[boot-wrapper PATCH 5/5] aarch64: Introduce EL2 boot code for Armv8-R AArch64

[Andre Przywara]

On Tue, 20 Apr 2021 15:24:38 +0800
Jaxson Han <jaxson.han at arm.com> wrote:

Hi,

> The Armv8-R AArch64 profile does not support the EL3 exception level.
> The Armv8-R AArch64 profile allows for an (optional) VMSAv8-64 MMU
> at EL1, which allows to run off-the-shelf Linux. However EL2 only
> supports a PMSA, which is not supported by Linux, so we need to drop
> into EL1 before entering the kernel.
> 
> The boot sequence is:
> If CurrentEL == EL3, then goto EL3 initialisation and drop to lower EL
>   before entering the kernel.
> If CurrentEL == EL2 && id_aa64mmfr0_el1.MSA == 0xf (Armv8-R aarch64),
>   then goto Armv8-R AArch64 initialisation and drop to EL1 before
>   entering the kernel.
> Else, no initialisation and keep the current EL before entering the
>   kernel.
> 
> Signed-off-by: Jaxson Han <jaxson.han at arm.com>
> ---
>  arch/aarch64/boot.S | 51 +++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 51 insertions(+)
> 
> diff --git a/arch/aarch64/boot.S b/arch/aarch64/boot.S
> index f7dbf3f..6961a2a 100644
> --- a/arch/aarch64/boot.S
> +++ b/arch/aarch64/boot.S
> @@ -25,16 +25,22 @@ _start:
>  	 * Boot sequence
>  	 * If CurrentEL == EL3, then goto EL3 initialisation and drop to
>  	 *   lower EL before entering the kernel.
> +	 * If CurrentEL == EL2 && id_aa64mmfr0_el1.MSA == 0xf, then goto
> +	 *   Armv8-R AArch64 initialisation and drop to EL1 before
> +	 *   entering the kernel.
>  	 * Else, no initialisation and keep the current EL before
>  	 *   entering the kernel.
>  	 */
>  	mrs	x0, CurrentEL
>  	cmp	x0, #CURRENTEL_EL3
>  	beq	el3_init
> +	cmp	x0, #CURRENTEL_EL2
> +	beq	el2_init

nitpick: I tend to compare against EL2, then use b.gt for EL3, b.lt for
EL1 and b.eq for EL2 code. Saves you an extra cmp here.

>  	/*
>  	 * We stay in the current EL for entering the kernel
>  	 */
> +keep_el:
>  	mov	w0, #1
>  	ldr	x1, =flag_keep_el
>  	str	w0, [x1]
> @@ -112,6 +118,43 @@ el3_init:
>  	str	w0, [x1]
>  	b	el_max_init
>  
> +	/*
> +	 * EL2 Armv8-R AArch64 initialisation
> +	 */
> +el2_init:
> +	/* Detect Armv8-R AArch64 */
> +	mrs	x1, id_aa64mmfr0_el1
> +	ubfx	x1, x1, #48, #4			// MSA
> +	/* 0xf means Armv8-R AArch64 */
> +	cmp	x1, 0xf
> +	bne	keep_el

Don't we need to also check bits[55:52], to have at least 0b0010?
IIUC the support for VMSA in EL1&0 is optional, and should be checked
before we proceed? VTCR_EL2[31] can only be set in the 0b0010 case.

> +
> +	mrs	x0, midr_el1
> +	msr	vpidr_el2, x0
> +
> +	mrs	x0, mpidr_el1
> +	msr	vmpidr_el2, x0
> +
> +	mov	x0, #(1 << 31)			// VTCR_MSA: VMSAv8-64 support
> +	msr	vtcr_el2, x0
> +
> +	/* Enable pointer authentication if present */
> +	mrs	x1, id_aa64isar1_el1
> +	ldr	x2, =(((0xff) << 24) | (0xff << 4))

Each feature only holds four bits, so the mask you shift should be 0xf.

> +	and	x1, x1, x2
> +	cbz	x1, 1f
> +
> +	mrs	x0, hcr_el2

Shouldn't we force HCR_EL2, instead of modifying it? Just to make sure
nothing unexpected traps into EL2, which we don't handle very well?
So basically just set bit 31 (RES1), plus those two bits on top, if
needed. But I also wonder about FIEN[47] and EnSCXT[53] ...


Rest looks alright.

Cheers,
Andre

> +	orr	x0, x0, #(1 << 40)		// AP key enable
> +	orr	x0, x0, #(1 << 41)		// AP insn enable
> +	msr	hcr_el2, x0
> +
> +1:	isb
> +	mov	w0, #SPSR_KERNEL_EL1
> +	ldr	x1, =spsr_to_elx
> +	str	w0, [x1]
> +	b	el_max_init
> +
>  el_max_init:
>  	ldr	x0, =CNTFRQ
>  	msr	cntfrq_el0, x0
> @@ -169,10 +212,18 @@ jump_kernel:
>  	 */
>  	bfi	x4, x19, #5, #1
>  
> +	mrs	x5, CurrentEL
> +	cmp	x5, #CURRENTEL_EL2
> +	b.eq	1f
> +
>  	msr	elr_el3, x19
>  	msr	spsr_el3, x4
>  	eret
>  
> +1:	msr	elr_el2, x19
> +	msr	spsr_el2, x4
> +	eret
> +
>  	.ltorg
>  
>  	.data