macvtap: rework object lifetime rules

This reworks the change done by the previous patch
in a more complete way.

The original macvtap code has a number of problems
resulting from the use of RCU for protecting the
access to struct macvtap_queue from open files.

This includes
- need for GFP_ATOMIC allocations for skbs
- potential deadlocks when copy_*_user sleeps
- inability to work with vhost-net

Changing the lifetime of macvtap_queue to always
depend on the open file solves all these. The
RCU reference simply moves one step down to
the reference on the macvlan_dev, which we
only need for nonblocking operations.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Sridhar Samudrala <sri@us.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 91 insertions, 92 deletions

diff --git a/drivers/net/macvtap.c b/drivers/net/macvtap.c
index fe7656bf68c6..705099749766 100644
--- a/drivers/net/macvtap.c
+++ b/drivers/net/macvtap.c
@@ -60,30 +60,19 @@ static struct cdev macvtap_cdev;
 
 /*
  * RCU usage:
- * The macvtap_queue is referenced both from the chardev struct file
- * and from the struct macvlan_dev using rcu_read_lock.
+ * The macvtap_queue and the macvlan_dev are loosely coupled, the
+ * pointers from one to the other can only be read while rcu_read_lock
+ * or macvtap_lock is held.
  *
- * We never actually update the contents of a macvtap_queue atomically
- * with RCU but it is used for race-free destruction of a queue when
- * either the file or the macvlan_dev goes away. Pointers back to
- * the dev and the file are implicitly valid as long as the queue
- * exists.
+ * Both the file and the macvlan_dev hold a reference on the macvtap_queue
+ * through sock_hold(&q->sk). When the macvlan_dev goes away first,
+ * q->vlan becomes inaccessible. When the files gets closed,
+ * macvtap_get_queue() fails.
  *
- * The callbacks from macvlan are always done with rcu_read_lock held
- * already. For calls from file_operations, we use the rcu_read_lock_bh
- * to get a reference count on the socket and the device.
- *
- * When destroying a queue, we remove the pointers from the file and
- * from the dev and then synchronize_rcu to make sure no thread is
- * still using the queue. There may still be references to the struct
- * sock inside of the queue from outbound SKBs, but these never
- * reference back to the file or the dev. The data structure is freed
- * through __sk_free when both our references and any pending SKBs
- * are gone.
- *
- * macvtap_lock is only used to prevent multiple concurrent open()
- * calls to assign a new vlan->tap pointer. It could be moved into
- * the macvlan_dev itself but is extremely rarely used.
+ * There may still be references to the struct sock inside of the
+ * queue from outbound SKBs, but these never reference back to the
+ * file or the dev. The data structure is freed through __sk_free
+ * when both our references and any pending SKBs are gone.
  */
 static DEFINE_SPINLOCK(macvtap_lock);
 
@@ -101,11 +90,12 @@ static int macvtap_set_queue(struct net_device *dev, struct file *file,
                 goto out;
 
         err = 0;
-        q->vlan = vlan;
+        rcu_assign_pointer(q->vlan, vlan);
         rcu_assign_pointer(vlan->tap, q);
+        sock_hold(&q->sk);
 
         q->file = file;
-        rcu_assign_pointer(file->private_data, q);
+        file->private_data = q;
 
 out:
         spin_unlock(&macvtap_lock);
@@ -113,28 +103,25 @@ out:
 }
 
 /*
- * We must destroy each queue exactly once, when either
- * the netdev or the file go away.
+ * The file owning the queue got closed, give up both
+ * the reference that the files holds as well as the
+ * one from the macvlan_dev if that still exists.
  *
  * Using the spinlock makes sure that we don't get
  * to the queue again after destroying it.
- *
- * synchronize_rcu serializes with the packet flow
- * that uses rcu_read_lock.
  */
-static void macvtap_del_queue(struct macvtap_queue **qp)
+static void macvtap_put_queue(struct macvtap_queue *q)
 {
-        struct macvtap_queue *q;
+        struct macvlan_dev *vlan;
 
         spin_lock(&macvtap_lock);
-        q = rcu_dereference(*qp);
-        if (!q) {
-                spin_unlock(&macvtap_lock);
-                return;
+        vlan = rcu_dereference(q->vlan);
+        if (vlan) {
+                rcu_assign_pointer(vlan->tap, NULL);
+                rcu_assign_pointer(q->vlan, NULL);
+                sock_put(&q->sk);
         }
 
-        rcu_assign_pointer(q->vlan->tap, NULL);
-        rcu_assign_pointer(q->file->private_data, NULL);
         spin_unlock(&macvtap_lock);
 
         synchronize_rcu();
@@ -152,29 +139,29 @@ static struct macvtap_queue *macvtap_get_queue(struct net_device *dev,
         return rcu_dereference(vlan->tap);
 }
 
+/*
+ * The net_device is going away, give up the reference
+ * that it holds on the queue (all the queues one day)
+ * and safely set the pointer from the queues to NULL.
+ */
 static void macvtap_del_queues(struct net_device *dev)
 {
         struct macvlan_dev *vlan = netdev_priv(dev);
-        macvtap_del_queue(&vlan->tap);
-}
-
-static inline struct macvtap_queue *macvtap_file_get_queue(struct file *file)
-{
         struct macvtap_queue *q;
-        rcu_read_lock_bh();
-        q = rcu_dereference(file->private_data);
-        if (q) {
-                sock_hold(&q->sk);
-                dev_hold(q->vlan->dev);
+
+        spin_lock(&macvtap_lock);
+        q = rcu_dereference(vlan->tap);
+        if (!q) {
+                spin_unlock(&macvtap_lock);
+                return;
         }
-        rcu_read_unlock_bh();
-        return q;
-}
 
-static inline void macvtap_file_put_queue(struct macvtap_queue *q)
-{
+        rcu_assign_pointer(vlan->tap, NULL);
+        rcu_assign_pointer(q->vlan, NULL);
+        spin_unlock(&macvtap_lock);
+
+        synchronize_rcu();
         sock_put(&q->sk);
-        dev_put(q->vlan->dev);
 }
 
 /*
@@ -284,7 +271,6 @@ static int macvtap_open(struct inode *inode, struct file *file)
         q->sock.type = SOCK_RAW;
         q->sock.state = SS_CONNECTED;
         sock_init_data(&q->sock, &q->sk);
-        q->sk.sk_allocation = GFP_ATOMIC; /* for now */
         q->sk.sk_write_space = macvtap_sock_write_space;
 
         err = macvtap_set_queue(dev, file, q);
@@ -300,13 +286,14 @@ out:
 
 static int macvtap_release(struct inode *inode, struct file *file)
 {
-        macvtap_del_queue((struct macvtap_queue **)&file->private_data);
+        struct macvtap_queue *q = file->private_data;
+        macvtap_put_queue(q);
         return 0;
 }
 
 static unsigned int macvtap_poll(struct file *file, poll_table * wait)
 {
-        struct macvtap_queue *q = macvtap_file_get_queue(file);
+        struct macvtap_queue *q = file->private_data;
         unsigned int mask = POLLERR;
 
         if (!q)
@@ -323,7 +310,6 @@ static unsigned int macvtap_poll(struct file *file, poll_table * wait)
              sock_writeable(&q->sk)))
                 mask |= POLLOUT | POLLWRNORM;
 
-        macvtap_file_put_queue(q);
 out:
         return mask;
 }
@@ -334,6 +320,7 @@ static ssize_t macvtap_get_user(struct macvtap_queue *q,
                                 int noblock)