Re: [PATCH v3] f2fs: fix to avoid data update racing between GC and DIO

From: Jaegeuk Kim
Date: Tue Jul 12 2016 - 13:09:12 EST

On Tue, Jul 12, 2016 at 09:28:26AM +0800, Chao Yu wrote:
> On 2016/7/10 0:22, Jaegeuk Kim wrote:
> > On Fri, Jul 08, 2016 at 11:50:02PM +0800, Chao Yu wrote:
> >> Hi Jaegeuk,
> >>
> >> On 2016/7/8 11:19, Jaegeuk Kim wrote:
> >>> Hi Chao,
> >>>
> >>> Could you take a look at this in xfstests/generic/013?
> >>>
> >>> [ 502.480850] ======================================================
> >>> [ 502.480864] [ INFO: possible circular locking dependency detected ]
> >>> [ 502.480877] 4.7.0-rc1+ #124 Tainted: G OE
> >>> [ 502.480886] -------------------------------------------------------
> >>> [ 502.480897] fsstress/10729 is trying to acquire lock:
> >>> [ 502.480906] (&sb->s_type->i_mutex_key#18){+.+.+.}, at: [<ffffffff81299c3b>] do_blockdev_direct_IO+0x1db/0x2310
> >>> [ 502.480948]
> >>> [ 502.480948] but task is already holding lock:
> >>> [ 502.480959] (&fi->dio_rwsem){.+.+.+}, at: [<ffffffffc081e2b1>] f2fs_direct_IO+0xd1/0x3d0 [f2fs]
> >>> [ 502.481003]
> >>> [ 502.481003] which lock already depends on the new lock.
> >>> [ 502.481003]
> >>> [ 502.481018]
> >>> [ 502.481018] the existing dependency chain (in reverse order) is:
> >>> [ 502.481030]
> >>> [ 502.481030] -> #1 (&fi->dio_rwsem){.+.+.+}:
> >>> [ 502.481054] [<ffffffff810e51c3>] lock_acquire+0xd3/0x220
> >>> [ 502.481071] [<ffffffff818d1921>] down_read+0x51/0xa0
> >>> [ 502.481089] [<ffffffffc081e2b1>] f2fs_direct_IO+0xd1/0x3d0 [f2fs]
> >>> [ 502.481114] [<ffffffff811c34c7>] generic_file_direct_write+0xa7/0x160
> >>> [ 502.481133] [<ffffffff811c363d>] __generic_file_write_iter+0xbd/0x1e0
> >>> [ 502.481149] [<ffffffffc080437b>] f2fs_file_write_iter+0xdb/0x100 [f2fs]
> >>> [ 502.481173] [<ffffffff81253a88>] __vfs_write+0xc8/0x140
> >>> [ 502.481190] [<ffffffff81254c55>] vfs_write+0xb5/0x1b0
> >>> [ 502.481205] [<ffffffff81255fe9>] SyS_write+0x49/0xa0
> >>> [ 502.481220] [<ffffffff818d4100>] entry_SYSCALL_64_fastpath+0x23/0xc1
> >>> [ 502.481236]
> >>> [ 502.481236] -> #0 (&sb->s_type->i_mutex_key#18){+.+.+.}:
> >>> [ 502.481264] [<ffffffff810e481c>] __lock_acquire+0x161c/0x1940
> >>> [ 502.481280] [<ffffffff810e51c3>] lock_acquire+0xd3/0x220
> >>> [ 502.481296] [<ffffffff818d1b9a>] down_write+0x5a/0xc0
> >>> [ 502.481312] [<ffffffff81299c3b>] do_blockdev_direct_IO+0x1db/0x2310
> >>> [ 502.481328] [<ffffffff8129bdaa>] __blockdev_direct_IO+0x3a/0x40
> >>> [ 502.481344] [<ffffffffc081e2e4>] f2fs_direct_IO+0x104/0x3d0 [f2fs]
> >>> [ 502.481368] [<ffffffff811c40a9>] generic_file_read_iter+0x689/0x7e0
> >>> [ 502.481384] [<ffffffff812545d1>] __vfs_read+0xc1/0x130
> >>> [ 502.481399] [<ffffffff81254af1>] vfs_read+0x91/0x140
> >>> [ 502.481414] [<ffffffff81255f49>] SyS_read+0x49/0xa0
> >>> [ 502.481429] [<ffffffff818d4100>] entry_SYSCALL_64_fastpath+0x23/0xc1
> >>> [ 502.481445]
> >>> [ 502.481445] other info that might help us debug this:
> >>> [ 502.481445]
> >>> [ 502.481459] Possible unsafe locking scenario:
> >>> [ 502.481459]
> >>> [ 502.481726] CPU0 CPU1
> >>> [ 502.481987] ---- ----
> >>> [ 502.482242] lock(&fi->dio_rwsem);
> >>> [ 502.482501] lock(&sb->s_type->i_mutex_key#18);
> >>> [ 502.482765] lock(&fi->dio_rwsem);
> >>> [ 502.483025] lock(&sb->s_type->i_mutex_key#18);
> >>
> >> Seems we will suffer ABBA deadlock:
> >>
> >> writer reader
> >> - f2fs_file_write_iter
> >> - down_write(&inode->i_rwsem)
> >> - __generic_file_write_iter
> >> - generic_file_direct_write
> >> - f2fs_direct_IO
> >> - generic_file_read_iter
> >> - f2fs_direct_IO
> >> - down_read(&fi->dio_rwsem)
> >> - __blockdev_direct_IO
> >> - do_blockdev_direct_IO
> >> - down_write(&inode->i_rwsem)
> >>
> >> - down_read(&fi->dio_rwsem)
> >>
> >> What about splitting dio_rwsem to rdio_rwsem/wdio_rwsem for reader/writer to
> >> avoid deadlock?
> >
> > Hmm, how about inode_trylock in GC?
>
> If we reuse inode->i_rwsem here, we will suffer the same issue when we remove
> i_rwsem lock in dio writer or dio reader for better concurrency.
>
> So I think it's better to use separate lock to just fix this issue.

Got it.

>
> Thanks,
>
> >
> >>
> >> Thanks,
> >>
> >>> [ 502.483285]
> >>> [ 502.483285] *** DEADLOCK ***
> >>> [ 502.483285]
> >>> [ 502.484018] 1 lock held by fsstress/10729:
> >>> [ 502.484262] #0: (&fi->dio_rwsem){.+.+.+}, at: [<ffffffffc081e2b1>] f2fs_direct_IO+0xd1/0x3d0 [f2fs]
> >>>
> >>> Thanks,
> >>>
> >>> On Thu, Jul 07, 2016 at 12:49:12PM +0800, Chao Yu wrote:
> >>>> From: Chao Yu <yuchao0@xxxxxxxxxx>
> >>>>
> >>>> Datas in file can be operated by GC and DIO simultaneously, so we will
> >>>> face race case as below:
> >>>>
> >>>> For write case:
> >>>> Thread A Thread B
> >>>> - generic_file_direct_write
> >>>> - invalidate_inode_pages2_range
> >>>> - f2fs_direct_IO
> >>>> - do_blockdev_direct_IO
> >>>> - do_direct_IO
> >>>> - get_more_blocks
> >>>> - f2fs_gc
> >>>> - do_garbage_collect
> >>>> - gc_data_segment
> >>>> - move_data_page
> >>>> - do_write_data_page
> >>>> migrate data block to new block address
> >>>> - dio_bio_submit
> >>>> update user data to old block address
> >>>>
> >>>> For read case:
> >>>> Thread A Thread B
> >>>> - generic_file_direct_write
> >>>> - invalidate_inode_pages2_range
> >>>> - f2fs_direct_IO
> >>>> - do_blockdev_direct_IO
> >>>> - do_direct_IO
> >>>> - get_more_blocks
> >>>> - f2fs_balance_fs
> >>>> - f2fs_gc
> >>>> - do_garbage_collect
> >>>> - gc_data_segment
> >>>> - move_data_page
> >>>> - do_write_data_page
> >>>> migrate data block to new block address
> >>>> - write_checkpoint
> >>>> - do_checkpoint
> >>>> - clear_prefree_segments
> >>>> - f2fs_issue_discard
> >>>> discard old block adress
> >>>> - dio_bio_submit
> >>>> update user buffer from obsolete block address
> >>>>
> >>>> In order to fix this, for one file, we should let DIO and GC getting exclusion
> >>>> against with each other.
> >>>>
> >>>> Signed-off-by: Chao Yu <yuchao0@xxxxxxxxxx>
> >>>> ---
> >>>> v3: use semaphore to avoid racing in between read dio and write dio.
> >>>> fs/f2fs/data.c | 4 ++++
> >>>> fs/f2fs/f2fs.h | 1 +
> >>>> fs/f2fs/gc.c | 13 +++++++++++++
> >>>> fs/f2fs/super.c | 1 +
> >>>> 4 files changed, 19 insertions(+)
> >>>>
> >>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> >>>> index b6fd5bd..19197bb 100644
> >>>> --- a/fs/f2fs/data.c
> >>>> +++ b/fs/f2fs/data.c
> >>>> @@ -1712,6 +1712,7 @@ static ssize_t f2fs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
> >>>> {
> >>>> struct address_space *mapping = iocb->ki_filp->f_mapping;
> >>>> struct inode *inode = mapping->host;
> >>>> + struct f2fs_inode_info *fi = F2FS_I(inode);
> >>>> size_t count = iov_iter_count(iter);
> >>>> loff_t offset = iocb->ki_pos;
> >>>> int err;
> >>>> @@ -1727,7 +1728,10 @@ static ssize_t f2fs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
> >>>>
> >>>> trace_f2fs_direct_IO_enter(inode, offset, count, iov_iter_rw(iter));
> >>>>
> >>>> + down_read(&fi->dio_rwsem);
> >>>> err = blockdev_direct_IO(iocb, inode, iter, get_data_block_dio);
> >>>> + up_read(&fi->dio_rwsem);
> >>>> +
> >>>> if (iov_iter_rw(iter) == WRITE) {
> >>>> if (err > 0)
> >>>> set_inode_flag(inode, FI_UPDATE_WRITE);
> >>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> >>>> index bf9a13a..2e439ec 100644
> >>>> --- a/fs/f2fs/f2fs.h
> >>>> +++ b/fs/f2fs/f2fs.h
> >>>> @@ -474,6 +474,7 @@ struct f2fs_inode_info {
> >>>> struct list_head inmem_pages; /* inmemory pages managed by f2fs */
> >>>> struct mutex inmem_lock; /* lock for inmemory pages */
> >>>> struct extent_tree *extent_tree; /* cached extent_tree entry */
> >>>> + struct rw_semaphore dio_rwsem; /* avoid racing between dio and gc */
> >>>> };
> >>>>
> >>>> static inline void get_extent_info(struct extent_info *ext,
> >>>> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> >>>> index c612137..a9bfb8d 100644
> >>>> --- a/fs/f2fs/gc.c
> >>>> +++ b/fs/f2fs/gc.c
> >>>> @@ -755,12 +755,25 @@ next_step:
> >>>> /* phase 3 */
> >>>> inode = find_gc_inode(gc_list, dni.ino);
> >>>> if (inode) {
> >>>> + struct f2fs_inode_info *fi = F2FS_I(inode);
> >>>> + bool locked = false;
> >>>> +
> >>>> + if (S_ISREG(inode->i_mode)) {
> >>>> + if (!down_write_trylock(&fi->dio_rwsem))
> >>>> + continue;
> >>>> + locked = true;
> >>>> + }
> >>>> +
> >>>> start_bidx = start_bidx_of_node(nofs, inode)
> >>>> + ofs_in_node;
> >>>> if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
> >>>> move_encrypted_block(inode, start_bidx);
> >>>> else
> >>>> move_data_page(inode, start_bidx, gc_type);
> >>>> +
> >>>> + if (locked)
> >>>> + up_write(&fi->dio_rwsem);
> >>>> +
> >>>> stat_inc_data_blk_count(sbi, 1, gc_type);
> >>>> }
> >>>> }
> >>>> diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> >>>> index edd1b35..dde57fb 100644
> >>>> --- a/fs/f2fs/super.c
> >>>> +++ b/fs/f2fs/super.c
> >>>> @@ -579,6 +579,7 @@ static struct inode *f2fs_alloc_inode(struct super_block *sb)
> >>>> INIT_LIST_HEAD(&fi->gdirty_list);
> >>>> INIT_LIST_HEAD(&fi->inmem_pages);
> >>>> mutex_init(&fi->inmem_lock);
> >>>> + init_rwsem(&fi->dio_rwsem);
> >>>>
> >>>> /* Will be used by directory only */
> >>>> fi->i_dir_level = F2FS_SB(sb)->dir_level;
> >>>> --
> >>>> 2.7.2
> >
> > .
> >