Merge bitcoin/bitcoin#16981: Improve runtime performance of --reindex

db92989 Faster -reindex by initially deserializing only headers (Larry Ruane)
c72de99 util: add CBufferedFile::SkipTo() to move ahead in the stream (Larry Ruane)
48a6890 Add LoadExternalBlockFile() benchmark (Larry Ruane)

Pull request description:

  ### Background
  During the first part of reindexing, `LoadExternalBlockFile()` sequentially reads raw blocks from the `blocks/blk00nnn.dat` files (rather than receiving them from peers, as with initial block download) and eventually adds all of them to the block index. When an individual block is initially read, it can't be immediately added unless all its ancestors have been added, which is rare (only about 8% of the time), because the blocks are not sorted by height. When the block can't be immediately added to the block index, its disk location is saved in a map so it can be added later. When its parent is later added to the block index, `LoadExternalBlockFile()` reads and deserializes the block from disk a second time and adds it to the block index. Most blocks (92%) get deserialized twice.

  ### This PR
  During the initial read, it's rarely useful to deserialize the entire block; only the header is needed to determine if the block can be added to the block index immediately. This change to `LoadExternalBlockFile()` initially deserializes only a block's header, then deserializes the entire block only if it can be added immediately. This reduces reindex time on mainnet by 7 hours on a Raspberry Pi, which translates to around a 25% reduction in the first part of reindexing (adding blocks to the index), and about a 6% reduction in overall reindex time.

  Summary: The performance gain is the result of deserializing each block only once, except its header which is deserialized twice, but the header is only 80 bytes.

ACKs for top commit:
  andrewtoth:
    ACK db92989
  achow101:
    ACK db92989
  aureleoules:
    ACK db92989 - minor changes and new benchmark since last review
  theStack:
    re-ACK db92989
  stickies-v:
    re-ACK db92989

Tree-SHA512: 5a5377192c11edb5b662e18f511c9beb8f250bc88aeadf2f404c92c3232a7617bade50477ebf16c0602b9bd3b68306d3ee7615de58acfd8cae664d28bb7b0136

Author: achow101

src/Makefile.bench.include

@@ -32,6 +32,7 @@ bench_bench_bitcoin_SOURCES = \
   bench/examples.cpp \
   bench/gcs_filter.cpp \
   bench/hashpadding.cpp \
+  bench/load_external.cpp \
   bench/lockedpool.cpp \
   bench/logging.cpp \
   bench/mempool_eviction.cpp \

src/bench/load_external.cpp

@@ -0,0 +1,63 @@
+// Copyright (c) 2022 The Bitcoin Core developers
+// Distributed under the MIT software license, see the accompanying
+// file COPYING or https://www.opensource.org/licenses/mit-license.php.
+
+#include <bench/bench.h>
+#include <bench/data.h>
+#include <chainparams.h>
+#include <test/util/setup_common.h>
+#include <validation.h>
+
+/**
+ * The LoadExternalBlockFile() function is used during -reindex and -loadblock.
+ *
+ * Create a test file that's similar to a datadir/blocks/blk?????.dat file,
+ * It contains around 134 copies of the same block (typical size of real block files).
+ * For each block in the file, LoadExternalBlockFile() won't find its parent,
+ * and so will skip the block. (In the real system, it will re-read the block
+ * from disk later when it encounters its parent.)
+ *
+ * This benchmark measures the performance of deserializing the block (or just
+ * its header, beginning with PR 16981).
+ */
+static void LoadExternalBlockFile(benchmark::Bench& bench)
+{
+    const auto testing_setup{MakeNoLogFileContext<const TestingSetup>(CBaseChainParams::MAIN)};
+
+    // Create a single block as in the blocks files (magic bytes, block size,
+    // block data) as a stream object.
+    const fs::path blkfile{testing_setup.get()->m_path_root / "blk.dat"};
+    CDataStream ss(SER_DISK, 0);
+    auto params{testing_setup->m_node.chainman->GetParams()};
+    ss << params.MessageStart();
+    ss << static_cast<uint32_t>(benchmark::data::block413567.size());
+    // We can't use the streaming serialization (ss << benchmark::data::block413567)
+    // because that first writes a compact size.
+    ss.write(MakeByteSpan(benchmark::data::block413567));
+
+    // Create the test file.
+    {
+        // "wb+" is "binary, O_RDWR | O_CREAT | O_TRUNC".
+        FILE* file{fsbridge::fopen(blkfile, "wb+")};
+        // Make the test block file about 128 MB in length.
+        for (size_t i = 0; i < node::MAX_BLOCKFILE_SIZE / ss.size(); ++i) {
+            if (fwrite(ss.data(), 1, ss.size(), file) != ss.size()) {
+                throw std::runtime_error("write to test file failed\n");
+            }
+        }
+        fclose(file);
+    }
+
+    Chainstate& chainstate{testing_setup->m_node.chainman->ActiveChainstate()};
+    std::multimap<uint256, FlatFilePos> blocks_with_unknown_parent;
+    FlatFilePos pos;
+    bench.run([&] {
+        // "rb" is "binary, O_RDONLY", positioned to the start of the file.
+        // The file will be closed by LoadExternalBlockFile().
+        FILE* file{fsbridge::fopen(blkfile, "rb")};
+        chainstate.LoadExternalBlockFile(file, &pos, &blocks_with_unknown_parent);
+    });
+    fs::remove(blkfile);
+}
+
+BENCHMARK(LoadExternalBlockFile, benchmark::PriorityLevel::HIGH);

src/streams.h

@@ -612,7 +612,6 @@ class CBufferedFile
     uint64_t nRewind;     //!< how many bytes we guarantee to rewind
     std::vector<std::byte> vchBuf; //!< the buffer
 
-protected:
     //! read data from the source to fill the buffer
     bool Fill() {
         unsigned int pos = nSrcPos % vchBuf.size();
@@ -630,6 +629,28 @@ class CBufferedFile
         return true;
     }
 
+    //! Advance the stream's read pointer (m_read_pos) by up to 'length' bytes,
+    //! filling the buffer from the file so that at least one byte is available.
+    //! Return a pointer to the available buffer data and the number of bytes
+    //! (which may be less than the requested length) that may be accessed
+    //! beginning at that pointer.
+    std::pair<std::byte*, size_t> AdvanceStream(size_t length)
+    {
+        assert(m_read_pos <= nSrcPos);
+        if (m_read_pos + length > nReadLimit) {
+            throw std::ios_base::failure("Attempt to position past buffer limit");
+        }
+        // If there are no bytes available, read from the file.
+        if (m_read_pos == nSrcPos && length > 0) Fill();
+
+        size_t buffer_offset{static_cast<size_t>(m_read_pos % vchBuf.size())};
+        size_t buffer_available{static_cast<size_t>(vchBuf.size() - buffer_offset)};
+        size_t bytes_until_source_pos{static_cast<size_t>(nSrcPos - m_read_pos)};
+        size_t advance{std::min({length, buffer_available, bytes_until_source_pos})};
+        m_read_pos += advance;
+        return std::make_pair(&vchBuf[buffer_offset], advance);
+    }
+
 public:
     CBufferedFile(FILE* fileIn, uint64_t nBufSize, uint64_t nRewindIn, int nTypeIn, int nVersionIn)
         : nType(nTypeIn), nVersion(nVersionIn), nSrcPos(0), m_read_pos(0), nReadLimit(std::numeric_limits<uint64_t>::max()), nRewind(nRewindIn), vchBuf(nBufSize, std::byte{0})
@@ -667,24 +688,21 @@ class CBufferedFile
     //! read a number of bytes
     void read(Span<std::byte> dst)
     {
-        if (dst.size() + m_read_pos > nReadLimit) {
-            throw std::ios_base::failure("Read attempted past buffer limit");
-        }
         while (dst.size() > 0) {
-            if (m_read_pos == nSrcPos)
-                Fill();
-            unsigned int pos = m_read_pos % vchBuf.size();
-            size_t nNow = dst.size();
-            if (nNow + pos > vchBuf.size())
-                nNow = vchBuf.size() - pos;
-            if (nNow + m_read_pos > nSrcPos)
-                nNow = nSrcPos - m_read_pos;
-            memcpy(dst.data(), &vchBuf[pos], nNow);
-            m_read_pos += nNow;
-            dst = dst.subspan(nNow);
+            auto [buffer_pointer, length]{AdvanceStream(dst.size())};
+            memcpy(dst.data(), buffer_pointer, length);
+            dst = dst.subspan(length);
         }
     }
 
+    //! Move the read position ahead in the stream to the given position.
+    //! Use SetPos() to back up in the stream, not SkipTo().
+    void SkipTo(const uint64_t file_pos)
+    {
+        assert(file_pos >= m_read_pos);
+        while (m_read_pos < file_pos) AdvanceStream(file_pos - m_read_pos);
+    }
+
     //! return the current reading position
     uint64_t GetPos() const {
         return m_read_pos;

src/test/streams_tests.cpp

@@ -253,7 +253,7 @@ BOOST_AUTO_TEST_CASE(streams_buffered_file)
         BOOST_CHECK(false);
     } catch (const std::exception& e) {
         BOOST_CHECK(strstr(e.what(),
-                        "Read attempted past buffer limit") != nullptr);
+                           "Attempt to position past buffer limit") != nullptr);
     }
     // The default argument removes the limit completely.
     BOOST_CHECK(bf.SetLimit());
@@ -322,14 +322,63 @@ BOOST_AUTO_TEST_CASE(streams_buffered_file)
     BOOST_CHECK(!bf.SetPos(0));
     // But we should now be positioned at least as far back as allowed
     // by the rewind window (relative to our farthest read position, 40).
-    BOOST_CHECK(bf.GetPos() <= 30);
+    BOOST_CHECK(bf.GetPos() <= 30U);
 
     // We can explicitly close the file, or the destructor will do it.
     bf.fclose();
 
     fs::remove(streams_test_filename);
 }
 
+BOOST_AUTO_TEST_CASE(streams_buffered_file_skip)
+{
+    fs::path streams_test_filename = m_args.GetDataDirBase() / "streams_test_tmp";
+    FILE* file = fsbridge::fopen(streams_test_filename, "w+b");
+    // The value at each offset is the byte offset (e.g. byte 1 in the file has the value 0x01).
+    for (uint8_t j = 0; j < 40; ++j) {
+        fwrite(&j, 1, 1, file);
+    }
+    rewind(file);
+
+    // The buffer is 25 bytes, allow rewinding 10 bytes.
+    CBufferedFile bf(file, 25, 10, 222, 333);
+
+    uint8_t i;
+    // This is like bf >> (7-byte-variable), in that it will cause data
+    // to be read from the file into memory, but it's not copied to us.
+    bf.SkipTo(7);
+    BOOST_CHECK_EQUAL(bf.GetPos(), 7U);
+    bf >> i;
+    BOOST_CHECK_EQUAL(i, 7);
+
+    // The bytes in the buffer up to offset 7 are valid and can be read.
+    BOOST_CHECK(bf.SetPos(0));
+    bf >> i;
+    BOOST_CHECK_EQUAL(i, 0);
+    bf >> i;
+    BOOST_CHECK_EQUAL(i, 1);
+
+    bf.SkipTo(11);
+    bf >> i;
+    BOOST_CHECK_EQUAL(i, 11);
+
+    // SkipTo() honors the transfer limit; we can't position beyond the limit.
+    bf.SetLimit(13);
+    try {
+        bf.SkipTo(14);
+        BOOST_CHECK(false);
+    } catch (const std::exception& e) {
+        BOOST_CHECK(strstr(e.what(), "Attempt to position past buffer limit") != nullptr);
+    }
+
+    // We can position exactly to the transfer limit.
+    bf.SkipTo(13);
+    BOOST_CHECK_EQUAL(bf.GetPos(), 13U);
+
+    bf.fclose();
+    fs::remove(streams_test_filename);
+}
+
 BOOST_AUTO_TEST_CASE(streams_buffered_file_rand)
 {
     // Make this test deterministic.
@@ -361,7 +410,7 @@ BOOST_AUTO_TEST_CASE(streams_buffered_file_rand)
             // sizes; the boundaries of the objects can interact arbitrarily
             // with the CBufferFile's internal buffer. These first three
             // cases simulate objects of various sizes (1, 2, 5 bytes).
-            switch (InsecureRandRange(5)) {
+            switch (InsecureRandRange(6)) {
             case 0: {
                 uint8_t a[1];
                 if (currentPos + 1 > fileSize)
@@ -399,6 +448,16 @@ BOOST_AUTO_TEST_CASE(streams_buffered_file_rand)
                 break;
             }
             case 3: {
+                // SkipTo is similar to the "read" cases above, except
+                // we don't receive the data.
+                size_t skip_length{static_cast<size_t>(InsecureRandRange(5))};
+                if (currentPos + skip_length > fileSize) continue;
+                bf.SetLimit(currentPos + skip_length);
+                bf.SkipTo(currentPos + skip_length);
+                currentPos += skip_length;
+                break;
+            }
+            case 4: {
                 // Find a byte value (that is at or ahead of the current position).
                 size_t find = currentPos + InsecureRandRange(8);
                 if (find >= fileSize)
@@ -415,7 +474,7 @@ BOOST_AUTO_TEST_CASE(streams_buffered_file_rand)
                 currentPos++;
                 break;
             }
-            case 4: {
+            case 5: {
                 size_t requestPos = InsecureRandRange(maxPos + 4);
                 bool okay = bf.SetPos(requestPos);
                 // The new position may differ from the requested position

src/validation.cpp

@@ -4389,6 +4389,8 @@ void Chainstate::LoadExternalBlockFile(
     try {
         // This takes over fileIn and calls fclose() on it in the CBufferedFile destructor
         CBufferedFile blkdat(fileIn, 2*MAX_BLOCK_SERIALIZED_SIZE, MAX_BLOCK_SERIALIZED_SIZE+8, SER_DISK, CLIENT_VERSION);
+        // nRewind indicates where to resume scanning in case something goes wrong,
+        // such as a block fails to deserialize.
         uint64_t nRewind = blkdat.GetPos();
         while (!blkdat.eof()) {
             if (ShutdownRequested()) return;
@@ -4412,42 +4414,50 @@ void Chainstate::LoadExternalBlockFile(
                     continue;
             } catch (const std::exception&) {
                 // no valid block header found; don't complain
+                // (this happens at the end of every blk.dat file)
                 break;
             }
             try {
-                // read block
-                uint64_t nBlockPos = blkdat.GetPos();
+                // read block header
+                const uint64_t nBlockPos{blkdat.GetPos()};
                 if (dbp)
                     dbp->nPos = nBlockPos;
                 blkdat.SetLimit(nBlockPos + nSize);
-                std::shared_ptr<CBlock> pblock = std::make_shared<CBlock>();
-                CBlock& block = *pblock;
-                blkdat >> block;
-                nRewind = blkdat.GetPos();
-
-                uint256 hash = block.GetHash();
+                CBlockHeader header;
+                blkdat >> header;
+                const uint256 hash{header.GetHash()};
+                // Skip the rest of this block (this may read from disk into memory); position to the marker before the
+                // next block, but it's still possible to rewind to the start of the current block (without a disk read).
+                nRewind = nBlockPos + nSize;
+                blkdat.SkipTo(nRewind);
                 {
                     LOCK(cs_main);
                     // detect out of order blocks, and store them for later
-                    if (hash != params.GetConsensus().hashGenesisBlock && !m_blockman.LookupBlockIndex(block.hashPrevBlock)) {
+                    if (hash != params.GetConsensus().hashGenesisBlock && !m_blockman.LookupBlockIndex(header.hashPrevBlock)) {
                         LogPrint(BCLog::REINDEX, "%s: Out of order block %s, parent %s not known\n", __func__, hash.ToString(),
-                                block.hashPrevBlock.ToString());
+                                 header.hashPrevBlock.ToString());
                         if (dbp && blocks_with_unknown_parent) {
-                            blocks_with_unknown_parent->emplace(block.hashPrevBlock, *dbp);
+                            blocks_with_unknown_parent->emplace(header.hashPrevBlock, *dbp);
                         }
                         continue;
                     }
 
                     // process in case the block isn't known yet
                     const CBlockIndex* pindex = m_blockman.LookupBlockIndex(hash);
                     if (!pindex || (pindex->nStatus & BLOCK_HAVE_DATA) == 0) {
-                      BlockValidationState state;
-                      if (AcceptBlock(pblock, state, nullptr, true, dbp, nullptr, true)) {
-                          nLoaded++;
-                      }
-                      if (state.IsError()) {
-                          break;
-                      }
+                        // This block can be processed immediately; rewind to its start, read and deserialize it.
+                        blkdat.SetPos(nBlockPos);
+                        std::shared_ptr<CBlock> pblock{std::make_shared<CBlock>()};
+                        blkdat >> *pblock;
+                        nRewind = blkdat.GetPos();
+
+                        BlockValidationState state;
+                        if (AcceptBlock(pblock, state, nullptr, true, dbp, nullptr, true)) {
+                            nLoaded++;
+                        }
+                        if (state.IsError()) {
+                            break;
+                        }
                     } else if (hash != params.GetConsensus().hashGenesisBlock && pindex->nHeight % 1000 == 0) {
                         LogPrint(BCLog::REINDEX, "Block Import: already had block %s at height %d\n", hash.ToString(), pindex->nHeight);
                     }