mdbx: исправление и рефакторинг цикла чтения мета-страниц при старте читающих транзакций.

Сценарий достаточно запутанный/сложный.

src/api-env.c

@@ -1274,7 +1274,7 @@ __cold int mdbx_env_set_geometry(MDBX_env *env, intptr_t size_lower,
       uint64_t timestamp = 0;
       while ("workaround for "
              "https://libmdbx.dqdkfa.ru/dead-github/issues/269") {
-        rc = coherency_check_head(env->basal_txn, head, &timestamp);
+        rc = coherency_fetch_head(env->basal_txn, head, &timestamp);
         if (likely(rc == MDBX_SUCCESS))
           break;
         if (unlikely(rc != MDBX_RESULT_TRUE))

src/coherency.c

@@ -152,18 +152,20 @@ __cold int coherency_timeout(uint64_t *timestamp, intptr_t pgno,
 
 /* check with timeout as the workaround
  * for https://libmdbx.dqdkfa.ru/dead-github/issues/269 */
-__hot int coherency_check_head(MDBX_txn *txn, const meta_ptr_t head,
+__hot int coherency_fetch_head(MDBX_txn *txn, const meta_ptr_t head,
                                uint64_t *timestamp) {
   /* Copy the DB info and flags */
-  txn->geo = head.ptr_v->geometry;
+  txn->txnid = head.txnid;
+  txn->geo = head.ptr_c->geometry;
   memcpy(txn->dbs, &head.ptr_c->trees, sizeof(head.ptr_c->trees));
   STATIC_ASSERT(sizeof(head.ptr_c->trees) == CORE_DBS * sizeof(tree_t));
   VALGRIND_MAKE_MEM_UNDEFINED(txn->dbs + CORE_DBS,
                               txn->env->max_dbi - CORE_DBS);
-  txn->canary = head.ptr_v->canary;
+  txn->canary = head.ptr_c->canary;
 
   if (unlikely(!coherency_check(txn->env, head.txnid, txn->dbs, head.ptr_v,
-                                *timestamp == 0)))
+                                *timestamp == 0) ||
+               txn->txnid != meta_txnid(head.ptr_v)))
     return coherency_timeout(timestamp, -1, txn->env);
 
   if (unlikely(txn->dbs[FREE_DBI].flags != MDBX_INTEGERKEY)) {

src/proto.h

@@ -113,7 +113,7 @@ MDBX_INTERNAL int __must_check_result tbl_setup(const MDBX_env *env,
 MDBX_INTERNAL bool coherency_check_meta(const MDBX_env *env,
                                         const volatile meta_t *meta,
                                         bool report);
-MDBX_INTERNAL int coherency_check_head(MDBX_txn *txn, const meta_ptr_t head,
+MDBX_INTERNAL int coherency_fetch_head(MDBX_txn *txn, const meta_ptr_t head,
                                        uint64_t *timestamp);
 MDBX_INTERNAL int coherency_check_written(const MDBX_env *env,
                                           const txnid_t txnid,

src/txn.c

@@ -982,7 +982,7 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
           likely(env->stuck_meta < 0)
               ? /* regular */ meta_recent(env, &troika)
               : /* recovery mode */ meta_ptr(env, env->stuck_meta);
-      if (likely(r)) {
+      if (likely(r != nullptr)) {
         safe64_reset(&r->txnid, true);
         atomic_store32(&r->snapshot_pages_used,
                        head.ptr_v->geometry.first_unallocated, mo_Relaxed);
@@ -1005,46 +1005,57 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
       }
       jitter4testing(true);
 
-      /* Snap the state from current meta-head */
+      if (unlikely(meta_should_retry(env, &troika))) {
-      txn->txnid = head.txnid;
+      retry:
-      if (likely(env->stuck_meta < 0) &&
+        if (likely(++loop < 42)) {
-          unlikely(meta_should_retry(env, &troika) ||
-                   head.txnid < atomic_load64(&env->lck->cached_oldest,
-                                              mo_AcquireRelease))) {
-        if (unlikely(++loop > 42)) {
-          ERROR("bailout waiting for valid snapshot (%s)",
-                "metapages are too volatile");
-          rc = MDBX_PROBLEM;
-          txn->txnid = INVALID_TXNID;
-          if (likely(r))
-            safe64_reset(&r->txnid, true);
-          goto bailout;
-        }
           timestamp = 0;
           continue;
         }
-
+        ERROR("bailout waiting for valid snapshot (%s)",
-      rc = coherency_check_head(txn, head, &timestamp);
+              "meta-pages are too volatile");
-      jitter4testing(false);
+        rc = MDBX_PROBLEM;
-      if (likely(rc == MDBX_SUCCESS))
+        goto read_failed;
-        break;
-
-      if (unlikely(rc != MDBX_RESULT_TRUE)) {
-        txn->txnid = INVALID_TXNID;
-        if (likely(r))
-          safe64_reset(&r->txnid, true);
-        goto bailout;
       }
+
+      /* Snap the state from current meta-head */
+      rc = coherency_fetch_head(txn, head, &timestamp);
+      jitter4testing(false);
+      if (unlikely(rc != MDBX_SUCCESS)) {
+        if (rc == MDBX_RESULT_TRUE)
+          goto retry;
+        else
+          goto read_failed;
+      }
+
+      const uint64_t snap_oldest =
+          atomic_load64(&env->lck->cached_oldest, mo_AcquireRelease);
+      if (unlikely(txn->txnid < snap_oldest)) {
+        if (env->stuck_meta < 0)
+          goto retry;
+        ERROR("target meta-page %i is referenced to an obsolete MVCC-snapshot "
+              "%" PRIaTXN " < cached-oldest %" PRIaTXN,
+              env->stuck_meta, txn->txnid, snap_oldest);
+        rc = MDBX_MVCC_RETARDED;
+        goto read_failed;
+      }
+
+      if (likely(r != nullptr) &&
+          unlikely(txn->txnid != atomic_load64(&r->txnid, mo_Relaxed)))
+        goto retry;
+      break;
     }
 
     if (unlikely(txn->txnid < MIN_TXNID || txn->txnid > MAX_TXNID)) {
       ERROR("%s", "environment corrupted by died writer, must shutdown!");
-      if (likely(r))
-        safe64_reset(&r->txnid, true);
-      txn->txnid = INVALID_TXNID;
       rc = MDBX_CORRUPTED;
+    read_failed:
+      txn->txnid = INVALID_TXNID;
+      if (likely(r != nullptr))
+        safe64_reset(&r->txnid, true);
       goto bailout;
     }
+
+    tASSERT(txn, rc == MDBX_SUCCESS);
     ENSURE(env,
            txn->txnid >=
                /* paranoia is appropriate here */ env->lck->cached_oldest.weak);
@@ -1092,14 +1103,14 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
     const meta_ptr_t head = meta_recent(env, &txn->tw.troika);
     uint64_t timestamp = 0;
     while ("workaround for https://libmdbx.dqdkfa.ru/dead-github/issues/269") {
-      rc = coherency_check_head(txn, head, &timestamp);
+      rc = coherency_fetch_head(txn, head, &timestamp);
       if (likely(rc == MDBX_SUCCESS))
         break;
       if (unlikely(rc != MDBX_RESULT_TRUE))
         goto bailout;
     }
-    eASSERT(env, meta_txnid(head.ptr_v) == head.txnid);
+    eASSERT(env, meta_txnid(head.ptr_v) == txn->txnid);
-    txn->txnid = safe64_txnid_next(head.txnid);
+    txn->txnid = safe64_txnid_next(txn->txnid);
     if (unlikely(txn->txnid > MAX_TXNID)) {
       rc = MDBX_TXN_FULL;
       ERROR("txnid overflow, raise %d", rc);