[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
|
2017-07-15 23:03:42 +00:00
|
|
|
// This source code is licensed under both the GPLv2 (found in the
|
|
|
|
// COPYING file in the root directory) and Apache 2.0 License
|
|
|
|
// (found in the LICENSE.Apache file in the root directory).
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
//
|
|
|
|
// Copyright 2014 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.
|
|
|
|
|
|
|
|
// This test uses a custom Env to keep track of the state of a filesystem as of
|
|
|
|
// the last "sync". It then checks for data loss errors by purposely dropping
|
|
|
|
// file data (or entire files) not protected by a "sync".
|
|
|
|
|
2020-07-09 21:33:42 +00:00
|
|
|
#include "utilities/fault_injection_env.h"
|
|
|
|
|
2016-12-16 19:17:26 +00:00
|
|
|
#include <functional>
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
#include <utility>
|
|
|
|
|
2020-07-09 21:33:42 +00:00
|
|
|
#include "util/random.h"
|
2020-02-20 20:07:53 +00:00
|
|
|
namespace ROCKSDB_NAMESPACE {
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
|
|
|
|
// Assume a filename, and not a directory name like "/foo/bar/"
|
|
|
|
std::string GetDirName(const std::string filename) {
|
|
|
|
size_t found = filename.find_last_of("/\\");
|
|
|
|
if (found == std::string::npos) {
|
|
|
|
return "";
|
|
|
|
} else {
|
|
|
|
return filename.substr(0, found);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// A basic file truncation function suitable for this test.
|
|
|
|
Status Truncate(Env* env, const std::string& filename, uint64_t length) {
|
2018-11-09 19:17:34 +00:00
|
|
|
std::unique_ptr<SequentialFile> orig_file;
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
const EnvOptions options;
|
|
|
|
Status s = env->NewSequentialFile(filename, &orig_file, options);
|
|
|
|
if (!s.ok()) {
|
2018-10-04 21:51:44 +00:00
|
|
|
fprintf(stderr, "Cannot open file %s for truncation: %s\n",
|
|
|
|
filename.c_str(), s.ToString().c_str());
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::unique_ptr<char[]> scratch(new char[length]);
|
2020-02-20 20:07:53 +00:00
|
|
|
ROCKSDB_NAMESPACE::Slice result;
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
s = orig_file->Read(length, &result, scratch.get());
|
|
|
|
#ifdef OS_WIN
|
|
|
|
orig_file.reset();
|
|
|
|
#endif
|
|
|
|
if (s.ok()) {
|
|
|
|
std::string tmp_name = GetDirName(filename) + "/truncate.tmp";
|
2018-11-09 19:17:34 +00:00
|
|
|
std::unique_ptr<WritableFile> tmp_file;
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
s = env->NewWritableFile(tmp_name, &tmp_file, options);
|
|
|
|
if (s.ok()) {
|
|
|
|
s = tmp_file->Append(result);
|
|
|
|
if (s.ok()) {
|
|
|
|
s = env->RenameFile(tmp_name, filename);
|
|
|
|
} else {
|
|
|
|
fprintf(stderr, "Cannot rename file %s to %s: %s\n", tmp_name.c_str(),
|
|
|
|
filename.c_str(), s.ToString().c_str());
|
|
|
|
env->DeleteFile(tmp_name);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (!s.ok()) {
|
|
|
|
fprintf(stderr, "Cannot truncate file %s: %s\n", filename.c_str(),
|
|
|
|
s.ToString().c_str());
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Trim the tailing "/" in the end of `str`
|
|
|
|
std::string TrimDirname(const std::string& str) {
|
|
|
|
size_t found = str.find_last_not_of("/");
|
|
|
|
if (found == std::string::npos) {
|
|
|
|
return str;
|
|
|
|
}
|
|
|
|
return str.substr(0, found + 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Return pair <parent directory name, file name> of a full path.
|
|
|
|
std::pair<std::string, std::string> GetDirAndName(const std::string& name) {
|
|
|
|
std::string dirname = GetDirName(name);
|
|
|
|
std::string fname = name.substr(dirname.size() + 1);
|
|
|
|
return std::make_pair(dirname, fname);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status FileState::DropUnsyncedData(Env* env) const {
|
|
|
|
ssize_t sync_pos = pos_at_last_sync_ == -1 ? 0 : pos_at_last_sync_;
|
|
|
|
return Truncate(env, filename_, sync_pos);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status FileState::DropRandomUnsyncedData(Env* env, Random* rand) const {
|
|
|
|
ssize_t sync_pos = pos_at_last_sync_ == -1 ? 0 : pos_at_last_sync_;
|
|
|
|
assert(pos_ >= sync_pos);
|
|
|
|
int range = static_cast<int>(pos_ - sync_pos);
|
|
|
|
uint64_t truncated_size =
|
|
|
|
static_cast<uint64_t>(sync_pos) + rand->Uniform(range);
|
|
|
|
return Truncate(env, filename_, truncated_size);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestDirectory::Fsync() {
|
2019-06-21 17:12:29 +00:00
|
|
|
if (!env_->IsFilesystemActive()) {
|
|
|
|
return env_->GetError();
|
|
|
|
}
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
env_->SyncDir(dirname_);
|
|
|
|
return dir_->Fsync();
|
|
|
|
}
|
|
|
|
|
Explicitly closing all directory file descriptors (#10049)
Summary:
Currently, the DB directory file descriptor is left open until the deconstruction process (`DB::Close()` does not close the file descriptor). To verify this, comment out the lines between `db_ = nullptr` and `db_->Close()` (line 512, 513, 514, 515 in ldb_cmd.cc) to leak the ``db_'' object, build `ldb` tool and run
```
strace --trace=open,openat,close ./ldb --db=$TEST_TMPDIR --ignore_unknown_options put K1 V1 --create_if_missing
```
There is one directory file descriptor that is not closed in the strace log.
Pull Request resolved: https://github.com/facebook/rocksdb/pull/10049
Test Plan: Add a new unit test DBBasicTest.DBCloseAllDirectoryFDs: Open a database with different WAL directory and three different data directories, and all directory file descriptors should be closed after calling Close(). Explicitly call Close() after a directory file descriptor is not used so that the counter of directory open and close should be equivalent.
Reviewed By: ajkr, hx235
Differential Revision: D36722135
Pulled By: littlepig2013
fbshipit-source-id: 07bdc2abc417c6b30997b9bbef1f79aa757b21ff
2022-06-02 01:03:34 +00:00
|
|
|
Status TestDirectory::Close() {
|
|
|
|
if (!env_->IsFilesystemActive()) {
|
|
|
|
return env_->GetError();
|
|
|
|
}
|
|
|
|
return dir_->Close();
|
|
|
|
}
|
|
|
|
|
2020-09-29 00:30:22 +00:00
|
|
|
TestRandomAccessFile::TestRandomAccessFile(
|
|
|
|
std::unique_ptr<RandomAccessFile>&& target, FaultInjectionTestEnv* env)
|
|
|
|
: target_(std::move(target)), env_(env) {
|
|
|
|
assert(target_);
|
|
|
|
assert(env_);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestRandomAccessFile::Read(uint64_t offset, size_t n, Slice* result,
|
|
|
|
char* scratch) const {
|
|
|
|
assert(env_);
|
|
|
|
if (!env_->IsFilesystemActive()) {
|
|
|
|
return env_->GetError();
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(target_);
|
|
|
|
return target_->Read(offset, n, result, scratch);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestRandomAccessFile::Prefetch(uint64_t offset, size_t n) {
|
|
|
|
assert(env_);
|
|
|
|
if (!env_->IsFilesystemActive()) {
|
|
|
|
return env_->GetError();
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(target_);
|
|
|
|
return target_->Prefetch(offset, n);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestRandomAccessFile::MultiRead(ReadRequest* reqs, size_t num_reqs) {
|
|
|
|
assert(env_);
|
|
|
|
if (!env_->IsFilesystemActive()) {
|
|
|
|
const Status s = env_->GetError();
|
|
|
|
|
|
|
|
assert(reqs);
|
|
|
|
for (size_t i = 0; i < num_reqs; ++i) {
|
|
|
|
reqs[i].status = s;
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(target_);
|
|
|
|
return target_->MultiRead(reqs, num_reqs);
|
|
|
|
}
|
|
|
|
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
TestWritableFile::TestWritableFile(const std::string& fname,
|
2018-11-09 19:17:34 +00:00
|
|
|
std::unique_ptr<WritableFile>&& f,
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
FaultInjectionTestEnv* env)
|
|
|
|
: state_(fname),
|
|
|
|
target_(std::move(f)),
|
|
|
|
writable_file_opened_(true),
|
|
|
|
env_(env) {
|
|
|
|
assert(target_ != nullptr);
|
|
|
|
state_.pos_ = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
TestWritableFile::~TestWritableFile() {
|
|
|
|
if (writable_file_opened_) {
|
2020-12-24 00:54:05 +00:00
|
|
|
Close().PermitUncheckedError();
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestWritableFile::Append(const Slice& data) {
|
|
|
|
if (!env_->IsFilesystemActive()) {
|
2018-06-28 19:23:57 +00:00
|
|
|
return env_->GetError();
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
}
|
|
|
|
Status s = target_->Append(data);
|
|
|
|
if (s.ok()) {
|
|
|
|
state_.pos_ += data.size();
|
2019-02-13 03:07:25 +00:00
|
|
|
env_->WritableFileAppended(state_);
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestWritableFile::Close() {
|
|
|
|
writable_file_opened_ = false;
|
|
|
|
Status s = target_->Close();
|
|
|
|
if (s.ok()) {
|
|
|
|
env_->WritableFileClosed(state_);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestWritableFile::Flush() {
|
|
|
|
Status s = target_->Flush();
|
|
|
|
if (s.ok() && env_->IsFilesystemActive()) {
|
|
|
|
state_.pos_at_last_flush_ = state_.pos_;
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestWritableFile::Sync() {
|
|
|
|
if (!env_->IsFilesystemActive()) {
|
2017-01-20 07:03:45 +00:00
|
|
|
return Status::IOError("FaultInjectionTestEnv: not active");
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
}
|
|
|
|
// No need to actual sync.
|
|
|
|
state_.pos_at_last_sync_ = state_.pos_;
|
2019-02-13 03:07:25 +00:00
|
|
|
env_->WritableFileSynced(state_);
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2019-06-21 17:12:29 +00:00
|
|
|
TestRandomRWFile::TestRandomRWFile(const std::string& /*fname*/,
|
|
|
|
std::unique_ptr<RandomRWFile>&& f,
|
|
|
|
FaultInjectionTestEnv* env)
|
|
|
|
: target_(std::move(f)), file_opened_(true), env_(env) {
|
|
|
|
assert(target_ != nullptr);
|
|
|
|
}
|
|
|
|
|
|
|
|
TestRandomRWFile::~TestRandomRWFile() {
|
|
|
|
if (file_opened_) {
|
2020-12-10 05:19:55 +00:00
|
|
|
Close().PermitUncheckedError();
|
2019-06-21 17:12:29 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestRandomRWFile::Write(uint64_t offset, const Slice& data) {
|
|
|
|
if (!env_->IsFilesystemActive()) {
|
|
|
|
return env_->GetError();
|
|
|
|
}
|
|
|
|
return target_->Write(offset, data);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestRandomRWFile::Read(uint64_t offset, size_t n, Slice* result,
|
|
|
|
char* scratch) const {
|
|
|
|
if (!env_->IsFilesystemActive()) {
|
|
|
|
return env_->GetError();
|
|
|
|
}
|
|
|
|
return target_->Read(offset, n, result, scratch);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestRandomRWFile::Close() {
|
|
|
|
file_opened_ = false;
|
|
|
|
return target_->Close();
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestRandomRWFile::Flush() {
|
|
|
|
if (!env_->IsFilesystemActive()) {
|
|
|
|
return env_->GetError();
|
|
|
|
}
|
|
|
|
return target_->Flush();
|
|
|
|
}
|
|
|
|
|
|
|
|
Status TestRandomRWFile::Sync() {
|
|
|
|
if (!env_->IsFilesystemActive()) {
|
|
|
|
return env_->GetError();
|
|
|
|
}
|
|
|
|
return target_->Sync();
|
|
|
|
}
|
|
|
|
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
Status FaultInjectionTestEnv::NewDirectory(const std::string& name,
|
2018-11-09 19:17:34 +00:00
|
|
|
std::unique_ptr<Directory>* result) {
|
|
|
|
std::unique_ptr<Directory> r;
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
Status s = target()->NewDirectory(name, &r);
|
|
|
|
assert(s.ok());
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
result->reset(new TestDirectory(this, TrimDirname(name), r.release()));
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2018-11-09 19:17:34 +00:00
|
|
|
Status FaultInjectionTestEnv::NewWritableFile(
|
|
|
|
const std::string& fname, std::unique_ptr<WritableFile>* result,
|
|
|
|
const EnvOptions& soptions) {
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
if (!IsFilesystemActive()) {
|
2018-06-28 19:23:57 +00:00
|
|
|
return GetError();
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
}
|
|
|
|
// Not allow overwriting files
|
|
|
|
Status s = target()->FileExists(fname);
|
|
|
|
if (s.ok()) {
|
|
|
|
return Status::Corruption("File already exists.");
|
|
|
|
} else if (!s.IsNotFound()) {
|
|
|
|
assert(s.IsIOError());
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
s = target()->NewWritableFile(fname, result, soptions);
|
|
|
|
if (s.ok()) {
|
|
|
|
result->reset(new TestWritableFile(fname, std::move(*result), this));
|
|
|
|
// WritableFileWriter* file is opened
|
|
|
|
// again then it will be truncated - so forget our saved state.
|
|
|
|
UntrackFile(fname);
|
|
|
|
MutexLock l(&mutex_);
|
2021-10-11 23:22:10 +00:00
|
|
|
open_managed_files_.insert(fname);
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
auto dir_and_name = GetDirAndName(fname);
|
|
|
|
auto& list = dir_to_new_files_since_last_sync_[dir_and_name.first];
|
|
|
|
list.insert(dir_and_name.second);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2018-10-23 22:01:36 +00:00
|
|
|
Status FaultInjectionTestEnv::ReopenWritableFile(
|
2018-11-09 19:17:34 +00:00
|
|
|
const std::string& fname, std::unique_ptr<WritableFile>* result,
|
2018-10-23 22:01:36 +00:00
|
|
|
const EnvOptions& soptions) {
|
|
|
|
if (!IsFilesystemActive()) {
|
|
|
|
return GetError();
|
|
|
|
}
|
2021-10-11 23:22:10 +00:00
|
|
|
|
|
|
|
bool exists;
|
|
|
|
Status s, exists_s = target()->FileExists(fname);
|
|
|
|
if (exists_s.IsNotFound()) {
|
|
|
|
exists = false;
|
|
|
|
} else if (exists_s.ok()) {
|
|
|
|
exists = true;
|
|
|
|
} else {
|
|
|
|
s = exists_s;
|
|
|
|
exists = false;
|
|
|
|
}
|
|
|
|
|
2018-10-23 22:01:36 +00:00
|
|
|
if (s.ok()) {
|
2021-10-11 23:22:10 +00:00
|
|
|
s = target()->ReopenWritableFile(fname, result, soptions);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Only track files we created. Files created outside of this
|
|
|
|
// `FaultInjectionTestEnv` are not eligible for tracking/data dropping
|
|
|
|
// (for example, they may contain data a previous db_stress run expects to
|
|
|
|
// be recovered). This could be extended to track/drop data appended once
|
|
|
|
// the file is under `FaultInjectionTestEnv`'s control.
|
|
|
|
if (s.ok()) {
|
|
|
|
bool should_track;
|
|
|
|
{
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
if (db_file_state_.find(fname) != db_file_state_.end()) {
|
|
|
|
// It was written by this `Env` earlier.
|
|
|
|
assert(exists);
|
|
|
|
should_track = true;
|
|
|
|
} else if (!exists) {
|
|
|
|
// It was created by this `Env` just now.
|
|
|
|
should_track = true;
|
|
|
|
open_managed_files_.insert(fname);
|
|
|
|
auto dir_and_name = GetDirAndName(fname);
|
|
|
|
auto& list = dir_to_new_files_since_last_sync_[dir_and_name.first];
|
|
|
|
list.insert(dir_and_name.second);
|
|
|
|
} else {
|
|
|
|
should_track = false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (should_track) {
|
|
|
|
result->reset(new TestWritableFile(fname, std::move(*result), this));
|
|
|
|
}
|
2018-10-23 22:01:36 +00:00
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2019-06-21 17:12:29 +00:00
|
|
|
Status FaultInjectionTestEnv::NewRandomRWFile(
|
|
|
|
const std::string& fname, std::unique_ptr<RandomRWFile>* result,
|
|
|
|
const EnvOptions& soptions) {
|
|
|
|
if (!IsFilesystemActive()) {
|
|
|
|
return GetError();
|
|
|
|
}
|
|
|
|
Status s = target()->NewRandomRWFile(fname, result, soptions);
|
|
|
|
if (s.ok()) {
|
|
|
|
result->reset(new TestRandomRWFile(fname, std::move(*result), this));
|
|
|
|
// WritableFileWriter* file is opened
|
|
|
|
// again then it will be truncated - so forget our saved state.
|
|
|
|
UntrackFile(fname);
|
|
|
|
MutexLock l(&mutex_);
|
2021-10-11 23:22:10 +00:00
|
|
|
open_managed_files_.insert(fname);
|
2019-06-21 17:12:29 +00:00
|
|
|
auto dir_and_name = GetDirAndName(fname);
|
|
|
|
auto& list = dir_to_new_files_since_last_sync_[dir_and_name.first];
|
|
|
|
list.insert(dir_and_name.second);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2018-09-20 23:50:07 +00:00
|
|
|
Status FaultInjectionTestEnv::NewRandomAccessFile(
|
|
|
|
const std::string& fname, std::unique_ptr<RandomAccessFile>* result,
|
|
|
|
const EnvOptions& soptions) {
|
|
|
|
if (!IsFilesystemActive()) {
|
|
|
|
return GetError();
|
|
|
|
}
|
2020-09-29 00:30:22 +00:00
|
|
|
|
|
|
|
assert(target());
|
|
|
|
const Status s = target()->NewRandomAccessFile(fname, result, soptions);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(result);
|
|
|
|
result->reset(new TestRandomAccessFile(std::move(*result), this));
|
|
|
|
|
|
|
|
return Status::OK();
|
2018-09-20 23:50:07 +00:00
|
|
|
}
|
|
|
|
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
Status FaultInjectionTestEnv::DeleteFile(const std::string& f) {
|
|
|
|
if (!IsFilesystemActive()) {
|
2018-06-28 19:23:57 +00:00
|
|
|
return GetError();
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
}
|
|
|
|
Status s = EnvWrapper::DeleteFile(f);
|
|
|
|
if (s.ok()) {
|
|
|
|
UntrackFile(f);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status FaultInjectionTestEnv::RenameFile(const std::string& s,
|
|
|
|
const std::string& t) {
|
|
|
|
if (!IsFilesystemActive()) {
|
2018-06-28 19:23:57 +00:00
|
|
|
return GetError();
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
}
|
|
|
|
Status ret = EnvWrapper::RenameFile(s, t);
|
|
|
|
|
|
|
|
if (ret.ok()) {
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
if (db_file_state_.find(s) != db_file_state_.end()) {
|
|
|
|
db_file_state_[t] = db_file_state_[s];
|
|
|
|
db_file_state_.erase(s);
|
|
|
|
}
|
|
|
|
|
|
|
|
auto sdn = GetDirAndName(s);
|
|
|
|
auto tdn = GetDirAndName(t);
|
|
|
|
if (dir_to_new_files_since_last_sync_[sdn.first].erase(sdn.second) != 0) {
|
|
|
|
auto& tlist = dir_to_new_files_since_last_sync_[tdn.first];
|
|
|
|
assert(tlist.find(tdn.second) == tlist.end());
|
|
|
|
tlist.insert(tdn.second);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2021-10-11 23:22:10 +00:00
|
|
|
Status FaultInjectionTestEnv::LinkFile(const std::string& s,
|
|
|
|
const std::string& t) {
|
|
|
|
if (!IsFilesystemActive()) {
|
|
|
|
return GetError();
|
|
|
|
}
|
|
|
|
Status ret = EnvWrapper::LinkFile(s, t);
|
|
|
|
|
|
|
|
if (ret.ok()) {
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
if (db_file_state_.find(s) != db_file_state_.end()) {
|
|
|
|
db_file_state_[t] = db_file_state_[s];
|
|
|
|
}
|
|
|
|
|
|
|
|
auto sdn = GetDirAndName(s);
|
|
|
|
auto tdn = GetDirAndName(t);
|
|
|
|
if (dir_to_new_files_since_last_sync_[sdn.first].find(sdn.second) !=
|
|
|
|
dir_to_new_files_since_last_sync_[sdn.first].end()) {
|
|
|
|
auto& tlist = dir_to_new_files_since_last_sync_[tdn.first];
|
|
|
|
assert(tlist.find(tdn.second) == tlist.end());
|
|
|
|
tlist.insert(tdn.second);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
void FaultInjectionTestEnv::WritableFileClosed(const FileState& state) {
|
|
|
|
MutexLock l(&mutex_);
|
2021-10-11 23:22:10 +00:00
|
|
|
if (open_managed_files_.find(state.filename_) != open_managed_files_.end()) {
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
db_file_state_[state.filename_] = state;
|
2021-10-11 23:22:10 +00:00
|
|
|
open_managed_files_.erase(state.filename_);
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-02-13 03:07:25 +00:00
|
|
|
void FaultInjectionTestEnv::WritableFileSynced(const FileState& state) {
|
|
|
|
MutexLock l(&mutex_);
|
2021-10-11 23:22:10 +00:00
|
|
|
if (open_managed_files_.find(state.filename_) != open_managed_files_.end()) {
|
2019-02-13 03:07:25 +00:00
|
|
|
if (db_file_state_.find(state.filename_) == db_file_state_.end()) {
|
|
|
|
db_file_state_.insert(std::make_pair(state.filename_, state));
|
|
|
|
} else {
|
|
|
|
db_file_state_[state.filename_] = state;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void FaultInjectionTestEnv::WritableFileAppended(const FileState& state) {
|
|
|
|
MutexLock l(&mutex_);
|
2021-10-11 23:22:10 +00:00
|
|
|
if (open_managed_files_.find(state.filename_) != open_managed_files_.end()) {
|
2019-02-13 03:07:25 +00:00
|
|
|
if (db_file_state_.find(state.filename_) == db_file_state_.end()) {
|
|
|
|
db_file_state_.insert(std::make_pair(state.filename_, state));
|
|
|
|
} else {
|
|
|
|
db_file_state_[state.filename_] = state;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
// For every file that is not fully synced, make a call to `func` with
|
|
|
|
// FileState of the file as the parameter.
|
|
|
|
Status FaultInjectionTestEnv::DropFileData(
|
|
|
|
std::function<Status(Env*, FileState)> func) {
|
|
|
|
Status s;
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
for (std::map<std::string, FileState>::const_iterator it =
|
|
|
|
db_file_state_.begin();
|
|
|
|
s.ok() && it != db_file_state_.end(); ++it) {
|
|
|
|
const FileState& state = it->second;
|
|
|
|
if (!state.IsFullySynced()) {
|
|
|
|
s = func(target(), state);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status FaultInjectionTestEnv::DropUnsyncedFileData() {
|
|
|
|
return DropFileData([&](Env* env, const FileState& state) {
|
|
|
|
return state.DropUnsyncedData(env);
|
|
|
|
});
|
|
|
|
}
|
|
|
|
|
|
|
|
Status FaultInjectionTestEnv::DropRandomUnsyncedFileData(Random* rnd) {
|
|
|
|
return DropFileData([&](Env* env, const FileState& state) {
|
|
|
|
return state.DropRandomUnsyncedData(env, rnd);
|
|
|
|
});
|
|
|
|
}
|
|
|
|
|
|
|
|
Status FaultInjectionTestEnv::DeleteFilesCreatedAfterLastDirSync() {
|
|
|
|
// Because DeleteFile access this container make a copy to avoid deadlock
|
|
|
|
std::map<std::string, std::set<std::string>> map_copy;
|
|
|
|
{
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
map_copy.insert(dir_to_new_files_since_last_sync_.begin(),
|
|
|
|
dir_to_new_files_since_last_sync_.end());
|
|
|
|
}
|
|
|
|
|
|
|
|
for (auto& pair : map_copy) {
|
|
|
|
for (std::string name : pair.second) {
|
|
|
|
Status s = DeleteFile(pair.first + "/" + name);
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
void FaultInjectionTestEnv::ResetState() {
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
db_file_state_.clear();
|
|
|
|
dir_to_new_files_since_last_sync_.clear();
|
|
|
|
SetFilesystemActiveNoLock(true);
|
|
|
|
}
|
|
|
|
|
|
|
|
void FaultInjectionTestEnv::UntrackFile(const std::string& f) {
|
|
|
|
MutexLock l(&mutex_);
|
|
|
|
auto dir_and_name = GetDirAndName(f);
|
|
|
|
dir_to_new_files_since_last_sync_[dir_and_name.first].erase(
|
|
|
|
dir_and_name.second);
|
|
|
|
db_file_state_.erase(f);
|
2021-10-11 23:22:10 +00:00
|
|
|
open_managed_files_.erase(f);
|
[rocksdb] Recovery path sequence miscount fix
Summary:
Consider the following WAL with 4 batch entries prefixed with their sequence at time of memtable insert.
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(a)]
[1: BEGIN_PREPARE, PUT, PUT, PUT, PUT, END_PREPARE(b)]
[4: COMMIT(a)]
[7: COMMIT(b)]
The first two batches do not consume any sequence numbers so are both prefixed with seq=1.
For 2pc commit, memtable insertion takes place before COMMIT batch is written to WAL.
We can see that sequence number consumption takes place between WAL entries giving us the seemingly sparse sequence prefix for WAL entries.
This is a valid WAL.
Because with 2PC markers one WriteBatch points to another batch containing its inserts a writebatch can consume more or less sequence numbers than the number of sequence consuming entries that it contains.
We can see that, given the entries in the WAL, 6 sequence ids were consumed. Yet on recovery the maximum sequence consumed would be 7 + 3 (the number of sequence numbers consumed by COMMIT(b))
So, now upon recovery we must track the actual consumption of sequence numbers.
In the provided scenario there will be no sequence gaps, but it is possible to produce a sequence gap. This should not be a problem though. correct?
Test Plan: provided test.
Reviewers: sdong
Subscribers: andrewkr, leveldb, dhruba, hermanlee4
Differential Revision: https://reviews.facebook.net/D57645
2016-05-04 21:02:27 +00:00
|
|
|
}
|
2020-02-20 20:07:53 +00:00
|
|
|
} // namespace ROCKSDB_NAMESPACE
|