Re: [PATCH 2/4] xfs: convert delayed extents to unwritten when zeroing post eof blocks

[Zhang Yi]

On 2024/3/13 15:07, Zhang Yi wrote:
> On 2024/3/13 0:21, Darrick J. Wong wrote:
>> On Tue, Mar 12, 2024 at 08:31:58PM +0800, Zhang Yi wrote:
>>> On 2024/3/11 23:37, Darrick J. Wong wrote:
>>>> On Mon, Mar 11, 2024 at 08:22:53PM +0800, Zhang Yi wrote:
>>>>> From: Zhang Yi <yi.zhang@xxxxxxxxxx>
>>>>>
>>>>> Current clone operation could be non-atomic if the destination of a file
>>>>> is beyond EOF, user could get a file with corrupted (zeroed) data on
>>>>> crash.
>>>>>
>>>>> The problem is about to pre-alloctions. If you write some data into a
>>>>> file [A, B) (the position letters are increased one by one), and xfs
>>>>> could pre-allocate some blocks, then we get a delayed extent [A, D).
>>>>> Then the writeback path allocate blocks and convert this delayed extent
>>>>> [A, C) since lack of enough contiguous physical blocks, so the extent
>>>>> [C, D) is still delayed. After that, both the in-memory and the on-disk
>>>>> file size are B. If we clone file range into [E, F) from another file,
>>>>> xfs_reflink_zero_posteof() would call iomap_zero_range() to zero out the
>>>>> range [B, E) beyond EOF and flush range. Since [C, D) is still a delayed
>>>>> extent, it will be zeroed and the file's in-memory && on-disk size will
>>>>> be updated to D after flushing and before doing the clone operation.
>>>>> This is wrong, because user can user can see the size change and read
>>>>> zeros in the middle of the clone operation.
>>>>>
>>>>> We need to keep the in-memory and on-disk size before the clone
>>>>> operation starts, so instead of writing zeroes through the page cache
>>>>> for delayed ranges beyond EOF, we convert these ranges to unwritten and
>>>>> invalidating any cached data over that range beyond EOF.
>>>>>
>>>>> Suggested-by: Dave Chinner <david@xxxxxxxxxxxxx>
>>>>> Signed-off-by: Zhang Yi <yi.zhang@xxxxxxxxxx>
>>>>> ---
>>>>>  fs/xfs/xfs_iomap.c | 29 +++++++++++++++++++++++++++++
>>>>>  1 file changed, 29 insertions(+)
>>>>>
>>>>> diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
>>>>> index ccf83e72d8ca..2b2aace25355 100644
>>>>> --- a/fs/xfs/xfs_iomap.c
>>>>> +++ b/fs/xfs/xfs_iomap.c
>>>>> @@ -957,6 +957,7 @@ xfs_buffered_write_iomap_begin(
>>>>>  	struct xfs_mount	*mp = ip->i_mount;
>>>>>  	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
>>>>>  	xfs_fileoff_t		end_fsb = xfs_iomap_end_fsb(mp, offset, count);
>>>>> +	xfs_fileoff_t		eof_fsb = XFS_B_TO_FSBT(mp, XFS_ISIZE(ip));
>>>>>  	struct xfs_bmbt_irec	imap, cmap;
>>>>>  	struct xfs_iext_cursor	icur, ccur;
>>>>>  	xfs_fsblock_t		prealloc_blocks = 0;
>>>>> @@ -1035,6 +1036,22 @@ xfs_buffered_write_iomap_begin(
>>>>>  	}
>>>>>  
>>>>>  	if (imap.br_startoff <= offset_fsb) {
>>>>> +		/*
>>>>> +		 * For zeroing out delayed allocation extent, we trim it if
>>>>> +		 * it's partial beyonds EOF block, or convert it to unwritten
>>>>> +		 * extent if it's all beyonds EOF block.
>>>>> +		 */
>>>>> +		if ((flags & IOMAP_ZERO) &&
>>>>> +		    isnullstartblock(imap.br_startblock)) {
>>>>> +			if (offset_fsb > eof_fsb)
>>>>> +				goto convert_delay;
>>>>> +			if (end_fsb > eof_fsb) {
>>>>> +				end_fsb = eof_fsb + 1;
>>>>> +				xfs_trim_extent(&imap, offset_fsb,
>>>>> +						end_fsb - offset_fsb);
>>>>> +			}
>>>>> +		}
>>>>> +
>>>>>  		/*
>>>>>  		 * For reflink files we may need a delalloc reservation when
>>>>>  		 * overwriting shared extents.   This includes zeroing of
>>>>> @@ -1158,6 +1175,18 @@ xfs_buffered_write_iomap_begin(
>>>>>  	xfs_iunlock(ip, lockmode);
>>>>>  	return xfs_bmbt_to_iomap(ip, iomap, &imap, flags, 0, seq);
>>>>>  
>>>>> +convert_delay:
>>>>> +	end_fsb = min(end_fsb, imap.br_startoff + imap.br_blockcount);
>>>>> +	xfs_iunlock(ip, lockmode);
>>>>> +	truncate_pagecache_range(inode, offset, XFS_FSB_TO_B(mp, end_fsb));
>>>>> +	error = xfs_iomap_write_direct(ip, offset_fsb, end_fsb - offset_fsb,
>>>>> +				       flags, &imap, &seq);
>>>>
>>>> I expected this to be a direct call to xfs_bmapi_convert_delalloc.
>>>> What was the reason not for using that?
>>>>
>>>
>>> It's because xfs_bmapi_convert_delalloc() isn't guarantee to convert
>>> enough blocks once a time, it may convert insufficient blocks since lack
>>> of enough contiguous free physical blocks. If we are going to use it, I
>>> suppose we need to introduce a new helper something like
>>> xfs_convert_blocks(), add a loop to do the conversion.
>>
>> I thought xfs_bmapi_convert_delalloc passes out (via @iomap) the extent
>> that xfs_bmapi_allocate (or anyone else) allocated (bma.got).  If that
>> mapping is shorter, won't xfs_buffered_write_iomap_begin pass the
>> shortened mapping out to the iomap machinery?  In which case that
>> iomap_iter loop will call ->iomap_begin on the unfinished delalloc
>> conversion work?
> 
> Yeah, make sense, it works, I forgot this loop in iomap_iter().

Sorry, I've found that it doesn't always work. Think about a special case,
If we have a file below:

	A          B           C                    D
	+wwwwwwwwww+DDDDDDDDDDD+dddddddddddddddddddd+
	          EOF         EOF
               (on disk)  (in memory)

where 'd' is a delalloc block with no data and 'D' is a delalloc
block with dirty folios over it.

xfs_bmapi_convert_delalloc() might only convert some blocks from B to B',

	A          B   B'       C                    D
	+wwwwwwwwww+UUU+DDDDDDD+dddddddddddddddddddd+
	          EOF         EOF
               (on disk)  (in memory)

After that, it will trigger below warning in iomap_iter_done():

 WARN_ON_ONCE(iter->iomap.offset + iter->iomap.length <= iter->pos);

So I guess the loop is still needed, I plane to revise and use
xfs_convert_blocks() here.

Yi.