Re: Regression caused by using node_to_bdi()

[Jan Kara <jack@suse.cz>]

On Sun 12-04-15 14:33:12, Boaz Harrosh wrote:
> On 04/10/2015 02:25 PM, Zhao Lei wrote:
> > Hi, Christoph Hellwig
> > 
> > resend: + cc lkml, linux-fsdevel
> > 
> > Since there is no response for my last mail, I worry that some problem in
> > the mail system, please allow me to resend it.
> > 
> > I found regression in v4.0-rc1 caused by this patch:
> >  Author: Christoph Hellwig <hch@lst.de>
> >  Date:   Wed Jan 14 10:42:36 2015 +0100
> >  fs: export inode_to_bdi and use it in favor of mapping->backing_dev_info
> > 
> <>
> > Result is following:
> >  v3.19-rc1: testcnt=40 average=135.677 range=[132.460,139.130] stdev=1.610 cv=1.19%
> >  v4.0-rc1: testcnt=40 average=130.970 range=[127.980,132.050] stdev=1.012 cv=0.77%
> > 
> > Then I bisect above case between v3.19-rc1 and v4.0-rc1, and found
> > this patch caused the regresstion.
> > 
> > Maybe it is because kernel need more time to call node_to_bdi(),
> > compared with "using inode->i_mapping->backing_dev_info directly" in
> > old code.
> > 
> > Is there some way to speed up it(inline, or some access some variant
> > in struct directly, ...)?
> > 
> 
> Christoph hi
> 
> Both node_to_bdi() and sb_is_blkdev_sb() 
>  (and I_BDEV() && blk_get_backing_dev_info())
> Are an exported function calls.
> 
> Can we not make blockdev_superblock->s_bdi == NULL,
> and then optimize-out the call to sb_is_blkdev_sb() to only
> that case. Something like:
> 
> ---
> 
> diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
> index 32a8bbd..e0375e1 100644
> --- a/fs/fs-writeback.c
> +++ b/fs/fs-writeback.c
> @@ -78,7 +78,7 @@ int writeback_in_progress(struct backing_dev_info *bdi)
>  }
>  EXPORT_SYMBOL(writeback_in_progress);
>  
> -struct backing_dev_info *inode_to_bdi(struct inode *inode)
> +struct backing_dev_info *__inode_to_bdi(struct inode *inode)
>  {
>  	struct super_block *sb;
>  
> @@ -92,7 +92,7 @@ struct backing_dev_info *inode_to_bdi(struct inode *inode)
>  #endif
>  	return sb->s_bdi;
>  }
> -EXPORT_SYMBOL_GPL(inode_to_bdi);
> +EXPORT_SYMBOL_GPL(__inode_to_bdi);
>  
>  static inline struct inode *wb_inode(struct list_head *head)
>  {
> diff --git a/include/linux/backing-dev.h b/include/linux/backing-dev.h
> index aff923a..7d172f5 100644
> --- a/include/linux/backing-dev.h
> +++ b/include/linux/backing-dev.h
> @@ -107,7 +107,16 @@ struct backing_dev_info {
>  #endif
>  };
>  
> -struct backing_dev_info *inode_to_bdi(struct inode *inode);
> +struct backing_dev_info *__inode_to_bdi(struct inode *inode);
> +
> +static inline
> +struct backing_dev_info *inode_to_bdi(struct inode *inode)
> +{
> +	if (!inode || !inode->i_sb)
> +		return __inode_to_bdi(inode);
> +
> +	return inode->i_sb->s_bdi;
> +}
  This is wrong for block-device inodes, isn't it?

							Honza

-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR