Re: [f2fs-dev] [PATCH 2/2] f2fs: support revoking atomic written pages

From: Chao Yu
Date: Fri Jan 01 2016 - 07:14:39 EST


Hi Jaegeuk,

On 1/1/16 11:50 AM, Jaegeuk Kim wrote:
> Hi Chao,
>
> ...
>
>>>>> On Tue, Dec 29, 2015 at 11:12:36AM +0800, Chao Yu wrote:
>>>>>> f2fs support atomic write with following semantics:
>>>>>> 1. open db file
>>>>>> 2. ioctl start atomic write
>>>>>> 3. (write db file) * n
>>>>>> 4. ioctl commit atomic write
>>>>>> 5. close db file
>>>>>>
>>>>>> With this flow we can avoid file becoming corrupted when abnormal power
>>>>>> cut, because we hold data of transaction in referenced pages linked in
>>>>>> inmem_pages list of inode, but without setting them dirty, so these data
>>>>>> won't be persisted unless we commit them in step 4.
>>>>>>
>>>>>> But we should still hold journal db file in memory by using volatile write,
>>>>>> because our semantics of 'atomic write support' is not full, in step 4, we
>>>>>> could be fail to submit all dirty data of transaction, once partial dirty
>>>>>> data was committed in storage, db file should be corrupted, in this case,
>>>>>> we should use journal db to recover the original data in db file.
>>>>>
>>>>> Originally, IOC_ABORT_VOLATILE_WRITE was supposed to handle commit failures,
>>>>> since database should get its error literally.
>>>>>
>>>>> So, the only thing that we need to do is keeping journal data for further db
>>>>> recovery.
>>>>
>>>> IMO, if we really support *atomic* interface, we don't need any journal data
>>>> kept by user, because f2fs already have it in its storage since we always
>>>> trigger OPU for pages written in atomic-write opened file, f2fs can easily try
>>>> to revoke (replace old to new in metadata) when any failure exist in atomic
>>>> write process.
>>>
>>> Yeah, so current design does not fully support atomic writes. IOWs, volatile
>>> writes for journal files should be used together to minimize sqlite change as
>>> much as possible.
>>>
>>>> But in current design, we still hold journal data in memory for recovering for
>>>> *rare* failure case. I think there are several issues:
>>>> a) most of time, we are in concurrent scenario, so if large number of journal
>>>> db files were opened simultaneously, we are under big memory pressure.
>>>
>>> In current android, I've seen that this is not a big concern. Even there is
>>> memory pressure, f2fs flushes volatile pages.
>>
>> When I change to redirty all volatile pages in ->writepage, android seems go
>> into an infinite loop when doing recovery flow of f2fs data partition in startup.
>>
>> if (f2fs_is_volatile_file(inode))
>> goto redirty_out;
>
> Where did you put this? It doesn't flush at all? Why?

Original place in ->writepage, just remove two other conditions.

To avoid potential random writebacking of dirty page in journal which
cause unpredicted corrupting in journal.

> Practically, the peak amount of journal writes depend on how many transactions
> are processing concurrently.
> I mean, in-memory pages are dropped at the end of every transaction.
> You can check the number of pages through f2fs_stat on your phone.
>
>> I didn't dig details, but I think there may be a little risk for this design.
>>
>>>
>>>> b) If we are out of memory, reclaimer tries to write page of journal db into
>>>> disk, it will destroy db file.
>>>
>>> I don't understand. Could you elaborate why journal writes can corrupt db?
>>
>> Normally, we keep pages of journal in memory, but partial page in journal
>> will be write out to device by reclaimer when out of memory. So this journal
>> may have valid data in its log head, but with corrupted data, then after
>> abnormal powe-cut, recovery with this journal before a transaction will
>> destroy db. Right?
>
> Just think about sqlite without this feature.
> Broken journal is pretty normal case for sqlite.

Maybe, if it is caused by bug or design issue of software, no matter db system
or filesystem, we should try our best to fix it to avoid generating broken journals.

>
>>>
>>>> c) Though, we have journal db file, we will face failure of recovering db file
>>>> from journal db due to ENOMEM or EIO, then db file will be corrupted.
>>>
>>> Do you mean the failure of recovering db with a complete journal?
>>> Why do we have to handle that? That's a database stuff, IMO.
>>
>> Yes, just list for indicating we will face the same issue which is hard to
>> handle both in original design and new design, so the inner revoking failure
>> issue would not be a weak point or flaw of new design.
>>
>>>
>>>> d) Recovery flow will make data page dirty, triggering both data stream and
>>>> metadata stream, there should be more IOs than in inner revoking in
>>>> atomic-interface.
>>>
>>> Well, do you mean there is no need to recover db after revoking?
>>
>> Yes, revoking make the same effect like the recovery of sqlite, so after
>> revoking, recovery is no need.
>
> Logically, it doesn't make sense. If there is a valid journal file, it should
> redo the previous transaction. No?

As we know, in sqlite, before we commit a transaction, we will use journal to
record original data of pages which will be updated in following transaction, so
in following if a) abnormal power-cut, b) commit error, c) redo command was
triggered by user, we will recover db with journal.

Ideally, if we support atomic write interface, in there should always return two
status in atomic write interface: success or fail. If success, transaction was
committed, otherwise, it looks like nothing happened, user will be told
transaction was failed. Then, journals in sqlite could no longer be used,
eventually no journal, no recovery.

The only thing we should concern is inner failure (e.g. ENOMEM, ENOSPC) of
revoking in commit interface since it could destroy db file permanently w/o
journal. IMO, some optimization could be done for these cases:
1. ENOMEM: enable retrying or mark accessed flag in page in advance.
2. ENOSPC: preallocate blocks for node blocks and data blocks.

These optimizations couldn't guarantee no failure in revoking operation
completely, luckily, those are not common cases, and they also happen in sqlite
w/o atomic feature.

One more possible proposal is: if we support reflink feature like ocfs2/xfs, I
guess we can optimize DB like:
1. reflink db to db.ref
2. do transaction in db.ref
- failed, rm db.ref
- power-cut rm db.ref
3. rename db.ref to db

>
>> One more case is that user can send a command to abort current transaction,
>> it should be happened before atomic_commit operation, which could easily
>> handle with abort_commit ioctl.
>>
>>>
>>>> e) Moreover, there should be a hole between 1) commit fail and 2) abort write &
>>>> recover, checkpoint will persist the corrupt data in db file, following abnormal
>>>> power-cut will leave that data in disk.
>>>
>>> Yes, in that case, database should recover corrupted db with its journal file.
>>
>> Journal could be corrupted as I descripted in b).
>
> Okay, so what I'm thinking is like this.
> It seems there are two corruption cases after journal writes.
>
> 1. power cut during atomic writes
> - broken journal file and clean db file -> give up
> - luckily, valid journal file and clean db file -> recover db
>
> 2. error during atomic writes
> a. power-cut before abort completion
> - broken journal file and broken db file -> revoking is needed!
>
> b. after abort
> - valid journal file and broken db file -> recover db (likewise plain sqlite)
>
> Indeed, in the 2.a. case, we need revoking; I guess that's what you mentioned.
> But, I think, even if revoking is done, we should notify an error to abort and
> recover db by 2.b.
>
> Something like this after successful revoking.
>
> 1. power cut during atomic writes
> - broken journal file and clean db file -> give up
> - luckily, valid journal file and clean db file -> recover db
>
> 2. error during atomic writes w/ revoking
> a. power-cut before abort completion
> - broken journal file and clean db file -> give up
> - luckily, valid journal file and clean db file -> recover db
>
> b. after abort
> - valid journal file and clean db file -> recover db

That's right.

>
> Let me verify these scenarios first. :)

OK. :)

Thanks,

>
> Thanks,
>
>>>
>>>> With revoking supported design, we can not solve all above issues, we will still
>>>> face the same issue like c), but it will be a big improve if we can apply this
>>>> in our interface, since it provide a way to fix the issue a) b) d). And also for
>>>> e) case, we try to rescue data in first time that our revoking operation would be
>>>> protected by f2fs_lock_op to avoid checkpoint + power-cut.
>>>>
>>>> If you don't want to have a big change in this interface or recovery flow, how
>>>> about keep them both, and add a mount option to control inner recovery flow?
>>>
>>> Hmm, okay. I believe the current design is fine for sqlite in android.
>>
>> I believe new design will enhance in memory usage and error handling of sqlite
>> in android, and hope this can be applied. But, I can understand that if you
>> were considerring about risk control and backward compatibility, since this
>> change affects all atomic related ioctls.
>>
>>> For other databases, I can understand that they can use atomic_write without
>>> journal control, which is a sort of stand-alone atomic_write.
>>>
>>> It'd better to add a new ioctl for that, but before adding it, can we find
>>> any usecase for this feature? (e.g., postgresql, mysql, mariadb, couchdb?)
>>
>> You mean investigating or we can only start when there is a clear commercial
>> demand ?
>>
>>> Then, I expect that we can define a more appropriate and powerful ioctl.
>>
>> Agreed :)
>>
>> Thanks,
>>
>>>
>>> Thanks,
>>>
>>>>
>>>> How do you think? :)
>>>>
>>>> Thanks,
>>>>
>>>>> But, unfortunately, it seems that something is missing in the
>>>>> current implementation.
>>>>>
>>>>> So simply how about this?
>>>>>
>>>>> A possible flow would be:
>>>>> 1. write journal data to volatile space
>>>>> 2. write db data to atomic space
>>>>> 3. in the error case, call ioc_abort_volatile_writes for both journal and db
>>>>> - flush/fsync journal data to disk
>>>>> - drop atomic data, and will be recovered by database with journal
>>>>>
>>>>> From cb33fc8bc30981c370ec70fe68871130109793ec Mon Sep 17 00:00:00 2001
>>>>> From: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
>>>>> Date: Tue, 29 Dec 2015 15:46:33 -0800
>>>>> Subject: [PATCH] f2fs: fix f2fs_ioc_abort_volatile_write
>>>>>
>>>>> There are two rules to handle aborting volatile or atomic writes.
>>>>>
>>>>> 1. drop atomic writes
>>>>> - we don't need to keep any stale db data.
>>>>>
>>>>> 2. write journal data
>>>>> - we should keep the journal data with fsync for db recovery.
>>>>>
>>>>> Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
>>>>> ---
>>>>> fs/f2fs/file.c | 13 ++++++++++---
>>>>> 1 file changed, 10 insertions(+), 3 deletions(-)
>>>>>
>>>>> diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
>>>>> index 91f576a..d16438a 100644
>>>>> --- a/fs/f2fs/file.c
>>>>> +++ b/fs/f2fs/file.c
>>>>> @@ -1433,9 +1433,16 @@ static int f2fs_ioc_abort_volatile_write(struct file *filp)
>>>>> if (ret)
>>>>> return ret;
>>>>>
>>>>> - clear_inode_flag(F2FS_I(inode), FI_ATOMIC_FILE);
>>>>> - clear_inode_flag(F2FS_I(inode), FI_VOLATILE_FILE);
>>>>> - commit_inmem_pages(inode, true);
>>>>> + if (f2fs_is_atomic_file(inode)) {
>>>>> + clear_inode_flag(F2FS_I(inode), FI_ATOMIC_FILE);
>>>>> + commit_inmem_pages(inode, true);
>>>>> + }
>>>>> + if (f2fs_is_volatile_file(inode)) {
>>>>> + clear_inode_flag(F2FS_I(inode), FI_VOLATILE_FILE);
>>>>> + ret = commit_inmem_pages(inode, false);
>>>>> + if (!ret)
>>>>> + ret = f2fs_sync_file(filp, 0, LLONG_MAX, 0);
>>>>> + }
>>>>>
>>>>> mnt_drop_write_file(filp);
>>>>> return ret;
>>>>> --
>>>>> 2.6.3
>>>>
>
> ------------------------------------------------------------------------------
> _______________________________________________
> Linux-f2fs-devel mailing list
> Linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx
> https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/