Re: [Qemu-devel] target/arm: Fix abort on exception return from AArch64 to AArch32

["Alex Bennée" <alex.bennee@linaro.org>]

Peter Maydell <peter.maydell@linaro.org> writes:

> On 12 July 2017 at 12:39, Jaroslaw Pelczar <j.pelczar@samsung.com> wrote:
>> Problem manifests itself when we handle the following sequence:
>>
>> 1. 64-bit Secure EL3 returns to 32-bit Secure EL1
>> 2. 32-bit Secure EL1 performs SMC call to 64-bit Secure EL3
>> 3. 64-bit Secure EL3 performs return ERET to 32-bit Secure EL1]
>> 4. 32-bit Secure EL1 receives prefetch abort
>>
>> If CPU's env->pc is not set to the same value as env->regs[15],
>> during ERET the simulator will try to fetch instruction from EL3's
>> virtual address inside 32-bit Secure EL1 virtual address space.
>> This will cause Prefetch Abort in 32-bit Secure EL1.
>>
>> Problem occurs because of the following code generation scheme:
>>
>> 1. disas_uncond_b_reg will decode ERET
>> 2. gen_helper_exception_return(cpu_env) will generate thunk to
>>    helper_exception_return
>> 3. s->is_jmp is set to DISAS_JUMP
>> 4. gen_intermediate_code_a64 will see dc->is_jmp == DISAS_JUMP
>> 5. tcg_gen_lookup_and_goto_ptr(cpu_pc) will be called which trigger
>>    access cpu->pc to fetch next opcode
>>
>> At this point cpu->pc will have EL3's PC value.
>>
>> Signed-off-by: Jaroslaw Pelczar <j.pelczar@samsung.com>
>> ---
>>  target/arm/op_helper.c | 3 +++
>>  1 file changed, 3 insertions(+)
>>
>> diff --git a/target/arm/op_helper.c b/target/arm/op_helper.c
>> index 2a85666..8222c19 100644
>> --- a/target/arm/op_helper.c
>> +++ b/target/arm/op_helper.c
>> @@ -1027,6 +1027,9 @@ void HELPER(exception_return)(CPUARMState *env)
>>          } else {
>>              env->regs[15] = env->elr_el[cur_el] & ~0x3;
>>          }
>> +
>> +        env->pc = env->regs[15];
>> +
>>          qemu_log_mask(CPU_LOG_INT, "Exception return from AArch64 EL%d to "
>>                        "AArch32 EL%d PC 0x%" PRIx32 "\n",
>>                        cur_el, new_el, env->regs[15]);
>
> Thanks for the bug report. I don't think this is the best
> fix, though. What should happen is that for ERET we do
> not attempt to do a lookup_and_goto_ptr(), we should just
> go back out to the main loop and get the next TB from there.
>
> I think this may be fixed as one of the things Alex's
> recent patchset tackles?

Yes, see:

  https://lists.nongnu.org/archive/html/qemu-devel/2017-07/msg02963.html

I'll probably get v4 out tomorrow at which point you can merge it ;-)

>
> thanks
> -- PMM


--
Alex Bennée