When SELinux security options are passed to btrfs via fsconfig(2) rather than via mount(2), the operation aborts with an error. What happens is roughly this sequence: 1. vfs_parse_fs_param() eats away the LSM options and parses them into fc->security. 2. legacy_get_tree() finds nothing in ctx->legacy_data, passes this nothing to btrfs. [here btrfs calls another layer of vfs_kern_mount(), but let's ignore that for simplicity] 3. btrfs calls security_sb_set_mnt_opts() with empty options. 4. vfs_get_tree() then calls its own security_sb_set_mnt_opts() with the options stashed in fc->security. 5. SELinux doesn't like that different options were used for the same superblock and returns -EINVAL. In the case of mount(2), the options are parsed by legacy_parse_monolithic(), which skips the eating away of security opts because of the FS_BINARY_MOUNTDATA flag, so they are passed to the FS via ctx->legacy_data. The second call to security_sb_set_mnt_opts() (from vfs_get_tree()) now passes empty opts, but the non-empty -> empty sequence is allowed by SELinux for the FS_BINARY_MOUNTDATA case. It is a total mess, but the only sane fix for now seems to be to skip processing the security opts in vfs_parse_fs_param() if the fc has legacy opts set AND the fs specfies the FS_BINARY_MOUNTDATA flag. This combination currently matches only btrfs and coda. For btrfs this fixes the fsconfig(2) behavior, and for coda it makes setting security opts via fsconfig(2) fail the same way as it would with mount(2) (because FS_BINARY_MOUNTDATA filesystems are expected to call the mount opts LSM hooks themselves, but coda never cared enough to do that). I believe that is an acceptable state until both filesystems (or at least btrfs) are converted to the new mount API (at which point btrfs won't need to pretend it takes binary mount data any more and also won't need to call the LSM hooks itself, assuming it will pass the fc->security information properly). Note that we can't skip LSM opts handling in vfs_parse_fs_param() solely based on FS_BINARY_MOUNTDATA because that would break NFS. See here for the original report and reproducer: https://lore.kernel.org/selinux/c02674c970fa292610402aa866c4068772d9ad4e.camel@xxxxxxxxxxxxxx/ Reported-by: Richard Haines <richard_c_haines@xxxxxxxxxxxxxx> Fixes: 3e1aeb00e6d1 ("vfs: Implement a filesystem superblock creation/configuration context") Signed-off-by: Ondrej Mosnacek <omosnace@xxxxxxxxxx> --- fs/fs_context.c | 28 ++++++++++++++++++++++------ 1 file changed, 22 insertions(+), 6 deletions(-) diff --git a/fs/fs_context.c b/fs/fs_context.c index 2834d1afa6e80..cfc5ee2e381ef 100644 --- a/fs/fs_context.c +++ b/fs/fs_context.c @@ -106,12 +106,28 @@ int vfs_parse_fs_param(struct fs_context *fc, struct fs_parameter *param) if (ret != -ENOPARAM) return ret; - ret = security_fs_context_parse_param(fc, param); - if (ret != -ENOPARAM) - /* Param belongs to the LSM or is disallowed by the LSM; so - * don't pass to the FS. - */ - return ret; + /* + * In the legacy+binary mode, skip the security_fs_context_parse_param() + * call and let the legacy handler process also the security options. + * It will format them into the monolithic string, where the FS can + * process them (with FS_BINARY_MOUNTDATA it is expected to do it). + * + * Currently, this matches only btrfs and coda. Coda is broken with + * fsconfig(2) anyway, because it does actually take binary data. Btrfs + * only *pretends* to take binary data to work around the SELinux's + * no-remount-with-different-options check, so this allows it to work + * with fsconfig(2) properly. + * + * Once btrfs is ported to the new mount API, this hack can be reverted. + */ + if (fc->ops != &legacy_fs_context_ops || !(fc->fs_type->fs_flags & FS_BINARY_MOUNTDATA)) { + ret = security_fs_context_parse_param(fc, param); + if (ret != -ENOPARAM) + /* Param belongs to the LSM or is disallowed by the LSM; + * so don't pass to the FS. + */ + return ret; + } if (fc->ops->parse_param) { ret = fc->ops->parse_param(fc, param); -- 2.26.2
