[edk2-devel] [PATCH] UefiCpuPkg/MpInitLib: Add check for CR3/GDT/IDT.

[Laszlo Ersek]

On 09/03/20 03:47, Dong, Eric wrote:
> Hi Laszlo,
>
> Thanks for your detail review and good comments, add my reply in
> below.
>
> This issue was reported by tester. They use a shell application to do
> the test. In the test, it first moves the page table to above 4G range
> then calls StartUpThisAp to let AP run the test procedure.  The system
> will hang during the test.
>
> Add more comments in below inline.

Thanks. Let me summarize my understanding:

(1) During normal boot, and normal MP services PPI usage, and normal MP
    services protocol usage, there is no issue.

(2) Right now, we have not identified the exact edk2 core modules and
    locations that are responsible for allocating the IDT / GDT / page
    tables. Nonetheless, the allocations are all placed in the 32-bit
    address space, and so there is no problem with AP startup, during
    normal usage (see point (1)).

(3) There is a UEFI shell application that isn't more closely identified
    in this discussion. It first moves at least one of the IDT / GDT /
    page tables above 4GB, and then calls some member functions of the
    MP services protocol. This causes a hang, experienced during the
    execution of the UEFI shell application.

(4) The present patch recognizes the issue in FillExchangeInfoData(),
    and returns an error. In the relevant use case (UEFI shell
    application), this causes the called MP services protocol member
    function to fail, synchronously. This means the application still
    doesn't work, but there is no hang at least.

Is my understanding correct?

If it is, then I have the following comments:

- Please file a TianoCore BZ for this issue, and capture the use case in
  it (= UEFI shell application that changes the IDT / GDT / page tables
  base). Please feel free to copy parts of the above description, if you
  think that's useful.

- I'm quite doubtful that the use case is valid, in the first place. A
  UEFI application is supposed to consume the services described in the
  UEFI specification. Messing with the page tables is something that a
  UEFI application should *not* do, in my opinion. Such page table
  manipulation is expected to interfere with various DXE drivers in the
  platform firmware.

- Another comment on the UEFI shell application, and the present patch,
  is that their *combination* will still break ExitBootServices().
  Assume that the patch is applied, and the UEFI shell application is
  invoked. The application now fails gracefully, and exits. Then we
  attempt to boot an OS (this is a valid thing to do). Because the
  application moved the IDT / GDT / page tables "out of range",
  MpInitChangeApLoopCallback() will do the wrong thing.

- Most importantly: in MpInitLib, there is a *large* amount of call
  sites, and a large number of call paths that lead to WakeUpAP(), and
  ultimately to FillExchangeInfoData(). This means that changing the
  return type of FillExchangeInfoData() from VOID to EFI_STATUS has a
  "ripple effect" -- many call paths would have to deal with error
  checking, error propagation, and resource release (!!!) along those
  error paths.

  - For example, your current patch leaks resources in WakeUpAP(), when
    FillExchangeInfoData() fails -- see AllocateResetVector() and
    AllocateSevEsAPMemory().

  - For another example, your current patch does not handle several
    WakeUpAP() call sites, such as the ones in
    ResetProcessorToIdleState() and CheckAllAPs().

  Whereas in reality, from the applications point of view, we only need
  the MP services protocol member functions to fail cleanly. Therefore
  we should address this problem *early*; that is, *much less deep* in
  the call stack.


I suggest the following:

- introduce a feature PCD (default value FALSE)

- modify the following functions in MpInitLib:
  - MpInitLibGetNumberOfProcessors
  - MpInitLibGetProcessorInfo
  - MpInitLibStartupAllAPs
  - MpInitLibStartupThisAP
  - MpInitLibSwitchBSP
  - MpInitLibEnableDisableAP
  - MpInitLibWhoAmI

- each modified function should check the feature PCD *very early*. If
  the PCD is true, then the function should pre-emptively verify the IDT
  / GDT / root page table  location. If any one of those objects is
  outside of the 32-bit address space, then the function should fail
  immediately.

- We need to consider the event handlers in DxeMpInitLib:

  - CheckApsStatus() is not a problem, because the above short-circuits
    in the public MpInitLib functions will guarantee that
    "mStopCheckAllApsStatus" is always TRUE.

  - For exit-boot-services and legacy-boot, we need to modify
    MpInitChangeApLoopCallback() however. MpInitChangeApLoopCallback()
    should do the same as the above-listed functions (check the PCD,
    check the IDT / GDT / root page table), and if one of them is above
    4GB, then MpInitChangeApLoopCallback() should hang *hard* -- call
    CpuDeadLoop() --, even in RELEASE builds.

    This is because the UEFI application in question makes OS boot
    impossible, so we should stop the system as early as we realize
    that.

- Of course, I suggest a helper function to encapsulate the PCD check,
  and the GDT / IDT / CR3 checks.

This will let most platforms ignore the special UEFI shell application
use case (which I do feel is invalid). In addition, platforms that do
want to handle this use case, can set the PCD to TRUE for the CpuDxe
module *only* (so that CpuMpPei is not penalized). Finally, this
approach keeps the code clean; long call paths to FillExchangeInfoData()
are not made more complex.

Thanks
Laszlo


-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.

View/Reply Online (#64997): https://edk2.groups.io/g/devel/message/64997
Mute This Topic: https://groups.io/mt/76573401/1813853
Group Owner: devel+owner at edk2.groups.io
Unsubscribe: https://edk2.groups.io/g/devel/unsub  [edk2-devel-archive at redhat.com]
-=-=-=-=-=-=-=-=-=-=-=-