On Fri, 12 Apr 2024 23:42:19 -0700 Amir Goldstein > On Sat, Apr 13, 2024 at 4:41=E2=80=AFAM Hillf Danton <hdanton@xxxxxxxx> wrote: > > > > On Thu, 11 Apr 2024 01:11:20 -0700 > > > syzbot found the following issue on: > > > > > > HEAD commit: 6ebf211bb11d Add linux-next specific files for 20240410 > > > git tree: linux-next > > > C reproducer: https://syzkaller.appspot.com/x/repro.c?x=3D1621af9d180000 > > > > #syz test https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git 6ebf211bb11d > > > > --- x/fs/notify/fsnotify.c > > +++ y/fs/notify/fsnotify.c > > @@ -101,8 +101,8 @@ void fsnotify_sb_delete(struct super_blo > > wait_var_event(fsnotify_sb_watched_objects(sb), > > !atomic_long_read(fsnotify_sb_watched_objects(sb))); > > WARN_ON(fsnotify_sb_has_priority_watchers(sb, FSNOTIFY_PRIO_CONTENT)); > > - WARN_ON(fsnotify_sb_has_priority_watchers(sb, > > - FSNOTIFY_PRIO_PRE_CONTENT)); > > + WARN_ON(fsnotify_sb_has_priority_watchers(sb, FSNOTIFY_PRIO_PRE_CONTENT)); > > + synchronize_srcu(&fsnotify_mark_srcu); > > kfree(sbinfo); > > } > > > > @@ -499,7 +499,7 @@ int fsnotify(__u32 mask, const void *dat > > { > > const struct path *path =3D fsnotify_data_path(data, data_type); > > struct super_block *sb =3D fsnotify_data_sb(data, data_type); > > - struct fsnotify_sb_info *sbinfo =3D fsnotify_sb_info(sb); > > + struct fsnotify_sb_info *sbinfo; > > struct fsnotify_iter_info iter_info = {}; > > struct mount *mnt =3D NULL; > > struct inode *inode2 =3D NULL; > > @@ -529,6 +529,8 @@ int fsnotify(__u32 mask, const void *dat > > inode2_type =3D FSNOTIFY_ITER_TYPE_PARENT; > > } > > > > + iter_info.srcu_idx =3D srcu_read_lock(&fsnotify_mark_srcu); > > + sbinfo =3D fsnotify_sb_info(sb); > > /* > > * Optimization: srcu_read_lock() has a memory barrier which can > > * be expensive. It protects walking the *_fsnotify_marks lists. > > > See comment above. This kills the optimization. > It is not worth letting all the fsnotify hooks suffer the consequence > for the edge case of calling fsnotify hook during fs shutdown. Say nothing before reading your fix. > > Also, fsnotify_sb_info(sb) in fsnotify_sb_has_priority_watchers() > is also not protected and using srcu_read_lock() there completely > nullifies the purpose of fsnotify_sb_info. > > Here is a simplified fix for fsnotify_sb_error() rebased on the > pending mm fixes for this syzbot boot failure: > > #syz test: https://github.com/amir73il/linux fsnotify-fixes Feel free to post your patch at lore because not everyone has access to sites like github. > > Jan, > > I think that all the functions called from fs shutdown context > should observe that SB_ACTIVE is cleared but wasn't sure? If you composed fix based on SB_ACTIVE that is cleared in generic_shutdown_super() with &sb->s_umount held for write, I wonder what simpler serialization than srcu you could find/create in fsnotify.
