On Tue, Mar 21, 2023 at 06:03:14PM +0300, Arseniy Krasnov wrote:
This adds small optimization for tx path: instead of allocating single skbuff on every call to transport, allocate multiple skbuff's until credit space allows, thus trying to send as much as possible data without return to af_vsock.c. Signed-off-by: Arseniy Krasnov <AVKrasnov@xxxxxxxxxxxxxx> --- Link to v1: https://lore.kernel.org/netdev/2c52aa26-8181-d37a-bccd-a86bd3cbc6e1@xxxxxxxxxxxxxx/ Link to v2: https://lore.kernel.org/netdev/ea5725eb-6cb5-cf15-2938-34e335a442fa@xxxxxxxxxxxxxx/ Link to v3: https://lore.kernel.org/netdev/f33ef593-982e-2b3f-0986-6d537a3aaf08@xxxxxxxxxxxxxx/ Changelog: v1 -> v2: - If sent something, return number of bytes sent (even in case of error). Return error only if failed to sent first skbuff. v2 -> v3: - Handle case when transport callback returns unexpected value which is not equal to 'skb->len'. Break loop. - Don't check for zero value of 'rest_len' before calling 'virtio_transport_put_credit()'. Decided to add this check directly to 'virtio_transport_put_credit()' in separate patch. v3 -> v4: - Use WARN_ONCE() to handle case when transport callback returns unexpected value. - Remove useless 'ret = -EFAULT;' assignment for case above. net/vmw_vsock/virtio_transport_common.c | 59 +++++++++++++++++++------ 1 file changed, 45 insertions(+), 14 deletions(-) diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c index 6564192e7f20..a300f25749ea 100644 --- a/net/vmw_vsock/virtio_transport_common.c +++ b/net/vmw_vsock/virtio_transport_common.c @@ -196,7 +196,8 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk, const struct virtio_transport *t_ops; struct virtio_vsock_sock *vvs; u32 pkt_len = info->pkt_len; - struct sk_buff *skb; + u32 rest_len; + int ret; info->type = virtio_transport_get_type(sk_vsock(vsk)); @@ -216,10 +217,6 @@ 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_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); @@ -227,17 +224,51 @@ static int virtio_transport_send_pkt_info(struct vsock_sock *vsk, if (pkt_len == 0 && info->op == VIRTIO_VSOCK_OP_RW) return pkt_len; - skb = virtio_transport_alloc_skb(info, pkt_len, - src_cid, src_port, - dst_cid, dst_port); - if (!skb) { - virtio_transport_put_credit(vvs, pkt_len); - return -ENOMEM; - } + ret = 0;
nit: this initialization seems superfluous since `ret` is overwritten later ...
+ rest_len = pkt_len; + + do { + struct sk_buff *skb; + size_t skb_len; + + skb_len = min_t(u32, VIRTIO_VSOCK_MAX_PKT_BUF_SIZE, rest_len); + + skb = virtio_transport_alloc_skb(info, skb_len, + src_cid, src_port, + dst_cid, dst_port); + if (!skb) { + ret = -ENOMEM; + break; + } - virtio_transport_inc_tx_pkt(vvs, skb); + virtio_transport_inc_tx_pkt(vvs, skb); - return t_ops->send_pkt(skb); + ret = t_ops->send_pkt(skb);
... here.
+
nit: we can remove this extra line
+ if (ret < 0) + break; + + /* Both virtio and vhost 'send_pkt()' returns 'skb_len', + * but for reliability use 'ret' instead of 'skb_len'. + * Also if partial send happens (e.g. 'ret' != 'skb_len') + * somehow, we break this loop, but account such returned + * value in 'virtio_transport_put_credit()'. + */ + rest_len -= ret; + + if (WARN_ONCE(ret != skb_len, + "'send_pkt()' returns %i, but %zu expected\n", + ret, skb_len)) + break; + } while (rest_len); + + virtio_transport_put_credit(vvs, rest_len); + + /* Return number of bytes, if any data has been sent. */ + if (rest_len != pkt_len) + ret = pkt_len - rest_len; + + return ret; } static bool virtio_transport_inc_rx_pkt(struct virtio_vsock_sock *vvs, -- 2.25.1
The patch LGTM: Reviewed-by: Stefano Garzarella <sgarzare@xxxxxxxxxx> Anyway, feel free to include in the same series or as separate patch also the changes to avoid useless lock in virtio_transport_put_credit() and virtio_transport_get_credit(). I would include it in this series, because before these changes, we used to call virtio_transport_put_credit() only in the error path, while now we always call it, even when rest_len is 0. Thanks, Stefano