bpf: support deferring bpf_link dealloc to after RCU grace period

BPF link for some program types is passed as a "context" which can be
used by those BPF programs to look up additional information. E.g., for
multi-kprobes and multi-uprobes, link is used to fetch BPF cookie values.

Because of this runtime dependency, when bpf_link refcnt drops to zero
there could still be active BPF programs running accessing link data.

This patch adds generic support to defer bpf_link dealloc callback to
after RCU GP, if requested. This is done by exposing two different
deallocation callbacks, one synchronous and one deferred. If deferred
one is provided, bpf_link_free() will schedule dealloc_deferred()
callback to happen after RCU GP.

BPF is using two flavors of RCU: "classic" non-sleepable one and RCU
tasks trace one. The latter is used when sleepable BPF programs are
used. bpf_link_free() accommodates that by checking underlying BPF
program's sleepable flag, and goes either through normal RCU GP only for
non-sleepable, or through RCU tasks trace GP *and* then normal RCU GP
(taking into account rcu_trace_implies_rcu_gp() optimization), if BPF
program is sleepable.

We use this for multi-kprobe and multi-uprobe links, which dereference
link during program run. We also preventively switch raw_tp link to use
deferred dealloc callback, as upcoming changes in bpf-next tree expose
raw_tp link data (specifically, cookie value) to BPF program at runtime
as well.

Fixes: 0dcac27 ("bpf: Add multi kprobe link")
Fixes: 89ae89f ("bpf: Add multi uprobe link")
Reported-by: [email protected]
Reported-by: [email protected]
Reported-by: [email protected]
Signed-off-by: Andrii Nakryiko <[email protected]>
Acked-by: Jiri Olsa <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Alexei Starovoitov <[email protected]>

Author: anakryiko

include/linux/bpf.h

@@ -1574,12 +1574,26 @@ struct bpf_link {
 	enum bpf_link_type type;
 	const struct bpf_link_ops *ops;
 	struct bpf_prog *prog;
-	struct work_struct work;
+	/* rcu is used before freeing, work can be used to schedule that
+	 * RCU-based freeing before that, so they never overlap
+	 */
+	union {
+		struct rcu_head rcu;
+		struct work_struct work;
+	};
 };
 
 struct bpf_link_ops {
 	void (*release)(struct bpf_link *link);
+	/* deallocate link resources callback, called without RCU grace period
+	 * waiting
+	 */
 	void (*dealloc)(struct bpf_link *link);
+	/* deallocate link resources callback, called after RCU grace period;
+	 * if underlying BPF program is sleepable we go through tasks trace
+	 * RCU GP and then "classic" RCU GP
+	 */
+	void (*dealloc_deferred)(struct bpf_link *link);
 	int (*detach)(struct bpf_link *link);
 	int (*update_prog)(struct bpf_link *link, struct bpf_prog *new_prog,
 			   struct bpf_prog *old_prog);

kernel/bpf/syscall.c

@@ -3024,17 +3024,46 @@ void bpf_link_inc(struct bpf_link *link)
 	atomic64_inc(&link->refcnt);
 }
 
+static void bpf_link_defer_dealloc_rcu_gp(struct rcu_head *rcu)
+{
+	struct bpf_link *link = container_of(rcu, struct bpf_link, rcu);
+
+	/* free bpf_link and its containing memory */
+	link->ops->dealloc_deferred(link);
+}
+
+static void bpf_link_defer_dealloc_mult_rcu_gp(struct rcu_head *rcu)
+{
+	if (rcu_trace_implies_rcu_gp())
+		bpf_link_defer_dealloc_rcu_gp(rcu);
+	else
+		call_rcu(rcu, bpf_link_defer_dealloc_rcu_gp);
+}
+
 /* bpf_link_free is guaranteed to be called from process context */
 static void bpf_link_free(struct bpf_link *link)
 {
+	bool sleepable = false;
+
 	bpf_link_free_id(link->id);
 	if (link->prog) {
+		sleepable = link->prog->sleepable;
 		/* detach BPF program, clean up used resources */
 		link->ops->release(link);
 		bpf_prog_put(link->prog);
 	}
-	/* free bpf_link and its containing memory */
-	link->ops->dealloc(link);
+	if (link->ops->dealloc_deferred) {
+		/* schedule BPF link deallocation; if underlying BPF program
+		 * is sleepable, we need to first wait for RCU tasks trace
+		 * sync, then go through "classic" RCU grace period
+		 */
+		if (sleepable)
+			call_rcu_tasks_trace(&link->rcu, bpf_link_defer_dealloc_mult_rcu_gp);
+		else
+			call_rcu(&link->rcu, bpf_link_defer_dealloc_rcu_gp);
+	}
+	if (link->ops->dealloc)
+		link->ops->dealloc(link);
 }
 
 static void bpf_link_put_deferred(struct work_struct *work)
@@ -3544,7 +3573,7 @@ static int bpf_raw_tp_link_fill_link_info(const struct bpf_link *link,
 
 static const struct bpf_link_ops bpf_raw_tp_link_lops = {
 	.release = bpf_raw_tp_link_release,
-	.dealloc = bpf_raw_tp_link_dealloc,
+	.dealloc_deferred = bpf_raw_tp_link_dealloc,
 	.show_fdinfo = bpf_raw_tp_link_show_fdinfo,
 	.fill_link_info = bpf_raw_tp_link_fill_link_info,
 };

kernel/trace/bpf_trace.c

@@ -2728,7 +2728,7 @@ static int bpf_kprobe_multi_link_fill_link_info(const struct bpf_link *link,
 
 static const struct bpf_link_ops bpf_kprobe_multi_link_lops = {
 	.release = bpf_kprobe_multi_link_release,
-	.dealloc = bpf_kprobe_multi_link_dealloc,
+	.dealloc_deferred = bpf_kprobe_multi_link_dealloc,
 	.fill_link_info = bpf_kprobe_multi_link_fill_link_info,
 };
 
@@ -3242,7 +3242,7 @@ static int bpf_uprobe_multi_link_fill_link_info(const struct bpf_link *link,
 
 static const struct bpf_link_ops bpf_uprobe_multi_link_lops = {
 	.release = bpf_uprobe_multi_link_release,
-	.dealloc = bpf_uprobe_multi_link_dealloc,
+	.dealloc_deferred = bpf_uprobe_multi_link_dealloc,
 	.fill_link_info = bpf_uprobe_multi_link_fill_link_info,
 };