Skip to contentSkip to conversation

fs-private data in struct block_device

PATCH In review Christoph Hellwig 5 participants 10 messages Jan 1, 2004 – Jan 4, 2004 VFS

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fs-private data in struct block_device

Al had some critism about a per-blockdev XFS datastructure a while ago and so I tried to follow the suggestion in favour of using struct block_device directly. Unfortunately XFS requires a little bit more information than what's available in struct block_device, in particular whether the blocksize (not sector size) for that devices is identical to the page size, and we can't really get that from elsewhere due to layering constraints.

Nathan Scott scott suggested turning the bd_invalidated fields into a flags field where the filesystem uses the so far unused flags for it's purposed (similar to struct buffer_head), but I think maybe the variant below that just adds a unsigned long for filesystem usage might be more useful as it also allows storing a pointer if needed. Comments?

include/linux/fs.h +1 -0
@@ -353,6 +353,7 @@
 	int			bd_invalidated;
 	struct gendisk *	bd_disk;
 	struct list_head	bd_list;
+	unsigned long		bd_fspriv; /* file-system private data */
 };
 
 /*
#1 lore raw
Quoted 8 lines · Christoph Hellwig

Al had some critism about a per-blockdev XFS datastructure a while ago and so I tried to follow the suggestion in favour of using struct block_device directly. Unfortunately XFS requires a little bit more information than what's available in struct block_device, in particular whether the blocksize (not sector size) for that devices is identical to the page size, and we can't really get that from elsewhere due to layering constraints.

Why doesn't `if (bdev->bd_block_size == PAGE_SIZE)' work?

Or sb->s_blocksize?

Generally, redundant info like this is a hassle because it has to be kept coherent with the thing which it is duplicating.

I suspect I don't understand the problem.

Quoted 10 lines

Nathan Scott scott suggested turning the bd_invalidated fields into a flags field where the filesystem uses the so far unused flags for it's purposed (similar to struct buffer_head), but I think maybe the variant below that just adds a unsigned long for filesystem usage might be more useful as it also allows storing a pointer if needed. Comments?

include/linux/fs.h line 359

Quoted 7 lines
include/linux/fs.h +1 -0
@@ -353,6 +353,7 @@
 	int			bd_invalidated;
 	struct gendisk *	bd_disk;
 	struct list_head	bd_list;
+	unsigned long		bd_fspriv; /* file-system private data */
 };
 
 /*

Seems sane. You'll need to decide whether that field should be operated upon with atomic ops. If so, document it. If not, document its locking, please.

#2 lore raw
Quoted 10 lines · Andrew Morton
Quoted 8 lines · Christoph Hellwig

Al had some critism about a per-blockdev XFS datastructure a while ago and so I tried to follow the suggestion in favour of using struct block_device directly. Unfortunately XFS requires a little bit more information than what's available in struct block_device, in particular whether the blocksize (not sector size) for that devices is identical to the page size, and we can't really get that from elsewhere due to layering constraints.

Quoted 14 lines

Why doesn't `if (bdev->bd_block_size == PAGE_SIZE)' work?

Or sb->s_blocksize?

Generally, redundant info like this is a hassle because it has to be kept coherent with the thing which it is duplicating.

I suspect I don't understand the problem.

Quoted 5 lines

Nathan Scott scott suggested turning the bd_invalidated fields into a flags field where the filesystem uses the so far unused flags for it's purposed (similar to struct buffer_head), but I think maybe the variant below that just adds a unsigned long for filesystem usage might be more useful as it also allows storing a pointer if needed. Comments?

Obvious: we need at least some rules regarding who can and who can not play with that field. IOW, which filesystem?

#7 lore raw
In reply to Andrew Morton
Quoted 11 lines · Andrew Morton
Quoted 8 lines · Christoph Hellwig

Al had some critism about a per-blockdev XFS datastructure a while ago and so I tried to follow the suggestion in favour of using struct block_device directly. Unfortunately XFS requires a little bit more information than what's available in struct block_device, in particular whether the blocksize (not sector size) for that devices is identical to the page size, and we can't really get that from elsewhere due to layering constraints.

Why doesn't `if (bdev->bd_block_size == PAGE_SIZE)' work?

Because it's different from the XFS block size. In XFS terms it's the sector size.

Quoted 1 line

Or sb->s_blocksize?

Would work, except that we don't get at per-superblock data in the place we need that information - it's in pagebuf which is the IRIX buffercache interface reimplemented ontop of the Linux pagecache, and that has neither access to Linux VFS nor XFS-internal data structures.

Quoted 3 lines

Seems sane. You'll need to decide whether that field should be operated upon with atomic ops. If so, document it. If not, document its locking, please.

That's completly left to the filesystem. For XFS it would be set once at mount time and then left constant.

#3 lore raw
In reply to [email protected]
Quoted 2 lines · On Thu, Jan 01, 2004 at 11:21:37PM +0000…

Obvious: we need at least some rules regarding who can and who can not play with that field. IOW, which filesystem?

Okay, let's make it more general. Throw the filesystem completly out of the few and say the private data can be used by the owner of that blockdevice as established by bd_claim, okay?

#4 lore raw
Quoted 7 lines · Christoph Hellwig
Quoted 2 lines · On Thu, Jan 01, 2004 at 11:21:37PM +0000…

Obvious: we need at least some rules regarding who can and who can not play with that field. IOW, which filesystem?

Okay, let's make it more general. Throw the filesystem completly out of the few and say the private data can be used by the owner of that blockdevice as established by bd_claim, okay?

OK... One question: what rules do you want wrt multiple claims with the same owner?

#5 lore raw
Quoted 6 lines · On Thu, Jan 01, 2004 at 11:28:32PM +0000…
Quoted 3 lines

Okay, let's make it more general. Throw the filesystem completly out of the few and say the private data can be used by the owner of that blockdevice as established by bd_claim, okay?

OK... One question: what rules do you want wrt multiple claims with the same owner?

I'd say this is the owners problem and add a big comment about that. What about something like the version below?:

include/linux/fs.h +7 -0
@@ -353,6 +353,13 @@
 	int			bd_invalidated;
 	struct gendisk *	bd_disk;
 	struct list_head	bd_list;
+	/*
+	 * Private data.  You must have bd_claimed the block_device
+	 * to use this.  NOTE:  bd_claims allows an owner to claim
+	 * the same device multiple timers, the owner must take special
+	 * care to not mess up bd_private for that case.
+	 */
+	unsigned long		bd_private;
 };
 
 /*
#8 lore raw
In reply to Andrew Morton
Quoted 9 lines · Andrew Morton
Quoted 9 lines
include/linux/fs.h +1 -0
@@ -353,6 +353,7 @@
 	int			bd_invalidated;
 	struct gendisk *	bd_disk;
 	struct list_head	bd_list;
+	unsigned long		bd_fspriv; /* file-system private data */
 };
 
Quoted 5 lines

include/linux/fs.h line 359

Quoted 1 line
include/linux/fs.h +0 -0
 /*

Seems sane. You'll need to decide whether that field should be operated upon with atomic ops. If so, document it. If not, document its locking, please.

It looks like the standard normally-a-pointer thing that's completely managed opaquely to the block device layer. In fact, I actually prefer that these be (void*) unless there is a common use as an integer.

Jeff

#9 lore raw
Quoted 3 lines · Jeff Garzik

It looks like the standard normally-a-pointer thing that's completely managed opaquely to the block device layer. In fact, I actually prefer that these be (void*) unless there is a common use as an integer.

The XFS usage would be as int (actually as boolean)

#6 lore raw
In reply to Christoph Hellwig

On Fri, 2 Jan 2004 00:32:27 +0100 Christoph Hellwig <[email protected]> wrote:

| On Thu, Jan 01, 2004 at 11:28:32PM +0000, [email protected] wrote:
| > > Okay, let's make it more general.  Throw the filesystem completly out of
| > > the few and say the private data can be used by the owner of that
| > > blockdevice as established by bd_claim, okay?
| > 
| > OK...  One question: what rules do you want wrt multiple claims with the
| > same owner?
| 
| I'd say this is the owners problem and add a big comment about that.
| What about something like the version below?:
| 
| 
| --- 1.276/include/linux/fs.h	Mon Dec 29 22:37:20 2003
| +++ edited/include/linux/fs.h	Fri Jan  2 01:30:42 2004
| @@ -353,6 +353,13 @@
|  	int			bd_invalidated;
|  	struct gendisk *	bd_disk;
|  	struct list_head	bd_list;
| +	/*
| +	 * Private data.  You must have bd_claimed the block_device
| +	 * to use this.  NOTE:  bd_claims allows an owner to claim
| +	 * the same device multiple timers, the owner must take special
                                    times;

| + * care to not mess up bd_private for that case.
preferably: not to mess up

| +	 */
| +	unsigned long		bd_private;
|  };
|  
|  /*
| -

just nits.
--
~Randy