/*
- * Copyright (c) 2008, 2009 Nicira Networks.
+ * Copyright (c) 2008, 2009, 2010 Nicira Networks.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
#include <fcntl.h>
#include <inttypes.h>
#include <net/if.h>
+#include <linux/types.h>
#include <linux/ethtool.h>
+#include <linux/pkt_sched.h>
#include <linux/rtnetlink.h>
#include <linux/sockios.h>
#include <stdlib.h>
#include <sys/ioctl.h>
+#include <sys/stat.h>
#include <unistd.h>
#include "dpif-provider.h"
-#include "netdev-linux.h"
+#include "netdev.h"
+#include "netdev-vport.h"
#include "ofpbuf.h"
#include "poll-loop.h"
+#include "rtnetlink.h"
+#include "shash.h"
#include "svec.h"
#include "util.h"
-
#include "vlog.h"
-#define THIS_MODULE VLM_dpif_linux
+
+VLOG_DEFINE_THIS_MODULE(dpif_linux);
/* Datapath interface for the openvswitch Linux kernel module. */
struct dpif_linux {
struct dpif dpif;
int fd;
+ /* Used by dpif_linux_get_all_names(). */
+ char *local_ifname;
+ int minor;
+
/* Change notification. */
int local_ifindex; /* Ifindex of local port. */
- struct svec changed_ports; /* Ports that have changed. */
- struct linux_netdev_notifier port_notifier;
+ struct shash changed_ports; /* Ports that have changed. */
+ struct rtnetlink_notifier port_notifier;
+ bool change_error;
};
static struct vlog_rate_limit error_rl = VLOG_RATE_LIMIT_INIT(9999, 5);
static int do_ioctl(const struct dpif *, int cmd, const void *arg);
static int lookup_minor(const char *name, int *minor);
static int finish_open(struct dpif *, const char *local_ifname);
+static int get_openvswitch_major(void);
static int create_minor(const char *name, int minor, struct dpif **dpifp);
static int open_minor(int minor, struct dpif **dpifp);
static int make_openvswitch_device(int minor, char **fnp);
-static void dpif_linux_port_changed(const struct linux_netdev_change *,
+static void dpif_linux_port_changed(const struct rtnetlink_change *,
void *dpif);
static struct dpif_linux *
return CONTAINER_OF(dpif, struct dpif_linux, dpif);
}
-static void
-dpif_linux_run(void)
+static int
+dpif_linux_enumerate(struct svec *all_dps)
{
- linux_netdev_notifier_run();
-}
+ int major;
+ int error;
+ int i;
-static void
-dpif_linux_wait(void)
-{
- linux_netdev_notifier_wait();
+ /* Check that the Open vSwitch module is loaded. */
+ major = get_openvswitch_major();
+ if (major < 0) {
+ return -major;
+ }
+
+ error = 0;
+ for (i = 0; i < ODP_MAX; i++) {
+ struct dpif *dpif;
+ char devname[16];
+ int retval;
+
+ sprintf(devname, "dp%d", i);
+ retval = dpif_open(devname, "system", &dpif);
+ if (!retval) {
+ svec_add(all_dps, devname);
+ dpif_uninit(dpif, true);
+ } else if (retval != ENODEV && !error) {
+ error = retval;
+ }
+ }
+ return error;
}
static int
-dpif_linux_open(const char *name UNUSED, char *suffix, bool create,
- struct dpif **dpifp)
+dpif_linux_open(const struct dpif_class *class OVS_UNUSED, const char *name,
+ bool create, struct dpif **dpifp)
{
int minor;
- minor = !strncmp(name, "dp", 2) && isdigit(name[2]) ? atoi(name + 2) : -1;
+ minor = !strncmp(name, "dp", 2)
+ && isdigit((unsigned char)name[2]) ? atoi(name + 2) : -1;
if (create) {
if (minor >= 0) {
- return create_minor(suffix, minor, dpifp);
+ return create_minor(name, minor, dpifp);
} else {
/* Scan for unused minor number. */
for (minor = 0; minor < ODP_MAX; minor++) {
- int error = create_minor(suffix, minor, dpifp);
+ int error = create_minor(name, minor, dpifp);
if (error != EBUSY) {
return error;
}
int error;
if (minor < 0) {
- error = lookup_minor(suffix, &minor);
+ error = lookup_minor(name, &minor);
if (error) {
return error;
}
* getting the local port's name. */
memset(&port, 0, sizeof port);
port.port = ODPP_LOCAL;
- if (ioctl(dpif->fd, ODP_PORT_QUERY, &port)) {
+ if (ioctl(dpif->fd, ODP_VPORT_QUERY, &port)) {
error = errno;
if (error != ENODEV) {
VLOG_WARN("%s: probe returned unexpected error: %s",
dpif_name(*dpifp), strerror(error));
}
- dpif_close(*dpifp);
+ dpif_uninit(*dpifp, true);
return error;
}
dpif_linux_close(struct dpif *dpif_)
{
struct dpif_linux *dpif = dpif_linux_cast(dpif_);
- linux_netdev_notifier_unregister(&dpif->port_notifier);
- svec_destroy(&dpif->changed_ports);
+ rtnetlink_notifier_unregister(&dpif->port_notifier);
+ shash_destroy(&dpif->changed_ports);
+ free(dpif->local_ifname);
close(dpif->fd);
free(dpif);
}
static int
-dpif_linux_delete(struct dpif *dpif_)
+dpif_linux_get_all_names(const struct dpif *dpif_, struct svec *all_names)
+{
+ struct dpif_linux *dpif = dpif_linux_cast(dpif_);
+
+ svec_add_nocopy(all_names, xasprintf("dp%d", dpif->minor));
+ svec_add(all_names, dpif->local_ifname);
+ return 0;
+}
+
+static int
+dpif_linux_destroy(struct dpif *dpif_)
{
return do_ioctl(dpif_, ODP_DP_DESTROY, NULL);
}
static int
dpif_linux_get_stats(const struct dpif *dpif_, struct odp_stats *stats)
{
+ memset(stats, 0, sizeof *stats);
return do_ioctl(dpif_, ODP_DP_STATS, stats);
}
return do_ioctl(dpif_, ODP_SET_DROP_FRAGS, &drop_frags_int);
}
+static void
+translate_vport_type_to_netdev_type(char *type, size_t size)
+{
+ if (!strcmp(type, "netdev")) {
+ ovs_strlcpy(type, "system", size);
+ }
+}
+
+static void
+translate_netdev_type_to_vport_type(char *type, size_t size)
+{
+ if (!strcmp(type, "system")) {
+ ovs_strlcpy(type, "netdev", size);
+ }
+}
+
static int
-dpif_linux_port_add(struct dpif *dpif_, const char *devname, uint16_t flags,
- uint16_t *port_no)
+dpif_linux_port_add(struct dpif *dpif, struct netdev *netdev,
+ uint16_t *port_nop)
{
+ const char *name = netdev_get_name(netdev);
+ const char *type = netdev_get_type(netdev);
struct odp_port port;
int error;
memset(&port, 0, sizeof port);
- strncpy(port.devname, devname, sizeof port.devname);
- port.flags = flags;
- error = do_ioctl(dpif_, ODP_PORT_ADD, &port);
+ strncpy(port.devname, name, sizeof port.devname);
+ strncpy(port.type, type, sizeof port.type);
+ translate_netdev_type_to_vport_type(port.type, sizeof port.type);
+ netdev_vport_get_config(netdev, port.config);
+
+ error = do_ioctl(dpif, ODP_VPORT_ATTACH, &port);
if (!error) {
- *port_no = port.port;
+ *port_nop = port.port;
}
+
return error;
}
static int
-dpif_linux_port_del(struct dpif *dpif_, uint16_t port_no)
+dpif_linux_port_del(struct dpif *dpif_, uint16_t port_no_)
{
- int tmp = port_no;
- return do_ioctl(dpif_, ODP_PORT_DEL, &tmp);
+ int port_no = port_no_; /* Kernel expects an "int". */
+ return do_ioctl(dpif_, ODP_VPORT_DETACH, &port_no);
}
static int
-dpif_linux_port_query_by_number(const struct dpif *dpif_, uint16_t port_no,
- struct odp_port *port)
+dpif_linux_port_query__(const struct dpif *dpif, struct odp_port *port)
+{
+ int error = do_ioctl(dpif, ODP_VPORT_QUERY, port);
+ if (!error) {
+ translate_vport_type_to_netdev_type(port->type, sizeof port->type);
+ }
+ return error;
+}
+
+static int
+dpif_linux_port_query_by_number(const struct dpif *dpif, uint16_t port_no,
+ struct odp_port *port)
{
memset(port, 0, sizeof *port);
port->port = port_no;
- return do_ioctl(dpif_, ODP_PORT_QUERY, port);
+ return dpif_linux_port_query__(dpif, port);
}
static int
-dpif_linux_port_query_by_name(const struct dpif *dpif_, const char *devname,
+dpif_linux_port_query_by_name(const struct dpif *dpif, const char *devname,
struct odp_port *port)
{
memset(port, 0, sizeof *port);
strncpy(port->devname, devname, sizeof port->devname);
- return do_ioctl(dpif_, ODP_PORT_QUERY, port);
+ return dpif_linux_port_query__(dpif, port);
}
static int
dpif_linux_port_list(const struct dpif *dpif_, struct odp_port *ports, int n)
{
struct odp_portvec pv;
+ unsigned int i;
int error;
pv.ports = ports;
pv.n_ports = n;
- error = do_ioctl(dpif_, ODP_PORT_LIST, &pv);
- return error ? -error : pv.n_ports;
+ error = do_ioctl(dpif_, ODP_VPORT_LIST, &pv);
+ if (error) {
+ return -error;
+ }
+
+ for (i = 0; i < pv.n_ports; i++) {
+ struct odp_port *port = &pv.ports[i];
+
+ translate_vport_type_to_netdev_type(port->type, sizeof port->type);
+ }
+ return pv.n_ports;
}
static int
dpif_linux_port_poll(const struct dpif *dpif_, char **devnamep)
{
struct dpif_linux *dpif = dpif_linux_cast(dpif_);
- int error;
- error = linux_netdev_notifier_get_error(&dpif->port_notifier);
- if (!error) {
- if (!dpif->changed_ports.n) {
- return EAGAIN;
- }
- *devnamep = dpif->changed_ports.names[--dpif->changed_ports.n];
+ if (dpif->change_error) {
+ dpif->change_error = false;
+ shash_clear(&dpif->changed_ports);
+ return ENOBUFS;
+ } else if (!shash_is_empty(&dpif->changed_ports)) {
+ struct shash_node *node = shash_first(&dpif->changed_ports);
+ *devnamep = shash_steal(&dpif->changed_ports, node);
+ return 0;
} else {
- svec_clear(&dpif->changed_ports);
+ return EAGAIN;
}
- return error;
}
static void
dpif_linux_port_poll_wait(const struct dpif *dpif_)
{
struct dpif_linux *dpif = dpif_linux_cast(dpif_);
- if (dpif->changed_ports.n
- || linux_netdev_notifier_peek_error(&dpif->port_notifier)) {
+ if (!shash_is_empty(&dpif->changed_ports) || dpif->change_error) {
poll_immediate_wake();
} else {
- linux_netdev_notifier_wait();
+ rtnetlink_notifier_wait();
}
}
-static int
-dpif_linux_port_group_get(const struct dpif *dpif_, int group,
- uint16_t ports[], int n)
-{
- struct odp_port_group pg;
- int error;
-
- assert(n <= UINT16_MAX);
- pg.group = group;
- pg.ports = ports;
- pg.n_ports = n;
- error = do_ioctl(dpif_, ODP_PORT_GROUP_GET, &pg);
- return error ? -error : pg.n_ports;
-}
-
-static int
-dpif_linux_port_group_set(struct dpif *dpif_, int group,
- const uint16_t ports[], int n)
-{
- struct odp_port_group pg;
-
- assert(n <= UINT16_MAX);
- pg.group = group;
- pg.ports = (uint16_t *) ports;
- pg.n_ports = n;
- return do_ioctl(dpif_, ODP_PORT_GROUP_SET, &pg);
-}
-
static int
dpif_linux_flow_get(const struct dpif *dpif_, struct odp_flow flows[], int n)
{
}
static int
-dpif_linux_execute(struct dpif *dpif_, uint16_t in_port,
- const union odp_action actions[], int n_actions,
+dpif_linux_execute(struct dpif *dpif_,
+ const struct nlattr *actions, unsigned int actions_len,
const struct ofpbuf *buf)
{
struct odp_execute execute;
memset(&execute, 0, sizeof execute);
- execute.in_port = in_port;
- execute.actions = (union odp_action *) actions;
- execute.n_actions = n_actions;
+ execute.actions = (struct nlattr *) actions;
+ execute.actions_len = actions_len;
execute.data = buf->data;
execute.length = buf->size;
return do_ioctl(dpif_, ODP_EXECUTE, &execute);
return do_ioctl(dpif_, ODP_SET_LISTEN_MASK, &listen_mask);
}
+static int
+dpif_linux_get_sflow_probability(const struct dpif *dpif_,
+ uint32_t *probability)
+{
+ return do_ioctl(dpif_, ODP_GET_SFLOW_PROBABILITY, probability);
+}
+
+static int
+dpif_linux_set_sflow_probability(struct dpif *dpif_, uint32_t probability)
+{
+ return do_ioctl(dpif_, ODP_SET_SFLOW_PROBABILITY, &probability);
+}
+
+static int
+dpif_linux_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
+ uint32_t queue_id, uint32_t *priority)
+{
+ if (queue_id < 0xf000) {
+ *priority = TC_H_MAKE(1 << 16, queue_id + 1);
+ return 0;
+ } else {
+ return EINVAL;
+ }
+}
+
static int
dpif_linux_recv(struct dpif *dpif_, struct ofpbuf **bufp)
{
int retval;
int error;
- buf = ofpbuf_new(65536);
+ buf = ofpbuf_new_with_headroom(65536, DPIF_RECV_MSG_PADDING);
retval = read(dpif->fd, ofpbuf_tail(buf), ofpbuf_tailroom(buf));
if (retval < 0) {
error = errno;
return 0;
} else {
VLOG_WARN_RL(&error_rl, "%s: discarding message truncated "
- "from %zu bytes to %d",
+ "from %"PRIu32" bytes to %d",
dpif_name(dpif_), msg->length, retval);
error = ERANGE;
}
}
const struct dpif_class dpif_linux_class = {
- "", /* This is the default class. */
- "linux",
- dpif_linux_run,
- dpif_linux_wait,
+ "system",
+ NULL,
+ NULL,
+ dpif_linux_enumerate,
dpif_linux_open,
dpif_linux_close,
- dpif_linux_delete,
+ dpif_linux_get_all_names,
+ dpif_linux_destroy,
dpif_linux_get_stats,
dpif_linux_get_drop_frags,
dpif_linux_set_drop_frags,
dpif_linux_port_list,
dpif_linux_port_poll,
dpif_linux_port_poll_wait,
- dpif_linux_port_group_get,
- dpif_linux_port_group_set,
dpif_linux_flow_get,
dpif_linux_flow_put,
dpif_linux_flow_del,
dpif_linux_execute,
dpif_linux_recv_get_mask,
dpif_linux_recv_set_mask,
+ dpif_linux_get_sflow_probability,
+ dpif_linux_set_sflow_probability,
+ dpif_linux_queue_to_priority,
dpif_linux_recv,
dpif_linux_recv_wait,
};
\f
static int get_openvswitch_major(void);
-static int get_major(const char *target, int default_major);
+static int get_major(const char *target);
static int
do_ioctl(const struct dpif *dpif_, int cmd, const void *arg)
}
static int
-lookup_minor(const char *name, int *minor)
+lookup_minor(const char *name, int *minorp)
{
struct ethtool_drvinfo drvinfo;
+ int minor, port_no;
struct ifreq ifr;
int error;
int sock;
goto error_close_sock;
}
- if (!isdigit(drvinfo.bus_info[0])) {
- VLOG_WARN("%s ethtool info does not contain an openvswitch minor",
- name);
+ if (sscanf(drvinfo.bus_info, "%d.%d", &minor, &port_no) != 2) {
+ VLOG_WARN("%s ethtool bus_info has unexpected format", name);
error = EPROTOTYPE;
goto error_close_sock;
+ } else if (port_no != ODPP_LOCAL) {
+ /* This is an Open vSwitch device but not the local port. We
+ * intentionally support only using the name of the local port as the
+ * name of a datapath; otherwise, it would be too difficult to
+ * enumerate all the names of a datapath. */
+ error = EOPNOTSUPP;
+ goto error_close_sock;
}
- *minor = atoi(drvinfo.bus_info);
+ *minorp = minor;
close(sock);
return 0;
static int
make_openvswitch_device(int minor, char **fnp)
{
- dev_t dev = makedev(get_openvswitch_major(), minor);
const char dirname[] = "/dev/net";
+ int major;
+ dev_t dev;
struct stat s;
char fn[128];
*fnp = NULL;
+
+ major = get_openvswitch_major();
+ if (major < 0) {
+ return -major;
+ }
+ dev = makedev(major, minor);
+
sprintf(fn, "%s/dp%d", dirname, minor);
if (!stat(fn, &s)) {
if (!S_ISCHR(s.st_mode)) {
fn);
} else if (s.st_rdev != dev) {
VLOG_WARN_RL(&error_rl,
- "%s is device %u:%u instead of %u:%u, fixing",
+ "%s is device %u:%u but should be %u:%u, fixing",
fn, major(s.st_rdev), minor(s.st_rdev),
major(dev), minor(dev));
} else {
return 0;
}
-
+/* Return the major device number of the Open vSwitch device. If it
+ * cannot be determined, a negative errno is returned. */
static int
get_openvswitch_major(void)
{
- static unsigned int openvswitch_major;
- if (!openvswitch_major) {
- enum { DEFAULT_MAJOR = 248 };
- openvswitch_major = get_major("openvswitch", DEFAULT_MAJOR);
+ static int openvswitch_major = -1;
+ if (openvswitch_major < 0) {
+ openvswitch_major = get_major("openvswitch");
}
return openvswitch_major;
}
static int
-get_major(const char *target, int default_major)
+get_major(const char *target)
{
const char fn[] = "/proc/devices";
char line[128];
file = fopen(fn, "r");
if (!file) {
VLOG_ERR("opening %s failed (%s)", fn, strerror(errno));
- goto error;
+ return -errno;
}
for (ln = 1; fgets(line, sizeof line, file); ln++) {
return major;
}
} else {
- static bool warned;
- if (!warned) {
- VLOG_WARN("%s:%d: syntax error", fn, ln);
- }
- warned = true;
+ VLOG_WARN_ONCE("%s:%d: syntax error", fn, ln);
}
}
- VLOG_ERR("%s: %s major not found (is the module loaded?), using "
- "default major %d", fn, target, default_major);
-error:
- VLOG_INFO("using default major %d for %s", default_major, target);
- return default_major;
+ fclose(file);
+
+ VLOG_ERR("%s: %s major not found (is the module loaded?)", fn, target);
+ return -ENODEV;
}
static int
finish_open(struct dpif *dpif_, const char *local_ifname)
{
struct dpif_linux *dpif = dpif_linux_cast(dpif_);
+ dpif->local_ifname = xstrdup(local_ifname);
dpif->local_ifindex = if_nametoindex(local_ifname);
if (!dpif->local_ifindex) {
int error = errno;
- dpif_close(dpif_);
+ dpif_uninit(dpif_, true);
VLOG_WARN("could not get ifindex of %s device: %s",
local_ifname, strerror(errno));
return error;
if (!error) {
error = finish_open(*dpifp, name);
} else {
- dpif_close(*dpifp);
+ dpif_uninit(*dpifp, true);
}
}
return error;
fd = open(fn, O_RDONLY | O_NONBLOCK);
if (fd >= 0) {
struct dpif_linux *dpif = xmalloc(sizeof *dpif);
- error = linux_netdev_notifier_register(&dpif->port_notifier,
- dpif_linux_port_changed, dpif);
+ error = rtnetlink_notifier_register(&dpif->port_notifier,
+ dpif_linux_port_changed, dpif);
if (!error) {
char *name;
free(name);
dpif->fd = fd;
+ dpif->local_ifname = NULL;
+ dpif->minor = minor;
dpif->local_ifindex = 0;
- svec_init(&dpif->changed_ports);
+ shash_init(&dpif->changed_ports);
+ dpif->change_error = false;
*dpifp = &dpif->dpif;
} else {
free(dpif);
}
static void
-dpif_linux_port_changed(const struct linux_netdev_change *change, void *dpif_)
+dpif_linux_port_changed(const struct rtnetlink_change *change, void *dpif_)
{
struct dpif_linux *dpif = dpif_;
- if (change->master_ifindex == dpif->local_ifindex
- && (change->nlmsg_type == RTM_NEWLINK
- || change->nlmsg_type == RTM_DELLINK))
- {
- /* Our datapath changed, either adding a new port or deleting an
- * existing one. */
- if (!svec_contains(&dpif->changed_ports, change->ifname)) {
- svec_add(&dpif->changed_ports, change->ifname);
- svec_sort(&dpif->changed_ports);
+ if (change) {
+ if (change->master_ifindex == dpif->local_ifindex
+ && (change->nlmsg_type == RTM_NEWLINK
+ || change->nlmsg_type == RTM_DELLINK))
+ {
+ /* Our datapath changed, either adding a new port or deleting an
+ * existing one. */
+ shash_add_once(&dpif->changed_ports, change->ifname, NULL);
}
+ } else {
+ dpif->change_error = true;
}
}