Virtualization - May 2019 - [PATCH v2 0/8] vsock/virtio: optimizations to increase the throughput

While I was testing this new series (v2) I discovered an huge use of memory
and a memory leak in the virtio-vsock driver in the guest when I sent
1-byte packets to the guest.

These issues are present since the introduction of the virtio-vsock
driver. I added the patches 1 and 2 to fix them in this series in order
to better track the performance trends.

v1: https://patchwork.kernel.org/cover/10885431/

v2:
- Add patch 1 to limit the memory usage
- Add patch 2 to avoid memory leak during the socket release
- Add patch 3 to fix locking of fwd_cnt and buf_alloc
- Patch 4: fix 'free_space' type (u32 instead of s64) [Stefan]
- Patch 5: Avoid integer underflow of iov_len [Stefan]
- Patch 5: Fix packet capture in order to see the exact packets that are
           delivered. [Stefan]
- Add patch 8 to make the RX buffer size tunable [Stefan]

Below are the benchmarks step by step. I used iperf3 [1] modified with VSOCK
support.
As Micheal suggested in the v1, I booted host and guest with 'nosmap',
and I
added a column with virtio-net+vhost-net performance.

A brief description of patches:
- Patches 1+2: limit the memory usage with an extra copy and avoid memory leak
- Patches 3+4: fix locking and reduce the number of credit update messages sent
               to the transmitter
- Patches 5+6: allow the host to split packets on multiple buffers and use
               VIRTIO_VSOCK_MAX_PKT_BUF_SIZE as the max packet size allowed
- Patches 7+8: increase RX buffer size to 64 KiB

                    host -> guest [Gbps]
pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net + vhost
                                                                     TCP_NODELAY
64         0.068     0.063    0.130    0.131    0.128         0.188     0.187
256        0.274     0.236    0.392    0.338    0.282         0.749     0.654
512        0.531     0.457    0.862    0.725    0.602         1.419     1.414
1K         0.954     0.827    1.591    1.598    1.548         2.599     2.640
2K         1.783     1.543    3.731    3.637    3.469         4.530     4.754
4K         3.332     3.436    7.164    7.124    6.494         7.738     7.696
8K         5.792     5.530   11.653   11.787   11.444        12.307    11.850
16K        8.405     8.462   16.372   16.855   17.562        16.936    16.954
32K       14.208    13.669   18.945   20.009   23.128        21.980    23.015
64K       21.082    18.893   20.266   20.903   30.622        27.290    27.383
128K      20.696    20.148   20.112   21.746   32.152        30.446    30.990
256K      20.801    20.589   20.725   22.685   34.721        33.151    32.745
512K      21.220    20.465   20.432   22.106   34.496        36.847    31.096

                    guest -> host [Gbps]
pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net + vhost
                                                                     TCP_NODELAY
64         0.089     0.091    0.120    0.115    0.117         0.274     0.272
256        0.352     0.354    0.452    0.445    0.451         1.085     1.136
512        0.705     0.704    0.893    0.858    0.898         2.131     1.882
1K         1.394     1.433    1.721    1.669    1.691         3.984     3.576
2K         2.818     2.874    3.316    3.249    3.303         6.719     6.359
4K         5.293     5.397    6.129    5.933    6.082        10.105     9.860
8K         8.890     9.151   10.990   10.545   10.519        15.239    14.868
16K       11.444    11.018   12.074   15.255   15.577        20.551    20.848
32K       11.229    10.875   10.857   24.401   25.227        26.294    26.380
64K       10.832    10.545   10.816   39.487   39.616        34.996    32.041
128K      10.435    10.241   10.500   39.813   40.012        38.379    35.055
256K      10.263     9.866    9.845   34.971   35.143        36.559    37.232
512K      10.224    10.060   10.092   35.469   34.627        34.963    33.401

As Stefan suggested in the v1, this time I measured also the efficiency in this
way:
    efficiency = Mbps / (%CPU_Host + %CPU_Guest)

The '%CPU_Guest' is taken inside the VM. I know that it is not the best
way,
but it's provided for free from iperf3 and could be an indication.

        host -> guest efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net + vhost
                                                                     TCP_NODELAY
64          0.94      0.59     3.96     4.06     4.09          2.82      2.11
256         2.62      2.50     6.45     6.09     5.81          9.64      8.73
512         5.16      4.87    13.16    12.39    11.67         17.83     17.76
1K          9.16      8.85    24.98    24.97    25.01         32.57     32.04
2K         17.41     17.03    49.09    48.59    49.22         55.31     57.14
4K         32.99     33.62    90.80    90.98    91.72         91.79     91.40
8K         58.51     59.98   153.53   170.83   167.31        137.51    132.85
16K        89.32     95.29   216.98   264.18   260.95        176.05    176.05
32K       152.94    167.10   285.75   387.02   360.81        215.49    226.30
64K       250.38    307.20   317.65   489.53   472.70        238.97    244.27
128K      327.99    335.24   335.76   523.71   486.41        253.29    260.86
256K      327.06    334.24   338.64   533.76   509.85        267.78    266.22
512K      337.36    330.61   334.95   512.90   496.35        280.42    241.43

        guest -> host efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net + vhost
                                                                     TCP_NODELAY
64          0.90      0.91     1.37     1.32     1.35          2.15      2.13
256         3.59      3.55     5.23     5.19     5.29          8.50      8.89
512         7.19      7.08    10.21     9.95    10.38         16.74     14.71
1K         14.15     14.34    19.85    19.06    19.33         31.44     28.11
2K         28.44     29.09    37.78    37.18    37.49         53.07     50.63
4K         55.37     57.60    71.02    69.27    70.97         81.56     79.32
8K        105.58    100.45   111.95   124.68   123.61        120.85    118.66
16K       141.63    138.24   137.67   187.41   190.20        160.43    163.00
32K       147.56    143.09   138.48   296.41   301.04        214.64    223.94
64K       144.81    143.27   138.49   433.98   462.26        298.86    269.71
128K      150.14    147.99   146.85   511.36   514.29        350.17    298.09
256K      156.69    152.25   148.69   542.19   549.97        326.42    333.32
512K      157.29    153.35   152.22   546.52   533.24        315.55    302.27

[1] https://github.com/stefano-garzarella/iperf/

Stefano Garzarella (8):
  vsock/virtio: limit the memory used per-socket
  vsock/virtio: free packets during the socket release
  vsock/virtio: fix locking for fwd_cnt and buf_alloc
  vsock/virtio: reduce credit update messages
  vhost/vsock: split packets to send using multiple buffers
  vsock/virtio: change the maximum packet size allowed
  vsock/virtio: increase RX buffer size to 64 KiB
  vsock/virtio: make the RX buffer size tunable

 drivers/vhost/vsock.c                   |  53 +++++++--
 include/linux/virtio_vsock.h            |  14 ++-
 net/vmw_vsock/virtio_transport.c        |  28 ++++-
 net/vmw_vsock/virtio_transport_common.c | 144 ++++++++++++++++++------
 4 files changed, 190 insertions(+), 49 deletions(-)

-- 
2.20.1

Since virtio-vsock was introduced, the buffers filled by the host
and pushed to the guest using the vring, are directly queued in
a per-socket list avoiding to copy it.
These buffers are preallocated by the guest with a fixed
size (4 KB).

The maximum amount of memory used by each socket should be
controlled by the credit mechanism.
The default credit available per-socket is 256 KB, but if we use
only 1 byte per packet, the guest can queue up to 262144 of 4 KB
buffers, using up to 1 GB of memory per-socket. In addition, the
guest will continue to fill the vring with new 4 KB free buffers
to avoid starvation of other sockets.

This patch solves this issue copying the payload in a new buffer.
Then it is queued in the per-socket list, and the 4KB buffer used
by the host is freed.

In this way, the memory used by each socket respects the credit
available, and we still avoid starvation, paying the cost of an
extra memory copy. When the buffer is completely full we do a
"zero-copy", moving the buffer directly in the per-socket list.

Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 drivers/vhost/vsock.c                   |  2 +
 include/linux/virtio_vsock.h            |  8 +++
 net/vmw_vsock/virtio_transport.c        |  1 +
 net/vmw_vsock/virtio_transport_common.c | 95 ++++++++++++++++++-------
 4 files changed, 81 insertions(+), 25 deletions(-)

diff --git a/drivers/vhost/vsock.c b/drivers/vhost/vsock.c
index bb5fc0e9fbc2..7964e2daee09 100644
--- a/drivers/vhost/vsock.c
+++ b/drivers/vhost/vsock.c
@@ -320,6 +320,8 @@ vhost_vsock_alloc_pkt(struct vhost_virtqueue *vq,
 		return NULL;
 	}
 
+	pkt->buf_len = pkt->len;
+
 	nbytes = copy_from_iter(pkt->buf, pkt->len, &iov_iter);
 	if (nbytes != pkt->len) {
 		vq_err(vq, "Expected %u byte payload, got %zu bytes\n",
diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
index e223e2632edd..345f04ee9193 100644
--- a/include/linux/virtio_vsock.h
+++ b/include/linux/virtio_vsock.h
@@ -54,9 +54,17 @@ struct virtio_vsock_pkt {
 	void *buf;
 	u32 len;
 	u32 off;
+	u32 buf_len;
 	bool reply;
 };
 
+struct virtio_vsock_buf {
+	struct list_head list;
+	void *addr;
+	u32 len;
+	u32 off;
+};
+
 struct virtio_vsock_pkt_info {
 	u32 remote_cid, remote_port;
 	struct vsock_sock *vsk;
diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
index 15eb5d3d4750..af1d2ce12f54 100644
--- a/net/vmw_vsock/virtio_transport.c
+++ b/net/vmw_vsock/virtio_transport.c
@@ -280,6 +280,7 @@ static void virtio_vsock_rx_fill(struct virtio_vsock *vsock)
 			break;
 		}
 
+		pkt->buf_len = buf_len;
 		pkt->len = buf_len;
 
 		sg_init_one(&hdr, &pkt->hdr, sizeof(pkt->hdr));
diff --git a/net/vmw_vsock/virtio_transport_common.c
b/net/vmw_vsock/virtio_transport_common.c
index 602715fc9a75..0248d6808755 100644
--- a/net/vmw_vsock/virtio_transport_common.c
+++ b/net/vmw_vsock/virtio_transport_common.c
@@ -65,6 +65,9 @@ virtio_transport_alloc_pkt(struct virtio_vsock_pkt_info *info,
 		pkt->buf = kmalloc(len, GFP_KERNEL);
 		if (!pkt->buf)
 			goto out_pkt;
+
+		pkt->buf_len = len;
+
 		err = memcpy_from_msg(pkt->buf, info->msg, len);
 		if (err)
 			goto out;
@@ -86,6 +89,46 @@ virtio_transport_alloc_pkt(struct virtio_vsock_pkt_info
*info,
 	return NULL;
 }
 
+static struct virtio_vsock_buf *
+virtio_transport_alloc_buf(struct virtio_vsock_pkt *pkt, bool zero_copy)
+{
+	struct virtio_vsock_buf *buf;
+
+	if (pkt->len == 0)
+		return NULL;
+
+	buf = kzalloc(sizeof(*buf), GFP_KERNEL);
+	if (!buf)
+		return NULL;
+
+	/* If the buffer in the virtio_vsock_pkt is full, we can move it to
+	 * the new virtio_vsock_buf avoiding the copy, because we are sure that
+	 * we are not use more memory than that counted by the credit mechanism.
+	 */
+	if (zero_copy && pkt->len == pkt->buf_len) {
+		buf->addr = pkt->buf;
+		pkt->buf = NULL;
+	} else {
+		buf->addr = kmalloc(pkt->len, GFP_KERNEL);
+		if (!buf->addr) {
+			kfree(buf);
+			return NULL;
+		}
+
+		memcpy(buf->addr, pkt->buf, pkt->len);
+	}
+
+	buf->len = pkt->len;
+
+	return buf;
+}
+
+static void virtio_transport_free_buf(struct virtio_vsock_buf *buf)
+{
+	kfree(buf->addr);
+	kfree(buf);
+}
+
 /* Packet capture */
 static struct sk_buff *virtio_transport_build_skb(void *opaque)
 {
@@ -190,17 +233,15 @@ static int virtio_transport_send_pkt_info(struct
vsock_sock *vsk,
 	return virtio_transport_get_ops()->send_pkt(pkt);
 }
 
-static void virtio_transport_inc_rx_pkt(struct virtio_vsock_sock *vvs,
-					struct virtio_vsock_pkt *pkt)
+static void virtio_transport_inc_rx_pkt(struct virtio_vsock_sock *vvs, u32 len)
 {
-	vvs->rx_bytes += pkt->len;
+	vvs->rx_bytes += len;
 }
 
-static void virtio_transport_dec_rx_pkt(struct virtio_vsock_sock *vvs,
-					struct virtio_vsock_pkt *pkt)
+static void virtio_transport_dec_rx_pkt(struct virtio_vsock_sock *vvs, u32 len)
 {
-	vvs->rx_bytes -= pkt->len;
-	vvs->fwd_cnt += pkt->len;
+	vvs->rx_bytes -= len;
+	vvs->fwd_cnt += len;
 }
 
 void virtio_transport_inc_tx_pkt(struct virtio_vsock_sock *vvs, struct
virtio_vsock_pkt *pkt)
@@ -254,36 +295,36 @@ virtio_transport_stream_do_dequeue(struct vsock_sock *vsk,
 				   size_t len)
 {
 	struct virtio_vsock_sock *vvs = vsk->trans;
-	struct virtio_vsock_pkt *pkt;
+	struct virtio_vsock_buf *buf;
 	size_t bytes, total = 0;
 	int err = -EFAULT;
 
 	spin_lock_bh(&vvs->rx_lock);
 	while (total < len && !list_empty(&vvs->rx_queue)) {
-		pkt = list_first_entry(&vvs->rx_queue,
-				       struct virtio_vsock_pkt, list);
+		buf = list_first_entry(&vvs->rx_queue,
+				       struct virtio_vsock_buf, list);
 
 		bytes = len - total;
-		if (bytes > pkt->len - pkt->off)
-			bytes = pkt->len - pkt->off;
+		if (bytes > buf->len - buf->off)
+			bytes = buf->len - buf->off;
 
 		/* sk_lock is held by caller so no one else can dequeue.
 		 * Unlock rx_lock since memcpy_to_msg() may sleep.
 		 */
 		spin_unlock_bh(&vvs->rx_lock);
 
-		err = memcpy_to_msg(msg, pkt->buf + pkt->off, bytes);
+		err = memcpy_to_msg(msg, buf->addr + buf->off, bytes);
 		if (err)
 			goto out;
 
 		spin_lock_bh(&vvs->rx_lock);
 
 		total += bytes;
-		pkt->off += bytes;
-		if (pkt->off == pkt->len) {
-			virtio_transport_dec_rx_pkt(vvs, pkt);
-			list_del(&pkt->list);
-			virtio_transport_free_pkt(pkt);
+		buf->off += bytes;
+		if (buf->off == buf->len) {
+			virtio_transport_dec_rx_pkt(vvs, buf->len);
+			list_del(&buf->list);
+			virtio_transport_free_buf(buf);
 		}
 	}
 	spin_unlock_bh(&vvs->rx_lock);
@@ -841,20 +882,24 @@ virtio_transport_recv_connected(struct sock *sk,
 {
 	struct vsock_sock *vsk = vsock_sk(sk);
 	struct virtio_vsock_sock *vvs = vsk->trans;
+	struct virtio_vsock_buf *buf;
 	int err = 0;
 
 	switch (le16_to_cpu(pkt->hdr.op)) {
 	case VIRTIO_VSOCK_OP_RW:
 		pkt->len = le32_to_cpu(pkt->hdr.len);
-		pkt->off = 0;
+		buf = virtio_transport_alloc_buf(pkt, true);
 
-		spin_lock_bh(&vvs->rx_lock);
-		virtio_transport_inc_rx_pkt(vvs, pkt);
-		list_add_tail(&pkt->list, &vvs->rx_queue);
-		spin_unlock_bh(&vvs->rx_lock);
+		if (buf) {
+			spin_lock_bh(&vvs->rx_lock);
+			virtio_transport_inc_rx_pkt(vvs, pkt->len);
+			list_add_tail(&buf->list, &vvs->rx_queue);
+			spin_unlock_bh(&vvs->rx_lock);
 
-		sk->sk_data_ready(sk);
-		return err;
+			sk->sk_data_ready(sk);
+		}
+
+		break;
 	case VIRTIO_VSOCK_OP_CREDIT_UPDATE:
 		sk->sk_write_space(sk);
 		break;
-- 
2.20.1

When the socket is released, we should free all packets
queued in the per-socket list in order to avoid a memory
leak.

Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 net/vmw_vsock/virtio_transport_common.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/net/vmw_vsock/virtio_transport_common.c
b/net/vmw_vsock/virtio_transport_common.c
index 0248d6808755..65c8b4a23f2b 100644
--- a/net/vmw_vsock/virtio_transport_common.c
+++ b/net/vmw_vsock/virtio_transport_common.c
@@ -827,12 +827,20 @@ static bool virtio_transport_close(struct vsock_sock *vsk)
 
 void virtio_transport_release(struct vsock_sock *vsk)
 {
+	struct virtio_vsock_sock *vvs = vsk->trans;
+	struct virtio_vsock_buf *buf;
 	struct sock *sk = &vsk->sk;
 	bool remove_sock = true;
 
 	lock_sock(sk);
 	if (sk->sk_type == SOCK_STREAM)
 		remove_sock = virtio_transport_close(vsk);
+	while (!list_empty(&vvs->rx_queue)) {
+		buf = list_first_entry(&vvs->rx_queue,
+				       struct virtio_vsock_buf, list);
+		list_del(&buf->list);
+		virtio_transport_free_buf(buf);
+	}
 	release_sock(sk);
 
 	if (remove_sock)
-- 
2.20.1

fwd_cnt is written with rx_lock, so we should read it using
the same spinlock also if we are in the TX path.

Move also buf_alloc under rx_lock and add a missing locking
when we modify it.

Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 include/linux/virtio_vsock.h            | 2 +-
 net/vmw_vsock/virtio_transport_common.c | 6 ++++--
 2 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
index 345f04ee9193..fb5954fc85c8 100644
--- a/include/linux/virtio_vsock.h
+++ b/include/linux/virtio_vsock.h
@@ -35,11 +35,11 @@ struct virtio_vsock_sock {
 
 	/* Protected by tx_lock */
 	u32 tx_cnt;
-	u32 buf_alloc;
 	u32 peer_fwd_cnt;
 	u32 peer_buf_alloc;
 
 	/* Protected by rx_lock */
+	u32 buf_alloc;
 	u32 fwd_cnt;
 	u32 rx_bytes;
 	struct list_head rx_queue;
diff --git a/net/vmw_vsock/virtio_transport_common.c
b/net/vmw_vsock/virtio_transport_common.c
index 65c8b4a23f2b..f2e4e128bc86 100644
--- a/net/vmw_vsock/virtio_transport_common.c
+++ b/net/vmw_vsock/virtio_transport_common.c
@@ -246,10 +246,10 @@ static void virtio_transport_dec_rx_pkt(struct
virtio_vsock_sock *vvs, u32 len)
 
 void virtio_transport_inc_tx_pkt(struct virtio_vsock_sock *vvs, struct
virtio_vsock_pkt *pkt)
 {
-	spin_lock_bh(&vvs->tx_lock);
+	spin_lock_bh(&vvs->rx_lock);
 	pkt->hdr.fwd_cnt = cpu_to_le32(vvs->fwd_cnt);
 	pkt->hdr.buf_alloc = cpu_to_le32(vvs->buf_alloc);
-	spin_unlock_bh(&vvs->tx_lock);
+	spin_unlock_bh(&vvs->rx_lock);
 }
 EXPORT_SYMBOL_GPL(virtio_transport_inc_tx_pkt);
 
@@ -469,7 +469,9 @@ void virtio_transport_set_buffer_size(struct vsock_sock
*vsk, u64 val)
 	if (val > vvs->buf_size_max)
 		vvs->buf_size_max = val;
 	vvs->buf_size = val;
+	spin_lock_bh(&vvs->rx_lock);
 	vvs->buf_alloc = val;
+	spin_unlock_bh(&vvs->rx_lock);
 }
 EXPORT_SYMBOL_GPL(virtio_transport_set_buffer_size);
 
-- 
2.20.1

In order to reduce the number of credit update messages,
we send them only when the space available seen by the
transmitter is less than VIRTIO_VSOCK_MAX_PKT_BUF_SIZE.

Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 include/linux/virtio_vsock.h            |  1 +
 net/vmw_vsock/virtio_transport_common.c | 16 +++++++++++++---
 2 files changed, 14 insertions(+), 3 deletions(-)

diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
index fb5954fc85c8..84b72026d327 100644
--- a/include/linux/virtio_vsock.h
+++ b/include/linux/virtio_vsock.h
@@ -41,6 +41,7 @@ struct virtio_vsock_sock {
 	/* Protected by rx_lock */
 	u32 buf_alloc;
 	u32 fwd_cnt;
+	u32 last_fwd_cnt;
 	u32 rx_bytes;
 	struct list_head rx_queue;
 };
diff --git a/net/vmw_vsock/virtio_transport_common.c
b/net/vmw_vsock/virtio_transport_common.c
index f2e4e128bc86..b61fd5e29a1f 100644
--- a/net/vmw_vsock/virtio_transport_common.c
+++ b/net/vmw_vsock/virtio_transport_common.c
@@ -247,6 +247,7 @@ static void virtio_transport_dec_rx_pkt(struct
virtio_vsock_sock *vvs, u32 len)
 void virtio_transport_inc_tx_pkt(struct virtio_vsock_sock *vvs, struct
virtio_vsock_pkt *pkt)
 {
 	spin_lock_bh(&vvs->rx_lock);
+	vvs->last_fwd_cnt = vvs->fwd_cnt;
 	pkt->hdr.fwd_cnt = cpu_to_le32(vvs->fwd_cnt);
 	pkt->hdr.buf_alloc = cpu_to_le32(vvs->buf_alloc);
 	spin_unlock_bh(&vvs->rx_lock);
@@ -297,6 +298,7 @@ virtio_transport_stream_do_dequeue(struct vsock_sock *vsk,
 	struct virtio_vsock_sock *vvs = vsk->trans;
 	struct virtio_vsock_buf *buf;
 	size_t bytes, total = 0;
+	u32 free_space;
 	int err = -EFAULT;
 
 	spin_lock_bh(&vvs->rx_lock);
@@ -327,11 +329,19 @@ virtio_transport_stream_do_dequeue(struct vsock_sock *vsk,
 			virtio_transport_free_buf(buf);
 		}
 	}
+
+	free_space = vvs->buf_alloc - (vvs->fwd_cnt - vvs->last_fwd_cnt);
+
 	spin_unlock_bh(&vvs->rx_lock);
 
-	/* Send a credit pkt to peer */
-	virtio_transport_send_credit_update(vsk, VIRTIO_VSOCK_TYPE_STREAM,
-					    NULL);
+	/* We send a credit update only when the space available seen
+	 * by the transmitter is less than VIRTIO_VSOCK_MAX_PKT_BUF_SIZE
+	 */
+	if (free_space < VIRTIO_VSOCK_MAX_PKT_BUF_SIZE) {
+		virtio_transport_send_credit_update(vsk,
+						    VIRTIO_VSOCK_TYPE_STREAM,
+						    NULL);
+	}
 
 	return total;
 
-- 
2.20.1

If the packets to sent to the guest are bigger than the buffer
available, we can split them, using multiple buffers and fixing
the length in the packet header.
This is safe since virtio-vsock supports only stream sockets.

Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 drivers/vhost/vsock.c                   | 51 +++++++++++++++++++------
 net/vmw_vsock/virtio_transport_common.c | 15 ++++++--
 2 files changed, 52 insertions(+), 14 deletions(-)

diff --git a/drivers/vhost/vsock.c b/drivers/vhost/vsock.c
index 7964e2daee09..fb731d09f5f1 100644
--- a/drivers/vhost/vsock.c
+++ b/drivers/vhost/vsock.c
@@ -94,7 +94,7 @@ vhost_transport_do_send_pkt(struct vhost_vsock *vsock,
 		struct iov_iter iov_iter;
 		unsigned out, in;
 		size_t nbytes;
-		size_t len;
+		size_t iov_len, payload_len;
 		int head;
 
 		spin_lock_bh(&vsock->send_pkt_list_lock);
@@ -139,8 +139,24 @@ vhost_transport_do_send_pkt(struct vhost_vsock *vsock,
 			break;
 		}
 
-		len = iov_length(&vq->iov[out], in);
-		iov_iter_init(&iov_iter, READ, &vq->iov[out], in, len);
+		iov_len = iov_length(&vq->iov[out], in);
+		if (iov_len < sizeof(pkt->hdr)) {
+			virtio_transport_free_pkt(pkt);
+			vq_err(vq, "Buffer len [%zu] too small\n", iov_len);
+			break;
+		}
+
+		iov_iter_init(&iov_iter, READ, &vq->iov[out], in, iov_len);
+		payload_len = pkt->len - pkt->off;
+
+		/* If the packet is greater than the space available in the
+		 * buffer, we split it using multiple buffers.
+		 */
+		if (payload_len > iov_len - sizeof(pkt->hdr))
+			payload_len = iov_len - sizeof(pkt->hdr);
+
+		/* Set the correct length in the header */
+		pkt->hdr.len = cpu_to_le32(payload_len);
 
 		nbytes = copy_to_iter(&pkt->hdr, sizeof(pkt->hdr), &iov_iter);
 		if (nbytes != sizeof(pkt->hdr)) {
@@ -149,16 +165,34 @@ vhost_transport_do_send_pkt(struct vhost_vsock *vsock,
 			break;
 		}
 
-		nbytes = copy_to_iter(pkt->buf, pkt->len, &iov_iter);
-		if (nbytes != pkt->len) {
+		nbytes = copy_to_iter(pkt->buf + pkt->off, payload_len,
+				      &iov_iter);
+		if (nbytes != payload_len) {
 			virtio_transport_free_pkt(pkt);
 			vq_err(vq, "Faulted on copying pkt buf\n");
 			break;
 		}
 
-		vhost_add_used(vq, head, sizeof(pkt->hdr) + pkt->len);
+		vhost_add_used(vq, head, sizeof(pkt->hdr) + payload_len);
 		added = true;
 
+		/* Deliver to monitoring devices all correctly transmitted
+		 * packets.
+		 */
+		virtio_transport_deliver_tap_pkt(pkt);
+
+		pkt->off += payload_len;
+
+		/* If we didn't send all the payload we can requeue the packet
+		 * to send it with the next available buffer.
+		 */
+		if (pkt->off < pkt->len) {
+			spin_lock_bh(&vsock->send_pkt_list_lock);
+			list_add(&pkt->list, &vsock->send_pkt_list);
+			spin_unlock_bh(&vsock->send_pkt_list_lock);
+			continue;
+		}
+
 		if (pkt->reply) {
 			int val;
 
@@ -169,11 +203,6 @@ vhost_transport_do_send_pkt(struct vhost_vsock *vsock,
 				restart_tx = true;
 		}
 
-		/* Deliver to monitoring devices all correctly transmitted
-		 * packets.
-		 */
-		virtio_transport_deliver_tap_pkt(pkt);
-
 		virtio_transport_free_pkt(pkt);
 	}
 	if (added)
diff --git a/net/vmw_vsock/virtio_transport_common.c
b/net/vmw_vsock/virtio_transport_common.c
index b61fd5e29a1f..3f313bcd6a26 100644
--- a/net/vmw_vsock/virtio_transport_common.c
+++ b/net/vmw_vsock/virtio_transport_common.c
@@ -135,8 +135,17 @@ static struct sk_buff *virtio_transport_build_skb(void
*opaque)
 	struct virtio_vsock_pkt *pkt = opaque;
 	struct af_vsockmon_hdr *hdr;
 	struct sk_buff *skb;
+	size_t payload_len;
+	void *payload_buf;
 
-	skb = alloc_skb(sizeof(*hdr) + sizeof(pkt->hdr) + pkt->len,
+	/* A packet could be split to fit the RX buffer, so we can retrieve
+	 * the payload length from the header and the buffer pointer taking
+	 * care of the offset in the original packet.
+	 */
+	payload_len = le32_to_cpu(pkt->hdr.len);
+	payload_buf = pkt->buf + pkt->off;
+
+	skb = alloc_skb(sizeof(*hdr) + sizeof(pkt->hdr) + payload_len,
 			GFP_ATOMIC);
 	if (!skb)
 		return NULL;
@@ -176,8 +185,8 @@ static struct sk_buff *virtio_transport_build_skb(void
*opaque)
 
 	skb_put_data(skb, &pkt->hdr, sizeof(pkt->hdr));
 
-	if (pkt->len) {
-		skb_put_data(skb, pkt->buf, pkt->len);
+	if (payload_len) {
+		skb_put_data(skb, payload_buf, payload_len);
 	}
 
 	return skb;
-- 
2.20.1

Since now we are able to split packets, we can avoid limiting
their sizes to VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE.
Instead, we can use VIRTIO_VSOCK_MAX_PKT_BUF_SIZE as the max
packet size.

Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 net/vmw_vsock/virtio_transport_common.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/vmw_vsock/virtio_transport_common.c
b/net/vmw_vsock/virtio_transport_common.c
index 3f313bcd6a26..63606525755d 100644
--- a/net/vmw_vsock/virtio_transport_common.c
+++ b/net/vmw_vsock/virtio_transport_common.c
@@ -219,8 +219,8 @@ static int virtio_transport_send_pkt_info(struct vsock_sock
*vsk,
 	vvs = vsk->trans;
 
 	/* we can send less than pkt_len bytes */
-	if (pkt_len > VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE)
-		pkt_len = VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE;
+	if (pkt_len > VIRTIO_VSOCK_MAX_PKT_BUF_SIZE)
+		pkt_len = VIRTIO_VSOCK_MAX_PKT_BUF_SIZE;
 
 	/* virtio_transport_get_credit might return less than pkt_len credit */
 	pkt_len = virtio_transport_get_credit(vvs, pkt_len);
-- 
2.20.1

In order to increase host -> guest throughput with large packets,
we can use 64 KiB RX buffers.

Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 include/linux/virtio_vsock.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
index 84b72026d327..5a9d25be72df 100644
--- a/include/linux/virtio_vsock.h
+++ b/include/linux/virtio_vsock.h
@@ -10,7 +10,7 @@
 #define VIRTIO_VSOCK_DEFAULT_MIN_BUF_SIZE	128
 #define VIRTIO_VSOCK_DEFAULT_BUF_SIZE		(1024 * 256)
 #define VIRTIO_VSOCK_DEFAULT_MAX_BUF_SIZE	(1024 * 256)
-#define VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE	(1024 * 4)
+#define VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE	(1024 * 64)
 #define VIRTIO_VSOCK_MAX_BUF_SIZE		0xFFFFFFFFUL
 #define VIRTIO_VSOCK_MAX_PKT_BUF_SIZE		(1024 * 64)
 
-- 
2.20.1

The RX buffer size determines the memory consumption of the
vsock/virtio guest driver, so we make it tunable through
a module parameter.

The size allowed are between 4 KB and 64 KB in order to be
compatible with old host drivers.

Suggested-by: Stefan Hajnoczi <stefanha at redhat.com>
Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
---
 include/linux/virtio_vsock.h     |  1 +
 net/vmw_vsock/virtio_transport.c | 27 ++++++++++++++++++++++++++-
 2 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
index 5a9d25be72df..b9f8c3d91f80 100644
--- a/include/linux/virtio_vsock.h
+++ b/include/linux/virtio_vsock.h
@@ -13,6 +13,7 @@
 #define VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE	(1024 * 64)
 #define VIRTIO_VSOCK_MAX_BUF_SIZE		0xFFFFFFFFUL
 #define VIRTIO_VSOCK_MAX_PKT_BUF_SIZE		(1024 * 64)
+#define VIRTIO_VSOCK_MIN_PKT_BUF_SIZE		(1024 * 4)
 
 enum {
 	VSOCK_VQ_RX     = 0, /* for host to guest data */
diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
index af1d2ce12f54..732398b4e28f 100644
--- a/net/vmw_vsock/virtio_transport.c
+++ b/net/vmw_vsock/virtio_transport.c
@@ -66,6 +66,31 @@ struct virtio_vsock {
 	u32 guest_cid;
 };
 
+static unsigned int rx_buf_size = VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE;
+
+static int param_set_rx_buf_size(const char *val, const struct kernel_param
*kp)
+{
+	unsigned int size;
+	int ret;
+
+	ret = kstrtouint(val, 0, &size);
+	if (ret)
+		return ret;
+
+	if (size < VIRTIO_VSOCK_MIN_PKT_BUF_SIZE ||
+	    size > VIRTIO_VSOCK_MAX_PKT_BUF_SIZE)
+		return -EINVAL;
+
+	return param_set_uint(val, kp);
+};
+
+static const struct kernel_param_ops param_ops_rx_buf_size = {
+	.set = param_set_rx_buf_size,
+	.get = param_get_uint,
+};
+
+module_param_cb(rx_buf_size, &param_ops_rx_buf_size, &rx_buf_size,
0644);
+
 static struct virtio_vsock *virtio_vsock_get(void)
 {
 	return the_virtio_vsock;
@@ -261,7 +286,7 @@ virtio_transport_cancel_pkt(struct vsock_sock *vsk)
 
 static void virtio_vsock_rx_fill(struct virtio_vsock *vsock)
 {
-	int buf_len = VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE;
+	int buf_len = rx_buf_size;
 	struct virtio_vsock_pkt *pkt;
 	struct scatterlist hdr, buf, *sgs[2];
 	struct virtqueue *vq;
-- 
2.20.1

From: Stefano Garzarella <sgarzare at redhat.com>
Date: Fri, 10 May 2019 14:58:37 +0200
> @@ -827,12 +827,20 @@ static bool virtio_transport_close(struct vsock_sock
*vsk)
>  
>  void virtio_transport_release(struct vsock_sock *vsk)
>  {
> +	struct virtio_vsock_sock *vvs = vsk->trans;
> +	struct virtio_vsock_buf *buf;
>  	struct sock *sk = &vsk->sk;
>  	bool remove_sock = true;
>  
>  	lock_sock(sk);
>  	if (sk->sk_type == SOCK_STREAM)
>  		remove_sock = virtio_transport_close(vsk);
> +	while (!list_empty(&vvs->rx_queue)) {
> +		buf = list_first_entry(&vvs->rx_queue,
> +				       struct virtio_vsock_buf, list);
Please use list_for_each_entry_safe().

On Fri, May 10, 2019 at 02:58:36PM +0200, Stefano Garzarella
wrote:
> Since virtio-vsock was introduced, the buffers filled by the host
> and pushed to the guest using the vring, are directly queued in
> a per-socket list avoiding to copy it.
> These buffers are preallocated by the guest with a fixed
> size (4 KB).
> 
> The maximum amount of memory used by each socket should be
> controlled by the credit mechanism.
> The default credit available per-socket is 256 KB, but if we use
> only 1 byte per packet, the guest can queue up to 262144 of 4 KB
> buffers, using up to 1 GB of memory per-socket. In addition, the
> guest will continue to fill the vring with new 4 KB free buffers
> to avoid starvation of other sockets.
> 
> This patch solves this issue copying the payload in a new buffer.
> Then it is queued in the per-socket list, and the 4KB buffer used
> by the host is freed.
> 
> In this way, the memory used by each socket respects the credit
> available, and we still avoid starvation, paying the cost of an
> extra memory copy. When the buffer is completely full we do a
> "zero-copy", moving the buffer directly in the per-socket list.
> 
> Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
> ---
>  drivers/vhost/vsock.c                   |  2 +
>  include/linux/virtio_vsock.h            |  8 +++
>  net/vmw_vsock/virtio_transport.c        |  1 +
>  net/vmw_vsock/virtio_transport_common.c | 95 ++++++++++++++++++-------
>  4 files changed, 81 insertions(+), 25 deletions(-)
> 
> diff --git a/drivers/vhost/vsock.c b/drivers/vhost/vsock.c
> index bb5fc0e9fbc2..7964e2daee09 100644
> --- a/drivers/vhost/vsock.c
> +++ b/drivers/vhost/vsock.c
> @@ -320,6 +320,8 @@ vhost_vsock_alloc_pkt(struct vhost_virtqueue *vq,
>  		return NULL;
>  	}
>  
> +	pkt->buf_len = pkt->len;
> +
>  	nbytes = copy_from_iter(pkt->buf, pkt->len, &iov_iter);
>  	if (nbytes != pkt->len) {
>  		vq_err(vq, "Expected %u byte payload, got %zu bytes\n",
> diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
> index e223e2632edd..345f04ee9193 100644
> --- a/include/linux/virtio_vsock.h
> +++ b/include/linux/virtio_vsock.h
> @@ -54,9 +54,17 @@ struct virtio_vsock_pkt {
>  	void *buf;
>  	u32 len;
>  	u32 off;
> +	u32 buf_len;
>  	bool reply;
>  };
>  
> +struct virtio_vsock_buf {
> +	struct list_head list;
> +	void *addr;
> +	u32 len;
> +	u32 off;
> +};
> +
>  struct virtio_vsock_pkt_info {
>  	u32 remote_cid, remote_port;
>  	struct vsock_sock *vsk;
> diff --git a/net/vmw_vsock/virtio_transport.c
b/net/vmw_vsock/virtio_transport.c
> index 15eb5d3d4750..af1d2ce12f54 100644
> --- a/net/vmw_vsock/virtio_transport.c
> +++ b/net/vmw_vsock/virtio_transport.c
> @@ -280,6 +280,7 @@ static void virtio_vsock_rx_fill(struct virtio_vsock
*vsock)
>  			break;
>  		}
>  
> +		pkt->buf_len = buf_len;
>  		pkt->len = buf_len;
>  
>  		sg_init_one(&hdr, &pkt->hdr, sizeof(pkt->hdr));
> diff --git a/net/vmw_vsock/virtio_transport_common.c
b/net/vmw_vsock/virtio_transport_common.c
> index 602715fc9a75..0248d6808755 100644
> --- a/net/vmw_vsock/virtio_transport_common.c
> +++ b/net/vmw_vsock/virtio_transport_common.c
> @@ -65,6 +65,9 @@ virtio_transport_alloc_pkt(struct virtio_vsock_pkt_info
*info,
>  		pkt->buf = kmalloc(len, GFP_KERNEL);
>  		if (!pkt->buf)
>  			goto out_pkt;
> +
> +		pkt->buf_len = len;
> +
>  		err = memcpy_from_msg(pkt->buf, info->msg, len);
>  		if (err)
>  			goto out;
> @@ -86,6 +89,46 @@ virtio_transport_alloc_pkt(struct virtio_vsock_pkt_info
*info,
>  	return NULL;
>  }
>  
> +static struct virtio_vsock_buf *
> +virtio_transport_alloc_buf(struct virtio_vsock_pkt *pkt, bool zero_copy)
> +{
> +	struct virtio_vsock_buf *buf;
> +
> +	if (pkt->len == 0)
> +		return NULL;
> +
> +	buf = kzalloc(sizeof(*buf), GFP_KERNEL);
> +	if (!buf)
> +		return NULL;
> +
> +	/* If the buffer in the virtio_vsock_pkt is full, we can move it to
> +	 * the new virtio_vsock_buf avoiding the copy, because we are sure that
> +	 * we are not use
we do not use
> more memory than that counted by the credit mechanism.
> +	 */
> +	if (zero_copy && pkt->len == pkt->buf_len) {
> +		buf->addr = pkt->buf;
> +		pkt->buf = NULL;
> +	} else {
> +		buf->addr = kmalloc(pkt->len, GFP_KERNEL);
> +		if (!buf->addr) {
> +			kfree(buf);
> +			return NULL;
> +		}
> +
> +		memcpy(buf->addr, pkt->buf, pkt->len);
> +	}
> +
> +	buf->len = pkt->len;
> +
> +	return buf;
> +}
> +
> +static void virtio_transport_free_buf(struct virtio_vsock_buf *buf)
> +{
> +	kfree(buf->addr);
> +	kfree(buf);
> +}
> +
>  /* Packet capture */
>  static struct sk_buff *virtio_transport_build_skb(void *opaque)
>  {
> @@ -190,17 +233,15 @@ static int virtio_transport_send_pkt_info(struct
vsock_sock *vsk,
>  	return virtio_transport_get_ops()->send_pkt(pkt);
>  }
>  
> -static void virtio_transport_inc_rx_pkt(struct virtio_vsock_sock *vvs,
> -					struct virtio_vsock_pkt *pkt)
> +static void virtio_transport_inc_rx_pkt(struct virtio_vsock_sock *vvs, u32
len)
>  {
> -	vvs->rx_bytes += pkt->len;
> +	vvs->rx_bytes += len;
>  }
>  
> -static void virtio_transport_dec_rx_pkt(struct virtio_vsock_sock *vvs,
> -					struct virtio_vsock_pkt *pkt)
> +static void virtio_transport_dec_rx_pkt(struct virtio_vsock_sock *vvs, u32
len)
>  {
> -	vvs->rx_bytes -= pkt->len;
> -	vvs->fwd_cnt += pkt->len;
> +	vvs->rx_bytes -= len;
> +	vvs->fwd_cnt += len;
>  }
>  
>  void virtio_transport_inc_tx_pkt(struct virtio_vsock_sock *vvs, struct
virtio_vsock_pkt *pkt)
> @@ -254,36 +295,36 @@ virtio_transport_stream_do_dequeue(struct vsock_sock
*vsk,
>  				   size_t len)
>  {
>  	struct virtio_vsock_sock *vvs = vsk->trans;
> -	struct virtio_vsock_pkt *pkt;
> +	struct virtio_vsock_buf *buf;
>  	size_t bytes, total = 0;
>  	int err = -EFAULT;
>  
>  	spin_lock_bh(&vvs->rx_lock);
>  	while (total < len && !list_empty(&vvs->rx_queue)) {
> -		pkt = list_first_entry(&vvs->rx_queue,
> -				       struct virtio_vsock_pkt, list);
> +		buf = list_first_entry(&vvs->rx_queue,
> +				       struct virtio_vsock_buf, list);
>  
>  		bytes = len - total;
> -		if (bytes > pkt->len - pkt->off)
> -			bytes = pkt->len - pkt->off;
> +		if (bytes > buf->len - buf->off)
> +			bytes = buf->len - buf->off;
>  
>  		/* sk_lock is held by caller so no one else can dequeue.
>  		 * Unlock rx_lock since memcpy_to_msg() may sleep.
>  		 */
>  		spin_unlock_bh(&vvs->rx_lock);
>  
> -		err = memcpy_to_msg(msg, pkt->buf + pkt->off, bytes);
> +		err = memcpy_to_msg(msg, buf->addr + buf->off, bytes);
>  		if (err)
>  			goto out;
>  
>  		spin_lock_bh(&vvs->rx_lock);
>  
>  		total += bytes;
> -		pkt->off += bytes;
> -		if (pkt->off == pkt->len) {
> -			virtio_transport_dec_rx_pkt(vvs, pkt);
> -			list_del(&pkt->list);
> -			virtio_transport_free_pkt(pkt);
> +		buf->off += bytes;
> +		if (buf->off == buf->len) {
> +			virtio_transport_dec_rx_pkt(vvs, buf->len);
> +			list_del(&buf->list);
> +			virtio_transport_free_buf(buf);
>  		}
>  	}
>  	spin_unlock_bh(&vvs->rx_lock);
> @@ -841,20 +882,24 @@ virtio_transport_recv_connected(struct sock *sk,
>  {
>  	struct vsock_sock *vsk = vsock_sk(sk);
>  	struct virtio_vsock_sock *vvs = vsk->trans;
> +	struct virtio_vsock_buf *buf;
>  	int err = 0;
>  
>  	switch (le16_to_cpu(pkt->hdr.op)) {
>  	case VIRTIO_VSOCK_OP_RW:
>  		pkt->len = le32_to_cpu(pkt->hdr.len);
> -		pkt->off = 0;
> +		buf = virtio_transport_alloc_buf(pkt, true);

This seems to be the only callers and second parameter
is always true. So why is it needed?
>  
> -		spin_lock_bh(&vvs->rx_lock);
> -		virtio_transport_inc_rx_pkt(vvs, pkt);
> -		list_add_tail(&pkt->list, &vvs->rx_queue);
> -		spin_unlock_bh(&vvs->rx_lock);
> +		if (buf) {
> +			spin_lock_bh(&vvs->rx_lock);
> +			virtio_transport_inc_rx_pkt(vvs, pkt->len);
> +			list_add_tail(&buf->list, &vvs->rx_queue);
> +			spin_unlock_bh(&vvs->rx_lock);
>  
> -		sk->sk_data_ready(sk);
> -		return err;
> +			sk->sk_data_ready(sk);
> +		}
> +
> +		break;
>  	case VIRTIO_VSOCK_OP_CREDIT_UPDATE:
>  		sk->sk_write_space(sk);
>  		break;
> -- 
> 2.20.1

On 2019/5/10 ??8:58, Stefano Garzarella wrote:
> While I was testing this new series (v2) I discovered an huge use of memory
> and a memory leak in the virtio-vsock driver in the guest when I sent
> 1-byte packets to the guest.
>
> These issues are present since the introduction of the virtio-vsock
> driver. I added the patches 1 and 2 to fix them in this series in order
> to better track the performance trends.
>
> v1: https://patchwork.kernel.org/cover/10885431/
>
> v2:
> - Add patch 1 to limit the memory usage
> - Add patch 2 to avoid memory leak during the socket release
> - Add patch 3 to fix locking of fwd_cnt and buf_alloc
> - Patch 4: fix 'free_space' type (u32 instead of s64) [Stefan]
> - Patch 5: Avoid integer underflow of iov_len [Stefan]
> - Patch 5: Fix packet capture in order to see the exact packets that are
>             delivered. [Stefan]
> - Add patch 8 to make the RX buffer size tunable [Stefan]
>
> Below are the benchmarks step by step. I used iperf3 [1] modified with
VSOCK
> support.
> As Micheal suggested in the v1, I booted host and guest with
'nosmap', and I
> added a column with virtio-net+vhost-net performance.
>
> A brief description of patches:
> - Patches 1+2: limit the memory usage with an extra copy and avoid memory
leak
> - Patches 3+4: fix locking and reduce the number of credit update messages
sent
>                 to the transmitter
> - Patches 5+6: allow the host to split packets on multiple buffers and use
>                 VIRTIO_VSOCK_MAX_PKT_BUF_SIZE as the max packet size
allowed
> - Patches 7+8: increase RX buffer size to 64 KiB
>
>                      host -> guest [Gbps]
> pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net +
vhost
>                                                                      
TCP_NODELAY
> 64         0.068     0.063    0.130    0.131    0.128         0.188    
0.187
> 256        0.274     0.236    0.392    0.338    0.282         0.749    
0.654
> 512        0.531     0.457    0.862    0.725    0.602         1.419    
1.414
> 1K         0.954     0.827    1.591    1.598    1.548         2.599    
2.640
> 2K         1.783     1.543    3.731    3.637    3.469         4.530    
4.754
> 4K         3.332     3.436    7.164    7.124    6.494         7.738    
7.696
> 8K         5.792     5.530   11.653   11.787   11.444        12.307   
11.850
> 16K        8.405     8.462   16.372   16.855   17.562        16.936   
16.954
> 32K       14.208    13.669   18.945   20.009   23.128        21.980   
23.015
> 64K       21.082    18.893   20.266   20.903   30.622        27.290   
27.383
> 128K      20.696    20.148   20.112   21.746   32.152        30.446   
30.990
> 256K      20.801    20.589   20.725   22.685   34.721        33.151   
32.745
> 512K      21.220    20.465   20.432   22.106   34.496        36.847   
31.096
>
>                      guest -> host [Gbps]
> pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net +
vhost
>                                                                      
TCP_NODELAY
> 64         0.089     0.091    0.120    0.115    0.117         0.274    
0.272
> 256        0.352     0.354    0.452    0.445    0.451         1.085    
1.136
> 512        0.705     0.704    0.893    0.858    0.898         2.131    
1.882
> 1K         1.394     1.433    1.721    1.669    1.691         3.984    
3.576
> 2K         2.818     2.874    3.316    3.249    3.303         6.719    
6.359
> 4K         5.293     5.397    6.129    5.933    6.082        10.105    
9.860
> 8K         8.890     9.151   10.990   10.545   10.519        15.239   
14.868
> 16K       11.444    11.018   12.074   15.255   15.577        20.551   
20.848
> 32K       11.229    10.875   10.857   24.401   25.227        26.294   
26.380
> 64K       10.832    10.545   10.816   39.487   39.616        34.996   
32.041
> 128K      10.435    10.241   10.500   39.813   40.012        38.379   
35.055
> 256K      10.263     9.866    9.845   34.971   35.143        36.559   
37.232
> 512K      10.224    10.060   10.092   35.469   34.627        34.963   
33.401
>
> As Stefan suggested in the v1, this time I measured also the efficiency in
this
> way:
>      efficiency = Mbps / (%CPU_Host + %CPU_Guest)
>
> The '%CPU_Guest' is taken inside the VM. I know that it is not the
best way,
> but it's provided for free from iperf3 and could be an indication.
>
>          host -> guest efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
> pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net +
vhost
>                                                                      
TCP_NODELAY
> 64          0.94      0.59     3.96     4.06     4.09          2.82     
2.11
> 256         2.62      2.50     6.45     6.09     5.81          9.64     
8.73
> 512         5.16      4.87    13.16    12.39    11.67         17.83    
17.76
> 1K          9.16      8.85    24.98    24.97    25.01         32.57    
32.04
> 2K         17.41     17.03    49.09    48.59    49.22         55.31    
57.14
> 4K         32.99     33.62    90.80    90.98    91.72         91.79    
91.40
> 8K         58.51     59.98   153.53   170.83   167.31        137.51   
132.85
> 16K        89.32     95.29   216.98   264.18   260.95        176.05   
176.05
> 32K       152.94    167.10   285.75   387.02   360.81        215.49   
226.30
> 64K       250.38    307.20   317.65   489.53   472.70        238.97   
244.27
> 128K      327.99    335.24   335.76   523.71   486.41        253.29   
260.86
> 256K      327.06    334.24   338.64   533.76   509.85        267.78   
266.22
> 512K      337.36    330.61   334.95   512.90   496.35        280.42   
241.43
>
>          guest -> host efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
> pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net +
vhost
>                                                                      
TCP_NODELAY
> 64          0.90      0.91     1.37     1.32     1.35          2.15     
2.13
> 256         3.59      3.55     5.23     5.19     5.29          8.50     
8.89
> 512         7.19      7.08    10.21     9.95    10.38         16.74    
14.71
> 1K         14.15     14.34    19.85    19.06    19.33         31.44    
28.11
> 2K         28.44     29.09    37.78    37.18    37.49         53.07    
50.63
> 4K         55.37     57.60    71.02    69.27    70.97         81.56    
79.32
> 8K        105.58    100.45   111.95   124.68   123.61        120.85   
118.66
> 16K       141.63    138.24   137.67   187.41   190.20        160.43   
163.00
> 32K       147.56    143.09   138.48   296.41   301.04        214.64   
223.94
> 64K       144.81    143.27   138.49   433.98   462.26        298.86   
269.71
> 128K      150.14    147.99   146.85   511.36   514.29        350.17   
298.09
> 256K      156.69    152.25   148.69   542.19   549.97        326.42   
333.32
> 512K      157.29    153.35   152.22   546.52   533.24        315.55   
302.27
>
> [1] https://github.com/stefano-garzarella/iperf/

Hi:

Do you have any explanation that vsock is better here? Is this because 
of the mergeable buffer? If you, we need test with mrg_rxbuf=off.

Thanks

>
> Stefano Garzarella (8):
>    vsock/virtio: limit the memory used per-socket
>    vsock/virtio: free packets during the socket release
>    vsock/virtio: fix locking for fwd_cnt and buf_alloc
>    vsock/virtio: reduce credit update messages
>    vhost/vsock: split packets to send using multiple buffers
>    vsock/virtio: change the maximum packet size allowed
>    vsock/virtio: increase RX buffer size to 64 KiB
>    vsock/virtio: make the RX buffer size tunable
>
>   drivers/vhost/vsock.c                   |  53 +++++++--
>   include/linux/virtio_vsock.h            |  14 ++-
>   net/vmw_vsock/virtio_transport.c        |  28 ++++-
>   net/vmw_vsock/virtio_transport_common.c | 144 ++++++++++++++++++------
>   4 files changed, 190 insertions(+), 49 deletions(-)
>

On 2019/5/10 ??8:58, Stefano Garzarella wrote:
> Since virtio-vsock was introduced, the buffers filled by the host
> and pushed to the guest using the vring, are directly queued in
> a per-socket list avoiding to copy it.
> These buffers are preallocated by the guest with a fixed
> size (4 KB).
>
> The maximum amount of memory used by each socket should be
> controlled by the credit mechanism.
> The default credit available per-socket is 256 KB, but if we use
> only 1 byte per packet, the guest can queue up to 262144 of 4 KB
> buffers, using up to 1 GB of memory per-socket. In addition, the
> guest will continue to fill the vring with new 4 KB free buffers
> to avoid starvation of other sockets.
>
> This patch solves this issue copying the payload in a new buffer.
> Then it is queued in the per-socket list, and the 4KB buffer used
> by the host is freed.
>
> In this way, the memory used by each socket respects the credit
> available, and we still avoid starvation, paying the cost of an
> extra memory copy. When the buffer is completely full we do a
> "zero-copy", moving the buffer directly in the per-socket list.

I wonder in the long run we should use generic socket accouting 
mechanism provided by kernel (e.g socket, skb, sndbuf, recvbug, 
truesize) instead of vsock specific thing to avoid duplicating efforts.

>
> Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
> ---
>   drivers/vhost/vsock.c                   |  2 +
>   include/linux/virtio_vsock.h            |  8 +++
>   net/vmw_vsock/virtio_transport.c        |  1 +
>   net/vmw_vsock/virtio_transport_common.c | 95 ++++++++++++++++++-------
>   4 files changed, 81 insertions(+), 25 deletions(-)
>
> diff --git a/drivers/vhost/vsock.c b/drivers/vhost/vsock.c
> index bb5fc0e9fbc2..7964e2daee09 100644
> --- a/drivers/vhost/vsock.c
> +++ b/drivers/vhost/vsock.c
> @@ -320,6 +320,8 @@ vhost_vsock_alloc_pkt(struct vhost_virtqueue *vq,
>   		return NULL;
>   	}
>   
> +	pkt->buf_len = pkt->len;
> +
>   	nbytes = copy_from_iter(pkt->buf, pkt->len, &iov_iter);
>   	if (nbytes != pkt->len) {
>   		vq_err(vq, "Expected %u byte payload, got %zu bytes\n",
> diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
> index e223e2632edd..345f04ee9193 100644
> --- a/include/linux/virtio_vsock.h
> +++ b/include/linux/virtio_vsock.h
> @@ -54,9 +54,17 @@ struct virtio_vsock_pkt {
>   	void *buf;
>   	u32 len;
>   	u32 off;
> +	u32 buf_len;
>   	bool reply;
>   };
>   
> +struct virtio_vsock_buf {
> +	struct list_head list;
> +	void *addr;
> +	u32 len;
> +	u32 off;
> +};
> +
>   struct virtio_vsock_pkt_info {
>   	u32 remote_cid, remote_port;
>   	struct vsock_sock *vsk;
> diff --git a/net/vmw_vsock/virtio_transport.c
b/net/vmw_vsock/virtio_transport.c
> index 15eb5d3d4750..af1d2ce12f54 100644
> --- a/net/vmw_vsock/virtio_transport.c
> +++ b/net/vmw_vsock/virtio_transport.c
> @@ -280,6 +280,7 @@ static void virtio_vsock_rx_fill(struct virtio_vsock
*vsock)
>   			break;
>   		}
>   
> +		pkt->buf_len = buf_len;
>   		pkt->len = buf_len;
>   
>   		sg_init_one(&hdr, &pkt->hdr, sizeof(pkt->hdr));
> diff --git a/net/vmw_vsock/virtio_transport_common.c
b/net/vmw_vsock/virtio_transport_common.c
> index 602715fc9a75..0248d6808755 100644
> --- a/net/vmw_vsock/virtio_transport_common.c
> +++ b/net/vmw_vsock/virtio_transport_common.c
> @@ -65,6 +65,9 @@ virtio_transport_alloc_pkt(struct virtio_vsock_pkt_info
*info,
>   		pkt->buf = kmalloc(len, GFP_KERNEL);
>   		if (!pkt->buf)
>   			goto out_pkt;
> +
> +		pkt->buf_len = len;
> +
>   		err = memcpy_from_msg(pkt->buf, info->msg, len);
>   		if (err)
>   			goto out;
> @@ -86,6 +89,46 @@ virtio_transport_alloc_pkt(struct virtio_vsock_pkt_info
*info,
>   	return NULL;
>   }
>   
> +static struct virtio_vsock_buf *
> +virtio_transport_alloc_buf(struct virtio_vsock_pkt *pkt, bool zero_copy)
> +{
> +	struct virtio_vsock_buf *buf;
> +
> +	if (pkt->len == 0)
> +		return NULL;
> +
> +	buf = kzalloc(sizeof(*buf), GFP_KERNEL);
> +	if (!buf)
> +		return NULL;
> +
> +	/* If the buffer in the virtio_vsock_pkt is full, we can move it to
> +	 * the new virtio_vsock_buf avoiding the copy, because we are sure that
> +	 * we are not use more memory than that counted by the credit mechanism.
> +	 */
> +	if (zero_copy && pkt->len == pkt->buf_len) {
> +		buf->addr = pkt->buf;
> +		pkt->buf = NULL;
> +	} else {

Is the copy still needed if we're just few bytes less? We meet similar 
issue for virito-net, and virtio-net solve this by always copy first 
128bytes for big packets.

See receive_big().

Thanks

> +		buf->addr = kmalloc(pkt->len, GFP_KERNEL);
> +		if (!buf->addr) {
> +			kfree(buf);
> +			return NULL;
> +		}
> +
> +		memcpy(buf->addr, pkt->buf, pkt->len);
> +	}
> +
> +	buf->len = pkt->len;
> +
> +	return buf;
> +}
> +
> +static void virtio_transport_free_buf(struct virtio_vsock_buf *buf)
> +{
> +	kfree(buf->addr);
> +	kfree(buf);
> +}
> +
>   /* Packet capture */
>   static struct sk_buff *virtio_transport_build_skb(void *opaque)
>   {
> @@ -190,17 +233,15 @@ static int virtio_transport_send_pkt_info(struct
vsock_sock *vsk,
>   	return virtio_transport_get_ops()->send_pkt(pkt);
>   }
>   
> -static void virtio_transport_inc_rx_pkt(struct virtio_vsock_sock *vvs,
> -					struct virtio_vsock_pkt *pkt)
> +static void virtio_transport_inc_rx_pkt(struct virtio_vsock_sock *vvs, u32
len)
>   {
> -	vvs->rx_bytes += pkt->len;
> +	vvs->rx_bytes += len;
>   }
>   
> -static void virtio_transport_dec_rx_pkt(struct virtio_vsock_sock *vvs,
> -					struct virtio_vsock_pkt *pkt)
> +static void virtio_transport_dec_rx_pkt(struct virtio_vsock_sock *vvs, u32
len)
>   {
> -	vvs->rx_bytes -= pkt->len;
> -	vvs->fwd_cnt += pkt->len;
> +	vvs->rx_bytes -= len;
> +	vvs->fwd_cnt += len;
>   }
>   
>   void virtio_transport_inc_tx_pkt(struct virtio_vsock_sock *vvs, struct
virtio_vsock_pkt *pkt)
> @@ -254,36 +295,36 @@ virtio_transport_stream_do_dequeue(struct vsock_sock
*vsk,
>   				   size_t len)
>   {
>   	struct virtio_vsock_sock *vvs = vsk->trans;
> -	struct virtio_vsock_pkt *pkt;
> +	struct virtio_vsock_buf *buf;
>   	size_t bytes, total = 0;
>   	int err = -EFAULT;
>   
>   	spin_lock_bh(&vvs->rx_lock);
>   	while (total < len && !list_empty(&vvs->rx_queue)) {
> -		pkt = list_first_entry(&vvs->rx_queue,
> -				       struct virtio_vsock_pkt, list);
> +		buf = list_first_entry(&vvs->rx_queue,
> +				       struct virtio_vsock_buf, list);
>   
>   		bytes = len - total;
> -		if (bytes > pkt->len - pkt->off)
> -			bytes = pkt->len - pkt->off;
> +		if (bytes > buf->len - buf->off)
> +			bytes = buf->len - buf->off;
>   
>   		/* sk_lock is held by caller so no one else can dequeue.
>   		 * Unlock rx_lock since memcpy_to_msg() may sleep.
>   		 */
>   		spin_unlock_bh(&vvs->rx_lock);
>   
> -		err = memcpy_to_msg(msg, pkt->buf + pkt->off, bytes);
> +		err = memcpy_to_msg(msg, buf->addr + buf->off, bytes);
>   		if (err)
>   			goto out;
>   
>   		spin_lock_bh(&vvs->rx_lock);
>   
>   		total += bytes;
> -		pkt->off += bytes;
> -		if (pkt->off == pkt->len) {
> -			virtio_transport_dec_rx_pkt(vvs, pkt);
> -			list_del(&pkt->list);
> -			virtio_transport_free_pkt(pkt);
> +		buf->off += bytes;
> +		if (buf->off == buf->len) {
> +			virtio_transport_dec_rx_pkt(vvs, buf->len);
> +			list_del(&buf->list);
> +			virtio_transport_free_buf(buf);
>   		}
>   	}
>   	spin_unlock_bh(&vvs->rx_lock);
> @@ -841,20 +882,24 @@ virtio_transport_recv_connected(struct sock *sk,
>   {
>   	struct vsock_sock *vsk = vsock_sk(sk);
>   	struct virtio_vsock_sock *vvs = vsk->trans;
> +	struct virtio_vsock_buf *buf;
>   	int err = 0;
>   
>   	switch (le16_to_cpu(pkt->hdr.op)) {
>   	case VIRTIO_VSOCK_OP_RW:
>   		pkt->len = le32_to_cpu(pkt->hdr.len);
> -		pkt->off = 0;
> +		buf = virtio_transport_alloc_buf(pkt, true);
>   
> -		spin_lock_bh(&vvs->rx_lock);
> -		virtio_transport_inc_rx_pkt(vvs, pkt);
> -		list_add_tail(&pkt->list, &vvs->rx_queue);
> -		spin_unlock_bh(&vvs->rx_lock);
> +		if (buf) {
> +			spin_lock_bh(&vvs->rx_lock);
> +			virtio_transport_inc_rx_pkt(vvs, pkt->len);
> +			list_add_tail(&buf->list, &vvs->rx_queue);
> +			spin_unlock_bh(&vvs->rx_lock);
>   
> -		sk->sk_data_ready(sk);
> -		return err;
> +			sk->sk_data_ready(sk);
> +		}
> +
> +		break;
>   	case VIRTIO_VSOCK_OP_CREDIT_UPDATE:
>   		sk->sk_write_space(sk);
>   		break;

On 2019/5/10 ??8:58, Stefano Garzarella wrote:
> In order to increase host -> guest throughput with large packets,
> we can use 64 KiB RX buffers.
>
> Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
> ---
>   include/linux/virtio_vsock.h | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
> index 84b72026d327..5a9d25be72df 100644
> --- a/include/linux/virtio_vsock.h
> +++ b/include/linux/virtio_vsock.h
> @@ -10,7 +10,7 @@
>   #define VIRTIO_VSOCK_DEFAULT_MIN_BUF_SIZE	128
>   #define VIRTIO_VSOCK_DEFAULT_BUF_SIZE		(1024 * 256)
>   #define VIRTIO_VSOCK_DEFAULT_MAX_BUF_SIZE	(1024 * 256)
> -#define VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE	(1024 * 4)
> +#define VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE	(1024 * 64)
>   #define VIRTIO_VSOCK_MAX_BUF_SIZE		0xFFFFFFFFUL
>   #define VIRTIO_VSOCK_MAX_PKT_BUF_SIZE		(1024 * 64)
>   

We probably don't want such high order allocation. It's better to switch
to use order 0 pages in this case. See add_recvbuf_big() for virtio-net. 
If we get datapath unified, we will get more stuffs set.

Thanks

On 2019/5/10 ??8:58, Stefano Garzarella wrote:
> The RX buffer size determines the memory consumption of the
> vsock/virtio guest driver, so we make it tunable through
> a module parameter.
>
> The size allowed are between 4 KB and 64 KB in order to be
> compatible with old host drivers.
>
> Suggested-by: Stefan Hajnoczi <stefanha at redhat.com>
> Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>

I don't see much value of doing this through kernel command line. We 
should deal with them automatically like what virtio-net did. Or even a 
module parameter is better.

Thanks

> ---
>   include/linux/virtio_vsock.h     |  1 +
>   net/vmw_vsock/virtio_transport.c | 27 ++++++++++++++++++++++++++-
>   2 files changed, 27 insertions(+), 1 deletion(-)
>
> diff --git a/include/linux/virtio_vsock.h b/include/linux/virtio_vsock.h
> index 5a9d25be72df..b9f8c3d91f80 100644
> --- a/include/linux/virtio_vsock.h
> +++ b/include/linux/virtio_vsock.h
> @@ -13,6 +13,7 @@
>   #define VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE	(1024 * 64)
>   #define VIRTIO_VSOCK_MAX_BUF_SIZE		0xFFFFFFFFUL
>   #define VIRTIO_VSOCK_MAX_PKT_BUF_SIZE		(1024 * 64)
> +#define VIRTIO_VSOCK_MIN_PKT_BUF_SIZE		(1024 * 4)
>   
>   enum {
>   	VSOCK_VQ_RX     = 0, /* for host to guest data */
> diff --git a/net/vmw_vsock/virtio_transport.c
b/net/vmw_vsock/virtio_transport.c
> index af1d2ce12f54..732398b4e28f 100644
> --- a/net/vmw_vsock/virtio_transport.c
> +++ b/net/vmw_vsock/virtio_transport.c
> @@ -66,6 +66,31 @@ struct virtio_vsock {
>   	u32 guest_cid;
>   };
>   
> +static unsigned int rx_buf_size = VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE;
> +
> +static int param_set_rx_buf_size(const char *val, const struct
kernel_param *kp)
> +{
> +	unsigned int size;
> +	int ret;
> +
> +	ret = kstrtouint(val, 0, &size);
> +	if (ret)
> +		return ret;
> +
> +	if (size < VIRTIO_VSOCK_MIN_PKT_BUF_SIZE ||
> +	    size > VIRTIO_VSOCK_MAX_PKT_BUF_SIZE)
> +		return -EINVAL;
> +
> +	return param_set_uint(val, kp);
> +};
> +
> +static const struct kernel_param_ops param_ops_rx_buf_size = {
> +	.set = param_set_rx_buf_size,
> +	.get = param_get_uint,
> +};
> +
> +module_param_cb(rx_buf_size, &param_ops_rx_buf_size, &rx_buf_size,
0644);
> +
>   static struct virtio_vsock *virtio_vsock_get(void)
>   {
>   	return the_virtio_vsock;
> @@ -261,7 +286,7 @@ virtio_transport_cancel_pkt(struct vsock_sock *vsk)
>   
>   static void virtio_vsock_rx_fill(struct virtio_vsock *vsock)
>   {
> -	int buf_len = VIRTIO_VSOCK_DEFAULT_RX_BUF_SIZE;
> +	int buf_len = rx_buf_size;
>   	struct virtio_vsock_pkt *pkt;
>   	struct scatterlist hdr, buf, *sgs[2];
>   	struct virtqueue *vq;

On Mon, May 13, 2019 at 05:33:40PM +0800, Jason Wang
wrote:
> 
> On 2019/5/10 ??8:58, Stefano Garzarella wrote:
> > While I was testing this new series (v2) I discovered an huge use of
memory
> > and a memory leak in the virtio-vsock driver in the guest when I sent
> > 1-byte packets to the guest.
> > 
> > These issues are present since the introduction of the virtio-vsock
> > driver. I added the patches 1 and 2 to fix them in this series in
order
> > to better track the performance trends.
> > 
> > v1: https://patchwork.kernel.org/cover/10885431/
> > 
> > v2:
> > - Add patch 1 to limit the memory usage
> > - Add patch 2 to avoid memory leak during the socket release
> > - Add patch 3 to fix locking of fwd_cnt and buf_alloc
> > - Patch 4: fix 'free_space' type (u32 instead of s64) [Stefan]
> > - Patch 5: Avoid integer underflow of iov_len [Stefan]
> > - Patch 5: Fix packet capture in order to see the exact packets that
are
> >             delivered. [Stefan]
> > - Add patch 8 to make the RX buffer size tunable [Stefan]
> > 
> > Below are the benchmarks step by step. I used iperf3 [1] modified with
VSOCK
> > support.
> > As Micheal suggested in the v1, I booted host and guest with
'nosmap', and I
> > added a column with virtio-net+vhost-net performance.
> > 
> > A brief description of patches:
> > - Patches 1+2: limit the memory usage with an extra copy and avoid
memory leak
> > - Patches 3+4: fix locking and reduce the number of credit update
messages sent
> >                 to the transmitter
> > - Patches 5+6: allow the host to split packets on multiple buffers and
use
> >                 VIRTIO_VSOCK_MAX_PKT_BUF_SIZE as the max packet size
allowed
> > - Patches 7+8: increase RX buffer size to 64 KiB
> > 
> >                      host -> guest [Gbps]
> > pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net
+ vhost
> >                                                                      
TCP_NODELAY
> > 64         0.068     0.063    0.130    0.131    0.128         0.188   
0.187
> > 256        0.274     0.236    0.392    0.338    0.282         0.749   
0.654
> > 512        0.531     0.457    0.862    0.725    0.602         1.419   
1.414
> > 1K         0.954     0.827    1.591    1.598    1.548         2.599   
2.640
> > 2K         1.783     1.543    3.731    3.637    3.469         4.530   
4.754
> > 4K         3.332     3.436    7.164    7.124    6.494         7.738   
7.696
> > 8K         5.792     5.530   11.653   11.787   11.444        12.307   
11.850
> > 16K        8.405     8.462   16.372   16.855   17.562        16.936   
16.954
> > 32K       14.208    13.669   18.945   20.009   23.128        21.980   
23.015
> > 64K       21.082    18.893   20.266   20.903   30.622        27.290   
27.383
> > 128K      20.696    20.148   20.112   21.746   32.152        30.446   
30.990
> > 256K      20.801    20.589   20.725   22.685   34.721        33.151   
32.745
> > 512K      21.220    20.465   20.432   22.106   34.496        36.847   
31.096
> > 
> >                      guest -> host [Gbps]
> > pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net
+ vhost
> >                                                                      
TCP_NODELAY
> > 64         0.089     0.091    0.120    0.115    0.117         0.274   
0.272
> > 256        0.352     0.354    0.452    0.445    0.451         1.085   
1.136
> > 512        0.705     0.704    0.893    0.858    0.898         2.131   
1.882
> > 1K         1.394     1.433    1.721    1.669    1.691         3.984   
3.576
> > 2K         2.818     2.874    3.316    3.249    3.303         6.719   
6.359
> > 4K         5.293     5.397    6.129    5.933    6.082        10.105   
9.860
> > 8K         8.890     9.151   10.990   10.545   10.519        15.239   
14.868
> > 16K       11.444    11.018   12.074   15.255   15.577        20.551   
20.848
> > 32K       11.229    10.875   10.857   24.401   25.227        26.294   
26.380
> > 64K       10.832    10.545   10.816   39.487   39.616        34.996   
32.041
> > 128K      10.435    10.241   10.500   39.813   40.012        38.379   
35.055
> > 256K      10.263     9.866    9.845   34.971   35.143        36.559   
37.232
> > 512K      10.224    10.060   10.092   35.469   34.627        34.963   
33.401
> > 
> > As Stefan suggested in the v1, this time I measured also the
efficiency in this
> > way:
> >      efficiency = Mbps / (%CPU_Host + %CPU_Guest)
> > 
> > The '%CPU_Guest' is taken inside the VM. I know that it is not
the best way,
> > but it's provided for free from iperf3 and could be an indication.
> > 
> >          host -> guest efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
> > pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net
+ vhost
> >                                                                      
TCP_NODELAY
> > 64          0.94      0.59     3.96     4.06     4.09          2.82   
2.11
> > 256         2.62      2.50     6.45     6.09     5.81          9.64   
8.73
> > 512         5.16      4.87    13.16    12.39    11.67         17.83   
17.76
> > 1K          9.16      8.85    24.98    24.97    25.01         32.57   
32.04
> > 2K         17.41     17.03    49.09    48.59    49.22         55.31   
57.14
> > 4K         32.99     33.62    90.80    90.98    91.72         91.79   
91.40
> > 8K         58.51     59.98   153.53   170.83   167.31        137.51   
132.85
> > 16K        89.32     95.29   216.98   264.18   260.95        176.05   
176.05
> > 32K       152.94    167.10   285.75   387.02   360.81        215.49   
226.30
> > 64K       250.38    307.20   317.65   489.53   472.70        238.97   
244.27
> > 128K      327.99    335.24   335.76   523.71   486.41        253.29   
260.86
> > 256K      327.06    334.24   338.64   533.76   509.85        267.78   
266.22
> > 512K      337.36    330.61   334.95   512.90   496.35        280.42   
241.43
> > 
> >          guest -> host efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
> > pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net
+ vhost
> >                                                                      
TCP_NODELAY
> > 64          0.90      0.91     1.37     1.32     1.35          2.15   
2.13
> > 256         3.59      3.55     5.23     5.19     5.29          8.50   
8.89
> > 512         7.19      7.08    10.21     9.95    10.38         16.74   
14.71
> > 1K         14.15     14.34    19.85    19.06    19.33         31.44   
28.11
> > 2K         28.44     29.09    37.78    37.18    37.49         53.07   
50.63
> > 4K         55.37     57.60    71.02    69.27    70.97         81.56   
79.32
> > 8K        105.58    100.45   111.95   124.68   123.61        120.85   
118.66
> > 16K       141.63    138.24   137.67   187.41   190.20        160.43   
163.00
> > 32K       147.56    143.09   138.48   296.41   301.04        214.64   
223.94
> > 64K       144.81    143.27   138.49   433.98   462.26        298.86   
269.71
> > 128K      150.14    147.99   146.85   511.36   514.29        350.17   
298.09
> > 256K      156.69    152.25   148.69   542.19   549.97        326.42   
333.32
> > 512K      157.29    153.35   152.22   546.52   533.24        315.55   
302.27
> > 
> > [1] https://github.com/stefano-garzarella/iperf/
> 
> 
> Hi:
> 
> Do you have any explanation that vsock is better here? Is this because of
> the mergeable buffer? If you, we need test with mrg_rxbuf=off.
Hi Jason,

virtio-net stays faster for packets with size up tp 16K/32K, maybe, as
you suggested, could be releated to mergeable buffer.

I'll try to disable it and re-run the tests.

Thanks,
Stefano

On Fri, May 10, 2019 at 02:58:36PM +0200, Stefano Garzarella
wrote:
> +struct virtio_vsock_buf {
Please add a comment describing the purpose of this struct and to
differentiate its use from struct virtio_vsock_pkt.
> +static struct virtio_vsock_buf *
> +virtio_transport_alloc_buf(struct virtio_vsock_pkt *pkt, bool zero_copy)
> +{
> +	struct virtio_vsock_buf *buf;
> +
> +	if (pkt->len == 0)
> +		return NULL;
> +
> +	buf = kzalloc(sizeof(*buf), GFP_KERNEL);
> +	if (!buf)
> +		return NULL;
> +
> +	/* If the buffer in the virtio_vsock_pkt is full, we can move it to
> +	 * the new virtio_vsock_buf avoiding the copy, because we are sure that
> +	 * we are not use more memory than that counted by the credit mechanism.
> +	 */
> +	if (zero_copy && pkt->len == pkt->buf_len) {
> +		buf->addr = pkt->buf;
> +		pkt->buf = NULL;
> +	} else {
> +		buf->addr = kmalloc(pkt->len, GFP_KERNEL);
buf and buf->addr could be allocated in a single call, though I'm not
sure how big an optimization this is.
> @@ -841,20 +882,24 @@ virtio_transport_recv_connected(struct sock *sk,
>  {
>  	struct vsock_sock *vsk = vsock_sk(sk);
>  	struct virtio_vsock_sock *vvs = vsk->trans;
> +	struct virtio_vsock_buf *buf;
>  	int err = 0;
>  
>  	switch (le16_to_cpu(pkt->hdr.op)) {
>  	case VIRTIO_VSOCK_OP_RW:
>  		pkt->len = le32_to_cpu(pkt->hdr.len);
> -		pkt->off = 0;
> +		buf = virtio_transport_alloc_buf(pkt, true);
>  
> -		spin_lock_bh(&vvs->rx_lock);
> -		virtio_transport_inc_rx_pkt(vvs, pkt);
> -		list_add_tail(&pkt->list, &vvs->rx_queue);
> -		spin_unlock_bh(&vvs->rx_lock);
> +		if (buf) {
> +			spin_lock_bh(&vvs->rx_lock);
> +			virtio_transport_inc_rx_pkt(vvs, pkt->len);
> +			list_add_tail(&buf->list, &vvs->rx_queue);
> +			spin_unlock_bh(&vvs->rx_lock);
>  
> -		sk->sk_data_ready(sk);
> -		return err;
> +			sk->sk_data_ready(sk);
> +		}
The return value of this function isn't used but the code still makes an
effort to return errors.  Please return -ENOMEM when buf == NULL.

If you'd like to remove the return value that's fine too, but please do
it for the whole function to be consistent.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 488 bytes
Desc: not available
URL:
<http://lists.linuxfoundation.org/pipermail/virtualization/attachments/20190516/1570bbca/attachment-0001.sig>

On Fri, May 10, 2019 at 02:58:37PM +0200, Stefano Garzarella
wrote:
> When the socket is released, we should free all packets
> queued in the per-socket list in order to avoid a memory
> leak.
> 
> Signed-off-by: Stefano Garzarella <sgarzare at redhat.com>
> ---
>  net/vmw_vsock/virtio_transport_common.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
Ouch, this would be nice as a separate patch that can be merged right
away (with s/virtio_vsock_buf/virtio_vsock_pkt/).
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 488 bytes
Desc: not available
URL:
<http://lists.linuxfoundation.org/pipermail/virtualization/attachments/20190516/70384c84/attachment.sig>

On Mon, May 13, 2019 at 05:33:40PM +0800, Jason Wang
wrote:
> 
> On 2019/5/10 ??8:58, Stefano Garzarella wrote:
> > While I was testing this new series (v2) I discovered an huge use of
memory
> > and a memory leak in the virtio-vsock driver in the guest when I sent
> > 1-byte packets to the guest.
> > 
> > These issues are present since the introduction of the virtio-vsock
> > driver. I added the patches 1 and 2 to fix them in this series in
order
> > to better track the performance trends.
> > 
> > v1: https://patchwork.kernel.org/cover/10885431/
> > 
> > v2:
> > - Add patch 1 to limit the memory usage
> > - Add patch 2 to avoid memory leak during the socket release
> > - Add patch 3 to fix locking of fwd_cnt and buf_alloc
> > - Patch 4: fix 'free_space' type (u32 instead of s64) [Stefan]
> > - Patch 5: Avoid integer underflow of iov_len [Stefan]
> > - Patch 5: Fix packet capture in order to see the exact packets that
are
> >             delivered. [Stefan]
> > - Add patch 8 to make the RX buffer size tunable [Stefan]
> > 
> > Below are the benchmarks step by step. I used iperf3 [1] modified with
VSOCK
> > support.
> > As Micheal suggested in the v1, I booted host and guest with
'nosmap', and I
> > added a column with virtio-net+vhost-net performance.
> > 
> > A brief description of patches:
> > - Patches 1+2: limit the memory usage with an extra copy and avoid
memory leak
> > - Patches 3+4: fix locking and reduce the number of credit update
messages sent
> >                 to the transmitter
> > - Patches 5+6: allow the host to split packets on multiple buffers and
use
> >                 VIRTIO_VSOCK_MAX_PKT_BUF_SIZE as the max packet size
allowed
> > - Patches 7+8: increase RX buffer size to 64 KiB
> > 
> >                      host -> guest [Gbps]
> > pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net
+ vhost
> >                                                                      
TCP_NODELAY
> > 64         0.068     0.063    0.130    0.131    0.128         0.188   
0.187
> > 256        0.274     0.236    0.392    0.338    0.282         0.749   
0.654
> > 512        0.531     0.457    0.862    0.725    0.602         1.419   
1.414
> > 1K         0.954     0.827    1.591    1.598    1.548         2.599   
2.640
> > 2K         1.783     1.543    3.731    3.637    3.469         4.530   
4.754
> > 4K         3.332     3.436    7.164    7.124    6.494         7.738   
7.696
> > 8K         5.792     5.530   11.653   11.787   11.444        12.307   
11.850
> > 16K        8.405     8.462   16.372   16.855   17.562        16.936   
16.954
> > 32K       14.208    13.669   18.945   20.009   23.128        21.980   
23.015
> > 64K       21.082    18.893   20.266   20.903   30.622        27.290   
27.383
> > 128K      20.696    20.148   20.112   21.746   32.152        30.446   
30.990
> > 256K      20.801    20.589   20.725   22.685   34.721        33.151   
32.745
> > 512K      21.220    20.465   20.432   22.106   34.496        36.847   
31.096
> > 
> >                      guest -> host [Gbps]
> > pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net
+ vhost
> >                                                                      
TCP_NODELAY
> > 64         0.089     0.091    0.120    0.115    0.117         0.274   
0.272
> > 256        0.352     0.354    0.452    0.445    0.451         1.085   
1.136
> > 512        0.705     0.704    0.893    0.858    0.898         2.131   
1.882
> > 1K         1.394     1.433    1.721    1.669    1.691         3.984   
3.576
> > 2K         2.818     2.874    3.316    3.249    3.303         6.719   
6.359
> > 4K         5.293     5.397    6.129    5.933    6.082        10.105   
9.860
> > 8K         8.890     9.151   10.990   10.545   10.519        15.239   
14.868
> > 16K       11.444    11.018   12.074   15.255   15.577        20.551   
20.848
> > 32K       11.229    10.875   10.857   24.401   25.227        26.294   
26.380
> > 64K       10.832    10.545   10.816   39.487   39.616        34.996   
32.041
> > 128K      10.435    10.241   10.500   39.813   40.012        38.379   
35.055
> > 256K      10.263     9.866    9.845   34.971   35.143        36.559   
37.232
> > 512K      10.224    10.060   10.092   35.469   34.627        34.963   
33.401
> > 
> > As Stefan suggested in the v1, this time I measured also the
efficiency in this
> > way:
> >      efficiency = Mbps / (%CPU_Host + %CPU_Guest)
> > 
> > The '%CPU_Guest' is taken inside the VM. I know that it is not
the best way,
> > but it's provided for free from iperf3 and could be an indication.
> > 
> >          host -> guest efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
> > pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net
+ vhost
> >                                                                      
TCP_NODELAY
> > 64          0.94      0.59     3.96     4.06     4.09          2.82   
2.11
> > 256         2.62      2.50     6.45     6.09     5.81          9.64   
8.73
> > 512         5.16      4.87    13.16    12.39    11.67         17.83   
17.76
> > 1K          9.16      8.85    24.98    24.97    25.01         32.57   
32.04
> > 2K         17.41     17.03    49.09    48.59    49.22         55.31   
57.14
> > 4K         32.99     33.62    90.80    90.98    91.72         91.79   
91.40
> > 8K         58.51     59.98   153.53   170.83   167.31        137.51   
132.85
> > 16K        89.32     95.29   216.98   264.18   260.95        176.05   
176.05
> > 32K       152.94    167.10   285.75   387.02   360.81        215.49   
226.30
> > 64K       250.38    307.20   317.65   489.53   472.70        238.97   
244.27
> > 128K      327.99    335.24   335.76   523.71   486.41        253.29   
260.86
> > 256K      327.06    334.24   338.64   533.76   509.85        267.78   
266.22
> > 512K      337.36    330.61   334.95   512.90   496.35        280.42   
241.43
> > 
> >          guest -> host efficiency [Mbps / (%CPU_Host + %CPU_Guest)]
> > pkt_size before opt  p 1+2    p 3+4    p 5+6    p 7+8       virtio-net
+ vhost
> >                                                                      
TCP_NODELAY
> > 64          0.90      0.91     1.37     1.32     1.35          2.15   
2.13
> > 256         3.59      3.55     5.23     5.19     5.29          8.50   
8.89
> > 512         7.19      7.08    10.21     9.95    10.38         16.74   
14.71
> > 1K         14.15     14.34    19.85    19.06    19.33         31.44   
28.11
> > 2K         28.44     29.09    37.78    37.18    37.49         53.07   
50.63
> > 4K         55.37     57.60    71.02    69.27    70.97         81.56   
79.32
> > 8K        105.58    100.45   111.95   124.68   123.61        120.85   
118.66
> > 16K       141.63    138.24   137.67   187.41   190.20        160.43   
163.00
> > 32K       147.56    143.09   138.48   296.41   301.04        214.64   
223.94
> > 64K       144.81    143.27   138.49   433.98   462.26        298.86   
269.71
> > 128K      150.14    147.99   146.85   511.36   514.29        350.17   
298.09
> > 256K      156.69    152.25   148.69   542.19   549.97        326.42   
333.32
> > 512K      157.29    153.35   152.22   546.52   533.24        315.55   
302.27
> > 
> > [1] https://github.com/stefano-garzarella/iperf/
> 
> 
> Hi:
> 
> Do you have any explanation that vsock is better here? Is this because of
> the mergeable buffer? If you, we need test with mrg_rxbuf=off.
> 
Hi Jason,
I tried to disable the mergeable buffer but I had even worst performance
with virtio-net.

Do you think the differences could be related to the TCP/IP stack?

Thanks,
Stefano