sfc: Use a separate workqueue for resets

This avoids deadlock in case a reset is triggered during self-test.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>

2 files changed, 21 insertions, 4 deletions

diff --git a/drivers/net/sfc/efx.c b/drivers/net/sfc/efx.c
index 7b2015f9e469..7b2a81866232 100644
--- a/drivers/net/sfc/efx.c
+++ b/drivers/net/sfc/efx.c
@@ -1762,7 +1762,7 @@ void efx_schedule_reset(struct efx_nic *efx, enum reset_type type)
 
         efx->reset_pending = method;
 
-        queue_work(efx->workqueue, &efx->reset_work);
+        queue_work(efx->reset_workqueue, &efx->reset_work);
 }
 
 /**************************************************************************
@@ -1907,14 +1907,28 @@ static int efx_init_struct(struct efx_nic *efx, struct efx_nic_type *type,
                 goto fail1;
         }
 
+        efx->reset_workqueue = create_singlethread_workqueue("sfc_reset");
+        if (!efx->reset_workqueue) {
+                rc = -ENOMEM;
+                goto fail2;
+        }
+
         return 0;
 
+ fail2:
+        destroy_workqueue(efx->workqueue);
+        efx->workqueue = NULL;
+
  fail1:
         return rc;
 }
 
 static void efx_fini_struct(struct efx_nic *efx)
 {
+        if (efx->reset_workqueue) {
+                destroy_workqueue(efx->reset_workqueue);
+                efx->reset_workqueue = NULL;
+        }
         if (efx->workqueue) {
                 destroy_workqueue(efx->workqueue);
                 efx->workqueue = NULL;
@@ -1977,7 +1991,7 @@ static void efx_pci_remove(struct pci_dev *pci_dev)
          * scheduled from this point because efx_stop_all() has been
          * called, we are no longer registered with driverlink, and
          * the net_device's have been removed. */
-        flush_workqueue(efx->workqueue);
+        flush_workqueue(efx->reset_workqueue);
 
         efx_pci_remove_main(efx);
 
@@ -2098,7 +2112,7 @@ static int __devinit efx_pci_probe(struct pci_dev *pci_dev,
                  * scheduled since efx_stop_all() has been called, and we
                  * have not and never have been registered with either
                  * the rtnetlink or driverlink layers. */
-                cancel_work_sync(&efx->reset_work);
+                flush_workqueue(efx->reset_workqueue);
 
                 /* Retry if a recoverably reset event has been scheduled */
                 if ((efx->reset_pending != RESET_TYPE_INVISIBLE) &&
diff --git a/drivers/net/sfc/net_driver.h b/drivers/net/sfc/net_driver.h
index d803b86c647c..219c74a772c3 100644
--- a/drivers/net/sfc/net_driver.h
+++ b/drivers/net/sfc/net_driver.h
@@ -616,7 +616,9 @@ union efx_multicast_hash {
  * @pci_dev: The PCI device
  * @type: Controller type attributes
  * @legacy_irq: IRQ number
- * @workqueue: Workqueue for resets, port reconfigures and the HW monitor
+ * @workqueue: Workqueue for port reconfigures and the HW monitor.
+ *      Work items do not hold and must not acquire RTNL.
+ * @reset_workqueue: Workqueue for resets.  Work item will acquire RTNL.
  * @reset_work: Scheduled reset workitem
  * @monitor_work: Hardware monitor workitem
  * @membase_phys: Memory BAR value as physical address
@@ -684,6 +686,7 @@ struct efx_nic {
         const struct efx_nic_type *type;
         int legacy_irq;
         struct workqueue_struct *workqueue;
+        struct workqueue_struct *reset_workqueue;
         struct work_struct reset_work;
         struct delayed_work monitor_work;
         resource_size_t membase_phys;