writeback: use RCU to protect bdi_list

Now that bdi_writeback_all() no longer handles integrity writeback, it doesn't have to block anymore. This means that we can switch bdi_list reader side protection to RCU. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

writeback: use RCU to protect bdi_list
Now that bdi_writeback_all() no longer handles integrity writeback, it doesn't have to block anymore. This means that we can switch bdi_list reader side protection to RCU. Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Jens Axboe
1 parent f11fcae840
Showing 4 changed files with 63 additions and 28 deletions Side-by-side Diff
fs/fs-writeback.c
include/linux/backing-dev.h
mm/backing-dev.c
mm/page-writeback.c
@@ -868,16 +868,16 @@
  
 	WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
  
-	spin_lock(&bdi_lock);
+	rcu_read_lock();
  
-	list_for_each_entry(bdi, &bdi_list, bdi_list) {
+	list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
 		if (!bdi_has_dirty_io(bdi))
 			continue;
  
 		bdi_alloc_queue_work(bdi, wbc);
 	}
  
-	spin_unlock(&bdi_lock);
+	rcu_read_unlock();
 }
  
 /*
@@ -59,6 +59,7 @@
  
 struct backing_dev_info {
 	struct list_head bdi_list;
+	struct rcu_head rcu_head;
 	unsigned long ra_pages;	/* max readahead in PAGE_CACHE_SIZE units */
 	unsigned long state;	/* Always use atomic bitops on this */
 	unsigned int capabilities; /* Device capabilities */
@@ -26,6 +26,12 @@
 EXPORT_SYMBOL_GPL(default_backing_dev_info);
  
 static struct class *bdi_class;
+
+/*
+ * bdi_lock protects updates to bdi_list and bdi_pending_list, as well as
+ * reader side protection for bdi_pending_list. bdi_list has RCU reader side
+ * locking.
+ */
 DEFINE_SPINLOCK(bdi_lock);
 LIST_HEAD(bdi_list);
 LIST_HEAD(bdi_pending_list);
@@ -284,9 +290,9 @@
 	/*
 	 * Add us to the active bdi_list
 	 */
-	spin_lock(&bdi_lock);
-	list_add(&bdi->bdi_list, &bdi_list);
-	spin_unlock(&bdi_lock);
+	spin_lock_bh(&bdi_lock);
+	list_add_rcu(&bdi->bdi_list, &bdi_list);
+	spin_unlock_bh(&bdi_lock);
  
 	bdi_task_init(bdi, wb);
  
@@ -389,7 +395,7 @@
 		if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list))
 			wb_do_writeback(me, 0);
  
-		spin_lock(&bdi_lock);
+		spin_lock_bh(&bdi_lock);
  
 		/*
 		 * Check if any existing bdi's have dirty data without
@@ -410,7 +416,7 @@
 		if (list_empty(&bdi_pending_list)) {
 			unsigned long wait;
  
-			spin_unlock(&bdi_lock);
+			spin_unlock_bh(&bdi_lock);
 			wait = msecs_to_jiffies(dirty_writeback_interval * 10);
 			schedule_timeout(wait);
 			try_to_freeze();
@@ -426,7 +432,7 @@
 		bdi = list_entry(bdi_pending_list.next, struct backing_dev_info,
 				 bdi_list);
 		list_del_init(&bdi->bdi_list);
-		spin_unlock(&bdi_lock);
+		spin_unlock_bh(&bdi_lock);
  
 		wb = &bdi->wb;
 		wb->task = kthread_run(bdi_start_fn, wb, "flush-%s",
  
@@ -445,9 +451,9 @@
 			 * a chance to flush other bdi's to free
 			 * memory.
 			 */
-			spin_lock(&bdi_lock);
+			spin_lock_bh(&bdi_lock);
 			list_add_tail(&bdi->bdi_list, &bdi_pending_list);
-			spin_unlock(&bdi_lock);
+			spin_unlock_bh(&bdi_lock);
  
 			bdi_flush_io(bdi);
 		}
@@ -456,6 +462,24 @@
 	return 0;
 }
  
+static void bdi_add_to_pending(struct rcu_head *head)
+{
+	struct backing_dev_info *bdi;
+
+	bdi = container_of(head, struct backing_dev_info, rcu_head);
+	INIT_LIST_HEAD(&bdi->bdi_list);
+
+	spin_lock(&bdi_lock);
+	list_add_tail(&bdi->bdi_list, &bdi_pending_list);
+	spin_unlock(&bdi_lock);
+
+	/*
+	 * We are now on the pending list, wake up bdi_forker_task()
+	 * to finish the job and add us back to the active bdi_list
+	 */
+	wake_up_process(default_backing_dev_info.wb.task);
+}
+
 /*
  * Add the default flusher task that gets created for any bdi
  * that has dirty data pending writeout
  
  
  
@@ -478,16 +502,29 @@
 	 * waiting for previous additions to finish.
 	 */
 	if (!test_and_set_bit(BDI_pending, &bdi->state)) {
-		list_move_tail(&bdi->bdi_list, &bdi_pending_list);
+		list_del_rcu(&bdi->bdi_list);
  
 		/*
-		 * We are now on the pending list, wake up bdi_forker_task()
-		 * to finish the job and add us back to the active bdi_list
+		 * We must wait for the current RCU period to end before
+		 * moving to the pending list. So schedule that operation
+		 * from an RCU callback.
 		 */
-		wake_up_process(default_backing_dev_info.wb.task);
+		call_rcu(&bdi->rcu_head, bdi_add_to_pending);
 	}
 }
  
+/*
+ * Remove bdi from bdi_list, and ensure that it is no longer visible
+ */
+static void bdi_remove_from_list(struct backing_dev_info *bdi)
+{
+	spin_lock_bh(&bdi_lock);
+	list_del_rcu(&bdi->bdi_list);
+	spin_unlock_bh(&bdi_lock);
+
+	synchronize_rcu();
+}
+
 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 		const char *fmt, ...)
 {
@@ -506,9 +543,9 @@
 		goto exit;
 	}
  
-	spin_lock(&bdi_lock);
-	list_add_tail(&bdi->bdi_list, &bdi_list);
-	spin_unlock(&bdi_lock);
+	spin_lock_bh(&bdi_lock);
+	list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
+	spin_unlock_bh(&bdi_lock);
  
 	bdi->dev = dev;
  
@@ -526,9 +563,7 @@
 			wb->task = NULL;
 			ret = -ENOMEM;
  
-			spin_lock(&bdi_lock);
-			list_del(&bdi->bdi_list);
-			spin_unlock(&bdi_lock);
+			bdi_remove_from_list(bdi);
 			goto exit;
 		}
 	}
@@ -565,9 +600,7 @@
 	/*
 	 * Make sure nobody finds us on the bdi_list anymore
 	 */
-	spin_lock(&bdi_lock);
-	list_del(&bdi->bdi_list);
-	spin_unlock(&bdi_lock);
+	bdi_remove_from_list(bdi);
  
 	/*
 	 * Finally, kill the kernel threads. We don't need to be RCU
@@ -599,6 +632,7 @@
 	bdi->max_ratio = 100;
 	bdi->max_prop_frac = PROP_FRAC_BASE;
 	spin_lock_init(&bdi->wb_lock);
+	INIT_RCU_HEAD(&bdi->rcu_head);
 	INIT_LIST_HEAD(&bdi->bdi_list);
 	INIT_LIST_HEAD(&bdi->wb_list);
 	INIT_LIST_HEAD(&bdi->work_list);
@@ -315,7 +315,7 @@
 {
 	int ret = 0;
  
-	spin_lock(&bdi_lock);
+	spin_lock_bh(&bdi_lock);
 	if (min_ratio > bdi->max_ratio) {
 		ret = -EINVAL;
 	} else {
@@ -327,7 +327,7 @@
 			ret = -EINVAL;
 		}
 	}
-	spin_unlock(&bdi_lock);
+	spin_unlock_bh(&bdi_lock);
  
 	return ret;
 }
  
@@ -339,14 +339,14 @@
 	if (max_ratio > 100)
 		return -EINVAL;
  
-	spin_lock(&bdi_lock);
+	spin_lock_bh(&bdi_lock);
 	if (bdi->min_ratio > max_ratio) {
 		ret = -EINVAL;
 	} else {
 		bdi->max_ratio = max_ratio;
 		bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
 	}
-	spin_unlock(&bdi_lock);
+	spin_unlock_bh(&bdi_lock);
  
 	return ret;
 }
...	...	@@ -868,16 +868,16 @@
868	868
869	869	WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
870	870
871		- spin_lock(&bdi_lock);
	871	+ rcu_read_lock();
872	872
873		- list_for_each_entry(bdi, &bdi_list, bdi_list) {
	873	+ list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
874	874	if (!bdi_has_dirty_io(bdi))
875	875	continue;
876	876
877	877	bdi_alloc_queue_work(bdi, wbc);
878	878	}
879	879
880		- spin_unlock(&bdi_lock);
	880	+ rcu_read_unlock();
881	881	}
882	882
883	883	/*
...	...	@@ -59,6 +59,7 @@
59	59
60	60	struct backing_dev_info {
61	61	struct list_head bdi_list;
	62	+ struct rcu_head rcu_head;
62	63	unsigned long ra_pages; /* max readahead in PAGE_CACHE_SIZE units */
63	64	unsigned long state; /* Always use atomic bitops on this */
64	65	unsigned int capabilities; /* Device capabilities */
...	...	@@ -26,6 +26,12 @@
26	26	EXPORT_SYMBOL_GPL(default_backing_dev_info);
27	27
28	28	static struct class *bdi_class;
	29	+
	30	+/*
	31	+ * bdi_lock protects updates to bdi_list and bdi_pending_list, as well as
	32	+ * reader side protection for bdi_pending_list. bdi_list has RCU reader side
	33	+ * locking.
	34	+ */
29	35	DEFINE_SPINLOCK(bdi_lock);
30	36	LIST_HEAD(bdi_list);
31	37	LIST_HEAD(bdi_pending_list);
...	...	@@ -284,9 +290,9 @@
284	290	/*
285	291	* Add us to the active bdi_list
286	292	*/
287		- spin_lock(&bdi_lock);
288		- list_add(&bdi->bdi_list, &bdi_list);
289		- spin_unlock(&bdi_lock);
	293	+ spin_lock_bh(&bdi_lock);
	294	+ list_add_rcu(&bdi->bdi_list, &bdi_list);
	295	+ spin_unlock_bh(&bdi_lock);
290	296
291	297	bdi_task_init(bdi, wb);
292	298
...	...	@@ -389,7 +395,7 @@
389	395	if (wb_has_dirty_io(me) \|\| !list_empty(&me->bdi->work_list))
390	396	wb_do_writeback(me, 0);
391	397
392		- spin_lock(&bdi_lock);
	398	+ spin_lock_bh(&bdi_lock);
393	399
394	400	/*
395	401	* Check if any existing bdi's have dirty data without
...	...	@@ -410,7 +416,7 @@
410	416	if (list_empty(&bdi_pending_list)) {
411	417	unsigned long wait;
412	418
413		- spin_unlock(&bdi_lock);
	419	+ spin_unlock_bh(&bdi_lock);
414	420	wait = msecs_to_jiffies(dirty_writeback_interval * 10);
415	421	schedule_timeout(wait);
416	422	try_to_freeze();
...	...	@@ -426,7 +432,7 @@
426	432	bdi = list_entry(bdi_pending_list.next, struct backing_dev_info,
427	433	bdi_list);
428	434	list_del_init(&bdi->bdi_list);
429		- spin_unlock(&bdi_lock);
	435	+ spin_unlock_bh(&bdi_lock);
430	436
431	437	wb = &bdi->wb;
432	438	wb->task = kthread_run(bdi_start_fn, wb, "flush-%s",
433	439
...	...	@@ -445,9 +451,9 @@
445	451	* a chance to flush other bdi's to free
446	452	* memory.
447	453	*/
448		- spin_lock(&bdi_lock);
	454	+ spin_lock_bh(&bdi_lock);
449	455	list_add_tail(&bdi->bdi_list, &bdi_pending_list);
450		- spin_unlock(&bdi_lock);
	456	+ spin_unlock_bh(&bdi_lock);
451	457
452	458	bdi_flush_io(bdi);
453	459	}
...	...	@@ -456,6 +462,24 @@
456	462	return 0;
457	463	}
458	464
	465	+static void bdi_add_to_pending(struct rcu_head *head)
	466	+{
	467	+ struct backing_dev_info *bdi;
	468	+
	469	+ bdi = container_of(head, struct backing_dev_info, rcu_head);
	470	+ INIT_LIST_HEAD(&bdi->bdi_list);
	471	+
	472	+ spin_lock(&bdi_lock);
	473	+ list_add_tail(&bdi->bdi_list, &bdi_pending_list);
	474	+ spin_unlock(&bdi_lock);
	475	+
	476	+ /*
	477	+ * We are now on the pending list, wake up bdi_forker_task()
	478	+ * to finish the job and add us back to the active bdi_list
	479	+ */
	480	+ wake_up_process(default_backing_dev_info.wb.task);
	481	+}
	482	+
459	483	/*
460	484	* Add the default flusher task that gets created for any bdi
461	485	* that has dirty data pending writeout
462	486
463	487
464	488
...	...	@@ -478,16 +502,29 @@
478	502	* waiting for previous additions to finish.
479	503	*/
480	504	if (!test_and_set_bit(BDI_pending, &bdi->state)) {
481		- list_move_tail(&bdi->bdi_list, &bdi_pending_list);
	505	+ list_del_rcu(&bdi->bdi_list);
482	506
483	507	/*
484		- * We are now on the pending list, wake up bdi_forker_task()
485		- * to finish the job and add us back to the active bdi_list
	508	+ * We must wait for the current RCU period to end before
	509	+ * moving to the pending list. So schedule that operation
	510	+ * from an RCU callback.
486	511	*/
487		- wake_up_process(default_backing_dev_info.wb.task);
	512	+ call_rcu(&bdi->rcu_head, bdi_add_to_pending);
488	513	}
489	514	}
490	515
	516	+/*
	517	+ * Remove bdi from bdi_list, and ensure that it is no longer visible
	518	+ */
	519	+static void bdi_remove_from_list(struct backing_dev_info *bdi)
	520	+{
	521	+ spin_lock_bh(&bdi_lock);
	522	+ list_del_rcu(&bdi->bdi_list);
	523	+ spin_unlock_bh(&bdi_lock);
	524	+
	525	+ synchronize_rcu();
	526	+}
	527	+
491	528	int bdi_register(struct backing_dev_info bdi, struct device parent,
492	529	const char *fmt, ...)
493	530	{
...	...	@@ -506,9 +543,9 @@
506	543	goto exit;
507	544	}
508	545
509		- spin_lock(&bdi_lock);
510		- list_add_tail(&bdi->bdi_list, &bdi_list);
511		- spin_unlock(&bdi_lock);
	546	+ spin_lock_bh(&bdi_lock);
	547	+ list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
	548	+ spin_unlock_bh(&bdi_lock);
512	549
513	550	bdi->dev = dev;
514	551
...	...	@@ -526,9 +563,7 @@
526	563	wb->task = NULL;
527	564	ret = -ENOMEM;
528	565
529		- spin_lock(&bdi_lock);
530		- list_del(&bdi->bdi_list);
531		- spin_unlock(&bdi_lock);
	566	+ bdi_remove_from_list(bdi);
532	567	goto exit;
533	568	}
534	569	}
...	...	@@ -565,9 +600,7 @@
565	600	/*
566	601	* Make sure nobody finds us on the bdi_list anymore
567	602	*/
568		- spin_lock(&bdi_lock);
569		- list_del(&bdi->bdi_list);
570		- spin_unlock(&bdi_lock);
	603	+ bdi_remove_from_list(bdi);
571	604
572	605	/*
573	606	* Finally, kill the kernel threads. We don't need to be RCU
...	...	@@ -599,6 +632,7 @@
599	632	bdi->max_ratio = 100;
600	633	bdi->max_prop_frac = PROP_FRAC_BASE;
601	634	spin_lock_init(&bdi->wb_lock);
	635	+ INIT_RCU_HEAD(&bdi->rcu_head);
602	636	INIT_LIST_HEAD(&bdi->bdi_list);
603	637	INIT_LIST_HEAD(&bdi->wb_list);
604	638	INIT_LIST_HEAD(&bdi->work_list);
...	...	@@ -315,7 +315,7 @@
315	315	{
316	316	int ret = 0;
317	317
318		- spin_lock(&bdi_lock);
	318	+ spin_lock_bh(&bdi_lock);
319	319	if (min_ratio > bdi->max_ratio) {
320	320	ret = -EINVAL;
321	321	} else {
...	...	@@ -327,7 +327,7 @@
327	327	ret = -EINVAL;
328	328	}
329	329	}
330		- spin_unlock(&bdi_lock);
	330	+ spin_unlock_bh(&bdi_lock);
331	331
332	332	return ret;
333	333	}
334	334
...	...	@@ -339,14 +339,14 @@
339	339	if (max_ratio > 100)
340	340	return -EINVAL;
341	341
342		- spin_lock(&bdi_lock);
	342	+ spin_lock_bh(&bdi_lock);
343	343	if (bdi->min_ratio > max_ratio) {
344	344	ret = -EINVAL;
345	345	} else {
346	346	bdi->max_ratio = max_ratio;
347	347	bdi->max_prop_frac = (PROP_FRAC_BASE * max_ratio) / 100;
348	348	}
349		- spin_unlock(&bdi_lock);
	349	+ spin_unlock_bh(&bdi_lock);
350	350
351	351	return ret;
352	352	}