Re: [PATCH 2/2] f2fs: fix to avoid data update racing between GC and DIO

From: Jaegeuk Kim
Date: Wed Jul 06 2016 - 18:38:13 EST


On Wed, Jul 06, 2016 at 10:10:57AM +0800, Chao Yu wrote:
> On 2016/7/6 8:24, Jaegeuk Kim wrote:
> > On Fri, Jul 01, 2016 at 02:03:17PM +0800, Chao Yu wrote:
> >> Hi Jaegeuk,
> >>
> >> On 2016/7/1 8:03, Jaegeuk Kim wrote:
> >>> Hi Chao,
> >>>
> >>> On Thu, Jun 30, 2016 at 04:42:48PM +0800, Chao Yu wrote:
> >>>> 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>
> >>>> ---
> >>>> fs/f2fs/data.c | 2 ++
> >>>> fs/f2fs/f2fs.h | 1 +
> >>>> fs/f2fs/gc.c | 14 +++++++++++++-
> >>>> fs/f2fs/super.c | 1 +
> >>>> 4 files changed, 17 insertions(+), 1 deletion(-)
> >>>>
> >>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> >>>> index ba4963f..08dc060 100644
> >>>> --- a/fs/f2fs/data.c
> >>>> +++ b/fs/f2fs/data.c
> >>>> @@ -1716,7 +1716,9 @@ 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));
> >>>>
> >>>> + mutex_lock(&F2FS_I(inode)->dio_mutex);
> >>>> err = blockdev_direct_IO(iocb, inode, iter, get_data_block_dio);
> >>>> + mutex_unlock(&F2FS_I(inode)->dio_mutex);
> >>>
> >>> This means we need to sacrifice entire parallism even in the normal cases?
> >>> Can we find another way?
> >>
> >> 1. For dio write vs dio write, writer will grab i_mutex before dio_mutex, so
> >> anyway, concurrent dio writes will be exclusive.
> >>
> >> 2. For dio write vs gc, keep using dio_mutex for making them exclusive.
> >>
> >> 3. For dio read vs dio read, and dio read vs gc, what about adding dio_rwsem to
> >> control the parallelism?
> >>
> >> 4. For dio write vs dio read, we grab different lock (write grabs dio_mutex,
> >> read grabs dio_rwsem), so there is no race condition.
> >
> > How about adding a flag in a dio inode and avoiding GCs for there-in blocks?
>
> Hmm.. IMO, without lock, it's hard to keep the sequence that let GC checking the
> flag after setting it, right?

Okay, could you add dio_rwsem for now?
Later, we may need to take a look at dio_overwrite case to mitigate inode_lock
contention likewise xfs. :)

Thanks,

> >>>
> >>>> 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 bd82b6d..a241576 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 mutex dio_mutex; /* 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 c2c4ac3..98e3763 100644
> >>>> --- a/fs/f2fs/gc.c
> >>>> +++ b/fs/f2fs/gc.c
> >>>> @@ -744,12 +744,24 @@ next_step:
> >>>> /* phase 3 */
> >>>> inode = find_gc_inode(gc_list, dni.ino);
> >>>> if (inode) {
> >>>> + bool locked = false;
> >>>> +
> >>>> + if (S_ISREG(inode->i_mode)) {
> >>>> + if (!mutex_trylock(&F2FS_I(inode)->dio_mutex))
> >>>> + continue;
> >>>> + locked = true;
> >>>> + }
> >>>> +
> >>>> start_bidx = start_bidx_of_node(nofs, inode)
> >>>> + ofs_in_node;
> >>>> - if (f2fs_encrypted_inode(inode) && S_ISREG(inode->i_mode))
> >>>> + 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)
> >>>> + mutex_unlock(&F2FS_I(inode)->dio_mutex);
> >>>> +
> >>>> stat_inc_data_blk_count(sbi, 1, gc_type);
> >>>> }
> >>>> }
> >>>> diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> >>>> index 8c698e1..24aab3f 100644
> >>>> --- a/fs/f2fs/super.c
> >>>> +++ b/fs/f2fs/super.c
> >>>> @@ -575,6 +575,7 @@ 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);
> >>>> + mutex_init(&fi->dio_mutex);
> >>>>
> >>>> /* Will be used by directory only */
> >>>> fi->i_dir_level = F2FS_SB(sb)->dir_level;
> >>>> --
> >>>> 2.8.2.311.gee88674
> >>>
> >>> .
> >>>
> >
> > .
> >