sb_edac.c, kernel linux-3.2-rc6.

While testing Sandy Bridge EDAC module I discovered a problem in the way sb_edac was registering itself for machine check notifications. The symptoms of the problem include:
1. Injecting a machine check exception can cause the system to hang for 10-15 seconds.
2. Removing and re-inserting the kernel module can cause panic.

The system hangs for 10-15 seconds because the sb_edac notifier gets called by the kernel (notifier_call_chain) 0xffffffff times ((u32)(-1)).

The problem occurs because sb_edac calls atomic_notifier_chain_register twice with the same static notifier_block structure. The function atomic_notifier_chain_register gets called once for each memory controller (MC) with the same structure. The patch, then, fixes this problem by making sure that sb_edac registers for machine check notifications only once.

Also copying Mauro Carvalho Chehab (maintainer of sb_edac) for the review of the patch.

Cheers,
Karan

@@ -1661,9 +1661,6 @@
     debugf0("MC: " __FILE__ ": %s(): mci = %p, dev = %p\n",
         __func__, mci, &sbridge_dev->pdev[0]->dev);
 
-    atomic_notifier_chain_unregister(&x86_mce_decoder_chain,
- &sbridge_mce_dec);
-
     /* Remove MC sysfs nodes */
     edac_mc_del_mc(mci->dev);
 

@@ -1731,8 +1728,6 @@
         goto fail0;
     }
 
-    atomic_notifier_chain_register(&x86_mce_decoder_chain,
- &sbridge_mce_dec);
     return 0;
 
 fail0:

@@ -1861,8 +1856,11 @@
 
     pci_rc = pci_register_driver(&sbridge_driver);
 
-    if (pci_rc >= 0)
+    if (pci_rc >= 0) {
+         atomic_notifier_chain_register(&x86_mce_decoder_chain,
+ &sbridge_mce_dec);
         return 0;
+    }
 
     sbridge_printk(KERN_ERR, "Failed to register device with error %d.\n",
               pci_rc);

@@ -1877,6 +1875,9 @@
 static void __exit sbridge_exit(void)
 {
     debugf2("MC: " __FILE__ ": %s()\n", __func__);
+    atomic_notifier_chain_unregister(&x86_mce_decoder_chain,
+ &sbridge_mce_dec);
+
     pci_unregister_driver(&sbridge_driver);
 }
 

@@ -1661,9 +1661,6 @@
 	debugf0("MC: " __FILE__ ": %s(): mci = %p, dev = %p\n",
 		__func__, mci, &sbridge_dev->pdev[0]->dev);
 
-	atomic_notifier_chain_unregister(&x86_mce_decoder_chain,
-					 &sbridge_mce_dec);
-
 	/* Remove MC sysfs nodes */
 	edac_mc_del_mc(mci->dev);
 

@@ -1731,8 +1728,6 @@
 		goto fail0;
 	}
 
-	atomic_notifier_chain_register(&x86_mce_decoder_chain,
-				       &sbridge_mce_dec);
 	return 0;
 
 fail0:

@@ -1861,8 +1856,11 @@
 
 	pci_rc = pci_register_driver(&sbridge_driver);
 
-	if (pci_rc >= 0)
+	if (pci_rc >= 0) {
+ 		atomic_notifier_chain_register(&x86_mce_decoder_chain,
+						  &sbridge_mce_dec);
 		return 0;
+	}
 
 	sbridge_printk(KERN_ERR, "Failed to register device with error %d.\n",
 		      pci_rc);

@@ -1877,6 +1875,9 @@
 static void __exit sbridge_exit(void)
 {
 	debugf2("MC: " __FILE__ ": %s()\n", __func__);
+	atomic_notifier_chain_unregister(&x86_mce_decoder_chain,
+					 &sbridge_mce_dec);
+
 	pci_unregister_driver(&sbridge_driver);
 }
 

[PATCH] sb_edac.c, kernel linux-3.2-rc6.
Karandeep Chahal <[email protected]>

The sb_edac patch fixes incorrect Sandy Bridge machine check notifier chain registration problem.