SimpleCache: merge the first and second stream in one file

net/disk_cache/simple/simple_entry_impl.cc

 // Copyright (c) 2013 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include "net/disk_cache/simple/simple_entry_impl.h"
 
 #include <algorithm>
 #include <cstring>
 #include <vector>
 
@@ skipping 317 matching lines @@
                               int buf_len,
                               const CompletionCallback& callback) {
   DCHECK(io_thread_checker_.CalledOnValidThread());
 
   if (net_log_.IsLoggingAllEvents()) {
     net_log_.AddEvent(net::NetLog::TYPE_SIMPLE_CACHE_ENTRY_READ_CALL,
         CreateNetLogReadWriteDataCallback(stream_index, offset, buf_len,
                                           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,
           CreateNetLogReadWriteCompleteCallback(net::ERR_INVALID_ARGUMENT));
     }
 
     RecordReadResult(cache_type_, READ_RESULT_INVALID_ARGUMENT);
     return net::ERR_INVALID_ARGUMENT;
   }
   if (pending_operations_.empty() && (offset >= GetDataSize(stream_index) ||
@@ skipping 25 matching lines @@
                                bool truncate) {
   DCHECK(io_thread_checker_.CalledOnValidThread());
 
   if (net_log_.IsLoggingAllEvents()) {
     net_log_.AddEvent(
         net::NetLog::TYPE_SIMPLE_CACHE_ENTRY_WRITE_CALL,
         CreateNetLogReadWriteDataCallback(stream_index, offset, buf_len,
                                           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,
           CreateNetLogReadWriteCompleteCallback(net::ERR_INVALID_ARGUMENT));
     }
     RecordWriteResult(cache_type_, WRITE_RESULT_INVALID_ARGUMENT);
     return net::ERR_INVALID_ARGUMENT;
   }
   if (backend_.get() && offset + buf_len > backend_->GetMaxFileSize()) {
     if (net_log_.IsLoggingAllEvents()) {
       net_log_.AddEvent(
           net::NetLog::TYPE_SIMPLE_CACHE_ENTRY_WRITE_END,
           CreateNetLogReadWriteCompleteCallback(net::ERR_FAILED));
     }
     RecordWriteResult(cache_type_, WRITE_RESULT_OVER_MAX_SIZE);
     return net::ERR_FAILED;
   }
   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
   // actually run the write operation that sets the stream size. It also
   // prevents from previous possibly-conflicting writes that could be stacked
   // in the |pending_operations_|. We could optimize this for when we have
   // only read operations enqueued.
   const bool optimistic =
       (use_optimistic_operations_ && state_ == STATE_READY &&
        pending_operations_.size() == 0);
@@ skipping 244 matching lines @@
   DCHECK_EQ(STATE_UNINITIALIZED, state_);
   DCHECK(!synchronous_entry_);
 
   state_ = STATE_IO_PENDING;
 
   // Since we don't know the correct values for |last_used_| and
   // |last_modified_| yet, we make this approximation.
   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();
   scoped_ptr<SimpleEntryCreationResults> results(
       new SimpleEntryCreationResults(
           SimpleEntryStat(last_used_, last_modified_, data_size_)));
   Closure task = base::Bind(&SimpleSynchronousEntry::CreateEntry,
                             cache_type_,
                             path_,
                             key_,
@@ skipping 14 matching lines @@
   DCHECK(io_thread_checker_.CalledOnValidThread());
   typedef SimpleSynchronousEntry::CRCRecord CRCRecord;
   scoped_ptr<std::vector<CRCRecord> >
       crc32s_to_write(new std::vector<CRCRecord>());
 
   net_log_.AddEvent(net::NetLog::TYPE_SIMPLE_CACHE_ENTRY_CLOSE_BEGIN);
 
   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];
           crc32s_to_write->push_back(CRCRecord(i, true, crc));
         } else {
           crc32s_to_write->push_back(CRCRecord(i, false, 0));
         }
       }
     }
   } else {
     DCHECK(STATE_UNINITIALIZED == state_ || STATE_FAILURE == state_);
   }
 
   if (synchronous_entry_) {
     Closure task =
         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_,
                          crc_check_state_[i], CRC_CHECK_MAX);
       }
     }
   } else {
     synchronous_entry_ = NULL;
     CloseOperationComplete();
   }
@@ skipping 31 matching lines @@
   if (offset >= GetDataSize(stream_index) || offset < 0 || !buf_len) {
     RecordReadResult(cache_type_, READ_RESULT_FAST_EMPTY_RETURN);
     // If there is nothing to read, we bail out before setting state_ to
     // STATE_IO_PENDING.
     if (!callback.is_null())
       MessageLoopProxy::current()->PostTask(FROM_HERE, base::Bind(
           callback, 0));
     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;
   if (backend_.get())
     backend_->index()->UseIfExists(entry_hash_);
 
   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,
                              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);
 }
 
 void SimpleEntryImpl::WriteDataInternal(int stream_index,
                                        int offset,
                                        net::IOBuffer* buf,
                                        int buf_len,
                                        const CompletionCallback& callback,
                                        bool truncate) {
@@ skipping 17 matching lines @@
     if (!callback.is_null()) {
       // We need to posttask so that we don't go in a loop when we call the
       // callback directly.
       MessageLoopProxy::current()->PostTask(FROM_HERE, base::Bind(
           callback, net::ERR_FAILED));
     }
     // |this| may be destroyed after return here.
     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())
     backend_->index()->UseIfExists(entry_hash_);
   // It is easy to incrementally compute the CRC from [0 .. |offset + buf_len|)
   // if |offset == 0| or we have already computed the CRC for [0 .. offset).
   // We rely on most write operations being sequential, start to end to compute
   // the crc of the data. When we write to an entry and close without having
   // done a sequential write, we don't check the CRC on read.
   if (offset == 0 || crc32s_end_offset_[stream_index] == offset) {
     uint32 initial_crc = (offset != 0) ? crc32s_[stream_index]
@@ skipping 14 matching lines @@
   } else {
     data_size_[stream_index] = std::max(offset + buf_len,
                                         GetDataSize(stream_index));
   }
 
   // Since we don't know the correct values for |last_used_| and
   // |last_modified_| yet, we make this approximation.
   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,
                             base::Unretained(synchronous_entry_),
                             SimpleSynchronousEntry::EntryOperationData(
                                 stream_index, offset, buf_len, truncate),
                             make_scoped_refptr(buf),
                             entry_stat.get(),
                             result.get());
   Closure reply = base::Bind(&SimpleEntryImpl::WriteOperationComplete,
@@ skipping 31 matching lines @@
     MakeUninitialized();
     return;
   }
   // If out_entry is NULL, it means we already called ReturnEntryToCaller from
   // the optimistic Create case.
   if (out_entry)
     ReturnEntryToCaller(out_entry);
 
   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 {
     // This should only be triggered when creating an entry. The key check in
     // the open case is handled in SimpleBackendImpl.
     DCHECK_EQ(key_, synchronous_entry_->key());
   }
   UpdateDataFromEntryStat(in_results->entry_stat);
   SIMPLE_CACHE_UMA(TIMES,
                    "EntryCreationTime", cache_type_,
@@ skipping 30 matching lines @@
         completion_callback, *result));
   }
   RunNextOperationIfNeeded();
 }
 
 void SimpleEntryImpl::ReadOperationComplete(
     int stream_index,
     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_);
   DCHECK_EQ(STATE_IO_PENDING, state_);
   DCHECK(read_crc32);
   DCHECK(result);
 
   if (*result > 0 &&
       crc_check_state_[stream_index] == CRC_CHECK_NEVER_READ_AT_ALL) {
     crc_check_state_[stream_index] = CRC_CHECK_NEVER_READ_TO_END;
@@ skipping 15 matching lines @@
       // entry, one reader can be behind the other. In this case we compute
       // the crc as the most advanced reader progresses, and check it for
       // both readers as they read the last byte.
 
       net_log_.AddEvent(net::NetLog::TYPE_SIMPLE_CACHE_ENTRY_CHECKSUM_BEGIN);
 
       scoped_ptr<int> new_result(new int());
       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,
                                  this, *result, stream_index,
                                  completion_callback,
                                  base::Passed(&new_result));
       worker_pool_->PostTaskAndReply(FROM_HERE, task, reply);
       crc_check_state_[stream_index] = CRC_CHECK_DONE;
       return;
     }
   }
 
   if (*result < 0) {
     RecordReadResult(cache_type_, READ_RESULT_SYNC_READ_FAILURE);
   } else {
     RecordReadResult(cache_type_, READ_RESULT_SUCCESS);
     if (crc_check_state_[stream_index] == CRC_CHECK_NEVER_READ_TO_END &&
         offset + *result == GetDataSize(stream_index)) {
       crc_check_state_[stream_index] = CRC_CHECK_NOT_DONE;
     }
   }
   if (net_log_.IsLoggingAllEvents()) {
     net_log_.AddEvent(
         net::NetLog::TYPE_SIMPLE_CACHE_ENTRY_READ_END,
         CreateNetLogReadWriteCompleteCallback(*result));
   }
 
   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(
     int stream_index,
     const CompletionCallback& completion_callback,
     scoped_ptr<SimpleEntryStat> entry_stat,
     scoped_ptr<int> result) {
   if (*result >= 0)
     RecordWriteResult(cache_type_, WRITE_RESULT_SUCCESS);
   else
@@ skipping 56 matching lines @@
 }
 
 void SimpleEntryImpl::UpdateDataFromEntryStat(
     const SimpleEntryStat& entry_stat) {
   DCHECK(io_thread_checker_.CalledOnValidThread());
   DCHECK(synchronous_entry_);
   DCHECK_EQ(STATE_READY, state_);
 
   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())
     backend_->index()->UpdateEntrySize(entry_hash_, GetDiskUsage());
 }
 
 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]);
   }
   return file_size;
 }
 
 void SimpleEntryImpl::RecordReadIsParallelizable(
     const SimpleEntryOperation& operation) const {
   if (!executing_operation_)
     return;
@@ skipping 57 matching lines @@
     } else {
       type = conflicting ? WRITE_FOLLOWS_CONFLICTING_WRITE
                          : WRITE_FOLLOWS_NON_CONFLICTING_WRITE;
     }
   }
   SIMPLE_CACHE_UMA(ENUMERATION,
                    "WriteDependencyType", cache_type_,
                    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