[ Alison, can you try this patch ]
This uses netpoll_poll_lock()/unlock() to synchronize netpoll and napi
poll operations. Without this method, the synchronization is done by
looping on NAPI_STATE_SCHED 'bitset'. This method works fine on a non-rt
kernel because a softirq can not be preempted, and the thread poll is
called with local_bh_disable() which prevents softirqs from running and
preempting it. But on rt, this code can be preempted. Thus, the code may
be preempted out while holding the NAPI_STATE_SCHED 'bitset', opening a
window for a livelock.
For example:
<interrupt thread (as all interrupts on RT are threaded>
napi_schedule_prep()
test_and_set_bit(NAPI_STATE_SCHED, &n->state)
<preempted by higher prio task that runs softirqs in its context>
sk_busy_loop()
do {
rc = busy_poll()
ret = napi_schedule_prep()
return !test_and_set_bit(NAPI_STATE_SCHED, &n->state)
<returns zero because NAPI_STATE_SCHED is set>
if (!ret) return 0
<rc is zero>
} while (...) /* for ever */
This isn't a problem in non PREEMPT_RT because the napi_schedule_prep()
can not be preempted. But because it can in PREEMPT_RT, we need to add
some extra locking. The netpoll_poll_lock() works well here, but they need
to be added around any call to busy_poll().
Using IS_ENABLED(CONFIG_PREEMPT_RT_FULL) will allow gcc to optimize out
the extra calls to poll_lock.
Tested-by: "Luis Claudio R. Goncalves" <lgonc...@redhat.com>
Reviewed-by: Daniel Bristot de Oliveira <bris...@redhat.com>
Signed-off-by: Steven Rostedt <rost...@goodmis.org>
---
include/linux/netpoll.h | 2 +-
include/net/busy_poll.h | 14 +++++++++++++-
2 files changed, 14 insertions(+), 2 deletions(-)
Index: linux-rt.git/include/linux/netpoll.h
===================================================================
--- linux-rt.git.orig/include/linux/netpoll.h 2016-05-26 18:31:09.183150389
-0400
+++ linux-rt.git/include/linux/netpoll.h 2016-05-26 18:52:02.657014280
-0400
@@ -77,7 +77,7 @@ static inline void *netpoll_poll_lock(st
{
struct net_device *dev = napi->dev;
- if (dev && dev->npinfo) {
+ if (dev && (IS_ENABLED(CONFIG_PREEMPT_RT_FULL) || dev->npinfo)) {
spin_lock(&napi->poll_lock);
napi->poll_owner = smp_processor_id();
return napi;
Index: linux-rt.git/include/net/busy_poll.h
===================================================================
--- linux-rt.git.orig/include/net/busy_poll.h 2016-05-26 18:31:09.183150389
-0400
+++ linux-rt.git/include/net/busy_poll.h 2016-05-26 19:10:12.134266713
-0400
@@ -25,6 +25,7 @@
#define _LINUX_NET_BUSY_POLL_H
#include <linux/netdevice.h>
+#include <linux/netpoll.h>
#include <net/ip.h>
#ifdef CONFIG_NET_RX_BUSY_POLL
@@ -97,7 +98,18 @@ static inline bool sk_busy_loop(struct s
goto out;
do {
- rc = ops->ndo_busy_poll(napi);
+ /* When RT is enabled, napi_schedule_prep() can be preempted
+ * with NAPI_STATE_SCHED set, causing the busy_poll() function
+ * to always return zero, and this loop may never exit.
+ * In that case, we must always take the netpoll_poll_lock.
+ */
+ if (IS_ENABLED(CONFIG_PREEMPT_RT_FULL)) {
+ void *have = netpoll_poll_lock(napi);
+ rc = ops->ndo_busy_poll(napi);
+ netpoll_poll_unlock(have);
+ } else {
+ rc = ops->ndo_busy_poll(napi);
+ }
if (rc == LL_FLUSH_FAILED)
break; /* permanent failure */
