Re: [PATCH v3] md: Fix the null-ptr-deref of 'mddev->private' while submitting IO

[yu kuai]

Hi,

在 2026/9/18 16:54, Abd-Alrhman Masalkhi 写道:
> Hi Zhihao,
>
> On Fri, Sep 18, 2026 at 16:34 +0800, Zhihao Cheng wrote:
>> 在 2026/9/18 16:20, Abd-Alrhman Masalkhi 写道:
>>> On Fri, Sep 18, 2026 at 15:38 +0800, Zhihao Cheng wrote:
>>>> 在 2026/9/18 15:29, Abd-Alrhman Masalkhi 写道:
>>>>> Hi Zhihao,
>>>>>
>>>>> On Fri, Sep 18, 2026 at 15:09 +0800, Zhihao Cheng wrote:
>>>>>> Concurrent processes md_stop and IO submitting could trigger a
>>>>>> null-ptr-deref of 'mddev->private':
>>>>>>
>>>>>>     BUG: kernel NULL pointer dereference, address: 0000000000000070
>>>>>>     RIP: 0010:_wait_barrier+0x2f/0x250
>>>>>>     Call Trace:
>>>>>>      raid1_make_request+0x150/0xf50
>>>>>>      md_handle_request+0x104/0x530
>>>>>>      md_submit_bio+0x76/0x130
>>>>>>      submit_bio+0xdd/0x250
>>>>>>      submit_bio_wait+0x1f/0x40
>>>>>>      __blkdev_direct_IO_simple+0x1f6/0x370
>>>>>>      blkdev_write_iter+0x3b2/0x520
>>>>>>      ksys_write+0x7d/0x190
>>>>>>
>>>>>>                  P1
>>>>>>      fd = open(/dev/md0, O_RDWR)
>>>>>> 	                      P2 (forked from P1, fd' <= fd)
>>>>>>      write(fd)
>>>>>>       submit_bio
>>>>>>        md_handle_request
>>>>>>         raid1_make_request
>>>>>>          raid1_write_request
>>>>>>                                      ioctl(fd, STOP_ARRAY)
>>>>>> 		                   mddev_set_closing_and_sync_blockdev
>>>>>> 				   // check passed, mddev->openers = 1,
>>>>>> 				   // because md_open() is only called
>>>>>> 				   // once in P1->open
>>>>>> 				   do_md_stop
>>>>>> 				    __md_stop
>>>>>> 				     mddev->private = NULL
>>>>>>
>>>>>>           conf = mddev->private // NULL
>>>>>>            wait_barrier(conf, sector) // null-ptr-deref !
>>>>>>
>>>>>> It is a common problem for raid0/1/10/5, and __md_stop could be triggered
>>>>>> by several paths(eg. ioctl, sysfs, ->dtr). Fix it by replacing
>>>>>> mddev_lock() with mddev_suspend_and_lock() for all __md_stop() callers.
>>>>>> The caller dm_table_destroy() is guaranteed being invoked with device
>>>>>> suspended, so raid_dtr() could keep using mddev_lock_nointr().
>>>>>> The caller array_state_store() is guaranteed by the check
>>>>>> mddev_set_closing_and_sync_blockdev(mddev, 0). For example, someone open
>>>>>> /dev/mdx, write something and close /dev/mdx, it won't trigger the
>>>>>> problem, all dirty pages can be flushed before mddev->openers decrement.
>>>>>> Besides, fail the submitting IO in md_handle_request() if the
>>>>>> 'mddev->pers' becomes NULL.
>>>>>>
>>>>>> Fetch a reproducer in https://bugzilla.kernel.org/show_bug.cgi?id=222020
>>>>>>
>>>>>> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
>>>>>> Reported-by: syzbot+3fe892ea5fc292e1353f@xxxxxxxxxxxxxxxxxxxxxxxxx
>>>>>> Closes: https://syzkaller.appspot.com/bug?extid=3fe892ea5fc292e1353f
>>>>>> Signed-off-by: Zhihao Cheng <chengzhihao1@xxxxxxxxxx>
>>>>>> ---
>>>>>>     v1->v2:
>>>>>>      1. Add 'mddev->pers != NULL' check before make_request
>>>>>>      2. Delete dm-raid caller(->dtr) modifications
>>>>>>      3. Move memalloc_noio_restore after mddev_unlock_and_resume
>>>>>>     v2->v3:
>>>>>>      1. Remove modifications in array_state_store()
>>>>>>      2. update commit msg
>>>>>>     drivers/md/md.c | 15 +++++++++++++++
>>>>>>     1 file changed, 15 insertions(+)
>>>>>>
>>>>>> diff --git a/drivers/md/md.c b/drivers/md/md.c
>>>>>> index 680b34a63cb3..e73338c26d53 100644
>>>>>> --- a/drivers/md/md.c
>>>>>> +++ b/drivers/md/md.c
>>>>>> @@ -414,6 +414,20 @@ bool md_handle_request(struct mddev *mddev, struct bio *bio)
>>>>>>     		if (!percpu_ref_tryget_live(&mddev->active_io))
>>>>>>     			goto check_suspended;
>>>>>>     	}
>>>>>> +	if (!mddev->pers) {
>>>>> Isn't this case already handled by md_submit_bio(). take a look in
>>>>> md_submit_bio()
>>>> Hi,Abd-Alrhman
>>>> I understand that 'mddev == NULL || mddev->pers == NULL' in
>>>> md_submit_bio() is a qiuck check, there still exists a small race window:
>>>>         P1                     P2
>>>> md_submit_bio
>>>>     if (mddev == NULL || mddev->pers == NULL)
>>>>                            md_ioctl
>>>>                             mddev_suspend_and_lock
>>>>                             do_md_stop->__md_stop
>>>>                             // set mddev->pers/private as NULL
>>>>                             mddev_unlock_and_resume
>>>>     md_handle_request
>>>>      if (is_suspended(mddev, bio))
>>>>      percpu_ref_tryget_live(&mddev->active_io)
>>>>      mddev->pers // null-ptr-deref
>>> It makes sense. It would solve only the null pointer reference, but I
>>> see another issue. Look at the comment before calling
>>> mddev_set_closing_and_sync_blockdev() in md_ioctl(). it says "Need to
>>> flush page cache, and ensure no-one else opens and writes". The
>>> suspending happens after flushing the page cache, in this case,
>>> there might be other writes in-flight.
>> At present, the current implementation does not align with the expected
>> annotations. Yu Kai privately gave me a feasible suggestion, which is to
>> have md_ioctl operate on a character device(like /dev/mapper/control)
>> instead of an md block device. This fix patch will serve as a temporary
>> solution before the official implementation is merged. By the way, I
>> don't have much time to implment it, do you have time to deal with the
>> official solution?
> Thanks for the summary. I can take on the official implementation.
> Before I get started, could we align on the interface details and any
> other relevant considerations?

I think the following procedures(details should be considered more):

1) add a new char device in mdraid to issue ioctl to array;
2) update mdadm to use the new ioctl;
3) a kernel message to warn the old block device ioctl is deprecated, and suggest
user to upgrade mdadm;
4) Since the new ioctl will introduce UAPI change, we should wait for a long time,
perhaps more than 1 year, before the deprecated old ioctl code in kernel can be removed.

>
>>>>>> +		/*
>>>>>> +		 * The __md_stop() sets 'mddev->private' to NULL during
>>>>>> +		 * the IO submitting, check 'mddev->pers' before the IO
>>>>>> +		 * being processed by specific driver to avoid the
>>>>>> +		 * null-ptr-deref of 'mddev-><member>'. The check is
>>>>>> +		 * safe because the IO has got the 'mddev->active_io'
>>>>>> +		 * reference, and all __md_stop() callers will wait for
>>>>>> +		 * the reference to be zero.
>>>>>> +		 */
>>>>>> +		bio_io_error(bio);
>>>>>> +		percpu_ref_put(&mddev->active_io);
>>>>>> +		return true;
>>>>>> +	}
>>>>>>     	if (!mddev->pers->make_request(mddev, bio)) {
>>>>>>     		percpu_ref_put(&mddev->active_io);
>>>>>>     		if (mddev_is_dm(mddev) && mddev->pers->prepare_suspend)
>>>>>> @@ -8299,6 +8313,7 @@ static bool md_ioctl_need_suspend(unsigned int cmd)
>>>>>>     	case HOT_REMOVE_DISK:
>>>>>>     	case SET_BITMAP_FILE:
>>>>>>     	case SET_ARRAY_INFO:
>>>>>> +	case STOP_ARRAY:
>>>>>>     		return true;
>>>>>>     	default:
>>>>>>     		return false;
>>>>>> -- 
>>>>>> 2.52.0
>>>>>>
-- 
Thanks,
Kuai