Skip to contentSkip to conversation

Reverting "bd_mount_mutex" to "bd_mount_sem"

RFC In review Srinivasa Ds 5 participants 6 messages Sep 27, 2006 – Oct 23, 2006

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

RFC 1/1 Reverting "bd_mount_mutex" to "bd_mount_sem"

Hi all
When I was executing "dmsetup resume <device-name>" command,I got the
error shown below. Which basically tells that, "bd_mount_mutex" in thaw_bdev() is not locked by "dmsetup resume" command and hence it is not allowing it to unlock also.
=========================================================
Badness in debug_mutex_unlock at kernel/mutex-debug.c:80

Call Trace:
[C0000000634DB260] [C000000000010948] .show_stack+0x68/0x1b0 (unreliable)
[C0000000634DB300] [C0000000003376C4] .program_check_exception+0x1cc/0x5b0
[C0000000634DB3D0] [C0000000000047EC] program_check_common+0xec/0x100
--- Exception: 700 at .debug_mutex_unlock+0x3c/0xc4
   LR = .debug_mutex_unlock+0x30/0xc4

[C0000000634DB6C0] [C0000000634DB750] 0xc0000000634db750 (unreliable) [C0000000634DB740] [C000000000335950] .__mutex_unlock_slowpath+0xd8/0x144 [C0000000634DB7E0] [C0000000000E2370] .thaw_bdev+0x9c/0xb8 [C0000000634DB870] [D000000000480830] .unlock_fs+0x34/0x70 [dm_mod] [C0000000634DB900] [D000000000481720] .dm_resume+0x110/0x1ac [dm_mod] [C0000000634DB9A0] [D000000000485C54] .dev_suspend+0x1b0/0x204 [dm_mod] [C0000000634DBA40] [D000000000486728] .ctl_ioctl+0x29c/0x318 [dm_mod] [C0000000634DBC30] [C0000000000F8310] .do_ioctl+0xbc/0xf0 [C0000000634DBCD0] [C0000000000F879C] .vfs_ioctl+0x458/0x498 [C0000000634DBD80] [C0000000000F8874] .sys_ioctl+0x98/0xe0 [C0000000634DBE30] [C00000000000871C] syscall_exit+0x0/0x40
======================================================================

On debugging I found out that,"dmsetup suspend <device name>" calls
"freeze_bdev()",which locks "bd_mount_mutex" to make sure that no new mounts happen on bdev until thaw_bdev() is called.
This "thaw_bdev()" is getting called when we resume the device through
"dmsetup resume <device-name>".
Hence we have 2 processes,one of which locks "bd_mount_mutex"(dmsetup
suspend) and Another(dmsetup resume) unlocks it.

Since this is not allowed in mutex,I reverted back to
bd_mount_sem(semaphore),It worked for me.

So need your comments for changing "bd_mount_mutex" to "bd_mount_sem".

This is the patch,which I have used.

Thanks
Srinivasa Ds
LTC-IBM
Bangalore

fs/block_dev.c +1 -1
diff -Naurp linux-2.6.18-rc6-orig/fs/block_dev.c linux-2.6.18-rc6-mod/fs/block_dev.c
@@ -261,7 +261,7 @@ static void init_once(void * foo, kmem_c
 	{
 		memset(bdev, 0, sizeof(*bdev));
 		mutex_init(&bdev->bd_mutex);
-		mutex_init(&bdev->bd_mount_mutex);
+		sema_init(&bdev->bd_mount_sem, 1);
 		INIT_LIST_HEAD(&bdev->bd_inodes);
 		INIT_LIST_HEAD(&bdev->bd_list);
 #ifdef CONFIG_SYSFS
fs/buffer.c +2 -2
diff -Naurp linux-2.6.18-rc6-orig/fs/buffer.c linux-2.6.18-rc6-mod/fs/buffer.c
@@ -213,7 +213,7 @@ struct super_block *freeze_bdev(struct b
 {
 	struct super_block *sb;
 
-	mutex_lock(&bdev->bd_mount_mutex);
+	down(&bdev->bd_mount_sem);
 	sb = get_super(bdev);
 	if (sb && !(sb->s_flags & MS_RDONLY)) {
 		sb->s_frozen = SB_FREEZE_WRITE;
@@ -255,7 +255,7 @@ void thaw_bdev(struct block_device *bdev
 		drop_super(sb);
 	}
 
-	mutex_unlock(&bdev->bd_mount_mutex);
+	up(&bdev->bd_mount_sem);
 }
 EXPORT_SYMBOL(thaw_bdev);
 
fs/super.c +2 -2
diff -Naurp linux-2.6.18-rc6-orig/fs/super.c linux-2.6.18-rc6-mod/fs/super.c
@@ -700,9 +700,9 @@ int get_sb_bdev(struct file_system_type 
 	 * will protect the lockfs code from trying to start a snapshot
 	 * while we are mounting
 	 */
-	mutex_lock(&bdev->bd_mount_mutex);
+	down(&bdev->bd_mount_sem);
 	s = sget(fs_type, test_bdev_super, set_bdev_super, bdev);
-	mutex_unlock(&bdev->bd_mount_mutex);
+	up(&bdev->bd_mount_sem);
 	if (IS_ERR(s))
 		goto error_s;
 
include/linux/fs.h +1 -1
diff -Naurp linux-2.6.18-rc6-orig/include/linux/fs.h linux-2.6.18-rc6-mod/include/linux/fs.h
@@ -414,7 +414,7 @@ struct block_device {
 	struct inode *		bd_inode;	/* will die */
 	int			bd_openers;
 	struct mutex		bd_mutex;	/* open/close mutex */
-	struct mutex		bd_mount_mutex;	/* mount mutex */
+	struct semaphore        bd_mount_sem;
 	struct list_head	bd_inodes;
 	void *			bd_holder;
 	int			bd_holders;
#1 lore raw
Quoted 8 lines · Srinivasa Ds

On debugging I found out that,"dmsetup suspend <device name>" calls
"freeze_bdev()",which locks "bd_mount_mutex" to make sure that no new mounts happen on bdev until thaw_bdev() is called.
This "thaw_bdev()" is getting called when we resume the device
through "dmsetup resume <device-name>".
Hence we have 2 processes,one of which locks
"bd_mount_mutex"(dmsetup suspend) and Another(dmsetup resume) unlocks it.

hm, to me this seems quite a fragile construct - even if the mutex-debugging warning is worked around by reverting to a semaphore.

Ingo

#2 lore raw
Quoted 11 lines · Ingo Molnar
Quoted 8 lines · Srinivasa Ds

On debugging I found out that,"dmsetup suspend <device name>" calls
"freeze_bdev()",which locks "bd_mount_mutex" to make sure that no new mounts happen on bdev until thaw_bdev() is called.
This "thaw_bdev()" is getting called when we resume the device
through "dmsetup resume <device-name>".
Hence we have 2 processes,one of which locks
"bd_mount_mutex"(dmsetup suspend) and Another(dmsetup resume) unlocks it.

Quoted 4 lines

hm, to me this seems quite a fragile construct - even if the mutex-debugging warning is worked around by reverting to a semaphore.

Ingo

Ingo, what do you feel is fragile about this? It seems like this is a reasonable way to go, except that maybe a down_trylock would be good if a 2nd process tries to freeze while it's already frozen...

Thanks,

-Eric

#3 lore raw
Quoted 17 lines · Eric Sandeen
Quoted 10 lines · Ingo Molnar
Quoted 7 lines · Srinivasa Ds

On debugging I found out that,"dmsetup suspend <device name>" calls
"freeze_bdev()",which locks "bd_mount_mutex" to make sure that no new mounts happen on bdev until thaw_bdev() is called.
This "thaw_bdev()" is getting called when we resume the device
through "dmsetup resume <device-name>".
Hence we have 2 processes,one of which locks
"bd_mount_mutex"(dmsetup suspend) and Another(dmsetup resume) unlocks it.

Quoted 4 lines

hm, to me this seems quite a fragile construct - even if the mutex-debugging warning is worked around by reverting to a semaphore.

Ingo

Quoted 8 lines

Ingo, what do you feel is fragile about this? It seems like this is a reasonable way to go, except that maybe a down_trylock would be good if a 2nd process tries to freeze while it's already frozen...

Thanks,

-Eric

Ingo, As per the discussion resending the patch with down_trylock.

---
 fs/block_dev.c     |    2 +-
 fs/buffer.c        |    6 ++++--
 fs/super.c         |    4 ++--
 include/linux/fs.h |    2 +-
 4 files changed, 8 insertions(+), 6 deletions(-)
fs/block_dev.c +1 -1
Index: linux-2.6.19-rc1/fs/block_dev.c
===================================================================
@@ -263,7 +263,7 @@ static void init_once(void * foo, kmem_c
 	{
 		memset(bdev, 0, sizeof(*bdev));
 		mutex_init(&bdev->bd_mutex);
-		mutex_init(&bdev->bd_mount_mutex);
+		sema_init(&bdev->bd_mount_sem, 1);
 		INIT_LIST_HEAD(&bdev->bd_inodes);
 		INIT_LIST_HEAD(&bdev->bd_list);
 #ifdef CONFIG_SYSFS
fs/buffer.c +4 -2
Index: linux-2.6.19-rc1/fs/buffer.c
===================================================================
@@ -188,7 +188,9 @@ struct super_block *freeze_bdev(struct b
 {
 	struct super_block *sb;
 
-	mutex_lock(&bdev->bd_mount_mutex);
+	if (down_trylock(&bdev->bd_mount_sem))
+		return -EBUSY;
+
 	sb = get_super(bdev);
 	if (sb && !(sb->s_flags & MS_RDONLY)) {
 		sb->s_frozen = SB_FREEZE_WRITE;
@@ -230,7 +232,7 @@ void thaw_bdev(struct block_device *bdev
 		drop_super(sb);
 	}
 
-	mutex_unlock(&bdev->bd_mount_mutex);
+	up(&bdev->bd_mount_sem);
 }
 EXPORT_SYMBOL(thaw_bdev);
 
fs/super.c +2 -2
Index: linux-2.6.19-rc1/fs/super.c
===================================================================
@@ -735,9 +735,9 @@ int get_sb_bdev(struct file_system_type 
 	 * will protect the lockfs code from trying to start a snapshot
 	 * while we are mounting
 	 */
-	mutex_lock(&bdev->bd_mount_mutex);
+	down(&bdev->bd_mount_sem);
 	s = sget(fs_type, test_bdev_super, set_bdev_super, bdev);
-	mutex_unlock(&bdev->bd_mount_mutex);
+	up(&bdev->bd_mount_sem);
 	if (IS_ERR(s))
 		goto error_s;
 
include/linux/fs.h +1 -1
Index: linux-2.6.19-rc1/include/linux/fs.h
===================================================================
@@ -456,7 +456,7 @@ struct block_device {
 	struct inode *		bd_inode;	/* will die */
 	int			bd_openers;
 	struct mutex		bd_mutex;	/* open/close mutex */
-	struct mutex		bd_mount_mutex;	/* mount mutex */
+	struct semaphore        bd_mount_sem;
 	struct list_head	bd_inodes;
 	void *			bd_holder;
 	int			bd_holders;
#4 lore raw
Quoted 25 lines · Srinivasa Ds
Quoted 24 lines · Eric Sandeen
Quoted 12 lines · Ingo Molnar
Quoted 9 lines · Srinivasa Ds

On debugging I found out that,"dmsetup suspend <device name>" calls
"freeze_bdev()",which locks "bd_mount_mutex" to make sure that no new mounts happen on bdev until thaw_bdev() is called.
This "thaw_bdev()" is getting called when we resume the device
through "dmsetup resume <device-name>".
Hence we have 2 processes,one of which locks
"bd_mount_mutex"(dmsetup suspend) and Another(dmsetup resume) unlocks it.

Quoted 5 lines

hm, to me this seems quite a fragile construct - even if the mutex-debugging warning is worked around by reverting to a semaphore.

Ingo

Ingo, what do you feel is fragile about this? It seems like this is a reasonable way to go, except that maybe a down_trylock would be good if a 2nd process tries to freeze while it's already frozen...

Quoted 5 lines
Quoted 4 lines

Thanks,

-Eric

Ingo, As per the discussion resending the patch with down_trylock.

Hi,

I still think that effectively exporting this semaphore to userspace is a big design mistake; but at least it can't be a mutex for this reason so the patch is sane in that regard...

Greetings,
Arjan van de Ven

#5 lore raw

Hi,

Checking through the code, I think that the semaphore -> mutex change was introduced in the original 2.6.17 kernel release. If it has problems, then it will affect all 2.6.17 and 2.6.18 kernels as the mutex code is still in the 2.6.18.1 release.

Can someone explain under what conditions this bug will occur and what its impact is? We are currently using the 2.6.16.20 kernel (which seems OK) but were keen to move to 2.6.18 as it has the new SATA hotplug ability. We use LVM and dmapper heavily, however, so don't want to switch if there are issues with it.

Alternatively, would locally updating the 2.6.18.1 kernel with Srinivasa's patch be sensible to restore the 2.6.16 style semaphore code (which does seem to be stable)?

Thanks,

Roger

Quoted 55 lines

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Arjan van de Ven
Sent: 10 October 2006 16:19
To: Srinivasa Ds
Cc: Eric Sandeen; [email protected]; [email protected]; [email protected]; Ingo Molnar; [email protected]
Subject: [linux-lvm] Re: [RFC] Reverting "bd_mount_mutex" to "bd_mount_sem"

Quoted 25 lines · Srinivasa Ds
Quoted 24 lines · Eric Sandeen
Quoted 12 lines · Ingo Molnar
Quoted 9 lines · Srinivasa Ds

On debugging I found out that,"dmsetup suspend <device name>" calls
"freeze_bdev()",which locks "bd_mount_mutex" to make sure that no new mounts happen on bdev until thaw_bdev() is called.
This "thaw_bdev()" is getting called when we resume the device
through "dmsetup resume <device-name>".
Hence we have 2 processes,one of which locks
"bd_mount_mutex"(dmsetup suspend) and Another(dmsetup resume) unlocks it.

Quoted 5 lines

hm, to me this seems quite a fragile construct - even if the mutex-debugging warning is worked around by reverting to a semaphore.

Ingo

Ingo, what do you feel is fragile about this? It seems like this is a reasonable way to go, except that maybe a down_trylock would be good if a 2nd process tries to freeze while it's already frozen...

Quoted 5 lines
Quoted 4 lines

Thanks,

-Eric

Ingo, As per the discussion resending the patch with down_trylock.

Hi,

I still think that effectively exporting this semaphore to userspace is a big design mistake; but at least it can't be a mutex for this reason so the patch is sane in that regard...

Greetings,
Arjan van de Ven

_______________________________________________
linux-lvm mailing list
[email protected]
https://www.redhat.com/mailman/listinfo/linux-lvm
read the LVM HOW-TO at http://tldp.org/HOWTO/LVM-HOWTO/