X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=ofproto%2Fconnmgr.c;h=9a5167de89100ed8e67fc4a62bdad625d2dbfe76;hb=dc723c447a797e555d400594133a35b9841eb1de;hp=6caad060012b266787f0e74f0be496cd814c27c8;hpb=5bd31620685801af66c022878b6029faa16d12b1;p=sliver-openvswitch.git diff --git a/ofproto/connmgr.c b/ofproto/connmgr.c index 6caad0600..9a5167de8 100644 --- a/ofproto/connmgr.c +++ b/ofproto/connmgr.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2009, 2010, 2011 Nicira Networks. + * Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -25,6 +25,8 @@ #include "fail-open.h" #include "in-band.h" #include "odp-util.h" +#include "ofp-actions.h" +#include "ofp-msgs.h" #include "ofp-util.h" #include "ofpbuf.h" #include "ofproto-provider.h" @@ -33,6 +35,7 @@ #include "pktbuf.h" #include "rconn.h" #include "shash.h" +#include "simap.h" #include "stream.h" #include "timeval.h" #include "vconn.h" @@ -41,14 +44,34 @@ VLOG_DEFINE_THIS_MODULE(connmgr); static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5); -/* An OpenFlow connection. */ +/* An OpenFlow connection. + * + * + * Thread-safety + * ============= + * + * 'ofproto_mutex' must be held whenever an ofconn is created or destroyed or, + * more or less equivalently, whenever an ofconn is added to or removed from a + * connmgr. 'ofproto_mutex' doesn't protect the data inside the ofconn, except + * as specifically noted below. */ struct ofconn { - struct connmgr *connmgr; /* Connection's manager. */ +/* Configuration that persists from one connection to the next. */ + struct list node; /* In struct connmgr's "all_conns" list. */ + struct hmap_node hmap_node; /* In struct connmgr's "controllers" map. */ + + struct connmgr *connmgr; /* Connection's manager. */ struct rconn *rconn; /* OpenFlow connection. */ enum ofconn_type type; /* Type. */ - enum nx_flow_format flow_format; /* Currently selected flow format. */ - bool flow_mod_table_id; /* NXT_FLOW_MOD_TABLE_ID enabled? */ + enum ofproto_band band; /* In-band or out-of-band? */ + bool enable_async_msgs; /* Initially enable async messages? */ + +/* State that should be cleared from one connection to the next. */ + + /* OpenFlow state. */ + enum ofp12_controller_role role; /* Role. */ + enum ofputil_protocol protocol; /* Current protocol variant. */ + enum nx_packet_in_format packet_in_format; /* OFPT_PACKET_IN format. */ /* Asynchronous flow table operation support. */ struct list opgroups; /* Contains pending "ofopgroups", if any. */ @@ -61,6 +84,7 @@ struct ofconn { struct pinsched *schedulers[N_SCHEDULERS]; struct pktbuf *pktbuf; /* OpenFlow packet buffers. */ int miss_send_len; /* Bytes to send of buffered packets. */ + uint16_t controller_id; /* Connection controller ID. */ /* Number of OpenFlow messages queued on 'rconn' as replies to OpenFlow * requests, and the maximum number before we stop reading OpenFlow @@ -68,22 +92,57 @@ struct ofconn { #define OFCONN_REPLY_MAX 100 struct rconn_packet_counter *reply_counter; - /* type == OFCONN_PRIMARY only. */ - enum nx_role role; /* Role. */ - struct hmap_node hmap_node; /* In struct connmgr's "controllers" map. */ - enum ofproto_band band; /* In-band or out-of-band? */ + /* Asynchronous message configuration in each possible roles. + * + * A 1-bit enables sending an asynchronous message for one possible reason + * that the message might be generated, a 0-bit disables it. */ + uint32_t master_async_config[OAM_N_TYPES]; /* master, other */ + uint32_t slave_async_config[OAM_N_TYPES]; /* slave */ + +/* Flow monitors (e.g. NXST_FLOW_MONITOR). */ + + /* Configuration. Contains "struct ofmonitor"s. */ + struct hmap monitors OVS_GUARDED_BY(ofproto_mutex); + + /* Flow control. + * + * When too many flow monitor notifications back up in the transmit buffer, + * we pause the transmission of further notifications. These members track + * the flow control state. + * + * When notifications are flowing, 'monitor_paused' is 0. When + * notifications are paused, 'monitor_paused' is the value of + * 'monitor_seqno' at the point we paused. + * + * 'monitor_counter' counts the OpenFlow messages and bytes currently in + * flight. This value growing too large triggers pausing. */ + uint64_t monitor_paused OVS_GUARDED_BY(ofproto_mutex); + struct rconn_packet_counter *monitor_counter OVS_GUARDED_BY(ofproto_mutex); + + /* State of monitors for a single ongoing flow_mod. + * + * 'updates' is a list of "struct ofpbuf"s that contain + * NXST_FLOW_MONITOR_REPLY messages representing the changes made by the + * current flow_mod. + * + * When 'updates' is nonempty, 'sent_abbrev_update' is true if 'updates' + * contains an update event of type NXFME_ABBREV and false otherwise.. */ + struct list updates OVS_GUARDED_BY(ofproto_mutex); + bool sent_abbrev_update OVS_GUARDED_BY(ofproto_mutex); }; static struct ofconn *ofconn_create(struct connmgr *, struct rconn *, - enum ofconn_type); -static void ofconn_destroy(struct ofconn *); + enum ofconn_type, bool enable_async_msgs) + OVS_REQUIRES(ofproto_mutex); +static void ofconn_destroy(struct ofconn *) OVS_REQUIRES(ofproto_mutex); +static void ofconn_flush(struct ofconn *) OVS_REQUIRES(ofproto_mutex); static void ofconn_reconfigure(struct ofconn *, const struct ofproto_controller *); static void ofconn_run(struct ofconn *, bool (*handle_openflow)(struct ofconn *, - struct ofpbuf *ofp_msg)); + const struct ofpbuf *ofp_msg)); static void ofconn_wait(struct ofconn *, bool handling_openflow); static const char *ofconn_get_target(const struct ofconn *); @@ -91,12 +150,10 @@ static char *ofconn_make_name(const struct connmgr *, const char *target); static void ofconn_set_rate_limit(struct ofconn *, int rate, int burst); -static bool ofconn_receives_async_msgs(const struct ofconn *); - static void ofconn_send(const struct ofconn *, struct ofpbuf *, struct rconn_packet_counter *); -static void do_send_packet_in(struct ofpbuf *, void *ofconn_); +static void do_send_packet_ins(struct ofconn *, struct list *txq); /* A listener for incoming OpenFlow "service" connections. */ struct ofservice { @@ -108,11 +165,16 @@ struct ofservice { int probe_interval; /* Max idle time before probing, in seconds. */ int rate_limit; /* Max packet-in rate in packets per second. */ int burst_limit; /* Limit on accumulating packet credits. */ + bool enable_async_msgs; /* Initially enable async messages? */ + uint8_t dscp; /* DSCP Value for controller connection */ + uint32_t allowed_versions; /* OpenFlow protocol versions that may + * be negotiated for a session. */ }; static void ofservice_reconfigure(struct ofservice *, const struct ofproto_controller *); -static int ofservice_create(struct connmgr *, const char *target); +static int ofservice_create(struct connmgr *mgr, const char *target, + uint32_t allowed_versions, uint8_t dscp); static void ofservice_destroy(struct connmgr *, struct ofservice *); static struct ofservice *ofservice_lookup(struct connmgr *, const char *target); @@ -126,6 +188,9 @@ struct connmgr { /* OpenFlow connections. */ struct hmap controllers; /* Controller "struct ofconn"s. */ struct list all_conns; /* Contains "struct ofconn"s. */ + uint64_t master_election_id; /* monotonically increasing sequence number + * for master election */ + bool master_election_id_defined; /* OpenFlow listeners. */ struct hmap services; /* Contains "struct ofservice"s. */ @@ -145,6 +210,8 @@ struct connmgr { static void update_in_band_remotes(struct connmgr *); static void add_snooper(struct connmgr *, struct vconn *); +static void ofmonitor_run(struct connmgr *); +static void ofmonitor_wait(struct connmgr *); /* Creates and returns a new connection manager owned by 'ofproto'. 'name' is * a name for the ofproto suitable for using in log messages. @@ -163,6 +230,8 @@ connmgr_create(struct ofproto *ofproto, hmap_init(&mgr->controllers); list_init(&mgr->all_conns); + mgr->master_election_id = 0; + mgr->master_election_id_defined = false; hmap_init(&mgr->services); mgr->snoops = NULL; @@ -191,9 +260,12 @@ connmgr_destroy(struct connmgr *mgr) return; } + ovs_mutex_lock(&ofproto_mutex); LIST_FOR_EACH_SAFE (ofconn, next_ofconn, node, &mgr->all_conns) { ofconn_destroy(ofconn); } + ovs_mutex_unlock(&ofproto_mutex); + hmap_destroy(&mgr->controllers); HMAP_FOR_EACH_SAFE (ofservice, next_ofservice, node, &mgr->services) { @@ -234,7 +306,9 @@ connmgr_destroy(struct connmgr *mgr) * fail-open processing) are suppressed too. */ void connmgr_run(struct connmgr *mgr, - bool (*handle_openflow)(struct ofconn *, struct ofpbuf *ofp_msg)) + bool (*handle_openflow)(struct ofconn *, + const struct ofpbuf *ofp_msg)) + OVS_EXCLUDED(ofproto_mutex) { struct ofconn *ofconn, *next_ofconn; struct ofservice *ofservice; @@ -250,6 +324,7 @@ connmgr_run(struct connmgr *mgr, LIST_FOR_EACH_SAFE (ofconn, next_ofconn, node, &mgr->all_conns) { ofconn_run(ofconn, handle_openflow); } + ofmonitor_run(mgr); /* Fail-open maintenance. Do this after processing the ofconns since * fail-open checks the status of the controller rconn. */ @@ -261,21 +336,27 @@ connmgr_run(struct connmgr *mgr, struct vconn *vconn; int retval; - retval = pvconn_accept(ofservice->pvconn, OFP_VERSION, &vconn); + retval = pvconn_accept(ofservice->pvconn, &vconn); if (!retval) { struct rconn *rconn; char *name; - rconn = rconn_create(ofservice->probe_interval, 0); + /* Passing default value for creation of the rconn */ + rconn = rconn_create(ofservice->probe_interval, 0, ofservice->dscp, + vconn_get_allowed_versions(vconn)); name = ofconn_make_name(mgr, vconn_get_name(vconn)); rconn_connect_unreliably(rconn, vconn, name); free(name); - ofconn = ofconn_create(mgr, rconn, OFCONN_SERVICE); + ovs_mutex_lock(&ofproto_mutex); + ofconn = ofconn_create(mgr, rconn, OFCONN_SERVICE, + ofservice->enable_async_msgs); + ovs_mutex_unlock(&ofproto_mutex); + ofconn_set_rate_limit(ofconn, ofservice->rate_limit, ofservice->burst_limit); } else if (retval != EAGAIN) { - VLOG_WARN_RL(&rl, "accept failed (%s)", strerror(retval)); + VLOG_WARN_RL(&rl, "accept failed (%s)", ovs_strerror(retval)); } } @@ -283,11 +364,11 @@ connmgr_run(struct connmgr *mgr, struct vconn *vconn; int retval; - retval = pvconn_accept(mgr->snoops[i], OFP_VERSION, &vconn); + retval = pvconn_accept(mgr->snoops[i], &vconn); if (!retval) { add_snooper(mgr, vconn); } else if (retval != EAGAIN) { - VLOG_WARN_RL(&rl, "accept failed (%s)", strerror(retval)); + VLOG_WARN_RL(&rl, "accept failed (%s)", ovs_strerror(retval)); } } } @@ -307,6 +388,7 @@ connmgr_wait(struct connmgr *mgr, bool handling_openflow) LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { ofconn_wait(ofconn, handling_openflow); } + ofmonitor_wait(mgr); if (handling_openflow && mgr->in_band) { in_band_wait(mgr->in_band); } @@ -321,6 +403,30 @@ connmgr_wait(struct connmgr *mgr, bool handling_openflow) } } +/* Adds some memory usage statistics for 'mgr' into 'usage', for use with + * memory_report(). */ +void +connmgr_get_memory_usage(const struct connmgr *mgr, struct simap *usage) +{ + const struct ofconn *ofconn; + unsigned int packets = 0; + unsigned int ofconns = 0; + + LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { + int i; + + ofconns++; + + packets += rconn_count_txqlen(ofconn->rconn); + for (i = 0; i < N_SCHEDULERS; i++) { + packets += pinsched_count_txqlen(ofconn->schedulers[i]); + } + packets += pktbuf_count_packets(ofconn->pktbuf); + } + simap_increase(usage, "ofconns", ofconns); + simap_increase(usage, "packets", packets); +} + /* Returns the ofproto that owns 'ofconn''s connmgr. */ struct ofproto * ofconn_get_ofproto(const struct ofconn *ofconn) @@ -343,10 +449,12 @@ connmgr_retry(struct connmgr *mgr) /* OpenFlow configuration. */ -static void add_controller(struct connmgr *, const char *target); +static void add_controller(struct connmgr *, const char *target, uint8_t dscp, + uint32_t allowed_versions) + OVS_REQUIRES(ofproto_mutex); static struct ofconn *find_controller_by_target(struct connmgr *, const char *target); -static void update_fail_open(struct connmgr *); +static void update_fail_open(struct connmgr *) OVS_EXCLUDED(ofproto_mutex); static int set_pvconns(struct pvconn ***pvconnsp, size_t *n_pvconnsp, const struct sset *); @@ -422,7 +530,7 @@ connmgr_free_controller_info(struct shash *info) SHASH_FOR_EACH (node, info) { struct ofproto_controller_info *cinfo = node->data; while (cinfo->pairs.n) { - free((char *) cinfo->pairs.values[--cinfo->pairs.n]); + free(CONST_CAST(char *, cinfo->pairs.values[--cinfo->pairs.n])); } free(cinfo); } @@ -434,7 +542,8 @@ connmgr_free_controller_info(struct shash *info) void connmgr_set_controllers(struct connmgr *mgr, const struct ofproto_controller *controllers, - size_t n_controllers) + size_t n_controllers, uint32_t allowed_versions) + OVS_EXCLUDED(ofproto_mutex) { bool had_controllers = connmgr_has_controllers(mgr); struct shash new_controllers; @@ -442,6 +551,10 @@ connmgr_set_controllers(struct connmgr *mgr, struct ofservice *ofservice, *next_ofservice; size_t i; + /* Required to add and remove ofconns. This could probably be narrowed to + * cover a smaller amount of code, if that yielded some benefit. */ + ovs_mutex_lock(&ofproto_mutex); + /* Create newly configured controllers and services. * Create a name to ofproto_controller mapping in 'new_controllers'. */ shash_init(&new_controllers); @@ -449,12 +562,37 @@ connmgr_set_controllers(struct connmgr *mgr, const struct ofproto_controller *c = &controllers[i]; if (!vconn_verify_name(c->target)) { - if (!find_controller_by_target(mgr, c->target)) { - add_controller(mgr, c->target); + bool add = false; + ofconn = find_controller_by_target(mgr, c->target); + if (!ofconn) { + VLOG_INFO("%s: added primary controller \"%s\"", + mgr->name, c->target); + add = true; + } else if (rconn_get_allowed_versions(ofconn->rconn) != + allowed_versions) { + VLOG_INFO("%s: re-added primary controller \"%s\"", + mgr->name, c->target); + add = true; + ofconn_destroy(ofconn); + } + if (add) { + add_controller(mgr, c->target, c->dscp, allowed_versions); } } else if (!pvconn_verify_name(c->target)) { - if (!ofservice_lookup(mgr, c->target)) { - ofservice_create(mgr, c->target); + bool add = false; + ofservice = ofservice_lookup(mgr, c->target); + if (!ofservice) { + VLOG_INFO("%s: added service controller \"%s\"", + mgr->name, c->target); + add = true; + } else if (ofservice->allowed_versions != allowed_versions) { + VLOG_INFO("%s: re-added service controller \"%s\"", + mgr->name, c->target); + ofservice_destroy(mgr, ofservice); + add = true; + } + if (add) { + ofservice_create(mgr, c->target, allowed_versions, c->dscp); } } else { VLOG_WARN_RL(&rl, "%s: unsupported controller \"%s\"", @@ -468,10 +606,13 @@ connmgr_set_controllers(struct connmgr *mgr, /* Delete controllers that are no longer configured. * Update configuration of all now-existing controllers. */ HMAP_FOR_EACH_SAFE (ofconn, next_ofconn, hmap_node, &mgr->controllers) { + const char *target = ofconn_get_target(ofconn); struct ofproto_controller *c; - c = shash_find_data(&new_controllers, ofconn_get_target(ofconn)); + c = shash_find_data(&new_controllers, target); if (!c) { + VLOG_INFO("%s: removed primary controller \"%s\"", + mgr->name, target); ofconn_destroy(ofconn); } else { ofconn_reconfigure(ofconn, c); @@ -481,11 +622,13 @@ connmgr_set_controllers(struct connmgr *mgr, /* Delete services that are no longer configured. * Update configuration of all now-existing services. */ HMAP_FOR_EACH_SAFE (ofservice, next_ofservice, node, &mgr->services) { + const char *target = pvconn_get_name(ofservice->pvconn); struct ofproto_controller *c; - c = shash_find_data(&new_controllers, - pvconn_get_name(ofservice->pvconn)); + c = shash_find_data(&new_controllers, target); if (!c) { + VLOG_INFO("%s: removed service controller \"%s\"", + mgr->name, target); ofservice_destroy(mgr, ofservice); } else { ofservice_reconfigure(ofservice, c); @@ -494,6 +637,8 @@ connmgr_set_controllers(struct connmgr *mgr, shash_destroy(&new_controllers); + ovs_mutex_unlock(&ofproto_mutex); + update_in_band_remotes(mgr); update_fail_open(mgr); if (had_controllers != connmgr_has_controllers(mgr)) { @@ -544,14 +689,16 @@ connmgr_has_snoops(const struct connmgr *mgr) /* Creates a new controller for 'target' in 'mgr'. update_controller() needs * to be called later to finish the new ofconn's configuration. */ static void -add_controller(struct connmgr *mgr, const char *target) +add_controller(struct connmgr *mgr, const char *target, uint8_t dscp, + uint32_t allowed_versions) + OVS_REQUIRES(ofproto_mutex) { char *name = ofconn_make_name(mgr, target); struct ofconn *ofconn; - ofconn = ofconn_create(mgr, rconn_create(5, 8), OFCONN_PRIMARY); + ofconn = ofconn_create(mgr, rconn_create(5, 8, dscp, allowed_versions), + OFCONN_PRIMARY, true); ofconn->pktbuf = pktbuf_create(); - ofconn->miss_send_len = OFP_DEFAULT_MISS_SEND_LEN; rconn_connect(ofconn->rconn, target, name); hmap_insert(&mgr->controllers, &ofconn->hmap_node, hash_string(target, 0)); @@ -587,18 +734,13 @@ update_in_band_remotes(struct connmgr *mgr) /* Add all the remotes. */ HMAP_FOR_EACH (ofconn, hmap_node, &mgr->controllers) { - struct sockaddr_in *sin = &addrs[n_addrs]; const char *target = rconn_get_target(ofconn->rconn); + struct sockaddr_storage ss; - if (ofconn->band == OFPROTO_OUT_OF_BAND) { - continue; - } - - if (stream_parse_target_with_default_ports(target, - OFP_TCP_PORT, - OFP_SSL_PORT, - sin)) { - n_addrs++; + if (ofconn->band == OFPROTO_IN_BAND + && stream_parse_target_with_default_port(target, OFP_OLD_PORT, &ss) + && ss.ss_family == AF_INET) { + addrs[n_addrs++] = *(struct sockaddr_in *) &ss; } } for (i = 0; i < mgr->n_extra_remotes; i++) { @@ -625,6 +767,7 @@ update_in_band_remotes(struct connmgr *mgr) static void update_fail_open(struct connmgr *mgr) + OVS_EXCLUDED(ofproto_mutex) { if (connmgr_has_controllers(mgr) && mgr->fail_mode == OFPROTO_FAIL_STANDALONE) { @@ -657,12 +800,11 @@ set_pvconns(struct pvconn ***pvconnsp, size_t *n_pvconnsp, SSET_FOR_EACH (name, sset) { struct pvconn *pvconn; int error; - - error = pvconn_open(name, &pvconn); + error = pvconn_open(name, 0, 0, &pvconn); if (!error) { pvconns[n_pvconns++] = pvconn; } else { - VLOG_ERR("failed to listen on %s: %s", name, strerror(error)); + VLOG_ERR("failed to listen on %s: %s", name, ovs_strerror(error)); if (!retval) { retval = error; } @@ -682,12 +824,13 @@ static int snoop_preference(const struct ofconn *ofconn) { switch (ofconn->role) { - case NX_ROLE_MASTER: + case OFPCR12_ROLE_MASTER: return 3; - case NX_ROLE_OTHER: + case OFPCR12_ROLE_EQUAL: return 2; - case NX_ROLE_SLAVE: + case OFPCR12_ROLE_SLAVE: return 1; + case OFPCR12_ROLE_NOCHANGE: default: /* Shouldn't happen. */ return 0; @@ -727,64 +870,157 @@ ofconn_get_type(const struct ofconn *ofconn) return ofconn->type; } +/* If a master election id is defined, stores it into '*idp' and returns + * true. Otherwise, stores UINT64_MAX into '*idp' and returns false. */ +bool +ofconn_get_master_election_id(const struct ofconn *ofconn, uint64_t *idp) +{ + *idp = (ofconn->connmgr->master_election_id_defined + ? ofconn->connmgr->master_election_id + : UINT64_MAX); + return ofconn->connmgr->master_election_id_defined; +} + +/* Sets the master election id. + * + * Returns true if successful, false if the id is stale + */ +bool +ofconn_set_master_election_id(struct ofconn *ofconn, uint64_t id) +{ + if (ofconn->connmgr->master_election_id_defined + && + /* Unsigned difference interpreted as a two's complement signed + * value */ + (int64_t)(id - ofconn->connmgr->master_election_id) < 0) { + return false; + } + ofconn->connmgr->master_election_id = id; + ofconn->connmgr->master_election_id_defined = true; + + return true; +} + /* Returns the role configured for 'ofconn'. * - * The default role, if no other role has been set, is NX_ROLE_OTHER. */ -enum nx_role + * The default role, if no other role has been set, is OFPCR12_ROLE_EQUAL. */ +enum ofp12_controller_role ofconn_get_role(const struct ofconn *ofconn) { return ofconn->role; } -/* Changes 'ofconn''s role to 'role'. If 'role' is NX_ROLE_MASTER then any - * existing master is demoted to a slave. */ void -ofconn_set_role(struct ofconn *ofconn, enum nx_role role) +ofconn_send_role_status(struct ofconn *ofconn, uint32_t role, uint8_t reason) +{ + struct ofputil_role_status status; + struct ofpbuf *buf; + + status.reason = reason; + status.role = role; + ofconn_get_master_election_id(ofconn, &status.generation_id); + + buf = ofputil_encode_role_status(&status, ofconn_get_protocol(ofconn)); + + ofconn_send(ofconn, buf, NULL); +} + +/* Changes 'ofconn''s role to 'role'. If 'role' is OFPCR12_ROLE_MASTER then + * any existing master is demoted to a slave. */ +void +ofconn_set_role(struct ofconn *ofconn, enum ofp12_controller_role role) { - if (role == NX_ROLE_MASTER) { + if (role != ofconn->role && role == OFPCR12_ROLE_MASTER) { struct ofconn *other; HMAP_FOR_EACH (other, hmap_node, &ofconn->connmgr->controllers) { - if (other->role == NX_ROLE_MASTER) { - other->role = NX_ROLE_SLAVE; + if (other->role == OFPCR12_ROLE_MASTER) { + other->role = OFPCR12_ROLE_SLAVE; + ofconn_send_role_status(other, OFPCR12_ROLE_SLAVE, OFPCRR_MASTER_REQUEST); } } } ofconn->role = role; } -/* Returns the currently configured flow format for 'ofconn', one of NXFF_*. - * - * The default, if no other format has been set, is NXFF_OPENFLOW10. */ -enum nx_flow_format -ofconn_get_flow_format(struct ofconn *ofconn) +void +ofconn_set_invalid_ttl_to_controller(struct ofconn *ofconn, bool enable) +{ + uint32_t bit = 1u << OFPR_INVALID_TTL; + if (enable) { + ofconn->master_async_config[OAM_PACKET_IN] |= bit; + } else { + ofconn->master_async_config[OAM_PACKET_IN] &= ~bit; + } +} + +bool +ofconn_get_invalid_ttl_to_controller(struct ofconn *ofconn) { - return ofconn->flow_format; + uint32_t bit = 1u << OFPR_INVALID_TTL; + return (ofconn->master_async_config[OAM_PACKET_IN] & bit) != 0; +} + +/* Returns the currently configured protocol for 'ofconn', one of OFPUTIL_P_*. + * + * Returns OFPUTIL_P_NONE, which is not a valid protocol, if 'ofconn' hasn't + * completed version negotiation. This can't happen if at least one OpenFlow + * message, other than OFPT_HELLO, has been received on the connection (such as + * in ofproto.c's message handling code), since version negotiation is a + * prerequisite for starting to receive messages. This means that + * OFPUTIL_P_NONE is a special case that most callers need not worry about. */ +enum ofputil_protocol +ofconn_get_protocol(const struct ofconn *ofconn) +{ + if (ofconn->protocol == OFPUTIL_P_NONE && + rconn_is_connected(ofconn->rconn)) { + int version = rconn_get_version(ofconn->rconn); + if (version > 0) { + ofconn_set_protocol(CONST_CAST(struct ofconn *, ofconn), + ofputil_protocol_from_ofp_version(version)); + } + } + + return ofconn->protocol; } -/* Sets the flow format for 'ofconn' to 'flow_format' (one of NXFF_*). */ +/* Sets the protocol for 'ofconn' to 'protocol' (one of OFPUTIL_P_*). + * + * (This doesn't actually send anything to accomplish this. Presumably the + * caller already did that.) */ void -ofconn_set_flow_format(struct ofconn *ofconn, enum nx_flow_format flow_format) +ofconn_set_protocol(struct ofconn *ofconn, enum ofputil_protocol protocol) { - ofconn->flow_format = flow_format; + ofconn->protocol = protocol; } -/* Returns true if the NXT_FLOW_MOD_TABLE_ID extension is enabled, false - * otherwise. +/* Returns the currently configured packet in format for 'ofconn', one of + * NXPIF_*. * - * By default the extension is not enabled. */ -bool -ofconn_get_flow_mod_table_id(const struct ofconn *ofconn) + * The default, if no other format has been set, is NXPIF_OPENFLOW10. */ +enum nx_packet_in_format +ofconn_get_packet_in_format(struct ofconn *ofconn) { - return ofconn->flow_mod_table_id; + return ofconn->packet_in_format; } -/* Enables or disables (according to 'enable') the NXT_FLOW_MOD_TABLE_ID - * extension on 'ofconn'. */ +/* Sets the packet in format for 'ofconn' to 'packet_in_format' (one of + * NXPIF_*). */ void -ofconn_set_flow_mod_table_id(struct ofconn *ofconn, bool enable) +ofconn_set_packet_in_format(struct ofconn *ofconn, + enum nx_packet_in_format packet_in_format) { - ofconn->flow_mod_table_id = enable; + ofconn->packet_in_format = packet_in_format; +} + +/* Sets the controller connection ID for 'ofconn' to 'controller_id'. + * + * The connection controller ID is used for OFPP_CONTROLLER and + * NXAST_CONTROLLER actions. See "struct nx_action_controller" for details. */ +void +ofconn_set_controller_id(struct ofconn *ofconn, uint16_t controller_id) +{ + ofconn->controller_id = controller_id; } /* Returns the default miss send length for 'ofconn'. */ @@ -801,6 +1037,25 @@ ofconn_set_miss_send_len(struct ofconn *ofconn, int miss_send_len) ofconn->miss_send_len = miss_send_len; } +void +ofconn_set_async_config(struct ofconn *ofconn, + const uint32_t master_masks[OAM_N_TYPES], + const uint32_t slave_masks[OAM_N_TYPES]) +{ + size_t size = sizeof ofconn->master_async_config; + memcpy(ofconn->master_async_config, master_masks, size); + memcpy(ofconn->slave_async_config, slave_masks, size); +} + +void +ofconn_get_async_config(struct ofconn *ofconn, + uint32_t *master_masks, uint32_t *slave_masks) +{ + size_t size = sizeof ofconn->master_async_config; + memcpy(master_masks, ofconn->master_async_config, size); + memcpy(slave_masks, ofconn->slave_async_config, size); +} + /* Sends 'msg' on 'ofconn', accounting it as a reply. (If there is a * sufficient number of OpenFlow replies in-flight on a single ofconn, then the * connmgr will stop accepting new OpenFlow requests on that ofconn until the @@ -824,45 +1079,38 @@ ofconn_send_replies(const struct ofconn *ofconn, struct list *replies) } } -/* Sends 'error', which should be an OpenFlow error created with - * e.g. ofp_mkerr(), on 'ofconn', as a reply to 'request'. Only at most the +/* Sends 'error' on 'ofconn', as a reply to 'request'. Only at most the * first 64 bytes of 'request' are used. */ void ofconn_send_error(const struct ofconn *ofconn, - const struct ofp_header *request, int error) -{ - struct ofpbuf *msg; - - msg = ofputil_encode_error_msg(error, request); - if (msg) { - static struct vlog_rate_limit err_rl = VLOG_RATE_LIMIT_INIT(10, 10); - - if (!VLOG_DROP_INFO(&err_rl)) { - const struct ofputil_msg_type *type; - const char *type_name; - size_t request_len; - char *error_s; - - request_len = ntohs(request->length); - type_name = (!ofputil_decode_msg_type_partial(request, - MIN(64, request_len), - &type) - ? ofputil_msg_type_name(type) - : "invalid"); - - error_s = ofputil_error_to_string(error); - VLOG_INFO("%s: sending %s error reply to %s message", - rconn_get_name(ofconn->rconn), error_s, type_name); - free(error_s); - } - ofconn_send_reply(ofconn, msg); + const struct ofp_header *request, enum ofperr error) +{ + static struct vlog_rate_limit err_rl = VLOG_RATE_LIMIT_INIT(10, 10); + struct ofpbuf *reply; + + reply = ofperr_encode_reply(error, request); + if (!VLOG_DROP_INFO(&err_rl)) { + const char *type_name; + size_t request_len; + enum ofpraw raw; + + request_len = ntohs(request->length); + type_name = (!ofpraw_decode_partial(&raw, request, + MIN(64, request_len)) + ? ofpraw_get_name(raw) + : "invalid"); + + VLOG_INFO("%s: sending %s error reply to %s message", + rconn_get_name(ofconn->rconn), ofperr_to_string(error), + type_name); } + ofconn_send_reply(ofconn, reply); } /* Same as pktbuf_retrieve(), using the pktbuf owned by 'ofconn'. */ -int +enum ofperr ofconn_pktbuf_retrieve(struct ofconn *ofconn, uint32_t id, - struct ofpbuf **bufferp, uint16_t *in_port) + struct ofpbuf **bufferp, ofp_port_t *in_port) { return pktbuf_retrieve(ofconn->pktbuf, id, bufferp, in_port); } @@ -898,42 +1146,116 @@ ofconn_get_target(const struct ofconn *ofconn) } static struct ofconn * -ofconn_create(struct connmgr *mgr, struct rconn *rconn, enum ofconn_type type) +ofconn_create(struct connmgr *mgr, struct rconn *rconn, enum ofconn_type type, + bool enable_async_msgs) { - struct ofconn *ofconn = xzalloc(sizeof *ofconn); + struct ofconn *ofconn; + + ofconn = xzalloc(sizeof *ofconn); ofconn->connmgr = mgr; list_push_back(&mgr->all_conns, &ofconn->node); ofconn->rconn = rconn; ofconn->type = type; - ofconn->flow_format = NXFF_OPENFLOW10; - ofconn->flow_mod_table_id = false; + ofconn->enable_async_msgs = enable_async_msgs; + list_init(&ofconn->opgroups); - ofconn->role = NX_ROLE_OTHER; - ofconn->packet_in_counter = rconn_packet_counter_create (); - ofconn->pktbuf = NULL; - ofconn->miss_send_len = 0; - ofconn->reply_counter = rconn_packet_counter_create (); + + hmap_init(&ofconn->monitors); + list_init(&ofconn->updates); + + ofconn_flush(ofconn); + return ofconn; } -/* Disassociates 'ofconn' from all of the ofopgroups that it initiated that - * have not yet completed. (Those ofopgroups will still run to completion in - * the usual way, but any errors that they run into will not be reported on any - * OpenFlow channel.) - * - * Also discards any blocked operation on 'ofconn'. */ +/* Clears all of the state in 'ofconn' that should not persist from one + * connection to the next. */ static void ofconn_flush(struct ofconn *ofconn) + OVS_REQUIRES(ofproto_mutex) { + struct ofmonitor *monitor, *next_monitor; + int i; + + ofconn->role = OFPCR12_ROLE_EQUAL; + ofconn_set_protocol(ofconn, OFPUTIL_P_NONE); + ofconn->packet_in_format = NXPIF_OPENFLOW10; + + /* Disassociate 'ofconn' from all of the ofopgroups that it initiated that + * have not yet completed. (Those ofopgroups will still run to completion + * in the usual way, but any errors that they run into will not be reported + * on any OpenFlow channel.) + * + * Also discard any blocked operation on 'ofconn'. */ while (!list_is_empty(&ofconn->opgroups)) { list_init(list_pop_front(&ofconn->opgroups)); } ofpbuf_delete(ofconn->blocked); ofconn->blocked = NULL; + + rconn_packet_counter_destroy(ofconn->packet_in_counter); + ofconn->packet_in_counter = rconn_packet_counter_create(); + for (i = 0; i < N_SCHEDULERS; i++) { + if (ofconn->schedulers[i]) { + int rate, burst; + + pinsched_get_limits(ofconn->schedulers[i], &rate, &burst); + pinsched_destroy(ofconn->schedulers[i]); + ofconn->schedulers[i] = pinsched_create(rate, burst); + } + } + if (ofconn->pktbuf) { + pktbuf_destroy(ofconn->pktbuf); + ofconn->pktbuf = pktbuf_create(); + } + ofconn->miss_send_len = (ofconn->type == OFCONN_PRIMARY + ? OFP_DEFAULT_MISS_SEND_LEN + : 0); + ofconn->controller_id = 0; + + rconn_packet_counter_destroy(ofconn->reply_counter); + ofconn->reply_counter = rconn_packet_counter_create(); + + if (ofconn->enable_async_msgs) { + uint32_t *master = ofconn->master_async_config; + uint32_t *slave = ofconn->slave_async_config; + + /* "master" and "other" roles get all asynchronous messages by default, + * except that the controller needs to enable nonstandard "packet-in" + * reasons itself. */ + master[OAM_PACKET_IN] = (1u << OFPR_NO_MATCH) | (1u << OFPR_ACTION); + master[OAM_PORT_STATUS] = ((1u << OFPPR_ADD) + | (1u << OFPPR_DELETE) + | (1u << OFPPR_MODIFY)); + master[OAM_FLOW_REMOVED] = ((1u << OFPRR_IDLE_TIMEOUT) + | (1u << OFPRR_HARD_TIMEOUT) + | (1u << OFPRR_DELETE)); + + /* "slave" role gets port status updates by default. */ + slave[OAM_PACKET_IN] = 0; + slave[OAM_PORT_STATUS] = ((1u << OFPPR_ADD) + | (1u << OFPPR_DELETE) + | (1u << OFPPR_MODIFY)); + slave[OAM_FLOW_REMOVED] = 0; + } else { + memset(ofconn->master_async_config, 0, + sizeof ofconn->master_async_config); + memset(ofconn->slave_async_config, 0, + sizeof ofconn->slave_async_config); + } + + HMAP_FOR_EACH_SAFE (monitor, next_monitor, ofconn_node, + &ofconn->monitors) { + ofmonitor_destroy(monitor); + } + rconn_packet_counter_destroy(ofconn->monitor_counter); + ofconn->monitor_counter = rconn_packet_counter_create(); + ofpbuf_list_delete(&ofconn->updates); /* ...but it should be empty. */ } static void ofconn_destroy(struct ofconn *ofconn) + OVS_REQUIRES(ofproto_mutex) { ofconn_flush(ofconn); @@ -941,11 +1263,13 @@ ofconn_destroy(struct ofconn *ofconn) hmap_remove(&ofconn->connmgr->controllers, &ofconn->hmap_node); } + hmap_destroy(&ofconn->monitors); list_remove(&ofconn->node); rconn_destroy(ofconn->rconn); rconn_packet_counter_destroy(ofconn->packet_in_counter); rconn_packet_counter_destroy(ofconn->reply_counter); pktbuf_destroy(ofconn->pktbuf); + rconn_packet_counter_destroy(ofconn->monitor_counter); free(ofconn); } @@ -957,6 +1281,7 @@ ofconn_reconfigure(struct ofconn *ofconn, const struct ofproto_controller *c) int probe_interval; ofconn->band = c->band; + ofconn->enable_async_msgs = c->enable_async_msgs; rconn_set_max_backoff(ofconn->rconn, c->max_backoff); @@ -964,6 +1289,12 @@ ofconn_reconfigure(struct ofconn *ofconn, const struct ofproto_controller *c) rconn_set_probe_interval(ofconn->rconn, probe_interval); ofconn_set_rate_limit(ofconn, c->rate_limit, c->burst_limit); + + /* If dscp value changed reconnect. */ + if (c->dscp != rconn_get_dscp(ofconn->rconn)) { + rconn_set_dscp(ofconn->rconn, c->dscp); + rconn_reconnect(ofconn->rconn); + } } /* Returns true if it makes sense for 'ofconn' to receive and process OpenFlow @@ -971,19 +1302,23 @@ ofconn_reconfigure(struct ofconn *ofconn, const struct ofproto_controller *c) static bool ofconn_may_recv(const struct ofconn *ofconn) { - int count = rconn_packet_counter_read (ofconn->reply_counter); + int count = rconn_packet_counter_n_packets(ofconn->reply_counter); return (!ofconn->blocked || ofconn->retry) && count < OFCONN_REPLY_MAX; } static void ofconn_run(struct ofconn *ofconn, - bool (*handle_openflow)(struct ofconn *, struct ofpbuf *ofp_msg)) + bool (*handle_openflow)(struct ofconn *, + const struct ofpbuf *ofp_msg)) { struct connmgr *mgr = ofconn->connmgr; size_t i; for (i = 0; i < N_SCHEDULERS; i++) { - pinsched_run(ofconn->schedulers[i], do_send_packet_in, ofconn); + struct list txq; + + pinsched_run(ofconn->schedulers[i], &txq); + do_send_packet_ins(ofconn, &txq); } rconn_run(ofconn->rconn); @@ -1013,11 +1348,13 @@ ofconn_run(struct ofconn *ofconn, } } + ovs_mutex_lock(&ofproto_mutex); if (!rconn_is_alive(ofconn->rconn)) { ofconn_destroy(ofconn); } else if (!rconn_is_connected(ofconn->rconn)) { ofconn_flush(ofconn); } + ovs_mutex_unlock(&ofproto_mutex); } static void @@ -1034,21 +1371,102 @@ ofconn_wait(struct ofconn *ofconn, bool handling_openflow) } } -/* Returns true if 'ofconn' should receive asynchronous messages. */ +/* Returns true if 'ofconn' should receive asynchronous messages of the given + * OAM_* 'type' and 'reason', which should be a OFPR_* value for OAM_PACKET_IN, + * a OFPPR_* value for OAM_PORT_STATUS, or an OFPRR_* value for + * OAM_FLOW_REMOVED. Returns false if the message should not be sent on + * 'ofconn'. */ static bool -ofconn_receives_async_msgs(const struct ofconn *ofconn) +ofconn_receives_async_msg(const struct ofconn *ofconn, + enum ofconn_async_msg_type type, + unsigned int reason) { - if (!rconn_is_connected(ofconn->rconn)) { + const uint32_t *async_config; + + ovs_assert(reason < 32); + ovs_assert((unsigned int) type < OAM_N_TYPES); + + if (ofconn_get_protocol(ofconn) == OFPUTIL_P_NONE + || !rconn_is_connected(ofconn->rconn)) { return false; - } else if (ofconn->type == OFCONN_PRIMARY) { - /* Primary controllers always get asynchronous messages unless they - * have configured themselves as "slaves". */ - return ofconn->role != NX_ROLE_SLAVE; - } else { + } + + /* Keep the following code in sync with the documentation in the + * "Asynchronous Messages" section in DESIGN. */ + + if (ofconn->type == OFCONN_SERVICE && !ofconn->miss_send_len) { /* Service connections don't get asynchronous messages unless they have * explicitly asked for them by setting a nonzero miss send length. */ - return ofconn->miss_send_len > 0; + return false; + } + + async_config = (ofconn->role == OFPCR12_ROLE_SLAVE + ? ofconn->slave_async_config + : ofconn->master_async_config); + if (!(async_config[type] & (1u << reason))) { + return false; + } + + return true; +} + +/* The default "table-miss" behaviour for OpenFlow1.3+ is to drop the + * packet rather than to send the packet to the controller. + * + * This function returns false to indicate the packet should be dropped if + * the controller action was the result of the default table-miss behaviour + * and the controller is using OpenFlow1.3+. + * + * Otherwise true is returned to indicate the packet should be forwarded to + * the controller */ +static bool +ofconn_wants_packet_in_on_miss(struct ofconn *ofconn, + const struct ofproto_packet_in *pin) +{ + if (pin->miss_type == OFPROTO_PACKET_IN_MISS_WITHOUT_FLOW) { + enum ofputil_protocol protocol = ofconn_get_protocol(ofconn); + + if (protocol != OFPUTIL_P_NONE + && ofputil_protocol_to_ofp_version(protocol) >= OFP13_VERSION) { + enum ofproto_table_config config; + + config = ofproto_table_get_config(ofconn->connmgr->ofproto, + pin->up.table_id); + if (config == OFPROTO_TABLE_MISS_DEFAULT) { + return false; + } + } } + return true; +} + +/* The default "table-miss" behaviour for OpenFlow1.3+ is to drop the + * packet rather than to send the packet to the controller. + * + * This function returns false to indicate that a packet_in message + * for a "table-miss" should be sent to at least one controller. + * That is there is at least one controller with controller_id 0 + * which connected using an OpenFlow version earlier than OpenFlow1.3. + * + * False otherwise. + * + * This logic assumes that "table-miss" packet_in messages + * are always sent to controller_id 0. */ +bool +connmgr_wants_packet_in_on_miss(struct connmgr *mgr) +{ + struct ofconn *ofconn; + + LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { + enum ofputil_protocol protocol = ofconn_get_protocol(ofconn); + + if (ofconn->controller_id == 0 && + (protocol == OFPUTIL_P_NONE || + ofputil_protocol_to_ofp_version(protocol) < OFP13_VERSION)) { + return true; + } + } + return false; } /* Returns a human-readable name for an OpenFlow connection between 'mgr' and @@ -1087,41 +1505,60 @@ static void ofconn_send(const struct ofconn *ofconn, struct ofpbuf *msg, struct rconn_packet_counter *counter) { - update_openflow_length(msg); - if (rconn_send(ofconn->rconn, msg, counter)) { - ofpbuf_delete(msg); - } + ofpmsg_update_length(msg); + rconn_send(ofconn->rconn, msg, counter); } /* Sending asynchronous messages. */ -static void schedule_packet_in(struct ofconn *, struct ofputil_packet_in, - const struct flow *, struct ofpbuf *rw_packet); +static void schedule_packet_in(struct ofconn *, struct ofproto_packet_in, + enum ofp_packet_in_reason wire_reason); /* Sends an OFPT_PORT_STATUS message with 'opp' and 'reason' to appropriate - * controllers managed by 'mgr'. */ + * controllers managed by 'mgr'. For messages caused by a controller + * OFPT_PORT_MOD, specify 'source' as the controller connection that sent the + * request; otherwise, specify 'source' as NULL. */ void -connmgr_send_port_status(struct connmgr *mgr, const struct ofp_phy_port *opp, - uint8_t reason) +connmgr_send_port_status(struct connmgr *mgr, struct ofconn *source, + const struct ofputil_phy_port *pp, uint8_t reason) { /* XXX Should limit the number of queued port status change messages. */ + struct ofputil_port_status ps; struct ofconn *ofconn; + ps.reason = reason; + ps.desc = *pp; LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { - struct ofp_port_status *ops; - struct ofpbuf *b; + if (ofconn_receives_async_msg(ofconn, OAM_PORT_STATUS, reason)) { + struct ofpbuf *msg; + + /* Before 1.5, OpenFlow specified that OFPT_PORT_MOD should not + * generate OFPT_PORT_STATUS messages. That requirement was a + * relic of how OpenFlow originally supported a single controller, + * so that one could expect the controller to already know the + * changes it had made. + * + * EXT-338 changes OpenFlow 1.5 OFPT_PORT_MOD to send + * OFPT_PORT_STATUS messages to every controller. This is + * obviously more useful in the multi-controller case. We could + * always implement it that way in OVS, but that would risk + * confusing controllers that are intended for single-controller + * use only. (Imagine a controller that generates an OFPT_PORT_MOD + * in response to any OFPT_PORT_STATUS!) + * + * So this compromises: for OpenFlow 1.4 and earlier, it generates + * OFPT_PORT_STATUS for OFPT_PORT_MOD, but not back to the + * originating controller. In a single-controller environment, in + * particular, this means that it will never generate + * OFPT_PORT_STATUS for OFPT_PORT_MOD at all. */ + if (ofconn == source + && rconn_get_version(ofconn->rconn) < OFP15_VERSION) { + continue; + } - /* Primary controllers, even slaves, should always get port status - updates. Otherwise obey ofconn_receives_async_msgs(). */ - if (ofconn->type != OFCONN_PRIMARY - && !ofconn_receives_async_msgs(ofconn)) { - continue; + msg = ofputil_encode_port_status(&ps, ofconn_get_protocol(ofconn)); + ofconn_send(ofconn, msg, NULL); } - - ops = make_openflow_xid(sizeof *ops, OFPT_PORT_STATUS, 0, &b); - ops->reason = reason; - ops->desc = *opp; - ofconn_send(ofconn, b, NULL); } } @@ -1134,102 +1571,129 @@ connmgr_send_flow_removed(struct connmgr *mgr, struct ofconn *ofconn; LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { - struct ofpbuf *msg; - - if (!ofconn_receives_async_msgs(ofconn)) { - continue; + if (ofconn_receives_async_msg(ofconn, OAM_FLOW_REMOVED, fr->reason)) { + struct ofpbuf *msg; + + /* Account flow expirations as replies to OpenFlow requests. That + * works because preventing OpenFlow requests from being processed + * also prevents new flows from being added (and expiring). (It + * also prevents processing OpenFlow requests that would not add + * new flows, so it is imperfect.) */ + msg = ofputil_encode_flow_removed(fr, ofconn_get_protocol(ofconn)); + ofconn_send_reply(ofconn, msg); } + } +} + +/* Normally a send-to-controller action uses reason OFPR_ACTION. However, in + * OpenFlow 1.3 and later, packet_ins generated by a send-to-controller action + * in a "table-miss" flow (one with priority 0 and completely wildcarded) are + * sent as OFPR_NO_MATCH. This function returns the reason that should + * actually be sent on 'ofconn' for 'pin'. */ +static enum ofp_packet_in_reason +wire_reason(struct ofconn *ofconn, const struct ofproto_packet_in *pin) +{ + if (pin->miss_type == OFPROTO_PACKET_IN_MISS_FLOW + && pin->up.reason == OFPR_ACTION) { + enum ofputil_protocol protocol = ofconn_get_protocol(ofconn); - /* Account flow expirations as replies to OpenFlow requests. That - * works because preventing OpenFlow requests from being processed also - * prevents new flows from being added (and expiring). (It also - * prevents processing OpenFlow requests that would not add new flows, - * so it is imperfect.) */ - msg = ofputil_encode_flow_removed(fr, ofconn->flow_format); - ofconn_send_reply(ofconn, msg); + if (protocol != OFPUTIL_P_NONE + && ofputil_protocol_to_ofp_version(protocol) >= OFP13_VERSION) { + return OFPR_NO_MATCH; + } } + return pin->up.reason; } /* Given 'pin', sends an OFPT_PACKET_IN message to each OpenFlow controller as * necessary according to their individual configurations. * - * 'rw_packet' may be NULL. Otherwise, 'rw_packet' must contain the same data - * as pin->packet. (rw_packet == pin->packet is also valid.) Ownership of - * 'rw_packet' is transferred to this function. */ + * The caller doesn't need to fill in pin->buffer_id or pin->total_len. */ void connmgr_send_packet_in(struct connmgr *mgr, - const struct ofputil_packet_in *pin, - const struct flow *flow, struct ofpbuf *rw_packet) + const struct ofproto_packet_in *pin) { - struct ofconn *ofconn, *prev; + struct ofconn *ofconn; - prev = NULL; LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { - if (ofconn_receives_async_msgs(ofconn)) { - if (prev) { - schedule_packet_in(prev, *pin, flow, NULL); - } - prev = ofconn; + enum ofp_packet_in_reason reason = wire_reason(ofconn, pin); + + if (ofconn_wants_packet_in_on_miss(ofconn, pin) + && ofconn_receives_async_msg(ofconn, OAM_PACKET_IN, pin->up.reason) + && ofconn->controller_id == pin->controller_id) { + schedule_packet_in(ofconn, *pin, reason); } } - if (prev) { - schedule_packet_in(prev, *pin, flow, rw_packet); - } else { - ofpbuf_delete(rw_packet); - } } -/* pinsched callback for sending 'ofp_packet_in' on 'ofconn'. */ static void -do_send_packet_in(struct ofpbuf *ofp_packet_in, void *ofconn_) +do_send_packet_ins(struct ofconn *ofconn, struct list *txq) { - struct ofconn *ofconn = ofconn_; + struct ofpbuf *pin, *next_pin; + + LIST_FOR_EACH_SAFE (pin, next_pin, list_node, txq) { + list_remove(&pin->list_node); - rconn_send_with_limit(ofconn->rconn, ofp_packet_in, - ofconn->packet_in_counter, 100); + if (rconn_send_with_limit(ofconn->rconn, pin, + ofconn->packet_in_counter, 100) == EAGAIN) { + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 5); + + VLOG_INFO_RL(&rl, "%s: dropping packet-in due to queue overflow", + rconn_get_name(ofconn->rconn)); + } + } } -/* Takes 'pin', whose packet has the flow specified by 'flow', composes an - * OpenFlow packet-in message from it, and passes it to 'ofconn''s packet - * scheduler for sending. - * - * 'rw_packet' may be NULL. Otherwise, 'rw_packet' must contain the same data - * as pin->packet. (rw_packet == pin->packet is also valid.) Ownership of - * 'rw_packet' is transferred to this function. */ +/* Takes 'pin', composes an OpenFlow packet-in message from it, and passes it + * to 'ofconn''s packet scheduler for sending. */ static void -schedule_packet_in(struct ofconn *ofconn, struct ofputil_packet_in pin, - const struct flow *flow, struct ofpbuf *rw_packet) +schedule_packet_in(struct ofconn *ofconn, struct ofproto_packet_in pin, + enum ofp_packet_in_reason wire_reason) { struct connmgr *mgr = ofconn->connmgr; + uint16_t controller_max_len; + struct list txq; - /* Get OpenFlow buffer_id. */ - if (pin.reason == OFPR_ACTION) { - pin.buffer_id = UINT32_MAX; - } else if (mgr->fail_open && fail_open_is_active(mgr->fail_open)) { - pin.buffer_id = pktbuf_get_null(); - } else if (!ofconn->pktbuf) { - pin.buffer_id = UINT32_MAX; + pin.up.total_len = pin.up.packet_len; + + pin.up.reason = wire_reason; + if (pin.up.reason == OFPR_ACTION) { + controller_max_len = pin.send_len; /* max_len */ } else { - pin.buffer_id = pktbuf_save(ofconn->pktbuf, pin.packet, flow->in_port); + controller_max_len = ofconn->miss_send_len; } - /* Figure out how much of the packet to send. */ - if (pin.reason == OFPR_NO_MATCH) { - pin.send_len = pin.packet->size; + /* Get OpenFlow buffer_id. + * For OpenFlow 1.2+, OFPCML_NO_BUFFER (== UINT16_MAX) specifies + * unbuffered. This behaviour doesn't violate prior versions, too. */ + if (controller_max_len == UINT16_MAX) { + pin.up.buffer_id = UINT32_MAX; + } else if (mgr->fail_open && fail_open_is_active(mgr->fail_open)) { + pin.up.buffer_id = pktbuf_get_null(); + } else if (!ofconn->pktbuf) { + pin.up.buffer_id = UINT32_MAX; } else { - /* Caller should have initialized 'send_len' to 'max_len' specified in - * struct ofp_action_output. */ + pin.up.buffer_id = pktbuf_save(ofconn->pktbuf, + pin.up.packet, pin.up.packet_len, + pin.up.fmd.in_port); } - if (pin.buffer_id != UINT32_MAX) { - pin.send_len = MIN(pin.send_len, ofconn->miss_send_len); + + /* Figure out how much of the packet to send. + * If not buffered, send the entire packet. Otherwise, depending on + * the reason of packet-in, send what requested by the controller. */ + if (pin.up.buffer_id != UINT32_MAX + && controller_max_len < pin.up.packet_len) { + pin.up.packet_len = controller_max_len; } - /* Make OFPT_PACKET_IN and hand over to packet scheduler. It might - * immediately call into do_send_packet_in() or it might buffer it for a - * while (until a later call to pinsched_run()). */ - pinsched_send(ofconn->schedulers[pin.reason == OFPR_NO_MATCH ? 0 : 1], - flow->in_port, ofputil_encode_packet_in(&pin, rw_packet), - do_send_packet_in, ofconn); + /* Make OFPT_PACKET_IN and hand over to packet scheduler. */ + pinsched_send(ofconn->schedulers[pin.up.reason == OFPR_NO_MATCH ? 0 : 1], + pin.up.fmd.in_port, + ofputil_encode_packet_in(&pin.up, + ofconn_get_protocol(ofconn), + ofconn->packet_in_format), + &txq); + do_send_packet_ins(ofconn, &txq); } /* Fail-open settings. */ @@ -1324,29 +1788,6 @@ connmgr_is_any_controller_admitted(const struct connmgr *mgr) } return false; } - -/* Sends 'packet' to each controller connected to 'mgr'. Takes ownership of - * 'packet'. */ -void -connmgr_broadcast(struct connmgr *mgr, struct ofpbuf *packet) -{ - struct ofconn *ofconn, *prev; - - prev = NULL; - LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { - if (prev) { - ofconn_send_reply(ofconn, ofpbuf_clone(packet)); - } - if (rconn_is_connected(ofconn->rconn)) { - prev = ofconn; - } - } - if (prev) { - ofconn_send_reply(prev, packet); - } else { - ofpbuf_delete(packet); - } -} /* In-band configuration. */ @@ -1409,18 +1850,9 @@ any_extras_changed(const struct connmgr *mgr, /* In-band implementation. */ bool -connmgr_msg_in_hook(struct connmgr *mgr, const struct flow *flow, - const struct ofpbuf *packet) -{ - return mgr->in_band && in_band_msg_in_hook(mgr->in_band, flow, packet); -} - -bool -connmgr_may_set_up_flow(struct connmgr *mgr, const struct flow *flow, - const struct nlattr *odp_actions, - size_t actions_len) +connmgr_has_in_band(struct connmgr *mgr) { - return !mgr->in_band || in_band_rule_check(flow, odp_actions, actions_len); + return mgr->in_band != NULL; } /* Fail-open and in-band implementation. */ @@ -1431,6 +1863,7 @@ connmgr_may_set_up_flow(struct connmgr *mgr, const struct flow *flow, * In-band control has more sophisticated code that manages flows itself. */ void connmgr_flushed(struct connmgr *mgr) + OVS_EXCLUDED(ofproto_mutex) { if (mgr->fail_open) { fail_open_flushed(mgr->fail_open); @@ -1442,15 +1875,18 @@ connmgr_flushed(struct connmgr *mgr) * traffic until a controller has been defined and it tells us to do so. */ if (!connmgr_has_controllers(mgr) && mgr->fail_mode == OFPROTO_FAIL_STANDALONE) { - union ofp_action action; - struct cls_rule rule; + struct ofpbuf ofpacts; + struct match match; + + ofpbuf_init(&ofpacts, OFPACT_OUTPUT_SIZE); + ofpact_put_OUTPUT(&ofpacts)->port = OFPP_NORMAL; + ofpact_pad(&ofpacts); - memset(&action, 0, sizeof action); - action.type = htons(OFPAT_OUTPUT); - action.output.len = htons(sizeof action); - action.output.port = htons(OFPP_NORMAL); - cls_rule_init_catchall(&rule, 0); - ofproto_add_flow(mgr->ofproto, &rule, &action, 1); + match_init_catchall(&match); + ofproto_add_flow(mgr->ofproto, &match, 0, ofpbuf_data(&ofpacts), + ofpbuf_size(&ofpacts)); + + ofpbuf_uninit(&ofpacts); } } @@ -1460,13 +1896,14 @@ connmgr_flushed(struct connmgr *mgr) * ofservice_reconfigure() must be called to fully configure the new * ofservice. */ static int -ofservice_create(struct connmgr *mgr, const char *target) +ofservice_create(struct connmgr *mgr, const char *target, + uint32_t allowed_versions, uint8_t dscp) { struct ofservice *ofservice; struct pvconn *pvconn; int error; - error = pvconn_open(target, &pvconn); + error = pvconn_open(target, allowed_versions, dscp, &pvconn); if (error) { return error; } @@ -1474,6 +1911,7 @@ ofservice_create(struct connmgr *mgr, const char *target) ofservice = xzalloc(sizeof *ofservice); hmap_insert(&mgr->services, &ofservice->node, hash_string(target, 0)); ofservice->pvconn = pvconn; + ofservice->allowed_versions = allowed_versions; return 0; } @@ -1493,6 +1931,8 @@ ofservice_reconfigure(struct ofservice *ofservice, ofservice->probe_interval = c->probe_interval; ofservice->rate_limit = c->rate_limit; ofservice->burst_limit = c->burst_limit; + ofservice->enable_async_msgs = c->enable_async_msgs; + ofservice->dscp = c->dscp; } /* Finds and returns the ofservice within 'mgr' that has the given @@ -1510,3 +1950,268 @@ ofservice_lookup(struct connmgr *mgr, const char *target) } return NULL; } + +/* Flow monitors (NXST_FLOW_MONITOR). */ + +/* A counter incremented when something significant happens to an OpenFlow + * rule. + * + * - When a rule is added, its 'add_seqno' and 'modify_seqno' are set to + * the current value (which is then incremented). + * + * - When a rule is modified, its 'modify_seqno' is set to the current + * value (which is then incremented). + * + * Thus, by comparing an old value of monitor_seqno against a rule's + * 'add_seqno', one can tell whether the rule was added before or after the old + * value was read, and similarly for 'modify_seqno'. + * + * 32 bits should normally be sufficient (and would be nice, to save space in + * each rule) but then we'd have to have some special cases for wraparound. + * + * We initialize monitor_seqno to 1 to allow 0 to be used as an invalid + * value. */ +static uint64_t monitor_seqno = 1; + +COVERAGE_DEFINE(ofmonitor_pause); +COVERAGE_DEFINE(ofmonitor_resume); + +enum ofperr +ofmonitor_create(const struct ofputil_flow_monitor_request *request, + struct ofconn *ofconn, struct ofmonitor **monitorp) + OVS_REQUIRES(ofproto_mutex) +{ + struct ofmonitor *m; + + *monitorp = NULL; + + m = ofmonitor_lookup(ofconn, request->id); + if (m) { + return OFPERR_NXBRC_FM_DUPLICATE_ID; + } + + m = xmalloc(sizeof *m); + m->ofconn = ofconn; + hmap_insert(&ofconn->monitors, &m->ofconn_node, hash_int(request->id, 0)); + m->id = request->id; + m->flags = request->flags; + m->out_port = request->out_port; + m->table_id = request->table_id; + minimatch_init(&m->match, &request->match); + + *monitorp = m; + return 0; +} + +struct ofmonitor * +ofmonitor_lookup(struct ofconn *ofconn, uint32_t id) + OVS_REQUIRES(ofproto_mutex) +{ + struct ofmonitor *m; + + HMAP_FOR_EACH_IN_BUCKET (m, ofconn_node, hash_int(id, 0), + &ofconn->monitors) { + if (m->id == id) { + return m; + } + } + return NULL; +} + +void +ofmonitor_destroy(struct ofmonitor *m) + OVS_REQUIRES(ofproto_mutex) +{ + if (m) { + minimatch_destroy(&m->match); + hmap_remove(&m->ofconn->monitors, &m->ofconn_node); + free(m); + } +} + +void +ofmonitor_report(struct connmgr *mgr, struct rule *rule, + enum nx_flow_update_event event, + enum ofp_flow_removed_reason reason, + const struct ofconn *abbrev_ofconn, ovs_be32 abbrev_xid) + OVS_REQUIRES(ofproto_mutex) +{ + enum nx_flow_monitor_flags update; + struct ofconn *ofconn; + + switch (event) { + case NXFME_ADDED: + update = NXFMF_ADD; + rule->add_seqno = rule->modify_seqno = monitor_seqno++; + break; + + case NXFME_DELETED: + update = NXFMF_DELETE; + break; + + case NXFME_MODIFIED: + update = NXFMF_MODIFY; + rule->modify_seqno = monitor_seqno++; + break; + + default: + case NXFME_ABBREV: + OVS_NOT_REACHED(); + } + + LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { + enum nx_flow_monitor_flags flags = 0; + struct ofmonitor *m; + + if (ofconn->monitor_paused) { + /* Only send NXFME_DELETED notifications for flows that were added + * before we paused. */ + if (event != NXFME_DELETED + || rule->add_seqno > ofconn->monitor_paused) { + continue; + } + } + + HMAP_FOR_EACH (m, ofconn_node, &ofconn->monitors) { + if (m->flags & update + && (m->table_id == 0xff || m->table_id == rule->table_id) + && ofoperation_has_out_port(rule->pending, m->out_port) + && cls_rule_is_loose_match(&rule->cr, &m->match)) { + flags |= m->flags; + } + } + + if (flags) { + if (list_is_empty(&ofconn->updates)) { + ofputil_start_flow_update(&ofconn->updates); + ofconn->sent_abbrev_update = false; + } + + if (ofconn != abbrev_ofconn || ofconn->monitor_paused) { + struct ofputil_flow_update fu; + struct match match; + + fu.event = event; + fu.reason = event == NXFME_DELETED ? reason : 0; + fu.table_id = rule->table_id; + fu.cookie = rule->flow_cookie; + minimatch_expand(&rule->cr.match, &match); + fu.match = &match; + fu.priority = rule->cr.priority; + + ovs_mutex_lock(&rule->mutex); + fu.idle_timeout = rule->idle_timeout; + fu.hard_timeout = rule->hard_timeout; + ovs_mutex_unlock(&rule->mutex); + + if (flags & NXFMF_ACTIONS) { + const struct rule_actions *actions = rule_get_actions(rule); + fu.ofpacts = actions->ofpacts; + fu.ofpacts_len = actions->ofpacts_len; + } else { + fu.ofpacts = NULL; + fu.ofpacts_len = 0; + } + ofputil_append_flow_update(&fu, &ofconn->updates); + } else if (!ofconn->sent_abbrev_update) { + struct ofputil_flow_update fu; + + fu.event = NXFME_ABBREV; + fu.xid = abbrev_xid; + ofputil_append_flow_update(&fu, &ofconn->updates); + + ofconn->sent_abbrev_update = true; + } + } + } +} + +void +ofmonitor_flush(struct connmgr *mgr) + OVS_REQUIRES(ofproto_mutex) +{ + struct ofconn *ofconn; + + LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { + struct ofpbuf *msg, *next; + + LIST_FOR_EACH_SAFE (msg, next, list_node, &ofconn->updates) { + unsigned int n_bytes; + + list_remove(&msg->list_node); + ofconn_send(ofconn, msg, ofconn->monitor_counter); + n_bytes = rconn_packet_counter_n_bytes(ofconn->monitor_counter); + if (!ofconn->monitor_paused && n_bytes > 128 * 1024) { + struct ofpbuf *pause; + + COVERAGE_INC(ofmonitor_pause); + ofconn->monitor_paused = monitor_seqno++; + pause = ofpraw_alloc_xid(OFPRAW_NXT_FLOW_MONITOR_PAUSED, + OFP10_VERSION, htonl(0), 0); + ofconn_send(ofconn, pause, ofconn->monitor_counter); + } + } + } +} + +static void +ofmonitor_resume(struct ofconn *ofconn) + OVS_REQUIRES(ofproto_mutex) +{ + struct rule_collection rules; + struct ofpbuf *resumed; + struct ofmonitor *m; + struct list msgs; + + rule_collection_init(&rules); + HMAP_FOR_EACH (m, ofconn_node, &ofconn->monitors) { + ofmonitor_collect_resume_rules(m, ofconn->monitor_paused, &rules); + } + + list_init(&msgs); + ofmonitor_compose_refresh_updates(&rules, &msgs); + + resumed = ofpraw_alloc_xid(OFPRAW_NXT_FLOW_MONITOR_RESUMED, OFP10_VERSION, + htonl(0), 0); + list_push_back(&msgs, &resumed->list_node); + ofconn_send_replies(ofconn, &msgs); + + ofconn->monitor_paused = 0; +} + +static bool +ofmonitor_may_resume(const struct ofconn *ofconn) + OVS_REQUIRES(ofproto_mutex) +{ + return (ofconn->monitor_paused != 0 + && !rconn_packet_counter_n_packets(ofconn->monitor_counter)); +} + +static void +ofmonitor_run(struct connmgr *mgr) +{ + struct ofconn *ofconn; + + ovs_mutex_lock(&ofproto_mutex); + LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { + if (ofmonitor_may_resume(ofconn)) { + COVERAGE_INC(ofmonitor_resume); + ofmonitor_resume(ofconn); + } + } + ovs_mutex_unlock(&ofproto_mutex); +} + +static void +ofmonitor_wait(struct connmgr *mgr) +{ + struct ofconn *ofconn; + + ovs_mutex_lock(&ofproto_mutex); + LIST_FOR_EACH (ofconn, node, &mgr->all_conns) { + if (ofmonitor_may_resume(ofconn)) { + poll_immediate_wake(); + } + } + ovs_mutex_unlock(&ofproto_mutex); +}