mtd: blkdevs: protect tr->devs list by mtd_table_mutex
diff mbox series

Message ID 20190929141157.115845-1-houtao1@huawei.com
State New
Headers show
Series
  • mtd: blkdevs: protect tr->devs list by mtd_table_mutex
Related show

Commit Message

Hou Tao Sept. 29, 2019, 2:11 p.m. UTC
There may be list corruption if there are concurrent list traversal
and list deletion on tr->devs as showed in the following case:

CPU 0                               CPU 1

open /dev/mtdblock1

// remove mtd1
blktrans_notify_remove()
    del_mtd_blktrans_dev()

close /dev/mtdblock1
  blktrans_release
    blktrans_dev_put
      acquire blktrans_ref_mutex     // remove mtd0
      // the final release           acquire mtd_table_mutex
      blktrans_dev_release()         blktrans_notify_remove()
        // remove mtdblock1            // next is mtdblock1
        list_del(&dev->list)           list_for_each_entry_safe()

We could fix the problem by acquiring blktrans_ref_mutex during
the traversal of tr->devs, but blktrans_ref_mutex needs to be released
before invoking tr->remote_dev(), so we also need to increase the kref
of current device else the device may be freed and decrease the kref
after the removal.

Or we could move the list deletion to del_mtd_blktrans_dev(), and protect
the operations on tr->devs by mtd_table_mutex which has already be taken.

The latter fix is simpler. We also can remove the unnecessary acquisitions
of blktrans_ref_mutex in add_mtd_blktrans_dev() because operations on
tr->devs have already been protected by mtd_table_mutex.

Fixes: 048d87199566 ("mtd: blktrans: Hotplug fixes")
Signed-off-by: Hou Tao <houtao1@huawei.com>
---
I found the problem by code review, and could not find a way to
ensure the problem, because the removal of mtd devices almost
comes from the removal of modules, and the open of /dev/mtdblockX
will prevent the module from removing.
---
 drivers/mtd/mtd_blkdevs.c | 9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)

Comments

Hou Tao Oct. 8, 2019, 1:12 p.m. UTC | #1
ping ?

On 2019/9/29 22:11, Hou Tao wrote:
> There may be list corruption if there are concurrent list traversal
> and list deletion on tr->devs as showed in the following case:
> 
> CPU 0                               CPU 1
> 
> open /dev/mtdblock1
> 
> // remove mtd1
> blktrans_notify_remove()
>     del_mtd_blktrans_dev()
> 
> close /dev/mtdblock1
>   blktrans_release
>     blktrans_dev_put
>       acquire blktrans_ref_mutex     // remove mtd0
>       // the final release           acquire mtd_table_mutex
>       blktrans_dev_release()         blktrans_notify_remove()
>         // remove mtdblock1            // next is mtdblock1
>         list_del(&dev->list)           list_for_each_entry_safe()
> 
> We could fix the problem by acquiring blktrans_ref_mutex during
> the traversal of tr->devs, but blktrans_ref_mutex needs to be released
> before invoking tr->remote_dev(), so we also need to increase the kref
> of current device else the device may be freed and decrease the kref
> after the removal.
> 
> Or we could move the list deletion to del_mtd_blktrans_dev(), and protect
> the operations on tr->devs by mtd_table_mutex which has already be taken.
> 
> The latter fix is simpler. We also can remove the unnecessary acquisitions
> of blktrans_ref_mutex in add_mtd_blktrans_dev() because operations on
> tr->devs have already been protected by mtd_table_mutex.
> 
> Fixes: 048d87199566 ("mtd: blktrans: Hotplug fixes")
> Signed-off-by: Hou Tao <houtao1@huawei.com>
> ---
> I found the problem by code review, and could not find a way to
> ensure the problem, because the removal of mtd devices almost
> comes from the removal of modules, and the open of /dev/mtdblockX
> will prevent the module from removing.
> ---
>  drivers/mtd/mtd_blkdevs.c | 9 +++------
>  1 file changed, 3 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/mtd/mtd_blkdevs.c b/drivers/mtd/mtd_blkdevs.c
> index 0c05f77f9b21..e04260237a25 100644
> --- a/drivers/mtd/mtd_blkdevs.c
> +++ b/drivers/mtd/mtd_blkdevs.c
> @@ -35,7 +35,6 @@ static void blktrans_dev_release(struct kref *kref)
>  	blk_mq_free_tag_set(dev->tag_set);
>  	kfree(dev->tag_set);
>  	put_disk(dev->disk);
> -	list_del(&dev->list);
>  	kfree(dev);
>  }
>  
> @@ -350,7 +349,6 @@ int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new)
>  		BUG();
>  	}
>  
> -	mutex_lock(&blktrans_ref_mutex);
>  	list_for_each_entry(d, &tr->devs, list) {
>  		if (new->devnum == -1) {
>  			/* Use first free number */
> @@ -362,7 +360,6 @@ int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new)
>  			}
>  		} else if (d->devnum == new->devnum) {
>  			/* Required number taken */
> -			mutex_unlock(&blktrans_ref_mutex);
>  			return -EBUSY;
>  		} else if (d->devnum > new->devnum) {
>  			/* Required number was free */
> @@ -381,14 +378,12 @@ int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new)
>  	 * with this number. */
>  	if (new->devnum > (MINORMASK >> tr->part_bits) ||
>  	    (tr->part_bits && new->devnum >= 27 * 26)) {
> -		mutex_unlock(&blktrans_ref_mutex);
>  		goto error1;
>  	}
>  
>  	list_add_tail(&new->list, &tr->devs);
> - added:
> -	mutex_unlock(&blktrans_ref_mutex);
>  
> + added:
>  	mutex_init(&new->lock);
>  	kref_init(&new->ref);
>  	if (!tr->writesect)
> @@ -484,6 +479,8 @@ int del_mtd_blktrans_dev(struct mtd_blktrans_dev *old)
>  		BUG();
>  	}
>  
> +	list_del(&old->list);
> +
>  	if (old->disk_attributes)
>  		sysfs_remove_group(&disk_to_dev(old->disk)->kobj,
>  						old->disk_attributes);
>
Richard Weinberger Oct. 15, 2019, 8:10 p.m. UTC | #2
On Sun, Sep 29, 2019 at 4:05 PM Hou Tao <houtao1@huawei.com> wrote:
>
> There may be list corruption if there are concurrent list traversal
> and list deletion on tr->devs as showed in the following case:
>
> CPU 0                               CPU 1
>
> open /dev/mtdblock1
>
> // remove mtd1
> blktrans_notify_remove()
>     del_mtd_blktrans_dev()
>
> close /dev/mtdblock1
>   blktrans_release
>     blktrans_dev_put
>       acquire blktrans_ref_mutex     // remove mtd0
>       // the final release           acquire mtd_table_mutex
>       blktrans_dev_release()         blktrans_notify_remove()
>         // remove mtdblock1            // next is mtdblock1
>         list_del(&dev->list)           list_for_each_entry_safe()
>
> We could fix the problem by acquiring blktrans_ref_mutex during
> the traversal of tr->devs, but blktrans_ref_mutex needs to be released
> before invoking tr->remote_dev(), so we also need to increase the kref
> of current device else the device may be freed and decrease the kref
> after the removal.
>
> Or we could move the list deletion to del_mtd_blktrans_dev(), and protect
> the operations on tr->devs by mtd_table_mutex which has already be taken.
>
> The latter fix is simpler. We also can remove the unnecessary acquisitions
> of blktrans_ref_mutex in add_mtd_blktrans_dev() because operations on
> tr->devs have already been protected by mtd_table_mutex.
>
> Fixes: 048d87199566 ("mtd: blktrans: Hotplug fixes")
> Signed-off-by: Hou Tao <houtao1@huawei.com>
> ---
> I found the problem by code review, and could not find a way to
> ensure the problem, because the removal of mtd devices almost
> comes from the removal of modules, and the open of /dev/mtdblockX
> will prevent the module from removing.

I'm confused. Can the problem only happen if you remove a mtd while
it is open?
Hou Tao Oct. 16, 2019, 10:55 a.m. UTC | #3
Hi,

On 2019/10/16 4:10, Richard Weinberger wrote:
> On Sun, Sep 29, 2019 at 4:05 PM Hou Tao <houtao1@huawei.com> wrote:
>>
>> There may be list corruption if there are concurrent list traversal
>> and list deletion on tr->devs as showed in the following case:
>>
>> CPU 0                               CPU 1
>>
>> open /dev/mtdblock1
>>
>> // remove mtd1
>> blktrans_notify_remove()
>>     del_mtd_blktrans_dev()
>>
>> close /dev/mtdblock1
>>   blktrans_release
>>     blktrans_dev_put
>>       acquire blktrans_ref_mutex     // remove mtd0
>>       // the final release           acquire mtd_table_mutex
>>       blktrans_dev_release()         blktrans_notify_remove()
>>         // remove mtdblock1            // next is mtdblock1
>>         list_del(&dev->list)           list_for_each_entry_safe()
>>
>> We could fix the problem by acquiring blktrans_ref_mutex during
>> the traversal of tr->devs, but blktrans_ref_mutex needs to be released
>> before invoking tr->remote_dev(), so we also need to increase the kref
>> of current device else the device may be freed and decrease the kref
>> after the removal.
>>
>> Or we could move the list deletion to del_mtd_blktrans_dev(), and protect
>> the operations on tr->devs by mtd_table_mutex which has already be taken.
>>
>> The latter fix is simpler. We also can remove the unnecessary acquisitions
>> of blktrans_ref_mutex in add_mtd_blktrans_dev() because operations on
>> tr->devs have already been protected by mtd_table_mutex.
>>
>> Fixes: 048d87199566 ("mtd: blktrans: Hotplug fixes")
>> Signed-off-by: Hou Tao <houtao1@huawei.com>
>> ---
>> I found the problem by code review, and could not find a way to
>> ensure the problem, because the removal of mtd devices almost
>> comes from the removal of modules, and the open of /dev/mtdblockX
>> will prevent the module from removing.
> 
> I'm confused. Can the problem only happen if you remove a mtd while
> it is open?
> 
No. The problem may happen when closing a mtd block device (instead of
the mtd char device) for which its mtd device had already been removed.

The reason why I can not confirm the problem is that I am trying to
confirm the problem by the following steps:
(1) insmod block2mtd.ko to create a mtd device (e.g., /dev/mtd0)
(2) open /dev/mtdblock0
(3) remove /dev/mtd0 by removing block2mtd.ko

step (3) always fails because the opening of /dev/mtdblock0 has already
increased the reference count of block2mtd.ko, so /dev/mtd0 can not be removed.

If the removal of mtd device doesn't need the removal of the module (for example,
the removal comes from unbinding from its driver), the problem may occur.

And according the implementation of del_mtd_blktrans_dev(), the scenario is possible
as show in the following lines:

    /* If the device is currently open, tell trans driver to close it,
        then put mtd device, and don't touch it again */
    mutex_lock(&old->lock);
    if (old->open) {
        if (old->tr->release)
            old->tr->release(old);
        __put_mtd_device(old->mtd);
    }

Regards,
Tao










> .
>
Richard Weinberger Oct. 16, 2019, 11:06 a.m. UTC | #4
----- Urspr√ľngliche Mail -----
> Von: "Hou Tao" <houtao1@huawei.com>
> An: "Richard Weinberger" <richard.weinberger@gmail.com>
> CC: "linux-mtd" <linux-mtd@lists.infradead.org>, "Miquel Raynal" <miquel.raynal@bootlin.com>, "Vignesh Raghavendra"
> <vigneshr@ti.com>, "richard" <richard@nod.at>, "Maxim Levitsky" <maximlevitsky@gmail.com>, "Marek Vasut"
> <marek.vasut@gmail.com>, "Brian Norris" <computersforpeace@gmail.com>, "David Woodhouse" <dwmw2@infradead.org>
> Gesendet: Mittwoch, 16. Oktober 2019 12:55:12
> Betreff: Re: [PATCH] mtd: blkdevs: protect tr->devs list by mtd_table_mutex

> Hi,
> 
> On 2019/10/16 4:10, Richard Weinberger wrote:
>> On Sun, Sep 29, 2019 at 4:05 PM Hou Tao <houtao1@huawei.com> wrote:
>>>
>>> There may be list corruption if there are concurrent list traversal
>>> and list deletion on tr->devs as showed in the following case:
>>>
>>> CPU 0                               CPU 1
>>>
>>> open /dev/mtdblock1
>>>
>>> // remove mtd1
>>> blktrans_notify_remove()
>>>     del_mtd_blktrans_dev()
>>>
>>> close /dev/mtdblock1
>>>   blktrans_release
>>>     blktrans_dev_put
>>>       acquire blktrans_ref_mutex     // remove mtd0
>>>       // the final release           acquire mtd_table_mutex
>>>       blktrans_dev_release()         blktrans_notify_remove()
>>>         // remove mtdblock1            // next is mtdblock1
>>>         list_del(&dev->list)           list_for_each_entry_safe()
>>>
>>> We could fix the problem by acquiring blktrans_ref_mutex during
>>> the traversal of tr->devs, but blktrans_ref_mutex needs to be released
>>> before invoking tr->remote_dev(), so we also need to increase the kref
>>> of current device else the device may be freed and decrease the kref
>>> after the removal.
>>>
>>> Or we could move the list deletion to del_mtd_blktrans_dev(), and protect
>>> the operations on tr->devs by mtd_table_mutex which has already be taken.
>>>
>>> The latter fix is simpler. We also can remove the unnecessary acquisitions
>>> of blktrans_ref_mutex in add_mtd_blktrans_dev() because operations on
>>> tr->devs have already been protected by mtd_table_mutex.
>>>
>>> Fixes: 048d87199566 ("mtd: blktrans: Hotplug fixes")
>>> Signed-off-by: Hou Tao <houtao1@huawei.com>
>>> ---
>>> I found the problem by code review, and could not find a way to
>>> ensure the problem, because the removal of mtd devices almost
>>> comes from the removal of modules, and the open of /dev/mtdblockX
>>> will prevent the module from removing.
>> 
>> I'm confused. Can the problem only happen if you remove a mtd while
>> it is open?
>> 
> No. The problem may happen when closing a mtd block device (instead of
> the mtd char device) for which its mtd device had already been removed.
> 
> The reason why I can not confirm the problem is that I am trying to
> confirm the problem by the following steps:
> (1) insmod block2mtd.ko to create a mtd device (e.g., /dev/mtd0)
> (2) open /dev/mtdblock0
> (3) remove /dev/mtd0 by removing block2mtd.ko
> 
> step (3) always fails because the opening of /dev/mtdblock0 has already
> increased the reference count of block2mtd.ko, so /dev/mtd0 can not be removed.

Ok. But yeah, the problem is real and I'm sure with ubi+gluebi it can be triggered.

Thanks,
//richard
Hou Tao Oct. 16, 2019, 1:06 p.m. UTC | #5
Hi,

On 2019/10/16 19:06, Richard Weinberger wrote:
> ----- Urspr√ľngliche Mail -----
>> Von: "Hou Tao" <houtao1@huawei.com>
>> An: "Richard Weinberger" <richard.weinberger@gmail.com>
>> CC: "linux-mtd" <linux-mtd@lists.infradead.org>, "Miquel Raynal" <miquel.raynal@bootlin.com>, "Vignesh Raghavendra"
>> <vigneshr@ti.com>, "richard" <richard@nod.at>, "Maxim Levitsky" <maximlevitsky@gmail.com>, "Marek Vasut"
>> <marek.vasut@gmail.com>, "Brian Norris" <computersforpeace@gmail.com>, "David Woodhouse" <dwmw2@infradead.org>
>> Gesendet: Mittwoch, 16. Oktober 2019 12:55:12
>> Betreff: Re: [PATCH] mtd: blkdevs: protect tr->devs list by mtd_table_mutex
> 
>> Hi,
>>
>> On 2019/10/16 4:10, Richard Weinberger wrote:
>>> On Sun, Sep 29, 2019 at 4:05 PM Hou Tao <houtao1@huawei.com> wrote:
>>>>
>>>> There may be list corruption if there are concurrent list traversal
>>>> and list deletion on tr->devs as showed in the following case:
>>>>
>>>> CPU 0                               CPU 1
>>>>
>>>> open /dev/mtdblock1
>>>>
>>>> // remove mtd1
>>>> blktrans_notify_remove()
>>>>     del_mtd_blktrans_dev()
>>>>
>>>> close /dev/mtdblock1
>>>>   blktrans_release
>>>>     blktrans_dev_put
>>>>       acquire blktrans_ref_mutex     // remove mtd0
>>>>       // the final release           acquire mtd_table_mutex
>>>>       blktrans_dev_release()         blktrans_notify_remove()
>>>>         // remove mtdblock1            // next is mtdblock1
>>>>         list_del(&dev->list)           list_for_each_entry_safe()
>>>>
>>>> We could fix the problem by acquiring blktrans_ref_mutex during
>>>> the traversal of tr->devs, but blktrans_ref_mutex needs to be released
>>>> before invoking tr->remote_dev(), so we also need to increase the kref
>>>> of current device else the device may be freed and decrease the kref
>>>> after the removal.
>>>>
>>>> Or we could move the list deletion to del_mtd_blktrans_dev(), and protect
>>>> the operations on tr->devs by mtd_table_mutex which has already be taken.
>>>>
>>>> The latter fix is simpler. We also can remove the unnecessary acquisitions
>>>> of blktrans_ref_mutex in add_mtd_blktrans_dev() because operations on
>>>> tr->devs have already been protected by mtd_table_mutex.
>>>>
>>>> Fixes: 048d87199566 ("mtd: blktrans: Hotplug fixes")
>>>> Signed-off-by: Hou Tao <houtao1@huawei.com>
>>>> ---
>>>> I found the problem by code review, and could not find a way to
>>>> ensure the problem, because the removal of mtd devices almost
>>>> comes from the removal of modules, and the open of /dev/mtdblockX
>>>> will prevent the module from removing.
>>>
>>> I'm confused. Can the problem only happen if you remove a mtd while
>>> it is open?
>>>
>> No. The problem may happen when closing a mtd block device (instead of
>> the mtd char device) for which its mtd device had already been removed.
>>
>> The reason why I can not confirm the problem is that I am trying to
>> confirm the problem by the following steps:
>> (1) insmod block2mtd.ko to create a mtd device (e.g., /dev/mtd0)
>> (2) open /dev/mtdblock0
>> (3) remove /dev/mtd0 by removing block2mtd.ko
>>
>> step (3) always fails because the opening of /dev/mtdblock0 has already
>> increased the reference count of block2mtd.ko, so /dev/mtd0 can not be removed.
> 
> Ok. But yeah, the problem is real and I'm sure with ubi+gluebi it can be triggered.
> 
Thanks for your suggestions. I will try to use ubi+gluebi to confirm the problem.

Regards,
Tao
> Thanks,
> //richard
> 
> ______________________________________________________
> Linux MTD discussion mailing list
> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>

Patch
diff mbox series

diff --git a/drivers/mtd/mtd_blkdevs.c b/drivers/mtd/mtd_blkdevs.c
index 0c05f77f9b21..e04260237a25 100644
--- a/drivers/mtd/mtd_blkdevs.c
+++ b/drivers/mtd/mtd_blkdevs.c
@@ -35,7 +35,6 @@  static void blktrans_dev_release(struct kref *kref)
 	blk_mq_free_tag_set(dev->tag_set);
 	kfree(dev->tag_set);
 	put_disk(dev->disk);
-	list_del(&dev->list);
 	kfree(dev);
 }
 
@@ -350,7 +349,6 @@  int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new)
 		BUG();
 	}
 
-	mutex_lock(&blktrans_ref_mutex);
 	list_for_each_entry(d, &tr->devs, list) {
 		if (new->devnum == -1) {
 			/* Use first free number */
@@ -362,7 +360,6 @@  int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new)
 			}
 		} else if (d->devnum == new->devnum) {
 			/* Required number taken */
-			mutex_unlock(&blktrans_ref_mutex);
 			return -EBUSY;
 		} else if (d->devnum > new->devnum) {
 			/* Required number was free */
@@ -381,14 +378,12 @@  int add_mtd_blktrans_dev(struct mtd_blktrans_dev *new)
 	 * with this number. */
 	if (new->devnum > (MINORMASK >> tr->part_bits) ||
 	    (tr->part_bits && new->devnum >= 27 * 26)) {
-		mutex_unlock(&blktrans_ref_mutex);
 		goto error1;
 	}
 
 	list_add_tail(&new->list, &tr->devs);
- added:
-	mutex_unlock(&blktrans_ref_mutex);
 
+ added:
 	mutex_init(&new->lock);
 	kref_init(&new->ref);
 	if (!tr->writesect)
@@ -484,6 +479,8 @@  int del_mtd_blktrans_dev(struct mtd_blktrans_dev *old)
 		BUG();
 	}
 
+	list_del(&old->list);
+
 	if (old->disk_attributes)
 		sysfs_remove_group(&disk_to_dev(old->disk)->kobj,
 						old->disk_attributes);