[PATCH] md: modify locking when accessing subdevices in md

Each md personality keeps a list of devices that are currently active in the
array (mdk_rdev_t).  As these can potentially be removed at any time, some
locking is needed when accessing entries in the list.  Currently this involves
a spinlock, but all the spinlocking this imposed in unplug_slaves bothered me.

So, I have changed it to use rcu locking.  This is more appropriate as objects
are removed only very rarely, and there is no cost in waiting a little while
for a remove to succeed.

Also, all changes to the list of devices are performed by the per-array thread
(calling md_check_recovery) and so are completely single threaded, so no
locking between writers is needed at all.

Finally, devices are never added or removed while resync is happening, so
resync doesn't need to worry about locking at all.

So with this patch, the spinlocking is replaced with rcu read locking and
synchronize_kernel.  The rcu_read_lock is held while dereferencing a possible
device, and the nr_pending count is atomically incremented if the device is to
be held outside of the rcu_read_lock.

When removing a device, if the nr_pending count appears to be zero, we set the
list entry to NULL and call synchronize_kernel().  If the count is still zero
after this, we have a safe removal.  If it is non-zero, then someone has just
started using it so we put the pointer back and fail the removal.  When the
new user finally drops it's reference, that will cause the per-array thread to
wake up again and retry the removal.
Signed-off-by: Neil Brown <neilb@cse.unsw.edu.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

drivers/md/multipath.c

@@ -61,16 +61,16 @@ static int multipath_map (multipath_conf_t *conf)
 	 * now we use the first available disk.
 	 */
 
-	spin_lock_irq(&conf->device_lock);
+	rcu_read_lock();
 	for (i = 0; i < disks; i++) {
 		mdk_rdev_t *rdev = conf->multipaths[i].rdev;
 		if (rdev && rdev->in_sync) {
 			atomic_inc(&rdev->nr_pending);
-			spin_unlock_irq(&conf->device_lock);
+			rcu_read_unlock();
 			return i;
 		}
 	}
-	spin_unlock_irq(&conf->device_lock);
+	rcu_read_unlock();
 
 	printk(KERN_ERR "multipath_map(): no more operational IO paths?\n");
 	return (-1);
@@ -135,26 +135,26 @@ static void unplug_slaves(mddev_t *mddev)
 {
 	multipath_conf_t *conf = mddev_to_conf(mddev);
 	int i;
-	unsigned long flags;
 
-	spin_lock_irqsave(&conf->device_lock, flags);
+	rcu_read_lock();
 	for (i=0; i<mddev->raid_disks; i++) {
 		mdk_rdev_t *rdev = conf->multipaths[i].rdev;
 		if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
 			request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 
 			atomic_inc(&rdev->nr_pending);
-			spin_unlock_irqrestore(&conf->device_lock, flags);
+			rcu_read_unlock();
 
 			if (r_queue->unplug_fn)
 				r_queue->unplug_fn(r_queue);
 
-			spin_lock_irqsave(&conf->device_lock, flags);
 			rdev_dec_pending(rdev, mddev);
+			rcu_read_lock();
 		}
 	}
-	spin_unlock_irqrestore(&conf->device_lock, flags);
+	rcu_read_unlock();
 }
+
 static void multipath_unplug(request_queue_t *q)
 {
 	unplug_slaves(q->queuedata);
@@ -219,6 +219,7 @@ static int multipath_issue_flush(request_queue_t *q, struct gendisk *disk,
 	multipath_conf_t *conf = mddev_to_conf(mddev);
 	int i, ret = 0;
 
+	rcu_read_lock();
 	for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 		mdk_rdev_t *rdev = conf->multipaths[i].rdev;
 		if (rdev && !rdev->faulty) {
@@ -227,11 +228,17 @@ static int multipath_issue_flush(request_queue_t *q, struct gendisk *disk,
 
 			if (!r_queue->issue_flush_fn)
 				ret = -EOPNOTSUPP;
-			else
+			else {
+				atomic_inc(&rdev->nr_pending);
+				rcu_read_unlock();
 				ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 							      error_sector);
+				rdev_dec_pending(rdev, mddev);
+				rcu_read_lock();
 			}
 		}
+	}
+	rcu_read_unlock();
 	return ret;
 }
 
@@ -302,10 +309,9 @@ static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 	struct multipath_info *p;
 
 	print_multipath_conf(conf);
-	spin_lock_irq(&conf->device_lock);
+
 	for (path=0; path<mddev->raid_disks; path++) 
 		if ((p=conf->multipaths+path)->rdev == NULL) {
-			p->rdev = rdev;
 			blk_queue_stack_limits(mddev->queue,
 					       rdev->bdev->bd_disk->queue);
 
@@ -322,9 +328,9 @@ static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 			conf->working_disks++;
 			rdev->raid_disk = path;
 			rdev->in_sync = 1;
+			p->rdev = rdev;
 			found = 1;
 		}
-	spin_unlock_irq(&conf->device_lock);
 
 	print_multipath_conf(conf);
 	return found;
@@ -333,26 +339,29 @@ static int multipath_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 static int multipath_remove_disk(mddev_t *mddev, int number)
 {
 	multipath_conf_t *conf = mddev->private;
-	int err = 1;
+	int err = 0;
+	mdk_rdev_t *rdev;
 	struct multipath_info *p = conf->multipaths + number;
 
 	print_multipath_conf(conf);
-	spin_lock_irq(&conf->device_lock);
 
-	if (p->rdev) {
-		if (p->rdev->in_sync ||
-		    atomic_read(&p->rdev->nr_pending)) {
+	rdev = p->rdev;
+	if (rdev) {
+		if (rdev->in_sync ||
+		    atomic_read(&rdev->nr_pending)) {
 			printk(KERN_ERR "hot-remove-disk, slot %d is identified"				" but is still operational!\n", number);
 			err = -EBUSY;
 			goto abort;
 		}
 		p->rdev = NULL;
-		err = 0;
+		synchronize_kernel();
+		if (atomic_read(&rdev->nr_pending)) {
+			/* lost the race, try later */
+			err = -EBUSY;
+			p->rdev = rdev;
+		}
 	}
-	if (err)
-		MD_BUG();
 abort:
-	spin_unlock_irq(&conf->device_lock);
 
 	print_multipath_conf(conf);
 	return err;

drivers/md/raid1.c

@@ -339,7 +339,7 @@ static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 	const int sectors = r1_bio->sectors;
 	sector_t new_distance, current_distance;
 
-	spin_lock_irq(&conf->device_lock);
+	rcu_read_lock();
 	/*
 	 * Check if it if we can balance. We can balance on the whole
 	 * device if no resync is going on, or below the resync window.
@@ -416,7 +416,7 @@ static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 		conf->last_used = new_disk;
 		atomic_inc(&conf->mirrors[new_disk].rdev->nr_pending);
 	}
-	spin_unlock_irq(&conf->device_lock);
+	rcu_read_unlock();
 
 	return new_disk;
 }
@@ -425,26 +425,26 @@ static void unplug_slaves(mddev_t *mddev)
 {
 	conf_t *conf = mddev_to_conf(mddev);
 	int i;
-	unsigned long flags;
 
-	spin_lock_irqsave(&conf->device_lock, flags);
+	rcu_read_lock();
 	for (i=0; i<mddev->raid_disks; i++) {
 		mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 		if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
 			request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 
 			atomic_inc(&rdev->nr_pending);
-			spin_unlock_irqrestore(&conf->device_lock, flags);
+			rcu_read_unlock();
 
 			if (r_queue->unplug_fn)
 				r_queue->unplug_fn(r_queue);
 
-			spin_lock_irqsave(&conf->device_lock, flags);
 			rdev_dec_pending(rdev, mddev);
+			rcu_read_lock();
 		}
 	}
-	spin_unlock_irqrestore(&conf->device_lock, flags);
+	rcu_read_unlock();
 }
+
 static void raid1_unplug(request_queue_t *q)
 {
 	unplug_slaves(q->queuedata);
@@ -455,10 +455,9 @@ static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
 {
 	mddev_t *mddev = q->queuedata;
 	conf_t *conf = mddev_to_conf(mddev);
-	unsigned long flags;
 	int i, ret = 0;
 
-	spin_lock_irqsave(&conf->device_lock, flags);
+	rcu_read_lock();
 	for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 		mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 		if (rdev && !rdev->faulty) {
@@ -467,12 +466,17 @@ static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
 
 			if (!r_queue->issue_flush_fn)
 				ret = -EOPNOTSUPP;
-			else
+			else {
+				atomic_inc(&rdev->nr_pending);
+				rcu_read_unlock();
 				ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 							      error_sector);
+				rdev_dec_pending(rdev, mddev);
+				rcu_read_lock();
 			}
 		}
-	spin_unlock_irqrestore(&conf->device_lock, flags);
+	}
+	rcu_read_unlock();
 	return ret;
 }
 
@@ -579,7 +583,7 @@ static int make_request(request_queue_t *q, struct bio * bio)
 	 * bios[x] to bio
 	 */
 	disks = conf->raid_disks;
-	spin_lock_irq(&conf->device_lock);
+	rcu_read_lock();
 	for (i = 0;  i < disks; i++) {
 		if (conf->mirrors[i].rdev &&
 		    !conf->mirrors[i].rdev->faulty) {
@@ -588,7 +592,7 @@ static int make_request(request_queue_t *q, struct bio * bio)
 		} else
 			r1_bio->bios[i] = NULL;
 	}
-	spin_unlock_irq(&conf->device_lock);
+	rcu_read_unlock();
 
 	atomic_set(&r1_bio->remaining, 1);
 	md_write_start(mddev);
@@ -710,7 +714,6 @@ static int raid1_spare_active(mddev_t *mddev)
 	conf_t *conf = mddev->private;
 	mirror_info_t *tmp;
 
-	spin_lock_irq(&conf->device_lock);
 	/*
 	 * Find all failed disks within the RAID1 configuration 
 	 * and mark them readable
@@ -725,7 +728,6 @@ static int raid1_spare_active(mddev_t *mddev)
 			tmp->rdev->in_sync = 1;
 		}
 	}
-	spin_unlock_irq(&conf->device_lock);
 
 	print_conf(conf);
 	return 0;
@@ -739,10 +741,8 @@ static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 	int mirror;
 	mirror_info_t *p;
 
-	spin_lock_irq(&conf->device_lock);
 	for (mirror=0; mirror < mddev->raid_disks; mirror++)
 		if ( !(p=conf->mirrors+mirror)->rdev) {
-			p->rdev = rdev;
 
 			blk_queue_stack_limits(mddev->queue,
 					       rdev->bdev->bd_disk->queue);
@@ -757,9 +757,9 @@ static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 			p->head_position = 0;
 			rdev->raid_disk = mirror;
 			found = 1;
+			p->rdev = rdev;
 			break;
 		}
-	spin_unlock_irq(&conf->device_lock);
 
 	print_conf(conf);
 	return found;
@@ -768,24 +768,27 @@ static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 static int raid1_remove_disk(mddev_t *mddev, int number)
 {
 	conf_t *conf = mddev->private;
-	int err = 1;
+	int err = 0;
+	mdk_rdev_t *rdev;
 	mirror_info_t *p = conf->mirrors+ number;
 
 	print_conf(conf);
-	spin_lock_irq(&conf->device_lock);
-	if (p->rdev) {
-		if (p->rdev->in_sync ||
-		    atomic_read(&p->rdev->nr_pending)) {
+	rdev = p->rdev;
+	if (rdev) {
+		if (rdev->in_sync ||
+		    atomic_read(&rdev->nr_pending)) {
 			err = -EBUSY;
 			goto abort;
 		}
 		p->rdev = NULL;
-		err = 0;
+		synchronize_kernel();
+		if (atomic_read(&rdev->nr_pending)) {
+			/* lost the race, try later */
+			err = -EBUSY;
+			p->rdev = rdev;
+		}
 	}
-	if (err)
-		MD_BUG();
 abort:
-	spin_unlock_irq(&conf->device_lock);
 
 	print_conf(conf);
 	return err;
@@ -1022,7 +1025,7 @@ static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
 	 */
 	disk = conf->last_used;
 	/* make sure disk is operational */
-	spin_lock_irq(&conf->device_lock);
+
 	while (conf->mirrors[disk].rdev == NULL ||
 	       !conf->mirrors[disk].rdev->in_sync) {
 		if (disk <= 0)
@@ -1033,7 +1036,7 @@ static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
 	}
 	conf->last_used = disk;
 	atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
-	spin_unlock_irq(&conf->device_lock);
+
 
 	mirror = conf->mirrors + disk;
 

drivers/md/raid10.c

@@ -498,7 +498,7 @@ static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 	sector_t new_distance, current_distance;
 
 	raid10_find_phys(conf, r10_bio);
-	spin_lock_irq(&conf->device_lock);
+	rcu_read_lock();
 	/*
 	 * Check if we can balance. We can balance on the whole
 	 * device if no resync is going on, or below the resync window.
@@ -570,7 +570,7 @@ static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 
 	if (disk >= 0 && conf->mirrors[disk].rdev)
 		atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
-	spin_unlock_irq(&conf->device_lock);
+	rcu_read_unlock();
 
 	return disk;
 }
@@ -579,26 +579,26 @@ static void unplug_slaves(mddev_t *mddev)
 {
 	conf_t *conf = mddev_to_conf(mddev);
 	int i;
-	unsigned long flags;
 
-	spin_lock_irqsave(&conf->device_lock, flags);
+	rcu_read_lock();
 	for (i=0; i<mddev->raid_disks; i++) {
 		mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 		if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
 			request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 
 			atomic_inc(&rdev->nr_pending);
-			spin_unlock_irqrestore(&conf->device_lock, flags);
+			rcu_read_unlock();
 
 			if (r_queue->unplug_fn)
 				r_queue->unplug_fn(r_queue);
 
-			spin_lock_irqsave(&conf->device_lock, flags);
 			rdev_dec_pending(rdev, mddev);
+			rcu_read_lock();
 		}
 	}
-	spin_unlock_irqrestore(&conf->device_lock, flags);
+	rcu_read_unlock();
 }
+
 static void raid10_unplug(request_queue_t *q)
 {
 	unplug_slaves(q->queuedata);
@@ -609,10 +609,9 @@ static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
 {
 	mddev_t *mddev = q->queuedata;
 	conf_t *conf = mddev_to_conf(mddev);
-	unsigned long flags;
 	int i, ret = 0;
 
-	spin_lock_irqsave(&conf->device_lock, flags);
+	rcu_read_lock();
 	for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 		mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 		if (rdev && !rdev->faulty) {
@@ -621,12 +620,17 @@ static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
 
 			if (!r_queue->issue_flush_fn)
 				ret = -EOPNOTSUPP;
-			else
+			else {
+				atomic_inc(&rdev->nr_pending);
+				rcu_read_unlock();
 				ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 							      error_sector);
+				rdev_dec_pending(rdev, mddev);
+				rcu_read_lock();
 			}
 		}
-	spin_unlock_irqrestore(&conf->device_lock, flags);
+	}
+	rcu_read_unlock();
 	return ret;
 }
 
@@ -757,7 +761,7 @@ static int make_request(request_queue_t *q, struct bio * bio)
 	 * bios[x] to bio
 	 */
 	raid10_find_phys(conf, r10_bio);
-	spin_lock_irq(&conf->device_lock);
+	rcu_read_lock();
 	for (i = 0;  i < conf->copies; i++) {
 		int d = r10_bio->devs[i].devnum;
 		if (conf->mirrors[d].rdev &&
@@ -767,7 +771,7 @@ static int make_request(request_queue_t *q, struct bio * bio)
 		} else
 			r10_bio->devs[i].bio = NULL;
 	}
-	spin_unlock_irq(&conf->device_lock);
+	rcu_read_unlock();
 
 	atomic_set(&r10_bio->remaining, 1);
 	md_write_start(mddev);
@@ -900,7 +904,6 @@ static int raid10_spare_active(mddev_t *mddev)
 	conf_t *conf = mddev->private;
 	mirror_info_t *tmp;
 
-	spin_lock_irq(&conf->device_lock);
 	/*
 	 * Find all non-in_sync disks within the RAID10 configuration
 	 * and mark them in_sync
@@ -915,7 +918,6 @@ static int raid10_spare_active(mddev_t *mddev)
 			tmp->rdev->in_sync = 1;
 		}
 	}
-	spin_unlock_irq(&conf->device_lock);
 
 	print_conf(conf);
 	return 0;
@@ -934,10 +936,9 @@ static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 		 * very different from resync
 		 */
 		return 0;
-	spin_lock_irq(&conf->device_lock);
+
 	for (mirror=0; mirror < mddev->raid_disks; mirror++)
 		if ( !(p=conf->mirrors+mirror)->rdev) {
-			p->rdev = rdev;
 
 			blk_queue_stack_limits(mddev->queue,
 					       rdev->bdev->bd_disk->queue);
@@ -952,9 +953,9 @@ static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 			p->head_position = 0;
 			rdev->raid_disk = mirror;
 			found = 1;
+			p->rdev = rdev;
 			break;
 		}
-	spin_unlock_irq(&conf->device_lock);
 
 	print_conf(conf);
 	return found;
@@ -963,24 +964,27 @@ static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 static int raid10_remove_disk(mddev_t *mddev, int number)
 {
 	conf_t *conf = mddev->private;
-	int err = 1;
+	int err = 0;
+	mdk_rdev_t *rdev;
 	mirror_info_t *p = conf->mirrors+ number;
 
 	print_conf(conf);
-	spin_lock_irq(&conf->device_lock);
-	if (p->rdev) {
-		if (p->rdev->in_sync ||
-		    atomic_read(&p->rdev->nr_pending)) {
+	rdev = p->rdev;
+	if (rdev) {
+		if (rdev->in_sync ||
+		    atomic_read(&rdev->nr_pending)) {
 			err = -EBUSY;
 			goto abort;
 		}
 		p->rdev = NULL;
-		err = 0;
+		synchronize_kernel();
+		if (atomic_read(&rdev->nr_pending)) {
+			/* lost the race, try later */
+			err = -EBUSY;
+			p->rdev = rdev;
+		}
 	}
-	if (err)
-		MD_BUG();
 abort:
-	spin_unlock_irq(&conf->device_lock);
 
 	print_conf(conf);
 	return err;
@@ -1468,7 +1472,7 @@ static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
 		set_bit(R10BIO_IsSync, &r10_bio->state);
 		raid10_find_phys(conf, r10_bio);
 		r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
-		spin_lock_irq(&conf->device_lock);
+
 		for (i=0; i<conf->copies; i++) {
 			int d = r10_bio->devs[i].devnum;
 			bio = r10_bio->devs[i].bio;
@@ -1488,7 +1492,7 @@ static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
 			bio->bi_bdev = conf->mirrors[d].rdev->bdev;
 			count++;
 		}
-		spin_unlock_irq(&conf->device_lock);
+
 		if (count < 2) {
 			for (i=0; i<conf->copies; i++) {
 				int d = r10_bio->devs[i].devnum;

drivers/md/raid5.c

@@ -1247,13 +1247,13 @@ static void handle_stripe(struct stripe_head *sh)
 		else
 			bi->bi_end_io = raid5_end_read_request;
  
-		spin_lock_irq(&conf->device_lock);
+		rcu_read_lock();
 		rdev = conf->disks[i].rdev;
 		if (rdev && rdev->faulty)
 			rdev = NULL;
 		if (rdev)
 			atomic_inc(&rdev->nr_pending);
-		spin_unlock_irq(&conf->device_lock);
+		rcu_read_unlock();
  
 		if (rdev) {
 			if (test_bit(R5_Syncio, &sh->dev[i].flags))
@@ -1302,25 +1302,24 @@ static void unplug_slaves(mddev_t *mddev)
 {
 	raid5_conf_t *conf = mddev_to_conf(mddev);
 	int i;
-	unsigned long flags;
 
-	spin_lock_irqsave(&conf->device_lock, flags);
+	rcu_read_lock();
 	for (i=0; i<mddev->raid_disks; i++) {
 		mdk_rdev_t *rdev = conf->disks[i].rdev;
 		if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
 			request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 
 			atomic_inc(&rdev->nr_pending);
-			spin_unlock_irqrestore(&conf->device_lock, flags);
+			rcu_read_unlock();
 
 			if (r_queue->unplug_fn)
 				r_queue->unplug_fn(r_queue);
 
-			spin_lock_irqsave(&conf->device_lock, flags);
 			rdev_dec_pending(rdev, mddev);
+			rcu_read_lock();
 		}
 	}
-	spin_unlock_irqrestore(&conf->device_lock, flags);
+	rcu_read_unlock();
 }
 
 static void raid5_unplug_device(request_queue_t *q)
@@ -1347,6 +1346,7 @@ static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
 	raid5_conf_t *conf = mddev_to_conf(mddev);
 	int i, ret = 0;
 
+	rcu_read_lock();
 	for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 		mdk_rdev_t *rdev = conf->disks[i].rdev;
 		if (rdev && !rdev->faulty) {
@@ -1355,11 +1355,17 @@ static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
 
 			if (!r_queue->issue_flush_fn)
 				ret = -EOPNOTSUPP;
-			else
+			else {
+				atomic_inc(&rdev->nr_pending);
+				rcu_read_unlock();
 				ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 							      error_sector);
+				rdev_dec_pending(rdev, mddev);
+				rcu_read_lock();
 			}
 		}
+	}
+	rcu_read_unlock();
 	return ret;
 }
 
@@ -1789,7 +1795,6 @@ static int raid5_spare_active(mddev_t *mddev)
 	raid5_conf_t *conf = mddev->private;
 	struct disk_info *tmp;
 
-	spin_lock_irq(&conf->device_lock);
 	for (i = 0; i < conf->raid_disks; i++) {
 		tmp = conf->disks + i;
 		if (tmp->rdev
@@ -1801,7 +1806,6 @@ static int raid5_spare_active(mddev_t *mddev)
 			tmp->rdev->in_sync = 1;
 		}
 	}
-	spin_unlock_irq(&conf->device_lock);
 	print_raid5_conf(conf);
 	return 0;
 }
@@ -1809,25 +1813,28 @@ static int raid5_spare_active(mddev_t *mddev)
 static int raid5_remove_disk(mddev_t *mddev, int number)
 {
 	raid5_conf_t *conf = mddev->private;
-	int err = 1;
+	int err = 0;
+	mdk_rdev_t *rdev;
 	struct disk_info *p = conf->disks + number;
 
 	print_raid5_conf(conf);
-	spin_lock_irq(&conf->device_lock);
-
-	if (p->rdev) {
-		if (p->rdev->in_sync || 
-		    atomic_read(&p->rdev->nr_pending)) {
+	rdev = p->rdev;
+	if (rdev) {
+		if (rdev->in_sync ||
+		    atomic_read(&rdev->nr_pending)) {
 			err = -EBUSY;
 			goto abort;
 		}
 		p->rdev = NULL;
-		err = 0;
+		synchronize_kernel();
+		if (atomic_read(&rdev->nr_pending)) {
+			/* lost the race, try later */
+			err = -EBUSY;
+			p->rdev = rdev;
+		}
 	}
-	if (err)
-		MD_BUG();
 abort:
-	spin_unlock_irq(&conf->device_lock);
+
 	print_raid5_conf(conf);
 	return err;
 }
@@ -1839,19 +1846,17 @@ static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 	int disk;
 	struct disk_info *p;
 
-	spin_lock_irq(&conf->device_lock);
 	/*
 	 * find the disk ...
 	 */
 	for (disk=0; disk < mddev->raid_disks; disk++)
 		if ((p=conf->disks + disk)->rdev == NULL) {
-			p->rdev = rdev;
 			rdev->in_sync = 0;
 			rdev->raid_disk = disk;
 			found = 1;
+			p->rdev = rdev;
 			break;
 		}
-	spin_unlock_irq(&conf->device_lock);
 	print_raid5_conf(conf);
 	return found;
 }

drivers/md/raid6main.c

@@ -1409,13 +1409,13 @@ static void handle_stripe(struct stripe_head *sh)
 		else
 			bi->bi_end_io = raid6_end_read_request;
 
-		spin_lock_irq(&conf->device_lock);
+		rcu_read_lock();
 		rdev = conf->disks[i].rdev;
 		if (rdev && rdev->faulty)
 			rdev = NULL;
 		if (rdev)
 			atomic_inc(&rdev->nr_pending);
-		spin_unlock_irq(&conf->device_lock);
+		rcu_read_unlock();
 
 		if (rdev) {
 			if (test_bit(R5_Syncio, &sh->dev[i].flags))
@@ -1464,25 +1464,24 @@ static void unplug_slaves(mddev_t *mddev)
 {
 	raid6_conf_t *conf = mddev_to_conf(mddev);
 	int i;
-	unsigned long flags;
 
-	spin_lock_irqsave(&conf->device_lock, flags);
+	rcu_read_lock();
 	for (i=0; i<mddev->raid_disks; i++) {
 		mdk_rdev_t *rdev = conf->disks[i].rdev;
 		if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
 			request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 
 			atomic_inc(&rdev->nr_pending);
-			spin_unlock_irqrestore(&conf->device_lock, flags);
+			rcu_read_unlock();
 
 			if (r_queue->unplug_fn)
 				r_queue->unplug_fn(r_queue);
 
-			spin_lock_irqsave(&conf->device_lock, flags);
 			rdev_dec_pending(rdev, mddev);
+			rcu_read_lock();
 		}
 	}
-	spin_unlock_irqrestore(&conf->device_lock, flags);
+	rcu_read_unlock();
 }
 
 static void raid6_unplug_device(request_queue_t *q)
@@ -1509,6 +1508,7 @@ static int raid6_issue_flush(request_queue_t *q, struct gendisk *disk,
 	raid6_conf_t *conf = mddev_to_conf(mddev);
 	int i, ret = 0;
 
+	rcu_read_lock();
 	for (i=0; i<mddev->raid_disks && ret == 0; i++) {
 		mdk_rdev_t *rdev = conf->disks[i].rdev;
 		if (rdev && !rdev->faulty) {
@@ -1517,11 +1517,17 @@ static int raid6_issue_flush(request_queue_t *q, struct gendisk *disk,
 
 			if (!r_queue->issue_flush_fn)
 				ret = -EOPNOTSUPP;
-			else
+			else {
+				atomic_inc(&rdev->nr_pending);
+				rcu_read_unlock();
 				ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
 							      error_sector);
+				rdev_dec_pending(rdev, mddev);
+				rcu_read_lock();
 			}
 		}
+	}
+	rcu_read_unlock();
 	return ret;
 }
 
@@ -1958,7 +1964,6 @@ static int raid6_spare_active(mddev_t *mddev)
 	raid6_conf_t *conf = mddev->private;
 	struct disk_info *tmp;
 
-	spin_lock_irq(&conf->device_lock);
 	for (i = 0; i < conf->raid_disks; i++) {
 		tmp = conf->disks + i;
 		if (tmp->rdev
@@ -1970,7 +1975,6 @@ static int raid6_spare_active(mddev_t *mddev)
 			tmp->rdev->in_sync = 1;
 		}
 	}
-	spin_unlock_irq(&conf->device_lock);
 	print_raid6_conf(conf);
 	return 0;
 }
@@ -1978,25 +1982,29 @@ static int raid6_spare_active(mddev_t *mddev)
 static int raid6_remove_disk(mddev_t *mddev, int number)
 {
 	raid6_conf_t *conf = mddev->private;
-	int err = 1;
+	int err = 0;
+	mdk_rdev_t *rdev;
 	struct disk_info *p = conf->disks + number;
 
 	print_raid6_conf(conf);
-	spin_lock_irq(&conf->device_lock);
-
-	if (p->rdev) {
-		if (p->rdev->in_sync ||
-		    atomic_read(&p->rdev->nr_pending)) {
+	rdev = p->rdev;
+	if (rdev) {
+		if (rdev->in_sync ||
+		    atomic_read(&rdev->nr_pending)) {
 			err = -EBUSY;
 			goto abort;
 		}
 		p->rdev = NULL;
-		err = 0;
+		synchronize_kernel();
+		if (atomic_read(&rdev->nr_pending)) {
+			/* lost the race, try later */
+			err = -EBUSY;
+			p->rdev = rdev;
+		}
 	}
-	if (err)
-		MD_BUG();
+
 abort:
-	spin_unlock_irq(&conf->device_lock);
+
 	print_raid6_conf(conf);
 	return err;
 }
@@ -2008,19 +2016,17 @@ static int raid6_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 	int disk;
 	struct disk_info *p;
 
-	spin_lock_irq(&conf->device_lock);
 	/*
 	 * find the disk ...
 	 */
 	for (disk=0; disk < mddev->raid_disks; disk++)
 		if ((p=conf->disks + disk)->rdev == NULL) {
-			p->rdev = rdev;
 			rdev->in_sync = 0;
 			rdev->raid_disk = disk;
 			found = 1;
+			p->rdev = rdev;
 			break;
 		}
-	spin_unlock_irq(&conf->device_lock);
 	print_raid6_conf(conf);
 	return found;
 }