2013-10-16 23:59:46 +02:00
|
|
|
// Copyright (c) 2013, Facebook, Inc. All rights reserved.
|
|
|
|
// This source code is licensed under the BSD-style license found in the
|
|
|
|
// LICENSE file in the root directory of this source tree. An additional grant
|
|
|
|
// of patent rights can be found in the PATENTS file in the same directory.
|
|
|
|
//
|
2011-03-18 23:37:00 +01:00
|
|
|
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
|
|
|
|
// Use of this source code is governed by a BSD-style license that can be
|
|
|
|
// found in the LICENSE file. See the AUTHORS file for names of contributors.
|
|
|
|
|
|
|
|
#include "db/log_reader.h"
|
|
|
|
#include "db/log_writer.h"
|
2013-08-23 17:38:13 +02:00
|
|
|
#include "rocksdb/env.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "util/coding.h"
|
|
|
|
#include "util/crc32c.h"
|
|
|
|
#include "util/random.h"
|
|
|
|
#include "util/testharness.h"
|
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
namespace rocksdb {
|
2011-03-18 23:37:00 +01:00
|
|
|
namespace log {
|
|
|
|
|
|
|
|
// Construct a string of the specified length made out of the supplied
|
|
|
|
// partial string.
|
|
|
|
static std::string BigString(const std::string& partial_string, size_t n) {
|
|
|
|
std::string result;
|
|
|
|
while (result.size() < n) {
|
|
|
|
result.append(partial_string);
|
|
|
|
}
|
|
|
|
result.resize(n);
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Construct a string from a number
|
|
|
|
static std::string NumberString(int n) {
|
|
|
|
char buf[50];
|
|
|
|
snprintf(buf, sizeof(buf), "%d.", n);
|
|
|
|
return std::string(buf);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Return a skewed potentially long string
|
|
|
|
static std::string RandomSkewedString(int i, Random* rnd) {
|
|
|
|
return BigString(NumberString(i), rnd->Skewed(17));
|
|
|
|
}
|
|
|
|
|
|
|
|
class LogTest {
|
|
|
|
private:
|
|
|
|
class StringDest : public WritableFile {
|
|
|
|
public:
|
|
|
|
std::string contents_;
|
|
|
|
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
explicit StringDest(Slice& reader_contents) :
|
|
|
|
WritableFile(),
|
|
|
|
contents_(""),
|
|
|
|
reader_contents_(reader_contents),
|
|
|
|
last_flush_(0) {
|
|
|
|
reader_contents_ = Slice(contents_.data(), 0);
|
|
|
|
};
|
|
|
|
|
2015-02-26 20:28:41 +01:00
|
|
|
virtual Status Close() override { return Status::OK(); }
|
|
|
|
virtual Status Flush() override {
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
ASSERT_TRUE(reader_contents_.size() <= last_flush_);
|
|
|
|
size_t offset = last_flush_ - reader_contents_.size();
|
|
|
|
reader_contents_ = Slice(
|
|
|
|
contents_.data() + offset,
|
|
|
|
contents_.size() - offset);
|
|
|
|
last_flush_ = contents_.size();
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
2015-02-26 20:28:41 +01:00
|
|
|
virtual Status Sync() override { return Status::OK(); }
|
|
|
|
virtual Status Append(const Slice& slice) override {
|
2011-03-18 23:37:00 +01:00
|
|
|
contents_.append(slice.data(), slice.size());
|
|
|
|
return Status::OK();
|
|
|
|
}
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
void Drop(size_t bytes) {
|
|
|
|
contents_.resize(contents_.size() - bytes);
|
|
|
|
reader_contents_ = Slice(
|
|
|
|
reader_contents_.data(), reader_contents_.size() - bytes);
|
|
|
|
last_flush_ = contents_.size();
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
Slice& reader_contents_;
|
|
|
|
size_t last_flush_;
|
2011-03-18 23:37:00 +01:00
|
|
|
};
|
|
|
|
|
|
|
|
class StringSource : public SequentialFile {
|
|
|
|
public:
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
Slice& contents_;
|
2011-03-18 23:37:00 +01:00
|
|
|
bool force_error_;
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
size_t force_error_position_;
|
|
|
|
bool force_eof_;
|
|
|
|
size_t force_eof_position_;
|
2011-03-18 23:37:00 +01:00
|
|
|
bool returned_partial_;
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
explicit StringSource(Slice& contents) :
|
|
|
|
contents_(contents),
|
|
|
|
force_error_(false),
|
|
|
|
force_error_position_(0),
|
|
|
|
force_eof_(false),
|
|
|
|
force_eof_position_(0),
|
|
|
|
returned_partial_(false) { }
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2015-02-26 20:28:41 +01:00
|
|
|
virtual Status Read(size_t n, Slice* result, char* scratch) override {
|
2011-03-18 23:37:00 +01:00
|
|
|
ASSERT_TRUE(!returned_partial_) << "must not Read() after eof/error";
|
|
|
|
|
|
|
|
if (force_error_) {
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
if (force_error_position_ >= n) {
|
|
|
|
force_error_position_ -= n;
|
|
|
|
} else {
|
|
|
|
*result = Slice(contents_.data(), force_error_position_);
|
|
|
|
contents_.remove_prefix(force_error_position_);
|
|
|
|
force_error_ = false;
|
|
|
|
returned_partial_ = true;
|
|
|
|
return Status::Corruption("read error");
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
if (contents_.size() < n) {
|
|
|
|
n = contents_.size();
|
|
|
|
returned_partial_ = true;
|
|
|
|
}
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
|
|
|
|
if (force_eof_) {
|
|
|
|
if (force_eof_position_ >= n) {
|
|
|
|
force_eof_position_ -= n;
|
|
|
|
} else {
|
|
|
|
force_eof_ = false;
|
|
|
|
n = force_eof_position_;
|
|
|
|
returned_partial_ = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// By using scratch we ensure that caller has control over the
|
|
|
|
// lifetime of result.data()
|
|
|
|
memcpy(scratch, contents_.data(), n);
|
|
|
|
*result = Slice(scratch, n);
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
contents_.remove_prefix(n);
|
|
|
|
return Status::OK();
|
|
|
|
}
|
2011-05-21 04:17:43 +02:00
|
|
|
|
2015-02-26 20:28:41 +01:00
|
|
|
virtual Status Skip(uint64_t n) override {
|
2011-05-21 04:17:43 +02:00
|
|
|
if (n > contents_.size()) {
|
|
|
|
contents_.clear();
|
|
|
|
return Status::NotFound("in-memory file skipepd past end");
|
|
|
|
}
|
|
|
|
|
|
|
|
contents_.remove_prefix(n);
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
};
|
|
|
|
|
|
|
|
class ReportCollector : public Reader::Reporter {
|
|
|
|
public:
|
|
|
|
size_t dropped_bytes_;
|
|
|
|
std::string message_;
|
|
|
|
|
|
|
|
ReportCollector() : dropped_bytes_(0) { }
|
2015-02-26 20:28:41 +01:00
|
|
|
virtual void Corruption(size_t bytes, const Status& status) override {
|
2011-03-18 23:37:00 +01:00
|
|
|
dropped_bytes_ += bytes;
|
|
|
|
message_.append(status.ToString());
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2013-01-20 11:07:13 +01:00
|
|
|
std::string& dest_contents() {
|
|
|
|
auto dest = dynamic_cast<StringDest*>(writer_.file());
|
|
|
|
assert(dest);
|
|
|
|
return dest->contents_;
|
|
|
|
}
|
|
|
|
|
|
|
|
const std::string& dest_contents() const {
|
|
|
|
auto dest = dynamic_cast<const StringDest*>(writer_.file());
|
|
|
|
assert(dest);
|
|
|
|
return dest->contents_;
|
|
|
|
}
|
|
|
|
|
|
|
|
void reset_source_contents() {
|
|
|
|
auto src = dynamic_cast<StringSource*>(reader_.file());
|
|
|
|
assert(src);
|
|
|
|
src->contents_ = dest_contents();
|
|
|
|
}
|
|
|
|
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
Slice reader_contents_;
|
2013-01-20 11:07:13 +01:00
|
|
|
unique_ptr<StringDest> dest_holder_;
|
|
|
|
unique_ptr<StringSource> source_holder_;
|
2011-03-18 23:37:00 +01:00
|
|
|
ReportCollector report_;
|
|
|
|
Writer writer_;
|
|
|
|
Reader reader_;
|
|
|
|
|
2011-05-21 04:17:43 +02:00
|
|
|
// Record metadata for testing initial offset functionality
|
|
|
|
static size_t initial_offset_record_sizes_[];
|
|
|
|
static uint64_t initial_offset_last_record_offsets_[];
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
public:
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
LogTest() : reader_contents_(),
|
|
|
|
dest_holder_(new StringDest(reader_contents_)),
|
|
|
|
source_holder_(new StringSource(reader_contents_)),
|
2013-01-20 11:07:13 +01:00
|
|
|
writer_(std::move(dest_holder_)),
|
|
|
|
reader_(std::move(source_holder_), &report_, true/*checksum*/,
|
2011-05-21 04:17:43 +02:00
|
|
|
0/*initial_offset*/) {
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
void Write(const std::string& msg) {
|
|
|
|
writer_.AddRecord(Slice(msg));
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t WrittenBytes() const {
|
2013-01-20 11:07:13 +01:00
|
|
|
return dest_contents().size();
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
std::string Read() {
|
|
|
|
std::string scratch;
|
|
|
|
Slice record;
|
|
|
|
if (reader_.ReadRecord(&record, &scratch)) {
|
|
|
|
return record.ToString();
|
|
|
|
} else {
|
|
|
|
return "EOF";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void IncrementByte(int offset, int delta) {
|
2013-01-20 11:07:13 +01:00
|
|
|
dest_contents()[offset] += delta;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
void SetByte(int offset, char new_byte) {
|
2013-01-20 11:07:13 +01:00
|
|
|
dest_contents()[offset] = new_byte;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
void ShrinkSize(int bytes) {
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
auto dest = dynamic_cast<StringDest*>(writer_.file());
|
|
|
|
assert(dest);
|
|
|
|
dest->Drop(bytes);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
void FixChecksum(int header_offset, int len) {
|
|
|
|
// Compute crc of type/len/data
|
2013-01-20 11:07:13 +01:00
|
|
|
uint32_t crc = crc32c::Value(&dest_contents()[header_offset+6], 1 + len);
|
2011-03-18 23:37:00 +01:00
|
|
|
crc = crc32c::Mask(crc);
|
2013-01-20 11:07:13 +01:00
|
|
|
EncodeFixed32(&dest_contents()[header_offset], crc);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
void ForceError(size_t position = 0) {
|
2013-01-20 11:07:13 +01:00
|
|
|
auto src = dynamic_cast<StringSource*>(reader_.file());
|
|
|
|
src->force_error_ = true;
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
src->force_error_position_ = position;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
size_t DroppedBytes() const {
|
|
|
|
return report_.dropped_bytes_;
|
|
|
|
}
|
|
|
|
|
2011-05-21 04:17:43 +02:00
|
|
|
std::string ReportMessage() const {
|
|
|
|
return report_.message_;
|
|
|
|
}
|
|
|
|
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
void ForceEOF(size_t position = 0) {
|
|
|
|
auto src = dynamic_cast<StringSource*>(reader_.file());
|
|
|
|
src->force_eof_ = true;
|
|
|
|
src->force_eof_position_ = position;
|
|
|
|
}
|
|
|
|
|
|
|
|
void UnmarkEOF() {
|
|
|
|
auto src = dynamic_cast<StringSource*>(reader_.file());
|
|
|
|
src->returned_partial_ = false;
|
|
|
|
reader_.UnmarkEOF();
|
|
|
|
}
|
|
|
|
|
|
|
|
bool IsEOF() {
|
|
|
|
return reader_.IsEOF();
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
// Returns OK iff recorded error message contains "msg"
|
|
|
|
std::string MatchError(const std::string& msg) const {
|
|
|
|
if (report_.message_.find(msg) == std::string::npos) {
|
|
|
|
return report_.message_;
|
|
|
|
} else {
|
|
|
|
return "OK";
|
|
|
|
}
|
|
|
|
}
|
2011-05-21 04:17:43 +02:00
|
|
|
|
|
|
|
void WriteInitialOffsetLog() {
|
|
|
|
for (int i = 0; i < 4; i++) {
|
|
|
|
std::string record(initial_offset_record_sizes_[i],
|
|
|
|
static_cast<char>('a' + i));
|
|
|
|
Write(record);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void CheckOffsetPastEndReturnsNoRecords(uint64_t offset_past_end) {
|
|
|
|
WriteInitialOffsetLog();
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
unique_ptr<StringSource> source(new StringSource(reader_contents_));
|
2013-01-20 11:07:13 +01:00
|
|
|
unique_ptr<Reader> offset_reader(
|
|
|
|
new Reader(std::move(source), &report_, true/*checksum*/,
|
|
|
|
WrittenBytes() + offset_past_end));
|
2011-05-21 04:17:43 +02:00
|
|
|
Slice record;
|
|
|
|
std::string scratch;
|
|
|
|
ASSERT_TRUE(!offset_reader->ReadRecord(&record, &scratch));
|
|
|
|
}
|
|
|
|
|
|
|
|
void CheckInitialOffsetRecord(uint64_t initial_offset,
|
|
|
|
int expected_record_offset) {
|
|
|
|
WriteInitialOffsetLog();
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
unique_ptr<StringSource> source(new StringSource(reader_contents_));
|
2013-01-20 11:07:13 +01:00
|
|
|
unique_ptr<Reader> offset_reader(
|
|
|
|
new Reader(std::move(source), &report_, true/*checksum*/,
|
|
|
|
initial_offset));
|
2011-05-21 04:17:43 +02:00
|
|
|
Slice record;
|
|
|
|
std::string scratch;
|
|
|
|
ASSERT_TRUE(offset_reader->ReadRecord(&record, &scratch));
|
|
|
|
ASSERT_EQ(initial_offset_record_sizes_[expected_record_offset],
|
|
|
|
record.size());
|
|
|
|
ASSERT_EQ(initial_offset_last_record_offsets_[expected_record_offset],
|
|
|
|
offset_reader->LastRecordOffset());
|
|
|
|
ASSERT_EQ((char)('a' + expected_record_offset), record.data()[0]);
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
};
|
|
|
|
|
2011-05-21 04:17:43 +02:00
|
|
|
size_t LogTest::initial_offset_record_sizes_[] =
|
|
|
|
{10000, // Two sizable records in first block
|
|
|
|
10000,
|
|
|
|
2 * log::kBlockSize - 1000, // Span three blocks
|
|
|
|
1};
|
|
|
|
|
|
|
|
uint64_t LogTest::initial_offset_last_record_offsets_[] =
|
|
|
|
{0,
|
|
|
|
kHeaderSize + 10000,
|
|
|
|
2 * (kHeaderSize + 10000),
|
|
|
|
2 * (kHeaderSize + 10000) +
|
|
|
|
(2 * log::kBlockSize - 1000) + 3 * kHeaderSize};
|
|
|
|
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
TEST(LogTest, Empty) {
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadWrite) {
|
|
|
|
Write("foo");
|
|
|
|
Write("bar");
|
|
|
|
Write("");
|
|
|
|
Write("xxxx");
|
|
|
|
ASSERT_EQ("foo", Read());
|
|
|
|
ASSERT_EQ("bar", Read());
|
|
|
|
ASSERT_EQ("", Read());
|
|
|
|
ASSERT_EQ("xxxx", Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
ASSERT_EQ("EOF", Read()); // Make sure reads at eof work
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ManyBlocks) {
|
|
|
|
for (int i = 0; i < 100000; i++) {
|
|
|
|
Write(NumberString(i));
|
|
|
|
}
|
|
|
|
for (int i = 0; i < 100000; i++) {
|
|
|
|
ASSERT_EQ(NumberString(i), Read());
|
|
|
|
}
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, Fragmentation) {
|
|
|
|
Write("small");
|
|
|
|
Write(BigString("medium", 50000));
|
|
|
|
Write(BigString("large", 100000));
|
|
|
|
ASSERT_EQ("small", Read());
|
|
|
|
ASSERT_EQ(BigString("medium", 50000), Read());
|
|
|
|
ASSERT_EQ(BigString("large", 100000), Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, MarginalTrailer) {
|
|
|
|
// Make a trailer that is exactly the same length as an empty record.
|
|
|
|
const int n = kBlockSize - 2*kHeaderSize;
|
|
|
|
Write(BigString("foo", n));
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ((unsigned int)(kBlockSize - kHeaderSize), WrittenBytes());
|
2011-03-18 23:37:00 +01:00
|
|
|
Write("");
|
|
|
|
Write("bar");
|
|
|
|
ASSERT_EQ(BigString("foo", n), Read());
|
|
|
|
ASSERT_EQ("", Read());
|
|
|
|
ASSERT_EQ("bar", Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
}
|
|
|
|
|
2011-05-21 04:17:43 +02:00
|
|
|
TEST(LogTest, MarginalTrailer2) {
|
|
|
|
// Make a trailer that is exactly the same length as an empty record.
|
|
|
|
const int n = kBlockSize - 2*kHeaderSize;
|
|
|
|
Write(BigString("foo", n));
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ((unsigned int)(kBlockSize - kHeaderSize), WrittenBytes());
|
2011-05-21 04:17:43 +02:00
|
|
|
Write("bar");
|
|
|
|
ASSERT_EQ(BigString("foo", n), Read());
|
|
|
|
ASSERT_EQ("bar", Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ(0U, DroppedBytes());
|
2011-05-21 04:17:43 +02:00
|
|
|
ASSERT_EQ("", ReportMessage());
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
TEST(LogTest, ShortTrailer) {
|
|
|
|
const int n = kBlockSize - 2*kHeaderSize + 4;
|
|
|
|
Write(BigString("foo", n));
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ((unsigned int)(kBlockSize - kHeaderSize + 4), WrittenBytes());
|
2011-03-18 23:37:00 +01:00
|
|
|
Write("");
|
|
|
|
Write("bar");
|
|
|
|
ASSERT_EQ(BigString("foo", n), Read());
|
|
|
|
ASSERT_EQ("", Read());
|
|
|
|
ASSERT_EQ("bar", Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, AlignedEof) {
|
|
|
|
const int n = kBlockSize - 2*kHeaderSize + 4;
|
|
|
|
Write(BigString("foo", n));
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ((unsigned int)(kBlockSize - kHeaderSize + 4), WrittenBytes());
|
2011-03-18 23:37:00 +01:00
|
|
|
ASSERT_EQ(BigString("foo", n), Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, RandomRead) {
|
|
|
|
const int N = 500;
|
|
|
|
Random write_rnd(301);
|
|
|
|
for (int i = 0; i < N; i++) {
|
|
|
|
Write(RandomSkewedString(i, &write_rnd));
|
|
|
|
}
|
|
|
|
Random read_rnd(301);
|
|
|
|
for (int i = 0; i < N; i++) {
|
|
|
|
ASSERT_EQ(RandomSkewedString(i, &read_rnd), Read());
|
|
|
|
}
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
}
|
|
|
|
|
|
|
|
// Tests of all the error paths in log_reader.cc follow:
|
|
|
|
|
|
|
|
TEST(LogTest, ReadError) {
|
|
|
|
Write("foo");
|
|
|
|
ForceError();
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ((unsigned int)kBlockSize, DroppedBytes());
|
2011-03-18 23:37:00 +01:00
|
|
|
ASSERT_EQ("OK", MatchError("read error"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, BadRecordType) {
|
|
|
|
Write("foo");
|
|
|
|
// Type is stored in header[6]
|
|
|
|
IncrementByte(6, 100);
|
|
|
|
FixChecksum(0, 3);
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ(3U, DroppedBytes());
|
2011-03-18 23:37:00 +01:00
|
|
|
ASSERT_EQ("OK", MatchError("unknown record type"));
|
|
|
|
}
|
|
|
|
|
2014-02-28 22:19:47 +01:00
|
|
|
TEST(LogTest, TruncatedTrailingRecordIsIgnored) {
|
2011-03-18 23:37:00 +01:00
|
|
|
Write("foo");
|
|
|
|
ShrinkSize(4); // Drop all payload as well as a header byte
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2014-02-28 22:19:47 +01:00
|
|
|
// Truncated last record is ignored, not treated as an error
|
2014-03-14 23:44:35 +01:00
|
|
|
ASSERT_EQ(0U, DroppedBytes());
|
2014-02-28 22:19:47 +01:00
|
|
|
ASSERT_EQ("", ReportMessage());
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, BadLength) {
|
2014-02-28 22:19:47 +01:00
|
|
|
const int kPayloadSize = kBlockSize - kHeaderSize;
|
|
|
|
Write(BigString("bar", kPayloadSize));
|
|
|
|
Write("foo");
|
|
|
|
// Least significant size byte is stored in header[4].
|
|
|
|
IncrementByte(4, 1);
|
|
|
|
ASSERT_EQ("foo", Read());
|
|
|
|
ASSERT_EQ(kBlockSize, DroppedBytes());
|
|
|
|
ASSERT_EQ("OK", MatchError("bad record length"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, BadLengthAtEndIsIgnored) {
|
2011-03-18 23:37:00 +01:00
|
|
|
Write("foo");
|
|
|
|
ShrinkSize(1);
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2014-03-14 23:44:35 +01:00
|
|
|
ASSERT_EQ(0U, DroppedBytes());
|
2014-02-28 22:19:47 +01:00
|
|
|
ASSERT_EQ("", ReportMessage());
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ChecksumMismatch) {
|
|
|
|
Write("foo");
|
|
|
|
IncrementByte(0, 10);
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ(10U, DroppedBytes());
|
2011-03-18 23:37:00 +01:00
|
|
|
ASSERT_EQ("OK", MatchError("checksum mismatch"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, UnexpectedMiddleType) {
|
|
|
|
Write("foo");
|
|
|
|
SetByte(6, kMiddleType);
|
|
|
|
FixChecksum(0, 3);
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ(3U, DroppedBytes());
|
2011-03-18 23:37:00 +01:00
|
|
|
ASSERT_EQ("OK", MatchError("missing start"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, UnexpectedLastType) {
|
|
|
|
Write("foo");
|
|
|
|
SetByte(6, kLastType);
|
|
|
|
FixChecksum(0, 3);
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ(3U, DroppedBytes());
|
2011-03-18 23:37:00 +01:00
|
|
|
ASSERT_EQ("OK", MatchError("missing start"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, UnexpectedFullType) {
|
|
|
|
Write("foo");
|
|
|
|
Write("bar");
|
|
|
|
SetByte(6, kFirstType);
|
|
|
|
FixChecksum(0, 3);
|
|
|
|
ASSERT_EQ("bar", Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ(3U, DroppedBytes());
|
2011-03-18 23:37:00 +01:00
|
|
|
ASSERT_EQ("OK", MatchError("partial record without end"));
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, UnexpectedFirstType) {
|
|
|
|
Write("foo");
|
|
|
|
Write(BigString("bar", 100000));
|
|
|
|
SetByte(6, kFirstType);
|
|
|
|
FixChecksum(0, 3);
|
|
|
|
ASSERT_EQ(BigString("bar", 100000), Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2012-11-06 21:02:18 +01:00
|
|
|
ASSERT_EQ(3U, DroppedBytes());
|
2011-03-18 23:37:00 +01:00
|
|
|
ASSERT_EQ("OK", MatchError("partial record without end"));
|
|
|
|
}
|
|
|
|
|
2014-02-28 22:19:47 +01:00
|
|
|
TEST(LogTest, MissingLastIsIgnored) {
|
|
|
|
Write(BigString("bar", kBlockSize));
|
|
|
|
// Remove the LAST block, including header.
|
|
|
|
ShrinkSize(14);
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
ASSERT_EQ("", ReportMessage());
|
2014-03-14 23:44:35 +01:00
|
|
|
ASSERT_EQ(0U, DroppedBytes());
|
2014-02-28 22:19:47 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, PartialLastIsIgnored) {
|
|
|
|
Write(BigString("bar", kBlockSize));
|
|
|
|
// Cause a bad record length in the LAST block.
|
|
|
|
ShrinkSize(1);
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
ASSERT_EQ("", ReportMessage());
|
2014-03-14 23:44:35 +01:00
|
|
|
ASSERT_EQ(0U, DroppedBytes());
|
2014-02-28 22:19:47 +01:00
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
TEST(LogTest, ErrorJoinsRecords) {
|
|
|
|
// Consider two fragmented records:
|
|
|
|
// first(R1) last(R1) first(R2) last(R2)
|
|
|
|
// where the middle two fragments disappear. We do not want
|
|
|
|
// first(R1),last(R2) to get joined and returned as a valid record.
|
|
|
|
|
|
|
|
// Write records that span two blocks
|
|
|
|
Write(BigString("foo", kBlockSize));
|
|
|
|
Write(BigString("bar", kBlockSize));
|
|
|
|
Write("correct");
|
|
|
|
|
|
|
|
// Wipe the middle block
|
2013-03-19 22:53:22 +01:00
|
|
|
for (unsigned int offset = kBlockSize; offset < 2*kBlockSize; offset++) {
|
2011-03-18 23:37:00 +01:00
|
|
|
SetByte(offset, 'x');
|
|
|
|
}
|
|
|
|
|
|
|
|
ASSERT_EQ("correct", Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
2014-11-11 22:47:22 +01:00
|
|
|
size_t dropped = DroppedBytes();
|
|
|
|
ASSERT_LE(dropped, 2 * kBlockSize + 100);
|
|
|
|
ASSERT_GE(dropped, 2 * kBlockSize);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2011-05-21 04:17:43 +02:00
|
|
|
TEST(LogTest, ReadStart) {
|
|
|
|
CheckInitialOffsetRecord(0, 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadSecondOneOff) {
|
|
|
|
CheckInitialOffsetRecord(1, 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadSecondTenThousand) {
|
|
|
|
CheckInitialOffsetRecord(10000, 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadSecondStart) {
|
|
|
|
CheckInitialOffsetRecord(10007, 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadThirdOneOff) {
|
|
|
|
CheckInitialOffsetRecord(10008, 2);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadThirdStart) {
|
|
|
|
CheckInitialOffsetRecord(20014, 2);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadFourthOneOff) {
|
|
|
|
CheckInitialOffsetRecord(20015, 3);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadFourthFirstBlockTrailer) {
|
|
|
|
CheckInitialOffsetRecord(log::kBlockSize - 4, 3);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadFourthMiddleBlock) {
|
|
|
|
CheckInitialOffsetRecord(log::kBlockSize + 1, 3);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadFourthLastBlock) {
|
|
|
|
CheckInitialOffsetRecord(2 * log::kBlockSize + 1, 3);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadFourthStart) {
|
|
|
|
CheckInitialOffsetRecord(
|
|
|
|
2 * (kHeaderSize + 1000) + (2 * log::kBlockSize - 1000) + 3 * kHeaderSize,
|
|
|
|
3);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadEnd) {
|
|
|
|
CheckOffsetPastEndReturnsNoRecords(0);
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ReadPastEnd) {
|
|
|
|
CheckOffsetPastEndReturnsNoRecords(5);
|
|
|
|
}
|
|
|
|
|
Fix UnmarkEOF for partial blocks
Summary:
Blocks in the transaction log are a fixed size, but the last block in the transaction log file is usually a partial block. When a new record is added after the reader hit the end of the file, a new physical record will be appended to the last block. ReadPhysicalRecord can only read full blocks and assumes that the file position indicator is aligned to the start of a block. If the reader is forced to read further by simply clearing the EOF flag, ReadPhysicalRecord will read a full block starting from somewhere in the middle of a real block, causing it to lose alignment and to have a partial physical record at the end of the read buffer. This will result in length mismatches and checksum failures. When the log file is tailed for replication this will cause the log iterator to become invalid, necessitating the creation of a new iterator which will have to read the log file from scratch.
This diff fixes this issue by reading the remaining portion of the last block we read from. This is done when the reader is forced to read further (UnmarkEOF is called).
Test Plan:
- Added unit tests
- Stress test (with replication). Check dbdir/LOG file for corruptions.
- Test on test tier
Reviewers: emayanke, haobo, dhruba
Reviewed By: haobo
CC: vamsi, sheki, dhruba, kailiu, igor
Differential Revision: https://reviews.facebook.net/D15249
2014-01-27 23:49:10 +01:00
|
|
|
TEST(LogTest, ClearEofSingleBlock) {
|
|
|
|
Write("foo");
|
|
|
|
Write("bar");
|
|
|
|
ForceEOF(3 + kHeaderSize + 2);
|
|
|
|
ASSERT_EQ("foo", Read());
|
|
|
|
UnmarkEOF();
|
|
|
|
ASSERT_EQ("bar", Read());
|
|
|
|
ASSERT_TRUE(IsEOF());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
Write("xxx");
|
|
|
|
UnmarkEOF();
|
|
|
|
ASSERT_EQ("xxx", Read());
|
|
|
|
ASSERT_TRUE(IsEOF());
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ClearEofMultiBlock) {
|
|
|
|
size_t num_full_blocks = 5;
|
|
|
|
size_t n = (kBlockSize - kHeaderSize) * num_full_blocks + 25;
|
|
|
|
Write(BigString("foo", n));
|
|
|
|
Write(BigString("bar", n));
|
|
|
|
ForceEOF(n + num_full_blocks * kHeaderSize + 10);
|
|
|
|
ASSERT_EQ(BigString("foo", n), Read());
|
|
|
|
ASSERT_TRUE(IsEOF());
|
|
|
|
UnmarkEOF();
|
|
|
|
ASSERT_EQ(BigString("bar", n), Read());
|
|
|
|
ASSERT_TRUE(IsEOF());
|
|
|
|
Write(BigString("xxx", n));
|
|
|
|
UnmarkEOF();
|
|
|
|
ASSERT_EQ(BigString("xxx", n), Read());
|
|
|
|
ASSERT_TRUE(IsEOF());
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ClearEofError) {
|
|
|
|
// If an error occurs during Read() in UnmarkEOF(), the records contained
|
|
|
|
// in the buffer should be returned on subsequent calls of ReadRecord()
|
|
|
|
// until no more full records are left, whereafter ReadRecord() should return
|
|
|
|
// false to indicate that it cannot read any further.
|
|
|
|
|
|
|
|
Write("foo");
|
|
|
|
Write("bar");
|
|
|
|
UnmarkEOF();
|
|
|
|
ASSERT_EQ("foo", Read());
|
|
|
|
ASSERT_TRUE(IsEOF());
|
|
|
|
Write("xxx");
|
|
|
|
ForceError(0);
|
|
|
|
UnmarkEOF();
|
|
|
|
ASSERT_EQ("bar", Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
}
|
|
|
|
|
|
|
|
TEST(LogTest, ClearEofError2) {
|
|
|
|
Write("foo");
|
|
|
|
Write("bar");
|
|
|
|
UnmarkEOF();
|
|
|
|
ASSERT_EQ("foo", Read());
|
|
|
|
Write("xxx");
|
|
|
|
ForceError(3);
|
|
|
|
UnmarkEOF();
|
|
|
|
ASSERT_EQ("bar", Read());
|
|
|
|
ASSERT_EQ("EOF", Read());
|
|
|
|
ASSERT_EQ(3U, DroppedBytes());
|
|
|
|
ASSERT_EQ("OK", MatchError("read error"));
|
|
|
|
}
|
|
|
|
|
2011-10-31 18:22:06 +01:00
|
|
|
} // namespace log
|
2013-10-04 06:49:15 +02:00
|
|
|
} // namespace rocksdb
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
int main(int argc, char** argv) {
|
2013-10-04 06:49:15 +02:00
|
|
|
return rocksdb::test::RunAllTests();
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|