eBPF is often used for kernel debugging, and one of the widely used and
powerful debugging techniques is post-mortem debugging with a full memory dump.
Triggering a panic at exactly the right moment allows the user to get such a
dump and thus a better view at the system's state. Right now the only way to
do this in BPF is to signal userspace to trigger kexec/panic. This is
suboptimal as going through userspace requires context changes and adds
significant delays taking system further away from "the right moment". On a
single-cpu system the situation is even worse because BPF program won't even be
able to block the thread of interest.
This patchset tries to solve this problem by allowing properly marked tracing
bpf programs to call crash_kexec() kernel function. The only requirement for
now to run programs calling crash_kexec() or other destructive kfuncs is
CAP_SYS_BOOT capability. When signature checking for bpf programs is available
it is possible that stricter rules will be applied to programs utilizing
destructive kfuncs.
Changes in v5:
- documentation numbering fixed
- no more warning on failed kfunc registration
Changes in v4:
- added description for KF_DESTRUCTIVE flag to documentation
Changes in v3:
- moved kfunc set registration to kernel/bpf/helpers.c
Changes in v2:
- BPF_PROG_LOAD flag dropped as it doesn't fully achieve it's aim of
preventing accidental execution of destructive bpf programs
- selftest moved to the end of patchset
- switched to kfunc destructive flag instead of a separate set
Changes from RFC:
- sysctl knob dropped
- using crash_kexec() instead of panic()
- using kfuncs instead of adding a new helper
Artem Savkov (3):
bpf: add destructive kfunc flag
bpf: export crash_kexec() as destructive kfunc
selftests/bpf: add destructive kfunc test
Documentation/bpf/kfuncs.rst | 9 +++++
include/linux/btf.h | 3 +-
kernel/bpf/helpers.c | 18 ++++++++++
kernel/bpf/verifier.c | 5 +++
net/bpf/test_run.c | 5 +++
.../selftests/bpf/prog_tests/kfunc_call.c | 36 +++++++++++++++++++
.../bpf/progs/kfunc_call_destructive.c | 14 ++++++++
7 files changed, 89 insertions(+), 1 deletion(-)
create mode 100644 tools/testing/selftests/bpf/progs/kfunc_call_destructive.c
--
2.37.1
Add a test checking that programs calling destructive kfuncs can only do
so if they have CAP_SYS_BOOT capabilities.
Signed-off-by: Artem Savkov <[email protected]>
---
net/bpf/test_run.c | 5 +++
.../selftests/bpf/prog_tests/kfunc_call.c | 36 +++++++++++++++++++
.../bpf/progs/kfunc_call_destructive.c | 14 ++++++++
3 files changed, 55 insertions(+)
create mode 100644 tools/testing/selftests/bpf/progs/kfunc_call_destructive.c
diff --git a/net/bpf/test_run.c b/net/bpf/test_run.c
index cbc9cd5058cb..afa7125252f6 100644
--- a/net/bpf/test_run.c
+++ b/net/bpf/test_run.c
@@ -695,6 +695,10 @@ noinline void bpf_kfunc_call_test_ref(struct prog_test_ref_kfunc *p)
{
}
+noinline void bpf_kfunc_call_test_destructive(void)
+{
+}
+
__diag_pop();
ALLOW_ERROR_INJECTION(bpf_modify_return_test, ERRNO);
@@ -719,6 +723,7 @@ BTF_ID_FLAGS(func, bpf_kfunc_call_test_mem_len_pass1)
BTF_ID_FLAGS(func, bpf_kfunc_call_test_mem_len_fail1)
BTF_ID_FLAGS(func, bpf_kfunc_call_test_mem_len_fail2)
BTF_ID_FLAGS(func, bpf_kfunc_call_test_ref, KF_TRUSTED_ARGS)
+BTF_ID_FLAGS(func, bpf_kfunc_call_test_destructive, KF_DESTRUCTIVE)
BTF_SET8_END(test_sk_check_kfunc_ids)
static void *bpf_test_init(const union bpf_attr *kattr, u32 user_size,
diff --git a/tools/testing/selftests/bpf/prog_tests/kfunc_call.c b/tools/testing/selftests/bpf/prog_tests/kfunc_call.c
index c00eb974eb85..351fafa006fb 100644
--- a/tools/testing/selftests/bpf/prog_tests/kfunc_call.c
+++ b/tools/testing/selftests/bpf/prog_tests/kfunc_call.c
@@ -5,6 +5,9 @@
#include "kfunc_call_test.lskel.h"
#include "kfunc_call_test_subprog.skel.h"
#include "kfunc_call_test_subprog.lskel.h"
+#include "kfunc_call_destructive.skel.h"
+
+#include "cap_helpers.h"
static void test_main(void)
{
@@ -86,6 +89,36 @@ static void test_subprog_lskel(void)
kfunc_call_test_subprog_lskel__destroy(skel);
}
+static int test_destructive_open_and_load(void)
+{
+ struct kfunc_call_destructive *skel;
+ int err;
+
+ skel = kfunc_call_destructive__open();
+ if (!ASSERT_OK_PTR(skel, "prog_open"))
+ return -1;
+
+ err = kfunc_call_destructive__load(skel);
+
+ kfunc_call_destructive__destroy(skel);
+
+ return err;
+}
+
+static void test_destructive(void)
+{
+ __u64 save_caps = 0;
+
+ ASSERT_OK(test_destructive_open_and_load(), "succesful_load");
+
+ if (!ASSERT_OK(cap_disable_effective(1ULL << CAP_SYS_BOOT, &save_caps), "drop_caps"))
+ return;
+
+ ASSERT_EQ(test_destructive_open_and_load(), -13, "no_caps_failure");
+
+ cap_enable_effective(save_caps, NULL);
+}
+
void test_kfunc_call(void)
{
if (test__start_subtest("main"))
@@ -96,4 +129,7 @@ void test_kfunc_call(void)
if (test__start_subtest("subprog_lskel"))
test_subprog_lskel();
+
+ if (test__start_subtest("destructive"))
+ test_destructive();
}
diff --git a/tools/testing/selftests/bpf/progs/kfunc_call_destructive.c b/tools/testing/selftests/bpf/progs/kfunc_call_destructive.c
new file mode 100644
index 000000000000..767472bc5a97
--- /dev/null
+++ b/tools/testing/selftests/bpf/progs/kfunc_call_destructive.c
@@ -0,0 +1,14 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <vmlinux.h>
+#include <bpf/bpf_helpers.h>
+
+extern void bpf_kfunc_call_test_destructive(void) __ksym;
+
+SEC("tc")
+int kfunc_destructive_test(void)
+{
+ bpf_kfunc_call_test_destructive();
+ return 0;
+}
+
+char _license[] SEC("license") = "GPL";
--
2.37.1
Hello:
This series was applied to bpf/bpf-next.git (master)
by Alexei Starovoitov <[email protected]>:
On Wed, 10 Aug 2022 08:59:02 +0200 you wrote:
> eBPF is often used for kernel debugging, and one of the widely used and
> powerful debugging techniques is post-mortem debugging with a full memory dump.
> Triggering a panic at exactly the right moment allows the user to get such a
> dump and thus a better view at the system's state. Right now the only way to
> do this in BPF is to signal userspace to trigger kexec/panic. This is
> suboptimal as going through userspace requires context changes and adds
> significant delays taking system further away from "the right moment". On a
> single-cpu system the situation is even worse because BPF program won't even be
> able to block the thread of interest.
>
> [...]
Here is the summary with links:
- [bpf-next,v5,1/3] bpf: add destructive kfunc flag
https://git.kernel.org/bpf/bpf-next/c/4dd48c6f1f83
- [bpf-next,v5,2/3] bpf: export crash_kexec() as destructive kfunc
https://git.kernel.org/bpf/bpf-next/c/133790596406
- [bpf-next,v5,3/3] selftests/bpf: add destructive kfunc test
https://git.kernel.org/bpf/bpf-next/c/e33894581675
You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html