commit 72cd4064fcca ("ARM: 8830/1: NOMMU: Toggle only bits in
EXC_RETURN we are really care of") only sets BIT[3] for Thread mode
and BIT[2] for PSP, it leaves BIT[4] untouched. But there's such a
case: the pre-linux env makes use of FPU then the BIT[4] in 'lr' is
cleared, this brings an umatch issue since the NOMMU kernel doesn't
support FPU yet. Before the above commit, we hardcode the exc_ret
as 0xfffffffd, so the flow works fine.

Fix this issue by explicitly set BIT[4] which means using standard
stack frame.

Fixes: 72cd4064fcca ("ARM: 8830/1: NOMMU: Toggle only bits in EXC_RETURN we are really care of")
Signed-off-by: Jisheng Zhang <jszhang@kernel.org>
---
 arch/arm/include/asm/v7m.h | 1 +
 arch/arm/mm/proc-v7m.S     | 1 +
 2 files changed, 2 insertions(+)

diff --git a/arch/arm/include/asm/v7m.h b/arch/arm/include/asm/v7m.h
index 4512f7e1918f..3aea6d3c97ee 100644
--- a/arch/arm/include/asm/v7m.h
+++ b/arch/arm/include/asm/v7m.h
@@ -51,6 +51,7 @@
  */
 #define EXC_RET_STACK_MASK			0x00000004
 #define EXC_RET_THREADMODE_PROCESSSTACK		(3 << 2)
+#define EXC_RET_FTYPE				(1 << 4)
 
 /* Cache related definitions */
 
diff --git a/arch/arm/mm/proc-v7m.S b/arch/arm/mm/proc-v7m.S
index ed7781c84341..fdae077d2654 100644
--- a/arch/arm/mm/proc-v7m.S
+++ b/arch/arm/mm/proc-v7m.S
@@ -138,6 +138,7 @@ __v7m_setup_cont:
 1:	cpsid	i
 	/* Calculate exc_ret */
 	orr	r10, lr, #EXC_RET_THREADMODE_PROCESSSTACK
+	orr	r10, #EXC_RET_FTYPE
 	ldmia	sp, {r0-r3, r12}
 	str	r5, [r12, #11 * 4]	@ restore the original SVC vector entry
 	mov	lr, r6			@ restore LR
-- 
2.45.2

On 10/13/24 14:25, Jisheng Zhang wrote:
> commit 72cd4064fcca ("ARM: 8830/1: NOMMU: Toggle only bits in
> EXC_RETURN we are really care of") only sets BIT[3] for Thread mode
> and BIT[2] for PSP, it leaves BIT[4] untouched. But there's such a
> case: the pre-linux env makes use of FPU then the BIT[4] in 'lr' is
> cleared, this brings an umatch issue since the NOMMU kernel doesn't

Can pre-linux env disable FPU before passing control to kernel (which
is, as correctly pointed, doesn't know how to use FPU)?

Cheers
Vladimir

On Mon, Oct 14, 2024 at 10:54:00AM +0100, Vladimir Murzin wrote:
> On 10/13/24 14:25, Jisheng Zhang wrote:
> > commit 72cd4064fcca ("ARM: 8830/1: NOMMU: Toggle only bits in
> > EXC_RETURN we are really care of") only sets BIT[3] for Thread mode
> > and BIT[2] for PSP, it leaves BIT[4] untouched. But there's such a
> > case: the pre-linux env makes use of FPU then the BIT[4] in 'lr' is
> > cleared, this brings an umatch issue since the NOMMU kernel doesn't
> 
> Can pre-linux env disable FPU before passing control to kernel (which
> is, as correctly pointed, doesn't know how to use FPU)?

IIRC, I did a experiment like this by clearing the SCB CPACR related
bits, but the stack frame type is still not correct. I searched in the
armv8m arm or cortex-mN's TRM I didn't find the relation between the
returned "lr" and the pre FPU usage.

What's more, IMHO, kernel needs to use the correct exc_ret no matter
the pre-linux env does, I.E if kernel/userspace supports FPU, then
clear BIT[4]; if no, set BIT[4].

PS: this is a regression: before the commit, the have-used-fpu pre-linux
env + linux nommu combination works; after the commit, it fails to
execute the init due to wrong stack frame type.

Thanks

On 10/15/24 01:02, Jisheng Zhang wrote:
> On Mon, Oct 14, 2024 at 10:54:00AM +0100, Vladimir Murzin wrote:
>> On 10/13/24 14:25, Jisheng Zhang wrote:
>>> commit 72cd4064fcca ("ARM: 8830/1: NOMMU: Toggle only bits in
>>> EXC_RETURN we are really care of") only sets BIT[3] for Thread mode
>>> and BIT[2] for PSP, it leaves BIT[4] untouched. But there's such a
>>> case: the pre-linux env makes use of FPU then the BIT[4] in 'lr' is
>>> cleared, this brings an umatch issue since the NOMMU kernel doesn't
>> Can pre-linux env disable FPU before passing control to kernel (which
>> is, as correctly pointed, doesn't know how to use FPU)?
> IIRC, I did a experiment like this by clearing the SCB CPACR related
> bits, but the stack frame type is still not correct. I searched in the
> armv8m arm or cortex-mN's TRM I didn't find the relation between the
> returned "lr" and the pre FPU usage.
> 
> What's more, IMHO, kernel needs to use the correct exc_ret no matter
> the pre-linux env does, I.E if kernel/userspace supports FPU, then
> clear BIT[4]; if no, set BIT[4].
> 
> PS: this is a regression: before the commit, the have-used-fpu pre-linux
> env + linux nommu combination works; after the commit, it fails to
> execute the init due to wrong stack frame type.
> 
> Thanks
> 

Fair enough.

Reviewed-by: Vladimir Murzin <vladimir.murzin@arm.com>

Please, upload the patch in RMK's patch system [1]

[1] https://www.arm.linux.org.uk/developer/patches/


Vladimir