Fix error happening when online compaction kicks in on async commits

CHANGELOG.md

@@ -8,6 +8,7 @@
 * Running a query on @keys in a Dictionary would throw an exception ([#6831](https://github.com/realm/realm-core/issues/6831), since v13.15.1)
 * Change JSON selialization format back to follow ISO 8601 - and add output of nanoseconds ([#6855](https://github.com/realm/realm-core/issues/6855), since 13.17.0)
 * Testing the size of a collection of links against zero would sometimes fail (sometimes = "difficult to explain"). In particular: ([#6850](https://github.com/realm/realm-core/issues/6850), since v13.15.1)
+* Fixed crash in slab allocator (Assertion failed: ref + size <= next->first) Many issues like ([#6340](https://github.com/realm/realm-core/issues/6340), since 13.0.0)
 
 ### Breaking changes
 * None.

src/realm/alloc_slab.hpp

@@ -721,6 +721,7 @@ class SlabAlloc : public Allocator {
     friend class DB;
     friend class Group;
     friend class GroupWriter;
+    friend class GroupComitter;
 };
 
 

src/realm/group.hpp

@@ -834,6 +834,7 @@ class Group : public ArrayParent {
 
     friend class Table;
     friend class GroupWriter;
+    friend class GroupComitter;
     friend class DB;
     friend class _impl::GroupFriend;
     friend class metrics::QueryInfo;

src/realm/group_writer.cpp

@@ -70,7 +70,7 @@ class InMemoryWriter : public _impl::ArrayWriterBase {
 
 
 // Class controlling a memory mapped window into a file
-class GroupWriter::MapWindow {
+class GroupComitter::MapWindow {
 public:
     MapWindow(size_t alignment, util::File& f, ref_type start_ref, size_t initial_size,
               util::WriteMarker* write_marker = nullptr);
@@ -101,7 +101,7 @@ class GroupWriter::MapWindow {
 };
 
 // True if a requested block fall within a memory mapping.
-bool GroupWriter::MapWindow::matches(ref_type start_ref, size_t size)
+bool GroupComitter::MapWindow::matches(ref_type start_ref, size_t size)
 {
     if (start_ref < m_base_ref)
         return false;
@@ -117,14 +117,14 @@ bool GroupWriter::MapWindow::matches(ref_type start_ref, size_t size)
 //
 // In cases where a 1MB window would stretch beyond the end of the file, we choose
 // a smaller window. Anything mapped after the end of file would be undefined anyways.
-ref_type GroupWriter::MapWindow::aligned_to_mmap_block(ref_type start_ref)
+ref_type GroupComitter::MapWindow::aligned_to_mmap_block(ref_type start_ref)
 {
     // align to 1MB boundary
     size_t page_mask = m_alignment - 1;
     return start_ref & ~page_mask;
 }
 
-size_t GroupWriter::MapWindow::get_window_size(util::File& f, ref_type start_ref, size_t size)
+size_t GroupComitter::MapWindow::get_window_size(util::File& f, ref_type start_ref, size_t size)
 {
     size_t window_size = start_ref + size - m_base_ref;
     // always map at least to match alignment
@@ -146,7 +146,7 @@ size_t GroupWriter::MapWindow::get_window_size(util::File& f, ref_type start_ref
 //
 // extends_to_match() will extend an existing mapping to accomodate a new request if possible
 // and return true. If the request falls in a different 1MB window, it'll return false.
-bool GroupWriter::MapWindow::extends_to_match(util::File& f, ref_type start_ref, size_t size)
+bool GroupComitter::MapWindow::extends_to_match(util::File& f, ref_type start_ref, size_t size)
 {
     size_t aligned_ref = aligned_to_mmap_block(start_ref);
     if (aligned_ref != m_base_ref)
@@ -158,8 +158,8 @@ bool GroupWriter::MapWindow::extends_to_match(util::File& f, ref_type start_ref,
     return true;
 }
 
-GroupWriter::MapWindow::MapWindow(size_t alignment, util::File& f, ref_type start_ref, size_t size,
-                                  util::WriteMarker* write_marker)
+GroupComitter::MapWindow::MapWindow(size_t alignment, util::File& f, ref_type start_ref, size_t size,
+                                    util::WriteMarker* write_marker)
     : m_alignment(alignment)
 {
     m_base_ref = aligned_to_mmap_block(start_ref);
@@ -173,47 +173,43 @@ GroupWriter::MapWindow::MapWindow(size_t alignment, util::File& f, ref_type star
 #endif
 }
 
-GroupWriter::MapWindow::~MapWindow()
+GroupComitter::MapWindow::~MapWindow()
 {
     m_map.sync();
     m_map.unmap();
 }
 
-void GroupWriter::MapWindow::flush()
+void GroupComitter::MapWindow::flush()
 {
     m_map.flush();
 }
 
-void GroupWriter::MapWindow::sync()
+void GroupComitter::MapWindow::sync()
 {
     flush();
     m_map.sync();
 }
 
-char* GroupWriter::MapWindow::translate(ref_type ref)
+char* GroupComitter::MapWindow::translate(ref_type ref)
 {
     return m_map.get_addr() + (ref - m_base_ref);
 }
 
-void GroupWriter::MapWindow::encryption_read_barrier(void* start_addr, size_t size)
+void GroupComitter::MapWindow::encryption_read_barrier(void* start_addr, size_t size)
 {
     realm::util::encryption_read_barrier_for_write(start_addr, size, m_map.get_encrypted_mapping());
 }
 
-void GroupWriter::MapWindow::encryption_write_barrier(void* start_addr, size_t size)
+void GroupComitter::MapWindow::encryption_write_barrier(void* start_addr, size_t size)
 {
     realm::util::encryption_write_barrier(start_addr, size, m_map.get_encrypted_mapping());
 }
 
-
-GroupWriter::GroupWriter(Group& group, Durability dura, WriteMarker* write_marker)
+GroupComitter::GroupComitter(Group& group, Durability dura, WriteMarker* write_marker)
     : m_group(group)
     , m_alloc(group.m_alloc)
-    , m_free_positions(m_alloc)
-    , m_free_lengths(m_alloc)
-    , m_free_versions(m_alloc)
-    , m_durability(dura)
     , m_write_marker(write_marker)
+    , m_durability(dura)
 {
     m_map_windows.reserve(num_map_windows);
 #if REALM_PLATFORM_APPLE && REALM_MOBILE
@@ -235,6 +231,16 @@ GroupWriter::GroupWriter(Group& group, Durability dura, WriteMarker* write_marke
         m_window_alignment = wanted_size;
     }
 #endif
+}
+
+GroupComitter::~GroupComitter() = default;
+
+GroupWriter::GroupWriter(Group& group, Durability dura, WriteMarker* write_marker)
+    : GroupComitter(group, dura, write_marker)
+    , m_free_positions(m_alloc)
+    , m_free_lengths(m_alloc)
+    , m_free_versions(m_alloc)
+{
     Array& top = m_group.m_top;
     m_logical_size = size_t(top.get_as_ref_or_tagged(Group::s_file_size_ndx).get_as_int());
 
@@ -332,14 +338,14 @@ size_t GroupWriter::get_file_size() const noexcept
     return sz;
 }
 
-void GroupWriter::flush_all_mappings()
+void GroupComitter::flush_all_mappings()
 {
     for (const auto& window : m_map_windows) {
         window->flush();
     }
 }
 
-void GroupWriter::sync_all_mappings()
+void GroupComitter::sync_all_mappings()
 {
     if (m_durability == Durability::Unsafe)
         return;
@@ -352,7 +358,7 @@ void GroupWriter::sync_all_mappings()
 // existing one (possibly extended to accomodate the new request). Maintain a
 // cache of open windows which are sync'ed and closed following a least recently
 // used policy. Entries in the cache are kept in MRU order.
-GroupWriter::MapWindow* GroupWriter::get_window(ref_type start_ref, size_t size)
+GroupComitter::MapWindow* GroupComitter::get_window(ref_type start_ref, size_t size)
 {
     auto match = std::find_if(m_map_windows.begin(), m_map_windows.end(), [&](const auto& window) {
         return window->matches(start_ref, size) || window->extends_to_match(m_alloc.get_file(), start_ref, size);
@@ -1335,7 +1341,7 @@ void GroupWriter::write_array_at(T* translator, ref_type ref, const char* data,
 }
 
 
-void GroupWriter::commit(ref_type new_top_ref)
+void GroupComitter::commit(ref_type new_top_ref)
 {
     using _impl::SimulatedFailure;
     SimulatedFailure::trigger(SimulatedFailure::group_writer__commit); // Throws

src/realm/group_writer.hpp

@@ -57,9 +57,46 @@ class VersionInfo {
 using TopRefMap = std::map<uint64_t, VersionInfo>;
 using VersionVector = std::vector<uint64_t>;
 
+class GroupComitter {
+public:
+    using Durability = DBOptions::Durability;
+    GroupComitter(Group&, Durability dura = Durability::Full, util::WriteMarker* write_marker = nullptr);
+    ~GroupComitter();
+    /// Flush changes to physical medium, then write the new top ref
+    /// to the file header, then flush again. Pass the top ref
+    /// returned by write_group().
+    void commit(ref_type new_top_ref);
+    // Flush all cached memory mappings
+    // Sync all cached memory mappings to disk - includes flush if needed
+    void sync_all_mappings();
+    // Flush all cached memory mappings from private to shared cache.
+    void flush_all_mappings();
+
+protected:
+    class MapWindow;
+    Group& m_group;
+    SlabAlloc& m_alloc;
+    // Currently cached memory mappings. We keep as many as 16 1MB windows
+    // open for writing. The allocator will favor sequential allocation
+    // from a modest number of windows, depending upon fragmentation, so
+    // 16 windows should be more than enough. If more than 16 windows are
+    // needed, the least recently used is sync'ed and closed to make room
+    // for a new one. The windows are kept in MRU (most recently used) order.
+    const static int num_map_windows = 16;
+    std::vector<std::unique_ptr<MapWindow>> m_map_windows;
+    size_t m_window_alignment;
+    util::WriteMarker* m_write_marker;
+    Durability m_durability;
+
+    // Get a suitable memory mapping for later access:
+    // potentially adding it to the cache, potentially closing
+    // the least recently used and sync'ing it to disk
+    MapWindow* get_window(ref_type start_ref, size_t size);
+};
+
 /// This class is not supposed to be reused for multiple write sessions. In
 /// particular, do not reuse it in case any of the functions throw.
-class GroupWriter : public _impl::ArrayWriterBase {
+class GroupWriter : public GroupComitter, public _impl::ArrayWriterBase {
 public:
     enum class EvacuationStage { idle, evacuating, waiting, blocked };
     // For groups in transactional mode (Group::m_is_shared), this constructor
@@ -71,7 +108,6 @@ class GroupWriter : public _impl::ArrayWriterBase {
     // (Group::m_is_shared), the constructor also adds version tracking
     // information to the group, if it is not already present (6th and 7th entry
     // in Group::m_top).
-    using Durability = DBOptions::Durability;
     GroupWriter(Group&, Durability dura = Durability::Full, util::WriteMarker* write_marker = nullptr);
     ~GroupWriter();
 
@@ -83,10 +119,6 @@ class GroupWriter : public _impl::ArrayWriterBase {
     /// commit() with the returned top ref.
     ref_type write_group();
 
-    /// Flush changes to physical medium, then write the new top ref
-    /// to the file header, then flush again. Pass the top ref
-    /// returned by write_group().
-    void commit(ref_type new_top_ref);
 
     size_t get_file_size() const noexcept;
 
@@ -146,13 +178,6 @@ class GroupWriter : public _impl::ArrayWriterBase {
         }
     }
 
-
-    // Flush all cached memory mappings
-    // Sync all cached memory mappings to disk - includes flush if needed
-    void sync_all_mappings();
-    // Flush all cached memory mappings from private to shared cache.
-    void flush_all_mappings();
-
 private:
     friend class InMemoryWriter;
     struct FreeSpaceEntry {
@@ -171,24 +196,18 @@ class GroupWriter : public _impl::ArrayWriterBase {
     static void move_free_in_file_to_size_map(const std::vector<GroupWriter::FreeSpaceEntry>& list,
                                               std::multimap<size_t, size_t>& size_map);
 
-    class MapWindow;
-    Group& m_group;
-    SlabAlloc& m_alloc;
     Array m_free_positions; // 4th slot in Group::m_top
     Array m_free_lengths;   // 5th slot in Group::m_top
     Array m_free_versions;  // 6th slot in Group::m_top
     uint64_t m_current_version = 0;
     uint64_t m_oldest_reachable_version;
     TopRefMap m_top_ref_map;
     bool m_any_new_unreachables;
-    size_t m_window_alignment;
     size_t m_free_space_size = 0;
     size_t m_locked_space_size = 0;
     size_t m_evacuation_limit;
     int64_t m_backoff;
     size_t m_logical_size = 0;
-    Durability m_durability;
-    util::WriteMarker* m_write_marker = nullptr;
 
     //  m_free_in_file;
     std::vector<FreeSpaceEntry> m_not_free_in_file;
@@ -199,19 +218,6 @@ class GroupWriter : public _impl::ArrayWriterBase {
 
     void read_in_freelist();
     size_t recreate_freelist(size_t reserve_pos);
-    // Currently cached memory mappings. We keep as many as 16 1MB windows
-    // open for writing. The allocator will favor sequential allocation
-    // from a modest number of windows, depending upon fragmentation, so
-    // 16 windows should be more than enough. If more than 16 windows are
-    // needed, the least recently used is sync'ed and closed to make room
-    // for a new one. The windows are kept in MRU (most recently used) order.
-    const static int num_map_windows = 16;
-    std::vector<std::unique_ptr<MapWindow>> m_map_windows;
-
-    // Get a suitable memory mapping for later access:
-    // potentially adding it to the cache, potentially closing
-    // the least recently used and sync'ing it to disk
-    MapWindow* get_window(ref_type start_ref, size_t size);
 
     /// Allocate a chunk of free space of the specified size. The
     /// specified size must be 8-byte aligned. Extend the file if

src/realm/transaction.cpp

@@ -725,7 +725,7 @@ void Transaction::complete_async_commit()
         if (db->m_logger) {
             db->m_logger->log(util::Logger::Level::trace, "Comitting ref %1 to disk", read_lock.m_top_ref);
         }
-        GroupWriter out(*this);
+        GroupComitter out(*this);
         out.commit(read_lock.m_top_ref); // Throws
         // we must release the write mutex before the callback, because the callback
         // is allowed to re-request it.