X-Git-Url: http://git.cascardo.eti.br/?a=blobdiff_plain;f=ofproto%2Fofproto-dpif-upcall.c;h=b505206093b9acaacadd34ad0e5401e7e73d6a0d;hb=9bfe933472a784cd6b3f946afc3eed73d74f6a7d;hp=3ef1b9a64eb90083429bbe1cefac9979a142cee9;hpb=27130224cd5c3e5d49771d0284edc4042d2cf520;p=cascardo%2Fovs.git diff --git a/ofproto/ofproto-dpif-upcall.c b/ofproto/ofproto-dpif-upcall.c index 3ef1b9a64..b50520609 100644 --- a/ofproto/ofproto-dpif-upcall.c +++ b/ofproto/ofproto-dpif-upcall.c @@ -1,4 +1,4 @@ -/* Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2015 Nicira, Inc. +/* Copyright (c) 2009, 2010, 2011, 2012, 2013, 2014, 2015, 2016 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -61,8 +61,8 @@ struct handler { }; /* In the absence of a multiple-writer multiple-reader datastructure for - * storing ukeys, we use a large number of cmaps, each with its own lock for - * writing. */ + * storing udpif_keys ("ukeys"), we use a large number of cmaps, each with its + * own lock for writing. */ #define N_UMAPS 512 /* per udpif. */ struct umap { struct ovs_mutex mutex; /* Take for writing to the following. */ @@ -70,7 +70,29 @@ struct umap { }; /* A thread that processes datapath flows, updates OpenFlow statistics, and - * updates or removes them if necessary. */ + * updates or removes them if necessary. + * + * Revalidator threads operate in two phases: "dump" and "sweep". In between + * each phase, all revalidators sync up so that all revalidator threads are + * either in one phase or the other, but not a combination. + * + * During the dump phase, revalidators fetch flows from the datapath and + * attribute the statistics to OpenFlow rules. Each datapath flow has a + * corresponding ukey which caches the most recently seen statistics. If + * a flow needs to be deleted (for example, because it is unused over a + * period of time), revalidator threads may delete the flow during the + * dump phase. The datapath is not guaranteed to reliably dump all flows + * from the datapath, and there is no mapping between datapath flows to + * revalidators, so a particular flow may be handled by zero or more + * revalidators during a single dump phase. To avoid duplicate attribution + * of statistics, ukeys are never deleted during this phase. + * + * During the sweep phase, each revalidator takes ownership of a different + * slice of umaps and sweeps through all ukeys in those umaps to figure out + * whether they need to be deleted. During this phase, revalidators may + * fetch individual flows which were not dumped during the dump phase to + * validate them and attribute statistics. + */ struct revalidator { struct udpif *udpif; /* Parent udpif. */ pthread_t thread; /* Thread ID. */ @@ -261,9 +283,8 @@ struct udpif_key { struct nlattr nla; } keybuf, maskbuf; - /* Recirculation IDs with references held by the ukey. */ - unsigned n_recircs; - uint32_t recircs[]; /* 'n_recircs' id's for which references are held. */ + uint32_t key_recirc_id; /* Non-zero if reference is held by the ukey. */ + struct recirc_refs recircs; /* Action recirc IDs with references held. */ }; /* Datapath operation with optional ukey attached. */ @@ -1069,36 +1090,11 @@ upcall_xlate(struct udpif *udpif, struct upcall *upcall, xlate_actions(&xin, &upcall->xout); upcall->xout_initialized = true; - /* Special case for fail-open mode. - * - * If we are in fail-open mode, but we are connected to a controller too, - * then we should send the packet up to the controller in the hope that it - * will try to set up a flow and thereby allow us to exit fail-open. - * - * See the top-level comment in fail-open.c for more information. - * - * Copy packets before they are modified by execution. */ - if (upcall->xout.fail_open) { - const struct dp_packet *packet = upcall->packet; - struct ofproto_packet_in *pin; - - pin = xmalloc(sizeof *pin); - pin->up.packet = xmemdup(dp_packet_data(packet), dp_packet_size(packet)); - pin->up.packet_len = dp_packet_size(packet); - pin->up.reason = OFPR_NO_MATCH; - pin->up.table_id = 0; - pin->up.cookie = OVS_BE64_MAX; - flow_get_metadata(upcall->flow, &pin->up.flow_metadata); - pin->send_len = 0; /* Not used for flow table misses. */ - pin->miss_type = OFPROTO_PACKET_IN_NO_MISS; - ofproto_dpif_send_packet_in(upcall->ofproto, pin); - } - if (!upcall->xout.slow) { ofpbuf_use_const(&upcall->put_actions, odp_actions->data, odp_actions->size); } else { - ofpbuf_init(&upcall->put_actions, 0); + /* upcall->put_actions already initialized by upcall_receive(). */ compose_slow_path(udpif, &upcall->xout, upcall->flow, upcall->flow->in_port.odp_port, &upcall->put_actions); @@ -1293,10 +1289,8 @@ handle_upcalls(struct udpif *udpif, struct upcall *upcalls, unsigned int flow_limit; size_t n_ops, n_opsp, i; bool may_put; - bool megaflow; atomic_read_relaxed(&udpif->flow_limit, &flow_limit); - atomic_read_relaxed(&enable_megaflows, &megaflow); may_put = udpif_get_n_flows(udpif) < flow_limit; @@ -1448,12 +1442,10 @@ ukey_create__(const struct nlattr *key, size_t key_len, bool ufid_present, const ovs_u128 *ufid, const unsigned pmd_id, const struct ofpbuf *actions, uint64_t dump_seq, uint64_t reval_seq, long long int used, - const struct recirc_id_node *key_recirc, struct xlate_out *xout) + uint32_t key_recirc_id, struct xlate_out *xout) OVS_NO_THREAD_SAFETY_ANALYSIS { - unsigned n_recircs = (key_recirc ? 1 : 0) + (xout ? xout->n_recircs : 0); - struct udpif_key *ukey = xmalloc(sizeof *ukey + - n_recircs * sizeof *ukey->recircs); + struct udpif_key *ukey = xmalloc(sizeof *ukey); memcpy(&ukey->keybuf, key, key_len); ukey->key = &ukey->keybuf.nla; @@ -1478,17 +1470,13 @@ ukey_create__(const struct nlattr *key, size_t key_len, ukey->stats.used = used; ukey->xcache = NULL; - ukey->n_recircs = n_recircs; - if (key_recirc) { - ukey->recircs[0] = key_recirc->id; + ukey->key_recirc_id = key_recirc_id; + recirc_refs_init(&ukey->recircs); + if (xout) { + /* Take ownership of the action recirc id references. */ + recirc_refs_swap(&ukey->recircs, &xout->recircs); } - if (xout && xout->n_recircs) { - const uint32_t *act_recircs = xlate_out_get_recircs(xout); - memcpy(ukey->recircs + (key_recirc ? 1 : 0), act_recircs, - xout->n_recircs * sizeof *ukey->recircs); - xlate_out_take_recircs(xout); - } return ukey; } @@ -1527,7 +1515,7 @@ ukey_create_from_upcall(struct upcall *upcall, struct flow_wildcards *wc) true, upcall->ufid, upcall->pmd_id, &upcall->put_actions, upcall->dump_seq, upcall->reval_seq, 0, - upcall->have_recirc_ref ? upcall->recirc : NULL, + upcall->have_recirc_ref ? upcall->recirc->id : 0, &upcall->xout); } @@ -1560,7 +1548,14 @@ ukey_create_from_dpif_flow(const struct udpif *udpif, /* Check the flow actions for recirculation action. As recirculation * relies on OVS userspace internal state, we need to delete all old - * datapath flows with recirculation upon OVS restart. */ + * datapath flows with either a non-zero recirc_id in the key, or any + * recirculation actions upon OVS restart. */ + NL_ATTR_FOR_EACH_UNSAFE (a, left, flow->key, flow->key_len) { + if (nl_attr_type(a) == OVS_KEY_ATTR_RECIRC_ID + && nl_attr_get_u32(a) != 0) { + return EINVAL; + } + } NL_ATTR_FOR_EACH_UNSAFE (a, left, flow->actions, flow->actions_len) { if (nl_attr_type(a) == OVS_ACTION_ATTR_RECIRC) { return EINVAL; @@ -1573,7 +1568,7 @@ ukey_create_from_dpif_flow(const struct udpif *udpif, *ukey = ukey_create__(flow->key, flow->key_len, flow->mask, flow->mask_len, flow->ufid_present, &flow->ufid, flow->pmd_id, &actions, dump_seq, - reval_seq, flow->stats.used, NULL, NULL); + reval_seq, flow->stats.used, 0, NULL); return 0; } @@ -1716,9 +1711,10 @@ ukey_delete__(struct udpif_key *ukey) OVS_NO_THREAD_SAFETY_ANALYSIS { if (ukey) { - for (int i = 0; i < ukey->n_recircs; i++) { - recirc_free_id(ukey->recircs[i]); + if (ukey->key_recirc_id) { + recirc_free_id(ukey->key_recirc_id); } + recirc_refs_unref(&ukey->recircs); xlate_cache_delete(ukey->xcache); ofpbuf_delete(ovsrcu_get(struct ofpbuf *, &ukey->actions)); ovs_mutex_destroy(&ukey->mutex); @@ -1775,11 +1771,21 @@ should_revalidate(const struct udpif *udpif, uint64_t packets, * UKEY_KEEP The ukey is fine as is. * UKEY_MODIFY The ukey's actions should be changed but is otherwise * fine. Callers should change the actions to those found - * in the caller supplied 'odp_actions' buffer. */ + * in the caller supplied 'odp_actions' buffer. The + * recirculation references can be found in 'recircs' and + * must be handled by the caller. + * + * If the result is UKEY_MODIFY, then references to all recirc_ids used by the + * new flow will be held within 'recircs' (which may be none). + * + * The caller is responsible for both initializing 'recircs' prior this call, + * and ensuring any references are eventually freed. + */ static enum reval_result revalidate_ukey(struct udpif *udpif, struct udpif_key *ukey, const struct dpif_flow_stats *stats, - struct ofpbuf *odp_actions, uint64_t reval_seq) + struct ofpbuf *odp_actions, uint64_t reval_seq, + struct recirc_refs *recircs) OVS_REQUIRES(ukey->mutex) { struct xlate_out xout, *xoutp; @@ -1869,7 +1875,7 @@ revalidate_ukey(struct udpif *udpif, struct udpif_key *ukey, } if (odp_flow_key_to_mask(ukey->mask, ukey->mask_len, ukey->key, - ukey->key_len, &dp_mask.masks, &flow) + ukey->key_len, &dp_mask, &flow) == ODP_FIT_ERROR) { goto exit; } @@ -1888,6 +1894,8 @@ revalidate_ukey(struct udpif *udpif, struct udpif_key *ukey, /* The datapath mask was OK, but the actions seem to have changed. * Let's modify it in place. */ result = UKEY_MODIFY; + /* Transfer recirc action ID references to the caller. */ + recirc_refs_swap(recircs, &xoutp->recircs); goto exit; } @@ -1948,8 +1956,10 @@ modify_op_init(struct ukey_op *op, struct udpif_key *ukey) &op->dop.u.flow_put.actions_len); } +/* Executes datapath operations 'ops' and attributes stats retrieved from the + * datapath as part of those operations. */ static void -push_ukey_ops__(struct udpif *udpif, struct ukey_op *ops, size_t n_ops) +push_dp_ops(struct udpif *udpif, struct ukey_op *ops, size_t n_ops) { struct dpif_op *opsp[REVALIDATE_MAX_BATCH]; size_t i; @@ -2033,16 +2043,20 @@ push_ukey_ops__(struct udpif *udpif, struct ukey_op *ops, size_t n_ops) } } +/* Executes datapath operations 'ops', attributes stats retrieved from the + * datapath, and deletes ukeys corresponding to deleted flows. */ static void push_ukey_ops(struct udpif *udpif, struct umap *umap, struct ukey_op *ops, size_t n_ops) { int i; - push_ukey_ops__(udpif, ops, n_ops); + push_dp_ops(udpif, ops, n_ops); ovs_mutex_lock(&umap->mutex); for (i = 0; i < n_ops; i++) { - ukey_delete(umap, ops[i].ukey); + if (ops[i].dop.type == DPIF_OP_FLOW_DEL) { + ukey_delete(umap, ops[i].ukey); + } } ovs_mutex_unlock(&umap->mutex); } @@ -2059,6 +2073,25 @@ log_unexpected_flow(const struct dpif_flow *flow, int error) VLOG_WARN_RL(&rl, "%s", ds_cstr(&ds)); } +static void +reval_op_init(struct ukey_op *op, enum reval_result result, + struct udpif *udpif, struct udpif_key *ukey, + struct recirc_refs *recircs, struct ofpbuf *odp_actions) +{ + if (result == UKEY_DELETE) { + delete_op_init(udpif, op, ukey); + } else if (result == UKEY_MODIFY) { + /* Store the new recircs. */ + recirc_refs_swap(&ukey->recircs, recircs); + /* Release old recircs. */ + recirc_refs_unref(recircs); + /* ukey->key_recirc_id remains, as the key is the same as before. */ + + ukey_set_actions(ukey, odp_actions); + modify_op_init(op, ukey); + } +} + static void revalidate(struct revalidator *revalidator) { @@ -2112,6 +2145,7 @@ revalidate(struct revalidator *revalidator) for (f = flows; f < &flows[n_dumped]; f++) { long long int used = f->stats.used; + struct recirc_refs recircs = RECIRC_REFS_EMPTY_INITIALIZER; enum reval_result result; struct udpif_key *ukey; bool already_dumped; @@ -2151,22 +2185,22 @@ revalidate(struct revalidator *revalidator) result = UKEY_DELETE; } else { result = revalidate_ukey(udpif, ukey, &f->stats, &odp_actions, - reval_seq); + reval_seq, &recircs); } ukey->dump_seq = dump_seq; ukey->flow_exists = result != UKEY_DELETE; - if (result == UKEY_DELETE) { - delete_op_init(udpif, &ops[n_ops++], ukey); - } else if (result == UKEY_MODIFY) { - ukey_set_actions(ukey, &odp_actions); - modify_op_init(&ops[n_ops++], ukey); + if (result != UKEY_KEEP) { + /* Takes ownership of 'recircs'. */ + reval_op_init(&ops[n_ops++], result, udpif, ukey, &recircs, + &odp_actions); } ovs_mutex_unlock(&ukey->mutex); } if (n_ops) { - push_ukey_ops__(udpif, ops, n_ops); + /* Push datapath ops but defer ukey deletion to 'sweep' phase. */ + push_dp_ops(udpif, ops, n_ops); } ovsrcu_quiesce(); } @@ -2208,8 +2242,7 @@ revalidator_sweep__(struct revalidator *revalidator, bool purge) size_t n_ops = 0; CMAP_FOR_EACH(ukey, cmap_node, &umap->cmap) { - bool flow_exists, seq_mismatch; - enum reval_result result; + bool flow_exists; /* Handler threads could be holding a ukey lock while it installs a * new flow, so don't hang around waiting for access to it. */ @@ -2217,39 +2250,45 @@ revalidator_sweep__(struct revalidator *revalidator, bool purge) continue; } flow_exists = ukey->flow_exists; - seq_mismatch = (ukey->dump_seq != dump_seq - && ukey->reval_seq != reval_seq); - - if (purge) { - result = UKEY_DELETE; - } else if (!seq_mismatch) { - result = UKEY_KEEP; - } else { - struct dpif_flow_stats stats; - COVERAGE_INC(revalidate_missed_dp_flow); - memset(&stats, 0, sizeof stats); - result = revalidate_ukey(udpif, ukey, &stats, &odp_actions, - reval_seq); + if (flow_exists) { + struct recirc_refs recircs = RECIRC_REFS_EMPTY_INITIALIZER; + bool seq_mismatch = (ukey->dump_seq != dump_seq + && ukey->reval_seq != reval_seq); + enum reval_result result; + + if (purge) { + result = UKEY_DELETE; + } else if (!seq_mismatch) { + result = UKEY_KEEP; + } else { + struct dpif_flow_stats stats; + COVERAGE_INC(revalidate_missed_dp_flow); + memset(&stats, 0, sizeof stats); + result = revalidate_ukey(udpif, ukey, &stats, &odp_actions, + reval_seq, &recircs); + } + if (result != UKEY_KEEP) { + /* Clears 'recircs' if filled by revalidate_ukey(). */ + reval_op_init(&ops[n_ops++], result, udpif, ukey, &recircs, + &odp_actions); + } } ovs_mutex_unlock(&ukey->mutex); - if (result == UKEY_DELETE) { - delete_op_init(udpif, &ops[n_ops++], ukey); - } else if (result == UKEY_MODIFY) { - ukey_set_actions(ukey, &odp_actions); - modify_op_init(&ops[n_ops++], ukey); + if (!flow_exists) { + /* The common flow deletion case involves deletion of the flow + * during the dump phase and ukey deletion here. */ + ovs_mutex_lock(&umap->mutex); + ukey_delete(umap, ukey); + ovs_mutex_unlock(&umap->mutex); } if (n_ops == REVALIDATE_MAX_BATCH) { + /* Update/delete missed flows and clean up corresponding ukeys + * if necessary. */ push_ukey_ops(udpif, umap, ops, n_ops); n_ops = 0; } - - if (!flow_exists) { - ovs_mutex_lock(&umap->mutex); - ukey_delete(umap, ukey); - ovs_mutex_unlock(&umap->mutex); - } } if (n_ops) {