Re: [PATCH RFC net-next/mm V2 1/2] page_pool: Remove workqueue in new shutdown scheme - Toke Høiland-Jørgensen

From: "Toke Høiland-Jørgensen" <toke@redhat.com>
To: Jesper Dangaard Brouer <jbrouer@redhat.com>,
	Ilias Apalodimas <ilias.apalodimas@linaro.org>,
	netdev@vger.kernel.org, Eric Dumazet <eric.dumazet@gmail.com>,
	linux-mm@kvack.org, Mel Gorman <mgorman@techsingularity.net>
Cc: brouer@redhat.com, lorenzo@kernel.org, linyunsheng@huawei.com,
	bpf@vger.kernel.org, "David S. Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>, Paolo Abeni <pabeni@redhat.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	willy@infradead.org
Subject: Re: [PATCH RFC net-next/mm V2 1/2] page_pool: Remove workqueue in new shutdown scheme
Date: Fri, 28 Apr 2023 12:52:50 +0200	[thread overview]
Message-ID: <87mt2s6zy5.fsf@toke.dk> (raw)
In-Reply-To: <4eab92af-251a-a9aa-e270-179634d0345b@redhat.com>

Jesper Dangaard Brouer <jbrouer@redhat.com> writes:

> On 27/04/2023 22.53, Toke Høiland-Jørgensen wrote:
>>> +noinline
>>>   static void page_pool_empty_ring(struct page_pool *pool)
>>>   {
>>>   	struct page *page;
>>> @@ -796,39 +828,29 @@ static void page_pool_scrub(struct page_pool *pool)
>>>   	page_pool_empty_ring(pool);
>>>   }
>> So this is not in the diff context, but page_pool_empty_ring() does
>> this:
>> 
>> static void page_pool_empty_ring(struct page_pool *pool)
>> {
>> 	struct page *page;
>> 
>> 	/* Empty recycle ring */
>> 	while ((page = ptr_ring_consume_bh(&pool->ring))) {
>> 		/* Verify the refcnt invariant of cached pages */
>> 		if (!(page_ref_count(page) == 1))
>> 			pr_crit("%s() page_pool refcnt %d violation\n",
>> 				__func__, page_ref_count(page));
>> 
>> 		page_pool_return_page(pool, page);
>> 	}
>> }
>> 
>> ...and with this patch, that page_pool_return_page() call will now free
>> the pool memory entirely when the last page is returned. When it does
>> this, the condition in the while loop will still execute afterwards; it
>> would return false, but if the pool was freed, it's now referencing
>> freed memory when trying to read from pool->ring.
>
> Yes, that sounds like a problem.
>
>> So I think page_pool_empty_ring needs to either pull out all the pages
>> in the ring to an on-stack buffer before calling page_pool_return_page()
>> on them, or there needs to be some other way to break the loop early.
>
> Let me address this one first, I'll get back to the other in another
> reply.  The usual/idiom way of doing this is to have a next pointer that
> is populated inside the loop before freeing the object.
> It should look like this (only compile tested):
>
>   static void page_pool_empty_ring(struct page_pool *pool)
>   {
> 	struct page *page, *next;
>
> 	next = ptr_ring_consume_bh(&pool->ring);
>
> 	/* Empty recycle ring */
> 	while (next) {
> 		page = next;
> 		next = ptr_ring_consume_bh(&pool->ring);
>
> 		/* Verify the refcnt invariant of cached pages */
> 		if (!(page_ref_count(page) == 1))
> 			pr_crit("%s() page_pool refcnt %d violation\n",
> 				__func__, page_ref_count(page));
>
> 		page_pool_return_page(pool, page);
> 	}
>   }

Yup, that works!

>> There are a couple of other places where page_pool_return_page() is
>> called in a loop where the loop variable lives inside struct page_pool,
>> so we need to be absolutely sure they will never be called in the
>> shutdown stage, or they'll have to be fixed as well.
>
> The other loops are okay, but I spotted another problem in 
> __page_pool_put_page() in "Fallback/non-XDP mode", but that is fixable.

Alright, great!

-Toke