Hi Andrew,

I agree DXE drivers could use a PCD to make it configurable and prevent collisions with UEFI defined TPL levels.

Bret’s suggestion of adding a DXE scoped services to create events using non-UEFI defined TPL levels could be used with these TPL levels from PCDs.  Would also allow DXE drivers to use TPL levels associated with the firmware interrupts in the range 17..30.  Perhaps extensions to the DXE Services Table?

Still does not address my concern that many DXE drivers using these extra TPL levels may run into race conditions if more than one DXE driver selects the same TPL level.  Platform integrators will need to understand the relative priorities of all DXE drivers that use extra TPL levels so they can assign values that both avoid collisions with future UEFI specs and prevent race conditions among DXE drivers.

Mike

From: Andrew Fish <afish@apple.com>
Sent: Tuesday, October 6, 2020 7:18 PM
To: edk2-devel-groups-io <devel@edk2.groups.io>; Kinney, Michael D <michael.d.kinney@intel.com>
Cc: bret.barkelew@microsoft.com
Subject: Re: [edk2-devel] VariablePolicy: Final Changes Thread 1 - TPL Ordering

Mike,

When I’ve done things like this in the past I think of making them configurable like via a PCD.

In terms of the #defines I think it makes more sense to just do math on the spec defined values. So TPL_CALLBACK + 1 or TPL_CALLBACK - 1 etc.  I’ve got an lldb type formatter for TPL and it prints out <UEFI Spec TPL> [+ <extra>] as I think this is the clearest way to do it.

Thanks,

Andrew Fish


On Oct 6, 2020, at 6:54 PM, Michael D Kinney <michael.d.kinney@intel.com<mailto:michael.d.kinney@intel.com>> wrote:

Bret,

It is likely best to go with the first approach.  The discussion on TPL levels can continue and you could adopt it in the future if a general solution is identified.

TPL 17..30 are reserved by the UEFI Spec for firmware interrupts.  So TPL_NOTIFY_HIGH as defined would not be allowed.

I agree that the use of TPL values other than those defined by the UEFI Spec can potentially be used by DXE.  However, that DXE usage must be flexible enough to handle a future extension to the UEFI Spec for new TPL levels without a collision.

Instead of defining specific TPL values, you could add a DXE scoped service to allocate the use of a new TPL level that is not being used by UEFI or other DXE drivers.  I will point out that these approaches (defining new TPL levels or allocating unused TPL levels) just moves the same problem.  You can solve it for the first driver that needs something special.  As soon as there is more than one driver that need something special at the same TPL level, the potential for a race condition for ordering will show up again.  So I do not consider adding TPL levels to be a good general solution to this problem.

Best regards,

Mike

From: devel@edk2.groups.io<mailto:devel@edk2.groups.io> <devel@edk2.groups.io<mailto:devel@edk2.groups.io>> On Behalf Of Bret Barkelew via groups.io<http://groups.io/>
Sent: Tuesday, October 6, 2020 5:24 PM
To: devel@edk2.groups.io<mailto:devel@edk2.groups.io>
Subject: [edk2-devel] VariablePolicy: Final Changes Thread 1 - TPL Ordering

As many will be aware, I’m in the final stages of having Variable Policy ready for commit. However, after moving the “Lock” event back to EndOfDxe, this exposed a race condition in variable services.

A quick synopsis of the problem:

  *   Previously, MorLock abused a privileged position by being tightly coupled to Variable Services, and its lock callback was called directly so that it could be strongly ordered with the internal property lock that disables future RequestToLock calls.
  *   VariablePolicy attempted to decouple this (without realizing it was a problem) and go through formalized interfaces that could ultimately be broken out of Variable Services altogether.
  *   However, doing so triggered the race condition, causing an ASSERT when MorLock attempted to lock its variables.
  *   I current have a reimplementation of the strong ordering workaround that can be previewed at the link below. I have tested that it works the same as the OLD system.

     *   Take a stab at solving the lock ordering problem · corthon/edk2@e7d0164 (github.com)<https://github.com/corthon/edk2/commit/e7d0164c8263b1fbfb8b4e289851fbedaa8997f1>

However, replacing one bad design with another is not what this community is about (when we can help it), so we’d like to take a moment to revisit a conversation that has come up before: expanding the number of supported TPL levels.

Now, I know that the current TPL levels are defined in the UEFI spec and we don’t want to have to change those, but there’s no reason that we can come up with not to add some more granularity in the PI spec, dedicated to platform and implementation ordering (the UEFI spec events will have to remain on UEFI spec TPLs). Basically there would be a set of DXE Services that allow WaitForEvent, CheckEvent, Event registration at TPLs other than notify/callback.  The UEFI system table versions of the functions would still have this restriction but code built with the platform could use the DXE Services. Right now, any attempt to use a non-UEFI TPL will ASSERT, so we can keep that in place on the SystemTable interface, but allow the platform to go around it with DXE Services. Similar functionality would have to be provided by the Mmst, but that’s already platform-specific and can probably allow it in all cases.

We’re suggesting something like the below:

//
// Task priority level
//
#define TPL_APPLICATION       4
#define TPL_CALLBACK_LOW      7
#define TPL_CALLBACK          8
#define TPL_CALLBACK_HIGH     9
#define TPL_NOTIFY_LOW        15
#define TPL_NOTIFY            16
#define TPL_NOTIFY_HIGH       17
#define TPL_HIGH_LEVEL        31

There’s already a long-in-the-tooth bug tracking a similar issue:
https://bugzilla.tianocore.org/show_bug.cgi?id=1676<https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fbugzilla.tianocore.org%2Fshow_bug.cgi%3Fid%3D1676&data=02%7C01%7CBret.Barkelew%40microsoft.com%7C9f3aa3e42bee4dd3a39208d8669a00e2%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C637372158932158036&sdata=vlC%2FB5f%2BDAMMp%2F3ECfBPukWiinHAmQj%2Flz1EPbSDKEA%3D&reserved=0>

This proposal is simpler than what’s in that bug, and would greatly simplify some of our event ordering (and code).

Thoughts?

If this conversation takes too long, I will publish a set of patches that just go with the lesser solution posted above, but I’d much rather work the problem this way.

- Bret