X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Frconn.c;h=c8bddab52efb104b752ef6e83adb90444cda096d;hb=refs%2Fheads%2Ffor-nox%2F0.4;hp=ee995e499d7ec80310687fbb12d0cd071e045582;hpb=394270a0e5bf42986f806c258591919eca2b473f;p=sliver-openvswitch.git diff --git a/lib/rconn.c b/lib/rconn.c index ee995e499..c8bddab52 100644 --- a/lib/rconn.c +++ b/lib/rconn.c @@ -38,9 +38,9 @@ #include #include #include -#include "buffer.h" +#include "ofpbuf.h" +#include "openflow/openflow.h" #include "poll-loop.h" -#include "ofp-print.h" #include "sat-math.h" #include "timeval.h" #include "util.h" @@ -83,7 +83,7 @@ struct rconn { char *name; bool reliable; - struct queue txq; + struct ofp_queue txq; int backoff; int max_backoff; @@ -92,6 +92,17 @@ struct rconn { time_t last_connected; unsigned int packets_sent; + /* In S_ACTIVE and S_IDLE, probably_admitted reports whether we believe + * that the peer has made a (positive) admission control decision on our + * connection. If we have not yet been (probably) admitted, then the + * connection does not reset the timer used for deciding whether the switch + * should go into fail-open mode. + * + * last_admitted reports the last time we believe such a positive admission + * control decision was made. */ + bool probably_admitted; + time_t last_admitted; + /* These values are simply for statistics reporting, not used directly by * anything internal to the rconn (or the secchan for that matter). */ unsigned int packets_received; @@ -114,6 +125,11 @@ struct rconn { * an echo request as an inactivity probe packet. We should receive back * a response. */ int probe_interval; /* Secs of inactivity before sending probe. */ + + /* Messages sent or received are copied to the monitor connections. */ +#define MAX_MONITORS 8 + struct vconn *monitors[8]; + size_t n_monitors; }; static unsigned int elapsed_in_this_state(const struct rconn *); @@ -125,6 +141,9 @@ static int reconnect(struct rconn *); static void disconnect(struct rconn *, int error); static void flush_queue(struct rconn *); static void question_connectivity(struct rconn *); +static void copy_to_monitor(struct rconn *, const struct ofpbuf *); +static bool is_connected_state(enum state); +static bool is_admitted_msg(const struct ofpbuf *); /* Creates a new rconn, connects it (reliably) to 'name', and returns it. */ struct rconn * @@ -178,6 +197,9 @@ rconn_create(int probe_interval, int max_backoff) rc->packets_sent = 0; + rc->probably_admitted = false; + rc->last_admitted = time_now(); + rc->packets_received = 0; rc->n_attempted_connections = 0; rc->n_successful_connections = 0; @@ -189,6 +211,8 @@ rconn_create(int probe_interval, int max_backoff) rc->probe_interval = probe_interval ? MAX(5, probe_interval) : 0; + rc->n_monitors = 0; + return rc; } @@ -219,18 +243,20 @@ rconn_connect_unreliably(struct rconn *rc, void rconn_disconnect(struct rconn *rc) { - if (rc->vconn) { - vconn_close(rc->vconn); - rc->vconn = NULL; - } - free(rc->name); - rc->name = xstrdup("void"); - rc->reliable = false; + if (rc->state != S_VOID) { + if (rc->vconn) { + vconn_close(rc->vconn); + rc->vconn = NULL; + } + free(rc->name); + rc->name = xstrdup("void"); + rc->reliable = false; - rc->backoff = 0; - rc->backoff_deadline = TIME_MIN; + rc->backoff = 0; + rc->backoff_deadline = TIME_MIN; - state_transition(rc, S_VOID); + state_transition(rc, S_VOID); + } } /* Disconnects 'rc' and frees the underlying storage. */ @@ -238,10 +264,15 @@ void rconn_destroy(struct rconn *rc) { if (rc) { + size_t i; + free(rc->name); vconn_close(rc->vconn); flush_queue(rc); queue_destroy(&rc->txq); + for (i = 0; i < rc->n_monitors; i++) { + vconn_close(rc->monitors[i]); + } free(rc); } } @@ -265,12 +296,13 @@ reconnect(struct rconn *rc) VLOG_WARN("%s: connecting...", rc->name); rc->n_attempted_connections++; - retval = vconn_open(rc->name, &rc->vconn); + retval = vconn_open(rc->name, OFP_VERSION, &rc->vconn); if (!retval) { rc->backoff_deadline = time_now() + rc->backoff; state_transition(rc, S_CONNECTING); } else { VLOG_WARN("%s: connection failed (%s)", rc->name, strerror(retval)); + rc->backoff_deadline = TIME_MAX; /* Prevent resetting backoff. */ disconnect(rc, 0); } return retval; @@ -303,13 +335,8 @@ run_CONNECTING(struct rconn *rc) if (!retval) { VLOG_WARN("%s: connected", rc->name); rc->n_successful_connections++; - if (vconn_is_passive(rc->vconn)) { - error(0, "%s: passive vconn not supported", rc->name); - state_transition(rc, S_VOID); - } else { - state_transition(rc, S_ACTIVE); - rc->last_connected = rc->state_entered; - } + state_transition(rc, S_ACTIVE); + rc->last_connected = rc->state_entered; } else if (retval != EAGAIN) { VLOG_WARN("%s: connection failed (%s)", rc->name, strerror(retval)); disconnect(rc, retval); @@ -353,10 +380,14 @@ run_ACTIVE(struct rconn *rc) { if (timed_out(rc)) { unsigned int base = MAX(rc->last_received, rc->state_entered); - rconn_send(rc, make_echo_request(), NULL); VLOG_DBG("%s: idle %u seconds, sending inactivity probe", rc->name, (unsigned int) (time_now() - base)); + + /* Ordering is important here: rconn_send() can transition to BACKOFF, + * and we don't want to transition back to IDLE if so, because then we + * can end up queuing a packet with vconn == NULL and then *boom*. */ state_transition(rc, S_IDLE); + rconn_send(rc, make_echo_request(), NULL); return; } @@ -421,14 +452,20 @@ rconn_run_wait(struct rconn *rc) /* Attempts to receive a packet from 'rc'. If successful, returns the packet; * otherwise, returns a null pointer. The caller is responsible for freeing - * the packet (with buffer_delete()). */ -struct buffer * + * the packet (with ofpbuf_delete()). */ +struct ofpbuf * rconn_recv(struct rconn *rc) { if (rc->state & (S_ACTIVE | S_IDLE)) { - struct buffer *buffer; + struct ofpbuf *buffer; int error = vconn_recv(rc->vconn, &buffer); if (!error) { + copy_to_monitor(rc, buffer); + if (is_admitted_msg(buffer) + || time_now() - rc->last_connected >= 30) { + rc->probably_admitted = true; + rc->last_admitted = time_now(); + } rc->last_received = time_now(); rc->packets_received++; if (rc->state == S_IDLE) { @@ -466,9 +503,10 @@ rconn_recv_wait(struct rconn *rc) * takes care of sending if you call rconn_run(), which will have the side * effect of waking up poll_block(). */ int -rconn_send(struct rconn *rc, struct buffer *b, int *n_queued) +rconn_send(struct rconn *rc, struct ofpbuf *b, int *n_queued) { if (rconn_is_connected(rc)) { + copy_to_monitor(rc, b); b->private = n_queued; if (n_queued) { ++*n_queued; @@ -486,7 +524,7 @@ rconn_send(struct rconn *rc, struct buffer *b, int *n_queued) /* Sends 'b' on 'rc'. Increments '*n_queued' while the packet is in flight; it * will be decremented when it has been sent (or discarded due to * disconnection). Returns 0 if successful, EAGAIN if '*n_queued' is already - * at least as large of 'queue_limit', or ENOTCONN if 'rc' is not currently + * at least as large as 'queue_limit', or ENOTCONN if 'rc' is not currently * connected. Regardless of return value, 'b' is destroyed. * * Because 'b' may be sent (or discarded) before this function returns, the @@ -496,13 +534,13 @@ rconn_send(struct rconn *rc, struct buffer *b, int *n_queued) * takes care of sending if you call rconn_run(), which will have the side * effect of waking up poll_block(). */ int -rconn_send_with_limit(struct rconn *rc, struct buffer *b, +rconn_send_with_limit(struct rconn *rc, struct ofpbuf *b, int *n_queued, int queue_limit) { int retval; retval = *n_queued >= queue_limit ? EAGAIN : rconn_send(rc, b, n_queued); if (retval) { - buffer_delete(b); + ofpbuf_delete(b); } return retval; } @@ -516,6 +554,21 @@ rconn_packets_sent(const struct rconn *rc) return rc->packets_sent; } +/* Adds 'vconn' to 'rc' as a monitoring connection, to which all messages sent + * and received on 'rconn' will be copied. 'rc' takes ownership of 'vconn'. */ +void +rconn_add_monitor(struct rconn *rc, struct vconn *vconn) +{ + if (rc->n_monitors < ARRAY_SIZE(rc->monitors)) { + VLOG_WARN("new monitor connection from %s", vconn_get_name(vconn)); + rc->monitors[rc->n_monitors++] = vconn; + } else { + VLOG_DBG("too many monitor connections, discarding %s", + vconn_get_name(vconn)); + vconn_close(vconn); + } +} + /* Returns 'rc''s name (the 'name' argument passed to rconn_new()). */ const char * rconn_get_name(const struct rconn *rc) @@ -535,15 +588,18 @@ rconn_is_alive(const struct rconn *rconn) bool rconn_is_connected(const struct rconn *rconn) { - return rconn->state & (S_ACTIVE | S_IDLE); + return is_connected_state(rconn->state); } -/* Returns 0 if 'rconn' is connected, otherwise the number of seconds that it - * has been disconnected. */ +/* Returns 0 if 'rconn' is connected. Otherwise, if 'rconn' is in a "failure + * mode" (that is, it is not connected), returns the number of seconds that it + * has been in failure mode, ignoring any times that it connected but the + * controller's admission control policy caused it to be quickly + * disconnected. */ int -rconn_disconnected_duration(const struct rconn *rconn) +rconn_failure_duration(const struct rconn *rconn) { - return rconn_is_connected(rconn) ? 0 : time_now() - rconn->last_received; + return rconn_is_connected(rconn) ? 0 : time_now() - rconn->last_admitted; } /* Returns the IP address of the peer, or 0 if the peer is not connected over @@ -623,6 +679,21 @@ rconn_get_total_time_connected(const struct rconn *rc) return (rc->total_time_connected + (rconn_is_connected(rc) ? elapsed_in_this_state(rc) : 0)); } + +/* Returns the current amount of backoff, in seconds. This is the amount of + * time after which the rconn will transition from BACKOFF to CONNECTING. */ +int +rconn_get_backoff(const struct rconn *rc) +{ + return rc->backoff; +} + +/* Returns the number of seconds spent in this state so far. */ +unsigned int +rconn_get_state_elapsed(const struct rconn *rc) +{ + return elapsed_in_this_state(rc); +} /* Tries to send a packet from 'rc''s send buffer. Returns 0 if successful, * otherwise a positive errno value. */ @@ -630,7 +701,7 @@ static int try_send(struct rconn *rc) { int retval = 0; - struct buffer *next = rc->txq.head->next; + struct ofpbuf *next = rc->txq.head->next; int *n_queued = rc->txq.head->private; retval = vconn_send(rc->vconn, rc->txq.head); if (retval) { @@ -662,7 +733,7 @@ disconnect(struct rconn *rc, int error) rc->name, strerror(error)); } else if (error == EOF) { if (rc->reliable) { - VLOG_WARN("%s: connection closed", rc->name); + VLOG_WARN("%s: connection closed by peer", rc->name); } } else { VLOG_WARN("%s: connection dropped", rc->name); @@ -698,12 +769,12 @@ flush_queue(struct rconn *rc) return; } while (rc->txq.n > 0) { - struct buffer *b = queue_pop_head(&rc->txq); + struct ofpbuf *b = queue_pop_head(&rc->txq); int *n_queued = b->private; if (n_queued) { --*n_queued; } - buffer_delete(b); + ofpbuf_delete(b); } poll_immediate_wake(); } @@ -735,6 +806,9 @@ timed_out(const struct rconn *rc) static void state_transition(struct rconn *rc, enum state state) { + if (is_connected_state(state) && !is_connected_state(rc->state)) { + rc->probably_admitted = false; + } if (rconn_is_connected(rc)) { rc->total_time_connected += elapsed_in_this_state(rc); } @@ -752,3 +826,55 @@ question_connectivity(struct rconn *rc) rc->last_questioned = now; } } + +static void +copy_to_monitor(struct rconn *rc, const struct ofpbuf *b) +{ + struct ofpbuf *clone = NULL; + int retval; + size_t i; + + for (i = 0; i < rc->n_monitors; ) { + struct vconn *vconn = rc->monitors[i]; + + if (!clone) { + clone = ofpbuf_clone(b); + } + retval = vconn_send(vconn, clone); + if (!retval) { + clone = NULL; + } else if (retval != EAGAIN) { + VLOG_DBG("%s: closing monitor connection to %s: %s", + rconn_get_name(rc), vconn_get_name(vconn), + strerror(retval)); + rc->monitors[i] = rc->monitors[--rc->n_monitors]; + continue; + } + i++; + } + ofpbuf_delete(clone); +} + +static bool +is_connected_state(enum state state) +{ + return (state & (S_ACTIVE | S_IDLE)) != 0; +} + +static bool +is_admitted_msg(const struct ofpbuf *b) +{ + struct ofp_header *oh = b->data; + uint8_t type = oh->type; + return !(type < 32 + && (1u << type) & ((1u << OFPT_HELLO) | + (1u << OFPT_ERROR) | + (1u << OFPT_ECHO_REQUEST) | + (1u << OFPT_ECHO_REPLY) | + (1u << OFPT_VENDOR) | + (1u << OFPT_FEATURES_REQUEST) | + (1u << OFPT_FEATURES_REPLY) | + (1u << OFPT_GET_CONFIG_REQUEST) | + (1u << OFPT_GET_CONFIG_REPLY) | + (1u << OFPT_SET_CONFIG))); +}