VLOG_DEFINE_THIS_MODULE(connmgr);
static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 5);
-COVERAGE_DEFINE(ofconn_stuck);
-
/* An OpenFlow connection. */
struct ofconn {
struct connmgr *connmgr; /* Connection's manager. */
enum nx_flow_format flow_format; /* Currently selected flow format. */
bool flow_mod_table_id; /* NXT_FLOW_MOD_TABLE_ID enabled? */
+ /* Asynchronous flow table operation support. */
+ struct list opgroups; /* Contains pending "ofopgroups", if any. */
+ struct ofpbuf *blocked; /* Postponed OpenFlow message, if any. */
+ bool retry; /* True if 'blocked' is ready to try again. */
+
/* OFPT_PACKET_IN related data. */
struct rconn_packet_counter *packet_in_counter; /* # queued on 'rconn'. */
#define N_SCHEDULERS 2
const struct ofproto_controller *);
static void ofconn_run(struct ofconn *,
- void (*handle_openflow)(struct ofconn *,
+ bool (*handle_openflow)(struct ofconn *,
struct ofpbuf *ofp_msg));
-static void ofconn_wait(struct ofconn *);
+static void ofconn_wait(struct ofconn *, bool handling_openflow);
static const char *ofconn_get_target(const struct ofconn *);
static char *ofconn_make_name(const struct connmgr *, const char *target);
free(mgr);
}
-/* Does all of the periodic maintenance required by 'mgr'. Calls
- * 'handle_openflow' for each message received on an OpenFlow connection,
- * passing along the OpenFlow connection itself and the message that was sent.
- * The 'handle_openflow' callback must not free the message. */
+/* Does all of the periodic maintenance required by 'mgr'.
+ *
+ * If 'handle_openflow' is nonnull, calls 'handle_openflow' for each message
+ * received on an OpenFlow connection, passing along the OpenFlow connection
+ * itself and the message that was sent. If 'handle_openflow' returns true,
+ * the message is considered to be fully processed. If 'handle_openflow'
+ * returns false, the message is considered not to have been processed at all;
+ * it will be stored and re-presented to 'handle_openflow' following the next
+ * call to connmgr_retry(). 'handle_openflow' must not modify or free the
+ * message.
+ *
+ * If 'handle_openflow' is NULL, no OpenFlow messages will be processed and
+ * other activities that could affect the flow table (in-band processing,
+ * fail-open processing) are suppressed too. */
void
connmgr_run(struct connmgr *mgr,
- void (*handle_openflow)(struct ofconn *, struct ofpbuf *ofp_msg))
+ bool (*handle_openflow)(struct ofconn *, struct ofpbuf *ofp_msg))
{
struct ofconn *ofconn, *next_ofconn;
struct ofservice *ofservice;
size_t i;
- if (mgr->in_band) {
+ if (handle_openflow && mgr->in_band) {
if (time_msec() >= mgr->next_in_band_update) {
update_in_band_remotes(mgr);
}
/* Fail-open maintenance. Do this after processing the ofconns since
* fail-open checks the status of the controller rconn. */
- if (mgr->fail_open) {
+ if (handle_openflow && mgr->fail_open) {
fail_open_run(mgr->fail_open);
}
}
}
-/* Causes the poll loop to wake up when connmgr_run() needs to run. */
+/* Causes the poll loop to wake up when connmgr_run() needs to run.
+ *
+ * If 'handling_openflow' is true, arriving OpenFlow messages and other
+ * activities that affect the flow table will wake up the poll loop. If
+ * 'handling_openflow' is false, they will not. */
void
-connmgr_wait(struct connmgr *mgr)
+connmgr_wait(struct connmgr *mgr, bool handling_openflow)
{
struct ofservice *ofservice;
struct ofconn *ofconn;
size_t i;
LIST_FOR_EACH (ofconn, node, &mgr->all_conns) {
- ofconn_wait(ofconn);
+ ofconn_wait(ofconn, handling_openflow);
}
- if (mgr->in_band) {
+ if (handling_openflow && mgr->in_band) {
poll_timer_wait_until(mgr->next_in_band_update);
in_band_wait(mgr->in_band);
}
- if (mgr->fail_open) {
+ if (handling_openflow && mgr->fail_open) {
fail_open_wait(mgr->fail_open);
}
HMAP_FOR_EACH (ofservice, node, &mgr->services) {
{
return ofconn->connmgr->ofproto;
}
+
+/* If processing of OpenFlow messages was blocked on any 'mgr' ofconns by
+ * returning false to the 'handle_openflow' callback to connmgr_run(), this
+ * re-enables them. */
+void
+connmgr_retry(struct connmgr *mgr)
+{
+ struct ofconn *ofconn;
+
+ LIST_FOR_EACH (ofconn, node, &mgr->all_conns) {
+ ofconn->retry = true;
+ }
+}
\f
/* OpenFlow configuration. */
{
return pktbuf_retrieve(ofconn->pktbuf, id, bufferp, in_port);
}
+
+/* Returns true if 'ofconn' has any pending opgroups. */
+bool
+ofconn_has_pending_opgroups(const struct ofconn *ofconn)
+{
+ return !list_is_empty(&ofconn->opgroups);
+}
+
+/* Returns the number of pending opgroups on 'ofconn'. */
+size_t
+ofconn_n_pending_opgroups(const struct ofconn *ofconn)
+{
+ return list_size(&ofconn->opgroups);
+}
+
+/* Adds 'ofconn_node' to 'ofconn''s list of pending opgroups.
+ *
+ * If 'ofconn' is destroyed or its connection drops, then 'ofconn' will remove
+ * 'ofconn_node' from the list and re-initialize it with list_init(). The
+ * client may, therefore, use list_is_empty(ofconn_node) to determine whether
+ * 'ofconn_node' is still associated with an active ofconn.
+ *
+ * The client may also remove ofconn_node from the list itself, with
+ * list_remove(). */
+void
+ofconn_add_opgroup(struct ofconn *ofconn, struct list *ofconn_node)
+{
+ list_push_back(&ofconn->opgroups, ofconn_node);
+}
\f
/* Private ofconn functions. */
ofconn->type = type;
ofconn->flow_format = NXFF_OPENFLOW10;
ofconn->flow_mod_table_id = false;
+ list_init(&ofconn->opgroups);
ofconn->role = NX_ROLE_OTHER;
ofconn->packet_in_counter = rconn_packet_counter_create ();
ofconn->pktbuf = NULL;
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'. */
+static void
+ofconn_flush(struct ofconn *ofconn)
+{
+ while (!list_is_empty(&ofconn->opgroups)) {
+ list_init(list_pop_front(&ofconn->opgroups));
+ }
+ ofpbuf_delete(ofconn->blocked);
+ ofconn->blocked = NULL;
+}
+
static void
ofconn_destroy(struct ofconn *ofconn)
{
+ ofconn_flush(ofconn);
+
if (ofconn->type == OFCONN_PRIMARY) {
hmap_remove(&ofconn->connmgr->controllers, &ofconn->hmap_node);
}
ofconn_set_rate_limit(ofconn, c->rate_limit, c->burst_limit);
}
+/* Returns true if it makes sense for 'ofconn' to receive and process OpenFlow
+ * messages. */
+static bool
+ofconn_may_recv(const struct ofconn *ofconn)
+{
+ int count = rconn_packet_counter_read (ofconn->reply_counter);
+ return (!ofconn->blocked || ofconn->retry) && count < OFCONN_REPLY_MAX;
+}
+
static void
ofconn_run(struct ofconn *ofconn,
- void (*handle_openflow)(struct ofconn *, struct ofpbuf *ofp_msg))
+ bool (*handle_openflow)(struct ofconn *, struct ofpbuf *ofp_msg))
{
struct connmgr *mgr = ofconn->connmgr;
- int iteration;
size_t i;
for (i = 0; i < N_SCHEDULERS; i++) {
rconn_run(ofconn->rconn);
- if (rconn_packet_counter_read (ofconn->reply_counter) < OFCONN_REPLY_MAX) {
- /* Limit the number of iterations to prevent other tasks from
- * starving. */
- for (iteration = 0; iteration < 50; iteration++) {
- struct ofpbuf *of_msg = rconn_recv(ofconn->rconn);
+ if (handle_openflow) {
+ /* Limit the number of iterations to avoid starving other tasks. */
+ for (i = 0; i < 50 && ofconn_may_recv(ofconn); i++) {
+ struct ofpbuf *of_msg;
+
+ of_msg = (ofconn->blocked
+ ? ofconn->blocked
+ : rconn_recv(ofconn->rconn));
if (!of_msg) {
break;
}
if (mgr->fail_open) {
fail_open_maybe_recover(mgr->fail_open);
}
- handle_openflow(ofconn, of_msg);
- ofpbuf_delete(of_msg);
+
+ if (handle_openflow(ofconn, of_msg)) {
+ ofpbuf_delete(of_msg);
+ ofconn->blocked = NULL;
+ } else {
+ ofconn->blocked = of_msg;
+ ofconn->retry = false;
+ }
}
}
if (!rconn_is_alive(ofconn->rconn)) {
ofconn_destroy(ofconn);
+ } else if (!rconn_is_connected(ofconn->rconn)) {
+ ofconn_flush(ofconn);
}
}
static void
-ofconn_wait(struct ofconn *ofconn)
+ofconn_wait(struct ofconn *ofconn, bool handling_openflow)
{
int i;
pinsched_wait(ofconn->schedulers[i]);
}
rconn_run_wait(ofconn->rconn);
- if (rconn_packet_counter_read (ofconn->reply_counter) < OFCONN_REPLY_MAX) {
+ if (handling_openflow && ofconn_may_recv(ofconn)) {
rconn_recv_wait(ofconn->rconn);
- } else {
- COVERAGE_INC(ofconn_stuck);
}
}
/* Fail-open and in-band implementation. */
/* Called by 'ofproto' after all flows have been flushed, to allow fail-open
- * and in-band control to re-create their flows. */
+ * and standalone mode to re-create their flows.
+ *
+ * In-band control has more sophisticated code that manages flows itself. */
void
connmgr_flushed(struct connmgr *mgr)
{
- if (mgr->in_band) {
- in_band_flushed(mgr->in_band);
- }
if (mgr->fail_open) {
fail_open_flushed(mgr->fail_open);
}
#include "openvswitch/types.h"
struct ofconn;
+struct ofopgroup;
struct ofputil_flow_removed;
struct ofputil_packet_in;
struct sset;
void connmgr_destroy(struct connmgr *);
void connmgr_run(struct connmgr *,
- void (*handle_openflow)(struct ofconn *,
+ bool (*handle_openflow)(struct ofconn *,
struct ofpbuf *ofp_msg));
-void connmgr_wait(struct connmgr *);
+void connmgr_wait(struct connmgr *, bool handling_openflow);
struct ofproto *ofconn_get_ofproto(const struct ofconn *);
+void connmgr_retry(struct connmgr *);
+
/* OpenFlow configuration. */
bool connmgr_has_controllers(const struct connmgr *);
void connmgr_get_controller_info(struct connmgr *, struct shash *);
int ofconn_pktbuf_retrieve(struct ofconn *, uint32_t id,
struct ofpbuf **bufferp, uint16_t *in_port);
+size_t ofconn_n_pending_opgroups(const struct ofconn *);
+bool ofconn_has_pending_opgroups(const struct ofconn *);
+void ofconn_add_opgroup(struct ofconn *, struct list *);
+void ofconn_remove_opgroup(struct ofconn *, struct list *,
+ const struct ofp_header *request, int error);
+
/* Sending asynchronous messages. */
void connmgr_send_port_status(struct connmgr *, const struct ofp_phy_port *,
uint8_t reason);
struct netdev *remote_netdev; /* Device to send to next-hop MAC. */
};
+/* What to do to an in_band_rule. */
+enum in_band_op {
+ ADD, /* Add the rule to ofproto's flow table. */
+ DELETE /* Delete the rule from ofproto's flow table. */
+};
+
+/* A rule to add to or delete from ofproto's flow table. */
+struct in_band_rule {
+ struct cls_rule cls_rule;
+ enum in_band_op op;
+};
+
struct in_band {
struct ofproto *ofproto;
- int queue_id, prev_queue_id;
+ int queue_id;
/* Remote information. */
time_t next_remote_refresh; /* Refresh timer. */
uint8_t local_mac[ETH_ADDR_LEN]; /* Current MAC. */
struct netdev *local_netdev; /* Local port's network device. */
- /* Local and remote addresses that are installed as flows. */
- uint8_t installed_local_mac[ETH_ADDR_LEN];
- struct sockaddr_in *remote_addrs;
- size_t n_remote_addrs;
- uint8_t *remote_macs;
- size_t n_remote_macs;
+ /* Flow tracking. */
+ struct hmap rules; /* Contains "struct in_band_rule"s. */
};
static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(60, 60);
}
static void
-make_rules(struct in_band *ib,
- void (*cb)(struct in_band *, const struct cls_rule *))
+add_rule(struct in_band *ib, const struct cls_rule *cls_rule)
{
+ uint32_t hash = cls_rule_hash(cls_rule, 0);
+ struct in_band_rule *rule;
+
+ HMAP_FOR_EACH_WITH_HASH (rule, cls_rule.hmap_node, hash, &ib->rules) {
+ if (cls_rule_equal(&rule->cls_rule, cls_rule)) {
+ rule->op = ADD;
+ return;
+ }
+ }
+
+ rule = xmalloc(sizeof *rule);
+ rule->cls_rule = *cls_rule;
+ rule->op = ADD;
+ hmap_insert(&ib->rules, &rule->cls_rule.hmap_node, hash);
+}
+
+static void
+update_rules(struct in_band *ib)
+{
+ struct in_band_rule *ib_rule;
+ struct in_band_remote *r;
struct cls_rule rule;
- size_t i;
- if (!eth_addr_is_zero(ib->installed_local_mac)) {
+ /* Mark all the existing rules for deletion. (Afterward we will re-add any
+ * rules that are still valid.) */
+ HMAP_FOR_EACH (ib_rule, cls_rule.hmap_node, &ib->rules) {
+ ib_rule->op = DELETE;
+ }
+
+ if (!eth_addr_is_zero(ib->local_mac)) {
/* (a) Allow DHCP requests sent from the local port. */
cls_rule_init_catchall(&rule, IBR_FROM_LOCAL_DHCP);
cls_rule_set_in_port(&rule, ODPP_LOCAL);
cls_rule_set_dl_type(&rule, htons(ETH_TYPE_IP));
- cls_rule_set_dl_src(&rule, ib->installed_local_mac);
+ cls_rule_set_dl_src(&rule, ib->local_mac);
cls_rule_set_nw_proto(&rule, IPPROTO_UDP);
cls_rule_set_tp_src(&rule, htons(DHCP_CLIENT_PORT));
cls_rule_set_tp_dst(&rule, htons(DHCP_SERVER_PORT));
- cb(ib, &rule);
+ add_rule(ib, &rule);
/* (b) Allow ARP replies to the local port's MAC address. */
cls_rule_init_catchall(&rule, IBR_TO_LOCAL_ARP);
cls_rule_set_dl_type(&rule, htons(ETH_TYPE_ARP));
- cls_rule_set_dl_dst(&rule, ib->installed_local_mac);
+ cls_rule_set_dl_dst(&rule, ib->local_mac);
cls_rule_set_nw_proto(&rule, ARP_OP_REPLY);
- cb(ib, &rule);
+ add_rule(ib, &rule);
/* (c) Allow ARP requests from the local port's MAC address. */
cls_rule_init_catchall(&rule, IBR_FROM_LOCAL_ARP);
cls_rule_set_dl_type(&rule, htons(ETH_TYPE_ARP));
- cls_rule_set_dl_src(&rule, ib->installed_local_mac);
+ cls_rule_set_dl_src(&rule, ib->local_mac);
cls_rule_set_nw_proto(&rule, ARP_OP_REQUEST);
- cb(ib, &rule);
+ add_rule(ib, &rule);
}
- for (i = 0; i < ib->n_remote_macs; i++) {
- const uint8_t *remote_mac = &ib->remote_macs[i * ETH_ADDR_LEN];
+ for (r = ib->remotes; r < &ib->remotes[ib->n_remotes]; r++) {
+ const uint8_t *remote_mac = r->remote_mac;
- if (i > 0) {
- const uint8_t *prev_mac = &ib->remote_macs[(i - 1) * ETH_ADDR_LEN];
- if (eth_addr_equals(remote_mac, prev_mac)) {
- /* Skip duplicates. */
- continue;
- }
+ if (eth_addr_is_zero(remote_mac)) {
+ continue;
}
/* (d) Allow ARP replies to the next hop's MAC address. */
cls_rule_set_dl_type(&rule, htons(ETH_TYPE_ARP));
cls_rule_set_dl_dst(&rule, remote_mac);
cls_rule_set_nw_proto(&rule, ARP_OP_REPLY);
- cb(ib, &rule);
+ add_rule(ib, &rule);
/* (e) Allow ARP requests from the next hop's MAC address. */
cls_rule_init_catchall(&rule, IBR_FROM_NEXT_HOP_ARP);
cls_rule_set_dl_type(&rule, htons(ETH_TYPE_ARP));
cls_rule_set_dl_src(&rule, remote_mac);
cls_rule_set_nw_proto(&rule, ARP_OP_REQUEST);
- cb(ib, &rule);
+ add_rule(ib, &rule);
}
- for (i = 0; i < ib->n_remote_addrs; i++) {
- const struct sockaddr_in *a = &ib->remote_addrs[i];
-
- if (!i || a->sin_addr.s_addr != a[-1].sin_addr.s_addr) {
- /* (f) Allow ARP replies containing the remote's IP address as a
- * target. */
- cls_rule_init_catchall(&rule, IBR_TO_REMOTE_ARP);
- cls_rule_set_dl_type(&rule, htons(ETH_TYPE_ARP));
- cls_rule_set_nw_proto(&rule, ARP_OP_REPLY);
- cls_rule_set_nw_dst(&rule, a->sin_addr.s_addr);
- cb(ib, &rule);
-
- /* (g) Allow ARP requests containing the remote's IP address as a
- * source. */
- cls_rule_init_catchall(&rule, IBR_FROM_REMOTE_ARP);
- cls_rule_set_dl_type(&rule, htons(ETH_TYPE_ARP));
- cls_rule_set_nw_proto(&rule, ARP_OP_REQUEST);
- cls_rule_set_nw_src(&rule, a->sin_addr.s_addr);
- cb(ib, &rule);
- }
-
- if (!i
- || a->sin_addr.s_addr != a[-1].sin_addr.s_addr
- || a->sin_port != a[-1].sin_port) {
- /* (h) Allow TCP traffic to the remote's IP and port. */
- cls_rule_init_catchall(&rule, IBR_TO_REMOTE_TCP);
- cls_rule_set_dl_type(&rule, htons(ETH_TYPE_IP));
- cls_rule_set_nw_proto(&rule, IPPROTO_TCP);
- cls_rule_set_nw_dst(&rule, a->sin_addr.s_addr);
- cls_rule_set_tp_dst(&rule, a->sin_port);
- cb(ib, &rule);
-
- /* (i) Allow TCP traffic from the remote's IP and port. */
- cls_rule_init_catchall(&rule, IBR_FROM_REMOTE_TCP);
- cls_rule_set_dl_type(&rule, htons(ETH_TYPE_IP));
- cls_rule_set_nw_proto(&rule, IPPROTO_TCP);
- cls_rule_set_nw_src(&rule, a->sin_addr.s_addr);
- cls_rule_set_tp_src(&rule, a->sin_port);
- cb(ib, &rule);
- }
- }
-}
-
-static void
-drop_rule(struct in_band *ib, const struct cls_rule *rule)
-{
- ofproto_delete_flow(ib->ofproto, rule);
-}
+ for (r = ib->remotes; r < &ib->remotes[ib->n_remotes]; r++) {
+ const struct sockaddr_in *a = &r->remote_addr;
-/* Drops from the flow table all of the flows set up by 'ib', then clears out
- * the information about the installed flows so that they can be filled in
- * again if necessary. */
-static void
-drop_rules(struct in_band *ib)
-{
- /* Drop rules. */
- make_rules(ib, drop_rule);
+ /* (f) Allow ARP replies containing the remote's IP address as a
+ * target. */
+ cls_rule_init_catchall(&rule, IBR_TO_REMOTE_ARP);
+ cls_rule_set_dl_type(&rule, htons(ETH_TYPE_ARP));
+ cls_rule_set_nw_proto(&rule, ARP_OP_REPLY);
+ cls_rule_set_nw_dst(&rule, a->sin_addr.s_addr);
+ add_rule(ib, &rule);
- /* Clear out state. */
- memset(ib->installed_local_mac, 0, sizeof ib->installed_local_mac);
+ /* (g) Allow ARP requests containing the remote's IP address as a
+ * source. */
+ cls_rule_init_catchall(&rule, IBR_FROM_REMOTE_ARP);
+ cls_rule_set_dl_type(&rule, htons(ETH_TYPE_ARP));
+ cls_rule_set_nw_proto(&rule, ARP_OP_REQUEST);
+ cls_rule_set_nw_src(&rule, a->sin_addr.s_addr);
+ add_rule(ib, &rule);
- free(ib->remote_addrs);
- ib->remote_addrs = NULL;
- ib->n_remote_addrs = 0;
+ /* (h) Allow TCP traffic to the remote's IP and port. */
+ cls_rule_init_catchall(&rule, IBR_TO_REMOTE_TCP);
+ cls_rule_set_dl_type(&rule, htons(ETH_TYPE_IP));
+ cls_rule_set_nw_proto(&rule, IPPROTO_TCP);
+ cls_rule_set_nw_dst(&rule, a->sin_addr.s_addr);
+ cls_rule_set_tp_dst(&rule, a->sin_port);
+ add_rule(ib, &rule);
- free(ib->remote_macs);
- ib->remote_macs = NULL;
- ib->n_remote_macs = 0;
+ /* (i) Allow TCP traffic from the remote's IP and port. */
+ cls_rule_init_catchall(&rule, IBR_FROM_REMOTE_TCP);
+ cls_rule_set_dl_type(&rule, htons(ETH_TYPE_IP));
+ cls_rule_set_nw_proto(&rule, IPPROTO_TCP);
+ cls_rule_set_nw_src(&rule, a->sin_addr.s_addr);
+ cls_rule_set_tp_src(&rule, a->sin_port);
+ add_rule(ib, &rule);
+ }
}
-static void
-add_rule(struct in_band *ib, const struct cls_rule *rule)
+void
+in_band_run(struct in_band *ib)
{
struct {
struct nx_action_set_queue nxsq;
union ofp_action oa;
} actions;
+ const void *a;
+ size_t na;
- memset(&actions, 0, sizeof actions);
+ struct in_band_rule *rule, *next;
+ memset(&actions, 0, sizeof actions);
actions.oa.output.type = htons(OFPAT_OUTPUT);
actions.oa.output.len = htons(sizeof actions.oa);
actions.oa.output.port = htons(OFPP_NORMAL);
actions.oa.output.max_len = htons(0);
-
if (ib->queue_id < 0) {
- ofproto_add_flow(ib->ofproto, rule, &actions.oa, 1);
+ a = &actions.oa;
+ na = sizeof actions.oa / sizeof(union ofp_action);
} else {
actions.nxsq.type = htons(OFPAT_VENDOR);
actions.nxsq.len = htons(sizeof actions.nxsq);
actions.nxsq.vendor = htonl(NX_VENDOR_ID);
actions.nxsq.subtype = htons(NXAST_SET_QUEUE);
actions.nxsq.queue_id = htonl(ib->queue_id);
-
- ofproto_add_flow(ib->ofproto, rule, (union ofp_action *) &actions,
- sizeof actions / sizeof(union ofp_action));
+ a = &actions;
+ na = sizeof actions / sizeof(union ofp_action);
}
-}
-/* Inserts flows into the flow table for the current state of 'ib'. */
-static void
-add_rules(struct in_band *ib)
-{
- make_rules(ib, add_rule);
-}
+ refresh_local(ib);
+ refresh_remotes(ib);
-static int
-compare_addrs(const void *a_, const void *b_)
-{
- const struct sockaddr_in *a = a_;
- const struct sockaddr_in *b = b_;
- int cmp;
-
- cmp = memcmp(&a->sin_addr.s_addr,
- &b->sin_addr.s_addr,
- sizeof a->sin_addr.s_addr);
- if (cmp) {
- return cmp;
- }
- return memcmp(&a->sin_port, &b->sin_port, sizeof a->sin_port);
-}
+ update_rules(ib);
-static int
-compare_macs(const void *a, const void *b)
-{
- return eth_addr_compare_3way(a, b);
-}
+ HMAP_FOR_EACH_SAFE (rule, next, cls_rule.hmap_node, &ib->rules) {
+ switch (rule->op) {
+ case ADD:
+ ofproto_add_flow(ib->ofproto, &rule->cls_rule, a, na);
+ break;
-void
-in_band_run(struct in_band *ib)
-{
- bool local_change, remote_change, queue_id_change;
- struct in_band_remote *r;
-
- local_change = refresh_local(ib);
- remote_change = refresh_remotes(ib);
- queue_id_change = ib->queue_id != ib->prev_queue_id;
- if (!local_change && !remote_change && !queue_id_change) {
- /* Nothing changed, nothing to do. */
- return;
- }
- ib->prev_queue_id = ib->queue_id;
-
- /* Drop old rules. */
- drop_rules(ib);
-
- /* Figure out new rules. */
- memcpy(ib->installed_local_mac, ib->local_mac, ETH_ADDR_LEN);
- ib->remote_addrs = xmalloc(ib->n_remotes * sizeof *ib->remote_addrs);
- ib->n_remote_addrs = 0;
- ib->remote_macs = xmalloc(ib->n_remotes * ETH_ADDR_LEN);
- ib->n_remote_macs = 0;
- for (r = ib->remotes; r < &ib->remotes[ib->n_remotes]; r++) {
- ib->remote_addrs[ib->n_remote_addrs++] = r->remote_addr;
- if (!eth_addr_is_zero(r->remote_mac)) {
- memcpy(&ib->remote_macs[ib->n_remote_macs * ETH_ADDR_LEN],
- r->remote_mac, ETH_ADDR_LEN);
- ib->n_remote_macs++;
+ case DELETE:
+ if (ofproto_delete_flow(ib->ofproto, &rule->cls_rule)) {
+ /* ofproto doesn't have the rule anymore so there's no reason
+ * for us to track it any longer. */
+ hmap_remove(&ib->rules, &rule->cls_rule.hmap_node);
+ free(rule);
+ }
+ break;
}
}
-
- /* Sort, to allow make_rules() to easily skip duplicates. */
- qsort(ib->remote_addrs, ib->n_remote_addrs, sizeof *ib->remote_addrs,
- compare_addrs);
- qsort(ib->remote_macs, ib->n_remote_macs, ETH_ADDR_LEN, compare_macs);
-
- /* Add new rules. */
- add_rules(ib);
}
void
poll_timer_wait_until(wakeup * 1000);
}
-/* ofproto has flushed all flows from the flow table and it is calling us back
- * to allow us to reinstall the ones that are important to us. */
-void
-in_band_flushed(struct in_band *in_band)
-{
- add_rules(in_band);
-}
-
int
in_band_create(struct ofproto *ofproto, const char *local_name,
struct in_band **in_bandp)
in_band = xzalloc(sizeof *in_band);
in_band->ofproto = ofproto;
- in_band->queue_id = in_band->prev_queue_id = -1;
+ in_band->queue_id = -1;
in_band->next_remote_refresh = TIME_MIN;
in_band->next_local_refresh = TIME_MIN;
in_band->local_netdev = local_netdev;
+ hmap_init(&in_band->rules);
*in_bandp = in_band;
in_band_destroy(struct in_band *ib)
{
if (ib) {
- drop_rules(ib);
+ struct in_band_rule *rule, *next;
+
+ HMAP_FOR_EACH_SAFE (rule, next, cls_rule.hmap_node, &ib->rules) {
+ hmap_remove(&ib->rules, &rule->cls_rule.hmap_node);
+ free(rule);
+ }
+ hmap_destroy(&ib->rules);
in_band_set_remotes(ib, NULL, 0);
netdev_close(ib->local_netdev);
free(ib);
const struct ofpbuf *packet);
bool in_band_rule_check(const struct flow *,
const struct nlattr *odp_actions, size_t actions_len);
-void in_band_flushed(struct in_band *);
#endif /* in-band.h */
static void port_wait(struct ofport_dpif *);
static int set_cfm(struct ofport *, const struct cfm_settings *);
+struct dpif_completion {
+ struct list list_node;
+ struct ofoperation *op;
+};
+
struct ofproto_dpif {
struct ofproto up;
struct dpif *dpif;
struct hmap facets;
bool need_revalidate;
struct tag_set revalidate_set;
+
+ /* Support for debugging async flow mods. */
+ struct list completions;
};
+/* Defer flow mod completion until "ovs-appctl ofproto/unclog"? (Useful only
+ * for debugging the asynchronous flow_mod implementation.) */
+static bool clogged;
+
static void ofproto_dpif_unixctl_init(void);
static struct ofproto_dpif *
ofproto->need_revalidate = false;
tag_set_init(&ofproto->revalidate_set);
+ list_init(&ofproto->completions);
+
ofproto->up.tables = xmalloc(sizeof *ofproto->up.tables);
classifier_init(&ofproto->up.tables[0]);
ofproto->up.n_tables = 1;
return 0;
}
+static void
+complete_operations(struct ofproto_dpif *ofproto)
+{
+ struct dpif_completion *c, *next;
+
+ LIST_FOR_EACH_SAFE (c, next, list_node, &ofproto->completions) {
+ ofoperation_complete(c->op, 0);
+ list_remove(&c->list_node);
+ free(c);
+ }
+}
+
static void
destruct(struct ofproto *ofproto_)
{
struct ofproto_dpif *ofproto = ofproto_dpif_cast(ofproto_);
+ struct rule_dpif *rule, *next_rule;
+ struct cls_cursor cursor;
int i;
+ complete_operations(ofproto);
+
+ cls_cursor_init(&cursor, &ofproto->up.tables[0], NULL);
+ CLS_CURSOR_FOR_EACH_SAFE (rule, next_rule, up.cr, &cursor) {
+ ofproto_rule_destroy(&rule->up);
+ }
+
for (i = 0; i < MAX_MIRRORS; i++) {
mirror_destroy(ofproto->mirrors[i]);
}
struct ofbundle *bundle;
int i;
+ if (!clogged) {
+ complete_operations(ofproto);
+ }
dpif_run(ofproto->dpif);
for (i = 0; i < 50; i++) {
struct ofport_dpif *ofport;
struct ofbundle *bundle;
+ if (!clogged && !list_is_empty(&ofproto->completions)) {
+ poll_immediate_wake();
+ }
+
dpif_wait(ofproto->dpif);
dpif_recv_wait(ofproto->dpif);
if (ofproto->sflow) {
flow)));
}
+static void
+complete_operation(struct rule_dpif *rule)
+{
+ struct ofproto_dpif *ofproto = ofproto_dpif_cast(rule->up.ofproto);
+
+ ofproto->need_revalidate = true;
+ if (clogged) {
+ struct dpif_completion *c = xmalloc(sizeof *c);
+ c->op = rule->up.pending;
+ list_push_back(&ofproto->completions, &c->list_node);
+ } else {
+ ofoperation_complete(rule->up.pending, 0);
+ }
+}
+
static struct rule *
rule_alloc(void)
{
{
struct rule_dpif *rule = rule_dpif_cast(rule_);
struct ofproto_dpif *ofproto = ofproto_dpif_cast(rule->up.ofproto);
- struct rule_dpif *old_rule;
+ struct rule_dpif *victim;
int error;
error = validate_actions(rule->up.actions, rule->up.n_actions,
return error;
}
- old_rule = rule_dpif_cast(rule_from_cls_rule(classifier_find_rule_exactly(
- &ofproto->up.tables[0],
- &rule->up.cr)));
- if (old_rule) {
- ofproto_rule_destroy(&old_rule->up);
- }
-
rule->used = rule->up.created;
rule->packet_count = 0;
rule->byte_count = 0;
- list_init(&rule->facets);
- classifier_insert(&ofproto->up.tables[0], &rule->up.cr);
- ofproto->need_revalidate = true;
+ victim = rule_dpif_cast(ofoperation_get_victim(rule->up.pending));
+ if (victim && !list_is_empty(&victim->facets)) {
+ struct facet *facet;
+
+ rule->facets = victim->facets;
+ list_moved(&rule->facets);
+ LIST_FOR_EACH (facet, list_node, &rule->facets) {
+ facet->rule = rule;
+ }
+ } else {
+ /* Must avoid list_moved() in this case. */
+ list_init(&rule->facets);
+ }
+ complete_operation(rule);
return 0;
}
struct ofproto_dpif *ofproto = ofproto_dpif_cast(rule->up.ofproto);
struct facet *facet, *next_facet;
- classifier_remove(&ofproto->up.tables[0], &rule->up.cr);
LIST_FOR_EACH_SAFE (facet, next_facet, list_node, &rule->facets) {
facet_revalidate(ofproto, facet);
}
- ofproto->need_revalidate = true;
+
+ complete_operation(rule);
}
static void
return 0;
}
-static int
-rule_modify_actions(struct rule *rule_,
- const union ofp_action *actions, size_t n_actions)
+static void
+rule_modify_actions(struct rule *rule_)
{
struct rule_dpif *rule = rule_dpif_cast(rule_);
struct ofproto_dpif *ofproto = ofproto_dpif_cast(rule->up.ofproto);
int error;
- error = validate_actions(actions, n_actions, &rule->up.cr.flow,
- ofproto->max_ports);
- if (!error) {
- ofproto->need_revalidate = true;
+ error = validate_actions(rule->up.actions, rule->up.n_actions,
+ &rule->up.cr.flow, ofproto->max_ports);
+ if (error) {
+ ofoperation_complete(rule->up.pending, error);
+ return;
}
- return error;
+
+ complete_operation(rule);
}
\f
/* Sends 'packet' out of port 'odp_port' within 'p'.
free(args);
}
+static void
+ofproto_dpif_clog(struct unixctl_conn *conn OVS_UNUSED,
+ const char *args_ OVS_UNUSED, void *aux OVS_UNUSED)
+{
+ clogged = true;
+ unixctl_command_reply(conn, 200, NULL);
+}
+
+static void
+ofproto_dpif_unclog(struct unixctl_conn *conn OVS_UNUSED,
+ const char *args_ OVS_UNUSED, void *aux OVS_UNUSED)
+{
+ clogged = false;
+ unixctl_command_reply(conn, 200, NULL);
+}
+
static void
ofproto_dpif_unixctl_init(void)
{
unixctl_command_register("ofproto/trace", ofproto_unixctl_trace, NULL);
unixctl_command_register("fdb/show", ofproto_unixctl_fdb_show, NULL);
+
+ unixctl_command_register("ofproto/clog", ofproto_dpif_clog, NULL);
+ unixctl_command_register("ofproto/unclog", ofproto_dpif_unclog, NULL);
}
\f
const struct ofproto_class ofproto_dpif_class = {
COVERAGE_DEFINE(ofproto_uninstallable);
COVERAGE_DEFINE(ofproto_update_port);
+enum ofproto_state {
+ S_OPENFLOW, /* Processing OpenFlow commands. */
+ S_FLUSH, /* Deleting all flow table rules. */
+};
+
+enum ofoperation_type {
+ OFOPERATION_ADD,
+ OFOPERATION_DELETE,
+ OFOPERATION_MODIFY
+};
+
+/* A single OpenFlow request can execute any number of operations. The
+ * ofopgroup maintain OpenFlow state common to all of the operations, e.g. the
+ * ofconn to which an error reply should be sent if necessary.
+ *
+ * ofproto initiates some operations internally. These operations are still
+ * assigned to groups but will not have an associated ofconn. */
+struct ofopgroup {
+ struct ofproto *ofproto; /* Owning ofproto. */
+ struct list ofproto_node; /* In ofproto's "pending" list. */
+ struct list ops; /* List of "struct ofoperation"s. */
+
+ /* Data needed to send OpenFlow reply on failure or to send a buffered
+ * packet on success.
+ *
+ * If list_is_empty(ofconn_node) then this ofopgroup never had an
+ * associated ofconn or its ofconn's connection dropped after it initiated
+ * the operation. In the latter case 'ofconn' is a wild pointer that
+ * refers to freed memory, so the 'ofconn' member must be used only if
+ * !list_is_empty(ofconn_node).
+ */
+ struct list ofconn_node; /* In ofconn's list of pending opgroups. */
+ struct ofconn *ofconn; /* ofconn for reply (but see note above). */
+ struct ofp_header *request; /* Original request (truncated at 64 bytes). */
+ uint32_t buffer_id; /* Buffer id from original request. */
+ int error; /* 0 if no error yet, otherwise error code. */
+};
+
+static struct ofopgroup *ofopgroup_create(struct ofproto *);
+static struct ofopgroup *ofopgroup_create_for_ofconn(struct ofconn *,
+ const struct ofp_header *,
+ uint32_t buffer_id);
+static void ofopgroup_submit(struct ofopgroup *);
+static void ofopgroup_destroy(struct ofopgroup *);
+
+/* A single flow table operation. */
+struct ofoperation {
+ struct ofopgroup *group; /* Owning group. */
+ struct list group_node; /* In ofopgroup's "ops" list. */
+ struct hmap_node hmap_node; /* In ofproto's "deletions" hmap. */
+ struct rule *rule; /* Rule being operated upon. */
+ enum ofoperation_type type; /* Type of operation. */
+ int status; /* -1 if pending, otherwise 0 or error code. */
+ struct rule *victim; /* OFOPERATION_ADDING: Replaced rule. */
+ union ofp_action *actions; /* OFOPERATION_MODIFYING: Replaced actions. */
+ int n_actions; /* OFOPERATION_MODIFYING: # of old actions. */
+ ovs_be64 flow_cookie; /* Rule's old flow cookie. */
+};
+
+static void ofoperation_create(struct ofopgroup *, struct rule *,
+ enum ofoperation_type);
+static void ofoperation_destroy(struct ofoperation *);
+
static void ofport_destroy__(struct ofport *);
static void ofport_destroy(struct ofport *);
-static int rule_create(struct ofproto *,
- const struct cls_rule *, uint8_t table_id,
- const union ofp_action *, size_t n_actions,
- uint16_t idle_timeout, uint16_t hard_timeout,
- ovs_be64 flow_cookie, bool send_flow_removed,
- struct rule **rulep);
-
static uint64_t pick_datapath_id(const struct ofproto *);
static uint64_t pick_fallback_dpid(void);
static void ofproto_destroy__(struct ofproto *);
-static void ofproto_flush_flows__(struct ofproto *);
static void ofproto_rule_destroy__(struct rule *);
static void ofproto_rule_send_removed(struct rule *, uint8_t reason);
-static void handle_openflow(struct ofconn *, struct ofpbuf *);
+static void ofopgroup_destroy(struct ofopgroup *);
+
+static int add_flow(struct ofproto *, struct ofconn *, struct flow_mod *,
+ const struct ofp_header *);
+
+/* This return value tells handle_openflow() that processing of the current
+ * OpenFlow message must be postponed until some ongoing operations have
+ * completed.
+ *
+ * This particular value is a good choice because it is negative (so it won't
+ * collide with any errno value or any value returned by ofp_mkerr()) and large
+ * (so it won't accidentally collide with EOF or a negative errno value). */
+enum { OFPROTO_POSTPONE = -100000 };
+
+static bool handle_openflow(struct ofconn *, struct ofpbuf *);
static void update_port(struct ofproto *, const char *devname);
static int init_ports(struct ofproto *);
ofproto->tables = NULL;
ofproto->n_tables = 0;
ofproto->connmgr = connmgr_create(ofproto, datapath_name, datapath_name);
+ ofproto->state = S_OPENFLOW;
+ list_init(&ofproto->pending);
+ hmap_init(&ofproto->deletions);
error = ofproto->ofproto_class->construct(ofproto);
if (error) {
connmgr_get_snoops(ofproto->connmgr, snoops);
}
+static void
+ofproto_flush__(struct ofproto *ofproto)
+{
+ struct classifier *table;
+ struct ofopgroup *group;
+
+ if (ofproto->ofproto_class->flush) {
+ ofproto->ofproto_class->flush(ofproto);
+ }
+
+ group = ofopgroup_create(ofproto);
+ for (table = ofproto->tables; table < &ofproto->tables[ofproto->n_tables];
+ table++) {
+ struct rule *rule, *next_rule;
+ struct cls_cursor cursor;
+
+ cls_cursor_init(&cursor, table, NULL);
+ CLS_CURSOR_FOR_EACH_SAFE (rule, next_rule, cr, &cursor) {
+ if (!rule->pending) {
+ ofoperation_create(group, rule, OFOPERATION_DELETE);
+ classifier_remove(table, &rule->cr);
+ ofproto->ofproto_class->rule_destruct(rule);
+ }
+ }
+ }
+ ofopgroup_submit(group);
+}
+
static void
ofproto_destroy__(struct ofproto *ofproto)
{
size_t i;
+ assert(list_is_empty(&ofproto->pending));
+
connmgr_destroy(ofproto->connmgr);
hmap_remove(&all_ofprotos, &ofproto->hmap_node);
shash_destroy(&ofproto->port_by_name);
for (i = 0; i < ofproto->n_tables; i++) {
+ assert(classifier_is_empty(&ofproto->tables[i]));
classifier_destroy(&ofproto->tables[i]);
}
free(ofproto->tables);
+ hmap_destroy(&ofproto->deletions);
+
ofproto->ofproto_class->dealloc(ofproto);
}
return;
}
- ofproto_flush_flows__(p);
+ ofproto_flush__(p);
HMAP_FOR_EACH_SAFE (ofport, next_ofport, hmap_node, &p->ports) {
ofport_destroy(ofport);
}
}
}
- connmgr_run(p->connmgr, handle_openflow);
+
+ switch (p->state) {
+ case S_OPENFLOW:
+ connmgr_run(p->connmgr, handle_openflow);
+ break;
+
+ case S_FLUSH:
+ connmgr_run(p->connmgr, NULL);
+ ofproto_flush__(p);
+ if (list_is_empty(&p->pending) && hmap_is_empty(&p->deletions)) {
+ connmgr_flushed(p->connmgr);
+ p->state = S_OPENFLOW;
+ }
+ break;
+
+ default:
+ NOT_REACHED();
+ }
return 0;
}
poll_immediate_wake();
}
}
- connmgr_wait(p->connmgr);
+
+ switch (p->state) {
+ case S_OPENFLOW:
+ connmgr_wait(p->connmgr, true);
+ break;
+
+ case S_FLUSH:
+ connmgr_wait(p->connmgr, false);
+ if (list_is_empty(&p->pending) && hmap_is_empty(&p->deletions)) {
+ poll_immediate_wake();
+ }
+ break;
+ }
}
bool
*
* This is a helper function for in-band control and fail-open. */
void
-ofproto_add_flow(struct ofproto *p, const struct cls_rule *cls_rule,
+ofproto_add_flow(struct ofproto *ofproto, const struct cls_rule *cls_rule,
const union ofp_action *actions, size_t n_actions)
{
- struct rule *rule;
- rule_create(p, cls_rule, 0, actions, n_actions, 0, 0, 0, false, &rule);
+ const struct rule *rule;
+
+ rule = rule_from_cls_rule(classifier_find_rule_exactly(
+ &ofproto->tables[0], cls_rule));
+ if (!rule || !ofputil_actions_equal(rule->actions, rule->n_actions,
+ actions, n_actions)) {
+ struct flow_mod fm;
+
+ memset(&fm, 0, sizeof fm);
+ fm.cr = *cls_rule;
+ fm.buffer_id = UINT32_MAX;
+ fm.actions = (union ofp_action *) actions;
+ fm.n_actions = n_actions;
+ add_flow(ofproto, NULL, &fm, NULL);
+ }
}
/* Searches for a rule with matching criteria exactly equal to 'target' in
* ofproto's table 0 and, if it finds one, deletes it.
*
* This is a helper function for in-band control and fail-open. */
-void
+bool
ofproto_delete_flow(struct ofproto *ofproto, const struct cls_rule *target)
{
struct rule *rule;
rule = rule_from_cls_rule(classifier_find_rule_exactly(
&ofproto->tables[0], target));
- ofproto_rule_destroy(rule);
-}
-
-static void
-ofproto_flush_flows__(struct ofproto *ofproto)
-{
- size_t i;
-
- COVERAGE_INC(ofproto_flush);
-
- if (ofproto->ofproto_class->flush) {
- ofproto->ofproto_class->flush(ofproto);
+ if (!rule) {
+ /* No such rule -> success. */
+ return true;
+ } else if (rule->pending) {
+ /* An operation on the rule is already pending -> failure.
+ * Caller must retry later if it's important. */
+ return false;
+ } else {
+ /* Initiate deletion -> success. */
+ struct ofopgroup *group = ofopgroup_create(ofproto);
+ ofoperation_create(group, rule, OFOPERATION_DELETE);
+ classifier_remove(&ofproto->tables[rule->table_id], &rule->cr);
+ rule->ofproto->ofproto_class->rule_destruct(rule);
+ ofopgroup_submit(group);
+ return true;
}
- for (i = 0; i < ofproto->n_tables; i++) {
- struct rule *rule, *next_rule;
- struct cls_cursor cursor;
-
- cls_cursor_init(&cursor, &ofproto->tables[i], NULL);
- CLS_CURSOR_FOR_EACH_SAFE (rule, next_rule, cr, &cursor) {
- ofproto_rule_destroy(rule);
- }
- }
}
-/* Deletes all of the flows from all of ofproto's flow tables, then
- * reintroduces rules required by in-band control and fail open. */
+/* Starts the process of deleting all of the flows from all of ofproto's flow
+ * tables and then reintroducing the flows required by in-band control and
+ * fail-open. The process will complete in a later call to ofproto_run(). */
void
ofproto_flush_flows(struct ofproto *ofproto)
{
- ofproto_flush_flows__(ofproto);
- connmgr_flushed(ofproto->connmgr);
+ COVERAGE_INC(ofproto_flush);
+ ofproto->state = S_FLUSH;
}
\f
static void
return 0;
}
\f
-/* Creates a new rule initialized as specified, inserts it into 'ofproto''s
- * flow table, and stores the new rule into '*rulep'. Returns 0 on success,
- * otherwise a positive errno value or OpenFlow error code. */
-static int
-rule_create(struct ofproto *ofproto,
- const struct cls_rule *cls_rule, uint8_t table_id,
- const union ofp_action *actions, size_t n_actions,
- uint16_t idle_timeout, uint16_t hard_timeout,
- ovs_be64 flow_cookie, bool send_flow_removed,
- struct rule **rulep)
-{
- struct rule *rule;
- int error;
-
- if (table_id == 0xff) {
- if (ofproto->n_tables > 1) {
- error = ofproto->ofproto_class->rule_choose_table(ofproto,
- cls_rule,
- &table_id);
- if (error) {
- return error;
- }
- assert(table_id < ofproto->n_tables);
- } else {
- table_id = 0;
- }
- }
-
- rule = ofproto->ofproto_class->rule_alloc();
- if (!rule) {
- error = ENOMEM;
- goto error;
- }
-
- rule->ofproto = ofproto;
- rule->cr = *cls_rule;
- rule->table_id = table_id;
- rule->flow_cookie = flow_cookie;
- rule->created = time_msec();
- rule->idle_timeout = idle_timeout;
- rule->hard_timeout = hard_timeout;
- rule->send_flow_removed = send_flow_removed;
- if (n_actions > 0) {
- rule->actions = xmemdup(actions, n_actions * sizeof *actions);
- } else {
- rule->actions = NULL;
- }
- rule->n_actions = n_actions;
-
- error = ofproto->ofproto_class->rule_construct(rule);
- if (error) {
- ofproto_rule_destroy__(rule);
- goto error;
- }
-
- *rulep = rule;
- return 0;
-
-error:
- VLOG_WARN_RL(&rl, "%s: failed to create rule (%s)",
- ofproto->name, strerror(error));
- *rulep = NULL;
- return error;
-}
-
static void
ofproto_rule_destroy__(struct rule *rule)
{
rule->ofproto->ofproto_class->rule_dealloc(rule);
}
-/* Destroys 'rule' and removes it from the flow table and the datapath. */
+/* This function allows an ofproto implementation to destroy any rules that
+ * remain when its ->destruct() function is called. The caller must have
+ * already uninitialized any derived members of 'rule' (step 5 described in the
+ * large comment in ofproto/private.h titled "Life Cycle"). This function
+ * implements steps 6 and 7.
+ *
+ * This function should only be called from an ofproto implementation's
+ * ->destruct() function. It is not suitable elsewhere. */
void
ofproto_rule_destroy(struct rule *rule)
{
- if (rule) {
- rule->ofproto->ofproto_class->rule_destruct(rule);
- ofproto_rule_destroy__(rule);
- }
+ assert(!rule->pending);
+ classifier_remove(&rule->ofproto->tables[rule->table_id], &rule->cr);
+ ofproto_rule_destroy__(rule);
}
/* Returns true if 'rule' has an OpenFlow OFPAT_OUTPUT or OFPAT_ENQUEUE action
cls_cursor_init(&cursor, cls, match);
CLS_CURSOR_FOR_EACH (rule, cr, &cursor) {
+ if (rule->pending) {
+ return OFPROTO_POSTPONE;
+ }
if (!rule_is_hidden(rule) && rule_has_out_port(rule, out_port)) {
list_push_back(rules, &rule->ofproto_node);
}
rule = rule_from_cls_rule(classifier_find_rule_exactly(cls, match));
if (rule) {
+ if (rule->pending) {
+ return OFPROTO_POSTPONE;
+ }
if (!rule_is_hidden(rule) && rule_has_out_port(rule, out_port)) {
list_push_back(rules, &rule->ofproto_node);
}
return 0;
}
-\f
+
+static bool
+is_flow_deletion_pending(const struct ofproto *ofproto,
+ const struct cls_rule *cls_rule,
+ uint8_t table_id)
+{
+ if (!hmap_is_empty(&ofproto->deletions)) {
+ struct ofoperation *op;
+
+ HMAP_FOR_EACH_WITH_HASH (op, hmap_node,
+ cls_rule_hash(cls_rule, table_id),
+ &ofproto->deletions) {
+ if (cls_rule_equal(cls_rule, &op->rule->cr)) {
+ return true;
+ }
+ }
+ }
+
+ return false;
+}
+
/* Implements OFPFC_ADD and the cases for OFPFC_MODIFY and OFPFC_MODIFY_STRICT
* in which no matching flow already exists in the flow table.
*
* 'ofconn' is used to retrieve the packet buffer specified in ofm->buffer_id,
* if any. */
static int
-add_flow(struct ofconn *ofconn, struct flow_mod *fm)
+add_flow(struct ofproto *ofproto, struct ofconn *ofconn, struct flow_mod *fm,
+ const struct ofp_header *request)
{
- struct ofproto *p = ofconn_get_ofproto(ofconn);
- struct ofpbuf *packet;
+ struct classifier *table;
+ struct ofopgroup *group;
+ struct rule *victim;
struct rule *rule;
- uint16_t in_port;
- int buf_err;
int error;
+ /* Check for overlap, if requested. */
if (fm->flags & OFPFF_CHECK_OVERLAP) {
struct classifier *cls;
- FOR_EACH_MATCHING_TABLE (cls, fm->table_id, p) {
+ FOR_EACH_MATCHING_TABLE (cls, fm->table_id, ofproto) {
if (classifier_rule_overlaps(cls, &fm->cr)) {
return ofp_mkerr(OFPET_FLOW_MOD_FAILED, OFPFMFC_OVERLAP);
}
}
}
- buf_err = ofconn_pktbuf_retrieve(ofconn, fm->buffer_id, &packet, &in_port);
- error = rule_create(p, &fm->cr, fm->table_id, fm->actions, fm->n_actions,
- fm->idle_timeout, fm->hard_timeout, fm->cookie,
- fm->flags & OFPFF_SEND_FLOW_REM, &rule);
- if (error) {
- ofpbuf_delete(packet);
- return error;
+ /* Pick table. */
+ if (fm->table_id == 0xff) {
+ uint8_t table_id;
+ if (ofproto->n_tables > 1) {
+ error = ofproto->ofproto_class->rule_choose_table(ofproto, &fm->cr,
+ &table_id);
+ if (error) {
+ return error;
+ }
+ assert(table_id < ofproto->n_tables);
+ table = &ofproto->tables[table_id];
+ } else {
+ table = &ofproto->tables[0];
+ }
+ } else if (fm->table_id < ofproto->n_tables) {
+ table = &ofproto->tables[fm->table_id];
+ } else {
+ return ofp_mkerr_nicira(OFPET_FLOW_MOD_FAILED, NXFMFC_BAD_TABLE_ID);
}
- if (packet) {
- assert(!buf_err);
- return rule_execute(rule, in_port, packet);
+ /* Serialize against pending deletion. */
+ if (is_flow_deletion_pending(ofproto, &fm->cr, table - ofproto->tables)) {
+ return OFPROTO_POSTPONE;
}
- return buf_err;
-}
-static int
-send_buffered_packet(struct ofconn *ofconn,
- struct rule *rule, uint32_t buffer_id)
-{
- struct ofpbuf *packet;
- uint16_t in_port;
- int error;
-
- if (buffer_id == UINT32_MAX) {
- return 0;
+ /* Allocate new rule. */
+ rule = ofproto->ofproto_class->rule_alloc();
+ if (!rule) {
+ VLOG_WARN_RL(&rl, "%s: failed to create rule (%s)",
+ ofproto->name, strerror(error));
+ return ENOMEM;
+ }
+ rule->ofproto = ofproto;
+ rule->cr = fm->cr;
+ rule->pending = NULL;
+ rule->flow_cookie = fm->cookie;
+ rule->created = time_msec();
+ rule->idle_timeout = fm->idle_timeout;
+ rule->hard_timeout = fm->hard_timeout;
+ rule->table_id = table - ofproto->tables;
+ rule->send_flow_removed = (fm->flags & OFPFF_SEND_FLOW_REM) != 0;
+ rule->actions = ofputil_actions_clone(fm->actions, fm->n_actions);
+ rule->n_actions = fm->n_actions;
+
+ /* Insert new rule. */
+ victim = rule_from_cls_rule(classifier_replace(table, &rule->cr));
+ if (victim && victim->pending) {
+ error = OFPROTO_POSTPONE;
+ } else {
+ group = (ofconn
+ ? ofopgroup_create_for_ofconn(ofconn, request, fm->buffer_id)
+ : ofopgroup_create(ofproto));
+ ofoperation_create(group, rule, OFOPERATION_ADD);
+ rule->pending->victim = victim;
+
+ error = ofproto->ofproto_class->rule_construct(rule);
+ if (error) {
+ ofoperation_destroy(rule->pending);
+ }
+ ofopgroup_submit(group);
}
- error = ofconn_pktbuf_retrieve(ofconn, buffer_id, &packet, &in_port);
+ /* Back out if an error occurred. */
if (error) {
- return error;
+ if (victim) {
+ classifier_replace(table, &victim->cr);
+ } else {
+ classifier_remove(table, &rule->cr);
+ }
+ ofproto_rule_destroy__(rule);
}
-
- return rule_execute(rule, in_port, packet);
+ return error;
}
\f
/* OFPFC_MODIFY and OFPFC_MODIFY_STRICT. */
* Returns 0 on success, otherwise an OpenFlow error code. */
static int
modify_flows__(struct ofconn *ofconn, const struct flow_mod *fm,
- struct list *rules)
+ const struct ofp_header *request, struct list *rules)
{
- struct rule *match;
+ struct ofopgroup *group;
struct rule *rule;
- int error;
- error = 0;
- match = NULL;
+ group = ofopgroup_create_for_ofconn(ofconn, request, fm->buffer_id);
LIST_FOR_EACH (rule, ofproto_node, rules) {
if (!ofputil_actions_equal(fm->actions, fm->n_actions,
rule->actions, rule->n_actions)) {
- int retval;
-
- retval = rule->ofproto->ofproto_class->rule_modify_actions(
- rule, fm->actions, fm->n_actions);
- if (!retval) {
- match = rule;
- free(rule->actions);
- rule->actions = ofputil_actions_clone(fm->actions,
- fm->n_actions);
- rule->n_actions = fm->n_actions;
- } else if (!error) {
- error = retval;
- }
+ ofoperation_create(group, rule, OFOPERATION_MODIFY);
+ rule->pending->actions = rule->actions;
+ rule->pending->n_actions = rule->n_actions;
+ rule->actions = ofputil_actions_clone(fm->actions, fm->n_actions);
+ rule->n_actions = fm->n_actions;
+ rule->ofproto->ofproto_class->rule_modify_actions(rule);
}
rule->flow_cookie = fm->cookie;
}
+ ofopgroup_submit(group);
- if (!error && match) {
- /* This credits the packet to whichever flow happened to match last.
- * That's weird. Maybe we should do a lookup for the flow that
- * actually matches the packet? Who knows. */
- send_buffered_packet(ofconn, match, fm->buffer_id);
- }
-
- return error;
+ return 0;
}
/* Implements OFPFC_MODIFY. Returns 0 on success or an OpenFlow error code as
* 'ofconn' is used to retrieve the packet buffer specified in fm->buffer_id,
* if any. */
static int
-modify_flows_loose(struct ofconn *ofconn, struct flow_mod *fm)
+modify_flows_loose(struct ofconn *ofconn, struct flow_mod *fm,
+ const struct ofp_header *request)
{
struct ofproto *p = ofconn_get_ofproto(ofconn);
struct list rules;
error = collect_rules_loose(p, fm->table_id, &fm->cr, OFPP_NONE, &rules);
return (error ? error
- : list_is_empty(&rules) ? add_flow(ofconn, fm)
- : modify_flows__(ofconn, fm, &rules));
+ : list_is_empty(&rules) ? add_flow(p, ofconn, fm, request)
+ : modify_flows__(ofconn, fm, request, &rules));
}
/* Implements OFPFC_MODIFY_STRICT. Returns 0 on success or an OpenFlow error
* 'ofconn' is used to retrieve the packet buffer specified in fm->buffer_id,
* if any. */
static int
-modify_flow_strict(struct ofconn *ofconn, struct flow_mod *fm)
+modify_flow_strict(struct ofconn *ofconn, struct flow_mod *fm,
+ const struct ofp_header *request)
{
struct ofproto *p = ofconn_get_ofproto(ofconn);
struct list rules;
error = collect_rules_strict(p, fm->table_id, &fm->cr, OFPP_NONE, &rules);
return (error ? error
- : list_is_empty(&rules) ? add_flow(ofconn, fm)
- : list_is_singleton(&rules) ? modify_flows__(ofconn, fm, &rules)
+ : list_is_empty(&rules) ? add_flow(p, ofconn, fm, request)
+ : list_is_singleton(&rules) ? modify_flows__(ofconn, fm, request,
+ &rules)
: 0);
}
\f
*
* Returns 0 on success, otherwise an OpenFlow error code. */
static int
-delete_flows__(struct list *rules)
+delete_flows__(struct ofconn *ofconn, const struct ofp_header *request,
+ struct list *rules)
{
+ struct ofproto *ofproto = ofconn_get_ofproto(ofconn);
struct rule *rule, *next;
+ struct ofopgroup *group;
+ group = ofopgroup_create_for_ofconn(ofconn, request, UINT32_MAX);
LIST_FOR_EACH_SAFE (rule, next, ofproto_node, rules) {
ofproto_rule_send_removed(rule, OFPRR_DELETE);
- ofproto_rule_destroy(rule);
+
+ ofoperation_create(group, rule, OFOPERATION_DELETE);
+ classifier_remove(&ofproto->tables[rule->table_id], &rule->cr);
+ rule->ofproto->ofproto_class->rule_destruct(rule);
}
+ ofopgroup_submit(group);
return 0;
}
/* Implements OFPFC_DELETE. */
static int
-delete_flows_loose(struct ofproto *p, const struct flow_mod *fm)
+delete_flows_loose(struct ofconn *ofconn, const struct flow_mod *fm,
+ const struct ofp_header *request)
{
+ struct ofproto *p = ofconn_get_ofproto(ofconn);
struct list rules;
int error;
error = collect_rules_loose(p, fm->table_id, &fm->cr, fm->out_port,
&rules);
return (error ? error
- : !list_is_empty(&rules) ? delete_flows__(&rules)
+ : !list_is_empty(&rules) ? delete_flows__(ofconn, request, &rules)
: 0);
}
/* Implements OFPFC_DELETE_STRICT. */
static int
-delete_flow_strict(struct ofproto *p, struct flow_mod *fm)
+delete_flow_strict(struct ofconn *ofconn, struct flow_mod *fm,
+ const struct ofp_header *request)
{
+ struct ofproto *p = ofconn_get_ofproto(ofconn);
struct list rules;
int error;
error = collect_rules_strict(p, fm->table_id, &fm->cr, fm->out_port,
&rules);
return (error ? error
- : list_is_singleton(&rules) ? delete_flows__(&rules)
+ : list_is_singleton(&rules) ? delete_flows__(ofconn, request,
+ &rules)
: 0);
}
void
ofproto_rule_expire(struct rule *rule, uint8_t reason)
{
+ struct ofproto *ofproto = rule->ofproto;
+ struct ofopgroup *group;
+
assert(reason == OFPRR_HARD_TIMEOUT || reason == OFPRR_IDLE_TIMEOUT);
+
ofproto_rule_send_removed(rule, reason);
- ofproto_rule_destroy(rule);
+
+ group = ofopgroup_create(ofproto);
+ ofoperation_create(group, rule, OFOPERATION_DELETE);
+ classifier_remove(&ofproto->tables[rule->table_id], &rule->cr);
+ rule->ofproto->ofproto_class->rule_destruct(rule);
+ ofopgroup_submit(group);
}
\f
static int
handle_flow_mod(struct ofconn *ofconn, const struct ofp_header *oh)
{
- struct ofproto *p = ofconn_get_ofproto(ofconn);
+ struct ofproto *ofproto = ofconn_get_ofproto(ofconn);
struct flow_mod fm;
int error;
return error;
}
+ if (list_size(&ofproto->pending) >= 50) {
+ return OFPROTO_POSTPONE;
+ }
+
error = ofputil_decode_flow_mod(&fm, oh,
ofconn_get_flow_mod_table_id(ofconn));
if (error) {
switch (fm.command) {
case OFPFC_ADD:
- return add_flow(ofconn, &fm);
+ return add_flow(ofproto, ofconn, &fm, oh);
case OFPFC_MODIFY:
- return modify_flows_loose(ofconn, &fm);
+ return modify_flows_loose(ofconn, &fm, oh);
case OFPFC_MODIFY_STRICT:
- return modify_flow_strict(ofconn, &fm);
+ return modify_flow_strict(ofconn, &fm, oh);
case OFPFC_DELETE:
- delete_flows_loose(p, &fm);
- return 0;
+ return delete_flows_loose(ofconn, &fm, oh);
case OFPFC_DELETE_STRICT:
- delete_flow_strict(p, &fm);
- return 0;
+ return delete_flow_strict(ofconn, &fm, oh);
default:
if (fm.command > 0xff) {
return ofp_mkerr(OFPET_BAD_REQUEST, -1);
}
+ if (ofconn_get_role(ofconn) != role
+ && ofconn_has_pending_opgroups(ofconn)) {
+ return OFPROTO_POSTPONE;
+ }
+
ofconn_set_role(ofconn, role);
reply = make_nxmsg_xid(sizeof *reply, NXT_ROLE_REPLY, oh->xid, &buf);
uint32_t format;
format = ntohl(msg->format);
- if (format == NXFF_OPENFLOW10
- || format == NXFF_NXM) {
- ofconn_set_flow_format(ofconn, format);
- return 0;
- } else {
+ if (format != NXFF_OPENFLOW10 && format != NXFF_NXM) {
return ofp_mkerr(OFPET_BAD_REQUEST, OFPBRC_EPERM);
}
+
+ if (format != ofconn_get_flow_format(ofconn)
+ && ofconn_has_pending_opgroups(ofconn)) {
+ /* Avoid sending async messages in surprising flow format. */
+ return OFPROTO_POSTPONE;
+ }
+
+ ofconn_set_flow_format(ofconn, format);
+ return 0;
}
static int
struct ofp_header *ob;
struct ofpbuf *buf;
- /* Currently, everything executes synchronously, so we can just
- * immediately send the barrier reply. */
+ if (ofconn_has_pending_opgroups(ofconn)) {
+ return OFPROTO_POSTPONE;
+ }
+
ob = make_openflow_xid(sizeof *ob, OFPT_BARRIER_REPLY, oh->xid, &buf);
ofconn_send_reply(ofconn, buf);
return 0;
}
}
-static void
+static bool
handle_openflow(struct ofconn *ofconn, struct ofpbuf *ofp_msg)
{
int error = handle_openflow__(ofconn, ofp_msg);
- if (error) {
+ if (error && error != OFPROTO_POSTPONE) {
ofconn_send_error(ofconn, ofp_msg->data, error);
}
COVERAGE_INC(ofproto_recv_openflow);
+ return error != OFPROTO_POSTPONE;
+}
+\f
+/* Asynchronous operations. */
+
+/* Creates and returns a new ofopgroup that is not associated with any
+ * OpenFlow connection.
+ *
+ * The caller should add operations to the returned group with
+ * ofoperation_create() and then submit it with ofopgroup_submit(). */
+static struct ofopgroup *
+ofopgroup_create(struct ofproto *ofproto)
+{
+ struct ofopgroup *group = xzalloc(sizeof *group);
+ group->ofproto = ofproto;
+ list_init(&group->ofproto_node);
+ list_init(&group->ops);
+ list_init(&group->ofconn_node);
+ return group;
+}
+
+/* Creates and returns a new ofopgroup that is associated with 'ofconn'. If
+ * the ofopgroup eventually fails, then the error reply will include 'request'.
+ * If the ofopgroup eventually succeeds, then the packet with buffer id
+ * 'buffer_id' on 'ofconn' will be sent by 'ofconn''s ofproto.
+ *
+ * The caller should add operations to the returned group with
+ * ofoperation_create() and then submit it with ofopgroup_submit(). */
+static struct ofopgroup *
+ofopgroup_create_for_ofconn(struct ofconn *ofconn,
+ const struct ofp_header *request,
+ uint32_t buffer_id)
+{
+ struct ofopgroup *group = ofopgroup_create(ofconn_get_ofproto(ofconn));
+ size_t request_len = ntohs(request->length);
+
+ ofconn_add_opgroup(ofconn, &group->ofconn_node);
+ group->ofconn = ofconn;
+ group->request = xmemdup(request, MIN(request_len, 64));
+ group->buffer_id = buffer_id;
+
+ return group;
+}
+
+/* Submits 'group' for processing.
+ *
+ * If 'group' contains no operations (e.g. none were ever added, or all of the
+ * ones that were added completed synchronously), then it is destroyed
+ * immediately. Otherwise it is added to the ofproto's list of pending
+ * groups. */
+static void
+ofopgroup_submit(struct ofopgroup *group)
+{
+ if (list_is_empty(&group->ops)) {
+ ofopgroup_destroy(group);
+ } else {
+ list_push_back(&group->ofproto->pending, &group->ofproto_node);
+ }
+}
+
+static void
+ofopgroup_destroy(struct ofopgroup *group)
+{
+ assert(list_is_empty(&group->ops));
+ if (!list_is_empty(&group->ofproto_node)) {
+ list_remove(&group->ofproto_node);
+ }
+ if (!list_is_empty(&group->ofconn_node)) {
+ list_remove(&group->ofconn_node);
+ if (group->error) {
+ ofconn_send_error(group->ofconn, group->request, group->error);
+ }
+ connmgr_retry(group->ofproto->connmgr);
+ }
+ free(group->request);
+ free(group);
+}
+
+/* Initiates a new operation on 'rule', of the specified 'type', within
+ * 'group'. Prior to calling, 'rule' must not have any pending operation. */
+static void
+ofoperation_create(struct ofopgroup *group, struct rule *rule,
+ enum ofoperation_type type)
+{
+ struct ofoperation *op;
+
+ assert(!rule->pending);
+
+ op = rule->pending = xzalloc(sizeof *op);
+ op->group = group;
+ list_push_back(&group->ops, &op->group_node);
+ op->rule = rule;
+ op->type = type;
+ op->status = -1;
+ op->flow_cookie = rule->flow_cookie;
+
+ if (type == OFOPERATION_DELETE) {
+ hmap_insert(&op->group->ofproto->deletions, &op->hmap_node,
+ cls_rule_hash(&rule->cr, rule->table_id));
+ }
+}
+
+static void
+ofoperation_destroy(struct ofoperation *op)
+{
+ struct ofopgroup *group = op->group;
+
+ if (op->rule) {
+ op->rule->pending = NULL;
+ }
+ if (op->type == OFOPERATION_DELETE) {
+ hmap_remove(&group->ofproto->deletions, &op->hmap_node);
+ }
+ list_remove(&op->group_node);
+ free(op->actions);
+ free(op);
+
+ if (list_is_empty(&group->ops) && !list_is_empty(&group->ofproto_node)) {
+ ofopgroup_destroy(group);
+ }
+}
+
+/* Indicates that 'op' completed with status 'error', which is either 0 to
+ * indicate success or an OpenFlow error code (constructed with
+ * e.g. ofp_mkerr()).
+ *
+ * If 'op' is a "delete flow" operation, 'error' must be 0. That is, flow
+ * deletions are not allowed to fail.
+ *
+ * Please see the large comment in ofproto/private.h titled "Asynchronous
+ * Operation Support" for more information. */
+void
+ofoperation_complete(struct ofoperation *op, int error)
+{
+ struct ofopgroup *group = op->group;
+ struct rule *rule = op->rule;
+ struct classifier *table = &rule->ofproto->tables[rule->table_id];
+
+ assert(rule->pending == op);
+ assert(op->status < 0);
+ assert(error >= 0);
+
+ if (!error
+ && !group->error
+ && op->type != OFOPERATION_DELETE
+ && group->ofconn
+ && group->buffer_id != UINT32_MAX
+ && list_is_singleton(&op->group_node)) {
+ struct ofpbuf *packet;
+ uint16_t in_port;
+
+ error = ofconn_pktbuf_retrieve(group->ofconn, group->buffer_id,
+ &packet, &in_port);
+ if (packet) {
+ assert(!error);
+ error = rule_execute(rule, in_port, packet);
+ }
+ }
+ if (!group->error) {
+ group->error = error;
+ }
+
+ switch (op->type) {
+ case OFOPERATION_ADD:
+ if (!error) {
+ if (op->victim) {
+ ofproto_rule_destroy__(op->victim);
+ }
+ } else {
+ if (op->victim) {
+ classifier_replace(table, &op->victim->cr);
+ op->victim = NULL;
+ } else {
+ classifier_remove(table, &rule->cr);
+ }
+ ofproto_rule_destroy__(rule);
+ }
+ op->victim = NULL;
+ break;
+
+ case OFOPERATION_DELETE:
+ assert(!error);
+ ofproto_rule_destroy__(rule);
+ op->rule = NULL;
+ break;
+
+ case OFOPERATION_MODIFY:
+ if (error) {
+ free(rule->actions);
+ rule->actions = op->actions;
+ rule->n_actions = op->n_actions;
+ op->actions = NULL;
+ }
+ break;
+
+ default:
+ NOT_REACHED();
+ }
+ ofoperation_destroy(op);
+}
+
+struct rule *
+ofoperation_get_victim(struct ofoperation *op)
+{
+ assert(op->type == OFOPERATION_ADD);
+ return op->victim;
}
\f
static uint64_t
/* OpenFlow connections. */
struct connmgr *connmgr;
+
+ /* Flow table operation tracking. */
+ int state; /* Internal state. */
+ struct list pending; /* List of "struct ofopgroup"s. */
+ struct hmap deletions; /* All OFOPERATION_DELETE "ofoperation"s. */
};
struct ofproto *ofproto_lookup(const char *name);
struct list ofproto_node; /* Owned by ofproto base code. */
struct cls_rule cr; /* In owning ofproto's classifier. */
+ struct ofoperation *pending; /* Operation now in progress, if nonnull. */
+
ovs_be64 flow_cookie; /* Controller-issued identifier. */
long long int created; /* Creation time. */
void ofproto_rule_expire(struct rule *, uint8_t reason);
void ofproto_rule_destroy(struct rule *);
+void ofoperation_complete(struct ofoperation *, int status);
+struct rule *ofoperation_get_victim(struct ofoperation *);
+
/* ofproto class structure, to be defined by each ofproto implementation.
*
*
/* ## --------------------------- ## */
/* Life-cycle functions for an "ofproto" (see "Life Cycle" above).
+ *
+ *
+ * Construction
+ * ============
*
* ->construct() should not modify most base members of the ofproto. In
* particular, the client will initialize the ofproto's 'ports' member
* allowed to fail with an error.
*
* ->construct() returns 0 if successful, otherwise a positive errno
- * value. */
+ * value.
+ *
+ *
+ * Destruction
+ * ===========
+ *
+ * ->destruct() must do at least the following:
+ *
+ * - If 'ofproto' has any pending asynchronous operations, ->destruct()
+ * must complete all of them by calling ofoperation_complete().
+ *
+ * - If 'ofproto' has any rules left in any of its flow tables, ->
+ */
struct ofproto *(*alloc)(void);
int (*construct)(struct ofproto *ofproto);
void (*destruct)(struct ofproto *ofproto);
/* ## OpenFlow Rule Functions ## */
/* ## ----------------------- ## */
+
+
/* Chooses an appropriate table for 'cls_rule' within 'ofproto'. On
* success, stores the table ID into '*table_idp' and returns 0. On
* failure, returns an OpenFlow error code (as returned by ofp_mkerr()).
/* Life-cycle functions for a "struct rule" (see "Life Cycle" above).
*
- * ->rule_construct() should first check whether the rule is acceptable:
*
- * - Validate that the matching rule in 'rule->cr' is supported by the
- * datapath. If not, then return an OpenFlow error code (as returned
- * by ofp_mkerr()).
+ * Asynchronous Operation Support
+ * ==============================
+ *
+ * The life-cycle operations on rules can operate asynchronously, meaning
+ * that ->rule_construct() and ->rule_destruct() only need to initiate
+ * their respective operations and do not need to wait for them to complete
+ * before they return. ->rule_modify_actions() also operates
+ * asynchronously.
+ *
+ * An ofproto implementation reports the success or failure of an
+ * asynchronous operation on a rule using the rule's 'pending' member,
+ * which points to a opaque "struct ofoperation" that represents the
+ * ongoing opreation. When the operation completes, the ofproto
+ * implementation calls ofoperation_complete(), passing the ofoperation and
+ * an error indication.
+ *
+ * Only the following contexts may call ofoperation_complete():
+ *
+ * - The function called to initiate the operation,
+ * e.g. ->rule_construct() or ->rule_destruct(). This is the best
+ * choice if the operation completes quickly.
+ *
+ * - The implementation's ->run() function.
+ *
+ * - The implementation's ->destruct() function.
+ *
+ * The ofproto base code updates the flow table optimistically, assuming
+ * that the operation will probably succeed:
+ *
+ * - ofproto adds or replaces the rule in the flow table before calling
+ * ->rule_construct().
+ *
+ * - ofproto updates the rule's actions before calling
+ * ->rule_modify_actions().
+ *
+ * - ofproto removes the rule before calling ->rule_destruct().
*
- * For example, if the datapath does not support registers, then it
- * should return an error if 'rule->cr' does not wildcard all
+ * With one exception, when an asynchronous operation completes with an
+ * error, ofoperation_complete() backs out the already applied changes:
+ *
+ * - If adding or replacing a rule in the flow table fails, ofproto
+ * removes the new rule or restores the original rule.
+ *
+ * - If modifying a rule's actions fails, ofproto restores the original
+ * actions.
+ *
+ * - Removing a rule is not allowed to fail. It must always succeed.
+ *
+ * The ofproto base code serializes operations: if any operation is in
+ * progress on a given rule, ofproto postpones initiating any new operation
+ * on that rule until the pending operation completes. Therefore, every
+ * operation must eventually complete through a call to
+ * ofoperation_complete() to avoid delaying new operations indefinitely
+ * (including any OpenFlow request that affects the rule in question, even
+ * just to query its statistics).
+ *
+ *
+ * Construction
+ * ============
+ *
+ * When ->rule_construct() is called, the caller has already inserted
+ * 'rule' into 'rule->ofproto''s flow table numbered 'rule->table_id'.
+ * There are two cases:
+ *
+ * - 'rule' is a new rule in its flow table. In this case,
+ * ofoperation_get_victim(rule) returns NULL.
+ *
+ * - 'rule' is replacing an existing rule in its flow table that had the
+ * same matching criteria and priority. In this case,
+ * ofoperation_get_victim(rule) returns the rule being replaced.
+ *
+ * ->rule_construct() should set the following in motion:
+ *
+ * - Validate that the matching rule in 'rule->cr' is supported by the
+ * datapath. For example, if the rule's table does not support
+ * registers, then it is an error if 'rule->cr' does not wildcard all
* registers.
*
* - Validate that 'rule->actions' and 'rule->n_actions' are well-formed
- * OpenFlow actions that can be correctly implemented by the datapath.
- * If not, then return an OpenFlow error code (as returned by
- * ofp_mkerr()).
- *
- * The validate_actions() function (in ofp-util.c) can be useful as a
- * model for action validation, but it accepts all of the OpenFlow
- * actions that OVS understands. If your ofproto implementation only
+ * OpenFlow actions that the datapath can correctly implement. The
+ * validate_actions() function (in ofp-util.c) can be useful as a model
+ * for action validation, but it accepts all of the OpenFlow actions
+ * that OVS understands. If your ofproto implementation only
* implements a subset of those, then you should implement your own
* action validation.
*
- * If the rule is acceptable, then ->rule_construct() should modify the
- * flow table:
+ * - If the rule is valid, update the datapath flow table, adding the new
+ * rule or replacing the existing one.
*
- * - If there was already a rule with exactly the same matching criteria
- * and priority in the classifier, then it should destroy it (with
- * ofproto_rule_destroy()).
+ * (On failure, the ofproto code will roll back the insertion from the flow
+ * table, either removing 'rule' or replacing it by the flow that was
+ * originally in its place.)
*
- * To the greatest extent possible, the old rule should be destroyed
- * only if inserting the new rule succeeds; that is, ->rule_construct()
- * should be transactional.
+ * ->rule_construct() must act in one of the following ways:
*
- * The function classifier_find_rule_exactly() can locate such a rule.
+ * - If it succeeds, it must call ofoperation_complete() and return 0.
*
- * - Insert the new rule into the ofproto's 'cls' classifier, and into
- * the datapath flow table.
+ * - If it fails, it must act in one of the following ways:
*
- * The function classifier_insert() inserts a rule into the classifier.
+ * * Call ofoperation_complete() and return 0.
*
- * Other than inserting 'rule->cr' into the classifier, ->rule_construct()
- * should not modify any base members of struct rule.
+ * * Return an OpenFlow error code (as returned by ofp_mkerr()). (Do
+ * not call ofoperation_complete() in this case.)
*
- * ->rule_destruct() should remove 'rule' from the ofproto's 'cls'
- * classifier (e.g. with classifier_remove()) and from the datapath flow
- * table. */
+ * In the former case, ->rule_destruct() will be called; in the latter
+ * case, it will not. ->rule_dealloc() will be called in either case.
+ *
+ * - If the operation is only partially complete, then it must return 0.
+ * Later, when the operation is complete, the ->run() or ->destruct()
+ * function must call ofoperation_complete() to report success or
+ * failure.
+ *
+ * ->rule_construct() should not modify any base members of struct rule.
+ *
+ *
+ * Destruction
+ * ===========
+ *
+ * When ->rule_destruct() is called, the caller has already removed 'rule'
+ * from 'rule->ofproto''s flow table. ->rule_destruct() should set in
+ * motion removing 'rule' from the datapath flow table. If removal
+ * completes synchronously, it should call ofoperation_complete().
+ * Otherwise, the ->run() or ->destruct() function must later call
+ * ofoperation_complete() after the operation completes.
+ *
+ * Rule destruction must not fail. */
struct rule *(*rule_alloc)(void);
int (*rule_construct)(struct rule *rule);
void (*rule_destruct)(struct rule *rule);
int (*rule_execute)(struct rule *rule, struct flow *flow,
struct ofpbuf *packet);
- /* Validates that the 'n' elements in 'actions' are well-formed OpenFlow
- * actions that can be correctly implemented by the datapath. If not, then
- * return an OpenFlow error code (as returned by ofp_mkerr()). If so,
- * then update the datapath to implement the new actions and return 0.
+ /* When ->rule_modify_actions() is called, the caller has already replaced
+ * the OpenFlow actions in 'rule' by a new set. (The original actions are
+ * in rule->pending->actions.)
+ *
+ * ->rule_modify_actions() should set the following in motion:
+ *
+ * - Validate that the actions now in 'rule' are well-formed OpenFlow
+ * actions that the datapath can correctly implement.
+ *
+ * - Update the datapath flow table with the new actions.
+ *
+ * If the operation synchronously completes, ->rule_modify_actions() may
+ * call ofoperation_complete() before it returns. Otherwise, ->run()
+ * should call ofoperation_complete() later, after the operation does
+ * complete.
+ *
+ * If the operation fails, then the base ofproto code will restore the
+ * original 'actions' and 'n_actions' of 'rule'.
*
- * When this function runs, 'rule' still has its original actions. If this
- * function returns 0, then the caller will update 'rule' with the new
- * actions and free the old ones. */
- int (*rule_modify_actions)(struct rule *rule,
- const union ofp_action *actions, size_t n);
+ * ->rule_modify_actions() should not modify any base members of struct
+ * rule. */
+ void (*rule_modify_actions)(struct rule *rule);
/* These functions implement the OpenFlow IP fragment handling policy. By
* default ('drop_frags' == false), an OpenFlow switch should treat IP
void ofproto_add_flow(struct ofproto *, const struct cls_rule *,
const union ofp_action *, size_t n_actions);
-void ofproto_delete_flow(struct ofproto *, const struct cls_rule *);
+bool ofproto_delete_flow(struct ofproto *, const struct cls_rule *);
void ofproto_flush_flows(struct ofproto *);
#endif /* ofproto/private.h */
projects / openvswitch / commitdiff