BUG: OvmfPkgX64 is broken with cee5b0441af3 UefiCpuPkg/CpuDxe: Fix boot error

BUG: OvmfPkgX64 is broken with cee5b0441af3 UefiCpuPkg/CpuDxe: Fix boot error

[James Bottomley]

I found this trying to test out the upstream SEV secret injection on an
AMD rome system.  However, I rebuilt the OvmfPkgX64 (still on a rome
system) just to check.  I get a boot loop here if I leave this commit
unreverted:

InstallProtocolInterface: BC62157E-3E33-4FEC-9920-2D3B36D750DF 7F1AD918
ProtectUefiImageCommon - 0x7F1AD2C0
  - 0x000000007F068000 - 0x000000000000FCC0
Paging: added 512 pages to page table pool
CurrentPagingContext:
  MachineType   - 0x8664
  PageTableBase - 0x7F801000
  Attributes    - 0xC0000006
SecCoreStartupWithStack(0xFFFCC000, 0x820000)

If I revert this commit, it will boot up normally.

The memory specification for qemu-kvm I'm using is

-m 4096M,slots=5,maxmem=30G

but reducing it to under 4GB doesn't help

James

Re: [edk2-devel] BUG: OvmfPkgX64 is broken with cee5b0441af3 UefiCpuPkg/CpuDxe: Fix boot error

[James Bottomley]

On Mon, 2020-12-14 at 15:51 -0800, James Bottomley wrote:
> I found this trying to test out the upstream SEV secret injection on
> an
> AMD rome system.  However, I rebuilt the OvmfPkgX64 (still on a rome
> system) just to check.  I get a boot loop here if I leave this commit
> unreverted:
> 
> InstallProtocolInterface: BC62157E-3E33-4FEC-9920-2D3B36D750DF
> 7F1AD918
> ProtectUefiImageCommon - 0x7F1AD2C0
>   - 0x000000007F068000 - 0x000000000000FCC0
> Paging: added 512 pages to page table pool
> CurrentPagingContext:
>   MachineType   - 0x8664
>   PageTableBase - 0x7F801000
>   Attributes    - 0xC0000006
> SecCoreStartupWithStack(0xFFFCC000, 0x820000)
> 
> If I revert this commit, it will boot up normally.
> 
> The memory specification for qemu-kvm I'm using is
> 
> -m 4096M,slots=5,maxmem=30G
> 
> but reducing it to under 4GB doesn't help

I think I found the source of the problem: nasm is generating this
assembly sequence (disassembled by objdump):

 14:	48 ff 2c 24          	rex.W ljmp *(%rsp)

However, on AMD the rex.W prefix to a ljmp can be ineffective and the
recommended way to do this is with lretq (or retfq in nasm speak).  If
I apply the patch below, I can get the long word version to work for me
on SEV.

James

---

diff --git a/UefiCpuPkg/CpuDxe/X64/CpuAsm.nasm b/UefiCpuPkg/CpuDxe/X64/CpuAsm.nasm
index 6ad32b49f4ef..a2490073cef6 100644
--- a/UefiCpuPkg/CpuDxe/X64/CpuAsm.nasm
+++ b/UefiCpuPkg/CpuDxe/X64/CpuAsm.nasm
@@ -20,13 +20,11 @@
 ;------------------------------------------------------------------------------
 global ASM_PFX(SetCodeSelector)
 ASM_PFX(SetCodeSelector):
-    sub     rsp, 0x10
     lea     rax, [setCodeSelectorLongJump]
-    mov     [rsp], rax
-    mov     [rsp+8], cx
-    jmp     qword far [rsp]
+    push     rcx
+    push     rax
+    retfq
 setCodeSelectorLongJump:
-    add     rsp, 0x10
     ret
 
 ;------------------------------------------------------------------------------