On Tue, 2009-08-25 at 16:01 +0200, Jan Kara wrote:

On Mon 24-08-09 14:38:13, Mingming wrote:

quoted

On Thu, 2009-08-20 at 13:52 +0200, Jan Kara wrote:

quoted

On Wed 19-08-09 14:26:16, Mingming wrote:

quoted

On Wed, 2009-08-19 at 16:15 +0200, Jan Kara wrote:

quoted

quoted

For direct IO write to the end of file, we now also could get rid of
using orphan list to protect expose stale data out after crash, when
direct write to end of file isn't complete. We now fallocate blocks for
the direct IO write to the end of file as well, and convert those
fallocated blocks at the end of file to written when IO is complete. If
fs crashed before the IO complete, it will only seen the file tail has
been fallocated but won't get the stale data out.

  But you still probably need orphan list to truncate blocks allocated
beyond file end during extending direct write. So I'd just remove this
paragraph...

Do we still need to truncate blocks allocated at the end of file? Those
blocks are marked as uninitialized extents (as any block allocation from
DIO are flagged as uninitialized extents, before IO is complete), so
that after recover from crash, the stale data won't get exposed. 

I guess I missed the cases you concerned that we need the orphan list to
protect, could you plain a little more?

  Well, you are right it's not a security problem since no data gets
exposed. It's just that after a crash, there will be blocks allocated to
a file and it's not trivial to it find out unless you specifically stat the
file. So it's just *not nice* kind of thing. If it brings us some
advantage to not put the inode on the orphan list, then sure it might be
a tradeoff we are willing to make. But otherwise I see no point.

Hmm... I see what you concerned now...user probably don't like allocated
blocks at the end...

  Yes, it's kind of counterintuitive that blocks can get allocated but
don't really "show up" in the file (because they are unwritten and beyond
i_size).

quoted

I am trying to think of the ext3 case. In ext3 case, inode will be
removed from orphan list once the IO is complete, but that is before the
i_disksize get updated and journal handle being closed. What if the
system crash after inode removed from orphan list but before the
i_disksize get updated?

  As you write below, until we stop the handle, the transaction cannot be
committed and so no change actually gets written to disk.

Then in case this, if no change actually gets written to disk, then
i_disksize hasn't get updated on disk, why we need the orphan list?

quoted

But since the journal handled has not marked as stopped, even without
putting the inode on the orpah list, wouldn't the open journal handle
and delayed i_disksize update until IO complete time, prevent we see
half DIO data on disk? Though blocks are allocated to the file, but
those block mapping wouldn't show up on disk, no? Did I miss anything?

In the ext4 +patch case (non AIO case), we also close the handle when
end_io completed. If the system crashed we would see the blocks are
allocated but marked as unwritten.

  Exactly, you end up with allocated and unwritten blocks beyond i_size and
that's what I'd like to avoid if user did not explicitely fallocate() these
blocks.

Yes but same as ext3, nothing should write to disk until we stop the
handle, so the unwritten blocks flag should also been seen as the
i_disksize has not been updated on disk yet.

quoted

quoted

quoted

quoted

quoted

1) Block allocation needed for DIO write are fallocated, including holes
and file tail write, marked as unwritten extents after block allocation.

2) those unwritten extents, and fallocate extents, will be converted to
written extents (and update disk size when write to end of file)when the
IO is complete. The conversion is triggered using end_io call back
function passing from ext4 fs to direct IO.

3) For already fallocated extent, at the time try to map the fallocated
extent, we will split the fallocated extent as necessary, mark the
to-write fallocated extent mapped but still remains unwritten,
insert the splitted extents, to prevent ENOSPC later.

This first patch does 1) and 2), the second patch does 3)

Patch against ext4 patch queue.

Comments? 

Singed-Off-By: Mingming Cao [off-list ref]
---
 fs/ext4/ext4.h  |   18 ++++
 fs/ext4/inode.c |  211 +++++++++++++++++++++++++++++++++++++++++++++++++++++++-
 fs/ext4/super.c |   11 ++
 3 files changed, 237 insertions(+), 3 deletions(-)

Index: linux-2.6.31-rc4/fs/ext4/ext4.h
===================================================================

--- linux-2.6.31-rc4.orig/fs/ext4/ext4.h	2009-08-09 14:46:10.000000000 -0700
+++ linux-2.6.31-rc4/fs/ext4/ext4.h	2009-08-09 23:13:15.000000000 -0700

@@ -111,6 +111,15 @@ struct ext4_allocation_request {
 	unsigned int flags;
 };
 
+typedef struct ext4_io_end{
+	struct inode		*inode;		/* file being written to */
+	unsigned int		type;		/* unwritten or written */
+	int			error;		/* I/O error code */
+	ext4_lblk_t		offset;		/* offset in the file */
+	size_t			size;		/* size of the extent */
+	struct work_struct	work;		/* data work queue */
+}ext4_io_end_t;
+
 /*
  * Special inodes numbers
  */

@@ -330,8 +339,8 @@ struct ext4_new_group_data {
 	/* Call ext4_da_update_reserve_space() after successfully 
 	   allocating the blocks */
 #define EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE	0x0008
-
-
+#define EXT4_GET_BLOCKS_DIO_CREATE_EXT		0x0011
+#define EXT4_GET_BLOCKS_DIO_CONVERT_EXT		0x0021
 /*
  * ioctl commands
  */

@@ -960,6 +969,9 @@ struct ext4_sb_info {
 
 	unsigned int s_log_groups_per_flex;
 	struct flex_groups *s_flex_groups;
+
+	/* workqueue for dio unwritten */
+	struct workqueue_struct *dio_unwritten_wq;
 };
 
 static inline struct ext4_sb_info *EXT4_SB(struct super_block *sb)

@@ -1650,6 +1662,8 @@ extern void ext4_ext_init(struct super_b
 extern void ext4_ext_release(struct super_block *);
 extern long ext4_fallocate(struct inode *inode, int mode, loff_t offset,
 			  loff_t len);
+extern int ext4_convert_unwritten_extents(struct inode *inode, loff_t offset,
+			  loff_t len);
 extern int ext4_get_blocks(handle_t *handle, struct inode *inode,
 			   sector_t block, unsigned int max_blocks,
 			   struct buffer_head *bh, int flags);

Index: linux-2.6.31-rc4/fs/ext4/super.c
===================================================================

--- linux-2.6.31-rc4.orig/fs/ext4/super.c	2009-08-09 14:51:08.000000000 -0700
+++ linux-2.6.31-rc4/fs/ext4/super.c	2009-08-09 14:51:14.000000000 -0700

@@ -578,6 +578,9 @@ static void ext4_put_super(struct super_
 	struct ext4_super_block *es = sbi->s_es;
 	int i, err;
 
+	flush_workqueue(sbi->dio_unwritten_wq);
+	destroy_workqueue(sbi->dio_unwritten_wq);
+
 	lock_super(sb);
 	lock_kernel();
 	if (sb->s_dirt)

@@ -2781,6 +2784,12 @@ no_journal:
 			clear_opt(sbi->s_mount_opt, NOBH);
 		}
 	}
+	EXT4_SB(sb)->dio_unwritten_wq = create_workqueue("ext4-dio-unwritten");
+	if (!EXT4_SB(sb)->dio_unwritten_wq) {
+		printk(KERN_ERR "EXT4-fs: failed to create DIO workqueue\n");
+		goto failed_mount_wq;
+	}
+
 	/*
 	 * The jbd2_journal_load will have done any necessary log recovery,
 	 * so we can safely mount the rest of the filesystem now.

@@ -2893,6 +2902,8 @@ cantfind_ext4:
 
 failed_mount4:
 	ext4_msg(sb, KERN_ERR, "mount failed");
+	destroy_workqueue(EXT4_SB(sb)->dio_unwritten_wq);
+failed_mount_wq:
 	ext4_release_system_zone(sb);
 	if (sbi->s_journal) {
 		jbd2_journal_destroy(sbi->s_journal);

Index: linux-2.6.31-rc4/fs/ext4/inode.c
===================================================================

--- linux-2.6.31-rc4.orig/fs/ext4/inode.c	2009-08-09 14:46:32.000000000 -0700
+++ linux-2.6.31-rc4/fs/ext4/inode.c	2009-08-09 14:56:40.000000000 -0700

@@ -37,6 +37,7 @@
 #include <linux/namei.h>
 #include <linux/uio.h>
 #include <linux/bio.h>
+#include <linux/workqueue.h>
 
 #include "ext4_jbd2.h"
 #include "xattr.h"

@@ -3279,6 +3280,8 @@ static int ext4_releasepage(struct page 
 }
 
 /*
+ * O_DIRECT for ext3 (or indirect map) based files
+ *
  * If the O_DIRECT write will extend the file then add this inode to the
  * orphan list.  So recovery will truncate it back to the original size
  * if the machine crashes during the write.

@@ -3287,7 +3290,7 @@ static int ext4_releasepage(struct page 
  * crashes then stale disk data _may_ be exposed inside the file. But current
  * VFS code falls back into buffered path in that case so we are safe.
  */
-static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
+static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
 			      const struct iovec *iov, loff_t offset,
 			      unsigned long nr_segs)
 {

@@ -3361,6 +3364,212 @@ out:
 	return ret;
 }
 
+struct workqueue_struct *ext4_unwritten_queue;
+
+/* Maximum number of blocks we map for direct IO at once. */
+
+static int ext4_get_block_dio_write(struct inode *inode, sector_t iblock,
+		   struct buffer_head *bh_result, int create)
+{
+	handle_t *handle = NULL;
+	int ret = 0;
+	unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
+	int dio_credits;
+
+	/*
+	 * DIO VFS code passes create = 0 flag for write to
+	 * the middle of file. It does this to avoid block
+	 * allocation for holes, to prevent expose stale data
+	 * out when there is parallel buffered read (which does
+	 * not hold the i_mutex lock) while direct IO write has
+	 * not completed. DIO request on holes finally falls back
+	 * to buffered IO for this reason.
+	 *
+	 * For ext4 extent based file, since we support fallocate,
+	 * new allocated extent as uninitialized, for holes, we
+	 * could fallocate blocks for holes, thus parallel
+	 * buffered IO read will zero out the page when read on
+	 * a hole while parallel DIO write to the hole has not completed.
+	 *
+	 * when we come here, we know it's a direct IO write,
+	 * so it's safe to override the create flag from VFS.
+	 */
+	create = EXT4_GET_BLOCKS_DIO_CREATE_EXT;
+
+	if (max_blocks > DIO_MAX_BLOCKS)
+		max_blocks = DIO_MAX_BLOCKS;
+	dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
+	handle = ext4_journal_start(inode, dio_credits);
+	if (IS_ERR(handle)) {
+		ret = PTR_ERR(handle);
+		goto out;
+	}
+	ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
+			      create);
+	if (ret > 0) {
+		bh_result->b_size = (ret << inode->i_blkbits);
+		ret = 0;
+	}
+	ext4_journal_stop(handle);
+out:
+	return ret;
+}
+
+static int ext4_get_block_dio_read(struct inode *inode, sector_t iblock,
+		   struct buffer_head *bh_result, int create)
+{
+	int ret = 0;
+	unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
+	handle_t *handle = NULL;
+
+	ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
+			      create);
+	if (ret > 0) {
+		bh_result->b_size = (ret << inode->i_blkbits);
+		ret = 0;
+	}
+	return ret;
+}

  Huh, what's the purpose of the above function? We can use normal
get_block, can't we?

This is pretty much a wrapper of this normal get_block function, except
here we need to store the space mapped in b_size, DIO will check that.

  But ext4_get_block() will do exactly the same, won't it?

Ah, it does the same, with a little more check for "read" or "write"...I
am fine to use ext4_get_block() directly. Will change it as you
suggested.

quoted

quoted

quoted

quoted

+
+
+#define		DIO_UNWRITTEN	0x1
+
+/*
+ * IO write completion for unwritten extents.
+ *
+ * check a range of space and convert unwritten extents to written.
+ */
+static void ext4_end_dio_unwritten(struct work_struct *work)
+{
+	ext4_io_end_t *io = container_of(work, ext4_io_end_t, work);
+	struct inode *inode = io->inode;
+	loff_t offset = io->offset;
+	size_t size = io->size;
+	int ret = 0;
+
+	ret = ext4_convert_unwritten_extents(inode, offset, size);
+
+	if (ret < 0)
+		printk(KERN_EMERG "%s: failed to convert unwritten"
+			"extents to written extents, error is %d\n",
+                       __func__, ret);
+	kfree(io);
+}

  Looking at ext4_convert_unwritten_extents(), you definitely miss some
locking. Since this is called completely asynchronously, you have to
protect against racing truncates and basically anything can happen with
the inode in the mean time. 

extents tree update is protected by the i_data_sem which will be hold at
the ext4_get_blocks() called from ext4_convert_unwritten_extents. 

  Ah, I've missed that.

quoted

perhaps should grab the i_mutex() which protects the inode update?

  Probably we should because we update i_disksize inside
ext4_convert_unwritten_extents(). BTW, where do we update i_size? The same
place where we update i_disksize is probably right and that definitely
needs i_mutex.

We do update i_size in generic_file_direct_write(), there I think proper
locking (i_mutex lock) is hold.

For i_disksize update, with this patch, we call ext4_update_i_disksize()
from ext4_convert_unwritten_extents(), at the end of IO is completed.
ext4_update_i_disksize() grab the i_data_sem to prevent race with
truncate.

Now I think we are fine with race with truncate... no?

  I don't think we are fine - i_disksize gets updated without i_data_sem
being held in ext4_setattr() so you can race with it. I'm not sure how
exactly locking rules for ext4 look like wrt. i_disksize update but either
your code or ext4_setattr() need to be changed.

then ext4_setattr() race with other places using
ext4_update_i_disksize() (e.g. fallocate) already. we need to fix
ext4_setattr and probably other places where i_disksize is updated
without holding i_data_sem.

  Actually, probably you should hold i_mutex until the io is finished -
otherwise a truncate against the file can be started before the io is
finished and actually nothing prevents it from finishing before your end_io
callback gets processed. Then the extent conversion will get confused I
expect because it won't find extents it should convert.

DIO already did that:) i_mutex is hold during the whole direct IO,
before return back to user, in the non AIO case.

quoted

quoted

quoted

quoted

It needn't be cached in the memory anymore!

Right, we probably need to increase the reference to the inode before
submit the IO, so the inode would not be push out of cache before IO
completed.

  Yes, that's possible but then you have to be prepared to handle deletes
of an inode from your worker thread because it can drop the last reference
to an already deleted inode.

quoted

quoted

  Also fsync() has to flush all the updates for the inode it has in the
workqueue... Ditto for ext4_sync_fs().

I think we discussed about this before, and it seems there is no clear
defination the DIO forces metadata update sync to disk before returns
back to user apps... If we do force this in ext4 &DIO, doing fsync() on
every DIO call is expensive.

  No, I meant something else:
fd = open("file", O_DIRECT | O_RDWR);
write(fd, buf, size);
fsync(fd)

  After fsync(fd) we *must* have everything on disk and reachable even if
we crashed just after fsync(). So conversion of extents from uninitialized
to initialized must happen during fsync() and it does not so far because we
have no connection from an inode to the work queued to the worker thread.

Are you worried about AIO DIO case?

Because for non AIO case, when DIO returns back to user, the data IO has
already completed, and the conversion of extents from uninitialized to
initialized has already being kicked by flush_workqueue(). Since the
conversion is being journaled, the aftweward fsync() should able to
commit all transactions, thus force the conversion to be complete on
disk after fsync.

  Yes, you are right, this case is fine.

quoted

For the AIO DIO case, yeah I agree there might be a issue....  Hmm.. if
we do fsync() after AIO DIO, currently I couldn't see any way fsync()
could ensure all the pending data IOs from AIO DIO path would reach to
disk...

if there is a way fsycn() could wait for all pending data IOs  from AIO
DIO to complete, then end_io callback would able to trigger the
conversion, and fsycn() would able to flush the metedata update hit to
disk.

  Well, thinking about this again, fsync() does not have to wait for
pending AIO - we never guaranteed anything about what fsync() does to such
requests. But once you complete the AIO (and thus userspace can know that
it is done), you have to make sure that fsync() flushes all the work
queued in the workqueue connected with this AIO.

I just realize, why don't we flush the work in the workqueue with this
AIO at the IO completion time (but not ext4_direct_IO time)? Since the
work just does conversion but not the full journal commit. That would
work if there is fsync() after the AIO is completed.

quoted

quoted

quoted

quoted

quoted

+
+static ext4_io_end_t *ext4_init_io_end (struct inode *inode, unsigned int type)
+{
+	ext4_io_end_t *io = NULL;
+
+	io = kmalloc(sizeof(*io), GFP_NOFS);
+
+	if (io) {
+		io->inode = inode;
+		io->type = type;
+		io->offset = 0;
+		io->size = 0;
+		io->error = 0;
+		INIT_WORK(&io->work, ext4_end_dio_unwritten);
+	}
+
+	return io;
+}
+
+static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
+			    ssize_t size, void *private)
+{
+        ext4_io_end_t *io_end = iocb->private;
+	struct workqueue_struct *wq;
+
+	/* if not hole or unwritten extents, just simple return */
+	if (!io_end || !size)
+		return;
+	io_end->offset = offset;
+	io_end->size = size;
+	wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
+
+	/* We need to convert unwritten extents to written */
+	queue_work(wq, &io_end->work);
+
+        if (is_sync_kiocb(iocb))
+		flush_workqueue(wq);

  I don't think you can flush_workqueue here. end_io is called from
interrupt context and flush_workqueue blocks for a long time...
The wait should be done in ext4_direct_IO IMHO...

Okay, I will move it to ext4_direct_IO(), hmm..I think that is fine with
AIO path as the is_sync_kiocb(iocb) will avoid that.

  Yes, this concerns standard IO.

quoted

If flush workqueue just kick off the work on the workqueue but not wait
for it to complete (no fsync()), would that still be a big concern? BTW,

  Waking up the worker thread is certainly fine even from the end_io.

Okay, I guess I didn't explain this well with my patch. The
flush_workqueue() here will call ext4_convert_unwritten_extents(), which
just does the conversion, but ext4_convert_unwritten_extents() does not
force the journal commit.

quoted

quoted

the flush workqueue only called when file is opened with sync mode.

  I don't think so - is_sync_kiocb() just means that it is a standard
read/write and not one submitted via the aio interface (io_submit syscall
etc.).

ah okay, I responsed too soon. yeah, for the DIO regular case (even if
file doesn't open with sync mode), we do ensure the worker thread kick
off the conversion and log the conversion. That's right behavior, I
think.

  Yes, this part of your patch is correct. Ted explained to me what I have
missed.

quoted

quoted

quoted

quoted

quoted

+
+	iocb->private = NULL;
+}
+/*
+ * For ext4 extent files, ext4 will do direct-io write to holes,
+ * preallocated extents, and those write extend the file, no need to
+ * fall back to buffered IO.
+ *
+ * If there is block allocation needed(holes, EOF write), we fallocate
+ * those blocks, mark them as unintialized
+ * If those blocks were preallocated, we mark sure they are splited, but
+ * still keep the range to write as unintialized.
+ *
+ * When end_io call back function called at the last IO complete time,
+ * those extents will be converted to written extents.
+ *
+ */
+static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
+			      const struct iovec *iov, loff_t offset,
+			      unsigned long nr_segs)
+{
+	struct file *file = iocb->ki_filp;
+	struct inode *inode = file->f_mapping->host;
+	ssize_t ret;
+
+	if (rw == WRITE) {
+		/*
+ 		 * For DIO we fallocate blocks for holes and end of file
+ 		 * write. Those fallocated extents are marked as uninitialized
+ 		 * to prevent paralel buffered read to expose the stale data
+ 		 * before DIO complete the data IO.
+ 		 * as for previously fallocated extents, ext4 get_block
+ 		 * will just simply mark the buffer mapped but still
+ 		 * keep the extents uninitialized.
+ 		 *
+ 		 * At the end of IO, the ext4 end_io callback function
+ 		 * will convert those unwritten extents to written,
+ 		 * and update on disk file size if the DIO expands the file.
+ 		 *
+ 		 */
+		iocb->private = ext4_init_io_end(inode, DIO_UNWRITTEN);
+		if (!iocb->private)
+			return -ENOMEM;
+
+		ret = blockdev_direct_IO(rw, iocb, inode,
+					 inode->i_sb->s_bdev, iov,
+					 offset, nr_segs,
+					 ext4_get_block_dio_write,
+					 ext4_end_io_dio);
+	} else
+		ret = blockdev_direct_IO(rw, iocb, inode,
+					 inode->i_sb->s_bdev, iov,
+					 offset, nr_segs,
+					 ext4_get_block_dio_read, NULL);
+
+ 	/*
+	 * In the case of AIO DIO, VFS dio submitted the IO, but it
+ 	 * does not wait for io complete. To prevent expose stale
+ 	 * data after crash before IO complete,
+ 	 * i_disksize needs to be updated at the
+ 	 * time all the IO is completed, not here
+ 	 */

  Yeah, but so far at least ext3_direct_IO and ext4_ind_direct_IO
happily update i_size here and noone reported the problem (yet) ;).  The
thing is that the current transaction has to commit for stale data to be
visible and that sends a barrier which also forces blocks written by
direct IO to the persistent storage. So at least if the underlying
storage supports barriers, we are fine. If it does not, it could maybe
reorder direct IO writes after the journal commit (it need not have
reported the direct IO as done so far) and after a crash we would have
a problem...
  It's just that I'm not sure that all the trouble with end_io and
workqueues is worth it... More code = more bugs ;)

the end_io is there for direct write to fallocate. I am completely sure
if there is better way? We need to convert the extents to written at the
end of IO is completed, end_io call back seems works for this purpose.

The workqueue is for AIO case,it would be nice if we could do without it
for AIO. But read comments from xfs code it seems this is needed if we
use end_io call back.

  Yes, end_io callback is useful for triggering the conversion or
acknowledging that the transaction with the conversion can be committed
and I'm not against using it. Actually, looking at the XFS code, they also
do flush_workqueue() from the end_io callback. Interesting. I guess I'll
ask them why it's not a problem...
  What I'm concerned about is the fact that we'd do the conversion from
completely asynchronous context and that opens all sorts of races. 

Yeah in the AIO case we do the conversion  from asynchronous context,
but for the non AIO case, the conversion is done under process context
still. I suspect the AIO DIO today for ext3 update the on disk size
before IO is complete is a bigger issue...:)

  Well, ext3 has a problem that if you do AIO DIO and power fails after
the transaction is committed but before the data really gets written, then
we expose stale data after a crash. I didn't see a single report of this
but in theory the race is there.
  Your patch closes this hole but we should better not introduce other
races :).

quoted

quoted

So I'd
find simpler to do the conversion from ext4_direct_IO and just make sure the
transaction is not committed before the IO is done (essentially what we
currently do with ordered buffers).

True, that might workable....I am concerned this won't work for no
journal mode, though this mode is not commonly used now.:)

  Well, if you are running without a journal, you can see stale data after
a crash - there's no way around it. So just not waiting for anything works
well.

But in the direct IO case, with the end_io call back, we won't update
i_disksize until IO is completed, you should not see stale data after a
crash, even without journal, am I miss anything?

Mingming

									Honza