Re: [PATCH net-next v2 3/3] tun: AF_XDP Tx zero-copy support

["Michael S. Tsirkin" <mst@redhat.com>]

On Fri, Mar 01, 2024 at 11:45:52AM +0000, wangyunjian wrote:
> > -----Original Message-----
> > From: Paolo Abeni [mailto:pabeni@redhat.com]
> > Sent: Thursday, February 29, 2024 7:13 PM
> > To: wangyunjian <wangyunjian@huawei.com>; mst@redhat.com;
> > willemdebruijn.kernel@gmail.com; jasowang@redhat.com; kuba@kernel.org;
> > bjorn@kernel.org; magnus.karlsson@intel.com; maciej.fijalkowski@intel.com;
> > jonathan.lemon@gmail.com; davem@davemloft.net
> > Cc: bpf@vger.kernel.org; netdev@vger.kernel.org;
> > linux-kernel@vger.kernel.org; kvm@vger.kernel.org;
> > virtualization@lists.linux.dev; xudingke <xudingke@huawei.com>; liwei (DT)
> > <liwei395@huawei.com>
> > Subject: Re: [PATCH net-next v2 3/3] tun: AF_XDP Tx zero-copy support
> > 
> > On Wed, 2024-02-28 at 19:05 +0800, Yunjian Wang wrote:
> > > @@ -2661,6 +2776,54 @@ static int tun_ptr_peek_len(void *ptr)
> > >  	}
> > >  }
> > >
> > > +static void tun_peek_xsk(struct tun_file *tfile) {
> > > +	struct xsk_buff_pool *pool;
> > > +	u32 i, batch, budget;
> > > +	void *frame;
> > > +
> > > +	if (!ptr_ring_empty(&tfile->tx_ring))
> > > +		return;
> > > +
> > > +	spin_lock(&tfile->pool_lock);
> > > +	pool = tfile->xsk_pool;
> > > +	if (!pool) {
> > > +		spin_unlock(&tfile->pool_lock);
> > > +		return;
> > > +	}
> > > +
> > > +	if (tfile->nb_descs) {
> > > +		xsk_tx_completed(pool, tfile->nb_descs);
> > > +		if (xsk_uses_need_wakeup(pool))
> > > +			xsk_set_tx_need_wakeup(pool);
> > > +	}
> > > +
> > > +	spin_lock(&tfile->tx_ring.producer_lock);
> > > +	budget = min_t(u32, tfile->tx_ring.size, TUN_XDP_BATCH);
> > > +
> > > +	batch = xsk_tx_peek_release_desc_batch(pool, budget);
> > > +	if (!batch) {
> > 
> > This branch looks like an unneeded "optimization". The generic loop below
> > should have the same effect with no measurable perf delta - and smaller code.
> > Just remove this.
> > 
> > > +		tfile->nb_descs = 0;
> > > +		spin_unlock(&tfile->tx_ring.producer_lock);
> > > +		spin_unlock(&tfile->pool_lock);
> > > +		return;
> > > +	}
> > > +
> > > +	tfile->nb_descs = batch;
> > > +	for (i = 0; i < batch; i++) {
> > > +		/* Encode the XDP DESC flag into lowest bit for consumer to differ
> > > +		 * XDP desc from XDP buffer and sk_buff.
> > > +		 */
> > > +		frame = tun_xdp_desc_to_ptr(&pool->tx_descs[i]);
> > > +		/* The budget must be less than or equal to tx_ring.size,
> > > +		 * so enqueuing will not fail.
> > > +		 */
> > > +		__ptr_ring_produce(&tfile->tx_ring, frame);
> > > +	}
> > > +	spin_unlock(&tfile->tx_ring.producer_lock);
> > > +	spin_unlock(&tfile->pool_lock);
> > 
> > More related to the general design: it looks wrong. What if
> > get_rx_bufs() will fail (ENOBUF) after successful peeking? With no more
> > incoming packets, later peek will return 0 and it looks like that the
> > half-processed packets will stay in the ring forever???
> > 
> > I think the 'ring produce' part should be moved into tun_do_read().
> 
> Currently, the vhost-net obtains a batch descriptors/sk_buffs from the
> ptr_ring and enqueue the batch descriptors/sk_buffs to the virtqueue'queue,
> and then consumes the descriptors/sk_buffs from the virtqueue'queue in
> sequence. As a result, TUN does not know whether the batch descriptors have
> been used up, and thus does not know when to return the batch descriptors.
> 
> So, I think it's reasonable that when vhost-net checks ptr_ring is empty,
> it calls peek_len to get new xsk's descs and return the descriptors.
> 
> Thanks

What you need to think about is that if you peek, another call
in parallel can get the same value at the same time.


> > 
> > Cheers,
> > 
> > Paolo
>