*/
#include <config.h>
-#include "dpif.h"
+#include "dpif-netdev.h"
#include <ctype.h>
#include <errno.h>
#include <unistd.h>
#include "classifier.h"
+#include "cmap.h"
#include "csum.h"
#include "dpif.h"
#include "dpif-provider.h"
#include "dummy.h"
#include "dynamic-string.h"
+#include "fat-rwlock.h"
#include "flow.h"
-#include "hmap.h"
+#include "cmap.h"
#include "latch.h"
#include "list.h"
#include "meta-flow.h"
#include "netdev.h"
+#include "netdev-dpdk.h"
#include "netdev-vport.h"
#include "netlink.h"
#include "odp-execute.h"
#include "ofp-print.h"
#include "ofpbuf.h"
#include "ovs-rcu.h"
+#include "packet-dpif.h"
#include "packets.h"
#include "poll-loop.h"
#include "random.h"
/* By default, choose a priority in the middle. */
#define NETDEV_RULE_PRIORITY 0x8000
+#define FLOW_DUMP_MAX_BATCH 50
+/* Use per thread recirc_depth to prevent recirculation loop. */
+#define MAX_RECIRC_DEPTH 5
+DEFINE_STATIC_PER_THREAD_DATA(uint32_t, recirc_depth, 0)
+
/* Configuration parameters. */
enum { MAX_FLOWS = 65536 }; /* Maximum number of flows in flow table. */
* Acquisition order is, from outermost to innermost:
*
* dp_netdev_mutex (global)
- * port_rwlock
+ * port_mutex
* flow_mutex
- * cls.rwlock
* queue_rwlock
*/
struct dp_netdev {
/* Flows.
*
- * Readers of 'cls' and 'flow_table' must take a 'cls->rwlock' read lock.
- *
- * Writers of 'cls' and 'flow_table' must take the 'flow_mutex' and then
- * the 'cls->rwlock' write lock. (The outer 'flow_mutex' allows writers to
- * atomically perform multiple operations on 'cls' and 'flow_table'.)
+ * Writers of 'flow_table' must take the 'flow_mutex'. Corresponding
+ * changes to 'cls' must be made while still holding the 'flow_mutex'.
*/
struct ovs_mutex flow_mutex;
- struct classifier cls; /* Classifier. Protected by cls.rwlock. */
- struct hmap flow_table OVS_GUARDED; /* Flow table. */
+ struct classifier cls;
+ struct cmap flow_table OVS_GUARDED; /* Flow table. */
/* Queues.
*
/* Ports.
*
- * Any lookup into 'ports' or any access to the dp_netdev_ports found
- * through 'ports' requires taking 'port_rwlock'. */
- struct ovs_rwlock port_rwlock;
- struct hmap ports OVS_GUARDED;
+ * Protected by RCU. Take the mutex to add or remove ports. */
+ struct ovs_mutex port_mutex;
+ struct cmap ports;
struct seq *port_seq; /* Incremented whenever a port changes. */
/* 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,
- odp_port_t)
- OVS_REQ_RDLOCK(dp->port_rwlock);
+ odp_port_t);
enum dp_stat_type {
DP_STAT_HIT, /* Packets that matched in the flow table. */
/* A port in a netdev-based datapath. */
struct dp_netdev_port {
- struct hmap_node node; /* Node in dp_netdev's 'ports'. */
+ struct cmap_node node; /* Node in dp_netdev's 'ports'. */
odp_port_t port_no;
struct netdev *netdev;
struct netdev_saved_flags *sf;
- struct netdev_rx *rx;
+ struct netdev_rxq **rxq;
+ struct ovs_refcount ref_cnt;
char *type; /* Port type as requested by user. */
};
+
+/* Stores a miniflow */
+
+/* There are fields in the flow structure that we never use. Therefore we can
+ * save a few words of memory */
+#define NETDEV_KEY_BUF_SIZE_U32 (FLOW_U32S - MINI_N_INLINE \
+ - FLOW_U32_SIZE(regs) \
+ - FLOW_U32_SIZE(metadata) \
+ )
+struct netdev_flow_key {
+ struct miniflow flow;
+ uint32_t buf[NETDEV_KEY_BUF_SIZE_U32];
+};
+
/* A flow in dp_netdev's 'flow_table'.
*
*
*
* 'flow->ref_cnt' protects 'flow' from being freed. It doesn't protect the
* flow from being deleted from 'cls' (that's 'cls->rwlock') and it doesn't
- * protect members of 'flow' from modification (that's 'flow->mutex').
- *
- * 'flow->mutex' protects the members of 'flow' from modification. It doesn't
- * protect the flow from being deleted from 'cls' (that's 'cls->rwlock') and it
- * doesn't prevent the flow from being freed (that's 'flow->ref_cnt').
+ * protect members of 'flow' from modification.
*
* Some members, marked 'const', are immutable. Accessing other members
* requires synchronization, as noted in more detail below.
const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
/* Hash table index by unmasked flow. */
- const struct hmap_node node; /* In owning dp_netdev's 'flow_table'. */
+ const struct cmap_node node; /* In owning dp_netdev's 'flow_table'. */
const struct flow flow; /* The flow that created this entry. */
- /* Protects members marked OVS_GUARDED.
- *
- * Acquire after datapath's flow_mutex. */
- struct ovs_mutex mutex OVS_ACQ_AFTER(dp_netdev_mutex);
-
/* Statistics.
*
* Reading or writing these members requires 'mutex'. */
struct ovsthread_stats stats; /* Contains "struct dp_netdev_flow_stats". */
- /* Actions.
- *
- * Reading 'actions' requires 'mutex'.
- * Writing 'actions' requires 'mutex' and (to allow for transactions) the
- * datapath's flow_mutex. */
+ /* Actions. */
OVSRCU_TYPE(struct dp_netdev_actions *) actions;
};
* Thread-safety
* =============
*
- * A struct dp_netdev_actions 'actions' may be accessed without a risk of being
- * freed by code that holds a read-lock or write-lock on 'flow->mutex' (where
- * 'flow' is the dp_netdev_flow for which 'flow->actions == actions') or that
- * owns a reference to 'actions->ref_cnt' (or both). */
+ * A struct dp_netdev_actions 'actions' is protected with RCU. */
struct dp_netdev_actions {
/* These members are immutable: they do not change during the struct's
* lifetime. */
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;
- char *name;
- uint32_t min_hash, max_hash;
+ int id;
+ atomic_uint change_seq;
};
/* Interface to netdev-based datapath. */
};
static int get_port_by_number(struct dp_netdev *dp, odp_port_t port_no,
- struct dp_netdev_port **portp)
- OVS_REQ_RDLOCK(dp->port_rwlock);
+ struct dp_netdev_port **portp);
static int get_port_by_name(struct dp_netdev *dp, const char *devname,
- struct dp_netdev_port **portp)
- OVS_REQ_RDLOCK(dp->port_rwlock);
+ struct dp_netdev_port **portp);
static void dp_netdev_free(struct dp_netdev *)
OVS_REQUIRES(dp_netdev_mutex);
static void dp_netdev_flow_flush(struct dp_netdev *);
static int do_add_port(struct dp_netdev *dp, const char *devname,
const char *type, odp_port_t port_no)
- OVS_REQ_WRLOCK(dp->port_rwlock);
-static int do_del_port(struct dp_netdev *dp, odp_port_t port_no)
- OVS_REQ_WRLOCK(dp->port_rwlock);
+ OVS_REQUIRES(dp->port_mutex);
+static void do_del_port(struct dp_netdev *dp, struct dp_netdev_port *)
+ OVS_REQUIRES(dp->port_mutex);
static void dp_netdev_destroy_all_queues(struct dp_netdev *dp)
OVS_REQ_WRLOCK(dp->queue_rwlock);
static int dpif_netdev_open(const struct dpif_class *, const char *name,
bool create, struct dpif **);
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 miniflow *,
+ 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 *,
+ struct dpif_packet **, int c,
+ bool may_steal, struct pkt_metadata *,
const struct nlattr *actions,
- size_t actions_len)
- OVS_REQ_RDLOCK(dp->port_rwlock);
-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);
+ size_t actions_len);
+static void dp_netdev_port_input(struct dp_netdev *dp,
+ struct dpif_packet **packets, int cnt,
+ odp_port_t port_no);
+
+static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
static struct dpif_netdev *
dpif_netdev_cast(const struct dpif *dpif)
}
static int
-dpif_netdev_enumerate(struct sset *all_dps)
+dpif_netdev_enumerate(struct sset *all_dps,
+ const struct dpif_class *dpif_class)
{
struct shash_node *node;
ovs_mutex_lock(&dp_netdev_mutex);
SHASH_FOR_EACH(node, &dp_netdevs) {
+ struct dp_netdev *dp = node->data;
+ if (dpif_class != dp->class) {
+ /* 'dp_netdevs' contains both "netdev" and "dummy" dpifs.
+ * If the class doesn't match, skip this dpif. */
+ continue;
+ }
sset_add(all_dps, node->name);
}
ovs_mutex_unlock(&dp_netdev_mutex);
* Return ODPP_NONE on failure. */
static odp_port_t
choose_port(struct dp_netdev *dp, const char *name)
- OVS_REQ_RDLOCK(dp->port_rwlock)
+ OVS_REQUIRES(dp->port_mutex)
{
uint32_t port_no;
ovs_mutex_init(&dp->flow_mutex);
classifier_init(&dp->cls, NULL);
- hmap_init(&dp->flow_table);
+ cmap_init(&dp->flow_table);
fat_rwlock_init(&dp->queue_rwlock);
ovsthread_stats_init(&dp->stats);
- ovs_rwlock_init(&dp->port_rwlock);
- hmap_init(&dp->ports);
+ ovs_mutex_init(&dp->port_mutex);
+ cmap_init(&dp->ports);
dp->port_seq = seq_create();
latch_init(&dp->exit_latch);
- ovs_rwlock_wrlock(&dp->port_rwlock);
+ ovs_mutex_lock(&dp->port_mutex);
error = do_add_port(dp, name, "internal", ODPP_LOCAL);
- ovs_rwlock_unlock(&dp->port_rwlock);
+ ovs_mutex_unlock(&dp->port_mutex);
if (error) {
dp_netdev_free(dp);
return error;
}
- dp_netdev_set_threads(dp, 2);
*dpp = dp;
return 0;
dp_netdev_free(struct dp_netdev *dp)
OVS_REQUIRES(dp_netdev_mutex)
{
- struct dp_netdev_port *port, *next;
+ struct dp_netdev_port *port;
struct dp_netdev_stats *bucket;
int i;
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);
- HMAP_FOR_EACH_SAFE (port, next, node, &dp->ports) {
- do_del_port(dp, port->port_no);
+ ovs_mutex_lock(&dp->port_mutex);
+ CMAP_FOR_EACH (port, node, &dp->ports) {
+ do_del_port(dp, port);
}
- ovs_rwlock_unlock(&dp->port_rwlock);
+ ovs_mutex_unlock(&dp->port_mutex);
OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
ovs_mutex_destroy(&bucket->mutex);
fat_rwlock_destroy(&dp->queue_rwlock);
classifier_destroy(&dp->cls);
- hmap_destroy(&dp->flow_table);
+ cmap_destroy(&dp->flow_table);
ovs_mutex_destroy(&dp->flow_mutex);
seq_destroy(dp->port_seq);
- hmap_destroy(&dp->ports);
+ cmap_destroy(&dp->ports);
latch_destroy(&dp->exit_latch);
free(CONST_CAST(char *, dp->name));
free(dp);
/* Take dp_netdev_mutex so that, if dp->ref_cnt falls to zero, we can't
* get a new reference to 'dp' through the 'dp_netdevs' shash. */
ovs_mutex_lock(&dp_netdev_mutex);
- if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
+ if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
dp_netdev_free(dp);
}
ovs_mutex_unlock(&dp_netdev_mutex);
struct dp_netdev *dp = get_dp_netdev(dpif);
if (!atomic_flag_test_and_set(&dp->destroyed)) {
- if (ovs_refcount_unref(&dp->ref_cnt) == 1) {
+ if (ovs_refcount_unref_relaxed(&dp->ref_cnt) == 1) {
/* Can't happen: 'dpif' still owns a reference to 'dp'. */
OVS_NOT_REACHED();
}
struct dp_netdev_stats *bucket;
size_t i;
- fat_rwlock_rdlock(&dp->cls.rwlock);
- stats->n_flows = hmap_count(&dp->flow_table);
- fat_rwlock_unlock(&dp->cls.rwlock);
+ stats->n_flows = cmap_count(&dp->flow_table);
stats->n_hit = stats->n_missed = stats->n_lost = 0;
OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->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 uint32_t
+hash_port_no(odp_port_t port_no)
+{
+ return hash_int(odp_to_u32(port_no), 0);
+}
+
static int
do_add_port(struct dp_netdev *dp, const char *devname, const char *type,
odp_port_t port_no)
- OVS_REQ_WRLOCK(dp->port_rwlock)
+ OVS_REQUIRES(dp->port_mutex)
{
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;
+ int i;
/* XXX reject devices already in some dp_netdev. */
return EINVAL;
}
- error = netdev_rx_open(netdev, &rx);
- if (error
- && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
- VLOG_ERR("%s: cannot receive packets on this network device (%s)",
- devname, ovs_strerror(errno));
- netdev_close(netdev);
- return error;
+ port = xzalloc(sizeof *port);
+ port->port_no = port_no;
+ port->netdev = netdev;
+ port->rxq = xmalloc(sizeof *port->rxq * netdev_n_rxq(netdev));
+ port->type = xstrdup(type);
+ for (i = 0; i < netdev_n_rxq(netdev); i++) {
+ error = netdev_rxq_open(netdev, &port->rxq[i], i);
+ if (error
+ && !(error == EOPNOTSUPP && dpif_netdev_class_is_dummy(dp->class))) {
+ VLOG_ERR("%s: cannot receive packets on this network device (%s)",
+ devname, ovs_strerror(errno));
+ netdev_close(netdev);
+ return error;
+ }
}
error = netdev_turn_flags_on(netdev, NETDEV_PROMISC, &sf);
if (error) {
- netdev_rx_close(rx);
+ for (i = 0; i < netdev_n_rxq(netdev); i++) {
+ netdev_rxq_close(port->rxq[i]);
+ }
netdev_close(netdev);
+ free(port->rxq);
+ 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);
- hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0));
+ if (netdev_is_pmd(netdev)) {
+ dp->pmd_count++;
+ dp_netdev_set_pmd_threads(dp, NR_PMD_THREADS);
+ dp_netdev_reload_pmd_threads(dp);
+ }
+ ovs_refcount_init(&port->ref_cnt);
+
+ cmap_insert(&dp->ports, &port->node, hash_port_no(port_no));
seq_change(dp->port_seq);
return 0;
odp_port_t port_no;
int error;
- ovs_rwlock_wrlock(&dp->port_rwlock);
+ ovs_mutex_lock(&dp->port_mutex);
dpif_port = netdev_vport_get_dpif_port(netdev, namebuf, sizeof namebuf);
if (*port_nop != ODPP_NONE) {
port_no = *port_nop;
*port_nop = port_no;
error = do_add_port(dp, dpif_port, netdev_get_type(netdev), port_no);
}
- ovs_rwlock_unlock(&dp->port_rwlock);
+ ovs_mutex_unlock(&dp->port_mutex);
return error;
}
struct dp_netdev *dp = get_dp_netdev(dpif);
int error;
- ovs_rwlock_wrlock(&dp->port_rwlock);
- error = port_no == ODPP_LOCAL ? EINVAL : do_del_port(dp, port_no);
- ovs_rwlock_unlock(&dp->port_rwlock);
+ ovs_mutex_lock(&dp->port_mutex);
+ if (port_no == ODPP_LOCAL) {
+ error = EINVAL;
+ } else {
+ struct dp_netdev_port *port;
+
+ error = get_port_by_number(dp, port_no, &port);
+ if (!error) {
+ do_del_port(dp, port);
+ }
+ }
+ ovs_mutex_unlock(&dp->port_mutex);
return error;
}
static struct dp_netdev_port *
dp_netdev_lookup_port(const struct dp_netdev *dp, odp_port_t port_no)
- OVS_REQ_RDLOCK(dp->port_rwlock)
{
struct dp_netdev_port *port;
- HMAP_FOR_EACH_IN_BUCKET (port, node, hash_int(odp_to_u32(port_no), 0),
- &dp->ports) {
+ CMAP_FOR_EACH_WITH_HASH (port, node, hash_port_no(port_no), &dp->ports) {
if (port->port_no == port_no) {
return port;
}
static int
get_port_by_number(struct dp_netdev *dp,
odp_port_t port_no, struct dp_netdev_port **portp)
- OVS_REQ_RDLOCK(dp->port_rwlock)
{
if (!is_valid_port_number(port_no)) {
*portp = NULL;
}
}
+static void
+port_ref(struct dp_netdev_port *port)
+{
+ if (port) {
+ ovs_refcount_ref(&port->ref_cnt);
+ }
+}
+
+static void
+port_destroy__(struct dp_netdev_port *port)
+{
+ int n_rxq = netdev_n_rxq(port->netdev);
+ int i;
+
+ netdev_close(port->netdev);
+ netdev_restore_flags(port->sf);
+
+ for (i = 0; i < n_rxq; i++) {
+ netdev_rxq_close(port->rxq[i]);
+ }
+ free(port->rxq);
+ free(port->type);
+ free(port);
+}
+
+static void
+port_unref(struct dp_netdev_port *port)
+{
+ if (port && ovs_refcount_unref_relaxed(&port->ref_cnt) == 1) {
+ ovsrcu_postpone(port_destroy__, port);
+ }
+}
+
static int
get_port_by_name(struct dp_netdev *dp,
const char *devname, struct dp_netdev_port **portp)
- OVS_REQ_RDLOCK(dp->port_rwlock)
+ OVS_REQUIRES(dp->port_mutex)
{
struct dp_netdev_port *port;
- HMAP_FOR_EACH (port, node, &dp->ports) {
+ CMAP_FOR_EACH (port, node, &dp->ports) {
if (!strcmp(netdev_get_name(port->netdev), devname)) {
*portp = port;
return 0;
return ENOENT;
}
-static int
-do_del_port(struct dp_netdev *dp, odp_port_t port_no)
- OVS_REQ_WRLOCK(dp->port_rwlock)
+static void
+do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
+ OVS_REQUIRES(dp->port_mutex)
{
- struct dp_netdev_port *port;
- int error;
-
- error = get_port_by_number(dp, port_no, &port);
- if (error) {
- return error;
- }
-
- hmap_remove(&dp->ports, &port->node);
+ cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
seq_change(dp->port_seq);
+ if (netdev_is_pmd(port->netdev)) {
+ dp_netdev_reload_pmd_threads(dp);
+ }
- netdev_close(port->netdev);
- netdev_restore_flags(port->sf);
- netdev_rx_close(port->rx);
- free(port->type);
- free(port);
-
- return 0;
+ port_unref(port);
}
static void
struct dp_netdev_port *port;
int error;
- ovs_rwlock_rdlock(&dp->port_rwlock);
error = get_port_by_number(dp, port_no, &port);
if (!error && dpif_port) {
answer_port_query(port, dpif_port);
}
- ovs_rwlock_unlock(&dp->port_rwlock);
return error;
}
struct dp_netdev_port *port;
int error;
- ovs_rwlock_rdlock(&dp->port_rwlock);
+ ovs_mutex_lock(&dp->port_mutex);
error = get_port_by_name(dp, devname, &port);
if (!error && dpif_port) {
answer_port_query(port, dpif_port);
}
- ovs_rwlock_unlock(&dp->port_rwlock);
+ ovs_mutex_unlock(&dp->port_mutex);
return error;
}
cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
- ovs_mutex_destroy(&flow->mutex);
free(flow);
}
static void
dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
- OVS_REQ_WRLOCK(dp->cls.rwlock)
OVS_REQUIRES(dp->flow_mutex)
{
struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
- struct hmap_node *node = CONST_CAST(struct hmap_node *, &flow->node);
+ struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
classifier_remove(&dp->cls, cr);
- hmap_remove(&dp->flow_table, node);
+ cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
ovsrcu_postpone(dp_netdev_flow_free, flow);
}
static void
dp_netdev_flow_flush(struct dp_netdev *dp)
{
- struct dp_netdev_flow *netdev_flow, *next;
+ struct dp_netdev_flow *netdev_flow;
ovs_mutex_lock(&dp->flow_mutex);
- fat_rwlock_wrlock(&dp->cls.rwlock);
- HMAP_FOR_EACH_SAFE (netdev_flow, next, node, &dp->flow_table) {
+ CMAP_FOR_EACH_SAFE (netdev_flow, node, &dp->flow_table) {
dp_netdev_remove_flow(dp, netdev_flow);
}
- fat_rwlock_unlock(&dp->cls.rwlock);
ovs_mutex_unlock(&dp->flow_mutex);
}
}
struct dp_netdev_port_state {
- uint32_t bucket;
- uint32_t offset;
+ struct cmap_position position;
char *name;
};
{
struct dp_netdev_port_state *state = state_;
struct dp_netdev *dp = get_dp_netdev(dpif);
- struct hmap_node *node;
+ struct cmap_node *node;
int retval;
- ovs_rwlock_rdlock(&dp->port_rwlock);
- node = hmap_at_position(&dp->ports, &state->bucket, &state->offset);
+ node = cmap_next_position(&dp->ports, &state->position);
if (node) {
struct dp_netdev_port *port;
} else {
retval = EOF;
}
- ovs_rwlock_unlock(&dp->port_rwlock);
return retval;
}
}
static struct dp_netdev_flow *
-dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct flow *flow)
- OVS_EXCLUDED(dp->cls.rwlock)
+dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct miniflow *key)
{
struct dp_netdev_flow *netdev_flow;
+ struct cls_rule *rule;
- fat_rwlock_rdlock(&dp->cls.rwlock);
- netdev_flow = dp_netdev_flow_cast(classifier_lookup(&dp->cls, flow, NULL));
- fat_rwlock_unlock(&dp->cls.rwlock);
+ classifier_lookup_miniflow_batch(&dp->cls, &key, &rule, 1);
+ netdev_flow = dp_netdev_flow_cast(rule);
return netdev_flow;
}
static struct dp_netdev_flow *
dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
- OVS_REQ_RDLOCK(dp->cls.rwlock)
{
struct dp_netdev_flow *netdev_flow;
- HMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
+ CMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
&dp->flow_table) {
if (flow_equal(&netdev_flow->flow, flow)) {
return netdev_flow;
return EINVAL;
}
- /* Force unwildcard the in_port. */
- mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
} else {
enum mf_field_id id;
/* No mask key, unwildcard everything except fields whose
}
}
+ /* Force unwildcard the in_port.
+ *
+ * We need to do this even in the case where we unwildcard "everything"
+ * above because "everything" only includes the 16-bit OpenFlow port number
+ * mask->in_port.ofp_port, which only covers half of the 32-bit datapath
+ * port number mask->in_port.odp_port. */
+ mask->in_port.odp_port = u32_to_odp(UINT32_MAX);
+
return 0;
}
static int
dpif_netdev_flow_get(const struct dpif *dpif,
const struct nlattr *nl_key, size_t nl_key_len,
- struct ofpbuf **actionsp, struct dpif_flow_stats *stats)
+ struct ofpbuf **bufp,
+ struct nlattr **maskp, size_t *mask_len,
+ struct nlattr **actionsp, size_t *actions_len,
+ struct dpif_flow_stats *stats)
{
struct dp_netdev *dp = get_dp_netdev(dpif);
struct dp_netdev_flow *netdev_flow;
return error;
}
- fat_rwlock_rdlock(&dp->cls.rwlock);
netdev_flow = dp_netdev_find_flow(dp, &key);
- fat_rwlock_unlock(&dp->cls.rwlock);
if (netdev_flow) {
if (stats) {
get_dpif_flow_stats(netdev_flow, stats);
}
+ if (maskp) {
+ struct flow_wildcards wc;
+
+ *bufp = ofpbuf_new(sizeof(struct odputil_keybuf));
+ minimask_expand(&netdev_flow->cr.match.mask, &wc);
+ odp_flow_key_from_mask(*bufp, &wc.masks, &netdev_flow->flow,
+ odp_to_u32(wc.masks.in_port.odp_port),
+ SIZE_MAX, true);
+ *maskp = ofpbuf_data(*bufp);
+ *mask_len = ofpbuf_size(*bufp);
+ }
if (actionsp) {
struct dp_netdev_actions *actions;
actions = dp_netdev_flow_get_actions(netdev_flow);
- *actionsp = ofpbuf_clone_data(actions->actions, actions->size);
+ *actionsp = actions->actions;
+ *actions_len = actions->size;
}
} else {
error = ENOENT;
netdev_flow = xzalloc(sizeof *netdev_flow);
*CONST_CAST(struct flow *, &netdev_flow->flow) = *flow;
- ovs_mutex_init(&netdev_flow->mutex);
-
ovsthread_stats_init(&netdev_flow->stats);
ovsrcu_set(&netdev_flow->actions,
match_init(&match, flow, wc);
cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
&match, NETDEV_RULE_PRIORITY);
- fat_rwlock_wrlock(&dp->cls.rwlock);
+ cmap_insert(&dp->flow_table,
+ CONST_CAST(struct cmap_node *, &netdev_flow->node),
+ flow_hash(flow, 0));
classifier_insert(&dp->cls,
CONST_CAST(struct cls_rule *, &netdev_flow->cr));
- hmap_insert(&dp->flow_table,
- CONST_CAST(struct hmap_node *, &netdev_flow->node),
- flow_hash(flow, 0));
- fat_rwlock_unlock(&dp->cls.rwlock);
return 0;
}
struct dp_netdev *dp = get_dp_netdev(dpif);
struct dp_netdev_flow *netdev_flow;
struct flow flow;
+ struct miniflow miniflow;
struct flow_wildcards wc;
int error;
if (error) {
return error;
}
+ miniflow_init(&miniflow, &flow);
ovs_mutex_lock(&dp->flow_mutex);
- netdev_flow = dp_netdev_lookup_flow(dp, &flow);
+ netdev_flow = dp_netdev_lookup_flow(dp, &miniflow);
if (!netdev_flow) {
if (put->flags & DPIF_FP_CREATE) {
- if (hmap_count(&dp->flow_table) < MAX_FLOWS) {
+ if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
if (put->stats) {
memset(put->stats, 0, sizeof *put->stats);
}
}
}
ovs_mutex_unlock(&dp->flow_mutex);
+ miniflow_destroy(&miniflow);
return error;
}
}
ovs_mutex_lock(&dp->flow_mutex);
- fat_rwlock_wrlock(&dp->cls.rwlock);
netdev_flow = dp_netdev_find_flow(dp, &key);
if (netdev_flow) {
if (del->stats) {
} else {
error = ENOENT;
}
- fat_rwlock_unlock(&dp->cls.rwlock);
ovs_mutex_unlock(&dp->flow_mutex);
return error;
}
-struct dp_netdev_flow_state {
- struct dp_netdev_actions *actions;
- struct odputil_keybuf keybuf;
- struct odputil_keybuf maskbuf;
- struct dpif_flow_stats stats;
-};
-
-struct dp_netdev_flow_iter {
- uint32_t bucket;
- uint32_t offset;
+struct dpif_netdev_flow_dump {
+ struct dpif_flow_dump up;
+ struct cmap_position pos;
int status;
struct ovs_mutex mutex;
};
-static void
-dpif_netdev_flow_dump_state_init(void **statep)
+static struct dpif_netdev_flow_dump *
+dpif_netdev_flow_dump_cast(struct dpif_flow_dump *dump)
{
- struct dp_netdev_flow_state *state;
-
- *statep = state = xmalloc(sizeof *state);
- state->actions = NULL;
+ return CONTAINER_OF(dump, struct dpif_netdev_flow_dump, up);
}
-static void
-dpif_netdev_flow_dump_state_uninit(void *state_)
+static struct dpif_flow_dump *
+dpif_netdev_flow_dump_create(const struct dpif *dpif_)
{
- struct dp_netdev_flow_state *state = state_;
+ struct dpif_netdev_flow_dump *dump;
- free(state);
+ dump = xmalloc(sizeof *dump);
+ dpif_flow_dump_init(&dump->up, dpif_);
+ memset(&dump->pos, 0, sizeof dump->pos);
+ dump->status = 0;
+ ovs_mutex_init(&dump->mutex);
+
+ return &dump->up;
}
static int
-dpif_netdev_flow_dump_start(const struct dpif *dpif OVS_UNUSED, void **iterp)
+dpif_netdev_flow_dump_destroy(struct dpif_flow_dump *dump_)
{
- struct dp_netdev_flow_iter *iter;
+ struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
- *iterp = iter = xmalloc(sizeof *iter);
- iter->bucket = 0;
- iter->offset = 0;
- iter->status = 0;
- ovs_mutex_init(&iter->mutex);
+ ovs_mutex_destroy(&dump->mutex);
+ free(dump);
return 0;
}
-/* XXX the caller must use 'actions' without quiescing */
-static int
-dpif_netdev_flow_dump_next(const struct dpif *dpif, void *iter_, void *state_,
- const struct nlattr **key, size_t *key_len,
- const struct nlattr **mask, size_t *mask_len,
- const struct nlattr **actions, size_t *actions_len,
- const struct dpif_flow_stats **stats)
-{
- struct dp_netdev_flow_iter *iter = iter_;
- struct dp_netdev_flow_state *state = state_;
- struct dp_netdev *dp = get_dp_netdev(dpif);
- struct dp_netdev_flow *netdev_flow;
- int error;
+struct dpif_netdev_flow_dump_thread {
+ struct dpif_flow_dump_thread up;
+ struct dpif_netdev_flow_dump *dump;
+ struct odputil_keybuf keybuf[FLOW_DUMP_MAX_BATCH];
+ struct odputil_keybuf maskbuf[FLOW_DUMP_MAX_BATCH];
+};
- ovs_mutex_lock(&iter->mutex);
- error = iter->status;
- if (!error) {
- struct hmap_node *node;
+static struct dpif_netdev_flow_dump_thread *
+dpif_netdev_flow_dump_thread_cast(struct dpif_flow_dump_thread *thread)
+{
+ return CONTAINER_OF(thread, struct dpif_netdev_flow_dump_thread, up);
+}
- fat_rwlock_rdlock(&dp->cls.rwlock);
- node = hmap_at_position(&dp->flow_table, &iter->bucket, &iter->offset);
- if (node) {
- netdev_flow = CONTAINER_OF(node, struct dp_netdev_flow, node);
- }
- fat_rwlock_unlock(&dp->cls.rwlock);
- if (!node) {
- iter->status = error = EOF;
- }
- }
- ovs_mutex_unlock(&iter->mutex);
- if (error) {
- return error;
- }
+static struct dpif_flow_dump_thread *
+dpif_netdev_flow_dump_thread_create(struct dpif_flow_dump *dump_)
+{
+ struct dpif_netdev_flow_dump *dump = dpif_netdev_flow_dump_cast(dump_);
+ struct dpif_netdev_flow_dump_thread *thread;
- if (key) {
- struct ofpbuf buf;
+ thread = xmalloc(sizeof *thread);
+ dpif_flow_dump_thread_init(&thread->up, &dump->up);
+ thread->dump = dump;
+ return &thread->up;
+}
+
+static void
+dpif_netdev_flow_dump_thread_destroy(struct dpif_flow_dump_thread *thread_)
+{
+ struct dpif_netdev_flow_dump_thread *thread
+ = dpif_netdev_flow_dump_thread_cast(thread_);
- ofpbuf_use_stack(&buf, &state->keybuf, sizeof state->keybuf);
- odp_flow_key_from_flow(&buf, &netdev_flow->flow,
- netdev_flow->flow.in_port.odp_port);
+ free(thread);
+}
+
+static int
+dpif_netdev_flow_dump_next(struct dpif_flow_dump_thread *thread_,
+ struct dpif_flow *flows, int max_flows)
+{
+ struct dpif_netdev_flow_dump_thread *thread
+ = dpif_netdev_flow_dump_thread_cast(thread_);
+ struct dpif_netdev_flow_dump *dump = thread->dump;
+ struct dpif_netdev *dpif = dpif_netdev_cast(thread->up.dpif);
+ struct dp_netdev_flow *netdev_flows[FLOW_DUMP_MAX_BATCH];
+ struct dp_netdev *dp = get_dp_netdev(&dpif->dpif);
+ int n_flows = 0;
+ int i;
+
+ ovs_mutex_lock(&dump->mutex);
+ if (!dump->status) {
+ for (n_flows = 0; n_flows < MIN(max_flows, FLOW_DUMP_MAX_BATCH);
+ n_flows++) {
+ struct cmap_node *node;
- *key = buf.data;
- *key_len = buf.size;
+ node = cmap_next_position(&dp->flow_table, &dump->pos);
+ if (!node) {
+ dump->status = EOF;
+ break;
+ }
+ netdev_flows[n_flows] = CONTAINER_OF(node, struct dp_netdev_flow,
+ node);
+ }
}
+ ovs_mutex_unlock(&dump->mutex);
- if (key && mask) {
- struct ofpbuf buf;
+ for (i = 0; i < n_flows; i++) {
+ struct odputil_keybuf *maskbuf = &thread->maskbuf[i];
+ struct odputil_keybuf *keybuf = &thread->keybuf[i];
+ struct dp_netdev_flow *netdev_flow = netdev_flows[i];
+ struct dpif_flow *f = &flows[i];
+ struct dp_netdev_actions *dp_actions;
struct flow_wildcards wc;
+ struct ofpbuf buf;
- ofpbuf_use_stack(&buf, &state->maskbuf, sizeof state->maskbuf);
minimask_expand(&netdev_flow->cr.match.mask, &wc);
- odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
- odp_to_u32(wc.masks.in_port.odp_port),
- SIZE_MAX);
- *mask = buf.data;
- *mask_len = buf.size;
- }
+ /* Key. */
+ ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
+ odp_flow_key_from_flow(&buf, &netdev_flow->flow, &wc.masks,
+ netdev_flow->flow.in_port.odp_port, true);
+ f->key = ofpbuf_data(&buf);
+ f->key_len = ofpbuf_size(&buf);
- if (actions || stats) {
- state->actions = NULL;
+ /* Mask. */
+ ofpbuf_use_stack(&buf, maskbuf, sizeof *maskbuf);
+ odp_flow_key_from_mask(&buf, &wc.masks, &netdev_flow->flow,
+ odp_to_u32(wc.masks.in_port.odp_port),
+ SIZE_MAX, true);
+ f->mask = ofpbuf_data(&buf);
+ f->mask_len = ofpbuf_size(&buf);
- if (actions) {
- state->actions = dp_netdev_flow_get_actions(netdev_flow);
- *actions = state->actions->actions;
- *actions_len = state->actions->size;
- }
+ /* Actions. */
+ dp_actions = dp_netdev_flow_get_actions(netdev_flow);
+ f->actions = dp_actions->actions;
+ f->actions_len = dp_actions->size;
- if (stats) {
- get_dpif_flow_stats(netdev_flow, &state->stats);
- *stats = &state->stats;
- }
+ /* Stats. */
+ get_dpif_flow_stats(netdev_flow, &f->stats);
}
- return 0;
-}
-
-static int
-dpif_netdev_flow_dump_done(const struct dpif *dpif OVS_UNUSED, void *iter_)
-{
- struct dp_netdev_flow_iter *iter = iter_;
-
- ovs_mutex_destroy(&iter->mutex);
- free(iter);
- return 0;
+ return n_flows;
}
static int
dpif_netdev_execute(struct dpif *dpif, struct dpif_execute *execute)
{
struct dp_netdev *dp = get_dp_netdev(dpif);
+ struct dpif_packet packet, *pp;
struct pkt_metadata *md = &execute->md;
- struct flow key;
- if (execute->packet->size < ETH_HEADER_LEN ||
- execute->packet->size > UINT16_MAX) {
+ if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
+ ofpbuf_size(execute->packet) > UINT16_MAX) {
return EINVAL;
}
- /* Extract flow key. */
- flow_extract(execute->packet, md, &key);
+ packet.ofpbuf = *execute->packet;
+ pp = &packet;
- ovs_rwlock_rdlock(&dp->port_rwlock);
- dp_netdev_execute_actions(dp, &key, execute->packet, false, md,
+ dp_netdev_execute_actions(dp, &pp, 1, false, md,
execute->actions, execute->actions_len);
- ovs_rwlock_unlock(&dp->port_rwlock);
+
+ /* Even though may_steal is set to false, some actions could modify or
+ * reallocate the ofpbuf memory. We need to pass those changes to the
+ * caller */
+ *execute->packet = packet.ofpbuf;
return 0;
}
free(actions);
}
\f
-static void *
-dp_forwarder_main(void *f_)
+
+static void
+dp_netdev_process_rxq_port(struct dp_netdev *dp,
+ struct dp_netdev_port *port,
+ struct netdev_rxq *rxq)
{
- struct dp_forwarder *f = f_;
- struct dp_netdev *dp = f->dp;
+ struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
+ int error, cnt;
- f->name = xasprintf("forwarder_%u", ovsthread_id_self());
- set_subprogram_name("%s", f->name);
+ error = netdev_rxq_recv(rxq, packets, &cnt);
+ if (!error) {
+ dp_netdev_port_input(dp, packets, cnt, port->port_no);
+ } else if (error != EAGAIN && error != EOPNOTSUPP) {
+ static struct vlog_rate_limit rl
+ = VLOG_RATE_LIMIT_INIT(1, 5);
- while (!latch_is_set(&dp->exit_latch)) {
- bool received_anything;
- int i;
+ VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
+ netdev_get_name(port->netdev),
+ ovs_strerror(error));
+ }
+}
- 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));
- }
- }
+static void
+dpif_netdev_run(struct dpif *dpif)
+{
+ struct dp_netdev_port *port;
+ struct dp_netdev *dp = get_dp_netdev(dpif);
+
+ CMAP_FOR_EACH (port, node, &dp->ports) {
+ if (!netdev_is_pmd(port->netdev)) {
+ int i;
+
+ for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
+ dp_netdev_process_rxq_port(dp, port, port->rxq[i]);
}
+ }
+ }
+}
- if (!received_anything) {
- break;
+static void
+dpif_netdev_wait(struct dpif *dpif)
+{
+ struct dp_netdev_port *port;
+ struct dp_netdev *dp = get_dp_netdev(dpif);
+
+ ovs_mutex_lock(&dp_netdev_mutex);
+ CMAP_FOR_EACH (port, node, &dp->ports) {
+ if (!netdev_is_pmd(port->netdev)) {
+ int i;
+
+ for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
+ netdev_rxq_wait(port->rxq[i]);
}
}
+ }
+ ovs_mutex_unlock(&dp_netdev_mutex);
+}
- if (received_anything) {
- poll_immediate_wake();
- } else {
- struct dp_netdev_port *port;
+struct rxq_poll {
+ struct dp_netdev_port *port;
+ struct netdev_rxq *rx;
+};
+
+static int
+pmd_load_queues(struct pmd_thread *f,
+ struct rxq_poll **ppoll_list, int poll_cnt)
+{
+ struct dp_netdev *dp = f->dp;
+ struct rxq_poll *poll_list = *ppoll_list;
+ struct dp_netdev_port *port;
+ int id = f->id;
+ int index;
+ int i;
+
+ /* Simple scheduler for netdev rx polling. */
+ for (i = 0; i < poll_cnt; i++) {
+ port_unref(poll_list[i].port);
+ }
+
+ poll_cnt = 0;
+ index = 0;
+
+ CMAP_FOR_EACH (port, node, &f->dp->ports) {
+ if (netdev_is_pmd(port->netdev)) {
+ int i;
- 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);
+ for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
+ 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;
+ poll_list[poll_cnt].rx = port->rxq[i];
+ poll_cnt++;
}
- seq_wait(dp->port_seq, seq_read(dp->port_seq));
- latch_wait(&dp->exit_latch);
+ index++;
+ }
+ }
+ }
+
+ *ppoll_list = poll_list;
+ return poll_cnt;
+}
+
+static void *
+pmd_thread_main(void *f_)
+{
+ struct pmd_thread *f = f_;
+ struct dp_netdev *dp = f->dp;
+ unsigned int lc = 0;
+ struct rxq_poll *poll_list;
+ unsigned int port_seq;
+ int poll_cnt;
+ int i;
+
+ poll_cnt = 0;
+ poll_list = NULL;
+
+ pmd_thread_setaffinity_cpu(f->id);
+reload:
+ poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
+ atomic_read(&f->change_seq, &port_seq);
+
+ for (;;) {
+ unsigned int c_port_seq;
+ int i;
+
+ for (i = 0; i < poll_cnt; i++) {
+ dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
+ }
+
+ if (lc++ > 1024) {
+ ovsrcu_quiesce();
+
+ /* 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;
+ }
}
- ovs_rwlock_unlock(&dp->port_rwlock);
+ }
- poll_block();
+ if (!latch_is_set(&f->dp->exit_latch)){
+ goto reload;
}
- free(f->name);
+ for (i = 0; i < poll_cnt; i++) {
+ port_unref(poll_list[i].port);
+ }
+ free(poll_list);
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. */
+ f->thread = ovs_thread_create("pmd", pmd_thread_main, f);
}
}
+
\f
static void *
dp_netdev_flow_stats_new_cb(void)
static void
dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
- const struct ofpbuf *packet,
- const struct flow *key)
+ int cnt, int size,
+ uint16_t tcp_flags)
{
- uint16_t tcp_flags = ntohs(key->tcp_flags);
long long int now = time_msec();
struct dp_netdev_flow_stats *bucket;
ovs_mutex_lock(&bucket->mutex);
bucket->used = MAX(now, bucket->used);
- bucket->packet_count++;
- bucket->byte_count += packet->size;
+ bucket->packet_count += cnt;
+ bucket->byte_count += size;
bucket->tcp_flags |= tcp_flags;
ovs_mutex_unlock(&bucket->mutex);
}
}
static void
-dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type)
+dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
{
struct dp_netdev_stats *bucket;
bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
ovs_mutex_lock(&bucket->mutex);
- bucket->n[type]++;
+ bucket->n[type] += cnt;
ovs_mutex_unlock(&bucket->mutex);
}
-static void
-dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
- struct pkt_metadata *md)
- OVS_REQ_RDLOCK(dp->port_rwlock)
+struct packet_batch {
+ unsigned int packet_count;
+ unsigned int byte_count;
+ uint16_t tcp_flags;
+
+ struct dp_netdev_flow *flow;
+
+ struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
+ struct pkt_metadata md;
+};
+
+static inline void
+packet_batch_update(struct packet_batch *batch,
+ struct dpif_packet *packet, const struct miniflow *mf)
{
- struct dp_netdev_flow *netdev_flow;
- struct flow key;
+ batch->tcp_flags |= miniflow_get_tcp_flags(mf);
+ batch->packets[batch->packet_count++] = packet;
+ batch->byte_count += ofpbuf_size(&packet->ofpbuf);
+}
- if (packet->size < ETH_HEADER_LEN) {
- ofpbuf_delete(packet);
- return;
+static inline void
+packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow,
+ struct pkt_metadata *md)
+{
+ batch->flow = flow;
+ batch->md = *md;
+
+ batch->packet_count = 0;
+ batch->byte_count = 0;
+ batch->tcp_flags = 0;
+}
+
+static inline void
+packet_batch_execute(struct packet_batch *batch, struct dp_netdev *dp)
+{
+ struct dp_netdev_actions *actions;
+ struct dp_netdev_flow *flow = batch->flow;
+
+ dp_netdev_flow_used(batch->flow, batch->packet_count, batch->byte_count,
+ batch->tcp_flags);
+
+ actions = dp_netdev_flow_get_actions(flow);
+
+ dp_netdev_execute_actions(dp, batch->packets,
+ batch->packet_count, true, &batch->md,
+ actions->actions, actions->size);
+
+ dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
+}
+
+static void
+dp_netdev_input(struct dp_netdev *dp, struct dpif_packet **packets, int cnt,
+ struct pkt_metadata *md)
+{
+ struct packet_batch batches[NETDEV_MAX_RX_BATCH];
+ struct netdev_flow_key keys[NETDEV_MAX_RX_BATCH];
+ const struct miniflow *mfs[NETDEV_MAX_RX_BATCH]; /* NULL at bad packets. */
+ struct cls_rule *rules[NETDEV_MAX_RX_BATCH];
+ size_t n_batches, i;
+
+ for (i = 0; i < cnt; i++) {
+ if (OVS_UNLIKELY(ofpbuf_size(&packets[i]->ofpbuf) < ETH_HEADER_LEN)) {
+ dpif_packet_delete(packets[i]);
+ mfs[i] = NULL;
+ continue;
+ }
+
+ miniflow_initialize(&keys[i].flow, keys[i].buf);
+ miniflow_extract(&packets[i]->ofpbuf, md, &keys[i].flow);
+ mfs[i] = &keys[i].flow;
}
- flow_extract(packet, md, &key);
- netdev_flow = dp_netdev_lookup_flow(dp, &key);
- if (netdev_flow) {
- struct dp_netdev_actions *actions;
- dp_netdev_flow_used(netdev_flow, packet, &key);
+ classifier_lookup_miniflow_batch(&dp->cls, mfs, rules, cnt);
- actions = dp_netdev_flow_get_actions(netdev_flow);
- dp_netdev_execute_actions(dp, &key, packet, true, md,
- actions->actions, actions->size);
- dp_netdev_count_packet(dp, DP_STAT_HIT);
- } else if (dp->handler_queues) {
- dp_netdev_count_packet(dp, DP_STAT_MISS);
- dp_netdev_output_userspace(dp, packet,
- flow_hash_5tuple(&key, 0) % dp->n_handlers,
- DPIF_UC_MISS, &key, NULL);
- ofpbuf_delete(packet);
+ n_batches = 0;
+ for (i = 0; i < cnt; i++) {
+ struct dp_netdev_flow *flow;
+ struct packet_batch *batch;
+ size_t j;
+
+ if (OVS_UNLIKELY(!mfs[i])) {
+ continue;
+ }
+
+ if (OVS_UNLIKELY(!rules[i])) {
+
+ dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
+
+ if (OVS_LIKELY(dp->handler_queues)) {
+ uint32_t hash = miniflow_hash_5tuple(mfs[i], 0);
+ struct ofpbuf *buf = &packets[i]->ofpbuf;
+
+ dp_netdev_output_userspace(dp, buf, hash % dp->n_handlers,
+ DPIF_UC_MISS, mfs[i], NULL);
+ }
+
+ dpif_packet_delete(packets[i]);
+ continue;
+ }
+
+ /* XXX: This O(n^2) algortihm makes sense if we're operating under the
+ * assumption that the number of distinct flows (and therefore the
+ * number of distinct batches) is quite small. If this turns out not
+ * to be the case, it may make sense to pre sort based on the
+ * netdev_flow pointer. That done we can get the appropriate batching
+ * in O(n * log(n)) instead. */
+ batch = NULL;
+ flow = dp_netdev_flow_cast(rules[i]);
+ for (j = 0; j < n_batches; j++) {
+ if (batches[j].flow == flow) {
+ batch = &batches[j];
+ break;
+ }
+ }
+
+ if (!batch) {
+ batch = &batches[n_batches++];
+ packet_batch_init(batch, flow, md);
+ }
+ packet_batch_update(batch, packets[i], mfs[i]);
+ }
+
+ for (i = 0; i < n_batches; i++) {
+ packet_batch_execute(&batches[i], dp);
}
}
-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)
+static void
+dp_netdev_port_input(struct dp_netdev *dp, struct dpif_packet **packets,
+ int cnt, odp_port_t port_no)
{
- struct dp_netdev_queue *q;
- int error;
+ uint32_t *recirc_depth = recirc_depth_get();
+ struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
- fat_rwlock_rdlock(&dp->queue_rwlock);
- q = &dp->handler_queues[queue_no];
- ovs_mutex_lock(&q->mutex);
+ *recirc_depth = 0;
+ dp_netdev_input(dp, packets, cnt, &md);
+}
+
+static int
+dp_netdev_queue_userspace_packet(struct dp_netdev_queue *q,
+ struct ofpbuf *packet, int type,
+ const struct miniflow *key,
+ const struct nlattr *userdata)
+OVS_REQUIRES(q->mutex)
+{
if (q->head - q->tail < MAX_QUEUE_LEN) {
struct dp_netdev_upcall *u = &q->upcalls[q->head++ & QUEUE_MASK];
struct dpif_upcall *upcall = &u->upcall;
struct ofpbuf *buf = &u->buf;
size_t buf_size;
+ struct flow flow;
+ void *data;
upcall->type = type;
if (userdata) {
buf_size += NLA_ALIGN(userdata->nla_len);
}
- buf_size += packet->size;
+ buf_size += ofpbuf_size(packet);
ofpbuf_init(buf, buf_size);
/* Put ODP flow. */
- odp_flow_key_from_flow(buf, flow, flow->in_port.odp_port);
- upcall->key = buf->data;
- upcall->key_len = buf->size;
+ miniflow_expand(key, &flow);
+ odp_flow_key_from_flow(buf, &flow, NULL, flow.in_port.odp_port, true);
+ upcall->key = ofpbuf_data(buf);
+ upcall->key_len = ofpbuf_size(buf);
/* Put userdata. */
if (userdata) {
upcall->userdata = ofpbuf_put(buf, userdata,
- NLA_ALIGN(userdata->nla_len));
+ NLA_ALIGN(userdata->nla_len));
}
- upcall->packet.data = ofpbuf_put(buf, packet->data, packet->size);
- upcall->packet.size = packet->size;
+ /* We have to perform a copy of the packet, because we cannot send DPDK
+ * mbufs to a non pmd thread. When the upcall processing will be done
+ * in the pmd thread, this copy can be avoided */
+ data = ofpbuf_put(buf, ofpbuf_data(packet), ofpbuf_size(packet));
+ ofpbuf_use_stub(&upcall->packet, data, ofpbuf_size(packet));
+ ofpbuf_set_size(&upcall->packet, ofpbuf_size(packet));
seq_change(q->seq);
- error = 0;
+ return 0;
} else {
- dp_netdev_count_packet(dp, DP_STAT_LOST);
- error = ENOBUFS;
+ return ENOBUFS;
+ }
+}
+
+static int
+dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet,
+ int queue_no, int type,
+ const struct miniflow *key,
+ const struct nlattr *userdata)
+{
+ struct dp_netdev_queue *q;
+ int error;
+
+ fat_rwlock_rdlock(&dp->queue_rwlock);
+ q = &dp->handler_queues[queue_no];
+ ovs_mutex_lock(&q->mutex);
+ error = dp_netdev_queue_userspace_packet(q, packet, type, key,
+ userdata);
+ if (error == ENOBUFS) {
+ dp_netdev_count_packet(dp, DP_STAT_LOST, 1);
}
ovs_mutex_unlock(&q->mutex);
fat_rwlock_unlock(&dp->queue_rwlock);
struct dp_netdev_execute_aux {
struct dp_netdev *dp;
- const struct flow *key;
};
static void
-dp_execute_cb(void *aux_, struct ofpbuf *packet,
- const struct pkt_metadata *md OVS_UNUSED,
+dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
+ struct pkt_metadata *md,
const struct nlattr *a, bool may_steal)
OVS_NO_THREAD_SAFETY_ANALYSIS
{
struct dp_netdev_execute_aux *aux = aux_;
int type = nl_attr_type(a);
struct dp_netdev_port *p;
+ uint32_t *depth = recirc_depth_get();
+ int i;
switch ((enum ovs_action_attr)type) {
case OVS_ACTION_ATTR_OUTPUT:
p = dp_netdev_lookup_port(aux->dp, u32_to_odp(nl_attr_get_u32(a)));
- if (p) {
- netdev_send(p->netdev, packet, may_steal);
+ if (OVS_LIKELY(p)) {
+ netdev_send(p->netdev, packets, cnt, may_steal);
+ } else if (may_steal) {
+ for (i = 0; i < cnt; i++) {
+ dpif_packet_delete(packets[i]);
+ }
}
break;
case OVS_ACTION_ATTR_USERSPACE: {
const struct nlattr *userdata;
+ struct netdev_flow_key key;
userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
- dp_netdev_output_userspace(aux->dp, packet,
- flow_hash_5tuple(aux->key, 0)
- % aux->dp->n_handlers,
- DPIF_UC_ACTION, aux->key,
- userdata);
+ miniflow_initialize(&key.flow, key.buf);
+
+ for (i = 0; i < cnt; i++) {
+ struct ofpbuf *packet;
+
+ packet = &packets[i]->ofpbuf;
+
+ miniflow_extract(packet, md, &key.flow);
+
+ dp_netdev_output_userspace(aux->dp, packet,
+ miniflow_hash_5tuple(&key.flow, 0)
+ % aux->dp->n_handlers,
+ DPIF_UC_ACTION, &key.flow,
+ userdata);
+ if (may_steal) {
+ dpif_packet_delete(packets[i]);
+ }
+ }
+ break;
+ }
+
+ case OVS_ACTION_ATTR_HASH: {
+ const struct ovs_action_hash *hash_act;
+ struct netdev_flow_key key;
+ uint32_t hash;
+
+ hash_act = nl_attr_get(a);
+
+ miniflow_initialize(&key.flow, key.buf);
- if (may_steal) {
- ofpbuf_delete(packet);
+ for (i = 0; i < cnt; i++) {
+
+ /* TODO: this is slow. Use RSS hash in the future */
+ miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
+
+ if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
+ /* Hash need not be symmetric, nor does it need to include
+ * L2 fields. */
+ hash = miniflow_hash_5tuple(&key.flow, hash_act->hash_basis);
+ } else {
+ VLOG_WARN("Unknown hash algorithm specified "
+ "for the hash action.");
+ hash = 2;
+ }
+
+ if (!hash) {
+ hash = 1; /* 0 is not valid */
+ }
+
+ if (i == 0) {
+ md->dp_hash = hash;
+ }
+ packets[i]->dp_hash = hash;
}
break;
}
+
+ case OVS_ACTION_ATTR_RECIRC:
+ if (*depth < MAX_RECIRC_DEPTH) {
+
+ (*depth)++;
+ for (i = 0; i < cnt; i++) {
+ struct dpif_packet *recirc_pkt;
+ struct pkt_metadata recirc_md = *md;
+
+ recirc_pkt = (may_steal) ? packets[i]
+ : dpif_packet_clone(packets[i]);
+
+ recirc_md.recirc_id = nl_attr_get_u32(a);
+
+ /* Hash is private to each packet */
+ recirc_md.dp_hash = packets[i]->dp_hash;
+
+ dp_netdev_input(aux->dp, &recirc_pkt, 1, &recirc_md);
+ }
+ (*depth)--;
+
+ break;
+ } else {
+ VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
+ if (may_steal) {
+ for (i = 0; i < cnt; i++) {
+ dpif_packet_delete(packets[i]);
+ }
+ }
+ }
+ break;
+
case OVS_ACTION_ATTR_PUSH_VLAN:
case OVS_ACTION_ATTR_POP_VLAN:
case OVS_ACTION_ATTR_PUSH_MPLS:
case __OVS_ACTION_ATTR_MAX:
OVS_NOT_REACHED();
}
-
}
static void
-dp_netdev_execute_actions(struct dp_netdev *dp, const struct flow *key,
- struct ofpbuf *packet, bool may_steal,
- struct pkt_metadata *md,
+dp_netdev_execute_actions(struct dp_netdev *dp,
+ struct dpif_packet **packets, int cnt,
+ 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};
+ struct dp_netdev_execute_aux aux = {dp};
- odp_execute_actions(&aux, packet, may_steal, md,
- actions, actions_len, dp_execute_cb);
+ odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
+ actions_len, dp_execute_cb);
}
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,
dpif_netdev_flow_put,
dpif_netdev_flow_del,
dpif_netdev_flow_flush,
- dpif_netdev_flow_dump_state_init,
- dpif_netdev_flow_dump_start,
+ dpif_netdev_flow_dump_create,
+ dpif_netdev_flow_dump_destroy,
+ dpif_netdev_flow_dump_thread_create,
+ dpif_netdev_flow_dump_thread_destroy,
dpif_netdev_flow_dump_next,
- NULL,
- dpif_netdev_flow_dump_done,
- dpif_netdev_flow_dump_state_uninit,
dpif_netdev_execute,
NULL, /* operate */
dpif_netdev_recv_set,
dpif_dummy_change_port_number(struct unixctl_conn *conn, int argc OVS_UNUSED,
const char *argv[], void *aux OVS_UNUSED)
{
- struct dp_netdev_port *port;
+ struct dp_netdev_port *old_port;
+ struct dp_netdev_port *new_port;
struct dp_netdev *dp;
odp_port_t port_no;
ovs_refcount_ref(&dp->ref_cnt);
ovs_mutex_unlock(&dp_netdev_mutex);
- ovs_rwlock_wrlock(&dp->port_rwlock);
- if (get_port_by_name(dp, argv[2], &port)) {
+ ovs_mutex_lock(&dp->port_mutex);
+ if (get_port_by_name(dp, argv[2], &old_port)) {
unixctl_command_reply_error(conn, "unknown port");
goto exit;
}
unixctl_command_reply_error(conn, "port number already in use");
goto exit;
}
- hmap_remove(&dp->ports, &port->node);
- port->port_no = port_no;
- hmap_insert(&dp->ports, &port->node, hash_int(odp_to_u32(port_no), 0));
+
+ /* Remove old port. */
+ cmap_remove(&dp->ports, &old_port->node, hash_port_no(old_port->port_no));
+ ovsrcu_postpone(free, old_port);
+
+ /* Insert new port (cmap semantics mean we cannot re-insert 'old_port'). */
+ new_port = xmemdup(old_port, sizeof *old_port);
+ new_port->port_no = port_no;
+ cmap_insert(&dp->ports, &new_port->node, hash_port_no(port_no));
+
seq_change(dp->port_seq);
unixctl_command_reply(conn, NULL);
exit:
- ovs_rwlock_unlock(&dp->port_rwlock);
+ ovs_mutex_unlock(&dp->port_mutex);
+ dp_netdev_unref(dp);
+}
+
+static void
+dpif_dummy_delete_port(struct unixctl_conn *conn, int argc OVS_UNUSED,
+ const char *argv[], void *aux OVS_UNUSED)
+{
+ struct dp_netdev_port *port;
+ struct dp_netdev *dp;
+
+ ovs_mutex_lock(&dp_netdev_mutex);
+ dp = shash_find_data(&dp_netdevs, argv[1]);
+ if (!dp || !dpif_netdev_class_is_dummy(dp->class)) {
+ ovs_mutex_unlock(&dp_netdev_mutex);
+ unixctl_command_reply_error(conn, "unknown datapath or not a dummy");
+ return;
+ }
+ ovs_refcount_ref(&dp->ref_cnt);
+ ovs_mutex_unlock(&dp_netdev_mutex);
+
+ ovs_mutex_lock(&dp->port_mutex);
+ if (get_port_by_name(dp, argv[2], &port)) {
+ unixctl_command_reply_error(conn, "unknown port");
+ } else if (port->port_no == ODPP_LOCAL) {
+ unixctl_command_reply_error(conn, "can't delete local port");
+ } else {
+ do_del_port(dp, port);
+ unixctl_command_reply(conn, NULL);
+ }
+ ovs_mutex_unlock(&dp->port_mutex);
+
dp_netdev_unref(dp);
}
unixctl_command_register("dpif-dummy/change-port-number",
"DP PORT NEW-NUMBER",
3, 3, dpif_dummy_change_port_number, NULL);
+ unixctl_command_register("dpif-dummy/delete-port", "DP PORT",
+ 2, 2, dpif_dummy_delete_port, NULL);
}
projects / cascardo / ovs.git / blobdiff