ARM: HYP/non-sec: allow relocation to secure RAM

The current non-sec switching code suffers from one major issue:
it cannot run in secure RAM, as a large part of u-boot still needs
to be run while we're switched to non-secure.

This patch reworks the whole HYP/non-secure strategy by:
- making sure the secure code is the *last* thing u-boot executes
  before entering the payload
- performing an exception return from secure mode directly into
  the payload
- allowing the code to be dynamically relocated to secure RAM
  before switching to non-secure.

This involves quite a bit of horrible code, specially as u-boot
relocation is quite primitive.
Signed-off-by: NMarc Zyngier <marc.zyngier@arm.com>
Acked-by: NIan Campbell <ijc@hellion.org.uk>

arch/arm/cpu/armv7/nonsec_virt.S

@@ -10,10 +10,13 @@
 #include <linux/linkage.h>
 #include <asm/gic.h>
 #include <asm/armv7.h>
+#include <asm/proc-armv/ptrace.h>
 
 .arch_extension sec
 .arch_extension virt
 
+	.pushsection ._secure.text, "ax"
+
 	.align	5
 /* the vector table for secure state and HYP mode */
 _monitor_vectors:
@@ -22,51 +25,86 @@ _monitor_vectors:
 	adr pc, _secure_monitor
 	.word 0
 	.word 0
-	adr pc, _hyp_trap
+	.word 0
 	.word 0
 	.word 0
 
+.macro is_cpu_virt_capable	tmp
+	mrc	p15, 0, \tmp, c0, c1, 1		@ read ID_PFR1
+	and	\tmp, \tmp, #CPUID_ARM_VIRT_MASK	@ mask virtualization bits
+	cmp	\tmp, #(1 << CPUID_ARM_VIRT_SHIFT)
+.endm
+
 /*
  * secure monitor handler
  * U-boot calls this "software interrupt" in start.S
  * This is executed on a "smc" instruction, we use a "smc #0" to switch
  * to non-secure state.
- * We use only r0 and r1 here, due to constraints in the caller.
+ * r0, r1, r2: passed to the callee
+ * ip: target PC
  */
 _secure_monitor:
-	mrc	p15, 0, r1, c1, c1, 0		@ read SCR
-	bic	r1, r1, #0x4e			@ clear IRQ, FIQ, EA, nET bits
-	orr	r1, r1, #0x31			@ enable NS, AW, FW bits
+	mrc	p15, 0, r5, c1, c1, 0		@ read SCR
+	bic	r5, r5, #0x4e			@ clear IRQ, FIQ, EA, nET bits
+	orr	r5, r5, #0x31			@ enable NS, AW, FW bits
 
-	mrc	p15, 0, r0, c0, c1, 1		@ read ID_PFR1
-	and	r0, r0, #CPUID_ARM_VIRT_MASK	@ mask virtualization bits
-	cmp	r0, #(1 << CPUID_ARM_VIRT_SHIFT)
+	mov	r6, #SVC_MODE			@ default mode is SVC
+	is_cpu_virt_capable r4
 #ifdef CONFIG_ARMV7_VIRT
-	orreq	r1, r1, #0x100			@ allow HVC instruction
+	orreq	r5, r5, #0x100			@ allow HVC instruction
+	moveq	r6, #HYP_MODE			@ Enter the kernel as HYP
 #endif
 
-	mcr	p15, 0, r1, c1, c1, 0		@ write SCR (with NS bit set)
+	mcr	p15, 0, r5, c1, c1, 0		@ write SCR (with NS bit set)
 	isb
 
-#ifdef CONFIG_ARMV7_VIRT
-	mrceq	p15, 0, r0, c12, c0, 1		@ get MVBAR value
-	mcreq	p15, 4, r0, c12, c0, 0		@ write HVBAR
-#endif
 	bne	1f
 
 	@ Reset CNTVOFF to 0 before leaving monitor mode
-	mrc	p15, 0, r0, c0, c1, 1		@ read ID_PFR1
-	ands	r0, r0, #CPUID_ARM_GENTIMER_MASK	@ test arch timer bits
-	movne	r0, #0
-	mcrrne	p15, 4, r0, r0, c14		@ Reset CNTVOFF to zero
+	mrc	p15, 0, r4, c0, c1, 1		@ read ID_PFR1
+	ands	r4, r4, #CPUID_ARM_GENTIMER_MASK	@ test arch timer bits
+	movne	r4, #0
+	mcrrne	p15, 4, r4, r4, c14		@ Reset CNTVOFF to zero
 1:
-	movs	pc, lr				@ return to non-secure SVC
-
-_hyp_trap:
-	mrs	lr, elr_hyp	@ for older asm: .byte 0x00, 0xe3, 0x0e, 0xe1
-	mov pc, lr				@ do no switch modes, but
-						@ return to caller
-
+	mov	lr, ip
+	mov	ip, #(F_BIT | I_BIT | A_BIT)	@ Set A, I and F
+	tst	lr, #1				@ Check for Thumb PC
+	orrne	ip, ip, #T_BIT			@ Set T if Thumb
+	orr	ip, ip, r6			@ Slot target mode in
+	msr	spsr_cxfs, ip			@ Set full SPSR
+	movs	pc, lr				@ ERET to non-secure
+
+ENTRY(_do_nonsec_entry)
+	mov	ip, r0
+	mov	r0, r1
+	mov	r1, r2
+	mov	r2, r3
+	smc	#0
+ENDPROC(_do_nonsec_entry)
+
+.macro get_cbar_addr	addr
+#ifdef CONFIG_ARM_GIC_BASE_ADDRESS
+	ldr	\addr, =CONFIG_ARM_GIC_BASE_ADDRESS
+#else
+	mrc	p15, 4, \addr, c15, c0, 0	@ read CBAR
+	bfc	\addr, #0, #15			@ clear reserved bits
+#endif
+.endm
+
+.macro get_gicd_addr	addr
+	get_cbar_addr	\addr
+	add	\addr, \addr, #GIC_DIST_OFFSET	@ GIC dist i/f offset
+.endm
+
+.macro get_gicc_addr	addr, tmp
+	get_cbar_addr	\addr
+	is_cpu_virt_capable \tmp
+	movne	\tmp, #GIC_CPU_OFFSET_A9	@ GIC CPU offset for A9
+	moveq	\tmp, #GIC_CPU_OFFSET_A15	@ GIC CPU offset for A15/A7
+	add	\addr, \addr, \tmp
+.endm
+
+#ifndef CONFIG_ARMV7_PSCI
 /*
  * Secondary CPUs start here and call the code for the core specific parts
  * of the non-secure and HYP mode transition. The GIC distributor specific
@@ -74,31 +112,21 @@ _hyp_trap:
  * Then they go back to wfi and wait to be woken up by the kernel again.
  */
 ENTRY(_smp_pen)
-	mrs	r0, cpsr
-	orr	r0, r0, #0xc0
-	msr	cpsr, r0			@ disable interrupts
-	ldr	r1, =_start
-	mcr	p15, 0, r1, c12, c0, 0		@ set VBAR
+	cpsid	i
+	cpsid	f
 
 	bl	_nonsec_init
-	mov	r12, r0				@ save GICC address
-#ifdef CONFIG_ARMV7_VIRT
-	bl	_switch_to_hyp
-#endif
-
-	ldr	r1, [r12, #GICC_IAR]		@ acknowledge IPI
-	str	r1, [r12, #GICC_EOIR]		@ signal end of interrupt
 
 	adr	r0, _smp_pen			@ do not use this address again
 	b	smp_waitloop			@ wait for IPIs, board specific
 ENDPROC(_smp_pen)
+#endif
 
 /*
  * Switch a core to non-secure state.
  *
  *  1. initialize the GIC per-core interface
  *  2. allow coprocessor access in non-secure modes
- *  3. switch the cpu mode (by calling "smc #0")
  *
  * Called from smp_pen by secondary cores and directly by the BSP.
  * Do not assume that the stack is available and only use registers
@@ -108,38 +136,23 @@ ENDPROC(_smp_pen)
  * though, but we check this in C before calling this function.
  */
 ENTRY(_nonsec_init)
-#ifdef CONFIG_ARM_GIC_BASE_ADDRESS
-	ldr	r2, =CONFIG_ARM_GIC_BASE_ADDRESS
-#else
-	mrc	p15, 4, r2, c15, c0, 0		@ read CBAR
-	bfc	r2, #0, #15			@ clear reserved bits
-#endif
-	add	r3, r2, #GIC_DIST_OFFSET	@ GIC dist i/f offset
+	get_gicd_addr	r3
+
 	mvn	r1, #0				@ all bits to 1
 	str	r1, [r3, #GICD_IGROUPRn]	@ allow private interrupts
 
-	mrc	p15, 0, r0, c0, c0, 0		@ read MIDR
-	ldr	r1, =MIDR_PRIMARY_PART_MASK
-	and	r0, r0, r1			@ mask out variant and revision
+	get_gicc_addr	r3, r1
 
-	ldr	r1, =MIDR_CORTEX_A7_R0P0 & MIDR_PRIMARY_PART_MASK
-	cmp	r0, r1				@ check for Cortex-A7
-
-	ldr	r1, =MIDR_CORTEX_A15_R0P0 & MIDR_PRIMARY_PART_MASK
-	cmpne	r0, r1				@ check for Cortex-A15
-
-	movne	r1, #GIC_CPU_OFFSET_A9		@ GIC CPU offset for A9
-	moveq	r1, #GIC_CPU_OFFSET_A15		@ GIC CPU offset for A15/A7
-	add	r3, r2, r1			@ r3 = GIC CPU i/f addr
-
-	mov	r1, #1				@ set GICC_CTLR[enable]
+	mov	r1, #3				@ Enable both groups
 	str	r1, [r3, #GICC_CTLR]		@ and clear all other bits
 	mov	r1, #0xff
 	str	r1, [r3, #GICC_PMR]		@ set priority mask register
 
+	mrc	p15, 0, r0, c1, c1, 2
 	movw	r1, #0x3fff
-	movt	r1, #0x0006
-	mcr	p15, 0, r1, c1, c1, 2		@ NSACR = all copros to non-sec
+	movt	r1, #0x0004
+	orr	r0, r0, r1
+	mcr	p15, 0, r0, c1, c1, 2		@ NSACR = all copros to non-sec
 
 /* The CNTFRQ register of the generic timer needs to be
  * programmed in secure state. Some primary bootloaders / firmware
@@ -157,21 +170,9 @@ ENTRY(_nonsec_init)
 
 	adr	r1, _monitor_vectors
 	mcr	p15, 0, r1, c12, c0, 1		@ set MVBAR to secure vectors
-
-	mrc	p15, 0, ip, c12, c0, 0		@ save secure copy of VBAR
-
 	isb
-	smc	#0				@ call into MONITOR mode
-
-	mcr	p15, 0, ip, c12, c0, 0		@ write non-secure copy of VBAR
-
-	mov	r1, #1
-	str	r1, [r3, #GICC_CTLR]		@ enable non-secure CPU i/f
-	add	r2, r2, #GIC_DIST_OFFSET
-	str	r1, [r2, #GICD_CTLR]		@ allow private interrupts
 
 	mov	r0, r3				@ return GICC address
-
 	bx	lr
 ENDPROC(_nonsec_init)
 
@@ -183,18 +184,10 @@ ENTRY(smp_waitloop)
 	ldr	r1, [r1]
 	cmp	r0, r1			@ make sure we dont execute this code
 	beq	smp_waitloop		@ again (due to a spurious wakeup)
-	mov	pc, r1
+	mov	r0, r1
+	b	_do_nonsec_entry
 ENDPROC(smp_waitloop)
 .weak smp_waitloop
 #endif
 
-ENTRY(_switch_to_hyp)
-	mov	r0, lr
-	mov	r1, sp				@ save SVC copy of LR and SP
-	isb
-	hvc #0			 @ for older asm: .byte 0x70, 0x00, 0x40, 0xe1
-	mov	sp, r1
-	mov	lr, r0				@ restore SVC copy of LR and SP
-
-	bx	lr
-ENDPROC(_switch_to_hyp)
+	.popsection

arch/arm/cpu/armv7/virt-v7.c

@@ -13,17 +13,10 @@
 #include <asm/armv7.h>
 #include <asm/gic.h>
 #include <asm/io.h>
+#include <asm/secure.h>
 
 unsigned long gic_dist_addr;
 
-static unsigned int read_cpsr(void)
-{
-	unsigned int reg;
-
-	asm volatile ("mrs %0, cpsr\n" : "=r" (reg));
-	return reg;
-}
-
 static unsigned int read_id_pfr1(void)
 {
 	unsigned int reg;
@@ -72,6 +65,18 @@ static unsigned long get_gicd_base_address(void)
 #endif
 }
 
+static void relocate_secure_section(void)
+{
+#ifdef CONFIG_ARMV7_SECURE_BASE
+	size_t sz = __secure_end - __secure_start;
+
+	memcpy((void *)CONFIG_ARMV7_SECURE_BASE, __secure_start, sz);
+	flush_dcache_range(CONFIG_ARMV7_SECURE_BASE,
+			   CONFIG_ARMV7_SECURE_BASE + sz + 1);
+	invalidate_icache_all();
+#endif
+}
+
 static void kick_secondary_cpus_gic(unsigned long gicdaddr)
 {
 	/* kick all CPUs (except this one) by writing to GICD_SGIR */
@@ -83,35 +88,7 @@ void __weak smp_kick_all_cpus(void)
 	kick_secondary_cpus_gic(gic_dist_addr);
 }
 
-int armv7_switch_hyp(void)
-{
-	unsigned int reg;
-
-	/* check whether we are in HYP mode already */
-	if ((read_cpsr() & 0x1f) == 0x1a) {
-		debug("CPU already in HYP mode\n");
-		return 0;
-	}
-
-	/* check whether the CPU supports the virtualization extensions */
-	reg = read_id_pfr1();
-	if ((reg & CPUID_ARM_VIRT_MASK) != 1 << CPUID_ARM_VIRT_SHIFT) {
-		printf("HYP mode: Virtualization extensions not implemented.\n");
-		return -1;
-	}
-
-	/* call the HYP switching code on this CPU also */
-	_switch_to_hyp();
-
-	if ((read_cpsr() & 0x1F) != 0x1a) {
-		printf("HYP mode: switch not successful.\n");
-		return -1;
-	}
-
-	return 0;
-}
-
-int armv7_switch_nonsec(void)
+int armv7_init_nonsec(void)
 {
 	unsigned int reg;
 	unsigned itlinesnr, i;
@@ -147,11 +124,13 @@ int armv7_switch_nonsec(void)
 	for (i = 1; i <= itlinesnr; i++)
 		writel((unsigned)-1, gic_dist_addr + GICD_IGROUPRn + 4 * i);
 
-	smp_set_core_boot_addr((unsigned long)_smp_pen, -1);
+#ifndef CONFIG_ARMV7_PSCI
+	smp_set_core_boot_addr((unsigned long)secure_ram_addr(_smp_pen), -1);
 	smp_kick_all_cpus();
+#endif
 
 	/* call the non-sec switching code on this CPU also */
-	_nonsec_init();
-
+	relocate_secure_section();
+	secure_ram_addr(_nonsec_init)();
 	return 0;
 }

arch/arm/include/asm/armv7.h

@@ -78,13 +78,17 @@ void v7_outer_cache_inval_range(u32 start, u32 end);
 
 #if defined(CONFIG_ARMV7_NONSEC) || defined(CONFIG_ARMV7_VIRT)
 
-int armv7_switch_nonsec(void);
-int armv7_switch_hyp(void);
+int armv7_init_nonsec(void);
 
 /* defined in assembly file */
 unsigned int _nonsec_init(void);
+void _do_nonsec_entry(void *target_pc, unsigned long r0,
+		      unsigned long r1, unsigned long r2);
 void _smp_pen(void);
-void _switch_to_hyp(void);
+
+extern char __secure_start[];
+extern char __secure_end[];
+
 #endif /* CONFIG_ARMV7_NONSEC || CONFIG_ARMV7_VIRT */
 
 #endif /* ! __ASSEMBLY__ */

arch/arm/include/asm/secure.h

+#ifndef __ASM_SECURE_H
+#define __ASM_SECURE_H
+
+#include <config.h>
+
+#ifdef CONFIG_ARMV7_SECURE_BASE
+/*
+ * Warning, horror ahead.
+ *
+ * The target code lives in our "secure ram", but u-boot doesn't know
+ * that, and has blindly added reloc_off to every relocation
+ * entry. Gahh. Do the opposite conversion. This hack also prevents
+ * GCC from generating code veeners, which u-boot doesn't relocate at
+ * all...
+ */
+#define secure_ram_addr(_fn) ({						\
+			DECLARE_GLOBAL_DATA_PTR;			\
+			void *__fn = _fn;				\
+			typeof(_fn) *__tmp = (__fn - gd->reloc_off);	\
+			__tmp;						\
+		})
+#else
+#define secure_ram_addr(_fn)	(_fn)
+#endif
+
+#endif

arch/arm/lib/bootm.c

@@ -20,6 +20,7 @@
 #include <libfdt.h>
 #include <fdt_support.h>
 #include <asm/bootm.h>
+#include <asm/secure.h>
 #include <linux/compiler.h>
 
 #if defined(CONFIG_ARMV7_NONSEC) || defined(CONFIG_ARMV7_VIRT)
@@ -184,27 +185,17 @@ static void setup_end_tag(bd_t *bd)
 
 __weak void setup_board_tags(struct tag **in_params) {}
 
+#ifdef CONFIG_ARM64
 static void do_nonsec_virt_switch(void)
 {
-#if defined(CONFIG_ARMV7_NONSEC) || defined(CONFIG_ARMV7_VIRT)
-	if (armv7_switch_nonsec() == 0)
-#ifdef CONFIG_ARMV7_VIRT
-		if (armv7_switch_hyp() == 0)
-			debug("entered HYP mode\n");
-#else
-		debug("entered non-secure state\n");
-#endif
-#endif
-
-#ifdef CONFIG_ARM64
 	smp_kick_all_cpus();
 	flush_dcache_all();	/* flush cache before swtiching to EL2 */
 	armv8_switch_to_el2();
 #ifdef CONFIG_ARMV8_SWITCH_TO_EL1
 	armv8_switch_to_el1();
 #endif
-#endif
 }
+#endif
 
 /* Subcommand: PREP */
 static void boot_prep_linux(bootm_headers_t *images)
@@ -289,8 +280,13 @@ static void boot_jump_linux(bootm_headers_t *images, int flag)
 		r2 = gd->bd->bi_boot_params;
 
 	if (!fake) {
-		do_nonsec_virt_switch();
+#if defined(CONFIG_ARMV7_NONSEC) || defined(CONFIG_ARMV7_VIRT)
+		armv7_init_nonsec();
+		secure_ram_addr(_do_nonsec_entry)(kernel_entry,
+						  0, machid, r2);
+#else
 		kernel_entry(0, machid, r2);
+#endif
 	}
 #endif
 }