fix: correct GDT, segment loads, and ISR stack alignment

Co-authored-by: aider (openrouter/anthropic/claude-opus-4.7) <aider@aider.chat>

kernel_entry.S

@@ -80,8 +80,21 @@ _start:
 
 .code64
 long_mode_start:
+    /* Load data segment registers with the data descriptor (0x10).
+     * In long mode most of these are ignored for access checks, but
+     * %ss must be a valid (or null) selector so that iretq from an
+     * interrupt can safely reload it. */
+    movw $0x10, %ax
+    movw %ax, %ds
+    movw %ax, %es
+    movw %ax, %fs
+    movw %ax, %gs
+    movw %ax, %ss
+
     movabsq $stack_top, %rax
     movq %rax, %rsp
+    /* Ensure 16-byte alignment for the System V AMD64 ABI. */
+    andq $-16, %rsp
 
     call kernel_main
     cli
@@ -93,6 +106,7 @@ long_mode_start:
 .global isr_keyboard
 .align 16
 isr_keyboard:
+    /* Save all caller-saved registers used by the SysV AMD64 ABI. */
     pushq %rax
     pushq %rcx
     pushq %rdx
@@ -102,8 +116,14 @@ isr_keyboard:
     pushq %r9
     pushq %r10
     pushq %r11
+    /* On interrupt entry the CPU pushed 5 qwords (40 bytes), and we
+     * pushed 9 more (72 bytes) = 112 bytes total. That leaves %rsp
+     * misaligned by 8 relative to a 16-byte boundary after the
+     * upcoming `call`. Push one dummy qword to realign. */
+    subq $8, %rsp
     cld
     call keyboard_handler
+    addq $8, %rsp
     popq %r11
     popq %r10
     popq %r9
@@ -127,8 +147,10 @@ isr_keyboard:
 .section .rodata
 .align 8
 gdt64:
-    .quad 0x0000000000000000
+    .quad 0x0000000000000000         /* 0x00: null descriptor         */
-    .quad 0x00209A0000000000
+    .quad 0x00209A0000000000         /* 0x08: 64-bit code, DPL=0      */
+    .quad 0x0000920000000000         /* 0x10: data, writable, present */
+gdt64_end:
 gdt64_pointer:
-    .word gdt64_pointer - gdt64 - 1
+    .word gdt64_end - gdt64 - 1
     .quad gdt64