linux-kernel - Re: [PATCH v5 3/7] fs: iomap: Atomic write support

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <112ec3a6-48b3-4596-9c20-e23288581ffd@oracle.com>
Date: Fri, 30 Aug 2024 16:48:36 +0100
From: John Garry <john.g.garry@...cle.com>
To: "Darrick J. Wong" <djwong@...nel.org>
Cc: axboe@...nel.dk, brauner@...nel.org, viro@...iv.linux.org.uk, jack@...e.cz,
        chandan.babu@...cle.com, dchinner@...hat.com, hch@....de,
        linux-block@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-xfs@...r.kernel.org, linux-fsdevel@...r.kernel.org, hare@...e.de,
        martin.petersen@...cle.com, catherine.hoang@...cle.com,
        kbusch@...nel.org
Subject: Re: [PATCH v5 3/7] fs: iomap: Atomic write support

On 22/08/2024 21:30, Darrick J. Wong wrote:
>> Then, the iomap->type/flag is either IOMAP_UNWRITTEN/IOMAP_F_DIRTY or
>> IOMAP_MAPPED/IOMAP_F_DIRTY per iter. So the type is not consistent. However
>> we will set IOMAP_DIO_UNWRITTEN in dio->flags, so call xfs_dio_write_endio()
>> -> xfs_iomap_write_unwritten() for the complete FSB range.
>>
>> Do you see a problem with this?

Sorry again for the slow response.

>>
>> Please see this also for some more background:
>> https://urldefense.com/v3/__https://lore.kernel.org/linux- 
>> xfs/20240726171358.GA27612@....de/__;!!ACWV5N9M2RV99hQ! 
>> P5jeP96F8wAtRAblbm8NvRo8nlpil03vA26UMMX8qrYa4IzKecAAk7x1l1M45bBshC3Czxn1CkDXypNSAg$ 
> Yes -- if you have a mix of written and unwritten blocks for the same
> chunk of physical space:
> 
> 0      7
> WUWUWUWU
> 
> the directio ioend function will start four separate transactions to
> convert blocks 1, 3, 5, and 7 to written status.  If the system crashes
> midway through, they will see this afterwards:
> 
> WWWWW0W0
> 
> IOWs, although the*disk write* was completed successfully, the mapping
> updates were torn, and the user program sees a torn write.
 > > The most performant/painful way to fix this would be to make the whole
> ioend completion a logged operation so that we could commit to updating
> all the unwritten mappings and restart it after a crash.

could we make it logged for those special cases which we are interested 
in only?

> 
> The least performant of course is to write zeroes at allocation time,
> like we do for fsdax.

That idea was already proposed:
https://lore.kernel.org/linux-xfs/ZcGIPlNCkL6EDx3Z@dread.disaster.area/

> 
> A possible middle ground would be to detect IOMAP_ATOMIC in the
> ->iomap_begin method, notice that there are mixed mappings under the
> proposed untorn IO, and pre-convert the unwritten blocks by writing
> zeroes to disk and updating the mappings 

Won't that have the same issue as using XFS_BMAPI_ZERO, above i.e. 
zeroing during allocation?

> before handing the one single
> mapping back to iomap_dio_rw to stage the untorn writes bio.  At least
> you'd only be suffering that penalty for the (probable) corner case of
> someone creating mixed mappings.

BTW, one issue I have with the sub-extent(or -alloc unit) zeroing from 
v4 series is how the unwritten conversion has changed, like:

xfs_iomap_write_unwritten()
{
	unsigned int rounding;

	/* when converting anything unwritten, we must be spanning an alloc 
unit, so round up/down */
	if (rounding > 1) {
		offset_fsb = rounddown(rounding);
		count_fsb = roundup(rounding);
	}

	...
	do {
		xfs_bmapi_write();
		...
		xfs_trans_commit();
	} while ();
}

I'm not too happy with it and it seems a bit of a bodge, as I would 
rather we report the complete size written (user data and zeroes); then 
xfs_iomap_write_unwritten() would do proper individual block conversion. 
However, we do something similar for zeroing for sub-FSB writes. I am 
not sure if that is the same thing really, as we only round up to FSB 
size. Opinion?

Thanks,
John