Re: [PATCH V4] f2fs: Avoid double lock for cp_rwsem during checkpoint

From: Jaegeuk Kim
Date: Mon May 11 2020 - 18:11:04 EST


On 05/11, Chao Yu wrote:
> On 2020/5/10 3:03, Jaegeuk Kim wrote:
> > On 05/09, Chao Yu wrote:
> >> On 2020/5/9 0:10, Jaegeuk Kim wrote:
> >>> Hi Sayali,
> >>>
> >>> In order to address the perf regression, how about this?
> >>>
> >>> >From 48418af635884803ffb35972df7958a2e6649322 Mon Sep 17 00:00:00 2001
> >>> From: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> >>> Date: Fri, 8 May 2020 09:08:37 -0700
> >>> Subject: [PATCH] f2fs: avoid double lock for cp_rwsem during checkpoint
> >>>
> >>> There could be a scenario where f2fs_sync_node_pages gets
> >>> called during checkpoint, which in turn tries to flush
> >>> inline data and calls iput(). This results in deadlock as
> >>> iput() tries to hold cp_rwsem, which is already held at the
> >>> beginning by checkpoint->block_operations().
> >>>
> >>> Call stack :
> >>>
> >>> Thread A Thread B
> >>> f2fs_write_checkpoint()
> >>> - block_operations(sbi)
> >>> - f2fs_lock_all(sbi);
> >>> - down_write(&sbi->cp_rwsem);
> >>>
> >>> - open()
> >>> - igrab()
> >>> - write() write inline data
> >>> - unlink()
> >>> - f2fs_sync_node_pages()
> >>> - if (is_inline_node(page))
> >>> - flush_inline_data()
> >>> - ilookup()
> >>> page = f2fs_pagecache_get_page()
> >>> if (!page)
> >>> goto iput_out;
> >>> iput_out:
> >>> -close()
> >>> -iput()
> >>> iput(inode);
> >>> - f2fs_evict_inode()
> >>> - f2fs_truncate_blocks()
> >>> - f2fs_lock_op()
> >>> - down_read(&sbi->cp_rwsem);
> >>>
> >>> Fixes: 2049d4fcb057 ("f2fs: avoid multiple node page writes due to inline_data")
> >>> Signed-off-by: Sayali Lokhande <sayalil@xxxxxxxxxxxxxx>
> >>> Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> >>> ---
> >>> fs/f2fs/node.c | 4 ++--
> >>> 1 file changed, 2 insertions(+), 2 deletions(-)
> >>>
> >>> diff --git a/fs/f2fs/node.c b/fs/f2fs/node.c
> >>> index 1db8cabf727ef..626d7daca09de 100644
> >>> --- a/fs/f2fs/node.c
> >>> +++ b/fs/f2fs/node.c
> >>> @@ -1870,8 +1870,8 @@ int f2fs_sync_node_pages(struct f2fs_sb_info *sbi,
> >>> goto continue_unlock;
> >>> }
> >>>
> >>> - /* flush inline_data */
> >>> - if (is_inline_node(page)) {
> >>> + /* flush inline_data, if it's not sync path. */
> >>> + if (do_balance && is_inline_node(page)) {
> >>
> >> IIRC, this flow was designed to avoid running out of free space issue
> >> during checkpoint:
> >>
> >> 2049d4fcb057 ("f2fs: avoid multiple node page writes due to inline_data")
> >>
> >> The sceanrio is:
> >> 1. create fully node blocks
> >> 2. flush node blocks
> >> 3. write inline_data for all the node blocks again
> >> 4. flush node blocks redundantly
> >>
> >> I guess this may cause failing one case of fstest.
> >
> > Yeah, actually I was hitting 204 failure, and thus, revised like this.
> > Now, I don't see any regression in fstest.
> >
> >>From 8f1882acfb0a5fc43e5a2bbd576a8f3c681a7d2c Mon Sep 17 00:00:00 2001
> > From: Sayali Lokhande <sayalil@xxxxxxxxxxxxxx>
> > Date: Thu, 30 Apr 2020 16:28:29 +0530
> > Subject: [PATCH] f2fs: Avoid double lock for cp_rwsem during checkpoint
> >
> > There could be a scenario where f2fs_sync_node_pages gets
> > called during checkpoint, which in turn tries to flush
> > inline data and calls iput(). This results in deadlock as
> > iput() tries to hold cp_rwsem, which is already held at the
> > beginning by checkpoint->block_operations().
> >
> > Call stack :
> >
> > Thread A Thread B
> > f2fs_write_checkpoint()
> > - block_operations(sbi)
> > - f2fs_lock_all(sbi);
> > - down_write(&sbi->cp_rwsem);
> >
> > - open()
> > - igrab()
> > - write() write inline data
> > - unlink()
> > - f2fs_sync_node_pages()
> > - if (is_inline_node(page))
> > - flush_inline_data()
> > - ilookup()
> > page = f2fs_pagecache_get_page()
> > if (!page)
> > goto iput_out;
> > iput_out:
> > -close()
> > -iput()
> > iput(inode);
> > - f2fs_evict_inode()
> > - f2fs_truncate_blocks()
> > - f2fs_lock_op()
> > - down_read(&sbi->cp_rwsem);
> >
> > Fixes: 2049d4fcb057 ("f2fs: avoid multiple node page writes due to inline_data")
> > Signed-off-by: Sayali Lokhande <sayalil@xxxxxxxxxxxxxx>
> > Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > ---
> > fs/f2fs/checkpoint.c | 9 ++++++++-
> > fs/f2fs/f2fs.h | 4 ++--
> > fs/f2fs/node.c | 10 +++++-----
> > 3 files changed, 15 insertions(+), 8 deletions(-)
> >
> > diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c
> > index d49f7a01d8a26..928aea4ff663d 100644
> > --- a/fs/f2fs/checkpoint.c
> > +++ b/fs/f2fs/checkpoint.c
> > @@ -1168,6 +1168,12 @@ static int block_operations(struct f2fs_sb_info *sbi)
> > };
> > int err = 0, cnt = 0;
> >
> > + /*
> > + * Let's flush node pages first to flush inline_data.
> > + * We'll actually guarantee everything below under f2fs_lock_all.
> > + */
> > + f2fs_sync_node_pages(sbi, &wbc, false, false, FS_CP_NODE_IO);
>
> It is possible that user write a large number of inline data in between
> f2fs_sync_node_pages() and f2fs_lock_all(), it will cause the no-space issue in
> race condition.
>
> Also, if there is huge number of F2FS_DIRTY_IMETA, after this change, we will
> flush inode page twice which is unneeded.
>
> f2fs_sync_node_pages() --- flush dirty inode page
> f2fs_lock_all()
> ...
> f2fs_sync_inode_meta() --- update dirty inode page
> f2fs_sync_node_pages() --- flush dirty inode page again.
>

Another version: