How to setup a buffer_head in a driver

How to setup a buffer_head in a driver

[Brian Kelly]

Hi,
I'm writing a device driver that among other things needs to write data,
manufactured by the driver itself, to a block device.

I have this data in block sized kmalloc()'d chunks. So what I'm doing is
allocating a struct buffer_head, initialising it, fill out it's various
fields and send it to generic_make_request(). Something like the following
[inspired by looking at various drivers like loop and ram]:

	do {
		if((bh = kmem_cache_alloc(bh_cachep, SLAB_NOIO)) != NULL){
			break;
		}
		run_task_queue(&tq_disk);
		schedule_timeout(HZ);
	} while (1);

	memset(bh, 0, sizeof(*bh));

	bh->b_size = size;
	bh->b_dev = dev;
	bh->b_rdev = dev;
	bh->b_data = data;
	init_waitqueue_head(&bh->b_wait);
	bh->b_rsector = sect;
	bh->b_end_io = write_done;
	bh->b_private = NULL;
	generic_make_request(WRITE, bh);


Now this causes a panic in ll_rw_blk.c because the b_state isn't set
correctly. I experimented with this a little but decided I needed some
proper direction on this whole endeavour.

Soooo, what do I really need to do? What's the correct way to do what I want
to do or could someone point me at an existing driver that does this
sort of thing.

Thanks,

Brian
-- 
bkelly@sulaco.com

Re: How to setup a buffer_head in a driver

[Andrew Morton]

On Tuesday 14 January 2003 02:35 pm, Brian Kelly wrote:
>
> Hi,
> I'm writing a device driver that among other things needs to write data,
> manufactured by the driver itself, to a block device.
> 
> I have this data in block sized kmalloc()'d chunks. So what I'm doing is
> allocating a struct buffer_head, initialising it, fill out it's various
> fields and send it to generic_make_request().

It's probably better to use submit_bh().  Set the BH_Lock and BH_Mapped bits,
also set up b_end_io.  Then do a wait_on_buffer(), wait for the IO to
complete.  There's some similar code in
fs/jbd/journal.c:journal_write_metadata_buffer().

However, what you're doing is an odd thing.  If there is already pagecache
against that block device then the kernel doesn't know that you've changed
the bytes on-disk and will cheerfully proceed to use (and write out) the
cached data.  You'll lose your modifications..

It would be better to use sb_getblk() or bread(), to lock the returned
buffer_head, then copy your data into it and to then write it back with
submit_bh() or ll_rw_block().  Or just leave it dirty and let the kernel
write it out in due course.

Re: How to setup a buffer_head in a driver

[Brian Kelly]

I wrote:
 > I'm writing a device driver that among other things needs to write data,
 > manufactured by the driver itself, to a block device.
 >
 > I have this data in block sized kmalloc()'d chunks. So what I'm doing is
 > allocating a struct buffer_head, initialising it, fill out it's various
 > fields and send it to generic_make_request().

Andrew Morton wrote:
 >It's probably better to use submit_bh().  Set the BH_Lock and BH_Mapped bits,
 >also set up b_end_io.  Then do a wait_on_buffer(), wait for the IO to
 >complete.  There's some similar code in
 >fs/jbd/journal.c:journal_write_metadata_buffer().

Thanks, that's exactly the sort of thing I was looking for.

 >However, what you're doing is an odd thing.  If there is already pagecache
 >against that block device then the kernel doesn't know that you've changed
 >the bytes on-disk and will cheerfully proceed to use (and write out) the
 >cached data.  You'll lose your modifications..
 >
 >It would be better to use sb_getblk() or bread(), to lock the returned
 >buffer_head, then copy your data into it and to then write it back with
 >submit_bh() or ll_rw_block().  Or just leave it dirty and let the kernel
 >write it out in due course.

Fair enough, that seems like the right thing to do so I'll look into it.

Thanks,

Brian
-- 
bkelly@sulaco.com