x86 - 32-bit ptrace emulation mishandles 6th arg

From: Chuck Ebbert <[email protected]>

[ jdike - Pushing Chuck's patch - see
http://lkml.org/lkml/2005/9/16/261 for some history and a test program. UML is also broken without this patch - its processes get SIGBUS from the corrupt 6th argument to mmap being interpretted as a file offset ]

When the 32-bit vDSO is used to make a system call, the %ebp register for the 6th syscall arg has to be loaded from the user stack (where it's pushed by the vDSO user code). The native i386 kernel always does this before stopping for syscall tracing, so %ebp can be seen and modified via ptrace to access the 6th syscall argument. The x86-64 kernel fails to do this, presenting the stack address to ptrace instead. This makes the %rbp value seen by 64-bit ptrace of a 32-bit process, and the %ebp value seen by a 32-bit caller of ptrace, both differ from the native i386 behavior.

This patch fixes the problem by putting the word loaded from the user stack into %rbp before calling syscall_trace_enter, and reloading the 6th syscall argument from there afterwards (so ptrace can change it). This makes the behavior match that of i386 kernels.

---
 arch/x86_64/ia32/ia32entry.S |   19 ++++++-------------
 1 file changed, 6 insertions(+), 13 deletions(-)

index 18b2318..df588f0 100644
@@ -159,20 +159,16 @@ sysenter_do_call:
 
 sysenter_tracesys:
 	CFI_RESTORE_STATE
+	xchgl	%r9d,%ebp
 	SAVE_REST
 	CLEAR_RREGS
+	movq	%r9,R9(%rsp)
 	movq	$-ENOSYS,RAX(%rsp)	/* really needed? */
 	movq	%rsp,%rdi        /* &pt_regs -> arg1 */
 	call	syscall_trace_enter
 	LOAD_ARGS32 ARGOFFSET  /* reload args from stack in case ptrace changed it */
 	RESTORE_REST
-	movl	%ebp, %ebp
-	/* no need to do an access_ok check here because rbp has been
-	   32bit zero extended */ 
-1:	movl	(%rbp),%r9d
-	.section __ex_table,"a"
-	.quad 1b,ia32_badarg
-	.previous
+	xchgl	%ebp,%r9d
 	jmp	sysenter_do_call
 	CFI_ENDPROC
 ENDPROC(ia32_sysenter_target)

@@ -262,20 +258,17 @@ cstar_do_call:
 	
 cstar_tracesys:	
 	CFI_RESTORE_STATE
+	xchgl %r9d,%ebp
 	SAVE_REST
 	CLEAR_RREGS
+	movq %r9,R9(%rsp)
 	movq $-ENOSYS,RAX(%rsp)	/* really needed? */
 	movq %rsp,%rdi        /* &pt_regs -> arg1 */
 	call syscall_trace_enter
 	LOAD_ARGS32 ARGOFFSET  /* reload args from stack in case ptrace changed it */
 	RESTORE_REST
+	xchgl %ebp,%r9d
 	movl RSP-ARGOFFSET(%rsp), %r8d
-	/* no need to do an access_ok check here because r8 has been
-	   32bit zero extended */ 
-1:	movl	(%r8),%r9d
-	.section __ex_table,"a"
-	.quad 1b,ia32_badarg
-	.previous
 	jmp cstar_do_call
 END(ia32_cstar_target)
 				

Sure has my ACK.
I never really understood why my old patch was not taken 2.5 years ago.

Thanks,
Roland

Nor I. It's needed.

As is your PTRACE_SET_THREAD_INFO patch from yesterday - with these two fixes, I can boot a 32-bit UML on a 64-bit host.

Jeff

Work email - jdike at linux dot intel dot com

FYI, http://sourceware.org/systemtap/wiki/utrace/tests has details on the ptrace-tests suite we're collecting. A test I added there is how I noticed the PTRACE_GET_THREAD_AREA regression. A regression test for the ebp bug should be easy to add too.

Thanks,
Roland

I forget the details, but I had to make some kind of trivial change to make it work in some corner cases.

If I remember correctly the old patch back then broke some programs, so it was reverted.

It's certainly ugly that it leaks r9 internals to ptrace. Also it generates redundant code with CLEAR_RREGS.

-Andi