From 72ff0be6857c0f714b2d588bd108afb2cab3de43 Mon Sep 17 00:00:00 2001
From: Lucian Petrut <lpetrut@cloudbasesolutions.com>
Date: Thu, 5 Sep 2024 07:57:33 +0000
Subject: [PATCH] Propagate raft_recover errors and improve tracing

At the moment, the raft_recover error messages are not propagated,
so the caller gets a cryptic error code (1), without further
details.

This change ensures that the error messages are propagated and adds
a few extra trace messages.

Furthermore, tracing is disabled unless we start the node (not the
case when we're actually trying to recover it). For now, we're adding a
"dqliteTracingMaybeEnable" call in "dqlite_node_recover_ext".
---
 src/raft/raft.c |  1 +
 src/raft/uv.c   | 11 +++++++++++
 src/server.c    |  7 +++++++
 3 files changed, 19 insertions(+)

diff --git a/src/raft/raft.c b/src/raft/raft.c
index ce3996b90..b0dd7d39f 100644
--- a/src/raft/raft.c
+++ b/src/raft/raft.c
@@ -227,6 +227,7 @@ int raft_recover(struct raft *r, const struct raft_configuration *conf)
 
 	rv = r->io->recover(r->io, conf);
 	if (rv != 0) {
+		ErrMsgTransfer(r->io->errmsg, r->errmsg, "io");
 		return rv;
 	}
 
diff --git a/src/raft/uv.c b/src/raft/uv.c
index f1450e742..b093715e1 100644
--- a/src/raft/uv.c
+++ b/src/raft/uv.c
@@ -324,6 +324,7 @@ static int uvFilterSegments(struct uv *uv,
 		ErrMsgPrintf(uv->io->errmsg,
 			     "closed segment %s is past last snapshot %s",
 			     segment->filename, snapshot_filename);
+		tracef("corrupted raft state, error: %s", uv->io->errmsg);
 		return RAFT_CORRUPT;
 	}
 
@@ -369,6 +370,7 @@ static int uvLoadSnapshotAndEntries(struct uv *uv,
 	rv = UvList(uv, &snapshots, &n_snapshots, &segments, &n_segments,
 		    uv->io->errmsg);
 	if (rv != 0) {
+		tracef("failed to list snapshots and segments, error: %d", rv);
 		goto err;
 	}
 
@@ -377,12 +379,14 @@ static int uvLoadSnapshotAndEntries(struct uv *uv,
 		char snapshot_filename[UV__FILENAME_LEN];
 		*snapshot = RaftHeapMalloc(sizeof **snapshot);
 		if (*snapshot == NULL) {
+			tracef("malloc failed");
 			rv = RAFT_NOMEM;
 			goto err;
 		}
 		rv = UvSnapshotLoad(uv, &snapshots[n_snapshots - 1], *snapshot,
 				    uv->io->errmsg);
 		if (rv != 0) {
+			tracef("snapshot load failed: %d", rv);
 			RaftHeapFree(*snapshot);
 			*snapshot = NULL;
 			goto err;
@@ -401,6 +405,7 @@ static int uvLoadSnapshotAndEntries(struct uv *uv,
 		rv = uvFilterSegments(uv, (*snapshot)->index, snapshot_filename,
 				      &segments, &n_segments);
 		if (rv != 0) {
+			tracef("failed to filter segments: %d", rv);
 			goto err;
 		}
 		if (segments != NULL) {
@@ -420,6 +425,7 @@ static int uvLoadSnapshotAndEntries(struct uv *uv,
 		rv = uvSegmentLoadAll(uv, *start_index, segments, n_segments,
 				      entries, n);
 		if (rv != 0) {
+			tracef("failed to load all segments: %d", rv);
 			goto err;
 		}
 
@@ -447,6 +453,9 @@ static int uvLoadSnapshotAndEntries(struct uv *uv,
 
 err:
 	assert(rv != 0);
+	tracef("auto-recovery: %d, load depth: %d, error: %s",
+		   uv->auto_recovery, depth, uv->io->errmsg);
+
 	if (*snapshot != NULL) {
 		snapshotDestroy(*snapshot);
 		*snapshot = NULL;
@@ -583,6 +592,7 @@ static int uvRecover(struct raft_io *io, const struct raft_configuration *conf)
 	rv = uvLoadSnapshotAndEntries(uv, &snapshot, &start_index, &entries,
 				      &n_entries, 0);
 	if (rv != 0) {
+		tracef("failed to load raft snapshot and entries, error: %d", rv);
 		return rv;
 	}
 
@@ -599,6 +609,7 @@ static int uvRecover(struct raft_io *io, const struct raft_configuration *conf)
 
 	rv = uvSegmentCreateClosedWithConfiguration(uv, next_index, conf);
 	if (rv != 0) {
+		tracef("failed to create segment, error: %d", rv);
 		return rv;
 	}
 
diff --git a/src/server.c b/src/server.c
index 7e44b1c49..3f4f33fa5 100644
--- a/src/server.c
+++ b/src/server.c
@@ -1022,6 +1022,7 @@ int dqlite_node_recover_ext(dqlite_node *n,
 			    struct dqlite_node_info_ext infos[],
 			    int n_info)
 {
+	dqliteTracingMaybeEnable(true);
 	tracef("dqlite node recover ext");
 	struct raft_configuration configuration;
 	int i;
@@ -1031,6 +1032,7 @@ int dqlite_node_recover_ext(dqlite_node *n,
 	for (i = 0; i < n_info; i++) {
 		struct dqlite_node_info_ext *info = &infos[i];
 		if (!node_info_valid(info)) {
+			tracef("invalid node info");
 			rv = DQLITE_MISUSE;
 			goto out;
 		}
@@ -1040,6 +1042,7 @@ int dqlite_node_recover_ext(dqlite_node *n,
 		rv = raft_configuration_add(&configuration, info->id, address,
 					    raft_role);
 		if (rv != 0) {
+			tracef("unable to add server to raft configuration, error: %d", rv);
 			assert(rv == RAFT_NOMEM);
 			rv = DQLITE_NOMEM;
 			goto out;
@@ -1049,11 +1052,15 @@ int dqlite_node_recover_ext(dqlite_node *n,
 	int lock_fd;
 	rv = acquire_dir(n->config.raft_dir, &lock_fd);
 	if (rv != 0) {
+		tracef("couldn't acquire lock, error: %d", rv);
 		goto out;
 	}
 
 	rv = raft_recover(&n->raft, &configuration);
 	if (rv != 0) {
+		tracef("raft recovery failed, error: %d", rv);
+		snprintf(n->errmsg, DQLITE_ERRMSG_BUF_SIZE, "raft_recover(): %s",
+			 raft_errmsg(&n->raft));
 		rv = DQLITE_ERROR;
 		goto out;
 	}