答复: [PATCH] fs-writeback: writeback_sb_inodes: Do not increase 'total_wrote' when nothing is written

From: 郭纯海
Date: Thu Sep 14 2023 - 00:12:41 EST


> On Wed 13-09-23 07:15:01, Chunhai Guo wrote:
> > > On Wed 13-09-23 10:42:21, Christian Brauner wrote:
> > > > [+Cc Jan]
> > >
> > > Thanks!
> > >
> > > > On Tue, Sep 12, 2023 at 08:20:43AM -0600, Chunhai Guo wrote:
> > > > > I am encountering a deadlock issue as shown below. There is a
> > > > > commit 344150999b7f ("f2fs: fix to avoid potential deadlock")
> > > > > can fix this issue.
> > > > > However, from log analysis, it appears that this is more likely
> > > > > a fake progress issue similar to commit 68f4c6eba70d ("fs-writeback:
> > > > > writeback_sb_inodes: Recalculate 'wrote' according skipped pages").
> > > > > In each writeback iteration, nothing is written, while
> > > > > writeback_sb_inodes() increases 'total_wrote' each time, causing
> > > > > an infinite loop. This patch fixes this issue by not increasing
> > > > > 'total_wrote' when nothing is written.
> > > > >
> > > > > wb_writeback fsync (inode-Y)
> > > > > blk_start_plug(&plug)
> > > > > for (;;) {
> > > > > iter i-1: some reqs with page-X added into plug->mq_list // f2fs node
> > > > > page-X with PG_writeback
> > > > > filemap_fdatawrite
> > > > > __filemap_fdatawrite_range // write inode-Y
> > > > > with sync_mode WB_SYNC_ALL
> > > > > do_writepages
> > > > > f2fs_write_data_pages
> > > > > __f2fs_write_data_pages //
> > > > > wb_sync_req[DATA]++ for WB_SYNC_ALL
> > > > > f2fs_write_cache_pages
> > > > > f2fs_write_single_data_page
> > > > > f2fs_do_write_data_page
> > > > > f2fs_outplace_write_data
> > > > > f2fs_update_data_blkaddr
> > > > > f2fs_wait_on_page_writeback
> > > > > wait_on_page_writeback // wait for
> > > > > f2fs node page-X
> > > > > iter i:
> > > > > progress = __writeback_inodes_wb(wb, work)
> > > > > . writeback_sb_inodes
> > > > > . __writeback_single_inode // write inode-Y with sync_mode
> > > > > WB_SYNC_NONE
> > > > > . . do_writepages
> > > > > . . f2fs_write_data_pages
> > > > > . . . __f2fs_write_data_pages // skip writepages due to
> > > > > (wb_sync_req[DATA]>0)
> > > > > . . . wbc->pages_skipped += get_dirty_pages(inode) //
> > > > > wbc->pages_skipped = 1
> > > > > . if (!(inode->i_state & I_DIRTY_ALL)) // i_state = I_SYNC |
> > > > > I_SYNC_QUEUED
> > > > > . total_wrote++; // total_wrote = 1
> > > > > . requeue_inode // requeue inode-Y to wb->b_dirty queue due to
> > > > > non-zero pages_skipped
> > > > > if (progress) // progress = 1
> > > > > continue;
> > > > > iter i+1:
> > > > > queue_io
> > > > > // similar process with iter i, infinite for-loop !
> > > > > }
> > > > > blk_finish_plug(&plug) // flush plug won't be called
> > > > >
> > > > > Signed-off-by: Chunhai Guo <guochunhai@xxxxxxxx>
> > >
> > > Thanks for the patch but did you test this patch fixed your deadlock?
> > > Because the patch seems like a noop to me. Look:
> >
> > Yes. I have tested this patch and it indeed fixed this deadlock issue, too.
>
> OK, thanks for letting me know!
>
> > > > > diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c index
> > > > > 969ce991b0b0..54cdee906be9 100644
> > > > > --- a/fs/fs-writeback.c
> > > > > +++ b/fs/fs-writeback.c
> > > > > @@ -1820,6 +1820,7 @@ static long writeback_sb_inodes(struct
> > > > > super_block *sb,
> > > > > struct inode *inode = wb_inode(wb->b_io.prev);
> > > > > struct bdi_writeback *tmp_wb;
> > > > > long wrote;
> > > > > + bool is_dirty_before;
> > > > >
> > > > > if (inode->i_sb != sb) {
> > > > > if (work->sb) { @@ -1881,6 +1882,7 @@ static
> > > > > long writeback_sb_inodes(struct super_block *sb,
> > > > > continue;
> > > > > }
> > > > > inode->i_state |= I_SYNC;
> > > > > + is_dirty_before = inode->i_state & I_DIRTY_ALL;
> > >
> > > is_dirty_before is going to be set if there's anything dirty -
> > > inode, page, timestamp. So it can be unset only if there are no
> > > dirty pages, in which case there are no pages that can be skipped
> > > during page writeback, which means that requeue_inode() will go and
> > > remove inode from b_io/b_dirty lists and it will not participate in writeback
> anymore.
> > >
> > > So I don't see how this patch can be helping anything... Please
> > > correct me if I'm missing anything.
> > >
> > > Honza
> >
> > From the dump info, there are only two pages as shown below. One is
> > updated and another is under writeback. Maybe f2fs counts the
> > writeback page as a dirty one, so get_dirty_pages() got one. As you
> > said, maybe this is unreasonable.
> >
> > Jaegeuk & Chao, what do you think of this?
> >
> >
> > crash_32> files -p 0xE5A44678
> > INODE NRPAGES
> > e5a44678 2
> >
> > PAGE PHYSICAL MAPPING INDEX CNT FLAGS
> > e8d0e338 641de000 e5a44810 0 5 a095
> locked,waiters,uptodate,lru,private,writeback
> > e8ad59a0 54528000 e5a44810 1 2 2036
> referenced,uptodate,lru,active,private
>
> Indeed, incrementing pages_skipped when there's no dirty page is a bit odd.
> That being said we could also harden requeue_inode() - in particular we could do
> there:
>
> if (wbc->pages_skipped) {
> /*
> * Writeback is not making progress due to locked buffers.
> * Skip this inode for now. Although having skipped pages
> * is odd for clean inodes, it can happen for some
> * filesystems so handle that gracefully.
> */
> if (inode->i_state & I_DIRTY_ALL)
> redirty_tail_locked(inode, wb);
> else
> inode_cgwb_move_to_attached(inode, wb);
> }
>
> Does this fix your problem as well?
>
> Honza

Thank you for your reply. Did you forget the 'return' statement? Since I encountered this issue on the 4.19 kernel and there is not inode_cgwb_move_to_attached() yet, I replaced it with inode_io_list_del_locked(). So, below is the test patch I am applying. Please have a check. By the way, the test will take some time. I will provide feedback when it is finished. Thanks.

if (wbc->pages_skipped) {
/*
* writeback is not making progress due to locked
* buffers. Skip this inode for now.
*/
- redirty_tail_locked(inode, wb);
+ if (inode->i_state & I_DIRTY_ALL)
+ redirty_tail_locked(inode, wb);
+ else
+ inode_io_list_del_locked(inode, wb);
return;
}

> >
> > Thanks,
> >
> > >
> > >
> > > > > wbc_attach_and_unlock_inode(&wbc, inode);
> > > > >
> > > > > write_chunk = writeback_chunk_size(wb, work); @@
> > > > > -1918,7
> > > > > +1920,7 @@ static long writeback_sb_inodes(struct super_block
> > > > > +*sb,
> > > > > */
> > > > > tmp_wb = inode_to_wb_and_lock_list(inode);
> > > > > spin_lock(&inode->i_lock);
> > > > > - if (!(inode->i_state & I_DIRTY_ALL))
> > > > > + if (!(inode->i_state & I_DIRTY_ALL) &&
> > > > > + is_dirty_before)
> > > > > total_wrote++;
> > > > > requeue_inode(inode, tmp_wb, &wbc);
> > > > > inode_sync_complete(inode);
> > > > > --
> > > > > 2.25.1
> > > > >
> > > --
> > > Jan Kara <jack@xxxxxxxx>
> > > SUSE Labs, CR
> >
> --
> Jan Kara <jack@xxxxxxxx>
> SUSE Labs, CR