xen/pciback: Fine-grain the spinlocks and fix BUG: scheduling while atomic cases.
authorKonrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Fri, 23 Jul 2010 18:35:47 +0000 (14:35 -0400)
committerKonrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Wed, 20 Jul 2011 00:58:32 +0000 (20:58 -0400)
We were using coarse spinlocks that could end up with a deadlock.
This patch fixes that and makes the spinlocks much more fine-grained.

We also drop be->watchding state spinlocks as they are already
guarded by the xenwatch_thread against multiple customers. Without
that we would trigger the BUG: scheduling while atomic.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
drivers/xen/xen-pciback/passthrough.c
drivers/xen/xen-pciback/xenbus.c

index 5386bebf7f9a22c6823b427952ca1a7c3b36b042..6e3999b997d486334319f41e1902a1313bbf28ce 100644 (file)
@@ -113,14 +113,14 @@ int pciback_publish_pci_roots(struct pciback_device *pdev,
 {
        int err = 0;
        struct passthrough_dev_data *dev_data = pdev->pci_dev_data;
-       struct pci_dev_entry *dev_entry, *e;
+       struct pci_dev_entry *dev_entry, *e, *tmp;
        struct pci_dev *dev;
        int found;
        unsigned int domain, bus;
 
        spin_lock(&dev_data->lock);
 
-       list_for_each_entry(dev_entry, &dev_data->dev_list, list) {
+       list_for_each_entry_safe(dev_entry, tmp, &dev_data->dev_list, list) {
                /* Only publish this device as a root if none of its
                 * parent bridges are exported
                 */
@@ -139,13 +139,16 @@ int pciback_publish_pci_roots(struct pciback_device *pdev,
                bus = (unsigned int)dev_entry->dev->bus->number;
 
                if (!found) {
+                       spin_unlock(&dev_data->lock);
                        err = publish_root_cb(pdev, domain, bus);
                        if (err)
                                break;
+                       spin_lock(&dev_data->lock);
                }
        }
 
-       spin_unlock(&dev_data->lock);
+       if (!err)
+               spin_unlock(&dev_data->lock);
 
        return err;
 }
index a0cf7285d320c3138555099dfdc0e04211399b89..70030c4092128c736f2ab087a21991fc213daf6e 100644 (file)
@@ -54,23 +54,29 @@ static void pciback_disconnect(struct pciback_device *pdev)
                unbind_from_irqhandler(pdev->evtchn_irq, pdev);
                pdev->evtchn_irq = INVALID_EVTCHN_IRQ;
        }
+       spin_unlock(&pdev->dev_lock);
 
        /* If the driver domain started an op, make sure we complete it
         * before releasing the shared memory */
+
+       /* Note, the workqueue does not use spinlocks at all.*/
        flush_workqueue(pciback_wq);
 
+       spin_lock(&pdev->dev_lock);
        if (pdev->sh_info != NULL) {
                xenbus_unmap_ring_vfree(pdev->xdev, pdev->sh_info);
                pdev->sh_info = NULL;
        }
-
        spin_unlock(&pdev->dev_lock);
+
 }
 
 static void free_pdev(struct pciback_device *pdev)
 {
-       if (pdev->be_watching)
+       if (pdev->be_watching) {
                unregister_xenbus_watch(&pdev->be_watch);
+               pdev->be_watching = 0;
+       }
 
        pciback_disconnect(pdev);
 
@@ -98,7 +104,10 @@ static int pciback_do_attach(struct pciback_device *pdev, int gnt_ref,
                                "Error mapping other domain page in ours.");
                goto out;
        }
+
+       spin_lock(&pdev->dev_lock);
        pdev->sh_info = vaddr;
+       spin_unlock(&pdev->dev_lock);
 
        err = bind_interdomain_evtchn_to_irqhandler(
                pdev->xdev->otherend_id, remote_evtchn, pciback_handle_event,
@@ -108,7 +117,10 @@ static int pciback_do_attach(struct pciback_device *pdev, int gnt_ref,
                                 "Error binding event channel to IRQ");
                goto out;
        }
+
+       spin_lock(&pdev->dev_lock);
        pdev->evtchn_irq = err;
+       spin_unlock(&pdev->dev_lock);
        err = 0;
 
        dev_dbg(&pdev->xdev->dev, "Attached!\n");
@@ -122,7 +134,6 @@ static int pciback_attach(struct pciback_device *pdev)
        int gnt_ref, remote_evtchn;
        char *magic = NULL;
 
-       spin_lock(&pdev->dev_lock);
 
        /* Make sure we only do this setup once */
        if (xenbus_read_driver_state(pdev->xdev->nodename) !=
@@ -168,7 +179,6 @@ static int pciback_attach(struct pciback_device *pdev)
 
        dev_dbg(&pdev->xdev->dev, "Connected? %d\n", err);
 out:
-       spin_unlock(&pdev->dev_lock);
 
        kfree(magic);
 
@@ -340,7 +350,6 @@ static int pciback_reconfigure(struct pciback_device *pdev)
        char state_str[64];
        char dev_str[64];
 
-       spin_lock(&pdev->dev_lock);
 
        dev_dbg(&pdev->xdev->dev, "Reconfiguring device ...\n");
 
@@ -481,8 +490,6 @@ static int pciback_reconfigure(struct pciback_device *pdev)
        }
 
 out:
-       spin_unlock(&pdev->dev_lock);
-
        return 0;
 }
 
@@ -539,8 +546,6 @@ static int pciback_setup_backend(struct pciback_device *pdev)
        char dev_str[64];
        char state_str[64];
 
-       spin_lock(&pdev->dev_lock);
-
        /* It's possible we could get the call to setup twice, so make sure
         * we're not already connected.
         */
@@ -621,8 +626,6 @@ static int pciback_setup_backend(struct pciback_device *pdev)
                                 "Error switching to initialised state!");
 
 out:
-       spin_unlock(&pdev->dev_lock);
-
        if (!err)
                /* see if pcifront is already configured (if not, we'll wait) */
                pciback_attach(pdev);
@@ -669,6 +672,7 @@ static int pciback_xenbus_probe(struct xenbus_device *dev,
                                pciback_be_watch);
        if (err)
                goto out;
+
        pdev->be_watching = 1;
 
        /* We need to force a call to our callback here in case