Hi,
On 8/10/2023 6:16 PM, Toke Høiland-Jørgensen wrote:
> Hou Tao <houtao@xxxxxxxxxxxxxxx> writes:
>
>> From: Hou Tao <houtao1@xxxxxxxxxx>
>>
>> As for now __cpu_map_entry_replace() uses call_rcu() to wait for the
>> inflight xdp program and NAPI poll to exit the RCU read critical
>> section, and then launch kworker cpu_map_kthread_stop() to call
>> kthread_stop() to handle all pending xdp frames or skbs.
>>
>> But it is unnecessary to use rcu_barrier() in cpu_map_kthread_stop() to
>> wait for the completion of __cpu_map_entry_free(), because rcu_barrier()
>> will wait for all pending RCU callbacks and cpu_map_kthread_stop() only
>> needs to wait for the completion of a specific __cpu_map_entry_free().
>>
>> So use queue_rcu_work() to replace call_rcu(), schedule_work() and
>> rcu_barrier(). queue_rcu_work() will queue a __cpu_map_entry_free()
>> kworker after a RCU grace period. Because __cpu_map_entry_free() is
>> running in a kworker context, so it is OK to do all of these freeing
>> procedures include kthread_stop() in it.
>>
>> After the update, there is no need to do reference-counting for
>> bpf_cpu_map_entry, because bpf_cpu_map_entry is freed directly in
>> __cpu_map_entry_free(), so just remove it.
>>
>> Signed-off-by: Hou Tao <houtao1@xxxxxxxxxx>
> I think your analysis is correct, and this is a nice cleanup of what is
> really a bit of an over-complicated cleanup flow - well done!
>
> I have a few nits below, but with those feel free to resend as non-RFC
> and add my:
>
> Reviewed-by: Toke Høiland-Jørgensen <toke@xxxxxxxxxx>
Thanks for the review.
>
>> ---
>> kernel/bpf/cpumap.c | 93 +++++++++++----------------------------------
>> 1 file changed, 23 insertions(+), 70 deletions(-)
>>
SNIP
>> -static void __cpu_map_entry_free(struct rcu_head *rcu)
>> +static void __cpu_map_entry_free(struct work_struct *work)
>> {
>> struct bpf_cpu_map_entry *rcpu;
>>
>> @@ -503,30 +454,33 @@ static void __cpu_map_entry_free(struct rcu_head *rcu)
>> * new packets and cannot change/set flush_needed that can
>> * find this entry.
>> */
>> - rcpu = container_of(rcu, struct bpf_cpu_map_entry, rcu);
>> + rcpu = container_of(to_rcu_work(work), struct bpf_cpu_map_entry, free_work);
>>
>> free_percpu(rcpu->bulkq);
> Let's move this free down to the end along with the others.
Will do in v1.
>
>> - /* Cannot kthread_stop() here, last put free rcpu resources */
>> - put_cpu_map_entry(rcpu);
>> +
>> + /* kthread_stop will wake_up_process and wait for it to complete */
> Suggest adding to this comment: "cpu_map_kthread_run() makes sure the
> pointer ring is empty before exiting."
Will do in v1.
>
>> + kthread_stop(rcpu->kthread);
>> +
>> + if (rcpu->prog)
>> + bpf_prog_put(rcpu->prog);
>> + /* The queue should be empty at this point */
>> + __cpu_map_ring_cleanup(rcpu->queue);
>> + ptr_ring_cleanup(rcpu->queue, NULL);
>> + kfree(rcpu->queue);
>> + kfree(rcpu);
>> }
>>
>> /* After xchg pointer to bpf_cpu_map_entry, use the call_rcu() to
>> - * ensure any driver rcu critical sections have completed, but this
>> - * does not guarantee a flush has happened yet. Because driver side
>> - * rcu_read_lock/unlock only protects the running XDP program. The
>> - * atomic xchg and NULL-ptr check in __cpu_map_flush() makes sure a
>> - * pending flush op doesn't fail.
>> + * ensure both any driver rcu critical sections and xdp_do_flush()
>> + * have completed.
>> *
>> * The bpf_cpu_map_entry is still used by the kthread, and there can
>> - * still be pending packets (in queue and percpu bulkq). A refcnt
>> - * makes sure to last user (kthread_stop vs. call_rcu) free memory
>> - * resources.
>> + * still be pending packets (in queue and percpu bulkq).
>> *
>> - * The rcu callback __cpu_map_entry_free flush remaining packets in
>> - * percpu bulkq to queue. Due to caller map_delete_elem() disable
>> - * preemption, cannot call kthread_stop() to make sure queue is empty.
>> - * Instead a work_queue is started for stopping kthread,
>> - * cpu_map_kthread_stop, which waits for an RCU grace period before
>> + * Due to caller map_delete_elem() is in RCU read critical section,
>> + * cannot call kthread_stop() to make sure queue is empty. Instead
>> + * a work_struct is started for stopping kthread,
>> + * __cpu_map_entry_free, which waits for a RCU grace period before
>> * stopping kthread, emptying the queue.
>> */
> I think the above comment is a bit too convoluted, still. I'd suggest
> just replacing the whole thing with this:
>
> /* After the xchg of the bpf_cpu_map_entry pointer, we need to make sure the old
> * entry is no longer in use before freeing. We use queue_rcu_work() to call
> * __cpu_map_entry_free() in a separate workqueue after waiting for an RCU grace
> * period. This means that (a) all pending enqueue and flush operations have
> * completed (because or the RCU callback), and (b) we are in a workqueue
> * context where we can stop the kthread and wait for it to exit before freeing
> * everything.
> */
Much better. Thanks for the rephrasing. Will update it in v1.
>> static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
>> @@ -536,9 +490,8 @@ static void __cpu_map_entry_replace(struct bpf_cpu_map *cmap,
>>
>> old_rcpu = unrcu_pointer(xchg(&cmap->cpu_map[key_cpu], RCU_INITIALIZER(rcpu)));
>> if (old_rcpu) {
>> - call_rcu(&old_rcpu->rcu, __cpu_map_entry_free);
>> - INIT_WORK(&old_rcpu->kthread_stop_wq, cpu_map_kthread_stop);
>> - schedule_work(&old_rcpu->kthread_stop_wq);
>> + INIT_RCU_WORK(&old_rcpu->free_work, __cpu_map_entry_free);
>> + queue_rcu_work(system_wq, &old_rcpu->free_work);
>> }
>> }
>>
>> --
>> 2.29.2
