Buffer head async write

Buffer head async write

[Gorka Guardiola]

I am trying to 	mark a buffer for async write, but I am  having a race condition
to unlock the buffer. I am doing.

        bh = getbmapst(bdev, nsect); //this function calls bread
	if (!bh) {
		printk(KERN_ALERT "I/O error, bitmap cannot be trusted\n");
		return -1;
	}
	
	lock_buffer(bh);	

	byte = bh->b_data + byteoffset;
	*byte |= 1 << bitoffset;
	lock_page(bh->b_page);
	mark_page_accessed(bh->b_page);
	__set_page_dirty_nobuffers(bh->b_page);
	unlock_page(bh->b_page);
	set_buffer_uptodate(bh);
	mark_buffer_dirty(bh);
	mark_buffer_async_write(bh);
	unlock_buffer(bh);
	brelse(bh);

What I want is the buffer head not to write synchronously when I do
the brelse, to
coalesce writes on it. It works normally, but this code has some form
of race condition.
The problem is that (in some circumstances, for example while using a
loop device against a file) this makes the fs/buffer.c:603
BUG_ON break. I am using the latest version of the kernel with SMP on a machine
with 4 processors.

Can anybody tell me what am I doing wrong?. I tried some other schemes, like
bit_spin_locking the BH_Uptodate_Lock, but it didn't work either.

Re: Buffer head async write

[Andrew Morton]

On Wed, 30 Aug 2006 11:35:44 -0500
"Gorka Guardiola" <paurealkml@gmail.com> wrote:

> I am trying to 	mark a buffer for async write, but I am  having a race condition
> to unlock the buffer. I am doing.
> 
>         bh = getbmapst(bdev, nsect); //this function calls bread
> 	if (!bh) {
> 		printk(KERN_ALERT "I/O error, bitmap cannot be trusted\n");
> 		return -1;
> 	}
> 	
> 	lock_buffer(bh);	
> 
> 	byte = bh->b_data + byteoffset;
> 	*byte |= 1 << bitoffset;
> 	lock_page(bh->b_page);
> 	mark_page_accessed(bh->b_page);
> 	__set_page_dirty_nobuffers(bh->b_page);
> 	unlock_page(bh->b_page);
> 	set_buffer_uptodate(bh);
> 	mark_buffer_dirty(bh);
> 	mark_buffer_async_write(bh);
> 	unlock_buffer(bh);
> 	brelse(bh);

There are rather a lot of mistakes in there.

More like this:

	lock_page(bh->b_page);
	lock_buffer(bh);
	byte = bh->b_data + byteoffset;
	*byte |= 1 << bitoffset;
	mark_buffer_dirty(bh);
	unlock_buffer(bh);
	flush_dcache_page(page);
	unlock_page(page);
	put_bh(bh);

now, the buffer is dirty and will be written back by pdflush, sync, etc.

If you really want to start async IO against that buffer now, do

	ll_rw_block(WRITE, 1, &bh);

just before the final put_bh().

	

Buffer head async write

[Gorka Guardiola]

I am trying to  mark a buffer for async write, but I am  having a race condition
to unlock the buffer. I am doing.

       bh = getbmapst(bdev, nsect); //this function calls bread
       if (!bh) {
               printk(KERN_ALERT "I/O error, bitmap cannot be trusted\n");
               return -1;
       }

       lock_buffer(bh);

       byte = bh->b_data + byteoffset;
       *byte |= 1 << bitoffset;
       lock_page(bh->b_page);
       mark_page_accessed(bh->b_page);
       __set_page_dirty_nobuffers(bh->b_page);
       unlock_page(bh->b_page);
       set_buffer_uptodate(bh);
       mark_buffer_dirty(bh);
       mark_buffer_async_write(bh);
       unlock_buffer(bh);
       brelse(bh);

What I want is the buffer head not to write synchronously when I do
the brelse, to
coalesce writes on it. It works normally, but this code has some form
of race condition.
The problem is that (in some circumstances, for example while using a
loop device against a file) this makes the fs/buffer.c:603
BUG_ON break. I am using the latest version of the kernel with SMP on a machine
with 4 processors.

Can anybody tell me what am I doing wrong?. I tried some other schemes, like
bit_spin_locking the BH_Uptodate_Lock, but they didn't work either.