RE: [PATCH net-next, 3/4] hv_netvsc: Add XDP support

[Haiyang Zhang]

> -----Original Message-----
> From: Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx>
> Sent: Monday, October 28, 2019 5:33 PM
> To: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>
> Cc: sashal@xxxxxxxxxx; linux-hyperv@xxxxxxxxxxxxxxx;
> netdev@xxxxxxxxxxxxxxx; KY Srinivasan <kys@xxxxxxxxxxxxx>; Stephen
> Hemminger <sthemmin@xxxxxxxxxxxxx>; olaf@xxxxxxxxx; vkuznets
> <vkuznets@xxxxxxxxxx>; davem@xxxxxxxxxxxxx; linux-
> kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH net-next, 3/4] hv_netvsc: Add XDP support
> 
> On Mon, 28 Oct 2019 21:07:04 +0000, Haiyang Zhang wrote:
> > This patch adds support of XDP in native mode for hv_netvsc driver, and
> > transparently sets the XDP program on the associated VF NIC as well.
> >
> > XDP program cannot run with LRO (RSC) enabled, so you need to disable
> LRO
> > before running XDP:
> >         ethtool -K eth0 lro off
> >
> > XDP actions not yet supported:
> >         XDP_TX, XDP_REDIRECT
> 
> I don't think we want to merge support without at least XDP_TX these
> days..
Thanks for your detailed comments --
I'm working on the XDP_TX...

> 
> And without the ability to prepend headers this may be the least
> complete initial XDP implementation we've seen :(
The RNDIS packet buffer received by netvsc doesn't have a head room, but I'm
considering copy the packets to the page buffer, with a head room space 
reserved for XDP.

> 
> > Signed-off-by: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>
> 
> > diff --git a/drivers/net/hyperv/netvsc.c b/drivers/net/hyperv/netvsc.c
> > index d22a36f..688487b 100644
> > --- a/drivers/net/hyperv/netvsc.c
> > +++ b/drivers/net/hyperv/netvsc.c
> > @@ -122,8 +122,10 @@ static void free_netvsc_device(struct rcu_head
> *head)
> >  	vfree(nvdev->send_buf);
> >  	kfree(nvdev->send_section_map);
> >
> > -	for (i = 0; i < VRSS_CHANNEL_MAX; i++)
> > +	for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
> > +		xdp_rxq_info_unreg(&nvdev->chan_table[i].xdp_rxq);
> >  		vfree(nvdev->chan_table[i].mrc.slots);
> > +	}
> >
> >  	kfree(nvdev);
> >  }
> > @@ -1370,6 +1372,10 @@ struct netvsc_device *netvsc_device_add(struct
> hv_device *device,
> >  		nvchan->net_device = net_device;
> >  		u64_stats_init(&nvchan->tx_stats.syncp);
> >  		u64_stats_init(&nvchan->rx_stats.syncp);
> > +
> > +		xdp_rxq_info_reg(&nvchan->xdp_rxq, ndev, i);
> > +		xdp_rxq_info_reg_mem_model(&nvchan->xdp_rxq,
> > +					   MEM_TYPE_PAGE_SHARED, NULL);
> 
> These can fail.
I will add error handling.

> 
> >  	}
> >
> >  	/* Enable NAPI handler before init callbacks */
> > diff --git a/drivers/net/hyperv/netvsc_bpf.c
> b/drivers/net/hyperv/netvsc_bpf.c
> > new file mode 100644
> > index 0000000..4d235ac
> > --- /dev/null
> > +++ b/drivers/net/hyperv/netvsc_bpf.c
> > @@ -0,0 +1,211 @@
> > +// SPDX-License-Identifier: GPL-2.0-only
> > +/* Copyright (c) 2019, Microsoft Corporation.
> > + *
> > + * Author:
> > + *   Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>
> > + */
> > +
> > +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
> > +
> > +#include <linux/netdevice.h>
> > +#include <linux/etherdevice.h>
> > +#include <linux/ethtool.h>
> > +#include <linux/bpf.h>
> > +#include <linux/bpf_trace.h>
> > +#include <linux/kernel.h>
> > +#include <net/xdp.h>
> > +
> > +#include <linux/mutex.h>
> > +#include <linux/rtnetlink.h>
> > +
> > +#include "hyperv_net.h"
> > +
> > +u32 netvsc_run_xdp(struct net_device *ndev, struct netvsc_channel
> *nvchan,
> > +		   void **p_pbuf)
> > +{
> > +	struct page *page = NULL;
> > +	void *data = nvchan->rsc.data[0];
> > +	u32 len = nvchan->rsc.len[0];
> > +	void *pbuf = data;
> > +	struct bpf_prog *prog;
> > +	struct xdp_buff xdp;
> > +	u32 act = XDP_PASS;
> > +
> > +	*p_pbuf = NULL;
> > +
> > +	rcu_read_lock();
> > +	prog = rcu_dereference(nvchan->bpf_prog);
> > +
> > +	if (!prog || nvchan->rsc.cnt > 1)
> 
> Can rsc.cnt == 1 not be ensured at setup time? This looks quite
> limiting if random frames could be forced to bypass the filter.
Yes, the setup code already check/ensure LRO is disabled. So rsc.cnt > 1
is NOT expected here. Just an error check. I will change the return value
to XDP_ABORTED for this.

> 
> > +		goto out;
> > +
> > +	/* copy to a new page buffer if data are not within a page */
> > +	if (virt_to_page(data) != virt_to_page(data + len - 1)) {
> > +		page = alloc_page(GFP_ATOMIC);
> > +		if (!page)
> > +			goto out;
> 
> Returning XDP_PASS on allocation failure seems highly questionable.
I will change the return value to XDP_ABORTED for this too.

> 
> > +		pbuf = page_address(page);
> > +		memcpy(pbuf, nvchan->rsc.data[0], len);
> > +
> > +		*p_pbuf = pbuf;
> > +	}
> > +
> > +	xdp.data_hard_start = pbuf;
> > +	xdp.data = xdp.data_hard_start;
> 
> This patch also doesn't add any headroom for XDP to prepend data :(
I'm considering to add the headroom to the start of the page.

> 
> > +	xdp_set_data_meta_invalid(&xdp);
> > +	xdp.data_end = xdp.data + len;
> > +	xdp.rxq = &nvchan->xdp_rxq;
> > +	xdp.handle = 0;
> > +
> > +	act = bpf_prog_run_xdp(prog, &xdp);
> > +
> > +	switch (act) {
> > +	case XDP_PASS:
> > +		/* Pass to upper layers */
> > +		break;
> > +
> > +	case XDP_ABORTED:
> > +		trace_xdp_exception(ndev, prog, act);
> > +		break;
> > +
> > +	case XDP_DROP:
> > +		break;
> > +
> > +	default:
> > +		bpf_warn_invalid_xdp_action(act);
> > +	}
> > +
> > +out:
> > +	rcu_read_unlock();
> > +
> > +	if (page && act != XDP_PASS) {
> > +		*p_pbuf = NULL;
> > +		__free_page(page);
> > +	}
> > +
> > +	return act;
> > +}
> > +
> > +unsigned int netvsc_xdp_fraglen(unsigned int len)
> > +{
> > +	return SKB_DATA_ALIGN(len) +
> > +	       SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
> > +}
> > +
> > +struct bpf_prog *netvsc_xdp_get(struct netvsc_device *nvdev)
> > +{
> > +	return rtnl_dereference(nvdev->chan_table[0].bpf_prog);
> > +}
> > +
> > +int netvsc_xdp_set(struct net_device *dev, struct bpf_prog *prog,
> > +		   struct netvsc_device *nvdev)
> > +{
> > +	struct bpf_prog *old_prog;
> > +	int frag_max, i;
> > +
> > +	old_prog = netvsc_xdp_get(nvdev);
> > +
> > +	if (!old_prog && !prog)
> > +		return 0;
> 
> I think this case is now handled by the core.
Thanks for the reminder. I saw the code in dev_change_xdp_fd(), so the upper layer
doesn't call XDP_SETUP_PROG with old/new prog both NULL.
But this function is also called by other functions in our driver, like netvsc_detach(),
netvsc_remove(), etc. Instead of checking for NULL in each place, I still keep the check inside
netvsc_xdp_set().

> 
> > +	frag_max = netvsc_xdp_fraglen(dev->mtu + ETH_HLEN);
> > +	if (prog && frag_max > PAGE_SIZE) {
> > +		netdev_err(dev, "XDP: mtu:%u too large, frag:%u\n",
> > +			   dev->mtu, frag_max);
> > +		return -EOPNOTSUPP;
> > +	}
> > +
> > +	if (prog && (dev->features & NETIF_F_LRO)) {
> > +		netdev_err(dev, "XDP: not support LRO\n");
> 
> Please report this via extack, that way users will see it in the console
> in which they're installing the program.
I will.

> 
> > +		return -EOPNOTSUPP;
> > +	}
> > +
> > +	if (prog) {
> > +		prog = bpf_prog_add(prog, nvdev->num_chn);
> > +		if (IS_ERR(prog))
> > +			return PTR_ERR(prog);
> > +	}
> > +
> > +	for (i = 0; i < nvdev->num_chn; i++)
> > +		rcu_assign_pointer(nvdev->chan_table[i].bpf_prog, prog);
> > +
> > +	if (old_prog)
> > +		for (i = 0; i < nvdev->num_chn; i++)
> > +			bpf_prog_put(old_prog);
> > +
> > +	return 0;
> > +}
> > +
> > +int netvsc_vf_setxdp(struct net_device *vf_netdev, struct bpf_prog
> *prog)
> > +{
> > +	struct netdev_bpf xdp;
> > +	bpf_op_t ndo_bpf;
> > +
> > +	ASSERT_RTNL();
> > +
> > +	if (!vf_netdev)
> > +		return 0;
> > +
> > +	ndo_bpf = vf_netdev->netdev_ops->ndo_bpf;
> > +	if (!ndo_bpf)
> > +		return 0;
> > +
> > +	memset(&xdp, 0, sizeof(xdp));
> > +
> > +	xdp.command = XDP_SETUP_PROG;
> > +	xdp.prog = prog;
> > +
> > +	return ndo_bpf(vf_netdev, &xdp);
> 
> IMHO the automatic propagation is not a good idea. Especially if the
> propagation doesn't make the entire installation fail if VF doesn't
> have ndo_bpf.

On Hyperv and Azure hosts, VF is always acting as a slave below netvsc.
And they are both active -- most data packets go to VF, but broadcast,
multicast, and TCP SYN packets go to netvsc synthetic data path. The synthetic 
NIC (netvsc) is also a failover NIC when VF is not available.
We ask customers to only use the synthetic NIC directly. So propagation
of XDP setting to VF NIC is desired. 
But, I will change the return code to error, so the entire installation fails if VF is 
present but unable to set XDP prog.

Thanks,
- Haiyang