Re: [syzbot] possible deadlock in hugetlb_fault

[Dmitry Vyukov]

On Sat, 12 Nov 2022 at 05:02, 'Miaohe Lin' via syzkaller-bugs
<syzkaller-bugs@xxxxxxxxxxxxxxxx> wrote:
>
> On 2022/11/12 8:07, Mike Kravetz wrote:
> > On 11/04/22 09:00, syzbot wrote:
> >> Hello,
> >>
> >> syzbot found the following issue on:
> >>
> >> HEAD commit:    f2f32f8af2b0 Merge tag 'for-6.1-rc3-tag' of git://git.kern..
> >> git tree:       upstream
> >> console output: https://syzkaller.appspot.com/x/log.txt?x=137d52ca880000
> >> kernel config:  https://syzkaller.appspot.com/x/.config?x=d080a4bd239918dd
> >> dashboard link: https://syzkaller.appspot.com/bug?extid=ca56f14c500045350f93
> >> compiler:       gcc (Debian 10.2.1-6) 10.2.1 20210110, GNU ld (GNU Binutils for Debian) 2.35.2
> >> userspace arch: i386
> >>
> >> Unfortunately, I don't have any reproducer for this issue yet.
> >>
> >> Downloadable assets:
> >> disk image: https://storage.googleapis.com/syzbot-assets/b4f72e7a4c11/disk-f2f32f8a.raw.xz
> >> vmlinux: https://storage.googleapis.com/syzbot-assets/3f88997ad7c9/vmlinux-f2f32f8a.xz
> >> kernel image: https://storage.googleapis.com/syzbot-assets/b4b5b3963e2d/bzImage-f2f32f8a.xz
> >>
> >> IMPORTANT: if you fix the issue, please add the following tag to the commit:
> >> Reported-by: syzbot+ca56f14c500045350f93@xxxxxxxxxxxxxxxxxxxxxxxxx
> >>
> >> ======================================================
> >> WARNING: possible circular locking dependency detected
> >> 6.1.0-rc3-syzkaller-00152-gf2f32f8af2b0 #0 Not tainted
> >> ------------------------------------------------------
> >> syz-executor.2/5665 is trying to acquire lock:
> >> ffff88801c74c298 (&mm->mmap_lock#2){++++}-{3:3}, at: __might_fault+0xa1/0x170 mm/memory.c:5645
> >>
> >> but task is already holding lock:
> >> ffff88801c4f3078 (&vma_lock->rw_sema){++++}-{3:3}, at: hugetlb_vma_lock_read mm/hugetlb.c:6816 [inline]
> >> ffff88801c4f3078 (&vma_lock->rw_sema){++++}-{3:3}, at: hugetlb_fault+0x40a/0x2060 mm/hugetlb.c:5859
> >>
> >> which lock already depends on the new lock.
> >>
> >>
> >> the existing dependency chain (in reverse order) is:
> >>
> >> -> #1 (&vma_lock->rw_sema){++++}-{3:3}:
> >>        down_write+0x90/0x220 kernel/locking/rwsem.c:1562
> >>        hugetlb_vma_lock_write mm/hugetlb.c:6834 [inline]
> >>        __unmap_hugepage_range_final+0x97/0x340 mm/hugetlb.c:5202
> >>        unmap_single_vma+0x23d/0x2a0 mm/memory.c:1690
> >>        unmap_vmas+0x21e/0x370 mm/memory.c:1733
> >>        exit_mmap+0x189/0x7a0 mm/mmap.c:3090
> >>        __mmput+0x128/0x4c0 kernel/fork.c:1185
> >>        mmput+0x5c/0x70 kernel/fork.c:1207
> >>        exit_mm kernel/exit.c:516 [inline]
> >>        do_exit+0xa39/0x2a20 kernel/exit.c:807
> >>        do_group_exit+0xd0/0x2a0 kernel/exit.c:950
> >>        get_signal+0x21a1/0x2430 kernel/signal.c:2858
> >>        arch_do_signal_or_restart+0x82/0x2300 arch/x86/kernel/signal.c:869
> >>        exit_to_user_mode_loop kernel/entry/common.c:168 [inline]
> >>        exit_to_user_mode_prepare+0x15f/0x250 kernel/entry/common.c:203
> >>        __syscall_exit_to_user_mode_work kernel/entry/common.c:285 [inline]
> >>        syscall_exit_to_user_mode+0x19/0x50 kernel/entry/common.c:296
> >>        __do_fast_syscall_32+0x72/0xf0 arch/x86/entry/common.c:181
> >>        do_fast_syscall_32+0x2f/0x70 arch/x86/entry/common.c:203
> >>        entry_SYSENTER_compat_after_hwframe+0x70/0x82
> >>
> >> -> #0 (&mm->mmap_lock#2){++++}-{3:3}:
> >>        check_prev_add kernel/locking/lockdep.c:3097 [inline]
> >>        check_prevs_add kernel/locking/lockdep.c:3216 [inline]
> >>        validate_chain kernel/locking/lockdep.c:3831 [inline]
> >>        __lock_acquire+0x2a43/0x56d0 kernel/locking/lockdep.c:5055
> >>        lock_acquire kernel/locking/lockdep.c:5668 [inline]
> >>        lock_acquire+0x1df/0x630 kernel/locking/lockdep.c:5633
> >>        __might_fault mm/memory.c:5646 [inline]
> >>        __might_fault+0x104/0x170 mm/memory.c:5639
> >>        _copy_from_user+0x25/0x170 lib/usercopy.c:13
> >>        copy_from_user include/linux/uaccess.h:161 [inline]
> >>        snd_rawmidi_kernel_write1+0x366/0x880 sound/core/rawmidi.c:1549
> >>        snd_rawmidi_write+0x273/0xbb0 sound/core/rawmidi.c:1618
> >>        vfs_write+0x2d7/0xdd0 fs/read_write.c:582
> >>        ksys_write+0x1e8/0x250 fs/read_write.c:637
> >>        do_syscall_32_irqs_on arch/x86/entry/common.c:112 [inline]
> >>        __do_fast_syscall_32+0x65/0xf0 arch/x86/entry/common.c:178
> >>        do_fast_syscall_32+0x2f/0x70 arch/x86/entry/common.c:203
> >>        entry_SYSENTER_compat_after_hwframe+0x70/0x82
> >>
> >> other info that might help us debug this:
> >>
> >>  Possible unsafe locking scenario:
> >>
> >>        CPU0                    CPU1
> >>        ----                    ----
> >>   lock(&vma_lock->rw_sema);
> >>                                lock(&mm->mmap_lock#2);
> >>                                lock(&vma_lock->rw_sema);
> >>   lock(&mm->mmap_lock#2);
> >
> > I may not be reading the report correctly, but I can not see how we acquire the
> > hugetlb vma_lock before trying to acquire mmap_lock in stack 0.  We would not
> > acquire the vma_lock until we enter hugetlb fault processing (not in the stack).

The unlock of vma_lock is conditional:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/mm/hugetlb.c?id=f2f32f8af2b0ca9d619e5183eae3eed431793baf#n6840

and the condition is:

static bool __vma_shareable_flags_pmd(struct vm_area_struct *vma)
{
    return vma->vm_flags & (VM_MAYSHARE | VM_SHARED) &&
         vma->vm_private_data;
}

Is it possible that the condition has changed between vma lock and unlock?
What mutexes protect vma->vm_flags/vm_private_data?

That would make the report perfectly sensible.

FWIW the test case that was running is this, that's the syscalls that
were running concurrently:

07:56:56 executing program 2:
r0 = syz_open_dev$sndmidi(&(0x7f0000000040), 0x2, 0x141101)
r1 = dup(r0)
setsockopt$inet_sctp_SCTP_I_WANT_MAPPED_V4_ADDR(r1, 0x84, 0xc,
&(0x7f0000000080), 0x4) (async)
write$6lowpan_enable(r1, &(0x7f0000000000)='0', 0xc86ade39) (async)
mmap(&(0x7f0000000000/0xb36000)=nil, 0xb36000, 0x3, 0x68831,
0xffffffffffffffff, 0x0) (async)
madvise(&(0x7f0000000000/0x600000)=nil, 0x600003, 0x4) (async, rerun: 32)
mremap(&(0x7f00007a0000/0x3000)=nil, 0x3000, 0x2000, 0x7,
&(0x7f0000835000/0x2000)=nil) (rerun: 32)





> > Adding Miaohe Lin on Cc due to previous help with vma_lock potential deadlock
> > situations.  Miaohe, does this make sense to you?
> >
>
> Hi Mike,
>   This doesn't make sense for me too. Stack #1 shows that syz-executor is releasing
> its address space while stack #0 shows another thread is serving the write syscall.
> In this case, mm->mm_users is 0 and all threads in this process should be serving
> do_exit()? But I could be easily wrong. Also I can't see how vma_lock is locked before
> trying to acquire mmap_lock in above stacks. Might this be a false positive?