Qemu and ARM secure state.

Qemu and ARM secure state.

[Jean-Christophe DUBOIS]

[-- Attachment #1: Type: text/plain, Size: 1699 bytes --]

Hi,

I have a little application that is designed to work on the i.MX6UL 
processor.

I developed it and tested it on the mcimx6ul-evk platform emulated by Qemu.

This application used to work "flawlessly" on Qemu 5.0.50 and is working 
on Qemu 6.0.0 (available as a pre-built package on the latest Ubuntu).

But when I try to run the exact same command line on a Qemu version I 
compile myself from main/latest of github (Qemu 6.1.50), my application 
fails to start.

So a little background:

My application expects to start in "secure" state and supervisor mode 
(which is the default state of i.MX6UL when booting barebone [without 
u-boot]).

 From this state the application tries to get to "non secure" / 
hypervisor mode which imply going to the "secure" / monitor state before 
being able to drop to "non secure" / hypervisor. To do so is runs a "smc 
0" operand (from "secure" / supervisor).

This "smc" instruction is processed "as expected" by Qemu 5.0.50 and 
Qemu 6.0.0 (getting to "secure" / monitor mode) but on Qemu 6.1.50 
(latest from github) it is as if the smc operand was a no-op. It doesn't 
trigger any exception and the processor just get to the next instruction 
after the "smc" instruction. So I am a bit puzzled.

  * Is there something that changed in Qemu (since Qemu 6.0.0) when it
    comes to the "secure" world/state?
  * Is there some additional command line parameters to use (I search in
    the documentation but without luck) to get secure world behavior ?
  * Is it necessary to "adapt" the emulated platform
    (i.MX6UL/mcimx6ul-evk) in some way (it looks like the "virt" machine
    with "secure=on" does work for arm platform)?

Thanks for your help.

JC

[-- Attachment #2: Type: text/html, Size: 2128 bytes --]

Re: Qemu and ARM secure state.

[Peter Maydell]

On Wed, 3 Nov 2021 at 13:27, Jean-Christophe DUBOIS <jcd@tribudubois.net> wrote:
> I have a little application that is designed to work on the i.MX6UL processor.
>
> I developed it and tested it on the mcimx6ul-evk platform emulated by Qemu.
>
> This application used to work "flawlessly" on Qemu 5.0.50 and is working on Qemu 6.0.0 (available as a pre-built package on the latest Ubuntu).
>
> But when I try to run the exact same command line on a Qemu version I compile myself from main/latest of github (Qemu 6.1.50), my application fails to start.
>
> So a little background:
>
> My application expects to start in "secure" state and supervisor mode (which is the default state of i.MX6UL when booting barebone [without u-boot]).
>
> From this state the application tries to get to "non secure" / hypervisor mode which imply going to the "secure" / monitor state before being able to drop to "non secure" / hypervisor. To do so is runs a "smc 0" operand (from "secure" / supervisor).
>
> This "smc" instruction is processed "as expected" by Qemu 5.0.50 and Qemu 6.0.0 (getting to "secure" / monitor mode) but on Qemu 6.1.50 (latest from github) it is as if the smc operand was a no-op. It doesn't trigger any exception and the processor just get to the next instruction after the "smc" instruction. So I am a bit puzzled.
>
> Is there something that changed in Qemu (since Qemu 6.0.0) when it comes to the "secure" world/state?
> Is there some additional command line parameters to use (I search in the documentation but without luck) to get secure world behavior ?
> Is it necessary to "adapt" the emulated platform (i.MX6UL/mcimx6ul-evk) in some way (it looks like the "virt" machine with "secure=on" does work for arm platform)?

Could you try doing a bisect to find the QEMU commit that caused
your guest to stop working ?

thanks
-- PMM

Re: Qemu and ARM secure state.

[Jean-Christophe DUBOIS]

[-- Attachment #1: Type: text/plain, Size: 2244 bytes --]

Le 04/11/2021 à 12:11, Peter Maydell a écrit :
> On Wed, 3 Nov 2021 at 13:27, Jean-Christophe DUBOIS<jcd@tribudubois.net>  wrote:
>> I have a little application that is designed to work on the i.MX6UL processor.
>>
>> I developed it and tested it on the mcimx6ul-evk platform emulated by Qemu.
>>
>> This application used to work "flawlessly" on Qemu 5.0.50 and is working on Qemu 6.0.0 (available as a pre-built package on the latest Ubuntu).
>>
>> But when I try to run the exact same command line on a Qemu version I compile myself from main/latest of github (Qemu 6.1.50), my application fails to start.
>>
>> So a little background:
>>
>> My application expects to start in "secure" state and supervisor mode (which is the default state of i.MX6UL when booting barebone [without u-boot]).
>>
>>  From this state the application tries to get to "non secure" / hypervisor mode which imply going to the "secure" / monitor state before being able to drop to "non secure" / hypervisor. To do so is runs a "smc 0" operand (from "secure" / supervisor).
>>
>> This "smc" instruction is processed "as expected" by Qemu 5.0.50 and Qemu 6.0.0 (getting to "secure" / monitor mode) but on Qemu 6.1.50 (latest from github) it is as if the smc operand was a no-op. It doesn't trigger any exception and the processor just get to the next instruction after the "smc" instruction. So I am a bit puzzled.
>>
>> Is there something that changed in Qemu (since Qemu 6.0.0) when it comes to the "secure" world/state?
>> Is there some additional command line parameters to use (I search in the documentation but without luck) to get secure world behavior ?
>> Is it necessary to "adapt" the emulated platform (i.MX6UL/mcimx6ul-evk) in some way (it looks like the "virt" machine with "secure=on" does work for arm platform)?
> Could you try doing a bisect to find the QEMU commit that caused
> your guest to stop working ?

OK, I did the bisect and the commit that break the i.MX6UL behavior for 
my program is commit 9fcd15b9193e819b6cc2fd0a45e3506148812bb4 (arm: tcg: 
Adhere to SMCCC 1.3 section 5.2).

Before it the SMC instruction would trigger a monitor exception.

After it the SMC instruction is acting like a no-op.

Thanks

JC


>
> thanks
> -- PMM
>

[-- Attachment #2: Type: text/html, Size: 3309 bytes --]

Re: Qemu and ARM secure state.

[Jean-Christophe DUBOIS]

[-- Attachment #1: Type: text/plain, Size: 3843 bytes --]

So, I am trying to understand:

Contrary to what I said, In my case the SMC instruction is not really a 
"no-op" as it sets R0 to -1 (0xffffffff) to indicate an unknown PSCI 
service (by the Qemu internal PSCI handler).

With the new code introduced by the "arm: tcg: Adhere to SMCCC 1.3 
section 5.2" commit, once a processor/platform configure things to have 
PSCI requests handled by Qemu code (with "psci-conduit" attribute set to 
QEMU_PSCI_CONDUIT_SMC for example), then any exception raised by an 
"SMC" instruction will be only handled by the Qemu internal code and 
will no call the monitor related code in the guest/OS application. This 
seems to be why my SMC monitor handler is not called anymore in my case.

As my i.MX6UL is a mono-processor platform I don't really need to set 
the "psci-conduit" attribute (which really makes sense when you have a 
cluster of 2 or more cores I guess). As a matter of fact if I remove the 
"psci-conduit" attribute setting from the i.MX6UL processor file, my 
application is working again on main/latest.

But this still raises the question to know if the current behavior for 
processors with "psci-conduit" set to SMC or HVC is correct. For example 
an i.MX7 based platform (with up to 4 cortex A7 cores) would not be able 
to trigger OS SMC handler as the exception would be entirely processed 
by the Qemu internal code (with CR generally set to -1 in R0 to indicate 
unknown PSCI request).

Is there something I am missing?

Regards

JC

Le 04/11/2021 à 22:11, Jean-Christophe DUBOIS a écrit :
> Le 04/11/2021 à 12:11, Peter Maydell a écrit :
>> On Wed, 3 Nov 2021 at 13:27, Jean-Christophe DUBOIS<jcd@tribudubois.net>  wrote:
>>> I have a little application that is designed to work on the i.MX6UL processor.
>>>
>>> I developed it and tested it on the mcimx6ul-evk platform emulated by Qemu.
>>>
>>> This application used to work "flawlessly" on Qemu 5.0.50 and is working on Qemu 6.0.0 (available as a pre-built package on the latest Ubuntu).
>>>
>>> But when I try to run the exact same command line on a Qemu version I compile myself from main/latest of github (Qemu 6.1.50), my application fails to start.
>>>
>>> So a little background:
>>>
>>> My application expects to start in "secure" state and supervisor mode (which is the default state of i.MX6UL when booting barebone [without u-boot]).
>>>
>>> >From this state the application tries to get to "non secure" / hypervisor mode which imply going to the "secure" / monitor state before being able to drop to "non secure" / hypervisor. To do so is runs a "smc 0" operand (from "secure" / supervisor).
>>>
>>> This "smc" instruction is processed "as expected" by Qemu 5.0.50 and Qemu 6.0.0 (getting to "secure" / monitor mode) but on Qemu 6.1.50 (latest from github) it is as if the smc operand was a no-op. It doesn't trigger any exception and the processor just get to the next instruction after the "smc" instruction. So I am a bit puzzled.
>>>
>>> Is there something that changed in Qemu (since Qemu 6.0.0) when it comes to the "secure" world/state?
>>> Is there some additional command line parameters to use (I search in the documentation but without luck) to get secure world behavior ?
>>> Is it necessary to "adapt" the emulated platform (i.MX6UL/mcimx6ul-evk) in some way (it looks like the "virt" machine with "secure=on" does work for arm platform)?
>> Could you try doing a bisect to find the QEMU commit that caused
>> your guest to stop working ?
>
> OK, I did the bisect and the commit that break the i.MX6UL behavior 
> for my program is commit 9fcd15b9193e819b6cc2fd0a45e3506148812bb4 
> (arm: tcg: Adhere to SMCCC 1.3 section 5.2).
>
> Before it the SMC instruction would trigger a monitor exception.
>
> After it the SMC instruction is acting like a no-op.
>
> Thanks
>
> JC
>
>
>> thanks
>> -- PMM
>>
>

[-- Attachment #2: Type: text/html, Size: 6052 bytes --]

Re: Qemu and ARM secure state.

[Jean-Christophe DUBOIS]

[-- Attachment #1: Type: text/plain, Size: 4399 bytes --]

So it seems that what is needed is a way to choose on the command line 
if we want to enable the Qemu built-in PSCI implementation (because we 
are booting linux for example) or if we really want a bare metal 
processor (because we are booting a trustedOS like optee).

The "virt" platform allows to dynamically choose one or the other. Other 
platforms seems to need the same feature.

JC

Le 06/11/2021 à 11:04, Jean-Christophe DUBOIS a écrit :
> So, I am trying to understand:
>
> Contrary to what I said, In my case the SMC instruction is not really 
> a "no-op" as it sets R0 to -1 (0xffffffff) to indicate an unknown PSCI 
> service (by the Qemu internal PSCI handler).
>
> With the new code introduced by the "arm: tcg: Adhere to SMCCC 1.3 
> section 5.2" commit, once a processor/platform configure things to 
> have PSCI requests handled by Qemu code (with "psci-conduit" attribute 
> set to QEMU_PSCI_CONDUIT_SMC for example), then any exception raised 
> by an "SMC" instruction will be only handled by the Qemu internal code 
> and will no call the monitor related code in the guest/OS application. 
> This seems to be why my SMC monitor handler is not called anymore in 
> my case.
>
> As my i.MX6UL is a mono-processor platform I don't really need to set 
> the "psci-conduit" attribute (which really makes sense when you have a 
> cluster of 2 or more cores I guess). As a matter of fact if I remove 
> the "psci-conduit" attribute setting from the i.MX6UL processor file, 
> my application is working again on main/latest.
>
> But this still raises the question to know if the current behavior for 
> processors with "psci-conduit" set to SMC or HVC is correct. For 
> example an i.MX7 based platform (with up to 4 cortex A7 cores) would 
> not be able to trigger OS SMC handler as the exception would be 
> entirely processed by the Qemu internal code (with CR generally set to 
> -1 in R0 to indicate unknown PSCI request).
>
> Is there something I am missing?
>
> Regards
>
> JC
>
> Le 04/11/2021 à 22:11, Jean-Christophe DUBOIS a écrit :
>> Le 04/11/2021 à 12:11, Peter Maydell a écrit :
>>> On Wed, 3 Nov 2021 at 13:27, Jean-Christophe DUBOIS<jcd@tribudubois.net>  wrote:
>>>> I have a little application that is designed to work on the i.MX6UL processor.
>>>>
>>>> I developed it and tested it on the mcimx6ul-evk platform emulated by Qemu.
>>>>
>>>> This application used to work "flawlessly" on Qemu 5.0.50 and is working on Qemu 6.0.0 (available as a pre-built package on the latest Ubuntu).
>>>>
>>>> But when I try to run the exact same command line on a Qemu version I compile myself from main/latest of github (Qemu 6.1.50), my application fails to start.
>>>>
>>>> So a little background:
>>>>
>>>> My application expects to start in "secure" state and supervisor mode (which is the default state of i.MX6UL when booting barebone [without u-boot]).
>>>>
>>>> >From this state the application tries to get to "non secure" / hypervisor mode which imply going to the "secure" / monitor state before being able to drop to "non secure" / hypervisor. To do so is runs a "smc 0" operand (from "secure" / supervisor).
>>>>
>>>> This "smc" instruction is processed "as expected" by Qemu 5.0.50 and Qemu 6.0.0 (getting to "secure" / monitor mode) but on Qemu 6.1.50 (latest from github) it is as if the smc operand was a no-op. It doesn't trigger any exception and the processor just get to the next instruction after the "smc" instruction. So I am a bit puzzled.
>>>>
>>>> Is there something that changed in Qemu (since Qemu 6.0.0) when it comes to the "secure" world/state?
>>>> Is there some additional command line parameters to use (I search in the documentation but without luck) to get secure world behavior ?
>>>> Is it necessary to "adapt" the emulated platform (i.MX6UL/mcimx6ul-evk) in some way (it looks like the "virt" machine with "secure=on" does work for arm platform)?
>>> Could you try doing a bisect to find the QEMU commit that caused
>>> your guest to stop working ?
>>
>> OK, I did the bisect and the commit that break the i.MX6UL behavior 
>> for my program is commit 9fcd15b9193e819b6cc2fd0a45e3506148812bb4 
>> (arm: tcg: Adhere to SMCCC 1.3 section 5.2).
>>
>> Before it the SMC instruction would trigger a monitor exception.
>>
>> After it the SMC instruction is acting like a no-op.
>>
>> Thanks
>>
>> JC
>>
>>
>>> thanks
>>> -- PMM
>>>
>>
>

[-- Attachment #2: Type: text/html, Size: 7245 bytes --]

Re: Qemu and ARM secure state.

[Jean-Christophe DUBOIS]

[-- Attachment #1: Type: text/plain, Size: 4959 bytes --]

One small question/remark:

According to the the "Arm Power State Coordinate Interface" (DEN0022D.b) 
document (chapter 5) PSCI calls can only be issued by "normal world" 
(EL1 or EL2). Therefore, should we be adding a test for the current 
secure state in the arm_is_psci_call() function? This would prevent 
calling the built-in Qemu PSCI function if SMC is issued  from secure state.

Regards.

JC

Le 06/11/2021 à 14:04, Jean-Christophe DUBOIS a écrit :
> So it seems that what is needed is a way to choose on the command line 
> if we want to enable the Qemu built-in PSCI implementation (because we 
> are booting linux for example) or if we really want a bare metal 
> processor (because we are booting a trustedOS like optee).
>
> The "virt" platform allows to dynamically choose one or the other. 
> Other platforms seems to need the same feature.
>
> JC
>
> Le 06/11/2021 à 11:04, Jean-Christophe DUBOIS a écrit :
>> So, I am trying to understand:
>>
>> Contrary to what I said, In my case the SMC instruction is not really 
>> a "no-op" as it sets R0 to -1 (0xffffffff) to indicate an unknown 
>> PSCI service (by the Qemu internal PSCI handler).
>>
>> With the new code introduced by the "arm: tcg: Adhere to SMCCC 1.3 
>> section 5.2" commit, once a processor/platform configure things to 
>> have PSCI requests handled by Qemu code (with "psci-conduit" 
>> attribute set to QEMU_PSCI_CONDUIT_SMC for example), then any 
>> exception raised by an "SMC" instruction will be only handled by the 
>> Qemu internal code and will no call the monitor related code in the 
>> guest/OS application. This seems to be why my SMC monitor handler is 
>> not called anymore in my case.
>>
>> As my i.MX6UL is a mono-processor platform I don't really need to set 
>> the "psci-conduit" attribute (which really makes sense when you have 
>> a cluster of 2 or more cores I guess). As a matter of fact if I 
>> remove the "psci-conduit" attribute setting from the i.MX6UL 
>> processor file, my application is working again on main/latest.
>>
>> But this still raises the question to know if the current behavior 
>> for processors with "psci-conduit" set to SMC or HVC is correct. For 
>> example an i.MX7 based platform (with up to 4 cortex A7 cores) would 
>> not be able to trigger OS SMC handler as the exception would be 
>> entirely processed by the Qemu internal code (with CR generally set 
>> to -1 in R0 to indicate unknown PSCI request).
>>
>> Is there something I am missing?
>>
>> Regards
>>
>> JC
>>
>> Le 04/11/2021 à 22:11, Jean-Christophe DUBOIS a écrit :
>>> Le 04/11/2021 à 12:11, Peter Maydell a écrit :
>>>> On Wed, 3 Nov 2021 at 13:27, Jean-Christophe DUBOIS<jcd@tribudubois.net>  wrote:
>>>>> I have a little application that is designed to work on the i.MX6UL processor.
>>>>>
>>>>> I developed it and tested it on the mcimx6ul-evk platform emulated by Qemu.
>>>>>
>>>>> This application used to work "flawlessly" on Qemu 5.0.50 and is working on Qemu 6.0.0 (available as a pre-built package on the latest Ubuntu).
>>>>>
>>>>> But when I try to run the exact same command line on a Qemu version I compile myself from main/latest of github (Qemu 6.1.50), my application fails to start.
>>>>>
>>>>> So a little background:
>>>>>
>>>>> My application expects to start in "secure" state and supervisor mode (which is the default state of i.MX6UL when booting barebone [without u-boot]).
>>>>>
>>>>> >From this state the application tries to get to "non secure" / hypervisor mode which imply going to the "secure" / monitor state before being able to drop to "non secure" / hypervisor. To do so is runs a "smc 0" operand (from "secure" / supervisor).
>>>>>
>>>>> This "smc" instruction is processed "as expected" by Qemu 5.0.50 and Qemu 6.0.0 (getting to "secure" / monitor mode) but on Qemu 6.1.50 (latest from github) it is as if the smc operand was a no-op. It doesn't trigger any exception and the processor just get to the next instruction after the "smc" instruction. So I am a bit puzzled.
>>>>>
>>>>> Is there something that changed in Qemu (since Qemu 6.0.0) when it comes to the "secure" world/state?
>>>>> Is there some additional command line parameters to use (I search in the documentation but without luck) to get secure world behavior ?
>>>>> Is it necessary to "adapt" the emulated platform (i.MX6UL/mcimx6ul-evk) in some way (it looks like the "virt" machine with "secure=on" does work for arm platform)?
>>>> Could you try doing a bisect to find the QEMU commit that caused
>>>> your guest to stop working ?
>>>
>>> OK, I did the bisect and the commit that break the i.MX6UL behavior 
>>> for my program is commit 9fcd15b9193e819b6cc2fd0a45e3506148812bb4 
>>> (arm: tcg: Adhere to SMCCC 1.3 section 5.2).
>>>
>>> Before it the SMC instruction would trigger a monitor exception.
>>>
>>> After it the SMC instruction is acting like a no-op.
>>>
>>> Thanks
>>>
>>> JC
>>>
>>>
>>>> thanks
>>>> -- PMM
>>>>
>>>
>>
>

[-- Attachment #2: Type: text/html, Size: 8634 bytes --]

Re: Qemu and ARM secure state.

[Alex Bennée]

Jean-Christophe DUBOIS <jcd@tribudubois.net> writes:

> One small question/remark:
>
> According to the the "Arm Power State Coordinate Interface" (DEN0022D.b) document (chapter 5) PSCI calls can only be issued by
> "normal world" (EL1 or EL2). Therefore, should we be adding a test for the current secure state in the arm_is_psci_call() function? This
> would prevent calling the built-in Qemu PSCI function if SMC is issued
> from secure state.

All that should be handled in:

  void HELPER(pre_smc)(CPUARMState *env, uint32_t syndrome)

which should cause things to be trapped if the CPU is in an invalid
state to execute the SMC instruction. If the exception is a valid SMC we
end up in arm_cpu_do_interrupt where we may divert to
arm_handle_psci_call.

-- 
Alex Bennée

Re: Qemu and ARM secure state.

[Alex Bennée]

Jean-Christophe DUBOIS <jcd@tribudubois.net> writes:

> One small question/remark:
>
> According to the the "Arm Power State Coordinate Interface" (DEN0022D.b) document (chapter 5) PSCI calls can only be issued by
> "normal world" (EL1 or EL2). Therefore, should we be adding a test for the current secure state in the arm_is_psci_call() function? This
> would prevent calling the built-in Qemu PSCI function if SMC is issued
> from secure state.

All that should be handled in:

  void HELPER(pre_smc)(CPUARMState *env, uint32_t syndrome)

which should cause things to be trapped if the CPU is in an invalid
state to execute the SMC instruction. If the exception is a valid SMC we
end up in arm_cpu_do_interrupt where we may divert to
arm_handle_psci_call.

-- 
Alex Bennée

Re: Qemu and ARM secure state.

[Jean-Christophe DUBOIS]

Le 08/11/2021 à 15:14, Alex Bennée a écrit :
> Jean-Christophe DUBOIS <jcd@tribudubois.net> writes:
>
>> One small question/remark:
>>
>> According to the the "Arm Power State Coordinate Interface" (DEN0022D.b) document (chapter 5) PSCI calls can only be issued by
>> "normal world" (EL1 or EL2). Therefore, should we be adding a test for the current secure state in the arm_is_psci_call() function? This
>> would prevent calling the built-in Qemu PSCI function if SMC is issued
>> from secure state.
> All that should be handled in:
>
>    void HELPER(pre_smc)(CPUARMState *env, uint32_t syndrome)
>
> which should cause things to be trapped if the CPU is in an invalid
> state to execute the SMC instruction. If the exception is a valid SMC we
> end up in arm_cpu_do_interrupt where we may divert to
> arm_handle_psci_call.

The problem I have is that is seems that once the "psci-conduit" is set 
to SMC, all SMC exception would be handled only by the diverted  
arm_handle_psci_call().

It seems there is no way to handle SMC exception in my software once 
"psci-conduit" is set to SMC on a platform.

It used to be that only some PSCI services were diverted and other SMC 
services were handled by the EL3 software if any (and this split was 
actually also a bit awkward).

>

Re: Qemu and ARM secure state.

[Peter Maydell]

On Sat, 6 Nov 2021 at 18:11, Jean-Christophe DUBOIS <jcd@tribudubois.net> wrote:
>
> One small question/remark:
>
> According to the the "Arm Power State Coordinate Interface" (DEN0022D.b) document (chapter 5) PSCI calls can only be issued by "normal world" (EL1 or EL2). Therefore, should we be adding a test for the current secure state in the arm_is_psci_call() function? This would prevent calling the built-in Qemu PSCI function if SMC is issued  from secure state.

This shouldn't matter, because if the machine model is configured
to execute guest code in EL3 at all then it should not be enabling
QEMU's internal PSCI support. The internal PSCI stuff is only
there as a kind of "emulated firmware" for when we're running
guest code that starts at EL2 (notably, when directly booting
a Linux kernel).

The problem seems to be that fsl_imx6ul_realize() and
fsl_imx7_realize() unconditionally enable PSCI-via-SMC.
The imx7 code also puts all the secondaries into
PSCI-powered-off mode -- this should be checked to
work out what the right thing is if we're not doing
emulated PSCI and instead starting the guest at EL3.

-- PMM

Re: Qemu and ARM secure state.

[Jean-Christophe DUBOIS]

Le 08/11/2021 à 15:50, Peter Maydell a écrit :
> On Sat, 6 Nov 2021 at 18:11, Jean-Christophe DUBOIS <jcd@tribudubois.net> wrote:
>> One small question/remark:
>>
>> According to the the "Arm Power State Coordinate Interface" (DEN0022D.b) document (chapter 5) PSCI calls can only be issued by "normal world" (EL1 or EL2). Therefore, should we be adding a test for the current secure state in the arm_is_psci_call() function? This would prevent calling the built-in Qemu PSCI function if SMC is issued  from secure state.
> This shouldn't matter, because if the machine model is configured
> to execute guest code in EL3 at all then it should not be enabling
> QEMU's internal PSCI support. The internal PSCI stuff is only
> there as a kind of "emulated firmware" for when we're running
> guest code that starts at EL2 (notably, when directly booting
> a Linux kernel).
>
> The problem seems to be that fsl_imx6ul_realize() and
> fsl_imx7_realize() unconditionally enable PSCI-via-SMC.
> The imx7 code also puts all the secondaries into
> PSCI-powered-off mode -- this should be checked to
> work out what the right thing is if we're not doing
> emulated PSCI and instead starting the guest at EL3.

OK, so one problem seems to be that PSCI-via-SMC is enabled on i.MX6UL 
when there is no built in PSCI related function on this processor.

According the Linux DTS, i.MX7 (solo and dual) processors have a 
somewhat PSCI related "entry-method" 
(https://github.com/torvalds/linux/blob/master/arch/arm/boot/dts/imx7s.dtsi). 
But it is not clear to me how this is used and this seems a bit strange 
as "entry-method" seems to be mostly used on arm64 and there is no other 
PSCI related information in the i.MX7 DTS files. As a matter of fact 
previous quad or dual i.MX6 were not supporting PSCI. Instead they were 
using a proprietary method through the internal SRC device (and i.MX7 
also has a similar internal SRC device). But let's assume Linux on i.mx7 
is actually using PSCI to handle processors.

Thinking about it, I guess this might be u-boot that sets an EL3 monitor 
software that is able to handle PSCI requests for the Linux kernel. If 
this is the case, it make sense that Qemu emulates the PSCI services 
normally provided by u-boot to be able to boot linux directly (without 
booting a real u-boot prior to linux). All  is well and nice.

But then if I want to boot and test the u-boot binary (or any trusted OS 
for the matter) on a Qemu emulated i.MX7 (to later boot an hypervisor or 
an OS), it would be rather strange that any PSCI related service 
requested by the hypervisor/OS  would be handled by Qemu directly and 
not by the u-boot code (or any other EL3 code) running on the processor. 
How is it supposed to work? How can I tell Qemu (dynamically?) if I want 
it to provide (or not) the PSCI services (and more generally SMC/HVC 
services). How can I tell it that I want to handle all SMC/EL3 services 
by myself even if the "psci-conduit" is already set to SMC in Qemu?

Am I missing something?

thanks.

JC

>
> -- PMM
>

Re: Qemu and ARM secure state.

[Peter Maydell]

On Mon, 8 Nov 2021 at 22:09, Jean-Christophe DUBOIS <jcd@tribudubois.net> wrote:
> OK, so one problem seems to be that PSCI-via-SMC is enabled on i.MX6UL
> when there is no built in PSCI related function on this processor.
>
> According the Linux DTS, i.MX7 (solo and dual) processors have a
> somewhat PSCI related "entry-method"
> (https://github.com/torvalds/linux/blob/master/arch/arm/boot/dts/imx7s.dtsi).
> But it is not clear to me how this is used and this seems a bit strange
> as "entry-method" seems to be mostly used on arm64 and there is no other
> PSCI related information in the i.MX7 DTS files.

Yeah, PSCI was an interface introduced mostly with aarch64. In the
32-bit world bringing up multiple CPUs was complete anarchy -- the
way the kernel told the secure firmware that it should start up
a secondary core was entirely determined by the firmware, and
the kernel had to have board-specific code to do this. (For the
32-bit imx boards I think this is in arch/arm/mach-imx/src.c.)
For aarch64 we had a clean slate and took the opportunity to insist
that all boards did it the same way, ie using PSCI. (There are other
useful things PSCI allows, but standardising secondary boot is the
one that matters for this discussion.) So if a platform's firmware
implements PSCI, all the dts file has to do is say so, and then
there's no need for board-specific "start secondary CPUs" code.
PSCI does define an aarch32 interface, but there are a lot of
legacy older boards (and new flavours of boards in long-standing
design families) which still do things the old way in aarch32 land.

Typically the top-level "PSCI is available" node is added by the firmware.
(QEMU will do this too when it's emulating PSCI firmware) -- if the
board code enables the psci-conduit it will add an appropriate psci
node in the hw/arm/boot.c code.)

> As a matter of fact
> previous quad or dual i.MX6 were not supporting PSCI. Instead they were
> using a proprietary method through the internal SRC device (and i.MX7
> also has a similar internal SRC device). But let's assume Linux on i.mx7
> is actually using PSCI to handle processors.
>
> Thinking about it, I guess this might be u-boot that sets an EL3 monitor
> software that is able to handle PSCI requests for the Linux kernel. If
> this is the case, it make sense that Qemu emulates the PSCI services
> normally provided by u-boot to be able to boot linux directly (without
> booting a real u-boot prior to linux). All  is well and nice.

Yes, that's the way it works. The EL3 firmware is supposed to provide
PSCI.

For aarch64 the kernel is never entered in EL3 -- it will always run at EL2
or EL1. (This is unlike aarch32, where in some cases you might run the
kernel in secure-SVC, although even there starting the kernel in
NS-SVC or NS-Hyp is more common.)

> But then if I want to boot and test the u-boot binary (or any trusted OS
> for the matter) on a Qemu emulated i.MX7 (to later boot an hypervisor or
> an OS), it would be rather strange that any PSCI related service
> requested by the hypervisor/OS  would be handled by Qemu directly and
> not by the u-boot code (or any other EL3 code) running on the processor.

Exactly. This is why the board code is supposed to set things
up so that if we are starting the guest code in EL1 or EL2
then we enable the PSCI-via-SMC support, and if we're starting
the guest code in EL3 then we do not.

> How is it supposed to work? How can I tell Qemu (dynamically?) if I want
> it to provide (or not) the PSCI services (and more generally SMC/HVC
> services).

If you want PSCI via SMC or HVC, then set the psci-conduit
property on the CPUs to QEMU_PSCI_CONDUIT_SMC or QEMU_PSCI_CONDUIT_HVC.
If you do not want QEMU to provide PSCI, then leave psci-conduit
at its default (which is QEMU_PSCI_CONDUIT_DISABLED).

How can I tell it that I want to handle all SMC/EL3 services
> by myself even if the "psci-conduit" is already set to SMC in Qemu?

It's the imx7 code that's setting psci-conduit, so it should
not do that if it doesn't want that (and also should either
start or not start the secondary cpus powered off, depending
on what the hardware-to-firmware interface is supposed to be.)
This is a bit awkward, because the boards we initially wanted
PSCI for (notably virt) don't have an SoC object, so the code
creating CPUs is in the same source file as the code that knows
whether it's booting a kernel directly or not, and so it just
open-codes the decision logic. With the imx, the CPU creation
is in the source code for the SoC object which is abstracted
away from the board model code. So we'd need to sort out how
to plumb that information into the SoC object (or have the SoC
object's code that creates the CPUs call some function to find out).

-- PMM

Re: Qemu and ARM secure state.

[Jean-Christophe DUBOIS]

[-- Attachment #1: Type: text/plain, Size: 7122 bytes --]

Le 09/11/2021 à 11:55, Peter Maydell a écrit :
> On Mon, 8 Nov 2021 at 22:09, Jean-Christophe DUBOIS<jcd@tribudubois.net>  wrote:
>> OK, so one problem seems to be that PSCI-via-SMC is enabled on i.MX6UL
>> when there is no built in PSCI related function on this processor.
>>
>> According the Linux DTS, i.MX7 (solo and dual) processors have a
>> somewhat PSCI related "entry-method"
>> (https://github.com/torvalds/linux/blob/master/arch/arm/boot/dts/imx7s.dtsi).
>> But it is not clear to me how this is used and this seems a bit strange
>> as "entry-method" seems to be mostly used on arm64 and there is no other
>> PSCI related information in the i.MX7 DTS files.
> Yeah, PSCI was an interface introduced mostly with aarch64. In the
> 32-bit world bringing up multiple CPUs was complete anarchy -- the
> way the kernel told the secure firmware that it should start up
> a secondary core was entirely determined by the firmware, and
> the kernel had to have board-specific code to do this. (For the
> 32-bit imx boards I think this is in arch/arm/mach-imx/src.c.)
> For aarch64 we had a clean slate and took the opportunity to insist
> that all boards did it the same way, ie using PSCI. (There are other
> useful things PSCI allows, but standardising secondary boot is the
> one that matters for this discussion.) So if a platform's firmware
> implements PSCI, all the dts file has to do is say so, and then
> there's no need for board-specific "start secondary CPUs" code.
> PSCI does define an aarch32 interface, but there are a lot of
> legacy older boards (and new flavours of boards in long-standing
> design families) which still do things the old way in aarch32 land.
>
> Typically the top-level "PSCI is available" node is added by the firmware.
> (QEMU will do this too when it's emulating PSCI firmware) -- if the
> board code enables the psci-conduit it will add an appropriate psci
> node in the hw/arm/boot.c code.)
>
>> As a matter of fact
>> previous quad or dual i.MX6 were not supporting PSCI. Instead they were
>> using a proprietary method through the internal SRC device (and i.MX7
>> also has a similar internal SRC device). But let's assume Linux on i.mx7
>> is actually using PSCI to handle processors.
>>
>> Thinking about it, I guess this might be u-boot that sets an EL3 monitor
>> software that is able to handle PSCI requests for the Linux kernel. If
>> this is the case, it make sense that Qemu emulates the PSCI services
>> normally provided by u-boot to be able to boot linux directly (without
>> booting a real u-boot prior to linux). All  is well and nice.
> Yes, that's the way it works. The EL3 firmware is supposed to provide
> PSCI.
>
> For aarch64 the kernel is never entered in EL3 -- it will always run at EL2
> or EL1. (This is unlike aarch32, where in some cases you might run the
> kernel in secure-SVC, although even there starting the kernel in
> NS-SVC or NS-Hyp is more common.)
>
>> But then if I want to boot and test the u-boot binary (or any trusted OS
>> for the matter) on a Qemu emulated i.MX7 (to later boot an hypervisor or
>> an OS), it would be rather strange that any PSCI related service
>> requested by the hypervisor/OS  would be handled by Qemu directly and
>> not by the u-boot code (or any other EL3 code) running on the processor.
> Exactly. This is why the board code is supposed to set things
> up so that if we are starting the guest code in EL1 or EL2
> then we enable the PSCI-via-SMC support, and if we're starting
> the guest code in EL3 then we do not.
>
>> How is it supposed to work? How can I tell Qemu (dynamically?) if I want
>> it to provide (or not) the PSCI services (and more generally SMC/HVC
>> services).
> If you want PSCI via SMC or HVC, then set the psci-conduit
> property on the CPUs to QEMU_PSCI_CONDUIT_SMC or QEMU_PSCI_CONDUIT_HVC.
> If you do not want QEMU to provide PSCI, then leave psci-conduit
> at its default (which is QEMU_PSCI_CONDUIT_DISABLED).
>
> How can I tell it that I want to handle all SMC/EL3 services
>> by myself even if the "psci-conduit" is already set to SMC in Qemu?
> It's the imx7 code that's setting psci-conduit, so it should
> not do that if it doesn't want that (and also should either
> start or not start the secondary cpus powered off, depending
> on what the hardware-to-firmware interface is supposed to be.)
> This is a bit awkward, because the boards we initially wanted
> PSCI for (notably virt) don't have an SoC object, so the code
> creating CPUs is in the same source file as the code that knows
> whether it's booting a kernel directly or not, and so it just
> open-codes the decision logic. With the imx, the CPU creation
> is in the source code for the SoC object which is abstracted
> away from the board model code. So we'd need to sort out how
> to plumb that information into the SoC object (or have the SoC
> object's code that creates the CPUs call some function to find out).

Thanks Peter,

So basically the Qemu i.MX7 processor code needs to set psci-conduit to 
SMC because we want to be able to boot the Linux kernel directly 
(without u-boot) with Qemu emulating the PSCI services when an SMC 
instruction is triggered.

As I see it we also need a way to disable this Qemu PSCI emulation in 
case we want to boot an EL3 software (u-boot, optee or other).

We could certainly pass a command line option to explicitly tell Qemu to 
disable the PCSI emulation. But this would be a bit cumbersome as the 
board level code would then need to call to the SOC code to disable the 
psci setting on all cores.

So I am wondering if we could be more "clever" solution and have things 
sorted out automatically. I would like to suggest 2 solutions;

 1. The first solution could be based on the mode of the processor when
    it starts executing the provided software. If I remember correctly
    what you explained, if an ELF file is provided then the processor
    boots in the highest level (EL3 for i.MX6/i.MX7) while if a non ELF
    file is provided (uImage ...) then the processor boots in EL2 (or
    lower depending on the processor) to emulate the uboot behavior.
     1. if the processor starts the software in EL3 then we need to
        disable the "psci-conduit" through SMC (if set) because the EL3
        software should install/provide the necessary firmware.
     2. If the processor starts the software in EL2 then we need to
        disable the "psci-conduit" through HVC (if set) because the EL2
        software should install/provide the necessary firmware.
 2. The other solution is to check if the booted software has installed
    an EL3 or EL2 exception vector table.
     1. if the software has installed an EL3 (monitor) exception vector
        table then we need to disable the "psci-conduit" through SMC (if
        set).
     2. if the software has installed an EL2 (hypervisor) exception
        vector table then we need to disable the "psci-conduit" through
        HVC (if set).

Is any of these proposals reasonable in your point of view? This would 
apply for any (ARM) target and this seems generic enough.

Thanks

JC

>
> -- PMM
>

[-- Attachment #2: Type: text/html, Size: 8746 bytes --]

Re: Qemu and ARM secure state.

[Peter Maydell]

On Tue, 9 Nov 2021 at 19:06, Jean-Christophe DUBOIS <jcd@tribudubois.net> wrote:
> So basically the Qemu i.MX7 processor code needs to set psci-conduit to SMC
> because we want to be able to boot the Linux kernel directly (without u-boot)
> with Qemu emulating the PSCI services when an SMC instruction is triggered.
>
> As I see it we also need a way to disable this Qemu PSCI emulation in case we
> want to boot an EL3 software (u-boot, optee or other).

Yes; we definitely want to support both of these use cases.

> We could certainly pass a command line option to explicitly tell Qemu to disable
> the PCSI emulation. But this would be a bit cumbersome as the board level
> code would then need to call to the SOC code to disable the psci setting on all cores.

We definitely don't want a new command line option. We just want to
use the information we already have (whether we're going to start the
guest code in secure/at EL3 or not) to make this decision.

> So I am wondering if we could be more "clever" solution and have things sorted out automatically. I would like to suggest 2 solutions;
>
> The first solution could be based on the mode of the processor when it starts executing the provided software. If I remember correctly what you explained, if an ELF file is provided then the processor boots in the highest level (EL3 for i.MX6/i.MX7) while if a non ELF file is provided (uImage ...) then the processor boots in EL2 (or lower depending on the processor) to emulate the uboot behavior.
>
> if the processor starts the software in EL3 then we need to disable the "psci-conduit" through SMC (if set) because the EL3 software should install/provide the necessary firmware.
> If the processor starts the software in EL2 then we need to disable the "psci-conduit" through HVC (if set) because the EL2 software should install/provide the necessary firmware.

This is the right approach. The idea I suggested in the previous email
was that rather than having the SoC object code setting psci-conduit
always, and then something else coming along later and changing it,
we should instead not set it in the first place. It's a simple QOM property
which means it's not writable after the CPU object has been realized, I think,
so we'd have to jump through some hoops to get it to be writable later.
On the other hand looking in hw/arm/boot.c we don't decide whether this
is a Linux boot (which means it will be to non-secure) or a non-Linux
boot (which means it will be to EL3) until quite late, so that's awkward
too (plus there's potentially duplication between board/SoC code of
the logic to decide whether to set psci-conduit or not). Let me have
a think about the cleanest way to arrange this.

-- PMM