*/
#include <config.h>
-#include "dpif.h"
+#include "dpif-netdev.h"
#include <ctype.h>
#include <errno.h>
#include <sys/stat.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 "match.h"
#include "meta-flow.h"
#include "netdev.h"
+#include "netdev-dpdk.h"
#include "netdev-vport.h"
#include "netlink.h"
#include "odp-execute.h"
#include "odp-util.h"
#include "ofp-print.h"
#include "ofpbuf.h"
+#include "ovs-numa.h"
+#include "ovs-rcu.h"
+#include "packet-dpif.h"
#include "packets.h"
#include "poll-loop.h"
+#include "pvector.h"
#include "random.h"
#include "seq.h"
#include "shash.h"
VLOG_DEFINE_THIS_MODULE(dpif_netdev);
-/* 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. */
-/* Enough headroom to add a vlan tag, plus an extra 2 bytes to allow IP
- * headers to be aligned on a 4-byte boundary. */
-enum { DP_NETDEV_HEADROOM = 2 + VLAN_HEADER_LEN };
-
-/* Queues. */
-enum { N_QUEUES = 2 }; /* Number of queues for dpif_recv(). */
-enum { MAX_QUEUE_LEN = 128 }; /* Maximum number of packets per queue. */
-enum { QUEUE_MASK = MAX_QUEUE_LEN - 1 };
-BUILD_ASSERT_DECL(IS_POW2(MAX_QUEUE_LEN));
-
/* Protects against changes to 'dp_netdevs'. */
static struct ovs_mutex dp_netdev_mutex = OVS_MUTEX_INITIALIZER;
static struct shash dp_netdevs OVS_GUARDED_BY(dp_netdev_mutex)
= SHASH_INITIALIZER(&dp_netdevs);
-struct dp_netdev_upcall {
- struct dpif_upcall upcall; /* Queued upcall information. */
- struct ofpbuf buf; /* ofpbuf instance for upcall.packet. */
+static struct vlog_rate_limit upcall_rl = VLOG_RATE_LIMIT_INIT(600, 600);
+
+/* Stores a miniflow with inline values */
+
+struct netdev_flow_key {
+ uint32_t hash; /* Hash function differs for different users. */
+ uint32_t len; /* Length of the following miniflow (incl. map). */
+ struct miniflow mf;
+ uint32_t buf[FLOW_MAX_PACKET_U32S - MINI_N_INLINE];
};
-/* A queue passing packets from a struct dp_netdev to its clients.
+/* Exact match cache for frequently used flows
+ *
+ * The cache uses a 32-bit hash of the packet (which can be the RSS hash) to
+ * search its entries for a miniflow that matches exactly the miniflow of the
+ * packet. It stores the 'dpcls_rule' (rule) that matches the miniflow.
+ *
+ * A cache entry holds a reference to its 'dp_netdev_flow'.
+ *
+ * A miniflow with a given hash can be in one of EM_FLOW_HASH_SEGS different
+ * entries. The 32-bit hash is split into EM_FLOW_HASH_SEGS values (each of
+ * them is EM_FLOW_HASH_SHIFT bits wide and the remainder is thrown away). Each
+ * value is the index of a cache entry where the miniflow could be.
*
*
* Thread-safety
* =============
*
- * Any access at all requires the owning 'dp_netdev''s queue_mutex. */
-struct dp_netdev_queue {
- struct dp_netdev_upcall upcalls[MAX_QUEUE_LEN] OVS_GUARDED;
- unsigned int head OVS_GUARDED;
- unsigned int tail OVS_GUARDED;
+ * Each pmd_thread has its own private exact match cache.
+ * If dp_netdev_input is not called from a pmd thread, a mutex is used.
+ */
+
+#define EM_FLOW_HASH_SHIFT 10
+#define EM_FLOW_HASH_ENTRIES (1u << EM_FLOW_HASH_SHIFT)
+#define EM_FLOW_HASH_MASK (EM_FLOW_HASH_ENTRIES - 1)
+#define EM_FLOW_HASH_SEGS 2
+
+struct emc_entry {
+ struct dp_netdev_flow *flow;
+ struct netdev_flow_key key; /* key.hash used for emc hash value. */
+};
+
+struct emc_cache {
+ struct emc_entry entries[EM_FLOW_HASH_ENTRIES];
+};
+
+/* Iterate in the exact match cache through every entry that might contain a
+ * miniflow with hash 'HASH'. */
+#define EMC_FOR_EACH_POS_WITH_HASH(EMC, CURRENT_ENTRY, HASH) \
+ for (uint32_t i__ = 0, srch_hash__ = (HASH); \
+ (CURRENT_ENTRY) = &(EMC)->entries[srch_hash__ & EM_FLOW_HASH_MASK], \
+ i__ < EM_FLOW_HASH_SEGS; \
+ i__++, srch_hash__ >>= EM_FLOW_HASH_SHIFT)
+\f
+/* Simple non-wildcarding single-priority classifier. */
+
+struct dpcls {
+ struct cmap subtables_map;
+ struct pvector subtables;
+};
+
+/* A rule to be inserted to the classifier. */
+struct dpcls_rule {
+ struct cmap_node cmap_node; /* Within struct dpcls_subtable 'rules'. */
+ struct netdev_flow_key *mask; /* Subtable's mask. */
+ struct netdev_flow_key flow; /* Matching key. */
+ /* 'flow' must be the last field, additional space is allocated here. */
};
+static void dpcls_init(struct dpcls *);
+static void dpcls_destroy(struct dpcls *);
+static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
+ const struct netdev_flow_key *mask);
+static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
+static bool dpcls_lookup(const struct dpcls *cls,
+ const struct netdev_flow_key keys[],
+ struct dpcls_rule **rules, size_t cnt);
+\f
/* Datapath based on the network device interface from netdev.h.
*
*
* Acquisition order is, from outermost to innermost:
*
* dp_netdev_mutex (global)
- * port_rwlock
+ * port_mutex
* flow_mutex
- * cls.rwlock
- * queue_mutex
*/
struct dp_netdev {
const struct dpif_class *const class;
const char *const name;
+ struct dpif *dpif;
struct ovs_refcount ref_cnt;
atomic_flag destroyed;
/* 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. */
-
- /* Queues.
- *
- * Everything in 'queues' is protected by 'queue_mutex'. */
- struct ovs_mutex queue_mutex;
- struct dp_netdev_queue queues[N_QUEUES];
- struct seq *queue_seq; /* Incremented whenever a packet is queued. */
+ struct dpcls cls;
+ struct cmap flow_table OVS_GUARDED; /* Flow table. */
/* Statistics.
*
- * ovsthread_counter is internally synchronized. */
- struct ovsthread_counter *n_hit; /* Number of flow table matches. */
- struct ovsthread_counter *n_missed; /* Number of flow table misses. */
- struct ovsthread_counter *n_lost; /* Number of misses not passed up. */
+ * ovsthread_stats is internally synchronized. */
+ struct ovsthread_stats stats; /* Contains 'struct dp_netdev_stats *'. */
/* 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;
+ /* Protects access to ofproto-dpif-upcall interface during revalidator
+ * thread synchronization. */
+ struct fat_rwlock upcall_rwlock;
+ upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
+ void *upcall_aux;
+
+ /* Stores all 'struct dp_netdev_pmd_thread's. */
+ struct cmap poll_threads;
+
+ /* Protects the access of the 'struct dp_netdev_pmd_thread'
+ * instance for non-pmd thread. */
+ struct ovs_mutex non_pmd_mutex;
+
+ /* Each pmd thread will store its pointer to
+ * 'struct dp_netdev_pmd_thread' in 'per_pmd_key'. */
+ ovsthread_key_t per_pmd_key;
+
+ /* Number of rx queues for each dpdk interface and the cpu mask
+ * for pin of pmd threads. */
+ size_t n_dpdk_rxqs;
+ char *pmd_cmask;
};
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. */
+ DP_STAT_MISS, /* Packets that did not match. */
+ DP_STAT_LOST, /* Packets not passed up to the client. */
+ DP_N_STATS
+};
+
+/* Contained by struct dp_netdev's 'stats' member. */
+struct dp_netdev_stats {
+ struct ovs_mutex mutex; /* Protects 'n'. */
+
+ /* Indexed by DP_STAT_*, protected by 'mutex'. */
+ unsigned long long int n[DP_N_STATS] OVS_GUARDED;
+};
+
/* 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. */
};
+\f
/* A flow in dp_netdev's 'flow_table'.
*
*
* Rules
* -----
*
- * A flow 'flow' may be accessed without a risk of being freed by code that
- * holds a read-lock or write-lock on 'cls->rwlock' or that owns a reference to
- * 'flow->ref_cnt' (or both). Code that needs to hold onto a flow for a while
- * should take 'cls->rwlock', find the flow it needs, increment 'flow->ref_cnt'
- * with dpif_netdev_flow_ref(), and drop 'cls->rwlock'.
+ * A flow 'flow' may be accessed without a risk of being freed during an RCU
+ * grace period. Code that needs to hold onto a flow for a while
+ * should try incrementing 'flow->ref_cnt' with dp_netdev_flow_ref().
*
* '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').
+ * flow from being deleted from 'cls' and it doesn't protect members of 'flow'
+ * from modification.
*
* Some members, marked 'const', are immutable. Accessing other members
* requires synchronization, as noted in more detail below.
*/
struct dp_netdev_flow {
- /* Packet classification. */
- const struct cls_rule cr; /* In owning dp_netdev's 'cls'. */
+ bool dead;
/* Hash table index by unmasked flow. */
- const struct hmap_node node; /* In owning dp_netdev's 'flow_table'. */
- const struct flow flow; /* The flow that created this entry. */
+ const struct cmap_node node; /* In owning dp_netdev's 'flow_table'. */
+ const struct flow flow; /* Unmasked flow that created this entry. */
/* Number of references.
* The classifier owns one reference.
* reference. */
struct ovs_refcount ref_cnt;
- /* 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. */
+ OVSRCU_TYPE(struct dp_netdev_actions *) actions;
+
+ /* Packet classification. */
+ struct dpcls_rule cr; /* In owning dp_netdev's 'cls'. */
+ /* 'cr' must be the last member. */
+};
+
+static void dp_netdev_flow_unref(struct dp_netdev_flow *);
+static bool dp_netdev_flow_ref(struct dp_netdev_flow *);
+
+/* Contained by struct dp_netdev_flow's 'stats' member. */
+struct dp_netdev_flow_stats {
+ struct ovs_mutex mutex; /* Guards all the other members. */
+
long long int used OVS_GUARDED; /* Last used time, in monotonic msecs. */
long long int packet_count OVS_GUARDED; /* Number of packets matched. */
long long int byte_count OVS_GUARDED; /* Number of bytes matched. */
uint16_t tcp_flags OVS_GUARDED; /* Bitwise-OR of seen tcp_flags values. */
-
- /* Actions.
- *
- * Reading 'actions' requires 'mutex'.
- * Writing 'actions' requires 'mutex' and (to allow for transactions) the
- * datapath's flow_mutex. */
- struct dp_netdev_actions *actions OVS_GUARDED;
};
-static struct dp_netdev_flow *dp_netdev_flow_ref(
- const struct dp_netdev_flow *);
-static void dp_netdev_flow_unref(struct dp_netdev_flow *);
-
/* A set of datapath actions within a "struct dp_netdev_flow".
*
*
* 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 {
- struct ovs_refcount ref_cnt;
-
/* These members are immutable: they do not change during the struct's
* lifetime. */
struct nlattr *actions; /* Sequence of OVS_ACTION_ATTR_* attributes. */
struct dp_netdev_actions *dp_netdev_actions_create(const struct nlattr *,
size_t);
-struct dp_netdev_actions *dp_netdev_actions_ref(
- const struct dp_netdev_actions *);
-void dp_netdev_actions_unref(struct dp_netdev_actions *);
+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.
+ *
+ * Note, instance with cpu core id NON_PMD_CORE_ID will be reserved for
+ * I/O of all non-pmd threads. There will be no actual thread created
+ * for the instance.
+ **/
+struct dp_netdev_pmd_thread {
struct dp_netdev *dp;
+ struct cmap_node node; /* In 'dp->poll_threads'. */
+ /* Per thread exact-match cache. Note, the instance for cpu core
+ * NON_PMD_CORE_ID can be accessed by multiple threads, and thusly
+ * need to be protected (e.g. by 'dp_netdev_mutex'). All other
+ * instances will only be accessed by its own pmd thread. */
+ struct emc_cache flow_cache;
+ struct latch exit_latch; /* For terminating the pmd thread. */
+ atomic_uint change_seq; /* For reloading pmd ports. */
pthread_t thread;
- char *name;
- uint32_t min_hash, max_hash;
+ int index; /* Idx of this pmd thread among pmd*/
+ /* threads on same numa node. */
+ int core_id; /* CPU core id of this pmd thread. */
+ int numa_id; /* numa node id of this pmd thread. */
};
+#define PMD_INITIAL_SEQ 1
+
/* Interface to netdev-based datapath. */
struct dpif_netdev {
struct dpif dpif;
};
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 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, const struct flow *,
- const struct nlattr *userdata)
- OVS_EXCLUDED(dp->queue_mutex);
-static void dp_netdev_execute_actions(struct dp_netdev *dp,
- const struct flow *, struct ofpbuf *,
- struct pkt_metadata *,
+static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
+ struct dpif_packet **, int c,
+ bool may_steal,
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_input(struct dp_netdev_pmd_thread *,
+ struct dpif_packet **, int cnt);
+
+static void dp_netdev_disable_upcall(struct dp_netdev *);
+static void dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd,
+ struct dp_netdev *dp, int index,
+ int core_id, int numa_id);
+static void dp_netdev_set_nonpmd(struct dp_netdev *dp);
+static struct dp_netdev_pmd_thread *dp_netdev_get_nonpmd(struct dp_netdev *dp);
+static void dp_netdev_destroy_all_pmds(struct dp_netdev *dp);
+static void dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id);
+static void dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id);
+static void dp_netdev_reset_pmd_threads(struct dp_netdev *dp);
+
+static void emc_clear_entry(struct emc_entry *ce);
+
+static void
+emc_cache_init(struct emc_cache *flow_cache)
+{
+ int i;
+
+ BUILD_ASSERT(offsetof(struct miniflow, inline_values) == sizeof(uint64_t));
+
+ for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
+ flow_cache->entries[i].flow = NULL;
+ flow_cache->entries[i].key.hash = 0;
+ flow_cache->entries[i].key.len
+ = offsetof(struct miniflow, inline_values);
+ miniflow_initialize(&flow_cache->entries[i].key.mf,
+ flow_cache->entries[i].key.buf);
+ }
+}
+
+static void
+emc_cache_uninit(struct emc_cache *flow_cache)
+{
+ int i;
+
+ for (i = 0; i < ARRAY_SIZE(flow_cache->entries); i++) {
+ emc_clear_entry(&flow_cache->entries[i]);
+ }
+}
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;
{
struct dp_netdev *dp;
int error;
- int i;
dp = xzalloc(sizeof *dp);
shash_add(&dp_netdevs, name, dp);
atomic_flag_clear(&dp->destroyed);
ovs_mutex_init(&dp->flow_mutex);
- classifier_init(&dp->cls, NULL);
- hmap_init(&dp->flow_table);
+ dpcls_init(&dp->cls);
+ cmap_init(&dp->flow_table);
- ovs_mutex_init(&dp->queue_mutex);
- ovs_mutex_lock(&dp->queue_mutex);
- for (i = 0; i < N_QUEUES; i++) {
- dp->queues[i].head = dp->queues[i].tail = 0;
- }
- ovs_mutex_unlock(&dp->queue_mutex);
- dp->queue_seq = seq_create();
-
- dp->n_hit = ovsthread_counter_create();
- dp->n_missed = ovsthread_counter_create();
- dp->n_lost = ovsthread_counter_create();
+ 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);
+ fat_rwlock_init(&dp->upcall_rwlock);
+
+ /* Disable upcalls by default. */
+ dp_netdev_disable_upcall(dp);
+ dp->upcall_aux = NULL;
+ dp->upcall_cb = NULL;
+
+ cmap_init(&dp->poll_threads);
+ ovs_mutex_init_recursive(&dp->non_pmd_mutex);
+ ovsthread_key_create(&dp->per_pmd_key, NULL);
- ovs_rwlock_wrlock(&dp->port_rwlock);
+ /* Reserves the core NON_PMD_CORE_ID for all non-pmd threads. */
+ ovs_numa_try_pin_core_specific(NON_PMD_CORE_ID);
+ dp_netdev_set_nonpmd(dp);
+ dp->n_dpdk_rxqs = NR_QUEUE;
+
+ 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;
}
if (!error) {
*dpifp = create_dpif_netdev(dp);
+ dp->dpif = *dpifp;
}
ovs_mutex_unlock(&dp_netdev_mutex);
}
static void
-dp_netdev_purge_queues(struct dp_netdev *dp)
+dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
+ OVS_NO_THREAD_SAFETY_ANALYSIS
{
- int i;
-
- ovs_mutex_lock(&dp->queue_mutex);
- for (i = 0; i < N_QUEUES; i++) {
- struct dp_netdev_queue *q = &dp->queues[i];
+ /* Check that upcalls are disabled, i.e. that the rwlock is taken */
+ ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
- while (q->tail != q->head) {
- struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
- ofpbuf_uninit(&u->upcall.packet);
- ofpbuf_uninit(&u->buf);
- }
- }
- ovs_mutex_unlock(&dp->queue_mutex);
+ /* Before freeing a lock we should release it */
+ fat_rwlock_unlock(&dp->upcall_rwlock);
+ fat_rwlock_destroy(&dp->upcall_rwlock);
}
/* Requires dp_netdev_mutex so that we can't get a new reference to 'dp'
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_destroy_all_pmds(dp);
+ cmap_destroy(&dp->poll_threads);
+ ovs_mutex_destroy(&dp->non_pmd_mutex);
+ ovsthread_key_delete(dp->per_pmd_key);
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);
- ovsthread_counter_destroy(dp->n_hit);
- ovsthread_counter_destroy(dp->n_missed);
- ovsthread_counter_destroy(dp->n_lost);
+ ovs_mutex_unlock(&dp->port_mutex);
- dp_netdev_purge_queues(dp);
- seq_destroy(dp->queue_seq);
- ovs_mutex_destroy(&dp->queue_mutex);
+ OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
+ ovs_mutex_destroy(&bucket->mutex);
+ free_cacheline(bucket);
+ }
+ ovsthread_stats_destroy(&dp->stats);
- classifier_destroy(&dp->cls);
- hmap_destroy(&dp->flow_table);
+ dpcls_destroy(&dp->cls);
+ cmap_destroy(&dp->flow_table);
ovs_mutex_destroy(&dp->flow_mutex);
seq_destroy(dp->port_seq);
- hmap_destroy(&dp->ports);
- latch_destroy(&dp->exit_latch);
+ cmap_destroy(&dp->ports);
+
+ /* Upcalls must be disabled at this point */
+ dp_netdev_destroy_upcall_lock(dp);
+
+ free(dp->pmd_cmask);
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();
}
dpif_netdev_get_stats(const struct dpif *dpif, struct dpif_dp_stats *stats)
{
struct dp_netdev *dp = get_dp_netdev(dpif);
+ 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 = ovsthread_counter_read(dp->n_hit);
- stats->n_missed = ovsthread_counter_read(dp->n_missed);
- stats->n_lost = ovsthread_counter_read(dp->n_lost);
+ stats->n_hit = stats->n_missed = stats->n_lost = 0;
+ OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &dp->stats) {
+ ovs_mutex_lock(&bucket->mutex);
+ stats->n_hit += bucket->n[DP_STAT_HIT];
+ stats->n_missed += bucket->n[DP_STAT_MISS];
+ stats->n_lost += bucket->n[DP_STAT_LOST];
+ ovs_mutex_unlock(&bucket->mutex);
+ }
stats->n_masks = UINT32_MAX;
stats->n_mask_hit = UINT64_MAX;
return 0;
}
+static void
+dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
+{
+ int old_seq;
+
+ atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
+}
+
+/* Causes all pmd threads to reload its tx/rx devices.
+ * Must be called after adding/removing ports. */
+static void
+dp_netdev_reload_pmds(struct dp_netdev *dp)
+{
+ struct dp_netdev_pmd_thread *pmd;
+
+ CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
+ dp_netdev_reload_pmd__(pmd);
+ }
+}
+
+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;
+ if (netdev_is_pmd(netdev)) {
+ int n_cores = ovs_numa_get_n_cores();
+
+ if (n_cores == OVS_CORE_UNSPEC) {
+ VLOG_ERR("%s, cannot get cpu core info", devname);
+ return ENOENT;
+ }
+ /* There can only be ovs_numa_get_n_cores() pmd threads,
+ * so creates a txq for each. */
+ error = netdev_set_multiq(netdev, n_cores, dp->n_dpdk_rxqs);
+ if (error && (error != EOPNOTSUPP)) {
+ VLOG_ERR("%s, cannot set multiq", devname);
+ return errno;
+ }
+ }
+ 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);
+ free(port->type);
+ free(port->rxq);
+ free(port);
+ 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->type);
+ 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_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
+ dp_netdev_reload_pmds(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 bool
+port_try_ref(struct dp_netdev_port *port)
+{
+ if (port) {
+ return ovs_refcount_try_ref_rcu(&port->ref_cnt);
+ }
+
+ return false;
+}
+
+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;
}
static int
-do_del_port(struct dp_netdev *dp, odp_port_t port_no)
- OVS_REQ_WRLOCK(dp->port_rwlock)
+get_n_pmd_threads_on_numa(struct dp_netdev *dp, int numa_id)
+{
+ struct dp_netdev_pmd_thread *pmd;
+ int n_pmds = 0;
+
+ CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
+ if (pmd->numa_id == numa_id) {
+ n_pmds++;
+ }
+ }
+
+ return n_pmds;
+}
+
+/* Returns 'true' if there is a port with pmd netdev and the netdev
+ * is on numa node 'numa_id'. */
+static bool
+has_pmd_port_for_numa(struct dp_netdev *dp, int numa_id)
{
struct dp_netdev_port *port;
- int error;
- error = get_port_by_number(dp, port_no, &port);
- if (error) {
- return error;
+ CMAP_FOR_EACH (port, node, &dp->ports) {
+ if (netdev_is_pmd(port->netdev)
+ && netdev_get_numa_id(port->netdev) == numa_id) {
+ return true;
+ }
}
- hmap_remove(&dp->ports, &port->node);
+ return false;
+}
+
+
+static void
+do_del_port(struct dp_netdev *dp, struct dp_netdev_port *port)
+ OVS_REQUIRES(dp->port_mutex)
+{
+ cmap_remove(&dp->ports, &port->node, hash_odp_port(port->port_no));
seq_change(dp->port_seq);
+ if (netdev_is_pmd(port->netdev)) {
+ int numa_id = netdev_get_numa_id(port->netdev);
- netdev_close(port->netdev);
- netdev_restore_flags(port->sf);
- netdev_rx_close(port->rx);
- free(port->type);
- free(port);
+ /* If there is no netdev on the numa node, deletes the pmd threads
+ * for that numa. Else, just reloads the queues. */
+ if (!has_pmd_port_for_numa(dp, numa_id)) {
+ dp_netdev_del_pmds_on_numa(dp, numa_id);
+ }
+ dp_netdev_reload_pmds(dp);
+ }
- 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;
}
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)
+dp_netdev_flow_free(struct dp_netdev_flow *flow)
{
- struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
- struct hmap_node *node = CONST_CAST(struct hmap_node *, &flow->node);
+ struct dp_netdev_flow_stats *bucket;
+ size_t i;
- classifier_remove(&dp->cls, cr);
- hmap_remove(&dp->flow_table, node);
- dp_netdev_flow_unref(flow);
+ OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &flow->stats) {
+ ovs_mutex_destroy(&bucket->mutex);
+ free_cacheline(bucket);
+ }
+ ovsthread_stats_destroy(&flow->stats);
+
+ dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
+ free(flow);
}
-static struct dp_netdev_flow *
-dp_netdev_flow_ref(const struct dp_netdev_flow *flow_)
+static void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
{
- struct dp_netdev_flow *flow = CONST_CAST(struct dp_netdev_flow *, flow_);
- if (flow) {
- ovs_refcount_ref(&flow->ref_cnt);
+ if (ovs_refcount_unref_relaxed(&flow->ref_cnt) == 1) {
+ ovsrcu_postpone(dp_netdev_flow_free, flow);
}
- return flow;
}
static void
-dp_netdev_flow_unref(struct dp_netdev_flow *flow)
+dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
+ OVS_REQUIRES(dp->flow_mutex)
{
- if (flow && ovs_refcount_unref(&flow->ref_cnt) == 1) {
- cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
- ovs_mutex_lock(&flow->mutex);
- dp_netdev_actions_unref(flow->actions);
- ovs_mutex_unlock(&flow->mutex);
- ovs_mutex_destroy(&flow->mutex);
- free(flow);
- }
+ struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
+
+ dpcls_remove(&dp->cls, &flow->cr);
+ cmap_remove(&dp->flow_table, node, flow_hash(&flow->flow, 0));
+ flow->dead = true;
+
+ dp_netdev_flow_unref(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 (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_flow_cast(const struct cls_rule *cr)
+dp_netdev_flow_cast(const struct dpcls_rule *cr)
{
return cr ? CONTAINER_OF(cr, struct dp_netdev_flow, cr) : NULL;
}
-static struct dp_netdev_flow *
-dp_netdev_lookup_flow(const struct dp_netdev *dp, const struct flow *flow)
- OVS_EXCLUDED(dp->cls.rwlock)
+static bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
{
- struct dp_netdev_flow *netdev_flow;
+ return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
+}
- fat_rwlock_rdlock(&dp->cls.rwlock);
- netdev_flow = dp_netdev_flow_cast(classifier_lookup(&dp->cls, flow, NULL));
- dp_netdev_flow_ref(netdev_flow);
- fat_rwlock_unlock(&dp->cls.rwlock);
+/* netdev_flow_key utilities.
+ *
+ * netdev_flow_key is basically a miniflow. We use these functions
+ * (netdev_flow_key_clone, netdev_flow_key_equal, ...) instead of the miniflow
+ * functions (miniflow_clone_inline, miniflow_equal, ...), because:
+ *
+ * - Since we are dealing exclusively with miniflows created by
+ * miniflow_extract(), if the map is different the miniflow is different.
+ * Therefore we can be faster by comparing the map and the miniflow in a
+ * single memcmp().
+ * _ netdev_flow_key's miniflow has always inline values.
+ * - These functions can be inlined by the compiler.
+ *
+ * The following assertions make sure that what we're doing with miniflow is
+ * safe
+ */
+BUILD_ASSERT_DECL(offsetof(struct miniflow, inline_values)
+ == sizeof(uint64_t));
- return netdev_flow;
+/* Given the number of bits set in the miniflow map, returns the size of the
+ * 'netdev_flow_key.mf' */
+static inline uint32_t
+netdev_flow_key_size(uint32_t flow_u32s)
+{
+ return offsetof(struct miniflow, inline_values) +
+ MINIFLOW_VALUES_SIZE(flow_u32s);
}
-static struct dp_netdev_flow *
-dp_netdev_find_flow(const struct dp_netdev *dp, const struct flow *flow)
- OVS_REQ_RDLOCK(dp->cls.rwlock)
+static inline bool
+netdev_flow_key_equal(const struct netdev_flow_key *a,
+ const struct netdev_flow_key *b)
{
- struct dp_netdev_flow *netdev_flow;
+ /* 'b->len' may be not set yet. */
+ return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
+}
- HMAP_FOR_EACH_WITH_HASH (netdev_flow, node, flow_hash(flow, 0),
- &dp->flow_table) {
- if (flow_equal(&netdev_flow->flow, flow)) {
- return dp_netdev_flow_ref(netdev_flow);
- }
- }
+/* Used to compare 'netdev_flow_key' in the exact match cache to a miniflow.
+ * The maps are compared bitwise, so both 'key->mf' 'mf' must have been
+ * generated by miniflow_extract. */
+static inline bool
+netdev_flow_key_equal_mf(const struct netdev_flow_key *key,
+ const struct miniflow *mf)
+{
+ return !memcmp(&key->mf, mf, key->len);
+}
- return NULL;
+static inline void
+netdev_flow_key_clone(struct netdev_flow_key *dst,
+ const struct netdev_flow_key *src)
+{
+ memcpy(dst, src,
+ offsetof(struct netdev_flow_key, mf) + src->len);
}
+/* Slow. */
static void
-get_dpif_flow_stats(struct dp_netdev_flow *netdev_flow,
- struct dpif_flow_stats *stats)
- OVS_REQ_RDLOCK(netdev_flow->mutex)
+netdev_flow_key_from_flow(struct netdev_flow_key *dst,
+ const struct flow *src)
{
- stats->n_packets = netdev_flow->packet_count;
- stats->n_bytes = netdev_flow->byte_count;
- stats->used = netdev_flow->used;
- stats->tcp_flags = netdev_flow->tcp_flags;
+ struct ofpbuf packet;
+ uint64_t buf_stub[512 / 8];
+ struct pkt_metadata md = pkt_metadata_from_flow(src);
+
+ miniflow_initialize(&dst->mf, dst->buf);
+
+ ofpbuf_use_stub(&packet, buf_stub, sizeof buf_stub);
+ flow_compose(&packet, src);
+ miniflow_extract(&packet, &md, &dst->mf);
+ ofpbuf_uninit(&packet);
+
+ dst->len = netdev_flow_key_size(count_1bits(dst->mf.map));
+ dst->hash = 0; /* Not computed yet. */
}
-static int
-dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
- const struct nlattr *mask_key,
- uint32_t mask_key_len, const struct flow *flow,
- struct flow *mask)
+/* Initialize a netdev_flow_key 'mask' from 'match'. */
+static inline void
+netdev_flow_mask_init(struct netdev_flow_key *mask,
+ const struct match *match)
{
- if (mask_key_len) {
- enum odp_key_fitness fitness;
-
- fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
- if (fitness) {
- /* This should not happen: it indicates that
- * odp_flow_key_from_mask() and odp_flow_key_to_mask()
- * disagree on the acceptable form of a mask. Log the problem
- * as an error, with enough details to enable debugging. */
- static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
+ const uint32_t *mask_u32 = (const uint32_t *) &match->wc.masks;
+ uint32_t *dst = mask->mf.inline_values;
+ uint64_t map, mask_map = 0;
+ uint32_t hash = 0;
+ int n;
- if (!VLOG_DROP_ERR(&rl)) {
- struct ds s;
+ /* Only check masks that make sense for the flow. */
+ map = flow_wc_map(&match->flow);
- ds_init(&s);
- odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
- true);
- VLOG_ERR("internal error parsing flow mask %s (%s)",
- ds_cstr(&s), odp_key_fitness_to_string(fitness));
- ds_destroy(&s);
- }
+ while (map) {
+ uint64_t rm1bit = rightmost_1bit(map);
+ int i = raw_ctz(map);
- return EINVAL;
+ if (mask_u32[i]) {
+ mask_map |= rm1bit;
+ *dst++ = mask_u32[i];
+ hash = hash_add(hash, mask_u32[i]);
}
- /* 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
- * prerequisities are not met. */
- memset(mask, 0x0, sizeof *mask);
+ map -= rm1bit;
+ }
- for (id = 0; id < MFF_N_IDS; ++id) {
- /* Skip registers and metadata. */
+ mask->mf.values_inline = true;
+ mask->mf.map = mask_map;
+
+ hash = hash_add(hash, mask_map);
+ hash = hash_add(hash, mask_map >> 32);
+
+ n = dst - mask->mf.inline_values;
+
+ mask->hash = hash_finish(hash, n * 4);
+ mask->len = netdev_flow_key_size(n);
+}
+
+/* Initializes 'dst' as a copy of 'src' masked with 'mask'. */
+static inline void
+netdev_flow_key_init_masked(struct netdev_flow_key *dst,
+ const struct flow *flow,
+ const struct netdev_flow_key *mask)
+{
+ uint32_t *dst_u32 = dst->mf.inline_values;
+ const uint32_t *mask_u32 = mask->mf.inline_values;
+ uint32_t hash = 0;
+ uint32_t value;
+
+ dst->len = mask->len;
+ dst->mf.values_inline = true;
+ dst->mf.map = mask->mf.map;
+
+ FLOW_FOR_EACH_IN_MAP(value, flow, mask->mf.map) {
+ *dst_u32 = value & *mask_u32++;
+ hash = hash_add(hash, *dst_u32++);
+ }
+ dst->hash = hash_finish(hash, (dst_u32 - dst->mf.inline_values) * 4);
+}
+
+/* Iterate through all netdev_flow_key u32 values specified by 'MAP' */
+#define NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(VALUE, KEY, MAP) \
+ for (struct mf_for_each_in_map_aux aux__ \
+ = { (KEY)->mf.inline_values, (KEY)->mf.map, MAP }; \
+ mf_get_next_in_map(&aux__, &(VALUE)); \
+ )
+
+/* Returns a hash value for the bits of 'key' where there are 1-bits in
+ * 'mask'. */
+static inline uint32_t
+netdev_flow_key_hash_in_mask(const struct netdev_flow_key *key,
+ const struct netdev_flow_key *mask)
+{
+ const uint32_t *p = mask->mf.inline_values;
+ uint32_t hash = 0;
+ uint32_t key_u32;
+
+ NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(key_u32, key, mask->mf.map) {
+ hash = hash_add(hash, key_u32 & *p++);
+ }
+
+ return hash_finish(hash, (p - mask->mf.inline_values) * 4);
+}
+
+static inline bool
+emc_entry_alive(struct emc_entry *ce)
+{
+ return ce->flow && !ce->flow->dead;
+}
+
+static void
+emc_clear_entry(struct emc_entry *ce)
+{
+ if (ce->flow) {
+ dp_netdev_flow_unref(ce->flow);
+ ce->flow = NULL;
+ }
+}
+
+static inline void
+emc_change_entry(struct emc_entry *ce, struct dp_netdev_flow *flow,
+ const struct netdev_flow_key *key)
+{
+ if (ce->flow != flow) {
+ if (ce->flow) {
+ dp_netdev_flow_unref(ce->flow);
+ }
+
+ if (dp_netdev_flow_ref(flow)) {
+ ce->flow = flow;
+ } else {
+ ce->flow = NULL;
+ }
+ }
+ if (key) {
+ netdev_flow_key_clone(&ce->key, key);
+ }
+}
+
+static inline void
+emc_insert(struct emc_cache *cache, const struct netdev_flow_key *key,
+ struct dp_netdev_flow *flow)
+{
+ struct emc_entry *to_be_replaced = NULL;
+ struct emc_entry *current_entry;
+
+ EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
+ if (netdev_flow_key_equal(¤t_entry->key, key)) {
+ /* We found the entry with the 'mf' miniflow */
+ emc_change_entry(current_entry, flow, NULL);
+ return;
+ }
+
+ /* Replacement policy: put the flow in an empty (not alive) entry, or
+ * in the first entry where it can be */
+ if (!to_be_replaced
+ || (emc_entry_alive(to_be_replaced)
+ && !emc_entry_alive(current_entry))
+ || current_entry->key.hash < to_be_replaced->key.hash) {
+ to_be_replaced = current_entry;
+ }
+ }
+ /* We didn't find the miniflow in the cache.
+ * The 'to_be_replaced' entry is where the new flow will be stored */
+
+ emc_change_entry(to_be_replaced, flow, key);
+}
+
+static inline struct dp_netdev_flow *
+emc_lookup(struct emc_cache *cache, const struct netdev_flow_key *key)
+{
+ struct emc_entry *current_entry;
+
+ EMC_FOR_EACH_POS_WITH_HASH(cache, current_entry, key->hash) {
+ if (current_entry->key.hash == key->hash
+ && emc_entry_alive(current_entry)
+ && netdev_flow_key_equal_mf(¤t_entry->key, &key->mf)) {
+
+ /* We found the entry with the 'key->mf' miniflow */
+ return current_entry->flow;
+ }
+ }
+
+ return NULL;
+}
+
+static struct dp_netdev_flow *
+dp_netdev_lookup_flow(const struct dp_netdev *dp,
+ const struct netdev_flow_key *key)
+{
+ struct dp_netdev_flow *netdev_flow;
+ struct dpcls_rule *rule;
+
+ dpcls_lookup(&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)
+{
+ struct dp_netdev_flow *netdev_flow;
+
+ 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 NULL;
+}
+
+static void
+get_dpif_flow_stats(const struct dp_netdev_flow *netdev_flow,
+ struct dpif_flow_stats *stats)
+{
+ struct dp_netdev_flow_stats *bucket;
+ size_t i;
+
+ memset(stats, 0, sizeof *stats);
+ OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
+ ovs_mutex_lock(&bucket->mutex);
+ stats->n_packets += bucket->packet_count;
+ stats->n_bytes += bucket->byte_count;
+ stats->used = MAX(stats->used, bucket->used);
+ stats->tcp_flags |= bucket->tcp_flags;
+ ovs_mutex_unlock(&bucket->mutex);
+ }
+}
+
+static void
+dp_netdev_flow_to_dpif_flow(const struct dp_netdev_flow *netdev_flow,
+ struct ofpbuf *buffer, struct dpif_flow *flow)
+{
+ struct flow_wildcards wc;
+ struct dp_netdev_actions *actions;
+
+ miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
+ odp_flow_key_from_mask(buffer, &wc.masks, &netdev_flow->flow,
+ odp_to_u32(wc.masks.in_port.odp_port),
+ SIZE_MAX, true);
+ flow->mask = ofpbuf_data(buffer);
+ flow->mask_len = ofpbuf_size(buffer);
+
+ actions = dp_netdev_flow_get_actions(netdev_flow);
+ flow->actions = actions->actions;
+ flow->actions_len = actions->size;
+
+ get_dpif_flow_stats(netdev_flow, &flow->stats);
+}
+
+static int
+dpif_netdev_mask_from_nlattrs(const struct nlattr *key, uint32_t key_len,
+ const struct nlattr *mask_key,
+ uint32_t mask_key_len, const struct flow *flow,
+ struct flow *mask)
+{
+ if (mask_key_len) {
+ enum odp_key_fitness fitness;
+
+ fitness = odp_flow_key_to_mask(mask_key, mask_key_len, mask, flow);
+ if (fitness) {
+ /* This should not happen: it indicates that
+ * odp_flow_key_from_mask() and odp_flow_key_to_mask()
+ * disagree on the acceptable form of a mask. Log the problem
+ * as an error, with enough details to enable debugging. */
+ static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
+
+ if (!VLOG_DROP_ERR(&rl)) {
+ struct ds s;
+
+ ds_init(&s);
+ odp_flow_format(key, key_len, mask_key, mask_key_len, NULL, &s,
+ true);
+ VLOG_ERR("internal error parsing flow mask %s (%s)",
+ ds_cstr(&s), odp_key_fitness_to_string(fitness));
+ ds_destroy(&s);
+ }
+
+ return EINVAL;
+ }
+ } else {
+ enum mf_field_id id;
+ /* No mask key, unwildcard everything except fields whose
+ * prerequisities are not met. */
+ memset(mask, 0x0, sizeof *mask);
+
+ for (id = 0; id < MFF_N_IDS; ++id) {
+ /* Skip registers and metadata. */
if (!(id >= MFF_REG0 && id < MFF_REG0 + FLOW_N_REGS)
&& id != MFF_METADATA) {
const struct mf_field *mf = mf_from_id(id);
}
}
+ /* 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)
+dpif_netdev_flow_get(const struct dpif *dpif, const struct dpif_flow_get *get)
{
struct dp_netdev *dp = get_dp_netdev(dpif);
struct dp_netdev_flow *netdev_flow;
struct flow key;
int error;
- error = dpif_netdev_flow_from_nlattrs(nl_key, nl_key_len, &key);
+ error = dpif_netdev_flow_from_nlattrs(get->key, get->key_len, &key);
if (error) {
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) {
- struct dp_netdev_actions *actions = NULL;
-
- ovs_mutex_lock(&netdev_flow->mutex);
- if (stats) {
- get_dpif_flow_stats(netdev_flow, stats);
- }
- if (actionsp) {
- actions = dp_netdev_actions_ref(netdev_flow->actions);
- }
- ovs_mutex_unlock(&netdev_flow->mutex);
-
- dp_netdev_flow_unref(netdev_flow);
-
- if (actionsp) {
- *actionsp = ofpbuf_clone_data(actions->actions, actions->size);
- dp_netdev_actions_unref(actions);
- }
- } else {
+ dp_netdev_flow_to_dpif_flow(netdev_flow, get->buffer, get->flow);
+ } else {
error = ENOENT;
}
return error;
}
-static int
-dp_netdev_flow_add(struct dp_netdev *dp, const struct flow *flow,
- const struct flow_wildcards *wc,
- const struct nlattr *actions,
- size_t actions_len)
+static struct dp_netdev_flow *
+dp_netdev_flow_add(struct dp_netdev *dp, struct match *match,
+ const struct nlattr *actions, size_t actions_len)
OVS_REQUIRES(dp->flow_mutex)
{
- struct dp_netdev_flow *netdev_flow;
- struct match match;
+ struct dp_netdev_flow *flow;
+ struct netdev_flow_key mask;
- netdev_flow = xzalloc(sizeof *netdev_flow);
- *CONST_CAST(struct flow *, &netdev_flow->flow) = *flow;
- ovs_refcount_init(&netdev_flow->ref_cnt);
+ netdev_flow_mask_init(&mask, match);
+ /* Make sure wc does not have metadata. */
+ ovs_assert(!(mask.mf.map & (MINIFLOW_MAP(metadata) | MINIFLOW_MAP(regs))));
- ovs_mutex_init(&netdev_flow->mutex);
- ovs_mutex_lock(&netdev_flow->mutex);
+ /* Do not allocate extra space. */
+ flow = xmalloc(sizeof *flow - sizeof flow->cr.flow.mf + mask.len);
+ flow->dead = false;
+ *CONST_CAST(struct flow *, &flow->flow) = match->flow;
+ ovs_refcount_init(&flow->ref_cnt);
+ ovsthread_stats_init(&flow->stats);
+ ovsrcu_set(&flow->actions, dp_netdev_actions_create(actions, actions_len));
- netdev_flow->actions = dp_netdev_actions_create(actions, actions_len);
+ cmap_insert(&dp->flow_table,
+ CONST_CAST(struct cmap_node *, &flow->node),
+ flow_hash(&flow->flow, 0));
+ netdev_flow_key_init_masked(&flow->cr.flow, &match->flow, &mask);
+ dpcls_insert(&dp->cls, &flow->cr, &mask);
- 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);
- 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);
+ if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
+ struct match match;
+ struct ds ds = DS_EMPTY_INITIALIZER;
- ovs_mutex_unlock(&netdev_flow->mutex);
+ match.flow = flow->flow;
+ miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
- return 0;
+ ds_put_cstr(&ds, "flow_add: ");
+ match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
+ ds_put_cstr(&ds, ", actions:");
+ format_odp_actions(&ds, actions, actions_len);
+
+ VLOG_DBG_RL(&upcall_rl, "%s", ds_cstr(&ds));
+
+ ds_destroy(&ds);
+ }
+
+ return flow;
}
static void
clear_stats(struct dp_netdev_flow *netdev_flow)
- OVS_REQUIRES(netdev_flow->mutex)
{
- netdev_flow->used = 0;
- netdev_flow->packet_count = 0;
- netdev_flow->byte_count = 0;
- netdev_flow->tcp_flags = 0;
+ struct dp_netdev_flow_stats *bucket;
+ size_t i;
+
+ OVSTHREAD_STATS_FOR_EACH_BUCKET (bucket, i, &netdev_flow->stats) {
+ ovs_mutex_lock(&bucket->mutex);
+ bucket->used = 0;
+ bucket->packet_count = 0;
+ bucket->byte_count = 0;
+ bucket->tcp_flags = 0;
+ ovs_mutex_unlock(&bucket->mutex);
+ }
}
static int
{
struct dp_netdev *dp = get_dp_netdev(dpif);
struct dp_netdev_flow *netdev_flow;
- struct flow flow;
- struct flow_wildcards wc;
+ struct netdev_flow_key key;
+ struct match match;
int error;
- error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &flow);
+ error = dpif_netdev_flow_from_nlattrs(put->key, put->key_len, &match.flow);
if (error) {
return error;
}
error = dpif_netdev_mask_from_nlattrs(put->key, put->key_len,
put->mask, put->mask_len,
- &flow, &wc.masks);
+ &match.flow, &match.wc.masks);
if (error) {
return error;
}
+ /* Must produce a netdev_flow_key for lookup.
+ * This interface is no longer performance critical, since it is not used
+ * for upcall processing any more. */
+ netdev_flow_key_from_flow(&key, &match.flow);
+
ovs_mutex_lock(&dp->flow_mutex);
- netdev_flow = dp_netdev_lookup_flow(dp, &flow);
+ netdev_flow = dp_netdev_lookup_flow(dp, &key);
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);
}
- error = dp_netdev_flow_add(dp, &flow, &wc, put->actions,
- put->actions_len);
+ dp_netdev_flow_add(dp, &match, put->actions, put->actions_len);
+ error = 0;
} else {
error = EFBIG;
}
}
} else {
if (put->flags & DPIF_FP_MODIFY
- && flow_equal(&flow, &netdev_flow->flow)) {
+ && flow_equal(&match.flow, &netdev_flow->flow)) {
struct dp_netdev_actions *new_actions;
struct dp_netdev_actions *old_actions;
new_actions = dp_netdev_actions_create(put->actions,
put->actions_len);
- ovs_mutex_lock(&netdev_flow->mutex);
- old_actions = netdev_flow->actions;
- netdev_flow->actions = new_actions;
+ old_actions = dp_netdev_flow_get_actions(netdev_flow);
+ ovsrcu_set(&netdev_flow->actions, new_actions);
+
if (put->stats) {
get_dpif_flow_stats(netdev_flow, put->stats);
}
if (put->flags & DPIF_FP_ZERO_STATS) {
clear_stats(netdev_flow);
}
- ovs_mutex_unlock(&netdev_flow->mutex);
- dp_netdev_actions_unref(old_actions);
+ ovsrcu_postpone(dp_netdev_actions_free, old_actions);
} else if (put->flags & DPIF_FP_CREATE) {
error = EEXIST;
} else {
/* Overlapping flow. */
error = EINVAL;
}
- dp_netdev_flow_unref(netdev_flow);
}
ovs_mutex_unlock(&dp->flow_mutex);
}
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) {
- ovs_mutex_lock(&netdev_flow->mutex);
get_dpif_flow_stats(netdev_flow, del->stats);
- ovs_mutex_unlock(&netdev_flow->mutex);
}
dp_netdev_remove_flow(dp, netdev_flow);
- dp_netdev_flow_unref(netdev_flow);
} 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;
- dp_netdev_actions_unref(state->actions);
- 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;
}
+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];
+};
+
+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);
+}
+
+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;
+
+ 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_);
+
+ free(thread);
+}
+
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;
+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(&iter->mutex);
- error = iter->status;
- if (!error) {
- struct hmap_node *node;
+ 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;
- 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);
- dp_netdev_flow_ref(netdev_flow);
- }
- fat_rwlock_unlock(&dp->cls.rwlock);
- if (!node) {
- iter->status = error = EOF;
+ 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(&iter->mutex);
- if (error) {
- return error;
- }
+ ovs_mutex_unlock(&dump->mutex);
- if (key) {
+ 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->keybuf, sizeof state->keybuf);
- odp_flow_key_from_flow(&buf, &netdev_flow->flow,
- netdev_flow->flow.in_port.odp_port);
-
- *key = buf.data;
- *key_len = buf.size;
- }
+ miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
- if (key && mask) {
- struct ofpbuf buf;
- struct flow_wildcards wc;
+ /* 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);
- ofpbuf_use_stack(&buf, &state->maskbuf, sizeof state->maskbuf);
- minimask_expand(&netdev_flow->cr.match.mask, &wc);
+ /* 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);
-
- *mask = buf.data;
- *mask_len = buf.size;
- }
+ SIZE_MAX, true);
+ f->mask = ofpbuf_data(&buf);
+ f->mask_len = ofpbuf_size(&buf);
- if (actions || stats) {
- dp_netdev_actions_unref(state->actions);
- state->actions = NULL;
+ /* Actions. */
+ dp_actions = dp_netdev_flow_get_actions(netdev_flow);
+ f->actions = dp_actions->actions;
+ f->actions_len = dp_actions->size;
- ovs_mutex_lock(&netdev_flow->mutex);
- if (actions) {
- state->actions = dp_netdev_actions_ref(netdev_flow->actions);
- *actions = state->actions->actions;
- *actions_len = state->actions->size;
- }
- if (stats) {
- get_dpif_flow_stats(netdev_flow, &state->stats);
- *stats = &state->stats;
- }
- ovs_mutex_unlock(&netdev_flow->mutex);
+ /* Stats. */
+ get_dpif_flow_stats(netdev_flow, &f->stats);
}
- dp_netdev_flow_unref(netdev_flow);
-
- 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)
+ OVS_NO_THREAD_SAFETY_ANALYSIS
{
struct dp_netdev *dp = get_dp_netdev(dpif);
- struct pkt_metadata *md = &execute->md;
- struct flow key;
+ struct dp_netdev_pmd_thread *pmd;
+ struct dpif_packet packet, *pp;
- 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;
+ packet.md = execute->md;
+ pp = &packet;
- ovs_rwlock_rdlock(&dp->port_rwlock);
- dp_netdev_execute_actions(dp, &key, execute->packet, md, execute->actions,
+ /* Tries finding the 'pmd'. If NULL is returned, that means
+ * the current thread is a non-pmd thread and should use
+ * dp_netdev_get_nonpmd(). */
+ pmd = ovsthread_getspecific(dp->per_pmd_key);
+ if (!pmd) {
+ pmd = dp_netdev_get_nonpmd(dp);
+ }
+
+ /* If the current thread is non-pmd thread, acquires
+ * the 'non_pmd_mutex'. */
+ if (pmd->core_id == NON_PMD_CORE_ID) {
+ ovs_mutex_lock(&dp->non_pmd_mutex);
+ }
+ dp_netdev_execute_actions(pmd, &pp, 1, false, execute->actions,
execute->actions_len);
- ovs_rwlock_unlock(&dp->port_rwlock);
+ if (pmd->core_id == NON_PMD_CORE_ID) {
+ ovs_mutex_unlock(&dp->non_pmd_mutex);
+ }
+ /* 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;
+ execute->md = packet.md;
return 0;
}
-static int
-dpif_netdev_recv_set(struct dpif *dpif OVS_UNUSED, bool enable OVS_UNUSED)
+static void
+dpif_netdev_operate(struct dpif *dpif, struct dpif_op **ops, size_t n_ops)
{
- return 0;
-}
+ size_t i;
-static int
-dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
- uint32_t queue_id, uint32_t *priority)
-{
- *priority = queue_id;
- return 0;
+ for (i = 0; i < n_ops; i++) {
+ struct dpif_op *op = ops[i];
+
+ switch (op->type) {
+ case DPIF_OP_FLOW_PUT:
+ op->error = dpif_netdev_flow_put(dpif, &op->u.flow_put);
+ break;
+
+ case DPIF_OP_FLOW_DEL:
+ op->error = dpif_netdev_flow_del(dpif, &op->u.flow_del);
+ break;
+
+ case DPIF_OP_EXECUTE:
+ op->error = dpif_netdev_execute(dpif, &op->u.execute);
+ break;
+
+ case DPIF_OP_FLOW_GET:
+ op->error = dpif_netdev_flow_get(dpif, &op->u.flow_get);
+ break;
+ }
+ }
}
-static struct dp_netdev_queue *
-find_nonempty_queue(struct dp_netdev *dp)
- OVS_REQUIRES(dp->queue_mutex)
+/* Returns true if the configuration for rx queues or cpu mask
+ * is changed. */
+static bool
+pmd_config_changed(const struct dp_netdev *dp, size_t rxqs, const char *cmask)
{
- int i;
-
- for (i = 0; i < N_QUEUES; i++) {
- struct dp_netdev_queue *q = &dp->queues[i];
- if (q->head != q->tail) {
- return q;
+ if (dp->n_dpdk_rxqs != rxqs) {
+ return true;
+ } else {
+ if (dp->pmd_cmask != NULL && cmask != NULL) {
+ return strcmp(dp->pmd_cmask, cmask);
+ } else {
+ return (dp->pmd_cmask != NULL || cmask != NULL);
}
}
- return NULL;
}
+/* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
static int
-dpif_netdev_recv(struct dpif *dpif, struct dpif_upcall *upcall,
- struct ofpbuf *buf)
+dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
{
struct dp_netdev *dp = get_dp_netdev(dpif);
- struct dp_netdev_queue *q;
- int error;
- ovs_mutex_lock(&dp->queue_mutex);
- q = find_nonempty_queue(dp);
- if (q) {
- struct dp_netdev_upcall *u = &q->upcalls[q->tail++ & QUEUE_MASK];
+ if (pmd_config_changed(dp, n_rxqs, cmask)) {
+ struct dp_netdev_port *port;
- *upcall = u->upcall;
+ dp_netdev_destroy_all_pmds(dp);
- ofpbuf_uninit(buf);
- *buf = u->buf;
+ CMAP_FOR_EACH (port, node, &dp->ports) {
+ if (netdev_is_pmd(port->netdev)) {
+ int i, err;
- error = 0;
- } else {
- error = EAGAIN;
- }
- ovs_mutex_unlock(&dp->queue_mutex);
+ /* Closes the existing 'rxq's. */
+ for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
+ netdev_rxq_close(port->rxq[i]);
+ port->rxq[i] = NULL;
+ }
- return error;
-}
+ /* Sets the new rx queue config. */
+ err = netdev_set_multiq(port->netdev, ovs_numa_get_n_cores(),
+ n_rxqs);
+ if (err && (err != EOPNOTSUPP)) {
+ VLOG_ERR("Failed to set dpdk interface %s rx_queue to:"
+ " %u", netdev_get_name(port->netdev),
+ n_rxqs);
+ return err;
+ }
-static void
-dpif_netdev_recv_wait(struct dpif *dpif)
-{
- struct dp_netdev *dp = get_dp_netdev(dpif);
- uint64_t seq;
+ /* If the set_multiq() above succeeds, reopens the 'rxq's. */
+ port->rxq = xrealloc(port->rxq, sizeof *port->rxq
+ * netdev_n_rxq(port->netdev));
+ for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
+ netdev_rxq_open(port->netdev, &port->rxq[i], i);
+ }
+ }
+ }
+ dp->n_dpdk_rxqs = n_rxqs;
- ovs_mutex_lock(&dp->queue_mutex);
- seq = seq_read(dp->queue_seq);
- if (find_nonempty_queue(dp)) {
- poll_immediate_wake();
- } else {
- seq_wait(dp->queue_seq, seq);
- }
- ovs_mutex_unlock(&dp->queue_mutex);
-}
+ /* Reconfigures the cpu mask. */
+ ovs_numa_set_cpu_mask(cmask);
+ free(dp->pmd_cmask);
+ dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
-static void
-dpif_netdev_recv_purge(struct dpif *dpif)
-{
- struct dpif_netdev *dpif_netdev = dpif_netdev_cast(dpif);
+ /* Restores the non-pmd. */
+ dp_netdev_set_nonpmd(dp);
+ /* Restores all pmd threads. */
+ dp_netdev_reset_pmd_threads(dp);
+ }
- dp_netdev_purge_queues(dpif_netdev->dp);
+ return 0;
}
+
+static int
+dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
+ uint32_t queue_id, uint32_t *priority)
+{
+ *priority = queue_id;
+ return 0;
+}
+
\f
/* Creates and returns a new 'struct dp_netdev_actions', with a reference count
* of 1, whose actions are a copy of from the 'ofpacts_len' bytes of
struct dp_netdev_actions *netdev_actions;
netdev_actions = xmalloc(sizeof *netdev_actions);
- ovs_refcount_init(&netdev_actions->ref_cnt);
netdev_actions->actions = xmemdup(actions, size);
netdev_actions->size = size;
return netdev_actions;
}
-/* Increments 'actions''s refcount. */
struct dp_netdev_actions *
-dp_netdev_actions_ref(const struct dp_netdev_actions *actions_)
+dp_netdev_flow_get_actions(const struct dp_netdev_flow *flow)
{
- struct dp_netdev_actions *actions;
+ return ovsrcu_get(struct dp_netdev_actions *, &flow->actions);
+}
+
+static void
+dp_netdev_actions_free(struct dp_netdev_actions *actions)
+{
+ free(actions->actions);
+ free(actions);
+}
+\f
+
+static void
+dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
+ struct dp_netdev_port *port,
+ struct netdev_rxq *rxq)
+{
+ struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
+ int error, cnt;
+
+ error = netdev_rxq_recv(rxq, packets, &cnt);
+ if (!error) {
+ int i;
+
+ *recirc_depth_get() = 0;
+
+ /* XXX: initialize md in netdev implementation. */
+ for (i = 0; i < cnt; i++) {
+ packets[i]->md = PKT_METADATA_INITIALIZER(port->port_no);
+ }
+ dp_netdev_input(pmd, packets, cnt);
+ } else if (error != EAGAIN && error != EOPNOTSUPP) {
+ static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
- actions = CONST_CAST(struct dp_netdev_actions *, actions_);
- if (actions) {
- ovs_refcount_ref(&actions->ref_cnt);
+ VLOG_ERR_RL(&rl, "error receiving data from %s: %s",
+ netdev_get_name(port->netdev), ovs_strerror(error));
}
- return actions;
}
-/* Decrements 'actions''s refcount and frees 'actions' if the refcount reaches
- * 0. */
-void
-dp_netdev_actions_unref(struct dp_netdev_actions *actions)
+static void
+dpif_netdev_run(struct dpif *dpif)
{
- if (actions && ovs_refcount_unref(&actions->ref_cnt) == 1) {
- free(actions->actions);
- free(actions);
+ struct dp_netdev_port *port;
+ struct dp_netdev *dp = get_dp_netdev(dpif);
+ struct dp_netdev_pmd_thread *non_pmd = dp_netdev_get_nonpmd(dp);
+
+ ovs_mutex_lock(&dp->non_pmd_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++) {
+ dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
+ }
+ }
}
+ ovs_mutex_unlock(&dp->non_pmd_mutex);
}
-\f
-static void *
-dp_forwarder_main(void *f_)
+
+static void
+dpif_netdev_wait(struct dpif *dpif)
{
- struct dp_forwarder *f = f_;
- struct dp_netdev *dp = f->dp;
- struct ofpbuf packet;
+ struct dp_netdev_port *port;
+ struct dp_netdev *dp = get_dp_netdev(dpif);
- f->name = xasprintf("forwarder_%u", ovsthread_id_self());
- set_subprogram_name("%s", f->name);
+ ovs_mutex_lock(&dp_netdev_mutex);
+ CMAP_FOR_EACH (port, node, &dp->ports) {
+ if (!netdev_is_pmd(port->netdev)) {
+ int i;
- ofpbuf_init(&packet, 0);
- while (!latch_is_set(&dp->exit_latch)) {
- bool received_anything;
- int i;
+ for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
+ netdev_rxq_wait(port->rxq[i]);
+ }
+ }
+ }
+ ovs_mutex_unlock(&dp_netdev_mutex);
+}
- 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) {
- int buf_size;
- int error;
- int mtu;
-
- if (netdev_get_mtu(port->netdev, &mtu)) {
- mtu = ETH_PAYLOAD_MAX;
- }
- buf_size = DP_NETDEV_HEADROOM + VLAN_ETH_HEADER_LEN + mtu;
-
- ofpbuf_clear(&packet);
- ofpbuf_reserve_with_tailroom(&packet, DP_NETDEV_HEADROOM,
- buf_size);
-
- error = netdev_rx_recv(port->rx, &packet);
- if (!error) {
- struct pkt_metadata md
- = PKT_METADATA_INITIALIZER(port->port_no);
-
- dp_netdev_port_input(dp, &packet, &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));
+struct rxq_poll {
+ struct dp_netdev_port *port;
+ struct netdev_rxq *rx;
+};
+
+static int
+pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
+ struct rxq_poll **ppoll_list, int poll_cnt)
+{
+ struct rxq_poll *poll_list = *ppoll_list;
+ struct dp_netdev_port *port;
+ int n_pmds_on_numa, index, i;
+
+ /* Simple scheduler for netdev rx polling. */
+ for (i = 0; i < poll_cnt; i++) {
+ port_unref(poll_list[i].port);
+ }
+
+ poll_cnt = 0;
+ n_pmds_on_numa = get_n_pmd_threads_on_numa(pmd->dp, pmd->numa_id);
+ index = 0;
+
+ CMAP_FOR_EACH (port, node, &pmd->dp->ports) {
+ /* Calls port_try_ref() to prevent the main thread
+ * from deleting the port. */
+ if (port_try_ref(port)) {
+ if (netdev_is_pmd(port->netdev)
+ && netdev_get_numa_id(port->netdev) == pmd->numa_id) {
+ int i;
+
+ for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
+ if ((index % n_pmds_on_numa) == pmd->index) {
+ 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++;
}
+ index++;
}
}
+ /* Unrefs the port_try_ref(). */
+ port_unref(port);
+ }
+ }
- if (!received_anything) {
- break;
- }
+ *ppoll_list = poll_list;
+ return poll_cnt;
+}
+
+static void *
+pmd_thread_main(void *f_)
+{
+ struct dp_netdev_pmd_thread *pmd = f_;
+ unsigned int lc = 0;
+ struct rxq_poll *poll_list;
+ unsigned int port_seq = PMD_INITIAL_SEQ;
+ int poll_cnt;
+ int i;
+
+ poll_cnt = 0;
+ poll_list = NULL;
+
+ /* Stores the pmd thread's 'pmd' to 'per_pmd_key'. */
+ ovsthread_setspecific(pmd->dp->per_pmd_key, pmd);
+ pmd_thread_setaffinity_cpu(pmd->core_id);
+reload:
+ emc_cache_init(&pmd->flow_cache);
+ poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
+
+ for (;;) {
+ int i;
+
+ for (i = 0; i < poll_cnt; i++) {
+ dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
}
- if (received_anything) {
- poll_immediate_wake();
- } else {
- struct dp_netdev_port *port;
+ if (lc++ > 1024) {
+ unsigned int seq;
- 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);
+ lc = 0;
+
+ ovsrcu_quiesce();
+
+ atomic_read_relaxed(&pmd->change_seq, &seq);
+ if (seq != port_seq) {
+ port_seq = seq;
+ break;
+ }
}
- ovs_rwlock_unlock(&dp->port_rwlock);
+ }
+
+ emc_cache_uninit(&pmd->flow_cache);
- poll_block();
+ if (!latch_is_set(&pmd->exit_latch)){
+ goto reload;
}
- ofpbuf_uninit(&packet);
- 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_disable_upcall(struct dp_netdev *dp)
+ OVS_ACQUIRES(dp->upcall_rwlock)
{
- int i;
+ fat_rwlock_wrlock(&dp->upcall_rwlock);
+}
- if (n == dp->n_forwarders) {
- return;
+static void
+dpif_netdev_disable_upcall(struct dpif *dpif)
+ OVS_NO_THREAD_SAFETY_ANALYSIS
+{
+ struct dp_netdev *dp = get_dp_netdev(dpif);
+ dp_netdev_disable_upcall(dp);
+}
+
+static void
+dp_netdev_enable_upcall(struct dp_netdev *dp)
+ OVS_RELEASES(dp->upcall_rwlock)
+{
+ fat_rwlock_unlock(&dp->upcall_rwlock);
+}
+
+static void
+dpif_netdev_enable_upcall(struct dpif *dpif)
+ OVS_NO_THREAD_SAFETY_ANALYSIS
+{
+ struct dp_netdev *dp = get_dp_netdev(dpif);
+ dp_netdev_enable_upcall(dp);
+}
+
+/* Returns the pointer to the dp_netdev_pmd_thread for non-pmd threads. */
+static struct dp_netdev_pmd_thread *
+dp_netdev_get_nonpmd(struct dp_netdev *dp)
+{
+ struct dp_netdev_pmd_thread *pmd;
+ const struct cmap_node *pnode;
+
+ pnode = cmap_find(&dp->poll_threads, hash_int(NON_PMD_CORE_ID, 0));
+ ovs_assert(pnode);
+ pmd = CONTAINER_OF(pnode, struct dp_netdev_pmd_thread, node);
+
+ return pmd;
+}
+
+/* Sets the 'struct dp_netdev_pmd_thread' for non-pmd threads. */
+static void
+dp_netdev_set_nonpmd(struct dp_netdev *dp)
+{
+ struct dp_netdev_pmd_thread *non_pmd;
+
+ non_pmd = xzalloc(sizeof *non_pmd);
+ dp_netdev_configure_pmd(non_pmd, dp, 0, NON_PMD_CORE_ID,
+ OVS_NUMA_UNSPEC);
+}
+
+/* Configures the 'pmd' based on the input argument. */
+static void
+dp_netdev_configure_pmd(struct dp_netdev_pmd_thread *pmd, struct dp_netdev *dp,
+ int index, int core_id, int numa_id)
+{
+ pmd->dp = dp;
+ pmd->index = index;
+ pmd->core_id = core_id;
+ pmd->numa_id = numa_id;
+ latch_init(&pmd->exit_latch);
+ atomic_init(&pmd->change_seq, PMD_INITIAL_SEQ);
+ /* init the 'flow_cache' since there is no
+ * actual thread created for NON_PMD_CORE_ID. */
+ if (core_id == NON_PMD_CORE_ID) {
+ emc_cache_init(&pmd->flow_cache);
+ }
+ cmap_insert(&dp->poll_threads, CONST_CAST(struct cmap_node *, &pmd->node),
+ hash_int(core_id, 0));
+}
+
+/* Stops the pmd thread, removes it from the 'dp->poll_threads'
+ * and destroys the struct. */
+static void
+dp_netdev_del_pmd(struct dp_netdev_pmd_thread *pmd)
+{
+ /* Uninit the 'flow_cache' since there is
+ * no actual thread uninit it. */
+ if (pmd->core_id == NON_PMD_CORE_ID) {
+ emc_cache_uninit(&pmd->flow_cache);
+ } else {
+ latch_set(&pmd->exit_latch);
+ dp_netdev_reload_pmd__(pmd);
+ ovs_numa_unpin_core(pmd->core_id);
+ xpthread_join(pmd->thread, NULL);
+ }
+ cmap_remove(&pmd->dp->poll_threads, &pmd->node, hash_int(pmd->core_id, 0));
+ latch_destroy(&pmd->exit_latch);
+ free(pmd);
+}
+
+/* Destroys all pmd threads. */
+static void
+dp_netdev_destroy_all_pmds(struct dp_netdev *dp)
+{
+ struct dp_netdev_pmd_thread *pmd;
+
+ CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
+ dp_netdev_del_pmd(pmd);
+ }
+}
+
+/* Deletes all pmd threads on numa node 'numa_id'. */
+static void
+dp_netdev_del_pmds_on_numa(struct dp_netdev *dp, int numa_id)
+{
+ struct dp_netdev_pmd_thread *pmd;
+
+ CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
+ if (pmd->numa_id == numa_id) {
+ dp_netdev_del_pmd(pmd);
+ }
}
+}
- /* Stop existing threads. */
- latch_set(&dp->exit_latch);
- for (i = 0; i < dp->n_forwarders; i++) {
- struct dp_forwarder *f = &dp->forwarders[i];
+/* Checks the numa node id of 'netdev' and starts pmd threads for
+ * the numa node. */
+static void
+dp_netdev_set_pmds_on_numa(struct dp_netdev *dp, int numa_id)
+{
+ int n_pmds;
- xpthread_join(f->thread, NULL);
+ if (!ovs_numa_numa_id_is_valid(numa_id)) {
+ VLOG_ERR("Cannot create pmd threads due to numa id (%d)"
+ "invalid", numa_id);
+ return ;
}
- latch_poll(&dp->exit_latch);
- free(dp->forwarders);
- /* Start new threads. */
- dp->forwarders = xmalloc(n * sizeof *dp->forwarders);
- dp->n_forwarders = n;
- for (i = 0; i < n; i++) {
- struct dp_forwarder *f = &dp->forwarders[i];
+ n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
+
+ /* If there are already pmd threads created for the numa node
+ * in which 'netdev' is on, do nothing. Else, creates the
+ * pmd threads for the numa node. */
+ if (!n_pmds) {
+ int can_have, n_unpinned, 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;
+ n_unpinned = ovs_numa_get_n_unpinned_cores_on_numa(numa_id);
+ if (!n_unpinned) {
+ VLOG_ERR("Cannot create pmd threads due to out of unpinned "
+ "cores on numa node");
+ return;
}
- xpthread_create(&f->thread, NULL, dp_forwarder_main, f);
+
+ /* If cpu mask is specified, uses all unpinned cores, otherwise
+ * tries creating NR_PMD_THREADS pmd threads. */
+ can_have = dp->pmd_cmask ? n_unpinned : MIN(n_unpinned, NR_PMD_THREADS);
+ for (i = 0; i < can_have; i++) {
+ struct dp_netdev_pmd_thread *pmd = xzalloc(sizeof *pmd);
+ int core_id = ovs_numa_get_unpinned_core_on_numa(numa_id);
+
+ dp_netdev_configure_pmd(pmd, dp, i, core_id, numa_id);
+ /* Each thread will distribute all devices rx-queues among
+ * themselves. */
+ pmd->thread = ovs_thread_create("pmd", pmd_thread_main, pmd);
+ }
+ VLOG_INFO("Created %d pmd threads on numa node %d", can_have, numa_id);
}
}
+
\f
+static void *
+dp_netdev_flow_stats_new_cb(void)
+{
+ struct dp_netdev_flow_stats *bucket = xzalloc_cacheline(sizeof *bucket);
+ ovs_mutex_init(&bucket->mutex);
+ return bucket;
+}
+
+/* Called after pmd threads config change. Restarts pmd threads with
+ * new configuration. */
+static void
+dp_netdev_reset_pmd_threads(struct dp_netdev *dp)
+{
+ struct dp_netdev_port *port;
+
+ CMAP_FOR_EACH (port, node, &dp->ports) {
+ if (netdev_is_pmd(port->netdev)) {
+ int numa_id = netdev_get_numa_id(port->netdev);
+
+ dp_netdev_set_pmds_on_numa(dp, numa_id);
+ }
+ }
+}
+
static void
dp_netdev_flow_used(struct dp_netdev_flow *netdev_flow,
- const struct ofpbuf *packet)
- OVS_REQUIRES(netdev_flow->mutex)
+ int cnt, int size,
+ uint16_t tcp_flags)
+{
+ long long int now = time_msec();
+ struct dp_netdev_flow_stats *bucket;
+
+ bucket = ovsthread_stats_bucket_get(&netdev_flow->stats,
+ dp_netdev_flow_stats_new_cb);
+
+ ovs_mutex_lock(&bucket->mutex);
+ bucket->used = MAX(now, bucket->used);
+ bucket->packet_count += cnt;
+ bucket->byte_count += size;
+ bucket->tcp_flags |= tcp_flags;
+ ovs_mutex_unlock(&bucket->mutex);
+}
+
+static void *
+dp_netdev_stats_new_cb(void)
{
- netdev_flow->used = time_msec();
- netdev_flow->packet_count++;
- netdev_flow->byte_count += packet->size;
- netdev_flow->tcp_flags |= packet_get_tcp_flags(packet, &netdev_flow->flow);
+ struct dp_netdev_stats *bucket = xzalloc_cacheline(sizeof *bucket);
+ ovs_mutex_init(&bucket->mutex);
+ return bucket;
}
static void
-dp_netdev_port_input(struct dp_netdev *dp, struct ofpbuf *packet,
- struct pkt_metadata *md)
- OVS_REQ_RDLOCK(dp->port_rwlock)
+dp_netdev_count_packet(struct dp_netdev *dp, enum dp_stat_type type, int cnt)
{
- struct dp_netdev_flow *netdev_flow;
- struct flow key;
+ struct dp_netdev_stats *bucket;
- if (packet->size < ETH_HEADER_LEN) {
- return;
+ bucket = ovsthread_stats_bucket_get(&dp->stats, dp_netdev_stats_new_cb);
+ ovs_mutex_lock(&bucket->mutex);
+ bucket->n[type] += cnt;
+ ovs_mutex_unlock(&bucket->mutex);
+}
+
+static int
+dp_netdev_upcall(struct dp_netdev *dp, struct dpif_packet *packet_,
+ struct flow *flow, struct flow_wildcards *wc,
+ enum dpif_upcall_type type, const struct nlattr *userdata,
+ struct ofpbuf *actions, struct ofpbuf *put_actions)
+{
+ struct ofpbuf *packet = &packet_->ofpbuf;
+
+ if (type == DPIF_UC_MISS) {
+ dp_netdev_count_packet(dp, DP_STAT_MISS, 1);
}
- flow_extract(packet, md, &key);
- netdev_flow = dp_netdev_lookup_flow(dp, &key);
- if (netdev_flow) {
- struct dp_netdev_actions *actions;
-
- ovs_mutex_lock(&netdev_flow->mutex);
- dp_netdev_flow_used(netdev_flow, packet);
- actions = dp_netdev_actions_ref(netdev_flow->actions);
- ovs_mutex_unlock(&netdev_flow->mutex);
-
- dp_netdev_execute_actions(dp, &key, packet, md,
- actions->actions, actions->size);
- dp_netdev_actions_unref(actions);
- dp_netdev_flow_unref(netdev_flow);
- ovsthread_counter_inc(dp->n_hit, 1);
- } else {
- ovsthread_counter_inc(dp->n_missed, 1);
- dp_netdev_output_userspace(dp, packet, DPIF_UC_MISS, &key, NULL);
+
+ if (OVS_UNLIKELY(!dp->upcall_cb)) {
+ return ENODEV;
+ }
+
+ if (OVS_UNLIKELY(!VLOG_DROP_DBG(&upcall_rl))) {
+ struct ds ds = DS_EMPTY_INITIALIZER;
+ struct ofpbuf key;
+ char *packet_str;
+
+ ofpbuf_init(&key, 0);
+ odp_flow_key_from_flow(&key, flow, &wc->masks, flow->in_port.odp_port,
+ true);
+
+ packet_str = ofp_packet_to_string(ofpbuf_data(packet),
+ ofpbuf_size(packet));
+
+ odp_flow_key_format(ofpbuf_data(&key), ofpbuf_size(&key), &ds);
+
+ VLOG_DBG("%s: %s upcall:\n%s\n%s", dp->name,
+ dpif_upcall_type_to_string(type), ds_cstr(&ds), packet_str);
+
+ ofpbuf_uninit(&key);
+ free(packet_str);
+ ds_destroy(&ds);
}
+
+ return dp->upcall_cb(packet, flow, type, userdata, actions, wc,
+ put_actions, dp->upcall_aux);
}
-static int
-dp_netdev_output_userspace(struct dp_netdev *dp, struct ofpbuf *packet,
- int queue_no, const struct flow *flow,
- const struct nlattr *userdata)
- OVS_EXCLUDED(dp->queue_mutex)
+static inline uint32_t
+dpif_netdev_packet_get_dp_hash(struct dpif_packet *packet,
+ const struct miniflow *mf)
{
- struct dp_netdev_queue *q = &dp->queues[queue_no];
- int error;
+ uint32_t hash;
+
+ hash = dpif_packet_get_dp_hash(packet);
+ if (OVS_UNLIKELY(!hash)) {
+ hash = miniflow_hash_5tuple(mf, 0);
+ dpif_packet_set_dp_hash(packet, hash);
+ }
+ return hash;
+}
+
+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];
+};
- ovs_mutex_lock(&dp->queue_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;
+static inline void
+packet_batch_update(struct packet_batch *batch, struct dpif_packet *packet,
+ const struct miniflow *mf)
+{
+ batch->tcp_flags |= miniflow_get_tcp_flags(mf);
+ batch->packets[batch->packet_count++] = packet;
+ batch->byte_count += ofpbuf_size(&packet->ofpbuf);
+}
+
+static inline void
+packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
+{
+ batch->flow = flow;
+
+ batch->packet_count = 0;
+ batch->byte_count = 0;
+ batch->tcp_flags = 0;
+}
- upcall->type = queue_no;
+static inline void
+packet_batch_execute(struct packet_batch *batch,
+ struct dp_netdev_pmd_thread *pmd)
+{
+ 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);
- /* Allocate buffer big enough for everything. */
- buf_size = ODPUTIL_FLOW_KEY_BYTES;
- if (userdata) {
- buf_size += NLA_ALIGN(userdata->nla_len);
+ actions = dp_netdev_flow_get_actions(flow);
+
+ dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
+ actions->actions, actions->size);
+
+ dp_netdev_count_packet(pmd->dp, DP_STAT_HIT, batch->packet_count);
+}
+
+static inline bool
+dp_netdev_queue_batches(struct dpif_packet *pkt,
+ struct dp_netdev_flow *flow, const struct miniflow *mf,
+ struct packet_batch *batches, size_t *n_batches,
+ size_t max_batches)
+{
+ struct packet_batch *batch = NULL;
+ int j;
+
+ if (OVS_UNLIKELY(!flow)) {
+ return false;
+ }
+ /* 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. */
+ for (j = *n_batches - 1; j >= 0; j--) {
+ if (batches[j].flow == flow) {
+ batch = &batches[j];
+ packet_batch_update(batch, pkt, mf);
+ return true;
}
- ofpbuf_init(buf, buf_size);
+ }
+ if (OVS_UNLIKELY(*n_batches >= max_batches)) {
+ return false;
+ }
- /* Put ODP flow. */
- odp_flow_key_from_flow(buf, flow, flow->in_port.odp_port);
- upcall->key = buf->data;
- upcall->key_len = buf->size;
+ batch = &batches[(*n_batches)++];
+ packet_batch_init(batch, flow);
+ packet_batch_update(batch, pkt, mf);
+ return true;
+}
+
+static inline void
+dpif_packet_swap(struct dpif_packet **a, struct dpif_packet **b)
+{
+ struct dpif_packet *tmp = *a;
+ *a = *b;
+ *b = tmp;
+}
- /* Put userdata. */
- if (userdata) {
- upcall->userdata = ofpbuf_put(buf, userdata,
- NLA_ALIGN(userdata->nla_len));
+/* Try to process all ('cnt') the 'packets' using only the exact match cache
+ * 'flow_cache'. If a flow is not found for a packet 'packets[i]', or if there
+ * is no matching batch for a packet's flow, the miniflow is copied into 'keys'
+ * and the packet pointer is moved at the beginning of the 'packets' array.
+ *
+ * The function returns the number of packets that needs to be processed in the
+ * 'packets' array (they have been moved to the beginning of the vector).
+ */
+static inline size_t
+emc_processing(struct dp_netdev_pmd_thread *pmd, struct dpif_packet **packets,
+ size_t cnt, struct netdev_flow_key *keys)
+{
+ struct netdev_flow_key key;
+ struct packet_batch batches[4];
+ struct emc_cache *flow_cache = &pmd->flow_cache;
+ size_t n_batches, i;
+ size_t notfound_cnt = 0;
+
+ n_batches = 0;
+ miniflow_initialize(&key.mf, key.buf);
+ for (i = 0; i < cnt; i++) {
+ struct dp_netdev_flow *flow;
+
+ if (OVS_UNLIKELY(ofpbuf_size(&packets[i]->ofpbuf) < ETH_HEADER_LEN)) {
+ dpif_packet_delete(packets[i]);
+ continue;
}
- /* Steal packet data. */
- ovs_assert(packet->source == OFPBUF_MALLOC);
- upcall->packet = *packet;
- ofpbuf_use(packet, NULL, 0);
+ miniflow_extract(&packets[i]->ofpbuf, &packets[i]->md, &key.mf);
+ key.len = 0; /* Not computed yet. */
+ key.hash = dpif_netdev_packet_get_dp_hash(packets[i], &key.mf);
- seq_change(dp->queue_seq);
+ flow = emc_lookup(flow_cache, &key);
+ if (OVS_UNLIKELY(!dp_netdev_queue_batches(packets[i], flow, &key.mf,
+ batches, &n_batches,
+ ARRAY_SIZE(batches)))) {
+ if (i != notfound_cnt) {
+ dpif_packet_swap(&packets[i], &packets[notfound_cnt]);
+ }
- error = 0;
- } else {
- ovsthread_counter_inc(dp->n_lost, 1);
- error = ENOBUFS;
+ keys[notfound_cnt++] = key;
+ }
}
- ovs_mutex_unlock(&dp->queue_mutex);
- return error;
+ for (i = 0; i < n_batches; i++) {
+ packet_batch_execute(&batches[i], pmd);
+ }
+
+ return notfound_cnt;
+}
+
+static inline void
+fast_path_processing(struct dp_netdev_pmd_thread *pmd,
+ struct dpif_packet **packets, size_t cnt,
+ struct netdev_flow_key *keys)
+{
+#if !defined(__CHECKER__) && !defined(_WIN32)
+ const size_t PKT_ARRAY_SIZE = cnt;
+#else
+ /* Sparse or MSVC doesn't like variable length array. */
+ enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
+#endif
+ struct packet_batch batches[PKT_ARRAY_SIZE];
+ struct dpcls_rule *rules[PKT_ARRAY_SIZE];
+ struct dp_netdev *dp = pmd->dp;
+ struct emc_cache *flow_cache = &pmd->flow_cache;
+ size_t n_batches, i;
+ bool any_miss;
+
+ for (i = 0; i < cnt; i++) {
+ /* Key length is needed in all the cases, hash computed on demand. */
+ keys[i].len = netdev_flow_key_size(count_1bits(keys[i].mf.map));
+ }
+ any_miss = !dpcls_lookup(&dp->cls, keys, rules, cnt);
+ if (OVS_UNLIKELY(any_miss) && !fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
+ uint64_t actions_stub[512 / 8], slow_stub[512 / 8];
+ struct ofpbuf actions, put_actions;
+
+ ofpbuf_use_stub(&actions, actions_stub, sizeof actions_stub);
+ ofpbuf_use_stub(&put_actions, slow_stub, sizeof slow_stub);
+
+ for (i = 0; i < cnt; i++) {
+ struct dp_netdev_flow *netdev_flow;
+ struct ofpbuf *add_actions;
+ struct match match;
+ int error;
+
+ if (OVS_LIKELY(rules[i])) {
+ continue;
+ }
+
+ /* It's possible that an earlier slow path execution installed
+ * a rule covering this flow. In this case, it's a lot cheaper
+ * to catch it here than execute a miss. */
+ netdev_flow = dp_netdev_lookup_flow(dp, &keys[i]);
+ if (netdev_flow) {
+ rules[i] = &netdev_flow->cr;
+ continue;
+ }
+
+ miniflow_expand(&keys[i].mf, &match.flow);
+
+ ofpbuf_clear(&actions);
+ ofpbuf_clear(&put_actions);
+
+ error = dp_netdev_upcall(dp, packets[i], &match.flow, &match.wc,
+ DPIF_UC_MISS, NULL, &actions,
+ &put_actions);
+ if (OVS_UNLIKELY(error && error != ENOSPC)) {
+ continue;
+ }
+
+ /* We can't allow the packet batching in the next loop to execute
+ * the actions. Otherwise, if there are any slow path actions,
+ * we'll send the packet up twice. */
+ dp_netdev_execute_actions(pmd, &packets[i], 1, true,
+ ofpbuf_data(&actions),
+ ofpbuf_size(&actions));
+
+ add_actions = ofpbuf_size(&put_actions)
+ ? &put_actions
+ : &actions;
+
+ if (OVS_LIKELY(error != ENOSPC)) {
+ /* XXX: There's a race window where a flow covering this packet
+ * could have already been installed since we last did the flow
+ * lookup before upcall. This could be solved by moving the
+ * mutex lock outside the loop, but that's an awful long time
+ * to be locking everyone out of making flow installs. If we
+ * move to a per-core classifier, it would be reasonable. */
+ ovs_mutex_lock(&dp->flow_mutex);
+ netdev_flow = dp_netdev_lookup_flow(dp, &keys[i]);
+ if (OVS_LIKELY(!netdev_flow)) {
+ netdev_flow = dp_netdev_flow_add(dp, &match,
+ ofpbuf_data(add_actions),
+ ofpbuf_size(add_actions));
+ }
+ ovs_mutex_unlock(&dp->flow_mutex);
+
+ emc_insert(flow_cache, &keys[i], netdev_flow);
+ }
+ }
+
+ ofpbuf_uninit(&actions);
+ ofpbuf_uninit(&put_actions);
+ fat_rwlock_unlock(&dp->upcall_rwlock);
+ } else if (OVS_UNLIKELY(any_miss)) {
+ int dropped_cnt = 0;
+
+ for (i = 0; i < cnt; i++) {
+ if (OVS_UNLIKELY(!rules[i])) {
+ dpif_packet_delete(packets[i]);
+ dropped_cnt++;
+ }
+ }
+
+ dp_netdev_count_packet(dp, DP_STAT_LOST, dropped_cnt);
+ }
+
+ n_batches = 0;
+ for (i = 0; i < cnt; i++) {
+ struct dpif_packet *packet = packets[i];
+ struct dp_netdev_flow *flow;
+
+ if (OVS_UNLIKELY(!rules[i])) {
+ continue;
+ }
+
+ flow = dp_netdev_flow_cast(rules[i]);
+
+ emc_insert(flow_cache, &keys[i], flow);
+ dp_netdev_queue_batches(packet, flow, &keys[i].mf, batches,
+ &n_batches, ARRAY_SIZE(batches));
+ }
+
+ for (i = 0; i < n_batches; i++) {
+ packet_batch_execute(&batches[i], pmd);
+ }
+}
+
+static void
+dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
+ struct dpif_packet **packets, int cnt)
+{
+#if !defined(__CHECKER__) && !defined(_WIN32)
+ const size_t PKT_ARRAY_SIZE = cnt;
+#else
+ /* Sparse or MSVC doesn't like variable length array. */
+ enum { PKT_ARRAY_SIZE = NETDEV_MAX_RX_BATCH };
+#endif
+ struct netdev_flow_key keys[PKT_ARRAY_SIZE];
+ size_t newcnt;
+
+ newcnt = emc_processing(pmd, packets, cnt, keys);
+ if (OVS_UNLIKELY(newcnt)) {
+ fast_path_processing(pmd, packets, newcnt, keys);
+ }
}
struct dp_netdev_execute_aux {
- struct dp_netdev *dp;
- const struct flow *key;
+ struct dp_netdev_pmd_thread *pmd;
};
static void
-dp_execute_cb(void *aux_, struct ofpbuf *packet,
- const struct pkt_metadata *md OVS_UNUSED,
+dpif_netdev_register_upcall_cb(struct dpif *dpif, upcall_callback *cb,
+ void *aux)
+{
+ struct dp_netdev *dp = get_dp_netdev(dpif);
+ dp->upcall_aux = aux;
+ dp->upcall_cb = cb;
+}
+
+static void
+dp_netdev_drop_packets(struct dpif_packet ** packets, int cnt, bool may_steal)
+{
+ int i;
+
+ if (may_steal) {
+ for (i = 0; i < cnt; i++) {
+ dpif_packet_delete(packets[i]);
+ }
+ }
+}
+
+static void
+dp_execute_cb(void *aux_, struct dpif_packet **packets, int cnt,
const struct nlattr *a, bool may_steal)
OVS_NO_THREAD_SAFETY_ANALYSIS
{
struct dp_netdev_execute_aux *aux = aux_;
+ uint32_t *depth = recirc_depth_get();
+ struct dp_netdev_pmd_thread *pmd= aux->pmd;
+ struct dp_netdev *dp= pmd->dp;
int type = nl_attr_type(a);
struct dp_netdev_port *p;
+ 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);
+ p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
+ if (OVS_LIKELY(p)) {
+ netdev_send(p->netdev, pmd->core_id, packets, cnt, may_steal);
+ return;
+ }
+ break;
+
+ case OVS_ACTION_ATTR_USERSPACE:
+ if (!fat_rwlock_tryrdlock(&dp->upcall_rwlock)) {
+ const struct nlattr *userdata;
+ struct ofpbuf actions;
+ struct flow flow;
+
+ userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
+ ofpbuf_init(&actions, 0);
+
+ for (i = 0; i < cnt; i++) {
+ int error;
+
+ ofpbuf_clear(&actions);
+
+ flow_extract(&packets[i]->ofpbuf, &packets[i]->md, &flow);
+ error = dp_netdev_upcall(dp, packets[i], &flow, NULL,
+ DPIF_UC_ACTION, userdata, &actions,
+ NULL);
+ if (!error || error == ENOSPC) {
+ dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
+ ofpbuf_data(&actions),
+ ofpbuf_size(&actions));
+ } else if (may_steal) {
+ dpif_packet_delete(packets[i]);
+ }
+ }
+ ofpbuf_uninit(&actions);
+ fat_rwlock_unlock(&dp->upcall_rwlock);
+
+ return;
}
break;
- case OVS_ACTION_ATTR_USERSPACE: {
- const struct nlattr *userdata;
+ case OVS_ACTION_ATTR_HASH: {
+ const struct ovs_action_hash *hash_act;
+ uint32_t hash;
+
+ hash_act = nl_attr_get(a);
+
+ for (i = 0; i < cnt; i++) {
+
+ if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
+ /* Hash need not be symmetric, nor does it need to include
+ * L2 fields. */
+ hash = hash_2words(dpif_packet_get_dp_hash(packets[i]),
+ hash_act->hash_basis);
+ } else {
+ VLOG_WARN("Unknown hash algorithm specified "
+ "for the hash action.");
+ hash = 2;
+ }
- userdata = nl_attr_find_nested(a, OVS_USERSPACE_ATTR_USERDATA);
+ if (!hash) {
+ hash = 1; /* 0 is not valid */
+ }
- /* Make a copy if we are not allowed to steal the packet's data. */
- if (!may_steal) {
- packet = ofpbuf_clone_with_headroom(packet, DP_NETDEV_HEADROOM);
+ dpif_packet_set_dp_hash(packets[i], hash);
}
- dp_netdev_output_userspace(aux->dp, packet, DPIF_UC_ACTION, aux->key,
- userdata);
- if (!may_steal) {
- ofpbuf_uninit(packet);
+ return;
+ }
+
+ case OVS_ACTION_ATTR_RECIRC:
+ if (*depth < MAX_RECIRC_DEPTH) {
+
+ (*depth)++;
+ for (i = 0; i < cnt; i++) {
+ struct dpif_packet *recirc_pkt;
+
+ recirc_pkt = (may_steal) ? packets[i]
+ : dpif_packet_clone(packets[i]);
+
+ recirc_pkt->md.recirc_id = nl_attr_get_u32(a);
+
+ /* Hash is private to each packet */
+ recirc_pkt->md.dp_hash = dpif_packet_get_dp_hash(packets[i]);
+
+ dp_netdev_input(pmd, &recirc_pkt, 1);
+ }
+ (*depth)--;
+
+ return;
}
+
+ VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
break;
- }
+
case OVS_ACTION_ATTR_PUSH_VLAN:
case OVS_ACTION_ATTR_POP_VLAN:
case OVS_ACTION_ATTR_PUSH_MPLS:
case OVS_ACTION_ATTR_POP_MPLS:
case OVS_ACTION_ATTR_SET:
+ case OVS_ACTION_ATTR_SET_MASKED:
case OVS_ACTION_ATTR_SAMPLE:
case OVS_ACTION_ATTR_UNSPEC:
case __OVS_ACTION_ATTR_MAX:
OVS_NOT_REACHED();
}
+
+ dp_netdev_drop_packets(packets, cnt, may_steal);
}
static void
-dp_netdev_execute_actions(struct dp_netdev *dp, const struct flow *key,
- struct ofpbuf *packet, struct pkt_metadata *md,
+dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
+ struct dpif_packet **packets, int cnt,
+ bool may_steal,
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 = { pmd };
- odp_execute_actions(&aux, packet, md, actions, actions_len, dp_execute_cb);
+ odp_execute_actions(&aux, packets, cnt, may_steal, 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_port_dump_done,
dpif_netdev_port_poll,
dpif_netdev_port_poll_wait,
- dpif_netdev_flow_get,
- 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_netdev_operate,
+ NULL, /* recv_set */
+ NULL, /* handlers_set */
+ dpif_netdev_pmd_set,
dpif_netdev_queue_to_priority,
- dpif_netdev_recv,
- dpif_netdev_recv_wait,
- dpif_netdev_recv_purge,
+ NULL, /* recv */
+ NULL, /* recv_wait */
+ NULL, /* recv_purge */
+ dpif_netdev_register_upcall_cb,
+ dpif_netdev_enable_upcall,
+ dpif_netdev_disable_upcall,
};
static void
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);
}
dpif_dummy_register__("dummy");
unixctl_command_register("dpif-dummy/change-port-number",
- "DP PORT NEW-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);
+}
+\f
+/* Datapath Classifier. */
+
+/* A set of rules that all have the same fields wildcarded. */
+struct dpcls_subtable {
+ /* The fields are only used by writers. */
+ struct cmap_node cmap_node OVS_GUARDED; /* Within dpcls 'subtables_map'. */
+
+ /* These fields are accessed by readers. */
+ struct cmap rules; /* Contains "struct dpcls_rule"s. */
+ struct netdev_flow_key mask; /* Wildcards for fields (const). */
+ /* 'mask' must be the last field, additional space is allocated here. */
+};
+
+/* Initializes 'cls' as a classifier that initially contains no classification
+ * rules. */
+static void
+dpcls_init(struct dpcls *cls)
+{
+ cmap_init(&cls->subtables_map);
+ pvector_init(&cls->subtables);
+}
+
+static void
+dpcls_destroy_subtable(struct dpcls *cls, struct dpcls_subtable *subtable)
+{
+ pvector_remove(&cls->subtables, subtable);
+ cmap_remove(&cls->subtables_map, &subtable->cmap_node,
+ subtable->mask.hash);
+ cmap_destroy(&subtable->rules);
+ ovsrcu_postpone(free, subtable);
+}
+
+/* Destroys 'cls'. Rules within 'cls', if any, are not freed; this is the
+ * caller's responsibility.
+ * May only be called after all the readers have been terminated. */
+static void
+dpcls_destroy(struct dpcls *cls)
+{
+ if (cls) {
+ struct dpcls_subtable *subtable;
+
+ CMAP_FOR_EACH (subtable, cmap_node, &cls->subtables_map) {
+ dpcls_destroy_subtable(cls, subtable);
+ }
+ cmap_destroy(&cls->subtables_map);
+ pvector_destroy(&cls->subtables);
+ }
+}
+
+static struct dpcls_subtable *
+dpcls_create_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
+{
+ struct dpcls_subtable *subtable;
+
+ /* Need to add one. */
+ subtable = xmalloc(sizeof *subtable
+ - sizeof subtable->mask.mf + mask->len);
+ cmap_init(&subtable->rules);
+ netdev_flow_key_clone(&subtable->mask, mask);
+ cmap_insert(&cls->subtables_map, &subtable->cmap_node, mask->hash);
+ pvector_insert(&cls->subtables, subtable, 0);
+
+ return subtable;
+}
+
+static inline struct dpcls_subtable *
+dpcls_find_subtable(struct dpcls *cls, const struct netdev_flow_key *mask)
+{
+ struct dpcls_subtable *subtable;
+
+ CMAP_FOR_EACH_WITH_HASH (subtable, cmap_node, mask->hash,
+ &cls->subtables_map) {
+ if (netdev_flow_key_equal(&subtable->mask, mask)) {
+ return subtable;
+ }
+ }
+ return dpcls_create_subtable(cls, mask);
+}
+
+/* Insert 'rule' into 'cls'. */
+static void
+dpcls_insert(struct dpcls *cls, struct dpcls_rule *rule,
+ const struct netdev_flow_key *mask)
+{
+ struct dpcls_subtable *subtable = dpcls_find_subtable(cls, mask);
+
+ rule->mask = &subtable->mask;
+ cmap_insert(&subtable->rules, &rule->cmap_node, rule->flow.hash);
+}
+
+/* Removes 'rule' from 'cls', also destructing the 'rule'. */
+static void
+dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
+{
+ struct dpcls_subtable *subtable;
+
+ ovs_assert(rule->mask);
+
+ INIT_CONTAINER(subtable, rule->mask, mask);
+
+ if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
+ == 0) {
+ dpcls_destroy_subtable(cls, subtable);
+ }
+}
+
+/* Returns true if 'target' satisifies 'key' in 'mask', that is, if each 1-bit
+ * in 'mask' the values in 'key' and 'target' are the same.
+ *
+ * Note: 'key' and 'mask' have the same mask, and 'key' is already masked. */
+static inline bool
+dpcls_rule_matches_key(const struct dpcls_rule *rule,
+ const struct netdev_flow_key *target)
+{
+ const uint32_t *keyp = rule->flow.mf.inline_values;
+ const uint32_t *maskp = rule->mask->mf.inline_values;
+ uint32_t target_u32;
+
+ NETDEV_FLOW_KEY_FOR_EACH_IN_MAP(target_u32, target, rule->flow.mf.map) {
+ if (OVS_UNLIKELY((target_u32 & *maskp++) != *keyp++)) {
+ return false;
+ }
+ }
+ return true;
+}
+
+/* For each miniflow in 'flows' performs a classifier lookup writing the result
+ * into the corresponding slot in 'rules'. If a particular entry in 'flows' is
+ * NULL it is skipped.
+ *
+ * This function is optimized for use in the userspace datapath and therefore
+ * does not implement a lot of features available in the standard
+ * classifier_lookup() function. Specifically, it does not implement
+ * priorities, instead returning any rule which matches the flow.
+ *
+ * Returns true if all flows found a corresponding rule. */
+static bool
+dpcls_lookup(const struct dpcls *cls, const struct netdev_flow_key keys[],
+ struct dpcls_rule **rules, const size_t cnt)
+{
+ /* The batch size 16 was experimentally found faster than 8 or 32. */
+ typedef uint16_t map_type;
+#define MAP_BITS (sizeof(map_type) * CHAR_BIT)
+
+#if !defined(__CHECKER__) && !defined(_WIN32)
+ const int N_MAPS = DIV_ROUND_UP(cnt, MAP_BITS);
+#else
+ enum { N_MAPS = DIV_ROUND_UP(NETDEV_MAX_RX_BATCH, MAP_BITS) };
+#endif
+ map_type maps[N_MAPS];
+ struct dpcls_subtable *subtable;
+
+ memset(maps, 0xff, sizeof maps);
+ if (cnt % MAP_BITS) {
+ maps[N_MAPS - 1] >>= MAP_BITS - cnt % MAP_BITS; /* Clear extra bits. */
+ }
+ memset(rules, 0, cnt * sizeof *rules);
+
+ PVECTOR_FOR_EACH (subtable, &cls->subtables) {
+ const struct netdev_flow_key *mkeys = keys;
+ struct dpcls_rule **mrules = rules;
+ map_type remains = 0;
+ int m;
+
+ BUILD_ASSERT_DECL(sizeof remains == sizeof *maps);
+
+ for (m = 0; m < N_MAPS; m++, mkeys += MAP_BITS, mrules += MAP_BITS) {
+ uint32_t hashes[MAP_BITS];
+ const struct cmap_node *nodes[MAP_BITS];
+ unsigned long map = maps[m];
+ int i;
+
+ if (!map) {
+ continue; /* Skip empty maps. */
+ }
+
+ /* Compute hashes for the remaining keys. */
+ ULONG_FOR_EACH_1(i, map) {
+ hashes[i] = netdev_flow_key_hash_in_mask(&mkeys[i],
+ &subtable->mask);
+ }
+ /* Lookup. */
+ map = cmap_find_batch(&subtable->rules, map, hashes, nodes);
+ /* Check results. */
+ ULONG_FOR_EACH_1(i, map) {
+ struct dpcls_rule *rule;
+
+ CMAP_NODE_FOR_EACH (rule, cmap_node, nodes[i]) {
+ if (OVS_LIKELY(dpcls_rule_matches_key(rule, &mkeys[i]))) {
+ mrules[i] = rule;
+ goto next;
+ }
+ }
+ ULONG_SET0(map, i); /* Did not match. */
+ next:
+ ; /* Keep Sparse happy. */
+ }
+ maps[m] &= ~map; /* Clear the found rules. */
+ remains |= maps[m];
+ }
+ if (!remains) {
+ return true; /* All found. */
+ }
+ }
+ return false; /* Some misses. */
}
projects / cascardo / ovs.git / blobdiff