This implements arch_within_stack_frames() for MIPS that validates if an
object is wholly contained by a kernel stack frame.
With CONFIG_HARDENED_USERCOPY enabled, MIPS now passes the LKDTM tests
USERCOPY_STACK_FRAME_TO, USERCOPY_STACK_FRAME_FROM and
USERCOPY_STACK_BEYOND on a Creator Ci40.
Since the MIPS kernel does not use frame pointers, we re-use the MIPS
kernels stack frame unwinder which uses instruction inspection to deduce
the stack frame size. As such it introduces a larger performance penalty
than on arches which use the frame pointer.
On qemu, before this patch, hackbench gives:
Running with 10*40 (== 400) tasks.
Time: 5.484
Running with 10*40 (== 400) tasks.
Time: 4.039
Running with 10*40 (== 400) tasks.
Time: 3.908
Running with 10*40 (== 400) tasks.
Time: 3.955
Running with 10*40 (== 400) tasks.
Time: 4.185
Running with 10*40 (== 400) tasks.
Time: 4.497
Running with 10*40 (== 400) tasks.
Time: 3.980
Running with 10*40 (== 400) tasks.
Time: 4.078
Running with 10*40 (== 400) tasks.
Time: 4.219
Running with 10*40 (== 400) tasks.
Time: 4.026
Giving an average of 4.2371
With this patch, hackbench gives:
Running with 10*40 (== 400) tasks.
Time: 5.671
Running with 10*40 (== 400) tasks.
Time: 4.282
Running with 10*40 (== 400) tasks.
Time: 4.101
Running with 10*40 (== 400) tasks.
Time: 4.040
Running with 10*40 (== 400) tasks.
Time: 4.683
Running with 10*40 (== 400) tasks.
Time: 4.387
Running with 10*40 (== 400) tasks.
Time: 4.289
Running with 10*40 (== 400) tasks.
Time: 4.027
Running with 10*40 (== 400) tasks.
Time: 4.048
Running with 10*40 (== 400) tasks.
Time: 4.079
Giving an average of 4.3607
This indicates an additional 3% overhead for inspecting the kernel stack
when CONFIG_HARDENED_USERCOPY is enabled.
This patch is based on Linux v4.13-rc4, and for correct operation on
microMIPS depends on my series "MIPS: Further microMIPS stack unwinding
fixes"
Signed-off-by: Matt Redfearn <[email protected]>
Reviewed-by: James Hogan <[email protected]>
---
arch/mips/Kconfig | 1 +
arch/mips/include/asm/thread_info.h | 74 +++++++++++++++++++++++++++++++++++++
2 files changed, 75 insertions(+)
diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig
index 8dd20358464f..6cbf2d525c8d 100644
--- a/arch/mips/Kconfig
+++ b/arch/mips/Kconfig
@@ -35,6 +35,7 @@ config MIPS
select HAVE_ARCH_SECCOMP_FILTER
select HAVE_ARCH_TRACEHOOK
select HAVE_ARCH_TRANSPARENT_HUGEPAGE if CPU_SUPPORTS_HUGEPAGES && 64BIT
+ select HAVE_ARCH_WITHIN_STACK_FRAMES if KALLSYMS
select HAVE_CBPF_JIT if (!64BIT && !CPU_MICROMIPS)
select HAVE_EBPF_JIT if (64BIT && !CPU_MICROMIPS)
select HAVE_CC_STACKPROTECTOR
diff --git a/arch/mips/include/asm/thread_info.h b/arch/mips/include/asm/thread_info.h
index b439e512792b..931652460393 100644
--- a/arch/mips/include/asm/thread_info.h
+++ b/arch/mips/include/asm/thread_info.h
@@ -14,6 +14,80 @@
#include <asm/processor.h>
+#ifdef CONFIG_HAVE_ARCH_WITHIN_STACK_FRAMES
+
+/*
+ * Walks up the stack frames to make sure that the specified object is
+ * entirely contained by a single stack frame.
+ *
+ * Returns:
+ * GOOD_FRAME if within a frame
+ * BAD_STACK if placed across a frame boundary (or outside stack)
+ * NOT_STACK unable to determine
+ */
+static inline int arch_within_stack_frames(const void *const stack,
+ const void *const stackend,
+ const void *obj, unsigned long len)
+{
+ /* Avoid header recursion by just declaring this here */
+ extern unsigned long unwind_stack_by_address(
+ unsigned long stack_page,
+ unsigned long *sp,
+ unsigned long pc,
+ unsigned long *ra);
+ unsigned long sp, lastsp, ra, pc;
+ int skip_frames;
+
+ /* Get this frame's details */
+ sp = (unsigned long)__builtin_frame_address(0);
+ pc = (unsigned long)current_text_addr();
+
+ /*
+ * Skip initial frames to get back the function requesting the copy.
+ * Unwind the frames of:
+ * arch_within_stack_frames (inlined into check_stack_object)
+ * __check_object_size
+ * This leaves sp & pc in the frame associated with
+ * copy_{to,from}_user() (inlined into do_usercopy_stack)
+ */
+ for (skip_frames = 0; skip_frames < 2; skip_frames++) {
+ pc = unwind_stack_by_address((unsigned long)stack, &sp, pc, &ra);
+ if (!pc)
+ return BAD_STACK;
+ }
+
+ if ((unsigned long)obj < sp) {
+ /* obj is not in the frame of the requestor or it's callers */
+ return BAD_STACK;
+ }
+
+ /*
+ * low ---------------------------------------> high
+ * [local vars][saved regs][ra][local vars']
+ * ^ ^
+ * lastsp sp
+ * ^----------------------^
+ * allow copies only within here
+ */
+ do {
+ lastsp = sp;
+ pc = unwind_stack_by_address((unsigned long)stack, &sp, pc, &ra);
+ if ((((unsigned long)obj) >= lastsp) &&
+ (((unsigned long)obj + len) <= (sp - sizeof(void *)))) {
+ /* obj is entirely within this stack frame */
+ return GOOD_FRAME;
+ }
+ } while (pc);
+
+ /*
+ * We can't unwind any further. If we haven't found the object entirely
+ * within one of our callers frames, it must be a bad object.
+ */
+ return BAD_STACK;
+}
+
+#endif /* CONFIG_HAVE_ARCH_WITHIN_STACK_FRAMES */
+
/*
* low level task data that entry.S needs immediate access to
* - this struct should fit entirely inside of one cache line
--
2.7.4
On Tue, Aug 8, 2017 at 5:23 AM, Matt Redfearn <[email protected]> wrote:
> This implements arch_within_stack_frames() for MIPS that validates if an
> object is wholly contained by a kernel stack frame.
>
> With CONFIG_HARDENED_USERCOPY enabled, MIPS now passes the LKDTM tests
> USERCOPY_STACK_FRAME_TO, USERCOPY_STACK_FRAME_FROM and
> USERCOPY_STACK_BEYOND on a Creator Ci40.
>
> Since the MIPS kernel does not use frame pointers, we re-use the MIPS
> kernels stack frame unwinder which uses instruction inspection to deduce
> the stack frame size. As such it introduces a larger performance penalty
> than on arches which use the frame pointer.
Hmm, given x86's plans to drop the frame pointer, I wonder if the
inter-frame checking code should be gated by a CONFIG. This (3%) is a
rather high performance hit to take for a relatively small protection
(it's mainly about catching too-large-reads, since most
too-large-writes will be caught by the stack canary).
What do you think?
-Kees
>
> On qemu, before this patch, hackbench gives:
> Running with 10*40 (== 400) tasks.
> Time: 5.484
> Running with 10*40 (== 400) tasks.
> Time: 4.039
> Running with 10*40 (== 400) tasks.
> Time: 3.908
> Running with 10*40 (== 400) tasks.
> Time: 3.955
> Running with 10*40 (== 400) tasks.
> Time: 4.185
> Running with 10*40 (== 400) tasks.
> Time: 4.497
> Running with 10*40 (== 400) tasks.
> Time: 3.980
> Running with 10*40 (== 400) tasks.
> Time: 4.078
> Running with 10*40 (== 400) tasks.
> Time: 4.219
> Running with 10*40 (== 400) tasks.
> Time: 4.026
>
> Giving an average of 4.2371
>
> With this patch, hackbench gives:
> Running with 10*40 (== 400) tasks.
> Time: 5.671
> Running with 10*40 (== 400) tasks.
> Time: 4.282
> Running with 10*40 (== 400) tasks.
> Time: 4.101
> Running with 10*40 (== 400) tasks.
> Time: 4.040
> Running with 10*40 (== 400) tasks.
> Time: 4.683
> Running with 10*40 (== 400) tasks.
> Time: 4.387
> Running with 10*40 (== 400) tasks.
> Time: 4.289
> Running with 10*40 (== 400) tasks.
> Time: 4.027
> Running with 10*40 (== 400) tasks.
> Time: 4.048
> Running with 10*40 (== 400) tasks.
> Time: 4.079
>
> Giving an average of 4.3607
>
> This indicates an additional 3% overhead for inspecting the kernel stack
> when CONFIG_HARDENED_USERCOPY is enabled.
>
> This patch is based on Linux v4.13-rc4, and for correct operation on
> microMIPS depends on my series "MIPS: Further microMIPS stack unwinding
> fixes"
>
> Signed-off-by: Matt Redfearn <[email protected]>
> Reviewed-by: James Hogan <[email protected]>
> ---
>
> arch/mips/Kconfig | 1 +
> arch/mips/include/asm/thread_info.h | 74 +++++++++++++++++++++++++++++++++++++
> 2 files changed, 75 insertions(+)
>
> diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig
> index 8dd20358464f..6cbf2d525c8d 100644
> --- a/arch/mips/Kconfig
> +++ b/arch/mips/Kconfig
> @@ -35,6 +35,7 @@ config MIPS
> select HAVE_ARCH_SECCOMP_FILTER
> select HAVE_ARCH_TRACEHOOK
> select HAVE_ARCH_TRANSPARENT_HUGEPAGE if CPU_SUPPORTS_HUGEPAGES && 64BIT
> + select HAVE_ARCH_WITHIN_STACK_FRAMES if KALLSYMS
> select HAVE_CBPF_JIT if (!64BIT && !CPU_MICROMIPS)
> select HAVE_EBPF_JIT if (64BIT && !CPU_MICROMIPS)
> select HAVE_CC_STACKPROTECTOR
> diff --git a/arch/mips/include/asm/thread_info.h b/arch/mips/include/asm/thread_info.h
> index b439e512792b..931652460393 100644
> --- a/arch/mips/include/asm/thread_info.h
> +++ b/arch/mips/include/asm/thread_info.h
> @@ -14,6 +14,80 @@
>
> #include <asm/processor.h>
>
> +#ifdef CONFIG_HAVE_ARCH_WITHIN_STACK_FRAMES
> +
> +/*
> + * Walks up the stack frames to make sure that the specified object is
> + * entirely contained by a single stack frame.
> + *
> + * Returns:
> + * GOOD_FRAME if within a frame
> + * BAD_STACK if placed across a frame boundary (or outside stack)
> + * NOT_STACK unable to determine
> + */
> +static inline int arch_within_stack_frames(const void *const stack,
> + const void *const stackend,
> + const void *obj, unsigned long len)
> +{
> + /* Avoid header recursion by just declaring this here */
> + extern unsigned long unwind_stack_by_address(
> + unsigned long stack_page,
> + unsigned long *sp,
> + unsigned long pc,
> + unsigned long *ra);
> + unsigned long sp, lastsp, ra, pc;
> + int skip_frames;
> +
> + /* Get this frame's details */
> + sp = (unsigned long)__builtin_frame_address(0);
> + pc = (unsigned long)current_text_addr();
> +
> + /*
> + * Skip initial frames to get back the function requesting the copy.
> + * Unwind the frames of:
> + * arch_within_stack_frames (inlined into check_stack_object)
> + * __check_object_size
> + * This leaves sp & pc in the frame associated with
> + * copy_{to,from}_user() (inlined into do_usercopy_stack)
> + */
> + for (skip_frames = 0; skip_frames < 2; skip_frames++) {
> + pc = unwind_stack_by_address((unsigned long)stack, &sp, pc, &ra);
> + if (!pc)
> + return BAD_STACK;
> + }
> +
> + if ((unsigned long)obj < sp) {
> + /* obj is not in the frame of the requestor or it's callers */
> + return BAD_STACK;
> + }
> +
> + /*
> + * low ---------------------------------------> high
> + * [local vars][saved regs][ra][local vars']
> + * ^ ^
> + * lastsp sp
> + * ^----------------------^
> + * allow copies only within here
> + */
> + do {
> + lastsp = sp;
> + pc = unwind_stack_by_address((unsigned long)stack, &sp, pc, &ra);
> + if ((((unsigned long)obj) >= lastsp) &&
> + (((unsigned long)obj + len) <= (sp - sizeof(void *)))) {
> + /* obj is entirely within this stack frame */
> + return GOOD_FRAME;
> + }
> + } while (pc);
> +
> + /*
> + * We can't unwind any further. If we haven't found the object entirely
> + * within one of our callers frames, it must be a bad object.
> + */
> + return BAD_STACK;
> +}
> +
> +#endif /* CONFIG_HAVE_ARCH_WITHIN_STACK_FRAMES */
> +
> /*
> * low level task data that entry.S needs immediate access to
> * - this struct should fit entirely inside of one cache line
> --
> 2.7.4
>
--
Kees Cook
Pixel Security
Hi Kees,
On 08/08/17 20:11, Kees Cook wrote:
> On Tue, Aug 8, 2017 at 5:23 AM, Matt Redfearn <[email protected]> wrote:
>> This implements arch_within_stack_frames() for MIPS that validates if an
>> object is wholly contained by a kernel stack frame.
>>
>> With CONFIG_HARDENED_USERCOPY enabled, MIPS now passes the LKDTM tests
>> USERCOPY_STACK_FRAME_TO, USERCOPY_STACK_FRAME_FROM and
>> USERCOPY_STACK_BEYOND on a Creator Ci40.
>>
>> Since the MIPS kernel does not use frame pointers, we re-use the MIPS
>> kernels stack frame unwinder which uses instruction inspection to deduce
>> the stack frame size. As such it introduces a larger performance penalty
>> than on arches which use the frame pointer.
> Hmm, given x86's plans to drop the frame pointer, I wonder if the
> inter-frame checking code should be gated by a CONFIG. This (3%) is a
> rather high performance hit to take for a relatively small protection
> (it's mainly about catching too-large-reads, since most
> too-large-writes will be caught by the stack canary).
>
> What do you think?
If x86 is going to move to a more expensive stack unwinding method than
the frame pointer then I guess it may end up seeing a similar
performance hit to what we see on MIPS. In that case it might make sense
to add a CONFIG for this such that only those who wish to make the trade
off of performance for the added protection need enable it.
Thanks,
Matt
>
> -Kees
>
>> On qemu, before this patch, hackbench gives:
>> Running with 10*40 (== 400) tasks.
>> Time: 5.484
>> Running with 10*40 (== 400) tasks.
>> Time: 4.039
>> Running with 10*40 (== 400) tasks.
>> Time: 3.908
>> Running with 10*40 (== 400) tasks.
>> Time: 3.955
>> Running with 10*40 (== 400) tasks.
>> Time: 4.185
>> Running with 10*40 (== 400) tasks.
>> Time: 4.497
>> Running with 10*40 (== 400) tasks.
>> Time: 3.980
>> Running with 10*40 (== 400) tasks.
>> Time: 4.078
>> Running with 10*40 (== 400) tasks.
>> Time: 4.219
>> Running with 10*40 (== 400) tasks.
>> Time: 4.026
>>
>> Giving an average of 4.2371
>>
>> With this patch, hackbench gives:
>> Running with 10*40 (== 400) tasks.
>> Time: 5.671
>> Running with 10*40 (== 400) tasks.
>> Time: 4.282
>> Running with 10*40 (== 400) tasks.
>> Time: 4.101
>> Running with 10*40 (== 400) tasks.
>> Time: 4.040
>> Running with 10*40 (== 400) tasks.
>> Time: 4.683
>> Running with 10*40 (== 400) tasks.
>> Time: 4.387
>> Running with 10*40 (== 400) tasks.
>> Time: 4.289
>> Running with 10*40 (== 400) tasks.
>> Time: 4.027
>> Running with 10*40 (== 400) tasks.
>> Time: 4.048
>> Running with 10*40 (== 400) tasks.
>> Time: 4.079
>>
>> Giving an average of 4.3607
>>
>> This indicates an additional 3% overhead for inspecting the kernel stack
>> when CONFIG_HARDENED_USERCOPY is enabled.
>>
>> This patch is based on Linux v4.13-rc4, and for correct operation on
>> microMIPS depends on my series "MIPS: Further microMIPS stack unwinding
>> fixes"
>>
>> Signed-off-by: Matt Redfearn <[email protected]>
>> Reviewed-by: James Hogan <[email protected]>
>> ---
>>
>> arch/mips/Kconfig | 1 +
>> arch/mips/include/asm/thread_info.h | 74 +++++++++++++++++++++++++++++++++++++
>> 2 files changed, 75 insertions(+)
>>
>> diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig
>> index 8dd20358464f..6cbf2d525c8d 100644
>> --- a/arch/mips/Kconfig
>> +++ b/arch/mips/Kconfig
>> @@ -35,6 +35,7 @@ config MIPS
>> select HAVE_ARCH_SECCOMP_FILTER
>> select HAVE_ARCH_TRACEHOOK
>> select HAVE_ARCH_TRANSPARENT_HUGEPAGE if CPU_SUPPORTS_HUGEPAGES && 64BIT
>> + select HAVE_ARCH_WITHIN_STACK_FRAMES if KALLSYMS
>> select HAVE_CBPF_JIT if (!64BIT && !CPU_MICROMIPS)
>> select HAVE_EBPF_JIT if (64BIT && !CPU_MICROMIPS)
>> select HAVE_CC_STACKPROTECTOR
>> diff --git a/arch/mips/include/asm/thread_info.h b/arch/mips/include/asm/thread_info.h
>> index b439e512792b..931652460393 100644
>> --- a/arch/mips/include/asm/thread_info.h
>> +++ b/arch/mips/include/asm/thread_info.h
>> @@ -14,6 +14,80 @@
>>
>> #include <asm/processor.h>
>>
>> +#ifdef CONFIG_HAVE_ARCH_WITHIN_STACK_FRAMES
>> +
>> +/*
>> + * Walks up the stack frames to make sure that the specified object is
>> + * entirely contained by a single stack frame.
>> + *
>> + * Returns:
>> + * GOOD_FRAME if within a frame
>> + * BAD_STACK if placed across a frame boundary (or outside stack)
>> + * NOT_STACK unable to determine
>> + */
>> +static inline int arch_within_stack_frames(const void *const stack,
>> + const void *const stackend,
>> + const void *obj, unsigned long len)
>> +{
>> + /* Avoid header recursion by just declaring this here */
>> + extern unsigned long unwind_stack_by_address(
>> + unsigned long stack_page,
>> + unsigned long *sp,
>> + unsigned long pc,
>> + unsigned long *ra);
>> + unsigned long sp, lastsp, ra, pc;
>> + int skip_frames;
>> +
>> + /* Get this frame's details */
>> + sp = (unsigned long)__builtin_frame_address(0);
>> + pc = (unsigned long)current_text_addr();
>> +
>> + /*
>> + * Skip initial frames to get back the function requesting the copy.
>> + * Unwind the frames of:
>> + * arch_within_stack_frames (inlined into check_stack_object)
>> + * __check_object_size
>> + * This leaves sp & pc in the frame associated with
>> + * copy_{to,from}_user() (inlined into do_usercopy_stack)
>> + */
>> + for (skip_frames = 0; skip_frames < 2; skip_frames++) {
>> + pc = unwind_stack_by_address((unsigned long)stack, &sp, pc, &ra);
>> + if (!pc)
>> + return BAD_STACK;
>> + }
>> +
>> + if ((unsigned long)obj < sp) {
>> + /* obj is not in the frame of the requestor or it's callers */
>> + return BAD_STACK;
>> + }
>> +
>> + /*
>> + * low ---------------------------------------> high
>> + * [local vars][saved regs][ra][local vars']
>> + * ^ ^
>> + * lastsp sp
>> + * ^----------------------^
>> + * allow copies only within here
>> + */
>> + do {
>> + lastsp = sp;
>> + pc = unwind_stack_by_address((unsigned long)stack, &sp, pc, &ra);
>> + if ((((unsigned long)obj) >= lastsp) &&
>> + (((unsigned long)obj + len) <= (sp - sizeof(void *)))) {
>> + /* obj is entirely within this stack frame */
>> + return GOOD_FRAME;
>> + }
>> + } while (pc);
>> +
>> + /*
>> + * We can't unwind any further. If we haven't found the object entirely
>> + * within one of our callers frames, it must be a bad object.
>> + */
>> + return BAD_STACK;
>> +}
>> +
>> +#endif /* CONFIG_HAVE_ARCH_WITHIN_STACK_FRAMES */
>> +
>> /*
>> * low level task data that entry.S needs immediate access to
>> * - this struct should fit entirely inside of one cache line
>> --
>> 2.7.4
>>
>
>
On Thu, Aug 10, 2017 at 1:24 AM, Matt Redfearn <[email protected]> wrote:
> Hi Kees,
>
>
> On 08/08/17 20:11, Kees Cook wrote:
>>
>> On Tue, Aug 8, 2017 at 5:23 AM, Matt Redfearn <[email protected]>
>> wrote:
>>>
>>> This implements arch_within_stack_frames() for MIPS that validates if an
>>> object is wholly contained by a kernel stack frame.
>>>
>>> With CONFIG_HARDENED_USERCOPY enabled, MIPS now passes the LKDTM tests
>>> USERCOPY_STACK_FRAME_TO, USERCOPY_STACK_FRAME_FROM and
>>> USERCOPY_STACK_BEYOND on a Creator Ci40.
>>>
>>> Since the MIPS kernel does not use frame pointers, we re-use the MIPS
>>> kernels stack frame unwinder which uses instruction inspection to deduce
>>> the stack frame size. As such it introduces a larger performance penalty
>>> than on arches which use the frame pointer.
>>
>> Hmm, given x86's plans to drop the frame pointer, I wonder if the
>> inter-frame checking code should be gated by a CONFIG. This (3%) is a
>> rather high performance hit to take for a relatively small protection
>> (it's mainly about catching too-large-reads, since most
>> too-large-writes will be caught by the stack canary).
>>
>> What do you think?
>
>
> If x86 is going to move to a more expensive stack unwinding method than the
> frame pointer then I guess it may end up seeing a similar performance hit to
> what we see on MIPS. In that case it might make sense to add a CONFIG for
> this such that only those who wish to make the trade off of performance for
> the added protection need enable it.
Sounds good. Can you send a v2 that adds a CONFIG, maybe something
like CONFIG_HARDENED_USERCOPY_UNWINDER with a description of the
trade-offs? Then x86 can do this too when it drops frame pointers.
Thanks!
-Kees
--
Kees Cook
Pixel Security