On Sat, 23 Oct 2010, Al Viro wrote: > On Fri, Oct 22, 2010 at 06:14:01PM +0200, Miklos Szeredi wrote: > > > @@ -1782,6 +1844,14 @@ int do_add_mount(struct vfsmount *newmnt > > > > mnt_flags &= ~(MNT_SHARED | MNT_WRITE_HOLD | MNT_INTERNAL); > > Obviously not enough - you've just added a new flag that needs to be > trimmed from mnt_flags. > > > + /* Locking is necessary to prevent racing with remount r/o */ > > + down_read(&newmnt->mnt_sb->s_umount); > > + if (newmnt->mnt_sb->s_flags & MS_RDONLY) > > + mnt_flags |= MNT_READONLY; > > + > > + newmnt->mnt_flags = mnt_flags; > > + up_read(&newmnt->mnt_sb->s_umount); > > FWIW, I really don't like the way you are doing that; what we really need > there is a per-sb analog of mnt_want_write()/mnt_drop_write(). With > mnt_want_write() bumping per-sb write count, which would solve all these > problems quite nicely. > > NOTE: vfsmount being ro and sb being ro are *independent* things; Yes, except the mount(2) API which doens't quite let them be changed independently. > either > is enough to deny writes. Having remount ro + remount rw lose the state > of other vfsmounts is a Bad Thing(tm). Hmm. > > Another thing: > "If clone_mnt() happens while mnt_make_readonly() is running, the > cloned mount might have MNT_WRITE_HOLD flag set, which results in > mnt_want_write() spinning forever on this mount." > actually means > "neither clone_mnt() nor remounts should ever be done without > namespace_sem held exclusive; if that ever happens, we have a serious > bug that can lead to any number of bad things happening". > > Do you actually see such places? If so, that's what needs fixing. do_remount() takes s_umount, but not namespace_sem. Thanks, Miklos -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html