feat(cluster): GCS block dedup capacity + eager-GC (spec-7.2a)

.github/workflows/nightly.yml

@@ -192,6 +192,11 @@ jobs:
           - { name: stage7-marker-async, ranges: "363-363", unit: false, regress: false }
           # t/365 spec-7.1a cross-instance write-write MVCC coordination.
           - { name: stage7-write-write, ranges: "365-365", unit: false, regress: false }
+          # t/366 spec-7.2a GCS block dedup capacity + eager reclaim.  Own shard:
+          # 2-node shared_catalog bring-up + cross-node distinct-read dedup
+          # pressure + injected retransmit-dedup correctness needs the wall
+          # clock.  (t/364 reserved by the parallel spec-7.3 LMS-pool lane.)
+          - { name: stage7-gcs-dedup-capacity, ranges: "366-366", unit: false, regress: false }
     steps:
       - name: Checkout
         uses: actions/checkout@v4

src/backend/cluster/cluster_catalog_bootstrap.c

@@ -35,7 +35,6 @@
  */
 #include "postgres.h"
 
-#include <ctype.h>
 #include <sys/stat.h>
 
 #include "access/transam.h"
@@ -44,6 +43,7 @@
 #include "cluster/cluster_catalog_bootstrap.h"
 #include "cluster/cluster_catalog_migrate.h"
 #include "cluster/cluster_cf_authority.h"
+#include "cluster/cluster_conf.h"
 #include "cluster/cluster_guc.h"
 #include "cluster/cluster_inject.h"
 #include "cluster/cluster_mode.h"
@@ -76,40 +76,6 @@ cluster_catalog_xid_authority_corrupt_fatal(const char *detail)
 }
 
 
-/*
- * Early D6 topology sniff: cluster_conf_load() runs later when shmem exists,
- * but shared_catalog bootstrap must reject multi-node xid_striping=off before
- * it seeds/adopts any authority.  This deliberately counts only [node.N]
- * section headers; the real parser remains the startup SSOT and will still
- * validate syntax, required fields, and node_id consistency later.
- */
-static int
-cluster_catalog_declared_node_count_early(void)
-{
-	const char *path = (cluster_config_file != NULL && cluster_config_file[0] != '\0')
-						   ? cluster_config_file
-						   : "pgrac.conf";
-	FILE *f;
-	char line[1024];
-	int nodes = 0;
-
-	f = AllocateFile(path, "r");
-	if (f == NULL)
-		return 1;
-
-	while (fgets(line, sizeof(line), f) != NULL) {
-		const char *p = line;
-
-		while (*p != '\0' && isspace((unsigned char)*p))
-			p++;
-		if (strncmp(p, "[node.", 6) == 0)
-			nodes++;
-	}
-	FreeFile(f);
-
-	return nodes > 0 ? nodes : 1;
-}
-
 static void
 cluster_catalog_vet_xid_striping_for_shared_catalog(void)
 {
@@ -118,7 +84,14 @@ cluster_catalog_vet_xid_striping_for_shared_catalog(void)
 	if (!cluster_enabled || cluster_xid_striping)
 		return;
 
-	nodes = cluster_catalog_declared_node_count_early();
+	/*
+	 * Early D6 topology sniff: cluster_conf_load() runs later when shmem
+	 * exists, but shared_catalog bootstrap must reject multi-node
+	 * xid_striping=off before it seeds/adopts any authority.  The sniff
+	 * helper lives in cluster_conf.c (spec-7.2a hoisted it for a second
+	 * pre-shmem consumer).
+	 */
+	nodes = cluster_conf_declared_node_count_early();
 	if (nodes > 1)
 		ereport(
 			FATAL,

src/backend/cluster/cluster_conf.c

@@ -44,11 +44,13 @@
  */
 #include "postgres.h"
 
+#include <ctype.h>
 #include <sys/stat.h>
 
 #include "fmgr.h"
 #include "funcapi.h"
 #include "miscadmin.h"
+#include "storage/fd.h"
 #include "storage/shmem.h"
 #include "utils/builtins.h"
 #include "utils/tuplestore.h"
@@ -192,6 +194,51 @@ cluster_conf_node_count(void)
 	return ClusterConfShmem->node_count;
 }
 
+/*
+ * cluster_conf_declared_node_count_early
+ *
+ *	Pre-shmem topology sniff (see cluster_conf.h).  Deliberately counts only
+ *	[node.N] section headers; cluster_conf_load() remains the startup SSOT
+ *	and still validates syntax, required fields, and node_id consistency
+ *	later.  Cached per process: the file is postmaster-static for the life
+ *	of the instance, and shmem size_fn/init_fn must see the same value so
+ *	the reserved region always covers the initialised HTAB (spec-7.2a D4).
+ *	Originally a static helper in cluster_catalog_bootstrap.c (spec-6.14 D6);
+ *	hoisted here when spec-7.2a added a second pre-shmem consumer.
+ */
+int
+cluster_conf_declared_node_count_early(void)
+{
+	static int declared_nodes_cache = -1;
+	const char *path;
+	FILE *f;
+	char line[1024];
+	int nodes = 0;
+
+	if (declared_nodes_cache > 0)
+		return declared_nodes_cache;
+
+	path = (cluster_config_file != NULL && cluster_config_file[0] != '\0') ? cluster_config_file
+																		   : "pgrac.conf";
+
+	f = AllocateFile(path, "r");
+	if (f == NULL)
+		return 1; /* absent file → single-node fallback; do not cache ENOENT */
+
+	while (fgets(line, sizeof(line), f) != NULL) {
+		const char *p = line;
+
+		while (*p != '\0' && isspace((unsigned char)*p))
+			p++;
+		if (strncmp(p, "[node.", 6) == 0)
+			nodes++;
+	}
+	FreeFile(f);
+
+	declared_nodes_cache = nodes > 0 ? nodes : 1;
+	return declared_nodes_cache;
+}
+
 
 /* ============================================================
  * INI parser.

src/backend/cluster/cluster_debug.c

@@ -1938,6 +1938,15 @@ dump_gcs(ReturnSetInfo *rsinfo)
 			 fmt_int64((int64)cluster_gcs_get_block_dedup_collision_count()));
 	emit_row(rsinfo, "gcs", "dedup_full_count",
 			 fmt_int64((int64)cluster_gcs_get_block_dedup_full_count()));
+	/* spec-7.2a D5:  3 NEW dedup capacity/occupancy rows.  entry_count /
+	 * max_entries give the saturation ratio;  evict_count counts eager
+	 * reclaim + TTL-sweep removals (dump_gcs gcs-category 85 -> 88). */
+	emit_row(rsinfo, "gcs", "dedup_entry_count",
+			 fmt_int64((int64)cluster_gcs_get_block_dedup_entry_count()));
+	emit_row(rsinfo, "gcs", "dedup_evict_count",
+			 fmt_int64((int64)cluster_gcs_get_block_dedup_evict_count()));
+	emit_row(rsinfo, "gcs", "dedup_max_entries",
+			 fmt_int64((int64)cluster_gcs_get_block_dedup_max_entries()));
 	emit_row(rsinfo, "gcs", "epoch_invalidate_wake_count",
 			 fmt_int64((int64)cluster_gcs_get_block_epoch_invalidate_wake_count()));
 	emit_row(rsinfo, "gcs", "stale_reply_drop_count",

src/backend/cluster/cluster_gcs_block.c

@@ -4652,13 +4652,31 @@ build_and_send_reply: {
 		 * active on that retry).  Useful for driving the
 		 * retransmit_send_count + dedup_hit_count TAP surfaces.
 		 */
-	CLUSTER_INJECTION_POINT("cluster-gcs-block-drop-reply-before-send");
-	if (cluster_injection_should_skip("cluster-gcs-block-drop-reply-before-send")) {
-		gcs_block_release_ship_image(block_payload_release_cb, block_payload_release_arg);
-		block_payload_release_cb = NULL;
-		block_payload_release_arg = NULL;
-		pfree(buf);
-		return;
+	/*
+	 * spec-7.2a: gate the drop-reply dispatch on the test target relfilenode.
+	 * A :skipn:N count is per-process global; without this gate an unrelated
+	 * (catalog / internal) block ship consumes the countdown before the test's
+	 * user-relation ship reaches the point.  Gating the CLUSTER_INJECTION_POINT
+	 * itself (not just should_skip) ensures only matching ships consume the
+	 * count.  0 (default) keeps the un-targeted behaviour for spec-2.34 tests.
+	 *
+	 * Current TAP coverage uses target=0 only (shared_catalog remaps the
+	 * catalog-visible relfilenode to a different physical relNumber, so SQL
+	 * cannot name the shipped block); the non-zero filter is reserved for
+	 * precise spec-2.34-style targeting on non-shared-catalog rigs and is not
+	 * yet exercised by any test.
+	 */
+	if (cluster_gcs_block_drop_target_relfilenode == 0
+		|| BufTagGetRelNumber(&req->tag)
+			   == (RelFileNumber)cluster_gcs_block_drop_target_relfilenode) {
+		CLUSTER_INJECTION_POINT("cluster-gcs-block-drop-reply-before-send");
+		if (cluster_injection_should_skip("cluster-gcs-block-drop-reply-before-send")) {
+			gcs_block_release_ship_image(block_payload_release_cb, block_payload_release_arg);
+			block_payload_release_cb = NULL;
+			block_payload_release_arg = NULL;
+			pfree(buf);
+			return;
+		}
 	}
 
 	{
@@ -7070,6 +7088,31 @@ cluster_gcs_get_block_dedup_full_count(void)
 	return cluster_gcs_block_dedup_get_full_count();
 }
 
+/*
+ * spec-7.2a D5:  dedup capacity/occupancy observability wrappers.  The
+ * entry_count wrapper reads the historical _get_in_flight_count accessor,
+ * whose backing counter (entry_count) tracks every live entry (in-flight
+ * slots plus completed cached replies) -- that live total is what dump_gcs
+ * surfaces as dedup_entry_count for the saturation ratio.
+ */
+uint64
+cluster_gcs_get_block_dedup_entry_count(void)
+{
+	return cluster_gcs_block_dedup_get_in_flight_count();
+}
+
+uint64
+cluster_gcs_get_block_dedup_evict_count(void)
+{
+	return cluster_gcs_block_dedup_get_evict_count();
+}
+
+uint64
+cluster_gcs_get_block_dedup_max_entries(void)
+{
+	return cluster_gcs_block_dedup_get_max_entries();
+}
+
 
 /* ============================================================
  * PGRAC: spec-2.34 D4 — eager wake on epoch advance.