On 2025-05-19 21:37:04, Dave Chinner wrote:
> On Thu, May 15, 2025 at 12:33:31PM +0200, Amir Goldstein wrote:
> > On Thu, May 15, 2025 at 11:02 AM Christian Brauner <brau...@kernel.org>
> > wrote:
> > >
> > > On Tue, May 13, 2025 at 11:53:23AM +0200, Arnd Bergmann wrote:
> > > > On Tue, May 13, 2025, at 11:17, Andrey Albershteyn wrote:
> > > >
> > > > >
> > > > > long syscall(SYS_file_getattr, int dirfd, const char *pathname,
> > > > > struct fsxattr *fsx, size_t size, unsigned int at_flags);
> > > > > long syscall(SYS_file_setattr, int dirfd, const char *pathname,
> > > > > struct fsxattr *fsx, size_t size, unsigned int at_flags);
> > > >
> > > > I don't think we can have both the "struct fsxattr" from the uapi
> > > > headers, and a variable size as an additional argument. I would
> > > > still prefer not having the extensible structure at all and just
> > >
> > > We're not going to add new interfaces that are fixed size unless for the
> > > very basic cases. I don't care if we're doing that somewhere else in the
> > > kernel but we're not doing that for vfs apis.
> > >
> > > > use fsxattr, but if you want to make it extensible in this way,
> > > > it should use a different structure (name). Otherwise adding
> > > > fields after fsx_pad[] would break the ioctl interface.
> > >
> > > Would that really be a problem? Just along the syscall simply add
> > > something like:
> > >
> > > diff --git a/fs/ioctl.c b/fs/ioctl.c
> > > index c91fd2b46a77..d3943805c4be 100644
> > > --- a/fs/ioctl.c
> > > +++ b/fs/ioctl.c
> > > @@ -868,12 +868,6 @@ static int do_vfs_ioctl(struct file *filp, unsigned
> > > int fd,
> > > case FS_IOC_SETFLAGS:
> > > return ioctl_setflags(filp, argp);
> > >
> > > - case FS_IOC_FSGETXATTR:
> > > - return ioctl_fsgetxattr(filp, argp);
> > > -
> > > - case FS_IOC_FSSETXATTR:
> > > - return ioctl_fssetxattr(filp, argp);
> > > -
> > > case FS_IOC_GETFSUUID:
> > > return ioctl_getfsuuid(filp, argp);
> > >
> > > @@ -886,6 +880,20 @@ static int do_vfs_ioctl(struct file *filp, unsigned
> > > int fd,
> > > break;
> > > }
> > >
> > > + switch (_IOC_NR(cmd)) {
> > > + case _IOC_NR(FS_IOC_FSGETXATTR):
> > > + if (WARN_ON_ONCE(_IOC_TYPE(cmd) !=
> > > _IOC_TYPE(FS_IOC_FSGETXATTR)))
> > > + return SOMETHING_SOMETHING;
> > > + /* Only handle original size. */
> > > + return ioctl_fsgetxattr(filp, argp);
> > > +
> > > + case _IOC_NR(FFS_IOC_FSSETXATTR):
> > > + if (WARN_ON_ONCE(_IOC_TYPE(cmd) !=
> > > _IOC_TYPE(FFS_IOC_FSSETXATTR)))
> > > + return SOMETHING_SOMETHING;
> > > + /* Only handle original size. */
> > > + return ioctl_fssetxattr(filp, argp);
> > > + }
> > > +
> >
> > I think what Arnd means is that we will not be able to change struct
> > sfxattr in uapi
> > going forward, because we are not going to deprecate the ioctls and
>
> There's no need to deprecate anything to rev an ioctl API. We have
> had to solve this "changing struct size" problem previously in XFS
> ioctls. See XFS_IOC_FSGEOMETRY and the older XFS_IOC_FSGEOMETRY_V4
> and XFS_IOC_FSGEOMETRY_V1 versions of the API/ABI.
>
> If we need to increase the structure size, we can rename the existing
> ioctl and struct to fix the version in the API, then use the
> original name for the new ioctl and structure definition.
>
> The only thing we have to make sure of is that the old and new
> structures have exactly the same overlapping structure. i.e.
> extension must always be done by appending new varibles, they can't
> be put in the middle of the structure.
>
> This way applications being rebuild will pick up the new definition
> automatically when the system asserts that it is suppored, whilst
> existing binaries will always still be supported by the kernel.
>
> If the application wants/needs to support all possible kernels, then
> if XFS_IOC_FSGEOMETRY is not supported, call XFS_IOC_FSGEOMETRY_V4,
> and if that fails (only on really old irix!) or you only need
> something in that original subset, call XFS_IOC_FSGEOMETRY_V1 which
> will always succeed....
>
> > Should we will need to depart from this struct definition and we might
> > as well do it for the initial release of the syscall rather than later on,
> > e.g.:
> >
> > --- a/include/uapi/linux/fs.h
> > +++ b/include/uapi/linux/fs.h
> > @@ -148,6 +148,17 @@ struct fsxattr {
> > unsigned char fsx_pad[8];
> > };
> >
> > +/*
> > + * Variable size structure for file_[sg]et_attr().
> > + */
> > +struct fsx_fileattr {
> > + __u32 fsx_xflags; /* xflags field value (get/set) */
> > + __u32 fsx_extsize; /* extsize field value (get/set)*/
> > + __u32 fsx_nextents; /* nextents field value (get) */
> > + __u32 fsx_projid; /* project identifier (get/set) */
> > + __u32 fsx_cowextsize; /* CoW extsize field value
> > (get/set)*/
> > +};
> > +
> > +#define FSXATTR_SIZE_VER0 20
> > +#define FSXATTR_SIZE_LATEST FSXATTR_SIZE_VER0
>
> If all the structures overlap the same, all that is needed in the
> code is to define the structure size that should be copied in and
> parsed. i.e:
>
> case FSXATTR..._V1:
> return ioctl_fsxattr...(args, sizeof(fsx_fileattr_v1));
> case FSXATTR..._V2:
> return ioctl_fsxattr...(args, sizeof(fsx_fileattr_v2));
> case FSXATTR...:
> return ioctl_fsxattr...(args, sizeof(fsx_fileattr));
>
> -Dave.
> --
> Dave Chinner
> da...@fromorbit.com
>
So, looks like there's at least two solutions to this concern.
Considering also that we have a bit of space in fsxattr,
'fsx_pad[8]', I think it's fine to stick with the current fsxattr
for now.
--
- Andrey
