Re: [PATCH v0] kernfs: Skip kernfs_put of parent from child node

[Gaurav Kohli <gkohli@codeaurora.org>]

On 4/5/2019 5:03 PM, Greg KH wrote:
> On Fri, Apr 05, 2019 at 04:51:07PM +0530, Gaurav Kohli wrote:
>> While adding kernfs node for child to the parent kernfs
>> node and when child node founds that parent kn count is
>> zero, then below comes like:
>>
>> WARNING: fs/kernfs/dir.c:494 kernfs_get+0x64/0x88
>>
>> This indicates that parent is in kernfs_put path/ or already
>> freed, and if the child node keeps continue to
>> make new kernfs node, then there is chance of
>> below race for parent node:
>>
>> CPU0				         CPU1
>> //Parent node			         //child node
>> kernfs_put
>> atomic_dec_and_test(&kn->count)
>> //count is 0, so continue
>> 					  kernfs_new_node(child)
>> 					  kernfs_get(parent);
>> 					  //increment parent count to 1
>> 					  //warning come as parent count is 0
>> 					  /* link in */
>> 					  kernfs_add_one(kn);
>> 					  // this should fail as parent is
>> 					  //in free path.
>> 					  kernfs_put(child)
>> kmem_cache_free(parent)
>> 					  kmem_cache_free(child)
>> 					  kn = parent
>> 					  atomic_dec_and_test(&kn->count))
>> 					  //this is 0 now, so release will
>> 					  continue for parent.
>> 					  kmem_cache_free(parent)
>>
>> To prevent this race, child simply has to decrement count of parent
>> kernfs node and keep continue the free path for itself.
>>
>> Signed-off-by: Gaurav Kohli <gkohli@codeaurora.org>
>> Signed-off-by: Mukesh Ojha <mojha@codeaurora.org>
>>
>> diff --git a/fs/kernfs/dir.c b/fs/kernfs/dir.c
>> index b84d635..d5a36e8 100644
>> --- a/fs/kernfs/dir.c
>> +++ b/fs/kernfs/dir.c
>> @@ -515,7 +515,6 @@ void kernfs_put(struct kernfs_node *kn)
>>   	if (!kn || !atomic_dec_and_test(&kn->count))
>>   		return;
>>   	root = kernfs_root(kn);
>> - repeat:
>>   	/*
>>   	 * Moving/renaming is always done while holding reference.
>>   	 * kn->parent won't change beneath us.
>> @@ -545,8 +544,8 @@ void kernfs_put(struct kernfs_node *kn)
>>   
>>   	kn = parent;
>>   	if (kn) {
>> -		if (atomic_dec_and_test(&kn->count))
>> -			goto repeat;
>> +	/* Parent may be on free path, so simply decrement the count */
> That's the wrong indentation :(
>
> And how are you hitting this issue?  What user of kernfs is causing
> this?

Hi Greg,

Thanks,  will fix comment indentation, seen during sys-executor running:

We have only one instance , In logs below warning also came:

WARNING: CPU: 4 kernel/msm-4.14/fs/kernfs/dir.c:494 kernfs_get+0x64/0x88

which indicated parent is in put path.

[  160.125151] Disabling lock debugging due to kernel taint
[  160.130626] INFO: Allocated in __kernfs_new_node+0x8c/0x3c0 age=11 
cpu=2 pid=7098
[  160.138314]     kmem_cache_alloc+0x358/0x388
[  160.142445]     __kernfs_new_node+0x8c/0x3c0
[  160.146590]     kernfs_new_node+0x80/0xc8
[  160.150462]     kernfs_create_dir_ns+0x44/0xfc
[  160.154777]     sysfs_create_dir_ns+0xa8/0x130
[  160.158416] CPU5: update max cpu_capacity 1024
[  160.159085]     kobject_add_internal+0x278/0x650
[  160.163567]     kobject_add_varg+0xe0/0x130
[  160.167606]     kobject_add+0x15c/0x1d0
[  160.168452] CPU5: update max cpu_capacity 780
[  160.171287]     get_device_parent+0x2d0/0x34c
[  160.175510]     device_add+0x240/0xde0
[  160.178371] CPU6: update max cpu_capacity 916
[  160.179108]     input_register_device+0x5f4/0xa0c
[  160.183686]     uinput_ioctl_handler+0x1184/0x2198
[  160.202436] INFO: Freed in kernfs_put+0x2c8/0x434 age=14 cpu=0 pid=7096
[  160.209230]     kernfs_put+0x2c8/0x434
[  160.212825]     kobject_del+0x50/0xcc
[  160.216332]     cleanup_glue_dir+0x124/0x16c
[  160.220456]     device_del+0x55c/0x5c8
[  160.224047]     __input_unregister_device+0x274/0x2a8
[  160.228974]     input_unregister_device+0x90/0xd0
[  160.233553]     uinput_destroy_device+0x15c/0x1dc
[  160.238131]     uinput_release+0x44/0x5c
[  160.241898]     __fput+0x1f4/0x4e4
[  160.245127]     ____fput+0x20/0x2c


during code review, I have found race between kernfs parent put call and 
child get call.

Regards

Gaurav

>
> thanks,
>
> greg k-h

-- 
Qualcomm India Private Limited, on behalf of Qualcomm Innovation Center,
Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project.