[PATCH] x86_64 signal handling for 64-bit apps w/ mixed 32-bit code - trivial fix

[Bryan Ford <baford@mit.edu>]

[-- Attachment #1: Type: text/plain, Size: 1164 bytes --]

The following trivial patch fixes a bug in signal handling on x86-64: the 
kernel currently fails to save and restore the CS and SS segment registers on 
user-mode signal handler dispatch/return, which makes it impossible for 
64-bit applications to catch and handle signals properly that occur while 
running 32-bit code fragments in compatibility mode.

The proposed patch doesn't affect any performance-critical paths (e.g., 
syscall or interrupt entry/exit), and merely involves a couple more moves 
to/from user space on signal frame setup and sigreturn.  It also doesn't 
affect the size or shape of the sigcontext at all, since there already was an
(unused) slot for CS, and I've assigned the convenient __pad0 field as a slot 
for SS.  The existing, unused slots for FS and GS remain unused for now, and 
I don't see any urgent need to change that.  The only way this might break an 
existing app is if the app tries to cons up its own signal frame (not 
generated by the kernel) and pass it to sigreturn, but this is presumably a 
no-no anyway.

The patch is against linux-2.6.13.3.
Author: Bryan Ford, baford@mit.edu
No copyright claimed; public domain.

[-- Attachment #2: bar --]
[-- Type: text/x-diff, Size: 1866 bytes --]

diff -ur o13/include/asm-x86_64/sigcontext.h linux-2.6.13.3/include/asm-x86_64/sigcontext.h
--- o13/include/asm-x86_64/sigcontext.h	2005-10-03 19:27:35.000000000 -0400
+++ linux-2.6.13.3/include/asm-x86_64/sigcontext.h	2005-10-05 12:06:59.000000000 -0400
@@ -43,7 +43,7 @@
 	unsigned short cs;
 	unsigned short gs;
 	unsigned short fs;
-	unsigned short __pad0; 
+	unsigned short ss; 
 	unsigned long err;
 	unsigned long trapno;
 	unsigned long oldmask;
diff -ur o13/arch/x86_64/kernel/signal.c linux-2.6.13.3/arch/x86_64/kernel/signal.c
--- o13/arch/x86_64/kernel/signal.c	2005-10-03 19:27:35.000000000 -0400
+++ linux-2.6.13.3/arch/x86_64/kernel/signal.c	2005-10-05 12:13:22.000000000 -0400
@@ -110,6 +110,15 @@
 	COPY(r14);
 	COPY(r15);
 
+	/* Kernel saves and restores only CS and DS segments on signals,
+	 * which are the bare essentials needed to allow mixed 32/64-bit code.
+	 * App's signal handler can save/restore other segments if needed. */
+	unsigned short cs, ss;
+	err |= __get_user(cs, &sc->cs);
+	err |= __get_user(ss, &sc->ss);
+	regs->cs = cs | 3;	/* Force into user mode */
+	regs->ss = ss | 3;
+
 	{
 		unsigned int tmpflags;
 		err |= __get_user(tmpflags, &sc->eflags);
@@ -187,6 +196,8 @@
 {
 	int err = 0;
 
+	err |= __put_user(regs->cs, &sc->cs);
+	err |= __put_user(regs->ss, &sc->ss);
 	err |= __put_user(0, &sc->gs);
 	err |= __put_user(0, &sc->fs);
 
@@ -318,7 +329,15 @@
 
 	regs->rsp = (unsigned long)frame;
 
+	/* Set up segment registers to run signal handlers in 64-bit mode,
+	   even if the handler happens to be interrupting 32-bit code. */
+	regs->cs = __USER_CS;
+	regs->ss = __USER_DS;
+
+	/* This, by contrast, has nothing to do with segment registers -
+	   see include/asm-x86_64/uaccess.h for details. */
 	set_fs(USER_DS);
+
 	regs->eflags &= ~TF_MASK;
 	if (test_thread_flag(TIF_SINGLESTEP))
 		ptrace_notify(SIGTRAP);