ipv4: igmp: use alarmtimer to prevent delayed reports

On receiving a IGMPv2/v3 query, based on max_delay set in the header a timer is started to send out a response after a random time within max_delay. If the system then moves into suspend state, Report is delayed until system wakes up.

Use a alarmtimer instead of using a timer. Alarmtimer will wake the system up from suspend to send out the IGMP report.

Signed-off-by Tejaswi Tanikella <[email protected]>

---
 include/linux/ktime.h | 4 ++++
 1 file changed, 4 insertions(+)

include/linux/ktime.h +0 -0

index 5b9fddb..4881483 100644

Quoted 7 lines · Tejaswi Tanikella

On receiving a IGMPv2/v3 query, based on max_delay set in the header a timer is started to send out a response after a random time within max_delay. If the system then moves into suspend state, Report is delayed until system wakes up.

Use a alarmtimer instead of using a timer. Alarmtimer will wake the system up from suspend to send out the IGMP report.

Hi Tejaswi

I think i must be missing something here. If we are suspended, we are not receiving multicast frames. If we are not receiving frames, why do we need to reply to the query?

Once we resume, i expect we will reply to the next query. You could optimise restarting the flow by immediately sending a membership report, same as when the setsockopt is used to join the group.

Andrew

Quoted 10 lines · Tejaswi Tanikella

Signed-off-by Tejaswi Tanikella <[email protected]>

---
 include/linux/ktime.h | 4 ++++
 1 file changed, 4 insertions(+)

include/linux/ktime.h +0 -0

index 5b9fddb..4881483 100644

include/linux/ktime.h static inline ktime_t timeval_to_ktime(struct timeval tv) · line 101

Quoted 6 lines

include/linux/ktime.h +3 -0

@@ -96,6 +96,10 @@ static inline ktime_t timeval_to_ktime(struct timeval tv)
 /* Convert ktime_t to nanoseconds - NOP in the scalar storage format: */
 #define ktime_to_ns(kt)			(kt)
 
+/* ktime to jiffies and back */
+#define ktime_to_jiffies(kt)		nsecs_to_jiffies(kt)
+#define jiffies_to_ktime(j)		jiffies_to_nsecs(j)

Hi Tejaswi

You should also add some users of these new helpers.

Andrew

I'll try to explain my scenario. This was observed on a arm64 device. An application registers for a mcast group, and just listens to mcast packets. The connection is setup and mcast packets are being forwarded by the router. Multicast packets are sent out every few minutes. Not a very busy connection.

After some time the router sends out a IGMPv2 query. The max delay in the header is set to 10s. The system starts a timer to send out the response at 9s. But the device suspends and wakes up after 60s.

The response is sent out 50s late.

ftrace logs with boottime trace_clock and my igmp_logs:

4740693      kworker/0:3-395   [000] ..s2  4716.425695: igmp_log: skbaddr=ffffffc156fe6600 daddr=224.0.0.1, id=0, rc=4295217721
4740694      kworker/0:3-395   [000] d.s4  4716.425717: timer_start:timer=ffffffc217763140 function=igmp_timer_expire expires=4295218678 [timeout=957] flags=0x00000000

timer set for 9.57 seconds. 957 jiffies

< device suspends >

< wakes up after ~60s >

4781289           <idle>-0     [000] .ns2  4779.170886: timer_expire_entry: timer=ffffffc217763140 function=igmp_timer_expire now=4295218678
4781290           <idle>-0     [000] .ns2  4779.171045: igmp_log: skbaddr=ffffffc1559d0e00 daddr=227.226.228.225, id=0, rc=4295218678

Since the response was delayed, mcast packets are not forwarded by the router.

Quoted 20 lines · Andrew Lunn

Quoted 10 lines · Tejaswi Tanikella

Signed-off-by Tejaswi Tanikella <[email protected]>

---
 include/linux/ktime.h | 4 ++++
 1 file changed, 4 insertions(+)

include/linux/ktime.h +0 -0

index 5b9fddb..4881483 100644

include/linux/ktime.h static inline ktime_t timeval_to_ktime(struct timeval tv) · line 101

Quoted 6 lines

include/linux/ktime.h +3 -0

@@ -96,6 +96,10 @@ static inline ktime_t timeval_to_ktime(struct timeval tv)
 /* Convert ktime_t to nanoseconds - NOP in the scalar storage format: */
 #define ktime_to_ns(kt)			(kt)
 
+/* ktime to jiffies and back */
+#define ktime_to_jiffies(kt)		nsecs_to_jiffies(kt)
+#define jiffies_to_ktime(j)		jiffies_to_nsecs(j)

Hi Tejaswi

Quoted 3 lines

You should also add some users of these new helpers.

Andrew

Hi Andrew,

I used them in the first patch.

While it has been asleep, it has also been dropping any multicast traffic in the stream. So it does not really matter it has left the group. You were not receiving the packets anyway.

Thing about this from another angle. I have an NTP client running on my laptop, using multicast address 224.0.1.1. I suspend my laptop and walk away for two hours. When i come back, i find that 20 seconds after i suspended it, it resumed and send an group response message. And an hour later, since it was still running, the battery went flat.

It seems to me, the change you are proposing cannot be the default behaviour.

I actually think you need to be looking at some sort of WoL feature. You need the multicast stream data packets to wake you, and you also need to wake up the IGMP query message. And you need to wake up to send the group membership. Does your hardware have this sort of WoL support? You can then explicitly enable this WoL for your application.

Andrew