Patrick Rudolph has uploaded this change for review.
arch/x86/gdt: Work around assembler bug
The GDT loading did work fine a few month ago, but today it only works in
QEMU, but not on real hardware or KVM enabled. This might be related to
toolchain changes.
Use 64bit GDT loading on x86_64 and force the assembler to generate a
64bit address load on the GDT. This will make sure no 32bit (signed)
displacement op is being generated, which points to the wrong address
in longmode.
Veryfied using readelf and made sure no R_X86_64_32S relocation symbol is
emitted. Disassembled the romstage ELF and made sure the GDT address is
64bit in size.
Tested on QEMU and KVM enabled QEMU: Doesn't crash any more on KVM.
Signed-off-by: Patrick Rudolph <siro@das-labor.org>
Change-Id: Ia824f90d9611e6e8db09bd62a05e6f990581f09a
---
M src/arch/x86/assembly_entry.S
M src/arch/x86/gdt_init.S
2 files changed, 24 insertions(+), 1 deletion(-)
git pull ssh://review.coreboot.org:29418/coreboot refs/changes/36/43136/1
diff --git a/src/arch/x86/assembly_entry.S b/src/arch/x86/assembly_entry.S
index fb48469..9ebc09d 100644
--- a/src/arch/x86/assembly_entry.S
+++ b/src/arch/x86/assembly_entry.S
@@ -15,12 +15,22 @@
#define _STACK_TOP _ecar_stack
#endif
+#ifdef __x86_64__
+.code64
+#else
+.code32
+#endif
+
.section ".text._start", "ax", @progbits
.global _start
_start:
/* Migrate GDT to this text segment */
+#if defined(__x86_64__)
+ call gdt_init64
+#else
call gdt_init
+#endif
/* reset stack pointer to CAR/EARLYRAM stack */
mov $_STACK_TOP, %esp
diff --git a/src/arch/x86/gdt_init.S b/src/arch/x86/gdt_init.S
index 7dd4b94..1558ac6 100644
--- a/src/arch/x86/gdt_init.S
+++ b/src/arch/x86/gdt_init.S
@@ -20,7 +20,20 @@
.section ".text._gdt64_", "ax", @progbits
.globl gdt_init64
gdt_init64:
- lgdt gdtptr64
+ /* Workaround a bug in the assembler.
+ * The following code doesn't work:
+ * lgdt gdtptr64
+ *
+ * The assembler tries to save memory by using 32bit displacement addressing mode.
+ * Displacements are using signed integers.
+ * This is fine in protected mode, as the negative address points to the correct
+ * address > 2GiB, but in long mode this doesn't work at all.
+ * Tests showed that QEMU can gracefully handle it, but real CPUs can't.
+ *
+ * Use the movabs pseudo instruction to force using a 64bit absolute address.
+ */
+ movabs $gdtptr64, %rax
+ lgdt (%rax)
ret
.previous
To view, visit change 43136. To unsubscribe, or for help writing mail filters, visit settings.