diff mbox series

[v4,07/17] scs: add support for stack usage debugging

Message ID 20191101221150.116536-8-samitolvanen@google.com (mailing list archive)
State New, archived
Headers show
Series [v4,01/17] arm64: mm: avoid x18 in idmap_kpti_install_ng_mappings | expand

Commit Message

Sami Tolvanen Nov. 1, 2019, 10:11 p.m. UTC
Implements CONFIG_DEBUG_STACK_USAGE for shadow stacks. When enabled,
also prints out the highest shadow stack usage per process.

Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
---
 kernel/scs.c | 39 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 39 insertions(+)

Comments

Kees Cook Nov. 2, 2019, 5:31 p.m. UTC | #1
On Fri, Nov 01, 2019 at 03:11:40PM -0700, Sami Tolvanen wrote:
> Implements CONFIG_DEBUG_STACK_USAGE for shadow stacks. When enabled,
> also prints out the highest shadow stack usage per process.
> 
> Signed-off-by: Sami Tolvanen <samitolvanen@google.com>

Thanks for helping me find this Kconfig. :) :)

Reviewed-by: Kees Cook <keescook@chromium.org>

-Kees

> ---
>  kernel/scs.c | 39 +++++++++++++++++++++++++++++++++++++++
>  1 file changed, 39 insertions(+)
> 
> diff --git a/kernel/scs.c b/kernel/scs.c
> index 7780fc4e29ac..67c43af627d1 100644
> --- a/kernel/scs.c
> +++ b/kernel/scs.c
> @@ -167,6 +167,44 @@ int scs_prepare(struct task_struct *tsk, int node)
>  	return 0;
>  }
>  
> +#ifdef CONFIG_DEBUG_STACK_USAGE
> +static inline unsigned long scs_used(struct task_struct *tsk)
> +{
> +	unsigned long *p = __scs_base(tsk);
> +	unsigned long *end = scs_magic(tsk);
> +	uintptr_t s = (uintptr_t)p;
> +
> +	while (p < end && *p)
> +		p++;
> +
> +	return (uintptr_t)p - s;
> +}
> +
> +static void scs_check_usage(struct task_struct *tsk)
> +{
> +	static DEFINE_SPINLOCK(lock);
> +	static unsigned long highest;
> +	unsigned long used = scs_used(tsk);
> +
> +	if (used <= highest)
> +		return;
> +
> +	spin_lock(&lock);
> +
> +	if (used > highest) {
> +		pr_info("%s: highest shadow stack usage %lu bytes\n",
> +			__func__, used);
> +		highest = used;
> +	}
> +
> +	spin_unlock(&lock);
> +}
> +#else
> +static inline void scs_check_usage(struct task_struct *tsk)
> +{
> +}
> +#endif
> +
>  bool scs_corrupted(struct task_struct *tsk)
>  {
>  	return *scs_magic(tsk) != SCS_END_MAGIC;
> @@ -181,6 +219,7 @@ void scs_release(struct task_struct *tsk)
>  		return;
>  
>  	WARN_ON(scs_corrupted(tsk));
> +	scs_check_usage(tsk);
>  
>  	scs_account(tsk, -1);
>  	task_set_scs(tsk, NULL);
> -- 
> 2.24.0.rc1.363.gb1bccd3e3d-goog
>
Mark Rutland Nov. 4, 2019, 12:40 p.m. UTC | #2
On Fri, Nov 01, 2019 at 03:11:40PM -0700, Sami Tolvanen wrote:
> Implements CONFIG_DEBUG_STACK_USAGE for shadow stacks. When enabled,
> also prints out the highest shadow stack usage per process.
> 
> Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
> ---
>  kernel/scs.c | 39 +++++++++++++++++++++++++++++++++++++++
>  1 file changed, 39 insertions(+)
> 
> diff --git a/kernel/scs.c b/kernel/scs.c
> index 7780fc4e29ac..67c43af627d1 100644
> --- a/kernel/scs.c
> +++ b/kernel/scs.c
> @@ -167,6 +167,44 @@ int scs_prepare(struct task_struct *tsk, int node)
>  	return 0;
>  }
>  
> +#ifdef CONFIG_DEBUG_STACK_USAGE
> +static inline unsigned long scs_used(struct task_struct *tsk)
> +{
> +	unsigned long *p = __scs_base(tsk);
> +	unsigned long *end = scs_magic(tsk);
> +	uintptr_t s = (uintptr_t)p;

As previously, please use unsigned long for consistency.

> +
> +	while (p < end && *p)
> +		p++;

I think this is the only place where we legtimately access the shadow
call stack directly. When using SCS and KASAN, are the
compiler-generated accesses to the SCS instrumented?

If not, it might make sense to make this:

	while (p < end && READ_ONCE_NOCKECK(*p))

... and poison the allocation from KASAN's PoV, so that we can find
unintentional accesses more easily. 

Mark.

> +
> +	return (uintptr_t)p - s;
> +}
> +
> +static void scs_check_usage(struct task_struct *tsk)
> +{
> +	static DEFINE_SPINLOCK(lock);
> +	static unsigned long highest;
> +	unsigned long used = scs_used(tsk);
> +
> +	if (used <= highest)
> +		return;
> +
> +	spin_lock(&lock);
> +
> +	if (used > highest) {
> +		pr_info("%s: highest shadow stack usage %lu bytes\n",
> +			__func__, used);
> +		highest = used;
> +	}
> +
> +	spin_unlock(&lock);
> +}
> +#else
> +static inline void scs_check_usage(struct task_struct *tsk)
> +{
> +}
> +#endif
> +
>  bool scs_corrupted(struct task_struct *tsk)
>  {
>  	return *scs_magic(tsk) != SCS_END_MAGIC;
> @@ -181,6 +219,7 @@ void scs_release(struct task_struct *tsk)
>  		return;
>  
>  	WARN_ON(scs_corrupted(tsk));
> +	scs_check_usage(tsk);
>  
>  	scs_account(tsk, -1);
>  	task_set_scs(tsk, NULL);
> -- 
> 2.24.0.rc1.363.gb1bccd3e3d-goog
>
Sami Tolvanen Nov. 4, 2019, 9:35 p.m. UTC | #3
On Mon, Nov 4, 2019 at 4:40 AM Mark Rutland <mark.rutland@arm.com> wrote:
> > +#ifdef CONFIG_DEBUG_STACK_USAGE
> > +static inline unsigned long scs_used(struct task_struct *tsk)
> > +{
> > +     unsigned long *p = __scs_base(tsk);
> > +     unsigned long *end = scs_magic(tsk);
> > +     uintptr_t s = (uintptr_t)p;
>
> As previously, please use unsigned long for consistency.

Ack.

> > +     while (p < end && *p)
> > +             p++;
>
> I think this is the only place where we legtimately access the shadow
> call stack directly.

There's also scs_corrupted, which checks that the end magic is intact.

> When using SCS and KASAN, are the
> compiler-generated accesses to the SCS instrumented?
>
> If not, it might make sense to make this:
>
>         while (p < end && READ_ONCE_NOCKECK(*p))
>
> ... and poison the allocation from KASAN's PoV, so that we can find
> unintentional accesses more easily.

Sure, that makes sense. I can poison the allocation for the
non-vmalloc case, I'll just need to refactor scs_set_magic to happen
before the poisoning.

Sami
Mark Rutland Nov. 5, 2019, 9:17 a.m. UTC | #4
On Mon, Nov 04, 2019 at 01:35:28PM -0800, Sami Tolvanen wrote:
> On Mon, Nov 4, 2019 at 4:40 AM Mark Rutland <mark.rutland@arm.com> wrote:
> > > +#ifdef CONFIG_DEBUG_STACK_USAGE
> > > +static inline unsigned long scs_used(struct task_struct *tsk)
> > > +{
> > > +     unsigned long *p = __scs_base(tsk);
> > > +     unsigned long *end = scs_magic(tsk);
> > > +     uintptr_t s = (uintptr_t)p;
> >
> > As previously, please use unsigned long for consistency.
> 
> Ack.
> 
> > > +     while (p < end && *p)
> > > +             p++;
> >
> > I think this is the only place where we legtimately access the shadow
> > call stack directly.
> 
> There's also scs_corrupted, which checks that the end magic is intact.

Ah, true. I missed that.

> > When using SCS and KASAN, are the
> > compiler-generated accesses to the SCS instrumented?
> >
> > If not, it might make sense to make this:
> >
> >         while (p < end && READ_ONCE_NOCKECK(*p))
> >
> > ... and poison the allocation from KASAN's PoV, so that we can find
> > unintentional accesses more easily.
> 
> Sure, that makes sense. I can poison the allocation for the
> non-vmalloc case, I'll just need to refactor scs_set_magic to happen
> before the poisoning.

Sounds good!

Mark.
diff mbox series

Patch

diff --git a/kernel/scs.c b/kernel/scs.c
index 7780fc4e29ac..67c43af627d1 100644
--- a/kernel/scs.c
+++ b/kernel/scs.c
@@ -167,6 +167,44 @@  int scs_prepare(struct task_struct *tsk, int node)
 	return 0;
 }
 
+#ifdef CONFIG_DEBUG_STACK_USAGE
+static inline unsigned long scs_used(struct task_struct *tsk)
+{
+	unsigned long *p = __scs_base(tsk);
+	unsigned long *end = scs_magic(tsk);
+	uintptr_t s = (uintptr_t)p;
+
+	while (p < end && *p)
+		p++;
+
+	return (uintptr_t)p - s;
+}
+
+static void scs_check_usage(struct task_struct *tsk)
+{
+	static DEFINE_SPINLOCK(lock);
+	static unsigned long highest;
+	unsigned long used = scs_used(tsk);
+
+	if (used <= highest)
+		return;
+
+	spin_lock(&lock);
+
+	if (used > highest) {
+		pr_info("%s: highest shadow stack usage %lu bytes\n",
+			__func__, used);
+		highest = used;
+	}
+
+	spin_unlock(&lock);
+}
+#else
+static inline void scs_check_usage(struct task_struct *tsk)
+{
+}
+#endif
+
 bool scs_corrupted(struct task_struct *tsk)
 {
 	return *scs_magic(tsk) != SCS_END_MAGIC;
@@ -181,6 +219,7 @@  void scs_release(struct task_struct *tsk)
 		return;
 
 	WARN_ON(scs_corrupted(tsk));
+	scs_check_usage(tsk);
 
 	scs_account(tsk, -1);
 	task_set_scs(tsk, NULL);