Skip to contentSkip to conversation

bsd-user/x86_64/target_arch_thread.h: Align stack

PATCH In review Ilya Leoshkevich 3 participants 4 messages Jul 31, 2024

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bsd-user/x86_64/target_arch_thread.h: Align stack

bsd-user qemu-x86_64 almost immediately dies with:

    qemu: 0x4002201a68: unhandled CPU exception 0xd - aborting

on FreeBSD 14.1-RELEASE. This is an instruction that requires alignment:

    (gdb) x/i 0x4002201a68
       0x4002201a68:        movaps %xmm0,-0x40(%rbp)

and the argument is not aligned:

    (gdb) p/x env->regs[5]
    $1 = 0x822443b58

A quick experiment shows that the userspace entry point expects misaligned rsp:

    (gdb) starti
    (gdb) p/x $rsp
    $1 = 0x7fffffffeaa8

Emulate this behavior in bsd-user.

---
 bsd-user/x86_64/target_arch_thread.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
bsd-user/x86_64/target_arch_thread.h +1 -1
index 52c28906d6d..25233443c14 100644
@@ -31,7 +31,7 @@ static inline void target_thread_init(struct target_pt_regs *regs,
     struct image_info *infop)
 {
     regs->rax = 0;
-    regs->rsp = infop->start_stack;
+    regs->rsp = (infop->start_stack & ~0xfUL) - 8;
     regs->rip = infop->entry;
     regs->rdi = infop->start_stack;
 }
Signature · 1 line
2.45.2
Reviewed-by on v1 1/1
#1 lore raw
Quoted 34 lines · Ilya Leoshkevich

bsd-user qemu-x86_64 almost immediately dies with:

     qemu: 0x4002201a68: unhandled CPU exception 0xd - aborting

on FreeBSD 14.1-RELEASE. This is an instruction that requires alignment:

     (gdb) x/i 0x4002201a68
        0x4002201a68:        movaps %xmm0,-0x40(%rbp)

and the argument is not aligned:

     (gdb) p/x env->regs[5]
     $1 = 0x822443b58

A quick experiment shows that the userspace entry point expects misaligned rsp:

     (gdb) starti
     (gdb) p/x $rsp
     $1 = 0x7fffffffeaa8

Emulate this behavior in bsd-user.

---
  bsd-user/x86_64/target_arch_thread.h | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
bsd-user/x86_64/target_arch_thread.h +0 -0
index 52c28906d6d..25233443c14 100644

bsd-user/x86_64/target_arch_thread.h static inline void target_thread_init(struct target_pt_regs *regs, · line 34

Quoted 5 lines
bsd-user/x86_64/target_arch_thread.h +1 -1
@@ -31,7 +31,7 @@ static inline void target_thread_init(struct target_pt_regs *regs,
      struct image_info *infop)
  {
      regs->rax = 0;
-    regs->rsp = infop->start_stack;
+    regs->rsp = (infop->start_stack & ~0xfUL) - 8;

The formula in sys/amd64/amd64/exec_machdep.c, exec_setregs is

((stack - 8) & ~0xful) + 8

With that,

r~

Reviewed-by
#2 lore raw

Hmmm... All platforms likely need this....
Oh, they all have it except x86, both 32 and 64-bit.... but i386 is already properly aligned (at least in the
FreeBSD kernel), so maybe that's not needed.

Quoted 43 lines · Richard Henderson
Quoted 22 lines · Ilya Leoshkevich

bsd-user qemu-x86_64 almost immediately dies with:

     qemu: 0x4002201a68: unhandled CPU exception 0xd - aborting

on FreeBSD 14.1-RELEASE. This is an instruction that requires alignment:

     (gdb) x/i 0x4002201a68
        0x4002201a68:        movaps %xmm0,-0x40(%rbp)

and the argument is not aligned:

     (gdb) p/x env->regs[5]
     $1 = 0x822443b58

A quick experiment shows that the userspace entry point expects misaligned rsp:

     (gdb) starti
     (gdb) p/x $rsp
     $1 = 0x7fffffffeaa8

Review

Quoted 8 lines

Emulate this behavior in bsd-user.

---
  bsd-user/x86_64/target_arch_thread.h | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
(quoted diff) +0 -0

b/bsd-user/x86_64/target_arch_thread.h

bsd-user/x86_64/target_arch_thread.h static inline void target_thread_init(struct

Quoted 4 lines

index 52c28906d6d..25233443c14 100644

bsd-user/x86_64/target_arch_thread.h +0 -0
@@ -31,7 +31,7 @@ static inline void target_thread_init(struct

target_pt_regs *regs,

bsd-user/x86_64/target_arch_thread.h static inline void target_thread_init(struct target_pt_regs *regs, · line 34

Quoted 5 lines
bsd-user/x86_64/target_arch_thread.h +1 -1
      struct image_info *infop)
  {
      regs->rax = 0;
-    regs->rsp = infop->start_stack;
+    regs->rsp = (infop->start_stack & ~0xfUL) - 8;
Quoted 7 lines

The formula in sys/amd64/amd64/exec_machdep.c, exec_setregs is

   ((stack - 8) & ~0xful) + 8

With that,

Agreed.

There's a *lot* that's missing on for amd64 emulation relative to the other platforms.

Warner

#3 lore raw
Quoted 60 lines · Warner Losh

Hmmm... All platforms likely need this....
Oh, they all have it except x86, both 32 and 64-bit.... but i386 is already properly aligned (at least in the
FreeBSD kernel), so maybe that's not needed.

Quoted 43 lines · Richard Henderson
Quoted 22 lines · Ilya Leoshkevich

bsd-user qemu-x86_64 almost immediately dies with:

     qemu: 0x4002201a68: unhandled CPU exception 0xd - aborting

on FreeBSD 14.1-RELEASE. This is an instruction that requires alignment:

     (gdb) x/i 0x4002201a68
        0x4002201a68:        movaps %xmm0,-0x40(%rbp)

and the argument is not aligned:

     (gdb) p/x env->regs[5]
     $1 = 0x822443b58

A quick experiment shows that the userspace entry point expects misaligned rsp:

     (gdb) starti
     (gdb) p/x $rsp
     $1 = 0x7fffffffeaa8

Review

Quoted 8 lines

Emulate this behavior in bsd-user.

---
  bsd-user/x86_64/target_arch_thread.h | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
(quoted diff) +0 -0

b/bsd-user/x86_64/target_arch_thread.h

bsd-user/x86_64/target_arch_thread.h static inline void target_thread_init(struct

Quoted 4 lines

index 52c28906d6d..25233443c14 100644

bsd-user/x86_64/target_arch_thread.h +0 -0
@@ -31,7 +31,7 @@ static inline void target_thread_init(struct

target_pt_regs *regs,

bsd-user/x86_64/target_arch_thread.h static inline void target_thread_init(struct target_pt_regs *regs, · line 34

Quoted 5 lines
bsd-user/x86_64/target_arch_thread.h +1 -1
      struct image_info *infop)
  {
      regs->rax = 0;
-    regs->rsp = infop->start_stack;
+    regs->rsp = (infop->start_stack & ~0xfUL) - 8;
Quoted 7 lines

The formula in sys/amd64/amd64/exec_machdep.c, exec_setregs is

   ((stack - 8) & ~0xful) + 8

With that,

Quoted 7 lines

Agreed.

There's a *lot* that's missing on for amd64 emulation relative to the other platforms.

Forgot to mention that I've queued this change with Richard's suggested modification.

Warner