[Mac] Add a perf test that counts the number of Mach ports in the browser.

chrome/test/perf/mach_ports_test.cc

+// Copyright (c) 2012 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 <vector>
+
+#include "base/memory/ref_counted.h"
+#include "base/string_util.h"
+#include "chrome/common/automation_constants.h"
+#include "chrome/common/automation_messages.h"
+#include "chrome/test/automation/automation_proxy.h"
+#include "chrome/test/automation/browser_proxy.h"
+#include "chrome/test/automation/tab_proxy.h"
+#include "chrome/test/automation/window_proxy.h"
+#include "chrome/test/perf/perf_test.h"
+#include "chrome/test/ui/ui_perf_test.h"
+#include "net/test/test_server.h"
+
+namespace {
+
+// This test spawns a new browser and counts the number of open Mach ports in
+// the browser process. It navigates tabs and closes them, repeatedly measuring
+// the number of open ports. This is used to protect against leaking Mach ports,
+// which was the source of <http://crbug.com/105513>.
+class MachPortsTest : public UIPerfTest {
+ public:
+  MachPortsTest()
+      : server_(net::TestServer::TYPE_HTTP,
+                FilePath(FILE_PATH_LITERAL("data/page_cycler/moz"))) {

[cmp, 2012/01/23 21:20:51]

Please create a new data set that's named to match your test.  Feel free to base
it on data/page_cycler/moz.  Please keep it in data/mach_ports/moz, too, in the
internal repo.

[Robert Sesek, 2012/01/23 21:58:41]

Really? Why do I need a separate data set if this has exactly what I want (some
decently complex sample webpages)?

[cmp, 2012/01/23 22:20:56]

Sure, it would technically be a separate data set, but really it's just a fresh
copy of the moz data set.  My concern is that your test can be impacted by
changes to the moz data set that are related only to page cycler changes, and
that those changes will be unexpected by the person making them.  By copying the
data to a new location just for this test, this test will own that data and
there won't be any concern that some unrelated change breaks it.

Have you tried other data sets, like morejs?  morejs is even more complex than
moz, and that may be desirable.

[Robert Sesek, 2012/01/27 19:10:41]

Added a new data set per internal CL.

+  }
+
+  virtual void SetUp() OVERRIDE {

[cmp, 2012/01/23 21:20:51]

I believe SetUp and TearDown can both be defined as void NAME as opposed to
virtual void NAME OVERRIDE.  I would follow our current conventions for this
(see chrome/test/perf/*_test.cc, eg) where possible.

[Robert Sesek, 2012/01/23 21:58:41]

In UITest, they're defined as OVERRIDE.

[cmp, 2012/01/23 22:20:56]

cmp@terakrom:/work/chrome/trunk/src/chrome/test/perf$ git gs OVERRIDE
browser_perf_test.h:20:  virtual void SetUpCommandLine(CommandLine*
command_line) OVERRIDE;
rendering/latency_tests.cc:121:  virtual void SetUpCommandLine(CommandLine*
command_line) OVERRIDE;
rendering/throughput_tests.cc:40:  virtual void SetUpCommandLine(CommandLine*
command_line) OVERRIDE {

I am uncertain why all of these tests were written without that modifier, and
some docs confirm that the "overridden base method must have the same signature
as the virtual method."  I suspect g++ (for example) will just ignore the lack
of the virtual+OVERRIDE modifiers, but I'm guessing.  Mark, do you understand if
the tests in chrome/test/perf/ are doing something wrong as-written?

[Robert Sesek, 2012/01/27 19:10:41]

The modifier is rather new, so old code may not have it. New code should,
though.

+    UIPerfTest::SetUp();
+
+    ASSERT_TRUE(server_.Start());
+  }
+
+  virtual void TearDown() OVERRIDE {
+    std::string ports;
+    for (std::vector<int>::iterator it = port_counts_.begin();
+         it != port_counts_.end(); ++it) {
+      base::StringAppendF(&ports, "%d,", *it);
+    }
+    perf_test::PrintResultList("mach_ports", "", "", ports, "ports", true);
+
+    UIPerfTest::TearDown();
+  }
+
+  // Gets the browser's current number of Mach ports and records it.
+  void RecordPortCount() {
+    int port_count = 0;
+    ASSERT_TRUE(automation()->Send(
+        new AutomationMsg_GetMachPortCount(&port_count)));
+    port_counts_.push_back(port_count);
+  }
+
+  // Adds a tab from the page cycler data at the specified domain.
+  bool AddTab(scoped_refptr<BrowserProxy> browser, const std::string& domain) {
+    GURL url = server_.GetURL("files/" + domain + "/").Resolve("?skip");
+    return browser->AppendTab(url);
+  }
+
+ private:
+  net::TestServer server_;
+  std::vector<int> port_counts_;
+};
+
+TEST_F(MachPortsTest, GetCounts) {
+  ASSERT_EQ(AUTOMATION_SUCCESS, automation()->WaitForAppLaunch());
+
+  // Record startup number.
+  RecordPortCount();
+
+  // Create a browser and open a few tabs.
+  scoped_refptr<WindowProxy> window(automation()->GetActiveWindow());
+  ASSERT_TRUE(window.get());
+
+  scoped_refptr<BrowserProxy> browser(automation()->FindTabbedBrowserWindow());
+  ASSERT_TRUE(browser.get());
+
+  EXPECT_TRUE(AddTab(browser, "www.google.com"));
+  RecordPortCount();
+
+  EXPECT_TRUE(AddTab(browser, "www.cnn.com"));
+  RecordPortCount();
+
+  EXPECT_TRUE(AddTab(browser, "www.nytimes.com"));
+  RecordPortCount();
+
+  int tab_count = 0;
+  ASSERT_TRUE(browser->GetTabCount(&tab_count));
+  EXPECT_EQ(4, tab_count);  // Also count about:blank.
+
+  // Close each tab, recording the number of ports after each. Do not close the
+  // last tab, which is about:blank because it will be closed by the proxy.
+  for (int i = 0; i < tab_count - 1; ++i) {
+    scoped_refptr<TabProxy> tab(browser->GetActiveTab());
+    ASSERT_TRUE(tab.get());
+
+    EXPECT_TRUE(tab->Close(true));
+    RecordPortCount();
+  }

[cmp, 2012/01/23 21:20:51]

Can the values returned by this test vary across test runs in copies of Chrome
from the same source tree at the same revision?  If so, do they vary much under
system load?

Most of our tests define a RunTest method which allows us to pass a different
class name into the fixture.  We use this to allow creating a fixture that uses
the reference build.

If you need a reference build version of this test, then please move the
majority of 69-102 into a RunTest method in MachPortsTest, then call the RunTest
method from this fixture.  Let me know if you have any questions about that.

[Robert Sesek, 2012/01/23 21:58:41]

The values will vary slightly, but the mean (what I will end up measuring) is
pretty consistent in my own testing. There will be differences in the numbers
(of statistical significance) across OS versions, though.

[cmp, 2012/01/23 22:20:56]

Based on this, I will assume you won't need any reference builds running your
test.  That would be great!

[Robert Sesek, 2012/01/27 19:10:41]

I think that's correct, yes.

+}
+
+}  // namespace