virtio-crypto: wait ctrl queue instead of busy polling

Originally, after submitting request into virtio crypto control
queue, the guest side polls the result from the virt queue. This
works like following:
    CPU0   CPU1               ...             CPUx  CPUy
     |      |                                  |     |
     \      \                                  /     /
      \--------spin_lock(&vcrypto->ctrl_lock)-------/
                           |
                 virtqueue add & kick
                           |
                  busy poll virtqueue
                           |
              spin_unlock(&vcrypto->ctrl_lock)
                          ...

There are two problems:
1, The queue depth is always 1, the performance of a virtio crypto
   device gets limited. Multi user processes share a single control
   queue, and hit spin lock race from control queue. Test on Intel
   Platinum 8260, a single worker gets ~35K/s create/close session
   operations, and 8 workers get ~40K/s operations with 800% CPU
   utilization.
2, The control request is supposed to get handled immediately, but
   in the current implementation of QEMU(v6.2), the vCPU thread kicks
   another thread to do this work, the latency also gets unstable.
   Tracking latency of virtio_crypto_alg_akcipher_close_session in 5s:
        usecs               : count     distribution
         0 -> 1          : 0        |                        |
         2 -> 3          : 7        |                        |
         4 -> 7          : 72       |                        |
         8 -> 15         : 186485   |************************|
        16 -> 31         : 687      |                        |
        32 -> 63         : 5        |                        |
        64 -> 127        : 3        |                        |
       128 -> 255        : 1        |                        |
       256 -> 511        : 0        |                        |
       512 -> 1023       : 0        |                        |
      1024 -> 2047       : 0        |                        |
      2048 -> 4095       : 0        |                        |
      4096 -> 8191       : 0        |                        |
      8192 -> 16383      : 2        |                        |
This means that a CPU may hold vcrypto->ctrl_lock as long as 8192~16383us.

To improve the performance of control queue, a request on control queue
waits completion instead of busy polling to reduce lock racing, and gets
completed by control queue callback.
    CPU0   CPU1               ...             CPUx  CPUy
     |      |                                  |     |
     \      \                                  /     /
      \--------spin_lock(&vcrypto->ctrl_lock)-------/
                           |
                 virtqueue add & kick
                           |
      ---------spin_unlock(&vcrypto->ctrl_lock)------
     /      /                                  \     \
     |      |                                  |     |
    wait   wait                               wait  wait

Test this patch, the guest side get ~200K/s operations with 300% CPU
utilization.

Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Jason Wang <jasowang@redhat.com>
Cc: Gonglei <arei.gonglei@huawei.com>
Reviewed-by: Gonglei <arei.gonglei@huawei.com>
Signed-off-by: zhenwei pi <pizhenwei@bytedance.com>
Message-Id: <20220506131627.180784-4-pizhenwei@bytedance.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>

1 files changed, 51 insertions, 1 deletions

diff --git a/drivers/crypto/virtio/virtio_crypto_core.c b/drivers/crypto/virtio/virtio_crypto_core.c
index c6f482db0bc0..60490ffa3df1 100644
--- a/drivers/crypto/virtio/virtio_crypto_core.c
+++ b/drivers/crypto/virtio/virtio_crypto_core.c
@@ -22,6 +22,56 @@ virtcrypto_clear_request(struct virtio_crypto_request *vc_req)
 	}
 }
 
+static void virtio_crypto_ctrlq_callback(struct virtio_crypto_ctrl_request *vc_ctrl_req)
+{
+	complete(&vc_ctrl_req->compl);
+}
+
+static void virtcrypto_ctrlq_callback(struct virtqueue *vq)
+{
+	struct virtio_crypto *vcrypto = vq->vdev->priv;
+	struct virtio_crypto_ctrl_request *vc_ctrl_req;
+	unsigned long flags;
+	unsigned int len;
+
+	spin_lock_irqsave(&vcrypto->ctrl_lock, flags);
+	do {
+		virtqueue_disable_cb(vq);
+		while ((vc_ctrl_req = virtqueue_get_buf(vq, &len)) != NULL) {
+			spin_unlock_irqrestore(&vcrypto->ctrl_lock, flags);
+			virtio_crypto_ctrlq_callback(vc_ctrl_req);
+			spin_lock_irqsave(&vcrypto->ctrl_lock, flags);
+		}
+		if (unlikely(virtqueue_is_broken(vq)))
+			break;
+	} while (!virtqueue_enable_cb(vq));
+	spin_unlock_irqrestore(&vcrypto->ctrl_lock, flags);
+}
+
+int virtio_crypto_ctrl_vq_request(struct virtio_crypto *vcrypto, struct scatterlist *sgs[],
+		unsigned int out_sgs, unsigned int in_sgs,
+		struct virtio_crypto_ctrl_request *vc_ctrl_req)
+{
+	int err;
+	unsigned long flags;
+
+	init_completion(&vc_ctrl_req->compl);
+
+	spin_lock_irqsave(&vcrypto->ctrl_lock, flags);
+	err = virtqueue_add_sgs(vcrypto->ctrl_vq, sgs, out_sgs, in_sgs, vc_ctrl_req, GFP_ATOMIC);
+	if (err < 0) {
+		spin_unlock_irqrestore(&vcrypto->ctrl_lock, flags);
+		return err;
+	}
+
+	virtqueue_kick(vcrypto->ctrl_vq);
+	spin_unlock_irqrestore(&vcrypto->ctrl_lock, flags);
+
+	wait_for_completion(&vc_ctrl_req->compl);
+
+	return 0;
+}
+
 static void virtcrypto_dataq_callback(struct virtqueue *vq)
 {
 	struct virtio_crypto *vcrypto = vq->vdev->priv;
@@ -73,7 +123,7 @@ static int virtcrypto_find_vqs(struct virtio_crypto *vi)
 		goto err_names;
 
 	/* Parameters for control virtqueue */
-	callbacks[total_vqs - 1] = NULL;
+	callbacks[total_vqs - 1] = virtcrypto_ctrlq_callback;
 	names[total_vqs - 1] = "controlq";
 
 	/* Allocate/initialize parameters for data virtqueues */