Skip to contentSkip to conversation

objtool: Fix switch-table detection

PATCH In review Peter Zijlstra 4 participants 4 messages Feb 8, 2018 – Feb 9, 2018 objtool

lore
Share & tools
lore

Peter Zijlstra

author maintainer
#0 lore raw

PATCH 1/1 objtool: Fix switch-table detection

Linus reported that GCC-7.3 generated a switch-table construct that confused objtool. It turns out that, in particular due to KASAN, it is possible to have unrelated .rodata usage in between the .rodata setup for the switch-table and the following indirect jump.

The simple linear reverse search from the indirect jump would hit upon the KASAN .rodata usage first and fail to find a switch_table, resulting in a spurious 'sibling call with modified stack frame' warning.

Fix this by creating a 'jump-stack' which we can 'unwind' during reversal, thereby skipping over much of the in-between code.

This is not fool proof by any means, but is sufficient to make the known cases work. Future work would be to construct more comprehensive flow analysis code.

Also, since Josh keeps asking, add myself to MAINTAINERS.

---
 MAINTAINERS           |    1 +
 tools/objtool/check.c |   41 +++++++++++++++++++++++++++++++++++++++--
 tools/objtool/check.h |    1 +
 3 files changed, 41 insertions(+), 2 deletions(-)
MAINTAINERS +1 -0
@@ -9847,6 +9847,7 @@ F:	drivers/nfc/nxp-nci
 
 OBJTOOL
 M:	Josh Poimboeuf <[email protected]>
+M:	Peter Zijlstra <[email protected]>
 S:	Supported
 F:	tools/objtool/
 
tools/objtool/check.c +39 -2

73 lines not on this page · raw message

tools/objtool/check.h +1 -0
@@ -47,6 +47,7 @@ struct instruction {
 	bool alt_group, visited, dead_end, ignore, hint, save, restore, ignore_alts;
 	struct symbol *call_dest;
 	struct instruction *jump_dest;
+	struct instruction *first_jump_src;
 	struct list_head alts;
 	struct symbol *func;
 	struct stack_op stack_op;
#1 lore raw
Quoted 3 lines · Peter Zijlstra

Fix this by creating a 'jump-stack' which we can 'unwind' during reversal, thereby skipping over much of the in-between code.

I'm assuming this will come through the normal tip trees, since that's where objtool stuff generally comes from, and this is not very high priority.

Quoted 1 line

And assuming it's the same patch, just with the extra comment and jump_orig -> first_jump_src rename (and MAINTAINERS update), that can be a

since the previous version worked fine for me.

Thanks,

Linus

Acked-by on v1 1/1
#2 lore raw
In reply to Peter Zijlstra
Quoted 18 lines · Peter Zijlstra

Linus reported that GCC-7.3 generated a switch-table construct that confused objtool. It turns out that, in particular due to KASAN, it is possible to have unrelated .rodata usage in between the .rodata setup for the switch-table and the following indirect jump.

The simple linear reverse search from the indirect jump would hit upon the KASAN .rodata usage first and fail to find a switch_table, resulting in a spurious 'sibling call with modified stack frame' warning.

Fix this by creating a 'jump-stack' which we can 'unwind' during reversal, thereby skipping over much of the in-between code.

This is not fool proof by any means, but is sufficient to make the known cases work. Future work would be to construct more comprehensive flow analysis code.

Quoted 5 lines

Also, since Josh keeps asking, add myself to MAINTAINERS.

Thanks again for helping out with objtool maintainership!

No complaints from the 0-day bot, so:

Signature · 1 line
Josh
#3 lore raw
In reply to Peter Zijlstra
objtool: Fix switch-table detection
Commit-ID:  99ce7962d52d1948ad6f2785e308d48e76e0a6ef
Gitweb:     https://git.kernel.org/tip/99ce7962d52d1948ad6f2785e308d48e76e0a6ef
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Thu, 8 Feb 2018 14:02:32 +0100
Committer:  Ingo Molnar <[email protected]>
CommitDate: Fri, 9 Feb 2018 07:20:23 +0100

objtool: Fix switch-table detection

Linus reported that GCC-7.3 generated a switch-table construct that confused objtool. It turns out that, in particular due to KASAN, it is possible to have unrelated .rodata usage in between the .rodata setup for the switch-table and the following indirect jump.

The simple linear reverse search from the indirect jump would hit upon the KASAN .rodata usage first and fail to find a switch_table, resulting in a spurious 'sibling call with modified stack frame' warning.

Fix this by creating a 'jump-stack' which we can 'unwind' during reversal, thereby skipping over much of the in-between code.

This is not fool proof by any means, but is sufficient to make the known cases work. Future work would be to construct more comprehensive flow analysis code.

---
 tools/objtool/check.c | 41 +++++++++++++++++++++++++++++++++++++++--
 tools/objtool/check.h |  1 +
 2 files changed, 40 insertions(+), 2 deletions(-)
tools/objtool/check.c +39 -2
index 9cd028a..2e458eb 100644
@@ -851,8 +851,14 @@ static int add_switch_table(struct objtool_file *file, struct symbol *func,
  *    This is a fairly uncommon pattern which is new for GCC 6.  As of this
  *    writing, there are 11 occurrences of it in the allmodconfig kernel.
  *
+ *    As of GCC 7 there are quite a few more of these and the 'in between' code
+ *    is significant. Esp. with KASAN enabled some of the code between the mov
+ *    and jmpq uses .rodata itself, which can confuse things.
+ *
  *    TODO: Once we have DWARF CFI and smarter instruction decoding logic,
  *    ensure the same register is used in the mov and jump instructions.
+ *
+ *    NOTE: RETPOLINE made it harder still to decode dynamic jumps.
  */
 static struct rela *find_switch_table(struct objtool_file *file,
 				      struct symbol *func,
@@ -874,12 +880,25 @@ static struct rela *find_switch_table(struct objtool_file *file,
 						text_rela->addend + 4);
 		if (!rodata_rela)
 			return NULL;
+
 		file->ignore_unreachables = true;
 		return rodata_rela;
 	}
 
 	/* case 3 */
-	func_for_each_insn_continue_reverse(file, func, insn) {
+	/*
+	 * Backward search using the @first_jump_src links, these help avoid
+	 * much of the 'in between' code. Which avoids us getting confused by
+	 * it.
+	 */
+	for (insn = list_prev_entry(insn, list);
+
+	     &insn->list != &file->insn_list &&
+	     insn->sec == func->sec &&
+	     insn->offset >= func->offset;
+
+	     insn = insn->first_jump_src ?: list_prev_entry(insn, list)) {
+
 		if (insn->type == INSN_JUMP_DYNAMIC)
 			break;
 
@@ -909,14 +928,32 @@ static struct rela *find_switch_table(struct objtool_file *file,
 	return NULL;
 }
 
+
 static int add_func_switch_tables(struct objtool_file *file,
 				  struct symbol *func)
 {
-	struct instruction *insn, *prev_jump = NULL;
+	struct instruction *insn, *last = NULL, *prev_jump = NULL;
 	struct rela *rela, *prev_rela = NULL;
 	int ret;
 
 	func_for_each_insn(file, func, insn) {
+		if (!last)
+			last = insn;
+
+		/*
+		 * Store back-pointers for unconditional forward jumps such
+		 * that find_switch_table() can back-track using those and
+		 * avoid some potentially confusing code.
+		 */
+		if (insn->type == INSN_JUMP_UNCONDITIONAL && insn->jump_dest &&
+		    insn->offset > last->offset &&
+		    insn->jump_dest->offset > insn->offset &&
+		    !insn->jump_dest->first_jump_src) {
+
+			insn->jump_dest->first_jump_src = insn;
+			last = insn->jump_dest;
+		}
+
 		if (insn->type != INSN_JUMP_DYNAMIC)
 			continue;
 
tools/objtool/check.h +1 -0
index dbadb30..23a1d06 100644
@@ -47,6 +47,7 @@ struct instruction {
 	bool alt_group, visited, dead_end, ignore, hint, save, restore, ignore_alts;
 	struct symbol *call_dest;
 	struct instruction *jump_dest;
+	struct instruction *first_jump_src;
 	struct list_head alts;
 	struct symbol *func;
 	struct stack_op stack_op;