Performance evaluation of Linux CAN-related system calls

Transcription

1 Performance evaluation of Linux CAN-related system calls Michal Sojka, Pavel Píša, Zdeněk Hanzálek Czech Technical University in Prague, Faculty of Electrical Engineering SESAMO Security and Safety Modelling Workshop on Factory Communication Systems (WFCS) May 7, 2014 Toulouse, France M. Sojka, et al. Linux CAN-related system calls WFCS / 15

2 Introduction Motivation Security of on-board networks M. Sojka, et al. Linux CAN-related system calls WFCS / 15

3 Introduction Motivation Security of on-board networks IDS Intrusion detection system M. Sojka, et al. Linux CAN-related system calls WFCS / 15

4 Introduction Motivation Security of on-board networks IDS Intrusion detection system Needs to be "tuned" to avoid false positives M. Sojka, et al. Linux CAN-related system calls WFCS / 15

5 Introduction Motivation Security of on-board networks IDS Intrusion detection system Needs to be "tuned" to avoid false positives Data from outdoor experiments M. Sojka, et al. Linux CAN-related system calls WFCS / 15

6 Introduction Motivation Security of on-board networks IDS Intrusion detection system Needs to be "tuned" to avoid false positives Data from outdoor experiments Logging M. Sojka, et al. Linux CAN-related system calls WFCS / 15

7 Introduction Motivation Security of on-board networks IDS Intrusion detection system Needs to be "tuned" to avoid false positives Data from outdoor experiments Logging M. Sojka, et al. Linux CAN-related system calls WFCS / 15

8 Introduction Requirements All CAN traffic from a car has to be logged in real-time. Up to 20 CAN buses needs to be logged simultaneously. Logs are analyzed off-line. M. Sojka, et al. Linux CAN-related system calls WFCS / 15

9 Introduction Outline CAN support in Linux Virtual CAN interface method Gateway-based method Low-latency sockets Conclusion M. Sojka, et al. Linux CAN-related system calls WFCS / 15

10 CAN support in Linux Linux networking subsystem RX/TX path overheads: Interrupt Application (user space) Socket layer (can/raw.c) Socket, protocol, Qdisc System call We are interested in the interface between applications and the kernel. Kernel Protocol layer (af_can.c) Queuing discipline Poll ISR CAN controller (HW) Device driver M. Sojka, et al. Linux CAN-related system calls WFCS / 15

11 CAN support in Linux Linux networking subsystem RX/TX path overheads: Interrupt Application (user space) Socket layer (can/raw.c) Socket, protocol, Qdisc System call We are interested in the interface between applications and the kernel. Kernel Protocol layer (af_can.c) Queuing discipline Poll ISR CAN controller (HW) Device driver M. Sojka, et al. Linux CAN-related system calls WFCS / 15

12 CAN support in Linux Linux networking API Sockets: Raw CAN sockets most common Memory mapped socket (PF PACKET) Non-blocking I/O (O NONBLOCK) New Low-latency sockets (Linux /2013) Operations on sockets: read/write (send/recv) sendmmsg/recvmmsg M. Sojka, et al. Linux CAN-related system calls WFCS / 15

13 CAN support in Linux Memory-mapped sockets PF PACKET Reception no system call wait: poll() Transmission Notify the kernel about ready messages. Kernel Application (user space) Protocol & socket layer (af_packet.c) Queuing discipline Poll ISR CAN controller (HW) Device driver M. Sojka, et al. Linux CAN-related system calls WFCS / 15

14 Virtual CAN interface method Single/multiple messages per system call Comparison of recv() and recvmmsg() system calls Use of virtual CAN interface (vcan) Application Kernel TX socket RX socket read/recv write/send RX queue af_can.c, raw.c vcan Experiment 1. Send X messages kernel queues them in an RX queue 2. Receive X messages from the RX queue M. Sojka, et al. Linux CAN-related system calls WFCS / 15

15 Virtual CAN interface method With or without mmsg results Time [ms] PC read recvmmsg write sendmmsg Frames [ 1000] recvmmsg 7% faster sendmmsg 12% faster Embedded system (MPC5200B) Time [ms] read recvmmsg write sendmmsg Frames [ 1000] recvmmsg 19% faster sendmmsg 35% faster M. Sojka, et al. Linux CAN-related system calls WFCS / 15

16 Gateway-based method Embedded system used for experiments PowerPC Freescale MPC5200B 400 Mhz 128 MB RAM on-chip CAN controller Linux M. Sojka, et al. Linux CAN-related system calls WFCS / 15

17 Gateway-based method Gateway-based experiments CAN bus 0 msg 1 CAN gateway (Linux) CAN bus 1 msg 1' GW latency Duration time Total latency RX timestamp 1 RX timestamp 2 M. Sojka, et al. Linux CAN-related system calls WFCS / 15

18 Gateway-based method Gateway-based experiments CAN bus 0 msg 1 CAN gateway (Linux) CAN bus 1 msg 1' GW latency Duration time Total latency RX timestamp 1 RX timestamp 2 M. Sojka, et al. Linux CAN-related system calls WFCS / 15

19 Gateway-based method Gateway-based experiments CAN bus 0 msg 1 CAN gateway (Linux) CAN bus 1 msg 1' GW latency Duration time Total latency RX timestamp 1 RX timestamp 2 M. Sojka, et al. Linux CAN-related system calls WFCS / 15

20 Gateway-based method Gateway-based experiments CAN bus 0 msg 1 CAN gateway (Linux) CAN bus 1 msg 1' GW latency Duration time Total latency RX timestamp 1 RX timestamp 2 M. Sojka, et al. Linux CAN-related system calls WFCS / 15

21 Gateway-based method Gateway-based experiments CAN bus 0 msg 1 CAN gateway (Linux) CAN bus 1 msg 1' GW latency Duration time Total latency RX timestamp 1 RX timestamp 2 M. Sojka, et al. Linux CAN-related system calls WFCS / 15

22 Gateway-based method Gateway-based experiments Different implemntations CAN bus 0 msg 1 CAN gateway (Linux) CAN bus 1 msg 1' GW latency Duration time Total latency RX timestamp 1 RX timestamp 2 M. Sojka, et al. Linux CAN-related system calls WFCS / 15

23 Gateway-based method Results one message at a time GW latency [µs] mmsg-mmsg mmap-mmap mmap-write read-write Blocking system calls Non-blocking/busy waiting Linux kernel & RTEMS (no Linux) M. Sojka, et al. Linux CAN-related system calls WFCS / 15

24 Gateway-based method Results one message at a time GW latency [µs] readbusynoirq-write readbusy-write readnb-write mmapbusy-write mmapbusy-mmap mmsg-mmsg mmap-mmap mmap-write read-write readnb-mmap Blocking system calls Non-blocking/busy waiting Linux kernel & RTEMS (no Linux) M. Sojka, et al. Linux CAN-related system calls WFCS / 15

25 Gateway-based method Results one message at a time GW latency [µs] mmsg-mmsg mmap-mmap mmap-write read-write readnb-mmap readbusynoirq-write readbusy-write readnb-write mmapbusy-write mmapbusy-mmap kernel rtems Blocking system calls Non-blocking/busy waiting Linux kernel & RTEMS (no Linux) M. Sojka, et al. Linux CAN-related system calls WFCS / 15

26 Gateway-based method Results periodic messages Bitrate: 1 Mbps, Payload: 2 bytes Median of 3200 frames 70 µs 70 µs period GW latency [µs] ± packet loss byte frame length Frame period [µs] read-write M. Sojka, et al. Linux CAN-related system calls WFCS / 15

27 Gateway-based method Results periodic messages Bitrate: 1 Mbps, Payload: 2 bytes Median of 3200 frames 70 µs 70 µs period GW latency [µs] ± packet loss byte frame length Frame period [µs] read-write mmap-write M. Sojka, et al. Linux CAN-related system calls WFCS / 15

28 Gateway-based method Results periodic messages Bitrate: 1 Mbps, Payload: 2 bytes Median of 3200 frames 70 µs 70 µs period GW latency [µs] ± packet loss byte frame length Frame period [µs] read-write mmap-write mmsg-mmsg mmap-mmap M. Sojka, et al. Linux CAN-related system calls WFCS / 15

29 Gateway-based method Results periodic messages Bitrate: 1 Mbps, Payload: 2 bytes Median of 3200 frames 70 µs 70 µs period GW latency [µs] ± packet loss byte frame length Frame period [µs] read-write mmap-write mmsg-mmsg mmap-mmap kernel M. Sojka, et al. Linux CAN-related system calls WFCS / 15

30 Gateway-based method Results periodic messages Bitrate: 1 Mbps, Payload: 2 bytes Median of 3200 frames 70 µs 70 µs period GW latency [µs] ± packet loss byte frame length Frame period [µs] M. Sojka, et al. Linux CAN-related system calls WFCS / 15

31 Low-latency sockets Low-latency sockets aka Busy polling sockets Network frame frame frame frame frame CPU read() IRQ Not many CAN controllers offer interrupt coalescing. Avoiding IRQ overhead brings only negligible benefits. Busy polling is not useful for CAN. M. Sojka, et al. Linux CAN-related system calls WFCS / 15

32 Low-latency sockets Low-latency sockets aka Busy polling sockets Network frame frame frame frame frame IRQ CPU read() read() Not many CAN controllers offer interrupt coalescing. Avoiding IRQ overhead brings only negligible benefits. Busy polling is not useful for CAN. M. Sojka, et al. Linux CAN-related system calls WFCS / 15

33 Low-latency sockets Low-latency sockets aka Busy polling sockets Network frame frame frame frame frame minimum inter-irq time ( 100 µs) IRQ CPU read() read() IRQ Not many CAN controllers offer interrupt coalescing. Avoiding IRQ overhead brings only negligible benefits. Busy polling is not useful for CAN. M. Sojka, et al. Linux CAN-related system calls WFCS / 15

34 Low-latency sockets Low-latency sockets aka Busy polling sockets Network frame frame frame frame frame minimum inter-irq time ( 100 µs) IRQ CPU read() read() IRQ busy waiting Not many CAN controllers offer interrupt coalescing. Avoiding IRQ overhead brings only negligible benefits. Busy polling is not useful for CAN. M. Sojka, et al. Linux CAN-related system calls WFCS / 15

35 Low-latency sockets Low-latency sockets aka Busy polling sockets Network frame frame frame frame frame minimum inter-irq time ( 100 µs) IRQ CPU read() read() busy waiting read() timeout IRQ Not many CAN controllers offer interrupt coalescing. Avoiding IRQ overhead brings only negligible benefits. Busy polling is not useful for CAN. M. Sojka, et al. Linux CAN-related system calls WFCS / 15

36 Low-latency sockets Low-latency sockets aka Busy polling sockets Network frame frame frame frame frame minimum inter-irq time ( 100 µs) IRQ CPU read() read() busy waiting read() timeout IRQ Not many CAN controllers offer interrupt coalescing. Avoiding IRQ overhead brings only negligible benefits. Busy polling is not useful for CAN. M. Sojka, et al. Linux CAN-related system calls WFCS / 15

37 Conclusion Conclusion Experimental performance analysis of Linux network-related user space APIs in the context of CAN bus. Interfaces that can handle multiple messages per system call (recvmmsg(), memory-mapped sockets) are superior to single-message system calls (read()). recvmmsg() is simpler to use than memory-mapped sockets. New Low-latency sockets are not useful for CAN applications. Task wake-up latency. M. Sojka, et al. Linux CAN-related system calls WFCS / 15

38 Conclusion Conclusion Experimental performance analysis of Linux network-related user space APIs in the context of CAN bus. Interfaces that can handle multiple messages per system call (recvmmsg(), memory-mapped sockets) are superior to single-message system calls (read()). recvmmsg() is simpler to use than memory-mapped sockets. New Low-latency sockets are not useful for CAN applications. Task wake-up latency. Thank you! M. Sojka, et al. Linux CAN-related system calls WFCS / 15