Re: [PATCH] f2fs: avoid fi->i_gc_rwsem[WRITE] lock in f2fs_gc

[Jaegeuk Kim]

On 07/30, Chao Yu wrote:
> On 2018/7/30 17:08, Jaegeuk Kim wrote:
> > On 07/30, Chao Yu wrote:
> >> On 2018/7/30 12:18, Jaegeuk Kim wrote:
> >>> On 07/30, Chao Yu wrote:
> >>>> On 2018/7/30 9:32, Jaegeuk Kim wrote:
> >>>>> The f2fs_gc() called by f2fs_balance_fs() requires to be called outside of
> >>>>> fi->i_gc_rwsem[WRITE], since f2fs_gc() can try to grab it in a loop.
> >>>>>
> >>>>> If it hits the miximum retrials in GC, let's give a chance to release
> >>>>> gc_mutex for a short time in order not to go into live lock in the worst
> >>>>> case.
> >>>>>
> >>>>> Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> >>>>> ---
> >>>>>  fs/f2fs/f2fs.h    |  1 +
> >>>>>  fs/f2fs/file.c    | 62 ++++++++++++++++++++++-------------------------
> >>>>>  fs/f2fs/gc.c      | 22 ++++++++++++-----
> >>>>>  fs/f2fs/segment.c |  5 +++-
> >>>>>  fs/f2fs/segment.h |  2 +-
> >>>>>  5 files changed, 51 insertions(+), 41 deletions(-)
> >>>>>
> >>>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> >>>>> index a9447c7d6570..50349780001b 100644
> >>>>> --- a/fs/f2fs/f2fs.h
> >>>>> +++ b/fs/f2fs/f2fs.h
> >>>>> @@ -1223,6 +1223,7 @@ struct f2fs_sb_info {
> >>>>>  	unsigned int gc_mode;			/* current GC state */
> >>>>>  	/* for skip statistic */
> >>>>>  	unsigned long long skipped_atomic_files[2];	/* FG_GC and BG_GC */
> >>>>> +	unsigned long long skipped_gc_rwsem;		/* FG_GC only */
> >>>>>  
> >>>>>  	/* threshold for gc trials on pinned files */
> >>>>>  	u64 gc_pin_file_threshold;
> >>>>> diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> >>>>> index 78c1bd6b8497..2b7d26ebb294 100644
> >>>>> --- a/fs/f2fs/file.c
> >>>>> +++ b/fs/f2fs/file.c
> >>>>> @@ -1179,10 +1179,12 @@ static int __exchange_data_block(struct inode *src_inode,
> >>>>>  	return ret;
> >>>>>  }
> >>>>>  
> >>>>> -static int f2fs_do_collapse(struct inode *inode, pgoff_t start, pgoff_t end)
> >>>>> +static int f2fs_do_collapse(struct inode *inode, loff_t offset, loff_t len)
> >>>>>  {
> >>>>>  	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> >>>>>  	pgoff_t nrpages = (i_size_read(inode) + PAGE_SIZE - 1) / PAGE_SIZE;
> >>>>> +	pgoff_t start = offset >> PAGE_SHIFT;
> >>>>> +	pgoff_t end = (offset + len) >> PAGE_SHIFT;
> >>>>>  	int ret;
> >>>>>  
> >>>>>  	f2fs_balance_fs(sbi, true);
> >>>>> @@ -1190,14 +1192,18 @@ static int f2fs_do_collapse(struct inode *inode, pgoff_t start, pgoff_t end)
> >>>>>  
> >>>>>  	f2fs_drop_extent_tree(inode);
> >>>>>  
> >>>>> +	/* avoid gc operation during block exchange */
> >>>>> +	down_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
> >>>>> +	truncate_pagecache(inode, offset);
> >>>>>  	ret = __exchange_data_block(inode, inode, end, start, nrpages - end, true);
> >>>>> +	up_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
> >>>>> +
> >>>>>  	f2fs_unlock_op(sbi);
> >>>>>  	return ret;
> >>>>>  }
> >>>>>  
> >>>>>  static int f2fs_collapse_range(struct inode *inode, loff_t offset, loff_t len)
> >>>>>  {
> >>>>> -	pgoff_t pg_start, pg_end;
> >>>>>  	loff_t new_size;
> >>>>>  	int ret;
> >>>>>  
> >>>>> @@ -1212,21 +1218,13 @@ static int f2fs_collapse_range(struct inode *inode, loff_t offset, loff_t len)
> >>>>>  	if (ret)
> >>>>>  		return ret;
> >>>>>  
> >>>>> -	pg_start = offset >> PAGE_SHIFT;
> >>>>> -	pg_end = (offset + len) >> PAGE_SHIFT;
> >>>>> -
> >>>>> -	/* avoid gc operation during block exchange */
> >>>>> -	down_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
> >>>>> -
> >>>>>  	down_write(&F2FS_I(inode)->i_mmap_sem);
> >>>>>  	/* write out all dirty pages from offset */
> >>>>>  	ret = filemap_write_and_wait_range(inode->i_mapping, offset, LLONG_MAX);
> >>>>>  	if (ret)
> >>>>>  		goto out_unlock;
> >>>>>  
> >>>>> -	truncate_pagecache(inode, offset);
> >>>>> -
> >>>>> -	ret = f2fs_do_collapse(inode, pg_start, pg_end);
> >>>>> +	ret = f2fs_do_collapse(inode, offset, len);
> >>>>>  	if (ret)
> >>>>>  		goto out_unlock;
> >>>>>  
> >>>>> @@ -1242,7 +1240,6 @@ static int f2fs_collapse_range(struct inode *inode, loff_t offset, loff_t len)
> >>>>>  		f2fs_i_size_write(inode, new_size);
> >>>>>  out_unlock:
> >>>>>  	up_write(&F2FS_I(inode)->i_mmap_sem);
> >>>>> -	up_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
> >>>>>  	return ret;
> >>>>>  }
> >>>>>  
> >>>>> @@ -1417,9 +1414,6 @@ static int f2fs_insert_range(struct inode *inode, loff_t offset, loff_t len)
> >>>>>  
> >>>>>  	f2fs_balance_fs(sbi, true);
> >>>>>  
> >>>>> -	/* avoid gc operation during block exchange */
> >>>>> -	down_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
> >>>>> -
> >>>>>  	down_write(&F2FS_I(inode)->i_mmap_sem);
> >>>>>  	ret = f2fs_truncate_blocks(inode, i_size_read(inode), true);
> >>>>>  	if (ret)
> >>>>> @@ -1430,13 +1424,15 @@ static int f2fs_insert_range(struct inode *inode, loff_t offset, loff_t len)
> >>>>>  	if (ret)
> >>>>>  		goto out;
> >>>>>  
> >>>>> -	truncate_pagecache(inode, offset);
> >>>>> -
> >>>>>  	pg_start = offset >> PAGE_SHIFT;
> >>>>>  	pg_end = (offset + len) >> PAGE_SHIFT;
> >>>>>  	delta = pg_end - pg_start;
> >>>>>  	idx = (i_size_read(inode) + PAGE_SIZE - 1) / PAGE_SIZE;
> >>>>>  
> >>>>> +	/* avoid gc operation during block exchange */
> >>>>> +	down_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
> >>>>> +	truncate_pagecache(inode, offset);
> >>>>> +
> >>>>>  	while (!ret && idx > pg_start) {
> >>>>>  		nr = idx - pg_start;
> >>>>>  		if (nr > delta)
> >>>>> @@ -1450,6 +1446,7 @@ static int f2fs_insert_range(struct inode *inode, loff_t offset, loff_t len)
> >>>>>  					idx + delta, nr, false);
> >>>>>  		f2fs_unlock_op(sbi);
> >>>>>  	}
> >>>>> +	up_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
> >>>>>  
> >>>>>  	/* write out all moved pages, if possible */
> >>>>>  	filemap_write_and_wait_range(inode->i_mapping, offset, LLONG_MAX);
> >>>>> @@ -1459,7 +1456,6 @@ static int f2fs_insert_range(struct inode *inode, loff_t offset, loff_t len)
> >>>>>  		f2fs_i_size_write(inode, new_size);
> >>>>>  out:
> >>>>>  	up_write(&F2FS_I(inode)->i_mmap_sem);
> >>>>> -	up_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
> >>>>>  	return ret;
> >>>>>  }
> >>>>>  
> >>>>> @@ -1706,8 +1702,6 @@ static int f2fs_ioc_start_atomic_write(struct file *filp)
> >>>>>  
> >>>>>  	inode_lock(inode);
> >>>>>  
> >>>>> -	down_write(&F2FS_I(inode)->i_gc_rwsem[WRITE]);
> >>>>
> >>>> After removing lock here, how can we handle below condition:
> >>>>
> >>>> commit 27319ba4044c0c67d62ae39e53c0118c89f0a029
> >>>> Author: Chao Yu <yuchao0@xxxxxxxxxx>
> >>>> Date:   Tue Apr 17 17:51:28 2018 +0800
> >>>>
> >>>>     f2fs: fix race in between GC and atomic open
> >>>>
> >>>>     Thread                                  GC thread
> >>>>     - f2fs_ioc_start_atomic_write
> >>>>      - get_dirty_pages
> >>>>      - filemap_write_and_wait_range
> >>>>                                             - f2fs_gc
> >>>>                                              - do_garbage_collect
> >>>>                                               - gc_data_segment
> >>>>                                                - move_data_page
> >>>>                                                 - f2fs_is_atomic_file
> >>>>                                                 - set_page_dirty
> >>>>      - set_inode_flag(, FI_ATOMIC_FILE)
> >>>>
> >>>>     Dirty data page can still be generated by GC in race condition as
> >>>>     above call stack.
> >>>>
> >>>>     This patch adds fi->dio_rwsem[WRITE] in f2fs_ioc_start_atomic_write
> >>>>     to avoid such race.
> >>>
> >>> "f2fs: don't allow any writes on aborted atomic writes" disallows any writes
> >>> on atomic file which has the revoking flag. So, this won't happen. In GC,
> >>
> >> Hmmm... In above condition, it's not related to FI_ATOMIC_REVOKE_REQUEST flag
> >> since we do not drop any inmem pages for atomic file.
> >>
> >> That patch was trying to eliminate a hole which exists in between
> >> filemap_write_and_wait_range and set_inode_flag(, FI_ATOMIC_FILE), where GC can
> >> still dirty page in the inode, it can pollute isolation of database transaction,
> >> so that is why we need this lock.
> > 
> > Ah, GC can generate any dirty pages of atomic_written data before starting
> > another transaction, right?
> 
> Yes,
> 
> > 
> > I think we can do
> >  - set_inode_flag() first, followed by
> >  - filemap_write_and_wait_range().
> 
> If there is redirty flow during filemap_write_and_wait_range, the page can be
> register as inmem one?
> 
> f2fs_set_data_page_dirty()
> ...
> 	if (f2fs_is_atomic_file(inode) && !f2fs_is_commit_atomic_write(inode)) {
> 		if (!IS_ATOMIC_WRITTEN_PAGE(page)) {
> 			f2fs_register_inmem_page(inode, page);
> 			return 1;
> 		}
> 
> Another concern is set_inode_flag and filemap_write_and_wait_range can be
> reorder by CPU pipeline, so the serial should be?
> 
> set_inode_flag(, FI_ATOMIC_COMMIT)
> smp_mb()
> set_inode_flag(, FI_ATOMIC_FILE)
> smp_mb()
> 
> ret = filemap_write_and_wait_range
> if (ret)
> 	goto err_out;
> 
> clear_inode_flag(, FI_ATOMIC_COMMIT)
> 
> Is that right?

Getting to the wrong way. Let me try rwsem again in the narrowed scope like
this.