diff mbox series

[v2,18/39] mm: Add guard pages around a shadow stack.

Message ID 20220929222936.14584-19-rick.p.edgecombe@intel.com (mailing list archive)
State New
Headers show
Series Shadowstacks for userspace | expand

Commit Message

Rick Edgecombe Sept. 29, 2022, 10:29 p.m. UTC 
From: Yu-cheng Yu <yu-cheng.yu@intel.com>

The architecture of shadow stack constrains the ability of userspace to
move the shadow stack pointer (SSP) in order to  prevent corrupting or
switching to other shadow stacks. The RSTORSSP can move the spp to
different shadow stacks, but it requires a specially placed token in order
to do this. However, the architecture does not prevent incrementing the
stack pointer to wander onto an adjacent shadow stack. To prevent this in
software, enforce guard pages at the beginning of shadow stack vmas, such
that there will always be a gap between adjacent shadow stacks.

Make the gap big enough so that no userspace SSP changing operations
(besides RSTORSSP), can move the SSP from one stack to the next. The
SSP can increment or decrement by CALL, RET  and INCSSP. CALL and RET
can move the SSP by a maximum of 8 bytes, at which point the shadow
stack would be accessed.

The INCSSP instruction can also increment the shadow stack pointer. It
is the shadow stack analog of an instruction like:

	addq    $0x80, %rsp

However, there is one important difference between an ADD on %rsp and
INCSSP. In addition to modifying SSP, INCSSP also reads from the memory
of the first and last elements that were "popped". It can be thought of
as acting like this:

READ_ONCE(ssp);       // read+discard top element on stack
ssp += nr_to_pop * 8; // move the shadow stack
READ_ONCE(ssp-8);     // read+discard last popped stack element

The maximum distance INCSSP can move the SSP is 2040 bytes, before it
would read the memory. Therefore a single page gap will be enough to
prevent any operation from shifting the SSP to an adjacent stack, since
it would have to land in the gap at least once, causing a fault.

This could be accomplished by using VM_GROWSDOWN, but this has a
downside. The behavior would allow shadow stack's to grow, which is
unneeded and adds a strange difference to how most regular stacks work.

Signed-off-by: Yu-cheng Yu <yu-cheng.yu@intel.com>
Co-developed-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Cc: Kees Cook <keescook@chromium.org>

---

v2:
 - Use __weak instead of #ifdef (Dave Hansen)
 - Only have start gap on shadow stack (Andy Luto)
 - Create stack_guard_start_gap() to not duplicate code
   in an arch version of vm_start_gap() (Dave Hansen)
 - Improve commit log partly with verbiage from (Dave Hansen)

Yu-cheng v25:
 - Move SHADOW_STACK_GUARD_GAP to arch/x86/mm/mmap.c.

Yu-cheng v24:
 - Instead changing vm_*_gap(), create x86-specific versions.

 arch/x86/mm/mmap.c | 23 +++++++++++++++++++++++
 include/linux/mm.h | 11 ++++++-----
 mm/mmap.c          |  7 +++++++
 3 files changed, 36 insertions(+), 5 deletions(-)

Comments

Kees Cook Oct. 3, 2022, 6:30 p.m. UTC | #1 
On Thu, Sep 29, 2022 at 03:29:15PM -0700, Rick Edgecombe wrote:
> [...]
> +unsigned long stack_guard_start_gap(struct vm_area_struct *vma)
> +{
> +	if (vma->vm_flags & VM_GROWSDOWN)
> +		return stack_guard_gap;
> +
> +	/*
> +	 * Shadow stack pointer is moved by CALL, RET, and INCSSP(Q/D).
> +	 * INCSSPQ moves shadow stack pointer up to 255 * 8 = ~2 KB
> +	 * (~1KB for INCSSPD) and touches the first and the last element
> +	 * in the range, which triggers a page fault if the range is not
> +	 * in a shadow stack. Because of this, creating 4-KB guard pages
> +	 * around a shadow stack prevents these instructions from going
> +	 * beyond.
> +	 *
> +	 * Creation of VM_SHADOW_STACK is tightly controlled, so a vma
> +	 * can't be both VM_GROWSDOWN and VM_SHADOW_STACK
> +	 */

Thank you for the details on how the size choice is made here! :)

> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index fef14ab3abcb..09458e77bf52 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -2775,15 +2775,16 @@ struct vm_area_struct *vma_lookup(struct mm_struct *mm, unsigned long addr)
>  	return vma;
>  }
>  
> +unsigned long stack_guard_start_gap(struct vm_area_struct *vma);
> +
>  static inline unsigned long vm_start_gap(struct vm_area_struct *vma)
>  {
> +	unsigned long gap = stack_guard_start_gap(vma);
>  	unsigned long vm_start = vma->vm_start;
>  
> -	if (vma->vm_flags & VM_GROWSDOWN) {
> -		vm_start -= stack_guard_gap;
> -		if (vm_start > vma->vm_start)
> -			vm_start = 0;
> -	}
> +	vm_start -= gap;
> +	if (vm_start > vma->vm_start)
> +		vm_start = 0;
>  	return vm_start;
>  }
>  
> diff --git a/mm/mmap.c b/mm/mmap.c
> index 9d780f415be3..f0d2e9143bd0 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -247,6 +247,13 @@ SYSCALL_DEFINE1(brk, unsigned long, brk)
>  	return origbrk;
>  }
>  

I feel like something could be done with this definitions to make them
inline, instead of __weak:

#ifndef stack_guard_start_gap
> +unsigned long __weak stack_guard_start_gap(struct vm_area_struct *vma)
> +{
> +	if (vma->vm_flags & VM_GROWSDOWN)
> +		return stack_guard_gap;
> +	return 0;
> +}
#endif

And then move the x86 stack_guard_start_gap to a header?

It's not exactly fast-path, but it feels a little weird. Regardlesss:

Reviewed-by: Kees Cook <keescook@chromium.org>
Andrew Cooper Oct. 5, 2022, 2:30 a.m. UTC | #2 
On 03/10/2022 19:30, Kees Cook wrote:
> On Thu, Sep 29, 2022 at 03:29:15PM -0700, Rick Edgecombe wrote:
>> [...]
>> +unsigned long stack_guard_start_gap(struct vm_area_struct *vma)
>> +{
>> +	if (vma->vm_flags & VM_GROWSDOWN)
>> +		return stack_guard_gap;
>> +
>> +	/*
>> +	 * Shadow stack pointer is moved by CALL, RET, and INCSSP(Q/D).
>> +	 * INCSSPQ moves shadow stack pointer up to 255 * 8 = ~2 KB
>> +	 * (~1KB for INCSSPD) and touches the first and the last element
>> +	 * in the range, which triggers a page fault if the range is not
>> +	 * in a shadow stack. Because of this, creating 4-KB guard pages
>> +	 * around a shadow stack prevents these instructions from going
>> +	 * beyond.
>> +	 *
>> +	 * Creation of VM_SHADOW_STACK is tightly controlled, so a vma
>> +	 * can't be both VM_GROWSDOWN and VM_SHADOW_STACK
>> +	 */
> Thank you for the details on how the size choice is made here! :)

(In case anyone is hankering for some premature optimisation...)

You don't actually need a hole to create a guard.  Any mapping of type
!= shstk will do.

If you've got a load of threads, you can tightly pack stack / shstk /
stack / shstk with no holes, and they each act as each other guard pages.

~Andrew
Florian Weimer Oct. 10, 2022, 12:33 p.m. UTC | #3 
* Andrew Cooper:

> You don't actually need a hole to create a guard.  Any mapping of type
> != shstk will do.
>
> If you've got a load of threads, you can tightly pack stack / shstk /
> stack / shstk with no holes, and they each act as each other guard pages.

Can userspace read the shadow stack directly?  Writing is obviously
blocked, but reading?

GCC's stack-clash probing uses OR instructions, so it would be fine with
a readable mapping.  POSIX does not appear to require PROT_NONE mappings
for the stack guard region, either.  However, the
pthread_attr_setguardsize manual page pretty clearly says that it's got
to be unreadable and unwriteable.  Hence my question.

Thanks,
Florian
Andrew Cooper Oct. 10, 2022, 1:32 p.m. UTC | #4 
On 10/10/2022 13:33, Florian Weimer wrote:
> * Andrew Cooper:
>
>> You don't actually need a hole to create a guard.  Any mapping of type
>> != shstk will do.
>>
>> If you've got a load of threads, you can tightly pack stack / shstk /
>> stack / shstk with no holes, and they each act as each other guard pages.
> Can userspace read the shadow stack directly?  Writing is obviously
> blocked, but reading?

Yes - regular reads are permitted to shstk memory.

It's actually a great way to get backtraces with no extra metadata needed.

> GCC's stack-clash probing uses OR instructions, so it would be fine with
> a readable mapping.

It's `or $0, (%rsp)` which is a read/modify/write and will fault when
hitting a shstk mapping.

> POSIX does not appear to require PROT_NONE mappings
> for the stack guard region, either.  However, the
> pthread_attr_setguardsize manual page pretty clearly says that it's got
> to be unreadable and unwriteable.  Hence my question.

Hmm.  If that's what the manuals say, then fine.

But honestly, you don't get very far at all without faulting on a
read-only stack.

~Andrew
Florian Weimer Oct. 10, 2022, 1:40 p.m. UTC | #5 
* Andrew Cooper:

> On 10/10/2022 13:33, Florian Weimer wrote:
>> * Andrew Cooper:
>>
>>> You don't actually need a hole to create a guard.  Any mapping of type
>>> != shstk will do.
>>>
>>> If you've got a load of threads, you can tightly pack stack / shstk /
>>> stack / shstk with no holes, and they each act as each other guard pages.
>> Can userspace read the shadow stack directly?  Writing is obviously
>> blocked, but reading?
>
> Yes - regular reads are permitted to shstk memory.
>
> It's actually a great way to get backtraces with no extra metadata
> needed.

Indeed, I hope shadow stacks can be used to put the discussion around
frame pointers to a rest, at least when it comes to profiling. 8-)

>> POSIX does not appear to require PROT_NONE mappings
>> for the stack guard region, either.  However, the
>> pthread_attr_setguardsize manual page pretty clearly says that it's got
>> to be unreadable and unwriteable.  Hence my question.
>
> Hmm.  If that's what the manuals say, then fine.
>
> But honestly, you don't get very far at all without faulting on a
> read-only stack.

I guess we can update the manual page proactively.  It does look like a
tempting optimization.

Thanks,
Florian
Andrew Cooper Oct. 10, 2022, 1:56 p.m. UTC | #6 
On 10/10/2022 14:40, Florian Weimer wrote:
> * Andrew Cooper:
>
>>> POSIX does not appear to require PROT_NONE mappings
>>> for the stack guard region, either.  However, the
>>> pthread_attr_setguardsize manual page pretty clearly says that it's got
>>> to be unreadable and unwriteable.  Hence my question.
>> Hmm.  If that's what the manuals say, then fine.
>>
>> But honestly, you don't get very far at all without faulting on a
>> read-only stack.
> I guess we can update the manual page proactively.  It does look like a
> tempting optimization.

Here's one I prepared earlier, discussing getting supervisor shadow
stacks working in Xen.

http://xenbits.xen.org/people/andrewcoop/Xen-CET-SS.pdf

This optimisation turned out to be very helpful by being able to put the
shadow stacks in what were previously the guard holes, meaning we didn't
actually need to allocate any more memory for the stacks.

~Andrew
diff mbox series

Patch

diff --git a/arch/x86/mm/mmap.c b/arch/x86/mm/mmap.c
index f3f52c5e2fd6..b0427bd2da30 100644
--- a/arch/x86/mm/mmap.c
+++ b/arch/x86/mm/mmap.c
@@ -250,3 +250,26 @@  bool pfn_modify_allowed(unsigned long pfn, pgprot_t prot)
 		return false;
 	return true;
 }
+
+unsigned long stack_guard_start_gap(struct vm_area_struct *vma)
+{
+	if (vma->vm_flags & VM_GROWSDOWN)
+		return stack_guard_gap;
+
+	/*
+	 * Shadow stack pointer is moved by CALL, RET, and INCSSP(Q/D).
+	 * INCSSPQ moves shadow stack pointer up to 255 * 8 = ~2 KB
+	 * (~1KB for INCSSPD) and touches the first and the last element
+	 * in the range, which triggers a page fault if the range is not
+	 * in a shadow stack. Because of this, creating 4-KB guard pages
+	 * around a shadow stack prevents these instructions from going
+	 * beyond.
+	 *
+	 * Creation of VM_SHADOW_STACK is tightly controlled, so a vma
+	 * can't be both VM_GROWSDOWN and VM_SHADOW_STACK
+	 */
+	if (vma->vm_flags & VM_SHADOW_STACK)
+		return PAGE_SIZE;
+
+	return 0;
+}
diff --git a/include/linux/mm.h b/include/linux/mm.h
index fef14ab3abcb..09458e77bf52 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -2775,15 +2775,16 @@  struct vm_area_struct *vma_lookup(struct mm_struct *mm, unsigned long addr)
 	return vma;
 }
 
+unsigned long stack_guard_start_gap(struct vm_area_struct *vma);
+
 static inline unsigned long vm_start_gap(struct vm_area_struct *vma)
 {
+	unsigned long gap = stack_guard_start_gap(vma);
 	unsigned long vm_start = vma->vm_start;
 
-	if (vma->vm_flags & VM_GROWSDOWN) {
-		vm_start -= stack_guard_gap;
-		if (vm_start > vma->vm_start)
-			vm_start = 0;
-	}
+	vm_start -= gap;
+	if (vm_start > vma->vm_start)
+		vm_start = 0;
 	return vm_start;
 }
 
diff --git a/mm/mmap.c b/mm/mmap.c
index 9d780f415be3..f0d2e9143bd0 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -247,6 +247,13 @@  SYSCALL_DEFINE1(brk, unsigned long, brk)
 	return origbrk;
 }
 
+unsigned long __weak stack_guard_start_gap(struct vm_area_struct *vma)
+{
+	if (vma->vm_flags & VM_GROWSDOWN)
+		return stack_guard_gap;
+	return 0;
+}
+
 static inline unsigned long vma_compute_gap(struct vm_area_struct *vma)
 {
 	unsigned long gap, prev_end;