Skip to contentSkip to conversation

2.6.21-rc7 new aops patchset

DISCUSSION Nick Piggin 2 participants 3 messages Apr 21, 2007 – Apr 24, 2007

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

http://www.kernel.org/pub/linux/kernel/people/npiggin/patches/new-aops/

2.6.21-rc7-new-aops*

New aops patchset against 2.6.21-rc7. I was going to rebase against -mm and submit the patchset, but I ended up doing quite a lot of new filesystem conversions and fixes so I stay with mainline so as to not introduce so much churn.

The plan is to next rebase against rc6-mm1 and submit all 47 patches to linux-fsdevel and Andrew Morton to be merged. OK? Should each fs conversion also go to the filesystem specific maintainer/list?

I have done some testing, but I would like testing and revuew abd acks from filesystem maintainers. Lots of filesystems I can't or don't know how to mount and test. Well that is a copout, but it is a hassle for me to do so many when presumably maintainers will have a test setup ready to go. And I have tried to minimise the load on maintainers by doing most conversions myself.

cc'ed reiserfs lists: could somebody volunteer to do the resierfs conversion? It's too complex for me to do a reasonable job, and reiserfs prepare_write also does weird things with generic_cont_expand which is hard to follow (this could be made much simpler by using the AOP_FLAG_CONT_EXPAND I added there -- can you also look at using that, or let me know why it doesn't work).

Changes:
- Cut perform_write batch AOP out of this series. It will be back, in a future discussion. Kept the iov_iter stuff because it's nice anyway.

- Added patch to inject short or zero length pagecache writes to better exercise filesystems. Another one to poision new and !uptodate buffers.

- Kept the old ocfs2 patch from Mark, so I don't have to suck up git-ocfs2.patch which keeps patch size down. I'll use Mark's new patch when rebasing to -mm.

- ext4 patch from Badari.

- Fixed ext3 data=journal (and the copied code in ext4).

- Converted cifs + fixed information leak.

- Converted udf seem to have fixed pagecache corruption(?) -- page was marked uptodate when it is not. Also, fixed the silly setup where prepare_write was doing a kmap to be used in commit_write: just do kmap_atomic in write_end.

- Fixed some earlier problems with my ext2 multi-page directory handling.

- Had a crack at converting all the others. reiserfs is the only one left unconverted!

- Tested (under UML, loopback, PAGE_SIZE=4096) with fsx-linux the following:
ext2 (bs=1024,4096)
ext3 (bs=1024,4096; data=writeback,ordered,journal)
minix (v2,3)
xfs (bs=1024,4096)
fuse (with ssh,passthough fs -- note fsx-linux fails as expected) jfs
reiserfs (bs=1024,4096) (unconverted)
vfat (bs=1024,4096)
tmpfs
rd (with ext2)
hostfs (it now survives much longer than with the old code!) nfs

Nick

maintainer
#1 lore raw

Nick,

Thanks for converting fuse, and testing. Here's a minor update to fs-fuse-aops.patch.

Miklos

----
Convert fuse to new aops.

[mszeredi]
- don't send zero length write requests
- it is not legal for the filesystem to return with zero written bytes

fs/fuse/file.c +33 -15
Index: linux/fs/fuse/file.c
===================================================================
@@ -443,22 +443,25 @@ static size_t fuse_send_write(struct fus
 	return outarg.size;
 }
 
-static int fuse_prepare_write(struct file *file, struct page *page,
-			      unsigned offset, unsigned to)
-{
-	/* No op */
+static int fuse_write_begin(struct file *file, struct address_space *mapping,
+			loff_t pos, unsigned len, unsigned flags,
+			struct page **pagep, void **fsdata)
+{
+	pgoff_t index = pos >> PAGE_CACHE_SHIFT;
+
+	*pagep = __grab_cache_page(mapping, index);
+	if (!*pagep)
+		return -ENOMEM;
 	return 0;
 }
 
-static int fuse_commit_write(struct file *file, struct page *page,
-			     unsigned offset, unsigned to)
+static int fuse_buffered_write(struct file *file, struct inode *inode,
+			       loff_t pos, unsigned count, struct page *page)
 {
 	int err;
 	size_t nres;
-	unsigned count = to - offset;
-	struct inode *inode = page->mapping->host;
 	struct fuse_conn *fc = get_fuse_conn(inode);
-	loff_t pos = page_offset(page) + offset;
+	unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
 	struct fuse_req *req;
 
 	if (is_bad_inode(inode))
@@ -474,20 +477,35 @@ static int fuse_commit_write(struct file
 	nres = fuse_send_write(req, file, inode, pos, count);
 	err = req->out.h.error;
 	fuse_put_request(fc, req);
-	if (!err && nres != count)
+	if (!err && !nres)
 		err = -EIO;
 	if (!err) {
-		pos += count;
+		pos += nres;
 		spin_lock(&fc->lock);
 		if (pos > inode->i_size)
 			i_size_write(inode, pos);
 		spin_unlock(&fc->lock);
 
-		if (offset == 0 && to == PAGE_CACHE_SIZE)
+		if (count == PAGE_CACHE_SIZE)
 			SetPageUptodate(page);
 	}
 	fuse_invalidate_attr(inode);
-	return err;
+	return err ? err : nres;
+}
+
+static int fuse_write_end(struct file *file, struct address_space *mapping,
+			loff_t pos, unsigned len, unsigned copied,
+			struct page *page, void *fsdata)
+{
+	struct inode *inode = mapping->host;
+	int res = 0;
+
+	if (copied)
+		res = fuse_buffered_write(file, inode, pos, copied, page);
+
+	unlock_page(page);
+	page_cache_release(page);
+	return res;
 }
 
 static void fuse_release_user_pages(struct fuse_req *req, int write)
@@ -817,8 +835,8 @@ static const struct file_operations fuse
 
 static const struct address_space_operations fuse_file_aops  = {
 	.readpage	= fuse_readpage,
-	.prepare_write	= fuse_prepare_write,
-	.commit_write	= fuse_commit_write,
+	.write_begin	= fuse_write_begin,
+	.write_end	= fuse_write_end,
 	.readpages	= fuse_readpages,
 	.set_page_dirty	= fuse_set_page_dirty,
 	.bmap		= fuse_bmap,
author Applied
#2 lore raw
Quoted 10 lines · Miklos Szeredi

Nick,

Thanks for converting fuse, and testing. Here's a minor update to fs-fuse-aops.patch.

Miklos

----
Convert fuse to new aops.

Quoted 6 lines

[mszeredi]
- don't send zero length write requests
- it is not legal for the filesystem to return with zero written bytes

Thanks, applied.