Bug 7691: Send a probe cell down certain types of circs.

In general, if we tried to use a circ for a stream, but then decided to place
that stream on a different circuit, we need to probe the original circuit
before deciding it was a "success".

We also need to do the same for cannibalized circuits that go unused.

8 files changed, 218 insertions, 21 deletions

diff --git a/src/or/circuitbuild.c b/src/or/circuitbuild.c
index 7d94b2bb1d..73bd3d4e40 100644
--- a/src/or/circuitbuild.c
+++ b/src/or/circuitbuild.c
@@ -39,6 +39,7 @@
 #include "routerparse.h"
 #include "routerset.h"
 #include "crypto.h"
+#include "connection_edge.h"
 
 #ifndef MIN
 #define MIN(a,b) ((a)<(b)?(a):(b))
@@ -1504,6 +1505,149 @@ pathbias_count_build_success(origin_circuit_t *circ)
 }
 
 /**
+ * Send a probe down a circuit that wasn't usable.
+ *
+ * Returns -1 if we couldn't probe, 0 otherwise.
+ */
+static int
+pathbias_send_usable_probe(circuit_t *circ)
+{
+  /* Based on connection_ap_handshake_send_begin() */
+  char payload[CELL_PAYLOAD_SIZE];
+  int payload_len;
+  origin_circuit_t *ocirc = TO_ORIGIN_CIRCUIT(circ);
+  crypt_path_t *cpath_layer = NULL;
+  // XXX: Generate a random 0.a.b.c adddress
+  const char *probe_nonce = "0.1.2.3";
+
+  tor_assert(ocirc);
+
+  cpath_layer = ocirc->cpath->prev;
+
+  if (cpath_layer->state != CPATH_STATE_OPEN) {
+    /* This can happen for cannibalized circuits. Their
+     * last hop isn't yet open */
+    log_info(LD_CIRC,
+             "Got pathbias probe request for unopened circuit %d. "
+             "Opened %d, len %d", ocirc->global_identifier,
+             ocirc->has_opened, ocirc->build_state->desired_path_len);
+    return -1;
+  }
+
+  /* We already went down this road. */
+  if (circ->purpose == CIRCUIT_PURPOSE_PATH_BIAS_TESTING &&
+      ocirc->pathbias_probe_id) {
+    log_info(LD_CIRC,
+             "Got pathbias probe request for circuit %d with "
+             "outstanding probe", ocirc->global_identifier);
+    return -1;
+  }
+
+  circuit_change_purpose(circ, CIRCUIT_PURPOSE_PATH_BIAS_TESTING);
+
+  /* Update timestamp for circuit_expire_building to kill us */
+  tor_gettimeofday(&circ->timestamp_began);
+
+  tor_snprintf(payload,RELAY_PAYLOAD_SIZE, "%s:25", probe_nonce);
+  tor_addr_parse(&ocirc->pathbias_probe_nonce, probe_nonce);
+
+  payload_len = (int)strlen(payload)+1;
+
+  // XXX: need this? Can we assume ipv4 will always be supported?
+  // If not, how do we tell?
+  //if (payload_len <= RELAY_PAYLOAD_SIZE - 4 && edge_conn->begincell_flags) {
+  //  set_uint32(payload + payload_len, htonl(edge_conn->begincell_flags));
+  //  payload_len += 4;
+  //}
+
+  /* Generate+Store stream id, make sure it's non-zero */
+  ocirc->pathbias_probe_id = get_unique_stream_id_by_circ(ocirc);
+
+  if (ocirc->pathbias_probe_id==0) {
+    log_warn(LD_CIRC,
+             "Ran out of stream IDs on circuit %u during "
+             "pathbias probe attempt.", ocirc->global_identifier);
+    return -1;
+  }
+
+  log_info(LD_CIRC,
+           "Sending pathbias testing cell to %s:25 on stream %d for circ %d.",
+           probe_nonce, ocirc->pathbias_probe_id, ocirc->global_identifier);
+
+  /* Send a test relay cell */
+  if (relay_send_command_from_edge(ocirc->pathbias_probe_id, circ,
+                               RELAY_COMMAND_BEGIN, payload,
+                               payload_len, cpath_layer) < 0) {
+    log_notice(LD_CIRC,
+             "Failed to send pathbias probe cell on circuit %d.",
+             ocirc->global_identifier);
+    return -1;
+  }
+
+  /* Mark it freshly dirty so it doesn't get expired in the meantime */
+  circ->timestamp_dirty = time(NULL);
+
+  return 0;
+}
+
+/**
+ * Check the response to a pathbias probe.
+ *
+ * If the response is valid, return 0. Otherwise return < 0.
+ */
+int
+pathbias_check_probe_response(circuit_t *circ, cell_t *cell)
+{
+  /* Based on connection_edge_process_relay_cell() */
+  relay_header_t rh;
+  int reason;
+  uint32_t ipv4_host;
+  tor_addr_t host;
+  origin_circuit_t *ocirc = TO_ORIGIN_CIRCUIT(circ);
+
+  tor_assert(cell);
+  tor_assert(ocirc);
+  tor_assert(circ->purpose == CIRCUIT_PURPOSE_PATH_BIAS_TESTING);
+
+  relay_header_unpack(&rh, cell->payload);
+
+  reason = rh.length > 0 ?
+        get_uint8(cell->payload+RELAY_HEADER_SIZE) : END_STREAM_REASON_MISC;
+
+  if (rh.command == RELAY_COMMAND_END &&
+      reason == END_STREAM_REASON_EXITPOLICY &&
+      ocirc->pathbias_probe_id == rh.stream_id) {
+
+    /* Check length+extract host: It is in network order after the reason code.
+     * See connection_edge_end(). */
+    if (rh.length != 9) { /* reason+ipv4+dns_ttl */
+      log_fn(LOG_PROTOCOL_WARN, LD_PROTOCOL,
+             "Path bias probe response length field is insane (%d).",
+             rh.length);
+      return - END_CIRC_REASON_TORPROTOCOL;
+    }
+
+    ipv4_host = get_uint32(cell->payload+RELAY_HEADER_SIZE+1);
+    tor_addr_from_ipv4n(&host, ipv4_host);
+
+    /* Check nonce */
+    if (memcmp(&host, &ocirc->pathbias_probe_nonce, sizeof(tor_addr_t)) == 0) {
+      ocirc->path_state = PATH_STATE_USE_SUCCEEDED;
+      circuit_mark_for_close(circ, END_CIRC_REASON_FINISHED);
+      log_info(LD_CIRC,
+               "Got valid path bias probe back for circ %d, stream %d.",
+               ocirc->global_identifier, ocirc->pathbias_probe_id);
+      return 0;
+    }
+  }
+  log_info(LD_CIRC,
+             "Got another cell back back on pathbias probe circuit %d: "
+             "Command: %d, Reason: %d, Stream-id: %d",
+             ocirc->global_identifier, rh.command, reason, rh.stream_id);
+  return -1;
+}
+
+/**
  * Check if a circuit was used and/or closed successfully.
  *
  * If we attempted to use the circuit to carry a stream but failed
@@ -1512,18 +1656,26 @@ pathbias_count_build_success(origin_circuit_t *circ)
  *
  * If we *have* successfully used the circuit, or it appears to
  * have been closed by us locally, count it as a success.
+ *
+ * Returns 0 if we're done making decisions with the circ,
+ * or -1 if we want to probe it first.
  */
-void
+int
 pathbias_check_close(origin_circuit_t *ocirc, int reason)
 {
   circuit_t *circ = &ocirc->base_;
 
   if (!pathbias_should_count(ocirc)) {
-    return;
+    return 0;
   }
 
   if (ocirc->path_state == PATH_STATE_BUILD_SUCCEEDED) {
     if (circ->timestamp_dirty) {
+      if (pathbias_send_usable_probe(circ) == 0)
+        return -1;
+      else
+        pathbias_count_unusable(ocirc);
+
       /* Any circuit where there were attempted streams but no successful
        * streams could be bias */
       log_info(LD_CIRC,
@@ -1533,7 +1685,7 @@ pathbias_check_close(origin_circuit_t *ocirc, int reason)
             reason, circ->purpose, ocirc->has_opened,
             circuit_state_to_string(circ->state),
             ocirc->build_state->desired_path_len);
-      pathbias_count_unusable(ocirc);
+
     } else {
       if (reason & END_CIRC_REASON_FLAG_REMOTE) {
         /* Unused remote circ close reasons all could be bias */
@@ -1569,6 +1721,8 @@ pathbias_check_close(origin_circuit_t *ocirc, int reason)
   } else if (ocirc->path_state == PATH_STATE_USE_SUCCEEDED) {
     pathbias_count_successful_close(ocirc);
   }
+
+  return 0;
 }
 
 /**
diff --git a/src/or/circuitbuild.h b/src/or/circuitbuild.h
index 029bdaa47d..b1cedecb94 100644
--- a/src/or/circuitbuild.h
+++ b/src/or/circuitbuild.h
@@ -60,7 +60,8 @@ const node_t *choose_good_entry_server(uint8_t purpose,
 double pathbias_get_extreme_rate(const or_options_t *options);
 int pathbias_get_dropguards(const or_options_t *options);
 void pathbias_count_timeout(origin_circuit_t *circ);
-void pathbias_check_close(origin_circuit_t *circ, int reason);
+int pathbias_check_close(origin_circuit_t *circ, int reason);
+int pathbias_check_probe_response(circuit_t *circ, cell_t *cell);
 
 #endif
 
diff --git a/src/or/circuitlist.c b/src/or/circuitlist.c
index 1a7306292f..90719d2b63 100644
--- a/src/or/circuitlist.c
+++ b/src/or/circuitlist.c
@@ -414,6 +414,8 @@ circuit_purpose_to_controller_string(uint8_t purpose)
       return "MEASURE_TIMEOUT";
     case CIRCUIT_PURPOSE_CONTROLLER:
       return "CONTROLLER";
+    case CIRCUIT_PURPOSE_PATH_BIAS_TESTING:
+      return "PATH_BIAS_TESTING";
 
     default:
       tor_snprintf(buf, sizeof(buf), "UNKNOWN_%d", (int)purpose);
@@ -441,6 +443,7 @@ circuit_purpose_to_controller_hs_state_string(uint8_t purpose)
     case CIRCUIT_PURPOSE_C_MEASURE_TIMEOUT:
     case CIRCUIT_PURPOSE_TESTING:
     case CIRCUIT_PURPOSE_CONTROLLER:
+    case CIRCUIT_PURPOSE_PATH_BIAS_TESTING:
       return NULL;
 
     case CIRCUIT_PURPOSE_INTRO_POINT:
@@ -1356,7 +1359,10 @@ circuit_mark_for_close_(circuit_t *circ, int reason, int line,
   }
 
   if (CIRCUIT_IS_ORIGIN(circ)) {
-    pathbias_check_close(TO_ORIGIN_CIRCUIT(circ), reason);
+    if (pathbias_check_close(TO_ORIGIN_CIRCUIT(circ), reason) == -1) {
+      /* Don't close it yet, we need to test it first */
+      return;
+    }
 
     /* We don't send reasons when closing circuits at the origin. */
     reason = END_CIRC_REASON_NONE;
diff --git a/src/or/circuituse.c b/src/or/circuituse.c
index ffcebfd412..de62100919 100644
--- a/src/or/circuituse.c
+++ b/src/or/circuituse.c
@@ -493,6 +493,8 @@ circuit_expire_building(void)
       cutoff = s_intro_cutoff;
     else if (victim->purpose == CIRCUIT_PURPOSE_C_ESTABLISH_REND)
       cutoff = stream_cutoff;
+    else if (victim->purpose == CIRCUIT_PURPOSE_PATH_BIAS_TESTING)
+      cutoff = close_cutoff;
     else if (TO_ORIGIN_CIRCUIT(victim)->has_opened &&
              victim->state != CIRCUIT_STATE_OPEN)
       cutoff = cannibalized_cutoff;
@@ -581,6 +583,11 @@ circuit_expire_building(void)
               victim->timestamp_dirty > cutoff.tv_sec)
             continue;
           break;
+        case CIRCUIT_PURPOSE_PATH_BIAS_TESTING:
+          /* Open path bias testing circuits are given a long
+           * time to complete the test, but not forever */
+          TO_ORIGIN_CIRCUIT(victim)->path_state = PATH_STATE_USE_FAILED;
+          break;
         case CIRCUIT_PURPOSE_C_INTRODUCING:
           /* We keep old introducing circuits around for
            * a while in parallel, and they can end up "opened".
@@ -652,6 +659,18 @@ circuit_expire_building(void)
           circuit_build_times_set_timeout(&circ_times);
         }
       }
+
+      if (TO_ORIGIN_CIRCUIT(victim)->has_opened &&
+          victim->purpose != CIRCUIT_PURPOSE_PATH_BIAS_TESTING) {
+        /* For path bias: we want to let these guys live for a while
+         * so we get a chance to test them. */
+        log_info(LD_CIRC,
+                 "Allowing cannibalized circuit %d time to finish building as a "
+                 "pathbias testing circ.",
+                 TO_ORIGIN_CIRCUIT(victim)->global_identifier);
+        circuit_change_purpose(victim, CIRCUIT_PURPOSE_PATH_BIAS_TESTING);
+        continue; /* It now should have a longer timeout next time */
+      }
     }
 
     /* If this is a hidden service client circuit which is far enough
@@ -1232,18 +1251,6 @@ circuit_has_opened(origin_circuit_t *circ)
 {
   control_event_circuit_status(circ, CIRC_EVENT_BUILT, 0);
 
-  /* Cannibalized circuits count as used for path bias.
-   * (PURPOSE_GENERAL circs especially, since they are
-   * marked dirty and often go unused after preemptive
-   * building). */
-  // XXX: Cannibalized now use RELAY_EARLY, which is visible
-  // to taggers end-to-end! We really need to probe these instead.
-  // Don't forget to remove this check once that's done!
-  if (circ->has_opened &&
-      circ->build_state->desired_path_len > DEFAULT_ROUTE_LEN) {
-    circ->path_state = PATH_STATE_USE_SUCCEEDED;
-  }
-
   /* Remember that this circuit has finished building. Now if we start
    * it building again later (e.g. by extending it), we will know not
    * to consider its build time. */
diff --git a/src/or/connection_edge.c b/src/or/connection_edge.c
index a0ebfd1397..272955f165 100644
--- a/src/or/connection_edge.c
+++ b/src/or/connection_edge.c
@@ -1661,7 +1661,7 @@ connection_ap_process_natd(entry_connection_t *conn)
 /** Iterate over the two bytes of stream_id until we get one that is not
  * already in use; return it. Return 0 if can't get a unique stream_id.
  */
-static streamid_t
+streamid_t
 get_unique_stream_id_by_circ(origin_circuit_t *circ)
 {
   edge_connection_t *tmpconn;
diff --git a/src/or/connection_edge.h b/src/or/connection_edge.h
index 9f38951f40..82cd95bedf 100644
--- a/src/or/connection_edge.h
+++ b/src/or/connection_edge.h
@@ -90,6 +90,7 @@ int connection_edge_update_circuit_isolation(const entry_connection_t *conn,
                                              origin_circuit_t *circ,
                                              int dry_run);
 void circuit_clear_isolation(origin_circuit_t *circ);
+streamid_t get_unique_stream_id_by_circ(origin_circuit_t *circ);
 
 /** @name Begin-cell flags
  *
diff --git a/src/or/or.h b/src/or/or.h
index 7b8ff705a4..4b4806c62e 100644
--- a/src/or/or.h
+++ b/src/or/or.h
@@ -522,7 +522,9 @@ typedef enum {
 #define CIRCUIT_PURPOSE_TESTING 18
 /** A controller made this circuit and Tor should not use it. */
 #define CIRCUIT_PURPOSE_CONTROLLER 19
-#define CIRCUIT_PURPOSE_MAX_ 19
+/** This circuit is used for path bias probing only */
+#define CIRCUIT_PURPOSE_PATH_BIAS_TESTING 20
+#define CIRCUIT_PURPOSE_MAX_ 20
 /** A catch-all for unrecognized purposes. Currently we don't expect
  * to make or see any circuits with this purpose. */
 #define CIRCUIT_PURPOSE_UNKNOWN 255
@@ -2887,6 +2889,14 @@ typedef struct origin_circuit_t {
    * debug why we are not seeing first hops in some cases. */
   path_state_t path_state : 3;
 
+  /** For path probing. Store the temporary probe stream ID
+   * for response comparison */
+  streamid_t pathbias_probe_id;
+
+  /** For path probing. Store the temporary probe address nonce
+   * for response comparison. */
+  tor_addr_t pathbias_probe_nonce;
+
   /** Set iff this is a hidden-service circuit which has timed out
    * according to our current circuit-build timeout, but which has
    * been kept around because it might still succeed in connecting to
diff --git a/src/or/relay.c b/src/or/relay.c
index f58c5c9c55..a942e44651 100644
--- a/src/or/relay.c
+++ b/src/or/relay.c
@@ -186,7 +186,17 @@ circuit_receive_relay_cell(cell_t *cell, circuit_t *circ,
   }
 
   if (recognized) {
-    edge_connection_t *conn = relay_lookup_conn(circ, cell, cell_direction,
+    edge_connection_t *conn = NULL;
+
+    if (circ->purpose == CIRCUIT_PURPOSE_PATH_BIAS_TESTING) {
+      pathbias_check_probe_response(circ, cell);
+
+      /* We need to drop this cell no matter what to avoid code that expects
+       * a certain purpose (such as the hidserv code). */
+      return 0;
+    }
+
+    conn = relay_lookup_conn(circ, cell, cell_direction,
                                                 layer_hint);
     if (cell_direction == CELL_DIRECTION_OUT) {
       ++stats_n_relay_cells_delivered;
@@ -222,7 +232,15 @@ circuit_receive_relay_cell(cell_t *cell, circuit_t *circ,
   } else {
     log_fn(LOG_PROTOCOL_WARN, LD_OR,
            "Dropping unrecognized inbound cell on origin circuit.");
-    return 0;
+    /* If we see unrecognized cells on path bias testing circs,
+     * it's bad mojo. Those circuits need to die.
+     * XXX: Shouldn't they always die? */
+    if (circ->purpose == CIRCUIT_PURPOSE_PATH_BIAS_TESTING) {
+      TO_ORIGIN_CIRCUIT(circ)->path_state = PATH_STATE_USE_FAILED;
+      return -END_CIRC_REASON_TORPROTOCOL;
+    } else {
+      return 0;
+    }
   }
 
   if (!chan) {