Optimize fallocation

Summary:
Based on my recent findings (posted in our internal group), if we use fallocate without KEEP_SIZE flag, we get superior performance of fdatasync() in append-only workloads.

This diff provides an option for user to not use KEEP_SIZE flag, thus optimizing his sync performance by up to 2x-3x.

At one point we also just called posix_fallocate instead of fallocate, which isn't very fast: http://code.woboq.org/userspace/glibc/sysdeps/posix/posix_fallocate.c.html (tl;dr it manually writes out zero bytes to allocate storage). This diff also fixes that, by first calling fallocate and then posix_fallocate if fallocate is not supported.

Test Plan: make check

Reviewers: dhruba, sdong, haobo, ljin

Reviewed By: dhruba

CC: leveldb

Differential Revision: https://reviews.facebook.net/D16761

Author: igorcanadi

db/db_impl.cc

@@ -456,8 +456,8 @@ Status DBImpl::NewDB() {
 
   const std::string manifest = DescriptorFileName(dbname_, 1);
   unique_ptr<WritableFile> file;
-  Status s = env_->NewWritableFile(manifest, &file,
-                                   storage_options_.AdaptForLogWrite());
+  Status s = env_->NewWritableFile(
+      manifest, &file, env_->OptimizeForManifestWrite(storage_options_));
   if (!s.ok()) {
     return s;
   }
@@ -3626,7 +3626,8 @@ Status DBImpl::MakeRoomForWrite(bool force,
       {
         DelayLoggingAndReset();
         s = env_->NewWritableFile(LogFileName(options_.wal_dir, new_log_number),
-                                  &lfile, storage_options_.AdaptForLogWrite());
+                                  &lfile,
+                                  env_->OptimizeForLogWrite(storage_options_));
         if (s.ok()) {
           // Our final size should be less than write_buffer_size
           // (compression, etc) but err on the side of caution.
@@ -3912,7 +3913,7 @@ Status DB::Open(const Options& options, const std::string& dbname, DB** dbptr) {
     EnvOptions soptions(options);
     s = impl->options_.env->NewWritableFile(
         LogFileName(impl->options_.wal_dir, new_log_number), &lfile,
-        soptions.AdaptForLogWrite());
+        impl->options_.env->OptimizeForLogWrite(soptions));
     if (s.ok()) {
       lfile->SetPreallocationBlockSize(1.1 * impl->options_.write_buffer_size);
       VersionEdit edit;

db/repair.cc

@@ -306,8 +306,8 @@ class Repairer {
   Status WriteDescriptor() {
     std::string tmp = TempFileName(dbname_, 1);
     unique_ptr<WritableFile> file;
-    Status status =
-        env_->NewWritableFile(tmp, &file, storage_options_.AdaptForLogWrite());
+    Status status = env_->NewWritableFile(
+        tmp, &file, env_->OptimizeForManifestWrite(storage_options_));
     if (!status.ok()) {
       return status;
     }

db/version_set.cc

@@ -1564,7 +1564,7 @@ Status VersionSet::LogAndApply(VersionEdit* edit, port::Mutex* mu,
       unique_ptr<WritableFile> descriptor_file;
       s = env_->NewWritableFile(
           DescriptorFileName(dbname_, pending_manifest_file_number_),
-          &descriptor_file, storage_options_.AdaptForLogWrite());
+          &descriptor_file, env_->OptimizeForManifestWrite(storage_options_));
       if (s.ok()) {
         descriptor_log_.reset(new log::Writer(std::move(descriptor_file)));
         s = WriteSnapshot(descriptor_log_.get());

include/rocksdb/env.h

@@ -49,8 +49,6 @@ struct EnvOptions {
   // construct from Options
   explicit EnvOptions(const Options& options);
 
-  EnvOptions AdaptForLogWrite() const;
-
   // If true, then allow caching of data in environment buffers
   bool use_os_buffer = true;
 
@@ -61,13 +59,21 @@ struct EnvOptions {
   bool use_mmap_writes = true;
 
   // If true, set the FD_CLOEXEC on open fd.
-  bool set_fd_cloexec= true;
+  bool set_fd_cloexec = true;
 
   // Allows OS to incrementally sync files to disk while they are being
   // written, in the background. Issue one request for every bytes_per_sync
   // written. 0 turns it off.
   // Default: 0
   uint64_t bytes_per_sync = 0;
+
+  // If true, we will preallocate the file with FALLOC_FL_KEEP_SIZE flag, which
+  // means that file size won't change as part of preallocation.
+  // If false, preallocation will also change the file size. This option will
+  // improve the performance in workloads where you sync the data on every
+  // write. By default, we set it to true for MANIFEST writes and false for
+  // WAL writes
+  bool fallocate_with_keep_size = true;
 };
 
 class Env {
@@ -260,6 +266,16 @@ class Env {
   // Generates a unique id that can be used to identify a db
   virtual std::string GenerateUniqueId();
 
+  // OptimizeForLogWrite will create a new EnvOptions object that is a copy of
+  // the EnvOptions in the parameters, but is optimized for writing log files.
+  // Default implementation returns the copy of the same object.
+  virtual EnvOptions OptimizeForLogWrite(const EnvOptions& env_options) const;
+  // OptimizeForManifestWrite will create a new EnvOptions object that is a copy
+  // of the EnvOptions in the parameters, but is optimized for writing manifest
+  // files. Default implementation returns the copy of the same object.
+  virtual EnvOptions OptimizeForManifestWrite(const EnvOptions& env_options)
+      const;
+
  private:
   // No copying allowed
   Env(const Env&);

util/env.cc

@@ -241,10 +241,12 @@ void AssignEnvOptions(EnvOptions* env_options, const Options& options) {
 
 }
 
-EnvOptions EnvOptions::AdaptForLogWrite() const {
-  EnvOptions adapted = *this;
-  adapted.use_mmap_writes = false;
-  return adapted;
+EnvOptions Env::OptimizeForLogWrite(const EnvOptions& env_options) const {
+  return env_options;
+}
+
+EnvOptions Env::OptimizeForManifestWrite(const EnvOptions& env_options) const {
+  return env_options;
 }
 
 EnvOptions::EnvOptions(const Options& options) {

util/env_posix.cc

@@ -354,9 +354,9 @@ class PosixMmapFile : public WritableFile {
   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_;
+  bool fallocate_with_keep_size_;
 
   // Roundup x to a multiple of y
   static size_t Roundup(size_t x, size_t y) {
@@ -399,7 +399,12 @@ class PosixMmapFile : public WritableFile {
     assert(base_ == nullptr);
 
     TEST_KILL_RANDOM(rocksdb_kill_odds);
-    int alloc_status = posix_fallocate(fd_, file_offset_, map_size_);
+    // we can't fallocate with FALLOC_FL_KEEP_SIZE here
+    int alloc_status = fallocate(fd_, 0, file_offset_, map_size_);
+    if (alloc_status != 0) {
+      // fallback to posix_fallocate
+      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));
@@ -436,7 +441,8 @@ class PosixMmapFile : public WritableFile {
         dst_(nullptr),
         last_sync_(nullptr),
         file_offset_(0),
-        pending_sync_(false) {
+        pending_sync_(false),
+        fallocate_with_keep_size_(options.fallocate_with_keep_size) {
     assert((page_size & (page_size - 1)) == 0);
     assert(options.use_mmap_writes);
   }
@@ -584,7 +590,9 @@ class PosixMmapFile : public WritableFile {
 #ifdef ROCKSDB_FALLOCATE_PRESENT
   virtual Status Allocate(off_t offset, off_t len) {
     TEST_KILL_RANDOM(rocksdb_kill_odds);
-    if (!fallocate(fd_, FALLOC_FL_KEEP_SIZE, offset, len)) {
+    int alloc_status = fallocate(
+        fd_, fallocate_with_keep_size_ ? FALLOC_FL_KEEP_SIZE : 0, offset, len);
+    if (alloc_status == 0) {
       return Status::OK();
     } else {
       return IOError(filename_, errno);
@@ -606,20 +614,22 @@ class PosixWritableFile : public WritableFile {
   bool pending_fsync_;
   uint64_t last_sync_size_;
   uint64_t bytes_per_sync_;
+  bool fallocate_with_keep_size_;
 
  public:
   PosixWritableFile(const std::string& fname, int fd, size_t capacity,
-                    const EnvOptions& options) :
-    filename_(fname),
-    fd_(fd),
-    cursize_(0),
-    capacity_(capacity),
-    buf_(new char[capacity]),
-    filesize_(0),
-    pending_sync_(false),
-    pending_fsync_(false),
-    last_sync_size_(0),
-    bytes_per_sync_(options.bytes_per_sync) {
+                    const EnvOptions& options)
+      : filename_(fname),
+        fd_(fd),
+        cursize_(0),
+        capacity_(capacity),
+        buf_(new char[capacity]),
+        filesize_(0),
+        pending_sync_(false),
+        pending_fsync_(false),
+        last_sync_size_(0),
+        bytes_per_sync_(options.bytes_per_sync),
+        fallocate_with_keep_size_(options.fallocate_with_keep_size) {
     assert(!options.use_mmap_writes);
   }
 
@@ -771,7 +781,9 @@ class PosixWritableFile : public WritableFile {
 #ifdef ROCKSDB_FALLOCATE_PRESENT
   virtual Status Allocate(off_t offset, off_t len) {
     TEST_KILL_RANDOM(rocksdb_kill_odds);
-    if (!fallocate(fd_, FALLOC_FL_KEEP_SIZE, offset, len)) {
+    int alloc_status = fallocate(
+        fd_, fallocate_with_keep_size_ ? FALLOC_FL_KEEP_SIZE : 0, offset, len);
+    if (alloc_status == 0) {
       return Status::OK();
     } else {
       return IOError(filename_, errno);
@@ -797,14 +809,15 @@ class PosixRandomRWFile : public RandomRWFile {
   int fd_;
   bool pending_sync_;
   bool pending_fsync_;
+  bool fallocate_with_keep_size_;
 
  public:
-  PosixRandomRWFile(const std::string& fname, int fd,
-                    const EnvOptions& options) :
-      filename_(fname),
-      fd_(fd),
-      pending_sync_(false),
-      pending_fsync_(false) {
+  PosixRandomRWFile(const std::string& fname, int fd, const EnvOptions& options)
+      : filename_(fname),
+        fd_(fd),
+        pending_sync_(false),
+        pending_fsync_(false),
+        fallocate_with_keep_size_(options.fallocate_with_keep_size) {
     assert(!options.use_mmap_writes && !options.use_mmap_reads);
   }
 
@@ -874,7 +887,10 @@ class PosixRandomRWFile : public RandomRWFile {
 
 #ifdef ROCKSDB_FALLOCATE_PRESENT
   virtual Status Allocate(off_t offset, off_t len) {
-    if (!fallocate(fd_, FALLOC_FL_KEEP_SIZE, offset, len)) {
+    TEST_KILL_RANDOM(rocksdb_kill_odds);
+    int alloc_status = fallocate(
+        fd_, fallocate_with_keep_size_ ? FALLOC_FL_KEEP_SIZE : 0, offset, len);
+    if (alloc_status == 0) {
       return Status::OK();
     } else {
       return IOError(filename_, errno);
@@ -1332,6 +1348,20 @@ class PosixEnv : public Env {
     return dummy;
   }
 
+  EnvOptions OptimizeForLogWrite(const EnvOptions& env_options) const {
+    EnvOptions optimized = env_options;
+    optimized.use_mmap_writes = false;
+    optimized.fallocate_with_keep_size = true;
+    return optimized;
+  }
+
+  EnvOptions OptimizeForManifestWrite(const EnvOptions& env_options) const {
+    EnvOptions optimized = env_options;
+    optimized.use_mmap_writes = false;
+    optimized.fallocate_with_keep_size = true;
+    return optimized;
+  }
+
  private:
   bool checkedDiskForMmap_;
   bool forceMmapOff; // do we override Env options?