[Android] Introduce libheap_profiler for memory profiling.

tools/android/heap_profiler/heap_profiler_integrationtest.cc

Index: tools/android/heap_profiler/heap_profiler_integrationtest.cc
diff --git a/tools/android/heap_profiler/heap_profiler_integrationtest.cc b/tools/android/heap_profiler/heap_profiler_integrationtest.cc
new file mode 100644
index 0000000000000000000000000000000000000000..e3754ebe11db8fa3019bf4df879381b33bcb9722
--- /dev/null
+++ b/tools/android/heap_profiler/heap_profiler_integrationtest.cc
@@ -0,0 +1,211 @@
+// Copyright 2014 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 <dlfcn.h>
+#include <fcntl.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <unistd.h>
+#include <map>
+
+#include "testing/gtest/include/gtest/gtest.h"
+#include "tools/android/heap_profiler/heap_profiler.h"
+
+
+namespace {
+
+// The purpose of the four functions below is to create watermarked allocations,
+// so the test fixture can ascertain that the hooks work end-to-end.
+__attribute__ ((noinline))
+void* MallocInner(size_t size) {
+  void* ptr = malloc(size);
+  // The memset below is to avoid tail-call elimination optimizations and ensure
+  // that this function will be part of the stack trace.

[pasko, 2014/06/25 16:39:50]

this trick is not done for DoMmap, is that for some particular reason or just
forgotten?

[Primiano Tucci (use gerrit), 2014/06/26 09:00:54]

the *trick* is essentially a way to avoid that the function loses the stack
frame, which I definitely don't want for the outer/inner functions.
I don't care, instead, if DoMmap loses the stack frame due to tail call optim (I
never look for the presence of DoMmap in the stack frames).
At this point you might wonder: why did you introduce DoMmap at all so? It's
simple: if I do the mmap() directly inside MallocInner, the generated code is
large (because of all the arguments passed to mmap) and the 16-byte offset
assumpion will not hold on all platforms. ;-)

+  memset(ptr, 0, size);
+  return ptr;
+}
+
+__attribute__ ((noinline))
+void* MallocOuter(size_t size) {
+  void* ptr = MallocInner(size);
+  memset(ptr, 0, size);
+  return ptr;
+}
+
+__attribute__ ((noinline))
+void* DoMmap(size_t size) {
+  return mmap(
+      0, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
+}
+
+__attribute__ ((noinline))
+void* MmapInner(size_t size) {
+  void* ptr = DoMmap(size);
+  memset(ptr, 0, size);
+  return ptr;
+}
+
+__attribute__ ((noinline))
+void* MmapOuter(size_t size) {
+  void* ptr = MmapInner(size);
+  memset(ptr, 0, size);
+  return ptr;
+}
+
+bool StackTraceContains(const StacktraceEntry* s, void* (*fn)(size_t)) {
+  // kExpectedFnLength is a gross estimation of the size of the Step* functions
+  // above. It tries to address the following problem: the addrs in the unwound
+  // stack frames will NOT point to the beginning of the functions, but to the
+  // PC after the call to malloc/mmap.
+  const size_t kExpectedFnLength = 16;
+  const uintptr_t fn_addr = reinterpret_cast<uintptr_t>(fn);
+
+  for (size_t i = 0; i < HEAP_PROFILER_MAX_DEPTH; ++i) {
+    if (s->frames[i] >= fn_addr && s->frames[i] <= fn_addr + kExpectedFnLength)
+      return true;
+  }
+  return false;
+}
+
+const HeapStats* GetHeapProfilerStats() {
+  HeapStats* const* stats_ptr = reinterpret_cast<HeapStats* const*>(
+      dlsym(RTLD_DEFAULT, "heap_profiler_stats_for_tests"));
+  EXPECT_TRUE(stats_ptr != NULL);
+  const HeapStats* stats = *stats_ptr;
+  EXPECT_TRUE(stats != NULL);
+  EXPECT_EQ(HEAP_PROFILER_MAGIC_MARKER, stats->magic_start);
+  return stats;
+}
+
+const StacktraceEntry* LookupStackTraceBySize(size_t size) {

[pasko, 2014/06/25 16:39:50]

maybe search by size _and_ a function to filter out spurious matches, possible
causes of flake?

[Primiano Tucci (use gerrit), 2014/06/26 09:00:54]

Done.

+  const HeapStats* stats = GetHeapProfilerStats();
+  for (size_t i = 0; i < stats->max_stack_traces; ++i) {
+    const StacktraceEntry* st = &stats->stack_traces[i];
+    if (st->alloc_bytes == size)
+      return st;
+  }
+  return NULL;
+}
+
+TEST(HeapProfilerIntegrationTest, TestMallocStackTraces) {
+  const HeapStats* stats = GetHeapProfilerStats();
+
+  const size_t kSize1 = 1000000;
+  const size_t kSize2 = 2000000;
+  const size_t kSize3 = 4000000;
+
+  void* m1 = MallocOuter(kSize1);
+  void* m2 = MallocInner(kSize2);
+  void* m3 = MallocInner(kSize3);
+
+  ASSERT_TRUE(stats->stack_traces != NULL);
+
+  free(m3);
+
+  const StacktraceEntry* st1 = LookupStackTraceBySize(kSize1);
+  const StacktraceEntry* st2 = LookupStackTraceBySize(kSize2);
+  const StacktraceEntry* st3 = LookupStackTraceBySize(kSize3);
+
+  EXPECT_TRUE(st1 != NULL);
+  EXPECT_TRUE(StackTraceContains(st1, &MallocOuter));
+  EXPECT_TRUE(StackTraceContains(st1, &MallocInner));
+
+  EXPECT_TRUE(st2 != NULL);
+  EXPECT_FALSE(StackTraceContains(st2, &MallocOuter));
+  EXPECT_TRUE(StackTraceContains(st2, &MallocInner));
+
+  EXPECT_EQ(NULL, st3);
+
+  const size_t total_alloc_start = stats->total_alloc_bytes;
+  const size_t num_stack_traces_start = stats->num_stack_traces;
+  free(m1);
+  free(m2);
+  const size_t total_alloc_end = stats->total_alloc_bytes;
+  const size_t num_stack_traces_end = stats->num_stack_traces;
+
+  EXPECT_EQ(kSize1 + kSize2, total_alloc_start - total_alloc_end);
+  EXPECT_EQ(2, num_stack_traces_start - num_stack_traces_end);
+  EXPECT_EQ(NULL, LookupStackTraceBySize(kSize1));
+  EXPECT_EQ(NULL, LookupStackTraceBySize(kSize2));
+  EXPECT_EQ(NULL, LookupStackTraceBySize(kSize3));
+}
+
+TEST(HeapProfilerIntegrationTest, TestMmapStackTraces) {

[bulach, 2014/06/25 11:19:33]

nit: I still think it'd be clearer to have one inner function taking the sizes /
allocators / free'ers as params, the whole body is *exactly* the same except for
those params..

[Primiano Tucci (use gerrit), 2014/06/26 09:00:54]

Ok, what about now?

+  const HeapStats* stats = GetHeapProfilerStats();
+
+  static const size_t kSize1 = 499 * PAGE_SIZE;
+  static const size_t kSize2 = 503 * PAGE_SIZE;
+  static const size_t kSize3 = 509 * PAGE_SIZE;
+
+  void* m1 = MmapOuter(kSize1);
+  void* m2 = MmapInner(kSize2);
+  void* m3 = MmapInner(kSize3);
+
+  ASSERT_TRUE(stats->stack_traces != NULL);
+
+  munmap(m3, kSize3);
+
+  const StacktraceEntry* st1 = LookupStackTraceBySize(kSize1);
+  const StacktraceEntry* st2 = LookupStackTraceBySize(kSize2);
+  const StacktraceEntry* st3 = LookupStackTraceBySize(kSize3);
+
+  EXPECT_TRUE(st1 != NULL);
+  EXPECT_TRUE(StackTraceContains(st1, &MmapOuter));
+  EXPECT_TRUE(StackTraceContains(st1, &MmapInner));
+
+  EXPECT_TRUE(st2 != NULL);
+  EXPECT_FALSE(StackTraceContains(st2, &MmapOuter));
+  EXPECT_TRUE(StackTraceContains(st2, &MmapInner));
+
+  EXPECT_EQ(NULL, st3);
+
+  const size_t total_alloc_start = stats->total_alloc_bytes;
+  const size_t num_stack_traces_start = stats->num_stack_traces;
+  munmap(m1, kSize1);
+  munmap(m2, kSize2);
+  const size_t total_alloc_end = stats->total_alloc_bytes;
+  const size_t num_stack_traces_end = stats->num_stack_traces;
+
+  EXPECT_EQ(kSize1 + kSize2, total_alloc_start - total_alloc_end);
+  EXPECT_EQ(2, num_stack_traces_start - num_stack_traces_end);
+  EXPECT_EQ(NULL, LookupStackTraceBySize(kSize1));
+  EXPECT_EQ(NULL, LookupStackTraceBySize(kSize2));
+  EXPECT_EQ(NULL, LookupStackTraceBySize(kSize3));
+}
+
+// Returns the path of the directory containing the current executable.
+std::string GetExePath() {
+  char buf[1024];
+  ssize_t len = readlink("/proc/self/exe", buf, sizeof(buf) - 1);
+  if (len == -1)
+    return std::string();
+  std::string path(buf, len);
+  size_t sep = path.find_last_of('/');
+  if (sep == std::string::npos)
+    return std::string();
+  path.erase(sep);
+  return path;
+}
+
+}  // namespace
+
+
+int main(int argc, char** argv) {
+  // Re-launch the process itself forcing the preload of the libheap_profiler.
+  if (getenv("LD_PRELOAD") == NULL) {
+    char env_ld_lib_path[256];
+    strlcpy(env_ld_lib_path, "LD_LIBRARY_PATH=", sizeof(env_ld_lib_path));
+    strlcat(env_ld_lib_path, GetExePath().c_str(), sizeof(env_ld_lib_path));
+    char env_ld_preload[] = "LD_PRELOAD=libheap_profiler.so";
+    char* const env[] = {env_ld_preload, env_ld_lib_path, 0};
+    execve("/proc/self/exe", argv, env);
+    // execve() never returns, unless something goes wrong.
+    perror("execve");
+    assert(false);
+  }
+
+  testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}