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 <deque>
|
2012-08-18 09:26:50 +02:00
|
|
|
#include <set>
|
2011-03-18 23:37:00 +01:00
|
|
|
#include <dirent.h>
|
|
|
|
#include <errno.h>
|
|
|
|
#include <fcntl.h>
|
|
|
|
#include <pthread.h>
|
|
|
|
#include <stdio.h>
|
|
|
|
#include <stdlib.h>
|
|
|
|
#include <string.h>
|
2013-02-01 00:20:24 +01:00
|
|
|
#include <sys/ioctl.h>
|
2011-03-18 23:37:00 +01:00
|
|
|
#include <sys/mman.h>
|
|
|
|
#include <sys/stat.h>
|
2013-11-17 08:44:39 +01:00
|
|
|
#ifdef OS_LINUX
|
2013-03-13 21:50:26 +01:00
|
|
|
#include <sys/statfs.h>
|
2013-11-17 08:44:39 +01:00
|
|
|
#endif
|
2011-03-18 23:37:00 +01:00
|
|
|
#include <sys/time.h>
|
|
|
|
#include <sys/types.h>
|
|
|
|
#include <time.h>
|
|
|
|
#include <unistd.h>
|
2013-02-01 00:20:24 +01:00
|
|
|
#if defined(OS_LINUX)
|
|
|
|
#include <linux/fs.h>
|
[RocksDB] Sync file to disk incrementally
Summary:
During compaction, we sync the output files after they are fully written out. This causes unnecessary blocking of the compaction thread and burstiness of the write traffic.
This diff simply asks the OS to sync data incrementally as they are written, on the background. The hope is that, at the final sync, most of the data are already on disk and we would block less on the sync call. Thus, each compaction runs faster and we could use fewer number of compaction threads to saturate IO.
In addition, the write traffic will be smoothed out, hopefully reducing the IO P99 latency too.
Some quick tests show 10~20% improvement in per thread compaction throughput. Combined with posix advice on compaction read, just 5 threads are enough to almost saturate the udb flash bandwidth for 800 bytes write only benchmark.
What's more promising is that, with saturated IO, iostat shows average wait time is actually smoother and much smaller.
For the write only test 800bytes test:
Before the change: await occillate between 10ms and 3ms
After the change: await ranges 1-3ms
Will test against read-modify-write workload too, see if high read latency P99 could be resolved.
Will introduce a parameter to control the sync interval in a follow up diff after cleaning up EnvOptions.
Test Plan: make check; db_bench; db_stress
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11115
2013-06-05 00:51:50 +02:00
|
|
|
#include <fcntl.h>
|
2013-02-01 00:20:24 +01:00
|
|
|
#endif
|
2011-03-18 23:37:00 +01:00
|
|
|
#if defined(LEVELDB_PLATFORM_ANDROID)
|
|
|
|
#include <sys/stat.h>
|
|
|
|
#endif
|
2013-08-23 17:38:13 +02:00
|
|
|
#include "rocksdb/env.h"
|
|
|
|
#include "rocksdb/slice.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "port/port.h"
|
2013-02-01 00:20:24 +01:00
|
|
|
#include "util/coding.h"
|
2011-03-18 23:37:00 +01:00
|
|
|
#include "util/logging.h"
|
2011-07-21 04:40:18 +02:00
|
|
|
#include "util/posix_logger.h"
|
2013-04-05 08:49:43 +02:00
|
|
|
#include "util/random.h"
|
|
|
|
#include <signal.h>
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-11-17 08:44:39 +01:00
|
|
|
// Get nano time for mach systems
|
|
|
|
#ifdef __MACH__
|
|
|
|
#include <mach/clock.h>
|
|
|
|
#include <mach/mach.h>
|
|
|
|
#endif
|
|
|
|
|
2013-03-13 21:50:26 +01:00
|
|
|
#if !defined(TMPFS_MAGIC)
|
|
|
|
#define TMPFS_MAGIC 0x01021994
|
|
|
|
#endif
|
|
|
|
#if !defined(XFS_SUPER_MAGIC)
|
|
|
|
#define XFS_SUPER_MAGIC 0x58465342
|
|
|
|
#endif
|
|
|
|
#if !defined(EXT4_SUPER_MAGIC)
|
|
|
|
#define EXT4_SUPER_MAGIC 0xEF53
|
|
|
|
#endif
|
|
|
|
|
2013-11-17 08:44:39 +01:00
|
|
|
// For non linux platform, the following macros are used only as place
|
|
|
|
// holder.
|
|
|
|
#ifndef OS_LINUX
|
|
|
|
#define POSIX_FADV_NORMAL 0 /* [MC1] no further special treatment */
|
|
|
|
#define POSIX_FADV_RANDOM 1 /* [MC1] expect random page refs */
|
|
|
|
#define POSIX_FADV_SEQUENTIAL 2 /* [MC1] expect sequential page refs */
|
|
|
|
#define POSIX_FADV_WILLNEED 3 /* [MC1] will need these pages */
|
|
|
|
#define POSIX_FADV_DONTNEED 4 /* [MC1] dont need these pages */
|
|
|
|
#endif
|
|
|
|
|
2013-04-05 08:49:43 +02:00
|
|
|
// This is only set from db_stress.cc and for testing only.
|
|
|
|
// If non-zero, kill at various points in source code with probability 1/this
|
2013-10-05 07:32:05 +02:00
|
|
|
int rocksdb_kill_odds = 0;
|
2013-04-05 08:49:43 +02:00
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
namespace rocksdb {
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
namespace {
|
|
|
|
|
2013-11-17 08:44:39 +01:00
|
|
|
// A wrapper for fadvise, if the platform doesn't support fadvise,
|
|
|
|
// it will simply return Status::NotSupport.
|
|
|
|
int Fadvise(int fd, off_t offset, size_t len, int advice) {
|
|
|
|
#ifdef OS_LINUX
|
|
|
|
return posix_fadvise(fd, offset, len, advice);
|
|
|
|
#else
|
|
|
|
return 0; // simply do nothing.
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
2012-08-18 09:26:50 +02:00
|
|
|
// list of pathnames that are locked
|
|
|
|
static std::set<std::string> lockedFiles;
|
|
|
|
static port::Mutex mutex_lockedFiles;
|
|
|
|
|
2011-07-15 02:20:57 +02:00
|
|
|
static Status IOError(const std::string& context, int err_number) {
|
|
|
|
return Status::IOError(context, strerror(err_number));
|
|
|
|
}
|
|
|
|
|
2013-04-05 08:49:43 +02:00
|
|
|
#ifdef NDEBUG
|
|
|
|
// empty in release build
|
2013-10-05 07:32:05 +02:00
|
|
|
#define TEST_KILL_RANDOM(rocksdb_kill_odds)
|
2013-04-05 08:49:43 +02:00
|
|
|
#else
|
|
|
|
|
|
|
|
// Kill the process with probablity 1/odds for testing.
|
|
|
|
static void TestKillRandom(int odds, const std::string& srcfile,
|
|
|
|
int srcline) {
|
|
|
|
time_t curtime = time(nullptr);
|
|
|
|
Random r((uint32_t)curtime);
|
|
|
|
|
|
|
|
assert(odds > 0);
|
|
|
|
bool crash = r.OneIn(odds);
|
|
|
|
if (crash) {
|
|
|
|
fprintf(stdout, "Crashing at %s:%d\n", srcfile.c_str(), srcline);
|
|
|
|
fflush(stdout);
|
|
|
|
kill(getpid(), SIGTERM);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// To avoid crashing always at some frequently executed codepaths (during
|
|
|
|
// kill random test), use this factor to reduce odds
|
|
|
|
#define REDUCE_ODDS 2
|
|
|
|
#define REDUCE_ODDS2 4
|
|
|
|
|
2013-10-05 07:32:05 +02:00
|
|
|
#define TEST_KILL_RANDOM(rocksdb_kill_odds) { \
|
|
|
|
if (rocksdb_kill_odds > 0) { \
|
|
|
|
TestKillRandom(rocksdb_kill_odds, __FILE__, __LINE__); \
|
2013-04-05 08:49:43 +02:00
|
|
|
} \
|
|
|
|
}
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
2013-09-02 08:23:40 +02:00
|
|
|
#if defined(OS_LINUX)
|
|
|
|
namespace {
|
|
|
|
static size_t GetUniqueIdFromFile(int fd, char* id, size_t max_size) {
|
|
|
|
if (max_size < kMaxVarint64Length*3) {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
struct stat buf;
|
|
|
|
int result = fstat(fd, &buf);
|
|
|
|
if (result == -1) {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
long version = 0;
|
|
|
|
result = ioctl(fd, FS_IOC_GETVERSION, &version);
|
|
|
|
if (result == -1) {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
uint64_t uversion = (uint64_t)version;
|
|
|
|
|
|
|
|
char* rid = id;
|
|
|
|
rid = EncodeVarint64(rid, buf.st_dev);
|
|
|
|
rid = EncodeVarint64(rid, buf.st_ino);
|
|
|
|
rid = EncodeVarint64(rid, uversion);
|
|
|
|
assert(rid >= id);
|
|
|
|
return static_cast<size_t>(rid-id);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
class PosixSequentialFile: public SequentialFile {
|
|
|
|
private:
|
|
|
|
std::string filename_;
|
|
|
|
FILE* file_;
|
2013-03-15 01:00:04 +01:00
|
|
|
int fd_;
|
2013-04-22 19:41:41 +02:00
|
|
|
bool use_os_buffer_;
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
public:
|
2013-03-15 01:00:04 +01:00
|
|
|
PosixSequentialFile(const std::string& fname, FILE* f,
|
|
|
|
const EnvOptions& options)
|
2013-04-22 19:41:41 +02:00
|
|
|
: filename_(fname), file_(f), fd_(fileno(f)),
|
2013-06-08 00:35:17 +02:00
|
|
|
use_os_buffer_(options.use_os_buffer) {
|
2013-03-15 01:00:04 +01:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
virtual ~PosixSequentialFile() { fclose(file_); }
|
|
|
|
|
|
|
|
virtual Status Read(size_t n, Slice* result, char* scratch) {
|
|
|
|
Status s;
|
|
|
|
size_t r = fread_unlocked(scratch, 1, n, file_);
|
|
|
|
*result = Slice(scratch, r);
|
|
|
|
if (r < n) {
|
|
|
|
if (feof(file_)) {
|
|
|
|
// We leave status as ok if we hit the end of the file
|
|
|
|
} else {
|
|
|
|
// A partial read with an error: return a non-ok status
|
2011-07-15 02:20:57 +02:00
|
|
|
s = IOError(filename_, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
2013-04-22 19:41:41 +02:00
|
|
|
if (!use_os_buffer_) {
|
2013-03-15 01:00:04 +01:00
|
|
|
// we need to fadvise away the entire range of pages because
|
|
|
|
// we do not want readahead pages to be cached.
|
2013-11-17 08:44:39 +01:00
|
|
|
Fadvise(fd_, 0, 0, POSIX_FADV_DONTNEED); // free OS pages
|
2013-03-15 01:00:04 +01:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
return s;
|
|
|
|
}
|
2011-05-21 04:17:43 +02:00
|
|
|
|
|
|
|
virtual Status Skip(uint64_t n) {
|
|
|
|
if (fseek(file_, n, SEEK_CUR)) {
|
2011-07-15 02:20:57 +02:00
|
|
|
return IOError(filename_, errno);
|
2011-05-21 04:17:43 +02:00
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
2013-09-21 08:00:13 +02:00
|
|
|
|
|
|
|
virtual Status InvalidateCache(size_t offset, size_t length) {
|
2013-11-17 08:44:39 +01:00
|
|
|
#ifndef OS_LINUX
|
|
|
|
return Status::OK();
|
|
|
|
#else
|
2013-09-21 08:00:13 +02:00
|
|
|
// free OS pages
|
2013-11-17 08:44:39 +01:00
|
|
|
int ret = Fadvise(fd_, offset, length, POSIX_FADV_DONTNEED);
|
2013-09-21 08:00:13 +02:00
|
|
|
if (ret == 0) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
return IOError(filename_, errno);
|
2013-11-17 08:44:39 +01:00
|
|
|
#endif
|
2013-09-21 08:00:13 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
};
|
|
|
|
|
2012-03-15 17:14:00 +01:00
|
|
|
// pread() based random-access
|
2011-03-18 23:37:00 +01:00
|
|
|
class PosixRandomAccessFile: public RandomAccessFile {
|
|
|
|
private:
|
|
|
|
std::string filename_;
|
|
|
|
int fd_;
|
2013-04-22 19:41:41 +02:00
|
|
|
bool use_os_buffer_;
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
public:
|
2013-03-15 01:00:04 +01:00
|
|
|
PosixRandomAccessFile(const std::string& fname, int fd,
|
|
|
|
const EnvOptions& options)
|
2013-06-08 00:35:17 +02:00
|
|
|
: filename_(fname), fd_(fd), use_os_buffer_(options.use_os_buffer) {
|
|
|
|
assert(!options.use_mmap_reads);
|
2012-09-13 20:54:53 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
virtual ~PosixRandomAccessFile() { close(fd_); }
|
|
|
|
|
|
|
|
virtual Status Read(uint64_t offset, size_t n, Slice* result,
|
|
|
|
char* scratch) const {
|
|
|
|
Status s;
|
|
|
|
ssize_t r = pread(fd_, scratch, n, static_cast<off_t>(offset));
|
|
|
|
*result = Slice(scratch, (r < 0) ? 0 : r);
|
|
|
|
if (r < 0) {
|
|
|
|
// An error: return a non-ok status
|
2011-07-15 02:20:57 +02:00
|
|
|
s = IOError(filename_, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-04-22 19:41:41 +02:00
|
|
|
if (!use_os_buffer_) {
|
2012-09-13 19:10:51 +02:00
|
|
|
// we need to fadvise away the entire range of pages because
|
|
|
|
// we do not want readahead pages to be cached.
|
2013-11-17 08:44:39 +01:00
|
|
|
Fadvise(fd_, 0, 0, POSIX_FADV_DONTNEED); // free OS pages
|
2012-05-25 00:54:02 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
return s;
|
|
|
|
}
|
2013-02-01 00:20:24 +01:00
|
|
|
|
2013-11-17 08:44:39 +01:00
|
|
|
#ifdef OS_LINUX
|
2013-02-01 00:20:24 +01:00
|
|
|
virtual size_t GetUniqueId(char* id, size_t max_size) const {
|
2013-09-02 08:23:40 +02:00
|
|
|
return GetUniqueIdFromFile(fd_, id, max_size);
|
2013-02-01 00:20:24 +01:00
|
|
|
}
|
|
|
|
#endif
|
2013-05-18 00:53:01 +02:00
|
|
|
|
|
|
|
virtual void Hint(AccessPattern pattern) {
|
|
|
|
switch(pattern) {
|
|
|
|
case NORMAL:
|
2013-11-17 08:44:39 +01:00
|
|
|
Fadvise(fd_, 0, 0, POSIX_FADV_NORMAL);
|
2013-05-18 00:53:01 +02:00
|
|
|
break;
|
|
|
|
case RANDOM:
|
2013-11-17 08:44:39 +01:00
|
|
|
Fadvise(fd_, 0, 0, POSIX_FADV_RANDOM);
|
2013-05-18 00:53:01 +02:00
|
|
|
break;
|
|
|
|
case SEQUENTIAL:
|
2013-11-17 08:44:39 +01:00
|
|
|
Fadvise(fd_, 0, 0, POSIX_FADV_SEQUENTIAL);
|
2013-05-18 00:53:01 +02:00
|
|
|
break;
|
|
|
|
case WILLNEED:
|
2013-11-17 08:44:39 +01:00
|
|
|
Fadvise(fd_, 0, 0, POSIX_FADV_WILLNEED);
|
2013-05-18 00:53:01 +02:00
|
|
|
break;
|
|
|
|
case DONTNEED:
|
2013-11-17 08:44:39 +01:00
|
|
|
Fadvise(fd_, 0, 0, POSIX_FADV_DONTNEED);
|
2013-05-18 00:53:01 +02:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
assert(false);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-09-21 08:00:13 +02:00
|
|
|
virtual Status InvalidateCache(size_t offset, size_t length) {
|
2013-11-17 08:44:39 +01:00
|
|
|
#ifndef OS_LINUX
|
|
|
|
return Status::OK();
|
|
|
|
#else
|
2013-09-21 08:00:13 +02:00
|
|
|
// free OS pages
|
2013-11-17 08:44:39 +01:00
|
|
|
int ret = Fadvise(fd_, offset, length, POSIX_FADV_DONTNEED);
|
2013-09-21 08:00:13 +02:00
|
|
|
if (ret == 0) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
return IOError(filename_, errno);
|
2013-11-17 08:44:39 +01:00
|
|
|
#endif
|
2013-09-21 08:00:13 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
};
|
|
|
|
|
2012-03-15 17:14:00 +01:00
|
|
|
// mmap() based random-access
|
|
|
|
class PosixMmapReadableFile: public RandomAccessFile {
|
|
|
|
private:
|
2013-09-21 08:00:13 +02:00
|
|
|
int fd_;
|
2012-03-15 17:14:00 +01:00
|
|
|
std::string filename_;
|
|
|
|
void* mmapped_region_;
|
|
|
|
size_t length_;
|
|
|
|
|
|
|
|
public:
|
|
|
|
// base[0,length-1] contains the mmapped contents of the file.
|
2013-09-21 08:00:13 +02:00
|
|
|
PosixMmapReadableFile(const int fd, const std::string& fname,
|
|
|
|
void* base, size_t length,
|
2013-03-15 01:00:04 +01:00
|
|
|
const EnvOptions& options)
|
2013-09-21 08:00:13 +02:00
|
|
|
: fd_(fd), filename_(fname), mmapped_region_(base), length_(length) {
|
2013-11-17 08:44:39 +01:00
|
|
|
fd_ = fd_ + 0; // suppress the warning for used variables
|
2013-06-08 00:35:17 +02:00
|
|
|
assert(options.use_mmap_reads);
|
|
|
|
assert(options.use_os_buffer);
|
2013-03-15 01:00:04 +01:00
|
|
|
}
|
2013-10-29 04:34:02 +01:00
|
|
|
virtual ~PosixMmapReadableFile() {
|
2013-11-26 21:23:02 +01:00
|
|
|
int ret = munmap(mmapped_region_, length_);
|
|
|
|
if (ret != 0) {
|
|
|
|
fprintf(stdout, "failed to munmap %p length %zu \n",
|
|
|
|
mmapped_region_, length_);
|
|
|
|
}
|
2013-10-29 04:34:02 +01:00
|
|
|
}
|
2012-03-15 17:14:00 +01:00
|
|
|
|
|
|
|
virtual Status Read(uint64_t offset, size_t n, Slice* result,
|
|
|
|
char* scratch) const {
|
|
|
|
Status s;
|
|
|
|
if (offset + n > length_) {
|
|
|
|
*result = Slice();
|
|
|
|
s = IOError(filename_, EINVAL);
|
|
|
|
} else {
|
|
|
|
*result = Slice(reinterpret_cast<char*>(mmapped_region_) + offset, n);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
2013-09-21 08:00:13 +02:00
|
|
|
virtual Status InvalidateCache(size_t offset, size_t length) {
|
2013-11-17 08:44:39 +01:00
|
|
|
#ifndef OS_LINUX
|
|
|
|
return Status::OK();
|
|
|
|
#else
|
2013-09-21 08:00:13 +02:00
|
|
|
// free OS pages
|
2013-11-17 08:44:39 +01:00
|
|
|
int ret = Fadvise(fd_, offset, length, POSIX_FADV_DONTNEED);
|
2013-09-21 08:00:13 +02:00
|
|
|
if (ret == 0) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
return IOError(filename_, errno);
|
2013-11-17 08:44:39 +01:00
|
|
|
#endif
|
2013-09-21 08:00:13 +02:00
|
|
|
}
|
2012-03-15 17:14:00 +01:00
|
|
|
};
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
// We preallocate up to an extra megabyte and use memcpy to append new
|
|
|
|
// data to the file. This is safe since we either properly close the
|
|
|
|
// file before reading from it, or for log files, the reading code
|
|
|
|
// knows enough to skip zero suffixes.
|
|
|
|
class PosixMmapFile : public WritableFile {
|
|
|
|
private:
|
|
|
|
std::string filename_;
|
|
|
|
int fd_;
|
|
|
|
size_t page_size_;
|
|
|
|
size_t map_size_; // How much extra memory to map at a time
|
|
|
|
char* base_; // The mapped region
|
|
|
|
char* limit_; // Limit of the mapped region
|
|
|
|
char* dst_; // Where to write next (in range [base_,limit_])
|
|
|
|
char* last_sync_; // Where have we synced up to
|
|
|
|
uint64_t file_offset_; // Offset of base_ in file
|
|
|
|
|
|
|
|
// Have we done an munmap of unsynced data?
|
|
|
|
bool pending_sync_;
|
|
|
|
|
|
|
|
// Roundup x to a multiple of y
|
|
|
|
static size_t Roundup(size_t x, size_t y) {
|
|
|
|
return ((x + y - 1) / y) * y;
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t TruncateToPageBoundary(size_t s) {
|
|
|
|
s -= (s & (page_size_ - 1));
|
|
|
|
assert((s % page_size_) == 0);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2011-07-15 02:20:57 +02:00
|
|
|
bool UnmapCurrentRegion() {
|
|
|
|
bool result = true;
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2013-03-01 03:04:58 +01:00
|
|
|
if (base_ != nullptr) {
|
2011-03-18 23:37:00 +01:00
|
|
|
if (last_sync_ < limit_) {
|
|
|
|
// Defer syncing this data until next Sync() call, if any
|
|
|
|
pending_sync_ = true;
|
|
|
|
}
|
2011-07-15 02:20:57 +02:00
|
|
|
if (munmap(base_, limit_ - base_) != 0) {
|
|
|
|
result = false;
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
file_offset_ += limit_ - base_;
|
2013-03-01 03:04:58 +01:00
|
|
|
base_ = nullptr;
|
|
|
|
limit_ = nullptr;
|
|
|
|
last_sync_ = nullptr;
|
|
|
|
dst_ = nullptr;
|
2011-03-18 23:37:00 +01:00
|
|
|
|
|
|
|
// Increase the amount we map the next time, but capped at 1MB
|
|
|
|
if (map_size_ < (1<<20)) {
|
|
|
|
map_size_ *= 2;
|
|
|
|
}
|
|
|
|
}
|
2011-07-15 02:20:57 +02:00
|
|
|
return result;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2013-03-13 21:50:26 +01:00
|
|
|
Status MapNewRegion() {
|
2013-12-11 07:34:19 +01:00
|
|
|
#ifdef ROCKSDB_FALLOCATE_PRESENT
|
2013-03-01 03:04:58 +01:00
|
|
|
assert(base_ == nullptr);
|
2013-03-13 21:50:26 +01:00
|
|
|
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2013-03-13 21:50:26 +01:00
|
|
|
int alloc_status = posix_fallocate(fd_, file_offset_, map_size_);
|
|
|
|
if (alloc_status != 0) {
|
|
|
|
return Status::IOError("Error allocating space to file : " + filename_ +
|
|
|
|
"Error : " + strerror(alloc_status));
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-03-13 21:50:26 +01:00
|
|
|
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2013-03-01 03:04:58 +01:00
|
|
|
void* ptr = mmap(nullptr, map_size_, PROT_READ | PROT_WRITE, MAP_SHARED,
|
2011-03-18 23:37:00 +01:00
|
|
|
fd_, file_offset_);
|
|
|
|
if (ptr == MAP_FAILED) {
|
2013-03-13 21:50:26 +01:00
|
|
|
return Status::IOError("MMap failed on " + filename_);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-04-05 08:49:43 +02:00
|
|
|
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2013-04-05 08:49:43 +02:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
base_ = reinterpret_cast<char*>(ptr);
|
|
|
|
limit_ = base_ + map_size_;
|
|
|
|
dst_ = base_;
|
|
|
|
last_sync_ = base_;
|
2013-03-13 21:50:26 +01:00
|
|
|
return Status::OK();
|
2013-11-17 08:44:39 +01:00
|
|
|
#else
|
|
|
|
return Status::NotSupported("This platform doesn't support fallocate()");
|
|
|
|
#endif
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
public:
|
2013-03-15 01:00:04 +01:00
|
|
|
PosixMmapFile(const std::string& fname, int fd, size_t page_size,
|
|
|
|
const EnvOptions& options)
|
2011-03-18 23:37:00 +01:00
|
|
|
: filename_(fname),
|
|
|
|
fd_(fd),
|
|
|
|
page_size_(page_size),
|
|
|
|
map_size_(Roundup(65536, page_size)),
|
2013-03-01 03:04:58 +01:00
|
|
|
base_(nullptr),
|
|
|
|
limit_(nullptr),
|
|
|
|
dst_(nullptr),
|
|
|
|
last_sync_(nullptr),
|
2011-03-18 23:37:00 +01:00
|
|
|
file_offset_(0),
|
|
|
|
pending_sync_(false) {
|
|
|
|
assert((page_size & (page_size - 1)) == 0);
|
2013-06-08 00:35:17 +02:00
|
|
|
assert(options.use_mmap_writes);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
~PosixMmapFile() {
|
|
|
|
if (fd_ >= 0) {
|
|
|
|
PosixMmapFile::Close();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Append(const Slice& data) {
|
|
|
|
const char* src = data.data();
|
|
|
|
size_t left = data.size();
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds * REDUCE_ODDS);
|
2013-01-15 23:05:42 +01:00
|
|
|
PrepareWrite(GetFileSize(), left);
|
2011-03-18 23:37:00 +01:00
|
|
|
while (left > 0) {
|
|
|
|
assert(base_ <= dst_);
|
|
|
|
assert(dst_ <= limit_);
|
|
|
|
size_t avail = limit_ - dst_;
|
|
|
|
if (avail == 0) {
|
2013-03-13 21:50:26 +01:00
|
|
|
if (UnmapCurrentRegion()) {
|
|
|
|
Status s = MapNewRegion();
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2011-07-15 02:20:57 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
size_t n = (left <= avail) ? left : avail;
|
|
|
|
memcpy(dst_, src, n);
|
|
|
|
dst_ += n;
|
|
|
|
src += n;
|
|
|
|
left -= n;
|
|
|
|
}
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2011-03-18 23:37:00 +01:00
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Close() {
|
|
|
|
Status s;
|
|
|
|
size_t unused = limit_ - dst_;
|
2013-04-05 08:49:43 +02:00
|
|
|
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2013-04-05 08:49:43 +02:00
|
|
|
|
2011-07-15 02:20:57 +02:00
|
|
|
if (!UnmapCurrentRegion()) {
|
|
|
|
s = IOError(filename_, errno);
|
|
|
|
} else if (unused > 0) {
|
2011-03-18 23:37:00 +01:00
|
|
|
// Trim the extra space at the end of the file
|
|
|
|
if (ftruncate(fd_, file_offset_ - unused) < 0) {
|
2011-07-15 02:20:57 +02:00
|
|
|
s = IOError(filename_, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2013-04-05 08:49:43 +02:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
if (close(fd_) < 0) {
|
|
|
|
if (s.ok()) {
|
2011-07-15 02:20:57 +02:00
|
|
|
s = IOError(filename_, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
fd_ = -1;
|
2013-03-01 03:04:58 +01:00
|
|
|
base_ = nullptr;
|
|
|
|
limit_ = nullptr;
|
2011-03-18 23:37:00 +01:00
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Flush() {
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2011-03-18 23:37:00 +01:00
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Sync() {
|
|
|
|
Status s;
|
|
|
|
|
|
|
|
if (pending_sync_) {
|
|
|
|
// Some unmapped data was not synced
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2011-03-18 23:37:00 +01:00
|
|
|
pending_sync_ = false;
|
|
|
|
if (fdatasync(fd_) < 0) {
|
2011-07-15 02:20:57 +02:00
|
|
|
s = IOError(filename_, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds * REDUCE_ODDS);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
if (dst_ > last_sync_) {
|
|
|
|
// Find the beginnings of the pages that contain the first and last
|
|
|
|
// bytes to be synced.
|
|
|
|
size_t p1 = TruncateToPageBoundary(last_sync_ - base_);
|
|
|
|
size_t p2 = TruncateToPageBoundary(dst_ - base_ - 1);
|
|
|
|
last_sync_ = dst_;
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2011-03-18 23:37:00 +01:00
|
|
|
if (msync(base_ + p1, p2 - p1 + page_size_, MS_SYNC) < 0) {
|
2011-07-15 02:20:57 +02:00
|
|
|
s = IOError(filename_, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
2012-08-27 21:10:26 +02:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Flush data as well as metadata to stable storage.
|
|
|
|
*/
|
|
|
|
virtual Status Fsync() {
|
|
|
|
if (pending_sync_) {
|
|
|
|
// Some unmapped data was not synced
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2012-08-27 21:10:26 +02:00
|
|
|
pending_sync_ = false;
|
|
|
|
if (fsync(fd_) < 0) {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2012-08-27 21:10:26 +02:00
|
|
|
}
|
|
|
|
// This invocation to Sync will not issue the call to
|
|
|
|
// fdatasync because pending_sync_ has already been cleared.
|
|
|
|
return Sync();
|
|
|
|
}
|
2012-09-24 23:01:01 +02:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Get the size of valid data in the file. This will not match the
|
|
|
|
* size that is returned from the filesystem because we use mmap
|
|
|
|
* to extend file by map_size every time.
|
|
|
|
*/
|
|
|
|
virtual uint64_t GetFileSize() {
|
|
|
|
size_t used = dst_ - base_;
|
|
|
|
return file_offset_ + used;
|
|
|
|
}
|
2013-01-15 23:05:42 +01:00
|
|
|
|
2013-09-21 08:00:13 +02:00
|
|
|
virtual Status InvalidateCache(size_t offset, size_t length) {
|
2013-11-17 08:44:39 +01:00
|
|
|
#ifndef OS_LINUX
|
|
|
|
return Status::OK();
|
|
|
|
#else
|
2013-09-21 08:00:13 +02:00
|
|
|
// free OS pages
|
2013-11-17 08:44:39 +01:00
|
|
|
int ret = Fadvise(fd_, offset, length, POSIX_FADV_DONTNEED);
|
2013-09-21 08:00:13 +02:00
|
|
|
if (ret == 0) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
return IOError(filename_, errno);
|
2013-11-17 08:44:39 +01:00
|
|
|
#endif
|
2013-09-21 08:00:13 +02:00
|
|
|
}
|
|
|
|
|
2013-12-11 07:34:19 +01:00
|
|
|
#ifdef ROCKSDB_FALLOCATE_PRESENT
|
2013-01-15 23:05:42 +01:00
|
|
|
virtual Status Allocate(off_t offset, off_t len) {
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2013-01-15 23:05:42 +01:00
|
|
|
if (!fallocate(fd_, FALLOC_FL_KEEP_SIZE, offset, len)) {
|
|
|
|
return Status::OK();
|
|
|
|
} else {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
|
|
|
}
|
2013-01-28 20:18:50 +01:00
|
|
|
#endif
|
2011-03-18 23:37:00 +01:00
|
|
|
};
|
|
|
|
|
2012-10-02 00:41:44 +02:00
|
|
|
// Use posix write to write data to a file.
|
|
|
|
class PosixWritableFile : public WritableFile {
|
|
|
|
private:
|
|
|
|
const std::string filename_;
|
|
|
|
int fd_;
|
|
|
|
size_t cursize_; // current size of cached data in buf_
|
|
|
|
size_t capacity_; // max size of buf_
|
2013-04-09 04:10:24 +02:00
|
|
|
unique_ptr<char[]> buf_; // a buffer to cache writes
|
2012-10-02 00:41:44 +02:00
|
|
|
uint64_t filesize_;
|
|
|
|
bool pending_sync_;
|
|
|
|
bool pending_fsync_;
|
[RocksDB] Sync file to disk incrementally
Summary:
During compaction, we sync the output files after they are fully written out. This causes unnecessary blocking of the compaction thread and burstiness of the write traffic.
This diff simply asks the OS to sync data incrementally as they are written, on the background. The hope is that, at the final sync, most of the data are already on disk and we would block less on the sync call. Thus, each compaction runs faster and we could use fewer number of compaction threads to saturate IO.
In addition, the write traffic will be smoothed out, hopefully reducing the IO P99 latency too.
Some quick tests show 10~20% improvement in per thread compaction throughput. Combined with posix advice on compaction read, just 5 threads are enough to almost saturate the udb flash bandwidth for 800 bytes write only benchmark.
What's more promising is that, with saturated IO, iostat shows average wait time is actually smoother and much smaller.
For the write only test 800bytes test:
Before the change: await occillate between 10ms and 3ms
After the change: await ranges 1-3ms
Will test against read-modify-write workload too, see if high read latency P99 could be resolved.
Will introduce a parameter to control the sync interval in a follow up diff after cleaning up EnvOptions.
Test Plan: make check; db_bench; db_stress
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11115
2013-06-05 00:51:50 +02:00
|
|
|
uint64_t last_sync_size_;
|
2013-06-14 07:49:46 +02:00
|
|
|
uint64_t bytes_per_sync_;
|
2012-10-02 00:41:44 +02:00
|
|
|
|
|
|
|
public:
|
2013-03-15 01:00:04 +01:00
|
|
|
PosixWritableFile(const std::string& fname, int fd, size_t capacity,
|
|
|
|
const EnvOptions& options) :
|
2012-10-02 00:41:44 +02:00
|
|
|
filename_(fname),
|
|
|
|
fd_(fd),
|
|
|
|
cursize_(0),
|
|
|
|
capacity_(capacity),
|
|
|
|
buf_(new char[capacity]),
|
|
|
|
filesize_(0),
|
|
|
|
pending_sync_(false),
|
[RocksDB] Sync file to disk incrementally
Summary:
During compaction, we sync the output files after they are fully written out. This causes unnecessary blocking of the compaction thread and burstiness of the write traffic.
This diff simply asks the OS to sync data incrementally as they are written, on the background. The hope is that, at the final sync, most of the data are already on disk and we would block less on the sync call. Thus, each compaction runs faster and we could use fewer number of compaction threads to saturate IO.
In addition, the write traffic will be smoothed out, hopefully reducing the IO P99 latency too.
Some quick tests show 10~20% improvement in per thread compaction throughput. Combined with posix advice on compaction read, just 5 threads are enough to almost saturate the udb flash bandwidth for 800 bytes write only benchmark.
What's more promising is that, with saturated IO, iostat shows average wait time is actually smoother and much smaller.
For the write only test 800bytes test:
Before the change: await occillate between 10ms and 3ms
After the change: await ranges 1-3ms
Will test against read-modify-write workload too, see if high read latency P99 could be resolved.
Will introduce a parameter to control the sync interval in a follow up diff after cleaning up EnvOptions.
Test Plan: make check; db_bench; db_stress
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11115
2013-06-05 00:51:50 +02:00
|
|
|
pending_fsync_(false),
|
2013-06-14 07:49:46 +02:00
|
|
|
last_sync_size_(0),
|
|
|
|
bytes_per_sync_(options.bytes_per_sync) {
|
2013-06-08 00:35:17 +02:00
|
|
|
assert(!options.use_mmap_writes);
|
2012-10-02 00:41:44 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
~PosixWritableFile() {
|
|
|
|
if (fd_ >= 0) {
|
|
|
|
PosixWritableFile::Close();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Append(const Slice& data) {
|
2013-10-10 09:03:08 +02:00
|
|
|
const char* src = data.data();
|
2012-10-02 00:41:44 +02:00
|
|
|
size_t left = data.size();
|
|
|
|
Status s;
|
|
|
|
pending_sync_ = true;
|
|
|
|
pending_fsync_ = true;
|
|
|
|
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds * REDUCE_ODDS2);
|
2013-04-05 08:49:43 +02:00
|
|
|
|
2013-01-15 23:05:42 +01:00
|
|
|
PrepareWrite(GetFileSize(), left);
|
2012-10-02 00:41:44 +02:00
|
|
|
// if there is no space in the cache, then flush
|
|
|
|
if (cursize_ + left > capacity_) {
|
|
|
|
s = Flush();
|
|
|
|
if (!s.ok()) {
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
// Increase the buffer size, but capped at 1MB
|
|
|
|
if (capacity_ < (1<<20)) {
|
|
|
|
capacity_ *= 2;
|
2013-04-09 04:10:24 +02:00
|
|
|
buf_.reset(new char[capacity_]);
|
2012-10-02 00:41:44 +02:00
|
|
|
}
|
|
|
|
assert(cursize_ == 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
// if the write fits into the cache, then write to cache
|
|
|
|
// otherwise do a write() syscall to write to OS buffers.
|
|
|
|
if (cursize_ + left <= capacity_) {
|
2013-04-09 04:10:24 +02:00
|
|
|
memcpy(buf_.get()+cursize_, src, left);
|
2012-10-02 00:41:44 +02:00
|
|
|
cursize_ += left;
|
|
|
|
} else {
|
|
|
|
while (left != 0) {
|
2013-01-15 03:37:01 +01:00
|
|
|
ssize_t done = write(fd_, src, left);
|
2012-10-02 00:41:44 +02:00
|
|
|
if (done < 0) {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2013-04-05 08:49:43 +02:00
|
|
|
|
2012-10-02 00:41:44 +02:00
|
|
|
left -= done;
|
|
|
|
src += done;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
filesize_ += data.size();
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Close() {
|
|
|
|
Status s;
|
|
|
|
s = Flush(); // flush cache to OS
|
|
|
|
if (!s.ok()) {
|
2014-03-07 00:59:27 +01:00
|
|
|
return s;
|
2012-10-02 00:41:44 +02:00
|
|
|
}
|
2013-04-05 08:49:43 +02:00
|
|
|
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2013-04-05 08:49:43 +02:00
|
|
|
|
2014-03-07 00:59:27 +01:00
|
|
|
size_t block_size;
|
|
|
|
size_t last_allocated_block;
|
|
|
|
GetPreallocationStatus(&block_size, &last_allocated_block);
|
|
|
|
if (last_allocated_block > 0) {
|
|
|
|
// trim the extra space preallocated at the end of the file
|
2014-03-07 18:14:34 +01:00
|
|
|
int dummy __attribute__((unused));
|
|
|
|
dummy = ftruncate(fd_, filesize_); // ignore errors
|
2014-03-07 00:59:27 +01:00
|
|
|
}
|
|
|
|
|
2012-10-02 00:41:44 +02:00
|
|
|
if (close(fd_) < 0) {
|
|
|
|
if (s.ok()) {
|
|
|
|
s = IOError(filename_, errno);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
fd_ = -1;
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
// write out the cached data to the OS cache
|
|
|
|
virtual Status Flush() {
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds * REDUCE_ODDS2);
|
2012-10-02 00:41:44 +02:00
|
|
|
size_t left = cursize_;
|
2013-04-09 04:10:24 +02:00
|
|
|
char* src = buf_.get();
|
2012-10-02 00:41:44 +02:00
|
|
|
while (left != 0) {
|
2013-01-15 03:37:01 +01:00
|
|
|
ssize_t done = write(fd_, src, left);
|
2012-10-02 00:41:44 +02:00
|
|
|
if (done < 0) {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds * REDUCE_ODDS2);
|
2012-10-02 00:41:44 +02:00
|
|
|
left -= done;
|
|
|
|
src += done;
|
|
|
|
}
|
|
|
|
cursize_ = 0;
|
[RocksDB] Sync file to disk incrementally
Summary:
During compaction, we sync the output files after they are fully written out. This causes unnecessary blocking of the compaction thread and burstiness of the write traffic.
This diff simply asks the OS to sync data incrementally as they are written, on the background. The hope is that, at the final sync, most of the data are already on disk and we would block less on the sync call. Thus, each compaction runs faster and we could use fewer number of compaction threads to saturate IO.
In addition, the write traffic will be smoothed out, hopefully reducing the IO P99 latency too.
Some quick tests show 10~20% improvement in per thread compaction throughput. Combined with posix advice on compaction read, just 5 threads are enough to almost saturate the udb flash bandwidth for 800 bytes write only benchmark.
What's more promising is that, with saturated IO, iostat shows average wait time is actually smoother and much smaller.
For the write only test 800bytes test:
Before the change: await occillate between 10ms and 3ms
After the change: await ranges 1-3ms
Will test against read-modify-write workload too, see if high read latency P99 could be resolved.
Will introduce a parameter to control the sync interval in a follow up diff after cleaning up EnvOptions.
Test Plan: make check; db_bench; db_stress
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11115
2013-06-05 00:51:50 +02:00
|
|
|
|
2013-06-14 07:49:46 +02:00
|
|
|
// sync OS cache to disk for every bytes_per_sync_
|
|
|
|
// TODO: give log file and sst file different options (log
|
|
|
|
// files could be potentially cached in OS for their whole
|
|
|
|
// life time, thus we might not want to flush at all).
|
|
|
|
if (bytes_per_sync_ &&
|
|
|
|
filesize_ - last_sync_size_ >= bytes_per_sync_) {
|
[RocksDB] Sync file to disk incrementally
Summary:
During compaction, we sync the output files after they are fully written out. This causes unnecessary blocking of the compaction thread and burstiness of the write traffic.
This diff simply asks the OS to sync data incrementally as they are written, on the background. The hope is that, at the final sync, most of the data are already on disk and we would block less on the sync call. Thus, each compaction runs faster and we could use fewer number of compaction threads to saturate IO.
In addition, the write traffic will be smoothed out, hopefully reducing the IO P99 latency too.
Some quick tests show 10~20% improvement in per thread compaction throughput. Combined with posix advice on compaction read, just 5 threads are enough to almost saturate the udb flash bandwidth for 800 bytes write only benchmark.
What's more promising is that, with saturated IO, iostat shows average wait time is actually smoother and much smaller.
For the write only test 800bytes test:
Before the change: await occillate between 10ms and 3ms
After the change: await ranges 1-3ms
Will test against read-modify-write workload too, see if high read latency P99 could be resolved.
Will introduce a parameter to control the sync interval in a follow up diff after cleaning up EnvOptions.
Test Plan: make check; db_bench; db_stress
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11115
2013-06-05 00:51:50 +02:00
|
|
|
RangeSync(last_sync_size_, filesize_ - last_sync_size_);
|
|
|
|
last_sync_size_ = filesize_;
|
|
|
|
}
|
|
|
|
|
2012-10-02 00:41:44 +02:00
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Sync() {
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2012-10-02 00:41:44 +02:00
|
|
|
if (pending_sync_ && fdatasync(fd_) < 0) {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2012-10-02 00:41:44 +02:00
|
|
|
pending_sync_ = false;
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Fsync() {
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2012-10-02 00:41:44 +02:00
|
|
|
if (pending_fsync_ && fsync(fd_) < 0) {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2012-10-02 00:41:44 +02:00
|
|
|
pending_fsync_ = false;
|
|
|
|
pending_sync_ = false;
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual uint64_t GetFileSize() {
|
|
|
|
return filesize_;
|
|
|
|
}
|
2013-01-15 23:05:42 +01:00
|
|
|
|
2013-09-21 08:00:13 +02:00
|
|
|
virtual Status InvalidateCache(size_t offset, size_t length) {
|
2013-11-17 08:44:39 +01:00
|
|
|
#ifndef OS_LINUX
|
|
|
|
return Status::OK();
|
|
|
|
#else
|
2013-09-21 08:00:13 +02:00
|
|
|
// free OS pages
|
2013-11-17 08:44:39 +01:00
|
|
|
int ret = Fadvise(fd_, offset, length, POSIX_FADV_DONTNEED);
|
2013-09-21 08:00:13 +02:00
|
|
|
if (ret == 0) {
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
return IOError(filename_, errno);
|
2013-11-17 08:44:39 +01:00
|
|
|
#endif
|
2013-09-21 08:00:13 +02:00
|
|
|
}
|
|
|
|
|
2013-12-11 07:34:19 +01:00
|
|
|
#ifdef ROCKSDB_FALLOCATE_PRESENT
|
2013-01-15 23:05:42 +01:00
|
|
|
virtual Status Allocate(off_t offset, off_t len) {
|
2013-10-05 07:32:05 +02:00
|
|
|
TEST_KILL_RANDOM(rocksdb_kill_odds);
|
2013-01-15 23:05:42 +01:00
|
|
|
if (!fallocate(fd_, FALLOC_FL_KEEP_SIZE, offset, len)) {
|
|
|
|
return Status::OK();
|
|
|
|
} else {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
|
|
|
}
|
[RocksDB] Sync file to disk incrementally
Summary:
During compaction, we sync the output files after they are fully written out. This causes unnecessary blocking of the compaction thread and burstiness of the write traffic.
This diff simply asks the OS to sync data incrementally as they are written, on the background. The hope is that, at the final sync, most of the data are already on disk and we would block less on the sync call. Thus, each compaction runs faster and we could use fewer number of compaction threads to saturate IO.
In addition, the write traffic will be smoothed out, hopefully reducing the IO P99 latency too.
Some quick tests show 10~20% improvement in per thread compaction throughput. Combined with posix advice on compaction read, just 5 threads are enough to almost saturate the udb flash bandwidth for 800 bytes write only benchmark.
What's more promising is that, with saturated IO, iostat shows average wait time is actually smoother and much smaller.
For the write only test 800bytes test:
Before the change: await occillate between 10ms and 3ms
After the change: await ranges 1-3ms
Will test against read-modify-write workload too, see if high read latency P99 could be resolved.
Will introduce a parameter to control the sync interval in a follow up diff after cleaning up EnvOptions.
Test Plan: make check; db_bench; db_stress
Reviewers: dhruba
CC: leveldb
Differential Revision: https://reviews.facebook.net/D11115
2013-06-05 00:51:50 +02:00
|
|
|
|
|
|
|
virtual Status RangeSync(off64_t offset, off64_t nbytes) {
|
|
|
|
if (sync_file_range(fd_, offset, nbytes, SYNC_FILE_RANGE_WRITE) == 0) {
|
|
|
|
return Status::OK();
|
|
|
|
} else {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
|
|
|
}
|
2013-09-02 08:23:40 +02:00
|
|
|
virtual size_t GetUniqueId(char* id, size_t max_size) const {
|
|
|
|
return GetUniqueIdFromFile(fd_, id, max_size);
|
|
|
|
}
|
2013-01-28 20:18:50 +01:00
|
|
|
#endif
|
2012-10-02 00:41:44 +02:00
|
|
|
};
|
|
|
|
|
2013-10-10 09:03:08 +02:00
|
|
|
class PosixRandomRWFile : public RandomRWFile {
|
|
|
|
private:
|
|
|
|
const std::string filename_;
|
|
|
|
int fd_;
|
|
|
|
bool pending_sync_;
|
|
|
|
bool pending_fsync_;
|
|
|
|
|
|
|
|
public:
|
|
|
|
PosixRandomRWFile(const std::string& fname, int fd,
|
|
|
|
const EnvOptions& options) :
|
|
|
|
filename_(fname),
|
|
|
|
fd_(fd),
|
|
|
|
pending_sync_(false),
|
|
|
|
pending_fsync_(false) {
|
|
|
|
assert(!options.use_mmap_writes && !options.use_mmap_reads);
|
|
|
|
}
|
|
|
|
|
|
|
|
~PosixRandomRWFile() {
|
|
|
|
if (fd_ >= 0) {
|
|
|
|
Close();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Write(uint64_t offset, const Slice& data) {
|
|
|
|
const char* src = data.data();
|
|
|
|
size_t left = data.size();
|
|
|
|
Status s;
|
|
|
|
pending_sync_ = true;
|
|
|
|
pending_fsync_ = true;
|
|
|
|
|
|
|
|
while (left != 0) {
|
|
|
|
ssize_t done = pwrite(fd_, src, left, offset);
|
|
|
|
if (done < 0) {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
|
|
|
|
|
|
|
left -= done;
|
|
|
|
src += done;
|
|
|
|
offset += done;
|
|
|
|
}
|
|
|
|
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Read(uint64_t offset, size_t n, Slice* result,
|
|
|
|
char* scratch) const {
|
|
|
|
Status s;
|
|
|
|
ssize_t r = pread(fd_, scratch, n, static_cast<off_t>(offset));
|
|
|
|
*result = Slice(scratch, (r < 0) ? 0 : r);
|
|
|
|
if (r < 0) {
|
|
|
|
s = IOError(filename_, errno);
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Close() {
|
|
|
|
Status s = Status::OK();
|
|
|
|
if (fd_ >= 0 && close(fd_) < 0) {
|
|
|
|
s = IOError(filename_, errno);
|
|
|
|
}
|
|
|
|
fd_ = -1;
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Sync() {
|
|
|
|
if (pending_sync_ && fdatasync(fd_) < 0) {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
|
|
|
pending_sync_ = false;
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Fsync() {
|
|
|
|
if (pending_fsync_ && fsync(fd_) < 0) {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
|
|
|
pending_fsync_ = false;
|
|
|
|
pending_sync_ = false;
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2013-12-11 07:34:19 +01:00
|
|
|
#ifdef ROCKSDB_FALLOCATE_PRESENT
|
2013-10-10 09:03:08 +02:00
|
|
|
virtual Status Allocate(off_t offset, off_t len) {
|
|
|
|
if (!fallocate(fd_, FALLOC_FL_KEEP_SIZE, offset, len)) {
|
|
|
|
return Status::OK();
|
|
|
|
} else {
|
|
|
|
return IOError(filename_, errno);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
};
|
|
|
|
|
2014-01-27 20:02:21 +01:00
|
|
|
class PosixDirectory : public Directory {
|
|
|
|
public:
|
|
|
|
explicit PosixDirectory(int fd) : fd_(fd) {}
|
|
|
|
~PosixDirectory() {
|
|
|
|
close(fd_);
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status Fsync() {
|
|
|
|
if (fsync(fd_) == -1) {
|
|
|
|
return IOError("directory", errno);
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
private:
|
|
|
|
int fd_;
|
|
|
|
};
|
|
|
|
|
2012-08-18 09:26:50 +02:00
|
|
|
static int LockOrUnlock(const std::string& fname, int fd, bool lock) {
|
|
|
|
mutex_lockedFiles.Lock();
|
|
|
|
if (lock) {
|
|
|
|
// If it already exists in the lockedFiles set, then it is already locked,
|
|
|
|
// and fail this lock attempt. Otherwise, insert it into lockedFiles.
|
|
|
|
// This check is needed because fcntl() does not detect lock conflict
|
|
|
|
// if the fcntl is issued by the same thread that earlier acquired
|
|
|
|
// this lock.
|
|
|
|
if (lockedFiles.insert(fname).second == false) {
|
|
|
|
mutex_lockedFiles.Unlock();
|
|
|
|
errno = ENOLCK;
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
// If we are unlocking, then verify that we had locked it earlier,
|
|
|
|
// it should already exist in lockedFiles. Remove it from lockedFiles.
|
|
|
|
if (lockedFiles.erase(fname) != 1) {
|
|
|
|
mutex_lockedFiles.Unlock();
|
|
|
|
errno = ENOLCK;
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
errno = 0;
|
|
|
|
struct flock f;
|
|
|
|
memset(&f, 0, sizeof(f));
|
|
|
|
f.l_type = (lock ? F_WRLCK : F_UNLCK);
|
|
|
|
f.l_whence = SEEK_SET;
|
|
|
|
f.l_start = 0;
|
|
|
|
f.l_len = 0; // Lock/unlock entire file
|
2012-08-18 09:26:50 +02:00
|
|
|
int value = fcntl(fd, F_SETLK, &f);
|
|
|
|
if (value == -1 && lock) {
|
|
|
|
// if there is an error in locking, then remove the pathname from lockedfiles
|
|
|
|
lockedFiles.erase(fname);
|
|
|
|
}
|
|
|
|
mutex_lockedFiles.Unlock();
|
|
|
|
return value;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
class PosixFileLock : public FileLock {
|
|
|
|
public:
|
|
|
|
int fd_;
|
2012-08-18 09:26:50 +02:00
|
|
|
std::string filename;
|
2011-03-18 23:37:00 +01:00
|
|
|
};
|
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
|
|
|
|
namespace {
|
|
|
|
void PthreadCall(const char* label, int result) {
|
|
|
|
if (result != 0) {
|
|
|
|
fprintf(stderr, "pthread %s: %s\n", label, strerror(result));
|
|
|
|
exit(1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
class PosixEnv : public Env {
|
|
|
|
public:
|
|
|
|
PosixEnv();
|
2013-03-19 22:39:28 +01:00
|
|
|
|
|
|
|
virtual ~PosixEnv(){
|
2013-09-12 09:53:30 +02:00
|
|
|
for (const auto tid : threads_to_join_) {
|
|
|
|
pthread_join(tid, nullptr);
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2013-04-10 04:42:07 +02:00
|
|
|
void SetFD_CLOEXEC(int fd, const EnvOptions* options) {
|
2013-06-08 00:35:17 +02:00
|
|
|
if ((options == nullptr || options->set_fd_cloexec) && fd > 0) {
|
2013-04-10 04:42:07 +02:00
|
|
|
fcntl(fd, F_SETFD, fcntl(fd, F_GETFD) | FD_CLOEXEC);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
virtual Status NewSequentialFile(const std::string& fname,
|
2013-03-15 01:00:04 +01:00
|
|
|
unique_ptr<SequentialFile>* result,
|
|
|
|
const EnvOptions& options) {
|
2013-01-20 11:07:13 +01:00
|
|
|
result->reset();
|
2011-03-18 23:37:00 +01:00
|
|
|
FILE* f = fopen(fname.c_str(), "r");
|
2013-03-01 03:04:58 +01:00
|
|
|
if (f == nullptr) {
|
|
|
|
*result = nullptr;
|
2011-07-15 02:20:57 +02:00
|
|
|
return IOError(fname, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
} else {
|
2013-04-10 04:42:07 +02:00
|
|
|
int fd = fileno(f);
|
|
|
|
SetFD_CLOEXEC(fd, &options);
|
2013-03-15 01:00:04 +01:00
|
|
|
result->reset(new PosixSequentialFile(fname, f, options));
|
2011-03-18 23:37:00 +01:00
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status NewRandomAccessFile(const std::string& fname,
|
2013-03-15 01:00:04 +01:00
|
|
|
unique_ptr<RandomAccessFile>* result,
|
|
|
|
const EnvOptions& options) {
|
2013-01-20 11:07:13 +01:00
|
|
|
result->reset();
|
2012-03-15 17:14:00 +01:00
|
|
|
Status s;
|
2011-03-18 23:37:00 +01:00
|
|
|
int fd = open(fname.c_str(), O_RDONLY);
|
2013-04-10 04:42:07 +02:00
|
|
|
SetFD_CLOEXEC(fd, &options);
|
2011-03-18 23:37:00 +01:00
|
|
|
if (fd < 0) {
|
2012-03-15 17:14:00 +01:00
|
|
|
s = IOError(fname, errno);
|
2013-06-08 00:35:17 +02:00
|
|
|
} else if (options.use_mmap_reads && sizeof(void*) >= 8) {
|
2012-09-15 05:57:15 +02:00
|
|
|
// Use of mmap for random reads has been removed because it
|
|
|
|
// kills performance when storage is fast.
|
|
|
|
// Use mmap when virtual address-space is plentiful.
|
|
|
|
uint64_t size;
|
|
|
|
s = GetFileSize(fname, &size);
|
|
|
|
if (s.ok()) {
|
2013-03-01 03:04:58 +01:00
|
|
|
void* base = mmap(nullptr, size, PROT_READ, MAP_SHARED, fd, 0);
|
2012-09-15 05:57:15 +02:00
|
|
|
if (base != MAP_FAILED) {
|
2013-09-21 08:00:13 +02:00
|
|
|
result->reset(new PosixMmapReadableFile(fd, fname, base,
|
|
|
|
size, options));
|
2012-09-15 05:57:15 +02:00
|
|
|
} else {
|
|
|
|
s = IOError(fname, errno);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
close(fd);
|
2012-03-15 17:14:00 +01:00
|
|
|
} else {
|
2013-03-15 01:00:04 +01:00
|
|
|
result->reset(new PosixRandomAccessFile(fname, fd, options));
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2012-03-15 17:14:00 +01:00
|
|
|
return s;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status NewWritableFile(const std::string& fname,
|
2013-03-15 01:00:04 +01:00
|
|
|
unique_ptr<WritableFile>* result,
|
|
|
|
const EnvOptions& options) {
|
2013-01-20 11:07:13 +01:00
|
|
|
result->reset();
|
2011-03-18 23:37:00 +01:00
|
|
|
Status s;
|
|
|
|
const int fd = open(fname.c_str(), O_CREAT | O_RDWR | O_TRUNC, 0644);
|
|
|
|
if (fd < 0) {
|
2011-07-15 02:20:57 +02:00
|
|
|
s = IOError(fname, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
} else {
|
2013-04-10 04:42:07 +02:00
|
|
|
SetFD_CLOEXEC(fd, &options);
|
2013-06-08 00:35:17 +02:00
|
|
|
if (options.use_mmap_writes) {
|
2013-03-15 01:00:04 +01:00
|
|
|
if (!checkedDiskForMmap_) {
|
|
|
|
// this will be executed once in the program's lifetime.
|
2013-03-13 21:50:26 +01:00
|
|
|
// do not use mmapWrite on non ext-3/xfs/tmpfs systems.
|
2013-03-15 01:00:04 +01:00
|
|
|
if (!SupportsFastAllocate(fname)) {
|
|
|
|
forceMmapOff = true;
|
|
|
|
}
|
|
|
|
checkedDiskForMmap_ = true;
|
2013-03-13 21:50:26 +01:00
|
|
|
}
|
|
|
|
}
|
2013-06-08 00:35:17 +02:00
|
|
|
if (options.use_mmap_writes && !forceMmapOff) {
|
2013-03-15 01:00:04 +01:00
|
|
|
result->reset(new PosixMmapFile(fname, fd, page_size_, options));
|
2012-10-02 00:41:44 +02:00
|
|
|
} else {
|
2013-11-17 08:44:39 +01:00
|
|
|
// disable mmap writes
|
|
|
|
EnvOptions no_mmap_writes_options = options;
|
|
|
|
no_mmap_writes_options.use_mmap_writes = false;
|
|
|
|
|
|
|
|
result->reset(
|
|
|
|
new PosixWritableFile(fname, fd, 65536, no_mmap_writes_options)
|
|
|
|
);
|
2012-10-02 00:41:44 +02:00
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2013-10-10 09:03:08 +02:00
|
|
|
virtual Status NewRandomRWFile(const std::string& fname,
|
|
|
|
unique_ptr<RandomRWFile>* result,
|
|
|
|
const EnvOptions& options) {
|
|
|
|
result->reset();
|
2014-02-07 06:59:44 +01:00
|
|
|
// no support for mmap yet
|
|
|
|
if (options.use_mmap_writes || options.use_mmap_reads) {
|
|
|
|
return Status::NotSupported("No support for mmap read/write yet");
|
|
|
|
}
|
2013-10-10 09:03:08 +02:00
|
|
|
Status s;
|
|
|
|
const int fd = open(fname.c_str(), O_CREAT | O_RDWR, 0644);
|
|
|
|
if (fd < 0) {
|
|
|
|
s = IOError(fname, errno);
|
|
|
|
} else {
|
|
|
|
SetFD_CLOEXEC(fd, &options);
|
|
|
|
result->reset(new PosixRandomRWFile(fname, fd, options));
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2014-01-27 20:02:21 +01:00
|
|
|
virtual Status NewDirectory(const std::string& name,
|
|
|
|
unique_ptr<Directory>* result) {
|
|
|
|
result->reset();
|
|
|
|
const int fd = open(name.c_str(), 0);
|
|
|
|
if (fd < 0) {
|
|
|
|
return IOError(name, errno);
|
|
|
|
} else {
|
|
|
|
result->reset(new PosixDirectory(fd));
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
virtual bool FileExists(const std::string& fname) {
|
|
|
|
return access(fname.c_str(), F_OK) == 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status GetChildren(const std::string& dir,
|
|
|
|
std::vector<std::string>* result) {
|
|
|
|
result->clear();
|
|
|
|
DIR* d = opendir(dir.c_str());
|
2013-03-01 03:04:58 +01:00
|
|
|
if (d == nullptr) {
|
2011-07-15 02:20:57 +02:00
|
|
|
return IOError(dir, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
struct dirent* entry;
|
2013-03-01 03:04:58 +01:00
|
|
|
while ((entry = readdir(d)) != nullptr) {
|
2011-03-18 23:37:00 +01:00
|
|
|
result->push_back(entry->d_name);
|
|
|
|
}
|
|
|
|
closedir(d);
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status DeleteFile(const std::string& fname) {
|
|
|
|
Status result;
|
|
|
|
if (unlink(fname.c_str()) != 0) {
|
2011-07-15 02:20:57 +02:00
|
|
|
result = IOError(fname, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
return result;
|
|
|
|
};
|
|
|
|
|
|
|
|
virtual Status CreateDir(const std::string& name) {
|
|
|
|
Status result;
|
|
|
|
if (mkdir(name.c_str(), 0755) != 0) {
|
2011-07-15 02:20:57 +02:00
|
|
|
result = IOError(name, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
return result;
|
|
|
|
};
|
|
|
|
|
2012-11-26 22:56:45 +01:00
|
|
|
virtual Status CreateDirIfMissing(const std::string& name) {
|
|
|
|
Status result;
|
|
|
|
if (mkdir(name.c_str(), 0755) != 0) {
|
|
|
|
if (errno != EEXIST) {
|
|
|
|
result = IOError(name, errno);
|
2013-01-07 19:11:18 +01:00
|
|
|
} else if (!DirExists(name)) { // Check that name is actually a
|
|
|
|
// directory.
|
|
|
|
// Message is taken from mkdir
|
|
|
|
result = Status::IOError("`"+name+"' exists but is not a directory");
|
2012-11-26 22:56:45 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
};
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
virtual Status DeleteDir(const std::string& name) {
|
|
|
|
Status result;
|
|
|
|
if (rmdir(name.c_str()) != 0) {
|
2011-07-15 02:20:57 +02:00
|
|
|
result = IOError(name, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
return result;
|
|
|
|
};
|
|
|
|
|
|
|
|
virtual Status GetFileSize(const std::string& fname, uint64_t* size) {
|
|
|
|
Status s;
|
|
|
|
struct stat sbuf;
|
|
|
|
if (stat(fname.c_str(), &sbuf) != 0) {
|
|
|
|
*size = 0;
|
2011-07-15 02:20:57 +02:00
|
|
|
s = IOError(fname, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
} else {
|
|
|
|
*size = sbuf.st_size;
|
|
|
|
}
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
2012-11-26 22:56:45 +01:00
|
|
|
virtual Status GetFileModificationTime(const std::string& fname,
|
|
|
|
uint64_t* file_mtime) {
|
|
|
|
struct stat s;
|
|
|
|
if (stat(fname.c_str(), &s) !=0) {
|
|
|
|
return IOError(fname, errno);
|
|
|
|
}
|
|
|
|
*file_mtime = static_cast<uint64_t>(s.st_mtime);
|
|
|
|
return Status::OK();
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
virtual Status RenameFile(const std::string& src, const std::string& target) {
|
|
|
|
Status result;
|
|
|
|
if (rename(src.c_str(), target.c_str()) != 0) {
|
2011-07-15 02:20:57 +02:00
|
|
|
result = IOError(src, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status LockFile(const std::string& fname, FileLock** lock) {
|
2013-03-01 03:04:58 +01:00
|
|
|
*lock = nullptr;
|
2011-03-18 23:37:00 +01:00
|
|
|
Status result;
|
|
|
|
int fd = open(fname.c_str(), O_RDWR | O_CREAT, 0644);
|
|
|
|
if (fd < 0) {
|
2011-07-15 02:20:57 +02:00
|
|
|
result = IOError(fname, errno);
|
2012-08-18 09:26:50 +02:00
|
|
|
} else if (LockOrUnlock(fname, fd, true) == -1) {
|
2011-07-15 02:20:57 +02:00
|
|
|
result = IOError("lock " + fname, errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
close(fd);
|
|
|
|
} else {
|
2013-04-10 04:42:07 +02:00
|
|
|
SetFD_CLOEXEC(fd, nullptr);
|
2011-03-18 23:37:00 +01:00
|
|
|
PosixFileLock* my_lock = new PosixFileLock;
|
|
|
|
my_lock->fd_ = fd;
|
2012-08-18 09:26:50 +02:00
|
|
|
my_lock->filename = fname;
|
2011-03-18 23:37:00 +01:00
|
|
|
*lock = my_lock;
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status UnlockFile(FileLock* lock) {
|
|
|
|
PosixFileLock* my_lock = reinterpret_cast<PosixFileLock*>(lock);
|
|
|
|
Status result;
|
2012-08-18 09:26:50 +02:00
|
|
|
if (LockOrUnlock(my_lock->filename, my_lock->fd_, false) == -1) {
|
2011-07-15 02:20:57 +02:00
|
|
|
result = IOError("unlock", errno);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
close(my_lock->fd_);
|
|
|
|
delete my_lock;
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
virtual void Schedule(void (*function)(void*), void* arg, Priority pri = LOW);
|
2013-03-19 22:39:28 +01:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
virtual void StartThread(void (*function)(void* arg), void* arg);
|
|
|
|
|
2014-02-26 02:47:37 +01:00
|
|
|
virtual void WaitForJoin();
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
virtual Status GetTestDirectory(std::string* result) {
|
|
|
|
const char* env = getenv("TEST_TMPDIR");
|
|
|
|
if (env && env[0] != '\0') {
|
|
|
|
*result = env;
|
|
|
|
} else {
|
|
|
|
char buf[100];
|
2013-10-05 07:32:05 +02:00
|
|
|
snprintf(buf, sizeof(buf), "/tmp/rocksdbtest-%d", int(geteuid()));
|
2011-03-18 23:37:00 +01:00
|
|
|
*result = buf;
|
|
|
|
}
|
|
|
|
// Directory may already exist
|
|
|
|
CreateDir(*result);
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2011-07-21 04:40:18 +02:00
|
|
|
static uint64_t gettid() {
|
2011-03-18 23:37:00 +01:00
|
|
|
pthread_t tid = pthread_self();
|
|
|
|
uint64_t thread_id = 0;
|
2011-03-21 20:40:57 +01:00
|
|
|
memcpy(&thread_id, &tid, std::min(sizeof(thread_id), sizeof(tid)));
|
2011-07-21 04:40:18 +02:00
|
|
|
return thread_id;
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-01-20 11:07:13 +01:00
|
|
|
virtual Status NewLogger(const std::string& fname,
|
|
|
|
shared_ptr<Logger>* result) {
|
2011-07-21 04:40:18 +02:00
|
|
|
FILE* f = fopen(fname.c_str(), "w");
|
2013-03-01 03:04:58 +01:00
|
|
|
if (f == nullptr) {
|
2013-01-20 11:07:13 +01:00
|
|
|
result->reset();
|
2011-07-21 04:40:18 +02:00
|
|
|
return IOError(fname, errno);
|
|
|
|
} else {
|
2013-04-10 04:42:07 +02:00
|
|
|
int fd = fileno(f);
|
|
|
|
SetFD_CLOEXEC(fd, nullptr);
|
2013-10-31 23:36:40 +01:00
|
|
|
result->reset(new PosixLogger(f, &PosixEnv::gettid, this));
|
2011-07-21 04:40:18 +02:00
|
|
|
return Status::OK();
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual uint64_t NowMicros() {
|
|
|
|
struct timeval tv;
|
2013-11-17 08:44:39 +01:00
|
|
|
// TODO(kailiu) MAC DON'T HAVE THIS
|
2013-03-01 03:04:58 +01:00
|
|
|
gettimeofday(&tv, nullptr);
|
2011-03-18 23:37:00 +01:00
|
|
|
return static_cast<uint64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
|
|
|
|
}
|
|
|
|
|
[RocksDB] Added nano second stopwatch and new perf counters to track block read cost
Summary: The pupose of this diff is to expose per user-call level precise timing of block read, so that we can answer questions like: a Get() costs me 100ms, is that somehow related to loading blocks from file system, or sth else? We will answer that with EXACTLY how many blocks have been read, how much time was spent on transfering the bytes from os, how much time was spent on checksum verification and how much time was spent on block decompression, just for that one Get. A nano second stopwatch was introduced to track time with higher precision. The cost/precision of the stopwatch is also measured in unit-test. On my dev box, retrieving one time instance costs about 30ns, on average. The deviation of timing results is good enough to track 100ns-1us level events. And the overhead could be safely ignored for 100us level events (10000 instances/s), for example, a viewstate thrift call.
Test Plan: perf_context_test, also testing with viewstate shadow traffic.
Reviewers: dhruba
Reviewed By: dhruba
CC: leveldb, xjin
Differential Revision: https://reviews.facebook.net/D12351
2013-06-04 08:09:15 +02:00
|
|
|
virtual uint64_t NowNanos() {
|
2013-11-17 08:44:39 +01:00
|
|
|
#ifdef OS_LINUX
|
[RocksDB] Added nano second stopwatch and new perf counters to track block read cost
Summary: The pupose of this diff is to expose per user-call level precise timing of block read, so that we can answer questions like: a Get() costs me 100ms, is that somehow related to loading blocks from file system, or sth else? We will answer that with EXACTLY how many blocks have been read, how much time was spent on transfering the bytes from os, how much time was spent on checksum verification and how much time was spent on block decompression, just for that one Get. A nano second stopwatch was introduced to track time with higher precision. The cost/precision of the stopwatch is also measured in unit-test. On my dev box, retrieving one time instance costs about 30ns, on average. The deviation of timing results is good enough to track 100ns-1us level events. And the overhead could be safely ignored for 100us level events (10000 instances/s), for example, a viewstate thrift call.
Test Plan: perf_context_test, also testing with viewstate shadow traffic.
Reviewers: dhruba
Reviewed By: dhruba
CC: leveldb, xjin
Differential Revision: https://reviews.facebook.net/D12351
2013-06-04 08:09:15 +02:00
|
|
|
struct timespec ts;
|
|
|
|
clock_gettime(CLOCK_MONOTONIC, &ts);
|
|
|
|
return static_cast<uint64_t>(ts.tv_sec) * 1000000000 + ts.tv_nsec;
|
2013-11-17 08:44:39 +01:00
|
|
|
#elif __MACH__
|
|
|
|
clock_serv_t cclock;
|
|
|
|
mach_timespec_t ts;
|
|
|
|
host_get_clock_service(mach_host_self(), CALENDAR_CLOCK, &cclock);
|
|
|
|
clock_get_time(cclock, &ts);
|
|
|
|
mach_port_deallocate(mach_task_self(), cclock);
|
|
|
|
#endif
|
|
|
|
return static_cast<uint64_t>(ts.tv_sec) * 1000000000 + ts.tv_nsec;
|
[RocksDB] Added nano second stopwatch and new perf counters to track block read cost
Summary: The pupose of this diff is to expose per user-call level precise timing of block read, so that we can answer questions like: a Get() costs me 100ms, is that somehow related to loading blocks from file system, or sth else? We will answer that with EXACTLY how many blocks have been read, how much time was spent on transfering the bytes from os, how much time was spent on checksum verification and how much time was spent on block decompression, just for that one Get. A nano second stopwatch was introduced to track time with higher precision. The cost/precision of the stopwatch is also measured in unit-test. On my dev box, retrieving one time instance costs about 30ns, on average. The deviation of timing results is good enough to track 100ns-1us level events. And the overhead could be safely ignored for 100us level events (10000 instances/s), for example, a viewstate thrift call.
Test Plan: perf_context_test, also testing with viewstate shadow traffic.
Reviewers: dhruba
Reviewed By: dhruba
CC: leveldb, xjin
Differential Revision: https://reviews.facebook.net/D12351
2013-06-04 08:09:15 +02:00
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
virtual void SleepForMicroseconds(int micros) {
|
|
|
|
usleep(micros);
|
|
|
|
}
|
|
|
|
|
2012-09-12 18:54:22 +02:00
|
|
|
virtual Status GetHostName(char* name, uint64_t len) {
|
2012-08-15 00:20:36 +02:00
|
|
|
int ret = gethostname(name, len);
|
|
|
|
if (ret < 0) {
|
|
|
|
if (errno == EFAULT || errno == EINVAL)
|
|
|
|
return Status::InvalidArgument(strerror(errno));
|
|
|
|
else
|
|
|
|
return IOError("GetHostName", errno);
|
|
|
|
}
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status GetCurrentTime(int64_t* unix_time) {
|
2013-03-01 03:04:58 +01:00
|
|
|
time_t ret = time(nullptr);
|
2012-08-15 00:20:36 +02:00
|
|
|
if (ret == (time_t) -1) {
|
|
|
|
return IOError("GetCurrentTime", errno);
|
|
|
|
}
|
|
|
|
*unix_time = (int64_t) ret;
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual Status GetAbsolutePath(const std::string& db_path,
|
|
|
|
std::string* output_path) {
|
|
|
|
if (db_path.find('/') == 0) {
|
|
|
|
*output_path = db_path;
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
|
|
|
char the_path[256];
|
|
|
|
char* ret = getcwd(the_path, 256);
|
2013-03-01 03:04:58 +01:00
|
|
|
if (ret == nullptr) {
|
2012-08-15 00:20:36 +02:00
|
|
|
return Status::IOError(strerror(errno));
|
|
|
|
}
|
|
|
|
|
|
|
|
*output_path = ret;
|
|
|
|
return Status::OK();
|
|
|
|
}
|
|
|
|
|
2012-11-29 01:42:36 +01:00
|
|
|
// Allow increasing the number of worker threads.
|
2013-09-12 09:53:30 +02:00
|
|
|
virtual void SetBackgroundThreads(int num, Priority pri) {
|
|
|
|
assert(pri >= Priority::LOW && pri <= Priority::HIGH);
|
|
|
|
thread_pools_[pri].SetBackgroundThreads(num);
|
2012-09-20 00:21:09 +02:00
|
|
|
}
|
|
|
|
|
2012-10-19 23:00:53 +02:00
|
|
|
virtual std::string TimeToString(uint64_t secondsSince1970) {
|
|
|
|
const time_t seconds = (time_t)secondsSince1970;
|
|
|
|
struct tm t;
|
|
|
|
int maxsize = 64;
|
|
|
|
std::string dummy;
|
|
|
|
dummy.reserve(maxsize);
|
|
|
|
dummy.resize(maxsize);
|
|
|
|
char* p = &dummy[0];
|
|
|
|
localtime_r(&seconds, &t);
|
|
|
|
snprintf(p, maxsize,
|
|
|
|
"%04d/%02d/%02d-%02d:%02d:%02d ",
|
|
|
|
t.tm_year + 1900,
|
|
|
|
t.tm_mon + 1,
|
|
|
|
t.tm_mday,
|
|
|
|
t.tm_hour,
|
|
|
|
t.tm_min,
|
|
|
|
t.tm_sec);
|
|
|
|
return dummy;
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
private:
|
2013-03-21 19:12:42 +01:00
|
|
|
bool checkedDiskForMmap_;
|
|
|
|
bool forceMmapOff; // do we override Env options?
|
2013-03-13 21:50:26 +01:00
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-01-07 19:11:18 +01:00
|
|
|
// Returns true iff the named directory exists and is a directory.
|
|
|
|
virtual bool DirExists(const std::string& dname) {
|
|
|
|
struct stat statbuf;
|
|
|
|
if (stat(dname.c_str(), &statbuf) == 0) {
|
|
|
|
return S_ISDIR(statbuf.st_mode);
|
|
|
|
}
|
|
|
|
return false; // stat() failed return false
|
|
|
|
}
|
|
|
|
|
2013-03-13 21:50:26 +01:00
|
|
|
bool SupportsFastAllocate(const std::string& path) {
|
2013-12-11 20:18:00 +01:00
|
|
|
#ifdef ROCKSDB_FALLOCATE_PRESENT
|
2013-03-13 21:50:26 +01:00
|
|
|
struct statfs s;
|
|
|
|
if (statfs(path.c_str(), &s)){
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
switch (s.f_type) {
|
|
|
|
case EXT4_SUPER_MAGIC:
|
|
|
|
return true;
|
|
|
|
case XFS_SUPER_MAGIC:
|
|
|
|
return true;
|
|
|
|
case TMPFS_MAGIC:
|
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
2013-11-17 08:44:39 +01:00
|
|
|
#else
|
|
|
|
return false;
|
|
|
|
#endif
|
2013-03-13 21:50:26 +01:00
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
size_t page_size_;
|
|
|
|
|
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
class ThreadPool {
|
|
|
|
public:
|
2013-03-19 22:39:28 +01:00
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
ThreadPool() :
|
|
|
|
total_threads_limit_(1),
|
|
|
|
bgthreads_(0),
|
|
|
|
queue_(),
|
|
|
|
exit_all_threads_(false) {
|
|
|
|
PthreadCall("mutex_init", pthread_mutex_init(&mu_, nullptr));
|
|
|
|
PthreadCall("cvar_init", pthread_cond_init(&bgsignal_, nullptr));
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
~ThreadPool() {
|
|
|
|
PthreadCall("lock", pthread_mutex_lock(&mu_));
|
|
|
|
assert(!exit_all_threads_);
|
|
|
|
exit_all_threads_ = true;
|
|
|
|
PthreadCall("signalall", pthread_cond_broadcast(&bgsignal_));
|
|
|
|
PthreadCall("unlock", pthread_mutex_unlock(&mu_));
|
|
|
|
for (const auto tid : bgthreads_) {
|
|
|
|
pthread_join(tid, nullptr);
|
|
|
|
}
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
void BGThread() {
|
|
|
|
while (true) {
|
|
|
|
// Wait until there is an item that is ready to run
|
|
|
|
PthreadCall("lock", pthread_mutex_lock(&mu_));
|
|
|
|
while (queue_.empty() && !exit_all_threads_) {
|
|
|
|
PthreadCall("wait", pthread_cond_wait(&bgsignal_, &mu_));
|
|
|
|
}
|
|
|
|
if (exit_all_threads_) { // mechanism to let BG threads exit safely
|
|
|
|
PthreadCall("unlock", pthread_mutex_unlock(&mu_));
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
void (*function)(void*) = queue_.front().function;
|
|
|
|
void* arg = queue_.front().arg;
|
|
|
|
queue_.pop_front();
|
2013-03-29 20:59:15 +01:00
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
PthreadCall("unlock", pthread_mutex_unlock(&mu_));
|
|
|
|
(*function)(arg);
|
|
|
|
}
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
static void* BGThreadWrapper(void* arg) {
|
|
|
|
reinterpret_cast<ThreadPool*>(arg)->BGThread();
|
|
|
|
return nullptr;
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
void SetBackgroundThreads(int num) {
|
|
|
|
PthreadCall("lock", pthread_mutex_lock(&mu_));
|
|
|
|
if (num > total_threads_limit_) {
|
|
|
|
total_threads_limit_ = num;
|
|
|
|
}
|
|
|
|
assert(total_threads_limit_ > 0);
|
|
|
|
PthreadCall("unlock", pthread_mutex_unlock(&mu_));
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
2013-09-12 09:53:30 +02:00
|
|
|
|
|
|
|
void Schedule(void (*function)(void*), void* arg) {
|
|
|
|
PthreadCall("lock", pthread_mutex_lock(&mu_));
|
|
|
|
|
|
|
|
if (exit_all_threads_) {
|
|
|
|
PthreadCall("unlock", pthread_mutex_unlock(&mu_));
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
// Start background thread if necessary
|
|
|
|
while ((int)bgthreads_.size() < total_threads_limit_) {
|
|
|
|
pthread_t t;
|
|
|
|
PthreadCall(
|
|
|
|
"create thread",
|
|
|
|
pthread_create(&t,
|
|
|
|
nullptr,
|
|
|
|
&ThreadPool::BGThreadWrapper,
|
|
|
|
this));
|
2013-11-27 03:00:43 +01:00
|
|
|
|
|
|
|
// Set the thread name to aid debugging
|
2013-12-05 08:00:33 +01:00
|
|
|
#if defined(_GNU_SOURCE) && defined(__GLIBC_PREREQ)
|
|
|
|
#if __GLIBC_PREREQ(2, 12)
|
2013-11-27 03:00:43 +01:00
|
|
|
char name_buf[16];
|
|
|
|
snprintf(name_buf, sizeof name_buf, "rocksdb:bg%zu", bgthreads_.size());
|
|
|
|
name_buf[sizeof name_buf - 1] = '\0';
|
|
|
|
pthread_setname_np(t, name_buf);
|
2013-12-05 08:00:33 +01:00
|
|
|
#endif
|
2013-11-27 03:00:43 +01:00
|
|
|
#endif
|
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
bgthreads_.push_back(t);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Add to priority queue
|
|
|
|
queue_.push_back(BGItem());
|
|
|
|
queue_.back().function = function;
|
|
|
|
queue_.back().arg = arg;
|
|
|
|
|
|
|
|
// always wake up at least one waiting thread.
|
|
|
|
PthreadCall("signal", pthread_cond_signal(&bgsignal_));
|
|
|
|
|
2013-03-19 22:39:28 +01:00
|
|
|
PthreadCall("unlock", pthread_mutex_unlock(&mu_));
|
|
|
|
}
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-09-12 09:53:30 +02:00
|
|
|
private:
|
|
|
|
// Entry per Schedule() call
|
|
|
|
struct BGItem { void* arg; void (*function)(void*); };
|
|
|
|
typedef std::deque<BGItem> BGQueue;
|
|
|
|
|
|
|
|
pthread_mutex_t mu_;
|
|
|
|
pthread_cond_t bgsignal_;
|
|
|
|
int total_threads_limit_;
|
|
|
|
std::vector<pthread_t> bgthreads_;
|
|
|
|
BGQueue queue_;
|
|
|
|
bool exit_all_threads_;
|
|
|
|
};
|
|
|
|
|
|
|
|
std::vector<ThreadPool> thread_pools_;
|
|
|
|
|
|
|
|
pthread_mutex_t mu_;
|
|
|
|
std::vector<pthread_t> threads_to_join_;
|
|
|
|
|
|
|
|
};
|
|
|
|
|
|
|
|
PosixEnv::PosixEnv() : checkedDiskForMmap_(false),
|
|
|
|
forceMmapOff(false),
|
|
|
|
page_size_(getpagesize()),
|
|
|
|
thread_pools_(Priority::TOTAL) {
|
|
|
|
PthreadCall("mutex_init", pthread_mutex_init(&mu_, nullptr));
|
|
|
|
}
|
|
|
|
|
|
|
|
void PosixEnv::Schedule(void (*function)(void*), void* arg, Priority pri) {
|
|
|
|
assert(pri >= Priority::LOW && pri <= Priority::HIGH);
|
|
|
|
thread_pools_[pri].Schedule(function, arg);
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
struct StartThreadState {
|
|
|
|
void (*user_function)(void*);
|
|
|
|
void* arg;
|
|
|
|
};
|
|
|
|
}
|
|
|
|
static void* StartThreadWrapper(void* arg) {
|
|
|
|
StartThreadState* state = reinterpret_cast<StartThreadState*>(arg);
|
|
|
|
state->user_function(state->arg);
|
|
|
|
delete state;
|
2013-03-01 03:04:58 +01:00
|
|
|
return nullptr;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
void PosixEnv::StartThread(void (*function)(void* arg), void* arg) {
|
|
|
|
pthread_t t;
|
|
|
|
StartThreadState* state = new StartThreadState;
|
|
|
|
state->user_function = function;
|
|
|
|
state->arg = arg;
|
|
|
|
PthreadCall("start thread",
|
2013-03-01 03:04:58 +01:00
|
|
|
pthread_create(&t, nullptr, &StartThreadWrapper, state));
|
2013-09-12 09:53:30 +02:00
|
|
|
PthreadCall("lock", pthread_mutex_lock(&mu_));
|
2013-03-19 22:39:28 +01:00
|
|
|
threads_to_join_.push_back(t);
|
2013-09-12 09:53:30 +02:00
|
|
|
PthreadCall("unlock", pthread_mutex_unlock(&mu_));
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2014-02-26 02:47:37 +01:00
|
|
|
void PosixEnv::WaitForJoin() {
|
|
|
|
for (const auto tid : threads_to_join_) {
|
|
|
|
pthread_join(tid, nullptr);
|
|
|
|
}
|
|
|
|
threads_to_join_.clear();
|
|
|
|
}
|
|
|
|
|
2011-10-31 18:22:06 +01:00
|
|
|
} // namespace
|
2011-03-18 23:37:00 +01:00
|
|
|
|
2013-10-18 23:50:54 +02:00
|
|
|
std::string Env::GenerateUniqueId() {
|
|
|
|
std::string uuid_file = "/proc/sys/kernel/random/uuid";
|
|
|
|
if (FileExists(uuid_file)) {
|
|
|
|
std::string uuid;
|
|
|
|
Status s = ReadFileToString(this, uuid_file, &uuid);
|
|
|
|
if (s.ok()) {
|
|
|
|
return uuid;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
// Could not read uuid_file - generate uuid using "nanos-random"
|
|
|
|
Random64 r(time(nullptr));
|
|
|
|
uint64_t random_uuid_portion =
|
|
|
|
r.Uniform(std::numeric_limits<uint64_t>::max());
|
|
|
|
uint64_t nanos_uuid_portion = NowNanos();
|
|
|
|
char uuid2[200];
|
2013-11-17 08:44:39 +01:00
|
|
|
snprintf(uuid2,
|
|
|
|
200,
|
|
|
|
"%lx-%lx",
|
|
|
|
(unsigned long)nanos_uuid_portion,
|
|
|
|
(unsigned long)random_uuid_portion);
|
2013-10-18 23:50:54 +02:00
|
|
|
return uuid2;
|
|
|
|
}
|
|
|
|
|
2011-03-18 23:37:00 +01:00
|
|
|
Env* Env::Default() {
|
2013-04-23 03:10:28 +02:00
|
|
|
static PosixEnv default_env;
|
2013-03-19 22:39:28 +01:00
|
|
|
return &default_env;
|
2011-03-18 23:37:00 +01:00
|
|
|
}
|
|
|
|
|
2013-10-04 06:49:15 +02:00
|
|
|
} // namespace rocksdb
|