bpf: Synchronize dispatcher update with bpf_dispatcher_xdp_func
authorJiri Olsa <jolsa@kernel.org>
Wed, 14 Dec 2022 12:35:42 +0000 (13:35 +0100)
committerMartin KaFai Lau <martin.lau@kernel.org>
Wed, 14 Dec 2022 20:02:14 +0000 (12:02 -0800)
Hao Sun reported crash in dispatcher image [1].

Currently we don't have any sync between bpf_dispatcher_update and
bpf_dispatcher_xdp_func, so following race is possible:

 cpu 0:                               cpu 1:

 bpf_prog_run_xdp
   ...
   bpf_dispatcher_xdp_func
     in image at offset 0x0

                                      bpf_dispatcher_update
                                        update image at offset 0x800
                                      bpf_dispatcher_update
                                        update image at offset 0x0

     in image at offset 0x0 -> crash

Fixing this by synchronizing dispatcher image update (which is done
in bpf_dispatcher_update function) with bpf_dispatcher_xdp_func that
reads and execute the dispatcher image.

Calling synchronize_rcu after updating and installing new image ensures
that readers leave old image before it's changed in the next dispatcher
update. The update itself is locked with dispatcher's mutex.

The bpf_prog_run_xdp is called under local_bh_disable and synchronize_rcu
will wait for it to leave [2].

[1] https://lore.kernel.org/bpf/Y5SFho7ZYXr9ifRn@krava/T/#m00c29ece654bc9f332a17df493bbca33e702896c
[2] https://lore.kernel.org/bpf/0B62D35A-E695-4B7A-A0D4-774767544C1A@gmail.com/T/#mff43e2c003ae99f4a38f353c7969be4c7162e877

Reported-by: Hao Sun <sunhao.th@gmail.com>
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Yonghong Song <yhs@fb.com>
Acked-by: Paul E. McKenney <paulmck@kernel.org>
Link: https://lore.kernel.org/r/20221214123542.1389719-1-jolsa@kernel.org
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
kernel/bpf/dispatcher.c

index c19719f..fa3e922 100644 (file)
@@ -125,6 +125,11 @@ static void bpf_dispatcher_update(struct bpf_dispatcher *d, int prev_num_progs)
 
        __BPF_DISPATCHER_UPDATE(d, new ?: (void *)&bpf_dispatcher_nop_func);
 
+       /* Make sure all the callers executing the previous/old half of the
+        * image leave it, so following update call can modify it safely.
+        */
+       synchronize_rcu();
+
        if (new)
                d->image_off = noff;
 }