This patch adds PMD type netdev for netdevice with poll-mode
drivers. Since there is no way to get signal on a packet recv
from these devices we need to poll them in busy loop. So minimize
system call overhead this patch uses dpif-thread exclusively
for PMD devices and rest of devices which needs system calls to
do IO are moved to dpif-netdev-run().
PMD device like DPDK work in userspace so there is no system call
overhead for them.
Signed-off-by: Pravin B Shelar <pshe...@nicira.com>
---
v1-v2:
Fixed according to comments from Ben.
Since we have ref on port no need to take port_rwlock. So Removed
port_rwlock thread sefty checks from fast-path functions.
---
lib/dpif-netdev.c | 313 ++++++++++++++++++++++++++++++++++++-----------------
lib/netdev.c | 6 +
lib/netdev.h | 5 +-
3 files changed, 222 insertions(+), 102 deletions(-)
diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index a4bf8a3..efbd156 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -66,6 +66,8 @@ VLOG_DEFINE_THIS_MODULE(dpif_netdev);
/* By default, choose a priority in the middle. */
#define NETDEV_RULE_PRIORITY 0x8000
+#define NR_THREADS 1
+
/* Configuration parameters. */
enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
@@ -161,8 +163,9 @@ struct dp_netdev {
/* Forwarding threads. */
struct latch exit_latch;
- struct dp_forwarder *forwarders;
- size_t n_forwarders;
+ struct pmd_thread *pmd_threads;
+ size_t n_pmd_threads;
+ int pmd_count;
};
static struct dp_netdev_port *dp_netdev_lookup_port(const struct dp_netdev *dp,
@@ -300,13 +303,23 @@ struct dp_netdev_actions *dp_netdev_flow_get_actions(
const struct dp_netdev_flow *);
static void dp_netdev_actions_free(struct dp_netdev_actions *);
-/* A thread that receives packets from some ports, looks them up in the flow
- * table, and executes the actions it finds. */
-struct dp_forwarder {
+/* PMD: Poll modes drivers. PMD accesses devices via polling to eliminate
+ * the performance overhead of interrupt processing. Therefore netdev can
+ * not implement rx-wait for these devices. dpif-netdev needs to poll
+ * these device to check for recv buffer. pmd-thread does polling for
+ * devices assigned to itself thread.
+ *
+ * DPDK used PMD for accessing NIC.
+ *
+ * A thread that receives packets from PMD ports, looks them up in the flow
+ * table, and executes the actions it finds.
+ **/
+struct pmd_thread {
struct dp_netdev *dp;
pthread_t thread;
+ int id;
+ atomic_uint change_seq;
char *name;
- uint32_t min_hash, max_hash;
};
/* Interface to netdev-based datapath. */
@@ -337,18 +350,16 @@ static int dpif_netdev_open(const struct dpif_class *,
const char *name,
static int dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *,
int queue_no, int type,
const struct flow *,
- const struct nlattr *userdata)
- OVS_EXCLUDED(dp->queue_rwlock);
+ const struct nlattr *userdata);
static void dp_netdev_execute_actions(struct dp_netdev *dp,
const struct flow *, struct ofpbuf *,
bool may_steal,
struct pkt_metadata *,
const struct nlattr *actions,
- size_t actions_len)
- OVS_REQ_RDLOCK(dp->port_rwlock);
+ size_t actions_len);
static void dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
- struct pkt_metadata *)
- OVS_REQ_RDLOCK(dp->port_rwlock);
-static void dp_netdev_set_threads(struct dp_netdev *, int n);
+ struct pkt_metadata *);
+
+static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
static struct dpif_netdev *
dpif_netdev_cast(const struct dpif *dpif)
@@ -485,7 +496,6 @@ create_dp_netdev(const char *name, const struct dpif_class
*class,
dp_netdev_free(dp);
return error;
}
- dp_netdev_set_threads(dp, 2);
*dpp = dp;
return 0;
@@ -546,8 +556,8 @@ dp_netdev_free(struct dp_netdev *dp)
shash_find_and_delete(&dp_netdevs, dp->name);
- dp_netdev_set_threads(dp, 0);
- free(dp->forwarders);
+ dp_netdev_set_pmd_threads(dp, 0);
+ free(dp->pmd_threads);
dp_netdev_flow_flush(dp);
ovs_rwlock_wrlock(&dp->port_rwlock);
@@ -641,6 +651,19 @@ dpif_netdev_get_stats(const struct dpif *dpif, struct
dpif_dp_stats *stats)
return 0;
}
+static void
+dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
+{
+ int i;
+
+ for (i = 0; i < dp->n_pmd_threads; i++) {
+ struct pmd_thread *f = &dp->pmd_threads[i];
+ int id;
+
+ atomic_add(&f->change_seq, 1, &id);
+ }
+}
+
static int
do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
odp_port_t port_no)
@@ -649,7 +672,6 @@ do_add_port(struct dp_netdev *dp, const char *devname,
const char *type,
struct netdev_saved_flags *sf;
struct dp_netdev_port *port;
struct netdev *netdev;
- struct netdev_rx *rx;
enum netdev_flags flags;
const char *open_type;
int error;
@@ -671,7 +693,11 @@ do_add_port(struct dp_netdev *dp, const char *devname,
const char *type,
return EINVAL;
}
- error = netdev_rx_open(netdev, &rx);
+ port = xzalloc(sizeof *port);
+ port->port_no = port_no;
+ port->netdev = netdev;
+ port->type = xstrdup(type);
+ error = netdev_rx_open(netdev, &port->rx);
if (error
&& !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
VLOG_ERR("%s: cannot receive packets on this network device (%s)",
@@ -682,21 +708,23 @@ do_add_port(struct dp_netdev *dp, const char *devname,
const char *type,
error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
if (error) {
- netdev_rx_close(rx);
+ netdev_rx_close(port->rx);
netdev_close(netdev);
+ free(port->rx);
+ free(port);
return error;
}
-
- port = xmalloc(sizeof *port);
- port->port_no = port_no;
- port->netdev = netdev;
port->sf = sf;
- port->rx = rx;
- port->type = xstrdup(type);
+
+ if (netdev_is_pmd(netdev)) {
+ dp->pmd_count++;
+ dp_netdev_set_pmd_threads(dp, NR_THREADS);
+ dp_netdev_reload_pmd_threads(dp);
+ }
+ ovs_refcount_init(&port->ref_cnt);
hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0));
seq_change(dp->port_seq);
- ovs_refcount_init(&port->ref_cnt);
return 0;
}
@@ -827,6 +855,9 @@ do_del_port(struct dp_netdev *dp, odp_port_t port_no)
hmap_remove(&dp->ports, &port->node);
seq_change(dp->port_seq);
+ if (netdev_is_pmd(port->netdev)) {
+ dp_netdev_reload_pmd_threads(dp);
+ }
port_unref(port);
return 0;
@@ -1690,116 +1721,199 @@ dp_netdev_actions_free(struct dp_netdev_actions
*actions)
free(actions);
}
+
+inline static void
+dp_netdev_process_rx_port(struct dp_netdev *dp,
+ struct dp_netdev_port *port,
+ struct netdev_rx *queue)
+{
+ struct ofpbuf *packet[NETDEV_MAX_RX_BATCH];
+ int error, c;
+
+ error = netdev_rx_recv(queue, packet, &c);
+ if (!error) {
+ struct pkt_metadata md = PKT_METADATA_INITIALIZER(port->port_no);
+ int i;
+
+ for (i = 0; i < c; i++) {
+ dp_netdev_port_input(dp, packet[i], &md);
+ }
+ } else if (error != EAGAIN && error != EOPNOTSUPP) {
+ static struct vlog_rate_limit rl
+ = VLOG_RATE_LIMIT_INIT(1, 5);
+
+ VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
+ netdev_get_name(port->netdev),
+ ovs_strerror(error));
+ }
+}
+
+static void
+dpif_netdev_run(struct dpif *dpif)
+{
+ struct dp_netdev_port *port;
+ struct dp_netdev *dp = get_dp_netdev(dpif);
+
+ ovs_rwlock_rdlock(&dp->port_rwlock);
+
+ HMAP_FOR_EACH (port, node, &dp->ports) {
+ if (port->rx && !netdev_is_pmd(port->netdev)) {
+ dp_netdev_process_rx_port(dp, port, port->rx);
+ }
+ }
+
+ ovs_rwlock_unlock(&dp->port_rwlock);
+}
+
+static void
+dpif_netdev_wait(struct dpif *dpif)
+{
+ struct dp_netdev_port *port;
+ struct dp_netdev *dp = get_dp_netdev(dpif);
+
+ ovs_rwlock_rdlock(&dp->port_rwlock);
+
+ HMAP_FOR_EACH (port, node, &dp->ports) {
+ if (port->rx && !netdev_is_pmd(port->netdev)) {
+ netdev_rx_wait(port->rx);
+ }
+ }
+ ovs_rwlock_unlock(&dp->port_rwlock);
+}
+
+struct rx_poll {
+ struct dp_netdev_port *port;
+};
+
+static int
+pmd_load_queues(struct pmd_thread *f,
+ struct rx_poll **ppoll_list, int poll_cnt)
+{
+ struct dp_netdev *dp = f->dp;
+ struct rx_poll *poll_list = *ppoll_list;
+ struct dp_netdev_port *port;
+ int id = f->id;
+ int index;
+ int i;
+
+ /* Simple scheduler for netdev rx polling. */
+ ovs_rwlock_rdlock(&dp->port_rwlock);
+ for (i = 0; i < poll_cnt; i++) {
+ port_unref(poll_list[i].port);
+ }
+
+ poll_cnt = 0;
+ index = 0;
+
+ HMAP_FOR_EACH (port, node, &f->dp->ports) {
+ if (netdev_is_pmd(port->netdev)) {
+ if ((index % dp->n_pmd_threads) == id) {
+ poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt
+ 1));
+
+ port_ref(port);
+ poll_list[poll_cnt++].port = port;
+ }
+ index++;
+ }
+ }
+
+ ovs_rwlock_unlock(&dp->port_rwlock);
+ *ppoll_list = poll_list;
+ return poll_cnt;
+}
+
static void *
-dp_forwarder_main(void *f_)
+pmd_thread_main(void *f_)
{
- struct dp_forwarder *f = f_;
+ struct pmd_thread *f = f_;
struct dp_netdev *dp = f->dp;
+ unsigned int lc = 0;
+ struct rx_poll *poll_list;
+ unsigned int port_seq;
+ int poll_cnt;
+ int i;
- f->name = xasprintf("forwarder_%u", ovsthread_id_self());
+ f->name = xasprintf("pmd_%u", ovsthread_id_self());
set_subprogram_name("%s", f->name);
+ poll_cnt = 0;
+ poll_list = NULL;
+
+reload:
+ poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
+ atomic_read(&f->change_seq, &port_seq);
- while (!latch_is_set(&dp->exit_latch)) {
- bool received_anything;
+ for (;;) {
+ unsigned int c_port_seq;
int i;
- ovs_rwlock_rdlock(&dp->port_rwlock);
- for (i = 0; i < 50; i++) {
- struct dp_netdev_port *port;
-
- received_anything = false;
- HMAP_FOR_EACH (port, node, &f->dp->ports) {
- if (port->rx
- && port->node.hash >= f->min_hash
- && port->node.hash <= f->max_hash) {
- struct ofpbuf *packets[NETDEV_MAX_RX_BATCH];
- int count;
- int error;
-
- error = netdev_rx_recv(port->rx, packets, &count);
- if (!error) {
- int i;
- struct pkt_metadata md
- = PKT_METADATA_INITIALIZER(port->port_no);
-
- for (i = 0; i < count; i++) {
- dp_netdev_port_input(dp, packets[i], &md);
- }
- received_anything = true;
- } else if (error != EAGAIN && error != EOPNOTSUPP) {
- static struct vlog_rate_limit rl
- = VLOG_RATE_LIMIT_INIT(1, 5);
-
- VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
- netdev_get_name(port->netdev),
- ovs_strerror(error));
- }
- }
- }
+ for (i = 0; i < poll_cnt; i++) {
+ dp_netdev_process_rx_port(dp, poll_list[i].port,
poll_list[i].port->rx);
+ }
+
+ if (lc++ > 1024) {
+ ovsrcu_quiesce();
- if (!received_anything) {
+ /* TODO: need completely userspace based signaling method.
+ * to keep this thread entirely in userspace.
+ * For now using atomic counter. */
+ lc = 0;
+ atomic_read_explicit(&f->change_seq, &c_port_seq,
memory_order_consume);
+ if (c_port_seq != port_seq) {
break;
}
}
+ }
- if (received_anything) {
- poll_immediate_wake();
- } else {
- struct dp_netdev_port *port;
-
- HMAP_FOR_EACH (port, node, &f->dp->ports)
- if (port->rx
- && port->node.hash >= f->min_hash
- && port->node.hash <= f->max_hash) {
- netdev_rx_wait(port->rx);
- }
- seq_wait(dp->port_seq, seq_read(dp->port_seq));
- latch_wait(&dp->exit_latch);
- }
- ovs_rwlock_unlock(&dp->port_rwlock);
+ if (!latch_is_set(&f->dp->exit_latch)){
+ goto reload;
+ }
- poll_block();
+ for (i = 0; i < poll_cnt; i++) {
+ port_unref(poll_list[i].port);
}
+ free(poll_list);
free(f->name);
-
return NULL;
}
static void
-dp_netdev_set_threads(struct dp_netdev *dp, int n)
+dp_netdev_set_pmd_threads(struct dp_netdev *dp, int n)
{
int i;
- if (n == dp->n_forwarders) {
+ if (n == dp->n_pmd_threads) {
return;
}
/* Stop existing threads. */
latch_set(&dp->exit_latch);
- for (i = 0; i < dp->n_forwarders; i++) {
- struct dp_forwarder *f = &dp->forwarders[i];
+ dp_netdev_reload_pmd_threads(dp);
+ for (i = 0; i < dp->n_pmd_threads; i++) {
+ struct pmd_thread *f = &dp->pmd_threads[i];
xpthread_join(f->thread, NULL);
}
latch_poll(&dp->exit_latch);
- free(dp->forwarders);
+ free(dp->pmd_threads);
/* Start new threads. */
- dp->forwarders = xmalloc(n * sizeof *dp->forwarders);
- dp->n_forwarders = n;
+ dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
+ dp->n_pmd_threads = n;
+
for (i = 0; i < n; i++) {
- struct dp_forwarder *f = &dp->forwarders[i];
+ struct pmd_thread *f = &dp->pmd_threads[i];
f->dp = dp;
- f->min_hash = UINT32_MAX / n * i;
- f->max_hash = UINT32_MAX / n * (i + 1) - 1;
- if (i == n - 1) {
- f->max_hash = UINT32_MAX;
- }
- xpthread_create(&f->thread, NULL, dp_forwarder_main, f);
+ f->id = i;
+ atomic_store(&f->change_seq, 1);
+
+ /* Each thread will distribute all devices rx-queues among
+ * themselves. */
+ xpthread_create(&f->thread, NULL, pmd_thread_main, f);
}
}
+
static void *
dp_netdev_flow_stats_new_cb(void)
@@ -1851,7 +1965,6 @@ dp_netdev_count_packet(struct dp_netdev *dp, enum
dp_stat_type type)
static void
dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
struct pkt_metadata *md)
- OVS_REQ_RDLOCK(dp->port_rwlock)
{
struct dp_netdev_flow *netdev_flow;
struct flow key;
@@ -1883,7 +1996,6 @@ static int
dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet,
int queue_no, int type, const struct flow *flow,
const struct nlattr *userdata)
- OVS_EXCLUDED(dp->queue_rwlock)
{
struct dp_netdev_queue *q;
int error;
@@ -1991,7 +2103,6 @@ dp_netdev_execute_actions(struct dp_netdev *dp, const
struct flow *key,
struct ofpbuf *packet, bool may_steal,
struct pkt_metadata *md,
const struct nlattr *actions, size_t actions_len)
- OVS_REQ_RDLOCK(dp->port_rwlock)
{
struct dp_netdev_execute_aux aux = {dp, key};
@@ -2006,8 +2117,8 @@ const struct dpif_class dpif_netdev_class = {
dpif_netdev_open,
dpif_netdev_close,
dpif_netdev_destroy,
- NULL, /* run */
- NULL, /* wait */
+ dpif_netdev_run,
+ dpif_netdev_wait,
dpif_netdev_get_stats,
dpif_netdev_port_add,
dpif_netdev_port_del,
diff --git a/lib/netdev.c b/lib/netdev.c
index 0d408bf..91c9ea5 100644
--- a/lib/netdev.c
+++ b/lib/netdev.c
@@ -90,6 +90,12 @@ static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5,
20);
static void restore_all_flags(void *aux OVS_UNUSED);
void update_device_args(struct netdev *, const struct shash *args);
+bool
+netdev_is_pmd(const struct netdev *netdev)
+{
+ return !strcmp(netdev->netdev_class->type, "dpdk");
+}
+
static void
netdev_initialize(void)
OVS_EXCLUDED(netdev_class_mutex, netdev_mutex)
diff --git a/lib/netdev.h b/lib/netdev.h
index bbd717f..cc1a45a 100644
--- a/lib/netdev.h
+++ b/lib/netdev.h
@@ -134,8 +134,11 @@ void netdev_wait(void);
void netdev_enumerate_types(struct sset *types);
bool netdev_is_reserved_name(const char *name);
+bool netdev_is_pmd(const struct netdev *netdev);
+
/* Open and close. */
-int netdev_open(const char *name, const char *type, struct netdev **);
+int netdev_open(const char *name, const char *type, struct netdev **netdevp);
+
struct netdev *netdev_ref(const struct netdev *);
void netdev_close(struct netdev *);
--
1.7.9.5
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
