Hi Manish,
On Tue, 14 Nov 2023 at 16:47, Manish Pandey2 Manish.Pandey2@arm.com wrote:
Hi Brian,
Your usecase is that you want to use SDEI framework to deliver some EL3 interrupts to NS and if your only concern is about system reset the easiest/temp solution would be to disable the G1NS interrupts in OP-TEE dispatcher's system RESET hook.
The generic solution would take some time to implement as i need to analyse/discuss following scenarios before making a decision
- Does secure world really needs EL3 interrupts or it can handle NS-interrupts as FIQ (pre-EHF, do not modify routing model for secure world).
I don't think secure world should handle any EL3 interrupts, the EL3 firmware should directly trap and handle EL3 interrupts.
Notification to EL3 can be done using some standard or IMPDEF mechanism. FFA has an ABI for this FFA_EL3_INTR_HANDLE. Current Hafnium implementation follows this for G0 interrupts. 2. The PSCI calls for CPU power down/up, migrate, system reset/off do invoke SPD hooks. It need to be carefully analysed whether we really need to disable G1NS interrupts in these paths or not ? For e.g in CPU power down path the PSCI spec mandates that its caller's responsbility to migrate interrupts from the core (so we don't need to do anything special in TF-A code). Similarly for reset/off and migrate we need to analyse what should be the best way to avoid this situation.
As per SMCCC spec [1], PSCI calls are part of Standard Secure Service Calls which are essentially fast calls defined as below in the spec:
"Fast Calls execute atomic operations. The call appears to be atomic from the perspective of the calling PE, and returns when the requested operation has completed"
These are non-preemptible. So we should disable all interrupts (EL3 G0 or EL1 G1NS or EL1 G1S) while handling PSCI calls. I don't think it's the responsibility of the OP-TEE dispatcher but rather the generic EL3 firmware itself. However, if a particular platform is unmasking interrupts in OP-TEE while handling PSCI notifications then it's an issue which doesn't seem to be the case here.
[1] https://developer.arm.com/documentation/den0028/f/?lang=en
-Sumit
Also, i need to work on test infrastructure to simulate these scenarios. It will be useful if your setup can be recreated using OP-TEE/TF-A ?
Thanks Manish
From: Olivier Deprez Olivier.Deprez@arm.com Sent: 14 November 2023 10:47 To: Manish Pandey2 Manish.Pandey2@arm.com; tf-a@lists.trustedfirmware.org tf-a@lists.trustedfirmware.org; Ethridge, Caleb Caleb.Ethridge@analog.com; Neely, Brian Brian.Neely@analog.com Cc: op-tee@lists.trustedfirmware.org op-tee@lists.trustedfirmware.org Subject: Re: [TF-A] Re: Handling of normal world interrupts with BL31 PSCI handler
Hi,
Cc the OP-TEE mailing list, as I sense this might be more of an OP-TEE (/ OPTEED) design question than TF-A generic.
I agree with Manish/Achin the best way to eliminate the return by NS interrupt while OP-TEE kernel runs handling a PSCI request is to mask interrupts globally in the GIC e.g. from within opteed_system_reset?
Additional questions:
On https://github.com/ARM-software/arm-trusted-firmware/blob/master/lib/psci/ps... The PSCI layer calls into OP-TEE SPD and then resumes OP-TEE to handle the PSCI call in a platform defined manner.
As suggested here https://github.com/ARM-software/arm-trusted-firmware/blob/master/services/sp... , it looks this invocation is not supposed to return into TF-A?
Did you redefine the weak psci_system_reset function within OP-TEE for your platform? Is this function designed to end up in a loop and never returns to TF-A?
As a side question, is this an issue that the TF-A's PSCI platform hook is possibly never called https://github.com/ARM-software/arm-trusted-firmware/blob/master/lib/psci/ps... ?
Regards, Olivier.
From: Neely, Brian via TF-A tf-a@lists.trustedfirmware.org Sent: 13 November 2023 19:01 To: Manish Pandey2 Manish.Pandey2@arm.com; tf-a@lists.trustedfirmware.org tf-a@lists.trustedfirmware.org; Ethridge, Caleb Caleb.Ethridge@analog.com Subject: [TF-A] Re: Handling of normal world interrupts with BL31 PSCI handler
Hi Manish,
Just following up on your comment below. Are you planning to provide guidance as to which PSCI calls should have NS interrupts masked? Are these changes that should go in the mainline TF-A repo, or do you believe they are specific to our use case?
Thanks,
Brian
From: Manish Pandey2 via TF-A tf-a@lists.trustedfirmware.org Sent: Friday, November 10, 2023 11:31 AM To: tf-a@lists.trustedfirmware.org; Ethridge, Caleb Caleb.Ethridge@analog.com Subject: [TF-A] Re: Handling of normal world interrupts with BL31 PSCI handler
[External]
Hi Caleb,
The quick answer to your query is to mask the NS interrupts in BL31 (option a), You should not remove the callback to OPTEE as it may need to do its own state maintained before system RESET.
There are other scenarios in PSCI (e.g. CPU power down) path which need to consider disabling NS interrupts before invoking Secure world hooks, will provide further analysis on those later.
Thanks Manish Pandey
From: Caleb Ethridge via TF-A <tf-a@lists.trustedfirmware.orgmailto:tf-a@lists.trustedfirmware.org> Sent: 10 November 2023 15:12 To: tf-a@lists.trustedfirmware.orgmailto:tf-a@lists.trustedfirmware.org <tf-a@lists.trustedfirmware.orgmailto:tf-a@lists.trustedfirmware.org> Subject: [TF-A] Handling of normal world interrupts with BL31 PSCI handler
Hello,
If normal world interrupts are received while invoking TF-A's PSCI reset handler, we have observed that the reset can be aborted.
In TF-A's PSCI reset handler, a call out to OP-TEE is made before performing the platform-specific reset: https://github.com/ARM-software/arm-trusted-firmware/blob/master/lib/psci/ps...https://urldefense.com/v3/__https:/github.com/ARM-software/arm-trusted-firmware/blob/master/lib/psci/psci_system_off.c*L44__;Iw!!A3Ni8CS0y2Y!7wdssXjEhDud9Tm5QRek7Mw71eVJJfoimjH1u-u9fLXfaoLqv1ECsW3KakFsaMwnp3xZzyiBgwAyngBZMHyWqcCFuQ$ https://github.com/ARM-software/arm-trusted-firmware/blob/master/services/sp...https://urldefense.com/v3/__https:/github.com/ARM-software/arm-trusted-firmware/blob/master/services/spd/opteed/opteed_pm.c*L233__;Iw!!A3Ni8CS0y2Y!7wdssXjEhDud9Tm5QRek7Mw71eVJJfoimjH1u-u9fLXfaoLqv1ECsW3KakFsaMwnp3xZzyiBgwAyngBZMHypZWQwUQ$
When OP-TEE is entered, it is possible to receive foreign (normal world) interrupts, which invokes the procedure described here: https://optee.readthedocs.io/en/3.16.0/architecture/core.html#deliver-non-se...https://urldefense.com/v3/__https:/optee.readthedocs.io/en/3.16.0/architecture/core.html*deliver-non-secure-interrupts-to-normal-world__;Iw!!A3Ni8CS0y2Y!7wdssXjEhDud9Tm5QRek7Mw71eVJJfoimjH1u-u9fLXfaoLqv1ECsW3KakFsaMwnp3xZzyiBgwAyngBZMHzqKZ6_0Q$ When the SMC call is aborted as described above, this results in the reboot failing. Linux does not retry the PSCI reset (https://github.com/torvalds/linux/blob/master/drivers/firmware/psci/psci.c#L...https://urldefense.com/v3/__https:/github.com/torvalds/linux/blob/master/drivers/firmware/psci/psci.c*L308__;Iw!!A3Ni8CS0y2Y!7wdssXjEhDud9Tm5QRek7Mw71eVJJfoimjH1u-u9fLXfaoLqv1ECsW3KakFsaMwnp3xZzyiBgwAyngBZMHwH0cQx8g$). This makes sense because it is not expecting the SMC call to fail (it expected to make an uninterruptable SMC call into the secure monitor, not a call into OP-TEE).
If OP-TEE itself is setup to handle PSCI reset calls, it also handles them in (uninterruptable) SMC context: https://github.com/OP-TEE/optee_os/blob/master/core/arch/arm/sm/psci.c#L140https://urldefense.com/v3/__https:/github.com/OP-TEE/optee_os/blob/master/core/arch/arm/sm/psci.c*L140__;Iw!!A3Ni8CS0y2Y!7wdssXjEhDud9Tm5QRek7Mw71eVJJfoimjH1u-u9fLXfaoLqv1ECsW3KakFsaMwnp3xZzyiBgwAyngBZMHwHJcYAAQ$ https://github.com/OP-TEE/optee_os/blob/master/core/arch/arm/sm/sm_a32.S#L96https://urldefense.com/v3/__https:/github.com/OP-TEE/optee_os/blob/master/core/arch/arm/sm/sm_a32.S*L96__;Iw!!A3Ni8CS0y2Y!7wdssXjEhDud9Tm5QRek7Mw71eVJJfoimjH1u-u9fLXfaoLqv1ECsW3KakFsaMwnp3xZzyiBgwAyngBZMHxqZFTGrA$
Based on this, we see two possible solutions: a) Masking the non-secure interrupts in BL31 while we are doing a reset b) Removing the call to OPTEE in the reset handler so that we never leave the SMC context
Which option do you suggest? Or are we missing an important detail here?
Thanks,
Caleb
TF-A mailing list -- tf-a@lists.trustedfirmware.orgmailto:tf-a@lists.trustedfirmware.org To unsubscribe send an email to tf-a-leave@lists.trustedfirmware.orgmailto:tf-a-leave@lists.trustedfirmware.org