[PATCH] udf: avoid redundant memcpy when writing data in ICB

[PATCH] udf: avoid redundant memcpy when writing data in ICB

[Chao Yu]

We have already copied the changed data of page to extent area in ICB when call
->write_end() of adinicb aops, so we do not need to copy them another time in
->writepage() of adinicb aops. SetPageUptodate is redundant too, so it's better
to remove these redundant codes.

Also it looks more reasonable to set inode dirty after extent area are modified,
we'd better move mark_inode_dirty() from ->writepage() to ->write_end().

Signed-off-by: Chao Yu <chao2.yu@samsung.com>
---
 fs/udf/file.c |   12 ++----------
 1 file changed, 2 insertions(+), 10 deletions(-)

diff --git a/fs/udf/file.c b/fs/udf/file.c
index d80738f..3cd0a50 100644
--- a/fs/udf/file.c
+++ b/fs/udf/file.c
@@ -65,17 +65,7 @@ static int udf_adinicb_readpage(struct file *file, struct page *page)
 static int udf_adinicb_writepage(struct page *page,
 				 struct writeback_control *wbc)
 {
-	struct inode *inode = page->mapping->host;
-	char *kaddr;
-	struct udf_inode_info *iinfo = UDF_I(inode);
-
 	BUG_ON(!PageLocked(page));
-
-	kaddr = kmap(page);
-	memcpy(iinfo->i_ext.i_data + iinfo->i_lenEAttr, kaddr, inode->i_size);
-	mark_inode_dirty(inode);
-	SetPageUptodate(page);
-	kunmap(page);
 	unlock_page(page);
 
 	return 0;
@@ -115,6 +105,8 @@ static int udf_adinicb_write_end(struct file *file,
 		kaddr + offset, copied);
 	kunmap_atomic(kaddr);
 
+	mark_inode_dirty(inode);
+
 	return simple_write_end(file, mapping, pos, len, copied, page, fsdata);
 }
 
-- 
1.7.9.5

Re: [PATCH] udf: avoid redundant memcpy when writing data in ICB

[Jan Kara]

On Mon 30-06-14 13:37:15, Chao Yu wrote:
> We have already copied the changed data of page to extent area in ICB when call
> ->write_end() of adinicb aops, so we do not need to copy them another time in
> ->writepage() of adinicb aops. SetPageUptodate is redundant too, so it's better
> to remove these redundant codes.
> 
> Also it looks more reasonable to set inode dirty after extent area are modified,
> we'd better move mark_inode_dirty() from ->writepage() to ->write_end().
  This doesn't look right. The copy is there because the page can be
modified via mmap as well, not only via buffered write... So if anything,
we may be able to remove the copy in udf_adinicb_write_end() (actually
using directly simple_write_end() for that), not the one in
udf_adinicb_write_begin().

								Honza
> 
> Signed-off-by: Chao Yu <chao2.yu@samsung.com>
> ---
>  fs/udf/file.c |   12 ++----------
>  1 file changed, 2 insertions(+), 10 deletions(-)
> 
> diff --git a/fs/udf/file.c b/fs/udf/file.c
> index d80738f..3cd0a50 100644
> --- a/fs/udf/file.c
> +++ b/fs/udf/file.c
> @@ -65,17 +65,7 @@ static int udf_adinicb_readpage(struct file *file, struct page *page)
>  static int udf_adinicb_writepage(struct page *page,
>  				 struct writeback_control *wbc)
>  {
> -	struct inode *inode = page->mapping->host;
> -	char *kaddr;
> -	struct udf_inode_info *iinfo = UDF_I(inode);
> -
>  	BUG_ON(!PageLocked(page));
> -
> -	kaddr = kmap(page);
> -	memcpy(iinfo->i_ext.i_data + iinfo->i_lenEAttr, kaddr, inode->i_size);
> -	mark_inode_dirty(inode);
> -	SetPageUptodate(page);
> -	kunmap(page);
>  	unlock_page(page);
>  
>  	return 0;
> @@ -115,6 +105,8 @@ static int udf_adinicb_write_end(struct file *file,
>  		kaddr + offset, copied);
>  	kunmap_atomic(kaddr);
>  
> +	mark_inode_dirty(inode);
> +
>  	return simple_write_end(file, mapping, pos, len, copied, page, fsdata);
>  }
>  
> -- 
> 1.7.9.5
> 
> 
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR

RE: [PATCH] udf: avoid redundant memcpy when writing data in ICB

[Chao Yu]

> -----Original Message-----
> From: Jan Kara [mailto:jack@suse.cz]
> Sent: Monday, July 14, 2014 4:46 PM
> To: Chao Yu
> Cc: jack@suse.cz; linux-kernel@vger.kernel.org
> Subject: Re: [PATCH] udf: avoid redundant memcpy when writing data in ICB
> 
> On Mon 30-06-14 13:37:15, Chao Yu wrote:
> > We have already copied the changed data of page to extent area in ICB when call
> > ->write_end() of adinicb aops, so we do not need to copy them another time in
> > ->writepage() of adinicb aops. SetPageUptodate is redundant too, so it's better
> > to remove these redundant codes.
> >
> > Also it looks more reasonable to set inode dirty after extent area are modified,
> > we'd better move mark_inode_dirty() from ->writepage() to ->write_end().
>   This doesn't look right. The copy is there because the page can be
> modified via mmap as well, not only via buffered write... So if anything,

You're right, I have made a mistake that I have not considered the 'mmap' case,
Thank you for reminding me about that. :)

> we may be able to remove the copy in udf_adinicb_write_end() (actually
> using directly simple_write_end() for that), not the one in
> udf_adinicb_write_begin().

Agreed, valid data within i_size will be copied to ICB cache when we writeback
the page by invoking udf_adinicb_writepage, so the copy in udf_adinicb_write_end
is redundant. So our patch should be modified like the following one, is this right?

---
 fs/udf/file.c | 10 ----------
 1 file changed, 10 deletions(-)

diff --git a/fs/udf/file.c b/fs/udf/file.c
index d80738f..a8dc0f1 100644
--- a/fs/udf/file.c
+++ b/fs/udf/file.c
@@ -105,16 +105,6 @@ static int udf_adinicb_write_end(struct file *file,
 			loff_t pos, unsigned len, unsigned copied,
 			struct page *page, void *fsdata)
 {
-	struct inode *inode = mapping->host;
-	unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
-	char *kaddr;
-	struct udf_inode_info *iinfo = UDF_I(inode);
-
-	kaddr = kmap_atomic(page);
-	memcpy(iinfo->i_ext.i_data + iinfo->i_lenEAttr + offset,
-		kaddr + offset, copied);
-	kunmap_atomic(kaddr);
-
 	return simple_write_end(file, mapping, pos, len, copied, page, fsdata);
 }
 
-- 
2.0.1.474.g72c7794


> 
> 								Honza
> >
> > Signed-off-by: Chao Yu <chao2.yu@samsung.com>
> > ---
> >  fs/udf/file.c |   12 ++----------
> >  1 file changed, 2 insertions(+), 10 deletions(-)
> >
> > diff --git a/fs/udf/file.c b/fs/udf/file.c
> > index d80738f..3cd0a50 100644
> > --- a/fs/udf/file.c
> > +++ b/fs/udf/file.c
> > @@ -65,17 +65,7 @@ static int udf_adinicb_readpage(struct file *file, struct page *page)
> >  static int udf_adinicb_writepage(struct page *page,
> >  				 struct writeback_control *wbc)
> >  {
> > -	struct inode *inode = page->mapping->host;
> > -	char *kaddr;
> > -	struct udf_inode_info *iinfo = UDF_I(inode);
> > -
> >  	BUG_ON(!PageLocked(page));
> > -
> > -	kaddr = kmap(page);
> > -	memcpy(iinfo->i_ext.i_data + iinfo->i_lenEAttr, kaddr, inode->i_size);
> > -	mark_inode_dirty(inode);
> > -	SetPageUptodate(page);
> > -	kunmap(page);
> >  	unlock_page(page);
> >
> >  	return 0;
> > @@ -115,6 +105,8 @@ static int udf_adinicb_write_end(struct file *file,
> >  		kaddr + offset, copied);
> >  	kunmap_atomic(kaddr);
> >
> > +	mark_inode_dirty(inode);
> > +
> >  	return simple_write_end(file, mapping, pos, len, copied, page, fsdata);
> >  }
> >
> > --
> > 1.7.9.5
> >
> >
> --
> Jan Kara <jack@suse.cz>
> SUSE Labs, CR

Re: [PATCH] udf: avoid redundant memcpy when writing data in ICB

[Jan Kara]

On Mon 14-07-14 19:14:39, Chao Yu wrote:
> > -----Original Message-----
> > From: Jan Kara [mailto:jack@suse.cz]
> > Sent: Monday, July 14, 2014 4:46 PM
> > To: Chao Yu
> > Cc: jack@suse.cz; linux-kernel@vger.kernel.org
> > Subject: Re: [PATCH] udf: avoid redundant memcpy when writing data in ICB
> > 
> > On Mon 30-06-14 13:37:15, Chao Yu wrote:
> > > We have already copied the changed data of page to extent area in ICB when call
> > > ->write_end() of adinicb aops, so we do not need to copy them another time in
> > > ->writepage() of adinicb aops. SetPageUptodate is redundant too, so it's better
> > > to remove these redundant codes.
> > >
> > > Also it looks more reasonable to set inode dirty after extent area are modified,
> > > we'd better move mark_inode_dirty() from ->writepage() to ->write_end().
> >   This doesn't look right. The copy is there because the page can be
> > modified via mmap as well, not only via buffered write... So if anything,
> 
> You're right, I have made a mistake that I have not considered the 'mmap' case,
> Thank you for reminding me about that. :)
> 
> > we may be able to remove the copy in udf_adinicb_write_end() (actually
> > using directly simple_write_end() for that), not the one in
> > udf_adinicb_write_begin().
> 
> Agreed, valid data within i_size will be copied to ICB cache when we writeback
> the page by invoking udf_adinicb_writepage, so the copy in udf_adinicb_write_end
> is redundant. So our patch should be modified like the following one, is this right?
> 
> ---
>  fs/udf/file.c | 10 ----------
>  1 file changed, 10 deletions(-)
> 
> diff --git a/fs/udf/file.c b/fs/udf/file.c
> index d80738f..a8dc0f1 100644
> --- a/fs/udf/file.c
> +++ b/fs/udf/file.c
> @@ -105,16 +105,6 @@ static int udf_adinicb_write_end(struct file *file,
>  			loff_t pos, unsigned len, unsigned copied,
>  			struct page *page, void *fsdata)
>  {
> -	struct inode *inode = mapping->host;
> -	unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
> -	char *kaddr;
> -	struct udf_inode_info *iinfo = UDF_I(inode);
> -
> -	kaddr = kmap_atomic(page);
> -	memcpy(iinfo->i_ext.i_data + iinfo->i_lenEAttr + offset,
> -		kaddr + offset, copied);
> -	kunmap_atomic(kaddr);
> -
>  	return simple_write_end(file, mapping, pos, len, copied, page, fsdata);
>  }
  You can completely remove udf_adinicb_write_end() function and just use
simple_write_end() in udf_adinicb_aops. 
 								Honza
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR