datapath: Avoid system freeze due to ovs-flow-rehash softlockup.

OVS datapath does periodic flow table rehash which takes genl_lock
in workq context.
In some cases, like ports add or delete, genl_lock can cause softlockup
as vswitchd would take and succeed with genl_lock and rehash workq
would block on the lock. Eventually rehash will proceed, flow rehash
is low priority task so this is not problem for rehashing.
But it is blocking workq thread; some other workq item from other
kernel subsystem would be blocked and can cause system freeze.
To avoid workq blocking and system freeze, we can use OVS compat workq.
It runs in separate kernel thread thus does not block any non-ovs
deferred workq work item.

We will fix it by making genetlink lockless and having fine granular
locking in OVS.

Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Acked-by: Jesse Gross <jesse@nicira.com>

diff --git a/datapath/linux/compat/include/linux/workqueue.h b/datapath/linux/compat/include/linux/workqueue.h
index 919afe350b1036657f5221e563014bd36bf0d1f8..79158f965b699f4e1538598fa6caae9d90e78eb1 100644 (file)
--- a/datapath/linux/compat/include/linux/workqueue.h
+++ b/datapath/linux/compat/include/linux/workqueue.h
@@ -1,23 +1,19 @@
 #ifndef __LINUX_WORKQUEUE_WRAPPER_H
 #define __LINUX_WORKQUEUE_WRAPPER_H 1
 
-#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,23)
-#include_next <linux/workqueue.h>
-static inline int __init ovs_workqueues_init(void) { return 0; }
-static inline void  ovs_workqueues_exit(void) {}
-
-#else
 #include <linux/timer.h>
 
 int __init ovs_workqueues_init(void);
 void ovs_workqueues_exit(void);
 
-
 /* Older kernels have an implementation of work queues with some very bad
  * characteristics when trying to cancel work (potential deadlocks, use after
  * free, etc.  Therefore we implement simple ovs specific work queue using
  * single worker thread. work-queue API are kept similar for compatibility.
+ * It seems it is useful even on newer kernel. As it can avoid system wide
+ * freeze in event of softlockup due to workq blocked on genl_lock.
  */
+
 struct work_struct;
 
 typedef void (*work_func_t)(struct work_struct *work);
@@ -68,6 +64,6 @@ int cancel_delayed_work_sync(struct delayed_work *dwork);
                (_work)->func = (_func);                        \
        } while (0)
 
-#endif /* kernel version < 2.6.23 */
+extern void flush_scheduled_work(void);
 
 #endif

diff --git a/datapath/linux/compat/workqueue.c b/datapath/linux/compat/workqueue.c
index 883665bf45e3b2e0d6519a01ec0556443341f627..9934f1a397b4ccd9e096f16561940f083bfc2248 100644 (file)
--- a/datapath/linux/compat/workqueue.c
+++ b/datapath/linux/compat/workqueue.c
@@ -23,8 +23,6 @@
 #include <linux/lockdep.h>
 #include <linux/idr.h>
 
-#if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,23)
-
 static spinlock_t wq_lock;
 static struct list_head workq;
 static wait_queue_head_t more_work;
@@ -210,4 +208,3 @@ void  ovs_workqueues_exit(void)
        BUG_ON(!list_empty(&workq));
        kthread_stop(workq_thread);
 }
-#endif