Re: [PATCH v4 0/3] fs: introduce getfsxattrat and setfsxattrat syscalls

[Amir Goldstein]

On Thu, Mar 27, 2025 at 10:13 PM Pali Rohár <pali@xxxxxxxxxx> wrote:
>
> On Thursday 27 March 2025 21:57:34 Amir Goldstein wrote:
> > On Thu, Mar 27, 2025 at 8:26 PM Pali Rohár <pali@xxxxxxxxxx> wrote:
> > >
> > > On Thursday 27 March 2025 12:47:02 Amir Goldstein wrote:
> > > > On Sun, Mar 23, 2025 at 11:32 AM Pali Rohár <pali@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Sunday 23 March 2025 09:45:06 Amir Goldstein wrote:
> > > > > > On Fri, Mar 21, 2025 at 8:50 PM Andrey Albershteyn <aalbersh@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > This patchset introduced two new syscalls getfsxattrat() and
> > > > > > > setfsxattrat(). These syscalls are similar to FS_IOC_FSSETXATTR ioctl()
> > > > > > > except they use *at() semantics. Therefore, there's no need to open the
> > > > > > > file to get an fd.
> > > > > > >
> > > > > > > These syscalls allow userspace to set filesystem inode attributes on
> > > > > > > special files. One of the usage examples is XFS quota projects.
> > > > > > >
> > > > > > > XFS has project quotas which could be attached to a directory. All
> > > > > > > new inodes in these directories inherit project ID set on parent
> > > > > > > directory.
> > > > > > >
> > > > > > > The project is created from userspace by opening and calling
> > > > > > > FS_IOC_FSSETXATTR on each inode. This is not possible for special
> > > > > > > files such as FIFO, SOCK, BLK etc. Therefore, some inodes are left
> > > > > > > with empty project ID. Those inodes then are not shown in the quota
> > > > > > > accounting but still exist in the directory. This is not critical but in
> > > > > > > the case when special files are created in the directory with already
> > > > > > > existing project quota, these new inodes inherit extended attributes.
> > > > > > > This creates a mix of special files with and without attributes.
> > > > > > > Moreover, special files with attributes don't have a possibility to
> > > > > > > become clear or change the attributes. This, in turn, prevents userspace
> > > > > > > from re-creating quota project on these existing files.
> > > > > > >
> > > > > > > Christian, if this get in some mergeable state, please don't merge it
> > > > > > > yet. Amir suggested these syscalls better to use updated struct fsxattr
> > > > > > > with masking from Pali Rohár patchset, so, let's see how it goes.
> > > > > >
> > > > > > Andrey,
> > > > > >
> > > > > > To be honest I don't think it would be fair to delay your syscalls more
> > > > > > than needed.
> > > > >
> > > > > I agree.
> > > > >
> > > > > > If Pali can follow through and post patches on top of your syscalls for
> > > > > > next merge window that would be great, but otherwise, I think the
> > > > > > minimum requirement is that the syscalls return EINVAL if fsx_pad
> > > > > > is not zero. we can take it from there later.
> > > > >
> > > > > IMHO SYS_getfsxattrat is fine in this form.
> > > > >
> > > > > For SYS_setfsxattrat I think there are needed some modifications
> > > > > otherwise we would have problem again with backward compatibility as
> > > > > is with ioctl if the syscall wants to be extended in future.
> > > > >
> > > > > I would suggest for following modifications for SYS_setfsxattrat:
> > > > >
> > > > > - return EINVAL if fsx_xflags contains some reserved or unsupported flag
> > > > >
> > > > > - add some flag to completely ignore fsx_extsize, fsx_projid, and
> > > > >   fsx_cowextsize fields, so SYS_setfsxattrat could be used just to
> > > > >   change fsx_xflags, and so could be used without the preceding
> > > > >   SYS_getfsxattrat call.
> > > > >
> > > > > What do you think about it?
> > > >
> > > > I think all Andrey needs to do now is return -EINVAL if fsx_pad is not zero.
> > > >
> > > > You can use this later to extend for the semantics of flags/fields mask
> > > > and we can have a long discussion later on what this semantics should be.
> > > >
> > > > Right?
> > > >
> > > > Amir.
> > >
> > > It is really enough?
> >
> > I don't know. Let's see...
> >
> > > All new extensions later would have to be added
> > > into fsx_pad fields, and currently unused bits in fsx_xflags would be
> > > unusable for extensions.
> >
> > I am working under the assumption that the first extension would be
> > to support fsx_xflags_mask and from there, you could add filesystem
> > flags support checks and then new flags. Am I wrong?
> >
> > Obviously, fsx_xflags_mask would be taken from fsx_pad space.
> > After that extension is implemented, calling SYS_setfsxattrat() with
> > a zero fsx_xflags_mask would be silly for programs that do not do
> > the legacy get+set.
> >
> > So when we introduce  fsx_xflags_mask, we could say that a value
> > of zero means that the mask is not being checked at all and unknown
> > flags in set syscall are ignored (a.k.a legacy ioctl behavior).
> >
> > Programs that actually want to try and set without get will have to set
> > a non zero fsx_xflags_mask to do something useful.
>
> Here we need to also solve the problem that without GET call we do not
> have valid values for fsx_extsize, fsx_projid, and fsx_cowextsize. So
> maybe we would need some flag in fsx_pad that fsx_extsize, fsx_projid,
> or fsx_cowextsize are ignored/masked.
>
> > I don't think this is great.
> > I would rather that the first version of syscalls will require the mask
> > and will always enforce filesystems supported flags.
>
> It is not great... But what about this? In a first step (part of this
> syscall patch series) would be just a check that fsx_pad is zero.
> Non-zero will return -EINVAL.
>
> In next changes would added fsx_filter bit field, which for each
> fsx_xflags and also for fsx_extsize, fsx_projid, and fsx_cowextsize
> fields would add a new bit flag which would say (when SET) that the
> particular thing has to be ignored.

1. I don't like the inverse mask. statx already has the stx_mask
    and stx_attributes_mask, so I rather stick to same semantics
    because some of those attributes are exposed via statx as well
2. fsx_*extsize already have a bit that says if that the particular
    attribute is valid or not, so setting a zero fsx_cowextsize with the
    flag FS_XFLAG_COWEXTSIZE has no effect in xfs:

        /*
         * Only set the extent size hint if we've already determined that the
         * extent size hint should be set on the inode. If no extent size flags
         * are set on the inode then unconditionally clear the extent size hint.
         */
        if (ip->i_diflags & (XFS_DIFLAG_EXTSIZE | XFS_DIFLAG_EXTSZINHERIT))
                ip->i_extsize = XFS_B_TO_FSB(mp, fa->fsx_extsize);
        else
                ip->i_extsize = 0;

        if (xfs_has_v3inodes(mp)) {
                if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
                        ip->i_cowextsize = XFS_B_TO_FSB(mp, fa->fsx_cowextsize);
                else
                        ip->i_cowextsize = 0;
        }

I think we need to enforce this logic in fileattr_set_prepare()
and I think we need to add a flag FS_XFLAG_PROJID
that will be set in GET when fsx_projid != 0 and similarly
required when setting fsx_projid != 0.

Probably will need to add some backward compat glue for this
flag in GET ioctl to avoid breaking out of tree fs and fuse.

>
> So when fsx_pad is all-zeros then fsx_filter (first field in fsx_pad)
> would say that nothing in fsx_xflags, fsx_extsize, fsx_projid, and
> fsx_cowextsize is ignored, and hence behave like before.
>
> And when something in fsx_pad/fsx_filter is set then it says which
> fields are ignored/filtered-out.
>
> > If you can get those patches (on top of current series) posted and
> > reviewed in time for the next merge window, including consensus
> > on the actual semantics, that would be the best IMO.
>
> I think that this starting to be more complicated to rebase my patches
> in a way that they do not affect IOCTL path but implement it properly
> for new syscall path. It does not sounds like a trivial thing which I
> would finish in merge window time and having proper review and consensus
> on this.
>

Yes, it is better to separate the two efforts.

wrt erroring on unsupported SET flags, all fs other than xfs already
have some variant of fileattr_has_fsx(), so xfs is the only filesystem
that requires special care with the new syscalls.
It's easier to write a patch than it is to explain what I mean, so
I'll try to write a patch.

Thanks,
Amir.