Re: [BUG] bpf: syscall: a possible sleep-in-atomic-context bug in map_update_elem()

From: Jia-Ju Bai
Date: Fri Aug 10 2018 - 21:57:31 EST




On 2018/8/10 22:22, Daniel Borkmann wrote:
On 08/10/2018 04:07 PM, Jia-Ju Bai wrote:
The kernel may sleep with holding a rcu read lock.

The function call paths (from bottom to top) in Linux-4.16 are:

[FUNC] kmalloc(GFP_KERNEL)
kernel/kthread.c, 283: kmalloc in __kthread_create_on_node
kernel/kthread.c, 365: __kthread_create_on_node in kthread_create_on_node
kernel/bpf/cpumap.c, 368: kthread_create_on_node in __cpu_map_entry_alloc
kernel/bpf/cpumap.c, 490: __cpu_map_entry_alloc in cpu_map_update_elem
kernel/bpf/syscall.c, 724: [FUNC_PTR]cpu_map_update_elem in map_update_elem
kernel/bpf/syscall.c, 723: rcu_read_lock in map_update_elem

Note that [FUNC_PTR] means a function pointer call is used.

I do not find a good way to fix it, so I only report.
This is found by my static analysis tool (DSAC).
Thanks for the report Jia-Ju! In the map_update_elem() from syscall
path there's a check map->map_type == BPF_MAP_TYPE_CPUMAP, where we
call the cpumap's map->ops->map_update_elem() while /not/ being under
rcu_read_lock() as in other cases, so looks okay to me. Could you point
out the case for being under rcu_read_lock() more specifically which
the tool found?

Thanks for your reply :)
My tool cannot accurately track the case of map->map_type at present...

According to my code review, there is a indeed check on line 697 in Linux-4.16:
else if (map->map_type == BPF_MAP_TYPE_CPUMAP) {
err = map->ops->map_update_elem(map, key, value, attr->flags);
goto out;
}
But there is a call to map->ops->map_update_elem() that is under rcu_read_lock on line 724:
rcu_read_lock();
err = map->ops->map_update_elem(map, key, value, attr->flags);
rcu_read_unlock();

So I think if map->map_type is not equal to BPF_MAP_TYPE_CPUMAP, map->ops->map_update_elem() can still be called under rcu_read_lock, is it right?


Best wishes,
Jia-Ju Bai