]> git.ipfire.org Git - thirdparty/kernel/linux.git/commitdiff
selftests: net: Fix checksums in xdp_native
authorNimrod Oren <noren@nvidia.com>
Wed, 20 May 2026 15:39:28 +0000 (18:39 +0300)
committerJakub Kicinski <kuba@kernel.org>
Thu, 21 May 2026 14:47:00 +0000 (07:47 -0700)
Data adjustment cases failed with "Data exchange failed" when using IPv4
because the program did not update the IP and UDP checksums in the IPv4
branch. The issue was masked when both IPv4 and IPv6 were configured,
since the test harness prefers IPv6.

While here, generalize csum_fold_helper() to fold twice so it works for
any 32-bit input.

Fixes: 0b65cfcef9c5 ("selftests: drv-net: Test tail-adjustment support")
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Reviewed-by: Dragos Tatulea <dtatulea@nvidia.com>
Signed-off-by: Nimrod Oren <noren@nvidia.com>
Link: https://patch.msgid.link/20260520153928.3371765-1-noren@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
tools/testing/selftests/net/lib/xdp_native.bpf.c

index 64f05229ab243781a9016dc9ce4b5b50890b12dc..ded3f896e6224b2a91b4312883e0fc6783ba3c0f 100644 (file)
@@ -268,6 +268,17 @@ static int xdp_mode_tx_handler(struct xdp_md *ctx, __u16 port)
        return XDP_PASS;
 }
 
+static __always_inline __u16 csum_fold_helper(__u32 csum)
+{
+       csum = (csum & 0xffff) + (csum >> 16);
+       return ~((csum & 0xffff) + (csum >> 16));
+}
+
+static __always_inline __u16 csum_fold_udp_helper(__u32 csum)
+{
+       return csum_fold_helper(csum) ? : 0xffff;
+}
+
 static void *update_pkt(struct xdp_md *ctx, __s16 offset, __u32 *udp_csum)
 {
        void *data_end = (void *)(long)ctx->data_end;
@@ -281,21 +292,22 @@ static void *update_pkt(struct xdp_md *ctx, __s16 offset, __u32 *udp_csum)
 
        if (eth->h_proto == bpf_htons(ETH_P_IP)) {
                struct iphdr *iph = data + sizeof(*eth);
-               __u16 total_len;
 
                if (iph + 1 > (struct iphdr *)data_end)
                        return NULL;
 
-               iph->tot_len = bpf_htons(bpf_ntohs(iph->tot_len) + offset);
-
                udph = (void *)eth + sizeof(*iph) + sizeof(*eth);
                if (!udph || udph + 1 > (struct udphdr *)data_end)
                        return NULL;
 
-               len_new = bpf_htons(bpf_ntohs(udph->len) + offset);
+               len = iph->tot_len;
+               len_new = bpf_htons(bpf_ntohs(len) + offset);
+               iph->tot_len = len_new;
+               iph->check = csum_fold_helper(
+                       bpf_csum_diff(&len, sizeof(len), &len_new,
+                                     sizeof(len_new), ~((__u32)iph->check)));
        } else if (eth->h_proto  == bpf_htons(ETH_P_IPV6)) {
                struct ipv6hdr *ipv6h = data + sizeof(*eth);
-               __u16 payload_len;
 
                if (ipv6h + 1 > (struct ipv6hdr *)data_end)
                        return NULL;
@@ -304,33 +316,27 @@ static void *update_pkt(struct xdp_md *ctx, __s16 offset, __u32 *udp_csum)
                if (!udph || udph + 1 > (struct udphdr *)data_end)
                        return NULL;
 
-               *udp_csum = ~((__u32)udph->check);
-
                len = ipv6h->payload_len;
                len_new = bpf_htons(bpf_ntohs(len) + offset);
                ipv6h->payload_len = len_new;
-
-               *udp_csum = bpf_csum_diff(&len, sizeof(len), &len_new,
-                                         sizeof(len_new), *udp_csum);
-
-               len = udph->len;
-               len_new = bpf_htons(bpf_ntohs(udph->len) + offset);
-               *udp_csum = bpf_csum_diff(&len, sizeof(len), &len_new,
-                                         sizeof(len_new), *udp_csum);
        } else {
                return NULL;
        }
 
+       len = udph->len;
+       len_new = bpf_htons(bpf_ntohs(len) + offset);
+
+       *udp_csum = ~((__u32)udph->check);
+       *udp_csum = bpf_csum_diff(&len, sizeof(len), &len_new,
+                                 sizeof(len_new), *udp_csum);
+       *udp_csum = bpf_csum_diff(&len, sizeof(len), &len_new,
+                                 sizeof(len_new), *udp_csum);
+
        udph->len = len_new;
 
        return udph;
 }
 
-static __u16 csum_fold_helper(__u32 csum)
-{
-       return ~((csum & 0xffff) + (csum >> 16)) ? : 0xffff;
-}
-
 static int xdp_adjst_tail_shrnk_data(struct xdp_md *ctx, __u16 offset,
                                     unsigned long hdr_len)
 {
@@ -359,7 +365,7 @@ static int xdp_adjst_tail_shrnk_data(struct xdp_md *ctx, __u16 offset,
                return -1;
 
        udp_csum = bpf_csum_diff((__be32 *)tmp_buff, offset, 0, 0, udp_csum);
-       udph->check = (__u16)csum_fold_helper(udp_csum);
+       udph->check = (__u16)csum_fold_udp_helper(udp_csum);
 
        if (bpf_xdp_adjust_tail(ctx, 0 - offset) < 0)
                return -1;
@@ -403,7 +409,7 @@ static int xdp_adjst_tail_grow_data(struct xdp_md *ctx, __u16 offset)
                return -1;
 
        udp_csum = bpf_csum_diff(0, 0, (__be32 *)tmp_buff, offset, udp_csum);
-       udph->check = (__u16)csum_fold_helper(udp_csum);
+       udph->check = (__u16)csum_fold_udp_helper(udp_csum);
 
        buff_len = bpf_xdp_get_buff_len(ctx);
 
@@ -484,8 +490,7 @@ static int xdp_adjst_head_shrnk_data(struct xdp_md *ctx, __u64 hdr_len,
                return -1;
 
        udp_csum = bpf_csum_diff((__be32 *)tmp_buff, offset, 0, 0, udp_csum);
-
-       udph->check = (__u16)csum_fold_helper(udp_csum);
+       udph->check = (__u16)csum_fold_udp_helper(udp_csum);
 
        if (bpf_xdp_load_bytes(ctx, 0, tmp_buff, MAX_ADJST_OFFSET) < 0)
                return -1;
@@ -542,7 +547,7 @@ static int xdp_adjst_head_grow_data(struct xdp_md *ctx, __u64 hdr_len,
                return -1;
 
        udp_csum = bpf_csum_diff(0, 0, (__be32 *)data_buff, offset, udp_csum);
-       udph->check = (__u16)csum_fold_helper(udp_csum);
+       udph->check = (__u16)csum_fold_udp_helper(udp_csum);
 
        if (hdr_len > MAX_ADJST_OFFSET || hdr_len == 0)
                return -1;