bpf: implement sleepable uprobes by chaining gps

uprobes work by raising a trap, setting a task flag from within the
interrupt handler, and processing the actual work for the uprobe on the
way back to userspace. As a result, uprobe handlers already execute in a
might_fault/_sleep context. The primary obstacle to sleepable bpf uprobe
programs is therefore on the bpf side.

Namely, the bpf_prog_array attached to the uprobe is protected by normal
rcu. In order for uprobe bpf programs to become sleepable, it has to be
protected by the tasks_trace rcu flavor instead (and kfree() called after
a corresponding grace period).

Therefore, the free path for bpf_prog_array now chains a tasks_trace and
normal grace periods one after the other.

Users who iterate under tasks_trace read section would
be safe, as would users who iterate under normal read sections (from
non-sleepable locations).

The downside is that the tasks_trace latency affects all perf_event-attached
bpf programs (and not just uprobe ones). This is deemed safe given the
possible attach rates for kprobe/uprobe/tp programs.

Separately, non-sleepable programs need access to dynamically sized
rcu-protected maps, so bpf_run_prog_array_sleepables now conditionally takes
an rcu read section, in addition to the overarching tasks_trace section.

Signed-off-by: Delyan Kratunov <delyank@fb.com>
Link: https://lore.kernel.org/r/ce844d62a2fd0443b08c5ab02e95bc7149f9aeb1.1655248076.git.delyank@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

1 files changed, 2 insertions, 2 deletions

diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
index 10b157a6d73e..d1c22594dbf9 100644
--- a/kernel/trace/bpf_trace.c
+++ b/kernel/trace/bpf_trace.c
@@ -1936,7 +1936,7 @@ int perf_event_attach_bpf_prog(struct perf_event *event,
 	event->prog = prog;
 	event->bpf_cookie = bpf_cookie;
 	rcu_assign_pointer(event->tp_event->prog_array, new_array);
-	bpf_prog_array_free(old_array);
+	bpf_prog_array_free_sleepable(old_array);
 
 unlock:
 	mutex_unlock(&bpf_event_mutex);
@@ -1962,7 +1962,7 @@ void perf_event_detach_bpf_prog(struct perf_event *event)
 		bpf_prog_array_delete_safe(old_array, event->prog);
 	} else {
 		rcu_assign_pointer(event->tp_event->prog_array, new_array);
-		bpf_prog_array_free(old_array);
+		bpf_prog_array_free_sleepable(old_array);
 	}
 
 	bpf_prog_put(event->prog);