Dear friends, This patch series aimed to provide support for idmapped mounts for fuse & virtiofs. We already have idmapped mounts support for almost all widely-used filesystems: * local (ext4, btrfs, xfs, fat, vfat, ntfs3, squashfs, f2fs, erofs, ZFS (out-of-tree)) * network (ceph) Git tree (based on torvalds/master): v3: https://github.com/mihalicyn/linux/commits/fuse_idmapped_mounts.v3 current: https://github.com/mihalicyn/linux/commits/fuse_idmapped_mounts Changelog for version 3: - introduce and use a new SB_I_NOIDMAP flag (suggested by Christian) - add support for virtiofs (+user space virtiofsd conversion) Changelog for version 2: - removed "fs/namespace: introduce fs_type->allow_idmap hook" and simplified logic to return -EIO if a fuse daemon does not support idmapped mounts (suggested by Christian Brauner) - passed an "idmap" in more cases even when it's not necessary to simplify things (suggested by Christian Brauner) - take ->rename() RENAME_WHITEOUT into account and forbid it for idmapped mount case Links to previous versions: v2: https://lore.kernel.org/linux-fsdevel/20240814114034.113953-1-aleksandr.mikhalitsyn@xxxxxxxxxxxxx tree: https://github.com/mihalicyn/linux/commits/fuse_idmapped_mounts.v2 v1: https://lore.kernel.org/all/20240108120824.122178-1-aleksandr.mikhalitsyn@xxxxxxxxxxxxx/#r tree: https://github.com/mihalicyn/linux/commits/fuse_idmapped_mounts.v1 Having fuse (+virtiofs) supported looks like a good next step. At the same time fuse conceptually close to the network filesystems and supporting it is a quite challenging task. Let me briefly explain what was done in this series and which obstacles we have. With this series, you can use idmapped mounts with fuse if the following conditions are met: 1. The filesystem daemon declares idmap support (new FUSE_INIT response feature flags FUSE_OWNER_UID_GID_EXT and FUSE_ALLOW_IDMAP) 2. The filesystem superblock was mounted with the "default_permissions" parameter 3. The filesystem fuse daemon does not perform any UID/GID-based checks internally and fully trusts the kernel to do that (yes, it's almost the same as 2.) I have prepared a bunch of real-world examples of the user space modifications that can be done to use this extension: - libfuse support https://github.com/mihalicyn/libfuse/commits/idmap_support - fuse-overlayfs support: https://github.com/mihalicyn/fuse-overlayfs/commits/idmap_support - cephfs-fuse conversion example https://github.com/mihalicyn/ceph/commits/fuse_idmap - glusterfs conversion example (there is a conceptual issue) https://github.com/mihalicyn/glusterfs/commits/fuse_idmap - virtiofsd conversion example https://gitlab.com/virtio-fs/virtiofsd/-/merge_requests/245 The glusterfs is a bit problematic, unfortunately, because even if the glusterfs superblock was mounted with the "default_permissions" parameter (1 and 2 conditions are satisfied), it fails to satisfy the 3rd condition. The glusterfs fuse daemon sends caller UIDs/GIDs over the wire and all the permission checks are done twice (first on the client side (in the fuse kernel module) and second on the glusterfs server side). Just for demonstration's sake, I found a hacky (but working) solution for glusterfs that disables these server-side permission checks (see [1]). This allows you to play with the filesystem and idmapped mounts and it works just fine. The problem described above is the main problem that we can meet when working on idmapped mounts support for network-based filesystems (or network-like filesystems like fuse). When people look at the idmapped mounts feature at first they tend to think that idmaps are for faking caller UIDs/GIDs, but that's not the case. There was a big discussion about this in the "ceph: support idmapped mounts" patch series [2], [3]. The brief outcome from this discussion is that we don't want and don't have to fool filesystem code and map a caller's UID/GID everywhere, but only in VFS i_op's which are provided with a "struct mnt_idmap *idmap"). For example ->lookup() callback is not provided with it and that's on purpose! We don't expect the low-level filesystem code to do any permissions checks inside this callback because everything was already checked on the higher level (see may_lookup() helper). For local filesystems this assumption works like a charm, but for network-based, unfortunately, not. For example, the cephfs kernel client *always* send called UID/GID with *any* request (->lookup included!) and then *may* (depending on the MDS configuration) perform any permissions checks on the server side based on these values, which obviously leads to issues/inconsistencies if VFS idmaps are involved. Fuse filesystem very-very close to cephfs example, because we have req->in.h.uid/req->in.h.gid and these values are present in all fuse requests and userspace may use them as it wants. All of the above explains why we have a "default_permissions" requirement. If filesystem does not use it, then permission checks will be widespread across all the i_op's like ->lookup, ->unlink, ->readlink instead of being consolidated in the one place (->permission callback). In this series, my approach is the same as in cephfs [4], [5]. Don't touch req->in.h.uid/req->in.h.gid values at all (because we can't properly idmap them as we don't have "struct mnt_idmap *idmap" everywhere), instead, provide the userspace with a new optional (FUSE_OWNER_UID_GID_EXT) UID/GID suitable only for ->mknod, ->mkdir, ->symlink, ->atomic_open and these values have to be used as the owner UID and GID for newly created inodes. Things to discuss: - we enable idmapped mounts support only if "default_permissions" mode is enabled, because otherwise, we would need to deal with UID/GID mappings on the userspace side OR provide the userspace with idmapped req->in.h.uid/req->in.h.gid values which is not something that we probably want to do. Idmapped mounts philosophy is not about faking caller uid/gid. How to play with it: 1. take any patched filesystem from the list (fuse-overlayfs, cephfs-fuse, glusterfs) and mount it 2. ./mount-idmapped --map-mount b:1000:0:2 /mnt/my_fuse_mount /mnt/my_fuse_mount_idmapped (maps UID/GIDs as 1000 -> 0, 1001 -> 1) [ taken from https://raw.githubusercontent.com/brauner/mount-idmapped/master/mount-idmapped.c ] [1] https://github.com/mihalicyn/glusterfs/commit/ab3ec2c7cbe22618cba9cc94a52a492b1904d0b2 [2] https://lore.kernel.org/lkml/20230608154256.562906-1-aleksandr.mikhalitsyn@xxxxxxxxxxxxx/ [3] https://lore.kernel.org/lkml/CAEivzxfw1fHO2TFA4dx3u23ZKK6Q+EThfzuibrhA3RKM=ZOYLg@xxxxxxxxxxxxxx/ [4] https://github.com/ceph/ceph/pull/52575 [5] https://lore.kernel.org/all/20230807132626.182101-4-aleksandr.mikhalitsyn@xxxxxxxxxxxxx/ Thanks! Alex Cc: Christian Brauner <brauner@xxxxxxxxxx> Cc: Seth Forshee <sforshee@xxxxxxxxxx> Cc: Miklos Szeredi <miklos@xxxxxxxxxx> Cc: Vivek Goyal <vgoyal@xxxxxxxxxx> Cc: German Maglione <gmaglione@xxxxxxxxxx> Cc: Amir Goldstein <amir73il@xxxxxxxxx> Cc: Bernd Schubert <bschubert@xxxxxxx> Cc: <linux-fsdevel@xxxxxxxxxxxxxxx> Signed-off-by: Alexander Mikhalitsyn <aleksandr.mikhalitsyn@xxxxxxxxxxxxx> Alexander Mikhalitsyn (11): fs/namespace: introduce SB_I_NOIDMAP flag fs/fuse: add FUSE_OWNER_UID_GID_EXT extension fs/fuse: support idmap for mkdir/mknod/symlink/create fs/fuse: support idmapped getattr inode op fs/fuse: support idmapped ->permission inode op fs/fuse: support idmapped ->setattr op fs/fuse: drop idmap argument from __fuse_get_acl fs/fuse: support idmapped ->set_acl fs/fuse: properly handle idmapped ->rename op fs/fuse: allow idmapped mounts fs/fuse/virtio_fs: allow idmapped mounts fs/fuse/acl.c | 10 ++- fs/fuse/dir.c | 146 +++++++++++++++++++++++++------------- fs/fuse/file.c | 2 +- fs/fuse/fuse_i.h | 7 +- fs/fuse/inode.c | 16 ++++- fs/fuse/virtio_fs.c | 1 + fs/namespace.c | 4 ++ include/linux/fs.h | 1 + include/uapi/linux/fuse.h | 24 ++++++- 9 files changed, 148 insertions(+), 63 deletions(-) -- 2.34.1