Count m(un)map for each stacktrace in MemoryRegionMap instead of HeapProfileTable.

third_party/tcmalloc/chromium/src/deep-heap-profile.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.
 
 // ---
 // Author: Sainbayar Sukhbaatar
 //         Dai Mikurube
 //
 
 #include "deep-heap-profile.h"
 
 #ifdef DEEP_HEAP_PROFILE
 #include <algorithm>
 #include <fcntl.h>
 #include <sys/stat.h>
 #include <sys/types.h>
 #ifdef HAVE_UNISTD_H
 #include <unistd.h>  // for getpagesize and getpid
 #endif  // HAVE_UNISTD_H
 
 #include "base/cycleclock.h"
 #include "base/low_level_alloc.h"
 #include "base/sysinfo.h"
 #include "internal_logging.h"  // for ASSERT, etc
+#include "memory_region_map.h"
 
 static const int kProfilerBufferSize = 1 << 20;
 static const int kHashTableSize = 179999;  // Same as heap-profile-table.cc.
 
 static const int PAGEMAP_BYTES = 8;
 static const uint64 MAX_ADDRESS = kuint64max;
 
 // Tag strings in heap profile dumps.
 static const char kProfileHeader[] = "heap profile: ";
 static const char kProfileVersion[] = "DUMP_DEEP_6";
@@ skipping 69 matching lines @@
   uint64 page_address = (first_address / page_size) * page_size;
   size_t committed_size = 0;
 
   Seek(first_address);
 
   // Check every page on which the allocation resides.
   while (page_address <= last_address) {
     // Read corresponding physical page.
     State state;
     // TODO(dmikurube): Read pagemap in bulk for speed.
+    // TODO(dmikurube): Consider using mincore(2).

[Alexander Potapenko, 2013/03/07 06:08:25]

Just wondering - how can mincore help you?

[Dai Mikurube (NOT FULLTIME), 2013/03/07 17:14:24]

This function is reading /proc/$PID/pagemap, which has similar information with
mincore(2).  I'm wondering that mincore(2) could be a little faster than reading
via procfs.

     if (Read(&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) {
@@ skipping 119 matching lines @@
     memory_residence_info_getter_->Initialize();
     deep_table_.ResetIsLogged();
 
     // Write maps into "|filename_prefix_|.<pid>.maps".
     WriteProcMaps(filename_prefix_, kProfilerBufferSize, profiler_buffer_);
   }
 
   // Reset committed sizes of buckets.
   deep_table_.ResetCommittedSize();
 
-  // Allocate a list for mmap'ed regions.
-  num_mmap_allocations_ = 0;
-  if (heap_profile_->mmap_address_map_) {
-    heap_profile_->mmap_address_map_->Iterate(CountMMap, this);
-
-    mmap_list_length_ = 0;
-    mmap_list_ = reinterpret_cast<MMapListEntry*>(heap_profile_->alloc_(
-        sizeof(MMapListEntry) * num_mmap_allocations_));
-
-    // Touch all the allocated pages.  Touching is required to avoid new page
-    // commitment while filling the list in SnapshotProcMaps.
-    for (int i = 0;
-         i < num_mmap_allocations_;
-         i += getpagesize() / 2 / sizeof(MMapListEntry))
-      mmap_list_[i].first_address = 0;
-    mmap_list_[num_mmap_allocations_ - 1].last_address = 0;
-  }
-
-  stats_.SnapshotProcMaps(memory_residence_info_getter_, NULL, 0, NULL);
+  // Record committed sizes.
+  stats_.SnapshotAllocations(this);
 
   // TODO(dmikurube): Eliminate dynamic memory allocation caused by snprintf.
   // glibc's snprintf internally allocates memory by alloca normally, but it
   // allocates memory by malloc if large memory is required.
 
-  // Record committed sizes.
-  stats_.SnapshotAllocations(this);
-
   buffer.AppendString(kProfileHeader, 0);
   buffer.AppendString(kProfileVersion, 0);
   buffer.AppendString("\n", 0);
 
   // Fill buffer with the global stats.
   buffer.AppendString(kMMapListHeader, 0);
 
-  // Check if committed bytes changed during SnapshotAllocations.
-  stats_.SnapshotProcMaps(memory_residence_info_getter_,
-                          mmap_list_,
-                          mmap_list_length_,
-                          &buffer);
+  stats_.SnapshotMaps(memory_residence_info_getter_, this, &buffer);
 
   // Fill buffer with the global stats.
   buffer.AppendString(kGlobalStatsHeader, 0);
 
   stats_.Unparse(&buffer);
 
   buffer.AppendString(kStacktraceHeader, 0);
   buffer.AppendString(kVirtualLabel, 10);
   buffer.AppendChar(' ');
   buffer.AppendString(kCommittedLabel, 10);
   buffer.AppendString("\n", 0);
 
   // Fill buffer.
   deep_table_.UnparseForStats(&buffer);
 
   RAW_DCHECK(buffer.FilledBytes() < buffer_size, "");
 
-  heap_profile_->dealloc_(mmap_list_);
-  mmap_list_ = NULL;
-
   // Write the bucket listing into a .bucket file.
   deep_table_.WriteForBucketFile(filename_prefix_, dump_count_, &global_buffer);
 
 #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 buffer.FilledBytes();
@@ skipping 300 matching lines @@
                                            TextBuffer* buffer) {
   buffer->AppendString(name, 25);
   buffer->AppendChar(' ');
   buffer->AppendLong(virtual_bytes_, 12);
   buffer->AppendChar(' ');
   buffer->AppendLong(committed_bytes_, 12);
   buffer->AppendString("\n", 0);
 }
 
 // TODO(dmikurube): Eliminate dynamic memory allocation caused by snprintf.
-void DeepHeapProfile::GlobalStats::SnapshotProcMaps(
+void DeepHeapProfile::GlobalStats::SnapshotMaps(
     const MemoryResidenceInfoGetterInterface* memory_residence_info_getter,
-    MMapListEntry* mmap_list,
-    int mmap_list_length,
+    DeepHeapProfile* deep_profile,
     TextBuffer* mmap_dump_buffer) {
-  ProcMapsIterator::Buffer iterator_buffer;
-  ProcMapsIterator iterator(0, &iterator_buffer);
+  ProcMapsIterator::Buffer procmaps_iter_buffer;
+  ProcMapsIterator procmaps_iter(0, &procmaps_iter_buffer);
   uint64 first_address, last_address, offset;
   int64 inode;
   char* flags;
   char* filename;
-  int mmap_list_index = 0;
   enum MapsRegionType type;
-
   for (int i = 0; i < NUMBER_OF_MAPS_REGION_TYPES; ++i) {
     all_[i].Initialize();
     unhooked_[i].Initialize();
   }
+  profiled_mmap_.Initialize();
 
   size_t unhooked_arena_size = LowLevelAlloc::GetSizeOfUnhookedArena();
   unhooked_arena_.Initialize();
   unhooked_arena_.AddToVirtualBytes(unhooked_arena_size);
   unhooked_arena_.AddToCommittedBytes(unhooked_arena_size);
 
-  while (iterator.Next(&first_address, &last_address,
-                       &flags, &offset, &inode, &filename)) {
+  MemoryRegionMap::Lock();
+  MemoryRegionMap::RegionIterator mmap_iter =
+      MemoryRegionMap::BeginRegionLocked();
+
+  while (procmaps_iter.Next(&first_address, &last_address,
+                            &flags, &offset, &inode, &filename)) {
     if (mmap_dump_buffer) {
       char buffer[1024];
-      int written = iterator.FormatLine(buffer, sizeof(buffer),
-                                        first_address, last_address, flags,
-                                        offset, inode, filename, 0);
+      int written = procmaps_iter.FormatLine(buffer, sizeof(buffer),
+                                             first_address, last_address, flags,
+                                             offset, inode, filename, 0);
       mmap_dump_buffer->AppendString(buffer, 0);
     }
 
     // '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].
     }
 
     type = ABSENT;
     if (filename[0] == '/') {
       if (flags[2] == 'x')
         type = FILE_EXEC;
       else
         type = FILE_NONEXEC;
     } else if (filename[0] == '\0' || filename[0] == '\n') {
       type = ANONYMOUS;
     } else if (strcmp(filename, "[stack]") == 0) {
       type = STACK;
     } else {
       type = OTHER;
     }
     all_[type].Record(
         memory_residence_info_getter, first_address, last_address);
 
     // TODO(dmikurube): Stop double-counting pagemap.
     // Counts unhooked memory regions in /proc/<pid>/maps.
-    if (mmap_list != NULL) {
+    if (MemoryRegionMap::IsWorking()) {
       // It assumes that every mmap'ed region is included in one maps line.
       uint64 cursor = first_address;
       bool first = true;
 
       do {
+        Bucket* bucket = NULL;
+        DeepBucket* deep_bucket = NULL;
         if (!first) {
-          mmap_list[mmap_list_index].type = type;
-          cursor = mmap_list[mmap_list_index].last_address + 1;
-          ++mmap_list_index;
+          size_t committed = deep_profile->memory_residence_info_getter_->
+              CommittedSize(mmap_iter->start_addr, mmap_iter->end_addr - 1);
+          // TODO(dmikurube): Store a reference to the bucket in region.
+          Bucket* bucket = MemoryRegionMap::GetBucket(
+              mmap_iter->call_stack_depth, mmap_iter->call_stack);
+          DeepBucket* deep_bucket = NULL;
+          if (bucket != NULL) {
+            deep_bucket = deep_profile->deep_table_.Lookup(
+                bucket,
+#if defined(TYPE_PROFILING)
+                NULL,
+#endif
+                /* is_mmap */ true);
+          }
+
+          if (deep_bucket != NULL)
+            deep_bucket->committed_size += committed;
+          profiled_mmap_.AddToVirtualBytes(
+              mmap_iter->end_addr - mmap_iter->start_addr);
+          profiled_mmap_.AddToCommittedBytes(committed);
+
+          cursor = mmap_iter->end_addr;
+          ++mmap_iter;
+          // Don't break here even if mmap_iter == EndRegionLocked().
         }
         first = false;
 
         uint64 last_address_of_unhooked;
         // If the next mmap entry is away from the current maps line.
-        if (mmap_list_index >= mmap_list_length ||
-            mmap_list[mmap_list_index].first_address > last_address) {
+        if (mmap_iter == MemoryRegionMap::EndRegionLocked() ||
+            mmap_iter->start_addr > last_address) {
           last_address_of_unhooked = last_address;
         } else {
-          last_address_of_unhooked =
-              mmap_list[mmap_list_index].first_address - 1;
+          last_address_of_unhooked = mmap_iter->start_addr - 1;
         }
 
         if (last_address_of_unhooked + 1 > cursor) {
           uint64 committed_size = unhooked_[type].Record(
               memory_residence_info_getter,
               cursor,
               last_address_of_unhooked);
           if (mmap_dump_buffer) {
             mmap_dump_buffer->AppendString("  ", 0);
             mmap_dump_buffer->AppendPtr(cursor, 0);
             mmap_dump_buffer->AppendString(" - ", 0);
             mmap_dump_buffer->AppendPtr(last_address_of_unhooked + 1, 0);
             mmap_dump_buffer->AppendString("  unhooked ", 0);
             mmap_dump_buffer->AppendString(kMapsRegionTypeDict[type], 0);
             mmap_dump_buffer->AppendString(" ", 0);
             mmap_dump_buffer->AppendInt64(committed_size, 0);
             mmap_dump_buffer->AppendString("\n", 0);
           }
           cursor = last_address_of_unhooked + 1;
         }
 
-        if (mmap_list_index < mmap_list_length &&
-            mmap_list[mmap_list_index].first_address <= last_address &&
+        if (mmap_iter != MemoryRegionMap::EndRegionLocked() &&
+            mmap_iter->start_addr <= last_address &&
             mmap_dump_buffer) {
-          bool trailing =
-            mmap_list[mmap_list_index].first_address < first_address;
-          bool continued =
-            mmap_list[mmap_list_index].last_address > last_address;
+          bool trailing = mmap_iter->start_addr < first_address;
+          bool continued = mmap_iter->end_addr - 1 > last_address;
           mmap_dump_buffer->AppendString(trailing ? " (" : "  ", 0);
-          mmap_dump_buffer->AppendPtr(
-              mmap_list[mmap_list_index].first_address, 0);
+          mmap_dump_buffer->AppendPtr(mmap_iter->start_addr, 0);
           mmap_dump_buffer->AppendString(trailing ? ")" : " ", 0);
           mmap_dump_buffer->AppendString("-", 0);
           mmap_dump_buffer->AppendString(continued ? "(" : " ", 0);
-          mmap_dump_buffer->AppendPtr(
-              mmap_list[mmap_list_index].last_address + 1, 0);
+          mmap_dump_buffer->AppendPtr(mmap_iter->end_addr, 0);
           mmap_dump_buffer->AppendString(continued ? ")" : " ", 0);
           mmap_dump_buffer->AppendString(" hooked ", 0);
           mmap_dump_buffer->AppendString(kMapsRegionTypeDict[type], 0);
           mmap_dump_buffer->AppendString(" @ ", 0);
-          mmap_dump_buffer->AppendInt(
-              mmap_list[mmap_list_index].deep_bucket->id, 0);
+          if (deep_bucket != NULL)
+            mmap_dump_buffer->AppendInt(deep_bucket->id, 0);
+          else
+            mmap_dump_buffer->AppendInt(0, 0);
           mmap_dump_buffer->AppendString("\n", 0);
         }
-      } while (mmap_list_index < mmap_list_length &&
-               mmap_list[mmap_list_index].last_address <= last_address);
+      } while (mmap_iter != MemoryRegionMap::EndRegionLocked() &&
+               mmap_iter->end_addr - 1 <= last_address);
     }
   }
+  MemoryRegionMap::Unlock();
 }
 
 void DeepHeapProfile::GlobalStats::SnapshotAllocations(
     DeepHeapProfile* deep_profile) {
-  profiled_mmap_.Initialize();
   profiled_malloc_.Initialize();
 
-  // malloc allocations.
-  deep_profile->heap_profile_->alloc_address_map_->Iterate(RecordAlloc,
-                                                           deep_profile);
-
-  // mmap allocations.
-  if (deep_profile->heap_profile_->mmap_address_map_) {
-    deep_profile->heap_profile_->mmap_address_map_->Iterate(RecordMMap,
-                                                            deep_profile);
-    std::sort(deep_profile->mmap_list_,
-              deep_profile->mmap_list_ + deep_profile->mmap_list_length_,
-              ByFirstAddress);
-  }
+  deep_profile->heap_profile_->address_map_->Iterate(RecordAlloc, deep_profile);
 }
 
 void DeepHeapProfile::GlobalStats::Unparse(TextBuffer* buffer) {
   RegionStats all_total;
   RegionStats unhooked_total;
   for (int i = 0; i < NUMBER_OF_MAPS_REGION_TYPES; ++i) {
     all_total.AddAnotherRegionStat(all_[i]);
     unhooked_total.AddAnotherRegionStat(unhooked_[i]);
   }
 
@@ skipping 29 matching lines @@
   unhooked_[FILE_EXEC].Unparse("nonprofiled-file-exec", buffer);
   unhooked_[FILE_NONEXEC].Unparse("nonprofiled-file-nonexec", buffer);
   unhooked_[STACK].Unparse("nonprofiled-stack", buffer);
   unhooked_[OTHER].Unparse("nonprofiled-other", buffer);
   profiled_mmap_.Unparse("profiled-mmap", buffer);
   profiled_malloc_.Unparse("profiled-malloc", buffer);
   unhooked_arena_.Unparse("unhooked-arena", buffer);
 }
 
 // static
-bool DeepHeapProfile::GlobalStats::ByFirstAddress(const MMapListEntry& a,
-                                                  const MMapListEntry& b) {
-  return a.first_address < b.first_address;
-}
-
-// static
 void DeepHeapProfile::GlobalStats::RecordAlloc(const void* pointer,
                                                AllocValue* alloc_value,
                                                DeepHeapProfile* deep_profile) {
   uint64 address = reinterpret_cast<uintptr_t>(pointer);
   size_t committed = deep_profile->memory_residence_info_getter_->CommittedSize(
       address, address + alloc_value->bytes - 1);
 
   DeepBucket* deep_bucket = deep_profile->deep_table_.Lookup(
       alloc_value->bucket(),
 #if defined(TYPE_PROFILING)
       LookupType(pointer),
 #endif
       /* is_mmap */ false);
   deep_bucket->committed_size += committed;
   deep_profile->stats_.profiled_malloc_.AddToVirtualBytes(alloc_value->bytes);
   deep_profile->stats_.profiled_malloc_.AddToCommittedBytes(committed);
 }
 
 // static
-void DeepHeapProfile::GlobalStats::RecordMMap(const void* pointer,
-                                              AllocValue* alloc_value,
-                                              DeepHeapProfile* deep_profile) {
-  uint64 address = reinterpret_cast<uintptr_t>(pointer);
-  size_t committed = deep_profile->memory_residence_info_getter_->CommittedSize(
-      address, address + alloc_value->bytes - 1);
-
-  DeepBucket* deep_bucket = deep_profile->deep_table_.Lookup(
-      alloc_value->bucket(),
-#if defined(TYPE_PROFILING)
-      NULL,
-#endif
-      /* is_mmap */ true);
-  deep_bucket->committed_size += committed;
-  deep_profile->stats_.profiled_mmap_.AddToVirtualBytes(alloc_value->bytes);
-  deep_profile->stats_.profiled_mmap_.AddToCommittedBytes(committed);
-
-  if (deep_profile->mmap_list_length_ < deep_profile->num_mmap_allocations_) {
-    deep_profile->mmap_list_[deep_profile->mmap_list_length_].first_address =
-        address;
-    deep_profile->mmap_list_[deep_profile->mmap_list_length_].last_address =
-        address - 1 + alloc_value->bytes;
-    deep_profile->mmap_list_[deep_profile->mmap_list_length_].type = ABSENT;
-    deep_profile->mmap_list_[deep_profile->mmap_list_length_].deep_bucket =
-        deep_bucket;
-    ++deep_profile->mmap_list_length_;
-  } else {
-    RAW_LOG(0, "Unexpected number of mmap entries: %d/%d",
-            deep_profile->mmap_list_length_,
-            deep_profile->num_mmap_allocations_);
-  }
-}
-
-// static
 void DeepHeapProfile::WriteProcMaps(const char* prefix,
                                     int buffer_size,
                                     char raw_buffer[]) {
   char filename[100];
   snprintf(filename, sizeof(filename),
            "%s.%05d.maps", prefix, static_cast<int>(getpid()));
 
   RawFD fd = RawOpenForWriting(filename);
   RAW_DCHECK(fd != kIllegalRawFD, "");
 
   int length;
   bool wrote_all;
   length = tcmalloc::FillProcSelfMaps(raw_buffer, buffer_size, &wrote_all);
   RAW_DCHECK(wrote_all, "");
   RAW_DCHECK(length <= buffer_size, "");
   RawWrite(fd, raw_buffer, length);
   RawClose(fd);
 }
-
-// static
-void DeepHeapProfile::CountMMap(const void* pointer,
-                                AllocValue* alloc_value,
-                                DeepHeapProfile* deep_profile) {
-  ++deep_profile->num_mmap_allocations_;
-}
 #else  // DEEP_HEAP_PROFILE
 
 DeepHeapProfile::DeepHeapProfile(HeapProfileTable* heap_profile,
                                  const char* prefix)
     : heap_profile_(heap_profile) {
 }
 
 DeepHeapProfile::~DeepHeapProfile() {
 }
 
 int DeepHeapProfile::FillOrderedProfile(char raw_buffer[], int buffer_size) {
   return heap_profile_->FillOrderedProfile(raw_buffer, buffer_size);
 }
 
 #endif  // DEEP_HEAP_PROFILE