Fixed endless loop for ARC_EVICT_ALL eviction

module/zfs/arc.c

@@ -4077,22 +4077,21 @@ typedef struct evict_arg {
 	arc_buf_hdr_t	*marker;
 	uint64_t	spa;
 	uint64_t	bytes;
-	uint64_t	*evicted_ptr;
+	uint64_t	evicted;
 } evict_arg_t;
 
 static void
 arc_evict_task(void *arg)
 {
 	evict_arg_t *eva = arg;
-	uint64_t *evictedp = eva->evicted_ptr;
 	multilist_t *ml = eva->ml;
 	arc_buf_hdr_t *marker = eva->marker;
 	int idx = eva->idx;
 	uint64_t spa = eva->spa;
 	uint64_t evict = eva->bytes;
-	uint64_t bytes_evicted = arc_evict_state_impl(ml, idx, marker, spa,
+
+	eva->evicted = arc_evict_state_impl(ml, idx, marker, spa,
 	    evict);
-	atomic_add_64(evictedp, bytes_evicted);
 }
 
 /*
@@ -4172,21 +4171,33 @@ arc_evict_state(arc_state_t *state, arc_buf_contents_t type, uint64_t spa,
 		    bytes - total_evicted);
 		uint64_t evict;
 		uint_t ntasks;
+		uint_t m = 0;
 
 		if (use_evcttq) {
-			if (left > nthreads * MIN_EVICT_SIZE) {
-				evict = DIV_ROUND_UP(left, nthreads);
+			/*
+			 * Scaling algorithm:
+			 * 1. Divide the left bytes by chunks each
+			 *    MIN_EVICT_SIZE.
+			 * 2. Equaly distribute the chunks between the threads.
+			 * 3. First (m) threads get an additional chunk each.
+			 */
+			if (left == ARC_EVICT_ALL) {
+				ntasks = nthreads;
+				evict = left;
+			} else if (left > nthreads * MIN_EVICT_SIZE) {
 				ntasks = nthreads;
+				uint64_t n = DIV_ROUND_UP(left, MIN_EVICT_SIZE);
+				evict = (n / nthreads) * MIN_EVICT_SIZE;
+				m = n % nthreads;	 /* additional chunks */
 			} else {
-				evict = MIN_EVICT_SIZE;
 				ntasks = DIV_ROUND_UP(left, MIN_EVICT_SIZE);
+				evict = MIN_EVICT_SIZE;
 			}
 		} else {
 			ntasks = num_sublists;
 			evict = left;
 		}
 
-
 		/*
 		 * Start eviction using a randomly selected sublist,
 		 * this is to try and evenly balance eviction across all
@@ -4207,9 +4218,10 @@ arc_evict_state(arc_state_t *state, arc_buf_contents_t type, uint64_t spa,
 				evarg[i].ml = ml;
 				evarg[i].marker = markers[sublist_idx];
 				evarg[i].spa = spa;
-				evarg[i].evicted_ptr = &scan_evicted;
 				evarg[i].idx = sublist_idx;
 				evarg[i].bytes = evict;
+				if (i < m)
+					evarg[i].bytes += MIN_EVICT_SIZE;
 
 				taskq_dispatch_ent(arc_evict_taskq,
 				    arc_evict_task,
@@ -4231,7 +4243,10 @@ arc_evict_state(arc_state_t *state, arc_buf_contents_t type, uint64_t spa,
 
 		if (use_evcttq) {
 			taskq_wait(arc_evict_taskq);
-			total_evicted += scan_evicted;
+			for (unsigned i = 0; i < ntasks; i++) {
+				scan_evicted += evarg[i].evicted;
+				total_evicted += evarg[i].evicted;
+			}
 		}
 
 		/*