SimpleCache: merge the first and second stream in one file

net/disk_cache/simple/simple_entry_impl.cc

Index: net/disk_cache/simple/simple_entry_impl.cc
diff --git a/net/disk_cache/simple/simple_entry_impl.cc b/net/disk_cache/simple/simple_entry_impl.cc
index cfc2689622b7b4495aaafc278584f8131c590237..4e8243a96b5e6fd01eaa95bb1b4027e8c8d618cb 100644
--- a/net/disk_cache/simple/simple_entry_impl.cc
+++ b/net/disk_cache/simple/simple_entry_impl.cc
@@ -335,7 +335,7 @@ int SimpleEntryImpl::ReadData(int stream_index,
                                           false));
   }
 
-  if (stream_index < 0 || stream_index >= kSimpleEntryFileCount ||
+  if (stream_index < 0 || stream_index >= kSimpleEntryStreamCount ||
       buf_len < 0) {
     if (net_log_.IsLoggingAllEvents()) {
       net_log_.AddEvent(net::NetLog::TYPE_SIMPLE_CACHE_ENTRY_READ_END,
@@ -381,8 +381,8 @@ int SimpleEntryImpl::WriteData(int stream_index,
                                           truncate));
   }
 
-  if (stream_index < 0 || stream_index >= kSimpleEntryFileCount || offset < 0 ||
-      buf_len < 0) {
+  if (stream_index < 0 || stream_index >= kSimpleEntryStreamCount ||
+      offset < 0 || buf_len < 0) {
     if (net_log_.IsLoggingAllEvents()) {
       net_log_.AddEvent(
           net::NetLog::TYPE_SIMPLE_CACHE_ENTRY_WRITE_END,
@@ -402,16 +402,11 @@ int SimpleEntryImpl::WriteData(int stream_index,
   }
   ScopedOperationRunner operation_runner(this);
 
-  // Currently, Simple Cache is only used for HTTP, which stores the headers in
-  // stream 0 and always writes them with a single, truncating write.  Detect
-  // these writes and record the size and size changes of the headers.  Also,
-  // note writes to stream 0 that violate those assumptions.
-  if (stream_index == 0) {
-    if (offset == 0 && truncate)
-      RecordHeaderSizeChange(cache_type_, data_size_[0], buf_len);
-    else
-      RecordUnexpectedStream0Write(cache_type_);
-  }
+  // Stream 0 data is kept in memory, so can be written immediatly if there are
+  // no IO operations pending.
+  if (stream_index == 0 && state_ == STATE_READY &&
+      pending_operations_.size() == 0)
+    return WriteStream0Data(buf, offset, buf_len, truncate);

[pasko, 2013/09/09 20:01:22]

This is an optimization over copying the contents of the buffer twice in the
optimistic write. That is neat. However, reading this and verifying correctness
is not trivial.

I would recommend disabling optimistic operations on stream 0 below, this would
merge this codepath with the main one at the cost of an extra push/pop at
entry_operations_. I think this cost is small.

I would even argue that just dropping these lines and going with double copying
is nicer because it's simpler.

[clamy, 2013/09/10 14:35:59]

Done.

 
   // We can only do optimistic Write if there is no pending operations, so
   // that we are sure that the next call to RunNextOperationIfNeeded will
@@ -676,7 +671,7 @@ void SimpleEntryImpl::CreateEntryInternal(bool have_index,
   last_used_ = last_modified_ = base::Time::Now();
 
   // If creation succeeds, we should mark all streams to be saved on close.
-  for (int i = 0; i < kSimpleEntryFileCount; ++i)
+  for (int i = 0; i < kSimpleEntryStreamCount; ++i)
     have_written_[i] = true;
 
   const base::TimeTicks start_time = base::TimeTicks::Now();
@@ -711,7 +706,7 @@ void SimpleEntryImpl::CloseInternal() {
   if (state_ == STATE_READY) {
     DCHECK(synchronous_entry_);
     state_ = STATE_IO_PENDING;
-    for (int i = 0; i < kSimpleEntryFileCount; ++i) {
+    for (int i = 0; i < kSimpleEntryStreamCount; ++i) {
       if (have_written_[i]) {
         if (GetDataSize(i) == crc32s_end_offset_[i]) {
           int32 crc = GetDataSize(i) == 0 ? crc32(0, Z_NULL, 0) : crc32s_[i];
@@ -730,12 +725,13 @@ void SimpleEntryImpl::CloseInternal() {
         base::Bind(&SimpleSynchronousEntry::Close,
                    base::Unretained(synchronous_entry_),
                    SimpleEntryStat(last_used_, last_modified_, data_size_),
-                   base::Passed(&crc32s_to_write));
+                   base::Passed(&crc32s_to_write),
+                   stream_0_data_);
     Closure reply = base::Bind(&SimpleEntryImpl::CloseOperationComplete, this);
     synchronous_entry_ = NULL;
     worker_pool_->PostTaskAndReply(FROM_HERE, task, reply);
 
-    for (int i = 0; i < kSimpleEntryFileCount; ++i) {
+    for (int i = 0; i < kSimpleEntryStreamCount; ++i) {
       if (!have_written_[i]) {
         SIMPLE_CACHE_UMA(ENUMERATION,
                          "CheckCRCResult", cache_type_,
@@ -787,6 +783,16 @@ void SimpleEntryImpl::ReadDataInternal(int stream_index,
     return;
   }
 
+  // Stream 0 data is kept in memory, so can be read immediately.

[pasko, 2013/09/09 20:01:22]

s/so can be//

we do not have an option to post a task to worker pool to perform the operation,
so the comment should not read as if this is an optimization.

[clamy, 2013/09/10 14:35:59]

Done.

+  if (stream_index == 0) {
+    int ret_value = ReadStream0Data(buf, offset, buf_len);
+    if (!callback.is_null()) {
+      MessageLoopProxy::current()
+          ->PostTask(FROM_HERE, base::Bind(callback, ret_value));
+    }
+    return;
+  }
+
   buf_len = std::min(buf_len, GetDataSize(stream_index) - offset);
 
   state_ = STATE_IO_PENDING;
@@ -795,14 +801,15 @@ void SimpleEntryImpl::ReadDataInternal(int stream_index,
 
   scoped_ptr<uint32> read_crc32(new uint32());
   scoped_ptr<int> result(new int());
-  scoped_ptr<base::Time> last_used(new base::Time());
+  scoped_ptr<SimpleEntryStat> entry_stat(
+      new SimpleEntryStat(last_used_, last_modified_, data_size_));
   Closure task = base::Bind(
       &SimpleSynchronousEntry::ReadData,
       base::Unretained(synchronous_entry_),
       SimpleSynchronousEntry::EntryOperationData(stream_index, offset, buf_len),
       make_scoped_refptr(buf),
       read_crc32.get(),
-      last_used.get(),
+      entry_stat.get(),
       result.get());
   Closure reply = base::Bind(&SimpleEntryImpl::ReadOperationComplete,
                              this,
@@ -810,7 +817,7 @@ void SimpleEntryImpl::ReadDataInternal(int stream_index,
                              offset,
                              callback,
                              base::Passed(&read_crc32),
-                             base::Passed(&last_used),
+                             base::Passed(&entry_stat),
                              base::Passed(&result));
   worker_pool_->PostTaskAndReply(FROM_HERE, task, reply);
 }
@@ -848,6 +855,17 @@ void SimpleEntryImpl::WriteDataInternal(int stream_index,
     return;
   }
 
+  // Stream 0 data is kept in memory, so can be written immediatly.
+  if (stream_index == 0) {
+    int ret_value = WriteStream0Data(buf, offset, buf_len, truncate);
+    if (!callback.is_null()) {
+      // Postask prevents creating a loop when calling the callback directly.
+      MessageLoopProxy::current()
+          ->PostTask(FROM_HERE, base::Bind(callback, ret_value));
+    }
+    return;
+  }
+
   DCHECK_EQ(STATE_READY, state_);

[pasko, 2013/09/09 20:01:22]

this DCHECK belongs better to a place _before_ we do anything with data

[clamy, 2013/09/10 14:35:59]

Done.

   state_ = STATE_IO_PENDING;
   if (backend_.get())
@@ -883,6 +901,9 @@ void SimpleEntryImpl::WriteDataInternal(int stream_index,
   last_used_ = last_modified_ = base::Time::Now();
 
   have_written_[stream_index] = true;
+  // Writing on stream 1 affects the placement of stream 0 in the file.
+  if (stream_index == 1)
+    have_written_[0] = true;
 
   scoped_ptr<int> result(new int());
   Closure task = base::Bind(&SimpleSynchronousEntry::WriteData,
@@ -934,6 +955,9 @@ void SimpleEntryImpl::CreationOperationComplete(
 
   state_ = STATE_READY;
   synchronous_entry_ = in_results->sync_entry;
+  stream_0_data_ = in_results->stream_0_data;
+  // The crc was read in SimpleSynchronousEntry.
+  crc_check_state_[0] = CRC_CHECK_DONE;
   if (key_.empty()) {
     SetKey(synchronous_entry_->key());
   } else {
@@ -984,7 +1008,7 @@ void SimpleEntryImpl::ReadOperationComplete(
     int offset,
     const CompletionCallback& completion_callback,
     scoped_ptr<uint32> read_crc32,
-    scoped_ptr<base::Time> last_used,
+    scoped_ptr<SimpleEntryStat> entry_stat,
     scoped_ptr<int> result) {
   DCHECK(io_thread_checker_.CalledOnValidThread());
   DCHECK(synchronous_entry_);
@@ -1020,7 +1044,7 @@ void SimpleEntryImpl::ReadOperationComplete(
       Closure task = base::Bind(&SimpleSynchronousEntry::CheckEOFRecord,
                                 base::Unretained(synchronous_entry_),
                                 stream_index,
-                                data_size_[stream_index],
+                                data_size_,
                                 crc32s_[stream_index],
                                 new_result.get());
       Closure reply = base::Bind(&SimpleEntryImpl::ChecksumOperationComplete,
@@ -1049,10 +1073,7 @@ void SimpleEntryImpl::ReadOperationComplete(
   }
 
   EntryOperationComplete(
-      stream_index,
-      completion_callback,
-      SimpleEntryStat(*last_used, last_modified_, data_size_),
-      result.Pass());
+      stream_index, completion_callback, *entry_stat, result.Pass());

[pasko, 2013/09/09 20:01:22]

probably .Pass() would be more stylish

[clamy, 2013/09/10 14:35:59]

EntryOperationComplete expects a const SimpleEntryStat&.

 }
 
 void SimpleEntryImpl::WriteOperationComplete(
@@ -1129,7 +1150,7 @@ void SimpleEntryImpl::UpdateDataFromEntryStat(
 
   last_used_ = entry_stat.last_used;
   last_modified_ = entry_stat.last_modified;
-  for (int i = 0; i < kSimpleEntryFileCount; ++i) {
+  for (int i = 0; i < kSimpleEntryStreamCount; ++i) {
     data_size_[i] = entry_stat.data_size[i];
   }
   if (backend_.get())
@@ -1138,7 +1159,7 @@ void SimpleEntryImpl::UpdateDataFromEntryStat(
 
 int64 SimpleEntryImpl::GetDiskUsage() const {
   int64 file_size = 0;
-  for (int i = 0; i < kSimpleEntryFileCount; ++i) {
+  for (int i = 0; i < kSimpleEntryStreamCount; ++i) {
     file_size +=
         simple_util::GetFileSizeFromKeyAndDataSize(key_, data_size_[i]);
   }
@@ -1216,4 +1237,66 @@ void SimpleEntryImpl::RecordWriteDependencyType(
                    type, WRITE_DEPENDENCY_TYPE_MAX);
 }
 
+int SimpleEntryImpl::ReadStream0Data(net::IOBuffer* buf,
+                                     int offset,
+                                     int buf_len) {
+  int read_size = std::min(data_size_[0] - offset, buf_len);
+  if (read_size < 0) {
+    RecordReadResult(cache_type_, READ_RESULT_SUCCESS);
+    return 0;
+  }
+  memcpy(buf->data(), stream_0_data_->data() + offset, read_size);
+  UpdateDataFromEntryStat(
+      SimpleEntryStat(base::Time::Now(), last_modified_, data_size_));
+  RecordReadResult(cache_type_, READ_RESULT_SUCCESS);
+  return read_size;
+}
+
+int SimpleEntryImpl::WriteStream0Data(net::IOBuffer* buf,
+                                      int offset,
+                                      int buf_len,
+                                      bool truncate) {
+  // Currently, Simple Cache is only used for HTTP, which stores the headers in
+  // stream 0 and always writes them with a single, truncating write.  Detect
+  // these writes and record the size and size changes of the headers.  Also,
+  // note writes to stream 0 that violate those assumptions.

[pasko, 2013/09/09 20:01:22]

nit: not sure: s/note/support/ ?

[clamy, 2013/09/10 14:35:59]

Done.

+  have_written_[0] = true;
+  if (offset == 0 && truncate) {
+    RecordHeaderSizeChange(cache_type_, data_size_[0], buf_len);
+    stream_0_data_ = new IOBuffer(buf_len);

[pasko, 2013/09/09 20:01:22]

please use a GrowableIOBuffer, it might be less optimal in some cases, but
certainly shorter to write and easier to read.

[clamy, 2013/09/10 14:35:59]

Done.

+    memcpy(stream_0_data_->data(), buf->data(), buf_len);
+    data_size_[0] = buf_len;
+  } else {
+    RecordUnexpectedStream0Write(cache_type_);
+    int buffer_size;
+    if (truncate)
+      buffer_size = offset + buf_len;
+    else
+      buffer_size = std::max(offset + buf_len, data_size_[0]);
+    scoped_refptr<IOBuffer> new_data = new IOBuffer(buffer_size);
+    memset(new_data->data(), 0, buffer_size);
+    if (stream_0_data_) {
+      memcpy(new_data->data(),
+             stream_0_data_->data(),
+             std::min(data_size_[0], buffer_size));
+    }
+    if (buf)
+      memcpy(new_data->data() + offset, buf->data(), buf_len);
+    stream_0_data_ = new_data;
+    data_size_[0] = buffer_size;
+  }
+  base::Time modification_time = base::Time::Now();
+  UpdateDataFromEntryStat(
+      SimpleEntryStat(modification_time, modification_time, data_size_));
+  if (stream_0_data_) {
+    crc32s_[0] = crc32(crc32(0L, Z_NULL, 0),
+                       reinterpret_cast<const Bytef*>(stream_0_data_->data()),
+                       data_size_[0]);
+  } else {
+    crc32s_[0] = crc32(0L, Z_NULL, 0);
+  }
+  RecordWriteResult(cache_type_, WRITE_RESULT_SUCCESS);
+  return buf_len;
+}
+
 }  // namespace disk_cache