	/*
+	 * It can be dangerous to call this function with an
+	 * out of bounds index.  Ensure `idx` is valid.
+	 */
+	mov	x2, x0
+	lsr	x2, x2, #2
+	cmp	x2, #0
+	ASM_ASSERT(eq)
+#endif
+
+	/*
+	 * Given `idx` calculate address of mrs/ret instruction pair
+	 * in the table below.
+	 */
+	adr	x2, 1f
+	lsl	x0, x0, #3		/* each msr/ret sequence is 8 bytes */
+	add	x2, x2, x0
+	br	x2
+
+1:
+	msr	AMEVCNTR00_EL0, x1	/* index 0 */
+	ret
+	msr	AMEVCNTR01_EL0, x1	/* index 1 */
+	ret
+	msr	AMEVCNTR02_EL0, x1	/* index 2 */
+	ret
+	msr	AMEVCNTR03_EL0, x1	/* index 3 */
+	ret
+endfunc amu_group0_cnt_write_internal
+
+/*
+ * uint64_t amu_group1_cnt_read_internal(int idx);
+ *
+ * Given `idx`, read the corresponding AMU counter
+ * and return it in `x0`.
+ */
+func amu_group1_cnt_read_internal
+#if ENABLE_ASSERTIONS
+	/*
+	 * It can be dangerous to call this function with an
+	 * out of bounds index.  Ensure `idx` is valid.
+	 */
+	mov	x1, x0
+	lsr	x1, x1, #4
+	cmp	x1, #0
+	ASM_ASSERT(eq)
+#endif
+
+	/*
+	 * Given `idx` calculate address of mrs/ret instruction pair
+	 * in the table below.
+	 */
+	adr	x1, 1f
+	lsl	x0, x0, #3		/* each mrs/ret sequence is 8 bytes */
+	add	x1, x1, x0
+	br	x1
+
+1:
+	mrs	x0, AMEVCNTR10_EL0	/* index 0 */
+	ret
+	mrs	x0, AMEVCNTR11_EL0	/* index 1 */
+	ret
+	mrs	x0, AMEVCNTR12_EL0	/* index 2 */
+	ret
+	mrs	x0, AMEVCNTR13_EL0	/* index 3 */
+	ret
+	mrs	x0, AMEVCNTR14_EL0	/* index 4 */
+	ret
+	mrs	x0, AMEVCNTR15_EL0	/* index 5 */
+	ret
+	mrs	x0, AMEVCNTR16_EL0	/* index 6 */
+	ret
+	mrs	x0, AMEVCNTR17_EL0	/* index 7 */
+	ret
+	mrs	x0, AMEVCNTR18_EL0	/* index 8 */
+	ret
+	mrs	x0, AMEVCNTR19_EL0	/* index 9 */
+	ret
+	mrs	x0, AMEVCNTR1A_EL0	/* index 10 */
+	ret
+	mrs	x0, AMEVCNTR1B_EL0	/* index 11 */
+	ret
+	mrs	x0, AMEVCNTR1C_EL0	/* index 12 */
+	ret
+	mrs	x0, AMEVCNTR1D_EL0	/* index 13 */
+	ret
+	mrs	x0, AMEVCNTR1E_EL0	/* index 14 */
+	ret
+	mrs	x0, AMEVCNTR1F_EL0	/* index 15 */
+	ret
+endfunc amu_group1_cnt_read_internal
+
+/*
+ * void amu_group1_cnt_write_internal(int idx, uint64_t val);
+ *
+ * Given `idx`, write `val` to the corresponding AMU counter.
+ */
+func amu_group1_cnt_write_internal
+#if ENABLE_ASSERTIONS
+	/*
+	 * It can be dangerous to call this function with an
+	 * out of bounds index.  Ensure `idx` is valid.
+	 */
+	mov	x2, x0
+	lsr	x2, x2, #4
+	cmp	x2, #0
+	ASM_ASSERT(eq)
+#endif
+
+	/*
+	 * Given `idx` calculate address of mrs/ret instruction pair
+	 * in the table below.
+	 */
+	adr	x2, 1f
+	lsl	x0, x0, #3		/* each msr/ret sequence is 8 bytes */
+	add	x2, x2, x0
+	br	x2
+
+1:
+	msr	AMEVCNTR10_EL0, x1	/* index 0 */
+	ret
+	msr	AMEVCNTR11_EL0, x1	/* index 1 */
+	ret
+	msr	AMEVCNTR12_EL0, x1	/* index 2 */
+	ret
+	msr	AMEVCNTR13_EL0, x1	/* index 3 */
+	ret
+	msr	AMEVCNTR14_EL0, x1	/* index 4 */
+	ret
+	msr	AMEVCNTR15_EL0, x1	/* index 5 */
+	ret
+	msr	AMEVCNTR16_EL0, x1	/* index 6 */
+	ret
+	msr	AMEVCNTR17_EL0, x1	/* index 7 */
+	ret
+	msr	AMEVCNTR18_EL0, x1	/* index 8 */
+	ret
+	msr	AMEVCNTR19_EL0, x1	/* index 9 */
+	ret
+	msr	AMEVCNTR1A_EL0, x1	/* index 10 */
+	ret
+	msr	AMEVCNTR1B_EL0, x1	/* index 11 */
+	ret
+	msr	AMEVCNTR1C_EL0, x1	/* index 12 */
+	ret
+	msr	AMEVCNTR1D_EL0, x1	/* index 13 */
+	ret
+	msr	AMEVCNTR1E_EL0, x1	/* index 14 */
+	ret
+	msr	AMEVCNTR1F_EL0, x1	/* index 15 */
+	ret
+endfunc amu_group1_cnt_write_internal
+
+/*
+ * void amu_group1_set_evtype_internal(int idx, unsigned int val);
+ *
+ * Program the AMU event type register indexed by `idx`
+ * with the value `val`.
+ */
+func amu_group1_set_evtype_internal
+#if ENABLE_ASSERTIONS
+	/*
+	 * It can be dangerous to call this function with an
+	 * out of bounds index.  Ensure `idx` is valid.
+	 */
+	mov	x2, x0
+	lsr	x2, x2, #4
+	cmp	x2, #0
+	ASM_ASSERT(eq)
+
+	/* val should be between [0, 65535] */
+	mov	x2, x1
+	lsr	x2, x2, #16
+	cmp	x2, #0
+	ASM_ASSERT(eq)
+#endif
+
+	/*
+	 * Given `idx` calculate address of msr/ret instruction pair
+	 * in the table below.
+	 */
+	adr	x2, 1f
+	lsl	x0, x0, #3		/* each msr/ret sequence is 8 bytes */
+	add	x2, x2, x0
+	br	x2
+
+1:
+	msr	AMEVTYPER10_EL0, x1	/* index 0 */
+	ret
+	msr	AMEVTYPER11_EL0, x1	/* index 1 */
+	ret
+	msr	AMEVTYPER12_EL0, x1	/* index 2 */
+	ret
+	msr	AMEVTYPER13_EL0, x1	/* index 3 */
+	ret
+	msr	AMEVTYPER14_EL0, x1	/* index 4 */
+	ret
+	msr	AMEVTYPER15_EL0, x1	/* index 5 */
+	ret
+	msr	AMEVTYPER16_EL0, x1	/* index 6 */
+	ret
+	msr	AMEVTYPER17_EL0, x1	/* index 7 */
+	ret
+	msr	AMEVTYPER18_EL0, x1	/* index 8 */
+	ret
+	msr	AMEVTYPER19_EL0, x1	/* index 9 */
+	ret
+	msr	AMEVTYPER1A_EL0, x1	/* index 10 */
+	ret
+	msr	AMEVTYPER1B_EL0, x1	/* index 11 */
+	ret
+	msr	AMEVTYPER1C_EL0, x1	/* index 12 */
+	ret
+	msr	AMEVTYPER1D_EL0, x1	/* index 13 */
+	ret
+	msr	AMEVTYPER1E_EL0, x1	/* index 14 */
+	ret
+	msr	AMEVTYPER1F_EL0, x1	/* index 15 */
+	ret
+endfunc amu_group1_set_evtype_internal
diff --git a/lib/psci/psci_suspend.c b/lib/psci/psci_suspend.c
index d9490672..a77972d3 100644
--- a/lib/psci/psci_suspend.c
+++ b/lib/psci/psci_suspend.c
@@ -14,6 +14,7 @@
 #include <debug.h>
 #include <platform.h>
 #include <pmf.h>
+#include <pubsub_events.h>
 #include <runtime_instr.h>
 #include <stddef.h>
 #include "psci_private.h"
@@ -68,6 +69,8 @@ static void psci_suspend_to_pwrdown_start(unsigned int end_pwrlvl,
 {
 	unsigned int max_off_lvl = psci_find_max_off_lvl(state_info);
 
+	PUBLISH_EVENT(psci_suspend_pwrdown_start);
+
 	/* Save PSCI target power level for the suspend finisher handler */
 	psci_set_suspend_pwrlvl(end_pwrlvl);
 
@@ -308,6 +311,8 @@ void psci_cpu_suspend_finish(unsigned int cpu_idx,
 	/* Invalidate the suspend level for the cpu */
 	psci_set_suspend_pwrlvl(PSCI_INVALID_PWR_LVL);
 
+	PUBLISH_EVENT(psci_suspend_pwrdown_finish);
+
 	/*
 	 * Generic management: Now we just need to retrieve the
 	 * information that we had stashed away during the suspend
diff --git a/plat/arm/board/fvp/platform.mk b/plat/arm/board/fvp/platform.mk
index 7edbd3df..9d3c5f6b 100644
--- a/plat/arm/board/fvp/platform.mk
+++ b/plat/arm/board/fvp/platform.mk
@@ -150,6 +150,10 @@ ENABLE_PLAT_COMPAT	:= 	0
 # Enable Activity Monitor Unit extensions by default
 ENABLE_AMU			:=	1
 
+ifeq (${ENABLE_AMU},1)
+BL31_SOURCES		+= lib/cpus/aarch64/cortex_a75_pubsub.c
+endif
+
 ifneq (${ENABLE_STACK_PROTECTOR},0)
 PLAT_BL_COMMON_SOURCES	+=	plat/arm/board/fvp/fvp_stack_protector.c
 endif