Commit 2e4be0d011f2 ("x86/show_trace_log_lvl: Ensure stack pointer is
aligned, again") was intended to ensure alignment of the stack pointer; but
it also moved the initialization of the "stack" variable down into the loop
header. This was likely intended as a no-op cleanup, since the commit
message does not mention it; however, this caused a behavioral change
because the value of "regs" is different between the two places.

Originally, get_stack_pointer() used the regs provided by the caller; after
that commit, get_stack_pointer() instead uses the regs at the top of the
stack frame the unwinder is looking at. Often, there are no such regs at
all, and "regs" is NULL, causing get_stack_pointer() to fall back to the
task's current stack pointer, which is not what we want here, but probably
happens to mostly work. Other times, the original regs will point to
another regs frame - in that case, the linear guess unwind logic in
show_trace_log_lvl() will start unwinding too far up the stack, causing the
first frame found by the proper unwinder to never be visited, resulting in
a stack trace consisting purely of guess lines.

Fix it by moving the "stack = " assignment back where it belongs.

Fixes: 2e4be0d011f2 ("x86/show_trace_log_lvl: Ensure stack pointer is aligned, again")
Signed-off-by: Jann Horn <jannh@google.com>
---
 arch/x86/kernel/dumpstack.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/arch/x86/kernel/dumpstack.c b/arch/x86/kernel/dumpstack.c
index a7d562697e50..b2b118a8c09b 100644
--- a/arch/x86/kernel/dumpstack.c
+++ b/arch/x86/kernel/dumpstack.c
@@ -195,6 +195,7 @@ static void show_trace_log_lvl(struct task_struct *task, struct pt_regs *regs,
 	printk("%sCall Trace:\n", log_lvl);
 
 	unwind_start(&state, task, regs, stack);
+	stack = stack ?: get_stack_pointer(task, regs);
 	regs = unwind_get_entry_regs(&state, &partial);
 
 	/*
@@ -213,9 +214,7 @@ static void show_trace_log_lvl(struct task_struct *task, struct pt_regs *regs,
 	 * - hardirq stack
 	 * - entry stack
 	 */
-	for (stack = stack ?: get_stack_pointer(task, regs);
-	     stack;
-	     stack = stack_info.next_sp) {
+	for (; stack; stack = stack_info.next_sp) {
 		const char *stack_name;
 
 		stack = PTR_ALIGN(stack, sizeof(long));

-- 
2.49.0.395.g12beb8f557-goog

On 03/25, Jann Horn wrote:
>
> Commit 2e4be0d011f2 ("x86/show_trace_log_lvl: Ensure stack pointer is
> aligned, again") was intended to ensure alignment of the stack pointer; but
> it also moved the initialization of the "stack" variable down into the loop
> header. This was likely intended as a no-op cleanup, since the commit
> message does not mention it;

Yes... initial version didn't do this, this was requested during review
and I didn't realize this adds another problem.

Thanks Jann!

Oleg.

On Tue, Mar 25, 2025 at 3:01 AM Jann Horn <jannh@google.com> wrote:
> Originally, get_stack_pointer() used the regs provided by the caller; after
> that commit, get_stack_pointer() instead uses the regs at the top of the
> stack frame the unwinder is looking at. Often, there are no such regs at
> all, and "regs" is NULL, causing get_stack_pointer() to fall back to the
> task's current stack pointer, which is not what we want here, but probably
> happens to mostly work. Other times, the original regs will point to
> another regs frame - in that case, the linear guess unwind logic in
> show_trace_log_lvl() will start unwinding too far up the stack, causing the
> first frame found by the proper unwinder to never be visited, resulting in
> a stack trace consisting purely of guess lines.

I guess the subject line is kind of misleading - maybe "x86/dumpstack:
Fix misplaced assignment in unwinder" would be better?

* Jann Horn <jannh@google.com> wrote:

> On Tue, Mar 25, 2025 at 3:01 AM Jann Horn <jannh@google.com> wrote:
> > Originally, get_stack_pointer() used the regs provided by the caller; after
> > that commit, get_stack_pointer() instead uses the regs at the top of the
> > stack frame the unwinder is looking at. Often, there are no such regs at
> > all, and "regs" is NULL, causing get_stack_pointer() to fall back to the
> > task's current stack pointer, which is not what we want here, but probably
> > happens to mostly work. Other times, the original regs will point to
> > another regs frame - in that case, the linear guess unwind logic in
> > show_trace_log_lvl() will start unwinding too far up the stack, causing the
> > first frame found by the proper unwinder to never be visited, resulting in
> > a stack trace consisting purely of guess lines.
> 
> I guess the subject line is kind of misleading - maybe "x86/dumpstack:
> Fix misplaced assignment in unwinder" would be better?

Well, it's a bug and the code is broken that results in subpar stack 
dumps from exception contexts that fall back to the guess-dumper, 
right?

So I've edited the subject line to:

   x86/dumpstack: Fix inaccurate unwinding from exception stacks due to misplaced assignment

But I'd have no problem calling it broken either - even if the bug 
doesn't crash anything.

Thanks,

	Ingo

The following commit has been merged into the x86/urgent branch of tip:

Commit-ID:     2c118f50d7fd4d9aefc4533a26f83338b2906b7a
Gitweb:        https://git.kernel.org/tip/2c118f50d7fd4d9aefc4533a26f83338b2906b7a
Author:        Jann Horn <jannh@google.com>
AuthorDate:    Tue, 25 Mar 2025 03:01:23 +01:00
Committer:     Ingo Molnar <mingo@kernel.org>
CommitterDate: Tue, 25 Mar 2025 08:30:43 +01:00

x86/dumpstack: Fix inaccurate unwinding from exception stacks due to misplaced assignment

Commit:

  2e4be0d011f2 ("x86/show_trace_log_lvl: Ensure stack pointer is aligned, again")

was intended to ensure alignment of the stack pointer; but it also moved
the initialization of the "stack" variable down into the loop header.

This was likely intended as a no-op cleanup, since the commit
message does not mention it; however, this caused a behavioral change
because the value of "regs" is different between the two places.

Originally, get_stack_pointer() used the regs provided by the caller; after
that commit, get_stack_pointer() instead uses the regs at the top of the
stack frame the unwinder is looking at. Often, there are no such regs at
all, and "regs" is NULL, causing get_stack_pointer() to fall back to the
task's current stack pointer, which is not what we want here, but probably
happens to mostly work. Other times, the original regs will point to
another regs frame - in that case, the linear guess unwind logic in
show_trace_log_lvl() will start unwinding too far up the stack, causing the
first frame found by the proper unwinder to never be visited, resulting in
a stack trace consisting purely of guess lines.

Fix it by moving the "stack = " assignment back where it belongs.

Fixes: 2e4be0d011f2 ("x86/show_trace_log_lvl: Ensure stack pointer is aligned, again")
Signed-off-by: Jann Horn <jannh@google.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20250325-2025-03-unwind-fixes-v1-2-acd774364768@google.com
---
 arch/x86/kernel/dumpstack.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/arch/x86/kernel/dumpstack.c b/arch/x86/kernel/dumpstack.c
index 91639d1..c6fefd4 100644
--- a/arch/x86/kernel/dumpstack.c
+++ b/arch/x86/kernel/dumpstack.c
@@ -195,6 +195,7 @@ static void show_trace_log_lvl(struct task_struct *task, struct pt_regs *regs,
 	printk("%sCall Trace:\n", log_lvl);
 
 	unwind_start(&state, task, regs, stack);
+	stack = stack ?: get_stack_pointer(task, regs);
 	regs = unwind_get_entry_regs(&state, &partial);
 
 	/*
@@ -213,9 +214,7 @@ static void show_trace_log_lvl(struct task_struct *task, struct pt_regs *regs,
 	 * - hardirq stack
 	 * - entry stack
 	 */
-	for (stack = stack ?: get_stack_pointer(task, regs);
-	     stack;
-	     stack = stack_info.next_sp) {
+	for (; stack; stack = stack_info.next_sp) {
 		const char *stack_name;
 
 		stack = PTR_ALIGN(stack, sizeof(long));