Re: [PATCH] bpf: Convert lpm_trie::lock to 'raw_spinlock_t'

[Thomas Gleixner]

On Fri, Nov 08 2024 at 14:32, Kunwu Chan wrote:
> When PREEMPT_RT is enabled, 'spinlock_t' becomes preemptible
> and bpf program has owned a raw_spinlock under a interrupt handler,
> which results in invalid lock acquire context.

This explanation is just wrong.

The problem has nothing to do with an interrupt handler. Interrupt
handlers on RT kernels are force threaded.

>  __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline]
>  _raw_spin_lock_irqsave+0xd5/0x120 kernel/locking/spinlock.c:162
>  trie_delete_elem+0x96/0x6a0 kernel/bpf/lpm_trie.c:462
>  bpf_prog_2c29ac5cdc6b1842+0x43/0x47
>  bpf_dispatcher_nop_func include/linux/bpf.h:1290 [inline]
>  __bpf_prog_run include/linux/filter.h:701 [inline]
>  bpf_prog_run include/linux/filter.h:708 [inline]
>  __bpf_trace_run kernel/trace/bpf_trace.c:2340 [inline]
>  bpf_trace_run1+0x2ca/0x520 kernel/trace/bpf_trace.c:2380
>  trace_workqueue_activate_work+0x186/0x1f0 include/trace/events/workqueue.h:59
>  __queue_work+0xc7b/0xf50 kernel/workqueue.c:2338

The problematic lock nesting is the work queue pool lock, which is a raw
spinlock.

> @@ -330,7 +330,7 @@ static long trie_update_elem(struct bpf_map *map,
>  	if (key->prefixlen > trie->max_prefixlen)
>  		return -EINVAL;
>  
> -	spin_lock_irqsave(&trie->lock, irq_flags);
> +	raw_spin_lock_irqsave(&trie->lock, irq_flags);
>  
>  	/* Allocate and fill a new node */

Making this a raw spinlock moves the problem from the BPF trie code into
the memory allocator. On RT the memory allocator cannot be invoked under
a raw spinlock.

Thanks,

        tglx