x86/stackframe/32: Provide consistent pt_regs

Currently pt_regs on x86_32 has an oddity in that kernel regs
(!user_mode(regs)) are short two entries (esp/ss). This means that any
code trying to use them (typically: regs->sp) needs to jump through
some unfortunate hoops.

Change the entry code to fix this up and create a full pt_regs frame.

This then simplifies various trampolines in ftrace and kprobes, the
stack unwinder, ptrace, kdump and kgdb.

Much thanks to Josh for help with the cleanups!

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Josh Poimboeuf <jpoimboe@redhat.com>
Acked-by: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>

arch/x86/entry/entry_32.S

@@ -202,9 +202,102 @@
 .Lend_\@:
 .endm
 
+#define CS_FROM_ENTRY_STACK	(1 << 31)
+#define CS_FROM_USER_CR3	(1 << 30)
+#define CS_FROM_KERNEL		(1 << 29)
+
+.macro FIXUP_FRAME
+	/*
+	 * The high bits of the CS dword (__csh) are used for CS_FROM_*.
+	 * Clear them in case hardware didn't do this for us.
+	 */
+	andl	$0x0000ffff, 3*4(%esp)
+
+#ifdef CONFIG_VM86
+	testl	$X86_EFLAGS_VM, 4*4(%esp)
+	jnz	.Lfrom_usermode_no_fixup_\@
+#endif
+	testl	$SEGMENT_RPL_MASK, 3*4(%esp)
+	jnz	.Lfrom_usermode_no_fixup_\@
+
+	orl	$CS_FROM_KERNEL, 3*4(%esp)
+
+	/*
+	 * When we're here from kernel mode; the (exception) stack looks like:
+	 *
+	 *  5*4(%esp) - <previous context>
+	 *  4*4(%esp) - flags
+	 *  3*4(%esp) - cs
+	 *  2*4(%esp) - ip
+	 *  1*4(%esp) - orig_eax
+	 *  0*4(%esp) - gs / function
+	 *
+	 * Lets build a 5 entry IRET frame after that, such that struct pt_regs
+	 * is complete and in particular regs->sp is correct. This gives us
+	 * the original 5 enties as gap:
+	 *
+	 * 12*4(%esp) - <previous context>
+	 * 11*4(%esp) - gap / flags
+	 * 10*4(%esp) - gap / cs
+	 *  9*4(%esp) - gap / ip
+	 *  8*4(%esp) - gap / orig_eax
+	 *  7*4(%esp) - gap / gs / function
+	 *  6*4(%esp) - ss
+	 *  5*4(%esp) - sp
+	 *  4*4(%esp) - flags
+	 *  3*4(%esp) - cs
+	 *  2*4(%esp) - ip
+	 *  1*4(%esp) - orig_eax
+	 *  0*4(%esp) - gs / function
+	 */
+
+	pushl	%ss		# ss
+	pushl	%esp		# sp (points at ss)
+	addl	$6*4, (%esp)	# point sp back at the previous context
+	pushl	6*4(%esp)	# flags
+	pushl	6*4(%esp)	# cs
+	pushl	6*4(%esp)	# ip
+	pushl	6*4(%esp)	# orig_eax
+	pushl	6*4(%esp)	# gs / function
+.Lfrom_usermode_no_fixup_\@:
+.endm
+
+.macro IRET_FRAME
+	testl $CS_FROM_KERNEL, 1*4(%esp)
+	jz .Lfinished_frame_\@
+
+	/*
+	 * Reconstruct the 3 entry IRET frame right after the (modified)
+	 * regs->sp without lowering %esp in between, such that an NMI in the
+	 * middle doesn't scribble our stack.
+	 */
+	pushl	%eax
+	pushl	%ecx
+	movl	5*4(%esp), %eax		# (modified) regs->sp
+
+	movl	4*4(%esp), %ecx		# flags
+	movl	%ecx, -4(%eax)
+
+	movl	3*4(%esp), %ecx		# cs
+	andl	$0x0000ffff, %ecx
+	movl	%ecx, -8(%eax)
+
+	movl	2*4(%esp), %ecx		# ip
+	movl	%ecx, -12(%eax)
+
+	movl	1*4(%esp), %ecx		# eax
+	movl	%ecx, -16(%eax)
+
+	popl	%ecx
+	lea	-16(%eax), %esp
+	popl	%eax
+.Lfinished_frame_\@:
+.endm
+
 .macro SAVE_ALL pt_regs_ax=%eax switch_stacks=0
 	cld
 	PUSH_GS
+	FIXUP_FRAME
 	pushl	%fs
 	pushl	%es
 	pushl	%ds
@@ -358,9 +451,6 @@
  * switch to it before we do any copying.
  */
 
-#define CS_FROM_ENTRY_STACK	(1 << 31)
-#define CS_FROM_USER_CR3	(1 << 30)
-
 .macro SWITCH_TO_KERNEL_STACK
 
 	ALTERNATIVE     "", "jmp .Lend_\@", X86_FEATURE_XENPV
@@ -374,13 +464,6 @@
 	 * that register for the time this macro runs
 	 */
 
-	/*
-	 * The high bits of the CS dword (__csh) are used for
-	 * CS_FROM_ENTRY_STACK and CS_FROM_USER_CR3. Clear them in case
-	 * hardware didn't do this for us.
-	 */
-	andl	$(0x0000ffff), PT_CS(%esp)
-
 	/* Are we on the entry stack? Bail out if not! */
 	movl	PER_CPU_VAR(cpu_entry_area), %ecx
 	addl	$CPU_ENTRY_AREA_entry_stack + SIZEOF_entry_stack, %ecx
@@ -990,6 +1073,7 @@ restore_all:
 	/* Restore user state */
 	RESTORE_REGS pop=4			# skip orig_eax/error_code
 .Lirq_return:
+	IRET_FRAME
 	/*
 	 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization
 	 * when returning from IPI handler and when returning from
@@ -1340,6 +1424,7 @@ END(page_fault)
 
 common_exception:
 	/* the function address is in %gs's slot on the stack */
+	FIXUP_FRAME
 	pushl	%fs
 	pushl	%es
 	pushl	%ds

arch/x86/include/asm/kexec.h

@@ -70,22 +70,6 @@ struct kimage;
 #define KEXEC_BACKUP_SRC_START	(0UL)
 #define KEXEC_BACKUP_SRC_END	(640 * 1024UL - 1)	/* 640K */
 
-/*
- * CPU does not save ss and sp on stack if execution is already
- * running in kernel mode at the time of NMI occurrence. This code
- * fixes it.
- */
-static inline void crash_fixup_ss_esp(struct pt_regs *newregs,
-				      struct pt_regs *oldregs)
-{
-#ifdef CONFIG_X86_32
-	newregs->sp = (unsigned long)&(oldregs->sp);
-	asm volatile("xorl %%eax, %%eax\n\t"
-		     "movw %%ss, %%ax\n\t"
-		     :"=a"(newregs->ss));
-#endif
-}
-
 /*
  * This function is responsible for capturing register states if coming
  * via panic otherwise just fix up the ss and sp if coming via kernel
@@ -96,7 +80,6 @@ static inline void crash_setup_regs(struct pt_regs *newregs,
 {
 	if (oldregs) {
 		memcpy(newregs, oldregs, sizeof(*newregs));
-		crash_fixup_ss_esp(newregs, oldregs);
 	} else {
 #ifdef CONFIG_X86_32
 		asm volatile("movl %%ebx,%0" : "=m"(newregs->bx));

arch/x86/include/asm/ptrace.h

@@ -166,14 +166,10 @@ static inline bool user_64bit_mode(struct pt_regs *regs)
 #define compat_user_stack_pointer()	current_pt_regs()->sp
 #endif
 
-#ifdef CONFIG_X86_32
-extern unsigned long kernel_stack_pointer(struct pt_regs *regs);
-#else
 static inline unsigned long kernel_stack_pointer(struct pt_regs *regs)
 {
 	return regs->sp;
 }
-#endif
 
 #define GET_IP(regs) ((regs)->ip)
 #define GET_FP(regs) ((regs)->bp)
@@ -201,14 +197,6 @@ static inline unsigned long regs_get_register(struct pt_regs *regs,
 	if (unlikely(offset > MAX_REG_OFFSET))
 		return 0;
 #ifdef CONFIG_X86_32
-	/*
-	 * Traps from the kernel do not save sp and ss.
-	 * Use the helper function to retrieve sp.
-	 */
-	if (offset == offsetof(struct pt_regs, sp) &&
-	    regs->cs == __KERNEL_CS)
-		return kernel_stack_pointer(regs);
-
 	/* The selector fields are 16-bit. */
 	if (offset == offsetof(struct pt_regs, cs) ||
 	    offset == offsetof(struct pt_regs, ss) ||
@@ -234,8 +222,7 @@ static inline unsigned long regs_get_register(struct pt_regs *regs,
 static inline int regs_within_kernel_stack(struct pt_regs *regs,
 					   unsigned long addr)
 {
-	return ((addr & ~(THREAD_SIZE - 1))  ==
-		(kernel_stack_pointer(regs) & ~(THREAD_SIZE - 1)));
+	return ((addr & ~(THREAD_SIZE - 1)) == (regs->sp & ~(THREAD_SIZE - 1)));
 }
 
 /**
@@ -249,7 +236,7 @@ static inline int regs_within_kernel_stack(struct pt_regs *regs,
  */
 static inline unsigned long *regs_get_kernel_stack_nth_addr(struct pt_regs *regs, unsigned int n)
 {
-	unsigned long *addr = (unsigned long *)kernel_stack_pointer(regs);
+	unsigned long *addr = (unsigned long *)regs->sp;
 
 	addr += n;
 	if (regs_within_kernel_stack(regs, (unsigned long)addr))

arch/x86/include/asm/stacktrace.h

@@ -78,7 +78,7 @@ static inline unsigned long *
 get_stack_pointer(struct task_struct *task, struct pt_regs *regs)
 {
 	if (regs)
-		return (unsigned long *)kernel_stack_pointer(regs);
+		return (unsigned long *)regs->sp;
 
 	if (task == current)
 		return __builtin_frame_address(0);

arch/x86/kernel/crash.c

@@ -73,14 +73,6 @@ static inline void cpu_crash_vmclear_loaded_vmcss(void)
 
 static void kdump_nmi_callback(int cpu, struct pt_regs *regs)
 {
-#ifdef CONFIG_X86_32
-	struct pt_regs fixed_regs;
-
-	if (!user_mode(regs)) {
-		crash_fixup_ss_esp(&fixed_regs, regs);
-		regs = &fixed_regs;
-	}
-#endif
 	crash_save_cpu(regs, cpu);
 
 	/*

arch/x86/kernel/ftrace_32.S

@@ -10,6 +10,7 @@
 #include <asm/ftrace.h>
 #include <asm/nospec-branch.h>
 #include <asm/frame.h>
+#include <asm/asm-offsets.h>
 
 # define function_hook	__fentry__
 EXPORT_SYMBOL(__fentry__)
@@ -90,26 +91,38 @@ END(ftrace_caller)
 
 ENTRY(ftrace_regs_caller)
 	/*
-	 * i386 does not save SS and ESP when coming from kernel.
-	 * Instead, to get sp, &regs->sp is used (see ptrace.h).
-	 * Unfortunately, that means eflags must be at the same location
-	 * as the current return ip is. We move the return ip into the
-	 * regs->ip location, and move flags into the return ip location.
+	 * We're here from an mcount/fentry CALL, and the stack frame looks like:
+	 *
+	 *  <previous context>
+	 *  RET-IP
+	 *
+	 * The purpose of this function is to call out in an emulated INT3
+	 * environment with a stack frame like:
+	 *
+	 *  <previous context>
+	 *  gap / RET-IP
+	 *  gap
+	 *  gap
+	 *  gap
+	 *  pt_regs
+	 *
+	 * We do _NOT_ restore: ss, flags, cs, gs, fs, es, ds
 	 */
-	pushl	$__KERNEL_CS
-	pushl	4(%esp)				/* Save the return ip */
-	pushl	$0				/* Load 0 into orig_ax */
+	subl	$3*4, %esp	# RET-IP + 3 gaps
+	pushl	%ss		# ss
+	pushl	%esp		# points at ss
+	addl	$5*4, (%esp)	#   make it point at <previous context>
+	pushfl			# flags
+	pushl	$__KERNEL_CS	# cs
+	pushl	7*4(%esp)	# ip <- RET-IP
+	pushl	$0		# orig_eax
+
 	pushl	%gs
 	pushl	%fs
 	pushl	%es
 	pushl	%ds
-	pushl	%eax
-
-	/* Get flags and place them into the return ip slot */
-	pushf
-	popl	%eax
-	movl	%eax, 8*4(%esp)
 
+	pushl	%eax
 	pushl	%ebp
 	pushl	%edi
 	pushl	%esi
@@ -119,39 +132,35 @@ ENTRY(ftrace_regs_caller)
 
 	ENCODE_FRAME_POINTER
 
-	movl	12*4(%esp), %eax		/* Load ip (1st parameter) */
-	subl	$MCOUNT_INSN_SIZE, %eax		/* Adjust ip */
-	movl	15*4(%esp), %edx		/* Load parent ip (2nd parameter) */
-	movl	function_trace_op, %ecx		/* Save ftrace_pos in 3rd parameter */
-	pushl	%esp				/* Save pt_regs as 4th parameter */
+	movl	PT_EIP(%esp), %eax	# 1st argument: IP
+	subl	$MCOUNT_INSN_SIZE, %eax
+	movl	21*4(%esp), %edx	# 2nd argument: parent ip
+	movl	function_trace_op, %ecx	# 3rd argument: ftrace_pos
+	pushl	%esp			# 4th argument: pt_regs
 
 GLOBAL(ftrace_regs_call)
 	call	ftrace_stub
 
-	addl	$4, %esp			/* Skip pt_regs */
+	addl	$4, %esp		# skip 4th argument
 
-	/* restore flags */
-	push	14*4(%esp)
-	popf
+	/* place IP below the new SP */
+	movl	PT_OLDESP(%esp), %eax
+	movl	PT_EIP(%esp), %ecx
+	movl	%ecx, -4(%eax)
 
-	/* Move return ip back to its original location */
-	movl	12*4(%esp), %eax
-	movl	%eax, 14*4(%esp)
+	/* place EAX below that */
+	movl	PT_EAX(%esp), %ecx
+	movl	%ecx, -8(%eax)
 
 	popl	%ebx
 	popl	%ecx
 	popl	%edx
 	popl	%esi
 	popl	%edi
 	popl	%ebp
-	popl	%eax
-	popl	%ds
-	popl	%es
-	popl	%fs
-	popl	%gs
 
-	/* use lea to not affect flags */
-	lea	3*4(%esp), %esp			/* Skip orig_ax, ip and cs */
+	lea	-8(%eax), %esp
+	popl	%eax
 
 	jmp	.Lftrace_ret
 

arch/x86/kernel/kgdb.c

@@ -118,14 +118,6 @@ char *dbg_get_reg(int regno, void *mem, struct pt_regs *regs)
 
 #ifdef CONFIG_X86_32
 	switch (regno) {
-	case GDB_SS:
-		if (!user_mode(regs))
-			*(unsigned long *)mem = __KERNEL_DS;
-		break;
-	case GDB_SP:
-		if (!user_mode(regs))
-			*(unsigned long *)mem = kernel_stack_pointer(regs);
-		break;
 	case GDB_GS:
 	case GDB_FS:
 		*(unsigned long *)mem = 0xFFFF;

arch/x86/kernel/kprobes/common.h

@@ -72,8 +72,8 @@
 	"	popl %edi\n"			\
 	"	popl %ebp\n"			\
 	"	popl %eax\n"			\
-	/* Skip ds, es, fs, gs, orig_ax, and ip. Note: don't pop cs here*/\
-	"	addl $24, %esp\n"
+	/* Skip ds, es, fs, gs, orig_ax, ip, and cs. */\
+	"	addl $7*4, %esp\n"
 #endif
 
 /* Ensure if the instruction can be boostable */