Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/gi…

…t/rdma/rdma

Pull rdma fixes from Jason Gunthorpe:
 "Nothing profound here, just a last set of long standing bug fixes:

   - Incorrect error unwind in qib and pvrdma

   - User triggerable NULL pointer crash in mlx5 with ODP prefetch

   - syzkaller RCU race in uverbs

   - Rare double free crash in ipoib"

* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
  IB/ipoib: Fix double free of skb in case of multicast traffic in CM mode
  RDMA/core: Fix double destruction of uobject
  RDMA/pvrdma: Fix missing pci disable in pvrdma_pci_probe()
  RDMA/mlx5: Fix NULL pointer dereference in destroy_prefetch_work
  IB/qib: Call kobject_put() when kobject_init_and_add() fails

drivers/infiniband/core/rdma_core.c

@@ -153,9 +153,9 @@ static int uverbs_destroy_uobject(struct ib_uobject *uobj,
 	uobj->context = NULL;
 
 	/*
-	 * For DESTROY the usecnt is held write locked, the caller is expected
-	 * to put it unlock and put the object when done with it. Only DESTROY
-	 * can remove the IDR handle.
+	 * For DESTROY the usecnt is not changed, the caller is expected to
+	 * manage it via uobj_put_destroy(). Only DESTROY can remove the IDR
+	 * handle.
 	 */
 	if (reason != RDMA_REMOVE_DESTROY)
 		atomic_set(&uobj->usecnt, 0);
@@ -187,7 +187,7 @@ static int uverbs_destroy_uobject(struct ib_uobject *uobj,
 /*
  * This calls uverbs_destroy_uobject() using the RDMA_REMOVE_DESTROY
  * sequence. It should only be used from command callbacks. On success the
- * caller must pair this with rdma_lookup_put_uobject(LOOKUP_WRITE). This
+ * caller must pair this with uobj_put_destroy(). This
  * version requires the caller to have already obtained an
  * LOOKUP_DESTROY uobject kref.
  */
@@ -198,6 +198,13 @@ int uobj_destroy(struct ib_uobject *uobj, struct uverbs_attr_bundle *attrs)
 
 	down_read(&ufile->hw_destroy_rwsem);
 
+	/*
+	 * Once the uobject is destroyed by RDMA_REMOVE_DESTROY then it is left
+	 * write locked as the callers put it back with UVERBS_LOOKUP_DESTROY.
+	 * This is because any other concurrent thread can still see the object
+	 * in the xarray due to RCU. Leaving it locked ensures nothing else will
+	 * touch it.
+	 */
 	ret = uverbs_try_lock_object(uobj, UVERBS_LOOKUP_WRITE);
 	if (ret)
 		goto out_unlock;
@@ -216,7 +223,7 @@ int uobj_destroy(struct ib_uobject *uobj, struct uverbs_attr_bundle *attrs)
 /*
  * uobj_get_destroy destroys the HW object and returns a handle to the uobj
  * with a NULL object pointer. The caller must pair this with
- * uverbs_put_destroy.
+ * uobj_put_destroy().
  */
 struct ib_uobject *__uobj_get_destroy(const struct uverbs_api_object *obj,
 				      u32 id, struct uverbs_attr_bundle *attrs)
@@ -250,8 +257,7 @@ int __uobj_perform_destroy(const struct uverbs_api_object *obj, u32 id,
 	uobj = __uobj_get_destroy(obj, id, attrs);
 	if (IS_ERR(uobj))
 		return PTR_ERR(uobj);
-
-	rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_WRITE);
+	uobj_put_destroy(uobj);
 	return 0;
 }
 

drivers/infiniband/hw/mlx5/mr.c

@@ -1439,6 +1439,7 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 
 	if (is_odp_mr(mr)) {
 		to_ib_umem_odp(mr->umem)->private = mr;
+		init_waitqueue_head(&mr->q_deferred_work);
 		atomic_set(&mr->num_deferred_work, 0);
 		err = xa_err(xa_store(&dev->odp_mkeys,
 				      mlx5_base_mkey(mr->mmkey.key), &mr->mmkey,

drivers/infiniband/hw/qib/qib_sysfs.c

@@ -760,7 +760,7 @@ int qib_create_port_files(struct ib_device *ibdev, u8 port_num,
 		qib_dev_err(dd,
 			"Skipping linkcontrol sysfs info, (err %d) port %u\n",
 			ret, port_num);
-		goto bail;
+		goto bail_link;
 	}
 	kobject_uevent(&ppd->pport_kobj, KOBJ_ADD);
 
@@ -770,7 +770,7 @@ int qib_create_port_files(struct ib_device *ibdev, u8 port_num,
 		qib_dev_err(dd,
 			"Skipping sl2vl sysfs info, (err %d) port %u\n",
 			ret, port_num);
-		goto bail_link;
+		goto bail_sl;
 	}
 	kobject_uevent(&ppd->sl2vl_kobj, KOBJ_ADD);
 
@@ -780,7 +780,7 @@ int qib_create_port_files(struct ib_device *ibdev, u8 port_num,
 		qib_dev_err(dd,
 			"Skipping diag_counters sysfs info, (err %d) port %u\n",
 			ret, port_num);
-		goto bail_sl;
+		goto bail_diagc;
 	}
 	kobject_uevent(&ppd->diagc_kobj, KOBJ_ADD);
 
@@ -793,7 +793,7 @@ int qib_create_port_files(struct ib_device *ibdev, u8 port_num,
 		qib_dev_err(dd,
 		 "Skipping Congestion Control sysfs info, (err %d) port %u\n",
 		 ret, port_num);
-		goto bail_diagc;
+		goto bail_cc;
 	}
 
 	kobject_uevent(&ppd->pport_cc_kobj, KOBJ_ADD);
@@ -854,6 +854,7 @@ void qib_verbs_unregister_sysfs(struct qib_devdata *dd)
 				&cc_table_bin_attr);
 			kobject_put(&ppd->pport_cc_kobj);
 		}
+		kobject_put(&ppd->diagc_kobj);
 		kobject_put(&ppd->sl2vl_kobj);
 		kobject_put(&ppd->pport_kobj);
 	}

drivers/infiniband/hw/vmw_pvrdma/pvrdma_main.c

@@ -829,7 +829,7 @@ static int pvrdma_pci_probe(struct pci_dev *pdev,
 	    !(pci_resource_flags(pdev, 1) & IORESOURCE_MEM)) {
 		dev_err(&pdev->dev, "PCI BAR region not MMIO\n");
 		ret = -ENOMEM;
-		goto err_free_device;
+		goto err_disable_pdev;
 	}
 
 	ret = pci_request_regions(pdev, DRV_NAME);

drivers/infiniband/ulp/ipoib/ipoib.h

@@ -377,8 +377,12 @@ struct ipoib_dev_priv {
 	struct ipoib_rx_buf *rx_ring;
 
 	struct ipoib_tx_buf *tx_ring;
+	/* cyclic ring variables for managing tx_ring, for UD only */
 	unsigned int	     tx_head;
 	unsigned int	     tx_tail;
+	/* cyclic ring variables for counting overall outstanding send WRs */
+	unsigned int	     global_tx_head;
+	unsigned int	     global_tx_tail;
 	struct ib_sge	     tx_sge[MAX_SKB_FRAGS + 1];
 	struct ib_ud_wr      tx_wr;
 	struct ib_wc	     send_wc[MAX_SEND_CQE];

drivers/infiniband/ulp/ipoib/ipoib_cm.c

@@ -756,7 +756,8 @@ void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_
 		return;
 	}
 
-	if ((priv->tx_head - priv->tx_tail) == ipoib_sendq_size - 1) {
+	if ((priv->global_tx_head - priv->global_tx_tail) ==
+	    ipoib_sendq_size - 1) {
 		ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 			  tx->qp->qp_num);
 		netif_stop_queue(dev);
@@ -786,7 +787,7 @@ void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_
 	} else {
 		netif_trans_update(dev);
 		++tx->tx_head;
-		++priv->tx_head;
+		++priv->global_tx_head;
 	}
 }
 
@@ -820,10 +821,11 @@ void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 	netif_tx_lock(dev);
 
 	++tx->tx_tail;
-	++priv->tx_tail;
+	++priv->global_tx_tail;
 
 	if (unlikely(netif_queue_stopped(dev) &&
-		     (priv->tx_head - priv->tx_tail) <= ipoib_sendq_size >> 1 &&
+		     ((priv->global_tx_head - priv->global_tx_tail) <=
+		      ipoib_sendq_size >> 1) &&
 		     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
 		netif_wake_queue(dev);
 
@@ -1232,8 +1234,9 @@ static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
 		dev_kfree_skb_any(tx_req->skb);
 		netif_tx_lock_bh(p->dev);
 		++p->tx_tail;
-		++priv->tx_tail;
-		if (unlikely(priv->tx_head - priv->tx_tail == ipoib_sendq_size >> 1) &&
+		++priv->global_tx_tail;
+		if (unlikely((priv->global_tx_head - priv->global_tx_tail) <=
+			     ipoib_sendq_size >> 1) &&
 		    netif_queue_stopped(p->dev) &&
 		    test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 			netif_wake_queue(p->dev);

drivers/infiniband/ulp/ipoib/ipoib_ib.c

@@ -407,9 +407,11 @@ static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 	dev_kfree_skb_any(tx_req->skb);
 
 	++priv->tx_tail;
+	++priv->global_tx_tail;
 
 	if (unlikely(netif_queue_stopped(dev) &&
-		     ((priv->tx_head - priv->tx_tail) <= ipoib_sendq_size >> 1) &&
+		     ((priv->global_tx_head - priv->global_tx_tail) <=
+		      ipoib_sendq_size >> 1) &&
 		     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
 		netif_wake_queue(dev);
 
@@ -634,7 +636,8 @@ int ipoib_send(struct net_device *dev, struct sk_buff *skb,
 	else
 		priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
 	/* increase the tx_head after send success, but use it for queue state */
-	if (priv->tx_head - priv->tx_tail == ipoib_sendq_size - 1) {
+	if ((priv->global_tx_head - priv->global_tx_tail) ==
+	    ipoib_sendq_size - 1) {
 		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");
 		netif_stop_queue(dev);
 	}
@@ -662,6 +665,7 @@ int ipoib_send(struct net_device *dev, struct sk_buff *skb,
 
 		rc = priv->tx_head;
 		++priv->tx_head;
+		++priv->global_tx_head;
 	}
 	return rc;
 }
@@ -807,6 +811,7 @@ int ipoib_ib_dev_stop_default(struct net_device *dev)
 				ipoib_dma_unmap_tx(priv, tx_req);
 				dev_kfree_skb_any(tx_req->skb);
 				++priv->tx_tail;
+				++priv->global_tx_tail;
 			}
 
 			for (i = 0; i < ipoib_recvq_size; ++i) {

drivers/infiniband/ulp/ipoib/ipoib_main.c

@@ -1184,9 +1184,11 @@ static void ipoib_timeout(struct net_device *dev, unsigned int txqueue)
 
 	ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
 		   jiffies_to_msecs(jiffies - dev_trans_start(dev)));
-	ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
-		   netif_queue_stopped(dev),
-		   priv->tx_head, priv->tx_tail);
+	ipoib_warn(priv,
+		   "queue stopped %d, tx_head %u, tx_tail %u, global_tx_head %u, global_tx_tail %u\n",
+		   netif_queue_stopped(dev), priv->tx_head, priv->tx_tail,
+		   priv->global_tx_head, priv->global_tx_tail);
+
 	/* XXX reset QP, etc. */
 }
 
@@ -1701,7 +1703,7 @@ static int ipoib_dev_init_default(struct net_device *dev)
 		goto out_rx_ring_cleanup;
 	}
 
-	/* priv->tx_head, tx_tail & tx_outstanding are already 0 */
+	/* priv->tx_head, tx_tail and global_tx_tail/head are already 0 */
 
 	if (ipoib_transport_dev_init(dev, priv->ca)) {
 		pr_warn("%s: ipoib_transport_dev_init failed\n",

include/rdma/uverbs_std_types.h

@@ -88,7 +88,7 @@ struct ib_uobject *__uobj_get_destroy(const struct uverbs_api_object *obj,
 
 static inline void uobj_put_destroy(struct ib_uobject *uobj)
 {
-	rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_WRITE);
+	rdma_lookup_put_uobject(uobj, UVERBS_LOOKUP_DESTROY);
 }
 
 static inline void uobj_put_read(struct ib_uobject *uobj)