sfc: Remove rcu_read_lock() around XDP program invocation
authorToke Høiland-Jørgensen <toke@redhat.com>
Thu, 24 Jun 2021 16:06:06 +0000 (18:06 +0200)
committerDaniel Borkmann <daniel@iogearbox.net>
Thu, 24 Jun 2021 17:46:01 +0000 (19:46 +0200)
The sfc driver has rcu_read_lock()/rcu_read_unlock() pairs around XDP
program invocations. However, the actual lifetime of the objects referred
by the XDP program invocation is longer, all the way through to the call to
xdp_do_flush(), making the scope of the rcu_read_lock() too small. This
turns out to be harmless because it all happens in a single NAPI poll
cycle (and thus under local_bh_disable()), but it makes the rcu_read_lock()
misleading.

Rather than extend the scope of the rcu_read_lock(), just get rid of it
entirely. With the addition of RCU annotations to the XDP_REDIRECT map
types that take bh execution into account, lockdep even understands this to
be safe, so there's really no reason to keep it around.

Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Edward Cree <ecree.xilinx@gmail.com>
Cc: Martin Habets <habetsm.xilinx@gmail.com>
Link: https://lore.kernel.org/bpf/20210624160609.292325-17-toke@redhat.com
drivers/net/ethernet/sfc/rx.c

index 17b8119..6067509 100644 (file)
@@ -260,18 +260,14 @@ static bool efx_do_xdp(struct efx_nic *efx, struct efx_channel *channel,
        s16 offset;
        int err;
 
-       rcu_read_lock();
-       xdp_prog = rcu_dereference(efx->xdp_prog);
-       if (!xdp_prog) {
-               rcu_read_unlock();
+       xdp_prog = rcu_dereference_bh(efx->xdp_prog);
+       if (!xdp_prog)
                return true;
-       }
 
        rx_queue = efx_channel_get_rx_queue(channel);
 
        if (unlikely(channel->rx_pkt_n_frags > 1)) {
                /* We can't do XDP on fragmented packets - drop. */
-               rcu_read_unlock();
                efx_free_rx_buffers(rx_queue, rx_buf,
                                    channel->rx_pkt_n_frags);
                if (net_ratelimit())
@@ -296,7 +292,6 @@ static bool efx_do_xdp(struct efx_nic *efx, struct efx_channel *channel,
                         rx_buf->len, false);
 
        xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
-       rcu_read_unlock();
 
        offset = (u8 *)xdp.data - *ehp;