Infiniband support for ISC DHCP
This patch adds Infiniband support to DHCP using the Linux Packet Filter
interface.
Concerning the dhcp-client-identifier, I have chosen to diverge from
current IB practice by setting it to the pkey followed by the port GUID:
ff:ff:08:00:38:00:01:37:A8:B1
| | |
+----+-----------------------+
pkey port GUID
This allows having DHCP working over OpenSM's QoS.
Signed-off-by: Sebastien Dugue <[email protected]>
client/dhclient.c | 71 ++++++++++++++++
common/bpf.c | 33 +++++++
common/lpf.c | 239 ++++++++++++++++++++++++++++++++++++++++++++----------
common/socket.c | 4
includes/dhcp.h | 1
includes/dhcpd.h | 3
6 files changed, 308 insertions(+), 43 deletions(-)
Index: dhcp-4.1.0p1/common/lpf.c
===================================================================
--- dhcp-4.1.0p1.orig/common/lpf.c 2010-05-17 16:03:17.000000000 +0200
+++ dhcp-4.1.0p1/common/lpf.c 2010-05-17 16:03:33.000000000 +0200
@@ -42,6 +42,7 @@
#include "includes/netinet/udp.h"
#include "includes/netinet/if_ether.h"
#include <net/if.h>
+#include <ifaddrs.h>
#ifndef PACKET_AUXDATA
#define PACKET_AUXDATA 8
@@ -59,6 +60,15 @@
/* Reinitializes the specified interface after an address change. This
is not required for packet-filter APIs. */
+/* Default broadcast address for IPoIB */
+unsigned char default_ib_bcast_addr[20] = {
+ 0x00, 0xff, 0xff, 0xff,
+ 0xff, 0x12, 0x40, 0x1b,
+ 0x00, 0x00, 0x00, 0x00,
+ 0x00, 0x00, 0x00, 0x00,
+ 0xff, 0xff, 0xff, 0xff
+};
+
#ifdef USE_LPF_SEND
void if_reinitialize_send (info)
struct interface_info *info;
@@ -86,10 +96,21 @@
struct sockaddr common;
} sa;
struct ifreq ifr;
+ int type;
+ int protocol;
/* Make an LPF socket. */
- if ((sock = socket(PF_PACKET, SOCK_RAW,
- htons((short)ETH_P_ALL))) < 0) {
+ get_hw_addr(info);
+
+ if (info->hw_address.hbuf[0] == HTYPE_INFINIBAND) {
+ type = SOCK_DGRAM;
+ protocol = ETHERTYPE_IP;
+ } else {
+ type = SOCK_RAW;
+ protocol = ETH_P_ALL;
+ }
+
+ if ((sock = socket(PF_PACKET, type, htons((short)protocol))) < 0) {
if (errno == ENOPROTOOPT || errno == EPROTONOSUPPORT ||
errno == ESOCKTNOSUPPORT || errno == EPFNOSUPPORT ||
errno == EAFNOSUPPORT || errno == EINVAL) {
@@ -111,6 +132,7 @@
/* Bind to the interface name */
memset (&sa, 0, sizeof sa);
sa.ll.sll_family = AF_PACKET;
+ sa.ll.sll_protocol = htons(protocol);
sa.ll.sll_ifindex = ifr.ifr_ifindex;
if (bind (sock, &sa.common, sizeof sa)) {
if (errno == ENOPROTOOPT || errno == EPROTONOSUPPORT ||
@@ -126,8 +148,6 @@
log_fatal ("Bind socket to interface: %m");
}
- get_hw_addr(info->name, &info->hw_address);
-
return sock;
}
#endif /* USE_LPF_SEND || USE_LPF_RECEIVE */
@@ -182,6 +202,8 @@
in bpf includes... */
extern struct sock_filter dhcp_bpf_filter [];
extern int dhcp_bpf_filter_len;
+extern struct sock_filter dhcp_ib_bpf_filter [];
+extern int dhcp_ib_bpf_filter_len;
#if defined (HAVE_TR_SUPPORT)
extern struct sock_filter dhcp_bpf_tr_filter [];
@@ -199,11 +221,13 @@
/* Open a LPF device and hang it on this interface... */
info -> rfdesc = if_register_lpf (info);
- val = 1;
- if (setsockopt (info -> rfdesc, SOL_PACKET, PACKET_AUXDATA, &val,
- sizeof val) < 0) {
- if (errno != ENOPROTOOPT)
- log_fatal ("Failed to set auxiliary packet data: %m");
+ if (info->hw_address.hbuf[0] != HTYPE_INFINIBAND) {
+ val = 1;
+ if (setsockopt (info -> rfdesc, SOL_PACKET, PACKET_AUXDATA,
+ &val, sizeof val) < 0) {
+ if (errno != ENOPROTOOPT)
+ log_fatal ("Failed to set auxiliary packet
data: %m");
+ }
}
#if defined (HAVE_TR_SUPPORT)
@@ -249,15 +273,28 @@
memset(&p, 0, sizeof(p));
- /* Set up the bpf filter program structure. This is defined in
- bpf.c */
- p.len = dhcp_bpf_filter_len;
- p.filter = dhcp_bpf_filter;
-
- /* Patch the server port into the LPF program...
- XXX changes to filter program may require changes
- to the insn number(s) used below! XXX */
- dhcp_bpf_filter [8].k = ntohs ((short)local_port);
+ if (info->hw_address.hbuf[0] == HTYPE_INFINIBAND) {
+ /* Set up the bpf filter program structure. */
+ p.len = dhcp_ib_bpf_filter_len;
+ p.filter = dhcp_ib_bpf_filter;
+
+ /* Patch the server port into the LPF program...
+ XXX
+ changes to filter program may require changes
+ to the insn number(s) used below!
+ XXX */
+ dhcp_ib_bpf_filter[6].k = ntohs ((short)local_port);
+ } else {
+ /* Set up the bpf filter program structure.
+ This is defined in bpf.c */
+ p.len = dhcp_bpf_filter_len;
+ p.filter = dhcp_bpf_filter;
+
+ /* Patch the server port into the LPF program...
+ XXX changes to filter program may require changes
+ to the insn number(s) used below! XXX */
+ dhcp_bpf_filter [8].k = ntohs ((short)local_port);
+ }
if (setsockopt (info -> rfdesc, SOL_SOCKET, SO_ATTACH_FILTER, &p,
sizeof p) < 0) {
@@ -313,6 +350,54 @@
#endif /* USE_LPF_RECEIVE */
#ifdef USE_LPF_SEND
+ssize_t send_packet_ib(interface, packet, raw, len, from, to, hto)
+ struct interface_info *interface;
+ struct packet *packet;
+ struct dhcp_packet *raw;
+ size_t len;
+ struct in_addr from;
+ struct sockaddr_in *to;
+ struct hardware *hto;
+{
+ unsigned ibufp = 0;
+ double ih [1536 / sizeof (double)];
+ unsigned char *buf = (unsigned char *)ih;
+ ssize_t result;
+
+ union sockunion {
+ struct sockaddr sa;
+ struct sockaddr_ll sll;
+ struct sockaddr_storage ss;
+ } su;
+
+ assemble_udp_ip_header (interface, buf, &ibufp, from.s_addr,
+ to->sin_addr.s_addr, to->sin_port,
+ (unsigned char *)raw, len);
+ memcpy (buf + ibufp, raw, len);
+
+ memset(&su, 0, sizeof(su));
+ su.sll.sll_family = AF_PACKET;
+ su.sll.sll_protocol = htons(ETHERTYPE_IP);
+
+ if (!(su.sll.sll_ifindex = if_nametoindex(interface->name))) {
+ errno = ENOENT;
+ log_error ("send_packet_ib: %m - failed to get if index");
+ return -1;
+ }
+
+ su.sll.sll_hatype = htons(HTYPE_INFINIBAND);
+ su.sll.sll_halen = interface->hw_address.hlen;
+ memcpy(&su.sll.sll_addr, interface->bcast_addr, 20);
+
+ result = sendto(interface->wfdesc, buf, ibufp + len, 0,
+ &su.sa, sizeof(su));
+
+ if (result < 0)
+ log_error ("send_packet_ib: %m");
+
+ return result;
+}
+
ssize_t send_packet (interface, packet, raw, len, from, to, hto)
struct interface_info *interface;
struct packet *packet;
@@ -333,6 +418,11 @@
return send_fallback (interface, packet, raw,
len, from, to, hto);
+ if (interface->hw_address.hbuf[0] == HTYPE_INFINIBAND) {
+ return send_packet_ib(interface, packet, raw, len, from,
+ to, hto);
+ }
+
if (hto == NULL && interface->anycast_mac_addr.hlen)
hto = &interface->anycast_mac_addr;
@@ -354,6 +444,42 @@
#endif /* USE_LPF_SEND */
#ifdef USE_LPF_RECEIVE
+ssize_t receive_packet_ib (interface, buf, len, from, hfrom)
+ struct interface_info *interface;
+ unsigned char *buf;
+ size_t len;
+ struct sockaddr_in *from;
+ struct hardware *hfrom;
+{
+ int length = 0;
+ int offset = 0;
+ unsigned char ibuf [1536];
+ unsigned bufix = 0;
+ unsigned paylen;
+
+ length = read(interface->rfdesc, ibuf, sizeof(ibuf));
+
+ if (length <= 0)
+ return length;
+
+ offset = decode_udp_ip_header(interface, ibuf, bufix, from,
+ (unsigned)length, &paylen, 0);
+
+ if (offset < 0)
+ return 0;
+
+ bufix += offset;
+ length -= offset;
+
+ if (length < paylen)
+ log_fatal("Internal inconsistency at %s:%d.", MDL);
+
+ /* Copy out the data in the packet... */
+ memcpy(buf, &ibuf[bufix], paylen);
+
+ return (ssize_t)paylen;
+}
+
ssize_t receive_packet (interface, buf, len, from, hfrom)
struct interface_info *interface;
unsigned char *buf;
@@ -380,6 +506,10 @@
};
struct cmsghdr *cmsg;
+ if (interface->hw_address.hbuf[0] == HTYPE_INFINIBAND) {
+ return receive_packet_ib(interface, buf, len, from, hfrom);
+ }
+
length = recvmsg (interface -> rfdesc, &msg, 0);
if (length <= 0)
return length;
@@ -460,33 +590,41 @@
}
void
-get_hw_addr(const char *name, struct hardware *hw) {
- int sock;
- struct ifreq tmp;
- struct sockaddr *sa;
+get_hw_addr(struct interface_info *info)
+{
+ struct hardware *hw = &info->hw_address;
+ char *name = info->name;
+ struct ifaddrs *ifaddrs;
+ struct ifaddrs *ifa;
+ struct sockaddr_ll *sll = NULL;
- if (strlen(name) >= sizeof(tmp.ifr_name)) {
- log_fatal("Device name too long: \"%s\"", name);
- }
+ if (getifaddrs(&ifaddrs) == -1)
+ log_fatal("Failed to get interfaces");
+
+ for (ifa = ifaddrs; ifa != NULL; ifa = ifa->ifa_next) {
- sock = socket(AF_INET, SOCK_DGRAM, 0);
- if (sock < 0) {
- log_fatal("Can't create socket for \"%s\": %m", name);
+ if (ifa->ifa_addr->sa_family != AF_PACKET)
+ continue;
+
+ if (ifa->ifa_flags & IFF_LOOPBACK)
+ continue;
+
+ if (strcmp(ifa->ifa_name, name) == 0) {
+ sll = (struct sockaddr_ll *)(void *)ifa->ifa_addr;
+ break;
+ }
}
- memset(&tmp, 0, sizeof(tmp));
- strcpy(tmp.ifr_name, name);
- if (ioctl(sock, SIOCGIFHWADDR, &tmp) < 0) {
- log_fatal("Error getting hardware address for \"%s\": %m",
- name);
+ if (sll == NULL) {
+ freeifaddrs(ifaddrs);
+ log_fatal("Failed to get HW address for %s\n", name);
}
- sa = &tmp.ifr_hwaddr;
- switch (sa->sa_family) {
+ switch (sll->sll_hatype) {
case ARPHRD_ETHER:
hw->hlen = 7;
hw->hbuf[0] = HTYPE_ETHER;
- memcpy(&hw->hbuf[1], sa->sa_data, 6);
+ memcpy(&hw->hbuf[1], sll->sll_addr, 6);
break;
case ARPHRD_IEEE802:
#ifdef ARPHRD_IEEE802_TR
@@ -494,18 +632,39 @@
#endif /* ARPHRD_IEEE802_TR */
hw->hlen = 7;
hw->hbuf[0] = HTYPE_IEEE802;
- memcpy(&hw->hbuf[1], sa->sa_data, 6);
+ memcpy(&hw->hbuf[1], sll->sll_addr, 6);
break;
case ARPHRD_FDDI:
hw->hlen = 17;
hw->hbuf[0] = HTYPE_FDDI;
- memcpy(&hw->hbuf[1], sa->sa_data, 16);
+ memcpy(&hw->hbuf[1], sll->sll_addr, 16);
+ break;
+ case ARPHRD_INFINIBAND:
+ /* For Infiniband, save the broadcast address and store
+ * the port GUID into the hardware address.
+ */
+ if (ifa->ifa_flags & IFF_BROADCAST) {
+ struct sockaddr_ll *bll;
+
+ bll = (struct sockaddr_ll *)ifa->ifa_broadaddr;
+ memcpy(&info->bcast_addr, bll->sll_addr, 20);
+ } else {
+ memcpy(&info->bcast_addr, default_ib_bcast_addr,
+ 20);
+ }
+
+
+ hw->hlen = 9;
+ hw->hbuf[0] = HTYPE_INFINIBAND;
+ memcpy(&hw->hbuf[1],
+ &sll->sll_addr[sll->sll_halen - 8], 8);
break;
default:
+ freeifaddrs(ifaddrs);
log_fatal("Unsupported device type %ld for \"%s\"",
- (long int)sa->sa_family, name);
+ (long int)sll->sll_family, name);
}
- close(sock);
+ freeifaddrs(ifaddrs);
}
#endif
Index: dhcp-4.1.0p1/includes/dhcp.h
===================================================================
--- dhcp-4.1.0p1.orig/includes/dhcp.h 2010-05-17 16:03:17.000000000 +0200
+++ dhcp-4.1.0p1/includes/dhcp.h 2010-05-17 16:03:33.000000000 +0200
@@ -79,6 +79,7 @@
#define HTYPE_ETHER 1 /* Ethernet 10Mbps */
#define HTYPE_IEEE802 6 /* IEEE 802.2 Token Ring... */
#define HTYPE_FDDI 8 /* FDDI... */
+#define HTYPE_INFINIBAND 32 /* Infiniband IPoIB */
/* Magic cookie validating dhcp options field (and bootp vendor
extensions field). */
Index: dhcp-4.1.0p1/client/dhclient.c
===================================================================
--- dhcp-4.1.0p1.orig/client/dhclient.c 2010-05-17 16:03:19.000000000 +0200
+++ dhcp-4.1.0p1/client/dhclient.c 2010-05-17 16:03:33.000000000 +0200
@@ -37,6 +37,8 @@
#include <sys/time.h>
#include <sys/wait.h>
#include <limits.h>
+#include <ifaddrs.h>
+#include <netpacket/packet.h>
#ifdef HAVE_LIBCAP_NG
#include <cap-ng.h>
#endif
@@ -99,6 +101,67 @@
static isc_result_t write_duid(struct data_string *duid);
+static void setup_ib_interface(struct interface_info *ip)
+{
+ struct group *g;
+ struct hardware *hw = &ip->hw_address;
+ char client_id[32];
+ int i;
+ char *arg_conf = NULL;
+ int arg_conf_len = 0;
+ isc_result_t status;
+ struct parse *cfile = (struct parse *)0;
+ const char *val = NULL;
+ int token;
+
+ /* Set the broadcast flag */
+ ip->client->config->bootp_broadcast_always = 1;
+
+ /*
+ * Find out if a dhcp-client-identifier option was specified either
+ * in the config file or on the command line
+ */
+ for (g = ip->client->config->on_transmission; g != NULL; g = g->next) {
+ if ((g->statements != NULL) &&
+ (strcmp(g->statements->data.option->option->name,
+ "dhcp-client-identifier") == 0)) {
+ return;
+ }
+ }
+
+ /*
+ * No client ID specified, make up one based on the PKEY and
+ * port GUID.
+ *
+ * NOTE: The pkey is located at offsets 8 & 9 in the Infiniband
+ * broadcast address.
+ */
+ sprintf(client_id, "%.2x:%.2x:%.2x:%.2x:%.2x:%.2x:%.2x:%.2x:%.2x:%.2x",
+ ip->bcast_addr[8], ip->bcast_addr[9],
+ hw->hbuf[1], hw->hbuf[2], hw->hbuf[3], hw->hbuf[4],
+ hw->hbuf[5], hw->hbuf[6], hw->hbuf[7], hw->hbuf[8]);
+
+ arg_conf_len = asprintf(&arg_conf,
+ "send dhcp-client-identifier %s;",
+ client_id);
+
+ if ((arg_conf == 0) || (arg_conf_len <= 0))
+ log_fatal("Unable to send option dhcp-client-identifier");
+
+ status = new_parse(&cfile, -1, arg_conf, arg_conf_len,
+ "Automatic Infiniband client identifier", 0);
+
+ if ((status != ISC_R_SUCCESS) || (cfile -> warnings_occurred))
+ log_fatal("Failed to parse Infiniband client identifier");
+
+ parse_client_statement(cfile, NULL, ip->client->config);
+
+ if (cfile -> warnings_occurred)
+ log_fatal("Failed to parse Infiniband client identifier");
+
+ end_parse(&cfile);
+}
+
int
main(int argc, char **argv) {
int fd;
@@ -866,6 +929,14 @@
}
srandom(seed + cur_time);
+ /* Setup specific Infiniband options */
+ for (ip = interfaces; ip; ip = ip->next) {
+ if (ip->client &&
+ (ip->hw_address.hbuf[0] == HTYPE_INFINIBAND)) {
+ setup_ib_interface(ip);
+ }
+ }
+
/* Start a configuration state machine for each interface. */
#ifdef DHCPv6
if (local_family == AF_INET6) {
Index: dhcp-4.1.0p1/common/bpf.c
===================================================================
--- dhcp-4.1.0p1.orig/common/bpf.c 2010-05-17 16:03:17.000000000 +0200
+++ dhcp-4.1.0p1/common/bpf.c 2010-05-17 16:03:33.000000000 +0200
@@ -198,11 +198,44 @@
BPF_STMT(BPF_RET+BPF_K, 0),
};
+/* Packet filter program for DHCP over Infiniband.
+ *
+ * XXX
+ * Changes to the filter program may require changes to the constant offsets
+ * used in lpf_gen_filter_setup to patch the port in the BPF program!
+ * XXX
+ */
+struct bpf_insn dhcp_ib_bpf_filter [] = {
+ /* Packet filter for Infiniband */
+ /* Make sure it's a UDP packet... */
+ BPF_STMT(BPF_LD + BPF_B + BPF_ABS, 9),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, IPPROTO_UDP, 0, 6),
+
+ /* Make sure this isn't a fragment... */
+ BPF_STMT(BPF_LD + BPF_H + BPF_ABS, 6),
+ BPF_JUMP(BPF_JMP + BPF_JSET + BPF_K, 0x1fff, 4, 0),
+
+ /* Get the IP header length... */
+ BPF_STMT(BPF_LDX + BPF_B + BPF_MSH, 0),
+
+ /* Make sure it's to the right port... */
+ BPF_STMT(BPF_LD + BPF_H + BPF_IND, 2),
+ BPF_JUMP(BPF_JMP + BPF_JEQ + BPF_K, 67, 0, 1),
+
+ /* If we passed all the tests, ask for the whole packet. */
+ BPF_STMT(BPF_RET + BPF_K, (u_int)-1),
+
+ /* Otherwise, drop it. */
+ BPF_STMT(BPF_RET + BPF_K, 0),
+};
+
#if defined (DEC_FDDI)
struct bpf_insn *bpf_fddi_filter;
#endif
int dhcp_bpf_filter_len = sizeof dhcp_bpf_filter / sizeof (struct bpf_insn);
+int dhcp_ib_bpf_filter_len = sizeof dhcp_ib_bpf_filter / sizeof (struct
bpf_insn);
+
#if defined (HAVE_TR_SUPPORT)
struct bpf_insn dhcp_bpf_tr_filter [] = {
/* accept all token ring packets due to variable length header */
Index: dhcp-4.1.0p1/common/socket.c
===================================================================
--- dhcp-4.1.0p1.orig/common/socket.c 2010-05-17 16:03:17.000000000 +0200
+++ dhcp-4.1.0p1/common/socket.c 2010-05-17 16:03:33.000000000 +0200
@@ -276,7 +276,7 @@
/* If this is a normal IPv4 address, get the hardware address. */
if ((local_family == AF_INET) && (strcmp(info->name, "fallback") != 0))
- get_hw_addr(info->name, &info->hw_address);
+ get_hw_addr(info);
return sock;
}
@@ -422,7 +422,7 @@
if (req_multi)
if_register_multicast(info);
- get_hw_addr(info->name, &info->hw_address);
+ get_hw_addr(info);
if (!quiet_interface_discovery) {
if (info->shared_network != NULL) {
Index: dhcp-4.1.0p1/includes/dhcpd.h
===================================================================
--- dhcp-4.1.0p1.orig/includes/dhcpd.h 2010-05-17 16:03:17.000000000 +0200
+++ dhcp-4.1.0p1/includes/dhcpd.h 2010-05-17 16:03:33.000000000 +0200
@@ -1168,6 +1168,7 @@
struct shared_network *shared_network;
/* Networks connected to this interface. */
struct hardware hw_address; /* Its physical address. */
+ u_int8_t bcast_addr[20]; /* Infiniband broadcast address */
struct in_addr *addresses; /* Addresses associated with this
* interface.
*/
@@ -2197,7 +2198,7 @@
#endif
const char *print_time(TIME);
-void get_hw_addr(const char *name, struct hardware *hw);
+void get_hw_addr(struct interface_info *info);
/* socket.c */
#if defined (USE_SOCKET_SEND) || defined (USE_SOCKET_RECEIVE) \
_______________________________________________
ewg mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
