Re: [PATCH v7 3/3] udp: Pass iov_tail to udp_update_hdr4()/udp_update_hdr6()

On Fri, Apr 03, 2026 at 06:42:41PM +0200, Laurent Vivier wrote:

Change udp_update_hdr4() and udp_update_hdr6() to take an iov_tail pointing at the UDP frame instead of a contiguous udp_payload_t buffer and explicit data length. This lets vhost-user pass scatter-gather virtqueue buffers directly without an intermediate copy.

The UDP header is built into a local struct udphdr and written back with IOV_PUSH_HEADER(). On the tap side, udp_tap_prepare() wraps the existing udp_payload_t in a two-element iov to match the new interface.

Signed-off-by: Laurent Vivier --- iov.c | 1 - udp.c | 65 +++++++++++++++++++++------------------ udp_internal.h | 4 +-- udp_vu.c | 82 +++++++++++++++++++++++++++----------------------- 4 files changed, 82 insertions(+), 70 deletions(-)

diff --git a/iov.c b/iov.c index b1bcdc4649df..c0d9c6d21322 100644 --- a/iov.c +++ b/iov.c @@ -368,7 +368,6 @@ void *iov_peek_header_(struct iov_tail *tail, void *v, size_t len, size_t align) * * Return: number of bytes written */ -/* cppcheck-suppress unusedFunction */ size_t iov_push_header_(struct iov_tail *tail, const void *v, size_t len) { size_t l; diff --git a/udp.c b/udp.c index e113b26bc726..46eb199c74c9 100644 --- a/udp.c +++ b/udp.c @@ -255,21 +255,22 @@ static void udp_iov_init(const struct ctx *c) /** * udp_update_hdr4() - Update headers for one IPv4 datagram * @ip4h: Pre-filled IPv4 header (except for tot_len and saddr) - * @bp: Pointer to udp_payload_t to update + * @payload: iov_tail with UDP payload to update

To me "UDP payload" implies it doesn't include the UDP header, but I think from the code this *will* include the UDP header.

* @toside: Flowside for destination side * @dlen: Length of UDP payload * @no_udp_csum: Do not set UDP checksum * * Return: size of IPv4 payload (UDP header + data) */ -size_t udp_update_hdr4(struct iphdr *ip4h, struct udp_payload_t *bp, +size_t udp_update_hdr4(struct iphdr *ip4h, struct iov_tail *payload, const struct flowside *toside, size_t dlen, bool no_udp_csum) { const struct in_addr *src = inany_v4(&toside->oaddr); const struct in_addr *dst = inany_v4(&toside->eaddr); - size_t l4len = dlen + sizeof(bp->uh); + size_t l4len = dlen + sizeof(struct udphdr);

dlen + sizeof(uh) would break the reverse christmas tree a bit, but I think I'd still prefer it on balance.

size_t l3len = l4len + sizeof(*ip4h); + struct udphdr uh;

assert(src && dst);

@@ -278,19 +279,18 @@ size_t udp_update_hdr4(struct iphdr *ip4h, struct udp_payload_t *bp, ip4h->saddr = src->s_addr; ip4h->check = csum_ip4_header(l3len, IPPROTO_UDP, *src, *dst);

- bp->uh.source = htons(toside->oport); - bp->uh.dest = htons(toside->eport); - bp->uh.len = htons(l4len); + uh.source = htons(toside->oport); + uh.dest = htons(toside->eport); + uh.len = htons(l4len);

This is done unconditionally, so it could be an initializer for uh.

if (no_udp_csum) { - bp->uh.check = 0; + uh.check = 0; } else { - const struct iovec iov = { - .iov_base = bp->data, - .iov_len = dlen - }; - struct iov_tail data = IOV_TAIL(&iov, 1, 0); - csum_udp4(&bp->uh, *src, *dst, &data, l4len); + struct iov_tail data = *payload;

Clunky, but I can't see a nicer way around this.

+ + IOV_DROP_HEADER(&data, struct udphdr); + csum_udp4(&uh, *src, *dst, &data, l4len); } + IOV_PUSH_HEADER(payload, uh);

return l4len; } @@ -299,18 +299,19 @@ size_t udp_update_hdr4(struct iphdr *ip4h, struct udp_payload_t *bp, * udp_update_hdr6() - Update headers for one IPv6 datagram * @ip6h: Pre-filled IPv6 header (except for payload_len and * addresses) - * @bp: Pointer to udp_payload_t to update + * @payload: iov_tail with UDP payload to update * @toside: Flowside for destination side * @dlen: Length of UDP payload * @no_udp_csum: Do not set UDP checksum * * Return: size of IPv6 payload (UDP header + data) */ -size_t udp_update_hdr6(struct ipv6hdr *ip6h, struct udp_payload_t *bp, +size_t udp_update_hdr6(struct ipv6hdr *ip6h, struct iov_tail *payload, const struct flowside *toside, size_t dlen, bool no_udp_csum)

Similar comments to the IPv4 version.

{ - uint16_t l4len = dlen + sizeof(bp->uh); + uint16_t l4len = dlen + sizeof(struct udphdr); + struct udphdr uh;

ip6h->payload_len = htons(l4len); ip6h->daddr = toside->eaddr.a6; @@ -319,24 +320,23 @@ size_t udp_update_hdr6(struct ipv6hdr *ip6h, struct udp_payload_t *bp, ip6h->nexthdr = IPPROTO_UDP; ip6h->hop_limit = 255;

- bp->uh.source = htons(toside->oport); - bp->uh.dest = htons(toside->eport); - bp->uh.len = ip6h->payload_len; + uh.source = htons(toside->oport); + uh.dest = htons(toside->eport); + uh.len = htons(l4len); if (no_udp_csum) { /* 0 is an invalid checksum for UDP IPv6 and dropped by * the kernel stack, even if the checksum is disabled by virtio * flags. We need to put any non-zero value here. */ - bp->uh.check = 0xffff; + uh.check = 0xffff; } else { - const struct iovec iov = { - .iov_base = bp->data, - .iov_len = dlen - }; - struct iov_tail data = IOV_TAIL(&iov, 1, 0); - csum_udp6(&bp->uh, &toside->oaddr.a6, &toside->eaddr.a6, &data, - l4len); + struct iov_tail data = *payload; + + IOV_DROP_HEADER(&data, struct udphdr); + csum_udp6(&uh, &toside->oaddr.a6, &toside->eaddr.a6, + &data, l4len); } + IOV_PUSH_HEADER(payload, uh);

return l4len; } @@ -375,11 +375,18 @@ static void udp_tap_prepare(const struct mmsghdr *mmh, struct ethhdr *eh = (*tap_iov)[UDP_IOV_ETH].iov_base; struct udp_payload_t *bp = &udp_payload[idx]; struct udp_meta_t *bm = &udp_meta[idx]; + struct iovec iov[2]; + struct iov_tail payload = IOV_TAIL(iov, ARRAY_SIZE(iov), 0); size_t l4len, l2len;

+ iov[0].iov_base = &bp->uh; + iov[0].iov_len = sizeof(bp->uh); + iov[1].iov_base = bp->data; + iov[1].iov_len = mmh[idx].msg_len; + eth_update_mac(eh, NULL, tap_omac); if (!inany_v4(&toside->eaddr) || !inany_v4(&toside->oaddr)) { - l4len = udp_update_hdr6(&bm->ip6h, bp, toside, + l4len = udp_update_hdr6(&bm->ip6h, &payload, toside, mmh[idx].msg_len, no_udp_csum);

l2len = MAX(l4len + sizeof(bm->ip6h) + ETH_HLEN, ETH_ZLEN); @@ -388,7 +395,7 @@ static void udp_tap_prepare(const struct mmsghdr *mmh, eh->h_proto = htons_constant(ETH_P_IPV6); (*tap_iov)[UDP_IOV_IP] = IOV_OF_LVALUE(bm->ip6h); } else { - l4len = udp_update_hdr4(&bm->ip4h, bp, toside, + l4len = udp_update_hdr4(&bm->ip4h, &payload, toside, mmh[idx].msg_len, no_udp_csum);

l2len = MAX(l4len + sizeof(bm->ip4h) + ETH_HLEN, ETH_ZLEN); diff --git a/udp_internal.h b/udp_internal.h index 64e457748324..e6cbaab79519 100644 --- a/udp_internal.h +++ b/udp_internal.h @@ -25,10 +25,10 @@ struct udp_payload_t { } __attribute__ ((packed, aligned(__alignof__(unsigned int)))); #endif

-size_t udp_update_hdr4(struct iphdr *ip4h, struct udp_payload_t *bp, +size_t udp_update_hdr4(struct iphdr *ip4h, struct iov_tail *payload, const struct flowside *toside, size_t dlen, bool no_udp_csum); -size_t udp_update_hdr6(struct ipv6hdr *ip6h, struct udp_payload_t *bp, +size_t udp_update_hdr6(struct ipv6hdr *ip6h, struct iov_tail *payload, const struct flowside *toside, size_t dlen, bool no_udp_csum); void udp_sock_fwd(const struct ctx *c, int s, int rule_hint, diff --git a/udp_vu.c b/udp_vu.c index 16086e6c0c03..989751deafe0 100644 --- a/udp_vu.c +++ b/udp_vu.c @@ -96,43 +96,52 @@ static ssize_t udp_vu_sock_recv(struct iovec *iov, size_t *cnt, int s, bool v6) /** * udp_vu_prepare() - Prepare the packet header * @c: Execution context - * @iov: IO vector for the frame (including vnet header) + * @data: IO vector tail for the frame, on return points to the L4 header

This doesn't make it clear which headers @data includes on entry.

* @toside: Address information for one side of the flow * @dlen: Packet data length * * Return: Layer-4 length */ -static size_t udp_vu_prepare(const struct ctx *c, const struct iovec *iov, - const struct flowside *toside, ssize_t dlen) +static size_t udp_vu_prepare(const struct ctx *c, struct iov_tail *data, + const struct flowside *toside, size_t dlen) { - struct ethhdr *eh; + bool ipv4 = inany_v4(&toside->eaddr) && inany_v4(&toside->oaddr); + struct ethhdr eh; size_t l4len;

/* ethernet header */ - eh = vu_eth(iov[0].iov_base); + memcpy(eh.h_dest, c->guest_mac, sizeof(eh.h_dest)); + memcpy(eh.h_source, c->our_tap_mac, sizeof(eh.h_source));

- memcpy(eh->h_dest, c->guest_mac, sizeof(eh->h_dest)); - memcpy(eh->h_source, c->our_tap_mac, sizeof(eh->h_source)); + if (ipv4) + eh.h_proto = htons(ETH_P_IP); + else + eh.h_proto = htons(ETH_P_IPV6); + IOV_PUSH_HEADER(data, eh);

From this, I'm assuming it does include the ethernet header, but not the vnet header.

/* initialize header */ - if (inany_v4(&toside->eaddr) && inany_v4(&toside->oaddr)) { - struct iphdr *iph = vu_ip(iov[0].iov_base); - struct udp_payload_t *bp = vu_payloadv4(iov[0].iov_base); + if (ipv4) { + struct iov_tail udp_frame; + struct iphdr iph;

- eh->h_proto = htons(ETH_P_IP); + iph = (struct iphdr)L2_BUF_IP4_INIT(IPPROTO_UDP);

This could be an initializer, since iph is a local now.

- *iph = (struct iphdr)L2_BUF_IP4_INIT(IPPROTO_UDP); + udp_frame = *data;

It's not strict, but I generally see the convention as being "frame" for the L2 object, "packet" for the L3 object and "datagram" for the L4 object. So "udp_frame" reads a bit oddly here, I'd call it "datagram".

+ IOV_DROP_HEADER(&udp_frame, struct iphdr); + l4len = udp_update_hdr4(&iph, &udp_frame, toside, dlen, true);

- l4len = udp_update_hdr4(iph, bp, toside, dlen, true); + IOV_PUSH_HEADER(data, iph); } else { - struct ipv6hdr *ip6h = vu_ip(iov[0].iov_base); - struct udp_payload_t *bp = vu_payloadv6(iov[0].iov_base); + struct iov_tail udp_frame; + struct ipv6hdr ip6h;

- eh->h_proto = htons(ETH_P_IPV6); + ip6h = (struct ipv6hdr)L2_BUF_IP6_INIT(IPPROTO_UDP);

- *ip6h = (struct ipv6hdr)L2_BUF_IP6_INIT(IPPROTO_UDP); + udp_frame = *data; + IOV_DROP_HEADER(&udp_frame, struct ipv6hdr); + l4len = udp_update_hdr6(&ip6h, &udp_frame, toside, dlen, true);

- l4len = udp_update_hdr6(ip6h, bp, toside, dlen, true); + IOV_PUSH_HEADER(data, ip6h); }

return l4len; @@ -141,29 +150,25 @@ static size_t udp_vu_prepare(const struct ctx *c, const struct iovec *iov, /** * udp_vu_csum() - Calculate and set checksum for a UDP packet * @toside: Address information for one side of the flow - * @iov: IO vector for the frame - * @cnt: Number of IO vector entries + * @data: IO vector tail for the L4 frame (point to the UDP header) * @l4len: L4 length */ -static void udp_vu_csum(const struct flowside *toside, const struct iovec *iov, - size_t cnt, size_t l4len) +static void udp_vu_csum(const struct flowside *toside, struct iov_tail *data, + size_t l4len) { const struct in_addr *src4 = inany_v4(&toside->oaddr); const struct in_addr *dst4 = inany_v4(&toside->eaddr); - char *base = iov[0].iov_base; - struct udp_payload_t *bp; - struct iov_tail data; - - if (src4 && dst4) { - bp = vu_payloadv4(base); - data = IOV_TAIL(iov, cnt, (char *)&bp->data - base); - csum_udp4(&bp->uh, *src4, *dst4, &data, l4len); - } else { - bp = vu_payloadv6(base); - data = IOV_TAIL(iov, cnt, (char *)&bp->data - base); - csum_udp6(&bp->uh, &toside->oaddr.a6, &toside->eaddr.a6, &data, - l4len); - } + struct udphdr *uh, uh_storage; + bool ipv4 = src4 && dst4; + + uh = IOV_REMOVE_HEADER(data, uh_storage); + if (!uh) + return; + + if (ipv4) + csum_udp4(uh, *src4, *dst4, data, l4len);

This is still implicitly requiring a contiguous UDP header, since you're writing to the pointer from IOV_REMOVE_HEADER. Which makes me think: maybe we should alter iov_peek_header() to return a const pointer to catch cases like that.

+ else + csum_udp6(uh, &toside->oaddr.a6, &toside->eaddr.a6, data, l4len); }

/** @@ -230,9 +235,10 @@ void udp_vu_sock_to_tap(const struct ctx *c, int s, int n, flow_sidx_t tosidx) vu_queue_rewind(vq, elem_cnt - elem_used);

if (iov_cnt > 0) { - size_t l4len = udp_vu_prepare(c, iov_vu, toside, dlen); + struct iov_tail data = IOV_TAIL(iov_vu, iov_cnt, VNET_HLEN); + size_t l4len = udp_vu_prepare(c, &data, toside, dlen); if (*c->pcap) { - udp_vu_csum(toside, iov_vu, iov_cnt, l4len); + udp_vu_csum(toside, &data, l4len); pcap_iov(iov_vu, iov_cnt, VNET_HLEN, hdrlen + dlen - VNET_HLEN); } -- 2.53.0

-- David Gibson (he or they) | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you, not the other way | around. http://www.ozlabs.org/~dgibson