mbox series

[0/9] ext4: Stop using ext4_writepage() for writeout of ordered data

Message ID 20221130162435.2324-1-jack@suse.cz
Headers show
Series ext4: Stop using ext4_writepage() for writeout of ordered data | expand

Message

Jan Kara Nov. 30, 2022, 4:35 p.m. UTC 
Hello,

this patch series modifies ext4 so that we stop using ext4_writepage() for
writeout of ordered data during transaction commit (through
generic_writepages() from jbd2_journal_submit_inode_data_buffers()). Instead we
directly call ext4_writepages() from the
ext4_journal_submit_inode_data_buffers(). This is part of Christoph's effort
to get rid of the .writepage() callback in all filesystems.

Note that we still keep ext4_writepage() function for writeout of journalled
data. That can be also dealt with but let's untangle things one by one.  Also
note that jbd2 still uses generic_writepages() in its
jbd2_journal_submit_inode_data_buffers() helper because it is still used from
OCFS2. Again, something to be dealt with in a separate patchset.

Patches pass for me tests on both 1k and 4k filesystems with dioreadlock and
without it.

								Honza

Comments

Ritesh Harjani (IBM) Dec. 1, 2022, 11:42 a.m. UTC | #1 
On 22/11/30 05:35PM, Jan Kara wrote:
> Hello,
>
> this patch series modifies ext4 so that we stop using ext4_writepage() for
> writeout of ordered data during transaction commit (through
> generic_writepages() from jbd2_journal_submit_inode_data_buffers()). Instead we
> directly call ext4_writepages() from the
> ext4_journal_submit_inode_data_buffers().

Hello Jan,

Do you think we should add a WARN_ON_ONCE() or something in
ext4_do_writepages() function where we might try to start a transaction
at [J]. Since we can now enter into ext4_do_writepages() from two places:
1. writeback
2. jbd2_journal_commit_transaction()
	mpage_submit_page
	mpage_prepare_extent_to_map
	ext4_do_writepages
	ext4_normal_submit_inode_data_buffers
	ext4_journal_submit_inode_data_buffers
	journal_submit_data_buffers
	jbd2_journal_commit_transaction
	kjournald2

So IIUC, we will call mpage_submit_page() in the first call to
mpage_prepare_extent_to_map() [1] itself. That may set mpd->scanned_until_end = 1
at the end of it. So then we should never enter into the while loop where we
start a journal txn.

But can it ever happen that we ever into this while loop for writing out
pages for journal_submit_data_buffers()?

ext4_do_writepages()
{
<...>
	mpd->do_map = 0;
	mpd->scanned_until_end = 0;
	mpd->io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
<...>
[1]	ret = mpage_prepare_extent_to_map(mpd);
	/* Unlock pages we didn't use */
	mpage_release_unused_pages(mpd, false);
	/* Submit prepared bio */
	ext4_io_submit(&mpd->io_submit);
	ext4_put_io_end_defer(mpd->io_submit.io_end);
	mpd->io_submit.io_end = NULL;
<...>
	while (!mpd->scanned_until_end && wbc->nr_to_write > 0) {
		/* For each extent of pages we use new io_end */
		mpd->io_submit.io_end = ext4_init_io_end(inode, GFP_KERNEL);
<....>
		BUG_ON(ext4_should_journal_data(inode));
		needed_blocks = ext4_da_writepages_trans_blocks(inode);

		/* start a new transaction */
[J]		handle = ext4_journal_start_with_reserve(inode,
				EXT4_HT_WRITE_PAGE, needed_blocks, rsv_blocks);
<...>
		mpd->do_map = 1;

		trace_ext4_da_write_pages(inode, mpd->first_page, wbc);
[2]		ret = mpage_prepare_extent_to_map(mpd);
		if (!ret && mpd->map.m_len)
			ret = mpage_map_and_submit_extent(handle, mpd,
					&give_up_on_write);
<...>
	}
<...>
}

-ritesh
Jan Kara Dec. 1, 2022, 11:55 a.m. UTC | #2 
On Thu 01-12-22 17:12:05, Ritesh Harjani (IBM) wrote:
> On 22/11/30 05:35PM, Jan Kara wrote:
> > Hello,
> >
> > this patch series modifies ext4 so that we stop using ext4_writepage() for
> > writeout of ordered data during transaction commit (through
> > generic_writepages() from jbd2_journal_submit_inode_data_buffers()). Instead we
> > directly call ext4_writepages() from the
> > ext4_journal_submit_inode_data_buffers().
> 
> Hello Jan,
> 
> Do you think we should add a WARN_ON_ONCE() or something in
> ext4_do_writepages() function where we might try to start a transaction
> at [J]. Since we can now enter into ext4_do_writepages() from two places:
> 1. writeback
> 2. jbd2_journal_commit_transaction()
> 	mpage_submit_page
> 	mpage_prepare_extent_to_map
> 	ext4_do_writepages
> 	ext4_normal_submit_inode_data_buffers
> 	ext4_journal_submit_inode_data_buffers
> 	journal_submit_data_buffers
> 	jbd2_journal_commit_transaction
> 	kjournald2
> 
> So IIUC, we will call mpage_submit_page() in the first call to
> mpage_prepare_extent_to_map() [1] itself. That may set mpd->scanned_until_end = 1
> at the end of it. So then we should never enter into the while loop where we
> start a journal txn.

Yes, correct. But I agree a WARN_ON_ONCE in the loop might be useful just
in case. I'll add it. Thanks for the idea.

								Honza