Fulcrum/Storage.cpp
Calin Culianu c2c17f5e36 nits
2019-12-03 22:47:03 +02:00

447 lines
19 KiB
C++

#include "BTC.h"
#include "Storage.h"
#include "rocksdb/db.h"
#include "rocksdb/options.h"
#include "rocksdb/slice.h"
#include <QByteArray>
#include <QDir>
#include <atomic>
#include <optional>
#include <shared_mutex>
#include <type_traits>
#include <vector>
DatabaseError::~DatabaseError(){} // weak vtable warning suppression
DatabaseSerializationError::~DatabaseSerializationError() {} // weak vtable warning suppression
DatabaseFormatError::~DatabaseFormatError() {} // weak vtable warning suppression
namespace {
/// Encapsulates the 'meta' db table
struct Meta {
uint32_t magic = 0xf33db33f, version = 0x1;
QString chain; ///< "test", "main", etc
};
// some database keys we use -- todo: if this grows large, move it elsewhere
static const rocksdb::Slice kMeta{"meta"}, kNumHeaders{"num_headers"};
static constexpr size_t MAX_HEADERS = 100000000; // 100 mln max headers for now.
/// NOTE: the byte array should live as long as the slice does. slice is just a weak ref into the byte array
inline rocksdb::Slice ToSlice(const QByteArray &ba) { return rocksdb::Slice(ba.constData(), size_t(ba.size())); }
/// NOTE: The slice should live as long as the returned QByteArray does. The QByteArray is a weak pointer into the slice!
inline QByteArray FromSlice(const rocksdb::Slice &s) { return QByteArray::fromRawData(s.data(), int(s.size())); }
// serialize/deser -- for basic types we use QDataStream, but we also have specializations at the end of this file
template <typename Type>
QByteArray Serialize(const Type & n) {
QByteArray ba;
QDataStream ds(&ba, QIODevice::WriteOnly|QIODevice::Truncate);
ds << n;
return ba;
}
template <typename Type>
Type Deserialize(const QByteArray &ba, bool *ok = nullptr) {
QDataStream ds(ba);
Type ret{};
ds >> ret;
if (ok)
*ok = ds.status() == QDataStream::Status::Ok;
return ret;
}
/// Serialize a simple value such as an int directly, without using the space overhead that QDataStream imposes.
/// This is less safe but is more compact since the bytes of the passed-in value are written directly to the
/// returned QByteArray, without any encapsulation. Note that use of this mechanism makes all data in the database
/// no longer platform-neutral, which is ok. The presumption is users can re-synch their DB if switching
/// architectures.
template <typename Scalar,
std::enable_if_t<std::is_scalar_v<Scalar> && !std::is_pointer_v<Scalar>, int> = 0>
QByteArray SerializeScalar(const Scalar & s) {
return QByteArray(reinterpret_cast<const char *>(&s), sizeof(s));
}
/// Inverse of above. Pass in an optional 'pos' pointer if you wish to continue reading raw scalars from the same
/// QByteArray during subsequent calls to this template function. *ok, if specified, is set to false if we ran off
/// the QByteArray's bounds, and a default-constructed value of 'Scalar' is returned. No other safety checking is
/// done. On successful deserialization of the scalar, *pos (if specified) is updated to point just past the
/// last byte of the successuflly converted item. On failure, *pos is always set to point past the end of the
/// QByteArray.
template <typename Scalar,
std::enable_if_t<std::is_scalar_v<Scalar> && !std::is_pointer_v<Scalar>, int> = 0>
Scalar DeserializeScalar(const QByteArray &ba, bool *ok = nullptr, int *pos_out = nullptr) {
Scalar ret{};
int dummy = 0;
int & pos = pos_out ? *pos_out : dummy;
if (pos >= 0 && pos + int(sizeof(ret)) <= ba.size()) {
if (ok) *ok = true;
ret = *reinterpret_cast<const Scalar *>(ba.data() + pos);
pos += sizeof(ret);
} else {
if (ok) *ok = false;
pos = ba.size();
}
return ret;
}
// specializations
template <> QByteArray Serialize(const Meta &);
template <> Meta Deserialize(const QByteArray &, bool *);
/// DB read/write helpers
/// NOTE: these may throw DatabaseError
/// If missingOk=false, then the returned optional is guaranteed to have a value if this function returns without throwing.
/// If missingOk=true, then if there was no other database error and the key was not found, the returned optional !has_value()
///
/// Template arg "safeScalar", if true, will deserialize even scalar int, float, etc data using the Deserialize<>
/// function (uses QDataStream, is platform neutral, but is slightly slower). If false, we will use the
/// DeserializeScalar<> fast function for scalars such as ints. It's important to read from the DB in the same
/// 'safeScalar' mode as was written!
template <typename RetType, bool safeScalar = false>
std::optional<RetType> GenericDBGet(rocksdb::DB *db, const rocksdb::Slice & key, bool missingOk = false,
const QString & errorMsgPrefix = QString(), ///< used to specify a custom error message in the thrown exception
bool acceptExtraBytesAtEndOfData = false) ///< if true, we are ok with extra unparsed bytes in data. otherwise we throw. (this check is only done for !safeScalar mode on basic types)
{
rocksdb::PinnableSlice datum;
std::optional<RetType> ret;
if (UNLIKELY(!db)) throw InternalError("GenericDBGet was passed a null pointer!");
const auto status = db->Get(rocksdb::ReadOptions(), db->DefaultColumnFamily(), key, &datum);
if (missingOk && status.IsNotFound()) {
return ret; // optional will not has_value() to indicate missing key
} else if (!status.ok()) {
throw DatabaseError(QString("%1: %2")
.arg(!errorMsgPrefix.isEmpty() ? errorMsgPrefix : "Error reading a key from the db")
.arg(status.ToString().c_str()));
} else {
// ok status
if constexpr (std::is_base_of_v<QByteArray, std::remove_cv_t<RetType> >) {
static_assert (!safeScalar, "safeScalar=true mode is not supported for QByteArrays (it only is useful for scalar types)" );
// special compile-time case for QByteArray subclasses -- return a deep copy of the data bytes directly.
// TODO: figure out a way to do this without the 1 extra copy! (PinnableSlice -> ret).
ret.emplace( reinterpret_cast<const char *>(datum.data()), QByteArray::size_type(datum.size()) );
} else if constexpr (!safeScalar && std::is_scalar_v<RetType> && !std::is_pointer_v<RetType>) {
if (!acceptExtraBytesAtEndOfData && datum.size() > sizeof(RetType)) {
// reject extra stuff at end of data stream
throw DatabaseFormatError(QString("%1: Extra bytes at the end of data")
.arg(!errorMsgPrefix.isEmpty() ? errorMsgPrefix : "Database format error"));
}
bool ok;
ret.emplace( DeserializeScalar<RetType>(FromSlice(datum), &ok) );
if (!ok) {
throw DatabaseSerializationError(
QString("%1: Key was retrieved ok, but data could not be deserialized as a scalar '%2'")
.arg(!errorMsgPrefix.isEmpty() ? errorMsgPrefix : "Error deserializing a scalar from db")
.arg(typeid (RetType).name()));
}
} else {
if (UNLIKELY(acceptExtraBytesAtEndOfData))
Debug() << "Warning: Caller misuse of function '" << __FUNCTION__
<< "'. 'acceptExtraBytesAtEndOfData=true' is ignored when deserializing using QDataStream.";
bool ok;
ret.emplace( Deserialize<RetType>(FromSlice(datum), &ok) );
if (!ok) {
throw DatabaseSerializationError(
QString("%1: Key was retrieved ok, but data could not be deserialized")
.arg(!errorMsgPrefix.isEmpty() ? errorMsgPrefix : "Error deserializing an object from db"));
}
}
}
return ret;
}
/// Conveneience for above with the missingOk flag set to false. Will always throw or return a real value.
template <typename RetType, bool safeScalar = false>
RetType GenericDBGetFailIfMissing(rocksdb::DB * db, const rocksdb::Slice &k, const QString &errMsgPrefix = QString(), bool extraDataOk = false)
{
return GenericDBGet<RetType, safeScalar>(db, k, false, errMsgPrefix, extraDataOk).value();
}
}
struct Storage::Pvt
{
Meta meta;
Lock metaLock;
Headers headers;
std::pair<unsigned, QByteArray> lastHeaderSaved; ///< remember the last header saved to disk, so subsequent saves leave off from this index
RWLock headersLock;
std::atomic<std::underlying_type_t<SaveItem>> pendingSaves{0};
struct RocksDBs {
std::unique_ptr<rocksdb::DB> meta, headers;
} db;
BTC::HeaderVerifier headerVerifier;
Lock headerVerifierLock;
};
Storage::Storage(const std::shared_ptr<Options> & options)
: Mgr(nullptr), options(options), p(new Pvt)
{
setObjectName("Storage");
_thread.setObjectName(objectName());
}
Storage::~Storage() { Debug("%s", __FUNCTION__); cleanup(); }
void Storage::startup()
{
rocksdb::Options opts;
// Optimize RocksDB. This is the easiest way to get RocksDB to perform well
opts.IncreaseParallelism();
opts.OptimizeLevelStyleCompaction();
// create the DB if it's not already present
opts.create_if_missing = true;
opts.error_if_exists = false;
opts.compression = rocksdb::CompressionType::kNoCompression; // for now we test without compression. TODO: characterize what is fastest and best..
QString metaPath = options->datadir + QDir::separator() + "meta",
headersPath = options->datadir + QDir::separator() + "headers";
rocksdb::DB *db = nullptr;
rocksdb::Status s;
// meta database
s = rocksdb::DB::Open(opts, metaPath.toStdString(), &db);
if (!s.ok() || !db)
throw DatabaseError(QString("Error opening meta database: %1 (path: %2)").arg(s.ToString().c_str()).arg(metaPath));
p->db.meta.reset(db); db = nullptr;
// headers database
s = rocksdb::DB::Open(opts, headersPath.toStdString(), &db);
if (!s.ok() || !db)
throw DatabaseError(QString("Error opening headers database: %1 (path: %2)").arg(s.ToString().c_str()).arg(headersPath));
p->db.headers.reset(db); db = nullptr;
// load/check meta
{
Meta m_db;
static const QString errMsg{"Incompatible database format -- delete the datadir and resynch. RocksDB error"};
if (auto opt = GenericDBGet<Meta>(p->db.meta.get(), kMeta, true, errMsg);
opt.has_value())
{
m_db = opt.value();
if (m_db.magic != p->meta.magic || m_db.version != p->meta.version) {
throw DatabaseFormatError(errMsg);
}
p->meta = m_db;
Debug () << "Read meta from db ok";
} else {
// ok, did not exist .. write a new one to db
saveMeta_impl();
}
}
Log() << "Loading database ...";
// load headers -- may throw
loadHeadersFromDB();
start(); // starts our thread
}
void Storage::cleanup()
{
stop(); // joins our thread
save_impl(); // writes all "dirty" data immediately to disk
}
auto Storage::stats() const -> Stats
{
// TODO ...
return Stats();
}
auto Storage::mutableHeaders() -> std::pair<Headers &, ExclusiveLockGuard>
{
return std::pair<Headers &, ExclusiveLockGuard>{ p->headers, p->headersLock };
}
auto Storage::headers() const -> std::pair<const Headers &, SharedLockGuard>
{
return std::pair<const Headers &, SharedLockGuard>{ p->headers, p->headersLock };
}
// Keep returned LockGuard in scope while you use the HeaderVerifier
auto Storage::headerVerifier() -> std::pair<BTC::HeaderVerifier &, LockGuard>
{
return std::pair<BTC::HeaderVerifier &, LockGuard>( p->headerVerifier, p->headerVerifierLock );
}
QString Storage::getChain() const
{
LockGuard l(p->metaLock);
return p->meta.chain;
}
void Storage::setChain(const QString &chain)
{
{
LockGuard l(p->metaLock);
p->meta.chain = chain;
}
save(SaveItem::Meta);
}
void Storage::save(SaveSpec typed_spec)
{
using IntType = decltype(p->pendingSaves.load());
// enqueue save on event loop if not previously enqueued (we know it was previously enqueued if the p->pendingSaves
// atomic variable is not 0).
if (const auto spec = IntType(typed_spec); ! p->pendingSaves.fetch_or(spec))
{
QTimer::singleShot(0, this, &Storage::save_impl);
}
}
void Storage::save_impl()
{
if (const auto flags = SaveSpec(p->pendingSaves.exchange(0)); flags) { // atomic clear of flags, grab prev val
try {
if (flags & SaveItem::Hdrs) { // headers
// TODO: See if this copy is inefficient. So far it seems preferable to the alternative.
// We would rather do this here than hold the lock for the duration of the db save.
const Headers copy = headers().first;
saveHeaders_impl(copy);
}
if (flags & SaveItem::Meta) { // Meta
LockGuard l(p->metaLock);
saveMeta_impl();
}
} catch (const std::exception & e) {
Fatal() << e.what(); // will abort app...
}
}
}
void Storage::saveMeta_impl()
{
if (!p->db.meta) return;
if (auto status = p->db.meta->Put(rocksdb::WriteOptions(), kMeta, ToSlice(Serialize(p->meta))); !status.ok()) {
throw DatabaseError("Failed to write meta to db");
}
Debug() << "Wrote new metadata to db";
}
void Storage::saveHeaders_impl(const Headers &h)
{
if (!p->db.headers) return;
Debug() << "Saving headers ...";
size_t start = 0;
const auto t0 = Util::getTimeNS();
{
// figure out where to resume from
if (size_t idx = p->lastHeaderSaved.first; idx && idx < h.size() && p->lastHeaderSaved.second == h[idx]) {
start = idx + 1;
}
if (start < h.size()) {
Debug() << "Saving from height " << start;
rocksdb::WriteBatch batch;
for (size_t i = start; i < h.size(); ++i) {
// headers are keyed off of uint32_t index (height) -- note we save the raw int value here without any QDataStream encapsulation
if (auto stat = batch.Put(ToSlice(SerializeScalar(uint32_t(i))), ToSlice(h[i])); !stat.ok())
throw DatabaseError(QString("Error writing header %1: %2").arg(i).arg(stat.ToString().c_str()));
}
// save height to db *after* we successfully wrote all the headers
if (auto stat = batch.Put(kNumHeaders, ToSlice(Serialize(uint32_t(h.size())))); !stat.ok())
throw DatabaseError(QString("Error writing header size key: %1").arg(stat.ToString().c_str()));
auto opts = rocksdb::WriteOptions();
opts.sync = true;
opts.low_pri = true;
if (auto stat = p->db.headers->Write(opts, &batch); !stat.ok())
throw DatabaseError(QString("Error writing headers: %1").arg(stat.ToString().c_str()));
auto fopts = rocksdb::FlushOptions();
fopts.wait = true; fopts.allow_write_stall = true;
if (auto stat = p->db.headers->Flush(fopts); !stat.ok())
throw DatabaseError(QString("Flush error while writing headers: %1").arg(stat.ToString().c_str()));
p->lastHeaderSaved = { h.size()-1, h.back() }; // remember last
}
}
const auto elapsed = Util::getTimeNS();
const auto ct = (h.size()-start);
Debug() << "Wrote " << ct << " " << Util::Pluralize("header", ct) << " to db in " << QString::number((elapsed-t0)/1e6, 'f', 3) << " msec";
}
void Storage::loadHeadersFromDB()
{
FatalAssert(!!p->db.headers) << __FUNCTION__ << ": Headers db is not open";
Debug() << "Loading headers ...";
uint32_t num = 0;
const auto t0 = Util::getTimeNS();
{
auto * const db = p->db.headers.get();
num = GenericDBGet<uint32_t, true>(db, kNumHeaders, true, "Error reading header count from database").value_or(0); // missing ok
if (num > MAX_HEADERS)
throw DatabaseFormatError(QString("Header count (%1) in database exceeds MAX_HEADERS! This is likely due to"
" a database format mistmatch. Delete the datadir and resynch it.")
.arg(num));
Headers h;
h.reserve(num);
const QString errMsg("Error retrieving header from db");
const int hsz = BTC::GetBlockHeaderSize();
// read db
for (uint32_t i = 0; i < num; ++i) {
// guaranteed to return a value or throw
auto bytes = GenericDBGetFailIfMissing<QByteArray>(db, ToSlice(SerializeScalar(uint32_t(i))), errMsg);
if (UNLIKELY(bytes.size() != hsz))
throw DatabaseFormatError(QString("Error reading header %1, wrong size: %2").arg(i).arg(bytes.size()));
h.emplace_back(std::move(bytes));
}
// verify headers: hashPrevBlock must match what we actually read from db
if (num) {
Debug() << "Verifying " << num << " " << Util::Pluralize("header", num) << " ...";
auto [verif, lock] = headerVerifier();
QString err;
for (unsigned i = 0; i < num; ++i) {
// verify headers hash chains match by checking hashPrevBlock versus actual previous hash.
// we use the helper functor HeaderVerifier for this.
if (!verif(h[i], &err))
throw DatabaseError(QString("%1. Possible databaase corruption. Delete the datadir and resynch.").arg(err));
}
p->lastHeaderSaved = verif.lastHeaderProcessed(); // remember last
}
// locked until scope end...
auto [headers, lock] = mutableHeaders();
headers.swap(h);
}
if (num) {
const auto elapsed = Util::getTimeNS();
Debug() << "Read & verified " << num << " " << Util::Pluralize("header", num) << " from db in " << QString::number((elapsed-t0)/1e6, 'f', 3) << " msec";
}
}
namespace {
// specializations of Serialize/Deserialize
template <> QByteArray Serialize(const Meta &m)
{
QByteArray ba;
{
QDataStream ds(&ba, QIODevice::WriteOnly|QIODevice::Truncate);
// we serialize the 'magic' value as a simple scalar as a sort of endian check for the DB
ds << SerializeScalar(m.magic) << m.version << m.chain;
}
return ba;
}
template <> Meta Deserialize(const QByteArray &ba, bool *ok_ptr)
{
bool dummy;
bool &ok (ok_ptr ? *ok_ptr : dummy);
ok = false;
Meta m{0, 0, {}};
{
QDataStream ds(ba);
QByteArray magicBytes;
ds >> magicBytes; // read magic as raw bytes.
if ((ok = ds.status() == QDataStream::Status::Ok)) {
m.magic = DeserializeScalar<decltype (m.magic)>(magicBytes, &ok);
if (ok) {
ds >> m.version >> m.chain;
ok = ds.status() == QDataStream::Status::Ok;
}
}
}
return m;
}
}