x86/unwind/orc: Fix premature unwind stoppage due to IRET frames

The following execution path is possible:

  fsnotify()
    [ realign the stack and store previous SP in R10 ]
    <IRQ>
      [ only IRET regs saved ]
      common_interrupt()
        interrupt_entry()
	  <NMI>
	    [ full pt_regs saved ]
	    ...
	    [ unwind stack ]

When the unwinder goes through the NMI and the IRQ on the stack, and
then sees fsnotify(), it doesn't have access to the value of R10,
because it only has the five IRET registers.  So the unwind stops
prematurely.

However, because the interrupt_entry() code is careful not to clobber
R10 before saving the full regs, the unwinder should be able to read R10
from the previously saved full pt_regs associated with the NMI.

Handle this case properly.  When encountering an IRET regs frame
immediately after a full pt_regs frame, use the pt_regs as a backup
which can be used to get the C register values.

Also, note that a call frame resets the 'prev_regs' value, because a
function is free to clobber the registers.  For this fix to work, the
IRET and full regs frames must be adjacent, with no FUNC frames in
between.  So replace the FUNC hint in interrupt_entry() with an
IRET_REGS hint.

Fixes: ee9f8fc ("x86/unwind: Add the ORC unwinder")
Reviewed-by: Miroslav Benes <[email protected]>
Signed-off-by: Josh Poimboeuf <[email protected]>
Signed-off-by: Ingo Molnar <[email protected]>
Cc: Andy Lutomirski <[email protected]>
Cc: Dave Jones <[email protected]>
Cc: Jann Horn <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Vince Weaver <[email protected]>
Link: https://lore.kernel.org/r/97a408167cc09f1cfa0de31a7b70dd88868d743f.1587808742.git.jpoimboe@redhat.com

Author: jpoimboe

arch/x86/entry/entry_64.S

@@ -511,7 +511,7 @@ SYM_CODE_END(spurious_entries_start)
  * +----------------------------------------------------+
  */
 SYM_CODE_START(interrupt_entry)
-	UNWIND_HINT_FUNC
+	UNWIND_HINT_IRET_REGS offset=16
 	ASM_CLAC
 	cld
 
@@ -543,9 +543,9 @@ SYM_CODE_START(interrupt_entry)
 	pushq	5*8(%rdi)		/* regs->eflags */
 	pushq	4*8(%rdi)		/* regs->cs */
 	pushq	3*8(%rdi)		/* regs->ip */
+	UNWIND_HINT_IRET_REGS
 	pushq	2*8(%rdi)		/* regs->orig_ax */
 	pushq	8(%rdi)			/* return address */
-	UNWIND_HINT_FUNC
 
 	movq	(%rdi), %rdi
 	jmp	2f

arch/x86/include/asm/unwind.h

@@ -19,7 +19,7 @@ struct unwind_state {
 #if defined(CONFIG_UNWINDER_ORC)
 	bool signal, full_regs;
 	unsigned long sp, bp, ip;
-	struct pt_regs *regs;
+	struct pt_regs *regs, *prev_regs;
 #elif defined(CONFIG_UNWINDER_FRAME_POINTER)
 	bool got_irq;
 	unsigned long *bp, *orig_sp, ip;

arch/x86/kernel/unwind_orc.c

@@ -384,9 +384,38 @@ static bool deref_stack_iret_regs(struct unwind_state *state, unsigned long addr
 	return true;
 }
 
+/*
+ * If state->regs is non-NULL, and points to a full pt_regs, just get the reg
+ * value from state->regs.
+ *
+ * Otherwise, if state->regs just points to IRET regs, and the previous frame
+ * had full regs, it's safe to get the value from the previous regs.  This can
+ * happen when early/late IRQ entry code gets interrupted by an NMI.
+ */
+static bool get_reg(struct unwind_state *state, unsigned int reg_off,
+		    unsigned long *val)
+{
+	unsigned int reg = reg_off/8;
+
+	if (!state->regs)
+		return false;
+
+	if (state->full_regs) {
+		*val = ((unsigned long *)state->regs)[reg];
+		return true;
+	}
+
+	if (state->prev_regs) {
+		*val = ((unsigned long *)state->prev_regs)[reg];
+		return true;
+	}
+
+	return false;
+}
+
 bool unwind_next_frame(struct unwind_state *state)
 {
-	unsigned long ip_p, sp, orig_ip = state->ip, prev_sp = state->sp;
+	unsigned long ip_p, sp, tmp, orig_ip = state->ip, prev_sp = state->sp;
 	enum stack_type prev_type = state->stack_info.type;
 	struct orc_entry *orc;
 	bool indirect = false;
@@ -448,39 +477,35 @@ bool unwind_next_frame(struct unwind_state *state)
 		break;
 
 	case ORC_REG_R10:
-		if (!state->regs || !state->full_regs) {
+		if (!get_reg(state, offsetof(struct pt_regs, r10), &sp)) {
 			orc_warn_current("missing R10 value at %pB\n",
 					 (void *)state->ip);
 			goto err;
 		}
-		sp = state->regs->r10;
 		break;
 
 	case ORC_REG_R13:
-		if (!state->regs || !state->full_regs) {
+		if (!get_reg(state, offsetof(struct pt_regs, r13), &sp)) {
 			orc_warn_current("missing R13 value at %pB\n",
 					 (void *)state->ip);
 			goto err;
 		}
-		sp = state->regs->r13;
 		break;
 
 	case ORC_REG_DI:
-		if (!state->regs || !state->full_regs) {
+		if (!get_reg(state, offsetof(struct pt_regs, di), &sp)) {
 			orc_warn_current("missing RDI value at %pB\n",
 					 (void *)state->ip);
 			goto err;
 		}
-		sp = state->regs->di;
 		break;
 
 	case ORC_REG_DX:
-		if (!state->regs || !state->full_regs) {
+		if (!get_reg(state, offsetof(struct pt_regs, dx), &sp)) {
 			orc_warn_current("missing DX value at %pB\n",
 					 (void *)state->ip);
 			goto err;
 		}
-		sp = state->regs->dx;
 		break;
 
 	default:
@@ -507,6 +532,7 @@ bool unwind_next_frame(struct unwind_state *state)
 
 		state->sp = sp;
 		state->regs = NULL;
+		state->prev_regs = NULL;
 		state->signal = false;
 		break;
 
@@ -518,6 +544,7 @@ bool unwind_next_frame(struct unwind_state *state)
 		}
 
 		state->regs = (struct pt_regs *)sp;
+		state->prev_regs = NULL;
 		state->full_regs = true;
 		state->signal = true;
 		break;
@@ -529,6 +556,8 @@ bool unwind_next_frame(struct unwind_state *state)
 			goto err;
 		}
 
+		if (state->full_regs)
+			state->prev_regs = state->regs;
 		state->regs = (void *)sp - IRET_FRAME_OFFSET;
 		state->full_regs = false;
 		state->signal = true;
@@ -543,8 +572,8 @@ bool unwind_next_frame(struct unwind_state *state)
 	/* Find BP: */
 	switch (orc->bp_reg) {
 	case ORC_REG_UNDEFINED:
-		if (state->regs && state->full_regs)
-			state->bp = state->regs->bp;
+		if (get_reg(state, offsetof(struct pt_regs, bp), &tmp))
+			state->bp = tmp;
 		break;
 
 	case ORC_REG_PREV_SP: