gossipd: remove gossip entirely if we hit a problem on load.

The crashes in #2750 are mostly caused by us trying to partially truncate
the store.  The simplest fix for release is to discard the whole thing if
we detect a problem.

This is a workaround: it'd be far nicer to try to recover.

Fixes: #2750
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

gossipd/gossip_store.c

@@ -525,44 +525,6 @@ int gossip_store_readonly_fd(struct gossip_store *gs)
 	return fd;
 }
 
-/* If we ever truncated, we might have a dangling entries. */
-static void cleanup_truncated_store(struct routing_state *rstate,
-				    struct gossip_store *gs,
-				    u32 chan_ann_off)
-{
-	size_t num;
-	u32 index;
-
-	/* channel_announce with no channel_amount. */
-	if (chan_ann_off) {
-		status_unusual("Deleting un-amounted channel_announcement @%u",
-			       chan_ann_off);
-		delete_by_index(gs, chan_ann_off, WIRE_CHANNEL_ANNOUNCEMENT);
-	}
-
-	num = 0;
-	while ((index = remove_unfinalized_node_announce(rstate)) != 0) {
-		delete_by_index(gs, index, WIRE_NODE_ANNOUNCEMENT);
-		num++;
-	}
-	if (num)
-		status_unusual("Deleted %zu unfinalized node_announcements",
-			       num);
-
-	num = 0;
-	while ((index = remove_unupdated_channel_announce(rstate)) != 0) {
-		u32 next;
-
-		/* Delete announcement and channel amount, too */
-		next = delete_by_index(gs, index, WIRE_CHANNEL_ANNOUNCEMENT);
-		delete_by_index(gs, next, WIRE_GOSSIP_STORE_CHANNEL_AMOUNT);
-		num++;
-	}
-	if (num)
-		status_unusual("Deleted %zu unupdated channel_announcements",
-			       num);
-}
-
 bool gossip_store_load(struct routing_state *rstate, struct gossip_store *gs)
 {
 	struct gossip_hdr hdr;
@@ -672,25 +634,36 @@ bool gossip_store_load(struct routing_state *rstate, struct gossip_store *gs)
 		clean_tmpctx();
 	}
 
+	if (chan_ann) {
+		status_unusual("gossip_store: dangling channel_announcement");
+		goto truncate_nomsg;
+	}
+
+	bad = unfinalized_entries(tmpctx, rstate);
+	if (bad) {
+		status_unusual("gossip_store: %s", bad);
+		goto truncate_nomsg;
+	}
+
 	/* If last timestamp is within 24 hours, say we're OK. */
 	contents_ok = (last_timestamp >= time_now().ts.tv_sec - 24*3600);
 	goto out;
 
 truncate:
-	status_unusual("gossip_store: %s (%s) truncating to %"PRIu64,
-		       bad, tal_hex(msg, msg), gs->len);
+	status_unusual("gossip_store: %s (%s) truncating",
+		       bad, tal_hex(msg, msg));
+
 truncate_nomsg:
-	/* FIXME: We would like to truncate to known_good, except we would
-	 * miss channel_delete msgs.  If we put block numbers into the store
-	 * as we process them, we can know how far we need to roll back if we
-	 * truncate the store */
-	if (ftruncate(gs->fd, gs->len) != 0)
+	/* FIXME: Debug partial truncate case. */
+	if (ftruncate(gs->fd, 1) != 0)
 		status_failed(STATUS_FAIL_INTERNAL_ERROR,
 			      "Truncating store: %s", strerror(errno));
+	remove_all_gossip(rstate);
+	gs->count = gs->deleted = 0;
+	gs->len = 1;
 	contents_ok = false;
 out:
 	gs->writable = true;
-	cleanup_truncated_store(rstate, gs, chan_ann ? chan_ann_off : 0);
 	status_trace("total store load time: %"PRIu64" msec",
 		     time_to_msec(time_between(time_now(), start)));
 	status_trace("gossip_store: Read %zu/%zu/%zu/%zu cannounce/cupdate/nannounce/cdelete from store (%zu deleted) in %"PRIu64" bytes",

gossipd/routing.c

@@ -2583,42 +2583,63 @@ struct timeabs gossip_time_now(const struct routing_state *rstate)
 	return time_now();
 }
 
-/* gossip_store wants to delete any dangling node_announcement msgs */
-u32 remove_unfinalized_node_announce(struct routing_state *rstate)
-{
-	/* We're only interested in node_announcement we caught. */
-	for (;;) {
-		struct pending_node_announce *pna;
-		struct pending_node_map_iter it;
-
-		pna = pending_node_map_first(rstate->pending_node_map, &it);
-		if (!pna)
-			return 0;
-
-		/* This will be deleted by the associated unupdated_channel; just
-		 * remove from map for now. */
-		pending_node_map_del(rstate->pending_node_map, pna);
-		if (!pna->node_announcement)
-			continue;
-
-		assert(pna->index);
-		return pna->index;
-	}
-}
-
-/* gossip_store wants to delete any dangling channel_announcement msgs */
-u32 remove_unupdated_channel_announce(struct routing_state *rstate)
+const char *unfinalized_entries(const tal_t *ctx, struct routing_state *rstate)
 {
 	struct unupdated_channel *uc;
 	u64 index;
+	struct pending_node_announce *pna;
+	struct pending_node_map_iter it;
 
 	uc = uintmap_first(&rstate->unupdated_chanmap, &index);
-	if (!uc)
-		return 0;
+	if (uc)
+		return tal_fmt(ctx, "Unupdated channel_announcement at %u",
+			       uc->index);
 
-	assert(uc->index);
-	index = uc->index;
+	pna = pending_node_map_first(rstate->pending_node_map, &it);
+	if (pna)
+		return tal_fmt(ctx, "Waiting node_announcement at %u",
+			       pna->index);
 
-	tal_free(uc);
-	return index;
+	return NULL;
+}
+
+/* Gossip store was corrupt, forget anything we loaded. */
+void remove_all_gossip(struct routing_state *rstate)
+{
+	struct node *n;
+	struct node_map_iter nit;
+	struct chan *c;
+	struct unupdated_channel *uc;
+	u64 index;
+	struct pending_cannouncement *pca;
+	struct pending_cannouncement_map_iter pit;
+	struct pending_node_map_iter pnait;
+
+	/* We don't want them to try to delete from store, so do this
+	 * manually. */
+	while ((n = node_map_first(rstate->nodes, &nit)) != NULL) {
+		tal_del_destructor2(n, destroy_node, rstate);
+		if (node_uses_chan_map(n))
+			chan_map_clear(&n->chans.map);
+		node_map_del(rstate->nodes, n);
+		tal_free(n);
+	}
+
+	/* Now free all the channels. */
+	while ((c = uintmap_first(&rstate->chanmap, &index)) != NULL) {
+		uintmap_del(&rstate->chanmap, index);
+
+		/* Remove from local_disabled_map if it's there. */
+		chan_map_del(&rstate->local_disabled_map, c);
+		tal_free(c);
+	}
+
+	while ((uc = uintmap_first(&rstate->unupdated_chanmap, &index)) != NULL)
+		tal_free(uc);
+
+	while ((pca = pending_cannouncement_map_first(&rstate->pending_cannouncements, &pit)) != NULL)
+		tal_free(pca);
+
+	/* Freeing unupdated chanmaps should empty this */
+	assert(pending_node_map_first(rstate->pending_node_map, &pnait) == NULL);
 }

gossipd/routing.h

@@ -437,12 +437,8 @@ struct wireaddr *read_addresses(const tal_t *ctx, const u8 *ser);
 void remove_channel_from_store(struct routing_state *rstate,
 			       struct chan *chan);
 
-/* gossip_store wants to delete any dangling entries immediately after
- * load; return 0 if no more, otherwise index into store.
- *
- * Must call remove_unfinalized_node_announce first, because removing
- * unupdated channels may delete associatd node_announcements. */
-u32 remove_unfinalized_node_announce(struct routing_state *rstate);
-u32 remove_unupdated_channel_announce(struct routing_state *rstate);
+/* Returns an error string if there are unfinalized entries after load */
+const char *unfinalized_entries(const tal_t *ctx, struct routing_state *rstate);
 
+void remove_all_gossip(struct routing_state *rstate);
 #endif /* LIGHTNING_GOSSIPD_ROUTING_H */

tests/test_gossip.py

@@ -945,8 +945,8 @@ def test_gossip_store_load_amount_truncated(node_factory):
 
     l1.start()
     # May preceed the Started msg waited for in 'start'.
-    wait_for(lambda: l1.daemon.is_in_log(r'Deleting un-amounted channel_announcement @1'))
-    wait_for(lambda: l1.daemon.is_in_log(r'gossip_store: Read 0/0/0/0 cannounce/cupdate/nannounce/cdelete from store \(1 deleted\) in 445 bytes'))
+    wait_for(lambda: l1.daemon.is_in_log(r'gossip_store: dangling channel_announcement'))
+    wait_for(lambda: l1.daemon.is_in_log(r'gossip_store: Read 0/0/0/0 cannounce/cupdate/nannounce/cdelete from store \(0 deleted\) in 1 bytes'))
     assert not l1.daemon.is_in_log('gossip_store.*truncating')
 
     # Extra sanity check if we can.