diff mbox

[dpdk-dev,v2] vhost: Add indirect descriptors support to the TX path

Message ID 1474615009-26626-1-git-send-email-maxime.coquelin@redhat.com (mailing list archive)
State Superseded, archived
Delegated to: Yuanhan Liu
Headers

Commit Message

Maxime Coquelin Sept. 23, 2016, 7:16 a.m. UTC 
  Indirect descriptors are usually supported by virtio-net devices,
allowing to dispatch a larger number of requests.

When the virtio device sends a packet using indirect descriptors,
only one slot is used in the ring, even for large packets.

The main effect is to improve the 0% packet loss benchmark.
A PVP benchmark using Moongen (64 bytes) on the TE, and testpmd
(fwd io for host, macswap for VM) on DUT shows a +50% gain for
zero loss.

On the downside, micro-benchmark using testpmd txonly in VM and
rxonly on host shows a loss between 1 and 4%.i But depending on
the needs, feature can be disabled at VM boot time by passing
indirect_desc=off argument to vhost-user device in Qemu.

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
---
Changes since v2:
-----------------
 - Enrich commit message with figures
 - Rebased on top of dpdk-next-virtio's master
 - Add feature check to ensure we don't receive an indirect desc
if not supported by the virtio driver

 lib/librte_vhost/vhost.c      |  3 ++-
 lib/librte_vhost/virtio_net.c | 38 +++++++++++++++++++++++++++++---------
 2 files changed, 31 insertions(+), 10 deletions(-)

Comments

Yuanhan Liu Sept. 23, 2016, 7:29 a.m. UTC | #1 
On Fri, Sep 23, 2016 at 09:16:49AM +0200, Maxime Coquelin wrote:
> +		if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
> +			if (unlikely(!(dev->features &
> +					(1ULL << VIRTIO_RING_F_INDIRECT_DESC)))) {
> +				RTE_LOG(ERR, VHOST_DATA,
> +						"Indirect desc but feature not negotiated.\n");
> +				break;
> +			}

I thought the alignment we got before was to follow linux kernel: check
nested indirect only?

> +
> +			desc = (struct vring_desc *)gpa_to_vva(dev,
> +					vq->desc[desc_indexes[i]].addr);

I think we should check the desc addr here. Otherwise we may crash here
if a malicious guest fills some bad addresses.

	--yliu
Maxime Coquelin Sept. 23, 2016, 7:33 a.m. UTC | #2 
On 09/23/2016 09:29 AM, Yuanhan Liu wrote:
> On Fri, Sep 23, 2016 at 09:16:49AM +0200, Maxime Coquelin wrote:
>> +		if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
>> +			if (unlikely(!(dev->features &
>> +					(1ULL << VIRTIO_RING_F_INDIRECT_DESC)))) {
>> +				RTE_LOG(ERR, VHOST_DATA,
>> +						"Indirect desc but feature not negotiated.\n");
>> +				break;
>> +			}
>
> I thought the alignment we got before was to follow linux kernel: check
> nested indirect only?
Right... I did the opposite..
Fixing this right now.

>
>> +
>> +			desc = (struct vring_desc *)gpa_to_vva(dev,
>> +					vq->desc[desc_indexes[i]].addr);
>
> I think we should check the desc addr here. Otherwise we may crash here
> if a malicious guest fills some bad addresses.
Good point!

Thanks,
Maxime
>
> 	--yliu
>
Stephen Hemminger Sept. 23, 2016, 6:24 p.m. UTC | #3 
On Fri, 23 Sep 2016 09:16:49 +0200
Maxime Coquelin <maxime.coquelin@redhat.com> wrote:

> Indirect descriptors are usually supported by virtio-net devices,
> allowing to dispatch a larger number of requests.
> 
> When the virtio device sends a packet using indirect descriptors,
> only one slot is used in the ring, even for large packets.
> 
> The main effect is to improve the 0% packet loss benchmark.
> A PVP benchmark using Moongen (64 bytes) on the TE, and testpmd
> (fwd io for host, macswap for VM) on DUT shows a +50% gain for
> zero loss.
> 
> On the downside, micro-benchmark using testpmd txonly in VM and
> rxonly on host shows a loss between 1 and 4%.i But depending on
> the needs, feature can be disabled at VM boot time by passing
> indirect_desc=off argument to vhost-user device in Qemu.
> 
> Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>

What about supporting VIRTIO_F_ANY_LAYOUT?
Michael S. Tsirkin Sept. 23, 2016, 6:31 p.m. UTC | #4 
On Fri, Sep 23, 2016 at 11:24:16AM -0700, Stephen Hemminger wrote:
> On Fri, 23 Sep 2016 09:16:49 +0200
> Maxime Coquelin <maxime.coquelin@redhat.com> wrote:
> 
> > Indirect descriptors are usually supported by virtio-net devices,
> > allowing to dispatch a larger number of requests.
> > 
> > When the virtio device sends a packet using indirect descriptors,
> > only one slot is used in the ring, even for large packets.
> > 
> > The main effect is to improve the 0% packet loss benchmark.
> > A PVP benchmark using Moongen (64 bytes) on the TE, and testpmd
> > (fwd io for host, macswap for VM) on DUT shows a +50% gain for
> > zero loss.
> > 
> > On the downside, micro-benchmark using testpmd txonly in VM and
> > rxonly on host shows a loss between 1 and 4%.i But depending on
> > the needs, feature can be disabled at VM boot time by passing
> > indirect_desc=off argument to vhost-user device in Qemu.
> > 
> > Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
> 
> What about supporting VIRTIO_F_ANY_LAYOUT?

I thought it's already supported.
That's required by virtio 1 and dpdk claims support for that.
Stephen Hemminger Sept. 23, 2016, 8:28 p.m. UTC | #5 
On Fri, 23 Sep 2016 21:31:27 +0300
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> On Fri, Sep 23, 2016 at 11:24:16AM -0700, Stephen Hemminger wrote:
> > On Fri, 23 Sep 2016 09:16:49 +0200
> > Maxime Coquelin <maxime.coquelin@redhat.com> wrote:
> >   
> > > Indirect descriptors are usually supported by virtio-net devices,
> > > allowing to dispatch a larger number of requests.
> > > 
> > > When the virtio device sends a packet using indirect descriptors,
> > > only one slot is used in the ring, even for large packets.
> > > 
> > > The main effect is to improve the 0% packet loss benchmark.
> > > A PVP benchmark using Moongen (64 bytes) on the TE, and testpmd
> > > (fwd io for host, macswap for VM) on DUT shows a +50% gain for
> > > zero loss.
> > > 
> > > On the downside, micro-benchmark using testpmd txonly in VM and
> > > rxonly on host shows a loss between 1 and 4%.i But depending on
> > > the needs, feature can be disabled at VM boot time by passing
> > > indirect_desc=off argument to vhost-user device in Qemu.
> > > 
> > > Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>  
> > 
> > What about supporting VIRTIO_F_ANY_LAYOUT?  
> 
> I thought it's already supported.
> That's required by virtio 1 and dpdk claims support for that.
> 

I don't see the flag set in the DPDK vhost driver feature bits
(at least in the source). 

/* Features supported by this lib. */
#define VHOST_SUPPORTED_FEATURES ((1ULL << VIRTIO_NET_F_MRG_RXBUF) | \
				(1ULL << VIRTIO_NET_F_CTRL_VQ) | \
				(1ULL << VIRTIO_NET_F_CTRL_RX) | \
				(1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) | \
				(VHOST_SUPPORTS_MQ)            | \
				(1ULL << VIRTIO_F_VERSION_1)   | \
				(1ULL << VHOST_F_LOG_ALL)      | \
				(1ULL << VHOST_USER_F_PROTOCOL_FEATURES) | \
				(1ULL << VIRTIO_NET_F_HOST_TSO4) | \
				(1ULL << VIRTIO_NET_F_HOST_TSO6) | \
				(1ULL << VIRTIO_NET_F_CSUM)    | \
				(1ULL << VIRTIO_NET_F_GUEST_CSUM) | \
				(1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
				(1ULL << VIRTIO_NET_F_GUEST_TSO6))
Michael S. Tsirkin Sept. 25, 2016, 1:02 a.m. UTC | #6 
On Fri, Sep 23, 2016 at 01:28:05PM -0700, Stephen Hemminger wrote:
> On Fri, 23 Sep 2016 21:31:27 +0300
> "Michael S. Tsirkin" <mst@redhat.com> wrote:
> 
> > On Fri, Sep 23, 2016 at 11:24:16AM -0700, Stephen Hemminger wrote:
> > > On Fri, 23 Sep 2016 09:16:49 +0200
> > > Maxime Coquelin <maxime.coquelin@redhat.com> wrote:
> > >   
> > > > Indirect descriptors are usually supported by virtio-net devices,
> > > > allowing to dispatch a larger number of requests.
> > > > 
> > > > When the virtio device sends a packet using indirect descriptors,
> > > > only one slot is used in the ring, even for large packets.
> > > > 
> > > > The main effect is to improve the 0% packet loss benchmark.
> > > > A PVP benchmark using Moongen (64 bytes) on the TE, and testpmd
> > > > (fwd io for host, macswap for VM) on DUT shows a +50% gain for
> > > > zero loss.
> > > > 
> > > > On the downside, micro-benchmark using testpmd txonly in VM and
> > > > rxonly on host shows a loss between 1 and 4%.i But depending on
> > > > the needs, feature can be disabled at VM boot time by passing
> > > > indirect_desc=off argument to vhost-user device in Qemu.
> > > > 
> > > > Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>  
> > > 
> > > What about supporting VIRTIO_F_ANY_LAYOUT?  
> > 
> > I thought it's already supported.
> > That's required by virtio 1 and dpdk claims support for that.
> > 
> 
> I don't see the flag set in the DPDK vhost driver feature bits
> (at least in the source). 
> 
> /* Features supported by this lib. */
> #define VHOST_SUPPORTED_FEATURES ((1ULL << VIRTIO_NET_F_MRG_RXBUF) | \
> 				(1ULL << VIRTIO_NET_F_CTRL_VQ) | \
> 				(1ULL << VIRTIO_NET_F_CTRL_RX) | \
> 				(1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) | \
> 				(VHOST_SUPPORTS_MQ)            | \
> 				(1ULL << VIRTIO_F_VERSION_1)   | \
> 				(1ULL << VHOST_F_LOG_ALL)      | \
> 				(1ULL << VHOST_USER_F_PROTOCOL_FEATURES) | \
> 				(1ULL << VIRTIO_NET_F_HOST_TSO4) | \
> 				(1ULL << VIRTIO_NET_F_HOST_TSO6) | \
> 				(1ULL << VIRTIO_NET_F_CSUM)    | \
> 				(1ULL << VIRTIO_NET_F_GUEST_CSUM) | \
> 				(1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
> 				(1ULL << VIRTIO_NET_F_GUEST_TSO6))


I see. It's implied by VERSION_1 in fact.
In other words if VERSION_1 is negotiated then bit 27
isn't ANY_LAYOUT, it's in fact reserved.
Stephen Hemminger Sept. 25, 2016, 1:50 a.m. UTC | #7 
On Sun, 25 Sep 2016 04:02:28 +0300
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> I see. It's implied by VERSION_1 in fact.
> In other words if VERSION_1 is negotiated then bit 27
> isn't ANY_LAYOUT, it's in fact reserved.


But what if guest isn't using Version 1? Legacy distro's certainly
won't have it enabled.
Michael S. Tsirkin Sept. 25, 2016, 1:53 a.m. UTC | #8 
On Sat, Sep 24, 2016 at 06:50:07PM -0700, Stephen Hemminger wrote:
> On Sun, 25 Sep 2016 04:02:28 +0300
> "Michael S. Tsirkin" <mst@redhat.com> wrote:
> 
> > I see. It's implied by VERSION_1 in fact.
> > In other words if VERSION_1 is negotiated then bit 27
> > isn't ANY_LAYOUT, it's in fact reserved.
> 
> 
> But what if guest isn't using Version 1? Legacy distro's certainly
> won't have it enabled.

We probably can just expose ANY_LAYOUT to guest if backend
has VERSION_1.
diff mbox

Patch

diff --git a/lib/librte_vhost/vhost.c b/lib/librte_vhost/vhost.c
index 46095c3..30bb0ce 100644
--- a/lib/librte_vhost/vhost.c
+++ b/lib/librte_vhost/vhost.c
@@ -65,7 +65,8 @@ 
 				(1ULL << VIRTIO_NET_F_CSUM)    | \
 				(1ULL << VIRTIO_NET_F_GUEST_CSUM) | \
 				(1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
-				(1ULL << VIRTIO_NET_F_GUEST_TSO6))
+				(1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
+				(1ULL << VIRTIO_RING_F_INDIRECT_DESC))
 
 uint64_t VHOST_FEATURES = VHOST_SUPPORTED_FEATURES;
 
diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
index 8a151af..0f7dd81 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -679,8 +679,8 @@  make_rarp_packet(struct rte_mbuf *rarp_mbuf, const struct ether_addr *mac)
 }
 
 static inline int __attribute__((always_inline))
-copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
-		  struct rte_mbuf *m, uint16_t desc_idx,
+copy_desc_to_mbuf(struct virtio_net *dev, struct vring_desc *descs,
+		  uint16_t max_desc, struct rte_mbuf *m, uint16_t desc_idx,
 		  struct rte_mempool *mbuf_pool)
 {
 	struct vring_desc *desc;
@@ -693,7 +693,7 @@  copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 	/* A counter to avoid desc dead loop chain */
 	uint32_t nr_desc = 1;
 
-	desc = &vq->desc[desc_idx];
+	desc = &descs[desc_idx];
 	if (unlikely(desc->len < dev->vhost_hlen))
 		return -1;
 
@@ -711,7 +711,7 @@  copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 	 */
 	if (likely((desc->len == dev->vhost_hlen) &&
 		   (desc->flags & VRING_DESC_F_NEXT) != 0)) {
-		desc = &vq->desc[desc->next];
+		desc = &descs[desc->next];
 
 		desc_addr = gpa_to_vva(dev, desc->addr);
 		if (unlikely(!desc_addr))
@@ -747,10 +747,10 @@  copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 			if ((desc->flags & VRING_DESC_F_NEXT) == 0)
 				break;
 
-			if (unlikely(desc->next >= vq->size ||
-				     ++nr_desc > vq->size))
+			if (unlikely(desc->next >= max_desc ||
+				     ++nr_desc > max_desc))
 				return -1;
-			desc = &vq->desc[desc->next];
+			desc = &descs[desc->next];
 
 			desc_addr = gpa_to_vva(dev, desc->addr);
 			if (unlikely(!desc_addr))
@@ -878,19 +878,39 @@  rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 	/* Prefetch descriptor index. */
 	rte_prefetch0(&vq->desc[desc_indexes[0]]);
 	for (i = 0; i < count; i++) {
+		struct vring_desc *desc;
+		uint16_t sz, idx;
 		int err;
 
 		if (likely(i + 1 < count))
 			rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
 
+		if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
+			if (unlikely(!(dev->features &
+					(1ULL << VIRTIO_RING_F_INDIRECT_DESC)))) {
+				RTE_LOG(ERR, VHOST_DATA,
+						"Indirect desc but feature not negotiated.\n");
+				break;
+			}
+
+			desc = (struct vring_desc *)gpa_to_vva(dev,
+					vq->desc[desc_indexes[i]].addr);
+			rte_prefetch0(desc);
+			sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
+			idx = 0;
+		} else {
+			desc = vq->desc;
+			sz = vq->size;
+			idx = desc_indexes[i];
+		}
+
 		pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
 		if (unlikely(pkts[i] == NULL)) {
 			RTE_LOG(ERR, VHOST_DATA,
 				"Failed to allocate memory for mbuf.\n");
 			break;
 		}
-		err = copy_desc_to_mbuf(dev, vq, pkts[i], desc_indexes[i],
-					mbuf_pool);
+		err = copy_desc_to_mbuf(dev, desc, sz, pkts[i], idx, mbuf_pool);
 		if (unlikely(err)) {
 			rte_pktmbuf_free(pkts[i]);
 			break;