Date:   Tue, 26 Nov 2019 06:19:46 +0000
From:   Damien Le Moal <Damien.LeMoal@....com>
To:     Javier González <javier@...igon.com>
CC:     "jaegeuk@...nel.org" <jaegeuk@...nel.org>,
        "yuchao0@...wei.com" <yuchao0@...wei.com>,
        "linux-f2fs-devel@...ts.sourceforge.net" 
        <linux-f2fs-devel@...ts.sourceforge.net>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Javier González <javier.gonz@...sung.com>
Subject: Re: [PATCH] f2fs: disble physical prealloc in LSF mount

On 2019/11/26 12:58, Javier González wrote:
> On 26.11.2019 02:06, Damien Le Moal wrote:
>> On 2019/11/26 4:03, Javier González wrote:
>>> On 25.11.2019 00:48, Damien Le Moal wrote:
>>>> On 2019/11/22 18:00, Javier González wrote:
>>>>> From: Javier González <javier.gonz@...sung.com>
>>>>>
>>>>> Fix file system corruption when using LFS mount (e.g., in zoned
>>>>> devices). Seems like the fallback into buffered I/O creates an
>>>>> inconsistency if the application is assuming both read and write DIO. I
>>>>> can easily reproduce a corruption with a simple RocksDB test.
>>>>>
>>>>> Might be that the f2fs_forced_buffered_io path brings some problems too,
>>>>> but I have not seen other failures besides this one.
>>>>>
>>>>> Problem reproducible without a zoned block device, simply by forcing
>>>>> LFS mount:
>>>>>
>>>>>   $ sudo mkfs.f2fs -f -m /dev/nvme0n1
>>>>>   $ sudo mount /dev/nvme0n1 /mnt/f2fs
>>>>>   $ sudo  /opt/rocksdb/db_bench  --benchmarks=fillseq --use_existing_db=0
>>>>>   --use_direct_reads=true --use_direct_io_for_flush_and_compaction=true
>>>>>   --db=/mnt/f2fs --num=5000 --value_size=1048576 --verify_checksum=1
>>>>>   --block_size=65536
>>>>>
>>>>> Note that the options that cause the problem are:
>>>>>   --use_direct_reads=true --use_direct_io_for_flush_and_compaction=true
>>>>>
>>>>> Fixes: f9d6d0597698 ("f2fs: fix out-place-update DIO write")
>>>>>
>>>>> Signed-off-by: Javier González <javier.gonz@...sung.com>
>>>>> ---
>>>>>  fs/f2fs/data.c | 3 ---
>>>>>  1 file changed, 3 deletions(-)
>>>>>
>>>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
>>>>> index 5755e897a5f0..b045dd6ab632 100644
>>>>> --- a/fs/f2fs/data.c
>>>>> +++ b/fs/f2fs/data.c
>>>>> @@ -1081,9 +1081,6 @@ int f2fs_preallocate_blocks(struct kiocb *iocb, struct iov_iter *from)
>>>>>  			return err;
>>>>>  	}
>>>>>
>>>>> -	if (direct_io && allow_outplace_dio(inode, iocb, from))
>>>>> -		return 0;
>>>>
>>>> Since for LFS mode, all DIOs can end up out of place, I think that it
>>>> may be better to change allow_outplace_dio() to always return true in
>>>> the case of LFS mode. So may be something like:
>>>>
>>>> static inline int allow_outplace_dio(struct inode *inode,
>>>> 			struct kiocb *iocb, struct iov_iter *iter)
>>>> {
>>>> 	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
>>>> 	int rw = iov_iter_rw(iter);
>>>>
>>>> 	return test_opt(sbi, LFS) ||
>>>> 	 	(rw == WRITE && !block_unaligned_IO(inode, iocb, iter));
>>>> }
>>>>
>>>> instead of the original:
>>>>
>>>> static inline int allow_outplace_dio(struct inode *inode,
>>>> 			struct kiocb *iocb, struct iov_iter *iter)
>>>> {
>>>> 	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
>>>> 	int rw = iov_iter_rw(iter);
>>>>
>>>> 	return (test_opt(sbi, LFS) && (rw == WRITE) &&
>>>> 				!block_unaligned_IO(inode, iocb, iter));
>>>> }
>>>>
>>>> Thoughts ?
>>>>
>>>
>>> I see what you mean and it makes sense. However, the problem I am seeing
>>> occurs when allow_outplace_dio() returns true, as this is what creates
>>> the inconsistency between the write being buffered and the read being
>>> DIO.
>>
>> But if the write is switched to buffered, the DIO read should use the
>> buffered path too, no ? Since this is all happening under VFS, the
>> generic DIO read path will not ensure that the buffered writes are
>> flushed to disk before issuing the direct read, I think. So that would
>> explain your data corruption, i.e. you are reading stale data on the
>> device before the buffered writes make it to the media.
>>
> 
> As far as I can see, the read is always sent DIO, so yes, I also believe
> that we are reading stale data. This is why the corruption is not seen
> if preventing allow_outplace_dio() from sending the write to the
> buffered path.
> 
> What surprises me is that this is very easy to trigger (see commit), so
> I assume you must have seen this with SMR in the past.

We just did. Shin'Ichiro in my team finally succeeded in recreating the
problem. The cause seems to be:

bool direct_io = iocb->ki_flags & IOCB_DIRECT;

being true on entry of f2fs_preallocate_blocks() whereas
f2fs_direct_IO() forces buffered IO path for DIO on zoned devices with:

if (f2fs_force_buffered_io(inode, iocb, iter))
		return 0;

which has:

	if (f2fs_sb_has_blkzoned(sbi))
		return true;

So the top DIO code says "do buffered IOs", but lower in the write path,
the IO is still assumed to be a DIO because of the iocb flag... That's
inconsistent.

Note that for the non-zoned device LFS case, f2fs_force_buffered_io()
returns true only for unaligned write DIOs... But that will still trip
on the iocb flag test. So the proper fix is likely something like:

int f2fs_preallocate_blocks(struct kiocb *iocb, struct iov_iter *from)
{
	struct inode *inode = file_inode(iocb->ki_filp);
	struct f2fs_map_blocks map;
	int flag;
	int err = 0;
-	bool direct_io = iocb->ki_flags & IOCB_DIRECT;
+	bool direct_io = (iocb->ki_flags & IOCB_DIRECT) &&
+		!2fs_force_buffered_io(inode, iocb, iter);

	/* convert inline data for Direct I/O*/
	if (direct_io) {
		err = f2fs_convert_inline_inode(inode);
		if (err)
			return err;
	}

Shin'Ichiro tried this on SMR disks and the failure is gone...

Cheers.


> 
> Does it make sense to leave the LFS check out of the
> allow_outplace_dio()? Or in other words, is there a hard requirement for
> writes to take this path on a zoned device that I am not seeing?
> Something like:
> 
>    static inline int allow_outplace_dio(struct inode *inode,
>    			struct kiocb *iocb, struct iov_iter *iter)
>    {
>    	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
>    	int rw = iov_iter_rw(iter);
> 
>    	return (rw == WRITE && !block_unaligned_IO(inode, iocb, iter));
>    }
> 
> Thanks,
> Javier
> 


-- 
Damien Le Moal
Western Digital Research

Powered by blists - more mailing lists