mdbx: rework mdbx_filesync().

src/mdbx.c

@@ -3041,7 +3041,7 @@ __cold static int mdbx_env_sync_ex(MDBX_env *env, int force, int nonblock) {
       int rc = (flags & MDBX_WRITEMAP)
                    ? mdbx_msync(&env->me_dxb_mmap, 0, usedbytes,
                                 flags & MDBX_MAPASYNC)
-                   : mdbx_filesync(env->me_fd, false);
+                   : mdbx_filesync(env->me_fd, MDBX_SYNC_DATA);
       if (unlikely(rc != MDBX_SUCCESS))
         return rc;
 
@@ -5422,14 +5422,16 @@ static int mdbx_sync_locked(MDBX_env *env, unsigned flags,
         goto fail;
       if ((flags & MDBX_MAPASYNC) == 0) {
         if (unlikely(pending->mm_geo.next > steady->mm_geo.now)) {
-          rc = mdbx_filesize_sync(env->me_fd);
+          rc = mdbx_filesync(env->me_fd, MDBX_SYNC_SIZE);
           if (unlikely(rc != MDBX_SUCCESS))
             goto fail;
         }
         env->me_sync_pending = 0;
       }
     } else {
-      rc = mdbx_filesync(env->me_fd, pending->mm_geo.next > steady->mm_geo.now);
+      rc = mdbx_filesync(env->me_fd, (pending->mm_geo.next > steady->mm_geo.now)
+                                         ? MDBX_SYNC_DATA | MDBX_SYNC_SIZE
+                                         : MDBX_SYNC_DATA);
       if (unlikely(rc != MDBX_SUCCESS))
         goto fail;
       env->me_sync_pending = 0;
@@ -5577,7 +5579,7 @@ static int mdbx_sync_locked(MDBX_env *env, unsigned flags,
       if (unlikely(rc != MDBX_SUCCESS))
         goto fail;
     } else {
-      rc = mdbx_filesync(env->me_fd, false);
+      rc = mdbx_filesync(env->me_fd, MDBX_SYNC_DATA | MDBX_SYNC_IODQ);
       if (rc != MDBX_SUCCESS)
         goto undo;
     }
@@ -11825,12 +11827,15 @@ int __cold mdbx_env_copy2fd(MDBX_env *env, mdbx_filehandle_t fd,
   mdbx_txn_abort(read_txn);
 
   if (likely(rc == MDBX_SUCCESS))
-    rc = mdbx_filesync(fd, true);
+    rc = mdbx_filesync(fd, MDBX_SYNC_DATA | MDBX_SYNC_SIZE);
 
   /* Write actual meta */
   if (likely(rc == MDBX_SUCCESS))
     rc = mdbx_pwrite(fd, buffer, pgno2bytes(env, NUM_METAS), 0);
 
+  if (likely(rc == MDBX_SUCCESS))
+    rc = mdbx_filesync(fd, MDBX_SYNC_DATA | MDBX_SYNC_IODQ);
+
   mdbx_memalign_free(buffer);
   return rc;
 }
@@ -12871,7 +12876,7 @@ int __cold mdbx_setup_debug(int flags, MDBX_debug_func *logger) {
   unsigned ret = mdbx_runtime_flags;
   mdbx_runtime_flags = flags;
 
-#ifdef __linux__
+#if defined(__linux__) || defined(__gnu_linux__)
   if (flags & MDBX_DBG_DUMP) {
     int core_filter_fd = open("/proc/self/coredump_filter", O_TRUNC | O_RDWR);
     if (core_filter_fd >= 0) {
@@ -12894,7 +12899,7 @@ int __cold mdbx_setup_debug(int flags, MDBX_debug_func *logger) {
       close(core_filter_fd);
     }
   }
-#endif /* __linux__ */
+#endif /* Linux */
 
   mdbx_debug_logger = logger;
   return ret;

src/osal.c

@@ -660,14 +660,22 @@ int mdbx_pwritev(mdbx_filehandle_t fd, struct iovec *iov, int iovcnt,
 #endif
 }
 
-int mdbx_filesync(mdbx_filehandle_t fd, bool filesize_changed) {
+int mdbx_filesync(mdbx_filehandle_t fd, enum mdbx_syncmode_bits mode_bits) {
 #if defined(_WIN32) || defined(_WIN64)
-  (void)filesize_changed;
+  return ((mode_bits & (MDBX_SYNC_DATA | MDBX_SYNC_IODQ)) == 0 ||
-  return FlushFileBuffers(fd) ? MDBX_SUCCESS : GetLastError();
+          FlushFileBuffers(fd))
-#elif defined(__APPLE__)
+             ? MDBX_SUCCESS
-  (void)filesize_changed;
+             : GetLastError();
-  return likely(fcntl(fd, F_FULLFSYNC) != -1) ? MDBX_SUCCESS : errno;
 #else
+
+#ifdef __APPLE__
+  if (mode_bits & MDBX_SYNC_IODQ)
+    return likely(fcntl(fd, F_FULLFSYNC) != -1) ? MDBX_SUCCESS : errno;
+#endif /* MacOS */
+#if defined(__linux__) || defined(__gnu_linux__)
+  if (mode_bits == MDBX_SYNC_SIZE && linux_kernel_version >= 0x03060000)
+    return MDBX_SUCCESS;
+#endif /* Linux */
   int rc;
   do {
 #if defined(_POSIX_SYNCHRONIZED_IO) && _POSIX_SYNCHRONIZED_IO > 0
@@ -676,12 +684,12 @@ int mdbx_filesync(mdbx_filehandle_t fd, bool filesize_changed) {
      *
      * For more info about of a corresponding fdatasync() bug
      * see http://www.spinics.net/lists/linux-ext4/msg33714.html */
-    if (!filesize_changed) {
+    if ((mode_bits & MDBX_SYNC_SIZE) == 0) {
       if (fdatasync(fd) == 0)
         return MDBX_SUCCESS;
     } else
 #else
-    (void)filesize_changed;
+    (void)mode_bits;
 #endif
         if (fsync(fd) == 0)
       return MDBX_SUCCESS;
@@ -691,22 +699,6 @@ int mdbx_filesync(mdbx_filehandle_t fd, bool filesize_changed) {
 #endif
 }
 
-int mdbx_filesize_sync(mdbx_filehandle_t fd) {
-#if defined(_WIN32) || defined(_WIN64)
-  (void)fd;
-  /* Nothing on Windows (i.e. newer 100% steady) */
-  return MDBX_SUCCESS;
-#else
-  for (;;) {
-    if (fsync(fd) == 0)
-      return MDBX_SUCCESS;
-    int rc = errno;
-    if (rc != EINTR)
-      return rc;
-  }
-#endif
-}
-
 int mdbx_filesize(mdbx_filehandle_t fd, uint64_t *length) {
 #if defined(_WIN32) || defined(_WIN64)
   BY_HANDLE_FILE_INFORMATION info;

src/osal.h

@@ -522,7 +522,13 @@ int mdbx_thread_create(mdbx_thread_t *thread,
                        void *arg);
 int mdbx_thread_join(mdbx_thread_t thread);
 
-int mdbx_filesync(mdbx_filehandle_t fd, bool fullsync);
+enum mdbx_syncmode_bits {
+  MDBX_SYNC_DATA = 1,
+  MDBX_SYNC_SIZE = 2,
+  MDBX_SYNC_IODQ = 4
+};
+
+int mdbx_filesync(mdbx_filehandle_t fd, enum mdbx_syncmode_bits mode_bits);
 int mdbx_filesize_sync(mdbx_filehandle_t fd);
 int mdbx_ftruncate(mdbx_filehandle_t fd, uint64_t length);
 int mdbx_fseek(mdbx_filehandle_t fd, uint64_t pos);