Re: [PATCH net] gve: fix checksum state and gso_segs of HW-GRO packets

From
Eric Dumazet <edumazet@kernel.org>
Date
2026-11-14 15:14:55
Also in
lkml, stable
Reply to
Joshua Washington, 2026-11-13
Le ven. 13 nov. 2026 à 20:56, Joshua Washington [off-list ref] a écrit :
quoted hunk ↗ jump to hunk
From: Ankit Garg <redacted>

With HW-GRO enabled, gve_rx_complete_rsc() turns a coalesced packet into
a GSO skb by setting gso_size, gso_segs and gso_type, but leaves it
CHECKSUM_UNNECESSARY with th->check as the device wrote it. Unlike
tcp4_gro_complete() and tcp6_gro_complete() for a SW-GRO packet, it does
not seed th->check with the TCP pseudo-header checksum or set
CHECKSUM_PARTIAL.

The GRO stack completes a HW-GRO skb itself only if it holds the skb for
merging (since commit 5eddb24901ee ("gro: add support of (hw)gro packets
to gro stack")). Otherwise the skb is delivered as is, e.g. with rx-gro
off or for a packet that is flushed immediately (e.g. with PSH set). If
such an skb is forwarded out of a TSO-capable device, netif_needs_gso()
accepts it as CHECKSUM_UNNECESSARY and the device segments it with a
bogus checksum seed, so every segment is dropped by the receiver. This
blackholes forwarded TCP traffic on VMs acting as routers or NAT
gateways and was observed on VMs in real use.

Fix it in a manner consistent with other drivers with HW-GRO: set the
network and transport headers and, for packets whose checksum the device
verified, seed th->check with the pseudo-header checksum and call
tcp_gro_complete(), passing gso_segs through NAPI_GRO_CB(skb)->count.
skb->data points at the Ethernet header on the napi_gro_frags() path and
at the IP header on the header-split and copybreak paths, so account for
that when setting the network header. Since the header layout now
matters, tell IPv4 from IPv6 by the EtherType, as qede and bnx2x do,
rather than by the packet type.

Also compute gso_segs from the TCP payload length. It was computed from
skb->data_len, which is 0 for a packet that rx_copybreak copied whole
into the linear part, leaving such packets with gso_segs 0 and a
non-zero gso_size.

Fixes: e637c244b954 ("gve: Advertise NETIF_F_GRO_HW instead of NETIF_F_LRO")
Cc: <redacted>
Reviewed-by: Joshua Washington <joshwash@google.com>
Reviewed-by: Jordan Rhee <redacted>
Reviewed-by: Praveen Kaligineedi <redacted>
Signed-off-by: Ankit Garg <redacted>
---
 drivers/net/ethernet/google/gve/gve_rx_dqo.c | 69 +++++++++++++++-----
 1 file changed, 53 insertions(+), 16 deletions(-)
diff --git a/drivers/net/ethernet/google/gve/gve_rx_dqo.c b/drivers/net/ethernet/google/gve/gve_rx_dqo.c
index 82ed5e48b00a..58696d26c698 100644
--- a/drivers/net/ethernet/google/gve/gve_rx_dqo.c
+++ b/drivers/net/ethernet/google/gve/gve_rx_dqo.c
@@ -13,6 +13,7 @@
 #include <linux/ipv6.h>
 #include <linux/skbuff.h>
 #include <linux/slab.h>
+#include <net/gro.h>
 #include <net/ip6_checksum.h>
 #include <net/ipv6.h>
 #include <net/tcp.h>
@@ -1000,13 +1001,34 @@ static int gve_rx_dqo(struct napi_struct *napi, struct gve_rx_ring *rx,
        return -ENOMEM;
 }

+/* Prepare a HW-GRO packet the way tcp4/6_gro_complete() prepare a SW-GRO
+ * packet, so that it can be resegmented by GSO or TSO if it is forwarded:
+ * seed th->check with the TCP pseudo-header checksum and set CHECKSUM_PARTIAL.
+ */
+static void gve_rx_fixup_gro_checksum(struct sk_buff *skb)
+{
+       int len = skb->len - skb_transport_offset(skb);
+       struct tcphdr *th = tcp_hdr(skb);
+
+       if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
+               th->check = ~tcp_v4_check(len, ip_hdr(skb)->saddr,
+                                         ip_hdr(skb)->daddr, 0);
+       else
+               th->check = ~tcp_v6_check(len, &ipv6_hdr(skb)->saddr,
+                                         &ipv6_hdr(skb)->daddr, 0);
+
+       NAPI_GRO_CB(skb)->count = skb_shinfo(skb)->gso_segs;
+       tcp_gro_complete(skb);
+}
+
 static int gve_rx_complete_rsc(struct sk_buff *skb,
                               const struct gve_rx_compl_desc_dqo *desc,
                               struct gve_ptype ptype)
 {
        struct skb_shared_info *shinfo = skb_shinfo(skb);
-       int rsc_segments, rsc_seg_len, hdr_len;
+       int rsc_seg_len, hdr_len, nhoff, thoff;
        skb_frag_t *frag;
+       __be16 proto;
        void *va;

        /* HW-GRO only coalesces TCP. */
@@ -1017,20 +1039,11 @@ static int gve_rx_complete_rsc(struct sk_buff *skb,
        if (!rsc_seg_len)
                return 0;

-       switch (ptype.l3_type) {
-       case GVE_L3_TYPE_IPV4:
-               shinfo->gso_type = SKB_GSO_TCPV4;
-               break;
-       case GVE_L3_TYPE_IPV6:
-               shinfo->gso_type = SKB_GSO_TCPV6;
-               break;
-       default:
-               return -EINVAL;
Before this patch the L3 type came from ptype.l3_type, now it comes from
the EtherType. A VLAN tagged frame (gve does not strip tags) would have
been accepted before, and is now dropped and accounted in
rx_desc_err_dropped_pkt. nhoff = ETH_HLEN has the same assumption.

Can the device coalesce VLAN tagged frames ?

If not, please say so in the changelog. Otherwise this needs to be
handled, like idpf_rx_rsc() does with eth_type_vlan().
quoted hunk ↗ jump to hunk
-       }
-
        if (skb_headlen(skb)) {
-               /* With header-split, payload is in the non-linear part */
-               rsc_segments = DIV_ROUND_UP(skb->data_len, rsc_seg_len);
+               /* With header-split or copybreak, eth_type_trans() has already
+                * pulled the Ethernet header.
+                */
+               nhoff = 0;
        } else {
                /* HW-GRO packets are guaranteed to have complete TCP/IP
                 * headers in frag[0] when header-split is not enabled.
@@ -1039,7 +1052,6 @@ static int gve_rx_complete_rsc(struct sk_buff *skb,
                va = skb_frag_address(frag);
                hdr_len =
                        eth_get_headlen(skb->dev, va, skb_frag_size(frag));
-               rsc_segments = DIV_ROUND_UP(skb->len - hdr_len, rsc_seg_len);
                skb_copy_to_linear_data(skb, va, hdr_len);
                skb_frag_size_sub(frag, hdr_len);
                /* Verify we didn't empty the fragment completely as that could
@@ -1049,10 +1061,35 @@ static int gve_rx_complete_rsc(struct sk_buff *skb,
                skb_frag_off_add(frag, hdr_len);
                skb->data_len -= hdr_len;
                skb->tail += hdr_len;
+               /* napi_gro_frags() pulls the Ethernet header later. */
+               nhoff = ETH_HLEN;
        }
+
+       skb_set_network_header(skb, nhoff);
+       proto = nhoff ? ((struct ethhdr *)skb->data)->h_proto : skb->protocol;
+       switch (proto) {
+       case htons(ETH_P_IP):
+               shinfo->gso_type = SKB_GSO_TCPV4;
+               thoff = nhoff + ip_hdrlen(skb);
+               break;
+       case htons(ETH_P_IPV6):
+               shinfo->gso_type = SKB_GSO_TCPV6;
+               thoff = nhoff + sizeof(struct ipv6hdr);
+               break;
+       default:
+               return -EINVAL;
+       }
+       skb_set_transport_header(skb, thoff);
+
        shinfo->gso_size = rsc_seg_len;
-       shinfo->gso_segs = rsc_segments;
+       shinfo->gso_segs = DIV_ROUND_UP(skb->len - thoff - tcp_hdrlen(skb),
+                                       rsc_seg_len);
+
+       if (!IS_ENABLED(CONFIG_INET) ||
+           unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
+               return 0;

+       gve_rx_fixup_gro_checksum(skb);
thoff is derived from iph->ihl, and tcp_hdrlen() from th->doff, both read
from the packet, and nothing checks that the TCP header is actually in
the linear part before we read th->doff and write th->check.

On the copybreak path the head was allocated by napi_alloc_skb(buf_len),
so a short frame with desc->rsc set and a large ihl ends up reading and
writing past the copied data. Same thing on the napi_get_frags() path if
eth_get_headlen() returned less than the full headers.

The old code was trusting the device as well, but it was not writing
into the headers.

I think we want something like this, before using tcp_hdr(skb):

        if (unlikely(thoff + sizeof(struct tcphdr) > skb_headlen(skb)))
                return -EINVAL;

        if (unlikely(skb->len < thoff + tcp_hdrlen(skb)))
                return -EINVAL;

(skb_headlen() is hdr_len on the napi_get_frags() path at this point,
so one check covers both cases.)

This also makes sure gso_segs can not be computed from a negative length.

Thanks.
Thread (5 messages) 5 messages, 4 authors, last reply 21h ago
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help