Re: [syzbot] possible deadlock in ovl_maybe_copy_up

[Amir Goldstein <amir73il@gmail.com>]

On Tue, Jun 22, 2021 at 9:53 AM Hillf Danton <hdanton@sina.com> wrote:
>
> On Mon, 21 Jun 2021 22:32:28 -0400 Mimi Zohar wrote:
> >On Fri, 2021-06-18 at 12:01 +0800, Hillf Danton wrote:
> >> On Sun, 4 Apr 2021 11:10:48 +0300 Amir Goldstein wrote:
> >> >On Sat, Apr 3, 2021 at 10:18 PM syzbot wrote:
> >> >>
> >> >> syzbot has found a reproducer for the following issue on:
> >> >>
> >> >> HEAD commit:    454c576c Add linux-next specific files for 20210401
> >> >> git tree:       linux-next
> >> >> console output: https://syzkaller.appspot.com/x/log.txt?x=1616e07ed00000
> >> >> kernel config:  https://syzkaller.appspot.com/x/.config?x=920cc274cae812a5
> >> >> dashboard link: https://syzkaller.appspot.com/bug?extid=c18f2f6a7b08c51e3025
> >> >> syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=13da365ed00000
> >> >> C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=13ca9d16d00000
> >> >>
> >> >> IMPORTANT: if you fix the issue, please add the following tag to the commit:
> >> >> Reported-by: syzbot+c18f2f6a7b08c51e3025@syzkaller.appspotmail.com
> >> >>
> >> >> ======================================================
> >> >> WARNING: possible circular locking dependency detected
> >> >> 5.12.0-rc5-next-20210401-syzkaller #0 Not tainted
> >> >> ------------------------------------------------------
> >> >> syz-executor144/9166 is trying to acquire lock:
> >> >> ffff888144cf0460 (sb_writers#5){.+.+}-{0:0}, at: ovl_maybe_copy_up+0x11f/0x190 fs/overlayfs/copy_up.c:995
> >> >>
> >> >> but task is already holding lock:
> >> >> ffff8880256d42c0 (&iint->mutex){+.+.}-{3:3}, at: process_measurement+0x3a8/0x17e0 security/integrity/ima/ima_main.c:253
> >> >>
> >> >> which lock already depends on the new lock.
> >> >>
> >> >>
> >> >> the existing dependency chain (in reverse order) is:
> >> >>
> >> >> -> #1 (&iint->mutex){+.+.}-{3:3}:
> >> >>        __mutex_lock_common kernel/locking/mutex.c:949 [inline]
> >> >>        __mutex_lock+0x139/0x1120 kernel/locking/mutex.c:1096
> >> >>        process_measurement+0x3a8/0x17e0 security/integrity/ima/ima_main.c:253
> >> >>        ima_file_check+0xb9/0x100 security/integrity/ima/ima_main.c:499
> >> >>        do_open fs/namei.c:3361 [inline]
> >> >>        path_openat+0x15b5/0x27e0 fs/namei.c:3492
> >> >>        do_filp_open+0x17e/0x3c0 fs/namei.c:3519
> >> >>        do_sys_openat2+0x16d/0x420 fs/open.c:1187
> >> >>        do_sys_open fs/open.c:1203 [inline]
> >> >>        __do_sys_open fs/open.c:1211 [inline]
> >> >>        __se_sys_open fs/open.c:1207 [inline]
> >> >>        __x64_sys_open+0x119/0x1c0 fs/open.c:1207
> >> >>        do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46
> >> >>        entry_SYSCALL_64_after_hwframe+0x44/0xae
> >> >>
> >> >> -> #0 (sb_writers#5){.+.+}-{0:0}:
> >> >>        check_prev_add kernel/locking/lockdep.c:2938 [inline]
> >> >>        check_prevs_add kernel/locking/lockdep.c:3061 [inline]
> >> >>        validate_chain kernel/locking/lockdep.c:3676 [inline]
> >> >>        __lock_acquire+0x2a17/0x5230 kernel/locking/lockdep.c:4902
> >> >>        lock_acquire kernel/locking/lockdep.c:5512 [inline]
> >> >>        lock_acquire+0x1ab/0x740 kernel/locking/lockdep.c:5477
> >> >>        percpu_down_read include/linux/percpu-rwsem.h:51 [inline]
> >> >>        __sb_start_write include/linux/fs.h:1758 [inline]
> >> >>        sb_start_write include/linux/fs.h:1828 [inline]
> >> >>        mnt_want_write+0x6e/0x3e0 fs/namespace.c:375
> >> >>        ovl_maybe_copy_up+0x11f/0x190 fs/overlayfs/copy_up.c:995
> >> >>        ovl_open+0xba/0x270 fs/overlayfs/file.c:149
> >> >>        do_dentry_open+0x4b9/0x11b0 fs/open.c:826
> >> >>        vfs_open fs/open.c:940 [inline]
> >> >>        dentry_open+0x132/0x1d0 fs/open.c:956
> >> >>        ima_calc_file_hash+0x2d2/0x4b0 security/integrity/ima/ima_crypto.c:557
> >> >>        ima_collect_measurement+0x4ca/0x570 security/integrity/ima/ima_api.c:252
> >> >>        process_measurement+0xd1c/0x17e0 security/integrity/ima/ima_main.c:330
> >> >>        ima_file_check+0xb9/0x100 security/integrity/ima/ima_main.c:499
> >> >>        do_open fs/namei.c:3361 [inline]
> >> >>        path_openat+0x15b5/0x27e0 fs/namei.c:3492
> >> >>        do_filp_open+0x17e/0x3c0 fs/namei.c:3519
> >> >>        do_sys_openat2+0x16d/0x420 fs/open.c:1187
> >> >>        do_sys_open fs/open.c:1203 [inline]
> >> >>        __do_sys_open fs/open.c:1211 [inline]
> >> >>        __se_sys_open fs/open.c:1207 [inline]
> >> >>        __x64_sys_open+0x119/0x1c0 fs/open.c:1207
> >> >>        do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46
> >> >>        entry_SYSCALL_64_after_hwframe+0x44/0xae
> >> >>
> >> >> other info that might help us debug this:
> >> >>
> >> >>  Possible unsafe locking scenario:
> >> >>
> >> >>        CPU0                    CPU1
> >> >>        ----                    ----
> >> >>   lock(&iint->mutex);
> >> >>                                lock(sb_writers#5);
> >> >>                                lock(&iint->mutex);
> >> >>   lock(sb_writers#5);
> >> >>
> >> >>  *** DEADLOCK ***
> >> >>
> >> >> 1 lock held by syz-executor144/9166:
> >> >>  #0: ffff8880256d42c0 (&iint->mutex){+.+.}-{3:3}, at: process_measurement+0x3a8/0x17e0 security/integrity/ima/ima_main.c:253
> >> >>
> >>
> >> It is reported again.
> >>   https://lore.kernel.org/lkml/00000000000067d24205c4d0e599@google.com/
> >> >
> >> >It's a false positive lockdep warning due to missing annotation of
> >> >stacking layer on iint->mutex in IMA code.
> >>
> >> Add it by copying what's created for ovl, see below.
> >> >
> >> >To fix it properly, iint->mutex, which can be taken in any of the
> >> >stacking fs layers, should be annotated with stacking depth like
> >> >ovl_lockdep_annotate_inode_mutex_key()
> >> >
> >> >I think it's the same root cause as:
> >> >https://syzkaller.appspot.com/bug?extid=18a1619cceea30ed45af
> >> >https://syzkaller.appspot.com/bug?extid=ae82084b07d0297e566b
> >> >
> >> >I think both of the above were marked "fixed" by a paper over.
> >> >The latter was marked "fixed" by "ovl: detect overlapping layers"
> >> >but if you look at the repro, the fact that 'workdir' overlaps with
> >> >'lowerdir' has nothing to do with the lockdep warning, so said
> >> >"fix" just papered over the IMA lockdep warning.
> >> >
> >> >Thanks,
> >> >Amir.
> >>
> >> +++ x/security/integrity/iint.c
> >> @@ -85,6 +85,45 @@ static void iint_free(struct integrity_i
> >>      kmem_cache_free(iint_cache, iint);
> >>  }
> >>
> >> +/*
> >> + * a copy from ovl_lockdep_annotate_inode_mutex_key() in a bit to fix
> >> +
> >> +   Possible unsafe locking scenario:
> >> +
> >> +    CPU0                    CPU1
> >> +       ----                    ----
> >> +     lock(&iint->mutex);
> >> +                               lock(sb_writers#5);
> >> +                               lock(&iint->mutex);
> >> +     lock(sb_writers#5);
> >> +
> >> +     *** DEADLOCK ***
> >> +
> >> +It's a false positive lockdep warning due to missing annotation of
> >> +stacking layer on iint->mutex in IMA code. [1]
> >> +
> >> +[1] https://lore.kernel.org/linux-unionfs/CAOQ4uxjk4XYuwz5HCmN-Ge=Ld=tM1f7ZxVrd5U1AC2Wisc9MTA@mail.gmail.com/
> >> +*/
> >> +static void iint_annotate_mutex_key(struct integrity_iint_cache *iint,
> >> +                                    struct inode *inode)
> >> +{
> >> +#ifdef CONFIG_LOCKDEP
> >> +    static struct lock_class_key
> >> +            iint_mutex_key[FILESYSTEM_MAX_STACK_DEPTH],
> >> +            iint_mutex_dir_key[FILESYSTEM_MAX_STACK_DEPTH];
> >> +
> >> +    int depth = inode->i_sb->s_stack_depth - 1;
> >> +
> >> +    if (WARN_ON_ONCE(depth < 0 || depth >= FILESYSTEM_MAX_STACK_DEPTH))
> >> +            depth = 0;
> >> +
> >> +    if (S_ISDIR(inode->i_mode))
> >> +            lockdep_set_class(&iint->mutex, &iint_mutex_dir_key[depth]);
> >> +    else
> >> +            lockdep_set_class(&iint->mutex, &iint_mutex_key[depth]);
> >> +#endif
> >> +}
> >
> >The iint cache is only for regular files.
>
> Yes you are right.
> >
> >> +
> >>  /**
> >>   * integrity_inode_get - find or allocate an iint associated with an inode
> >>   * @inode: pointer to the inode
> >> @@ -113,6 +152,7 @@ struct integrity_iint_cache *integrity_i
> >>      iint = kmem_cache_alloc(iint_cache, GFP_NOFS);
> >>      if (!iint)
> >>              return NULL;
> >> +    iint_annotate_mutex_key(iint, inode);
> >>
> >>      write_lock(&integrity_iint_lock);
> >
> >Should annotating the iint be limited to files on overlay filesystems?
>
> Yes but it is more difficult to address than thought without adding to
> the vfs is_ovl_inode(inode).
>
> Aside from adding is_xfs_inode(inode), another option is move the
> dentry_open() in ima_calc_file_hash() out of and before the iint->mutex in
> process_measurement(), and that will keep the AB locking order intact.
>
> Thoughts are welcome.

There is no need to detect overlayfs check for (inode->i_sb->s_stack_depth > 0)

Thanks,
Amir.