[PATCH] objtool: Fix stack-swizzle for FRAME_POINTER=y

[Peter Zijlstra]

On Wed, Feb 10, 2021 at 08:01:23PM -0000, tip-bot2 for Peter Zijlstra wrote:
> objtool: Support stack-swizzle

---
Subject: objtool: Fix stack-swizzle for FRAME_POINTER=y
From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Date: Thu Feb 18 17:14:10 CET 2021

When objtool encounters the stack-swizzle:

	mov %rsp, (%[tos])
	mov %[tos], %rsp
	...
	pop %rsp

Inside a FRAME_POINTER=y build, things go a little screwy because
clearly we're not adjusting the cfa->base. This then results in the
pop %rsp not being detected as a restore of cfa->base so it will turn
into a regular POP and offset the stack, resulting in:

  kernel/softirq.o: warning: objtool: do_softirq()+0xdb: return with modified stack frame

Therefore, have "mov %[tos], %rsp" act like a PUSH (it sorta is
anyway) to balance the things out. We're not too concerned with the
actual stack_size for frame-pointer builds, since we don't generate
ORC data for them anyway.

Fixes: aafeb14e9da2 ("objtool: Support stack-swizzle")
Reported-by: kernel test robot <lkp@xxxxxxxxx>
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
 tools/objtool/check.c |   14 ++++++++++++++
 1 file changed, 14 insertions(+)

--- a/tools/objtool/check.c
+++ b/tools/objtool/check.c
@@ -1996,6 +1996,20 @@ static int update_cfi_state(struct instr
 				}
 			}
 
+			else if (op->dest.reg == CFI_SP &&
+				 cfi->vals[op->src.reg].base == CFI_SP_INDIRECT &&
+				 cfi->vals[op->src.reg].offset == cfa->offset) {
+
+				/*
+				 * The same stack swizzle case 2) as above. But
+				 * because we can't change cfa->base, case 3)
+				 * will become a regular POP. Pretend we're a
+				 * PUSH so things don't go unbalanced.
+				 */
+				cfi->stack_size += 8;
+			}
+
+
 			break;
 
 		case OP_SRC_ADD: