buffer head read/write

buffer head read/write

[David Chow]

Hi all,

Is there a way to make sure I submit a buffer head read/write and make 
sure it is commited immediately? (sychronized) . Please give direction 
on examples in the kernel code? Thanks.

regards,
David Chow

Re: buffer head read/write

[Bryan Henderson]

>Is there a way to make sure I submit a buffer head read/write and make
>sure it is commited immediately? (sychronized)

ll_rw_block() to submit it, followed by wait_on_buffer() to wait for the
I/O to complete.

>Please give direction on examples in the kernel code? Thanks.

I can't think of an example of where those are called together, but if you
search the fs/ directory for those, you should find plenty of examples of
using them individually.

Re: buffer head read/write

[Neil Brown]

On Tuesday December 10, davidchow@shaolinmicro.com wrote:
> Hi all,
> 
> Is there a way to make sure I submit a buffer head read/write and make 
> sure it is commited immediately? (sychronized) . Please give direction 
> on examples in the kernel code? Thanks.

md does (or will) read/write it's super-blocks that way.
In 2.5, see drivers/md/md.c:sync_page_io
In 2.4, see
    http://cgi.cse.unsw.edu.au/~neilb/patches/linux-stable/2.4.leading/2002-12-09:01/005MDSyncIo

... I haven't submitted it yet.

NeilBrown

Re: buffer head read/write

[David Chow]

Neil Brown wrote:

>On Tuesday December 10, davidchow@shaolinmicro.com wrote:
>  
>
>>Hi all,
>>
>>Is there a way to make sure I submit a buffer head read/write and make 
>>sure it is commited immediately? (sychronized) . Please give direction 
>>on examples in the kernel code? Thanks.
>>    
>>
>
>md does (or will) read/write it's super-blocks that way.
>In 2.5, see drivers/md/md.c:sync_page_io
>In 2.4, see
>    http://cgi.cse.unsw.edu.au/~neilb/patches/linux-stable/2.4.leading/2002-12-09:01/005MDSyncIo
>
>... I haven't submitted it yet.
>
>NeilBrown
>  
>
Thanks to all who give suggestions. I know how to make a call to 
general_make_request() but wait for the buffer to complete which is too 
slow. If I am doing a continuous sychronous read/write (which I already 
have my own way of buffer management), can I put those "dirty 
buffers"/"non-up-to-date buffers" in the head of the buffer queue so 
that they get written by their respective handler immediately? I am not 
familiar with buffer cache so would like to seek advice on doing this. 
Is there anyway to lock a buffer and get its buffer_head request handler 
and call end_io() immediately? Thanks for advice.

regards,
David Chow

Re: buffer head read/write

[Andrew Morton]

David Chow wrote:
> 
> Hi all,
> 
> Is there a way to make sure I submit a buffer head read/write and make
> sure it is commited immediately? (sychronized) . Please give direction
> on examples in the kernel code? Thanks.

grep wait_on_buffer fs/ext2/*.c - you'll see lots of examples.  eg:

        mark_buffer_dirty(bh);
        if (sb->s_flags & MS_SYNCHRONOUS) {
                ll_rw_block (WRITE, 1, &bh);
                wait_on_buffer (bh);
        }

Re: buffer head read/write

[Bryan Henderson]

>I know how to make a call to general_make_request() but
>wait for the buffer to complete which is too slow.
>If I am doing a continuous sychronous read/write ..., can I put those
"dirty
>buffers"/"non-up-to-date buffers" in the head of the buffer queue so
>that they get written by their respective handler immediately?

"immediately" in computer systems normally means "while I wait" rather than
"before doing anything else."  That's why you got a bunch of answers that
weren't what you were looking for.

What you want is a way to assign priority to a block I/O request, so that
a particular one gets executed first regardless of other factors, like
that it's a mile away from where the head is now or 100 other requests
came in before it.  There's no way to do priority in today's block I/O
layer.

Incidentally, this has nothing to do with the buffer cache.  The queue in
question is for all I/O to a block device; it is below the buffer cache.

Re: buffer head read/write

[David Chow]

  OK, I think you know what I want, and you've just told me this is 
impossible, then the best I could do is try to execute the task as best 
as the CPU can. This is stolen from Neils md super block writing routing 
that he just gave me a few days ago and merged with 
wait_for_complettion. Does this code make sense?

static int sync_buffer_io(kdev_t dev, unsigned long sector, int size,
			struct page *page, int rw)


	struct buffer_head bh;
	struct completion event;

	init_completion(&event);
	init_buffer(&bh, bh_complete, &event);
	bh.b_rdev = dev;
	bh.b_rsector = sector;
	bh.b_state	= (1 << BH_Req) | (1 << BH_Mapped);
	bh.b_size = size;
	bh.b_page = page;
	bh.b_reqnext = NULL;
	bh.b_data = page_address(page);
	generic_make_request(rw, &bh);

	run_task_queue(&tq_disk);

	spin_lock_irq(even.wait.lock);

	if (!event.done) {
		DECLARE_WAITQUEUE(wait, current);

		wait.flags |= WQ_FLAG_EXCLUSIVE;
		__add_wait_queue_tail(event.wait, &wait);
		do {
			__set_current_state(TASK_UNINTERRUPTIBLE);
			spin_unlock_irq(event.wait.lock);
			run_task_queue(&tq_disk);
			spin_lock_irq(event.wait.lock);
		} while (!x->done);
		__remove_wait_queue(event.wait, &wait);
	}
	event.done--;
	spin_unlock_irq(&x->wait.lock);

	return test_bit(BH_Uptodate, &bh.b_state);
}

This is dedicate to run the task of tq_disk instead of just schedule() which doesn't know
which task is going to run. Please comment. Thanks.

regards,
David




Bryan Henderson wrote:

>
>
>  
>
>>I know how to make a call to general_make_request() but
>>wait for the buffer to complete which is too slow.
>>If I am doing a continuous sychronous read/write ..., can I put those
>>    
>>
>"dirty
>  
>
>>buffers"/"non-up-to-date buffers" in the head of the buffer queue so
>>that they get written by their respective handler immediately?
>>    
>>
>
>"immediately" in computer systems normally means "while I wait" rather than
>"before doing anything else."  That's why you got a bunch of answers that
>weren't what you were looking for.
>
>What you want is a way to assign priority to a block I/O request, so that
>a particular one gets executed first regardless of other factors, like
>that it's a mile away from where the head is now or 100 other requests
>came in before it.  There's no way to do priority in today's block I/O
>layer.
>
>Incidentally, this has nothing to do with the buffer cache.  The queue in
>question is for all I/O to a block device; it is below the buffer cache.
>  
>

-- 

-----------------------------------------------------------------------------
David Chow	 | Managing Director
http://www.shaolinmicro.com
ShaoLin Microsystems - Multiply Efficiency
-----------------------------------------------------------------------------