2013/1/21, Jan Kara <jack@xxxxxxx>:
> On Sat 19-01-13 11:17:14, Namjae Jeon wrote:
>> From: Namjae Jeon <namjae.jeon@xxxxxxxxxxx>
>>
>> This patch implements extent caching in case of file reading.
>> While reading a file, currently, UDF reads metadata serially
>> which takes a lot of time depending on the number of extents present
>> in the file. Caching last accessd extent improves metadata read time.
>> Instead of reading file metadata from start, now we read from
>> the cached extent.
>>
>> This patch considerably improves the time spent by CPU in kernel mode.
>> For example, while reading a 10.9 GB file using dd:
>> Time before applying patch:
>> 11677022208 bytes (10.9GB) copied, 1529.748921 seconds, 7.3MB/s
>> real 25m 29.85s
>> user 0m 12.41s
>> sys 15m 34.75s
>>
>> Time after applying patch:
>> 11677022208 bytes (10.9GB) copied, 1469.338231 seconds, 7.6MB/s
>> real 24m 29.44s
>> user 0m 15.73s
>> sys 3m 27.61s
>>
>> Signed-off-by: Namjae Jeon <namjae.jeon@xxxxxxxxxxx>
>> Signed-off-by: Ashish Sangwan <a.sangwan@xxxxxxxxxxx>
>> Signed-off-by: Bonggil Bak <bgbak@xxxxxxxxxxx>
> Thanks for the patch Namjae. I did a few more changes to the patch.
> Please check them whether you think they are OK.
>
> diff --git a/fs/udf/ialloc.c b/fs/udf/ialloc.c
> index 0cb208e..7e5aae4 100644
> --- a/fs/udf/ialloc.c
> +++ b/fs/udf/ialloc.c
> @@ -117,10 +117,6 @@ struct inode *udf_new_inode(struct inode *dir, umode_t
> mode, int *err)
> iinfo->i_lenAlloc = 0;
> iinfo->i_use = 0;
> iinfo->i_checkpoint = 1;
> - memset(&iinfo->cached_extent, 0, sizeof(struct udf_ext_cache));
> - spin_lock_init(&(iinfo->i_extent_cache_lock));
> - /* Mark extent cache as invalid for now */
> - iinfo->cached_extent.lstart = -1;
> if (UDF_QUERY_FLAG(inode->i_sb, UDF_FLAG_USE_AD_IN_ICB))
> iinfo->i_alloc_type = ICBTAG_FLAG_AD_IN_ICB;
> else if (UDF_QUERY_FLAG(inode->i_sb, UDF_FLAG_USE_SHORT_AD))
> diff --git a/fs/udf/inode.c b/fs/udf/inode.c
> index 8494b8c..fb0c4c4 100644
> --- a/fs/udf/inode.c
> +++ b/fs/udf/inode.c
> @@ -1305,9 +1305,6 @@ static void udf_fill_inode(struct inode *inode, struct
> buffer_head *bh)
> iinfo->i_lenAlloc = 0;
> iinfo->i_next_alloc_block = 0;
> iinfo->i_next_alloc_goal = 0;
> - memset(&iinfo->cached_extent, 0, sizeof(struct udf_ext_cache));
> - spin_lock_init(&(iinfo->i_extent_cache_lock));
> - iinfo->cached_extent.lstart = -1;
> if (fe->descTag.tagIdent == cpu_to_le16(TAG_IDENT_EFE)) {
> iinfo->i_efe = 1;
> iinfo->i_use = 0;
>
Hi Jan.
> Initialization now happens in udf_alloc_inode(). Also it's not necessary
> to initialized cached_extent.epos when lstart == -1 - noone should look at
> that.
This change is okay.
>
> @@ -2222,6 +2219,8 @@ int udf_read_extent_cache(struct inode *inode, loff_t
> bcount,
> *lbcount = iinfo->cached_extent.lstart;
> memcpy(pos, &iinfo->cached_extent.epos,
> sizeof(struct extent_position));
> + if (pos->bh)
> + get_bh(pos->bh);
> spin_unlock(&iinfo->i_extent_cache_lock);
> return 1;
> } else
> This is the most important - we should give buffer reference to pos->bh.
> Caller will eventually free it right?
This change is not required as we give buffer reference to pos->bh at
the time of cache update.
When we start reading a file, first we try to read the cache which
will lead to cache miss.
So, we would really access the pos->bh in udf_update_extent_cache for
the first time, and this is where the buffer reference is incremented.
Calling get_bh at 2 places will eventually lead to mem leak.
Let me know your opinion.
Thanks for review and change!
>
> @@ -2236,8 +2235,7 @@ void udf_update_extent_cache(struct inode *inode,
> loff_t estart,
> struct udf_inode_info *iinfo = UDF_I(inode);
>
> spin_lock(&iinfo->i_extent_cache_lock);
> - if (pos->bh != NULL)
> - /* Increase ref count */
> + if (pos->bh)
> get_bh(pos->bh);
> memcpy(&iinfo->cached_extent.epos, pos,
> sizeof(struct extent_position));
> @@ -2266,4 +2264,3 @@ void udf_clear_extent_cache(struct udf_inode_info
> *iinfo)
> iinfo->cached_extent.lstart = -1;
> }
> }
> -
> diff --git a/fs/udf/super.c b/fs/udf/super.c
> index 186adbf..da8ce9f 100644
> --- a/fs/udf/super.c
> +++ b/fs/udf/super.c
> @@ -134,6 +134,8 @@ static struct inode *udf_alloc_inode(struct super_block
> *sb)
> ei->i_next_alloc_goal = 0;
> ei->i_strat4096 = 0;
> init_rwsem(&ei->i_data_sem);
> + ei->cached_extent.lstart = -1;
> + spin_lock_init(&ei->i_extent_cache_lock);
>
> return &ei->vfs_inode;
> }
>
> Honza
> --
> Jan Kara <jack@xxxxxxx>
> SUSE Labs, CR
>
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
