Enforce a memory limit on MappedMemoryManager

gpu/command_buffer/client/mapped_memory_unittest.cc

Index: gpu/command_buffer/client/mapped_memory_unittest.cc
diff --git a/gpu/command_buffer/client/mapped_memory_unittest.cc b/gpu/command_buffer/client/mapped_memory_unittest.cc
index 99e6e8d3ec27eea55b7c42fbb3cd2b4f160fd55a..4ba1337af9028be4f3afd81b841f094f70ec7a36 100644
--- a/gpu/command_buffer/client/mapped_memory_unittest.cc
+++ b/gpu/command_buffer/client/mapped_memory_unittest.cc
@@ -306,4 +306,37 @@ TEST_F(MappedMemoryManagerTest, ChunkSizeMultiple) {
   EXPECT_EQ(0u, offset3);
 }
 
+TEST_F(MappedMemoryManagerTest, MemoryLimit) {
+  const unsigned int kSize = 1024;
+  // Reset the manager with a memory limit.
+  manager_.reset(new MappedMemoryManager(helper_.get(), 2 * kSize));
+  manager_->set_chunk_size_multiple(2 * kSize);
+
+  // Allocate half a chunk worth of memory.
+  int32 id1 = -1;
+  unsigned int offset1 = 0xFFFFFFFFU;
+  void* mem1 = manager_->Alloc(kSize, &id1, &offset1);
+  ASSERT_TRUE(mem1);
+  EXPECT_NE(-1, id1);
+  EXPECT_EQ(0u, offset1);
+
+  // Allocate half a chunk worth of memory again.
+  // The same chunk will be used.
+  int32 id2 = -1;
+  unsigned int offset2 = 0xFFFFFFFFU;
+  void* mem2 = manager_->Alloc(kSize, &id2, &offset2);
+  ASSERT_TRUE(mem2);
+  EXPECT_NE(-1, id2);
+  EXPECT_EQ(kSize, offset2);
+
+  // Allocate half a chunk worth of memory. Since we reached
+  // the memory limit we'll wait tokens on existing chunks.

[piman, 2013/08/15 02:49:03]

Wait for what token? You didn't free anything pending token.

[kaanb, 2013/08/16 22:50:44]

Done.

+  int32 id3 = -1;
+  unsigned int offset3 = 0xFFFFFFFFU;
+  void* mem3 = manager_->Alloc(kSize, &id3, &offset3);
+  ASSERT_TRUE(mem3);

[no sievers, 2013/08/15 01:43:42]

I think this test shows that your patch does not necessarily work as expected.
I assume it's allocating again here instead of reusing.

What you are missing is the case where the mapped memory is IN_USE.

I'd keep this test to specify what the behavior should be for the corresponding
calling sequence (multiple Map calls without Unmap).

Maybe add another that mimicks the Map/Unmap, Map/Unmap, .. pattern. You would
explicitly have to insert the token and eventually signal it, see
FreePendingToken test above.

[no sievers, 2013/08/15 01:49:06]

Btw, which usage case are you trying to optimize again?
async uploads or not?

[kaanb, 2013/08/16 22:50:44]

Done.

[kaanb, 2013/08/16 22:50:44]

I think this optimization is oblivious to whether we use async uploads or not,
right?

+  EXPECT_NE(-1, id3);
+  EXPECT_EQ(0u, offset3);
+}
+
 }  // namespace gpu