[PATCH] Remove all register pops before sysexit

[Luca Barbieri <ldb@ldb.ods.org>]

[-- Attachment #1: Type: text/plain, Size: 1551 bytes --]

This patch, which depends on the previous %ebx -> %ebp patch, removes
all pop instruction in the sysenter return path.

This leaks the thread_info address to user mode but this shouldn't be
a security problem.

This is what happens to the various registers:

%eax: return value from system call: already in place
%ebx, %esi, %edi: saved by the C compiler
%ecx, %edx, %ebp: restored by user mode
%esp, eip: copied to %ecx/%edx and restored by sysexit
%ds, %es: initialized to __USER_DS on kernel entry
%cs, %ss: restored by sysexit based on msr
%fs, %gs: not modified by the kernel (saved around context switch)
eflags: not preserved
FP, XMM: any code that modifies them must save/restore them

Note that while it is possible to change %ebx, %esi, %edi, %ecx, %edx
or %ebp via struct pt_regs, anything that does should set TIF_IRET or
another work flag (and it hopefully already does).


diff --exclude-from=/home/ldb/src/exclude -urNdp linux-2.5.54-preldb/arch/i386/kernel/entry.S linux-2.5.54-ldb/arch/i386/kernel/entry.S
--- linux-2.5.54-preldb/arch/i386/kernel/entry.S	2003-01-06 16:01:40.000000000 +0100
+++ linux-2.5.54-ldb/arch/i386/kernel/entry.S	2003-01-06 04:54:58.000000000 +0100
@@ -276,9 +276,9 @@ ENTRY(sysenter_entry)
 	movl TI_FLAGS(%ebp), %ecx
 	testw $_TIF_ALLWORK_MASK, %cx
 	jne syscall_exit_work
-	RESTORE_INT_REGS
-	movl 12(%esp),%edx
-	movl 24(%esp),%ecx
+/* if something modifies registers it must also disable sysexit */
+	movl EIP(%esp), %edx
+	movl OLDESP(%esp), %ecx
 	sti
 	sysexit
 

[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]