Deep-Memory-Profiler (DMP) implementation

third_party/tcmalloc/chromium/src/deep-memory-profiler.cc

+// 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.
+
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <unistd.h>
+
+#include "deep-memory-profiler.h"
+#include "base/sysinfo.h"
+#include "base/cycleclock.h"
+
+using tcmalloc::FillProcSelfMaps;   // from sysinfo.h
+using tcmalloc::DumpProcSelfMaps;   // from sysinfo.h
+
+// Those are used in parsing /proc/self/pagemaps
+#define PAGE_SIZE          4096
+#define U64_1              ((uint64)1)
+#define PFN_FILTER         ((U64_1 << 55) - U64_1)
+#define PAGE_PRESENT       U64_1 << 63
+#define PAGE_SWAP          U64_1 << 62
+#define PAGE_RESERVED      U64_1 << 61
+#define FLAG_NOPAGE        U64_1 << 20
+#define FLAG_KSM           U64_1 << 21
+#define FLAG_MMAP          U64_1 << 11
+#define PAGEMAP_BYTES      8
+
+static const int kProfilerBufferSize = 1 << 20;
+static const int kHashTableSize = 179999;   // Have to same as heap-profile-table.cc
+
+// header of the dumped heap profile
+static const char kProfileHeader[] = "Deep Memory Profile\n";
+static const char kGlobalStatsHeader[] = "GLOBAL_STATS:\n";
+static const char kStacktraceHeader[] = "STACKTRACES:\n";
+static const char kProcSelfMapsHeader[] = "\nMAPPED_LIBRARIES:\n";
+
+DeepMemoryProfiler::DeepMemoryProfiler(HeapProfileTable* heap_profile, const char* prefix) {
+  heap_profile_ = heap_profile;
+
+  //kpageflags_fd_ = open("/proc/kpageflags", O_RDONLY);
+  //RAW_DCHECK(kpageflags_fd_ != -1, "Failed to open /proc/kpageflags");
+  kpageflags_fd_ = -1; // Not currently using this
+
+  page_map_ = new(heap_profile_->alloc_(sizeof(PageStateMap))) 
+      PageStateMap(heap_profile_->alloc_, heap_profile_->dealloc_);
+
+  // Copy filename prefix
+  RAW_DCHECK(filename_prefix_ == NULL, "");
+  const int prefix_length = strlen(prefix);
+  filename_prefix_ = reinterpret_cast<char*>(heap_profile_->alloc_(prefix_length + 1));
+  memcpy(filename_prefix_, prefix, prefix_length);
+  filename_prefix_[prefix_length] = '\0';
+
+  profiler_buffer_ = reinterpret_cast<char*>(heap_profile_->alloc_(kProfilerBufferSize));
+  dump_count_ = 0;
+}
+
+DeepMemoryProfiler::~DeepMemoryProfiler() {
+  page_map_->~PageStateMap();
+  heap_profile_->dealloc_(page_map_);
+  heap_profile_->dealloc_(filename_prefix_);
+  heap_profile_->dealloc_(profiler_buffer_);
+}
+
+// This function need to be called after each fork
+void DeepMemoryProfiler::OpenPageMap() {
+  char filename[100];
+  sprintf(filename, "/proc/%d/pagemap", getpid());
+  pagemap_fd_ = open(filename, O_RDONLY);
+  RAW_DCHECK(pagemap_fd_ != -1, "Failed to open /proc/self/pagemap");
+}
+
+bool DeepMemoryProfiler::PageMapSeek(uint64 addr) {
+  uint64 index = (addr / PAGE_SIZE) * PAGEMAP_BYTES;
+  uint64 o = lseek64(pagemap_fd_, index, SEEK_SET);
+  RAW_DCHECK(o == index, "");
+  return true;
+}
+
+bool DeepMemoryProfiler::PageMapRead(PageState* state) {
+  uint64 pa;
+  int t = read(pagemap_fd_, &pa, PAGEMAP_BYTES);
+  if(t != PAGEMAP_BYTES)
+    return false;
+
+  // Check if the page is committed
+  state->is_committed = (pa & (PAGE_PRESENT | PAGE_SWAP));
+
+  state->is_present = (pa & PAGE_PRESENT);
+  state->is_swapped = (pa & PAGE_SWAP);
+
+  // We can get more detailed stats from kPageflags
+  if(state->is_present && kpageflags_fd_ != -1) {
+    uint64 pfn = pa & PFN_FILTER;
+    int64 index = pfn * sizeof(uint64);
+    if (lseek64(kpageflags_fd_, index, SEEK_SET) != index) {
+      RAW_LOG(ERROR, "kpageflags seek failed. errno %d",errno);
+      return false;
+    }
+    uint64 flags;
+    if (read(kpageflags_fd_, &flags, sizeof(uint64)) < 0) {
+      RAW_LOG(ERROR, "kpageflags read failed. errno %d",errno);
+      return false;
+    }
+    if(flags & FLAG_NOPAGE) RAW_LOG(ERROR,"NOPAGE at present page frame %"PRId64"", pfn);
+    state->is_shared = (flags & FLAG_KSM);
+    state->is_mmap = (flags & FLAG_MMAP);
+  }else{
+    state->is_shared = false;
+  }
+
+  return true;
+}
+
+uint64 DeepMemoryProfiler::GetCommittedSize(uint64 addr, uint64 size) {
+  uint64 page_addr = (addr / PAGE_SIZE) * PAGE_SIZE;
+  uint64 committed_size = 0;
+
+  PageMapSeek(addr);
+  // Check every pages on which the allocation reside
+  while(page_addr < addr + size) {
+    // Read corresponding physical page
+    PageState state;
+    if(PageMapRead(&state) == false){
+      // We can't read the last region (e.g vsyscall)
+      RAW_LOG(0, "pagemap read failed @ %#llx %"PRId64" bytes", addr, size);
+      return 0;
+    }
+
+    if(state.is_committed){
+      // Calculate the size of the allocation part in this page
+      uint64 bytes = PAGE_SIZE;
+      if(page_addr < addr)
+        bytes -= addr - page_addr;
+      if(addr + size < page_addr + PAGE_SIZE)
+        bytes -= PAGE_SIZE - (addr + size - page_addr);
+
+      committed_size += bytes;
+    }
+    page_addr += PAGE_SIZE;
+  }
+
+  return committed_size;
+}
+
+void DeepMemoryProfiler::InitRegionStats(RegionStats* stats) {
+  stats->virtual_bytes = 0;
+  stats->committed_bytes = 0;
+}
+
+void DeepMemoryProfiler::RecordRegionStats(uint64 start, uint64 end,
+                                               RegionStats* stats) {
+  stats->virtual_bytes += end - start;
+  stats->committed_bytes += GetCommittedSize(start, end - start);
+}
+
+void DeepMemoryProfiler::GetGlobalStats() {
+  ProcMapsIterator::Buffer iterbuf;
+  ProcMapsIterator it(0, &iterbuf);
+  uint64 start, end, offset;
+  int64 inode;
+  char *flags, *filename;
+
+  InitRegionStats(&(stats_.total));
+  InitRegionStats(&(stats_.file_mapped));
+  InitRegionStats(&(stats_.anonymous));
+  InitRegionStats(&(stats_.other));
+
+  /*
+  char file_name[1000];
+  snprintf(file_name, sizeof(file_name), "%s.%05d.%04d.maps",
+           filename_prefix_, getpid(), dump_count_);
+
+  RawFD maps_fd = RawOpenForWriting(file_name);
+  RAW_DCHECK(maps_fd != kIllegalRawFD, "");
+
+  // We use global buffer here
+  char* buf = profiler_buffer_;
+  int bufsize = kProfilerBufferSize;
+  int buflen = 0;
+  int64 mmap_end = 0;
+  const AllocValue* v = NULL;
+  */
+
+  while (it.Next(&start, &end, &flags, &offset, &inode, &filename)) {
+    if(strcmp("[vsyscall]", filename) == 0) continue;  // pagemap read fails in this region
+
+    int64 committed_bytes = stats_.total.committed_bytes;
+    RecordRegionStats(start, end, &(stats_.total));
+    committed_bytes = stats_.total.committed_bytes - committed_bytes;
+
+    if(filename[0] == '/') {
+      RecordRegionStats(start, end, &(stats_.file_mapped));
+    }else if(filename[0] == '\0' || filename[0] == '\n' || filename[0] == EOF){ 
+      RecordRegionStats(start, end, &(stats_.anonymous));
+    }else{
+      RecordRegionStats(start, end, &(stats_.other));
+    }
+    /*
+    RegionValue rv;
+    rv.start = start;
+    rv.end = end;
+    rv.size = end - start;
+    rv.committed_size = committed_bytes;
+    rv.recorded_size = 0;
+    rv.recorded_committed_size = 0;
+    memcpy(rv.permissions, flags, 4);
+    rv.permissions[4] = '\0';
+    memcpy(rv.filename, filename, strlen(filename) + 1);
+    regions_->Insert(reinterpret_cast<void*>(start), rv);
+
+    if(rv.size > max_region_size_)
+      max_region_size_ = rv.size;
+    */
+
+    /*
+    uint64 recorded_bytes = 0;
+    if(mmap_end > start && v != NULL){
+      // This region is a part of the previous mmap allocation
+      if(mmap_end >= end)
+        recorded_bytes = end - start;
+      else
+        recorded_bytes = mmap_end - start;
+    }else{
+      v = heap_profile_->allocation_mmap_->Find(reinterpret_cast<void*>(start));
+      if(v != NULL){
+        mmap_end = start + v->bytes;
+        if(mmap_end >= end)
+          recorded_bytes = end - start;
+        else
+          recorded_bytes = mmap_end - start;
+        while(mmap_end < end){
+          // There might be multiple mmap allocations in this region
+          v = heap_profile_->allocation_mmap_->Find(reinterpret_cast<void*>(mmap_end));
+          if(v == NULL)
+            break;
+          else{
+            mmap_end += v->bytes;
+            recorded_bytes += v->bytes;
+          }
+        }
+      }
+    }
+
+    buflen += snprintf(buf + buflen, bufsize - buflen,
+                       "%#llx-%#llx %10"PRId64" %10"PRId64" %10"PRId64" %s\n",
+                       start, end, (end - start), committed_bytes, recorded_bytes,
+                       filename);
+    */
+  }
+
+  //RawWrite(maps_fd, buf, buflen);
+  //RawClose(maps_fd);
+}
+
+void DeepMemoryProfiler::RecordAllocInRegions(uint64 addr, uint64 size) {
+  const void* regions_start;
+  const RegionValue* rv_const = regions_->FindInside(&GetRegionSize,
+                                                     (size_t)max_region_size_,
+                                                     (void*)addr, &regions_start);
+
+  if(rv_const == NULL) {
+    RAW_LOG(0, "Could't find a region for allocation!");
+    return;
+  }
+  RegionValue* rv = regions_->FindMutable(regions_start);
+  RAW_DCHECK(rv->start <= addr, "");
+  RAW_DCHECK(rv->end > addr, "");
+
+  if(rv->end >= addr + size) {
+    // This region includes the whole allocation
+    rv->recorded_size += size;
+    rv->recorded_committed_size += GetCommittedSize(addr, size);
+  }else{
+    // This region includes only a part of the allocation.
+    // Need to find the other regions that include this allocation.
+    rv->recorded_size += rv->end - addr;
+    rv->recorded_committed_size += GetCommittedSize(addr, rv->end - addr);
+    RecordAllocInRegions(rv->end, addr + size - rv->end);
+  }
+}
+
+void DeepMemoryProfiler::RecordAlloc(const void* ptr,  AllocValue* v,
+                                     DeepMemoryProfiler* deep_profiler) {
+  uint64 alloc_addr = (uint64)ptr;
+  uint64 committed = deep_profiler->GetCommittedSize(alloc_addr, v->bytes);
+
+  v->bucket()->committed_size += committed;
+  if(deep_profiler->recording_mmap_){
+    deep_profiler->stats_.record_mmap.virtual_bytes += v->bytes;
+    deep_profiler->stats_.record_mmap.committed_bytes += committed;
+    // Record this allocation in the region map
+    //deep_profiler->RecordAllocInRegions(alloc_addr, v->bytes);
+  }else{
+    deep_profiler->stats_.record_tcmalloc.virtual_bytes += v->bytes;
+    deep_profiler->stats_.record_tcmalloc.committed_bytes += committed;
+  }
+}
+
+void DeepMemoryProfiler::RecordAllAllocs() {
+  stats_.record_mmap.virtual_bytes = 0;
+  stats_.record_mmap.committed_bytes = 0;
+  stats_.record_tcmalloc.virtual_bytes = 0;
+  stats_.record_tcmalloc.committed_bytes = 0;
+
+  // Tcmalloc allocs
+  recording_mmap_ = false;
+  heap_profile_->allocation_->Iterate(RecordAlloc, this);
+
+  // Mmap allocs
+  recording_mmap_ = true;
+  heap_profile_->allocation_mmap_->Iterate(RecordAlloc, this);
+}
+
+
+void DeepMemoryProfiler::WriteLeakyRegion(const void* ptr,
+                                          RegionValue* rv,
+                                          BufferArgs* buffer) {
+  if(rv->committed_size > rv->recorded_committed_size){
+    //if(rv->filename[0] != '\0')
+    //  return;
+    if(buffer->len >= buffer->size)
+      return;
+
+    int printed = snprintf(buffer->buf + buffer->len, buffer->size - buffer->len,
+                           "%#llx-%#llx %s %10"PRId64" %10"PRId64""
+                            " %10"PRId64" %10"PRId64" %s\n",
+                           rv->start, rv->end, rv->permissions,
+                           rv->size - rv->recorded_size, rv->size,
+                           rv->committed_size - rv->recorded_committed_size, 
+                           rv->committed_size,
+                           rv->filename);
+
+    if(printed >= buffer->size - buffer->len)
+      return;
+    buffer->len += printed;
+  }
+}
+
+// Write leaked regions to a file
+void DeepMemoryProfiler::WriteAllLeakyRegions() {
+  BufferArgs buffer;
+  buffer.buf = profiler_buffer_;
+  buffer.size = kProfilerBufferSize;
+  buffer.len = 0;
+  regions_->Iterate(WriteLeakyRegion, &buffer);
+
+  char leaks_file_name[1000];
+  snprintf(leaks_file_name, sizeof(leaks_file_name), "%s.%05d.%04d.leaks",
+           filename_prefix_, getpid(), dump_count_);
+
+  RawFD leaks_fd = RawOpenForWriting(leaks_file_name);
+  RAW_DCHECK(leaks_fd != kIllegalRawFD, "");
+  RawWrite(leaks_fd, buffer.buf, buffer.len);
+  RawClose(leaks_fd);
+}
+
+void DeepMemoryProfiler::WriteMapsToFile(char buf[], int size) {
+  char file_name[100];
+  snprintf(file_name, sizeof(file_name), "%s.%05d.maps", filename_prefix_, getpid());
+
+  RawFD maps_fd = RawOpenForWriting(file_name);
+  RAW_DCHECK(maps_fd != kIllegalRawFD, "");
+
+  int map_length;
+  bool dummy;   // "wrote_all" -- did /proc/self/maps fit in its entirety?
+  map_length = FillProcSelfMaps(profiler_buffer_, kProfilerBufferSize, &dummy);
+  RAW_DCHECK(map_length <= kProfilerBufferSize, "");
+  RawWrite(maps_fd, profiler_buffer_, map_length);
+  RawClose(maps_fd);
+}
+
+int DeepMemoryProfiler::WriteBucket(const Bucket* b, char buf[], int bufsize) {
+  int buflen = 0;
+  buflen += snprintf(buf + buflen, bufsize - buflen, "%05d", b->id);
+  for (int d = 0; d < b->depth; d++) {
+    buflen += snprintf(buf + buflen, bufsize - buflen, " 0x%08" PRIxPTR,
+                       reinterpret_cast<uintptr_t>(b->stack[d]));
+  }
+  buflen += snprintf(buf + buflen, bufsize - buflen, "\n");
+  return buflen;
+}
+
+void DeepMemoryProfiler::WriteBucketsToFile() {
+  char file_name[100];
+  snprintf(file_name, sizeof(file_name), "%s.%05d.%04d.buckets",
+           filename_prefix_, getpid(), dump_count_);
+  RawFD bucket_fd = RawOpenForWriting(file_name);
+  RAW_DCHECK(bucket_fd != kIllegalRawFD, "");
+
+  // We will use the global buffer here
+  char* buf = profiler_buffer_;
+  int size = kProfilerBufferSize;
+  int buflen = 0;
+  for (int b = 0; b < kHashTableSize; b++) {
+    for (Bucket* x = heap_profile_->table_[b]; x != 0; x = x->next) {
+      if(x->is_logged)
+        continue; // Skip the bucket if it is already logged
+      if(x->alloc_size - x->free_size <= 64)
+        continue; // Skip small buckets
+
+      buflen += WriteBucket(x, buf + buflen, size - buflen);
+      x->is_logged = true;
+
+      // Write to file if buffer 80% full
+      if(buflen > size * 0.8){
+        RawWrite(bucket_fd, buf, buflen);
+        buflen = 0;
+      }
+    }
+  }
+  RawWrite(bucket_fd, buf, buflen);
+  RawClose(bucket_fd);
+}
+
+int DeepMemoryProfiler::UnparseBucket(const Bucket& b,
+                                      char* buf, int buflen, int bufsize,
+                                      const char* extra,
+                                      Stats* profile_stats) {
+  if (profile_stats != NULL) {
+    profile_stats->allocs += b.allocs;
+    profile_stats->alloc_size += b.alloc_size;
+    profile_stats->frees += b.frees;
+    profile_stats->free_size += b.free_size;
+  }
+  int printed =
+      snprintf(buf + buflen, bufsize - buflen, 
+               "%10"PRId64" %10"PRId64" %6d %6d @%s %d\n",
+               b.alloc_size - b.free_size,
+               b.committed_size,
+               b.allocs, b.frees,
+               extra,
+               b.id);
+  // If it looks like the snprintf failed, ignore the fact we printed anything
+  if (printed < 0 || printed >= bufsize - buflen) return buflen;
+  buflen += printed;
+
+  return buflen;
+}
+
+int DeepMemoryProfiler::UnparseRegionStats(const RegionStats* stats, const char* name, 
+                                           char* buf, int buflen, int bufsize) {
+  int printed = snprintf(buf + buflen, bufsize - buflen, 
+                         "%15s %10"PRId64" %10"PRId64"\n",
+                         name,
+                         stats->virtual_bytes,
+                         stats->committed_bytes);
+
+  return buflen + printed;
+}
+
+int DeepMemoryProfiler::UnparseGlobalStats(char* buf, int buflen, int bufsize) {
+  buflen += snprintf(buf + buflen, bufsize - buflen, 
+                         "%15s %10s %10s\n",
+                         "", "virtual", "committed");
+
+  buflen = UnparseRegionStats(&(stats_.total), "total", buf, buflen, bufsize);
+  buflen = UnparseRegionStats(&(stats_.file_mapped), "file mapped", buf, buflen, bufsize);
+  buflen = UnparseRegionStats(&(stats_.anonymous), "anonymous", buf, buflen, bufsize);
+  buflen = UnparseRegionStats(&(stats_.other), "other", buf, buflen, bufsize);
+  buflen = UnparseRegionStats(&(stats_.record_mmap), "mmap", buf, buflen, bufsize);
+  buflen = UnparseRegionStats(&(stats_.record_tcmalloc), "tcmalloc", buf, buflen, bufsize);
+  return buflen;
+}
+
+// Takes snapshot of current memory usage.
+// We avoid any memory allocations during snapshots.
+//void DeepMemoryProfiler::TakeMemorySnapshot() {
+//}
+
+int DeepMemoryProfiler::FillOrderedProfile(char buf[], int size) {
+  int64 start_time = CycleClock::Now();
+  dump_count_++;
+
+  // We need to re-open files in /proc/pid/ if pid is created
+  if(most_recent_pid_ != getpid()) {
+    most_recent_pid_ = getpid();
+    OpenPageMap();
+    WriteMapsToFile(profiler_buffer_, kProfilerBufferSize);  // Using the global buffer
+  }
+
+
+  //regions_ = new(heap_profile_->alloc_(sizeof(RegionMap))) 
+  //    RegionMap(heap_profile_->alloc_, heap_profile_->dealloc_);
+  //max_region_size_ = 0;
+  GetGlobalStats();
+  uint64 anonymous_committed = stats_.anonymous.committed_bytes;
+
+  // Reset committed size of buckets
+  for (int b = 0; b < kHashTableSize; b++) {
+    for (Bucket* x = heap_profile_->table_[b]; x != 0; x = x->next) {
+      x->committed_size = 0;
+    }
+  }
+
+  //Bucket** list = heap_profile_->MakeBucketList();
+  //for (int i = 0; i < heap_profile_->num_buckets_; i++) {
+  //  list[i]->committed_size = 0;
+  //}
+
+  // No allocation zone starts -----------------------------------
+
+  // Record committed sizes
+  RecordAllAllocs();
+
+  // Check if committed bytes changed during RecordAllAllocs.
+  GetGlobalStats();
+  uint64 comm_diff = stats_.anonymous.committed_bytes - anonymous_committed;
+  if(comm_diff != 0)
+    RAW_LOG(0, "committed diff: %"PRId64"", comm_diff);
+
+  HeapProfileTable::Stats stats;
+  memset(&stats, 0, sizeof(stats));
+
+  int bucket_length = snprintf(buf, size, kProfileHeader);
+  if (bucket_length < 0 || bucket_length >= size) return 0;
+
+  // Printing Global Stats
+  bucket_length += snprintf(buf + bucket_length, size - bucket_length, kGlobalStatsHeader);
+  bucket_length = UnparseGlobalStats(buf, bucket_length, size);
+
+  // Printing Stacktraces
+  bucket_length += snprintf(buf + bucket_length, size - bucket_length, kStacktraceHeader);
+  bucket_length += snprintf(buf + bucket_length, size - bucket_length, 
+                            "%10s %10s\n", "virtual", "committed");
+
+  for (int b = 0; b < kHashTableSize; b++) {
+    for (Bucket* x = heap_profile_->table_[b]; x != 0; x = x->next) {
+      if(x->alloc_size - x->free_size == 0)
+        continue; // Skip empty buckets
+      bucket_length = UnparseBucket(*x, buf, bucket_length, size, "",
+                                    &stats);
+    }
+  }
+
+  RAW_DCHECK(bucket_length < size, "");
+  // No allocation zone ends -----------------------------------
+
+  //WriteAllLeakyRegions();
+  // Write stacktraces
+  WriteBucketsToFile();
+
+  //regions_->~RegionMap();
+  //heap_profile_->dealloc_(regions_);
+
+  int64 dt = CycleClock::Now() - start_time;
+  double dtf = dt / CyclesPerSecond();
+  RAW_LOG(0, "Time spent on DeepProfiler: %.3f sec\n", dtf);
+
+  return bucket_length;
+}