X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=ofproto%2Fofproto-dpif-upcall.c;h=cc4982fe8a560eed69d9f7c02d8b87250b66f7e6;hb=f72e3073d70232ad6bebb9186cbf37a14793fb95;hp=bfe85fe91a8bcaaba04646df49ce50727f7fa193;hpb=0a8763fcb31bfca0d8d854c235c531005088fcb9;p=sliver-openvswitch.git diff --git a/ofproto/ofproto-dpif-upcall.c b/ofproto/ofproto-dpif-upcall.c index bfe85fe91..cc4982fe8 100644 --- a/ofproto/ofproto-dpif-upcall.c +++ b/ofproto/ofproto-dpif-upcall.c @@ -41,7 +41,6 @@ #define MAX_QUEUE_LENGTH 512 #define FLOW_MISS_MAX_BATCH 50 #define REVALIDATE_MAX_BATCH 50 -#define MAX_IDLE 1500 VLOG_DEFINE_THIS_MODULE(ofproto_dpif_upcall); @@ -202,7 +201,6 @@ struct flow_miss { struct ofproto_dpif *ofproto; struct flow flow; - enum odp_key_fitness key_fitness; const struct nlattr *key; size_t key_len; enum dpif_upcall_type upcall_type; @@ -231,12 +229,15 @@ static void *udpif_revalidator(void *); static uint64_t udpif_get_n_flows(struct udpif *); static void revalidate_udumps(struct revalidator *, struct list *udumps); static void revalidator_sweep(struct revalidator *); +static void revalidator_purge(struct revalidator *); static void upcall_unixctl_show(struct unixctl_conn *conn, int argc, const char *argv[], void *aux); static void upcall_unixctl_disable_megaflows(struct unixctl_conn *, int argc, const char *argv[], void *aux); static void upcall_unixctl_enable_megaflows(struct unixctl_conn *, int argc, const char *argv[], void *aux); +static void upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, int argc, + const char *argv[], void *aux); static void ukey_delete(struct revalidator *, struct udpif_key *); static atomic_bool enable_megaflows = ATOMIC_VAR_INIT(true); @@ -254,6 +255,8 @@ udpif_create(struct dpif_backer *backer, struct dpif *dpif) upcall_unixctl_disable_megaflows, NULL); unixctl_command_register("upcall/enable-megaflows", "", 0, 0, upcall_unixctl_enable_megaflows, NULL); + unixctl_command_register("upcall/set-flow-limit", "", 1, 1, + upcall_unixctl_set_flow_limit, NULL); ovsthread_once_done(&once); } @@ -276,15 +279,12 @@ void udpif_destroy(struct udpif *udpif) { udpif_set_threads(udpif, 0, 0); - udpif_flush(); + udpif_flush(udpif); list_remove(&udpif->list_node); latch_destroy(&udpif->exit_latch); seq_destroy(udpif->reval_seq); seq_destroy(udpif->dump_seq); - atomic_destroy(&udpif->flow_limit); - atomic_destroy(&udpif->n_flows); - atomic_destroy(&udpif->n_flows_timestamp); ovs_mutex_destroy(&udpif->n_flows_mutex); free(udpif); } @@ -329,7 +329,6 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, for (i = 0; i < udpif->n_revalidators; i++) { struct revalidator *revalidator = &udpif->revalidators[i]; struct udpif_flow_dump *udump, *next_udump; - struct udpif_key *ukey, *next_ukey; LIST_FOR_EACH_SAFE (udump, next_udump, list_node, &revalidator->udumps) { @@ -337,10 +336,9 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, free(udump); } - HMAP_FOR_EACH_SAFE (ukey, next_ukey, hmap_node, - &revalidator->ukeys) { - ukey_delete(revalidator, ukey); - } + /* Delete ukeys, and delete all flows from the datapath to prevent + * double-counting stats. */ + revalidator_purge(revalidator); hmap_destroy(&revalidator->ukeys); ovs_mutex_destroy(&revalidator->mutex); @@ -409,6 +407,22 @@ udpif_set_threads(struct udpif *udpif, size_t n_handlers, } } +/* Waits for all ongoing upcall translations to complete. This ensures that + * there are no transient references to any removed ofprotos (or other + * objects). In particular, this should be called after an ofproto is removed + * (e.g. via xlate_remove_ofproto()) but before it is destroyed. */ +void +udpif_synchronize(struct udpif *udpif) +{ + /* This is stronger than necessary. It would be sufficient to ensure + * (somehow) that each handler and revalidator thread had passed through + * its main loop once. */ + size_t n_handlers = udpif->n_handlers; + size_t n_revalidators = udpif->n_revalidators; + udpif_set_threads(udpif, 0, 0); + udpif_set_threads(udpif, n_handlers, n_revalidators); +} + /* Notifies 'udpif' that something changed which may render previous * xlate_actions() results invalid. */ void @@ -456,16 +470,31 @@ udpif_get_memory_usage(struct udpif *udpif, struct simap *usage) } } -/* Removes all flows from all datapaths. */ +/* Remove flows from a single datapath. */ void -udpif_flush(void) +udpif_flush(struct udpif *udpif) +{ + size_t n_handlers, n_revalidators; + + n_handlers = udpif->n_handlers; + n_revalidators = udpif->n_revalidators; + + udpif_set_threads(udpif, 0, 0); + dpif_flow_flush(udpif->dpif); + udpif_set_threads(udpif, n_handlers, n_revalidators); +} + +/* Removes all flows from all datapaths. */ +static void +udpif_flush_all_datapaths(void) { struct udpif *udpif; LIST_FOR_EACH (udpif, list_node, &all_udpifs) { - dpif_flow_flush(udpif->dpif); + udpif_flush(udpif); } } + /* Destroys and deallocates 'upcall'. */ static void @@ -534,6 +563,8 @@ udpif_flow_dumper(void *arg) bool need_revalidate; uint64_t reval_seq; size_t n_flows, i; + int error; + void *state = NULL; reval_seq = seq_read(udpif->reval_seq); need_revalidate = udpif->last_reval_seq != reval_seq; @@ -544,9 +575,14 @@ udpif_flow_dumper(void *arg) udpif->avg_n_flows = (udpif->avg_n_flows + n_flows) / 2; start_time = time_msec(); - dpif_flow_dump_start(&dump, udpif->dpif); - while (dpif_flow_dump_next(&dump, &key, &key_len, &mask, &mask_len, - NULL, NULL, &stats) + error = dpif_flow_dump_start(&dump, udpif->dpif); + if (error) { + VLOG_INFO("Failed to start flow dump (%s)", ovs_strerror(error)); + goto skip; + } + dpif_flow_dump_state_init(udpif->dpif, &state); + while (dpif_flow_dump_next(&dump, state, &key, &key_len, + &mask, &mask_len, NULL, NULL, &stats) && !latch_is_set(&udpif->exit_latch)) { struct udpif_flow_dump *udump = xmalloc(sizeof *udump); struct revalidator *revalidator; @@ -577,6 +613,7 @@ udpif_flow_dumper(void *arg) xpthread_cond_signal(&revalidator->wake_cond); ovs_mutex_unlock(&revalidator->mutex); } + dpif_flow_dump_state_uninit(udpif->dpif, state); dpif_flow_dump_done(&dump); /* Let all the revalidators finish and garbage collect. */ @@ -619,7 +656,8 @@ udpif_flow_dumper(void *arg) duration); } - poll_timer_wait_until(start_time + MIN(MAX_IDLE, 500)); +skip: + poll_timer_wait_until(start_time + MIN(ofproto_max_idle, 500)); seq_wait(udpif->reval_seq, udpif->last_reval_seq); latch_wait(&udpif->exit_latch); poll_block(); @@ -747,16 +785,16 @@ classify_upcall(const struct upcall *upcall) } memset(&cookie, 0, sizeof cookie); memcpy(&cookie, nl_attr_get(dpif_upcall->userdata), userdata_len); - if (userdata_len == sizeof cookie.sflow + if (userdata_len == MAX(8, sizeof cookie.sflow) && cookie.type == USER_ACTION_COOKIE_SFLOW) { return SFLOW_UPCALL; - } else if (userdata_len == sizeof cookie.slow_path + } else if (userdata_len == MAX(8, sizeof cookie.slow_path) && cookie.type == USER_ACTION_COOKIE_SLOW_PATH) { return MISS_UPCALL; - } else if (userdata_len == sizeof cookie.flow_sample + } else if (userdata_len == MAX(8, sizeof cookie.flow_sample) && cookie.type == USER_ACTION_COOKIE_FLOW_SAMPLE) { return FLOW_SAMPLE_UPCALL; - } else if (userdata_len == sizeof cookie.ipfix + } else if (userdata_len == MAX(8, sizeof cookie.ipfix) && cookie.type == USER_ACTION_COOKIE_IPFIX) { return IPFIX_UPCALL; } else { @@ -941,7 +979,7 @@ handle_upcalls(struct handler *handler, struct list *upcalls) int error; error = xlate_receive(udpif->backer, packet, dupcall->key, - dupcall->key_len, &flow, &miss->key_fitness, + dupcall->key_len, &flow, &ofproto, &ipfix, &sflow, NULL, &odp_in_port); if (error) { if (error == ENODEV) { @@ -965,9 +1003,10 @@ handle_upcalls(struct handler *handler, struct list *upcalls) type = classify_upcall(upcall); if (type == MISS_UPCALL) { uint32_t hash; + struct pkt_metadata md; - flow_extract(packet, flow.skb_priority, flow.pkt_mark, - &flow.tunnel, &flow.in_port, &miss->flow); + pkt_metadata_from_flow(&md, &flow); + flow_extract(packet, &md, &miss->flow); hash = flow_hash(&miss->flow, 0); existing_miss = flow_miss_find(&misses, ofproto, &miss->flow, @@ -1139,8 +1178,11 @@ handle_upcalls(struct handler *handler, struct list *upcalls) atomic_read(&enable_megaflows, &megaflow); ofpbuf_use_stack(&mask, &miss->mask_buf, sizeof miss->mask_buf); if (megaflow) { + size_t max_mpls; + + max_mpls = ofproto_dpif_get_max_mpls_depth(miss->ofproto); odp_flow_key_from_mask(&mask, &miss->xout.wc.masks, - &miss->flow, UINT32_MAX); + &miss->flow, UINT32_MAX, max_mpls); } op = &ops[n_ops++]; @@ -1246,6 +1288,22 @@ ukey_lookup(struct revalidator *revalidator, struct udpif_flow_dump *udump) return NULL; } +static struct udpif_key * +ukey_create(const struct nlattr *key, size_t key_len, long long int used) +{ + struct udpif_key *ukey = xmalloc(sizeof *ukey); + + ukey->key = (struct nlattr *) &ukey->key_buf; + memcpy(&ukey->key_buf, key, key_len); + ukey->key_len = key_len; + + ukey->mark = false; + ukey->created = used ? used : time_msec(); + memset(&ukey->stats, 0, sizeof ukey->stats); + + return ukey; +} + static void ukey_delete(struct revalidator *revalidator, struct udpif_key *ukey) { @@ -1300,7 +1358,7 @@ revalidate_ukey(struct udpif *udpif, struct udpif_flow_dump *udump, } error = xlate_receive(udpif->backer, NULL, ukey->key, ukey->key_len, &flow, - NULL, &ofproto, NULL, NULL, NULL, &odp_in_port); + &ofproto, NULL, NULL, NULL, &odp_in_port); if (error) { goto exit; } @@ -1354,20 +1412,106 @@ exit: return ok; } +struct dump_op { + struct udpif_key *ukey; + struct udpif_flow_dump *udump; + struct dpif_flow_stats stats; /* Stats for 'op'. */ + struct dpif_op op; /* Flow del operation. */ +}; + static void -revalidate_udumps(struct revalidator *revalidator, struct list *udumps) +dump_op_init(struct dump_op *op, const struct nlattr *key, size_t key_len, + struct udpif_key *ukey, struct udpif_flow_dump *udump) +{ + op->ukey = ukey; + op->udump = udump; + op->op.type = DPIF_OP_FLOW_DEL; + op->op.u.flow_del.key = key; + op->op.u.flow_del.key_len = key_len; + op->op.u.flow_del.stats = &op->stats; +} + +static void +push_dump_ops(struct revalidator *revalidator, + struct dump_op *ops, size_t n_ops) { struct udpif *udpif = revalidator->udpif; + struct dpif_op *opsp[REVALIDATE_MAX_BATCH]; + size_t i; - struct { - struct dpif_flow_stats ukey_stats; /* Stats stored in the ukey. */ - struct dpif_flow_stats stats; /* Stats for 'op'. */ - struct dpif_op op; /* Flow del operation. */ - } ops[REVALIDATE_MAX_BATCH]; + ovs_assert(n_ops <= REVALIDATE_MAX_BATCH); + for (i = 0; i < n_ops; i++) { + opsp[i] = &ops[i].op; + } + dpif_operate(udpif->dpif, opsp, n_ops); - struct dpif_op *opsp[REVALIDATE_MAX_BATCH]; + for (i = 0; i < n_ops; i++) { + struct dump_op *op = &ops[i]; + struct dpif_flow_stats *push, *stats, push_buf; + + stats = op->op.u.flow_del.stats; + if (op->ukey) { + push = &push_buf; + push->used = MAX(stats->used, op->ukey->stats.used); + push->tcp_flags = stats->tcp_flags | op->ukey->stats.tcp_flags; + push->n_packets = stats->n_packets - op->ukey->stats.n_packets; + push->n_bytes = stats->n_bytes - op->ukey->stats.n_bytes; + } else { + push = stats; + } + + if (push->n_packets || netflow_exists()) { + struct ofproto_dpif *ofproto; + struct netflow *netflow; + struct flow flow; + + if (!xlate_receive(udpif->backer, NULL, op->op.u.flow_del.key, + op->op.u.flow_del.key_len, &flow, &ofproto, + NULL, NULL, &netflow, NULL)) { + struct xlate_in xin; + + xlate_in_init(&xin, ofproto, &flow, NULL, push->tcp_flags, + NULL); + xin.resubmit_stats = push->n_packets ? push : NULL; + xin.may_learn = push->n_packets > 0; + xin.skip_wildcards = true; + xlate_actions_for_side_effects(&xin); + + if (netflow) { + netflow_expire(netflow, &flow); + netflow_flow_clear(netflow, &flow); + netflow_unref(netflow); + } + } + } + } + + for (i = 0; i < n_ops; i++) { + struct udpif_key *ukey; + + /* If there's a udump, this ukey came directly from a datapath flow + * dump. Sometimes a datapath can send duplicates in flow dumps, in + * which case we wouldn't want to double-free a ukey, so avoid that by + * looking up the ukey again. + * + * If there's no udump then we know what we're doing. */ + ukey = (ops[i].udump + ? ukey_lookup(revalidator, ops[i].udump) + : ops[i].ukey); + if (ukey) { + ukey_delete(revalidator, ukey); + } + } +} + +static void +revalidate_udumps(struct revalidator *revalidator, struct list *udumps) +{ + struct udpif *udpif = revalidator->udpif; + + struct dump_op ops[REVALIDATE_MAX_BATCH]; struct udpif_flow_dump *udump, *next_udump; - size_t n_ops, i, n_flows; + size_t n_ops, n_flows; unsigned int flow_limit; long long int max_idle; bool must_del; @@ -1377,7 +1521,7 @@ revalidate_udumps(struct revalidator *revalidator, struct list *udumps) n_flows = udpif_get_n_flows(udpif); must_del = false; - max_idle = MAX_IDLE; + max_idle = ofproto_max_idle; if (n_flows > flow_limit) { must_del = n_flows > 2 * flow_limit; max_idle = 100; @@ -1397,37 +1541,14 @@ revalidate_udumps(struct revalidator *revalidator, struct list *udumps) } if (must_del || (used && used < now - max_idle)) { - struct dpif_flow_stats *ukey_stats = &ops[n_ops].ukey_stats; - struct dpif_op *op = &ops[n_ops].op; - - op->type = DPIF_OP_FLOW_DEL; - op->u.flow_del.key = udump->key; - op->u.flow_del.key_len = udump->key_len; - op->u.flow_del.stats = &ops[n_ops].stats; - n_ops++; - - if (ukey) { - *ukey_stats = ukey->stats; - ukey_delete(revalidator, ukey); - } else { - memset(ukey_stats, 0, sizeof *ukey_stats); - } + struct dump_op *dop = &ops[n_ops++]; + dump_op_init(dop, udump->key, udump->key_len, ukey, udump); continue; } if (!ukey) { - ukey = xmalloc(sizeof *ukey); - - ukey->key = (struct nlattr *) &ukey->key_buf; - memcpy(ukey->key, udump->key, udump->key_len); - ukey->key_len = udump->key_len; - - ukey->created = used ? used : now; - memset(&ukey->stats, 0, sizeof ukey->stats); - - ukey->mark = false; - + ukey = ukey_create(udump->key, udump->key_len, used); hmap_insert(&revalidator->ukeys, &ukey->hmap_node, udump->key_hash); } @@ -1442,46 +1563,7 @@ revalidate_udumps(struct revalidator *revalidator, struct list *udumps) free(udump); } - for (i = 0; i < n_ops; i++) { - opsp[i] = &ops[i].op; - } - dpif_operate(udpif->dpif, opsp, n_ops); - - for (i = 0; i < n_ops; i++) { - struct dpif_flow_stats push, *stats, *ukey_stats; - - ukey_stats = &ops[i].ukey_stats; - stats = ops[i].op.u.flow_del.stats; - push.used = MAX(stats->used, ukey_stats->used); - push.tcp_flags = stats->tcp_flags | ukey_stats->tcp_flags; - push.n_packets = stats->n_packets - ukey_stats->n_packets; - push.n_bytes = stats->n_bytes - ukey_stats->n_bytes; - - if (push.n_packets || netflow_exists()) { - struct ofproto_dpif *ofproto; - struct netflow *netflow; - struct flow flow; - - if (!xlate_receive(udpif->backer, NULL, ops[i].op.u.flow_del.key, - ops[i].op.u.flow_del.key_len, &flow, NULL, - &ofproto, NULL, NULL, &netflow, NULL)) { - struct xlate_in xin; - - xlate_in_init(&xin, ofproto, &flow, NULL, push.tcp_flags, - NULL); - xin.resubmit_stats = push.n_packets ? &push : NULL; - xin.may_learn = push.n_packets > 0; - xin.skip_wildcards = true; - xlate_actions_for_side_effects(&xin); - - if (netflow) { - netflow_expire(netflow, &flow); - netflow_flow_clear(netflow, &flow); - netflow_unref(netflow); - } - } - } - } + push_dump_ops(revalidator, ops, n_ops); LIST_FOR_EACH_SAFE (udump, next_udump, list_node, udumps) { list_remove(&udump->list_node); @@ -1490,17 +1572,46 @@ revalidate_udumps(struct revalidator *revalidator, struct list *udumps) } static void -revalidator_sweep(struct revalidator *revalidator) +revalidator_sweep__(struct revalidator *revalidator, bool purge) { + struct dump_op ops[REVALIDATE_MAX_BATCH]; struct udpif_key *ukey, *next; + size_t n_ops; + + n_ops = 0; HMAP_FOR_EACH_SAFE (ukey, next, hmap_node, &revalidator->ukeys) { - if (ukey->mark) { + if (!purge && ukey->mark) { ukey->mark = false; } else { - ukey_delete(revalidator, ukey); + struct dump_op *op = &ops[n_ops++]; + + /* If we have previously seen a flow in the datapath, but didn't + * see it during the most recent dump, delete it. This allows us + * to clean up the ukey and keep the statistics consistent. */ + dump_op_init(op, ukey->key, ukey->key_len, ukey, NULL); + if (n_ops == REVALIDATE_MAX_BATCH) { + push_dump_ops(revalidator, ops, n_ops); + n_ops = 0; + } } } + + if (n_ops) { + push_dump_ops(revalidator, ops, n_ops); + } +} + +static void +revalidator_sweep(struct revalidator *revalidator) +{ + revalidator_sweep__(revalidator, false); +} + +static void +revalidator_purge(struct revalidator *revalidator) +{ + revalidator_sweep__(revalidator, true); } static void @@ -1561,7 +1672,7 @@ upcall_unixctl_disable_megaflows(struct unixctl_conn *conn, void *aux OVS_UNUSED) { atomic_store(&enable_megaflows, false); - udpif_flush(); + udpif_flush_all_datapaths(); unixctl_command_reply(conn, "megaflows disabled"); } @@ -1576,6 +1687,28 @@ upcall_unixctl_enable_megaflows(struct unixctl_conn *conn, void *aux OVS_UNUSED) { atomic_store(&enable_megaflows, true); - udpif_flush(); + udpif_flush_all_datapaths(); unixctl_command_reply(conn, "megaflows enabled"); } + +/* Set the flow limit. + * + * This command is only needed for advanced debugging, so it's not + * documented in the man page. */ +static void +upcall_unixctl_set_flow_limit(struct unixctl_conn *conn, + int argc OVS_UNUSED, + const char *argv[] OVS_UNUSED, + void *aux OVS_UNUSED) +{ + struct ds ds = DS_EMPTY_INITIALIZER; + struct udpif *udpif; + unsigned int flow_limit = atoi(argv[1]); + + LIST_FOR_EACH (udpif, list_node, &all_udpifs) { + atomic_store(&udpif->flow_limit, flow_limit); + } + ds_put_format(&ds, "set flow_limit to %u\n", flow_limit); + unixctl_command_reply(conn, ds_cstr(&ds)); + ds_destroy(&ds); +}