[dpdk-dev] RTE TIMER LIBRARY

[dpdk-dev] RTE TIMER LIBRARY

[Yeddula, Avinash]

Hello Dpdk-team,

I have few very basic questions.  In the DPDK timer sample application, the periodic timer0's callback function is not called unless "rte_timer_manage()" is called. I understand rte_timer_manage() function looks at all the expired timers in the core and runs them.

Here are my questions:

1.      If I comment out the rte_timer_manage() function in the code, time0_cb function is not getting called at all.  Why is this ?
       I see some documentation saying if they are in Running state the timer would fail, what exactly does this mean, could you please provide more details.

2.      Is it mandatory to call rte_timer_manage to trigger for the dpdk timers  when they expire?

3.      I see this point "The Timer library provides a timer service to Intel(r) DPDK execution units to enable execution of callback functions asynchronously." What does asynchronously means ?

For example, I just have 1 worker thread  in my application and the periodic timer expired and the corresponding call back function has to be executed. Now does the worker thread  stop what it is doing now and starts executing timer's callback function  or how does it work.

I apologies if the questions are very basic and doesn't make sense.

Thanks
-Avinash

PS: I modified the app to make it run only on 1 core. Here is the output when it runs on one core. Let me know if I'm doing something wrong.

[root@localhost timer]# ./build/timer -c 1 -n 1
EAL: Requesting 1024 pages of size 2MB from socket 0
EAL: TSC frequency is ~2399986 KHz
EAL: Master lcore 0 is ready (tid=adaa8c0;cpuset=[0])
EAL: PCI device 0000:00:14.0 on NUMA socket -1
EAL:   probe driver: 8086:1f41 rte_igb_pmd
EAL:   PCI memory mapped at 0x7f4b08e00000
EAL:   PCI memory mapped at 0x7f4b08e20000
PMD: eth_igb_dev_init(): port_id 0 vendorID=0x8086 deviceID=0x1f41
EAL: PCI device 0000:00:14.1 on NUMA socket -1
EAL:   probe driver: 8086:1f41 rte_igb_pmd
EAL:   PCI memory mapped at 0x7f4b08e24000
EAL:   PCI memory mapped at 0x7f4b08e44000
PMD: eth_igb_dev_init(): port_id 1 vendorID=0x8086 deviceID=0x1f41
EAL: PCI device 0000:00:14.2 on NUMA socket -1
EAL:   probe driver: 8086:1f41 rte_igb_pmd
EAL:   PCI memory mapped at 0x7f4b08e48000
EAL:   PCI memory mapped at 0x7f4b08e68000
PMD: eth_igb_dev_init(): port_id 2 vendorID=0x8086 deviceID=0x1f41
EAL: PCI device 0000:00:14.3 on NUMA socket -1
EAL:   probe driver: 8086:1f41 rte_igb_pmd
EAL:   Not managed by a supported kernel driver, skipped
Starting mainloop on core 0
timer0_cb() on lcore 0
timer0_cb() on lcore 0
timer0_cb() on lcore 0

Re: [dpdk-dev] RTE TIMER LIBRARY

[Bruce Richardson]

On Thu, Jun 18, 2015 at 08:57:45PM -0400, Yeddula, Avinash wrote:
> Hello Dpdk-team,
> 
> I have few very basic questions.  In the DPDK timer sample application, the periodic timer0's callback function is not called unless "rte_timer_manage()" is called. I understand rte_timer_manage() function looks at all the expired timers in the core and runs them.
> 
> Here are my questions:
> 
> 1.      If I comment out the rte_timer_manage() function in the code, time0_cb function is not getting called at all.  Why is this ?
>        I see some documentation saying if they are in Running state the timer would fail, what exactly does this mean, could you please provide more details.
> 
> 2.      Is it mandatory to call rte_timer_manage to trigger for the dpdk timers  when they expire?
> 
> 3.      I see this point "The Timer library provides a timer service to Intel(r) DPDK execution units to enable execution of callback functions asynchronously." What does asynchronously means ?
> 
> For example, I just have 1 worker thread  in my application and the periodic timer expired and the corresponding call back function has to be executed. Now does the worker thread  stop what it is doing now and starts executing timer's callback function  or how does it work.
> 
> I apologies if the questions are very basic and doesn't make sense.
> 
> Thanks
> -Avinash
>
The timers in DPDK are asynchronous in the sense that the timers will fire at a
time given by the timeout command, and not at a guaranteed point in the code
execution. On the other hand, they are also synchronous in that the timers' callbacks
can only be executed at specific designated points in the code - the rte_timer_manage
calls.

This arrangement allows greater predictability of execution - as timer callbacks
fire at known points - and helps apps avoid race conditions e.g. if you are modifying
a setting on a port/queue in a callback, you can know that the core on which this
callback is executing is not performing packet IO on that port/queue at the same
time.

So overall, the rte_timer_manage functions are required in order to have timer
callbacks executed - since it is that function which itself triggers the callbacks,
at known points in the code.

/Bruce

Re: [dpdk-dev] RTE TIMER LIBRARY

[Stephen Hemminger]

Read the code??