Reintroduce a fixed version of 3d70088. Fixes #87

Author: mintsuki

src/Thunk16.asm32

@@ -237,10 +237,10 @@ BITS    32
     sidt    [esp + 36]        ; save IDT stack in argument space
     mov     eax, cr0
     mov     [edx - 4], eax                  ; save CR0 in _BackFromUserCode.SavedCr0End - 4
-    and     eax, 7ffffffeh              ; clear PE, PG bits
+    mov     eax, 00000010h              ; clear all bits except ET
     mov     ebp, cr4
     mov     [edx + (_BackFromUserCode.SavedCr4End - 4 - _BackFromUserCode.SavedCr0End)], ebp
-    and     ebp, ~30h                ; clear PAE, PSE bits
+    xor     ebp, ebp                    ; zero out CR4
     push    10h
     pop     ecx                         ; ecx <- selector for data segments
     lgdt    [edx + (_16Gdtr - _BackFromUserCode.SavedCr0End)]

src/Thunk16.asm64

@@ -125,8 +125,10 @@ BITS    16
     mov     cr4, eax
 o32 lgdt [cs:bx + (SavedGdt - .Base)]
     mov     ecx, 0c0000080h
-    rdmsr
-    or      ah, 1
+    mov     eax, strict dword 0
+.SavedEFERLowEnd:
+    mov     edx, strict dword 0
+.SavedEFERHighEnd:
     wrmsr
     mov     eax, strict dword 0
 .SavedCr0End:
@@ -164,8 +166,8 @@ BITS    16
     mov     gs, dx
     mov     ecx, 0c0000080h
     mov     cr0, eax                    ; real mode starts at next instruction
-    rdmsr
-    and     ah, ~1
+    xor     eax, eax
+    xor     edx, edx
     wrmsr
     mov     cr4, ebp
     mov     ss, si                      ; set up 16-bit stack segment
@@ -283,10 +285,23 @@ BITS    64
     sidt    [rsp + 50h]       ; save IDT stack in argument space
     mov     rax, cr0
     mov     [rcx + (_BackFromUserCode.SavedCr0End - 4 - _BackFromUserCode.SavedCr4End)], eax
-    and     eax, 7ffffffeh              ; clear PE, PG bits
+    mov     eax, 00000010h              ; clear all bits except ET
+    push    rax
+    push    rcx
+    push    rdx
+    mov     ecx, 0c0000080h
+    rdmsr
+    btr     eax, 10                     ; Reset LMA flag to avoid crash on AMD Zen CPUs
+    mov     r10d, eax
+    mov     r11d, edx
+    pop     rdx
+    pop     rcx
+    pop     rax
+    mov     [rcx + (_BackFromUserCode.SavedEFERLowEnd - 4 - _BackFromUserCode.SavedCr4End)], r10d
+    mov     [rcx + (_BackFromUserCode.SavedEFERHighEnd - 4 - _BackFromUserCode.SavedCr4End)], r11d
     mov     rbp, cr4
     mov     [rcx - 4], ebp              ; save CR4 in _BackFromUserCode.SavedCr4End - 4
-    and     ebp, ~30h                ; clear PAE, PSE bits
+    xor     ebp, ebp                    ; zero out CR4
     mov     esi, r8d                    ; esi <- 16-bit stack segment
     push    DATA32
     pop     rdx                         ; rdx <- 32-bit data segment selector