Skip to contentSkip to conversation

new bd_mutex lockdep annotation

PATCH In review Peter Zijlstra 3 participants 7 messages Sep 13, 2006 – Sep 29, 2006

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover new bd_mutex lockdep annotation

--

Andrew thought the old bd_mutex lockdep annotations were too complex, Arjan too said something like this several weeks ago and suggested I do something with lockdep_set_class() instead of mutex_lock_nested().

#2 lore raw

PATCH 2/2 new bd_mutex lockdep annotation

Use the gendisk partition number to set a lock class.

Cc Neil Brown <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Arjan van de Ven <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Jason Baron <[email protected]>
---
 fs/block_dev.c |    9 +++++++++
 1 file changed, 9 insertions(+)
fs/block_dev.c +9 -0
Index: linux-2.6-mm/fs/block_dev.c
===================================================================
@@ -357,10 +357,14 @@ static int bdev_set(struct inode *inode,
 
 static LIST_HEAD(all_bdevs);
 
+static struct lock_class_key bdev_part_lock_key;
+
 struct block_device *bdget(dev_t dev)
 {
 	struct block_device *bdev;
 	struct inode *inode;
+	struct gendisk *disk;
+	int part = 0;
 
 	inode = iget5_locked(bd_mnt->mnt_sb, hash(dev),
 			bdev_test, bdev_set, &dev);
@@ -386,6 +390,11 @@ struct block_device *bdget(dev_t dev)
 		list_add(&bdev->bd_list, &all_bdevs);
 		spin_unlock(&bdev_lock);
 		unlock_new_inode(inode);
+		mutex_init(&bdev->bd_mutex);
+		disk = get_gendisk(dev, &part);
+		if (part)
+			lockdep_set_class(&bdev->bd_mutex, &bdev_part_lock_key);
+		put_disk(disk);
 	}
 	return bdev;
 }

--

#1 lore raw

PATCH 1/2 remove the old bd_mutex lockdep annotation

In reply to Peter Zijlstra

Remove the old complex and crufty bd_mutex annotation.

Cc Neil Brown <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Arjan van de Ven <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Jason Baron <[email protected]>
---
 block/ioctl.c      |    4 -
 drivers/md/md.c    |    6 -
 fs/block_dev.c     |  180 ++++++++++++++++-------------------------------------
 include/linux/fs.h |   17 -----
 4 files changed, 60 insertions(+), 147 deletions(-)
drivers/md/md.c +3 -3
Index: linux-2.6-mm/drivers/md/md.c
===================================================================
@@ -1408,7 +1408,7 @@ static int lock_rdev(mdk_rdev_t *rdev, d
 	struct block_device *bdev;
 	char b[BDEVNAME_SIZE];
 
-	bdev = open_partition_by_devnum(dev, FMODE_READ|FMODE_WRITE);
+	bdev = open_by_devnum(dev, FMODE_READ|FMODE_WRITE);
 	if (IS_ERR(bdev)) {
 		printk(KERN_ERR "md: could not open %s.\n",
 			__bdevname(dev, b));
@@ -1418,7 +1418,7 @@ static int lock_rdev(mdk_rdev_t *rdev, d
 	if (err) {
 		printk(KERN_ERR "md: could not bd_claim %s.\n",
 			bdevname(bdev, b));
-		blkdev_put_partition(bdev);
+		blkdev_put(bdev);
 		return err;
 	}
 	rdev->bdev = bdev;
@@ -1432,7 +1432,7 @@ static void unlock_rdev(mdk_rdev_t *rdev
 	if (!bdev)
 		MD_BUG();
 	bd_release(bdev);
-	blkdev_put_partition(bdev);
+	blkdev_put(bdev);
 }
 
 void md_autodetect_dev(dev_t dev);
fs/block_dev.c +55 -125

253 lines not on this page · raw message

include/linux/fs.h +0 -17
Index: linux-2.6-mm/include/linux/fs.h
===================================================================
@@ -504,21 +504,6 @@ struct block_device {
 };
 
 /*
- * bdev->bd_mutex nesting subclasses for the lock validator:
- *
- * 0: normal
- * 1: 'whole'
- * 2: 'partition'
- */
-enum bdev_bd_mutex_lock_class
-{
-	BD_MUTEX_NORMAL,
-	BD_MUTEX_WHOLE,
-	BD_MUTEX_PARTITION
-};
-
-
-/*
  * Radix-tree tags, for tagging dirty and writeback pages within the pagecache
  * radix trees
  */
@@ -1596,7 +1581,6 @@ extern void bd_set_size(struct block_dev
 extern void bd_forget(struct inode *inode);
 extern void bdput(struct block_device *);
 extern struct block_device *open_by_devnum(dev_t, unsigned);
-extern struct block_device *open_partition_by_devnum(dev_t, unsigned);
 extern const struct address_space_operations def_blk_aops;
 #else
 static inline void bd_forget(struct inode *inode) {}
@@ -1614,7 +1598,6 @@ extern int blkdev_driver_ioctl(struct in
 extern long compat_blkdev_ioctl(struct file *, unsigned, unsigned long);
 extern int blkdev_get(struct block_device *, mode_t, unsigned);
 extern int blkdev_put(struct block_device *);
-extern int blkdev_put_partition(struct block_device *);
 extern int bd_claim(struct block_device *, void *);
 extern void bd_release(struct block_device *);
 #ifdef CONFIG_SYSFS
block/ioctl.c +2 -2
Index: linux-2.6-mm/block/ioctl.c
===================================================================
@@ -72,7 +72,7 @@ static int blkpg_ioctl(struct block_devi
 			bdevp = bdget_disk(disk, part);
 			if (!bdevp)
 				return -ENOMEM;
-			mutex_lock_nested(&bdevp->bd_mutex, BD_MUTEX_PARTITION);
+			mutex_lock(&bdevp->bd_mutex);
 			if (bdevp->bd_openers) {
 				mutex_unlock(&bdevp->bd_mutex);
 				bdput(bdevp);
@@ -82,7 +82,7 @@ static int blkpg_ioctl(struct block_devi
 			fsync_bdev(bdevp);
 			invalidate_bdev(bdevp, 0);
 
-			mutex_lock_nested(&bdev->bd_mutex, BD_MUTEX_WHOLE);
+			mutex_lock(&bdev->bd_mutex);
 			delete_partition(disk, part);
 			mutex_unlock(&bdev->bd_mutex);
 			mutex_unlock(&bdevp->bd_mutex);

--

#4 lore raw
In reply to Peter Zijlstra
Quoted 1 line · On Wednesday September 13, [email protected] wrote:

Use the gendisk partition number to set a lock class.

Yes, this does look a lot nicer, thanks.

Two observations.
1/ I was confused that you added a call to mutex_init. One would
normally expect to only have one of these for any given mutex, so adding one was a surprise.
I now realise that the purpose of this call is not exactly to init the mutex, but to init the lockdep class in case this inode was previously used for a partition but is now being used for a whole device. This makes sense, but renders the mutex_init in init_once pointless. Maybe that should be removed?

2/ You are introducing a new call to get_gendisk.
This bothers me for two reasons. Both relate to a comparison with the call to get_gendisk in block_dev.c:do_open. a/ That call is protected by lock_kernel. Your call is not. b/ That call is followed by a test for '!disk' implying that it
can return NULL. Yours is not - at least not obviously (put_disk does have the check).

I'm not sure if these are actually problems, but the do bother me.

Thinking through the possibly reasons for the lock_kernel, I wonder it the current device number mapping scheme actually allows you to determine if something is partitioned or not in a static sense. Maybe that is only guaranteed to be stable while the device is open...
I wonder if Al Viro could put my mind at rest .... Al - do you have a moment to look at this? Thanks.

NeilBrown

Quoted 44 lines
Cc Neil Brown <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Arjan van de Ven <[email protected]>
Cc Andrew Morton <[email protected]>
Cc Jason Baron <[email protected]>
---
 fs/block_dev.c |    9 +++++++++
 1 file changed, 9 insertions(+)

Index: linux-2.6-mm/fs/block_dev.c
===================================================================

fs/block_dev.c +9 -1
@@ -357,10 +357,14 @@ static int bdev_set(struct inode *inode,
 
 static LIST_HEAD(all_bdevs);
 
+static struct lock_class_key bdev_part_lock_key;
+
 struct block_device *bdget(dev_t dev)
 {
 	struct block_device *bdev;
 	struct inode *inode;
+	struct gendisk *disk;
+	int part = 0;
 
 	inode = iget5_locked(bd_mnt->mnt_sb, hash(dev),
 			bdev_test, bdev_set, &dev);
@@ -386,6 +390,11 @@ struct block_device *bdget(dev_t dev)
 		list_add(&bdev->bd_list, &all_bdevs);
 		spin_unlock(&bdev_lock);
 		unlock_new_inode(inode);
+		mutex_init(&bdev->bd_mutex);
+		disk = get_gendisk(dev, &part);
+		if (part)
+			lockdep_set_class(&bdev->bd_mutex, &bdev_part_lock_key);
+		put_disk(disk);
 	}
 	return bdev;
 }
 
--
author Acked-by Arjan van de Ven
#5 lore raw
Quoted 5 lines · Neil Brown
Quoted 1 line · On Wednesday September 13, [email protected] wrote:

Use the gendisk partition number to set a lock class.

Yes, this does look a lot nicer, thanks.

Quoted 9 lines

Two observations.
1/ I was confused that you added a call to mutex_init. One would
normally expect to only have one of these for any given mutex, so adding one was a surprise.
I now realise that the purpose of this call is not exactly to init the mutex, but to init the lockdep class in case this inode was previously used for a partition but is now being used for a whole device. This makes sense, but renders the mutex_init in init_once pointless. Maybe that should be removed?

Yes, that would be quite redundant now, new patch attached.

Quoted 7 lines

2/ You are introducing a new call to get_gendisk.
This bothers me for two reasons. Both relate to a comparison with the call to get_gendisk in block_dev.c:do_open. a/ That call is protected by lock_kernel. Your call is not. b/ That call is followed by a test for '!disk' implying that it
can return NULL. Yours is not - at least not obviously (put_disk does have the check).

a) kobj_lookup() vs kobj_(un)map() use the domain lock.

Not all calls to blk_register_region() were under lock_kernel() afaicf.

So I don't think this is needed, but I'll gladly take advise otherwise, I'm not well versed with the kobj stuff.

b) from quick inspection yesterday I reached two (false) conclusions
- &part would not be changed when !disk
- disk would have to exists at the time we call bdget()
Now I can't seem to validate either of them. Added disk to the if statement just to be safe.

Quoted 7 lines

I'm not sure if these are actually problems, but the do bother me.

Thinking through the possibly reasons for the lock_kernel, I wonder it the current device number mapping scheme actually allows you to determine if something is partitioned or not in a static sense. Maybe that is only guaranteed to be stable while the device is open...

Hmm, yes I think I see what you mean...

Quoted 2 lines

I wonder if Al Viro could put my mind at rest .... Al - do you have a moment to look at this? Thanks.

+1

---

Use the gendisk partition number to set a lock class.

---
 fs/block_dev.c |   10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
fs/block_dev.c +9 -1
Index: linux-2.6-mm/fs/block_dev.c
===================================================================
@@ -264,7 +264,6 @@ static void init_once(void * foo, kmem_c
 	    SLAB_CTOR_CONSTRUCTOR)
 	{
 		memset(bdev, 0, sizeof(*bdev));
-		mutex_init(&bdev->bd_mutex);
 		mutex_init(&bdev->bd_mount_mutex);
 		INIT_LIST_HEAD(&bdev->bd_inodes);
 		INIT_LIST_HEAD(&bdev->bd_list);
@@ -357,10 +356,14 @@ static int bdev_set(struct inode *inode,
 
 static LIST_HEAD(all_bdevs);
 
+static struct lock_class_key bdev_part_lock_key;
+
 struct block_device *bdget(dev_t dev)
 {
 	struct block_device *bdev;
 	struct inode *inode;
+	struct gendisk *disk;
+	int part = 0;
 
 	inode = iget5_locked(bd_mnt->mnt_sb, hash(dev),
 			bdev_test, bdev_set, &dev);
@@ -386,6 +389,11 @@ struct block_device *bdget(dev_t dev)
 		list_add(&bdev->bd_list, &all_bdevs);
 		spin_unlock(&bdev_lock);
 		unlock_new_inode(inode);
+		mutex_init(&bdev->bd_mutex);
+		disk = get_gendisk(dev, &part);
+		if (disk && part)
+			lockdep_set_class(&bdev->bd_mutex, &bdev_part_lock_key);
+		put_disk(disk);
 	}
 	return bdev;
 }
#6 lore raw

Andrew, please hold on this, Al Viro doesn't agree with it.

Calling get_gendisk() from bdget() changes the bdget() semantics too much. For one it enables bdget() to load modules.

Al proposed the following approach. Neil can you agree with this too?

---
Avoid the nesting of bd_mutex by serializing the locks. This is made easier by changing the ->bd_part_count rules, its now only changed for the first openers/closers.

---

fs/block_dev.c +42 -25
index 92de28d..0ffc4f0 100644
@@ -898,12 +889,15 @@ void bd_set_size(struct block_device *bd
 }
 EXPORT_SYMBOL(bd_set_size);
 
+static int __blkdev_put(struct block_device *bdev, int part);
+
 static int do_open(struct block_device *bdev, struct file *file)
 {
 	struct module *owner = NULL;
 	struct gendisk *disk;
 	int ret = -ENXIO;
 	int part;
+	struct block_device *whole = NULL;
 
 	file->f_mapping = bdev->bd_inode->i_mapping;
 	lock_kernel();
@@ -937,30 +931,42 @@ static int do_open(struct block_device *
 				rescan_partitions(disk, bdev);
 		} else {
 			struct hd_struct *p;
-			struct block_device *whole;
+
+			mutex_unlock(&bdev->bd_mutex);
+
 			whole = bdget_disk(disk, 0);
 			ret = -ENOMEM;
 			if (!whole)
-				goto out_first;
+				goto out_first_lock;
 			ret = blkdev_get(whole, file->f_mode, file->f_flags);
 			if (ret)
-				goto out_first;
-			bdev->bd_contains = whole;
+				goto out_first_lock;
+
 			mutex_lock(&whole->bd_mutex);
 			whole->bd_part_count++;
 			p = disk->part[part - 1];
-			bdev->bd_inode->i_data.backing_dev_info =
-			   whole->bd_inode->i_data.backing_dev_info;
 			if (!(disk->flags & GENHD_FL_UP) || !p || !p->nr_sects) {
-				whole->bd_part_count--;
 				mutex_unlock(&whole->bd_mutex);
 				ret = -ENXIO;
-				goto out_first;
+				goto out_first_lock;
 			}
 			kobject_get(&p->kobj);
-			bdev->bd_part = p;
-			bd_set_size(bdev, (loff_t) p->nr_sects << 9);
 			mutex_unlock(&whole->bd_mutex);
+
+			mutex_lock(&bdev->bd_mutex);
+			if (bdev->bd_contains != whole) {
+				bdev->bd_contains = whole;
+				bdev->bd_inode->i_data.backing_dev_info =
+				  whole->bd_inode->i_data.backing_dev_info;
+				bdev->bd_part = p;
+				bd_set_size(bdev, (loff_t) p->nr_sects << 9);
+				whole = NULL;
+			} else {
+				mutex_unlock(&bdev->bd_mutex);
+				kobject_put(&p->kobj);
+				__blkdev_put(whole, 1);
+				mutex_lock(&bdev->bd_mutex);
+			}
 		}
 	} else {
 		put_disk(disk);
@@ -973,10 +979,6 @@ static int do_open(struct block_device *
 			}
 			if (bdev->bd_invalidated)
 				rescan_partitions(bdev->bd_disk, bdev);
-		} else {
-			mutex_lock(&bdev->bd_contains->bd_mutex);
-			bdev->bd_contains->bd_part_count++;
-			mutex_unlock(&bdev->bd_contains->bd_mutex);
 		}
 	}
 	bdev->bd_openers++;
@@ -984,11 +986,12 @@ static int do_open(struct block_device *
 	unlock_kernel();
 	return 0;
 
+out_first_lock:
+	if (whole)
+		__blkdev_put(whole, 1);
+	mutex_lock(&bdev->bd_mutex);
 out_first:
 	bdev->bd_disk = NULL;
-	bdev->bd_inode->i_data.backing_dev_info = &default_backing_dev_info;
-	if (bdev != bdev->bd_contains)
-		blkdev_put(bdev->bd_contains);
 	bdev->bd_contains = NULL;
 	put_disk(disk);
 	module_put(owner);
@@ -1049,14 +1052,17 @@ static int blkdev_open(struct inode * in
 	return res;
 }
 
-int blkdev_put(struct block_device *bdev)
+static int __blkdev_put(struct block_device *bdev, int part)
 {
 	int ret = 0;
 	struct inode *bd_inode = bdev->bd_inode;
 	struct gendisk *disk = bdev->bd_disk;
+	struct block_device *victim = NULL;
 
 	mutex_lock(&bdev->bd_mutex);
 	lock_kernel();
+	if (part)
+		bdev->bd_part_count--;
 	if (!--bdev->bd_openers) {
 		sync_blockdev(bdev);
 		kill_bdev(bdev);
@@ -1064,10 +1070,6 @@ int blkdev_put(struct block_device *bdev
 	if (bdev->bd_contains == bdev) {
 		if (disk->fops->release)
 			ret = disk->fops->release(bd_inode, NULL);
-	} else {
-		mutex_lock(&bdev->bd_contains->bd_mutex);
-		bdev->bd_contains->bd_part_count--;
-		mutex_unlock(&bdev->bd_contains->bd_mutex);
 	}
 	if (!bdev->bd_openers) {
 		struct module *owner = disk->fops->owner;
@@ -1081,17 +1083,23 @@ int blkdev_put(struct block_device *bdev
 		}
 		bdev->bd_disk = NULL;
 		bdev->bd_inode->i_data.backing_dev_info = &default_backing_dev_info;
-		if (bdev != bdev->bd_contains) {
-			blkdev_put(bdev->bd_contains);
-		}
+		if (bdev != bdev->bd_contains)
+			victim = bdev->bd_contains;
 		bdev->bd_contains = NULL;
 	}
 	unlock_kernel();
 	mutex_unlock(&bdev->bd_mutex);
+	if (victim)
+		__blkdev_put(victim, 1);
 	bdput(bdev);
 	return ret;
 }
 
+int blkdev_put(struct block_device *bdev)
+{
+	return __blkdev_put(bdev, 0);
+}
+
 EXPORT_SYMBOL(blkdev_put);
 
 static int blkdev_close(struct inode * inode, struct file * filp)