Re: [PATCH net v2 1/2] page_pool: fix timing for checking and disabling napi_local

On 2024/10/9 8:40, Jakub Kicinski wrote:

On Wed, 25 Sep 2024 15:57:06 +0800 Yunsheng Lin wrote:

quoted

Use rcu mechanism to avoid the above concurrent access problem.

Note, the above was found during code reviewing on how to fix
the problem in [1].

The driver must make sure NAPI cannot be running while
page_pool_destroy() is called. There's even an WARN()
checking this.. if you know what to look for.

I am guessing you are referring to the WARN() in
page_pool_disable_direct_recycling(), right?
If yes, I am aware of that WARN().

The problem is that at least from the skb_defer_free_flush()
case, it is not tied to any specific napi instance. When
skb_attempt_defer_free() calls kick_defer_list_purge() to
trigger running of net_rx_action(), skb_defer_free_flush() can
be called without tieing to any specific napi instance as my
understanding:
https://elixir.bootlin.com/linux/v6.7-rc8/source/net/core/dev.c#L6719

Or I am missing something obvious here? I even used below diff to
verify that and it did trigger without any napi in the sd->poll_list:

@@ -6313,6 +6313,9 @@ static void skb_defer_free_flush(struct softnet_data *sd)
        spin_unlock(&sd->defer_lock);

        while (skb != NULL) {
+               if (list_empty(&sd->poll_list))
+                       pr_err("defer freeing: %px with empty napi list\n", skb);
+
                next = skb->next;
                napi_consume_skb(skb, 1);
                skb = next;