[tip:x86/urgent] x86/unwind: Add end-of-stack check for ftrace handlers

From: tip-bot for Josh Poimboeuf
Date: Wed May 24 2017 - 06:25:45 EST


Commit-ID: 519fb5c3350d1b5225b27b1cac55144f79351718
Gitweb: http://git.kernel.org/tip/519fb5c3350d1b5225b27b1cac55144f79351718
Author: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>
AuthorDate: Tue, 23 May 2017 10:37:30 -0500
Committer: Ingo Molnar <mingo@xxxxxxxxxx>
CommitDate: Wed, 24 May 2017 09:05:16 +0200

x86/unwind: Add end-of-stack check for ftrace handlers

Dave Jones and Steven Rostedt reported unwinder warnings like the
following:

WARNING: kernel stack frame pointer at ffff8800bda0ff30 in sshd:1090 has bad value 000055b32abf1fa8

In both cases, the unwinder was attempting to unwind from an ftrace
handler into entry code. The callchain was something like:

syscall entry code
C function
ftrace handler
save_stack_trace()

The problem is that the unwinder's end-of-stack logic gets confused by
the way ftrace lays out the stack frame (with fentry enabled).

I was able to recreate this warning with:

echo call_usermodehelper_exec_async:stacktrace > /sys/kernel/debug/tracing/set_ftrace_filter
(exit login session)

I considered fixing this by changing the ftrace code to rewrite the
stack to make the unwinder happy. But that seemed too intrusive after I
implemented it. Instead, just add another check to the unwinder's
end-of-stack logic to detect this special case.

Side note: We could probably get rid of these end-of-stack checks by
encoding the frame pointer for syscall entry just like we do for
interrupt entry. That would be simpler, but it would also be a lot more
intrusive since it would slightly affect the performance of every
syscall.

Reported-by: Dave Jones <davej@xxxxxxxxxxxxxxxxx>
Reported-by: Steven Rostedt <rostedt@xxxxxxxxxxx>
Signed-off-by: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>
Acked-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Petr Mladek <pmladek@xxxxxxxx>
Cc: live-patching@xxxxxxxxxxxxxxx
Fixes: c32c47c68a0a ("x86/unwind: Warn on bad frame pointer")
Link: http://lkml.kernel.org/r/671ba22fbc0156b8f7e0cfa5ab2a795e08bc37e1.1495553739.git.jpoimboe@xxxxxxxxxx
Signed-off-by: Ingo Molnar <mingo@xxxxxxxxxx>
---
arch/x86/kernel/unwind_frame.c | 49 ++++++++++++++++++++++++++++++++++--------
1 file changed, 40 insertions(+), 9 deletions(-)

diff --git a/arch/x86/kernel/unwind_frame.c b/arch/x86/kernel/unwind_frame.c
index 82c6d7f..b9389d7 100644
--- a/arch/x86/kernel/unwind_frame.c
+++ b/arch/x86/kernel/unwind_frame.c
@@ -104,6 +104,11 @@ static inline unsigned long *last_frame(struct unwind_state *state)
return (unsigned long *)task_pt_regs(state->task) - 2;
}

+static bool is_last_frame(struct unwind_state *state)
+{
+ return state->bp == last_frame(state);
+}
+
#ifdef CONFIG_X86_32
#define GCC_REALIGN_WORDS 3
#else
@@ -115,16 +120,15 @@ static inline unsigned long *last_aligned_frame(struct unwind_state *state)
return last_frame(state) - GCC_REALIGN_WORDS;
}

-static bool is_last_task_frame(struct unwind_state *state)
+static bool is_last_aligned_frame(struct unwind_state *state)
{
unsigned long *last_bp = last_frame(state);
unsigned long *aligned_bp = last_aligned_frame(state);

/*
- * We have to check for the last task frame at two different locations
- * because gcc can occasionally decide to realign the stack pointer and
- * change the offset of the stack frame in the prologue of a function
- * called by head/entry code. Examples:
+ * GCC can occasionally decide to realign the stack pointer and change
+ * the offset of the stack frame in the prologue of a function called
+ * by head/entry code. Examples:
*
* <start_secondary>:
* push %edi
@@ -141,11 +145,38 @@ static bool is_last_task_frame(struct unwind_state *state)
* push %rbp
* mov %rsp,%rbp
*
- * Note that after aligning the stack, it pushes a duplicate copy of
- * the return address before pushing the frame pointer.
+ * After aligning the stack, it pushes a duplicate copy of the return
+ * address before pushing the frame pointer.
+ */
+ return (state->bp == aligned_bp && *(aligned_bp + 1) == *(last_bp + 1));
+}
+
+static bool is_last_ftrace_frame(struct unwind_state *state)
+{
+ unsigned long *last_bp = last_frame(state);
+ unsigned long *last_ftrace_bp = last_bp - 3;
+
+ /*
+ * When unwinding from an ftrace handler of a function called by entry
+ * code, the stack layout of the last frame is:
+ *
+ * bp
+ * parent ret addr
+ * bp
+ * function ret addr
+ * parent ret addr
+ * pt_regs
+ * -----------------
*/
- return (state->bp == last_bp ||
- (state->bp == aligned_bp && *(aligned_bp+1) == *(last_bp+1)));
+ return (state->bp == last_ftrace_bp &&
+ *state->bp == *(state->bp + 2) &&
+ *(state->bp + 1) == *(state->bp + 4));
+}
+
+static bool is_last_task_frame(struct unwind_state *state)
+{
+ return is_last_frame(state) || is_last_aligned_frame(state) ||
+ is_last_ftrace_frame(state);
}

/*