#include <sys/stat.h>
#include <unistd.h>
-#include "classifier.h"
#include "cmap.h"
#include "csum.h"
#include "dpif.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 "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
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];
+};
+
+/* 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
+ * =============
+ *
+ * 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.
*
*
* changes to 'cls' must be made while still holding the 'flow_mutex'.
*/
struct ovs_mutex flow_mutex;
- struct classifier cls;
+ struct dpcls cls;
struct cmap flow_table OVS_GUARDED; /* Flow table. */
/* Statistics.
upcall_callback *upcall_cb; /* Callback function for executing upcalls. */
void *upcall_aux;
- /* Forwarding threads. */
- struct latch exit_latch;
- struct pmd_thread *pmd_threads;
- size_t n_pmd_threads;
- int pmd_count;
+ /* 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,
char *type; /* Port type as requested by user. */
};
-
-/* Stores a miniflow */
-
-/* There are fields in the flow structure that we never use. Therefore we can
- * save a few words of memory */
-#define NETDEV_KEY_BUF_SIZE_U32 (FLOW_U32S - MINI_N_INLINE \
- - FLOW_U32_SIZE(regs) \
- - FLOW_U32_SIZE(metadata) \
- )
-struct netdev_flow_key {
- struct miniflow flow;
- uint32_t buf[NETDEV_KEY_BUF_SIZE_U32];
-};
-
+\f
/* A flow in dp_netdev's 'flow_table'.
*
*
* 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 cmap_node node; /* In owning dp_netdev's 'flow_table'. */
- const struct flow flow; /* The flow that created this entry. */
+ const struct flow flow; /* Unmasked flow that created this entry. */
/* Number of references.
* The classifier owns one reference.
/* 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 {
*
* DPDK used PMD for accessing NIC.
*
- * A thread that receives packets from PMD ports, looks them up in the flow
- * table, and executes the actions it finds.
+ * 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 pmd_thread {
+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;
- int id;
- atomic_uint change_seq;
+ 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;
OVS_REQUIRES(dp->port_mutex);
static int dpif_netdev_open(const struct dpif_class *, const char *name,
bool create, struct dpif **);
-static void dp_netdev_execute_actions(struct dp_netdev *dp,
+static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
struct dpif_packet **, int c,
- bool may_steal, struct pkt_metadata *,
+ bool may_steal,
const struct nlattr *actions,
size_t actions_len);
-static void dp_netdev_port_input(struct dp_netdev *dp,
- struct dpif_packet **packets, int cnt,
- odp_port_t port_no);
+static void dp_netdev_input(struct dp_netdev_pmd_thread *,
+ struct dpif_packet **, int cnt);
-static void dp_netdev_set_pmd_threads(struct dp_netdev *, int n);
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)
atomic_flag_clear(&dp->destroyed);
ovs_mutex_init(&dp->flow_mutex);
- classifier_init(&dp->cls, NULL);
+ dpcls_init(&dp->cls);
cmap_init(&dp->flow_table);
ovsthread_stats_init(&dp->stats);
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->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);
+
+ /* 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_mutex_unlock(&dp->port_mutex);
return error;
}
+static void
+dp_netdev_destroy_upcall_lock(struct dp_netdev *dp)
+ OVS_NO_THREAD_SAFETY_ANALYSIS
+{
+ /* Check that upcalls are disabled, i.e. that the rwlock is taken */
+ ovs_assert(fat_rwlock_tryrdlock(&dp->upcall_rwlock));
+
+ /* 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'
* through the 'dp_netdevs' shash while freeing 'dp'. */
static void
shash_find_and_delete(&dp_netdevs, dp->name);
- dp_netdev_set_pmd_threads(dp, 0);
- free(dp->pmd_threads);
+ 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_mutex_lock(&dp->port_mutex);
}
ovsthread_stats_destroy(&dp->stats);
- classifier_destroy(&dp->cls);
+ dpcls_destroy(&dp->cls);
cmap_destroy(&dp->flow_table);
ovs_mutex_destroy(&dp->flow_mutex);
seq_destroy(dp->port_seq);
cmap_destroy(&dp->ports);
- fat_rwlock_destroy(&dp->upcall_rwlock);
- latch_destroy(&dp->exit_latch);
+
+ /* 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);
}
}
static void
-dp_netdev_reload_pmd_threads(struct dp_netdev *dp)
+dp_netdev_reload_pmd__(struct dp_netdev_pmd_thread *pmd)
{
- int i;
+ int old_seq;
+
+ atomic_add_relaxed(&pmd->change_seq, 1, &old_seq);
+}
- for (i = 0; i < dp->n_pmd_threads; i++) {
- struct pmd_thread *f = &dp->pmd_threads[i];
- int id;
+/* 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;
- atomic_add(&f->change_seq, 1, &id);
- }
+ CMAP_FOR_EACH (pmd, node, &dp->poll_threads) {
+ dp_netdev_reload_pmd__(pmd);
+ }
}
static uint32_t
return EINVAL;
}
+ 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;
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;
}
}
netdev_rxq_close(port->rxq[i]);
}
netdev_close(netdev);
+ free(port->type);
free(port->rxq);
free(port);
return error;
port->sf = sf;
if (netdev_is_pmd(netdev)) {
- dp->pmd_count++;
- dp_netdev_set_pmd_threads(dp, NR_PMD_THREADS);
- dp_netdev_reload_pmd_threads(dp);
+ dp_netdev_set_pmds_on_numa(dp, netdev_get_numa_id(netdev));
+ dp_netdev_reload_pmds(dp);
}
ovs_refcount_init(&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)
{
return ENOENT;
}
+static int
+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;
+
+ CMAP_FOR_EACH (port, node, &dp->ports) {
+ if (netdev_is_pmd(port->netdev)
+ && netdev_get_numa_id(port->netdev) == numa_id) {
+ return true;
+ }
+ }
+
+ 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)) {
- dp_netdev_reload_pmd_threads(dp);
+ int numa_id = netdev_get_numa_id(port->netdev);
+
+ /* 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);
}
port_unref(port);
}
ovsthread_stats_destroy(&flow->stats);
- cls_rule_destroy(CONST_CAST(struct cls_rule *, &flow->cr));
dp_netdev_actions_free(dp_netdev_flow_get_actions(flow));
free(flow);
}
dp_netdev_remove_flow(struct dp_netdev *dp, struct dp_netdev_flow *flow)
OVS_REQUIRES(dp->flow_mutex)
{
- struct cls_rule *cr = CONST_CAST(struct cls_rule *, &flow->cr);
struct cmap_node *node = CONST_CAST(struct cmap_node *, &flow->node);
- classifier_remove(&dp->cls, cr);
+ 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 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 bool dp_netdev_flow_ref(struct dp_netdev_flow *flow)
+{
+ return ovs_refcount_try_ref_rcu(&flow->ref_cnt);
+}
+
+/* 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));
+
+/* 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 inline bool
+netdev_flow_key_equal(const struct netdev_flow_key *a,
+ const struct netdev_flow_key *b)
+{
+ /* 'b->len' may be not set yet. */
+ return a->hash == b->hash && !memcmp(&a->mf, &b->mf, a->len);
+}
+
+/* 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);
+}
+
+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
+netdev_flow_key_from_flow(struct netdev_flow_key *dst,
+ const struct flow *src)
+{
+ 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. */
+}
+
+/* Initialize a netdev_flow_key 'mask' from 'match'. */
+static inline void
+netdev_flow_mask_init(struct netdev_flow_key *mask,
+ const struct match *match)
+{
+ 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;
+
+ /* Only check masks that make sense for the flow. */
+ map = flow_wc_map(&match->flow);
+
+ while (map) {
+ uint64_t rm1bit = rightmost_1bit(map);
+ int i = raw_ctz(map);
+
+ if (mask_u32[i]) {
+ mask_map |= rm1bit;
+ *dst++ = mask_u32[i];
+ hash = hash_add(hash, mask_u32[i]);
+ }
+ map -= rm1bit;
+ }
+
+ 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 miniflow *key)
+dp_netdev_lookup_flow(const struct dp_netdev *dp,
+ const struct netdev_flow_key *key)
{
struct dp_netdev_flow *netdev_flow;
- struct cls_rule *rule;
+ struct dpcls_rule *rule;
- classifier_lookup_miniflow_batch(&dp->cls, &key, &rule, 1);
+ dpcls_lookup(&dp->cls, key, &rule, 1);
netdev_flow = dp_netdev_flow_cast(rule);
return netdev_flow;
struct flow_wildcards wc;
struct dp_netdev_actions *actions;
- minimask_expand(&netdev_flow->cr.match.mask, &wc);
+ 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);
return error;
}
-static int
+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;
-
- netdev_flow = xzalloc(sizeof *netdev_flow);
- *CONST_CAST(struct flow *, &netdev_flow->flow) = match->flow;
-
- ovs_refcount_init(&netdev_flow->ref_cnt);
+ struct dp_netdev_flow *flow;
+ struct netdev_flow_key mask;
- ovsthread_stats_init(&netdev_flow->stats);
+ netdev_flow_mask_init(&mask, match);
+ /* Make sure wc does not have metadata. */
+ ovs_assert(!(mask.mf.map & (MINIFLOW_MAP(metadata) | MINIFLOW_MAP(regs))));
- ovsrcu_set(&netdev_flow->actions,
- dp_netdev_actions_create(actions, actions_len));
+ /* 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));
- cls_rule_init(CONST_CAST(struct cls_rule *, &netdev_flow->cr),
- match, NETDEV_RULE_PRIORITY);
cmap_insert(&dp->flow_table,
- CONST_CAST(struct cmap_node *, &netdev_flow->node),
- flow_hash(&match->flow, 0));
- classifier_insert(&dp->cls,
- CONST_CAST(struct cls_rule *, &netdev_flow->cr));
+ 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);
if (OVS_UNLIKELY(VLOG_IS_DBG_ENABLED())) {
+ struct match match;
struct ds ds = DS_EMPTY_INITIALIZER;
+ match.flow = flow->flow;
+ miniflow_expand(&flow->cr.mask->mf, &match.wc.masks);
+
ds_put_cstr(&ds, "flow_add: ");
- match_format(match, &ds, OFP_DEFAULT_PRIORITY);
+ match_format(&match, &ds, OFP_DEFAULT_PRIORITY);
ds_put_cstr(&ds, ", actions:");
format_odp_actions(&ds, actions, actions_len);
ds_destroy(&ds);
}
- return 0;
+ return flow;
}
static void
{
struct dp_netdev *dp = get_dp_netdev(dpif);
struct dp_netdev_flow *netdev_flow;
- struct miniflow miniflow;
+ struct netdev_flow_key key;
struct match match;
int error;
if (error) {
return error;
}
- miniflow_init(&miniflow, &match.flow);
+
+ /* 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, &miniflow);
+ netdev_flow = dp_netdev_lookup_flow(dp, &key);
if (!netdev_flow) {
if (put->flags & DPIF_FP_CREATE) {
if (cmap_count(&dp->flow_table) < MAX_FLOWS) {
if (put->stats) {
memset(put->stats, 0, sizeof *put->stats);
}
- error = dp_netdev_flow_add(dp, &match, put->actions,
- put->actions_len);
+ dp_netdev_flow_add(dp, &match, put->actions, put->actions_len);
+ error = 0;
} else {
error = EFBIG;
}
}
}
ovs_mutex_unlock(&dp->flow_mutex);
- miniflow_destroy(&miniflow);
return error;
}
struct flow_wildcards wc;
struct ofpbuf buf;
- minimask_expand(&netdev_flow->cr.match.mask, &wc);
+ miniflow_expand(&netdev_flow->cr.mask->mf, &wc.masks);
/* Key. */
ofpbuf_use_stack(&buf, keybuf, sizeof *keybuf);
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 dp_netdev_pmd_thread *pmd;
struct dpif_packet packet, *pp;
- struct pkt_metadata *md = &execute->md;
if (ofpbuf_size(execute->packet) < ETH_HEADER_LEN ||
ofpbuf_size(execute->packet) > UINT16_MAX) {
}
packet.ofpbuf = *execute->packet;
+ packet.md = execute->md;
pp = &packet;
- dp_netdev_execute_actions(dp, &pp, 1, false, md,
- execute->actions, execute->actions_len);
+ /* 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);
+ 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;
}
}
}
+/* 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)
+{
+ 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);
+ }
+ }
+}
+
+/* Resets pmd threads if the configuration for 'rxq's or cpu mask changes. */
+static int
+dpif_netdev_pmd_set(struct dpif *dpif, unsigned int n_rxqs, const char *cmask)
+{
+ struct dp_netdev *dp = get_dp_netdev(dpif);
+
+ if (pmd_config_changed(dp, n_rxqs, cmask)) {
+ struct dp_netdev_port *port;
+
+ dp_netdev_destroy_all_pmds(dp);
+
+ CMAP_FOR_EACH (port, node, &dp->ports) {
+ if (netdev_is_pmd(port->netdev)) {
+ int i, err;
+
+ /* 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;
+ }
+
+ /* 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;
+ }
+
+ /* 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;
+
+ /* Reconfigures the cpu mask. */
+ ovs_numa_set_cpu_mask(cmask);
+ free(dp->pmd_cmask);
+ dp->pmd_cmask = cmask ? xstrdup(cmask) : NULL;
+
+ /* Restores the non-pmd. */
+ dp_netdev_set_nonpmd(dp);
+ /* Restores all pmd threads. */
+ dp_netdev_reset_pmd_threads(dp);
+ }
+
+ return 0;
+}
+
static int
dpif_netdev_queue_to_priority(const struct dpif *dpif OVS_UNUSED,
uint32_t queue_id, uint32_t *priority)
\f
static void
-dp_netdev_process_rxq_port(struct dp_netdev *dp,
- struct dp_netdev_port *port,
- struct netdev_rxq *rxq)
+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) {
- dp_netdev_port_input(dp, packets, cnt, port->port_no);
+ 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);
+ 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));
+ netdev_get_name(port->netdev), ovs_strerror(error));
}
}
{
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(dp, port, port->rxq[i]);
+ dp_netdev_process_rxq_port(non_pmd, port, port->rxq[i]);
}
}
}
+ ovs_mutex_unlock(&dp->non_pmd_mutex);
}
static void
};
static int
-pmd_load_queues(struct pmd_thread *f,
+pmd_load_queues(struct dp_netdev_pmd_thread *pmd,
struct rxq_poll **ppoll_list, int poll_cnt)
{
- struct dp_netdev *dp = f->dp;
struct rxq_poll *poll_list = *ppoll_list;
struct dp_netdev_port *port;
- int id = f->id;
- int index;
- int i;
+ 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);
+ 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, &f->dp->ports) {
- if (netdev_is_pmd(port->netdev)) {
- int i;
-
- for (i = 0; i < netdev_n_rxq(port->netdev); i++) {
- if ((index % dp->n_pmd_threads) == id) {
- poll_list = xrealloc(poll_list, sizeof *poll_list * (poll_cnt + 1));
-
- port_ref(port);
- poll_list[poll_cnt].port = port;
- poll_list[poll_cnt].rx = port->rxq[i];
- poll_cnt++;
+ 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++;
}
- index++;
}
+ /* Unrefs the port_try_ref(). */
+ port_unref(port);
}
}
static void *
pmd_thread_main(void *f_)
{
- struct pmd_thread *f = f_;
- struct dp_netdev *dp = f->dp;
+ struct dp_netdev_pmd_thread *pmd = f_;
unsigned int lc = 0;
struct rxq_poll *poll_list;
- unsigned int port_seq;
+ unsigned int port_seq = PMD_INITIAL_SEQ;
int poll_cnt;
int i;
poll_cnt = 0;
poll_list = NULL;
- pmd_thread_setaffinity_cpu(f->id);
+ /* 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:
- poll_cnt = pmd_load_queues(f, &poll_list, poll_cnt);
- atomic_read(&f->change_seq, &port_seq);
+ emc_cache_init(&pmd->flow_cache);
+ poll_cnt = pmd_load_queues(pmd, &poll_list, poll_cnt);
for (;;) {
- unsigned int c_port_seq;
int i;
for (i = 0; i < poll_cnt; i++) {
- dp_netdev_process_rxq_port(dp, poll_list[i].port, poll_list[i].rx);
+ dp_netdev_process_rxq_port(pmd, poll_list[i].port, poll_list[i].rx);
}
if (lc++ > 1024) {
- ovsrcu_quiesce();
+ unsigned int seq;
- /* XXX: need completely userspace based signaling method.
- * to keep this thread entirely in userspace.
- * For now using atomic counter. */
lc = 0;
- atomic_read_explicit(&f->change_seq, &c_port_seq, memory_order_consume);
- if (c_port_seq != port_seq) {
+
+ ovsrcu_quiesce();
+
+ atomic_read_relaxed(&pmd->change_seq, &seq);
+ if (seq != port_seq) {
+ port_seq = seq;
break;
}
}
}
- if (!latch_is_set(&f->dp->exit_latch)){
+ emc_cache_uninit(&pmd->flow_cache);
+
+ if (!latch_is_set(&pmd->exit_latch)){
goto reload;
}
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_pmd_threads(struct dp_netdev *dp, int n)
+dp_netdev_set_nonpmd(struct dp_netdev *dp)
{
- int i;
+ struct dp_netdev_pmd_thread *non_pmd;
- if (n == dp->n_pmd_threads) {
- return;
+ 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);
- dp_netdev_reload_pmd_threads(dp);
- for (i = 0; i < dp->n_pmd_threads; i++) {
- struct pmd_thread *f = &dp->pmd_threads[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->pmd_threads);
- /* Start new threads. */
- dp->pmd_threads = xmalloc(n * sizeof *dp->pmd_threads);
- dp->n_pmd_threads = n;
+ n_pmds = get_n_pmd_threads_on_numa(dp, numa_id);
- for (i = 0; i < n; i++) {
- struct pmd_thread *f = &dp->pmd_threads[i];
+ /* 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->id = i;
- atomic_store(&f->change_seq, 1);
+ 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;
+ }
- /* Each thread will distribute all devices rx-queues among
- * themselves. */
- f->thread = ovs_thread_create("pmd", pmd_thread_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);
}
}
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,
int cnt, int size,
put_actions, dp->upcall_aux);
}
+static inline uint32_t
+dpif_netdev_packet_get_dp_hash(struct dpif_packet *packet,
+ const struct miniflow *mf)
+{
+ 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;
struct dp_netdev_flow *flow;
struct dpif_packet *packets[NETDEV_MAX_RX_BATCH];
- struct pkt_metadata md;
};
static inline void
-packet_batch_update(struct packet_batch *batch,
- struct dpif_packet *packet, const struct miniflow *mf)
+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;
}
static inline void
-packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow,
- struct pkt_metadata *md)
+packet_batch_init(struct packet_batch *batch, struct dp_netdev_flow *flow)
{
batch->flow = flow;
- batch->md = *md;
batch->packet_count = 0;
batch->byte_count = 0;
}
static inline void
-packet_batch_execute(struct packet_batch *batch, struct dp_netdev *dp)
+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;
actions = dp_netdev_flow_get_actions(flow);
- dp_netdev_execute_actions(dp, batch->packets,
- batch->packet_count, true, &batch->md,
+ dp_netdev_execute_actions(pmd, batch->packets, batch->packet_count, true,
actions->actions, actions->size);
- dp_netdev_count_packet(dp, DP_STAT_HIT, batch->packet_count);
+ dp_netdev_count_packet(pmd->dp, DP_STAT_HIT, batch->packet_count);
}
-static void
-dp_netdev_input(struct dp_netdev *dp, struct dpif_packet **packets, int cnt,
- struct pkt_metadata *md)
+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;
+ }
+ }
+ if (OVS_UNLIKELY(*n_batches >= max_batches)) {
+ return false;
+ }
+
+ 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;
+}
+
+/* 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 packet_batch batches[NETDEV_MAX_RX_BATCH];
- struct netdev_flow_key keys[NETDEV_MAX_RX_BATCH];
- const struct miniflow *mfs[NETDEV_MAX_RX_BATCH]; /* NULL at bad packets. */
- struct cls_rule *rules[NETDEV_MAX_RX_BATCH];
+ struct netdev_flow_key key;
+ struct packet_batch batches[4];
+ struct emc_cache *flow_cache = &pmd->flow_cache;
size_t n_batches, i;
- bool any_miss;
+ 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]);
- mfs[i] = NULL;
continue;
}
- miniflow_initialize(&keys[i].flow, keys[i].buf);
- miniflow_extract(&packets[i]->ofpbuf, md, &keys[i].flow);
- mfs[i] = &keys[i].flow;
+ 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);
+
+ 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]);
+ }
+
+ keys[notfound_cnt++] = key;
+ }
+ }
+
+ for (i = 0; i < n_batches; i++) {
+ packet_batch_execute(&batches[i], pmd);
}
- any_miss = !classifier_lookup_miniflow_batch(&dp->cls, mfs, rules, cnt);
+ 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;
- struct match match;
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++) {
- const struct dp_netdev_flow *netdev_flow;
+ struct dp_netdev_flow *netdev_flow;
struct ofpbuf *add_actions;
+ struct match match;
int error;
- if (OVS_LIKELY(rules[i] || !mfs[i])) {
+ if (OVS_LIKELY(rules[i])) {
continue;
}
/* It's possible that an earlier slow path execution installed
- * the rule this flow needs. In this case, it's a lot cheaper
+ * 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, mfs[i]);
+ netdev_flow = dp_netdev_lookup_flow(dp, &keys[i]);
if (netdev_flow) {
- rules[i] = CONST_CAST(struct cls_rule *, &netdev_flow->cr);
+ rules[i] = &netdev_flow->cr;
continue;
}
- miniflow_expand(mfs[i], &match.flow);
+ 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);
+ 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(dp, &packets[i], 1, false, md,
+ dp_netdev_execute_actions(pmd, &packets[i], 1, true,
ofpbuf_data(&actions),
ofpbuf_size(&actions));
? &put_actions
: &actions;
- ovs_mutex_lock(&dp->flow_mutex);
- /* XXX: There's a brief race where this flow could have already
- * been installed since we last did the flow lookup. 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. */
- if (OVS_LIKELY(error != ENOSPC)
- && !dp_netdev_lookup_flow(dp, mfs[i])) {
- dp_netdev_flow_add(dp, &match, ofpbuf_data(add_actions),
- ofpbuf_size(add_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);
}
- ovs_mutex_unlock(&dp->flow_mutex);
}
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;
- struct packet_batch *batch;
- size_t j;
- if (OVS_UNLIKELY(!rules[i] || !mfs[i])) {
+ if (OVS_UNLIKELY(!rules[i])) {
continue;
}
- /* XXX: This O(n^2) algortihm makes sense if we're operating under the
- * assumption that the number of distinct flows (and therefore the
- * number of distinct batches) is quite small. If this turns out not
- * to be the case, it may make sense to pre sort based on the
- * netdev_flow pointer. That done we can get the appropriate batching
- * in O(n * log(n)) instead. */
- batch = NULL;
flow = dp_netdev_flow_cast(rules[i]);
- for (j = 0; j < n_batches; j++) {
- if (batches[j].flow == flow) {
- batch = &batches[j];
- break;
- }
- }
- if (!batch) {
- batch = &batches[n_batches++];
- packet_batch_init(batch, flow, md);
- }
- packet_batch_update(batch, packets[i], mfs[i]);
+ 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], dp);
+ packet_batch_execute(&batches[i], pmd);
}
}
static void
-dp_netdev_port_input(struct dp_netdev *dp, struct dpif_packet **packets,
- int cnt, odp_port_t port_no)
+dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
+ struct dpif_packet **packets, int cnt)
{
- uint32_t *recirc_depth = recirc_depth_get();
- struct pkt_metadata md = PKT_METADATA_INITIALIZER(port_no);
+#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;
- *recirc_depth = 0;
- dp_netdev_input(dp, packets, cnt, &md);
+ 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;
+ struct dp_netdev_pmd_thread *pmd;
};
static void
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,
- struct pkt_metadata *md,
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 *dp = aux->dp;
+ 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;
case OVS_ACTION_ATTR_OUTPUT:
p = dp_netdev_lookup_port(dp, u32_to_odp(nl_attr_get_u32(a)));
if (OVS_LIKELY(p)) {
- netdev_send(p->netdev, packets, cnt, may_steal);
- } else if (may_steal) {
- for (i = 0; i < cnt; i++) {
- dpif_packet_delete(packets[i]);
- }
+ netdev_send(p->netdev, pmd->core_id, packets, cnt, may_steal);
+ return;
}
break;
ofpbuf_clear(&actions);
- flow_extract(&packets[i]->ofpbuf, md, &flow);
+ 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(dp, &packets[i], 1, false, md,
+ dp_netdev_execute_actions(pmd, &packets[i], 1, may_steal,
ofpbuf_data(&actions),
ofpbuf_size(&actions));
- }
-
- if (may_steal) {
+ } else if (may_steal) {
dpif_packet_delete(packets[i]);
}
}
ofpbuf_uninit(&actions);
fat_rwlock_unlock(&dp->upcall_rwlock);
- }
+ return;
+ }
break;
case OVS_ACTION_ATTR_HASH: {
const struct ovs_action_hash *hash_act;
- struct netdev_flow_key key;
uint32_t hash;
hash_act = nl_attr_get(a);
- miniflow_initialize(&key.flow, key.buf);
-
for (i = 0; i < cnt; i++) {
- /* XXX: this is slow. Use RSS hash in the future */
- miniflow_extract(&packets[i]->ofpbuf, md, &key.flow);
-
if (hash_act->hash_alg == OVS_HASH_ALG_L4) {
/* Hash need not be symmetric, nor does it need to include
* L2 fields. */
- hash = miniflow_hash_5tuple(&key.flow, hash_act->hash_basis);
+ 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 = 1; /* 0 is not valid */
}
- if (i == 0) {
- md->dp_hash = hash;
- }
- packets[i]->dp_hash = hash;
+ dpif_packet_set_dp_hash(packets[i], hash);
}
- break;
+ return;
}
case OVS_ACTION_ATTR_RECIRC:
(*depth)++;
for (i = 0; i < cnt; i++) {
struct dpif_packet *recirc_pkt;
- struct pkt_metadata recirc_md = *md;
recirc_pkt = (may_steal) ? packets[i]
: dpif_packet_clone(packets[i]);
- recirc_md.recirc_id = nl_attr_get_u32(a);
+ recirc_pkt->md.recirc_id = nl_attr_get_u32(a);
/* Hash is private to each packet */
- recirc_md.dp_hash = packets[i]->dp_hash;
+ recirc_pkt->md.dp_hash = dpif_packet_get_dp_hash(packets[i]);
- dp_netdev_input(dp, &recirc_pkt, 1, &recirc_md);
+ dp_netdev_input(pmd, &recirc_pkt, 1);
}
(*depth)--;
- break;
- } else {
- VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
- if (may_steal) {
- for (i = 0; i < cnt; i++) {
- dpif_packet_delete(packets[i]);
- }
- }
+ return;
}
+
+ VLOG_WARN("Packet dropped. Max recirculation depth exceeded.");
break;
case OVS_ACTION_ATTR_PUSH_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,
+dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
struct dpif_packet **packets, int cnt,
- bool may_steal, struct pkt_metadata *md,
+ bool may_steal,
const struct nlattr *actions, size_t actions_len)
{
- struct dp_netdev_execute_aux aux = {dp};
+ struct dp_netdev_execute_aux aux = { pmd };
- odp_execute_actions(&aux, packets, cnt, may_steal, md, actions,
+ odp_execute_actions(&aux, packets, cnt, may_steal, actions,
actions_len, dp_execute_cb);
}
dpif_netdev_operate,
NULL, /* recv_set */
NULL, /* handlers_set */
+ dpif_netdev_pmd_set,
dpif_netdev_queue_to_priority,
NULL, /* recv */
NULL, /* recv_wait */
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",
+ 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. */
+}