Ajay,
In general the order of dispatch when all the DEPEX evaluate to TRUE and the oder of events are not defined by the specification.
I think you are trying to say stuff that dispatches later is higher up the stack and it would make more sense to send the rest to things higher up the stack?
I’m not sure how this would work in your example as the Tcg2Dxe driver has a Depex of gEfiVariableArchProtocolGuid and the NVMe driver is a UEFI_DRIVER so it has an implied Depex [1] that is more complex?
To really fix we would probably need to add some kind of new protocol tot he NVMe driver that lets drivers request the reset get delayed. It would have a register, unregister, I’m Done.
The NVMe driver would do something like:
1) Register for the reset notification
2) Publish the new protocol
3) When the 1st driver registers the NVMe driver unregisters the reset notification, or I guess it could just ignore it.
When the shutdown happens and the last registered driver calls back in to say I’m Done then the NVMe driver can shutdown.
The other drivers will need to RegisterProtocolNotify so there is not dispatch sequence dependency.
I’m not sure that is the best solution, that is just the best I have right now.
[1] UEFI_DRIVER depends on all the protocols required to impelemnt all the UEFI Boot and Runtime Services.
Thanks,
Andrew FIsh
On Jun 28, 2023, at 5:23 PM, Ajay Iyengar (QUIC) <quic_aiyengar@quicinc.com> wrote:
Hello,
This is a comment and query on the Reset Notification Protocol defined by the UEFI specification.
For example, this is used for gracefully shutting off NVMe and TPM through their respective spec defined power off notifications:
If we assume a platform implementing fTPM using NVMe-RPMB, the boot order would require NVMe as a dependency for TPM, but that might result in an issue with the reset notification path. The default EDK2 implementation dispatches these notifications in boot order: https://github.com/tianocore/edk2/blob/master/MdeModulePkg/Universal/ResetSystemRuntimeDxe/ResetSystem.c#L78. Here NVMe would power off, before TPM commits to NVMe-RPMB within its reset notification callback. Reversing the order of dispatch (i.e. reverse-boot order) would help with this situation.
We have several similar situations on our platform -- drivers with a Depex ordering registering for ResetNotify which would benefit from being signaled in reverse-boot order. Are there any provisions that could be made with regards to ensuring the ordering requirement? Could we mimic the behavior of the ExitBootServices event dispatch for ResetNotify:
[The proposal would be to use InsertHeadList() instead of InsertTailList() in RegisterResetNotify()]
Thanks,
Ajay