merge bitcoin#21726: Improve Indices on pruned nodes via prune blockers

Author: kwvg

src/index/base.cpp

@@ -62,9 +62,9 @@ bool BaseIndex::Init()
     LOCK(cs_main);
     CChain& active_chain = m_chainstate->m_chain;
     if (locator.IsNull()) {
-        m_best_block_index = nullptr;
+        SetBestBlockIndex(nullptr);
     } else {
-        m_best_block_index = m_chainstate->FindForkInGlobalIndex(locator);
+        SetBestBlockIndex(m_chainstate->FindForkInGlobalIndex(locator));
     }
 
     // Note: this will latch to true immediately if the user starts up with an empty
@@ -76,11 +76,7 @@ bool BaseIndex::Init()
         if (!m_best_block_index) {
             // index is not built yet
             // make sure we have all block data back to the genesis
-            const CBlockIndex* block = active_chain.Tip();
-            while (block->pprev && (block->pprev->nStatus & BLOCK_HAVE_DATA)) {
-                block = block->pprev;
-            }
-            prune_violation = block != active_chain.Genesis();
+            prune_violation = GetFirstStoredBlock(active_chain.Tip()) != active_chain.Genesis();
         }
         // in case the index has a best block set and is not fully synced
         // check if we have the required blocks to continue building the index
@@ -138,7 +134,7 @@ void BaseIndex::ThreadSync()
         std::chrono::steady_clock::time_point last_locator_write_time{0s};
         while (true) {
             if (m_interrupt) {
-                m_best_block_index = pindex;
+                SetBestBlockIndex(pindex);
                 // No need to handle errors in Commit. If it fails, the error will be already be
                 // logged. The best way to recover is to continue, as index cannot be corrupted by
                 // a missed commit to disk for an advanced index state.
@@ -150,7 +146,7 @@ void BaseIndex::ThreadSync()
                 LOCK(cs_main);
                 const CBlockIndex* pindex_next = NextSyncBlock(pindex, m_chainstate->m_chain);
                 if (!pindex_next) {
-                    m_best_block_index = pindex;
+                    SetBestBlockIndex(pindex);
                     m_synced = true;
                     // No need to handle errors in Commit. See rationale above.
                     Commit();
@@ -172,7 +168,7 @@ void BaseIndex::ThreadSync()
             }
 
             if (last_locator_write_time + SYNC_LOCATOR_WRITE_INTERVAL < current_time) {
-                m_best_block_index = pindex;
+                SetBestBlockIndex(pindex);
                 last_locator_write_time = current_time;
                 // No need to handle errors in Commit. See rationale above.
                 Commit();
@@ -230,10 +226,10 @@ bool BaseIndex::Rewind(const CBlockIndex* current_tip, const CBlockIndex* new_ti
     // out of sync may be possible but a users fault.
     // In case we reorg beyond the pruned depth, ReadBlockFromDisk would
     // throw and lead to a graceful shutdown
-    m_best_block_index = new_tip;
+    SetBestBlockIndex(new_tip);
     if (!Commit()) {
         // If commit fails, revert the best block index to avoid corruption.
-        m_best_block_index = current_tip;
+        SetBestBlockIndex(current_tip);
         return false;
     }
 
@@ -274,7 +270,7 @@ void BaseIndex::BlockConnected(const std::shared_ptr<const CBlock>& block, const
     }
 
     if (WriteBlock(*block, pindex)) {
-        m_best_block_index = pindex;
+        SetBestBlockIndex(pindex);
     } else {
         FatalError("%s: Failed to write block %s to index",
                    __func__, pindex->GetBlockHash().ToString());
@@ -381,3 +377,14 @@ IndexSummary BaseIndex::GetSummary() const
     summary.best_block_height = m_best_block_index ? m_best_block_index.load()->nHeight : 0;
     return summary;
 }
+
+void BaseIndex::SetBestBlockIndex(const CBlockIndex* block) {
+    assert(!fPruneMode || AllowPrune());
+
+    m_best_block_index = block;
+    if (AllowPrune() && block) {
+        PruneLockInfo prune_lock;
+        prune_lock.height_first = block->nHeight;
+        WITH_LOCK(::cs_main, m_chainstate->m_blockman.UpdatePruneLock(GetName(), prune_lock));
+    }
+}

src/index/base.h

@@ -81,6 +81,9 @@ class BaseIndex : public CValidationInterface
     /// to a chain reorganization), the index must halt until Commit succeeds or else it could end up
     /// getting corrupted.
     bool Commit();
+
+    virtual bool AllowPrune() const = 0;
+
 protected:
     CChainState* m_chainstate{nullptr};
 
@@ -109,6 +112,9 @@ class BaseIndex : public CValidationInterface
     /// Get the name of the index for display in logs.
     virtual const char* GetName() const = 0;
 
+    /// Update the internal best block index as well as the prune lock.
+    void SetBestBlockIndex(const CBlockIndex* block);
+
 public:
     /// Destructor interrupts sync thread if running and blocks until it exits.
     virtual ~BaseIndex();

src/index/blockfilterindex.h

@@ -39,6 +39,8 @@ class BlockFilterIndex final : public BaseIndex
     /** cache of block hash to filter header, to avoid disk access when responding to getcfcheckpt. */
     std::unordered_map<uint256, uint256, FilterHeaderHasher> m_headers_cache GUARDED_BY(m_cs_headers_cache);
 
+    bool AllowPrune() const override { return true; }
+
 protected:
     bool Init() override;
 

src/index/coinstatsindex.h

@@ -36,6 +36,8 @@ class CoinStatsIndex final : public BaseIndex
 
     bool ReverseBlock(const CBlock& block, const CBlockIndex* pindex);
 
+    bool AllowPrune() const override { return true; }
+
 protected:
     bool Init() override;
 

src/index/txindex.h

@@ -20,6 +20,8 @@ class TxIndex final : public BaseIndex
 private:
     const std::unique_ptr<DB> m_db;
 
+    bool AllowPrune() const override { return false; }
+
 protected:
     bool WriteBlock(const CBlock& block, const CBlockIndex* pindex) override;
 

src/init.cpp

@@ -538,7 +538,7 @@ void SetupServerArgs(ArgsManager& argsman)
         -GetNumCores(), MAX_SCRIPTCHECK_THREADS, DEFAULT_SCRIPTCHECK_THREADS), ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
     argsman.AddArg("-persistmempool", strprintf("Whether to save the mempool on shutdown and load on restart (default: %u)", DEFAULT_PERSIST_MEMPOOL), ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
     argsman.AddArg("-pid=<file>", strprintf("Specify pid file. Relative paths will be prefixed by a net-specific datadir location. (default: %s)", BITCOIN_PID_FILENAME), ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
-    argsman.AddArg("-prune=<n>", strprintf("Reduce storage requirements by enabling pruning (deleting) of old blocks. This allows the pruneblockchain RPC to be called to delete specific blocks, and enables automatic pruning of old blocks if a target size in MiB is provided. This mode is incompatible with -txindex, -coinstatsindex, -rescan and -disablegovernance=false. "
+    argsman.AddArg("-prune=<n>", strprintf("Reduce storage requirements by enabling pruning (deleting) of old blocks. This allows the pruneblockchain RPC to be called to delete specific blocks, and enables automatic pruning of old blocks if a target size in MiB is provided. This mode is incompatible with -txindex, -rescan and -disablegovernance=false. "
             "Warning: Reverting this setting requires re-downloading the entire blockchain. "
             "(default: 0 = disable pruning blocks, 1 = allow manual pruning via RPC, >%u = automatically prune block files to stay under the specified target size in MiB)", MIN_DISK_SPACE_FOR_BLOCK_FILES / 1024 / 1024), ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
     argsman.AddArg("-settings=<file>", strprintf("Specify path to dynamic settings data file. Can be disabled with -nosettings. File is written at runtime and not meant to be edited by users (use %s instead for custom settings). Relative paths will be prefixed by datadir location. (default: %s)", BITCOIN_CONF_FILENAME, BITCOIN_SETTINGS_FILENAME), ArgsManager::ALLOW_ANY, OptionsCategory::OPTIONS);
@@ -1159,8 +1159,6 @@ bool AppInitParameterInteraction(const ArgsManager& args)
     if (args.GetArg("-prune", 0)) {
         if (args.GetBoolArg("-txindex", DEFAULT_TXINDEX))
             return InitError(_("Prune mode is incompatible with -txindex."));
-        if (args.GetBoolArg("-coinstatsindex", DEFAULT_COINSTATSINDEX))
-            return InitError(_("Prune mode is incompatible with -coinstatsindex."));
         if (args.GetBoolArg("-reindex-chainstate", false)) {
             return InitError(_("Prune mode is incompatible with -reindex-chainstate. Use full -reindex instead."));
         }

src/node/blockstorage.cpp

@@ -24,6 +24,7 @@
 #include <walletinitinterface.h>
 
 #include <map>
+#include <unordered_map>
 
 std::atomic_bool fImporting(false);
 std::atomic_bool fReindex(false);
@@ -249,6 +250,11 @@ void BlockManager::FindFilesToPrune(std::set<int>& setFilesToPrune, uint64_t nPr
            nLastBlockWeCanPrune, count);
 }
 
+void BlockManager::UpdatePruneLock(const std::string& name, const PruneLockInfo& lock_info) {
+    AssertLockHeld(::cs_main);
+    m_prune_locks[name] = lock_info;
+}
+
 CBlockIndex* BlockManager::InsertBlockIndex(const uint256& hash)
 {
     AssertLockHeld(cs_main);
@@ -421,6 +427,16 @@ bool BlockManager::IsBlockPruned(const CBlockIndex* pblockindex)
     return (m_have_pruned && !(pblockindex->nStatus & BLOCK_HAVE_DATA) && pblockindex->nTx > 0);
 }
 
+const CBlockIndex* GetFirstStoredBlock(const CBlockIndex* start_block) {
+    AssertLockHeld(::cs_main);
+    assert(start_block);
+    const CBlockIndex* last_block = start_block;
+    while (last_block->pprev && (last_block->pprev->nStatus & BLOCK_HAVE_DATA)) {
+        last_block = last_block->pprev;
+    }
+    return last_block;
+}
+
 // If we're using -prune with -reindex, then delete block files that will be ignored by the
 // reindex.  Since reindexing works by starting at block file 0 and looping until a blockfile
 // is missing, do the same here to delete any later block files after a gap.  Also delete all

src/node/blockstorage.h

@@ -12,6 +12,7 @@
 #include <txdb.h>
 
 #include <cstdint>
+#include <unordered_map>
 #include <vector>
 
 extern RecursiveMutex cs_main;
@@ -77,6 +78,10 @@ struct CBlockIndexHeightOnlyComparator {
     bool operator()(const CBlockIndex* pa, const CBlockIndex* pb) const;
 };
 
+struct PruneLockInfo {
+    int height_first{std::numeric_limits<int>::max()}; //! Height of earliest block that should be kept and not pruned
+};
+
 /**
  * Maintains a tree of blocks (stored in `m_block_index`) which is consulted
  * to determine where the most-work tip is.
@@ -137,6 +142,14 @@ class BlockManager
     /** Dirty block file entries. */
     std::set<int> m_dirty_fileinfo;
 
+    /**
+     * Map from external index name to oldest block that must not be pruned.
+     *
+     * @note Internally, only blocks at height (height_first - PRUNE_LOCK_BUFFER - 1) and
+     * below will be pruned, but callers should avoid assuming any particular buffer size.
+     */
+    std::unordered_map<std::string, PruneLockInfo> m_prune_locks GUARDED_BY(::cs_main);
+
 public:
     BlockMap m_block_index GUARDED_BY(cs_main);
     PrevBlockMap m_prev_block_index GUARDED_BY(cs_main);
@@ -185,8 +198,14 @@ class BlockManager
 
     //! Check whether the block associated with this index entry is pruned or not.
     bool IsBlockPruned(const CBlockIndex* pblockindex) EXCLUSIVE_LOCKS_REQUIRED(::cs_main);
+
+    //! Create or update a prune lock identified by its name
+    void UpdatePruneLock(const std::string& name, const PruneLockInfo& lock_info) EXCLUSIVE_LOCKS_REQUIRED(::cs_main);
 };
 
+//! Find the first block that is not pruned
+const CBlockIndex* GetFirstStoredBlock(const CBlockIndex* start_block) EXCLUSIVE_LOCKS_REQUIRED(::cs_main);
+
 void CleanupBlockRevFiles();
 
 /** Open a block file (blk?????.dat) */

src/rpc/blockchain.cpp

@@ -1361,10 +1361,9 @@ static RPCHelpMan pruneblockchain()
 
     PruneBlockFilesManual(active_chainstate, height);
     const CBlockIndex* block = CHECK_NONFATAL(active_chain.Tip());
-    while (block->pprev && (block->pprev->nStatus & BLOCK_HAVE_DATA)) {
-        block = block->pprev;
-    }
-    return uint64_t(block->nHeight);
+    const CBlockIndex* last_block = GetFirstStoredBlock(block);
+
+    return static_cast<uint64_t>(last_block->nHeight);
 },
     };
 }
@@ -1806,11 +1805,7 @@ RPCHelpMan getblockchaininfo()
     obj.pushKV("pruned", fPruneMode);
     if (fPruneMode) {
         const CBlockIndex* block = CHECK_NONFATAL(tip);
-        while (block->pprev && (block->pprev->nStatus & BLOCK_HAVE_DATA)) {
-            block = block->pprev;
-        }
-
-        obj.pushKV("pruneheight",        block->nHeight);
+        obj.pushKV("pruneheight", GetFirstStoredBlock(block)->nHeight);
 
         // if 0, execution bypasses the whole if block.
         bool automatic_pruning{args.GetArg("-prune", 0) != 1};

src/validation.cpp

@@ -19,7 +19,6 @@
 #include <deploymentstatus.h>
 #include <flatfile.h>
 #include <hash.h>
-#include <index/blockfilterindex.h>
 #include <logging.h>
 #include <logging/timer.h>
 #include <node/blockstorage.h>
@@ -93,6 +92,12 @@ const std::vector<std::string> CHECKLEVEL_DOC {
     "level 4 tries to reconnect the blocks",
     "each level includes the checks of the previous levels",
 };
+/** The number of blocks to keep below the deepest prune lock.
+ *  There is nothing special about this number. It is higher than what we
+ *  expect to see in regular mainnet reorgs, but not so high that it would
+ *  noticeably interfere with the pruning mechanism.
+ * */
+static constexpr int PRUNE_LOCK_BUFFER{10};
 
 /**
  * Mutex to guard access to validation specific variables, such as reading
@@ -2368,12 +2373,24 @@ bool CChainState::FlushStateToDisk(
         CoinsCacheSizeState cache_state = GetCoinsCacheSizeState();
         LOCK(m_blockman.cs_LastBlockFile);
         if (fPruneMode && (m_blockman.m_check_for_pruning || nManualPruneHeight > 0) && !fReindex) {
-            // make sure we don't prune above the blockfilterindexes bestblocks
+            // make sure we don't prune above any of the prune locks bestblocks
             // pruning is height-based
-            int last_prune = m_chain.Height(); // last height we can prune
-            ForEachBlockFilterIndex([&](BlockFilterIndex& index) {
-               last_prune = std::max(1, std::min(last_prune, index.GetSummary().best_block_height));
-            });
+            int last_prune{m_chain.Height()}; // last height we can prune
+            std::optional<std::string> limiting_lock; // prune lock that actually was the limiting factor, only used for logging
+
+            for (const auto& prune_lock : m_blockman.m_prune_locks) {
+                if (prune_lock.second.height_first == std::numeric_limits<int>::max()) continue;
+                // Remove the buffer and one additional block here to get actual height that is outside of the buffer
+                const int lock_height{prune_lock.second.height_first - PRUNE_LOCK_BUFFER - 1};
+                last_prune = std::max(1, std::min(last_prune, lock_height));
+                if (last_prune == lock_height) {
+                    limiting_lock = prune_lock.first;
+                }
+            }
+
+            if (limiting_lock) {
+                LogPrint(BCLog::PRUNE, "%s limited pruning to height %d\n", limiting_lock.value(), last_prune);
+            }
 
             if (nManualPruneHeight > 0) {
                 LOG_TIME_MILLIS_WITH_CATEGORY("find files to prune (manual)", BCLog::BENCHMARK);
@@ -2628,6 +2645,18 @@ bool CChainState::DisconnectTip(BlockValidationState& state, DisconnectedBlockTr
         dbTx->Commit();
     }
     LogPrint(BCLog::BENCHMARK, "- Disconnect block: %.2fms\n", (GetTimeMicros() - nStart) * MILLI);
+
+    {
+        // Prune locks that began at or after the tip should be moved backward so they get a chance to reorg
+        const int max_height_first{pindexDelete->nHeight - 1};
+        for (auto& prune_lock : m_blockman.m_prune_locks) {
+            if (prune_lock.second.height_first <= max_height_first) continue;
+
+            prune_lock.second.height_first = max_height_first;
+            LogPrint(BCLog::PRUNE, "%s prune lock moved back to %d\n", prune_lock.first, max_height_first);
+        }
+    }
+
     // Write the chain state to disk, if necessary.
     if (!FlushStateToDisk(state, FlushStateMode::IF_NEEDED)) {
         return false;