A deeper heap profile dumper in third_party/tcmalloc/chromium.

third_party/tcmalloc/chromium/src/deep-heap-profile.cc

Index: third_party/tcmalloc/chromium/src/deep-heap-profile.cc
diff --git a/third_party/tcmalloc/chromium/src/deep-heap-profile.cc b/third_party/tcmalloc/chromium/src/deep-heap-profile.cc
new file mode 100644
index 0000000000000000000000000000000000000000..ed8bf9369138432b7d98440ebf9775670169d289
--- /dev/null
+++ b/third_party/tcmalloc/chromium/src/deep-heap-profile.cc
@@ -0,0 +1,588 @@
+// Copyright (c) 2011 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.
+
+// ---
+// Author: Sainbayar Sukhbaatar
+//         Dai Mikurube
+//
+
+#include "deep-heap-profile.h"
+
+#ifdef DEEP_HEAP_PROFILE
+#include <fcntl.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#ifdef HAVE_UNISTD_H
+#include <unistd.h>  // for getpid()
+#endif

[jar (doing other things), 2012/01/27 00:46:58]

For endif, please add comment showing what is terminated:
#endif   // HAVE_UNISTD_H

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Done.

+
+#include "base/cycleclock.h"
+#include "base/sysinfo.h"
+
+static const int kProfilerBufferSize = 1 << 20;
+static const int kHashTableSize = 179999;  // The same as heap-profile-table.cc.
+
+static const int PAGE_SIZE = 4096;

[jar (doing other things), 2012/01/27 00:46:58]

It surprises me that we don't have the moral equivalent to this elsewhere in
TCMalloc.  For instance, why aren't you using getpagesize()?

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

getpagesize() is better.  Fixed.

+static const int PAGEMAP_BYTES = 8;
+static const uint64 TOP_ADDRESS = kuint64max;

[jar (doing other things), 2012/01/27 00:46:58]

Why do you use uint64 for addresses, when size_t is defined to handle this
nicely? (and not waste space and performance on a 32bit OS).

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Simply, its' sbecause ProcMapsIterator::Next(Ext) (base/sysinfo.h) uses uint64
for addresses even if 32-bits.  One reason is written in a comment for
ProcMapsIterator::Next, and another reason is mixing uint64 and size_t is
confusing.

No uint64 values are stored in DeepHeapProfile.  They're used only in stacks or
as constants.  I think it's not a problem.

+
+// Header strings of the dumped heap profile.
+static const char kProfileHeader[] = "heap profile: ";
+static const char kProfileVersion[] = "DUMP_DEEP_3";
+static const char kGlobalStatsHeader[] = "GLOBAL_STATS:\n";
+static const char kMMapStacktraceHeader[] = "MMAP_STACKTRACES:\n";
+static const char kAllocStacktraceHeader[] = "MALLOC_STACKTRACES:\n";
+static const char kProcSelfMapsHeader[] = "\nMAPPED_LIBRARIES:\n";
+
+DeepHeapProfile::DeepHeapProfile(HeapProfileTable* heap_profile,
+                                 const char* prefix)
+    : heap_profile_(heap_profile),
+      pagemap_fd_(-1),
+      most_recent_pid_(-1),
+      stats_(),
+      dump_count_(0),
+      filename_prefix_(NULL),
+      profiler_buffer_(NULL),
+      bucket_id_(0) {
+  deep_bucket_map_ = new(heap_profile_->alloc_(sizeof(DeepBucketMap)))
+      DeepBucketMap(heap_profile_->alloc_, heap_profile_->dealloc_);
+
+  // Copy filename prefix.
+  const int prefix_length = strlen(prefix);
+  filename_prefix_ =
+      reinterpret_cast<char*>(heap_profile_->alloc_(prefix_length + 1));
+  memcpy(filename_prefix_, prefix, prefix_length);

[jar (doing other things), 2012/01/27 00:46:58]

Why not just strcpy(), given that you base this all on strlen in line 52?

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

I guess we thought that memcpy is faster in that case.  strlen is required to
allocate memory.  strcpy checks null-termination again.

+  filename_prefix_[prefix_length] = '\0';
+
+  profiler_buffer_ =
+      reinterpret_cast<char*>(heap_profile_->alloc_(kProfilerBufferSize));
+}
+
+DeepHeapProfile::~DeepHeapProfile() {
+  heap_profile_->dealloc_(profiler_buffer_);
+  heap_profile_->dealloc_(filename_prefix_);
+  deep_bucket_map_->~DeepBucketMap();
+  heap_profile_->dealloc_(deep_bucket_map_);
+}
+
+int DeepHeapProfile::FillOrderedProfile(char buffer[], int buffer_size) {
+#ifndef NDEBUG
+  int64 starting_cycles = CycleClock::Now();
+#endif
+  ++dump_count_;
+
+  // Re-open files in /proc/pid/ if the process is newly forked one.
+  if (most_recent_pid_ != getpid()) {
+    most_recent_pid_ = getpid();
+    pagemap_fd_ = OpenProcPagemap();
+
+    deep_bucket_map_->Iterate(ClearIsLogged, this);
+
+    // Write maps into a .maps file with using the global buffer.
+    WriteMapsToFile(profiler_buffer_, kProfilerBufferSize, filename_prefix_);
+  }
+
+  // Reset committed sizes of buckets.
+  ResetCommittedSize(heap_profile_->alloc_table_);
+  ResetCommittedSize(heap_profile_->mmap_table_);
+
+  GetGlobalStats(pagemap_fd_, &stats_);
+  size_t anonymous_committed = stats_.anonymous.committed_bytes;
+
+  // Note: Don't allocate any memory from here.

[jar (doing other things), 2012/01/27 00:46:58]

Could you motivate the comment?  I'm guessing you're start to snapshot data, and
don't want anything to happen during that snapshot.  When I've implemented such
stuff, I work very hard to make the snapshotting region as small as possible,
which also makes it clear there is compliance.  In this case, you call a pile of
code, and it is really hard to see what is happening where.   It might even be
nice to include the assertion into the function names so that I'd remember abide
by the restriction when I review the called functions.  For instance, perhaps:

RecordAllAllocs() SnapshotAllAllocs()
GetGlobalStats()  SnapshotGlobalStats()

or maybe you should put on a postfix:

RecordAllAllocs --> RecordAllAllocsSansAllocating
GetGlobalStats --> GetGlobalStatsSansAllocating

All I can tell is that it is very hard to review the code and assertions here.

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Renamed snapshotting functions here into Snapshot...WithoutMalloc().

Calling malloc here, actually, doesn't cause violation, but it causes some gap
from actual memory allocation.  I found snprintf may call malloc if large memory
is required internally, but, I decided to ignore it since it's a rare case. 
Exact numbers are not required in our analysis.

(The original heap-profiler looks line ingoring it, too. 
HeapProfileTable::FillOrderedBucket calls HeapProfileTable::UnparseBucket which
calls snprintf while sweeping a bucket table.  Please look at the comment for
HeapProfileTable::UnparseBucket.)

Finally, I modified this comment, renamed snapshotting functions and added TODO.
 What do you think?

+
+  // Record committed sizes.
+  RecordAllAllocs();
+
+  // Check if committed bytes changed during RecordAllAllocs.
+  GetGlobalStats(pagemap_fd_, &stats_);
+#ifndef NDEBUG
+  size_t committed_difference =
+      stats_.anonymous.committed_bytes - anonymous_committed;
+  if (committed_difference != 0) {
+    RAW_LOG(0, "Difference in committed size: %ld", committed_difference);
+  }
+#endif
+
+  HeapProfileTable::Stats stats;

[jar (doing other things), 2012/01/27 00:46:58]

Please move closer to first (real) use.... around line 145.

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Done.

+  memset(&stats, 0, sizeof(stats));
+
+  // Start filling buffer with the ordered profile.
+  int printed = snprintf(buffer, buffer_size,

[jar (doing other things), 2012/01/27 00:46:58]

How are you sure than snprintf() doesn't allocate any memory during performance
of its function?  

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Please look at the above reply.

+                         "%s%s\n", kProfileHeader, kProfileVersion);
+  if (printed < 0 || printed >= buffer_size) {
+    return 0;
+  }
+  int used_in_buffer = printed;
+
+  // Fill buffer with the global stats.
+  printed = snprintf(buffer + used_in_buffer, buffer_size - used_in_buffer,
+                     kGlobalStatsHeader);
+  if (printed < 0 || printed >= buffer_size - used_in_buffer) {
+    return used_in_buffer;
+  }
+  used_in_buffer += printed;
+
+  used_in_buffer = UnparseGlobalStats(buffer, used_in_buffer, buffer_size);
+
+  // Fill buffer with the header for buckets of mmap'ed regions.
+  printed = snprintf(buffer + used_in_buffer, buffer_size - used_in_buffer,
+                     kMMapStacktraceHeader);
+  if (printed < 0 || printed >= buffer_size - used_in_buffer) {
+    return used_in_buffer;
+  }
+  used_in_buffer += printed;
+
+  printed = snprintf(buffer + used_in_buffer, buffer_size - used_in_buffer,
+                     "%10s %10s\n", "virtual", "committed");
+  if (printed < 0 || printed >= buffer_size - used_in_buffer) {
+    return used_in_buffer;
+  }
+  used_in_buffer += printed;
+
+  // Fill buffer with stack trace buckets of mmap'ed regions.
+  used_in_buffer = FillBucketTable(heap_profile_->mmap_table_, buffer,
+                                   buffer_size, used_in_buffer, &stats);

[jar (doing other things), 2012/01/27 00:46:58]

This sure looks like a method on stats, but perhaps that is not code you control
(otherwise it should be made a method).

Maybe it should be a method on a derived class, that you define??

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

HeapProfileTable::Stats is just a public struct (not a class) which contains
four integers.  No additional members are required here.  Why is deriving
required?

But, before all, I found that this stats is not required in the current code. 
Removed it for now.  Maybe added in an extended version in future.

+
+  // Fill buffer with the header for buckets of allocated regions.
+  printed = snprintf(buffer + used_in_buffer, buffer_size - used_in_buffer,
+                     kAllocStacktraceHeader);
+  if (printed < 0 || printed >= buffer_size - used_in_buffer) {
+    return used_in_buffer;
+  }
+  used_in_buffer += printed;
+
+  printed = snprintf(buffer + used_in_buffer, buffer_size - used_in_buffer,
+                     "%10s %10s\n", "virtual", "committed");
+  if (printed < 0 || printed >= buffer_size - used_in_buffer) {
+    return used_in_buffer;
+  }
+  used_in_buffer += printed;
+
+  // Fill buffer with stack trace buckets of allocated regions.
+  used_in_buffer = FillBucketTable(heap_profile_->alloc_table_, buffer,
+                                   buffer_size, used_in_buffer, &stats);
+
+  RAW_DCHECK(used_in_buffer < buffer_size, "");
+
+  // Note: Don't allocate any memory until here.
+
+  // Write the bucket listing into a .bucket file.
+  WriteBucketsToBucketFile();
+
+#ifndef NDEBUG
+  int64 elapsed_cycles = CycleClock::Now() - starting_cycles;
+  double elapsed_seconds = elapsed_cycles / CyclesPerSecond();
+  RAW_LOG(0, "Time spent on DeepProfiler: %.3f sec\n", elapsed_seconds);
+#endif
+
+  return used_in_buffer;
+}
+
+void DeepHeapProfile::RegionStats::Initialize() {
+  virtual_bytes = 0;
+  committed_bytes = 0;
+}
+
+void DeepHeapProfile::RegionStats::Record(
+    int pagemap_fd, uint64 first_address, uint64 last_address) {
+  virtual_bytes += static_cast<size_t>(last_address - first_address + 1);
+  committed_bytes += GetCommittedSize(pagemap_fd, first_address, last_address);
+}
+
+// TODO(dmikurube): Avoid calling ClearIsLogged to rewrite buckets by add a
+// reference to a previous file in a .heap file.
+// static
+void DeepHeapProfile::ClearIsLogged(const void* pointer,
+                                    DeepHeapProfile::DeepBucket* db,
+                                    DeepHeapProfile* deep_profile) {
+  db->is_logged = false;
+}
+
+// static
+int DeepHeapProfile::OpenProcPagemap() {
+  char filename[100];
+  sprintf(filename, "/proc/%d/pagemap", getpid());

[jar (doing other things), 2012/01/27 00:46:58]

use snprintf

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Done.

+  int pagemap_fd = open(filename, O_RDONLY);
+  RAW_DCHECK(pagemap_fd != -1, "Failed to open /proc/self/pagemap");
+  return pagemap_fd;
+}
+
+// static
+bool DeepHeapProfile::SeekProcPagemap(int pagemap_fd, uint64 address) {
+  int64 index = (address / PAGE_SIZE) * PAGEMAP_BYTES;
+  int64 offset = lseek64(pagemap_fd, index, SEEK_SET);
+  RAW_DCHECK(offset == index, "");
+  if (offset < 0) {

[jar (doing other things), 2012/01/27 00:46:58]

nit: return offset >= 0;

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Done.

+    return false;
+  }
+  return true;
+}
+
+// static
+bool DeepHeapProfile::ReadProcPagemap(int pagemap_fd, PageState* state) {
+  static const uint64 U64_1 = 1;
+  static const uint64 PFN_FILTER = (U64_1 << 55) - U64_1;
+  static const uint64 PAGE_PRESENT = U64_1 << 63;
+  static const uint64 PAGE_SWAP = U64_1 << 62;
+  static const uint64 PAGE_RESERVED = U64_1 << 61;
+  static const uint64 FLAG_NOPAGE = U64_1 << 20;
+  static const uint64 FLAG_KSM = U64_1 << 21;
+  static const uint64 FLAG_MMAP = U64_1 << 11;
+
+  uint64 pagemap_value;
+  int result = read(pagemap_fd, &pagemap_value, PAGEMAP_BYTES);
+  if (result != PAGEMAP_BYTES) {
+    return false;
+  }
+
+  // Check if the page is committed.
+  state->is_committed = (pagemap_value & (PAGE_PRESENT | PAGE_SWAP));
+
+  state->is_present = (pagemap_value & PAGE_PRESENT);
+  state->is_swapped = (pagemap_value & PAGE_SWAP);
+  state->is_shared = false;
+
+  return true;
+}
+
+// static
+size_t DeepHeapProfile::GetCommittedSize(
+    int pagemap_fd, uint64 first_address, uint64 last_address) {
+  uint64 page_address = (first_address / PAGE_SIZE) * PAGE_SIZE;
+  size_t committed_size = 0;
+
+  SeekProcPagemap(pagemap_fd, first_address);
+
+  // Check every page on which the allocation resides.
+  while (page_address <= last_address) {
+    // Read corresponding physical page.
+    PageState state;
+    if (ReadProcPagemap(pagemap_fd, &state) == false) {
+      // We can't read the last region (e.g vsyscall).
+#ifndef NDEBUG
+      RAW_LOG(0, "pagemap read failed @ %#llx %"PRId64" bytes",
+              first_address, last_address - first_address + 1);
+#endif
+      return 0;
+    }
+
+    if (state.is_committed) {
+      // Calculate the size of the allocation part in this page.
+      size_t bytes = PAGE_SIZE;
+
+      // If looking at the last page in a given region.
+      if (last_address <= page_address - 1 + PAGE_SIZE) {
+        bytes = last_address - page_address + 1;
+      }
+
+      // If looking at the first page in a given region.
+      if (page_address < first_address) {
+        bytes -= first_address - page_address;
+      }
+
+      committed_size += bytes;
+    }
+    if (page_address > TOP_ADDRESS - PAGE_SIZE) {
+      break;
+    }
+    page_address += PAGE_SIZE;
+  }
+
+  return committed_size;
+}
+
+// static
+void DeepHeapProfile::WriteMapsToFile(char buffer[], int buffer_size,
+                                      char* filename_prefix) {
+  char filename[100];
+  snprintf(filename, sizeof(filename),
+           "%s.%05d.maps", filename_prefix, getpid());
+
+  RawFD maps_fd = RawOpenForWriting(filename);
+  RAW_DCHECK(maps_fd != kIllegalRawFD, "");
+
+  int map_length;
+  bool wrote_all;
+  map_length = tcmalloc::FillProcSelfMaps(buffer, buffer_size, &wrote_all);
+  RAW_DCHECK(wrote_all, "");
+  RAW_DCHECK(map_length <= buffer_size, "");
+  RawWrite(maps_fd, buffer, map_length);
+  RawClose(maps_fd);
+}
+
+// static
+void DeepHeapProfile::GetGlobalStats(int pagemap_fd, GlobalStats* stats) {
+  ProcMapsIterator::Buffer iterator_buffer;
+  ProcMapsIterator it(0, &iterator_buffer);
+  uint64 first_address, last_address, offset;
+  int64 inode;
+  char *flags, *filename;

[jar (doing other things), 2012/01/27 00:46:58]

nit: Chrome style puts the "*" next to the type.  As a result, it is probably
better to put the above on two lines as well.

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Done.

+
+  stats->total.Initialize();
+  stats->file_mapped.Initialize();
+  stats->anonymous.Initialize();
+  stats->other.Initialize();
+
+  while (it.Next(&first_address, &last_address,
+                 &flags, &offset, &inode, &filename)) {
+    // 'last_address' should be the last inclusive address of the region.
+    last_address -= 1;
+    if (strcmp("[vsyscall]", filename) == 0) {
+      continue;  // Reading pagemap will fail in [vsyscall].
+    }
+
+    int64 committed_bytes = stats->total.committed_bytes;
+    stats->total.Record(pagemap_fd, first_address, last_address);
+    committed_bytes = stats->total.committed_bytes - committed_bytes;

[jar (doing other things), 2012/01/27 00:46:58]

Why isn't this merely committed_bytes = 0

In addition, since committed_bytes is never again used in this iteration of the
loop, why are you even updating it?

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Removed.  It was a code to be used in a planned extension...

On 2012/01/27 00:46:58, jar wrote:

+
+    if (filename[0] == '/') {
+      stats->file_mapped.Record(pagemap_fd, first_address, last_address);
+    } else if (filename[0] == '\0' || filename[0] == '\n') {
+      stats->anonymous.Record(pagemap_fd, first_address, last_address);
+    } else {
+      stats->other.Record(pagemap_fd, first_address, last_address);
+    }
+  }
+}
+
+DeepHeapProfile::DeepBucket*
+DeepHeapProfile::GetDeepBucket(Bucket* bucket) {

[jar (doing other things), 2012/01/27 00:46:58]

nit: Try to put function definition on one line.  Wrap at the open paren if you
must.

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Done.

+  DeepBucket* found = deep_bucket_map_->FindMutable(bucket);
+  if (found == NULL) {
+    DeepBucket created;
+    created.bucket = bucket;
+    created.committed_size = 0;
+    created.id = (bucket_id_++);
+    created.is_logged = false;
+    deep_bucket_map_->Insert(bucket, created);
+    return deep_bucket_map_->FindMutable(bucket);
+  } else {

[jar (doing other things), 2012/01/27 00:46:58]

style nit: both sides of the if have a return... so you don't need an else.  If
you put the tiny branch |return found| first, then you don't need curlies, and
don't need to indent the code in lines 354-360.

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Done.

+    return found;
+  }
+}
+
+void DeepHeapProfile::ResetCommittedSize(Bucket** bucket_table) {
+  for (int i = 0; i < kHashTableSize; i++) {
+    for (Bucket* b = bucket_table[i]; b != 0; b = b->next) {

[jar (doing other things), 2012/01/27 00:46:58]

nit: use NULL for pointers:
b != NULL

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Done.

+      DeepBucket* db = GetDeepBucket(b);
+      db->committed_size = 0;
+    }
+  }
+}
+
+int DeepHeapProfile::FillBucketTable(Bucket** bucket_table,
+                                     char buffer[],
+                                     int buffer_size,
+                                     int used_in_buffer,
+                                     HeapProfileTable::Stats* stats) {
+  for (int i = 0; i < kHashTableSize; i++) {
+    for (Bucket* b = bucket_table[i]; b != 0; b = b->next) {

[jar (doing other things), 2012/01/27 00:46:58]

nit: NULL instead of 0

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Done.

+      if (b->alloc_size - b->free_size == 0) {
+        continue;  // Skip empty buckets.
+      }
+      const DeepBucket& db = *GetDeepBucket(b);

[jar (doing other things), 2012/01/27 00:46:58]

nit: suggest:
const DeepBucket db = GetDeepBucket(b);

then pass "*db" as an argument to the function.

nit: don't use abbreviations like "db" (which has a lot of loading meaning
besides).

[Dai Mikurube (NOT FULLTIME), 2012/01/30 12:54:53]

Did you mean:
  const DeepBucket *db = GetDeepBucket(b);
?

Renamed db -> deep_bucket, and b -> bucket.

[jar (doing other things), 2012/03/16 01:20:02]

Yes... thanks... the point was to avoid using references inside the method.

On 2012/01/30 12:54:53, Dai Mikurube wrote:

+      used_in_buffer =
+          UnparseBucket(db, buffer, used_in_buffer, buffer_size, "", stats);
+    }
+  }
+  return used_in_buffer;
+}
+
+void DeepHeapProfile::RecordAlloc(const void* pointer,
+                                  AllocValue* alloc_value,
+                                  DeepHeapProfile* deep_profile) {
+  uint64 address = reinterpret_cast<uintptr_t>(pointer);
+  size_t committed = GetCommittedSize(deep_profile->pagemap_fd_,
+      address, address + alloc_value->bytes - 1);
+
+  DeepBucket* db = deep_profile->GetDeepBucket(alloc_value->bucket());
+  db->committed_size += committed;
+  deep_profile->stats_.record_malloc.virtual_bytes += alloc_value->bytes;
+  deep_profile->stats_.record_malloc.committed_bytes += committed;
+}
+
+void DeepHeapProfile::RecordMMap(const void* pointer,
+                                 AllocValue* alloc_value,
+                                 DeepHeapProfile* deep_profile) {
+  uint64 address = reinterpret_cast<uintptr_t>(pointer);
+  size_t committed = GetCommittedSize(deep_profile->pagemap_fd_,
+      address, address + alloc_value->bytes - 1);
+
+  DeepBucket* db = deep_profile->GetDeepBucket(alloc_value->bucket());
+  db->committed_size += committed;
+  deep_profile->stats_.record_mmap.virtual_bytes += alloc_value->bytes;
+  deep_profile->stats_.record_mmap.committed_bytes += committed;
+}
+
+void DeepHeapProfile::RecordAllAllocs() {
+  stats_.record_mmap.virtual_bytes = 0;
+  stats_.record_mmap.committed_bytes = 0;
+  stats_.record_malloc.virtual_bytes = 0;
+  stats_.record_malloc.committed_bytes = 0;
+
+  // malloc allocations.
+  heap_profile_->alloc_address_map_->Iterate(RecordAlloc, this);
+
+  // mmap allocations.
+  heap_profile_->mmap_address_map_->Iterate(RecordMMap, this);
+}
+
+int DeepHeapProfile::FillBucketForBucketFile(const DeepBucket* deep_bucket,
+                                             char buffer[],
+                                             int buffer_size) {
+  const Bucket* bucket = deep_bucket->bucket;
+  int printed = snprintf(buffer, buffer_size, "%05d", deep_bucket->id);
+  if (printed < 0 || printed >= buffer_size) {
+    return 0;
+  }
+  int used_in_buffer = printed;
+
+  for (int d = 0; d < bucket->depth; d++) {
+    printed = snprintf(buffer + used_in_buffer, buffer_size - used_in_buffer,
+                       " 0x%08" PRIxPTR,
+                       reinterpret_cast<uintptr_t>(bucket->stack[d]));
+    if (printed < 0 || printed >= buffer_size - used_in_buffer) {
+      return used_in_buffer;
+    }
+    used_in_buffer += printed;
+  }
+  printed = snprintf(buffer + used_in_buffer, buffer_size - used_in_buffer,
+                     "\n");
+  if (printed < 0 || printed >= buffer_size - used_in_buffer) {
+    return used_in_buffer;
+  }
+  used_in_buffer += printed;
+
+  return used_in_buffer;
+}
+
+void DeepHeapProfile::WriteBucketsTableToBucketFile(Bucket** bucket_table,
+                                                    RawFD bucket_fd) {
+  // We will use the global buffer here.
+  char* buffer = profiler_buffer_;
+  int buffer_size = kProfilerBufferSize;
+  int used_in_buffer = 0;
+
+  for (int i = 0; i < kHashTableSize; i++) {
+    for (Bucket* b = bucket_table[i]; b != 0; b = b->next) {
+      DeepBucket* db = GetDeepBucket(b);
+      if (db->is_logged) {
+        continue;  // Skip the bucket if it is already logged.
+      }
+      if (b->alloc_size - b->free_size <= 64) {
+        continue;  // Skip small buckets.
+      }
+
+      used_in_buffer += FillBucketForBucketFile(
+          db, buffer + used_in_buffer, buffer_size - used_in_buffer);
+      db->is_logged = true;
+
+      // Write to file if buffer 80% full.
+      if (used_in_buffer > buffer_size * 0.8) {
+        RawWrite(bucket_fd, buffer, used_in_buffer);
+        used_in_buffer = 0;
+      }
+    }
+  }
+
+  RawWrite(bucket_fd, buffer, used_in_buffer);
+}
+
+void DeepHeapProfile::WriteBucketsToBucketFile() {
+  char filename[100];
+  snprintf(filename, sizeof(filename),
+           "%s.%05d.%04d.buckets", filename_prefix_, getpid(), dump_count_);
+  RawFD bucket_fd = RawOpenForWriting(filename);
+  RAW_DCHECK(bucket_fd != kIllegalRawFD, "");
+
+  WriteBucketsTableToBucketFile(heap_profile_->alloc_table_, bucket_fd);
+  WriteBucketsTableToBucketFile(heap_profile_->mmap_table_, bucket_fd);
+
+  RawClose(bucket_fd);
+}
+
+int DeepHeapProfile::UnparseBucket(const DeepBucket& deep_bucket,
+                                   char* buffer,
+                                   int used_in_buffer,
+                                   int buffer_size,
+                                   const char* extra,
+                                   Stats* profile_stats) {
+  const Bucket& bucket = *deep_bucket.bucket;
+  if (profile_stats != NULL) {
+    profile_stats->allocs += bucket.allocs;
+    profile_stats->alloc_size += bucket.alloc_size;
+    profile_stats->frees += bucket.frees;
+    profile_stats->free_size += bucket.free_size;
+  }
+
+  int printed = snprintf(buffer + used_in_buffer, buffer_size - used_in_buffer,
+                         "%10"PRId64" %10"PRId64" %6d %6d @%s %d\n",
+                         bucket.alloc_size - bucket.free_size,
+                         deep_bucket.committed_size,
+                         bucket.allocs, bucket.frees, extra, deep_bucket.id);
+  // If it looks like the snprintf failed, ignore the fact we printed anything.
+  if (printed < 0 || printed >= buffer_size - used_in_buffer) {
+    return used_in_buffer;
+  }
+  used_in_buffer += printed;
+
+  return used_in_buffer;
+}
+
+int DeepHeapProfile::UnparseRegionStats(const RegionStats* stats,
+                                        const char* name,
+                                        char* buffer,
+                                        int used_in_buffer,
+                                        int buffer_size) {
+  int printed = snprintf(buffer + used_in_buffer, buffer_size - used_in_buffer,
+                         "%15s %10ld %10ld\n",
+                         name, stats->virtual_bytes, stats->committed_bytes);
+  if (printed < 0 || printed >= buffer_size - used_in_buffer) {
+    return used_in_buffer;
+  }
+  used_in_buffer += printed;
+
+  return used_in_buffer;
+}
+
+int DeepHeapProfile::UnparseGlobalStats(char* buffer,
+                                        int used_in_buffer,
+                                        int buffer_size) {
+  int printed = snprintf(buffer + used_in_buffer, buffer_size - used_in_buffer,
+                         "%15s %10s %10s\n", "", "virtual", "committed");
+  if (printed < 0 || printed >= buffer_size - used_in_buffer) {
+    return used_in_buffer;
+  }
+  used_in_buffer += printed;
+
+  used_in_buffer = UnparseRegionStats(&(stats_.total), "total",
+      buffer, used_in_buffer, buffer_size);
+  used_in_buffer = UnparseRegionStats(&(stats_.file_mapped), "file mapped",
+      buffer, used_in_buffer, buffer_size);
+  used_in_buffer = UnparseRegionStats(&(stats_.anonymous), "anonymous",
+      buffer, used_in_buffer, buffer_size);
+  used_in_buffer = UnparseRegionStats(&(stats_.other), "other",
+      buffer, used_in_buffer, buffer_size);
+  used_in_buffer = UnparseRegionStats(&(stats_.record_mmap), "mmap",
+      buffer, used_in_buffer, buffer_size);
+  used_in_buffer = UnparseRegionStats(&(stats_.record_malloc), "tcmalloc",
+      buffer, used_in_buffer, buffer_size);
+  return used_in_buffer;
+}
+#else  // DEEP_HEAP_PROFILE
+
+DeepHeapProfile::DeepHeapProfile(HeapProfileTable* heap_profile,
+                                 const char* prefix)
+    : heap_profile_(heap_profile) {
+}
+
+DeepHeapProfile::~DeepHeapProfile() {
+}
+
+int DeepHeapProfile::FillOrderedProfile(char buffer[], int buffer_size) {
+  return heap_profile_->FillOrderedProfile(buffer, buffer_size);
+}
+
+#endif  // DEEP_HEAP_PROFILE