In order to do kernel builds with the bounds checker individually
available, introduce CONFIG_UBSAN_BOUNDS, with the remaining options
under CONFIG_UBSAN_MISC.
For example, using this, we can start to expand the coverage syzkaller is
providing. Right now, all of UBSan is disabled for syzbot builds because
taken as a whole, it is too noisy. This will let us focus on one feature
at a time.
For the bounds checker specifically, this provides a mechanism to
eliminate an entire class of array overflows with close to zero
performance overhead (I cannot measure a difference). In my (mostly)
defconfig, enabling bounds checking adds ~4200 checks to the kernel.
Performance changes are in the noise, likely due to the branch predictors
optimizing for the non-fail path.
Some notes on the bounds checker:
- it does not instrument {mem,str}*()-family functions, it only
instruments direct indexed accesses (e.g. "foo[i]"). Dealing with
the {mem,str}*()-family functions is a work-in-progress around
CONFIG_FORTIFY_SOURCE[1].
- it ignores flexible array members, including the very old single
byte (e.g. "int foo[1];") declarations. (Note that GCC's
implementation appears to ignore _all_ trailing arrays, but Clang only
ignores empty, 0, and 1 byte arrays[2].)
[1] https://github.com/KSPP/linux/issues/6
[2] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92589
Suggested-by: Elena Petrova <[email protected]>
Signed-off-by: Kees Cook <[email protected]>
---
lib/Kconfig.ubsan | 19 +++++++++++++++++++
scripts/Makefile.ubsan | 7 ++++++-
2 files changed, 25 insertions(+), 1 deletion(-)
diff --git a/lib/Kconfig.ubsan b/lib/Kconfig.ubsan
index d69e8b21ebae..f5ed2dceef30 100644
--- a/lib/Kconfig.ubsan
+++ b/lib/Kconfig.ubsan
@@ -22,6 +22,25 @@ config UBSAN_TRAP
can just issue a trap. This reduces the kernel size overhead but
turns all warnings into full thread-killing exceptions.
+config UBSAN_BOUNDS
+ bool "Perform array bounds checking"
+ depends on UBSAN
+ default UBSAN
+ help
+ This option enables detection of direct out of bounds array
+ accesses, where the array size is known at compile time. Note
+ that this does not protect character array overflows due to
+ bad calls to the {str,mem}*cpy() family of functions.
+
+config UBSAN_MISC
+ bool "Enable all other Undefined Behavior sanity checks"
+ depends on UBSAN
+ default UBSAN
+ help
+ This option enables all sanity checks that don't have their
+ own Kconfig options. Disable this if you only want to have
+ individually selected checks.
+
config UBSAN_SANITIZE_ALL
bool "Enable instrumentation for the entire kernel"
depends on UBSAN
diff --git a/scripts/Makefile.ubsan b/scripts/Makefile.ubsan
index 668a91510bfe..5b15bc425ec9 100644
--- a/scripts/Makefile.ubsan
+++ b/scripts/Makefile.ubsan
@@ -5,14 +5,19 @@ ifdef CONFIG_UBSAN_ALIGNMENT
CFLAGS_UBSAN += $(call cc-option, -fsanitize=alignment)
endif
+ifdef CONFIG_UBSAN_BOUNDS
+ CFLAGS_UBSAN += $(call cc-option, -fsanitize=bounds)
+endif
+
+ifdef CONFIG_UBSAN_MISC
CFLAGS_UBSAN += $(call cc-option, -fsanitize=shift)
CFLAGS_UBSAN += $(call cc-option, -fsanitize=integer-divide-by-zero)
CFLAGS_UBSAN += $(call cc-option, -fsanitize=unreachable)
CFLAGS_UBSAN += $(call cc-option, -fsanitize=signed-integer-overflow)
- CFLAGS_UBSAN += $(call cc-option, -fsanitize=bounds)
CFLAGS_UBSAN += $(call cc-option, -fsanitize=object-size)
CFLAGS_UBSAN += $(call cc-option, -fsanitize=bool)
CFLAGS_UBSAN += $(call cc-option, -fsanitize=enum)
+endif
ifdef CONFIG_UBSAN_TRAP
CFLAGS_UBSAN += $(call cc-option, -fsanitize-undefined-trap-on-error)
--
2.17.1
On 11/20/19 4:06 AM, Kees Cook wrote:
> In order to do kernel builds with the bounds checker individually
> available, introduce CONFIG_UBSAN_BOUNDS, with the remaining options
> under CONFIG_UBSAN_MISC.
>
> For example, using this, we can start to expand the coverage syzkaller is
> providing. Right now, all of UBSan is disabled for syzbot builds because
> taken as a whole, it is too noisy. This will let us focus on one feature
> at a time.
>
> For the bounds checker specifically, this provides a mechanism to
> eliminate an entire class of array overflows with close to zero
> performance overhead (I cannot measure a difference). In my (mostly)
> defconfig, enabling bounds checking adds ~4200 checks to the kernel.
> Performance changes are in the noise, likely due to the branch predictors
> optimizing for the non-fail path.
>
> Some notes on the bounds checker:
>
> - it does not instrument {mem,str}*()-family functions, it only
> instruments direct indexed accesses (e.g. "foo[i]"). Dealing with
> the {mem,str}*()-family functions is a work-in-progress around
> CONFIG_FORTIFY_SOURCE[1].
>
> - it ignores flexible array members, including the very old single
> byte (e.g. "int foo[1];") declarations. (Note that GCC's
> implementation appears to ignore _all_ trailing arrays, but Clang only
> ignores empty, 0, and 1 byte arrays[2].)
>
> [1] https://github.com/KSPP/linux/issues/6
> [2] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92589
>
> Suggested-by: Elena Petrova <[email protected]>
> Signed-off-by: Kees Cook <[email protected]>
Reviewed-by: Andrey Ryabinin <[email protected]>