/*
* Similar to rss_m2cpuid, but designed to be used by the IP NETISR
* on incoming frames.
*
* If an existing RSS hash exists and it matches what the configured
* hashing is, then use it.
*
* If there's an existing RSS hash but the desired hash is different,
* or if there's no useful RSS hash, then calculate it via
* the software path.
*
* XXX TODO: definitely want statistics here!
*/
struct mbuf *
rss_soft_m2cpuid(struct mbuf *m, uintptr_t source, u_int *cpuid)
{
uint32_t hash_val, hash_type;
int ret;
M_ASSERTPKTHDR(m);
ret = rss_mbuf_software_hash_v4(m, RSS_HASH_PKT_INGRESS,
&hash_val, &hash_type);
if (ret > 0) {
/* mbuf has a valid hash already; don't need to modify it */
*cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
} else if (ret == 0) {
/* hash was done; update */
m->m_pkthdr.flowid = hash_val;
M_HASHTYPE_SET(m, hash_type);
m->m_flags |= M_FLOWID;
*cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
} else { /* ret < 0 */
/* no hash was done */
*cpuid = NETISR_CPUID_NONE;
}
return (m);
}
static struct inpcbgroup *
in_pcbgroup_bymbuf(struct inpcbinfo *pcbinfo, struct mbuf *m)
{
return (in_pcbgroup_byhash(pcbinfo, M_HASHTYPE_GET(m),
m->m_pkthdr.flowid));
}
/*
* netisr CPU affinity lookup routine for use by protocols.
*/
struct mbuf *
rss_m2cpuid(struct mbuf *m, uintptr_t source, u_int *cpuid)
{
M_ASSERTPKTHDR(m);
*cpuid = rss_hash2cpuid(m->m_pkthdr.flowid, M_HASHTYPE_GET(m));
return (m);
}
int
rss_m2bucket(struct mbuf *m, uint32_t *bucket_id)
{
M_ASSERTPKTHDR(m);
return(rss_hash2bucket(m->m_pkthdr.flowid, M_HASHTYPE_GET(m),
bucket_id));
}
static int
nicvf_if_transmit(struct ifnet *ifp, struct mbuf *mbuf)
{
struct nicvf *nic = if_getsoftc(ifp);
struct queue_set *qs = nic->qs;
struct snd_queue *sq;
struct mbuf *mtmp;
int qidx;
int err = 0;
if (__predict_false(qs == NULL)) {
panic("%s: missing queue set for %s", __func__,
device_get_nameunit(nic->dev));
}
/* Select queue */
if (M_HASHTYPE_GET(mbuf) != M_HASHTYPE_NONE)
qidx = mbuf->m_pkthdr.flowid % qs->sq_cnt;
else
qidx = curcpu % qs->sq_cnt;
sq = &qs->sq[qidx];
if (mbuf->m_next != NULL &&
(mbuf->m_pkthdr.csum_flags &
(CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP)) != 0) {
if (M_WRITABLE(mbuf) == 0) {
mtmp = m_dup(mbuf, M_NOWAIT);
m_freem(mbuf);
if (mtmp == NULL)
return (ENOBUFS);
mbuf = mtmp;
}
}
err = drbr_enqueue(ifp, sq->br, mbuf);
if (((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
IFF_DRV_RUNNING) || !nic->link_up || (err != 0)) {
/*
* Try to enqueue packet to the ring buffer.
* If the driver is not active, link down or enqueue operation
* failed, return with the appropriate error code.
*/
return (err);
}
if (NICVF_TX_TRYLOCK(sq) != 0) {
err = nicvf_xmit_locked(sq);
NICVF_TX_UNLOCK(sq);
return (err);
} else
taskqueue_enqueue(sq->snd_taskq, &sq->snd_task);
return (0);
}
/*
* Do a software calculation of the RSS for the given mbuf.
*
* This is typically used by the input path to recalculate the RSS after
* some form of packet processing (eg de-capsulation, IP fragment reassembly.)
*
* dir is the packet direction - RSS_HASH_PKT_INGRESS for incoming and
* RSS_HASH_PKT_EGRESS for outgoing.
*
* Returns 0 if a hash was done, -1 if no hash was done, +1 if
* the mbuf already had a valid RSS flowid.
*
* This function doesn't modify the mbuf. It's up to the caller to
* assign flowid/flowtype as appropriate.
*/
int
rss_mbuf_software_hash_v4(const struct mbuf *m, int dir, uint32_t *hashval,
uint32_t *hashtype)
{
const struct ip *ip;
const struct tcphdr *th;
const struct udphdr *uh;
uint32_t flowid;
uint32_t flowtype;
uint8_t proto;
int iphlen;
int is_frag = 0;
/*
* XXX For now this only handles hashing on incoming mbufs.
*/
if (dir != RSS_HASH_PKT_INGRESS) {
RSS_DEBUG("called on EGRESS packet!\n");
return (-1);
}
/*
* First, validate that the mbuf we have is long enough
* to have an IPv4 header in it.
*/
if (m->m_pkthdr.len < (sizeof(struct ip))) {
RSS_DEBUG("short mbuf pkthdr\n");
return (-1);
}
if (m->m_len < (sizeof(struct ip))) {
RSS_DEBUG("short mbuf len\n");
return (-1);
}
/* Ok, let's dereference that */
ip = mtod(m, struct ip *);
proto = ip->ip_p;
iphlen = ip->ip_hl << 2;
/*
* If this is a fragment then it shouldn't be four-tuple
* hashed just yet. Once it's reassembled into a full
* frame it should be re-hashed.
*/
if (ip->ip_off & htons(IP_MF | IP_OFFMASK))
is_frag = 1;
/*
* If the mbuf flowid/flowtype matches the packet type,
* and we don't support the 4-tuple version of the given protocol,
* then signal to the owner that it can trust the flowid/flowtype
* details.
*
* This is a little picky - eg, if TCPv4 / UDPv4 hashing
* is supported but we got a TCP/UDP frame only 2-tuple hashed,
* then we shouldn't just "trust" the 2-tuple hash. We need
* a 4-tuple hash.
*/
flowid = m->m_pkthdr.flowid;
flowtype = M_HASHTYPE_GET(m);
if (flowtype != M_HASHTYPE_NONE) {
switch (proto) {
case IPPROTO_UDP:
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) &&
(flowtype == M_HASHTYPE_RSS_UDP_IPV4) &&
(is_frag == 0)) {
return (1);
}
/*
* Only allow 2-tuple for UDP frames if we don't also
* support 4-tuple for UDP.
*/
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) &&
((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) == 0) &&
flowtype == M_HASHTYPE_RSS_IPV4) {
return (1);
}
break;
case IPPROTO_TCP:
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) &&
(flowtype == M_HASHTYPE_RSS_TCP_IPV4) &&
(is_frag == 0)) {
return (1);
}
/*
* Only allow 2-tuple for TCP frames if we don't also
* support 2-tuple for TCP.
*/
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) &&
((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) == 0) &&
flowtype == M_HASHTYPE_RSS_IPV4) {
return (1);
}
break;
default:
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) &&
flowtype == M_HASHTYPE_RSS_IPV4) {
return (1);
}
break;
}
}
/*
* Decode enough information to make a hash decision.
*
* XXX TODO: does the hardware hash on 4-tuple if IP
* options are present?
*/
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4) &&
(proto == IPPROTO_TCP) &&
(is_frag == 0)) {
if (m->m_len < iphlen + sizeof(struct tcphdr)) {
RSS_DEBUG("short TCP frame?\n");
return (-1);
}
th = (const struct tcphdr *)((c_caddr_t)ip + iphlen);
return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst,
th->th_sport,
th->th_dport,
proto,
hashval,
hashtype);
} else if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4) &&
(proto == IPPROTO_UDP) &&
(is_frag == 0)) {
uh = (const struct udphdr *)((c_caddr_t)ip + iphlen);
if (m->m_len < iphlen + sizeof(struct udphdr)) {
RSS_DEBUG("short UDP frame?\n");
return (-1);
}
return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst,
uh->uh_sport,
uh->uh_dport,
proto,
hashval,
hashtype);
} else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV4) {
/* Default to 2-tuple hash */
return rss_proto_software_hash_v4(ip->ip_src, ip->ip_dst,
0, /* source port */
0, /* destination port */
0, /* IPPROTO_IP */
hashval,
hashtype);
} else {
RSS_DEBUG("no available hashtypes!\n");
return (-1);
}
}
/*
* Do a software calculation of the RSS for the given mbuf.
*
* This is typically used by the input path to recalculate the RSS after
* some form of packet processing (eg de-capsulation, IP fragment reassembly.)
*
* dir is the packet direction - RSS_HASH_PKT_INGRESS for incoming and
* RSS_HASH_PKT_EGRESS for outgoing.
*
* Returns 0 if a hash was done, -1 if no hash was done, +1 if
* the mbuf already had a valid RSS flowid.
*
* This function doesn't modify the mbuf. It's up to the caller to
* assign flowid/flowtype as appropriate.
*/
int
rss_mbuf_software_hash_v6(const struct mbuf *m, int dir, uint32_t *hashval,
uint32_t *hashtype)
{
const struct ip6_hdr *ip6;
const struct tcphdr *th;
const struct udphdr *uh;
uint32_t flowtype;
uint8_t proto;
int off, newoff;
int nxt;
/*
* XXX For now this only handles hashing on incoming mbufs.
*/
if (dir != RSS_HASH_PKT_INGRESS) {
RSS_DEBUG("called on EGRESS packet!\n");
return (-1);
}
off = sizeof(struct ip6_hdr);
/*
* First, validate that the mbuf we have is long enough
* to have an IPv6 header in it.
*/
if (m->m_pkthdr.len < off) {
RSS_DEBUG("short mbuf pkthdr\n");
return (-1);
}
if (m->m_len < off) {
RSS_DEBUG("short mbuf len\n");
return (-1);
}
/* Ok, let's dereference that */
ip6 = mtod(m, struct ip6_hdr *);
proto = ip6->ip6_nxt;
/*
* Find the beginning of the TCP/UDP header.
*
* If this is a fragment then it shouldn't be four-tuple
* hashed just yet. Once it's reassembled into a full
* frame it should be re-hashed.
*/
while (proto != IPPROTO_FRAGMENT) {
newoff = ip6_nexthdr(m, off, proto, &nxt);
if (newoff < 0)
break;
off = newoff;
proto = nxt;
}
/*
* If the mbuf flowid/flowtype matches the packet type,
* and we don't support the 4-tuple version of the given protocol,
* then signal to the owner that it can trust the flowid/flowtype
* details.
*
* This is a little picky - eg, if TCPv6 / UDPv6 hashing
* is supported but we got a TCP/UDP frame only 2-tuple hashed,
* then we shouldn't just "trust" the 2-tuple hash. We need
* a 4-tuple hash.
*/
flowtype = M_HASHTYPE_GET(m);
if (flowtype != M_HASHTYPE_NONE) {
switch (proto) {
case IPPROTO_UDP:
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6) &&
(flowtype == M_HASHTYPE_RSS_UDP_IPV6)) {
return (1);
}
/*
* Only allow 2-tuple for UDP frames if we don't also
* support 4-tuple for UDP.
*/
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) &&
((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6) == 0) &&
flowtype == M_HASHTYPE_RSS_IPV6) {
return (1);
}
break;
case IPPROTO_TCP:
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6) &&
(flowtype == M_HASHTYPE_RSS_TCP_IPV6)) {
return (1);
}
/*
* Only allow 2-tuple for TCP frames if we don't also
* support 4-tuple for TCP.
*/
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) &&
((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6) == 0) &&
flowtype == M_HASHTYPE_RSS_IPV6) {
return (1);
}
break;
default:
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) &&
flowtype == M_HASHTYPE_RSS_IPV6) {
return (1);
}
break;
}
}
/*
* Decode enough information to make a hash decision.
*/
if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6) &&
(proto == IPPROTO_TCP)) {
if (m->m_len < off + sizeof(struct tcphdr)) {
RSS_DEBUG("short TCP frame?\n");
return (-1);
}
th = (const struct tcphdr *)((c_caddr_t)ip6 + off);
return rss_proto_software_hash_v6(&ip6->ip6_src, &ip6->ip6_dst,
th->th_sport,
th->th_dport,
proto,
hashval,
hashtype);
} else if ((rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6) &&
(proto == IPPROTO_UDP)) {
if (m->m_len < off + sizeof(struct udphdr)) {
RSS_DEBUG("short UDP frame?\n");
return (-1);
}
uh = (const struct udphdr *)((c_caddr_t)ip6 + off);
return rss_proto_software_hash_v6(&ip6->ip6_src, &ip6->ip6_dst,
uh->uh_sport,
uh->uh_dport,
proto,
hashval,
hashtype);
} else if (rss_gethashconfig() & RSS_HASHTYPE_RSS_IPV6) {
/* Default to 2-tuple hash */
return rss_proto_software_hash_v6(&ip6->ip6_src, &ip6->ip6_dst,
0, /* source port */
0, /* destination port */
0, /* IPPROTO_IP */
hashval,
hashtype);
} else {
RSS_DEBUG("no available hashtypes!\n");
return (-1);
}
}
