Bitmap transport

This patch reworks bitmap transport on all platforms. Linux and Mac
are switched from serialising bitmaps over the IPC channel to using
shared memory. All platforms gain a shared memory mapping cache on the
host side.

The concept of a TransportDIB (device independent bitmap) is added to
encapsulate most of the platform specifics.

On Linux, we use SysV shared memory. This is because X shared pixmaps,
which predate POSIX SHM, can only use SysV. By using SysV between
renderer and browser, we open up the possibility to map the shared
memory directly from the renderer to the X server.

On Mac, we use POSIX shared memory. However, since this needs
filesystem access and the Mac renderer is sandboxed from the
filesystem, we add two new messages from renderer -> browser:

The first, AllocTransportDIB, synchronously creates a transport DIB in
the browser and passes a handle back to the renderer. The second,
FreeTransportDIB, asynchronously, notifies the browser that it may
close its handle to the shared memory region.

On Mac, the shared memory regions are identified by their inode
numbers on the wire. This means that the browser must keep handles
open to all the allocated shared memory regions (since an inode number
is insufficient to map the region). The alternative design is that the
renderer passes the file descriptor with each paint operation. Since
passing file descriptors is special case in the code, I felt that it
would be best to minimise their use. Creating and freeing transport
DIBs are relatively rare operations relative to paints and scrolls.

On Windows, most of the code remains the same, except that Windows now
uses the mapping cache added in this patch. This allows the browser to
maintain a shared memory mapping for a transport DIB over several
paints. Previously it mapped and unmapped for every operation, causing
lots of TLB and VM churn.

Review URL: http://codereview.chromium.org/21485


git-svn-id: svn://svn.chromium.org/chrome/trunk/src@10071 0039d316-1c4b-4281-b951-d872f2087c98

1 files changed, 15 insertions, 2 deletions

diff --git a/base/shared_memory.h b/base/shared_memory.h
index b44367a..744b348 100644
--- a/base/shared_memory.h
+++ b/base/shared_memory.h
@@ -8,6 +8,7 @@
 #include "build/build_config.h"
 
 #if defined(OS_POSIX)
+#include <sys/types.h>
 #include <semaphore.h>
 #include "base/file_descriptor_posix.h"
 #endif
@@ -24,7 +25,10 @@ namespace base {
 typedef HANDLE SharedMemoryHandle;
 typedef HANDLE SharedMemoryLock;
 #elif defined(OS_POSIX)
+// A SharedMemoryId is sufficient to identify a given shared memory segment on a
+// system, but insufficient to map it.
 typedef FileDescriptor SharedMemoryHandle;
+typedef ino_t SharedMemoryId;
 // On POSIX, the lock is implemented as a lockf() on the mapped file,
 // so no additional member (or definition of SharedMemoryLock) is
 // needed.
@@ -99,6 +103,14 @@ class SharedMemory {
   // identifier is not portable.
   SharedMemoryHandle handle() const;
 
+#if defined(OS_POSIX)
+  // Return a unique identifier for this shared memory segment. Inode numbers
+  // are technically only unique to a single filesystem. However, we always
+  // allocate shared memory backing files from the same directory, so will end
+  // up on the same filesystem.
+  SharedMemoryId id() const { return inode_; }
+#endif
+
   // Closes the open shared memory segment.
   // It is safe to call Close repeatedly.
   void Close();
@@ -153,9 +165,10 @@ class SharedMemory {
 
   std::wstring       name_;
 #if defined(OS_WIN)
-  HANDLE mapped_file_;
+  HANDLE             mapped_file_;
 #elif defined(OS_POSIX)
-  int mapped_file_;
+  int                mapped_file_;
+  ino_t              inode_;
 #endif
   void*              memory_;
   bool               read_only_;