From: Igor Fedotov <ifedotov@suse.com>
Date: Thu, 13 Feb 2020 22:04:39 +0000 (+0300)
Subject: os/bluestore: add hybrid (avl+bitmap) allocator
X-Git-Tag: v14.2.11~31^2~5
X-Git-Url: http://git.apps.os.sepia.ceph.com/?a=commitdiff_plain;h=b19be107f4bcefce1380645d4958133d0e7ae31b;p=ceph.git

os/bluestore: add hybrid (avl+bitmap) allocator

Signed-off-by: Igor Fedotov <ifedotov@suse.com>
(cherry picked from commit c90eee34955734a5d4d2d4f0d1ff04b275374fb1)

 Conflicts:
	src/os/CMakeLists.txt
	src/test/objectstore/Allocator_bench.cc
	src/test/objectstore/Allocator_test.cc
	src/test/objectstore/CMakeLists.txt
  GTEST_SKIP not available, +trivial
---

diff --git a/src/common/options.cc b/src/common/options.cc
index 25ff532ee1bb3..f3b91c73bb110 100644
--- a/src/common/options.cc
+++ b/src/common/options.cc
@@ -4364,6 +4364,7 @@ std::vector<Option> get_global_options() {
 
     Option("bluefs_allocator", Option::TYPE_STR, Option::LEVEL_DEV)
     .set_default("bitmap")
+    .set_enum_allowed({"bitmap", "stupid", "avl", "hybrid"})
     .set_description(""),
 
     Option("bluefs_preextend_wal_files", Option::TYPE_BOOL, Option::LEVEL_ADVANCED)
@@ -4719,7 +4720,7 @@ std::vector<Option> get_global_options() {
 
     Option("bluestore_allocator", Option::TYPE_STR, Option::LEVEL_ADVANCED)
     .set_default("bitmap")
-    .set_enum_allowed({"bitmap", "stupid", "avl"})
+    .set_enum_allowed({"bitmap", "stupid", "avl", "hybrid"})
     .set_description("Allocator policy")
     .set_long_description("Allocator to use for bluestore.  Stupid should only be used for testing."),
 
diff --git a/src/os/CMakeLists.txt b/src/os/CMakeLists.txt
index f2767e39d4914..8fc8d3533edaf 100644
--- a/src/os/CMakeLists.txt
+++ b/src/os/CMakeLists.txt
@@ -33,6 +33,7 @@ if(WITH_BLUESTORE)
     bluestore/StupidAllocator.cc
     bluestore/BitmapAllocator.cc
     bluestore/AvlAllocator.cc
+    bluestore/HybridAllocator.cc
   )
 endif(WITH_BLUESTORE)
 
diff --git a/src/os/bluestore/Allocator.cc b/src/os/bluestore/Allocator.cc
index 25cb331880a53..8a638bdf81a37 100644
--- a/src/os/bluestore/Allocator.cc
+++ b/src/os/bluestore/Allocator.cc
@@ -5,6 +5,7 @@
 #include "StupidAllocator.h"
 #include "BitmapAllocator.h"
 #include "AvlAllocator.h"
+#include "HybridAllocator.h"
 #include "common/debug.h"
 #include "common/admin_socket.h"
 #define dout_subsys ceph_subsys_bluestore
@@ -12,6 +13,7 @@
 class Allocator::SocketHook : public AdminSocketHook {
   Allocator *alloc;
 
+  friend class Allocator;
   std::string name;
 public:
   explicit SocketHook(Allocator *alloc,
@@ -113,6 +115,9 @@ Allocator::~Allocator()
   delete asok_hook;
 }
 
+const string& Allocator::get_name() const {
+  return asok_hook->name;
+}
 
 Allocator *Allocator::create(CephContext* cct, string type,
                              int64_t size, int64_t block_size, const std::string& name)
@@ -124,6 +129,8 @@ Allocator *Allocator::create(CephContext* cct, string type,
     alloc = new BitmapAllocator(cct, size, block_size, name);
   } else if (type == "avl") {
     return new AvlAllocator(cct, size, block_size, name);
+  } else if (type == "hybrid") {
+    return new HybridAllocator(cct, size, block_size, 1024 /*FIXME*/, name);
   }
   if (alloc == nullptr) {
     lderr(cct) << "Allocator::" << __func__ << " unknown alloc type "
diff --git a/src/os/bluestore/Allocator.h b/src/os/bluestore/Allocator.h
index 9f679d584e1f8..7b571395f25f1 100644
--- a/src/os/bluestore/Allocator.h
+++ b/src/os/bluestore/Allocator.h
@@ -61,6 +61,9 @@ public:
 
   static Allocator *create(CephContext* cct, string type, int64_t size,
 			   int64_t block_size, const std::string& name = "");
+
+  const string& get_name() const;
+
 private:
   class SocketHook;
   SocketHook* asok_hook = nullptr;
diff --git a/src/os/bluestore/BitmapAllocator.h b/src/os/bluestore/BitmapAllocator.h
index 996663b49a2b0..cbf8c007591a5 100755
--- a/src/os/bluestore/BitmapAllocator.h
+++ b/src/os/bluestore/BitmapAllocator.h
@@ -30,6 +30,8 @@ public:
   void release(
     const interval_set<uint64_t>& release_set) override;
 
+  using Allocator::release;
+
   uint64_t get_free() override
   {
     return get_available();
diff --git a/src/os/bluestore/HybridAllocator.cc b/src/os/bluestore/HybridAllocator.cc
new file mode 100644
index 0000000000000..8563b6f25a5dc
--- /dev/null
+++ b/src/os/bluestore/HybridAllocator.cc
@@ -0,0 +1,219 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#include "HybridAllocator.h"
+
+#include <limits>
+
+#include "common/config_proxy.h"
+#include "common/debug.h"
+
+#define dout_context cct
+#define dout_subsys ceph_subsys_bluestore
+#undef  dout_prefix
+#define dout_prefix *_dout << "HybridAllocator "
+
+
+int64_t HybridAllocator::allocate(
+  uint64_t want,
+  uint64_t unit,
+  uint64_t max_alloc_size,
+  int64_t  hint,
+  PExtentVector* extents)
+{
+  ldout(cct, 10) << __func__ << std::hex
+                 << " want 0x" << want
+                 << " unit 0x" << unit
+                 << " max_alloc_size 0x" << max_alloc_size
+                 << " hint 0x" << hint
+                 << std::dec << dendl;
+  ceph_assert(isp2(unit));
+  ceph_assert(want % unit == 0);
+
+  if (max_alloc_size == 0) {
+    max_alloc_size = want;
+  }
+  if (constexpr auto cap = std::numeric_limits<decltype(bluestore_pextent_t::length)>::max();
+      max_alloc_size >= cap) {
+    max_alloc_size = p2align(uint64_t(cap), (uint64_t)get_block_size());
+  }
+
+  std::lock_guard l(lock);
+
+  int64_t res;
+  PExtentVector local_extents;
+
+  // preserve original 'extents' vector state
+  auto orig_size = extents->size();
+  auto orig_pos = extents->end();
+  if (orig_size) {
+    --orig_pos;
+  }
+
+  // try bitmap first to avoid unneeded contiguous extents split if
+  // desired amount is less than shortes range in AVL
+  if (bmap_alloc && bmap_alloc->get_free() &&
+    want < _lowest_size_available()) {
+    res = bmap_alloc->allocate(want, unit, max_alloc_size, hint, extents);
+    if (res < 0) {
+      // got a failure, release already allocated and
+      // start over allocation from avl
+      if (orig_size) {
+        local_extents.insert(
+          local_extents.end(), ++orig_pos, extents->end());
+        extents->resize(orig_size);
+      } else {
+        extents->swap(local_extents);
+      }
+      bmap_alloc->release(local_extents);
+      res = 0;
+    }
+    if ((uint64_t)res < want) {
+      auto res2 = _allocate(want - res, unit, max_alloc_size, hint, extents);
+      if (res2 < 0) {
+        res = res2; // caller to do the release
+      } else {
+        res += res2;
+      }
+    }
+  } else {
+    res = _allocate(want, unit, max_alloc_size, hint, extents);
+    if (res < 0) {
+      // got a failure, release already allocated and
+      // start over allocation from bitmap
+      if (orig_size) {
+        local_extents.insert(
+          local_extents.end(), ++orig_pos, extents->end());
+        extents->resize(orig_size);
+      } else {
+        extents->swap(local_extents);
+      }
+      _release(local_extents);
+      res = 0;
+    }
+    if ((uint64_t)res < want ) {
+      auto res2 = bmap_alloc ?
+        bmap_alloc->allocate(want - res, unit, max_alloc_size, hint, extents) :
+        0;
+      if (res2 < 0 ) {
+        res = res2; // caller to do the release
+      } else {
+        res += res2;
+      }
+    }
+  }
+  return res ? res : -ENOSPC;
+}
+
+void HybridAllocator::release(const interval_set<uint64_t>& release_set) {
+  std::lock_guard l(lock);
+  // this will attempt to put free ranges into AvlAllocator first and
+  // fallback to bitmap one via _try_insert_range call
+  _release(release_set);
+}
+
+uint64_t HybridAllocator::get_free()
+{
+  std::lock_guard l(lock);
+  return (bmap_alloc ? bmap_alloc->get_free() : 0) + _get_free();
+}
+
+double HybridAllocator::get_fragmentation()
+{
+  std::lock_guard l(lock);
+  auto f = AvlAllocator::_get_fragmentation();
+  auto bmap_free = bmap_alloc ? bmap_alloc->get_free() : 0;
+  if (bmap_free) {
+    auto _free = _get_free() + bmap_free;
+    auto bf = bmap_alloc->get_fragmentation();
+
+    f = f * _get_free() / _free + bf * bmap_free / _free;
+  }
+  return f;
+}
+
+void HybridAllocator::dump()
+{
+  std::lock_guard l(lock);
+  AvlAllocator::_dump();
+  if (bmap_alloc) {
+    bmap_alloc->dump();
+  }
+  ldout(cct, 0) << __func__
+    << " avl_free: " << _get_free()
+    << " bmap_free: " << (bmap_alloc ? bmap_alloc->get_free() : 0)
+    << dendl;
+}
+
+void HybridAllocator::dump(std::function<void(uint64_t offset, uint64_t length)> notify)
+{
+  AvlAllocator::dump(notify);
+  if (bmap_alloc) {
+    bmap_alloc->dump(notify);
+  }
+}
+
+void HybridAllocator::init_add_free(uint64_t offset, uint64_t length)
+{
+  std::lock_guard l(lock);
+  ldout(cct, 10) << __func__ << std::hex
+                 << " offset 0x" << offset
+                 << " length 0x" << length
+                 << std::dec << dendl;
+  _add_to_tree(offset, length);
+}
+
+void HybridAllocator::init_rm_free(uint64_t offset, uint64_t length)
+{
+  std::lock_guard l(lock);
+  ldout(cct, 10) << __func__ << std::hex
+                 << " offset 0x" << offset
+                 << " length 0x" << length
+                 << std::dec << dendl;
+  _try_remove_from_tree(offset, length,
+    [&](uint64_t o, uint64_t l, bool found) {
+      if (!found) {
+        if (bmap_alloc) {
+          bmap_alloc->init_rm_free(o, l);
+        } else {
+          lderr(cct) << "init_rm_free lambda" << std::hex
+            << "Uexpected extent: "
+            << " 0x" << o << "~" << l
+            << std::dec << dendl;
+          ceph_assert(false);
+        }
+      }
+    });
+}
+
+void HybridAllocator::shutdown()
+{
+  std::lock_guard l(lock);
+  _shutdown();
+  if (bmap_alloc) {
+    bmap_alloc->shutdown();
+    delete bmap_alloc;
+    bmap_alloc = nullptr;
+  }
+}
+
+void HybridAllocator::_spillover_range(uint64_t start, uint64_t end) {
+  auto size = end - start;
+  dout(20) << __func__
+    << std::hex << " "
+    << start << "~" << size
+    << std::dec
+    << dendl;
+  ceph_assert(size);
+  if (!bmap_alloc) {
+    dout(1) << __func__
+      << std::hex
+      << " constructing fallback allocator"
+      << dendl;
+    bmap_alloc = new BitmapAllocator(cct,
+      get_capacity(),
+      get_block_size(),
+      get_name());
+  }
+  bmap_alloc->init_add_free(start, size);
+}
diff --git a/src/os/bluestore/HybridAllocator.h b/src/os/bluestore/HybridAllocator.h
new file mode 100644
index 0000000000000..1c821f69c3b88
--- /dev/null
+++ b/src/os/bluestore/HybridAllocator.h
@@ -0,0 +1,46 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#pragma once
+
+#include <mutex>
+
+#include "AvlAllocator.h"
+#include "BitmapAllocator.h"
+
+class HybridAllocator : public AvlAllocator {
+  BitmapAllocator* bmap_alloc = nullptr;
+public:
+  HybridAllocator(CephContext* cct, int64_t device_size, int64_t _block_size,
+                  uint64_t max_entries,
+	          const std::string& name) :
+      AvlAllocator(cct, device_size, _block_size, max_entries, name) {
+  }
+  int64_t allocate(
+    uint64_t want,
+    uint64_t unit,
+    uint64_t max_alloc_size,
+    int64_t  hint,
+    PExtentVector *extents) override;
+  void release(const interval_set<uint64_t>& release_set) override;
+  uint64_t get_free() override;
+  double get_fragmentation() override;
+
+  void dump() override;
+  void dump(std::function<void(uint64_t offset, uint64_t length)> notify) override;
+  void init_add_free(uint64_t offset, uint64_t length) override;
+  void init_rm_free(uint64_t offset, uint64_t length) override;
+  void shutdown() override;
+
+protected:
+  // intended primarily for UT
+  BitmapAllocator* get_bmap() {
+    return bmap_alloc;
+  }
+  const BitmapAllocator* get_bmap() const {
+    return bmap_alloc;
+  }
+private:
+
+  void _spillover_range(uint64_t start, uint64_t end) override;
+};
diff --git a/src/test/objectstore/Allocator_bench.cc b/src/test/objectstore/Allocator_bench.cc
index cfadee5306d9c..2e45df04be95f 100755
--- a/src/test/objectstore/Allocator_bench.cc
+++ b/src/test/objectstore/Allocator_bench.cc
@@ -361,10 +361,10 @@ TEST_P(AllocTest, mempoolAccounting)
   ASSERT_EQ(mempool::bluestore_alloc::allocated_items(), items);
 }
 
-INSTANTIATE_TEST_SUITE_P(
+INSTANTIATE_TEST_CASE_P(
   Allocator,
   AllocTest,
-  ::testing::Values("stupid", "bitmap", "avl"));
+  ::testing::Values("stupid", "bitmap", "avl", "hybrid"));
 
 #else
 
diff --git a/src/test/objectstore/Allocator_test.cc b/src/test/objectstore/Allocator_test.cc
index 4659e00e94ef8..1881cfa822ce3 100644
--- a/src/test/objectstore/Allocator_test.cc
+++ b/src/test/objectstore/Allocator_test.cc
@@ -15,7 +15,6 @@
 #include "include/Context.h"
 #include "os/bluestore/Allocator.h"
 
-#include <boost/random/uniform_int.hpp>
 typedef boost::mt11213b gen_type;
 
 #if GTEST_HAS_PARAM_TEST
@@ -257,12 +256,17 @@ TEST_P(AllocTest, test_alloc_fragmentation)
       EXPECT_EQ(0.0, alloc->get_fragmentation());
     }
   }
+  tmp.clear();
   EXPECT_EQ(-ENOSPC, alloc->allocate(want_size, alloc_unit, 0, 0, &tmp));
 
   if (GetParam() == string("avl")) {
     // AVL allocator uses a different allocating strategy
     cout << "skipping for AVL allocator";
     return;
+  } else if (GetParam() == string("hybrid")) {
+    // AVL allocator uses a different allocating strategy
+    cout << "skipping for Hybrid allocator";
+    return;
   }
 
   for (size_t i = 0; i < allocated.size(); i += 2)
@@ -417,6 +421,7 @@ TEST_P(AllocTest, test_alloc_big2)
   EXPECT_EQ(need,
       alloc->allocate(need, mas, 0, &extents));
   need = block_size * blocks / 4; // 2GB
+  extents.clear();
   EXPECT_EQ(need,
       alloc->allocate(need, mas, 0, &extents));
   EXPECT_TRUE(extents[0].length > 0);
@@ -440,11 +445,40 @@ TEST_P(AllocTest, test_alloc_big3)
   EXPECT_TRUE(extents[0].length > 0);
 }
 
+TEST_P(AllocTest, test_alloc_contiguous)
+{
+  int64_t block_size = 0x1000;
+  int64_t capacity = block_size * 1024 * 1024;
+
+  {
+    init_alloc(capacity, block_size);
+
+    alloc->init_add_free(0, capacity);
+    PExtentVector extents;
+    uint64_t need = 4 * block_size;
+    EXPECT_EQ(need,
+      alloc->allocate(need, need,
+        0, (int64_t)0, &extents));
+    EXPECT_EQ(1u, extents.size());
+    EXPECT_EQ(extents[0].offset, 0);
+    EXPECT_EQ(extents[0].length, 4 * block_size);
+
+    extents.clear();
+    EXPECT_EQ(need,
+      alloc->allocate(need, need,
+        0, (int64_t)0, &extents));
+    EXPECT_EQ(1u, extents.size());
+    EXPECT_EQ(extents[0].offset, 4 * block_size);
+    EXPECT_EQ(extents[0].length, 4 * block_size);
+  }
+
+  alloc->shutdown();
+}
+
 INSTANTIATE_TEST_CASE_P(
   Allocator,
   AllocTest,
-  ::testing::Values("stupid", "bitmap", "avl"));
-
+  ::testing::Values("stupid", "bitmap", "avl", "hybrid"));
 #else
 
 TEST(DummyTest, ValueParameterizedTestsAreNotSupportedOnThisPlatform) {}
diff --git a/src/test/objectstore/CMakeLists.txt b/src/test/objectstore/CMakeLists.txt
index be9951d6ef908..53f1b0e1ddbd9 100644
--- a/src/test/objectstore/CMakeLists.txt
+++ b/src/test/objectstore/CMakeLists.txt
@@ -106,6 +106,20 @@ if(WITH_BLUESTORE)
   add_ceph_unittest(unittest_fastbmap_allocator)
   target_link_libraries(unittest_fastbmap_allocator os global)
 
+  set_target_properties(unittest_fastbmap_allocator PROPERTIES COMPILE_FLAGS
+  "${UNITTEST_CXX_FLAGS}")
+
+  add_executable(unittest_hybrid_allocator
+    hybrid_allocator_test.cc
+    $<TARGET_OBJECTS:unit-main>
+    )
+  add_ceph_unittest(unittest_hybrid_allocator)
+  target_link_libraries(unittest_hybrid_allocator os global)
+
+  set_target_properties(unittest_hybrid_allocator PROPERTIES COMPILE_FLAGS
+  "${UNITTEST_CXX_FLAGS}")
+
+  # unittest_bluefs
   add_executable(unittest_bluefs
     test_bluefs.cc
     )
diff --git a/src/test/objectstore/hybrid_allocator_test.cc b/src/test/objectstore/hybrid_allocator_test.cc
new file mode 100755
index 0000000000000..c21957ef88d8d
--- /dev/null
+++ b/src/test/objectstore/hybrid_allocator_test.cc
@@ -0,0 +1,230 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#include <iostream>
+#include <gtest/gtest.h>
+
+#include "os/bluestore/HybridAllocator.h"
+
+class TestHybridAllocator : public HybridAllocator {
+public:
+  TestHybridAllocator(CephContext* cct,
+                      int64_t device_size,
+                      int64_t _block_size,
+                      uint64_t max_entries,
+      const std::string& name) :
+    HybridAllocator(cct, device_size, _block_size,
+      max_entries,
+      name) {
+  }
+
+  uint64_t get_bmap_free() {
+    return get_bmap() ? get_bmap()->get_free() : 0;
+  }
+  uint64_t get_avl_free() {
+    return AvlAllocator::get_free();
+  }
+};
+
+const uint64_t _1m = 1024 * 1024;
+const uint64_t _4m = 4 * 1024 * 1024;
+
+TEST(HybridAllocator, basic)
+{
+  {
+    uint64_t block_size = 0x1000;
+    uint64_t capacity = 0x10000 * _1m; // = 64GB
+    TestHybridAllocator ha(g_ceph_context, capacity, block_size, 4, "test_hybrid_allocator");
+
+    ASSERT_EQ(0, ha.get_free());
+    ASSERT_EQ(0, ha.get_avl_free());
+    ASSERT_EQ(0, ha.get_bmap_free());
+
+    ha.init_add_free(0, _4m);
+    ASSERT_EQ(_4m, ha.get_free());
+    ASSERT_EQ(_4m, ha.get_avl_free());
+    ASSERT_EQ(0, ha.get_bmap_free());
+
+    ha.init_add_free(2 * _4m, _4m);
+    ASSERT_EQ(_4m * 2, ha.get_free());
+    ASSERT_EQ(_4m * 2, ha.get_avl_free());
+    ASSERT_EQ(0, ha.get_bmap_free());
+
+    ha.init_add_free(100 * _4m, _4m);
+    ha.init_add_free(102 * _4m, _4m);
+
+    ASSERT_EQ(_4m * 4, ha.get_free());
+    ASSERT_EQ(_4m * 4, ha.get_avl_free());
+    ASSERT_EQ(0, ha.get_bmap_free());
+
+    // next allocs will go to bitmap
+    ha.init_add_free(4 * _4m, _4m);
+    ASSERT_EQ(_4m * 5, ha.get_free());
+    ASSERT_EQ(_4m * 4, ha.get_avl_free());
+    ASSERT_EQ(_4m * 1, ha.get_bmap_free());
+
+    ha.init_add_free(6 * _4m, _4m);
+    ASSERT_EQ(_4m * 6, ha.get_free());
+    ASSERT_EQ(_4m * 4, ha.get_avl_free());
+    ASSERT_EQ(_4m * 2, ha.get_bmap_free());
+
+    // so we have 6x4M chunks, 4 chunks at AVL and 2 at bitmap
+
+    ha.init_rm_free(_1m, _1m); // take 1M from AVL
+    ASSERT_EQ(_1m * 23, ha.get_free());
+    ASSERT_EQ(_1m * 14, ha.get_avl_free());
+    ASSERT_EQ(_1m * 9, ha.get_bmap_free());
+
+    ha.init_rm_free(6 * _4m + _1m, _1m); // take 1M from bmap
+    ASSERT_EQ(_1m * 22, ha.get_free());
+    ASSERT_EQ(_1m * 14, ha.get_avl_free());
+    ASSERT_EQ(_1m * 8, ha.get_bmap_free());
+
+    // so we have at avl: 2M~2M, 8M~4M, 400M~4M , 408M~4M
+    // and at bmap: 0~1M, 16M~1M, 18M~2M, 24~4M
+
+    PExtentVector extents;
+    // allocate 4K, to be served from bitmap
+    EXPECT_EQ(block_size, ha.allocate(block_size, block_size,
+      0, (int64_t)0, &extents));
+    ASSERT_EQ(1, extents.size());
+    ASSERT_EQ(0, extents[0].offset);
+
+    ASSERT_EQ(_1m * 14, ha.get_avl_free());
+    ASSERT_EQ(_1m * 8 - block_size, ha.get_bmap_free());
+
+    interval_set<uint64_t> release_set;
+    // release 4K, to be returned to bitmap
+    release_set.insert(extents[0].offset, extents[0].length);
+    ha.release(release_set);
+
+    ASSERT_EQ(_1m * 14, ha.get_avl_free());
+    ASSERT_EQ(_1m * 8, ha.get_bmap_free());
+    extents.clear();
+    release_set.clear();
+
+    // again we have at avl: 2M~2M, 8M~4M, 400M~4M , 408M~4M
+    // and at bmap: 0~1M, 16M~1M, 18M~2M, 24~4M
+
+    // add 12M~3M which will go to avl
+    ha.init_add_free(3 * _4m, 3 * _1m);
+    ASSERT_EQ(_1m * 17, ha.get_avl_free());
+    ASSERT_EQ(_1m * 8, ha.get_bmap_free());
+
+
+    // add 15M~4K which will be appended to existing slot
+    ha.init_add_free(15 * _1m, 0x1000);
+    ASSERT_EQ(_1m * 17 + 0x1000, ha.get_avl_free());
+    ASSERT_EQ(_1m * 8, ha.get_bmap_free());
+
+
+    // again we have at avl: 2M~2M, 8M~(7M+4K), 400M~4M , 408M~4M
+    // and at bmap: 0~1M, 16M~1M, 18M~2M, 24~4M
+
+    //some removals from bmap
+    ha.init_rm_free(28 * _1m - 0x1000, 0x1000);
+    ASSERT_EQ(_1m * 17 + 0x1000, ha.get_avl_free());
+    ASSERT_EQ(_1m * 8 - 0x1000, ha.get_bmap_free());
+
+    ha.init_rm_free(24 * _1m + 0x1000, 0x1000);
+    ASSERT_EQ(_1m * 17 + 0x1000, ha.get_avl_free());
+    ASSERT_EQ(_1m * 8 - 0x2000, ha.get_bmap_free());
+
+    ha.init_rm_free(24 * _1m + 0x1000, _4m - 0x2000);
+    ASSERT_EQ(_1m * 17 + 0x1000, ha.get_avl_free());
+    ASSERT_EQ(_1m * 4, ha.get_bmap_free());
+
+    //4K removal from avl
+    ha.init_rm_free(15 * _1m, 0x1000);
+    ASSERT_EQ(_1m * 17, ha.get_avl_free());
+    ASSERT_EQ(_1m * 4, ha.get_bmap_free());
+
+    //remove highest 4Ms from avl
+    ha.init_rm_free(_1m * 400, _4m);
+    ha.init_rm_free(_1m * 408, _4m);
+    ASSERT_EQ(_1m * 9, ha.get_avl_free());
+    ASSERT_EQ(_1m * 4, ha.get_bmap_free());
+
+    // we have at avl: 2M~2M, 8M~7M
+    // and at bmap: 0~1M, 16M~1M, 18M~2M
+
+    // this will go to avl making 16M~4M span broken between both allocators
+    ha.init_add_free(17 * _1m, _1m);
+    ASSERT_EQ(_1m * 10, ha.get_avl_free());
+    ASSERT_EQ(_1m * 4, ha.get_bmap_free());
+
+    // we have at avl: 2M~2M, 8M~7M, 17M~1M
+    // and at bmap: 0~1M, 16M~1M, 18M~2M
+
+    // and now do some cutoffs from this "broken" 16M~4M span
+
+    //cut off 4K from bmap
+    ha.init_rm_free(16 * _1m, 0x1000);
+    ASSERT_EQ(_1m * 10, ha.get_avl_free());
+    ASSERT_EQ(_1m * 4 - 0x1000, ha.get_bmap_free());
+
+    //cut off 1M-4K from bmap and 4K from avl
+    ha.init_rm_free(16 * _1m + 0x1000, _1m);
+    ASSERT_EQ(_1m * 10 - 0x1000, ha.get_avl_free());
+    ASSERT_EQ(_1m * 3, ha.get_bmap_free());
+
+    //cut off 512K avl
+    ha.init_rm_free(17 * _1m + 0x1000, _1m / 2);
+    ASSERT_EQ(_1m * 10 - 0x1000 - _1m / 2, ha.get_avl_free());
+    ASSERT_EQ(_1m * 3, ha.get_bmap_free());
+
+    //cut off the rest from avl and 8K from bmap
+    ha.init_rm_free(17 * _1m + 0x1000 + _1m / 2, _1m / 2 + 0x1000);
+    ASSERT_EQ(_1m * 9, ha.get_avl_free());
+    ASSERT_EQ(_1m * 3 - 0x2000, ha.get_bmap_free());
+  }
+
+  {
+    uint64_t block_size = 0x1000;
+    uint64_t capacity = 0x10000 * _1m; // = 64GB
+    TestHybridAllocator ha(g_ceph_context, capacity, block_size,
+      4 * sizeof(range_seg_t), "test_hybrid_allocator");
+
+    ha.init_add_free(_1m, _1m);
+    ha.init_add_free(_1m * 3, _1m);
+    ha.init_add_free(_1m * 5, _1m);
+    ha.init_add_free(0x4000, 0x1000);
+
+    ASSERT_EQ(_1m * 3 + 0x1000, ha.get_free());
+    ASSERT_EQ(_1m * 3 + 0x1000, ha.get_avl_free());
+    ASSERT_EQ(0, ha.get_bmap_free());
+
+    // This will substitute chunk 0x4000~1000.
+    // Since new chunk insertion into into AvlAllocator:range_tree
+    // happens immediately before 0x4000~1000 chunk care should be taken
+    // to order operations properly and do not use already disposed iterator.
+    ha.init_add_free(0, 0x2000);
+
+    ASSERT_EQ(_1m * 3 + 0x3000, ha.get_free());
+    ASSERT_EQ(_1m * 3 + 0x2000, ha.get_avl_free());
+    ASSERT_EQ(0x1000, ha.get_bmap_free());
+  }
+}
+
+TEST(HybridAllocator, fragmentation)
+{
+  {
+    uint64_t block_size = 0x1000;
+    uint64_t capacity = 0x1000 * 0x1000; // = 16M
+    TestHybridAllocator ha(g_ceph_context, capacity, block_size,
+      4 * sizeof(range_seg_t), "test_hybrid_allocator");
+
+    ha.init_add_free(0, 0x2000);
+    ha.init_add_free(0x4000, 0x2000);
+    ha.init_add_free(0x8000, 0x2000);
+    ha.init_add_free(0xc000, 0x1000);
+
+    ASSERT_EQ(0.5, ha.get_fragmentation());
+
+    // this will got to bmap with fragmentation = 1
+    ha.init_add_free(0x10000, 0x1000);
+
+    // which results in the following total fragmentation
+    ASSERT_EQ(0.5 * 7 / 8 + 1.0 / 8, ha.get_fragmentation());
+  }
+}