Re: [PATCH v2] udf: extent cache implementation for manipulatingblock map

[Jan Kara]

  Hello,

On Fri 31-08-12 12:51:58, Namjae Jeon wrote:
> From: Namjae Jeon <namjae.jeon@xxxxxxxxxxx>
> 
> While mapping logical blocks of a file to physical blocks on the partition,
> everytime UDF read file metadata from the begining which decrease preformance.
> The drawback of this scheme is more prominent while reading large files.
> For example, while reading a large file of ~5GB, read speed will
> gradually become less as we near the end of file because of the time
> taken in calculating the corresponding physical block.
> 
> This patch implements caching and remembers the location of the last read
> extent. Instead of reading file metadata from begining, start from the
> cached location.
  Thanks for the patch. I like this much better than the previous one. Some
comments are below:

> diff --git a/fs/udf/inode.c b/fs/udf/inode.c
> index 1a0588e..f4cc4e0 100644
> --- a/fs/udf/inode.c
> +++ b/fs/udf/inode.c
> @@ -357,8 +359,14 @@ static int udf_get_block(struct inode *inode, sector_t block,
>  	if (!phys)
>  		goto abort;
>  
> -	if (new)
> +	if (new) {
> +		unsigned char lshift = inode->i_sb->s_blocksize_bits;
>  		set_buffer_new(bh_result);
> +		if (iinfo->cached_extent.sanity &&
> +		    (iinfo->cached_extent.eblock > (block << lshift)))
> +			/* Block allocated for hole, invalidate cache */
> +			udf_clear_extent_cache(iinfo);
> +	}
>  	map_bh(bh_result, inode->i_sb, phys);
  Instead of this, you should change inode_getblk() to also use the extent
cache (for initialization of prev_epos when appropriate) and then update
the extent cache to the just found / created extent when the function is
done.

>  abort:
> @@ -2132,11 +2144,12 @@ int8_t inode_bmap(struct inode *inode, sector_t block,
>  	struct udf_inode_info *iinfo;
>  
>  	iinfo = UDF_I(inode);
> -	pos->offset = 0;
> -	pos->block = iinfo->i_location;
> -	pos->bh = NULL;
> -	*elen = 0;
> -
> +	if (!udf_read_extent_cache(inode, bcount, &lbcount, pos)) {
> +		pos->offset = 0;
> +		pos->block = iinfo->i_location;
> +		pos->bh = NULL;
> +	}
> +		*elen = 0;
  There's TAB added here which doesn't make sense.

>  	do {
>  		etype = udf_next_aext(inode, pos, eloc, elen, 1);
>  		if (etype == -1) {
...
> @@ -2176,3 +2190,52 @@ long udf_block_map(struct inode *inode, sector_t block)
>  	else
>  		return ret;
>  }
> +int udf_read_extent_cache(struct inode *inode, loff_t bcount,
> +			  loff_t *lbcount, struct extent_position *pos)
> +{
> +	struct udf_inode_info *iinfo = UDF_I(inode);
> +	mutex_lock(&iinfo->i_extent_cache_lock);
> +	if ((iinfo->cached_extent.eblock <= bcount) &&
> +	    (iinfo->cached_extent.sanity)) {
> +		/* Cache hit */
> +		*lbcount = iinfo->cached_extent.eblock;
> +		memcpy(pos, &iinfo->cached_extent.epos,
> +		       sizeof(struct extent_position));
> +		mutex_unlock(&iinfo->i_extent_cache_lock);
> +		return 1;
> +	} else
> +		udf_clear_extent_cache(iinfo);
  It would be less confusing, if udf_read_extent_cache() didn't clear the
cache in case of a miss. Then in udf_update_extent_cache() just clear the
cache if it is still valid.

> +
> +	mutex_unlock(&iinfo->i_extent_cache_lock);
> +	return 0;
> +}
> +void udf_update_extent_cache(struct inode *inode, loff_t estart,
> +			     struct extent_position *pos)
> +{
> +	struct udf_inode_info *iinfo = UDF_I(inode);
> +	mutex_lock(&iinfo->i_extent_cache_lock);
> +	if (pos->bh != NULL)
> +		/* Increase ref count */
> +		get_bh(pos->bh);
> +	memcpy(&iinfo->cached_extent.epos, pos,
> +	       sizeof(struct extent_position));
> +	iinfo->cached_extent.eblock = estart;
> +	iinfo->cached_extent.sanity = 1;
> +	switch (iinfo->i_alloc_type) {
> +	case ICBTAG_FLAG_AD_SHORT:
> +		iinfo->cached_extent.epos.offset -= sizeof(struct short_ad);
> +		break;
> +	case ICBTAG_FLAG_AD_LONG:
> +		iinfo->cached_extent.epos.offset -= sizeof(struct long_ad);
> +	}
> +	mutex_unlock(&iinfo->i_extent_cache_lock);
> +}
> +
> +void udf_clear_extent_cache(struct udf_inode_info *iinfo)
> +{
> +	if (iinfo->cached_extent.sanity) {
> +		brelse(iinfo->cached_extent.epos.bh);
> +		memset(&iinfo->cached_extent, 0, sizeof(struct udf_ext_cache));
> +	}
> +}
> +
  I think udf_clear_entent_cache() should take i_extent_cache_lock. Or if
you are sure it's not needed, you need a good documentation why.

> diff --git a/fs/udf/udf_i.h b/fs/udf/udf_i.h
> index bb8309d..ec3878a 100644
> --- a/fs/udf/udf_i.h
> +++ b/fs/udf/udf_i.h
> @@ -1,6 +1,20 @@
>  #ifndef _UDF_I_H
>  #define _UDF_I_H
>  
> +struct extent_position {
> +	struct buffer_head *bh;
> +	uint32_t offset;
> +	struct kernel_lb_addr block;
> +};
> +
> +struct udf_ext_cache {
> +	/* Extent position */
> +	struct extent_position epos;
> +	/* Start logical block */
> +	loff_t eblock;
  Well, this is really in bytes not blocks, isn't it? Then call it like
'lstart' and add a comment that it's in bytes.

> +	int8_t sanity;
  Make this 'bool' and call it 'valid'. That's more common.

								Honza
-- 
Jan Kara <jack@xxxxxxx>
SUSE Labs, CR
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/