Merge dd97d59744 into 138263a1b4

src/dict.c

@@ -44,6 +44,18 @@
 static dictResizeEnable dict_can_resize = DICT_RESIZE_ENABLE;
 static unsigned int dict_force_resize_ratio = 4;
 
+/* Accumulator of microseconds spent in incremental rehash steps performed
+ * inside dict lookup/insert/delete operations. Read and reset by higher
+ * layers (server.c call()) to attribute per-command rehash cost. */
+uint64_t dictRehashStepElapsedUs = 0;
+
+/* Non-zero when latency monitoring is active
+ * (server.latency_monitor_threshold != 0). Set by server.c at the outermost
+ * call() entry and cleared at the outermost call() exit. While zero, the
+ * rehash timing path is skipped entirely: no clock reads, no accumulator
+ * writes - just one predictable branch. */
+int dictRehashStepTiming = 0;
+
 /* -------------------------- types ----------------------------------------- */
 struct dictEntry {
     struct dictEntry *next;  /* Must be first */
@@ -466,7 +478,19 @@ int dictRehashMicroseconds(dict *d, uint64_t us) {
  * dictionary so that the hash table automatically migrates from H1 to H2
  * while it is actively used. */
 static void _dictRehashStep(dict *d) {
-    if (d->pauserehash == 0) dictRehash(d,1);
+    if (d->pauserehash == 0) {
+        /* Skip the timing path when latency monitoring is off, and also
+         * when getMonotonicUs is still NULL during early startup (core
+         * module registration runs dictAdd before monotonicInit). */
+        if (!dictRehashStepTiming || getMonotonicUs == NULL) {
+            dictRehash(d,1);
+            return;
+        }
+        monotime t;
+        elapsedStart(&t);
+        dictRehash(d,1);
+        dictRehashStepElapsedUs += elapsedUs(t);
+    }
 }
 
 /* Performs rehashing on a single bucket. */
@@ -1705,6 +1729,12 @@ static void _dictShrinkIfNeeded(dict *d)
 static void _dictRehashStepIfNeeded(dict *d, uint64_t visitedIdx) {
     if ((!dictIsRehashing(d)) || (d->pauserehash != 0))
         return;
+    /* Skip the timing path when latency monitoring is off, and also when
+     * getMonotonicUs is still NULL during early startup (core module
+     * registration runs dictAdd before monotonicInit). */
+    const int time_it = dictRehashStepTiming && (getMonotonicUs != NULL);
+    monotime t;
+    if (time_it) elapsedStart(&t);
     /* rehashing not in progress if rehashidx == -1 */
     if ((long)visitedIdx >= d->rehashidx && d->ht_table[0][visitedIdx]) {
         /* If we have a valid hash entry at `idx` in ht0, we perform
@@ -1715,6 +1745,7 @@ static void _dictRehashStepIfNeeded(dict *d, uint64_t visitedIdx) {
          * on the rehashidx (not CPU cache friendly). */
         dictRehash(d,1);
     }
+    if (time_it) dictRehashStepElapsedUs += elapsedUs(t);
 }
 
 /* Our hash table capability is a power of two */

src/dict.h

@@ -292,6 +292,8 @@ void dictEmpty(dict *d, void(callback)(dict*));
 void dictSetResizeEnabled(dictResizeEnable enable);
 int dictRehash(dict *d, int n);
 int dictRehashMicroseconds(dict *d, uint64_t us);
+extern uint64_t dictRehashStepElapsedUs;
+extern int dictRehashStepTiming;
 void dictSetHashFunctionSeed(uint8_t *seed);
 unsigned long dictScan(dict *d, unsigned long v, dictScanFunction *fn, void *privdata);
 unsigned long dictScanDefrag(dict *d, unsigned long v, dictScanFunction *fn, dictDefragFunctions *defragfns, void *privdata);

src/server.c

@@ -3983,6 +3983,17 @@ void call(client *c, int flags) {
     /* Pass current server.ustime to avoid ustime() call if monotonic clock is used
      * and time will be updated before command execution based on monotonic clock. */
     const long long call_timer = use_hw_clock ? server.ustime : ustime();
+    /* Arm dict-rehash attribution at the outermost call() only. The
+     * accumulator starts at zero so that only rehash work performed during
+     * this top-level command's proc is attributed to it, and the timing
+     * flag is set so that the dict-side path starts collecting samples
+     * while this command runs. When the feature is disabled
+     * (latency_monitor_threshold == 0), the flag stays 0 and the entire
+     * timing path is a single predictable branch. */
+    if (server.execution_nesting == 0) {
+        dictRehashStepElapsedUs = 0;
+        dictRehashStepTiming = (server.latency_monitor_threshold != 0);
+    }
     enterExecutionUnit(1, call_timer);
 
     /* setting the CLIENT_EXECUTING_COMMAND flag so we will avoid
@@ -4039,10 +4050,20 @@ void call(client *c, int flags) {
         char *latency_event = (real_cmd->flags & CMD_FAST) ?
                                "fast-command" : "command";
         latencyAddSampleIfNeeded(latency_event,duration/1000);
-        if (server.execution_nesting == 0)
+        if (server.execution_nesting == 0) {
             durationAddSample(EL_DURATION_TYPE_CMD, duration);
+            latencyAddSampleIfNeeded("dict-rehash-during-command",
+                                     (long long)(dictRehashStepElapsedUs/1000));
+        }
     }
 
+    /* Disarm the dict-rehash timing path unconditionally at outermost call()
+     * exit, so cron work between commands does not pay the timing cost. This
+     * must run regardless of update_command_stats (e.g. during AOF loading)
+     * so the flag cannot stay armed across paths that skip the stats block. */
+    if (server.execution_nesting == 0)
+        dictRehashStepTiming = 0;
+
     /* Log the command into the Slow log if needed.
      * If the client is blocked we will handle slowlog when it is unblocked. */
     if (update_command_stats && !(c->flags & CLIENT_BLOCKED))