Hello,
On Fri 31-08-12 12:51:58, Namjae Jeon wrote:
> From: Namjae Jeon <namjae.j...@samsung.com>
>
> 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 <j...@suse.cz>
SUSE Labs, CR
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
