IB/mlx5: Disable BME on panic

There are two reasons for disabling BME on an RDMA capable device on a
panic. The first one is Exadata's FNDD logic, the other is quite
simple; we do not want a vmcore to be generated whilst there are
incoming RDMA Write/Atomics, modifying the image while it is
generated.

Exadata has implemented Fast Node Death Detection (FNDD) by means of
posting RDMA operation between nodes. Now, since the RDMA responder is
handled by the remote HCA, the RDMA requests will be satisfied, even
though the host OS is crashing, generating vmcore, or other. It will
do so until the PCIe PRST signal (PCI Reset) has been raised.

Exadata has tried to circumvent this situation by having a process at
the target, regularly incrementing a variable in the MR used by FNDD,
then performing RDMA Read to read the variable and make sure it
increments.

This mechanism has proven to give too many false negatives, as the
process incrementing the variable may be suspended for several
seconds, due to a very high number of processes.

Hence, the idea is to revoke the HCA's ability to perform host memory
accesses, by simply resetting the Bus Master Enable (BME) bit, when
the host OS panics or reboots.

Here is an excerpt of the PCI Express specification:

<quote>
Bus Master Enable - Controls the ability of a PCI Express Endpoint to
issue Memory and I/O Read/Write Requests, and the ability of a Root or
Switch Port to forward Memory and I/O Read/Write Requests in the
Upstream direction

Endpoints:
When this bit is Set, the PCI Express Function is allowed to issue
Memory or I/O Requests.

When this bit is Clear, the PCI Express Function is not allowed to
issue any Memory or I/O Requests.

Note that as MSI/MSI-X interrupt Messages are in-band memory writes,
setting the Bus Master Enable bit to 0b disables MSI/MSI-X interrupt
Messages as well.

Requests other than Memory or I/O Requests are not controlled by this
bit.

Default value of this bit is 0b.

This bit is hardwired to 0b if a Function does not generate Memory or
I/O Requests.
</quote>

To accommodate Exadata's requirement here, we install a panic-notifier
than when invoked, revokes BME for the function.

Orabug: 31556128

UEK5 => UEK6

(cherry picked from commit 6baf337)
cherry-pick-repo=UEK/production/linux-uek.git

Conflicts:
       drivers/infiniband/hw/mlx5/mlx5_ib.h
       drivers/infiniband/hw/mlx5/main.c

       * The conflict arouse because in UEK5, the device struct
       * contained an array of mlx5_roce structs, whereas in UEK6, the
       * mlx5_roce struct is a sub-ordinate to the device's port
       * array.

Signed-off-by: Håkon Bugge <[email protected]>
Reviewed-by: Sharath Srinivasan <[email protected]>
Signed-off-by: Aron Silverton <[email protected]>

Author: Hakon-Bugge

drivers/infiniband/hw/mlx5/main.c

@@ -5281,6 +5281,75 @@ static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
 	}
 }
 
+#ifndef WITHOUT_ORACLE_EXTENSIONS
+static int mlx5_ib_panic_handler(struct notifier_block *nb,
+				 unsigned long action, void *unused)
+{
+	struct mlx5_ib_dev *dev;
+
+	list_for_each_entry(dev, &mlx5_ib_dev_list, ib_dev_list) {
+		struct pci_dev *pdev = dev->mdev->pdev;
+
+		pci_crit(pdev, "Revoke Bus_Mastership_Enable (BME)\n");
+		pci_clear_master(pdev);
+	}
+
+	return NOTIFY_DONE;
+}
+
+static bool panic_handler_added;
+static int mlx5_ib_add_panic_notifier(struct mlx5_ib_dev *dev, u8 port_num)
+{
+	int err;
+
+	/* We only need to add the handler once, as it iterates
+	 * through all the devices
+	 */
+	if (panic_handler_added)
+		return 0;
+
+	dev->port[port_num].roce.panic_nb.notifier_call = mlx5_ib_panic_handler;
+	err = atomic_notifier_chain_register(&panic_notifier_list,
+					     &dev->port[port_num].roce.panic_nb);
+	if (err) {
+		pci_crit(dev->mdev->pdev,
+			 "Failed registering panic handler for port %d, err %d\n",
+			 port_num + 1, err);
+		dev->port[port_num].roce.panic_nb.notifier_call = NULL;
+		return err;
+	}
+
+	pci_crit(dev->mdev->pdev,
+		 "Successfully registered panic handler for port %d\n", port_num + 1);
+
+	panic_handler_added = true;
+
+	return 0;
+}
+
+static void mlx5_ib_remove_panic_notifier(struct mlx5_ib_dev *dev, u8 port_num)
+{
+	int err = 0;
+
+	if (dev->port[port_num].roce.panic_nb.notifier_call) {
+		err = atomic_notifier_chain_unregister(&panic_notifier_list,
+						       &dev->port[port_num].roce.panic_nb);
+		if (err) {
+			pci_crit(dev->mdev->pdev,
+				 "Unregistering the panic handler for port %d failed, err %d\n",
+				 port_num + 1, err);
+		} else {
+			pci_crit(dev->mdev->pdev,
+				 "Successfully unregistered panic handler for port %d",
+				 port_num + 1);
+			panic_handler_added = false;
+		}
+
+		dev->port[port_num].roce.panic_nb.notifier_call = NULL;
+	}
+}
+#endif /* !WITHOUT_ORACLE_EXTENSIONS */
+
 static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
 {
 	int err;
@@ -6569,6 +6638,9 @@ static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = {
 static int mlx5_ib_stage_common_roce_init(struct mlx5_ib_dev *dev)
 {
 	u8 port_num;
+#ifndef WITHOUT_ORACLE_EXTENSIONS
+	int err;
+#endif /* !WITHOUT_ORACLE_EXTENSIONS */
 
 	dev->ib_dev.uverbs_ex_cmd_mask |=
 			(1ull << IB_USER_VERBS_EX_CMD_CREATE_WQ) |
@@ -6581,13 +6653,32 @@ static int mlx5_ib_stage_common_roce_init(struct mlx5_ib_dev *dev)
 	port_num = mlx5_core_native_port_num(dev->mdev) - 1;
 
 	/* Register only for native ports */
-	return mlx5_add_netdev_notifier(dev, port_num);
+#ifdef WITHOUT_ORACLE_EXTENSIONS
+	return mlx5_add_netdev_notifier(dev, port_num)
+#else
+	err = mlx5_add_netdev_notifier(dev, port_num);
+	if (err)
+		return err;
+
+	err = mlx5_ib_add_panic_notifier(dev, port_num);
+	if (err)
+		goto err_remove_netdev_notifier;
+
+	return 0;
+
+err_remove_netdev_notifier:
+	mlx5_remove_netdev_notifier(dev, port_num);
+	return err;
+#endif /* WITHOUT_ORACLE_EXTENSIONS */
 }
 
 static void mlx5_ib_stage_common_roce_cleanup(struct mlx5_ib_dev *dev)
 {
 	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
 
+#ifndef WITHOUT_ORACLE_EXTENSIONS
+	mlx5_ib_remove_panic_notifier(dev, port_num);
+#endif /* !WITHOUT_ORACLE_EXTENSIONS */
 	mlx5_remove_netdev_notifier(dev, port_num);
 }
 

drivers/infiniband/hw/mlx5/mlx5_ib.h

@@ -744,6 +744,9 @@ struct mlx5_roce {
 	rwlock_t		netdev_lock;
 	struct net_device	*netdev;
 	struct notifier_block	nb;
+#ifndef WITHOUT_ORACLE_EXTENSIONS
+	struct notifier_block   panic_nb;
+#endif /* !WITHOUT_ORACLE_EXTENSIONS */
 	atomic_t		tx_port_affinity;
 	enum ib_port_state last_port_state;
 	struct mlx5_ib_dev	*dev;