From: Damien Le Moal <Damien.LeMoal@wdc.com>
To: "Javier González" <javier@javigon.com>
Cc: "jaegeuk@kernel.org" <jaegeuk@kernel.org>,
"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
"Javier González" <javier.gonz@samsung.com>,
"linux-f2fs-devel@lists.sourceforge.net"
<linux-f2fs-devel@lists.sourceforge.net>
Subject: Re: [f2fs-dev] [PATCH] f2fs: disble physical prealloc in LSF mount
Date: Tue, 26 Nov 2019 06:19:46 +0000 [thread overview]
Message-ID: <BYAPR04MB581676157DCF909EDF1AAAFCE7450@BYAPR04MB5816.namprd04.prod.outlook.com> (raw)
In-Reply-To: 20191126035726.xj7pierxsck6adow@mpHalley
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@samsung.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@samsung.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
_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
next prev parent reply other threads:[~2019-11-26 6:19 UTC|newest]
Thread overview: 8+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-11-22 8:59 [f2fs-dev] [PATCH] f2fs: disble physical prealloc in LSF mount Javier González
2019-11-25 0:48 ` Damien Le Moal
2019-11-25 19:03 ` Javier González
2019-11-26 2:06 ` Damien Le Moal
2019-11-26 3:57 ` Javier González
2019-11-26 6:19 ` Damien Le Moal [this message]
2019-11-26 6:20 ` Damien Le Moal
2019-11-26 6:43 ` Javier González
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=BYAPR04MB581676157DCF909EDF1AAAFCE7450@BYAPR04MB5816.namprd04.prod.outlook.com \
--to=damien.lemoal@wdc.com \
--cc=jaegeuk@kernel.org \
--cc=javier.gonz@samsung.com \
--cc=javier@javigon.com \
--cc=linux-f2fs-devel@lists.sourceforge.net \
--cc=linux-kernel@vger.kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).