mdbx: add timed auto-sync feature.

Change-Id: Ia9b8529fda321d5f78b306f270d157a78f708916

mdbx.h

@@ -1601,6 +1601,32 @@ LIBMDBX_API int mdbx_env_close_ex(MDBX_env *env, int dont_sync);
  * Returns A non-zero error value on failure and 0 on success. */
 LIBMDBX_API int mdbx_env_set_syncbytes(MDBX_env *env, size_t bytes);
 
+/* Sets relative period since the last unsteay commit to force flush the data
+ * buffers to disk, even of MDBX_NOSYNC, MDBX_NOMETASYNC and MDBX_MAPASYNC flags
+ * in the environment. The value affects all processes which operates with given
+ * DB until the last process close DB or a new value will be settled.
+ *
+ * Data is always written to disk when mdbx_txn_commit() is called,
+ * but the operating system may keep it buffered. MDBX always flushes
+ * the OS buffers upon commit as well, unless the environment was
+ * opened with MDBX_NOSYNC, MDBX_MAPASYNC or in part MDBX_NOMETASYNC.
+ *
+ * Settled period don't checked asynchronously, but only inside the functions.
+ * mdbx_txn_commit() and mdbx_env_sync(). Therefore, in cases where transactions
+ * are committed infrequently and/or irregularly, polling by mdbx_env_sync() may
+ * be a reasonable solution to timeout enforcement.
+ *
+ * The default is 0, than mean no any timeout checked, and no additional
+ * flush will be made.
+ *
+ * [in] env               An environment handle returned by mdbx_env_create()
+ * [in] seconds_16dot16   The period in 1/65536 of second when a synchronous
+ *                        flush would be made since the last unsteay commit.
+ *
+ * Returns A non-zero error value on failure and 0 on success. */
+LIBMDBX_API int mdbx_env_set_syncperiod(MDBX_env *env,
+                                        unsigned seconds_16dot16);
+
 /* Returns a lag of the reading for the given transaction.
  *
  * Returns an information for estimate how much given read-only

src/bits.h

@@ -810,6 +810,8 @@ struct MDBX_env {
   unsigned me_maxkey_limit;  /* max size of a key */
   mdbx_pid_t me_live_reader; /* have liveness lock in reader table */
   void *me_userctx;          /* User-settable context */
+  volatile uint64_t *me_unsynced_timeout;
+  volatile uint64_t *me_autosync_period;
   volatile pgno_t *me_unsynced_pages;
   volatile pgno_t *me_autosync_threshold;
   MDBX_oom_func *me_oom_func; /* Callback for kicking laggard readers */
@@ -818,6 +820,8 @@ struct MDBX_env {
     MDBX_OSAL_LOCK wmutex;
 #endif
     txnid_t oldest;
+    uint64_t unsynced_timeout;
+    uint64_t autosync_period;
     pgno_t autosync_pending;
     pgno_t autosync_threshold;
   } me_lckless_stub;

src/mdbx.c

@@ -3025,7 +3025,9 @@ __cold static int mdbx_env_sync_ex(MDBX_env *env, int force, int nonblock) {
   pgno_t unsynced_pages = *env->me_unsynced_pages;
   if (!META_IS_STEADY(head) || unsynced_pages) {
     const pgno_t autosync_threshold = *env->me_autosync_threshold;
-    if (force || (autosync_threshold && unsynced_pages >= autosync_threshold))
+    const uint64_t unsynced_timeout = *env->me_unsynced_timeout;
+    if (force || (autosync_threshold && unsynced_pages >= autosync_threshold) ||
+        (unsynced_timeout && mdbx_osal_monotime() >= unsynced_timeout))
       flags &= MDBX_WRITEMAP /* clear flags for full steady sync */;
 
     if (outside_txn) {
@@ -5390,7 +5392,9 @@ static int mdbx_sync_locked(MDBX_env *env, unsigned flags,
   if (flags & (MDBX_NOSYNC | MDBX_MAPASYNC)) {
     /* Check auto-sync conditions */
     const pgno_t autosync_threshold = *env->me_autosync_threshold;
-    if (autosync_threshold && *env->me_unsynced_pages >= autosync_threshold)
+    const uint64_t unsynced_timeout = *env->me_unsynced_timeout;
+    if ((autosync_threshold && *env->me_unsynced_pages >= autosync_threshold) ||
+        (unsynced_timeout && mdbx_osal_monotime() >= unsynced_timeout))
       flags &= MDBX_WRITEMAP | MDBX_SHRINK_ALLOWED; /* force steady */
   }
 
@@ -5457,8 +5461,12 @@ static int mdbx_sync_locked(MDBX_env *env, unsigned flags,
   if (rc == MDBX_RESULT_FALSE /* carry steady */) {
     pending->mm_datasync_sign = mdbx_meta_sign(pending);
     *env->me_unsynced_pages = 0;
+    *env->me_unsynced_timeout = 0;
   } else {
     assert(rc == MDBX_RESULT_TRUE /* carry non-steady */);
+    const uint64_t autosync_period = *env->me_autosync_period;
+    if (autosync_period && *env->me_unsynced_timeout == 0)
+      *env->me_unsynced_timeout = mdbx_osal_monotime() + autosync_period;
     pending->mm_datasync_sign =
         (flags & MDBX_UTTERLY_NOSYNC) == MDBX_UTTERLY_NOSYNC
             ? MDBX_DATASIGN_NONE
@@ -6430,6 +6438,8 @@ static int __cold mdbx_setup_lck(MDBX_env *env, char *lck_pathname,
       return rc;
 
     env->me_oldest = &env->me_lckless_stub.oldest;
+    env->me_unsynced_timeout = &env->me_lckless_stub.unsynced_timeout;
+    env->me_autosync_period = &env->me_lckless_stub.autosync_period;
     env->me_unsynced_pages = &env->me_lckless_stub.autosync_pending;
     env->me_autosync_threshold = &env->me_lckless_stub.autosync_threshold;
     env->me_maxreaders = UINT_MAX;
@@ -6540,6 +6550,8 @@ static int __cold mdbx_setup_lck(MDBX_env *env, char *lck_pathname,
 
   mdbx_assert(env, !MDBX_IS_ERROR(rc));
   env->me_oldest = &env->me_lck->mti_oldest_reader;
+  env->me_unsynced_timeout = &env->me_lck->mti_unsynced_timeout;
+  env->me_autosync_period = &env->me_lck->mti_autosync_period;
   env->me_unsynced_pages = &env->me_lck->mti_unsynced_pages;
   env->me_autosync_threshold = &env->me_lck->mti_autosync_threshold;
 #ifdef MDBX_OSAL_LOCK
@@ -6788,6 +6800,8 @@ static void __cold mdbx_env_close0(MDBX_env *env) {
   if (env->me_lck)
     mdbx_munmap(&env->me_lck_mmap);
   env->me_oldest = nullptr;
+  env->me_unsynced_timeout = nullptr;
+  env->me_autosync_period = nullptr;
   env->me_unsynced_pages = nullptr;
   env->me_autosync_threshold = nullptr;
 
@@ -13028,6 +13042,23 @@ int __cold mdbx_env_set_syncbytes(MDBX_env *env, size_t bytes) {
   return bytes ? mdbx_env_sync(env, false) : MDBX_SUCCESS;
 }
 
+int __cold mdbx_env_set_syncperiod(MDBX_env *env, unsigned seconds_16dot16) {
+  if (unlikely(!env))
+    return MDBX_EINVAL;
+
+  if (unlikely(env->me_signature != MDBX_ME_SIGNATURE))
+    return MDBX_EBADSIGN;
+
+  if (unlikely(env->me_flags & (MDBX_RDONLY | MDBX_FATAL_ERROR)))
+    return MDBX_EACCESS;
+
+  if (unlikely(!env->me_map))
+    return MDBX_EPERM;
+
+  *env->me_autosync_period = mdbx_osal_16dot16_to_monotime(seconds_16dot16);
+  return seconds_16dot16 ? mdbx_env_sync(env, false) : MDBX_SUCCESS;
+}
+
 int __cold mdbx_env_set_oomfunc(MDBX_env *env, MDBX_oom_func *oomfunc) {
   if (unlikely(!env))
     return MDBX_EINVAL;

src/osal.c

@@ -1223,3 +1223,71 @@ __cold void mdbx_osal_jitter(bool tiny) {
 #endif
   }
 }
+
+#if defined(_WIN32) || defined(_WIN64)
+#elif defined(__APPLE__) || defined(__MACH__)
+#include <mach/mach_time.h>
+#elif defined(__linux__) || defined(__gnu_linux__)
+static __cold clockid_t choise_monoclock() {
+  struct timespec probe;
+#if defined(CLOCK_BOOTTIME)
+  if (clock_gettime(CLOCK_BOOTTIME, &probe) == 0)
+    return CLOCK_BOOTTIME;
+#elif defined(CLOCK_MONOTONIC_RAW)
+  if (clock_gettime(CLOCK_MONOTONIC_RAW, &probe) == 0)
+    return CLOCK_MONOTONIC_RAW;
+#elif defined(CLOCK_MONOTONIC_COARSE)
+  if (clock_gettime(CLOCK_MONOTONIC_COARSE, &probe) == 0)
+    return CLOCK_MONOTONIC_COARSE;
+#endif
+  return CLOCK_MONOTONIC;
+}
+#endif
+
+uint64_t mdbx_osal_16dot16_to_monotime(uint32_t seconds_16dot16) {
+#if defined(_WIN32) || defined(_WIN64)
+  static LARGE_INTEGER performance_frequency;
+  if (performance_frequency.QuadPart == 0)
+    QueryPerformanceFrequency(&performance_frequency);
+  const uint64_t ratio = performance_frequency.QuadPart;
+#elif defined(__APPLE__) || defined(__MACH__)
+  static uint64_t ratio;
+  if (!ratio) {
+    mach_timebase_info_data_t ti;
+    mach_timebase_info(&ti);
+    ratio = UINT64_C(1000000000) * ti.denom / ti.numer;
+  }
+#else
+  const uint64_t ratio = UINT64_C(1000000000);
+#endif
+  return (ratio * seconds_16dot16 + 32768) >> 16;
+}
+
+uint64_t mdbx_osal_monotime(void) {
+#if defined(_WIN32) || defined(_WIN64)
+  LARGE_INTEGER counter;
+  counter.QuadPart = 0;
+  QueryPerformanceCounter(&counter);
+  return counter.QuadPart;
+#elif defined(__APPLE__) || defined(__MACH__)
+  return mach_absolute_time();
+#else
+
+#if defined(__linux__) || defined(__gnu_linux__)
+  static clockid_t posix_clockid = -1;
+  if (unlikely(posix_clockid < 0))
+    posix_clockid = choise_monoclock();
+#elif defined(CLOCK_MONOTONIC)
+#define posix_clockid CLOCK_MONOTONIC
+#else
+#define posix_clockid CLOCK_REALTIME
+#endif
+
+  struct timespec ts;
+  if (unlikely(clock_gettime(posix_clockid, &ts) != 0)) {
+    ts.tv_nsec = 0;
+    ts.tv_sec = 0;
+  }
+  return ts.tv_sec * UINT64_C(1000000000) + ts.tv_nsec;
+#endif
+}

src/osal.h

@@ -599,6 +599,8 @@ static __inline mdbx_tid_t mdbx_thread_self(void) {
 }
 
 void mdbx_osal_jitter(bool tiny);
+uint64_t mdbx_osal_monotime(void);
+uint64_t mdbx_osal_16dot16_to_monotime(uint32_t seconds_16dot16);
 
 /*----------------------------------------------------------------------------*/
 /* lck stuff */