From: James Sedgwick <jsedgwick@fb.com>
Date: Mon, 23 Oct 2017 14:27:04 +0000 (-0700)
Subject: move Arena, ThreadCachedArena, and Malloc to memory/
X-Git-Tag: v2017.10.30.00~31
X-Git-Url: http://plrg.eecs.uci.edu/git/?p=folly.git;a=commitdiff_plain;h=e8252dd7b5590427839eae4e4b2d76d2c4d9cff5

move Arena, ThreadCachedArena, and Malloc to memory/

Summary: all memory/-destined components besides hphp includes

Reviewed By: yfeldblum, mzlee

Differential Revision: D6121822

fbshipit-source-id: 6c6214d84dcdefe4789ed5200399ae27203d6340
---

diff --git a/CMakeLists.txt b/CMakeLists.txt
index d3b14ea6..530d6a2a 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -463,6 +463,10 @@ if (BUILD_TESTS)
     DIRECTORY io/async/ssl/test/
       TEST ssl_errors_test SOURCES SSLErrorsTest.cpp
 
+    DIRECTORY memory/test/
+      TEST thread_cached_arena_test SOURCES ThreadCachedArenaTest.cpp
+      TEST arena_test SOURCES ArenaTest.cpp
+
     DIRECTORY portability/test/
       TEST constexpr_test SOURCES ConstexprTest.cpp
       TEST libgen-test SOURCES LibgenTest.cpp
@@ -489,7 +493,6 @@ if (BUILD_TESTS)
 
     DIRECTORY test/
       TEST ahm_int_stress_test SOURCES AHMIntStressTest.cpp
-      TEST arena_test SOURCES ArenaTest.cpp
       TEST arena_smartptr_test SOURCES ArenaSmartPtrTest.cpp
       TEST array_test SOURCES ArrayTest.cpp
       TEST ascii_check_test SOURCES AsciiCaseInsensitiveTest.cpp
@@ -594,7 +597,6 @@ if (BUILD_TESTS)
       TEST sparse_byte_set_test SOURCES SparseByteSetTest.cpp
       TEST string_test SOURCES StringTest.cpp
       TEST synchronized_test SOURCES SynchronizedTest.cpp
-      TEST thread_cached_arena_test SOURCES ThreadCachedArenaTest.cpp
       TEST thread_cached_int_test SOURCES ThreadCachedIntTest.cpp
       TEST thread_local_test SOURCES ThreadLocalTest.cpp
       TEST timeout_queue_test SOURCES TimeoutQueueTest.cpp
diff --git a/folly/Arena-inl.h b/folly/Arena-inl.h
deleted file mode 100644
index 58059c9f..00000000
--- a/folly/Arena-inl.h
+++ /dev/null
@@ -1,93 +0,0 @@
-/*
- * Copyright 2017 Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#ifndef FOLLY_ARENA_H_
-#error This file may only be included from Arena.h
-#endif
-
-// Implementation of Arena.h functions
-
-namespace folly {
-
-template <class Alloc>
-std::pair<typename Arena<Alloc>::Block*, size_t>
-Arena<Alloc>::Block::allocate(Alloc& alloc, size_t size, bool allowSlack) {
-  size_t allocSize = sizeof(Block) + size;
-  if (allowSlack) {
-    allocSize = ArenaAllocatorTraits<Alloc>::goodSize(alloc, allocSize);
-  }
-
-  void* mem = alloc.allocate(allocSize);
-  return std::make_pair(new (mem) Block(), allocSize - sizeof(Block));
-}
-
-template <class Alloc>
-void Arena<Alloc>::Block::deallocate(Alloc& alloc) {
-  this->~Block();
-  alloc.deallocate(this);
-}
-
-template <class Alloc>
-void* Arena<Alloc>::allocateSlow(size_t size) {
-  std::pair<Block*, size_t> p;
-  char* start;
-
-  size_t allocSize = std::max(size, minBlockSize()) + sizeof(Block);
-  if (sizeLimit_ != kNoSizeLimit &&
-      allocSize > sizeLimit_ - totalAllocatedSize_) {
-    throw std::bad_alloc();
-  }
-
-  if (size > minBlockSize()) {
-    // Allocate a large block for this chunk only, put it at the back of the
-    // list so it doesn't get used for small allocations; don't change ptr_
-    // and end_, let them point into a normal block (or none, if they're
-    // null)
-    p = Block::allocate(alloc(), size, false);
-    start = p.first->start();
-    blocks_.push_back(*p.first);
-  } else {
-    // Allocate a normal sized block and carve out size bytes from it
-    p = Block::allocate(alloc(), minBlockSize(), true);
-    start = p.first->start();
-    blocks_.push_front(*p.first);
-    ptr_ = start + size;
-    end_ = start + p.second;
-  }
-
-  assert(p.second >= size);
-  totalAllocatedSize_ += p.second + sizeof(Block);
-  return start;
-}
-
-template <class Alloc>
-void Arena<Alloc>::merge(Arena<Alloc>&& other) {
-  blocks_.splice_after(blocks_.before_begin(), other.blocks_);
-  other.blocks_.clear();
-  other.ptr_ = other.end_ = nullptr;
-  totalAllocatedSize_ += other.totalAllocatedSize_;
-  other.totalAllocatedSize_ = 0;
-}
-
-template <class Alloc>
-Arena<Alloc>::~Arena() {
-  auto disposer = [this] (Block* b) { b->deallocate(this->alloc()); };
-  while (!blocks_.empty()) {
-    blocks_.pop_front_and_dispose(disposer);
-  }
-}
-
-} // namespace folly
diff --git a/folly/Arena.h b/folly/Arena.h
deleted file mode 100644
index 25279a7b..00000000
--- a/folly/Arena.h
+++ /dev/null
@@ -1,246 +0,0 @@
-/*
- * Copyright 2017 Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#pragma once
-#define FOLLY_ARENA_H_
-
-#include <cassert>
-#include <limits>
-#include <stdexcept>
-#include <utility>
-
-#include <boost/intrusive/slist.hpp>
-
-#include <folly/Conv.h>
-#include <folly/Likely.h>
-#include <folly/Malloc.h>
-#include <folly/Memory.h>
-
-namespace folly {
-
-/**
- * Simple arena: allocate memory which gets freed when the arena gets
- * destroyed.
- *
- * The arena itself allocates memory using a custom allocator which provides
- * the following interface (same as required by StlAllocator in StlAllocator.h)
- *
- *   void* allocate(size_t size);
- *      Allocate a block of size bytes, properly aligned to the maximum
- *      alignment required on your system; throw std::bad_alloc if the
- *      allocation can't be satisfied.
- *
- *   void deallocate(void* ptr);
- *      Deallocate a previously allocated block.
- *
- * You may also specialize ArenaAllocatorTraits for your allocator type to
- * provide:
- *
- *   size_t goodSize(const Allocator& alloc, size_t size) const;
- *      Return a size (>= the provided size) that is considered "good" for your
- *      allocator (for example, if your allocator allocates memory in 4MB
- *      chunks, size should be rounded up to 4MB).  The provided value is
- *      guaranteed to be rounded up to a multiple of the maximum alignment
- *      required on your system; the returned value must be also.
- *
- * An implementation that uses malloc() / free() is defined below, see SysArena.
- */
-template <class Alloc> struct ArenaAllocatorTraits;
-template <class Alloc>
-class Arena {
- public:
-  explicit Arena(const Alloc& alloc,
-                 size_t minBlockSize = kDefaultMinBlockSize,
-                 size_t sizeLimit = kNoSizeLimit,
-                 size_t maxAlign = kDefaultMaxAlign)
-    : allocAndSize_(alloc, minBlockSize)
-    , ptr_(nullptr)
-    , end_(nullptr)
-    , totalAllocatedSize_(0)
-    , bytesUsed_(0)
-    , sizeLimit_(sizeLimit)
-    , maxAlign_(maxAlign) {
-    if ((maxAlign_ & (maxAlign_ - 1)) || maxAlign_ > alignof(Block)) {
-      throw std::invalid_argument(
-          folly::to<std::string>("Invalid maxAlign: ", maxAlign_));
-    }
-  }
-
-  ~Arena();
-
-  void* allocate(size_t size) {
-    size = roundUp(size);
-    bytesUsed_ += size;
-
-    assert(ptr_ <= end_);
-    if (LIKELY((size_t)(end_ - ptr_) >= size)) {
-      // Fast path: there's enough room in the current block
-      char* r = ptr_;
-      ptr_ += size;
-      assert(isAligned(r));
-      return r;
-    }
-
-    // Not enough room in the current block
-    void* r = allocateSlow(size);
-    assert(isAligned(r));
-    return r;
-  }
-
-  void deallocate(void* /* p */) {
-    // Deallocate? Never!
-  }
-
-  // Transfer ownership of all memory allocated from "other" to "this".
-  void merge(Arena&& other);
-
-  // Gets the total memory used by the arena
-  size_t totalSize() const {
-    return totalAllocatedSize_ + sizeof(Arena);
-  }
-
-  // Gets the total number of "used" bytes, i.e. bytes that the arena users
-  // allocated via the calls to `allocate`. Doesn't include fragmentation, e.g.
-  // if block size is 4KB and you allocate 2 objects of 3KB in size,
-  // `bytesUsed()` will be 6KB, while `totalSize()` will be 8KB+.
-  size_t bytesUsed() const {
-    return bytesUsed_;
-  }
-
-  // not copyable
-  Arena(const Arena&) = delete;
-  Arena& operator=(const Arena&) = delete;
-
-  // movable
-  Arena(Arena&&) = default;
-  Arena& operator=(Arena&&) = default;
-
- private:
-  struct Block;
-  typedef boost::intrusive::slist_member_hook<
-    boost::intrusive::tag<Arena>> BlockLink;
-
-  struct FOLLY_ALIGNED_MAX Block {
-    BlockLink link;
-
-    // Allocate a block with at least size bytes of storage.
-    // If allowSlack is true, allocate more than size bytes if convenient
-    // (via ArenaAllocatorTraits::goodSize()) as we'll try to pack small
-    // allocations in this block.
-    static std::pair<Block*, size_t> allocate(
-        Alloc& alloc, size_t size, bool allowSlack);
-    void deallocate(Alloc& alloc);
-
-    char* start() {
-      return reinterpret_cast<char*>(this + 1);
-    }
-
-   private:
-    Block() = default;
-    ~Block() = default;
-  };
-
- public:
-  static constexpr size_t kDefaultMinBlockSize = 4096 - sizeof(Block);
-  static constexpr size_t kNoSizeLimit = 0;
-  static constexpr size_t kDefaultMaxAlign = alignof(Block);
-  static constexpr size_t kBlockOverhead = sizeof(Block);
-
- private:
-  bool isAligned(uintptr_t address) const {
-    return (address & (maxAlign_ - 1)) == 0;
-  }
-  bool isAligned(void* p) const {
-    return isAligned(reinterpret_cast<uintptr_t>(p));
-  }
-
-  // Round up size so it's properly aligned
-  size_t roundUp(size_t size) const {
-    return (size + maxAlign_ - 1) & ~(maxAlign_ - 1);
-  }
-
-  // cache_last<true> makes the list keep a pointer to the last element, so we
-  // have push_back() and constant time splice_after()
-  typedef boost::intrusive::slist<
-    Block,
-    boost::intrusive::member_hook<Block, BlockLink, &Block::link>,
-    boost::intrusive::constant_time_size<false>,
-    boost::intrusive::cache_last<true>> BlockList;
-
-  void* allocateSlow(size_t size);
-
-  // Empty member optimization: package Alloc with a non-empty member
-  // in case Alloc is empty (as it is in the case of SysAlloc).
-  struct AllocAndSize : public Alloc {
-    explicit AllocAndSize(const Alloc& a, size_t s)
-      : Alloc(a), minBlockSize(s) {
-    }
-
-    size_t minBlockSize;
-  };
-
-  size_t minBlockSize() const {
-    return allocAndSize_.minBlockSize;
-  }
-  Alloc& alloc() { return allocAndSize_; }
-  const Alloc& alloc() const { return allocAndSize_; }
-
-  AllocAndSize allocAndSize_;
-  BlockList blocks_;
-  char* ptr_;
-  char* end_;
-  size_t totalAllocatedSize_;
-  size_t bytesUsed_;
-  const size_t sizeLimit_;
-  const size_t maxAlign_;
-};
-
-template <class Alloc>
-struct IsArenaAllocator<Arena<Alloc>> : std::true_type { };
-
-/**
- * By default, don't pad the given size.
- */
-template <class Alloc>
-struct ArenaAllocatorTraits {
-  static size_t goodSize(const Alloc& /* alloc */, size_t size) { return size; }
-};
-
-template <>
-struct ArenaAllocatorTraits<SysAlloc> {
-  static size_t goodSize(const SysAlloc& /* alloc */, size_t size) {
-    return goodMallocSize(size);
-  }
-};
-
-/**
- * Arena that uses the system allocator (malloc / free)
- */
-class SysArena : public Arena<SysAlloc> {
- public:
-  explicit SysArena(size_t minBlockSize = kDefaultMinBlockSize,
-                    size_t sizeLimit = kNoSizeLimit,
-                    size_t maxAlign = kDefaultMaxAlign)
-    : Arena<SysAlloc>(SysAlloc(), minBlockSize, sizeLimit, maxAlign) {
-  }
-};
-
-template <>
-struct IsArenaAllocator<SysArena> : std::true_type { };
-
-} // namespace folly
-
-#include <folly/Arena-inl.h>
diff --git a/folly/Demangle.cpp b/folly/Demangle.cpp
index 12564693..d1f53cfc 100644
--- a/folly/Demangle.cpp
+++ b/folly/Demangle.cpp
@@ -19,7 +19,7 @@
 #include <algorithm>
 #include <cstring>
 
-#include <folly/Malloc.h>
+#include <folly/memory/Malloc.h>
 #include <folly/portability/Config.h>
 
 #if FOLLY_HAVE_CPLUS_DEMANGLE_V3_CALLBACK
diff --git a/folly/FBString.h b/folly/FBString.h
index af49ebe0..9c9f670b 100644
--- a/folly/FBString.h
+++ b/folly/FBString.h
@@ -55,8 +55,8 @@
 #include <utility>
 
 #include <folly/Hash.h>
-#include <folly/Malloc.h>
 #include <folly/Traits.h>
+#include <folly/memory/Malloc.h>
 #include <folly/portability/BitsFunctexcept.h>
 
 // When used in folly, assertions are not disabled.
diff --git a/folly/FBVector.h b/folly/FBVector.h
index c52c3501..dec58fc0 100644
--- a/folly/FBVector.h
+++ b/folly/FBVector.h
@@ -37,8 +37,8 @@
 
 #include <folly/FormatTraits.h>
 #include <folly/Likely.h>
-#include <folly/Malloc.h>
 #include <folly/Traits.h>
+#include <folly/memory/Malloc.h>
 #include <folly/portability/BitsFunctexcept.h>
 
 //=============================================================================
diff --git a/folly/Makefile.am b/folly/Makefile.am
index eac08817..b6ac6821 100644
--- a/folly/Makefile.am
+++ b/folly/Makefile.am
@@ -27,8 +27,6 @@ lib_LTLIBRARIES = \
 follyincludedir = $(includedir)/folly
 
 nobase_follyinclude_HEADERS = \
-	Arena.h \
-	Arena-inl.h \
 	Array.h \
 	Assume.h \
 	AtomicBitSet.h \
@@ -72,7 +70,6 @@ nobase_follyinclude_HEADERS = \
 	detail/GroupVarintDetail.h \
 	detail/IPAddress.h \
 	detail/IPAddressSource.h \
-	detail/MallocImpl.h \
 	detail/MemoryIdler.h \
 	detail/MPMCPipelineDetail.h \
 	detail/RangeCommon.h \
@@ -322,12 +319,16 @@ nobase_follyinclude_HEADERS = \
 	LockTraitsBoost.h \
 	Logging.h \
 	MacAddress.h \
-	Malloc.h \
 	MapUtil.h \
 	Math.h \
 	Memory.h \
+	memory/Arena.h \
+	memory/Arena-inl.h \
 	memory/MallctlHelper.h \
+	memory/Malloc.h \
+	memory/ThreadCachedArena.h \
 	memory/UninitializedMemoryHacks.h \
+	memory/detail/MallocImpl.h \
 	MicroSpinLock.h \
 	MicroLock.h \
 	MoveWrapper.h \
@@ -434,7 +435,6 @@ nobase_follyinclude_HEADERS = \
 	test/SynchronizedTestLib.h \
 	test/SynchronizedTestLib-inl.h \
 	test/TestUtils.h \
-	ThreadCachedArena.h \
 	ThreadCachedInt.h \
 	ThreadLocal.h \
 	TimeoutQueue.h \
@@ -562,6 +562,7 @@ libfolly_la_SOURCES = \
 	detail/MemoryIdler.cpp \
 	detail/SocketFastOpen.cpp \
 	MacAddress.cpp \
+	memory/ThreadCachedArena.cpp \
 	portability/Dirent.cpp \
 	portability/Fcntl.cpp \
 	portability/Libgen.cpp \
@@ -604,7 +605,6 @@ libfolly_la_SOURCES = \
 	system/ThreadName.cpp \
 	system/VersionCheck.cpp \
 	Subprocess.cpp \
-	ThreadCachedArena.cpp \
 	TimeoutQueue.cpp \
 	Try.cpp \
 	Uri.cpp \
@@ -633,7 +633,7 @@ libfolly_la_SOURCES += \
 endif
 
 if !HAVE_WEAK_SYMBOLS
-libfollybase_la_SOURCES += detail/MallocImpl.cpp
+libfollybase_la_SOURCES += memory/detail/MallocImpl.cpp
 endif
 
 if HAVE_BOOST_CONTEXT
diff --git a/folly/Malloc.h b/folly/Malloc.h
deleted file mode 100644
index 71f37ca2..00000000
--- a/folly/Malloc.h
+++ /dev/null
@@ -1,285 +0,0 @@
-/*
- * Copyright 2017 Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-// Functions to provide smarter use of jemalloc, if jemalloc is being used.
-// http://www.canonware.com/download/jemalloc/jemalloc-latest/doc/jemalloc.html
-
-#pragma once
-
-#include <folly/portability/Config.h>
-
-/**
- * Define various MALLOCX_* macros normally provided by jemalloc.  We define
- * them so that we don't have to include jemalloc.h, in case the program is
- * built without jemalloc support.
- */
-#if defined(USE_JEMALLOC) || defined(FOLLY_USE_JEMALLOC)
-// We have JEMalloc, so use it.
-# include <jemalloc/jemalloc.h>
-#else
-# ifndef MALLOCX_LG_ALIGN
-#  define MALLOCX_LG_ALIGN(la) (la)
-# endif
-# ifndef MALLOCX_ZERO
-#  define MALLOCX_ZERO (static_cast<int>(0x40))
-# endif
-#endif
-
-// If using fbstring from libstdc++ (see comment in FBString.h), then
-// just define stub code here to typedef the fbstring type into the
-// folly namespace.
-// This provides backwards compatibility for code that explicitly
-// includes and uses fbstring.
-#if defined(_GLIBCXX_USE_FB) && !defined(_LIBSTDCXX_FBSTRING)
-
-#include <folly/detail/MallocImpl.h>
-#include <folly/portability/BitsFunctexcept.h>
-
-#include <string>
-
-namespace folly {
-  using std::goodMallocSize;
-  using std::jemallocMinInPlaceExpandable;
-  using std::usingJEMalloc;
-  using std::smartRealloc;
-  using std::checkedMalloc;
-  using std::checkedCalloc;
-  using std::checkedRealloc;
-}
-
-#else // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
-
-#ifdef _LIBSTDCXX_FBSTRING
-#pragma GCC system_header
-
-/**
- * Declare *allocx() and mallctl*() as weak symbols. These will be provided by
- * jemalloc if we are using jemalloc, or will be nullptr if we are using another
- * malloc implementation.
- */
-extern "C" void* mallocx(size_t, int)
-__attribute__((__weak__));
-extern "C" void* rallocx(void*, size_t, int)
-__attribute__((__weak__));
-extern "C" size_t xallocx(void*, size_t, size_t, int)
-__attribute__((__weak__));
-extern "C" size_t sallocx(const void*, int)
-__attribute__((__weak__));
-extern "C" void dallocx(void*, int)
-__attribute__((__weak__));
-extern "C" void sdallocx(void*, size_t, int)
-__attribute__((__weak__));
-extern "C" size_t nallocx(size_t, int)
-__attribute__((__weak__));
-extern "C" int mallctl(const char*, void*, size_t*, void*, size_t)
-__attribute__((__weak__));
-extern "C" int mallctlnametomib(const char*, size_t*, size_t*)
-__attribute__((__weak__));
-extern "C" int mallctlbymib(const size_t*, size_t, void*, size_t*, void*,
-                            size_t)
-__attribute__((__weak__));
-
-#include <bits/functexcept.h>
-
-#define FOLLY_HAVE_MALLOC_H 1
-
-#else // !defined(_LIBSTDCXX_FBSTRING)
-
-#include <folly/detail/MallocImpl.h> /* nolint */
-#include <folly/portability/BitsFunctexcept.h> /* nolint */
-
-#endif
-
-// for malloc_usable_size
-// NOTE: FreeBSD 9 doesn't have malloc.h.  Its definitions
-// are found in stdlib.h.
-#if FOLLY_HAVE_MALLOC_H
-#include <malloc.h>
-#else
-#include <stdlib.h>
-#endif
-
-#include <cassert>
-#include <cstddef>
-#include <cstdint>
-#include <cstdlib>
-#include <cstring>
-
-#include <atomic>
-#include <new>
-
-#ifdef _LIBSTDCXX_FBSTRING
-namespace std _GLIBCXX_VISIBILITY(default) {
-_GLIBCXX_BEGIN_NAMESPACE_VERSION
-#else
-namespace folly {
-#endif
-
-// Cannot depend on Portability.h when _LIBSTDCXX_FBSTRING.
-#if defined(__GNUC__)
-#define FOLLY_MALLOC_NOINLINE __attribute__((__noinline__))
-#if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL) >= 40900
-// This is for checked malloc-like functions (returns non-null pointer
-// which cannot alias any outstanding pointer).
-#define FOLLY_MALLOC_CHECKED_MALLOC                     \
-  __attribute__((__returns_nonnull__, __malloc__))
-#else
-#define FOLLY_MALLOC_CHECKED_MALLOC __attribute__((__malloc__))
-#endif
-#else
-#define FOLLY_MALLOC_NOINLINE
-#define FOLLY_MALLOC_CHECKED_MALLOC
-#endif
-
-/**
- * Determine if we are using jemalloc or not.
- */
-FOLLY_MALLOC_NOINLINE inline bool usingJEMalloc() noexcept {
-  // Checking for rallocx != nullptr is not sufficient; we may be in a
-  // dlopen()ed module that depends on libjemalloc, so rallocx is resolved, but
-  // the main program might be using a different memory allocator.
-  // How do we determine that we're using jemalloc? In the hackiest
-  // way possible. We allocate memory using malloc() and see if the
-  // per-thread counter of allocated memory increases. This makes me
-  // feel dirty inside. Also note that this requires jemalloc to have
-  // been compiled with --enable-stats.
-  static const bool result = [] () noexcept {
-    // Some platforms (*cough* OSX *cough*) require weak symbol checks to be
-    // in the form if (mallctl != nullptr). Not if (mallctl) or if (!mallctl)
-    // (!!). http://goo.gl/xpmctm
-    if (mallocx == nullptr || rallocx == nullptr || xallocx == nullptr
-        || sallocx == nullptr || dallocx == nullptr || sdallocx == nullptr
-        || nallocx == nullptr || mallctl == nullptr
-        || mallctlnametomib == nullptr || mallctlbymib == nullptr) {
-      return false;
-    }
-
-    // "volatile" because gcc optimizes out the reads from *counter, because
-    // it "knows" malloc doesn't modify global state...
-    /* nolint */ volatile uint64_t* counter;
-    size_t counterLen = sizeof(uint64_t*);
-
-    if (mallctl("thread.allocatedp", static_cast<void*>(&counter), &counterLen,
-                nullptr, 0) != 0) {
-      return false;
-    }
-
-    if (counterLen != sizeof(uint64_t*)) {
-      return false;
-    }
-
-    uint64_t origAllocated = *counter;
-
-    const void* ptr = malloc(1);
-    if (!ptr) {
-      // wtf, failing to allocate 1 byte
-      return false;
-    }
-
-    /* Avoid optimizing away the malloc.  */
-    asm volatile("" ::"m"(ptr) : "memory");
-
-    return (origAllocated != *counter);
-  }();
-
-  return result;
-}
-
-inline size_t goodMallocSize(size_t minSize) noexcept {
-  if (minSize == 0) {
-    return 0;
-  }
-
-  if (!usingJEMalloc()) {
-    // Not using jemalloc - no smarts
-    return minSize;
-  }
-
-  return nallocx(minSize, 0);
-}
-
-// We always request "good" sizes for allocation, so jemalloc can
-// never grow in place small blocks; they're already occupied to the
-// brim.  Blocks larger than or equal to 4096 bytes can in fact be
-// expanded in place, and this constant reflects that.
-static const size_t jemallocMinInPlaceExpandable = 4096;
-
-/**
- * Trivial wrappers around malloc, calloc, realloc that check for allocation
- * failure and throw std::bad_alloc in that case.
- */
-inline void* checkedMalloc(size_t size) {
-  void* p = malloc(size);
-  if (!p) {
-    std::__throw_bad_alloc();
-  }
-  return p;
-}
-
-inline void* checkedCalloc(size_t n, size_t size) {
-  void* p = calloc(n, size);
-  if (!p) {
-    std::__throw_bad_alloc();
-  }
-  return p;
-}
-
-inline void* checkedRealloc(void* ptr, size_t size) {
-  void* p = realloc(ptr, size);
-  if (!p) {
-    std::__throw_bad_alloc();
-  }
-  return p;
-}
-
-/**
- * This function tries to reallocate a buffer of which only the first
- * currentSize bytes are used. The problem with using realloc is that
- * if currentSize is relatively small _and_ if realloc decides it
- * needs to move the memory chunk to a new buffer, then realloc ends
- * up copying data that is not used. It's generally not a win to try
- * to hook in to realloc() behavior to avoid copies - at least in
- * jemalloc, realloc() almost always ends up doing a copy, because
- * there is little fragmentation / slack space to take advantage of.
- */
-FOLLY_MALLOC_CHECKED_MALLOC FOLLY_MALLOC_NOINLINE inline void* smartRealloc(
-    void* p,
-    const size_t currentSize,
-    const size_t currentCapacity,
-    const size_t newCapacity) {
-  assert(p);
-  assert(currentSize <= currentCapacity &&
-         currentCapacity < newCapacity);
-
-  auto const slack = currentCapacity - currentSize;
-  if (slack * 2 > currentSize) {
-    // Too much slack, malloc-copy-free cycle:
-    auto const result = checkedMalloc(newCapacity);
-    std::memcpy(result, p, currentSize);
-    free(p);
-    return result;
-  }
-  // If there's not too much slack, we realloc in hope of coalescing
-  return checkedRealloc(p, newCapacity);
-}
-
-#ifdef _LIBSTDCXX_FBSTRING
-_GLIBCXX_END_NAMESPACE_VERSION
-#endif
-
-} // namespace folly
-
-#endif // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
diff --git a/folly/ThreadCachedArena.cpp b/folly/ThreadCachedArena.cpp
deleted file mode 100644
index 3869d6c2..00000000
--- a/folly/ThreadCachedArena.cpp
+++ /dev/null
@@ -1,53 +0,0 @@
-/*
- * Copyright 2017 Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <folly/ThreadCachedArena.h>
-
-#include <memory>
-
-namespace folly {
-
-ThreadCachedArena::ThreadCachedArena(size_t minBlockSize, size_t maxAlign)
-  : minBlockSize_(minBlockSize), maxAlign_(maxAlign) {
-}
-
-SysArena* ThreadCachedArena::allocateThreadLocalArena() {
-  SysArena* arena =
-    new SysArena(minBlockSize_, SysArena::kNoSizeLimit, maxAlign_);
-  auto disposer = [this] (SysArena* t, TLPDestructionMode mode) {
-    std::unique_ptr<SysArena> tp(t);  // ensure it gets deleted
-    if (mode == TLPDestructionMode::THIS_THREAD) {
-      zombify(std::move(*t));
-    }
-  };
-  arena_.reset(arena, disposer);
-  return arena;
-}
-
-void ThreadCachedArena::zombify(SysArena&& arena) {
-  zombies_->merge(std::move(arena));
-}
-
-size_t ThreadCachedArena::totalSize() const {
-  size_t result = sizeof(ThreadCachedArena);
-  for (const auto& arena : arena_.accessAllThreads()) {
-    result += arena.totalSize();
-  }
-  result += zombies_->totalSize() - sizeof(SysArena);
-  return result;
-}
-
-} // namespace folly
diff --git a/folly/ThreadCachedArena.h b/folly/ThreadCachedArena.h
deleted file mode 100644
index d4a5b23b..00000000
--- a/folly/ThreadCachedArena.h
+++ /dev/null
@@ -1,87 +0,0 @@
-/*
- * Copyright 2017 Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#pragma once
-
-#include <type_traits>
-
-#include <folly/Arena.h>
-#include <folly/Likely.h>
-#include <folly/Synchronized.h>
-#include <folly/ThreadLocal.h>
-
-namespace folly {
-
-/**
- * Thread-caching arena: allocate memory which gets freed when the arena gets
- * destroyed.
- *
- * The arena itself allocates memory using malloc() in blocks of
- * at least minBlockSize bytes.
- *
- * For speed, each thread gets its own Arena (see Arena.h); when threads
- * exit, the Arena gets merged into a "zombie" Arena, which will be deallocated
- * when the ThreadCachedArena object is destroyed.
- */
-class ThreadCachedArena {
- public:
-  explicit ThreadCachedArena(
-      size_t minBlockSize = SysArena::kDefaultMinBlockSize,
-      size_t maxAlign = SysArena::kDefaultMaxAlign);
-
-  void* allocate(size_t size) {
-    SysArena* arena = arena_.get();
-    if (UNLIKELY(!arena)) {
-      arena = allocateThreadLocalArena();
-    }
-
-    return arena->allocate(size);
-  }
-
-  void deallocate(void* /* p */) {
-    // Deallocate? Never!
-  }
-
-  // Gets the total memory used by the arena
-  size_t totalSize() const;
-
- private:
-  struct ThreadLocalPtrTag {};
-
-  ThreadCachedArena(const ThreadCachedArena&) = delete;
-  ThreadCachedArena(ThreadCachedArena&&) = delete;
-  ThreadCachedArena& operator=(const ThreadCachedArena&) = delete;
-  ThreadCachedArena& operator=(ThreadCachedArena&&) = delete;
-
-  SysArena* allocateThreadLocalArena();
-
-  // Zombify the blocks in arena, saving them for deallocation until
-  // the ThreadCachedArena is destroyed.
-  void zombify(SysArena&& arena);
-
-  const size_t minBlockSize_;
-  const size_t maxAlign_;
-
-  ThreadLocalPtr<SysArena, ThreadLocalPtrTag> arena_;  // Per-thread arena.
-
-  // Allocations from threads that are now dead.
-  Synchronized<SysArena> zombies_;
-};
-
-template <>
-struct IsArenaAllocator<ThreadCachedArena> : std::true_type { };
-
-} // namespace folly
diff --git a/folly/detail/MallocImpl.cpp b/folly/detail/MallocImpl.cpp
deleted file mode 100644
index 3daa120a..00000000
--- a/folly/detail/MallocImpl.cpp
+++ /dev/null
@@ -1,48 +0,0 @@
-/*
- * Copyright 2017 Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <folly/detail/MallocImpl.h>
-
-extern "C" {
-
-#ifdef _MSC_VER
-// MSVC doesn't have weak symbols, so do some linker magic
-// to emulate them. (the magic is in the header)
-const char* mallocxWeak = nullptr;
-const char* rallocxWeak = nullptr;
-const char* xallocxWeak = nullptr;
-const char* sallocxWeak = nullptr;
-const char* dallocxWeak = nullptr;
-const char* sdallocxWeak = nullptr;
-const char* nallocxWeak = nullptr;
-const char* mallctlWeak = nullptr;
-const char* mallctlnametomibWeak = nullptr;
-const char* mallctlbymibWeak = nullptr;
-#elif !FOLLY_HAVE_WEAK_SYMBOLS
-void* (*mallocx)(size_t, int) = nullptr;
-void* (*rallocx)(void*, size_t, int) = nullptr;
-size_t (*xallocx)(void*, size_t, size_t, int) = nullptr;
-size_t (*sallocx)(const void*, int) = nullptr;
-void (*dallocx)(void*, int) = nullptr;
-void (*sdallocx)(void*, size_t, int) = nullptr;
-size_t (*nallocx)(size_t, int) = nullptr;
-int (*mallctl)(const char*, void*, size_t*, void*, size_t) = nullptr;
-int (*mallctlnametomib)(const char*, size_t*, size_t*) = nullptr;
-int (*mallctlbymib)(const size_t*, size_t, void*, size_t*, void*, size_t) =
-    nullptr;
-#endif
-
-}
diff --git a/folly/detail/MallocImpl.h b/folly/detail/MallocImpl.h
deleted file mode 100644
index 74358f1c..00000000
--- a/folly/detail/MallocImpl.h
+++ /dev/null
@@ -1,66 +0,0 @@
-/*
- * Copyright 2017 Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#pragma once
-
-#include <stdlib.h>
-
-#include <folly/Portability.h>
-
-extern "C" {
-
-#if FOLLY_HAVE_WEAK_SYMBOLS
-void* mallocx(size_t, int) __attribute__((__weak__));
-void* rallocx(void*, size_t, int) __attribute__((__weak__));
-size_t xallocx(void*, size_t, size_t, int) __attribute__((__weak__));
-size_t sallocx(const void*, int) __attribute__((__weak__));
-void dallocx(void*, int) __attribute__((__weak__));
-void sdallocx(void*, size_t, int) __attribute__((__weak__));
-size_t nallocx(size_t, int) __attribute__((__weak__));
-int mallctl(const char*, void*, size_t*, void*, size_t)
-      __attribute__((__weak__));
-int mallctlnametomib(const char*, size_t*, size_t*) __attribute__((__weak__));
-int mallctlbymib(const size_t*, size_t, void*, size_t*, void*, size_t)
-      __attribute__((__weak__));
-#else
-extern void* (*mallocx)(size_t, int);
-extern void* (*rallocx)(void*, size_t, int);
-extern size_t (*xallocx)(void*, size_t, size_t, int);
-extern size_t (*sallocx)(const void*, int);
-extern void (*dallocx)(void*, int);
-extern void (*sdallocx)(void*, size_t, int);
-extern size_t (*nallocx)(size_t, int);
-extern int (*mallctl)(const char*, void*, size_t*, void*, size_t);
-extern int (*mallctlnametomib)(const char*, size_t*, size_t*);
-extern int (*mallctlbymib)(const size_t*, size_t, void*, size_t*, void*,
-                           size_t);
-#ifdef _MSC_VER
-// We emulate weak linkage for MSVC. The symbols we're
-// aliasing to are hiding in MallocImpl.cpp
-#pragma comment(linker, "/alternatename:mallocx=mallocxWeak")
-#pragma comment(linker, "/alternatename:rallocx=rallocxWeak")
-#pragma comment(linker, "/alternatename:xallocx=xallocxWeak")
-#pragma comment(linker, "/alternatename:sallocx=sallocxWeak")
-#pragma comment(linker, "/alternatename:dallocx=dallocxWeak")
-#pragma comment(linker, "/alternatename:sdallocx=sdallocxWeak")
-#pragma comment(linker, "/alternatename:nallocx=nallocxWeak")
-#pragma comment(linker, "/alternatename:mallctl=mallctlWeak")
-#pragma comment(linker, "/alternatename:mallctlnametomib=mallctlnametomibWeak")
-#pragma comment(linker, "/alternatename:mallctlbymib=mallctlbymibWeak")
-#endif
-#endif
-
-}
diff --git a/folly/detail/MemoryIdler.cpp b/folly/detail/MemoryIdler.cpp
index 0171939c..cef21a13 100644
--- a/folly/detail/MemoryIdler.cpp
+++ b/folly/detail/MemoryIdler.cpp
@@ -17,11 +17,11 @@
 #include <folly/detail/MemoryIdler.h>
 
 #include <folly/Logging.h>
-#include <folly/Malloc.h>
 #include <folly/Portability.h>
 #include <folly/ScopeGuard.h>
 #include <folly/concurrency/CacheLocality.h>
 #include <folly/memory/MallctlHelper.h>
+#include <folly/memory/Malloc.h>
 #include <folly/portability/PThread.h>
 #include <folly/portability/SysMman.h>
 #include <folly/portability/Unistd.h>
diff --git a/folly/detail/ThreadLocalDetail.h b/folly/detail/ThreadLocalDetail.h
index 363196b0..b0a2ae9f 100644
--- a/folly/detail/ThreadLocalDetail.h
+++ b/folly/detail/ThreadLocalDetail.h
@@ -29,11 +29,11 @@
 #include <folly/Exception.h>
 #include <folly/Foreach.h>
 #include <folly/Function.h>
-#include <folly/Malloc.h>
 #include <folly/MicroSpinLock.h>
 #include <folly/Portability.h>
 #include <folly/ScopeGuard.h>
 #include <folly/SharedMutex.h>
+#include <folly/memory/Malloc.h>
 #include <folly/portability/PThread.h>
 
 #include <folly/detail/StaticSingletonManager.h>
diff --git a/folly/experimental/JemallocNodumpAllocator.cpp b/folly/experimental/JemallocNodumpAllocator.cpp
index 91625a79..5824c11a 100644
--- a/folly/experimental/JemallocNodumpAllocator.cpp
+++ b/folly/experimental/JemallocNodumpAllocator.cpp
@@ -17,8 +17,8 @@
 #include <folly/experimental/JemallocNodumpAllocator.h>
 
 #include <folly/Conv.h>
-#include <folly/Malloc.h>
 #include <folly/String.h>
+#include <folly/memory/Malloc.h>
 #include <glog/logging.h>
 
 namespace folly {
diff --git a/folly/experimental/test/JemallocNodumpAllocatorTest.cpp b/folly/experimental/test/JemallocNodumpAllocatorTest.cpp
index 5c665031..cf51106c 100644
--- a/folly/experimental/test/JemallocNodumpAllocatorTest.cpp
+++ b/folly/experimental/test/JemallocNodumpAllocatorTest.cpp
@@ -16,8 +16,8 @@
 
 #include <folly/experimental/JemallocNodumpAllocator.h>
 
-#include <folly/Malloc.h>
 #include <folly/io/IOBuf.h>
+#include <folly/memory/Malloc.h>
 #include <folly/portability/GTest.h>
 
 TEST(JemallocNodumpAllocatorTest, Basic) {
diff --git a/folly/io/IOBuf.cpp b/folly/io/IOBuf.cpp
index 40f74e7c..7206fc46 100644
--- a/folly/io/IOBuf.cpp
+++ b/folly/io/IOBuf.cpp
@@ -27,11 +27,11 @@
 
 #include <folly/Conv.h>
 #include <folly/Likely.h>
-#include <folly/Malloc.h>
 #include <folly/Memory.h>
 #include <folly/ScopeGuard.h>
 #include <folly/hash/SpookyHashV2.h>
 #include <folly/io/Cursor.h>
+#include <folly/memory/Malloc.h>
 
 using std::unique_ptr;
 
@@ -791,7 +791,7 @@ void IOBuf::reserveSlow(uint64_t minHeadroom, uint64_t minTailroom) {
   //   - If using jemalloc, we can try to expand in place, avoiding a memcpy()
   //   - If not using jemalloc and we don't have too much to copy,
   //     we'll use realloc() (note that realloc might have to copy
-  //     headroom + data + tailroom, see smartRealloc in folly/Malloc.h)
+  //     headroom + data + tailroom, see smartRealloc in folly/memory/Malloc.h)
   // - Otherwise, bite the bullet and reallocate.
   if (headroom() + tailroom() >= minHeadroom + minTailroom) {
     uint8_t* newData = writableBuffer() + minHeadroom;
diff --git a/folly/io/TypedIOBuf.h b/folly/io/TypedIOBuf.h
index 86df07c4..382151e7 100644
--- a/folly/io/TypedIOBuf.h
+++ b/folly/io/TypedIOBuf.h
@@ -20,8 +20,8 @@
 #include <iterator>
 #include <type_traits>
 
-#include <folly/Malloc.h>
 #include <folly/io/IOBuf.h>
+#include <folly/memory/Malloc.h>
 
 namespace folly {
 
diff --git a/folly/io/test/IOBufTest.cpp b/folly/io/test/IOBufTest.cpp
index 9fd353f4..0235563d 100644
--- a/folly/io/test/IOBufTest.cpp
+++ b/folly/io/test/IOBufTest.cpp
@@ -21,8 +21,8 @@
 
 #include <boost/random.hpp>
 
-#include <folly/Malloc.h>
 #include <folly/Range.h>
+#include <folly/memory/Malloc.h>
 #include <folly/portability/GTest.h>
 
 using folly::fbstring;
diff --git a/folly/memory/Arena-inl.h b/folly/memory/Arena-inl.h
new file mode 100644
index 00000000..58059c9f
--- /dev/null
+++ b/folly/memory/Arena-inl.h
@@ -0,0 +1,93 @@
+/*
+ * Copyright 2017 Facebook, Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FOLLY_ARENA_H_
+#error This file may only be included from Arena.h
+#endif
+
+// Implementation of Arena.h functions
+
+namespace folly {
+
+template <class Alloc>
+std::pair<typename Arena<Alloc>::Block*, size_t>
+Arena<Alloc>::Block::allocate(Alloc& alloc, size_t size, bool allowSlack) {
+  size_t allocSize = sizeof(Block) + size;
+  if (allowSlack) {
+    allocSize = ArenaAllocatorTraits<Alloc>::goodSize(alloc, allocSize);
+  }
+
+  void* mem = alloc.allocate(allocSize);
+  return std::make_pair(new (mem) Block(), allocSize - sizeof(Block));
+}
+
+template <class Alloc>
+void Arena<Alloc>::Block::deallocate(Alloc& alloc) {
+  this->~Block();
+  alloc.deallocate(this);
+}
+
+template <class Alloc>
+void* Arena<Alloc>::allocateSlow(size_t size) {
+  std::pair<Block*, size_t> p;
+  char* start;
+
+  size_t allocSize = std::max(size, minBlockSize()) + sizeof(Block);
+  if (sizeLimit_ != kNoSizeLimit &&
+      allocSize > sizeLimit_ - totalAllocatedSize_) {
+    throw std::bad_alloc();
+  }
+
+  if (size > minBlockSize()) {
+    // Allocate a large block for this chunk only, put it at the back of the
+    // list so it doesn't get used for small allocations; don't change ptr_
+    // and end_, let them point into a normal block (or none, if they're
+    // null)
+    p = Block::allocate(alloc(), size, false);
+    start = p.first->start();
+    blocks_.push_back(*p.first);
+  } else {
+    // Allocate a normal sized block and carve out size bytes from it
+    p = Block::allocate(alloc(), minBlockSize(), true);
+    start = p.first->start();
+    blocks_.push_front(*p.first);
+    ptr_ = start + size;
+    end_ = start + p.second;
+  }
+
+  assert(p.second >= size);
+  totalAllocatedSize_ += p.second + sizeof(Block);
+  return start;
+}
+
+template <class Alloc>
+void Arena<Alloc>::merge(Arena<Alloc>&& other) {
+  blocks_.splice_after(blocks_.before_begin(), other.blocks_);
+  other.blocks_.clear();
+  other.ptr_ = other.end_ = nullptr;
+  totalAllocatedSize_ += other.totalAllocatedSize_;
+  other.totalAllocatedSize_ = 0;
+}
+
+template <class Alloc>
+Arena<Alloc>::~Arena() {
+  auto disposer = [this] (Block* b) { b->deallocate(this->alloc()); };
+  while (!blocks_.empty()) {
+    blocks_.pop_front_and_dispose(disposer);
+  }
+}
+
+} // namespace folly
diff --git a/folly/memory/Arena.h b/folly/memory/Arena.h
new file mode 100644
index 00000000..ffe9f38b
--- /dev/null
+++ b/folly/memory/Arena.h
@@ -0,0 +1,246 @@
+/*
+ * Copyright 2017 Facebook, Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+#define FOLLY_ARENA_H_
+
+#include <cassert>
+#include <limits>
+#include <stdexcept>
+#include <utility>
+
+#include <boost/intrusive/slist.hpp>
+
+#include <folly/Conv.h>
+#include <folly/Likely.h>
+#include <folly/Memory.h>
+#include <folly/memory/Malloc.h>
+
+namespace folly {
+
+/**
+ * Simple arena: allocate memory which gets freed when the arena gets
+ * destroyed.
+ *
+ * The arena itself allocates memory using a custom allocator which provides
+ * the following interface (same as required by StlAllocator in StlAllocator.h)
+ *
+ *   void* allocate(size_t size);
+ *      Allocate a block of size bytes, properly aligned to the maximum
+ *      alignment required on your system; throw std::bad_alloc if the
+ *      allocation can't be satisfied.
+ *
+ *   void deallocate(void* ptr);
+ *      Deallocate a previously allocated block.
+ *
+ * You may also specialize ArenaAllocatorTraits for your allocator type to
+ * provide:
+ *
+ *   size_t goodSize(const Allocator& alloc, size_t size) const;
+ *      Return a size (>= the provided size) that is considered "good" for your
+ *      allocator (for example, if your allocator allocates memory in 4MB
+ *      chunks, size should be rounded up to 4MB).  The provided value is
+ *      guaranteed to be rounded up to a multiple of the maximum alignment
+ *      required on your system; the returned value must be also.
+ *
+ * An implementation that uses malloc() / free() is defined below, see SysArena.
+ */
+template <class Alloc> struct ArenaAllocatorTraits;
+template <class Alloc>
+class Arena {
+ public:
+  explicit Arena(const Alloc& alloc,
+                 size_t minBlockSize = kDefaultMinBlockSize,
+                 size_t sizeLimit = kNoSizeLimit,
+                 size_t maxAlign = kDefaultMaxAlign)
+    : allocAndSize_(alloc, minBlockSize)
+    , ptr_(nullptr)
+    , end_(nullptr)
+    , totalAllocatedSize_(0)
+    , bytesUsed_(0)
+    , sizeLimit_(sizeLimit)
+    , maxAlign_(maxAlign) {
+    if ((maxAlign_ & (maxAlign_ - 1)) || maxAlign_ > alignof(Block)) {
+      throw std::invalid_argument(
+          folly::to<std::string>("Invalid maxAlign: ", maxAlign_));
+    }
+  }
+
+  ~Arena();
+
+  void* allocate(size_t size) {
+    size = roundUp(size);
+    bytesUsed_ += size;
+
+    assert(ptr_ <= end_);
+    if (LIKELY((size_t)(end_ - ptr_) >= size)) {
+      // Fast path: there's enough room in the current block
+      char* r = ptr_;
+      ptr_ += size;
+      assert(isAligned(r));
+      return r;
+    }
+
+    // Not enough room in the current block
+    void* r = allocateSlow(size);
+    assert(isAligned(r));
+    return r;
+  }
+
+  void deallocate(void* /* p */) {
+    // Deallocate? Never!
+  }
+
+  // Transfer ownership of all memory allocated from "other" to "this".
+  void merge(Arena&& other);
+
+  // Gets the total memory used by the arena
+  size_t totalSize() const {
+    return totalAllocatedSize_ + sizeof(Arena);
+  }
+
+  // Gets the total number of "used" bytes, i.e. bytes that the arena users
+  // allocated via the calls to `allocate`. Doesn't include fragmentation, e.g.
+  // if block size is 4KB and you allocate 2 objects of 3KB in size,
+  // `bytesUsed()` will be 6KB, while `totalSize()` will be 8KB+.
+  size_t bytesUsed() const {
+    return bytesUsed_;
+  }
+
+  // not copyable
+  Arena(const Arena&) = delete;
+  Arena& operator=(const Arena&) = delete;
+
+  // movable
+  Arena(Arena&&) = default;
+  Arena& operator=(Arena&&) = default;
+
+ private:
+  struct Block;
+  typedef boost::intrusive::slist_member_hook<
+    boost::intrusive::tag<Arena>> BlockLink;
+
+  struct FOLLY_ALIGNED_MAX Block {
+    BlockLink link;
+
+    // Allocate a block with at least size bytes of storage.
+    // If allowSlack is true, allocate more than size bytes if convenient
+    // (via ArenaAllocatorTraits::goodSize()) as we'll try to pack small
+    // allocations in this block.
+    static std::pair<Block*, size_t> allocate(
+        Alloc& alloc, size_t size, bool allowSlack);
+    void deallocate(Alloc& alloc);
+
+    char* start() {
+      return reinterpret_cast<char*>(this + 1);
+    }
+
+   private:
+    Block() = default;
+    ~Block() = default;
+  };
+
+ public:
+  static constexpr size_t kDefaultMinBlockSize = 4096 - sizeof(Block);
+  static constexpr size_t kNoSizeLimit = 0;
+  static constexpr size_t kDefaultMaxAlign = alignof(Block);
+  static constexpr size_t kBlockOverhead = sizeof(Block);
+
+ private:
+  bool isAligned(uintptr_t address) const {
+    return (address & (maxAlign_ - 1)) == 0;
+  }
+  bool isAligned(void* p) const {
+    return isAligned(reinterpret_cast<uintptr_t>(p));
+  }
+
+  // Round up size so it's properly aligned
+  size_t roundUp(size_t size) const {
+    return (size + maxAlign_ - 1) & ~(maxAlign_ - 1);
+  }
+
+  // cache_last<true> makes the list keep a pointer to the last element, so we
+  // have push_back() and constant time splice_after()
+  typedef boost::intrusive::slist<
+    Block,
+    boost::intrusive::member_hook<Block, BlockLink, &Block::link>,
+    boost::intrusive::constant_time_size<false>,
+    boost::intrusive::cache_last<true>> BlockList;
+
+  void* allocateSlow(size_t size);
+
+  // Empty member optimization: package Alloc with a non-empty member
+  // in case Alloc is empty (as it is in the case of SysAlloc).
+  struct AllocAndSize : public Alloc {
+    explicit AllocAndSize(const Alloc& a, size_t s)
+      : Alloc(a), minBlockSize(s) {
+    }
+
+    size_t minBlockSize;
+  };
+
+  size_t minBlockSize() const {
+    return allocAndSize_.minBlockSize;
+  }
+  Alloc& alloc() { return allocAndSize_; }
+  const Alloc& alloc() const { return allocAndSize_; }
+
+  AllocAndSize allocAndSize_;
+  BlockList blocks_;
+  char* ptr_;
+  char* end_;
+  size_t totalAllocatedSize_;
+  size_t bytesUsed_;
+  const size_t sizeLimit_;
+  const size_t maxAlign_;
+};
+
+template <class Alloc>
+struct IsArenaAllocator<Arena<Alloc>> : std::true_type { };
+
+/**
+ * By default, don't pad the given size.
+ */
+template <class Alloc>
+struct ArenaAllocatorTraits {
+  static size_t goodSize(const Alloc& /* alloc */, size_t size) { return size; }
+};
+
+template <>
+struct ArenaAllocatorTraits<SysAlloc> {
+  static size_t goodSize(const SysAlloc& /* alloc */, size_t size) {
+    return goodMallocSize(size);
+  }
+};
+
+/**
+ * Arena that uses the system allocator (malloc / free)
+ */
+class SysArena : public Arena<SysAlloc> {
+ public:
+  explicit SysArena(size_t minBlockSize = kDefaultMinBlockSize,
+                    size_t sizeLimit = kNoSizeLimit,
+                    size_t maxAlign = kDefaultMaxAlign)
+    : Arena<SysAlloc>(SysAlloc(), minBlockSize, sizeLimit, maxAlign) {
+  }
+};
+
+template <>
+struct IsArenaAllocator<SysArena> : std::true_type { };
+
+} // namespace folly
+
+#include <folly/memory/Arena-inl.h>
diff --git a/folly/memory/MallctlHelper.h b/folly/memory/MallctlHelper.h
index f883a5f6..7a6b200d 100644
--- a/folly/memory/MallctlHelper.h
+++ b/folly/memory/MallctlHelper.h
@@ -19,7 +19,7 @@
 #pragma once
 
 #include <folly/Likely.h>
-#include <folly/Malloc.h>
+#include <folly/memory/Malloc.h>
 
 #include <stdexcept>
 
diff --git a/folly/memory/Malloc.h b/folly/memory/Malloc.h
new file mode 100644
index 00000000..2726c9d8
--- /dev/null
+++ b/folly/memory/Malloc.h
@@ -0,0 +1,285 @@
+/*
+ * Copyright 2017 Facebook, Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// Functions to provide smarter use of jemalloc, if jemalloc is being used.
+// http://www.canonware.com/download/jemalloc/jemalloc-latest/doc/jemalloc.html
+
+#pragma once
+
+#include <folly/portability/Config.h>
+
+/**
+ * Define various MALLOCX_* macros normally provided by jemalloc.  We define
+ * them so that we don't have to include jemalloc.h, in case the program is
+ * built without jemalloc support.
+ */
+#if defined(USE_JEMALLOC) || defined(FOLLY_USE_JEMALLOC)
+// We have JEMalloc, so use it.
+# include <jemalloc/jemalloc.h>
+#else
+# ifndef MALLOCX_LG_ALIGN
+#  define MALLOCX_LG_ALIGN(la) (la)
+# endif
+# ifndef MALLOCX_ZERO
+#  define MALLOCX_ZERO (static_cast<int>(0x40))
+# endif
+#endif
+
+// If using fbstring from libstdc++ (see comment in FBString.h), then
+// just define stub code here to typedef the fbstring type into the
+// folly namespace.
+// This provides backwards compatibility for code that explicitly
+// includes and uses fbstring.
+#if defined(_GLIBCXX_USE_FB) && !defined(_LIBSTDCXX_FBSTRING)
+
+#include <folly/memory/detail/MallocImpl.h>
+#include <folly/portability/BitsFunctexcept.h>
+
+#include <string>
+
+namespace folly {
+  using std::goodMallocSize;
+  using std::jemallocMinInPlaceExpandable;
+  using std::usingJEMalloc;
+  using std::smartRealloc;
+  using std::checkedMalloc;
+  using std::checkedCalloc;
+  using std::checkedRealloc;
+}
+
+#else // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
+
+#ifdef _LIBSTDCXX_FBSTRING
+#pragma GCC system_header
+
+/**
+ * Declare *allocx() and mallctl*() as weak symbols. These will be provided by
+ * jemalloc if we are using jemalloc, or will be nullptr if we are using another
+ * malloc implementation.
+ */
+extern "C" void* mallocx(size_t, int)
+__attribute__((__weak__));
+extern "C" void* rallocx(void*, size_t, int)
+__attribute__((__weak__));
+extern "C" size_t xallocx(void*, size_t, size_t, int)
+__attribute__((__weak__));
+extern "C" size_t sallocx(const void*, int)
+__attribute__((__weak__));
+extern "C" void dallocx(void*, int)
+__attribute__((__weak__));
+extern "C" void sdallocx(void*, size_t, int)
+__attribute__((__weak__));
+extern "C" size_t nallocx(size_t, int)
+__attribute__((__weak__));
+extern "C" int mallctl(const char*, void*, size_t*, void*, size_t)
+__attribute__((__weak__));
+extern "C" int mallctlnametomib(const char*, size_t*, size_t*)
+__attribute__((__weak__));
+extern "C" int mallctlbymib(const size_t*, size_t, void*, size_t*, void*,
+                            size_t)
+__attribute__((__weak__));
+
+#include <bits/functexcept.h>
+
+#define FOLLY_HAVE_MALLOC_H 1
+
+#else // !defined(_LIBSTDCXX_FBSTRING)
+
+#include <folly/memory/detail/MallocImpl.h> /* nolint */
+#include <folly/portability/BitsFunctexcept.h> /* nolint */
+
+#endif
+
+// for malloc_usable_size
+// NOTE: FreeBSD 9 doesn't have malloc.h.  Its definitions
+// are found in stdlib.h.
+#if FOLLY_HAVE_MALLOC_H
+#include <malloc.h>
+#else
+#include <stdlib.h>
+#endif
+
+#include <cassert>
+#include <cstddef>
+#include <cstdint>
+#include <cstdlib>
+#include <cstring>
+
+#include <atomic>
+#include <new>
+
+#ifdef _LIBSTDCXX_FBSTRING
+namespace std _GLIBCXX_VISIBILITY(default) {
+_GLIBCXX_BEGIN_NAMESPACE_VERSION
+#else
+namespace folly {
+#endif
+
+// Cannot depend on Portability.h when _LIBSTDCXX_FBSTRING.
+#if defined(__GNUC__)
+#define FOLLY_MALLOC_NOINLINE __attribute__((__noinline__))
+#if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL) >= 40900
+// This is for checked malloc-like functions (returns non-null pointer
+// which cannot alias any outstanding pointer).
+#define FOLLY_MALLOC_CHECKED_MALLOC                     \
+  __attribute__((__returns_nonnull__, __malloc__))
+#else
+#define FOLLY_MALLOC_CHECKED_MALLOC __attribute__((__malloc__))
+#endif
+#else
+#define FOLLY_MALLOC_NOINLINE
+#define FOLLY_MALLOC_CHECKED_MALLOC
+#endif
+
+/**
+ * Determine if we are using jemalloc or not.
+ */
+FOLLY_MALLOC_NOINLINE inline bool usingJEMalloc() noexcept {
+  // Checking for rallocx != nullptr is not sufficient; we may be in a
+  // dlopen()ed module that depends on libjemalloc, so rallocx is resolved, but
+  // the main program might be using a different memory allocator.
+  // How do we determine that we're using jemalloc? In the hackiest
+  // way possible. We allocate memory using malloc() and see if the
+  // per-thread counter of allocated memory increases. This makes me
+  // feel dirty inside. Also note that this requires jemalloc to have
+  // been compiled with --enable-stats.
+  static const bool result = [] () noexcept {
+    // Some platforms (*cough* OSX *cough*) require weak symbol checks to be
+    // in the form if (mallctl != nullptr). Not if (mallctl) or if (!mallctl)
+    // (!!). http://goo.gl/xpmctm
+    if (mallocx == nullptr || rallocx == nullptr || xallocx == nullptr
+        || sallocx == nullptr || dallocx == nullptr || sdallocx == nullptr
+        || nallocx == nullptr || mallctl == nullptr
+        || mallctlnametomib == nullptr || mallctlbymib == nullptr) {
+      return false;
+    }
+
+    // "volatile" because gcc optimizes out the reads from *counter, because
+    // it "knows" malloc doesn't modify global state...
+    /* nolint */ volatile uint64_t* counter;
+    size_t counterLen = sizeof(uint64_t*);
+
+    if (mallctl("thread.allocatedp", static_cast<void*>(&counter), &counterLen,
+                nullptr, 0) != 0) {
+      return false;
+    }
+
+    if (counterLen != sizeof(uint64_t*)) {
+      return false;
+    }
+
+    uint64_t origAllocated = *counter;
+
+    const void* ptr = malloc(1);
+    if (!ptr) {
+      // wtf, failing to allocate 1 byte
+      return false;
+    }
+
+    /* Avoid optimizing away the malloc.  */
+    asm volatile("" ::"m"(ptr) : "memory");
+
+    return (origAllocated != *counter);
+  }();
+
+  return result;
+}
+
+inline size_t goodMallocSize(size_t minSize) noexcept {
+  if (minSize == 0) {
+    return 0;
+  }
+
+  if (!usingJEMalloc()) {
+    // Not using jemalloc - no smarts
+    return minSize;
+  }
+
+  return nallocx(minSize, 0);
+}
+
+// We always request "good" sizes for allocation, so jemalloc can
+// never grow in place small blocks; they're already occupied to the
+// brim.  Blocks larger than or equal to 4096 bytes can in fact be
+// expanded in place, and this constant reflects that.
+static const size_t jemallocMinInPlaceExpandable = 4096;
+
+/**
+ * Trivial wrappers around malloc, calloc, realloc that check for allocation
+ * failure and throw std::bad_alloc in that case.
+ */
+inline void* checkedMalloc(size_t size) {
+  void* p = malloc(size);
+  if (!p) {
+    std::__throw_bad_alloc();
+  }
+  return p;
+}
+
+inline void* checkedCalloc(size_t n, size_t size) {
+  void* p = calloc(n, size);
+  if (!p) {
+    std::__throw_bad_alloc();
+  }
+  return p;
+}
+
+inline void* checkedRealloc(void* ptr, size_t size) {
+  void* p = realloc(ptr, size);
+  if (!p) {
+    std::__throw_bad_alloc();
+  }
+  return p;
+}
+
+/**
+ * This function tries to reallocate a buffer of which only the first
+ * currentSize bytes are used. The problem with using realloc is that
+ * if currentSize is relatively small _and_ if realloc decides it
+ * needs to move the memory chunk to a new buffer, then realloc ends
+ * up copying data that is not used. It's generally not a win to try
+ * to hook in to realloc() behavior to avoid copies - at least in
+ * jemalloc, realloc() almost always ends up doing a copy, because
+ * there is little fragmentation / slack space to take advantage of.
+ */
+FOLLY_MALLOC_CHECKED_MALLOC FOLLY_MALLOC_NOINLINE inline void* smartRealloc(
+    void* p,
+    const size_t currentSize,
+    const size_t currentCapacity,
+    const size_t newCapacity) {
+  assert(p);
+  assert(currentSize <= currentCapacity &&
+         currentCapacity < newCapacity);
+
+  auto const slack = currentCapacity - currentSize;
+  if (slack * 2 > currentSize) {
+    // Too much slack, malloc-copy-free cycle:
+    auto const result = checkedMalloc(newCapacity);
+    std::memcpy(result, p, currentSize);
+    free(p);
+    return result;
+  }
+  // If there's not too much slack, we realloc in hope of coalescing
+  return checkedRealloc(p, newCapacity);
+}
+
+#ifdef _LIBSTDCXX_FBSTRING
+_GLIBCXX_END_NAMESPACE_VERSION
+#endif
+
+} // namespace folly
+
+#endif // !defined(_GLIBCXX_USE_FB) || defined(_LIBSTDCXX_FBSTRING)
diff --git a/folly/memory/ThreadCachedArena.cpp b/folly/memory/ThreadCachedArena.cpp
new file mode 100644
index 00000000..f8764642
--- /dev/null
+++ b/folly/memory/ThreadCachedArena.cpp
@@ -0,0 +1,53 @@
+/*
+ * Copyright 2017 Facebook, Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <folly/memory/ThreadCachedArena.h>
+
+#include <memory>
+
+namespace folly {
+
+ThreadCachedArena::ThreadCachedArena(size_t minBlockSize, size_t maxAlign)
+  : minBlockSize_(minBlockSize), maxAlign_(maxAlign) {
+}
+
+SysArena* ThreadCachedArena::allocateThreadLocalArena() {
+  SysArena* arena =
+    new SysArena(minBlockSize_, SysArena::kNoSizeLimit, maxAlign_);
+  auto disposer = [this] (SysArena* t, TLPDestructionMode mode) {
+    std::unique_ptr<SysArena> tp(t);  // ensure it gets deleted
+    if (mode == TLPDestructionMode::THIS_THREAD) {
+      zombify(std::move(*t));
+    }
+  };
+  arena_.reset(arena, disposer);
+  return arena;
+}
+
+void ThreadCachedArena::zombify(SysArena&& arena) {
+  zombies_->merge(std::move(arena));
+}
+
+size_t ThreadCachedArena::totalSize() const {
+  size_t result = sizeof(ThreadCachedArena);
+  for (const auto& arena : arena_.accessAllThreads()) {
+    result += arena.totalSize();
+  }
+  result += zombies_->totalSize() - sizeof(SysArena);
+  return result;
+}
+
+} // namespace folly
diff --git a/folly/memory/ThreadCachedArena.h b/folly/memory/ThreadCachedArena.h
new file mode 100644
index 00000000..f4562447
--- /dev/null
+++ b/folly/memory/ThreadCachedArena.h
@@ -0,0 +1,87 @@
+/*
+ * Copyright 2017 Facebook, Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include <type_traits>
+
+#include <folly/Likely.h>
+#include <folly/Synchronized.h>
+#include <folly/ThreadLocal.h>
+#include <folly/memory/Arena.h>
+
+namespace folly {
+
+/**
+ * Thread-caching arena: allocate memory which gets freed when the arena gets
+ * destroyed.
+ *
+ * The arena itself allocates memory using malloc() in blocks of
+ * at least minBlockSize bytes.
+ *
+ * For speed, each thread gets its own Arena (see Arena.h); when threads
+ * exit, the Arena gets merged into a "zombie" Arena, which will be deallocated
+ * when the ThreadCachedArena object is destroyed.
+ */
+class ThreadCachedArena {
+ public:
+  explicit ThreadCachedArena(
+      size_t minBlockSize = SysArena::kDefaultMinBlockSize,
+      size_t maxAlign = SysArena::kDefaultMaxAlign);
+
+  void* allocate(size_t size) {
+    SysArena* arena = arena_.get();
+    if (UNLIKELY(!arena)) {
+      arena = allocateThreadLocalArena();
+    }
+
+    return arena->allocate(size);
+  }
+
+  void deallocate(void* /* p */) {
+    // Deallocate? Never!
+  }
+
+  // Gets the total memory used by the arena
+  size_t totalSize() const;
+
+ private:
+  struct ThreadLocalPtrTag {};
+
+  ThreadCachedArena(const ThreadCachedArena&) = delete;
+  ThreadCachedArena(ThreadCachedArena&&) = delete;
+  ThreadCachedArena& operator=(const ThreadCachedArena&) = delete;
+  ThreadCachedArena& operator=(ThreadCachedArena&&) = delete;
+
+  SysArena* allocateThreadLocalArena();
+
+  // Zombify the blocks in arena, saving them for deallocation until
+  // the ThreadCachedArena is destroyed.
+  void zombify(SysArena&& arena);
+
+  const size_t minBlockSize_;
+  const size_t maxAlign_;
+
+  ThreadLocalPtr<SysArena, ThreadLocalPtrTag> arena_;  // Per-thread arena.
+
+  // Allocations from threads that are now dead.
+  Synchronized<SysArena> zombies_;
+};
+
+template <>
+struct IsArenaAllocator<ThreadCachedArena> : std::true_type { };
+
+} // namespace folly
diff --git a/folly/memory/detail/MallocImpl.cpp b/folly/memory/detail/MallocImpl.cpp
new file mode 100644
index 00000000..3c4f6dd9
--- /dev/null
+++ b/folly/memory/detail/MallocImpl.cpp
@@ -0,0 +1,48 @@
+/*
+ * Copyright 2017 Facebook, Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <folly/memory/detail/MallocImpl.h>
+
+extern "C" {
+
+#ifdef _MSC_VER
+// MSVC doesn't have weak symbols, so do some linker magic
+// to emulate them. (the magic is in the header)
+const char* mallocxWeak = nullptr;
+const char* rallocxWeak = nullptr;
+const char* xallocxWeak = nullptr;
+const char* sallocxWeak = nullptr;
+const char* dallocxWeak = nullptr;
+const char* sdallocxWeak = nullptr;
+const char* nallocxWeak = nullptr;
+const char* mallctlWeak = nullptr;
+const char* mallctlnametomibWeak = nullptr;
+const char* mallctlbymibWeak = nullptr;
+#elif !FOLLY_HAVE_WEAK_SYMBOLS
+void* (*mallocx)(size_t, int) = nullptr;
+void* (*rallocx)(void*, size_t, int) = nullptr;
+size_t (*xallocx)(void*, size_t, size_t, int) = nullptr;
+size_t (*sallocx)(const void*, int) = nullptr;
+void (*dallocx)(void*, int) = nullptr;
+void (*sdallocx)(void*, size_t, int) = nullptr;
+size_t (*nallocx)(size_t, int) = nullptr;
+int (*mallctl)(const char*, void*, size_t*, void*, size_t) = nullptr;
+int (*mallctlnametomib)(const char*, size_t*, size_t*) = nullptr;
+int (*mallctlbymib)(const size_t*, size_t, void*, size_t*, void*, size_t) =
+    nullptr;
+#endif
+
+}
diff --git a/folly/memory/detail/MallocImpl.h b/folly/memory/detail/MallocImpl.h
new file mode 100644
index 00000000..74358f1c
--- /dev/null
+++ b/folly/memory/detail/MallocImpl.h
@@ -0,0 +1,66 @@
+/*
+ * Copyright 2017 Facebook, Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include <stdlib.h>
+
+#include <folly/Portability.h>
+
+extern "C" {
+
+#if FOLLY_HAVE_WEAK_SYMBOLS
+void* mallocx(size_t, int) __attribute__((__weak__));
+void* rallocx(void*, size_t, int) __attribute__((__weak__));
+size_t xallocx(void*, size_t, size_t, int) __attribute__((__weak__));
+size_t sallocx(const void*, int) __attribute__((__weak__));
+void dallocx(void*, int) __attribute__((__weak__));
+void sdallocx(void*, size_t, int) __attribute__((__weak__));
+size_t nallocx(size_t, int) __attribute__((__weak__));
+int mallctl(const char*, void*, size_t*, void*, size_t)
+      __attribute__((__weak__));
+int mallctlnametomib(const char*, size_t*, size_t*) __attribute__((__weak__));
+int mallctlbymib(const size_t*, size_t, void*, size_t*, void*, size_t)
+      __attribute__((__weak__));
+#else
+extern void* (*mallocx)(size_t, int);
+extern void* (*rallocx)(void*, size_t, int);
+extern size_t (*xallocx)(void*, size_t, size_t, int);
+extern size_t (*sallocx)(const void*, int);
+extern void (*dallocx)(void*, int);
+extern void (*sdallocx)(void*, size_t, int);
+extern size_t (*nallocx)(size_t, int);
+extern int (*mallctl)(const char*, void*, size_t*, void*, size_t);
+extern int (*mallctlnametomib)(const char*, size_t*, size_t*);
+extern int (*mallctlbymib)(const size_t*, size_t, void*, size_t*, void*,
+                           size_t);
+#ifdef _MSC_VER
+// We emulate weak linkage for MSVC. The symbols we're
+// aliasing to are hiding in MallocImpl.cpp
+#pragma comment(linker, "/alternatename:mallocx=mallocxWeak")
+#pragma comment(linker, "/alternatename:rallocx=rallocxWeak")
+#pragma comment(linker, "/alternatename:xallocx=xallocxWeak")
+#pragma comment(linker, "/alternatename:sallocx=sallocxWeak")
+#pragma comment(linker, "/alternatename:dallocx=dallocxWeak")
+#pragma comment(linker, "/alternatename:sdallocx=sdallocxWeak")
+#pragma comment(linker, "/alternatename:nallocx=nallocxWeak")
+#pragma comment(linker, "/alternatename:mallctl=mallctlWeak")
+#pragma comment(linker, "/alternatename:mallctlnametomib=mallctlnametomibWeak")
+#pragma comment(linker, "/alternatename:mallctlbymib=mallctlbymibWeak")
+#endif
+#endif
+
+}
diff --git a/folly/memory/test/ArenaTest.cpp b/folly/memory/test/ArenaTest.cpp
new file mode 100644
index 00000000..f6a1ed47
--- /dev/null
+++ b/folly/memory/test/ArenaTest.cpp
@@ -0,0 +1,176 @@
+/*
+ * Copyright 2017 Facebook, Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <folly/memory/Arena.h>
+#include <folly/Memory.h>
+#include <folly/portability/GTest.h>
+
+#include <set>
+#include <vector>
+
+#include <glog/logging.h>
+
+using namespace folly;
+
+static_assert(IsArenaAllocator<SysArena>::value, "");
+
+TEST(Arena, SizeSanity) {
+  std::set<size_t*> allocatedItems;
+
+  static const size_t requestedBlockSize = 64;
+  SysArena arena(requestedBlockSize);
+  size_t minimum_size = sizeof(SysArena), maximum_size = minimum_size;
+  EXPECT_EQ(arena.totalSize(), minimum_size);
+
+  // Insert a single small element to get a new block
+  size_t* ptr = static_cast<size_t*>(arena.allocate(sizeof(long)));
+  allocatedItems.insert(ptr);
+  minimum_size += requestedBlockSize;
+  maximum_size += goodMallocSize(requestedBlockSize + SysArena::kBlockOverhead);
+  EXPECT_TRUE(arena.totalSize() >= minimum_size);
+  EXPECT_TRUE(arena.totalSize() <= maximum_size);
+  VLOG(4) << minimum_size << " < " << arena.totalSize() << " < "
+          << maximum_size;
+
+  // Insert a larger element, size should be the same
+  ptr = static_cast<size_t*>(arena.allocate(requestedBlockSize / 2));
+  allocatedItems.insert(ptr);
+  EXPECT_TRUE(arena.totalSize() >= minimum_size);
+  EXPECT_TRUE(arena.totalSize() <= maximum_size);
+  VLOG(4) << minimum_size << " < " << arena.totalSize() << " < "
+          << maximum_size;
+
+  // Insert 10 full block sizes to get 10 new blocks
+  for (int i = 0; i < 10; i++) {
+    ptr = static_cast<size_t*>(arena.allocate(requestedBlockSize));
+    allocatedItems.insert(ptr);
+  }
+  minimum_size += 10 * requestedBlockSize;
+  maximum_size += 10 * goodMallocSize(requestedBlockSize
+                                      + SysArena::kBlockOverhead);
+  EXPECT_TRUE(arena.totalSize() >= minimum_size);
+  EXPECT_TRUE(arena.totalSize() <= maximum_size);
+  VLOG(4) << minimum_size << " < " << arena.totalSize() << " < "
+          << maximum_size;
+
+  // Insert something huge
+  ptr = static_cast<size_t*>(arena.allocate(10 * requestedBlockSize));
+  allocatedItems.insert(ptr);
+  minimum_size += 10 * requestedBlockSize;
+  maximum_size += goodMallocSize(10 * requestedBlockSize
+                                 + SysArena::kBlockOverhead);
+  EXPECT_TRUE(arena.totalSize() >= minimum_size);
+  EXPECT_TRUE(arena.totalSize() <= maximum_size);
+  VLOG(4) << minimum_size << " < " << arena.totalSize() << " < "
+          << maximum_size;
+
+  // Nuke 'em all
+  for (const auto& item : allocatedItems) {
+    arena.deallocate(item);
+  }
+  //The total size should be the same
+  EXPECT_TRUE(arena.totalSize() >= minimum_size);
+  EXPECT_TRUE(arena.totalSize() <= maximum_size);
+  VLOG(4) << minimum_size << " < " << arena.totalSize() << " < "
+          << maximum_size;
+}
+
+TEST(Arena, BytesUsedSanity) {
+  static const size_t smallChunkSize = 1024;
+  static const size_t blockSize = goodMallocSize(16 * smallChunkSize);
+  const size_t bigChunkSize = blockSize - 4 * smallChunkSize;
+
+  size_t bytesUsed = 0;
+
+  SysArena arena(blockSize);
+  EXPECT_EQ(arena.bytesUsed(), bytesUsed);
+
+  // Insert 2 small chunks
+  arena.allocate(smallChunkSize);
+  arena.allocate(smallChunkSize);
+  bytesUsed += 2 * smallChunkSize;
+  EXPECT_EQ(arena.bytesUsed(), bytesUsed);
+  EXPECT_TRUE(arena.totalSize() >= blockSize);
+  EXPECT_TRUE(arena.totalSize() <= 2 * blockSize);
+
+  // Insert big chunk, should still fit in one block
+  arena.allocate(bigChunkSize);
+  bytesUsed += bigChunkSize;
+  EXPECT_EQ(arena.bytesUsed(), bytesUsed);
+  EXPECT_TRUE(arena.totalSize() >= blockSize);
+  EXPECT_TRUE(arena.totalSize() <= 2 * blockSize);
+
+  // Insert big chunk once more, should trigger new block allocation
+  arena.allocate(bigChunkSize);
+  bytesUsed += bigChunkSize;
+  EXPECT_EQ(arena.bytesUsed(), bytesUsed);
+  EXPECT_TRUE(arena.totalSize() >= 2 * blockSize);
+  EXPECT_TRUE(arena.totalSize() <= 3 * blockSize);
+
+  // Test that bytesUsed() accounts for alignment
+  static const size_t tinyChunkSize = 7;
+  arena.allocate(tinyChunkSize);
+  EXPECT_TRUE(arena.bytesUsed() >= bytesUsed + tinyChunkSize);
+  size_t delta = arena.bytesUsed() - bytesUsed;
+  EXPECT_EQ(delta & (delta - 1), 0);
+}
+
+TEST(Arena, Vector) {
+  static const size_t requestedBlockSize = 64;
+  SysArena arena(requestedBlockSize);
+
+  EXPECT_EQ(arena.totalSize(), sizeof(SysArena));
+
+  std::vector<size_t, StlAllocator<SysArena, size_t>>
+    vec { {}, StlAllocator<SysArena, size_t>(&arena) };
+
+  for (size_t i = 0; i < 1000; i++) {
+    vec.push_back(i);
+  }
+
+  for (size_t i = 0; i < 1000; i++) {
+    EXPECT_EQ(i, vec[i]);
+  }
+}
+
+TEST(Arena, SizeLimit) {
+  static const size_t requestedBlockSize = sizeof(size_t);
+  static const size_t maxSize = 10 * requestedBlockSize;
+
+  SysArena arena(requestedBlockSize, maxSize);
+
+  void* a = arena.allocate(sizeof(size_t));
+  EXPECT_TRUE(a != nullptr);
+  EXPECT_THROW(arena.allocate(maxSize + 1), std::bad_alloc);
+}
+
+TEST(Arena, MoveArena) {
+  SysArena arena(sizeof(size_t) * 2);
+  arena.allocate(sizeof(size_t));
+  auto totalSize = arena.totalSize();
+  auto bytesUsed = arena.bytesUsed();
+
+  SysArena moved(std::move(arena));
+  EXPECT_EQ(totalSize, moved.totalSize());
+  EXPECT_EQ(bytesUsed, moved.bytesUsed());
+}
+
+int main(int argc, char *argv[]) {
+  testing::InitGoogleTest(&argc, argv);
+  gflags::ParseCommandLineFlags(&argc, &argv, true);
+  auto ret = RUN_ALL_TESTS();
+  return ret;
+}
diff --git a/folly/memory/test/MallctlHelperTest.cpp b/folly/memory/test/MallctlHelperTest.cpp
index e2d3c2a7..dc447a7a 100644
--- a/folly/memory/test/MallctlHelperTest.cpp
+++ b/folly/memory/test/MallctlHelperTest.cpp
@@ -15,8 +15,8 @@
  */
 
 #include <folly/memory/MallctlHelper.h>
-#include <folly/Malloc.h>
 #include <folly/init/Init.h>
+#include <folly/memory/Malloc.h>
 #include <folly/portability/GTest.h>
 
 #ifdef FOLLY_HAVE_LIBJEMALLOC
diff --git a/folly/memory/test/ThreadCachedArenaTest.cpp b/folly/memory/test/ThreadCachedArenaTest.cpp
new file mode 100644
index 00000000..4401ae0a
--- /dev/null
+++ b/folly/memory/test/ThreadCachedArenaTest.cpp
@@ -0,0 +1,265 @@
+/*
+ * Copyright 2017 Facebook, Inc.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <folly/memory/ThreadCachedArena.h>
+
+#include <algorithm>
+#include <iterator>
+#include <map>
+#include <mutex>
+#include <random>
+#include <thread>
+#include <unordered_map>
+
+#include <glog/logging.h>
+
+#include <folly/Benchmark.h>
+#include <folly/Memory.h>
+#include <folly/Range.h>
+#include <folly/portability/GTest.h>
+
+using namespace folly;
+
+namespace {
+
+class ArenaTester {
+ public:
+  explicit ArenaTester(ThreadCachedArena& arena) : arena_(&arena) { }
+
+  void allocate(size_t count, size_t maxSize);
+  void verify();
+  void merge(ArenaTester&& other);
+
+ private:
+  std::mutex mergeMutex_;
+  std::vector<std::pair<uint8_t, Range<uint8_t*>>> areas_;
+  ThreadCachedArena* arena_;
+};
+
+void ArenaTester::allocate(size_t count, size_t maxSize) {
+  // Allocate chunks of memory of random sizes
+  std::mt19937 rnd;
+  std::uniform_int_distribution<uint32_t> sizeDist(1, maxSize - 1);
+  areas_.clear();
+  areas_.reserve(count);
+  for (size_t i = 0; i < count; i++) {
+    size_t size = sizeDist(rnd);
+    uint8_t* p = static_cast<uint8_t*>(arena_->allocate(size));
+    areas_.emplace_back(uint8_t(rnd() & 0xff), Range<uint8_t*>(p, size));
+  }
+
+  // Fill each area with a different value, to prove that they don't overlap
+  // Fill in random order.
+  std::random_shuffle(areas_.begin(), areas_.end(), [&rnd](ptrdiff_t n) {
+    return std::uniform_int_distribution<uint32_t>(0, n - 1)(rnd);
+  });
+
+  for (auto& p : areas_) {
+    std::fill(p.second.begin(), p.second.end(), p.first);
+  }
+}
+
+void ArenaTester::verify() {
+  for (auto& p : areas_) {
+    for (auto v : p.second) {
+      EXPECT_EQ(p.first, v);
+    }
+  }
+}
+
+void ArenaTester::merge(ArenaTester&& other) {
+  {
+    std::lock_guard<std::mutex> lock(mergeMutex_);
+    std::move(other.areas_.begin(), other.areas_.end(),
+              std::back_inserter(areas_));
+  }
+  other.areas_.clear();
+}
+
+} // namespace
+
+TEST(ThreadCachedArena, BlockSize) {
+  static const size_t alignment = folly::max_align_v;
+  static const size_t requestedBlockSize = 64;
+
+  ThreadCachedArena arena(requestedBlockSize);
+  size_t blockSize = alignment;
+  uint8_t* prev = static_cast<uint8_t*>(arena.allocate(1));
+
+  // Keep allocating until we're no longer one single alignment away from the
+  // previous allocation -- that's when we've gotten to the next block.
+  uint8_t* p;
+  while ((p = static_cast<uint8_t*>(arena.allocate(1))) ==
+         prev + alignment) {
+    prev = p;
+    blockSize += alignment;
+  }
+
+  VLOG(1) << "Requested block size: " << requestedBlockSize << ", actual: "
+          << blockSize;
+  EXPECT_LE(requestedBlockSize, blockSize);
+}
+
+TEST(ThreadCachedArena, SingleThreaded) {
+  static const size_t requestedBlockSize = 64;
+  ThreadCachedArena arena(requestedBlockSize);
+  EXPECT_EQ(arena.totalSize(), sizeof(ThreadCachedArena));
+
+  ArenaTester tester(arena);
+  tester.allocate(100, 100 << 10);
+  tester.verify();
+
+  EXPECT_GT(arena.totalSize(), sizeof(ThreadCachedArena));
+}
+
+TEST(ThreadCachedArena, MultiThreaded) {
+  static const size_t requestedBlockSize = 64;
+  ThreadCachedArena arena(requestedBlockSize);
+  ArenaTester mainTester(arena);
+
+  // Do this twice, to catch the possibility that memory from the first
+  // round gets freed
+  static const size_t numThreads = 20;
+  for (size_t i = 0; i < 2; i++) {
+    std::vector<std::thread> threads;
+    threads.reserve(numThreads);
+    for (size_t j = 0; j < numThreads; j++) {
+      threads.emplace_back(
+          [&arena, &mainTester] () {
+            ArenaTester tester(arena);
+            tester.allocate(500, 1 << 10);
+            tester.verify();
+            mainTester.merge(std::move(tester));
+          });
+    }
+    for (auto& t : threads) {
+      t.join();
+    }
+  }
+
+  mainTester.verify();
+}
+
+TEST(ThreadCachedArena, StlAllocator) {
+  typedef std::unordered_map<
+    int, int, std::hash<int>, std::equal_to<int>,
+    StlAllocator<ThreadCachedArena, std::pair<const int, int>>> Map;
+
+  static const size_t requestedBlockSize = 64;
+  ThreadCachedArena arena(requestedBlockSize);
+
+  Map map {0, std::hash<int>(), std::equal_to<int>(),
+           StlAllocator<ThreadCachedArena, std::pair<const int, int>>(&arena)};
+
+  for (int i = 0; i < 1000; i++) {
+    map[i] = i;
+  }
+
+  for (int i = 0; i < 1000; i++) {
+    EXPECT_EQ(i, map[i]);
+  }
+}
+
+namespace {
+
+static const int kNumValues = 10000;
+
+BENCHMARK(bmUMStandard, iters) {
+  typedef std::unordered_map<int, int> Map;
+
+  while (iters--) {
+    Map map {0};
+    for (int i = 0; i < kNumValues; i++) {
+      map[i] = i;
+    }
+  }
+}
+
+BENCHMARK(bmUMArena, iters) {
+  typedef std::unordered_map<
+    int, int, std::hash<int>, std::equal_to<int>,
+    StlAllocator<ThreadCachedArena, std::pair<const int, int>>> Map;
+
+  while (iters--) {
+    ThreadCachedArena arena;
+
+    Map map {0, std::hash<int>(), std::equal_to<int>(),
+             StlAllocator<ThreadCachedArena, std::pair<const int, int>>(
+                 &arena)};
+
+    for (int i = 0; i < kNumValues; i++) {
+      map[i] = i;
+    }
+  }
+}
+
+BENCHMARK_DRAW_LINE()
+
+BENCHMARK(bmMStandard, iters) {
+  typedef std::map<int, int> Map;
+
+  while (iters--) {
+    Map map;
+    for (int i = 0; i < kNumValues; i++) {
+      map[i] = i;
+    }
+  }
+}
+
+BENCHMARK_DRAW_LINE()
+
+BENCHMARK(bmMArena, iters) {
+  typedef std::map<
+    int, int, std::less<int>,
+    StlAllocator<ThreadCachedArena, std::pair<const int, int>>> Map;
+
+  while (iters--) {
+    ThreadCachedArena arena;
+
+    Map map {std::less<int>(),
+             StlAllocator<ThreadCachedArena, std::pair<const int, int>>(
+                 &arena)};
+
+    for (int i = 0; i < kNumValues; i++) {
+      map[i] = i;
+    }
+  }
+}
+
+BENCHMARK_DRAW_LINE()
+
+} // namespace
+
+// Benchmark                               Iters   Total t    t/iter iter/sec
+// ----------------------------------------------------------------------------
+// Comparing benchmarks: bmUMStandard,bmUMArena
+//  + 143% bmUMStandard                     1570  2.005 s   1.277 ms  782.9
+// *       bmUMArena                        3817  2.003 s   524.7 us  1.861 k
+// ----------------------------------------------------------------------------
+// Comparing benchmarks: bmMStandard,bmMArena
+//  +79.0% bmMStandard                      1197  2.009 s   1.678 ms  595.8
+// *       bmMArena                         2135  2.002 s   937.6 us  1.042 k
+// ----------------------------------------------------------------------------
+
+int main(int argc, char *argv[]) {
+  testing::InitGoogleTest(&argc, argv);
+  gflags::ParseCommandLineFlags(&argc, &argv, true);
+  auto ret = RUN_ALL_TESTS();
+  if (!ret && FLAGS_benchmark) {
+    folly::runBenchmarks();
+  }
+  return ret;
+}
diff --git a/folly/small_vector.h b/folly/small_vector.h
index 71440437..59dd3e54 100644
--- a/folly/small_vector.h
+++ b/folly/small_vector.h
@@ -48,10 +48,10 @@
 #include <folly/Assume.h>
 #include <folly/ConstexprMath.h>
 #include <folly/FormatTraits.h>
-#include <folly/Malloc.h>
 #include <folly/Portability.h>
 #include <folly/SmallLocks.h>
 #include <folly/Traits.h>
+#include <folly/memory/Malloc.h>
 #include <folly/portability/BitsFunctexcept.h>
 #include <folly/portability/Malloc.h>
 #include <folly/portability/TypeTraits.h>
diff --git a/folly/test/ArenaSmartPtrTest.cpp b/folly/test/ArenaSmartPtrTest.cpp
index 911c3c37..ee94db0e 100644
--- a/folly/test/ArenaSmartPtrTest.cpp
+++ b/folly/test/ArenaSmartPtrTest.cpp
@@ -18,8 +18,8 @@
  * @author: Marcelo Juchem <marcelo@fb.com>
  */
 
-#include <folly/Arena.h>
 #include <folly/Memory.h>
+#include <folly/memory/Arena.h>
 #include <folly/portability/GTest.h>
 
 using namespace folly;
diff --git a/folly/test/ArenaTest.cpp b/folly/test/ArenaTest.cpp
deleted file mode 100644
index a1d324a4..00000000
--- a/folly/test/ArenaTest.cpp
+++ /dev/null
@@ -1,176 +0,0 @@
-/*
- * Copyright 2017 Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <folly/Arena.h>
-#include <folly/Memory.h>
-#include <folly/portability/GTest.h>
-
-#include <set>
-#include <vector>
-
-#include <glog/logging.h>
-
-using namespace folly;
-
-static_assert(IsArenaAllocator<SysArena>::value, "");
-
-TEST(Arena, SizeSanity) {
-  std::set<size_t*> allocatedItems;
-
-  static const size_t requestedBlockSize = 64;
-  SysArena arena(requestedBlockSize);
-  size_t minimum_size = sizeof(SysArena), maximum_size = minimum_size;
-  EXPECT_EQ(arena.totalSize(), minimum_size);
-
-  // Insert a single small element to get a new block
-  size_t* ptr = static_cast<size_t*>(arena.allocate(sizeof(long)));
-  allocatedItems.insert(ptr);
-  minimum_size += requestedBlockSize;
-  maximum_size += goodMallocSize(requestedBlockSize + SysArena::kBlockOverhead);
-  EXPECT_TRUE(arena.totalSize() >= minimum_size);
-  EXPECT_TRUE(arena.totalSize() <= maximum_size);
-  VLOG(4) << minimum_size << " < " << arena.totalSize() << " < "
-          << maximum_size;
-
-  // Insert a larger element, size should be the same
-  ptr = static_cast<size_t*>(arena.allocate(requestedBlockSize / 2));
-  allocatedItems.insert(ptr);
-  EXPECT_TRUE(arena.totalSize() >= minimum_size);
-  EXPECT_TRUE(arena.totalSize() <= maximum_size);
-  VLOG(4) << minimum_size << " < " << arena.totalSize() << " < "
-          << maximum_size;
-
-  // Insert 10 full block sizes to get 10 new blocks
-  for (int i = 0; i < 10; i++) {
-    ptr = static_cast<size_t*>(arena.allocate(requestedBlockSize));
-    allocatedItems.insert(ptr);
-  }
-  minimum_size += 10 * requestedBlockSize;
-  maximum_size += 10 * goodMallocSize(requestedBlockSize
-                                      + SysArena::kBlockOverhead);
-  EXPECT_TRUE(arena.totalSize() >= minimum_size);
-  EXPECT_TRUE(arena.totalSize() <= maximum_size);
-  VLOG(4) << minimum_size << " < " << arena.totalSize() << " < "
-          << maximum_size;
-
-  // Insert something huge
-  ptr = static_cast<size_t*>(arena.allocate(10 * requestedBlockSize));
-  allocatedItems.insert(ptr);
-  minimum_size += 10 * requestedBlockSize;
-  maximum_size += goodMallocSize(10 * requestedBlockSize
-                                 + SysArena::kBlockOverhead);
-  EXPECT_TRUE(arena.totalSize() >= minimum_size);
-  EXPECT_TRUE(arena.totalSize() <= maximum_size);
-  VLOG(4) << minimum_size << " < " << arena.totalSize() << " < "
-          << maximum_size;
-
-  // Nuke 'em all
-  for (const auto& item : allocatedItems) {
-    arena.deallocate(item);
-  }
-  //The total size should be the same
-  EXPECT_TRUE(arena.totalSize() >= minimum_size);
-  EXPECT_TRUE(arena.totalSize() <= maximum_size);
-  VLOG(4) << minimum_size << " < " << arena.totalSize() << " < "
-          << maximum_size;
-}
-
-TEST(Arena, BytesUsedSanity) {
-  static const size_t smallChunkSize = 1024;
-  static const size_t blockSize = goodMallocSize(16 * smallChunkSize);
-  const size_t bigChunkSize = blockSize - 4 * smallChunkSize;
-
-  size_t bytesUsed = 0;
-
-  SysArena arena(blockSize);
-  EXPECT_EQ(arena.bytesUsed(), bytesUsed);
-
-  // Insert 2 small chunks
-  arena.allocate(smallChunkSize);
-  arena.allocate(smallChunkSize);
-  bytesUsed += 2 * smallChunkSize;
-  EXPECT_EQ(arena.bytesUsed(), bytesUsed);
-  EXPECT_TRUE(arena.totalSize() >= blockSize);
-  EXPECT_TRUE(arena.totalSize() <= 2 * blockSize);
-
-  // Insert big chunk, should still fit in one block
-  arena.allocate(bigChunkSize);
-  bytesUsed += bigChunkSize;
-  EXPECT_EQ(arena.bytesUsed(), bytesUsed);
-  EXPECT_TRUE(arena.totalSize() >= blockSize);
-  EXPECT_TRUE(arena.totalSize() <= 2 * blockSize);
-
-  // Insert big chunk once more, should trigger new block allocation
-  arena.allocate(bigChunkSize);
-  bytesUsed += bigChunkSize;
-  EXPECT_EQ(arena.bytesUsed(), bytesUsed);
-  EXPECT_TRUE(arena.totalSize() >= 2 * blockSize);
-  EXPECT_TRUE(arena.totalSize() <= 3 * blockSize);
-
-  // Test that bytesUsed() accounts for alignment
-  static const size_t tinyChunkSize = 7;
-  arena.allocate(tinyChunkSize);
-  EXPECT_TRUE(arena.bytesUsed() >= bytesUsed + tinyChunkSize);
-  size_t delta = arena.bytesUsed() - bytesUsed;
-  EXPECT_EQ(delta & (delta - 1), 0);
-}
-
-TEST(Arena, Vector) {
-  static const size_t requestedBlockSize = 64;
-  SysArena arena(requestedBlockSize);
-
-  EXPECT_EQ(arena.totalSize(), sizeof(SysArena));
-
-  std::vector<size_t, StlAllocator<SysArena, size_t>>
-    vec { {}, StlAllocator<SysArena, size_t>(&arena) };
-
-  for (size_t i = 0; i < 1000; i++) {
-    vec.push_back(i);
-  }
-
-  for (size_t i = 0; i < 1000; i++) {
-    EXPECT_EQ(i, vec[i]);
-  }
-}
-
-TEST(Arena, SizeLimit) {
-  static const size_t requestedBlockSize = sizeof(size_t);
-  static const size_t maxSize = 10 * requestedBlockSize;
-
-  SysArena arena(requestedBlockSize, maxSize);
-
-  void* a = arena.allocate(sizeof(size_t));
-  EXPECT_TRUE(a != nullptr);
-  EXPECT_THROW(arena.allocate(maxSize + 1), std::bad_alloc);
-}
-
-TEST(Arena, MoveArena) {
-  SysArena arena(sizeof(size_t) * 2);
-  arena.allocate(sizeof(size_t));
-  auto totalSize = arena.totalSize();
-  auto bytesUsed = arena.bytesUsed();
-
-  SysArena moved(std::move(arena));
-  EXPECT_EQ(totalSize, moved.totalSize());
-  EXPECT_EQ(bytesUsed, moved.bytesUsed());
-}
-
-int main(int argc, char *argv[]) {
-  testing::InitGoogleTest(&argc, argv);
-  gflags::ParseCommandLineFlags(&argc, &argv, true);
-  auto ret = RUN_ALL_TESTS();
-  return ret;
-}
diff --git a/folly/test/ConcurrentSkipListTest.cpp b/folly/test/ConcurrentSkipListTest.cpp
index 671a8341..dc3deda7 100644
--- a/folly/test/ConcurrentSkipListTest.cpp
+++ b/folly/test/ConcurrentSkipListTest.cpp
@@ -27,10 +27,10 @@
 
 #include <glog/logging.h>
 
-#include <folly/Arena.h>
 #include <folly/Foreach.h>
 #include <folly/Memory.h>
 #include <folly/String.h>
+#include <folly/memory/Arena.h>
 #include <folly/portability/GFlags.h>
 #include <folly/portability/GTest.h>
 
diff --git a/folly/test/MemoryTest.cpp b/folly/test/MemoryTest.cpp
index 5ad28d5a..f3e6f8d3 100644
--- a/folly/test/MemoryTest.cpp
+++ b/folly/test/MemoryTest.cpp
@@ -21,8 +21,8 @@
 
 #include <glog/logging.h>
 
-#include <folly/Arena.h>
 #include <folly/String.h>
+#include <folly/memory/Arena.h>
 #include <folly/portability/GTest.h>
 
 using namespace folly;
diff --git a/folly/test/ThreadCachedArenaTest.cpp b/folly/test/ThreadCachedArenaTest.cpp
deleted file mode 100644
index 1d24c77a..00000000
--- a/folly/test/ThreadCachedArenaTest.cpp
+++ /dev/null
@@ -1,265 +0,0 @@
-/*
- * Copyright 2017 Facebook, Inc.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *   http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <folly/ThreadCachedArena.h>
-
-#include <algorithm>
-#include <iterator>
-#include <map>
-#include <mutex>
-#include <random>
-#include <thread>
-#include <unordered_map>
-
-#include <glog/logging.h>
-
-#include <folly/Benchmark.h>
-#include <folly/Memory.h>
-#include <folly/Range.h>
-#include <folly/portability/GTest.h>
-
-using namespace folly;
-
-namespace {
-
-class ArenaTester {
- public:
-  explicit ArenaTester(ThreadCachedArena& arena) : arena_(&arena) { }
-
-  void allocate(size_t count, size_t maxSize);
-  void verify();
-  void merge(ArenaTester&& other);
-
- private:
-  std::mutex mergeMutex_;
-  std::vector<std::pair<uint8_t, Range<uint8_t*>>> areas_;
-  ThreadCachedArena* arena_;
-};
-
-void ArenaTester::allocate(size_t count, size_t maxSize) {
-  // Allocate chunks of memory of random sizes
-  std::mt19937 rnd;
-  std::uniform_int_distribution<uint32_t> sizeDist(1, maxSize - 1);
-  areas_.clear();
-  areas_.reserve(count);
-  for (size_t i = 0; i < count; i++) {
-    size_t size = sizeDist(rnd);
-    uint8_t* p = static_cast<uint8_t*>(arena_->allocate(size));
-    areas_.emplace_back(uint8_t(rnd() & 0xff), Range<uint8_t*>(p, size));
-  }
-
-  // Fill each area with a different value, to prove that they don't overlap
-  // Fill in random order.
-  std::random_shuffle(areas_.begin(), areas_.end(), [&rnd](ptrdiff_t n) {
-    return std::uniform_int_distribution<uint32_t>(0, n - 1)(rnd);
-  });
-
-  for (auto& p : areas_) {
-    std::fill(p.second.begin(), p.second.end(), p.first);
-  }
-}
-
-void ArenaTester::verify() {
-  for (auto& p : areas_) {
-    for (auto v : p.second) {
-      EXPECT_EQ(p.first, v);
-    }
-  }
-}
-
-void ArenaTester::merge(ArenaTester&& other) {
-  {
-    std::lock_guard<std::mutex> lock(mergeMutex_);
-    std::move(other.areas_.begin(), other.areas_.end(),
-              std::back_inserter(areas_));
-  }
-  other.areas_.clear();
-}
-
-} // namespace
-
-TEST(ThreadCachedArena, BlockSize) {
-  static const size_t alignment = folly::max_align_v;
-  static const size_t requestedBlockSize = 64;
-
-  ThreadCachedArena arena(requestedBlockSize);
-  size_t blockSize = alignment;
-  uint8_t* prev = static_cast<uint8_t*>(arena.allocate(1));
-
-  // Keep allocating until we're no longer one single alignment away from the
-  // previous allocation -- that's when we've gotten to the next block.
-  uint8_t* p;
-  while ((p = static_cast<uint8_t*>(arena.allocate(1))) ==
-         prev + alignment) {
-    prev = p;
-    blockSize += alignment;
-  }
-
-  VLOG(1) << "Requested block size: " << requestedBlockSize << ", actual: "
-          << blockSize;
-  EXPECT_LE(requestedBlockSize, blockSize);
-}
-
-TEST(ThreadCachedArena, SingleThreaded) {
-  static const size_t requestedBlockSize = 64;
-  ThreadCachedArena arena(requestedBlockSize);
-  EXPECT_EQ(arena.totalSize(), sizeof(ThreadCachedArena));
-
-  ArenaTester tester(arena);
-  tester.allocate(100, 100 << 10);
-  tester.verify();
-
-  EXPECT_GT(arena.totalSize(), sizeof(ThreadCachedArena));
-}
-
-TEST(ThreadCachedArena, MultiThreaded) {
-  static const size_t requestedBlockSize = 64;
-  ThreadCachedArena arena(requestedBlockSize);
-  ArenaTester mainTester(arena);
-
-  // Do this twice, to catch the possibility that memory from the first
-  // round gets freed
-  static const size_t numThreads = 20;
-  for (size_t i = 0; i < 2; i++) {
-    std::vector<std::thread> threads;
-    threads.reserve(numThreads);
-    for (size_t j = 0; j < numThreads; j++) {
-      threads.emplace_back(
-          [&arena, &mainTester] () {
-            ArenaTester tester(arena);
-            tester.allocate(500, 1 << 10);
-            tester.verify();
-            mainTester.merge(std::move(tester));
-          });
-    }
-    for (auto& t : threads) {
-      t.join();
-    }
-  }
-
-  mainTester.verify();
-}
-
-TEST(ThreadCachedArena, StlAllocator) {
-  typedef std::unordered_map<
-    int, int, std::hash<int>, std::equal_to<int>,
-    StlAllocator<ThreadCachedArena, std::pair<const int, int>>> Map;
-
-  static const size_t requestedBlockSize = 64;
-  ThreadCachedArena arena(requestedBlockSize);
-
-  Map map {0, std::hash<int>(), std::equal_to<int>(),
-           StlAllocator<ThreadCachedArena, std::pair<const int, int>>(&arena)};
-
-  for (int i = 0; i < 1000; i++) {
-    map[i] = i;
-  }
-
-  for (int i = 0; i < 1000; i++) {
-    EXPECT_EQ(i, map[i]);
-  }
-}
-
-namespace {
-
-static const int kNumValues = 10000;
-
-BENCHMARK(bmUMStandard, iters) {
-  typedef std::unordered_map<int, int> Map;
-
-  while (iters--) {
-    Map map {0};
-    for (int i = 0; i < kNumValues; i++) {
-      map[i] = i;
-    }
-  }
-}
-
-BENCHMARK(bmUMArena, iters) {
-  typedef std::unordered_map<
-    int, int, std::hash<int>, std::equal_to<int>,
-    StlAllocator<ThreadCachedArena, std::pair<const int, int>>> Map;
-
-  while (iters--) {
-    ThreadCachedArena arena;
-
-    Map map {0, std::hash<int>(), std::equal_to<int>(),
-             StlAllocator<ThreadCachedArena, std::pair<const int, int>>(
-                 &arena)};
-
-    for (int i = 0; i < kNumValues; i++) {
-      map[i] = i;
-    }
-  }
-}
-
-BENCHMARK_DRAW_LINE()
-
-BENCHMARK(bmMStandard, iters) {
-  typedef std::map<int, int> Map;
-
-  while (iters--) {
-    Map map;
-    for (int i = 0; i < kNumValues; i++) {
-      map[i] = i;
-    }
-  }
-}
-
-BENCHMARK_DRAW_LINE()
-
-BENCHMARK(bmMArena, iters) {
-  typedef std::map<
-    int, int, std::less<int>,
-    StlAllocator<ThreadCachedArena, std::pair<const int, int>>> Map;
-
-  while (iters--) {
-    ThreadCachedArena arena;
-
-    Map map {std::less<int>(),
-             StlAllocator<ThreadCachedArena, std::pair<const int, int>>(
-                 &arena)};
-
-    for (int i = 0; i < kNumValues; i++) {
-      map[i] = i;
-    }
-  }
-}
-
-BENCHMARK_DRAW_LINE()
-
-} // namespace
-
-// Benchmark                               Iters   Total t    t/iter iter/sec
-// ----------------------------------------------------------------------------
-// Comparing benchmarks: bmUMStandard,bmUMArena
-//  + 143% bmUMStandard                     1570  2.005 s   1.277 ms  782.9
-// *       bmUMArena                        3817  2.003 s   524.7 us  1.861 k
-// ----------------------------------------------------------------------------
-// Comparing benchmarks: bmMStandard,bmMArena
-//  +79.0% bmMStandard                      1197  2.009 s   1.678 ms  595.8
-// *       bmMArena                         2135  2.002 s   937.6 us  1.042 k
-// ----------------------------------------------------------------------------
-
-int main(int argc, char *argv[]) {
-  testing::InitGoogleTest(&argc, argv);
-  gflags::ParseCommandLineFlags(&argc, &argv, true);
-  auto ret = RUN_ALL_TESTS();
-  if (!ret && FLAGS_benchmark) {
-    folly::runBenchmarks();
-  }
-  return ret;
-}