tcm_vhost: Use vq->private_data to indicate if the endpoint is setup
authorAsias He <asias@redhat.com>
Wed, 3 Apr 2013 06:17:37 +0000 (14:17 +0800)
committerNicholas Bellinger <nab@linux-iscsi.org>
Mon, 8 Apr 2013 21:07:00 +0000 (14:07 -0700)
Currently, vs->vs_endpoint is used indicate if the endpoint is setup or
not. It is set or cleared in vhost_scsi_set_endpoint() or
vhost_scsi_clear_endpoint() under the vs->dev.mutex lock. However, when
we check it in vhost_scsi_handle_vq(), we ignored the lock.

Instead of using the vs->vs_endpoint and the vs->dev.mutex lock to
indicate the status of the endpoint, we use per virtqueue
vq->private_data to indicate it. In this way, we can only take the
vq->mutex lock which is per queue and make the concurrent multiqueue
process having less lock contention. Further, in the read side of
vq->private_data, we can even do not take the lock if it is accessed in
the vhost worker thread, because it is protected by "vhost rcu".

(nab: Do s/VHOST_FEATURES/~VHOST_SCSI_FEATURES)

Signed-off-by: Asias He <asias@redhat.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
drivers/vhost/tcm_vhost.c

index dd9614e..6cda137 100644 (file)
@@ -74,9 +74,8 @@ enum {
 
 struct vhost_scsi {
        /* Protected by vhost_scsi->dev.mutex */
-       struct tcm_vhost_tpg *vs_tpg[VHOST_SCSI_MAX_TARGET];
+       struct tcm_vhost_tpg **vs_tpg;
        char vs_vhost_wwpn[TRANSPORT_IQN_LEN];
-       bool vs_endpoint;
 
        struct vhost_dev dev;
        struct vhost_virtqueue vqs[VHOST_SCSI_MAX_VQ];
@@ -582,6 +581,7 @@ static void tcm_vhost_submission_work(struct work_struct *work)
 static void vhost_scsi_handle_vq(struct vhost_scsi *vs,
        struct vhost_virtqueue *vq)
 {
+       struct tcm_vhost_tpg **vs_tpg;
        struct virtio_scsi_cmd_req v_req;
        struct tcm_vhost_tpg *tv_tpg;
        struct tcm_vhost_cmd *tv_cmd;
@@ -590,8 +590,16 @@ static void vhost_scsi_handle_vq(struct vhost_scsi *vs,
        int head, ret;
        u8 target;
 
-       /* Must use ioctl VHOST_SCSI_SET_ENDPOINT */
-       if (unlikely(!vs->vs_endpoint))
+       /*
+        * We can handle the vq only after the endpoint is setup by calling the
+        * VHOST_SCSI_SET_ENDPOINT ioctl.
+        *
+        * TODO: Check that we are running from vhost_worker which acts
+        * as read-side critical section for vhost kind of RCU.
+        * See the comments in struct vhost_virtqueue in drivers/vhost/vhost.h
+        */
+       vs_tpg = rcu_dereference_check(vq->private_data, 1);
+       if (!vs_tpg)
                return;
 
        mutex_lock(&vq->mutex);
@@ -661,7 +669,7 @@ static void vhost_scsi_handle_vq(struct vhost_scsi *vs,
 
                /* Extract the tpgt */
                target = v_req.lun[1];
-               tv_tpg = ACCESS_ONCE(vs->vs_tpg[target]);
+               tv_tpg = ACCESS_ONCE(vs_tpg[target]);
 
                /* Target does not exist, fail the request */
                if (unlikely(!tv_tpg)) {
@@ -780,6 +788,20 @@ static void vhost_scsi_handle_kick(struct vhost_work *work)
        vhost_scsi_handle_vq(vs, vq);
 }
 
+static void vhost_scsi_flush_vq(struct vhost_scsi *vs, int index)
+{
+       vhost_poll_flush(&vs->dev.vqs[index].poll);
+}
+
+static void vhost_scsi_flush(struct vhost_scsi *vs)
+{
+       int i;
+
+       for (i = 0; i < VHOST_SCSI_MAX_VQ; i++)
+               vhost_scsi_flush_vq(vs, i);
+       vhost_work_flush(&vs->dev, &vs->vs_completion_work);
+}
+
 /*
  * Called from vhost_scsi_ioctl() context to walk the list of available
  * tcm_vhost_tpg with an active struct tcm_vhost_nexus
@@ -790,8 +812,10 @@ static int vhost_scsi_set_endpoint(
 {
        struct tcm_vhost_tport *tv_tport;
        struct tcm_vhost_tpg *tv_tpg;
+       struct tcm_vhost_tpg **vs_tpg;
+       struct vhost_virtqueue *vq;
+       int index, ret, i, len;
        bool match = false;
-       int index, ret;
 
        mutex_lock(&vs->dev.mutex);
        /* Verify that ring has been setup correctly. */
@@ -803,6 +827,15 @@ static int vhost_scsi_set_endpoint(
                }
        }
 
+       len = sizeof(vs_tpg[0]) * VHOST_SCSI_MAX_TARGET;
+       vs_tpg = kzalloc(len, GFP_KERNEL);
+       if (!vs_tpg) {
+               mutex_unlock(&vs->dev.mutex);
+               return -ENOMEM;
+       }
+       if (vs->vs_tpg)
+               memcpy(vs_tpg, vs->vs_tpg, len);
+
        mutex_lock(&tcm_vhost_mutex);
        list_for_each_entry(tv_tpg, &tcm_vhost_list, tv_tpg_list) {
                mutex_lock(&tv_tpg->tv_tpg_mutex);
@@ -817,14 +850,15 @@ static int vhost_scsi_set_endpoint(
                tv_tport = tv_tpg->tport;
 
                if (!strcmp(tv_tport->tport_name, t->vhost_wwpn)) {
-                       if (vs->vs_tpg[tv_tpg->tport_tpgt]) {
+                       if (vs->vs_tpg && vs->vs_tpg[tv_tpg->tport_tpgt]) {
                                mutex_unlock(&tv_tpg->tv_tpg_mutex);
                                mutex_unlock(&tcm_vhost_mutex);
                                mutex_unlock(&vs->dev.mutex);
+                               kfree(vs_tpg);
                                return -EEXIST;
                        }
                        tv_tpg->tv_tpg_vhost_count++;
-                       vs->vs_tpg[tv_tpg->tport_tpgt] = tv_tpg;
+                       vs_tpg[tv_tpg->tport_tpgt] = tv_tpg;
                        smp_mb__after_atomic_inc();
                        match = true;
                }
@@ -835,12 +869,26 @@ static int vhost_scsi_set_endpoint(
        if (match) {
                memcpy(vs->vs_vhost_wwpn, t->vhost_wwpn,
                       sizeof(vs->vs_vhost_wwpn));
-               vs->vs_endpoint = true;
+               for (i = 0; i < VHOST_SCSI_MAX_VQ; i++) {
+                       vq = &vs->vqs[i];
+                       /* Flushing the vhost_work acts as synchronize_rcu */
+                       mutex_lock(&vq->mutex);
+                       rcu_assign_pointer(vq->private_data, vs_tpg);
+                       mutex_unlock(&vq->mutex);
+               }
                ret = 0;
        } else {
                ret = -EEXIST;
        }
 
+       /*
+        * Act as synchronize_rcu to make sure access to
+        * old vs->vs_tpg is finished.
+        */
+       vhost_scsi_flush(vs);
+       kfree(vs->vs_tpg);
+       vs->vs_tpg = vs_tpg;
+
        mutex_unlock(&vs->dev.mutex);
        return ret;
 }
@@ -851,6 +899,8 @@ static int vhost_scsi_clear_endpoint(
 {
        struct tcm_vhost_tport *tv_tport;
        struct tcm_vhost_tpg *tv_tpg;
+       struct vhost_virtqueue *vq;
+       bool match = false;
        int index, ret, i;
        u8 target;
 
@@ -862,9 +912,14 @@ static int vhost_scsi_clear_endpoint(
                        goto err_dev;
                }
        }
+
+       if (!vs->vs_tpg) {
+               mutex_unlock(&vs->dev.mutex);
+               return 0;
+       }
+
        for (i = 0; i < VHOST_SCSI_MAX_TARGET; i++) {
                target = i;
-
                tv_tpg = vs->vs_tpg[target];
                if (!tv_tpg)
                        continue;
@@ -886,10 +941,27 @@ static int vhost_scsi_clear_endpoint(
                }
                tv_tpg->tv_tpg_vhost_count--;
                vs->vs_tpg[target] = NULL;
-               vs->vs_endpoint = false;
+               match = true;
                mutex_unlock(&tv_tpg->tv_tpg_mutex);
        }
+       if (match) {
+               for (i = 0; i < VHOST_SCSI_MAX_VQ; i++) {
+                       vq = &vs->vqs[i];
+                       /* Flushing the vhost_work acts as synchronize_rcu */
+                       mutex_lock(&vq->mutex);
+                       rcu_assign_pointer(vq->private_data, NULL);
+                       mutex_unlock(&vq->mutex);
+               }
+       }
+       /*
+        * Act as synchronize_rcu to make sure access to
+        * old vs->vs_tpg is finished.
+        */
+       vhost_scsi_flush(vs);
+       kfree(vs->vs_tpg);
+       vs->vs_tpg = NULL;
        mutex_unlock(&vs->dev.mutex);
+
        return 0;
 
 err_tpg:
@@ -899,6 +971,24 @@ err_dev:
        return ret;
 }
 
+static int vhost_scsi_set_features(struct vhost_scsi *vs, u64 features)
+{
+       if (features & ~VHOST_SCSI_FEATURES)
+               return -EOPNOTSUPP;
+
+       mutex_lock(&vs->dev.mutex);
+       if ((features & (1 << VHOST_F_LOG_ALL)) &&
+           !vhost_log_access_ok(&vs->dev)) {
+               mutex_unlock(&vs->dev.mutex);
+               return -EFAULT;
+       }
+       vs->dev.acked_features = features;
+       smp_wmb();
+       vhost_scsi_flush(vs);
+       mutex_unlock(&vs->dev.mutex);
+       return 0;
+}
+
 static int vhost_scsi_open(struct inode *inode, struct file *f)
 {
        struct vhost_scsi *s;
@@ -939,38 +1029,6 @@ static int vhost_scsi_release(struct inode *inode, struct file *f)
        return 0;
 }
 
-static void vhost_scsi_flush_vq(struct vhost_scsi *vs, int index)
-{
-       vhost_poll_flush(&vs->dev.vqs[index].poll);
-}
-
-static void vhost_scsi_flush(struct vhost_scsi *vs)
-{
-       int i;
-
-       for (i = 0; i < VHOST_SCSI_MAX_VQ; i++)
-               vhost_scsi_flush_vq(vs, i);
-       vhost_work_flush(&vs->dev, &vs->vs_completion_work);
-}
-
-static int vhost_scsi_set_features(struct vhost_scsi *vs, u64 features)
-{
-       if (features & ~VHOST_SCSI_FEATURES)
-               return -EOPNOTSUPP;
-
-       mutex_lock(&vs->dev.mutex);
-       if ((features & (1 << VHOST_F_LOG_ALL)) &&
-           !vhost_log_access_ok(&vs->dev)) {
-               mutex_unlock(&vs->dev.mutex);
-               return -EFAULT;
-       }
-       vs->dev.acked_features = features;
-       smp_wmb();
-       vhost_scsi_flush(vs);
-       mutex_unlock(&vs->dev.mutex);
-       return 0;
-}
-
 static long vhost_scsi_ioctl(struct file *f, unsigned int ioctl,
                                unsigned long arg)
 {