[NOT TO COMMIT!] Merge Chromium-specific changes in tcmalloc thru. the original gperftools r136.

third_party/tcmalloc/chromium/src/tcmalloc.cc

Index: third_party/tcmalloc/chromium/src/tcmalloc.cc
diff --git a/third_party/tcmalloc/chromium/src/tcmalloc.cc b/third_party/tcmalloc/chromium/src/tcmalloc.cc
index 425e90b9b240eec8c42c1677d228d67ff7e10931..89220e1b2fa8ae4c42bee82f19ea6211964f596d 100644
--- a/third_party/tcmalloc/chromium/src/tcmalloc.cc
+++ b/third_party/tcmalloc/chromium/src/tcmalloc.cc
@@ -119,8 +119,8 @@
 #include "base/spinlock.h"              // for SpinLockHolder
 #include "central_freelist.h"  // for CentralFreeListPadded
 #include "common.h"            // for StackTrace, kPageShift, etc
+#include "free_list.h"         // for FL_Init
 #include "internal_logging.h"  // for ASSERT, TCMalloc_Printer, etc
-#include "linked_list.h"       // for SLL_SetNext
 #include "malloc_hook-inl.h"       // for MallocHook::InvokeNewHook, etc
 #include "page_heap.h"         // for PageHeap, PageHeap::Stats
 #include "page_heap_allocator.h"  // for PageHeapAllocator
@@ -150,6 +150,13 @@
 // Some windows file somewhere (at least on cygwin) #define's small (!)
 #undef small
 
+// GLibc 2.14+ requires the hook functions be declared volatile, based on the
+// value of the define __MALLOC_HOOK_VOLATILE. For compatibility with
+// older/non-GLibc implementations, provide an empty definition.
+#if !defined(__MALLOC_HOOK_VOLATILE)
+#define __MALLOC_HOOK_VOLATILE
+#endif
+
 using STL_NAMESPACE::max;
 using STL_NAMESPACE::numeric_limits;
 using STL_NAMESPACE::vector;
@@ -177,6 +184,14 @@ using tcmalloc::StackTrace;
 using tcmalloc::Static;
 using tcmalloc::ThreadCache;
 
+// ---- Double free debug declarations
+static size_t ExcludeSpaceForMark(size_t size);
+static void AddRoomForMark(size_t* size);
+static void ExcludeMarkFromSize(size_t* new_size);
+static void MarkAllocatedRegion(void* ptr);
+static void ValidateAllocatedRegion(void* ptr, size_t cl);
+// ---- End Double free debug declarations
+
 DECLARE_int64(tcmalloc_sample_parameter);
 DECLARE_double(tcmalloc_release_rate);
 
@@ -211,7 +226,6 @@ DEFINE_int64(tcmalloc_large_alloc_report_threshold,
 // put all callers of MallocHook::Invoke* in this module into
 // ATTRIBUTE_SECTION(google_malloc) section, so that
 // MallocHook::GetCallerStackTrace can function accurately.
-#ifndef _WIN32   // windows doesn't have attribute_section, so don't bother
 extern "C" {
   void* tc_malloc(size_t size) __THROW
       ATTRIBUTE_SECTION(google_malloc);
@@ -273,7 +287,6 @@ extern "C" {
   size_t tc_malloc_size(void* p) __THROW
       ATTRIBUTE_SECTION(google_malloc);
 }  // extern "C"
-#endif  // #ifndef _WIN32
 
 // ----------------------- IMPLEMENTATION -------------------------------
 
@@ -373,6 +386,15 @@ static void DumpStats(TCMalloc_Printer* out, int level) {
                                         - stats.transfer_bytes
                                         - stats.thread_bytes);
 
+  out->printf(
+      "WASTE: %7.1f MiB committed but not used\n"
+      "WASTE: %7.1f MiB bytes committed, %7.1f MiB bytes in use\n"
+      "WASTE: committed/used ratio of %f\n",
+      (stats.pageheap.committed_bytes - bytes_in_use_by_app) / MiB,
+      stats.pageheap.committed_bytes / MiB,
+      bytes_in_use_by_app / MiB,
+      stats.pageheap.committed_bytes / static_cast<double>(bytes_in_use_by_app)
+      );
 #ifdef TCMALLOC_SMALL_BUT_SLOW
   out->printf(
       "NOTE:  SMALL MEMORY MODEL IS IN USE, PERFORMANCE MAY SUFFER.\n");
@@ -380,6 +402,7 @@ static void DumpStats(TCMalloc_Printer* out, int level) {
   out->printf(
       "------------------------------------------------\n"
       "MALLOC:   %12" PRIu64 " (%7.1f MiB) Bytes in use by application\n"
+      "MALLOC: %12" PRIu64 " (%7.1f MB) Bytes committed\n"
       "MALLOC: + %12" PRIu64 " (%7.1f MiB) Bytes in page heap freelist\n"
       "MALLOC: + %12" PRIu64 " (%7.1f MiB) Bytes in central cache freelist\n"
       "MALLOC: + %12" PRIu64 " (%7.1f MiB) Bytes in transfer cache freelist\n"
@@ -400,6 +423,7 @@ static void DumpStats(TCMalloc_Printer* out, int level) {
       "Bytes released to the OS take up virtual address space"
       " but no physical memory.\n",
       bytes_in_use_by_app, bytes_in_use_by_app / MiB,
+      stats.pageheap.committed_bytes, stats.pageheap.committed_bytes / MiB,
       stats.pageheap.free_bytes, stats.pageheap.free_bytes / MiB,
       stats.central_bytes, stats.central_bytes / MiB,
       stats.transfer_bytes, stats.transfer_bytes / MiB,
@@ -933,6 +957,7 @@ static inline bool CheckCachedSizeClass(void *ptr) {
 
 static inline void* CheckedMallocResult(void *result) {
   ASSERT(result == NULL || CheckCachedSizeClass(result));
+  MarkAllocatedRegion(result);
   return result;
 }
 
@@ -997,7 +1022,7 @@ static void ReportLargeAlloc(Length num_pages, void* result) {
 inline void* cpp_alloc(size_t size, bool nothrow);
 inline void* do_malloc(size_t size);
 
-// TODO(willchan): Investigate whether or not lining this much is harmful to
+// TODO(willchan): Investigate whether or not inlining this much is harmful to
 // performance.
 // This is equivalent to do_malloc() except when tc_new_mode is set to true.
 // Otherwise, it will run the std::new_handler if set.
@@ -1052,6 +1077,8 @@ inline void* do_malloc_pages(ThreadCache* heap, size_t size) {
 }
 
 inline void* do_malloc(size_t size) {
+  AddRoomForMark(&size);
+
   void* ret = NULL;
 
   // The following call forces module initialization
@@ -1062,13 +1089,15 @@ inline void* do_malloc(size_t size) {
 
     if ((FLAGS_tcmalloc_sample_parameter > 0) && heap->SampleAllocation(size)) {
       ret = DoSampledAllocation(size);
+      MarkAllocatedRegion(ret);
     } else {
-      // The common case, and also the simplest.  This just pops the
+      // The common case, and also the simplest. This just pops the
       // size-appropriate freelist, after replenishing it if it's empty.
       ret = CheckedMallocResult(heap->Allocate(size, cl));
     }
   } else {
     ret = do_malloc_pages(heap, size);
+    MarkAllocatedRegion(ret);
   }
   if (ret == NULL) errno = ENOMEM;
   return ret;
@@ -1124,6 +1153,9 @@ inline void do_free_with_callback(void* ptr, void (*invalid_free_fn)(void*)) {
     cl = span->sizeclass;
     Static::pageheap()->CacheSizeClass(p, cl);
   }
+
+  ValidateAllocatedRegion(ptr, cl);
+
   if (cl != 0) {
     ASSERT(!Static::pageheap()->GetDescriptor(p)->sample);
     ThreadCache* heap = GetCacheIfPresent();
@@ -1131,7 +1163,7 @@ inline void do_free_with_callback(void* ptr, void (*invalid_free_fn)(void*)) {
       heap->Deallocate(ptr, cl);
     } else {
       // Delete directly into central cache
-      tcmalloc::SLL_SetNext(ptr, NULL);
+      tcmalloc::FL_Init(ptr);
       Static::central_cache()[cl].InsertRange(ptr, ptr, 1);
     }
   } else {
@@ -1182,6 +1214,7 @@ inline void* do_realloc_with_callback(
     void* old_ptr, size_t new_size,
     void (*invalid_free_fn)(void*),
     size_t (*invalid_get_size_fn)(const void*)) {
+  AddRoomForMark(&new_size);
   // Get the size of the old entry
   const size_t old_size = GetSizeWithCallback(old_ptr, invalid_get_size_fn);
 
@@ -1200,6 +1233,7 @@ inline void* do_realloc_with_callback(
     if (new_size > old_size && new_size < lower_bound_to_grow) {
       new_ptr = do_malloc_or_cpp_alloc(lower_bound_to_grow);
     }
+    ExcludeMarkFromSize(&new_size);  // do_malloc will add space if needed.
     if (new_ptr == NULL) {
       // Either new_size is not a tiny increment, or last do_malloc failed.
       new_ptr = do_malloc_or_cpp_alloc(new_size);
@@ -1218,6 +1252,7 @@ inline void* do_realloc_with_callback(
   } else {
     // We still need to call hooks to report the updated size:
     MallocHook::InvokeDeleteHook(old_ptr);
+    ExcludeMarkFromSize(&new_size);
     MallocHook::InvokeNewHook(old_ptr, new_size);
     return old_ptr;
   }
@@ -1238,6 +1273,8 @@ inline void* do_realloc(void* old_ptr, size_t new_size) {
 void* do_memalign(size_t align, size_t size) {
   ASSERT((align & (align - 1)) == 0);
   ASSERT(align > 0);
+  // Marked in CheckMallocResult(), which is also inside SpanToMallocResult(). 
+  AddRoomForMark(&size);
   if (size + align < size) return NULL;         // Overflow
 
   // Fall back to malloc if we would already align this memory access properly.
@@ -1449,7 +1486,8 @@ void* cpp_memalign(size_t align, size_t size) {
 size_t TCMallocImplementation::GetAllocatedSize(const void* ptr) {
   ASSERT(TCMallocImplementation::GetOwnership(ptr)
          != TCMallocImplementation::kNotOwned);
-  return GetSizeWithCallback(ptr, &InvalidGetAllocatedSize);
+  return ExcludeSpaceForMark(
+      GetSizeWithCallback(ptr, &InvalidGetAllocatedSize));
 }
 
 void TCMallocImplementation::MarkThreadBusy() {
@@ -1650,3 +1688,172 @@ extern "C" PERFTOOLS_DLL_DECL size_t tc_malloc_size(void* ptr) __THROW {
 }
 
 #endif  // TCMALLOC_USING_DEBUGALLOCATION
+
+// ---Double free() debugging implementation -----------------------------------
+// We will put a mark at the extreme end of each allocation block.  We make
+// sure that we always allocate enough "extra memory" that we can fit in the
+// mark, and still provide the requested usable region.  If ever that mark is
+// not as expected, then we know that the user is corrupting memory beyond their
+// request size, or that they have called free a second time without having
+// the memory allocated (again).  This allows us to spot most double free()s,
+// but some can "slip by" or confuse our logic if the caller reallocates memory
+// (for a second use) before performing an evil double-free of a first
+// allocation
+
+// This code can be optimized, but for now, it is written to be most easily
+// understood, and flexible (since it is evolving a bit). Potential
+// optimizations include using other calculated data, such as class size, or
+// allocation size, which is known in the code above, but then is recalculated
+// below.  Another potential optimization would be careful manual inlining of
+// code, but I *think* that the compile will probably do this for me, and I've
+// been careful to avoid aliasing issues that might make a compiler back-off.
+
+// Evolution includes experimenting with different marks, to minimize the chance
+// that a mark would be misunderstood (missed corruption).  The marks are meant
+// to be hashed encoding of the location, so that they can't be copied over a
+// different region (by accident) without being detected (most of the time).
+
+// Enable the following define to turn on all the TCMalloc checking.
+// It will cost about 2% in performance, but it will catch double frees (most of
+// the time), and will often catch allocated-buffer overrun errors.  This
+// validation is only active when TCMalloc is used as the allocator.
+#ifndef NDEBUG
+#define TCMALLOC_VALIDATION
+#endif
+
+#if !defined(TCMALLOC_VALIDATION)
+
+static size_t ExcludeSpaceForMark(size_t size) { return size; }
+static void AddRoomForMark(size_t* size) {}
+static void ExcludeMarkFromSize(size_t* new_size) {}
+static void MarkAllocatedRegion(void* ptr) {}
+static void ValidateAllocatedRegion(void* ptr, size_t cl) {}
+
+#else  // TCMALLOC_VALIDATION
+
+static void DieFromDoubleFree() {
+  char* p = NULL;
+  p++;
+  *p += 1;  // Segv.
+}
+
+static size_t DieFromBadFreePointer(const void* unused) {
+  char* p = NULL;
+  p += 2;
+  *p += 2;  // Segv.
+  return 0;
+}
+
+static void DieFromMemoryCorruption() {
+  char* p = NULL;
+  p += 3;
+  *p += 3;  // Segv.
+}
+
+// We can either do byte marking, or whole word marking based on the following
+// define.  char is as small as we can get, and word marking probably provides
+// more than enough bits that we won't miss a corruption. Any sized integral
+// type can be used, but we just define two examples.
+
+//  #define TCMALLOC_SMALL_VALIDATION
+#if defined (TCMALLOC_SMALL_VALIDATION)
+
+typedef char MarkType;  // char saves memory... int is more complete.
+static const MarkType kAllocationMarkMask = static_cast<MarkType>(0x36);
+
+#else 
+
+typedef int MarkType;  // char saves memory... int is more complete.
+static const MarkType kAllocationMarkMask = static_cast<MarkType>(0xE1AB9536);
+
+#endif
+
+// TODO(jar): See if use of reference rather than pointer gets better inlining,
+// or if macro is needed.  My fear is that taking address map preclude register
+// allocation :-(.
+inline static void AddRoomForMark(size_t* size) {
+  *size += sizeof(kAllocationMarkMask);
+}
+
+inline static void ExcludeMarkFromSize(size_t* new_size) {
+  *new_size -= sizeof(kAllocationMarkMask);
+}
+
+inline static size_t ExcludeSpaceForMark(size_t size) {
+  return size - sizeof(kAllocationMarkMask);  // Lie about size when asked.
+}
+
+inline static MarkType* GetMarkLocation(void* ptr) {
+  size_t class_size = GetSizeWithCallback(ptr, DieFromBadFreePointer);
+  ASSERT(class_size % sizeof(kAllocationMarkMask) == 0);
+  size_t last_index = (class_size / sizeof(kAllocationMarkMask)) - 1;
+  return static_cast<MarkType*>(ptr) + last_index;
+}
+
+// We hash in the mark location plus the pointer so that we effectively mix in
+// the size of the block.  This means that if a span is used for different sizes
+// that the mark will be different. It would be good to hash in the size (which
+// we effectively get by using both mark location and pointer), but even better
+// would be to also include the class, as it concisely contains the entropy
+// found in the size (when we don't have large allocation), and there is less
+// risk of losing those bits to truncation. It would probably be good to combine
+// the high bits of size (capturing info about large blocks) with the class
+// (which is a 6 bit number).
+inline static MarkType GetMarkValue(void* ptr, MarkType* mark) {
+  void* ptr2 = static_cast<void*>(mark);
+  size_t offset1 = static_cast<char*>(ptr) - static_cast<char*>(NULL);
+  size_t offset2 = static_cast<char*>(ptr2) - static_cast<char*>(NULL);
+  static const int kInvariantBits = 2;
+  ASSERT((offset1 >> kInvariantBits) << kInvariantBits == offset1);
+  // Note: low bits of both offsets are invariants due to alignment.  High bits
+  // of both offsets are the same (unless we have a large allocation).  Avoid
+  // XORing high bits together, as they will cancel for most small allocations.
+
+  MarkType ret = kAllocationMarkMask;
+  // Using a little shift, we can safely XOR together both offsets.
+  ret ^= static_cast<MarkType>(offset1 >> kInvariantBits) ^
+         static_cast<MarkType>(offset2);
+  if (sizeof(ret) == 1) {
+    // Try to bring some high level bits into the mix.
+    ret += static_cast<MarkType>(offset1 >> 8) ^
+           static_cast<MarkType>(offset1 >> 16) ^
+           static_cast<MarkType>(offset1 >> 24) ;
+  }
+  // Hash in high bits on a 64 bit architecture.
+  if (sizeof(size_t) == 8 && sizeof(ret) == 4)
+    ret += offset1 >> 16;
+  if (ret == 0)
+    ret = kAllocationMarkMask;  // Avoid common pattern of all zeros.
+  return ret;
+}
+
+// TODO(jar): Use the passed in TCmalloc Class Index to calculate mark location
+// faster.  The current implementation calls general functions, which have to
+// recalculate this in order to get the Class Size.  This is a slow and wasteful
+// recomputation... but it is much more readable this way (for now).
+static void ValidateAllocatedRegion(void* ptr, size_t cl) {
+  if (ptr == NULL) return;
+  MarkType* mark = GetMarkLocation(ptr);
+  MarkType allocated_mark = GetMarkValue(ptr, mark);
+  MarkType current_mark = *mark;
+
+  if (current_mark == ~allocated_mark)
+    DieFromDoubleFree();
+  if (current_mark != allocated_mark)
+    DieFromMemoryCorruption();
+#ifndef NDEBUG
+  // In debug mode, copy the mark into all the free'd region.
+  size_t class_size = static_cast<size_t>(reinterpret_cast<char*>(mark) -
+                                          reinterpret_cast<char*>(ptr));
+  memset(ptr, static_cast<char>(0x36), class_size);
+#endif
+  *mark = ~allocated_mark;  //  Distinctively not allocated.
+}
+
+static void MarkAllocatedRegion(void* ptr) {
+  if (ptr == NULL) return;
+  MarkType* mark = GetMarkLocation(ptr);
+  *mark = GetMarkValue(ptr, mark);
+}
+
+#endif  // TCMALLOC_VALIDATION