Hi all!
I would like to kindly ask your help to understand this subject, since my
familiarity with the network stack is limited. I'm still
trying to find a solution for the latency problems with Realtek cards I
reported earlier.
Why does the GSO have to be forced on the socket level
if drivers for high performance chips will have it enabled by default?
A consequence of forcing GSO is that TSO is also forced [1] in
sk_setup_caps() via the NETIF_F_GSO_SOFTWARE flag (list of features with
software fallbacks). I'm sorry for my ignorance, but I wonder if TSO
will really be done in software in case the card does have NETIF_F_TSO
capability but "might not work properly" [2]. (Plus tx-checksum and SG
are all off)
Effectively for me, the following patch shows the same good performance
as the 3.14 kernel (or kernel 4.19 with reverted "tcp: switch to GSO
being always on").
diff --git a/net/core/sock.c b/net/core/sock.c
index 9c32e8eb64da..d792d12e0f66 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -1770,7 +1770,7 @@ void sk_setup_caps(struct sock *sk, struct
dst_entry *dst)
sk_dst_set(sk, dst);
sk->sk_route_caps = dst->dev->features | sk->sk_route_forced_caps;
if (sk->sk_route_caps & NETIF_F_GSO)
- sk->sk_route_caps |= NETIF_F_GSO_SOFTWARE;
+ sk->sk_route_caps |= (NETIF_F_GSO_SOFTWARE &
~NETIF_F_ALL_TSO);
sk->sk_route_caps &= ~sk->sk_route_nocaps;
if (sk_can_gso(sk)) {
if (dst->header_len && !xfrm_dst_offload_ok(dst)) {
Any help is highly appreciated.
Best regards,
Juliana.
[1]
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/net/core/sock.c?h=linux-4.19.y#n1773
[2]
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/drivers/net/ethernet/realtek/r8169.c?h=linux-4.19.y#n7590
