MIME-Version: 1.0
References: <20230202024417.4477-1-dqfext@gmail.com> <ec42f238-8fc7-2ea4-c1a7-e4c3c4b8f512@redhat.com>
In-Reply-To: <ec42f238-8fc7-2ea4-c1a7-e4c3c4b8f512@redhat.com>
From:   DENG Qingfang <dqfext@gmail.com>
Date:   Fri, 3 Feb 2023 21:05:41 +0800
Message-ID: <CALW65ja+P+E0fjEsZfm1XWb_dn_snuRoFA5i_i+_1K9j0+wi7Q@mail.gmail.com>
Subject: Re: [PATCH net] net: page_pool: use in_softirq() instead
To:     Jesper Dangaard Brouer <jbrouer@redhat.com>
Cc:     Jesper Dangaard Brouer <hawk@kernel.org>,
        Ilias Apalodimas <ilias.apalodimas@linaro.org>,
        "David S. Miller" <davem@davemloft.net>,
        Eric Dumazet <edumazet@google.com>,
        Jakub Kicinski <kuba@kernel.org>,
        Paolo Abeni <pabeni@redhat.com>,
        Lorenzo Bianconi <lorenzo@kernel.org>,
        Daniel Borkmann <daniel@iogearbox.net>,
        John Fastabend <john.fastabend@gmail.com>,
        netdev@vger.kernel.org, linux-kernel@vger.kernel.org,
        brouer@redhat.com
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

Hi Jesper,

On Fri, Feb 3, 2023 at 7:15 PM Jesper Dangaard Brouer
<jbrouer@redhat.com> wrote:
> How can I enable threaded NAPI on my system?

dev_set_threaded(napi_dev, true);

You can also enable it at runtime by writing 1 to
/sys/class/net/<devname>/threaded, but it only works if the driver
does _not_ use a dummy netdev for NAPI poll.

> I think other cases (above) are likely safe, but I worry a little about
> this case, as the page_pool_recycle_in_cache() rely on RX-NAPI protection.
> Meaning it is only the CPU that handles RX-NAPI for this RX-queue that
> is allowed to access this lockless array.

The major changes to the threaded NAPI is that instead of scheduling a
NET_RX softirq, it wakes up a kthread which then does the polling,
allowing it to be scheduled to another CPU. The driver's poll function
is called with BH disabled so it's still considered BH context.

> We do have the 'allow_direct' boolean, and if every driver/user uses
> this correctly, then this should be safe.  Changing this makes it
> possible for drivers to use page_pool API incorrectly and this leads to
> hard-to-debug errors.

"incorrectly", like, calling it outside RX-NAPI?