Message-ID: <20080416103531.GC6116@duck.suse.cz>
Date:	Wed, 16 Apr 2008 12:35:31 +0200
From:	Jan Kara <jack@...e.cz>
To:	Mingming Cao <cmm@...ibm.com>
Cc:	linux-ext4@...r.kernel.org, sandeen@...hat.com
Subject: Re: Delayed allocation and page_lock vs transaction start ordering

On Tue 15-04-08 16:33:17, Mingming Cao wrote:
> On Tue, 2008-04-15 at 16:28 -0700, Mingming Cao wrote:
> > On Tue, 2008-04-15 at 11:08 -0700, Mingming Cao wrote:
> > > On Tue, 2008-04-15 at 18:14 +0200, Jan Kara wrote:
> > > >   Hi,
> > > > 
> > > >   I've ported my patch inversing locking ordering of page_lock and
> > > > transaction start to ext4 (on top of ext4 patch queue). Everything except
> > > > delayed allocation is converted (the patch is below for interested
> > > > readers). The question is how to proceed with delayed allocation. Its
> > > > current implementation in VFS is designed to work well with the old
> > > > ordering (page lock first, then start a transaction). We could bend it to
> > > > work with the new locking ordering but I really see no point since ext4 is
> > > > the only user. 
> > > 
> > > I think the plan is port the changes to ext2/3/JFS and support delayed
> > > allocation on those filesystems. 
> > > 
> > > > Also XFS has AFAIK ordering first start transaction, then
> > > > lock pages so if we should ever merge delayed alloc implementations the new
> > > > ordering would make it easier.
> > > >   So what do people think here? Do you agree with reimplementing current
> > > > mpage_da_... functions?
> > > 
> > > It worth a try, but I could not see how to bend delayed allocation to
> > > work the new ordering:( With delayed allocation Ext4 gets into
> > > writepage() directly with page locked, but we need to start transaction
> > > to do block allocation...:(
> > 
> > Looked again it seems possible to reservse the order with delayed
> > allocation. with ext3_da_writepgaes() we could start the journal before
> > calling mpage_da_writepages()(which will lock the pages), instead of
> > start the journal inside ext4_da_get_block_write(). So that we could get
> > the locking order right. Just need to taking care of the estimated
> > credits right.
> > 
> > How about this? (untested, just throw out for comment)
> 
> Seems sent out an old version, this version compiles
  Thanks for the patch. Some comments are below.

> ---
>  fs/ext4/inode.c |   53 ++++++++++++++++++++++++++++++++++++++++-------------
>  1 file changed, 40 insertions(+), 13 deletions(-)
> 
> Index: linux-2.6.25-rc9/fs/ext4/inode.c
> ===================================================================
> --- linux-2.6.25-rc9.orig/fs/ext4/inode.c	2008-04-15 15:40:33.000000000 -0700
> +++ linux-2.6.25-rc9/fs/ext4/inode.c	2008-04-15 16:32:10.000000000 -0700
> @@ -1437,18 +1437,12 @@ static int ext4_da_get_block_prep(struct
>  static int ext4_da_get_block_write(struct inode *inode, sector_t iblock,
>  				   struct buffer_head *bh_result, int create)
>  {
> -	int ret, needed_blocks = ext4_writepage_trans_blocks(inode);
> +	int ret;
>  	unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
>  	loff_t disksize = EXT4_I(inode)->i_disksize;
>  	handle_t *handle = NULL;
>  
> -	if (create) {
> -		handle = ext4_journal_start(inode, needed_blocks);
> -		if (IS_ERR(handle)) {
> -			ret = PTR_ERR(handle);
> -			goto out;
> -		}
> -	}
> +	handle = ext4_journal_current_handle();
  Maybe we could assert that handle != NULL? When using delayed allocation,
a transaction should always be started.

>  	ret = ext4_get_blocks_wrap(handle, inode, iblock, max_blocks,
>  				   bh_result, create, 0);
> @@ -1483,17 +1477,51 @@ static int ext4_da_get_block_write(struc
>  		ret = 0;
>  	}
>  
> -out:
> -	if (handle && !IS_ERR(handle))
> -		ext4_journal_stop(handle);
> -
>  	return ret;
>  }
>  
> +/*
> + * For now just follow the DIO way to estimate the max credits
> + * needed to write out EXT4_MAX_BUF_BLOCKS pages.
> + * todo: need to calculate the max credits need for
> + * extent based files, currently the DIO credits is based on
> + * indirect-blocks mapping way.
> + *
> + * Probably should have a generic way to calculate credits
> + * for DIO, writepages, and truncate
> + */
> +#define EXT4_MAX_BUF_BLOCKS	DIO_MAX_BLOCKS
> +#define EXT4_MAX_BUF_CREDITS	DIO_CREDITS
> +
>  static int ext4_da_writepages(struct address_space *mapping,
>  				struct writeback_control *wbc)
>  {
> -	return mpage_da_writepages(mapping, wbc, ext4_da_get_block_write);
> +	struct inode *inode = mapping->host;
> +	handle_t *handle = NULL;
> +	int needed_blocks;
> +	int ret;
> +
> +	/*
> +	 * Estimate the worse case needed credits to write out
> +	 * EXT4_MAX_BUF_BLOCKS pages
> +	 */
> +	needed_blocks = EXT4_MAX_BUF_CREDITS;
> +
> +	/* start the transaction with credits*/
> +	handle = ext4_journal_start(inode, needed_blocks);
> +	if (IS_ERR(handle)) {
> +		ret = PTR_ERR(handle);
> +		return ret;
> +	}
> +
> +	/* set the max pages could be write-out at a time */
> +	wbc->range_end = wbc->range_start +
> +			EXT4_MAX_BUF_BLOCKS << PAGE_CACHE_SHIFT - 1;
  I think limiting mpage_da_writepages through nr_to_write is better than
through range_end. That way you don't count clean pages...

> +
> +	ret = mpage_da_writepages(mapping, wbc, ext4_da_get_block_write);
> +	ext4_journal_stop(handle);
  But here we can't just stop. We have to write everything original caller
has asked about (at least in WB_SYNC_ALL mode). But the question is where
to resume because scanning the whole range again is kind-of excessive and
prone do livelock with other process dirtying the file via mmap. Maybe if
we slightly modified write_cache_pages() to always store in writeback_index
where they finished, we could use this value.

> +
> +	return ret;
>  }
>  
>  static int ext4_da_write_begin(struct file *file, struct address_space *mapping,

									Honza
-- 
Jan Kara <jack@...e.cz>
SUSE Labs, CR
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists