
On Fri, Apr 26, 2013 at 6:14 AM, Andre Przywara andre.przywara@linaro.org wrote:
Currently the non-secure switch is only done for the boot processor. To later allow full SMP support, we have to switch all secondary cores into non-secure state also.
So we add an entry point for secondary CPUs coming out of low-power state and make sure we put them into WFI again after having switched to non-secure state. For this we acknowledge and EOI the wake-up IPI, then go into WFI. Once being kicked out of it later, we sanity check that the start address has actually been changed (since another attempt to switch to non-secure would block the core) and jump to the new address.
The actual CPU kick is done by sending an inter-processor interrupt via the GIC to all CPU interfaces except the requesting processor. The secondary cores will then setup their respective GIC CPU interface.
The address secondary cores jump to is board specific, we provide the value here for the Versatile Express board.
Signed-off-by: Andre Przywara andre.przywara@linaro.org
arch/arm/cpu/armv7/start.S | 27 ++++++++++++++++++++++++++- arch/arm/lib/virt-v7.c | 10 +++++++++- include/configs/vexpress_ca15_tc2.h | 1 + 3 files changed, 36 insertions(+), 2 deletions(-)
diff --git a/arch/arm/cpu/armv7/start.S b/arch/arm/cpu/armv7/start.S index 401b0eb..2b47881 100644 --- a/arch/arm/cpu/armv7/start.S +++ b/arch/arm/cpu/armv7/start.S @@ -563,8 +563,19 @@ fiq: #endif /* CONFIG_SPL_BUILD */
/* Routine to initialize GIC CPU interface and switch to nonsecure state.
- Will be executed directly by secondary CPUs after coming out of
- WFI, or can be called directly by C code for CPU 0.
- Those two paths mandate to not use any stack and to only use registers
- r0-r3 to comply with both the C ABI and the requirement of SMP startup
*/
- code.
.globl _nonsec_gic_switch +.globl _smp_pen +_smp_pen:
mrs r0, cpsr
orr r0, r0, #0xc0
msr cpsr, r0 @ disable interrupts
mov lr, #0 @ clear LR to mark secondary
_nonsec_gic_switch: mrc p15, 4, r2, c15, c0, 0 @ r2 = PERIPHBASE add r3, r2, #0x1000 @ GIC dist i/f offset @@ -605,4 +616,18 @@ _nonsec_gic_switch: add r2, r2, #0x1000 @ GIC dist i/f offset str r1, [r2] @ allow private interrupts
mov pc, lr
cmp lr, #0
movne pc, lr @ CPU 0 to return
@ all others: go to sleep
+_ack_int:
ldr r1, [r3, #0x0c] @ read GICD acknowledge
str r1, [r3, #0x10] @ write GICD EOI
adr r1, _smp_pen
+waitloop:
wfi
ldr r0, =CONFIG_SYSFLAGS_ADDR @ load start address
ldr r0, [r0]
cmp r0, r1 @ make sure we dont execute this code
beq waitloop @ again (due to a spurious wakeup)
mov pc, r0
diff --git a/arch/arm/lib/virt-v7.c b/arch/arm/lib/virt-v7.c index 416ca29..5ca093a 100644 --- a/arch/arm/lib/virt-v7.c +++ b/arch/arm/lib/virt-v7.c @@ -29,6 +29,7 @@
/* the assembly routine doing the actual work in start.S */ void _nonsec_gic_switch(void); +void _smp_pen(void);
#define GICD_CTLR 0x000 #define GICD_TYPER 0x004 @@ -51,6 +52,7 @@ int armv7_switch_nonsec(void) unsigned int reg; volatile unsigned int *gicdptr; unsigned itlinesnr, i;
unsigned int *sysflags; /* check whether the CPU supports the security extensions */ asm("mrc p15, 0, %0, c0, c1, 1\n" : "=r"(reg));
@@ -109,7 +111,13 @@ int armv7_switch_nonsec(void) for (i = 0; i <= itlinesnr; i++) gicdptr[GICD_IGROUPR0 / 4 + i] = (unsigned)-1;
/* call the non-sec switching code on this CPU */
/* now kick all CPUs (expect this one) by writing to GICD_SIGR */
sysflags = (void *)CONFIG_SYSFLAGS_ADDR;
sysflags[1] = (unsigned)-1;
sysflags[0] = (uintptr_t)_smp_pen;
gicdptr[GICD_SGIR / 4] = 1U << 24;
/* call the non-sec switching code on this CPU also */ _nonsec_gic_switch(); return 0;
diff --git a/include/configs/vexpress_ca15_tc2.h b/include/configs/vexpress_ca15_tc2.h index 9e230ad..210a27c 100644 --- a/include/configs/vexpress_ca15_tc2.h +++ b/include/configs/vexpress_ca15_tc2.h @@ -32,5 +32,6 @@ #define CONFIG_BOOTP_VCI_STRING "U-boot.armv7.vexpress_ca15x2_tc2"
#define CONFIG_SYS_CLK_FREQ 24000000 +#define CONFIG_SYSFLAGS_ADDR 0x1c010030
#endif
hmm, is all this likely to work across all armv7 cores? I imagined that the SMP boot protocol could be potentially vastly different on different implementations and this stuff would have to call into board specific code...?