[Question] Is there race between swapoff and swapout

[Question] Is there race between swapoff and swapout

[Chen Wandun]

I analysied the code about swapoff and swapout, and I suspected there may be a race.
The kernel version is 4.14 stable.

CPU0						CPU1
swapoff						swap out
						add_to_swap
							get_swap_page
	......							get_swap_pages						
									spin_lock(&swap_avail_lock)
									get swap_info_struct
									spin_unlock(&swap_avail_lock)		
	spin_lock(&swap_avail_lock)									
	__def_from_avail_list(swap_info_struct)								
     	spin_unlock(&swap_avail_lock)					......	
	try_to_unuse  // unuse all slot
									/* get a free slot from swap_info_struct,
									 * and write data to slot later
									 */	
									scan_swap_map_slots
	free swap_info_struct
	.......

	
If CPU1 get the swap_info_struct first, then CPU0 delete it from list and
unuse all slot in swap_info_struct, before CPU0 free swap_info_struct CPU1
call scan_swap_map_slots to alloc a free slot.

I am not sure the analysis above is correct,
Please let me know if there is any mistake

Thanks
ChenWandun

Re: [Question] Is there race between swapoff and swapout

[Huang, Ying]

Chen Wandun <chenwandun@huawei.com> writes:

> I analysied the code about swapoff and swapout, and I suspected there may be a race.
> The kernel version is 4.14 stable.
>
> CPU0						CPU1
> swapoff						swap out
> 						add_to_swap
> 							get_swap_page
> 	......							get_swap_pages						
> 									spin_lock(&swap_avail_lock)
> 									get swap_info_struct
> 									spin_unlock(&swap_avail_lock)		
> 	spin_lock(&swap_avail_lock)									
> 	__def_from_avail_list(swap_info_struct)								
>     	spin_unlock(&swap_avail_lock)					......	
> 	try_to_unuse  // unuse all slot
> 									/* get a free slot from swap_info_struct,
> 									 * and write data to slot later
> 									 */	
> 									scan_swap_map_slots
> 	free swap_info_struct
> 	.......
>
> 	
> If CPU1 get the swap_info_struct first, then CPU0 delete it from list and
> unuse all slot in swap_info_struct, before CPU0 free swap_info_struct CPU1
> call scan_swap_map_slots to alloc a free slot.
>
> I am not sure the analysis above is correct,
> Please let me know if there is any mistake

SWP_WRITEOK will be cleared during swapoff, and it is checked during
swap slots allocation.

Best Regards,
Huang, Ying

> Thanks
> ChenWandun