Commit 8b401f9ed244 ("bpf: implement bpf_send_signal() helper")
introduced bpf_send_signal() helper and Commit 8482941f0906
("bpf: Add bpf_send_signal_thread() helper") added bpf_send_signal_thread()
helper. Both helpers try to send a signel to current process or thread.
When bpf_prog is called with scheduler rq_lock held, a deadlock
could happen since bpf_send_signal() and bpf_send_signal_thread()
will call group_send_sig_info() which may ultimately want to acquire
rq_lock() again. This happens in 5.2 and 4.16 based kernels in our
production environment with perf_sw_event.
In a different scenario, the following is also possible in the last kernel:
cpu 1:
do_task_stat <- holding sighand->siglock
...
task_sched_runtime <- trying to grab rq_lock
cpu 2:
__schedule <- holding rq_lock
...
do_send_sig_info <- trying to grab sighand->siglock
Commit eac9153f2b58 ("bpf/stackmap: Fix deadlock with
rq_lock in bpf_get_stack()") has a similar issue with above
rq_lock() deadlock. This patch set addressed the issue
in a similar way. Patch #1 provided kernel solution and
Patch #2 added a selftest.
Changelogs:
v1 -> v2:
. previous fix using task_work in nmi() is incorrect.
there is no nmi() involvement here. Using task_work
in all cases might be a solution. But decided to
use a similar fix as in Commit eac9153f2b58.
Yonghong Song (2):
bpf: Fix deadlock with rq_lock in bpf_send_signal()
selftests/bpf: add send_signal_sched_switch test
kernel/trace/bpf_trace.c | 5 +-
.../bpf/prog_tests/send_signal_sched_switch.c | 89 +++++++++++++++++++
.../bpf/progs/test_send_signal_kern.c | 6 ++
3 files changed, 99 insertions(+), 1 deletion(-)
create mode 100644 tools/testing/selftests/bpf/prog_tests/send_signal_sched_switch.c
--
2.17.1
