af_packet: do not call packet_read_pending() from tpacket_destruct_skb()

trafgen performance considerably sank on hosts with many cores
after the blamed commit.

packet_read_pending() is very expensive, and calling it
in af_packet fast path defeats Daniel intent in commit
b013840810c2 ("packet: use percpu mmap tx frame pending refcount")

tpacket_destruct_skb() makes room for one packet, we can immediately
wakeup a producer, no need to completely drain the tx ring.

Fixes: 89ed5b519004 ("af_packet: Block execution of tasks waiting for transmit to complete in AF_PACKET")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neil Horman <nhorman@tuxdriver.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://lore.kernel.org/r/20240515163358.4105915-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
index 8c6d3fbb4ed87f17c2e365810106a05fe9b8ff0c..ea3ebc160e25cc661901717a755f47db927c304d 100644 (file)
--- a/net/packet/af_packet.c
+++ b/net/packet/af_packet.c
@@ -2522,8 +2522,7 @@ static void tpacket_destruct_skb(struct sk_buff *skb)
                ts = __packet_set_timestamp(po, ph, skb);
                __packet_set_status(po, ph, TP_STATUS_AVAILABLE | ts);
 
-               if (!packet_read_pending(&po->tx_ring))
-                       complete(&po->skb_completion);
+               complete(&po->skb_completion);
        }
 
        sock_wfree(skb);