When an M-profile CPU is restoring registers from the stack on
exception return, the stack pointer to use is determined based on
bits in the magic exception return type value.  We were not getting
this logic entirely correct.

Whether we use one of the Secure stack pointers or one of the
Non-Secure stack pointers depends on the EXCRET.S bit.  However,
whether we use the MSP or the PSP then depends on the SPSEL bit in
either the CONTROL_S or CONTROL_NS register.  We were incorrectly
selecting MSP vs PSP based on the EXCRET.SPSEL bit.

(In the pseudocode this is in the PopStack() function, which calls
LookUpSp_with_security_mode() which in turn looks at the relevant
CONTROL.SPSEL bit.)

The buggy behaviour wasn't noticeable in most cases, because we write
EXCRET.SPSEL to the CONTROL.SPSEL bit for the S/NS register selected
by EXCRET.ES, so we only do the wrong thing when EXCRET.S and
EXCRET.ES are different.  This will happen when secure code takes a
secure exception, which then tail-chains to a non-secure exception
which finally returns to the original secure code.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
This was noticed by the Arm TF-M folks, who have a pending change
to their code which will run into the situation we mishandle.
---
 target/arm/m_helper.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/target/arm/m_helper.c b/target/arm/m_helper.c
index d63ae465e1e..eda74e55450 100644
--- a/target/arm/m_helper.c
+++ b/target/arm/m_helper.c
@@ -1597,10 +1597,11 @@ static void do_v7m_exception_exit(ARMCPU *cpu)
          * We use this limited C variable scope so we don't accidentally
          * use 'frame_sp_p' after we do something that makes it invalid.
          */
+        bool spsel = env->v7m.control[return_to_secure] & R_V7M_CONTROL_SPSEL_MASK;
         uint32_t *frame_sp_p = get_v7m_sp_ptr(env,
                                               return_to_secure,
                                               !return_to_handler,
-                                              return_to_sp_process);
+                                              spsel);
         uint32_t frameptr = *frame_sp_p;
         bool pop_ok = true;
         ARMMMUIdx mmu_idx;
-- 
2.20.1

On 5/20/21 6:09 AM, Peter Maydell wrote:
> When an M-profile CPU is restoring registers from the stack on
> exception return, the stack pointer to use is determined based on
> bits in the magic exception return type value.  We were not getting
> this logic entirely correct.
> 
> Whether we use one of the Secure stack pointers or one of the
> Non-Secure stack pointers depends on the EXCRET.S bit.  However,
> whether we use the MSP or the PSP then depends on the SPSEL bit in
> either the CONTROL_S or CONTROL_NS register.  We were incorrectly
> selecting MSP vs PSP based on the EXCRET.SPSEL bit.
> 
> (In the pseudocode this is in the PopStack() function, which calls
> LookUpSp_with_security_mode() which in turn looks at the relevant
> CONTROL.SPSEL bit.)
> 
> The buggy behaviour wasn't noticeable in most cases, because we write
> EXCRET.SPSEL to the CONTROL.SPSEL bit for the S/NS register selected
> by EXCRET.ES, so we only do the wrong thing when EXCRET.S and
> EXCRET.ES are different.  This will happen when secure code takes a
> secure exception, which then tail-chains to a non-secure exception
> which finally returns to the original secure code.
> 
> Signed-off-by: Peter Maydell<peter.maydell@linaro.org>
> ---
> This was noticed by the Arm TF-M folks, who have a pending change
> to their code which will run into the situation we mishandle.
> ---

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>

r~