In case a distribution enables branch protection by default do as we do for
the main kernel and explicitly disable branch protection when building the
test case for having BTI disabled to ensure it doesn't get turned on by the
toolchain defaults.
Signed-off-by: Mark Brown <broonie(a)kernel.org>
---
v2:
- Rebase on -next
tools/testing/selftests/arm64/bti/Makefile | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tools/testing/selftests/arm64/bti/Makefile b/tools/testing/selftests/arm64/bti/Makefile
index 73e013c082a6..25c66e622b86 100644
--- a/tools/testing/selftests/arm64/bti/Makefile
+++ b/tools/testing/selftests/arm64/bti/Makefile
@@ -10,7 +10,7 @@ PROGS := $(patsubst %,gen/%,$(TEST_GEN_PROGS))
# cases for statically linked and dynamically lined binaries are
# slightly different.
-CFLAGS_NOBTI = -DBTI=0
+CFLAGS_NOBTI = -mbranch-protection=none -DBTI=0
CFLAGS_BTI = -mbranch-protection=standard -DBTI=1
CFLAGS_COMMON = -ffreestanding -Wall -Wextra $(CFLAGS)
--
2.30.2
The "bti" selftests are built with -nostdlib, which apparently
automatically creates a statically linked binary, which is what we want
and need for BTI (to avoid interactions with the dynamic linker).
However this is not true when building a PIE binary, which some
toolchains (Ubuntu) configure as the default.
When compiling btitest with such a toolchain, it will create a
dynamically linked binary, which will probably fail some tests, as the
dynamic linker might not support BTI:
===================
TAP version 13
1..18
not ok 1 nohint_func/call_using_br_x0
not ok 2 nohint_func/call_using_br_x16
not ok 3 nohint_func/call_using_blr
....
===================
To make sure we create static binaries, add an explicit -static on the
linker command line. This forces static linking even if the toolchain
defaults to PIE builds, and fixes btitest runs on BTI enabled machines.
Signed-off-by: Andre Przywara <andre.przywara(a)arm.com>
---
tools/testing/selftests/arm64/bti/Makefile | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/tools/testing/selftests/arm64/bti/Makefile b/tools/testing/selftests/arm64/bti/Makefile
index 73e013c082a65..dafa1c2aa5c47 100644
--- a/tools/testing/selftests/arm64/bti/Makefile
+++ b/tools/testing/selftests/arm64/bti/Makefile
@@ -39,7 +39,7 @@ BTI_OBJS = \
teststubs-bti.o \
trampoline-bti.o
gen/btitest: $(BTI_OBJS)
- $(CC) $(CFLAGS_BTI) $(CFLAGS_COMMON) -nostdlib -o $@ $^
+ $(CC) $(CFLAGS_BTI) $(CFLAGS_COMMON) -nostdlib -static -o $@ $^
NOBTI_OBJS = \
test-nobti.o \
@@ -50,7 +50,7 @@ NOBTI_OBJS = \
teststubs-nobti.o \
trampoline-nobti.o
gen/nobtitest: $(NOBTI_OBJS)
- $(CC) $(CFLAGS_BTI) $(CFLAGS_COMMON) -nostdlib -o $@ $^
+ $(CC) $(CFLAGS_BTI) $(CFLAGS_COMMON) -nostdlib -static -o $@ $^
# Including KSFT lib.mk here will also mangle the TEST_GEN_PROGS list
# to account for any OUTPUT target-dirs optionally provided by
--
2.25.1
As suggested in [1], the kprobe_multi interface is to be fixed for 32-bit
architectures and compat, rather then disabled. As it turned out,
there are a couple of additional problems that are to be addressed:
- the absence of size overflow checks, leading to possible
out-of-bounds writes (addressed by the first patch);
- the assumption that long has the same size as u64, which would make
cookies arrays size calculation incorrect on 32-bit architectures
(addressed by the second patch);
- the addrs array passing API, that is incompatible with compat and has
to be changed (addressed in the fourth patch): those are kernel
addresses and not user ones (as was incorrectly stated in [2]);
this change is only semantical for 64-bit user/kernelspace,
so it shouldn't impact ABI there, at least.
[1] https://lore.kernel.org/lkml/CAADnVQ+2gwhcMht4PuDnDOFKY68Wsq8QFz4Y69NBX_TLa…
[2] https://lore.kernel.org/lkml/20220510184155.GA8295@asgard.redhat.com/
Eugene Syromiatnikov (4):
bpf_trace: check size for overflow in bpf_kprobe_multi_link_attach
bpf_trace: support 32-bit kernels in bpf_kprobe_multi_link_attach
bpf_trace: handle compat in kprobe_multi_resolve_syms
bpf_trace: pass array of u64 values in kprobe_multi.addrs
kernel/trace/bpf_trace.c | 63 ++++++++++++++++------
tools/lib/bpf/bpf.h | 2 +-
tools/lib/bpf/libbpf.c | 8 +--
tools/lib/bpf/libbpf.h | 2 +-
.../testing/selftests/bpf/prog_tests/bpf_cookie.c | 2 +-
.../selftests/bpf/prog_tests/kprobe_multi_test.c | 2 +-
6 files changed, 54 insertions(+), 25 deletions(-)
--
2.1.4
As suggested in [1], the kprobe_multi interface is to be fixed for 32-bit
architectures and compat, rather then disabled. As it turned out,
there are a couple of additional problems that are to be addressed:
- the absence of size overflow checks, leading to possible
out-of-bounds writes (addressed by the first patch);
- the assumption that long has the same size as u64, which would make
cookies arrays size calculation incorrect on 32-bit architectures
(addressed by the second patch);
- the addrs array passing API, that is incompatible with compat and has
to be changed (addressed in the fourth patch): those are kernel
addresses and not user ones (as was incorrectly stated in [2]);
this change is only semantical for 64-bit user/kernelspace,
so it shouldn't impact ABI there, at least.
[1] https://lore.kernel.org/lkml/CAADnVQ+2gwhcMht4PuDnDOFKY68Wsq8QFz4Y69NBX_TLa…
[2] https://lore.kernel.org/lkml/20220510184155.GA8295@asgard.redhat.com/
Eugene Syromiatnikov (4):
bpf_trace: check size for overflow in bpf_kprobe_multi_link_attach
bpf_trace: support 32-bit kernels in bpf_kprobe_multi_link_attach
bpf_trace: handle compat in kprobe_multi_resolve_syms
bpf_trace: pass array of u64 values in kprobe_multi.addrs
kernel/trace/bpf_trace.c | 63 ++++++++++++++++------
tools/lib/bpf/bpf.h | 2 +-
tools/lib/bpf/libbpf.c | 8 +--
tools/lib/bpf/libbpf.h | 2 +-
.../testing/selftests/bpf/prog_tests/bpf_cookie.c | 2 +-
.../selftests/bpf/prog_tests/kprobe_multi_test.c | 2 +-
6 files changed, 54 insertions(+), 25 deletions(-)
--
2.1.4
In case a distribution enables branch protection by default do as we do for
the main kernel and explicitly disable branch protection when building the
test case for having BTI disabled to ensure it doesn't get turned on by the
toolchain defaults.
Signed-off-by: Mark Brown <broonie(a)kernel.org>
---
tools/testing/selftests/arm64/bti/Makefile | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tools/testing/selftests/arm64/bti/Makefile b/tools/testing/selftests/arm64/bti/Makefile
index 10a12c5e05e6..a3e59d7910b3 100644
--- a/tools/testing/selftests/arm64/bti/Makefile
+++ b/tools/testing/selftests/arm64/bti/Makefile
@@ -14,7 +14,7 @@ CFLAGS += -I../../../../../usr/include/
CFLAGS_STATIC = -ffreestanding
-CFLAGS_NOBTI = -DBTI=0
+CFLAGS_NOBTI = -mbranch-protection=none -DBTI=0
CFLAGS_BTI = -mbranch-protection=standard -DBTI=1
CFLAGS_COMMON = -Wall -Wextra $(CFLAGS)
--
2.30.2