aarch64: Enable Statistical Profiling Extensions for lower ELs

SPE is only supported in non-secure state.  Accesses to SPE specific
registers from SEL1 will trap to EL3.  During a world switch, before
`TTBR` is modified the SPE profiling buffers are drained.  This is to
avoid a potential invalid memory access in SEL1.

SPE is architecturally specified only for AArch64.

Change-Id: I04a96427d9f9d586c331913d815fdc726855f6b0
Signed-off-by: dp-arm <dimitris.papastamos@arm.com>

Makefile

@@ -454,6 +454,7 @@ $(eval $(call assert_boolean,TRUSTED_BOARD_BOOT))
 $(eval $(call assert_boolean,USE_COHERENT_MEM))
 $(eval $(call assert_boolean,USE_TBBR_DEFS))
 $(eval $(call assert_boolean,WARMBOOT_ENABLE_DCACHE_EARLY))
+$(eval $(call assert_boolean,ENABLE_SPE_FOR_LOWER_ELS))
 
 $(eval $(call assert_numeric,ARM_ARCH_MAJOR))
 $(eval $(call assert_numeric,ARM_ARCH_MINOR))
@@ -493,6 +494,7 @@ $(eval $(call add_define,TRUSTED_BOARD_BOOT))
 $(eval $(call add_define,USE_COHERENT_MEM))
 $(eval $(call add_define,USE_TBBR_DEFS))
 $(eval $(call add_define,WARMBOOT_ENABLE_DCACHE_EARLY))
+$(eval $(call add_define,ENABLE_SPE_FOR_LOWER_ELS))
 
 # Define the EL3_PAYLOAD_BASE flag only if it is provided.
 ifdef EL3_PAYLOAD_BASE

docs/user-guide.md

@@ -542,6 +542,11 @@ performed.
     cluster platforms). If this option is enabled, then warm boot path
     enables D-caches immediately after enabling MMU. This option defaults to 0.
 
+*   `ENABLE_SPE_FOR_LOWER_ELS` : Boolean option to enable Statistical Profiling
+     extensions.  This is an optional architectural feature available only for
+     AArch64 8.2 onwards.  This option defaults to 1 but is automatically
+     disabled when the target architecture is AArch32 or AArch64 8.0/8.1.
+
 #### ARM development platform specific build options
 
 *   `ARM_BL31_IN_DRAM`: Boolean option to select loading of BL31 in TZC secured

include/common/aarch64/el3_common_macros.S

@@ -95,6 +95,10 @@
 	 * MDCR_EL3.SPD32: Set to 0b10 to disable AArch32 Secure self-hosted
 	 *  privileged debug from S-EL1.
 	 *
+	 * MDCR_EL3.NSPB (ARM v8.2): SPE enabled in non-secure state and
+	 * disabled in secure state. Accesses to SPE registers at SEL1 generate
+	 * trap exceptions to EL3.
+	 *
 	 * MDCR_EL3.TDOSA: Set to zero so that EL2 and EL2 System register
 	 *  access to the powerdown debug registers do not trap to EL3.
 	 *
@@ -108,6 +112,19 @@
 	 */
 	mov_imm	x0, ((MDCR_EL3_RESET_VAL | MDCR_SDD_BIT | MDCR_SPD32(MDCR_SPD32_DISABLE)) \
 			& ~(MDCR_TDOSA_BIT | MDCR_TDA_BIT | MDCR_TPM_BIT))
+
+#if ENABLE_SPE_FOR_LOWER_ELS
+	/* Detect if SPE is implemented */
+	mrs	x1, id_aa64dfr0_el1
+	ubfx	x1, x1, #ID_AA64DFR0_PMS_SHIFT, #ID_AA64DFR0_PMS_LENGTH
+	cmp	x1, #0x1
+	b.ne	1f
+
+	/* Enable SPE for use by normal world */
+	orr	x0, x0, #MDCR_NSPB(MDCR_NSPB_EL1)
+1:
+#endif
+
 	msr	mdcr_el3, x0
 
 	/* ---------------------------------------------------------------------

include/lib/aarch64/arch.h

@@ -110,6 +110,11 @@
 #define ID_AA64PFR0_EL3_SHIFT	U(12)
 #define ID_AA64PFR0_ELX_MASK	U(0xf)
 
+/* 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	U(0xf)
+
 #define EL_IMPL_NONE		U(0)
 #define EL_IMPL_A64ONLY		U(1)
 #define EL_IMPL_A64_A32		U(2)
@@ -189,6 +194,8 @@
 #define MDCR_SPD32_DISABLE	U(0x2)
 #define MDCR_SPD32_ENABLE	U(0x3)
 #define MDCR_SDD_BIT		(U(1) << 16)
+#define MDCR_NSPB(x)		((x) << 12)
+#define MDCR_NSPB_EL1		U(0x3)
 #define MDCR_TDOSA_BIT		(U(1) << 10)
 #define MDCR_TDA_BIT		(U(1) << 9)
 #define MDCR_TPM_BIT		(U(1) << 6)
@@ -199,6 +206,9 @@
 #endif
 
 /* MDCR_EL2 definitions */
+#define MDCR_EL2_TPMS		(U(1) << 14)
+#define MDCR_EL2_E2PB(x)	((x) << 12)
+#define MDCR_EL2_E2PB_EL1	U(0x3)
 #define MDCR_EL2_TDRA_BIT	(U(1) << 11)
 #define MDCR_EL2_TDOSA_BIT	(U(1) << 10)
 #define MDCR_EL2_TDA_BIT	(U(1) << 9)

include/lib/aarch64/arch_helpers.h

@@ -184,6 +184,7 @@ DEFINE_SYSREG_WRITE_CONST_FUNC(daifclr)
 DEFINE_SYSREG_READ_FUNC(par_el1)
 DEFINE_SYSREG_READ_FUNC(id_pfr1_el1)
 DEFINE_SYSREG_READ_FUNC(id_aa64pfr0_el1)
+DEFINE_SYSREG_READ_FUNC(id_aa64dfr0_el1)
 DEFINE_SYSREG_READ_FUNC(CurrentEl)
 DEFINE_SYSREG_RW_FUNCS(daif)
 DEFINE_SYSREG_RW_FUNCS(spsr_el1)

include/lib/el3_runtime/aarch64/context.h

@@ -308,6 +308,7 @@ CASSERT(CTX_EL3STATE_OFFSET == __builtin_offsetof(cpu_context_t, el3state_ctx),
  * Function prototypes
  ******************************************************************************/
 void el1_sysregs_context_save(el1_sys_regs_t *regs);
+void el1_sysregs_context_save_post_ops(void);
 void el1_sysregs_context_restore(el1_sys_regs_t *regs);
 #if CTX_INCLUDE_FPREGS
 void fpregs_context_save(fp_regs_t *regs);

include/plat/arm/common/plat_arm.h

@@ -218,4 +218,7 @@ int arm_execution_state_switch(unsigned int smc_fid,
 		uint32_t cookie_lo,
 		void *handle);
 
+/* Disable Statistical Profiling Extensions helper */
+void arm_disable_spe(void);
+
 #endif /* __PLAT_ARM_H__ */

lib/el3_runtime/aarch64/context.S

 /*
- * Copyright (c) 2013-2016, ARM Limited and Contributors. All rights reserved.
+ * Copyright (c) 2013-2017, ARM Limited and Contributors. All rights reserved.
  *
  * SPDX-License-Identifier: BSD-3-Clause
  */
@@ -9,6 +9,7 @@
 #include <context.h>
 
 	.global	el1_sysregs_context_save
+	.global el1_sysregs_context_save_post_ops
 	.global	el1_sysregs_context_restore
 #if CTX_INCLUDE_FPREGS
 	.global	fpregs_context_save
@@ -108,6 +109,36 @@ func el1_sysregs_context_save
 	ret
 endfunc el1_sysregs_context_save
 
+/* -----------------------------------------------------
+ * The following function strictly follows the AArch64
+ * PCS to use x9-x17 (temporary caller-saved registers)
+ * to do post operations after saving the EL1 system
+ * register context.
+ * -----------------------------------------------------
+ */
+func el1_sysregs_context_save_post_ops
+#if ENABLE_SPE_FOR_LOWER_ELS
+	/* Detect if SPE is implemented */
+	mrs	x9, id_aa64dfr0_el1
+	ubfx	x9, x9, #ID_AA64DFR0_PMS_SHIFT, #ID_AA64DFR0_PMS_LENGTH
+	cmp	x9, #0x1
+	b.ne	1f
+
+	/*
+	 * Before switching from normal world to secure world
+	 * the profiling buffers need to be drained out to memory.  This is
+	 * required to avoid an invalid memory access when TTBR is switched
+	 * for entry to SEL1.
+	 */
+	.arch	armv8.2-a+profile
+	psb	csync
+	dsb	nsh
+	.arch	armv8-a
+1:
+#endif
+	ret
+endfunc el1_sysregs_context_save_post_ops
+
 /* -----------------------------------------------------
  * The following function strictly follows the AArch64
  * PCS to use x9-x17 (temporary caller-saved registers)
@@ -343,7 +374,7 @@ func restore_gp_registers_callee_eret
 	ldp	x24, x25, [sp, #CTX_GPREGS_OFFSET + CTX_GPREG_X24]
 	ldp	x26, x27, [sp, #CTX_GPREGS_OFFSET + CTX_GPREG_X26]
 	ldp	x28, x29, [sp, #CTX_GPREGS_OFFSET + CTX_GPREG_X28]
-	ldp	 x30, x17, [sp, #CTX_GPREGS_OFFSET + CTX_GPREG_LR]
+	ldp	x30, x17, [sp, #CTX_GPREGS_OFFSET + CTX_GPREG_LR]
 	msr	sp_el0, x17
 	ldp	x16, x17, [sp, #CTX_GPREGS_OFFSET + CTX_GPREG_X16]
 	eret

lib/el3_runtime/aarch64/context_mgmt.c

@@ -218,7 +218,7 @@ void cm_init_my_context(const entry_point_info_t *ep)
  ******************************************************************************/
 void cm_prepare_el3_exit(uint32_t security_state)
 {
-	uint32_t sctlr_elx, scr_el3;
+	uint32_t sctlr_elx, scr_el3, mdcr_el2;
 	cpu_context_t *ctx = cm_get_context(security_state);
 
 	assert(ctx);
@@ -315,6 +315,13 @@ void cm_prepare_el3_exit(uint32_t security_state)
 			 * relying on hw. Some fields are architecturally
 			 * UNKNOWN on reset.
 			 *
+			 * MDCR_EL2.TPMS (ARM v8.2): Do not trap statistical
+			 * profiling controls to EL2.
+			 *
+			 * MDCR_EL2.E2PB (ARM v8.2): SPE enabled in non-secure
+			 * state. Accesses to profiling buffer controls at
+			 * non-secure EL1 are not trapped to EL2.
+			 *
 			 * MDCR_EL2.TDRA: Set to zero so that Non-secure EL0 and
 			 *  EL1 System register accesses to the Debug ROM
 			 *  registers are not trapped to EL2.
@@ -343,13 +350,32 @@ void cm_prepare_el3_exit(uint32_t security_state)
 			 * MDCR_EL2.HPMN: Set to value of PMCR_EL0.N which is the
 			 *  architecturally-defined reset value.
 			 */
-			write_mdcr_el2((MDCR_EL2_RESET_VAL |
+			mdcr_el2 = ((MDCR_EL2_RESET_VAL |
 					((read_pmcr_el0() & PMCR_EL0_N_BITS)
 					>> PMCR_EL0_N_SHIFT)) &
 					~(MDCR_EL2_TDRA_BIT | MDCR_EL2_TDOSA_BIT
 					| MDCR_EL2_TDA_BIT | MDCR_EL2_TDE_BIT
 					| MDCR_EL2_HPME_BIT | MDCR_EL2_TPM_BIT
 					| MDCR_EL2_TPMCR_BIT));
+
+#if ENABLE_SPE_FOR_LOWER_ELS
+			uint64_t id_aa64dfr0_el1;
+
+			/* Detect if SPE is implemented */
+			id_aa64dfr0_el1 = read_id_aa64dfr0_el1() >>
+				ID_AA64DFR0_PMS_SHIFT;
+			if ((id_aa64dfr0_el1 & ID_AA64DFR0_PMS_MASK) == 1) {
+				/*
+				 * Make sure traps to EL2 are not generated if
+				 * EL2 is implemented but not used.
+				 */
+				mdcr_el2 &= ~MDCR_EL2_TPMS;
+				mdcr_el2 |= MDCR_EL2_E2PB(MDCR_EL2_E2PB_EL1);
+			}
+#endif
+
+			write_mdcr_el2(mdcr_el2);
+
 			/*
 			 * Initialise HSTR_EL2. All fields are architecturally
 			 * UNKNOWN on reset.
@@ -389,6 +415,7 @@ void cm_el1_sysregs_context_save(uint32_t security_state)
 	assert(ctx);
 
 	el1_sysregs_context_save(get_sysregs_ctx(ctx));
+	el1_sysregs_context_save_post_ops();
 }
 
 void cm_el1_sysregs_context_restore(uint32_t security_state)

make_helpers/defaults.mk

@@ -136,3 +136,20 @@ V				:= 0
 # required to enable cache coherency after warm reset (eg: single cluster
 # platforms).
 WARMBOOT_ENABLE_DCACHE_EARLY	:= 0
+
+# By default, enable Statistical Profiling Extensions.
+# The top level Makefile will disable this feature depending on
+# the target architecture and version number.
+ENABLE_SPE_FOR_LOWER_ELS	:= 1
+
+# SPE is enabled by default but only supported on AArch64 8.2 onwards.
+# Disable it in all other cases.
+ifeq (${ARCH},aarch32)
+    override ENABLE_SPE_FOR_LOWER_ELS := 0
+else
+    ifeq (${ARM_ARCH_MAJOR},8)
+        ifeq ($(ARM_ARCH_MINOR),$(filter $(ARM_ARCH_MINOR),0 1))
+            ENABLE_SPE_FOR_LOWER_ELS := 0
+        endif
+    endif
+endif

plat/arm/board/fvp/fvp_pm.c

@@ -48,6 +48,14 @@ static void fvp_cluster_pwrdwn_common(void)
 {
 	uint64_t mpidr = read_mpidr_el1();
 
+#if ENABLE_SPE_FOR_LOWER_ELS
+	/*
+	 * On power down we need to disable statistical profiling extensions
+	 * before exiting coherency.
+	 */
+	arm_disable_spe();
+#endif
+
 	/* Disable coherency if this cluster is to be turned off */
 	fvp_interconnect_disable();
 

plat/arm/common/aarch64/arm_helpers.S

@@ -12,6 +12,7 @@
 	.globl	plat_crash_console_putc
 	.globl	plat_crash_console_flush
 	.globl	platform_mem_init
+	.globl	arm_disable_spe
 
 
 	/* -----------------------------------------------------
@@ -86,3 +87,31 @@ endfunc plat_crash_console_flush
 func platform_mem_init
 	ret
 endfunc platform_mem_init
+
+	/* -----------------------------------------------------
+	 * void arm_disable_spe (void);
+	 * -----------------------------------------------------
+	 */
+#if ENABLE_SPE_FOR_LOWER_ELS
+func arm_disable_spe
+	/* Detect if SPE is implemented */
+	mrs	x0, id_aa64dfr0_el1
+	ubfx	x0, x0, #ID_AA64DFR0_PMS_SHIFT, #ID_AA64DFR0_PMS_LENGTH
+	cmp	x0, #0x1
+	b.ne	1f
+
+	/* Drain buffered data */
+	.arch	armv8.2-a+profile
+	psb	csync
+	dsb	nsh
+
+	/* Disable Profiling Buffer */
+	mrs	x0, pmblimitr_el1
+	bic	x0, x0, #1
+	msr	pmblimitr_el1, x0
+	isb
+	.arch	armv8-a
+1:
+	ret
+endfunc arm_disable_spe
+#endif