Ocfs2 devel - Jul 2010 - [PATCH 0/5 v2] Add readahead support in CoW for reflinked files.

Hi all,
	When we try to CoW some clusters for a reflinked file, we have to read 
the contents first, allocate some clusters and then map these pages to 
the new clusters. Currently, we use block_read_full_page to read it, but 
it is a little bit slower.
	So this patch set try to add readahead support for CoW. Before we 
start, we call readahead first so that the pages can be read at the very 
first. And during CoW, when we find a readahead page, we know that we 
need to move the readahead window, so a new asyncreadahead is called.

I have a small test to show how readahead speed up CoW.
readahead_test()
{
MNT_DIR=/mnt/ocfs2
DEVICE=/dev/sda8
echo 'y'|mkfs.ocfs2 --fs-features=local,refcount $DEVICE
mount -t ocfs2 $DEVICE $MNT_DIR
FILE=$MNT_DIR/$RANDOM
REFLINK=$MNT_DIR/$RANDOM
dd if=/dev/zero of=$FILE bs=1M count=1000
reflink $FILE $REFLINK
dd if=/dev/zero of=$REFLINK bs=1M count=1000 conv=notrunc
umount $MNT_DIR
}

Without these patch set, the 2nd dd has a i/o speed of 22MB/s.
with the patch set, the i/o speed is increased to about 40MB/s.

Any comments are welcomed.

Regards,
Tao

struct file * has file_ra_state to store the readahead state
and data. So pass this to ocfs2_write_begin_nolock so that
it can be used in ocfs2_refcount_cow.

Signed-off-by: Tao Ma <tao.ma at oracle.com>
---
 fs/ocfs2/aops.c |    5 +++--
 fs/ocfs2/aops.h |    3 ++-
 fs/ocfs2/mmap.c |    7 ++++---
 3 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c
index 3623ca2..13a41ef 100644
--- a/fs/ocfs2/aops.c
+++ b/fs/ocfs2/aops.c
@@ -1644,7 +1644,8 @@ static int ocfs2_expand_nonsparse_inode(struct inode
*inode, loff_t pos,
 	return ret;
 }
 
-int ocfs2_write_begin_nolock(struct address_space *mapping,
+int ocfs2_write_begin_nolock(struct file *filp,
+			     struct address_space *mapping,
 			     loff_t pos, unsigned len, unsigned flags,
 			     struct page **pagep, void **fsdata,
 			     struct buffer_head *di_bh, struct page *mmap_page)
@@ -1852,7 +1853,7 @@ static int ocfs2_write_begin(struct file *file, struct
address_space *mapping,
 	 */
 	down_write(&OCFS2_I(inode)->ip_alloc_sem);
 
-	ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
+	ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
 				       fsdata, di_bh, NULL);
 	if (ret) {
 		mlog_errno(ret);
diff --git a/fs/ocfs2/aops.h b/fs/ocfs2/aops.h
index c48e93f..7606f66 100644
--- a/fs/ocfs2/aops.h
+++ b/fs/ocfs2/aops.h
@@ -48,7 +48,8 @@ int ocfs2_write_end_nolock(struct address_space *mapping,
 			   loff_t pos, unsigned len, unsigned copied,
 			   struct page *page, void *fsdata);
 
-int ocfs2_write_begin_nolock(struct address_space *mapping,
+int ocfs2_write_begin_nolock(struct file *filp,
+			     struct address_space *mapping,
 			     loff_t pos, unsigned len, unsigned flags,
 			     struct page **pagep, void **fsdata,
 			     struct buffer_head *di_bh, struct page *mmap_page);
diff --git a/fs/ocfs2/mmap.c b/fs/ocfs2/mmap.c
index af2b8fe..b04d696 100644
--- a/fs/ocfs2/mmap.c
+++ b/fs/ocfs2/mmap.c
@@ -59,10 +59,11 @@ static int ocfs2_fault(struct vm_area_struct *area, struct
vm_fault *vmf)
 	return ret;
 }
 
-static int __ocfs2_page_mkwrite(struct inode *inode, struct buffer_head *di_bh,
+static int __ocfs2_page_mkwrite(struct file *file, struct buffer_head *di_bh,
 				struct page *page)
 {
 	int ret;
+	struct inode *inode = file->f_path.dentry->d_inode;
 	struct address_space *mapping = inode->i_mapping;
 	loff_t pos = page_offset(page);
 	unsigned int len = PAGE_CACHE_SIZE;
@@ -109,7 +110,7 @@ static int __ocfs2_page_mkwrite(struct inode *inode, struct
buffer_head *di_bh,
 	if (page->index == last_index)
 		len = size & ~PAGE_CACHE_MASK;
 
-	ret = ocfs2_write_begin_nolock(mapping, pos, len, 0, &locked_page,
+	ret = ocfs2_write_begin_nolock(file, mapping, pos, len, 0, &locked_page,
 				       &fsdata, di_bh, page);
 	if (ret) {
 		if (ret != -ENOSPC)
@@ -157,7 +158,7 @@ static int ocfs2_page_mkwrite(struct vm_area_struct *vma,
struct vm_fault *vmf)
 	 */
 	down_write(&OCFS2_I(inode)->ip_alloc_sem);
 
-	ret = __ocfs2_page_mkwrite(inode, di_bh, page);
+	ret = __ocfs2_page_mkwrite(vma->vm_file, di_bh, page);
 
 	up_write(&OCFS2_I(inode)->ip_alloc_sem);
 
-- 
1.7.1.GIT

struct file * has file_ra_state to store the readahead state
and data. So pass this to ocfs2_prepare_inode_for_write. so
that it can be used in ocfs2_refcount_cow.

Signed-off-by: Tao Ma <tao.ma at oracle.com>
---
 fs/ocfs2/file.c |    9 ++++++---
 1 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c
index 6a13ea6..d61137e 100644
--- a/fs/ocfs2/file.c
+++ b/fs/ocfs2/file.c
@@ -1864,6 +1864,7 @@ out:
 }
 
 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
+					    struct file *file,
 					    loff_t pos, size_t count,
 					    int *meta_level)
 {
@@ -1889,7 +1890,7 @@ out:
 	return ret;
 }
 
-static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
+static int ocfs2_prepare_inode_for_write(struct file *file,
 					 loff_t *ppos,
 					 size_t count,
 					 int appending,
@@ -1897,6 +1898,7 @@ static int ocfs2_prepare_inode_for_write(struct dentry
*dentry,
 					 int *has_refcount)
 {
 	int ret = 0, meta_level = 0;
+	struct dentry *dentry = file->f_path.dentry;
 	struct inode *inode = dentry->d_inode;
 	loff_t saved_pos, end;
 
@@ -1952,6 +1954,7 @@ static int ocfs2_prepare_inode_for_write(struct dentry
*dentry,
 			meta_level = -1;
 
 			ret = ocfs2_prepare_inode_for_refcount(inode,
+							       file,
 							       saved_pos,
 							       count,
 							       &meta_level);
@@ -2066,7 +2069,7 @@ relock:
 	}
 
 	can_do_direct = direct_io;
-	ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
+	ret = ocfs2_prepare_inode_for_write(file, ppos,
 					    iocb->ki_left, appending,
 					    &can_do_direct, &has_refcount);
 	if (ret < 0) {
@@ -2196,7 +2199,7 @@ static int ocfs2_splice_to_file(struct pipe_inode_info
*pipe,
 {
 	int ret;
 
-	ret = ocfs2_prepare_inode_for_write(out->f_path.dentry,	&sd->pos,
+	ret = ocfs2_prepare_inode_for_write(out, &sd->pos,
 					    sd->total_len, 0, NULL, NULL);
 	if (ret < 0) {
 		mlog_errno(ret);
-- 
1.7.1.GIT

Add a new parameter 'struct file *' to ocfs2_refcount_cow
so that we can add readahead support later.

Signed-off-by: Tao Ma <tao.ma at oracle.com>
---
 fs/ocfs2/aops.c         |    2 +-
 fs/ocfs2/file.c         |    4 ++--
 fs/ocfs2/refcounttree.c |    4 +++-
 fs/ocfs2/refcounttree.h |    3 ++-
 4 files changed, 8 insertions(+), 5 deletions(-)

diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c
index 13a41ef..968ae3d 100644
--- a/fs/ocfs2/aops.c
+++ b/fs/ocfs2/aops.c
@@ -1691,7 +1691,7 @@ int ocfs2_write_begin_nolock(struct file *filp,
 		mlog_errno(ret);
 		goto out;
 	} else if (ret == 1) {
-		ret = ocfs2_refcount_cow(inode, di_bh,
+		ret = ocfs2_refcount_cow(inode, filp, di_bh,
 					 wc->w_cpos, wc->w_clen, UINT_MAX);
 		if (ret) {
 			mlog_errno(ret);
diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c
index d61137e..ab1d0ce 100644
--- a/fs/ocfs2/file.c
+++ b/fs/ocfs2/file.c
@@ -361,7 +361,7 @@ static int ocfs2_cow_file_pos(struct inode *inode,
 	if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
 		goto out;
 
-	return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
+	return ocfs2_refcount_cow(inode, NULL, fe_bh, cpos, 1, cpos+1);
 
 out:
 	return status;
@@ -1882,7 +1882,7 @@ static int ocfs2_prepare_inode_for_refcount(struct inode
*inode,
 
 	*meta_level = 1;
 
-	ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
+	ret = ocfs2_refcount_cow(inode, file, di_bh, cpos, clusters, UINT_MAX);
 	if (ret)
 		mlog_errno(ret);
 out:
diff --git a/fs/ocfs2/refcounttree.c b/fs/ocfs2/refcounttree.c
index 4793f36..7636174 100644
--- a/fs/ocfs2/refcounttree.c
+++ b/fs/ocfs2/refcounttree.c
@@ -3398,6 +3398,7 @@ static int ocfs2_replace_cow(struct ocfs2_cow_context
*context)
  * unrefcounted extent.
  */
 static int ocfs2_refcount_cow_hunk(struct inode *inode,
+				   struct file *file,
 				   struct buffer_head *di_bh,
 				   u32 cpos, u32 write_len, u32 max_cpos)
 {
@@ -3475,6 +3476,7 @@ out:
  * clusters between cpos and cpos+write_len are safe to modify.
  */
 int ocfs2_refcount_cow(struct inode *inode,
+		       struct file *file,
 		       struct buffer_head *di_bh,
 		       u32 cpos, u32 write_len, u32 max_cpos)
 {
@@ -3494,7 +3496,7 @@ int ocfs2_refcount_cow(struct inode *inode,
 			num_clusters = write_len;
 
 		if (ext_flags & OCFS2_EXT_REFCOUNTED) {
-			ret = ocfs2_refcount_cow_hunk(inode, di_bh, cpos,
+			ret = ocfs2_refcount_cow_hunk(inode, file, di_bh, cpos,
 						      num_clusters, max_cpos);
 			if (ret) {
 				mlog_errno(ret);
diff --git a/fs/ocfs2/refcounttree.h b/fs/ocfs2/refcounttree.h
index 9983ba1..29cba0e 100644
--- a/fs/ocfs2/refcounttree.h
+++ b/fs/ocfs2/refcounttree.h
@@ -52,7 +52,8 @@ int ocfs2_prepare_refcount_change_for_del(struct inode *inode,
 					  u32 clusters,
 					  int *credits,
 					  int *ref_blocks);
-int ocfs2_refcount_cow(struct inode *inode, struct buffer_head *di_bh,
+int ocfs2_refcount_cow(struct inode *inode,
+		       struct file *filep, struct buffer_head *di_bh,
 		       u32 cpos, u32 write_len, u32 max_cpos);
 
 typedef int (ocfs2_post_refcount_func)(struct inode *inode,
-- 
1.7.1.GIT

Add a new function ocfs2_readahead_for_cow so that
we start readahead before we start our CoW.

Signed-off-by: Tao Ma <tao.ma at oracle.com>
---
 fs/ocfs2/refcounttree.c |   24 ++++++++++++++++++++++++
 1 files changed, 24 insertions(+), 0 deletions(-)

diff --git a/fs/ocfs2/refcounttree.c b/fs/ocfs2/refcounttree.c
index 7636174..03ec6ac 100644
--- a/fs/ocfs2/refcounttree.c
+++ b/fs/ocfs2/refcounttree.c
@@ -3392,6 +3392,28 @@ static int ocfs2_replace_cow(struct ocfs2_cow_context
*context)
 	return ret;
 }
 
+static void ocfs2_readahead_for_cow(struct inode *inode,
+				    struct file *file,
+				    u32 start, u32 len)
+{
+	struct address_space *mapping;
+	pgoff_t index;
+	unsigned long num_pages;
+	int cs_bits = OCFS2_SB(inode->i_sb)->s_clustersize_bits;
+
+	if (!file)
+		return;
+
+	mapping = file->f_mapping;
+	num_pages = (len << cs_bits) >> PAGE_CACHE_SHIFT;
+	if (!num_pages)
+		num_pages = 1;
+
+	index = ((loff_t)start << cs_bits) >> PAGE_CACHE_SHIFT;
+	page_cache_sync_readahead(mapping, &file->f_ra, file,
+				  index, num_pages);
+}
+
 /*
  * Starting at cpos, try to CoW write_len clusters.  Don't CoW
  * past max_cpos.  This will stop when it runs into a hole or an
@@ -3427,6 +3449,8 @@ static int ocfs2_refcount_cow_hunk(struct inode *inode,
 
 	BUG_ON(cow_len == 0);
 
+	ocfs2_readahead_for_cow(inode, file, cow_start, cow_len);
+
 	context = kzalloc(sizeof(struct ocfs2_cow_context), GFP_NOFS);
 	if (!context) {
 		ret = -ENOMEM;
-- 
1.7.1.GIT

In CoW, when we meet with a readahead page, we know
it is time to move the readahead window. So carry
out a new readahead.

Signed-off-by: Tao Ma <tao.ma at oracle.com>
---
 fs/ocfs2/refcounttree.c |   15 ++++++++++++++-
 1 files changed, 14 insertions(+), 1 deletions(-)

diff --git a/fs/ocfs2/refcounttree.c b/fs/ocfs2/refcounttree.c
index 03ec6ac..df47182 100644
--- a/fs/ocfs2/refcounttree.c
+++ b/fs/ocfs2/refcounttree.c
@@ -49,6 +49,7 @@
 
 struct ocfs2_cow_context {
 	struct inode *inode;
+	struct file *file;
 	u32 cow_start;
 	u32 cow_len;
 	struct ocfs2_extent_tree data_et;
@@ -2922,13 +2923,16 @@ static int ocfs2_duplicate_clusters_by_page(handle_t
*handle,
 	u64 new_block = ocfs2_clusters_to_blocks(sb, new_cluster);
 	struct page *page;
 	pgoff_t page_index;
-	unsigned int from, to;
+	unsigned int from, to, readahead_pages;
 	loff_t offset, end, map_end;
 	struct address_space *mapping = context->inode->i_mapping;
 
 	mlog(0, "old_cluster %u, new %u, len %u at offset %u\n",
old_cluster,
 	     new_cluster, new_len, cpos);
 
+	readahead_pages +		(ocfs2_cow_contig_clusters(sb) <<
+		 OCFS2_SB(sb)->s_clustersize_bits) >> PAGE_CACHE_SHIFT;
 	offset = ((loff_t)cpos) << OCFS2_SB(sb)->s_clustersize_bits;
 	end = offset + (new_len << OCFS2_SB(sb)->s_clustersize_bits);
 
@@ -2953,6 +2957,14 @@ static int ocfs2_duplicate_clusters_by_page(handle_t
*handle,
 		if (PAGE_CACHE_SIZE <= OCFS2_SB(sb)->s_clustersize)
 			BUG_ON(PageDirty(page));
 
+		if (PageReadahead(page) && context->file) {
+			page_cache_async_readahead(mapping,
+						   &context->file->f_ra,
+						   context->file,
+						   page, page_index,
+						   readahead_pages);
+		}
+
 		if (!PageUptodate(page)) {
 			ret = block_read_full_page(page, ocfs2_get_block);
 			if (ret) {
@@ -3472,6 +3484,7 @@ static int ocfs2_refcount_cow_hunk(struct inode *inode,
 	context->ref_root_bh = ref_root_bh;
 	context->cow_duplicate_clusters = ocfs2_duplicate_clusters_by_page;
 	context->get_clusters = ocfs2_di_get_clusters;
+	context->file = file;
 
 	ocfs2_init_dinode_extent_tree(&context->data_et,
 				      INODE_CACHE(inode), di_bh);
-- 
1.7.1.GIT

On Wed, Jul 14, 2010 at 03:52:26PM +0800, Tao Ma wrote:
> 	When we try to CoW some clusters for a reflinked file, we have to
> read the contents first, allocate some clusters and then map these
> pages to the new clusters. Currently, we use block_read_full_page to
> read it, but it is a little bit slower.
> 	So this patch set try to add readahead support for CoW. Before we
> start, we call readahead first so that the pages can be read at the
> very first. And during CoW, when we find a readahead page, we know
> that we need to move the readahead window, so a new asyncreadahead
> is called.
	I think this looks good.  Can you create a branch in your git
tree for it?  I'll pull when I build merge-window.

Joel

-- 

"Conservative, n.  A statesman who is enamoured of existing evils,
 as distinguished from the Liberal, who wishes to replace them
 with others."
	- Ambrose Bierce, The Devil's Dictionary

Joel Becker
Consulting Software Developer
Oracle
E-mail: joel.becker at oracle.com
Phone: (650) 506-8127