Re: [PATCH RFC v2 0/6] ext4: yet another project quota

[Konstantin Khlebnikov]

On Mon, Mar 16, 2015 at 7:52 PM, Jan Kara <jack@xxxxxxx> wrote:
>   Hello,
>
> On Tue 10-03-15 20:22:04, Konstantin Khlebnikov wrote:
>> Projects quota allows to enforce disk quota for several subtrees or even
>> individual files on the filesystem. Each inode is marked with project-id
>> (independently from uid and gid) and accounted into corresponding project
>> quota. New files inherits project id from directory where they are created.
>>
>> This is must-have feature for deploying lightweight containers.
>> Also project quota can tell size of subtree without doing recursive 'du'.
>>
>> This patchset adds project id and quota into ext4.
>>
>> This time I've prepared patches also for e2fsprogs and quota-tools.
>>
>> All patches are available at github:
>> https://github.com/koct9i/linux --branch project
>> https://github.com/koct9i/e2fsprogs --branch project
>> https://github.com/koct9i/quota-tools --branch project
>>
>> Porposed behavior is similar to project quota in XFS:
>>
>> * all inode are marked with project id
>> * new files inherits project id from parent directory
>> * project quota accounts inodes and enforces limits
>> * cross-project link and rename operations are restricted
>
>   Thanks for the patches and sorry for taking a long time to look into
> them. I like your patches and they looks mostly fine to me but can we
> *please* start with making things completely compatible with how XFS works?
> I understand that may seem broken / not useful for your usecase but
> consistency among filesystems is really important.
>
> I was talking with Dave Chinner last week and we agreed that the direction
> you are changing things makes sense but getting things compatible with how
> they were for 15 years in XFS is an important first step (because there may
> be people with other usecases depending on the old behavior and they would
> get confused by ext4 behaving differently). After we have compatible code
> in, we can add your fs.protected_projects thing on top of that (and also
> support it in XFS to stay compatible).
>
>
>> Differences:
>>
>> There is no flag similar to XFS_XFLAG_PROJINHERIT (which allows to disable
>> project id inheritance), instead of that project which userspace sees as '0'
>> (in nested user-name space that might be actually non-zero project) acts as
>> default project where restrictions for link/rename are ignored.
>> (also see below, in "why new ioctl" part)
>>
>> This implementation adds shortcut for moving files from one project into
>> another: non-directory inodes with n_link == 1 are transferred without
>> copying (XFS in this case returns -EXDEV and userspace have to copy file).
>>
>> In XFS file owner (or process with CAP_FOWNER) can set set any project id,
>> XFS permits changing project id only from init user-namespace.
>>
>> This patchset adds sysctl fs.protected_projects. By default it's 0 and project
>> id acts as XFS project. Setting it to 1 makes chaning project id priviliged
>> operation which requires CAP_SYS_RESOURCE in current user-namespace, changing
>> project id mapping for nested user-namespace also requires that capability.
>> Thus there are two levels of control: project id mapping in user-ns defines set
>> of permitted projects and capability protects operations within this set.
>>
>> I see no problems with supporting all this in XFS, all difference in interface.
>>
>> Ext4 layout
>> -----------
>>
>> Project id introduce ro-compatible feature 'project'.
>>
>> Inode project id is stored in place of obsolete field 'i_faddr' (that trick was
>> suggested by Darrick J. Wong in previous discussions of project quota).
>> Linux never used that field and present fsck checks that it contains zero.
>>
>> Quota information is stored in special inode â11 (by default only 10 inodes are
>> reserved for special usage, I've add option to resize2fs to reserve more).
>> (see e2fsprogs patches for details) For symmetry with other quotas inode number
>> is stored in superblock.
>>
>> Project quota supports only modern 'hidden' journaled mode.
>>
>> Interface
>> ---------
>>
>> Interface for changing limits / query current usage is common vfs quotactl()
>> where quotatype = PRJQUOTA = 2. User can query current state of any project
>> mapped into user-ns, changing of limits requires CAP_SYS_ADMIN in init user-ns.
>>
>> Two new ioctls for getting / changing inode project id:
>> int ioctl(fd, FS_IOC_GETPROJECT, unsigned *project);
>> int ioctl(fd, FS_IOC_SETPROJECT, unsigned *project);
>>
>> They acts as interface for super-block methods get_project / set_project
>> Generic code checks permissions, does project id translation in user-namespace
>> mapping, grabs write-access to the filesystem, locks i_mutex for set opetaion.
>> Filesystem method only updates inode and transfers project quota.
>>
>> No new mount options added. Disk usage tracking is enabled at mount.
>> Limits are enabeld later by "quotaon".
>>
>> (BTW why journalled quota doesn't enable limits right at the time of mounting?)
>   Well, mostly historical reasons :) The biggest probably was that to
> properly initialize quotas (or fix them if quota file gets corrupted) you
> have to still run quotacheck and for that kernel mustn't touch the files.
> At the time when I was writing journaled quotas I didn't feel like adding
> quotacheck support into e2fsck... We eventually did that anyway for support
> of quota in system files but how we did that was actually based on the
> experience I had from the journaled quota lesson ;).
>
>> Why new ioctls?
>> ---------------
>>
>> XFS has it's own interface for that: XFS_IOC_FSGETXATTR / XFS_IOC_FSSETXATTR.
>> But it has several flaws and doesn't fit for a role of generic interface.
>>
>> It contains a lot of xfs-specific flags and racy by design: set operation
>> commits all fields at once thus it's used in sequence get-change-set without
>> any lock, Concurrent updates from user space will collide.

>   Sure but that's the case for generic GETFLAGS / SETFLAGS interface as
> well.

In this case we care only about project id. We can make interface clear,
non-racy and do it in just one syscall instead of two.

>
>> Also xfs has flag XFS_XFLAG_PROJINHERIT which tells should new files inherit
>> project id from parent directory or not. This flag is barely useful and only
>> makes everything complicated. Even tools in xfsprogs don't use it: they always
>> set it together with project id and clears when set project id back to zero.
>   I think this is about balance - adding the support for the PROJINHERIT flag
> costs us close to nothing and we get a compatibility with XFS. So even though
> the usefulness of that flag is doubtful, I think the additional
> compatibility is worth it.

There're only few free bits left. This bit gives nothing, xfsprogs can change it
but project quota management doesn't uses it. It gives nothing but complication.

Also this bit makes difference only for accounting directories itself
and it's internal
structures. This is pretty fs-specific thing.

>
>> And the main reason: this compatibility gives nothing. The only user of xfs
>> ioctl which I've found is the xfsprogs. And these tools check filesystem name
>> and don't work anywhere except 'xfs'.
>   The fact that you didn't find any other user doesn't mean there isn't
> any. And historically if we though "nobody could be relying on this", we
> were sometimes proven wrong - a few years later when it was *much* more
> painful to make things compatible.
>
> Here we speak about new feature for the filesystem so compatibility
> requirements aren't that strong but still I find the case that the same
> ioctl will just work on ext4 as it does on xfs more appealing than creating
> a new simpler generic ioctl. That way userspace doesn't have to care on
> which filesystem it is working or whether the current kernel already
> supports the new ioctl for XFS. So please use the XFS interface is Li Xi
> does in his patches.

Please consider this as _new_ feature which has compatible behaviour
from user's point of view. That's perfect time to fix its birth defects and
redesign behavior to satisfy expectations of new users. I do not care
about migration from xfs to ext4, that's not my case. I do not care about
mythical closed source software which might use this interface.

As I already said ioctl compatibility gives nothing but legacy problems.
Unexpected apperance of partial compatibility might be a nightmare.
And I've seen a lot when I worked with mixed containerized evironments.
it emerges untested and unexpected combinations of old and new software.
But, of course not in this case because nobody uses this interface except
xfsprogs. And as I said without update it works only with XFS.

Also that's ioctl is so poorly designed for generic usage so resulting code
looks like a shit. I just cannot add this into the kernel without any
good reason.

I've implemented interface and behaviour xfs-compatible enough to add
support of that into xfsprogs/xfstests in few lines. That's my next step after
polishing support in quota-tools.

>
>                                                                 Honza
>
>> Links
>> -----
>>
>> [1] 2014-12-09  ext4: add project quota support by Li Xi
>> http://marc.info/?l=linux-fsdevel&m=141810265603565&w=2
>>
>> [2] 2014-01-28  A draft for making ext4 support project quota by Zheng Liu
>> http://marc.info/?l=linux-ext4&m=139089109605795&w=2
>>
>> [3] 2012-07-09  introduce extended inode owner identifier v10 by Dmitry Monakhov
>> http://thread.gmane.org/gmane.linux.file-systems/65752
>>
>> [4] 2010-02-08  Introduce subtree quota support by Dmitry Monakhov
>> http://thread.gmane.org/gmane.comp.file-systems.ext4/17530
>>
>>
>> ---
>>
>> Konstantin Khlebnikov (6):
>>       fs: vfs ioctls for managing project id
>>       fs: protected project id
>>       quota: generic project quota
>>       ext4: support project id and project quota
>>       ext4: add shortcut for moving files across projects
>>       ext4: mangle statfs results accourding to project quota usage and limits
>>
>>
>>  Documentation/filesystems/Locking |    4 +
>>  Documentation/filesystems/vfs.txt |    8 +++
>>  Documentation/sysctl/fs.txt       |   16 ++++++
>>  fs/compat_ioctl.c                 |    2 +
>>  fs/ext4/ext4.h                    |   15 ++++-
>>  fs/ext4/ialloc.c                  |    3 +
>>  fs/ext4/inode.c                   |   15 +++++
>>  fs/ext4/namei.c                   |  102 ++++++++++++++++++++++++++++++++++++-
>>  fs/ext4/super.c                   |   61 ++++++++++++++++++++--
>>  fs/ioctl.c                        |   62 ++++++++++++++++++++++
>>  fs/quota/dquot.c                  |   96 +++++++++++++++++++++++++++++++++--
>>  fs/quota/quota.c                  |    8 ++-
>>  fs/quota/quotaio_v2.h             |    6 +-
>>  include/linux/fs.h                |    3 +
>>  include/linux/quota.h             |    1
>>  include/linux/quotaops.h          |   16 ++++++
>>  include/uapi/linux/capability.h   |    1
>>  include/uapi/linux/fs.h           |    3 +
>>  include/uapi/linux/quota.h        |    6 +-
>>  kernel/sysctl.c                   |    9 +++
>>  kernel/user_namespace.c           |    4 +
>>  21 files changed, 416 insertions(+), 25 deletions(-)
>>
>> --
>> Signature
> --
> Jan Kara <jack@xxxxxxx>
> SUSE Labs, CR
> --
> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
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/